Add TRIM support

UNMAP/TRIM support is a frequently-requested feature to help
prevent performance from degrading on SSDs and on various other
SAN-like storage back-ends.  By issuing UNMAP/TRIM commands for
sectors which are no longer allocated the underlying device can
often more efficiently manage itself.

This TRIM implementation is modeled on the `zpool initialize`
feature which writes a pattern to all unallocated space in the
pool.  The new `zpool trim` command uses the same vdev_xlate()
code to calculate what sectors are unallocated, the same per-
vdev TRIM thread model and locking, and the same basic CLI for
a consistent user experience.  The core difference is that
instead of writing a pattern it will issue UNMAP/TRIM commands
for those extents.

The zio pipeline was updated to accommodate this by adding a new
ZIO_TYPE_TRIM type and associated spa taskq.  This new type makes
is straight forward to add the platform specific TRIM/UNMAP calls
to vdev_disk.c and vdev_file.c.  These new ZIO_TYPE_TRIM zios are
handled largely the same way as ZIO_TYPE_READs or ZIO_TYPE_WRITEs.
This makes it possible to largely avoid changing the pipieline,
one exception is that TRIM zio's may exceed the 16M block size
limit since they contain no data.

In addition to the manual `zpool trim` command, a background
automatic TRIM was added and is controlled by the 'autotrim'
property.  It relies on the exact same infrastructure as the
manual TRIM.  However, instead of relying on the extents in a
metaslab's ms_allocatable range tree, a ms_trim tree is kept
per metaslab.  When 'autotrim=on', ranges added back to the
ms_allocatable tree are also added to the ms_free tree.  The
ms_free tree is then periodically consumed by an autotrim
thread which systematically walks a top level vdev's metaslabs.

Since the automatic TRIM will skip ranges it considers too small
there is value in occasionally running a full `zpool trim`.  This
may occur when the freed blocks are small and not enough time
was allowed to aggregate them.  An automatic TRIM and a manual
`zpool trim` may be run concurrently, in which case the automatic
TRIM will yield to the manual TRIM.

Reviewed-by: Jorgen Lundman <[email protected]>
Reviewed-by: Tim Chase <[email protected]>
Reviewed-by: Matt Ahrens <[email protected]>
Reviewed-by: George Wilson <[email protected]>
Reviewed-by: Serapheim Dimitropoulos <[email protected]>
Contributions-by: Saso Kiselkov <[email protected]>
Contributions-by: Tim Chase <[email protected]>
Contributions-by: Chunwei Chen <[email protected]>
Signed-off-by: Brian Behlendorf <[email protected]>
Closes #8419 
Closes #598 

1 files changed, 185 insertions, 21 deletions

diff --git a/module/zfs/spa.c b/module/zfs/spa.c
index 71744139e..5392e3547 100644
--- a/module/zfs/spa.c
+++ b/module/zfs/spa.c
@@ -57,6 +57,7 @@
 #include <sys/vdev_indirect_mapping.h>
 #include <sys/vdev_indirect_births.h>
 #include <sys/vdev_initialize.h>
+#include <sys/vdev_trim.h>
 #include <sys/vdev_disk.h>
 #include <sys/metaslab.h>
 #include <sys/metaslab_impl.h>
@@ -132,7 +133,7 @@ static const char *const zio_taskq_types[ZIO_TASKQ_TYPES] = {
  * number of threads assigned to their taskqs using the ZTI_N(#) or ZTI_ONE
  * macros. Other operations process a large amount of data; the ZTI_BATCH
  * macro causes us to create a taskq oriented for throughput. Some operations
- * are so high frequency and short-lived that the taskq itself can become a a
+ * are so high frequency and short-lived that the taskq itself can become a
  * point of lock contention. The ZTI_P(#, #) macro indicates that we need an
  * additional degree of parallelism specified by the number of threads per-
  * taskq and the number of taskqs; when dispatching an event in this case, the
@@ -150,6 +151,7 @@ const zio_taskq_info_t zio_taskqs[ZIO_TYPES][ZIO_TASKQ_TYPES] = {
 	{ ZTI_P(12, 8),	ZTI_NULL,	ZTI_ONE,	ZTI_NULL }, /* FREE */
 	{ ZTI_ONE,	ZTI_NULL,	ZTI_ONE,	ZTI_NULL }, /* CLAIM */
 	{ ZTI_ONE,	ZTI_NULL,	ZTI_ONE,	ZTI_NULL }, /* IOCTL */
+	{ ZTI_N(4),	ZTI_NULL,	ZTI_ONE,	ZTI_NULL }, /* TRIM */
 };
 
