Re: [PATCH vfio 3/3] pds/vfio: Fix possible sleep while in atomic context

[Brett Creeley <bcreeley@amd.com>]

On 9/19/2023 11:59 AM, Jason Gunthorpe wrote:
> Caution: This message originated from an External Source. Use proper caution when opening attachments, clicking links, or responding.
> 
> 
> On Thu, Sep 14, 2023 at 04:38:37PM -0600, Alex Williamson wrote:
>> On Thu, 14 Sep 2023 12:15:40 -0700
>> Brett Creeley <brett.creeley@amd.com> wrote:
>>
>>> The driver could possibly sleep while in atomic context resulting
>>> in the following call trace while CONFIG_DEBUG_ATOMIC_SLEEP=y is
>>> set:
>>>
>>> [  227.229806] BUG: sleeping function called from invalid context at kernel/locking/mutex.c:283
>>> [  227.229818] in_atomic(): 1, irqs_disabled(): 0, non_block: 0, pid: 2817, name: bash
>>> [  227.229824] preempt_count: 1, expected: 0
>>> [  227.229827] RCU nest depth: 0, expected: 0
>>> [  227.229832] CPU: 5 PID: 2817 Comm: bash Tainted: G S         OE      6.6.0-rc1-next-20230911 #1
>>> [  227.229839] Hardware name: HPE ProLiant DL360 Gen10/ProLiant DL360 Gen10, BIOS U32 01/23/2021
>>> [  227.229843] Call Trace:
>>> [  227.229848]  <TASK>
>>> [  227.229853]  dump_stack_lvl+0x36/0x50
>>> [  227.229865]  __might_resched+0x123/0x170
>>> [  227.229877]  mutex_lock+0x1e/0x50
>>> [  227.229891]  pds_vfio_put_lm_file+0x1e/0xa0 [pds_vfio_pci]
>>> [  227.229909]  pds_vfio_put_save_file+0x19/0x30 [pds_vfio_pci]
>>> [  227.229923]  pds_vfio_state_mutex_unlock+0x2e/0x80 [pds_vfio_pci]
>>> [  227.229937]  pci_reset_function+0x4b/0x70
>>> [  227.229948]  reset_store+0x5b/0xa0
>>> [  227.229959]  kernfs_fop_write_iter+0x137/0x1d0
>>> [  227.229972]  vfs_write+0x2de/0x410
>>> [  227.229986]  ksys_write+0x5d/0xd0
>>> [  227.229996]  do_syscall_64+0x3b/0x90
>>> [  227.230004]  entry_SYSCALL_64_after_hwframe+0x6e/0xd8
>>> [  227.230017] RIP: 0033:0x7fb202b1fa28
>>> [  227.230023] Code: 89 02 48 c7 c0 ff ff ff ff eb b3 0f 1f 80 00 00 00 00 f3 0f 1e fa 48 8d 05 15 4d 2a 00 8b 00 85 c0 75 17 b8 01 00 00 00 0f 05 <48> 3d 00 f0 ff ff 77 58 c3 0f 1f 80 00 00 00 00 41 54 49 89 d4 55
>>> [  227.230028] RSP: 002b:00007fff6915fbd8 EFLAGS: 00000246 ORIG_RAX: 0000000000000001
>>> [  227.230036] RAX: ffffffffffffffda RBX: 0000000000000002 RCX: 00007fb202b1fa28
>>> [  227.230040] RDX: 0000000000000002 RSI: 000055f3834d5aa0 RDI: 0000000000000001
>>> [  227.230044] RBP: 000055f3834d5aa0 R08: 000000000000000a R09: 00007fb202b7fae0
>>> [  227.230047] R10: 000000000000000a R11: 0000000000000246 R12: 00007fb202dc06e0
>>> [  227.230050] R13: 0000000000000002 R14: 00007fb202dbb860 R15: 0000000000000002
>>> [  227.230056]  </TASK>
> 
> I usually encourage people to trim the oops, remove the time stamp at least.

Makes sense. I will remember that going forward. Thanks for the suggestion.

Brett

>>>
>>> This can happen if pds_vfio_put_restore_file() and/or
>>> pds_vfio_put_save_file() grab the mutex_lock(&lm_file->lock)
>>> while the spin_lock(&pds_vfio->reset_lock) is held, which can
>>> happen during while calling pds_vfio_state_mutex_unlock().
>>>
>>> Fix this by releasing the spin_unlock(&pds_vfio->reset_lock) before
>>> calling pds_vfio_put_restore_file() and pds_vfio_put_save_file() and
>>> re-acquiring spin_lock(&pds_vfio->reset_lock) after the previously
>>> mentioned functions are called to protect setting the subsequent
>>> state/deferred reset settings.
>>>
>>> The only possible concerns are other threads that may call
>>> pds_vfio_put_restore_file() and/or pds_vfio_put_save_file(). However,
>>> those paths are already protected by the state mutex_lock().
>>
>> Is there another viable solution to change reset_lock to a mutex?
>>
>> I think this is the origin of this algorithm:
>>
>> https://lore.kernel.org/all/20211019191025.GA4072278@nvidia.com/
>>
>> But it's not clear to me why Jason chose an example with a spinlock and
>> if some subtlety here requires it.  Thanks,
> 
> I think there was no specific reason it must be a spinlock
> 
> Certainly I'm not feeling comfortable just unlocking and relocking
> like that. It would need a big explanation why it is safe in a
> comment.

This follows the example in mlx5vf_state_mutex_unlock(), which releases 
the spinlock before calling mlx5vf_disable_fds().

However, there is a small difference where 
pds_vfio->deferred_reset_state could change in the window where the 
reset_lock isn't held. It seems this can be fixed this by a local 
deferred_reset_state in pds_vfio_state_mutex_unlock() that I set before 
unlocking to clear the fds.

Thanks,

Brett
> 
> Jason