X-Git-Url: https://git.camperquake.de/gitweb.cgi?a=blobdiff_plain;f=module%2Fzfs%2Fzio.c;h=ccefaf8ac1cb82eb541bac20c60c1a07dc5d22e3;hb=34e143323e359b42bc9d06dd19cc4b1f13091283;hp=f0f0cff6b6e5f5492b364e0d057f16ad67d8763c;hpb=bf701a83c5ec192be6d3afe87ebeee45ce9127f4;p=zfs.git

diff --git a/module/zfs/zio.c b/module/zfs/zio.c
index f0f0cff..ccefaf8 100644
--- a/module/zfs/zio.c
+++ b/module/zfs/zio.c
@@ -20,6 +20,8 @@
  */
 /*
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2012 by Delphix. All rights reserved.
+ * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
  */
 
 #include <sys/zfs_context.h>
@@ -61,8 +63,7 @@ uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = {
  * ==========================================================================
  */
 char *zio_type_name[ZIO_TYPES] = {
-	"zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
-	"zio_ioctl"
+	"z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl"
 };
 
 /*
@@ -72,12 +73,29 @@ char *zio_type_name[ZIO_TYPES] = {
  */
 kmem_cache_t *zio_cache;
 kmem_cache_t *zio_link_cache;
+kmem_cache_t *zio_vdev_cache;
 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
+int zio_bulk_flags = 0;
+int zio_delay_max = ZIO_DELAY_MAX;
 
-#ifdef _KERNEL
-extern vmem_t *zio_alloc_arena;
-#endif
+extern int zfs_mg_alloc_failures;
+
+/*
+ * The following actions directly effect the spa's sync-to-convergence logic.
+ * The values below define the sync pass when we start performing the action.
+ * Care should be taken when changing these values as they directly impact
+ * spa_sync() performance. Tuning these values may introduce subtle performance
+ * pathologies and should only be done in the context of performance analysis.
+ * These tunables will eventually be removed and replaced with #defines once
+ * enough analysis has been done to determine optimal values.
+ *
+ * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
+ * regular blocks are not deferred.
+ */
+int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
+int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
+int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
 
 /*
  * An allocating zio is one that either currently has the DVA allocate
@@ -85,7 +103,7 @@ extern vmem_t *zio_alloc_arena;
  */
 #define	IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
 
-boolean_t	zio_requeue_io_start_cut_in_line = B_TRUE;
+int zio_requeue_io_start_cut_in_line = 1;
 
 #ifdef ZFS_DEBUG
 int zio_buf_debug_limit = 16384;
@@ -93,19 +111,49 @@ int zio_buf_debug_limit = 16384;
 int zio_buf_debug_limit = 0;
 #endif
 
+static inline void __zio_execute(zio_t *zio);
+
+static int
+zio_cons(void *arg, void *unused, int kmflag)
+{
+	zio_t *zio = arg;
+
+	bzero(zio, sizeof (zio_t));
+
+	mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
+	cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
+
+	list_create(&zio->io_parent_list, sizeof (zio_link_t),
+	    offsetof(zio_link_t, zl_parent_node));
+	list_create(&zio->io_child_list, sizeof (zio_link_t),
+	    offsetof(zio_link_t, zl_child_node));
+
+	return (0);
+}
+
+static void
+zio_dest(void *arg, void *unused)
+{
+	zio_t *zio = arg;
+
+	mutex_destroy(&zio->io_lock);
+	cv_destroy(&zio->io_cv);
+	list_destroy(&zio->io_parent_list);
+	list_destroy(&zio->io_child_list);
+}
+
 void
 zio_init(void)
 {
 	size_t c;
 	vmem_t *data_alloc_arena = NULL;
 
-#ifdef _KERNEL
-	data_alloc_arena = zio_alloc_arena;
-#endif
-	zio_cache = kmem_cache_create("zio_cache",
-	    sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
+	zio_cache = kmem_cache_create("zio_cache", sizeof (zio_t), 0,
+	    zio_cons, zio_dest, NULL, NULL, NULL, KMC_KMEM);
 	zio_link_cache = kmem_cache_create("zio_link_cache",
-	    sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
+	    sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, KMC_KMEM);
+	zio_vdev_cache = kmem_cache_create("zio_vdev_cache", sizeof(vdev_io_t),
+	    PAGESIZE, NULL, NULL, NULL, NULL, NULL, KMC_VMEM);
 
