OpenZFS 7578 - Fix/improve some aspects of ZIL writing

- After some ZIL changes 6 years ago zil_slog_limit got partially broken
due to zl_itx_list_sz not updated when async itx'es upgraded to sync.
Actually because of other changes about that time zl_itx_list_sz is not
really required to implement the functionality, so this patch removes
some unneeded broken code and variables.

 - Original idea of zil_slog_limit was to reduce chance of SLOG abuse by
single heavy logger, that increased latency for other (more latency critical)
loggers, by pushing heavy log out into the main pool instead of SLOG.  Beside
huge latency increase for heavy writers, this implementation caused double
write of all data, since the log records were explicitly prepared for SLOG.
Since we now have I/O scheduler, I've found it can be much more efficient
to reduce priority of heavy logger SLOG writes from ZIO_PRIORITY_SYNC_WRITE
to ZIO_PRIORITY_ASYNC_WRITE, while still leave them on SLOG.

 - Existing ZIL implementation had problem with space efficiency when it
has to write large chunks of data into log blocks of limited size.  In some
cases efficiency stopped to almost as low as 50%.  In case of ZIL stored on
spinning rust, that also reduced log write speed in half, since head had to
uselessly fly over allocated but not written areas.  This change improves
the situation by offloading problematic operations from z*_log_write() to
zil_lwb_commit(), which knows real situation of log blocks allocation and
can split large requests into pieces much more efficiently.  Also as side
effect it removes one of two data copy operations done by ZIL code WR_COPIED
case.

 - While there, untangle and unify code of z*_log_write() functions.
Also zfs_log_write() alike to zvol_log_write() can now handle writes crossing
block boundary, that may also improve efficiency if ZPL is made to do that.

Sponsored by:   iXsystems, Inc.

Authored by: Alexander Motin <[email protected]>
Reviewed by: Matthew Ahrens <[email protected]>
Reviewed by: Prakash Surya <[email protected]>
Reviewed by: Andriy Gapon <[email protected]>
Reviewed by: Steven Hartland <[email protected]>
Reviewed by: Brad Lewis <[email protected]>
Reviewed by: Richard Elling <[email protected]>
Approved by: Robert Mustacchi <[email protected]>
Reviewed-by: Brian Behlendorf <[email protected]>
Reviewed-by: Richard Yao <[email protected]>
Ported-by: Giuseppe Di Natale <[email protected]>

OpenZFS-issue: https://www.illumos.org/issues/7578
OpenZFS-commit: https://github.com/openzfs/openzfs/commit/aeb13ac
Closes #6191 

10 files changed, 133 insertions, 128 deletions

diff --git a/cmd/ztest/ztest.c b/cmd/ztest/ztest.c
index ea1a3030b..39fbfd993 100644
--- a/cmd/ztest/ztest.c
+++ b/cmd/ztest/ztest.c
@@ -1620,7 +1620,6 @@ ztest_log_write(ztest_ds_t *zd, dmu_tx_t *tx, lr_write_t *lr)
 	itx->itx_private = zd;
 	itx->itx_wr_state = write_state;
 	itx->itx_sync = (ztest_random(8) == 0);
-	itx->itx_sod += (write_state == WR_NEED_COPY ? lr->lr_length : 0);
 
