Add 'feature-commit-cb' branch for DMU commit callbacks.

9 files changed, 415 insertions, 11 deletions

diff --git a/.topdeps b/.topdeps
new file mode 100644
index 000000000..1f7391f92
--- /dev/null
+++ b/.topdeps
@@ -0,0 +1 @@
+master
diff --git a/.topmsg b/.topmsg
new file mode 100644
index 000000000..33c1e1c91
--- /dev/null
+++ b/.topmsg
@@ -0,0 +1,8 @@
+From: Brian Behlendorf <[email protected]>
+Subject: [PATCH] feature commit cb
+
+DMU commit callbacks. Provides an API to attach callbacks to transactions,
+which are called by the DMU when they are safely committed to disk.
+See Lustre bug 14117 for details.
+
+Signed-off-by: Brian Behlendorf <[email protected]>
diff --git a/cmd/ztest/ztest.c b/cmd/ztest/ztest.c
index 4503a3d02..efe4fd06b 100644
--- a/cmd/ztest/ztest.c
+++ b/cmd/ztest/ztest.c
@@ -164,6 +164,7 @@ typedef void ztest_func_t(ztest_args_t *);
 ztest_func_t ztest_dmu_read_write;
 ztest_func_t ztest_dmu_write_parallel;
 ztest_func_t ztest_dmu_object_alloc_free;
+ztest_func_t ztest_dmu_commit_callbacks;
 ztest_func_t ztest_zap;
 ztest_func_t ztest_zap_parallel;
 ztest_func_t ztest_traverse;
@@ -198,6 +199,7 @@ ztest_info_t ztest_info[] = {
 	{ ztest_dmu_read_write,			1,	&zopt_always	},
 	{ ztest_dmu_write_parallel,		30,	&zopt_always	},
 	{ ztest_dmu_object_alloc_free,		1,	&zopt_always	},
+	{ ztest_dmu_commit_callbacks,		10,	&zopt_always	},
 	{ ztest_zap,				30,	&zopt_always	},
 	{ ztest_zap_parallel,			100,	&zopt_always	},
 	{ ztest_dsl_prop_get_set,		1,	&zopt_sometimes	},
@@ -218,6 +220,16 @@ ztest_info_t ztest_info[] = {
 #define	ZTEST_SYNC_LOCKS	16
 
 /*
+ * The following struct is used to hold a list of uncalled commit callbacks.
+ *
+ * The callbacks are ordered by txg number.
+ */
+typedef struct ztest_cb_list {
+	mutex_t	zcl_callbacks_lock;
+	list_t	zcl_callbacks;
+} ztest_cb_list_t;
+
+/*
  * Stuff we need to share writably between parent and child.
  */
 typedef struct ztest_shared {
@@ -233,6 +245,7 @@ typedef struct ztest_shared {
 	ztest_info_t	zs_info[ZTEST_FUNCS];
 	mutex_t		zs_sync_lock[ZTEST_SYNC_LOCKS];
 	uint64_t	zs_seq[ZTEST_SYNC_LOCKS];
+	ztest_cb_list_t	zs_cb_list;
 } ztest_shared_t;
 
 static char ztest_dev_template[] = "%s/%s.%llua";
@@ -2433,6 +2446,205 @@ ztest_zap_parallel(ztest_args_t *za)
 		dmu_tx_commit(tx);
 }
 
