nvme-pci: fix a NULL pointer dereference in nvme_alloc_admin_tags

[ Upstream commit da42761 ]

In nvme_alloc_admin_tags, the admin_q can be set to an error (typically
-ENOMEM) if the blk_mq_init_queue call fails to set up the queue, which
is checked immediately after the call. However, when we return the error
message up the stack, to nvme_reset_work the error takes us to
nvme_remove_dead_ctrl()
  nvme_dev_disable()
   nvme_suspend_queue(&dev->queues[0]).

Here, we only check that the admin_q is non-NULL, rather than not
an error or NULL, and begin quiescing a queue that never existed, leading
to bad / NULL pointer dereference.

Signed-off-by: Kyle Smith <kyles@hpe.com>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Sasha Levin <sashal@kernel.org>

drivers/nvme/host/pci.c

@@ -1680,6 +1680,7 @@ static int nvme_alloc_admin_tags(struct nvme_dev *dev)
 		dev->ctrl.admin_q = blk_mq_init_queue(&dev->admin_tagset);
 		if (IS_ERR(dev->ctrl.admin_q)) {
 			blk_mq_free_tag_set(&dev->admin_tagset);
+			dev->ctrl.admin_q = NULL;
 			return -ENOMEM;
 		}
 		if (!blk_get_queue(dev->ctrl.admin_q)) {