Limit stripesize reported from nvd(4) to 4K

Intel NVMe controllers have a slow path for I/Os that span a 128KB stripe boundary but ZFS limits ashift, which is derived from d_stripesize, to 13 (8KB) so we limit the stripesize reported to geom(8) to 4KB.

This may result in a small number of additional I/Os to require splitting in nvme(4), however the NVMe I/O path is very efficient so these additional I/Os will cause very minimal (if any) difference in performance or CPU utilisation.

This can be controller by the new sysctl kern.nvme.max_optimal_sectorsize.

MFC after:	1 week
Sponsored by:	Multiplay
Differential Revision:	https://reviews.freebsd.org/D4446

4 files changed, 36 insertions, 1 deletions

diff --git a/sys/dev/nvd/nvd.c b/sys/dev/nvd/nvd.c
index 5d75876..f459e06 100644
--- a/sys/dev/nvd/nvd.c
+++ b/sys/dev/nvd/nvd.c
@@ -279,7 +279,7 @@ nvd_new_disk(struct nvme_namespace *ns, void *ctrlr_arg)
 	disk->d_sectorsize = nvme_ns_get_sector_size(ns);
 	disk->d_mediasize = (off_t)nvme_ns_get_size(ns);
 	disk->d_delmaxsize = (off_t)nvme_ns_get_size(ns);
-	disk->d_stripesize = nvme_ns_get_stripesize(ns);
+	disk->d_stripesize = nvme_ns_get_optimal_sector_size(ns);
 
 	if (TAILQ_EMPTY(&disk_head))
 		disk->d_unit = 0;
diff --git a/sys/dev/nvme/nvme.h b/sys/dev/nvme/nvme.h
index 7e41e77..227a89e 100644
--- a/sys/dev/nvme/nvme.h
+++ b/sys/dev/nvme/nvme.h
@@ -870,6 +870,7 @@ const char *	nvme_ns_get_serial_number(struct nvme_namespace *ns);
 const char *	nvme_ns_get_model_number(struct nvme_namespace *ns);
 const struct nvme_namespace_data *
 		nvme_ns_get_data(struct nvme_namespace *ns);
+uint32_t	nvme_ns_get_optimal_sector_size(struct nvme_namespace *ns);
 uint32_t	nvme_ns_get_stripesize(struct nvme_namespace *ns);
 
 int	nvme_ns_bio_process(struct nvme_namespace *ns, struct bio *bp,
diff --git a/sys/dev/nvme/nvme_ns.c b/sys/dev/nvme/nvme_ns.c
index 754d074..4580e66 100644
--- a/sys/dev/nvme/nvme_ns.c
+++ b/sys/dev/nvme/nvme_ns.c
@@ -45,6 +45,8 @@ __FBSDID("$FreeBSD$");
 
 #include "nvme_private.h"
 
+extern int		nvme_max_optimal_sectorsize;
+
 static void		nvme_bio_child_inbed(struct bio *parent, int bio_error);
 static void		nvme_bio_child_done(void *arg,
 					    const struct nvme_completion *cpl);
@@ -217,6 +219,22 @@ nvme_ns_get_stripesize(struct nvme_namespace *ns)
 	return (ns->stripesize);
 }
 
+uint32_t
+nvme_ns_get_optimal_sector_size(struct nvme_namespace *ns)
+{
+	uint32_t stripesize;
+
+	stripesize = nvme_ns_get_stripesize(ns);
+
+	if (stripesize == 0)
+		return nvme_ns_get_sector_size(ns);
+		
+	if (nvme_max_optimal_sectorsize == 0) 
+		return (stripesize);
+
+	return (MIN(stripesize, nvme_max_optimal_sectorsize));
+}
+
 static void
 nvme_ns_bio_done(void *arg, const struct nvme_completion *status)
 {
diff --git a/sys/dev/nvme/nvme_sysctl.c b/sys/dev/nvme/nvme_sysctl.c
index 0ebbbf7..8b99111 100644
--- a/sys/dev/nvme/nvme_sysctl.c
+++ b/sys/dev/nvme/nvme_sysctl.c
@@ -33,6 +33,22 @@ __FBSDID("$FreeBSD$");
 
 #include "nvme_private.h"
 
+SYSCTL_NODE(_kern, OID_AUTO, nvme, CTLFLAG_RD, 0, "NVM Express");
+/*
+ * Intel NVMe controllers have a slow path for I/Os that span a 128KB
+ * stripe boundary but ZFS limits ashift, which is derived from
+ * d_stripesize, to 13 (8KB) so we limit the stripesize reported to
+ * geom(8) to 4KB by default.
+ *
+ * This may result in a small number of additional I/Os to require
+ * splitting in nvme(4), however the NVMe I/O path is very efficient
+ * so these additional I/Os will cause very minimal (if any) difference
+ * in performance or CPU utilisation.
+ */
+int nvme_max_optimal_sectorsize = 1<<12;
+SYSCTL_INT(_kern_nvme, OID_AUTO, max_optimal_sectorsize, CTLFLAG_RWTUN,
+    &nvme_max_optimal_sectorsize, 0, "The maximum optimal sectorsize reported");
+
 /*
  * CTLTYPE_S64 and sysctl_handle_64 were added in r217616.  Define these
  *  explicitly here for older kernels that don't include the r217616