2019-05-19 20:08:55 +08:00
|
|
|
// SPDX-License-Identifier: GPL-2.0-only
|
2005-04-17 06:20:36 +08:00
|
|
|
/*
|
|
|
|
|
* mm/mmap.c
|
|
|
|
|
*
|
|
|
|
|
* Written by obz.
|
|
|
|
|
*
|
2009-01-05 22:06:29 +08:00
|
|
|
* Address space accounting code <alan@lxorguk.ukuu.org.uk>
|
2005-04-17 06:20:36 +08:00
|
|
|
*/
|
|
|
|
|
|
2014-06-07 05:38:30 +08:00
|
|
|
#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
|
|
|
|
|
|
2013-04-30 06:08:33 +08:00
|
|
|
#include <linux/kernel.h>
|
2005-04-17 06:20:36 +08:00
|
|
|
#include <linux/slab.h>
|
2007-10-17 14:29:23 +08:00
|
|
|
#include <linux/backing-dev.h>
|
2005-04-17 06:20:36 +08:00
|
|
|
#include <linux/mm.h>
|
2022-01-15 06:06:07 +08:00
|
|
|
#include <linux/mm_inline.h>
|
mm: per-thread vma caching
This patch is a continuation of efforts trying to optimize find_vma(),
avoiding potentially expensive rbtree walks to locate a vma upon faults.
The original approach (https://lkml.org/lkml/2013/11/1/410), where the
largest vma was also cached, ended up being too specific and random,
thus further comparison with other approaches were needed. There are
two things to consider when dealing with this, the cache hit rate and
the latency of find_vma(). Improving the hit-rate does not necessarily
translate in finding the vma any faster, as the overhead of any fancy
caching schemes can be too high to consider.
We currently cache the last used vma for the whole address space, which
provides a nice optimization, reducing the total cycles in find_vma() by
up to 250%, for workloads with good locality. On the other hand, this
simple scheme is pretty much useless for workloads with poor locality.
Analyzing ebizzy runs shows that, no matter how many threads are
running, the mmap_cache hit rate is less than 2%, and in many situations
below 1%.
The proposed approach is to replace this scheme with a small per-thread
cache, maximizing hit rates at a very low maintenance cost.
Invalidations are performed by simply bumping up a 32-bit sequence
number. The only expensive operation is in the rare case of a seq
number overflow, where all caches that share the same address space are
flushed. Upon a miss, the proposed replacement policy is based on the
page number that contains the virtual address in question. Concretely,
the following results are seen on an 80 core, 8 socket x86-64 box:
1) System bootup: Most programs are single threaded, so the per-thread
scheme does improve ~50% hit rate by just adding a few more slots to
the cache.
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 50.61% | 19.90 |
| patched | 73.45% | 13.58 |
+----------------+----------+------------------+
2) Kernel build: This one is already pretty good with the current
approach as we're dealing with good locality.
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 75.28% | 11.03 |
| patched | 88.09% | 9.31 |
+----------------+----------+------------------+
3) Oracle 11g Data Mining (4k pages): Similar to the kernel build workload.
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 70.66% | 17.14 |
| patched | 91.15% | 12.57 |
+----------------+----------+------------------+
4) Ebizzy: There's a fair amount of variation from run to run, but this
approach always shows nearly perfect hit rates, while baseline is just
about non-existent. The amounts of cycles can fluctuate between
anywhere from ~60 to ~116 for the baseline scheme, but this approach
reduces it considerably. For instance, with 80 threads:
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 1.06% | 91.54 |
| patched | 99.97% | 14.18 |
+----------------+----------+------------------+
[akpm@linux-foundation.org: fix nommu build, per Davidlohr]
[akpm@linux-foundation.org: document vmacache_valid() logic]
[akpm@linux-foundation.org: attempt to untangle header files]
[akpm@linux-foundation.org: add vmacache_find() BUG_ON]
[hughd@google.com: add vmacache_valid_mm() (from Oleg)]
[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: adjust and enhance comments]
Signed-off-by: Davidlohr Bueso <davidlohr@hp.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Reviewed-by: Michel Lespinasse <walken@google.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Tested-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-04-08 06:37:25 +08:00
|
|
|
#include <linux/vmacache.h>
|
2005-04-17 06:20:36 +08:00
|
|
|
#include <linux/shm.h>
|
|
|
|
|
#include <linux/mman.h>
|
|
|
|
|
#include <linux/pagemap.h>
|
|
|
|
|
#include <linux/swap.h>
|
|
|
|
|
#include <linux/syscalls.h>
|
2006-01-12 04:17:46 +08:00
|
|
|
#include <linux/capability.h>
|
2005-04-17 06:20:36 +08:00
|
|
|
#include <linux/init.h>
|
|
|
|
|
#include <linux/file.h>
|
|
|
|
|
#include <linux/fs.h>
|
|
|
|
|
#include <linux/personality.h>
|
|
|
|
|
#include <linux/security.h>
|
|
|
|
|
#include <linux/hugetlb.h>
|
2016-07-27 06:26:15 +08:00
|
|
|
#include <linux/shmem_fs.h>
|
2005-04-17 06:20:36 +08:00
|
|
|
#include <linux/profile.h>
|
2011-10-16 14:01:52 +08:00
|
|
|
#include <linux/export.h>
|
2005-04-17 06:20:36 +08:00
|
|
|
#include <linux/mount.h>
|
|
|
|
|
#include <linux/mempolicy.h>
|
|
|
|
|
#include <linux/rmap.h>
|
mmu-notifiers: core
With KVM/GFP/XPMEM there isn't just the primary CPU MMU pointing to pages.
There are secondary MMUs (with secondary sptes and secondary tlbs) too.
sptes in the kvm case are shadow pagetables, but when I say spte in
mmu-notifier context, I mean "secondary pte". In GRU case there's no
actual secondary pte and there's only a secondary tlb because the GRU
secondary MMU has no knowledge about sptes and every secondary tlb miss
event in the MMU always generates a page fault that has to be resolved by
the CPU (this is not the case of KVM where the a secondary tlb miss will
walk sptes in hardware and it will refill the secondary tlb transparently
to software if the corresponding spte is present). The same way
zap_page_range has to invalidate the pte before freeing the page, the spte
(and secondary tlb) must also be invalidated before any page is freed and
reused.
Currently we take a page_count pin on every page mapped by sptes, but that
means the pages can't be swapped whenever they're mapped by any spte
because they're part of the guest working set. Furthermore a spte unmap
event can immediately lead to a page to be freed when the pin is released
(so requiring the same complex and relatively slow tlb_gather smp safe
logic we have in zap_page_range and that can be avoided completely if the
spte unmap event doesn't require an unpin of the page previously mapped in
the secondary MMU).
The mmu notifiers allow kvm/GRU/XPMEM to attach to the tsk->mm and know
when the VM is swapping or freeing or doing anything on the primary MMU so
that the secondary MMU code can drop sptes before the pages are freed,
avoiding all page pinning and allowing 100% reliable swapping of guest
physical address space. Furthermore it avoids the code that teardown the
mappings of the secondary MMU, to implement a logic like tlb_gather in
zap_page_range that would require many IPI to flush other cpu tlbs, for
each fixed number of spte unmapped.
To make an example: if what happens on the primary MMU is a protection
downgrade (from writeable to wrprotect) the secondary MMU mappings will be
invalidated, and the next secondary-mmu-page-fault will call
get_user_pages and trigger a do_wp_page through get_user_pages if it
called get_user_pages with write=1, and it'll re-establishing an updated
spte or secondary-tlb-mapping on the copied page. Or it will setup a
readonly spte or readonly tlb mapping if it's a guest-read, if it calls
get_user_pages with write=0. This is just an example.
This allows to map any page pointed by any pte (and in turn visible in the
primary CPU MMU), into a secondary MMU (be it a pure tlb like GRU, or an
full MMU with both sptes and secondary-tlb like the shadow-pagetable layer
with kvm), or a remote DMA in software like XPMEM (hence needing of
schedule in XPMEM code to send the invalidate to the remote node, while no
need to schedule in kvm/gru as it's an immediate event like invalidating
primary-mmu pte).
At least for KVM without this patch it's impossible to swap guests
reliably. And having this feature and removing the page pin allows
several other optimizations that simplify life considerably.
Dependencies:
1) mm_take_all_locks() to register the mmu notifier when the whole VM
isn't doing anything with "mm". This allows mmu notifier users to keep
track if the VM is in the middle of the invalidate_range_begin/end
critical section with an atomic counter incraese in range_begin and
decreased in range_end. No secondary MMU page fault is allowed to map
any spte or secondary tlb reference, while the VM is in the middle of
range_begin/end as any page returned by get_user_pages in that critical
section could later immediately be freed without any further
->invalidate_page notification (invalidate_range_begin/end works on
ranges and ->invalidate_page isn't called immediately before freeing
the page). To stop all page freeing and pagetable overwrites the
mmap_sem must be taken in write mode and all other anon_vma/i_mmap
locks must be taken too.
2) It'd be a waste to add branches in the VM if nobody could possibly
run KVM/GRU/XPMEM on the kernel, so mmu notifiers will only enabled if
CONFIG_KVM=m/y. In the current kernel kvm won't yet take advantage of
mmu notifiers, but this already allows to compile a KVM external module
against a kernel with mmu notifiers enabled and from the next pull from
kvm.git we'll start using them. And GRU/XPMEM will also be able to
continue the development by enabling KVM=m in their config, until they
submit all GRU/XPMEM GPLv2 code to the mainline kernel. Then they can
also enable MMU_NOTIFIERS in the same way KVM does it (even if KVM=n).
This guarantees nobody selects MMU_NOTIFIER=y if KVM and GRU and XPMEM
are all =n.
The mmu_notifier_register call can fail because mm_take_all_locks may be
interrupted by a signal and return -EINTR. Because mmu_notifier_reigster
is used when a driver startup, a failure can be gracefully handled. Here
an example of the change applied to kvm to register the mmu notifiers.
Usually when a driver startups other allocations are required anyway and
-ENOMEM failure paths exists already.
struct kvm *kvm_arch_create_vm(void)
{
struct kvm *kvm = kzalloc(sizeof(struct kvm), GFP_KERNEL);
+ int err;
if (!kvm)
return ERR_PTR(-ENOMEM);
INIT_LIST_HEAD(&kvm->arch.active_mmu_pages);
+ kvm->arch.mmu_notifier.ops = &kvm_mmu_notifier_ops;
+ err = mmu_notifier_register(&kvm->arch.mmu_notifier, current->mm);
+ if (err) {
+ kfree(kvm);
+ return ERR_PTR(err);
+ }
+
return kvm;
}
mmu_notifier_unregister returns void and it's reliable.
The patch also adds a few needed but missing includes that would prevent
kernel to compile after these changes on non-x86 archs (x86 didn't need
them by luck).
[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: fix mm/filemap_xip.c build]
[akpm@linux-foundation.org: fix mm/mmu_notifier.c build]
Signed-off-by: Andrea Arcangeli <andrea@qumranet.com>
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Christoph Lameter <cl@linux-foundation.org>
Cc: Jack Steiner <steiner@sgi.com>
Cc: Robin Holt <holt@sgi.com>
Cc: Nick Piggin <npiggin@suse.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Kanoj Sarcar <kanojsarcar@yahoo.com>
Cc: Roland Dreier <rdreier@cisco.com>
Cc: Steve Wise <swise@opengridcomputing.com>
Cc: Avi Kivity <avi@qumranet.com>
Cc: Hugh Dickins <hugh@veritas.com>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Cc: Anthony Liguori <aliguori@us.ibm.com>
Cc: Chris Wright <chrisw@redhat.com>
Cc: Marcelo Tosatti <marcelo@kvack.org>
Cc: Eric Dumazet <dada1@cosmosbay.com>
Cc: "Paul E. McKenney" <paulmck@us.ibm.com>
Cc: Izik Eidus <izike@qumranet.com>
Cc: Anthony Liguori <aliguori@us.ibm.com>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-07-29 06:46:29 +08:00
|
|
|
#include <linux/mmu_notifier.h>
|
2014-08-07 07:06:36 +08:00
|
|
|
#include <linux/mmdebug.h>
|
perf: Do the big rename: Performance Counters -> Performance Events
Bye-bye Performance Counters, welcome Performance Events!
In the past few months the perfcounters subsystem has grown out its
initial role of counting hardware events, and has become (and is
becoming) a much broader generic event enumeration, reporting, logging,
monitoring, analysis facility.
Naming its core object 'perf_counter' and naming the subsystem
'perfcounters' has become more and more of a misnomer. With pending
code like hw-breakpoints support the 'counter' name is less and
less appropriate.
All in one, we've decided to rename the subsystem to 'performance
events' and to propagate this rename through all fields, variables
and API names. (in an ABI compatible fashion)
The word 'event' is also a bit shorter than 'counter' - which makes
it slightly more convenient to write/handle as well.
Thanks goes to Stephane Eranian who first observed this misnomer and
suggested a rename.
User-space tooling and ABI compatibility is not affected - this patch
should be function-invariant. (Also, defconfigs were not touched to
keep the size down.)
This patch has been generated via the following script:
FILES=$(find * -type f | grep -vE 'oprofile|[^K]config')
sed -i \
-e 's/PERF_EVENT_/PERF_RECORD_/g' \
-e 's/PERF_COUNTER/PERF_EVENT/g' \
-e 's/perf_counter/perf_event/g' \
-e 's/nb_counters/nb_events/g' \
-e 's/swcounter/swevent/g' \
-e 's/tpcounter_event/tp_event/g' \
$FILES
for N in $(find . -name perf_counter.[ch]); do
M=$(echo $N | sed 's/perf_counter/perf_event/g')
mv $N $M
done
FILES=$(find . -name perf_event.*)
sed -i \
-e 's/COUNTER_MASK/REG_MASK/g' \
-e 's/COUNTER/EVENT/g' \
-e 's/\<event\>/event_id/g' \
-e 's/counter/event/g' \
-e 's/Counter/Event/g' \
$FILES
... to keep it as correct as possible. This script can also be
used by anyone who has pending perfcounters patches - it converts
a Linux kernel tree over to the new naming. We tried to time this
change to the point in time where the amount of pending patches
is the smallest: the end of the merge window.
Namespace clashes were fixed up in a preparatory patch - and some
stylistic fallout will be fixed up in a subsequent patch.
( NOTE: 'counters' are still the proper terminology when we deal
with hardware registers - and these sed scripts are a bit
over-eager in renaming them. I've undone some of that, but
in case there's something left where 'counter' would be
better than 'event' we can undo that on an individual basis
instead of touching an otherwise nicely automated patch. )
Suggested-by: Stephane Eranian <eranian@google.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Paul Mackerras <paulus@samba.org>
Reviewed-by: Arjan van de Ven <arjan@linux.intel.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: David Howells <dhowells@redhat.com>
Cc: Kyle McMartin <kyle@mcmartin.ca>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: <linux-arch@vger.kernel.org>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-09-21 18:02:48 +08:00
|
|
|
#include <linux/perf_event.h>
|
2010-10-30 14:54:44 +08:00
|
|
|
#include <linux/audit.h>
|
2011-01-14 07:46:59 +08:00
|
|
|
#include <linux/khugepaged.h>
|
uprobes, mm, x86: Add the ability to install and remove uprobes breakpoints
Add uprobes support to the core kernel, with x86 support.
This commit adds the kernel facilities, the actual uprobes
user-space ABI and perf probe support comes in later commits.
General design:
Uprobes are maintained in an rb-tree indexed by inode and offset
(the offset here is from the start of the mapping). For a unique
(inode, offset) tuple, there can be at most one uprobe in the
rb-tree.
Since the (inode, offset) tuple identifies a unique uprobe, more
than one user may be interested in the same uprobe. This provides
the ability to connect multiple 'consumers' to the same uprobe.
Each consumer defines a handler and a filter (optional). The
'handler' is run every time the uprobe is hit, if it matches the
'filter' criteria.
The first consumer of a uprobe causes the breakpoint to be
inserted at the specified address and subsequent consumers are
appended to this list. On subsequent probes, the consumer gets
appended to the existing list of consumers. The breakpoint is
removed when the last consumer unregisters. For all other
unregisterations, the consumer is removed from the list of
consumers.
Given a inode, we get a list of the mms that have mapped the
inode. Do the actual registration if mm maps the page where a
probe needs to be inserted/removed.
We use a temporary list to walk through the vmas that map the
inode.
- The number of maps that map the inode, is not known before we
walk the rmap and keeps changing.
- extending vm_area_struct wasn't recommended, it's a
size-critical data structure.
- There can be more than one maps of the inode in the same mm.
We add callbacks to the mmap methods to keep an eye on text vmas
that are of interest to uprobes. When a vma of interest is mapped,
we insert the breakpoint at the right address.
Uprobe works by replacing the instruction at the address defined
by (inode, offset) with the arch specific breakpoint
instruction. We save a copy of the original instruction at the
uprobed address.
This is needed for:
a. executing the instruction out-of-line (xol).
b. instruction analysis for any subsequent fixups.
c. restoring the instruction back when the uprobe is unregistered.
We insert or delete a breakpoint instruction, and this
breakpoint instruction is assumed to be the smallest instruction
available on the platform. For fixed size instruction platforms
this is trivially true, for variable size instruction platforms
the breakpoint instruction is typically the smallest (often a
single byte).
Writing the instruction is done by COWing the page and changing
the instruction during the copy, this even though most platforms
allow atomic writes of the breakpoint instruction. This also
mirrors the behaviour of a ptrace() memory write to a PRIVATE
file map.
The core worker is derived from KSM's replace_page() logic.
In essence, similar to KSM:
a. allocate a new page and copy over contents of the page that
has the uprobed vaddr
b. modify the copy and insert the breakpoint at the required
address
c. switch the original page with the copy containing the
breakpoint
d. flush page tables.
replace_page() is being replicated here because of some minor
changes in the type of pages and also because Hugh Dickins had
plans to improve replace_page() for KSM specific work.
Instruction analysis on x86 is based on instruction decoder and
determines if an instruction can be probed and determines the
necessary fixups after singlestep. Instruction analysis is done
at probe insertion time so that we avoid having to repeat the
same analysis every time a probe is hit.
A lot of code here is due to the improvement/suggestions/inputs
from Peter Zijlstra.
Changelog:
(v10):
- Add code to clear REX.B prefix as suggested by Denys Vlasenko
and Masami Hiramatsu.
(v9):
- Use insn_offset_modrm as suggested by Masami Hiramatsu.
(v7):
Handle comments from Peter Zijlstra:
- Dont take reference to inode. (expect inode to uprobe_register to be sane).
- Use PTR_ERR to set the return value.
- No need to take reference to inode.
- use PTR_ERR to return error value.
- register and uprobe_unregister share code.
(v5):
- Modified del_consumer as per comments from Peter.
- Drop reference to inode before dropping reference to uprobe.
- Use i_size_read(inode) instead of inode->i_size.
- Ensure uprobe->consumers is NULL, before __uprobe_unregister() is called.
- Includes errno.h as recommended by Stephen Rothwell to fix a build issue
on sparc defconfig
- Remove restrictions while unregistering.
- Earlier code leaked inode references under some conditions while
registering/unregistering.
- Continue the vma-rmap walk even if the intermediate vma doesnt
meet the requirements.
- Validate the vma found by find_vma before inserting/removing the
breakpoint
- Call del_consumer under mutex_lock.
- Use hash locks.
- Handle mremap.
- Introduce find_least_offset_node() instead of close match logic in
find_uprobe
- Uprobes no more depends on MM_OWNER; No reference to task_structs
while inserting/removing a probe.
- Uses read_mapping_page instead of grab_cache_page so that the pages
have valid content.
- pass NULL to get_user_pages for the task parameter.
- call SetPageUptodate on the new page allocated in write_opcode.
- fix leaking a reference to the new page under certain conditions.
- Include Instruction Decoder if Uprobes gets defined.
- Remove const attributes for instruction prefix arrays.
- Uses mm_context to know if the application is 32 bit.
Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Also-written-by: Jim Keniston <jkenisto@us.ibm.com>
Reviewed-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Roland McGrath <roland@hack.frob.com>
Cc: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
Cc: Anton Arapov <anton@redhat.com>
Cc: Ananth N Mavinakayanahalli <ananth@in.ibm.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Denys Vlasenko <vda.linux@googlemail.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Linux-mm <linux-mm@kvack.org>
Link: http://lkml.kernel.org/r/20120209092642.GE16600@linux.vnet.ibm.com
[ Made various small edits to the commit log ]
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2012-02-09 17:26:42 +08:00
|
|
|
#include <linux/uprobes.h>
|
2012-12-12 08:01:38 +08:00
|
|
|
#include <linux/rbtree_augmented.h>
|
2013-04-30 06:08:12 +08:00
|
|
|
#include <linux/notifier.h>
|
|
|
|
|
#include <linux/memory.h>
|
2014-06-07 05:38:30 +08:00
|
|
|
#include <linux/printk.h>
|
2015-09-05 06:46:24 +08:00
|
|
|
#include <linux/userfaultfd_k.h>
|
2016-02-03 08:57:43 +08:00
|
|
|
#include <linux/moduleparam.h>
|
mm/core, x86/mm/pkeys: Add execute-only protection keys support
Protection keys provide new page-based protection in hardware.
But, they have an interesting attribute: they only affect data
accesses and never affect instruction fetches. That means that
if we set up some memory which is set as "access-disabled" via
protection keys, we can still execute from it.
This patch uses protection keys to set up mappings to do just that.
If a user calls:
mmap(..., PROT_EXEC);
or
mprotect(ptr, sz, PROT_EXEC);
(note PROT_EXEC-only without PROT_READ/WRITE), the kernel will
notice this, and set a special protection key on the memory. It
also sets the appropriate bits in the Protection Keys User Rights
(PKRU) register so that the memory becomes unreadable and
unwritable.
I haven't found any userspace that does this today. With this
facility in place, we expect userspace to move to use it
eventually. Userspace _could_ start doing this today. Any
PROT_EXEC calls get converted to PROT_READ inside the kernel, and
would transparently be upgraded to "true" PROT_EXEC with this
code. IOW, userspace never has to do any PROT_EXEC runtime
detection.
This feature provides enhanced protection against leaking
executable memory contents. This helps thwart attacks which are
attempting to find ROP gadgets on the fly.
But, the security provided by this approach is not comprehensive.
The PKRU register which controls access permissions is a normal
user register writable from unprivileged userspace. An attacker
who can execute the 'wrpkru' instruction can easily disable the
protection provided by this feature.
The protection key that is used for execute-only support is
permanently dedicated at compile time. This is fine for now
because there is currently no API to set a protection key other
than this one.
Despite there being a constant PKRU value across the entire
system, we do not set it unless this feature is in use in a
process. That is to preserve the PKRU XSAVE 'init state',
which can lead to faster context switches.
PKRU *is* a user register and the kernel is modifying it. That
means that code doing:
pkru = rdpkru()
pkru |= 0x100;
mmap(..., PROT_EXEC);
wrpkru(pkru);
could lose the bits in PKRU that enforce execute-only
permissions. To avoid this, we suggest avoiding ever calling
mmap() or mprotect() when the PKRU value is expected to be
unstable.
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Borislav Petkov <bp@suse.de>
Cc: Brian Gerst <brgerst@gmail.com>
Cc: Chen Gang <gang.chen.5i5j@gmail.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Dave Hansen <dave@sr71.net>
Cc: David Hildenbrand <dahi@linux.vnet.ibm.com>
Cc: Denys Vlasenko <dvlasenk@redhat.com>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Kees Cook <keescook@chromium.org>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Konstantin Khlebnikov <koct9i@gmail.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Piotr Kwapulinski <kwapulinski.piotr@gmail.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Stephen Smalley <sds@tycho.nsa.gov>
Cc: Vladimir Murzin <vladimir.murzin@arm.com>
Cc: Will Deacon <will.deacon@arm.com>
Cc: keescook@google.com
Cc: linux-kernel@vger.kernel.org
Cc: linux-mm@kvack.org
Link: http://lkml.kernel.org/r/20160212210240.CB4BB5CA@viggo.jf.intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-02-13 05:02:40 +08:00
|
|
|
#include <linux/pkeys.h>
|
2017-09-07 07:25:00 +08:00
|
|
|
#include <linux/oom.h>
|
2019-04-19 08:50:52 +08:00
|
|
|
#include <linux/sched/mm.h>
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2016-12-25 03:46:01 +08:00
|
|
|
#include <linux/uaccess.h>
|
2005-04-17 06:20:36 +08:00
|
|
|
#include <asm/cacheflush.h>
|
|
|
|
|
#include <asm/tlb.h>
|
2007-05-03 01:27:14 +08:00
|
|
|
#include <asm/mmu_context.h>
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2020-04-02 12:09:13 +08:00
|
|
|
#define CREATE_TRACE_POINTS
|
|
|
|
|
#include <trace/events/mmap.h>
|
|
|
|
|
|
2008-07-24 12:27:10 +08:00
|
|
|
#include "internal.h"
|
|
|
|
|
|
2006-09-07 18:17:04 +08:00
|
|
|
#ifndef arch_mmap_check
|
|
|
|
|
#define arch_mmap_check(addr, len, flags) (0)
|
|
|
|
|
#endif
|
|
|
|
|
|
mm: mmap: add new /proc tunable for mmap_base ASLR
Address Space Layout Randomization (ASLR) provides a barrier to
exploitation of user-space processes in the presence of security
vulnerabilities by making it more difficult to find desired code/data
which could help an attack. This is done by adding a random offset to
the location of regions in the process address space, with a greater
range of potential offset values corresponding to better protection/a
larger search-space for brute force, but also to greater potential for
fragmentation.
The offset added to the mmap_base address, which provides the basis for
the majority of the mappings for a process, is set once on process exec
in arch_pick_mmap_layout() and is done via hard-coded per-arch values,
which reflect, hopefully, the best compromise for all systems. The
trade-off between increased entropy in the offset value generation and
the corresponding increased variability in address space fragmentation
is not absolute, however, and some platforms may tolerate higher amounts
of entropy. This patch introduces both new Kconfig values and a sysctl
interface which may be used to change the amount of entropy used for
offset generation on a system.
The direct motivation for this change was in response to the
libstagefright vulnerabilities that affected Android, specifically to
information provided by Google's project zero at:
http://googleprojectzero.blogspot.com/2015/09/stagefrightened.html
The attack presented therein, by Google's project zero, specifically
targeted the limited randomness used to generate the offset added to the
mmap_base address in order to craft a brute-force-based attack.
Concretely, the attack was against the mediaserver process, which was
limited to respawning every 5 seconds, on an arm device. The hard-coded
8 bits used resulted in an average expected success rate of defeating
the mmap ASLR after just over 10 minutes (128 tries at 5 seconds a
piece). With this patch, and an accompanying increase in the entropy
value to 16 bits, the same attack would take an average expected time of
over 45 hours (32768 tries), which makes it both less feasible and more
likely to be noticed.
The introduced Kconfig and sysctl options are limited by per-arch
minimum and maximum values, the minimum of which was chosen to match the
current hard-coded value and the maximum of which was chosen so as to
give the greatest flexibility without generating an invalid mmap_base
address, generally a 3-4 bits less than the number of bits in the
user-space accessible virtual address space.
When decided whether or not to change the default value, a system
developer should consider that mmap_base address could be placed
anywhere up to 2^(value) bits away from the non-randomized location,
which would introduce variable-sized areas above and below the mmap_base
address such that the maximum vm_area_struct size may be reduced,
preventing very large allocations.
This patch (of 4):
ASLR only uses as few as 8 bits to generate the random offset for the
mmap base address on 32 bit architectures. This value was chosen to
prevent a poorly chosen value from dividing the address space in such a
way as to prevent large allocations. This may not be an issue on all
platforms. Allow the specification of a minimum number of bits so that
platforms desiring greater ASLR protection may determine where to place
the trade-off.
Signed-off-by: Daniel Cashman <dcashman@google.com>
Cc: Russell King <linux@arm.linux.org.uk>
Acked-by: Kees Cook <keescook@chromium.org>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Don Zickus <dzickus@redhat.com>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Cc: Heinrich Schuchardt <xypron.glpk@gmx.de>
Cc: Josh Poimboeuf <jpoimboe@redhat.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Mark Salyzyn <salyzyn@android.com>
Cc: Jeff Vander Stoep <jeffv@google.com>
Cc: Nick Kralevich <nnk@google.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Will Deacon <will.deacon@arm.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Hector Marco-Gisbert <hecmargi@upv.es>
Cc: Borislav Petkov <bp@suse.de>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-15 07:19:53 +08:00
|
|
|
#ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
|
|
|
|
|
const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
|
|
|
|
|
const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
|
|
|
|
|
int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
|
|
|
|
|
#endif
|
|
|
|
|
#ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
|
|
|
|
|
const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
|
|
|
|
|
const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
|
|
|
|
|
int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
|
|
|
|
|
#endif
|
|
|
|
|
|
2016-05-21 07:57:45 +08:00
|
|
|
static bool ignore_rlimit_data;
|
2016-02-03 08:57:43 +08:00
|
|
|
core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
|
mm: mmap: add new /proc tunable for mmap_base ASLR
Address Space Layout Randomization (ASLR) provides a barrier to
exploitation of user-space processes in the presence of security
vulnerabilities by making it more difficult to find desired code/data
which could help an attack. This is done by adding a random offset to
the location of regions in the process address space, with a greater
range of potential offset values corresponding to better protection/a
larger search-space for brute force, but also to greater potential for
fragmentation.
The offset added to the mmap_base address, which provides the basis for
the majority of the mappings for a process, is set once on process exec
in arch_pick_mmap_layout() and is done via hard-coded per-arch values,
which reflect, hopefully, the best compromise for all systems. The
trade-off between increased entropy in the offset value generation and
the corresponding increased variability in address space fragmentation
is not absolute, however, and some platforms may tolerate higher amounts
of entropy. This patch introduces both new Kconfig values and a sysctl
interface which may be used to change the amount of entropy used for
offset generation on a system.
The direct motivation for this change was in response to the
libstagefright vulnerabilities that affected Android, specifically to
information provided by Google's project zero at:
http://googleprojectzero.blogspot.com/2015/09/stagefrightened.html
The attack presented therein, by Google's project zero, specifically
targeted the limited randomness used to generate the offset added to the
mmap_base address in order to craft a brute-force-based attack.
Concretely, the attack was against the mediaserver process, which was
limited to respawning every 5 seconds, on an arm device. The hard-coded
8 bits used resulted in an average expected success rate of defeating
the mmap ASLR after just over 10 minutes (128 tries at 5 seconds a
piece). With this patch, and an accompanying increase in the entropy
value to 16 bits, the same attack would take an average expected time of
over 45 hours (32768 tries), which makes it both less feasible and more
likely to be noticed.
The introduced Kconfig and sysctl options are limited by per-arch
minimum and maximum values, the minimum of which was chosen to match the
current hard-coded value and the maximum of which was chosen so as to
give the greatest flexibility without generating an invalid mmap_base
address, generally a 3-4 bits less than the number of bits in the
user-space accessible virtual address space.
When decided whether or not to change the default value, a system
developer should consider that mmap_base address could be placed
anywhere up to 2^(value) bits away from the non-randomized location,
which would introduce variable-sized areas above and below the mmap_base
address such that the maximum vm_area_struct size may be reduced,
preventing very large allocations.
This patch (of 4):
ASLR only uses as few as 8 bits to generate the random offset for the
mmap base address on 32 bit architectures. This value was chosen to
prevent a poorly chosen value from dividing the address space in such a
way as to prevent large allocations. This may not be an issue on all
platforms. Allow the specification of a minimum number of bits so that
platforms desiring greater ASLR protection may determine where to place
the trade-off.
Signed-off-by: Daniel Cashman <dcashman@google.com>
Cc: Russell King <linux@arm.linux.org.uk>
Acked-by: Kees Cook <keescook@chromium.org>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Don Zickus <dzickus@redhat.com>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Cc: Heinrich Schuchardt <xypron.glpk@gmx.de>
Cc: Josh Poimboeuf <jpoimboe@redhat.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Mark Salyzyn <salyzyn@android.com>
Cc: Jeff Vander Stoep <jeffv@google.com>
Cc: Nick Kralevich <nnk@google.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Will Deacon <will.deacon@arm.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Hector Marco-Gisbert <hecmargi@upv.es>
Cc: Borislav Petkov <bp@suse.de>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-15 07:19:53 +08:00
|
|
|
|
[PATCH] freepgt: free_pgtables use vma list
Recent woes with some arches needing their own pgd_addr_end macro; and 4-level
clear_page_range regression since 2.6.10's clear_page_tables; and its
long-standing well-known inefficiency in searching throughout the higher-level
page tables for those few entries to clear and free: all can be blamed on
ignoring the list of vmas when we free page tables.
Replace exit_mmap's clear_page_range of the total user address space by
free_pgtables operating on the mm's vma list; unmap_region use it in the same
way, giving floor and ceiling beyond which it may not free tables. This
brings lmbench fork/exec/sh numbers back to 2.6.10 (unless preempt is enabled,
in which case latency fixes spoil unmap_vmas throughput).
Beware: the do_mmap_pgoff driver failure case must now use unmap_region
instead of zap_page_range, since a page table might have been allocated, and
can only be freed while it is touched by some vma.
Move free_pgtables from mmap.c to memory.c, where its lower levels are adapted
from the clear_page_range levels. (Most of free_pgtables' old code was
actually for a non-existent case, prev not properly set up, dating from before
hch gave us split_vma.) Pass mmu_gather** in the public interfaces, since we
might want to add latency lockdrops later; but no attempt to do so yet, going
by vma should itself reduce latency.
But what if is_hugepage_only_range? Those ia64 and ppc64 cases need careful
examination: put that off until a later patch of the series.
What of x86_64's 32bit vdso page __map_syscall32 maps outside any vma?
And the range to sparc64's flush_tlb_pgtables? It's less clear to me now that
we need to do more than is done here - every PMD_SIZE ever occupied will be
flushed, do we really have to flush every PGDIR_SIZE ever partially occupied?
A shame to complicate it unnecessarily.
Special thanks to David Miller for time spent repairing my ceilings.
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-04-20 04:29:15 +08:00
|
|
|
static void unmap_region(struct mm_struct *mm,
|
|
|
|
|
struct vm_area_struct *vma, struct vm_area_struct *prev,
|
|
|
|
|
unsigned long start, unsigned long end);
|
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
/* description of effects of mapping type and prot in current implementation.
|
|
|
|
|
* this is due to the limited x86 page protection hardware. The expected
|
|
|
|
|
* behavior is in parens:
|
|
|
|
|
*
|
|
|
|
|
* map_type prot
|
|
|
|
|
* PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
|
|
|
|
|
* MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
|
|
|
|
|
* w: (no) no w: (no) no w: (yes) yes w: (no) no
|
|
|
|
|
* x: (no) no x: (no) yes x: (no) yes x: (yes) yes
|
2014-10-10 06:26:29 +08:00
|
|
|
*
|
2005-04-17 06:20:36 +08:00
|
|
|
* MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
|
|
|
|
|
* w: (no) no w: (no) no w: (copy) copy w: (no) no
|
|
|
|
|
* x: (no) no x: (no) yes x: (no) yes x: (yes) yes
|
2021-03-13 01:38:10 +08:00
|
|
|
*
|
|
|
|
|
* On arm64, PROT_EXEC has the following behaviour for both MAP_SHARED and
|
|
|
|
|
* MAP_PRIVATE (with Enhanced PAN supported):
|
|
|
|
|
* r: (no) no
|
|
|
|
|
* w: (no) no
|
|
|
|
|
* x: (yes) yes
|
2005-04-17 06:20:36 +08:00
|
|
|
*/
|
2017-07-07 06:36:47 +08:00
|
|
|
pgprot_t protection_map[16] __ro_after_init = {
|
2005-04-17 06:20:36 +08:00
|
|
|
__P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
|
|
|
|
|
__S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
|
|
|
|
|
};
|
|
|
|
|
|
2018-04-21 06:20:28 +08:00
|
|
|
#ifndef CONFIG_ARCH_HAS_FILTER_PGPROT
|
|
|
|
|
static inline pgprot_t arch_filter_pgprot(pgprot_t prot)
|
|
|
|
|
{
|
|
|
|
|
return prot;
|
|
|
|
|
}
|
|
|
|
|
#endif
|
|
|
|
|
|
2006-07-27 04:39:49 +08:00
|
|
|
pgprot_t vm_get_page_prot(unsigned long vm_flags)
|
|
|
|
|
{
|
2018-04-21 06:20:28 +08:00
|
|
|
pgprot_t ret = __pgprot(pgprot_val(protection_map[vm_flags &
|
2008-07-07 22:28:51 +08:00
|
|
|
(VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
|
|
|
|
|
pgprot_val(arch_vm_get_page_prot(vm_flags)));
|
2018-04-21 06:20:28 +08:00
|
|
|
|
|
|
|
|
return arch_filter_pgprot(ret);
|
2006-07-27 04:39:49 +08:00
|
|
|
}
|
|
|
|
|
EXPORT_SYMBOL(vm_get_page_prot);
|
|
|
|
|
|
mm: softdirty: enable write notifications on VMAs after VM_SOFTDIRTY cleared
For VMAs that don't want write notifications, PTEs created for read faults
have their write bit set. If the read fault happens after VM_SOFTDIRTY is
cleared, then the PTE's softdirty bit will remain clear after subsequent
writes.
Here's a simple code snippet to demonstrate the bug:
char* m = mmap(NULL, getpagesize(), PROT_READ | PROT_WRITE,
MAP_ANONYMOUS | MAP_SHARED, -1, 0);
system("echo 4 > /proc/$PPID/clear_refs"); /* clear VM_SOFTDIRTY */
assert(*m == '\0'); /* new PTE allows write access */
assert(!soft_dirty(x));
*m = 'x'; /* should dirty the page */
assert(soft_dirty(x)); /* fails */
With this patch, write notifications are enabled when VM_SOFTDIRTY is
cleared. Furthermore, to avoid unnecessary faults, write notifications
are disabled when VM_SOFTDIRTY is set.
As a side effect of enabling and disabling write notifications with
care, this patch fixes a bug in mprotect where vm_page_prot bits set by
drivers were zapped on mprotect. An analogous bug was fixed in mmap by
commit c9d0bf241451 ("mm: uncached vma support with writenotify").
Signed-off-by: Peter Feiner <pfeiner@google.com>
Reported-by: Peter Feiner <pfeiner@google.com>
Suggested-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Jamie Liu <jamieliu@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Bjorn Helgaas <bhelgaas@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-10-14 06:55:46 +08:00
|
|
|
static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
|
|
|
|
|
{
|
|
|
|
|
return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Update vma->vm_page_prot to reflect vma->vm_flags. */
|
|
|
|
|
void vma_set_page_prot(struct vm_area_struct *vma)
|
|
|
|
|
{
|
|
|
|
|
unsigned long vm_flags = vma->vm_flags;
|
2016-10-08 08:01:22 +08:00
|
|
|
pgprot_t vm_page_prot;
|
mm: softdirty: enable write notifications on VMAs after VM_SOFTDIRTY cleared
For VMAs that don't want write notifications, PTEs created for read faults
have their write bit set. If the read fault happens after VM_SOFTDIRTY is
cleared, then the PTE's softdirty bit will remain clear after subsequent
writes.
Here's a simple code snippet to demonstrate the bug:
char* m = mmap(NULL, getpagesize(), PROT_READ | PROT_WRITE,
MAP_ANONYMOUS | MAP_SHARED, -1, 0);
system("echo 4 > /proc/$PPID/clear_refs"); /* clear VM_SOFTDIRTY */
assert(*m == '\0'); /* new PTE allows write access */
assert(!soft_dirty(x));
*m = 'x'; /* should dirty the page */
assert(soft_dirty(x)); /* fails */
With this patch, write notifications are enabled when VM_SOFTDIRTY is
cleared. Furthermore, to avoid unnecessary faults, write notifications
are disabled when VM_SOFTDIRTY is set.
As a side effect of enabling and disabling write notifications with
care, this patch fixes a bug in mprotect where vm_page_prot bits set by
drivers were zapped on mprotect. An analogous bug was fixed in mmap by
commit c9d0bf241451 ("mm: uncached vma support with writenotify").
Signed-off-by: Peter Feiner <pfeiner@google.com>
Reported-by: Peter Feiner <pfeiner@google.com>
Suggested-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Jamie Liu <jamieliu@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Bjorn Helgaas <bhelgaas@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-10-14 06:55:46 +08:00
|
|
|
|
2016-10-08 08:01:22 +08:00
|
|
|
vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
|
|
|
|
|
if (vma_wants_writenotify(vma, vm_page_prot)) {
|
mm: softdirty: enable write notifications on VMAs after VM_SOFTDIRTY cleared
For VMAs that don't want write notifications, PTEs created for read faults
have their write bit set. If the read fault happens after VM_SOFTDIRTY is
cleared, then the PTE's softdirty bit will remain clear after subsequent
writes.
Here's a simple code snippet to demonstrate the bug:
char* m = mmap(NULL, getpagesize(), PROT_READ | PROT_WRITE,
MAP_ANONYMOUS | MAP_SHARED, -1, 0);
system("echo 4 > /proc/$PPID/clear_refs"); /* clear VM_SOFTDIRTY */
assert(*m == '\0'); /* new PTE allows write access */
assert(!soft_dirty(x));
*m = 'x'; /* should dirty the page */
assert(soft_dirty(x)); /* fails */
With this patch, write notifications are enabled when VM_SOFTDIRTY is
cleared. Furthermore, to avoid unnecessary faults, write notifications
are disabled when VM_SOFTDIRTY is set.
As a side effect of enabling and disabling write notifications with
care, this patch fixes a bug in mprotect where vm_page_prot bits set by
drivers were zapped on mprotect. An analogous bug was fixed in mmap by
commit c9d0bf241451 ("mm: uncached vma support with writenotify").
Signed-off-by: Peter Feiner <pfeiner@google.com>
Reported-by: Peter Feiner <pfeiner@google.com>
Suggested-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Jamie Liu <jamieliu@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Bjorn Helgaas <bhelgaas@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-10-14 06:55:46 +08:00
|
|
|
vm_flags &= ~VM_SHARED;
|
2016-10-08 08:01:22 +08:00
|
|
|
vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
|
mm: softdirty: enable write notifications on VMAs after VM_SOFTDIRTY cleared
For VMAs that don't want write notifications, PTEs created for read faults
have their write bit set. If the read fault happens after VM_SOFTDIRTY is
cleared, then the PTE's softdirty bit will remain clear after subsequent
writes.
Here's a simple code snippet to demonstrate the bug:
char* m = mmap(NULL, getpagesize(), PROT_READ | PROT_WRITE,
MAP_ANONYMOUS | MAP_SHARED, -1, 0);
system("echo 4 > /proc/$PPID/clear_refs"); /* clear VM_SOFTDIRTY */
assert(*m == '\0'); /* new PTE allows write access */
assert(!soft_dirty(x));
*m = 'x'; /* should dirty the page */
assert(soft_dirty(x)); /* fails */
With this patch, write notifications are enabled when VM_SOFTDIRTY is
cleared. Furthermore, to avoid unnecessary faults, write notifications
are disabled when VM_SOFTDIRTY is set.
As a side effect of enabling and disabling write notifications with
care, this patch fixes a bug in mprotect where vm_page_prot bits set by
drivers were zapped on mprotect. An analogous bug was fixed in mmap by
commit c9d0bf241451 ("mm: uncached vma support with writenotify").
Signed-off-by: Peter Feiner <pfeiner@google.com>
Reported-by: Peter Feiner <pfeiner@google.com>
Suggested-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Jamie Liu <jamieliu@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Bjorn Helgaas <bhelgaas@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-10-14 06:55:46 +08:00
|
|
|
}
|
2020-06-09 12:33:54 +08:00
|
|
|
/* remove_protection_ptes reads vma->vm_page_prot without mmap_lock */
|
2016-10-08 08:01:22 +08:00
|
|
|
WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
|
mm: softdirty: enable write notifications on VMAs after VM_SOFTDIRTY cleared
For VMAs that don't want write notifications, PTEs created for read faults
have their write bit set. If the read fault happens after VM_SOFTDIRTY is
cleared, then the PTE's softdirty bit will remain clear after subsequent
writes.
Here's a simple code snippet to demonstrate the bug:
char* m = mmap(NULL, getpagesize(), PROT_READ | PROT_WRITE,
MAP_ANONYMOUS | MAP_SHARED, -1, 0);
system("echo 4 > /proc/$PPID/clear_refs"); /* clear VM_SOFTDIRTY */
assert(*m == '\0'); /* new PTE allows write access */
assert(!soft_dirty(x));
*m = 'x'; /* should dirty the page */
assert(soft_dirty(x)); /* fails */
With this patch, write notifications are enabled when VM_SOFTDIRTY is
cleared. Furthermore, to avoid unnecessary faults, write notifications
are disabled when VM_SOFTDIRTY is set.
As a side effect of enabling and disabling write notifications with
care, this patch fixes a bug in mprotect where vm_page_prot bits set by
drivers were zapped on mprotect. An analogous bug was fixed in mmap by
commit c9d0bf241451 ("mm: uncached vma support with writenotify").
Signed-off-by: Peter Feiner <pfeiner@google.com>
Reported-by: Peter Feiner <pfeiner@google.com>
Suggested-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Jamie Liu <jamieliu@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Bjorn Helgaas <bhelgaas@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-10-14 06:55:46 +08:00
|
|
|
}
|
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
/*
|
2014-12-13 08:54:24 +08:00
|
|
|
* Requires inode->i_mapping->i_mmap_rwsem
|
2005-04-17 06:20:36 +08:00
|
|
|
*/
|
|
|
|
|
static void __remove_shared_vm_struct(struct vm_area_struct *vma,
|
|
|
|
|
struct file *file, struct address_space *mapping)
|
|
|
|
|
{
|
|
|
|
|
if (vma->vm_flags & VM_SHARED)
|
2014-08-09 05:25:25 +08:00
|
|
|
mapping_unmap_writable(mapping);
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
|
flush_dcache_mmap_lock(mapping);
|
2015-02-11 06:09:59 +08:00
|
|
|
vma_interval_tree_remove(vma, &mapping->i_mmap);
|
2005-04-17 06:20:36 +08:00
|
|
|
flush_dcache_mmap_unlock(mapping);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
2012-10-09 07:31:25 +08:00
|
|
|
* Unlink a file-based vm structure from its interval tree, to hide
|
2005-10-30 09:15:57 +08:00
|
|
|
* vma from rmap and vmtruncate before freeing its page tables.
|
2005-04-17 06:20:36 +08:00
|
|
|
*/
|
2005-10-30 09:15:57 +08:00
|
|
|
void unlink_file_vma(struct vm_area_struct *vma)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
|
|
|
|
struct file *file = vma->vm_file;
|
|
|
|
|
|
|
|
|
|
if (file) {
|
|
|
|
|
struct address_space *mapping = file->f_mapping;
|
2014-12-13 08:54:21 +08:00
|
|
|
i_mmap_lock_write(mapping);
|
2005-04-17 06:20:36 +08:00
|
|
|
__remove_shared_vm_struct(vma, file, mapping);
|
2014-12-13 08:54:21 +08:00
|
|
|
i_mmap_unlock_write(mapping);
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
2005-10-30 09:15:57 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Close a vm structure and free it, returning the next.
|
|
|
|
|
*/
|
|
|
|
|
static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
|
|
|
|
|
{
|
|
|
|
|
struct vm_area_struct *next = vma->vm_next;
|
|
|
|
|
|
|
|
|
|
might_sleep();
|
2005-04-17 06:20:36 +08:00
|
|
|
if (vma->vm_ops && vma->vm_ops->close)
|
|
|
|
|
vma->vm_ops->close(vma);
|
2012-10-09 07:28:54 +08:00
|
|
|
if (vma->vm_file)
|
2005-10-30 09:15:57 +08:00
|
|
|
fput(vma->vm_file);
|
2008-04-28 17:13:08 +08:00
|
|
|
mpol_put(vma_policy(vma));
|
2018-07-22 04:48:51 +08:00
|
|
|
vm_area_free(vma);
|
2005-10-30 09:15:57 +08:00
|
|
|
return next;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
|
2018-07-14 07:59:20 +08:00
|
|
|
static int do_brk_flags(unsigned long addr, unsigned long request, unsigned long flags,
|
|
|
|
|
struct list_head *uf);
|
2009-01-14 21:14:15 +08:00
|
|
|
SYSCALL_DEFINE1(brk, unsigned long, brk)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
2018-10-27 06:08:54 +08:00
|
|
|
unsigned long newbrk, oldbrk, origbrk;
|
2005-04-17 06:20:36 +08:00
|
|
|
struct mm_struct *mm = current->mm;
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
struct vm_area_struct *next;
|
2008-06-06 13:46:05 +08:00
|
|
|
unsigned long min_brk;
|
2013-02-23 08:32:40 +08:00
|
|
|
bool populate;
|
2018-10-27 06:08:54 +08:00
|
|
|
bool downgraded = false;
|
2017-02-25 06:58:22 +08:00
|
|
|
LIST_HEAD(uf);
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2020-06-09 12:33:25 +08:00
|
|
|
if (mmap_write_lock_killable(mm))
|
2016-05-24 07:25:27 +08:00
|
|
|
return -EINTR;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2018-10-27 06:08:54 +08:00
|
|
|
origbrk = mm->brk;
|
|
|
|
|
|
2008-06-06 13:46:05 +08:00
|
|
|
#ifdef CONFIG_COMPAT_BRK
|
2011-01-14 07:47:23 +08:00
|
|
|
/*
|
|
|
|
|
* CONFIG_COMPAT_BRK can still be overridden by setting
|
|
|
|
|
* randomize_va_space to 2, which will still cause mm->start_brk
|
|
|
|
|
* to be arbitrarily shifted
|
|
|
|
|
*/
|
2011-04-15 06:22:09 +08:00
|
|
|
if (current->brk_randomized)
|
2011-01-14 07:47:23 +08:00
|
|
|
min_brk = mm->start_brk;
|
|
|
|
|
else
|
|
|
|
|
min_brk = mm->end_data;
|
2008-06-06 13:46:05 +08:00
|
|
|
#else
|
|
|
|
|
min_brk = mm->start_brk;
|
|
|
|
|
#endif
|
|
|
|
|
if (brk < min_brk)
|
2005-04-17 06:20:36 +08:00
|
|
|
goto out;
|
2006-04-11 13:52:57 +08:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Check against rlimit here. If this check is done later after the test
|
|
|
|
|
* of oldbrk with newbrk then it can escape the test and let the data
|
|
|
|
|
* segment grow beyond its set limit the in case where the limit is
|
|
|
|
|
* not page aligned -Ram Gupta
|
|
|
|
|
*/
|
2014-10-10 06:27:32 +08:00
|
|
|
if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
|
|
|
|
|
mm->end_data, mm->start_data))
|
2006-04-11 13:52:57 +08:00
|
|
|
goto out;
|
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
newbrk = PAGE_ALIGN(brk);
|
|
|
|
|
oldbrk = PAGE_ALIGN(mm->brk);
|
2018-10-27 06:08:54 +08:00
|
|
|
if (oldbrk == newbrk) {
|
|
|
|
|
mm->brk = brk;
|
|
|
|
|
goto success;
|
|
|
|
|
}
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2018-10-27 06:08:54 +08:00
|
|
|
/*
|
|
|
|
|
* Always allow shrinking brk.
|
2020-06-09 12:33:54 +08:00
|
|
|
* __do_munmap() may downgrade mmap_lock to read.
|
2018-10-27 06:08:54 +08:00
|
|
|
*/
|
2005-04-17 06:20:36 +08:00
|
|
|
if (brk <= mm->brk) {
|
2018-10-27 06:08:54 +08:00
|
|
|
int ret;
|
|
|
|
|
|
|
|
|
|
/*
|
2020-06-09 12:33:54 +08:00
|
|
|
* mm->brk must to be protected by write mmap_lock so update it
|
|
|
|
|
* before downgrading mmap_lock. When __do_munmap() fails,
|
2018-10-27 06:08:54 +08:00
|
|
|
* mm->brk will be restored from origbrk.
|
|
|
|
|
*/
|
|
|
|
|
mm->brk = brk;
|
|
|
|
|
ret = __do_munmap(mm, newbrk, oldbrk-newbrk, &uf, true);
|
|
|
|
|
if (ret < 0) {
|
|
|
|
|
mm->brk = origbrk;
|
|
|
|
|
goto out;
|
|
|
|
|
} else if (ret == 1) {
|
|
|
|
|
downgraded = true;
|
|
|
|
|
}
|
|
|
|
|
goto success;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Check against existing mmap mappings. */
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
next = find_vma(mm, oldbrk);
|
|
|
|
|
if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
|
2005-04-17 06:20:36 +08:00
|
|
|
goto out;
|
|
|
|
|
|
|
|
|
|
/* Ok, looks good - let it rip. */
|
2018-07-14 07:59:20 +08:00
|
|
|
if (do_brk_flags(oldbrk, newbrk-oldbrk, 0, &uf) < 0)
|
2005-04-17 06:20:36 +08:00
|
|
|
goto out;
|
|
|
|
|
mm->brk = brk;
|
2018-10-27 06:08:54 +08:00
|
|
|
|
|
|
|
|
success:
|
2013-02-23 08:32:40 +08:00
|
|
|
populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
|
2018-10-27 06:08:54 +08:00
|
|
|
if (downgraded)
|
2020-06-09 12:33:25 +08:00
|
|
|
mmap_read_unlock(mm);
|
2018-10-27 06:08:54 +08:00
|
|
|
else
|
2020-06-09 12:33:25 +08:00
|
|
|
mmap_write_unlock(mm);
|
2017-02-25 06:58:22 +08:00
|
|
|
userfaultfd_unmap_complete(mm, &uf);
|
2013-02-23 08:32:40 +08:00
|
|
|
if (populate)
|
|
|
|
|
mm_populate(oldbrk, newbrk - oldbrk);
|
|
|
|
|
return brk;
|
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
out:
|
2020-06-09 12:33:25 +08:00
|
|
|
mmap_write_unlock(mm);
|
2021-02-25 04:04:29 +08:00
|
|
|
return origbrk;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
|
2019-09-26 07:46:07 +08:00
|
|
|
static inline unsigned long vma_compute_gap(struct vm_area_struct *vma)
|
2012-12-12 08:01:38 +08:00
|
|
|
{
|
2019-09-26 07:46:07 +08:00
|
|
|
unsigned long gap, prev_end;
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
|
|
|
|
|
* allow two stack_guard_gaps between them here, and when choosing
|
|
|
|
|
* an unmapped area; whereas when expanding we only require one.
|
|
|
|
|
* That's a little inconsistent, but keeps the code here simpler.
|
|
|
|
|
*/
|
2019-09-26 07:46:07 +08:00
|
|
|
gap = vm_start_gap(vma);
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
if (vma->vm_prev) {
|
|
|
|
|
prev_end = vm_end_gap(vma->vm_prev);
|
2019-09-26 07:46:07 +08:00
|
|
|
if (gap > prev_end)
|
|
|
|
|
gap -= prev_end;
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
else
|
2019-09-26 07:46:07 +08:00
|
|
|
gap = 0;
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
}
|
2019-09-26 07:46:07 +08:00
|
|
|
return gap;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
#ifdef CONFIG_DEBUG_VM_RB
|
|
|
|
|
static unsigned long vma_compute_subtree_gap(struct vm_area_struct *vma)
|
|
|
|
|
{
|
|
|
|
|
unsigned long max = vma_compute_gap(vma), subtree_gap;
|
2012-12-12 08:01:38 +08:00
|
|
|
if (vma->vm_rb.rb_left) {
|
|
|
|
|
subtree_gap = rb_entry(vma->vm_rb.rb_left,
|
|
|
|
|
struct vm_area_struct, vm_rb)->rb_subtree_gap;
|
|
|
|
|
if (subtree_gap > max)
|
|
|
|
|
max = subtree_gap;
|
|
|
|
|
}
|
|
|
|
|
if (vma->vm_rb.rb_right) {
|
|
|
|
|
subtree_gap = rb_entry(vma->vm_rb.rb_right,
|
|
|
|
|
struct vm_area_struct, vm_rb)->rb_subtree_gap;
|
|
|
|
|
if (subtree_gap > max)
|
|
|
|
|
max = subtree_gap;
|
|
|
|
|
}
|
|
|
|
|
return max;
|
|
|
|
|
}
|
|
|
|
|
|
2016-02-06 07:36:13 +08:00
|
|
|
static int browse_rb(struct mm_struct *mm)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
2016-02-06 07:36:13 +08:00
|
|
|
struct rb_root *root = &mm->mm_rb;
|
2012-12-12 08:01:42 +08:00
|
|
|
int i = 0, j, bug = 0;
|
2005-04-17 06:20:36 +08:00
|
|
|
struct rb_node *nd, *pn = NULL;
|
|
|
|
|
unsigned long prev = 0, pend = 0;
|
|
|
|
|
|
|
|
|
|
for (nd = rb_first(root); nd; nd = rb_next(nd)) {
|
|
|
|
|
struct vm_area_struct *vma;
|
|
|
|
|
vma = rb_entry(nd, struct vm_area_struct, vm_rb);
|
2012-12-12 08:01:42 +08:00
|
|
|
if (vma->vm_start < prev) {
|
2014-10-10 06:28:19 +08:00
|
|
|
pr_emerg("vm_start %lx < prev %lx\n",
|
|
|
|
|
vma->vm_start, prev);
|
2012-12-12 08:01:42 +08:00
|
|
|
bug = 1;
|
|
|
|
|
}
|
|
|
|
|
if (vma->vm_start < pend) {
|
2014-10-10 06:28:19 +08:00
|
|
|
pr_emerg("vm_start %lx < pend %lx\n",
|
|
|
|
|
vma->vm_start, pend);
|
2012-12-12 08:01:42 +08:00
|
|
|
bug = 1;
|
|
|
|
|
}
|
|
|
|
|
if (vma->vm_start > vma->vm_end) {
|
2014-10-10 06:28:19 +08:00
|
|
|
pr_emerg("vm_start %lx > vm_end %lx\n",
|
|
|
|
|
vma->vm_start, vma->vm_end);
|
2012-12-12 08:01:42 +08:00
|
|
|
bug = 1;
|
|
|
|
|
}
|
2016-02-06 07:36:13 +08:00
|
|
|
spin_lock(&mm->page_table_lock);
|
2012-12-12 08:01:42 +08:00
|
|
|
if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
|
2014-09-10 05:50:59 +08:00
|
|
|
pr_emerg("free gap %lx, correct %lx\n",
|
2012-12-12 08:01:42 +08:00
|
|
|
vma->rb_subtree_gap,
|
|
|
|
|
vma_compute_subtree_gap(vma));
|
|
|
|
|
bug = 1;
|
|
|
|
|
}
|
2016-02-06 07:36:13 +08:00
|
|
|
spin_unlock(&mm->page_table_lock);
|
2005-04-17 06:20:36 +08:00
|
|
|
i++;
|
|
|
|
|
pn = nd;
|
2007-03-01 12:13:13 +08:00
|
|
|
prev = vma->vm_start;
|
|
|
|
|
pend = vma->vm_end;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
j = 0;
|
2012-12-12 08:01:42 +08:00
|
|
|
for (nd = pn; nd; nd = rb_prev(nd))
|
2005-04-17 06:20:36 +08:00
|
|
|
j++;
|
2012-12-12 08:01:42 +08:00
|
|
|
if (i != j) {
|
2014-09-10 05:50:59 +08:00
|
|
|
pr_emerg("backwards %d, forwards %d\n", j, i);
|
2012-12-12 08:01:42 +08:00
|
|
|
bug = 1;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
2012-12-12 08:01:42 +08:00
|
|
|
return bug ? -1 : i;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
|
2012-12-12 08:01:38 +08:00
|
|
|
static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
|
|
|
|
|
{
|
|
|
|
|
struct rb_node *nd;
|
|
|
|
|
|
|
|
|
|
for (nd = rb_first(root); nd; nd = rb_next(nd)) {
|
|
|
|
|
struct vm_area_struct *vma;
|
|
|
|
|
vma = rb_entry(nd, struct vm_area_struct, vm_rb);
|
2014-10-10 06:28:39 +08:00
|
|
|
VM_BUG_ON_VMA(vma != ignore &&
|
|
|
|
|
vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
|
|
|
|
|
vma);
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2014-04-04 05:48:03 +08:00
|
|
|
static void validate_mm(struct mm_struct *mm)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
|
|
|
|
int bug = 0;
|
|
|
|
|
int i = 0;
|
2012-12-12 08:01:42 +08:00
|
|
|
unsigned long highest_address = 0;
|
2012-10-09 07:31:45 +08:00
|
|
|
struct vm_area_struct *vma = mm->mmap;
|
2014-10-10 06:28:19 +08:00
|
|
|
|
2012-10-09 07:31:45 +08:00
|
|
|
while (vma) {
|
2016-02-06 07:36:50 +08:00
|
|
|
struct anon_vma *anon_vma = vma->anon_vma;
|
2012-10-09 07:31:45 +08:00
|
|
|
struct anon_vma_chain *avc;
|
2014-10-10 06:28:19 +08:00
|
|
|
|
2016-02-06 07:36:50 +08:00
|
|
|
if (anon_vma) {
|
|
|
|
|
anon_vma_lock_read(anon_vma);
|
|
|
|
|
list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
|
|
|
|
|
anon_vma_interval_tree_verify(avc);
|
|
|
|
|
anon_vma_unlock_read(anon_vma);
|
|
|
|
|
}
|
|
|
|
|
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
highest_address = vm_end_gap(vma);
|
2012-10-09 07:31:45 +08:00
|
|
|
vma = vma->vm_next;
|
2005-04-17 06:20:36 +08:00
|
|
|
i++;
|
|
|
|
|
}
|
2012-12-12 08:01:42 +08:00
|
|
|
if (i != mm->map_count) {
|
2014-09-10 05:50:59 +08:00
|
|
|
pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
|
2012-12-12 08:01:42 +08:00
|
|
|
bug = 1;
|
|
|
|
|
}
|
|
|
|
|
if (highest_address != mm->highest_vm_end) {
|
2014-09-10 05:50:59 +08:00
|
|
|
pr_emerg("mm->highest_vm_end %lx, found %lx\n",
|
2014-10-10 06:28:19 +08:00
|
|
|
mm->highest_vm_end, highest_address);
|
2012-12-12 08:01:42 +08:00
|
|
|
bug = 1;
|
|
|
|
|
}
|
2016-02-06 07:36:13 +08:00
|
|
|
i = browse_rb(mm);
|
2012-12-12 08:01:42 +08:00
|
|
|
if (i != mm->map_count) {
|
2014-10-10 06:28:19 +08:00
|
|
|
if (i != -1)
|
|
|
|
|
pr_emerg("map_count %d rb %d\n", mm->map_count, i);
|
2012-12-12 08:01:42 +08:00
|
|
|
bug = 1;
|
|
|
|
|
}
|
2014-10-10 06:28:39 +08:00
|
|
|
VM_BUG_ON_MM(bug, mm);
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
#else
|
2012-12-12 08:01:38 +08:00
|
|
|
#define validate_mm_rb(root, ignore) do { } while (0)
|
2005-04-17 06:20:36 +08:00
|
|
|
#define validate_mm(mm) do { } while (0)
|
|
|
|
|
#endif
|
|
|
|
|
|
2019-09-26 07:46:07 +08:00
|
|
|
RB_DECLARE_CALLBACKS_MAX(static, vma_gap_callbacks,
|
|
|
|
|
struct vm_area_struct, vm_rb,
|
|
|
|
|
unsigned long, rb_subtree_gap, vma_compute_gap)
|
2012-12-12 08:01:38 +08:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Update augmented rbtree rb_subtree_gap values after vma->vm_start or
|
|
|
|
|
* vma->vm_prev->vm_end values changed, without modifying the vma's position
|
|
|
|
|
* in the rbtree.
|
|
|
|
|
*/
|
|
|
|
|
static void vma_gap_update(struct vm_area_struct *vma)
|
|
|
|
|
{
|
|
|
|
|
/*
|
2019-09-26 07:46:07 +08:00
|
|
|
* As it turns out, RB_DECLARE_CALLBACKS_MAX() already created
|
|
|
|
|
* a callback function that does exactly what we want.
|
2012-12-12 08:01:38 +08:00
|
|
|
*/
|
|
|
|
|
vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static inline void vma_rb_insert(struct vm_area_struct *vma,
|
|
|
|
|
struct rb_root *root)
|
|
|
|
|
{
|
|
|
|
|
/* All rb_subtree_gap values must be consistent prior to insertion */
|
|
|
|
|
validate_mm_rb(root, NULL);
|
|
|
|
|
|
|
|
|
|
rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
|
|
|
|
|
}
|
|
|
|
|
|
2016-10-08 08:01:37 +08:00
|
|
|
static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
|
2012-12-12 08:01:38 +08:00
|
|
|
{
|
|
|
|
|
/*
|
|
|
|
|
* Note rb_erase_augmented is a fairly large inline function,
|
|
|
|
|
* so make sure we instantiate it only once with our desired
|
|
|
|
|
* augmented rbtree callbacks.
|
|
|
|
|
*/
|
|
|
|
|
rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
|
|
|
|
|
}
|
|
|
|
|
|
2016-10-08 08:01:37 +08:00
|
|
|
static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
|
|
|
|
|
struct rb_root *root,
|
|
|
|
|
struct vm_area_struct *ignore)
|
|
|
|
|
{
|
|
|
|
|
/*
|
|
|
|
|
* All rb_subtree_gap values must be consistent prior to erase,
|
2020-10-14 07:53:35 +08:00
|
|
|
* with the possible exception of
|
|
|
|
|
*
|
|
|
|
|
* a. the "next" vma being erased if next->vm_start was reduced in
|
|
|
|
|
* __vma_adjust() -> __vma_unlink()
|
|
|
|
|
* b. the vma being erased in detach_vmas_to_be_unmapped() ->
|
|
|
|
|
* vma_rb_erase()
|
2016-10-08 08:01:37 +08:00
|
|
|
*/
|
|
|
|
|
validate_mm_rb(root, ignore);
|
|
|
|
|
|
|
|
|
|
__vma_rb_erase(vma, root);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
|
|
|
|
|
struct rb_root *root)
|
|
|
|
|
{
|
2020-10-14 07:53:35 +08:00
|
|
|
vma_rb_erase_ignore(vma, root, vma);
|
2016-10-08 08:01:37 +08:00
|
|
|
}
|
|
|
|
|
|
mm anon rmap: replace same_anon_vma linked list with an interval tree.
When a large VMA (anon or private file mapping) is first touched, which
will populate its anon_vma field, and then split into many regions through
the use of mprotect(), the original anon_vma ends up linking all of the
vmas on a linked list. This can cause rmap to become inefficient, as we
have to walk potentially thousands of irrelevent vmas before finding the
one a given anon page might fall into.
By replacing the same_anon_vma linked list with an interval tree (where
each avc's interval is determined by its vma's start and last pgoffs), we
can make rmap efficient for this use case again.
While the change is large, all of its pieces are fairly simple.
Most places that were walking the same_anon_vma list were looking for a
known pgoff, so they can just use the anon_vma_interval_tree_foreach()
interval tree iterator instead. The exception here is ksm, where the
page's index is not known. It would probably be possible to rework ksm so
that the index would be known, but for now I have decided to keep things
simple and just walk the entirety of the interval tree there.
When updating vma's that already have an anon_vma assigned, we must take
care to re-index the corresponding avc's on their interval tree. This is
done through the use of anon_vma_interval_tree_pre_update_vma() and
anon_vma_interval_tree_post_update_vma(), which remove the avc's from
their interval tree before the update and re-insert them after the update.
The anon_vma stays locked during the update, so there is no chance that
rmap would miss the vmas that are being updated.
Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-10-09 07:31:39 +08:00
|
|
|
/*
|
|
|
|
|
* vma has some anon_vma assigned, and is already inserted on that
|
|
|
|
|
* anon_vma's interval trees.
|
|
|
|
|
*
|
|
|
|
|
* Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
|
|
|
|
|
* vma must be removed from the anon_vma's interval trees using
|
|
|
|
|
* anon_vma_interval_tree_pre_update_vma().
|
|
|
|
|
*
|
|
|
|
|
* After the update, the vma will be reinserted using
|
|
|
|
|
* anon_vma_interval_tree_post_update_vma().
|
|
|
|
|
*
|
2020-06-09 12:33:54 +08:00
|
|
|
* The entire update must be protected by exclusive mmap_lock and by
|
mm anon rmap: replace same_anon_vma linked list with an interval tree.
When a large VMA (anon or private file mapping) is first touched, which
will populate its anon_vma field, and then split into many regions through
the use of mprotect(), the original anon_vma ends up linking all of the
vmas on a linked list. This can cause rmap to become inefficient, as we
have to walk potentially thousands of irrelevent vmas before finding the
one a given anon page might fall into.
By replacing the same_anon_vma linked list with an interval tree (where
each avc's interval is determined by its vma's start and last pgoffs), we
can make rmap efficient for this use case again.
While the change is large, all of its pieces are fairly simple.
Most places that were walking the same_anon_vma list were looking for a
known pgoff, so they can just use the anon_vma_interval_tree_foreach()
interval tree iterator instead. The exception here is ksm, where the
page's index is not known. It would probably be possible to rework ksm so
that the index would be known, but for now I have decided to keep things
simple and just walk the entirety of the interval tree there.
When updating vma's that already have an anon_vma assigned, we must take
care to re-index the corresponding avc's on their interval tree. This is
done through the use of anon_vma_interval_tree_pre_update_vma() and
anon_vma_interval_tree_post_update_vma(), which remove the avc's from
their interval tree before the update and re-insert them after the update.
The anon_vma stays locked during the update, so there is no chance that
rmap would miss the vmas that are being updated.
Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-10-09 07:31:39 +08:00
|
|
|
* the root anon_vma's mutex.
|
|
|
|
|
*/
|
|
|
|
|
static inline void
|
|
|
|
|
anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
|
|
|
|
|
{
|
|
|
|
|
struct anon_vma_chain *avc;
|
|
|
|
|
|
|
|
|
|
list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
|
|
|
|
|
anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static inline void
|
|
|
|
|
anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
|
|
|
|
|
{
|
|
|
|
|
struct anon_vma_chain *avc;
|
|
|
|
|
|
|
|
|
|
list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
|
|
|
|
|
anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
|
|
|
|
|
}
|
|
|
|
|
|
2012-10-09 07:29:07 +08:00
|
|
|
static int find_vma_links(struct mm_struct *mm, unsigned long addr,
|
|
|
|
|
unsigned long end, struct vm_area_struct **pprev,
|
|
|
|
|
struct rb_node ***rb_link, struct rb_node **rb_parent)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
2012-10-09 07:29:07 +08:00
|
|
|
struct rb_node **__rb_link, *__rb_parent, *rb_prev;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2021-09-03 05:56:46 +08:00
|
|
|
mmap_assert_locked(mm);
|
2005-04-17 06:20:36 +08:00
|
|
|
__rb_link = &mm->mm_rb.rb_node;
|
|
|
|
|
rb_prev = __rb_parent = NULL;
|
|
|
|
|
|
|
|
|
|
while (*__rb_link) {
|
|
|
|
|
struct vm_area_struct *vma_tmp;
|
|
|
|
|
|
|
|
|
|
__rb_parent = *__rb_link;
|
|
|
|
|
vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
|
|
|
|
|
|
|
|
|
|
if (vma_tmp->vm_end > addr) {
|
2012-10-09 07:29:07 +08:00
|
|
|
/* Fail if an existing vma overlaps the area */
|
|
|
|
|
if (vma_tmp->vm_start < end)
|
|
|
|
|
return -ENOMEM;
|
2005-04-17 06:20:36 +08:00
|
|
|
__rb_link = &__rb_parent->rb_left;
|
|
|
|
|
} else {
|
|
|
|
|
rb_prev = __rb_parent;
|
|
|
|
|
__rb_link = &__rb_parent->rb_right;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
*pprev = NULL;
|
|
|
|
|
if (rb_prev)
|
|
|
|
|
*pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
|
|
|
|
|
*rb_link = __rb_link;
|
|
|
|
|
*rb_parent = __rb_parent;
|
2012-10-09 07:29:07 +08:00
|
|
|
return 0;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
|
2020-10-18 07:14:06 +08:00
|
|
|
/*
|
|
|
|
|
* vma_next() - Get the next VMA.
|
|
|
|
|
* @mm: The mm_struct.
|
|
|
|
|
* @vma: The current vma.
|
|
|
|
|
*
|
|
|
|
|
* If @vma is NULL, return the first vma in the mm.
|
|
|
|
|
*
|
|
|
|
|
* Returns: The next VMA after @vma.
|
|
|
|
|
*/
|
|
|
|
|
static inline struct vm_area_struct *vma_next(struct mm_struct *mm,
|
|
|
|
|
struct vm_area_struct *vma)
|
|
|
|
|
{
|
|
|
|
|
if (!vma)
|
|
|
|
|
return mm->mmap;
|
|
|
|
|
|
|
|
|
|
return vma->vm_next;
|
|
|
|
|
}
|
2020-10-18 07:14:09 +08:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* munmap_vma_range() - munmap VMAs that overlap a range.
|
|
|
|
|
* @mm: The mm struct
|
|
|
|
|
* @start: The start of the range.
|
|
|
|
|
* @len: The length of the range.
|
|
|
|
|
* @pprev: pointer to the pointer that will be set to previous vm_area_struct
|
|
|
|
|
* @rb_link: the rb_node
|
|
|
|
|
* @rb_parent: the parent rb_node
|
|
|
|
|
*
|
|
|
|
|
* Find all the vm_area_struct that overlap from @start to
|
|
|
|
|
* @end and munmap them. Set @pprev to the previous vm_area_struct.
|
|
|
|
|
*
|
|
|
|
|
* Returns: -ENOMEM on munmap failure or 0 on success.
|
|
|
|
|
*/
|
|
|
|
|
static inline int
|
|
|
|
|
munmap_vma_range(struct mm_struct *mm, unsigned long start, unsigned long len,
|
|
|
|
|
struct vm_area_struct **pprev, struct rb_node ***link,
|
|
|
|
|
struct rb_node **parent, struct list_head *uf)
|
|
|
|
|
{
|
|
|
|
|
|
|
|
|
|
while (find_vma_links(mm, start, start + len, pprev, link, parent))
|
|
|
|
|
if (do_munmap(mm, start, len, uf))
|
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
2013-04-30 06:08:33 +08:00
|
|
|
static unsigned long count_vma_pages_range(struct mm_struct *mm,
|
|
|
|
|
unsigned long addr, unsigned long end)
|
|
|
|
|
{
|
|
|
|
|
unsigned long nr_pages = 0;
|
|
|
|
|
struct vm_area_struct *vma;
|
|
|
|
|
|
2021-05-07 09:06:47 +08:00
|
|
|
/* Find first overlapping mapping */
|
2013-04-30 06:08:33 +08:00
|
|
|
vma = find_vma_intersection(mm, addr, end);
|
|
|
|
|
if (!vma)
|
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
|
|
nr_pages = (min(end, vma->vm_end) -
|
|
|
|
|
max(addr, vma->vm_start)) >> PAGE_SHIFT;
|
|
|
|
|
|
|
|
|
|
/* Iterate over the rest of the overlaps */
|
|
|
|
|
for (vma = vma->vm_next; vma; vma = vma->vm_next) {
|
|
|
|
|
unsigned long overlap_len;
|
|
|
|
|
|
|
|
|
|
if (vma->vm_start > end)
|
|
|
|
|
break;
|
|
|
|
|
|
|
|
|
|
overlap_len = min(end, vma->vm_end) - vma->vm_start;
|
|
|
|
|
nr_pages += overlap_len >> PAGE_SHIFT;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return nr_pages;
|
|
|
|
|
}
|
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
|
|
|
|
|
struct rb_node **rb_link, struct rb_node *rb_parent)
|
|
|
|
|
{
|
2012-12-12 08:01:38 +08:00
|
|
|
/* Update tracking information for the gap following the new vma. */
|
|
|
|
|
if (vma->vm_next)
|
|
|
|
|
vma_gap_update(vma->vm_next);
|
|
|
|
|
else
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
mm->highest_vm_end = vm_end_gap(vma);
|
2012-12-12 08:01:38 +08:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* vma->vm_prev wasn't known when we followed the rbtree to find the
|
|
|
|
|
* correct insertion point for that vma. As a result, we could not
|
|
|
|
|
* update the vma vm_rb parents rb_subtree_gap values on the way down.
|
|
|
|
|
* So, we first insert the vma with a zero rb_subtree_gap value
|
|
|
|
|
* (to be consistent with what we did on the way down), and then
|
|
|
|
|
* immediately update the gap to the correct value. Finally we
|
|
|
|
|
* rebalance the rbtree after all augmented values have been set.
|
|
|
|
|
*/
|
2005-04-17 06:20:36 +08:00
|
|
|
rb_link_node(&vma->vm_rb, rb_parent, rb_link);
|
2012-12-12 08:01:38 +08:00
|
|
|
vma->rb_subtree_gap = 0;
|
|
|
|
|
vma_gap_update(vma);
|
|
|
|
|
vma_rb_insert(vma, &mm->mm_rb);
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
|
2008-10-19 11:27:01 +08:00
|
|
|
static void __vma_link_file(struct vm_area_struct *vma)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
2009-01-07 06:40:21 +08:00
|
|
|
struct file *file;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
|
file = vma->vm_file;
|
|
|
|
|
if (file) {
|
|
|
|
|
struct address_space *mapping = file->f_mapping;
|
|
|
|
|
|
|
|
|
|
if (vma->vm_flags & VM_SHARED)
|
2020-10-14 07:54:10 +08:00
|
|
|
mapping_allow_writable(mapping);
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
|
flush_dcache_mmap_lock(mapping);
|
2015-02-11 06:09:59 +08:00
|
|
|
vma_interval_tree_insert(vma, &mapping->i_mmap);
|
2005-04-17 06:20:36 +08:00
|
|
|
flush_dcache_mmap_unlock(mapping);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void
|
|
|
|
|
__vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
|
|
|
|
|
struct vm_area_struct *prev, struct rb_node **rb_link,
|
|
|
|
|
struct rb_node *rb_parent)
|
|
|
|
|
{
|
2019-12-01 09:50:53 +08:00
|
|
|
__vma_link_list(mm, vma, prev);
|
2005-04-17 06:20:36 +08:00
|
|
|
__vma_link_rb(mm, vma, rb_link, rb_parent);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
|
|
|
|
|
struct vm_area_struct *prev, struct rb_node **rb_link,
|
|
|
|
|
struct rb_node *rb_parent)
|
|
|
|
|
{
|
|
|
|
|
struct address_space *mapping = NULL;
|
|
|
|
|
|
2014-06-05 07:07:33 +08:00
|
|
|
if (vma->vm_file) {
|
2005-04-17 06:20:36 +08:00
|
|
|
mapping = vma->vm_file->f_mapping;
|
2014-12-13 08:54:21 +08:00
|
|
|
i_mmap_lock_write(mapping);
|
2014-06-05 07:07:33 +08:00
|
|
|
}
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
|
__vma_link(mm, vma, prev, rb_link, rb_parent);
|
|
|
|
|
__vma_link_file(vma);
|
|
|
|
|
|
|
|
|
|
if (mapping)
|
2014-12-13 08:54:21 +08:00
|
|
|
i_mmap_unlock_write(mapping);
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
|
mm->map_count++;
|
|
|
|
|
validate_mm(mm);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
2012-03-22 07:34:16 +08:00
|
|
|
* Helper for vma_adjust() in the split_vma insert case: insert a vma into the
|
2012-10-09 07:31:25 +08:00
|
|
|
* mm's list and rbtree. It has already been inserted into the interval tree.
|
2005-04-17 06:20:36 +08:00
|
|
|
*/
|
2009-01-07 06:40:21 +08:00
|
|
|
static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
2012-10-09 07:29:07 +08:00
|
|
|
struct vm_area_struct *prev;
|
2009-01-07 06:40:21 +08:00
|
|
|
struct rb_node **rb_link, *rb_parent;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2012-10-09 07:29:07 +08:00
|
|
|
if (find_vma_links(mm, vma->vm_start, vma->vm_end,
|
|
|
|
|
&prev, &rb_link, &rb_parent))
|
|
|
|
|
BUG();
|
2005-04-17 06:20:36 +08:00
|
|
|
__vma_link(mm, vma, prev, rb_link, rb_parent);
|
|
|
|
|
mm->map_count++;
|
|
|
|
|
}
|
|
|
|
|
|
2020-10-14 07:53:32 +08:00
|
|
|
static __always_inline void __vma_unlink(struct mm_struct *mm,
|
mm: vma_merge: fix vm_page_prot SMP race condition against rmap_walk
The rmap_walk can access vm_page_prot (and potentially vm_flags in the
pte/pmd manipulations). So it's not safe to wait the caller to update
the vm_page_prot/vm_flags after vma_merge returned potentially removing
the "next" vma and extending the "current" vma over the
next->vm_start,vm_end range, but still with the "current" vma
vm_page_prot, after releasing the rmap locks.
The vm_page_prot/vm_flags must be transferred from the "next" vma to the
current vma while vma_merge still holds the rmap locks.
The side effect of this race condition is pte corruption during migrate
as remove_migration_ptes when run on a address of the "next" vma that
got removed, used the vm_page_prot of the current vma.
migrate mprotect
------------ -------------
migrating in "next" vma
vma_merge() # removes "next" vma and
# extends "current" vma
# current vma is not with
# vm_page_prot updated
remove_migration_ptes
read vm_page_prot of current "vma"
establish pte with wrong permissions
vm_set_page_prot(vma) # too late!
change_protection in the old vma range
only, next range is not updated
This caused segmentation faults and potentially memory corruption in
heavy mprotect loads with some light page migration caused by compaction
in the background.
Hugh Dickins pointed out the comment about the Odd case 8 in vma_merge
which confirms the case 8 is only buggy one where the race can trigger,
in all other vma_merge cases the above cannot happen.
This fix removes the oddness factor from case 8 and it converts it from:
AAAA
PPPPNNNNXXXX -> PPPPNNNNNNNN
to:
AAAA
PPPPNNNNXXXX -> PPPPXXXXXXXX
XXXX has the right vma properties for the whole merged vma returned by
vma_adjust, so it solves the problem fully. It has the added benefits
that the callers could stop updating vma properties when vma_merge
succeeds however the callers are not updated by this patch (there are
bits like VM_SOFTDIRTY that still need special care for the whole range,
as the vma merging ignores them, but as long as they're not processed by
rmap walks and instead they're accessed with the mmap_sem at least for
reading, they are fine not to be updated within vma_adjust before
releasing the rmap_locks).
Link: http://lkml.kernel.org/r/1474309513-20313-1-git-send-email-aarcange@redhat.com
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Aditya Mandaleeka <adityam@microsoft.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Vorlicek <janvorli@microsoft.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-10-08 08:01:28 +08:00
|
|
|
struct vm_area_struct *vma,
|
2016-10-08 08:01:37 +08:00
|
|
|
struct vm_area_struct *ignore)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
2016-10-08 08:01:37 +08:00
|
|
|
vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
|
2019-12-01 09:50:49 +08:00
|
|
|
__vma_unlink_list(mm, vma);
|
mm: per-thread vma caching
This patch is a continuation of efforts trying to optimize find_vma(),
avoiding potentially expensive rbtree walks to locate a vma upon faults.
The original approach (https://lkml.org/lkml/2013/11/1/410), where the
largest vma was also cached, ended up being too specific and random,
thus further comparison with other approaches were needed. There are
two things to consider when dealing with this, the cache hit rate and
the latency of find_vma(). Improving the hit-rate does not necessarily
translate in finding the vma any faster, as the overhead of any fancy
caching schemes can be too high to consider.
We currently cache the last used vma for the whole address space, which
provides a nice optimization, reducing the total cycles in find_vma() by
up to 250%, for workloads with good locality. On the other hand, this
simple scheme is pretty much useless for workloads with poor locality.
Analyzing ebizzy runs shows that, no matter how many threads are
running, the mmap_cache hit rate is less than 2%, and in many situations
below 1%.
The proposed approach is to replace this scheme with a small per-thread
cache, maximizing hit rates at a very low maintenance cost.
Invalidations are performed by simply bumping up a 32-bit sequence
number. The only expensive operation is in the rare case of a seq
number overflow, where all caches that share the same address space are
flushed. Upon a miss, the proposed replacement policy is based on the
page number that contains the virtual address in question. Concretely,
the following results are seen on an 80 core, 8 socket x86-64 box:
1) System bootup: Most programs are single threaded, so the per-thread
scheme does improve ~50% hit rate by just adding a few more slots to
the cache.
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 50.61% | 19.90 |
| patched | 73.45% | 13.58 |
+----------------+----------+------------------+
2) Kernel build: This one is already pretty good with the current
approach as we're dealing with good locality.
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 75.28% | 11.03 |
| patched | 88.09% | 9.31 |
+----------------+----------+------------------+
3) Oracle 11g Data Mining (4k pages): Similar to the kernel build workload.
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 70.66% | 17.14 |
| patched | 91.15% | 12.57 |
+----------------+----------+------------------+
4) Ebizzy: There's a fair amount of variation from run to run, but this
approach always shows nearly perfect hit rates, while baseline is just
about non-existent. The amounts of cycles can fluctuate between
anywhere from ~60 to ~116 for the baseline scheme, but this approach
reduces it considerably. For instance, with 80 threads:
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 1.06% | 91.54 |
| patched | 99.97% | 14.18 |
+----------------+----------+------------------+
[akpm@linux-foundation.org: fix nommu build, per Davidlohr]
[akpm@linux-foundation.org: document vmacache_valid() logic]
[akpm@linux-foundation.org: attempt to untangle header files]
[akpm@linux-foundation.org: add vmacache_find() BUG_ON]
[hughd@google.com: add vmacache_valid_mm() (from Oleg)]
[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: adjust and enhance comments]
Signed-off-by: Davidlohr Bueso <davidlohr@hp.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Reviewed-by: Michel Lespinasse <walken@google.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Tested-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-04-08 06:37:25 +08:00
|
|
|
/* Kill the cache */
|
|
|
|
|
vmacache_invalidate(mm);
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
|
|
|
|
|
* is already present in an i_mmap tree without adjusting the tree.
|
|
|
|
|
* The following helper function should be used when such adjustments
|
|
|
|
|
* are necessary. The "insert" vma (if any) is to be inserted
|
|
|
|
|
* before we drop the necessary locks.
|
|
|
|
|
*/
|
mm: vma_merge: fix vm_page_prot SMP race condition against rmap_walk
The rmap_walk can access vm_page_prot (and potentially vm_flags in the
pte/pmd manipulations). So it's not safe to wait the caller to update
the vm_page_prot/vm_flags after vma_merge returned potentially removing
the "next" vma and extending the "current" vma over the
next->vm_start,vm_end range, but still with the "current" vma
vm_page_prot, after releasing the rmap locks.
The vm_page_prot/vm_flags must be transferred from the "next" vma to the
current vma while vma_merge still holds the rmap locks.
The side effect of this race condition is pte corruption during migrate
as remove_migration_ptes when run on a address of the "next" vma that
got removed, used the vm_page_prot of the current vma.
migrate mprotect
------------ -------------
migrating in "next" vma
vma_merge() # removes "next" vma and
# extends "current" vma
# current vma is not with
# vm_page_prot updated
remove_migration_ptes
read vm_page_prot of current "vma"
establish pte with wrong permissions
vm_set_page_prot(vma) # too late!
change_protection in the old vma range
only, next range is not updated
This caused segmentation faults and potentially memory corruption in
heavy mprotect loads with some light page migration caused by compaction
in the background.
Hugh Dickins pointed out the comment about the Odd case 8 in vma_merge
which confirms the case 8 is only buggy one where the race can trigger,
in all other vma_merge cases the above cannot happen.
This fix removes the oddness factor from case 8 and it converts it from:
AAAA
PPPPNNNNXXXX -> PPPPNNNNNNNN
to:
AAAA
PPPPNNNNXXXX -> PPPPXXXXXXXX
XXXX has the right vma properties for the whole merged vma returned by
vma_adjust, so it solves the problem fully. It has the added benefits
that the callers could stop updating vma properties when vma_merge
succeeds however the callers are not updated by this patch (there are
bits like VM_SOFTDIRTY that still need special care for the whole range,
as the vma merging ignores them, but as long as they're not processed by
rmap walks and instead they're accessed with the mmap_sem at least for
reading, they are fine not to be updated within vma_adjust before
releasing the rmap_locks).
Link: http://lkml.kernel.org/r/1474309513-20313-1-git-send-email-aarcange@redhat.com
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Aditya Mandaleeka <adityam@microsoft.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Vorlicek <janvorli@microsoft.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-10-08 08:01:28 +08:00
|
|
|
int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
|
|
|
|
|
unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
|
|
|
|
|
struct vm_area_struct *expand)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
|
|
|
|
struct mm_struct *mm = vma->vm_mm;
|
mm: vma_merge: fix vm_page_prot SMP race condition against rmap_walk
The rmap_walk can access vm_page_prot (and potentially vm_flags in the
pte/pmd manipulations). So it's not safe to wait the caller to update
the vm_page_prot/vm_flags after vma_merge returned potentially removing
the "next" vma and extending the "current" vma over the
next->vm_start,vm_end range, but still with the "current" vma
vm_page_prot, after releasing the rmap locks.
The vm_page_prot/vm_flags must be transferred from the "next" vma to the
current vma while vma_merge still holds the rmap locks.
The side effect of this race condition is pte corruption during migrate
as remove_migration_ptes when run on a address of the "next" vma that
got removed, used the vm_page_prot of the current vma.
migrate mprotect
------------ -------------
migrating in "next" vma
vma_merge() # removes "next" vma and
# extends "current" vma
# current vma is not with
# vm_page_prot updated
remove_migration_ptes
read vm_page_prot of current "vma"
establish pte with wrong permissions
vm_set_page_prot(vma) # too late!
change_protection in the old vma range
only, next range is not updated
This caused segmentation faults and potentially memory corruption in
heavy mprotect loads with some light page migration caused by compaction
in the background.
Hugh Dickins pointed out the comment about the Odd case 8 in vma_merge
which confirms the case 8 is only buggy one where the race can trigger,
in all other vma_merge cases the above cannot happen.
This fix removes the oddness factor from case 8 and it converts it from:
AAAA
PPPPNNNNXXXX -> PPPPNNNNNNNN
to:
AAAA
PPPPNNNNXXXX -> PPPPXXXXXXXX
XXXX has the right vma properties for the whole merged vma returned by
vma_adjust, so it solves the problem fully. It has the added benefits
that the callers could stop updating vma properties when vma_merge
succeeds however the callers are not updated by this patch (there are
bits like VM_SOFTDIRTY that still need special care for the whole range,
as the vma merging ignores them, but as long as they're not processed by
rmap walks and instead they're accessed with the mmap_sem at least for
reading, they are fine not to be updated within vma_adjust before
releasing the rmap_locks).
Link: http://lkml.kernel.org/r/1474309513-20313-1-git-send-email-aarcange@redhat.com
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Aditya Mandaleeka <adityam@microsoft.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Vorlicek <janvorli@microsoft.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-10-08 08:01:28 +08:00
|
|
|
struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
|
2005-04-17 06:20:36 +08:00
|
|
|
struct address_space *mapping = NULL;
|
2017-09-09 07:15:08 +08:00
|
|
|
struct rb_root_cached *root = NULL;
|
2010-08-10 08:18:40 +08:00
|
|
|
struct anon_vma *anon_vma = NULL;
|
2005-04-17 06:20:36 +08:00
|
|
|
struct file *file = vma->vm_file;
|
2012-12-12 08:01:38 +08:00
|
|
|
bool start_changed = false, end_changed = false;
|
2005-04-17 06:20:36 +08:00
|
|
|
long adjust_next = 0;
|
|
|
|
|
int remove_next = 0;
|
|
|
|
|
|
|
|
|
|
if (next && !insert) {
|
2016-07-29 06:49:01 +08:00
|
|
|
struct vm_area_struct *exporter = NULL, *importer = NULL;
|
2010-04-11 06:22:30 +08:00
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
if (end >= next->vm_end) {
|
|
|
|
|
/*
|
|
|
|
|
* vma expands, overlapping all the next, and
|
|
|
|
|
* perhaps the one after too (mprotect case 6).
|
2016-10-08 08:01:34 +08:00
|
|
|
* The only other cases that gets here are
|
mm: vma_merge: fix vm_page_prot SMP race condition against rmap_walk
The rmap_walk can access vm_page_prot (and potentially vm_flags in the
pte/pmd manipulations). So it's not safe to wait the caller to update
the vm_page_prot/vm_flags after vma_merge returned potentially removing
the "next" vma and extending the "current" vma over the
next->vm_start,vm_end range, but still with the "current" vma
vm_page_prot, after releasing the rmap locks.
The vm_page_prot/vm_flags must be transferred from the "next" vma to the
current vma while vma_merge still holds the rmap locks.
The side effect of this race condition is pte corruption during migrate
as remove_migration_ptes when run on a address of the "next" vma that
got removed, used the vm_page_prot of the current vma.
migrate mprotect
------------ -------------
migrating in "next" vma
vma_merge() # removes "next" vma and
# extends "current" vma
# current vma is not with
# vm_page_prot updated
remove_migration_ptes
read vm_page_prot of current "vma"
establish pte with wrong permissions
vm_set_page_prot(vma) # too late!
change_protection in the old vma range
only, next range is not updated
This caused segmentation faults and potentially memory corruption in
heavy mprotect loads with some light page migration caused by compaction
in the background.
Hugh Dickins pointed out the comment about the Odd case 8 in vma_merge
which confirms the case 8 is only buggy one where the race can trigger,
in all other vma_merge cases the above cannot happen.
This fix removes the oddness factor from case 8 and it converts it from:
AAAA
PPPPNNNNXXXX -> PPPPNNNNNNNN
to:
AAAA
PPPPNNNNXXXX -> PPPPXXXXXXXX
XXXX has the right vma properties for the whole merged vma returned by
vma_adjust, so it solves the problem fully. It has the added benefits
that the callers could stop updating vma properties when vma_merge
succeeds however the callers are not updated by this patch (there are
bits like VM_SOFTDIRTY that still need special care for the whole range,
as the vma merging ignores them, but as long as they're not processed by
rmap walks and instead they're accessed with the mmap_sem at least for
reading, they are fine not to be updated within vma_adjust before
releasing the rmap_locks).
Link: http://lkml.kernel.org/r/1474309513-20313-1-git-send-email-aarcange@redhat.com
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Aditya Mandaleeka <adityam@microsoft.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Vorlicek <janvorli@microsoft.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-10-08 08:01:28 +08:00
|
|
|
* case 1, case 7 and case 8.
|
2005-04-17 06:20:36 +08:00
|
|
|
*/
|
mm: vma_merge: fix vm_page_prot SMP race condition against rmap_walk
The rmap_walk can access vm_page_prot (and potentially vm_flags in the
pte/pmd manipulations). So it's not safe to wait the caller to update
the vm_page_prot/vm_flags after vma_merge returned potentially removing
the "next" vma and extending the "current" vma over the
next->vm_start,vm_end range, but still with the "current" vma
vm_page_prot, after releasing the rmap locks.
The vm_page_prot/vm_flags must be transferred from the "next" vma to the
current vma while vma_merge still holds the rmap locks.
The side effect of this race condition is pte corruption during migrate
as remove_migration_ptes when run on a address of the "next" vma that
got removed, used the vm_page_prot of the current vma.
migrate mprotect
------------ -------------
migrating in "next" vma
vma_merge() # removes "next" vma and
# extends "current" vma
# current vma is not with
# vm_page_prot updated
remove_migration_ptes
read vm_page_prot of current "vma"
establish pte with wrong permissions
vm_set_page_prot(vma) # too late!
change_protection in the old vma range
only, next range is not updated
This caused segmentation faults and potentially memory corruption in
heavy mprotect loads with some light page migration caused by compaction
in the background.
Hugh Dickins pointed out the comment about the Odd case 8 in vma_merge
which confirms the case 8 is only buggy one where the race can trigger,
in all other vma_merge cases the above cannot happen.
This fix removes the oddness factor from case 8 and it converts it from:
AAAA
PPPPNNNNXXXX -> PPPPNNNNNNNN
to:
AAAA
PPPPNNNNXXXX -> PPPPXXXXXXXX
XXXX has the right vma properties for the whole merged vma returned by
vma_adjust, so it solves the problem fully. It has the added benefits
that the callers could stop updating vma properties when vma_merge
succeeds however the callers are not updated by this patch (there are
bits like VM_SOFTDIRTY that still need special care for the whole range,
as the vma merging ignores them, but as long as they're not processed by
rmap walks and instead they're accessed with the mmap_sem at least for
reading, they are fine not to be updated within vma_adjust before
releasing the rmap_locks).
Link: http://lkml.kernel.org/r/1474309513-20313-1-git-send-email-aarcange@redhat.com
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Aditya Mandaleeka <adityam@microsoft.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Vorlicek <janvorli@microsoft.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-10-08 08:01:28 +08:00
|
|
|
if (next == expand) {
|
|
|
|
|
/*
|
|
|
|
|
* The only case where we don't expand "vma"
|
|
|
|
|
* and we expand "next" instead is case 8.
|
|
|
|
|
*/
|
|
|
|
|
VM_WARN_ON(end != next->vm_end);
|
|
|
|
|
/*
|
|
|
|
|
* remove_next == 3 means we're
|
|
|
|
|
* removing "vma" and that to do so we
|
|
|
|
|
* swapped "vma" and "next".
|
|
|
|
|
*/
|
|
|
|
|
remove_next = 3;
|
|
|
|
|
VM_WARN_ON(file != next->vm_file);
|
|
|
|
|
swap(vma, next);
|
|
|
|
|
} else {
|
|
|
|
|
VM_WARN_ON(expand != vma);
|
|
|
|
|
/*
|
|
|
|
|
* case 1, 6, 7, remove_next == 2 is case 6,
|
|
|
|
|
* remove_next == 1 is case 1 or 7.
|
|
|
|
|
*/
|
|
|
|
|
remove_next = 1 + (end > next->vm_end);
|
|
|
|
|
VM_WARN_ON(remove_next == 2 &&
|
|
|
|
|
end != next->vm_next->vm_end);
|
|
|
|
|
/* trim end to next, for case 6 first pass */
|
|
|
|
|
end = next->vm_end;
|
|
|
|
|
}
|
|
|
|
|
|
2010-04-11 06:22:30 +08:00
|
|
|
exporter = next;
|
2005-04-17 06:20:36 +08:00
|
|
|
importer = vma;
|
2016-07-29 06:49:01 +08:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* If next doesn't have anon_vma, import from vma after
|
|
|
|
|
* next, if the vma overlaps with it.
|
|
|
|
|
*/
|
2016-10-08 08:01:31 +08:00
|
|
|
if (remove_next == 2 && !next->anon_vma)
|
2016-07-29 06:49:01 +08:00
|
|
|
exporter = next->vm_next;
|
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
} else if (end > next->vm_start) {
|
|
|
|
|
/*
|
|
|
|
|
* vma expands, overlapping part of the next:
|
|
|
|
|
* mprotect case 5 shifting the boundary up.
|
|
|
|
|
*/
|
2020-10-14 07:53:57 +08:00
|
|
|
adjust_next = (end - next->vm_start);
|
2010-04-11 06:22:30 +08:00
|
|
|
exporter = next;
|
2005-04-17 06:20:36 +08:00
|
|
|
importer = vma;
|
mm: vma_merge: fix vm_page_prot SMP race condition against rmap_walk
The rmap_walk can access vm_page_prot (and potentially vm_flags in the
pte/pmd manipulations). So it's not safe to wait the caller to update
the vm_page_prot/vm_flags after vma_merge returned potentially removing
the "next" vma and extending the "current" vma over the
next->vm_start,vm_end range, but still with the "current" vma
vm_page_prot, after releasing the rmap locks.
