mbox series

[v6,00/12] HWPOISON: soft offline rework

Message ID 20200806184923.7007-1-nao.horiguchi@gmail.com
Headers show
Series HWPOISON: soft offline rework | expand

Message

Naoya Horiguchi Aug. 6, 2020, 6:49 p.m. UTC
Hi,

This patchset is the latest version of soft offline rework patchset
targetted for v5.9.

Since v5, I dropped some patches which tweak refcount handling in
madvise_inject_error() to avoid the "unknown refcount page" error.
I don't confirm the fix (that didn't reproduce with v5 in my environment),
but this change surely call soft_offline_page() after holding refcount,
so the error should not happen any more.

Dropped patches
- mm,madvise: call soft_offline_page() without MF_COUNT_INCREASED
- mm,madvise: Refactor madvise_inject_error
- mm,hwpoison: remove MF_COUNT_INCREASED
- mm,hwpoison: remove flag argument from soft offline functions

Thanks,
Naoya Horiguchi

Quoting cover letter of v5:
----
Main focus of this series is to stabilize soft offline.  Historically soft
offlined pages have suffered from racy conditions because PageHWPoison is
used to a little too aggressively, which (directly or indirectly) invades
other mm code which cares little about hwpoison.  This results in unexpected
behavior or kernel panic, which is very far from soft offline's "do not
disturb userspace or other kernel component" policy.

Main point of this change set is to contain target page "via buddy allocator",
where we first free the target page as we do for normal pages, and remove
from buddy only when we confirm that it reaches free list. There is surely
race window of page allocation, but that's fine because someone really want
that page and the page is still working, so soft offline can happily give up.

v4 from Oscar tries to handle the race around reallocation, but that part
seems still work in progress, so I decide to separate it for changes into
v5.9.  Thank you for your contribution, Oscar.

---
Previous versions:
  v1: https://lore.kernel.org/linux-mm/1541746035-13408-1-git-send-email-n-horiguchi@ah.jp.nec.com/
  v2: https://lore.kernel.org/linux-mm/20191017142123.24245-1-osalvador@suse.de/
  v3: https://lore.kernel.org/linux-mm/20200624150137.7052-1-nao.horiguchi@gmail.com/
  v4: https://lore.kernel.org/linux-mm/20200716123810.25292-1-osalvador@suse.de/
  v5: https://lore.kernel.org/linux-mm/20200805204354.GA16406@hori.linux.bs1.fc.nec.co.jp/T/#t
---
Summary:

Naoya Horiguchi (5):
      mm,hwpoison: cleanup unused PageHuge() check
      mm, hwpoison: remove recalculating hpage
      mm,hwpoison-inject: don't pin for hwpoison_filter
      mm,hwpoison: introduce MF_MSG_UNSPLIT_THP
      mm,hwpoison: double-check page count in __get_any_page()

Oscar Salvador (7):
      mm,hwpoison: Un-export get_hwpoison_page and make it static
      mm,hwpoison: Kill put_hwpoison_page
      mm,hwpoison: Unify THP handling for hard and soft offline
      mm,hwpoison: Rework soft offline for free pages
      mm,hwpoison: Rework soft offline for in-use pages
      mm,hwpoison: Refactor soft_offline_huge_page and __soft_offline_page
      mm,hwpoison: Return 0 if the page is already poisoned in soft-offline

 include/linux/mm.h         |   3 +-
 include/linux/page-flags.h |   6 +-
 include/ras/ras_event.h    |   3 +
 mm/hwpoison-inject.c       |  18 +--
 mm/madvise.c               |   5 -
 mm/memory-failure.c        | 307 +++++++++++++++++++++------------------------
 mm/migrate.c               |  11 +-
 mm/page_alloc.c            |  60 +++++++--
 8 files changed, 203 insertions(+), 210 deletions(-)

