Skip to content

L
linux

xiphi1978 / linux

通知 2
Star 0
Fork 0
L
linux

体验新版 GitCode，发现更多精彩内容 >>

切换分支/标签
  1. 01 2月, 2016 1 次提交
  3. 21 1月, 2016 1 次提交
  5. 16 1月, 2016 12 次提交
  7. 15 1月, 2016 2 次提交
    • M
      mm: allow GFP_{FS,IO} for page_cache_read page cache allocation · c20cd45e
      由 Michal Hocko 提交于 
      page_cache_read has been historically using page_cache_alloc_cold to
allocate a new page.  This means that mapping_gfp_mask is used as the
base for the gfp_mask.  Many filesystems are setting this mask to
GFP_NOFS to prevent from fs recursion issues.  page_cache_read is called
from the vm_operations_struct::fault() context during the page fault.
This context doesn't need the reclaim protection normally.

ceph and ocfs2 which call filemap_fault from their fault handlers seem
to be OK because they are not taking any fs lock before invoking generic
implementation.  xfs which takes XFS_MMAPLOCK_SHARED is safe from the
reclaim recursion POV because this lock serializes truncate and punch
hole with the page faults and it doesn't get involved in the reclaim.

There is simply no reason to deliberately use a weaker allocation
context when a __GFP_FS | __GFP_IO can be used.  The GFP_NOFS protection
might be even harmful.  There is a push to fail GFP_NOFS allocations
rather than loop within allocator indefinitely with a very limited
reclaim ability.  Once we start failing those requests the OOM killer
might be triggered prematurely because the page cache allocation failure
is propagated up the page fault path and end up in
pagefault_out_of_memory.

We cannot play with mapping_gfp_mask directly because that would be racy
wrt.  parallel page faults and it might interfere with other users who
really rely on NOFS semantic from the stored gfp_mask.  The mask is also
inode proper so it would even be a layering violation.  What we can do
instead is to push the gfp_mask into struct vm_fault and allow fs layer
to overwrite it should the callback need to be called with a different
allocation context.

Initialize the default to (mapping_gfp_mask | __GFP_FS | __GFP_IO)
because this should be safe from the page fault path normally.  Why do
we care about mapping_gfp_mask at all then? Because this doesn't hold
only reclaim protection flags but it also might contain zone and
movability restrictions (GFP_DMA32, __GFP_MOVABLE and others) so we have
to respect those.
Signed-off-by: NMichal Hocko <mhocko@suse.com>
Reported-by: NTetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Acked-by: NJan Kara <jack@suse.com>
Acked-by: NVlastimil Babka <vbabka@suse.cz>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Mark Fasheh <mfasheh@suse.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>
      c20cd45e
    • J
      mm, shmem: add internal shmem resident memory accounting · eca56ff9
      由 Jerome Marchand 提交于 
      Currently looking at /proc/<pid>/status or statm, there is no way to
distinguish shmem pages from pages mapped to a regular file (shmem pages
are mapped to /dev/zero), even though their implication in actual memory
use is quite different.

The internal accounting currently counts shmem pages together with
regular files.  As a preparation to extend the userspace interfaces,
this patch adds MM_SHMEMPAGES counter to mm_rss_stat to account for
shmem pages separately from MM_FILEPAGES.  The next patch will expose it
to userspace - this patch doesn't change the exported values yet, by
adding up MM_SHMEMPAGES to MM_FILEPAGES at places where MM_FILEPAGES was
used before.  The only user-visible change after this patch is the OOM
killer message that separates the reported "shmem-rss" from "file-rss".

[vbabka@suse.cz: forward-porting, tweak changelog]
Signed-off-by: NJerome Marchand <jmarchan@redhat.com>
Signed-off-by: NVlastimil Babka <vbabka@suse.cz>
Acked-by: NKonstantin Khlebnikov <khlebnikov@yandex-team.ru>
Acked-by: NMichal Hocko <mhocko@suse.com>
Acked-by: NHugh Dickins <hughd@google.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>
      eca56ff9
  9. 19 11月, 2015 1 次提交
    • Y
      mm, dax: fix DAX deadlocks (COW fault) · 0df9d41a
      由 Yigal Korman 提交于 
      DAX handling of COW faults has wrong locking sequence:
	dax_fault does i_mmap_lock_read
	do_cow_fault does i_mmap_unlock_write

Ross's commit[1] missed a fix[2] that Kirill added to Matthew's
commit[3].