The vm_page_prot/vm_flags must be transferred from the "next" vma to the
current vma while vma_merge still holds the rmap locks.
The side effect of this race condition is pte corruption during migrate
as remove_migration_ptes when run on a address of the "next" vma that
got removed, used the vm_page_prot of the current vma.
migrate mprotect
------------ -------------
migrating in "next" vma
vma_merge() # removes "next" vma and
# extends "current" vma
# current vma is not with
# vm_page_prot updated
remove_migration_ptes
read vm_page_prot of current "vma"
establish pte with wrong permissions
vm_set_page_prot(vma) # too late!
change_protection in the old vma range
only, next range is not updated
This caused segmentation faults and potentially memory corruption in
heavy mprotect loads with some light page migration caused by compaction
in the background.
Hugh Dickins pointed out the comment about the Odd case 8 in vma_merge
which confirms the case 8 is only buggy one where the race can trigger,
in all other vma_merge cases the above cannot happen.
This fix removes the oddness factor from case 8 and it converts it from:
AAAA
PPPPNNNNXXXX -> PPPPNNNNNNNN
to:
AAAA
PPPPNNNNXXXX -> PPPPXXXXXXXX
XXXX has the right vma properties for the whole merged vma returned by
vma_adjust, so it solves the problem fully. It has the added benefits
that the callers could stop updating vma properties when vma_merge
succeeds however the callers are not updated by this patch (there are
bits like VM_SOFTDIRTY that still need special care for the whole range,
as the vma merging ignores them, but as long as they're not processed by
rmap walks and instead they're accessed with the mmap_sem at least for
reading, they are fine not to be updated within vma_adjust before
releasing the rmap_locks).
Link: http://lkml.kernel.org/r/1474309513-20313-1-git-send-email-aarcange@redhat.com
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Aditya Mandaleeka <adityam@microsoft.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Vorlicek <janvorli@microsoft.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-10-08 08:01:28 +08:00
|
|
|
VM_WARN_ON(expand != importer);
|
2005-04-17 06:20:36 +08:00
|
|
|
} else if (end < vma->vm_end) {
|
|
|
|
|
/*
|
|
|
|
|
* vma shrinks, and !insert tells it's not
|
|
|
|
|
* split_vma inserting another: so it must be
|
|
|
|
|
* mprotect case 4 shifting the boundary down.
|
|
|
|
|
*/
|
2020-10-14 07:53:57 +08:00
|
|
|
adjust_next = -(vma->vm_end - end);
|
2010-04-11 06:22:30 +08:00
|
|
|
exporter = vma;
|
2005-04-17 06:20:36 +08:00
|
|
|
importer = next;
|
mm: vma_merge: fix vm_page_prot SMP race condition against rmap_walk
The rmap_walk can access vm_page_prot (and potentially vm_flags in the
pte/pmd manipulations). So it's not safe to wait the caller to update
the vm_page_prot/vm_flags after vma_merge returned potentially removing
the "next" vma and extending the "current" vma over the
next->vm_start,vm_end range, but still with the "current" vma
vm_page_prot, after releasing the rmap locks.
The vm_page_prot/vm_flags must be transferred from the "next" vma to the
current vma while vma_merge still holds the rmap locks.
The side effect of this race condition is pte corruption during migrate
as remove_migration_ptes when run on a address of the "next" vma that
got removed, used the vm_page_prot of the current vma.
migrate mprotect
------------ -------------
migrating in "next" vma
vma_merge() # removes "next" vma and
# extends "current" vma
# current vma is not with
# vm_page_prot updated
remove_migration_ptes
read vm_page_prot of current "vma"
establish pte with wrong permissions
vm_set_page_prot(vma) # too late!
change_protection in the old vma range
only, next range is not updated
This caused segmentation faults and potentially memory corruption in
heavy mprotect loads with some light page migration caused by compaction
in the background.
Hugh Dickins pointed out the comment about the Odd case 8 in vma_merge
which confirms the case 8 is only buggy one where the race can trigger,
in all other vma_merge cases the above cannot happen.
This fix removes the oddness factor from case 8 and it converts it from:
AAAA
PPPPNNNNXXXX -> PPPPNNNNNNNN
to:
AAAA
PPPPNNNNXXXX -> PPPPXXXXXXXX
XXXX has the right vma properties for the whole merged vma returned by
vma_adjust, so it solves the problem fully. It has the added benefits
that the callers could stop updating vma properties when vma_merge
succeeds however the callers are not updated by this patch (there are
bits like VM_SOFTDIRTY that still need special care for the whole range,
as the vma merging ignores them, but as long as they're not processed by
rmap walks and instead they're accessed with the mmap_sem at least for
reading, they are fine not to be updated within vma_adjust before
releasing the rmap_locks).
Link: http://lkml.kernel.org/r/1474309513-20313-1-git-send-email-aarcange@redhat.com
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Aditya Mandaleeka <adityam@microsoft.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Vorlicek <janvorli@microsoft.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-10-08 08:01:28 +08:00
|
|
|
VM_WARN_ON(expand != importer);
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
|
mm: change anon_vma linking to fix multi-process server scalability issue
The old anon_vma code can lead to scalability issues with heavily forking
workloads. Specifically, each anon_vma will be shared between the parent
process and all its child processes.
In a workload with 1000 child processes and a VMA with 1000 anonymous
pages per process that get COWed, this leads to a system with a million
anonymous pages in the same anon_vma, each of which is mapped in just one
of the 1000 processes. However, the current rmap code needs to walk them
all, leading to O(N) scanning complexity for each page.
This can result in systems where one CPU is walking the page tables of
1000 processes in page_referenced_one, while all other CPUs are stuck on
the anon_vma lock. This leads to catastrophic failure for a benchmark
like AIM7, where the total number of processes can reach in the tens of
thousands. Real workloads are still a factor 10 less process intensive
than AIM7, but they are catching up.
This patch changes the way anon_vmas and VMAs are linked, which allows us
to associate multiple anon_vmas with a VMA. At fork time, each child
process gets its own anon_vmas, in which its COWed pages will be
instantiated. The parents' anon_vma is also linked to the VMA, because
non-COWed pages could be present in any of the children.
This reduces rmap scanning complexity to O(1) for the pages of the 1000
child processes, with O(N) complexity for at most 1/N pages in the system.
This reduces the average scanning cost in heavily forking workloads from
O(N) to 2.
The only real complexity in this patch stems from the fact that linking a
VMA to anon_vmas now involves memory allocations. This means vma_adjust
can fail, if it needs to attach a VMA to anon_vma structures. This in
turn means error handling needs to be added to the calling functions.
A second source of complexity is that, because there can be multiple
anon_vmas, the anon_vma linking in vma_adjust can no longer be done under
"the" anon_vma lock. To prevent the rmap code from walking up an
incomplete VMA, this patch introduces the VM_LOCK_RMAP VMA flag. This bit
flag uses the same slot as the NOMMU VM_MAPPED_COPY, with an ifdef in mm.h
to make sure it is impossible to compile a kernel that needs both symbolic
values for the same bitflag.
Some test results:
Without the anon_vma changes, when AIM7 hits around 9.7k users (on a test
box with 16GB RAM and not quite enough IO), the system ends up running
>99% in system time, with every CPU on the same anon_vma lock in the
pageout code.
With these changes, AIM7 hits the cross-over point around 29.7k users.
This happens with ~99% IO wait time, there never seems to be any spike in
system time. The anon_vma lock contention appears to be resolved.
[akpm@linux-foundation.org: cleanups]
Signed-off-by: Rik van Riel <riel@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Larry Woodman <lwoodman@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-06 05:42:07 +08:00
|
|
|
/*
|
|
|
|
|
* Easily overlooked: when mprotect shifts the boundary,
|
|
|
|
|
* make sure the expanding vma has anon_vma set if the
|
|
|
|
|
* shrinking vma had, to cover any anon pages imported.
|
|
|
|
|
*/
|
2010-04-11 06:22:30 +08:00
|
|
|
if (exporter && exporter->anon_vma && !importer->anon_vma) {
|
2014-12-03 07:59:42 +08:00
|
|
|
int error;
|
|
|
|
|
|
2015-01-11 21:54:06 +08:00
|
|
|
importer->anon_vma = exporter->anon_vma;
|
2014-12-03 07:59:42 +08:00
|
|
|
error = anon_vma_clone(importer, exporter);
|
2015-03-26 06:55:11 +08:00
|
|
|
if (error)
|
2014-12-03 07:59:42 +08:00
|
|
|
return error;
|
mm: change anon_vma linking to fix multi-process server scalability issue
The old anon_vma code can lead to scalability issues with heavily forking
workloads. Specifically, each anon_vma will be shared between the parent
process and all its child processes.
In a workload with 1000 child processes and a VMA with 1000 anonymous
pages per process that get COWed, this leads to a system with a million
anonymous pages in the same anon_vma, each of which is mapped in just one
of the 1000 processes. However, the current rmap code needs to walk them
all, leading to O(N) scanning complexity for each page.
This can result in systems where one CPU is walking the page tables of
1000 processes in page_referenced_one, while all other CPUs are stuck on
the anon_vma lock. This leads to catastrophic failure for a benchmark
like AIM7, where the total number of processes can reach in the tens of
thousands. Real workloads are still a factor 10 less process intensive
than AIM7, but they are catching up.
This patch changes the way anon_vmas and VMAs are linked, which allows us
to associate multiple anon_vmas with a VMA. At fork time, each child
process gets its own anon_vmas, in which its COWed pages will be
instantiated. The parents' anon_vma is also linked to the VMA, because
non-COWed pages could be present in any of the children.
This reduces rmap scanning complexity to O(1) for the pages of the 1000
child processes, with O(N) complexity for at most 1/N pages in the system.
This reduces the average scanning cost in heavily forking workloads from
O(N) to 2.
The only real complexity in this patch stems from the fact that linking a
VMA to anon_vmas now involves memory allocations. This means vma_adjust
can fail, if it needs to attach a VMA to anon_vma structures. This in
turn means error handling needs to be added to the calling functions.
A second source of complexity is that, because there can be multiple
anon_vmas, the anon_vma linking in vma_adjust can no longer be done under
"the" anon_vma lock. To prevent the rmap code from walking up an
incomplete VMA, this patch introduces the VM_LOCK_RMAP VMA flag. This bit
flag uses the same slot as the NOMMU VM_MAPPED_COPY, with an ifdef in mm.h
to make sure it is impossible to compile a kernel that needs both symbolic
values for the same bitflag.
Some test results:
Without the anon_vma changes, when AIM7 hits around 9.7k users (on a test
box with 16GB RAM and not quite enough IO), the system ends up running
>99% in system time, with every CPU on the same anon_vma lock in the
pageout code.
With these changes, AIM7 hits the cross-over point around 29.7k users.
This happens with ~99% IO wait time, there never seems to be any spike in
system time. The anon_vma lock contention appears to be resolved.
[akpm@linux-foundation.org: cleanups]
Signed-off-by: Rik van Riel <riel@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Larry Woodman <lwoodman@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-06 05:42:07 +08:00
|
|
|
}
|
|
|
|
|
}
|
2016-07-29 06:49:01 +08:00
|
|
|
again:
|
mm: vma_merge: fix vm_page_prot SMP race condition against rmap_walk
The rmap_walk can access vm_page_prot (and potentially vm_flags in the
pte/pmd manipulations). So it's not safe to wait the caller to update
the vm_page_prot/vm_flags after vma_merge returned potentially removing
the "next" vma and extending the "current" vma over the
next->vm_start,vm_end range, but still with the "current" vma
vm_page_prot, after releasing the rmap locks.
The vm_page_prot/vm_flags must be transferred from the "next" vma to the
current vma while vma_merge still holds the rmap locks.
The side effect of this race condition is pte corruption during migrate
as remove_migration_ptes when run on a address of the "next" vma that
got removed, used the vm_page_prot of the current vma.
migrate mprotect
------------ -------------
migrating in "next" vma
vma_merge() # removes "next" vma and
# extends "current" vma
# current vma is not with
# vm_page_prot updated
remove_migration_ptes
read vm_page_prot of current "vma"
establish pte with wrong permissions
vm_set_page_prot(vma) # too late!
change_protection in the old vma range
only, next range is not updated
This caused segmentation faults and potentially memory corruption in
heavy mprotect loads with some light page migration caused by compaction
in the background.
Hugh Dickins pointed out the comment about the Odd case 8 in vma_merge
which confirms the case 8 is only buggy one where the race can trigger,
in all other vma_merge cases the above cannot happen.
This fix removes the oddness factor from case 8 and it converts it from:
AAAA
PPPPNNNNXXXX -> PPPPNNNNNNNN
to:
AAAA
PPPPNNNNXXXX -> PPPPXXXXXXXX
XXXX has the right vma properties for the whole merged vma returned by
vma_adjust, so it solves the problem fully. It has the added benefits
that the callers could stop updating vma properties when vma_merge
succeeds however the callers are not updated by this patch (there are
bits like VM_SOFTDIRTY that still need special care for the whole range,
as the vma merging ignores them, but as long as they're not processed by
rmap walks and instead they're accessed with the mmap_sem at least for
reading, they are fine not to be updated within vma_adjust before
releasing the rmap_locks).
Link: http://lkml.kernel.org/r/1474309513-20313-1-git-send-email-aarcange@redhat.com
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Aditya Mandaleeka <adityam@microsoft.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Vorlicek <janvorli@microsoft.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-10-08 08:01:28 +08:00
|
|
|
vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
|
2016-07-27 06:25:48 +08:00
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
if (file) {
|
|
|
|
|
mapping = file->f_mapping;
|
2015-02-11 06:09:59 +08:00
|
|
|
root = &mapping->i_mmap;
|
|
|
|
|
uprobe_munmap(vma, vma->vm_start, vma->vm_end);
|
2012-03-31 02:26:46 +08:00
|
|
|
|
2015-02-11 06:09:59 +08:00
|
|
|
if (adjust_next)
|
|
|
|
|
uprobe_munmap(next, next->vm_start, next->vm_end);
|
2012-03-31 02:26:46 +08:00
|
|
|
|
2014-12-13 08:54:21 +08:00
|
|
|
i_mmap_lock_write(mapping);
|
2005-04-17 06:20:36 +08:00
|
|
|
if (insert) {
|
|
|
|
|
/*
|
2012-10-09 07:31:25 +08:00
|
|
|
* Put into interval tree now, so instantiated pages
|
2005-04-17 06:20:36 +08:00
|
|
|
* are visible to arm/parisc __flush_dcache_page
|
|
|
|
|
* throughout; but we cannot insert into address
|
|
|
|
|
* space until vma start or end is updated.
|
|
|
|
|
*/
|
|
|
|
|
__vma_link_file(insert);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
mm anon rmap: replace same_anon_vma linked list with an interval tree.
When a large VMA (anon or private file mapping) is first touched, which
will populate its anon_vma field, and then split into many regions through
the use of mprotect(), the original anon_vma ends up linking all of the
vmas on a linked list. This can cause rmap to become inefficient, as we
have to walk potentially thousands of irrelevent vmas before finding the
one a given anon page might fall into.
By replacing the same_anon_vma linked list with an interval tree (where
each avc's interval is determined by its vma's start and last pgoffs), we
can make rmap efficient for this use case again.
While the change is large, all of its pieces are fairly simple.
Most places that were walking the same_anon_vma list were looking for a
known pgoff, so they can just use the anon_vma_interval_tree_foreach()
interval tree iterator instead. The exception here is ksm, where the
page's index is not known. It would probably be possible to rework ksm so
that the index would be known, but for now I have decided to keep things
simple and just walk the entirety of the interval tree there.
When updating vma's that already have an anon_vma assigned, we must take
care to re-index the corresponding avc's on their interval tree. This is
done through the use of anon_vma_interval_tree_pre_update_vma() and
anon_vma_interval_tree_post_update_vma(), which remove the avc's from
their interval tree before the update and re-insert them after the update.
The anon_vma stays locked during the update, so there is no chance that
rmap would miss the vmas that are being updated.
Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-10-09 07:31:39 +08:00
|
|
|
anon_vma = vma->anon_vma;
|
|
|
|
|
if (!anon_vma && adjust_next)
|
|
|
|
|
anon_vma = next->anon_vma;
|
|
|
|
|
if (anon_vma) {
|
mm: vma_merge: fix vm_page_prot SMP race condition against rmap_walk
The rmap_walk can access vm_page_prot (and potentially vm_flags in the
pte/pmd manipulations). So it's not safe to wait the caller to update
the vm_page_prot/vm_flags after vma_merge returned potentially removing
the "next" vma and extending the "current" vma over the
next->vm_start,vm_end range, but still with the "current" vma
vm_page_prot, after releasing the rmap locks.
The vm_page_prot/vm_flags must be transferred from the "next" vma to the
current vma while vma_merge still holds the rmap locks.
The side effect of this race condition is pte corruption during migrate
as remove_migration_ptes when run on a address of the "next" vma that
got removed, used the vm_page_prot of the current vma.
migrate mprotect
------------ -------------
migrating in "next" vma
vma_merge() # removes "next" vma and
# extends "current" vma
# current vma is not with
# vm_page_prot updated
remove_migration_ptes
read vm_page_prot of current "vma"
establish pte with wrong permissions
vm_set_page_prot(vma) # too late!
change_protection in the old vma range
only, next range is not updated
This caused segmentation faults and potentially memory corruption in
heavy mprotect loads with some light page migration caused by compaction
in the background.
Hugh Dickins pointed out the comment about the Odd case 8 in vma_merge
which confirms the case 8 is only buggy one where the race can trigger,
in all other vma_merge cases the above cannot happen.
This fix removes the oddness factor from case 8 and it converts it from:
AAAA
PPPPNNNNXXXX -> PPPPNNNNNNNN
to:
AAAA
PPPPNNNNXXXX -> PPPPXXXXXXXX
XXXX has the right vma properties for the whole merged vma returned by
vma_adjust, so it solves the problem fully. It has the added benefits
that the callers could stop updating vma properties when vma_merge
succeeds however the callers are not updated by this patch (there are
bits like VM_SOFTDIRTY that still need special care for the whole range,
as the vma merging ignores them, but as long as they're not processed by
rmap walks and instead they're accessed with the mmap_sem at least for
reading, they are fine not to be updated within vma_adjust before
releasing the rmap_locks).
Link: http://lkml.kernel.org/r/1474309513-20313-1-git-send-email-aarcange@redhat.com
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Aditya Mandaleeka <adityam@microsoft.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Vorlicek <janvorli@microsoft.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-10-08 08:01:28 +08:00
|
|
|
VM_WARN_ON(adjust_next && next->anon_vma &&
|
|
|
|
|
anon_vma != next->anon_vma);
|
2012-12-03 03:56:50 +08:00
|
|
|
anon_vma_lock_write(anon_vma);
|
mm anon rmap: replace same_anon_vma linked list with an interval tree.
When a large VMA (anon or private file mapping) is first touched, which
will populate its anon_vma field, and then split into many regions through
the use of mprotect(), the original anon_vma ends up linking all of the
vmas on a linked list. This can cause rmap to become inefficient, as we
have to walk potentially thousands of irrelevent vmas before finding the
one a given anon page might fall into.
By replacing the same_anon_vma linked list with an interval tree (where
each avc's interval is determined by its vma's start and last pgoffs), we
can make rmap efficient for this use case again.
While the change is large, all of its pieces are fairly simple.
Most places that were walking the same_anon_vma list were looking for a
known pgoff, so they can just use the anon_vma_interval_tree_foreach()
interval tree iterator instead. The exception here is ksm, where the
page's index is not known. It would probably be possible to rework ksm so
that the index would be known, but for now I have decided to keep things
simple and just walk the entirety of the interval tree there.
When updating vma's that already have an anon_vma assigned, we must take
care to re-index the corresponding avc's on their interval tree. This is
done through the use of anon_vma_interval_tree_pre_update_vma() and
anon_vma_interval_tree_post_update_vma(), which remove the avc's from
their interval tree before the update and re-insert them after the update.
The anon_vma stays locked during the update, so there is no chance that
rmap would miss the vmas that are being updated.
Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-10-09 07:31:39 +08:00
|
|
|
anon_vma_interval_tree_pre_update_vma(vma);
|
|
|
|
|
if (adjust_next)
|
|
|
|
|
anon_vma_interval_tree_pre_update_vma(next);
|
|
|
|
|
}
|
2010-08-10 08:18:40 +08:00
|
|
|
|
2020-10-14 07:54:07 +08:00
|
|
|
if (file) {
|
2005-04-17 06:20:36 +08:00
|
|
|
flush_dcache_mmap_lock(mapping);
|
2012-10-09 07:31:25 +08:00
|
|
|
vma_interval_tree_remove(vma, root);
|
2005-04-17 06:20:36 +08:00
|
|
|
if (adjust_next)
|
2012-10-09 07:31:25 +08:00
|
|
|
vma_interval_tree_remove(next, root);
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
|
2012-12-12 08:01:38 +08:00
|
|
|
if (start != vma->vm_start) {
|
|
|
|
|
vma->vm_start = start;
|
|
|
|
|
start_changed = true;
|
|
|
|
|
}
|
|
|
|
|
if (end != vma->vm_end) {
|
|
|
|
|
vma->vm_end = end;
|
|
|
|
|
end_changed = true;
|
|
|
|
|
}
|
2005-04-17 06:20:36 +08:00
|
|
|
vma->vm_pgoff = pgoff;
|
|
|
|
|
if (adjust_next) {
|
2020-10-14 07:53:57 +08:00
|
|
|
next->vm_start += adjust_next;
|
|
|
|
|
next->vm_pgoff += adjust_next >> PAGE_SHIFT;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
|
2020-10-14 07:54:07 +08:00
|
|
|
if (file) {
|
2005-04-17 06:20:36 +08:00
|
|
|
if (adjust_next)
|
2012-10-09 07:31:25 +08:00
|
|
|
vma_interval_tree_insert(next, root);
|
|
|
|
|
vma_interval_tree_insert(vma, root);
|
2005-04-17 06:20:36 +08:00
|
|
|
flush_dcache_mmap_unlock(mapping);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (remove_next) {
|
|
|
|
|
/*
|
|
|
|
|
* vma_merge has merged next into vma, and needs
|
|
|
|
|
* us to remove next before dropping the locks.
|
|
|
|
|
*/
|
mm: vma_merge: fix vm_page_prot SMP race condition against rmap_walk
The rmap_walk can access vm_page_prot (and potentially vm_flags in the
pte/pmd manipulations). So it's not safe to wait the caller to update
the vm_page_prot/vm_flags after vma_merge returned potentially removing
the "next" vma and extending the "current" vma over the
next->vm_start,vm_end range, but still with the "current" vma
vm_page_prot, after releasing the rmap locks.
The vm_page_prot/vm_flags must be transferred from the "next" vma to the
current vma while vma_merge still holds the rmap locks.
The side effect of this race condition is pte corruption during migrate
as remove_migration_ptes when run on a address of the "next" vma that
got removed, used the vm_page_prot of the current vma.
migrate mprotect
------------ -------------
migrating in "next" vma
vma_merge() # removes "next" vma and
# extends "current" vma
# current vma is not with
# vm_page_prot updated
remove_migration_ptes
read vm_page_prot of current "vma"
establish pte with wrong permissions
vm_set_page_prot(vma) # too late!
change_protection in the old vma range
only, next range is not updated
This caused segmentation faults and potentially memory corruption in
heavy mprotect loads with some light page migration caused by compaction
in the background.
Hugh Dickins pointed out the comment about the Odd case 8 in vma_merge
which confirms the case 8 is only buggy one where the race can trigger,
in all other vma_merge cases the above cannot happen.
This fix removes the oddness factor from case 8 and it converts it from:
AAAA
PPPPNNNNXXXX -> PPPPNNNNNNNN
to:
AAAA
PPPPNNNNXXXX -> PPPPXXXXXXXX
XXXX has the right vma properties for the whole merged vma returned by
vma_adjust, so it solves the problem fully. It has the added benefits
that the callers could stop updating vma properties when vma_merge
succeeds however the callers are not updated by this patch (there are
bits like VM_SOFTDIRTY that still need special care for the whole range,
as the vma merging ignores them, but as long as they're not processed by
rmap walks and instead they're accessed with the mmap_sem at least for
reading, they are fine not to be updated within vma_adjust before
releasing the rmap_locks).
Link: http://lkml.kernel.org/r/1474309513-20313-1-git-send-email-aarcange@redhat.com
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Aditya Mandaleeka <adityam@microsoft.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Vorlicek <janvorli@microsoft.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-10-08 08:01:28 +08:00
|
|
|
if (remove_next != 3)
|
2020-10-14 07:53:32 +08:00
|
|
|
__vma_unlink(mm, next, next);
|
mm: vma_merge: fix vm_page_prot SMP race condition against rmap_walk
The rmap_walk can access vm_page_prot (and potentially vm_flags in the
pte/pmd manipulations). So it's not safe to wait the caller to update
the vm_page_prot/vm_flags after vma_merge returned potentially removing
the "next" vma and extending the "current" vma over the
next->vm_start,vm_end range, but still with the "current" vma
vm_page_prot, after releasing the rmap locks.
The vm_page_prot/vm_flags must be transferred from the "next" vma to the
current vma while vma_merge still holds the rmap locks.
The side effect of this race condition is pte corruption during migrate
as remove_migration_ptes when run on a address of the "next" vma that
got removed, used the vm_page_prot of the current vma.
migrate mprotect
------------ -------------
migrating in "next" vma
vma_merge() # removes "next" vma and
# extends "current" vma
# current vma is not with
# vm_page_prot updated
remove_migration_ptes
read vm_page_prot of current "vma"
establish pte with wrong permissions
vm_set_page_prot(vma) # too late!
change_protection in the old vma range
only, next range is not updated
This caused segmentation faults and potentially memory corruption in
heavy mprotect loads with some light page migration caused by compaction
in the background.
Hugh Dickins pointed out the comment about the Odd case 8 in vma_merge
which confirms the case 8 is only buggy one where the race can trigger,
in all other vma_merge cases the above cannot happen.
This fix removes the oddness factor from case 8 and it converts it from:
AAAA
PPPPNNNNXXXX -> PPPPNNNNNNNN
to:
AAAA
PPPPNNNNXXXX -> PPPPXXXXXXXX
XXXX has the right vma properties for the whole merged vma returned by
vma_adjust, so it solves the problem fully. It has the added benefits
that the callers could stop updating vma properties when vma_merge
succeeds however the callers are not updated by this patch (there are
bits like VM_SOFTDIRTY that still need special care for the whole range,
as the vma merging ignores them, but as long as they're not processed by
rmap walks and instead they're accessed with the mmap_sem at least for
reading, they are fine not to be updated within vma_adjust before
releasing the rmap_locks).
Link: http://lkml.kernel.org/r/1474309513-20313-1-git-send-email-aarcange@redhat.com
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Aditya Mandaleeka <adityam@microsoft.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Vorlicek <janvorli@microsoft.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-10-08 08:01:28 +08:00
|
|
|
else
|
2016-10-08 08:01:37 +08:00
|
|
|
/*
|
|
|
|
|
* vma is not before next if they've been
|
|
|
|
|
* swapped.
|
|
|
|
|
*
|
|
|
|
|
* pre-swap() next->vm_start was reduced so
|
|
|
|
|
* tell validate_mm_rb to ignore pre-swap()
|
|
|
|
|
* "next" (which is stored in post-swap()
|
|
|
|
|
* "vma").
|
|
|
|
|
*/
|
2020-10-14 07:53:32 +08:00
|
|
|
__vma_unlink(mm, next, vma);
|
2005-04-17 06:20:36 +08:00
|
|
|
if (file)
|
|
|
|
|
__remove_shared_vm_struct(next, file, mapping);
|
|
|
|
|
} else if (insert) {
|
|
|
|
|
/*
|
|
|
|
|
* split_vma has split insert from vma, and needs
|
|
|
|
|
* us to insert it before dropping the locks
|
|
|
|
|
* (it may either follow vma or precede it).
|
|
|
|
|
*/
|
|
|
|
|
__insert_vm_struct(mm, insert);
|
2012-12-12 08:01:38 +08:00
|
|
|
} else {
|
|
|
|
|
if (start_changed)
|
|
|
|
|
vma_gap_update(vma);
|
|
|
|
|
if (end_changed) {
|
|
|
|
|
if (!next)
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
mm->highest_vm_end = vm_end_gap(vma);
|
2012-12-12 08:01:38 +08:00
|
|
|
else if (!adjust_next)
|
|
|
|
|
vma_gap_update(next);
|
|
|
|
|
}
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
|
mm anon rmap: replace same_anon_vma linked list with an interval tree.
When a large VMA (anon or private file mapping) is first touched, which
will populate its anon_vma field, and then split into many regions through
the use of mprotect(), the original anon_vma ends up linking all of the
vmas on a linked list. This can cause rmap to become inefficient, as we
have to walk potentially thousands of irrelevent vmas before finding the
one a given anon page might fall into.
By replacing the same_anon_vma linked list with an interval tree (where
each avc's interval is determined by its vma's start and last pgoffs), we
can make rmap efficient for this use case again.
While the change is large, all of its pieces are fairly simple.
Most places that were walking the same_anon_vma list were looking for a
known pgoff, so they can just use the anon_vma_interval_tree_foreach()
interval tree iterator instead. The exception here is ksm, where the
page's index is not known. It would probably be possible to rework ksm so
that the index would be known, but for now I have decided to keep things
simple and just walk the entirety of the interval tree there.
When updating vma's that already have an anon_vma assigned, we must take
care to re-index the corresponding avc's on their interval tree. This is
done through the use of anon_vma_interval_tree_pre_update_vma() and
anon_vma_interval_tree_post_update_vma(), which remove the avc's from
their interval tree before the update and re-insert them after the update.
The anon_vma stays locked during the update, so there is no chance that
rmap would miss the vmas that are being updated.
Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-10-09 07:31:39 +08:00
|
|
|
if (anon_vma) {
|
|
|
|
|
anon_vma_interval_tree_post_update_vma(vma);
|
|
|
|
|
if (adjust_next)
|
|
|
|
|
anon_vma_interval_tree_post_update_vma(next);
|
2013-02-23 08:34:40 +08:00
|
|
|
anon_vma_unlock_write(anon_vma);
|
mm anon rmap: replace same_anon_vma linked list with an interval tree.
When a large VMA (anon or private file mapping) is first touched, which
will populate its anon_vma field, and then split into many regions through
the use of mprotect(), the original anon_vma ends up linking all of the
vmas on a linked list. This can cause rmap to become inefficient, as we
have to walk potentially thousands of irrelevent vmas before finding the
one a given anon page might fall into.
By replacing the same_anon_vma linked list with an interval tree (where
each avc's interval is determined by its vma's start and last pgoffs), we
can make rmap efficient for this use case again.
While the change is large, all of its pieces are fairly simple.
Most places that were walking the same_anon_vma list were looking for a
known pgoff, so they can just use the anon_vma_interval_tree_foreach()
interval tree iterator instead. The exception here is ksm, where the
page's index is not known. It would probably be possible to rework ksm so
that the index would be known, but for now I have decided to keep things
simple and just walk the entirety of the interval tree there.
When updating vma's that already have an anon_vma assigned, we must take
care to re-index the corresponding avc's on their interval tree. This is
done through the use of anon_vma_interval_tree_pre_update_vma() and
anon_vma_interval_tree_post_update_vma(), which remove the avc's from
their interval tree before the update and re-insert them after the update.
The anon_vma stays locked during the update, so there is no chance that
rmap would miss the vmas that are being updated.
Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-10-09 07:31:39 +08:00
|
|
|
}
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2020-10-14 07:54:07 +08:00
|
|
|
if (file) {
|
2020-10-14 07:54:04 +08:00
|
|
|
i_mmap_unlock_write(mapping);
|
2012-02-17 16:27:41 +08:00
|
|
|
uprobe_mmap(vma);
|
uprobes, mm, x86: Add the ability to install and remove uprobes breakpoints
Add uprobes support to the core kernel, with x86 support.
This commit adds the kernel facilities, the actual uprobes
user-space ABI and perf probe support comes in later commits.
General design:
Uprobes are maintained in an rb-tree indexed by inode and offset
(the offset here is from the start of the mapping). For a unique
(inode, offset) tuple, there can be at most one uprobe in the
rb-tree.
Since the (inode, offset) tuple identifies a unique uprobe, more
than one user may be interested in the same uprobe. This provides
the ability to connect multiple 'consumers' to the same uprobe.
Each consumer defines a handler and a filter (optional). The
'handler' is run every time the uprobe is hit, if it matches the
'filter' criteria.
The first consumer of a uprobe causes the breakpoint to be
inserted at the specified address and subsequent consumers are
appended to this list. On subsequent probes, the consumer gets
appended to the existing list of consumers. The breakpoint is
removed when the last consumer unregisters. For all other
unregisterations, the consumer is removed from the list of
consumers.
Given a inode, we get a list of the mms that have mapped the
inode. Do the actual registration if mm maps the page where a
probe needs to be inserted/removed.
We use a temporary list to walk through the vmas that map the
inode.
- The number of maps that map the inode, is not known before we
walk the rmap and keeps changing.
- extending vm_area_struct wasn't recommended, it's a
size-critical data structure.
- There can be more than one maps of the inode in the same mm.
We add callbacks to the mmap methods to keep an eye on text vmas
that are of interest to uprobes. When a vma of interest is mapped,
we insert the breakpoint at the right address.
Uprobe works by replacing the instruction at the address defined
by (inode, offset) with the arch specific breakpoint
instruction. We save a copy of the original instruction at the
uprobed address.
This is needed for:
a. executing the instruction out-of-line (xol).
b. instruction analysis for any subsequent fixups.
c. restoring the instruction back when the uprobe is unregistered.
We insert or delete a breakpoint instruction, and this
breakpoint instruction is assumed to be the smallest instruction
available on the platform. For fixed size instruction platforms
this is trivially true, for variable size instruction platforms
the breakpoint instruction is typically the smallest (often a
single byte).
Writing the instruction is done by COWing the page and changing
the instruction during the copy, this even though most platforms
allow atomic writes of the breakpoint instruction. This also
mirrors the behaviour of a ptrace() memory write to a PRIVATE
file map.
The core worker is derived from KSM's replace_page() logic.
In essence, similar to KSM:
a. allocate a new page and copy over contents of the page that
has the uprobed vaddr
b. modify the copy and insert the breakpoint at the required
address
c. switch the original page with the copy containing the
breakpoint
d. flush page tables.
replace_page() is being replicated here because of some minor
changes in the type of pages and also because Hugh Dickins had
plans to improve replace_page() for KSM specific work.
Instruction analysis on x86 is based on instruction decoder and
determines if an instruction can be probed and determines the
necessary fixups after singlestep. Instruction analysis is done
at probe insertion time so that we avoid having to repeat the
same analysis every time a probe is hit.
A lot of code here is due to the improvement/suggestions/inputs
from Peter Zijlstra.
Changelog:
(v10):
- Add code to clear REX.B prefix as suggested by Denys Vlasenko
and Masami Hiramatsu.
(v9):
- Use insn_offset_modrm as suggested by Masami Hiramatsu.
(v7):
Handle comments from Peter Zijlstra:
- Dont take reference to inode. (expect inode to uprobe_register to be sane).
- Use PTR_ERR to set the return value.
- No need to take reference to inode.
- use PTR_ERR to return error value.
- register and uprobe_unregister share code.
(v5):
- Modified del_consumer as per comments from Peter.
- Drop reference to inode before dropping reference to uprobe.
- Use i_size_read(inode) instead of inode->i_size.
- Ensure uprobe->consumers is NULL, before __uprobe_unregister() is called.
- Includes errno.h as recommended by Stephen Rothwell to fix a build issue
on sparc defconfig
- Remove restrictions while unregistering.
- Earlier code leaked inode references under some conditions while
registering/unregistering.
- Continue the vma-rmap walk even if the intermediate vma doesnt
meet the requirements.
- Validate the vma found by find_vma before inserting/removing the
breakpoint
- Call del_consumer under mutex_lock.
- Use hash locks.
- Handle mremap.
- Introduce find_least_offset_node() instead of close match logic in
find_uprobe
- Uprobes no more depends on MM_OWNER; No reference to task_structs
while inserting/removing a probe.
- Uses read_mapping_page instead of grab_cache_page so that the pages
have valid content.
- pass NULL to get_user_pages for the task parameter.
- call SetPageUptodate on the new page allocated in write_opcode.
- fix leaking a reference to the new page under certain conditions.
- Include Instruction Decoder if Uprobes gets defined.
- Remove const attributes for instruction prefix arrays.
- Uses mm_context to know if the application is 32 bit.
Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Also-written-by: Jim Keniston <jkenisto@us.ibm.com>
Reviewed-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Roland McGrath <roland@hack.frob.com>
Cc: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
Cc: Anton Arapov <anton@redhat.com>
Cc: Ananth N Mavinakayanahalli <ananth@in.ibm.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Denys Vlasenko <vda.linux@googlemail.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Linux-mm <linux-mm@kvack.org>
Link: http://lkml.kernel.org/r/20120209092642.GE16600@linux.vnet.ibm.com
[ Made various small edits to the commit log ]
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2012-02-09 17:26:42 +08:00
|
|
|
|
|
|
|
|
if (adjust_next)
|
2012-02-17 16:27:41 +08:00
|
|
|
uprobe_mmap(next);
|
uprobes, mm, x86: Add the ability to install and remove uprobes breakpoints
Add uprobes support to the core kernel, with x86 support.
This commit adds the kernel facilities, the actual uprobes
user-space ABI and perf probe support comes in later commits.
General design:
Uprobes are maintained in an rb-tree indexed by inode and offset
(the offset here is from the start of the mapping). For a unique
(inode, offset) tuple, there can be at most one uprobe in the
rb-tree.
Since the (inode, offset) tuple identifies a unique uprobe, more
than one user may be interested in the same uprobe. This provides
the ability to connect multiple 'consumers' to the same uprobe.
Each consumer defines a handler and a filter (optional). The
'handler' is run every time the uprobe is hit, if it matches the
'filter' criteria.
The first consumer of a uprobe causes the breakpoint to be
inserted at the specified address and subsequent consumers are
appended to this list. On subsequent probes, the consumer gets
appended to the existing list of consumers. The breakpoint is
removed when the last consumer unregisters. For all other
unregisterations, the consumer is removed from the list of
consumers.
Given a inode, we get a list of the mms that have mapped the
inode. Do the actual registration if mm maps the page where a
probe needs to be inserted/removed.
We use a temporary list to walk through the vmas that map the
inode.
- The number of maps that map the inode, is not known before we
walk the rmap and keeps changing.
- extending vm_area_struct wasn't recommended, it's a
size-critical data structure.
- There can be more than one maps of the inode in the same mm.
We add callbacks to the mmap methods to keep an eye on text vmas
that are of interest to uprobes. When a vma of interest is mapped,
we insert the breakpoint at the right address.
Uprobe works by replacing the instruction at the address defined
by (inode, offset) with the arch specific breakpoint
instruction. We save a copy of the original instruction at the
uprobed address.
This is needed for:
a. executing the instruction out-of-line (xol).
b. instruction analysis for any subsequent fixups.
c. restoring the instruction back when the uprobe is unregistered.
We insert or delete a breakpoint instruction, and this
breakpoint instruction is assumed to be the smallest instruction
available on the platform. For fixed size instruction platforms
this is trivially true, for variable size instruction platforms
the breakpoint instruction is typically the smallest (often a
single byte).
Writing the instruction is done by COWing the page and changing
the instruction during the copy, this even though most platforms
allow atomic writes of the breakpoint instruction. This also
mirrors the behaviour of a ptrace() memory write to a PRIVATE
file map.
The core worker is derived from KSM's replace_page() logic.
In essence, similar to KSM:
a. allocate a new page and copy over contents of the page that
has the uprobed vaddr
b. modify the copy and insert the breakpoint at the required
address
c. switch the original page with the copy containing the
breakpoint
d. flush page tables.
replace_page() is being replicated here because of some minor
changes in the type of pages and also because Hugh Dickins had
plans to improve replace_page() for KSM specific work.
Instruction analysis on x86 is based on instruction decoder and
determines if an instruction can be probed and determines the
necessary fixups after singlestep. Instruction analysis is done
at probe insertion time so that we avoid having to repeat the
same analysis every time a probe is hit.
A lot of code here is due to the improvement/suggestions/inputs
from Peter Zijlstra.
Changelog:
(v10):
- Add code to clear REX.B prefix as suggested by Denys Vlasenko
and Masami Hiramatsu.
(v9):
- Use insn_offset_modrm as suggested by Masami Hiramatsu.
(v7):
Handle comments from Peter Zijlstra:
- Dont take reference to inode. (expect inode to uprobe_register to be sane).
- Use PTR_ERR to set the return value.
- No need to take reference to inode.
- use PTR_ERR to return error value.
- register and uprobe_unregister share code.
(v5):
- Modified del_consumer as per comments from Peter.
- Drop reference to inode before dropping reference to uprobe.
- Use i_size_read(inode) instead of inode->i_size.
- Ensure uprobe->consumers is NULL, before __uprobe_unregister() is called.
- Includes errno.h as recommended by Stephen Rothwell to fix a build issue
on sparc defconfig
- Remove restrictions while unregistering.
- Earlier code leaked inode references under some conditions while
registering/unregistering.
- Continue the vma-rmap walk even if the intermediate vma doesnt
meet the requirements.
- Validate the vma found by find_vma before inserting/removing the
breakpoint
- Call del_consumer under mutex_lock.
- Use hash locks.
- Handle mremap.
- Introduce find_least_offset_node() instead of close match logic in
find_uprobe
- Uprobes no more depends on MM_OWNER; No reference to task_structs
while inserting/removing a probe.
- Uses read_mapping_page instead of grab_cache_page so that the pages
have valid content.
- pass NULL to get_user_pages for the task parameter.
- call SetPageUptodate on the new page allocated in write_opcode.
- fix leaking a reference to the new page under certain conditions.
- Include Instruction Decoder if Uprobes gets defined.
- Remove const attributes for instruction prefix arrays.
- Uses mm_context to know if the application is 32 bit.
Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Also-written-by: Jim Keniston <jkenisto@us.ibm.com>
Reviewed-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Roland McGrath <roland@hack.frob.com>
Cc: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
Cc: Anton Arapov <anton@redhat.com>
Cc: Ananth N Mavinakayanahalli <ananth@in.ibm.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Denys Vlasenko <vda.linux@googlemail.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Linux-mm <linux-mm@kvack.org>
Link: http://lkml.kernel.org/r/20120209092642.GE16600@linux.vnet.ibm.com
[ Made various small edits to the commit log ]
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2012-02-09 17:26:42 +08:00
|
|
|
}
|
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
if (remove_next) {
|
2008-04-29 16:01:36 +08:00
|
|
|
if (file) {
|
2012-04-11 18:35:27 +08:00
|
|
|
uprobe_munmap(next, next->vm_start, next->vm_end);
|
2005-04-17 06:20:36 +08:00
|
|
|
fput(file);
|
2008-04-29 16:01:36 +08:00
|
|
|
}
|
mm: change anon_vma linking to fix multi-process server scalability issue
The old anon_vma code can lead to scalability issues with heavily forking
workloads. Specifically, each anon_vma will be shared between the parent
process and all its child processes.
In a workload with 1000 child processes and a VMA with 1000 anonymous
pages per process that get COWed, this leads to a system with a million
anonymous pages in the same anon_vma, each of which is mapped in just one
of the 1000 processes. However, the current rmap code needs to walk them
all, leading to O(N) scanning complexity for each page.
This can result in systems where one CPU is walking the page tables of
1000 processes in page_referenced_one, while all other CPUs are stuck on
the anon_vma lock. This leads to catastrophic failure for a benchmark
like AIM7, where the total number of processes can reach in the tens of
thousands. Real workloads are still a factor 10 less process intensive
than AIM7, but they are catching up.
This patch changes the way anon_vmas and VMAs are linked, which allows us
to associate multiple anon_vmas with a VMA. At fork time, each child
process gets its own anon_vmas, in which its COWed pages will be
instantiated. The parents' anon_vma is also linked to the VMA, because
non-COWed pages could be present in any of the children.
This reduces rmap scanning complexity to O(1) for the pages of the 1000
child processes, with O(N) complexity for at most 1/N pages in the system.
This reduces the average scanning cost in heavily forking workloads from
O(N) to 2.
The only real complexity in this patch stems from the fact that linking a
VMA to anon_vmas now involves memory allocations. This means vma_adjust
can fail, if it needs to attach a VMA to anon_vma structures. This in
turn means error handling needs to be added to the calling functions.
A second source of complexity is that, because there can be multiple
anon_vmas, the anon_vma linking in vma_adjust can no longer be done under
"the" anon_vma lock. To prevent the rmap code from walking up an
incomplete VMA, this patch introduces the VM_LOCK_RMAP VMA flag. This bit
flag uses the same slot as the NOMMU VM_MAPPED_COPY, with an ifdef in mm.h
to make sure it is impossible to compile a kernel that needs both symbolic
values for the same bitflag.
Some test results:
Without the anon_vma changes, when AIM7 hits around 9.7k users (on a test
box with 16GB RAM and not quite enough IO), the system ends up running
>99% in system time, with every CPU on the same anon_vma lock in the
pageout code.
With these changes, AIM7 hits the cross-over point around 29.7k users.
This happens with ~99% IO wait time, there never seems to be any spike in
system time. The anon_vma lock contention appears to be resolved.
[akpm@linux-foundation.org: cleanups]
Signed-off-by: Rik van Riel <riel@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Larry Woodman <lwoodman@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-06 05:42:07 +08:00
|
|
|
if (next->anon_vma)
|
|
|
|
|
anon_vma_merge(vma, next);
|
2005-04-17 06:20:36 +08:00
|
|
|
mm->map_count--;
|
2013-08-01 04:53:28 +08:00
|
|
|
mpol_put(vma_policy(next));
|
2018-07-22 04:48:51 +08:00
|
|
|
vm_area_free(next);
|
2005-04-17 06:20:36 +08:00
|
|
|
/*
|
|
|
|
|
* In mprotect's case 6 (see comments on vma_merge),
|
|
|
|
|
* we must remove another next too. It would clutter
|
|
|
|
|
* up the code too much to do both in one go.
|
|
|
|
|
*/
|
mm: vma_merge: fix vm_page_prot SMP race condition against rmap_walk
The rmap_walk can access vm_page_prot (and potentially vm_flags in the
pte/pmd manipulations). So it's not safe to wait the caller to update
the vm_page_prot/vm_flags after vma_merge returned potentially removing
the "next" vma and extending the "current" vma over the
next->vm_start,vm_end range, but still with the "current" vma
vm_page_prot, after releasing the rmap locks.
The vm_page_prot/vm_flags must be transferred from the "next" vma to the
current vma while vma_merge still holds the rmap locks.
The side effect of this race condition is pte corruption during migrate
as remove_migration_ptes when run on a address of the "next" vma that
got removed, used the vm_page_prot of the current vma.
migrate mprotect
------------ -------------
migrating in "next" vma
vma_merge() # removes "next" vma and
# extends "current" vma
# current vma is not with
# vm_page_prot updated
remove_migration_ptes
read vm_page_prot of current "vma"
establish pte with wrong permissions
vm_set_page_prot(vma) # too late!
change_protection in the old vma range
only, next range is not updated
This caused segmentation faults and potentially memory corruption in
heavy mprotect loads with some light page migration caused by compaction
in the background.
Hugh Dickins pointed out the comment about the Odd case 8 in vma_merge
which confirms the case 8 is only buggy one where the race can trigger,
in all other vma_merge cases the above cannot happen.
This fix removes the oddness factor from case 8 and it converts it from:
AAAA
PPPPNNNNXXXX -> PPPPNNNNNNNN
to:
AAAA
PPPPNNNNXXXX -> PPPPXXXXXXXX
XXXX has the right vma properties for the whole merged vma returned by
vma_adjust, so it solves the problem fully. It has the added benefits
that the callers could stop updating vma properties when vma_merge
succeeds however the callers are not updated by this patch (there are
bits like VM_SOFTDIRTY that still need special care for the whole range,
as the vma merging ignores them, but as long as they're not processed by
rmap walks and instead they're accessed with the mmap_sem at least for
reading, they are fine not to be updated within vma_adjust before
releasing the rmap_locks).
Link: http://lkml.kernel.org/r/1474309513-20313-1-git-send-email-aarcange@redhat.com
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Aditya Mandaleeka <adityam@microsoft.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Vorlicek <janvorli@microsoft.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-10-08 08:01:28 +08:00
|
|
|
if (remove_next != 3) {
|
|
|
|
|
/*
|
|
|
|
|
* If "next" was removed and vma->vm_end was
|
|
|
|
|
* expanded (up) over it, in turn
|
|
|
|
|
* "next->vm_prev->vm_end" changed and the
|
|
|
|
|
* "vma->vm_next" gap must be updated.
|
|
|
|
|
*/
|
|
|
|
|
next = vma->vm_next;
|
|
|
|
|
} else {
|
|
|
|
|
/*
|
|
|
|
|
* For the scope of the comment "next" and
|
|
|
|
|
* "vma" considered pre-swap(): if "vma" was
|
|
|
|
|
* removed, next->vm_start was expanded (down)
|
|
|
|
|
* over it and the "next" gap must be updated.
|
|
|
|
|
* Because of the swap() the post-swap() "vma"
|
|
|
|
|
* actually points to pre-swap() "next"
|
|
|
|
|
* (post-swap() "next" as opposed is now a
|
|
|
|
|
* dangling pointer).
|
|
|
|
|
*/
|
|
|
|
|
next = vma;
|
|
|
|
|
}
|
2016-07-29 06:49:01 +08:00
|
|
|
if (remove_next == 2) {
|
|
|
|
|
remove_next = 1;
|
|
|
|
|
end = next->vm_end;
|
2005-04-17 06:20:36 +08:00
|
|
|
goto again;
|
2016-07-29 06:49:01 +08:00
|
|
|
}
|
2012-12-12 08:01:38 +08:00
|
|
|
else if (next)
|
|
|
|
|
vma_gap_update(next);
|
2016-10-08 08:01:25 +08:00
|
|
|
else {
|
|
|
|
|
/*
|
|
|
|
|
* If remove_next == 2 we obviously can't
|
|
|
|
|
* reach this path.
|
|
|
|
|
*
|
|
|
|
|
* If remove_next == 3 we can't reach this
|
|
|
|
|
* path because pre-swap() next is always not
|
|
|
|
|
* NULL. pre-swap() "next" is not being
|
|
|
|
|
* removed and its next->vm_end is not altered
|
|
|
|
|
* (and furthermore "end" already matches
|
|
|
|
|
* next->vm_end in remove_next == 3).
|
|
|
|
|
*
|
|
|
|
|
* We reach this only in the remove_next == 1
|
|
|
|
|
* case if the "next" vma that was removed was
|
|
|
|
|
* the highest vma of the mm. However in such
|
|
|
|
|
* case next->vm_end == "end" and the extended
|
|
|
|
|
* "vma" has vma->vm_end == next->vm_end so
|
|
|
|
|
* mm->highest_vm_end doesn't need any update
|
|
|
|
|
* in remove_next == 1 case.
|
|
|
|
|
*/
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
|
2016-10-08 08:01:25 +08:00
|
|
|
}
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
uprobes, mm, x86: Add the ability to install and remove uprobes breakpoints
Add uprobes support to the core kernel, with x86 support.
This commit adds the kernel facilities, the actual uprobes
user-space ABI and perf probe support comes in later commits.
General design:
Uprobes are maintained in an rb-tree indexed by inode and offset
(the offset here is from the start of the mapping). For a unique
(inode, offset) tuple, there can be at most one uprobe in the
rb-tree.
Since the (inode, offset) tuple identifies a unique uprobe, more
than one user may be interested in the same uprobe. This provides
the ability to connect multiple 'consumers' to the same uprobe.
Each consumer defines a handler and a filter (optional). The
'handler' is run every time the uprobe is hit, if it matches the
'filter' criteria.
The first consumer of a uprobe causes the breakpoint to be
inserted at the specified address and subsequent consumers are
appended to this list. On subsequent probes, the consumer gets
appended to the existing list of consumers. The breakpoint is
removed when the last consumer unregisters. For all other
unregisterations, the consumer is removed from the list of
consumers.
Given a inode, we get a list of the mms that have mapped the
inode. Do the actual registration if mm maps the page where a
probe needs to be inserted/removed.
We use a temporary list to walk through the vmas that map the
inode.
- The number of maps that map the inode, is not known before we
walk the rmap and keeps changing.
- extending vm_area_struct wasn't recommended, it's a
size-critical data structure.
- There can be more than one maps of the inode in the same mm.
We add callbacks to the mmap methods to keep an eye on text vmas
that are of interest to uprobes. When a vma of interest is mapped,
we insert the breakpoint at the right address.
Uprobe works by replacing the instruction at the address defined
by (inode, offset) with the arch specific breakpoint
instruction. We save a copy of the original instruction at the
uprobed address.
This is needed for:
a. executing the instruction out-of-line (xol).
b. instruction analysis for any subsequent fixups.
c. restoring the instruction back when the uprobe is unregistered.
We insert or delete a breakpoint instruction, and this
breakpoint instruction is assumed to be the smallest instruction
available on the platform. For fixed size instruction platforms
this is trivially true, for variable size instruction platforms
the breakpoint instruction is typically the smallest (often a
single byte).
Writing the instruction is done by COWing the page and changing
the instruction during the copy, this even though most platforms
allow atomic writes of the breakpoint instruction. This also
mirrors the behaviour of a ptrace() memory write to a PRIVATE
file map.
The core worker is derived from KSM's replace_page() logic.
In essence, similar to KSM:
a. allocate a new page and copy over contents of the page that
has the uprobed vaddr
b. modify the copy and insert the breakpoint at the required
address
c. switch the original page with the copy containing the
breakpoint
d. flush page tables.
replace_page() is being replicated here because of some minor
changes in the type of pages and also because Hugh Dickins had
plans to improve replace_page() for KSM specific work.
Instruction analysis on x86 is based on instruction decoder and
determines if an instruction can be probed and determines the
necessary fixups after singlestep. Instruction analysis is done
at probe insertion time so that we avoid having to repeat the
same analysis every time a probe is hit.
A lot of code here is due to the improvement/suggestions/inputs
from Peter Zijlstra.
Changelog:
(v10):
- Add code to clear REX.B prefix as suggested by Denys Vlasenko
and Masami Hiramatsu.
(v9):
- Use insn_offset_modrm as suggested by Masami Hiramatsu.
(v7):
Handle comments from Peter Zijlstra:
- Dont take reference to inode. (expect inode to uprobe_register to be sane).
- Use PTR_ERR to set the return value.
- No need to take reference to inode.
- use PTR_ERR to return error value.
- register and uprobe_unregister share code.
(v5):
- Modified del_consumer as per comments from Peter.
- Drop reference to inode before dropping reference to uprobe.
- Use i_size_read(inode) instead of inode->i_size.
- Ensure uprobe->consumers is NULL, before __uprobe_unregister() is called.
- Includes errno.h as recommended by Stephen Rothwell to fix a build issue
on sparc defconfig
- Remove restrictions while unregistering.
- Earlier code leaked inode references under some conditions while
registering/unregistering.
- Continue the vma-rmap walk even if the intermediate vma doesnt
meet the requirements.
- Validate the vma found by find_vma before inserting/removing the
breakpoint
- Call del_consumer under mutex_lock.
- Use hash locks.
- Handle mremap.
- Introduce find_least_offset_node() instead of close match logic in
find_uprobe
- Uprobes no more depends on MM_OWNER; No reference to task_structs
while inserting/removing a probe.
- Uses read_mapping_page instead of grab_cache_page so that the pages
have valid content.
- pass NULL to get_user_pages for the task parameter.
- call SetPageUptodate on the new page allocated in write_opcode.
- fix leaking a reference to the new page under certain conditions.
- Include Instruction Decoder if Uprobes gets defined.
- Remove const attributes for instruction prefix arrays.
- Uses mm_context to know if the application is 32 bit.
Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Also-written-by: Jim Keniston <jkenisto@us.ibm.com>
Reviewed-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Roland McGrath <roland@hack.frob.com>
Cc: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
Cc: Anton Arapov <anton@redhat.com>
Cc: Ananth N Mavinakayanahalli <ananth@in.ibm.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Denys Vlasenko <vda.linux@googlemail.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Linux-mm <linux-mm@kvack.org>
Link: http://lkml.kernel.org/r/20120209092642.GE16600@linux.vnet.ibm.com
[ Made various small edits to the commit log ]
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2012-02-09 17:26:42 +08:00
|
|
|
if (insert && file)
|
2012-02-17 16:27:41 +08:00
|
|
|
uprobe_mmap(insert);
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
|
validate_mm(mm);
|
mm: change anon_vma linking to fix multi-process server scalability issue
The old anon_vma code can lead to scalability issues with heavily forking
workloads. Specifically, each anon_vma will be shared between the parent
process and all its child processes.
In a workload with 1000 child processes and a VMA with 1000 anonymous
pages per process that get COWed, this leads to a system with a million
anonymous pages in the same anon_vma, each of which is mapped in just one
of the 1000 processes. However, the current rmap code needs to walk them
all, leading to O(N) scanning complexity for each page.
This can result in systems where one CPU is walking the page tables of
1000 processes in page_referenced_one, while all other CPUs are stuck on
the anon_vma lock. This leads to catastrophic failure for a benchmark
like AIM7, where the total number of processes can reach in the tens of
thousands. Real workloads are still a factor 10 less process intensive
than AIM7, but they are catching up.
This patch changes the way anon_vmas and VMAs are linked, which allows us
to associate multiple anon_vmas with a VMA. At fork time, each child
process gets its own anon_vmas, in which its COWed pages will be
instantiated. The parents' anon_vma is also linked to the VMA, because
non-COWed pages could be present in any of the children.
This reduces rmap scanning complexity to O(1) for the pages of the 1000
child processes, with O(N) complexity for at most 1/N pages in the system.
This reduces the average scanning cost in heavily forking workloads from
O(N) to 2.
The only real complexity in this patch stems from the fact that linking a
VMA to anon_vmas now involves memory allocations. This means vma_adjust
can fail, if it needs to attach a VMA to anon_vma structures. This in
turn means error handling needs to be added to the calling functions.
A second source of complexity is that, because there can be multiple
anon_vmas, the anon_vma linking in vma_adjust can no longer be done under
"the" anon_vma lock. To prevent the rmap code from walking up an
incomplete VMA, this patch introduces the VM_LOCK_RMAP VMA flag. This bit
flag uses the same slot as the NOMMU VM_MAPPED_COPY, with an ifdef in mm.h
to make sure it is impossible to compile a kernel that needs both symbolic
values for the same bitflag.
Some test results:
Without the anon_vma changes, when AIM7 hits around 9.7k users (on a test
box with 16GB RAM and not quite enough IO), the system ends up running
>99% in system time, with every CPU on the same anon_vma lock in the
pageout code.
With these changes, AIM7 hits the cross-over point around 29.7k users.
This happens with ~99% IO wait time, there never seems to be any spike in
system time. The anon_vma lock contention appears to be resolved.
[akpm@linux-foundation.org: cleanups]
Signed-off-by: Rik van Riel <riel@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Larry Woodman <lwoodman@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-06 05:42:07 +08:00
|
|
|
|
|
|
|
|
return 0;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* If the vma has a ->close operation then the driver probably needs to release
|
|
|
|
|
* per-vma resources, so we don't attempt to merge those.
|
|
|
|
|
*/
|
|
|
|
|
static inline int is_mergeable_vma(struct vm_area_struct *vma,
|
2015-09-05 06:46:24 +08:00
|
|
|
struct file *file, unsigned long vm_flags,
|
mm: add a field to store names for private anonymous memory
In many userspace applications, and especially in VM based applications
like Android uses heavily, there are multiple different allocators in
use. At a minimum there is libc malloc and the stack, and in many cases
there are libc malloc, the stack, direct syscalls to mmap anonymous
memory, and multiple VM heaps (one for small objects, one for big
objects, etc.). Each of these layers usually has its own tools to
inspect its usage; malloc by compiling a debug version, the VM through
heap inspection tools, and for direct syscalls there is usually no way
to track them.
On Android we heavily use a set of tools that use an extended version of
the logic covered in Documentation/vm/pagemap.txt to walk all pages
mapped in userspace and slice their usage by process, shared (COW) vs.
unique mappings, backing, etc. This can account for real physical
memory usage even in cases like fork without exec (which Android uses
heavily to share as many private COW pages as possible between
processes), Kernel SamePage Merging, and clean zero pages. It produces
a measurement of the pages that only exist in that process (USS, for
unique), and a measurement of the physical memory usage of that process
with the cost of shared pages being evenly split between processes that
share them (PSS).
If all anonymous memory is indistinguishable then figuring out the real
physical memory usage (PSS) of each heap requires either a pagemap
walking tool that can understand the heap debugging of every layer, or
for every layer's heap debugging tools to implement the pagemap walking
logic, in which case it is hard to get a consistent view of memory
across the whole system.
Tracking the information in userspace leads to all sorts of problems.
It either needs to be stored inside the process, which means every
process has to have an API to export its current heap information upon
request, or it has to be stored externally in a filesystem that somebody
needs to clean up on crashes. It needs to be readable while the process
is still running, so it has to have some sort of synchronization with
every layer of userspace. Efficiently tracking the ranges requires
reimplementing something like the kernel vma trees, and linking to it
from every layer of userspace. It requires more memory, more syscalls,
more runtime cost, and more complexity to separately track regions that
the kernel is already tracking.
This patch adds a field to /proc/pid/maps and /proc/pid/smaps to show a
userspace-provided name for anonymous vmas. The names of named
anonymous vmas are shown in /proc/pid/maps and /proc/pid/smaps as
[anon:<name>].
Userspace can set the name for a region of memory by calling
prctl(PR_SET_VMA, PR_SET_VMA_ANON_NAME, start, len, (unsigned long)name)
Setting the name to NULL clears it. The name length limit is 80 bytes
including NUL-terminator and is checked to contain only printable ascii
characters (including space), except '[',']','\','$' and '`'.
Ascii strings are being used to have a descriptive identifiers for vmas,
which can be understood by the users reading /proc/pid/maps or
/proc/pid/smaps. Names can be standardized for a given system and they
can include some variable parts such as the name of the allocator or a
library, tid of the thread using it, etc.
The name is stored in a pointer in the shared union in vm_area_struct
that points to a null terminated string. Anonymous vmas with the same
name (equivalent strings) and are otherwise mergeable will be merged.
The name pointers are not shared between vmas even if they contain the
same name. The name pointer is stored in a union with fields that are
only used on file-backed mappings, so it does not increase memory usage.
CONFIG_ANON_VMA_NAME kernel configuration is introduced to enable this
feature. It keeps the feature disabled by default to prevent any
additional memory overhead and to avoid confusing procfs parsers on
systems which are not ready to support named anonymous vmas.
The patch is based on the original patch developed by Colin Cross, more
specifically on its latest version [1] posted upstream by Sumit Semwal.
It used a userspace pointer to store vma names. In that design, name
pointers could be shared between vmas. However during the last
upstreaming attempt, Kees Cook raised concerns [2] about this approach
and suggested to copy the name into kernel memory space, perform
validity checks [3] and store as a string referenced from
vm_area_struct.
One big concern is about fork() performance which would need to strdup
anonymous vma names. Dave Hansen suggested experimenting with
worst-case scenario of forking a process with 64k vmas having longest
possible names [4]. I ran this experiment on an ARM64 Android device
and recorded a worst-case regression of almost 40% when forking such a
process.
This regression is addressed in the followup patch which replaces the
pointer to a name with a refcounted structure that allows sharing the
name pointer between vmas of the same name. Instead of duplicating the
string during fork() or when splitting a vma it increments the refcount.
[1] https://lore.kernel.org/linux-mm/20200901161459.11772-4-sumit.semwal@linaro.org/
[2] https://lore.kernel.org/linux-mm/202009031031.D32EF57ED@keescook/
[3] https://lore.kernel.org/linux-mm/202009031022.3834F692@keescook/
[4] https://lore.kernel.org/linux-mm/5d0358ab-8c47-2f5f-8e43-23b89d6a8e95@intel.com/
Changes for prctl(2) manual page (in the options section):
PR_SET_VMA
Sets an attribute specified in arg2 for virtual memory areas
starting from the address specified in arg3 and spanning the
size specified in arg4. arg5 specifies the value of the attribute
to be set. Note that assigning an attribute to a virtual memory
area might prevent it from being merged with adjacent virtual
memory areas due to the difference in that attribute's value.
Currently, arg2 must be one of:
PR_SET_VMA_ANON_NAME
Set a name for anonymous virtual memory areas. arg5 should
be a pointer to a null-terminated string containing the
name. The name length including null byte cannot exceed
80 bytes. If arg5 is NULL, the name of the appropriate
anonymous virtual memory areas will be reset. The name
can contain only printable ascii characters (including
space), except '[',']','\','$' and '`'.
This feature is available only if the kernel is built with
the CONFIG_ANON_VMA_NAME option enabled.
[surenb@google.com: docs: proc.rst: /proc/PID/maps: fix malformed table]
Link: https://lkml.kernel.org/r/20211123185928.2513763-1-surenb@google.com
[surenb: rebased over v5.15-rc6, replaced userpointer with a kernel copy,
added input sanitization and CONFIG_ANON_VMA_NAME config. The bulk of the
work here was done by Colin Cross, therefore, with his permission, keeping
him as the author]
Link: https://lkml.kernel.org/r/20211019215511.3771969-2-surenb@google.com
Signed-off-by: Colin Cross <ccross@google.com>
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Kees Cook <keescook@chromium.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Jan Glauber <jan.glauber@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: John Stultz <john.stultz@linaro.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rob Landley <rob@landley.net>
Cc: "Serge E. Hallyn" <serge.hallyn@ubuntu.com>
Cc: Shaohua Li <shli@fusionio.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-15 06:05:59 +08:00
|
|
|
struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
|
2022-03-05 12:28:51 +08:00
|
|
|
struct anon_vma_name *anon_name)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
2014-01-24 07:53:42 +08:00
|
|
|
/*
|
|
|
|
|
* VM_SOFTDIRTY should not prevent from VMA merging, if we
|
|
|
|
|
* match the flags but dirty bit -- the caller should mark
|
|
|
|
|
* merged VMA as dirty. If dirty bit won't be excluded from
|
2019-03-06 07:46:22 +08:00
|
|
|
* comparison, we increase pressure on the memory system forcing
|
2014-01-24 07:53:42 +08:00
|
|
|
* the kernel to generate new VMAs when old one could be
|
|
|
|
|
* extended instead.
|
|
|
|
|
*/
|
|
|
|
|
if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
|
2005-04-17 06:20:36 +08:00
|
|
|
return 0;
|
|
|
|
|
if (vma->vm_file != file)
|
|
|
|
|
return 0;
|
|
|
|
|
if (vma->vm_ops && vma->vm_ops->close)
|
|
|
|
|
return 0;
|
2015-09-05 06:46:24 +08:00
|
|
|
if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
|
|
|
|
|
return 0;
|
2022-03-05 12:28:51 +08:00
|
|
|
if (!anon_vma_name_eq(anon_vma_name(vma), anon_name))
|
mm: add a field to store names for private anonymous memory
In many userspace applications, and especially in VM based applications
like Android uses heavily, there are multiple different allocators in
use. At a minimum there is libc malloc and the stack, and in many cases
there are libc malloc, the stack, direct syscalls to mmap anonymous
memory, and multiple VM heaps (one for small objects, one for big
objects, etc.). Each of these layers usually has its own tools to
inspect its usage; malloc by compiling a debug version, the VM through
heap inspection tools, and for direct syscalls there is usually no way
to track them.
On Android we heavily use a set of tools that use an extended version of
the logic covered in Documentation/vm/pagemap.txt to walk all pages
mapped in userspace and slice their usage by process, shared (COW) vs.
unique mappings, backing, etc. This can account for real physical
memory usage even in cases like fork without exec (which Android uses
heavily to share as many private COW pages as possible between
processes), Kernel SamePage Merging, and clean zero pages. It produces
a measurement of the pages that only exist in that process (USS, for
unique), and a measurement of the physical memory usage of that process
with the cost of shared pages being evenly split between processes that
share them (PSS).
If all anonymous memory is indistinguishable then figuring out the real
physical memory usage (PSS) of each heap requires either a pagemap
walking tool that can understand the heap debugging of every layer, or
for every layer's heap debugging tools to implement the pagemap walking
logic, in which case it is hard to get a consistent view of memory
across the whole system.
Tracking the information in userspace leads to all sorts of problems.
It either needs to be stored inside the process, which means every
process has to have an API to export its current heap information upon
request, or it has to be stored externally in a filesystem that somebody
needs to clean up on crashes. It needs to be readable while the process
is still running, so it has to have some sort of synchronization with
every layer of userspace. Efficiently tracking the ranges requires
reimplementing something like the kernel vma trees, and linking to it
from every layer of userspace. It requires more memory, more syscalls,
more runtime cost, and more complexity to separately track regions that
the kernel is already tracking.
This patch adds a field to /proc/pid/maps and /proc/pid/smaps to show a
userspace-provided name for anonymous vmas. The names of named
anonymous vmas are shown in /proc/pid/maps and /proc/pid/smaps as
[anon:<name>].
Userspace can set the name for a region of memory by calling
prctl(PR_SET_VMA, PR_SET_VMA_ANON_NAME, start, len, (unsigned long)name)
Setting the name to NULL clears it. The name length limit is 80 bytes
including NUL-terminator and is checked to contain only printable ascii
characters (including space), except '[',']','\','$' and '`'.
Ascii strings are being used to have a descriptive identifiers for vmas,
which can be understood by the users reading /proc/pid/maps or
/proc/pid/smaps. Names can be standardized for a given system and they
can include some variable parts such as the name of the allocator or a
library, tid of the thread using it, etc.
The name is stored in a pointer in the shared union in vm_area_struct
that points to a null terminated string. Anonymous vmas with the same
name (equivalent strings) and are otherwise mergeable will be merged.
The name pointers are not shared between vmas even if they contain the
same name. The name pointer is stored in a union with fields that are
only used on file-backed mappings, so it does not increase memory usage.
CONFIG_ANON_VMA_NAME kernel configuration is introduced to enable this
feature. It keeps the feature disabled by default to prevent any
additional memory overhead and to avoid confusing procfs parsers on
systems which are not ready to support named anonymous vmas.
The patch is based on the original patch developed by Colin Cross, more
specifically on its latest version [1] posted upstream by Sumit Semwal.
It used a userspace pointer to store vma names. In that design, name
pointers could be shared between vmas. However during the last
upstreaming attempt, Kees Cook raised concerns [2] about this approach
and suggested to copy the name into kernel memory space, perform
validity checks [3] and store as a string referenced from
vm_area_struct.
One big concern is about fork() performance which would need to strdup
anonymous vma names. Dave Hansen suggested experimenting with
worst-case scenario of forking a process with 64k vmas having longest
possible names [4]. I ran this experiment on an ARM64 Android device
and recorded a worst-case regression of almost 40% when forking such a
process.
This regression is addressed in the followup patch which replaces the
pointer to a name with a refcounted structure that allows sharing the
name pointer between vmas of the same name. Instead of duplicating the
string during fork() or when splitting a vma it increments the refcount.
[1] https://lore.kernel.org/linux-mm/20200901161459.11772-4-sumit.semwal@linaro.org/
[2] https://lore.kernel.org/linux-mm/202009031031.D32EF57ED@keescook/
[3] https://lore.kernel.org/linux-mm/202009031022.3834F692@keescook/
[4] https://lore.kernel.org/linux-mm/5d0358ab-8c47-2f5f-8e43-23b89d6a8e95@intel.com/
Changes for prctl(2) manual page (in the options section):
PR_SET_VMA
Sets an attribute specified in arg2 for virtual memory areas
starting from the address specified in arg3 and spanning the
size specified in arg4. arg5 specifies the value of the attribute
to be set. Note that assigning an attribute to a virtual memory
area might prevent it from being merged with adjacent virtual
memory areas due to the difference in that attribute's value.
Currently, arg2 must be one of:
PR_SET_VMA_ANON_NAME
Set a name for anonymous virtual memory areas. arg5 should
be a pointer to a null-terminated string containing the
name. The name length including null byte cannot exceed
80 bytes. If arg5 is NULL, the name of the appropriate
anonymous virtual memory areas will be reset. The name
can contain only printable ascii characters (including
space), except '[',']','\','$' and '`'.
This feature is available only if the kernel is built with
the CONFIG_ANON_VMA_NAME option enabled.
[surenb@google.com: docs: proc.rst: /proc/PID/maps: fix malformed table]
Link: https://lkml.kernel.org/r/20211123185928.2513763-1-surenb@google.com
[surenb: rebased over v5.15-rc6, replaced userpointer with a kernel copy,
added input sanitization and CONFIG_ANON_VMA_NAME config. The bulk of the
work here was done by Colin Cross, therefore, with his permission, keeping
him as the author]
Link: https://lkml.kernel.org/r/20211019215511.3771969-2-surenb@google.com
Signed-off-by: Colin Cross <ccross@google.com>
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Kees Cook <keescook@chromium.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Jan Glauber <jan.glauber@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: John Stultz <john.stultz@linaro.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rob Landley <rob@landley.net>
Cc: "Serge E. Hallyn" <serge.hallyn@ubuntu.com>
Cc: Shaohua Li <shli@fusionio.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-15 06:05:59 +08:00
|
|
|
return 0;
|
2005-04-17 06:20:36 +08:00
|
|
|
return 1;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
|
2011-05-25 08:11:20 +08:00
|
|
|
struct anon_vma *anon_vma2,
|
|
|
|
|
struct vm_area_struct *vma)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
2011-05-25 08:11:20 +08:00
|
|
|
/*
|
|
|
|
|
* The list_is_singular() test is to avoid merging VMA cloned from
|
|
|
|
|
* parents. This can improve scalability caused by anon_vma lock.
|
|
|
|
|
*/
|
|
|
|
|
if ((!anon_vma1 || !anon_vma2) && (!vma ||
|
|
|
|
|
list_is_singular(&vma->anon_vma_chain)))
|
|
|
|
|
return 1;
|
|
|
|
|
return anon_vma1 == anon_vma2;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
|
|
|
|
|
* in front of (at a lower virtual address and file offset than) the vma.
|
|
|
|
|
*
|
|
|
|
|
* We cannot merge two vmas if they have differently assigned (non-NULL)
|
|
|
|
|
* anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
|
|
|
|
|
*
|
|
|
|
|
* We don't check here for the merged mmap wrapping around the end of pagecache
|
2020-08-07 14:23:37 +08:00
|
|
|
* indices (16TB on ia32) because do_mmap() does not permit mmap's which
|
2005-04-17 06:20:36 +08:00
|
|
|
* wrap, nor mmaps which cover the final page at index -1UL.
|
|
|
|
|
*/
|
|
|
|
|
static int
|
|
|
|
|
can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
|
2015-09-05 06:46:24 +08:00
|
|
|
struct anon_vma *anon_vma, struct file *file,
|
|
|
|
|
pgoff_t vm_pgoff,
|
mm: add a field to store names for private anonymous memory
In many userspace applications, and especially in VM based applications
like Android uses heavily, there are multiple different allocators in
use. At a minimum there is libc malloc and the stack, and in many cases
there are libc malloc, the stack, direct syscalls to mmap anonymous
memory, and multiple VM heaps (one for small objects, one for big
objects, etc.). Each of these layers usually has its own tools to
inspect its usage; malloc by compiling a debug version, the VM through
heap inspection tools, and for direct syscalls there is usually no way
to track them.
On Android we heavily use a set of tools that use an extended version of
the logic covered in Documentation/vm/pagemap.txt to walk all pages
mapped in userspace and slice their usage by process, shared (COW) vs.
unique mappings, backing, etc. This can account for real physical
memory usage even in cases like fork without exec (which Android uses
heavily to share as many private COW pages as possible between
processes), Kernel SamePage Merging, and clean zero pages. It produces
a measurement of the pages that only exist in that process (USS, for
unique), and a measurement of the physical memory usage of that process
with the cost of shared pages being evenly split between processes that
share them (PSS).
If all anonymous memory is indistinguishable then figuring out the real
physical memory usage (PSS) of each heap requires either a pagemap
walking tool that can understand the heap debugging of every layer, or
for every layer's heap debugging tools to implement the pagemap walking
logic, in which case it is hard to get a consistent view of memory
across the whole system.
Tracking the information in userspace leads to all sorts of problems.
It either needs to be stored inside the process, which means every
process has to have an API to export its current heap information upon
request, or it has to be stored externally in a filesystem that somebody
needs to clean up on crashes. It needs to be readable while the process
is still running, so it has to have some sort of synchronization with
every layer of userspace. Efficiently tracking the ranges requires
reimplementing something like the kernel vma trees, and linking to it
from every layer of userspace. It requires more memory, more syscalls,
more runtime cost, and more complexity to separately track regions that
the kernel is already tracking.
This patch adds a field to /proc/pid/maps and /proc/pid/smaps to show a
userspace-provided name for anonymous vmas. The names of named
anonymous vmas are shown in /proc/pid/maps and /proc/pid/smaps as
[anon:<name>].
Userspace can set the name for a region of memory by calling
prctl(PR_SET_VMA, PR_SET_VMA_ANON_NAME, start, len, (unsigned long)name)
Setting the name to NULL clears it. The name length limit is 80 bytes
including NUL-terminator and is checked to contain only printable ascii
characters (including space), except '[',']','\','$' and '`'.
Ascii strings are being used to have a descriptive identifiers for vmas,
which can be understood by the users reading /proc/pid/maps or
/proc/pid/smaps. Names can be standardized for a given system and they
can include some variable parts such as the name of the allocator or a
library, tid of the thread using it, etc.
The name is stored in a pointer in the shared union in vm_area_struct
that points to a null terminated string. Anonymous vmas with the same
name (equivalent strings) and are otherwise mergeable will be merged.
The name pointers are not shared between vmas even if they contain the
same name. The name pointer is stored in a union with fields that are
only used on file-backed mappings, so it does not increase memory usage.
CONFIG_ANON_VMA_NAME kernel configuration is introduced to enable this
feature. It keeps the feature disabled by default to prevent any
additional memory overhead and to avoid confusing procfs parsers on
systems which are not ready to support named anonymous vmas.
The patch is based on the original patch developed by Colin Cross, more
specifically on its latest version [1] posted upstream by Sumit Semwal.
It used a userspace pointer to store vma names. In that design, name
pointers could be shared between vmas. However during the last
upstreaming attempt, Kees Cook raised concerns [2] about this approach
and suggested to copy the name into kernel memory space, perform
validity checks [3] and store as a string referenced from
vm_area_struct.
One big concern is about fork() performance which would need to strdup
anonymous vma names. Dave Hansen suggested experimenting with
worst-case scenario of forking a process with 64k vmas having longest
possible names [4]. I ran this experiment on an ARM64 Android device
and recorded a worst-case regression of almost 40% when forking such a
process.
This regression is addressed in the followup patch which replaces the
pointer to a name with a refcounted structure that allows sharing the
name pointer between vmas of the same name. Instead of duplicating the
string during fork() or when splitting a vma it increments the refcount.
[1] https://lore.kernel.org/linux-mm/20200901161459.11772-4-sumit.semwal@linaro.org/
[2] https://lore.kernel.org/linux-mm/202009031031.D32EF57ED@keescook/
[3] https://lore.kernel.org/linux-mm/202009031022.3834F692@keescook/
[4] https://lore.kernel.org/linux-mm/5d0358ab-8c47-2f5f-8e43-23b89d6a8e95@intel.com/
Changes for prctl(2) manual page (in the options section):
PR_SET_VMA
Sets an attribute specified in arg2 for virtual memory areas
starting from the address specified in arg3 and spanning the
size specified in arg4. arg5 specifies the value of the attribute
to be set. Note that assigning an attribute to a virtual memory
area might prevent it from being merged with adjacent virtual
memory areas due to the difference in that attribute's value.
Currently, arg2 must be one of:
PR_SET_VMA_ANON_NAME
Set a name for anonymous virtual memory areas. arg5 should
be a pointer to a null-terminated string containing the
name. The name length including null byte cannot exceed
80 bytes. If arg5 is NULL, the name of the appropriate
anonymous virtual memory areas will be reset. The name
can contain only printable ascii characters (including
space), except '[',']','\','$' and '`'.
This feature is available only if the kernel is built with
the CONFIG_ANON_VMA_NAME option enabled.
[surenb@google.com: docs: proc.rst: /proc/PID/maps: fix malformed table]
Link: https://lkml.kernel.org/r/20211123185928.2513763-1-surenb@google.com
[surenb: rebased over v5.15-rc6, replaced userpointer with a kernel copy,
added input sanitization and CONFIG_ANON_VMA_NAME config. The bulk of the
work here was done by Colin Cross, therefore, with his permission, keeping
him as the author]
Link: https://lkml.kernel.org/r/20211019215511.3771969-2-surenb@google.com
Signed-off-by: Colin Cross <ccross@google.com>
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Kees Cook <keescook@chromium.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Jan Glauber <jan.glauber@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: John Stultz <john.stultz@linaro.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rob Landley <rob@landley.net>
Cc: "Serge E. Hallyn" <serge.hallyn@ubuntu.com>
Cc: Shaohua Li <shli@fusionio.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-15 06:05:59 +08:00
|
|
|
struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
|
2022-03-05 12:28:51 +08:00
|
|
|
struct anon_vma_name *anon_name)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
mm: add a field to store names for private anonymous memory
In many userspace applications, and especially in VM based applications
like Android uses heavily, there are multiple different allocators in
use. At a minimum there is libc malloc and the stack, and in many cases
there are libc malloc, the stack, direct syscalls to mmap anonymous
memory, and multiple VM heaps (one for small objects, one for big
objects, etc.). Each of these layers usually has its own tools to
inspect its usage; malloc by compiling a debug version, the VM through
heap inspection tools, and for direct syscalls there is usually no way
to track them.
On Android we heavily use a set of tools that use an extended version of
the logic covered in Documentation/vm/pagemap.txt to walk all pages
mapped in userspace and slice their usage by process, shared (COW) vs.
unique mappings, backing, etc. This can account for real physical
memory usage even in cases like fork without exec (which Android uses
heavily to share as many private COW pages as possible between
processes), Kernel SamePage Merging, and clean zero pages. It produces
a measurement of the pages that only exist in that process (USS, for
unique), and a measurement of the physical memory usage of that process
with the cost of shared pages being evenly split between processes that
share them (PSS).
If all anonymous memory is indistinguishable then figuring out the real
physical memory usage (PSS) of each heap requires either a pagemap
walking tool that can understand the heap debugging of every layer, or
for every layer's heap debugging tools to implement the pagemap walking
logic, in which case it is hard to get a consistent view of memory
across the whole system.
Tracking the information in userspace leads to all sorts of problems.
It either needs to be stored inside the process, which means every
process has to have an API to export its current heap information upon
request, or it has to be stored externally in a filesystem that somebody
needs to clean up on crashes. It needs to be readable while the process
is still running, so it has to have some sort of synchronization with
every layer of userspace. Efficiently tracking the ranges requires
reimplementing something like the kernel vma trees, and linking to it
from every layer of userspace. It requires more memory, more syscalls,
more runtime cost, and more complexity to separately track regions that
the kernel is already tracking.
This patch adds a field to /proc/pid/maps and /proc/pid/smaps to show a
userspace-provided name for anonymous vmas. The names of named
anonymous vmas are shown in /proc/pid/maps and /proc/pid/smaps as
[anon:<name>].
Userspace can set the name for a region of memory by calling
prctl(PR_SET_VMA, PR_SET_VMA_ANON_NAME, start, len, (unsigned long)name)
Setting the name to NULL clears it. The name length limit is 80 bytes
including NUL-terminator and is checked to contain only printable ascii
characters (including space), except '[',']','\','$' and '`'.
Ascii strings are being used to have a descriptive identifiers for vmas,
which can be understood by the users reading /proc/pid/maps or
/proc/pid/smaps. Names can be standardized for a given system and they
can include some variable parts such as the name of the allocator or a
library, tid of the thread using it, etc.
The name is stored in a pointer in the shared union in vm_area_struct
that points to a null terminated string. Anonymous vmas with the same
name (equivalent strings) and are otherwise mergeable will be merged.
The name pointers are not shared between vmas even if they contain the
same name. The name pointer is stored in a union with fields that are
only used on file-backed mappings, so it does not increase memory usage.
CONFIG_ANON_VMA_NAME kernel configuration is introduced to enable this
feature. It keeps the feature disabled by default to prevent any
additional memory overhead and to avoid confusing procfs parsers on
systems which are not ready to support named anonymous vmas.
The patch is based on the original patch developed by Colin Cross, more
specifically on its latest version [1] posted upstream by Sumit Semwal.
It used a userspace pointer to store vma names. In that design, name
pointers could be shared between vmas. However during the last
upstreaming attempt, Kees Cook raised concerns [2] about this approach
and suggested to copy the name into kernel memory space, perform
validity checks [3] and store as a string referenced from
vm_area_struct.
One big concern is about fork() performance which would need to strdup
anonymous vma names. Dave Hansen suggested experimenting with
worst-case scenario of forking a process with 64k vmas having longest
possible names [4]. I ran this experiment on an ARM64 Android device
and recorded a worst-case regression of almost 40% when forking such a
process.
This regression is addressed in the followup patch which replaces the
pointer to a name with a refcounted structure that allows sharing the
name pointer between vmas of the same name. Instead of duplicating the
string during fork() or when splitting a vma it increments the refcount.
[1] https://lore.kernel.org/linux-mm/20200901161459.11772-4-sumit.semwal@linaro.org/
[2] https://lore.kernel.org/linux-mm/202009031031.D32EF57ED@keescook/
[3] https://lore.kernel.org/linux-mm/202009031022.3834F692@keescook/
[4] https://lore.kernel.org/linux-mm/5d0358ab-8c47-2f5f-8e43-23b89d6a8e95@intel.com/
Changes for prctl(2) manual page (in the options section):
PR_SET_VMA
Sets an attribute specified in arg2 for virtual memory areas
starting from the address specified in arg3 and spanning the
size specified in arg4. arg5 specifies the value of the attribute
to be set. Note that assigning an attribute to a virtual memory
area might prevent it from being merged with adjacent virtual
memory areas due to the difference in that attribute's value.
Currently, arg2 must be one of:
PR_SET_VMA_ANON_NAME
Set a name for anonymous virtual memory areas. arg5 should
be a pointer to a null-terminated string containing the
name. The name length including null byte cannot exceed
80 bytes. If arg5 is NULL, the name of the appropriate
anonymous virtual memory areas will be reset. The name
can contain only printable ascii characters (including
space), except '[',']','\','$' and '`'.
This feature is available only if the kernel is built with
the CONFIG_ANON_VMA_NAME option enabled.
[surenb@google.com: docs: proc.rst: /proc/PID/maps: fix malformed table]
Link: https://lkml.kernel.org/r/20211123185928.2513763-1-surenb@google.com
[surenb: rebased over v5.15-rc6, replaced userpointer with a kernel copy,
added input sanitization and CONFIG_ANON_VMA_NAME config. The bulk of the
work here was done by Colin Cross, therefore, with his permission, keeping
him as the author]
Link: https://lkml.kernel.org/r/20211019215511.3771969-2-surenb@google.com
Signed-off-by: Colin Cross <ccross@google.com>
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Kees Cook <keescook@chromium.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Jan Glauber <jan.glauber@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: John Stultz <john.stultz@linaro.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rob Landley <rob@landley.net>
Cc: "Serge E. Hallyn" <serge.hallyn@ubuntu.com>
Cc: Shaohua Li <shli@fusionio.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-15 06:05:59 +08:00
|
|
|
if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name) &&
|
2011-05-25 08:11:20 +08:00
|
|
|
is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
|
2005-04-17 06:20:36 +08:00
|
|
|
if (vma->vm_pgoff == vm_pgoff)
|
|
|
|
|
return 1;
|
|
|
|
|
}
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
|
|
|
|
|
* beyond (at a higher virtual address and file offset than) the vma.
|
|
|
|
|
*
|
|
|
|
|
* We cannot merge two vmas if they have differently assigned (non-NULL)
|
|
|
|
|
* anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
|
|
|
|
|
*/
|
|
|
|
|
static int
|
|
|
|
|
can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
|
2015-09-05 06:46:24 +08:00
|
|
|
struct anon_vma *anon_vma, struct file *file,
|
|
|
|
|
pgoff_t vm_pgoff,
|
mm: add a field to store names for private anonymous memory
In many userspace applications, and especially in VM based applications
like Android uses heavily, there are multiple different allocators in
use. At a minimum there is libc malloc and the stack, and in many cases
there are libc malloc, the stack, direct syscalls to mmap anonymous
memory, and multiple VM heaps (one for small objects, one for big
objects, etc.). Each of these layers usually has its own tools to
inspect its usage; malloc by compiling a debug version, the VM through
heap inspection tools, and for direct syscalls there is usually no way
to track them.
On Android we heavily use a set of tools that use an extended version of
the logic covered in Documentation/vm/pagemap.txt to walk all pages
mapped in userspace and slice their usage by process, shared (COW) vs.
unique mappings, backing, etc. This can account for real physical
memory usage even in cases like fork without exec (which Android uses
heavily to share as many private COW pages as possible between
processes), Kernel SamePage Merging, and clean zero pages. It produces
a measurement of the pages that only exist in that process (USS, for
unique), and a measurement of the physical memory usage of that process
with the cost of shared pages being evenly split between processes that
share them (PSS).
If all anonymous memory is indistinguishable then figuring out the real
physical memory usage (PSS) of each heap requires either a pagemap
walking tool that can understand the heap debugging of every layer, or
for every layer's heap debugging tools to implement the pagemap walking
logic, in which case it is hard to get a consistent view of memory
across the whole system.
Tracking the information in userspace leads to all sorts of problems.
It either needs to be stored inside the process, which means every
process has to have an API to export its current heap information upon
request, or it has to be stored externally in a filesystem that somebody
needs to clean up on crashes. It needs to be readable while the process
is still running, so it has to have some sort of synchronization with
every layer of userspace. Efficiently tracking the ranges requires
reimplementing something like the kernel vma trees, and linking to it
from every layer of userspace. It requires more memory, more syscalls,
more runtime cost, and more complexity to separately track regions that
the kernel is already tracking.
This patch adds a field to /proc/pid/maps and /proc/pid/smaps to show a
userspace-provided name for anonymous vmas. The names of named
anonymous vmas are shown in /proc/pid/maps and /proc/pid/smaps as
[anon:<name>].
Userspace can set the name for a region of memory by calling
prctl(PR_SET_VMA, PR_SET_VMA_ANON_NAME, start, len, (unsigned long)name)
Setting the name to NULL clears it. The name length limit is 80 bytes
including NUL-terminator and is checked to contain only printable ascii
characters (including space), except '[',']','\','$' and '`'.
Ascii strings are being used to have a descriptive identifiers for vmas,
which can be understood by the users reading /proc/pid/maps or
/proc/pid/smaps. Names can be standardized for a given system and they
can include some variable parts such as the name of the allocator or a
library, tid of the thread using it, etc.
The name is stored in a pointer in the shared union in vm_area_struct
that points to a null terminated string. Anonymous vmas with the same
name (equivalent strings) and are otherwise mergeable will be merged.
The name pointers are not shared between vmas even if they contain the
same name. The name pointer is stored in a union with fields that are
only used on file-backed mappings, so it does not increase memory usage.
CONFIG_ANON_VMA_NAME kernel configuration is introduced to enable this
feature. It keeps the feature disabled by default to prevent any
additional memory overhead and to avoid confusing procfs parsers on
systems which are not ready to support named anonymous vmas.
The patch is based on the original patch developed by Colin Cross, more
specifically on its latest version [1] posted upstream by Sumit Semwal.
It used a userspace pointer to store vma names. In that design, name
pointers could be shared between vmas. However during the last
upstreaming attempt, Kees Cook raised concerns [2] about this approach
and suggested to copy the name into kernel memory space, perform
validity checks [3] and store as a string referenced from
vm_area_struct.
One big concern is about fork() performance which would need to strdup
anonymous vma names. Dave Hansen suggested experimenting with
worst-case scenario of forking a process with 64k vmas having longest
possible names [4]. I ran this experiment on an ARM64 Android device
and recorded a worst-case regression of almost 40% when forking such a
process.
This regression is addressed in the followup patch which replaces the
pointer to a name with a refcounted structure that allows sharing the
name pointer between vmas of the same name. Instead of duplicating the
string during fork() or when splitting a vma it increments the refcount.
[1] https://lore.kernel.org/linux-mm/20200901161459.11772-4-sumit.semwal@linaro.org/
[2] https://lore.kernel.org/linux-mm/202009031031.D32EF57ED@keescook/
[3] https://lore.kernel.org/linux-mm/202009031022.3834F692@keescook/
[4] https://lore.kernel.org/linux-mm/5d0358ab-8c47-2f5f-8e43-23b89d6a8e95@intel.com/
Changes for prctl(2) manual page (in the options section):
PR_SET_VMA
Sets an attribute specified in arg2 for virtual memory areas
starting from the address specified in arg3 and spanning the
size specified in arg4. arg5 specifies the value of the attribute
to be set. Note that assigning an attribute to a virtual memory
area might prevent it from being merged with adjacent virtual
memory areas due to the difference in that attribute's value.
Currently, arg2 must be one of:
PR_SET_VMA_ANON_NAME
Set a name for anonymous virtual memory areas. arg5 should
be a pointer to a null-terminated string containing the
name. The name length including null byte cannot exceed
80 bytes. If arg5 is NULL, the name of the appropriate
anonymous virtual memory areas will be reset. The name
can contain only printable ascii characters (including
space), except '[',']','\','$' and '`'.
This feature is available only if the kernel is built with
the CONFIG_ANON_VMA_NAME option enabled.
[surenb@google.com: docs: proc.rst: /proc/PID/maps: fix malformed table]
Link: https://lkml.kernel.org/r/20211123185928.2513763-1-surenb@google.com
[surenb: rebased over v5.15-rc6, replaced userpointer with a kernel copy,
added input sanitization and CONFIG_ANON_VMA_NAME config. The bulk of the
work here was done by Colin Cross, therefore, with his permission, keeping
him as the author]
Link: https://lkml.kernel.org/r/20211019215511.3771969-2-surenb@google.com
Signed-off-by: Colin Cross <ccross@google.com>
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Kees Cook <keescook@chromium.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Jan Glauber <jan.glauber@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: John Stultz <john.stultz@linaro.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rob Landley <rob@landley.net>
Cc: "Serge E. Hallyn" <serge.hallyn@ubuntu.com>
Cc: Shaohua Li <shli@fusionio.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-15 06:05:59 +08:00
|
|
|
struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
|
2022-03-05 12:28:51 +08:00
|
|
|
struct anon_vma_name *anon_name)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
mm: add a field to store names for private anonymous memory
In many userspace applications, and especially in VM based applications
like Android uses heavily, there are multiple different allocators in
use. At a minimum there is libc malloc and the stack, and in many cases
there are libc malloc, the stack, direct syscalls to mmap anonymous
memory, and multiple VM heaps (one for small objects, one for big
objects, etc.). Each of these layers usually has its own tools to
inspect its usage; malloc by compiling a debug version, the VM through
heap inspection tools, and for direct syscalls there is usually no way
to track them.
On Android we heavily use a set of tools that use an extended version of
the logic covered in Documentation/vm/pagemap.txt to walk all pages
mapped in userspace and slice their usage by process, shared (COW) vs.
unique mappings, backing, etc. This can account for real physical
memory usage even in cases like fork without exec (which Android uses
heavily to share as many private COW pages as possible between
processes), Kernel SamePage Merging, and clean zero pages. It produces
a measurement of the pages that only exist in that process (USS, for
unique), and a measurement of the physical memory usage of that process
with the cost of shared pages being evenly split between processes that
share them (PSS).
If all anonymous memory is indistinguishable then figuring out the real
physical memory usage (PSS) of each heap requires either a pagemap
walking tool that can understand the heap debugging of every layer, or
for every layer's heap debugging tools to implement the pagemap walking
logic, in which case it is hard to get a consistent view of memory
across the whole system.
Tracking the information in userspace leads to all sorts of problems.
It either needs to be stored inside the process, which means every
process has to have an API to export its current heap information upon
request, or it has to be stored externally in a filesystem that somebody
needs to clean up on crashes. It needs to be readable while the process
is still running, so it has to have some sort of synchronization with
every layer of userspace. Efficiently tracking the ranges requires
reimplementing something like the kernel vma trees, and linking to it
from every layer of userspace. It requires more memory, more syscalls,
more runtime cost, and more complexity to separately track regions that
the kernel is already tracking.
This patch adds a field to /proc/pid/maps and /proc/pid/smaps to show a
userspace-provided name for anonymous vmas. The names of named
anonymous vmas are shown in /proc/pid/maps and /proc/pid/smaps as
[anon:<name>].