 	bcopy(&lr->lr_common + 1, &itx->itx_lr + 1,
 	    sizeof (*lr) - sizeof (lr_t));
diff --git a/include/sys/trace_zil.h b/include/sys/trace_zil.h
index 5b43480ae..eaa391270 100644
--- a/include/sys/trace_zil.h
+++ b/include/sys/trace_zil.h
@@ -63,7 +63,6 @@ DECLARE_EVENT_CLASS(zfs_zil_class,
 	    __field(uint64_t,	zl_parse_lr_count)
 	    __field(uint64_t,	zl_next_batch)
 	    __field(uint64_t,	zl_com_batch)
-	    __field(uint64_t,	zl_itx_list_sz)
 	    __field(uint64_t,	zl_cur_used)
 	    __field(clock_t,	zl_replay_time)
 	    __field(uint64_t,	zl_replay_blks)
@@ -88,7 +87,6 @@ DECLARE_EVENT_CLASS(zfs_zil_class,
 	    __entry->zl_parse_lr_count	= zilog->zl_parse_lr_count;
 	    __entry->zl_next_batch	= zilog->zl_next_batch;
 	    __entry->zl_com_batch	= zilog->zl_com_batch;
-	    __entry->zl_itx_list_sz	= zilog->zl_itx_list_sz;
 	    __entry->zl_cur_used	= zilog->zl_cur_used;
 	    __entry->zl_replay_time	= zilog->zl_replay_time;
 	    __entry->zl_replay_blks	= zilog->zl_replay_blks;
@@ -98,8 +96,7 @@ DECLARE_EVENT_CLASS(zfs_zil_class,
 	    "replay %u stop_sync %u writer %u logbias %u sync %u "
 	    "parse_error %u parse_blk_seq %llu parse_lr_seq %llu "
 	    "parse_blk_count %llu parse_lr_count %llu next_batch %llu "
-	    "com_batch %llu itx_list_sz %llu cur_used %llu replay_time %lu "
-	    "replay_blks %llu }",
+	    "com_batch %llu cur_used %llu replay_time %lu replay_blks %llu }",
 	    __entry->zl_lr_seq, __entry->zl_commit_lr_seq,
 	    __entry->zl_destroy_txg, __entry->zl_replaying_seq,
 	    __entry->zl_suspend, __entry->zl_suspending, __entry->zl_keep_first,
@@ -107,8 +104,7 @@ DECLARE_EVENT_CLASS(zfs_zil_class,
 	    __entry->zl_logbias, __entry->zl_sync, __entry->zl_parse_error,
 	    __entry->zl_parse_blk_seq, __entry->zl_parse_lr_seq,
 	    __entry->zl_parse_blk_count, __entry->zl_parse_lr_count,
-	    __entry->zl_next_batch, __entry->zl_com_batch,
-	    __entry->zl_itx_list_sz, __entry->zl_cur_used,
+	    __entry->zl_next_batch, __entry->zl_com_batch, __entry->zl_cur_used,
 	    __entry->zl_replay_time, __entry->zl_replay_blks)
 );
 /* END CSTYLED */
diff --git a/include/sys/zil.h b/include/sys/zil.h
index 62572f894..95fd324b4 100644
--- a/include/sys/zil.h
+++ b/include/sys/zil.h
@@ -394,7 +394,6 @@ typedef struct itx {
 	uint8_t		itx_sync;	/* synchronous transaction */
 	zil_callback_t	itx_callback;   /* Called when the itx is persistent */
 	void		*itx_callback_data; /* User data for the callback */
-	uint64_t	itx_sod;	/* record size on disk */
 	uint64_t	itx_oid;	/* object id */
 	lr_t		itx_lr;		/* common part of log record */
 	/* followed by type-specific part of lr_xx_t and its immediate data */
diff --git a/include/sys/zil_impl.h b/include/sys/zil_impl.h
index 0c426a15d..13ecca3c8 100644
--- a/include/sys/zil_impl.h
+++ b/include/sys/zil_impl.h
@@ -42,6 +42,7 @@ typedef struct lwb {
 	zilog_t		*lwb_zilog;	/* back pointer to log struct */
 	blkptr_t	lwb_blk;	/* on disk address of this log blk */
 	boolean_t	lwb_fastwrite;	/* is blk marked for fastwrite? */
+	boolean_t	lwb_slog;	/* lwb_blk is on SLOG device */
 	int		lwb_nused;	/* # used bytes in buffer */
 	int		lwb_sz;		/* size of block and buffer */
 	char		*lwb_buf;	/* log write buffer */
@@ -62,7 +63,6 @@ typedef struct itxs {
 typedef struct itxg {
 	kmutex_t	itxg_lock;	/* lock for this structure */
 	uint64_t	itxg_txg;	/* txg for this chain */
-	uint64_t	itxg_sod;	/* total size on disk for this txg */
 	itxs_t		*itxg_itxs;	/* sync and async itxs */
 } itxg_t;
 