 	/*
 	 * For small buffers, we want a cache for each multiple of
@@ -131,15 +179,27 @@ zio_init(void)
 
 		if (align != 0) {
 			char name[36];
+			int flags = zio_bulk_flags;
+
+			/*
+			 * The smallest buffers (512b) are heavily used and
+			 * experience a lot of churn.  The slabs allocated
+			 * for them are also relatively small (32K).  Thus
+			 * in over to avoid expensive calls to vmalloc() we
+			 * make an exception to the usual slab allocation
+			 * policy and force these buffers to be kmem backed.
+			 */
+			if (size == (1 << SPA_MINBLOCKSHIFT))
+				flags |= KMC_KMEM;
+
 			(void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
 			zio_buf_cache[c] = kmem_cache_create(name, size,
-			    align, NULL, NULL, NULL, NULL, NULL,
-			    size > zio_buf_debug_limit ? KMC_NODEBUG : 0);
+			    align, NULL, NULL, NULL, NULL, NULL, flags);
 
 			(void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
 			zio_data_buf_cache[c] = kmem_cache_create(name, size,
-			    align, NULL, NULL, NULL, NULL, data_alloc_arena,
-			    size > zio_buf_debug_limit ? KMC_NODEBUG : 0);
+			    align, NULL, NULL, NULL, NULL,
+			    data_alloc_arena, flags);
 		}
 	}
 
@@ -153,7 +213,15 @@ zio_init(void)
 			zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
 	}
 
+	/*
+	 * The zio write taskqs have 1 thread per cpu, allow 1/2 of the taskqs
+	 * to fail 3 times per txg or 8 failures, whichever is greater.
+	 */
+	zfs_mg_alloc_failures = MAX((3 * max_ncpus / 2), 8);
+
 	zio_inject_init();
+
+	lz4_init();
 }
 
 void
@@ -177,10 +245,13 @@ zio_fini(void)
 		zio_data_buf_cache[c] = NULL;
 	}
 
+	kmem_cache_destroy(zio_vdev_cache);
 	kmem_cache_destroy(zio_link_cache);
 	kmem_cache_destroy(zio_cache);
 
 	zio_inject_fini();
+
+	lz4_fini();
 }
 
 /*
@@ -202,7 +273,7 @@ zio_buf_alloc(size_t size)
 
 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
 
-	return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
+	return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE | KM_NODEBUG));
 }
 
 /*
@@ -218,7 +289,8 @@ zio_data_buf_alloc(size_t size)
 
 	ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
 
-	return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
+	return (kmem_cache_alloc(zio_data_buf_cache[c],
+	    KM_PUSHPAGE | KM_NODEBUG));
 }
 
 void
@@ -242,6 +314,24 @@ zio_data_buf_free(void *buf, size_t size)
 }
 
 /*
+ * Dedicated I/O buffers to ensure that memory fragmentation never prevents
+ * or significantly delays the issuing of a zio.   These buffers are used
+ * to aggregate I/O and could be used for raidz stripes.
+ */
+void *
+zio_vdev_alloc(void)
+{
+	return (kmem_cache_alloc(zio_vdev_cache, KM_PUSHPAGE));
+}
+
+void
+zio_vdev_free(void *buf)
+{
+	kmem_cache_free(zio_vdev_cache, buf);
+
+}
+
+/*
  * ==========================================================================
  * Push and pop I/O transform buffers
  * ==========================================================================
@@ -250,7 +340,7 @@ static void
 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
 	zio_transform_func_t *transform)
 {
-	zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
+	zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_PUSHPAGE);
 
 	zt->zt_orig_data = zio->io_data;
 	zt->zt_orig_size = zio->io_size;
@@ -365,7 +455,7 @@ zio_unique_parent(zio_t *cio)
 void
 zio_add_child(zio_t *pio, zio_t *cio)
 {
-	zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
+	zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_PUSHPAGE);
 	int w;
 
 	/*
@@ -451,7 +541,7 @@ zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
 	if (--*countp == 0 && pio->io_stall == countp) {
 		pio->io_stall = NULL;
 		mutex_exit(&pio->io_lock);
-		zio_execute(pio);
+		__zio_execute(pio);
 	} else {
 		mutex_exit(&pio->io_lock);
 	}
@@ -486,16 +576,7 @@ zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
 	ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
 	ASSERT(vd || stage == ZIO_STAGE_OPEN);
 
