15 files changed, 550 insertions, 95 deletions

diff --git a/sys/contrib/openzfs/module/os/freebsd/zfs/zvol_os.c b/sys/contrib/openzfs/module/os/freebsd/zfs/zvol_os.c
index 6a7c2d2811b1..712ff1b837d7 100644
--- a/sys/contrib/openzfs/module/os/freebsd/zfs/zvol_os.c
+++ b/sys/contrib/openzfs/module/os/freebsd/zfs/zvol_os.c
@@ -1259,7 +1259,7 @@ zvol_os_rename_minor(zvol_state_t *zv, const char *newname)
 	ASSERT(MUTEX_HELD(&zv->zv_state_lock));
 
 	/* Move to a new hashtable entry.  */
-	zv->zv_hash = zvol_name_hash(zv->zv_name);
+	zv->zv_hash = zvol_name_hash(newname);
 	hlist_del(&zv->zv_hlink);
 	hlist_add_head(&zv->zv_hlink, ZVOL_HT_HEAD(zv->zv_hash));
 
diff --git a/sys/contrib/openzfs/module/os/linux/zfs/vdev_disk.c b/sys/contrib/openzfs/module/os/linux/zfs/vdev_disk.c
index 2cea61a6294c..463c5f705102 100644
--- a/sys/contrib/openzfs/module/os/linux/zfs/vdev_disk.c
+++ b/sys/contrib/openzfs/module/os/linux/zfs/vdev_disk.c
@@ -429,8 +429,11 @@ vdev_disk_open(vdev_t *v, uint64_t *psize, uint64_t *max_psize,
 	/*  Determine the logical block size */
 	int logical_block_size = bdev_logical_block_size(bdev);
 
-	/* Clear the nowritecache bit, causes vdev_reopen() to try again. */
-	v->vdev_nowritecache = B_FALSE;
+	/*
+	 * If the device has a write cache, clear the nowritecache flag,
+	 * so that we start issuing flush requests again.
+	 */
+	v->vdev_nowritecache = !zfs_bdev_has_write_cache(bdev);
 
 	/* Set when device reports it supports TRIM. */
 	v->vdev_has_trim = bdev_discard_supported(bdev);
diff --git a/sys/contrib/openzfs/module/os/linux/zfs/zvol_os.c b/sys/contrib/openzfs/module/os/linux/zfs/zvol_os.c
index 4b960daf89ee..2a036dc5136b 100644
--- a/sys/contrib/openzfs/module/os/linux/zfs/zvol_os.c
+++ b/sys/contrib/openzfs/module/os/linux/zfs/zvol_os.c
@@ -1571,7 +1571,7 @@ zvol_os_rename_minor(zvol_state_t *zv, const char *newname)
 	strlcpy(zv->zv_name, newname, sizeof (zv->zv_name));
 
 	/* move to new hashtable entry  */
-	zv->zv_hash = zvol_name_hash(zv->zv_name);
+	zv->zv_hash = zvol_name_hash(newname);
 	hlist_del(&zv->zv_hlink);
 	hlist_add_head(&zv->zv_hlink, ZVOL_HT_HEAD(zv->zv_hash));
 
diff --git a/sys/contrib/openzfs/module/zfs/dmu_objset.c b/sys/contrib/openzfs/module/zfs/dmu_objset.c
index 2ba26f68e398..f1818ae155bd 100644
--- a/sys/contrib/openzfs/module/zfs/dmu_objset.c
+++ b/sys/contrib/openzfs/module/zfs/dmu_objset.c
@@ -400,10 +400,10 @@ dnode_hash(const objset_t *os, uint64_t obj)
 
 	ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
 	/*
-	 * The low 6 bits of the pointer don't have much entropy, because
-	 * the objset_t is larger than 2^6 bytes long.
+	 * The lower 11 bits of the pointer don't have much entropy, because
+	 * the objset_t is more than 1KB long and so likely aligned to 2KB.
 	 */
-	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
+	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 11)) & 0xFF];
 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 16)) & 0xFF];
@@ -1664,12 +1664,14 @@ sync_dnodes_task(void *arg)
 	sync_objset_arg_t *soa = sda->sda_soa;
 	objset_t *os = soa->soa_os;
 
+	uint_t allocator = spa_acq_allocator(os->os_spa);
 	multilist_sublist_t *ms =
 	    multilist_sublist_lock_idx(sda->sda_list, sda->sda_sublist_idx);
 
 	dmu_objset_sync_dnodes(ms, soa->soa_tx);
 
 	multilist_sublist_unlock(ms);
+	spa_rel_allocator(os->os_spa, allocator);
 
 	kmem_free(sda, sizeof (*sda));
 
diff --git a/sys/contrib/openzfs/module/zfs/mmp.c b/sys/contrib/openzfs/module/zfs/mmp.c
index 66bc0ae60b10..71122542758d 100644
--- a/sys/contrib/openzfs/module/zfs/mmp.c
+++ b/sys/contrib/openzfs/module/zfs/mmp.c
@@ -664,12 +664,13 @@ mmp_thread(void *arg)
 		    (gethrtime() - mmp->mmp_last_write) > mmp_fail_ns) {
 			zfs_dbgmsg("MMP suspending pool '%s': gethrtime %llu "
 			    "mmp_last_write %llu mmp_interval %llu "
-			    "mmp_fail_intervals %llu mmp_fail_ns %llu",
+			    "mmp_fail_intervals %llu mmp_fail_ns %llu txg %llu",
 			    spa_name(spa), (u_longlong_t)gethrtime(),
 			    (u_longlong_t)mmp->mmp_last_write,
 			    (u_longlong_t)mmp_interval,
 			    (u_longlong_t)mmp_fail_intervals,
-			    (u_longlong_t)mmp_fail_ns);
+			    (u_longlong_t)mmp_fail_ns,
+			    (u_longlong_t)spa->spa_uberblock.ub_txg);
 			cmn_err(CE_WARN, "MMP writes to pool '%s' have not "
 			    "succeeded in over %llu ms; suspending pool. "
 			    "Hrtime %llu",
diff --git a/sys/contrib/openzfs/module/zfs/spa.c b/sys/contrib/openzfs/module/zfs/spa.c
index 96daf51b696a..ec2b674fb7ee 100644
--- a/sys/contrib/openzfs/module/zfs/spa.c
+++ b/sys/contrib/openzfs/module/zfs/spa.c
@@ -208,7 +208,7 @@ static const uint_t	zio_taskq_basedc = 80;	  /* base duty cycle */
 static const boolean_t spa_create_process = B_TRUE; /* no process => no sysdc */
 #endif
 
