nvme: fix double blk_mq_complete_request for timeout request with low probability

When nvme_cancel_tagset traverses all tagsets and executes
nvme_cancel_request, this request may be executing blk_mq_free_request
that is called by nvme_rdma_complete_timed_out/nvme_tcp_complete_timed_out.
When blk_mq_free_request executes to WRITE_ONCE(rq->state, MQ_RQ_IDLE) and
__blk_mq_free_request(rq), it will cause double blk_mq_complete_request for
this request, and it will cause a null pointer error in the second
execution of this function because rq->mq_hctx has set to NULL in first
execution.

Signed-off-by: Lei Yin <yinlei2@lenovo.com>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>

diff --git a/drivers/nvme/host/core.c b/drivers/nvme/host/core.c
index 0699365..6c1e7d6 100644 (file)
--- a/drivers/nvme/host/core.c
+++ b/drivers/nvme/host/core.c
@@ -450,8 +450,8 @@ bool nvme_cancel_request(struct request *req, void *data)
        dev_dbg_ratelimited(((struct nvme_ctrl *) data)->device,
                                "Cancelling I/O %d", req->tag);
 
-       /* don't abort one completed request */
-       if (blk_mq_request_completed(req))
+       /* don't abort one completed or idle request */
+       if (blk_mq_rq_state(req) != MQ_RQ_IN_FLIGHT)
                return true;
 
        nvme_req(req)->status = NVME_SC_HOST_ABORTED_CMD;