Re: [PATCH] nfsd: decrease cl_cb_inflight if fail to queue cb_work

[Chuck Lever]

On 2/18/25 10:51 AM, Jeff Layton wrote:
> On Tue, 2025-02-18 at 10:02 -0500, Chuck Lever wrote:
>> On 2/18/25 9:40 AM, Jeff Layton wrote:
>>> On Tue, 2025-02-18 at 09:31 -0500, Chuck Lever wrote:
>>>> On 2/18/25 9:29 AM, Jeff Layton wrote:
>>>>> On Tue, 2025-02-18 at 08:58 -0500, Jeff Layton wrote:
>>>>>> On Tue, 2025-02-18 at 21:54 +0800, Li Lingfeng wrote:
>>>>>>> In nfsd4_run_cb, cl_cb_inflight is increased before attempting to queue
>>>>>>> cb_work to callback_wq. This count can be decreased in three situations:
>>>>>>> 1) If queuing fails in nfsd4_run_cb, the count will be decremented
>>>>>>> accordingly.
>>>>>>> 2) After cb_work is running, the count is decreased in the exception
>>>>>>> branch of nfsd4_run_cb_work via nfsd41_destroy_cb.
>>>>>>> 3) The count is decreased in the release callback of rpc_task — either
>>>>>>> directly calling nfsd41_cb_inflight_end in nfsd4_cb_probe_release, or
>>>>>>> calling nfsd41_destroy_cb in 	.
>>>>>>>
>>>>>>> However, in nfsd4_cb_release, if the current cb_work needs to restart, the
>>>>>>> count will not be decreased, with the expectation that it will be reduced
>>>>>>> once cb_work is running.
>>>>>>> If queuing fails here, then the count will leak, ultimately causing the
>>>>>>> nfsd service to be unable to exit as shown below:
>>>>>>> [root@nfs_test2 ~]# cat /proc/2271/stack
>>>>>>> [<0>] nfsd4_shutdown_callback+0x22b/0x290
>>>>>>> [<0>] __destroy_client+0x3cd/0x5c0
>>>>>>> [<0>] nfs4_state_destroy_net+0xd2/0x330
>>>>>>> [<0>] nfs4_state_shutdown_net+0x2ad/0x410
>>>>>>> [<0>] nfsd_shutdown_net+0xb7/0x250
>>>>>>> [<0>] nfsd_last_thread+0x15f/0x2a0
>>>>>>> [<0>] nfsd_svc+0x388/0x3f0
>>>>>>> [<0>] write_threads+0x17e/0x2b0
>>>>>>> [<0>] nfsctl_transaction_write+0x91/0xf0
>>>>>>> [<0>] vfs_write+0x1c4/0x750
>>>>>>> [<0>] ksys_write+0xcb/0x170
>>>>>>> [<0>] do_syscall_64+0x70/0x120
>>>>>>> [<0>] entry_SYSCALL_64_after_hwframe+0x78/0xe2
>>>>>>> [root@nfs_test2 ~]#
>>>>>>>
>>>>>>> Fix this by decreasing cl_cb_inflight if the restart fails.
>>>>>>>
>>>>>>> Fixes: cba5f62b1830 ("nfsd: fix callback restarts")
>>>>>>> Signed-off-by: Li Lingfeng <lilingfeng3@xxxxxxxxxx>
>>>>>>> ---
>>>>>>>  fs/nfsd/nfs4callback.c | 10 +++++++---
>>>>>>>  1 file changed, 7 insertions(+), 3 deletions(-)
>>>>>>>
>>>>>>> diff --git a/fs/nfsd/nfs4callback.c b/fs/nfsd/nfs4callback.c
>>>>>>> index 484077200c5d..8a7d24efdd08 100644
>>>>>>> --- a/fs/nfsd/nfs4callback.c
>>>>>>> +++ b/fs/nfsd/nfs4callback.c
>>>>>>> @@ -1459,12 +1459,16 @@ static void nfsd4_cb_done(struct rpc_task *task, void *calldata)
>>>>>>>  static void nfsd4_cb_release(void *calldata)
>>>>>>>  {
>>>>>>>  	struct nfsd4_callback *cb = calldata;
>>>>>>> +	struct nfs4_client *clp = cb->cb_clp;
>>>>>>> +	int queued;
>>>>>>>  
>>>>>>>  	trace_nfsd_cb_rpc_release(cb->cb_clp);
>>>>>>>  
>>>>>>> -	if (cb->cb_need_restart)
>>>>>>> -		nfsd4_queue_cb(cb);
>>>>>>> -	else
>>>>>>> +	if (cb->cb_need_restart) {
>>>>>>> +		queued = nfsd4_queue_cb(cb);
>>>>>>> +		if (!queued)
>>>>>>> +			nfsd41_cb_inflight_end(clp);
>>>>>>> +	} else
>>>>>>>  		nfsd41_destroy_cb(cb);
>>>>>>>  
>>>>>>>  }
>>>>>>
>>>>>> Good catch!
>>>>>>
>>>>>> Reviewed-by: Jeff Layton <jlayton@xxxxxxxxxx>
>>>>>>
>>>>>
>>>>> Actually, I think this is not quite right. It's a bit more subtle than
>>>>> it first appears. The problem of course is that the callback workqueue
>>>>> jobs run in a different task than the RPC workqueue jobs, so they can
>>>>> race.
>>>>>
>>>>> cl_cb_inflight gets bumped when the callback is first queued, and only
>>>>> gets released in nfsd41_destroy_cb(). If it fails to be queued, it's
>>>>> because something else has queued the workqueue job in the meantime.
>>>>>
>>>>> There are two places that can occur: nfsd4_cb_release() and
>>>>> nfsd4_run_cb(). Since this is occurring in nfsd4_cb_release(), the only
>>>>> other option is that something raced in and queued it via
>>>>> nfsd4_run_cb().
>>>>
>>>> What would be the "something" that raced in?
>>>>
>>>
>>> I think we may be able to get there via multiple __break_lease() calls
>>> on the same layout or delegation. That could mean multiple calls to the
>>> ->lm_break operation on the same object.
>>
>> Makes sense.
>>
>> Out of curiosity, what would be the complexity/performance cost of
>> serializing the lm_break calls, or having each lm_break call register
>> an interest in the CB_RECALL callback? Maybe not worth it.
>>
> 
> I'd probably resist trying to put that logic in generic code.