-static uint_t	zio_taskq_wr_iss_ncpus = 0;
+static uint_t	zio_taskq_write_tpq = 16;
 
 /*
  * Report any spa_load_verify errors found, but do not fail spa_load.
@@ -1067,17 +1067,16 @@ spa_taskqs_init(spa_t *spa, zio_type_t t, zio_taskq_type_t q)
 	case ZTI_MODE_SYNC:
 
 		/*
-		 * Create one wr_iss taskq for every 'zio_taskq_wr_iss_ncpus',
-		 * not to exceed the number of spa allocators.
+		 * Create one wr_iss taskq for every 'zio_taskq_write_tpq' CPUs,
+		 * not to exceed the number of spa allocators, and align to it.
 		 */
-		if (zio_taskq_wr_iss_ncpus == 0) {
-			count = MAX(boot_ncpus / spa->spa_alloc_count, 1);
-		} else {
-			count = MAX(1,
-			    boot_ncpus / MAX(1, zio_taskq_wr_iss_ncpus));
-		}
+		cpus = MAX(1, boot_ncpus * zio_taskq_batch_pct / 100);
+		count = MAX(1, cpus / MAX(1, zio_taskq_write_tpq));
 		count = MAX(count, (zio_taskq_batch_pct + 99) / 100);
 		count = MIN(count, spa->spa_alloc_count);
+		while (spa->spa_alloc_count % count != 0 &&
+		    spa->spa_alloc_count < count * 2)
+			count--;
 
 		/*
 		 * zio_taskq_batch_pct is unbounded and may exceed 100%, but no
@@ -1495,15 +1494,11 @@ spa_taskq_dispatch_select(spa_t *spa, zio_type_t t, zio_taskq_type_t q,
 	ASSERT3P(tqs->stqs_taskq, !=, NULL);
 	ASSERT3U(tqs->stqs_count, !=, 0);
 
-	if ((t == ZIO_TYPE_WRITE) && (q == ZIO_TASKQ_ISSUE) &&
-	    (zio != NULL) && (zio->io_wr_iss_tq != NULL)) {
-		/* dispatch to assigned write issue taskq */
-		tq = zio->io_wr_iss_tq;
-		return (tq);
-	}
-
 	if (tqs->stqs_count == 1) {
 		tq = tqs->stqs_taskq[0];
+	} else if ((t == ZIO_TYPE_WRITE) && (q == ZIO_TASKQ_ISSUE) &&
+	    (zio != NULL) && ZIO_HAS_ALLOCATOR(zio)) {
+		tq = tqs->stqs_taskq[zio->io_allocator % tqs->stqs_count];
 	} else {
 		tq = tqs->stqs_taskq[((uint64_t)gethrtime()) % tqs->stqs_count];
 	}
@@ -3594,11 +3589,16 @@ spa_activity_check_duration(spa_t *spa, uberblock_t *ub)
 }
 
 /*
- * Perform the import activity check.  If the user canceled the import or
- * we detected activity then fail.
+ * Remote host activity check.
+ *
+ * error results:
+ *          0 - no activity detected
+ *  EREMOTEIO - remote activity detected
+ *      EINTR - user canceled the operation
  */
 static int
-spa_activity_check(spa_t *spa, uberblock_t *ub, nvlist_t *config)
+spa_activity_check(spa_t *spa, uberblock_t *ub, nvlist_t *config,
+    boolean_t importing)
 {
 	uint64_t txg = ub->ub_txg;
 	uint64_t timestamp = ub->ub_timestamp;
@@ -3643,19 +3643,23 @@ spa_activity_check(spa_t *spa, uberblock_t *ub, nvlist_t *config)
 
 	import_expire = gethrtime() + import_delay;
 
-	spa_import_progress_set_notes(spa, "Checking MMP activity, waiting "
-	    "%llu ms", (u_longlong_t)NSEC2MSEC(import_delay));
+	if (importing) {
+		spa_import_progress_set_notes(spa, "Checking MMP activity, "
+		    "waiting %llu ms", (u_longlong_t)NSEC2MSEC(import_delay));
+	}
 
-	int interations = 0;
+	int iterations = 0;
 	while ((now = gethrtime()) < import_expire) {
-		if (interations++ % 30 == 0) {
+		if (importing && iterations++ % 30 == 0) {
 			spa_import_progress_set_notes(spa, "Checking MMP "
 			    "activity, %llu ms remaining",
 			    (u_longlong_t)NSEC2MSEC(import_expire - now));
 		}
 
-		(void) spa_import_progress_set_mmp_check(spa_guid(spa),
-		    NSEC2SEC(import_expire - gethrtime()));
+		if (importing) {
+			(void) spa_import_progress_set_mmp_check(spa_guid(spa),
+			    NSEC2SEC(import_expire - gethrtime()));
+		}
 
 		vdev_uberblock_load(rvd, ub, &mmp_label);
 
@@ -3737,6 +3741,61 @@ out:
 	return (error);
 }
 
+/*
+ * Called from zfs_ioc_clear for a pool that was suspended
+ * after failing mmp write checks.
+ */
+boolean_t
+spa_mmp_remote_host_activity(spa_t *spa)
+{
+	ASSERT(spa_multihost(spa) && spa_suspended(spa));
+
+	nvlist_t *best_label;
+	uberblock_t best_ub;
+
+	/*
+	 * Locate the best uberblock on disk
+	 */
+	vdev_uberblock_load(spa->spa_root_vdev, &best_ub, &best_label);
+	if (best_label) {
+		/*
+		 * confirm that the best hostid matches our hostid
+		 */
+		if (nvlist_exists(best_label, ZPOOL_CONFIG_HOSTID) &&
+		    spa_get_hostid(spa) !=
+		    fnvlist_lookup_uint64(best_label, ZPOOL_CONFIG_HOSTID)) {
+			nvlist_free(best_label);
+			return (B_TRUE);
+		}
+		nvlist_free(best_label);
+	} else {
+		return (B_TRUE);
+	}
+
+	if (!MMP_VALID(&best_ub) ||
+	    !MMP_FAIL_INT_VALID(&best_ub) ||
+	    MMP_FAIL_INT(&best_ub) == 0) {
+		return (B_TRUE);
+	}
+
+	if (best_ub.ub_txg != spa->spa_uberblock.ub_txg ||
+	    best_ub.ub_timestamp != spa->spa_uberblock.ub_timestamp) {
+		zfs_dbgmsg("txg mismatch detected during pool clear "
+		    "txg %llu ub_txg %llu timestamp %llu ub_timestamp %llu",
+		    (u_longlong_t)spa->spa_uberblock.ub_txg,
+		    (u_longlong_t)best_ub.ub_txg,
+		    (u_longlong_t)spa->spa_uberblock.ub_timestamp,
+		    (u_longlong_t)best_ub.ub_timestamp);
+		return (B_TRUE);
+	}
+
+	/*
+	 * Perform an activity check looking for any remote writer
+	 */
+	return (spa_activity_check(spa, &spa->spa_uberblock, spa->spa_config,
+	    B_FALSE) != 0);
+}
+
 static int
 spa_verify_host(spa_t *spa, nvlist_t *mos_config)
 {
@@ -4063,7 +4122,8 @@ spa_ld_select_uberblock(spa_t *spa, spa_import_type_t type)
 			return (spa_vdev_err(rvd, VDEV_AUX_ACTIVE, EREMOTEIO));
 		}
 
