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

[Christoph Hellwig]

On Tue, Mar 23, 2021 at 11:53:30PM +0900, Keith Busch wrote:
> Failover might be overkill. We can run out of tags in a perfectly normal
> situation, and simply waiting may be the best option, or even scheduling
> on a different CPU may be sufficient to get a viable tag  rather than
> selecting a different path.

Indeed.  Then again IFF there are multiple optimized paths picking one
could also be a way to make progress.

> Does it make sense to just abort all allocated tags during a reset and
> let the original bio requeue for multipath IO?

Well, this would again hard code multipath information into the lower
levels.  But then again we could at least do it so that we check for
the REQ_NVME_MPATH to make it clear what is going on.