OpenZFS 6988 spa_sync() spends half its time in dmu_objset_do_userquota_updates

Using a benchmark which creates 2 million files in one TXG, I observe
that the thread running spa_sync() is on CPU almost the entire time we
are syncing, and therefore can be a performance bottleneck. About 50% of
the time in spa_sync() is in dmu_objset_do_userquota_updates().

The problem is that dmu_objset_do_userquota_updates() calls
zap_increment_int(DMU_USERUSED_OBJECT) once for every file that was
modified (or created). In this benchmark, all the files are owned by the
same user/group, so all 2 million calls to zap_increment_int() are
modifying the same entry in the zap. The same issue exists for the
DMU_GROUPUSED_OBJECT.

We should keep an in-memory map from user to space delta while we are
syncing, and when we finish, iterate over the in-memory map and modify
the ZAP once per entry. This reduces the number of calls to
zap_increment_int() from "number of objects modified" to "number of
owners/groups of modified files".

This reduced the time spent in spa_sync() in the file create benchmark
by ~33%, from 11 seconds to 7 seconds.

Upstream bugs: DLPX-44799
Ported by: Ned Bass <[email protected]>

OpenZFS-issue: https://www.illumos.org/issues/6988
ZFSonLinux-issue: https://github.com/zfsonlinux/zfs/issues/4642
OpenZFS-commit: unmerged

Porting notes:
- Added curly braces around declaration of userquota_cache_t cache to
  quiet compiler warning;
- Handled the userobj accounting the same way it proposed in this path.

Signed-off-by: Jinshan Xiong <[email protected]>

2 files changed, 107 insertions, 36 deletions

diff --git a/module/zfs/dmu_objset.c b/module/zfs/dmu_objset.c
index c6b6eb745..b1f05cb61 100644
--- a/module/zfs/dmu_objset.c
+++ b/module/zfs/dmu_objset.c
@@ -1338,37 +1338,113 @@ dmu_objset_userobjused_enabled(objset_t *os)
 	    spa_feature_is_enabled(os->os_spa, SPA_FEATURE_USEROBJ_ACCOUNTING));
 }
 
