Re: [for-4.14 RFC PATCH 1/2] dm rq: avoid deadlock if dm-mq is stacked on old .request_fn device(s)

[Christoph Hellwig <hch@lst.de>]

The problem here is the following:

blk_finish_request must always be called with the queue lock held,
it even has an assert.

Without blk-mq used by dm-rq, dm uses the block softirq to execute the
completion, which means we always have a different execution context and
can take the queue lock again without issuesi.

With blk-mq used by dm-rq, the the dm .complete handler that is the rough
equivalent of the softirq handler is called either directly if were are
on the same CPU, or using a IPI (hardirq) if not.  If this handler gets
called from a legacy request function it will be called with the
queue_lock held, but if it's called from a blk-mq driver or actually
uses the IPI no lock will be held.

When I did my blk-mq only for dm-mpath WIP patch my solution to that
was that I removed the ->complete handler entirely and just ran the
whole dm completion from the original hardirq context.  With that change
I know that for blk-mq we'll never hold the queue_lock (and the blk-mq
request free path doesn't care), and for legacy we always hold it,
so __blk_put_request can always be used.