-		int error = spa_activity_check(spa, ub, spa->spa_config);
+		int error =
+		    spa_activity_check(spa, ub, spa->spa_config, B_TRUE);
 		if (error) {
 			nvlist_free(label);
 			return (error);
@@ -8771,15 +8831,16 @@ spa_async_remove(spa_t *spa, vdev_t *vd)
 }
 
 static void
-spa_async_probe(spa_t *spa, vdev_t *vd)
+spa_async_fault_vdev(spa_t *spa, vdev_t *vd)
 {
-	if (vd->vdev_probe_wanted) {
-		vd->vdev_probe_wanted = B_FALSE;
-		vdev_reopen(vd);	/* vdev_open() does the actual probe */
+	if (vd->vdev_fault_wanted) {
+		vd->vdev_fault_wanted = B_FALSE;
+		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
+		    VDEV_AUX_ERR_EXCEEDED);
 	}
 
 	for (int c = 0; c < vd->vdev_children; c++)
-		spa_async_probe(spa, vd->vdev_child[c]);
+		spa_async_fault_vdev(spa, vd->vdev_child[c]);
 }
 
 static void
@@ -8867,11 +8928,11 @@ spa_async_thread(void *arg)
 	}
 
 	/*
-	 * See if any devices need to be probed.
+	 * See if any devices need to be marked faulted.
 	 */
-	if (tasks & SPA_ASYNC_PROBE) {
+	if (tasks & SPA_ASYNC_FAULT_VDEV) {
 		spa_vdev_state_enter(spa, SCL_NONE);
-		spa_async_probe(spa, spa->spa_root_vdev);
+		spa_async_fault_vdev(spa, spa->spa_root_vdev);
 		(void) spa_vdev_state_exit(spa, NULL, 0);
 	}
 
@@ -10167,16 +10228,10 @@ spa_sync_tq_create(spa_t *spa, const char *name)
 	VERIFY(spa->spa_sync_tq != NULL);
 	VERIFY(kthreads != NULL);
 
-	spa_taskqs_t *tqs =
-	    &spa->spa_zio_taskq[ZIO_TYPE_WRITE][ZIO_TASKQ_ISSUE];
-
 	spa_syncthread_info_t *ti = spa->spa_syncthreads;
-	for (int i = 0, w = 0; i < nthreads; i++, w++, ti++) {
+	for (int i = 0; i < nthreads; i++, ti++) {
 		ti->sti_thread = kthreads[i];
-		if (w == tqs->stqs_count) {
-			w = 0;
-		}
-		ti->sti_wr_iss_tq = tqs->stqs_taskq[w];
+		ti->sti_allocator = i;
 	}
 
 	kmem_free(kthreads, sizeof (*kthreads) * nthreads);
@@ -10195,6 +10250,42 @@ spa_sync_tq_destroy(spa_t *spa)
 	spa->spa_sync_tq = NULL;
 }
 
+uint_t
+spa_acq_allocator(spa_t *spa)
+{
+	int i;
+
+	if (spa->spa_alloc_count == 1)
+		return (0);
+
+	mutex_enter(&spa->spa_allocs_use->sau_lock);
+	uint_t r = spa->spa_allocs_use->sau_rotor;
+	do {
+		if (++r == spa->spa_alloc_count)
+			r = 0;
+	} while (spa->spa_allocs_use->sau_inuse[r]);
+	spa->spa_allocs_use->sau_inuse[r] = B_TRUE;
+	spa->spa_allocs_use->sau_rotor = r;
+	mutex_exit(&spa->spa_allocs_use->sau_lock);
+
+	spa_syncthread_info_t *ti = spa->spa_syncthreads;
+	for (i = 0; i < spa->spa_alloc_count; i++, ti++) {
+		if (ti->sti_thread == curthread) {
+			ti->sti_allocator = r;
+			break;
+		}
+	}
+	ASSERT3S(i, <, spa->spa_alloc_count);
+	return (r);
+}
+
+void
+spa_rel_allocator(spa_t *spa, uint_t allocator)
+{
+	if (spa->spa_alloc_count > 1)
+		spa->spa_allocs_use->sau_inuse[allocator] = B_FALSE;
+}
+
 void
 spa_select_allocator(zio_t *zio)
 {
@@ -10222,8 +10313,7 @@ spa_select_allocator(zio_t *zio)
 		spa_syncthread_info_t *ti = spa->spa_syncthreads;
 		for (int i = 0; i < spa->spa_alloc_count; i++, ti++) {
 			if (ti->sti_thread == curthread) {
-				zio->io_allocator = i;
-				zio->io_wr_iss_tq = ti->sti_wr_iss_tq;
+				zio->io_allocator = ti->sti_allocator;
 				return;
 			}
 		}
@@ -10240,7 +10330,6 @@ spa_select_allocator(zio_t *zio)
 	    bm->zb_blkid >> 20);
 
 	zio->io_allocator = (uint_t)hv % spa->spa_alloc_count;
-	zio->io_wr_iss_tq = NULL;
 }
 
 /*
@@ -10811,10 +10900,10 @@ ZFS_MODULE_PARAM(zfs_spa, spa_, load_verify_data, INT, ZMOD_RW,
 ZFS_MODULE_PARAM(zfs_spa, spa_, load_print_vdev_tree, INT, ZMOD_RW,
 	"Print vdev tree to zfs_dbgmsg during pool import");
 