Original COW locking logic was introduced by Matthew here[4].

This should be applied to v4.3 as well.

[1] 0f90cc66 mm, dax: fix DAX deadlocks
[2] 52a2b53f mm, dax: use i_mmap_unlock_write() in do_cow_fault()
[3] 84317297 dax: fix race between simultaneous faults
[4] 2e4cdab0 mm: allow page fault handlers to perform the COW

Cc: <stable@vger.kernel.org>
Cc: Boaz Harrosh <boaz@plexistor.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: Jan Kara <jack@suse.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Matthew Wilcox <matthew.r.wilcox@intel.com>
Acked-by: NRoss Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: NYigal Korman <yigal@plexistor.com>
Signed-off-by: NDan Williams <dan.j.williams@intel.com>
      0df9d41a
  11. 17 10月, 2015 1 次提交
  13. 11 9月, 2015 1 次提交
  15. 09 9月, 2015 5 次提交
  17. 05 9月, 2015 2 次提交
  19. 10 7月, 2015 1 次提交
  21. 25 6月, 2015 1 次提交
    • M
      mm, memcg: Try charging a page before setting page up to date · eb3c24f3
      由 Mel Gorman 提交于 
      Historically memcg overhead was high even if memcg was unused.  This has
improved a lot but it still showed up in a profile summary as being a
problem.

/usr/src/linux-4.0-vanilla/mm/memcontrol.c                           6.6441   395842
  mem_cgroup_try_charge                                                        2.950%   175781
  __mem_cgroup_count_vm_event                                                  1.431%    85239
  mem_cgroup_page_lruvec                                                       0.456%    27156
  mem_cgroup_commit_charge                                                     0.392%    23342
  uncharge_list                                                                0.323%    19256
  mem_cgroup_update_lru_size                                                   0.278%    16538
  memcg_check_events                                                           0.216%    12858
  mem_cgroup_charge_statistics.isra.22                                         0.188%    11172
  try_charge                                                                   0.150%     8928
  commit_charge                                                                0.141%     8388
  get_mem_cgroup_from_mm                                                       0.121%     7184

That is showing that 6.64% of system CPU cycles were in memcontrol.c and
dominated by mem_cgroup_try_charge.  The annotation shows that the bulk
of the cost was checking PageSwapCache which is expected to be cache hot
but is very expensive.  The problem appears to be that __SetPageUptodate
is called just before the check which is a write barrier.  It is
required to make sure struct page and page data is written before the
PTE is updated and the data visible to userspace.  memcg charging does
not require or need the barrier but gets unfairly hit with the cost so
this patch attempts the charging before the barrier.  Aside from the
accidental cost to memcg there is the added benefit that the barrier is
avoided if the page cannot be charged.  When applied the relevant
profile summary is as follows.

/usr/src/linux-4.0-chargefirst-v2r1/mm/memcontrol.c                  3.7907   223277
  __mem_cgroup_count_vm_event                                                  1.143%    67312
  mem_cgroup_page_lruvec                                                       0.465%    27403
  mem_cgroup_commit_charge                                                     0.381%    22452
  uncharge_list                                                                0.332%    19543
  mem_cgroup_update_lru_size                                                   0.284%    16704
  get_mem_cgroup_from_mm                                                       0.271%    15952
  mem_cgroup_try_charge                                                        0.237%    13982
  memcg_check_events                                                           0.222%    13058
  mem_cgroup_charge_statistics.isra.22                                         0.185%    10920
  commit_charge                                                                0.140%     8235
  try_charge                                                                   0.131%     7716

That brings the overhead down to 3.79% and leaves the memcg fault
accounting to the root cgroup but it's an improvement.  The difference
in headline performance of the page fault microbench is marginal as
memcg is such a small component of it.