Userspace can set the name for a region of memory by calling
prctl(PR_SET_VMA, PR_SET_VMA_ANON_NAME, start, len, (unsigned long)name)
Setting the name to NULL clears it. The name length limit is 80 bytes
including NUL-terminator and is checked to contain only printable ascii
characters (including space), except '[',']','\','$' and '`'.
Ascii strings are being used to have a descriptive identifiers for vmas,
which can be understood by the users reading /proc/pid/maps or
/proc/pid/smaps. Names can be standardized for a given system and they
can include some variable parts such as the name of the allocator or a
library, tid of the thread using it, etc.
The name is stored in a pointer in the shared union in vm_area_struct
that points to a null terminated string. Anonymous vmas with the same
name (equivalent strings) and are otherwise mergeable will be merged.
The name pointers are not shared between vmas even if they contain the
same name. The name pointer is stored in a union with fields that are
only used on file-backed mappings, so it does not increase memory usage.
CONFIG_ANON_VMA_NAME kernel configuration is introduced to enable this
feature. It keeps the feature disabled by default to prevent any
additional memory overhead and to avoid confusing procfs parsers on
systems which are not ready to support named anonymous vmas.
The patch is based on the original patch developed by Colin Cross, more
specifically on its latest version [1] posted upstream by Sumit Semwal.
It used a userspace pointer to store vma names. In that design, name
pointers could be shared between vmas. However during the last
upstreaming attempt, Kees Cook raised concerns [2] about this approach
and suggested to copy the name into kernel memory space, perform
validity checks [3] and store as a string referenced from
vm_area_struct.
One big concern is about fork() performance which would need to strdup
anonymous vma names. Dave Hansen suggested experimenting with
worst-case scenario of forking a process with 64k vmas having longest
possible names [4]. I ran this experiment on an ARM64 Android device
and recorded a worst-case regression of almost 40% when forking such a
process.
This regression is addressed in the followup patch which replaces the
pointer to a name with a refcounted structure that allows sharing the
name pointer between vmas of the same name. Instead of duplicating the
string during fork() or when splitting a vma it increments the refcount.
[1] https://lore.kernel.org/linux-mm/20200901161459.11772-4-sumit.semwal@linaro.org/
[2] https://lore.kernel.org/linux-mm/202009031031.D32EF57ED@keescook/
[3] https://lore.kernel.org/linux-mm/202009031022.3834F692@keescook/
[4] https://lore.kernel.org/linux-mm/5d0358ab-8c47-2f5f-8e43-23b89d6a8e95@intel.com/
Changes for prctl(2) manual page (in the options section):
PR_SET_VMA
Sets an attribute specified in arg2 for virtual memory areas
starting from the address specified in arg3 and spanning the
size specified in arg4. arg5 specifies the value of the attribute
to be set. Note that assigning an attribute to a virtual memory
area might prevent it from being merged with adjacent virtual
memory areas due to the difference in that attribute's value.
Currently, arg2 must be one of:
PR_SET_VMA_ANON_NAME
Set a name for anonymous virtual memory areas. arg5 should
be a pointer to a null-terminated string containing the
name. The name length including null byte cannot exceed
80 bytes. If arg5 is NULL, the name of the appropriate
anonymous virtual memory areas will be reset. The name
can contain only printable ascii characters (including
space), except '[',']','\','$' and '`'.
This feature is available only if the kernel is built with
the CONFIG_ANON_VMA_NAME option enabled.
[surenb@google.com: docs: proc.rst: /proc/PID/maps: fix malformed table]
Link: https://lkml.kernel.org/r/20211123185928.2513763-1-surenb@google.com
[surenb: rebased over v5.15-rc6, replaced userpointer with a kernel copy,
added input sanitization and CONFIG_ANON_VMA_NAME config. The bulk of the
work here was done by Colin Cross, therefore, with his permission, keeping
him as the author]
Link: https://lkml.kernel.org/r/20211019215511.3771969-2-surenb@google.com
Signed-off-by: Colin Cross <ccross@google.com>
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Kees Cook <keescook@chromium.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Jan Glauber <jan.glauber@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: John Stultz <john.stultz@linaro.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rob Landley <rob@landley.net>
Cc: "Serge E. Hallyn" <serge.hallyn@ubuntu.com>
Cc: Shaohua Li <shli@fusionio.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-15 06:05:59 +08:00
|
|
|
if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name) &&
|
2011-05-25 08:11:20 +08:00
|
|
|
is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
|
2005-04-17 06:20:36 +08:00
|
|
|
pgoff_t vm_pglen;
|
2013-07-04 06:01:26 +08:00
|
|
|
vm_pglen = vma_pages(vma);
|
2005-04-17 06:20:36 +08:00
|
|
|
if (vma->vm_pgoff + vm_pglen == vm_pgoff)
|
|
|
|
|
return 1;
|
|
|
|
|
}
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
mm: add a field to store names for private anonymous memory
In many userspace applications, and especially in VM based applications
like Android uses heavily, there are multiple different allocators in
use. At a minimum there is libc malloc and the stack, and in many cases
there are libc malloc, the stack, direct syscalls to mmap anonymous
memory, and multiple VM heaps (one for small objects, one for big
objects, etc.). Each of these layers usually has its own tools to
inspect its usage; malloc by compiling a debug version, the VM through
heap inspection tools, and for direct syscalls there is usually no way
to track them.
On Android we heavily use a set of tools that use an extended version of
the logic covered in Documentation/vm/pagemap.txt to walk all pages
mapped in userspace and slice their usage by process, shared (COW) vs.
unique mappings, backing, etc. This can account for real physical
memory usage even in cases like fork without exec (which Android uses
heavily to share as many private COW pages as possible between
processes), Kernel SamePage Merging, and clean zero pages. It produces
a measurement of the pages that only exist in that process (USS, for
unique), and a measurement of the physical memory usage of that process
with the cost of shared pages being evenly split between processes that
share them (PSS).
If all anonymous memory is indistinguishable then figuring out the real
physical memory usage (PSS) of each heap requires either a pagemap
walking tool that can understand the heap debugging of every layer, or
for every layer's heap debugging tools to implement the pagemap walking
logic, in which case it is hard to get a consistent view of memory
across the whole system.
Tracking the information in userspace leads to all sorts of problems.
It either needs to be stored inside the process, which means every
process has to have an API to export its current heap information upon
request, or it has to be stored externally in a filesystem that somebody
needs to clean up on crashes. It needs to be readable while the process
is still running, so it has to have some sort of synchronization with
every layer of userspace. Efficiently tracking the ranges requires
reimplementing something like the kernel vma trees, and linking to it
from every layer of userspace. It requires more memory, more syscalls,
more runtime cost, and more complexity to separately track regions that
the kernel is already tracking.
This patch adds a field to /proc/pid/maps and /proc/pid/smaps to show a
userspace-provided name for anonymous vmas. The names of named
anonymous vmas are shown in /proc/pid/maps and /proc/pid/smaps as
[anon:<name>].
Userspace can set the name for a region of memory by calling
prctl(PR_SET_VMA, PR_SET_VMA_ANON_NAME, start, len, (unsigned long)name)
Setting the name to NULL clears it. The name length limit is 80 bytes
including NUL-terminator and is checked to contain only printable ascii
characters (including space), except '[',']','\','$' and '`'.
Ascii strings are being used to have a descriptive identifiers for vmas,
which can be understood by the users reading /proc/pid/maps or
/proc/pid/smaps. Names can be standardized for a given system and they
can include some variable parts such as the name of the allocator or a
library, tid of the thread using it, etc.
The name is stored in a pointer in the shared union in vm_area_struct
that points to a null terminated string. Anonymous vmas with the same
name (equivalent strings) and are otherwise mergeable will be merged.
The name pointers are not shared between vmas even if they contain the
same name. The name pointer is stored in a union with fields that are
only used on file-backed mappings, so it does not increase memory usage.
CONFIG_ANON_VMA_NAME kernel configuration is introduced to enable this
feature. It keeps the feature disabled by default to prevent any
additional memory overhead and to avoid confusing procfs parsers on
systems which are not ready to support named anonymous vmas.
The patch is based on the original patch developed by Colin Cross, more
specifically on its latest version [1] posted upstream by Sumit Semwal.
It used a userspace pointer to store vma names. In that design, name
pointers could be shared between vmas. However during the last
upstreaming attempt, Kees Cook raised concerns [2] about this approach
and suggested to copy the name into kernel memory space, perform
validity checks [3] and store as a string referenced from
vm_area_struct.
One big concern is about fork() performance which would need to strdup
anonymous vma names. Dave Hansen suggested experimenting with
worst-case scenario of forking a process with 64k vmas having longest
possible names [4]. I ran this experiment on an ARM64 Android device
and recorded a worst-case regression of almost 40% when forking such a
process.
This regression is addressed in the followup patch which replaces the
pointer to a name with a refcounted structure that allows sharing the
name pointer between vmas of the same name. Instead of duplicating the
string during fork() or when splitting a vma it increments the refcount.
[1] https://lore.kernel.org/linux-mm/20200901161459.11772-4-sumit.semwal@linaro.org/
[2] https://lore.kernel.org/linux-mm/202009031031.D32EF57ED@keescook/
[3] https://lore.kernel.org/linux-mm/202009031022.3834F692@keescook/
[4] https://lore.kernel.org/linux-mm/5d0358ab-8c47-2f5f-8e43-23b89d6a8e95@intel.com/
Changes for prctl(2) manual page (in the options section):
PR_SET_VMA
Sets an attribute specified in arg2 for virtual memory areas
starting from the address specified in arg3 and spanning the
size specified in arg4. arg5 specifies the value of the attribute
to be set. Note that assigning an attribute to a virtual memory
area might prevent it from being merged with adjacent virtual
memory areas due to the difference in that attribute's value.
Currently, arg2 must be one of:
PR_SET_VMA_ANON_NAME
Set a name for anonymous virtual memory areas. arg5 should
be a pointer to a null-terminated string containing the
name. The name length including null byte cannot exceed
80 bytes. If arg5 is NULL, the name of the appropriate
anonymous virtual memory areas will be reset. The name
can contain only printable ascii characters (including
space), except '[',']','\','$' and '`'.
This feature is available only if the kernel is built with
the CONFIG_ANON_VMA_NAME option enabled.
[surenb@google.com: docs: proc.rst: /proc/PID/maps: fix malformed table]
Link: https://lkml.kernel.org/r/20211123185928.2513763-1-surenb@google.com
[surenb: rebased over v5.15-rc6, replaced userpointer with a kernel copy,
added input sanitization and CONFIG_ANON_VMA_NAME config. The bulk of the
work here was done by Colin Cross, therefore, with his permission, keeping
him as the author]
Link: https://lkml.kernel.org/r/20211019215511.3771969-2-surenb@google.com
Signed-off-by: Colin Cross <ccross@google.com>
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Kees Cook <keescook@chromium.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Jan Glauber <jan.glauber@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: John Stultz <john.stultz@linaro.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rob Landley <rob@landley.net>
Cc: "Serge E. Hallyn" <serge.hallyn@ubuntu.com>
Cc: Shaohua Li <shli@fusionio.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-15 06:05:59 +08:00
|
|
|
* Given a mapping request (addr,end,vm_flags,file,pgoff,anon_name),
|
|
|
|
|
* figure out whether that can be merged with its predecessor or its
|
|
|
|
|
* successor. Or both (it neatly fills a hole).
|
2005-04-17 06:20:36 +08:00
|
|
|
*
|
|
|
|
|
* In most cases - when called for mmap, brk or mremap - [addr,end) is
|
|
|
|
|
* certain not to be mapped by the time vma_merge is called; but when
|
|
|
|
|
* called for mprotect, it is certain to be already mapped (either at
|
|
|
|
|
* an offset within prev, or at the start of next), and the flags of
|
|
|
|
|
* this area are about to be changed to vm_flags - and the no-change
|
|
|
|
|
* case has already been eliminated.
|
|
|
|
|
*
|
|
|
|
|
* The following mprotect cases have to be considered, where AAAA is
|
|
|
|
|
* the area passed down from mprotect_fixup, never extending beyond one
|
|
|
|
|
* vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
|
|
|
|
|
*
|
2019-12-01 09:57:39 +08:00
|
|
|
* AAAA AAAA AAAA
|
|
|
|
|
* PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN
|
|
|
|
|
* cannot merge might become might become
|
|
|
|
|
* PPNNNNNNNNNN PPPPPPPPPPNN
|
|
|
|
|
* mmap, brk or case 4 below case 5 below
|
|
|
|
|
* mremap move:
|
|
|
|
|
* AAAA AAAA
|
|
|
|
|
* PPPP NNNN PPPPNNNNXXXX
|
|
|
|
|
* might become might become
|
|
|
|
|
* PPPPPPPPPPPP 1 or PPPPPPPPPPPP 6 or
|
|
|
|
|
* PPPPPPPPNNNN 2 or PPPPPPPPXXXX 7 or
|
|
|
|
|
* PPPPNNNNNNNN 3 PPPPXXXXXXXX 8
|
2005-04-17 06:20:36 +08:00
|
|
|
*
|
2019-03-06 07:46:22 +08:00
|
|
|
* It is important for case 8 that the vma NNNN overlapping the
|
mm: vma_merge: fix vm_page_prot SMP race condition against rmap_walk
The rmap_walk can access vm_page_prot (and potentially vm_flags in the
pte/pmd manipulations). So it's not safe to wait the caller to update
the vm_page_prot/vm_flags after vma_merge returned potentially removing
the "next" vma and extending the "current" vma over the
next->vm_start,vm_end range, but still with the "current" vma
vm_page_prot, after releasing the rmap locks.
The vm_page_prot/vm_flags must be transferred from the "next" vma to the
current vma while vma_merge still holds the rmap locks.
The side effect of this race condition is pte corruption during migrate
as remove_migration_ptes when run on a address of the "next" vma that
got removed, used the vm_page_prot of the current vma.
migrate mprotect
------------ -------------
migrating in "next" vma
vma_merge() # removes "next" vma and
# extends "current" vma
# current vma is not with
# vm_page_prot updated
remove_migration_ptes
read vm_page_prot of current "vma"
establish pte with wrong permissions
vm_set_page_prot(vma) # too late!
change_protection in the old vma range
only, next range is not updated
This caused segmentation faults and potentially memory corruption in
heavy mprotect loads with some light page migration caused by compaction
in the background.
Hugh Dickins pointed out the comment about the Odd case 8 in vma_merge
which confirms the case 8 is only buggy one where the race can trigger,
in all other vma_merge cases the above cannot happen.
This fix removes the oddness factor from case 8 and it converts it from:
AAAA
PPPPNNNNXXXX -> PPPPNNNNNNNN
to:
AAAA
PPPPNNNNXXXX -> PPPPXXXXXXXX
XXXX has the right vma properties for the whole merged vma returned by
vma_adjust, so it solves the problem fully. It has the added benefits
that the callers could stop updating vma properties when vma_merge
succeeds however the callers are not updated by this patch (there are
bits like VM_SOFTDIRTY that still need special care for the whole range,
as the vma merging ignores them, but as long as they're not processed by
rmap walks and instead they're accessed with the mmap_sem at least for
reading, they are fine not to be updated within vma_adjust before
releasing the rmap_locks).
Link: http://lkml.kernel.org/r/1474309513-20313-1-git-send-email-aarcange@redhat.com
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Aditya Mandaleeka <adityam@microsoft.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Vorlicek <janvorli@microsoft.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-10-08 08:01:28 +08:00
|
|
|
* region AAAA is never going to extended over XXXX. Instead XXXX must
|
|
|
|
|
* be extended in region AAAA and NNNN must be removed. This way in
|
|
|
|
|
* all cases where vma_merge succeeds, the moment vma_adjust drops the
|
|
|
|
|
* rmap_locks, the properties of the merged vma will be already
|
|
|
|
|
* correct for the whole merged range. Some of those properties like
|
|
|
|
|
* vm_page_prot/vm_flags may be accessed by rmap_walks and they must
|
|
|
|
|
* be correct for the whole merged range immediately after the
|
|
|
|
|
* rmap_locks are released. Otherwise if XXXX would be removed and
|
|
|
|
|
* NNNN would be extended over the XXXX range, remove_migration_ptes
|
|
|
|
|
* or other rmap walkers (if working on addresses beyond the "end"
|
|
|
|
|
* parameter) may establish ptes with the wrong permissions of NNNN
|
|
|
|
|
* instead of the right permissions of XXXX.
|
2005-04-17 06:20:36 +08:00
|
|
|
*/
|
|
|
|
|
struct vm_area_struct *vma_merge(struct mm_struct *mm,
|
|
|
|
|
struct vm_area_struct *prev, unsigned long addr,
|
|
|
|
|
unsigned long end, unsigned long vm_flags,
|
2014-10-10 06:26:29 +08:00
|
|
|
struct anon_vma *anon_vma, struct file *file,
|
2015-09-05 06:46:24 +08:00
|
|
|
pgoff_t pgoff, struct mempolicy *policy,
|
mm: add a field to store names for private anonymous memory
In many userspace applications, and especially in VM based applications
like Android uses heavily, there are multiple different allocators in
use. At a minimum there is libc malloc and the stack, and in many cases
there are libc malloc, the stack, direct syscalls to mmap anonymous
memory, and multiple VM heaps (one for small objects, one for big
objects, etc.). Each of these layers usually has its own tools to
inspect its usage; malloc by compiling a debug version, the VM through
heap inspection tools, and for direct syscalls there is usually no way
to track them.
On Android we heavily use a set of tools that use an extended version of
the logic covered in Documentation/vm/pagemap.txt to walk all pages
mapped in userspace and slice their usage by process, shared (COW) vs.
unique mappings, backing, etc. This can account for real physical
memory usage even in cases like fork without exec (which Android uses
heavily to share as many private COW pages as possible between
processes), Kernel SamePage Merging, and clean zero pages. It produces
a measurement of the pages that only exist in that process (USS, for
unique), and a measurement of the physical memory usage of that process
with the cost of shared pages being evenly split between processes that
share them (PSS).
If all anonymous memory is indistinguishable then figuring out the real
physical memory usage (PSS) of each heap requires either a pagemap
walking tool that can understand the heap debugging of every layer, or
for every layer's heap debugging tools to implement the pagemap walking
logic, in which case it is hard to get a consistent view of memory
across the whole system.
Tracking the information in userspace leads to all sorts of problems.
It either needs to be stored inside the process, which means every
process has to have an API to export its current heap information upon
request, or it has to be stored externally in a filesystem that somebody
needs to clean up on crashes. It needs to be readable while the process
is still running, so it has to have some sort of synchronization with
every layer of userspace. Efficiently tracking the ranges requires
reimplementing something like the kernel vma trees, and linking to it
from every layer of userspace. It requires more memory, more syscalls,
more runtime cost, and more complexity to separately track regions that
the kernel is already tracking.
This patch adds a field to /proc/pid/maps and /proc/pid/smaps to show a
userspace-provided name for anonymous vmas. The names of named
anonymous vmas are shown in /proc/pid/maps and /proc/pid/smaps as
[anon:<name>].
Userspace can set the name for a region of memory by calling
prctl(PR_SET_VMA, PR_SET_VMA_ANON_NAME, start, len, (unsigned long)name)
Setting the name to NULL clears it. The name length limit is 80 bytes
including NUL-terminator and is checked to contain only printable ascii
characters (including space), except '[',']','\','$' and '`'.
Ascii strings are being used to have a descriptive identifiers for vmas,
which can be understood by the users reading /proc/pid/maps or
/proc/pid/smaps. Names can be standardized for a given system and they
can include some variable parts such as the name of the allocator or a
library, tid of the thread using it, etc.
The name is stored in a pointer in the shared union in vm_area_struct
that points to a null terminated string. Anonymous vmas with the same
name (equivalent strings) and are otherwise mergeable will be merged.
The name pointers are not shared between vmas even if they contain the
same name. The name pointer is stored in a union with fields that are
only used on file-backed mappings, so it does not increase memory usage.
CONFIG_ANON_VMA_NAME kernel configuration is introduced to enable this
feature. It keeps the feature disabled by default to prevent any
additional memory overhead and to avoid confusing procfs parsers on
systems which are not ready to support named anonymous vmas.
The patch is based on the original patch developed by Colin Cross, more
specifically on its latest version [1] posted upstream by Sumit Semwal.
It used a userspace pointer to store vma names. In that design, name
pointers could be shared between vmas. However during the last
upstreaming attempt, Kees Cook raised concerns [2] about this approach
and suggested to copy the name into kernel memory space, perform
validity checks [3] and store as a string referenced from
vm_area_struct.
One big concern is about fork() performance which would need to strdup
anonymous vma names. Dave Hansen suggested experimenting with
worst-case scenario of forking a process with 64k vmas having longest
possible names [4]. I ran this experiment on an ARM64 Android device
and recorded a worst-case regression of almost 40% when forking such a
process.
This regression is addressed in the followup patch which replaces the
pointer to a name with a refcounted structure that allows sharing the
name pointer between vmas of the same name. Instead of duplicating the
string during fork() or when splitting a vma it increments the refcount.
[1] https://lore.kernel.org/linux-mm/20200901161459.11772-4-sumit.semwal@linaro.org/
[2] https://lore.kernel.org/linux-mm/202009031031.D32EF57ED@keescook/
[3] https://lore.kernel.org/linux-mm/202009031022.3834F692@keescook/
[4] https://lore.kernel.org/linux-mm/5d0358ab-8c47-2f5f-8e43-23b89d6a8e95@intel.com/
Changes for prctl(2) manual page (in the options section):
PR_SET_VMA
Sets an attribute specified in arg2 for virtual memory areas
starting from the address specified in arg3 and spanning the
size specified in arg4. arg5 specifies the value of the attribute
to be set. Note that assigning an attribute to a virtual memory
area might prevent it from being merged with adjacent virtual
memory areas due to the difference in that attribute's value.
Currently, arg2 must be one of:
PR_SET_VMA_ANON_NAME
Set a name for anonymous virtual memory areas. arg5 should
be a pointer to a null-terminated string containing the
name. The name length including null byte cannot exceed
80 bytes. If arg5 is NULL, the name of the appropriate
anonymous virtual memory areas will be reset. The name
can contain only printable ascii characters (including
space), except '[',']','\','$' and '`'.
This feature is available only if the kernel is built with
the CONFIG_ANON_VMA_NAME option enabled.
[surenb@google.com: docs: proc.rst: /proc/PID/maps: fix malformed table]
Link: https://lkml.kernel.org/r/20211123185928.2513763-1-surenb@google.com
[surenb: rebased over v5.15-rc6, replaced userpointer with a kernel copy,
added input sanitization and CONFIG_ANON_VMA_NAME config. The bulk of the
work here was done by Colin Cross, therefore, with his permission, keeping
him as the author]
Link: https://lkml.kernel.org/r/20211019215511.3771969-2-surenb@google.com
Signed-off-by: Colin Cross <ccross@google.com>
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Kees Cook <keescook@chromium.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Jan Glauber <jan.glauber@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: John Stultz <john.stultz@linaro.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rob Landley <rob@landley.net>
Cc: "Serge E. Hallyn" <serge.hallyn@ubuntu.com>
Cc: Shaohua Li <shli@fusionio.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-15 06:05:59 +08:00
|
|
|
struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
|
2022-03-05 12:28:51 +08:00
|
|
|
struct anon_vma_name *anon_name)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
|
|
|
|
pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
|
|
|
|
|
struct vm_area_struct *area, *next;
|
mm: change anon_vma linking to fix multi-process server scalability issue
The old anon_vma code can lead to scalability issues with heavily forking
workloads. Specifically, each anon_vma will be shared between the parent
process and all its child processes.
In a workload with 1000 child processes and a VMA with 1000 anonymous
pages per process that get COWed, this leads to a system with a million
anonymous pages in the same anon_vma, each of which is mapped in just one
of the 1000 processes. However, the current rmap code needs to walk them
all, leading to O(N) scanning complexity for each page.
This can result in systems where one CPU is walking the page tables of
1000 processes in page_referenced_one, while all other CPUs are stuck on
the anon_vma lock. This leads to catastrophic failure for a benchmark
like AIM7, where the total number of processes can reach in the tens of
thousands. Real workloads are still a factor 10 less process intensive
than AIM7, but they are catching up.
This patch changes the way anon_vmas and VMAs are linked, which allows us
to associate multiple anon_vmas with a VMA. At fork time, each child
process gets its own anon_vmas, in which its COWed pages will be
instantiated. The parents' anon_vma is also linked to the VMA, because
non-COWed pages could be present in any of the children.
This reduces rmap scanning complexity to O(1) for the pages of the 1000
child processes, with O(N) complexity for at most 1/N pages in the system.
This reduces the average scanning cost in heavily forking workloads from
O(N) to 2.
The only real complexity in this patch stems from the fact that linking a
VMA to anon_vmas now involves memory allocations. This means vma_adjust
can fail, if it needs to attach a VMA to anon_vma structures. This in
turn means error handling needs to be added to the calling functions.
A second source of complexity is that, because there can be multiple
anon_vmas, the anon_vma linking in vma_adjust can no longer be done under
"the" anon_vma lock. To prevent the rmap code from walking up an
incomplete VMA, this patch introduces the VM_LOCK_RMAP VMA flag. This bit
flag uses the same slot as the NOMMU VM_MAPPED_COPY, with an ifdef in mm.h
to make sure it is impossible to compile a kernel that needs both symbolic
values for the same bitflag.
Some test results:
Without the anon_vma changes, when AIM7 hits around 9.7k users (on a test
box with 16GB RAM and not quite enough IO), the system ends up running
>99% in system time, with every CPU on the same anon_vma lock in the
pageout code.
With these changes, AIM7 hits the cross-over point around 29.7k users.
This happens with ~99% IO wait time, there never seems to be any spike in
system time. The anon_vma lock contention appears to be resolved.
[akpm@linux-foundation.org: cleanups]
Signed-off-by: Rik van Riel <riel@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Larry Woodman <lwoodman@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-06 05:42:07 +08:00
|
|
|
int err;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* We later require that vma->vm_flags == vm_flags,
|
|
|
|
|
* so this tests vma->vm_flags & VM_SPECIAL, too.
|
|
|
|
|
*/
|
|
|
|
|
if (vm_flags & VM_SPECIAL)
|
|
|
|
|
return NULL;
|
|
|
|
|
|
2020-10-18 07:14:06 +08:00
|
|
|
next = vma_next(mm, prev);
|
2005-04-17 06:20:36 +08:00
|
|
|
area = next;
|
mm: vma_merge: fix vm_page_prot SMP race condition against rmap_walk
The rmap_walk can access vm_page_prot (and potentially vm_flags in the
pte/pmd manipulations). So it's not safe to wait the caller to update
the vm_page_prot/vm_flags after vma_merge returned potentially removing
the "next" vma and extending the "current" vma over the
next->vm_start,vm_end range, but still with the "current" vma
vm_page_prot, after releasing the rmap locks.
The vm_page_prot/vm_flags must be transferred from the "next" vma to the
current vma while vma_merge still holds the rmap locks.
The side effect of this race condition is pte corruption during migrate
as remove_migration_ptes when run on a address of the "next" vma that
got removed, used the vm_page_prot of the current vma.
migrate mprotect
------------ -------------
migrating in "next" vma
vma_merge() # removes "next" vma and
# extends "current" vma
# current vma is not with
# vm_page_prot updated
remove_migration_ptes
read vm_page_prot of current "vma"
establish pte with wrong permissions
vm_set_page_prot(vma) # too late!
change_protection in the old vma range
only, next range is not updated
This caused segmentation faults and potentially memory corruption in
heavy mprotect loads with some light page migration caused by compaction
in the background.
Hugh Dickins pointed out the comment about the Odd case 8 in vma_merge
which confirms the case 8 is only buggy one where the race can trigger,
in all other vma_merge cases the above cannot happen.
This fix removes the oddness factor from case 8 and it converts it from:
AAAA
PPPPNNNNXXXX -> PPPPNNNNNNNN
to:
AAAA
PPPPNNNNXXXX -> PPPPXXXXXXXX
XXXX has the right vma properties for the whole merged vma returned by
vma_adjust, so it solves the problem fully. It has the added benefits
that the callers could stop updating vma properties when vma_merge
succeeds however the callers are not updated by this patch (there are
bits like VM_SOFTDIRTY that still need special care for the whole range,
as the vma merging ignores them, but as long as they're not processed by
rmap walks and instead they're accessed with the mmap_sem at least for
reading, they are fine not to be updated within vma_adjust before
releasing the rmap_locks).
Link: http://lkml.kernel.org/r/1474309513-20313-1-git-send-email-aarcange@redhat.com
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Aditya Mandaleeka <adityam@microsoft.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Vorlicek <janvorli@microsoft.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-10-08 08:01:28 +08:00
|
|
|
if (area && area->vm_end == end) /* cases 6, 7, 8 */
|
2005-04-17 06:20:36 +08:00
|
|
|
next = next->vm_next;
|
|
|
|
|
|
mm: vma_merge: fix vm_page_prot SMP race condition against rmap_walk
The rmap_walk can access vm_page_prot (and potentially vm_flags in the
pte/pmd manipulations). So it's not safe to wait the caller to update
the vm_page_prot/vm_flags after vma_merge returned potentially removing
the "next" vma and extending the "current" vma over the
next->vm_start,vm_end range, but still with the "current" vma
vm_page_prot, after releasing the rmap locks.
The vm_page_prot/vm_flags must be transferred from the "next" vma to the
current vma while vma_merge still holds the rmap locks.
The side effect of this race condition is pte corruption during migrate
as remove_migration_ptes when run on a address of the "next" vma that
got removed, used the vm_page_prot of the current vma.
migrate mprotect
------------ -------------
migrating in "next" vma
vma_merge() # removes "next" vma and
# extends "current" vma
# current vma is not with
# vm_page_prot updated
remove_migration_ptes
read vm_page_prot of current "vma"
establish pte with wrong permissions
vm_set_page_prot(vma) # too late!
change_protection in the old vma range
only, next range is not updated
This caused segmentation faults and potentially memory corruption in
heavy mprotect loads with some light page migration caused by compaction
in the background.
Hugh Dickins pointed out the comment about the Odd case 8 in vma_merge
which confirms the case 8 is only buggy one where the race can trigger,
in all other vma_merge cases the above cannot happen.
This fix removes the oddness factor from case 8 and it converts it from:
AAAA
PPPPNNNNXXXX -> PPPPNNNNNNNN
to:
AAAA
PPPPNNNNXXXX -> PPPPXXXXXXXX
XXXX has the right vma properties for the whole merged vma returned by
vma_adjust, so it solves the problem fully. It has the added benefits
that the callers could stop updating vma properties when vma_merge
succeeds however the callers are not updated by this patch (there are
bits like VM_SOFTDIRTY that still need special care for the whole range,
as the vma merging ignores them, but as long as they're not processed by
rmap walks and instead they're accessed with the mmap_sem at least for
reading, they are fine not to be updated within vma_adjust before
releasing the rmap_locks).
Link: http://lkml.kernel.org/r/1474309513-20313-1-git-send-email-aarcange@redhat.com
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Aditya Mandaleeka <adityam@microsoft.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Vorlicek <janvorli@microsoft.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-10-08 08:01:28 +08:00
|
|
|
/* verify some invariant that must be enforced by the caller */
|
|
|
|
|
VM_WARN_ON(prev && addr <= prev->vm_start);
|
|
|
|
|
VM_WARN_ON(area && end > area->vm_end);
|
|
|
|
|
VM_WARN_ON(addr >= end);
|
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
/*
|
|
|
|
|
* Can it merge with the predecessor?
|
|
|
|
|
*/
|
|
|
|
|
if (prev && prev->vm_end == addr &&
|
2014-10-10 06:26:29 +08:00
|
|
|
mpol_equal(vma_policy(prev), policy) &&
|
2005-04-17 06:20:36 +08:00
|
|
|
can_vma_merge_after(prev, vm_flags,
|
2015-09-05 06:46:24 +08:00
|
|
|
anon_vma, file, pgoff,
|
mm: add a field to store names for private anonymous memory
In many userspace applications, and especially in VM based applications
like Android uses heavily, there are multiple different allocators in
use. At a minimum there is libc malloc and the stack, and in many cases
there are libc malloc, the stack, direct syscalls to mmap anonymous
memory, and multiple VM heaps (one for small objects, one for big
objects, etc.). Each of these layers usually has its own tools to
inspect its usage; malloc by compiling a debug version, the VM through
heap inspection tools, and for direct syscalls there is usually no way
to track them.
On Android we heavily use a set of tools that use an extended version of
the logic covered in Documentation/vm/pagemap.txt to walk all pages
mapped in userspace and slice their usage by process, shared (COW) vs.
unique mappings, backing, etc. This can account for real physical
memory usage even in cases like fork without exec (which Android uses
heavily to share as many private COW pages as possible between
processes), Kernel SamePage Merging, and clean zero pages. It produces
a measurement of the pages that only exist in that process (USS, for
unique), and a measurement of the physical memory usage of that process
with the cost of shared pages being evenly split between processes that
share them (PSS).
If all anonymous memory is indistinguishable then figuring out the real
physical memory usage (PSS) of each heap requires either a pagemap
walking tool that can understand the heap debugging of every layer, or
for every layer's heap debugging tools to implement the pagemap walking
logic, in which case it is hard to get a consistent view of memory
across the whole system.
Tracking the information in userspace leads to all sorts of problems.
It either needs to be stored inside the process, which means every
process has to have an API to export its current heap information upon
request, or it has to be stored externally in a filesystem that somebody
needs to clean up on crashes. It needs to be readable while the process
is still running, so it has to have some sort of synchronization with
every layer of userspace. Efficiently tracking the ranges requires
reimplementing something like the kernel vma trees, and linking to it
from every layer of userspace. It requires more memory, more syscalls,
more runtime cost, and more complexity to separately track regions that
the kernel is already tracking.
This patch adds a field to /proc/pid/maps and /proc/pid/smaps to show a
userspace-provided name for anonymous vmas. The names of named
anonymous vmas are shown in /proc/pid/maps and /proc/pid/smaps as
[anon:<name>].
Userspace can set the name for a region of memory by calling
prctl(PR_SET_VMA, PR_SET_VMA_ANON_NAME, start, len, (unsigned long)name)
Setting the name to NULL clears it. The name length limit is 80 bytes
including NUL-terminator and is checked to contain only printable ascii
characters (including space), except '[',']','\','$' and '`'.
Ascii strings are being used to have a descriptive identifiers for vmas,
which can be understood by the users reading /proc/pid/maps or
/proc/pid/smaps. Names can be standardized for a given system and they
can include some variable parts such as the name of the allocator or a
library, tid of the thread using it, etc.
The name is stored in a pointer in the shared union in vm_area_struct
that points to a null terminated string. Anonymous vmas with the same
name (equivalent strings) and are otherwise mergeable will be merged.
The name pointers are not shared between vmas even if they contain the
same name. The name pointer is stored in a union with fields that are
only used on file-backed mappings, so it does not increase memory usage.
CONFIG_ANON_VMA_NAME kernel configuration is introduced to enable this
feature. It keeps the feature disabled by default to prevent any
additional memory overhead and to avoid confusing procfs parsers on
systems which are not ready to support named anonymous vmas.
The patch is based on the original patch developed by Colin Cross, more
specifically on its latest version [1] posted upstream by Sumit Semwal.
It used a userspace pointer to store vma names. In that design, name
pointers could be shared between vmas. However during the last
upstreaming attempt, Kees Cook raised concerns [2] about this approach
and suggested to copy the name into kernel memory space, perform
validity checks [3] and store as a string referenced from
vm_area_struct.
One big concern is about fork() performance which would need to strdup
anonymous vma names. Dave Hansen suggested experimenting with
worst-case scenario of forking a process with 64k vmas having longest
possible names [4]. I ran this experiment on an ARM64 Android device
and recorded a worst-case regression of almost 40% when forking such a
process.
This regression is addressed in the followup patch which replaces the
pointer to a name with a refcounted structure that allows sharing the
name pointer between vmas of the same name. Instead of duplicating the
string during fork() or when splitting a vma it increments the refcount.
[1] https://lore.kernel.org/linux-mm/20200901161459.11772-4-sumit.semwal@linaro.org/
[2] https://lore.kernel.org/linux-mm/202009031031.D32EF57ED@keescook/
[3] https://lore.kernel.org/linux-mm/202009031022.3834F692@keescook/
[4] https://lore.kernel.org/linux-mm/5d0358ab-8c47-2f5f-8e43-23b89d6a8e95@intel.com/
Changes for prctl(2) manual page (in the options section):
PR_SET_VMA
Sets an attribute specified in arg2 for virtual memory areas
starting from the address specified in arg3 and spanning the
size specified in arg4. arg5 specifies the value of the attribute
to be set. Note that assigning an attribute to a virtual memory
area might prevent it from being merged with adjacent virtual
memory areas due to the difference in that attribute's value.
Currently, arg2 must be one of:
PR_SET_VMA_ANON_NAME
Set a name for anonymous virtual memory areas. arg5 should
be a pointer to a null-terminated string containing the
name. The name length including null byte cannot exceed
80 bytes. If arg5 is NULL, the name of the appropriate
anonymous virtual memory areas will be reset. The name
can contain only printable ascii characters (including
space), except '[',']','\','$' and '`'.
This feature is available only if the kernel is built with
the CONFIG_ANON_VMA_NAME option enabled.
[surenb@google.com: docs: proc.rst: /proc/PID/maps: fix malformed table]
Link: https://lkml.kernel.org/r/20211123185928.2513763-1-surenb@google.com
[surenb: rebased over v5.15-rc6, replaced userpointer with a kernel copy,
added input sanitization and CONFIG_ANON_VMA_NAME config. The bulk of the
work here was done by Colin Cross, therefore, with his permission, keeping
him as the author]
Link: https://lkml.kernel.org/r/20211019215511.3771969-2-surenb@google.com
Signed-off-by: Colin Cross <ccross@google.com>
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Kees Cook <keescook@chromium.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Jan Glauber <jan.glauber@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: John Stultz <john.stultz@linaro.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rob Landley <rob@landley.net>
Cc: "Serge E. Hallyn" <serge.hallyn@ubuntu.com>
Cc: Shaohua Li <shli@fusionio.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-15 06:05:59 +08:00
|
|
|
vm_userfaultfd_ctx, anon_name)) {
|
2005-04-17 06:20:36 +08:00
|
|
|
/*
|
|
|
|
|
* OK, it can. Can we now merge in the successor as well?
|
|
|
|
|
*/
|
|
|
|
|
if (next && end == next->vm_start &&
|
|
|
|
|
mpol_equal(policy, vma_policy(next)) &&
|
|
|
|
|
can_vma_merge_before(next, vm_flags,
|
2015-09-05 06:46:24 +08:00
|
|
|
anon_vma, file,
|
|
|
|
|
pgoff+pglen,
|
mm: add a field to store names for private anonymous memory
In many userspace applications, and especially in VM based applications
like Android uses heavily, there are multiple different allocators in
use. At a minimum there is libc malloc and the stack, and in many cases
there are libc malloc, the stack, direct syscalls to mmap anonymous
memory, and multiple VM heaps (one for small objects, one for big
objects, etc.). Each of these layers usually has its own tools to
inspect its usage; malloc by compiling a debug version, the VM through
heap inspection tools, and for direct syscalls there is usually no way
to track them.
On Android we heavily use a set of tools that use an extended version of
the logic covered in Documentation/vm/pagemap.txt to walk all pages
mapped in userspace and slice their usage by process, shared (COW) vs.
unique mappings, backing, etc. This can account for real physical
memory usage even in cases like fork without exec (which Android uses
heavily to share as many private COW pages as possible between
processes), Kernel SamePage Merging, and clean zero pages. It produces
a measurement of the pages that only exist in that process (USS, for
unique), and a measurement of the physical memory usage of that process
with the cost of shared pages being evenly split between processes that
share them (PSS).
If all anonymous memory is indistinguishable then figuring out the real
physical memory usage (PSS) of each heap requires either a pagemap
walking tool that can understand the heap debugging of every layer, or
for every layer's heap debugging tools to implement the pagemap walking
logic, in which case it is hard to get a consistent view of memory
across the whole system.
Tracking the information in userspace leads to all sorts of problems.
It either needs to be stored inside the process, which means every
process has to have an API to export its current heap information upon
request, or it has to be stored externally in a filesystem that somebody
needs to clean up on crashes. It needs to be readable while the process
is still running, so it has to have some sort of synchronization with
every layer of userspace. Efficiently tracking the ranges requires
reimplementing something like the kernel vma trees, and linking to it
from every layer of userspace. It requires more memory, more syscalls,
more runtime cost, and more complexity to separately track regions that
the kernel is already tracking.
This patch adds a field to /proc/pid/maps and /proc/pid/smaps to show a
userspace-provided name for anonymous vmas. The names of named
anonymous vmas are shown in /proc/pid/maps and /proc/pid/smaps as
[anon:<name>].
Userspace can set the name for a region of memory by calling
prctl(PR_SET_VMA, PR_SET_VMA_ANON_NAME, start, len, (unsigned long)name)
Setting the name to NULL clears it. The name length limit is 80 bytes
including NUL-terminator and is checked to contain only printable ascii
characters (including space), except '[',']','\','$' and '`'.
Ascii strings are being used to have a descriptive identifiers for vmas,
which can be understood by the users reading /proc/pid/maps or
/proc/pid/smaps. Names can be standardized for a given system and they
can include some variable parts such as the name of the allocator or a
library, tid of the thread using it, etc.
The name is stored in a pointer in the shared union in vm_area_struct
that points to a null terminated string. Anonymous vmas with the same
name (equivalent strings) and are otherwise mergeable will be merged.
The name pointers are not shared between vmas even if they contain the
same name. The name pointer is stored in a union with fields that are
only used on file-backed mappings, so it does not increase memory usage.
CONFIG_ANON_VMA_NAME kernel configuration is introduced to enable this
feature. It keeps the feature disabled by default to prevent any
additional memory overhead and to avoid confusing procfs parsers on
systems which are not ready to support named anonymous vmas.
The patch is based on the original patch developed by Colin Cross, more
specifically on its latest version [1] posted upstream by Sumit Semwal.
It used a userspace pointer to store vma names. In that design, name
pointers could be shared between vmas. However during the last
upstreaming attempt, Kees Cook raised concerns [2] about this approach
and suggested to copy the name into kernel memory space, perform
validity checks [3] and store as a string referenced from
vm_area_struct.
One big concern is about fork() performance which would need to strdup
anonymous vma names. Dave Hansen suggested experimenting with
worst-case scenario of forking a process with 64k vmas having longest
possible names [4]. I ran this experiment on an ARM64 Android device
and recorded a worst-case regression of almost 40% when forking such a
process.
This regression is addressed in the followup patch which replaces the
pointer to a name with a refcounted structure that allows sharing the
name pointer between vmas of the same name. Instead of duplicating the
string during fork() or when splitting a vma it increments the refcount.
[1] https://lore.kernel.org/linux-mm/20200901161459.11772-4-sumit.semwal@linaro.org/
[2] https://lore.kernel.org/linux-mm/202009031031.D32EF57ED@keescook/
[3] https://lore.kernel.org/linux-mm/202009031022.3834F692@keescook/
[4] https://lore.kernel.org/linux-mm/5d0358ab-8c47-2f5f-8e43-23b89d6a8e95@intel.com/
Changes for prctl(2) manual page (in the options section):
PR_SET_VMA
Sets an attribute specified in arg2 for virtual memory areas
starting from the address specified in arg3 and spanning the
size specified in arg4. arg5 specifies the value of the attribute
to be set. Note that assigning an attribute to a virtual memory
area might prevent it from being merged with adjacent virtual
memory areas due to the difference in that attribute's value.
Currently, arg2 must be one of:
PR_SET_VMA_ANON_NAME
Set a name for anonymous virtual memory areas. arg5 should
be a pointer to a null-terminated string containing the
name. The name length including null byte cannot exceed
80 bytes. If arg5 is NULL, the name of the appropriate
anonymous virtual memory areas will be reset. The name
can contain only printable ascii characters (including
space), except '[',']','\','$' and '`'.
This feature is available only if the kernel is built with
the CONFIG_ANON_VMA_NAME option enabled.
[surenb@google.com: docs: proc.rst: /proc/PID/maps: fix malformed table]
Link: https://lkml.kernel.org/r/20211123185928.2513763-1-surenb@google.com
[surenb: rebased over v5.15-rc6, replaced userpointer with a kernel copy,
added input sanitization and CONFIG_ANON_VMA_NAME config. The bulk of the
work here was done by Colin Cross, therefore, with his permission, keeping
him as the author]
Link: https://lkml.kernel.org/r/20211019215511.3771969-2-surenb@google.com
Signed-off-by: Colin Cross <ccross@google.com>
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Kees Cook <keescook@chromium.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Jan Glauber <jan.glauber@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: John Stultz <john.stultz@linaro.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rob Landley <rob@landley.net>
Cc: "Serge E. Hallyn" <serge.hallyn@ubuntu.com>
Cc: Shaohua Li <shli@fusionio.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-15 06:05:59 +08:00
|
|
|
vm_userfaultfd_ctx, anon_name) &&
|
2005-04-17 06:20:36 +08:00
|
|
|
is_mergeable_anon_vma(prev->anon_vma,
|
2011-05-25 08:11:20 +08:00
|
|
|
next->anon_vma, NULL)) {
|
2005-04-17 06:20:36 +08:00
|
|
|
/* cases 1, 6 */
|
mm: vma_merge: fix vm_page_prot SMP race condition against rmap_walk
The rmap_walk can access vm_page_prot (and potentially vm_flags in the
pte/pmd manipulations). So it's not safe to wait the caller to update
the vm_page_prot/vm_flags after vma_merge returned potentially removing
the "next" vma and extending the "current" vma over the
next->vm_start,vm_end range, but still with the "current" vma
vm_page_prot, after releasing the rmap locks.
The vm_page_prot/vm_flags must be transferred from the "next" vma to the
current vma while vma_merge still holds the rmap locks.
The side effect of this race condition is pte corruption during migrate
as remove_migration_ptes when run on a address of the "next" vma that
got removed, used the vm_page_prot of the current vma.
migrate mprotect
------------ -------------
migrating in "next" vma
vma_merge() # removes "next" vma and
# extends "current" vma
# current vma is not with
# vm_page_prot updated
remove_migration_ptes
read vm_page_prot of current "vma"
establish pte with wrong permissions
vm_set_page_prot(vma) # too late!
change_protection in the old vma range
only, next range is not updated
This caused segmentation faults and potentially memory corruption in
heavy mprotect loads with some light page migration caused by compaction
in the background.
Hugh Dickins pointed out the comment about the Odd case 8 in vma_merge
which confirms the case 8 is only buggy one where the race can trigger,
in all other vma_merge cases the above cannot happen.
This fix removes the oddness factor from case 8 and it converts it from:
AAAA
PPPPNNNNXXXX -> PPPPNNNNNNNN
to:
AAAA
PPPPNNNNXXXX -> PPPPXXXXXXXX
XXXX has the right vma properties for the whole merged vma returned by
vma_adjust, so it solves the problem fully. It has the added benefits
that the callers could stop updating vma properties when vma_merge
succeeds however the callers are not updated by this patch (there are
bits like VM_SOFTDIRTY that still need special care for the whole range,
as the vma merging ignores them, but as long as they're not processed by
rmap walks and instead they're accessed with the mmap_sem at least for
reading, they are fine not to be updated within vma_adjust before
releasing the rmap_locks).
Link: http://lkml.kernel.org/r/1474309513-20313-1-git-send-email-aarcange@redhat.com
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Aditya Mandaleeka <adityam@microsoft.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Vorlicek <janvorli@microsoft.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-10-08 08:01:28 +08:00
|
|
|
err = __vma_adjust(prev, prev->vm_start,
|
|
|
|
|
next->vm_end, prev->vm_pgoff, NULL,
|
|
|
|
|
prev);
|
2005-04-17 06:20:36 +08:00
|
|
|
} else /* cases 2, 5, 7 */
|
mm: vma_merge: fix vm_page_prot SMP race condition against rmap_walk
The rmap_walk can access vm_page_prot (and potentially vm_flags in the
pte/pmd manipulations). So it's not safe to wait the caller to update
the vm_page_prot/vm_flags after vma_merge returned potentially removing
the "next" vma and extending the "current" vma over the
next->vm_start,vm_end range, but still with the "current" vma
vm_page_prot, after releasing the rmap locks.
The vm_page_prot/vm_flags must be transferred from the "next" vma to the
current vma while vma_merge still holds the rmap locks.
The side effect of this race condition is pte corruption during migrate
as remove_migration_ptes when run on a address of the "next" vma that
got removed, used the vm_page_prot of the current vma.
migrate mprotect
------------ -------------
migrating in "next" vma
vma_merge() # removes "next" vma and
# extends "current" vma
# current vma is not with
# vm_page_prot updated
remove_migration_ptes
read vm_page_prot of current "vma"
establish pte with wrong permissions
vm_set_page_prot(vma) # too late!
change_protection in the old vma range
only, next range is not updated
This caused segmentation faults and potentially memory corruption in
heavy mprotect loads with some light page migration caused by compaction
in the background.
Hugh Dickins pointed out the comment about the Odd case 8 in vma_merge
which confirms the case 8 is only buggy one where the race can trigger,
in all other vma_merge cases the above cannot happen.
This fix removes the oddness factor from case 8 and it converts it from:
AAAA
PPPPNNNNXXXX -> PPPPNNNNNNNN
to:
AAAA
PPPPNNNNXXXX -> PPPPXXXXXXXX
XXXX has the right vma properties for the whole merged vma returned by
vma_adjust, so it solves the problem fully. It has the added benefits
that the callers could stop updating vma properties when vma_merge
succeeds however the callers are not updated by this patch (there are
bits like VM_SOFTDIRTY that still need special care for the whole range,
as the vma merging ignores them, but as long as they're not processed by
rmap walks and instead they're accessed with the mmap_sem at least for
reading, they are fine not to be updated within vma_adjust before
releasing the rmap_locks).
Link: http://lkml.kernel.org/r/1474309513-20313-1-git-send-email-aarcange@redhat.com
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Aditya Mandaleeka <adityam@microsoft.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Vorlicek <janvorli@microsoft.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-10-08 08:01:28 +08:00
|
|
|
err = __vma_adjust(prev, prev->vm_start,
|
|
|
|
|
end, prev->vm_pgoff, NULL, prev);
|
mm: change anon_vma linking to fix multi-process server scalability issue
The old anon_vma code can lead to scalability issues with heavily forking
workloads. Specifically, each anon_vma will be shared between the parent
process and all its child processes.
In a workload with 1000 child processes and a VMA with 1000 anonymous
pages per process that get COWed, this leads to a system with a million
anonymous pages in the same anon_vma, each of which is mapped in just one
of the 1000 processes. However, the current rmap code needs to walk them
all, leading to O(N) scanning complexity for each page.
This can result in systems where one CPU is walking the page tables of
1000 processes in page_referenced_one, while all other CPUs are stuck on
the anon_vma lock. This leads to catastrophic failure for a benchmark
like AIM7, where the total number of processes can reach in the tens of
thousands. Real workloads are still a factor 10 less process intensive
than AIM7, but they are catching up.
This patch changes the way anon_vmas and VMAs are linked, which allows us
to associate multiple anon_vmas with a VMA. At fork time, each child
process gets its own anon_vmas, in which its COWed pages will be
instantiated. The parents' anon_vma is also linked to the VMA, because
non-COWed pages could be present in any of the children.
This reduces rmap scanning complexity to O(1) for the pages of the 1000
child processes, with O(N) complexity for at most 1/N pages in the system.
This reduces the average scanning cost in heavily forking workloads from
O(N) to 2.
The only real complexity in this patch stems from the fact that linking a
VMA to anon_vmas now involves memory allocations. This means vma_adjust
can fail, if it needs to attach a VMA to anon_vma structures. This in
turn means error handling needs to be added to the calling functions.
A second source of complexity is that, because there can be multiple
anon_vmas, the anon_vma linking in vma_adjust can no longer be done under
"the" anon_vma lock. To prevent the rmap code from walking up an
incomplete VMA, this patch introduces the VM_LOCK_RMAP VMA flag. This bit
flag uses the same slot as the NOMMU VM_MAPPED_COPY, with an ifdef in mm.h
to make sure it is impossible to compile a kernel that needs both symbolic
values for the same bitflag.
Some test results:
Without the anon_vma changes, when AIM7 hits around 9.7k users (on a test
box with 16GB RAM and not quite enough IO), the system ends up running
>99% in system time, with every CPU on the same anon_vma lock in the
pageout code.
With these changes, AIM7 hits the cross-over point around 29.7k users.
This happens with ~99% IO wait time, there never seems to be any spike in
system time. The anon_vma lock contention appears to be resolved.
[akpm@linux-foundation.org: cleanups]
Signed-off-by: Rik van Riel <riel@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Larry Woodman <lwoodman@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-06 05:42:07 +08:00
|
|
|
if (err)
|
|
|
|
|
return NULL;
|
2014-10-30 05:50:31 +08:00
|
|
|
khugepaged_enter_vma_merge(prev, vm_flags);
|
2005-04-17 06:20:36 +08:00
|
|
|
return prev;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Can this new request be merged in front of next?
|
|
|
|
|
*/
|
|
|
|
|
if (next && end == next->vm_start &&
|
2014-10-10 06:26:29 +08:00
|
|
|
mpol_equal(policy, vma_policy(next)) &&
|
2005-04-17 06:20:36 +08:00
|
|
|
can_vma_merge_before(next, vm_flags,
|
2015-09-05 06:46:24 +08:00
|
|
|
anon_vma, file, pgoff+pglen,
|
mm: add a field to store names for private anonymous memory
In many userspace applications, and especially in VM based applications
like Android uses heavily, there are multiple different allocators in
use. At a minimum there is libc malloc and the stack, and in many cases
there are libc malloc, the stack, direct syscalls to mmap anonymous
memory, and multiple VM heaps (one for small objects, one for big
objects, etc.). Each of these layers usually has its own tools to
inspect its usage; malloc by compiling a debug version, the VM through
heap inspection tools, and for direct syscalls there is usually no way
to track them.
On Android we heavily use a set of tools that use an extended version of
the logic covered in Documentation/vm/pagemap.txt to walk all pages
mapped in userspace and slice their usage by process, shared (COW) vs.
unique mappings, backing, etc. This can account for real physical
memory usage even in cases like fork without exec (which Android uses
heavily to share as many private COW pages as possible between
processes), Kernel SamePage Merging, and clean zero pages. It produces
a measurement of the pages that only exist in that process (USS, for
unique), and a measurement of the physical memory usage of that process
with the cost of shared pages being evenly split between processes that
share them (PSS).
If all anonymous memory is indistinguishable then figuring out the real
physical memory usage (PSS) of each heap requires either a pagemap
walking tool that can understand the heap debugging of every layer, or
for every layer's heap debugging tools to implement the pagemap walking
logic, in which case it is hard to get a consistent view of memory
across the whole system.
Tracking the information in userspace leads to all sorts of problems.
It either needs to be stored inside the process, which means every
process has to have an API to export its current heap information upon
request, or it has to be stored externally in a filesystem that somebody
needs to clean up on crashes. It needs to be readable while the process
is still running, so it has to have some sort of synchronization with
every layer of userspace. Efficiently tracking the ranges requires
reimplementing something like the kernel vma trees, and linking to it
from every layer of userspace. It requires more memory, more syscalls,
more runtime cost, and more complexity to separately track regions that
the kernel is already tracking.
This patch adds a field to /proc/pid/maps and /proc/pid/smaps to show a
userspace-provided name for anonymous vmas. The names of named
anonymous vmas are shown in /proc/pid/maps and /proc/pid/smaps as
[anon:<name>].
Userspace can set the name for a region of memory by calling
prctl(PR_SET_VMA, PR_SET_VMA_ANON_NAME, start, len, (unsigned long)name)
Setting the name to NULL clears it. The name length limit is 80 bytes
including NUL-terminator and is checked to contain only printable ascii
characters (including space), except '[',']','\','$' and '`'.
Ascii strings are being used to have a descriptive identifiers for vmas,
which can be understood by the users reading /proc/pid/maps or
/proc/pid/smaps. Names can be standardized for a given system and they
can include some variable parts such as the name of the allocator or a
library, tid of the thread using it, etc.
The name is stored in a pointer in the shared union in vm_area_struct
that points to a null terminated string. Anonymous vmas with the same
name (equivalent strings) and are otherwise mergeable will be merged.
The name pointers are not shared between vmas even if they contain the
same name. The name pointer is stored in a union with fields that are
only used on file-backed mappings, so it does not increase memory usage.
CONFIG_ANON_VMA_NAME kernel configuration is introduced to enable this
feature. It keeps the feature disabled by default to prevent any
additional memory overhead and to avoid confusing procfs parsers on
systems which are not ready to support named anonymous vmas.
The patch is based on the original patch developed by Colin Cross, more
specifically on its latest version [1] posted upstream by Sumit Semwal.
It used a userspace pointer to store vma names. In that design, name
pointers could be shared between vmas. However during the last
upstreaming attempt, Kees Cook raised concerns [2] about this approach
and suggested to copy the name into kernel memory space, perform
validity checks [3] and store as a string referenced from
vm_area_struct.
One big concern is about fork() performance which would need to strdup
anonymous vma names. Dave Hansen suggested experimenting with
worst-case scenario of forking a process with 64k vmas having longest
possible names [4]. I ran this experiment on an ARM64 Android device
and recorded a worst-case regression of almost 40% when forking such a
process.
This regression is addressed in the followup patch which replaces the
pointer to a name with a refcounted structure that allows sharing the
name pointer between vmas of the same name. Instead of duplicating the
string during fork() or when splitting a vma it increments the refcount.
[1] https://lore.kernel.org/linux-mm/20200901161459.11772-4-sumit.semwal@linaro.org/
[2] https://lore.kernel.org/linux-mm/202009031031.D32EF57ED@keescook/
[3] https://lore.kernel.org/linux-mm/202009031022.3834F692@keescook/
[4] https://lore.kernel.org/linux-mm/5d0358ab-8c47-2f5f-8e43-23b89d6a8e95@intel.com/
Changes for prctl(2) manual page (in the options section):
PR_SET_VMA
Sets an attribute specified in arg2 for virtual memory areas
starting from the address specified in arg3 and spanning the
size specified in arg4. arg5 specifies the value of the attribute
to be set. Note that assigning an attribute to a virtual memory
area might prevent it from being merged with adjacent virtual
memory areas due to the difference in that attribute's value.
Currently, arg2 must be one of:
PR_SET_VMA_ANON_NAME
Set a name for anonymous virtual memory areas. arg5 should
be a pointer to a null-terminated string containing the
name. The name length including null byte cannot exceed
80 bytes. If arg5 is NULL, the name of the appropriate
anonymous virtual memory areas will be reset. The name
can contain only printable ascii characters (including
space), except '[',']','\','$' and '`'.
This feature is available only if the kernel is built with
the CONFIG_ANON_VMA_NAME option enabled.
[surenb@google.com: docs: proc.rst: /proc/PID/maps: fix malformed table]
Link: https://lkml.kernel.org/r/20211123185928.2513763-1-surenb@google.com
[surenb: rebased over v5.15-rc6, replaced userpointer with a kernel copy,
added input sanitization and CONFIG_ANON_VMA_NAME config. The bulk of the
work here was done by Colin Cross, therefore, with his permission, keeping
him as the author]
Link: https://lkml.kernel.org/r/20211019215511.3771969-2-surenb@google.com
Signed-off-by: Colin Cross <ccross@google.com>
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Kees Cook <keescook@chromium.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Jan Glauber <jan.glauber@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: John Stultz <john.stultz@linaro.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rob Landley <rob@landley.net>
Cc: "Serge E. Hallyn" <serge.hallyn@ubuntu.com>
Cc: Shaohua Li <shli@fusionio.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-15 06:05:59 +08:00
|
|
|
vm_userfaultfd_ctx, anon_name)) {
|
2005-04-17 06:20:36 +08:00
|
|
|
if (prev && addr < prev->vm_end) /* case 4 */
|
mm: vma_merge: fix vm_page_prot SMP race condition against rmap_walk
The rmap_walk can access vm_page_prot (and potentially vm_flags in the
pte/pmd manipulations). So it's not safe to wait the caller to update
the vm_page_prot/vm_flags after vma_merge returned potentially removing
the "next" vma and extending the "current" vma over the
next->vm_start,vm_end range, but still with the "current" vma
vm_page_prot, after releasing the rmap locks.
The vm_page_prot/vm_flags must be transferred from the "next" vma to the
current vma while vma_merge still holds the rmap locks.
The side effect of this race condition is pte corruption during migrate
as remove_migration_ptes when run on a address of the "next" vma that
got removed, used the vm_page_prot of the current vma.
migrate mprotect
------------ -------------
migrating in "next" vma
vma_merge() # removes "next" vma and
# extends "current" vma
# current vma is not with
# vm_page_prot updated
remove_migration_ptes
read vm_page_prot of current "vma"
establish pte with wrong permissions
vm_set_page_prot(vma) # too late!
change_protection in the old vma range
only, next range is not updated
This caused segmentation faults and potentially memory corruption in
heavy mprotect loads with some light page migration caused by compaction
in the background.
Hugh Dickins pointed out the comment about the Odd case 8 in vma_merge
which confirms the case 8 is only buggy one where the race can trigger,
in all other vma_merge cases the above cannot happen.
This fix removes the oddness factor from case 8 and it converts it from:
AAAA
PPPPNNNNXXXX -> PPPPNNNNNNNN
to:
AAAA
PPPPNNNNXXXX -> PPPPXXXXXXXX
XXXX has the right vma properties for the whole merged vma returned by
vma_adjust, so it solves the problem fully. It has the added benefits
that the callers could stop updating vma properties when vma_merge
succeeds however the callers are not updated by this patch (there are
bits like VM_SOFTDIRTY that still need special care for the whole range,
as the vma merging ignores them, but as long as they're not processed by
rmap walks and instead they're accessed with the mmap_sem at least for
reading, they are fine not to be updated within vma_adjust before
releasing the rmap_locks).
Link: http://lkml.kernel.org/r/1474309513-20313-1-git-send-email-aarcange@redhat.com
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Aditya Mandaleeka <adityam@microsoft.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Vorlicek <janvorli@microsoft.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-10-08 08:01:28 +08:00
|
|
|
err = __vma_adjust(prev, prev->vm_start,
|
|
|
|
|
addr, prev->vm_pgoff, NULL, next);
|
|
|
|
|
else { /* cases 3, 8 */
|
|
|
|
|
err = __vma_adjust(area, addr, next->vm_end,
|
|
|
|
|
next->vm_pgoff - pglen, NULL, next);
|
|
|
|
|
/*
|
|
|
|
|
* In case 3 area is already equal to next and
|
|
|
|
|
* this is a noop, but in case 8 "area" has
|
|
|
|
|
* been removed and next was expanded over it.
|
|
|
|
|
*/
|
|
|
|
|
area = next;
|
|
|
|
|
}
|
mm: change anon_vma linking to fix multi-process server scalability issue
The old anon_vma code can lead to scalability issues with heavily forking
workloads. Specifically, each anon_vma will be shared between the parent
process and all its child processes.
In a workload with 1000 child processes and a VMA with 1000 anonymous
pages per process that get COWed, this leads to a system with a million
anonymous pages in the same anon_vma, each of which is mapped in just one
of the 1000 processes. However, the current rmap code needs to walk them
all, leading to O(N) scanning complexity for each page.
This can result in systems where one CPU is walking the page tables of
1000 processes in page_referenced_one, while all other CPUs are stuck on
the anon_vma lock. This leads to catastrophic failure for a benchmark
like AIM7, where the total number of processes can reach in the tens of
thousands. Real workloads are still a factor 10 less process intensive
than AIM7, but they are catching up.
This patch changes the way anon_vmas and VMAs are linked, which allows us
to associate multiple anon_vmas with a VMA. At fork time, each child
process gets its own anon_vmas, in which its COWed pages will be
instantiated. The parents' anon_vma is also linked to the VMA, because
non-COWed pages could be present in any of the children.
This reduces rmap scanning complexity to O(1) for the pages of the 1000
child processes, with O(N) complexity for at most 1/N pages in the system.
This reduces the average scanning cost in heavily forking workloads from
O(N) to 2.
The only real complexity in this patch stems from the fact that linking a
VMA to anon_vmas now involves memory allocations. This means vma_adjust
can fail, if it needs to attach a VMA to anon_vma structures. This in
turn means error handling needs to be added to the calling functions.
A second source of complexity is that, because there can be multiple
anon_vmas, the anon_vma linking in vma_adjust can no longer be done under
"the" anon_vma lock. To prevent the rmap code from walking up an
incomplete VMA, this patch introduces the VM_LOCK_RMAP VMA flag. This bit
flag uses the same slot as the NOMMU VM_MAPPED_COPY, with an ifdef in mm.h
to make sure it is impossible to compile a kernel that needs both symbolic
values for the same bitflag.
Some test results:
Without the anon_vma changes, when AIM7 hits around 9.7k users (on a test
box with 16GB RAM and not quite enough IO), the system ends up running
>99% in system time, with every CPU on the same anon_vma lock in the
pageout code.
With these changes, AIM7 hits the cross-over point around 29.7k users.
This happens with ~99% IO wait time, there never seems to be any spike in
system time. The anon_vma lock contention appears to be resolved.
[akpm@linux-foundation.org: cleanups]
Signed-off-by: Rik van Riel <riel@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Larry Woodman <lwoodman@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-06 05:42:07 +08:00
|
|
|
if (err)
|
|
|
|
|
return NULL;
|
2014-10-30 05:50:31 +08:00
|
|
|
khugepaged_enter_vma_merge(area, vm_flags);
|
2005-04-17 06:20:36 +08:00
|
|
|
return area;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return NULL;
|
|
|
|
|
}
|
|
|
|
|
|
2010-04-11 01:36:19 +08:00
|
|
|
/*
|
2020-06-05 07:49:04 +08:00
|
|
|
* Rough compatibility check to quickly see if it's even worth looking
|
2010-04-11 01:36:19 +08:00
|
|
|
* at sharing an anon_vma.
|
|
|
|
|
*
|
|
|
|
|
* They need to have the same vm_file, and the flags can only differ
|
|
|
|
|
* in things that mprotect may change.
|
|
|
|
|
*
|
|
|
|
|
* NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
|
|
|
|
|
* we can merge the two vma's. For example, we refuse to merge a vma if
|
|
|
|
|
* there is a vm_ops->close() function, because that indicates that the
|
|
|
|
|
* driver is doing some kind of reference counting. But that doesn't
|
|
|
|
|
* really matter for the anon_vma sharing case.
|
|
|
|
|
*/
|
|
|
|
|
static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
|
|
|
|
|
{
|
|
|
|
|
return a->vm_end == b->vm_start &&
|
|
|
|
|
mpol_equal(vma_policy(a), vma_policy(b)) &&
|
|
|
|
|
a->vm_file == b->vm_file &&
|
2020-04-11 05:33:09 +08:00
|
|
|
!((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
|
2010-04-11 01:36:19 +08:00
|
|
|
b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Do some basic sanity checking to see if we can re-use the anon_vma
|
|
|
|
|
* from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
|
|
|
|
|
* the same as 'old', the other will be the new one that is trying
|
|
|
|
|
* to share the anon_vma.
|
|
|
|
|
*
|
|
|
|
|
* NOTE! This runs with mm_sem held for reading, so it is possible that
|
|
|
|
|
* the anon_vma of 'old' is concurrently in the process of being set up
|
|
|
|
|
* by another page fault trying to merge _that_. But that's ok: if it
|
|
|
|
|
* is being set up, that automatically means that it will be a singleton
|
|
|
|
|
* acceptable for merging, so we can do all of this optimistically. But
|
2015-04-16 07:14:08 +08:00
|
|
|
* we do that READ_ONCE() to make sure that we never re-load the pointer.
|
2010-04-11 01:36:19 +08:00
|
|
|
*
|
|
|
|
|
* IOW: that the "list_is_singular()" test on the anon_vma_chain only
|
|
|
|
|
* matters for the 'stable anon_vma' case (ie the thing we want to avoid
|
|
|
|
|
* is to return an anon_vma that is "complex" due to having gone through
|
|
|
|
|
* a fork).
|
|
|
|
|
*
|
|
|
|
|
* We also make sure that the two vma's are compatible (adjacent,
|
|
|
|
|
* and with the same memory policies). That's all stable, even with just
|
|
|
|
|
* a read lock on the mm_sem.
|
|
|
|
|
*/
|
|
|
|
|
static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
|
|
|
|
|
{
|
|
|
|
|
if (anon_vma_compatible(a, b)) {
|
2015-04-16 07:14:08 +08:00
|
|
|
struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
|
2010-04-11 01:36:19 +08:00
|
|
|
|
|
|
|
|
if (anon_vma && list_is_singular(&old->anon_vma_chain))
|
|
|
|
|
return anon_vma;
|
|
|
|
|
}
|
|
|
|
|
return NULL;
|
|
|
|
|
}
|
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
/*
|
|
|
|
|
* find_mergeable_anon_vma is used by anon_vma_prepare, to check
|
|
|
|
|
* neighbouring vmas for a suitable anon_vma, before it goes off
|
|
|
|
|
* to allocate a new anon_vma. It checks because a repetitive
|
|
|
|
|
* sequence of mprotects and faults may otherwise lead to distinct
|
|
|
|
|
* anon_vmas being allocated, preventing vma merge in subsequent
|
|
|
|
|
* mprotect.
|
|
|
|
|
*/
|
|
|
|
|
struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
|
|
|
|
|
{
|
2020-01-31 14:14:51 +08:00
|
|
|
struct anon_vma *anon_vma = NULL;
|
|
|
|
|
|
|
|
|
|
/* Try next first. */
|
|
|
|
|
if (vma->vm_next) {
|
|
|
|
|
anon_vma = reusable_anon_vma(vma->vm_next, vma, vma->vm_next);
|
|
|
|
|
if (anon_vma)
|
|
|
|
|
return anon_vma;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Try prev next. */
|
|
|
|
|
if (vma->vm_prev)
|
|
|
|
|
anon_vma = reusable_anon_vma(vma->vm_prev, vma->vm_prev, vma);
|
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
/*
|
2020-01-31 14:14:51 +08:00
|
|
|
* We might reach here with anon_vma == NULL if we can't find
|
|
|
|
|
* any reusable anon_vma.
|
2005-04-17 06:20:36 +08:00
|
|
|
* There's no absolute need to look only at touching neighbours:
|
|
|
|
|
* we could search further afield for "compatible" anon_vmas.
|
|
|
|
|
* But it would probably just be a waste of time searching,
|
|
|
|
|
* or lead to too many vmas hanging off the same anon_vma.
|
|
|
|
|
* We're trying to allow mprotect remerging later on,
|
|
|
|
|
* not trying to minimize memory used for anon_vmas.
|
|
|
|
|
*/
|
2020-01-31 14:14:51 +08:00
|
|
|
return anon_vma;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
|
2012-02-13 11:58:52 +08:00
|
|
|
/*
|
|
|
|
|
* If a hint addr is less than mmap_min_addr change hint to be as
|
|
|
|
|
* low as possible but still greater than mmap_min_addr
|
|
|
|
|
*/
|
|
|
|
|
static inline unsigned long round_hint_to_min(unsigned long hint)
|
|
|
|
|
{
|
|
|
|
|
hint &= PAGE_MASK;
|
|
|
|
|
if (((void *)hint != NULL) &&
|
|
|
|
|
(hint < mmap_min_addr))
|
|
|
|
|
return PAGE_ALIGN(mmap_min_addr);
|
|
|
|
|
return hint;
|
|
|
|
|
}
|
|
|
|
|
|
mmap: make mlock_future_check() global
Patch series "mm: introduce memfd_secret system call to create "secret" memory areas", v20.
This is an implementation of "secret" mappings backed by a file
descriptor.
The file descriptor backing secret memory mappings is created using a
dedicated memfd_secret system call The desired protection mode for the
memory is configured using flags parameter of the system call. The mmap()
of the file descriptor created with memfd_secret() will create a "secret"
memory mapping. The pages in that mapping will be marked as not present
in the direct map and will be present only in the page table of the owning
mm.
Although normally Linux userspace mappings are protected from other users,
such secret mappings are useful for environments where a hostile tenant is
trying to trick the kernel into giving them access to other tenants
mappings.
It's designed to provide the following protections:
* Enhanced protection (in conjunction with all the other in-kernel
attack prevention systems) against ROP attacks. Seceretmem makes
"simple" ROP insufficient to perform exfiltration, which increases the
required complexity of the attack. Along with other protections like
the kernel stack size limit and address space layout randomization which
make finding gadgets is really hard, absence of any in-kernel primitive
for accessing secret memory means the one gadget ROP attack can't work.
Since the only way to access secret memory is to reconstruct the missing
mapping entry, the attacker has to recover the physical page and insert
a PTE pointing to it in the kernel and then retrieve the contents. That
takes at least three gadgets which is a level of difficulty beyond most
standard attacks.
* Prevent cross-process secret userspace memory exposures. Once the
secret memory is allocated, the user can't accidentally pass it into the
kernel to be transmitted somewhere. The secreremem pages cannot be
accessed via the direct map and they are disallowed in GUP.
* Harden against exploited kernel flaws. In order to access secretmem,
a kernel-side attack would need to either walk the page tables and
create new ones, or spawn a new privileged uiserspace process to perform
secrets exfiltration using ptrace.
In the future the secret mappings may be used as a mean to protect guest
memory in a virtual machine host.
For demonstration of secret memory usage we've created a userspace library
https://git.kernel.org/pub/scm/linux/kernel/git/jejb/secret-memory-preloader.git
that does two things: the first is act as a preloader for openssl to
redirect all the OPENSSL_malloc calls to secret memory meaning any secret
keys get automatically protected this way and the other thing it does is
expose the API to the user who needs it. We anticipate that a lot of the
use cases would be like the openssl one: many toolkits that deal with
secret keys already have special handling for the memory to try to give
them greater protection, so this would simply be pluggable into the
toolkits without any need for user application modification.
Hiding secret memory mappings behind an anonymous file allows usage of the
page cache for tracking pages allocated for the "secret" mappings as well
as using address_space_operations for e.g. page migration callbacks.
The anonymous file may be also used implicitly, like hugetlb files, to
implement mmap(MAP_SECRET) and use the secret memory areas with "native"
mm ABIs in the future.
Removing of the pages from the direct map may cause its fragmentation on
architectures that use large pages to map the physical memory which
affects the system performance. However, the original Kconfig text for
CONFIG_DIRECT_GBPAGES said that gigabyte pages in the direct map "... can
improve the kernel's performance a tiny bit ..." (commit 00d1c5e05736
("x86: add gbpages switches")) and the recent report [1] showed that "...
although 1G mappings are a good default choice, there is no compelling
evidence that it must be the only choice". Hence, it is sufficient to
have secretmem disabled by default with the ability of a system
administrator to enable it at boot time.
In addition, there is also a long term goal to improve management of the
direct map.
[1] https://lore.kernel.org/linux-mm/213b4567-46ce-f116-9cdf-bbd0c884eb3c@linux.intel.com/
This patch (of 7):
It will be used by the upcoming secret memory implementation.
Link: https://lkml.kernel.org/r/20210518072034.31572-1-rppt@kernel.org
Link: https://lkml.kernel.org/r/20210518072034.31572-2-rppt@kernel.org
Signed-off-by: Mike Rapoport <rppt@linux.ibm.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
Acked-by: James Bottomley <James.Bottomley@HansenPartnership.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Christopher Lameter <cl@linux.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Elena Reshetova <elena.reshetova@intel.com>
Cc: Hagen Paul Pfeifer <hagen@jauu.net>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: James Bottomley <jejb@linux.ibm.com>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Palmer Dabbelt <palmer@dabbelt.com>
Cc: Palmer Dabbelt <palmerdabbelt@google.com>
Cc: Paul Walmsley <paul.walmsley@sifive.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rick Edgecombe <rick.p.edgecombe@intel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Tycho Andersen <tycho@tycho.ws>
Cc: Will Deacon <will@kernel.org>
Cc: kernel test robot <lkp@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-07-08 09:07:50 +08:00
|
|
|
int mlock_future_check(struct mm_struct *mm, unsigned long flags,
|
|
|
|
|
unsigned long len)
|
2014-01-22 07:49:15 +08:00
|
|
|
{
|
|
|
|
|
unsigned long locked, lock_limit;
|
|
|
|
|
|
|
|
|
|
/* mlock MCL_FUTURE? */
|
|
|
|
|
if (flags & VM_LOCKED) {
|
|
|
|
|
locked = len >> PAGE_SHIFT;
|
|
|
|
|
locked += mm->locked_vm;
|
|
|
|
|
lock_limit = rlimit(RLIMIT_MEMLOCK);
|
|
|
|
|
lock_limit >>= PAGE_SHIFT;
|
|
|
|
|
if (locked > lock_limit && !capable(CAP_IPC_LOCK))
|
|
|
|
|
return -EAGAIN;
|
|
|
|
|
}
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
mmap: introduce sane default mmap limits
The internal VM "mmap()" interfaces are based on the mmap target doing
everything using page indexes rather than byte offsets, because
traditionally (ie 32-bit) we had the situation that the byte offset
didn't fit in a register. So while the mmap virtual address was limited
by the word size of the architecture, the backing store was not.
So we're basically passing "pgoff" around as a page index, in order to
be able to describe backing store locations that are much bigger than
the word size (think files larger than 4GB etc).
But while this all makes a ton of sense conceptually, we've been dogged
by various drivers that don't really understand this, and internally
work with byte offsets, and then try to work with the page index by
turning it into a byte offset with "pgoff << PAGE_SHIFT".
Which obviously can overflow.
Adding the size of the mapping to it to get the byte offset of the end
of the backing store just exacerbates the problem, and if you then use
this overflow-prone value to check various limits of your device driver
mmap capability, you're just setting yourself up for problems.
The correct thing for drivers to do is to do their limit math in page
indices, the way the interface is designed. Because the generic mmap
code _does_ test that the index doesn't overflow, since that's what the
mmap code really cares about.
HOWEVER.
Finding and fixing various random drivers is a sisyphean task, so let's
just see if we can just make the core mmap() code do the limiting for
us. Realistically, the only "big" backing stores we need to care about
are regular files and block devices, both of which are known to do this
properly, and which have nice well-defined limits for how much data they
can access.
So let's special-case just those two known cases, and then limit other
random mmap users to a backing store that still fits in "unsigned long".
Realistically, that's not much of a limit at all on 64-bit, and on
32-bit architectures the only worry might be the GPU drivers, which can
have big physical address spaces.
To make it possible for drivers like that to say that they are 64-bit
clean, this patch does repurpose the "FMODE_UNSIGNED_OFFSET" bit in the
file flags to allow drivers to mark their file descriptors as safe in
the full 64-bit mmap address space.
[ The timing for doing this is less than optimal, and this should really
go in a merge window. But realistically, this needs wide testing more
than it needs anything else, and being main-line is the only way to do
that.
So the earlier the better, even if it's outside the proper development
cycle - Linus ]
Cc: Kees Cook <keescook@chromium.org>
Cc: Dan Carpenter <dan.carpenter@oracle.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Willy Tarreau <w@1wt.eu>
Cc: Dave Airlie <airlied@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-05-12 00:52:01 +08:00
|
|
|
static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
|
|
|
|
|
{
|
|
|
|
|
if (S_ISREG(inode->i_mode))
|
mmap: relax file size limit for regular files
Commit be83bbf80682 ("mmap: introduce sane default mmap limits") was
introduced to catch problems in various ad-hoc character device drivers
doing mmap and getting the size limits wrong. In the process, it used
"known good" limits for the normal cases of mapping regular files and
block device drivers.
It turns out that the "s_maxbytes" limit was less "known good" than I
thought. In particular, /proc doesn't set it, but exposes one regular
file to mmap: /proc/vmcore. As a result, that file got limited to the
default MAX_INT s_maxbytes value.
This went unnoticed for a while, because apparently the only thing that
needs it is the s390 kernel zfcpdump, but there might be other tools
that use this too.
Vasily suggested just changing s_maxbytes for all of /proc, which isn't
wrong, but makes me nervous at this stage. So instead, just make the
new mmap limit always be MAX_LFS_FILESIZE for regular files, which won't
affect anything else. It wasn't the regular file case I was worried
about.
I'd really prefer for maxsize to have been per-inode, but that is not
how things are today.
Fixes: be83bbf80682 ("mmap: introduce sane default mmap limits")
Reported-by: Vasily Gorbik <gor@linux.ibm.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-05-20 00:29:11 +08:00
|
|
|
return MAX_LFS_FILESIZE;
|
mmap: introduce sane default mmap limits
The internal VM "mmap()" interfaces are based on the mmap target doing
everything using page indexes rather than byte offsets, because
traditionally (ie 32-bit) we had the situation that the byte offset
didn't fit in a register. So while the mmap virtual address was limited
by the word size of the architecture, the backing store was not.
So we're basically passing "pgoff" around as a page index, in order to
be able to describe backing store locations that are much bigger than
the word size (think files larger than 4GB etc).
But while this all makes a ton of sense conceptually, we've been dogged
by various drivers that don't really understand this, and internally
work with byte offsets, and then try to work with the page index by
turning it into a byte offset with "pgoff << PAGE_SHIFT".
Which obviously can overflow.
Adding the size of the mapping to it to get the byte offset of the end
of the backing store just exacerbates the problem, and if you then use
this overflow-prone value to check various limits of your device driver
mmap capability, you're just setting yourself up for problems.
The correct thing for drivers to do is to do their limit math in page
indices, the way the interface is designed. Because the generic mmap
code _does_ test that the index doesn't overflow, since that's what the
mmap code really cares about.
HOWEVER.
Finding and fixing various random drivers is a sisyphean task, so let's
just see if we can just make the core mmap() code do the limiting for
us. Realistically, the only "big" backing stores we need to care about
are regular files and block devices, both of which are known to do this
properly, and which have nice well-defined limits for how much data they
can access.
So let's special-case just those two known cases, and then limit other
random mmap users to a backing store that still fits in "unsigned long".
Realistically, that's not much of a limit at all on 64-bit, and on
32-bit architectures the only worry might be the GPU drivers, which can
have big physical address spaces.
To make it possible for drivers like that to say that they are 64-bit
clean, this patch does repurpose the "FMODE_UNSIGNED_OFFSET" bit in the
file flags to allow drivers to mark their file descriptors as safe in
the full 64-bit mmap address space.
[ The timing for doing this is less than optimal, and this should really
go in a merge window. But realistically, this needs wide testing more
than it needs anything else, and being main-line is the only way to do
that.
So the earlier the better, even if it's outside the proper development
cycle - Linus ]
Cc: Kees Cook <keescook@chromium.org>
Cc: Dan Carpenter <dan.carpenter@oracle.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Willy Tarreau <w@1wt.eu>
Cc: Dave Airlie <airlied@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-05-12 00:52:01 +08:00
|
|
|
|
|
|
|
|
if (S_ISBLK(inode->i_mode))
|
|
|
|
|
return MAX_LFS_FILESIZE;
|
|
|
|
|
|
2019-09-24 06:39:28 +08:00
|
|
|
if (S_ISSOCK(inode->i_mode))
|
|
|
|
|
return MAX_LFS_FILESIZE;
|
|
|
|
|
|
mmap: introduce sane default mmap limits
The internal VM "mmap()" interfaces are based on the mmap target doing
everything using page indexes rather than byte offsets, because
traditionally (ie 32-bit) we had the situation that the byte offset
didn't fit in a register. So while the mmap virtual address was limited
by the word size of the architecture, the backing store was not.
So we're basically passing "pgoff" around as a page index, in order to
be able to describe backing store locations that are much bigger than
the word size (think files larger than 4GB etc).
But while this all makes a ton of sense conceptually, we've been dogged
by various drivers that don't really understand this, and internally
work with byte offsets, and then try to work with the page index by
turning it into a byte offset with "pgoff << PAGE_SHIFT".
Which obviously can overflow.
Adding the size of the mapping to it to get the byte offset of the end
of the backing store just exacerbates the problem, and if you then use
this overflow-prone value to check various limits of your device driver
mmap capability, you're just setting yourself up for problems.
The correct thing for drivers to do is to do their limit math in page
indices, the way the interface is designed. Because the generic mmap
code _does_ test that the index doesn't overflow, since that's what the
mmap code really cares about.
HOWEVER.
Finding and fixing various random drivers is a sisyphean task, so let's
just see if we can just make the core mmap() code do the limiting for
us. Realistically, the only "big" backing stores we need to care about
are regular files and block devices, both of which are known to do this
properly, and which have nice well-defined limits for how much data they
can access.
So let's special-case just those two known cases, and then limit other
random mmap users to a backing store that still fits in "unsigned long".
Realistically, that's not much of a limit at all on 64-bit, and on
32-bit architectures the only worry might be the GPU drivers, which can
have big physical address spaces.
To make it possible for drivers like that to say that they are 64-bit
clean, this patch does repurpose the "FMODE_UNSIGNED_OFFSET" bit in the
file flags to allow drivers to mark their file descriptors as safe in
the full 64-bit mmap address space.
[ The timing for doing this is less than optimal, and this should really
go in a merge window. But realistically, this needs wide testing more
than it needs anything else, and being main-line is the only way to do
that.
So the earlier the better, even if it's outside the proper development
cycle - Linus ]
Cc: Kees Cook <keescook@chromium.org>
Cc: Dan Carpenter <dan.carpenter@oracle.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Willy Tarreau <w@1wt.eu>
Cc: Dave Airlie <airlied@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-05-12 00:52:01 +08:00
|
|
|
/* Special "we do even unsigned file positions" case */
|
|
|
|
|
if (file->f_mode & FMODE_UNSIGNED_OFFSET)
|
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
|
|
/* Yes, random drivers might want more. But I'm tired of buggy drivers */
|
|
|
|
|
return ULONG_MAX;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static inline bool file_mmap_ok(struct file *file, struct inode *inode,
|
|
|
|
|
unsigned long pgoff, unsigned long len)
|
|
|
|
|
{
|
|
|
|
|
u64 maxsize = file_mmap_size_max(file, inode);
|
|
|
|
|
|
|
|
|
|
if (maxsize && len > maxsize)
|
|
|
|
|
return false;
|
|
|
|
|
maxsize -= len;
|
|
|
|
|
if (pgoff > maxsize >> PAGE_SHIFT)
|
|
|
|
|
return false;
|
|
|
|
|
return true;
|
|
|
|
|
}
|
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
/*
|
2020-06-09 12:33:51 +08:00
|
|
|
* The caller must write-lock current->mm->mmap_lock.
|
2005-04-17 06:20:36 +08:00
|
|
|
*/
|
2015-09-10 06:39:29 +08:00
|
|
|
unsigned long do_mmap(struct file *file, unsigned long addr,
|
2005-04-17 06:20:36 +08:00
|
|
|
unsigned long len, unsigned long prot,
|
2020-08-07 14:23:37 +08:00
|
|
|
unsigned long flags, unsigned long pgoff,
|
|
|
|
|
unsigned long *populate, struct list_head *uf)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
2014-10-10 06:26:29 +08:00
|
|
|
struct mm_struct *mm = current->mm;
|
2020-08-07 14:23:37 +08:00
|
|
|
vm_flags_t vm_flags;
|
mm/core, x86/mm/pkeys: Add execute-only protection keys support
Protection keys provide new page-based protection in hardware.
But, they have an interesting attribute: they only affect data
accesses and never affect instruction fetches. That means that
if we set up some memory which is set as "access-disabled" via
protection keys, we can still execute from it.
This patch uses protection keys to set up mappings to do just that.
If a user calls:
mmap(..., PROT_EXEC);
or
mprotect(ptr, sz, PROT_EXEC);
(note PROT_EXEC-only without PROT_READ/WRITE), the kernel will
notice this, and set a special protection key on the memory. It
also sets the appropriate bits in the Protection Keys User Rights
(PKRU) register so that the memory becomes unreadable and
unwritable.
I haven't found any userspace that does this today. With this
facility in place, we expect userspace to move to use it
eventually. Userspace _could_ start doing this today. Any
PROT_EXEC calls get converted to PROT_READ inside the kernel, and
would transparently be upgraded to "true" PROT_EXEC with this
code. IOW, userspace never has to do any PROT_EXEC runtime
detection.
This feature provides enhanced protection against leaking
executable memory contents. This helps thwart attacks which are
attempting to find ROP gadgets on the fly.
But, the security provided by this approach is not comprehensive.
The PKRU register which controls access permissions is a normal
user register writable from unprivileged userspace. An attacker
who can execute the 'wrpkru' instruction can easily disable the
protection provided by this feature.
The protection key that is used for execute-only support is
permanently dedicated at compile time. This is fine for now
because there is currently no API to set a protection key other
than this one.
Despite there being a constant PKRU value across the entire
system, we do not set it unless this feature is in use in a
process. That is to preserve the PKRU XSAVE 'init state',
which can lead to faster context switches.
PKRU *is* a user register and the kernel is modifying it. That
means that code doing:
pkru = rdpkru()
pkru |= 0x100;
mmap(..., PROT_EXEC);
wrpkru(pkru);
could lose the bits in PKRU that enforce execute-only
permissions. To avoid this, we suggest avoiding ever calling
mmap() or mprotect() when the PKRU value is expected to be
unstable.
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Borislav Petkov <bp@suse.de>
Cc: Brian Gerst <brgerst@gmail.com>
Cc: Chen Gang <gang.chen.5i5j@gmail.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Dave Hansen <dave@sr71.net>
Cc: David Hildenbrand <dahi@linux.vnet.ibm.com>
Cc: Denys Vlasenko <dvlasenk@redhat.com>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Kees Cook <keescook@chromium.org>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Konstantin Khlebnikov <koct9i@gmail.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Piotr Kwapulinski <kwapulinski.piotr@gmail.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Stephen Smalley <sds@tycho.nsa.gov>
Cc: Vladimir Murzin <vladimir.murzin@arm.com>
Cc: Will Deacon <will.deacon@arm.com>
Cc: keescook@google.com
Cc: linux-kernel@vger.kernel.org
Cc: linux-mm@kvack.org
Link: http://lkml.kernel.org/r/20160212210240.CB4BB5CA@viggo.jf.intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-02-13 05:02:40 +08:00
|
|
|
int pkey = 0;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2013-02-23 08:32:47 +08:00
|
|
|
*populate = 0;
|
2013-02-23 08:32:37 +08:00
|
|
|
|
2015-06-25 07:58:39 +08:00
|
|
|
if (!len)
|
|
|
|
|
return -EINVAL;
|
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
/*
|
|
|
|
|
* Does the application expect PROT_READ to imply PROT_EXEC?
|
|
|
|
|
*
|
|
|
|
|
* (the exception is when the underlying filesystem is noexec
|
|
|
|
|
* mounted, in which case we dont add PROT_EXEC.)
|
|
|
|
|
*/
|
|
|
|
|
if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
|
2015-06-30 03:42:03 +08:00
|
|
|
if (!(file && path_noexec(&file->f_path)))
|
2005-04-17 06:20:36 +08:00
|
|
|
prot |= PROT_EXEC;
|
|
|
|
|
|
mm: introduce MAP_FIXED_NOREPLACE
Patch series "mm: introduce MAP_FIXED_NOREPLACE", v2.
This has started as a follow up discussion [3][4] resulting in the
runtime failure caused by hardening patch [5] which removes MAP_FIXED
from the elf loader because MAP_FIXED is inherently dangerous as it
might silently clobber an existing underlying mapping (e.g. stack).
The reason for the failure is that some architectures enforce an
alignment for the given address hint without MAP_FIXED used (e.g. for
shared or file backed mappings).
One way around this would be excluding those archs which do alignment
tricks from the hardening [6]. The patch is really trivial but it has
been objected, rightfully so, that this screams for a more generic
solution. We basically want a non-destructive MAP_FIXED.
The first patch introduced MAP_FIXED_NOREPLACE which enforces the given
address but unlike MAP_FIXED it fails with EEXIST if the given range
conflicts with an existing one. The flag is introduced as a completely
new one rather than a MAP_FIXED extension because of the backward
compatibility. We really want a never-clobber semantic even on older
kernels which do not recognize the flag. Unfortunately mmap sucks
wrt flags evaluation because we do not EINVAL on unknown flags. On
those kernels we would simply use the traditional hint based semantic so
the caller can still get a different address (which sucks) but at least
not silently corrupt an existing mapping. I do not see a good way
around that. Except we won't export expose the new semantic to the
userspace at all.
It seems there are users who would like to have something like that.
Jemalloc has been mentioned by Michael Ellerman [7]
Florian Weimer has mentioned the following:
: glibc ld.so currently maps DSOs without hints. This means that the kernel
: will map right next to each other, and the offsets between them a completely
: predictable. We would like to change that and supply a random address in a
: window of the address space. If there is a conflict, we do not want the
: kernel to pick a non-random address. Instead, we would try again with a
: random address.
John Hubbard has mentioned CUDA example
: a) Searches /proc/<pid>/maps for a "suitable" region of available
: VA space. "Suitable" generally means it has to have a base address
: within a certain limited range (a particular device model might
: have odd limitations, for example), it has to be large enough, and
: alignment has to be large enough (again, various devices may have
: constraints that lead us to do this).
:
: This is of course subject to races with other threads in the process.
:
: Let's say it finds a region starting at va.
:
: b) Next it does:
: p = mmap(va, ...)
:
: *without* setting MAP_FIXED, of course (so va is just a hint), to
: attempt to safely reserve that region. If p != va, then in most cases,
: this is a failure (almost certainly due to another thread getting a
: mapping from that region before we did), and so this layer now has to
: call munmap(), before returning a "failure: retry" to upper layers.
:
: IMPROVEMENT: --> if instead, we could call this:
:
: p = mmap(va, ... MAP_FIXED_NOREPLACE ...)
:
: , then we could skip the munmap() call upon failure. This
: is a small thing, but it is useful here. (Thanks to Piotr
: Jaroszynski and Mark Hairgrove for helping me get that detail
: exactly right, btw.)
:
: c) After that, CUDA suballocates from p, via:
:
: q = mmap(sub_region_start, ... MAP_FIXED ...)
:
: Interestingly enough, "freeing" is also done via MAP_FIXED, and
: setting PROT_NONE to the subregion. Anyway, I just included (c) for
: general interest.
Atomic address range probing in the multithreaded programs in general
sounds like an interesting thing to me.
The second patch simply replaces MAP_FIXED use in elf loader by
MAP_FIXED_NOREPLACE. I believe other places which rely on MAP_FIXED
should follow. Actually real MAP_FIXED usages should be docummented
properly and they should be more of an exception.
[1] http://lkml.kernel.org/r/20171116101900.13621-1-mhocko@kernel.org
[2] http://lkml.kernel.org/r/20171129144219.22867-1-mhocko@kernel.org
[3] http://lkml.kernel.org/r/20171107162217.382cd754@canb.auug.org.au
[4] http://lkml.kernel.org/r/1510048229.12079.7.camel@abdul.in.ibm.com
[5] http://lkml.kernel.org/r/20171023082608.6167-1-mhocko@kernel.org
[6] http://lkml.kernel.org/r/20171113094203.aofz2e7kueitk55y@dhcp22.suse.cz
[7] http://lkml.kernel.org/r/87efp1w7vy.fsf@concordia.ellerman.id.au
This patch (of 2):
MAP_FIXED is used quite often to enforce mapping at the particular range.
The main problem of this flag is, however, that it is inherently dangerous
because it unmaps existing mappings covered by the requested range. This
can cause silent memory corruptions. Some of them even with serious
security implications. While the current semantic might be really
desiderable in many cases there are others which would want to enforce the
given range but rather see a failure than a silent memory corruption on a
clashing range. Please note that there is no guarantee that a given range
is obeyed by the mmap even when it is free - e.g. arch specific code is
allowed to apply an alignment.
Introduce a new MAP_FIXED_NOREPLACE flag for mmap to achieve this
behavior. It has the same semantic as MAP_FIXED wrt. the given address
request with a single exception that it fails with EEXIST if the requested
address is already covered by an existing mapping. We still do rely on
get_unmaped_area to handle all the arch specific MAP_FIXED treatment and
check for a conflicting vma after it returns.
The flag is introduced as a completely new one rather than a MAP_FIXED
extension because of the backward compatibility. We really want a
never-clobber semantic even on older kernels which do not recognize the
flag. Unfortunately mmap sucks wrt. flags evaluation because we do not
EINVAL on unknown flags. On those kernels we would simply use the
traditional hint based semantic so the caller can still get a different
address (which sucks) but at least not silently corrupt an existing
mapping. I do not see a good way around that.
[mpe@ellerman.id.au: fix whitespace]
[fail on clashing range with EEXIST as per Florian Weimer]
[set MAP_FIXED before round_hint_to_min as per Khalid Aziz]
Link: http://lkml.kernel.org/r/20171213092550.2774-2-mhocko@kernel.org
Reviewed-by: Khalid Aziz <khalid.aziz@oracle.com>
Signed-off-by: Michal Hocko <mhocko@suse.com>
Acked-by: Michael Ellerman <mpe@ellerman.id.au>
Cc: Khalid Aziz <khalid.aziz@oracle.com>
Cc: Russell King - ARM Linux <linux@armlinux.org.uk>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Florian Weimer <fweimer@redhat.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Abdul Haleem <abdhalee@linux.vnet.ibm.com>
Cc: Joel Stanley <joel@jms.id.au>
Cc: Kees Cook <keescook@chromium.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Jason Evans <jasone@google.com>
Cc: David Goldblatt <davidtgoldblatt@gmail.com>
Cc: Edward Tomasz Napierała <trasz@FreeBSD.org>
Cc: Anshuman Khandual <khandual@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-04-11 07:35:57 +08:00
|
|
|
/* force arch specific MAP_FIXED handling in get_unmapped_area */
|
|
|
|
|
if (flags & MAP_FIXED_NOREPLACE)
|
|
|
|
|
flags |= MAP_FIXED;
|
|
|
|
|
|
2007-11-27 07:47:40 +08:00
|
|
|
if (!(flags & MAP_FIXED))
|
|
|
|
|
addr = round_hint_to_min(addr);
|
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
/* Careful about overflows.. */
|
|
|
|
|
len = PAGE_ALIGN(len);
|
2009-12-04 04:23:11 +08:00
|
|
|
if (!len)
|
2005-04-17 06:20:36 +08:00
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
|
|
/* offset overflow? */
|
|
|
|
|
if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
|
2014-10-10 06:26:29 +08:00
|
|
|
return -EOVERFLOW;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
|
/* Too many mappings? */
|
|
|
|
|
if (mm->map_count > sysctl_max_map_count)
|
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
|
|
/* Obtain the address to map to. we verify (or select) it and ensure
|
|
|
|
|
* that it represents a valid section of the address space.
|
|
|
|
|
*/
|
|
|
|
|
addr = get_unmapped_area(file, addr, len, pgoff, flags);
|
2019-12-01 09:51:03 +08:00
|
|
|
if (IS_ERR_VALUE(addr))
|
2005-04-17 06:20:36 +08:00
|
|
|
return addr;
|
|
|
|
|
|
mm: introduce MAP_FIXED_NOREPLACE
Patch series "mm: introduce MAP_FIXED_NOREPLACE", v2.