+/*
+ * Commit callback data.
+ */
+typedef struct ztest_cb_data {
+	list_node_t		zcd_node;
+	ztest_cb_list_t		*zcd_zcl;
+	uint64_t		zcd_txg;
+	int			zcd_expected_err;
+	boolean_t		zcd_added;
+	boolean_t		zcd_called;
+	spa_t			*zcd_spa;
+} ztest_cb_data_t;
+
+static void
+ztest_commit_callback(void *arg, int error)
+{
+	ztest_cb_data_t *data = arg;
+	ztest_cb_list_t *zcl;
+	uint64_t synced_txg;
+
+	VERIFY(data != NULL);
+	VERIFY3S(data->zcd_expected_err, ==, error);
+	VERIFY(!data->zcd_called);
+
+	synced_txg = spa_last_synced_txg(data->zcd_spa);
+	if (data->zcd_txg > synced_txg)
+		fatal(0, "commit callback of txg %llu prematurely called, last"
+		    " synced txg = %llu\n", data->zcd_txg, synced_txg);
+
+	zcl = data->zcd_zcl;
+	data->zcd_called = B_TRUE;
+
+	if (error == ECANCELED) {
+		ASSERT3U(data->zcd_txg, ==, 0);
+		ASSERT(!data->zcd_added);
+
+		/*
+		 * The private callback data should be destroyed here, but
+		 * since we are going to check the zcd_called field after
+		 * dmu_tx_abort(), we will destroy it there.
+		 */
+		return;
+	}
+
+	if (!data->zcd_added)
+		goto out;
+
+	ASSERT3U(data->zcd_txg, !=, 0);
+
+	/* Remove our callback from the list */
+	(void) mutex_lock(&zcl->zcl_callbacks_lock);
+	list_remove(&zcl->zcl_callbacks, data);
+	(void) mutex_unlock(&zcl->zcl_callbacks_lock);
+
+out:
+	umem_free(data, sizeof (ztest_cb_data_t));
+}
+
+static ztest_cb_data_t *
+ztest_create_cb_data(objset_t *os, ztest_cb_list_t *zcl, uint64_t txg)
+{
+	ztest_cb_data_t *cb_data;
+
+	cb_data = umem_zalloc(sizeof (ztest_cb_data_t), UMEM_NOFAIL);
+
+	cb_data->zcd_zcl = zcl;
+	cb_data->zcd_txg = txg;
+	cb_data->zcd_spa = os->os->os_spa;
+
+	return (cb_data);
+}
+
+/*
+ * If a number of txgs equal to this threshold have been created after a commit
+ * callback has been registered but not called, then we assume there is an
+ * implementation bug.
+ */
+#define	ZTEST_COMMIT_CALLBACK_THRESH	3
+
+/*
+ * Commit callback test.
+ */
+void
+ztest_dmu_commit_callbacks(ztest_args_t *za)
+{
+	objset_t *os = za->za_os;
+	dmu_tx_t *tx;
+	ztest_cb_list_t *zcl = &ztest_shared->zs_cb_list;
+	ztest_cb_data_t *cb_data[3], *tmp_cb;
+	uint64_t old_txg, txg;
+	int i, error;
+
+	tx = dmu_tx_create(os);
+
+	cb_data[0] = ztest_create_cb_data(os, zcl, 0);
+	dmu_tx_callback_register(tx, ztest_commit_callback, cb_data[0]);
+
+	dmu_tx_hold_write(tx, ZTEST_DIROBJ, za->za_diroff, sizeof (uint64_t));
+
+	/* Every once in a while, abort the transaction on purpose */
+	if (ztest_random(100) == 0)
+		error = -1;
+
+	if (!error)
+		error = dmu_tx_assign(tx, TXG_NOWAIT);
+
+	txg = error ? 0 : dmu_tx_get_txg(tx);
+
+	cb_data[1] = ztest_create_cb_data(os, zcl, txg);
+	dmu_tx_callback_register(tx, ztest_commit_callback, cb_data[1]);
