As a preparational patch to bump up page allocator pass-through threshold,
introduce two new constants SLUB_MAX_SIZE and SLUB_PAGE_SHIFT and convert
mm/slub.c to use them.
Reported-by: "Zhang, Yanmin" <yanmin_zhang@linux.intel.com>
Tested-by: "Zhang, Yanmin" <yanmin_zhang@linux.intel.com>
Signed-off-by: Christoph Lameter <cl@linux-foundation.org>
Signed-off-by: Pekka Enberg <penberg@cs.helsinki.fi>
Increase the maximum object size in SLUB so that 8k objects are not
passed through to the page allocator anymore. The network stack uses 8k
objects for performance critical operations.
The patch is motivated by a SLAB vs. SLUB regression in the netperf
benchmark. The problem is that the kfree(skb->head) call in
skb_release_data() that is subject to page allocator pass-through as the
size passed to __alloc_skb() is larger than 4 KB in this test.
As explained by Yanmin Zhang:
I use 2.6.29-rc2 kernel to run netperf UDP-U-4k CPU_NUM client/server
pair loopback testing on x86-64 machines. Comparing with SLUB, SLAB's
result is about 2.3 times of SLUB's. After applying the reverting patch,
the result difference between SLUB and SLAB becomes 1% which we might
consider as fluctuation.
[ penberg@cs.helsinki.fi: fix oops in kmalloc() ]
Reported-by: "Zhang, Yanmin" <yanmin_zhang@linux.intel.com>
Tested-by: "Zhang, Yanmin" <yanmin_zhang@linux.intel.com>
Signed-off-by: Christoph Lameter <cl@linux-foundation.org>
Signed-off-by: Pekka Enberg <penberg@cs.helsinki.fi>
As a preparational patch to bump up page allocator pass-through threshold,
introduce two new constants SLUB_MAX_SIZE and SLUB_PAGE_SHIFT and convert
mm/slub.c to use them.
Reported-by: "Zhang, Yanmin" <yanmin_zhang@linux.intel.com>
Tested-by: "Zhang, Yanmin" <yanmin_zhang@linux.intel.com>
Signed-off-by: Christoph Lameter <cl@linux-foundation.org>
Signed-off-by: Pekka Enberg <penberg@cs.helsinki.fi>
Impact: new tracer plugin
This patch adapts kmemtrace raw events tracing to the unified tracing API.
To enable and use this tracer, just do the following:
echo kmemtrace > /debugfs/tracing/current_tracer
cat /debugfs/tracing/trace
You will have the following output:
# tracer: kmemtrace
#
#
# ALLOC TYPE REQ GIVEN FLAGS POINTER NODE CALLER
# FREE | | | | | | | |
# |
type_id 1 call_site 18446744071565527833 ptr 18446612134395152256
type_id 0 call_site 18446744071565585597 ptr 18446612134405955584 bytes_req 4096 bytes_alloc 4096 gfp_flags 208 node -1
type_id 1 call_site 18446744071565585534 ptr 18446612134405955584
type_id 0 call_site 18446744071565585597 ptr 18446612134405955584 bytes_req 4096 bytes_alloc 4096 gfp_flags 208 node -1
type_id 0 call_site 18446744071565636711 ptr 18446612134345164672 bytes_req 240 bytes_alloc 240 gfp_flags 208 node -1
type_id 1 call_site 18446744071565585534 ptr 18446612134405955584
type_id 0 call_site 18446744071565585597 ptr 18446612134405955584 bytes_req 4096 bytes_alloc 4096 gfp_flags 208 node -1
type_id 0 call_site 18446744071565636711 ptr 18446612134345164912 bytes_req 240 bytes_alloc 240 gfp_flags 208 node -1
type_id 1 call_site 18446744071565585534 ptr 18446612134405955584
type_id 0 call_site 18446744071565585597 ptr 18446612134405955584 bytes_req 4096 bytes_alloc 4096 gfp_flags 208 node -1
type_id 0 call_site 18446744071565636711 ptr 18446612134345165152 bytes_req 240 bytes_alloc 240 gfp_flags 208 node -1
type_id 0 call_site 18446744071566144042 ptr 18446612134346191680 bytes_req 1304 bytes_alloc 1312 gfp_flags 208 node -1
type_id 1 call_site 18446744071565585534 ptr 18446612134405955584
type_id 0 call_site 18446744071565585597 ptr 18446612134405955584 bytes_req 4096 bytes_alloc 4096 gfp_flags 208 node -1
type_id 1 call_site 18446744071565585534 ptr 18446612134405955584
That was to stay backward compatible with the format output produced in
inux/tracepoint.h.
This is the default ouput, but note that I tried something else.
