[PATCH RFC] nvme: avoid deadlock warning in sync_io_queues() path

[Sagi Grimberg]

>> The commit c0feea594e05 ("workqueue: don't skip lockdep work dependency
>> in cancel_work_sync()") restored a lockdep check and it triggered
>> deadlock warning in sync_io_queues() path. This function locks
>> namespaces_rwsem for read to traverse the namespace list and calls
>> blk_sync_queue() for each namespace. blk_sync_queue() calls
>> cancel_work_sync() to wait for nvme_timeout() completion which has paths
>> to lock the namespaces_rwsem again for read. These two locks are both
>> for read, but it can cause deadlock when another task has a lock for
>> write. Hence, the deadlock warning was reported.
>>
>> To avoid the deadlock warning, guard the namespace list by SRCU in place
>> of the namespaces_rwsem. This avoids the deadlock by the two read locks
>> since writes by other tasks do not stop the reads.
> 
> Based on the report this is triggered from the pci shutdown code path?
> I think the problem is that there is a dev disable running together
> with a reset (driven from the timeout handler). Perhaps this should
> be addressed.
> 
> I don't think that making the namespaces lock an srcu helps us here.

Just read the correspondence on the report. If this is just lockdep
false positive, then I don't think we need to change the namespaces
rwsem.

Can't there be some annotation that we can do to teach lockdep that
this one is a false positive?