> On May 11, 2022, at 1:31 PM, Trond Myklebust <trondmy@xxxxxxxxxxxxxxx> wrote:
>
> On Wed, 2022-05-11 at 16:27 +0000, Chuck Lever III wrote:
>>
>>
>>> On May 11, 2022, at 12:09 PM, Trond Myklebust
>>> <trondmy@xxxxxxxxxxxxxxx> wrote:
>>>
>>> On Wed, 2022-05-11 at 14:57 +0000, Chuck Lever III wrote:
>>>>
>>>>
>>>>> On May 10, 2022, at 10:24 AM, Chuck Lever III
>>>>> <chuck.lever@xxxxxxxxxx> wrote:
>>>>>
>>>>>
>>>>>
>>>>>> On May 3, 2022, at 3:11 PM, dai.ngo@xxxxxxxxxx wrote:
>>>>>>
>>>>>> Hi,
>>>>>>
>>>>>> I just noticed there were couple of oops in my Oracle6 in
>>>>>> nfs4.dev network.
>>>>>> I'm not sure who ran which tests (be useful to know) that
>>>>>> caused
>>>>>> these oops.
>>>>>>
>>>>>> Here is the stack traces:
>>>>>>
>>>>>> [286123.154006] BUG: sleeping function called from invalid
>>>>>> context at kernel/locking/rwsem.c:1585
>>>>>> [286123.155126] in_atomic(): 1, irqs_disabled(): 0,
>>>>>> non_block: 0,
>>>>>> pid: 3983, name: nfsd
>>>>>> [286123.155872] preempt_count: 1, expected: 0
>>>>>> [286123.156443] RCU nest depth: 0, expected: 0
>>>>>> [286123.156771] 1 lock held by nfsd/3983:
>>>>>> [286123.156786] #0: ffff888006762520 (&clp->cl_lock){+.+.}-
>>>>>> {2:2}, at: nfsd4_release_lockowner+0x306/0x850 [nfsd]
>>>>>> [286123.156949] Preemption disabled at:
>>>>>> [286123.156961] [<0000000000000000>] 0x0
>>>>>> [286123.157520] CPU: 1 PID: 3983 Comm: nfsd Kdump: loaded Not
>>>>>> tainted 5.18.0-rc4+ #1
>>>>>> [286123.157539] Hardware name: innotek GmbH
>>>>>> VirtualBox/VirtualBox, BIOS VirtualBox 12/01/2006
>>>>>> [286123.157552] Call Trace:
>>>>>> [286123.157565] <TASK>
>>>>>> [286123.157581] dump_stack_lvl+0x57/0x7d
>>>>>> [286123.157609] __might_resched.cold+0x222/0x26b
>>>>>> [286123.157644] down_read_nested+0x68/0x420
>>>>>> [286123.157671] ? down_write_nested+0x130/0x130
>>>>>> [286123.157686] ? rwlock_bug.part.0+0x90/0x90
>>>>>> [286123.157705] ? rcu_read_lock_sched_held+0x81/0xb0
>>>>>> [286123.157749] xfs_file_fsync+0x3b9/0x820
>>>>>> [286123.157776] ? lock_downgrade+0x680/0x680
>>>>>> [286123.157798] ? xfs_filemap_pfn_mkwrite+0x10/0x10
>>>>>> [286123.157823] ? nfsd_file_put+0x100/0x100 [nfsd]
>>>>>> [286123.157921] nfsd_file_flush.isra.0+0x1b/0x220 [nfsd]
>>>>>> [286123.158007] nfsd_file_put+0x79/0x100 [nfsd]
>>>>>> [286123.158088] check_for_locks+0x152/0x200 [nfsd]
>>>>>> [286123.158191] nfsd4_release_lockowner+0x4cf/0x850 [nfsd]
>>>>>> [286123.158393] ? nfsd4_locku+0xd10/0xd10 [nfsd]
>>>>>> [286123.158488] ? rcu_read_lock_bh_held+0x90/0x90
