blk-mq: drain I/O when all CPUs in a hctx are offline

Most of blk-mq drivers depend on managed IRQ's auto-affinity to setup
up queue mapping. Thomas mentioned the following point[1]:

"That was the constraint of managed interrupts from the very beginning:

 The driver/subsystem has to quiesce the interrupt line and the associated
 queue _before_ it gets shutdown in CPU unplug and not fiddle with it
 until it's restarted by the core when the CPU is plugged in again."

However, current blk-mq implementation doesn't quiesce hw queue before
the last CPU in the hctx is shutdown.  Even worse, CPUHP_BLK_MQ_DEAD is a
cpuhp state handled after the CPU is down, so there isn't any chance to
quiesce the hctx before shutting down the CPU.

Add new CPUHP_AP_BLK_MQ_ONLINE state to stop allocating from blk-mq hctxs
where the last CPU goes away, and wait for completion of in-flight
requests.  This guarantees that there is no inflight I/O before shutting
down the managed IRQ.

Add a BLK_MQ_F_STACKING and set it for dm-rq and loop, so we don't need
to wait for completion of in-flight requests from these drivers to avoid
a potential dead-lock. It is safe to do this for stacking drivers as those
do not use interrupts at all and their I/O completions are triggered by
underlying devices I/O completion.

[1] https://lore.kernel.org/linux-block/alpine.DEB.2.21.1904051331270.1802@nanos.tec.linutronix.de/

[hch: different retry mechanism, merged two patches, minor cleanups]

Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Reviewed-by: Daniel Wagner <dwagner@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

include/linux/cpuhotplug.h

@@ -152,6 +152,7 @@ enum cpuhp_state {
 	CPUHP_AP_SMPBOOT_THREADS,
 	CPUHP_AP_X86_VDSO_VMA_ONLINE,
 	CPUHP_AP_IRQ_AFFINITY_ONLINE,
+	CPUHP_AP_BLK_MQ_ONLINE,
 	CPUHP_AP_ARM_MVEBU_SYNC_CLOCKS,
 	CPUHP_AP_X86_INTEL_EPB_ONLINE,
 	CPUHP_AP_PERF_ONLINE,