PA-RISC uses a global spinlock to protect pagetable updates in the TLB
fault handlers. When multiple cores are taking TLB faults simultaneously,
the cache line containing the spinlock becomes a bottleneck.
This patch embeds the spinlock in the top level page directory, so that
every process has its own lock. It improves performance by 30% when
doing parallel compilations.
At least on the N class systems, only one PxTLB inter processor
broadcast can be active at any one time on the Merced bus. If a Merced
bus is found, this patch serializes the TLB flushes with the
pa_tlb_flush_lock spinlock.
v1: Initial patch by Mikulas
v2: Added Merced detection by Helge
v3: Revised TLB serialization by Dave & Helge
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: John David Anglin <dave.anglin@bell.net>
Signed-off-by: Helge Deller <deller@gmx.de>
The attached patch implements three optimizations:
1) Loops in flush_user_dcache_range_asm, flush_kernel_dcache_range_asm,
purge_kernel_dcache_range_asm, flush_user_icache_range_asm, and
flush_kernel_icache_range_asm are unrolled to reduce branch overhead.
2) The static branch prediction for cmpb instructions in pacache.S have
been reviewed and the operand order adjusted where necessary.
3) For flush routines in cache.c, we purge rather flush when we have no
context. The pdc instruction at level 0 is not required to write back
dirty lines to memory. This provides a performance improvement over the
fdc instruction if the feature is implemented.
Version 2 adds alternative patching.
The patch provides an average improvement of about 2%.
Signed-off-by: John David Anglin <dave.anglin@bell.net>
Signed-off-by: Helge Deller <deller@gmx.de>
This patch adds the necessary code to patch a running kernel at runtime
to improve performance.
The current implementation offers a few optimizations variants:
- When running a SMP kernel on a single UP processor, unwanted assembler
statements like locking functions are overwritten with NOPs. When
multiple instructions shall be skipped, one branch instruction is used
instead of multiple nop instructions.
- In the UP case, some pdtlb and pitlb instructions are patched to
become pdtlb,l and pitlb,l which only flushes the CPU-local tlb
entries instead of broadcasting the flush to other CPUs in the system
and thus may improve performance.
- fic and fdc instructions are skipped if no I- or D-caches are
installed. This should speed up qemu emulation and cacheless systems.
- If no cache coherence is needed for IO operations, the relevant fdc
and sync instructions in the sba and ccio drivers are replaced by
nops.
- On systems which share I- and D-TLBs and thus don't have a seperate
instruction TLB, the pitlb instruction is replaced by a nop.
Live-patching is done early in the boot process, just after having run
the system inventory. No drivers are running and thus no external
interrupts should arrive. So the hope is that no TLB exceptions will
occur during the patching. If this turns out to be wrong we will
probably need to do the patching in real-mode.
Signed-off-by: Helge Deller <deller@gmx.de>
On boot (mostly reboot), my c8000 sometimes crashes after it prints the
TLB flush threshold. The lockup is hard. The front LED flashes red and
the box must be unplugged to reset the error.
I noticed that when the crash occurs the TLB flush threshold is about
one quarter what it is on a successful boot. If I disabled the
calculation, the crash didn't occur. There also seemed to be a timing
dependency affecting the crash. I finally realized that the
flush_tlb_all() timing test runs just after the secondary CPUs are
started. There seems to be a problem with running flush_tlb_all() too
soon after the CPUs are started.
The timing for the range test always seemed okay. So, I reversed the
order of the two timing tests and I haven't had a crash at this point so
far.
I added a couple of information messages which I have left to help with
diagnosis if the problem should appear on another machine.
This version reduces the minimum TLB flush threshold to 16 KiB.
Signed-off-by: John David Anglin <dave.anglin@bell.net>
Signed-off-by: Helge Deller <deller@gmx.de>
Thanks to commit 4b3ef9daa4 ("mm/swap: split swap cache into 64MB
trunks"), after swapoff the address_space associated with the swap
device will be freed. So page_mapping() users which may touch the
address_space need some kind of mechanism to prevent the address_space
from being freed during accessing.
The dcache flushing functions (flush_dcache_page(), etc) in architecture
specific code may access the address_space of swap device for anonymous
pages in swap cache via page_mapping() function. But in some cases
there are no mechanisms to prevent the swap device from being swapoff,
for example,
CPU1 CPU2
__get_user_pages() swapoff()
flush_dcache_page()
mapping = page_mapping()
... exit_swap_address_space()
... kvfree(spaces)
mapping_mapped(mapping)
The address space may be accessed after being freed.