If you change an option:
echo kmem_minimalistic > /debugfs/trace_options
and then cat /debugfs/trace, you will have the following output:
# tracer: kmemtrace
#
#
# ALLOC TYPE REQ GIVEN FLAGS POINTER NODE CALLER
# FREE | | | | | | | |
# |
- C 0xffff88007c088780 file_free_rcu
+ K 4096 4096 000000d0 0xffff88007cad6000 -1 getname
- C 0xffff88007cad6000 putname
+ K 4096 4096 000000d0 0xffff88007cad6000 -1 getname
+ K 240 240 000000d0 0xffff8800790dc780 -1 d_alloc
- C 0xffff88007cad6000 putname
+ K 4096 4096 000000d0 0xffff88007cad6000 -1 getname
+ K 240 240 000000d0 0xffff8800790dc870 -1 d_alloc
- C 0xffff88007cad6000 putname
+ K 4096 4096 000000d0 0xffff88007cad6000 -1 getname
+ K 240 240 000000d0 0xffff8800790dc960 -1 d_alloc
+ K 1304 1312 000000d0 0xffff8800791d7340 -1 reiserfs_alloc_inode
- C 0xffff88007cad6000 putname
+ K 4096 4096 000000d0 0xffff88007cad6000 -1 getname
- C 0xffff88007cad6000 putname
+ K 992 1000 000000d0 0xffff880079045b58 -1 alloc_inode
+ K 768 1024 000080d0 0xffff88007c096400 -1 alloc_pipe_info
+ K 240 240 000000d0 0xffff8800790dca50 -1 d_alloc
+ K 272 320 000080d0 0xffff88007c088780 -1 get_empty_filp
+ K 272 320 000080d0 0xffff88007c088000 -1 get_empty_filp
Yeah I shall confess kmem_minimalistic should be: kmem_alternative.
Whatever, I find it more readable but this a personal opinion of course.
We can drop it if you want.
On the ALLOC/FREE column, + means an allocation and - a free.
On the type column, you have K = kmalloc, C = cache, P = page
I would like the flags to be GFP_* strings but that would not be easy to not
break the column with strings....
About the node...it seems to always be -1. I don't know why but that shouldn't
be difficult to find.
I moved linux/tracepoint.h to trace/tracepoint.h as well. I think that would
be more easy to find the tracer headers if they are all in their common
directory.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This adds hooks for the SLUB allocator, to allow tracing with kmemtrace.
Signed-off-by: Eduard - Gabriel Munteanu <eduard.munteanu@linux360.ro>
Signed-off-by: Pekka Enberg <penberg@cs.helsinki.fi>
This patch changes the static MIN_PARTIAL to a dynamic per-cache ->min_partial
value that is calculated from object size. The bigger the object size, the more
pages we keep on the partial list.
I tested SLAB, SLUB, and SLUB with this patch on Jens Axboe's 'netio' example
script of the fio benchmarking tool. The script stresses the networking
subsystem which should also give a fairly good beating of kmalloc() et al.
To run the test yourself, first clone the fio repository:
git clone git://git.kernel.dk/fio.git
and then run the following command n times on your machine:
time ./fio examples/netio
The results on my 2-way 64-bit x86 machine are as follows:
[ the minimum, maximum, and average are captured from 50 individual runs ]
real time (seconds)
min max avg sd
SLAB 22.76 23.38 22.98 0.17
SLUB 22.80 25.78 23.46 0.72
SLUB (dynamic) 22.74 23.54 23.00 0.20
sys time (seconds)
min max avg sd
SLAB 6.90 8.28 7.70 0.28
SLUB 7.42 16.95 8.89 2.28
SLUB (dynamic) 7.17 8.64 7.73 0.29
user time (seconds)
min max avg sd
SLAB 36.89 38.11 37.50 0.29
SLUB 30.85 37.99 37.06 1.67
SLUB (dynamic) 36.75 38.07 37.59 0.32
As you can see from the above numbers, this patch brings SLUB to the same level
as SLAB for this particular workload fixing a ~2% regression. I'd expect this
change to help similar workloads that allocate a lot of objects that are close
to the size of a page.
Cc: Matthew Wilcox <matthew@wil.cx>
Cc: Andrew Morton <akpm@linux-foundation.org>
Acked-by: Christoph Lameter <cl@linux-foundation.org>
Signed-off-by: Pekka Enberg <penberg@cs.helsinki.fi>
Kmem cache passed to constructor is only needed for constructors that are
themselves multiplexeres. Nobody uses this "feature", nor does anybody uses
passed kmem cache in non-trivial way, so pass only pointer to object.
Non-trivial places are:
arch/powerpc/mm/init_64.c
arch/powerpc/mm/hugetlbpage.c
This is flag day, yes.
Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Acked-by: Pekka Enberg <penberg@cs.helsinki.fi>
Acked-by: Christoph Lameter <cl@linux-foundation.org>
Cc: Jon Tollefson <kniht@linux.vnet.ibm.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Matt Mackall <mpm@selenic.com>
[akpm@linux-foundation.org: fix arch/powerpc/mm/hugetlbpage.c]
[akpm@linux-foundation.org: fix mm/slab.c]
[akpm@linux-foundation.org: fix ubifs]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Remove all clameter@sgi.com addresses from the kernel tree since they will
become invalid on June 27th. Change my maintainer email address for the
slab allocators to cl@linux-foundation.org (which will be the new email
address for the future).
