Illumos #3834

3834 incremental replication of 'holey' file systems is slow
Reviewed by: Adam Leventhal <ahl@delphix.com>
Reviewed by: George Wilson <george.wilson@delphix.com>
Approved by: Richard Lowe <richlowe@richlowe.net>

References:
  https://www.illumos.org/issues/3834
  illumos/illumos-gate@ca48f36f20

Ported-by: Richard Yao <ryao@gentoo.org>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #1775

include/sys/dmu_impl.h

@@ -21,7 +21,10 @@
 /*
  * Copyright 2010 Sun Microsystems, Inc.  All rights reserved.
  * Use is subject to license terms.
+ */
+/*
  * Copyright (c) 2012, Joyent, Inc. All rights reserved.
+ * Copyright (c) 2013 by Delphix. All rights reserved.
  */
 
 #ifndef _SYS_DMU_IMPL_H
@@ -265,6 +268,9 @@ typedef struct dmu_sendarg {
 	uint64_t dsa_toguid;
 	int dsa_err;
 	dmu_pendop_t dsa_pending_op;
+	boolean_t dsa_incremental;
+	uint64_t dsa_last_data_object;
+	uint64_t dsa_last_data_offset;
 } dmu_sendarg_t;
 
 #ifdef	__cplusplus

include/sys/dmu_send.h

@@ -21,7 +21,7 @@
 
 /*
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
- * Copyright (c) 2012 by Delphix. All rights reserved.
+ * Copyright (c) 2013 by Delphix. All rights reserved.
  * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
  * Copyright (c) 2012, Joyent, Inc. All rights reserved.
  */
@@ -63,5 +63,6 @@ int dmu_recv_begin(char *tofs, char *tosnap, struct drr_begin *drrb,
 int dmu_recv_stream(dmu_recv_cookie_t *drc, struct vnode *vp, offset_t *voffp,
     int cleanup_fd, uint64_t *action_handlep);
 int dmu_recv_end(dmu_recv_cookie_t *drc, void *owner);
+boolean_t dmu_objset_is_receiving(objset_t *os);
 
 #endif /* _DMU_SEND_H */

module/zfs/dbuf.c

@@ -28,6 +28,7 @@
 #include <sys/zfs_context.h>
 #include <sys/arc.h>
 #include <sys/dmu.h>
+#include <sys/dmu_send.h>
 #include <sys/dmu_impl.h>
 #include <sys/dbuf.h>
 #include <sys/dmu_objset.h>
@@ -846,9 +847,12 @@ dbuf_unoverride(dbuf_dirty_record_t *dr)
 /*
  * Evict (if its unreferenced) or clear (if its referenced) any level-0
  * data blocks in the free range, so that any future readers will find
- * empty blocks.  Also, if we happen accross any level-1 dbufs in the
+ * empty blocks.  Also, if we happen across any level-1 dbufs in the
  * range that have not already been marked dirty, mark them dirty so
  * they stay in memory.
+ *
+ * This is a no-op if the dataset is in the middle of an incremental
+ * receive; see comment below for details.
  */
 void
 dbuf_free_range(dnode_t *dn, uint64_t start, uint64_t end, dmu_tx_t *tx)
@@ -864,6 +868,20 @@ dbuf_free_range(dnode_t *dn, uint64_t start, uint64_t end, dmu_tx_t *tx)
 		last_l1 = end >> epbs;
 	}
 	dprintf_dnode(dn, "start=%llu end=%llu\n", start, end);
+
+	if (dmu_objset_is_receiving(dn->dn_objset)) {
+		/*
+		 * When processing a free record from a zfs receive,
+		 * there should have been no previous modifications to the
+		 * data in this range.  Therefore there should be no dbufs
+		 * in the range.  Searching dn_dbufs for these non-existent
+		 * dbufs can be very expensive, so simply ignore this.
+		 */
+		VERIFY3P(dbuf_find(dn, 0, start), ==, NULL);
+		VERIFY3P(dbuf_find(dn, 0, end), ==, NULL);
+		return;
+	}
+
 	mutex_enter(&dn->dn_dbufs_mtx);
 	for (db = list_head(&dn->dn_dbufs); db; db = db_next) {
 		db_next = list_next(&dn->dn_dbufs, db);

module/zfs/dmu_send.c

@@ -109,6 +109,32 @@ dump_free(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
 {
 	struct drr_free *drrf = &(dsp->dsa_drr->drr_u.drr_free);
 