 static void spa_sync_version(void *arg, dmu_tx_t *tx);
@@ -554,6 +556,7 @@ spa_prop_validate(spa_t *spa, nvlist_t *props)
 		case ZPOOL_PROP_AUTOREPLACE:
 		case ZPOOL_PROP_LISTSNAPS:
 		case ZPOOL_PROP_AUTOEXPAND:
+		case ZPOOL_PROP_AUTOTRIM:
 			error = nvpair_value_uint64(elem, &intval);
 			if (!error && intval > 1)
 				error = SET_ERROR(EINVAL);
@@ -1442,8 +1445,10 @@ spa_unload(spa_t *spa)
 	spa_async_suspend(spa);
 
 	if (spa->spa_root_vdev) {
-		vdev_initialize_stop_all(spa->spa_root_vdev,
-		    VDEV_INITIALIZE_ACTIVE);
+		vdev_t *root_vdev = spa->spa_root_vdev;
+		vdev_initialize_stop_all(root_vdev, VDEV_INITIALIZE_ACTIVE);
+		vdev_trim_stop_all(root_vdev, VDEV_TRIM_ACTIVE);
+		vdev_autotrim_stop_all(spa);
 	}
 
 	/*
@@ -3585,7 +3590,7 @@ spa_ld_get_props(spa_t *spa)
 		spa_prop_find(spa, ZPOOL_PROP_MULTIHOST, &spa->spa_multihost);
 		spa_prop_find(spa, ZPOOL_PROP_DEDUPDITTO,
 		    &spa->spa_dedup_ditto);
-
+		spa_prop_find(spa, ZPOOL_PROP_AUTOTRIM, &spa->spa_autotrim);
 		spa->spa_autoreplace = (autoreplace != 0);
 	}
 
@@ -4336,6 +4341,8 @@ spa_load_impl(spa_t *spa, spa_import_type_t type, char **ereport)
 
 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
 		vdev_initialize_restart(spa->spa_root_vdev);
+		vdev_trim_restart(spa->spa_root_vdev);
+		vdev_autotrim_restart(spa);
 		spa_config_exit(spa, SCL_CONFIG, FTAG);
 	}
 
@@ -5338,6 +5345,7 @@ spa_create(const char *pool, nvlist_t *nvroot, nvlist_t *props,
 	spa->spa_failmode = zpool_prop_default_numeric(ZPOOL_PROP_FAILUREMODE);
 	spa->spa_autoexpand = zpool_prop_default_numeric(ZPOOL_PROP_AUTOEXPAND);
 	spa->spa_multihost = zpool_prop_default_numeric(ZPOOL_PROP_MULTIHOST);
+	spa->spa_autotrim = zpool_prop_default_numeric(ZPOOL_PROP_AUTOTRIM);
 
 	if (props != NULL) {
 		spa_configfile_set(spa, props, B_FALSE);
@@ -5746,14 +5754,16 @@ spa_export_common(char *pool, int new_state, nvlist_t **oldconfig,
 