pft faults
                                       4.0.0                  4.0.0
                                     vanilla            chargefirst
Hmean    faults/cpu-1 1443258.1051 (  0.00%) 1509075.7561 (  4.56%)
Hmean    faults/cpu-3 1340385.9270 (  0.00%) 1339160.7113 ( -0.09%)
Hmean    faults/cpu-5  875599.0222 (  0.00%)  874174.1255 ( -0.16%)
Hmean    faults/cpu-7  601146.6726 (  0.00%)  601370.9977 (  0.04%)
Hmean    faults/cpu-8  510728.2754 (  0.00%)  510598.8214 ( -0.03%)
Hmean    faults/sec-1 1432084.7845 (  0.00%) 1497935.5274 (  4.60%)
Hmean    faults/sec-3 3943818.1437 (  0.00%) 3941920.1520 ( -0.05%)
Hmean    faults/sec-5 3877573.5867 (  0.00%) 3869385.7553 ( -0.21%)
Hmean    faults/sec-7 3991832.0418 (  0.00%) 3992181.4189 (  0.01%)
Hmean    faults/sec-8 3987189.8167 (  0.00%) 3986452.2204 ( -0.02%)

It's only visible at single threaded. The overhead is there for higher
threads but other factors dominate.
Signed-off-by: NMel Gorman <mgorman@suse.de>
Acked-by: NMichal Hocko <mhocko@suse.cz>
Acked-by: NJohannes Weiner <hannes@cmpxchg.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>
      eb3c24f3
  23. 24 6月, 2015 1 次提交
  25. 19 5月, 2015 1 次提交
    • D
      sched/preempt, mm/fault: Trigger might_sleep() in might_fault() with disabled pagefaults · 9ec23531
      由 David Hildenbrand 提交于 
      Commit 662bbcb2 ("mm, sched: Allow uaccess in atomic with
pagefault_disable()") removed might_sleep() checks for all user access
code (that uses might_fault()).

The reason was to disable wrong "sleep in atomic" warnings in the
following scenario:

    pagefault_disable()
    rc = copy_to_user(...)
    pagefault_enable()

Which is valid, as pagefault_disable() increments the preempt counter
and therefore disables the pagefault handler. copy_to_user() will not
sleep and return an error code if a page is not available.

However, as all might_sleep() checks are removed,
CONFIG_DEBUG_ATOMIC_SLEEP would no longer detect the following scenario:

    spin_lock(&lock);
    rc = copy_to_user(...)
    spin_unlock(&lock)

If the kernel is compiled with preemption turned on, preempt_disable()
will make in_atomic() detect disabled preemption. The fault handler would
correctly never sleep on user access.
However, with preemption turned off, preempt_disable() is usually a NOP
(with !CONFIG_PREEMPT_COUNT), therefore in_atomic() will not be able to
detect disabled preemption nor disabled pagefaults. The fault handler
could sleep.
We really want to enable CONFIG_DEBUG_ATOMIC_SLEEP checks for user access
functions again, otherwise we can end up with horrible deadlocks.

Root of all evil is that pagefault_disable() acts almost as
preempt_disable(), depending on preemption being turned on/off.

As we now have pagefault_disabled(), we can use it to distinguish
whether user acces functions might sleep.

Convert might_fault() into a makro that calls __might_fault(), to
allow proper file + line messages in case of a might_sleep() warning.
Reviewed-and-tested-by: NThomas Gleixner <tglx@linutronix.de>
Signed-off-by: NDavid Hildenbrand <dahi@linux.vnet.ibm.com>
Signed-off-by: NPeter Zijlstra (Intel) <peterz@infradead.org>
Cc: David.Laight@ACULAB.COM
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: airlied@linux.ie
Cc: akpm@linux-foundation.org
Cc: benh@kernel.crashing.org
Cc: bigeasy@linutronix.de
Cc: borntraeger@de.ibm.com
Cc: daniel.vetter@intel.com
Cc: heiko.carstens@de.ibm.com
Cc: herbert@gondor.apana.org.au
Cc: hocko@suse.cz
Cc: hughd@google.com
Cc: mst@redhat.com
Cc: paulus@samba.org
Cc: ralf@linux-mips.org
Cc: schwidefsky@de.ibm.com
Cc: yang.shi@windriver.com
Link: http://lkml.kernel.org/r/1431359540-32227-3-git-send-email-dahi@linux.vnet.ibm.comSigned-off-by: NIngo Molnar <mingo@kernel.org>
      9ec23531
  27. 16 4月, 2015 3 次提交
    • B
      mm: new pfn_mkwrite same as page_mkwrite for VM_PFNMAP · dd906184
      由 Boaz Harrosh 提交于 
      This will allow FS that uses VM_PFNMAP | VM_MIXEDMAP (no page structs) to
get notified when access is a write to a read-only PFN.