@@ -120,7 +120,6 @@ struct zilog {
 	kcondvar_t	zl_cv_batch[2];	/* batch condition variables */
 	itxg_t		zl_itxg[TXG_SIZE]; /* intent log txg chains */
 	list_t		zl_itx_commit_list; /* itx list to be committed */
-	uint64_t	zl_itx_list_sz;	/* total size of records on list */
 	uint64_t	zl_cur_used;	/* current commit log size used */
 	list_t		zl_lwb_list;	/* in-flight log write list */
 	kmutex_t	zl_vdev_lock;	/* protects zl_vdev_tree */
@@ -140,9 +139,26 @@ typedef struct zil_bp_node {
 	avl_node_t	zn_node;
 } zil_bp_node_t;
 
+/*
+ * Maximum amount of write data that can be put into single log block.
+ */
 #define	ZIL_MAX_LOG_DATA (SPA_OLD_MAXBLOCKSIZE - sizeof (zil_chain_t) - \
     sizeof (lr_write_t))
 
+/*
+ * Maximum amount of log space we agree to waste to reduce number of
+ * WR_NEED_COPY chunks to reduce zl_get_data() overhead (~12%).
+ */
+#define	ZIL_MAX_WASTE_SPACE (ZIL_MAX_LOG_DATA / 8)
+
+/*
+ * Maximum amount of write data for WR_COPIED.  Fall back to WR_NEED_COPY
+ * as more space efficient if we can't fit at least two log records into
+ * maximum sized log block.
+ */
+#define	ZIL_MAX_COPIED_DATA ((SPA_OLD_MAXBLOCKSIZE - \
+    sizeof (zil_chain_t)) / 2 - sizeof (lr_write_t))
+
 #ifdef	__cplusplus
 }
 #endif
diff --git a/include/sys/zio.h b/include/sys/zio.h
index 23c918af1..4eaabc38c 100644
--- a/include/sys/zio.h
+++ b/include/sys/zio.h
@@ -515,7 +515,7 @@ extern zio_t *zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg,
     const blkptr_t *bp, enum zio_flag flags);
 
 extern int zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp,
-    uint64_t size, boolean_t use_slog);
+    uint64_t size, boolean_t *slog);
 extern void zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp);
 extern void zio_flush(zio_t *zio, vdev_t *vd);
 extern void zio_shrink(zio_t *zio, uint64_t size);
diff --git a/man/man5/zfs-module-parameters.5 b/man/man5/zfs-module-parameters.5
index 51f1cd93f..d4bab8327 100644
--- a/man/man5/zfs-module-parameters.5
+++ b/man/man5/zfs-module-parameters.5
@@ -1935,12 +1935,14 @@ Use \fB1\fR for yes and \fB0\fR for no (default).
 .sp
 .ne 2
 .na
-\fBzil_slog_limit\fR (ulong)
+\fBzil_slog_bulk\fR (ulong)
 .ad
 .RS 12n
-Max commit bytes to separate log device
+Limit SLOG write size per commit executed with synchronous priority.
+Any writes above that will be executed with lower (asynchronous) priority
+to limit potential SLOG device abuse by single active ZIL writer.
 .sp
-Default value: \fB1,048,576\fR.
+Default value: \fB786,432\fR.
 .RE
 
 .sp
diff --git a/module/zfs/zfs_log.c b/module/zfs/zfs_log.c
index ca3699052..8887f037a 100644
--- a/module/zfs/zfs_log.c
+++ b/module/zfs/zfs_log.c
@@ -487,10 +487,9 @@ zfs_log_write(zilog_t *zilog, dmu_tx_t *tx, int txtype,
     znode_t *zp, offset_t off, ssize_t resid, int ioflag,
     zil_callback_t callback, void *callback_data)
 {
+	uint32_t blocksize = zp->z_blksz;
 	itx_wr_state_t write_state;
-	boolean_t slogging;
 	uintptr_t fsync_cnt;
-	ssize_t immediate_write_sz;
 
 	if (zil_replaying(zilog, tx) || zp->z_unlinked ||
 	    zfs_xattr_owner_unlinked(zp)) {
@@ -499,12 +498,10 @@ zfs_log_write(zilog_t *zilog, dmu_tx_t *tx, int txtype,
 		return;
 	}
 