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Christoph Lameter <cl@linux-foundation.org>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Matt Mackall <mpm@selenic.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The 192 byte cache is not necessary if we have a basic alignment of 128
byte. If it would be used then the 192 would be aligned to the next 128 byte
boundary which would result in another 256 byte cache. Two 256 kmalloc caches
cause sysfs to complain about a duplicate entry.
MIPS needs 128 byte aligned kmalloc caches and spits out warnings on boot without
this patch.
Signed-off-by: Christoph Lameter <cl@linux-foundation.org>
Signed-off-by: Pekka Enberg <penberg@cs.helsinki.fi>
If any higher order allocation fails then fall back the smallest order
necessary to contain at least one object. This enables fallback for all
allocations to order 0 pages. The fallback will waste more memory (objects
will not fit neatly) and the fallback slabs will be not as efficient as larger
slabs since they contain less objects.
Note that SLAB also depends on order 1 allocations for some slabs that waste
too much memory if forced into PAGE_SIZE'd page. SLUB now can now deal with
failing order 1 allocs which SLAB cannot do.
Add a new field min that will contain the objects for the smallest possible order
for a slab cache.
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Pekka Enberg <penberg@cs.helsinki.fi>
Change the statistics to consider that slabs of the same slabcache
can have different number of objects in them since they may be of
different order.
Provide a new sysfs field
total_objects
which shows the total objects that the allocated slabs of a slabcache
could hold.
Add a max field that holds the largest slab order that was ever used
for a slab cache.
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Pekka Enberg <penberg@cs.helsinki.fi>
Pack the order and the number of objects into a single word.
This saves some memory in the kmem_cache_structure and more importantly
allows us to fetch both values atomically.
Later the slab orders become runtime configurable and we need to fetch these
two items together in order to properly allocate a slab and initialize its
objects.
Fix the race by fetching the order and the number of objects in one word.
[penberg@cs.helsinki.fi: fix memset() page order in new_slab()]
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Pekka Enberg <penberg@cs.helsinki.fi>
The per node counters are used mainly for showing data through the sysfs API.
If that API is not compiled in then there is no point in keeping track of this
data. Disable counters for the number of slabs and the number of total slabs
if !SLUB_DEBUG. Incrementing the per node counters is also accessing a
potentially contended cacheline so this could actually be a performance
benefit to embedded systems.
SLABINFO support is also affected. It now must depends on SLUB_DEBUG (which
is on by default).
Patch also avoids a check for a NULL kmem_cache_node pointer in new_slab()
if the system is not compiled with NUMA support.
[penberg@cs.helsinki.fi: fix oops and move ->nr_slabs into CONFIG_SLUB_DEBUG]
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Pekka Enberg <penberg@cs.helsinki.fi>
Currently we hand off PAGE_SIZEd kmallocs to the page allocator in the
mistaken belief that the page allocator can handle these allocations
effectively. However, measurements indicate a minimum slowdown by the
factor of 8 (and that is only SMP, NUMA is much worse) vs the slub fastpath
which causes regressions in tbench.
Increase the number of kmalloc caches by one so that we again handle 4k
kmallocs directly from slub. 4k page buffering for the page allocator
will be performed by slub like done by slab.
At some point the page allocator fastpath should be fixed. A lot of the kernel
would benefit from a faster ability to allocate a single page. If that is
done then the 4k allocs may again be forwarded to the page allocator and this
patch could be reverted.
Reviewed-by: Pekka Enberg <penberg@cs.helsinki.fi>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Currently we determine the gfp flags to pass to the page allocator
each time a slab is being allocated.
Determine the bits to be set at the time the slab is created. Store
in a new allocflags field and add the flags in allocate_slab().
Acked-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Pekka Enberg <penberg@cs.helsinki.fi>
Signed-off-by: Christoph Lameter <clameter@sgi.com>
This adds a proper function for kmalloc page allocator pass-through. While it
simplifies any code that does slab tracing code a lot, I think it's a
worthwhile cleanup in itself.
Signed-off-by: Pekka Enberg <penberg@cs.helsinki.fi>
Signed-off-by: Christoph Lameter <clameter@sgi.com>
The statistics provided here allow the monitoring of allocator behavior but
at the cost of some (minimal) loss of performance. Counters are placed in
SLUB's per cpu data structure. The per cpu structure may be extended by the
statistics to grow larger than one cacheline which will increase the cache
footprint of SLUB.
There is a compile option to enable/disable the inclusion of the runtime
statistics and its off by default.
The slabinfo tool is enhanced to support these statistics via two options:
-D Switches the line of information displayed for a slab from size
mode to activity mode.
-A Sorts the slabs displayed by activity. This allows the display of
the slabs most important to the performance of a certain load.