>>>>>> [286123.158525] nfsd4_proc_compound+0xd15/0x25a0 [nfsd]
>>>>>> [286123.158699] nfsd_dispatch+0x4ed/0xc30 [nfsd]
>>>>>> [286123.158974] ? rcu_read_lock_bh_held+0x90/0x90
>>>>>> [286123.159010] svc_process_common+0xd8e/0x1b20 [sunrpc]
>>>>>> [286123.159043] ? svc_generic_rpcbind_set+0x450/0x450
>>>>>> [sunrpc]
>>>>>> [286123.159043] ? nfsd_svc+0xc50/0xc50 [nfsd]
>>>>>> [286123.159043] ? svc_sock_secure_port+0x27/0x40 [sunrpc]
>>>>>> [286123.159043] ? svc_recv+0x1100/0x2390 [sunrpc]
>>>>>> [286123.159043] svc_process+0x361/0x4f0 [sunrpc]
>>>>>> [286123.159043] nfsd+0x2d6/0x570 [nfsd]
>>>>>> [286123.159043] ? nfsd_shutdown_threads+0x2a0/0x2a0 [nfsd]
>>>>>> [286123.159043] kthread+0x29f/0x340
>>>>>> [286123.159043] ? kthread_complete_and_exit+0x20/0x20
>>>>>> [286123.159043] ret_from_fork+0x22/0x30
>>>>>> [286123.159043] </TASK>
>>>>>> [286123.187052] BUG: scheduling while atomic:
>>>>>> nfsd/3983/0x00000002
>>>>>> [286123.187551] INFO: lockdep is turned off.
>>>>>> [286123.187918] Modules linked in: nfsd auth_rpcgss nfs_acl
>>>>>> lockd
>>>>>> grace sunrpc
>>>>>> [286123.188527] Preemption disabled at:
>>>>>> [286123.188535] [<0000000000000000>] 0x0
>>>>>> [286123.189255] CPU: 1 PID: 3983 Comm: nfsd Kdump: loaded
>>>>>> Tainted: G W 5.18.0-rc4+ #1
>>>>>> [286123.190233] Hardware name: innotek GmbH
>>>>>> VirtualBox/VirtualBox, BIOS VirtualBox 12/01/2006
>>>>>> [286123.190910] Call Trace:
>>>>>> [286123.190910] <TASK>
>>>>>> [286123.190910] dump_stack_lvl+0x57/0x7d
>>>>>> [286123.190910] __schedule_bug.cold+0x133/0x143
>>>>>> [286123.190910] __schedule+0x16c9/0x20a0
>>>>>> [286123.190910] ? schedule_timeout+0x314/0x510
>>>>>> [286123.190910] ? __sched_text_start+0x8/0x8
>>>>>> [286123.190910] ? internal_add_timer+0xa4/0xe0
>>>>>> [286123.190910] schedule+0xd7/0x1f0
>>>>>> [286123.190910] schedule_timeout+0x319/0x510
>>>>>> [286123.190910] ? rcu_read_lock_held_common+0xe/0xa0
>>>>>> [286123.190910] ? usleep_range_state+0x150/0x150
>>>>>> [286123.190910] ? lock_acquire+0x331/0x490
>>>>>> [286123.190910] ? init_timer_on_stack_key+0x50/0x50
>>>>>> [286123.190910] ? do_raw_spin_lock+0x116/0x260
>>>>>> [286123.190910] ? rwlock_bug.part.0+0x90/0x90
>>>>>> [286123.190910] io_schedule_timeout+0x26/0x80
>>>>>> [286123.190910] wait_for_completion_io_timeout+0x16a/0x340
>>>>>> [286123.190910] ? rcu_read_lock_bh_held+0x90/0x90
>>>>>> [286123.190910] ? wait_for_completion+0x330/0x330
>>>>>> [286123.190910] submit_bio_wait+0x135/0x1d0