But from cachetlb.txt and Russell King, flush_dcache_page() only care
about file cache pages, for anonymous pages, flush_anon_page() should be
used. The implementation of flush_dcache_page() in all architectures
follows this too. They will check whether page_mapping() is NULL and
whether mapping_mapped() is true to determine whether to flush the
dcache immediately. And they will use interval tree (mapping->i_mmap)
to find all user space mappings. While mapping_mapped() and
mapping->i_mmap isn't used by anonymous pages in swap cache at all.
So, to fix the race between swapoff and flush dcache, __page_mapping()
is add to return the address_space for file cache pages and NULL
otherwise. All page_mapping() invoking in flush dcache functions are
replaced with page_mapping_file().
[akpm@linux-foundation.org: simplify page_mapping_file(), per Mike]
Link: http://lkml.kernel.org/r/20180305083634.15174-1-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Chen Liqin <liqin.linux@gmail.com>
Cc: Russell King <linux@armlinux.org.uk>
Cc: Yoshinori Sato <ysato@users.sourceforge.jp>
Cc: "James E.J. Bottomley" <jejb@parisc-linux.org>
Cc: Guan Xuetao <gxt@mprc.pku.edu.cn>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Chris Zankel <chris@zankel.net>
Cc: Vineet Gupta <vgupta@synopsys.com>
Cc: Ley Foon Tan <lftan@altera.com>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Mike Rapoport <rppt@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Just when I had decided that flush_cache_range() was always called with
a valid context, Helge reported two cases where the
"BUG_ON(!vma->vm_mm->context);" was hit on the phantom buildd:
kernel BUG at /mnt/sdb6/linux/linux-4.15.4/arch/parisc/kernel/cache.c:587!
CPU: 1 PID: 3254 Comm: kworker/1:2 Tainted: G D 4.15.0-1-parisc64-smp #1 Debian 4.15.4-1+b1
Workqueue: events free_ioctx
IAOQ[0]: flush_cache_range+0x164/0x168
IAOQ[1]: flush_cache_page+0x0/0x1c8
RP(r2): unmap_page_range+0xae8/0xb88
Backtrace:
[<00000000404a6980>] unmap_page_range+0xae8/0xb88
[<00000000404a6ae0>] unmap_single_vma+0xc0/0x188
[<00000000404a6cdc>] zap_page_range_single+0x134/0x1f8
[<00000000404a702c>] unmap_mapping_range+0x1cc/0x208
[<0000000040461518>] truncate_pagecache+0x98/0x108
[<0000000040461624>] truncate_setsize+0x9c/0xb8
[<00000000405d7f30>] put_aio_ring_file+0x80/0x100
[<00000000405d803c>] aio_free_ring+0x8c/0x290
[<00000000405d82c0>] free_ioctx+0x80/0x180
[<0000000040284e6c>] process_one_work+0x21c/0x668
[<00000000402854c4>] worker_thread+0x20c/0x778
[<0000000040291d44>] kthread+0x2d4/0x2e0
[<0000000040204020>] end_fault_vector+0x20/0xc0
This indicates that we need to handle the no context case in
flush_cache_range() as we do in flush_cache_mm().
In thinking about this, I realized that we don't need to flush the TLB
when there is no context. So, I added context checks to the large flush
cases in flush_cache_mm() and flush_cache_range(). The large flush case
occurs frequently in flush_cache_mm() and the change should improve fork
performance.
The v2 version of this change removes the BUG_ON from flush_cache_page()
by skipping the TLB flush when there is no context. I also added code
to flush the TLB in flush_cache_mm() and flush_cache_range() when we
have a context that's not current. Now all three routines handle TLB
flushes in a similar manner.
Signed-off-by: John David Anglin <dave.anglin@bell.net>
Cc: stable@vger.kernel.org # 4.9+
Signed-off-by: Helge Deller <deller@gmx.de>
The change to flush_kernel_vmap_range() wasn't sufficient to avoid the
SMP stalls. The problem is some drivers call these routines with
interrupts disabled. Interrupts need to be enabled for flush_tlb_all()
and flush_cache_all() to work. This version adds checks to ensure
interrupts are not disabled before calling routines that need IPI
interrupts. When interrupts are disabled, we now drop into slower code.
