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 <ldok...@redhat.com> Signed-off-by: Stefan Hajnoczi <stefa...@redhat.com> Tested-by: Lukas Doktor <ldok...@redhat.com> Message-id: 20230712191628.252806-1-stefa...@redhat.com Fixes: f2e590002bd6 ("block/nvme: convert to blk_io_plug_call() API") Signed-off-by: Stefan Hajnoczi <stefa...@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