Comments

HORIGUCHI NAOYA(堀口 直也) Aug. 11, 2020, 3:11 a.m. UTC | #1
On Mon, Aug 10, 2020 at 11:22:55AM -0400, Qian Cai wrote:
> On Thu, Aug 06, 2020 at 06:49:11PM +0000, nao.horiguchi@gmail.com wrote:
> > Hi,
> > 
> > This patchset is the latest version of soft offline rework patchset
> > targetted for v5.9.
> > 
> > Since v5, I dropped some patches which tweak refcount handling in
> > madvise_inject_error() to avoid the "unknown refcount page" error.
> > I don't confirm the fix (that didn't reproduce with v5 in my environment),
> > but this change surely call soft_offline_page() after holding refcount,
> > so the error should not happen any more.
> 
> With this patchset, arm64 is still suffering from premature 512M-size hugepages
> allocation failures.
> 
> # git clone https://gitlab.com/cailca/linux-mm
> # cd linux-mm; make
> # ./random 1
> - start: migrate_huge_offline
> - use NUMA nodes 0,1.
> - mmap and free 2147483648 bytes hugepages on node 0
> - mmap and free 2147483648 bytes hugepages on node 1
> madvise: Cannot allocate memory
> 
> [  292.456538][ T3685] soft offline: 0x8a000: hugepage isolation failed: 0, page count 2, type 7ffff80001000e (referenced|uptodate|dirty|head)
> [  292.469113][ T3685] Soft offlining pfn 0x8c000 at process virtual address 0xffff60000000
> [  292.983855][ T3685] Soft offlining pfn 0x88000 at process virtual address 0xffff40000000
> [  293.271369][ T3685] Soft offlining pfn 0x8a000 at process virtual address 0xffff60000000
> [  293.834030][ T3685] Soft offlining pfn 0xa000 at process virtual address 0xffff40000000
> [  293.851378][ T3685] soft offline: 0xa000: hugepage migration failed -12, type 7ffff80001000e (referenced|uptodate|dirty|head)
> 
> The fresh-booted system still had 40G+ memory free before running the test.

As I commented over v5, this failure is expected and it doesn't mean kernel
issue.  Once we successfully soft offline a hugepage, the memory range
covering the hugepage will never participate in hugepage because one of the
subpages is removed from buddy.  So if you iterate soft offlining hugepages,
all memory range are "holed" finally, and no hugepage will be available in
the system.

Please fix your test program to properly determine nubmer of loop (NR_LOOP)
so that you can assume that you can always allocate hugepage during testing.
For example, if you can use 40G memory and hugepage size is 512MB, NR_LOOP
should not be larger than 80.

> 
> Reverting the following commits allowed the test to run succesfully over and over again.
> 
> "mm, hwpoison: remove recalculating hpage"
> "mm,hwpoison-inject: don't pin for hwpoison_filter"
> "mm,hwpoison: Un-export get_hwpoison_page and make it static"
> "mm,hwpoison: kill put_hwpoison_page"
> "mm,hwpoison: unify THP handling for hard and soft offline"
> "mm,hwpoison: rework soft offline for free pages"
> "mm,hwpoison: rework soft offline for in-use pages"
> "mm,hwpoison: refactor soft_offline_huge_page and __soft_offline_page"

I'm still not sure why the test succeeded by reverting these because
current mainline kernel provides similar mechanism to prevent reuse of
soft offlined page. So this success seems to me something suspicious.

To investigate more, I want to have additional info about the page states
of the relevant pages after soft offlining.  Could you collect it by the
following steps?

  - modify random.c not to run hotplug_memory() in migrate_huge_hotplug_memory(),
  - compile it and run "./random 1" once,
  - to collect page state with hwpoisoned pages, run "./page-types -Nlr -b hwpoison",
    where page-types is available under tools/vm in kernel source tree.
  - choose a few pfns of soft offlined pages from kernel message
    "Soft offlining pfn ...", and run "./page-types -Nlr -a <pfn>".

