Re: [PATCH v3] ceph: defer flushing the capsnap if the Fb is used

[Jeff Layton <jlayton@kernel.org>]

On Mon, 2021-01-18 at 17:10 +0800, Xiubo Li wrote:
> On 2021/1/13 5:48, Jeff Layton wrote:
> > On Sun, 2021-01-10 at 10:01 +0800, xiubli@redhat.com wrote:
> > > From: Xiubo Li <xiubli@redhat.com>
> > > 
> > > If the Fb cap is used it means the current inode is flushing the
> > > dirty data to OSD, just defer flushing the capsnap.
> > > 
> > > URL: https://tracker.ceph.com/issues/48679
> > > URL: https://tracker.ceph.com/issues/48640
> > > Signed-off-by: Xiubo Li <xiubli@redhat.com>
> > > ---
> > > 
> > > V3:
> > > - Add more comments about putting the inode ref
> > > - A small change about the code style
> > > 
> > > V2:
> > > - Fix inode reference leak bug
> > > 
> > >   fs/ceph/caps.c | 32 +++++++++++++++++++-------------
> > >   fs/ceph/snap.c |  6 +++---
> > >   2 files changed, 22 insertions(+), 16 deletions(-)
> > > 
> > Hi Xiubo,
> > 
> > This patch seems to cause hangs in some xfstests (generic/013, in
> > particular). I'll take a closer look when I have a chance, but I'm
> > dropping this for now.
> 
> Okay.
> 
> BTW, what's your test commands to reproduce it ? I will take a look when 
> I am free these days or later.
> 


FWIW, I was able to trigger a hang with this patch by running one of the
tests that this patch was intended to fix (snaptest-git-ceph.sh). Here's
the stack trace of the hung task:

# cat /proc/1166/stack
[<0>] wait_woken+0x87/0xb0
[<0>] ceph_get_caps+0x405/0x6a0 [ceph]
[<0>] ceph_write_iter+0x2ca/0xd20 [ceph]
[<0>] new_sync_write+0x10b/0x190
[<0>] vfs_write+0x240/0x390
[<0>] ksys_write+0x58/0xd0
[<0>] do_syscall_64+0x33/0x40
[<0>] entry_SYSCALL_64_after_hwframe+0x44/0xa9

Without this patch I could run that test in a loop without issue. This
bug mentions that the original issue occurred during mds thrashing
though, and I haven't tried reproducing that scenario yet:

    https://tracker.ceph.com/issues/48640

Cheers,
-- 
Jeff Layton <jlayton@kernel.org>