>>>>>> [286123.190910] ? submit_bio_wait_endio+0x40/0x40
>>>>>> [286123.190910] ? xfs_iunlock+0xd5/0x300
>>>>>> [286123.190910] ? bio_init+0x295/0x470
>>>>>> [286123.190910] blkdev_issue_flush+0x69/0x80
>>>>>> [286123.190910] ? blk_unregister_queue+0x1e0/0x1e0
>>>>>> [286123.190910] ? bio_kmalloc+0x90/0x90
>>>>>> [286123.190910] ? xfs_iunlock+0x1b4/0x300
>>>>>> [286123.190910] xfs_file_fsync+0x354/0x820
>>>>>> [286123.190910] ? lock_downgrade+0x680/0x680
>>>>>> [286123.190910] ? xfs_filemap_pfn_mkwrite+0x10/0x10
>>>>>> [286123.190910] ? nfsd_file_put+0x100/0x100 [nfsd]
>>>>>> [286123.190910] nfsd_file_flush.isra.0+0x1b/0x220 [nfsd]
>>>>>> [286123.190910] nfsd_file_put+0x79/0x100 [nfsd]
>>>>>> [286123.190910] check_for_locks+0x152/0x200 [nfsd]
>>>>>> [286123.190910] nfsd4_release_lockowner+0x4cf/0x850 [nfsd]
>>>>>> [286123.190910] ? nfsd4_locku+0xd10/0xd10 [nfsd]
>>>>>> [286123.190910] ? rcu_read_lock_bh_held+0x90/0x90
>>>>>> [286123.190910] nfsd4_proc_compound+0xd15/0x25a0 [nfsd]
>>>>>> [286123.190910] nfsd_dispatch+0x4ed/0xc30 [nfsd]
>>>>>> [286123.190910] ? rcu_read_lock_bh_held+0x90/0x90
>>>>>> [286123.190910] svc_process_common+0xd8e/0x1b20 [sunrpc]
>>>>>> [286123.190910] ? svc_generic_rpcbind_set+0x450/0x450
>>>>>> [sunrpc]
>>>>>> [286123.190910] ? nfsd_svc+0xc50/0xc50 [nfsd]
>>>>>> [286123.190910] ? svc_sock_secure_port+0x27/0x40 [sunrpc]
>>>>>> [286123.190910] ? svc_recv+0x1100/0x2390 [sunrpc]
>>>>>> [286123.190910] svc_process+0x361/0x4f0 [sunrpc]
>>>>>> [286123.190910] nfsd+0x2d6/0x570 [nfsd]
>>>>>> [286123.190910] ? nfsd_shutdown_threads+0x2a0/0x2a0 [nfsd]
>>>>>> [286123.190910] kthread+0x29f/0x340
>>>>>> [286123.190910] ? kthread_complete_and_exit+0x20/0x20
>>>>>> [286123.190910] ret_from_fork+0x22/0x30
>>>>>> [286123.190910] </TASK>
>>>>>>
>>>>>> The problem is the process tries to sleep while holding the
>>>>>> cl_lock by nfsd4_release_lockowner. I think the problem was
>>>>>> introduced with the filemap_flush in nfsd_file_put since
>>>>>> 'b6669305d35a nfsd: Reduce the number of calls to
>>>>>> nfsd_file_gc()'.
>>>>>> The filemap_flush is later replaced by nfsd_file_flush by
>>>>>> '6b8a94332ee4f nfsd: Fix a write performance regression'.
>>>>>
>>>>> That seems plausible, given the traces above.
>>>>>
>>>>> But it begs the question: why was a vfs_fsync() needed by
>>>>> RELEASE_LOCKOWNER in this case? I've tried to reproduce the
>>>>> problem, and even added a might_sleep() call in
>>>>> nfsd_file_flush()
>>>>> but haven't been able to reproduce.
