bdev/nvme: Reset I/O cancels reconnect timer and starts reconnection

Previously, if a reconnect timer was registered when a reset request
came, the reset request failed with -EBUSY. However, this means the
reset request was queued for a long time until the reconnect timer was
expired.

When a reconnect timer is registered, reset is not actually in progress.
Hence, a new reset request can cancel the reconnect timer and can start
reconnection safely.

Add a unit test case to verify this change.

Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: Ied8dd0ad822d2fd6829d88cd56cb36bd4fad13f9
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/16823
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Community-CI: Mellanox Build Bot
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>

module/bdev/nvme/bdev_nvme.c

@@ -1863,7 +1863,11 @@ bdev_nvme_reconnect_delay_timer_expired(void *ctx)
 
 	spdk_poller_unregister(&nvme_ctrlr->reconnect_delay_timer);
 
-	assert(nvme_ctrlr->reconnect_is_delayed == true);
+	if (!nvme_ctrlr->reconnect_is_delayed) {
+		pthread_mutex_unlock(&nvme_ctrlr->mutex);
+		return SPDK_POLLER_BUSY;
+	}
+
 	nvme_ctrlr->reconnect_is_delayed = false;
 
 	if (nvme_ctrlr->destruct) {
@@ -2081,6 +2085,21 @@ bdev_nvme_reset_destroy_qpairs(struct nvme_ctrlr *nvme_ctrlr)
 			      bdev_nvme_reset_ctrlr);
 }
 
+static void
+_bdev_nvme_reconnect(void *ctx)
+{
+	struct nvme_ctrlr *nvme_ctrlr = ctx;
+
+	assert(nvme_ctrlr->resetting == true);
+	assert(nvme_ctrlr->thread == spdk_get_thread());
+
+	spdk_poller_unregister(&nvme_ctrlr->reconnect_delay_timer);
+
+	spdk_poller_resume(nvme_ctrlr->adminq_timer_poller);
+
+	bdev_nvme_reconnect_ctrlr(nvme_ctrlr);
+}
+
 static void
 _bdev_nvme_reset(void *ctx)
 {
@@ -2099,6 +2118,8 @@ _bdev_nvme_reset(void *ctx)
 static int
 bdev_nvme_reset(struct nvme_ctrlr *nvme_ctrlr)
 {
+	spdk_msg_fn msg_fn;
+
 	pthread_mutex_lock(&nvme_ctrlr->mutex);
 	if (nvme_ctrlr->destruct) {
 		pthread_mutex_unlock(&nvme_ctrlr->mutex);
@@ -2111,20 +2132,22 @@ bdev_nvme_reset(struct nvme_ctrlr *nvme_ctrlr)
 		return -EBUSY;
 	}
 
-	if (nvme_ctrlr->reconnect_is_delayed) {
-		pthread_mutex_unlock(&nvme_ctrlr->mutex);
-		SPDK_NOTICELOG("Reconnect is already scheduled.\n");
-		return -EBUSY;
-	}
-
 	nvme_ctrlr->resetting = true;
 
-	assert(nvme_ctrlr->reset_start_tsc == 0);
+	if (nvme_ctrlr->reconnect_is_delayed) {
+		SPDK_DEBUGLOG(bdev_nvme, "Reconnect is already scheduled.\n");
+		msg_fn = _bdev_nvme_reconnect;
+		nvme_ctrlr->reconnect_is_delayed = false;
+	} else {
+		msg_fn = _bdev_nvme_reset;
+		assert(nvme_ctrlr->reset_start_tsc == 0);
+	}
+
 	nvme_ctrlr->reset_start_tsc = spdk_get_ticks();
 
 	pthread_mutex_unlock(&nvme_ctrlr->mutex);
 
-	spdk_thread_send_msg(nvme_ctrlr->thread, _bdev_nvme_reset, nvme_ctrlr);
+	spdk_thread_send_msg(nvme_ctrlr->thread, msg_fn, nvme_ctrlr);
 	return 0;
 }
 

test/unit/lib/bdev/nvme/bdev_nvme.c/bdev_nvme_ut.c

@@ -5251,6 +5251,30 @@ test_reconnect_ctrlr(void)
 	CU_ASSERT(nvme_ctrlr->reconnect_delay_timer != NULL);
 	CU_ASSERT(nvme_ctrlr->reconnect_is_delayed == true);
 
+	/* A new reset starts from thread 0. */
+	set_thread(1);
+
+	/* The reset should cancel the reconnect timer and should start from reconnection.
+	 * Then, the reset should fail and a reconnect timer should be registered again.
+	 */
+	ctrlr.fail_reset = true;
+	ctrlr.is_failed = true;
+
+	rc = bdev_nvme_reset(nvme_ctrlr);
+	CU_ASSERT(rc == 0);
+	CU_ASSERT(nvme_ctrlr->resetting == true);
+	CU_ASSERT(nvme_ctrlr->reconnect_is_delayed == false);
+	CU_ASSERT(ctrlr.is_failed == true);
+
+	poll_threads();
+
+	CU_ASSERT(nvme_ctrlr->resetting == false);
+	CU_ASSERT(ctrlr.is_failed == false);
+	CU_ASSERT(ctrlr_ch1->qpair->qpair == NULL);
+	CU_ASSERT(ctrlr_ch2->qpair->qpair == NULL);
+	CU_ASSERT(nvme_ctrlr->reconnect_delay_timer != NULL);
+	CU_ASSERT(nvme_ctrlr->reconnect_is_delayed == true);
+
 	/* Then a reconnect retry should suceeed. */
 	ctrlr.fail_reset = false;