-	zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
-	bzero(zio, sizeof (zio_t));
-
-	mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
-	cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
-
-	list_create(&zio->io_parent_list, sizeof (zio_link_t),
-	    offsetof(zio_link_t, zl_parent_node));
-	list_create(&zio->io_child_list, sizeof (zio_link_t),
-	    offsetof(zio_link_t, zl_child_node));
+	zio = kmem_cache_alloc(zio_cache, KM_PUSHPAGE);
 
 	if (vd != NULL)
 		zio->io_child_type = ZIO_CHILD_VDEV;
@@ -507,6 +588,7 @@ zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
 		zio->io_child_type = ZIO_CHILD_LOGICAL;
 
 	if (bp != NULL) {
+		zio->io_logical = NULL;
 		zio->io_bp = (blkptr_t *)bp;
 		zio->io_bp_copy = *bp;
 		zio->io_bp_orig = *bp;
@@ -517,21 +599,54 @@ zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
 			zio->io_logical = zio;
 		if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
 			pipeline |= ZIO_GANG_STAGES;
+	} else {
+		zio->io_logical = NULL;
+		zio->io_bp = NULL;
+		bzero(&zio->io_bp_copy, sizeof (blkptr_t));
+		bzero(&zio->io_bp_orig, sizeof (blkptr_t));
 	}
 
 	zio->io_spa = spa;
 	zio->io_txg = txg;
+	zio->io_ready = NULL;
 	zio->io_done = done;
 	zio->io_private = private;
+	zio->io_prev_space_delta = 0;
 	zio->io_type = type;
 	zio->io_priority = priority;
 	zio->io_vd = vd;
+	zio->io_vsd = NULL;
+	zio->io_vsd_ops = NULL;
 	zio->io_offset = offset;
+	zio->io_deadline = 0;
+	zio->io_timestamp = 0;
+	zio->io_delta = 0;
+	zio->io_delay = 0;
 	zio->io_orig_data = zio->io_data = data;
 	zio->io_orig_size = zio->io_size = size;
 	zio->io_orig_flags = zio->io_flags = flags;
 	zio->io_orig_stage = zio->io_stage = stage;
 	zio->io_orig_pipeline = zio->io_pipeline = pipeline;
+	bzero(&zio->io_prop, sizeof (zio_prop_t));
+	zio->io_cmd = 0;
+	zio->io_reexecute = 0;
+	zio->io_bp_override = NULL;
+	zio->io_walk_link = NULL;
+	zio->io_transform_stack = NULL;
+	zio->io_error = 0;
+	zio->io_child_count = 0;
+	zio->io_parent_count = 0;
+	zio->io_stall = NULL;
+	zio->io_gang_leader = NULL;
+	zio->io_gang_tree = NULL;
+	zio->io_executor = NULL;
+	zio->io_waiter = NULL;
+	zio->io_cksum_report = NULL;
+	zio->io_ena = 0;
+	bzero(zio->io_child_error, sizeof (int) * ZIO_CHILD_TYPES);
+	bzero(zio->io_children,
+	    sizeof (uint64_t) * ZIO_CHILD_TYPES * ZIO_WAIT_TYPES);
+	bzero(&zio->io_bookmark, sizeof (zbookmark_t));
 
 	zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
 	zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
@@ -547,16 +662,14 @@ zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
 		zio_add_child(pio, zio);
 	}
 
+	taskq_init_ent(&zio->io_tqent);
+
 	return (zio);
 }
 
 static void
 zio_destroy(zio_t *zio)
 {
-	list_destroy(&zio->io_parent_list);
-	list_destroy(&zio->io_child_list);
-	mutex_destroy(&zio->io_lock);
-	cv_destroy(&zio->io_cv);
 	kmem_cache_free(zio_cache, zio);
 }
 
@@ -607,7 +720,7 @@ zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
 	    zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
 	    zp->zp_compress >= ZIO_COMPRESS_OFF &&
 	    zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
-	    zp->zp_type < DMU_OT_NUMTYPES &&
+	    DMU_OT_IS_VALID(zp->zp_type) &&
 	    zp->zp_level < 32 &&
 	    zp->zp_copies > 0 &&
 	    zp->zp_copies <= spa_max_replication(spa) &&
@@ -668,7 +781,9 @@ zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
 
 	ASSERT(!BP_IS_HOLE(bp));
 	ASSERT(spa_syncing_txg(spa) == txg);
-	ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
+	ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
+
+	arc_freed(spa, bp);
 
 	zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
 	    NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
@@ -891,7 +1006,7 @@ zio_read_bp_init(zio_t *zio)
 		zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
 	}
 
