Re: nfsd blocks indefinitely in nfsd4_destroy_session

[Chuck Lever <chuck.lever@oracle.com>]

On 1/9/25 7:42 AM, Jeff Layton wrote:
> On Thu, 2025-01-09 at 12:56 +0100, Christian Herzog wrote:
>> Dear Chuck,
>>
>> On Wed, Jan 08, 2025 at 10:07:49AM -0500, Chuck Lever wrote:
>>> On 1/8/25 9:54 AM, Christian Herzog wrote:
>>>> Dear Chuck,
>>>>
>>>> On Wed, Jan 08, 2025 at 08:33:23AM -0500, Chuck Lever wrote:
>>>>> On 1/7/25 4:17 PM, Salvatore Bonaccorso wrote:
>>>>>> Hi Chuck,
>>>>>>
>>>>>> Thanks for your time on this, much appreciated.
>>>>>>
>>>>>> On Wed, Jan 01, 2025 at 02:24:50PM -0500, Chuck Lever wrote:
>>>>>>> On 12/25/24 4:15 AM, Salvatore Bonaccorso wrote:
>>>>>>>> Hi Chuck, hi all,
>>>>>>>>
>>>>>>>> [it was not ideal to pick one of the message for this followup, let me
>>>>>>>> know if you want a complete new thread, adding as well Benjamin and
>>>>>>>> Trond as they are involved in one mentioned patch]
>>>>>>>>
>>>>>>>> On Mon, Jun 17, 2024 at 02:31:54PM +0000, Chuck Lever III wrote:
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>> On Jun 17, 2024, at 2:55 AM, Harald Dunkel <harald.dunkel@aixigo.com> wrote:
>>>>>>>>>>
>>>>>>>>>> Hi folks,
>>>>>>>>>>
>>>>>>>>>> what would be the reason for nfsd getting stuck somehow and becoming
>>>>>>>>>> an unkillable process? See
>>>>>>>>>>
>>>>>>>>>> - https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=1071562
>>>>>>>>>> - https://bugs.launchpad.net/ubuntu/+source/nfs-utils/+bug/2062568
>>>>>>>>>>
>>>>>>>>>> Doesn't this mean that something inside the kernel gets stuck as
>>>>>>>>>> well? Seems odd to me.
>>>>>>>>>
>>>>>>>>> I'm not familiar with the Debian or Ubuntu kernel packages. Can
>>>>>>>>> the kernel release numbers be translated to LTS kernel releases
>>>>>>>>> please? Need both "last known working" and "first broken" releases.
>>>>>>>>>
>>>>>>>>> This:
>>>>>>>>>
>>>>>>>>> [ 6596.911785] RPC: Could not send backchannel reply error: -110
>>>>>>>>> [ 6596.972490] RPC: Could not send backchannel reply error: -110
>>>>>>>>> [ 6837.281307] RPC: Could not send backchannel reply error: -110
>>>>>>>>>
>>>>>>>>> is a known set of client backchannel bugs. Knowing the LTS kernel
>>>>>>>>> releases (see above) will help us figure out what needs to be
>>>>>>>>> backported to the LTS kernels kernels in question.
>>>>>>>>>
>>>>>>>>> This:
>>>>>>>>>
>>>>>>>>> [11183.290619] wait_for_completion+0x88/0x150
>>>>>>>>> [11183.290623] __flush_workqueue+0x140/0x3e0
>>>>>>>>> [11183.290629] nfsd4_probe_callback_sync+0x1a/0x30 [nfsd]
>>>>>>>>> [11183.290689] nfsd4_destroy_session+0x186/0x260 [nfsd]
>>>>>>>>>
>>>>>>>>> is probably related to the backchannel errors on the client, but
>>>>>>>>> client bugs shouldn't cause the server to hang like this. We
>>>>>>>>> might be able to say more if you can provide the kernel release
>>>>>>>>> translations (see above).
>>>>>>>>
>>>>>>>> In Debian we hstill have the bug #1071562 open and one person notified
>>>>>>>> mye offlist that it appears that the issue get more frequent since
>>>>>>>> they updated on NFS client side from Ubuntu 20.04 to Debian bookworm
>>>>>>>> with a 6.1.y based kernel).
