Re: [PATCH] fs: xfs: xfs_log: Don't use KM_MAYFAIL at xfs_log_reserve().

[Luis Chamberlain <mcgrof@kernel.org>]

On Thu, Aug 01, 2019 at 07:06:35PM +0900, Tetsuo Handa wrote:
> When the system is close-to-OOM, fsync() may fail due to -ENOMEM because
> xfs_log_reserve() is using KM_MAYFAIL. It is a bad thing to fail writeback
> operation due to user-triggerable OOM condition. Since we are not using
> KM_MAYFAIL at xfs_trans_alloc() before calling xfs_log_reserve(), let's
> use the same flags at xfs_log_reserve().
> 
>   oom-torture: page allocation failure: order:0, mode:0x46c40(GFP_NOFS|__GFP_NOWARN|__GFP_RETRY_MAYFAIL|__GFP_COMP), nodemask=(null)
>   CPU: 7 PID: 1662 Comm: oom-torture Kdump: loaded Not tainted 5.3.0-rc2+ #925
>   Hardware name: VMware, Inc. VMware Virtual Platform/440BX Desktop Reference Platform, BIOS 6.00
>   Call Trace:
>    dump_stack+0x67/0x95
>    warn_alloc+0xa9/0x140
>    __alloc_pages_slowpath+0x9a8/0xbce
>    __alloc_pages_nodemask+0x372/0x3b0
>    alloc_slab_page+0x3a/0x8d0
>    new_slab+0x330/0x420
>    ___slab_alloc.constprop.94+0x879/0xb00
>    __slab_alloc.isra.89.constprop.93+0x43/0x6f
>    kmem_cache_alloc+0x331/0x390
>    kmem_zone_alloc+0x9f/0x110 [xfs]
>    kmem_zone_alloc+0x9f/0x110 [xfs]
>    xlog_ticket_alloc+0x33/0xd0 [xfs]
>    xfs_log_reserve+0xb4/0x410 [xfs]
>    xfs_trans_reserve+0x1d1/0x2b0 [xfs]
>    xfs_trans_alloc+0xc9/0x250 [xfs]
>    xfs_setfilesize_trans_alloc.isra.27+0x44/0xc0 [xfs]
>    xfs_submit_ioend.isra.28+0xa5/0x180 [xfs]
>    xfs_vm_writepages+0x76/0xa0 [xfs]
>    do_writepages+0x17/0x80
>    __filemap_fdatawrite_range+0xc1/0xf0
>    file_write_and_wait_range+0x53/0xa0
>    xfs_file_fsync+0x87/0x290 [xfs]
>    vfs_fsync_range+0x37/0x80
>    do_fsync+0x38/0x60
>    __x64_sys_fsync+0xf/0x20
>    do_syscall_64+0x4a/0x1c0
>    entry_SYSCALL_64_after_hwframe+0x49/0xbe
> 
> Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>

That's quite an opaque commit log for what started off as a severe email
thread of potential leak of information. As such, can you expand on this
commit log considerably to explain the situation a bit better? Your
initial thread here provided much clearer evidence of the issue. As-is
this commit log tells the reader *nothing* about the potential harm in
not applying this patch.

You had mentioned you identified this issue present on at least
4.18 till 5.3-rc1. So, I'm at least inclined to consider this for
stable for at least v4.19.

However, what about older kernels? Now that you have identified
a fix, were the flag changed in prior commits, is it a regression
that perhaps added KM_MAYFAIL at some point?

  Luis