+	/*
+	 * When we receive a free record, dbuf_free_range() assumes
+	 * that the receiving system doesn't have any dbufs in the range
+	 * being freed.  This is always true because there is a one-record
+	 * constraint: we only send one WRITE record for any given
+	 * object+offset.  We know that the one-record constraint is
+	 * true because we always send data in increasing order by
+	 * object,offset.
+	 *
+	 * If the increasing-order constraint ever changes, we should find
+	 * another way to assert that the one-record constraint is still
+	 * satisfied.
+	 */
+	ASSERT(object > dsp->dsa_last_data_object ||
+	    (object == dsp->dsa_last_data_object &&
+	    offset > dsp->dsa_last_data_offset));
+
+	/*
+	 * If we are doing a non-incremental send, then there can't
+	 * be any data in the dataset we're receiving into.  Therefore
+	 * a free record would simply be a no-op.  Save space by not
+	 * sending it to begin with.
+	 */
+	if (!dsp->dsa_incremental)
+		return (0);
+
 	if (length != -1ULL && offset + length < offset)
 		length = -1ULL;
 
@@ -175,6 +201,15 @@ dump_data(dmu_sendarg_t *dsp, dmu_object_type_t type,
 {
 	struct drr_write *drrw = &(dsp->dsa_drr->drr_u.drr_write);
 
+	/*
+	 * We send data in increasing object, offset order.
+	 * See comment in dump_free() for details.
+	 */
+	ASSERT(object > dsp->dsa_last_data_object ||
+	    (object == dsp->dsa_last_data_object &&
+	    offset > dsp->dsa_last_data_offset));
+	dsp->dsa_last_data_object = object;
+	dsp->dsa_last_data_offset = offset + blksz - 1;
 
 	/*
 	 * If there is any kind of pending aggregation (currently either
@@ -242,6 +277,10 @@ dump_freeobjects(dmu_sendarg_t *dsp, uint64_t firstobj, uint64_t numobjs)
 {
 	struct drr_freeobjects *drrfo = &(dsp->dsa_drr->drr_u.drr_freeobjects);
 
+	/* See comment in dump_free(). */
+	if (!dsp->dsa_incremental)
+		return (0);
+
 	/*
 	 * If there is a pending op, but it's not PENDING_FREEOBJECTS,
 	 * push it out, since free block aggregation can only be done for
@@ -318,9 +357,9 @@ dump_dnode(dmu_sendarg_t *dsp, uint64_t object, dnode_phys_t *dnp)
 	if (dump_bytes(dsp, DN_BONUS(dnp), P2ROUNDUP(dnp->dn_bonuslen, 8)) != 0)
 		return (SET_ERROR(EINTR));
 
-	/* free anything past the end of the file */
+	/* Free anything past the end of the file. */
 	if (dump_free(dsp, object, (dnp->dn_maxblkid + 1) *
-	    (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT), -1ULL))
+	    (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT), -1ULL) != 0)
 		return (SET_ERROR(EINTR));
 	if (dsp->dsa_err != 0)
 		return (SET_ERROR(EINTR));
@@ -503,6 +542,7 @@ dmu_send_impl(void *tag, dsl_pool_t *dp, dsl_dataset_t *ds,
 	dsp->dsa_toguid = ds->ds_phys->ds_guid;
 	ZIO_SET_CHECKSUM(&dsp->dsa_zc, 0, 0, 0, 0);
 	dsp->dsa_pending_op = PENDING_NONE;
+	dsp->dsa_incremental = (fromtxg != 0);
 
 	mutex_enter(&ds->ds_sendstream_lock);
 	list_insert_head(&ds->ds_sendstreams, dsp);
@@ -1799,3 +1839,13 @@ dmu_recv_end(dmu_recv_cookie_t *drc, void *owner)
 	else
 		return (dmu_recv_existing_end(drc));
 }
+
+/*
+ * Return TRUE if this objset is currently being received into.
+ */
+boolean_t
+dmu_objset_is_receiving(objset_t *os)
+{
+	return (os->os_dsl_dataset != NULL &&
+	    os->os_dsl_dataset->ds_owner == dmu_recv_tag);
+}

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;
-	int err, shift;
 	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
+	 * For i/o error checking, we read the first and last level-0
-	 * blocks, and all the level-1 blocks.  The above count_write's
+	 * blocks if they are not aligned, and all the level-1 blocks.
-	 * have already taken care of the level-0 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;
+		uint64_t start = off >> shift;
-		end = dn->dn_datablkshift ? ((off+len) >> shift) : 0;
+		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);