The attached change fixes the ordering of cache and TLB flushes in
several cases. When we flush the cache using the existing PTE/TLB
entries, we need to flush the TLB after doing the cache flush. We don't
need to do this when we flush the entire instruction and data caches as
these flushes don't use the existing TLB entries. The same is true for
tmpalias region flushes.
The flush_kernel_vmap_range() and invalidate_kernel_vmap_range()
routines have been updated.
Secondly, we added a new purge_kernel_dcache_range_asm() routine to
pacache.S and use it in invalidate_kernel_vmap_range(). Nominally,
purges are faster than flushes as the cache lines don't have to be
written back to memory.
Hopefully, this is sufficient to resolve the remaining problems due to
cache speculation. So far, testing indicates that this is the case. I
did work up a patch using tmpalias flushes, but there is a performance
hit because we need the physical address for each page, and we also need
to sequence access to the tmpalias flush code. This increases the
probability of stalls.
Signed-off-by: John David Anglin <dave.anglin@bell.net>
Cc: stable@vger.kernel.org # 4.9+
Signed-off-by: Helge Deller <deller@gmx.de>
In testing James' patch to drivers/parisc/pdc_stable.c, I hit the BUG
statement in flush_cache_range() during a system shutdown:
kernel BUG at arch/parisc/kernel/cache.c:595!
CPU: 2 PID: 6532 Comm: kworker/2:0 Not tainted 4.13.0-rc2+ #1
Workqueue: events free_ioctx
IAOQ[0]: flush_cache_range+0x144/0x148
IAOQ[1]: flush_cache_page+0x0/0x1a8
RP(r2): flush_cache_range+0xec/0x148
Backtrace:
[<00000000402910ac>] unmap_page_range+0x84/0x880
[<00000000402918f4>] unmap_single_vma+0x4c/0x60
[<0000000040291a18>] zap_page_range_single+0x110/0x160
[<0000000040291c34>] unmap_mapping_range+0x174/0x1a8
[<000000004026ccd8>] truncate_pagecache+0x50/0xa8
[<000000004026cd84>] truncate_setsize+0x54/0x70
[<000000004033d534>] put_aio_ring_file+0x44/0xb0
[<000000004033d5d8>] aio_free_ring+0x38/0x140
[<000000004033d714>] free_ioctx+0x34/0xa8
[<00000000401b0028>] process_one_work+0x1b8/0x4d0
[<00000000401b04f4>] worker_thread+0x1b4/0x648
[<00000000401b9128>] kthread+0x1b0/0x208
[<0000000040150020>] end_fault_vector+0x20/0x28
[<0000000040639518>] nf_ip_reroute+0x50/0xa8
[<0000000040638ed0>] nf_ip_route+0x10/0x78
[<0000000040638c90>] xfrm4_mode_tunnel_input+0x180/0x1f8
CPU: 2 PID: 6532 Comm: kworker/2:0 Not tainted 4.13.0-rc2+ #1
Workqueue: events free_ioctx
Backtrace:
[<0000000040163bf0>] show_stack+0x20/0x38
[<0000000040688480>] dump_stack+0xa8/0x120
[<0000000040163dc4>] die_if_kernel+0x19c/0x2b0
[<0000000040164d0c>] handle_interruption+0xa24/0xa48
This patch modifies flush_cache_range() to handle non current contexts.
In as much as this occurs infrequently, the simplest approach is to
flush the entire cache when this happens.
Signed-off-by: John David Anglin <dave.anglin@bell.net>
Cc: stable@vger.kernel.org # 4.9+
Signed-off-by: Helge Deller <deller@gmx.de>
It's always bothered me that we only disable preemption in
copy_user_page around the call to flush_dcache_page_asm.
This patch extends this to after the copy.
Signed-off-by: John David Anglin <dave.anglin@bell.net>
Cc: stable@vger.kernel.org # 4.9+
Signed-off-by: Helge Deller <deller@gmx.de>
Helge noticed that we flush the TLB page in flush_cache_page but not in
flush_cache_range or flush_cache_mm.
For a long time, we have had random segmentation faults building
packages on machines with PA8800/8900 processors. These machines only
support equivalent aliases. We don't see these faults on machines that
don't require strict coherency. So, it appears TLB speculation
sometimes leads to cache corruption on machines that require coherency.
This patch adds TLB flushes to flush_cache_range and flush_cache_mm when
coherency is required. We only flush the TLB in flush_cache_page when
coherency is required.