Hrm, I consider the callback code itself as generic, since all callback
operations use it. It makes sense for example to specifically constrain
the number of CB_RECALLs on a single inode to one.


> I think
> for now I'd rather just add this logic in the nfsd4_callback handling.

Fair enough. Let's see a specific change proposal before letting our
imaginations run wild. ;-)


> Longer term, I'm wondering if we really need the callback workqueue at
> all. It basically exists to just queue async RPC jobs (which are
> themselves workqueue tasks). Maybe it'd be simpler to just queue the
> RPC tasks directly at the points where we call nfsd4_run_cb() today?

The workqueue is single-threaded, so it serializes access to various
data structures, and prevents multiple concurrent attempts to start or
shutdown a backchannel. The WQ can be removed, but only after
alternative serialization has been added. Not impossible, but not a
dev priority either.


> That's a bit more of an overhaul though, particularly in the callback
> channel probing codepath.
> 
>>
>>>>> That will have incremented cl_cb_inflight() an extra
>>>>> time and so your patch will make sense for that.
>>>>>
>>>>> Unfortunately, the slot may leak in that case if nothing released it
>>>>> earlier. I think this probably needs to call nfsd41_destroy_cb() if the
>>>>> job can't be queued. That might, however, race with the callback
>>>>> workqueue job running.
>>>>>
>>>>> I think we might need to consider adding some sort of "this callback is
>>>>> running" atomic flag: do a test_and_set on the flag in nfsd4_run_cb()
>>>>> and only queue the workqueue job if that comes back false. Then, we can
>>>>> clear the bit in nfsd41_destroy_cb().
>>>>>
>>>>> That should ensure that you never fail to queue the workqueue job from
>>>>> nfsd4_cb_release().
>>>>>
>>>>> Thoughts?
>>>>
>>>>
>>>
>>
>>
> 


-- 
Chuck Lever