-ZFS_MODULE_PARAM(zfs_zio, zio_, taskq_batch_pct, UINT, ZMOD_RD,
+ZFS_MODULE_PARAM(zfs_zio, zio_, taskq_batch_pct, UINT, ZMOD_RW,
 	"Percentage of CPUs to run an IO worker thread");
 
-ZFS_MODULE_PARAM(zfs_zio, zio_, taskq_batch_tpq, UINT, ZMOD_RD,
+ZFS_MODULE_PARAM(zfs_zio, zio_, taskq_batch_tpq, UINT, ZMOD_RW,
 	"Number of threads per IO worker taskqueue");
 
 /* BEGIN CSTYLED */
@@ -10845,13 +10934,13 @@ ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, new_alloc, INT,
 
 #ifdef _KERNEL
 ZFS_MODULE_VIRTUAL_PARAM_CALL(zfs_zio, zio_, taskq_read,
-	spa_taskq_read_param_set, spa_taskq_read_param_get, ZMOD_RD,
+	spa_taskq_read_param_set, spa_taskq_read_param_get, ZMOD_RW,
 	"Configure IO queues for read IO");
 ZFS_MODULE_VIRTUAL_PARAM_CALL(zfs_zio, zio_, taskq_write,
-	spa_taskq_write_param_set, spa_taskq_write_param_get, ZMOD_RD,
+	spa_taskq_write_param_set, spa_taskq_write_param_get, ZMOD_RW,
 	"Configure IO queues for write IO");
 #endif
 /* END CSTYLED */
 
-ZFS_MODULE_PARAM(zfs_zio, zio_, taskq_wr_iss_ncpus, UINT, ZMOD_RW,
-	"Number of CPUs to run write issue taskqs");
+ZFS_MODULE_PARAM(zfs_zio, zio_, taskq_write_tpq, UINT, ZMOD_RW,
+	"Number of CPUs per write issue taskq");
diff --git a/sys/contrib/openzfs/module/zfs/spa_misc.c b/sys/contrib/openzfs/module/zfs/spa_misc.c
index 5fb7847b5d8b..e6d4a9bdb29c 100644
--- a/sys/contrib/openzfs/module/zfs/spa_misc.c
+++ b/sys/contrib/openzfs/module/zfs/spa_misc.c
@@ -394,6 +394,7 @@ static const uint64_t spa_max_slop = 128ULL * 1024 * 1024 * 1024;
  * Number of allocators to use, per spa instance
  */
 static int spa_num_allocators = 4;
+static int spa_cpus_per_allocator = 4;
 
 /*
  * Spa active allocator.
@@ -747,8 +748,9 @@ spa_add(const char *name, nvlist_t *config, const char *altroot)
 	if (altroot)
 		spa->spa_root = spa_strdup(altroot);
 
-	/* Do not allow more allocators than CPUs. */
-	spa->spa_alloc_count = MIN(MAX(spa_num_allocators, 1), boot_ncpus);
+	/* Do not allow more allocators than fraction of CPUs. */
+	spa->spa_alloc_count = MAX(MIN(spa_num_allocators,
+	    boot_ncpus / MAX(spa_cpus_per_allocator, 1)), 1);
 
 	spa->spa_allocs = kmem_zalloc(spa->spa_alloc_count *
 	    sizeof (spa_alloc_t), KM_SLEEP);
@@ -758,6 +760,12 @@ spa_add(const char *name, nvlist_t *config, const char *altroot)
 		avl_create(&spa->spa_allocs[i].spaa_tree, zio_bookmark_compare,
 		    sizeof (zio_t), offsetof(zio_t, io_queue_node.a));
 	}
+	if (spa->spa_alloc_count > 1) {
+		spa->spa_allocs_use = kmem_zalloc(offsetof(spa_allocs_use_t,
+		    sau_inuse[spa->spa_alloc_count]), KM_SLEEP);
+		mutex_init(&spa->spa_allocs_use->sau_lock, NULL, MUTEX_DEFAULT,
+		    NULL);
+	}
 
 	avl_create(&spa->spa_metaslabs_by_flushed, metaslab_sort_by_flushed,
 	    sizeof (metaslab_t), offsetof(metaslab_t, ms_spa_txg_node));
@@ -853,6 +861,11 @@ spa_remove(spa_t *spa)
 	}
 	kmem_free(spa->spa_allocs, spa->spa_alloc_count *
 	    sizeof (spa_alloc_t));
+	if (spa->spa_alloc_count > 1) {
+		mutex_destroy(&spa->spa_allocs_use->sau_lock);
+		kmem_free(spa->spa_allocs_use, offsetof(spa_allocs_use_t,
+		    sau_inuse[spa->spa_alloc_count]));
+	}
 
 	avl_destroy(&spa->spa_metaslabs_by_flushed);
 	avl_destroy(&spa->spa_sm_logs_by_txg);
@@ -3097,4 +3110,7 @@ ZFS_MODULE_PARAM_CALL(zfs_spa, spa_, slop_shift, param_set_slop_shift,
 	param_get_uint, ZMOD_RW, "Reserved free space in pool");
 
 ZFS_MODULE_PARAM(zfs, spa_, num_allocators, INT, ZMOD_RW,
-	"Number of allocators per spa, capped by ncpus");
+	"Number of allocators per spa");
+
+ZFS_MODULE_PARAM(zfs, spa_, cpus_per_allocator, INT, ZMOD_RW,
+	"Minimum number of CPUs per allocators");
diff --git a/sys/contrib/openzfs/module/zfs/txg.c b/sys/contrib/openzfs/module/zfs/txg.c
index a67c043446f5..5ce6be69be14 100644
--- a/sys/contrib/openzfs/module/zfs/txg.c
+++ b/sys/contrib/openzfs/module/zfs/txg.c
@@ -551,6 +551,15 @@ txg_sync_thread(void *arg)
 		}
 
