Illumos 5176 - lock contention on godfather zio

5176 lock contention on godfather zio
Reviewed by: Adam Leventhal <ahl@delphix.com>
Reviewed by: Alex Reece <alex.reece@delphix.com>
Reviewed by: Christopher Siden <christopher.siden@delphix.com>
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Richard Elling <richard.elling@gmail.com>
Reviewed by: Bayard Bell <Bayard.Bell@nexenta.com>
Approved by: Garrett D'Amore <garrett@damore.org>

References:
  https://www.illumos.org/issues/5176
  https://github.com/illumos/illumos-gate/commit/6f834bc

Porting notes:

Under Linux max_ncpus is defined as num_possible_cpus().  This is
largest number of cpu ids which might be available during the life
time of the system boot.  This value can be larger than the number
of present cpus if CONFIG_HOTPLUG_CPU is defined.

Ported by: Turbo Fredriksson <turbo@bayour.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #2711

cmd/zdb/zdb.c

@@ -2565,7 +2565,7 @@ dump_block_stats(spa_t *spa)
 	uint64_t norm_alloc, norm_space, total_alloc, total_found;
 	int flags = TRAVERSE_PRE | TRAVERSE_PREFETCH_METADATA | TRAVERSE_HARD;
 	boolean_t leaks = B_FALSE;
-	int e;
+	int e, c;
 	bp_embedded_type_t i;
 
 	(void) printf("\nTraversing all blocks %s%s%s%s%s...\n\n",
@@ -2614,10 +2614,12 @@ dump_block_stats(spa_t *spa)
 	 * all async I/Os to complete.
 	 */
 	if (dump_opt['c']) {
-		(void) zio_wait(spa->spa_async_zio_root);
+		for (c = 0; c < max_ncpus; c++) {
-		spa->spa_async_zio_root = zio_root(spa, NULL, NULL,
+			(void) zio_wait(spa->spa_async_zio_root[c]);
-		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
+			spa->spa_async_zio_root[c] = zio_root(spa, NULL, NULL,
-		    ZIO_FLAG_GODFATHER);
+			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
+			    ZIO_FLAG_GODFATHER);
+		}
 	}
 
 	if (zcb.zcb_haderrors) {

include/sys/spa_impl.h

@@ -204,7 +204,8 @@ struct spa {
 	uint64_t	spa_failmode;		/* failure mode for the pool */
 	uint64_t	spa_delegation;		/* delegation on/off */
 	list_t		spa_config_list;	/* previous cache file(s) */
-	zio_t		*spa_async_zio_root;	/* root of all async I/O */
+	/* per-CPU array of root of async I/O: */
+	zio_t		**spa_async_zio_root;
 	zio_t		*spa_suspend_zio_root;	/* root of all suspended I/O */
 	kmutex_t	spa_suspend_lock;	/* protects suspend_zio_root */
 	kcondvar_t	spa_suspend_cv;		/* notification of resume */

module/zfs/spa.c

@@ -1252,7 +1252,9 @@ spa_unload(spa_t *spa)
 	 * Wait for any outstanding async I/O to complete.
 	 */
 	if (spa->spa_async_zio_root != NULL) {
-		(void) zio_wait(spa->spa_async_zio_root);
+		for (i = 0; i < max_ncpus; i++)
+			(void) zio_wait(spa->spa_async_zio_root[i]);
+		kmem_free(spa->spa_async_zio_root, max_ncpus * sizeof (void *));
 		spa->spa_async_zio_root = NULL;
 	}
 
@@ -2165,7 +2167,7 @@ spa_load_impl(spa_t *spa, uint64_t pool_guid, nvlist_t *config,
 	uberblock_t *ub = &spa->spa_uberblock;
 	uint64_t children, config_cache_txg = spa->spa_config_txg;
 	int orig_mode = spa->spa_mode;
-	int parse;
+	int parse, i;
 	uint64_t obj;
 	boolean_t missing_feat_write = B_FALSE;
 
@@ -2189,8 +2191,13 @@ spa_load_impl(spa_t *spa, uint64_t pool_guid, nvlist_t *config,
 	/*
 	 * Create "The Godfather" zio to hold all async IOs
 	 */
-	spa->spa_async_zio_root = zio_root(spa, NULL, NULL,
+	spa->spa_async_zio_root = kmem_alloc(max_ncpus * sizeof (void *),
-	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_GODFATHER);
+	    KM_SLEEP);
+	for (i = 0; i < max_ncpus; i++) {
+		spa->spa_async_zio_root[i] = zio_root(spa, NULL, NULL,
+		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
+		    ZIO_FLAG_GODFATHER);
+	}
 
 	/*
 	 * Parse the configuration into a vdev tree.  We explicitly set the
@@ -3495,7 +3502,7 @@ spa_create(const char *pool, nvlist_t *nvroot, nvlist_t *props,
 	uint64_t version, obj;
 	boolean_t has_features;
 	nvpair_t *elem;
-	int c;
+	int c, i;
 	char *poolname;
 	nvlist_t *nvl;
 
@@ -3556,8 +3563,13 @@ spa_create(const char *pool, nvlist_t *nvroot, nvlist_t *props,
 	/*
 	 * Create "The Godfather" zio to hold all async IOs
 	 */
-	spa->spa_async_zio_root = zio_root(spa, NULL, NULL,
+	spa->spa_async_zio_root = kmem_alloc(max_ncpus * sizeof (void *),
-	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_GODFATHER);
+	    KM_SLEEP);
+	for (i = 0; i < max_ncpus; i++) {
+		spa->spa_async_zio_root[i] = zio_root(spa, NULL, NULL,
+		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
+		    ZIO_FLAG_GODFATHER);
+	}
 
 	/*
 	 * Create the root vdev.

module/zfs/zio.c

@@ -1474,7 +1474,7 @@ zio_nowait(zio_t *zio)
 		 */
 		spa_t *spa = zio->io_spa;
 
-		zio_add_child(spa->spa_async_zio_root, zio);
+		zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
 	}
 
 	__zio_execute(zio);