Reduce latency effects of non-interactive I/O.

Investigating influence of scrub (especially sequential) on random read
latency I've noticed that on some HDDs single 4KB read may take up to 4
seconds!  Deeper investigation shown that many HDDs heavily prioritize
sequential reads even when those are submitted with queue depth of 1.

This patch addresses the latency from two sides:
 - by using _min_active queue depths for non-interactive requests while
   the interactive request(s) are active and few requests after;
 - by throttling it further if no interactive requests has completed
   while configured amount of non-interactive did.

While there, I've also modified vdev_queue_class_to_issue() to give
more chances to schedule at least _min_active requests to the lowest
priorities.  It should reduce starvation if several non-interactive
processes are running same time with some interactive and I think should
make possible setting of zfs_vdev_max_active to as low as 1.

I've benchmarked this change with 4KB random reads from ZVOL with 16KB
block size on newly written non-fragmented pool.  On fragmented pool I
also saw improvements, but not so dramatic.  Below are log2 histograms
of the random read latency in milliseconds for different devices:

4 2x mirror vdevs of SATA HDD WDC WD20EFRX-68EUZN0 before:
0, 0, 2,  1,  12,  21,  19,  18, 10, 15, 17, 21
after:
0, 0, 0, 24, 101, 195, 419, 250, 47,  4,  0,  0
, that means maximum latency reduction from 2s to 500ms.

4 2x mirror vdevs of SATA HDD WDC WD80EFZX-68UW8N0 before:
0, 0,  2,  31,  38,  28,  18,  12, 17, 20, 24, 10, 3
after:
0, 0, 55, 247, 455, 470, 412, 181, 36,  0,  0,  0, 0
, i.e. from 4s to 250ms.

1 SAS HDD SEAGATE ST14000NM0048 before:
0,  0,  29,   70, 107,   45,  27, 1, 0, 0, 1, 4, 19
after:
1, 29, 681, 1261, 676, 1633,  67, 1, 0, 0, 0, 0,  0
, i.e. from 4s to 125ms.

1 SAS SSD SEAGATE XS3840TE70014 before (microseconds):
0, 0, 0, 0, 0, 0, 0, 0,  70, 18343, 82548, 618
after:
0, 0, 0, 0, 0, 0, 0, 0, 283, 92351, 34844,  90

I've also measured scrub time during the test and on idle pools.  On
idle fragmented pool I've measured scrub getting few percent faster
due to use of QD3 instead of QD2 before.  On idle non-fragmented pool
I've measured no difference.  On busy non-fragmented pool I've measured
scrub time increase about 1.5-1.7x, while IOPS increase reached 5-9x.

Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.

include/sys/vdev_impl.h

@@ -148,6 +148,9 @@ struct vdev_queue {
 	avl_tree_t	vq_write_offset_tree;
 	avl_tree_t	vq_trim_offset_tree;
 	uint64_t	vq_last_offset;
+	zio_priority_t	vq_last_prio;	/* Last sent I/O priority. */
+	uint32_t	vq_ia_active;	/* Active interactive I/Os. */
+	uint32_t	vq_nia_credit;	/* Non-interactive I/Os credit. */
 	hrtime_t	vq_io_complete_ts; /* time last i/o completed */
 	hrtime_t	vq_io_delta_ts;
 	zio_t		vq_io_search; /* used as local for stack reduction */

man/man5/zfs-module-parameters.5

@@ -2015,8 +2015,7 @@ Default value: \fB1\fR.
 .ad
 .RS 12n
 The maximum number of I/Os active to each device.  Ideally, this will be >=
-the sum of each queue's max_active.  It must be at least the sum of each
-queue's min_active.  See the section "ZFS I/O SCHEDULER".
+the sum of each queue's max_active.  See the section "ZFS I/O SCHEDULER".
 .sp
 Default value: \fB1,000\fR.
 .RE
@@ -2165,6 +2164,41 @@ See the section "ZFS I/O SCHEDULER".
 Default value: \fB1\fR.
 .RE
 
