Re: Killing process in D state on mount to dead NFS server. (when process is in fsync)

[Jeff Layton <jlayton@poochiereds.net>]

[-- Attachment #1: Type: text/plain, Size: 6034 bytes --]

On Fri, 1 Aug 2014 07:50:53 +1000
NeilBrown <neilb@suse.de> wrote:

> On Thu, 31 Jul 2014 14:20:07 -0700 Ben Greear <greearb@candelatech.com> wrote:
> 
> > -----BEGIN PGP SIGNED MESSAGE-----
> > Hash: SHA1
> > 
> > On 07/31/2014 01:42 PM, NeilBrown wrote:
> > > On Thu, 31 Jul 2014 11:00:35 -0700 Ben Greear <greearb@candelatech.com> wrote:
> > > 
> > >> So, this has been asked all over the interweb for years and years, but the best answer I can find is to reboot the system or create a fake NFS server
> > >> somewhere with the same IP as the gone-away NFS server.
> > >> 
> > >> The problem is:
> > >> 
> > >> I have some mounts to an NFS server that no longer exists (crashed/powered down).
> > >> 
> > >> I have some processes stuck trying to write to files open on these mounts.
> > >> 
> > >> I want to kill the process and unmount.
> > >> 
> > >> umount -l will make the mount go a way, sort of.  But process is still hung. umount -f complains: umount2:  Device or resource busy umount.nfs: /mnt/foo:
> > >> device is busy
> > >> 
> > >> kill -9 does not work on process.
> > > 
> > > Kill -1 should work (since about 2.6.25 or so).
> > 
> > That is -[ONE], right?  Assuming so, it did not work for me.
> 
> No, it was "-9" .... sorry, I really shouldn't be let out without my proof
> reader.
> 
> However the 'stack' is sufficient to see what is going on.
> 
> The problem is that it is blocked inside the "VM" well away from NFS and
> there is no way for NFS to say "give up and go home".
> 
> I'd suggest that is a bug.   I cannot see any justification for fsync to not
> be killable.
> It wouldn't be too hard to create a patch to make it so.
> It would be a little harder to examine all call paths and create a
> convincing case that the patch was safe.
> It might be herculean task to convince others that it was the right thing
> to do.... so let's start with that one.
> 
> Hi Linux-mm and fs-devel people.  What do people think of making "fsync" and
> variants "KILLABLE" ??
> 
> I probably only need a little bit of encouragement to write a patch....
> 
> Thanks,
> NeilBrown
> 


It would be good to fix this in some fashion once and for all, and the
wait_on_page_writeback wait is a major source of pain for a lot of
people.

So to summarize...

The problem in a nutshell is that Ben has some cached writes to the
NFS server, but the server has gone away (presumably forever). The
question is -- how do we communicate to the kernel that that server
isn't coming back and that those dirty pages should be invalidated so
that we can umount the filesystem?

Allowing fsync/close to be killable sounds reasonable to me as at least
a partial solution. Both close(2) and fsync(2) are allowed to return
EINTR according to the POSIX spec. Allowing a kill -9 there seems
like it should be fine, and maybe we ought to even consider letting it
be susceptible to lesser signals.

That still leaves some open questions though...

Is that enough to fix it? You'd still have the dirty pages lingering
around, right? Would a umount -f presumably work at that point?