+
+	if (error) {
+		/*
+		 * It's not a strict requirement to call the registered
+		 * callbacks from inside dmu_tx_abort(), but that's what
+		 * happens in the current implementation so we will check for
+		 * that.
+		 */
+		for (i = 0; i < 2; i++) {
+			cb_data[i]->zcd_expected_err = ECANCELED;
+			VERIFY(!cb_data[i]->zcd_called);
+		}
+
+		dmu_tx_abort(tx);
+
+		for (i = 0; i < 2; i++) {
+			VERIFY(cb_data[i]->zcd_called);
+			umem_free(cb_data[i], sizeof (ztest_cb_data_t));
+		}
+
+		return;
+	}
+
+	cb_data[0]->zcd_txg = txg;
+	cb_data[2] = ztest_create_cb_data(os, zcl, txg);
+	dmu_tx_callback_register(tx, ztest_commit_callback, cb_data[2]);
+
+	/*
+	 * Read existing data to make sure there isn't a future leak.
+	 */
+	VERIFY(0 == dmu_read(os, ZTEST_DIROBJ, za->za_diroff, sizeof (uint64_t),
+	    &old_txg));
+
+	if (old_txg > txg)
+		fatal(0, "future leak: got %llx, open txg is %llx", old_txg,
+		    txg);
+
+	dmu_write(os, ZTEST_DIROBJ, za->za_diroff, sizeof (uint64_t), &txg, tx);
+
+	(void) mutex_lock(&zcl->zcl_callbacks_lock);
+
+	/*
+	 * Since commit callbacks don't have any ordering requirement and since
+	 * it is theoretically possible for a commit callback to be called
+	 * after an arbitrary amount of time has elapsed since its txg has been
+	 * synced, it is difficult to reliably determine whether a commit
+	 * callback hasn't been called due to high load or due to a flawed
+	 * implementation.
+	 *
+	 * In practice, we will assume that if after a certain number of txgs a
+	 * commit callback hasn't been called, then most likely there's an
+	 * implementation bug..
+	 */
+	tmp_cb = list_head(&zcl->zcl_callbacks);
+	if (tmp_cb != NULL)
+		VERIFY3U(tmp_cb->zcd_txg, >,
+		    txg - ZTEST_COMMIT_CALLBACK_THRESH);
+
+	/*
+	 * Let's find the place to insert our callbacks.
+	 *
+	 * Even though the list is ordered by txg, it is possible for the
+	 * insertion point to not be the end because our txg may already be
+	 * quiescing at this point and other callbacks in the open txg may
+	 * have sneaked in.
+	 */
+	tmp_cb = list_tail(&zcl->zcl_callbacks);
+	while (tmp_cb != NULL && tmp_cb->zcd_txg > txg)
+		tmp_cb = list_prev(&zcl->zcl_callbacks, tmp_cb);
+
+	/* Add the 3 callbacks to the list */
+	for (i = 0; i < 3; i++) {
+		if (tmp_cb == NULL)
+			list_insert_head(&zcl->zcl_callbacks, cb_data[i]);
+		else
+			list_insert_after(&zcl->zcl_callbacks, tmp_cb,
+			    cb_data[i]);
+
+		cb_data[i]->zcd_added = B_TRUE;
+		VERIFY(!cb_data[i]->zcd_called);
+
+		tmp_cb = cb_data[i];
+	}
+
+	(void) mutex_unlock(&zcl->zcl_callbacks_lock);
+
+	dmu_tx_commit(tx);
+}
+
 void
 ztest_dsl_prop_get_set(ztest_args_t *za)
 {
@@ -3041,6 +3253,11 @@ ztest_run(char *pool)
 