This can happen if we mmap() a file then first mmap-read from it to
page-in a read-only PFN, than we mmap-write to the same page.

We need this functionality to fix a DAX bug, where in the scenario above
we fail to set ctime/mtime though we modified the file.  An xfstest is
attached to this patchset that shows the failure and the fix.  (A DAX
patch will follow)

This functionality is extra important for us, because upon dirtying of a
pmem page we also want to RDMA the page to a remote cluster node.

We define a new pfn_mkwrite and do not reuse page_mkwrite because
  1 - The name ;-)
  2 - But mainly because it would take a very long and tedious
      audit of all page_mkwrite functions of VM_MIXEDMAP/VM_PFNMAP
      users. To make sure they do not now CRASH. For example current
      DAX code (which this is for) would crash.
      If we would want to reuse page_mkwrite, We will need to first
      patch all users, so to not-crash-on-no-page. Then enable this
      patch. But even if I did that I would not sleep so well at night.
      Adding a new vector is the safest thing to do, and is not that
      expensive. an extra pointer at a static function vector per driver.
      Also the new vector is better for performance, because else we
      Will call all current Kernel vectors, so to:
        check-ha-no-page-do-nothing and return.

No need to call it from do_shared_fault because do_wp_page is called to
change pte permissions anyway.
Signed-off-by: NYigal Korman <yigal@plexistor.com>
Signed-off-by: NBoaz Harrosh <boaz@plexistor.com>
Acked-by: NKirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Matthew Wilcox <matthew.r.wilcox@intel.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Dave Chinner <david@fromorbit.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>
      dd906184
    • K
      mm/memory: also print a_ops->readpage in print_bad_pte() · 2682582a
      由 Konstantin Khlebnikov 提交于 
      A lot of filesystems use generic_file_mmap() and filemap_fault(),
f_op->mmap and vm_ops->fault aren't enough to identify filesystem.

This prints file name, vm_ops->fault, f_op->mmap and a_ops->readpage
(which is almost always implemented and filesystem-specific).

Example:

[   23.676410] BUG: Bad page map in process sh  pte:1b7e6025 pmd:19bbd067
[   23.676887] page:ffffea00006df980 count:4 mapcount:1 mapping:ffff8800196426c0 index:0x97
[   23.677481] flags: 0x10000000000000c(referenced|uptodate)
[   23.677896] page dumped because: bad pte
[   23.678205] addr:00007f52fcb17000 vm_flags:00000075 anon_vma:          (null) mapping:ffff8800196426c0 index:97
[   23.678922] file:libc-2.19.so fault:filemap_fault mmap:generic_file_readonly_mmap readpage:v9fs_vfs_readpage

[akpm@linux-foundation.org: use pr_alert, per Kirill]
Signed-off-by: NKonstantin Khlebnikov <khlebnikov@yandex-team.ru>
Cc: Sasha Levin <sasha.levin@oracle.com>
Acked-by: NKirill A. Shutemov <kirill@shutemov.name>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>
      2682582a
    • J
      mm: remove rest of ACCESS_ONCE() usages · 4db0c3c2
      由 Jason Low 提交于 
      We converted some of the usages of ACCESS_ONCE to READ_ONCE in the mm/
tree since it doesn't work reliably on non-scalar types.

