OpenZFS 9465 - ARC check for 'anon_size > arc_c/2' can stall the system

In the case of one pool being built on another pool, we want
to make sure we don't end up throttling the lower (backing)
pool when the upper pool is the majority contributor to dirty
data. To insure we make forward progress during throttling, we
also check the current pool's net dirty data and only throttle
if it exceeds zfs_arc_pool_dirty_percent of the anonymous dirty
data in the cache.

Authored by: Don Brady <[email protected]>
Reviewed by: Sebastien Roy <[email protected]>
Reviewed by: Matt Ahrens <[email protected]>
Reviewed by: Prashanth Sreenivasa <[email protected]>
Approved by: Robert Mustacchi <[email protected]>
Ported-by: Brian Behlendorf <[email protected]>

Porting Notes:
* The new global variables zfs_arc_dirty_limit_percent,
  zfs_arc_anon_limit_percent, and zfs_arc_pool_dirty_percent
  were intentially not added as tunable module parameters.

OpenZFS-issue: https://illumos.org/issues/9465
OpenZFS-commit: https://github.com/openzfs/openzfs/commit/d6a4c3ef
Closes #7749

6 files changed, 50 insertions, 17 deletions

diff --git a/include/sys/arc.h b/include/sys/arc.h
index 9d6bab505..a5bdefb56 100644
--- a/include/sys/arc.h
+++ b/include/sys/arc.h
@@ -289,7 +289,7 @@ void arc_freed(spa_t *spa, const blkptr_t *bp);
 
 void arc_flush(spa_t *spa, boolean_t retry);
 void arc_tempreserve_clear(uint64_t reserve);
-int arc_tempreserve_space(uint64_t reserve, uint64_t txg);
+int arc_tempreserve_space(spa_t *spa, uint64_t reserve, uint64_t txg);
 
 uint64_t arc_target_bytes(void);
 void arc_init(void);
diff --git a/include/sys/spa.h b/include/sys/spa.h
index 4a3fc71f7..82fe2c18c 100644
--- a/include/sys/spa.h
+++ b/include/sys/spa.h
@@ -990,6 +990,7 @@ extern uint64_t spa_delegation(spa_t *spa);
 extern objset_t *spa_meta_objset(spa_t *spa);
 extern uint64_t spa_deadman_synctime(spa_t *spa);
 extern uint64_t spa_deadman_ziotime(spa_t *spa);
+extern uint64_t spa_dirty_data(spa_t *spa);
 
 /* Miscellaneous support routines */
 extern void spa_load_failed(spa_t *spa, const char *fmt, ...);
diff --git a/include/sys/spa_impl.h b/include/sys/spa_impl.h
index 8d2a31961..8d2a20dbb 100644
--- a/include/sys/spa_impl.h
+++ b/include/sys/spa_impl.h
@@ -363,6 +363,11 @@ struct spa {
 	uint64_t	spa_errata;		/* errata issues detected */
 	spa_stats_t	spa_stats;		/* assorted spa statistics */
 	spa_keystore_t	spa_keystore;		/* loaded crypto keys */
+
+	/* arc_memory_throttle() parameters during low memory condition */
+	uint64_t	spa_lowmem_page_load;	/* memory load during txg */
+	uint64_t	spa_lowmem_last_txg;	/* txg window start */
+
 	hrtime_t	spa_ccw_fail_time;	/* Conf cache write fail time */
 	taskq_t		*spa_zvol_taskq;	/* Taskq for minor management */
 	taskq_t		*spa_prefetch_taskq;	/* Taskq for prefetch threads */
diff --git a/module/zfs/arc.c b/module/zfs/arc.c
index e49f8a547..7f2929c17 100644
--- a/module/zfs/arc.c
+++ b/module/zfs/arc.c
@@ -392,6 +392,16 @@ int zfs_arc_shrink_shift = 0;
 int zfs_arc_p_min_shift = 0;
 int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
 
+/*
+ * ARC dirty data constraints for arc_tempreserve_space() throttle.
+ */
+unsigned long zfs_arc_dirty_limit_percent = 50;	/* total dirty data limit */
+unsigned long zfs_arc_anon_limit_percent = 25;	/* anon block dirty limit */
+unsigned long zfs_arc_pool_dirty_percent = 20;	/* each pool's anon allowance */
+
+/*
+ * Enable or disable compressed arc buffers.
+ */
 int zfs_compressed_arc_enabled = B_TRUE;
 
 /*
@@ -7182,12 +7192,10 @@ arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
 }
 
 static int
-arc_memory_throttle(uint64_t reserve, uint64_t txg)
+arc_memory_throttle(spa_t *spa, uint64_t reserve, uint64_t txg)
 {
 #ifdef _KERNEL
 	uint64_t available_memory = arc_free_memory();
-	static uint64_t page_load = 0;
-	static uint64_t last_txg = 0;
 
 #if defined(_ILP32)
 	available_memory =
@@ -7197,9 +7205,9 @@ arc_memory_throttle(uint64_t reserve, uint64_t txg)
 	if (available_memory > arc_all_memory() * arc_lotsfree_percent / 100)
 		return (0);
 