-	immediate_write_sz = (zilog->zl_logbias == ZFS_LOGBIAS_THROUGHPUT)
-	    ? 0 : (ssize_t)zfs_immediate_write_sz;
-
-	slogging = spa_has_slogs(zilog->zl_spa) &&
-	    (zilog->zl_logbias == ZFS_LOGBIAS_LATENCY);
-	if (resid > immediate_write_sz && !slogging && resid <= zp->z_blksz)
+	if (zilog->zl_logbias == ZFS_LOGBIAS_THROUGHPUT)
+		write_state = WR_INDIRECT;
+	else if (!spa_has_slogs(zilog->zl_spa) &&
+	    resid >= zfs_immediate_write_sz)
 		write_state = WR_INDIRECT;
 	else if (ioflag & (FSYNC | FDSYNC))
 		write_state = WR_COPIED;
@@ -518,30 +515,26 @@ zfs_log_write(zilog_t *zilog, dmu_tx_t *tx, int txtype,
 	while (resid) {
 		itx_t *itx;
 		lr_write_t *lr;
-		ssize_t len;
+		itx_wr_state_t wr_state = write_state;
+		ssize_t len = resid;
 
-		/*
-		 * If the write would overflow the largest block then split it.
-		 */
-		if (write_state != WR_INDIRECT && resid > ZIL_MAX_LOG_DATA)
-			len = SPA_OLD_MAXBLOCKSIZE >> 1;
-		else
-			len = resid;
+		if (wr_state == WR_COPIED && resid > ZIL_MAX_COPIED_DATA)
+			wr_state = WR_NEED_COPY;
+		else if (wr_state == WR_INDIRECT)
+			len = MIN(blocksize - P2PHASE(off, blocksize), resid);
 
 		itx = zil_itx_create(txtype, sizeof (*lr) +
-		    (write_state == WR_COPIED ? len : 0));
+		    (wr_state == WR_COPIED ? len : 0));
 		lr = (lr_write_t *)&itx->itx_lr;
-		if (write_state == WR_COPIED && dmu_read(ZTOZSB(zp)->z_os,
+		if (wr_state == WR_COPIED && dmu_read(ZTOZSB(zp)->z_os,
 		    zp->z_id, off, len, lr + 1, DMU_READ_NO_PREFETCH) != 0) {
 			zil_itx_destroy(itx);
 			itx = zil_itx_create(txtype, sizeof (*lr));
 			lr = (lr_write_t *)&itx->itx_lr;
-			write_state = WR_NEED_COPY;
+			wr_state = WR_NEED_COPY;
 		}
 
-		itx->itx_wr_state = write_state;
-		if (write_state == WR_NEED_COPY)
-			itx->itx_sod += len;
+		itx->itx_wr_state = wr_state;
 		lr->lr_foid = zp->z_id;
 		lr->lr_offset = off;
 		lr->lr_length = len;
diff --git a/module/zfs/zil.c b/module/zfs/zil.c
index 12a034d5b..6a1f190f5 100644
--- a/module/zfs/zil.c
+++ b/module/zfs/zil.c
@@ -102,6 +102,13 @@ int zil_replay_disable = 0;
  */
 int zfs_nocacheflush = 0;
 
+/*
+ * Limit SLOG write size per commit executed with synchronous priority.
+ * Any writes above that will be executed with lower (asynchronous) priority
+ * to limit potential SLOG device abuse by single active ZIL writer.
+ */
+unsigned long zil_slog_bulk = 768 * 1024;
+
 static kmem_cache_t *zil_lwb_cache;
 
 static void zil_async_to_sync(zilog_t *zilog, uint64_t foid);
@@ -449,7 +456,8 @@ zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
 }
 
 static lwb_t *
-zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg, boolean_t fastwrite)
+zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, boolean_t slog, uint64_t txg,
+    boolean_t fastwrite)
 {
 	lwb_t *lwb;
 
@@ -457,6 +465,7 @@ zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg, boolean_t fastwrite)
 	lwb->lwb_zilog = zilog;
 	lwb->lwb_blk = *bp;
 	lwb->lwb_fastwrite = fastwrite;
+	lwb->lwb_slog = slog;
 	lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
 	lwb->lwb_max_txg = txg;
 	lwb->lwb_zio = NULL;
@@ -542,6 +551,7 @@ zil_create(zilog_t *zilog)
 	blkptr_t blk;
 	int error = 0;
 	boolean_t fastwrite = FALSE;
+	boolean_t slog = FALSE;
 
