Since CONFIG_HIGHMEM got enabled on ARMv5 Kirkwood, we have noticed a
very significant drop in networking performance. The test were
conducted on an OpenBlocks A7 board. Without this patch, the outgoing
performance measured with iperf are:
- highmem OFF, TSO OFF 544 Mbit/s
- highmem OFF, TSO ON 942 Mbit/s
- highmem ON, TSO OFF 306 Mbit/s
- highmem ON, TSO ON 246 Mbit/s
On this Kirkwood platform, the L2 cache is a Feroceon cache, and with
this cache, all the range operations have to be done on virtual
addresses and not physical addresses. Therefore, whenever
CONFIG_HIGHMEM is enabled, the cache maintenance operations call
kmap_atomic_pfn() and kunmap_atomic().
However, kmap_atomic_pfn() does not implement the same fast path for
non-highmem pages as the one implemented in kmap_atomic(), and this is
one of the reason for the performance drop. While this patch does not
fully restore the performances, it clearly improves them a lot:
without patch with patch
- highmem ON, TSO OFF 306 Mbit/s 387 Mbit/s
- highmem ON, TSO ON 246 Mbit/s 434 Mbit/s
We're still far from the !CONFIG_HIGHMEM performances, but it does
improve a bit the situation.
Thanks a lot to Ezequiel Garcia and Gregory Clement for all the
testing work around this topic.
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
In 32-bit ARM systems, the fixmap mapping region can support no more
than 14 CPUs(total: 896k; one CPU: 64K). And we can configure NR_CPUS
up to 32. So there is a mismatch.
This patch moves fixmapping region downwards to region 0xffc00000-
0xffe00000. Then the fixmap mapping region can support up to 32 CPUs.
Reviewed-by: Nicolas Pitre <nico@linaro.org>
Signed-off-by: Liu Hua <sdu.liu@huawei.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
It seems that these two macros are not used by non architecture
specific code. And on ARM FIX_KMAP_BEGIN equals zero.
This patch removes these two macros. Instead, using FIX_KMAP_NR_PTES to
tell the pte number belonged to fixmap mapping region. The code will
become clearer when I introduce a bugfix on fixmap mapping region.
Reviewed-by: Nicolas Pitre <nico@linaro.org>
Signed-off-by: Liu Hua <sdu.liu@huawei.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
Pull more ARM updates from Russell King.
This got a fair number of conflicts with the <asm/system.h> split, but
also with some other sparse-irq and header file include cleanups. They
all looked pretty trivial, though.
* 'for-linus' of git://git.linaro.org/people/rmk/linux-arm: (59 commits)
ARM: fix Kconfig warning for HAVE_BPF_JIT
ARM: 7361/1: provide XIP_VIRT_ADDR for no-MMU builds
ARM: 7349/1: integrator: convert to sparse irqs
ARM: 7259/3: net: JIT compiler for packet filters
ARM: 7334/1: add jump label support
ARM: 7333/2: jump label: detect %c support for ARM
ARM: 7338/1: add support for early console output via semihosting
ARM: use set_current_blocked() and block_sigmask()
ARM: exec: remove redundant set_fs(USER_DS)
ARM: 7332/1: extract out code patch function from kprobes
ARM: 7331/1: extract out insn generation code from ftrace
ARM: 7330/1: ftrace: use canonical Thumb-2 wide instruction format
ARM: 7351/1: ftrace: remove useless memory checks
ARM: 7316/1: kexec: EOI active and mask all interrupts in kexec crash path
ARM: Versatile Express: add NO_IOPORT
ARM: get rid of asm/irq.h in asm/prom.h
ARM: 7319/1: Print debug info for SIGBUS in user faults
ARM: 7318/1: gic: refactor irq_start assignment
ARM: 7317/1: irq: avoid NULL check in for_each_irq_desc loop
ARM: 7315/1: perf: add support for the Cortex-A7 PMU
...
[swarren@nvidia.com: highmem: Fix ARM build break due to __kmap_atomic rename]
Signed-off-by: Stephen Warren <swarren@nvidia.com>
Signed-off-by: Cong Wang <amwang@redhat.com>
Provide get_top_pte() to complement set_top_pte(), moving the only
users of TOP_PTE to arch/arm/mm/mm.h.
Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
A number of places establish a PTE in our top page table and
immediately flush the TLB. Rather than having this at every callsite,
provide an inline function for this purpose.
This changes some global tlb flushes to be local; each time we setup
one of these mappings, we always do it with preemption disabled which
would prevent us migrating to another CPU.
Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
Since commit 3e4d3af501 "mm: stack based kmap_atomic()", it is no longer
necessary to carry an ad hoc version of kmap_atomic() added in commit
7e5a69e83b "ARM: 6007/1: fix highmem with VIPT cache and DMA" to cope
with reentrancy.
