diff --git a/include/spdk/nvmf_transport.h b/include/spdk/nvmf_transport.h
index 828c5828e..810bbb02f 100644
--- a/include/spdk/nvmf_transport.h
+++ b/include/spdk/nvmf_transport.h
@@ -126,7 +126,10 @@ struct spdk_nvmf_qpair {
 	bool					connect_received;
 	bool					disconnect_started;
 
-	struct spdk_nvmf_request		*first_fused_req;
+	union {
+		struct spdk_nvmf_request	*first_fused_req;
+		struct spdk_nvmf_request	*connect_req;
+	};
 
 	TAILQ_HEAD(, spdk_nvmf_request)		outstanding;
 	TAILQ_ENTRY(spdk_nvmf_qpair)		link;
diff --git a/lib/nvmf/ctrlr.c b/lib/nvmf/ctrlr.c
index 3e57f3530..01ed56f69 100644
--- a/lib/nvmf/ctrlr.c
+++ b/lib/nvmf/ctrlr.c
@@ -31,6 +31,8 @@
 
 #define NVMF_CTRLR_RESET_SHN_TIMEOUT_IN_MS	(NVMF_CC_RESET_SHN_TIMEOUT_IN_MS + 5000)
 
+#define DUPLICATE_QID_RETRY_US 100
+
 /*
  * Report the SPDK version as the firmware revision.
  * SPDK_VERSION_STRING won't fit into FR (only 8 bytes), so try to fit the most important parts.
@@ -209,6 +211,8 @@ nvmf_ctrlr_start_keep_alive_timer(struct spdk_nvmf_ctrlr *ctrlr)
 	}
 }
 
+static int _retry_qid_check(void *ctx);
+
 static void
 ctrlr_add_qpair_and_send_rsp(struct spdk_nvmf_qpair *qpair,
 			     struct spdk_nvmf_ctrlr *ctrlr,
@@ -219,10 +223,22 @@ ctrlr_add_qpair_and_send_rsp(struct spdk_nvmf_qpair *qpair,
 	assert(ctrlr->admin_qpair->group->thread == spdk_get_thread());
 
 	if (spdk_bit_array_get(ctrlr->qpair_mask, qpair->qid)) {
-		SPDK_ERRLOG("Got I/O connect with duplicate QID %u\n", qpair->qid);
-		rsp->status.sct = SPDK_NVME_SCT_COMMAND_SPECIFIC;
-		rsp->status.sc = SPDK_NVME_SC_INVALID_QUEUE_IDENTIFIER;
-		spdk_nvmf_request_complete(req);
+		if (qpair->connect_req != NULL) {
+			SPDK_ERRLOG("Got I/O connect with duplicate QID %u\n", qpair->qid);
+			rsp->status.sct = SPDK_NVME_SCT_COMMAND_SPECIFIC;
+			rsp->status.sc = SPDK_NVME_SC_INVALID_QUEUE_IDENTIFIER;
+			qpair->connect_req = NULL;
+			qpair->ctrlr = NULL;
+			spdk_nvmf_request_complete(req);
+		} else {
+			SPDK_WARNLOG("Duplicate QID detected, re-check in %dus\n",
+				     DUPLICATE_QID_RETRY_US);
+			qpair->connect_req = req;
+			/* Set qpair->ctrlr here so that we'll have it when the poller expires. */
+			qpair->ctrlr = ctrlr;
+			req->poller = SPDK_POLLER_REGISTER(_retry_qid_check, qpair,
+							   DUPLICATE_QID_RETRY_US);
+		}
 		return;
 	}
 
@@ -239,6 +255,18 @@ ctrlr_add_qpair_and_send_rsp(struct spdk_nvmf_qpair *qpair,
 			   ctrlr->hostnqn);
 }
 
+static int
+_retry_qid_check(void *ctx)
+{
+	struct spdk_nvmf_qpair *qpair = ctx;
+	struct spdk_nvmf_request *req = qpair->connect_req;
+	struct spdk_nvmf_ctrlr *ctrlr = req->qpair->ctrlr;
+
+	spdk_poller_unregister(&req->poller);
+	ctrlr_add_qpair_and_send_rsp(qpair, ctrlr, req);
+	return SPDK_POLLER_BUSY;
+}
+
 static void
 _nvmf_ctrlr_add_admin_qpair(void *ctx)
 {
diff --git a/test/unit/lib/nvmf/ctrlr.c/ctrlr_ut.c b/test/unit/lib/nvmf/ctrlr.c/ctrlr_ut.c
index 17e1abbf3..b699b2249 100644
--- a/test/unit/lib/nvmf/ctrlr.c/ctrlr_ut.c
+++ b/test/unit/lib/nvmf/ctrlr.c/ctrlr_ut.c
@@ -811,13 +811,46 @@ test_connect(void)
 	sgroups[subsystem.id].mgmt_io_outstanding++;
 	TAILQ_INSERT_TAIL(&qpair.outstanding, &req, link);
 	rc = nvmf_ctrlr_cmd_connect(&req);
-	poll_threads();
 	CU_ASSERT(rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS);
+	poll_threads();
+	/* First time, it will detect duplicate QID and schedule a retry.  So for
+	 * now we should expect the response to still be all zeroes.
+	 */
+	CU_ASSERT(spdk_mem_all_zero(&rsp, sizeof(rsp)));
+	CU_ASSERT(sgroups[subsystem.id].mgmt_io_outstanding == 1);
+
+	/* Now advance the clock, so that the retry poller executes. */
+	spdk_delay_us(DUPLICATE_QID_RETRY_US * 2);
+	poll_threads();
 	CU_ASSERT(rsp.nvme_cpl.status.sct == SPDK_NVME_SCT_COMMAND_SPECIFIC);
 	CU_ASSERT(rsp.nvme_cpl.status.sc == SPDK_NVME_SC_INVALID_QUEUE_IDENTIFIER);
 	CU_ASSERT(qpair.ctrlr == NULL);
 	CU_ASSERT(sgroups[subsystem.id].mgmt_io_outstanding == 0);
 
+	/* I/O connect with temporarily duplicate queue ID. This covers race
+	 * where qpair_mask bit may not yet be cleared, even though initiator
+	 * has closed the connection.  See issue #2955. */
+	memset(&rsp, 0, sizeof(rsp));
+	sgroups[subsystem.id].mgmt_io_outstanding++;
+	TAILQ_INSERT_TAIL(&qpair.outstanding, &req, link);
+	rc = nvmf_ctrlr_cmd_connect(&req);
+	CU_ASSERT(rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS);
+	poll_threads();
+	/* First time, it will detect duplicate QID and schedule a retry.  So for
+	 * now we should expect the response to still be all zeroes.
+	 */
+	CU_ASSERT(spdk_mem_all_zero(&rsp, sizeof(rsp)));
+	CU_ASSERT(sgroups[subsystem.id].mgmt_io_outstanding == 1);
+
+	/* Now advance the clock, so that the retry poller executes. */
+	spdk_bit_array_clear(ctrlr.qpair_mask, 1);
+	spdk_delay_us(DUPLICATE_QID_RETRY_US * 2);
+	poll_threads();
+	CU_ASSERT(nvme_status_success(&rsp.nvme_cpl.status));
+	CU_ASSERT(qpair.ctrlr == &ctrlr);
+	CU_ASSERT(sgroups[subsystem.id].mgmt_io_outstanding == 0);
+	qpair.ctrlr = NULL;
+
 	/* I/O connect when admin qpair is being destroyed */
 	admin_qpair.group = NULL;
 	admin_qpair.state = SPDK_NVMF_QPAIR_DEACTIVATING;