This patch removes the rest of the usages of ACCESS_ONCE, and use the new
READ_ONCE API for the read accesses.  This makes things cleaner, instead
of using separate/multiple sets of APIs.
Signed-off-by: NJason Low <jason.low2@hp.com>
Acked-by: NMichal Hocko <mhocko@suse.cz>
Acked-by: NDavidlohr Bueso <dave@stgolabs.net>
Acked-by: NRik van Riel <riel@redhat.com>
Reviewed-by: NChristian Borntraeger <borntraeger@de.ibm.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>
      4db0c3c2
  29. 15 4月, 2015 4 次提交
  31. 26 3月, 2015 3 次提交
    • M
      mm: numa: slow PTE scan rate if migration failures occur · 074c2381
      由 Mel Gorman 提交于 
      Dave Chinner reported the following on https://lkml.org/lkml/2015/3/1/226

  Across the board the 4.0-rc1 numbers are much slower, and the degradation
  is far worse when using the large memory footprint configs. Perf points
  straight at the cause - this is from 4.0-rc1 on the "-o bhash=101073" config:

   -   56.07%    56.07%  [kernel]            [k] default_send_IPI_mask_sequence_phys
      - default_send_IPI_mask_sequence_phys
         - 99.99% physflat_send_IPI_mask
            - 99.37% native_send_call_func_ipi
                 smp_call_function_many
               - native_flush_tlb_others
                  - 99.85% flush_tlb_page
                       ptep_clear_flush
                       try_to_unmap_one
                       rmap_walk
                       try_to_unmap
                       migrate_pages
                       migrate_misplaced_page
                     - handle_mm_fault
                        - 99.73% __do_page_fault
                             trace_do_page_fault
                             do_async_page_fault
                           + async_page_fault
              0.63% native_send_call_func_single_ipi
                 generic_exec_single
                 smp_call_function_single

This is showing excessive migration activity even though excessive
migrations are meant to get throttled.  Normally, the scan rate is tuned
on a per-task basis depending on the locality of faults.  However, if
migrations fail for any reason then the PTE scanner may scan faster if
the faults continue to be remote.  This means there is higher system CPU
overhead and fault trapping at exactly the time we know that migrations
cannot happen.  This patch tracks when migration failures occur and
slows the PTE scanner.
Signed-off-by: NMel Gorman <mgorman@suse.de>
Reported-by: NDave Chinner <david@fromorbit.com>
Tested-by: NDave Chinner <david@fromorbit.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Aneesh Kumar <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>
      074c2381
    • M
      mm: numa: preserve PTE write permissions across a NUMA hinting fault · b191f9b1
      由 Mel Gorman 提交于 
      Protecting a PTE to trap a NUMA hinting fault clears the writable bit
and further faults are needed after trapping a NUMA hinting fault to set
the writable bit again.  This patch preserves the writable bit when
trapping NUMA hinting faults.  The impact is obvious from the number of
minor faults trapped during the basis balancing benchmark and the system
CPU usage;

  autonumabench
                                             4.0.0-rc4             4.0.0-rc4
                                              baseline              preserve
  Time System-NUMA01                  107.13 (  0.00%)      103.13 (  3.73%)
  Time System-NUMA01_THEADLOCAL       131.87 (  0.00%)       83.30 ( 36.83%)
  Time System-NUMA02                    8.95 (  0.00%)       10.72 (-19.78%)
  Time System-NUMA02_SMT                4.57 (  0.00%)        3.99 ( 12.69%)
  Time Elapsed-NUMA01                 515.78 (  0.00%)      517.26 ( -0.29%)
  Time Elapsed-NUMA01_THEADLOCAL      384.10 (  0.00%)      384.31 ( -0.05%)
  Time Elapsed-NUMA02                  48.86 (  0.00%)       48.78 (  0.16%)
  Time Elapsed-NUMA02_SMT              47.98 (  0.00%)       48.12 ( -0.29%)

               4.0.0-rc4   4.0.0-rc4
                baseline    preserve
  User          44383.95    43971.89
  System          252.61      201.24
  Elapsed         998.68     1000.94

  Minor Faults   2597249     1981230
  Major Faults       365         364

There is a similar drop in system CPU usage using Dave Chinner's xfsrepair
workload

                                      4.0.0-rc4             4.0.0-rc4
                                       baseline              preserve
  Amean    real-xfsrepair      454.14 (  0.00%)      442.36 (  2.60%)
  Amean    syst-xfsrepair      277.20 (  0.00%)      204.68 ( 26.16%)