-	if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
+	if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
 		zio->io_flags |= ZIO_FLAG_DONT_CACHE;
 
 	if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
@@ -965,7 +1080,7 @@ zio_write_bp_init(zio_t *zio)
 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
 		ASSERT(!BP_GET_DEDUP(bp));
 
-		if (pass > SYNC_PASS_DONT_COMPRESS)
+		if (pass >= zfs_sync_pass_dont_compress)
 			compress = ZIO_COMPRESS_OFF;
 
 		/* Make sure someone doesn't change their mind on overwrites */
@@ -994,7 +1109,7 @@ zio_write_bp_init(zio_t *zio)
 	 * There should only be a handful of blocks after pass 1 in any case.
 	 */
 	if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
-	    pass > SYNC_PASS_REWRITE) {
+	    pass >= zfs_sync_pass_rewrite) {
 		enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
 		ASSERT(psize != 0);
 		zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
@@ -1046,11 +1161,11 @@ zio_free_bp_init(zio_t *zio)
  */
 
 static void
-zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
+zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
 {
 	spa_t *spa = zio->io_spa;
 	zio_type_t t = zio->io_type;
-	int flags = TQ_NOSLEEP | (cutinline ? TQ_FRONT : 0);
+	int flags = (cutinline ? TQ_FRONT : 0);
 
 	/*
 	 * If we're a config writer or a probe, the normal issue and
@@ -1067,28 +1182,40 @@ zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
 		t = ZIO_TYPE_NULL;
 
 	/*
-	 * If this is a high priority I/O, then use the high priority taskq.
+	 * If this is a high priority I/O, then use the high priority taskq if
+	 * available.
 	 */
 	if (zio->io_priority == ZIO_PRIORITY_NOW &&
-	    spa->spa_zio_taskq[t][q + 1] != NULL)
+	    spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
 		q++;
 
 	ASSERT3U(q, <, ZIO_TASKQ_TYPES);
 
-	while (taskq_dispatch(spa->spa_zio_taskq[t][q],
-	    (task_func_t *)zio_execute, zio, flags) == 0); /* do nothing */
+	/*
+	 * NB: We are assuming that the zio can only be dispatched
+	 * to a single taskq at a time.  It would be a grievous error
+	 * to dispatch the zio to another taskq at the same time.
+	 */
+	ASSERT(taskq_empty_ent(&zio->io_tqent));
+	spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
+	    flags, &zio->io_tqent);
 }
 
 static boolean_t
-zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
+zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
 {
 	kthread_t *executor = zio->io_executor;
 	spa_t *spa = zio->io_spa;
 	zio_type_t t;
 
-	for (t = 0; t < ZIO_TYPES; t++)
-		if (taskq_member(spa->spa_zio_taskq[t][q], executor))
-			return (B_TRUE);
+	for (t = 0; t < ZIO_TYPES; t++) {
+		spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
+		uint_t i;
+		for (i = 0; i < tqs->stqs_count; i++) {
+			if (taskq_member(tqs->stqs_taskq[i], executor))
+				return (B_TRUE);
+		}
+	}
 
 	return (B_FALSE);
 }
@@ -1115,21 +1242,38 @@ zio_interrupt(zio_t *zio)
  * vdev-level caching or aggregation; (5) the I/O is deferred
  * due to vdev-level queueing; (6) the I/O is handed off to
  * another thread.  In all cases, the pipeline stops whenever
- * there's no CPU work; it never burns a thread in cv_wait().
+ * there's no CPU work; it never burns a thread in cv_wait_io().
  *
  * There's no locking on io_stage because there's no legitimate way
  * for multiple threads to be attempting to process the same I/O.
  */
 static zio_pipe_stage_t *zio_pipeline[];
 
+/*
+ * zio_execute() is a wrapper around the static function
+ * __zio_execute() so that we can force  __zio_execute() to be
+ * inlined.  This reduces stack overhead which is important
+ * because __zio_execute() is called recursively in several zio
+ * code paths.  zio_execute() itself cannot be inlined because
+ * it is externally visible.
+ */
 void
 zio_execute(zio_t *zio)
 {
+	__zio_execute(zio);
+}
+
+__attribute__((always_inline))
+static inline void
+__zio_execute(zio_t *zio)
+{
 	zio->io_executor = curthread;
 
 	while (zio->io_stage < ZIO_STAGE_DONE) {
 		enum zio_stage pipeline = zio->io_pipeline;
 		enum zio_stage stage = zio->io_stage;
+		dsl_pool_t *dp;
+		boolean_t cut;
 		int rv;
 