 		/*
 		 * We're about to export or destroy this pool. Make sure
-		 * we stop all initializtion activity here before we
-		 * set the spa_final_txg. This will ensure that all
+		 * we stop all initialization and trim activity here before
+		 * we set the spa_final_txg. This will ensure that all
 		 * dirty data resulting from the initialization is
 		 * committed to disk before we unload the pool.
 		 */
 		if (spa->spa_root_vdev != NULL) {
-			vdev_initialize_stop_all(spa->spa_root_vdev,
-			    VDEV_INITIALIZE_ACTIVE);
+			vdev_t *rvd = spa->spa_root_vdev;
+			vdev_initialize_stop_all(rvd, VDEV_INITIALIZE_ACTIVE);
+			vdev_trim_stop_all(rvd, VDEV_TRIM_ACTIVE);
+			vdev_autotrim_stop_all(spa);
 		}
 
 		/*
@@ -6376,7 +6386,6 @@ spa_vdev_detach(spa_t *spa, uint64_t guid, uint64_t pguid, int replace_done)
 		vdev_remove_parent(cvd);
 	}
 
-
 	/*
 	 * We don't set tvd until now because the parent we just removed
 	 * may have been the previous top-level vdev.
@@ -6490,7 +6499,7 @@ spa_vdev_initialize_impl(spa_t *spa, uint64_t guid, uint64_t cmd_type,
 	 * a previous initialization process which has completed but
 	 * the thread is not exited.
 	 */
-	if (cmd_type == POOL_INITIALIZE_DO &&
+	if (cmd_type == POOL_INITIALIZE_START &&
 	    (vd->vdev_initialize_thread != NULL ||
 	    vd->vdev_top->vdev_removing)) {
 		mutex_exit(&vd->vdev_initialize_lock);
@@ -6507,7 +6516,7 @@ spa_vdev_initialize_impl(spa_t *spa, uint64_t guid, uint64_t cmd_type,
 	}
 
 	switch (cmd_type) {
-	case POOL_INITIALIZE_DO:
+	case POOL_INITIALIZE_START:
 		vdev_initialize(vd);
 		break;
 	case POOL_INITIALIZE_CANCEL:
@@ -6571,6 +6580,126 @@ spa_vdev_initialize(spa_t *spa, nvlist_t *nv, uint64_t cmd_type,
 	return (total_errors);
 }
 