In fact, it is now actively wrong to rely on fixed kmap type indices
(namely KM_L1_CACHE) as kmap_atomic() totally ignores them now and a
concurrent instance of it may reuse any slot for any purpose.
Signed-off-by: Nicolas Pitre <nicolas.pitre@linaro.org>
Christoph reported a nice splat which illustrated a race in the new stack
based kmap_atomic implementation.
The problem is that we pop our stack slot before we're completely done
resetting its state -- in particular clearing the PTE (sometimes that's
CONFIG_DEBUG_HIGHMEM). If an interrupt happens before we actually clear
the PTE used for the last slot, that interrupt can reuse the slot in a
dirty state, which triggers a BUG in kmap_atomic().
Fix this by introducing kmap_atomic_idx() which reports the current slot
index without actually releasing it and use that to find the PTE and delay
the _pop() until after we're completely done.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Reported-by: Christoph Hellwig <hch@infradead.org>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Keep the current interface but ignore the KM_type and use a stack based
approach.
The advantage is that we get rid of crappy code like:
#define __KM_PTE \
(in_nmi() ? KM_NMI_PTE : \
in_irq() ? KM_IRQ_PTE : \
KM_PTE0)
and in general can stop worrying about what context we're in and what kmap
slots might be appropriate for that.
The downside is that FRV kmap_atomic() gets more expensive.
For now we use a CPP trick suggested by Andrew:
#define kmap_atomic(page, args...) __kmap_atomic(page)
to avoid having to touch all kmap_atomic() users in a single patch.
[ not compiled on:
- mn10300: the arch doesn't actually build with highmem to begin with ]
[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: fix up drivers/gpu/drm/i915/intel_overlay.c]
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Chris Metcalf <cmetcalf@tilera.com>
Cc: David Howells <dhowells@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Russell King <rmk@arm.linux.org.uk>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: David Miller <davem@davemloft.net>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Dave Airlie <airlied@linux.ie>
Cc: Li Zefan <lizf@cn.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
kunmap_atomic() is currently at level -4 on Rusty's "Hard To Misuse"
list[1] ("Follow common convention and you'll get it wrong"), except in
some architectures when CONFIG_DEBUG_HIGHMEM is set[2][3].
kunmap() takes a pointer to a struct page; kunmap_atomic(), however, takes
takes a pointer to within the page itself. This seems to once in a while
trip people up (the convention they are following is the one from
kunmap()).
Make it much harder to misuse, by moving it to level 9 on Rusty's list[4]
("The compiler/linker won't let you get it wrong"). This is done by
refusing to build if the type of its first argument is a pointer to a
struct page.
The real kunmap_atomic() is renamed to kunmap_atomic_notypecheck()
(which is what you would call in case for some strange reason calling it
with a pointer to a struct page is not incorrect in your code).
The previous version of this patch was compile tested on x86-64.
[1] http://ozlabs.org/~rusty/index.cgi/tech/2008-04-01.html
[2] In these cases, it is at level 5, "Do it right or it will always
break at runtime."
[3] At least mips and powerpc look very similar, and sparc also seems to
share a common ancestor with both; there seems to be quite some
degree of copy-and-paste coding here. The include/asm/highmem.h file
for these three archs mention x86 CPUs at its top.
[4] http://ozlabs.org/~rusty/index.cgi/tech/2008-03-30.html
[5] As an aside, could someone tell me why mn10300 uses unsigned long as
the first parameter of kunmap_atomic() instead of void *?
Signed-off-by: Cesar Eduardo Barros <cesarb@cesarb.net>
Cc: Russell King <linux@arm.linux.org.uk> (arch/arm)
Cc: Ralf Baechle <ralf@linux-mips.org> (arch/mips)
Cc: David Howells <dhowells@redhat.com> (arch/frv, arch/mn10300)
Cc: Koichi Yasutake <yasutake.koichi@jp.panasonic.com> (arch/mn10300)
Cc: Kyle McMartin <kyle@mcmartin.ca> (arch/parisc)
Cc: Helge Deller <deller@gmx.de> (arch/parisc)
Cc: "James E.J. Bottomley" <jejb@parisc-linux.org> (arch/parisc)
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> (arch/powerpc)
Cc: Paul Mackerras <paulus@samba.org> (arch/powerpc)
Cc: "David S. Miller" <davem@davemloft.net> (arch/sparc)
Cc: Thomas Gleixner <tglx@linutronix.de> (arch/x86)
Cc: Ingo Molnar <mingo@redhat.com> (arch/x86)
Cc: "H. Peter Anvin" <hpa@zytor.com> (arch/x86)
Cc: Arnd Bergmann <arnd@arndb.de> (include/asm-generic)
Cc: Rusty Russell <rusty@rustcorp.com.au> ("Hard To Misuse" list)
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
smp_processor_id() must not be called from a preemptible context (this
is checked by CONFIG_DEBUG_PREEMPT). kmap_high_l1_vipt() was doing so.