Thanks,
Naoya Horiguchi

> 
> i.e., it is not enough to only revert,
> 
> mm,hwpoison: double-check page count in __get_any_page()
> mm,hwpoison: introduce MF_MSG_UNSPLIT_THP
> mm,hwpoison: return 0 if the page is already poisoned in soft-offline
>
HORIGUCHI NAOYA(堀口 直也) Aug. 11, 2020, 3:56 a.m. UTC | #2
On Mon, Aug 10, 2020 at 11:45:36PM -0400, Qian Cai wrote:
> 
> 
> > On Aug 10, 2020, at 11:11 PM, HORIGUCHI NAOYA(堀口 直也) <naoya.horiguchi@nec.com> wrote:
> > 
> > I'm still not sure why the test succeeded by reverting these because
> > current mainline kernel provides similar mechanism to prevent reuse of
> > soft offlined page. So this success seems to me something suspicious.
> 
> Even if we call munmap() on the range, it still can’t be be reused? If so, how to recover those memory then?

No, it can't, because soft offline isolates the physical page.
so even after calling munmap(), the side effect remains on the page.
In your random.c, memory online/offline resets the status of hwpoison.
So you can reallocate hugepages in another run of the program.

> 
> > 
> > To investigate more, I want to have additional info about the page states
> > of the relevant pages after soft offlining.  Could you collect it by the
> > following steps?
> 
> Do you want to collect those from the failed or succeed kernel?

I'd like to check on the succeeded kernel.
Sorry for the lack of information.

Thanks,
Naoya Horiguchi
Naoya Horiguchi Aug. 11, 2020, 7:32 p.m. UTC | #3
On Tue, Aug 11, 2020 at 01:39:24PM -0400, Qian Cai wrote:
> On Tue, Aug 11, 2020 at 03:11:40AM +0000, HORIGUCHI NAOYA(堀口 直也) wrote:
> > I'm still not sure why the test succeeded by reverting these because
> > current mainline kernel provides similar mechanism to prevent reuse of
> > soft offlined page. So this success seems to me something suspicious.
> > 
> > To investigate more, I want to have additional info about the page states
> > of the relevant pages after soft offlining.  Could you collect it by the
> > following steps?
> > 
> >   - modify random.c not to run hotplug_memory() in migrate_huge_hotplug_memory(),
> >   - compile it and run "./random 1" once,
> >   - to collect page state with hwpoisoned pages, run "./page-types -Nlr -b hwpoison",
> >     where page-types is available under tools/vm in kernel source tree.
> >   - choose a few pfns of soft offlined pages from kernel message
> >     "Soft offlining pfn ...", and run "./page-types -Nlr -a <pfn>".
> 
> # ./page-types -Nlr -b hwpoison
> offset	len	flags
> 99a000	1	__________B________X_______________________
> 99c000	1	__________B________X_______________________
> 99e000	1	__________B________X_______________________
> 9a0000	1	__________B________X_______________________
> ba6000	1	__________B________X_______________________
> baa000	1	__________B________X_______________________

Thank you.  It only shows 6 lines of records, which is unexpected to me
because random.c iterates soft offline 2 hugepages with madvise() 1000 times.
Somehow (maybe in arch specific way?) other hwpoisoned pages might be cleared?
If they really are, the success of this test is a fake, and this patchset
can be considered as a fix.

> 
> Every single one of pfns was like this,
> 
> # ./page-types -Nlr -a 0x99a000
> offset	len	flags
> 99a000	1	__________B________X_______________________
> 
> # ./page-types -Nlr -a 0x99e000
> offset	len	flags
> 99e000	1	__________B________X_______________________
> 
> # ./page-types -Nlr -a 0x99c000
> offset	len	flags
> 99c000	1	__________B________X_______________________
osalvador Aug. 24, 2020, 12:21 p.m. UTC | #4
On Thu, Aug 06, 2020 at 06:49:23PM +0000, nao.horiguchi@gmail.com wrote:
> From: Naoya Horiguchi <naoya.horiguchi@nec.com>
> 
> Soft offlining could fail with EIO due to the race condition with
> hugepage migration. This issuse became visible due to the change by
> previous patch that makes soft offline handler take page refcount
> by its own.  We have no way to directly pin zero refcount page, and
> the page considered as a zero refcount page could be allocated just
> after the first check.
> 
> This patch adds the second check to find the race and gives us
> chance to handle it more reliably.
> 
> Reported-by: Qian Cai <cai@lca.pw>
> Signed-off-by: Naoya Horiguchi <naoya.horiguchi@nec.com>

Reviewed-by: Oscar Salvador <osalvador@suse.de>