This has started as a follow up discussion [3][4] resulting in the
runtime failure caused by hardening patch [5] which removes MAP_FIXED
from the elf loader because MAP_FIXED is inherently dangerous as it
might silently clobber an existing underlying mapping (e.g. stack).
The reason for the failure is that some architectures enforce an
alignment for the given address hint without MAP_FIXED used (e.g. for
shared or file backed mappings).
One way around this would be excluding those archs which do alignment
tricks from the hardening [6]. The patch is really trivial but it has
been objected, rightfully so, that this screams for a more generic
solution. We basically want a non-destructive MAP_FIXED.
The first patch introduced MAP_FIXED_NOREPLACE which enforces the given
address but unlike MAP_FIXED it fails with EEXIST if the given range
conflicts with an existing one. The flag is introduced as a completely
new one rather than a MAP_FIXED extension because of the backward
compatibility. We really want a never-clobber semantic even on older
kernels which do not recognize the flag. Unfortunately mmap sucks
wrt flags evaluation because we do not EINVAL on unknown flags. On
those kernels we would simply use the traditional hint based semantic so
the caller can still get a different address (which sucks) but at least
not silently corrupt an existing mapping. I do not see a good way
around that. Except we won't export expose the new semantic to the
userspace at all.
It seems there are users who would like to have something like that.
Jemalloc has been mentioned by Michael Ellerman [7]
Florian Weimer has mentioned the following:
: glibc ld.so currently maps DSOs without hints. This means that the kernel
: will map right next to each other, and the offsets between them a completely
: predictable. We would like to change that and supply a random address in a
: window of the address space. If there is a conflict, we do not want the
: kernel to pick a non-random address. Instead, we would try again with a
: random address.
John Hubbard has mentioned CUDA example
: a) Searches /proc/<pid>/maps for a "suitable" region of available
: VA space. "Suitable" generally means it has to have a base address
: within a certain limited range (a particular device model might
: have odd limitations, for example), it has to be large enough, and
: alignment has to be large enough (again, various devices may have
: constraints that lead us to do this).
:
: This is of course subject to races with other threads in the process.
:
: Let's say it finds a region starting at va.
:
: b) Next it does:
: p = mmap(va, ...)
:
: *without* setting MAP_FIXED, of course (so va is just a hint), to
: attempt to safely reserve that region. If p != va, then in most cases,
: this is a failure (almost certainly due to another thread getting a
: mapping from that region before we did), and so this layer now has to
: call munmap(), before returning a "failure: retry" to upper layers.
:
: IMPROVEMENT: --> if instead, we could call this:
:
: p = mmap(va, ... MAP_FIXED_NOREPLACE ...)
:
: , then we could skip the munmap() call upon failure. This
: is a small thing, but it is useful here. (Thanks to Piotr
: Jaroszynski and Mark Hairgrove for helping me get that detail
: exactly right, btw.)
:
: c) After that, CUDA suballocates from p, via:
:
: q = mmap(sub_region_start, ... MAP_FIXED ...)
:
: Interestingly enough, "freeing" is also done via MAP_FIXED, and
: setting PROT_NONE to the subregion. Anyway, I just included (c) for
: general interest.
Atomic address range probing in the multithreaded programs in general
sounds like an interesting thing to me.
The second patch simply replaces MAP_FIXED use in elf loader by
MAP_FIXED_NOREPLACE. I believe other places which rely on MAP_FIXED
should follow. Actually real MAP_FIXED usages should be docummented
properly and they should be more of an exception.
[1] http://lkml.kernel.org/r/20171116101900.13621-1-mhocko@kernel.org
[2] http://lkml.kernel.org/r/20171129144219.22867-1-mhocko@kernel.org
[3] http://lkml.kernel.org/r/20171107162217.382cd754@canb.auug.org.au
[4] http://lkml.kernel.org/r/1510048229.12079.7.camel@abdul.in.ibm.com
[5] http://lkml.kernel.org/r/20171023082608.6167-1-mhocko@kernel.org
[6] http://lkml.kernel.org/r/20171113094203.aofz2e7kueitk55y@dhcp22.suse.cz
[7] http://lkml.kernel.org/r/87efp1w7vy.fsf@concordia.ellerman.id.au
This patch (of 2):
MAP_FIXED is used quite often to enforce mapping at the particular range.
The main problem of this flag is, however, that it is inherently dangerous
because it unmaps existing mappings covered by the requested range. This
can cause silent memory corruptions. Some of them even with serious
security implications. While the current semantic might be really
desiderable in many cases there are others which would want to enforce the
given range but rather see a failure than a silent memory corruption on a
clashing range. Please note that there is no guarantee that a given range
is obeyed by the mmap even when it is free - e.g. arch specific code is
allowed to apply an alignment.
Introduce a new MAP_FIXED_NOREPLACE flag for mmap to achieve this
behavior. It has the same semantic as MAP_FIXED wrt. the given address
request with a single exception that it fails with EEXIST if the requested
address is already covered by an existing mapping. We still do rely on
get_unmaped_area to handle all the arch specific MAP_FIXED treatment and
check for a conflicting vma after it returns.
The flag is introduced as a completely new one rather than a MAP_FIXED
extension because of the backward compatibility. We really want a
never-clobber semantic even on older kernels which do not recognize the
flag. Unfortunately mmap sucks wrt. flags evaluation because we do not
EINVAL on unknown flags. On those kernels we would simply use the
traditional hint based semantic so the caller can still get a different
address (which sucks) but at least not silently corrupt an existing
mapping. I do not see a good way around that.
[mpe@ellerman.id.au: fix whitespace]
[fail on clashing range with EEXIST as per Florian Weimer]
[set MAP_FIXED before round_hint_to_min as per Khalid Aziz]
Link: http://lkml.kernel.org/r/20171213092550.2774-2-mhocko@kernel.org
Reviewed-by: Khalid Aziz <khalid.aziz@oracle.com>
Signed-off-by: Michal Hocko <mhocko@suse.com>
Acked-by: Michael Ellerman <mpe@ellerman.id.au>
Cc: Khalid Aziz <khalid.aziz@oracle.com>
Cc: Russell King - ARM Linux <linux@armlinux.org.uk>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Florian Weimer <fweimer@redhat.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Abdul Haleem <abdhalee@linux.vnet.ibm.com>
Cc: Joel Stanley <joel@jms.id.au>
Cc: Kees Cook <keescook@chromium.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Jason Evans <jasone@google.com>
Cc: David Goldblatt <davidtgoldblatt@gmail.com>
Cc: Edward Tomasz Napierała <trasz@FreeBSD.org>
Cc: Anshuman Khandual <khandual@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-04-11 07:35:57 +08:00
|
|
|
if (flags & MAP_FIXED_NOREPLACE) {
|
2021-06-29 10:38:44 +08:00
|
|
|
if (find_vma_intersection(mm, addr, addr + len))
|
mm: introduce MAP_FIXED_NOREPLACE
Patch series "mm: introduce MAP_FIXED_NOREPLACE", v2.
This has started as a follow up discussion [3][4] resulting in the
runtime failure caused by hardening patch [5] which removes MAP_FIXED
from the elf loader because MAP_FIXED is inherently dangerous as it
might silently clobber an existing underlying mapping (e.g. stack).
The reason for the failure is that some architectures enforce an
alignment for the given address hint without MAP_FIXED used (e.g. for
shared or file backed mappings).
One way around this would be excluding those archs which do alignment
tricks from the hardening [6]. The patch is really trivial but it has
been objected, rightfully so, that this screams for a more generic
solution. We basically want a non-destructive MAP_FIXED.
The first patch introduced MAP_FIXED_NOREPLACE which enforces the given
address but unlike MAP_FIXED it fails with EEXIST if the given range
conflicts with an existing one. The flag is introduced as a completely
new one rather than a MAP_FIXED extension because of the backward
compatibility. We really want a never-clobber semantic even on older
kernels which do not recognize the flag. Unfortunately mmap sucks
wrt flags evaluation because we do not EINVAL on unknown flags. On
those kernels we would simply use the traditional hint based semantic so
the caller can still get a different address (which sucks) but at least
not silently corrupt an existing mapping. I do not see a good way
around that. Except we won't export expose the new semantic to the
userspace at all.
It seems there are users who would like to have something like that.
Jemalloc has been mentioned by Michael Ellerman [7]
Florian Weimer has mentioned the following:
: glibc ld.so currently maps DSOs without hints. This means that the kernel
: will map right next to each other, and the offsets between them a completely
: predictable. We would like to change that and supply a random address in a
: window of the address space. If there is a conflict, we do not want the
: kernel to pick a non-random address. Instead, we would try again with a
: random address.
John Hubbard has mentioned CUDA example
: a) Searches /proc/<pid>/maps for a "suitable" region of available
: VA space. "Suitable" generally means it has to have a base address
: within a certain limited range (a particular device model might
: have odd limitations, for example), it has to be large enough, and
: alignment has to be large enough (again, various devices may have
: constraints that lead us to do this).
:
: This is of course subject to races with other threads in the process.
:
: Let's say it finds a region starting at va.
:
: b) Next it does:
: p = mmap(va, ...)
:
: *without* setting MAP_FIXED, of course (so va is just a hint), to
: attempt to safely reserve that region. If p != va, then in most cases,
: this is a failure (almost certainly due to another thread getting a
: mapping from that region before we did), and so this layer now has to
: call munmap(), before returning a "failure: retry" to upper layers.
:
: IMPROVEMENT: --> if instead, we could call this:
:
: p = mmap(va, ... MAP_FIXED_NOREPLACE ...)
:
: , then we could skip the munmap() call upon failure. This
: is a small thing, but it is useful here. (Thanks to Piotr
: Jaroszynski and Mark Hairgrove for helping me get that detail
: exactly right, btw.)
:
: c) After that, CUDA suballocates from p, via:
:
: q = mmap(sub_region_start, ... MAP_FIXED ...)
:
: Interestingly enough, "freeing" is also done via MAP_FIXED, and
: setting PROT_NONE to the subregion. Anyway, I just included (c) for
: general interest.
Atomic address range probing in the multithreaded programs in general
sounds like an interesting thing to me.
The second patch simply replaces MAP_FIXED use in elf loader by
MAP_FIXED_NOREPLACE. I believe other places which rely on MAP_FIXED
should follow. Actually real MAP_FIXED usages should be docummented
properly and they should be more of an exception.
[1] http://lkml.kernel.org/r/20171116101900.13621-1-mhocko@kernel.org
[2] http://lkml.kernel.org/r/20171129144219.22867-1-mhocko@kernel.org
[3] http://lkml.kernel.org/r/20171107162217.382cd754@canb.auug.org.au
[4] http://lkml.kernel.org/r/1510048229.12079.7.camel@abdul.in.ibm.com
[5] http://lkml.kernel.org/r/20171023082608.6167-1-mhocko@kernel.org
[6] http://lkml.kernel.org/r/20171113094203.aofz2e7kueitk55y@dhcp22.suse.cz
[7] http://lkml.kernel.org/r/87efp1w7vy.fsf@concordia.ellerman.id.au
This patch (of 2):
MAP_FIXED is used quite often to enforce mapping at the particular range.
The main problem of this flag is, however, that it is inherently dangerous
because it unmaps existing mappings covered by the requested range. This
can cause silent memory corruptions. Some of them even with serious
security implications. While the current semantic might be really
desiderable in many cases there are others which would want to enforce the
given range but rather see a failure than a silent memory corruption on a
clashing range. Please note that there is no guarantee that a given range
is obeyed by the mmap even when it is free - e.g. arch specific code is
allowed to apply an alignment.
Introduce a new MAP_FIXED_NOREPLACE flag for mmap to achieve this
behavior. It has the same semantic as MAP_FIXED wrt. the given address
request with a single exception that it fails with EEXIST if the requested
address is already covered by an existing mapping. We still do rely on
get_unmaped_area to handle all the arch specific MAP_FIXED treatment and
check for a conflicting vma after it returns.
The flag is introduced as a completely new one rather than a MAP_FIXED
extension because of the backward compatibility. We really want a
never-clobber semantic even on older kernels which do not recognize the
flag. Unfortunately mmap sucks wrt. flags evaluation because we do not
EINVAL on unknown flags. On those kernels we would simply use the
traditional hint based semantic so the caller can still get a different
address (which sucks) but at least not silently corrupt an existing
mapping. I do not see a good way around that.
[mpe@ellerman.id.au: fix whitespace]
[fail on clashing range with EEXIST as per Florian Weimer]
[set MAP_FIXED before round_hint_to_min as per Khalid Aziz]
Link: http://lkml.kernel.org/r/20171213092550.2774-2-mhocko@kernel.org
Reviewed-by: Khalid Aziz <khalid.aziz@oracle.com>
Signed-off-by: Michal Hocko <mhocko@suse.com>
Acked-by: Michael Ellerman <mpe@ellerman.id.au>
Cc: Khalid Aziz <khalid.aziz@oracle.com>
Cc: Russell King - ARM Linux <linux@armlinux.org.uk>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Florian Weimer <fweimer@redhat.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Abdul Haleem <abdhalee@linux.vnet.ibm.com>
Cc: Joel Stanley <joel@jms.id.au>
Cc: Kees Cook <keescook@chromium.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Jason Evans <jasone@google.com>
Cc: David Goldblatt <davidtgoldblatt@gmail.com>
Cc: Edward Tomasz Napierała <trasz@FreeBSD.org>
Cc: Anshuman Khandual <khandual@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-04-11 07:35:57 +08:00
|
|
|
return -EEXIST;
|
|
|
|
|
}
|
|
|
|
|
|
mm/core, x86/mm/pkeys: Add execute-only protection keys support
Protection keys provide new page-based protection in hardware.
But, they have an interesting attribute: they only affect data
accesses and never affect instruction fetches. That means that
if we set up some memory which is set as "access-disabled" via
protection keys, we can still execute from it.
This patch uses protection keys to set up mappings to do just that.
If a user calls:
mmap(..., PROT_EXEC);
or
mprotect(ptr, sz, PROT_EXEC);
(note PROT_EXEC-only without PROT_READ/WRITE), the kernel will
notice this, and set a special protection key on the memory. It
also sets the appropriate bits in the Protection Keys User Rights
(PKRU) register so that the memory becomes unreadable and
unwritable.
I haven't found any userspace that does this today. With this
facility in place, we expect userspace to move to use it
eventually. Userspace _could_ start doing this today. Any
PROT_EXEC calls get converted to PROT_READ inside the kernel, and
would transparently be upgraded to "true" PROT_EXEC with this
code. IOW, userspace never has to do any PROT_EXEC runtime
detection.
This feature provides enhanced protection against leaking
executable memory contents. This helps thwart attacks which are
attempting to find ROP gadgets on the fly.
But, the security provided by this approach is not comprehensive.
The PKRU register which controls access permissions is a normal
user register writable from unprivileged userspace. An attacker
who can execute the 'wrpkru' instruction can easily disable the
protection provided by this feature.
The protection key that is used for execute-only support is
permanently dedicated at compile time. This is fine for now
because there is currently no API to set a protection key other
than this one.
Despite there being a constant PKRU value across the entire
system, we do not set it unless this feature is in use in a
process. That is to preserve the PKRU XSAVE 'init state',
which can lead to faster context switches.
PKRU *is* a user register and the kernel is modifying it. That
means that code doing:
pkru = rdpkru()
pkru |= 0x100;
mmap(..., PROT_EXEC);
wrpkru(pkru);
could lose the bits in PKRU that enforce execute-only
permissions. To avoid this, we suggest avoiding ever calling
mmap() or mprotect() when the PKRU value is expected to be
unstable.
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Borislav Petkov <bp@suse.de>
Cc: Brian Gerst <brgerst@gmail.com>
Cc: Chen Gang <gang.chen.5i5j@gmail.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Dave Hansen <dave@sr71.net>
Cc: David Hildenbrand <dahi@linux.vnet.ibm.com>
Cc: Denys Vlasenko <dvlasenk@redhat.com>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Kees Cook <keescook@chromium.org>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Konstantin Khlebnikov <koct9i@gmail.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Piotr Kwapulinski <kwapulinski.piotr@gmail.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Stephen Smalley <sds@tycho.nsa.gov>
Cc: Vladimir Murzin <vladimir.murzin@arm.com>
Cc: Will Deacon <will.deacon@arm.com>
Cc: keescook@google.com
Cc: linux-kernel@vger.kernel.org
Cc: linux-mm@kvack.org
Link: http://lkml.kernel.org/r/20160212210240.CB4BB5CA@viggo.jf.intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-02-13 05:02:40 +08:00
|
|
|
if (prot == PROT_EXEC) {
|
|
|
|
|
pkey = execute_only_pkey(mm);
|
|
|
|
|
if (pkey < 0)
|
|
|
|
|
pkey = 0;
|
|
|
|
|
}
|
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
/* Do simple checking here so the lower-level routines won't have
|
|
|
|
|
* to. we assume access permissions have been handled by the open
|
|
|
|
|
* of the memory object, so we don't do any here.
|
|
|
|
|
*/
|
2020-08-07 14:23:37 +08:00
|
|
|
vm_flags = calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
|
2005-04-17 06:20:36 +08:00
|
|
|
mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
|
|
|
|
|
|
2009-09-22 08:03:36 +08:00
|
|
|
if (flags & MAP_LOCKED)
|
2005-04-17 06:20:36 +08:00
|
|
|
if (!can_do_mlock())
|
|
|
|
|
return -EPERM;
|
2008-10-19 11:26:50 +08:00
|
|
|
|
2014-01-22 07:49:15 +08:00
|
|
|
if (mlock_future_check(mm, vm_flags, len))
|
|
|
|
|
return -EAGAIN;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
|
if (file) {
|
2013-09-12 05:20:19 +08:00
|
|
|
struct inode *inode = file_inode(file);
|
mm: introduce MAP_SHARED_VALIDATE, a mechanism to safely define new mmap flags
The mmap(2) syscall suffers from the ABI anti-pattern of not validating
unknown flags. However, proposals like MAP_SYNC need a mechanism to
define new behavior that is known to fail on older kernels without the
support. Define a new MAP_SHARED_VALIDATE flag pattern that is
guaranteed to fail on all legacy mmap implementations.
It is worth noting that the original proposal was for a standalone
MAP_VALIDATE flag. However, when that could not be supported by all
archs Linus observed:
I see why you *think* you want a bitmap. You think you want
a bitmap because you want to make MAP_VALIDATE be part of MAP_SYNC
etc, so that people can do
ret = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_SHARED
| MAP_SYNC, fd, 0);
and "know" that MAP_SYNC actually takes.
And I'm saying that whole wish is bogus. You're fundamentally
depending on special semantics, just make it explicit. It's already
not portable, so don't try to make it so.
Rename that MAP_VALIDATE as MAP_SHARED_VALIDATE, make it have a value
of 0x3, and make people do
ret = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_SHARED_VALIDATE
| MAP_SYNC, fd, 0);
and then the kernel side is easier too (none of that random garbage
playing games with looking at the "MAP_VALIDATE bit", but just another
case statement in that map type thing.
Boom. Done.
Similar to ->fallocate() we also want the ability to validate the
support for new flags on a per ->mmap() 'struct file_operations'
instance basis. Towards that end arrange for flags to be generically
validated against a mmap_supported_flags exported by 'struct
file_operations'. By default all existing flags are implicitly
supported, but new flags require MAP_SHARED_VALIDATE and
per-instance-opt-in.
Cc: Jan Kara <jack@suse.cz>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Suggested-by: Christoph Hellwig <hch@lst.de>
Suggested-by: Linus Torvalds <torvalds@linux-foundation.org>
Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2017-11-01 23:36:30 +08:00
|
|
|
unsigned long flags_mask;
|
|
|
|
|
|
mmap: introduce sane default mmap limits
The internal VM "mmap()" interfaces are based on the mmap target doing
everything using page indexes rather than byte offsets, because
traditionally (ie 32-bit) we had the situation that the byte offset
didn't fit in a register. So while the mmap virtual address was limited
by the word size of the architecture, the backing store was not.
So we're basically passing "pgoff" around as a page index, in order to
be able to describe backing store locations that are much bigger than
the word size (think files larger than 4GB etc).
But while this all makes a ton of sense conceptually, we've been dogged
by various drivers that don't really understand this, and internally
work with byte offsets, and then try to work with the page index by
turning it into a byte offset with "pgoff << PAGE_SHIFT".
Which obviously can overflow.
Adding the size of the mapping to it to get the byte offset of the end
of the backing store just exacerbates the problem, and if you then use
this overflow-prone value to check various limits of your device driver
mmap capability, you're just setting yourself up for problems.
The correct thing for drivers to do is to do their limit math in page
indices, the way the interface is designed. Because the generic mmap
code _does_ test that the index doesn't overflow, since that's what the
mmap code really cares about.
HOWEVER.
Finding and fixing various random drivers is a sisyphean task, so let's
just see if we can just make the core mmap() code do the limiting for
us. Realistically, the only "big" backing stores we need to care about
are regular files and block devices, both of which are known to do this
properly, and which have nice well-defined limits for how much data they
can access.
So let's special-case just those two known cases, and then limit other
random mmap users to a backing store that still fits in "unsigned long".
Realistically, that's not much of a limit at all on 64-bit, and on
32-bit architectures the only worry might be the GPU drivers, which can
have big physical address spaces.
To make it possible for drivers like that to say that they are 64-bit
clean, this patch does repurpose the "FMODE_UNSIGNED_OFFSET" bit in the
file flags to allow drivers to mark their file descriptors as safe in
the full 64-bit mmap address space.
[ The timing for doing this is less than optimal, and this should really
go in a merge window. But realistically, this needs wide testing more
than it needs anything else, and being main-line is the only way to do
that.
So the earlier the better, even if it's outside the proper development
cycle - Linus ]
Cc: Kees Cook <keescook@chromium.org>
Cc: Dan Carpenter <dan.carpenter@oracle.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Willy Tarreau <w@1wt.eu>
Cc: Dave Airlie <airlied@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-05-12 00:52:01 +08:00
|
|
|
if (!file_mmap_ok(file, inode, pgoff, len))
|
|
|
|
|
return -EOVERFLOW;
|
|
|
|
|
|
mm: introduce MAP_SHARED_VALIDATE, a mechanism to safely define new mmap flags
The mmap(2) syscall suffers from the ABI anti-pattern of not validating
unknown flags. However, proposals like MAP_SYNC need a mechanism to
define new behavior that is known to fail on older kernels without the
support. Define a new MAP_SHARED_VALIDATE flag pattern that is
guaranteed to fail on all legacy mmap implementations.
It is worth noting that the original proposal was for a standalone
MAP_VALIDATE flag. However, when that could not be supported by all
archs Linus observed:
I see why you *think* you want a bitmap. You think you want
a bitmap because you want to make MAP_VALIDATE be part of MAP_SYNC
etc, so that people can do
ret = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_SHARED
| MAP_SYNC, fd, 0);
and "know" that MAP_SYNC actually takes.
And I'm saying that whole wish is bogus. You're fundamentally
depending on special semantics, just make it explicit. It's already
not portable, so don't try to make it so.
Rename that MAP_VALIDATE as MAP_SHARED_VALIDATE, make it have a value
of 0x3, and make people do
ret = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_SHARED_VALIDATE
| MAP_SYNC, fd, 0);
and then the kernel side is easier too (none of that random garbage
playing games with looking at the "MAP_VALIDATE bit", but just another
case statement in that map type thing.
Boom. Done.
Similar to ->fallocate() we also want the ability to validate the
support for new flags on a per ->mmap() 'struct file_operations'
instance basis. Towards that end arrange for flags to be generically
validated against a mmap_supported_flags exported by 'struct
file_operations'. By default all existing flags are implicitly
supported, but new flags require MAP_SHARED_VALIDATE and
per-instance-opt-in.
Cc: Jan Kara <jack@suse.cz>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Suggested-by: Christoph Hellwig <hch@lst.de>
Suggested-by: Linus Torvalds <torvalds@linux-foundation.org>
Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2017-11-01 23:36:30 +08:00
|
|
|
flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
|
2013-09-12 05:20:19 +08:00
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
switch (flags & MAP_TYPE) {
|
|
|
|
|
case MAP_SHARED:
|
mm: introduce MAP_SHARED_VALIDATE, a mechanism to safely define new mmap flags
The mmap(2) syscall suffers from the ABI anti-pattern of not validating
unknown flags. However, proposals like MAP_SYNC need a mechanism to
define new behavior that is known to fail on older kernels without the
support. Define a new MAP_SHARED_VALIDATE flag pattern that is
guaranteed to fail on all legacy mmap implementations.
It is worth noting that the original proposal was for a standalone
MAP_VALIDATE flag. However, when that could not be supported by all
archs Linus observed:
I see why you *think* you want a bitmap. You think you want
a bitmap because you want to make MAP_VALIDATE be part of MAP_SYNC
etc, so that people can do
ret = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_SHARED
| MAP_SYNC, fd, 0);
and "know" that MAP_SYNC actually takes.
And I'm saying that whole wish is bogus. You're fundamentally
depending on special semantics, just make it explicit. It's already
not portable, so don't try to make it so.
Rename that MAP_VALIDATE as MAP_SHARED_VALIDATE, make it have a value
of 0x3, and make people do
ret = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_SHARED_VALIDATE
| MAP_SYNC, fd, 0);
and then the kernel side is easier too (none of that random garbage
playing games with looking at the "MAP_VALIDATE bit", but just another
case statement in that map type thing.
Boom. Done.
Similar to ->fallocate() we also want the ability to validate the
support for new flags on a per ->mmap() 'struct file_operations'
instance basis. Towards that end arrange for flags to be generically
validated against a mmap_supported_flags exported by 'struct
file_operations'. By default all existing flags are implicitly
supported, but new flags require MAP_SHARED_VALIDATE and
per-instance-opt-in.
Cc: Jan Kara <jack@suse.cz>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Suggested-by: Christoph Hellwig <hch@lst.de>
Suggested-by: Linus Torvalds <torvalds@linux-foundation.org>
Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2017-11-01 23:36:30 +08:00
|
|
|
/*
|
|
|
|
|
* Force use of MAP_SHARED_VALIDATE with non-legacy
|
|
|
|
|
* flags. E.g. MAP_SYNC is dangerous to use with
|
|
|
|
|
* MAP_SHARED as you don't know which consistency model
|
|
|
|
|
* you will get. We silently ignore unsupported flags
|
|
|
|
|
* with MAP_SHARED to preserve backward compatibility.
|
|
|
|
|
*/
|
|
|
|
|
flags &= LEGACY_MAP_MASK;
|
2020-04-07 11:08:39 +08:00
|
|
|
fallthrough;
|
mm: introduce MAP_SHARED_VALIDATE, a mechanism to safely define new mmap flags
The mmap(2) syscall suffers from the ABI anti-pattern of not validating
unknown flags. However, proposals like MAP_SYNC need a mechanism to
define new behavior that is known to fail on older kernels without the
support. Define a new MAP_SHARED_VALIDATE flag pattern that is
guaranteed to fail on all legacy mmap implementations.
It is worth noting that the original proposal was for a standalone
MAP_VALIDATE flag. However, when that could not be supported by all
archs Linus observed:
I see why you *think* you want a bitmap. You think you want
a bitmap because you want to make MAP_VALIDATE be part of MAP_SYNC
etc, so that people can do
ret = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_SHARED
| MAP_SYNC, fd, 0);
and "know" that MAP_SYNC actually takes.
And I'm saying that whole wish is bogus. You're fundamentally
depending on special semantics, just make it explicit. It's already
not portable, so don't try to make it so.
Rename that MAP_VALIDATE as MAP_SHARED_VALIDATE, make it have a value
of 0x3, and make people do
ret = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_SHARED_VALIDATE
| MAP_SYNC, fd, 0);
and then the kernel side is easier too (none of that random garbage
playing games with looking at the "MAP_VALIDATE bit", but just another
case statement in that map type thing.
Boom. Done.
Similar to ->fallocate() we also want the ability to validate the
support for new flags on a per ->mmap() 'struct file_operations'
instance basis. Towards that end arrange for flags to be generically
validated against a mmap_supported_flags exported by 'struct
file_operations'. By default all existing flags are implicitly
supported, but new flags require MAP_SHARED_VALIDATE and
per-instance-opt-in.
Cc: Jan Kara <jack@suse.cz>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Suggested-by: Christoph Hellwig <hch@lst.de>
Suggested-by: Linus Torvalds <torvalds@linux-foundation.org>
Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2017-11-01 23:36:30 +08:00
|
|
|
case MAP_SHARED_VALIDATE:
|
|
|
|
|
if (flags & ~flags_mask)
|
|
|
|
|
return -EOPNOTSUPP;
|
2019-08-20 22:55:16 +08:00
|
|
|
if (prot & PROT_WRITE) {
|
|
|
|
|
if (!(file->f_mode & FMODE_WRITE))
|
|
|
|
|
return -EACCES;
|
|
|
|
|
if (IS_SWAPFILE(file->f_mapping->host))
|
|
|
|
|
return -ETXTBSY;
|
|
|
|
|
}
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Make sure we don't allow writing to an append-only
|
|
|
|
|
* file..
|
|
|
|
|
*/
|
|
|
|
|
if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
|
|
|
|
|
return -EACCES;
|
|
|
|
|
|
|
|
|
|
vm_flags |= VM_SHARED | VM_MAYSHARE;
|
|
|
|
|
if (!(file->f_mode & FMODE_WRITE))
|
|
|
|
|
vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
|
2020-04-07 11:08:39 +08:00
|
|
|
fallthrough;
|
2005-04-17 06:20:36 +08:00
|
|
|
case MAP_PRIVATE:
|
|
|
|
|
if (!(file->f_mode & FMODE_READ))
|
|
|
|
|
return -EACCES;
|
2015-06-30 03:42:03 +08:00
|
|
|
if (path_noexec(&file->f_path)) {
|
2006-10-16 05:09:55 +08:00
|
|
|
if (vm_flags & VM_EXEC)
|
|
|
|
|
return -EPERM;
|
|
|
|
|
vm_flags &= ~VM_MAYEXEC;
|
|
|
|
|
}
|
|
|
|
|
|
2013-09-23 04:27:52 +08:00
|
|
|
if (!file->f_op->mmap)
|
2006-10-16 05:09:55 +08:00
|
|
|
return -ENODEV;
|
2013-09-12 05:20:18 +08:00
|
|
|
if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
|
|
|
|
|
return -EINVAL;
|
2005-04-17 06:20:36 +08:00
|
|
|
break;
|
|
|
|
|
|
|
|
|
|
default:
|
|
|
|
|
return -EINVAL;
|
|
|
|
|
}
|
|
|
|
|
} else {
|
|
|
|
|
switch (flags & MAP_TYPE) {
|
|
|
|
|
case MAP_SHARED:
|
2013-09-12 05:20:18 +08:00
|
|
|
if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
|
|
|
|
|
return -EINVAL;
|
2008-09-03 22:09:47 +08:00
|
|
|
/*
|
|
|
|
|
* Ignore pgoff.
|
|
|
|
|
*/
|
|
|
|
|
pgoff = 0;
|
2005-04-17 06:20:36 +08:00
|
|
|
vm_flags |= VM_SHARED | VM_MAYSHARE;
|
|
|
|
|
break;
|
|
|
|
|
case MAP_PRIVATE:
|
|
|
|
|
/*
|
|
|
|
|
* Set pgoff according to addr for anon_vma.
|
|
|
|
|
*/
|
|
|
|
|
pgoff = addr >> PAGE_SHIFT;
|
|
|
|
|
break;
|
|
|
|
|
default:
|
|
|
|
|
return -EINVAL;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2013-02-23 08:32:43 +08:00
|
|
|
/*
|
|
|
|
|
* Set 'VM_NORESERVE' if we should not account for the
|
|
|
|
|
* memory use of this mapping.
|
|
|
|
|
*/
|
|
|
|
|
if (flags & MAP_NORESERVE) {
|
|
|
|
|
/* We honor MAP_NORESERVE if allowed to overcommit */
|
|
|
|
|
if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
|
|
|
|
|
vm_flags |= VM_NORESERVE;
|
|
|
|
|
|
|
|
|
|
/* hugetlb applies strict overcommit unless MAP_NORESERVE */
|
|
|
|
|
if (file && is_file_hugepages(file))
|
|
|
|
|
vm_flags |= VM_NORESERVE;
|
|
|
|
|
}
|
|
|
|
|
|
2017-02-25 06:58:22 +08:00
|
|
|
addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
|
2013-03-29 07:26:23 +08:00
|
|
|
if (!IS_ERR_VALUE(addr) &&
|
|
|
|
|
((vm_flags & VM_LOCKED) ||
|
|
|
|
|
(flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
|
2013-02-23 08:32:47 +08:00
|
|
|
*populate = len;
|
2013-02-23 08:32:37 +08:00
|
|
|
return addr;
|
2007-07-16 14:38:26 +08:00
|
|
|
}
|
2012-04-21 08:13:58 +08:00
|
|
|
|
2018-03-11 18:34:46 +08:00
|
|
|
unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
|
|
|
|
|
unsigned long prot, unsigned long flags,
|
|
|
|
|
unsigned long fd, unsigned long pgoff)
|
2009-12-31 04:17:34 +08:00
|
|
|
{
|
|
|
|
|
struct file *file = NULL;
|
2015-11-06 10:48:35 +08:00
|
|
|
unsigned long retval;
|
2009-12-31 04:17:34 +08:00
|
|
|
|
|
|
|
|
if (!(flags & MAP_ANONYMOUS)) {
|
2010-10-30 14:54:44 +08:00
|
|
|
audit_mmap_fd(fd, flags);
|
2009-12-31 04:17:34 +08:00
|
|
|
file = fget(fd);
|
|
|
|
|
if (!file)
|
2015-11-06 10:48:35 +08:00
|
|
|
return -EBADF;
|
mm/mmap: optimize a branch judgment in ksys_mmap_pgoff()
Look at the pseudo code below. It's very clear that, the judgement
"!is_file_hugepages(file)" at 3) is duplicated to the one at 1), we can
use "else if" to avoid it. And the assignment "retval = -EINVAL" at 2) is
only needed by the branch 3), because "retval" will be overwritten at 4).
No functional change, but it can reduce the code size. Maybe more clearer?
Before:
text data bss dec hex filename
28733 1590 1 30324 7674 mm/mmap.o
After:
text data bss dec hex filename
28701 1590 1 30292 7654 mm/mmap.o
====pseudo code====:
if (!(flags & MAP_ANONYMOUS)) {
...
1) if (is_file_hugepages(file))
len = ALIGN(len, huge_page_size(hstate_file(file)));
2) retval = -EINVAL;
3) if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
goto out_fput;
} else if (flags & MAP_HUGETLB) {
...
}
...
4) retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
out_fput:
...
return retval;
Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Link: http://lkml.kernel.org/r/20200705080112.1405-1-thunder.leizhen@huawei.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-07 14:22:59 +08:00
|
|
|
if (is_file_hugepages(file)) {
|
2013-05-08 07:18:13 +08:00
|
|
|
len = ALIGN(len, huge_page_size(hstate_file(file)));
|
mm/mmap: optimize a branch judgment in ksys_mmap_pgoff()
Look at the pseudo code below. It's very clear that, the judgement
"!is_file_hugepages(file)" at 3) is duplicated to the one at 1), we can
use "else if" to avoid it. And the assignment "retval = -EINVAL" at 2) is
only needed by the branch 3), because "retval" will be overwritten at 4).
No functional change, but it can reduce the code size. Maybe more clearer?
Before:
text data bss dec hex filename
28733 1590 1 30324 7674 mm/mmap.o
After:
text data bss dec hex filename
28701 1590 1 30292 7654 mm/mmap.o
====pseudo code====:
if (!(flags & MAP_ANONYMOUS)) {
...
1) if (is_file_hugepages(file))
len = ALIGN(len, huge_page_size(hstate_file(file)));
2) retval = -EINVAL;
3) if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
goto out_fput;
} else if (flags & MAP_HUGETLB) {
...
}
...
4) retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
out_fput:
...
return retval;
Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Link: http://lkml.kernel.org/r/20200705080112.1405-1-thunder.leizhen@huawei.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-07 14:22:59 +08:00
|
|
|
} else if (unlikely(flags & MAP_HUGETLB)) {
|
|
|
|
|
retval = -EINVAL;
|
2013-07-09 07:00:26 +08:00
|
|
|
goto out_fput;
|
mm/mmap: optimize a branch judgment in ksys_mmap_pgoff()
Look at the pseudo code below. It's very clear that, the judgement
"!is_file_hugepages(file)" at 3) is duplicated to the one at 1), we can
use "else if" to avoid it. And the assignment "retval = -EINVAL" at 2) is
only needed by the branch 3), because "retval" will be overwritten at 4).
No functional change, but it can reduce the code size. Maybe more clearer?
Before:
text data bss dec hex filename
28733 1590 1 30324 7674 mm/mmap.o
After:
text data bss dec hex filename
28701 1590 1 30292 7654 mm/mmap.o
====pseudo code====:
if (!(flags & MAP_ANONYMOUS)) {
...
1) if (is_file_hugepages(file))
len = ALIGN(len, huge_page_size(hstate_file(file)));
2) retval = -EINVAL;
3) if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
goto out_fput;
} else if (flags & MAP_HUGETLB) {
...
}
...
4) retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
out_fput:
...
return retval;
Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Link: http://lkml.kernel.org/r/20200705080112.1405-1-thunder.leizhen@huawei.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-07 14:22:59 +08:00
|
|
|
}
|
2009-12-31 04:17:34 +08:00
|
|
|
} else if (flags & MAP_HUGETLB) {
|
2013-07-09 07:01:08 +08:00
|
|
|
struct hstate *hs;
|
2013-05-08 07:18:13 +08:00
|
|
|
|
2017-05-04 05:55:00 +08:00
|
|
|
hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
|
2013-05-09 15:08:15 +08:00
|
|
|
if (!hs)
|
|
|
|
|
return -EINVAL;
|
|
|
|
|
|
|
|
|
|
len = ALIGN(len, huge_page_size(hs));
|
2009-12-31 04:17:34 +08:00
|
|
|
/*
|
|
|
|
|
* VM_NORESERVE is used because the reservations will be
|
|
|
|
|
* taken when vm_ops->mmap() is called
|
|
|
|
|
*/
|
2013-05-08 07:18:13 +08:00
|
|
|
file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
|
2012-12-12 08:01:34 +08:00
|
|
|
VM_NORESERVE,
|
2021-11-09 10:31:27 +08:00
|
|
|
HUGETLB_ANONHUGE_INODE,
|
2012-12-12 08:01:34 +08:00
|
|
|
(flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
|
2009-12-31 04:17:34 +08:00
|
|
|
if (IS_ERR(file))
|
|
|
|
|
return PTR_ERR(file);
|
|
|
|
|
}
|
|
|
|
|
|
2016-05-24 07:25:30 +08:00
|
|
|
retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
|
2013-07-09 07:00:26 +08:00
|
|
|
out_fput:
|
2009-12-31 04:17:34 +08:00
|
|
|
if (file)
|
|
|
|
|
fput(file);
|
|
|
|
|
return retval;
|
|
|
|
|
}
|
|
|
|
|
|
2018-03-11 18:34:46 +08:00
|
|
|
SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
|
|
|
|
|
unsigned long, prot, unsigned long, flags,
|
|
|
|
|
unsigned long, fd, unsigned long, pgoff)
|
|
|
|
|
{
|
|
|
|
|
return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
|
|
|
|
|
}
|
|
|
|
|
|
2010-03-11 07:21:15 +08:00
|
|
|
#ifdef __ARCH_WANT_SYS_OLD_MMAP
|
|
|
|
|
struct mmap_arg_struct {
|
|
|
|
|
unsigned long addr;
|
|
|
|
|
unsigned long len;
|
|
|
|
|
unsigned long prot;
|
|
|
|
|
unsigned long flags;
|
|
|
|
|
unsigned long fd;
|
|
|
|
|
unsigned long offset;
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
|
|
|
|
|
{
|
|
|
|
|
struct mmap_arg_struct a;
|
|
|
|
|
|
|
|
|
|
if (copy_from_user(&a, arg, sizeof(a)))
|
|
|
|
|
return -EFAULT;
|
2015-11-06 10:46:54 +08:00
|
|
|
if (offset_in_page(a.offset))
|
2010-03-11 07:21:15 +08:00
|
|
|
return -EINVAL;
|
|
|
|
|
|
2018-03-11 18:34:46 +08:00
|
|
|
return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
|
|
|
|
|
a.offset >> PAGE_SHIFT);
|
2010-03-11 07:21:15 +08:00
|
|
|
}
|
|
|
|
|
#endif /* __ARCH_WANT_SYS_OLD_MMAP */
|
|
|
|
|
|
2007-07-30 06:36:13 +08:00
|
|
|
/*
|
2019-03-06 07:46:22 +08:00
|
|
|
* Some shared mappings will want the pages marked read-only
|
2007-07-30 06:36:13 +08:00
|
|
|
* to track write events. If so, we'll downgrade vm_page_prot
|
|
|
|
|
* to the private version (using protection_map[] without the
|
|
|
|
|
* VM_SHARED bit).
|
|
|
|
|
*/
|
2016-10-08 08:01:22 +08:00
|
|
|
int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
|
2007-07-30 06:36:13 +08:00
|
|
|
{
|
2011-05-26 18:16:19 +08:00
|
|
|
vm_flags_t vm_flags = vma->vm_flags;
|
2015-09-23 05:59:12 +08:00
|
|
|
const struct vm_operations_struct *vm_ops = vma->vm_ops;
|
2007-07-30 06:36:13 +08:00
|
|
|
|
|
|
|
|
/* If it was private or non-writable, the write bit is already clear */
|
|
|
|
|
if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
|
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
|
|
/* The backer wishes to know when pages are first written to? */
|
2015-09-23 05:59:12 +08:00
|
|
|
if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
|
2007-07-30 06:36:13 +08:00
|
|
|
return 1;
|
|
|
|
|
|
mm: softdirty: enable write notifications on VMAs after VM_SOFTDIRTY cleared
For VMAs that don't want write notifications, PTEs created for read faults
have their write bit set. If the read fault happens after VM_SOFTDIRTY is
cleared, then the PTE's softdirty bit will remain clear after subsequent
writes.
Here's a simple code snippet to demonstrate the bug:
char* m = mmap(NULL, getpagesize(), PROT_READ | PROT_WRITE,
MAP_ANONYMOUS | MAP_SHARED, -1, 0);
system("echo 4 > /proc/$PPID/clear_refs"); /* clear VM_SOFTDIRTY */
assert(*m == '\0'); /* new PTE allows write access */
assert(!soft_dirty(x));
*m = 'x'; /* should dirty the page */
assert(soft_dirty(x)); /* fails */
With this patch, write notifications are enabled when VM_SOFTDIRTY is
cleared. Furthermore, to avoid unnecessary faults, write notifications
are disabled when VM_SOFTDIRTY is set.
As a side effect of enabling and disabling write notifications with
care, this patch fixes a bug in mprotect where vm_page_prot bits set by
drivers were zapped on mprotect. An analogous bug was fixed in mmap by
commit c9d0bf241451 ("mm: uncached vma support with writenotify").
Signed-off-by: Peter Feiner <pfeiner@google.com>
Reported-by: Peter Feiner <pfeiner@google.com>
Suggested-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Jamie Liu <jamieliu@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Bjorn Helgaas <bhelgaas@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-10-14 06:55:46 +08:00
|
|
|
/* The open routine did something to the protections that pgprot_modify
|
|
|
|
|
* won't preserve? */
|
2016-10-08 08:01:22 +08:00
|
|
|
if (pgprot_val(vm_page_prot) !=
|
|
|
|
|
pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
|
2007-07-30 06:36:13 +08:00
|
|
|
return 0;
|
|
|
|
|
|
mm: softdirty: enable write notifications on VMAs after VM_SOFTDIRTY cleared
For VMAs that don't want write notifications, PTEs created for read faults
have their write bit set. If the read fault happens after VM_SOFTDIRTY is
cleared, then the PTE's softdirty bit will remain clear after subsequent
writes.
Here's a simple code snippet to demonstrate the bug:
char* m = mmap(NULL, getpagesize(), PROT_READ | PROT_WRITE,
MAP_ANONYMOUS | MAP_SHARED, -1, 0);
system("echo 4 > /proc/$PPID/clear_refs"); /* clear VM_SOFTDIRTY */
assert(*m == '\0'); /* new PTE allows write access */
assert(!soft_dirty(x));
*m = 'x'; /* should dirty the page */
assert(soft_dirty(x)); /* fails */
With this patch, write notifications are enabled when VM_SOFTDIRTY is
cleared. Furthermore, to avoid unnecessary faults, write notifications
are disabled when VM_SOFTDIRTY is set.
As a side effect of enabling and disabling write notifications with
care, this patch fixes a bug in mprotect where vm_page_prot bits set by
drivers were zapped on mprotect. An analogous bug was fixed in mmap by
commit c9d0bf241451 ("mm: uncached vma support with writenotify").
Signed-off-by: Peter Feiner <pfeiner@google.com>
Reported-by: Peter Feiner <pfeiner@google.com>
Suggested-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Jamie Liu <jamieliu@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Bjorn Helgaas <bhelgaas@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-10-14 06:55:46 +08:00
|
|
|
/* Do we need to track softdirty? */
|
|
|
|
|
if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
|
|
|
|
|
return 1;
|
|
|
|
|
|
2007-07-30 06:36:13 +08:00
|
|
|
/* Specialty mapping? */
|
2012-10-09 07:28:40 +08:00
|
|
|
if (vm_flags & VM_PFNMAP)
|
2007-07-30 06:36:13 +08:00
|
|
|
return 0;
|
|
|
|
|
|
|
|
|
|
/* Can the mapping track the dirty pages? */
|
|
|
|
|
return vma->vm_file && vma->vm_file->f_mapping &&
|
2020-09-24 14:51:40 +08:00
|
|
|
mapping_can_writeback(vma->vm_file->f_mapping);
|
2007-07-30 06:36:13 +08:00
|
|
|
}
|
|
|
|
|
|
2009-02-01 07:08:56 +08:00
|
|
|
/*
|
|
|
|
|
* We account for memory if it's a private writeable mapping,
|
2009-02-10 22:02:27 +08:00
|
|
|
* not hugepages and VM_NORESERVE wasn't set.
|
2009-02-01 07:08:56 +08:00
|
|
|
*/
|
2011-05-26 18:16:19 +08:00
|
|
|
static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
|
2009-02-01 07:08:56 +08:00
|
|
|
{
|
2009-02-10 22:02:27 +08:00
|
|
|
/*
|
|
|
|
|
* hugetlb has its own accounting separate from the core VM
|
|
|
|
|
* VM_HUGETLB may not be set yet so we cannot check for that flag.
|
|
|
|
|
*/
|
|
|
|
|
if (file && is_file_hugepages(file))
|
|
|
|
|
return 0;
|
|
|
|
|
|
2009-02-01 07:08:56 +08:00
|
|
|
return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
|
|
|
|
|
}
|
|
|
|
|
|
2007-07-16 14:38:26 +08:00
|
|
|
unsigned long mmap_region(struct file *file, unsigned long addr,
|
2017-02-25 06:58:22 +08:00
|
|
|
unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
|
|
|
|
|
struct list_head *uf)
|
2007-07-16 14:38:26 +08:00
|
|
|
{
|
|
|
|
|
struct mm_struct *mm = current->mm;
|
2020-08-07 14:23:34 +08:00
|
|
|
struct vm_area_struct *vma, *prev, *merge;
|
2007-07-16 14:38:26 +08:00
|
|
|
int error;
|
|
|
|
|
struct rb_node **rb_link, *rb_parent;
|
|
|
|
|
unsigned long charged = 0;
|
|
|
|
|
|
2013-04-30 06:08:33 +08:00
|
|
|
/* Check against address space limit. */
|
2016-01-15 07:22:07 +08:00
|
|
|
if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
|
2013-04-30 06:08:33 +08:00
|
|
|
unsigned long nr_pages;
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* MAP_FIXED may remove pages of mappings that intersects with
|
|
|
|
|
* requested mapping. Account for the pages it would unmap.
|
|
|
|
|
*/
|
|
|
|
|
nr_pages = count_vma_pages_range(mm, addr, addr + len);
|
|
|
|
|
|
2016-01-15 07:22:07 +08:00
|
|
|
if (!may_expand_vm(mm, vm_flags,
|
|
|
|
|
(len >> PAGE_SHIFT) - nr_pages))
|
2013-04-30 06:08:33 +08:00
|
|
|
return -ENOMEM;
|
|
|
|
|
}
|
|
|
|
|
|
2020-10-18 07:14:09 +08:00
|
|
|
/* Clear old maps, set up prev, rb_link, rb_parent, and uf */
|
|
|
|
|
if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf))
|
|
|
|
|
return -ENOMEM;
|
2009-02-01 07:08:56 +08:00
|
|
|
/*
|
|
|
|
|
* Private writable mapping: check memory availability
|
|
|
|
|
*/
|
2009-02-10 22:02:27 +08:00
|
|
|
if (accountable_mapping(file, vm_flags)) {
|
2009-02-01 07:08:56 +08:00
|
|
|
charged = len >> PAGE_SHIFT;
|
2012-02-13 11:58:52 +08:00
|
|
|
if (security_vm_enough_memory_mm(mm, charged))
|
2009-02-01 07:08:56 +08:00
|
|
|
return -ENOMEM;
|
|
|
|
|
vm_flags |= VM_ACCOUNT;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
2009-01-30 09:46:42 +08:00
|
|
|
* Can we just expand an old mapping?
|
2005-04-17 06:20:36 +08:00
|
|
|
*/
|
2015-09-05 06:46:24 +08:00
|
|
|
vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
|
mm: add a field to store names for private anonymous memory
In many userspace applications, and especially in VM based applications
like Android uses heavily, there are multiple different allocators in
use. At a minimum there is libc malloc and the stack, and in many cases
there are libc malloc, the stack, direct syscalls to mmap anonymous
memory, and multiple VM heaps (one for small objects, one for big
objects, etc.). Each of these layers usually has its own tools to
inspect its usage; malloc by compiling a debug version, the VM through
heap inspection tools, and for direct syscalls there is usually no way
to track them.
On Android we heavily use a set of tools that use an extended version of
the logic covered in Documentation/vm/pagemap.txt to walk all pages
mapped in userspace and slice their usage by process, shared (COW) vs.
unique mappings, backing, etc. This can account for real physical
memory usage even in cases like fork without exec (which Android uses
heavily to share as many private COW pages as possible between
processes), Kernel SamePage Merging, and clean zero pages. It produces
a measurement of the pages that only exist in that process (USS, for
unique), and a measurement of the physical memory usage of that process
with the cost of shared pages being evenly split between processes that
share them (PSS).
If all anonymous memory is indistinguishable then figuring out the real
physical memory usage (PSS) of each heap requires either a pagemap
walking tool that can understand the heap debugging of every layer, or
for every layer's heap debugging tools to implement the pagemap walking
logic, in which case it is hard to get a consistent view of memory
across the whole system.
Tracking the information in userspace leads to all sorts of problems.
It either needs to be stored inside the process, which means every
process has to have an API to export its current heap information upon
request, or it has to be stored externally in a filesystem that somebody
needs to clean up on crashes. It needs to be readable while the process
is still running, so it has to have some sort of synchronization with
every layer of userspace. Efficiently tracking the ranges requires
reimplementing something like the kernel vma trees, and linking to it
from every layer of userspace. It requires more memory, more syscalls,
more runtime cost, and more complexity to separately track regions that
the kernel is already tracking.
This patch adds a field to /proc/pid/maps and /proc/pid/smaps to show a
userspace-provided name for anonymous vmas. The names of named
anonymous vmas are shown in /proc/pid/maps and /proc/pid/smaps as
[anon:<name>].
Userspace can set the name for a region of memory by calling
prctl(PR_SET_VMA, PR_SET_VMA_ANON_NAME, start, len, (unsigned long)name)
Setting the name to NULL clears it. The name length limit is 80 bytes
including NUL-terminator and is checked to contain only printable ascii
characters (including space), except '[',']','\','$' and '`'.
Ascii strings are being used to have a descriptive identifiers for vmas,
which can be understood by the users reading /proc/pid/maps or
/proc/pid/smaps. Names can be standardized for a given system and they
can include some variable parts such as the name of the allocator or a
library, tid of the thread using it, etc.
The name is stored in a pointer in the shared union in vm_area_struct
that points to a null terminated string. Anonymous vmas with the same
name (equivalent strings) and are otherwise mergeable will be merged.
The name pointers are not shared between vmas even if they contain the
same name. The name pointer is stored in a union with fields that are
only used on file-backed mappings, so it does not increase memory usage.
CONFIG_ANON_VMA_NAME kernel configuration is introduced to enable this
feature. It keeps the feature disabled by default to prevent any
additional memory overhead and to avoid confusing procfs parsers on
systems which are not ready to support named anonymous vmas.
The patch is based on the original patch developed by Colin Cross, more
specifically on its latest version [1] posted upstream by Sumit Semwal.
It used a userspace pointer to store vma names. In that design, name
pointers could be shared between vmas. However during the last
upstreaming attempt, Kees Cook raised concerns [2] about this approach
and suggested to copy the name into kernel memory space, perform
validity checks [3] and store as a string referenced from
vm_area_struct.
One big concern is about fork() performance which would need to strdup
anonymous vma names. Dave Hansen suggested experimenting with
worst-case scenario of forking a process with 64k vmas having longest
possible names [4]. I ran this experiment on an ARM64 Android device
and recorded a worst-case regression of almost 40% when forking such a
process.
This regression is addressed in the followup patch which replaces the
pointer to a name with a refcounted structure that allows sharing the
name pointer between vmas of the same name. Instead of duplicating the
string during fork() or when splitting a vma it increments the refcount.
[1] https://lore.kernel.org/linux-mm/20200901161459.11772-4-sumit.semwal@linaro.org/
[2] https://lore.kernel.org/linux-mm/202009031031.D32EF57ED@keescook/
[3] https://lore.kernel.org/linux-mm/202009031022.3834F692@keescook/
[4] https://lore.kernel.org/linux-mm/5d0358ab-8c47-2f5f-8e43-23b89d6a8e95@intel.com/
Changes for prctl(2) manual page (in the options section):
PR_SET_VMA
Sets an attribute specified in arg2 for virtual memory areas
starting from the address specified in arg3 and spanning the
size specified in arg4. arg5 specifies the value of the attribute
to be set. Note that assigning an attribute to a virtual memory
area might prevent it from being merged with adjacent virtual
memory areas due to the difference in that attribute's value.
Currently, arg2 must be one of:
PR_SET_VMA_ANON_NAME
Set a name for anonymous virtual memory areas. arg5 should
be a pointer to a null-terminated string containing the
name. The name length including null byte cannot exceed
80 bytes. If arg5 is NULL, the name of the appropriate
anonymous virtual memory areas will be reset. The name
can contain only printable ascii characters (including
space), except '[',']','\','$' and '`'.
This feature is available only if the kernel is built with
the CONFIG_ANON_VMA_NAME option enabled.
[surenb@google.com: docs: proc.rst: /proc/PID/maps: fix malformed table]
Link: https://lkml.kernel.org/r/20211123185928.2513763-1-surenb@google.com
[surenb: rebased over v5.15-rc6, replaced userpointer with a kernel copy,
added input sanitization and CONFIG_ANON_VMA_NAME config. The bulk of the
work here was done by Colin Cross, therefore, with his permission, keeping
him as the author]
Link: https://lkml.kernel.org/r/20211019215511.3771969-2-surenb@google.com
Signed-off-by: Colin Cross <ccross@google.com>
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Kees Cook <keescook@chromium.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Jan Glauber <jan.glauber@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: John Stultz <john.stultz@linaro.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rob Landley <rob@landley.net>
Cc: "Serge E. Hallyn" <serge.hallyn@ubuntu.com>
Cc: Shaohua Li <shli@fusionio.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-15 06:05:59 +08:00
|
|
|
NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX, NULL);
|
2009-01-30 09:46:42 +08:00
|
|
|
if (vma)
|
|
|
|
|
goto out;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Determine the object being mapped and call the appropriate
|
|
|
|
|
* specific mapper. the address has already been validated, but
|
|
|
|
|
* not unmapped, but the maps are removed from the list.
|
|
|
|
|
*/
|
2018-07-22 06:24:03 +08:00
|
|
|
vma = vm_area_alloc(mm);
|
2005-04-17 06:20:36 +08:00
|
|
|
if (!vma) {
|
|
|
|
|
error = -ENOMEM;
|
|
|
|
|
goto unacct_error;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
vma->vm_start = addr;
|
|
|
|
|
vma->vm_end = addr + len;
|
|
|
|
|
vma->vm_flags = vm_flags;
|
2007-10-19 14:39:15 +08:00
|
|
|
vma->vm_page_prot = vm_get_page_prot(vm_flags);
|
2005-04-17 06:20:36 +08:00
|
|
|
vma->vm_pgoff = pgoff;
|
|
|
|
|
|
|
|
|
|
if (file) {
|
2014-08-09 05:25:25 +08:00
|
|
|
if (vm_flags & VM_SHARED) {
|
|
|
|
|
error = mapping_map_writable(file->f_mapping);
|
|
|
|
|
if (error)
|
2021-04-22 18:08:20 +08:00
|
|
|
goto free_vma;
|
2014-08-09 05:25:25 +08:00
|
|
|
}
|
|
|
|
|
|
2012-08-28 02:48:26 +08:00
|
|
|
vma->vm_file = get_file(file);
|
2017-02-20 23:51:23 +08:00
|
|
|
error = call_mmap(file, vma);
|
2005-04-17 06:20:36 +08:00
|
|
|
if (error)
|
|
|
|
|
goto unmap_and_free_vma;
|
2009-09-22 08:03:41 +08:00
|
|
|
|
2020-12-06 14:15:15 +08:00
|
|
|
/* Can addr have changed??
|
|
|
|
|
*
|
|
|
|
|
* Answer: Yes, several device drivers can do it in their
|
|
|
|
|
* f_op->mmap method. -DaveM
|
|
|
|
|
* Bug: If addr is changed, prev, rb_link, rb_parent should
|
|
|
|
|
* be updated for vma_link()
|
|
|
|
|
*/
|
|
|
|
|
WARN_ON_ONCE(addr != vma->vm_start);
|
|
|
|
|
|
|
|
|
|
addr = vma->vm_start;
|
|
|
|
|
|
2020-08-07 14:23:34 +08:00
|
|
|
/* If vm_flags changed after call_mmap(), we should try merge vma again
|
|
|
|
|
* as we may succeed this time.
|
|
|
|
|
*/
|
|
|
|
|
if (unlikely(vm_flags != vma->vm_flags && prev)) {
|
|
|
|
|
merge = vma_merge(mm, prev, vma->vm_start, vma->vm_end, vma->vm_flags,
|
mm: add a field to store names for private anonymous memory
In many userspace applications, and especially in VM based applications
like Android uses heavily, there are multiple different allocators in
use. At a minimum there is libc malloc and the stack, and in many cases
there are libc malloc, the stack, direct syscalls to mmap anonymous
memory, and multiple VM heaps (one for small objects, one for big
objects, etc.). Each of these layers usually has its own tools to
inspect its usage; malloc by compiling a debug version, the VM through
heap inspection tools, and for direct syscalls there is usually no way
to track them.
On Android we heavily use a set of tools that use an extended version of
the logic covered in Documentation/vm/pagemap.txt to walk all pages
mapped in userspace and slice their usage by process, shared (COW) vs.
unique mappings, backing, etc. This can account for real physical
memory usage even in cases like fork without exec (which Android uses
heavily to share as many private COW pages as possible between
processes), Kernel SamePage Merging, and clean zero pages. It produces
a measurement of the pages that only exist in that process (USS, for
unique), and a measurement of the physical memory usage of that process
with the cost of shared pages being evenly split between processes that
share them (PSS).
If all anonymous memory is indistinguishable then figuring out the real
physical memory usage (PSS) of each heap requires either a pagemap
walking tool that can understand the heap debugging of every layer, or
for every layer's heap debugging tools to implement the pagemap walking
logic, in which case it is hard to get a consistent view of memory
across the whole system.
Tracking the information in userspace leads to all sorts of problems.
It either needs to be stored inside the process, which means every
process has to have an API to export its current heap information upon
request, or it has to be stored externally in a filesystem that somebody
needs to clean up on crashes. It needs to be readable while the process
is still running, so it has to have some sort of synchronization with
every layer of userspace. Efficiently tracking the ranges requires
reimplementing something like the kernel vma trees, and linking to it
from every layer of userspace. It requires more memory, more syscalls,
more runtime cost, and more complexity to separately track regions that
the kernel is already tracking.
This patch adds a field to /proc/pid/maps and /proc/pid/smaps to show a
userspace-provided name for anonymous vmas. The names of named
anonymous vmas are shown in /proc/pid/maps and /proc/pid/smaps as
[anon:<name>].
Userspace can set the name for a region of memory by calling
prctl(PR_SET_VMA, PR_SET_VMA_ANON_NAME, start, len, (unsigned long)name)
Setting the name to NULL clears it. The name length limit is 80 bytes
including NUL-terminator and is checked to contain only printable ascii
characters (including space), except '[',']','\','$' and '`'.
Ascii strings are being used to have a descriptive identifiers for vmas,
which can be understood by the users reading /proc/pid/maps or
/proc/pid/smaps. Names can be standardized for a given system and they
can include some variable parts such as the name of the allocator or a
library, tid of the thread using it, etc.
The name is stored in a pointer in the shared union in vm_area_struct
that points to a null terminated string. Anonymous vmas with the same
name (equivalent strings) and are otherwise mergeable will be merged.
The name pointers are not shared between vmas even if they contain the
same name. The name pointer is stored in a union with fields that are
only used on file-backed mappings, so it does not increase memory usage.
CONFIG_ANON_VMA_NAME kernel configuration is introduced to enable this
feature. It keeps the feature disabled by default to prevent any
additional memory overhead and to avoid confusing procfs parsers on
systems which are not ready to support named anonymous vmas.
The patch is based on the original patch developed by Colin Cross, more
specifically on its latest version [1] posted upstream by Sumit Semwal.
It used a userspace pointer to store vma names. In that design, name
pointers could be shared between vmas. However during the last
upstreaming attempt, Kees Cook raised concerns [2] about this approach
and suggested to copy the name into kernel memory space, perform
validity checks [3] and store as a string referenced from
vm_area_struct.
One big concern is about fork() performance which would need to strdup
anonymous vma names. Dave Hansen suggested experimenting with
worst-case scenario of forking a process with 64k vmas having longest
possible names [4]. I ran this experiment on an ARM64 Android device
and recorded a worst-case regression of almost 40% when forking such a
process.
This regression is addressed in the followup patch which replaces the
pointer to a name with a refcounted structure that allows sharing the
name pointer between vmas of the same name. Instead of duplicating the
string during fork() or when splitting a vma it increments the refcount.
[1] https://lore.kernel.org/linux-mm/20200901161459.11772-4-sumit.semwal@linaro.org/
[2] https://lore.kernel.org/linux-mm/202009031031.D32EF57ED@keescook/
[3] https://lore.kernel.org/linux-mm/202009031022.3834F692@keescook/
[4] https://lore.kernel.org/linux-mm/5d0358ab-8c47-2f5f-8e43-23b89d6a8e95@intel.com/
Changes for prctl(2) manual page (in the options section):
PR_SET_VMA
Sets an attribute specified in arg2 for virtual memory areas
starting from the address specified in arg3 and spanning the
size specified in arg4. arg5 specifies the value of the attribute
to be set. Note that assigning an attribute to a virtual memory
area might prevent it from being merged with adjacent virtual
memory areas due to the difference in that attribute's value.
Currently, arg2 must be one of:
PR_SET_VMA_ANON_NAME
Set a name for anonymous virtual memory areas. arg5 should
be a pointer to a null-terminated string containing the
name. The name length including null byte cannot exceed
80 bytes. If arg5 is NULL, the name of the appropriate
anonymous virtual memory areas will be reset. The name
can contain only printable ascii characters (including
space), except '[',']','\','$' and '`'.
This feature is available only if the kernel is built with
the CONFIG_ANON_VMA_NAME option enabled.
[surenb@google.com: docs: proc.rst: /proc/PID/maps: fix malformed table]
Link: https://lkml.kernel.org/r/20211123185928.2513763-1-surenb@google.com
[surenb: rebased over v5.15-rc6, replaced userpointer with a kernel copy,
added input sanitization and CONFIG_ANON_VMA_NAME config. The bulk of the
work here was done by Colin Cross, therefore, with his permission, keeping
him as the author]
Link: https://lkml.kernel.org/r/20211019215511.3771969-2-surenb@google.com
Signed-off-by: Colin Cross <ccross@google.com>
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Kees Cook <keescook@chromium.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Jan Glauber <jan.glauber@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: John Stultz <john.stultz@linaro.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rob Landley <rob@landley.net>
Cc: "Serge E. Hallyn" <serge.hallyn@ubuntu.com>
Cc: Shaohua Li <shli@fusionio.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-15 06:05:59 +08:00
|
|
|
NULL, vma->vm_file, vma->vm_pgoff, NULL, NULL_VM_UFFD_CTX, NULL);
|
2020-08-07 14:23:34 +08:00
|
|
|
if (merge) {
|
2020-10-11 14:16:34 +08:00
|
|
|
/* ->mmap() can change vma->vm_file and fput the original file. So
|
|
|
|
|
* fput the vma->vm_file here or we would add an extra fput for file
|
|
|
|
|
* and cause general protection fault ultimately.
|
|
|
|
|
*/
|
|
|
|
|
fput(vma->vm_file);
|
2020-08-07 14:23:34 +08:00
|
|
|
vm_area_free(vma);
|
|
|
|
|
vma = merge;
|
2020-12-06 14:15:15 +08:00
|
|
|
/* Update vm_flags to pick up the change. */
|
2020-08-07 14:23:34 +08:00
|
|
|
vm_flags = vma->vm_flags;
|
|
|
|
|
goto unmap_writable;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2009-09-22 08:03:41 +08:00
|
|
|
vm_flags = vma->vm_flags;
|
2005-04-17 06:20:36 +08:00
|
|
|
} else if (vm_flags & VM_SHARED) {
|
|
|
|
|
error = shmem_zero_setup(vma);
|
|
|
|
|
if (error)
|
|
|
|
|
goto free_vma;
|
mm: fix vma_is_anonymous() false-positives
vma_is_anonymous() relies on ->vm_ops being NULL to detect anonymous
VMA. This is unreliable as ->mmap may not set ->vm_ops.
False-positive vma_is_anonymous() may lead to crashes:
next ffff8801ce5e7040 prev ffff8801d20eca50 mm ffff88019c1e13c0
prot 27 anon_vma ffff88019680cdd8 vm_ops 0000000000000000
pgoff 0 file ffff8801b2ec2d00 private_data 0000000000000000
flags: 0xff(read|write|exec|shared|mayread|maywrite|mayexec|mayshare)
------------[ cut here ]------------
kernel BUG at mm/memory.c:1422!
invalid opcode: 0000 [#1] SMP KASAN
CPU: 0 PID: 18486 Comm: syz-executor3 Not tainted 4.18.0-rc3+ #136
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google
01/01/2011
RIP: 0010:zap_pmd_range mm/memory.c:1421 [inline]
RIP: 0010:zap_pud_range mm/memory.c:1466 [inline]
RIP: 0010:zap_p4d_range mm/memory.c:1487 [inline]
RIP: 0010:unmap_page_range+0x1c18/0x2220 mm/memory.c:1508
Call Trace:
unmap_single_vma+0x1a0/0x310 mm/memory.c:1553
zap_page_range_single+0x3cc/0x580 mm/memory.c:1644
unmap_mapping_range_vma mm/memory.c:2792 [inline]
unmap_mapping_range_tree mm/memory.c:2813 [inline]
unmap_mapping_pages+0x3a7/0x5b0 mm/memory.c:2845
unmap_mapping_range+0x48/0x60 mm/memory.c:2880
truncate_pagecache+0x54/0x90 mm/truncate.c:800
truncate_setsize+0x70/0xb0 mm/truncate.c:826
simple_setattr+0xe9/0x110 fs/libfs.c:409
notify_change+0xf13/0x10f0 fs/attr.c:335
do_truncate+0x1ac/0x2b0 fs/open.c:63
do_sys_ftruncate+0x492/0x560 fs/open.c:205
__do_sys_ftruncate fs/open.c:215 [inline]
__se_sys_ftruncate fs/open.c:213 [inline]
__x64_sys_ftruncate+0x59/0x80 fs/open.c:213
do_syscall_64+0x1b9/0x820 arch/x86/entry/common.c:290
entry_SYSCALL_64_after_hwframe+0x49/0xbe
Reproducer:
#include <stdio.h>
#include <stddef.h>
#include <stdint.h>
#include <stdlib.h>
#include <string.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <sys/ioctl.h>
#include <sys/mman.h>
#include <unistd.h>
#include <fcntl.h>
#define KCOV_INIT_TRACE _IOR('c', 1, unsigned long)
#define KCOV_ENABLE _IO('c', 100)
#define KCOV_DISABLE _IO('c', 101)
#define COVER_SIZE (1024<<10)
#define KCOV_TRACE_PC 0
#define KCOV_TRACE_CMP 1
int main(int argc, char **argv)
{
int fd;
unsigned long *cover;
system("mount -t debugfs none /sys/kernel/debug");
fd = open("/sys/kernel/debug/kcov", O_RDWR);
ioctl(fd, KCOV_INIT_TRACE, COVER_SIZE);
cover = mmap(NULL, COVER_SIZE * sizeof(unsigned long),
PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
munmap(cover, COVER_SIZE * sizeof(unsigned long));
cover = mmap(NULL, COVER_SIZE * sizeof(unsigned long),
PROT_READ | PROT_WRITE, MAP_PRIVATE, fd, 0);
memset(cover, 0, COVER_SIZE * sizeof(unsigned long));
ftruncate(fd, 3UL << 20);
return 0;
}
This can be fixed by assigning anonymous VMAs own vm_ops and not relying
on it being NULL.
If ->mmap() failed to set ->vm_ops, mmap_region() will set it to
dummy_vm_ops. This way we will have non-NULL ->vm_ops for all VMAs.
Link: http://lkml.kernel.org/r/20180724121139.62570-4-kirill.shutemov@linux.intel.com
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Reported-by: syzbot+3f84280d52be9b7083cc@syzkaller.appspotmail.com
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-07-27 07:37:35 +08:00
|
|
|
} else {
|
|
|
|
|
vma_set_anonymous(vma);
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
|
2019-11-26 01:27:06 +08:00
|
|
|
/* Allow architectures to sanity-check the vm_flags */
|
|
|
|
|
if (!arch_validate_flags(vma->vm_flags)) {
|
|
|
|
|
error = -EINVAL;
|
|
|
|
|
if (file)
|
|
|
|
|
goto unmap_and_free_vma;
|
|
|
|
|
else
|
|
|
|
|
goto free_vma;
|
|
|
|
|
}
|
|
|
|
|
|
2009-01-30 09:46:42 +08:00
|
|
|
vma_link(mm, vma, prev, rb_link, rb_parent);
|
2008-04-28 17:12:10 +08:00
|
|
|
/* Once vma denies write, undo our temporary denial count */
|
2020-08-07 14:23:34 +08:00
|
|
|
unmap_writable:
|
2021-04-22 18:08:20 +08:00
|
|
|
if (file && vm_flags & VM_SHARED)
|
|
|
|
|
mapping_unmap_writable(file->f_mapping);
|
2013-09-12 05:20:20 +08:00
|
|
|
file = vma->vm_file;
|
2008-04-28 17:12:10 +08:00
|
|
|
out:
|
perf: Do the big rename: Performance Counters -> Performance Events
Bye-bye Performance Counters, welcome Performance Events!
In the past few months the perfcounters subsystem has grown out its
initial role of counting hardware events, and has become (and is
becoming) a much broader generic event enumeration, reporting, logging,
monitoring, analysis facility.
Naming its core object 'perf_counter' and naming the subsystem
'perfcounters' has become more and more of a misnomer. With pending
code like hw-breakpoints support the 'counter' name is less and
less appropriate.
All in one, we've decided to rename the subsystem to 'performance
events' and to propagate this rename through all fields, variables
and API names. (in an ABI compatible fashion)
The word 'event' is also a bit shorter than 'counter' - which makes
it slightly more convenient to write/handle as well.
Thanks goes to Stephane Eranian who first observed this misnomer and
suggested a rename.
User-space tooling and ABI compatibility is not affected - this patch
should be function-invariant. (Also, defconfigs were not touched to
keep the size down.)
This patch has been generated via the following script:
FILES=$(find * -type f | grep -vE 'oprofile|[^K]config')
sed -i \
-e 's/PERF_EVENT_/PERF_RECORD_/g' \
-e 's/PERF_COUNTER/PERF_EVENT/g' \
-e 's/perf_counter/perf_event/g' \
-e 's/nb_counters/nb_events/g' \
-e 's/swcounter/swevent/g' \
-e 's/tpcounter_event/tp_event/g' \
$FILES
for N in $(find . -name perf_counter.[ch]); do
M=$(echo $N | sed 's/perf_counter/perf_event/g')
mv $N $M
done
FILES=$(find . -name perf_event.*)
sed -i \
-e 's/COUNTER_MASK/REG_MASK/g' \
-e 's/COUNTER/EVENT/g' \
-e 's/\<event\>/event_id/g' \
-e 's/counter/event/g' \
-e 's/Counter/Event/g' \
$FILES
... to keep it as correct as possible. This script can also be
used by anyone who has pending perfcounters patches - it converts
a Linux kernel tree over to the new naming. We tried to time this
change to the point in time where the amount of pending patches
is the smallest: the end of the merge window.
Namespace clashes were fixed up in a preparatory patch - and some
stylistic fallout will be fixed up in a subsequent patch.
( NOTE: 'counters' are still the proper terminology when we deal
with hardware registers - and these sed scripts are a bit
over-eager in renaming them. I've undone some of that, but
in case there's something left where 'counter' would be
better than 'event' we can undo that on an individual basis
instead of touching an otherwise nicely automated patch. )
Suggested-by: Stephane Eranian <eranian@google.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Paul Mackerras <paulus@samba.org>
Reviewed-by: Arjan van de Ven <arjan@linux.intel.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: David Howells <dhowells@redhat.com>
Cc: Kyle McMartin <kyle@mcmartin.ca>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: <linux-arch@vger.kernel.org>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-09-21 18:02:48 +08:00
|
|
|
perf_event_mmap(vma);
|
2009-03-31 01:07:05 +08:00
|
|
|
|
2016-01-15 07:22:07 +08:00
|
|
|
vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
|
2005-04-17 06:20:36 +08:00
|
|
|
if (vm_flags & VM_LOCKED) {
|
2018-08-18 06:43:40 +08:00
|
|
|
if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
|
|
|
|
|
is_vm_hugetlb_page(vma) ||
|
|
|
|
|
vma == get_gate_vma(current->mm))
|
2015-11-06 10:51:36 +08:00
|
|
|
vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
|
2018-08-18 06:43:40 +08:00
|
|
|
else
|
|
|
|
|
mm->locked_vm += (len >> PAGE_SHIFT);
|
2013-02-23 08:32:37 +08:00
|
|
|
}
|
uprobes, mm, x86: Add the ability to install and remove uprobes breakpoints
Add uprobes support to the core kernel, with x86 support.
This commit adds the kernel facilities, the actual uprobes
user-space ABI and perf probe support comes in later commits.
General design:
Uprobes are maintained in an rb-tree indexed by inode and offset
(the offset here is from the start of the mapping). For a unique
(inode, offset) tuple, there can be at most one uprobe in the
rb-tree.
Since the (inode, offset) tuple identifies a unique uprobe, more
than one user may be interested in the same uprobe. This provides
the ability to connect multiple 'consumers' to the same uprobe.
Each consumer defines a handler and a filter (optional). The
'handler' is run every time the uprobe is hit, if it matches the
'filter' criteria.
The first consumer of a uprobe causes the breakpoint to be
inserted at the specified address and subsequent consumers are
appended to this list. On subsequent probes, the consumer gets
appended to the existing list of consumers. The breakpoint is
removed when the last consumer unregisters. For all other
unregisterations, the consumer is removed from the list of
consumers.
Given a inode, we get a list of the mms that have mapped the
inode. Do the actual registration if mm maps the page where a
probe needs to be inserted/removed.
We use a temporary list to walk through the vmas that map the
inode.
- The number of maps that map the inode, is not known before we
walk the rmap and keeps changing.
- extending vm_area_struct wasn't recommended, it's a
size-critical data structure.
- There can be more than one maps of the inode in the same mm.
We add callbacks to the mmap methods to keep an eye on text vmas
that are of interest to uprobes. When a vma of interest is mapped,
we insert the breakpoint at the right address.
Uprobe works by replacing the instruction at the address defined
by (inode, offset) with the arch specific breakpoint
instruction. We save a copy of the original instruction at the
uprobed address.
This is needed for:
a. executing the instruction out-of-line (xol).
b. instruction analysis for any subsequent fixups.
c. restoring the instruction back when the uprobe is unregistered.
We insert or delete a breakpoint instruction, and this
breakpoint instruction is assumed to be the smallest instruction
available on the platform. For fixed size instruction platforms
this is trivially true, for variable size instruction platforms
the breakpoint instruction is typically the smallest (often a
single byte).
Writing the instruction is done by COWing the page and changing
the instruction during the copy, this even though most platforms
allow atomic writes of the breakpoint instruction. This also
mirrors the behaviour of a ptrace() memory write to a PRIVATE
file map.
The core worker is derived from KSM's replace_page() logic.
In essence, similar to KSM:
a. allocate a new page and copy over contents of the page that
has the uprobed vaddr
b. modify the copy and insert the breakpoint at the required
address
c. switch the original page with the copy containing the
breakpoint
d. flush page tables.
replace_page() is being replicated here because of some minor
changes in the type of pages and also because Hugh Dickins had
plans to improve replace_page() for KSM specific work.
Instruction analysis on x86 is based on instruction decoder and
determines if an instruction can be probed and determines the
necessary fixups after singlestep. Instruction analysis is done
at probe insertion time so that we avoid having to repeat the
same analysis every time a probe is hit.
A lot of code here is due to the improvement/suggestions/inputs
from Peter Zijlstra.
Changelog:
(v10):
- Add code to clear REX.B prefix as suggested by Denys Vlasenko
and Masami Hiramatsu.
(v9):
- Use insn_offset_modrm as suggested by Masami Hiramatsu.
(v7):
Handle comments from Peter Zijlstra:
- Dont take reference to inode. (expect inode to uprobe_register to be sane).
- Use PTR_ERR to set the return value.
- No need to take reference to inode.
- use PTR_ERR to return error value.
- register and uprobe_unregister share code.
(v5):
- Modified del_consumer as per comments from Peter.
- Drop reference to inode before dropping reference to uprobe.
- Use i_size_read(inode) instead of inode->i_size.
- Ensure uprobe->consumers is NULL, before __uprobe_unregister() is called.
- Includes errno.h as recommended by Stephen Rothwell to fix a build issue
on sparc defconfig
- Remove restrictions while unregistering.
- Earlier code leaked inode references under some conditions while
registering/unregistering.
- Continue the vma-rmap walk even if the intermediate vma doesnt
meet the requirements.
- Validate the vma found by find_vma before inserting/removing the
breakpoint
- Call del_consumer under mutex_lock.
- Use hash locks.
- Handle mremap.
- Introduce find_least_offset_node() instead of close match logic in
find_uprobe
- Uprobes no more depends on MM_OWNER; No reference to task_structs
while inserting/removing a probe.
- Uses read_mapping_page instead of grab_cache_page so that the pages
have valid content.
- pass NULL to get_user_pages for the task parameter.
- call SetPageUptodate on the new page allocated in write_opcode.
- fix leaking a reference to the new page under certain conditions.
- Include Instruction Decoder if Uprobes gets defined.
- Remove const attributes for instruction prefix arrays.
- Uses mm_context to know if the application is 32 bit.
Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Also-written-by: Jim Keniston <jkenisto@us.ibm.com>
Reviewed-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Roland McGrath <roland@hack.frob.com>
Cc: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
Cc: Anton Arapov <anton@redhat.com>
Cc: Ananth N Mavinakayanahalli <ananth@in.ibm.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Denys Vlasenko <vda.linux@googlemail.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Linux-mm <linux-mm@kvack.org>
Link: http://lkml.kernel.org/r/20120209092642.GE16600@linux.vnet.ibm.com
[ Made various small edits to the commit log ]
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2012-02-09 17:26:42 +08:00
|
|
|
|
2012-08-20 01:10:42 +08:00
|
|
|
if (file)
|
|
|
|
|
uprobe_mmap(vma);
|
uprobes, mm, x86: Add the ability to install and remove uprobes breakpoints
Add uprobes support to the core kernel, with x86 support.
This commit adds the kernel facilities, the actual uprobes
user-space ABI and perf probe support comes in later commits.
General design:
Uprobes are maintained in an rb-tree indexed by inode and offset
(the offset here is from the start of the mapping). For a unique
(inode, offset) tuple, there can be at most one uprobe in the
rb-tree.
Since the (inode, offset) tuple identifies a unique uprobe, more
than one user may be interested in the same uprobe. This provides
the ability to connect multiple 'consumers' to the same uprobe.
Each consumer defines a handler and a filter (optional). The
'handler' is run every time the uprobe is hit, if it matches the
'filter' criteria.
The first consumer of a uprobe causes the breakpoint to be
inserted at the specified address and subsequent consumers are
appended to this list. On subsequent probes, the consumer gets
appended to the existing list of consumers. The breakpoint is
removed when the last consumer unregisters. For all other
unregisterations, the consumer is removed from the list of
consumers.
Given a inode, we get a list of the mms that have mapped the
inode. Do the actual registration if mm maps the page where a
probe needs to be inserted/removed.
We use a temporary list to walk through the vmas that map the
inode.
- The number of maps that map the inode, is not known before we
walk the rmap and keeps changing.
- extending vm_area_struct wasn't recommended, it's a
size-critical data structure.
- There can be more than one maps of the inode in the same mm.
We add callbacks to the mmap methods to keep an eye on text vmas
that are of interest to uprobes. When a vma of interest is mapped,
we insert the breakpoint at the right address.
Uprobe works by replacing the instruction at the address defined
by (inode, offset) with the arch specific breakpoint
instruction. We save a copy of the original instruction at the
uprobed address.
This is needed for:
a. executing the instruction out-of-line (xol).
b. instruction analysis for any subsequent fixups.
c. restoring the instruction back when the uprobe is unregistered.
We insert or delete a breakpoint instruction, and this
breakpoint instruction is assumed to be the smallest instruction
available on the platform. For fixed size instruction platforms
this is trivially true, for variable size instruction platforms
the breakpoint instruction is typically the smallest (often a
single byte).
Writing the instruction is done by COWing the page and changing
the instruction during the copy, this even though most platforms
allow atomic writes of the breakpoint instruction. This also
mirrors the behaviour of a ptrace() memory write to a PRIVATE
file map.
The core worker is derived from KSM's replace_page() logic.
In essence, similar to KSM:
a. allocate a new page and copy over contents of the page that
has the uprobed vaddr
b. modify the copy and insert the breakpoint at the required
address
c. switch the original page with the copy containing the
breakpoint
d. flush page tables.
replace_page() is being replicated here because of some minor
changes in the type of pages and also because Hugh Dickins had
plans to improve replace_page() for KSM specific work.
Instruction analysis on x86 is based on instruction decoder and
determines if an instruction can be probed and determines the
necessary fixups after singlestep. Instruction analysis is done
at probe insertion time so that we avoid having to repeat the
same analysis every time a probe is hit.
A lot of code here is due to the improvement/suggestions/inputs
from Peter Zijlstra.
Changelog:
(v10):
- Add code to clear REX.B prefix as suggested by Denys Vlasenko
and Masami Hiramatsu.
(v9):
- Use insn_offset_modrm as suggested by Masami Hiramatsu.
(v7):
Handle comments from Peter Zijlstra:
- Dont take reference to inode. (expect inode to uprobe_register to be sane).
- Use PTR_ERR to set the return value.
- No need to take reference to inode.
- use PTR_ERR to return error value.
- register and uprobe_unregister share code.
(v5):
- Modified del_consumer as per comments from Peter.
- Drop reference to inode before dropping reference to uprobe.
- Use i_size_read(inode) instead of inode->i_size.
- Ensure uprobe->consumers is NULL, before __uprobe_unregister() is called.
- Includes errno.h as recommended by Stephen Rothwell to fix a build issue
on sparc defconfig
- Remove restrictions while unregistering.
- Earlier code leaked inode references under some conditions while
registering/unregistering.
- Continue the vma-rmap walk even if the intermediate vma doesnt
meet the requirements.
- Validate the vma found by find_vma before inserting/removing the
breakpoint
- Call del_consumer under mutex_lock.
- Use hash locks.
- Handle mremap.
- Introduce find_least_offset_node() instead of close match logic in
find_uprobe
- Uprobes no more depends on MM_OWNER; No reference to task_structs
while inserting/removing a probe.
- Uses read_mapping_page instead of grab_cache_page so that the pages
have valid content.
- pass NULL to get_user_pages for the task parameter.
- call SetPageUptodate on the new page allocated in write_opcode.
- fix leaking a reference to the new page under certain conditions.
- Include Instruction Decoder if Uprobes gets defined.
- Remove const attributes for instruction prefix arrays.
- Uses mm_context to know if the application is 32 bit.
Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Also-written-by: Jim Keniston <jkenisto@us.ibm.com>
Reviewed-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Roland McGrath <roland@hack.frob.com>
Cc: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
Cc: Anton Arapov <anton@redhat.com>
Cc: Ananth N Mavinakayanahalli <ananth@in.ibm.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Denys Vlasenko <vda.linux@googlemail.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Linux-mm <linux-mm@kvack.org>
Link: http://lkml.kernel.org/r/20120209092642.GE16600@linux.vnet.ibm.com
[ Made various small edits to the commit log ]
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2012-02-09 17:26:42 +08:00
|
|
|
|
2013-09-12 05:22:24 +08:00
|
|
|
/*
|
|
|
|
|
* New (or expanded) vma always get soft dirty status.
|
|
|
|
|
* Otherwise user-space soft-dirty page tracker won't
|
|
|
|
|
* be able to distinguish situation when vma area unmapped,
|
|
|
|
|
* then new mapped in-place (which must be aimed as
|
|
|
|
|
* a completely new data area).
|
|
|
|
|
*/
|
|
|
|
|
vma->vm_flags |= VM_SOFTDIRTY;
|
|
|
|
|
|
mm: softdirty: enable write notifications on VMAs after VM_SOFTDIRTY cleared
For VMAs that don't want write notifications, PTEs created for read faults
have their write bit set. If the read fault happens after VM_SOFTDIRTY is
cleared, then the PTE's softdirty bit will remain clear after subsequent
writes.
Here's a simple code snippet to demonstrate the bug:
char* m = mmap(NULL, getpagesize(), PROT_READ | PROT_WRITE,
MAP_ANONYMOUS | MAP_SHARED, -1, 0);
system("echo 4 > /proc/$PPID/clear_refs"); /* clear VM_SOFTDIRTY */
assert(*m == '\0'); /* new PTE allows write access */
assert(!soft_dirty(x));
*m = 'x'; /* should dirty the page */
assert(soft_dirty(x)); /* fails */
With this patch, write notifications are enabled when VM_SOFTDIRTY is
cleared. Furthermore, to avoid unnecessary faults, write notifications
are disabled when VM_SOFTDIRTY is set.
As a side effect of enabling and disabling write notifications with
care, this patch fixes a bug in mprotect where vm_page_prot bits set by
drivers were zapped on mprotect. An analogous bug was fixed in mmap by
commit c9d0bf241451 ("mm: uncached vma support with writenotify").
Signed-off-by: Peter Feiner <pfeiner@google.com>
Reported-by: Peter Feiner <pfeiner@google.com>
Suggested-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Jamie Liu <jamieliu@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Bjorn Helgaas <bhelgaas@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-10-14 06:55:46 +08:00
|
|
|
vma_set_page_prot(vma);
|
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
return addr;
|
|
|
|
|
|
|
|
|
|
unmap_and_free_vma:
|
2020-10-09 21:08:55 +08:00
|
|
|
fput(vma->vm_file);
|
2005-04-17 06:20:36 +08:00
|
|
|
vma->vm_file = NULL;
|
|
|
|
|
|
|
|
|
|
/* Undo any partial mapping done by a device driver. */
|
[PATCH] freepgt: free_pgtables use vma list
Recent woes with some arches needing their own pgd_addr_end macro; and 4-level
clear_page_range regression since 2.6.10's clear_page_tables; and its
long-standing well-known inefficiency in searching throughout the higher-level
page tables for those few entries to clear and free: all can be blamed on
ignoring the list of vmas when we free page tables.
Replace exit_mmap's clear_page_range of the total user address space by
free_pgtables operating on the mm's vma list; unmap_region use it in the same
way, giving floor and ceiling beyond which it may not free tables. This
brings lmbench fork/exec/sh numbers back to 2.6.10 (unless preempt is enabled,
in which case latency fixes spoil unmap_vmas throughput).
Beware: the do_mmap_pgoff driver failure case must now use unmap_region
instead of zap_page_range, since a page table might have been allocated, and
can only be freed while it is touched by some vma.
Move free_pgtables from mmap.c to memory.c, where its lower levels are adapted
from the clear_page_range levels. (Most of free_pgtables' old code was
actually for a non-existent case, prev not properly set up, dating from before
hch gave us split_vma.) Pass mmu_gather** in the public interfaces, since we
might want to add latency lockdrops later; but no attempt to do so yet, going
by vma should itself reduce latency.
But what if is_hugepage_only_range? Those ia64 and ppc64 cases need careful
examination: put that off until a later patch of the series.
What of x86_64's 32bit vdso page __map_syscall32 maps outside any vma?
And the range to sparc64's flush_tlb_pgtables? It's less clear to me now that
we need to do more than is done here - every PMD_SIZE ever occupied will be
flushed, do we really have to flush every PGDIR_SIZE ever partially occupied?
A shame to complicate it unnecessarily.
Special thanks to David Miller for time spent repairing my ceilings.
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-04-20 04:29:15 +08:00
|
|
|
unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
|
|
|
|
|
charged = 0;
|
2014-08-09 05:25:25 +08:00
|
|
|
if (vm_flags & VM_SHARED)
|
|
|
|
|
mapping_unmap_writable(file->f_mapping);
|
2005-04-17 06:20:36 +08:00
|
|
|
free_vma:
|
2018-07-22 04:48:51 +08:00
|
|
|
vm_area_free(vma);
|
2005-04-17 06:20:36 +08:00
|
|
|
unacct_error:
|
|
|
|
|
if (charged)
|
|
|
|
|
vm_unacct_memory(charged);
|
|
|
|
|
return error;
|
|
|
|
|
}
|
|
|
|
|
|
2020-04-02 12:09:10 +08:00
|
|
|
static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
|
2012-12-12 08:01:49 +08:00
|
|
|
{
|
|
|
|
|
/*
|
|
|
|
|
* We implement the search by looking for an rbtree node that
|
|
|
|
|
* immediately follows a suitable gap. That is,
|
|
|
|
|
* - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
|
|
|
|
|
* - gap_end = vma->vm_start >= info->low_limit + length;
|
|
|
|
|
* - gap_end - gap_start >= length
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
struct mm_struct *mm = current->mm;
|
|
|
|
|
struct vm_area_struct *vma;
|
|
|
|
|
unsigned long length, low_limit, high_limit, gap_start, gap_end;
|
|
|
|
|
|
|
|
|
|
/* Adjust search length to account for worst case alignment overhead */
|
|
|
|
|
length = info->length + info->align_mask;
|
|
|
|
|
if (length < info->length)
|
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
|
|
/* Adjust search limits by the desired length */
|
|
|
|
|
if (info->high_limit < length)
|
|
|
|
|
return -ENOMEM;
|
|
|
|
|
high_limit = info->high_limit - length;
|
|
|
|
|
|
|
|
|
|
if (info->low_limit > high_limit)
|
|
|
|
|
return -ENOMEM;
|
|
|
|
|
low_limit = info->low_limit + length;
|
|
|
|
|
|
|
|
|
|
/* Check if rbtree root looks promising */
|
|
|
|
|
if (RB_EMPTY_ROOT(&mm->mm_rb))
|
|
|
|
|
goto check_highest;
|
|
|
|
|
vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
|
|
|
|
|
if (vma->rb_subtree_gap < length)
|
|
|
|
|
goto check_highest;
|
|
|
|
|
|
|
|
|
|
while (true) {
|
|
|
|
|
/* Visit left subtree if it looks promising */
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
gap_end = vm_start_gap(vma);
|
2012-12-12 08:01:49 +08:00
|
|
|
if (gap_end >= low_limit && vma->vm_rb.rb_left) {
|
|
|
|
|
struct vm_area_struct *left =
|
|
|
|
|
rb_entry(vma->vm_rb.rb_left,
|
|
|
|
|
struct vm_area_struct, vm_rb);
|
|
|
|
|
if (left->rb_subtree_gap >= length) {
|
|
|
|
|
vma = left;
|
|
|
|
|
continue;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
|
2012-12-12 08:01:49 +08:00
|
|
|
check_current:
|
|
|
|
|
/* Check if current node has a suitable gap */
|
|
|
|
|
if (gap_start > high_limit)
|
|
|
|
|
return -ENOMEM;
|
2017-06-20 17:10:44 +08:00
|
|
|
if (gap_end >= low_limit &&
|
|
|
|
|
gap_end > gap_start && gap_end - gap_start >= length)
|
2012-12-12 08:01:49 +08:00
|
|
|
goto found;
|
|
|
|
|
|
|
|
|
|
/* Visit right subtree if it looks promising */
|
|
|
|
|
if (vma->vm_rb.rb_right) {
|
|
|
|
|
struct vm_area_struct *right =
|
|
|
|
|
rb_entry(vma->vm_rb.rb_right,
|
|
|
|
|
struct vm_area_struct, vm_rb);
|
|
|
|
|
if (right->rb_subtree_gap >= length) {
|
|
|
|
|
vma = right;
|
|
|
|
|
continue;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Go back up the rbtree to find next candidate node */
|
|
|
|
|
while (true) {
|
|
|
|
|
struct rb_node *prev = &vma->vm_rb;
|
|
|
|
|
if (!rb_parent(prev))
|
|
|
|
|
goto check_highest;
|
|
|
|
|
vma = rb_entry(rb_parent(prev),
|
|
|
|
|
struct vm_area_struct, vm_rb);
|
|
|
|
|
if (prev == vma->vm_rb.rb_left) {
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
gap_start = vm_end_gap(vma->vm_prev);
|
|
|
|
|
gap_end = vm_start_gap(vma);
|
2012-12-12 08:01:49 +08:00
|
|
|
goto check_current;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
check_highest:
|
|
|
|
|
/* Check highest gap, which does not precede any rbtree node */
|
|
|
|
|
gap_start = mm->highest_vm_end;
|
|
|
|
|
gap_end = ULONG_MAX; /* Only for VM_BUG_ON below */
|
|
|
|
|
if (gap_start > high_limit)
|
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
|
|
found:
|
|
|
|
|
/* We found a suitable gap. Clip it with the original low_limit. */
|
|
|
|
|
if (gap_start < info->low_limit)
|
|
|
|
|
gap_start = info->low_limit;
|
|
|
|
|
|
|
|
|
|
/* Adjust gap address to the desired alignment */
|
|
|
|
|
gap_start += (info->align_offset - gap_start) & info->align_mask;
|
|
|
|
|
|
|
|
|
|
VM_BUG_ON(gap_start + info->length > info->high_limit);
|
|
|
|
|
VM_BUG_ON(gap_start + info->length > gap_end);
|
|
|
|
|
return gap_start;
|
|
|
|
|
}
|
|
|
|
|
|
2020-04-02 12:09:10 +08:00
|
|
|
static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
|
2012-12-12 08:01:49 +08:00
|
|
|
{
|
|
|
|
|
struct mm_struct *mm = current->mm;
|
|
|
|
|
struct vm_area_struct *vma;
|
|
|
|
|
unsigned long length, low_limit, high_limit, gap_start, gap_end;
|
|
|
|
|
|
|
|
|
|
/* Adjust search length to account for worst case alignment overhead */
|
|
|
|
|
length = info->length + info->align_mask;
|
|
|
|
|
if (length < info->length)
|
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Adjust search limits by the desired length.
|
|
|
|
|
* See implementation comment at top of unmapped_area().
|
|
|
|
|
*/
|
|
|
|
|
gap_end = info->high_limit;
|
|
|
|
|
if (gap_end < length)
|
|
|
|
|
return -ENOMEM;
|
|
|
|
|
high_limit = gap_end - length;
|
|
|
|
|
|
|
|
|
|
if (info->low_limit > high_limit)
|
|
|
|
|
return -ENOMEM;
|
|
|
|
|
low_limit = info->low_limit + length;
|
|
|
|
|
|
|
|
|
|
/* Check highest gap, which does not precede any rbtree node */
|
|
|
|
|
gap_start = mm->highest_vm_end;
|
|
|
|
|
if (gap_start <= high_limit)
|
|
|
|
|
goto found_highest;
|
|
|
|
|
|
|
|
|
|
/* Check if rbtree root looks promising */
|
|
|
|
|
if (RB_EMPTY_ROOT(&mm->mm_rb))
|
|
|
|
|
return -ENOMEM;
|
|
|
|
|
vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
|
|
|
|
|
if (vma->rb_subtree_gap < length)
|
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
|
|
while (true) {
|
|
|
|
|
/* Visit right subtree if it looks promising */
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
|
2012-12-12 08:01:49 +08:00
|
|
|
if (gap_start <= high_limit && vma->vm_rb.rb_right) {
|
|
|
|
|
struct vm_area_struct *right =
|
|
|
|
|
rb_entry(vma->vm_rb.rb_right,
|
|
|
|
|
struct vm_area_struct, vm_rb);
|
|
|
|
|
if (right->rb_subtree_gap >= length) {
|
|
|
|
|
vma = right;
|
|
|
|
|
continue;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
check_current:
|
|
|
|
|
/* Check if current node has a suitable gap */
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
gap_end = vm_start_gap(vma);
|
2012-12-12 08:01:49 +08:00
|
|
|
if (gap_end < low_limit)
|
|
|
|
|
return -ENOMEM;
|
2017-06-20 17:10:44 +08:00
|
|
|
if (gap_start <= high_limit &&
|
|
|
|
|
gap_end > gap_start && gap_end - gap_start >= length)
|
2012-12-12 08:01:49 +08:00
|
|
|
goto found;
|
|
|
|
|
|
|
|
|
|
/* Visit left subtree if it looks promising */
|
|
|
|
|
if (vma->vm_rb.rb_left) {
|
|
|
|
|
struct vm_area_struct *left =
|
|
|
|
|
rb_entry(vma->vm_rb.rb_left,
|
|
|
|
|
struct vm_area_struct, vm_rb);
|
|
|
|
|
if (left->rb_subtree_gap >= length) {
|
|
|
|
|
vma = left;
|
|
|
|
|
continue;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Go back up the rbtree to find next candidate node */
|
|
|
|
|
while (true) {
|
|
|
|
|
struct rb_node *prev = &vma->vm_rb;
|
|
|
|
|
if (!rb_parent(prev))
|
|
|
|
|
return -ENOMEM;
|
|
|
|
|
vma = rb_entry(rb_parent(prev),
|
|
|
|
|
struct vm_area_struct, vm_rb);
|
|
|
|
|
if (prev == vma->vm_rb.rb_right) {
|
|
|
|
|
gap_start = vma->vm_prev ?
