Re: INFO: task hung in mpage_prepare_extent_to_map

[Johannes Weiner]

On Mon, Oct 28, 2019 at 12:52:09PM -0700, syzbot wrote:
> Hello,
> 
> syzbot found the following crash on:
> 
> HEAD commit:    12d61c69 Add linux-next specific files for 20191024
> git tree:       linux-next
> console output: https://syzkaller.appspot.com/x/log.txt?x=15a0fa97600000
> kernel config:  https://syzkaller.appspot.com/x/.config?x=afb75fd8c9fd5ed8
> dashboard link: https://syzkaller.appspot.com/bug?extid=efb9e48b9fbdc49bb34a
> compiler:       gcc (GCC) 9.0.0 20181231 (experimental)
> syz repro:      https://syzkaller.appspot.com/x/repro.syz?x=13a63dc4e00000
> 
> The bug was bisected to:
> 
> commit 9c61acffe2b8833152041f7b6a02d1d0a17fd378
> Author: Song Liu <songliubraving@xxxxxx>
> Date:   Wed Oct 23 00:24:28 2019 +0000
> 
>     mm,thp: recheck each page before collapsing file THP
> 
> bisection log:  https://syzkaller.appspot.com/x/bisect.txt?x=13eb6ec0e00000
> final crash:    https://syzkaller.appspot.com/x/report.txt?x=101b6ec0e00000
> console output: https://syzkaller.appspot.com/x/log.txt?x=17eb6ec0e00000
> 
> IMPORTANT: if you fix the bug, please add the following tag to the commit:
> Reported-by: syzbot+efb9e48b9fbdc49bb34a@xxxxxxxxxxxxxxxxxxxxxxxxx
> Fixes: 9c61acffe2b8 ("mm,thp: recheck each page before collapsing file THP")
> 
> INFO: task khugepaged:1084 blocked for more than 143 seconds.
>       Not tainted 5.4.0-rc4-next-20191024 #0
> "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
> khugepaged      D27568  1084      2 0x80004000
> Call Trace:
>  context_switch kernel/sched/core.c:3384 [inline]
>  __schedule+0x94a/0x1e70 kernel/sched/core.c:4069
>  schedule+0xd9/0x260 kernel/sched/core.c:4136
>  io_schedule+0x1c/0x70 kernel/sched/core.c:5780
>  wait_on_page_bit_common mm/filemap.c:1175 [inline]
>  __lock_page+0x422/0xab0 mm/filemap.c:1383
>  lock_page include/linux/pagemap.h:480 [inline]
>  mpage_prepare_extent_to_map+0xb3f/0xf90 fs/ext4/inode.c:2668
>  ext4_writepages+0xb6a/0x2e70 fs/ext4/inode.c:2866
>  ? 0xffffffff81000000
>  do_writepages+0xfa/0x2a0 mm/page-writeback.c:2344
>  __filemap_fdatawrite_range+0x2bc/0x3b0 mm/filemap.c:421
>  __filemap_fdatawrite mm/filemap.c:429 [inline]
>  filemap_flush+0x24/0x30 mm/filemap.c:456

This is a double locking deadlock. The page lock is already held when
we call into filemap_flush() here, and does another lock_page() in
write_cache_pages().

To fix it, we have to either initiate flushing before acquiring the
page lock, or simply skip over dirty pages.

Maybe doing vfs_fsync_range() from the madvise(HUGEPAGE) call isn't a
bad idea after all? (I had discussed this with Song off-list before.)