 	(void) _mutex_init(&zs->zs_vdev_lock, USYNC_THREAD, NULL);
 	(void) rwlock_init(&zs->zs_name_lock, USYNC_THREAD, NULL);
+	(void) _mutex_init(&zs->zs_cb_list.zcl_callbacks_lock, USYNC_THREAD,
+	    NULL);
+
+	list_create(&zs->zs_cb_list.zcl_callbacks, sizeof (ztest_cb_data_t),
+	    offsetof(ztest_cb_data_t, zcd_node));
 
 	for (t = 0; t < ZTEST_SYNC_LOCKS; t++)
 		(void) _mutex_init(&zs->zs_sync_lock[t], USYNC_THREAD, NULL);
@@ -3240,6 +3457,12 @@ ztest_run(char *pool)
 	spa_close(spa, FTAG);
 
 	kernel_fini();
+
+	list_destroy(&zs->zs_cb_list.zcl_callbacks);
+
+	(void) _mutex_destroy(&zs->zs_cb_list.zcl_callbacks_lock);
+	(void) rwlock_destroy(&zs->zs_name_lock);
+	(void) _mutex_destroy(&zs->zs_vdev_lock);
 }
 
 void
diff --git a/module/zfs/dmu_tx.c b/module/zfs/dmu_tx.c
index bf560e565..0af089198 100644
--- a/module/zfs/dmu_tx.c
+++ b/module/zfs/dmu_tx.c
@@ -48,6 +48,8 @@ dmu_tx_create_dd(dsl_dir_t *dd)
 		tx->tx_pool = dd->dd_pool;
 	list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t),
 	    offsetof(dmu_tx_hold_t, txh_node));
+	list_create(&tx->tx_callbacks, sizeof (dmu_tx_callback_t),
+	    offsetof(dmu_tx_callback_t, dcb_node));
 #ifdef ZFS_DEBUG
 	refcount_create(&tx->tx_space_written);
 	refcount_create(&tx->tx_space_freed);
@@ -1020,8 +1022,13 @@ dmu_tx_commit(dmu_tx_t *tx)
 	if (tx->tx_tempreserve_cookie)
 		dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
 
+	if (!list_is_empty(&tx->tx_callbacks))
+		txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
+
 	if (tx->tx_anyobj == FALSE)
 		txg_rele_to_sync(&tx->tx_txgh);
+
+	list_destroy(&tx->tx_callbacks);
 	list_destroy(&tx->tx_holds);
 #ifdef ZFS_DEBUG
 	dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
@@ -1050,6 +1057,14 @@ dmu_tx_abort(dmu_tx_t *tx)
 		if (dn != NULL)
 			dnode_rele(dn, tx);
 	}
+
+	/*
+	 * Call any registered callbacks with an error code.
+	 */
+	if (!list_is_empty(&tx->tx_callbacks))
+		dmu_tx_callback(&tx->tx_callbacks, ECANCELED);
+
+	list_destroy(&tx->tx_callbacks);
 	list_destroy(&tx->tx_holds);
 #ifdef ZFS_DEBUG
 	refcount_destroy_many(&tx->tx_space_written,
@@ -1066,3 +1081,31 @@ dmu_tx_get_txg(dmu_tx_t *tx)
 	ASSERT(tx->tx_txg != 0);
 	return (tx->tx_txg);
 }
+
+void
+dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
+{
+	dmu_tx_callback_t *dcb;
+
+	dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
+
+	dcb->dcb_func = func;
+	dcb->dcb_data = data;
+
+	list_insert_tail(&tx->tx_callbacks, dcb);
+}
+
+/*
+ * Call all the commit callbacks on a list, with a given error code.
+ */
+void
+dmu_tx_callback(list_t *cb_list, int error)
+{
+	dmu_tx_callback_t *dcb;
+
+	while (dcb = list_head(cb_list)) {
+		list_remove(cb_list, dcb);
+		dcb->dcb_func(dcb->dcb_data, error);
+		kmem_free(dcb, sizeof (dmu_tx_callback_t));
+	}
+}
diff --git a/module/zfs/include/sys/dmu.h b/module/zfs/include/sys/dmu.h
index 3b1e5c8fb..392431a18 100644
--- a/module/zfs/include/sys/dmu.h
+++ b/module/zfs/include/sys/dmu.h
@@ -430,6 +430,26 @@ void dmu_tx_wait(dmu_tx_t *tx);
 void dmu_tx_commit(dmu_tx_t *tx);
 