 		/*
+		 * When we're suspended, nothing should be changing and for
+		 * MMP we don't want to bump anything that would make it
+		 * harder to detect if another host is changing it when
+		 * resuming after a MMP suspend.
+		 */
+		if (spa_suspended(spa))
+			continue;
+
+		/*
 		 * Wait until the quiesce thread hands off a txg to us,
 		 * prompting it to do so if necessary.
 		 */
diff --git a/sys/contrib/openzfs/module/zfs/vdev.c b/sys/contrib/openzfs/module/zfs/vdev.c
index d97d0a8100c2..c5551eb6cf6e 100644
--- a/sys/contrib/openzfs/module/zfs/vdev.c
+++ b/sys/contrib/openzfs/module/zfs/vdev.c
@@ -1664,6 +1664,7 @@ vdev_metaslab_fini(vdev_t *vd)
 typedef struct vdev_probe_stats {
 	boolean_t	vps_readable;
 	boolean_t	vps_writeable;
+	boolean_t	vps_zio_done_probe;
 	int		vps_flags;
 } vdev_probe_stats_t;
 
@@ -1709,6 +1710,17 @@ vdev_probe_done(zio_t *zio)
 			(void) zfs_ereport_post(FM_EREPORT_ZFS_PROBE_FAILURE,
 			    spa, vd, NULL, NULL, 0);
 			zio->io_error = SET_ERROR(ENXIO);
+
+			/*
+			 * If this probe was initiated from zio pipeline, then
+			 * change the state in a spa_async_request. Probes that
+			 * were initiated from a vdev_open can change the state
+			 * as part of the open call.
+			 */
+			if (vps->vps_zio_done_probe) {
+				vd->vdev_fault_wanted = B_TRUE;
+				spa_async_request(spa, SPA_ASYNC_FAULT_VDEV);
+			}
 		}
 
 		mutex_enter(&vd->vdev_probe_lock);
@@ -1759,6 +1771,7 @@ vdev_probe(vdev_t *vd, zio_t *zio)
 
 		vps->vps_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_PROBE |
 		    ZIO_FLAG_DONT_AGGREGATE | ZIO_FLAG_TRYHARD;
+		vps->vps_zio_done_probe = (zio != NULL);
 
 		if (spa_config_held(spa, SCL_ZIO, RW_WRITER)) {
 			/*
@@ -1785,15 +1798,6 @@ vdev_probe(vdev_t *vd, zio_t *zio)
 		vd->vdev_probe_zio = pio = zio_null(NULL, spa, vd,
 		    vdev_probe_done, vps,
 		    vps->vps_flags | ZIO_FLAG_DONT_PROPAGATE);
-
-		/*
-		 * We can't change the vdev state in this context, so we
-		 * kick off an async task to do it on our behalf.
-		 */
-		if (zio != NULL) {
-			vd->vdev_probe_wanted = B_TRUE;
-			spa_async_request(spa, SPA_ASYNC_PROBE);
-		}
 	}
 
 	if (zio != NULL)
diff --git a/sys/contrib/openzfs/module/zfs/vdev_label.c b/sys/contrib/openzfs/module/zfs/vdev_label.c
index c31f48028bbc..ed592514fded 100644
--- a/sys/contrib/openzfs/module/zfs/vdev_label.c
+++ b/sys/contrib/openzfs/module/zfs/vdev_label.c
@@ -2027,6 +2027,7 @@ retry:
 	/*
 	 * If this isn't a resync due to I/O errors,
 	 * and nothing changed in this transaction group,
+	 * and multihost protection isn't enabled,
 	 * and the vdev configuration hasn't changed,
 	 * then there's nothing to do.
 	 */
@@ -2034,7 +2035,8 @@ retry:
 		boolean_t changed = uberblock_update(ub, spa->spa_root_vdev,
 		    txg, spa->spa_mmp.mmp_delay);
 
-		if (!changed && list_is_empty(&spa->spa_config_dirty_list))
+		if (!changed && list_is_empty(&spa->spa_config_dirty_list) &&
+		    !spa_multihost(spa))
 			return (0);
 	}
 
diff --git a/sys/contrib/openzfs/module/zfs/vdev_raidz.c b/sys/contrib/openzfs/module/zfs/vdev_raidz.c
index b03331ec69c6..de7d0fa79478 100644
--- a/sys/contrib/openzfs/module/zfs/vdev_raidz.c
+++ b/sys/contrib/openzfs/module/zfs/vdev_raidz.c
@@ -1891,8 +1891,9 @@ vdev_raidz_matrix_reconstruct(raidz_row_t *rr, int n, int nmissing,
 static void
 vdev_raidz_reconstruct_general(raidz_row_t *rr, int *tgts, int ntgts)
 {
-	int n, i, c, t, tt;
-	int nmissing_rows;
+	int i, c, t, tt;
+	unsigned int n;
+	unsigned int nmissing_rows;
 	int missing_rows[VDEV_RAIDZ_MAXPARITY];
 	int parity_map[VDEV_RAIDZ_MAXPARITY];
 	uint8_t *p, *pp;
diff --git a/sys/contrib/openzfs/module/zfs/zap.c b/sys/contrib/openzfs/module/zfs/zap.c
index da86defb445c..1b6b16fc6662 100644
--- a/sys/contrib/openzfs/module/zfs/zap.c
+++ b/sys/contrib/openzfs/module/zfs/zap.c
@@ -22,6 +22,8 @@
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  * Copyright (c) 2012, 2018 by Delphix. All rights reserved.
  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
+ * Copyright 2023 Alexander Stetsenko <alex.stetsenko@gmail.com>
+ * Copyright (c) 2023, Klara Inc.
  */
 
 /*
@@ -41,6 +43,7 @@
 
 #include <sys/spa.h>
 #include <sys/dmu.h>
+#include <sys/dnode.h>
 #include <sys/zfs_context.h>
 #include <sys/zfs_znode.h>
 #include <sys/fs/zfs.h>
@@ -78,9 +81,16 @@
  */
 static int zap_iterate_prefetch = B_TRUE;
 