+static int
+spa_vdev_trim_impl(spa_t *spa, uint64_t guid, uint64_t cmd_type,
+    uint64_t rate, boolean_t partial, boolean_t secure, list_t *vd_list)
+{
+	ASSERT(MUTEX_HELD(&spa_namespace_lock));
+
+	spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER);
+
+	/* Look up vdev and ensure it's a leaf. */
+	vdev_t *vd = spa_lookup_by_guid(spa, guid, B_FALSE);
+	if (vd == NULL || vd->vdev_detached) {
+		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
+		return (SET_ERROR(ENODEV));
+	} else if (!vd->vdev_ops->vdev_op_leaf || !vdev_is_concrete(vd)) {
+		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
+		return (SET_ERROR(EINVAL));
+	} else if (!vdev_writeable(vd)) {
+		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
+		return (SET_ERROR(EROFS));
+	} else if (!vd->vdev_has_trim) {
+		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
+		return (SET_ERROR(EOPNOTSUPP));
+	} else if (secure && !vd->vdev_has_securetrim) {
+		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
+		return (SET_ERROR(EOPNOTSUPP));
+	}
+	mutex_enter(&vd->vdev_trim_lock);
+	spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
+
+	/*
+	 * When we activate a TRIM action we check to see if the
+	 * vdev_trim_thread is NULL. We do this instead of using the
+	 * vdev_trim_state since there might be a previous TRIM process
+	 * which has completed but the thread is not exited.
+	 */
+	if (cmd_type == POOL_TRIM_START &&
+	    (vd->vdev_trim_thread != NULL || vd->vdev_top->vdev_removing)) {
+		mutex_exit(&vd->vdev_trim_lock);
+		return (SET_ERROR(EBUSY));
+	} else if (cmd_type == POOL_TRIM_CANCEL &&
+	    (vd->vdev_trim_state != VDEV_TRIM_ACTIVE &&
+	    vd->vdev_trim_state != VDEV_TRIM_SUSPENDED)) {
+		mutex_exit(&vd->vdev_trim_lock);
+		return (SET_ERROR(ESRCH));
+	} else if (cmd_type == POOL_TRIM_SUSPEND &&
+	    vd->vdev_trim_state != VDEV_TRIM_ACTIVE) {
+		mutex_exit(&vd->vdev_trim_lock);
+		return (SET_ERROR(ESRCH));
+	}
+
+	switch (cmd_type) {
+	case POOL_TRIM_START:
+		vdev_trim(vd, rate, partial, secure);
+		break;
+	case POOL_TRIM_CANCEL:
+		vdev_trim_stop(vd, VDEV_TRIM_CANCELED, vd_list);
+		break;
+	case POOL_TRIM_SUSPEND:
+		vdev_trim_stop(vd, VDEV_TRIM_SUSPENDED, vd_list);
+		break;
+	default:
+		panic("invalid cmd_type %llu", (unsigned long long)cmd_type);
+	}
+	mutex_exit(&vd->vdev_trim_lock);
+
+	return (0);
+}
+
+/*
+ * Initiates a manual TRIM for the requested vdevs. This kicks off individual
+ * TRIM threads for each child vdev.  These threads pass over all of the free
+ * space in the vdev's metaslabs and issues TRIM commands for that space.
+ */
+int
+spa_vdev_trim(spa_t *spa, nvlist_t *nv, uint64_t cmd_type, uint64_t rate,
+    boolean_t partial, boolean_t secure, nvlist_t *vdev_errlist)
+{
+	int total_errors = 0;
+	list_t vd_list;
+
+	list_create(&vd_list, sizeof (vdev_t),
+	    offsetof(vdev_t, vdev_trim_node));
+
+	/*
+	 * We hold the namespace lock through the whole function
+	 * to prevent any changes to the pool while we're starting or
+	 * stopping TRIM. The config and state locks are held so that
+	 * we can properly assess the vdev state before we commit to
+	 * the TRIM operation.
+	 */
+	mutex_enter(&spa_namespace_lock);
+
+	for (nvpair_t *pair = nvlist_next_nvpair(nv, NULL);
+	    pair != NULL; pair = nvlist_next_nvpair(nv, pair)) {
+		uint64_t vdev_guid = fnvpair_value_uint64(pair);
+
+		int error = spa_vdev_trim_impl(spa, vdev_guid, cmd_type,
+		    rate, partial, secure, &vd_list);
+		if (error != 0) {
+			char guid_as_str[MAXNAMELEN];
+
+			(void) snprintf(guid_as_str, sizeof (guid_as_str),
+			    "%llu", (unsigned long long)vdev_guid);
+			fnvlist_add_int64(vdev_errlist, guid_as_str, error);
+			total_errors++;
+		}
+	}
+
+	/* Wait for all TRIM threads to stop. */
+	vdev_trim_stop_wait(spa, &vd_list);
+
+	/* Sync out the TRIM state */
+	txg_wait_synced(spa->spa_dsl_pool, 0);
+	mutex_exit(&spa_namespace_lock);
+
+	list_destroy(&vd_list);
+
+	return (total_errors);
+}
+
 /*
  * Split a set of devices from their mirrors, and create a new pool from them.
  */
@@ -6780,24 +6909,36 @@ spa_vdev_split_mirror(spa_t *spa, char *newname, nvlist_t *config,
 	spa_async_suspend(newspa);
 
