Date:   Wed, 11 Sep 2019 05:29:56 +0000
From:   Naoya Horiguchi <n-horiguchi@...jp.nec.com>
To:     Oscar Salvador <osalvador@...e.de>
CC:     "mhocko@...nel.org" <mhocko@...nel.org>,
        "mike.kravetz@...cle.com" <mike.kravetz@...cle.com>,
        "linux-mm@...ck.org" <linux-mm@...ck.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH 00/10] Hwpoison soft-offline rework

Hi Oscar,

Thank you for your working on this.

My testing shows the following error:

  [ 1926.932435] ===> testcase 'mce_ksm_soft-offline_avoid_access.auto2' start
  [ 1927.155321] bash (15853): drop_caches: 3
  [ 1929.019094] page:ffffe5c384c4cd40 refcount:1 mapcount:0 mapping:0000000000000003 index:0x700000001
  [ 1929.021586] anon
  [ 1929.021588] flags: 0x57ffe00088000e(referenced|uptodate|dirty|swapbacked|hwpoison)
  [ 1929.024289] raw: 0057ffe00088000e dead000000000100 dead000000000122 0000000000000003
  [ 1929.026611] raw: 0000000700000001 0000000000000000 00000000ffffffff 0000000000000000
  [ 1929.028760] page dumped because: VM_BUG_ON_PAGE(page_ref_count(page))
  [ 1929.030559] ------------[ cut here ]------------
  [ 1929.031684] kernel BUG at mm/internal.h:73!
  [ 1929.032738] invalid opcode: 0000 [#1] SMP PTI
  [ 1929.033941] CPU: 3 PID: 16052 Comm: mceinj.sh Not tainted 5.3.0-rc8-v5.3-rc8-190911-1025-00010-ga436dbce8674+ #18
  [ 1929.037137] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.12.0-2.fc30 04/01/2014
  [ 1929.040066] RIP: 0010:page_set_poison+0xf9/0x160
  [ 1929.041665] Code: 63 02 7f 31 c0 5b 5d 41 5c c3 48 c7 c6 d0 d1 0c b0 48 89 df e8 88 bb f8 ff 0f 0b 48 c7 c6 f0 2a 0d b0 48 89 df e8 77 bb f8 ff <0f> 0b 48 8b 45 00 48 c1 e8 33 83 e0 07 83 f8 04 75 89 48 8b 45 08
  [ 1929.047773] RSP: 0018:ffffb4fb8a73bde0 EFLAGS: 00010246
  [ 1929.049511] RAX: 0000000000000039 RBX: ffffe5c384c4cd40 RCX: 0000000000000000
  [ 1929.051870] RDX: 0000000000000000 RSI: 0000000000000000 RDI: ffffffffb00d1814
  [ 1929.054238] RBP: ffffe5c384c4cd40 R08: 0000000000000596 R09: 0000000000000048
  [ 1929.056599] R10: 0000000000000000 R11: ffffb4fb8a73bc58 R12: 0000000000000000
  [ 1929.058986] R13: ffffb4fb8a73be10 R14: 0000000000131335 R15: 0000000000000001
  [ 1929.061366] FS:  00007fc9e208d740(0000) GS:ffff9fa9bdb00000(0000) knlGS:0000000000000000
  [ 1929.063842] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
  [ 1929.065429] CR2: 000055946c05d192 CR3: 00000001365f2000 CR4: 00000000001406e0
  [ 1929.067373] Call Trace:
  [ 1929.068094]  soft_offline_page+0x2be/0x600
  [ 1929.069246]  soft_offline_page_store+0xdf/0x110
  [ 1929.070510]  kernfs_fop_write+0x116/0x190
  [ 1929.071618]  vfs_write+0xa5/0x1a0
  [ 1929.072614]  ksys_write+0x59/0xd0
  [ 1929.073548]  do_syscall_64+0x5f/0x1a0
  [ 1929.074554]  entry_SYSCALL_64_after_hwframe+0x44/0xa9
  [ 1929.075957] RIP: 0033:0x7fc9e217ded8

It seems that soft-offlining on ksm pages is affected by this changeset.
Could you try to handle this?

