nvme-tcp: check if the queue is allocated before stopping it

When an error is detected and the host reconnects, the
nvme_tcp_error_recovery_work() function is called and starts
tearing down the io queues and de-allocating them;
If at the same time the "nvme" process deletes the controller via sysfs,
the nvme_tcp_delete_ctrl() gets called and waits until the
nvme_tcp_error_recovery_work() finishes its job; then starts
tearing down the io queues, but at this point they have already
been freed and the mutexes are destroyed.

Calling mutex_lock() against a destroyed mutex triggers a warning:

[ 1299.025575] nvme nvme1: Reconnecting in 10 seconds...
[ 1299.636449] nvme nvme1: Removing ctrl: NQN "blktests-subsystem-1"
[ 1299.645262] ------------[ cut here ]------------
[ 1299.649949] DEBUG_LOCKS_WARN_ON(lock->magic != lock)
[ 1299.649971] WARNING: CPU: 4 PID: 104150 at kernel/locking/mutex.c:579 __mutex_lock+0x2d0/0x7dc

[ 1299.717934] CPU: 4 PID: 104150 Comm: nvme
[ 1299.828075] Call trace:
[ 1299.830526]  __mutex_lock+0x2d0/0x7dc
[ 1299.834203]  mutex_lock_nested+0x64/0xd4
[ 1299.838139]  nvme_tcp_stop_queue+0x54/0xe0 [nvme_tcp]
[ 1299.843211]  nvme_tcp_teardown_io_queues.part.0+0x90/0x280 [nvme_tcp]
[ 1299.849672]  nvme_tcp_delete_ctrl+0x6c/0xf0 [nvme_tcp]
[ 1299.854831]  nvme_do_delete_ctrl+0x108/0x120 [nvme_core]
[ 1299.860181]  nvme_sysfs_delete+0xec/0xf0 [nvme_core]
[ 1299.865179]  dev_attr_store+0x40/0x70

Fix the warning by checking if the queues are allocated
in the nvme_tcp_stop_queue(). If they are not, it makes no
sense to try to stop them.

Signed-off-by: Maurizio Lombardi <mlombard@redhat.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>

diff --git a/drivers/nvme/host/tcp.c b/drivers/nvme/host/tcp.c
index e82dcfc..044da18 100644 (file)
--- a/drivers/nvme/host/tcp.c
+++ b/drivers/nvme/host/tcp.c
@@ -1660,6 +1660,9 @@ static void nvme_tcp_stop_queue(struct nvme_ctrl *nctrl, int qid)
        struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
        struct nvme_tcp_queue *queue = &ctrl->queues[qid];
 
+       if (!test_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
+               return;
+
        mutex_lock(&queue->queue_lock);
        if (test_and_clear_bit(NVME_TCP_Q_LIVE, &queue->flags))
                __nvme_tcp_stop_queue(queue);