 		ASSERT(!MUTEX_HELD(&zio->io_lock));
@@ -1142,6 +1286,10 @@ zio_execute(zio_t *zio)
 
 		ASSERT(stage <= ZIO_STAGE_DONE);
 
+		dp = spa_get_dsl(zio->io_spa);
+		cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
+		    zio_requeue_io_start_cut_in_line : B_FALSE;
+
 		/*
 		 * If we are in interrupt context and this pipeline stage
 		 * will grab a config lock that is held across I/O,
@@ -1153,12 +1301,27 @@ zio_execute(zio_t *zio)
 		 */
 		if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
 		    zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
-			boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
-			    zio_requeue_io_start_cut_in_line : B_FALSE;
 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
 			return;
 		}
 
+#ifdef _KERNEL
+		/*
+		 * If we executing in the context of the tx_sync_thread,
+		 * or we are performing pool initialization outside of a
+		 * zio_taskq[ZIO_TASKQ_ISSUE|ZIO_TASKQ_ISSUE_HIGH] context.
+		 * Then issue the zio asynchronously to minimize stack usage
+		 * for these deep call paths.
+		 */
+		if ((dp && curthread == dp->dp_tx.tx_sync_thread) ||
+		    (dp && spa_is_initializing(dp->dp_spa) &&
+		    !zio_taskq_member(zio, ZIO_TASKQ_ISSUE) &&
+		    !zio_taskq_member(zio, ZIO_TASKQ_ISSUE_HIGH))) {
+			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
+			return;
+		}
+#endif
+
 		zio->io_stage = stage;
 		rv = zio_pipeline[highbit(stage) - 1](zio);
 
@@ -1169,6 +1332,7 @@ zio_execute(zio_t *zio)
 	}
 }
 
+
 /*
  * ==========================================================================
  * Initiate I/O, either sync or async
@@ -1184,11 +1348,11 @@ zio_wait(zio_t *zio)
 
 	zio->io_waiter = curthread;
 
-	zio_execute(zio);
+	__zio_execute(zio);
 
 	mutex_enter(&zio->io_lock);
 	while (zio->io_executor != NULL)
-		cv_wait(&zio->io_cv, &zio->io_lock);
+		cv_wait_io(&zio->io_cv, &zio->io_lock);
 	mutex_exit(&zio->io_lock);
 
 	error = zio->io_error;
@@ -1214,7 +1378,7 @@ zio_nowait(zio_t *zio)
 		zio_add_child(spa->spa_async_zio_root, zio);
 	}
 
-	zio_execute(zio);
+	__zio_execute(zio);
 }
 
 /*
@@ -1269,7 +1433,7 @@ zio_reexecute(zio_t *pio)
 	 * responsibility of the caller to wait on him.
 	 */
 	if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
-		zio_execute(pio);
+		__zio_execute(pio);
 }
 
 void
@@ -1280,6 +1444,9 @@ zio_suspend(spa_t *spa, zio_t *zio)
 		    "failure and the failure mode property for this pool "
 		    "is set to panic.", spa_name(spa));
 
+	cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable I/O "
+	    "failure and has been suspended.\n", spa_name(spa));
+
 	zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
 
 	mutex_enter(&spa->spa_suspend_lock);
@@ -1482,7 +1649,7 @@ zio_gang_node_alloc(zio_gang_node_t **gnpp)
 
 	ASSERT(*gnpp == NULL);
 
-	gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
+	gn = kmem_zalloc(sizeof (*gn), KM_PUSHPAGE);
 	gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
 	*gnpp = gn;
 
@@ -1733,6 +1900,11 @@ zio_write_gang_block(zio_t *pio)
 	 */
 	pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
 
+	/*
+	 * We didn't allocate this bp, so make sure it doesn't get unmarked.
+	 */
+	pio->io_flags &= ~ZIO_FLAG_FASTWRITE;
+
 	zio_nowait(zio);
 
 	return (ZIO_PIPELINE_CONTINUE);
@@ -1878,7 +2050,7 @@ zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
 
 			ddt_exit(ddt);
 