>>>>
>>>> Trond, I'm assuming you switched to a synchronous flush here to
>>>> capture writeback errors. There's no other requirement for
>>>> waiting for the flush to complete, right?
>>>
>>> It's because the file is unhashed, so it is about to be closed and
>>> garbage collected as soon as the refcount goes to zero.
>>>
>>>>
>>>> To enable nfsd_file_put() to be invoked in atomic contexts again,
>>>> would the following be a reasonable short term fix:
>>>>
>>>> diff --git a/fs/nfsd/filecache.c b/fs/nfsd/filecache.c
>>>> index 2c1b027774d4..96c8d07788f4 100644
>>>> --- a/fs/nfsd/filecache.c
>>>> +++ b/fs/nfsd/filecache.c
>>>> @@ -304,7 +304,7 @@ nfsd_file_put(struct nfsd_file *nf)
>>>> {
>>>> set_bit(NFSD_FILE_REFERENCED, &nf->nf_flags);
>>>> if (test_bit(NFSD_FILE_HASHED, &nf->nf_flags) == 0) {
>>>> - nfsd_file_flush(nf);
>>>> + filemap_flush(nf->nf_file->f_mapping);
>>>> nfsd_file_put_noref(nf);
>>>> } else {
>>>> nfsd_file_put_noref(nf);
>>>>
>>>>
>>>
>>> filemap_flush() sleeps, and so does nfsd_file_put_noref() (it can
>>> call
>>> filp_close() + fput()), so this kind of change isn't going to work
>>> no
>>> matter how you massage it.
>>
>> Er. Wouldn't that mean we would have seen "sleep while spinlock is
>> held" BUGs since nfsd4_release_lockowner() was added? It's done
>> at least an fput() while holding clp->cl_lock since it was added,
>> I think.
>
>
> There is nothing magical about using WB_SYNC_NONE as far as the
> writeback code is concerned. write_cache_pages() will still happily
> call lock_page() and sleep on that lock if it is contended. The
> writepage/writepages code will happily try to allocate memory if
> necessary.
>
> The only difference is that it won't sleep waiting for the PG_writeback
> bit.
>
> So, no, you can't safely call filemap_flush() from a spin locked
> context, and
> yes, the bug has been there from day 1. It was not introduced by me.
>
> Also, as I said, nfsd_file_put_noref() is not safe to call from a spin
> locked context. Again, this was not introduced any time recently.
OK. I'm trying to figure out how bad the problem is and how
far back to apply the fix.
I agree that the arrangement of the code path means
nfsd4_release_lockowner() has always called fput() or
filemap_flush() while cl_lock was held.
But again, I'm not aware of recent instances of this particular
splat. So I'm wondering if a recent change has made this issue
easier to hit. We might not be able to answer that until we
find out how the bake-a-thon testers managed to trigger the
issue on Dai's server.
>>> Is there any reason why you need a reference to the nfs_file there?
>>> Wouldn't holding the fp->fi_lock be sufficient, since you're
>>> already in
>>> an atomic context? As long as one of the entries in fp->fi_fds[] is
>>> non-zero then you should be safe.
>>
>> Sure, check_for_locks() seems to be the only problematic caller.
>> Other callers appear to be careful to call nfsd_file_put() only
>> after releasing spinlocks.
>>
>> I would like a fix that can be backported without fuss. I was
>> thinking changing check_for_locks() might get a little too
>> hairy for that, but I'll check it out.
Notably: check_for_locks() needs to drop fi_lock before taking
flc_lock because the OPEN path can take flc_lock first, then
fi_lock, via nfsd_break_deleg_cb(). Holding a reference to the
nfsd_file guarantees that the inode won't go away while
check_for_locks() examines the flc_posix list without holding
fi_lock.
So my first take on this was we need nfsd4_release_lockowner()
to drop cl_lock before check_for_locks() is called.
--
Chuck Lever