-r Report option will report detailed statistics on
Example (tbench load):
slabinfo -AD ->Shows the most active slabs
Name Objects Alloc Free %Fast
skbuff_fclone_cache 33 111953835 111953835 99 99
:0000192 2666 5283688 5281047 99 99
:0001024 849 5247230 5246389 83 83
vm_area_struct 1349 119642 118355 91 22
:0004096 15 66753 66751 98 98
:0000064 2067 25297 23383 98 78
dentry 10259 28635 18464 91 45
:0000080 11004 18950 8089 98 98
:0000096 1703 12358 10784 99 98
:0000128 762 10582 9875 94 18
:0000512 184 9807 9647 95 81
:0002048 479 9669 9195 83 65
anon_vma 777 9461 9002 99 71
kmalloc-8 6492 9981 5624 99 97
:0000768 258 7174 6931 58 15
So the skbuff_fclone_cache is of highest importance for the tbench load.
Pretty high load on the 192 sized slab. Look for the aliases
slabinfo -a | grep 000192
:0000192 <- xfs_btree_cur filp kmalloc-192 uid_cache tw_sock_TCP
request_sock_TCPv6 tw_sock_TCPv6 skbuff_head_cache xfs_ili
Likely skbuff_head_cache.
Looking into the statistics of the skbuff_fclone_cache is possible through
slabinfo skbuff_fclone_cache ->-r option implied if cache name is mentioned
.... Usual output ...
Slab Perf Counter Alloc Free %Al %Fr
--------------------------------------------------
Fastpath 111953360 111946981 99 99
Slowpath 1044 7423 0 0
Page Alloc 272 264 0 0
Add partial 25 325 0 0
Remove partial 86 264 0 0
RemoteObj/SlabFrozen 350 4832 0 0
Total 111954404 111954404
Flushes 49 Refill 0
Deactivate Full=325(92%) Empty=0(0%) ToHead=24(6%) ToTail=1(0%)
Looks good because the fastpath is overwhelmingly taken.
skbuff_head_cache:
Slab Perf Counter Alloc Free %Al %Fr
--------------------------------------------------
Fastpath 5297262 5259882 99 99
Slowpath 4477 39586 0 0
Page Alloc 937 824 0 0
Add partial 0 2515 0 0
Remove partial 1691 824 0 0
RemoteObj/SlabFrozen 2621 9684 0 0
Total 5301739 5299468
Deactivate Full=2620(100%) Empty=0(0%) ToHead=0(0%) ToTail=0(0%)
Descriptions of the output:
Total: The total number of allocation and frees that occurred for a
slab
Fastpath: The number of allocations/frees that used the fastpath.
Slowpath: Other allocations
Page Alloc: Number of calls to the page allocator as a result of slowpath
processing
Add Partial: Number of slabs added to the partial list through free or
alloc (occurs during cpuslab flushes)
Remove Partial: Number of slabs removed from the partial list as a result of
allocations retrieving a partial slab or by a free freeing
the last object of a slab.
RemoteObj/Froz: How many times were remotely freed object encountered when a
slab was about to be deactivated. Frozen: How many times was
free able to skip list processing because the slab was in use
as the cpuslab of another processor.
Flushes: Number of times the cpuslab was flushed on request
(kmem_cache_shrink, may result from races in __slab_alloc)
Refill: Number of times we were able to refill the cpuslab from
remotely freed objects for the same slab.
Deactivate: Statistics how slabs were deactivated. Shows how they were
put onto the partial list.
In general fastpath is very good. Slowpath without partial list processing is
also desirable. Any touching of partial list uses node specific locks which
may potentially cause list lock contention.
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Add some comments explaining the fields of the kmem_cache_cpu structure.
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
The NUMA defrag works by allocating objects from partial slabs on remote
nodes. Rename it to
remote_node_defrag_ratio
to be clear about this.
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Both SLUB and SLAB really did almost exactly the same thing for
/proc/slabinfo setup, using duplicate code and per-allocator #ifdef's.
This just creates a common CONFIG_SLABINFO that is enabled by both SLUB
and SLAB, and shares all the setup code. Maybe SLOB will want this some
day too.
Reviewed-by: Pekka Enberg <penberg@cs.helsinki.fi>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Slab constructors currently have a flags parameter that is never used. And
the order of the arguments is opposite to other slab functions. The object
pointer is placed before the kmem_cache pointer.
Convert
ctor(void *object, struct kmem_cache *s, unsigned long flags)
to
ctor(struct kmem_cache *s, void *object)
throughout the kernel
[akpm@linux-foundation.org: coupla fixes]
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
We touch a cacheline in the kmem_cache structure for zeroing to get the
size. However, the hot paths in slab_alloc and slab_free do not reference
any other fields in kmem_cache, so we may have to just bring in the
cacheline for this one access.
Add a new field to kmem_cache_cpu that contains the object size. That
cacheline must already be used in the hotpaths. So we save one cacheline
on every slab_alloc if we zero.