>>>>>>>>
>>>>>>>> Some people around those issues, seem to claim that the change
>>>>>>>> mentioned in
>>>>>>>> https://lists.proxmox.com/pipermail/pve-devel/2024-July/064614.html
>>>>>>>> would fix the issue, which is as well backchannel related.
>>>>>>>>
>>>>>>>> This is upstream: 6ddc9deacc13 ("SUNRPC: Fix backchannel reply,
>>>>>>>> again"). While this commit fixes 57331a59ac0d ("NFSv4.1: Use the
>>>>>>>> nfs_client's rpc timeouts for backchannel") this is not something
>>>>>>>> which goes back to 6.1.y, could it be possible that hte backchannel
>>>>>>>> refactoring and this final fix indeeds fixes the issue?
>>>>>>>>
>>>>>>>> As people report it is not easily reproducible, so this makes it
>>>>>>>> harder to identify fixes correctly.
>>>>>>>>
>>>>>>>> I gave a (short) stance on trying to backport commits up to
>>>>>>>> 6ddc9deacc13 ("SUNRPC: Fix backchannel reply, again") but this quickly
>>>>>>>> seems to indicate it is probably still not the right thing for
>>>>>>>> backporting to the older stable series.
>>>>>>>>
>>>>>>>> As at least pre-requisites:
>>>>>>>>
>>>>>>>> 2009e32997ed568a305cf9bc7bf27d22e0f6ccda
>>>>>>>> 4119bd0306652776cb0b7caa3aea5b2a93aecb89
>>>>>>>> 163cdfca341b76c958567ae0966bd3575c5c6192
>>>>>>>> f4afc8fead386c81fda2593ad6162271d26667f8
>>>>>>>> 6ed8cdf967f7e9fc96cd1c129719ef99db2f9afc
>>>>>>>> 57331a59ac0d680f606403eb24edd3c35aecba31
>>>>>>>>
>>>>>>>> and still there would be conflicting codepaths (and does not seem
>>>>>>>> right).
>>>>>>>>
>>>>>>>> Chuck, Benjamin, Trond, is there anything we can provive on reporters
>>>>>>>> side that we can try to tackle this issue better?
>>>>>>>
>>>>>>> As I've indicated before, NFSD should not block no matter what the
>>>>>>> client may or may not be doing. I'd like to focus on the server first.
>>>>>>>
>>>>>>> What is the result of:
>>>>>>>
>>>>>>> $ cd <Bookworm's v6.1.90 kernel source >
>>>>>>> $ unset KBUILD_OUTPUT
>>>>>>> $ make -j `nproc`
>>>>>>> $ scripts/faddr2line \
>>>>>>> 	fs/nfsd/nfs4state.o \
>>>>>>> 	nfsd4_destroy_session+0x16d
>>>>>>>
>>>>>>> Since this issue appeared after v6.1.1, is it possible to bisect
>>>>>>> between v6.1.1 and v6.1.90 ?
>>>>>>
>>>>>> First please note, at least speaking of triggering the issue in
>>>>>> Debian, Debian has moved to 6.1.119 based kernel already (and soon in
>>>>>> the weekends point release move to 6.1.123).
>>>>>>
>>>>>> That said, one of the users which regularly seems to be hit by the
>>>>>> issue was able to provide the above requested information, based for
>>>>>> 6.1.119:
>>>>>>
>>>>>> ~/kernel/linux-source-6.1# make kernelversion
>>>>>> 6.1.119
>>>>>> ~/kernel/linux-source-6.1# scripts/faddr2line fs/nfsd/nfs4state.o nfsd4_destroy_session+0x16d
>>>>>> nfsd4_destroy_session+0x16d/0x250:
>>>>>> __list_del_entry at /root/kernel/linux-source-6.1/./include/linux/list.h:134
>>>>>> (inlined by) list_del at /root/kernel/linux-source-6.1/./include/linux/list.h:148
>>>>>> (inlined by) unhash_session at /root/kernel/linux-source-6.1/fs/nfsd/nfs4state.c:2062
>>>>>> (inlined by) nfsd4_destroy_session at /root/kernel/linux-source-6.1/fs/nfsd/nfs4state.c:3856
>>>>>>
>>>>>> They could provide as well a decode_stacktrace output for the recent
>>>>>> hit (if that is helpful for you):
>>>>>>
>>>>>> [Mon Jan 6 13:25:28 2025] INFO: task nfsd:55306 blocked for more than 6883 seconds.