The patch looks hacky but the alternatives looked worse.  The tidest was
to rewalk the page tables after a hinting fault but it was more complex
than this approach and the performance was worse.  It's not generally
safe to just mark the page writable during the fault if it's a write
fault as it may have been read-only for COW so that approach was
discarded.
Signed-off-by: NMel Gorman <mgorman@suse.de>
Reported-by: NDave Chinner <david@fromorbit.com>
Tested-by: NDave Chinner <david@fromorbit.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Aneesh Kumar <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>
      b191f9b1
    • M
      mm: numa: group related processes based on VMA flags instead of page table flags · bea66fbd
      由 Mel Gorman 提交于 
      These are three follow-on patches based on the xfsrepair workload Dave
Chinner reported was problematic in 4.0-rc1 due to changes in page table
management -- https://lkml.org/lkml/2015/3/1/226.

Much of the problem was reduced by commit 53da3bc2 ("mm: fix up numa
read-only thread grouping logic") and commit ba68bc01 ("mm: thp:
Return the correct value for change_huge_pmd").  It was known that the
performance in 3.19 was still better even if is far less safe.  This
series aims to restore the performance without compromising on safety.

For the test of this mail, I'm comparing 3.19 against 4.0-rc4 and the
three patches applied on top

  autonumabench
                                                3.19.0             4.0.0-rc4             4.0.0-rc4             4.0.0-rc4             4.0.0-rc4
                                               vanilla               vanilla          vmwrite-v5r8         preserve-v5r8         slowscan-v5r8
  Time System-NUMA01                  124.00 (  0.00%)      161.86 (-30.53%)      107.13 ( 13.60%)      103.13 ( 16.83%)      145.01 (-16.94%)
  Time System-NUMA01_THEADLOCAL       115.54 (  0.00%)      107.64 (  6.84%)      131.87 (-14.13%)       83.30 ( 27.90%)       92.35 ( 20.07%)
  Time System-NUMA02                    9.35 (  0.00%)       10.44 (-11.66%)        8.95 (  4.28%)       10.72 (-14.65%)        8.16 ( 12.73%)
  Time System-NUMA02_SMT                3.87 (  0.00%)        4.63 (-19.64%)        4.57 (-18.09%)        3.99 ( -3.10%)        3.36 ( 13.18%)
  Time Elapsed-NUMA01                 570.06 (  0.00%)      567.82 (  0.39%)      515.78 (  9.52%)      517.26 (  9.26%)      543.80 (  4.61%)
  Time Elapsed-NUMA01_THEADLOCAL      393.69 (  0.00%)      384.83 (  2.25%)      384.10 (  2.44%)      384.31 (  2.38%)      380.73 (  3.29%)
  Time Elapsed-NUMA02                  49.09 (  0.00%)       49.33 ( -0.49%)       48.86 (  0.47%)       48.78 (  0.63%)       50.94 ( -3.77%)
  Time Elapsed-NUMA02_SMT              47.51 (  0.00%)       47.15 (  0.76%)       47.98 ( -0.99%)       48.12 ( -1.28%)       49.56 ( -4.31%)

                3.19.0   4.0.0-rc4   4.0.0-rc4   4.0.0-rc4   4.0.0-rc4
               vanilla     vanillavmwrite-v5r8preserve-v5r8slowscan-v5r8
  User        46334.60    46391.94    44383.95    43971.89    44372.12
  System        252.84      284.66      252.61      201.24      249.00
  Elapsed      1062.14     1050.96      998.68     1000.94     1026.78

Overall the system CPU usage is comparable and the test is naturally a
bit variable.  The slowing of the scanner hurts numa01 but on this
machine it is an adverse workload and patches that dramatically help it
often hurt absolutely everything else.

Due to patch 2, the fault activity is interesting

                                  3.19.0   4.0.0-rc4   4.0.0-rc4   4.0.0-rc4   4.0.0-rc4
                                 vanilla     vanillavmwrite-v5r8preserve-v5r8slowscan-v5r8
  Minor Faults                   2097811     2656646     2597249     1981230     1636841
  Major Faults                       362         450         365         364         365

Note the impact preserving the write bit across protection updates and
fault reduces faults.

