by Angel Shtilianov

[permalink] [raw]

Subject: Re: [PATCH] mm, vmscan: Do not wait for page writeback for GFP_NOFS allocations

Hello Hugh,

On 08/04/2015 09:32 AM, Hugh Dickins wrote:
> Hi Michal,
>
> On Thu, 2 Jul 2015, Michal Hocko wrote:
>> On Thu 02-07-15 10:25:51, Theodore Ts'o wrote:
>>> On Wed, Jul 01, 2015 at 03:37:15PM +0200, Michal Hocko wrote:
>> From: Michal Hocko <[email protected]>
>> Date: Thu, 2 Jul 2015 17:05:05 +0200
>> Subject: [PATCH] mm, vmscan: Do not wait for page writeback for GFP_NOFS
>> allocations
>>
>> Nikolay has reported a hang when a memcg reclaim got stuck with the
>> following backtrace...
>
> Sorry, I couldn't manage more than to ignore you when you Cc'ed me on
> this a month ago. Dave's perfectly correct, we had ourselves come to
> notice that recently: although in an ideal world a filesystem would
> only mark PageWriteback once the IO is all ready to go, in the real
> world that's not quite so, and a memory allocation may stand between.
> Which leaves my v3.6 c3b94f44fcb0 in danger of deadlocking.
>
> And suddenly now, in v4.2-rc or perhaps in v4.1 also, that has started
> hitting me too (I don't know which release Nicolay noticed this on).
> And it has become urgent to fix: I've added Linus to the Cc because
> I believe his comment in the rc5 announcement, "There's also a pending
> question about some of the VM changes", reflects this. Twice when I
> was trying to verify fixes to the dcache issue which came up at the
> end of last week, I was frustrated by unrelated hangs in my load.
> The first time I didn't recognize it, but the second time I did,
> and then came to realize that your patch is just what is needed.
>
> But I have modified it a little, I don't think you'll mind. As you
> suggested yourself, I actually prefer to test may_enter_fs there, rather
> than __GFP_FS: not a big deal, I certainly wouldn't want to delay the
> fix if someone thinks differently; but I tend to feel that may_enter_fs
> is what we already use for such decisions there, so better to use it.
> (And the SwapCache case immune to ext4 or xfs IO submission pattern.)
>
> I've fixed up the patch and updated the comments, since Tejun has
> meanwhile introduced sane_reclaim(sc) - I'm staying on in the insane
> asylum for now (and sane_reclaim is clearly unaffected by the change).
>
> I've omitted your hunk unindenting Case 3 wait_on_page_writeback(page):
> I prefer your style too, but thought it better to minimize the patch,
> especially if this is heading to the stables. (I was tempted to add in
> my unlock_page there, that we discussed once before: but again thought
> it better to minimize the fix - it is "selfish" not to unlock_page,
> but I think that anything heading for deadlock on the locked page would
> in other circumstances be heading for deadlock on the writeback page -
> I've never found that change critical.)
>
> And I've done quite a bit of testing. The loads that hung at the
> weekend have been running nicely for 24 hours now, no problem with the
> writeback hang and no problem with the dcache ENOTDIR issue. Though
> I've no idea of what recent VM change turned this into a hot issue.
>

Are these production loads you are referring to that have been able to
reproduce the issue or are they some synthetic ones which? So far I
haven't been able to reproduce the issue using artifical loads so I'm
interested in incorporating this into my test set setup if it's available?

