blk_io_plug_call() is invoked outside a blk_io_plug()/blk_io_unplug()
section while opening the NVMe drive from:

  nvme_file_open() ->
  nvme_init() ->
  nvme_identify() ->
  nvme_admin_cmd_sync() ->
  nvme_submit_command() ->
  blk_io_plug_call()

blk_io_plug_call() immediately invokes the given callback when the
current thread is not plugged, as is the case during nvme_file_open().

Unfortunately, nvme_submit_command() calls blk_io_plug_call() with
q->lock still held:

    ...
    q->sq.tail = (q->sq.tail + 1) % NVME_QUEUE_SIZE;
    q->need_kick++;
    blk_io_plug_call(nvme_unplug_fn, q);
    qemu_mutex_unlock(&q->lock);
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^

nvme_unplug_fn() deadlocks trying to acquire q->lock because the lock is
already acquired by the same thread. The symptom is that QEMU hangs
during startup while opening the NVMe drive.

Fix this by moving the blk_io_plug_call() outside q->lock. This is safe
because no other thread runs code related to this queue and
blk_io_plug_call()'s internal state is immune to thread safety issues
since it is thread-local.

Reported-by: Lukáš Doktor <ldoktor@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Tested-by: Lukas Doktor <ldoktor@redhat.com>
Message-id: 20230712191628.252806-1-stefanha@redhat.com
Fixes: f2e590002bd6 ("block/nvme: convert to blk_io_plug_call() API")
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
---
 block/nvme.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/block/nvme.c b/block/nvme.c
index 7ca85bc44a..b6e95f0b7e 100644
--- a/block/nvme.c
+++ b/block/nvme.c
@@ -501,8 +501,9 @@ static void nvme_submit_command(NVMeQueuePair *q, NVMeRequest *req,
            q->sq.tail * NVME_SQ_ENTRY_BYTES, cmd, sizeof(*cmd));
     q->sq.tail = (q->sq.tail + 1) % NVME_QUEUE_SIZE;
     q->need_kick++;
+    qemu_mutex_unlock(&q->lock);
+
     blk_io_plug_call(nvme_unplug_fn, q);
-    qemu_mutex_unlock(&q->lock);
 }
 
 static void nvme_admin_cmd_sync_cb(void *opaque, int ret)
-- 
2.40.1