We need to update the kmem_cache_cpu object size if an aliasing operation
changes the objsize of an non debug slab.
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The kmem_cache_cpu structures introduced are currently an array placed in the
kmem_cache struct. Meaning the kmem_cache_cpu structures are overwhelmingly
on the wrong node for systems with a higher amount of nodes. These are
performance critical structures since the per node information has
to be touched for every alloc and free in a slab.
In order to place the kmem_cache_cpu structure optimally we put an array
of pointers to kmem_cache_cpu structs in kmem_cache (similar to SLAB).
However, the kmem_cache_cpu structures can now be allocated in a more
intelligent way.
We would like to put per cpu structures for the same cpu but different
slab caches in cachelines together to save space and decrease the cache
footprint. However, the slab allocators itself control only allocations
per node. We set up a simple per cpu array for every processor with
100 per cpu structures which is usually enough to get them all set up right.
If we run out then we fall back to kmalloc_node. This also solves the
bootstrap problem since we do not have to use slab allocator functions
early in boot to get memory for the small per cpu structures.
Pro:
- NUMA aware placement improves memory performance
- All global structures in struct kmem_cache become readonly
- Dense packing of per cpu structures reduces cacheline
footprint in SMP and NUMA.
- Potential avoidance of exclusive cacheline fetches
on the free and alloc hotpath since multiple kmem_cache_cpu
structures are in one cacheline. This is particularly important
for the kmalloc array.
Cons:
- Additional reference to one read only cacheline (per cpu
array of pointers to kmem_cache_cpu) in both slab_alloc()
and slab_free().
[akinobu.mita@gmail.com: fix cpu hotplug offline/online path]
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Cc: "Pekka Enberg" <penberg@cs.helsinki.fi>
Cc: Akinobu Mita <akinobu.mita@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
We need the offset from the page struct during slab_alloc and slab_free. In
both cases we also reference the cacheline of the kmem_cache_cpu structure.
We can therefore move the offset field into the kmem_cache_cpu structure
freeing up 16 bits in the page struct.
Moving the offset allows an allocation from slab_alloc() without touching the
page struct in the hot path.
The only thing left in slab_free() that touches the page struct cacheline for
per cpu freeing is the checking of SlabDebug(page). The next patch deals with
that.
Use the available 16 bits to broaden page->inuse. More than 64k objects per
slab become possible and we can get rid of the checks for that limitation.
No need anymore to shrink the order of slabs if we boot with 2M sized slabs
(slub_min_order=9).
No need anymore to switch off the offset calculation for very large slabs
since the field in the kmem_cache_cpu structure is 32 bits and so the offset
field can now handle slab sizes of up to 8GB.
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
A remote free may access the same page struct that also contains the lockless
freelist for the cpu slab. If objects have a short lifetime and are freed by
a different processor then remote frees back to the slab from which we are
currently allocating are frequent. The cacheline with the page struct needs
to be repeately acquired in exclusive mode by both the allocating thread and
the freeing thread. If this is frequent enough then performance will suffer
because of cacheline bouncing.
This patchset puts the lockless_freelist pointer in its own cacheline. In
order to make that happen we introduce a per cpu structure called
kmem_cache_cpu.
Instead of keeping an array of pointers to page structs we now keep an array
to a per cpu structure that--among other things--contains the pointer to the
lockless freelist. The freeing thread can then keep possession of exclusive
access to the page struct cacheline while the allocating thread keeps its
exclusive access to the cacheline containing the per cpu structure.
This works as long as the allocating cpu is able to service its request
from the lockless freelist. If the lockless freelist runs empty then the
allocating thread needs to acquire exclusive access to the cacheline with
the page struct lock the slab.
The allocating thread will then check if new objects were freed to the per
cpu slab. If so it will keep the slab as the cpu slab and continue with the
recently remote freed objects. So the allocating thread can take a series
of just freed remote pages and dish them out again. Ideally allocations
could be just recycling objects in the same slab this way which will lead
to an ideal allocation / remote free pattern.
The number of objects that can be handled in this way is limited by the
capacity of one slab. Increasing slab size via slub_min_objects/
slub_max_order may increase the number of objects and therefore performance.
If the allocating thread runs out of objects and finds that no objects were
put back by the remote processor then it will retrieve a new slab (from the
partial lists or from the page allocator) and start with a whole
new set of objects while the remote thread may still be freeing objects to
the old cpu slab. This may then repeat until the new slab is also exhausted.
If remote freeing has freed objects in the earlier slab then that earlier
slab will now be on the partial freelist and the allocating thread will
pick that slab next for allocation. So the loop is extended. However,
both threads need to take the list_lock to make the swizzling via
the partial list happen.
It is likely that this kind of scheme will keep the objects being passed
around to a small set that can be kept in the cpu caches leading to increased
performance.
More code cleanups become possible:
- Instead of passing a cpu we can now pass a kmem_cache_cpu structure around.
Allows reducing the number of parameters to various functions.