The patch also optimizes flush_cache_range. It turns out we always have
the right context to use flush_user_dcache_range_asm and
flush_user_icache_range_asm.
The patch has been tested for some time on rp3440, rp3410 and A500-44.
It's been boot tested on c8000. No random segmentation faults were
observed during testing.
Signed-off-by: John David Anglin <dave.anglin@bell.net>
Cc: stable@vger.kernel.org # 4.9+
Signed-off-by: Helge Deller <deller@gmx.de>
The previously submitted patch did not resolve the random segmentation
faults observed on the phantom buildd system. There are still
unresolved problems with the Debian 4.8 and 4.9 kernels on C8000.
The attached patch removes the flush of the offset map pages and does a
whole data cache flush for large ranges. No other arch flushes the
offset map in these routines as far as I can tell.
I have not observed any random segmentation faults on rp3440 in two
weeks of testing with 4.10.0 and 4.10.1.
Signed-off-by: John David Anglin <dave.anglin@bell.net>
Cc: stable@vger.kernel.org # v4.8+
Signed-off-by: Helge Deller <deller@gmx.de>
We are going to split more MM APIs out of <linux/sched.h>, which
will have to be picked up from a couple of .c files.
The APIs that we are going to move are:
arch_pick_mmap_layout()
arch_get_unmapped_area()
arch_get_unmapped_area_topdown()
mm_update_next_owner()
Include the header in the files that are going to need it.
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The functions flush_user_dcache_range() and flush_user_icache_range()
are only used by the parisc signal handling code. This code only needs
to flush a couple of lines, so the threshold check is unnecessary
overhead.
Signed-off-by: John David Anglin <dave.anglin@bell.net>
Signed-off-by: Helge Deller <deller@gmx.de>
At bootup we run measurements to calculate the best threshold for when we
should be using full TLB flushes instead of just flushing a specific amount of
TLB entries. This performance test is run over the kernel text segment.
But running this TLB performance test on the kernel text segment turned out to
crash some SMP machines when the kernel text pages were mapped as huge pages.
To avoid those crashes this patch simply skips this test on some SMP machines
and calculates an optimal threshold based on the maximum number of available
TLB entries and number of online CPUs.
On a technical side, this seems to happen:
The TLB measurement code uses flush_tlb_kernel_range() to flush specific TLB
entries with a page size of 4k (pdtlb 0(sr1,addr)). On UP systems this purge
instruction seems to work without problems even if the pages were mapped as
huge pages. But on SMP systems the TLB purge instruction is broadcasted to
other CPUs. Those CPUs then crash the machine because the page size is not as
expected. C8000 machines with PA8800/PA8900 CPUs were not affected by this
problem, because the required cache coherency prohibits to use huge pages at
all. Sadly I didn't found any documentation about this behaviour, so this
finding is purely based on testing with phyiscal SMP machines (A500-44 and
J5000, both were 2-way boxes).
Cc: <stable@vger.kernel.org> # v3.18+
Signed-off-by: Helge Deller <deller@gmx.de>
PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time
ago with promise that one day it will be possible to implement page
cache with bigger chunks than PAGE_SIZE.
This promise never materialized. And unlikely will.
We have many places where PAGE_CACHE_SIZE assumed to be equal to
PAGE_SIZE. And it's constant source of confusion on whether
PAGE_CACHE_* or PAGE_* constant should be used in a particular case,
especially on the border between fs and mm.
Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much
breakage to be doable.
Let's stop pretending that pages in page cache are special. They are
not.
The changes are pretty straight-forward:
- <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>;
- <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>;
- PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN};
- page_cache_get() -> get_page();
- page_cache_release() -> put_page();
This patch contains automated changes generated with coccinelle using
script below. For some reason, coccinelle doesn't patch header files.
I've called spatch for them manually.
The only adjustment after coccinelle is revert of changes to
PAGE_CAHCE_ALIGN definition: we are going to drop it later.
