Illumos #3834

3834 incremental replication of 'holey' file systems is slow
Reviewed by: Adam Leventhal <[email protected]>
Reviewed by: George Wilson <[email protected]>
Approved by: Richard Lowe <[email protected]>

References:
  https://www.illumos.org/issues/3834
  illumos/illumos-gate@ca48f36f20f6098ceb19d5b084b6b3d4b8eca9fa

Ported-by: Richard Yao <[email protected]>
Signed-off-by: Brian Behlendorf <[email protected]>
Issue #1775

1 files changed, 31 insertions, 17 deletions

diff --git a/module/zfs/dmu_tx.c b/module/zfs/dmu_tx.c
index caac60193..7ee214bb0 100644
--- a/module/zfs/dmu_tx.c
+++ b/module/zfs/dmu_tx.c
@@ -604,8 +604,7 @@ dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
 {
 	dmu_tx_hold_t *txh;
 	dnode_t *dn;
-	uint64_t start, end, i;
-	int err, shift;
+	int err;
 	zio_t *zio;
 
 	ASSERT(tx->tx_txg == 0);
@@ -616,30 +615,45 @@ dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
 		return;
 	dn = txh->txh_dnode;
 
-	/* first block */
-	if (off != 0)
-		dmu_tx_count_write(txh, off, 1);
-	/* last block */
-	if (len != DMU_OBJECT_END)
-		dmu_tx_count_write(txh, off+len, 1);
-
-	dmu_tx_count_dnode(txh);
-
 	if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
 		return;
 	if (len == DMU_OBJECT_END)
 		len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
 
+	dmu_tx_count_dnode(txh);
+
 	/*
-	 * For i/o error checking, read the first and last level-0
-	 * blocks, and all the level-1 blocks.  The above count_write's
-	 * have already taken care of the level-0 blocks.
+	 * For i/o error checking, we read the first and last level-0
+	 * blocks if they are not aligned, and all the level-1 blocks.
+	 *
+	 * Note:  dbuf_free_range() assumes that we have not instantiated
+	 * any level-0 dbufs that will be completely freed.  Therefore we must
+	 * exercise care to not read or count the first and last blocks
+	 * if they are blocksize-aligned.
+	 */
+	if (dn->dn_datablkshift == 0) {
+		dmu_tx_count_write(txh, off, len);
+	} else {
+		/* first block will be modified if it is not aligned */
+		if (!IS_P2ALIGNED(off, 1 << dn->dn_datablkshift))
+			dmu_tx_count_write(txh, off, 1);
+		/* last block will be modified if it is not aligned */
+		if (!IS_P2ALIGNED(off + len, 1 << dn->dn_datablkshift))
+			dmu_tx_count_write(txh, off+len, 1);
+	}
+
+	/*
+	 * Check level-1 blocks.
 	 */
 	if (dn->dn_nlevels > 1) {
-		shift = dn->dn_datablkshift + dn->dn_indblkshift -
+		int shift = dn->dn_datablkshift + dn->dn_indblkshift -
 		    SPA_BLKPTRSHIFT;
-		start = off >> shift;
-		end = dn->dn_datablkshift ? ((off+len) >> shift) : 0;
+		uint64_t start = off >> shift;
+		uint64_t end = (off + len) >> shift;
+		uint64_t i;
+
+		ASSERT(dn->dn_datablkshift != 0);
+		ASSERT(dn->dn_indblkshift != 0);
 
 		zio = zio_root(tx->tx_pool->dp_spa,
 		    NULL, NULL, ZIO_FLAG_CANFAIL);