-			error = arc_read_nolock(NULL, spa, &blk,
+			error = arc_read(NULL, spa, &blk,
 			    arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
 			    &aflags, &zio->io_bookmark);
@@ -2102,8 +2274,11 @@ zio_ddt_free(zio_t *zio)
 
 	ddt_enter(ddt);
 	freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
-	ddp = ddt_phys_select(dde, bp);
-	ddt_phys_decref(ddp);
+	if (dde) {
+		ddp = ddt_phys_select(dde, bp);
+		if (ddp)
+			ddt_phys_decref(ddp);
+	}
 	ddt_exit(ddt);
 
 	return (ZIO_PIPELINE_CONTINUE);
@@ -2121,6 +2296,7 @@ zio_dva_allocate(zio_t *zio)
 	metaslab_class_t *mc = spa_normal_class(spa);
 	blkptr_t *bp = zio->io_bp;
 	int error;
+	int flags = 0;
 
 	if (zio->io_gang_leader == NULL) {
 		ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
@@ -2128,15 +2304,27 @@ zio_dva_allocate(zio_t *zio)
 	}
 
 	ASSERT(BP_IS_HOLE(bp));
-	ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
+	ASSERT0(BP_GET_NDVAS(bp));
 	ASSERT3U(zio->io_prop.zp_copies, >, 0);
 	ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
 	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
 
+	/*
+	 * The dump device does not support gang blocks so allocation on
+	 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
+	 * the "fast" gang feature.
+	 */
+	flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
+	flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
+	    METASLAB_GANG_CHILD : 0;
+	flags |= (zio->io_flags & ZIO_FLAG_FASTWRITE) ? METASLAB_FASTWRITE : 0;
 	error = metaslab_alloc(spa, mc, zio->io_size, bp,
-	    zio->io_prop.zp_copies, zio->io_txg, NULL, 0);
+	    zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
 
 	if (error) {
+		spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
+		    "size %llu, error %d", spa_name(spa), zio, zio->io_size,
+		    error);
 		if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
 			return (zio_write_gang_block(zio));
 		zio->io_error = error;
@@ -2193,20 +2381,29 @@ zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
  * Try to allocate an intent log block.  Return 0 on success, errno on failure.
  */
 int
-zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
-    uint64_t size, boolean_t use_slog)
+zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, uint64_t size,
+    boolean_t use_slog)
 {
 	int error = 1;
 
 	ASSERT(txg > spa_syncing_txg(spa));
 
-	if (use_slog)
+	/*
+	 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
+	 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
+	 * when allocating them.
+	 */
+	if (use_slog) {
 		error = metaslab_alloc(spa, spa_log_class(spa), size,
-		    new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
+		    new_bp, 1, txg, NULL,
+		    METASLAB_FASTWRITE | METASLAB_GANG_AVOID);
+	}
 
-	if (error)
+	if (error) {
 		error = metaslab_alloc(spa, spa_normal_class(spa), size,
-		    new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
+		    new_bp, 1, txg, NULL,
+		    METASLAB_FASTWRITE | METASLAB_GANG_AVOID);
+	}
 
 	if (error == 0) {
 		BP_SET_LSIZE(new_bp, size);
@@ -2663,11 +2860,6 @@ zio_ready(zio_t *zio)
 static int
 zio_done(zio_t *zio)
 {
-	spa_t *spa = zio->io_spa;
-	zio_t *lio = zio->io_logical;
-	blkptr_t *bp = zio->io_bp;
-	vdev_t *vd = zio->io_vd;
-	uint64_t psize = zio->io_size;
 	zio_t *pio, *pio_next;
 	int c, w;
 
@@ -2685,18 +2877,18 @@ zio_done(zio_t *zio)
 		for (w = 0; w < ZIO_WAIT_TYPES; w++)
 			ASSERT(zio->io_children[c][w] == 0);
 
-	if (bp != NULL) {
-		ASSERT(bp->blk_pad[0] == 0);
-		ASSERT(bp->blk_pad[1] == 0);
-		ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
-		    (bp == zio_unique_parent(zio)->io_bp));
-		if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
+	if (zio->io_bp != NULL) {
+		ASSERT(zio->io_bp->blk_pad[0] == 0);
+		ASSERT(zio->io_bp->blk_pad[1] == 0);
+		ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
+		    (zio->io_bp == zio_unique_parent(zio)->io_bp));
+		if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) &&
 		    zio->io_bp_override == NULL &&
 		    !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
-			ASSERT(!BP_SHOULD_BYTESWAP(bp));
-			ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
-			ASSERT(BP_COUNT_GANG(bp) == 0 ||
-			    (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
+			ASSERT(!BP_SHOULD_BYTESWAP(zio->io_bp));
+			ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
+			ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 ||
+			    (BP_COUNT_GANG(zio->io_bp) == BP_GET_NDVAS(zio->io_bp)));
 		}
 	}
 