+.sp
+.ne 2
+.na
+\fBzfs_vdev_nia_delay\fR (int)
+.ad
+.RS 12n
+To reduce the effects of non-interactive I/O (scrub, resilver, removal,
+initialize and rebuild) on interactive I/O (both sync and async) latency
+they are limited to *_min_active while there are outstanding interactive
+I/Os, or until an additional zfs_vdev_nia_delay I/O's complete after the
+last interactive I/O.
+See the section "ZFS I/O SCHEDULER".
+.sp
+Default value: \fB5\fR.
+.RE
+
+.sp
+.ne 2
+.na
+\fBzfs_vdev_nia_credit\fR (int)
+.ad
+.RS 12n
+Some HDDs tend to prioritize sequential I/O so high, that concurrent
+random I/O latency reaches several seconds.  On some HDDs it happens
+even if sequential I/Os are submitted one at a time, and so setting
+*_max_active to 1 does not help.  To prevent non-interactive I/Os, like
+scrub, from monopolizing the device no more than zfs_vdev_nia_credit
+I/Os can be sent while there are outstanding incomplete interactive
+I/Os.  This enforced wait ensures the HDD services the interactive I/O
+within a reasonable amount of time.
+See the section "ZFS I/O SCHEDULER".
+.sp
+Default value: \fB5\fR.
+.RE
+
 .sp
 .ne 2
 .na

module/zfs/vdev_queue.c

@@ -121,16 +121,17 @@
 
 /*
  * The maximum number of i/os active to each device.  Ideally, this will be >=
- * the sum of each queue's max_active.  It must be at least the sum of each
- * queue's min_active.
+ * the sum of each queue's max_active.
  */
 uint32_t zfs_vdev_max_active = 1000;
 
 /*
  * Per-queue limits on the number of i/os active to each device.  If the
  * number of active i/os is < zfs_vdev_max_active, then the min_active comes
- * into play. We will send min_active from each queue, and then select from
- * queues in the order defined by zio_priority_t.
+ * into play.  We will send min_active from each queue round-robin, and then
+ * send from queues in the order defined by zio_priority_t up to max_active.
+ * Some queues have additional mechanisms to limit number of active I/Os in
+ * addition to min_active and max_active, see below.
  *
  * In general, smaller max_active's will lead to lower latency of synchronous
  * operations.  Larger max_active's may lead to higher overall throughput,
@@ -151,7 +152,7 @@ uint32_t zfs_vdev_async_read_max_active = 3;
 uint32_t zfs_vdev_async_write_min_active = 2;
 uint32_t zfs_vdev_async_write_max_active = 10;
 uint32_t zfs_vdev_scrub_min_active = 1;
-uint32_t zfs_vdev_scrub_max_active = 2;
+uint32_t zfs_vdev_scrub_max_active = 3;
 uint32_t zfs_vdev_removal_min_active = 1;
 uint32_t zfs_vdev_removal_max_active = 2;
 uint32_t zfs_vdev_initializing_min_active = 1;
@@ -171,6 +172,27 @@ uint32_t zfs_vdev_rebuild_max_active = 3;
 int zfs_vdev_async_write_active_min_dirty_percent = 30;
 int zfs_vdev_async_write_active_max_dirty_percent = 60;
 
+/*
+ * To reduce the effects of non-interactive I/O (scrub, resilver, removal,
+ * initialize and rebuild) on interactive I/O (both sync and async) latency
+ * they are limited to *_min_active while there are outstanding interactive
+ * I/Os, or until an additional zfs_vdev_nia_delay I/O's complete after the
+ * last interactive I/O.
+ */
+uint_t zfs_vdev_nia_delay = 5;
+
+/*
+ * Some HDDs tend to prioritize sequential I/O so high that concurrent
+ * random I/O latency reaches several seconds.  On some HDDs it happens
+ * even if sequential I/Os are submitted one at a time, and so setting
+ * *_max_active to 1 does not help.  To prevent non-interactive I/Os, like
+ * scrub, from monopolizing the device no more than zfs_vdev_nia_credit
+ * I/Os can be sent while there are outstanding incomplete interactive
+ * I/Os.  This enforced wait ensures the HDD services the interactive I/O
+ * within a reasonable amount of time.
+ */
+uint_t zfs_vdev_nia_credit = 5;
+
 /*
  * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O.
  * For read I/Os, we also aggregate across small adjacency gaps; for writes
@@ -261,7 +283,7 @@ vdev_queue_timestamp_compare(const void *x1, const void *x2)
 }
 
