diff options
author | Brian Behlendorf <[email protected]> | 2019-03-29 09:13:20 -0700 |
---|---|---|
committer | GitHub <[email protected]> | 2019-03-29 09:13:20 -0700 |
commit | 1b939560be5c51deecf875af9dada9d094633bf7 (patch) | |
tree | 2a780b838134636ddbc65f89d227e37c74abe17b /module/zfs | |
parent | f94b3cbf43d62f4962e71cfe7ba8c6f0602e2a45 (diff) |
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
Diffstat (limited to 'module/zfs')
-rw-r--r-- | module/zfs/Makefile.in | 1 | ||||
-rw-r--r-- | module/zfs/dmu.c | 2 | ||||
-rw-r--r-- | module/zfs/metaslab.c | 141 | ||||
-rw-r--r-- | module/zfs/spa.c | 206 | ||||
-rw-r--r-- | module/zfs/spa_misc.c | 21 | ||||
-rw-r--r-- | module/zfs/spa_stats.c | 101 | ||||
-rw-r--r-- | module/zfs/txg.c | 8 | ||||
-rw-r--r-- | module/zfs/vdev.c | 106 | ||||
-rw-r--r-- | module/zfs/vdev_disk.c | 29 | ||||
-rw-r--r-- | module/zfs/vdev_file.c | 35 | ||||
-rw-r--r-- | module/zfs/vdev_initialize.c | 145 | ||||
-rw-r--r-- | module/zfs/vdev_label.c | 18 | ||||
-rw-r--r-- | module/zfs/vdev_queue.c | 67 | ||||
-rw-r--r-- | module/zfs/vdev_raidz.c | 2 | ||||
-rw-r--r-- | module/zfs/vdev_removal.c | 17 | ||||
-rw-r--r-- | module/zfs/vdev_trim.c | 1460 | ||||
-rw-r--r-- | module/zfs/zfs_ioctl.c | 95 | ||||
-rw-r--r-- | module/zfs/zfs_sysfs.c | 3 | ||||
-rw-r--r-- | module/zfs/zio.c | 39 |
19 files changed, 2281 insertions, 215 deletions
diff --git a/module/zfs/Makefile.in b/module/zfs/Makefile.in index 193bdc510..b2460f0d6 100644 --- a/module/zfs/Makefile.in +++ b/module/zfs/Makefile.in @@ -99,6 +99,7 @@ $(MODULE)-objs += vdev_raidz_math.o $(MODULE)-objs += vdev_raidz_math_scalar.o $(MODULE)-objs += vdev_removal.o $(MODULE)-objs += vdev_root.o +$(MODULE)-objs += vdev_trim.o $(MODULE)-objs += zap.o $(MODULE)-objs += zap_leaf.o $(MODULE)-objs += zap_micro.o diff --git a/module/zfs/dmu.c b/module/zfs/dmu.c index 219703231..18328042c 100644 --- a/module/zfs/dmu.c +++ b/module/zfs/dmu.c @@ -842,7 +842,7 @@ dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset, if (dirty_frees_threshold != 0 && long_free_dirty_all_txgs >= dirty_frees_threshold) { DMU_TX_STAT_BUMP(dmu_tx_dirty_frees_delay); - txg_wait_open(dp, 0); + txg_wait_open(dp, 0, B_TRUE); continue; } diff --git a/module/zfs/metaslab.c b/module/zfs/metaslab.c index 8380897a9..06d8383f0 100644 --- a/module/zfs/metaslab.c +++ b/module/zfs/metaslab.c @@ -181,7 +181,6 @@ int metaslab_lba_weighting_enabled = B_TRUE; */ int metaslab_bias_enabled = B_TRUE; - /* * Enable/disable remapping of indirect DVAs to their concrete vdevs. */ @@ -219,6 +218,12 @@ boolean_t metaslab_trace_enabled = B_TRUE; uint64_t metaslab_trace_max_entries = 5000; #endif +/* + * Maximum number of metaslabs per group that can be disabled + * simultaneously. + */ +int max_disabled_ms = 3; + static uint64_t metaslab_weight(metaslab_t *); static void metaslab_set_fragmentation(metaslab_t *); static void metaslab_free_impl(vdev_t *, uint64_t, uint64_t, boolean_t); @@ -652,8 +657,8 @@ metaslab_group_create(metaslab_class_t *mc, vdev_t *vd, int