  NUMA alloc hit                 1229008     1217015     1191660     1178322     1199681
  NUMA alloc miss                      0           0           0           0           0
  NUMA interleave hit                  0           0           0           0           0
  NUMA alloc local               1228514     1216317     1190871     1177448     1199021
  NUMA base PTE updates        245706197   240041607   238195516   244704842   115012800
  NUMA huge PMD updates           479530      468448      464868      477573      224487
  NUMA page range updates      491225557   479886983   476207932   489222218   229950144
  NUMA hint faults                659753      656503      641678      656926      294842
  NUMA hint local faults          381604      373963      360478      337585      186249
  NUMA hint local percent             57          56          56          51          63
  NUMA pages migrated            5412140     6374899     6266530     5277468     5755096
  AutoNUMA cost                    5121%       5083%       4994%       5097%       2388%

Here the impact of slowing the PTE scanner on migratrion failures is
obvious as "NUMA base PTE updates" and "NUMA huge PMD updates" are
massively reduced even though the headline performance is very similar.

As xfsrepair was the reported workload here is the impact of the series
on it.

  xfsrepair
                                         3.19.0             4.0.0-rc4             4.0.0-rc4             4.0.0-rc4             4.0.0-rc4
                                        vanilla               vanilla          vmwrite-v5r8         preserve-v5r8         slowscan-v5r8
  Min      real-fsmark        1183.29 (  0.00%)     1165.73 (  1.48%)     1152.78 (  2.58%)     1153.64 (  2.51%)     1177.62 (  0.48%)
  Min      syst-fsmark        4107.85 (  0.00%)     4027.75 (  1.95%)     3986.74 (  2.95%)     3979.16 (  3.13%)     4048.76 (  1.44%)
  Min      real-xfsrepair      441.51 (  0.00%)      463.96 ( -5.08%)      449.50 ( -1.81%)      440.08 (  0.32%)      439.87 (  0.37%)
  Min      syst-xfsrepair      195.76 (  0.00%)      278.47 (-42.25%)      262.34 (-34.01%)      203.70 ( -4.06%)      143.64 ( 26.62%)
  Amean    real-fsmark        1188.30 (  0.00%)     1177.34 (  0.92%)     1157.97 (  2.55%)     1158.21 (  2.53%)     1182.22 (  0.51%)
  Amean    syst-fsmark        4111.37 (  0.00%)     4055.70 (  1.35%)     3987.19 (  3.02%)     3998.72 (  2.74%)     4061.69 (  1.21%)
  Amean    real-xfsrepair      450.88 (  0.00%)      468.32 ( -3.87%)      454.14 ( -0.72%)      442.36 (  1.89%)      440.59 (  2.28%)
  Amean    syst-xfsrepair      199.66 (  0.00%)      290.60 (-45.55%)      277.20 (-38.84%)      204.68 ( -2.51%)      150.55 ( 24.60%)
  Stddev   real-fsmark           4.12 (  0.00%)       10.82 (-162.29%)       4.14 ( -0.28%)        5.98 (-45.05%)        4.60 (-11.53%)
  Stddev   syst-fsmark           2.63 (  0.00%)       20.32 (-671.82%)       0.37 ( 85.89%)       16.47 (-525.59%)      15.05 (-471.79%)
  Stddev   real-xfsrepair        6.87 (  0.00%)        4.55 ( 33.75%)        3.46 ( 49.58%)        1.78 ( 74.12%)        0.52 ( 92.50%)
  Stddev   syst-xfsrepair        3.02 (  0.00%)       10.30 (-241.37%)      13.17 (-336.37%)       0.71 ( 76.63%)        5.00 (-65.61%)
  CoeffVar real-fsmark           0.35 (  0.00%)        0.92 (-164.73%)       0.36 ( -2.91%)        0.52 (-48.82%)        0.39 (-12.10%)
  CoeffVar syst-fsmark           0.06 (  0.00%)        0.50 (-682.41%)       0.01 ( 85.45%)        0.41 (-543.22%)       0.37 (-478.78%)
  CoeffVar real-xfsrepair        1.52 (  0.00%)        0.97 ( 36.21%)        0.76 ( 49.94%)        0.40 ( 73.62%)        0.12 ( 92.33%)
  CoeffVar syst-xfsrepair        1.51 (  0.00%)        3.54 (-134.54%)       4.75 (-214.31%)       0.34 ( 77.20%)        3.32 (-119.63%)
  Max      real-fsmark        1193.39 (  0.00%)     1191.77 (  0.14%)     1162.90 (  2.55%)     1166.66 (  2.24%)     1188.50 (  0.41%)
  Max      syst-fsmark        4114.18 (  0.00%)     4075.45 (  0.94%)     3987.65 (  3.08%)     4019.45 (  2.30%)     4082.80 (  0.76%)
  Max      real-xfsrepair      457.80 (  0.00%)      474.60 ( -3.67%)      457.82 ( -0.00%)      444.42 (  2.92%)      441.03 (  3.66%)
  Max      syst-xfsrepair      203.11 (  0.00%)      303.65 (-49.50%)      294.35 (-44.92%)      205.33 ( -1.09%)      155.28 ( 23.55%)