- Naoya

On Tue, Sep 10, 2019 at 12:30:06PM +0200, Oscar Salvador wrote:
>
> This patchset was based on Naoya's hwpoison rework [1], so thanks to him
> for the initial work.
>
> This patchset aims to fix some issues laying in soft-offline handling,
> but it also takes the chance and takes some further steps to perform
> cleanups and some refactoring as well.
>
>  - Motivation:
>
>    A customer and I were facing an issue where poisoned pages we returned
>    back to user-space after having offlined them properly.
>    This was only seend under some memory stress + soft offlining pages.
>    After some anaylsis, it became clear that the problem was that
>    when kcompactd kicked in to migrate pages over, compaction_alloc
>    callback was handing poisoned pages to the migrate routine.
>    Once this page was later on fault in, __do_page_fault returned
>    VM_FAULT_HWPOISON making the process being killed.
>
>    All this could happen because isolate_freepages_block and
>    fast_isolate_freepages just check for the page to be PageBuddy,
>    and since 1) poisoned pages can be part of a higher order page
>    and 2) poisoned pages are also Page Buddy, they can sneak in easily.
>
>    I also saw some problem with swap pages, but I suspected to be the
>    same sort of problem, so I did not follow that trace.
>
>    The full explanation can be see in [2].
>
>  - Approach:
>
>    The taken approach is to not let poisoned pages hit neither
>    pcplists nor buddy freelists.
>    This is achieved by:
>
> In-use pages:
>
>    * Normal pages
>
>    1) do not release the last reference count after the
>       invalidation/migration of the page.
>    2) the page is being handed to page_set_poison, which does:
>       2a) sets PageHWPoison flag
>       2b) calls put_page (only to be able to call __page_cache_release)
>           Since poisoned pages are skipped in free_pages_prepare,
>           this put_page is safe.
>       2c) Sets the refcount to 1
>
>    * Hugetlb pages
>
>    1) Hand the page to page_set_poison after migration
>    2) page_set_poison does:
>       2a) Calls dissolve_free_huge_page
>       2b) If ranged to be dissolved contains poisoned pages,
>           we free the rangeas order-0 pages (as we do with gigantic hugetlb page),
>           so free_pages_prepare will skip them accordingly.
>       2c) Sets the refcount to 1
>
> Free pages:
>
>    * Normal pages:
>
>    1) Take the page off the buddy freelist
>    2) Set PageHWPoison flag and set refcount to 1
>
>    * Hugetlb pages
>
>    1) Try to allocate a new hugetlb page to the pool
>    2) Take off the pool the poisoned hugetlb
>
>
> With this patchset, I no longer see the issues I faced before.
>
> Note:
> I presented this as RFC to open discussion of the taken aproach.
> I think that furthers cleanups and refactors could be made, but I would
> like to get some insight of the taken approach before touching more
> code.
>
> Thanks
>
> [1] https://lore.kernel.org/linux-mm/1541746035-13408-1-git-send-email-n-horiguchi@ah.jp.nec.com/
> [2] https://lore.kernel.org/linux-mm/20190826104144.GA7849@linux/T/#u
>
> Naoya Horiguchi (5):
>   mm,hwpoison: cleanup unused PageHuge() check
>   mm,madvise: call soft_offline_page() without MF_COUNT_INCREASED
>   mm,hwpoison-inject: don't pin for hwpoison_filter
>   mm,hwpoison: remove MF_COUNT_INCREASED
>   mm: remove flag argument from soft offline functions
>
> Oscar Salvador (5):
>   mm,hwpoison: Unify THP handling for hard and soft offline
>   mm,hwpoison: Rework soft offline for in-use pages
>   mm,hwpoison: Refactor soft_offline_huge_page and __soft_offline_page
>   mm,hwpoison: Rework soft offline for free pages
>   mm,hwpoison: Use hugetlb_replace_page to replace free hugetlb pages
>
>  drivers/base/memory.c      |   2 +-
>  include/linux/mm.h         |   9 +-
>  include/linux/page-flags.h |   5 -
>  mm/hugetlb.c               |  51 +++++++-
>  mm/hwpoison-inject.c       |  18 +--
>  mm/madvise.c               |  25 ++--
>  mm/memory-failure.c        | 319 +++++++++++++++++++++------------------------
>  mm/migrate.c               |  11 +-
>  mm/page_alloc.c            |  62 +++++++--
>  9 files changed, 267 insertions(+), 235 deletions(-)
>
> --
> 2.12.3
>
>

Powered by blists - more mailing lists