Allow and prefer special vdevs as ZIL

Before this change ZIL blocks were allocated only from normal or
SLOG vdevs.  In typical situation when special vdevs are SSDs and
normal are HDDs it could cause weird inversions when data blocks
are written to SSDs, but ZIL referencing them to HDDs.

This change assumes that special vdevs typically have much better
(or at least not worse) latency than normal, and so in absence of
SLOGs should store ZIL blocks.  It means similar to normal vdevs
introduction of special embedded log allocation class and updating
the allocation fallback order to: SLOG -> special embedded log ->
special -> normal embedded log -> normal.

The code tries to guess whether data block is going to be written
to normal or special vdev (it can not be done precisely before
compression) and prefer indirect writes for blocks written to a
special vdev to avoid double-write.  For blocks that are going to
be written to normal vdev, special vdev by default plays as SLOG,
reducing write latency by the cost of higher special vdev wear,
but it is tunable via module parameter.

This should allow HDD pools with decent SSD as special vdev to
work under synchronous workloads without requiring additional
SLOG SSD, impractical in many scenarios.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Rob Norris <rob.norris@klarasystems.com>
Reviewed-by: Paul Dagnelie <paul.dagnelie@klarasystems.com>
Signed-off-by:	Alexander Motin <mav@FreeBSD.org>
Sponsored by:	iXsystems, Inc.
Closes #17505

module/zfs/metaslab.c

@@ -750,7 +750,8 @@ metaslab_class_histogram_verify(metaslab_class_t *mc)
 		}
 
 		IMPLY(mg == mg->mg_vd->vdev_log_mg,
-		    mc == spa_embedded_log_class(mg->mg_vd->vdev_spa));
+		    mc == spa_embedded_log_class(mg->mg_vd->vdev_spa) ||
+		    mc == spa_special_embedded_log_class(mg->mg_vd->vdev_spa));
 
 		for (i = 0; i < ZFS_RANGE_TREE_HISTOGRAM_SIZE; i++)
 			mc_hist[i] += mg->mg_histogram[i];
@@ -1288,7 +1289,8 @@ metaslab_group_histogram_add(metaslab_group_t *mg, metaslab_t *msp)
 	mutex_enter(&mc->mc_lock);
 	for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
 		IMPLY(mg == mg->mg_vd->vdev_log_mg,
-		    mc == spa_embedded_log_class(mg->mg_vd->vdev_spa));
+		    mc == spa_embedded_log_class(mg->mg_vd->vdev_spa) ||
+		    mc == spa_special_embedded_log_class(mg->mg_vd->vdev_spa));
 		mg->mg_histogram[i + ashift] +=
 		    msp->ms_sm->sm_phys->smp_histogram[i];
 		mc->mc_histogram[i + ashift] +=
@@ -1316,7 +1318,8 @@ metaslab_group_histogram_remove(metaslab_group_t *mg, metaslab_t *msp)
 		ASSERT3U(mc->mc_histogram[i + ashift], >=,
 		    msp->ms_sm->sm_phys->smp_histogram[i]);
 		IMPLY(mg == mg->mg_vd->vdev_log_mg,
-		    mc == spa_embedded_log_class(mg->mg_vd->vdev_spa));
+		    mc == spa_embedded_log_class(mg->mg_vd->vdev_spa) ||
+		    mc == spa_special_embedded_log_class(mg->mg_vd->vdev_spa));
 
 		mg->mg_histogram[i + ashift] -=
 		    msp->ms_sm->sm_phys->smp_histogram[i];

module/zfs/spa.c

@@ -417,11 +417,15 @@ spa_prop_get_config(spa_t *spa, nvlist_t *nv)
 		alloc += metaslab_class_get_alloc(spa_special_class(spa));
 		alloc += metaslab_class_get_alloc(spa_dedup_class(spa));
 		alloc += metaslab_class_get_alloc(spa_embedded_log_class(spa));
+		alloc += metaslab_class_get_alloc(
+		    spa_special_embedded_log_class(spa));
 
 		size = metaslab_class_get_space(mc);
 		size += metaslab_class_get_space(spa_special_class(spa));
 		size += metaslab_class_get_space(spa_dedup_class(spa));
 		size += metaslab_class_get_space(spa_embedded_log_class(spa));
+		size += metaslab_class_get_space(
+		    spa_special_embedded_log_class(spa));
 
 		spa_prop_add_list(nv, ZPOOL_PROP_NAME, spa_name(spa), 0, src);
 		spa_prop_add_list(nv, ZPOOL_PROP_SIZE, NULL, size, src);
@@ -1679,6 +1683,8 @@ spa_activate(spa_t *spa, spa_mode_t mode)
 	    "embedded_log", msp, B_TRUE);
 	spa->spa_special_class = metaslab_class_create(spa, "special",
 	    msp, B_FALSE);
+	spa->spa_special_embedded_log_class = metaslab_class_create(spa,
+	    "special_embedded_log", msp, B_TRUE);
 	spa->spa_dedup_class = metaslab_class_create(spa, "dedup",
 	    msp, B_FALSE);
 
