7578 Fix/improve some aspects of ZIL writing.

Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: Prakash Surya <prakash.surya@delphix.com>
Reviewed by: Andriy Gapon <avg@FreeBSD.org>
Reviewed by: Steven Hartland <steven.hartland@multiplay.co.uk>
Reviewed by: Brad Lewis <brad.lewis@delphix.com>
Reviewed by: Richard Elling <Richard.Elling@RichardElling.com>
Approved by: Robert Mustacchi <rm@joyent.com>

usr/src/cmd/ztest/ztest.c

@@ -1377,7 +1377,6 @@ 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));

usr/src/uts/common/fs/zfs/sys/zil.h

@@ -378,7 +378,6 @@ typedef struct itx {
 	void		*itx_private;	/* type-specific opaque data */
 	itx_wr_state_t	itx_wr_state;	/* write state */
 	uint8_t		itx_sync;	/* synchronous transaction */
-	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 */

usr/src/uts/common/fs/zfs/sys/zil_impl.h

@@ -42,6 +42,7 @@ extern "C" {
 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_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

usr/src/uts/common/fs/zfs/sys/zio.h

@@ -505,7 +505,7 @@ 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,
-    blkptr_t *old_bp, uint64_t size, boolean_t use_slog);
+    blkptr_t *old_bp, 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);

usr/src/uts/common/fs/zfs/zfs_log.c

@@ -454,20 +454,17 @@ void
 zfs_log_write(zilog_t *zilog, dmu_tx_t *tx, int txtype,
     znode_t *zp, offset_t off, ssize_t resid, int ioflag)
 {
+	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)
 		return;
 
-	immediate_write_sz = (zilog->zl_logbias == ZFS_LOGBIAS_THROUGHPUT)
-	    ? 0 : 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;
@@ -481,30 +478,26 @@ 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(zp->z_zfsvfs->z_os,
+		if (wr_state == WR_COPIED && dmu_read(zp->z_zfsvfs->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;