+typedef struct userquota_node {
+	/* must be in the first filed, see userquota_update_cache() */
+	char		uqn_id[20 + DMU_OBJACCT_PREFIX_LEN];
+	int64_t		uqn_delta;
+	avl_node_t	uqn_node;
+} userquota_node_t;
+
+typedef struct userquota_cache {
+	avl_tree_t uqc_user_deltas;
+	avl_tree_t uqc_group_deltas;
+} userquota_cache_t;
+
+static int
+userquota_compare(const void *l, const void *r)
+{
+	const userquota_node_t *luqn = l;
+	const userquota_node_t *ruqn = r;
+
+	/*
+	 * NB: can only access uqn_id because userquota_update_cache() doesn't
+	 * pass in an entire userquota_node_t.
+	 */
+	return (strcmp(luqn->uqn_id, ruqn->uqn_id));
+}
+
+static void
+do_userquota_cacheflush(objset_t *os, userquota_cache_t *cache, dmu_tx_t *tx)
+{
+	void *cookie;
+	userquota_node_t *uqn;
+
+	ASSERT(dmu_tx_is_syncing(tx));
+
+	cookie = NULL;
+	while ((uqn = avl_destroy_nodes(&cache->uqc_user_deltas,
+	    &cookie)) != NULL) {
+		VERIFY0(zap_increment(os, DMU_USERUSED_OBJECT,
+		    uqn->uqn_id, uqn->uqn_delta, tx));
+		kmem_free(uqn, sizeof (*uqn));
+	}
+	avl_destroy(&cache->uqc_user_deltas);
+
+	cookie = NULL;
+	while ((uqn = avl_destroy_nodes(&cache->uqc_group_deltas,
+	    &cookie)) != NULL) {
+		VERIFY0(zap_increment(os, DMU_GROUPUSED_OBJECT,
+		    uqn->uqn_id, uqn->uqn_delta, tx));
+		kmem_free(uqn, sizeof (*uqn));
+	}
+	avl_destroy(&cache->uqc_group_deltas);
+}
+
+static void
+userquota_update_cache(avl_tree_t *avl, const char *id, int64_t delta)
+{
+	userquota_node_t *uqn;
+	avl_index_t idx;
+
+	ASSERT(strlen(id) < sizeof (uqn->uqn_id));
+	/*
+	 * Use id directly for searching because uqn_id is the first field of
+	 * userquota_node_t and fields after uqn_id won't be accessed in
+	 * avl_find().
+	 */
+	uqn = avl_find(avl, (const void *)id, &idx);
+	if (uqn == NULL) {
+		uqn = kmem_zalloc(sizeof (*uqn), KM_SLEEP);
+		strcpy(uqn->uqn_id, id);
+		avl_insert(avl, uqn, idx);
+	}
+	uqn->uqn_delta += delta;
+}
+
 static void
-do_userquota_update(objset_t *os, uint64_t used, uint64_t flags,
-    uint64_t user, uint64_t group, boolean_t subtract, dmu_tx_t *tx)
+do_userquota_update(userquota_cache_t *cache, uint64_t used, uint64_t flags,
+    uint64_t user, uint64_t group, boolean_t subtract)
 {
 	if ((flags & DNODE_FLAG_USERUSED_ACCOUNTED)) {
 		int64_t delta = DNODE_MIN_SIZE + used;
+		char name[20];
+
 		if (subtract)
 			delta = -delta;
-		VERIFY3U(0, ==, zap_increment_int(os, DMU_USERUSED_OBJECT,
-		    user, delta, tx));
-		VERIFY3U(0, ==, zap_increment_int(os, DMU_GROUPUSED_OBJECT,
-		    group, delta, tx));
+
+		(void) sprintf(name, "%llx", (longlong_t)user);
+		userquota_update_cache(&cache->uqc_user_deltas, name, delta);
+
+		(void) sprintf(name, "%llx", (longlong_t)group);
+		userquota_update_cache(&cache->uqc_group_deltas, name, delta);
 	}
 }
 
 static void
-do_userobjquota_update(objset_t *os, uint64_t flags, uint64_t user,
-    uint64_t group, boolean_t subtract, dmu_tx_t *tx)
+do_userobjquota_update(userquota_cache_t *cache, uint64_t flags,
+    uint64_t user, uint64_t group, boolean_t subtract)
 {
 	if (flags & DNODE_FLAG_USEROBJUSED_ACCOUNTED) {
 		char name[20 + DMU_OBJACCT_PREFIX_LEN];
+		int delta = subtract ? -1 : 1;
 
 		(void) snprintf(name, sizeof (name), DMU_OBJACCT_PREFIX "%llx",
 		    (longlong_t)user);
-		VERIFY0(zap_increment(os, DMU_USERUSED_OBJECT, name,
-		    subtract ? -1 : 1, tx));
+		userquota_update_cache(&cache->uqc_user_deltas, name, delta);
 
 		(void) snprintf(name, sizeof (name), DMU_OBJACCT_PREFIX "%llx",
 		    (longlong_t)group);
-		VERIFY0(zap_increment(os, DMU_GROUPUSED_OBJECT, name,
-		    subtract ? -1 : 1, tx));
+		userquota_update_cache(&cache->uqc_group_deltas, name, delta);
 	}
 }
 
@@ -1377,9 +1453,15 @@ dmu_objset_do_userquota_updates(objset_t *os, dmu_tx_t *tx)
 {
 	dnode_t *dn;
 	list_t *list = &os->os_synced_dnodes;
+	userquota_cache_t cache = { { 0 } };
 
 	ASSERT(list_head(list) == NULL || dmu_objset_userused_enabled(os));
 