+/*
+ * Enable ZAP shrinking. When enabled, empty sibling leaf blocks will be
+ * collapsed into a single block.
+ */
+int zap_shrink_enabled = B_TRUE;
+
 int fzap_default_block_shift = 14; /* 16k blocksize */
 
 static uint64_t zap_allocate_blocks(zap_t *zap, int nblocks);
+static int zap_shrink(zap_name_t *zn, zap_leaf_t *l, dmu_tx_t *tx);
 
 void
 fzap_byteswap(void *vbuf, size_t size)
@@ -587,6 +597,72 @@ zap_set_idx_to_blk(zap_t *zap, uint64_t idx, uint64_t blk, dmu_tx_t *tx)
 }
 
 static int
+zap_set_idx_range_to_blk(zap_t *zap, uint64_t idx, uint64_t nptrs, uint64_t blk,
+    dmu_tx_t *tx)
+{
+	int bs = FZAP_BLOCK_SHIFT(zap);
+	int epb = bs >> 3; /* entries per block */
+	int err = 0;
+
+	ASSERT(tx != NULL);
+	ASSERT(RW_WRITE_HELD(&zap->zap_rwlock));
+
+	/*
+	 * Check for i/o errors
+	 */
+	for (int i = 0; i < nptrs; i += epb) {
+		uint64_t blk;
+		err = zap_idx_to_blk(zap, idx + i, &blk);
+		if (err != 0) {
+			return (err);
+		}
+	}
+
+	for (int i = 0; i < nptrs; i++) {
+		err = zap_set_idx_to_blk(zap, idx + i, blk, tx);
+		ASSERT0(err); /* we checked for i/o errors above */
+		if (err != 0)
+			break;
+	}
+
+	return (err);
+}
+
+#define	ZAP_PREFIX_HASH(pref, pref_len)	((pref) << (64 - (pref_len)))
+
+/*
+ * Each leaf has single range of entries (block pointers) in the ZAP ptrtbl.
+ * If two leaves are siblings, their ranges are adjecent and contain the same
+ * number of entries. In order to find out if a leaf has a sibling, we need to
+ * check the range corresponding to the sibling leaf. There is no need to check
+ * all entries in the range, we only need to check the frist and the last one.
+ */
+static uint64_t
+check_sibling_ptrtbl_range(zap_t *zap, uint64_t prefix, uint64_t prefix_len)
+{
+	ASSERT(RW_LOCK_HELD(&zap->zap_rwlock));
+
+	uint64_t h = ZAP_PREFIX_HASH(prefix, prefix_len);
+	uint64_t idx = ZAP_HASH_IDX(h, zap_f_phys(zap)->zap_ptrtbl.zt_shift);
+	uint64_t pref_diff = zap_f_phys(zap)->zap_ptrtbl.zt_shift - prefix_len;
+	uint64_t nptrs = (1 << pref_diff);
+	uint64_t first;
+	uint64_t last;
+
+	ASSERT3U(idx+nptrs, <=, (1UL << zap_f_phys(zap)->zap_ptrtbl.zt_shift));
+
+	if (zap_idx_to_blk(zap, idx, &first) != 0)
+		return (0);
+
+	if (zap_idx_to_blk(zap, idx + nptrs - 1, &last) != 0)
+		return (0);
+
+	if (first != last)
+		return (0);
+	return (first);
+}
+
+static int
 zap_deref_leaf(zap_t *zap, uint64_t h, dmu_tx_t *tx, krw_t lt, zap_leaf_t **lp)
 {
 	uint64_t blk;
@@ -958,6 +1034,10 @@ fzap_remove(zap_name_t *zn, dmu_tx_t *tx)
 	if (err == 0) {
 		zap_entry_remove(&zeh);
 		zap_increment_num_entries(zn->zn_zap, -1, tx);
+
+		if (zap_leaf_phys(l)->l_hdr.lh_nentries == 0 &&
+		    zap_shrink_enabled)
+			return (zap_shrink(zn, l, tx));
 	}
 	zap_put_leaf(l);
 	return (err);
@@ -1222,13 +1302,19 @@ fzap_cursor_retrieve(zap_t *zap, zap_cursor_t *zc, zap_attribute_t *za)
 		    ZIO_PRIORITY_ASYNC_READ);
 	}
 
-	if (zc->zc_leaf &&
-	    (ZAP_HASH_IDX(zc->zc_hash,
-	    zap_leaf_phys(zc->zc_leaf)->l_hdr.lh_prefix_len) !=
-	    zap_leaf_phys(zc->zc_leaf)->l_hdr.lh_prefix)) {
+	if (zc->zc_leaf) {
 		rw_enter(&zc->zc_leaf->l_rwlock, RW_READER);
-		zap_put_leaf(zc->zc_leaf);
-		zc->zc_leaf = NULL;
+
+		/*
+		 * The leaf was either shrunk or split.
+		 */
+		if ((zap_leaf_phys(zc->zc_leaf)->l_hdr.lh_block_type == 0) ||
+		    (ZAP_HASH_IDX(zc->zc_hash,
+		    zap_leaf_phys(zc->zc_leaf)->l_hdr.lh_prefix_len) !=
+		    zap_leaf_phys(zc->zc_leaf)->l_hdr.lh_prefix)) {
+			zap_put_leaf(zc->zc_leaf);
+			zc->zc_leaf = NULL;
+		}
 	}
 
 again:
@@ -1237,8 +1323,6 @@ again:
 		    &zc->zc_leaf);
 		if (err != 0)
 			return (err);
-	} else {
-		rw_enter(&zc->zc_leaf->l_rwlock, RW_READER);
 	}
 	l = zc->zc_leaf;
 