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
vm_end_gap(vma->vm_prev) : 0;
|
2012-12-12 08:01:49 +08:00
|
|
|
goto check_current;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
found:
|
|
|
|
|
/* We found a suitable gap. Clip it with the original high_limit. */
|
|
|
|
|
if (gap_end > info->high_limit)
|
|
|
|
|
gap_end = info->high_limit;
|
|
|
|
|
|
|
|
|
|
found_highest:
|
|
|
|
|
/* Compute highest gap address at the desired alignment */
|
|
|
|
|
gap_end -= info->length;
|
|
|
|
|
gap_end -= (gap_end - info->align_offset) & info->align_mask;
|
|
|
|
|
|
|
|
|
|
VM_BUG_ON(gap_end < info->low_limit);
|
|
|
|
|
VM_BUG_ON(gap_end < gap_start);
|
|
|
|
|
return gap_end;
|
|
|
|
|
}
|
|
|
|
|
|
2020-04-02 12:09:10 +08:00
|
|
|
/*
|
|
|
|
|
* Search for an unmapped address range.
|
|
|
|
|
*
|
|
|
|
|
* We are looking for a range that:
|
|
|
|
|
* - does not intersect with any VMA;
|
|
|
|
|
* - is contained within the [low_limit, high_limit) interval;
|
|
|
|
|
* - is at least the desired size.
|
|
|
|
|
* - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
|
|
|
|
|
*/
|
|
|
|
|
unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
|
|
|
|
|
{
|
2020-04-02 12:09:13 +08:00
|
|
|
unsigned long addr;
|
|
|
|
|
|
2020-04-02 12:09:10 +08:00
|
|
|
if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
|
2020-04-02 12:09:13 +08:00
|
|
|
addr = unmapped_area_topdown(info);
|
2020-04-02 12:09:10 +08:00
|
|
|
else
|
2020-04-02 12:09:13 +08:00
|
|
|
addr = unmapped_area(info);
|
|
|
|
|
|
|
|
|
|
trace_vm_unmapped_area(addr, info);
|
|
|
|
|
return addr;
|
2020-04-02 12:09:10 +08:00
|
|
|
}
|
2018-12-07 06:50:36 +08:00
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
/* Get an address range which is currently unmapped.
|
|
|
|
|
* For shmat() with addr=0.
|
|
|
|
|
*
|
|
|
|
|
* Ugly calling convention alert:
|
|
|
|
|
* Return value with the low bits set means error value,
|
|
|
|
|
* ie
|
|
|
|
|
* if (ret & ~PAGE_MASK)
|
|
|
|
|
* error = ret;
|
|
|
|
|
*
|
|
|
|
|
* This function "knows" that -ENOMEM has the bits set.
|
|
|
|
|
*/
|
|
|
|
|
#ifndef HAVE_ARCH_UNMAPPED_AREA
|
|
|
|
|
unsigned long
|
|
|
|
|
arch_get_unmapped_area(struct file *filp, unsigned long addr,
|
|
|
|
|
unsigned long len, unsigned long pgoff, unsigned long flags)
|
|
|
|
|
{
|
|
|
|
|
struct mm_struct *mm = current->mm;
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
struct vm_area_struct *vma, *prev;
|
2012-12-12 08:01:49 +08:00
|
|
|
struct vm_unmapped_area_info info;
|
2018-12-07 06:50:36 +08:00
|
|
|
const unsigned long mmap_end = arch_get_mmap_end(addr);
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2018-12-07 06:50:36 +08:00
|
|
|
if (len > mmap_end - mmap_min_addr)
|
2005-04-17 06:20:36 +08:00
|
|
|
return -ENOMEM;
|
|
|
|
|
|
2007-05-07 05:50:13 +08:00
|
|
|
if (flags & MAP_FIXED)
|
|
|
|
|
return addr;
|
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
if (addr) {
|
|
|
|
|
addr = PAGE_ALIGN(addr);
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
vma = find_vma_prev(mm, addr, &prev);
|
2018-12-07 06:50:36 +08:00
|
|
|
if (mmap_end - len >= addr && addr >= mmap_min_addr &&
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
(!vma || addr + len <= vm_start_gap(vma)) &&
|
|
|
|
|
(!prev || addr >= vm_end_gap(prev)))
|
2005-04-17 06:20:36 +08:00
|
|
|
return addr;
|
|
|
|
|
}
|
|
|
|
|
|
2012-12-12 08:01:49 +08:00
|
|
|
info.flags = 0;
|
|
|
|
|
info.length = len;
|
2013-11-13 07:07:54 +08:00
|
|
|
info.low_limit = mm->mmap_base;
|
2018-12-07 06:50:36 +08:00
|
|
|
info.high_limit = mmap_end;
|
2012-12-12 08:01:49 +08:00
|
|
|
info.align_mask = 0;
|
2020-04-11 05:32:48 +08:00
|
|
|
info.align_offset = 0;
|
2012-12-12 08:01:49 +08:00
|
|
|
return vm_unmapped_area(&info);
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
2014-10-10 06:26:29 +08:00
|
|
|
#endif
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* This mmap-allocator allocates new areas top-down from below the
|
|
|
|
|
* stack's low limit (the base):
|
|
|
|
|
*/
|
|
|
|
|
#ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
|
|
|
|
|
unsigned long
|
2019-03-06 07:46:16 +08:00
|
|
|
arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
|
|
|
|
|
unsigned long len, unsigned long pgoff,
|
|
|
|
|
unsigned long flags)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
struct vm_area_struct *vma, *prev;
|
2005-04-17 06:20:36 +08:00
|
|
|
struct mm_struct *mm = current->mm;
|
2012-12-12 08:01:49 +08:00
|
|
|
struct vm_unmapped_area_info info;
|
2018-12-07 06:50:36 +08:00
|
|
|
const unsigned long mmap_end = arch_get_mmap_end(addr);
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
|
/* requested length too big for entire address space */
|
2018-12-07 06:50:36 +08:00
|
|
|
if (len > mmap_end - mmap_min_addr)
|
2005-04-17 06:20:36 +08:00
|
|
|
return -ENOMEM;
|
|
|
|
|
|
2007-05-07 05:50:13 +08:00
|
|
|
if (flags & MAP_FIXED)
|
|
|
|
|
return addr;
|
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
/* requesting a specific address */
|
|
|
|
|
if (addr) {
|
|
|
|
|
addr = PAGE_ALIGN(addr);
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
vma = find_vma_prev(mm, addr, &prev);
|
2018-12-07 06:50:36 +08:00
|
|
|
if (mmap_end - len >= addr && addr >= mmap_min_addr &&
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
(!vma || addr + len <= vm_start_gap(vma)) &&
|
|
|
|
|
(!prev || addr >= vm_end_gap(prev)))
|
2005-04-17 06:20:36 +08:00
|
|
|
return addr;
|
|
|
|
|
}
|
|
|
|
|
|
2012-12-12 08:01:49 +08:00
|
|
|
info.flags = VM_UNMAPPED_AREA_TOPDOWN;
|
|
|
|
|
info.length = len;
|
mm: ensure get_unmapped_area() returns higher address than mmap_min_addr
This patch fixes the problem that get_unmapped_area() can return illegal
address and result in failing mmap(2) etc.
In case that the address higher than PAGE_SIZE is set to
/proc/sys/vm/mmap_min_addr, the address lower than mmap_min_addr can be
returned by get_unmapped_area(), even if you do not pass any virtual
address hint (i.e. the second argument).
This is because the current get_unmapped_area() code does not take into
account mmap_min_addr.
This leads to two actual problems as follows:
1. mmap(2) can fail with EPERM on the process without CAP_SYS_RAWIO,
although any illegal parameter is not passed.
2. The bottom-up search path after the top-down search might not work in
arch_get_unmapped_area_topdown().
Note: The first and third chunk of my patch, which changes "len" check,
are for more precise check using mmap_min_addr, and not for solving the
above problem.
[How to reproduce]
--- test.c -------------------------------------------------
#include <stdio.h>
#include <unistd.h>
#include <sys/mman.h>
#include <sys/errno.h>
int main(int argc, char *argv[])
{
void *ret = NULL, *last_map;
size_t pagesize = sysconf(_SC_PAGESIZE);
do {
last_map = ret;
ret = mmap(0, pagesize, PROT_NONE,
MAP_PRIVATE|MAP_ANONYMOUS, -1, 0);
// printf("ret=%p\n", ret);
} while (ret != MAP_FAILED);
if (errno != ENOMEM) {
printf("ERR: unexpected errno: %d (last map=%p)\n",
errno, last_map);
}
return 0;
}
---------------------------------------------------------------
$ gcc -m32 -o test test.c
$ sudo sysctl -w vm.mmap_min_addr=65536
vm.mmap_min_addr = 65536
$ ./test (run as non-priviledge user)
ERR: unexpected errno: 1 (last map=0x10000)
Signed-off-by: Akira Takeuchi <takeuchi.akr@jp.panasonic.com>
Signed-off-by: Kiyoshi Owada <owada.kiyoshi@jp.panasonic.com>
Reviewed-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-11-13 07:08:21 +08:00
|
|
|
info.low_limit = max(PAGE_SIZE, mmap_min_addr);
|
2018-12-07 06:50:36 +08:00
|
|
|
info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
|
2012-12-12 08:01:49 +08:00
|
|
|
info.align_mask = 0;
|
2020-04-11 05:32:48 +08:00
|
|
|
info.align_offset = 0;
|
2012-12-12 08:01:49 +08:00
|
|
|
addr = vm_unmapped_area(&info);
|
2012-03-22 07:33:56 +08:00
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
/*
|
|
|
|
|
* A failed mmap() very likely causes application failure,
|
|
|
|
|
* so fall back to the bottom-up function here. This scenario
|
|
|
|
|
* can happen with large stack limits and large mmap()
|
|
|
|
|
* allocations.
|
|
|
|
|
*/
|
2015-11-06 10:46:54 +08:00
|
|
|
if (offset_in_page(addr)) {
|
2012-12-12 08:01:49 +08:00
|
|
|
VM_BUG_ON(addr != -ENOMEM);
|
|
|
|
|
info.flags = 0;
|
|
|
|
|
info.low_limit = TASK_UNMAPPED_BASE;
|
2018-12-07 06:50:36 +08:00
|
|
|
info.high_limit = mmap_end;
|
2012-12-12 08:01:49 +08:00
|
|
|
addr = vm_unmapped_area(&info);
|
|
|
|
|
}
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
|
return addr;
|
|
|
|
|
}
|
|
|
|
|
#endif
|
|
|
|
|
|
|
|
|
|
unsigned long
|
|
|
|
|
get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
|
|
|
|
|
unsigned long pgoff, unsigned long flags)
|
|
|
|
|
{
|
2007-05-07 05:50:13 +08:00
|
|
|
unsigned long (*get_area)(struct file *, unsigned long,
|
|
|
|
|
unsigned long, unsigned long, unsigned long);
|
|
|
|
|
|
2009-12-04 04:23:11 +08:00
|
|
|
unsigned long error = arch_mmap_check(addr, len, flags);
|
|
|
|
|
if (error)
|
|
|
|
|
return error;
|
|
|
|
|
|
|
|
|
|
/* Careful about overflows.. */
|
|
|
|
|
if (len > TASK_SIZE)
|
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
2007-05-07 05:50:13 +08:00
|
|
|
get_area = current->mm->get_unmapped_area;
|
2016-07-27 06:26:15 +08:00
|
|
|
if (file) {
|
|
|
|
|
if (file->f_op->get_unmapped_area)
|
|
|
|
|
get_area = file->f_op->get_unmapped_area;
|
|
|
|
|
} else if (flags & MAP_SHARED) {
|
|
|
|
|
/*
|
|
|
|
|
* mmap_region() will call shmem_zero_setup() to create a file,
|
|
|
|
|
* so use shmem's get_unmapped_area in case it can be huge.
|
2020-08-07 14:23:37 +08:00
|
|
|
* do_mmap() will clear pgoff, so match alignment.
|
2016-07-27 06:26:15 +08:00
|
|
|
*/
|
|
|
|
|
pgoff = 0;
|
|
|
|
|
get_area = shmem_get_unmapped_area;
|
|
|
|
|
}
|
|
|
|
|
|
2007-05-07 05:50:13 +08:00
|
|
|
addr = get_area(file, addr, len, pgoff, flags);
|
|
|
|
|
if (IS_ERR_VALUE(addr))
|
|
|
|
|
return addr;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2005-05-20 13:43:37 +08:00
|
|
|
if (addr > TASK_SIZE - len)
|
|
|
|
|
return -ENOMEM;
|
2015-11-06 10:46:54 +08:00
|
|
|
if (offset_in_page(addr))
|
2005-05-20 13:43:37 +08:00
|
|
|
return -EINVAL;
|
2007-05-07 05:50:13 +08:00
|
|
|
|
2012-05-31 05:13:15 +08:00
|
|
|
error = security_mmap_addr(addr);
|
|
|
|
|
return error ? error : addr;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
EXPORT_SYMBOL(get_unmapped_area);
|
|
|
|
|
|
|
|
|
|
/* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
|
2009-01-07 06:40:21 +08:00
|
|
|
struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
mm: per-thread vma caching
This patch is a continuation of efforts trying to optimize find_vma(),
avoiding potentially expensive rbtree walks to locate a vma upon faults.
The original approach (https://lkml.org/lkml/2013/11/1/410), where the
largest vma was also cached, ended up being too specific and random,
thus further comparison with other approaches were needed. There are
two things to consider when dealing with this, the cache hit rate and
the latency of find_vma(). Improving the hit-rate does not necessarily
translate in finding the vma any faster, as the overhead of any fancy
caching schemes can be too high to consider.
We currently cache the last used vma for the whole address space, which
provides a nice optimization, reducing the total cycles in find_vma() by
up to 250%, for workloads with good locality. On the other hand, this
simple scheme is pretty much useless for workloads with poor locality.
Analyzing ebizzy runs shows that, no matter how many threads are
running, the mmap_cache hit rate is less than 2%, and in many situations
below 1%.
The proposed approach is to replace this scheme with a small per-thread
cache, maximizing hit rates at a very low maintenance cost.
Invalidations are performed by simply bumping up a 32-bit sequence
number. The only expensive operation is in the rare case of a seq
number overflow, where all caches that share the same address space are
flushed. Upon a miss, the proposed replacement policy is based on the
page number that contains the virtual address in question. Concretely,
the following results are seen on an 80 core, 8 socket x86-64 box:
1) System bootup: Most programs are single threaded, so the per-thread
scheme does improve ~50% hit rate by just adding a few more slots to
the cache.
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 50.61% | 19.90 |
| patched | 73.45% | 13.58 |
+----------------+----------+------------------+
2) Kernel build: This one is already pretty good with the current
approach as we're dealing with good locality.
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 75.28% | 11.03 |
| patched | 88.09% | 9.31 |
+----------------+----------+------------------+
3) Oracle 11g Data Mining (4k pages): Similar to the kernel build workload.
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 70.66% | 17.14 |
| patched | 91.15% | 12.57 |
+----------------+----------+------------------+
4) Ebizzy: There's a fair amount of variation from run to run, but this
approach always shows nearly perfect hit rates, while baseline is just
about non-existent. The amounts of cycles can fluctuate between
anywhere from ~60 to ~116 for the baseline scheme, but this approach
reduces it considerably. For instance, with 80 threads:
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 1.06% | 91.54 |
| patched | 99.97% | 14.18 |
+----------------+----------+------------------+
[akpm@linux-foundation.org: fix nommu build, per Davidlohr]
[akpm@linux-foundation.org: document vmacache_valid() logic]
[akpm@linux-foundation.org: attempt to untangle header files]
[akpm@linux-foundation.org: add vmacache_find() BUG_ON]
[hughd@google.com: add vmacache_valid_mm() (from Oleg)]
[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: adjust and enhance comments]
Signed-off-by: Davidlohr Bueso <davidlohr@hp.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Reviewed-by: Michel Lespinasse <walken@google.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Tested-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-04-08 06:37:25 +08:00
|
|
|
struct rb_node *rb_node;
|
|
|
|
|
struct vm_area_struct *vma;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2021-09-03 05:56:46 +08:00
|
|
|
mmap_assert_locked(mm);
|
2012-05-30 06:06:21 +08:00
|
|
|
/* Check the cache first. */
|
mm: per-thread vma caching
This patch is a continuation of efforts trying to optimize find_vma(),
avoiding potentially expensive rbtree walks to locate a vma upon faults.
The original approach (https://lkml.org/lkml/2013/11/1/410), where the
largest vma was also cached, ended up being too specific and random,
thus further comparison with other approaches were needed. There are
two things to consider when dealing with this, the cache hit rate and
the latency of find_vma(). Improving the hit-rate does not necessarily
translate in finding the vma any faster, as the overhead of any fancy
caching schemes can be too high to consider.
We currently cache the last used vma for the whole address space, which
provides a nice optimization, reducing the total cycles in find_vma() by
up to 250%, for workloads with good locality. On the other hand, this
simple scheme is pretty much useless for workloads with poor locality.
Analyzing ebizzy runs shows that, no matter how many threads are
running, the mmap_cache hit rate is less than 2%, and in many situations
below 1%.
The proposed approach is to replace this scheme with a small per-thread
cache, maximizing hit rates at a very low maintenance cost.
Invalidations are performed by simply bumping up a 32-bit sequence
number. The only expensive operation is in the rare case of a seq
number overflow, where all caches that share the same address space are
flushed. Upon a miss, the proposed replacement policy is based on the
page number that contains the virtual address in question. Concretely,
the following results are seen on an 80 core, 8 socket x86-64 box:
1) System bootup: Most programs are single threaded, so the per-thread
scheme does improve ~50% hit rate by just adding a few more slots to
the cache.
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 50.61% | 19.90 |
| patched | 73.45% | 13.58 |
+----------------+----------+------------------+
2) Kernel build: This one is already pretty good with the current
approach as we're dealing with good locality.
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 75.28% | 11.03 |
| patched | 88.09% | 9.31 |
+----------------+----------+------------------+
3) Oracle 11g Data Mining (4k pages): Similar to the kernel build workload.
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 70.66% | 17.14 |
| patched | 91.15% | 12.57 |
+----------------+----------+------------------+
4) Ebizzy: There's a fair amount of variation from run to run, but this
approach always shows nearly perfect hit rates, while baseline is just
about non-existent. The amounts of cycles can fluctuate between
anywhere from ~60 to ~116 for the baseline scheme, but this approach
reduces it considerably. For instance, with 80 threads:
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 1.06% | 91.54 |
| patched | 99.97% | 14.18 |
+----------------+----------+------------------+
[akpm@linux-foundation.org: fix nommu build, per Davidlohr]
[akpm@linux-foundation.org: document vmacache_valid() logic]
[akpm@linux-foundation.org: attempt to untangle header files]
[akpm@linux-foundation.org: add vmacache_find() BUG_ON]
[hughd@google.com: add vmacache_valid_mm() (from Oleg)]
[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: adjust and enhance comments]
Signed-off-by: Davidlohr Bueso <davidlohr@hp.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Reviewed-by: Michel Lespinasse <walken@google.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Tested-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-04-08 06:37:25 +08:00
|
|
|
vma = vmacache_find(mm, addr);
|
|
|
|
|
if (likely(vma))
|
|
|
|
|
return vma;
|
2012-05-30 06:06:21 +08:00
|
|
|
|
mm: per-thread vma caching
This patch is a continuation of efforts trying to optimize find_vma(),
avoiding potentially expensive rbtree walks to locate a vma upon faults.
The original approach (https://lkml.org/lkml/2013/11/1/410), where the
largest vma was also cached, ended up being too specific and random,
thus further comparison with other approaches were needed. There are
two things to consider when dealing with this, the cache hit rate and
the latency of find_vma(). Improving the hit-rate does not necessarily
translate in finding the vma any faster, as the overhead of any fancy
caching schemes can be too high to consider.
We currently cache the last used vma for the whole address space, which
provides a nice optimization, reducing the total cycles in find_vma() by
up to 250%, for workloads with good locality. On the other hand, this
simple scheme is pretty much useless for workloads with poor locality.
Analyzing ebizzy runs shows that, no matter how many threads are
running, the mmap_cache hit rate is less than 2%, and in many situations
below 1%.
The proposed approach is to replace this scheme with a small per-thread
cache, maximizing hit rates at a very low maintenance cost.
Invalidations are performed by simply bumping up a 32-bit sequence
number. The only expensive operation is in the rare case of a seq
number overflow, where all caches that share the same address space are
flushed. Upon a miss, the proposed replacement policy is based on the
page number that contains the virtual address in question. Concretely,
the following results are seen on an 80 core, 8 socket x86-64 box:
1) System bootup: Most programs are single threaded, so the per-thread
scheme does improve ~50% hit rate by just adding a few more slots to
the cache.
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 50.61% | 19.90 |
| patched | 73.45% | 13.58 |
+----------------+----------+------------------+
2) Kernel build: This one is already pretty good with the current
approach as we're dealing with good locality.
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 75.28% | 11.03 |
| patched | 88.09% | 9.31 |
+----------------+----------+------------------+
3) Oracle 11g Data Mining (4k pages): Similar to the kernel build workload.
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 70.66% | 17.14 |
| patched | 91.15% | 12.57 |
+----------------+----------+------------------+
4) Ebizzy: There's a fair amount of variation from run to run, but this
approach always shows nearly perfect hit rates, while baseline is just
about non-existent. The amounts of cycles can fluctuate between
anywhere from ~60 to ~116 for the baseline scheme, but this approach
reduces it considerably. For instance, with 80 threads:
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 1.06% | 91.54 |
| patched | 99.97% | 14.18 |
+----------------+----------+------------------+
[akpm@linux-foundation.org: fix nommu build, per Davidlohr]
[akpm@linux-foundation.org: document vmacache_valid() logic]
[akpm@linux-foundation.org: attempt to untangle header files]
[akpm@linux-foundation.org: add vmacache_find() BUG_ON]
[hughd@google.com: add vmacache_valid_mm() (from Oleg)]
[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: adjust and enhance comments]
Signed-off-by: Davidlohr Bueso <davidlohr@hp.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Reviewed-by: Michel Lespinasse <walken@google.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Tested-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-04-08 06:37:25 +08:00
|
|
|
rb_node = mm->mm_rb.rb_node;
|
2012-05-30 06:06:21 +08:00
|
|
|
|
mm: per-thread vma caching
This patch is a continuation of efforts trying to optimize find_vma(),
avoiding potentially expensive rbtree walks to locate a vma upon faults.
The original approach (https://lkml.org/lkml/2013/11/1/410), where the
largest vma was also cached, ended up being too specific and random,
thus further comparison with other approaches were needed. There are
two things to consider when dealing with this, the cache hit rate and
the latency of find_vma(). Improving the hit-rate does not necessarily
translate in finding the vma any faster, as the overhead of any fancy
caching schemes can be too high to consider.
We currently cache the last used vma for the whole address space, which
provides a nice optimization, reducing the total cycles in find_vma() by
up to 250%, for workloads with good locality. On the other hand, this
simple scheme is pretty much useless for workloads with poor locality.
Analyzing ebizzy runs shows that, no matter how many threads are
running, the mmap_cache hit rate is less than 2%, and in many situations
below 1%.
The proposed approach is to replace this scheme with a small per-thread
cache, maximizing hit rates at a very low maintenance cost.
Invalidations are performed by simply bumping up a 32-bit sequence
number. The only expensive operation is in the rare case of a seq
number overflow, where all caches that share the same address space are
flushed. Upon a miss, the proposed replacement policy is based on the
page number that contains the virtual address in question. Concretely,
the following results are seen on an 80 core, 8 socket x86-64 box:
1) System bootup: Most programs are single threaded, so the per-thread
scheme does improve ~50% hit rate by just adding a few more slots to
the cache.
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 50.61% | 19.90 |
| patched | 73.45% | 13.58 |
+----------------+----------+------------------+
2) Kernel build: This one is already pretty good with the current
approach as we're dealing with good locality.
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 75.28% | 11.03 |
| patched | 88.09% | 9.31 |
+----------------+----------+------------------+
3) Oracle 11g Data Mining (4k pages): Similar to the kernel build workload.
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 70.66% | 17.14 |
| patched | 91.15% | 12.57 |
+----------------+----------+------------------+
4) Ebizzy: There's a fair amount of variation from run to run, but this
approach always shows nearly perfect hit rates, while baseline is just
about non-existent. The amounts of cycles can fluctuate between
anywhere from ~60 to ~116 for the baseline scheme, but this approach
reduces it considerably. For instance, with 80 threads:
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 1.06% | 91.54 |
| patched | 99.97% | 14.18 |
+----------------+----------+------------------+
[akpm@linux-foundation.org: fix nommu build, per Davidlohr]
[akpm@linux-foundation.org: document vmacache_valid() logic]
[akpm@linux-foundation.org: attempt to untangle header files]
[akpm@linux-foundation.org: add vmacache_find() BUG_ON]
[hughd@google.com: add vmacache_valid_mm() (from Oleg)]
[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: adjust and enhance comments]
Signed-off-by: Davidlohr Bueso <davidlohr@hp.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Reviewed-by: Michel Lespinasse <walken@google.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Tested-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-04-08 06:37:25 +08:00
|
|
|
while (rb_node) {
|
|
|
|
|
struct vm_area_struct *tmp;
|
|
|
|
|
|
|
|
|
|
tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
|
|
|
|
|
|
|
|
|
|
if (tmp->vm_end > addr) {
|
|
|
|
|
vma = tmp;
|
|
|
|
|
if (tmp->vm_start <= addr)
|
|
|
|
|
break;
|
|
|
|
|
rb_node = rb_node->rb_left;
|
|
|
|
|
} else
|
|
|
|
|
rb_node = rb_node->rb_right;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
mm: per-thread vma caching
This patch is a continuation of efforts trying to optimize find_vma(),
avoiding potentially expensive rbtree walks to locate a vma upon faults.
The original approach (https://lkml.org/lkml/2013/11/1/410), where the
largest vma was also cached, ended up being too specific and random,
thus further comparison with other approaches were needed. There are
two things to consider when dealing with this, the cache hit rate and
the latency of find_vma(). Improving the hit-rate does not necessarily
translate in finding the vma any faster, as the overhead of any fancy
caching schemes can be too high to consider.
We currently cache the last used vma for the whole address space, which
provides a nice optimization, reducing the total cycles in find_vma() by
up to 250%, for workloads with good locality. On the other hand, this
simple scheme is pretty much useless for workloads with poor locality.
Analyzing ebizzy runs shows that, no matter how many threads are
running, the mmap_cache hit rate is less than 2%, and in many situations
below 1%.
The proposed approach is to replace this scheme with a small per-thread
cache, maximizing hit rates at a very low maintenance cost.
Invalidations are performed by simply bumping up a 32-bit sequence
number. The only expensive operation is in the rare case of a seq
number overflow, where all caches that share the same address space are
flushed. Upon a miss, the proposed replacement policy is based on the
page number that contains the virtual address in question. Concretely,
the following results are seen on an 80 core, 8 socket x86-64 box:
1) System bootup: Most programs are single threaded, so the per-thread
scheme does improve ~50% hit rate by just adding a few more slots to
the cache.
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 50.61% | 19.90 |
| patched | 73.45% | 13.58 |
+----------------+----------+------------------+
2) Kernel build: This one is already pretty good with the current
approach as we're dealing with good locality.
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 75.28% | 11.03 |
| patched | 88.09% | 9.31 |
+----------------+----------+------------------+
3) Oracle 11g Data Mining (4k pages): Similar to the kernel build workload.
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 70.66% | 17.14 |
| patched | 91.15% | 12.57 |
+----------------+----------+------------------+
4) Ebizzy: There's a fair amount of variation from run to run, but this
approach always shows nearly perfect hit rates, while baseline is just
about non-existent. The amounts of cycles can fluctuate between
anywhere from ~60 to ~116 for the baseline scheme, but this approach
reduces it considerably. For instance, with 80 threads:
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 1.06% | 91.54 |
| patched | 99.97% | 14.18 |
+----------------+----------+------------------+
[akpm@linux-foundation.org: fix nommu build, per Davidlohr]
[akpm@linux-foundation.org: document vmacache_valid() logic]
[akpm@linux-foundation.org: attempt to untangle header files]
[akpm@linux-foundation.org: add vmacache_find() BUG_ON]
[hughd@google.com: add vmacache_valid_mm() (from Oleg)]
[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: adjust and enhance comments]
Signed-off-by: Davidlohr Bueso <davidlohr@hp.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Reviewed-by: Michel Lespinasse <walken@google.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Tested-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-04-08 06:37:25 +08:00
|
|
|
|
|
|
|
|
if (vma)
|
|
|
|
|
vmacache_update(addr, vma);
|
2005-04-17 06:20:36 +08:00
|
|
|
return vma;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
EXPORT_SYMBOL(find_vma);
|
|
|
|
|
|
2012-01-11 07:08:07 +08:00
|
|
|
/*
|
|
|
|
|
* Same as find_vma, but also return a pointer to the previous VMA in *pprev.
|
|
|
|
|
*/
|
2005-04-17 06:20:36 +08:00
|
|
|
struct vm_area_struct *
|
|
|
|
|
find_vma_prev(struct mm_struct *mm, unsigned long addr,
|
|
|
|
|
struct vm_area_struct **pprev)
|
|
|
|
|
{
|
2012-01-11 07:08:07 +08:00
|
|
|
struct vm_area_struct *vma;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2012-01-11 07:08:07 +08:00
|
|
|
vma = find_vma(mm, addr);
|
2012-03-05 08:52:03 +08:00
|
|
|
if (vma) {
|
|
|
|
|
*pprev = vma->vm_prev;
|
|
|
|
|
} else {
|
2019-09-24 06:39:25 +08:00
|
|
|
struct rb_node *rb_node = rb_last(&mm->mm_rb);
|
|
|
|
|
|
|
|
|
|
*pprev = rb_node ? rb_entry(rb_node, struct vm_area_struct, vm_rb) : NULL;
|
2012-03-05 08:52:03 +08:00
|
|
|
}
|
2012-01-11 07:08:07 +08:00
|
|
|
return vma;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Verify that the stack growth is acceptable and
|
|
|
|
|
* update accounting. This is shared with both the
|
|
|
|
|
* grow-up and grow-down cases.
|
|
|
|
|
*/
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
static int acct_stack_growth(struct vm_area_struct *vma,
|
|
|
|
|
unsigned long size, unsigned long grow)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
|
|
|
|
struct mm_struct *mm = vma->vm_mm;
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
unsigned long new_start;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
|
/* address space limit tests */
|
2016-01-15 07:22:07 +08:00
|
|
|
if (!may_expand_vm(mm, vma->vm_flags, grow))
|
2005-04-17 06:20:36 +08:00
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
|
|
/* Stack limit test */
|
2017-07-11 06:50:03 +08:00
|
|
|
if (size > rlimit(RLIMIT_STACK))
|
2005-04-17 06:20:36 +08:00
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
|
|
/* mlock limit tests */
|
|
|
|
|
if (vma->vm_flags & VM_LOCKED) {
|
|
|
|
|
unsigned long locked;
|
|
|
|
|
unsigned long limit;
|
|
|
|
|
locked = mm->locked_vm + grow;
|
2017-07-11 06:50:03 +08:00
|
|
|
limit = rlimit(RLIMIT_MEMLOCK);
|
2010-03-06 05:41:44 +08:00
|
|
|
limit >>= PAGE_SHIFT;
|
2005-04-17 06:20:36 +08:00
|
|
|
if (locked > limit && !capable(CAP_IPC_LOCK))
|
|
|
|
|
return -ENOMEM;
|
|
|
|
|
}
|
|
|
|
|
|
2007-01-31 06:35:39 +08:00
|
|
|
/* Check to ensure the stack will not grow into a hugetlb-only region */
|
|
|
|
|
new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
|
|
|
|
|
vma->vm_end - size;
|
|
|
|
|
if (is_hugepage_only_range(vma->vm_mm, new_start, size))
|
|
|
|
|
return -EFAULT;
|
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
/*
|
|
|
|
|
* Overcommit.. This must be the final test, as it will
|
|
|
|
|
* update security statistics.
|
|
|
|
|
*/
|
2009-04-17 04:58:12 +08:00
|
|
|
if (security_vm_enough_memory_mm(mm, grow))
|
2005-04-17 06:20:36 +08:00
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
2005-10-30 09:16:20 +08:00
|
|
|
#if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
|
2005-04-17 06:20:36 +08:00
|
|
|
/*
|
2005-10-30 09:16:20 +08:00
|
|
|
* PA-RISC uses this for its stack; IA64 for its Register Backing Store.
|
|
|
|
|
* vma is the last one with address > vma->vm_end. Have to extend vma.
|
2005-04-17 06:20:36 +08:00
|
|
|
*/
|
2005-10-30 09:16:20 +08:00
|
|
|
int expand_upwards(struct vm_area_struct *vma, unsigned long address)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
2015-11-06 10:48:17 +08:00
|
|
|
struct mm_struct *mm = vma->vm_mm;
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
struct vm_area_struct *next;
|
|
|
|
|
unsigned long gap_addr;
|
2016-02-06 07:36:50 +08:00
|
|
|
int error = 0;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
|
if (!(vma->vm_flags & VM_GROWSUP))
|
|
|
|
|
return -EFAULT;
|
|
|
|
|
|
2017-06-19 23:34:05 +08:00
|
|
|
/* Guard against exceeding limits of the address space. */
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
address &= PAGE_MASK;
|
2017-07-15 05:49:38 +08:00
|
|
|
if (address >= (TASK_SIZE & PAGE_MASK))
|
2016-02-06 07:36:50 +08:00
|
|
|
return -ENOMEM;
|
2017-06-19 23:34:05 +08:00
|
|
|
address += PAGE_SIZE;
|
2016-02-06 07:36:50 +08:00
|
|
|
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
/* Enforce stack_guard_gap */
|
|
|
|
|
gap_addr = address + stack_guard_gap;
|
2017-06-19 23:34:05 +08:00
|
|
|
|
|
|
|
|
/* Guard against overflow */
|
|
|
|
|
if (gap_addr < address || gap_addr > TASK_SIZE)
|
|
|
|
|
gap_addr = TASK_SIZE;
|
|
|
|
|
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
next = vma->vm_next;
|
2020-04-07 11:03:47 +08:00
|
|
|
if (next && next->vm_start < gap_addr && vma_is_accessible(next)) {
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
if (!(next->vm_flags & VM_GROWSUP))
|
|
|
|
|
return -ENOMEM;
|
|
|
|
|
/* Check that both stack segments have the same anon_vma? */
|
|
|
|
|
}
|
|
|
|
|
|
2016-02-06 07:36:50 +08:00
|
|
|
/* We must make sure the anon_vma is allocated. */
|
2005-04-17 06:20:36 +08:00
|
|
|
if (unlikely(anon_vma_prepare(vma)))
|
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* vma->vm_start/vm_end cannot change under us because the caller
|
2020-06-09 12:33:54 +08:00
|
|
|
* is required to hold the mmap_lock in read mode. We need the
|
2005-04-17 06:20:36 +08:00
|
|
|
* anon_vma lock to serialize against concurrent expand_stacks.
|
|
|
|
|
*/
|
2016-02-06 07:36:50 +08:00
|
|
|
anon_vma_lock_write(vma->anon_vma);
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
|
/* Somebody else might have raced and expanded it already */
|
|
|
|
|
if (address > vma->vm_end) {
|
|
|
|
|
unsigned long size, grow;
|
|
|
|
|
|
|
|
|
|
size = address - vma->vm_start;
|
|
|
|
|
grow = (address - vma->vm_end) >> PAGE_SHIFT;
|
|
|
|
|
|
2011-05-10 08:44:42 +08:00
|
|
|
error = -ENOMEM;
|
|
|
|
|
if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
|
|
|
|
|
error = acct_stack_growth(vma, size, grow);
|
|
|
|
|
if (!error) {
|
2012-12-13 05:52:25 +08:00
|
|
|
/*
|
|
|
|
|
* vma_gap_update() doesn't support concurrent
|
2020-06-09 12:33:54 +08:00
|
|
|
* updates, but we only hold a shared mmap_lock
|
2012-12-13 05:52:25 +08:00
|
|
|
* lock here, so we need to protect against
|
|
|
|
|
* concurrent vma expansions.
|
2016-02-06 07:36:50 +08:00
|
|
|
* anon_vma_lock_write() doesn't help here, as
|
2012-12-13 05:52:25 +08:00
|
|
|
* we don't guarantee that all growable vmas
|
|
|
|
|
* in a mm share the same root anon vma.
|
|
|
|
|
* So, we reuse mm->page_table_lock to guard
|
|
|
|
|
* against concurrent vma expansions.
|
|
|
|
|
*/
|
2015-11-06 10:48:17 +08:00
|
|
|
spin_lock(&mm->page_table_lock);
|
2015-11-06 10:48:14 +08:00
|
|
|
if (vma->vm_flags & VM_LOCKED)
|
2015-11-06 10:48:17 +08:00
|
|
|
mm->locked_vm += grow;
|
2016-01-15 07:22:07 +08:00
|
|
|
vm_stat_account(mm, vma->vm_flags, grow);
|
mm anon rmap: replace same_anon_vma linked list with an interval tree.
When a large VMA (anon or private file mapping) is first touched, which
will populate its anon_vma field, and then split into many regions through
the use of mprotect(), the original anon_vma ends up linking all of the
vmas on a linked list. This can cause rmap to become inefficient, as we
have to walk potentially thousands of irrelevent vmas before finding the
one a given anon page might fall into.
By replacing the same_anon_vma linked list with an interval tree (where
each avc's interval is determined by its vma's start and last pgoffs), we
can make rmap efficient for this use case again.
While the change is large, all of its pieces are fairly simple.
Most places that were walking the same_anon_vma list were looking for a
known pgoff, so they can just use the anon_vma_interval_tree_foreach()
interval tree iterator instead. The exception here is ksm, where the
page's index is not known. It would probably be possible to rework ksm so
that the index would be known, but for now I have decided to keep things
simple and just walk the entirety of the interval tree there.
When updating vma's that already have an anon_vma assigned, we must take
care to re-index the corresponding avc's on their interval tree. This is
done through the use of anon_vma_interval_tree_pre_update_vma() and
anon_vma_interval_tree_post_update_vma(), which remove the avc's from
their interval tree before the update and re-insert them after the update.
The anon_vma stays locked during the update, so there is no chance that
rmap would miss the vmas that are being updated.
Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-10-09 07:31:39 +08:00
|
|
|
anon_vma_interval_tree_pre_update_vma(vma);
|
2011-05-10 08:44:42 +08:00
|
|
|
vma->vm_end = address;
|
mm anon rmap: replace same_anon_vma linked list with an interval tree.
When a large VMA (anon or private file mapping) is first touched, which
will populate its anon_vma field, and then split into many regions through
the use of mprotect(), the original anon_vma ends up linking all of the
vmas on a linked list. This can cause rmap to become inefficient, as we
have to walk potentially thousands of irrelevent vmas before finding the
one a given anon page might fall into.
By replacing the same_anon_vma linked list with an interval tree (where
each avc's interval is determined by its vma's start and last pgoffs), we
can make rmap efficient for this use case again.
While the change is large, all of its pieces are fairly simple.
Most places that were walking the same_anon_vma list were looking for a
known pgoff, so they can just use the anon_vma_interval_tree_foreach()
interval tree iterator instead. The exception here is ksm, where the
page's index is not known. It would probably be possible to rework ksm so
that the index would be known, but for now I have decided to keep things
simple and just walk the entirety of the interval tree there.
When updating vma's that already have an anon_vma assigned, we must take
care to re-index the corresponding avc's on their interval tree. This is
done through the use of anon_vma_interval_tree_pre_update_vma() and
anon_vma_interval_tree_post_update_vma(), which remove the avc's from
their interval tree before the update and re-insert them after the update.
The anon_vma stays locked during the update, so there is no chance that
rmap would miss the vmas that are being updated.
Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-10-09 07:31:39 +08:00
|
|
|
anon_vma_interval_tree_post_update_vma(vma);
|
2012-12-12 08:01:38 +08:00
|
|
|
if (vma->vm_next)
|
|
|
|
|
vma_gap_update(vma->vm_next);
|
|
|
|
|
else
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
mm->highest_vm_end = vm_end_gap(vma);
|
2015-11-06 10:48:17 +08:00
|
|
|
spin_unlock(&mm->page_table_lock);
|
2012-12-13 05:52:25 +08:00
|
|
|
|
2011-05-10 08:44:42 +08:00
|
|
|
perf_event_mmap(vma);
|
|
|
|
|
}
|
2010-05-18 22:30:49 +08:00
|
|
|
}
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
2016-02-06 07:36:50 +08:00
|
|
|
anon_vma_unlock_write(vma->anon_vma);
|
2014-10-30 05:50:31 +08:00
|
|
|
khugepaged_enter_vma_merge(vma, vma->vm_flags);
|
2015-11-06 10:48:17 +08:00
|
|
|
validate_mm(mm);
|
2005-04-17 06:20:36 +08:00
|
|
|
return error;
|
|
|
|
|
}
|
2005-10-30 09:16:20 +08:00
|
|
|
#endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
|
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
/*
|
|
|
|
|
* vma is the first one with address < vma->vm_start. Have to extend vma.
|
|
|
|
|
*/
|
2011-05-25 08:11:44 +08:00
|
|
|
int expand_downwards(struct vm_area_struct *vma,
|
2007-07-19 16:48:16 +08:00
|
|
|
unsigned long address)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
2015-11-06 10:48:17 +08:00
|
|
|
struct mm_struct *mm = vma->vm_mm;
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
struct vm_area_struct *prev;
|
2019-02-28 04:29:52 +08:00
|
|
|
int error = 0;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2007-11-27 07:47:26 +08:00
|
|
|
address &= PAGE_MASK;
|
2019-02-28 04:29:52 +08:00
|
|
|
if (address < mmap_min_addr)
|
|
|
|
|
return -EPERM;
|
2007-11-27 07:47:26 +08:00
|
|
|
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
/* Enforce stack_guard_gap */
|
|
|
|
|
prev = vma->vm_prev;
|
2017-07-11 06:49:54 +08:00
|
|
|
/* Check that both stack segments have the same anon_vma? */
|
|
|
|
|
if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
|
2020-04-07 11:03:47 +08:00
|
|
|
vma_is_accessible(prev)) {
|
2017-07-11 06:49:54 +08:00
|
|
|
if (address - prev->vm_end < stack_guard_gap)
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
return -ENOMEM;
|
|
|
|
|
}
|
|
|
|
|
|
2016-02-06 07:36:50 +08:00
|
|
|
/* We must make sure the anon_vma is allocated. */
|
|
|
|
|
if (unlikely(anon_vma_prepare(vma)))
|
|
|
|
|
return -ENOMEM;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* vma->vm_start/vm_end cannot change under us because the caller
|
2020-06-09 12:33:54 +08:00
|
|
|
* is required to hold the mmap_lock in read mode. We need the
|
2005-04-17 06:20:36 +08:00
|
|
|
* anon_vma lock to serialize against concurrent expand_stacks.
|
|
|
|
|
*/
|
2016-02-06 07:36:50 +08:00
|
|
|
anon_vma_lock_write(vma->anon_vma);
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
|
/* Somebody else might have raced and expanded it already */
|
|
|
|
|
if (address < vma->vm_start) {
|
|
|
|
|
unsigned long size, grow;
|
|
|
|
|
|
|
|
|
|
size = vma->vm_end - address;
|
|
|
|
|
grow = (vma->vm_start - address) >> PAGE_SHIFT;
|
|
|
|
|
|
2011-04-13 23:07:28 +08:00
|
|
|
error = -ENOMEM;
|
|
|
|
|
if (grow <= vma->vm_pgoff) {
|
|
|
|
|
error = acct_stack_growth(vma, size, grow);
|
|
|
|
|
if (!error) {
|
2012-12-13 05:52:25 +08:00
|
|
|
/*
|
|
|
|
|
* vma_gap_update() doesn't support concurrent
|
2020-06-09 12:33:54 +08:00
|
|
|
* updates, but we only hold a shared mmap_lock
|
2012-12-13 05:52:25 +08:00
|
|
|
* lock here, so we need to protect against
|
|
|
|
|
* concurrent vma expansions.
|
2016-02-06 07:36:50 +08:00
|
|
|
* anon_vma_lock_write() doesn't help here, as
|
2012-12-13 05:52:25 +08:00
|
|
|
* we don't guarantee that all growable vmas
|
|
|
|
|
* in a mm share the same root anon vma.
|
|
|
|
|
* So, we reuse mm->page_table_lock to guard
|
|
|
|
|
* against concurrent vma expansions.
|
|
|
|
|
*/
|
2015-11-06 10:48:17 +08:00
|
|
|
spin_lock(&mm->page_table_lock);
|
2015-11-06 10:48:14 +08:00
|
|
|
if (vma->vm_flags & VM_LOCKED)
|
2015-11-06 10:48:17 +08:00
|
|
|
mm->locked_vm += grow;
|
2016-01-15 07:22:07 +08:00
|
|
|
vm_stat_account(mm, vma->vm_flags, grow);
|
mm anon rmap: replace same_anon_vma linked list with an interval tree.
When a large VMA (anon or private file mapping) is first touched, which
will populate its anon_vma field, and then split into many regions through
the use of mprotect(), the original anon_vma ends up linking all of the
vmas on a linked list. This can cause rmap to become inefficient, as we
have to walk potentially thousands of irrelevent vmas before finding the
one a given anon page might fall into.
By replacing the same_anon_vma linked list with an interval tree (where
each avc's interval is determined by its vma's start and last pgoffs), we
can make rmap efficient for this use case again.
While the change is large, all of its pieces are fairly simple.
Most places that were walking the same_anon_vma list were looking for a
known pgoff, so they can just use the anon_vma_interval_tree_foreach()
interval tree iterator instead. The exception here is ksm, where the
page's index is not known. It would probably be possible to rework ksm so
that the index would be known, but for now I have decided to keep things
simple and just walk the entirety of the interval tree there.
When updating vma's that already have an anon_vma assigned, we must take
care to re-index the corresponding avc's on their interval tree. This is
done through the use of anon_vma_interval_tree_pre_update_vma() and
anon_vma_interval_tree_post_update_vma(), which remove the avc's from
their interval tree before the update and re-insert them after the update.
The anon_vma stays locked during the update, so there is no chance that
rmap would miss the vmas that are being updated.
Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-10-09 07:31:39 +08:00
|
|
|
anon_vma_interval_tree_pre_update_vma(vma);
|
2011-04-13 23:07:28 +08:00
|
|
|
vma->vm_start = address;
|
|
|
|
|
vma->vm_pgoff -= grow;
|
mm anon rmap: replace same_anon_vma linked list with an interval tree.
When a large VMA (anon or private file mapping) is first touched, which
will populate its anon_vma field, and then split into many regions through
the use of mprotect(), the original anon_vma ends up linking all of the
vmas on a linked list. This can cause rmap to become inefficient, as we
have to walk potentially thousands of irrelevent vmas before finding the
one a given anon page might fall into.
By replacing the same_anon_vma linked list with an interval tree (where
each avc's interval is determined by its vma's start and last pgoffs), we
can make rmap efficient for this use case again.
While the change is large, all of its pieces are fairly simple.
Most places that were walking the same_anon_vma list were looking for a
known pgoff, so they can just use the anon_vma_interval_tree_foreach()
interval tree iterator instead. The exception here is ksm, where the
page's index is not known. It would probably be possible to rework ksm so
that the index would be known, but for now I have decided to keep things
simple and just walk the entirety of the interval tree there.
When updating vma's that already have an anon_vma assigned, we must take
care to re-index the corresponding avc's on their interval tree. This is
done through the use of anon_vma_interval_tree_pre_update_vma() and
anon_vma_interval_tree_post_update_vma(), which remove the avc's from
their interval tree before the update and re-insert them after the update.
The anon_vma stays locked during the update, so there is no chance that
rmap would miss the vmas that are being updated.
Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-10-09 07:31:39 +08:00
|
|
|
anon_vma_interval_tree_post_update_vma(vma);
|
2012-12-12 08:01:38 +08:00
|
|
|
vma_gap_update(vma);
|
2015-11-06 10:48:17 +08:00
|
|
|
spin_unlock(&mm->page_table_lock);
|
2012-12-13 05:52:25 +08:00
|
|
|
|
2011-04-13 23:07:28 +08:00
|
|
|
perf_event_mmap(vma);
|
|
|
|
|
}
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
}
|
2016-02-06 07:36:50 +08:00
|
|
|
anon_vma_unlock_write(vma->anon_vma);
|
2014-10-30 05:50:31 +08:00
|
|
|
khugepaged_enter_vma_merge(vma, vma->vm_flags);
|
2015-11-06 10:48:17 +08:00
|
|
|
validate_mm(mm);
|
2005-04-17 06:20:36 +08:00
|
|
|
return error;
|
|
|
|
|
}
|
|
|
|
|
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
/* enforced gap between the expanding stack and other mappings. */
|
|
|
|
|
unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
|
|
|
|
|
|
|
|
|
|
static int __init cmdline_parse_stack_guard_gap(char *p)
|
|
|
|
|
{
|
|
|
|
|
unsigned long val;
|
|
|
|
|
char *endptr;
|
|
|
|
|
|
|
|
|
|
val = simple_strtoul(p, &endptr, 10);
|
|
|
|
|
if (!*endptr)
|
|
|
|
|
stack_guard_gap = val << PAGE_SHIFT;
|
|
|
|
|
|
2022-03-23 05:42:27 +08:00
|
|
|
return 1;
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
}
|
|
|
|
|
__setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
|
|
|
|
|
|
2007-07-19 16:48:16 +08:00
|
|
|
#ifdef CONFIG_STACK_GROWSUP
|
|
|
|
|
int expand_stack(struct vm_area_struct *vma, unsigned long address)
|
|
|
|
|
{
|
|
|
|
|
return expand_upwards(vma, address);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
struct vm_area_struct *
|
|
|
|
|
find_extend_vma(struct mm_struct *mm, unsigned long addr)
|
|
|
|
|
{
|
|
|
|
|
struct vm_area_struct *vma, *prev;
|
|
|
|
|
|
|
|
|
|
addr &= PAGE_MASK;
|
|
|
|
|
vma = find_vma_prev(mm, addr, &prev);
|
|
|
|
|
if (vma && (vma->vm_start <= addr))
|
|
|
|
|
return vma;
|
2019-04-19 08:50:52 +08:00
|
|
|
/* don't alter vm_end if the coredump is running */
|
2020-10-16 11:13:00 +08:00
|
|
|
if (!prev || expand_stack(prev, addr))
|
2007-07-19 16:48:16 +08:00
|
|
|
return NULL;
|
2013-02-23 08:32:44 +08:00
|
|
|
if (prev->vm_flags & VM_LOCKED)
|
2015-04-15 06:44:39 +08:00
|
|
|
populate_vma_page_range(prev, addr, prev->vm_end, NULL);
|
2007-07-19 16:48:16 +08:00
|
|
|
return prev;
|
|
|
|
|
}
|
|
|
|
|
#else
|
|
|
|
|
int expand_stack(struct vm_area_struct *vma, unsigned long address)
|
|
|
|
|
{
|
|
|
|
|
return expand_downwards(vma, address);
|
|
|
|
|
}
|
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
struct vm_area_struct *
|
2014-10-10 06:26:29 +08:00
|
|
|
find_extend_vma(struct mm_struct *mm, unsigned long addr)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
2014-10-10 06:26:29 +08:00
|
|
|
struct vm_area_struct *vma;
|
2005-04-17 06:20:36 +08:00
|
|
|
unsigned long start;
|
|
|
|
|
|
|
|
|
|
addr &= PAGE_MASK;
|
2014-10-10 06:26:29 +08:00
|
|
|
vma = find_vma(mm, addr);
|
2005-04-17 06:20:36 +08:00
|
|
|
if (!vma)
|
|
|
|
|
return NULL;
|
|
|
|
|
if (vma->vm_start <= addr)
|
|
|
|
|
return vma;
|
|
|
|
|
if (!(vma->vm_flags & VM_GROWSDOWN))
|
|
|
|
|
return NULL;
|
|
|
|
|
start = vma->vm_start;
|
|
|
|
|
if (expand_stack(vma, addr))
|
|
|
|
|
return NULL;
|
2013-02-23 08:32:44 +08:00
|
|
|
if (vma->vm_flags & VM_LOCKED)
|
2015-04-15 06:44:39 +08:00
|
|
|
populate_vma_page_range(vma, addr, start, NULL);
|
2005-04-17 06:20:36 +08:00
|
|
|
return vma;
|
|
|
|
|
}
|
|
|
|
|
#endif
|
|
|
|
|
|
2014-12-13 08:55:27 +08:00
|
|
|
EXPORT_SYMBOL_GPL(find_extend_vma);
|
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
/*
|
2005-10-30 09:15:56 +08:00
|
|
|
* Ok - we have the memory areas we should free on the vma list,
|
2005-04-17 06:20:36 +08:00
|
|
|
* so release them, and do the vma updates.
|
2005-10-30 09:15:56 +08:00
|
|
|
*
|
|
|
|
|
* Called with the mm semaphore held.
|
2005-04-17 06:20:36 +08:00
|
|
|
*/
|
2005-10-30 09:15:56 +08:00
|
|
|
static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
2012-05-07 04:54:06 +08:00
|
|
|
unsigned long nr_accounted = 0;
|
|
|
|
|
|
[PATCH] mm: update_hiwaters just in time
update_mem_hiwater has attracted various criticisms, in particular from those
concerned with mm scalability. Originally it was called whenever rss or
total_vm got raised. Then many of those callsites were replaced by a timer
tick call from account_system_time. Now Frank van Maarseveen reports that to
be found inadequate. How about this? Works for Frank.
Replace update_mem_hiwater, a poor combination of two unrelated ops, by macros
update_hiwater_rss and update_hiwater_vm. Don't attempt to keep
mm->hiwater_rss up to date at timer tick, nor every time we raise rss (usually
by 1): those are hot paths. Do the opposite, update only when about to lower
rss (usually by many), or just before final accounting in do_exit. Handle
mm->hiwater_vm in the same way, though it's much less of an issue. Demand
that whoever collects these hiwater statistics do the work of taking the
maximum with rss or total_vm.
And there has been no collector of these hiwater statistics in the tree. The
new convention needs an example, so match Frank's usage by adding a VmPeak
line above VmSize to /proc/<pid>/status, and also a VmHWM line above VmRSS
(High-Water-Mark or High-Water-Memory).
There was a particular anomaly during mremap move, that hiwater_vm might be
captured too high. A fleeting such anomaly remains, but it's quickly
corrected now, whereas before it would stick.
What locking? None: if the app is racy then these statistics will be racy,
it's not worth any overhead to make them exact. But whenever it suits,
hiwater_vm is updated under exclusive mmap_sem, and hiwater_rss under
page_table_lock (for now) or with preemption disabled (later on): without
going to any trouble, minimize the time between reading current values and
updating, to minimize those occasions when a racing thread bumps a count up
and back down in between.
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-10-30 09:16:18 +08:00
|
|
|
/* Update high watermark before we lower total_vm */
|
|
|
|
|
update_hiwater_vm(mm);
|
2005-04-17 06:20:36 +08:00
|
|
|
do {
|
2005-10-30 09:15:56 +08:00
|
|
|
long nrpages = vma_pages(vma);
|
|
|
|
|
|
2012-05-07 04:54:06 +08:00
|
|
|
if (vma->vm_flags & VM_ACCOUNT)
|
|
|
|
|
nr_accounted += nrpages;
|
2016-01-15 07:22:07 +08:00
|
|
|
vm_stat_account(mm, vma->vm_flags, -nrpages);
|
2005-10-30 09:15:57 +08:00
|
|
|
vma = remove_vma(vma);
|
2005-04-20 04:29:18 +08:00
|
|
|
} while (vma);
|
2012-05-07 04:54:06 +08:00
|
|
|
vm_unacct_memory(nr_accounted);
|
2005-04-17 06:20:36 +08:00
|
|
|
validate_mm(mm);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Get rid of page table information in the indicated region.
|
|
|
|
|
*
|
2005-09-22 00:55:37 +08:00
|
|
|
* Called with the mm semaphore held.
|
2005-04-17 06:20:36 +08:00
|
|
|
*/
|
|
|
|
|
static void unmap_region(struct mm_struct *mm,
|
[PATCH] freepgt: free_pgtables use vma list
Recent woes with some arches needing their own pgd_addr_end macro; and 4-level
clear_page_range regression since 2.6.10's clear_page_tables; and its
long-standing well-known inefficiency in searching throughout the higher-level
page tables for those few entries to clear and free: all can be blamed on
ignoring the list of vmas when we free page tables.
Replace exit_mmap's clear_page_range of the total user address space by
free_pgtables operating on the mm's vma list; unmap_region use it in the same
way, giving floor and ceiling beyond which it may not free tables. This
brings lmbench fork/exec/sh numbers back to 2.6.10 (unless preempt is enabled,
in which case latency fixes spoil unmap_vmas throughput).
Beware: the do_mmap_pgoff driver failure case must now use unmap_region
instead of zap_page_range, since a page table might have been allocated, and
can only be freed while it is touched by some vma.
Move free_pgtables from mmap.c to memory.c, where its lower levels are adapted
from the clear_page_range levels. (Most of free_pgtables' old code was
actually for a non-existent case, prev not properly set up, dating from before
hch gave us split_vma.) Pass mmu_gather** in the public interfaces, since we
might want to add latency lockdrops later; but no attempt to do so yet, going
by vma should itself reduce latency.
But what if is_hugepage_only_range? Those ia64 and ppc64 cases need careful
examination: put that off until a later patch of the series.
What of x86_64's 32bit vdso page __map_syscall32 maps outside any vma?
And the range to sparc64's flush_tlb_pgtables? It's less clear to me now that
we need to do more than is done here - every PMD_SIZE ever occupied will be
flushed, do we really have to flush every PGDIR_SIZE ever partially occupied?
A shame to complicate it unnecessarily.
Special thanks to David Miller for time spent repairing my ceilings.
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-04-20 04:29:15 +08:00
|
|
|
struct vm_area_struct *vma, struct vm_area_struct *prev,
|
|
|
|
|
unsigned long start, unsigned long end)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
2020-10-18 07:14:06 +08:00
|
|
|
struct vm_area_struct *next = vma_next(mm, prev);
|
2011-05-25 08:11:45 +08:00
|
|
|
struct mmu_gather tlb;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
|
lru_add_drain();
|
2021-01-28 07:53:45 +08:00
|
|
|
tlb_gather_mmu(&tlb, mm);
|
[PATCH] mm: update_hiwaters just in time
update_mem_hiwater has attracted various criticisms, in particular from those
concerned with mm scalability. Originally it was called whenever rss or
total_vm got raised. Then many of those callsites were replaced by a timer
tick call from account_system_time. Now Frank van Maarseveen reports that to
be found inadequate. How about this? Works for Frank.
Replace update_mem_hiwater, a poor combination of two unrelated ops, by macros
update_hiwater_rss and update_hiwater_vm. Don't attempt to keep
mm->hiwater_rss up to date at timer tick, nor every time we raise rss (usually
by 1): those are hot paths. Do the opposite, update only when about to lower
rss (usually by many), or just before final accounting in do_exit. Handle
mm->hiwater_vm in the same way, though it's much less of an issue. Demand
that whoever collects these hiwater statistics do the work of taking the
maximum with rss or total_vm.
And there has been no collector of these hiwater statistics in the tree. The
new convention needs an example, so match Frank's usage by adding a VmPeak
line above VmSize to /proc/<pid>/status, and also a VmHWM line above VmRSS
(High-Water-Mark or High-Water-Memory).
There was a particular anomaly during mremap move, that hiwater_vm might be
captured too high. A fleeting such anomaly remains, but it's quickly
corrected now, whereas before it would stick.
What locking? None: if the app is racy then these statistics will be racy,
it's not worth any overhead to make them exact. But whenever it suits,
hiwater_vm is updated under exclusive mmap_sem, and hiwater_rss under
page_table_lock (for now) or with preemption disabled (later on): without
going to any trouble, minimize the time between reading current values and
updating, to minimize those occasions when a racing thread bumps a count up
and back down in between.
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-10-30 09:16:18 +08:00
|
|
|
update_hiwater_rss(mm);
|
2012-05-07 04:54:06 +08:00
|
|
|
unmap_vmas(&tlb, vma, start, end);
|
2011-05-25 08:11:45 +08:00
|
|
|
free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
|
2013-04-30 06:07:44 +08:00
|
|
|
next ? next->vm_start : USER_PGTABLES_CEILING);
|
2021-01-28 07:53:43 +08:00
|
|
|
tlb_finish_mmu(&tlb);
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Create a list of vma's touched by the unmap, removing them from the mm's
|
|
|
|
|
* vma list as we go..
|
|
|
|
|
*/
|
2020-07-24 12:15:11 +08:00
|
|
|
static bool
|
2005-04-17 06:20:36 +08:00
|
|
|
detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
|
|
|
|
|
struct vm_area_struct *prev, unsigned long end)
|
|
|
|
|
{
|
|
|
|
|
struct vm_area_struct **insertion_point;
|
|
|
|
|
struct vm_area_struct *tail_vma = NULL;
|
|
|
|
|
|
|
|
|
|
insertion_point = (prev ? &prev->vm_next : &mm->mmap);
|
2010-08-21 07:24:55 +08:00
|
|
|
vma->vm_prev = NULL;
|
2005-04-17 06:20:36 +08:00
|
|
|
do {
|
2012-12-12 08:01:38 +08:00
|
|
|
vma_rb_erase(vma, &mm->mm_rb);
|
2022-02-15 10:23:29 +08:00
|
|
|
if (vma->vm_flags & VM_LOCKED)
|
|
|
|
|
mm->locked_vm -= vma_pages(vma);
|
2005-04-17 06:20:36 +08:00
|
|
|
mm->map_count--;
|
|
|
|
|
tail_vma = vma;
|
|
|
|
|
vma = vma->vm_next;
|
|
|
|
|
} while (vma && vma->vm_start < end);
|
|
|
|
|
*insertion_point = vma;
|
2012-12-12 08:01:38 +08:00
|
|
|
if (vma) {
|
2010-08-21 07:24:55 +08:00
|
|
|
vma->vm_prev = prev;
|
2012-12-12 08:01:38 +08:00
|
|
|
vma_gap_update(vma);
|
|
|
|
|
} else
|
mm: larger stack guard gap, between vmas
Stack guard page is a useful feature to reduce a risk of stack smashing
into a different mapping. We have been using a single page gap which
is sufficient to prevent having stack adjacent to a different mapping.
But this seems to be insufficient in the light of the stack usage in
userspace. E.g. glibc uses as large as 64kB alloca() in many commonly
used functions. Others use constructs liks gid_t buffer[NGROUPS_MAX]
which is 256kB or stack strings with MAX_ARG_STRLEN.
This will become especially dangerous for suid binaries and the default
no limit for the stack size limit because those applications can be
tricked to consume a large portion of the stack and a single glibc call
could jump over the guard page. These attacks are not theoretical,
unfortunatelly.
Make those attacks less probable by increasing the stack guard gap
to 1MB (on systems with 4k pages; but make it depend on the page size
because systems with larger base pages might cap stack allocations in
the PAGE_SIZE units) which should cover larger alloca() and VLA stack
allocations. It is obviously not a full fix because the problem is
somehow inherent, but it should reduce attack space a lot.
One could argue that the gap size should be configurable from userspace,
but that can be done later when somebody finds that the new 1MB is wrong
for some special case applications. For now, add a kernel command line
option (stack_guard_gap) to specify the stack gap size (in page units).
Implementation wise, first delete all the old code for stack guard page:
because although we could get away with accounting one extra page in a
stack vma, accounting a larger gap can break userspace - case in point,
a program run with "ulimit -S -v 20000" failed when the 1MB gap was
counted for RLIMIT_AS; similar problems could come with RLIMIT_MLOCK
and strict non-overcommit mode.
Instead of keeping gap inside the stack vma, maintain the stack guard
gap as a gap between vmas: using vm_start_gap() in place of vm_start
(or vm_end_gap() in place of vm_end if VM_GROWSUP) in just those few
places which need to respect the gap - mainly arch_get_unmapped_area(),
and and the vma tree's subtree_gap support for that.
Original-patch-by: Oleg Nesterov <oleg@redhat.com>
Original-patch-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Tested-by: Helge Deller <deller@gmx.de> # parisc
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-19 19:03:24 +08:00
|
|
|
mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
|
2005-04-17 06:20:36 +08:00
|
|
|
tail_vma->vm_next = NULL;
|
mm: per-thread vma caching
This patch is a continuation of efforts trying to optimize find_vma(),
avoiding potentially expensive rbtree walks to locate a vma upon faults.
The original approach (https://lkml.org/lkml/2013/11/1/410), where the
largest vma was also cached, ended up being too specific and random,
thus further comparison with other approaches were needed. There are
two things to consider when dealing with this, the cache hit rate and
the latency of find_vma(). Improving the hit-rate does not necessarily
translate in finding the vma any faster, as the overhead of any fancy
caching schemes can be too high to consider.
We currently cache the last used vma for the whole address space, which
provides a nice optimization, reducing the total cycles in find_vma() by
up to 250%, for workloads with good locality. On the other hand, this
simple scheme is pretty much useless for workloads with poor locality.
Analyzing ebizzy runs shows that, no matter how many threads are
running, the mmap_cache hit rate is less than 2%, and in many situations
below 1%.
The proposed approach is to replace this scheme with a small per-thread
cache, maximizing hit rates at a very low maintenance cost.
Invalidations are performed by simply bumping up a 32-bit sequence
number. The only expensive operation is in the rare case of a seq
number overflow, where all caches that share the same address space are
flushed. Upon a miss, the proposed replacement policy is based on the
page number that contains the virtual address in question. Concretely,
the following results are seen on an 80 core, 8 socket x86-64 box:
1) System bootup: Most programs are single threaded, so the per-thread
scheme does improve ~50% hit rate by just adding a few more slots to
the cache.
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 50.61% | 19.90 |
| patched | 73.45% | 13.58 |
+----------------+----------+------------------+
2) Kernel build: This one is already pretty good with the current
approach as we're dealing with good locality.
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 75.28% | 11.03 |
| patched | 88.09% | 9.31 |
+----------------+----------+------------------+
3) Oracle 11g Data Mining (4k pages): Similar to the kernel build workload.
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 70.66% | 17.14 |
| patched | 91.15% | 12.57 |
+----------------+----------+------------------+
4) Ebizzy: There's a fair amount of variation from run to run, but this
approach always shows nearly perfect hit rates, while baseline is just
about non-existent. The amounts of cycles can fluctuate between
anywhere from ~60 to ~116 for the baseline scheme, but this approach
reduces it considerably. For instance, with 80 threads:
+----------------+----------+------------------+
| caching scheme | hit-rate | cycles (billion) |
+----------------+----------+------------------+
| baseline | 1.06% | 91.54 |
| patched | 99.97% | 14.18 |
+----------------+----------+------------------+
[akpm@linux-foundation.org: fix nommu build, per Davidlohr]
[akpm@linux-foundation.org: document vmacache_valid() logic]
[akpm@linux-foundation.org: attempt to untangle header files]
[akpm@linux-foundation.org: add vmacache_find() BUG_ON]
[hughd@google.com: add vmacache_valid_mm() (from Oleg)]
[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: adjust and enhance comments]
Signed-off-by: Davidlohr Bueso <davidlohr@hp.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Reviewed-by: Michel Lespinasse <walken@google.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Tested-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-04-08 06:37:25 +08:00
|
|
|
|
|
|
|
|
/* Kill the cache */
|
|
|
|
|
vmacache_invalidate(mm);
|
2020-07-24 12:15:11 +08:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Do not downgrade mmap_lock if we are next to VM_GROWSDOWN or
|
|
|
|
|
* VM_GROWSUP VMA. Such VMAs can change their size under
|
|
|
|
|
* down_read(mmap_lock) and collide with the VMA we are about to unmap.
|
|
|
|
|
*/
|
|
|
|
|
if (vma && (vma->vm_flags & VM_GROWSDOWN))
|
|
|
|
|
return false;
|
|
|
|
|
if (prev && (prev->vm_flags & VM_GROWSUP))
|
|
|
|
|
return false;
|
|
|
|
|
return true;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
2017-02-25 06:58:47 +08:00
|
|
|
* __split_vma() bypasses sysctl_max_map_count checking. We use this where it
|
|
|
|
|
* has already been checked or doesn't make sense to fail.
|
2005-04-17 06:20:36 +08:00
|
|
|
*/
|
2017-02-25 06:58:47 +08:00
|
|
|
int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
|
|
|
|
|
unsigned long addr, int new_below)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
|
|
|
|
struct vm_area_struct *new;
|
2015-09-09 06:03:38 +08:00
|
|
|
int err;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2020-12-15 11:08:17 +08:00
|
|
|
if (vma->vm_ops && vma->vm_ops->may_split) {
|
|
|
|
|
err = vma->vm_ops->may_split(vma, addr);
|
2017-11-30 08:10:28 +08:00
|
|
|
if (err)
|
|
|
|
|
return err;
|
|
|
|
|
}
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2018-07-22 04:48:51 +08:00
|
|
|
new = vm_area_dup(vma);
|
2005-04-17 06:20:36 +08:00
|
|
|
if (!new)
|
2015-09-09 06:03:38 +08:00
|
|
|
return -ENOMEM;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
|
if (new_below)
|
|
|
|
|
new->vm_end = addr;
|
|
|
|
|
else {
|
|
|
|
|
new->vm_start = addr;
|
|
|
|
|
new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
|
|
|
|
|
}
|
|
|
|
|
|
2013-09-12 05:20:14 +08:00
|
|
|
err = vma_dup_policy(vma, new);
|
|
|
|
|
if (err)
|
mm: change anon_vma linking to fix multi-process server scalability issue
The old anon_vma code can lead to scalability issues with heavily forking
workloads. Specifically, each anon_vma will be shared between the parent
process and all its child processes.
In a workload with 1000 child processes and a VMA with 1000 anonymous
pages per process that get COWed, this leads to a system with a million
anonymous pages in the same anon_vma, each of which is mapped in just one
of the 1000 processes. However, the current rmap code needs to walk them
all, leading to O(N) scanning complexity for each page.
This can result in systems where one CPU is walking the page tables of
1000 processes in page_referenced_one, while all other CPUs are stuck on
the anon_vma lock. This leads to catastrophic failure for a benchmark
like AIM7, where the total number of processes can reach in the tens of
thousands. Real workloads are still a factor 10 less process intensive
than AIM7, but they are catching up.
This patch changes the way anon_vmas and VMAs are linked, which allows us
to associate multiple anon_vmas with a VMA. At fork time, each child
process gets its own anon_vmas, in which its COWed pages will be
instantiated. The parents' anon_vma is also linked to the VMA, because
non-COWed pages could be present in any of the children.
This reduces rmap scanning complexity to O(1) for the pages of the 1000
child processes, with O(N) complexity for at most 1/N pages in the system.
This reduces the average scanning cost in heavily forking workloads from
O(N) to 2.
The only real complexity in this patch stems from the fact that linking a
VMA to anon_vmas now involves memory allocations. This means vma_adjust
can fail, if it needs to attach a VMA to anon_vma structures. This in
turn means error handling needs to be added to the calling functions.
A second source of complexity is that, because there can be multiple
anon_vmas, the anon_vma linking in vma_adjust can no longer be done under
"the" anon_vma lock. To prevent the rmap code from walking up an
incomplete VMA, this patch introduces the VM_LOCK_RMAP VMA flag. This bit
flag uses the same slot as the NOMMU VM_MAPPED_COPY, with an ifdef in mm.h
to make sure it is impossible to compile a kernel that needs both symbolic
values for the same bitflag.
Some test results:
Without the anon_vma changes, when AIM7 hits around 9.7k users (on a test
box with 16GB RAM and not quite enough IO), the system ends up running
>99% in system time, with every CPU on the same anon_vma lock in the
pageout code.
With these changes, AIM7 hits the cross-over point around 29.7k users.
This happens with ~99% IO wait time, there never seems to be any spike in
system time. The anon_vma lock contention appears to be resolved.
[akpm@linux-foundation.org: cleanups]
Signed-off-by: Rik van Riel <riel@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Larry Woodman <lwoodman@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-06 05:42:07 +08:00
|
|
|
goto out_free_vma;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2014-12-03 07:59:42 +08:00
|
|
|
err = anon_vma_clone(new, vma);
|
|
|
|
|
if (err)
|
mm: change anon_vma linking to fix multi-process server scalability issue
The old anon_vma code can lead to scalability issues with heavily forking
workloads. Specifically, each anon_vma will be shared between the parent
process and all its child processes.
In a workload with 1000 child processes and a VMA with 1000 anonymous
pages per process that get COWed, this leads to a system with a million
anonymous pages in the same anon_vma, each of which is mapped in just one
of the 1000 processes. However, the current rmap code needs to walk them
all, leading to O(N) scanning complexity for each page.
This can result in systems where one CPU is walking the page tables of
1000 processes in page_referenced_one, while all other CPUs are stuck on
the anon_vma lock. This leads to catastrophic failure for a benchmark
like AIM7, where the total number of processes can reach in the tens of
thousands. Real workloads are still a factor 10 less process intensive
than AIM7, but they are catching up.
This patch changes the way anon_vmas and VMAs are linked, which allows us
to associate multiple anon_vmas with a VMA. At fork time, each child
process gets its own anon_vmas, in which its COWed pages will be
instantiated. The parents' anon_vma is also linked to the VMA, because
non-COWed pages could be present in any of the children.
This reduces rmap scanning complexity to O(1) for the pages of the 1000
child processes, with O(N) complexity for at most 1/N pages in the system.
This reduces the average scanning cost in heavily forking workloads from
O(N) to 2.
The only real complexity in this patch stems from the fact that linking a
VMA to anon_vmas now involves memory allocations. This means vma_adjust
can fail, if it needs to attach a VMA to anon_vma structures. This in
turn means error handling needs to be added to the calling functions.
A second source of complexity is that, because there can be multiple
anon_vmas, the anon_vma linking in vma_adjust can no longer be done under
"the" anon_vma lock. To prevent the rmap code from walking up an
incomplete VMA, this patch introduces the VM_LOCK_RMAP VMA flag. This bit
flag uses the same slot as the NOMMU VM_MAPPED_COPY, with an ifdef in mm.h
to make sure it is impossible to compile a kernel that needs both symbolic
values for the same bitflag.
Some test results:
Without the anon_vma changes, when AIM7 hits around 9.7k users (on a test
box with 16GB RAM and not quite enough IO), the system ends up running
>99% in system time, with every CPU on the same anon_vma lock in the
pageout code.
With these changes, AIM7 hits the cross-over point around 29.7k users.
This happens with ~99% IO wait time, there never seems to be any spike in
system time. The anon_vma lock contention appears to be resolved.
[akpm@linux-foundation.org: cleanups]
Signed-off-by: Rik van Riel <riel@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Larry Woodman <lwoodman@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-06 05:42:07 +08:00
|
|
|
goto out_free_mpol;
|
|
|
|
|
|
2012-10-09 07:28:54 +08:00
|
|
|
if (new->vm_file)
|
2005-04-17 06:20:36 +08:00
|
|
|
get_file(new->vm_file);
|
|
|
|
|
|
|
|
|
|
if (new->vm_ops && new->vm_ops->open)
|
|
|
|
|
new->vm_ops->open(new);
|
|
|
|
|
|
|
|
|
|
if (new_below)
|
mm: change anon_vma linking to fix multi-process server scalability issue
The old anon_vma code can lead to scalability issues with heavily forking
workloads. Specifically, each anon_vma will be shared between the parent
process and all its child processes.
In a workload with 1000 child processes and a VMA with 1000 anonymous
pages per process that get COWed, this leads to a system with a million
anonymous pages in the same anon_vma, each of which is mapped in just one
of the 1000 processes. However, the current rmap code needs to walk them
all, leading to O(N) scanning complexity for each page.
This can result in systems where one CPU is walking the page tables of
1000 processes in page_referenced_one, while all other CPUs are stuck on
the anon_vma lock. This leads to catastrophic failure for a benchmark
like AIM7, where the total number of processes can reach in the tens of
thousands. Real workloads are still a factor 10 less process intensive
than AIM7, but they are catching up.
This patch changes the way anon_vmas and VMAs are linked, which allows us
to associate multiple anon_vmas with a VMA. At fork time, each child
process gets its own anon_vmas, in which its COWed pages will be
instantiated. The parents' anon_vma is also linked to the VMA, because
non-COWed pages could be present in any of the children.
This reduces rmap scanning complexity to O(1) for the pages of the 1000
child processes, with O(N) complexity for at most 1/N pages in the system.
This reduces the average scanning cost in heavily forking workloads from
O(N) to 2.
The only real complexity in this patch stems from the fact that linking a
VMA to anon_vmas now involves memory allocations. This means vma_adjust
can fail, if it needs to attach a VMA to anon_vma structures. This in
turn means error handling needs to be added to the calling functions.
A second source of complexity is that, because there can be multiple
anon_vmas, the anon_vma linking in vma_adjust can no longer be done under
"the" anon_vma lock. To prevent the rmap code from walking up an
incomplete VMA, this patch introduces the VM_LOCK_RMAP VMA flag. This bit
flag uses the same slot as the NOMMU VM_MAPPED_COPY, with an ifdef in mm.h
to make sure it is impossible to compile a kernel that needs both symbolic
values for the same bitflag.
Some test results:
Without the anon_vma changes, when AIM7 hits around 9.7k users (on a test
box with 16GB RAM and not quite enough IO), the system ends up running
>99% in system time, with every CPU on the same anon_vma lock in the
pageout code.
With these changes, AIM7 hits the cross-over point around 29.7k users.
This happens with ~99% IO wait time, there never seems to be any spike in
system time. The anon_vma lock contention appears to be resolved.
[akpm@linux-foundation.org: cleanups]
Signed-off-by: Rik van Riel <riel@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Larry Woodman <lwoodman@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-06 05:42:07 +08:00
|
|
|
err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
|
2005-04-17 06:20:36 +08:00
|
|
|
((addr - new->vm_start) >> PAGE_SHIFT), new);
|
|
|
|
|
else
|
mm: change anon_vma linking to fix multi-process server scalability issue
The old anon_vma code can lead to scalability issues with heavily forking
workloads. Specifically, each anon_vma will be shared between the parent
process and all its child processes.
In a workload with 1000 child processes and a VMA with 1000 anonymous
pages per process that get COWed, this leads to a system with a million
anonymous pages in the same anon_vma, each of which is mapped in just one
of the 1000 processes. However, the current rmap code needs to walk them
all, leading to O(N) scanning complexity for each page.
This can result in systems where one CPU is walking the page tables of
1000 processes in page_referenced_one, while all other CPUs are stuck on
the anon_vma lock. This leads to catastrophic failure for a benchmark
like AIM7, where the total number of processes can reach in the tens of
thousands. Real workloads are still a factor 10 less process intensive
than AIM7, but they are catching up.
This patch changes the way anon_vmas and VMAs are linked, which allows us
to associate multiple anon_vmas with a VMA. At fork time, each child
process gets its own anon_vmas, in which its COWed pages will be
instantiated. The parents' anon_vma is also linked to the VMA, because
non-COWed pages could be present in any of the children.
This reduces rmap scanning complexity to O(1) for the pages of the 1000
child processes, with O(N) complexity for at most 1/N pages in the system.
This reduces the average scanning cost in heavily forking workloads from
O(N) to 2.
The only real complexity in this patch stems from the fact that linking a
VMA to anon_vmas now involves memory allocations. This means vma_adjust
can fail, if it needs to attach a VMA to anon_vma structures. This in
turn means error handling needs to be added to the calling functions.
A second source of complexity is that, because there can be multiple
anon_vmas, the anon_vma linking in vma_adjust can no longer be done under
"the" anon_vma lock. To prevent the rmap code from walking up an
incomplete VMA, this patch introduces the VM_LOCK_RMAP VMA flag. This bit
flag uses the same slot as the NOMMU VM_MAPPED_COPY, with an ifdef in mm.h
to make sure it is impossible to compile a kernel that needs both symbolic
values for the same bitflag.
Some test results:
Without the anon_vma changes, when AIM7 hits around 9.7k users (on a test
box with 16GB RAM and not quite enough IO), the system ends up running
>99% in system time, with every CPU on the same anon_vma lock in the
pageout code.
With these changes, AIM7 hits the cross-over point around 29.7k users.
This happens with ~99% IO wait time, there never seems to be any spike in
system time. The anon_vma lock contention appears to be resolved.
[akpm@linux-foundation.org: cleanups]
Signed-off-by: Rik van Riel <riel@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Larry Woodman <lwoodman@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-06 05:42:07 +08:00
|
|
|
err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
|
2005-04-17 06:20:36 +08:00
|
|
|
|
mm: change anon_vma linking to fix multi-process server scalability issue
The old anon_vma code can lead to scalability issues with heavily forking
workloads. Specifically, each anon_vma will be shared between the parent
process and all its child processes.
In a workload with 1000 child processes and a VMA with 1000 anonymous
pages per process that get COWed, this leads to a system with a million
anonymous pages in the same anon_vma, each of which is mapped in just one
of the 1000 processes. However, the current rmap code needs to walk them
all, leading to O(N) scanning complexity for each page.
This can result in systems where one CPU is walking the page tables of
1000 processes in page_referenced_one, while all other CPUs are stuck on
the anon_vma lock. This leads to catastrophic failure for a benchmark
like AIM7, where the total number of processes can reach in the tens of
thousands. Real workloads are still a factor 10 less process intensive
than AIM7, but they are catching up.
This patch changes the way anon_vmas and VMAs are linked, which allows us
to associate multiple anon_vmas with a VMA. At fork time, each child
process gets its own anon_vmas, in which its COWed pages will be
instantiated. The parents' anon_vma is also linked to the VMA, because
non-COWed pages could be present in any of the children.
This reduces rmap scanning complexity to O(1) for the pages of the 1000
child processes, with O(N) complexity for at most 1/N pages in the system.
This reduces the average scanning cost in heavily forking workloads from
O(N) to 2.
The only real complexity in this patch stems from the fact that linking a
VMA to anon_vmas now involves memory allocations. This means vma_adjust
can fail, if it needs to attach a VMA to anon_vma structures. This in
turn means error handling needs to be added to the calling functions.
A second source of complexity is that, because there can be multiple
anon_vmas, the anon_vma linking in vma_adjust can no longer be done under
"the" anon_vma lock. To prevent the rmap code from walking up an
incomplete VMA, this patch introduces the VM_LOCK_RMAP VMA flag. This bit
flag uses the same slot as the NOMMU VM_MAPPED_COPY, with an ifdef in mm.h
to make sure it is impossible to compile a kernel that needs both symbolic
values for the same bitflag.
Some test results:
Without the anon_vma changes, when AIM7 hits around 9.7k users (on a test
box with 16GB RAM and not quite enough IO), the system ends up running
>99% in system time, with every CPU on the same anon_vma lock in the
pageout code.
With these changes, AIM7 hits the cross-over point around 29.7k users.
This happens with ~99% IO wait time, there never seems to be any spike in
system time. The anon_vma lock contention appears to be resolved.
[akpm@linux-foundation.org: cleanups]
Signed-off-by: Rik van Riel <riel@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Larry Woodman <lwoodman@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-06 05:42:07 +08:00
|
|
|
/* Success. */
|
|
|
|
|
if (!err)
|
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
|
|
/* Clean everything up if vma_adjust failed. */
|
2010-04-27 00:33:03 +08:00
|
|
|
if (new->vm_ops && new->vm_ops->close)
|
|
|
|
|
new->vm_ops->close(new);
|
2012-10-09 07:28:54 +08:00
|
|
|
if (new->vm_file)
|
mm: change anon_vma linking to fix multi-process server scalability issue
The old anon_vma code can lead to scalability issues with heavily forking
workloads. Specifically, each anon_vma will be shared between the parent
process and all its child processes.
In a workload with 1000 child processes and a VMA with 1000 anonymous
pages per process that get COWed, this leads to a system with a million
anonymous pages in the same anon_vma, each of which is mapped in just one
of the 1000 processes. However, the current rmap code needs to walk them
all, leading to O(N) scanning complexity for each page.
This can result in systems where one CPU is walking the page tables of
1000 processes in page_referenced_one, while all other CPUs are stuck on
the anon_vma lock. This leads to catastrophic failure for a benchmark
like AIM7, where the total number of processes can reach in the tens of
thousands. Real workloads are still a factor 10 less process intensive
than AIM7, but they are catching up.
This patch changes the way anon_vmas and VMAs are linked, which allows us
to associate multiple anon_vmas with a VMA. At fork time, each child
process gets its own anon_vmas, in which its COWed pages will be
instantiated. The parents' anon_vma is also linked to the VMA, because
non-COWed pages could be present in any of the children.
This reduces rmap scanning complexity to O(1) for the pages of the 1000
child processes, with O(N) complexity for at most 1/N pages in the system.
This reduces the average scanning cost in heavily forking workloads from
O(N) to 2.
The only real complexity in this patch stems from the fact that linking a
VMA to anon_vmas now involves memory allocations. This means vma_adjust
can fail, if it needs to attach a VMA to anon_vma structures. This in
turn means error handling needs to be added to the calling functions.
A second source of complexity is that, because there can be multiple
anon_vmas, the anon_vma linking in vma_adjust can no longer be done under
"the" anon_vma lock. To prevent the rmap code from walking up an
incomplete VMA, this patch introduces the VM_LOCK_RMAP VMA flag. This bit
flag uses the same slot as the NOMMU VM_MAPPED_COPY, with an ifdef in mm.h
to make sure it is impossible to compile a kernel that needs both symbolic
values for the same bitflag.
Some test results:
Without the anon_vma changes, when AIM7 hits around 9.7k users (on a test
box with 16GB RAM and not quite enough IO), the system ends up running
>99% in system time, with every CPU on the same anon_vma lock in the
pageout code.
With these changes, AIM7 hits the cross-over point around 29.7k users.
This happens with ~99% IO wait time, there never seems to be any spike in
system time. The anon_vma lock contention appears to be resolved.
[akpm@linux-foundation.org: cleanups]
Signed-off-by: Rik van Riel <riel@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Larry Woodman <lwoodman@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-06 05:42:07 +08:00
|
|
|
fput(new->vm_file);
|
2010-09-23 04:05:12 +08:00
|
|
|
unlink_anon_vmas(new);
|
mm: change anon_vma linking to fix multi-process server scalability issue
The old anon_vma code can lead to scalability issues with heavily forking
workloads. Specifically, each anon_vma will be shared between the parent
process and all its child processes.
In a workload with 1000 child processes and a VMA with 1000 anonymous
pages per process that get COWed, this leads to a system with a million
anonymous pages in the same anon_vma, each of which is mapped in just one
of the 1000 processes. However, the current rmap code needs to walk them
all, leading to O(N) scanning complexity for each page.
This can result in systems where one CPU is walking the page tables of
1000 processes in page_referenced_one, while all other CPUs are stuck on
the anon_vma lock. This leads to catastrophic failure for a benchmark
like AIM7, where the total number of processes can reach in the tens of
thousands. Real workloads are still a factor 10 less process intensive
than AIM7, but they are catching up.
This patch changes the way anon_vmas and VMAs are linked, which allows us
to associate multiple anon_vmas with a VMA. At fork time, each child
process gets its own anon_vmas, in which its COWed pages will be
instantiated. The parents' anon_vma is also linked to the VMA, because
non-COWed pages could be present in any of the children.
This reduces rmap scanning complexity to O(1) for the pages of the 1000
child processes, with O(N) complexity for at most 1/N pages in the system.
This reduces the average scanning cost in heavily forking workloads from
O(N) to 2.
The only real complexity in this patch stems from the fact that linking a
VMA to anon_vmas now involves memory allocations. This means vma_adjust
can fail, if it needs to attach a VMA to anon_vma structures. This in
turn means error handling needs to be added to the calling functions.
A second source of complexity is that, because there can be multiple
anon_vmas, the anon_vma linking in vma_adjust can no longer be done under
"the" anon_vma lock. To prevent the rmap code from walking up an
incomplete VMA, this patch introduces the VM_LOCK_RMAP VMA flag. This bit
flag uses the same slot as the NOMMU VM_MAPPED_COPY, with an ifdef in mm.h
to make sure it is impossible to compile a kernel that needs both symbolic
values for the same bitflag.
Some test results:
Without the anon_vma changes, when AIM7 hits around 9.7k users (on a test
box with 16GB RAM and not quite enough IO), the system ends up running
>99% in system time, with every CPU on the same anon_vma lock in the
pageout code.
With these changes, AIM7 hits the cross-over point around 29.7k users.
This happens with ~99% IO wait time, there never seems to be any spike in
system time. The anon_vma lock contention appears to be resolved.
[akpm@linux-foundation.org: cleanups]
Signed-off-by: Rik van Riel <riel@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Larry Woodman <lwoodman@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-06 05:42:07 +08:00
|
|
|
out_free_mpol:
|
2013-09-12 05:20:14 +08:00
|
|
|
mpol_put(vma_policy(new));
|
mm: change anon_vma linking to fix multi-process server scalability issue
The old anon_vma code can lead to scalability issues with heavily forking
workloads. Specifically, each anon_vma will be shared between the parent
process and all its child processes.
In a workload with 1000 child processes and a VMA with 1000 anonymous
pages per process that get COWed, this leads to a system with a million
anonymous pages in the same anon_vma, each of which is mapped in just one
of the 1000 processes. However, the current rmap code needs to walk them
all, leading to O(N) scanning complexity for each page.
This can result in systems where one CPU is walking the page tables of
1000 processes in page_referenced_one, while all other CPUs are stuck on
the anon_vma lock. This leads to catastrophic failure for a benchmark
like AIM7, where the total number of processes can reach in the tens of
thousands. Real workloads are still a factor 10 less process intensive
than AIM7, but they are catching up.
This patch changes the way anon_vmas and VMAs are linked, which allows us
to associate multiple anon_vmas with a VMA. At fork time, each child
process gets its own anon_vmas, in which its COWed pages will be
instantiated. The parents' anon_vma is also linked to the VMA, because
non-COWed pages could be present in any of the children.
This reduces rmap scanning complexity to O(1) for the pages of the 1000
child processes, with O(N) complexity for at most 1/N pages in the system.
This reduces the average scanning cost in heavily forking workloads from
O(N) to 2.
The only real complexity in this patch stems from the fact that linking a
VMA to anon_vmas now involves memory allocations. This means vma_adjust
can fail, if it needs to attach a VMA to anon_vma structures. This in
turn means error handling needs to be added to the calling functions.
A second source of complexity is that, because there can be multiple
anon_vmas, the anon_vma linking in vma_adjust can no longer be done under
"the" anon_vma lock. To prevent the rmap code from walking up an
incomplete VMA, this patch introduces the VM_LOCK_RMAP VMA flag. This bit
flag uses the same slot as the NOMMU VM_MAPPED_COPY, with an ifdef in mm.h
to make sure it is impossible to compile a kernel that needs both symbolic
values for the same bitflag.
Some test results:
Without the anon_vma changes, when AIM7 hits around 9.7k users (on a test
box with 16GB RAM and not quite enough IO), the system ends up running
>99% in system time, with every CPU on the same anon_vma lock in the
pageout code.
With these changes, AIM7 hits the cross-over point around 29.7k users.
This happens with ~99% IO wait time, there never seems to be any spike in
system time. The anon_vma lock contention appears to be resolved.
[akpm@linux-foundation.org: cleanups]
Signed-off-by: Rik van Riel <riel@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Larry Woodman <lwoodman@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-06 05:42:07 +08:00
|
|
|
out_free_vma:
|
2018-07-22 04:48:51 +08:00
|
|
|
vm_area_free(new);
|
mm: change anon_vma linking to fix multi-process server scalability issue
The old anon_vma code can lead to scalability issues with heavily forking
workloads. Specifically, each anon_vma will be shared between the parent
process and all its child processes.
In a workload with 1000 child processes and a VMA with 1000 anonymous
pages per process that get COWed, this leads to a system with a million
anonymous pages in the same anon_vma, each of which is mapped in just one
of the 1000 processes. However, the current rmap code needs to walk them
all, leading to O(N) scanning complexity for each page.
This can result in systems where one CPU is walking the page tables of
1000 processes in page_referenced_one, while all other CPUs are stuck on
the anon_vma lock. This leads to catastrophic failure for a benchmark
like AIM7, where the total number of processes can reach in the tens of
thousands. Real workloads are still a factor 10 less process intensive
than AIM7, but they are catching up.
This patch changes the way anon_vmas and VMAs are linked, which allows us
to associate multiple anon_vmas with a VMA. At fork time, each child
process gets its own anon_vmas, in which its COWed pages will be
instantiated. The parents' anon_vma is also linked to the VMA, because
non-COWed pages could be present in any of the children.
This reduces rmap scanning complexity to O(1) for the pages of the 1000
child processes, with O(N) complexity for at most 1/N pages in the system.
This reduces the average scanning cost in heavily forking workloads from
O(N) to 2.
The only real complexity in this patch stems from the fact that linking a
VMA to anon_vmas now involves memory allocations. This means vma_adjust
can fail, if it needs to attach a VMA to anon_vma structures. This in
turn means error handling needs to be added to the calling functions.
A second source of complexity is that, because there can be multiple
anon_vmas, the anon_vma linking in vma_adjust can no longer be done under
"the" anon_vma lock. To prevent the rmap code from walking up an
incomplete VMA, this patch introduces the VM_LOCK_RMAP VMA flag. This bit
flag uses the same slot as the NOMMU VM_MAPPED_COPY, with an ifdef in mm.h
to make sure it is impossible to compile a kernel that needs both symbolic
values for the same bitflag.
Some test results:
Without the anon_vma changes, when AIM7 hits around 9.7k users (on a test
box with 16GB RAM and not quite enough IO), the system ends up running
>99% in system time, with every CPU on the same anon_vma lock in the
pageout code.
With these changes, AIM7 hits the cross-over point around 29.7k users.
This happens with ~99% IO wait time, there never seems to be any spike in
system time. The anon_vma lock contention appears to be resolved.
[akpm@linux-foundation.org: cleanups]
Signed-off-by: Rik van Riel <riel@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Larry Woodman <lwoodman@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-06 05:42:07 +08:00
|
|
|
return err;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
|
mmap: don't return ENOMEM when mapcount is temporarily exceeded in munmap()
On ia64, the following test program exit abnormally, because glibc thread
library called abort().
========================================================
(gdb) bt
#0 0xa000000000010620 in __kernel_syscall_via_break ()
#1 0x20000000003208e0 in raise () from /lib/libc.so.6.1
#2 0x2000000000324090 in abort () from /lib/libc.so.6.1
#3 0x200000000027c3e0 in __deallocate_stack () from /lib/libpthread.so.0
#4 0x200000000027f7c0 in start_thread () from /lib/libpthread.so.0
#5 0x200000000047ef60 in __clone2 () from /lib/libc.so.6.1
========================================================
The fact is, glibc call munmap() when thread exitng time for freeing
stack, and it assume munlock() never fail. However, munmap() often make
vma splitting and it with many mapcount make -ENOMEM.
Oh well, that's crazy, because stack unmapping never increase mapcount.
The maxcount exceeding is only temporary. internal temporary exceeding
shouldn't make ENOMEM.