> And more testing on the history of it, considering your stable 3.6+
> designation that I wasn't satisfied with. Getting out that USB stick
> again, I find that 3.6, 3.7 and 3.8 all OOM if their __GFP_IO test
> is updated to a may_enter_fs test; but something happened in 3.9
> to make it and subsequent releases safe with the may_enter_fs test.
> You can certainly argue that the remote chance of a deadlock is
> worse than the fair chance of a spurious OOM; but if you insist
> on 3.6+, then I think it would have to go back even further,
> because we marked that commit for stable itself. I suggest 3.9+.
>
>
> [PATCH] mm, vmscan: Do not wait for page writeback for GFP_NOFS allocations
>
> From: Michal Hocko <[email protected]>
>
> Nikolay has reported a hang when a memcg reclaim got stuck with the
> following backtrace:
> PID: 18308 TASK: ffff883d7c9b0a30 CPU: 1 COMMAND: "rsync"
> #0 [ffff88177374ac60] __schedule at ffffffff815ab152
> #1 [ffff88177374acb0] schedule at ffffffff815ab76e
> #2 [ffff88177374acd0] schedule_timeout at ffffffff815ae5e5
> #3 [ffff88177374ad70] io_schedule_timeout at ffffffff815aad6a
> #4 [ffff88177374ada0] bit_wait_io at ffffffff815abfc6
> #5 [ffff88177374adb0] __wait_on_bit at ffffffff815abda5
> #6 [ffff88177374ae00] wait_on_page_bit at ffffffff8111fd4f
> #7 [ffff88177374ae50] shrink_page_list at ffffffff81135445
> #8 [ffff88177374af50] shrink_inactive_list at ffffffff81135845
> #9 [ffff88177374b060] shrink_lruvec at ffffffff81135ead
> #10 [ffff88177374b150] shrink_zone at ffffffff811360c3
> #11 [ffff88177374b220] shrink_zones at ffffffff81136eff
> #12 [ffff88177374b2a0] do_try_to_free_pages at ffffffff8113712f
> #13 [ffff88177374b300] try_to_free_mem_cgroup_pages at ffffffff811372be
> #14 [ffff88177374b380] try_charge at ffffffff81189423
> #15 [ffff88177374b430] mem_cgroup_try_charge at ffffffff8118c6f5
> #16 [ffff88177374b470] __add_to_page_cache_locked at ffffffff8112137d
> #17 [ffff88177374b4e0] add_to_page_cache_lru at ffffffff81121618
> #18 [ffff88177374b510] pagecache_get_page at ffffffff8112170b
> #19 [ffff88177374b560] grow_dev_page at ffffffff811c8297
> #20 [ffff88177374b5c0] __getblk_slow at ffffffff811c91d6
> #21 [ffff88177374b600] __getblk_gfp at ffffffff811c92c1
> #22 [ffff88177374b630] ext4_ext_grow_indepth at ffffffff8124565c
> #23 [ffff88177374b690] ext4_ext_create_new_leaf at ffffffff81246ca8
> #24 [ffff88177374b6e0] ext4_ext_insert_extent at ffffffff81246f09
> #25 [ffff88177374b750] ext4_ext_map_blocks at ffffffff8124a848
> #26 [ffff88177374b870] ext4_map_blocks at ffffffff8121a5b7
> #27 [ffff88177374b910] mpage_map_one_extent at ffffffff8121b1fa
> #28 [ffff88177374b950] mpage_map_and_submit_extent at ffffffff8121f07b
> #29 [ffff88177374b9b0] ext4_writepages at ffffffff8121f6d5
> #30 [ffff88177374bb20] do_writepages at ffffffff8112c490
> #31 [ffff88177374bb30] __filemap_fdatawrite_range at ffffffff81120199
> #32 [ffff88177374bb80] filemap_flush at ffffffff8112041c
> #33 [ffff88177374bb90] ext4_alloc_da_blocks at ffffffff81219da1
> #34 [ffff88177374bbb0] ext4_rename at ffffffff81229b91
> #35 [ffff88177374bcd0] ext4_rename2 at ffffffff81229e32
> #36 [ffff88177374bce0] vfs_rename at ffffffff811a08a5
> #37 [ffff88177374bd60] SYSC_renameat2 at ffffffff811a3ffc
> #38 [ffff88177374bf60] sys_renameat2 at ffffffff811a408e
> #39 [ffff88177374bf70] sys_rename at ffffffff8119e51e
> #40 [ffff88177374bf80] system_call_fastpath at ffffffff815afa89
>
> Dave Chinner has properly pointed out that this is a deadlock in the
> reclaim code because ext4 doesn't submit pages which are marked by
> PG_writeback right away. The heuristic was introduced by e62e384e9da8
> ("memcg: prevent OOM with too many dirty pages") and it was applied
> only when may_enter_fs was specified. The code has been changed by
> c3b94f44fcb0 ("memcg: further prevent OOM with too many dirty pages")
> which has removed the __GFP_FS restriction with a reasoning that we
> do not get into the fs code. But this is not sufficient apparently
> because the fs doesn't necessarily submit pages marked PG_writeback
> for IO right away.
>
> ext4_bio_write_page calls io_submit_add_bh but that doesn't necessarily
> submit the bio. Instead it tries to map more pages into the bio and
> mpage_map_one_extent might trigger memcg charge which might end up
> waiting on a page which is marked PG_writeback but hasn't been submitted
> yet so we would end up waiting for something that never finishes.
>
> Fix this issue by replacing __GFP_IO by may_enter_fs check (for case 2)
> before we go to wait on the writeback. The page fault path, which is the
> only path that triggers memcg oom killer since 3.12, shouldn't require
> GFP_NOFS and so we shouldn't reintroduce the premature OOM killer issue
> which was originally addressed by the heuristic.
>
> As per David Chinner the xfs is doing similar thing since 2.6.15 already
> so ext4 is not the only affected filesystem. Moreover he notes:
> : For example: IO completion might require unwritten extent conversion
> : which executes filesystem transactions and GFP_NOFS allocations. The
> : writeback flag on the pages can not be cleared until unwritten
> : extent conversion completes. Hence memory reclaim cannot wait on
> : page writeback to complete in GFP_NOFS context because it is not
> : safe to do so, memcg reclaim or otherwise.
>
> Cc: [email protected] # 3.9+
> [[email protected]: corrected the control flow]
> Fixes: c3b94f44fcb0 ("memcg: further prevent OOM with too many dirty pages")
> Reported-by: Nikolay Borisov <[email protected]>
> Signed-off-by: Michal Hocko <[email protected]>
> Signed-off-by: Hugh Dickins <[email protected]>
> ---
>
> mm/vmscan.c | 16 ++++++----------
> 1 file changed, 6 insertions(+), 10 deletions(-)
>
> --- 4.2-rc5/mm/vmscan.c 2015-07-05 19:25:02.856131170 -0700
> +++ linux/mm/vmscan.c 2015-08-02 21:24:03.000614050 -0700
> @@ -973,22 +973,18 @@ static unsigned long shrink_page_list(st
> * caller can stall after page list has been processed.
> *
> * 2) Global or new memcg reclaim encounters a page that is
> - * not marked for immediate reclaim or the caller does not
> - * have __GFP_IO. In this case mark the page for immediate
> + * not marked for immediate reclaim, or the caller does not
> + * have __GFP_FS (or __GFP_IO if it's simply going to swap,
> + * not to fs). In this case mark the page for immediate
> * reclaim and continue scanning.
> *
> - * __GFP_IO is checked because a loop driver thread might
> + * Require may_enter_fs because we would wait on fs, which
> + * may not have submitted IO yet. And the loop driver might
> * enter reclaim, and deadlock if it waits on a page for
> * which it is needed to do the write (loop masks off
> * __GFP_IO|__GFP_FS for this reason); but more thought
> * would probably show more reasons.
> *
> - * Don't require __GFP_FS, since we're not going into the
> - * FS, just waiting on its writeback completion. Worryingly,
> - * ext4 gfs2 and xfs allocate pages with
> - * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
> - * may_enter_fs here is liable to OOM on them.
> - *
> * 3) Legacy memcg encounters a page that is not already marked
> * PageReclaim. memcg does not have any dirty pages
> * throttling so we could easily OOM just because too many
> @@ -1005,7 +1001,7 @@ static unsigned long shrink_page_list(st
>
> /* Case 2 above */
> } else if (sane_reclaim(sc) ||
> - !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
> + !PageReclaim(page) || !may_enter_fs) {
> /*
> * This is slightly racy - end_page_writeback()
> * might have just cleared PageReclaim, then
>

