[PATCH 2/2] nvme-multipath: don't block on blk_queue_enter of the underlying device

[Sagi Grimberg]

>> The deadlock in this patchset reproduces upstream. It is not possible to
>> update the kernel in the env in the original report.
>>
>> So IFF we assume that this does not reproduce in upstream (pending
>> proof), is there something that we can do with stable fixes? This will
>> probably go back to everything that is before 5.8...
> 
> The direct_make_request removal should be pretty easily backportable.

Umm, the direct_make_request removal is based on the submit_bio_noacct
rework you've done. Are you suggesting that we replace it with
generic_make_request for these kernels?

> In old kernels without the streamlined normal path it might cause minor
> performance degradations, but the actual change should be trivial.

That is better than getting stuck in I/O failover...