 /*
+ * To register a commit callback, dmu_tx_callback_register() must be called.
+ *
+ * dcb_data is a pointer to caller private data that is passed on as a
+ * callback parameter. The caller is responsible for properly allocating and
+ * freeing it.
+ *
+ * When registering a callback, the transaction must be already created, but
+ * it cannot be committed or aborted. It can be assigned to a txg or not.
+ *
+ * The callback will be called after the transaction has been safely written
+ * to stable storage and will also be called if the dmu_tx is aborted.
+ * If there is any error which prevents the transaction from being committed to
+ * disk, the callback will be called with a value of error != 0.
+ */
+typedef void dmu_tx_callback_func_t(void *dcb_data, int error);
+
+void dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *dcb_func,
+    void *dcb_data);
+
+/*
  * Free up the data blocks for a defined range of a file.  If size is
  * zero, the range from offset to end-of-file is freed.
  */
diff --git a/module/zfs/include/sys/dmu_tx.h b/module/zfs/include/sys/dmu_tx.h
index 2727daaaa..7fcab936f 100644
--- a/module/zfs/include/sys/dmu_tx.h
+++ b/module/zfs/include/sys/dmu_tx.h
@@ -26,8 +26,6 @@
 #ifndef	_SYS_DMU_TX_H
 #define	_SYS_DMU_TX_H
 
-#pragma ident	"%Z%%M%	%I%	%E% SMI"
-
 #include <sys/inttypes.h>
 #include <sys/dmu.h>
 #include <sys/txg.h>
@@ -59,6 +57,7 @@ struct dmu_tx {
 	txg_handle_t tx_txgh;
 	void *tx_tempreserve_cookie;
 	struct dmu_tx_hold *tx_needassign_txh;
+	list_t tx_callbacks; /* list of dmu_tx_callback_t on this dmu_tx */
 	uint8_t tx_anyobj;
 	int tx_err;
 #ifdef ZFS_DEBUG
@@ -98,6 +97,11 @@ typedef struct dmu_tx_hold {
 #endif
 } dmu_tx_hold_t;
 
+typedef struct dmu_tx_callback {
+	list_node_t		dcb_node;    /* linked to tx_callbacks list */
+	dmu_tx_callback_func_t	*dcb_func;   /* caller function pointer */
+	void			*dcb_data;   /* caller private data */
+} dmu_tx_callback_t;
 
 /*
  * These routines are defined in dmu.h, and are called by the user.
@@ -109,6 +113,10 @@ void dmu_tx_abort(dmu_tx_t *tx);
 uint64_t dmu_tx_get_txg(dmu_tx_t *tx);
 void dmu_tx_wait(dmu_tx_t *tx);
 
+void dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *dcb_func,
+    void *dcb_data);
+void dmu_tx_callback(list_t *cb_list, int error);
+
 /*
  * These routines are defined in dmu_spa.h, and are called by the SPA.
  */
diff --git a/module/zfs/include/sys/txg.h b/module/zfs/include/sys/txg.h
index 23bdff211..e679898db 100644
--- a/module/zfs/include/sys/txg.h
+++ b/module/zfs/include/sys/txg.h
@@ -26,8 +26,6 @@
 #ifndef _SYS_TXG_H
 #define	_SYS_TXG_H
 
-#pragma ident	"%Z%%M%	%I%	%E% SMI"
-
 #include <sys/spa.h>
 #include <sys/zfs_context.h>
 
@@ -71,6 +69,7 @@ extern void txg_sync_stop(struct dsl_pool *dp);
 extern uint64_t txg_hold_open(struct dsl_pool *dp, txg_handle_t *txghp);
 extern void txg_rele_to_quiesce(txg_handle_t *txghp);
 extern void txg_rele_to_sync(txg_handle_t *txghp);
+extern void txg_register_callbacks(txg_handle_t *txghp, list_t *tx_callbacks);
 extern void txg_suspend(struct dsl_pool *dp);
 extern void txg_resume(struct dsl_pool *dp);
 
diff --git a/module/zfs/include/sys/txg_impl.h b/module/zfs/include/sys/txg_impl.h
index 7413c662b..bc7d7c7e5 100644
--- a/module/zfs/include/sys/txg_impl.h
+++ b/module/zfs/include/sys/txg_impl.h
@@ -33,10 +33,16 @@
 extern "C" {
 #endif
 