This patch does it.
test_max_mapcount.c
==================================================================
#include<stdio.h>
#include<stdlib.h>
#include<string.h>
#include<pthread.h>
#include<errno.h>
#include<unistd.h>
#define THREAD_NUM 30000
#define MAL_SIZE (8*1024*1024)
void *wait_thread(void *args)
{
void *addr;
addr = malloc(MAL_SIZE);
sleep(10);
return NULL;
}
void *wait_thread2(void *args)
{
sleep(60);
return NULL;
}
int main(int argc, char *argv[])
{
int i;
pthread_t thread[THREAD_NUM], th;
int ret, count = 0;
pthread_attr_t attr;
ret = pthread_attr_init(&attr);
if(ret) {
perror("pthread_attr_init");
}
ret = pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED);
if(ret) {
perror("pthread_attr_setdetachstate");
}
for (i = 0; i < THREAD_NUM; i++) {
ret = pthread_create(&th, &attr, wait_thread, NULL);
if(ret) {
fprintf(stderr, "[%d] ", count);
perror("pthread_create");
} else {
printf("[%d] create OK.\n", count);
}
count++;
ret = pthread_create(&thread[i], &attr, wait_thread2, NULL);
if(ret) {
fprintf(stderr, "[%d] ", count);
perror("pthread_create");
} else {
printf("[%d] create OK.\n", count);
}
count++;
}
sleep(3600);
return 0;
}
==================================================================
[akpm@linux-foundation.org: coding-style fixes]
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 09:57:56 +08:00
|
|
|
/*
|
|
|
|
|
* Split a vma into two pieces at address 'addr', a new vma is allocated
|
|
|
|
|
* either for the first part or the tail.
|
|
|
|
|
*/
|
|
|
|
|
int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
|
|
|
|
|
unsigned long addr, int new_below)
|
|
|
|
|
{
|
|
|
|
|
if (mm->map_count >= sysctl_max_map_count)
|
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
|
|
return __split_vma(mm, vma, addr, new_below);
|
|
|
|
|
}
|
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
/* Munmap is split into 2 main parts -- this part which finds
|
|
|
|
|
* what needs doing, and the areas themselves, which do the
|
|
|
|
|
* work. This now handles partial unmappings.
|
|
|
|
|
* Jeremy Fitzhardinge <jeremy@goop.org>
|
|
|
|
|
*/
|
2018-10-27 06:08:50 +08:00
|
|
|
int __do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
|
|
|
|
|
struct list_head *uf, bool downgrade)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
|
|
|
|
unsigned long end;
|
2005-04-20 04:29:18 +08:00
|
|
|
struct vm_area_struct *vma, *prev, *last;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2015-11-06 10:46:54 +08:00
|
|
|
if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
|
2005-04-17 06:20:36 +08:00
|
|
|
return -EINVAL;
|
|
|
|
|
|
2014-10-10 06:26:29 +08:00
|
|
|
len = PAGE_ALIGN(len);
|
x86/mpx, mm/core: Fix recursive munmap() corruption
This is a bit of a mess, to put it mildly. But, it's a bug
that only seems to have showed up in 4.20 but wasn't noticed
until now, because nobody uses MPX.
MPX has the arch_unmap() hook inside of munmap() because MPX
uses bounds tables that protect other areas of memory. When
memory is unmapped, there is also a need to unmap the MPX
bounds tables. Barring this, unused bounds tables can eat 80%
of the address space.
But, the recursive do_munmap() that gets called vi arch_unmap()
wreaks havoc with __do_munmap()'s state. It can result in
freeing populated page tables, accessing bogus VMA state,
double-freed VMAs and more.
See the "long story" further below for the gory details.
To fix this, call arch_unmap() before __do_unmap() has a chance
to do anything meaningful. Also, remove the 'vma' argument
and force the MPX code to do its own, independent VMA lookup.
== UML / unicore32 impact ==
Remove unused 'vma' argument to arch_unmap(). No functional
change.
I compile tested this on UML but not unicore32.
== powerpc impact ==
powerpc uses arch_unmap() well to watch for munmap() on the
VDSO and zeroes out 'current->mm->context.vdso_base'. Moving
arch_unmap() makes this happen earlier in __do_munmap(). But,
'vdso_base' seems to only be used in perf and in the signal
delivery that happens near the return to userspace. I can not
find any likely impact to powerpc, other than the zeroing
happening a little earlier.
powerpc does not use the 'vma' argument and is unaffected by
its removal.
I compile-tested a 64-bit powerpc defconfig.
== x86 impact ==
For the common success case this is functionally identical to
what was there before. For the munmap() failure case, it's
possible that some MPX tables will be zapped for memory that
continues to be in use. But, this is an extraordinarily
unlikely scenario and the harm would be that MPX provides no
protection since the bounds table got reset (zeroed).
I can't imagine anyone doing this:
ptr = mmap();
// use ptr
ret = munmap(ptr);
if (ret)
// oh, there was an error, I'll
// keep using ptr.
Because if you're doing munmap(), you are *done* with the
memory. There's probably no good data in there _anyway_.
This passes the original reproducer from Richard Biener as
well as the existing mpx selftests/.
The long story:
munmap() has a couple of pieces:
1. Find the affected VMA(s)
2. Split the start/end one(s) if neceesary
3. Pull the VMAs out of the rbtree
4. Actually zap the memory via unmap_region(), including
freeing page tables (or queueing them to be freed).
5. Fix up some of the accounting (like fput()) and actually
free the VMA itself.
This specific ordering was actually introduced by:
dd2283f2605e ("mm: mmap: zap pages with read mmap_sem in munmap")
during the 4.20 merge window. The previous __do_munmap() code
was actually safe because the only thing after arch_unmap() was
remove_vma_list(). arch_unmap() could not see 'vma' in the
rbtree because it was detached, so it is not even capable of
doing operations unsafe for remove_vma_list()'s use of 'vma'.
Richard Biener reported a test that shows this in dmesg:
[1216548.787498] BUG: Bad rss-counter state mm:0000000017ce560b idx:1 val:551
[1216548.787500] BUG: non-zero pgtables_bytes on freeing mm: 24576
What triggered this was the recursive do_munmap() called via
arch_unmap(). It was freeing page tables that has not been
properly zapped.
But, the problem was bigger than this. For one, arch_unmap()
can free VMAs. But, the calling __do_munmap() has variables
that *point* to VMAs and obviously can't handle them just
getting freed while the pointer is still in use.
I tried a couple of things here. First, I tried to fix the page
table freeing problem in isolation, but I then found the VMA
issue. I also tried having the MPX code return a flag if it
modified the rbtree which would force __do_munmap() to re-walk
to restart. That spiralled out of control in complexity pretty
fast.
Just moving arch_unmap() and accepting that the bonkers failure
case might eat some bounds tables seems like the simplest viable
fix.
This was also reported in the following kernel bugzilla entry:
https://bugzilla.kernel.org/show_bug.cgi?id=203123
There are some reports that this commit triggered this bug:
dd2283f2605 ("mm: mmap: zap pages with read mmap_sem in munmap")
While that commit certainly made the issues easier to hit, I believe
the fundamental issue has been with us as long as MPX itself, thus
the Fixes: tag below is for one of the original MPX commits.
[ mingo: Minor edits to the changelog and the patch. ]
Reported-by: Richard Biener <rguenther@suse.de>
Reported-by: H.J. Lu <hjl.tools@gmail.com>
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Reviewed-by Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Yang Shi <yang.shi@linux.alibaba.com>
Acked-by: Michael Ellerman <mpe@ellerman.id.au>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Anton Ivanov <anton.ivanov@cambridgegreys.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Guan Xuetao <gxt@pku.edu.cn>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Jeff Dike <jdike@addtoit.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Richard Weinberger <richard@nod.at>
Cc: Rik van Riel <riel@surriel.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: linux-arch@vger.kernel.org
Cc: linux-mm@kvack.org
Cc: linux-um@lists.infradead.org
Cc: linuxppc-dev@lists.ozlabs.org
Cc: stable@vger.kernel.org
Fixes: dd2283f2605e ("mm: mmap: zap pages with read mmap_sem in munmap")
Link: http://lkml.kernel.org/r/20190419194747.5E1AD6DC@viggo.jf.intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-20 03:47:47 +08:00
|
|
|
end = start + len;
|
2014-10-10 06:26:29 +08:00
|
|
|
if (len == 0)
|
2005-04-17 06:20:36 +08:00
|
|
|
return -EINVAL;
|
|
|
|
|
|
x86/mpx, mm/core: Fix recursive munmap() corruption
This is a bit of a mess, to put it mildly. But, it's a bug
that only seems to have showed up in 4.20 but wasn't noticed
until now, because nobody uses MPX.
MPX has the arch_unmap() hook inside of munmap() because MPX
uses bounds tables that protect other areas of memory. When
memory is unmapped, there is also a need to unmap the MPX
bounds tables. Barring this, unused bounds tables can eat 80%
of the address space.
But, the recursive do_munmap() that gets called vi arch_unmap()
wreaks havoc with __do_munmap()'s state. It can result in
freeing populated page tables, accessing bogus VMA state,
double-freed VMAs and more.
See the "long story" further below for the gory details.
To fix this, call arch_unmap() before __do_unmap() has a chance
to do anything meaningful. Also, remove the 'vma' argument
and force the MPX code to do its own, independent VMA lookup.
== UML / unicore32 impact ==
Remove unused 'vma' argument to arch_unmap(). No functional
change.
I compile tested this on UML but not unicore32.
== powerpc impact ==
powerpc uses arch_unmap() well to watch for munmap() on the
VDSO and zeroes out 'current->mm->context.vdso_base'. Moving
arch_unmap() makes this happen earlier in __do_munmap(). But,
'vdso_base' seems to only be used in perf and in the signal
delivery that happens near the return to userspace. I can not
find any likely impact to powerpc, other than the zeroing
happening a little earlier.
powerpc does not use the 'vma' argument and is unaffected by
its removal.
I compile-tested a 64-bit powerpc defconfig.
== x86 impact ==
For the common success case this is functionally identical to
what was there before. For the munmap() failure case, it's
possible that some MPX tables will be zapped for memory that
continues to be in use. But, this is an extraordinarily
unlikely scenario and the harm would be that MPX provides no
protection since the bounds table got reset (zeroed).
I can't imagine anyone doing this:
ptr = mmap();
// use ptr
ret = munmap(ptr);
if (ret)
// oh, there was an error, I'll
// keep using ptr.
Because if you're doing munmap(), you are *done* with the
memory. There's probably no good data in there _anyway_.
This passes the original reproducer from Richard Biener as
well as the existing mpx selftests/.
The long story:
munmap() has a couple of pieces:
1. Find the affected VMA(s)
2. Split the start/end one(s) if neceesary
3. Pull the VMAs out of the rbtree
4. Actually zap the memory via unmap_region(), including
freeing page tables (or queueing them to be freed).
5. Fix up some of the accounting (like fput()) and actually
free the VMA itself.
This specific ordering was actually introduced by:
dd2283f2605e ("mm: mmap: zap pages with read mmap_sem in munmap")
during the 4.20 merge window. The previous __do_munmap() code
was actually safe because the only thing after arch_unmap() was
remove_vma_list(). arch_unmap() could not see 'vma' in the
rbtree because it was detached, so it is not even capable of
doing operations unsafe for remove_vma_list()'s use of 'vma'.
Richard Biener reported a test that shows this in dmesg:
[1216548.787498] BUG: Bad rss-counter state mm:0000000017ce560b idx:1 val:551
[1216548.787500] BUG: non-zero pgtables_bytes on freeing mm: 24576
What triggered this was the recursive do_munmap() called via
arch_unmap(). It was freeing page tables that has not been
properly zapped.
But, the problem was bigger than this. For one, arch_unmap()
can free VMAs. But, the calling __do_munmap() has variables
that *point* to VMAs and obviously can't handle them just
getting freed while the pointer is still in use.
I tried a couple of things here. First, I tried to fix the page
table freeing problem in isolation, but I then found the VMA
issue. I also tried having the MPX code return a flag if it
modified the rbtree which would force __do_munmap() to re-walk
to restart. That spiralled out of control in complexity pretty
fast.
Just moving arch_unmap() and accepting that the bonkers failure
case might eat some bounds tables seems like the simplest viable
fix.
This was also reported in the following kernel bugzilla entry:
https://bugzilla.kernel.org/show_bug.cgi?id=203123
There are some reports that this commit triggered this bug:
dd2283f2605 ("mm: mmap: zap pages with read mmap_sem in munmap")
While that commit certainly made the issues easier to hit, I believe
the fundamental issue has been with us as long as MPX itself, thus
the Fixes: tag below is for one of the original MPX commits.
[ mingo: Minor edits to the changelog and the patch. ]
Reported-by: Richard Biener <rguenther@suse.de>
Reported-by: H.J. Lu <hjl.tools@gmail.com>
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Reviewed-by Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Yang Shi <yang.shi@linux.alibaba.com>
Acked-by: Michael Ellerman <mpe@ellerman.id.au>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Anton Ivanov <anton.ivanov@cambridgegreys.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Guan Xuetao <gxt@pku.edu.cn>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Jeff Dike <jdike@addtoit.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Richard Weinberger <richard@nod.at>
Cc: Rik van Riel <riel@surriel.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: linux-arch@vger.kernel.org
Cc: linux-mm@kvack.org
Cc: linux-um@lists.infradead.org
Cc: linuxppc-dev@lists.ozlabs.org
Cc: stable@vger.kernel.org
Fixes: dd2283f2605e ("mm: mmap: zap pages with read mmap_sem in munmap")
Link: http://lkml.kernel.org/r/20190419194747.5E1AD6DC@viggo.jf.intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-20 03:47:47 +08:00
|
|
|
/*
|
|
|
|
|
* arch_unmap() might do unmaps itself. It must be called
|
|
|
|
|
* and finish any rbtree manipulation before this code
|
|
|
|
|
* runs and also starts to manipulate the rbtree.
|
|
|
|
|
*/
|
|
|
|
|
arch_unmap(mm, start, end);
|
|
|
|
|
|
2021-06-29 10:38:39 +08:00
|
|
|
/* Find the first overlapping VMA where start < vma->vm_end */
|
|
|
|
|
vma = find_vma_intersection(mm, start, end);
|
2005-04-20 04:29:18 +08:00
|
|
|
if (!vma)
|
2005-04-17 06:20:36 +08:00
|
|
|
return 0;
|
2011-06-16 15:35:09 +08:00
|
|
|
prev = vma->vm_prev;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* If we need to split any vma, do it now to save pain later.
|
|
|
|
|
*
|
|
|
|
|
* Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
|
|
|
|
|
* unmapped vm_area_struct will remain in use: so lower split_vma
|
|
|
|
|
* places tmp vma above, and higher split_vma places tmp vma below.
|
|
|
|
|
*/
|
2005-04-20 04:29:18 +08:00
|
|
|
if (start > vma->vm_start) {
|
mmap: don't return ENOMEM when mapcount is temporarily exceeded in munmap()
On ia64, the following test program exit abnormally, because glibc thread
library called abort().
========================================================
(gdb) bt
#0 0xa000000000010620 in __kernel_syscall_via_break ()
#1 0x20000000003208e0 in raise () from /lib/libc.so.6.1
#2 0x2000000000324090 in abort () from /lib/libc.so.6.1
#3 0x200000000027c3e0 in __deallocate_stack () from /lib/libpthread.so.0
#4 0x200000000027f7c0 in start_thread () from /lib/libpthread.so.0
#5 0x200000000047ef60 in __clone2 () from /lib/libc.so.6.1
========================================================
The fact is, glibc call munmap() when thread exitng time for freeing
stack, and it assume munlock() never fail. However, munmap() often make
vma splitting and it with many mapcount make -ENOMEM.
Oh well, that's crazy, because stack unmapping never increase mapcount.
The maxcount exceeding is only temporary. internal temporary exceeding
shouldn't make ENOMEM.
This patch does it.
test_max_mapcount.c
==================================================================
#include<stdio.h>
#include<stdlib.h>
#include<string.h>
#include<pthread.h>
#include<errno.h>
#include<unistd.h>
#define THREAD_NUM 30000
#define MAL_SIZE (8*1024*1024)
void *wait_thread(void *args)
{
void *addr;
addr = malloc(MAL_SIZE);
sleep(10);
return NULL;
}
void *wait_thread2(void *args)
{
sleep(60);
return NULL;
}
int main(int argc, char *argv[])
{
int i;
pthread_t thread[THREAD_NUM], th;
int ret, count = 0;
pthread_attr_t attr;
ret = pthread_attr_init(&attr);
if(ret) {
perror("pthread_attr_init");
}
ret = pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED);
if(ret) {
perror("pthread_attr_setdetachstate");
}
for (i = 0; i < THREAD_NUM; i++) {
ret = pthread_create(&th, &attr, wait_thread, NULL);
if(ret) {
fprintf(stderr, "[%d] ", count);
perror("pthread_create");
} else {
printf("[%d] create OK.\n", count);
}
count++;
ret = pthread_create(&thread[i], &attr, wait_thread2, NULL);
if(ret) {
fprintf(stderr, "[%d] ", count);
perror("pthread_create");
} else {
printf("[%d] create OK.\n", count);
}
count++;
}
sleep(3600);
return 0;
}
==================================================================
[akpm@linux-foundation.org: coding-style fixes]
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 09:57:56 +08:00
|
|
|
int error;
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Make sure that map_count on return from munmap() will
|
|
|
|
|
* not exceed its limit; but let map_count go just above
|
|
|
|
|
* its limit temporarily, to help free resources as expected.
|
|
|
|
|
*/
|
|
|
|
|
if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
|
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
|
|
error = __split_vma(mm, vma, start, 0);
|
2005-04-17 06:20:36 +08:00
|
|
|
if (error)
|
|
|
|
|
return error;
|
2005-04-20 04:29:18 +08:00
|
|
|
prev = vma;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Does it split the last one? */
|
|
|
|
|
last = find_vma(mm, end);
|
|
|
|
|
if (last && end > last->vm_start) {
|
mmap: don't return ENOMEM when mapcount is temporarily exceeded in munmap()
On ia64, the following test program exit abnormally, because glibc thread
library called abort().
========================================================
(gdb) bt
#0 0xa000000000010620 in __kernel_syscall_via_break ()
#1 0x20000000003208e0 in raise () from /lib/libc.so.6.1
#2 0x2000000000324090 in abort () from /lib/libc.so.6.1
#3 0x200000000027c3e0 in __deallocate_stack () from /lib/libpthread.so.0
#4 0x200000000027f7c0 in start_thread () from /lib/libpthread.so.0
#5 0x200000000047ef60 in __clone2 () from /lib/libc.so.6.1
========================================================
The fact is, glibc call munmap() when thread exitng time for freeing
stack, and it assume munlock() never fail. However, munmap() often make
vma splitting and it with many mapcount make -ENOMEM.
Oh well, that's crazy, because stack unmapping never increase mapcount.
The maxcount exceeding is only temporary. internal temporary exceeding
shouldn't make ENOMEM.
This patch does it.
test_max_mapcount.c
==================================================================
#include<stdio.h>
#include<stdlib.h>
#include<string.h>
#include<pthread.h>
#include<errno.h>
#include<unistd.h>
#define THREAD_NUM 30000
#define MAL_SIZE (8*1024*1024)
void *wait_thread(void *args)
{
void *addr;
addr = malloc(MAL_SIZE);
sleep(10);
return NULL;
}
void *wait_thread2(void *args)
{
sleep(60);
return NULL;
}
int main(int argc, char *argv[])
{
int i;
pthread_t thread[THREAD_NUM], th;
int ret, count = 0;
pthread_attr_t attr;
ret = pthread_attr_init(&attr);
if(ret) {
perror("pthread_attr_init");
}
ret = pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED);
if(ret) {
perror("pthread_attr_setdetachstate");
}
for (i = 0; i < THREAD_NUM; i++) {
ret = pthread_create(&th, &attr, wait_thread, NULL);
if(ret) {
fprintf(stderr, "[%d] ", count);
perror("pthread_create");
} else {
printf("[%d] create OK.\n", count);
}
count++;
ret = pthread_create(&thread[i], &attr, wait_thread2, NULL);
if(ret) {
fprintf(stderr, "[%d] ", count);
perror("pthread_create");
} else {
printf("[%d] create OK.\n", count);
}
count++;
}
sleep(3600);
return 0;
}
==================================================================
[akpm@linux-foundation.org: coding-style fixes]
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 09:57:56 +08:00
|
|
|
int error = __split_vma(mm, last, end, 1);
|
2005-04-17 06:20:36 +08:00
|
|
|
if (error)
|
|
|
|
|
return error;
|
|
|
|
|
}
|
2020-10-18 07:14:06 +08:00
|
|
|
vma = vma_next(mm, prev);
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2017-09-07 07:23:53 +08:00
|
|
|
if (unlikely(uf)) {
|
|
|
|
|
/*
|
|
|
|
|
* If userfaultfd_unmap_prep returns an error the vmas
|
2021-05-07 09:06:47 +08:00
|
|
|
* will remain split, but userland will get a
|
2017-09-07 07:23:53 +08:00
|
|
|
* highly unexpected error anyway. This is no
|
|
|
|
|
* different than the case where the first of the two
|
|
|
|
|
* __split_vma fails, but we don't undo the first
|
|
|
|
|
* split, despite we could. This is unlikely enough
|
|
|
|
|
* failure that it's not worth optimizing it for.
|
|
|
|
|
*/
|
|
|
|
|
int error = userfaultfd_unmap_prep(vma, start, end, uf);
|
|
|
|
|
if (error)
|
|
|
|
|
return error;
|
|
|
|
|
}
|
|
|
|
|
|
mm: mmap: zap pages with read mmap_sem in munmap
Patch series "mm: zap pages with read mmap_sem in munmap for large
mapping", v11.
Background:
Recently, when we ran some vm scalability tests on machines with large memory,
we ran into a couple of mmap_sem scalability issues when unmapping large memory
space, please refer to https://lkml.org/lkml/2017/12/14/733 and
https://lkml.org/lkml/2018/2/20/576.
History:
Then akpm suggested to unmap large mapping section by section and drop mmap_sem
at a time to mitigate it (see https://lkml.org/lkml/2018/3/6/784).
V1 patch series was submitted to the mailing list per Andrew's suggestion
(see https://lkml.org/lkml/2018/3/20/786). Then I received a lot great
feedback and suggestions.
Then this topic was discussed on LSFMM summit 2018. In the summit, Michal
Hocko suggested (also in the v1 patches review) to try "two phases"
approach. Zapping pages with read mmap_sem, then doing via cleanup with
write mmap_sem (for discussion detail, see
https://lwn.net/Articles/753269/)
Approach:
Zapping pages is the most time consuming part, according to the suggestion from
Michal Hocko [1], zapping pages can be done with holding read mmap_sem, like
what MADV_DONTNEED does. Then re-acquire write mmap_sem to cleanup vmas.
But, we can't call MADV_DONTNEED directly, since there are two major drawbacks:
* The unexpected state from PF if it wins the race in the middle of munmap.
It may return zero page, instead of the content or SIGSEGV.
* Can't handle VM_LOCKED | VM_HUGETLB | VM_PFNMAP and uprobe mappings, which
is a showstopper from akpm
But, some part may need write mmap_sem, for example, vma splitting. So,
the design is as follows:
acquire write mmap_sem
lookup vmas (find and split vmas)
deal with special mappings
detach vmas
downgrade_write
zap pages
free page tables
release mmap_sem
The vm events with read mmap_sem may come in during page zapping, but
since vmas have been detached before, they, i.e. page fault, gup, etc,
will not be able to find valid vma, then just return SIGSEGV or -EFAULT as
expected.
If the vma has VM_HUGETLB | VM_PFNMAP, they are considered as special
mappings. They will be handled by falling back to regular do_munmap()
with exclusive mmap_sem held in this patch since they may update vm flags.
But, with the "detach vmas first" approach, the vmas have been detached
when vm flags are updated, so it sounds safe to update vm flags with read
mmap_sem for this specific case. So, VM_HUGETLB and VM_PFNMAP will be
handled by using the optimized path in the following separate patches for
bisectable sake.
Unmapping uprobe areas may need update mm flags (MMF_RECALC_UPROBES).
However it is fine to have false-positive MMF_RECALC_UPROBES according to
uprobes developer. So, uprobe unmap will not be handled by the regular
path.
With the "detach vmas first" approach we don't have to re-acquire mmap_sem
again to clean up vmas to avoid race window which might get the address
space changed since downgrade_write() doesn't release the lock to lead
regression, which simply downgrades to read lock.
And, since the lock acquire/release cost is managed to the minimum and
almost as same as before, the optimization could be extended to any size
of mapping without incurring significant penalty to small mappings.
For the time being, just do this in munmap syscall path. Other
vm_munmap() or do_munmap() call sites (i.e mmap, mremap, etc) remain
intact due to some implementation difficulties since they acquire write
mmap_sem from very beginning and hold it until the end, do_munmap() might
be called in the middle. But, the optimized do_munmap would like to be
called without mmap_sem held so that we can do the optimization. So, if
we want to do the similar optimization for mmap/mremap path, I'm afraid we
would have to redesign them. mremap might be called on very large area
depending on the usecases, the optimization to it will be considered in
the future.
This patch (of 3):
When running some mmap/munmap scalability tests with large memory (i.e.
> 300GB), the below hung task issue may happen occasionally.
INFO: task ps:14018 blocked for more than 120 seconds.
Tainted: G E 4.9.79-009.ali3000.alios7.x86_64 #1
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this
message.
ps D 0 14018 1 0x00000004
ffff885582f84000 ffff885e8682f000 ffff880972943000 ffff885ebf499bc0
ffff8828ee120000 ffffc900349bfca8 ffffffff817154d0 0000000000000040
00ffffff812f872a ffff885ebf499bc0 024000d000948300 ffff880972943000
Call Trace:
[<ffffffff817154d0>] ? __schedule+0x250/0x730
[<ffffffff817159e6>] schedule+0x36/0x80
[<ffffffff81718560>] rwsem_down_read_failed+0xf0/0x150
[<ffffffff81390a28>] call_rwsem_down_read_failed+0x18/0x30
[<ffffffff81717db0>] down_read+0x20/0x40
[<ffffffff812b9439>] proc_pid_cmdline_read+0xd9/0x4e0
[<ffffffff81253c95>] ? do_filp_open+0xa5/0x100
[<ffffffff81241d87>] __vfs_read+0x37/0x150
[<ffffffff812f824b>] ? security_file_permission+0x9b/0xc0
[<ffffffff81242266>] vfs_read+0x96/0x130
[<ffffffff812437b5>] SyS_read+0x55/0xc0
[<ffffffff8171a6da>] entry_SYSCALL_64_fastpath+0x1a/0xc5
It is because munmap holds mmap_sem exclusively from very beginning to all
the way down to the end, and doesn't release it in the middle. When
unmapping large mapping, it may take long time (take ~18 seconds to unmap
320GB mapping with every single page mapped on an idle machine).
Zapping pages is the most time consuming part, according to the suggestion
from Michal Hocko [1], zapping pages can be done with holding read
mmap_sem, like what MADV_DONTNEED does. Then re-acquire write mmap_sem to
cleanup vmas.
But, some part may need write mmap_sem, for example, vma splitting. So,
the design is as follows:
acquire write mmap_sem
lookup vmas (find and split vmas)
deal with special mappings
detach vmas
downgrade_write
zap pages
free page tables
release mmap_sem
The vm events with read mmap_sem may come in during page zapping, but
since vmas have been detached before, they, i.e. page fault, gup, etc,
will not be able to find valid vma, then just return SIGSEGV or -EFAULT as
expected.
If the vma has VM_HUGETLB | VM_PFNMAP, they are considered as special
mappings. They will be handled by without downgrading mmap_sem in this
patch since they may update vm flags.
But, with the "detach vmas first" approach, the vmas have been detached
when vm flags are updated, so it sounds safe to update vm flags with read
mmap_sem for this specific case. So, VM_HUGETLB and VM_PFNMAP will be
handled by using the optimized path in the following separate patches for
bisectable sake.
Unmapping uprobe areas may need update mm flags (MMF_RECALC_UPROBES).
However it is fine to have false-positive MMF_RECALC_UPROBES according to
uprobes developer.
With the "detach vmas first" approach we don't have to re-acquire mmap_sem
again to clean up vmas to avoid race window which might get the address
space changed since downgrade_write() doesn't release the lock to lead
regression, which simply downgrades to read lock.
And, since the lock acquire/release cost is managed to the minimum and
almost as same as before, the optimization could be extended to any size
of mapping without incurring significant penalty to small mappings.
For the time being, just do this in munmap syscall path. Other
vm_munmap() or do_munmap() call sites (i.e mmap, mremap, etc) remain
intact due to some implementation difficulties since they acquire write
mmap_sem from very beginning and hold it until the end, do_munmap() might
be called in the middle. But, the optimized do_munmap would like to be
called without mmap_sem held so that we can do the optimization. So, if
we want to do the similar optimization for mmap/mremap path, I'm afraid we
would have to redesign them. mremap might be called on very large area
depending on the usecases, the optimization to it will be considered in
the future.
With the patches, exclusive mmap_sem hold time when munmap a 80GB address
space on a machine with 32 cores of E5-2680 @ 2.70GHz dropped to us level
from second.
munmap_test-15002 [008] 594.380138: funcgraph_entry: |
__vm_munmap() {
munmap_test-15002 [008] 594.380146: funcgraph_entry: !2485684 us
| unmap_region();
munmap_test-15002 [008] 596.865836: funcgraph_exit: !2485692 us
| }
Here the execution time of unmap_region() is used to evaluate the time of
holding read mmap_sem, then the remaining time is used with holding
exclusive lock.
[1] https://lwn.net/Articles/753269/
Link: http://lkml.kernel.org/r/1537376621-51150-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>Suggested-by: Michal Hocko <mhocko@kernel.org>
Suggested-by: Kirill A. Shutemov <kirill@shutemov.name>
Suggested-by: Matthew Wilcox <willy@infradead.org>
Reviewed-by: Matthew Wilcox <willy@infradead.org>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Laurent Dufour <ldufour@linux.vnet.ibm.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-27 06:07:11 +08:00
|
|
|
/* Detach vmas from rbtree */
|
2020-07-24 12:15:11 +08:00
|
|
|
if (!detach_vmas_to_be_unmapped(mm, vma, prev, end))
|
|
|
|
|
downgrade = false;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
mm: mmap: zap pages with read mmap_sem in munmap
Patch series "mm: zap pages with read mmap_sem in munmap for large
mapping", v11.
Background:
Recently, when we ran some vm scalability tests on machines with large memory,
we ran into a couple of mmap_sem scalability issues when unmapping large memory
space, please refer to https://lkml.org/lkml/2017/12/14/733 and
https://lkml.org/lkml/2018/2/20/576.
History:
Then akpm suggested to unmap large mapping section by section and drop mmap_sem
at a time to mitigate it (see https://lkml.org/lkml/2018/3/6/784).
V1 patch series was submitted to the mailing list per Andrew's suggestion
(see https://lkml.org/lkml/2018/3/20/786). Then I received a lot great
feedback and suggestions.
Then this topic was discussed on LSFMM summit 2018. In the summit, Michal
Hocko suggested (also in the v1 patches review) to try "two phases"
approach. Zapping pages with read mmap_sem, then doing via cleanup with
write mmap_sem (for discussion detail, see
https://lwn.net/Articles/753269/)
Approach:
Zapping pages is the most time consuming part, according to the suggestion from
Michal Hocko [1], zapping pages can be done with holding read mmap_sem, like
what MADV_DONTNEED does. Then re-acquire write mmap_sem to cleanup vmas.
But, we can't call MADV_DONTNEED directly, since there are two major drawbacks:
* The unexpected state from PF if it wins the race in the middle of munmap.
It may return zero page, instead of the content or SIGSEGV.
* Can't handle VM_LOCKED | VM_HUGETLB | VM_PFNMAP and uprobe mappings, which
is a showstopper from akpm
But, some part may need write mmap_sem, for example, vma splitting. So,
the design is as follows:
acquire write mmap_sem
lookup vmas (find and split vmas)
deal with special mappings
detach vmas
downgrade_write
zap pages
free page tables
release mmap_sem
The vm events with read mmap_sem may come in during page zapping, but
since vmas have been detached before, they, i.e. page fault, gup, etc,
will not be able to find valid vma, then just return SIGSEGV or -EFAULT as
expected.
If the vma has VM_HUGETLB | VM_PFNMAP, they are considered as special
mappings. They will be handled by falling back to regular do_munmap()
with exclusive mmap_sem held in this patch since they may update vm flags.
But, with the "detach vmas first" approach, the vmas have been detached
when vm flags are updated, so it sounds safe to update vm flags with read
mmap_sem for this specific case. So, VM_HUGETLB and VM_PFNMAP will be
handled by using the optimized path in the following separate patches for
bisectable sake.
Unmapping uprobe areas may need update mm flags (MMF_RECALC_UPROBES).
However it is fine to have false-positive MMF_RECALC_UPROBES according to
uprobes developer. So, uprobe unmap will not be handled by the regular
path.
With the "detach vmas first" approach we don't have to re-acquire mmap_sem
again to clean up vmas to avoid race window which might get the address
space changed since downgrade_write() doesn't release the lock to lead
regression, which simply downgrades to read lock.
And, since the lock acquire/release cost is managed to the minimum and
almost as same as before, the optimization could be extended to any size
of mapping without incurring significant penalty to small mappings.
For the time being, just do this in munmap syscall path. Other
vm_munmap() or do_munmap() call sites (i.e mmap, mremap, etc) remain
intact due to some implementation difficulties since they acquire write
mmap_sem from very beginning and hold it until the end, do_munmap() might
be called in the middle. But, the optimized do_munmap would like to be
called without mmap_sem held so that we can do the optimization. So, if
we want to do the similar optimization for mmap/mremap path, I'm afraid we
would have to redesign them. mremap might be called on very large area
depending on the usecases, the optimization to it will be considered in
the future.
This patch (of 3):
When running some mmap/munmap scalability tests with large memory (i.e.
> 300GB), the below hung task issue may happen occasionally.
INFO: task ps:14018 blocked for more than 120 seconds.
Tainted: G E 4.9.79-009.ali3000.alios7.x86_64 #1
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this
message.
ps D 0 14018 1 0x00000004
ffff885582f84000 ffff885e8682f000 ffff880972943000 ffff885ebf499bc0
ffff8828ee120000 ffffc900349bfca8 ffffffff817154d0 0000000000000040
00ffffff812f872a ffff885ebf499bc0 024000d000948300 ffff880972943000
Call Trace:
[<ffffffff817154d0>] ? __schedule+0x250/0x730
[<ffffffff817159e6>] schedule+0x36/0x80
[<ffffffff81718560>] rwsem_down_read_failed+0xf0/0x150
[<ffffffff81390a28>] call_rwsem_down_read_failed+0x18/0x30
[<ffffffff81717db0>] down_read+0x20/0x40
[<ffffffff812b9439>] proc_pid_cmdline_read+0xd9/0x4e0
[<ffffffff81253c95>] ? do_filp_open+0xa5/0x100
[<ffffffff81241d87>] __vfs_read+0x37/0x150
[<ffffffff812f824b>] ? security_file_permission+0x9b/0xc0
[<ffffffff81242266>] vfs_read+0x96/0x130
[<ffffffff812437b5>] SyS_read+0x55/0xc0
[<ffffffff8171a6da>] entry_SYSCALL_64_fastpath+0x1a/0xc5
It is because munmap holds mmap_sem exclusively from very beginning to all
the way down to the end, and doesn't release it in the middle. When
unmapping large mapping, it may take long time (take ~18 seconds to unmap
320GB mapping with every single page mapped on an idle machine).
Zapping pages is the most time consuming part, according to the suggestion
from Michal Hocko [1], zapping pages can be done with holding read
mmap_sem, like what MADV_DONTNEED does. Then re-acquire write mmap_sem to
cleanup vmas.
But, some part may need write mmap_sem, for example, vma splitting. So,
the design is as follows:
acquire write mmap_sem
lookup vmas (find and split vmas)
deal with special mappings
detach vmas
downgrade_write
zap pages
free page tables
release mmap_sem
The vm events with read mmap_sem may come in during page zapping, but
since vmas have been detached before, they, i.e. page fault, gup, etc,
will not be able to find valid vma, then just return SIGSEGV or -EFAULT as
expected.
If the vma has VM_HUGETLB | VM_PFNMAP, they are considered as special
mappings. They will be handled by without downgrading mmap_sem in this
patch since they may update vm flags.
But, with the "detach vmas first" approach, the vmas have been detached
when vm flags are updated, so it sounds safe to update vm flags with read
mmap_sem for this specific case. So, VM_HUGETLB and VM_PFNMAP will be
handled by using the optimized path in the following separate patches for
bisectable sake.
Unmapping uprobe areas may need update mm flags (MMF_RECALC_UPROBES).
However it is fine to have false-positive MMF_RECALC_UPROBES according to
uprobes developer.
With the "detach vmas first" approach we don't have to re-acquire mmap_sem
again to clean up vmas to avoid race window which might get the address
space changed since downgrade_write() doesn't release the lock to lead
regression, which simply downgrades to read lock.
And, since the lock acquire/release cost is managed to the minimum and
almost as same as before, the optimization could be extended to any size
of mapping without incurring significant penalty to small mappings.
For the time being, just do this in munmap syscall path. Other
vm_munmap() or do_munmap() call sites (i.e mmap, mremap, etc) remain
intact due to some implementation difficulties since they acquire write
mmap_sem from very beginning and hold it until the end, do_munmap() might
be called in the middle. But, the optimized do_munmap would like to be
called without mmap_sem held so that we can do the optimization. So, if
we want to do the similar optimization for mmap/mremap path, I'm afraid we
would have to redesign them. mremap might be called on very large area
depending on the usecases, the optimization to it will be considered in
the future.
With the patches, exclusive mmap_sem hold time when munmap a 80GB address
space on a machine with 32 cores of E5-2680 @ 2.70GHz dropped to us level
from second.
munmap_test-15002 [008] 594.380138: funcgraph_entry: |
__vm_munmap() {
munmap_test-15002 [008] 594.380146: funcgraph_entry: !2485684 us
| unmap_region();
munmap_test-15002 [008] 596.865836: funcgraph_exit: !2485692 us
| }
Here the execution time of unmap_region() is used to evaluate the time of
holding read mmap_sem, then the remaining time is used with holding
exclusive lock.
[1] https://lwn.net/Articles/753269/
Link: http://lkml.kernel.org/r/1537376621-51150-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>Suggested-by: Michal Hocko <mhocko@kernel.org>
Suggested-by: Kirill A. Shutemov <kirill@shutemov.name>
Suggested-by: Matthew Wilcox <willy@infradead.org>
Reviewed-by: Matthew Wilcox <willy@infradead.org>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Laurent Dufour <ldufour@linux.vnet.ibm.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-27 06:07:11 +08:00
|
|
|
if (downgrade)
|
2020-06-09 12:33:25 +08:00
|
|
|
mmap_write_downgrade(mm);
|
mm: mmap: zap pages with read mmap_sem in munmap
Patch series "mm: zap pages with read mmap_sem in munmap for large
mapping", v11.
Background:
Recently, when we ran some vm scalability tests on machines with large memory,
we ran into a couple of mmap_sem scalability issues when unmapping large memory
space, please refer to https://lkml.org/lkml/2017/12/14/733 and
https://lkml.org/lkml/2018/2/20/576.
History:
Then akpm suggested to unmap large mapping section by section and drop mmap_sem
at a time to mitigate it (see https://lkml.org/lkml/2018/3/6/784).
V1 patch series was submitted to the mailing list per Andrew's suggestion
(see https://lkml.org/lkml/2018/3/20/786). Then I received a lot great
feedback and suggestions.
Then this topic was discussed on LSFMM summit 2018. In the summit, Michal
Hocko suggested (also in the v1 patches review) to try "two phases"
approach. Zapping pages with read mmap_sem, then doing via cleanup with
write mmap_sem (for discussion detail, see
https://lwn.net/Articles/753269/)
Approach:
Zapping pages is the most time consuming part, according to the suggestion from
Michal Hocko [1], zapping pages can be done with holding read mmap_sem, like
what MADV_DONTNEED does. Then re-acquire write mmap_sem to cleanup vmas.
But, we can't call MADV_DONTNEED directly, since there are two major drawbacks:
* The unexpected state from PF if it wins the race in the middle of munmap.
It may return zero page, instead of the content or SIGSEGV.
* Can't handle VM_LOCKED | VM_HUGETLB | VM_PFNMAP and uprobe mappings, which
is a showstopper from akpm
But, some part may need write mmap_sem, for example, vma splitting. So,
the design is as follows:
acquire write mmap_sem
lookup vmas (find and split vmas)
deal with special mappings
detach vmas
downgrade_write
zap pages
free page tables
release mmap_sem
The vm events with read mmap_sem may come in during page zapping, but
since vmas have been detached before, they, i.e. page fault, gup, etc,
will not be able to find valid vma, then just return SIGSEGV or -EFAULT as
expected.
If the vma has VM_HUGETLB | VM_PFNMAP, they are considered as special
mappings. They will be handled by falling back to regular do_munmap()
with exclusive mmap_sem held in this patch since they may update vm flags.
But, with the "detach vmas first" approach, the vmas have been detached
when vm flags are updated, so it sounds safe to update vm flags with read
mmap_sem for this specific case. So, VM_HUGETLB and VM_PFNMAP will be
handled by using the optimized path in the following separate patches for
bisectable sake.
Unmapping uprobe areas may need update mm flags (MMF_RECALC_UPROBES).
However it is fine to have false-positive MMF_RECALC_UPROBES according to
uprobes developer. So, uprobe unmap will not be handled by the regular
path.
With the "detach vmas first" approach we don't have to re-acquire mmap_sem
again to clean up vmas to avoid race window which might get the address
space changed since downgrade_write() doesn't release the lock to lead
regression, which simply downgrades to read lock.
And, since the lock acquire/release cost is managed to the minimum and
almost as same as before, the optimization could be extended to any size
of mapping without incurring significant penalty to small mappings.
For the time being, just do this in munmap syscall path. Other
vm_munmap() or do_munmap() call sites (i.e mmap, mremap, etc) remain
intact due to some implementation difficulties since they acquire write
mmap_sem from very beginning and hold it until the end, do_munmap() might
be called in the middle. But, the optimized do_munmap would like to be
called without mmap_sem held so that we can do the optimization. So, if
we want to do the similar optimization for mmap/mremap path, I'm afraid we
would have to redesign them. mremap might be called on very large area
depending on the usecases, the optimization to it will be considered in
the future.
This patch (of 3):
When running some mmap/munmap scalability tests with large memory (i.e.
> 300GB), the below hung task issue may happen occasionally.
INFO: task ps:14018 blocked for more than 120 seconds.
Tainted: G E 4.9.79-009.ali3000.alios7.x86_64 #1
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this
message.
ps D 0 14018 1 0x00000004
ffff885582f84000 ffff885e8682f000 ffff880972943000 ffff885ebf499bc0
ffff8828ee120000 ffffc900349bfca8 ffffffff817154d0 0000000000000040
00ffffff812f872a ffff885ebf499bc0 024000d000948300 ffff880972943000
Call Trace:
[<ffffffff817154d0>] ? __schedule+0x250/0x730
[<ffffffff817159e6>] schedule+0x36/0x80
[<ffffffff81718560>] rwsem_down_read_failed+0xf0/0x150
[<ffffffff81390a28>] call_rwsem_down_read_failed+0x18/0x30
[<ffffffff81717db0>] down_read+0x20/0x40
[<ffffffff812b9439>] proc_pid_cmdline_read+0xd9/0x4e0
[<ffffffff81253c95>] ? do_filp_open+0xa5/0x100
[<ffffffff81241d87>] __vfs_read+0x37/0x150
[<ffffffff812f824b>] ? security_file_permission+0x9b/0xc0
[<ffffffff81242266>] vfs_read+0x96/0x130
[<ffffffff812437b5>] SyS_read+0x55/0xc0
[<ffffffff8171a6da>] entry_SYSCALL_64_fastpath+0x1a/0xc5
It is because munmap holds mmap_sem exclusively from very beginning to all
the way down to the end, and doesn't release it in the middle. When
unmapping large mapping, it may take long time (take ~18 seconds to unmap
320GB mapping with every single page mapped on an idle machine).
Zapping pages is the most time consuming part, according to the suggestion
from Michal Hocko [1], zapping pages can be done with holding read
mmap_sem, like what MADV_DONTNEED does. Then re-acquire write mmap_sem to
cleanup vmas.
But, some part may need write mmap_sem, for example, vma splitting. So,
the design is as follows:
acquire write mmap_sem
lookup vmas (find and split vmas)
deal with special mappings
detach vmas
downgrade_write
zap pages
free page tables
release mmap_sem
The vm events with read mmap_sem may come in during page zapping, but
since vmas have been detached before, they, i.e. page fault, gup, etc,
will not be able to find valid vma, then just return SIGSEGV or -EFAULT as
expected.
If the vma has VM_HUGETLB | VM_PFNMAP, they are considered as special
mappings. They will be handled by without downgrading mmap_sem in this
patch since they may update vm flags.
But, with the "detach vmas first" approach, the vmas have been detached
when vm flags are updated, so it sounds safe to update vm flags with read
mmap_sem for this specific case. So, VM_HUGETLB and VM_PFNMAP will be
handled by using the optimized path in the following separate patches for
bisectable sake.
Unmapping uprobe areas may need update mm flags (MMF_RECALC_UPROBES).
However it is fine to have false-positive MMF_RECALC_UPROBES according to
uprobes developer.
With the "detach vmas first" approach we don't have to re-acquire mmap_sem
again to clean up vmas to avoid race window which might get the address
space changed since downgrade_write() doesn't release the lock to lead
regression, which simply downgrades to read lock.
And, since the lock acquire/release cost is managed to the minimum and
almost as same as before, the optimization could be extended to any size
of mapping without incurring significant penalty to small mappings.
For the time being, just do this in munmap syscall path. Other
vm_munmap() or do_munmap() call sites (i.e mmap, mremap, etc) remain
intact due to some implementation difficulties since they acquire write
mmap_sem from very beginning and hold it until the end, do_munmap() might
be called in the middle. But, the optimized do_munmap would like to be
called without mmap_sem held so that we can do the optimization. So, if
we want to do the similar optimization for mmap/mremap path, I'm afraid we
would have to redesign them. mremap might be called on very large area
depending on the usecases, the optimization to it will be considered in
the future.
With the patches, exclusive mmap_sem hold time when munmap a 80GB address
space on a machine with 32 cores of E5-2680 @ 2.70GHz dropped to us level
from second.
munmap_test-15002 [008] 594.380138: funcgraph_entry: |
__vm_munmap() {
munmap_test-15002 [008] 594.380146: funcgraph_entry: !2485684 us
| unmap_region();
munmap_test-15002 [008] 596.865836: funcgraph_exit: !2485692 us
| }
Here the execution time of unmap_region() is used to evaluate the time of
holding read mmap_sem, then the remaining time is used with holding
exclusive lock.
[1] https://lwn.net/Articles/753269/
Link: http://lkml.kernel.org/r/1537376621-51150-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>Suggested-by: Michal Hocko <mhocko@kernel.org>
Suggested-by: Kirill A. Shutemov <kirill@shutemov.name>
Suggested-by: Matthew Wilcox <willy@infradead.org>
Reviewed-by: Matthew Wilcox <willy@infradead.org>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Laurent Dufour <ldufour@linux.vnet.ibm.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-27 06:07:11 +08:00
|
|
|
|
|
|
|
|
unmap_region(mm, vma, prev, start, end);
|
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
/* Fix up all other VM information */
|
2005-10-30 09:15:56 +08:00
|
|
|
remove_vma_list(mm, vma);
|
2005-04-17 06:20:36 +08:00
|
|
|
|
mm: mmap: zap pages with read mmap_sem in munmap
Patch series "mm: zap pages with read mmap_sem in munmap for large
mapping", v11.
Background:
Recently, when we ran some vm scalability tests on machines with large memory,
we ran into a couple of mmap_sem scalability issues when unmapping large memory
space, please refer to https://lkml.org/lkml/2017/12/14/733 and
https://lkml.org/lkml/2018/2/20/576.
History:
Then akpm suggested to unmap large mapping section by section and drop mmap_sem
at a time to mitigate it (see https://lkml.org/lkml/2018/3/6/784).
V1 patch series was submitted to the mailing list per Andrew's suggestion
(see https://lkml.org/lkml/2018/3/20/786). Then I received a lot great
feedback and suggestions.
Then this topic was discussed on LSFMM summit 2018. In the summit, Michal
Hocko suggested (also in the v1 patches review) to try "two phases"
approach. Zapping pages with read mmap_sem, then doing via cleanup with
write mmap_sem (for discussion detail, see
https://lwn.net/Articles/753269/)
Approach:
Zapping pages is the most time consuming part, according to the suggestion from
Michal Hocko [1], zapping pages can be done with holding read mmap_sem, like
what MADV_DONTNEED does. Then re-acquire write mmap_sem to cleanup vmas.
But, we can't call MADV_DONTNEED directly, since there are two major drawbacks:
* The unexpected state from PF if it wins the race in the middle of munmap.
It may return zero page, instead of the content or SIGSEGV.
* Can't handle VM_LOCKED | VM_HUGETLB | VM_PFNMAP and uprobe mappings, which
is a showstopper from akpm
But, some part may need write mmap_sem, for example, vma splitting. So,
the design is as follows:
acquire write mmap_sem
lookup vmas (find and split vmas)
deal with special mappings
detach vmas
downgrade_write
zap pages
free page tables
release mmap_sem
The vm events with read mmap_sem may come in during page zapping, but
since vmas have been detached before, they, i.e. page fault, gup, etc,
will not be able to find valid vma, then just return SIGSEGV or -EFAULT as
expected.
If the vma has VM_HUGETLB | VM_PFNMAP, they are considered as special
mappings. They will be handled by falling back to regular do_munmap()
with exclusive mmap_sem held in this patch since they may update vm flags.
But, with the "detach vmas first" approach, the vmas have been detached
when vm flags are updated, so it sounds safe to update vm flags with read
mmap_sem for this specific case. So, VM_HUGETLB and VM_PFNMAP will be
handled by using the optimized path in the following separate patches for
bisectable sake.
Unmapping uprobe areas may need update mm flags (MMF_RECALC_UPROBES).
However it is fine to have false-positive MMF_RECALC_UPROBES according to
uprobes developer. So, uprobe unmap will not be handled by the regular
path.
With the "detach vmas first" approach we don't have to re-acquire mmap_sem
again to clean up vmas to avoid race window which might get the address
space changed since downgrade_write() doesn't release the lock to lead
regression, which simply downgrades to read lock.
And, since the lock acquire/release cost is managed to the minimum and
almost as same as before, the optimization could be extended to any size
of mapping without incurring significant penalty to small mappings.
For the time being, just do this in munmap syscall path. Other
vm_munmap() or do_munmap() call sites (i.e mmap, mremap, etc) remain
intact due to some implementation difficulties since they acquire write
mmap_sem from very beginning and hold it until the end, do_munmap() might
be called in the middle. But, the optimized do_munmap would like to be
called without mmap_sem held so that we can do the optimization. So, if
we want to do the similar optimization for mmap/mremap path, I'm afraid we
would have to redesign them. mremap might be called on very large area
depending on the usecases, the optimization to it will be considered in
the future.
This patch (of 3):
When running some mmap/munmap scalability tests with large memory (i.e.
> 300GB), the below hung task issue may happen occasionally.
INFO: task ps:14018 blocked for more than 120 seconds.
Tainted: G E 4.9.79-009.ali3000.alios7.x86_64 #1
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this
message.
ps D 0 14018 1 0x00000004
ffff885582f84000 ffff885e8682f000 ffff880972943000 ffff885ebf499bc0
ffff8828ee120000 ffffc900349bfca8 ffffffff817154d0 0000000000000040
00ffffff812f872a ffff885ebf499bc0 024000d000948300 ffff880972943000
Call Trace:
[<ffffffff817154d0>] ? __schedule+0x250/0x730
[<ffffffff817159e6>] schedule+0x36/0x80
[<ffffffff81718560>] rwsem_down_read_failed+0xf0/0x150
[<ffffffff81390a28>] call_rwsem_down_read_failed+0x18/0x30
[<ffffffff81717db0>] down_read+0x20/0x40
[<ffffffff812b9439>] proc_pid_cmdline_read+0xd9/0x4e0
[<ffffffff81253c95>] ? do_filp_open+0xa5/0x100
[<ffffffff81241d87>] __vfs_read+0x37/0x150
[<ffffffff812f824b>] ? security_file_permission+0x9b/0xc0
[<ffffffff81242266>] vfs_read+0x96/0x130
[<ffffffff812437b5>] SyS_read+0x55/0xc0
[<ffffffff8171a6da>] entry_SYSCALL_64_fastpath+0x1a/0xc5
It is because munmap holds mmap_sem exclusively from very beginning to all
the way down to the end, and doesn't release it in the middle. When
unmapping large mapping, it may take long time (take ~18 seconds to unmap
320GB mapping with every single page mapped on an idle machine).
Zapping pages is the most time consuming part, according to the suggestion
from Michal Hocko [1], zapping pages can be done with holding read
mmap_sem, like what MADV_DONTNEED does. Then re-acquire write mmap_sem to
cleanup vmas.
But, some part may need write mmap_sem, for example, vma splitting. So,
the design is as follows:
acquire write mmap_sem
lookup vmas (find and split vmas)
deal with special mappings
detach vmas
downgrade_write
zap pages
free page tables
release mmap_sem
The vm events with read mmap_sem may come in during page zapping, but
since vmas have been detached before, they, i.e. page fault, gup, etc,
will not be able to find valid vma, then just return SIGSEGV or -EFAULT as
expected.
If the vma has VM_HUGETLB | VM_PFNMAP, they are considered as special
mappings. They will be handled by without downgrading mmap_sem in this
patch since they may update vm flags.
But, with the "detach vmas first" approach, the vmas have been detached
when vm flags are updated, so it sounds safe to update vm flags with read
mmap_sem for this specific case. So, VM_HUGETLB and VM_PFNMAP will be
handled by using the optimized path in the following separate patches for
bisectable sake.
Unmapping uprobe areas may need update mm flags (MMF_RECALC_UPROBES).
However it is fine to have false-positive MMF_RECALC_UPROBES according to
uprobes developer.
With the "detach vmas first" approach we don't have to re-acquire mmap_sem
again to clean up vmas to avoid race window which might get the address
space changed since downgrade_write() doesn't release the lock to lead
regression, which simply downgrades to read lock.
And, since the lock acquire/release cost is managed to the minimum and
almost as same as before, the optimization could be extended to any size
of mapping without incurring significant penalty to small mappings.
For the time being, just do this in munmap syscall path. Other
vm_munmap() or do_munmap() call sites (i.e mmap, mremap, etc) remain
intact due to some implementation difficulties since they acquire write
mmap_sem from very beginning and hold it until the end, do_munmap() might
be called in the middle. But, the optimized do_munmap would like to be
called without mmap_sem held so that we can do the optimization. So, if
we want to do the similar optimization for mmap/mremap path, I'm afraid we
would have to redesign them. mremap might be called on very large area
depending on the usecases, the optimization to it will be considered in
the future.
With the patches, exclusive mmap_sem hold time when munmap a 80GB address
space on a machine with 32 cores of E5-2680 @ 2.70GHz dropped to us level
from second.
munmap_test-15002 [008] 594.380138: funcgraph_entry: |
__vm_munmap() {
munmap_test-15002 [008] 594.380146: funcgraph_entry: !2485684 us
| unmap_region();
munmap_test-15002 [008] 596.865836: funcgraph_exit: !2485692 us
| }
Here the execution time of unmap_region() is used to evaluate the time of
holding read mmap_sem, then the remaining time is used with holding
exclusive lock.
[1] https://lwn.net/Articles/753269/
Link: http://lkml.kernel.org/r/1537376621-51150-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>Suggested-by: Michal Hocko <mhocko@kernel.org>
Suggested-by: Kirill A. Shutemov <kirill@shutemov.name>
Suggested-by: Matthew Wilcox <willy@infradead.org>
Reviewed-by: Matthew Wilcox <willy@infradead.org>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Laurent Dufour <ldufour@linux.vnet.ibm.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-27 06:07:11 +08:00
|
|
|
return downgrade ? 1 : 0;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
|
mm: mmap: zap pages with read mmap_sem in munmap
Patch series "mm: zap pages with read mmap_sem in munmap for large
mapping", v11.
Background:
Recently, when we ran some vm scalability tests on machines with large memory,
we ran into a couple of mmap_sem scalability issues when unmapping large memory
space, please refer to https://lkml.org/lkml/2017/12/14/733 and
https://lkml.org/lkml/2018/2/20/576.
History:
Then akpm suggested to unmap large mapping section by section and drop mmap_sem
at a time to mitigate it (see https://lkml.org/lkml/2018/3/6/784).
V1 patch series was submitted to the mailing list per Andrew's suggestion
(see https://lkml.org/lkml/2018/3/20/786). Then I received a lot great
feedback and suggestions.
Then this topic was discussed on LSFMM summit 2018. In the summit, Michal
Hocko suggested (also in the v1 patches review) to try "two phases"
approach. Zapping pages with read mmap_sem, then doing via cleanup with
write mmap_sem (for discussion detail, see
https://lwn.net/Articles/753269/)
Approach:
Zapping pages is the most time consuming part, according to the suggestion from
Michal Hocko [1], zapping pages can be done with holding read mmap_sem, like
what MADV_DONTNEED does. Then re-acquire write mmap_sem to cleanup vmas.
But, we can't call MADV_DONTNEED directly, since there are two major drawbacks:
* The unexpected state from PF if it wins the race in the middle of munmap.
It may return zero page, instead of the content or SIGSEGV.
* Can't handle VM_LOCKED | VM_HUGETLB | VM_PFNMAP and uprobe mappings, which
is a showstopper from akpm
But, some part may need write mmap_sem, for example, vma splitting. So,
the design is as follows:
acquire write mmap_sem
lookup vmas (find and split vmas)
deal with special mappings
detach vmas
downgrade_write
zap pages
free page tables
release mmap_sem
The vm events with read mmap_sem may come in during page zapping, but
since vmas have been detached before, they, i.e. page fault, gup, etc,
will not be able to find valid vma, then just return SIGSEGV or -EFAULT as
expected.
If the vma has VM_HUGETLB | VM_PFNMAP, they are considered as special
mappings. They will be handled by falling back to regular do_munmap()
with exclusive mmap_sem held in this patch since they may update vm flags.
But, with the "detach vmas first" approach, the vmas have been detached
when vm flags are updated, so it sounds safe to update vm flags with read
mmap_sem for this specific case. So, VM_HUGETLB and VM_PFNMAP will be
handled by using the optimized path in the following separate patches for
bisectable sake.
Unmapping uprobe areas may need update mm flags (MMF_RECALC_UPROBES).
However it is fine to have false-positive MMF_RECALC_UPROBES according to
uprobes developer. So, uprobe unmap will not be handled by the regular
path.
With the "detach vmas first" approach we don't have to re-acquire mmap_sem
again to clean up vmas to avoid race window which might get the address
space changed since downgrade_write() doesn't release the lock to lead
regression, which simply downgrades to read lock.
And, since the lock acquire/release cost is managed to the minimum and
almost as same as before, the optimization could be extended to any size
of mapping without incurring significant penalty to small mappings.
For the time being, just do this in munmap syscall path. Other
vm_munmap() or do_munmap() call sites (i.e mmap, mremap, etc) remain
intact due to some implementation difficulties since they acquire write
mmap_sem from very beginning and hold it until the end, do_munmap() might
be called in the middle. But, the optimized do_munmap would like to be
called without mmap_sem held so that we can do the optimization. So, if
we want to do the similar optimization for mmap/mremap path, I'm afraid we
would have to redesign them. mremap might be called on very large area
depending on the usecases, the optimization to it will be considered in
the future.
This patch (of 3):
When running some mmap/munmap scalability tests with large memory (i.e.
> 300GB), the below hung task issue may happen occasionally.
INFO: task ps:14018 blocked for more than 120 seconds.
Tainted: G E 4.9.79-009.ali3000.alios7.x86_64 #1
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this
message.
ps D 0 14018 1 0x00000004
ffff885582f84000 ffff885e8682f000 ffff880972943000 ffff885ebf499bc0
ffff8828ee120000 ffffc900349bfca8 ffffffff817154d0 0000000000000040
00ffffff812f872a ffff885ebf499bc0 024000d000948300 ffff880972943000
Call Trace:
[<ffffffff817154d0>] ? __schedule+0x250/0x730
[<ffffffff817159e6>] schedule+0x36/0x80
[<ffffffff81718560>] rwsem_down_read_failed+0xf0/0x150
[<ffffffff81390a28>] call_rwsem_down_read_failed+0x18/0x30
[<ffffffff81717db0>] down_read+0x20/0x40
[<ffffffff812b9439>] proc_pid_cmdline_read+0xd9/0x4e0
[<ffffffff81253c95>] ? do_filp_open+0xa5/0x100
[<ffffffff81241d87>] __vfs_read+0x37/0x150
[<ffffffff812f824b>] ? security_file_permission+0x9b/0xc0
[<ffffffff81242266>] vfs_read+0x96/0x130
[<ffffffff812437b5>] SyS_read+0x55/0xc0
[<ffffffff8171a6da>] entry_SYSCALL_64_fastpath+0x1a/0xc5
It is because munmap holds mmap_sem exclusively from very beginning to all
the way down to the end, and doesn't release it in the middle. When
unmapping large mapping, it may take long time (take ~18 seconds to unmap
320GB mapping with every single page mapped on an idle machine).
Zapping pages is the most time consuming part, according to the suggestion
from Michal Hocko [1], zapping pages can be done with holding read
mmap_sem, like what MADV_DONTNEED does. Then re-acquire write mmap_sem to
cleanup vmas.
But, some part may need write mmap_sem, for example, vma splitting. So,
the design is as follows:
acquire write mmap_sem
lookup vmas (find and split vmas)
deal with special mappings
detach vmas
downgrade_write
zap pages
free page tables
release mmap_sem
The vm events with read mmap_sem may come in during page zapping, but
since vmas have been detached before, they, i.e. page fault, gup, etc,
will not be able to find valid vma, then just return SIGSEGV or -EFAULT as
expected.
If the vma has VM_HUGETLB | VM_PFNMAP, they are considered as special
mappings. They will be handled by without downgrading mmap_sem in this
patch since they may update vm flags.
But, with the "detach vmas first" approach, the vmas have been detached
when vm flags are updated, so it sounds safe to update vm flags with read
mmap_sem for this specific case. So, VM_HUGETLB and VM_PFNMAP will be
handled by using the optimized path in the following separate patches for
bisectable sake.
Unmapping uprobe areas may need update mm flags (MMF_RECALC_UPROBES).
However it is fine to have false-positive MMF_RECALC_UPROBES according to
uprobes developer.
With the "detach vmas first" approach we don't have to re-acquire mmap_sem
again to clean up vmas to avoid race window which might get the address
space changed since downgrade_write() doesn't release the lock to lead
regression, which simply downgrades to read lock.
And, since the lock acquire/release cost is managed to the minimum and
almost as same as before, the optimization could be extended to any size
of mapping without incurring significant penalty to small mappings.
For the time being, just do this in munmap syscall path. Other
vm_munmap() or do_munmap() call sites (i.e mmap, mremap, etc) remain
intact due to some implementation difficulties since they acquire write
mmap_sem from very beginning and hold it until the end, do_munmap() might
be called in the middle. But, the optimized do_munmap would like to be
called without mmap_sem held so that we can do the optimization. So, if
we want to do the similar optimization for mmap/mremap path, I'm afraid we
would have to redesign them. mremap might be called on very large area
depending on the usecases, the optimization to it will be considered in
the future.
With the patches, exclusive mmap_sem hold time when munmap a 80GB address
space on a machine with 32 cores of E5-2680 @ 2.70GHz dropped to us level
from second.
munmap_test-15002 [008] 594.380138: funcgraph_entry: |
__vm_munmap() {
munmap_test-15002 [008] 594.380146: funcgraph_entry: !2485684 us
| unmap_region();
munmap_test-15002 [008] 596.865836: funcgraph_exit: !2485692 us
| }
Here the execution time of unmap_region() is used to evaluate the time of
holding read mmap_sem, then the remaining time is used with holding
exclusive lock.
[1] https://lwn.net/Articles/753269/
Link: http://lkml.kernel.org/r/1537376621-51150-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>Suggested-by: Michal Hocko <mhocko@kernel.org>
Suggested-by: Kirill A. Shutemov <kirill@shutemov.name>
Suggested-by: Matthew Wilcox <willy@infradead.org>
Reviewed-by: Matthew Wilcox <willy@infradead.org>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Laurent Dufour <ldufour@linux.vnet.ibm.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-27 06:07:11 +08:00
|
|
|
int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
|
|
|
|
|
struct list_head *uf)
|
|
|
|
|
{
|
|
|
|
|
return __do_munmap(mm, start, len, uf, false);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static int __vm_munmap(unsigned long start, size_t len, bool downgrade)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
|
|
|
|
int ret;
|
2012-04-21 09:57:04 +08:00
|
|
|
struct mm_struct *mm = current->mm;
|
2017-02-25 06:58:22 +08:00
|
|
|
LIST_HEAD(uf);
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2020-06-09 12:33:25 +08:00
|
|
|
if (mmap_write_lock_killable(mm))
|
2016-05-24 07:25:33 +08:00
|
|
|
return -EINTR;
|
|
|
|
|
|
mm: mmap: zap pages with read mmap_sem in munmap
Patch series "mm: zap pages with read mmap_sem in munmap for large
mapping", v11.
Background:
Recently, when we ran some vm scalability tests on machines with large memory,
we ran into a couple of mmap_sem scalability issues when unmapping large memory
space, please refer to https://lkml.org/lkml/2017/12/14/733 and
https://lkml.org/lkml/2018/2/20/576.
History:
Then akpm suggested to unmap large mapping section by section and drop mmap_sem
at a time to mitigate it (see https://lkml.org/lkml/2018/3/6/784).
V1 patch series was submitted to the mailing list per Andrew's suggestion
(see https://lkml.org/lkml/2018/3/20/786). Then I received a lot great
feedback and suggestions.
Then this topic was discussed on LSFMM summit 2018. In the summit, Michal
Hocko suggested (also in the v1 patches review) to try "two phases"
approach. Zapping pages with read mmap_sem, then doing via cleanup with
write mmap_sem (for discussion detail, see
https://lwn.net/Articles/753269/)
Approach:
Zapping pages is the most time consuming part, according to the suggestion from
Michal Hocko [1], zapping pages can be done with holding read mmap_sem, like
what MADV_DONTNEED does. Then re-acquire write mmap_sem to cleanup vmas.
But, we can't call MADV_DONTNEED directly, since there are two major drawbacks:
* The unexpected state from PF if it wins the race in the middle of munmap.
It may return zero page, instead of the content or SIGSEGV.
* Can't handle VM_LOCKED | VM_HUGETLB | VM_PFNMAP and uprobe mappings, which
is a showstopper from akpm
But, some part may need write mmap_sem, for example, vma splitting. So,
the design is as follows:
acquire write mmap_sem
lookup vmas (find and split vmas)
deal with special mappings
detach vmas
downgrade_write
zap pages
free page tables
release mmap_sem
The vm events with read mmap_sem may come in during page zapping, but
since vmas have been detached before, they, i.e. page fault, gup, etc,
will not be able to find valid vma, then just return SIGSEGV or -EFAULT as
expected.
If the vma has VM_HUGETLB | VM_PFNMAP, they are considered as special
mappings. They will be handled by falling back to regular do_munmap()
with exclusive mmap_sem held in this patch since they may update vm flags.
But, with the "detach vmas first" approach, the vmas have been detached
when vm flags are updated, so it sounds safe to update vm flags with read
mmap_sem for this specific case. So, VM_HUGETLB and VM_PFNMAP will be
handled by using the optimized path in the following separate patches for
bisectable sake.
Unmapping uprobe areas may need update mm flags (MMF_RECALC_UPROBES).
However it is fine to have false-positive MMF_RECALC_UPROBES according to
uprobes developer. So, uprobe unmap will not be handled by the regular
path.
With the "detach vmas first" approach we don't have to re-acquire mmap_sem
again to clean up vmas to avoid race window which might get the address
space changed since downgrade_write() doesn't release the lock to lead
regression, which simply downgrades to read lock.
And, since the lock acquire/release cost is managed to the minimum and
almost as same as before, the optimization could be extended to any size
of mapping without incurring significant penalty to small mappings.
For the time being, just do this in munmap syscall path. Other
vm_munmap() or do_munmap() call sites (i.e mmap, mremap, etc) remain
intact due to some implementation difficulties since they acquire write
mmap_sem from very beginning and hold it until the end, do_munmap() might
be called in the middle. But, the optimized do_munmap would like to be
called without mmap_sem held so that we can do the optimization. So, if
we want to do the similar optimization for mmap/mremap path, I'm afraid we
would have to redesign them. mremap might be called on very large area
depending on the usecases, the optimization to it will be considered in
the future.
This patch (of 3):
When running some mmap/munmap scalability tests with large memory (i.e.
> 300GB), the below hung task issue may happen occasionally.
INFO: task ps:14018 blocked for more than 120 seconds.
Tainted: G E 4.9.79-009.ali3000.alios7.x86_64 #1
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this
message.
ps D 0 14018 1 0x00000004
ffff885582f84000 ffff885e8682f000 ffff880972943000 ffff885ebf499bc0
ffff8828ee120000 ffffc900349bfca8 ffffffff817154d0 0000000000000040
00ffffff812f872a ffff885ebf499bc0 024000d000948300 ffff880972943000
Call Trace:
[<ffffffff817154d0>] ? __schedule+0x250/0x730
[<ffffffff817159e6>] schedule+0x36/0x80
[<ffffffff81718560>] rwsem_down_read_failed+0xf0/0x150
[<ffffffff81390a28>] call_rwsem_down_read_failed+0x18/0x30
[<ffffffff81717db0>] down_read+0x20/0x40
[<ffffffff812b9439>] proc_pid_cmdline_read+0xd9/0x4e0
[<ffffffff81253c95>] ? do_filp_open+0xa5/0x100
[<ffffffff81241d87>] __vfs_read+0x37/0x150
[<ffffffff812f824b>] ? security_file_permission+0x9b/0xc0
[<ffffffff81242266>] vfs_read+0x96/0x130
[<ffffffff812437b5>] SyS_read+0x55/0xc0
[<ffffffff8171a6da>] entry_SYSCALL_64_fastpath+0x1a/0xc5
It is because munmap holds mmap_sem exclusively from very beginning to all
the way down to the end, and doesn't release it in the middle. When
unmapping large mapping, it may take long time (take ~18 seconds to unmap
320GB mapping with every single page mapped on an idle machine).
Zapping pages is the most time consuming part, according to the suggestion
from Michal Hocko [1], zapping pages can be done with holding read
mmap_sem, like what MADV_DONTNEED does. Then re-acquire write mmap_sem to
cleanup vmas.
But, some part may need write mmap_sem, for example, vma splitting. So,
the design is as follows:
acquire write mmap_sem
lookup vmas (find and split vmas)
deal with special mappings
detach vmas
downgrade_write
zap pages
free page tables
release mmap_sem
The vm events with read mmap_sem may come in during page zapping, but
since vmas have been detached before, they, i.e. page fault, gup, etc,
will not be able to find valid vma, then just return SIGSEGV or -EFAULT as
expected.
If the vma has VM_HUGETLB | VM_PFNMAP, they are considered as special
mappings. They will be handled by without downgrading mmap_sem in this
patch since they may update vm flags.
But, with the "detach vmas first" approach, the vmas have been detached
when vm flags are updated, so it sounds safe to update vm flags with read
mmap_sem for this specific case. So, VM_HUGETLB and VM_PFNMAP will be
handled by using the optimized path in the following separate patches for
bisectable sake.
Unmapping uprobe areas may need update mm flags (MMF_RECALC_UPROBES).
However it is fine to have false-positive MMF_RECALC_UPROBES according to
uprobes developer.
With the "detach vmas first" approach we don't have to re-acquire mmap_sem
again to clean up vmas to avoid race window which might get the address
space changed since downgrade_write() doesn't release the lock to lead
regression, which simply downgrades to read lock.
And, since the lock acquire/release cost is managed to the minimum and
almost as same as before, the optimization could be extended to any size
of mapping without incurring significant penalty to small mappings.
For the time being, just do this in munmap syscall path. Other
vm_munmap() or do_munmap() call sites (i.e mmap, mremap, etc) remain
intact due to some implementation difficulties since they acquire write
mmap_sem from very beginning and hold it until the end, do_munmap() might
be called in the middle. But, the optimized do_munmap would like to be
called without mmap_sem held so that we can do the optimization. So, if
we want to do the similar optimization for mmap/mremap path, I'm afraid we
would have to redesign them. mremap might be called on very large area
depending on the usecases, the optimization to it will be considered in
the future.
With the patches, exclusive mmap_sem hold time when munmap a 80GB address
space on a machine with 32 cores of E5-2680 @ 2.70GHz dropped to us level
from second.
munmap_test-15002 [008] 594.380138: funcgraph_entry: |
__vm_munmap() {
munmap_test-15002 [008] 594.380146: funcgraph_entry: !2485684 us
| unmap_region();
munmap_test-15002 [008] 596.865836: funcgraph_exit: !2485692 us
| }
Here the execution time of unmap_region() is used to evaluate the time of
holding read mmap_sem, then the remaining time is used with holding
exclusive lock.
[1] https://lwn.net/Articles/753269/
Link: http://lkml.kernel.org/r/1537376621-51150-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>Suggested-by: Michal Hocko <mhocko@kernel.org>
Suggested-by: Kirill A. Shutemov <kirill@shutemov.name>
Suggested-by: Matthew Wilcox <willy@infradead.org>
Reviewed-by: Matthew Wilcox <willy@infradead.org>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Laurent Dufour <ldufour@linux.vnet.ibm.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-27 06:07:11 +08:00
|
|
|
ret = __do_munmap(mm, start, len, &uf, downgrade);
|
|
|
|
|
/*
|
2020-06-09 12:33:54 +08:00
|
|
|
* Returning 1 indicates mmap_lock is downgraded.
|
mm: mmap: zap pages with read mmap_sem in munmap
Patch series "mm: zap pages with read mmap_sem in munmap for large
mapping", v11.
Background:
Recently, when we ran some vm scalability tests on machines with large memory,
we ran into a couple of mmap_sem scalability issues when unmapping large memory
space, please refer to https://lkml.org/lkml/2017/12/14/733 and
https://lkml.org/lkml/2018/2/20/576.
History:
Then akpm suggested to unmap large mapping section by section and drop mmap_sem
at a time to mitigate it (see https://lkml.org/lkml/2018/3/6/784).
V1 patch series was submitted to the mailing list per Andrew's suggestion
(see https://lkml.org/lkml/2018/3/20/786). Then I received a lot great
feedback and suggestions.
Then this topic was discussed on LSFMM summit 2018. In the summit, Michal
Hocko suggested (also in the v1 patches review) to try "two phases"
approach. Zapping pages with read mmap_sem, then doing via cleanup with
write mmap_sem (for discussion detail, see
https://lwn.net/Articles/753269/)
Approach:
Zapping pages is the most time consuming part, according to the suggestion from
Michal Hocko [1], zapping pages can be done with holding read mmap_sem, like
what MADV_DONTNEED does. Then re-acquire write mmap_sem to cleanup vmas.
But, we can't call MADV_DONTNEED directly, since there are two major drawbacks:
* The unexpected state from PF if it wins the race in the middle of munmap.
It may return zero page, instead of the content or SIGSEGV.
* Can't handle VM_LOCKED | VM_HUGETLB | VM_PFNMAP and uprobe mappings, which
is a showstopper from akpm
But, some part may need write mmap_sem, for example, vma splitting. So,
the design is as follows:
acquire write mmap_sem
lookup vmas (find and split vmas)
deal with special mappings
detach vmas
downgrade_write
zap pages
free page tables
release mmap_sem
The vm events with read mmap_sem may come in during page zapping, but
since vmas have been detached before, they, i.e. page fault, gup, etc,
will not be able to find valid vma, then just return SIGSEGV or -EFAULT as
expected.
If the vma has VM_HUGETLB | VM_PFNMAP, they are considered as special
mappings. They will be handled by falling back to regular do_munmap()
with exclusive mmap_sem held in this patch since they may update vm flags.
But, with the "detach vmas first" approach, the vmas have been detached
when vm flags are updated, so it sounds safe to update vm flags with read
mmap_sem for this specific case. So, VM_HUGETLB and VM_PFNMAP will be
handled by using the optimized path in the following separate patches for
bisectable sake.
Unmapping uprobe areas may need update mm flags (MMF_RECALC_UPROBES).
However it is fine to have false-positive MMF_RECALC_UPROBES according to
uprobes developer. So, uprobe unmap will not be handled by the regular
path.
With the "detach vmas first" approach we don't have to re-acquire mmap_sem
again to clean up vmas to avoid race window which might get the address
space changed since downgrade_write() doesn't release the lock to lead
regression, which simply downgrades to read lock.
And, since the lock acquire/release cost is managed to the minimum and
almost as same as before, the optimization could be extended to any size
of mapping without incurring significant penalty to small mappings.
For the time being, just do this in munmap syscall path. Other
vm_munmap() or do_munmap() call sites (i.e mmap, mremap, etc) remain
intact due to some implementation difficulties since they acquire write
mmap_sem from very beginning and hold it until the end, do_munmap() might
be called in the middle. But, the optimized do_munmap would like to be
called without mmap_sem held so that we can do the optimization. So, if
we want to do the similar optimization for mmap/mremap path, I'm afraid we
would have to redesign them. mremap might be called on very large area
depending on the usecases, the optimization to it will be considered in
the future.
This patch (of 3):
When running some mmap/munmap scalability tests with large memory (i.e.
> 300GB), the below hung task issue may happen occasionally.
INFO: task ps:14018 blocked for more than 120 seconds.
Tainted: G E 4.9.79-009.ali3000.alios7.x86_64 #1
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this
message.
ps D 0 14018 1 0x00000004
ffff885582f84000 ffff885e8682f000 ffff880972943000 ffff885ebf499bc0
ffff8828ee120000 ffffc900349bfca8 ffffffff817154d0 0000000000000040
00ffffff812f872a ffff885ebf499bc0 024000d000948300 ffff880972943000
Call Trace:
[<ffffffff817154d0>] ? __schedule+0x250/0x730
[<ffffffff817159e6>] schedule+0x36/0x80
[<ffffffff81718560>] rwsem_down_read_failed+0xf0/0x150
[<ffffffff81390a28>] call_rwsem_down_read_failed+0x18/0x30
[<ffffffff81717db0>] down_read+0x20/0x40
[<ffffffff812b9439>] proc_pid_cmdline_read+0xd9/0x4e0
[<ffffffff81253c95>] ? do_filp_open+0xa5/0x100
[<ffffffff81241d87>] __vfs_read+0x37/0x150
[<ffffffff812f824b>] ? security_file_permission+0x9b/0xc0
[<ffffffff81242266>] vfs_read+0x96/0x130
[<ffffffff812437b5>] SyS_read+0x55/0xc0
[<ffffffff8171a6da>] entry_SYSCALL_64_fastpath+0x1a/0xc5
It is because munmap holds mmap_sem exclusively from very beginning to all
the way down to the end, and doesn't release it in the middle. When
unmapping large mapping, it may take long time (take ~18 seconds to unmap
320GB mapping with every single page mapped on an idle machine).
Zapping pages is the most time consuming part, according to the suggestion
from Michal Hocko [1], zapping pages can be done with holding read
mmap_sem, like what MADV_DONTNEED does. Then re-acquire write mmap_sem to
cleanup vmas.
But, some part may need write mmap_sem, for example, vma splitting. So,
the design is as follows:
acquire write mmap_sem
lookup vmas (find and split vmas)
deal with special mappings
detach vmas
downgrade_write
zap pages
free page tables
release mmap_sem
The vm events with read mmap_sem may come in during page zapping, but
since vmas have been detached before, they, i.e. page fault, gup, etc,
will not be able to find valid vma, then just return SIGSEGV or -EFAULT as
expected.
If the vma has VM_HUGETLB | VM_PFNMAP, they are considered as special
mappings. They will be handled by without downgrading mmap_sem in this
patch since they may update vm flags.
But, with the "detach vmas first" approach, the vmas have been detached
when vm flags are updated, so it sounds safe to update vm flags with read
mmap_sem for this specific case. So, VM_HUGETLB and VM_PFNMAP will be
handled by using the optimized path in the following separate patches for
bisectable sake.
Unmapping uprobe areas may need update mm flags (MMF_RECALC_UPROBES).
However it is fine to have false-positive MMF_RECALC_UPROBES according to
uprobes developer.
With the "detach vmas first" approach we don't have to re-acquire mmap_sem
again to clean up vmas to avoid race window which might get the address
space changed since downgrade_write() doesn't release the lock to lead
regression, which simply downgrades to read lock.
And, since the lock acquire/release cost is managed to the minimum and
almost as same as before, the optimization could be extended to any size
of mapping without incurring significant penalty to small mappings.
For the time being, just do this in munmap syscall path. Other
vm_munmap() or do_munmap() call sites (i.e mmap, mremap, etc) remain
intact due to some implementation difficulties since they acquire write
mmap_sem from very beginning and hold it until the end, do_munmap() might
be called in the middle. But, the optimized do_munmap would like to be
called without mmap_sem held so that we can do the optimization. So, if
we want to do the similar optimization for mmap/mremap path, I'm afraid we
would have to redesign them. mremap might be called on very large area
depending on the usecases, the optimization to it will be considered in
the future.
With the patches, exclusive mmap_sem hold time when munmap a 80GB address
space on a machine with 32 cores of E5-2680 @ 2.70GHz dropped to us level
from second.
munmap_test-15002 [008] 594.380138: funcgraph_entry: |
__vm_munmap() {
munmap_test-15002 [008] 594.380146: funcgraph_entry: !2485684 us
| unmap_region();
munmap_test-15002 [008] 596.865836: funcgraph_exit: !2485692 us
| }
Here the execution time of unmap_region() is used to evaluate the time of
holding read mmap_sem, then the remaining time is used with holding
exclusive lock.
[1] https://lwn.net/Articles/753269/
Link: http://lkml.kernel.org/r/1537376621-51150-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>Suggested-by: Michal Hocko <mhocko@kernel.org>
Suggested-by: Kirill A. Shutemov <kirill@shutemov.name>
Suggested-by: Matthew Wilcox <willy@infradead.org>
Reviewed-by: Matthew Wilcox <willy@infradead.org>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Laurent Dufour <ldufour@linux.vnet.ibm.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-27 06:07:11 +08:00
|
|
|
* But 1 is not legal return value of vm_munmap() and munmap(), reset
|
|
|
|
|
* it to 0 before return.
|
|
|
|
|
*/
|
|
|
|
|
if (ret == 1) {
|
2020-06-09 12:33:25 +08:00
|
|
|
mmap_read_unlock(mm);
|
mm: mmap: zap pages with read mmap_sem in munmap
Patch series "mm: zap pages with read mmap_sem in munmap for large
mapping", v11.
Background:
Recently, when we ran some vm scalability tests on machines with large memory,
we ran into a couple of mmap_sem scalability issues when unmapping large memory
space, please refer to https://lkml.org/lkml/2017/12/14/733 and
https://lkml.org/lkml/2018/2/20/576.
History:
Then akpm suggested to unmap large mapping section by section and drop mmap_sem
at a time to mitigate it (see https://lkml.org/lkml/2018/3/6/784).
V1 patch series was submitted to the mailing list per Andrew's suggestion
(see https://lkml.org/lkml/2018/3/20/786). Then I received a lot great
feedback and suggestions.
Then this topic was discussed on LSFMM summit 2018. In the summit, Michal
Hocko suggested (also in the v1 patches review) to try "two phases"
approach. Zapping pages with read mmap_sem, then doing via cleanup with
write mmap_sem (for discussion detail, see
https://lwn.net/Articles/753269/)
Approach:
Zapping pages is the most time consuming part, according to the suggestion from
Michal Hocko [1], zapping pages can be done with holding read mmap_sem, like
what MADV_DONTNEED does. Then re-acquire write mmap_sem to cleanup vmas.
But, we can't call MADV_DONTNEED directly, since there are two major drawbacks:
* The unexpected state from PF if it wins the race in the middle of munmap.
It may return zero page, instead of the content or SIGSEGV.
* Can't handle VM_LOCKED | VM_HUGETLB | VM_PFNMAP and uprobe mappings, which
is a showstopper from akpm
But, some part may need write mmap_sem, for example, vma splitting. So,
the design is as follows:
acquire write mmap_sem
lookup vmas (find and split vmas)
deal with special mappings
detach vmas
downgrade_write
zap pages
free page tables
release mmap_sem
The vm events with read mmap_sem may come in during page zapping, but
since vmas have been detached before, they, i.e. page fault, gup, etc,
will not be able to find valid vma, then just return SIGSEGV or -EFAULT as
expected.
If the vma has VM_HUGETLB | VM_PFNMAP, they are considered as special
mappings. They will be handled by falling back to regular do_munmap()
with exclusive mmap_sem held in this patch since they may update vm flags.
But, with the "detach vmas first" approach, the vmas have been detached
when vm flags are updated, so it sounds safe to update vm flags with read
mmap_sem for this specific case. So, VM_HUGETLB and VM_PFNMAP will be
handled by using the optimized path in the following separate patches for
bisectable sake.
Unmapping uprobe areas may need update mm flags (MMF_RECALC_UPROBES).
However it is fine to have false-positive MMF_RECALC_UPROBES according to
uprobes developer. So, uprobe unmap will not be handled by the regular
path.
With the "detach vmas first" approach we don't have to re-acquire mmap_sem
again to clean up vmas to avoid race window which might get the address
space changed since downgrade_write() doesn't release the lock to lead
regression, which simply downgrades to read lock.
And, since the lock acquire/release cost is managed to the minimum and
almost as same as before, the optimization could be extended to any size
of mapping without incurring significant penalty to small mappings.
For the time being, just do this in munmap syscall path. Other
vm_munmap() or do_munmap() call sites (i.e mmap, mremap, etc) remain
intact due to some implementation difficulties since they acquire write
mmap_sem from very beginning and hold it until the end, do_munmap() might
be called in the middle. But, the optimized do_munmap would like to be
called without mmap_sem held so that we can do the optimization. So, if
we want to do the similar optimization for mmap/mremap path, I'm afraid we
would have to redesign them. mremap might be called on very large area
depending on the usecases, the optimization to it will be considered in
the future.
This patch (of 3):
When running some mmap/munmap scalability tests with large memory (i.e.
> 300GB), the below hung task issue may happen occasionally.
INFO: task ps:14018 blocked for more than 120 seconds.
Tainted: G E 4.9.79-009.ali3000.alios7.x86_64 #1
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this
message.
ps D 0 14018 1 0x00000004
ffff885582f84000 ffff885e8682f000 ffff880972943000 ffff885ebf499bc0
ffff8828ee120000 ffffc900349bfca8 ffffffff817154d0 0000000000000040
00ffffff812f872a ffff885ebf499bc0 024000d000948300 ffff880972943000
Call Trace:
[<ffffffff817154d0>] ? __schedule+0x250/0x730
[<ffffffff817159e6>] schedule+0x36/0x80
[<ffffffff81718560>] rwsem_down_read_failed+0xf0/0x150
[<ffffffff81390a28>] call_rwsem_down_read_failed+0x18/0x30
[<ffffffff81717db0>] down_read+0x20/0x40
[<ffffffff812b9439>] proc_pid_cmdline_read+0xd9/0x4e0
[<ffffffff81253c95>] ? do_filp_open+0xa5/0x100
[<ffffffff81241d87>] __vfs_read+0x37/0x150
[<ffffffff812f824b>] ? security_file_permission+0x9b/0xc0
[<ffffffff81242266>] vfs_read+0x96/0x130
[<ffffffff812437b5>] SyS_read+0x55/0xc0
[<ffffffff8171a6da>] entry_SYSCALL_64_fastpath+0x1a/0xc5
It is because munmap holds mmap_sem exclusively from very beginning to all
the way down to the end, and doesn't release it in the middle. When
unmapping large mapping, it may take long time (take ~18 seconds to unmap
320GB mapping with every single page mapped on an idle machine).
Zapping pages is the most time consuming part, according to the suggestion
from Michal Hocko [1], zapping pages can be done with holding read
mmap_sem, like what MADV_DONTNEED does. Then re-acquire write mmap_sem to
cleanup vmas.
But, some part may need write mmap_sem, for example, vma splitting. So,
the design is as follows:
acquire write mmap_sem
lookup vmas (find and split vmas)
deal with special mappings
detach vmas
downgrade_write
zap pages
free page tables
release mmap_sem
The vm events with read mmap_sem may come in during page zapping, but
since vmas have been detached before, they, i.e. page fault, gup, etc,
will not be able to find valid vma, then just return SIGSEGV or -EFAULT as
expected.
If the vma has VM_HUGETLB | VM_PFNMAP, they are considered as special
mappings. They will be handled by without downgrading mmap_sem in this
patch since they may update vm flags.
But, with the "detach vmas first" approach, the vmas have been detached
when vm flags are updated, so it sounds safe to update vm flags with read
mmap_sem for this specific case. So, VM_HUGETLB and VM_PFNMAP will be
handled by using the optimized path in the following separate patches for
bisectable sake.
Unmapping uprobe areas may need update mm flags (MMF_RECALC_UPROBES).
However it is fine to have false-positive MMF_RECALC_UPROBES according to
uprobes developer.
With the "detach vmas first" approach we don't have to re-acquire mmap_sem
again to clean up vmas to avoid race window which might get the address
space changed since downgrade_write() doesn't release the lock to lead
regression, which simply downgrades to read lock.
And, since the lock acquire/release cost is managed to the minimum and
almost as same as before, the optimization could be extended to any size
of mapping without incurring significant penalty to small mappings.
For the time being, just do this in munmap syscall path. Other
vm_munmap() or do_munmap() call sites (i.e mmap, mremap, etc) remain
intact due to some implementation difficulties since they acquire write
mmap_sem from very beginning and hold it until the end, do_munmap() might
be called in the middle. But, the optimized do_munmap would like to be
called without mmap_sem held so that we can do the optimization. So, if
we want to do the similar optimization for mmap/mremap path, I'm afraid we
would have to redesign them. mremap might be called on very large area
depending on the usecases, the optimization to it will be considered in
the future.
With the patches, exclusive mmap_sem hold time when munmap a 80GB address
space on a machine with 32 cores of E5-2680 @ 2.70GHz dropped to us level
from second.
munmap_test-15002 [008] 594.380138: funcgraph_entry: |
__vm_munmap() {
munmap_test-15002 [008] 594.380146: funcgraph_entry: !2485684 us
| unmap_region();
munmap_test-15002 [008] 596.865836: funcgraph_exit: !2485692 us
| }
Here the execution time of unmap_region() is used to evaluate the time of
holding read mmap_sem, then the remaining time is used with holding
exclusive lock.
[1] https://lwn.net/Articles/753269/
Link: http://lkml.kernel.org/r/1537376621-51150-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>Suggested-by: Michal Hocko <mhocko@kernel.org>
Suggested-by: Kirill A. Shutemov <kirill@shutemov.name>
Suggested-by: Matthew Wilcox <willy@infradead.org>
Reviewed-by: Matthew Wilcox <willy@infradead.org>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Laurent Dufour <ldufour@linux.vnet.ibm.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-27 06:07:11 +08:00
|
|
|
ret = 0;
|
|
|
|
|
} else
|
2020-06-09 12:33:25 +08:00
|
|
|
mmap_write_unlock(mm);
|
mm: mmap: zap pages with read mmap_sem in munmap
Patch series "mm: zap pages with read mmap_sem in munmap for large
mapping", v11.
Background:
Recently, when we ran some vm scalability tests on machines with large memory,
we ran into a couple of mmap_sem scalability issues when unmapping large memory
space, please refer to https://lkml.org/lkml/2017/12/14/733 and
https://lkml.org/lkml/2018/2/20/576.
History:
Then akpm suggested to unmap large mapping section by section and drop mmap_sem
at a time to mitigate it (see https://lkml.org/lkml/2018/3/6/784).
V1 patch series was submitted to the mailing list per Andrew's suggestion
(see https://lkml.org/lkml/2018/3/20/786). Then I received a lot great
feedback and suggestions.
Then this topic was discussed on LSFMM summit 2018. In the summit, Michal
Hocko suggested (also in the v1 patches review) to try "two phases"
approach. Zapping pages with read mmap_sem, then doing via cleanup with
write mmap_sem (for discussion detail, see
https://lwn.net/Articles/753269/)
Approach:
Zapping pages is the most time consuming part, according to the suggestion from
Michal Hocko [1], zapping pages can be done with holding read mmap_sem, like
what MADV_DONTNEED does. Then re-acquire write mmap_sem to cleanup vmas.
But, we can't call MADV_DONTNEED directly, since there are two major drawbacks:
* The unexpected state from PF if it wins the race in the middle of munmap.
It may return zero page, instead of the content or SIGSEGV.
* Can't handle VM_LOCKED | VM_HUGETLB | VM_PFNMAP and uprobe mappings, which
is a showstopper from akpm
But, some part may need write mmap_sem, for example, vma splitting. So,
the design is as follows:
acquire write mmap_sem
lookup vmas (find and split vmas)
deal with special mappings
detach vmas
downgrade_write
zap pages
free page tables
release mmap_sem
The vm events with read mmap_sem may come in during page zapping, but
since vmas have been detached before, they, i.e. page fault, gup, etc,
will not be able to find valid vma, then just return SIGSEGV or -EFAULT as
expected.
If the vma has VM_HUGETLB | VM_PFNMAP, they are considered as special
mappings. They will be handled by falling back to regular do_munmap()
with exclusive mmap_sem held in this patch since they may update vm flags.
But, with the "detach vmas first" approach, the vmas have been detached
when vm flags are updated, so it sounds safe to update vm flags with read
mmap_sem for this specific case. So, VM_HUGETLB and VM_PFNMAP will be
handled by using the optimized path in the following separate patches for
bisectable sake.
Unmapping uprobe areas may need update mm flags (MMF_RECALC_UPROBES).
However it is fine to have false-positive MMF_RECALC_UPROBES according to
uprobes developer. So, uprobe unmap will not be handled by the regular
path.
With the "detach vmas first" approach we don't have to re-acquire mmap_sem
again to clean up vmas to avoid race window which might get the address
space changed since downgrade_write() doesn't release the lock to lead
regression, which simply downgrades to read lock.
And, since the lock acquire/release cost is managed to the minimum and
almost as same as before, the optimization could be extended to any size
of mapping without incurring significant penalty to small mappings.
For the time being, just do this in munmap syscall path. Other
vm_munmap() or do_munmap() call sites (i.e mmap, mremap, etc) remain
intact due to some implementation difficulties since they acquire write
mmap_sem from very beginning and hold it until the end, do_munmap() might
be called in the middle. But, the optimized do_munmap would like to be
called without mmap_sem held so that we can do the optimization. So, if
we want to do the similar optimization for mmap/mremap path, I'm afraid we
would have to redesign them. mremap might be called on very large area
depending on the usecases, the optimization to it will be considered in
the future.
This patch (of 3):
When running some mmap/munmap scalability tests with large memory (i.e.
> 300GB), the below hung task issue may happen occasionally.
INFO: task ps:14018 blocked for more than 120 seconds.
Tainted: G E 4.9.79-009.ali3000.alios7.x86_64 #1
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this
message.
ps D 0 14018 1 0x00000004
ffff885582f84000 ffff885e8682f000 ffff880972943000 ffff885ebf499bc0
ffff8828ee120000 ffffc900349bfca8 ffffffff817154d0 0000000000000040
00ffffff812f872a ffff885ebf499bc0 024000d000948300 ffff880972943000
Call Trace:
[<ffffffff817154d0>] ? __schedule+0x250/0x730
[<ffffffff817159e6>] schedule+0x36/0x80
[<ffffffff81718560>] rwsem_down_read_failed+0xf0/0x150
[<ffffffff81390a28>] call_rwsem_down_read_failed+0x18/0x30
[<ffffffff81717db0>] down_read+0x20/0x40
[<ffffffff812b9439>] proc_pid_cmdline_read+0xd9/0x4e0
[<ffffffff81253c95>] ? do_filp_open+0xa5/0x100
[<ffffffff81241d87>] __vfs_read+0x37/0x150
[<ffffffff812f824b>] ? security_file_permission+0x9b/0xc0
[<ffffffff81242266>] vfs_read+0x96/0x130
[<ffffffff812437b5>] SyS_read+0x55/0xc0
[<ffffffff8171a6da>] entry_SYSCALL_64_fastpath+0x1a/0xc5
It is because munmap holds mmap_sem exclusively from very beginning to all
the way down to the end, and doesn't release it in the middle. When
unmapping large mapping, it may take long time (take ~18 seconds to unmap
320GB mapping with every single page mapped on an idle machine).
Zapping pages is the most time consuming part, according to the suggestion
from Michal Hocko [1], zapping pages can be done with holding read
mmap_sem, like what MADV_DONTNEED does. Then re-acquire write mmap_sem to
cleanup vmas.
But, some part may need write mmap_sem, for example, vma splitting. So,
the design is as follows:
acquire write mmap_sem
lookup vmas (find and split vmas)
deal with special mappings
detach vmas
downgrade_write
zap pages
free page tables
release mmap_sem
The vm events with read mmap_sem may come in during page zapping, but
since vmas have been detached before, they, i.e. page fault, gup, etc,
will not be able to find valid vma, then just return SIGSEGV or -EFAULT as
expected.
If the vma has VM_HUGETLB | VM_PFNMAP, they are considered as special
mappings. They will be handled by without downgrading mmap_sem in this
patch since they may update vm flags.
But, with the "detach vmas first" approach, the vmas have been detached
when vm flags are updated, so it sounds safe to update vm flags with read
mmap_sem for this specific case. So, VM_HUGETLB and VM_PFNMAP will be
handled by using the optimized path in the following separate patches for
bisectable sake.
Unmapping uprobe areas may need update mm flags (MMF_RECALC_UPROBES).
However it is fine to have false-positive MMF_RECALC_UPROBES according to
uprobes developer.
With the "detach vmas first" approach we don't have to re-acquire mmap_sem
again to clean up vmas to avoid race window which might get the address
space changed since downgrade_write() doesn't release the lock to lead
regression, which simply downgrades to read lock.
And, since the lock acquire/release cost is managed to the minimum and
almost as same as before, the optimization could be extended to any size
of mapping without incurring significant penalty to small mappings.
For the time being, just do this in munmap syscall path. Other
vm_munmap() or do_munmap() call sites (i.e mmap, mremap, etc) remain
intact due to some implementation difficulties since they acquire write
mmap_sem from very beginning and hold it until the end, do_munmap() might
be called in the middle. But, the optimized do_munmap would like to be
called without mmap_sem held so that we can do the optimization. So, if
we want to do the similar optimization for mmap/mremap path, I'm afraid we
would have to redesign them. mremap might be called on very large area
depending on the usecases, the optimization to it will be considered in
the future.
With the patches, exclusive mmap_sem hold time when munmap a 80GB address
space on a machine with 32 cores of E5-2680 @ 2.70GHz dropped to us level
from second.
munmap_test-15002 [008] 594.380138: funcgraph_entry: |
__vm_munmap() {
munmap_test-15002 [008] 594.380146: funcgraph_entry: !2485684 us
| unmap_region();
munmap_test-15002 [008] 596.865836: funcgraph_exit: !2485692 us
| }
Here the execution time of unmap_region() is used to evaluate the time of
holding read mmap_sem, then the remaining time is used with holding
exclusive lock.