@@ -2715,13 +2907,13 @@ zio_done(zio_t *zio)
 		while (zio->io_cksum_report != NULL) {
 			zio_cksum_report_t *zcr = zio->io_cksum_report;
 			uint64_t align = zcr->zcr_align;
-			uint64_t asize = P2ROUNDUP(psize, align);
+			uint64_t asize = P2ROUNDUP(zio->io_size, align);
 			char *abuf = zio->io_data;
 
-			if (asize != psize) {
+			if (asize != zio->io_size) {
 				abuf = zio_buf_alloc(asize);
-				bcopy(zio->io_data, abuf, psize);
-				bzero(abuf + psize, asize - psize);
+				bcopy(zio->io_data, abuf, zio->io_size);
+				bzero(abuf + zio->io_size, asize - zio->io_size);
 			}
 
 			zio->io_cksum_report = zcr->zcr_next;
@@ -2729,14 +2921,25 @@ zio_done(zio_t *zio)
 			zcr->zcr_finish(zcr, abuf);
 			zfs_ereport_free_checksum(zcr);
 
-			if (asize != psize)
+			if (asize != zio->io_size)
 				zio_buf_free(abuf, asize);
 		}
 	}
 
 	zio_pop_transforms(zio);	/* note: may set zio->io_error */
 
-	vdev_stat_update(zio, psize);
+	vdev_stat_update(zio, zio->io_size);
+
+	/*
+	 * If this I/O is attached to a particular vdev is slow, exceeding
+	 * 30 seconds to complete, post an error described the I/O delay.
+	 * We ignore these errors if the device is currently unavailable.
+	 */
+	if (zio->io_delay >= MSEC_TO_TICK(zio_delay_max)) {
+		if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd))
+			zfs_ereport_post(FM_EREPORT_ZFS_DELAY, zio->io_spa,
+                                         zio->io_vd, zio, 0, 0);
+	}
 
 	if (zio->io_error) {
 		/*
@@ -2745,28 +2948,30 @@ zio_done(zio_t *zio)
 		 * at the block level.  We ignore these errors if the
 		 * device is currently unavailable.
 		 */
-		if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
-			zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
+		if (zio->io_error != ECKSUM && zio->io_vd != NULL &&
+			!vdev_is_dead(zio->io_vd))
+			zfs_ereport_post(FM_EREPORT_ZFS_IO, zio->io_spa,
+						zio->io_vd, zio, 0, 0);
 
 		if ((zio->io_error == EIO || !(zio->io_flags &
 		    (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
-		    zio == lio) {
+		    zio == zio->io_logical) {
 			/*
 			 * For logical I/O requests, tell the SPA to log the
 			 * error and generate a logical data ereport.
 			 */
-			spa_log_error(spa, zio);
-			zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
+			spa_log_error(zio->io_spa, zio);
+			zfs_ereport_post(FM_EREPORT_ZFS_DATA, zio->io_spa, NULL, zio,
 			    0, 0);
 		}
 	}
 
-	if (zio->io_error && zio == lio) {
+	if (zio->io_error && zio == zio->io_logical) {
 		/*
 		 * Determine whether zio should be reexecuted.  This will
 		 * propagate all the way to the root via zio_notify_parent().
 		 */
-		ASSERT(vd == NULL && bp != NULL);
+		ASSERT(zio->io_vd == NULL && zio->io_bp != NULL);
 		ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
 
 		if (IO_IS_ALLOCATING(zio) &&
@@ -2781,8 +2986,8 @@ zio_done(zio_t *zio)
 		    zio->io_type == ZIO_TYPE_FREE) &&
 		    !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
 		    zio->io_error == ENXIO &&
-		    spa_load_state(spa) == SPA_LOAD_NONE &&
-		    spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
+		    spa_load_state(zio->io_spa) == SPA_LOAD_NONE &&
+		    spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE)
 			zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
 
 		if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
@@ -2808,7 +3013,7 @@ zio_done(zio_t *zio)
 	if ((zio->io_error || zio->io_reexecute) &&
 	    IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
 	    !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
-		zio_dva_unallocate(zio, zio->io_gang_tree, bp);
+		zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp);
 
 	zio_gang_tree_free(&zio->io_gang_tree);
 