The really relevant lines as syst-xfsrepair which is the system CPU
usage when running xfsrepair.  Note that on my machine the overhead was
45% higher on 4.0-rc4 which may be part of what Dave is seeing.  Once we
preserve the write bit across faults, it's only 2.51% higher on average.
With the full series applied, system CPU usage is 24.6% lower on
average.

Again, the impact of preserving the write bit on minor faults is obvious
and the impact of slowing scanning after migration failures is obvious
on the PTE updates.  Note also that the number of pages migrated is much
reduced even though the headline performance is comparable.

                                  3.19.0   4.0.0-rc4   4.0.0-rc4   4.0.0-rc4   4.0.0-rc4
                                 vanilla     vanillavmwrite-v5r8preserve-v5r8slowscan-v5r8
  Minor Faults                 153466827   254507978   249163829   153501373   105737890
  Major Faults                       610         702         690         649         724
  NUMA base PTE updates        217735049   210756527   217729596   216937111   144344993
  NUMA huge PMD updates           129294       85044      106921      127246       79887
  NUMA pages migrated           21938995    29705270    28594162    22687324    16258075

                        3.19.0   4.0.0-rc4   4.0.0-rc4   4.0.0-rc4   4.0.0-rc4
                       vanilla     vanillavmwrite-v5r8preserve-v5r8slowscan-v5r8
  Mean sdb-avgqusz       13.47        2.54        2.55        2.47        2.49
  Mean sdb-avgrqsz      202.32      140.22      139.50      139.02      138.12
  Mean sdb-await         25.92        5.09        5.33        5.02        5.22
  Mean sdb-r_await        4.71        0.19        0.83        0.51        0.11
  Mean sdb-w_await      104.13        5.21        5.38        5.05        5.32
  Mean sdb-svctm          0.59        0.13        0.14        0.13        0.14
  Mean sdb-rrqm           0.16        0.00        0.00        0.00        0.00
  Mean sdb-wrqm           3.59     1799.43     1826.84     1812.21     1785.67
  Max  sdb-avgqusz      111.06       12.13       14.05       11.66       15.60
  Max  sdb-avgrqsz      255.60      190.34      190.01      187.33      191.78
  Max  sdb-await        168.24       39.28       49.22       44.64       65.62
  Max  sdb-r_await      660.00       52.00      280.00       76.00       12.00
  Max  sdb-w_await     7804.00       39.28       49.22       44.64       65.62
  Max  sdb-svctm          4.00        2.82        2.86        1.98        2.84
  Max  sdb-rrqm           8.30        0.00        0.00        0.00        0.00
  Max  sdb-wrqm          34.20     5372.80     5278.60     5386.60     5546.15

FWIW, I also checked SPECjbb in different configurations but it's
similar observations -- minor faults lower, PTE update activity lower
and performance is roughly comparable against 3.19.

This patch (of 3):

Threads that share writable data within pages are grouped together as
related tasks.  This decision is based on whether the PTE is marked
dirty which is subject to timing races between the PTE scanner update
and when the application writes the page.  If the page is file-backed,
then background flushes and sync also affect placement.  This is
unpredictable behaviour which is impossible to reason about so this
patch makes grouping decisions based on the VMA flags.
Signed-off-by: NMel Gorman <mgorman@suse.de>
Reported-by: NDave Chinner <david@fromorbit.com>
Tested-by: NDave Chinner <david@fromorbit.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Aneesh Kumar <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>
      bea66fbd