NVMe: Fix locking on abort handling
authorKeith Busch <keith.busch@intel.com>
Thu, 8 Jan 2015 01:55:53 +0000 (18:55 -0700)
committerJens Axboe <axboe@fb.com>
Thu, 8 Jan 2015 16:02:23 +0000 (09:02 -0700)
The queues and device need to be locked when messing with them.

Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
drivers/block/nvme-core.c

index ff4ff0999f022d5368e20e3dd638c4c01a2f3156..cb529e9a82dd685b5b372bea2ed272c59fae5bc5 100644 (file)
@@ -1021,14 +1021,19 @@ static void nvme_abort_req(struct request *req)
        struct nvme_command cmd;
 
        if (!nvmeq->qid || cmd_rq->aborted) {
+               unsigned long flags;
+
+               spin_lock_irqsave(&dev_list_lock, flags);
                if (work_busy(&dev->reset_work))
-                       return;
+                       goto out;
                list_del_init(&dev->node);
                dev_warn(&dev->pci_dev->dev,
                        "I/O %d QID %d timeout, reset controller\n",
                                                        req->tag, nvmeq->qid);
                dev->reset_workfn = nvme_reset_failed_dev;
                queue_work(nvme_workq, &dev->reset_work);
+ out:
+               spin_unlock_irqrestore(&dev_list_lock, flags);
                return;
        }
 
@@ -1096,25 +1101,29 @@ static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
        struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(req);
        struct nvme_queue *nvmeq = cmd->nvmeq;
 
+       /*
+        * The aborted req will be completed on receiving the abort req.
+        * We enable the timer again. If hit twice, it'll cause a device reset,
+        * as the device then is in a faulty state.
+        */
+       int ret = BLK_EH_RESET_TIMER;
+
        dev_warn(nvmeq->q_dmadev, "Timeout I/O %d QID %d\n", req->tag,
                                                        nvmeq->qid);
 
+       spin_lock_irq(&nvmeq->q_lock);
        if (!nvmeq->dev->initialized) {
                /*
                 * Force cancelled command frees the request, which requires we
                 * return BLK_EH_NOT_HANDLED.
                 */
                nvme_cancel_queue_ios(nvmeq->hctx, req, nvmeq, reserved);
-               return BLK_EH_NOT_HANDLED;
-       }
-       nvme_abort_req(req);
+               ret = BLK_EH_NOT_HANDLED;
+       } else
+               nvme_abort_req(req);
+       spin_unlock_irq(&nvmeq->q_lock);
 
-       /*
-        * The aborted req will be completed on receiving the abort req.
-        * We enable the timer again. If hit twice, it'll cause a device reset,
-        * as the device then is in a faulty state.
-        */
-       return BLK_EH_RESET_TIMER;
+       return ret;
 }
 
 static void nvme_free_queue(struct nvme_queue *nvmeq)