- Can define a new node_match() function for NUMA to encapsulate locality
checks.
Effect on allocations:
Cachelines touched before this patch:
Write: page cache struct and first cacheline of object
Cachelines touched after this patch:
Write: kmem_cache_cpu cacheline and first cacheline of object
Read: page cache struct (but see later patch that avoids touching
that cacheline)
The handling when the lockless alloc list runs empty gets to be a bit more
complicated since another cacheline has now to be written to. But that is
halfway out of the hot path.
Effect on freeing:
Cachelines touched before this patch:
Write: page_struct and first cacheline of object
Cachelines touched after this patch depending on how we free:
Write(to cpu_slab): kmem_cache_cpu struct and first cacheline of object
Write(to other): page struct and first cacheline of object
Read(to cpu_slab): page struct to id slab etc. (but see later patch that
avoids touching the page struct on free)
Read(to other): cpu local kmem_cache_cpu struct to verify its not
the cpu slab.
Summary:
Pro:
- Distinct cachelines so that concurrent remote frees and local
allocs on a cpuslab can occur without cacheline bouncing.
- Avoids potential bouncing cachelines because of neighboring
per cpu pointer updates in kmem_cache's cpu_slab structure since
it now grows to a cacheline (Therefore remove the comment
that talks about that concern).
Cons:
- Freeing objects now requires the reading of one additional
cacheline. That can be mitigated for some cases by the following
patches but its not possible to completely eliminate these
references.
- Memory usage grows slightly.
The size of each per cpu object is blown up from one word
(pointing to the page_struct) to one cacheline with various data.
So this is NR_CPUS*NR_SLABS*L1_BYTES more memory use. Lets say
NR_SLABS is 100 and a cache line size of 128 then we have just
increased SLAB metadata requirements by 12.8k per cpu.
(Another later patch reduces these requirements)
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This gets rid of all kmalloc caches larger than page size. A kmalloc
request larger than PAGE_SIZE > 2 is going to be passed through to the page
allocator. This works both inline where we will call __get_free_pages
instead of kmem_cache_alloc and in __kmalloc.
kfree is modified to check if the object is in a slab page. If not then
the page is freed via the page allocator instead. Roughly similar to what
SLOB does.
Advantages:
- Reduces memory overhead for kmalloc array
- Large kmalloc operations are faster since they do not
need to pass through the slab allocator to get to the
page allocator.
- Performance increase of 10%-20% on alloc and 50% on free for
PAGE_SIZEd allocations.
SLUB must call page allocator for each alloc anyways since
the higher order pages which that allowed avoiding the page alloc calls
are not available in a reliable way anymore. So we are basically removing
useless slab allocator overhead.
- Large kmallocs yields page aligned object which is what
SLAB did. Bad things like using page sized kmalloc allocations to
stand in for page allocate allocs can be transparently handled and are not
distinguishable from page allocator uses.
- Checking for too large objects can be removed since
it is done by the page allocator.
Drawbacks:
- No accounting for large kmalloc slab allocations anymore
- No debugging of large kmalloc slab allocations.
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Some compilers (especially older gcc releases) may skip inlining
sometimes which will lead to link failures. Force the inlining of
keyfunctions in slub_def.h to avoid these issues.
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Acked-by: Jan Dittmer <jdi@l4x.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Since we have use like ~SLUB_DMA, we ought to have the type
set right in both cases.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
It becomes now easy to support the zeroing allocs with generic inline
functions in slab.h. Provide inline definitions to allow the continued use of
kzalloc, kmem_cache_zalloc etc but remove other definitions of zeroing
functions from the slab allocators and util.c.
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Add #ifdefs around data structures only needed if debugging is compiled into
SLUB.
Add inlines to small functions to reduce code size.
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Define ZERO_OR_NULL_PTR macro to be able to remove the checks from the
allocators. Move ZERO_SIZE_PTR related stuff into slab.h.
Make ZERO_SIZE_PTR work for all slab allocators and get rid of the
WARN_ON_ONCE(size == 0) that is still remaining in SLAB.
Make slub return NULL like the other allocators if a too large memory segment
is requested via __kmalloc.
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Acked-by: Pekka Enberg <penberg@cs.helsinki.fi>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This adds preliminary NUMA support to SLOB, primarily aimed at systems with
small nodes (tested all the way down to a 128kB SRAM block), whether
asymmetric or otherwise.
We follow the same conventions as SLAB/SLUB, preferring current node
placement for new pages, or with explicit placement, if a node has been
specified. Presently on UP NUMA this has the side-effect of preferring
node#0 allocations (since numa_node_id() == 0, though this could be
reworked if we could hand off a pfn to determine node placement), so
single-CPU NUMA systems will want to place smaller nodes further out in
terms of node id. Once a page has been bound to a node (via explicit node
id typing), we only do block allocations from partial free pages that have
a matching node id in the page flags.