@@ -1853,6 +1859,9 @@ spa_deactivate(spa_t *spa)
 	metaslab_class_destroy(spa->spa_special_class);
 	spa->spa_special_class = NULL;
 
+	metaslab_class_destroy(spa->spa_special_embedded_log_class);
+	spa->spa_special_embedded_log_class = NULL;
+
 	metaslab_class_destroy(spa->spa_dedup_class);
 	spa->spa_dedup_class = NULL;
 
@@ -9092,6 +9101,8 @@ spa_async_thread(void *arg)
 		old_space += metaslab_class_get_space(spa_dedup_class(spa));
 		old_space += metaslab_class_get_space(
 		    spa_embedded_log_class(spa));
+		old_space += metaslab_class_get_space(
+		    spa_special_embedded_log_class(spa));
 
 		spa_config_update(spa, SPA_CONFIG_UPDATE_POOL);
 
@@ -9100,6 +9111,8 @@ spa_async_thread(void *arg)
 		new_space += metaslab_class_get_space(spa_dedup_class(spa));
 		new_space += metaslab_class_get_space(
 		    spa_embedded_log_class(spa));
+		new_space += metaslab_class_get_space(
+		    spa_special_embedded_log_class(spa));
 		mutex_exit(&spa_namespace_lock);
 
 		/*
@@ -10309,7 +10322,7 @@ spa_sync(spa_t *spa, uint64_t txg)
 
 	metaslab_class_evict_old(spa->spa_normal_class, txg);
 	metaslab_class_evict_old(spa->spa_log_class, txg);
-	/* spa_embedded_log_class has only one metaslab per vdev. */
+	/* Embedded log classes have only one metaslab per vdev. */
 	metaslab_class_evict_old(spa->spa_special_class, txg);
 	metaslab_class_evict_old(spa->spa_dedup_class, txg);
 

module/zfs/spa_misc.c

@@ -1308,6 +1308,7 @@ spa_vdev_config_exit(spa_t *spa, vdev_t *vd, uint64_t txg, int error,
 	metaslab_class_validate(spa_log_class(spa));
 	metaslab_class_validate(spa_embedded_log_class(spa));
 	metaslab_class_validate(spa_special_class(spa));
+	metaslab_class_validate(spa_special_embedded_log_class(spa));
 	metaslab_class_validate(spa_dedup_class(spa));
 
 	spa_config_exit(spa, SCL_ALL, spa);
@@ -1896,6 +1897,8 @@ spa_get_slop_space(spa_t *spa)
 	 */
 	uint64_t embedded_log =
 	    metaslab_class_get_dspace(spa_embedded_log_class(spa));
+	embedded_log += metaslab_class_get_dspace(
+	    spa_special_embedded_log_class(spa));
 	slop -= MIN(embedded_log, slop >> 1);
 
 	/*
@@ -2000,6 +2003,12 @@ spa_special_class(spa_t *spa)
 	return (spa->spa_special_class);
 }
 
+metaslab_class_t *
+spa_special_embedded_log_class(spa_t *spa)
+{
+	return (spa->spa_special_embedded_log_class);
+}
+
 metaslab_class_t *
 spa_dedup_class(spa_t *spa)
 {

module/zfs/vdev.c

@@ -282,12 +282,15 @@ vdev_getops(const char *type)
  * Given a vdev and a metaslab class, find which metaslab group we're
  * interested in. All vdevs may belong to two different metaslab classes.
  * Dedicated slog devices use only the primary metaslab group, rather than a
- * separate log group. For embedded slogs, the vdev_log_mg will be non-NULL.
+ * separate log group.  For embedded slogs, vdev_log_mg will be non-NULL and
+ * will point to a metaslab group of either embedded_log_class (for normal
+ * vdevs) or special_embedded_log_class (for special vdevs).
  */
 metaslab_group_t *
 vdev_get_mg(vdev_t *vd, metaslab_class_t *mc)
 {
-	if (mc == spa_embedded_log_class(vd->vdev_spa) &&
+	if ((mc == spa_embedded_log_class(vd->vdev_spa) ||
+	    mc == spa_special_embedded_log_class(vd->vdev_spa)) &&
 	    vd->vdev_log_mg != NULL)
 		return (vd->vdev_log_mg);
 	else
@@ -1508,8 +1511,13 @@ vdev_metaslab_group_create(vdev_t *vd)
 		vd->vdev_mg = metaslab_group_create(mc, vd);
 
 		if (!vd->vdev_islog) {
-			vd->vdev_log_mg = metaslab_group_create(
-			    spa_embedded_log_class(spa), vd);
+			if (mc == spa_special_class(spa)) {
+				vd->vdev_log_mg = metaslab_group_create(
+				    spa_special_embedded_log_class(spa), vd);
+			} else {
+				vd->vdev_log_mg = metaslab_group_create(
+				    spa_embedded_log_class(spa), vd);
+			}
 		}
 