>>>>>> [Mon Jan 6 13:25:28 2025]       Not tainted 6.1.0-28-amd64 #1 Debian 6.1.119-1
>>>>>> [Mon Jan 6 13:25:28 2025] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
>>>>>> [Mon Jan 6 13:25:28 2025] task:nfsd            state:D stack:0     pid:55306 ppid:2      flags:0x00004000
>>>>>> [Mon Jan 6 13:25:28 2025] Call Trace:
>>>>>> [Mon Jan 6 13:25:28 2025]  <TASK>
>>>>>> [Mon Jan 6 13:25:28 2025] __schedule+0x34d/0x9e0
>>>>>> [Mon Jan 6 13:25:28 2025] schedule+0x5a/0xd0
>>>>>> [Mon Jan 6 13:25:28 2025] schedule_timeout+0x118/0x150
>>>>>> [Mon Jan 6 13:25:28 2025] wait_for_completion+0x86/0x160
>>>>>> [Mon Jan 6 13:25:28 2025] __flush_workqueue+0x152/0x420
>>>>>> [Mon Jan 6 13:25:28 2025] nfsd4_destroy_session (debian/build/build_amd64_none_amd64/include/linux/spinlock.h:351 debian/build/build_amd64_none_amd64/fs/nfsd/nfs4state.c:3861) nfsd
>>>>>> [Mon Jan 6 13:25:28 2025] nfsd4_proc_compound (debian/build/build_amd64_none_amd64/fs/nfsd/nfs4proc.c:2680) nfsd
>>>>>> [Mon Jan 6 13:25:28 2025] nfsd_dispatch (debian/build/build_amd64_none_amd64/fs/nfsd/nfssvc.c:1022) nfsd
>>>>>> [Mon Jan 6 13:25:28 2025] svc_process_common (debian/build/build_amd64_none_amd64/net/sunrpc/svc.c:1344) sunrpc
>>>>>> [Mon Jan 6 13:25:28 2025] ? svc_recv (debian/build/build_amd64_none_amd64/net/sunrpc/svc_xprt.c:897) sunrpc
>>>>>> [Mon Jan 6 13:25:28 2025] ? nfsd_svc (debian/build/build_amd64_none_amd64/fs/nfsd/nfssvc.c:983) nfsd
>>>>>> [Mon Jan 6 13:25:28 2025] ? nfsd_inet6addr_event (debian/build/build_amd64_none_amd64/fs/nfsd/nfssvc.c:922) nfsd
>>>>>> [Mon Jan 6 13:25:28 2025] svc_process (debian/build/build_amd64_none_amd64/net/sunrpc/svc.c:1474) sunrpc
>>>>>> [Mon Jan 6 13:25:28 2025] nfsd (debian/build/build_amd64_none_amd64/fs/nfsd/nfssvc.c:960) nfsd
>>>>>> [Mon Jan 6 13:25:28 2025] kthread+0xd7/0x100
>>>>>> [Mon Jan 6 13:25:28 2025] ? kthread_complete_and_exit+0x20/0x20
>>>>>> [Mon Jan 6 13:25:28 2025] ret_from_fork+0x1f/0x30
>>>>>> [Mon Jan  6 13:25:28 2025]  </TASK>
>>>>>>
>>>>>> The question about bisection is actually harder, those are production
>>>>>> systems and I understand it's not possible to do bisect there, while
>>>>>> unfortunately not reprodcing the issue on "lab conditions".
>>>>>>
>>>>>> Does the above give us still a clue on what you were looking for?
>>>>>
>>>>> Thanks for the update.
>>>>>
>>>>> It's possible that 38f080f3cd19 ("NFSD: Move callback_wq into struct
>>>>> nfs4_client"), while not a specific fix for this issue, might offer some
>>>>> relief by preventing the DESTROY_SESSION hang from affecting all other
>>>>> clients of the degraded server.
