blk_mq_unquiesce_queue is still very much safe to call as many times as we
want.
The only thing that comes in the way is this pesky WARN. How about we make
it go away and have
a debug print instead?

My preference would be to allow nvme to unquiesce queues that were not
previously quiesced (just
like it historically was) instead of having to block a controller reset
until the scan_work is completed (which
is admin I/O dependent, and may get stuck until admin timeout, which can be
changed by the user for 60
minutes or something arbitrarily long btw).

How about something like this patch instead:
--

diff --git a/block/blk-mq.c b/block/blk-mq.c
index c2697db59109..74f3ad16e812 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c

@@ -327,8 +327,10 @@ void blk_mq_unquiesce_queue(struct request_queue *q)

        bool run_queue = false;

        spin_lock_irqsave(&q->queue_lock, flags);
-       if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
-               ;
+       if (q->quiesce_depth <= 0) {
+               printk(KERN_DEBUG
+                       "dev %s: unquiescing a non-quiesced queue,
expected?\n",
+                       q->disk ? q->disk->disk_name : "?", );
        } else if (!--q->quiesce_depth) {
                blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
                run_queue = true;
--