 static int
-vdev_queue_class_min_active(zio_priority_t p)
+vdev_queue_class_min_active(vdev_queue_t *vq, zio_priority_t p)
 {
 	switch (p) {
 	case ZIO_PRIORITY_SYNC_READ:
@@ -273,15 +295,19 @@ vdev_queue_class_min_active(zio_priority_t p)
 	case ZIO_PRIORITY_ASYNC_WRITE:
 		return (zfs_vdev_async_write_min_active);
 	case ZIO_PRIORITY_SCRUB:
-		return (zfs_vdev_scrub_min_active);
+		return (vq->vq_ia_active == 0 ? zfs_vdev_scrub_min_active :
+		    MIN(vq->vq_nia_credit, zfs_vdev_scrub_min_active));
 	case ZIO_PRIORITY_REMOVAL:
-		return (zfs_vdev_removal_min_active);
+		return (vq->vq_ia_active == 0 ? zfs_vdev_removal_min_active :
+		    MIN(vq->vq_nia_credit, zfs_vdev_removal_min_active));
 	case ZIO_PRIORITY_INITIALIZING:
-		return (zfs_vdev_initializing_min_active);
+		return (vq->vq_ia_active == 0 ?zfs_vdev_initializing_min_active:
+		    MIN(vq->vq_nia_credit, zfs_vdev_initializing_min_active));
 	case ZIO_PRIORITY_TRIM:
 		return (zfs_vdev_trim_min_active);
 	case ZIO_PRIORITY_REBUILD:
-		return (zfs_vdev_rebuild_min_active);
+		return (vq->vq_ia_active == 0 ? zfs_vdev_rebuild_min_active :
+		    MIN(vq->vq_nia_credit, zfs_vdev_rebuild_min_active));
 	default:
 		panic("invalid priority %u", p);
 		return (0);
@@ -337,7 +363,7 @@ vdev_queue_max_async_writes(spa_t *spa)
 }
 
 static int
-vdev_queue_class_max_active(spa_t *spa, zio_priority_t p)
+vdev_queue_class_max_active(spa_t *spa, vdev_queue_t *vq, zio_priority_t p)
 {
 	switch (p) {
 	case ZIO_PRIORITY_SYNC_READ:
@@ -349,14 +375,34 @@ vdev_queue_class_max_active(spa_t *spa, zio_priority_t p)
 	case ZIO_PRIORITY_ASYNC_WRITE:
 		return (vdev_queue_max_async_writes(spa));
 	case ZIO_PRIORITY_SCRUB:
+		if (vq->vq_ia_active > 0) {
+			return (MIN(vq->vq_nia_credit,
+			    zfs_vdev_scrub_min_active));
+		} else if (vq->vq_nia_credit < zfs_vdev_nia_delay)
+			return (zfs_vdev_scrub_min_active);
 		return (zfs_vdev_scrub_max_active);
 	case ZIO_PRIORITY_REMOVAL:
+		if (vq->vq_ia_active > 0) {
+			return (MIN(vq->vq_nia_credit,
+			    zfs_vdev_removal_min_active));
+		} else if (vq->vq_nia_credit < zfs_vdev_nia_delay)
+			return (zfs_vdev_removal_min_active);
 		return (zfs_vdev_removal_max_active);
 	case ZIO_PRIORITY_INITIALIZING:
+		if (vq->vq_ia_active > 0) {
+			return (MIN(vq->vq_nia_credit,
+			    zfs_vdev_initializing_min_active));
+		} else if (vq->vq_nia_credit < zfs_vdev_nia_delay)
+			return (zfs_vdev_initializing_min_active);
 		return (zfs_vdev_initializing_max_active);
 	case ZIO_PRIORITY_TRIM:
 		return (zfs_vdev_trim_max_active);
 	case ZIO_PRIORITY_REBUILD:
+		if (vq->vq_ia_active > 0) {
+			return (MIN(vq->vq_nia_credit,
+			    zfs_vdev_rebuild_min_active));
+		} else if (vq->vq_nia_credit < zfs_vdev_nia_delay)
+			return (zfs_vdev_rebuild_min_active);
 		return (zfs_vdev_rebuild_max_active);
 	default:
 		panic("invalid priority %u", p);
@@ -372,17 +418,24 @@ static zio_priority_t
 vdev_queue_class_to_issue(vdev_queue_t *vq)
 {
 	spa_t *spa = vq->vq_vdev->vdev_spa;
-	zio_priority_t p;
+	zio_priority_t p, n;
 
 	if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active)
 		return (ZIO_PRIORITY_NUM_QUEUEABLE);
 