The current implementation does have some scalability problems, in that all
partial free pages are tracked in the global freelist (with contention due
to the single spinlock). However, these are things that are being reworked
for SMP scalability first, while things like per-node freelists can easily
be built on top of this sort of functionality once it's been added.
More background can be found in:
http://marc.info/?l=linux-mm&m=118117916022379&w=2http://marc.info/?l=linux-mm&m=118170446306199&w=2http://marc.info/?l=linux-mm&m=118187859420048&w=2
and subsequent threads.
Acked-by: Christoph Lameter <clameter@sgi.com>
Acked-by: Matt Mackall <mpm@selenic.com>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
Acked-by: Nick Piggin <nickpiggin@yahoo.com.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
If ARCH_KMALLOC_MINALIGN is set to a value greater than 8 (SLUBs smallest
kmalloc cache) then SLUB may generate duplicate slabs in sysfs (yes again)
because the object size is padded to reach ARCH_KMALLOC_MINALIGN. Thus the
size of the small slabs is all the same.
No arch sets ARCH_KMALLOC_MINALIGN larger than 8 though except mips which
for some reason wants a 128 byte alignment.
This patch increases the size of the smallest cache if
ARCH_KMALLOC_MINALIGN is greater than 8. In that case more and more of the
smallest caches are disabled.
If we do that then the count of the active general caches that is displayed
on boot is not correct anymore since we may skip elements of the kmalloc
array. So count them separately.
This approach was tested by Havard yesterday.
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Cc: Haavard Skinnemoen <hskinnemoen@atmel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Instead of returning the smallest available object return ZERO_SIZE_PTR.
A ZERO_SIZE_PTR can be legitimately used as an object pointer as long as it
is not deferenced. The dereference of ZERO_SIZE_PTR causes a distinctive
fault. kfree can handle a ZERO_SIZE_PTR in the same way as NULL.
This enables functions to use zero sized object. e.g. n = number of objects.
objects = kmalloc(n * sizeof(object));
for (i = 0; i < n; i++)
objects[i].x = y;
kfree(objects);
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Acked-by: Pekka Enberg <penberg@cs.helsinki.fi>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Currently we have a maze of configuration variables that determine the
maximum slab size. Worst of all it seems to vary between SLAB and SLUB.
So define a common maximum size for kmalloc. For conveniences sake we use
the maximum size ever supported which is 32 MB. We limit the maximum size
to a lower limit if MAX_ORDER does not allow such large allocations.
For many architectures this patch will have the effect of adding large
kmalloc sizes. x86_64 adds 5 new kmalloc sizes. So a small amount of
memory will be needed for these caches (contemporary SLAB has dynamically
sizeable node and cpu structure so the waste is less than in the past)
Most architectures will then be able to allocate object with sizes up to
MAX_ORDER. We have had repeated breakage (in fact whenever we doubled the
number of supported processors) on IA64 because one or the other struct
grew beyond what the slab allocators supported. This will avoid future
issues and f.e. avoid fixes for 2k and 4k cpu support.
CONFIG_LARGE_ALLOCS is no longer necessary so drop it.
It fixes sparc64 with SLAB.
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: "David S. Miller" <davem@davemloft.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
I'm getting zillions of undefined references to __kmalloc_size_too_large on
alpha. For some reason alpha is building out-of-line copies of kmalloc_slab()
into lots of compilation units.
It turns out that gcc just isn't smart enough to work out that
__builtin_contant_p(size)==true implies that __builtin_contant_p(index)==true.
So let's give it a bit of help.
Cc: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
There is no user of destructors left. There is no reason why we should keep
checking for destructors calls in the slab allocators.
The RFC for this patch was discussed at
http://marc.info/?l=linux-kernel&m=117882364330705&w=2
Destructors were mainly used for list management which required them to take a
spinlock. Taking a spinlock in a destructor is a bit risky since the slab
allocators may run the destructors anytime they decide a slab is no longer
needed.
Patch drops destructor support. Any attempt to use a destructor will BUG().
Acked-by: Pekka Enberg <penberg@cs.helsinki.fi>
Acked-by: Paul Mundt <lethal@linux-sh.org>
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Sorry I screwed up the comparison. It is only an error if we attempt
to allocate a slab larger than the maximum allowed size.
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Take MAX_ORDER into consideration when determining KMALLOC_SHIFT_HIGH.
Otherwise we may run into a situation where we attempt to create general
slabs larger than MAX_ORDER.
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Cc: "David S. Miller" <davem@davemloft.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
If slab tracking is on then build a list of full slabs so that we can verify
the integrity of all slabs and are also able to built list of alloc/free
callers.
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Makes SLUB behave like SLAB in this area to avoid issues....
Throw a stack dump to alert people.
At some point the behavior should be switched back. NULL is no memory as
far as I can tell and if the use asked for 0 bytes then he need to get no
memory.
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This is a new slab allocator which was motivated by the complexity of the
existing code in mm/slab.c. It attempts to address a variety of concerns
with the existing implementation.
A. Management of object queues
A particular concern was the complex management of the numerous object
queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for
each allocating CPU and use objects from a slab directly instead of
queueing them up.