allocators) mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP); mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL); - mutex_init(&mg->mg_ms_initialize_lock, NULL, MUTEX_DEFAULT, NULL); - cv_init(&mg->mg_ms_initialize_cv, NULL, CV_DEFAULT, NULL); + mutex_init(&mg->mg_ms_disabled_lock, NULL, MUTEX_DEFAULT, NULL); + cv_init(&mg->mg_ms_disabled_cv, NULL, CV_DEFAULT, NULL); mg->mg_primaries = kmem_zalloc(allocators * sizeof (metaslab_t *), KM_SLEEP); mg->mg_secondaries = kmem_zalloc(allocators * sizeof (metaslab_t *), @@ -700,8 +705,8 @@ metaslab_group_destroy(metaslab_group_t *mg) kmem_free(mg->mg_secondaries, mg->mg_allocators * sizeof (metaslab_t *)); mutex_destroy(&mg->mg_lock); - mutex_destroy(&mg->mg_ms_initialize_lock); - cv_destroy(&mg->mg_ms_initialize_cv); + mutex_destroy(&mg->mg_ms_disabled_lock); + cv_destroy(&mg->mg_ms_disabled_cv); for (int i = 0; i < mg->mg_allocators; i++) { zfs_refcount_destroy(&mg->mg_alloc_queue_depth[i]); @@ -1846,8 +1851,10 @@ metaslab_init(metaslab_group_t *mg, uint64_t id, uint64_t object, uint64_t txg, */ ms->ms_allocatable = range_tree_create_impl(&rt_avl_ops, &ms->ms_allocatable_by_size, metaslab_rangesize_compare, 0); - metaslab_group_add(mg, ms); + ms->ms_trim = range_tree_create(NULL, NULL); + + metaslab_group_add(mg, ms); metaslab_set_fragmentation(ms); /* @@ -1921,6 +1928,9 @@ metaslab_fini(metaslab_t *msp) for (int t = 0; t < TXG_SIZE; t++) ASSERT(!txg_list_member(&vd->vdev_ms_list, msp, t)); + range_tree_vacate(msp->ms_trim, NULL, NULL); + range_tree_destroy(msp->ms_trim); + mutex_exit(&msp->ms_lock); cv_destroy(&msp->ms_load_cv); mutex_destroy(&msp->ms_lock); @@ -2727,6 +2737,7 @@ metaslab_sync(metaslab_t *msp, uint64_t txg) ASSERT3P(msp->ms_freeing, !=, NULL); ASSERT3P(msp->ms_freed, !=, NULL); ASSERT3P(msp->ms_checkpointing, !=, NULL); + ASSERT3P(msp->ms_trim, !=, NULL); /* * Normally, we don't want to process a metaslab if there are no @@ -3000,6 +3011,24 @@ metaslab_sync_done(metaslab_t *msp, uint64_t txg) metaslab_load_wait(msp); /* + * When auto-trimming is enabled, free ranges which are added to + * ms_allocatable are also be added to ms_trim. The ms_trim tree is + * periodically consumed by the vdev_autotrim_thread() which issues + * trims for all ranges and then vacates the tree. The ms_trim tree + * can be discarded at any time with the sole consequence of recent + * frees not being trimmed. + */ + if (spa_get_autotrim(spa) == SPA_AUTOTRIM_ON) { + range_tree_walk(*defer_tree, range_tree_add, msp->ms_trim); + if (!defer_allowed) { + range_tree_walk(msp->ms_freed, range_tree_add, + msp->ms_trim); + } + } else { + range_tree_vacate(msp->ms_trim, NULL, NULL); + } + + /* * Move the frees from the defer_tree back to the free * range tree (if it's loaded). Swap the freed_tree and * the defer_tree -- this is safe to do because we've @@ -3047,7 +3076,7 @@ metaslab_sync_done(metaslab_t *msp, uint64_t txg) * from it in 'metaslab_unload_delay' txgs, then unload it. */ if (msp->ms_loaded && - msp->ms_initializing == 0 && + msp->ms_disabled == 0 && msp->ms_selected_txg + metaslab_unload_delay < txg) { for (int t = 1; t < TXG_CONCURRENT_STATES; t++) { @@ -3330,7 +3359,7 @@ metaslab_block_alloc(metaslab_t *msp, uint64_t size, uint64_t