@@ -1367,6 +1451,242 @@ fzap_get_stats(zap_t *zap, zap_stats_t *zs)
 	}
 }
 
+/*
+ * Find last allocated block and update freeblk.
+ */
+static void
+zap_trunc(zap_t *zap)
+{
+	uint64_t nentries;
+	uint64_t lastblk;
+
+	ASSERT(RW_WRITE_HELD(&zap->zap_rwlock));
+
+	if (zap_f_phys(zap)->zap_ptrtbl.zt_blk > 0) {
+		/* External ptrtbl */
+		nentries = (1 << zap_f_phys(zap)->zap_ptrtbl.zt_shift);
+		lastblk = zap_f_phys(zap)->zap_ptrtbl.zt_blk +
+		    zap_f_phys(zap)->zap_ptrtbl.zt_numblks - 1;
+	} else {
+		/* Embedded ptrtbl */
+		nentries = (1 << ZAP_EMBEDDED_PTRTBL_SHIFT(zap));
+		lastblk = 0;
+	}
+
+	for (uint64_t idx = 0; idx < nentries; idx++) {
+		uint64_t blk;
+		if (zap_idx_to_blk(zap, idx, &blk) != 0)
+			return;
+		if (blk > lastblk)
+			lastblk = blk;
+	}
+
+	ASSERT3U(lastblk, <, zap_f_phys(zap)->zap_freeblk);
+
+	zap_f_phys(zap)->zap_freeblk = lastblk + 1;
+}
+
+/*
+ * ZAP shrinking algorithm.
+ *
+ * We shrink ZAP recuresively removing empty leaves. We can remove an empty leaf
+ * only if it has a sibling. Sibling leaves have the same prefix length and
+ * their prefixes differ only by the least significant (sibling) bit. We require
+ * both siblings to be empty. This eliminates a need to rehash the non-empty
+ * remaining leaf. When we have removed one of two empty sibling, we set ptrtbl
+ * entries of the removed leaf to point out to the remaining leaf. Prefix length
+ * of the remaining leaf is decremented. As a result, it has a new prefix and it
+ * might have a new sibling. So, we repeat the process.
+ *
+ * Steps:
+ * 1. Check if a sibling leaf (sl) exists and it is empty.
+ * 2. Release the leaf (l) if it has the sibling bit (slbit) equal to 1.
+ * 3. Release the sibling (sl) to derefer it again with WRITER lock.
+ * 4. Upgrade zapdir lock to WRITER (once).
+ * 5. Derefer released leaves again.
+ * 6. If it is needed, recheck whether both leaves are still siblings and empty.
+ * 7. Set ptrtbl pointers of the removed leaf (slbit 1) to point out to blkid of
+ * the remaining leaf (slbit 0).
+ * 8. Free disk block of the removed leaf (dmu_free_range).
+ * 9. Decrement prefix_len of the remaining leaf.
+ * 10. Repeat the steps.
+ */
+static int
+zap_shrink(zap_name_t *zn, zap_leaf_t *l, dmu_tx_t *tx)
+{
+	zap_t *zap = zn->zn_zap;
+	int64_t zt_shift = zap_f_phys(zap)->zap_ptrtbl.zt_shift;
+	uint64_t hash = zn->zn_hash;
+	uint64_t prefix = zap_leaf_phys(l)->l_hdr.lh_prefix;
+	uint64_t prefix_len = zap_leaf_phys(l)->l_hdr.lh_prefix_len;
+	boolean_t trunc = B_FALSE;
+	int err = 0;
+
+	ASSERT3U(zap_leaf_phys(l)->l_hdr.lh_nentries, ==, 0);
+	ASSERT3U(prefix_len, <=, zap_f_phys(zap)->zap_ptrtbl.zt_shift);
+	ASSERT(RW_LOCK_HELD(&zap->zap_rwlock));
+	ASSERT3U(ZAP_HASH_IDX(hash, prefix_len), ==, prefix);
+
+	boolean_t writer = B_FALSE;
+
+	/*
+	 * To avoid deadlock always deref leaves in the same order -
+	 * sibling 0 first, then sibling 1.
+	 */
+	while (prefix_len) {
+		zap_leaf_t *sl;
+		int64_t prefix_diff = zt_shift - prefix_len;
+		uint64_t sl_prefix = prefix ^ 1;
+		uint64_t sl_hash = ZAP_PREFIX_HASH(sl_prefix, prefix_len);
+		int slbit = prefix & 1;
+
+		ASSERT3U(zap_leaf_phys(l)->l_hdr.lh_nentries, ==, 0);
+
+		/*
+		 * Check if there is a sibling by reading ptrtbl ptrs.
+		 */
+		if (check_sibling_ptrtbl_range(zap, sl_prefix, prefix_len) == 0)
+			break;
+
+		/*
+		 * sibling 1, unlock it - we haven't yet dereferenced sibling 0.
+		 */
+		if (slbit == 1) {
+			zap_put_leaf(l);
+			l = NULL;
+		}
+
+		/*
+		 * Dereference sibling leaf and check if it is empty.
+		 */
+		if ((err = zap_deref_leaf(zap, sl_hash, tx, RW_READER,
+		    &sl)) != 0)
+			break;
+
+		ASSERT3U(ZAP_HASH_IDX(sl_hash, prefix_len), ==, sl_prefix);
+
+		/*
+		 * Check if we have a sibling and it is empty.
+		 */
+		if (zap_leaf_phys(sl)->l_hdr.lh_prefix_len != prefix_len ||
+		    zap_leaf_phys(sl)->l_hdr.lh_nentries != 0) {
+			zap_put_leaf(sl);
+			break;
+		}
+
+		zap_put_leaf(sl);
+
+		/*
+		 * If there two empty sibling, we have work to do, so
+		 * we need to lock ZAP ptrtbl as WRITER.
+		 */
+		if (!writer && (writer = zap_tryupgradedir(zap, tx)) == 0) {
+			/* We failed to upgrade */
+			if (l != NULL) {
+				zap_put_leaf(l);
+				l = NULL;
+			}
+
+			/*
+			 * Usually, the right way to upgrade from a READER lock
+			 * to a WRITER lock is to call zap_unlockdir() and
+			 * zap_lockdir(), but we do not have a tag. Instead,
+			 * we do it in more sophisticated way.
+			 */
+			rw_exit(&zap->zap_rwlock);
+			rw_enter(&zap->zap_rwlock, RW_WRITER);
+			dmu_buf_will_dirty(zap->zap_dbuf, tx);
+
+			zt_shift = zap_f_phys(zap)->zap_ptrtbl.zt_shift;
+			writer = B_TRUE;
+		}
+
+		/*
+		 * Here we have WRITER lock for ptrtbl.
+		 * Now, we need a WRITER lock for both siblings leaves.
+		 * Also, we have to recheck if the leaves are still siblings
+		 * and still empty.
+		 */
+		if (l == NULL) {
+			/* sibling 0 */
+			if ((err = zap_deref_leaf(zap, (slbit ? sl_hash : hash),
+			    tx, RW_WRITER, &l)) != 0)
+				break;
+
+			/*
+			 * The leaf isn't empty anymore or
+			 * it was shrunk/split while our locks were down.
+			 */
+			if (zap_leaf_phys(l)->l_hdr.lh_nentries != 0 ||
+			    zap_leaf_phys(l)->l_hdr.lh_prefix_len != prefix_len)
+				break;
+		}
+
+		/* sibling 1 */
+		if ((err = zap_deref_leaf(zap, (slbit ? hash : sl_hash), tx,
+		    RW_WRITER, &sl)) != 0)
+			break;
+
+		/*
+		 * The leaf isn't empty anymore or
+		 * it was shrunk/split while our locks were down.
+		 */
+		if (zap_leaf_phys(sl)->l_hdr.lh_nentries != 0 ||
+		    zap_leaf_phys(sl)->l_hdr.lh_prefix_len != prefix_len) {
+			zap_put_leaf(sl);
+			break;
+		}
+
+		/* If we have gotten here, we have a leaf to collapse */
+		uint64_t idx = (slbit ? prefix : sl_prefix) << prefix_diff;
+		uint64_t nptrs = (1ULL << prefix_diff);
+		uint64_t sl_blkid = sl->l_blkid;
+
+		/*
+		 * Set ptrtbl entries to point out to the slibling 0 blkid
+		 */
+		if ((err = zap_set_idx_range_to_blk(zap, idx, nptrs, l->l_blkid,
+		    tx)) != 0) {
+			zap_put_leaf(sl);
+			break;
+		}
+
+		/*
+		 * Free sibling 1 disk block.
+		 */
+		int bs = FZAP_BLOCK_SHIFT(zap);
+		if (sl_blkid == zap_f_phys(zap)->zap_freeblk - 1)
+			trunc = B_TRUE;
+
+		(void) dmu_free_range(zap->zap_objset, zap->zap_object,
+		    sl_blkid << bs, 1 << bs, tx);
+		zap_put_leaf(sl);
+
+		zap_f_phys(zap)->zap_num_leafs--;
+
+		/*
+		 * Update prefix and prefix_len.
+		 */
+		zap_leaf_phys(l)->l_hdr.lh_prefix >>= 1;
+		zap_leaf_phys(l)->l_hdr.lh_prefix_len--;
+
+		prefix = zap_leaf_phys(l)->l_hdr.lh_prefix;
+		prefix_len = zap_leaf_phys(l)->l_hdr.lh_prefix_len;
+	}
+
+	if (trunc)
+		zap_trunc(zap);
+
+	if (l != NULL)
+		zap_put_leaf(l);
+
+	return (err);
+}
+
 /* CSTYLED */
 ZFS_MODULE_PARAM(zfs, , zap_iterate_prefetch, INT, ZMOD_RW,
 	"When iterating ZAP object, prefetch it");
+
+/* CSTYLED */
+ZFS_MODULE_PARAM(zfs, , zap_shrink_enabled, INT, ZMOD_RW,
+	"Enable ZAP shrinking");
diff --git a/sys/contrib/openzfs/module/zfs/zfs_ioctl.c b/sys/contrib/openzfs/module/zfs/zfs_ioctl.c
index 2ac1e34dccec..908b9efc1813 100644
--- a/sys/contrib/openzfs/module/zfs/zfs_ioctl.c
+++ b/sys/contrib/openzfs/module/zfs/zfs_ioctl.c
@@ -5823,10 +5823,13 @@ zfs_ioc_clear(zfs_cmd_t *zc)
 