 	/*
 	 * Wait for any previous destroy to complete.
@@ -570,7 +580,7 @@ zil_create(zilog_t *zilog)
 		}
 
 		error = zio_alloc_zil(zilog->zl_spa, txg, &blk,
-		    ZIL_MIN_BLKSZ, B_TRUE);
+		    ZIL_MIN_BLKSZ, &slog);
 		fastwrite = TRUE;
 
 		if (error == 0)
@@ -581,7 +591,7 @@ zil_create(zilog_t *zilog)
 	 * Allocate a log write buffer (lwb) for the first log block.
 	 */
 	if (error == 0)
-		lwb = zil_alloc_lwb(zilog, &blk, txg, fastwrite);
+		lwb = zil_alloc_lwb(zilog, &blk, slog, txg, fastwrite);
 
 	/*
 	 * If we just allocated the first log block, commit our transaction
@@ -915,6 +925,7 @@ static void
 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
 {
 	zbookmark_phys_t zb;
+	zio_priority_t prio;
 
 	SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
 	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
@@ -934,9 +945,13 @@ zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
 			metaslab_fastwrite_mark(zilog->zl_spa, &lwb->lwb_blk);
 			lwb->lwb_fastwrite = 1;
 		}
+		if (!lwb->lwb_slog || zilog->zl_cur_used <= zil_slog_bulk)
+			prio = ZIO_PRIORITY_SYNC_WRITE;
+		else
+			prio = ZIO_PRIORITY_ASYNC_WRITE;
 		lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
 		    0, &lwb->lwb_blk, lwb_abd, BP_GET_LSIZE(&lwb->lwb_blk),
-		    zil_lwb_write_done, lwb, ZIO_PRIORITY_SYNC_WRITE,
+		    zil_lwb_write_done, lwb, prio,
 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE |
 		    ZIO_FLAG_FASTWRITE, &zb);
 	}
@@ -958,15 +973,6 @@ uint64_t zil_block_buckets[] = {
 };
 
 /*
- * Use the slog as long as the current commit size is less than the
- * limit or the total list size is less than 2X the limit.  Limit
- * checking is disabled by setting zil_slog_limit to UINT64_MAX.
- */
-unsigned long zil_slog_limit = 1024 * 1024;
-#define	USE_SLOG(zilog) (((zilog)->zl_cur_used < zil_slog_limit) || \
-	((zilog)->zl_itx_list_sz < (zil_slog_limit << 1)))
-
-/*
  * Start a log block write and advance to the next log block.
  * Calls are serialized.
  */
@@ -981,7 +987,7 @@ zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
 	uint64_t txg;
 	uint64_t zil_blksz, wsz;
 	int i, error;
-	boolean_t use_slog;
+	boolean_t slog;
 
 	if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
 		zilc = (zil_chain_t *)lwb->lwb_buf;
@@ -1037,10 +1043,8 @@ zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
 	zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
 
 	BP_ZERO(bp);
-	use_slog = USE_SLOG(zilog);
-	error = zio_alloc_zil(spa, txg, bp, zil_blksz,
-	    USE_SLOG(zilog));
-	if (use_slog) {
+	error = zio_alloc_zil(spa, txg, bp, zil_blksz, &slog);
+	if (slog) {
 		ZIL_STAT_BUMP(zil_itx_metaslab_slog_count);
 		ZIL_STAT_INCR(zil_itx_metaslab_slog_bytes, lwb->lwb_nused);
 	} else {
@@ -1055,7 +1059,7 @@ zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
 		/*
 		 * Allocate a new log write buffer (lwb).
 		 */
-		nlwb = zil_alloc_lwb(zilog, bp, txg, TRUE);
+		nlwb = zil_alloc_lwb(zilog, bp, slog, txg, TRUE);
 
 		/* Record the block for later vdev flushing */
 		zil_add_block(zilog, &lwb->lwb_blk);
@@ -1092,45 +1096,53 @@ zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
 static lwb_t *
 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
 {
-	lr_t *lrc = &itx->itx_lr; /* common log record */
-	lr_write_t *lrw = (lr_write_t *)lrc;
+	lr_t *lrcb, *lrc;
+	lr_write_t *lrwb, *lrw;
 	char *lr_buf;
-	uint64_t txg = lrc->lrc_txg;
-	uint64_t reclen = lrc->lrc_reclen;
-	uint64_t dlen = 0;
+	uint64_t dlen, dnow, lwb_sp, reclen, txg;
 
 	if (lwb == NULL)
 		return (NULL);
 