txg) metaslab_class_t *mc = msp->ms_group->mg_class; VERIFY(!msp->ms_condensing); - VERIFY0(msp->ms_initializing); + VERIFY0(msp->ms_disabled); start = mc->mc_ops->msop_alloc(msp, size); if (start != -1ULL) { @@ -3341,6 +3370,7 @@ metaslab_block_alloc(metaslab_t *msp, uint64_t size, uint64_t txg) VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift)); VERIFY3U(range_tree_space(rt) - size, <=, msp->ms_size); range_tree_remove(rt, start, size); + range_tree_clear(msp->ms_trim, start, size); if (range_tree_is_empty(msp->ms_allocating[txg & TXG_MASK])) vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg); @@ -3391,10 +3421,10 @@ find_valid_metaslab(metaslab_group_t *mg, uint64_t activation_weight, } /* - * If the selected metaslab is condensing or being - * initialized, skip it. + * If the selected metaslab is condensing or disabled, + * skip it. */ - if (msp->ms_condensing || msp->ms_initializing > 0) + if (msp->ms_condensing || msp->ms_disabled > 0) continue; *was_active = msp->ms_allocator != -1; @@ -3566,9 +3596,9 @@ metaslab_group_alloc_normal(metaslab_group_t *mg, zio_alloc_list_t *zal, ~METASLAB_ACTIVE_MASK); mutex_exit(&msp->ms_lock); continue; - } else if (msp->ms_initializing > 0) { + } else if (msp->ms_disabled > 0) { metaslab_trace_add(zal, mg, msp, asize, d, - TRACE_INITIALIZING, allocator); + TRACE_DISABLED, allocator); metaslab_passivate(msp, msp->ms_weight & ~METASLAB_ACTIVE_MASK); mutex_exit(&msp->ms_lock); @@ -4294,6 +4324,7 @@ metaslab_claim_concrete(vdev_t *vd, uint64_t offset, uint64_t size, VERIFY3U(range_tree_space(msp->ms_allocatable) - size, <=, msp->ms_size); range_tree_remove(msp->ms_allocatable, offset, size); + range_tree_clear(msp->ms_trim, offset, size); if (spa_writeable(spa)) { /* don't dirty if we're zdb(1M) */ if (range_tree_is_empty(msp->ms_allocating[txg & TXG_MASK])) @@ -4606,6 +4637,7 @@ metaslab_check_free_impl(vdev_t *vd, uint64_t offset, uint64_t size) offset, size); } + range_tree_verify_not_present(msp->ms_trim, offset, size); range_tree_verify_not_present(msp->ms_freeing, offset, size); range_tree_verify_not_present(msp->ms_checkpointing, offset, size); range_tree_verify_not_present(msp->ms_freed, offset, size); @@ -4637,6 +4669,89 @@ metaslab_check_free(spa_t *spa, const blkptr_t *bp) spa_config_exit(spa, SCL_VDEV, FTAG); } +static void +metaslab_group_disable_wait(metaslab_group_t *mg) +{ + ASSERT(MUTEX_HELD(&mg->mg_ms_disabled_lock)); + while (mg->mg_disabled_updating) { + cv_wait(&mg->mg_ms_disabled_cv, &mg->mg_ms_disabled_lock); + } +} + +static void +metaslab_group_disabled_increment(metaslab_group_t *mg) +{ + ASSERT(MUTEX_HELD(&mg->mg_ms_disabled_lock)); + ASSERT(mg->mg_disabled_updating); + + while (mg->mg_ms_disabled >= max_disabled_ms) { + cv_wait(&mg->mg_ms_disabled_cv, &mg->mg_ms_disabled_lock); + } + mg->mg_ms_disabled++; + ASSERT3U(mg->mg_ms_disabled, <=, max_disabled_ms); +} + +/* + * Mark the metaslab as disabled to prevent any allocations on this metaslab. + * We must also track how many metaslabs are currently disabled within a + * metaslab group and limit them to prevent allocation failures from + * occurring