 		/*
@@ -1624,9 +1632,10 @@ vdev_metaslab_init(vdev_t *vd, uint64_t txg)
 	/*
 	 * Find the emptiest metaslab on the vdev and mark it for use for
 	 * embedded slog by moving it from the regular to the log metaslab
-	 * group.
+	 * group.  This works for normal and special vdevs.
 	 */
-	if (vd->vdev_mg->mg_class == spa_normal_class(spa) &&
+	if ((vd->vdev_mg->mg_class == spa_normal_class(spa) ||
+	    vd->vdev_mg->mg_class == spa_special_class(spa)) &&
 	    vd->vdev_ms_count > zfs_embedded_slog_min_ms &&
 	    avl_is_empty(&vd->vdev_log_mg->mg_metaslab_tree)) {
 		uint64_t slog_msid = 0;

module/zfs/zfs_log.c

@@ -607,8 +607,6 @@ zfs_log_rename_whiteout(zilog_t *zilog, dmu_tx_t *tx, uint64_t txtype,
  * called as soon as the write is on stable storage (be it via a DMU sync or a
  * ZIL commit).
  */
-static uint_t zfs_immediate_write_sz = 32768;
-
 void
 zfs_log_write(zilog_t *zilog, dmu_tx_t *tx, int txtype,
     znode_t *zp, offset_t off, ssize_t resid, boolean_t commit,
@@ -626,15 +624,8 @@ zfs_log_write(zilog_t *zilog, dmu_tx_t *tx, int txtype,
 		return;
 	}
 
-	if (zilog->zl_logbias == ZFS_LOGBIAS_THROUGHPUT || o_direct)
-		write_state = WR_INDIRECT;
-	else if (!spa_has_slogs(zilog->zl_spa) &&
-	    resid >= zfs_immediate_write_sz)
-		write_state = WR_INDIRECT;
-	else if (commit)
-		write_state = WR_COPIED;
-	else
-		write_state = WR_NEED_COPY;
+	write_state = zil_write_state(zilog, resid, blocksize, o_direct,
+	    commit);
 
 	(void) sa_lookup(zp->z_sa_hdl, SA_ZPL_GEN(ZTOZSB(zp)), &gen,
 	    sizeof (gen));
@@ -938,6 +929,3 @@ zfs_log_clone_range(zilog_t *zilog, dmu_tx_t *tx, int txtype, znode_t *zp,
 		len -= partlen;
 	}
 }
-
-ZFS_MODULE_PARAM(zfs, zfs_, immediate_write_sz, UINT, ZMOD_RW,
-	"Largest data block to write to zil");

module/zfs/zil.c

@@ -2095,6 +2095,19 @@ zil_max_waste_space(zilog_t *zilog)
  */
 static uint_t zil_maxcopied = 7680;
 
+/*
+ * Largest write size to store the data directly into ZIL.
+ */
+uint_t zfs_immediate_write_sz = 32768;
+
+/*
+ * When enabled and blocks go to normal vdev, treat special vdevs as SLOG,
+ * writing data to ZIL (WR_COPIED/WR_NEED_COPY).  Disabling this forces the
+ * indirect writes (WR_INDIRECT) to preserve special vdev throughput and
+ * endurance, likely at the cost of normal vdev latency.
+ */
+int zil_special_is_slog = 1;
+
 uint64_t
 zil_max_copied_data(zilog_t *zilog)
 {
@@ -2102,6 +2115,46 @@ zil_max_copied_data(zilog_t *zilog)
 	return (MIN(max_data, zil_maxcopied));
 }
 