@@ -2873,15 +3078,17 @@ zio_done(zio_t *zio)
 			 * We'd fail again if we reexecuted now, so suspend
 			 * until conditions improve (e.g. device comes online).
 			 */
-			zio_suspend(spa, zio);
+			zio_suspend(zio->io_spa, zio);
 		} else {
 			/*
 			 * Reexecution is potentially a huge amount of work.
 			 * Hand it off to the otherwise-unused claim taskq.
 			 */
-			(void) taskq_dispatch(
-			    spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
-			    (task_func_t *)zio_reexecute, zio, TQ_SLEEP);
+			ASSERT(taskq_empty_ent(&zio->io_tqent));
+			spa_taskq_dispatch_ent(zio->io_spa,
+			    ZIO_TYPE_CLAIM, ZIO_TASKQ_ISSUE,
+			    (task_func_t *)zio_reexecute, zio, 0,
+			    &zio->io_tqent);
 		}
 		return (ZIO_PIPELINE_STOP);
 	}
@@ -2901,6 +3108,11 @@ zio_done(zio_t *zio)
 		zfs_ereport_free_checksum(zcr);
 	}
 
+	if (zio->io_flags & ZIO_FLAG_FASTWRITE && zio->io_bp &&
+	    !BP_IS_HOLE(zio->io_bp)) {
+		metaslab_fastwrite_unmark(zio->io_spa, zio->io_bp);
+	}
+
 	/*
 	 * It is the responsibility of the done callback to ensure that this
 	 * particular zio is no longer discoverable for adoption, and as
@@ -2960,3 +3172,79 @@ static zio_pipe_stage_t *zio_pipeline[] = {
 	zio_checksum_verify,
 	zio_done
 };
+
+/* dnp is the dnode for zb1->zb_object */
+boolean_t
+zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_t *zb1,
+    const zbookmark_t *zb2)
+{
+	uint64_t zb1nextL0, zb2thisobj;
+
+	ASSERT(zb1->zb_objset == zb2->zb_objset);
+	ASSERT(zb2->zb_level == 0);
+
+	/*
+	 * A bookmark in the deadlist is considered to be after
+	 * everything else.
+	 */
+	if (zb2->zb_object == DMU_DEADLIST_OBJECT)
+		return (B_TRUE);
+
+	/* The objset_phys_t isn't before anything. */
+	if (dnp == NULL)
+		return (B_FALSE);
+
+	zb1nextL0 = (zb1->zb_blkid + 1) <<
+	    ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
+
+	zb2thisobj = zb2->zb_object ? zb2->zb_object :
+	    zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
+
+	if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
+		uint64_t nextobj = zb1nextL0 *
+		    (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
+		return (nextobj <= zb2thisobj);
+	}
+
+	if (zb1->zb_object < zb2thisobj)
+		return (B_TRUE);
+	if (zb1->zb_object > zb2thisobj)
+		return (B_FALSE);
+	if (zb2->zb_object == DMU_META_DNODE_OBJECT)
+		return (B_FALSE);
+	return (zb1nextL0 <= zb2->zb_blkid);
+}
+
+#if defined(_KERNEL) && defined(HAVE_SPL)
+/* Fault injection */
+EXPORT_SYMBOL(zio_injection_enabled);
+EXPORT_SYMBOL(zio_inject_fault);
+EXPORT_SYMBOL(zio_inject_list_next);
+EXPORT_SYMBOL(zio_clear_fault);
+EXPORT_SYMBOL(zio_handle_fault_injection);
+EXPORT_SYMBOL(zio_handle_device_injection);
+EXPORT_SYMBOL(zio_handle_label_injection);
+EXPORT_SYMBOL(zio_priority_table);
+EXPORT_SYMBOL(zio_type_name);
+
+module_param(zio_bulk_flags, int, 0644);
+MODULE_PARM_DESC(zio_bulk_flags, "Additional flags to pass to bulk buffers");
+
+module_param(zio_delay_max, int, 0644);
+MODULE_PARM_DESC(zio_delay_max, "Max zio millisec delay before posting event");
+
+module_param(zio_requeue_io_start_cut_in_line, int, 0644);
+MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line, "Prioritize requeued I/O");
+
+module_param(zfs_sync_pass_deferred_free, int, 0644);
+MODULE_PARM_DESC(zfs_sync_pass_deferred_free,
+    "defer frees starting in this pass");
+
+module_param(zfs_sync_pass_dont_compress, int, 0644);
+MODULE_PARM_DESC(zfs_sync_pass_dont_compress,
+    "don't compress starting in this pass");
+
+module_param(zfs_sync_pass_rewrite, int, 0644);
+MODULE_PARM_DESC(zfs_sync_pass_rewrite,
+    "rewrite new bps starting in this pass");
+#endif