There are few places in the code where coccinelle didn't reach. I'll
fix them manually in a separate patch. Comments and documentation also
will be addressed with the separate patch.
virtual patch
@@
expression E;
@@
- E << (PAGE_CACHE_SHIFT - PAGE_SHIFT)
+ E
@@
expression E;
@@
- E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT)
+ E
@@
@@
- PAGE_CACHE_SHIFT
+ PAGE_SHIFT
@@
@@
- PAGE_CACHE_SIZE
+ PAGE_SIZE
@@
@@
- PAGE_CACHE_MASK
+ PAGE_MASK
@@
expression E;
@@
- PAGE_CACHE_ALIGN(E)
+ PAGE_ALIGN(E)
@@
expression E;
@@
- page_cache_get(E)
+ get_page(E)
@@
expression E;
@@
- page_cache_release(E)
+ put_page(E)
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The increased use of pdtlb/pitlb instructions seemed to increase the
frequency of random segmentation faults building packages. Further, we
had a number of cases where TLB inserts would repeatedly fail and all
forward progress would stop. The Haskell ghc package caused a lot of
trouble in this area. The final indication of a race in pte handling was
this syslog entry on sibaris (C8000):
swap_free: Unused swap offset entry 00000004
BUG: Bad page map in process mysqld pte:00000100 pmd:019bbec5
addr:00000000ec464000 vm_flags:00100073 anon_vma:0000000221023828 mapping: (null) index:ec464
CPU: 1 PID: 9176 Comm: mysqld Not tainted 4.0.0-2-parisc64-smp #1 Debian 4.0.5-1
Backtrace:
[<0000000040173eb0>] show_stack+0x20/0x38
[<0000000040444424>] dump_stack+0x9c/0x110
[<00000000402a0d38>] print_bad_pte+0x1a8/0x278
[<00000000402a28b8>] unmap_single_vma+0x3d8/0x770
[<00000000402a4090>] zap_page_range+0xf0/0x198
[<00000000402ba2a4>] SyS_madvise+0x404/0x8c0
Note that the pte value is 0 except for the accessed bit 0x100. This bit
shouldn't be set without the present bit.
It should be noted that the madvise system call is probably a trigger for many
of the random segmentation faults.
In looking at the kernel code, I found the following problems:
1) The pte_clear define didn't take TLB lock when clearing a pte.
2) We didn't test pte present bit inside lock in exception support.
3) The pte and tlb locks needed to merged in order to ensure consistency
between page table and TLB. This also has the effect of serializing TLB
broadcasts on SMP systems.
The attached change implements the above and a few other tweaks to try
to improve performance. Based on the timing code, TLB purges are very
slow (e.g., ~ 209 cycles per page on rp3440). Thus, I think it
beneficial to test the split_tlb variable to avoid duplicate purges.
Probably, all PA 2.0 machines have combined TLBs.
I dropped using __flush_tlb_range in flush_tlb_mm as I realized all
applications and most threads have a stack size that is too large to
make this useful. I added some comments to this effect.
Since implementing 1 through 3, I haven't had any random segmentation
faults on mx3210 (rp3440) in about one week of building code and running
as a Debian buildd.
Signed-off-by: John David Anglin <dave.anglin@bell.net>
Cc: stable@vger.kernel.org # v3.18+
Signed-off-by: Helge Deller <deller@gmx.de>
On parisc, SHMLBA was defined to 0x00400000 (4MB) to reflect that we need to
take care of our caches for shared mappings. But actually, we can map a file at
any multiple address of PAGE_SIZE, so let us correct that now with a value of
PAGE_SIZE for SHMLBA. Instead we now take care of this cache colouring via the
constant SHM_COLOUR while we map shared pages.
Signed-off-by: Helge Deller <deller@gmx.de>
CC: Jeroen Roovers <jer@gentoo.org>
CC: John David Anglin <dave.anglin@bell.net>
CC: Carlos O'Donell <carlos@systemhalted.org>
Cc: stable@kernel.org [3.13+]
The attached change removes the unused and experimental
CONFIG_PARISC_TMPALIAS code. It doesn't work and I don't believe it will
ever be used.
Signed-off-by: John David Anglin <dave.anglin@bell.net>
Signed-off-by: Helge Deller <deller@gmx.de>
This commit:
f8dae00684d678afa13041ef170cecfd1297ed40: parisc: Ensure full cache coherency for kmap/kunmap
caused negative caching side-effects, e.g. hanging processes with expect and
too many inequivalent alias messages from flush_dcache_page() on Debian 5 systems.
This patch now partly reverts it and has been in production use on our debian buildd
makeservers since a week without any major problems.