>>>>>
>>>>> Not having a reproducer or the ability to bisect puts a damper on
>>>>> things. The next step, then, is to enable tracing on servers where this
>>>>> issue can come up, and wait for the hang to occur. The following command
>>>>> captures information only about callback operation, so it should not
>>>>> generate much data or impact server performance.
>>>>>
>>>>>     # trace-cmd record -e nfsd:nfsd_cb\*
>>>>>
>>>>> Let that run until the problem occurs, then ^C, compress the resulting
>>>>> trace.dat file, and either attach it to 1071562 or email it to me
>>>>> privately.
>>>> thanks for the follow-up.
>>>>
>>>> I am the "customer" with two affected file servers. We have since moved those
>>>> servers to the backports kernel (6.11.10) in the hope of forward fixing the
>>>> issue. If this kernel is stable, I'm afraid I can't go back to the 'bad'
>>>> kernel (700+ researchers affected..), and we're also not able to trigger the
>>>> issue on our test environment.
>>>
>>> Hello Dr. Herzog -
>>>
>>> If the problem recurs on 6.11, the trace-cmd I suggest above works
>>> there as well.
>> the bad news is: it just happened again with the bpo kernel.
>>
>> We then immediately started trace-cmd since usually there are several call
>> traces in a row and we did get a trace.dat:
>> http://people.phys.ethz.ch/~daduke/trace.dat
>>
>> we did notice however that the stack trace looked a bit different this time:
>>
>>      INFO: task nfsd:2566 blocked for more than 5799 seconds.
>>      Tainted: G        W          6.11.10+bpo-amd64 #1 Debia>
>>      "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables t>
>>      task:nfsd            state:D stack:0     pid:2566  tgid:2566 >
>>      Call Trace:
>>      <TASK>
>>      __schedule+0x400/0xad0
>>      schedule+0x27/0xf0
>>      nfsd4_shutdown_callback+0xfe/0x140 [nfsd]
>>      ? __pfx_var_wake_function+0x10/0x10
>>      __destroy_client+0x1f0/0x290 [nfsd]
>>      nfsd4_destroy_clientid+0xf1/0x1e0 [nfsd]
>>      ? svcauth_unix_set_client+0x586/0x5f0 [sunrpc]
>>      nfsd4_proc_compound+0x34d/0x670 [nfsd]
>>      nfsd_dispatch+0xfd/0x220 [nfsd]
>>      svc_process_common+0x2f7/0x700 [sunrpc]
>>      ? __pfx_nfsd_dispatch+0x10/0x10 [nfsd]
>>      svc_process+0x131/0x180 [sunrpc]
>>      svc_recv+0x830/0xa10 [sunrpc]
>>      ? __pfx_nfsd+0x10/0x10 [nfsd]
>>      nfsd+0x87/0xf0 [nfsd]
>>      kthread+0xcf/0x100
>>      ? __pfx_kthread+0x10/0x10
>>      ret_from_fork+0x31/0x50
>>      ? __pfx_kthread+0x10/0x10
>>      ret_from_fork_asm+0x1a/0x30
>>      </TASK>
>>
>> and also the state of the offending client in `/proc/fs/nfsd/clients/*/info`
>> used to be callback state: UNKNOWN while now it is DOWN or FAULT. No idea
>> what it means, but I thought it was worth mentioning.
>>
> 
> Looks like this is hung in nfsd41_cb_inflight_wait_complete() ? That
> probably means that the cl_cb_inflight counter is stuck at >0. I'm
> guessing that means that there is some callback that it's expecting to
> complete that isn't. From nfsd4_shutdown_callback():
> 
>          /*
>           * Note this won't actually result in a null callback;
>           * instead, nfsd4_run_cb_null() will detect the killed
>           * client, destroy the rpc client, and stop:
>           */
>          nfsd4_run_cb(&clp->cl_cb_null);
>          flush_workqueue(clp->cl_callback_wq);
>          nfsd41_cb_inflight_wait_complete(clp);
> 
> ...it sounds like that isn't happening properly though.
> 
> It might be interesting to see if you can track down the callback
> client in /sys/kernel/debug/sunrpc and see what it's doing.

Could be that the RPC client was shutdown while there were outstanding
callbacks, and that means that counter can now never go to zero?



-- 
Chuck Lever