+typedef struct tx_cb {
+	tx_cpu_t	*tcb_tc;
+	uint64_t	tcb_txg;
+} tx_cb_t;
+
 struct tx_cpu {
 	kmutex_t	tc_lock;
 	kcondvar_t	tc_cv[TXG_SIZE];
 	uint64_t	tc_count[TXG_SIZE];
+	list_t		tc_callbacks[TXG_SIZE]; /* commit cb list */
 	char		tc_pad[16];
 };
 
@@ -64,6 +70,8 @@ typedef struct tx_state {
 
 	kthread_t	*tx_sync_thread;
 	kthread_t	*tx_quiesce_thread;
+
+	taskq_t		*tx_commit_cb_taskq; /* commit callback taskq */
 } tx_state_t;
 
 #ifdef	__cplusplus
diff --git a/module/zfs/txg.c b/module/zfs/txg.c
index e3c0e2a13..b5fcc8c4a 100644
--- a/module/zfs/txg.c
+++ b/module/zfs/txg.c
@@ -26,6 +26,7 @@
 #include <sys/zfs_context.h>
 #include <sys/txg_impl.h>
 #include <sys/dmu_impl.h>
+#include <sys/dmu_tx.h>
 #include <sys/dsl_pool.h>
 #include <sys/callb.h>
 
@@ -57,6 +58,9 @@ txg_init(dsl_pool_t *dp, uint64_t txg)
 		for (i = 0; i < TXG_SIZE; i++) {
 			cv_init(&tx->tx_cpu[c].tc_cv[i], NULL, CV_DEFAULT,
 			    NULL);
+			list_create(&tx->tx_cpu[c].tc_callbacks[i],
+			    sizeof (dmu_tx_callback_t),
+			    offsetof(dmu_tx_callback_t, dcb_node));
 		}
 	}
 
@@ -96,10 +100,15 @@ txg_fini(dsl_pool_t *dp)
 		int i;
 
 		mutex_destroy(&tx->tx_cpu[c].tc_lock);
-		for (i = 0; i < TXG_SIZE; i++)
+		for (i = 0; i < TXG_SIZE; i++) {
 			cv_destroy(&tx->tx_cpu[c].tc_cv[i]);
+			list_destroy(&tx->tx_cpu[c].tc_callbacks[i]);
+		}
 	}
 
+	if (tx->tx_commit_cb_taskq != NULL)
+		taskq_destroy(tx->tx_commit_cb_taskq);
+
 	kmem_free(tx->tx_cpu, max_ncpus * sizeof (tx_cpu_t));
 
 	bzero(tx, sizeof (tx_state_t));
@@ -229,25 +238,55 @@ txg_rele_to_quiesce(txg_handle_t *th)
 }
 
 void
-txg_rele_to_sync(txg_handle_t *th)
+txg_register_callbacks(txg_handle_t *th, list_t *tx_callbacks)
 {
 	tx_cpu_t *tc = th->th_cpu;
 	int g = th->th_txg & TXG_MASK;
 
 	mutex_enter(&tc->tc_lock);
+	list_move_tail(&tc->tc_callbacks[g], tx_callbacks);
+	mutex_exit(&tc->tc_lock);
+}
+
+static void
+txg_exit(tx_cpu_t *tc, uint64_t txg)
+{
+	int g = txg & TXG_MASK;
+
+	mutex_enter(&tc->tc_lock);
 	ASSERT(tc->tc_count[g] != 0);
 	if (--tc->tc_count[g] == 0)
 		cv_broadcast(&tc->tc_cv[g]);
 	mutex_exit(&tc->tc_lock);
+}
+
+void
+txg_rele_to_sync(txg_handle_t *th)
+{
+	txg_exit(th->th_cpu, th->th_txg);
 