 	/*
-	 * Temporarily stop the initializing activity. We set the state to
-	 * ACTIVE so that we know to resume the initializing once the split
-	 * has completed.
+	 * Temporarily stop the initializing and TRIM activity.  We set the
+	 * state to ACTIVE so that we know to resume initializing or TRIM
+	 * once the split has completed.
 	 */
-	list_t vd_list;
-	list_create(&vd_list, sizeof (vdev_t),
+	list_t vd_initialize_list;
+	list_create(&vd_initialize_list, sizeof (vdev_t),
 	    offsetof(vdev_t, vdev_initialize_node));
 
+	list_t vd_trim_list;
+	list_create(&vd_trim_list, sizeof (vdev_t),
+	    offsetof(vdev_t, vdev_trim_node));
+
 	for (c = 0; c < children; c++) {
 		if (vml[c] != NULL) {
 			mutex_enter(&vml[c]->vdev_initialize_lock);
-			vdev_initialize_stop(vml[c], VDEV_INITIALIZE_ACTIVE,
-			    &vd_list);
+			vdev_initialize_stop(vml[c],
+			    VDEV_INITIALIZE_ACTIVE, &vd_initialize_list);
 			mutex_exit(&vml[c]->vdev_initialize_lock);
+
+			mutex_enter(&vml[c]->vdev_trim_lock);
+			vdev_trim_stop(vml[c], VDEV_TRIM_ACTIVE, &vd_trim_list);
+			mutex_exit(&vml[c]->vdev_trim_lock);
 		}
 	}
-	vdev_initialize_stop_wait(spa, &vd_list);
-	list_destroy(&vd_list);
+
+	vdev_initialize_stop_wait(spa, &vd_initialize_list);
+	vdev_trim_stop_wait(spa, &vd_trim_list);
+
+	list_destroy(&vd_initialize_list);
+	list_destroy(&vd_trim_list);
 
 	newspa->spa_config_source = SPA_CONFIG_SRC_SPLIT;
 
@@ -6899,8 +7040,10 @@ out:
 			vml[c]->vdev_offline = B_FALSE;
 	}
 
-	/* restart initializing disks as necessary */
+	/* restart initializing or trimming disks as necessary */
 	spa_async_request(spa, SPA_ASYNC_INITIALIZE_RESTART);
+	spa_async_request(spa, SPA_ASYNC_TRIM_RESTART);
+	spa_async_request(spa, SPA_ASYNC_AUTOTRIM_RESTART);
 
 	vdev_reopen(spa->spa_root_vdev);
 
@@ -7283,6 +7426,22 @@ spa_async_thread(void *arg)
 		mutex_exit(&spa_namespace_lock);
 	}
 
+	if (tasks & SPA_ASYNC_TRIM_RESTART) {
+		mutex_enter(&spa_namespace_lock);
+		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
+		vdev_trim_restart(spa->spa_root_vdev);
+		spa_config_exit(spa, SCL_CONFIG, FTAG);
+		mutex_exit(&spa_namespace_lock);
+	}
+
+	if (tasks & SPA_ASYNC_AUTOTRIM_RESTART) {
+		mutex_enter(&spa_namespace_lock);
+		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
+		vdev_autotrim_restart(spa);
+		spa_config_exit(spa, SCL_CONFIG, FTAG);
+		mutex_exit(&spa_namespace_lock);
+	}
+
 	/*
 	 * Let the world know that we're done.
 	 */
@@ -7782,6 +7941,11 @@ spa_sync_props(void *arg, dmu_tx_t *tx)
 			case ZPOOL_PROP_FAILUREMODE:
 				spa->spa_failmode = intval;
 				break;
+			case ZPOOL_PROP_AUTOTRIM:
+				spa->spa_autotrim = intval;
+				spa_async_request(spa,
+				    SPA_ASYNC_AUTOTRIM_RESTART);
+				break;
 			case ZPOOL_PROP_AUTOEXPAND:
 				spa->spa_autoexpand = intval;
 				if (tx->tx_txg != TXG_INITIAL)