2015-08-13 02:14:23

by Hugh Dickins

[permalink] [raw]

Subject: Re: [PATCH] mm, vmscan: Do not wait for page writeback for GFP_NOFS allocations

On Fri, 7 Aug 2015, Nikolay Borisov wrote:
> On 08/04/2015 09:32 AM, Hugh Dickins wrote:
> >
> > And I've done quite a bit of testing. The loads that hung at the
> > weekend have been running nicely for 24 hours now, no problem with the
> > writeback hang and no problem with the dcache ENOTDIR issue. Though
> > I've no idea of what recent VM change turned this into a hot issue.
> >
>
> Are these production loads you are referring to that have been able to
> reproduce the issue or are they some synthetic ones which? So far I
> haven't been able to reproduce the issue using artifical loads so I'm
> interested in incorporating this into my test set setup if it's available?

Not production loads, no, just an artificial load. But not very good
at reproducing the hang: variable, but took hours, and only showed up
on one faster machine; I had to run the load for 2 days, then again 2
days, to feel confident that this hang was fixed.

And I'm sorry, but describing it in full detail is not something I find
time to do, in days or in years - partly because once I try to detail it,
I need to simplify this and streamline that, and it turns into something
else. As happened when I sent it, offlist, to someone in 2009: I looked
back at that with a view to forwarding to you, but a lot of the details
don't match what I reverted to or advanced to doing since.

Broadly, it's a pair of repeated make -j20 kernel builds, one in tmpfs,
one in ext4 over loop over tmpfs, in limited memory 700M with 1.5G swap.
And to test this particular hang, it needed to use memcg (of what's now
branded an "insane" variety, CONFIG_CGROUP_WRITEBACK=n): I was using 1G
not 700M ram for this, but 300M memcg limit and 250M soft limit on each
memcg that was hosting one of the pair of repeated builds. It can be
difficult to tune the right balance, swapping heavily but not OOMing:
it's a 2.6.24 tree I've gone back to building, because that's so much
smaller than current, with a greater proportion active in the build.

Hugh