 	ASSERT(lwb->lwb_buf != NULL);
 
-	if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
+	lrc = &itx->itx_lr;		/* Common log record inside itx. */
+	lrw = (lr_write_t *)lrc;	/* Write log record inside itx. */
+	if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) {
 		dlen = P2ROUNDUP_TYPED(
 		    lrw->lr_length, sizeof (uint64_t), uint64_t);
-
+	} else {
+		dlen = 0;
+	}
+	reclen = lrc->lrc_reclen;
 	zilog->zl_cur_used += (reclen + dlen);
+	txg = lrc->lrc_txg;
 
 	zil_lwb_write_init(zilog, lwb);
 
+cont:
 	/*
 	 * If this record won't fit in the current log block, start a new one.
+	 * For WR_NEED_COPY optimize layout for minimal number of chunks.
 	 */
-	if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
+	lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
+	if (reclen > lwb_sp || (reclen + dlen > lwb_sp &&
+	    lwb_sp < ZIL_MAX_WASTE_SPACE && (dlen % ZIL_MAX_LOG_DATA == 0 ||
+	    lwb_sp < reclen + dlen % ZIL_MAX_LOG_DATA))) {
 		lwb = zil_lwb_write_start(zilog, lwb);
 		if (lwb == NULL)
 			return (NULL);
 		zil_lwb_write_init(zilog, lwb);
 		ASSERT(LWB_EMPTY(lwb));
-		if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
-			txg_wait_synced(zilog->zl_dmu_pool, txg);
-			return (lwb);
-		}
+		lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
+		ASSERT3U(reclen + MIN(dlen, sizeof (uint64_t)), <=, lwb_sp);
 	}
 
+	dnow = MIN(dlen, lwb_sp - reclen);
 	lr_buf = lwb->lwb_buf + lwb->lwb_nused;
 	bcopy(lrc, lr_buf, reclen);
-	lrc = (lr_t *)lr_buf;
-	lrw = (lr_write_t *)lrc;
+	lrcb = (lr_t *)lr_buf;		/* Like lrc, but inside lwb. */
+	lrwb = (lr_write_t *)lrcb;	/* Like lrw, but inside lwb. */
 
 	ZIL_STAT_BUMP(zil_itx_count);
 
@@ -1147,10 +1159,13 @@ zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
 			char *dbuf;
 			int error;
 
-			if (dlen) {
-				ASSERT(itx->itx_wr_state == WR_NEED_COPY);
+			if (itx->itx_wr_state == WR_NEED_COPY) {
 				dbuf = lr_buf + reclen;
-				lrw->lr_common.lrc_reclen += dlen;
+				lrcb->lrc_reclen += dnow;
+				if (lrwb->lr_length > dnow)
+					lrwb->lr_length = dnow;
+				lrw->lr_offset += dnow;
+				lrw->lr_length -= dnow;
 				ZIL_STAT_BUMP(zil_itx_needcopy_count);
 				ZIL_STAT_INCR(zil_itx_needcopy_bytes,
 				    lrw->lr_length);
@@ -1162,7 +1177,7 @@ zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
 				    lrw->lr_length);
 			}
 			error = zilog->zl_get_data(
-			    itx->itx_private, lrw, dbuf, lwb->lwb_zio);
+			    itx->itx_private, lrwb, dbuf, lwb->lwb_zio);
 			if (error == EIO) {
 				txg_wait_synced(zilog->zl_dmu_pool, txg);
 				return (lwb);
@@ -1181,12 +1196,18 @@ zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
 	 * equal to the itx sequence number because not all transactions
 	 * are synchronous, and sometimes spa_sync() gets there first.
 	 */
-	lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
-	lwb->lwb_nused += reclen + dlen;
+	lrcb->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
+	lwb->lwb_nused += reclen + dnow;
 	lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
 	ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
 	ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
 
+	dlen -= dnow;
+	if (dlen > 0) {
+		zilog->zl_cur_used += reclen;
+		goto cont;
+	}
+
 	return (lwb);
 }
 