B. Storage overhead of object queues
SLAB Object queues exist per node, per CPU. The alien cache queue even
has a queue array that contain a queue for each processor on each
node. For very large systems the number of queues and the number of
objects that may be caught in those queues grows exponentially. On our
systems with 1k nodes / processors we have several gigabytes just tied up
for storing references to objects for those queues This does not include
the objects that could be on those queues. One fears that the whole
memory of the machine could one day be consumed by those queues.
C. SLAB meta data overhead
SLAB has overhead at the beginning of each slab. This means that data
cannot be naturally aligned at the beginning of a slab block. SLUB keeps
all meta data in the corresponding page_struct. Objects can be naturally
aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte
boundaries and can fit tightly into a 4k page with no bytes left over.
SLAB cannot do this.
D. SLAB has a complex cache reaper
SLUB does not need a cache reaper for UP systems. On SMP systems
the per CPU slab may be pushed back into partial list but that
operation is simple and does not require an iteration over a list
of objects. SLAB expires per CPU, shared and alien object queues
during cache reaping which may cause strange hold offs.
E. SLAB has complex NUMA policy layer support
SLUB pushes NUMA policy handling into the page allocator. This means that
allocation is coarser (SLUB does interleave on a page level) but that
situation was also present before 2.6.13. SLABs application of
policies to individual slab objects allocated in SLAB is
certainly a performance concern due to the frequent references to
memory policies which may lead a sequence of objects to come from
one node after another. SLUB will get a slab full of objects
from one node and then will switch to the next.
F. Reduction of the size of partial slab lists
SLAB has per node partial lists. This means that over time a large
number of partial slabs may accumulate on those lists. These can
only be reused if allocator occur on specific nodes. SLUB has a global
pool of partial slabs and will consume slabs from that pool to
decrease fragmentation.
G. Tunables
SLAB has sophisticated tuning abilities for each slab cache. One can
manipulate the queue sizes in detail. However, filling the queues still
requires the uses of the spin lock to check out slabs. SLUB has a global
parameter (min_slab_order) for tuning. Increasing the minimum slab
order can decrease the locking overhead. The bigger the slab order the
less motions of pages between per CPU and partial lists occur and the
better SLUB will be scaling.
G. Slab merging
We often have slab caches with similar parameters. SLUB detects those
on boot up and merges them into the corresponding general caches. This
leads to more effective memory use. About 50% of all caches can
be eliminated through slab merging. This will also decrease
slab fragmentation because partial allocated slabs can be filled
up again. Slab merging can be switched off by specifying
slub_nomerge on boot up.
Note that merging can expose heretofore unknown bugs in the kernel
because corrupted objects may now be placed differently and corrupt
differing neighboring objects. Enable sanity checks to find those.
H. Diagnostics
The current slab diagnostics are difficult to use and require a
recompilation of the kernel. SLUB contains debugging code that
is always available (but is kept out of the hot code paths).
SLUB diagnostics can be enabled via the "slab_debug" option.
Parameters can be specified to select a single or a group of
slab caches for diagnostics. This means that the system is running
with the usual performance and it is much more likely that
race conditions can be reproduced.
I. Resiliency
If basic sanity checks are on then SLUB is capable of detecting
common error conditions and recover as best as possible to allow the
system to continue.
J. Tracing
Tracing can be enabled via the slab_debug=T,<slabcache> option
during boot. SLUB will then protocol all actions on that slabcache
and dump the object contents on free.
K. On demand DMA cache creation.
Generally DMA caches are not needed. If a kmalloc is used with
__GFP_DMA then just create this single slabcache that is needed.
For systems that have no ZONE_DMA requirement the support is
completely eliminated.
L. Performance increase
Some benchmarks have shown speed improvements on kernbench in the
range of 5-10%. The locking overhead of slub is based on the
underlying base allocation size. If we can reliably allocate
larger order pages then it is possible to increase slub
performance much further. The anti-fragmentation patches may
enable further performance increases.
Tested on:
i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator
SLUB Boot options
slub_nomerge Disable merging of slabs
slub_min_order=x Require a minimum order for slab caches. This
increases the managed chunk size and therefore
reduces meta data and locking overhead.
slub_min_objects=x Mininum objects per slab. Default is 8.
slub_max_order=x Avoid generating slabs larger than order specified.
slub_debug Enable all diagnostics for all caches
slub_debug=<options> Enable selective options for all caches
slub_debug=<o>,<cache> Enable selective options for a certain set of
caches
Available Debug options
F Double Free checking, sanity and resiliency
R Red zoning
P Object / padding poisoning
U Track last free / alloc
T Trace all allocs / frees (only use for individual slabs).
To use SLUB: Apply this patch and then select SLUB as the default slab
allocator.
[hugh@veritas.com: fix an oops-causing locking error]
[akpm@linux-foundation.org: various stupid cleanups and small fixes]
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Christoph Lameter