Signed-off-by: Helge Deller <deller@gmx.de>
Signed-off-by: John David Anglin <dave.anglin@bell.net>
Cc: stable@vger.kernel.org # v3.9+
Signed-off-by: Helge Deller <deller@gmx.de>
Helge Deller noted a few weeks ago problems with the AIO support on
parisc. This change is the result of numerous iterations on how best to
deal with this problem.
The solution adopted here is to provide full cache coherency in a
uniform manner on all parisc systems. This involves calling
flush_dcache_page() on kmap operations and flush_kernel_dcache_page() on
kunmap operations. As a result, the copy_user_page() and
clear_user_page() functions can be removed and the overall code is
simpler.
The change ensures that both userspace and kernel aliases to a mapped
page are invalidated and flushed. This is necessary for the correct
operation of PA8800 and PA8900 based systems which do not support
inequivalent aliases.
With this change, I have observed no cache related issues on c8000 and
rp3440. It is now possible for example to do kernel builds with "-j64"
on four way systems.
On systems using XFS file systems, the patch recently posted by Mikulas
Patocka to "fix crash using XFS on loopback" is needed to avoid a hang
caused by an uninitialized lock passed to flush_dcache_page() in the
page struct.
Signed-off-by: John David Anglin <dave.anglin@bell.net>
Cc: stable@vger.kernel.org # v3.9+
Signed-off-by: Helge Deller <deller@gmx.de>
This reverts commit 320c90be7b.
Christoph Hellwig <hch@infradead.org> commented:
This one shouldn't go in - Geert sent it a bit prematurely, as Lustre
shouldn't use it just to reimplement core VM functionality (which it
shouldn't use either, but that's a separate story).
Signed-off-by: Helge Deller <deller@gmx.de>
The parisc architecture does not have a pte special bit. As a result,
special mappings are handled with the VM_PFNMAP and VM_MIXEDMAP flags.
VM_MIXEDMAP mappings may or may not have a "struct page" backing. When
pfn_valid() is false, there is no "struct page" backing. Otherwise, they
are treated as normal pages.
The FireGL driver uses the VM_MIXEDMAP without a backing "struct page".
This treatment caused a panic due to a TLB data miss in
update_mmu_cache. This appeared to be in the code generated for
page_address(). We were in fact using a very circular bit of code to
determine the physical address of the PFN in various cache routines.
This wasn't valid when there was no "struct page" backing. The needed
address can in fact be determined simply from the PFN itself without
using the "struct page".
The attached patch updates update_mmu_cache(), flush_cache_mm(),
flush_cache_range() and flush_cache_page() to check pfn_valid() and to
directly compute the PFN physical and virtual addresses.
Signed-off-by: John David Anglin <dave.anglin@bell.net>
Cc: <stable@vger.kernel.org> # 3.10
Signed-off-by: Helge Deller <deller@gmx.de>
I still see the occasional random segv on rp3440. Looking at one of
these (a code 15), it appeared the problem must be with the cache
handling of anonymous pages. Reviewing this, I noticed that the space
register %sr1 might be being clobbered when we flush an anonymous page.
Register %sr1 is used for TLB purges in a couple of places. These
purges are needed on PA8800 and PA8900 processors to ensure cache
consistency of flushed cache lines.
The solution here is simply to move the %sr1 load into the TLB lock
region needed to ensure that one purge executes at a time on SMP
systems. This was already the case for one use. After a few days of
operation, I haven't had a random segv on my rp3440.
Signed-off-by: John David Anglin <dave.anglin@bell.net>
Cc: <stable@vger.kernel.org> # 3.10
Signed-off-by: Helge Deller <deller@gmx.de>
kmap_atomic allows only one argument now, just move the second.
Signed-off-by: Zhao Hongjiang <zhaohongjiang@huawei.com>
Signed-off-by: Helge Deller <deller@gmx.de>
User applications running on SMP kernels have long suffered from instability
and random segmentation faults. This patch improves the situation although
there is more work to be done.
One of the problems is the various routines in pgtable.h that update page table
entries use different locking mechanisms, or no lock at all (set_pte_at). This
change modifies the routines to all use the same lock pa_dbit_lock. This lock
is used for dirty bit updates in the interruption code. The patch also purges
the TLB entries associated with the PTE to ensure that inconsistent values are
not used after the page table entry is updated. The UP and SMP code are now
identical.
The change also includes a minor update to the purge_tlb_entries function in
cache.c to improve its efficiency.