> > 
> > Kernel is 3.14.4+, with some of extra patches, but probably nothing that
> > influences this particular behaviour.
> > 
> > [root@lf1005-14010010 ~]# cat /proc/3805/stack
> > [<ffffffff811371ba>] sleep_on_page+0x9/0xd
> > [<ffffffff8113738e>] wait_on_page_bit+0x71/0x78
> > [<ffffffff8113769a>] filemap_fdatawait_range+0xa2/0x16d
> > [<ffffffff8113780e>] filemap_write_and_wait_range+0x3b/0x77
> > [<ffffffffa0f04734>] nfs_file_fsync+0x37/0x83 [nfs]
> > [<ffffffff811a8d32>] vfs_fsync_range+0x19/0x1b
> > [<ffffffff811a8d4b>] vfs_fsync+0x17/0x19
> > [<ffffffffa0f05305>] nfs_file_flush+0x6b/0x6f [nfs]
> > [<ffffffff81183e46>] filp_close+0x3f/0x71
> > [<ffffffff8119c8ae>] __close_fd+0x80/0x98
> > [<ffffffff81183de5>] SyS_close+0x1c/0x3e
> > [<ffffffff815c55f9>] system_call_fastpath+0x16/0x1b
> > [<ffffffffffffffff>] 0xffffffffffffffff
> > [root@lf1005-14010010 ~]# kill -1 3805
> > [root@lf1005-14010010 ~]# cat /proc/3805/stack
> > [<ffffffff811371ba>] sleep_on_page+0x9/0xd
> > [<ffffffff8113738e>] wait_on_page_bit+0x71/0x78
> > [<ffffffff8113769a>] filemap_fdatawait_range+0xa2/0x16d
> > [<ffffffff8113780e>] filemap_write_and_wait_range+0x3b/0x77
> > [<ffffffffa0f04734>] nfs_file_fsync+0x37/0x83 [nfs]
> > [<ffffffff811a8d32>] vfs_fsync_range+0x19/0x1b
> > [<ffffffff811a8d4b>] vfs_fsync+0x17/0x19
> > [<ffffffffa0f05305>] nfs_file_flush+0x6b/0x6f [nfs]
> > [<ffffffff81183e46>] filp_close+0x3f/0x71
> > [<ffffffff8119c8ae>] __close_fd+0x80/0x98
> > [<ffffffff81183de5>] SyS_close+0x1c/0x3e
> > [<ffffffff815c55f9>] system_call_fastpath+0x16/0x1b
> > [<ffffffffffffffff>] 0xffffffffffffffff
> > 
> > Thanks,
> > Ben
> > 
> > > If it doesn't please report the kernel version and cat /proc/$PID/stack
> > > 
> > > for some processes that cannot be killed.
> > > 
> > > NeilBrown
> > > 
> > >> 
> > >> 
> > >> Aside from bringing a fake NFS server back up on the same IP, is there any other way to get these mounts unmounted and the processes killed without 
> > >> rebooting?
> > >> 
> > >> Thanks, Ben
> > >> 
> > > 
> > 
> > 
> > - -- 
> > Ben Greear <greearb@candelatech.com>
> > Candela Technologies Inc  http://www.candelatech.com
> > 
> > -----BEGIN PGP SIGNATURE-----
> > Version: GnuPG v1.4.13 (GNU/Linux)
> > Comment: Using GnuPG with Thunderbird - http://www.enigmail.net/
> > 
> > iQEcBAEBAgAGBQJT2rLiAAoJELbHqkYeJT4OqPgH/0taKW6Be90c1mETZf9yeqZF
> > YMLZk8XC2wloEd9nVz//mXREmiu18Hc+5p7Upd4Os21J2P4PBMGV6P/9DMxxehwH
> > YX1HKha0EoAsbO5ILQhbLf83cRXAPEpvJPgYHrq6xjlKB8Q8OxxND37rY7kl19Zz
> > sdAw6GiqHICF3Hq1ATa/jvixMluDnhER9Dln3wOdAGzmmuFYqpTsV4EwzbKKqInJ
> > 6C15q+cq/9aYh6usN6z2qJhbHgqM9EWcPL6jOrCwX4PbC1XjKHekpFN0t9oKQClx
> > qSPuweMQ7fP4IBd2Ke8L/QlyOVblAKSE7t+NdrjfzLmYPzyHTyfLABR/BI053to=
> > =/9FJ
> > -----END PGP SIGNATURE-----
> 


-- 
Jeff Layton <jlayton@poochiereds.net>

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 819 bytes --]