[1] https://lwn.net/Articles/753269/
Link: http://lkml.kernel.org/r/1537376621-51150-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>Suggested-by: Michal Hocko <mhocko@kernel.org>
Suggested-by: Kirill A. Shutemov <kirill@shutemov.name>
Suggested-by: Matthew Wilcox <willy@infradead.org>
Reviewed-by: Matthew Wilcox <willy@infradead.org>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Laurent Dufour <ldufour@linux.vnet.ibm.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-27 06:07:11 +08:00
|
|
|
|
2017-02-25 06:58:22 +08:00
|
|
|
userfaultfd_unmap_complete(mm, &uf);
|
2005-04-17 06:20:36 +08:00
|
|
|
return ret;
|
|
|
|
|
}
|
mm: mmap: zap pages with read mmap_sem in munmap
Patch series "mm: zap pages with read mmap_sem in munmap for large
mapping", v11.
Background:
Recently, when we ran some vm scalability tests on machines with large memory,
we ran into a couple of mmap_sem scalability issues when unmapping large memory
space, please refer to https://lkml.org/lkml/2017/12/14/733 and
https://lkml.org/lkml/2018/2/20/576.
History:
Then akpm suggested to unmap large mapping section by section and drop mmap_sem
at a time to mitigate it (see https://lkml.org/lkml/2018/3/6/784).
V1 patch series was submitted to the mailing list per Andrew's suggestion
(see https://lkml.org/lkml/2018/3/20/786). Then I received a lot great
feedback and suggestions.
Then this topic was discussed on LSFMM summit 2018. In the summit, Michal
Hocko suggested (also in the v1 patches review) to try "two phases"
approach. Zapping pages with read mmap_sem, then doing via cleanup with
write mmap_sem (for discussion detail, see
https://lwn.net/Articles/753269/)
Approach:
Zapping pages is the most time consuming part, according to the suggestion from
Michal Hocko [1], zapping pages can be done with holding read mmap_sem, like
what MADV_DONTNEED does. Then re-acquire write mmap_sem to cleanup vmas.
But, we can't call MADV_DONTNEED directly, since there are two major drawbacks:
* The unexpected state from PF if it wins the race in the middle of munmap.
It may return zero page, instead of the content or SIGSEGV.
* Can't handle VM_LOCKED | VM_HUGETLB | VM_PFNMAP and uprobe mappings, which
is a showstopper from akpm
But, some part may need write mmap_sem, for example, vma splitting. So,
the design is as follows:
acquire write mmap_sem
lookup vmas (find and split vmas)
deal with special mappings
detach vmas
downgrade_write
zap pages
free page tables
release mmap_sem
The vm events with read mmap_sem may come in during page zapping, but
since vmas have been detached before, they, i.e. page fault, gup, etc,
will not be able to find valid vma, then just return SIGSEGV or -EFAULT as
expected.
If the vma has VM_HUGETLB | VM_PFNMAP, they are considered as special
mappings. They will be handled by falling back to regular do_munmap()
with exclusive mmap_sem held in this patch since they may update vm flags.
But, with the "detach vmas first" approach, the vmas have been detached
when vm flags are updated, so it sounds safe to update vm flags with read
mmap_sem for this specific case. So, VM_HUGETLB and VM_PFNMAP will be
handled by using the optimized path in the following separate patches for
bisectable sake.
Unmapping uprobe areas may need update mm flags (MMF_RECALC_UPROBES).
However it is fine to have false-positive MMF_RECALC_UPROBES according to
uprobes developer. So, uprobe unmap will not be handled by the regular
path.
With the "detach vmas first" approach we don't have to re-acquire mmap_sem
again to clean up vmas to avoid race window which might get the address
space changed since downgrade_write() doesn't release the lock to lead
regression, which simply downgrades to read lock.
And, since the lock acquire/release cost is managed to the minimum and
almost as same as before, the optimization could be extended to any size
of mapping without incurring significant penalty to small mappings.
For the time being, just do this in munmap syscall path. Other
vm_munmap() or do_munmap() call sites (i.e mmap, mremap, etc) remain
intact due to some implementation difficulties since they acquire write
mmap_sem from very beginning and hold it until the end, do_munmap() might
be called in the middle. But, the optimized do_munmap would like to be
called without mmap_sem held so that we can do the optimization. So, if
we want to do the similar optimization for mmap/mremap path, I'm afraid we
would have to redesign them. mremap might be called on very large area
depending on the usecases, the optimization to it will be considered in
the future.
This patch (of 3):
When running some mmap/munmap scalability tests with large memory (i.e.
> 300GB), the below hung task issue may happen occasionally.
INFO: task ps:14018 blocked for more than 120 seconds.
Tainted: G E 4.9.79-009.ali3000.alios7.x86_64 #1
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this
message.
ps D 0 14018 1 0x00000004
ffff885582f84000 ffff885e8682f000 ffff880972943000 ffff885ebf499bc0
ffff8828ee120000 ffffc900349bfca8 ffffffff817154d0 0000000000000040
00ffffff812f872a ffff885ebf499bc0 024000d000948300 ffff880972943000
Call Trace:
[<ffffffff817154d0>] ? __schedule+0x250/0x730
[<ffffffff817159e6>] schedule+0x36/0x80
[<ffffffff81718560>] rwsem_down_read_failed+0xf0/0x150
[<ffffffff81390a28>] call_rwsem_down_read_failed+0x18/0x30
[<ffffffff81717db0>] down_read+0x20/0x40
[<ffffffff812b9439>] proc_pid_cmdline_read+0xd9/0x4e0
[<ffffffff81253c95>] ? do_filp_open+0xa5/0x100
[<ffffffff81241d87>] __vfs_read+0x37/0x150
[<ffffffff812f824b>] ? security_file_permission+0x9b/0xc0
[<ffffffff81242266>] vfs_read+0x96/0x130
[<ffffffff812437b5>] SyS_read+0x55/0xc0
[<ffffffff8171a6da>] entry_SYSCALL_64_fastpath+0x1a/0xc5
It is because munmap holds mmap_sem exclusively from very beginning to all
the way down to the end, and doesn't release it in the middle. When
unmapping large mapping, it may take long time (take ~18 seconds to unmap
320GB mapping with every single page mapped on an idle machine).
Zapping pages is the most time consuming part, according to the suggestion
from Michal Hocko [1], zapping pages can be done with holding read
mmap_sem, like what MADV_DONTNEED does. Then re-acquire write mmap_sem to
cleanup vmas.
But, some part may need write mmap_sem, for example, vma splitting. So,
the design is as follows:
acquire write mmap_sem
lookup vmas (find and split vmas)
deal with special mappings
detach vmas
downgrade_write
zap pages
free page tables
release mmap_sem
The vm events with read mmap_sem may come in during page zapping, but
since vmas have been detached before, they, i.e. page fault, gup, etc,
will not be able to find valid vma, then just return SIGSEGV or -EFAULT as
expected.
If the vma has VM_HUGETLB | VM_PFNMAP, they are considered as special
mappings. They will be handled by without downgrading mmap_sem in this
patch since they may update vm flags.
But, with the "detach vmas first" approach, the vmas have been detached
when vm flags are updated, so it sounds safe to update vm flags with read
mmap_sem for this specific case. So, VM_HUGETLB and VM_PFNMAP will be
handled by using the optimized path in the following separate patches for
bisectable sake.
Unmapping uprobe areas may need update mm flags (MMF_RECALC_UPROBES).
However it is fine to have false-positive MMF_RECALC_UPROBES according to
uprobes developer.
With the "detach vmas first" approach we don't have to re-acquire mmap_sem
again to clean up vmas to avoid race window which might get the address
space changed since downgrade_write() doesn't release the lock to lead
regression, which simply downgrades to read lock.
And, since the lock acquire/release cost is managed to the minimum and
almost as same as before, the optimization could be extended to any size
of mapping without incurring significant penalty to small mappings.
For the time being, just do this in munmap syscall path. Other
vm_munmap() or do_munmap() call sites (i.e mmap, mremap, etc) remain
intact due to some implementation difficulties since they acquire write
mmap_sem from very beginning and hold it until the end, do_munmap() might
be called in the middle. But, the optimized do_munmap would like to be
called without mmap_sem held so that we can do the optimization. So, if
we want to do the similar optimization for mmap/mremap path, I'm afraid we
would have to redesign them. mremap might be called on very large area
depending on the usecases, the optimization to it will be considered in
the future.
With the patches, exclusive mmap_sem hold time when munmap a 80GB address
space on a machine with 32 cores of E5-2680 @ 2.70GHz dropped to us level
from second.
munmap_test-15002 [008] 594.380138: funcgraph_entry: |
__vm_munmap() {
munmap_test-15002 [008] 594.380146: funcgraph_entry: !2485684 us
| unmap_region();
munmap_test-15002 [008] 596.865836: funcgraph_exit: !2485692 us
| }
Here the execution time of unmap_region() is used to evaluate the time of
holding read mmap_sem, then the remaining time is used with holding
exclusive lock.
[1] https://lwn.net/Articles/753269/
Link: http://lkml.kernel.org/r/1537376621-51150-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>Suggested-by: Michal Hocko <mhocko@kernel.org>
Suggested-by: Kirill A. Shutemov <kirill@shutemov.name>
Suggested-by: Matthew Wilcox <willy@infradead.org>
Reviewed-by: Matthew Wilcox <willy@infradead.org>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Laurent Dufour <ldufour@linux.vnet.ibm.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-27 06:07:11 +08:00
|
|
|
|
|
|
|
|
int vm_munmap(unsigned long start, size_t len)
|
|
|
|
|
{
|
|
|
|
|
return __vm_munmap(start, len, false);
|
|
|
|
|
}
|
2012-04-21 07:20:01 +08:00
|
|
|
EXPORT_SYMBOL(vm_munmap);
|
|
|
|
|
|
|
|
|
|
SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
|
|
|
|
|
{
|
2019-09-26 07:49:04 +08:00
|
|
|
addr = untagged_addr(addr);
|
mm: mmap: zap pages with read mmap_sem in munmap
Patch series "mm: zap pages with read mmap_sem in munmap for large
mapping", v11.
Background:
Recently, when we ran some vm scalability tests on machines with large memory,
we ran into a couple of mmap_sem scalability issues when unmapping large memory
space, please refer to https://lkml.org/lkml/2017/12/14/733 and
https://lkml.org/lkml/2018/2/20/576.
History:
Then akpm suggested to unmap large mapping section by section and drop mmap_sem
at a time to mitigate it (see https://lkml.org/lkml/2018/3/6/784).
V1 patch series was submitted to the mailing list per Andrew's suggestion
(see https://lkml.org/lkml/2018/3/20/786). Then I received a lot great
feedback and suggestions.
Then this topic was discussed on LSFMM summit 2018. In the summit, Michal
Hocko suggested (also in the v1 patches review) to try "two phases"
approach. Zapping pages with read mmap_sem, then doing via cleanup with
write mmap_sem (for discussion detail, see
https://lwn.net/Articles/753269/)
Approach:
Zapping pages is the most time consuming part, according to the suggestion from
Michal Hocko [1], zapping pages can be done with holding read mmap_sem, like
what MADV_DONTNEED does. Then re-acquire write mmap_sem to cleanup vmas.
But, we can't call MADV_DONTNEED directly, since there are two major drawbacks:
* The unexpected state from PF if it wins the race in the middle of munmap.
It may return zero page, instead of the content or SIGSEGV.
* Can't handle VM_LOCKED | VM_HUGETLB | VM_PFNMAP and uprobe mappings, which
is a showstopper from akpm
But, some part may need write mmap_sem, for example, vma splitting. So,
the design is as follows:
acquire write mmap_sem
lookup vmas (find and split vmas)
deal with special mappings
detach vmas
downgrade_write
zap pages
free page tables
release mmap_sem
The vm events with read mmap_sem may come in during page zapping, but
since vmas have been detached before, they, i.e. page fault, gup, etc,
will not be able to find valid vma, then just return SIGSEGV or -EFAULT as
expected.
If the vma has VM_HUGETLB | VM_PFNMAP, they are considered as special
mappings. They will be handled by falling back to regular do_munmap()
with exclusive mmap_sem held in this patch since they may update vm flags.
But, with the "detach vmas first" approach, the vmas have been detached
when vm flags are updated, so it sounds safe to update vm flags with read
mmap_sem for this specific case. So, VM_HUGETLB and VM_PFNMAP will be
handled by using the optimized path in the following separate patches for
bisectable sake.
Unmapping uprobe areas may need update mm flags (MMF_RECALC_UPROBES).
However it is fine to have false-positive MMF_RECALC_UPROBES according to
uprobes developer. So, uprobe unmap will not be handled by the regular
path.
With the "detach vmas first" approach we don't have to re-acquire mmap_sem
again to clean up vmas to avoid race window which might get the address
space changed since downgrade_write() doesn't release the lock to lead
regression, which simply downgrades to read lock.
And, since the lock acquire/release cost is managed to the minimum and
almost as same as before, the optimization could be extended to any size
of mapping without incurring significant penalty to small mappings.
For the time being, just do this in munmap syscall path. Other
vm_munmap() or do_munmap() call sites (i.e mmap, mremap, etc) remain
intact due to some implementation difficulties since they acquire write
mmap_sem from very beginning and hold it until the end, do_munmap() might
be called in the middle. But, the optimized do_munmap would like to be
called without mmap_sem held so that we can do the optimization. So, if
we want to do the similar optimization for mmap/mremap path, I'm afraid we
would have to redesign them. mremap might be called on very large area
depending on the usecases, the optimization to it will be considered in
the future.
This patch (of 3):
When running some mmap/munmap scalability tests with large memory (i.e.
> 300GB), the below hung task issue may happen occasionally.
INFO: task ps:14018 blocked for more than 120 seconds.
Tainted: G E 4.9.79-009.ali3000.alios7.x86_64 #1
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this
message.
ps D 0 14018 1 0x00000004
ffff885582f84000 ffff885e8682f000 ffff880972943000 ffff885ebf499bc0
ffff8828ee120000 ffffc900349bfca8 ffffffff817154d0 0000000000000040
00ffffff812f872a ffff885ebf499bc0 024000d000948300 ffff880972943000
Call Trace:
[<ffffffff817154d0>] ? __schedule+0x250/0x730
[<ffffffff817159e6>] schedule+0x36/0x80
[<ffffffff81718560>] rwsem_down_read_failed+0xf0/0x150
[<ffffffff81390a28>] call_rwsem_down_read_failed+0x18/0x30
[<ffffffff81717db0>] down_read+0x20/0x40
[<ffffffff812b9439>] proc_pid_cmdline_read+0xd9/0x4e0
[<ffffffff81253c95>] ? do_filp_open+0xa5/0x100
[<ffffffff81241d87>] __vfs_read+0x37/0x150
[<ffffffff812f824b>] ? security_file_permission+0x9b/0xc0
[<ffffffff81242266>] vfs_read+0x96/0x130
[<ffffffff812437b5>] SyS_read+0x55/0xc0
[<ffffffff8171a6da>] entry_SYSCALL_64_fastpath+0x1a/0xc5
It is because munmap holds mmap_sem exclusively from very beginning to all
the way down to the end, and doesn't release it in the middle. When
unmapping large mapping, it may take long time (take ~18 seconds to unmap
320GB mapping with every single page mapped on an idle machine).
Zapping pages is the most time consuming part, according to the suggestion
from Michal Hocko [1], zapping pages can be done with holding read
mmap_sem, like what MADV_DONTNEED does. Then re-acquire write mmap_sem to
cleanup vmas.
But, some part may need write mmap_sem, for example, vma splitting. So,
the design is as follows:
acquire write mmap_sem
lookup vmas (find and split vmas)
deal with special mappings
detach vmas
downgrade_write
zap pages
free page tables
release mmap_sem
The vm events with read mmap_sem may come in during page zapping, but
since vmas have been detached before, they, i.e. page fault, gup, etc,
will not be able to find valid vma, then just return SIGSEGV or -EFAULT as
expected.
If the vma has VM_HUGETLB | VM_PFNMAP, they are considered as special
mappings. They will be handled by without downgrading mmap_sem in this
patch since they may update vm flags.
But, with the "detach vmas first" approach, the vmas have been detached
when vm flags are updated, so it sounds safe to update vm flags with read
mmap_sem for this specific case. So, VM_HUGETLB and VM_PFNMAP will be
handled by using the optimized path in the following separate patches for
bisectable sake.
Unmapping uprobe areas may need update mm flags (MMF_RECALC_UPROBES).
However it is fine to have false-positive MMF_RECALC_UPROBES according to
uprobes developer.
With the "detach vmas first" approach we don't have to re-acquire mmap_sem
again to clean up vmas to avoid race window which might get the address
space changed since downgrade_write() doesn't release the lock to lead
regression, which simply downgrades to read lock.
And, since the lock acquire/release cost is managed to the minimum and
almost as same as before, the optimization could be extended to any size
of mapping without incurring significant penalty to small mappings.
For the time being, just do this in munmap syscall path. Other
vm_munmap() or do_munmap() call sites (i.e mmap, mremap, etc) remain
intact due to some implementation difficulties since they acquire write
mmap_sem from very beginning and hold it until the end, do_munmap() might
be called in the middle. But, the optimized do_munmap would like to be
called without mmap_sem held so that we can do the optimization. So, if
we want to do the similar optimization for mmap/mremap path, I'm afraid we
would have to redesign them. mremap might be called on very large area
depending on the usecases, the optimization to it will be considered in
the future.
With the patches, exclusive mmap_sem hold time when munmap a 80GB address
space on a machine with 32 cores of E5-2680 @ 2.70GHz dropped to us level
from second.
munmap_test-15002 [008] 594.380138: funcgraph_entry: |
__vm_munmap() {
munmap_test-15002 [008] 594.380146: funcgraph_entry: !2485684 us
| unmap_region();
munmap_test-15002 [008] 596.865836: funcgraph_exit: !2485692 us
| }
Here the execution time of unmap_region() is used to evaluate the time of
holding read mmap_sem, then the remaining time is used with holding
exclusive lock.
[1] https://lwn.net/Articles/753269/
Link: http://lkml.kernel.org/r/1537376621-51150-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>Suggested-by: Michal Hocko <mhocko@kernel.org>
Suggested-by: Kirill A. Shutemov <kirill@shutemov.name>
Suggested-by: Matthew Wilcox <willy@infradead.org>
Reviewed-by: Matthew Wilcox <willy@infradead.org>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Laurent Dufour <ldufour@linux.vnet.ibm.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-27 06:07:11 +08:00
|
|
|
return __vm_munmap(addr, len, true);
|
2012-04-21 07:20:01 +08:00
|
|
|
}
|
2005-04-17 06:20:36 +08:00
|
|
|
|
mm: replace remap_file_pages() syscall with emulation
remap_file_pages(2) was invented to be able efficiently map parts of
huge file into limited 32-bit virtual address space such as in database
workloads.
Nonlinear mappings are pain to support and it seems there's no
legitimate use-cases nowadays since 64-bit systems are widely available.
Let's drop it and get rid of all these special-cased code.
The patch replaces the syscall with emulation which creates new VMA on
each remap_file_pages(), unless they it can be merged with an adjacent
one.
I didn't find *any* real code that uses remap_file_pages(2) to test
emulation impact on. I've checked Debian code search and source of all
packages in ALT Linux. No real users: libc wrappers, mentions in
strace, gdb, valgrind and this kind of stuff.
There are few basic tests in LTP for the syscall. They work just fine
with emulation.
To test performance impact, I've written small test case which
demonstrate pretty much worst case scenario: map 4G shmfs file, write to
begin of every page pgoff of the page, remap pages in reverse order,
read every page.
The test creates 1 million of VMAs if emulation is in use, so I had to
set vm.max_map_count to 1100000 to avoid -ENOMEM.
Before: 23.3 ( +- 4.31% ) seconds
After: 43.9 ( +- 0.85% ) seconds
Slowdown: 1.88x
I believe we can live with that.
Test case:
#define _GNU_SOURCE
#include <assert.h>
#include <stdlib.h>
#include <stdio.h>
#include <sys/mman.h>
#define MB (1024UL * 1024)
#define SIZE (4096 * MB)
int main(int argc, char **argv)
{
unsigned long *p;
long i, pass;
for (pass = 0; pass < 10; pass++) {
p = mmap(NULL, SIZE, PROT_READ|PROT_WRITE,
MAP_SHARED | MAP_ANONYMOUS, -1, 0);
if (p == MAP_FAILED) {
perror("mmap");
return -1;
}
for (i = 0; i < SIZE / 4096; i++)
p[i * 4096 / sizeof(*p)] = i;
for (i = 0; i < SIZE / 4096; i++) {
if (remap_file_pages(p + i * 4096 / sizeof(*p), 4096,
0, (SIZE - 4096 * (i + 1)) >> 12, 0)) {
perror("remap_file_pages");
return -1;
}
}
for (i = SIZE / 4096 - 1; i >= 0; i--)
assert(p[i * 4096 / sizeof(*p)] == SIZE / 4096 - i - 1);
munmap(p, SIZE);
}
return 0;
}
[akpm@linux-foundation.org: fix spello]
[sasha.levin@oracle.com: initialize populate before usage]
[sasha.levin@oracle.com: grab file ref to prevent race while mmaping]
Signed-off-by: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Dave Jones <davej@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Armin Rigo <arigo@tunes.org>
Signed-off-by: Sasha Levin <sasha.levin@oracle.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 06:09:46 +08:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Emulation of deprecated remap_file_pages() syscall.
|
|
|
|
|
*/
|
|
|
|
|
SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
|
|
|
|
|
unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
|
|
|
|
|
{
|
|
|
|
|
|
|
|
|
|
struct mm_struct *mm = current->mm;
|
|
|
|
|
struct vm_area_struct *vma;
|
|
|
|
|
unsigned long populate = 0;
|
|
|
|
|
unsigned long ret = -EINVAL;
|
|
|
|
|
struct file *file;
|
|
|
|
|
|
2018-03-22 03:22:47 +08:00
|
|
|
pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.rst.\n",
|
2016-03-18 05:19:47 +08:00
|
|
|
current->comm, current->pid);
|
mm: replace remap_file_pages() syscall with emulation
remap_file_pages(2) was invented to be able efficiently map parts of
huge file into limited 32-bit virtual address space such as in database
workloads.
Nonlinear mappings are pain to support and it seems there's no
legitimate use-cases nowadays since 64-bit systems are widely available.
Let's drop it and get rid of all these special-cased code.
The patch replaces the syscall with emulation which creates new VMA on
each remap_file_pages(), unless they it can be merged with an adjacent
one.
I didn't find *any* real code that uses remap_file_pages(2) to test
emulation impact on. I've checked Debian code search and source of all
packages in ALT Linux. No real users: libc wrappers, mentions in
strace, gdb, valgrind and this kind of stuff.
There are few basic tests in LTP for the syscall. They work just fine
with emulation.
To test performance impact, I've written small test case which
demonstrate pretty much worst case scenario: map 4G shmfs file, write to
begin of every page pgoff of the page, remap pages in reverse order,
read every page.
The test creates 1 million of VMAs if emulation is in use, so I had to
set vm.max_map_count to 1100000 to avoid -ENOMEM.
Before: 23.3 ( +- 4.31% ) seconds
After: 43.9 ( +- 0.85% ) seconds
Slowdown: 1.88x
I believe we can live with that.
Test case:
#define _GNU_SOURCE
#include <assert.h>
#include <stdlib.h>
#include <stdio.h>
#include <sys/mman.h>
#define MB (1024UL * 1024)
#define SIZE (4096 * MB)
int main(int argc, char **argv)
{
unsigned long *p;
long i, pass;
for (pass = 0; pass < 10; pass++) {
p = mmap(NULL, SIZE, PROT_READ|PROT_WRITE,
MAP_SHARED | MAP_ANONYMOUS, -1, 0);
if (p == MAP_FAILED) {
perror("mmap");
return -1;
}
for (i = 0; i < SIZE / 4096; i++)
p[i * 4096 / sizeof(*p)] = i;
for (i = 0; i < SIZE / 4096; i++) {
if (remap_file_pages(p + i * 4096 / sizeof(*p), 4096,
0, (SIZE - 4096 * (i + 1)) >> 12, 0)) {
perror("remap_file_pages");
return -1;
}
}
for (i = SIZE / 4096 - 1; i >= 0; i--)
assert(p[i * 4096 / sizeof(*p)] == SIZE / 4096 - i - 1);
munmap(p, SIZE);
}
return 0;
}
[akpm@linux-foundation.org: fix spello]
[sasha.levin@oracle.com: initialize populate before usage]
[sasha.levin@oracle.com: grab file ref to prevent race while mmaping]
Signed-off-by: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Dave Jones <davej@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Armin Rigo <arigo@tunes.org>
Signed-off-by: Sasha Levin <sasha.levin@oracle.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 06:09:46 +08:00
|
|
|
|
|
|
|
|
if (prot)
|
|
|
|
|
return ret;
|
|
|
|
|
start = start & PAGE_MASK;
|
|
|
|
|
size = size & PAGE_MASK;
|
|
|
|
|
|
|
|
|
|
if (start + size <= start)
|
|
|
|
|
return ret;
|
|
|
|
|
|
|
|
|
|
/* Does pgoff wrap? */
|
|
|
|
|
if (pgoff + (size >> PAGE_SHIFT) < pgoff)
|
|
|
|
|
return ret;
|
|
|
|
|
|
2020-06-09 12:33:25 +08:00
|
|
|
if (mmap_write_lock_killable(mm))
|
2016-05-24 07:25:27 +08:00
|
|
|
return -EINTR;
|
|
|
|
|
|
2021-09-03 05:56:49 +08:00
|
|
|
vma = vma_lookup(mm, start);
|
mm: replace remap_file_pages() syscall with emulation
remap_file_pages(2) was invented to be able efficiently map parts of
huge file into limited 32-bit virtual address space such as in database
workloads.
Nonlinear mappings are pain to support and it seems there's no
legitimate use-cases nowadays since 64-bit systems are widely available.
Let's drop it and get rid of all these special-cased code.
The patch replaces the syscall with emulation which creates new VMA on
each remap_file_pages(), unless they it can be merged with an adjacent
one.
I didn't find *any* real code that uses remap_file_pages(2) to test
emulation impact on. I've checked Debian code search and source of all
packages in ALT Linux. No real users: libc wrappers, mentions in
strace, gdb, valgrind and this kind of stuff.
There are few basic tests in LTP for the syscall. They work just fine
with emulation.
To test performance impact, I've written small test case which
demonstrate pretty much worst case scenario: map 4G shmfs file, write to
begin of every page pgoff of the page, remap pages in reverse order,
read every page.
The test creates 1 million of VMAs if emulation is in use, so I had to
set vm.max_map_count to 1100000 to avoid -ENOMEM.
Before: 23.3 ( +- 4.31% ) seconds
After: 43.9 ( +- 0.85% ) seconds
Slowdown: 1.88x
I believe we can live with that.
Test case:
#define _GNU_SOURCE
#include <assert.h>
#include <stdlib.h>
#include <stdio.h>
#include <sys/mman.h>
#define MB (1024UL * 1024)
#define SIZE (4096 * MB)
int main(int argc, char **argv)
{
unsigned long *p;
long i, pass;
for (pass = 0; pass < 10; pass++) {
p = mmap(NULL, SIZE, PROT_READ|PROT_WRITE,
MAP_SHARED | MAP_ANONYMOUS, -1, 0);
if (p == MAP_FAILED) {
perror("mmap");
return -1;
}
for (i = 0; i < SIZE / 4096; i++)
p[i * 4096 / sizeof(*p)] = i;
for (i = 0; i < SIZE / 4096; i++) {
if (remap_file_pages(p + i * 4096 / sizeof(*p), 4096,
0, (SIZE - 4096 * (i + 1)) >> 12, 0)) {
perror("remap_file_pages");
return -1;
}
}
for (i = SIZE / 4096 - 1; i >= 0; i--)
assert(p[i * 4096 / sizeof(*p)] == SIZE / 4096 - i - 1);
munmap(p, SIZE);
}
return 0;
}
[akpm@linux-foundation.org: fix spello]
[sasha.levin@oracle.com: initialize populate before usage]
[sasha.levin@oracle.com: grab file ref to prevent race while mmaping]
Signed-off-by: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Dave Jones <davej@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Armin Rigo <arigo@tunes.org>
Signed-off-by: Sasha Levin <sasha.levin@oracle.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 06:09:46 +08:00
|
|
|
|
|
|
|
|
if (!vma || !(vma->vm_flags & VM_SHARED))
|
|
|
|
|
goto out;
|
|
|
|
|
|
mm: fix regression in remap_file_pages() emulation
Grazvydas Ignotas has reported a regression in remap_file_pages()
emulation.
Testcase:
#define _GNU_SOURCE
#include <assert.h>
#include <stdlib.h>
#include <stdio.h>
#include <sys/mman.h>
#define SIZE (4096 * 3)
int main(int argc, char **argv)
{
unsigned long *p;
long i;
p = mmap(NULL, SIZE, PROT_READ | PROT_WRITE,
MAP_SHARED | MAP_ANONYMOUS, -1, 0);
if (p == MAP_FAILED) {
perror("mmap");
return -1;
}
for (i = 0; i < SIZE / 4096; i++)
p[i * 4096 / sizeof(*p)] = i;
if (remap_file_pages(p, 4096, 0, 1, 0)) {
perror("remap_file_pages");
return -1;
}
if (remap_file_pages(p, 4096 * 2, 0, 1, 0)) {
perror("remap_file_pages");
return -1;
}
assert(p[0] == 1);
munmap(p, SIZE);
return 0;
}
The second remap_file_pages() fails with -EINVAL.
The reason is that remap_file_pages() emulation assumes that the target
vma covers whole area we want to over map. That assumption is broken by
first remap_file_pages() call: it split the area into two vma.
The solution is to check next adjacent vmas, if they map the same file
with the same flags.
Fixes: c8d78c1823f4 ("mm: replace remap_file_pages() syscall with emulation")
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Reported-by: Grazvydas Ignotas <notasas@gmail.com>
Tested-by: Grazvydas Ignotas <notasas@gmail.com>
Cc: <stable@vger.kernel.org> [4.0+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-02-18 05:11:15 +08:00
|
|
|
if (start + size > vma->vm_end) {
|
|
|
|
|
struct vm_area_struct *next;
|
|
|
|
|
|
|
|
|
|
for (next = vma->vm_next; next; next = next->vm_next) {
|
|
|
|
|
/* hole between vmas ? */
|
|
|
|
|
if (next->vm_start != next->vm_prev->vm_end)
|
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
|
|
if (next->vm_file != vma->vm_file)
|
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
|
|
if (next->vm_flags != vma->vm_flags)
|
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
|
|
if (start + size <= next->vm_end)
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (!next)
|
|
|
|
|
goto out;
|
mm: replace remap_file_pages() syscall with emulation
remap_file_pages(2) was invented to be able efficiently map parts of
huge file into limited 32-bit virtual address space such as in database
workloads.
Nonlinear mappings are pain to support and it seems there's no
legitimate use-cases nowadays since 64-bit systems are widely available.
Let's drop it and get rid of all these special-cased code.
The patch replaces the syscall with emulation which creates new VMA on
each remap_file_pages(), unless they it can be merged with an adjacent
one.
I didn't find *any* real code that uses remap_file_pages(2) to test
emulation impact on. I've checked Debian code search and source of all
packages in ALT Linux. No real users: libc wrappers, mentions in
strace, gdb, valgrind and this kind of stuff.
There are few basic tests in LTP for the syscall. They work just fine
with emulation.
To test performance impact, I've written small test case which
demonstrate pretty much worst case scenario: map 4G shmfs file, write to
begin of every page pgoff of the page, remap pages in reverse order,
read every page.
The test creates 1 million of VMAs if emulation is in use, so I had to
set vm.max_map_count to 1100000 to avoid -ENOMEM.
Before: 23.3 ( +- 4.31% ) seconds
After: 43.9 ( +- 0.85% ) seconds
Slowdown: 1.88x
I believe we can live with that.
Test case:
#define _GNU_SOURCE
#include <assert.h>
#include <stdlib.h>
#include <stdio.h>
#include <sys/mman.h>
#define MB (1024UL * 1024)
#define SIZE (4096 * MB)
int main(int argc, char **argv)
{
unsigned long *p;
long i, pass;
for (pass = 0; pass < 10; pass++) {
p = mmap(NULL, SIZE, PROT_READ|PROT_WRITE,
MAP_SHARED | MAP_ANONYMOUS, -1, 0);
if (p == MAP_FAILED) {
perror("mmap");
return -1;
}
for (i = 0; i < SIZE / 4096; i++)
p[i * 4096 / sizeof(*p)] = i;
for (i = 0; i < SIZE / 4096; i++) {
if (remap_file_pages(p + i * 4096 / sizeof(*p), 4096,
0, (SIZE - 4096 * (i + 1)) >> 12, 0)) {
perror("remap_file_pages");
return -1;
}
}
for (i = SIZE / 4096 - 1; i >= 0; i--)
assert(p[i * 4096 / sizeof(*p)] == SIZE / 4096 - i - 1);
munmap(p, SIZE);
}
return 0;
}
[akpm@linux-foundation.org: fix spello]
[sasha.levin@oracle.com: initialize populate before usage]
[sasha.levin@oracle.com: grab file ref to prevent race while mmaping]
Signed-off-by: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Dave Jones <davej@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Armin Rigo <arigo@tunes.org>
Signed-off-by: Sasha Levin <sasha.levin@oracle.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 06:09:46 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
|
|
|
|
|
prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
|
|
|
|
|
prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
|
|
|
|
|
|
|
|
|
|
flags &= MAP_NONBLOCK;
|
|
|
|
|
flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
|
2021-05-05 09:38:06 +08:00
|
|
|
if (vma->vm_flags & VM_LOCKED)
|
mm: replace remap_file_pages() syscall with emulation
remap_file_pages(2) was invented to be able efficiently map parts of
huge file into limited 32-bit virtual address space such as in database
workloads.
Nonlinear mappings are pain to support and it seems there's no
legitimate use-cases nowadays since 64-bit systems are widely available.
Let's drop it and get rid of all these special-cased code.
The patch replaces the syscall with emulation which creates new VMA on
each remap_file_pages(), unless they it can be merged with an adjacent
one.
I didn't find *any* real code that uses remap_file_pages(2) to test
emulation impact on. I've checked Debian code search and source of all
packages in ALT Linux. No real users: libc wrappers, mentions in
strace, gdb, valgrind and this kind of stuff.
There are few basic tests in LTP for the syscall. They work just fine
with emulation.
To test performance impact, I've written small test case which
demonstrate pretty much worst case scenario: map 4G shmfs file, write to
begin of every page pgoff of the page, remap pages in reverse order,
read every page.
The test creates 1 million of VMAs if emulation is in use, so I had to
set vm.max_map_count to 1100000 to avoid -ENOMEM.
Before: 23.3 ( +- 4.31% ) seconds
After: 43.9 ( +- 0.85% ) seconds
Slowdown: 1.88x
I believe we can live with that.
Test case:
#define _GNU_SOURCE
#include <assert.h>
#include <stdlib.h>
#include <stdio.h>
#include <sys/mman.h>
#define MB (1024UL * 1024)
#define SIZE (4096 * MB)
int main(int argc, char **argv)
{
unsigned long *p;
long i, pass;
for (pass = 0; pass < 10; pass++) {
p = mmap(NULL, SIZE, PROT_READ|PROT_WRITE,
MAP_SHARED | MAP_ANONYMOUS, -1, 0);
if (p == MAP_FAILED) {
perror("mmap");
return -1;
}
for (i = 0; i < SIZE / 4096; i++)
p[i * 4096 / sizeof(*p)] = i;
for (i = 0; i < SIZE / 4096; i++) {
if (remap_file_pages(p + i * 4096 / sizeof(*p), 4096,
0, (SIZE - 4096 * (i + 1)) >> 12, 0)) {
perror("remap_file_pages");
return -1;
}
}
for (i = SIZE / 4096 - 1; i >= 0; i--)
assert(p[i * 4096 / sizeof(*p)] == SIZE / 4096 - i - 1);
munmap(p, SIZE);
}
return 0;
}
[akpm@linux-foundation.org: fix spello]
[sasha.levin@oracle.com: initialize populate before usage]
[sasha.levin@oracle.com: grab file ref to prevent race while mmaping]
Signed-off-by: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Dave Jones <davej@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Armin Rigo <arigo@tunes.org>
Signed-off-by: Sasha Levin <sasha.levin@oracle.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 06:09:46 +08:00
|
|
|
flags |= MAP_LOCKED;
|
mm: fix regression in remap_file_pages() emulation
Grazvydas Ignotas has reported a regression in remap_file_pages()
emulation.
Testcase:
#define _GNU_SOURCE
#include <assert.h>
#include <stdlib.h>
#include <stdio.h>
#include <sys/mman.h>
#define SIZE (4096 * 3)
int main(int argc, char **argv)
{
unsigned long *p;
long i;
p = mmap(NULL, SIZE, PROT_READ | PROT_WRITE,
MAP_SHARED | MAP_ANONYMOUS, -1, 0);
if (p == MAP_FAILED) {
perror("mmap");
return -1;
}
for (i = 0; i < SIZE / 4096; i++)
p[i * 4096 / sizeof(*p)] = i;
if (remap_file_pages(p, 4096, 0, 1, 0)) {
perror("remap_file_pages");
return -1;
}
if (remap_file_pages(p, 4096 * 2, 0, 1, 0)) {
perror("remap_file_pages");
return -1;
}
assert(p[0] == 1);
munmap(p, SIZE);
return 0;
}
The second remap_file_pages() fails with -EINVAL.
The reason is that remap_file_pages() emulation assumes that the target
vma covers whole area we want to over map. That assumption is broken by
first remap_file_pages() call: it split the area into two vma.
The solution is to check next adjacent vmas, if they map the same file
with the same flags.
Fixes: c8d78c1823f4 ("mm: replace remap_file_pages() syscall with emulation")
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Reported-by: Grazvydas Ignotas <notasas@gmail.com>
Tested-by: Grazvydas Ignotas <notasas@gmail.com>
Cc: <stable@vger.kernel.org> [4.0+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-02-18 05:11:15 +08:00
|
|
|
|
mm: replace remap_file_pages() syscall with emulation
remap_file_pages(2) was invented to be able efficiently map parts of
huge file into limited 32-bit virtual address space such as in database
workloads.
Nonlinear mappings are pain to support and it seems there's no
legitimate use-cases nowadays since 64-bit systems are widely available.
Let's drop it and get rid of all these special-cased code.
The patch replaces the syscall with emulation which creates new VMA on
each remap_file_pages(), unless they it can be merged with an adjacent
one.
I didn't find *any* real code that uses remap_file_pages(2) to test
emulation impact on. I've checked Debian code search and source of all
packages in ALT Linux. No real users: libc wrappers, mentions in
strace, gdb, valgrind and this kind of stuff.
There are few basic tests in LTP for the syscall. They work just fine
with emulation.
To test performance impact, I've written small test case which
demonstrate pretty much worst case scenario: map 4G shmfs file, write to
begin of every page pgoff of the page, remap pages in reverse order,
read every page.
The test creates 1 million of VMAs if emulation is in use, so I had to
set vm.max_map_count to 1100000 to avoid -ENOMEM.
Before: 23.3 ( +- 4.31% ) seconds
After: 43.9 ( +- 0.85% ) seconds
Slowdown: 1.88x
I believe we can live with that.
Test case:
#define _GNU_SOURCE
#include <assert.h>
#include <stdlib.h>
#include <stdio.h>
#include <sys/mman.h>
#define MB (1024UL * 1024)
#define SIZE (4096 * MB)
int main(int argc, char **argv)
{
unsigned long *p;
long i, pass;
for (pass = 0; pass < 10; pass++) {
p = mmap(NULL, SIZE, PROT_READ|PROT_WRITE,
MAP_SHARED | MAP_ANONYMOUS, -1, 0);
if (p == MAP_FAILED) {
perror("mmap");
return -1;
}
for (i = 0; i < SIZE / 4096; i++)
p[i * 4096 / sizeof(*p)] = i;
for (i = 0; i < SIZE / 4096; i++) {
if (remap_file_pages(p + i * 4096 / sizeof(*p), 4096,
0, (SIZE - 4096 * (i + 1)) >> 12, 0)) {
perror("remap_file_pages");
return -1;
}
}
for (i = SIZE / 4096 - 1; i >= 0; i--)
assert(p[i * 4096 / sizeof(*p)] == SIZE / 4096 - i - 1);
munmap(p, SIZE);
}
return 0;
}
[akpm@linux-foundation.org: fix spello]
[sasha.levin@oracle.com: initialize populate before usage]
[sasha.levin@oracle.com: grab file ref to prevent race while mmaping]
Signed-off-by: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Dave Jones <davej@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Armin Rigo <arigo@tunes.org>
Signed-off-by: Sasha Levin <sasha.levin@oracle.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 06:09:46 +08:00
|
|
|
file = get_file(vma->vm_file);
|
2020-08-07 14:23:37 +08:00
|
|
|
ret = do_mmap(vma->vm_file, start, size,
|
2017-02-25 06:58:22 +08:00
|
|
|
prot, flags, pgoff, &populate, NULL);
|
mm: replace remap_file_pages() syscall with emulation
remap_file_pages(2) was invented to be able efficiently map parts of
huge file into limited 32-bit virtual address space such as in database
workloads.
Nonlinear mappings are pain to support and it seems there's no
legitimate use-cases nowadays since 64-bit systems are widely available.
Let's drop it and get rid of all these special-cased code.
The patch replaces the syscall with emulation which creates new VMA on
each remap_file_pages(), unless they it can be merged with an adjacent
one.
I didn't find *any* real code that uses remap_file_pages(2) to test
emulation impact on. I've checked Debian code search and source of all
packages in ALT Linux. No real users: libc wrappers, mentions in
strace, gdb, valgrind and this kind of stuff.
There are few basic tests in LTP for the syscall. They work just fine
with emulation.
To test performance impact, I've written small test case which
demonstrate pretty much worst case scenario: map 4G shmfs file, write to
begin of every page pgoff of the page, remap pages in reverse order,
read every page.
The test creates 1 million of VMAs if emulation is in use, so I had to
set vm.max_map_count to 1100000 to avoid -ENOMEM.
Before: 23.3 ( +- 4.31% ) seconds
After: 43.9 ( +- 0.85% ) seconds
Slowdown: 1.88x
I believe we can live with that.
Test case:
#define _GNU_SOURCE
#include <assert.h>
#include <stdlib.h>
#include <stdio.h>
#include <sys/mman.h>
#define MB (1024UL * 1024)
#define SIZE (4096 * MB)
int main(int argc, char **argv)
{
unsigned long *p;
long i, pass;
for (pass = 0; pass < 10; pass++) {
p = mmap(NULL, SIZE, PROT_READ|PROT_WRITE,
MAP_SHARED | MAP_ANONYMOUS, -1, 0);
if (p == MAP_FAILED) {
perror("mmap");
return -1;
}
for (i = 0; i < SIZE / 4096; i++)
p[i * 4096 / sizeof(*p)] = i;
for (i = 0; i < SIZE / 4096; i++) {
if (remap_file_pages(p + i * 4096 / sizeof(*p), 4096,
0, (SIZE - 4096 * (i + 1)) >> 12, 0)) {
perror("remap_file_pages");
return -1;
}
}
for (i = SIZE / 4096 - 1; i >= 0; i--)
assert(p[i * 4096 / sizeof(*p)] == SIZE / 4096 - i - 1);
munmap(p, SIZE);
}
return 0;
}
[akpm@linux-foundation.org: fix spello]
[sasha.levin@oracle.com: initialize populate before usage]
[sasha.levin@oracle.com: grab file ref to prevent race while mmaping]
Signed-off-by: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Dave Jones <davej@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Armin Rigo <arigo@tunes.org>
Signed-off-by: Sasha Levin <sasha.levin@oracle.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 06:09:46 +08:00
|
|
|
fput(file);
|
|
|
|
|
out:
|
2020-06-09 12:33:25 +08:00
|
|
|
mmap_write_unlock(mm);
|
mm: replace remap_file_pages() syscall with emulation
remap_file_pages(2) was invented to be able efficiently map parts of
huge file into limited 32-bit virtual address space such as in database
workloads.
Nonlinear mappings are pain to support and it seems there's no
legitimate use-cases nowadays since 64-bit systems are widely available.
Let's drop it and get rid of all these special-cased code.
The patch replaces the syscall with emulation which creates new VMA on
each remap_file_pages(), unless they it can be merged with an adjacent
one.
I didn't find *any* real code that uses remap_file_pages(2) to test
emulation impact on. I've checked Debian code search and source of all
packages in ALT Linux. No real users: libc wrappers, mentions in
strace, gdb, valgrind and this kind of stuff.
There are few basic tests in LTP for the syscall. They work just fine
with emulation.
To test performance impact, I've written small test case which
demonstrate pretty much worst case scenario: map 4G shmfs file, write to
begin of every page pgoff of the page, remap pages in reverse order,
read every page.
The test creates 1 million of VMAs if emulation is in use, so I had to
set vm.max_map_count to 1100000 to avoid -ENOMEM.
Before: 23.3 ( +- 4.31% ) seconds
After: 43.9 ( +- 0.85% ) seconds
Slowdown: 1.88x
I believe we can live with that.
Test case:
#define _GNU_SOURCE
#include <assert.h>
#include <stdlib.h>
#include <stdio.h>
#include <sys/mman.h>
#define MB (1024UL * 1024)
#define SIZE (4096 * MB)
int main(int argc, char **argv)
{
unsigned long *p;
long i, pass;
for (pass = 0; pass < 10; pass++) {
p = mmap(NULL, SIZE, PROT_READ|PROT_WRITE,
MAP_SHARED | MAP_ANONYMOUS, -1, 0);
if (p == MAP_FAILED) {
perror("mmap");
return -1;
}
for (i = 0; i < SIZE / 4096; i++)
p[i * 4096 / sizeof(*p)] = i;
for (i = 0; i < SIZE / 4096; i++) {
if (remap_file_pages(p + i * 4096 / sizeof(*p), 4096,
0, (SIZE - 4096 * (i + 1)) >> 12, 0)) {
perror("remap_file_pages");
return -1;
}
}
for (i = SIZE / 4096 - 1; i >= 0; i--)
assert(p[i * 4096 / sizeof(*p)] == SIZE / 4096 - i - 1);
munmap(p, SIZE);
}
return 0;
}
[akpm@linux-foundation.org: fix spello]
[sasha.levin@oracle.com: initialize populate before usage]
[sasha.levin@oracle.com: grab file ref to prevent race while mmaping]
Signed-off-by: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Dave Jones <davej@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Armin Rigo <arigo@tunes.org>
Signed-off-by: Sasha Levin <sasha.levin@oracle.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 06:09:46 +08:00
|
|
|
if (populate)
|
|
|
|
|
mm_populate(ret, populate);
|
|
|
|
|
if (!IS_ERR_VALUE(ret))
|
|
|
|
|
ret = 0;
|
|
|
|
|
return ret;
|
|
|
|
|
}
|
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
/*
|
|
|
|
|
* this is really a simplified "do_mmap". it only handles
|
|
|
|
|
* anonymous maps. eventually we may be able to do some
|
|
|
|
|
* brk-specific accounting here.
|
|
|
|
|
*/
|
2018-07-14 07:59:20 +08:00
|
|
|
static int do_brk_flags(unsigned long addr, unsigned long len, unsigned long flags, struct list_head *uf)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
2014-10-10 06:26:29 +08:00
|
|
|
struct mm_struct *mm = current->mm;
|
|
|
|
|
struct vm_area_struct *vma, *prev;
|
|
|
|
|
struct rb_node **rb_link, *rb_parent;
|
2005-04-17 06:20:36 +08:00
|
|
|
pgoff_t pgoff = addr >> PAGE_SHIFT;
|
2006-09-07 18:17:04 +08:00
|
|
|
int error;
|
2019-12-01 09:51:03 +08:00
|
|
|
unsigned long mapped_addr;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
powerpc: do not make the entire heap executable
On 32-bit powerpc the ELF PLT sections of binaries (built with
--bss-plt, or with a toolchain which defaults to it) look like this:
[17] .sbss NOBITS 0002aff8 01aff8 000014 00 WA 0 0 4
[18] .plt NOBITS 0002b00c 01aff8 000084 00 WAX 0 0 4
[19] .bss NOBITS 0002b090 01aff8 0000a4 00 WA 0 0 4
Which results in an ELF load header:
Type Offset VirtAddr PhysAddr FileSiz MemSiz Flg Align
LOAD 0x019c70 0x00029c70 0x00029c70 0x01388 0x014c4 RWE 0x10000
This is all correct, the load region containing the PLT is marked as
executable. Note that the PLT starts at 0002b00c but the file mapping
ends at 0002aff8, so the PLT falls in the 0 fill section described by
the load header, and after a page boundary.
Unfortunately the generic ELF loader ignores the X bit in the load
headers when it creates the 0 filled non-file backed mappings. It
assumes all of these mappings are RW BSS sections, which is not the case
for PPC.
gcc/ld has an option (--secure-plt) to not do this, this is said to
incur a small performance penalty.
Currently, to support 32-bit binaries with PLT in BSS kernel maps
*entire brk area* with executable rights for all binaries, even
--secure-plt ones.
Stop doing that.
Teach the ELF loader to check the X bit in the relevant load header and
create 0 filled anonymous mappings that are executable if the load
header requests that.
Test program showing the difference in /proc/$PID/maps:
int main() {
char buf[16*1024];
char *p = malloc(123); /* make "[heap]" mapping appear */
int fd = open("/proc/self/maps", O_RDONLY);
int len = read(fd, buf, sizeof(buf));
write(1, buf, len);
printf("%p\n", p);
return 0;
}
Compiled using: gcc -mbss-plt -m32 -Os test.c -otest
Unpatched ppc64 kernel:
00100000-00120000 r-xp 00000000 00:00 0 [vdso]
0fe10000-0ffd0000 r-xp 00000000 fd:00 67898094 /usr/lib/libc-2.17.so
0ffd0000-0ffe0000 r--p 001b0000 fd:00 67898094 /usr/lib/libc-2.17.so
0ffe0000-0fff0000 rw-p 001c0000 fd:00 67898094 /usr/lib/libc-2.17.so
10000000-10010000 r-xp 00000000 fd:00 100674505 /home/user/test
10010000-10020000 r--p 00000000 fd:00 100674505 /home/user/test
10020000-10030000 rw-p 00010000 fd:00 100674505 /home/user/test
10690000-106c0000 rwxp 00000000 00:00 0 [heap]
f7f70000-f7fa0000 r-xp 00000000 fd:00 67898089 /usr/lib/ld-2.17.so
f7fa0000-f7fb0000 r--p 00020000 fd:00 67898089 /usr/lib/ld-2.17.so
f7fb0000-f7fc0000 rw-p 00030000 fd:00 67898089 /usr/lib/ld-2.17.so
ffa90000-ffac0000 rw-p 00000000 00:00 0 [stack]
0x10690008
Patched ppc64 kernel:
00100000-00120000 r-xp 00000000 00:00 0 [vdso]
0fe10000-0ffd0000 r-xp 00000000 fd:00 67898094 /usr/lib/libc-2.17.so
0ffd0000-0ffe0000 r--p 001b0000 fd:00 67898094 /usr/lib/libc-2.17.so
0ffe0000-0fff0000 rw-p 001c0000 fd:00 67898094 /usr/lib/libc-2.17.so
10000000-10010000 r-xp 00000000 fd:00 100674505 /home/user/test
10010000-10020000 r--p 00000000 fd:00 100674505 /home/user/test
10020000-10030000 rw-p 00010000 fd:00 100674505 /home/user/test
10180000-101b0000 rw-p 00000000 00:00 0 [heap]
^^^^ this has changed
f7c60000-f7c90000 r-xp 00000000 fd:00 67898089 /usr/lib/ld-2.17.so
f7c90000-f7ca0000 r--p 00020000 fd:00 67898089 /usr/lib/ld-2.17.so
f7ca0000-f7cb0000 rw-p 00030000 fd:00 67898089 /usr/lib/ld-2.17.so
ff860000-ff890000 rw-p 00000000 00:00 0 [stack]
0x10180008
The patch was originally posted in 2012 by Jason Gunthorpe
and apparently ignored:
https://lkml.org/lkml/2012/9/30/138
Lightly run-tested.
Link: http://lkml.kernel.org/r/20161215131950.23054-1-dvlasenk@redhat.com
Signed-off-by: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
Signed-off-by: Denys Vlasenko <dvlasenk@redhat.com>
Acked-by: Kees Cook <keescook@chromium.org>
Acked-by: Michael Ellerman <mpe@ellerman.id.au>
Tested-by: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Florian Weimer <fweimer@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-23 07:45:16 +08:00
|
|
|
/* Until we need other flags, refuse anything except VM_EXEC. */
|
|
|
|
|
if ((flags & (~VM_EXEC)) != 0)
|
|
|
|
|
return -EINVAL;
|
|
|
|
|
flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
|
2006-09-07 18:17:04 +08:00
|
|
|
|
2019-12-01 09:51:03 +08:00
|
|
|
mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
|
|
|
|
|
if (IS_ERR_VALUE(mapped_addr))
|
|
|
|
|
return mapped_addr;
|
2006-09-07 18:17:04 +08:00
|
|
|
|
2014-01-22 07:49:15 +08:00
|
|
|
error = mlock_future_check(mm, mm->def_flags, len);
|
|
|
|
|
if (error)
|
|
|
|
|
return error;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2020-10-18 07:14:09 +08:00
|
|
|
/* Clear old maps, set up prev, rb_link, rb_parent, and uf */
|
|
|
|
|
if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf))
|
|
|
|
|
return -ENOMEM;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
|
/* Check against address space limits *after* clearing old maps... */
|
2016-01-15 07:22:07 +08:00
|
|
|
if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
|
2005-04-17 06:20:36 +08:00
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
|
|
if (mm->map_count > sysctl_max_map_count)
|
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
2012-02-13 11:58:52 +08:00
|
|
|
if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
|
2005-04-17 06:20:36 +08:00
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
|
|
/* Can we just expand an old private anonymous mapping? */
|
2008-10-19 11:26:50 +08:00
|
|
|
vma = vma_merge(mm, prev, addr, addr + len, flags,
|
mm: add a field to store names for private anonymous memory
In many userspace applications, and especially in VM based applications
like Android uses heavily, there are multiple different allocators in
use. At a minimum there is libc malloc and the stack, and in many cases
there are libc malloc, the stack, direct syscalls to mmap anonymous
memory, and multiple VM heaps (one for small objects, one for big
objects, etc.). Each of these layers usually has its own tools to
inspect its usage; malloc by compiling a debug version, the VM through
heap inspection tools, and for direct syscalls there is usually no way
to track them.
On Android we heavily use a set of tools that use an extended version of
the logic covered in Documentation/vm/pagemap.txt to walk all pages
mapped in userspace and slice their usage by process, shared (COW) vs.
unique mappings, backing, etc. This can account for real physical
memory usage even in cases like fork without exec (which Android uses
heavily to share as many private COW pages as possible between
processes), Kernel SamePage Merging, and clean zero pages. It produces
a measurement of the pages that only exist in that process (USS, for
unique), and a measurement of the physical memory usage of that process
with the cost of shared pages being evenly split between processes that
share them (PSS).
If all anonymous memory is indistinguishable then figuring out the real
physical memory usage (PSS) of each heap requires either a pagemap
walking tool that can understand the heap debugging of every layer, or
for every layer's heap debugging tools to implement the pagemap walking
logic, in which case it is hard to get a consistent view of memory
across the whole system.
Tracking the information in userspace leads to all sorts of problems.
It either needs to be stored inside the process, which means every
process has to have an API to export its current heap information upon
request, or it has to be stored externally in a filesystem that somebody
needs to clean up on crashes. It needs to be readable while the process
is still running, so it has to have some sort of synchronization with
every layer of userspace. Efficiently tracking the ranges requires
reimplementing something like the kernel vma trees, and linking to it
from every layer of userspace. It requires more memory, more syscalls,
more runtime cost, and more complexity to separately track regions that
the kernel is already tracking.
This patch adds a field to /proc/pid/maps and /proc/pid/smaps to show a
userspace-provided name for anonymous vmas. The names of named
anonymous vmas are shown in /proc/pid/maps and /proc/pid/smaps as
[anon:<name>].
Userspace can set the name for a region of memory by calling
prctl(PR_SET_VMA, PR_SET_VMA_ANON_NAME, start, len, (unsigned long)name)
Setting the name to NULL clears it. The name length limit is 80 bytes
including NUL-terminator and is checked to contain only printable ascii
characters (including space), except '[',']','\','$' and '`'.
Ascii strings are being used to have a descriptive identifiers for vmas,
which can be understood by the users reading /proc/pid/maps or
/proc/pid/smaps. Names can be standardized for a given system and they
can include some variable parts such as the name of the allocator or a
library, tid of the thread using it, etc.
The name is stored in a pointer in the shared union in vm_area_struct
that points to a null terminated string. Anonymous vmas with the same
name (equivalent strings) and are otherwise mergeable will be merged.
The name pointers are not shared between vmas even if they contain the
same name. The name pointer is stored in a union with fields that are
only used on file-backed mappings, so it does not increase memory usage.
CONFIG_ANON_VMA_NAME kernel configuration is introduced to enable this
feature. It keeps the feature disabled by default to prevent any
additional memory overhead and to avoid confusing procfs parsers on
systems which are not ready to support named anonymous vmas.
The patch is based on the original patch developed by Colin Cross, more
specifically on its latest version [1] posted upstream by Sumit Semwal.
It used a userspace pointer to store vma names. In that design, name
pointers could be shared between vmas. However during the last
upstreaming attempt, Kees Cook raised concerns [2] about this approach
and suggested to copy the name into kernel memory space, perform
validity checks [3] and store as a string referenced from
vm_area_struct.
One big concern is about fork() performance which would need to strdup
anonymous vma names. Dave Hansen suggested experimenting with
worst-case scenario of forking a process with 64k vmas having longest
possible names [4]. I ran this experiment on an ARM64 Android device
and recorded a worst-case regression of almost 40% when forking such a
process.
This regression is addressed in the followup patch which replaces the
pointer to a name with a refcounted structure that allows sharing the
name pointer between vmas of the same name. Instead of duplicating the
string during fork() or when splitting a vma it increments the refcount.
[1] https://lore.kernel.org/linux-mm/20200901161459.11772-4-sumit.semwal@linaro.org/
[2] https://lore.kernel.org/linux-mm/202009031031.D32EF57ED@keescook/
[3] https://lore.kernel.org/linux-mm/202009031022.3834F692@keescook/
[4] https://lore.kernel.org/linux-mm/5d0358ab-8c47-2f5f-8e43-23b89d6a8e95@intel.com/
Changes for prctl(2) manual page (in the options section):
PR_SET_VMA
Sets an attribute specified in arg2 for virtual memory areas
starting from the address specified in arg3 and spanning the
size specified in arg4. arg5 specifies the value of the attribute
to be set. Note that assigning an attribute to a virtual memory
area might prevent it from being merged with adjacent virtual
memory areas due to the difference in that attribute's value.
Currently, arg2 must be one of:
PR_SET_VMA_ANON_NAME
Set a name for anonymous virtual memory areas. arg5 should
be a pointer to a null-terminated string containing the
name. The name length including null byte cannot exceed
80 bytes. If arg5 is NULL, the name of the appropriate
anonymous virtual memory areas will be reset. The name
can contain only printable ascii characters (including
space), except '[',']','\','$' and '`'.
This feature is available only if the kernel is built with
the CONFIG_ANON_VMA_NAME option enabled.
[surenb@google.com: docs: proc.rst: /proc/PID/maps: fix malformed table]
Link: https://lkml.kernel.org/r/20211123185928.2513763-1-surenb@google.com
[surenb: rebased over v5.15-rc6, replaced userpointer with a kernel copy,
added input sanitization and CONFIG_ANON_VMA_NAME config. The bulk of the
work here was done by Colin Cross, therefore, with his permission, keeping
him as the author]
Link: https://lkml.kernel.org/r/20211019215511.3771969-2-surenb@google.com
Signed-off-by: Colin Cross <ccross@google.com>
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Kees Cook <keescook@chromium.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Jan Glauber <jan.glauber@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: John Stultz <john.stultz@linaro.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rob Landley <rob@landley.net>
Cc: "Serge E. Hallyn" <serge.hallyn@ubuntu.com>
Cc: Shaohua Li <shli@fusionio.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-15 06:05:59 +08:00
|
|
|
NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX, NULL);
|
2008-10-19 11:26:50 +08:00
|
|
|
if (vma)
|
2005-04-17 06:20:36 +08:00
|
|
|
goto out;
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* create a vma struct for an anonymous mapping
|
|
|
|
|
*/
|
2018-07-22 06:24:03 +08:00
|
|
|
vma = vm_area_alloc(mm);
|
2005-04-17 06:20:36 +08:00
|
|
|
if (!vma) {
|
|
|
|
|
vm_unacct_memory(len >> PAGE_SHIFT);
|
|
|
|
|
return -ENOMEM;
|
|
|
|
|
}
|
|
|
|
|
|
mm: fix vma_is_anonymous() false-positives
vma_is_anonymous() relies on ->vm_ops being NULL to detect anonymous
VMA. This is unreliable as ->mmap may not set ->vm_ops.
False-positive vma_is_anonymous() may lead to crashes:
next ffff8801ce5e7040 prev ffff8801d20eca50 mm ffff88019c1e13c0
prot 27 anon_vma ffff88019680cdd8 vm_ops 0000000000000000
pgoff 0 file ffff8801b2ec2d00 private_data 0000000000000000
flags: 0xff(read|write|exec|shared|mayread|maywrite|mayexec|mayshare)
------------[ cut here ]------------
kernel BUG at mm/memory.c:1422!
invalid opcode: 0000 [#1] SMP KASAN
CPU: 0 PID: 18486 Comm: syz-executor3 Not tainted 4.18.0-rc3+ #136
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google
01/01/2011
RIP: 0010:zap_pmd_range mm/memory.c:1421 [inline]
RIP: 0010:zap_pud_range mm/memory.c:1466 [inline]
RIP: 0010:zap_p4d_range mm/memory.c:1487 [inline]
RIP: 0010:unmap_page_range+0x1c18/0x2220 mm/memory.c:1508
Call Trace:
unmap_single_vma+0x1a0/0x310 mm/memory.c:1553
zap_page_range_single+0x3cc/0x580 mm/memory.c:1644
unmap_mapping_range_vma mm/memory.c:2792 [inline]
unmap_mapping_range_tree mm/memory.c:2813 [inline]
unmap_mapping_pages+0x3a7/0x5b0 mm/memory.c:2845
unmap_mapping_range+0x48/0x60 mm/memory.c:2880
truncate_pagecache+0x54/0x90 mm/truncate.c:800
truncate_setsize+0x70/0xb0 mm/truncate.c:826
simple_setattr+0xe9/0x110 fs/libfs.c:409
notify_change+0xf13/0x10f0 fs/attr.c:335
do_truncate+0x1ac/0x2b0 fs/open.c:63
do_sys_ftruncate+0x492/0x560 fs/open.c:205
__do_sys_ftruncate fs/open.c:215 [inline]
__se_sys_ftruncate fs/open.c:213 [inline]
__x64_sys_ftruncate+0x59/0x80 fs/open.c:213
do_syscall_64+0x1b9/0x820 arch/x86/entry/common.c:290
entry_SYSCALL_64_after_hwframe+0x49/0xbe
Reproducer:
#include <stdio.h>
#include <stddef.h>
#include <stdint.h>
#include <stdlib.h>
#include <string.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <sys/ioctl.h>
#include <sys/mman.h>
#include <unistd.h>
#include <fcntl.h>
#define KCOV_INIT_TRACE _IOR('c', 1, unsigned long)
#define KCOV_ENABLE _IO('c', 100)
#define KCOV_DISABLE _IO('c', 101)
#define COVER_SIZE (1024<<10)
#define KCOV_TRACE_PC 0
#define KCOV_TRACE_CMP 1
int main(int argc, char **argv)
{
int fd;
unsigned long *cover;
system("mount -t debugfs none /sys/kernel/debug");
fd = open("/sys/kernel/debug/kcov", O_RDWR);
ioctl(fd, KCOV_INIT_TRACE, COVER_SIZE);
cover = mmap(NULL, COVER_SIZE * sizeof(unsigned long),
PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
munmap(cover, COVER_SIZE * sizeof(unsigned long));
cover = mmap(NULL, COVER_SIZE * sizeof(unsigned long),
PROT_READ | PROT_WRITE, MAP_PRIVATE, fd, 0);
memset(cover, 0, COVER_SIZE * sizeof(unsigned long));
ftruncate(fd, 3UL << 20);
return 0;
}
This can be fixed by assigning anonymous VMAs own vm_ops and not relying
on it being NULL.
If ->mmap() failed to set ->vm_ops, mmap_region() will set it to
dummy_vm_ops. This way we will have non-NULL ->vm_ops for all VMAs.
Link: http://lkml.kernel.org/r/20180724121139.62570-4-kirill.shutemov@linux.intel.com
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Reported-by: syzbot+3f84280d52be9b7083cc@syzkaller.appspotmail.com
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-07-27 07:37:35 +08:00
|
|
|
vma_set_anonymous(vma);
|
2005-04-17 06:20:36 +08:00
|
|
|
vma->vm_start = addr;
|
|
|
|
|
vma->vm_end = addr + len;
|
|
|
|
|
vma->vm_pgoff = pgoff;
|
|
|
|
|
vma->vm_flags = flags;
|
2007-10-19 14:39:15 +08:00
|
|
|
vma->vm_page_prot = vm_get_page_prot(flags);
|
2005-04-17 06:20:36 +08:00
|
|
|
vma_link(mm, vma, prev, rb_link, rb_parent);
|
|
|
|
|
out:
|
2010-05-18 22:30:49 +08:00
|
|
|
perf_event_mmap(vma);
|
2005-04-17 06:20:36 +08:00
|
|
|
mm->total_vm += len >> PAGE_SHIFT;
|
2016-01-15 07:22:07 +08:00
|
|
|
mm->data_vm += len >> PAGE_SHIFT;
|
2013-02-23 08:32:40 +08:00
|
|
|
if (flags & VM_LOCKED)
|
|
|
|
|
mm->locked_vm += (len >> PAGE_SHIFT);
|
2013-09-12 05:22:24 +08:00
|
|
|
vma->vm_flags |= VM_SOFTDIRTY;
|
2016-05-28 06:57:31 +08:00
|
|
|
return 0;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
|
2018-07-14 07:59:20 +08:00
|
|
|
int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
|
2012-04-21 06:35:40 +08:00
|
|
|
{
|
|
|
|
|
struct mm_struct *mm = current->mm;
|
2018-07-14 07:59:20 +08:00
|
|
|
unsigned long len;
|
2016-05-28 06:57:31 +08:00
|
|
|
int ret;
|
2013-02-23 08:32:40 +08:00
|
|
|
bool populate;
|
2017-02-25 06:58:22 +08:00
|
|
|
LIST_HEAD(uf);
|
2012-04-21 06:35:40 +08:00
|
|
|
|
2018-07-14 07:59:20 +08:00
|
|
|
len = PAGE_ALIGN(request);
|
|
|
|
|
if (len < request)
|
|
|
|
|
return -ENOMEM;
|
|
|
|
|
if (!len)
|
|
|
|
|
return 0;
|
|
|
|
|
|
2020-06-09 12:33:25 +08:00
|
|
|
if (mmap_write_lock_killable(mm))
|
2016-05-24 07:25:42 +08:00
|
|
|
return -EINTR;
|
|
|
|
|
|
2017-02-25 06:58:22 +08:00
|
|
|
ret = do_brk_flags(addr, len, flags, &uf);
|
2013-02-23 08:32:40 +08:00
|
|
|
populate = ((mm->def_flags & VM_LOCKED) != 0);
|
2020-06-09 12:33:25 +08:00
|
|
|
mmap_write_unlock(mm);
|
2017-02-25 06:58:22 +08:00
|
|
|
userfaultfd_unmap_complete(mm, &uf);
|
2016-05-28 06:57:31 +08:00
|
|
|
if (populate && !ret)
|
2013-02-23 08:32:40 +08:00
|
|
|
mm_populate(addr, len);
|
2012-04-21 06:35:40 +08:00
|
|
|
return ret;
|
|
|
|
|
}
|
powerpc: do not make the entire heap executable
On 32-bit powerpc the ELF PLT sections of binaries (built with
--bss-plt, or with a toolchain which defaults to it) look like this:
[17] .sbss NOBITS 0002aff8 01aff8 000014 00 WA 0 0 4
[18] .plt NOBITS 0002b00c 01aff8 000084 00 WAX 0 0 4
[19] .bss NOBITS 0002b090 01aff8 0000a4 00 WA 0 0 4
Which results in an ELF load header:
Type Offset VirtAddr PhysAddr FileSiz MemSiz Flg Align
LOAD 0x019c70 0x00029c70 0x00029c70 0x01388 0x014c4 RWE 0x10000
This is all correct, the load region containing the PLT is marked as
executable. Note that the PLT starts at 0002b00c but the file mapping
ends at 0002aff8, so the PLT falls in the 0 fill section described by
the load header, and after a page boundary.
Unfortunately the generic ELF loader ignores the X bit in the load
headers when it creates the 0 filled non-file backed mappings. It
assumes all of these mappings are RW BSS sections, which is not the case
for PPC.
gcc/ld has an option (--secure-plt) to not do this, this is said to
incur a small performance penalty.
Currently, to support 32-bit binaries with PLT in BSS kernel maps
*entire brk area* with executable rights for all binaries, even
--secure-plt ones.
Stop doing that.
Teach the ELF loader to check the X bit in the relevant load header and
create 0 filled anonymous mappings that are executable if the load
header requests that.
Test program showing the difference in /proc/$PID/maps:
int main() {
char buf[16*1024];
char *p = malloc(123); /* make "[heap]" mapping appear */
int fd = open("/proc/self/maps", O_RDONLY);
int len = read(fd, buf, sizeof(buf));
write(1, buf, len);
printf("%p\n", p);
return 0;
}
Compiled using: gcc -mbss-plt -m32 -Os test.c -otest
Unpatched ppc64 kernel:
00100000-00120000 r-xp 00000000 00:00 0 [vdso]
0fe10000-0ffd0000 r-xp 00000000 fd:00 67898094 /usr/lib/libc-2.17.so
0ffd0000-0ffe0000 r--p 001b0000 fd:00 67898094 /usr/lib/libc-2.17.so
0ffe0000-0fff0000 rw-p 001c0000 fd:00 67898094 /usr/lib/libc-2.17.so
10000000-10010000 r-xp 00000000 fd:00 100674505 /home/user/test
10010000-10020000 r--p 00000000 fd:00 100674505 /home/user/test
10020000-10030000 rw-p 00010000 fd:00 100674505 /home/user/test
10690000-106c0000 rwxp 00000000 00:00 0 [heap]
f7f70000-f7fa0000 r-xp 00000000 fd:00 67898089 /usr/lib/ld-2.17.so
f7fa0000-f7fb0000 r--p 00020000 fd:00 67898089 /usr/lib/ld-2.17.so
f7fb0000-f7fc0000 rw-p 00030000 fd:00 67898089 /usr/lib/ld-2.17.so
ffa90000-ffac0000 rw-p 00000000 00:00 0 [stack]
0x10690008
Patched ppc64 kernel:
00100000-00120000 r-xp 00000000 00:00 0 [vdso]
0fe10000-0ffd0000 r-xp 00000000 fd:00 67898094 /usr/lib/libc-2.17.so
0ffd0000-0ffe0000 r--p 001b0000 fd:00 67898094 /usr/lib/libc-2.17.so
0ffe0000-0fff0000 rw-p 001c0000 fd:00 67898094 /usr/lib/libc-2.17.so
10000000-10010000 r-xp 00000000 fd:00 100674505 /home/user/test
10010000-10020000 r--p 00000000 fd:00 100674505 /home/user/test
10020000-10030000 rw-p 00010000 fd:00 100674505 /home/user/test
10180000-101b0000 rw-p 00000000 00:00 0 [heap]
^^^^ this has changed
f7c60000-f7c90000 r-xp 00000000 fd:00 67898089 /usr/lib/ld-2.17.so
f7c90000-f7ca0000 r--p 00020000 fd:00 67898089 /usr/lib/ld-2.17.so
f7ca0000-f7cb0000 rw-p 00030000 fd:00 67898089 /usr/lib/ld-2.17.so
ff860000-ff890000 rw-p 00000000 00:00 0 [stack]
0x10180008
The patch was originally posted in 2012 by Jason Gunthorpe
and apparently ignored:
https://lkml.org/lkml/2012/9/30/138
Lightly run-tested.
Link: http://lkml.kernel.org/r/20161215131950.23054-1-dvlasenk@redhat.com
Signed-off-by: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
Signed-off-by: Denys Vlasenko <dvlasenk@redhat.com>
Acked-by: Kees Cook <keescook@chromium.org>
Acked-by: Michael Ellerman <mpe@ellerman.id.au>
Tested-by: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Florian Weimer <fweimer@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-23 07:45:16 +08:00
|
|
|
EXPORT_SYMBOL(vm_brk_flags);
|
|
|
|
|
|
|
|
|
|
int vm_brk(unsigned long addr, unsigned long len)
|
|
|
|
|
{
|
|
|
|
|
return vm_brk_flags(addr, len, 0);
|
|
|
|
|
}
|
2012-04-21 06:35:40 +08:00
|
|
|
EXPORT_SYMBOL(vm_brk);
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
|
/* Release all mmaps. */
|
|
|
|
|
void exit_mmap(struct mm_struct *mm)
|
|
|
|
|
{
|
2011-05-25 08:11:45 +08:00
|
|
|
struct mmu_gather tlb;
|
2008-10-19 11:26:50 +08:00
|
|
|
struct vm_area_struct *vma;
|
2005-04-17 06:20:36 +08:00
|
|
|
unsigned long nr_accounted = 0;
|
|
|
|
|
|
2007-05-03 01:27:14 +08:00
|
|
|
/* mm's last user has gone, and its about to be pulled down */
|
mmu-notifiers: core
With KVM/GFP/XPMEM there isn't just the primary CPU MMU pointing to pages.
There are secondary MMUs (with secondary sptes and secondary tlbs) too.
sptes in the kvm case are shadow pagetables, but when I say spte in
mmu-notifier context, I mean "secondary pte". In GRU case there's no
actual secondary pte and there's only a secondary tlb because the GRU
secondary MMU has no knowledge about sptes and every secondary tlb miss
event in the MMU always generates a page fault that has to be resolved by
the CPU (this is not the case of KVM where the a secondary tlb miss will
walk sptes in hardware and it will refill the secondary tlb transparently
to software if the corresponding spte is present). The same way
zap_page_range has to invalidate the pte before freeing the page, the spte
(and secondary tlb) must also be invalidated before any page is freed and
reused.
Currently we take a page_count pin on every page mapped by sptes, but that
means the pages can't be swapped whenever they're mapped by any spte
because they're part of the guest working set. Furthermore a spte unmap
event can immediately lead to a page to be freed when the pin is released
(so requiring the same complex and relatively slow tlb_gather smp safe
logic we have in zap_page_range and that can be avoided completely if the
spte unmap event doesn't require an unpin of the page previously mapped in
the secondary MMU).
The mmu notifiers allow kvm/GRU/XPMEM to attach to the tsk->mm and know
when the VM is swapping or freeing or doing anything on the primary MMU so
that the secondary MMU code can drop sptes before the pages are freed,
avoiding all page pinning and allowing 100% reliable swapping of guest
physical address space. Furthermore it avoids the code that teardown the
mappings of the secondary MMU, to implement a logic like tlb_gather in
zap_page_range that would require many IPI to flush other cpu tlbs, for
each fixed number of spte unmapped.
To make an example: if what happens on the primary MMU is a protection
downgrade (from writeable to wrprotect) the secondary MMU mappings will be
invalidated, and the next secondary-mmu-page-fault will call
get_user_pages and trigger a do_wp_page through get_user_pages if it
called get_user_pages with write=1, and it'll re-establishing an updated
spte or secondary-tlb-mapping on the copied page. Or it will setup a
readonly spte or readonly tlb mapping if it's a guest-read, if it calls
get_user_pages with write=0. This is just an example.
This allows to map any page pointed by any pte (and in turn visible in the
primary CPU MMU), into a secondary MMU (be it a pure tlb like GRU, or an
full MMU with both sptes and secondary-tlb like the shadow-pagetable layer
with kvm), or a remote DMA in software like XPMEM (hence needing of
schedule in XPMEM code to send the invalidate to the remote node, while no
need to schedule in kvm/gru as it's an immediate event like invalidating
primary-mmu pte).
At least for KVM without this patch it's impossible to swap guests
reliably. And having this feature and removing the page pin allows
several other optimizations that simplify life considerably.
Dependencies:
1) mm_take_all_locks() to register the mmu notifier when the whole VM
isn't doing anything with "mm". This allows mmu notifier users to keep
track if the VM is in the middle of the invalidate_range_begin/end
critical section with an atomic counter incraese in range_begin and
decreased in range_end. No secondary MMU page fault is allowed to map
any spte or secondary tlb reference, while the VM is in the middle of
range_begin/end as any page returned by get_user_pages in that critical
section could later immediately be freed without any further
->invalidate_page notification (invalidate_range_begin/end works on
ranges and ->invalidate_page isn't called immediately before freeing
the page). To stop all page freeing and pagetable overwrites the
mmap_sem must be taken in write mode and all other anon_vma/i_mmap
locks must be taken too.
2) It'd be a waste to add branches in the VM if nobody could possibly
run KVM/GRU/XPMEM on the kernel, so mmu notifiers will only enabled if
CONFIG_KVM=m/y. In the current kernel kvm won't yet take advantage of
mmu notifiers, but this already allows to compile a KVM external module
against a kernel with mmu notifiers enabled and from the next pull from
kvm.git we'll start using them. And GRU/XPMEM will also be able to
continue the development by enabling KVM=m in their config, until they
submit all GRU/XPMEM GPLv2 code to the mainline kernel. Then they can
also enable MMU_NOTIFIERS in the same way KVM does it (even if KVM=n).
This guarantees nobody selects MMU_NOTIFIER=y if KVM and GRU and XPMEM
are all =n.
The mmu_notifier_register call can fail because mm_take_all_locks may be
interrupted by a signal and return -EINTR. Because mmu_notifier_reigster
is used when a driver startup, a failure can be gracefully handled. Here
an example of the change applied to kvm to register the mmu notifiers.
Usually when a driver startups other allocations are required anyway and
-ENOMEM failure paths exists already.
struct kvm *kvm_arch_create_vm(void)
{
struct kvm *kvm = kzalloc(sizeof(struct kvm), GFP_KERNEL);
+ int err;
if (!kvm)
return ERR_PTR(-ENOMEM);
INIT_LIST_HEAD(&kvm->arch.active_mmu_pages);
+ kvm->arch.mmu_notifier.ops = &kvm_mmu_notifier_ops;
+ err = mmu_notifier_register(&kvm->arch.mmu_notifier, current->mm);
+ if (err) {
+ kfree(kvm);
+ return ERR_PTR(err);
+ }
+
return kvm;
}
mmu_notifier_unregister returns void and it's reliable.
The patch also adds a few needed but missing includes that would prevent
kernel to compile after these changes on non-x86 archs (x86 didn't need
them by luck).
[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: fix mm/filemap_xip.c build]
[akpm@linux-foundation.org: fix mm/mmu_notifier.c build]
Signed-off-by: Andrea Arcangeli <andrea@qumranet.com>
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Christoph Lameter <cl@linux-foundation.org>
Cc: Jack Steiner <steiner@sgi.com>
Cc: Robin Holt <holt@sgi.com>
Cc: Nick Piggin <npiggin@suse.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Kanoj Sarcar <kanojsarcar@yahoo.com>
Cc: Roland Dreier <rdreier@cisco.com>
Cc: Steve Wise <swise@opengridcomputing.com>
Cc: Avi Kivity <avi@qumranet.com>
Cc: Hugh Dickins <hugh@veritas.com>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Cc: Anthony Liguori <aliguori@us.ibm.com>
Cc: Chris Wright <chrisw@redhat.com>
Cc: Marcelo Tosatti <marcelo@kvack.org>
Cc: Eric Dumazet <dada1@cosmosbay.com>
Cc: "Paul E. McKenney" <paulmck@us.ibm.com>
Cc: Izik Eidus <izike@qumranet.com>
Cc: Anthony Liguori <aliguori@us.ibm.com>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-07-29 06:46:29 +08:00
|
|
|
mmu_notifier_release(mm);
|
2007-05-03 01:27:14 +08:00
|
|
|
|
2018-05-12 07:02:04 +08:00
|
|
|
if (unlikely(mm_is_oom_victim(mm))) {
|
|
|
|
|
/*
|
|
|
|
|
* Manually reap the mm to free as much memory as possible.
|
|
|
|
|
* Then, as the oom reaper does, set MMF_OOM_SKIP to disregard
|
2020-06-09 12:33:54 +08:00
|
|
|
* this mm from further consideration. Taking mm->mmap_lock for
|
2018-05-12 07:02:04 +08:00
|
|
|
* write after setting MMF_OOM_SKIP will guarantee that the oom
|
2020-06-09 12:33:54 +08:00
|
|
|
* reaper will not run on this mm again after mmap_lock is
|
2018-05-12 07:02:04 +08:00
|
|
|
* dropped.
|
|
|
|
|
*
|
2020-06-09 12:33:54 +08:00
|
|
|
* Nothing can be holding mm->mmap_lock here and the above call
|
2018-05-12 07:02:04 +08:00
|
|
|
* to mmu_notifier_release(mm) ensures mmu notifier callbacks in
|
|
|
|
|
* __oom_reap_task_mm() will not block.
|
|
|
|
|
*/
|
2018-08-22 12:52:33 +08:00
|
|
|
(void)__oom_reap_task_mm(mm);
|
2018-05-12 07:02:04 +08:00
|
|
|
set_bit(MMF_OOM_SKIP, &mm->flags);
|
|
|
|
|
}
|
|
|
|
|
|
2022-01-15 06:06:14 +08:00
|
|
|
mmap_write_lock(mm);
|
mm: rearrange exit_mmap() to unlock before arch_exit_mmap
Christophe Saout reported [in precursor to:
http://marc.info/?l=linux-kernel&m=123209902707347&w=4]:
> Note that I also some a different issue with CONFIG_UNEVICTABLE_LRU.
> Seems like Xen tears down current->mm early on process termination, so
> that __get_user_pages in exit_mmap causes nasty messages when the
> process had any mlocked pages. (in fact, it somehow manages to get into
> the swapping code and produces a null pointer dereference trying to get
> a swap token)
Jeremy explained:
Yes. In the normal case under Xen, an in-use pagetable is "pinned",
meaning that it is RO to the kernel, and all updates must go via hypercall
(or writes are trapped and emulated, which is much the same thing). An
unpinned pagetable is not currently in use by any process, and can be
directly accessed as normal RW pages.
As an optimisation at process exit time, we unpin the pagetable as early
as possible (switching the process to init_mm), so that all the normal
pagetable teardown can happen with direct memory accesses.
This happens in exit_mmap() -> arch_exit_mmap(). The munlocking happens
a few lines below. The obvious thing to do would be to move
arch_exit_mmap() to below the munlock code, but I think we'd want to
call it even if mm->mmap is NULL, just to be on the safe side.
Thus, this patch:
exit_mmap() needs to unlock any locked vmas before calling arch_exit_mmap,
as the latter may switch the current mm to init_mm, which would cause the
former to fail.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christophe Saout <christophe@saout.de>
Cc: Keir Fraser <keir.fraser@eu.citrix.com>
Cc: Christophe Saout <christophe@saout.de>
Cc: Alex Williamson <alex.williamson@hp.com>
Cc: <stable@kernel.org> [2.6.28.x]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-02-12 05:04:41 +08:00
|
|
|
arch_exit_mmap(mm);
|
|
|
|
|
|
2008-10-19 11:26:50 +08:00
|
|
|
vma = mm->mmap;
|
2022-01-15 06:06:14 +08:00
|
|
|
if (!vma) {
|
|
|
|
|
/* Can happen if dup_mmap() received an OOM */
|
|
|
|
|
mmap_write_unlock(mm);
|
mm: rearrange exit_mmap() to unlock before arch_exit_mmap
Christophe Saout reported [in precursor to:
http://marc.info/?l=linux-kernel&m=123209902707347&w=4]:
> Note that I also some a different issue with CONFIG_UNEVICTABLE_LRU.
> Seems like Xen tears down current->mm early on process termination, so
> that __get_user_pages in exit_mmap causes nasty messages when the
> process had any mlocked pages. (in fact, it somehow manages to get into
> the swapping code and produces a null pointer dereference trying to get
> a swap token)
Jeremy explained:
Yes. In the normal case under Xen, an in-use pagetable is "pinned",
meaning that it is RO to the kernel, and all updates must go via hypercall
(or writes are trapped and emulated, which is much the same thing). An
unpinned pagetable is not currently in use by any process, and can be
directly accessed as normal RW pages.
As an optimisation at process exit time, we unpin the pagetable as early
as possible (switching the process to init_mm), so that all the normal
pagetable teardown can happen with direct memory accesses.
This happens in exit_mmap() -> arch_exit_mmap(). The munlocking happens
a few lines below. The obvious thing to do would be to move
arch_exit_mmap() to below the munlock code, but I think we'd want to
call it even if mm->mmap is NULL, just to be on the safe side.
Thus, this patch:
exit_mmap() needs to unlock any locked vmas before calling arch_exit_mmap,
as the latter may switch the current mm to init_mm, which would cause the
former to fail.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christophe Saout <christophe@saout.de>
Cc: Keir Fraser <keir.fraser@eu.citrix.com>
Cc: Christophe Saout <christophe@saout.de>
Cc: Alex Williamson <alex.williamson@hp.com>
Cc: <stable@kernel.org> [2.6.28.x]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-02-12 05:04:41 +08:00
|
|
|
return;
|
2022-01-15 06:06:14 +08:00
|
|
|
}
|
mm: rearrange exit_mmap() to unlock before arch_exit_mmap
Christophe Saout reported [in precursor to:
http://marc.info/?l=linux-kernel&m=123209902707347&w=4]:
> Note that I also some a different issue with CONFIG_UNEVICTABLE_LRU.
> Seems like Xen tears down current->mm early on process termination, so
> that __get_user_pages in exit_mmap causes nasty messages when the
> process had any mlocked pages. (in fact, it somehow manages to get into
> the swapping code and produces a null pointer dereference trying to get
> a swap token)
Jeremy explained:
Yes. In the normal case under Xen, an in-use pagetable is "pinned",
meaning that it is RO to the kernel, and all updates must go via hypercall
(or writes are trapped and emulated, which is much the same thing). An
unpinned pagetable is not currently in use by any process, and can be
directly accessed as normal RW pages.
As an optimisation at process exit time, we unpin the pagetable as early
as possible (switching the process to init_mm), so that all the normal
pagetable teardown can happen with direct memory accesses.
This happens in exit_mmap() -> arch_exit_mmap(). The munlocking happens
a few lines below. The obvious thing to do would be to move
arch_exit_mmap() to below the munlock code, but I think we'd want to
call it even if mm->mmap is NULL, just to be on the safe side.
Thus, this patch:
exit_mmap() needs to unlock any locked vmas before calling arch_exit_mmap,
as the latter may switch the current mm to init_mm, which would cause the
former to fail.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Christophe Saout <christophe@saout.de>
Cc: Keir Fraser <keir.fraser@eu.citrix.com>
Cc: Christophe Saout <christophe@saout.de>
Cc: Alex Williamson <alex.williamson@hp.com>
Cc: <stable@kernel.org> [2.6.28.x]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-02-12 05:04:41 +08:00
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
lru_add_drain();
|
|
|
|
|
flush_cache_mm(mm);
|
2021-01-28 07:53:44 +08:00
|
|
|
tlb_gather_mmu_fullmm(&tlb, mm);
|
2009-01-07 06:40:29 +08:00
|
|
|
/* update_hiwater_rss(mm) here? but nobody should be looking */
|
[PATCH] freepgt: free_pgtables use vma list
Recent woes with some arches needing their own pgd_addr_end macro; and 4-level
clear_page_range regression since 2.6.10's clear_page_tables; and its
long-standing well-known inefficiency in searching throughout the higher-level
page tables for those few entries to clear and free: all can be blamed on
ignoring the list of vmas when we free page tables.
Replace exit_mmap's clear_page_range of the total user address space by
free_pgtables operating on the mm's vma list; unmap_region use it in the same
way, giving floor and ceiling beyond which it may not free tables. This
brings lmbench fork/exec/sh numbers back to 2.6.10 (unless preempt is enabled,
in which case latency fixes spoil unmap_vmas throughput).
Beware: the do_mmap_pgoff driver failure case must now use unmap_region
instead of zap_page_range, since a page table might have been allocated, and
can only be freed while it is touched by some vma.
Move free_pgtables from mmap.c to memory.c, where its lower levels are adapted
from the clear_page_range levels. (Most of free_pgtables' old code was
actually for a non-existent case, prev not properly set up, dating from before
hch gave us split_vma.) Pass mmu_gather** in the public interfaces, since we
might want to add latency lockdrops later; but no attempt to do so yet, going
by vma should itself reduce latency.
But what if is_hugepage_only_range? Those ia64 and ppc64 cases need careful
examination: put that off until a later patch of the series.
What of x86_64's 32bit vdso page __map_syscall32 maps outside any vma?
And the range to sparc64's flush_tlb_pgtables? It's less clear to me now that
we need to do more than is done here - every PMD_SIZE ever occupied will be
flushed, do we really have to flush every PGDIR_SIZE ever partially occupied?
A shame to complicate it unnecessarily.
Special thanks to David Miller for time spent repairing my ceilings.
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-04-20 04:29:15 +08:00
|
|
|
/* Use -1 here to ensure all VMAs in the mm are unmapped */
|
2012-05-07 04:54:06 +08:00
|
|
|
unmap_vmas(&tlb, vma, 0, -1);
|
2013-04-30 06:07:44 +08:00
|
|
|
free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
|
2021-01-28 07:53:43 +08:00
|
|
|
tlb_finish_mmu(&tlb);
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2022-01-15 06:06:14 +08:00
|
|
|
/* Walk the list again, actually closing and freeing it. */
|
2012-05-07 04:54:06 +08:00
|
|
|
while (vma) {
|
|
|
|
|
if (vma->vm_flags & VM_ACCOUNT)
|
|
|
|
|
nr_accounted += vma_pages(vma);
|
2005-10-30 09:15:57 +08:00
|
|
|
vma = remove_vma(vma);
|
2020-04-17 07:46:10 +08:00
|
|
|
cond_resched();
|
2012-05-07 04:54:06 +08:00
|
|
|
}
|
2022-02-26 11:11:05 +08:00
|
|
|
mm->mmap = NULL;
|
2022-01-15 06:06:14 +08:00
|
|
|
mmap_write_unlock(mm);
|
2012-05-07 04:54:06 +08:00
|
|
|
vm_unacct_memory(nr_accounted);
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Insert vm structure into process list sorted by address
|
|
|
|
|
* and into the inode's i_mmap tree. If vm_file is non-NULL
|
2014-12-13 08:54:24 +08:00
|
|
|
* then i_mmap_rwsem is taken here.
|
2005-04-17 06:20:36 +08:00
|
|
|
*/
|
2012-10-09 07:29:07 +08:00
|
|
|
int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
2012-10-09 07:29:07 +08:00
|
|
|
struct vm_area_struct *prev;
|
|
|
|
|
struct rb_node **rb_link, *rb_parent;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2015-09-09 06:04:08 +08:00
|
|
|
if (find_vma_links(mm, vma->vm_start, vma->vm_end,
|
|
|
|
|
&prev, &rb_link, &rb_parent))
|
|
|
|
|
return -ENOMEM;
|
|
|
|
|
if ((vma->vm_flags & VM_ACCOUNT) &&
|
|
|
|
|
security_vm_enough_memory_mm(mm, vma_pages(vma)))
|
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
/*
|
|
|
|
|
* The vm_pgoff of a purely anonymous vma should be irrelevant
|
|
|
|
|
* until its first write fault, when page's anon_vma and index
|
|
|
|
|
* are set. But now set the vm_pgoff it will almost certainly
|
|
|
|
|
* end up with (unless mremap moves it elsewhere before that
|
|
|
|
|
* first wfault), so /proc/pid/maps tells a consistent story.
|
|
|
|
|
*
|
|
|
|
|
* By setting it to reflect the virtual start address of the
|
|
|
|
|
* vma, merges and splits can happen in a seamless way, just
|
|
|
|
|
* using the existing file pgoff checks and manipulations.
|
2020-10-14 07:54:18 +08:00
|
|
|
* Similarly in do_mmap and in do_brk_flags.
|
2005-04-17 06:20:36 +08:00
|
|
|
*/
|
mmap: fix the usage of ->vm_pgoff in special_mapping paths
Test-case:
#include <stdio.h>
#include <unistd.h>
#include <stdlib.h>
#include <string.h>
#include <sys/mman.h>
#include <assert.h>
void *find_vdso_vaddr(void)
{
FILE *perl;
char buf[32] = {};
perl = popen("perl -e 'open STDIN,qq|/proc/@{[getppid]}/maps|;"
"/^(.*?)-.*vdso/ && print hex $1 while <>'", "r");
fread(buf, sizeof(buf), 1, perl);
fclose(perl);
return (void *)atol(buf);
}
#define PAGE_SIZE 4096
int main(void)
{
void *vdso = find_vdso_vaddr();
assert(vdso);
// of course they should differ, and they do so far
printf("vdso pages differ: %d\n",
!!memcmp(vdso, vdso + PAGE_SIZE, PAGE_SIZE));
// split into 2 vma's
assert(mprotect(vdso, PAGE_SIZE, PROT_READ) == 0);
// force another fault on the next check
assert(madvise(vdso, 2 * PAGE_SIZE, MADV_DONTNEED) == 0);
// now they no longer differ, the 2nd vm_pgoff is wrong
printf("vdso pages differ: %d\n",
!!memcmp(vdso, vdso + PAGE_SIZE, PAGE_SIZE));
return 0;
}
Output:
vdso pages differ: 1
vdso pages differ: 0
This is because split_vma() correctly updates ->vm_pgoff, but the logic
in insert_vm_struct() and special_mapping_fault() is absolutely broken,
so the fault at vdso + PAGE_SIZE return the 1st page. The same happens
if you simply unmap the 1st page.
special_mapping_fault() does:
pgoff = vmf->pgoff - vma->vm_pgoff;
and this is _only_ correct if vma->vm_start mmaps the first page from
->vm_private_data array.
vdso or any other user of install_special_mapping() is not anonymous,
it has the "backing storage" even if it is just the array of pages.
So we actually need to make vm_pgoff work as an offset in this array.
Note: this also allows to fix another problem: currently gdb can't access
"[vvar]" memory because in this case special_mapping_fault() doesn't work.
Now that we can use ->vm_pgoff we can implement ->access() and fix this.
Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-09 05:58:31 +08:00
|
|
|
if (vma_is_anonymous(vma)) {
|
2005-04-17 06:20:36 +08:00
|
|
|
BUG_ON(vma->anon_vma);
|
|
|
|
|
vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
|
|
|
|
|
}
|
uprobes, mm, x86: Add the ability to install and remove uprobes breakpoints
Add uprobes support to the core kernel, with x86 support.
This commit adds the kernel facilities, the actual uprobes
user-space ABI and perf probe support comes in later commits.
General design:
Uprobes are maintained in an rb-tree indexed by inode and offset
(the offset here is from the start of the mapping). For a unique
(inode, offset) tuple, there can be at most one uprobe in the
rb-tree.
Since the (inode, offset) tuple identifies a unique uprobe, more
than one user may be interested in the same uprobe. This provides
the ability to connect multiple 'consumers' to the same uprobe.
Each consumer defines a handler and a filter (optional). The
'handler' is run every time the uprobe is hit, if it matches the
'filter' criteria.
The first consumer of a uprobe causes the breakpoint to be
inserted at the specified address and subsequent consumers are
appended to this list. On subsequent probes, the consumer gets
appended to the existing list of consumers. The breakpoint is
removed when the last consumer unregisters. For all other
unregisterations, the consumer is removed from the list of
consumers.
Given a inode, we get a list of the mms that have mapped the
inode. Do the actual registration if mm maps the page where a
probe needs to be inserted/removed.
We use a temporary list to walk through the vmas that map the
inode.
- The number of maps that map the inode, is not known before we
walk the rmap and keeps changing.
- extending vm_area_struct wasn't recommended, it's a
size-critical data structure.
- There can be more than one maps of the inode in the same mm.
We add callbacks to the mmap methods to keep an eye on text vmas
that are of interest to uprobes. When a vma of interest is mapped,
we insert the breakpoint at the right address.
Uprobe works by replacing the instruction at the address defined
by (inode, offset) with the arch specific breakpoint
instruction. We save a copy of the original instruction at the
uprobed address.
This is needed for:
a. executing the instruction out-of-line (xol).
b. instruction analysis for any subsequent fixups.
c. restoring the instruction back when the uprobe is unregistered.
We insert or delete a breakpoint instruction, and this
breakpoint instruction is assumed to be the smallest instruction
available on the platform. For fixed size instruction platforms
this is trivially true, for variable size instruction platforms
the breakpoint instruction is typically the smallest (often a
single byte).
Writing the instruction is done by COWing the page and changing
the instruction during the copy, this even though most platforms
allow atomic writes of the breakpoint instruction. This also
mirrors the behaviour of a ptrace() memory write to a PRIVATE
file map.
The core worker is derived from KSM's replace_page() logic.
In essence, similar to KSM:
a. allocate a new page and copy over contents of the page that
has the uprobed vaddr
b. modify the copy and insert the breakpoint at the required
address
c. switch the original page with the copy containing the
breakpoint
d. flush page tables.
replace_page() is being replicated here because of some minor
changes in the type of pages and also because Hugh Dickins had
plans to improve replace_page() for KSM specific work.
Instruction analysis on x86 is based on instruction decoder and
determines if an instruction can be probed and determines the
necessary fixups after singlestep. Instruction analysis is done
at probe insertion time so that we avoid having to repeat the
same analysis every time a probe is hit.
A lot of code here is due to the improvement/suggestions/inputs
from Peter Zijlstra.
Changelog:
(v10):
- Add code to clear REX.B prefix as suggested by Denys Vlasenko
and Masami Hiramatsu.
(v9):
- Use insn_offset_modrm as suggested by Masami Hiramatsu.
