Re: [PATCH v3] xfs: avoid deadlock when trigger memory reclaim in ->writepages

[Michal Hocko <mhocko@kernel.org>]

On Tue 16-06-20 17:39:33, Yafang Shao wrote:
> On Tue, Jun 16, 2020 at 5:27 PM Michal Hocko <mhocko@kernel.org> wrote:
> >
> > On Tue 16-06-20 17:05:25, Yafang Shao wrote:
> > > On Tue, Jun 16, 2020 at 4:16 PM Michal Hocko <mhocko@kernel.org> wrote:
> > > >
> > > > On Mon 15-06-20 07:56:21, Yafang Shao wrote:
> > > > > Recently there is a XFS deadlock on our server with an old kernel.
> > > > > This deadlock is caused by allocating memory in xfs_map_blocks() while
> > > > > doing writeback on behalf of memroy reclaim. Although this deadlock happens
> > > > > on an old kernel, I think it could happen on the upstream as well. This
> > > > > issue only happens once and can't be reproduced, so I haven't tried to
> > > > > reproduce it on upsteam kernel.
> > > > >
> > > > > Bellow is the call trace of this deadlock.
> > > > > [480594.790087] INFO: task redis-server:16212 blocked for more than 120 seconds.
> > > > > [480594.790087] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
> > > > > [480594.790088] redis-server    D ffffffff8168bd60     0 16212  14347 0x00000004
> > > > > [480594.790090]  ffff880da128f070 0000000000000082 ffff880f94a2eeb0 ffff880da128ffd8
> > > > > [480594.790092]  ffff880da128ffd8 ffff880da128ffd8 ffff880f94a2eeb0 ffff88103f9d6c40
> > > > > [480594.790094]  0000000000000000 7fffffffffffffff ffff88207ffc0ee8 ffffffff8168bd60
> > > > > [480594.790096] Call Trace:
> > > > > [480594.790101]  [<ffffffff8168dce9>] schedule+0x29/0x70
> > > > > [480594.790103]  [<ffffffff8168b749>] schedule_timeout+0x239/0x2c0
> > > > > [480594.790111]  [<ffffffff8168d28e>] io_schedule_timeout+0xae/0x130
> > > > > [480594.790114]  [<ffffffff8168d328>] io_schedule+0x18/0x20
> > > > > [480594.790116]  [<ffffffff8168bd71>] bit_wait_io+0x11/0x50
> > > > > [480594.790118]  [<ffffffff8168b895>] __wait_on_bit+0x65/0x90
> > > > > [480594.790121]  [<ffffffff811814e1>] wait_on_page_bit+0x81/0xa0
> > > > > [480594.790125]  [<ffffffff81196ad2>] shrink_page_list+0x6d2/0xaf0
> > > > > [480594.790130]  [<ffffffff811975a3>] shrink_inactive_list+0x223/0x710
> > > > > [480594.790135]  [<ffffffff81198225>] shrink_lruvec+0x3b5/0x810
> > > > > [480594.790139]  [<ffffffff8119873a>] shrink_zone+0xba/0x1e0
> > > > > [480594.790141]  [<ffffffff81198c20>] do_try_to_free_pages+0x100/0x510
> > > > > [480594.790143]  [<ffffffff8119928d>] try_to_free_mem_cgroup_pages+0xdd/0x170
> > > > > [480594.790145]  [<ffffffff811f32de>] mem_cgroup_reclaim+0x4e/0x120
> > > > > [480594.790147]  [<ffffffff811f37cc>] __mem_cgroup_try_charge+0x41c/0x670
> > > > > [480594.790153]  [<ffffffff811f5cb6>] __memcg_kmem_newpage_charge+0xf6/0x180
> > > > > [480594.790157]  [<ffffffff8118c72d>] __alloc_pages_nodemask+0x22d/0x420
> > > > > [480594.790162]  [<ffffffff811d0c7a>] alloc_pages_current+0xaa/0x170
> > > > > [480594.790165]  [<ffffffff811db8fc>] new_slab+0x30c/0x320
> > > > > [480594.790168]  [<ffffffff811dd17c>] ___slab_alloc+0x3ac/0x4f0
> > > > > [480594.790204]  [<ffffffff81685656>] __slab_alloc+0x40/0x5c
> > > > > [480594.790206]  [<ffffffff811dfc43>] kmem_cache_alloc+0x193/0x1e0
> > > > > [480594.790233]  [<ffffffffa04fab67>] kmem_zone_alloc+0x97/0x130 [xfs]
> > > > > [480594.790247]  [<ffffffffa04f90ba>] _xfs_trans_alloc+0x3a/0xa0 [xfs]
> > > > > [480594.790261]  [<ffffffffa04f915c>] xfs_trans_alloc+0x3c/0x50 [xfs]
> > > >
> > > > Now with a more explanation from Dave I have got back to the original
> > > > backtrace. Not sure which kernel version you are using but this path
> > > > should have passed xfs_trans_reserve which sets PF_MEMALLOC_NOFS and
> > > > this in turn should have made __alloc_pages_nodemask to use __GFP_NOFS
> > > > and the memcg reclaim shouldn't ever wait_on_page_writeback (pressumably
> > > > this is what the io_schedule is coming from).
> > >
> > > Hi Michal,
> > >
> > > The page is allocated before calling xfs_trans_reserve, so the
> > > PF_MEMALLOC_NOFS hasn't been set yet.
> > > See bellow,
> > >
> > > xfs_trans_alloc
> > >     kmem_zone_zalloc() <<< GPF_NOFS hasn't been set yet, but it may
> > > trigger memory reclaim
> > >     xfs_trans_reserve() <<<< GPF_NOFS is set here (for the kernel
> > > prior to commit 9070733b4efac, it is PF_FSTRANS)
> >
> > You are right, I have misread the actual allocation side. 8683edb7755b8
> > has added KM_NOFS and 73d30d48749f8 has removed it. I cannot really
> > judge the correctness here.
> >
> 
> The history is complicated, but it doesn't matter.
> Let's  turn back to the upstream kernel now. As I explained in the commit log,
> xfs_vm_writepages
>   -> iomap_writepages.
>      -> write_cache_pages
>         -> lock_page <<<< This page is locked.
>         -> writepages ( which is  iomap_do_writepage)
>            -> xfs_map_blocks
>               -> xfs_convert_blocks
>                  -> xfs_bmapi_convert_delalloc
>                     -> xfs_trans_alloc
>                          -> kmem_zone_zalloc //It should alloc page
> with GFP_NOFS
> 
> If GFP_NOFS isn't set in xfs_trans_alloc(), the kmem_zone_zalloc() may
> trigger the memory reclaim then it may wait on the page locked in
> write_cache_pages() ...

This cannot happen because the memory reclaim backs off on locked pages.
Have a look at trylock_page at the very beginning of the shrink_page_list
loop. You are likely waiting on a different page which is not being
flushed because of the FS ordering requirement or something like that.

> That means the ->writepages should be set with GFP_NOFS to avoid this
> recursive filesystem reclaim.

-- 
Michal Hocko
SUSE Labs