nvmf: Permit failing I/O requests while disconnected

Add a kern.nvmf.fail_on_disconnection sysctl similar to the
kern.iscsi.fail_on_disconnection sysctl.  This causes pending I/O
requests to fail with an error if an association is disconnected
instead of requeueing to be retried once the association is
reconnected.  As with iSCSI, the default is to queue and retry
operations.

Reviewed by:	imp
Sponsored by:	Chelsio Communications
Differential Revision:	https://reviews.freebsd.org/D45308

share/man/man4/nvmf.4

@@ -3,7 +3,7 @@
 .\"
 .\" Copyright (c) 2024 Chelsio Communications, Inc.
 .\"
-.Dd May 2, 2024
+.Dd June 5, 2024
 .Dt NVMF 4
 .Os
 .Sh NAME
@@ -65,6 +65,28 @@ disk driver.
 Associations require a supported transport such as
 .Xr nvmf_tcp 4
 for associations using TCP/IP.
+.Sh SYSCTL VARIABLES
+The following variables are available as both
+.Xr sysctl 8
+variables and
+.Xr loader 8
+tunables:
+.Bl -tag -width indent
+.It Va kern.nvmf.fail_on_disconnection
+Determines the behavior when an association's connection is interrupted.
+By default, input/output operations are suspended while a host is disconnected.
+This includes operations pending at the time the association's connection was
+interrupted as well as new requests submitted while the host is disconnected.
+Once a new association is established, suspended I/O requests are retried.
+When set to 1, input/output operations fail with
+.Er EIO
+while a host is disconnected and
+.Xr nda 4
+peripherals are destroyed after the first failed I/O request.
+Note that any destroyed
+.Xr nda 4
+peripherals will be recreated after a new association is established.
+.El
 .Sh SEE ALSO
 .Xr nda 4 ,
 .Xr nvme 4 ,

sys/dev/nvmf/host/nvmf.c

@@ -15,6 +15,7 @@
 #include <sys/module.h>
 #include <sys/mutex.h>
 #include <sys/sx.h>
+#include <sys/sysctl.h>
 #include <sys/taskqueue.h>
 #include <dev/nvme/nvme.h>
 #include <dev/nvmf/nvmf.h>
@@ -23,6 +24,10 @@
 
 static struct cdevsw nvmf_cdevsw;
 
+bool nvmf_fail_disconnect = false;
+SYSCTL_BOOL(_kern_nvmf, OID_AUTO, fail_on_disconnection, CTLFLAG_RWTUN,
+    &nvmf_fail_disconnect, 0, "Fail I/O requests on connection failure");
+
 MALLOC_DEFINE(M_NVMF, "nvmf", "NVMe over Fabrics host");
 
 static void	nvmf_disconnect_task(void *arg, int pending);

sys/dev/nvmf/host/nvmf_ns.c

@@ -84,13 +84,22 @@ nvmf_ns_biodone(struct bio *bio)
 	ns = bio->bio_dev->si_drv1;
 
 	/* If a request is aborted, resubmit or queue it for resubmission. */
-	if (bio->bio_error == ECONNABORTED) {
+	if (bio->bio_error == ECONNABORTED && !nvmf_fail_disconnect) {
 		bio->bio_error = 0;
 		bio->bio_driver2 = 0;
 		mtx_lock(&ns->lock);
 		if (ns->disconnected) {
-			TAILQ_INSERT_TAIL(&ns->pending_bios, bio, bio_queue);
-			mtx_unlock(&ns->lock);
+			if (nvmf_fail_disconnect) {
+				mtx_unlock(&ns->lock);
+				bio->bio_error = ECONNABORTED;
+				bio->bio_flags |= BIO_ERROR;
+				bio->bio_resid = bio->bio_bcount;
+				biodone(bio);
+			} else {
+				TAILQ_INSERT_TAIL(&ns->pending_bios, bio,
+				    bio_queue);
+				mtx_unlock(&ns->lock);
+			}
 		} else {
 			mtx_unlock(&ns->lock);
 			nvmf_ns_strategy(bio);
@@ -163,6 +172,7 @@ nvmf_ns_submit_bio(struct nvmf_namespace *ns, struct bio *bio)
 	struct nvme_dsm_range *dsm_range;
 	struct memdesc mem;
 	uint64_t lba, lba_count;
+	int error;
 
 	dsm_range = NULL;
 	memset(&cmd, 0, sizeof(cmd));
@@ -201,10 +211,15 @@ nvmf_ns_submit_bio(struct nvmf_namespace *ns, struct bio *bio)
 
 	mtx_lock(&ns->lock);
 	if (ns->disconnected) {
-		TAILQ_INSERT_TAIL(&ns->pending_bios, bio, bio_queue);
+		if (nvmf_fail_disconnect) {
+			error = ECONNABORTED;
+		} else {
+			TAILQ_INSERT_TAIL(&ns->pending_bios, bio, bio_queue);
+			error = 0;
+		}
 		mtx_unlock(&ns->lock);
 		free(dsm_range, M_NVMF);
-		return (0);
+		return (error);
 	}
 
 	req = nvmf_allocate_request(nvmf_select_io_queue(ns->sc), &cmd,

sys/dev/nvmf/host/nvmf_sim.c

@@ -40,7 +40,10 @@ nvmf_ccb_done(union ccb *ccb)
 		return;
 
 	if (nvmf_cqe_aborted(&ccb->nvmeio.cpl)) {
-		ccb->ccb_h.status = CAM_REQUEUE_REQ;
+		if (nvmf_fail_disconnect)
+			ccb->ccb_h.status = CAM_DEV_NOT_THERE;
+		else
+			ccb->ccb_h.status = CAM_REQUEUE_REQ;
 		xpt_done(ccb);
 	} else if (ccb->nvmeio.cpl.status != 0) {
 		ccb->ccb_h.status = CAM_NVME_STATUS_ERROR;
@@ -106,7 +109,10 @@ nvmf_sim_io(struct nvmf_softc *sc, union ccb *ccb)
 	mtx_lock(&sc->sim_mtx);
 	if (sc->sim_disconnected) {
 		mtx_unlock(&sc->sim_mtx);
-		nvmeio->ccb_h.status = CAM_REQUEUE_REQ;
+		if (nvmf_fail_disconnect)
+			nvmeio->ccb_h.status = CAM_DEV_NOT_THERE;
+		else
+			nvmeio->ccb_h.status = CAM_REQUEUE_REQ;
 		xpt_done(ccb);
 		return;
 	}

sys/dev/nvmf/host/nvmf_var.h

@@ -140,6 +140,9 @@ extern driver_t nvme_nvmf_driver;
 MALLOC_DECLARE(M_NVMF);
 #endif
 
+/* If true, I/O requests will fail while the host is disconnected. */
+extern bool nvmf_fail_disconnect;
+
 /* nvmf.c */
 void	nvmf_complete(void *arg, const struct nvme_completion *cqe);
 void	nvmf_io_complete(void *arg, size_t xfered, int error);