 	/*
 	 * If multihost is enabled, resuming I/O is unsafe as another
-	 * host may have imported the pool.
+	 * host may have imported the pool. Check for remote activity.
 	 */
-	if (spa_multihost(spa) && spa_suspended(spa))
-		return (SET_ERROR(EINVAL));
+	if (spa_multihost(spa) && spa_suspended(spa) &&
+	    spa_mmp_remote_host_activity(spa)) {
+		spa_close(spa, FTAG);
+		return (SET_ERROR(EREMOTEIO));
+	}
 
 	spa_vdev_state_enter(spa, SCL_NONE);
 
diff --git a/sys/contrib/openzfs/module/zfs/zio.c b/sys/contrib/openzfs/module/zfs/zio.c
index 1ba99f4d4624..870343bf4fa3 100644
--- a/sys/contrib/openzfs/module/zfs/zio.c
+++ b/sys/contrib/openzfs/module/zfs/zio.c
@@ -803,9 +803,10 @@ zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait,
 
 		/*
 		 * If we can tell the caller to execute this parent next, do
-		 * so. We only do this if the parent's zio type matches the
-		 * child's type. Otherwise dispatch the parent zio in its
-		 * own taskq.
+		 * so. We do this if the parent's zio type matches the child's
+		 * type, or if it's a zio_null() with no done callback, and so
+		 * has no actual work to do. Otherwise dispatch the parent zio
+		 * in its own taskq.
 		 *
 		 * Having the caller execute the parent when possible reduces
 		 * locking on the zio taskq's, reduces context switch
@@ -825,7 +826,8 @@ zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait,
 		 * of writes for spa_sync(), and the chain of ZIL blocks.
 		 */
 		if (next_to_executep != NULL && *next_to_executep == NULL &&
-		    pio->io_type == zio->io_type) {
+		    (pio->io_type == zio->io_type ||
+		    (pio->io_type == ZIO_TYPE_NULL && !pio->io_done))) {
 			*next_to_executep = pio;
 		} else {
 			zio_taskq_dispatch(pio, type, B_FALSE);
@@ -2532,8 +2534,10 @@ zio_suspend(spa_t *spa, zio_t *zio, zio_suspend_reason_t reason)
 		    "failure and the failure mode property for this pool "
 		    "is set to panic.", spa_name(spa));
 
-	cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable I/O "
-	    "failure and has been suspended.\n", spa_name(spa));
+	if (reason != ZIO_SUSPEND_MMP) {
+		cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable "
+		    "I/O failure and has been suspended.\n", spa_name(spa));
+	}
 
 	(void) zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL,
 	    NULL, NULL, 0);
@@ -2921,7 +2925,6 @@ static void
 zio_gang_inherit_allocator(zio_t *pio, zio_t *cio)
 {
 	cio->io_allocator = pio->io_allocator;
-	cio->io_wr_iss_tq = pio->io_wr_iss_tq;
 }
 
 static void
diff --git a/sys/contrib/openzfs/module/zfs/zio_inject.c b/sys/contrib/openzfs/module/zfs/zio_inject.c
index 3773e400d799..012a0e3c6c17 100644
--- a/sys/contrib/openzfs/module/zfs/zio_inject.c
+++ b/sys/contrib/openzfs/module/zfs/zio_inject.c
@@ -607,9 +607,11 @@ zio_handle_io_delay(zio_t *zio)
 		if (vd->vdev_guid != handler->zi_record.zi_guid)
 			continue;
 
+		/* also match on I/O type (e.g., -T read) */
 		if (handler->zi_record.zi_iotype != ZIO_TYPES &&
-		    handler->zi_record.zi_iotype != zio->io_type)
-				continue;
+		    handler->zi_record.zi_iotype != zio->io_type) {
+			continue;
+		}
 
 		/*
 		 * Defensive; should never happen as the array allocation