This lead to a problem where the wrong per_cpu kmap_high_l1_vipt_depth
could be incremented, causing a BUG_ON(*depth <= 0); in
kunmap_high_l1_vipt().
The solution is to move the call to smp_processor_id() after the call
to preempt_disable().
Originally by: Andrew Howe <ahowe@nvidia.com>
Signed-off-by: Gary King <gking@nvidia.com>
Acked-by: Nicolas Pitre <nico.as.pitre@linaro.org>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
When CONFIG_DEBUG_HIGHMEM is used, the fixmap entry used for a highmem page
by kmap_atomic() is always cleared by kunmap_atomic(). This helps find
bad usages such as dereferences after the unmap, or overflow into the
adjacent fixmap areas.
But this debugging aid is completely bypassed when a kmap for the same
page already exists as the kmap is reused instead. ON VIVT systems we
have no choice but to reuse that kmap due to cache coherency issues,
but on non VIVT systems we should always force the fixmap usage when
debugging is active.
Signed-off-by: Nicolas Pitre <nicolas.pitre@linaro.org>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
The VIVT cache of a highmem page is always flushed before the page
is unmapped. This cache flush is explicit through flush_cache_kmaps()
in flush_all_zero_pkmaps(), or through __cpuc_flush_dcache_area() in
kunmap_atomic(). There is also an implicit flush of those highmem pages
that were part of a process that just terminated making those pages free
as the whole VIVT cache has to be flushed on every task switch. Hence
unmapped highmem pages need no cache maintenance in that case.
However unmapped pages may still be cached with a VIPT cache because the
cache is tagged with physical addresses. There is no need for a whole
cache flush during task switching for that reason, and despite the
explicit cache flushes in flush_all_zero_pkmaps() and kunmap_atomic(),
some highmem pages that were mapped in user space end up still cached
even when they become unmapped.
So, we do have to perform cache maintenance on those unmapped highmem
pages in the context of DMA when using a VIPT cache. Unfortunately,
it is not possible to perform that cache maintenance using physical
addresses as all the L1 cache maintenance coprocessor functions accept
virtual addresses only. Therefore we have no choice but to set up a
temporary virtual mapping for that purpose.
And of course the explicit cache flushing when unmapping a highmem page
on a system with a VIPT cache now can go, which should increase
performance.
While at it, because the code in __flush_dcache_page() has to be modified
anyway, let's also make sure the mapped highmem pages are pinned with
kmap_high_get() for the duration of the cache maintenance operation.
Because kunmap() does unmap highmem pages lazily, it was reported by
Gary King <GKing@nvidia.com> that those pages ended up being unmapped
during cache maintenance on SMP causing segmentation faults.
Signed-off-by: Nicolas Pitre <nico@marvell.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
Seemingly this support was missed when highmem was added, so
DEBUG_HIGHMEM wouldn't have checked the kmap_atomic type.
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
Let's suppose a highmem page is kmap'd with kmap(). A pkmap entry is
used, the page mapped to it, and the virtual cache is dirtied. Then
kunmap() is used which does virtually nothing except for decrementing a
usage count.
Then, let's suppose the _same_ page gets mapped using kmap_atomic().
It is therefore mapped onto a fixmap entry instead, which has a
different virtual address unaware of the dirty cache data for that page
sitting in the pkmap mapping.
Fortunately it is easy to know if a pkmap mapping still exists for that
page and use it directly with kmap_atomic(), thanks to kmap_high_get().
And actual testing with a printk in the added code path shows that this
condition is actually met *extremely* frequently. Seems that we've been
quite lucky that things have worked so well with highmem so far.
Cc: stable@kernel.org
Signed-off-by: Nicolas Pitre <nico@marvell.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
The kmap virtual area borrows a 2MB range at the top of the 16MB area
below PAGE_OFFSET currently reserved for kernel modules and/or the
XIP kernel. This 2MB corresponds to the range covered by 2 consecutive
second-level page tables, or a single pmd entry as seen by the Linux
page table abstraction. Because XIP kernels are unlikely to be seen
on systems needing highmem support, there shouldn't be any shortage of
VM space for modules (14 MB for modules is still way more than twice the
typical usage).
Because the virtual mapping of highmem pages can go away at any moment
after kunmap() is called on them, we need to bypass the delayed cache
flushing provided by flush_dcache_page() in that case.
The atomic kmap versions are based on fixmaps, and
__cpuc_flush_dcache_page() is used directly in that case.
Signed-off-by: Nicolas Pitre <nico@marvell.com>
Thomas Petazzoni