+	avl_create(&cache.uqc_user_deltas, userquota_compare,
+	    sizeof (userquota_node_t), offsetof(userquota_node_t, uqn_node));
+	avl_create(&cache.uqc_group_deltas, userquota_compare,
+	    sizeof (userquota_node_t), offsetof(userquota_node_t, uqn_node));
+
 	while ((dn = list_head(list))) {
 		int flags;
 		ASSERT(!DMU_OBJECT_IS_SPECIAL(dn->dn_object));
@@ -1389,36 +1471,27 @@ dmu_objset_do_userquota_updates(objset_t *os, dmu_tx_t *tx)
 
 		/* Allocate the user/groupused objects if necessary. */
 		if (DMU_USERUSED_DNODE(os)->dn_type == DMU_OT_NONE) {
-			VERIFY(0 == zap_create_claim(os,
-			    DMU_USERUSED_OBJECT,
+			VERIFY0(zap_create_claim(os, DMU_USERUSED_OBJECT,
 			    DMU_OT_USERGROUP_USED, DMU_OT_NONE, 0, tx));
-			VERIFY(0 == zap_create_claim(os,
-			    DMU_GROUPUSED_OBJECT,
+			VERIFY0(zap_create_claim(os, DMU_GROUPUSED_OBJECT,
 			    DMU_OT_USERGROUP_USED, DMU_OT_NONE, 0, tx));
 		}
 
-		/*
-		 * We intentionally modify the zap object even if the
-		 * net delta is zero.  Otherwise
-		 * the block of the zap obj could be shared between
-		 * datasets but need to be different between them after
-		 * a bprewrite.
-		 */
-
 		flags = dn->dn_id_flags;
 		ASSERT(flags);
 		if (flags & DN_ID_OLD_EXIST)  {
-			do_userquota_update(os, dn->dn_oldused, dn->dn_oldflags,
-			    dn->dn_olduid, dn->dn_oldgid, B_TRUE, tx);
-			do_userobjquota_update(os, dn->dn_oldflags,
-			    dn->dn_olduid, dn->dn_oldgid, B_TRUE, tx);
+			do_userquota_update(&cache,
+			    dn->dn_oldused, dn->dn_oldflags,
+			    dn->dn_olduid, dn->dn_oldgid, B_TRUE);
+			do_userobjquota_update(&cache, dn->dn_oldflags,
+			    dn->dn_olduid, dn->dn_oldgid, B_TRUE);
 		}
 		if (flags & DN_ID_NEW_EXIST) {
-			do_userquota_update(os, DN_USED_BYTES(dn->dn_phys),
-			    dn->dn_phys->dn_flags,  dn->dn_newuid,
-			    dn->dn_newgid, B_FALSE, tx);
-			do_userobjquota_update(os, dn->dn_phys->dn_flags,
-			    dn->dn_newuid, dn->dn_newgid, B_FALSE, tx);
+			do_userquota_update(&cache,
+			    DN_USED_BYTES(dn->dn_phys), dn->dn_phys->dn_flags,
+			    dn->dn_newuid, dn->dn_newgid, B_FALSE);
+			do_userobjquota_update(&cache, dn->dn_phys->dn_flags,
+			    dn->dn_newuid, dn->dn_newgid, B_FALSE);
 		}
 
 		mutex_enter(&dn->dn_mtx);
@@ -1439,6 +1512,7 @@ dmu_objset_do_userquota_updates(objset_t *os, dmu_tx_t *tx)
 		list_remove(list, dn);
 		dnode_rele(dn, list);
 	}
+	do_userquota_cacheflush(os, &cache, tx);
 }
 
 /*
diff --git a/module/zfs/dnode.c b/module/zfs/dnode.c
index f0b03bbba..6ba8207e2 100644
--- a/module/zfs/dnode.c
+++ b/module/zfs/dnode.c
@@ -2001,9 +2001,6 @@ dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
 	boolean_t hole;
 	int i, inc, error, span;
 
-	dprintf("probing object %llu offset %llx level %d of %u\n",
-	    dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
-
 	hole = ((flags & DNODE_FIND_HOLE) != 0);
 	inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
 	ASSERT(txg == 0 || !hole);