+/*
+ * Determine the appropriate write state for ZIL transactions based on
+ * pool configuration, data placement, write size, and logbias settings.
+ */
+itx_wr_state_t
+zil_write_state(zilog_t *zilog, uint64_t size, uint32_t blocksize,
+    boolean_t o_direct, boolean_t commit)
+{
+	if (zilog->zl_logbias == ZFS_LOGBIAS_THROUGHPUT || o_direct)
+		return (WR_INDIRECT);
+
+	/*
+	 * Don't use indirect for too small writes to reduce overhead.
+	 * Don't use indirect if written less than a half of a block if
+	 * we are going to commit it immediately, since next write might
+	 * rewrite the same block again, causing inflation.  If commit
+	 * is not planned, then next writes might coalesce, and so the
+	 * indirect may be perfect.
+	 */
+	boolean_t indirect = (size >= zfs_immediate_write_sz &&
+	    (size >= blocksize / 2 || !commit));
+
+	if (spa_has_slogs(zilog->zl_spa)) {
+		/* Dedicated slogs: never use indirect */
+		indirect = B_FALSE;
+	} else if (spa_has_special(zilog->zl_spa)) {
+		/* Special vdevs: only when beneficial */
+		boolean_t on_special = (blocksize <=
+		    zilog->zl_os->os_zpl_special_smallblock);
+		indirect &= (on_special || !zil_special_is_slog);
+	}
+
+	if (indirect)
+		return (WR_INDIRECT);
+	else if (commit)
+		return (WR_COPIED);
+	else
+		return (WR_NEED_COPY);
+}
+
 static uint64_t
 zil_itx_record_size(itx_t *itx)
 {
@@ -4418,3 +4471,9 @@ ZFS_MODULE_PARAM(zfs_zil, zil_, maxblocksize, UINT, ZMOD_RW,
 
 ZFS_MODULE_PARAM(zfs_zil, zil_, maxcopied, UINT, ZMOD_RW,
 	"Limit in bytes WR_COPIED size");
+
+ZFS_MODULE_PARAM(zfs, zfs_, immediate_write_sz, UINT, ZMOD_RW,
+	"Largest write size to store data into ZIL");
+
+ZFS_MODULE_PARAM(zfs_zil, zil_, special_is_slog, INT, ZMOD_RW,
+	"Treat special vdevs as SLOG");

module/zfs/zio.c

@@ -4433,14 +4433,34 @@ zio_alloc_zil(spa_t *spa, objset_t *os, uint64_t txg, blkptr_t *new_bp,
 	int allocator = (uint_t)cityhash1(os->os_dsl_dataset->ds_object)
 	    % spa->spa_alloc_count;
 	ZIOSTAT_BUMP(ziostat_total_allocations);
+
+	/* Try log class (dedicated slog devices) first */
 	error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
 	    txg, NULL, flags, &io_alloc_list, allocator, NULL);
 	*slog = (error == 0);
+
+	/* Try special_embedded_log class (reserved on special vdevs) */
+	if (error != 0) {
+		error = metaslab_alloc(spa, spa_special_embedded_log_class(spa),
+		    size, new_bp, 1, txg, NULL, flags, &io_alloc_list,
+		    allocator, NULL);
+	}
+
+	/* Try special class (general special vdev allocation) */
+	if (error != 0) {
+		error = metaslab_alloc(spa, spa_special_class(spa), size,
+		    new_bp, 1, txg, NULL, flags, &io_alloc_list, allocator,
+		    NULL);
+	}
+
+	/* Try embedded_log class (reserved on normal vdevs) */
 	if (error != 0) {
 		error = metaslab_alloc(spa, spa_embedded_log_class(spa), size,
 		    new_bp, 1, txg, NULL, flags, &io_alloc_list, allocator,
 		    NULL);
 	}
+
+	/* Finally fall back to normal class */
 	if (error != 0) {
 		ZIOSTAT_BUMP(ziostat_alloc_class_fallbacks);
 		error = metaslab_alloc(spa, spa_normal_class(spa), size,

module/zfs/zvol.c

@@ -859,13 +859,8 @@ zil_replay_func_t *const zvol_replay_vector[TX_MAX_TYPE] = {
 };
 
 /*
- * zvol_log_write() handles synchronous writes using TX_WRITE ZIL transactions.
- *
- * We store data in the log buffers if it's small enough.
- * Otherwise we will later flush the data out via dmu_sync().
+ * zvol_log_write() handles TX_WRITE transactions.
  */
-static const ssize_t zvol_immediate_write_sz = 32768;
-
 void
 zvol_log_write(zvol_state_t *zv, dmu_tx_t *tx, uint64_t offset,
     uint64_t size, boolean_t commit)
@@ -878,15 +873,7 @@ zvol_log_write(zvol_state_t *zv, dmu_tx_t *tx, uint64_t offset,
 	if (zil_replaying(zilog, tx))
 		return;
 
-	if (zilog->zl_logbias == ZFS_LOGBIAS_THROUGHPUT)
-		write_state = WR_INDIRECT;
-	else if (!spa_has_slogs(zilog->zl_spa) &&
-	    size >= blocksize && blocksize > zvol_immediate_write_sz)
-		write_state = WR_INDIRECT;
-	else if (commit)
-		write_state = WR_COPIED;
-	else
-		write_state = WR_NEED_COPY;
+	write_state = zil_write_state(zilog, size, blocksize, B_FALSE, commit);
 
 	while (size) {
 		itx_t *itx;