ZFS send should use spill block prefetched from send_reader_thread

Currently, even though send_reader_thread prefetches spill block, do_dump() will not use it and issues its own blocking arc_read. This causes significant performance degradation when sending datasets with lots of spill blocks. Signed-off-by: Chunwei Chen <[email protected]>
openzfs · Oct 31, 2024 · 0ae990d · 0ae990d
1 parent ae93aeb
commit 0ae990d
Showing 1 changed file with 10 additions and 30 deletions.
diff --git a/module/zfs/dmu_send.c b/module/zfs/dmu_send.c
@@ -939,34 +939,7 @@ do_dump(dmu_send_cookie_t *dscp, struct send_range *range)
 
 		ASSERT3U(srdp->datablksz, ==, BP_GET_LSIZE(bp));
 		ASSERT3U(range->start_blkid + 1, ==, range->end_blkid);
-		if (BP_GET_TYPE(bp) == DMU_OT_SA) {
-			arc_flags_t aflags = ARC_FLAG_WAIT;
-			zio_flag_t zioflags = ZIO_FLAG_CANFAIL;
-
-			if (dscp->dsc_featureflags & DMU_BACKUP_FEATURE_RAW) {
-				ASSERT(BP_IS_PROTECTED(bp));
-				zioflags |= ZIO_FLAG_RAW;
-			}
 
-			zbookmark_phys_t zb;
-			ASSERT3U(range->start_blkid, ==, DMU_SPILL_BLKID);
-			zb.zb_objset = dmu_objset_id(dscp->dsc_os);
-			zb.zb_object = range->object;
-			zb.zb_level = 0;
-			zb.zb_blkid = range->start_blkid;
-
-			arc_buf_t *abuf = NULL;
-			if (!dscp->dsc_dso->dso_dryrun && arc_read(NULL, spa,
-			    bp, arc_getbuf_func, &abuf, ZIO_PRIORITY_ASYNC_READ,
-			    zioflags, &aflags, &zb) != 0)
-				return (SET_ERROR(EIO));
-
-			err = dump_spill(dscp, bp, zb.zb_object,
-			    (abuf == NULL ? NULL : abuf->b_data));
-			if (abuf != NULL)
-				arc_buf_destroy(abuf, &abuf);
-			return (err);
-		}
 		if (send_do_embed(bp, dscp->dsc_featureflags)) {
 			err = dump_write_embedded(dscp, range->object,
 			    range->start_blkid * srdp->datablksz,
@@ -975,8 +948,9 @@ do_dump(dmu_send_cookie_t *dscp, struct send_range *range)
 		}
 		ASSERT(range->object > dscp->dsc_resume_object ||
 		    (range->object == dscp->dsc_resume_object &&
+		    (range->start_blkid == DMU_SPILL_BLKID ||
 		    range->start_blkid * srdp->datablksz >=
-		    dscp->dsc_resume_offset));
+		    dscp->dsc_resume_offset)));
 		/* it's a level-0 block of a regular object */
 
 		mutex_enter(&srdp->lock);
@@ -1006,8 +980,6 @@ do_dump(dmu_send_cookie_t *dscp, struct send_range *range)
 		ASSERT(dscp->dsc_dso->dso_dryrun ||
 		    srdp->abuf != NULL || srdp->abd != NULL);
 
-		uint64_t offset = range->start_blkid * srdp->datablksz;
-
 		char *data = NULL;
 		if (srdp->abd != NULL) {
 			data = abd_to_buf(srdp->abd);
@@ -1016,6 +988,14 @@ do_dump(dmu_send_cookie_t *dscp, struct send_range *range)
 			data = srdp->abuf->b_data;
 		}
 
+		if (BP_GET_TYPE(bp) == DMU_OT_SA) {
+			ASSERT3U(range->start_blkid, ==, DMU_SPILL_BLKID);
+			err = dump_spill(dscp, bp, range->object, data);
+			return (err);
+		}
+
+		uint64_t offset = range->start_blkid * srdp->datablksz;
+
 		/*
 		 * If we have large blocks stored on disk but the send flags
 		 * don't allow us to send large blocks, we split the data from