Signed-off-by: John David Anglin <dave.anglin@bell.net>
Cc: Helge Deller <deller@gmx.de>
Signed-off-by: Helge Deller <deller@gmx.de>
It is necessary to disable preemption during cache flushes done through the
TMPALIAS region to ensure that the TLB setup is not clobbered by another flush.
Signed-off-by: John David Anglin <dave.anglin@bell.net>
Signed-off-by: Helge Deller <deller@gmx.de>
CONFIG_PARISC_TMPALIAS enables clear_user_highpage and copy_user_highpage.
These are essentially alternative implementations of clear_user_page and
copy_user_page. They don't have anything to do with x86 high pages, but they
build on the infrastructure to save a few instructions. Read the comment in
clear_user_highpage as it is very important to the implementation. For this
reason, there isn't any gain in using the TMPALIAS/highpage approach.
Signed-off-by: John David Anglin <dave.anglin@bell.net>
Signed-off-by: Helge Deller <deller@gmx.de>
flush_cache_mm, for the non current case also uses flush_dcache_page_asm
and flush_icache_page_asm which are TMPALIAS flushes.
For the non current case, the algorithm used by get_ptep is derived from the
vmalloc_to_page implementation in vmalloc.c. It is essentially a generic page
table lookup. The other alternative was to duplicate the lookup in entry.S.
The break point for switching to a full cache flush is somewhat arbitrary. The
same approach is used in flush_cache_range for non current case. In a GCC
build and check, many small programs are executed and this change provided a
significant performance enhancement, e.g. GCC build time was cut almost in half
on a rp3440 at j4. Previously, we always flushed the entire cache.
Signed-off-by: John David Anglin <dave.anglin@bell.net>
Signed-off-by: Helge Deller <deller@gmx.de>
This is the first patch in a series of 4, with which the page cache flushing of
parisc will gets fixed and enhanced. This even fixes the nasty "minifail" bug
(http://wiki.parisc-linux.org/TestCases?highlight=%28minifail%29) which
prevented parisc to stay an official debian port. Basically the flush in
copy_user_page together with the TLB patch from commit
7139bc1579 is what fixes the minifail bug.
This patch still uses the TMPALIAS approach. The new copy_user_page
implementation calls flush_dcache_page_asm to flush the user dcache page
(crucial for minifail fix) via a kernel TMPALIAS mapping. After that, it just
copies the page using the kernel mapping. It does a final flush if needed.
Generally it is hard to avoid doing some cache flushes using the kernel mapping
(e.g., copy_to_user_page and copy_from_user_page).
This patch depends on a subsequent change to pacache.S implementing
clear_page_asm and copy_page_asm. These are optimized routines to clear and
copy a page. The calls in clear_user_page and copy_user_page could be replaced
by calls to memset and memcpy, respectively. I tested prefetch optimizations
in clear_page_asm and copy_page_asm but didn't see any significant performance
improvement on rp3440. I'm not sure if these are routines are significantly
faster than memset and/or memcpy, but they are there for further performance
evaluation.
Signed-off-by: John David Anglin <dave.anglin@bell.net>
Signed-off-by: Helge Deller <deller@gmx.de>
This patch goes a long way toward fixing the minifail bug, and
it significantly improves the stability of SMP machines such as
the rp3440. When write protecting a page for COW, we need to
purge the existing translation. Otherwise, the COW break
doesn't occur as expected because the TLB may still have a stale entry
which allows writes.
[jejb: fix up checkpatch errors]
Signed-off-by: John David Anglin <dave.anglin@bell.net>
Cc: stable@vger.kernel.org
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
Implement an interval tree as a replacement for the VMA prio_tree. The
algorithms are similar to lib/interval_tree.c; however that code can't be
directly reused as the interval endpoints are not explicitly stored in the
VMA. So instead, the common algorithm is moved into a template and the
details (node type, how to get interval endpoints from the node, etc) are
filled in using the C preprocessor.
Once the interval tree functions are available, using them as a
replacement to the VMA prio tree is a relatively simple, mechanical job.
Signed-off-by: Michel Lespinasse <walken@google.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Hillf Danton <dhillf@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Woodhouse <dwmw2@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
According to Appendix F, the TLB is the primary arbiter of speculation.