@@ -1200,7 +1221,6 @@ zil_itx_create(uint64_t txtype, size_t lrsize)
 	itx = zio_data_buf_alloc(offsetof(itx_t, itx_lr) + lrsize);
 	itx->itx_lr.lrc_txtype = txtype;
 	itx->itx_lr.lrc_reclen = lrsize;
-	itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
 	itx->itx_lr.lrc_seq = 0;	/* defensive */
 	itx->itx_sync = B_TRUE;		/* default is synchronous */
 	itx->itx_callback = NULL;
@@ -1351,11 +1371,8 @@ zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
 			 */
 			zfs_dbgmsg("zil_itx_assign: missed itx cleanup for "
 			    "txg %llu", itxg->itxg_txg);
-			atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
-			itxg->itxg_sod = 0;
 			clean = itxg->itxg_itxs;
 		}
-		ASSERT(itxg->itxg_sod == 0);
 		itxg->itxg_txg = txg;
 		itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t),
 		    KM_SLEEP);
@@ -1368,8 +1385,6 @@ zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
 	}
 	if (itx->itx_sync) {
 		list_insert_tail(&itxs->i_sync_list, itx);
-		atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod);
-		itxg->itxg_sod += itx->itx_sod;
 	} else {
 		avl_tree_t *t = &itxs->i_async_tree;
 		uint64_t foid =
@@ -1419,8 +1434,6 @@ zil_clean(zilog_t *zilog, uint64_t synced_txg)
 	ASSERT3U(itxg->itxg_txg, <=, synced_txg);
 	ASSERT(itxg->itxg_txg != 0);
 	ASSERT(zilog->zl_clean_taskq != NULL);
-	atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
-	itxg->itxg_sod = 0;
 	clean_me = itxg->itxg_itxs;
 	itxg->itxg_itxs = NULL;
 	itxg->itxg_txg = 0;
@@ -1444,7 +1457,6 @@ zil_get_commit_list(zilog_t *zilog)
 {
 	uint64_t otxg, txg;
 	list_t *commit_list = &zilog->zl_itx_commit_list;
-	uint64_t push_sod = 0;
 
 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
 		otxg = ZILTEST_TXG;
@@ -1476,12 +1488,9 @@ zil_get_commit_list(zilog_t *zilog)
 		ASSERT(zilog_is_dirty_in_txg(zilog, txg) ||
 		    spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
 		list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
-		push_sod += itxg->itxg_sod;
-		itxg->itxg_sod = 0;
 
 		mutex_exit(&itxg->itxg_lock);
 	}
-	atomic_add_64(&zilog->zl_itx_list_sz, -push_sod);
 }
 
 /*
@@ -2304,13 +2313,14 @@ EXPORT_SYMBOL(zil_bp_tree_add);
 EXPORT_SYMBOL(zil_set_sync);
 EXPORT_SYMBOL(zil_set_logbias);
 
+/* BEGIN CSTYLED */
 module_param(zil_replay_disable, int, 0644);
 MODULE_PARM_DESC(zil_replay_disable, "Disable intent logging replay");
 
 module_param(zfs_nocacheflush, int, 0644);
 MODULE_PARM_DESC(zfs_nocacheflush, "Disable cache flushes");
 
-/* CSTYLED */
-module_param(zil_slog_limit, ulong, 0644);
-MODULE_PARM_DESC(zil_slog_limit, "Max commit bytes to separate log device");
+module_param(zil_slog_bulk, ulong, 0644);
+MODULE_PARM_DESC(zil_slog_bulk, "Limit in bytes slog sync writes per commit");
+/* END CSTYLED */
 #endif
diff --git a/module/zfs/zio.c b/module/zfs/zio.c
index 61eb575ef..acfc49eb5 100644
--- a/module/zfs/zio.c
+++ b/module/zfs/zio.c
@@ -3098,7 +3098,7 @@ zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
  */
 int
 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, uint64_t size,
-    boolean_t use_slog)
+    boolean_t *slog)
 {
 	int error = 1;
 	zio_alloc_list_t io_alloc_list;
@@ -3106,17 +3106,16 @@ zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, uint64_t size,
 	ASSERT(txg > spa_syncing_txg(spa));
 