-	/* find a queue that has not reached its minimum # outstanding i/os */
-	for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
+	/*
+	 * Find a queue that has not reached its minimum # outstanding i/os.
+	 * Do round-robin to reduce starvation due to zfs_vdev_max_active
+	 * and vq_nia_credit limits.
+	 */
+	for (n = 0; n < ZIO_PRIORITY_NUM_QUEUEABLE; n++) {
+		p = (vq->vq_last_prio + n + 1) % ZIO_PRIORITY_NUM_QUEUEABLE;
 		if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
 		    vq->vq_class[p].vqc_active <
-		    vdev_queue_class_min_active(p))
+		    vdev_queue_class_min_active(vq, p)) {
+			vq->vq_last_prio = p;
 			return (p);
+		}
 	}
 
 	/*
@@ -392,8 +445,10 @@ vdev_queue_class_to_issue(vdev_queue_t *vq)
 	for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
 		if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
 		    vq->vq_class[p].vqc_active <
-		    vdev_queue_class_max_active(spa, p))
+		    vdev_queue_class_max_active(spa, vq, p)) {
+			vq->vq_last_prio = p;
 			return (p);
+		}
 	}
 
 	/* No eligible queued i/os */
@@ -493,6 +548,20 @@ vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
 	}
 }
 
+static boolean_t
+vdev_queue_is_interactive(zio_priority_t p)
+{
+	switch (p) {
+	case ZIO_PRIORITY_SCRUB:
+	case ZIO_PRIORITY_REMOVAL:
+	case ZIO_PRIORITY_INITIALIZING:
+	case ZIO_PRIORITY_REBUILD:
+		return (B_FALSE);
+	default:
+		return (B_TRUE);
+	}
+}
+
 static void
 vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio)
 {
@@ -502,6 +571,11 @@ vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio)
 	ASSERT(MUTEX_HELD(&vq->vq_lock));
 	ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
 	vq->vq_class[zio->io_priority].vqc_active++;
+	if (vdev_queue_is_interactive(zio->io_priority)) {
+		if (++vq->vq_ia_active == 1)
+			vq->vq_nia_credit = 1;
+	} else if (vq->vq_ia_active > 0)
+			vq->vq_nia_credit--;
 	avl_add(&vq->vq_active_tree, zio);
 
 	if (shk->kstat != NULL) {
@@ -520,6 +594,13 @@ vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio)
 	ASSERT(MUTEX_HELD(&vq->vq_lock));
 	ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
 	vq->vq_class[zio->io_priority].vqc_active--;
+	if (vdev_queue_is_interactive(zio->io_priority)) {
+		if (--vq->vq_ia_active == 0)
+			vq->vq_nia_credit = 0;
+		else
+			vq->vq_nia_credit = zfs_vdev_nia_credit;
+	} else if (vq->vq_ia_active == 0)
+		vq->vq_nia_credit++;
 	avl_remove(&vq->vq_active_tree, zio);
 
 	if (shk->kstat != NULL) {
@@ -1065,6 +1146,12 @@ ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, rebuild_max_active, INT, ZMOD_RW,
 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, rebuild_min_active, INT, ZMOD_RW,
 	"Min active rebuild I/Os per vdev");
 
+ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, nia_credit, INT, ZMOD_RW,
+	"Number of non-interactive I/Os to allow in sequence");
+
+ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, nia_delay, INT, ZMOD_RW,
+	"Number of non-interactive I/Os before _max_active");
+
 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, queue_depth_pct, INT, ZMOD_RW,
 	"Queue depth percentage for each top-level vdev");
 /* END CSTYLED */