Thus, if a page has a TLB entry, it may be speculatively read into the
cache. On linux, this can cause us incoherencies because if we're about
to do a disk read, we call get_user_pages() to do the flush/invalidate
in user space, but we still potentially have the user TLB entries, and
the cache could speculate the lines back into userspace (thus causing
stale data to be used). This is fixed by purging the TLB entries before
we flush through the tmpalias space. Now, the only way the line could
be re-speculated is if the user actually tries to touch it (which is not
allowed).
Signed-off-by: James Bottomley <James.Bottomley@suse.de>
The kernel has an 8M tmpailas space (originally designed for copying
and clearing pages but now only used for clearing). The idea is
to place zeros into the cache above a physical page rather than into
the physical page and flush the cache, because often the zeros end up
being replaced quickly anyway.
We can also use the tmpalias space for flushing a page. The difference
here is that we have to do tmpalias processing in the non access data and
instruction traps. The principle is the same: as long as we know the physical
address and have a virtual address congruent to the real one, the flush will
be effective.
In order to use the tmpalias space, the icache miss path has to be enhanced to
check for the alias region to make the fic instruction effective.
Signed-off-by: James Bottomley <James.Bottomley@suse.de>
On VIVT ARM, when we have multiple shared mappings of the same file
in the same MM, we need to ensure that we have coherency across all
copies. We do this via make_coherent() by making the pages
uncacheable.
This used to work fine, until we allowed highmem with highpte - we
now have a page table which is mapped as required, and is not available
for modification via update_mmu_cache().
Ralf Beache suggested getting rid of the PTE value passed to
update_mmu_cache():
On MIPS update_mmu_cache() calls __update_tlb() which walks pagetables
to construct a pointer to the pte again. Passing a pte_t * is much
more elegant. Maybe we might even replace the pte argument with the
pte_t?
Ben Herrenschmidt would also like the pte pointer for PowerPC:
Passing the ptep in there is exactly what I want. I want that
-instead- of the PTE value, because I have issue on some ppc cases,
for I$/D$ coherency, where set_pte_at() may decide to mask out the
_PAGE_EXEC.
So, pass in the mapped page table pointer into update_mmu_cache(), and
remove the PTE value, updating all implementations and call sites to
suit.
Includes a fix from Stephen Rothwell:
sparc: fix fallout from update_mmu_cache API change
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Acked-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
The TLB flushing functions on hppa, which causes PxTLB broadcasts on the system
bus, needs to be protected by irq-safe spinlocks to avoid irq handlers to deadlock
the kernel. The deadlocks only happened during I/O intensive loads and triggered
pretty seldom, which is why this bug went so long unnoticed.
Signed-off-by: Helge Deller <deller@gmx.de>
[edited to use spin_lock_irqsave on UP as well since we'd been locking there
all this time anyway, --kyle]
Signed-off-by: Kyle McMartin <kyle@mcmartin.ca>
- convert a few "if (xx) BUG();" to BUG_ON(xx)
- remove a few printk()s, as we get a backtrace with BUG_ON() anyway
Signed-off-by: Helge Deller <deller@gmx.de>
Signed-off-by: Kyle McMartin <kyle@mcmartin.ca>
It's not even passed on to smp_call_function() anymore, since that
was removed. So kill it.
Acked-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Reviewed-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
First thing mm.h does is including sched.h solely for can_do_mlock() inline
function which has "current" dereference inside. By dealing with can_do_mlock()
mm.h can be detached from sched.h which is good. See below, why.
This patch
a) removes unconditional inclusion of sched.h from mm.h
b) makes can_do_mlock() normal function in mm/mlock.c
c) exports can_do_mlock() to not break compilation
d) adds sched.h inclusions back to files that were getting it indirectly.
e) adds less bloated headers to some files (asm/signal.h, jiffies.h) that were
getting them indirectly
Net result is:
a) mm.h users would get less code to open, read, preprocess, parse, ... if
they don't need sched.h
b) sched.h stops being dependency for significant number of files:
on x86_64 allmodconfig touching sched.h results in recompile of 4083 files,
after patch it's only 3744 (-8.3%).
Cross-compile tested on
all arm defconfigs, all mips defconfigs, all powerpc defconfigs,
alpha alpha-up
arm
i386 i386-up i386-defconfig i386-allnoconfig
ia64 ia64-up
m68k
mips
parisc parisc-up
powerpc powerpc-up
s390 s390-up
sparc sparc-up
sparc64 sparc64-up
um-x86_64
x86_64 x86_64-up x86_64-defconfig x86_64-allnoconfig
as well as my two usual configs.
Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Mikulas Patocka