 	metaslab_trace_init(&io_alloc_list);
-
-	if (use_slog) {
-		error = metaslab_alloc(spa, spa_log_class(spa), size,
-		    new_bp, 1, txg, NULL, METASLAB_FASTWRITE,
-		    &io_alloc_list, NULL);
-	}
-
-	if (error) {
+	error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
+	    txg, NULL, METASLAB_FASTWRITE, &io_alloc_list, NULL);
+	if (error == 0) {
+		*slog = TRUE;
+	} else {
 		error = metaslab_alloc(spa, spa_normal_class(spa), size,
 		    new_bp, 1, txg, NULL, METASLAB_FASTWRITE,
 		    &io_alloc_list, NULL);
+		if (error == 0)
+			*slog = FALSE;
 	}
 	metaslab_trace_fini(&io_alloc_list);
 
diff --git a/module/zfs/zvol.c b/module/zfs/zvol.c
index 1a86cd3cd..bf9f48adb 100644
--- a/module/zfs/zvol.c
+++ b/module/zfs/zvol.c
@@ -611,53 +611,44 @@ zvol_log_write(zvol_state_t *zv, dmu_tx_t *tx, uint64_t offset,
 {
 	uint32_t blocksize = zv->zv_volblocksize;
 	zilog_t *zilog = zv->zv_zilog;
-	boolean_t slogging;
-	ssize_t immediate_write_sz;
+	itx_wr_state_t write_state;
 
 	if (zil_replaying(zilog, tx))
 		return;
 
-	immediate_write_sz = (zilog->zl_logbias == ZFS_LOGBIAS_THROUGHPUT)
-	    ? 0 : zvol_immediate_write_sz;
-	slogging = spa_has_slogs(zilog->zl_spa) &&
-	    (zilog->zl_logbias == ZFS_LOGBIAS_LATENCY);
+	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 (sync)
+		write_state = WR_COPIED;
+	else
+		write_state = WR_NEED_COPY;
 
 	while (size) {
 		itx_t *itx;
 		lr_write_t *lr;
-		ssize_t len;
-		itx_wr_state_t write_state;
+		itx_wr_state_t wr_state = write_state;
+		ssize_t len = size;
 
-		/*
-		 * Unlike zfs_log_write() we can be called with
-		 * up to DMU_MAX_ACCESS/2 (5MB) writes.
-		 */
-		if (blocksize > immediate_write_sz && !slogging &&
-		    size >= blocksize && offset % blocksize == 0) {
-			write_state = WR_INDIRECT; /* uses dmu_sync */
-			len = blocksize;
-		} else if (sync) {
-			write_state = WR_COPIED;
-			len = MIN(ZIL_MAX_LOG_DATA, size);
-		} else {
-			write_state = WR_NEED_COPY;
-			len = MIN(ZIL_MAX_LOG_DATA, size);
-		}
+		if (wr_state == WR_COPIED && size > ZIL_MAX_COPIED_DATA)
+			wr_state = WR_NEED_COPY;
+		else if (wr_state == WR_INDIRECT)
+			len = MIN(blocksize - P2PHASE(offset, blocksize), size);
 
 		itx = zil_itx_create(TX_WRITE, sizeof (*lr) +
-		    (write_state == WR_COPIED ? len : 0));
+		    (wr_state == WR_COPIED ? len : 0));
 		lr = (lr_write_t *)&itx->itx_lr;
-		if (write_state == WR_COPIED && dmu_read(zv->zv_objset,
+		if (wr_state == WR_COPIED && dmu_read(zv->zv_objset,
 		    ZVOL_OBJ, offset, len, lr+1, DMU_READ_NO_PREFETCH) != 0) {
 			zil_itx_destroy(itx);
 			itx = zil_itx_create(TX_WRITE, sizeof (*lr));
 			lr = (lr_write_t *)&itx->itx_lr;
-			write_state = WR_NEED_COPY;
+			wr_state = WR_NEED_COPY;
 		}
 
-		itx->itx_wr_state = write_state;
-		if (write_state == WR_NEED_COPY)
-			itx->itx_sod += len;
+		itx->itx_wr_state = wr_state;
 		lr->lr_foid = ZVOL_OBJ;
 		lr->lr_offset = offset;
 		lr->lr_length = len;