OpenZFS 9442 - decrease indirect block size of spacemaps

Authored by: Matthew Ahrens <[email protected]>
Reviewed by: Serapheim Dimitropoulos <[email protected]>
Reviewed by: George Wilson <[email protected]>
Reviewed by: Albert Lee <[email protected]>
Reviewed by: Igor Kozhukhov <[email protected]>
Reviewed by: George Melikov <[email protected]>
Approved by: Dan McDonald <[email protected]>
Ported-by: Brian Behlendorf <[email protected]>

Updates to indirect blocks of spacemaps can contribute significantly to
write inflation.  Therefore we want to reduce the indirect block size of
spacemaps from 128K to 16K.

Porting notes:
* Refactored to allow the dmu_object_alloc(), dmu_object_alloc_ibs()
  and dmu_object_alloc_dnsize() functions to use a common shared
  dmu_object_alloc_impl() function.

OpenZFS-issue: https://www.illumos.org/issues/9442
OpenZFS-commit: https://github.com/openzfs/openzfs/commit/0c2e6408b
Closes #7712

2 files changed, 47 insertions, 18 deletions

diff --git a/module/zfs/dmu_object.c b/module/zfs/dmu_object.c
index 586a04b16..b9960782e 100644
--- a/module/zfs/dmu_object.c
+++ b/module/zfs/dmu_object.c
@@ -41,17 +41,10 @@
  */
 int dmu_object_alloc_chunk_shift = 7;
 
-uint64_t
-dmu_object_alloc(objset_t *os, dmu_object_type_t ot, int blocksize,
-    dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
-{
-	return dmu_object_alloc_dnsize(os, ot, blocksize, bonustype, bonuslen,
-	    0, tx);
-}
-
-uint64_t
-dmu_object_alloc_dnsize(objset_t *os, dmu_object_type_t ot, int blocksize,
-    dmu_object_type_t bonustype, int bonuslen, int dnodesize, dmu_tx_t *tx)
+static uint64_t
+dmu_object_alloc_impl(objset_t *os, dmu_object_type_t ot, int blocksize,
+    int indirect_blockshift, dmu_object_type_t bonustype, int bonuslen,
+    int dnodesize, dmu_tx_t *tx)
 {
 	uint64_t object;
 	uint64_t L1_dnode_count = DNODES_PER_BLOCK <<
@@ -182,8 +175,9 @@ dmu_object_alloc_dnsize(objset_t *os, dmu_object_type_t ot, int blocksize,
 			 * again now that we have the struct lock.
 			 */
 			if (dn->dn_type == DMU_OT_NONE) {
-				dnode_allocate(dn, ot, blocksize, 0,
-				    bonustype, bonuslen, dn_slots, tx);
+				dnode_allocate(dn, ot, blocksize,
+				    indirect_blockshift, bonustype,
+				    bonuslen, dn_slots, tx);
 				rw_exit(&dn->dn_struct_rwlock);
 				dmu_tx_add_new_object(tx, dn);
 				dnode_rele(dn, FTAG);
@@ -206,6 +200,31 @@ dmu_object_alloc_dnsize(objset_t *os, dmu_object_type_t ot, int blocksize,
 	}
 }
 
+uint64_t
+dmu_object_alloc(objset_t *os, dmu_object_type_t ot, int blocksize,
+    dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
+{
+	return dmu_object_alloc_impl(os, ot, blocksize, 0, bonustype,
+	    bonuslen, 0, tx);
+}
+
+uint64_t
+dmu_object_alloc_ibs(objset_t *os, dmu_object_type_t ot, int blocksize,
+    int indirect_blockshift, dmu_object_type_t bonustype, int bonuslen,
+    dmu_tx_t *tx)
+{
+	return dmu_object_alloc_impl(os, ot, blocksize, indirect_blockshift,
+	    bonustype, bonuslen, 0, tx);
+}
+
+uint64_t
+dmu_object_alloc_dnsize(objset_t *os, dmu_object_type_t ot, int blocksize,
+    dmu_object_type_t bonustype, int bonuslen, int dnodesize, dmu_tx_t *tx)
+{
+	return (dmu_object_alloc_impl(os, ot, blocksize, 0, bonustype,
+	    bonuslen, dnodesize, tx));
+}
+
 int
 dmu_object_claim(objset_t *os, uint64_t object, dmu_object_type_t ot,
     int blocksize, dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
@@ -423,6 +442,7 @@ dmu_object_free_zapified(objset_t *mos, uint64_t object, dmu_tx_t *tx)
 
 #if defined(_KERNEL)
 EXPORT_SYMBOL(dmu_object_alloc);
+EXPORT_SYMBOL(dmu_object_alloc_ibs);
 EXPORT_SYMBOL(dmu_object_alloc_dnsize);
 EXPORT_SYMBOL(dmu_object_claim);
 EXPORT_SYMBOL(dmu_object_claim_dnsize);
diff --git a/module/zfs/space_map.c b/module/zfs/space_map.c
index 5f67a7987..9ba6ff6ff 100644
--- a/module/zfs/space_map.c
+++ b/module/zfs/space_map.c
@@ -52,6 +52,14 @@
  */
 boolean_t zfs_force_some_double_word_sm_entries = B_FALSE;
 
+/*
+ * Override the default indirect block size of 128K, instead use 16K for
+ * spacemaps (2^14 bytes).  This dramatically reduces write inflation since
+ * appending to a spacemap typically has to write one data block (4KB) and one
+ * or two indirect blocks (16K-32K, rather than 128K).
+ */
+int space_map_ibs = 14;
+
 boolean_t
 sm_entry_is_debug(uint64_t e)
 {
@@ -674,8 +682,8 @@ space_map_write_impl(space_map_t *sm, range_tree_t *rt, maptype_t maptype,
 		 *
 		 * [1] The feature is enabled.
 		 * [2] The offset or run is too big for a single-word entry,
-		 * 	or the vdev_id is set (meaning not equal to
-		 * 	SM_NO_VDEVID).
+		 *	or the vdev_id is set (meaning not equal to
+		 *	SM_NO_VDEVID).
 		 *
 		 * Note that for purposes of testing we've added the case that
 		 * we write two-word entries occasionally when the feature is
@@ -837,7 +845,8 @@ space_map_truncate(space_map_t *sm, int blocksize, dmu_tx_t *tx)
 	 */
 	if ((spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM) &&
 	    doi.doi_bonus_size != sizeof (space_map_phys_t)) ||
-	    doi.doi_data_block_size != blocksize) {
+	    doi.doi_data_block_size != blocksize ||
+	    doi.doi_metadata_block_size != 1 << space_map_ibs) {
 		zfs_dbgmsg("txg %llu, spa %s, sm %p, reallocating "
 		    "object[%llu]: old bonus %u, old blocksz %u",
 		    dmu_tx_get_txg(tx), spa_name(spa), sm, sm->sm_object,
@@ -893,8 +902,8 @@ space_map_alloc(objset_t *os, int blocksize, dmu_tx_t *tx)
 		bonuslen = SPACE_MAP_SIZE_V0;
 	}
 
-	object = dmu_object_alloc(os, DMU_OT_SPACE_MAP, blocksize,
-	    DMU_OT_SPACE_MAP_HEADER, bonuslen, tx);
+	object = dmu_object_alloc_ibs(os, DMU_OT_SPACE_MAP, blocksize,
+	    space_map_ibs, DMU_OT_SPACE_MAP_HEADER, bonuslen, tx);
 
 	return (object);
 }