 	th->th_cpu = NULL;	/* defensive */
 }
 
 static void
+txg_wait_exit(tx_state_t *tx, uint64_t txg)
+{
+	int g = txg & TXG_MASK;
+	int c;
+
+	for (c = 0; c < max_ncpus; c++) {
+		tx_cpu_t *tc = &tx->tx_cpu[c];
+		mutex_enter(&tc->tc_lock);
+		while (tc->tc_count[g] != 0)
+			cv_wait(&tc->tc_cv[g], &tc->tc_lock);
+		mutex_exit(&tc->tc_lock);
+	}
+}
+
+static void
 txg_quiesce(dsl_pool_t *dp, uint64_t txg)
 {
 	tx_state_t *tx = &dp->dp_tx;
-	int g = txg & TXG_MASK;
 	int c;
 
 	/*
@@ -269,12 +308,60 @@ txg_quiesce(dsl_pool_t *dp, uint64_t txg)
 	/*
 	 * Quiesce the transaction group by waiting for everyone to txg_exit().
 	 */
+	txg_wait_exit(tx, txg);
+}
+
+static void
+txg_callback(tx_cb_t *tcb)
+{
+	tx_cpu_t *tc = tcb->tcb_tc;
+	int g = tcb->tcb_txg & TXG_MASK;
+
+	dmu_tx_callback(&tc->tc_callbacks[g], 0);
+
+	txg_exit(tc, tcb->tcb_txg);
+
+	kmem_free(tcb, sizeof (tx_cb_t));
+}
+
+/*
+ * Dispatch the commit callbacks registered on this txg to worker threads.
+ */
+static void
+txg_dispatch_callbacks(dsl_pool_t *dp, uint64_t txg)
+{
+	int c;
+	tx_state_t *tx = &dp->dp_tx;
+	tx_cb_t *tcb;
+
 	for (c = 0; c < max_ncpus; c++) {
 		tx_cpu_t *tc = &tx->tx_cpu[c];
-		mutex_enter(&tc->tc_lock);
-		while (tc->tc_count[g] != 0)
-			cv_wait(&tc->tc_cv[g], &tc->tc_lock);
-		mutex_exit(&tc->tc_lock);
+		/* No need to lock tx_cpu_t at this point */
+
+		int g = txg & TXG_MASK;
+
+		if (list_is_empty(&tc->tc_callbacks[g]))
+			continue;
+
+		if (tx->tx_commit_cb_taskq == NULL) {
+			/*
+			 * Commit callback taskq hasn't been created yet.
+			 */
+			tx->tx_commit_cb_taskq = taskq_create("tx_commit_cb",
+			    max_ncpus, minclsyspri, max_ncpus, max_ncpus * 4,
+			    TASKQ_PREPOPULATE);
+		}
+
+		tcb = kmem_alloc(sizeof (tx_cb_t), KM_SLEEP);
+		tcb->tcb_txg = txg;
+		tcb->tcb_tc = tc;
+
+		/* There shouldn't be any holders on this txg at this point */
+		ASSERT3U(tc->tc_count[g], ==, 0);
+		tc->tc_count[g]++;
+
+		(void) taskq_dispatch(tx->tx_commit_cb_taskq, (task_func_t *)
+		    txg_callback, tcb, TQ_SLEEP);
 	}
 }
 
@@ -345,6 +432,13 @@ txg_sync_thread(dsl_pool_t *dp)
 		spa_sync(dp->dp_spa, txg);
 		delta = lbolt - start;
 
+		/*
+		 * Dispatch commit callbacks to worker threads and wait for
+		 * them to finish.
+		 */
+		txg_dispatch_callbacks(dp, txg);
+		txg_wait_exit(tx, txg);
+
 		mutex_enter(&tx->tx_sync_lock);
 		rw_enter(&tx->tx_suspend, RW_WRITER);
 		tx->tx_synced_txg = txg;