(v7):
Handle comments from Peter Zijlstra:
- Dont take reference to inode. (expect inode to uprobe_register to be sane).
- Use PTR_ERR to set the return value.
- No need to take reference to inode.
- use PTR_ERR to return error value.
- register and uprobe_unregister share code.
(v5):
- Modified del_consumer as per comments from Peter.
- Drop reference to inode before dropping reference to uprobe.
- Use i_size_read(inode) instead of inode->i_size.
- Ensure uprobe->consumers is NULL, before __uprobe_unregister() is called.
- Includes errno.h as recommended by Stephen Rothwell to fix a build issue
on sparc defconfig
- Remove restrictions while unregistering.
- Earlier code leaked inode references under some conditions while
registering/unregistering.
- Continue the vma-rmap walk even if the intermediate vma doesnt
meet the requirements.
- Validate the vma found by find_vma before inserting/removing the
breakpoint
- Call del_consumer under mutex_lock.
- Use hash locks.
- Handle mremap.
- Introduce find_least_offset_node() instead of close match logic in
find_uprobe
- Uprobes no more depends on MM_OWNER; No reference to task_structs
while inserting/removing a probe.
- Uses read_mapping_page instead of grab_cache_page so that the pages
have valid content.
- pass NULL to get_user_pages for the task parameter.
- call SetPageUptodate on the new page allocated in write_opcode.
- fix leaking a reference to the new page under certain conditions.
- Include Instruction Decoder if Uprobes gets defined.
- Remove const attributes for instruction prefix arrays.
- Uses mm_context to know if the application is 32 bit.
Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Also-written-by: Jim Keniston <jkenisto@us.ibm.com>
Reviewed-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Roland McGrath <roland@hack.frob.com>
Cc: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
Cc: Anton Arapov <anton@redhat.com>
Cc: Ananth N Mavinakayanahalli <ananth@in.ibm.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Denys Vlasenko <vda.linux@googlemail.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Linux-mm <linux-mm@kvack.org>
Link: http://lkml.kernel.org/r/20120209092642.GE16600@linux.vnet.ibm.com
[ Made various small edits to the commit log ]
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2012-02-09 17:26:42 +08:00
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
vma_link(mm, vma, prev, rb_link, rb_parent);
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Copy the vma structure to a new location in the same mm,
|
|
|
|
|
* prior to moving page table entries, to effect an mremap move.
|
|
|
|
|
*/
|
|
|
|
|
struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
|
2012-10-09 07:31:50 +08:00
|
|
|
unsigned long addr, unsigned long len, pgoff_t pgoff,
|
|
|
|
|
bool *need_rmap_locks)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
|
|
|
|
struct vm_area_struct *vma = *vmap;
|
|
|
|
|
unsigned long vma_start = vma->vm_start;
|
|
|
|
|
struct mm_struct *mm = vma->vm_mm;
|
|
|
|
|
struct vm_area_struct *new_vma, *prev;
|
|
|
|
|
struct rb_node **rb_link, *rb_parent;
|
mremap: enforce rmap src/dst vma ordering in case of vma_merge() succeeding in copy_vma()
migrate was doing an rmap_walk with speculative lock-less access on
pagetables. That could lead it to not serializing properly against mremap
PT locks. But a second problem remains in the order of vmas in the
same_anon_vma list used by the rmap_walk.
If vma_merge succeeds in copy_vma, the src vma could be placed after the
dst vma in the same_anon_vma list. That could still lead to migrate
missing some pte.
This patch adds an anon_vma_moveto_tail() function to force the dst vma at
the end of the list before mremap starts to solve the problem.
If the mremap is very large and there are a lots of parents or childs
sharing the anon_vma root lock, this should still scale better than taking
the anon_vma root lock around every pte copy practically for the whole
duration of mremap.
Update: Hugh noticed special care is needed in the error path where
move_page_tables goes in the reverse direction, a second
anon_vma_moveto_tail() call is needed in the error path.
This program exercises the anon_vma_moveto_tail:
===
int main()
{
static struct timeval oldstamp, newstamp;
long diffsec;
char *p, *p2, *p3, *p4;
if (posix_memalign((void **)&p, 2*1024*1024, SIZE))
perror("memalign"), exit(1);
if (posix_memalign((void **)&p2, 2*1024*1024, SIZE))
perror("memalign"), exit(1);
if (posix_memalign((void **)&p3, 2*1024*1024, SIZE))
perror("memalign"), exit(1);
memset(p, 0xff, SIZE);
printf("%p\n", p);
memset(p2, 0xff, SIZE);
memset(p3, 0x77, 4096);
if (memcmp(p, p2, SIZE))
printf("error\n");
p4 = mremap(p+SIZE/2, SIZE/2, SIZE/2, MREMAP_FIXED|MREMAP_MAYMOVE, p3);
if (p4 != p3)
perror("mremap"), exit(1);
p4 = mremap(p4, SIZE/2, SIZE/2, MREMAP_FIXED|MREMAP_MAYMOVE, p+SIZE/2);
if (p4 != p+SIZE/2)
perror("mremap"), exit(1);
if (memcmp(p, p2, SIZE))
printf("error\n");
printf("ok\n");
return 0;
}
===
$ perf probe -a anon_vma_moveto_tail
Add new event:
probe:anon_vma_moveto_tail (on anon_vma_moveto_tail)
You can now use it on all perf tools, such as:
perf record -e probe:anon_vma_moveto_tail -aR sleep 1
$ perf record -e probe:anon_vma_moveto_tail -aR ./anon_vma_moveto_tail
0x7f2ca2800000
ok
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.043 MB perf.data (~1860 samples) ]
$ perf report --stdio
100.00% anon_vma_moveto [kernel.kallsyms] [k] anon_vma_moveto_tail
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Nai Xia <nai.xia@gmail.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Hugh Dickins <hughd@google.com>
Cc: Pawel Sikora <pluto@agmk.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-11 07:08:05 +08:00
|
|
|
bool faulted_in_anon_vma = true;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* If anonymous vma has not yet been faulted, update new pgoff
|
|
|
|
|
* to match new location, to increase its chance of merging.
|
|
|
|
|
*/
|
mremap: fix the wrong !vma->vm_file check in copy_vma()
Test-case:
#define _GNU_SOURCE
#include <stdio.h>
#include <unistd.h>
#include <stdlib.h>
#include <string.h>
#include <sys/mman.h>
#include <assert.h>
void *find_vdso_vaddr(void)
{
FILE *perl;
char buf[32] = {};
perl = popen("perl -e 'open STDIN,qq|/proc/@{[getppid]}/maps|;"
"/^(.*?)-.*vdso/ && print hex $1 while <>'", "r");
fread(buf, sizeof(buf), 1, perl);
fclose(perl);
return (void *)atol(buf);
}
#define PAGE_SIZE 4096
void *get_unmapped_area(void)
{
void *p = mmap(0, PAGE_SIZE, PROT_NONE,
MAP_PRIVATE|MAP_ANONYMOUS, -1,0);
assert(p != MAP_FAILED);
munmap(p, PAGE_SIZE);
return p;
}
char save[2][PAGE_SIZE];
int main(void)
{
void *vdso = find_vdso_vaddr();
void *page[2];
assert(vdso);
memcpy(save, vdso, sizeof (save));
// force another fault on the next check
assert(madvise(vdso, 2 * PAGE_SIZE, MADV_DONTNEED) == 0);
page[0] = mremap(vdso,
PAGE_SIZE, PAGE_SIZE, MREMAP_FIXED | MREMAP_MAYMOVE,
get_unmapped_area());
page[1] = mremap(vdso + PAGE_SIZE,
PAGE_SIZE, PAGE_SIZE, MREMAP_FIXED | MREMAP_MAYMOVE,
get_unmapped_area());
assert(page[0] != MAP_FAILED && page[1] != MAP_FAILED);
printf("match: %d %d\n",
!memcmp(save[0], page[0], PAGE_SIZE),
!memcmp(save[1], page[1], PAGE_SIZE));
return 0;
}
fails without this patch. Before the previous commit it gets the wrong
page, now it segfaults (which is imho better).
This is because copy_vma() wrongly assumes that if vma->vm_file == NULL
is irrelevant until the first fault which will use do_anonymous_page().
This is obviously wrong for the special mapping.
Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-09 05:58:34 +08:00
|
|
|
if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
|
2005-04-17 06:20:36 +08:00
|
|
|
pgoff = addr >> PAGE_SHIFT;
|
mremap: enforce rmap src/dst vma ordering in case of vma_merge() succeeding in copy_vma()
migrate was doing an rmap_walk with speculative lock-less access on
pagetables. That could lead it to not serializing properly against mremap
PT locks. But a second problem remains in the order of vmas in the
same_anon_vma list used by the rmap_walk.
If vma_merge succeeds in copy_vma, the src vma could be placed after the
dst vma in the same_anon_vma list. That could still lead to migrate
missing some pte.
This patch adds an anon_vma_moveto_tail() function to force the dst vma at
the end of the list before mremap starts to solve the problem.
If the mremap is very large and there are a lots of parents or childs
sharing the anon_vma root lock, this should still scale better than taking
the anon_vma root lock around every pte copy practically for the whole
duration of mremap.
Update: Hugh noticed special care is needed in the error path where
move_page_tables goes in the reverse direction, a second
anon_vma_moveto_tail() call is needed in the error path.
This program exercises the anon_vma_moveto_tail:
===
int main()
{
static struct timeval oldstamp, newstamp;
long diffsec;
char *p, *p2, *p3, *p4;
if (posix_memalign((void **)&p, 2*1024*1024, SIZE))
perror("memalign"), exit(1);
if (posix_memalign((void **)&p2, 2*1024*1024, SIZE))
perror("memalign"), exit(1);
if (posix_memalign((void **)&p3, 2*1024*1024, SIZE))
perror("memalign"), exit(1);
memset(p, 0xff, SIZE);
printf("%p\n", p);
memset(p2, 0xff, SIZE);
memset(p3, 0x77, 4096);
if (memcmp(p, p2, SIZE))
printf("error\n");
p4 = mremap(p+SIZE/2, SIZE/2, SIZE/2, MREMAP_FIXED|MREMAP_MAYMOVE, p3);
if (p4 != p3)
perror("mremap"), exit(1);
p4 = mremap(p4, SIZE/2, SIZE/2, MREMAP_FIXED|MREMAP_MAYMOVE, p+SIZE/2);
if (p4 != p+SIZE/2)
perror("mremap"), exit(1);
if (memcmp(p, p2, SIZE))
printf("error\n");
printf("ok\n");
return 0;
}
===
$ perf probe -a anon_vma_moveto_tail
Add new event:
probe:anon_vma_moveto_tail (on anon_vma_moveto_tail)
You can now use it on all perf tools, such as:
perf record -e probe:anon_vma_moveto_tail -aR sleep 1
$ perf record -e probe:anon_vma_moveto_tail -aR ./anon_vma_moveto_tail
0x7f2ca2800000
ok
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.043 MB perf.data (~1860 samples) ]
$ perf report --stdio
100.00% anon_vma_moveto [kernel.kallsyms] [k] anon_vma_moveto_tail
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Nai Xia <nai.xia@gmail.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Hugh Dickins <hughd@google.com>
Cc: Pawel Sikora <pluto@agmk.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-11 07:08:05 +08:00
|
|
|
faulted_in_anon_vma = false;
|
|
|
|
|
}
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2012-10-09 07:29:07 +08:00
|
|
|
if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
|
|
|
|
|
return NULL; /* should never get here */
|
2005-04-17 06:20:36 +08:00
|
|
|
new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
|
2015-09-05 06:46:24 +08:00
|
|
|
vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
|
2022-03-05 12:28:51 +08:00
|
|
|
vma->vm_userfaultfd_ctx, anon_vma_name(vma));
|
2005-04-17 06:20:36 +08:00
|
|
|
if (new_vma) {
|
|
|
|
|
/*
|
|
|
|
|
* Source vma may have been merged into new_vma
|
|
|
|
|
*/
|
mremap: enforce rmap src/dst vma ordering in case of vma_merge() succeeding in copy_vma()
migrate was doing an rmap_walk with speculative lock-less access on
pagetables. That could lead it to not serializing properly against mremap
PT locks. But a second problem remains in the order of vmas in the
same_anon_vma list used by the rmap_walk.
If vma_merge succeeds in copy_vma, the src vma could be placed after the
dst vma in the same_anon_vma list. That could still lead to migrate
missing some pte.
This patch adds an anon_vma_moveto_tail() function to force the dst vma at
the end of the list before mremap starts to solve the problem.
If the mremap is very large and there are a lots of parents or childs
sharing the anon_vma root lock, this should still scale better than taking
the anon_vma root lock around every pte copy practically for the whole
duration of mremap.
Update: Hugh noticed special care is needed in the error path where
move_page_tables goes in the reverse direction, a second
anon_vma_moveto_tail() call is needed in the error path.
This program exercises the anon_vma_moveto_tail:
===
int main()
{
static struct timeval oldstamp, newstamp;
long diffsec;
char *p, *p2, *p3, *p4;
if (posix_memalign((void **)&p, 2*1024*1024, SIZE))
perror("memalign"), exit(1);
if (posix_memalign((void **)&p2, 2*1024*1024, SIZE))
perror("memalign"), exit(1);
if (posix_memalign((void **)&p3, 2*1024*1024, SIZE))
perror("memalign"), exit(1);
memset(p, 0xff, SIZE);
printf("%p\n", p);
memset(p2, 0xff, SIZE);
memset(p3, 0x77, 4096);
if (memcmp(p, p2, SIZE))
printf("error\n");
p4 = mremap(p+SIZE/2, SIZE/2, SIZE/2, MREMAP_FIXED|MREMAP_MAYMOVE, p3);
if (p4 != p3)
perror("mremap"), exit(1);
p4 = mremap(p4, SIZE/2, SIZE/2, MREMAP_FIXED|MREMAP_MAYMOVE, p+SIZE/2);
if (p4 != p+SIZE/2)
perror("mremap"), exit(1);
if (memcmp(p, p2, SIZE))
printf("error\n");
printf("ok\n");
return 0;
}
===
$ perf probe -a anon_vma_moveto_tail
Add new event:
probe:anon_vma_moveto_tail (on anon_vma_moveto_tail)
You can now use it on all perf tools, such as:
perf record -e probe:anon_vma_moveto_tail -aR sleep 1
$ perf record -e probe:anon_vma_moveto_tail -aR ./anon_vma_moveto_tail
0x7f2ca2800000
ok
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.043 MB perf.data (~1860 samples) ]
$ perf report --stdio
100.00% anon_vma_moveto [kernel.kallsyms] [k] anon_vma_moveto_tail
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Reported-by: Nai Xia <nai.xia@gmail.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Hugh Dickins <hughd@google.com>
Cc: Pawel Sikora <pluto@agmk.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-11 07:08:05 +08:00
|
|
|
if (unlikely(vma_start >= new_vma->vm_start &&
|
|
|
|
|
vma_start < new_vma->vm_end)) {
|
|
|
|
|
/*
|
|
|
|
|
* The only way we can get a vma_merge with
|
|
|
|
|
* self during an mremap is if the vma hasn't
|
|
|
|
|
* been faulted in yet and we were allowed to
|
|
|
|
|
* reset the dst vma->vm_pgoff to the
|
|
|
|
|
* destination address of the mremap to allow
|
|
|
|
|
* the merge to happen. mremap must change the
|
|
|
|
|
* vm_pgoff linearity between src and dst vmas
|
|
|
|
|
* (in turn preventing a vma_merge) to be
|
|
|
|
|
* safe. It is only safe to keep the vm_pgoff
|
|
|
|
|
* linear if there are no pages mapped yet.
|
|
|
|
|
*/
|
2014-10-10 06:28:10 +08:00
|
|
|
VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
|
2012-10-09 07:31:50 +08:00
|
|
|
*vmap = vma = new_vma;
|
2012-10-09 07:31:36 +08:00
|
|
|
}
|
2012-10-09 07:31:50 +08:00
|
|
|
*need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
|
2005-04-17 06:20:36 +08:00
|
|
|
} else {
|
2018-07-22 04:48:51 +08:00
|
|
|
new_vma = vm_area_dup(vma);
|
2015-09-09 06:03:38 +08:00
|
|
|
if (!new_vma)
|
|
|
|
|
goto out;
|
|
|
|
|
new_vma->vm_start = addr;
|
|
|
|
|
new_vma->vm_end = addr + len;
|
|
|
|
|
new_vma->vm_pgoff = pgoff;
|
|
|
|
|
if (vma_dup_policy(vma, new_vma))
|
|
|
|
|
goto out_free_vma;
|
|
|
|
|
if (anon_vma_clone(new_vma, vma))
|
|
|
|
|
goto out_free_mempol;
|
|
|
|
|
if (new_vma->vm_file)
|
|
|
|
|
get_file(new_vma->vm_file);
|
|
|
|
|
if (new_vma->vm_ops && new_vma->vm_ops->open)
|
|
|
|
|
new_vma->vm_ops->open(new_vma);
|
|
|
|
|
vma_link(mm, new_vma, prev, rb_link, rb_parent);
|
|
|
|
|
*need_rmap_locks = false;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
return new_vma;
|
mm: change anon_vma linking to fix multi-process server scalability issue
The old anon_vma code can lead to scalability issues with heavily forking
workloads. Specifically, each anon_vma will be shared between the parent
process and all its child processes.
In a workload with 1000 child processes and a VMA with 1000 anonymous
pages per process that get COWed, this leads to a system with a million
anonymous pages in the same anon_vma, each of which is mapped in just one
of the 1000 processes. However, the current rmap code needs to walk them
all, leading to O(N) scanning complexity for each page.
This can result in systems where one CPU is walking the page tables of
1000 processes in page_referenced_one, while all other CPUs are stuck on
the anon_vma lock. This leads to catastrophic failure for a benchmark
like AIM7, where the total number of processes can reach in the tens of
thousands. Real workloads are still a factor 10 less process intensive
than AIM7, but they are catching up.
This patch changes the way anon_vmas and VMAs are linked, which allows us
to associate multiple anon_vmas with a VMA. At fork time, each child
process gets its own anon_vmas, in which its COWed pages will be
instantiated. The parents' anon_vma is also linked to the VMA, because
non-COWed pages could be present in any of the children.
This reduces rmap scanning complexity to O(1) for the pages of the 1000
child processes, with O(N) complexity for at most 1/N pages in the system.
This reduces the average scanning cost in heavily forking workloads from
O(N) to 2.
The only real complexity in this patch stems from the fact that linking a
VMA to anon_vmas now involves memory allocations. This means vma_adjust
can fail, if it needs to attach a VMA to anon_vma structures. This in
turn means error handling needs to be added to the calling functions.
A second source of complexity is that, because there can be multiple
anon_vmas, the anon_vma linking in vma_adjust can no longer be done under
"the" anon_vma lock. To prevent the rmap code from walking up an
incomplete VMA, this patch introduces the VM_LOCK_RMAP VMA flag. This bit
flag uses the same slot as the NOMMU VM_MAPPED_COPY, with an ifdef in mm.h
to make sure it is impossible to compile a kernel that needs both symbolic
values for the same bitflag.
Some test results:
Without the anon_vma changes, when AIM7 hits around 9.7k users (on a test
box with 16GB RAM and not quite enough IO), the system ends up running
>99% in system time, with every CPU on the same anon_vma lock in the
pageout code.
With these changes, AIM7 hits the cross-over point around 29.7k users.
This happens with ~99% IO wait time, there never seems to be any spike in
system time. The anon_vma lock contention appears to be resolved.
[akpm@linux-foundation.org: cleanups]
Signed-off-by: Rik van Riel <riel@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Larry Woodman <lwoodman@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-06 05:42:07 +08:00
|
|
|
|
2015-09-09 06:03:38 +08:00
|
|
|
out_free_mempol:
|
2013-09-12 05:20:14 +08:00
|
|
|
mpol_put(vma_policy(new_vma));
|
2015-09-09 06:03:38 +08:00
|
|
|
out_free_vma:
|
2018-07-22 04:48:51 +08:00
|
|
|
vm_area_free(new_vma);
|
2015-09-09 06:03:38 +08:00
|
|
|
out:
|
mm: change anon_vma linking to fix multi-process server scalability issue
The old anon_vma code can lead to scalability issues with heavily forking
workloads. Specifically, each anon_vma will be shared between the parent
process and all its child processes.
In a workload with 1000 child processes and a VMA with 1000 anonymous
pages per process that get COWed, this leads to a system with a million
anonymous pages in the same anon_vma, each of which is mapped in just one
of the 1000 processes. However, the current rmap code needs to walk them
all, leading to O(N) scanning complexity for each page.
This can result in systems where one CPU is walking the page tables of
1000 processes in page_referenced_one, while all other CPUs are stuck on
the anon_vma lock. This leads to catastrophic failure for a benchmark
like AIM7, where the total number of processes can reach in the tens of
thousands. Real workloads are still a factor 10 less process intensive
than AIM7, but they are catching up.
This patch changes the way anon_vmas and VMAs are linked, which allows us
to associate multiple anon_vmas with a VMA. At fork time, each child
process gets its own anon_vmas, in which its COWed pages will be
instantiated. The parents' anon_vma is also linked to the VMA, because
non-COWed pages could be present in any of the children.
This reduces rmap scanning complexity to O(1) for the pages of the 1000
child processes, with O(N) complexity for at most 1/N pages in the system.
This reduces the average scanning cost in heavily forking workloads from
O(N) to 2.
The only real complexity in this patch stems from the fact that linking a
VMA to anon_vmas now involves memory allocations. This means vma_adjust
can fail, if it needs to attach a VMA to anon_vma structures. This in
turn means error handling needs to be added to the calling functions.
A second source of complexity is that, because there can be multiple
anon_vmas, the anon_vma linking in vma_adjust can no longer be done under
"the" anon_vma lock. To prevent the rmap code from walking up an
incomplete VMA, this patch introduces the VM_LOCK_RMAP VMA flag. This bit
flag uses the same slot as the NOMMU VM_MAPPED_COPY, with an ifdef in mm.h
to make sure it is impossible to compile a kernel that needs both symbolic
values for the same bitflag.
Some test results:
Without the anon_vma changes, when AIM7 hits around 9.7k users (on a test
box with 16GB RAM and not quite enough IO), the system ends up running
>99% in system time, with every CPU on the same anon_vma lock in the
pageout code.
With these changes, AIM7 hits the cross-over point around 29.7k users.
This happens with ~99% IO wait time, there never seems to be any spike in
system time. The anon_vma lock contention appears to be resolved.
[akpm@linux-foundation.org: cleanups]
Signed-off-by: Rik van Riel <riel@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Larry Woodman <lwoodman@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-06 05:42:07 +08:00
|
|
|
return NULL;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
2005-05-01 23:58:35 +08:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Return true if the calling process may expand its vm space by the passed
|
|
|
|
|
* number of pages
|
|
|
|
|
*/
|
2016-01-15 07:22:07 +08:00
|
|
|
bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
|
2005-05-01 23:58:35 +08:00
|
|
|
{
|
2016-01-15 07:22:07 +08:00
|
|
|
if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
|
|
|
|
|
return false;
|
2005-05-01 23:58:35 +08:00
|
|
|
|
2016-02-03 08:57:43 +08:00
|
|
|
if (is_data_mapping(flags) &&
|
|
|
|
|
mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
|
2016-05-21 07:57:45 +08:00
|
|
|
/* Workaround for Valgrind */
|
|
|
|
|
if (rlimit(RLIMIT_DATA) == 0 &&
|
|
|
|
|
mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
|
|
|
|
|
return true;
|
2018-04-06 07:22:05 +08:00
|
|
|
|
|
|
|
|
pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
|
|
|
|
|
current->comm, current->pid,
|
|
|
|
|
(mm->data_vm + npages) << PAGE_SHIFT,
|
|
|
|
|
rlimit(RLIMIT_DATA),
|
|
|
|
|
ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
|
|
|
|
|
|
|
|
|
|
if (!ignore_rlimit_data)
|
2016-02-03 08:57:43 +08:00
|
|
|
return false;
|
|
|
|
|
}
|
2005-05-01 23:58:35 +08:00
|
|
|
|
2016-01-15 07:22:07 +08:00
|
|
|
return true;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
|
|
|
|
|
{
|
2021-11-06 04:38:12 +08:00
|
|
|
WRITE_ONCE(mm->total_vm, READ_ONCE(mm->total_vm)+npages);
|
2016-01-15 07:22:07 +08:00
|
|
|
|
2016-02-03 08:57:43 +08:00
|
|
|
if (is_exec_mapping(flags))
|
2016-01-15 07:22:07 +08:00
|
|
|
mm->exec_vm += npages;
|
2016-02-03 08:57:43 +08:00
|
|
|
else if (is_stack_mapping(flags))
|
2016-01-15 07:22:07 +08:00
|
|
|
mm->stack_vm += npages;
|
2016-02-03 08:57:43 +08:00
|
|
|
else if (is_data_mapping(flags))
|
2016-01-15 07:22:07 +08:00
|
|
|
mm->data_vm += npages;
|
2005-05-01 23:58:35 +08:00
|
|
|
}
|
2007-02-09 06:20:41 +08:00
|
|
|
|
2018-06-08 08:08:04 +08:00
|
|
|
static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
|
2014-05-20 06:58:33 +08:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Having a close hook prevents vma merging regardless of flags.
|
|
|
|
|
*/
|
|
|
|
|
static void special_mapping_close(struct vm_area_struct *vma)
|
|
|
|
|
{
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static const char *special_mapping_name(struct vm_area_struct *vma)
|
|
|
|
|
{
|
|
|
|
|
return ((struct vm_special_mapping *)vma->vm_private_data)->name;
|
|
|
|
|
}
|
|
|
|
|
|
2021-04-30 13:57:48 +08:00
|
|
|
static int special_mapping_mremap(struct vm_area_struct *new_vma)
|
2016-06-28 19:35:38 +08:00
|
|
|
{
|
|
|
|
|
struct vm_special_mapping *sm = new_vma->vm_private_data;
|
|
|
|
|
|
2017-06-20 00:32:42 +08:00
|
|
|
if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
|
|
|
|
|
return -EFAULT;
|
|
|
|
|
|
2016-06-28 19:35:38 +08:00
|
|
|
if (sm->mremap)
|
|
|
|
|
return sm->mremap(sm, new_vma);
|
2017-06-20 00:32:42 +08:00
|
|
|
|
2016-06-28 19:35:38 +08:00
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
2020-12-15 11:08:25 +08:00
|
|
|
static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr)
|
|
|
|
|
{
|
|
|
|
|
/*
|
|
|
|
|
* Forbid splitting special mappings - kernel has expectations over
|
|
|
|
|
* the number of pages in mapping. Together with VM_DONTEXPAND
|
|
|
|
|
* the size of vma should stay the same over the special mapping's
|
|
|
|
|
* lifetime.
|
|
|
|
|
*/
|
|
|
|
|
return -EINVAL;
|
|
|
|
|
}
|
|
|
|
|
|
2014-05-20 06:58:33 +08:00
|
|
|
static const struct vm_operations_struct special_mapping_vmops = {
|
|
|
|
|
.close = special_mapping_close,
|
|
|
|
|
.fault = special_mapping_fault,
|
2016-06-28 19:35:38 +08:00
|
|
|
.mremap = special_mapping_mremap,
|
2014-05-20 06:58:33 +08:00
|
|
|
.name = special_mapping_name,
|
2019-11-12 09:27:13 +08:00
|
|
|
/* vDSO code relies that VVAR can't be accessed remotely */
|
|
|
|
|
.access = NULL,
|
2020-12-15 11:08:25 +08:00
|
|
|
.may_split = special_mapping_split,
|
2014-05-20 06:58:33 +08:00
|
|
|
};
|
|
|
|
|
|
|
|
|
|
static const struct vm_operations_struct legacy_special_mapping_vmops = {
|
|
|
|
|
.close = special_mapping_close,
|
|
|
|
|
.fault = special_mapping_fault,
|
|
|
|
|
};
|
2007-02-09 06:20:41 +08:00
|
|
|
|
2018-06-08 08:08:04 +08:00
|
|
|
static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
|
2007-02-09 06:20:41 +08:00
|
|
|
{
|
2017-02-25 06:56:41 +08:00
|
|
|
struct vm_area_struct *vma = vmf->vma;
|
2008-02-09 08:15:19 +08:00
|
|
|
pgoff_t pgoff;
|
2007-02-09 06:20:41 +08:00
|
|
|
struct page **pages;
|
|
|
|
|
|
2015-12-30 12:12:19 +08:00
|
|
|
if (vma->vm_ops == &legacy_special_mapping_vmops) {
|
2014-05-20 06:58:33 +08:00
|
|
|
pages = vma->vm_private_data;
|
2015-12-30 12:12:19 +08:00
|
|
|
} else {
|
|
|
|
|
struct vm_special_mapping *sm = vma->vm_private_data;
|
|
|
|
|
|
|
|
|
|
if (sm->fault)
|
2017-02-25 06:56:41 +08:00
|
|
|
return sm->fault(sm, vmf->vma, vmf);
|
2015-12-30 12:12:19 +08:00
|
|
|
|
|
|
|
|
pages = sm->pages;
|
|
|
|
|
}
|
2014-05-20 06:58:33 +08:00
|
|
|
|
mmap: fix the usage of ->vm_pgoff in special_mapping paths
Test-case:
#include <stdio.h>
#include <unistd.h>
#include <stdlib.h>
#include <string.h>
#include <sys/mman.h>
#include <assert.h>
void *find_vdso_vaddr(void)
{
FILE *perl;
char buf[32] = {};
perl = popen("perl -e 'open STDIN,qq|/proc/@{[getppid]}/maps|;"
"/^(.*?)-.*vdso/ && print hex $1 while <>'", "r");
fread(buf, sizeof(buf), 1, perl);
fclose(perl);
return (void *)atol(buf);
}
#define PAGE_SIZE 4096
int main(void)
{
void *vdso = find_vdso_vaddr();
assert(vdso);
// of course they should differ, and they do so far
printf("vdso pages differ: %d\n",
!!memcmp(vdso, vdso + PAGE_SIZE, PAGE_SIZE));
// split into 2 vma's
assert(mprotect(vdso, PAGE_SIZE, PROT_READ) == 0);
// force another fault on the next check
assert(madvise(vdso, 2 * PAGE_SIZE, MADV_DONTNEED) == 0);
// now they no longer differ, the 2nd vm_pgoff is wrong
printf("vdso pages differ: %d\n",
!!memcmp(vdso, vdso + PAGE_SIZE, PAGE_SIZE));
return 0;
}
Output:
vdso pages differ: 1
vdso pages differ: 0
This is because split_vma() correctly updates ->vm_pgoff, but the logic
in insert_vm_struct() and special_mapping_fault() is absolutely broken,
so the fault at vdso + PAGE_SIZE return the 1st page. The same happens
if you simply unmap the 1st page.
special_mapping_fault() does:
pgoff = vmf->pgoff - vma->vm_pgoff;
and this is _only_ correct if vma->vm_start mmaps the first page from
->vm_private_data array.
vdso or any other user of install_special_mapping() is not anonymous,
it has the "backing storage" even if it is just the array of pages.
So we actually need to make vm_pgoff work as an offset in this array.
Note: this also allows to fix another problem: currently gdb can't access
"[vvar]" memory because in this case special_mapping_fault() doesn't work.
Now that we can use ->vm_pgoff we can implement ->access() and fix this.
Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-09 05:58:31 +08:00
|
|
|
for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
|
2008-02-09 08:15:19 +08:00
|
|
|
pgoff--;
|
2007-02-09 06:20:41 +08:00
|
|
|
|
|
|
|
|
if (*pages) {
|
|
|
|
|
struct page *page = *pages;
|
|
|
|
|
get_page(page);
|
2008-02-09 08:15:19 +08:00
|
|
|
vmf->page = page;
|
|
|
|
|
return 0;
|
2007-02-09 06:20:41 +08:00
|
|
|
}
|
|
|
|
|
|
2008-02-09 08:15:19 +08:00
|
|
|
return VM_FAULT_SIGBUS;
|
2007-02-09 06:20:41 +08:00
|
|
|
}
|
|
|
|
|
|
2014-05-20 06:58:33 +08:00
|
|
|
static struct vm_area_struct *__install_special_mapping(
|
|
|
|
|
struct mm_struct *mm,
|
|
|
|
|
unsigned long addr, unsigned long len,
|
2015-11-06 10:48:41 +08:00
|
|
|
unsigned long vm_flags, void *priv,
|
|
|
|
|
const struct vm_operations_struct *ops)
|
2007-02-09 06:20:41 +08:00
|
|
|
{
|
2010-12-09 22:29:42 +08:00
|
|
|
int ret;
|
2007-02-09 06:20:41 +08:00
|
|
|
struct vm_area_struct *vma;
|
|
|
|
|
|
2018-07-22 06:24:03 +08:00
|
|
|
vma = vm_area_alloc(mm);
|
2007-02-09 06:20:41 +08:00
|
|
|
if (unlikely(vma == NULL))
|
2014-03-18 06:22:02 +08:00
|
|
|
return ERR_PTR(-ENOMEM);
|
2007-02-09 06:20:41 +08:00
|
|
|
|
|
|
|
|
vma->vm_start = addr;
|
|
|
|
|
vma->vm_end = addr + len;
|
|
|
|
|
|
2013-09-12 05:22:24 +08:00
|
|
|
vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
|
2022-03-23 05:42:36 +08:00
|
|
|
vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
|
2007-10-19 14:39:15 +08:00
|
|
|
vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
|
2007-02-09 06:20:41 +08:00
|
|
|
|
2014-05-20 06:58:33 +08:00
|
|
|
vma->vm_ops = ops;
|
|
|
|
|
vma->vm_private_data = priv;
|
2007-02-09 06:20:41 +08:00
|
|
|
|
2010-12-09 22:29:42 +08:00
|
|
|
ret = insert_vm_struct(mm, vma);
|
|
|
|
|
if (ret)
|
|
|
|
|
goto out;
|
2007-02-09 06:20:41 +08:00
|
|
|
|
2016-01-15 07:22:07 +08:00
|
|
|
vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
|
2007-02-09 06:20:41 +08:00
|
|
|
|
perf: Do the big rename: Performance Counters -> Performance Events
Bye-bye Performance Counters, welcome Performance Events!
In the past few months the perfcounters subsystem has grown out its
initial role of counting hardware events, and has become (and is
becoming) a much broader generic event enumeration, reporting, logging,
monitoring, analysis facility.
Naming its core object 'perf_counter' and naming the subsystem
'perfcounters' has become more and more of a misnomer. With pending
code like hw-breakpoints support the 'counter' name is less and
less appropriate.
All in one, we've decided to rename the subsystem to 'performance
events' and to propagate this rename through all fields, variables
and API names. (in an ABI compatible fashion)
The word 'event' is also a bit shorter than 'counter' - which makes
it slightly more convenient to write/handle as well.
Thanks goes to Stephane Eranian who first observed this misnomer and
suggested a rename.
User-space tooling and ABI compatibility is not affected - this patch
should be function-invariant. (Also, defconfigs were not touched to
keep the size down.)
This patch has been generated via the following script:
FILES=$(find * -type f | grep -vE 'oprofile|[^K]config')
sed -i \
-e 's/PERF_EVENT_/PERF_RECORD_/g' \
-e 's/PERF_COUNTER/PERF_EVENT/g' \
-e 's/perf_counter/perf_event/g' \
-e 's/nb_counters/nb_events/g' \
-e 's/swcounter/swevent/g' \
-e 's/tpcounter_event/tp_event/g' \
$FILES
for N in $(find . -name perf_counter.[ch]); do
M=$(echo $N | sed 's/perf_counter/perf_event/g')
mv $N $M
done
FILES=$(find . -name perf_event.*)
sed -i \
-e 's/COUNTER_MASK/REG_MASK/g' \
-e 's/COUNTER/EVENT/g' \
-e 's/\<event\>/event_id/g' \
-e 's/counter/event/g' \
-e 's/Counter/Event/g' \
$FILES
... to keep it as correct as possible. This script can also be
used by anyone who has pending perfcounters patches - it converts
a Linux kernel tree over to the new naming. We tried to time this
change to the point in time where the amount of pending patches
is the smallest: the end of the merge window.
Namespace clashes were fixed up in a preparatory patch - and some
stylistic fallout will be fixed up in a subsequent patch.
( NOTE: 'counters' are still the proper terminology when we deal
with hardware registers - and these sed scripts are a bit
over-eager in renaming them. I've undone some of that, but
in case there's something left where 'counter' would be
better than 'event' we can undo that on an individual basis
instead of touching an otherwise nicely automated patch. )
Suggested-by: Stephane Eranian <eranian@google.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Paul Mackerras <paulus@samba.org>
Reviewed-by: Arjan van de Ven <arjan@linux.intel.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: David Howells <dhowells@redhat.com>
Cc: Kyle McMartin <kyle@mcmartin.ca>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: <linux-arch@vger.kernel.org>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-09-21 18:02:48 +08:00
|
|
|
perf_event_mmap(vma);
|
2009-06-05 20:04:55 +08:00
|
|
|
|
2014-03-18 06:22:02 +08:00
|
|
|
return vma;
|
2010-12-09 22:29:42 +08:00
|
|
|
|
|
|
|
|
out:
|
2018-07-22 04:48:51 +08:00
|
|
|
vm_area_free(vma);
|
2014-03-18 06:22:02 +08:00
|
|
|
return ERR_PTR(ret);
|
|
|
|
|
}
|
|
|
|
|
|
2016-09-05 21:33:05 +08:00
|
|
|
bool vma_is_special_mapping(const struct vm_area_struct *vma,
|
|
|
|
|
const struct vm_special_mapping *sm)
|
|
|
|
|
{
|
|
|
|
|
return vma->vm_private_data == sm &&
|
|
|
|
|
(vma->vm_ops == &special_mapping_vmops ||
|
|
|
|
|
vma->vm_ops == &legacy_special_mapping_vmops);
|
|
|
|
|
}
|
|
|
|
|
|
2014-05-20 06:58:33 +08:00
|
|
|
/*
|
2020-06-09 12:33:54 +08:00
|
|
|
* Called with mm->mmap_lock held for writing.
|
2014-05-20 06:58:33 +08:00
|
|
|
* Insert a new vma covering the given region, with the given flags.
|
|
|
|
|
* Its pages are supplied by the given array of struct page *.
|
|
|
|
|
* The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
|
|
|
|
|
* The region past the last page supplied will always produce SIGBUS.
|
|
|
|
|
* The array pointer and the pages it points to are assumed to stay alive
|
|
|
|
|
* for as long as this mapping might exist.
|
|
|
|
|
*/
|
|
|
|
|
struct vm_area_struct *_install_special_mapping(
|
|
|
|
|
struct mm_struct *mm,
|
|
|
|
|
unsigned long addr, unsigned long len,
|
|
|
|
|
unsigned long vm_flags, const struct vm_special_mapping *spec)
|
|
|
|
|
{
|
2015-11-06 10:48:41 +08:00
|
|
|
return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
|
|
|
|
|
&special_mapping_vmops);
|
2014-05-20 06:58:33 +08:00
|
|
|
}
|
|
|
|
|
|
2014-03-18 06:22:02 +08:00
|
|
|
int install_special_mapping(struct mm_struct *mm,
|
|
|
|
|
unsigned long addr, unsigned long len,
|
|
|
|
|
unsigned long vm_flags, struct page **pages)
|
|
|
|
|
{
|
2014-05-20 06:58:33 +08:00
|
|
|
struct vm_area_struct *vma = __install_special_mapping(
|
2015-11-06 10:48:41 +08:00
|
|
|
mm, addr, len, vm_flags, (void *)pages,
|
|
|
|
|
&legacy_special_mapping_vmops);
|
2014-03-18 06:22:02 +08:00
|
|
|
|
2014-06-05 07:07:05 +08:00
|
|
|
return PTR_ERR_OR_ZERO(vma);
|
2007-02-09 06:20:41 +08:00
|
|
|
}
|
2008-07-29 06:46:26 +08:00
|
|
|
|
|
|
|
|
static DEFINE_MUTEX(mm_all_locks_mutex);
|
|
|
|
|
|
2008-08-11 15:30:25 +08:00
|
|
|
static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
|
2008-07-29 06:46:26 +08:00
|
|
|
{
|
2017-09-09 07:15:08 +08:00
|
|
|
if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
|
2008-07-29 06:46:26 +08:00
|
|
|
/*
|
|
|
|
|
* The LSB of head.next can't change from under us
|
|
|
|
|
* because we hold the mm_all_locks_mutex.
|
|
|
|
|
*/
|
2020-06-09 12:33:47 +08:00
|
|
|
down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
|
2008-07-29 06:46:26 +08:00
|
|
|
/*
|
|
|
|
|
* We can safely modify head.next after taking the
|
2012-12-03 03:56:46 +08:00
|
|
|
* anon_vma->root->rwsem. If some other vma in this mm shares
|
2008-07-29 06:46:26 +08:00
|
|
|
* the same anon_vma we won't take it again.
|
|
|
|
|
*
|
|
|
|
|
* No need of atomic instructions here, head.next
|
|
|
|
|
* can't change from under us thanks to the
|
2012-12-03 03:56:46 +08:00
|
|
|
* anon_vma->root->rwsem.
|
2008-07-29 06:46:26 +08:00
|
|
|
*/
|
|
|
|
|
if (__test_and_set_bit(0, (unsigned long *)
|
2017-09-09 07:15:08 +08:00
|
|
|
&anon_vma->root->rb_root.rb_root.rb_node))
|
2008-07-29 06:46:26 +08:00
|
|
|
BUG();
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2008-08-11 15:30:25 +08:00
|
|
|
static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
|
2008-07-29 06:46:26 +08:00
|
|
|
{
|
|
|
|
|
if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
|
|
|
|
|
/*
|
|
|
|
|
* AS_MM_ALL_LOCKS can't change from under us because
|
|
|
|
|
* we hold the mm_all_locks_mutex.
|
|
|
|
|
*
|
|
|
|
|
* Operations on ->flags have to be atomic because
|
|
|
|
|
* even if AS_MM_ALL_LOCKS is stable thanks to the
|
|
|
|
|
* mm_all_locks_mutex, there may be other cpus
|
|
|
|
|
* changing other bitflags in parallel to us.
|
|
|
|
|
*/
|
|
|
|
|
if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
|
|
|
|
|
BUG();
|
2020-06-09 12:33:47 +08:00
|
|
|
down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
|
2008-07-29 06:46:26 +08:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* This operation locks against the VM for all pte/vma/mm related
|
|
|
|
|
* operations that could ever happen on a certain mm. This includes
|
|
|
|
|
* vmtruncate, try_to_unmap, and all page faults.
|
|
|
|
|
*
|
2020-06-09 12:33:54 +08:00
|
|
|
* The caller must take the mmap_lock in write mode before calling
|
2008-07-29 06:46:26 +08:00
|
|
|
* mm_take_all_locks(). The caller isn't allowed to release the
|
2020-06-09 12:33:54 +08:00
|
|
|
* mmap_lock until mm_drop_all_locks() returns.
|
2008-07-29 06:46:26 +08:00
|
|
|
*
|
2020-06-09 12:33:54 +08:00
|
|
|
* mmap_lock in write mode is required in order to block all operations
|
2008-07-29 06:46:26 +08:00
|
|
|
* that could modify pagetables and free pages without need of
|
2015-02-11 06:09:59 +08:00
|
|
|
* altering the vma layout. It's also needed in write mode to avoid new
|
2008-07-29 06:46:26 +08:00
|
|
|
* anon_vmas to be associated with existing vmas.
|
|
|
|
|
*
|
|
|
|
|
* A single task can't take more than one mm_take_all_locks() in a row
|
|
|
|
|
* or it would deadlock.
|
|
|
|
|
*
|
mm anon rmap: replace same_anon_vma linked list with an interval tree.
When a large VMA (anon or private file mapping) is first touched, which
will populate its anon_vma field, and then split into many regions through
the use of mprotect(), the original anon_vma ends up linking all of the
vmas on a linked list. This can cause rmap to become inefficient, as we
have to walk potentially thousands of irrelevent vmas before finding the
one a given anon page might fall into.
By replacing the same_anon_vma linked list with an interval tree (where
each avc's interval is determined by its vma's start and last pgoffs), we
can make rmap efficient for this use case again.
While the change is large, all of its pieces are fairly simple.
Most places that were walking the same_anon_vma list were looking for a
known pgoff, so they can just use the anon_vma_interval_tree_foreach()
interval tree iterator instead. The exception here is ksm, where the
page's index is not known. It would probably be possible to rework ksm so
that the index would be known, but for now I have decided to keep things
simple and just walk the entirety of the interval tree there.
When updating vma's that already have an anon_vma assigned, we must take
care to re-index the corresponding avc's on their interval tree. This is
done through the use of anon_vma_interval_tree_pre_update_vma() and
anon_vma_interval_tree_post_update_vma(), which remove the avc's from
their interval tree before the update and re-insert them after the update.
The anon_vma stays locked during the update, so there is no chance that
rmap would miss the vmas that are being updated.
Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-10-09 07:31:39 +08:00
|
|
|
* The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
|
2008-07-29 06:46:26 +08:00
|
|
|
* mapping->flags avoid to take the same lock twice, if more than one
|
|
|
|
|
* vma in this mm is backed by the same anon_vma or address_space.
|
|
|
|
|
*
|
2016-01-16 08:57:31 +08:00
|
|
|
* We take locks in following order, accordingly to comment at beginning
|
|
|
|
|
* of mm/rmap.c:
|
|
|
|
|
* - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
|
|
|
|
|
* hugetlb mapping);
|
|
|
|
|
* - all i_mmap_rwsem locks;
|
|
|
|
|
* - all anon_vma->rwseml
|
|
|
|
|
*
|
|
|
|
|
* We can take all locks within these types randomly because the VM code
|
|
|
|
|
* doesn't nest them and we protected from parallel mm_take_all_locks() by
|
|
|
|
|
* mm_all_locks_mutex.
|
2008-07-29 06:46:26 +08:00
|
|
|
*
|
|
|
|
|
* mm_take_all_locks() and mm_drop_all_locks are expensive operations
|
|
|
|
|
* that may have to take thousand of locks.
|
|
|
|
|
*
|
|
|
|
|
* mm_take_all_locks() can fail if it's interrupted by signals.
|
|
|
|
|
*/
|
|
|
|
|
int mm_take_all_locks(struct mm_struct *mm)
|
|
|
|
|
{
|
|
|
|
|
struct vm_area_struct *vma;
|
mm: change anon_vma linking to fix multi-process server scalability issue
The old anon_vma code can lead to scalability issues with heavily forking
workloads. Specifically, each anon_vma will be shared between the parent
process and all its child processes.
In a workload with 1000 child processes and a VMA with 1000 anonymous
pages per process that get COWed, this leads to a system with a million
anonymous pages in the same anon_vma, each of which is mapped in just one
of the 1000 processes. However, the current rmap code needs to walk them
all, leading to O(N) scanning complexity for each page.
This can result in systems where one CPU is walking the page tables of
1000 processes in page_referenced_one, while all other CPUs are stuck on
the anon_vma lock. This leads to catastrophic failure for a benchmark
like AIM7, where the total number of processes can reach in the tens of
thousands. Real workloads are still a factor 10 less process intensive
than AIM7, but they are catching up.
This patch changes the way anon_vmas and VMAs are linked, which allows us
to associate multiple anon_vmas with a VMA. At fork time, each child
process gets its own anon_vmas, in which its COWed pages will be
instantiated. The parents' anon_vma is also linked to the VMA, because
non-COWed pages could be present in any of the children.
This reduces rmap scanning complexity to O(1) for the pages of the 1000
child processes, with O(N) complexity for at most 1/N pages in the system.
This reduces the average scanning cost in heavily forking workloads from
O(N) to 2.
The only real complexity in this patch stems from the fact that linking a
VMA to anon_vmas now involves memory allocations. This means vma_adjust
can fail, if it needs to attach a VMA to anon_vma structures. This in
turn means error handling needs to be added to the calling functions.
A second source of complexity is that, because there can be multiple
anon_vmas, the anon_vma linking in vma_adjust can no longer be done under
"the" anon_vma lock. To prevent the rmap code from walking up an
incomplete VMA, this patch introduces the VM_LOCK_RMAP VMA flag. This bit
flag uses the same slot as the NOMMU VM_MAPPED_COPY, with an ifdef in mm.h
to make sure it is impossible to compile a kernel that needs both symbolic
values for the same bitflag.
Some test results:
Without the anon_vma changes, when AIM7 hits around 9.7k users (on a test
box with 16GB RAM and not quite enough IO), the system ends up running
>99% in system time, with every CPU on the same anon_vma lock in the
pageout code.
With these changes, AIM7 hits the cross-over point around 29.7k users.
This happens with ~99% IO wait time, there never seems to be any spike in
system time. The anon_vma lock contention appears to be resolved.
[akpm@linux-foundation.org: cleanups]
Signed-off-by: Rik van Riel <riel@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Larry Woodman <lwoodman@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-06 05:42:07 +08:00
|
|
|
struct anon_vma_chain *avc;
|
2008-07-29 06:46:26 +08:00
|
|
|
|
2020-06-09 12:33:25 +08:00
|
|
|
BUG_ON(mmap_read_trylock(mm));
|
2008-07-29 06:46:26 +08:00
|
|
|
|
|
|
|
|
mutex_lock(&mm_all_locks_mutex);
|
|
|
|
|
|
|
|
|
|
for (vma = mm->mmap; vma; vma = vma->vm_next) {
|
|
|
|
|
if (signal_pending(current))
|
|
|
|
|
goto out_unlock;
|
2016-01-16 08:57:31 +08:00
|
|
|
if (vma->vm_file && vma->vm_file->f_mapping &&
|
|
|
|
|
is_vm_hugetlb_page(vma))
|
|
|
|
|
vm_lock_mapping(mm, vma->vm_file->f_mapping);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
for (vma = mm->mmap; vma; vma = vma->vm_next) {
|
|
|
|
|
if (signal_pending(current))
|
|
|
|
|
goto out_unlock;
|
|
|
|
|
if (vma->vm_file && vma->vm_file->f_mapping &&
|
|
|
|
|
!is_vm_hugetlb_page(vma))
|
2008-08-11 15:30:25 +08:00
|
|
|
vm_lock_mapping(mm, vma->vm_file->f_mapping);
|
2008-07-29 06:46:26 +08:00
|
|
|
}
|
2008-08-11 15:30:25 +08:00
|
|
|
|
|
|
|
|
for (vma = mm->mmap; vma; vma = vma->vm_next) {
|
|
|
|
|
if (signal_pending(current))
|
|
|
|
|
goto out_unlock;
|
|
|
|
|
if (vma->anon_vma)
|
mm: change anon_vma linking to fix multi-process server scalability issue
The old anon_vma code can lead to scalability issues with heavily forking
workloads. Specifically, each anon_vma will be shared between the parent
process and all its child processes.
In a workload with 1000 child processes and a VMA with 1000 anonymous
pages per process that get COWed, this leads to a system with a million
anonymous pages in the same anon_vma, each of which is mapped in just one
of the 1000 processes. However, the current rmap code needs to walk them
all, leading to O(N) scanning complexity for each page.
This can result in systems where one CPU is walking the page tables of
1000 processes in page_referenced_one, while all other CPUs are stuck on
the anon_vma lock. This leads to catastrophic failure for a benchmark
like AIM7, where the total number of processes can reach in the tens of
thousands. Real workloads are still a factor 10 less process intensive
than AIM7, but they are catching up.
This patch changes the way anon_vmas and VMAs are linked, which allows us
to associate multiple anon_vmas with a VMA. At fork time, each child
process gets its own anon_vmas, in which its COWed pages will be
instantiated. The parents' anon_vma is also linked to the VMA, because
non-COWed pages could be present in any of the children.
This reduces rmap scanning complexity to O(1) for the pages of the 1000
child processes, with O(N) complexity for at most 1/N pages in the system.
This reduces the average scanning cost in heavily forking workloads from
O(N) to 2.
The only real complexity in this patch stems from the fact that linking a
VMA to anon_vmas now involves memory allocations. This means vma_adjust
can fail, if it needs to attach a VMA to anon_vma structures. This in
turn means error handling needs to be added to the calling functions.
A second source of complexity is that, because there can be multiple
anon_vmas, the anon_vma linking in vma_adjust can no longer be done under
"the" anon_vma lock. To prevent the rmap code from walking up an
incomplete VMA, this patch introduces the VM_LOCK_RMAP VMA flag. This bit
flag uses the same slot as the NOMMU VM_MAPPED_COPY, with an ifdef in mm.h
to make sure it is impossible to compile a kernel that needs both symbolic
values for the same bitflag.
Some test results:
Without the anon_vma changes, when AIM7 hits around 9.7k users (on a test
box with 16GB RAM and not quite enough IO), the system ends up running
>99% in system time, with every CPU on the same anon_vma lock in the
pageout code.
With these changes, AIM7 hits the cross-over point around 29.7k users.
This happens with ~99% IO wait time, there never seems to be any spike in
system time. The anon_vma lock contention appears to be resolved.
[akpm@linux-foundation.org: cleanups]
Signed-off-by: Rik van Riel <riel@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Larry Woodman <lwoodman@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-06 05:42:07 +08:00
|
|
|
list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
|
|
|
|
|
vm_lock_anon_vma(mm, avc->anon_vma);
|
2008-07-29 06:46:26 +08:00
|
|
|
}
|
2008-08-11 15:30:25 +08:00
|
|
|
|
2011-11-01 08:08:59 +08:00
|
|
|
return 0;
|
2008-07-29 06:46:26 +08:00
|
|
|
|
|
|
|
|
out_unlock:
|
2011-11-01 08:08:59 +08:00
|
|
|
mm_drop_all_locks(mm);
|
|
|
|
|
return -EINTR;
|
2008-07-29 06:46:26 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
|
|
|
|
|
{
|
2017-09-09 07:15:08 +08:00
|
|
|
if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
|
2008-07-29 06:46:26 +08:00
|
|
|
/*
|
|
|
|
|
* The LSB of head.next can't change to 0 from under
|
|
|
|
|
* us because we hold the mm_all_locks_mutex.
|
|
|
|
|
*
|
|
|
|
|
* We must however clear the bitflag before unlocking
|
mm anon rmap: replace same_anon_vma linked list with an interval tree.
When a large VMA (anon or private file mapping) is first touched, which
will populate its anon_vma field, and then split into many regions through
the use of mprotect(), the original anon_vma ends up linking all of the
vmas on a linked list. This can cause rmap to become inefficient, as we
have to walk potentially thousands of irrelevent vmas before finding the
one a given anon page might fall into.
By replacing the same_anon_vma linked list with an interval tree (where
each avc's interval is determined by its vma's start and last pgoffs), we
can make rmap efficient for this use case again.
While the change is large, all of its pieces are fairly simple.
Most places that were walking the same_anon_vma list were looking for a
known pgoff, so they can just use the anon_vma_interval_tree_foreach()
interval tree iterator instead. The exception here is ksm, where the
page's index is not known. It would probably be possible to rework ksm so
that the index would be known, but for now I have decided to keep things
simple and just walk the entirety of the interval tree there.
When updating vma's that already have an anon_vma assigned, we must take
care to re-index the corresponding avc's on their interval tree. This is
done through the use of anon_vma_interval_tree_pre_update_vma() and
anon_vma_interval_tree_post_update_vma(), which remove the avc's from
their interval tree before the update and re-insert them after the update.
The anon_vma stays locked during the update, so there is no chance that
rmap would miss the vmas that are being updated.
Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Daniel Santos <daniel.santos@pobox.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-10-09 07:31:39 +08:00
|
|
|
* the vma so the users using the anon_vma->rb_root will
|
2008-07-29 06:46:26 +08:00
|
|
|
* never see our bitflag.
|
|
|
|
|
*
|
|
|
|
|
* No need of atomic instructions here, head.next
|
|
|
|
|
* can't change from under us until we release the
|
2012-12-03 03:56:46 +08:00
|
|
|
* anon_vma->root->rwsem.
|
2008-07-29 06:46:26 +08:00
|
|
|
*/
|
|
|
|
|
if (!__test_and_clear_bit(0, (unsigned long *)
|
2017-09-09 07:15:08 +08:00
|
|
|
&anon_vma->root->rb_root.rb_root.rb_node))
|
2008-07-29 06:46:26 +08:00
|
|
|
BUG();
|
2013-02-23 08:34:40 +08:00
|
|
|
anon_vma_unlock_write(anon_vma);
|
2008-07-29 06:46:26 +08:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void vm_unlock_mapping(struct address_space *mapping)
|
|
|
|
|
{
|
|
|
|
|
if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
|
|
|
|
|
/*
|
|
|
|
|
* AS_MM_ALL_LOCKS can't change to 0 from under us
|
|
|
|
|
* because we hold the mm_all_locks_mutex.
|
|
|
|
|
*/
|
2014-12-13 08:54:21 +08:00
|
|
|
i_mmap_unlock_write(mapping);
|
2008-07-29 06:46:26 +08:00
|
|
|
if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
|
|
|
|
|
&mapping->flags))
|
|
|
|
|
BUG();
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
2020-06-09 12:33:54 +08:00
|
|
|
* The mmap_lock cannot be released by the caller until
|
2008-07-29 06:46:26 +08:00
|
|
|
* mm_drop_all_locks() returns.
|
|
|
|
|
*/
|
|
|
|
|
void mm_drop_all_locks(struct mm_struct *mm)
|
|
|
|
|
{
|
|
|
|
|
struct vm_area_struct *vma;
|
mm: change anon_vma linking to fix multi-process server scalability issue
The old anon_vma code can lead to scalability issues with heavily forking
workloads. Specifically, each anon_vma will be shared between the parent
process and all its child processes.
In a workload with 1000 child processes and a VMA with 1000 anonymous
pages per process that get COWed, this leads to a system with a million
anonymous pages in the same anon_vma, each of which is mapped in just one
of the 1000 processes. However, the current rmap code needs to walk them
all, leading to O(N) scanning complexity for each page.
This can result in systems where one CPU is walking the page tables of
1000 processes in page_referenced_one, while all other CPUs are stuck on
the anon_vma lock. This leads to catastrophic failure for a benchmark
like AIM7, where the total number of processes can reach in the tens of
thousands. Real workloads are still a factor 10 less process intensive
than AIM7, but they are catching up.
This patch changes the way anon_vmas and VMAs are linked, which allows us
to associate multiple anon_vmas with a VMA. At fork time, each child
process gets its own anon_vmas, in which its COWed pages will be
instantiated. The parents' anon_vma is also linked to the VMA, because
non-COWed pages could be present in any of the children.
This reduces rmap scanning complexity to O(1) for the pages of the 1000
child processes, with O(N) complexity for at most 1/N pages in the system.
This reduces the average scanning cost in heavily forking workloads from
O(N) to 2.
The only real complexity in this patch stems from the fact that linking a
VMA to anon_vmas now involves memory allocations. This means vma_adjust
can fail, if it needs to attach a VMA to anon_vma structures. This in
turn means error handling needs to be added to the calling functions.
A second source of complexity is that, because there can be multiple
anon_vmas, the anon_vma linking in vma_adjust can no longer be done under
"the" anon_vma lock. To prevent the rmap code from walking up an
incomplete VMA, this patch introduces the VM_LOCK_RMAP VMA flag. This bit
flag uses the same slot as the NOMMU VM_MAPPED_COPY, with an ifdef in mm.h
to make sure it is impossible to compile a kernel that needs both symbolic
values for the same bitflag.
Some test results:
Without the anon_vma changes, when AIM7 hits around 9.7k users (on a test
box with 16GB RAM and not quite enough IO), the system ends up running
>99% in system time, with every CPU on the same anon_vma lock in the
pageout code.
With these changes, AIM7 hits the cross-over point around 29.7k users.
This happens with ~99% IO wait time, there never seems to be any spike in
system time. The anon_vma lock contention appears to be resolved.
[akpm@linux-foundation.org: cleanups]
Signed-off-by: Rik van Riel <riel@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Larry Woodman <lwoodman@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-06 05:42:07 +08:00
|
|
|
struct anon_vma_chain *avc;
|
2008-07-29 06:46:26 +08:00
|
|
|
|
2020-06-09 12:33:25 +08:00
|
|
|
BUG_ON(mmap_read_trylock(mm));
|
2008-07-29 06:46:26 +08:00
|
|
|
BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
|
|
|
|
|
|
|
|
|
|
for (vma = mm->mmap; vma; vma = vma->vm_next) {
|
|
|
|
|
if (vma->anon_vma)
|
mm: change anon_vma linking to fix multi-process server scalability issue
The old anon_vma code can lead to scalability issues with heavily forking
workloads. Specifically, each anon_vma will be shared between the parent
process and all its child processes.
In a workload with 1000 child processes and a VMA with 1000 anonymous
pages per process that get COWed, this leads to a system with a million
anonymous pages in the same anon_vma, each of which is mapped in just one
of the 1000 processes. However, the current rmap code needs to walk them
all, leading to O(N) scanning complexity for each page.
This can result in systems where one CPU is walking the page tables of
1000 processes in page_referenced_one, while all other CPUs are stuck on
the anon_vma lock. This leads to catastrophic failure for a benchmark
like AIM7, where the total number of processes can reach in the tens of
thousands. Real workloads are still a factor 10 less process intensive
than AIM7, but they are catching up.
This patch changes the way anon_vmas and VMAs are linked, which allows us
to associate multiple anon_vmas with a VMA. At fork time, each child
process gets its own anon_vmas, in which its COWed pages will be
instantiated. The parents' anon_vma is also linked to the VMA, because
non-COWed pages could be present in any of the children.
This reduces rmap scanning complexity to O(1) for the pages of the 1000
child processes, with O(N) complexity for at most 1/N pages in the system.
This reduces the average scanning cost in heavily forking workloads from
O(N) to 2.
The only real complexity in this patch stems from the fact that linking a
VMA to anon_vmas now involves memory allocations. This means vma_adjust
can fail, if it needs to attach a VMA to anon_vma structures. This in
turn means error handling needs to be added to the calling functions.
A second source of complexity is that, because there can be multiple
anon_vmas, the anon_vma linking in vma_adjust can no longer be done under
"the" anon_vma lock. To prevent the rmap code from walking up an
incomplete VMA, this patch introduces the VM_LOCK_RMAP VMA flag. This bit
flag uses the same slot as the NOMMU VM_MAPPED_COPY, with an ifdef in mm.h
to make sure it is impossible to compile a kernel that needs both symbolic
values for the same bitflag.
Some test results:
Without the anon_vma changes, when AIM7 hits around 9.7k users (on a test
box with 16GB RAM and not quite enough IO), the system ends up running
>99% in system time, with every CPU on the same anon_vma lock in the
pageout code.
With these changes, AIM7 hits the cross-over point around 29.7k users.
This happens with ~99% IO wait time, there never seems to be any spike in
system time. The anon_vma lock contention appears to be resolved.
[akpm@linux-foundation.org: cleanups]
Signed-off-by: Rik van Riel <riel@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Larry Woodman <lwoodman@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-06 05:42:07 +08:00
|
|
|
list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
|
|
|
|
|
vm_unlock_anon_vma(avc->anon_vma);
|
2008-07-29 06:46:26 +08:00
|
|
|
if (vma->vm_file && vma->vm_file->f_mapping)
|
|
|
|
|
vm_unlock_mapping(vma->vm_file->f_mapping);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
mutex_unlock(&mm_all_locks_mutex);
|
|
|
|
|
}
|
2009-01-08 20:04:47 +08:00
|
|
|
|
|
|
|
|
/*
|
2017-02-25 06:56:44 +08:00
|
|
|
* initialise the percpu counter for VM
|
2009-01-08 20:04:47 +08:00
|
|
|
*/
|
|
|
|
|
void __init mmap_init(void)
|
|
|
|
|
{
|
2009-05-01 06:08:51 +08:00
|
|
|
int ret;
|
|
|
|
|
|
2014-09-08 08:51:29 +08:00
|
|
|
ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
|
2009-05-01 06:08:51 +08:00
|
|
|
VM_BUG_ON(ret);
|
2009-01-08 20:04:47 +08:00
|
|
|
}
|
mm: limit growth of 3% hardcoded other user reserve
Add user_reserve_kbytes knob.
Limit the growth of the memory reserved for other user processes to
min(3% current process size, user_reserve_pages). Only about 8MB is
necessary to enable recovery in the default mode, and only a few hundred
MB are required even when overcommit is disabled.
user_reserve_pages defaults to min(3% free pages, 128MB)
I arrived at 128MB by taking the max VSZ of sshd, login, bash, and top ...
then adding the RSS of each.
This only affects OVERCOMMIT_NEVER mode.
Background
1. user reserve
__vm_enough_memory reserves a hardcoded 3% of the current process size for
other applications when overcommit is disabled. This was done so that a
user could recover if they launched a memory hogging process. Without the
reserve, a user would easily run into a message such as:
bash: fork: Cannot allocate memory
2. admin reserve
Additionally, a hardcoded 3% of free memory is reserved for root in both
overcommit 'guess' and 'never' modes. This was intended to prevent a
scenario where root-cant-log-in and perform recovery operations.
Note that this reserve shrinks, and doesn't guarantee a useful reserve.
Motivation
The two hardcoded memory reserves should be updated to account for current
memory sizes.
Also, the admin reserve would be more useful if it didn't shrink too much.
When the current code was originally written, 1GB was considered
"enterprise". Now the 3% reserve can grow to multiple GB on large memory
systems, and it only needs to be a few hundred MB at most to enable a user
or admin to recover a system with an unwanted memory hogging process.
I've found that reducing these reserves is especially beneficial for a
specific type of application load:
* single application system
* one or few processes (e.g. one per core)
* allocating all available memory
* not initializing every page immediately
* long running
I've run scientific clusters with this sort of load. A long running job
sometimes failed many hours (weeks of CPU time) into a calculation. They
weren't initializing all of their memory immediately, and they weren't
using calloc, so I put systems into overcommit 'never' mode. These
clusters run diskless and have no swap.
However, with the current reserves, a user wishing to allocate as much
memory as possible to one process may be prevented from using, for
example, almost 2GB out of 32GB.
The effect is less, but still significant when a user starts a job with
one process per core. I have repeatedly seen a set of processes
requesting the same amount of memory fail because one of them could not
allocate the amount of memory a user would expect to be able to allocate.
For example, Message Passing Interfce (MPI) processes, one per core. And
it is similar for other parallel programming frameworks.
Changing this reserve code will make the overcommit never mode more useful
by allowing applications to allocate nearly all of the available memory.
Also, the new admin_reserve_kbytes will be safer than the current behavior
since the hardcoded 3% of available memory reserve can shrink to something
useless in the case where applications have grabbed all available memory.
Risks
* "bash: fork: Cannot allocate memory"
The downside of the first patch-- which creates a tunable user reserve
that is only used in overcommit 'never' mode--is that an admin can set
it so low that a user may not be able to kill their process, even if
they already have a shell prompt.
Of course, a user can get in the same predicament with the current 3%
reserve--they just have to launch processes until 3% becomes negligible.
* root-cant-log-in problem
The second patch, adding the tunable rootuser_reserve_pages, allows
the admin to shoot themselves in the foot by setting it too small. They
can easily get the system into a state where root-can't-log-in.
However, the new admin_reserve_kbytes will be safer than the current
behavior since the hardcoded 3% of available memory reserve can shrink
to something useless in the case where applications have grabbed all
available memory.
Alternatives
* Memory cgroups provide a more flexible way to limit application memory.
Not everyone wants to set up cgroups or deal with their overhead.
* We could create a fourth overcommit mode which provides smaller reserves.
The size of useful reserves may be drastically different depending
on the whether the system is embedded or enterprise.
* Force users to initialize all of their memory or use calloc.
Some users don't want/expect the system to overcommit when they malloc.
Overcommit 'never' mode is for this scenario, and it should work well.
The new user and admin reserve tunables are simple to use, with low
overhead compared to cgroups. The patches preserve current behavior where
3% of memory is less than 128MB, except that the admin reserve doesn't
shrink to an unusable size under pressure. The code allows admins to tune
for embedded and enterprise usage.
FAQ
* How is the root-cant-login problem addressed?
What happens if admin_reserve_pages is set to 0?
Root is free to shoot themselves in the foot by setting
admin_reserve_kbytes too low.
On x86_64, the minimum useful reserve is:
8MB for overcommit 'guess'
128MB for overcommit 'never'
admin_reserve_pages defaults to min(3% free memory, 8MB)
So, anyone switching to 'never' mode needs to adjust
admin_reserve_pages.
* How do you calculate a minimum useful reserve?
A user or the admin needs enough memory to login and perform
recovery operations, which includes, at a minimum:
sshd or login + bash (or some other shell) + top (or ps, kill, etc.)
For overcommit 'guess', we can sum resident set sizes (RSS)
because we only need enough memory to handle what the recovery
programs will typically use. On x86_64 this is about 8MB.
For overcommit 'never', we can take the max of their virtual sizes (VSZ)
and add the sum of their RSS. We use VSZ instead of RSS because mode
forces us to ensure we can fulfill all of the requested memory allocations--
even if the programs only use a fraction of what they ask for.
On x86_64 this is about 128MB.
When swap is enabled, reserves are useful even when they are as
small as 10MB, regardless of overcommit mode.
When both swap and overcommit are disabled, then the admin should
tune the reserves higher to be absolutley safe. Over 230MB each
was safest in my testing.
* What happens if user_reserve_pages is set to 0?
Note, this only affects overcomitt 'never' mode.
Then a user will be able to allocate all available memory minus
admin_reserve_kbytes.
However, they will easily see a message such as:
"bash: fork: Cannot allocate memory"
And they won't be able to recover/kill their application.
The admin should be able to recover the system if
admin_reserve_kbytes is set appropriately.
* What's the difference between overcommit 'guess' and 'never'?
"Guess" allows an allocation if there are enough free + reclaimable
pages. It has a hardcoded 3% of free pages reserved for root.
"Never" allows an allocation if there is enough swap + a configurable
percentage (default is 50) of physical RAM. It has a hardcoded 3% of
free pages reserved for root, like "Guess" mode. It also has a
hardcoded 3% of the current process size reserved for additional
applications.
* Why is overcommit 'guess' not suitable even when an app eventually
writes to every page? It takes free pages, file pages, available
swap pages, reclaimable slab pages into consideration. In other words,
these are all pages available, then why isn't overcommit suitable?
Because it only looks at the present state of the system. It
does not take into account the memory that other applications have
malloced, but haven't initialized yet. It overcommits the system.
Test Summary
There was little change in behavior in the default overcommit 'guess'
mode with swap enabled before and after the patch. This was expected.
Systems run most predictably (i.e. no oom kills) in overcommit 'never'
mode with swap enabled. This also allowed the most memory to be allocated
to a user application.
Overcommit 'guess' mode without swap is a bad idea. It is easy to
crash the system. None of the other tested combinations crashed.
This matches my experience on the Roadrunner supercomputer.
Without the tunable user reserve, a system in overcommit 'never' mode
and without swap does not allow the admin to recover, although the
admin can.
With the new tunable reserves, a system in overcommit 'never' mode
and without swap can be configured to:
1. maximize user-allocatable memory, running close to the edge of
recoverability
2. maximize recoverability, sacrificing allocatable memory to
ensure that a user cannot take down a system
Test Description
Fedora 18 VM - 4 x86_64 cores, 5725MB RAM, 4GB Swap
System is booted into multiuser console mode, with unnecessary services
turned off. Caches were dropped before each test.
Hogs are user memtester processes that attempt to allocate all free memory
as reported by /proc/meminfo
In overcommit 'never' mode, memory_ratio=100
Test Results
3.9.0-rc1-mm1
Overcommit | Swap | Hogs | MB Got/Wanted | OOMs | User Recovery | Admin Recovery
---------- ---- ---- ------------- ---- ------------- --------------
guess yes 1 5432/5432 no yes yes
guess yes 4 5444/5444 1 yes yes
guess no 1 5302/5449 no yes yes
guess no 4 - crash no no
never yes 1 5460/5460 1 yes yes
never yes 4 5460/5460 1 yes yes
never no 1 5218/5432 no no yes
never no 4 5203/5448 no no yes
3.9.0-rc1-mm1-tunablereserves
User and Admin Recovery show their respective reserves, if applicable.
Overcommit | Swap | Hogs | MB Got/Wanted | OOMs | User Recovery | Admin Recovery
---------- ---- ---- ------------- ---- ------------- --------------
guess yes 1 5419/5419 no - yes 8MB yes
guess yes 4 5436/5436 1 - yes 8MB yes
guess no 1 5440/5440 * - yes 8MB yes
guess no 4 - crash - no 8MB no
* process would successfully mlock, then the oom killer would pick it
never yes 1 5446/5446 no 10MB yes 20MB yes
never yes 4 5456/5456 no 10MB yes 20MB yes
never no 1 5387/5429 no 128MB no 8MB barely
never no 1 5323/5428 no 226MB barely 8MB barely
never no 1 5323/5428 no 226MB barely 8MB barely
never no 1 5359/5448 no 10MB no 10MB barely
never no 1 5323/5428 no 0MB no 10MB barely
never no 1 5332/5428 no 0MB no 50MB yes
never no 1 5293/5429 no 0MB no 90MB yes
never no 1 5001/5427 no 230MB yes 338MB yes
never no 4* 4998/5424 no 230MB yes 338MB yes
* more memtesters were launched, able to allocate approximately another 100MB
Future Work
- Test larger memory systems.
- Test an embedded image.
- Test other architectures.
- Time malloc microbenchmarks.
- Would it be useful to be able to set overcommit policy for
each memory cgroup?
- Some lines are slightly above 80 chars.
Perhaps define a macro to convert between pages and kb?
Other places in the kernel do this.
[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: make init_user_reserve() static]
Signed-off-by: Andrew Shewmaker <agshew@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-04-30 06:08:10 +08:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Initialise sysctl_user_reserve_kbytes.
|
|
|
|
|
*
|
|
|
|
|
* This is intended to prevent a user from starting a single memory hogging
|
|
|
|
|
* process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
|
|
|
|
|
* mode.
|
|
|
|
|
*
|
|
|
|
|
* The default value is min(3% of free memory, 128MB)
|
|
|
|
|
* 128MB is enough to recover with sshd/login, bash, and top/kill.
|
|
|
|
|
*/
|
2013-04-30 06:08:12 +08:00
|
|
|
static int init_user_reserve(void)
|
mm: limit growth of 3% hardcoded other user reserve
Add user_reserve_kbytes knob.
Limit the growth of the memory reserved for other user processes to
min(3% current process size, user_reserve_pages). Only about 8MB is
necessary to enable recovery in the default mode, and only a few hundred
MB are required even when overcommit is disabled.
user_reserve_pages defaults to min(3% free pages, 128MB)
I arrived at 128MB by taking the max VSZ of sshd, login, bash, and top ...
then adding the RSS of each.
This only affects OVERCOMMIT_NEVER mode.
Background
1. user reserve
__vm_enough_memory reserves a hardcoded 3% of the current process size for
other applications when overcommit is disabled. This was done so that a
user could recover if they launched a memory hogging process. Without the
reserve, a user would easily run into a message such as:
bash: fork: Cannot allocate memory
2. admin reserve
Additionally, a hardcoded 3% of free memory is reserved for root in both
overcommit 'guess' and 'never' modes. This was intended to prevent a
scenario where root-cant-log-in and perform recovery operations.
Note that this reserve shrinks, and doesn't guarantee a useful reserve.
Motivation
The two hardcoded memory reserves should be updated to account for current
memory sizes.
Also, the admin reserve would be more useful if it didn't shrink too much.
When the current code was originally written, 1GB was considered
"enterprise". Now the 3% reserve can grow to multiple GB on large memory
systems, and it only needs to be a few hundred MB at most to enable a user
or admin to recover a system with an unwanted memory hogging process.
I've found that reducing these reserves is especially beneficial for a
specific type of application load:
* single application system
* one or few processes (e.g. one per core)
* allocating all available memory
* not initializing every page immediately
* long running
I've run scientific clusters with this sort of load. A long running job
sometimes failed many hours (weeks of CPU time) into a calculation. They
weren't initializing all of their memory immediately, and they weren't
using calloc, so I put systems into overcommit 'never' mode. These
clusters run diskless and have no swap.
However, with the current reserves, a user wishing to allocate as much
memory as possible to one process may be prevented from using, for
example, almost 2GB out of 32GB.
The effect is less, but still significant when a user starts a job with
one process per core. I have repeatedly seen a set of processes
requesting the same amount of memory fail because one of them could not
allocate the amount of memory a user would expect to be able to allocate.
For example, Message Passing Interfce (MPI) processes, one per core. And
it is similar for other parallel programming frameworks.
Changing this reserve code will make the overcommit never mode more useful
by allowing applications to allocate nearly all of the available memory.
Also, the new admin_reserve_kbytes will be safer than the current behavior
since the hardcoded 3% of available memory reserve can shrink to something
useless in the case where applications have grabbed all available memory.
Risks
* "bash: fork: Cannot allocate memory"
The downside of the first patch-- which creates a tunable user reserve
that is only used in overcommit 'never' mode--is that an admin can set
it so low that a user may not be able to kill their process, even if
they already have a shell prompt.
Of course, a user can get in the same predicament with the current 3%
reserve--they just have to launch processes until 3% becomes negligible.
* root-cant-log-in problem
The second patch, adding the tunable rootuser_reserve_pages, allows
the admin to shoot themselves in the foot by setting it too small. They
can easily get the system into a state where root-can't-log-in.
However, the new admin_reserve_kbytes will be safer than the current
behavior since the hardcoded 3% of available memory reserve can shrink
to something useless in the case where applications have grabbed all
available memory.
Alternatives
* Memory cgroups provide a more flexible way to limit application memory.
Not everyone wants to set up cgroups or deal with their overhead.
* We could create a fourth overcommit mode which provides smaller reserves.
The size of useful reserves may be drastically different depending
on the whether the system is embedded or enterprise.
* Force users to initialize all of their memory or use calloc.
Some users don't want/expect the system to overcommit when they malloc.
Overcommit 'never' mode is for this scenario, and it should work well.
The new user and admin reserve tunables are simple to use, with low
overhead compared to cgroups. The patches preserve current behavior where
3% of memory is less than 128MB, except that the admin reserve doesn't
shrink to an unusable size under pressure. The code allows admins to tune
for embedded and enterprise usage.
FAQ
* How is the root-cant-login problem addressed?
What happens if admin_reserve_pages is set to 0?
Root is free to shoot themselves in the foot by setting
admin_reserve_kbytes too low.
On x86_64, the minimum useful reserve is:
8MB for overcommit 'guess'
128MB for overcommit 'never'
admin_reserve_pages defaults to min(3% free memory, 8MB)
So, anyone switching to 'never' mode needs to adjust
admin_reserve_pages.
* How do you calculate a minimum useful reserve?
A user or the admin needs enough memory to login and perform
recovery operations, which includes, at a minimum:
sshd or login + bash (or some other shell) + top (or ps, kill, etc.)
For overcommit 'guess', we can sum resident set sizes (RSS)
because we only need enough memory to handle what the recovery
programs will typically use. On x86_64 this is about 8MB.
For overcommit 'never', we can take the max of their virtual sizes (VSZ)
and add the sum of their RSS. We use VSZ instead of RSS because mode
forces us to ensure we can fulfill all of the requested memory allocations--
even if the programs only use a fraction of what they ask for.
On x86_64 this is about 128MB.
When swap is enabled, reserves are useful even when they are as
small as 10MB, regardless of overcommit mode.
When both swap and overcommit are disabled, then the admin should
tune the reserves higher to be absolutley safe. Over 230MB each
was safest in my testing.
* What happens if user_reserve_pages is set to 0?
Note, this only affects overcomitt 'never' mode.
Then a user will be able to allocate all available memory minus
admin_reserve_kbytes.
However, they will easily see a message such as:
"bash: fork: Cannot allocate memory"
And they won't be able to recover/kill their application.
The admin should be able to recover the system if
admin_reserve_kbytes is set appropriately.
* What's the difference between overcommit 'guess' and 'never'?
"Guess" allows an allocation if there are enough free + reclaimable
pages. It has a hardcoded 3% of free pages reserved for root.
"Never" allows an allocation if there is enough swap + a configurable
percentage (default is 50) of physical RAM. It has a hardcoded 3% of
free pages reserved for root, like "Guess" mode. It also has a
hardcoded 3% of the current process size reserved for additional
applications.
* Why is overcommit 'guess' not suitable even when an app eventually
writes to every page? It takes free pages, file pages, available
swap pages, reclaimable slab pages into consideration. In other words,
these are all pages available, then why isn't overcommit suitable?
Because it only looks at the present state of the system. It
does not take into account the memory that other applications have
malloced, but haven't initialized yet. It overcommits the system.
Test Summary
There was little change in behavior in the default overcommit 'guess'
mode with swap enabled before and after the patch. This was expected.
Systems run most predictably (i.e. no oom kills) in overcommit 'never'
mode with swap enabled. This also allowed the most memory to be allocated
to a user application.
Overcommit 'guess' mode without swap is a bad idea. It is easy to
crash the system. None of the other tested combinations crashed.
This matches my experience on the Roadrunner supercomputer.
Without the tunable user reserve, a system in overcommit 'never' mode
and without swap does not allow the admin to recover, although the
admin can.
With the new tunable reserves, a system in overcommit 'never' mode
and without swap can be configured to:
1. maximize user-allocatable memory, running close to the edge of
recoverability
2. maximize recoverability, sacrificing allocatable memory to
ensure that a user cannot take down a system
Test Description
Fedora 18 VM - 4 x86_64 cores, 5725MB RAM, 4GB Swap
System is booted into multiuser console mode, with unnecessary services
turned off. Caches were dropped before each test.
Hogs are user memtester processes that attempt to allocate all free memory
as reported by /proc/meminfo
In overcommit 'never' mode, memory_ratio=100
Test Results
3.9.0-rc1-mm1
Overcommit | Swap | Hogs | MB Got/Wanted | OOMs | User Recovery | Admin Recovery
---------- ---- ---- ------------- ---- ------------- --------------
guess yes 1 5432/5432 no yes yes
guess yes 4 5444/5444 1 yes yes
guess no 1 5302/5449 no yes yes
guess no 4 - crash no no
never yes 1 5460/5460 1 yes yes
never yes 4 5460/5460 1 yes yes
never no 1 5218/5432 no no yes
never no 4 5203/5448 no no yes
3.9.0-rc1-mm1-tunablereserves
User and Admin Recovery show their respective reserves, if applicable.
Overcommit | Swap | Hogs | MB Got/Wanted | OOMs | User Recovery | Admin Recovery
---------- ---- ---- ------------- ---- ------------- --------------
guess yes 1 5419/5419 no - yes 8MB yes
guess yes 4 5436/5436 1 - yes 8MB yes
guess no 1 5440/5440 * - yes 8MB yes
guess no 4 - crash - no 8MB no
* process would successfully mlock, then the oom killer would pick it
never yes 1 5446/5446 no 10MB yes 20MB yes
never yes 4 5456/5456 no 10MB yes 20MB yes
never no 1 5387/5429 no 128MB no 8MB barely
never no 1 5323/5428 no 226MB barely 8MB barely
never no 1 5323/5428 no 226MB barely 8MB barely
never no 1 5359/5448 no 10MB no 10MB barely
never no 1 5323/5428 no 0MB no 10MB barely
never no 1 5332/5428 no 0MB no 50MB yes
never no 1 5293/5429 no 0MB no 90MB yes
never no 1 5001/5427 no 230MB yes 338MB yes
never no 4* 4998/5424 no 230MB yes 338MB yes
* more memtesters were launched, able to allocate approximately another 100MB
Future Work
- Test larger memory systems.
- Test an embedded image.
- Test other architectures.
- Time malloc microbenchmarks.
- Would it be useful to be able to set overcommit policy for
each memory cgroup?
- Some lines are slightly above 80 chars.
Perhaps define a macro to convert between pages and kb?
Other places in the kernel do this.
[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: make init_user_reserve() static]
Signed-off-by: Andrew Shewmaker <agshew@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-04-30 06:08:10 +08:00
|
|
|
{
|
|
|
|
|
unsigned long free_kbytes;
|
|
|
|
|
|
2017-09-07 07:23:36 +08:00
|
|
|
free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
|
mm: limit growth of 3% hardcoded other user reserve
Add user_reserve_kbytes knob.
Limit the growth of the memory reserved for other user processes to
min(3% current process size, user_reserve_pages). Only about 8MB is
necessary to enable recovery in the default mode, and only a few hundred
MB are required even when overcommit is disabled.
user_reserve_pages defaults to min(3% free pages, 128MB)
I arrived at 128MB by taking the max VSZ of sshd, login, bash, and top ...
then adding the RSS of each.
This only affects OVERCOMMIT_NEVER mode.
Background
1. user reserve
__vm_enough_memory reserves a hardcoded 3% of the current process size for
other applications when overcommit is disabled. This was done so that a
user could recover if they launched a memory hogging process. Without the
reserve, a user would easily run into a message such as:
bash: fork: Cannot allocate memory
2. admin reserve
Additionally, a hardcoded 3% of free memory is reserved for root in both
overcommit 'guess' and 'never' modes. This was intended to prevent a
scenario where root-cant-log-in and perform recovery operations.
Note that this reserve shrinks, and doesn't guarantee a useful reserve.
Motivation
The two hardcoded memory reserves should be updated to account for current
memory sizes.
Also, the admin reserve would be more useful if it didn't shrink too much.
When the current code was originally written, 1GB was considered
"enterprise". Now the 3% reserve can grow to multiple GB on large memory
systems, and it only needs to be a few hundred MB at most to enable a user
or admin to recover a system with an unwanted memory hogging process.
I've found that reducing these reserves is especially beneficial for a
specific type of application load:
* single application system
* one or few processes (e.g. one per core)
* allocating all available memory
* not initializing every page immediately
* long running
I've run scientific clusters with this sort of load. A long running job
sometimes failed many hours (weeks of CPU time) into a calculation. They
weren't initializing all of their memory immediately, and they weren't
using calloc, so I put systems into overcommit 'never' mode. These
clusters run diskless and have no swap.
However, with the current reserves, a user wishing to allocate as much
memory as possible to one process may be prevented from using, for
example, almost 2GB out of 32GB.
The effect is less, but still significant when a user starts a job with
one process per core. I have repeatedly seen a set of processes
requesting the same amount of memory fail because one of them could not
allocate the amount of memory a user would expect to be able to allocate.
For example, Message Passing Interfce (MPI) processes, one per core. And
it is similar for other parallel programming frameworks.
Changing this reserve code will make the overcommit never mode more useful
by allowing applications to allocate nearly all of the available memory.
Also, the new admin_reserve_kbytes will be safer than the current behavior
since the hardcoded 3% of available memory reserve can shrink to something
useless in the case where applications have grabbed all available memory.
Risks
* "bash: fork: Cannot allocate memory"
The downside of the first patch-- which creates a tunable user reserve
that is only used in overcommit 'never' mode--is that an admin can set
it so low that a user may not be able to kill their process, even if
they already have a shell prompt.
Of course, a user can get in the same predicament with the current 3%
reserve--they just have to launch processes until 3% becomes negligible.
* root-cant-log-in problem
The second patch, adding the tunable rootuser_reserve_pages, allows
the admin to shoot themselves in the foot by setting it too small. They
can easily get the system into a state where root-can't-log-in.
However, the new admin_reserve_kbytes will be safer than the current
behavior since the hardcoded 3% of available memory reserve can shrink
to something useless in the case where applications have grabbed all
available memory.
Alternatives
* Memory cgroups provide a more flexible way to limit application memory.
Not everyone wants to set up cgroups or deal with their overhead.
* We could create a fourth overcommit mode which provides smaller reserves.
The size of useful reserves may be drastically different depending
on the whether the system is embedded or enterprise.
* Force users to initialize all of their memory or use calloc.
Some users don't want/expect the system to overcommit when they malloc.
Overcommit 'never' mode is for this scenario, and it should work well.
The new user and admin reserve tunables are simple to use, with low
overhead compared to cgroups. The patches preserve current behavior where
3% of memory is less than 128MB, except that the admin reserve doesn't
shrink to an unusable size under pressure. The code allows admins to tune
for embedded and enterprise usage.
FAQ
* How is the root-cant-login problem addressed?
What happens if admin_reserve_pages is set to 0?
Root is free to shoot themselves in the foot by setting
admin_reserve_kbytes too low.
On x86_64, the minimum useful reserve is:
8MB for overcommit 'guess'
128MB for overcommit 'never'
admin_reserve_pages defaults to min(3% free memory, 8MB)
So, anyone switching to 'never' mode needs to adjust
admin_reserve_pages.
* How do you calculate a minimum useful reserve?
A user or the admin needs enough memory to login and perform
recovery operations, which includes, at a minimum:
sshd or login + bash (or some other shell) + top (or ps, kill, etc.)
For overcommit 'guess', we can sum resident set sizes (RSS)
because we only need enough memory to handle what the recovery
programs will typically use. On x86_64 this is about 8MB.
For overcommit 'never', we can take the max of their virtual sizes (VSZ)
and add the sum of their RSS. We use VSZ instead of RSS because mode
forces us to ensure we can fulfill all of the requested memory allocations--
even if the programs only use a fraction of what they ask for.
On x86_64 this is about 128MB.
When swap is enabled, reserves are useful even when they are as
small as 10MB, regardless of overcommit mode.
When both swap and overcommit are disabled, then the admin should
tune the reserves higher to be absolutley safe. Over 230MB each
was safest in my testing.
* What happens if user_reserve_pages is set to 0?
Note, this only affects overcomitt 'never' mode.
Then a user will be able to allocate all available memory minus
admin_reserve_kbytes.
However, they will easily see a message such as:
"bash: fork: Cannot allocate memory"
And they won't be able to recover/kill their application.
The admin should be able to recover the system if
admin_reserve_kbytes is set appropriately.
* What's the difference between overcommit 'guess' and 'never'?
"Guess" allows an allocation if there are enough free + reclaimable
pages. It has a hardcoded 3% of free pages reserved for root.
"Never" allows an allocation if there is enough swap + a configurable
percentage (default is 50) of physical RAM. It has a hardcoded 3% of
free pages reserved for root, like "Guess" mode. It also has a
hardcoded 3% of the current process size reserved for additional
applications.
* Why is overcommit 'guess' not suitable even when an app eventually
writes to every page? It takes free pages, file pages, available
swap pages, reclaimable slab pages into consideration. In other words,
these are all pages available, then why isn't overcommit suitable?
Because it only looks at the present state of the system. It
does not take into account the memory that other applications have
malloced, but haven't initialized yet. It overcommits the system.
Test Summary
There was little change in behavior in the default overcommit 'guess'
mode with swap enabled before and after the patch. This was expected.
Systems run most predictably (i.e. no oom kills) in overcommit 'never'
mode with swap enabled. This also allowed the most memory to be allocated
to a user application.
Overcommit 'guess' mode without swap is a bad idea. It is easy to
crash the system. None of the other tested combinations crashed.
This matches my experience on the Roadrunner supercomputer.
Without the tunable user reserve, a system in overcommit 'never' mode
and without swap does not allow the admin to recover, although the
admin can.
With the new tunable reserves, a system in overcommit 'never' mode
and without swap can be configured to:
1. maximize user-allocatable memory, running close to the edge of
recoverability
2. maximize recoverability, sacrificing allocatable memory to
ensure that a user cannot take down a system
Test Description
Fedora 18 VM - 4 x86_64 cores, 5725MB RAM, 4GB Swap
System is booted into multiuser console mode, with unnecessary services
turned off. Caches were dropped before each test.
Hogs are user memtester processes that attempt to allocate all free memory
as reported by /proc/meminfo
In overcommit 'never' mode, memory_ratio=100
Test Results
3.9.0-rc1-mm1
Overcommit | Swap | Hogs | MB Got/Wanted | OOMs | User Recovery | Admin Recovery
---------- ---- ---- ------------- ---- ------------- --------------
guess yes 1 5432/5432 no yes yes
guess yes 4 5444/5444 1 yes yes
guess no 1 5302/5449 no yes yes
guess no 4 - crash no no
never yes 1 5460/5460 1 yes yes
never yes 4 5460/5460 1 yes yes
never no 1 5218/5432 no no yes
never no 4 5203/5448 no no yes
3.9.0-rc1-mm1-tunablereserves
User and Admin Recovery show their respective reserves, if applicable.
Overcommit | Swap | Hogs | MB Got/Wanted | OOMs | User Recovery | Admin Recovery
---------- ---- ---- ------------- ---- ------------- --------------
guess yes 1 5419/5419 no - yes 8MB yes
guess yes 4 5436/5436 1 - yes 8MB yes
guess no 1 5440/5440 * - yes 8MB yes
guess no 4 - crash - no 8MB no
* process would successfully mlock, then the oom killer would pick it
never yes 1 5446/5446 no 10MB yes 20MB yes
never yes 4 5456/5456 no 10MB yes 20MB yes
never no 1 5387/5429 no 128MB no 8MB barely
never no 1 5323/5428 no 226MB barely 8MB barely
never no 1 5323/5428 no 226MB barely 8MB barely
never no 1 5359/5448 no 10MB no 10MB barely
never no 1 5323/5428 no 0MB no 10MB barely
never no 1 5332/5428 no 0MB no 50MB yes
never no 1 5293/5429 no 0MB no 90MB yes
never no 1 5001/5427 no 230MB yes 338MB yes
never no 4* 4998/5424 no 230MB yes 338MB yes
* more memtesters were launched, able to allocate approximately another 100MB
Future Work
- Test larger memory systems.
- Test an embedded image.
- Test other architectures.
- Time malloc microbenchmarks.
- Would it be useful to be able to set overcommit policy for
each memory cgroup?
- Some lines are slightly above 80 chars.
Perhaps define a macro to convert between pages and kb?
Other places in the kernel do this.
[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: make init_user_reserve() static]
Signed-off-by: Andrew Shewmaker <agshew@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-04-30 06:08:10 +08:00
|
|
|
|
|
|
|
|
sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
2014-01-24 07:53:30 +08:00
|
|
|
subsys_initcall(init_user_reserve);
|
2013-04-30 06:08:11 +08:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Initialise sysctl_admin_reserve_kbytes.
|
|
|
|
|
*
|
|
|
|
|
* The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
|
|
|
|
|
* to log in and kill a memory hogging process.
|
|
|
|
|
*
|
|
|
|
|
* Systems with more than 256MB will reserve 8MB, enough to recover
|
|
|
|
|
* with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
|
|
|
|
|
* only reserve 3% of free pages by default.
|
|
|
|
|
*/
|
2013-04-30 06:08:12 +08:00
|
|
|
static int init_admin_reserve(void)
|
2013-04-30 06:08:11 +08:00
|
|
|
{
|
|
|
|
|
unsigned long free_kbytes;
|
|
|
|
|
|
2017-09-07 07:23:36 +08:00
|
|
|
free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
|
2013-04-30 06:08:11 +08:00
|
|
|
|
|
|
|
|
sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
2014-01-24 07:53:30 +08:00
|
|
|
subsys_initcall(init_admin_reserve);
|
2013-04-30 06:08:12 +08:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Reinititalise user and admin reserves if memory is added or removed.
|
|
|
|
|
*
|
|
|
|
|
* The default user reserve max is 128MB, and the default max for the
|
|
|
|
|
* admin reserve is 8MB. These are usually, but not always, enough to
|
|
|
|
|
* enable recovery from a memory hogging process using login/sshd, a shell,
|
|
|
|
|
* and tools like top. It may make sense to increase or even disable the
|
|
|
|
|
* reserve depending on the existence of swap or variations in the recovery
|
|
|
|
|
* tools. So, the admin may have changed them.
|
|
|
|
|
*
|
|
|
|
|
* If memory is added and the reserves have been eliminated or increased above
|
|
|
|
|
* the default max, then we'll trust the admin.
|
|
|
|
|
*
|
|
|
|
|
* If memory is removed and there isn't enough free memory, then we
|
|
|
|
|
* need to reset the reserves.
|
|
|
|
|
*
|
|
|
|
|
* Otherwise keep the reserve set by the admin.
|
|
|
|
|
*/
|
|
|
|
|
static int reserve_mem_notifier(struct notifier_block *nb,
|
|
|
|
|
unsigned long action, void *data)
|
|
|
|
|
{
|
|
|
|
|
unsigned long tmp, free_kbytes;
|
|
|
|
|
|
|
|
|
|
switch (action) {
|
|
|
|
|
case MEM_ONLINE:
|
|
|
|
|
/* Default max is 128MB. Leave alone if modified by operator. */
|
|
|
|
|
tmp = sysctl_user_reserve_kbytes;
|
|
|
|
|
if (0 < tmp && tmp < (1UL << 17))
|
|
|
|
|
init_user_reserve();
|
|
|
|
|
|
|
|
|
|
/* Default max is 8MB. Leave alone if modified by operator. */
|
|
|
|
|
tmp = sysctl_admin_reserve_kbytes;
|
|
|
|
|
if (0 < tmp && tmp < (1UL << 13))
|
|
|
|
|
init_admin_reserve();
|
|
|
|
|
|
|
|
|
|
break;
|
|
|
|
|
case MEM_OFFLINE:
|
2017-09-07 07:23:36 +08:00
|
|
|
free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
|
2013-04-30 06:08:12 +08:00
|
|
|
|
|
|
|
|
if (sysctl_user_reserve_kbytes > free_kbytes) {
|
|
|
|
|
init_user_reserve();
|
|
|
|
|
pr_info("vm.user_reserve_kbytes reset to %lu\n",
|
|
|
|
|
sysctl_user_reserve_kbytes);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (sysctl_admin_reserve_kbytes > free_kbytes) {
|
|
|
|
|
init_admin_reserve();
|
|
|
|
|
pr_info("vm.admin_reserve_kbytes reset to %lu\n",
|
|
|
|
|
sysctl_admin_reserve_kbytes);
|
|
|
|
|
}
|
|
|
|
|
break;
|
|
|
|
|
default:
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
return NOTIFY_OK;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static struct notifier_block reserve_mem_nb = {
|
|
|
|
|
.notifier_call = reserve_mem_notifier,
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
static int __meminit init_reserve_notifier(void)
|
|
|
|
|
{
|
|
|
|
|
if (register_hotmemory_notifier(&reserve_mem_nb))
|
2014-06-07 05:38:30 +08:00
|
|
|
pr_err("Failed registering memory add/remove notifier for admin reserve\n");
|
2013-04-30 06:08:12 +08:00
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
2014-01-24 07:53:30 +08:00
|
|
|
subsys_initcall(init_reserve_notifier);
|