-	if (txg > last_txg) {
-		last_txg = txg;
-		page_load = 0;
+	if (txg > spa->spa_lowmem_last_txg) {
+		spa->spa_lowmem_last_txg = txg;
+		spa->spa_lowmem_page_load = 0;
 	}
 	/*
 	 * If we are in pageout, we know that memory is already tight,
@@ -7207,21 +7215,22 @@ arc_memory_throttle(uint64_t reserve, uint64_t txg)
 	 * continue to let page writes occur as quickly as possible.
 	 */
 	if (current_is_kswapd()) {
-		if (page_load > MAX(arc_sys_free / 4, available_memory) / 4) {
+		if (spa->spa_lowmem_page_load >
+		    MAX(arc_sys_free / 4, available_memory) / 4) {
 			DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim);
 			return (SET_ERROR(ERESTART));
 		}
 		/* Note: reserve is inflated, so we deflate */
-		page_load += reserve / 8;
+		atomic_add_64(&spa->spa_lowmem_page_load, reserve / 8);
 		return (0);
-	} else if (page_load > 0 && arc_reclaim_needed()) {
+	} else if (spa->spa_lowmem_page_load > 0 && arc_reclaim_needed()) {
 		/* memory is low, delay before restarting */
 		ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
 		DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim);
 		return (SET_ERROR(EAGAIN));
 	}
-	page_load = 0;
-#endif
+	spa->spa_lowmem_page_load = 0;
+#endif /* _KERNEL */
 	return (0);
 }
 
@@ -7233,7 +7242,7 @@ arc_tempreserve_clear(uint64_t reserve)
 }
 
 int
-arc_tempreserve_space(uint64_t reserve, uint64_t txg)
+arc_tempreserve_space(spa_t *spa, uint64_t reserve, uint64_t txg)
 {
 	int error;
 	uint64_t anon_size;
@@ -7269,7 +7278,7 @@ arc_tempreserve_space(uint64_t reserve, uint64_t txg)
 	 * in order to compress/encrypt/etc the data.  We therefore need to
 	 * make sure that there is sufficient available memory for this.
 	 */
-	error = arc_memory_throttle(reserve, txg);
+	error = arc_memory_throttle(spa, reserve, txg);
 	if (error != 0)
 		return (error);
 
@@ -7277,12 +7286,24 @@ arc_tempreserve_space(uint64_t reserve, uint64_t txg)
 	 * Throttle writes when the amount of dirty data in the cache
 	 * gets too large.  We try to keep the cache less than half full
 	 * of dirty blocks so that our sync times don't grow too large.
+	 *
+	 * In the case of one pool being built on another pool, we want
+	 * to make sure we don't end up throttling the lower (backing)
+	 * pool when the upper pool is the majority contributor to dirty
+	 * data. To insure we make forward progress during throttling, we
+	 * also check the current pool's net dirty data and only throttle
+	 * if it exceeds zfs_arc_pool_dirty_percent of the anonymous dirty
+	 * data in the cache.
+	 *
 	 * Note: if two requests come in concurrently, we might let them
 	 * both succeed, when one of them should fail.  Not a huge deal.
 	 */
+	uint64_t total_dirty = reserve + arc_tempreserve + anon_size;
+	uint64_t spa_dirty_anon = spa_dirty_data(spa);
 
-	if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
-	    anon_size > arc_c / 4) {
+	if (total_dirty > arc_c * zfs_arc_dirty_limit_percent / 100 &&
+	    anon_size > arc_c * zfs_arc_anon_limit_percent / 100 &&
+	    spa_dirty_anon > anon_size * zfs_arc_pool_dirty_percent / 100) {
 #ifdef ZFS_DEBUG
 		uint64_t meta_esize =
 		    refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
diff --git a/module/zfs/dsl_dir.c b/module/zfs/dsl_dir.c
index 75c40c68c..9a43691e5 100644
--- a/module/zfs/dsl_dir.c
+++ b/module/zfs/dsl_dir.c
@@ -1416,7 +1416,7 @@ dsl_dir_tempreserve_space(dsl_dir_t *dd, uint64_t lsize, uint64_t asize,
 	    offsetof(struct tempreserve, tr_node));
 	ASSERT3S(asize, >, 0);
 
-	err = arc_tempreserve_space(lsize, tx->tx_txg);
+	err = arc_tempreserve_space(dd->dd_pool->dp_spa, lsize, tx->tx_txg);
 	if (err == 0) {
 		struct tempreserve *tr;
 
diff --git a/module/zfs/spa_misc.c b/module/zfs/spa_misc.c
index 9410fab07..f43a38ef1 100644
--- a/module/zfs/spa_misc.c
+++ b/module/zfs/spa_misc.c
@@ -1917,6 +1917,12 @@ bp_get_dsize(spa_t *spa, const blkptr_t *bp)
 	return (dsize);
 }
 
+uint64_t
+spa_dirty_data(spa_t *spa)
+{
+	return (spa->spa_dsl_pool->dp_dirty_total);
+}
+
 /*
  * ==========================================================================
  * Initialization and Termination