because all metaslabs are disabled. + */ +void +metaslab_disable(metaslab_t *msp) +{ + ASSERT(!MUTEX_HELD(&msp->ms_lock)); + metaslab_group_t *mg = msp->ms_group; + + mutex_enter(&mg->mg_ms_disabled_lock); + + /* + * To keep an accurate count of how many threads have disabled + * a specific metaslab group, we only allow one thread to mark + * the metaslab group at a time. This ensures that the value of + * ms_disabled will be accurate when we decide to mark a metaslab + * group as disabled. To do this we force all other threads + * to wait till the metaslab's mg_disabled_updating flag is no + * longer set. + */ + metaslab_group_disable_wait(mg); + mg->mg_disabled_updating = B_TRUE; + if (msp->ms_disabled == 0) { + metaslab_group_disabled_increment(mg); + } + mutex_enter(&msp->ms_lock); + msp->ms_disabled++; + mutex_exit(&msp->ms_lock); + + mg->mg_disabled_updating = B_FALSE; + cv_broadcast(&mg->mg_ms_disabled_cv); + mutex_exit(&mg->mg_ms_disabled_lock); +} + +void +metaslab_enable(metaslab_t *msp, boolean_t sync) +{ + metaslab_group_t *mg = msp->ms_group; + spa_t *spa = mg->mg_vd->vdev_spa; + + /* + * Wait for the outstanding IO to be synced to prevent newly + * allocated blocks from being overwritten. This used by + * initialize and TRIM which are modifying unallocated space. + */ + if (sync) + txg_wait_synced(spa_get_dsl(spa), 0); + + mutex_enter(&mg->mg_ms_disabled_lock); + mutex_enter(&msp->ms_lock); + if (--msp->ms_disabled == 0) { + mg->mg_ms_disabled--; + cv_broadcast(&mg->mg_ms_disabled_cv); + } + mutex_exit(&msp->ms_lock); + mutex_exit(&mg->mg_ms_disabled_lock); +} + #if defined(_KERNEL) /* BEGIN CSTYLED */ module_param(metaslab_aliquot, ulong, 0644); 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) diff --git a/module/zfs/spa_misc.c b/module/zfs/spa_misc.c index 71221b21b..b3a4a7b12 100644 --- a/module/zfs/spa_misc.c +++ b/module/zfs/spa_misc.c @@ -39,6 +39,7 @@ #include <sys/zil.h> #include <sys/vdev_impl.h> #include <sys/vdev_initialize.h> +#include <sys/vdev_trim.h> #include <sys/vdev_file.h> #include <sys/vdev_raidz.h> #include <sys/metaslab.h> @@ -1128,6 +1129,9 @@ spa_vdev_enter(spa_t *spa) { mutex_enter(&spa->spa_vdev_top_lock); mutex_enter(&spa_namespace_lock); + + vdev_autotrim_stop_all(spa); + return (spa_vdev_config_enter(spa)); } @@ -1204,8 +1208,17 @@ spa_vdev_config_exit(spa_t *spa, vdev_t *vd, uint64_t txg, int error, char *tag) vdev_initialize_stop(vd, VDEV_INITIALIZE_CANCELED, NULL); mutex_exit(&vd->vdev_initialize_lock); + + mutex_enter(&vd->vdev_trim_lock); + vdev_trim_stop(vd, VDEV_TRIM_CANCELED, NULL); + mutex_exit(&vd->vdev_trim_lock); } + /* + * The vdev may be both a leaf and top-level device. + */ + vdev_autotrim_stop_wait(vd); + spa_config_enter(spa, SCL_ALL, spa, RW_WRITER); vdev_free(vd); spa_config_exit(spa, SCL_ALL, spa); @@ -1227,6 +1240,8 @@ spa_vdev_config_exit(spa_t *spa, vdev_t *vd, uint64_t txg, int error, char *tag) int spa_vdev_exit(spa_t *spa, vdev_t *vd, uint64_t txg, int error) { + vdev_autotrim_restart(spa); + spa_vdev_config_exit(spa, vd, txg, error, FTAG); mutex_exit(&spa_namespace_lock); mutex_exit(&spa->spa_vdev_top_lock); @@ -1923,6 +1938,12 @@ spa_deadman_synctime(spa_t *spa) return (spa->spa_deadman_synctime); } +spa_autotrim_t +spa_get_autotrim(spa_t *spa) +{ + return (spa->spa_autotrim); +} + uint64_t spa_deadman_ziotime(spa_t *spa) { diff --git a/module/zfs/spa_stats.c b/module/zfs/spa_stats.c index e01d2d198..3b51250c6 100644 --- a/module/zfs/spa_stats.c +++ b/module/zfs/spa_stats.c @@ -887,6 +887,105 @@ spa_health_destroy(spa_t *spa) mutex_destroy(&shk->lock); } +static spa_iostats_t spa_iostats_template = { + { "trim_extents_written", KSTAT_DATA_UINT64 }, + { "trim_bytes_written", KSTAT_DATA_UINT64 }, + { "trim_extents_skipped", KSTAT_DATA_UINT64 }, + { "trim_bytes_skipped", KSTAT_DATA_UINT64 }, + { "trim_extents_failed", KSTAT_DATA_UINT64 }, + { "trim_bytes_failed", KSTAT_DATA_UINT64 }, + { "autotrim_extents_written", KSTAT_DATA_UINT64 }, + { "autotrim_bytes_written", KSTAT_DATA_UINT64 }, + { "autotrim_extents_skipped", KSTAT_DATA_UINT64 }, + { "autotrim_bytes_skipped", KSTAT_DATA_UINT64 }, + { "autotrim_extents_failed", KSTAT_DATA_UINT64 }, + { "autotrim_bytes_failed", KSTAT_DATA_UINT64 }, +}; + +#define SPA_IOSTATS_ADD(stat, val) \ + atomic_add_64(&iostats->stat.value.ui64, (val)); + +void +spa_iostats_trim_add(spa_t *spa, trim_type_t type, + uint64_t extents_written, uint64_t bytes_written, + uint64_t extents_skipped, uint64_t bytes_skipped, + uint64_t extents_failed, uint64_t bytes_failed) +{ + spa_history_kstat_t *shk = &spa->spa_stats.iostats; + kstat_t *ksp = shk->kstat; + spa_iostats_t *iostats; + + if (ksp == NULL) + return; + + iostats = ksp->ks_data; + if (type == TRIM_TYPE_MANUAL) { + SPA_IOSTATS_ADD(trim_extents_written, extents_written); + SPA_IOSTATS_ADD(trim_bytes_written, bytes_written); + SPA_IOSTATS_ADD(trim_extents_skipped, extents_skipped); + SPA_IOSTATS_ADD(trim_bytes_skipped, bytes_skipped); + SPA_IOSTATS_ADD(trim_extents_failed, extents_failed); + SPA_IOSTATS_ADD(trim_bytes_failed, bytes_failed); + } else { + SPA_IOSTATS_ADD(autotrim_extents_written, extents_written); + SPA_IOSTATS_ADD(autotrim_bytes_written, bytes_written); + SPA_IOSTATS_ADD(autotrim_extents_skipped, extents_skipped); + SPA_IOSTATS_ADD(autotrim_bytes_skipped, bytes_skipped); + SPA_IOSTATS_ADD(autotrim_extents_failed, extents_failed); + SPA_IOSTATS_ADD(autotrim_bytes_failed, bytes_failed); + } +} + +int +spa_iostats_update(kstat_t *ksp, int rw) +{ + if (rw == KSTAT_WRITE) { + memcpy(ksp->ks_data, &spa_iostats_template, + sizeof (spa_iostats_t)); + } + + return (0); +} + +static void +spa_iostats_init(spa_t *spa) +{ + spa_history_kstat_t *shk = &spa->spa_stats.iostats; + + mutex_init(&shk->lock, NULL, MUTEX_DEFAULT, NULL); + + char *name = kmem_asprintf("zfs/%s", spa_name(spa)); + kstat_t *ksp = kstat_create(name, 0, "iostats", "misc", + KSTAT_TYPE_NAMED, sizeof (spa_iostats_t) / sizeof (kstat_named_t), + KSTAT_FLAG_VIRTUAL); + + shk->kstat = ksp; + if (ksp) { + int size = sizeof (spa_iostats_t); + ksp->ks_lock = &shk->lock; + ksp->ks_private = spa; + ksp->ks_update = spa_iostats_update; + ksp->ks_data = kmem_alloc(size, KM_SLEEP); + memcpy(ksp->ks_data, &spa_iostats_template, size); + kstat_install(ksp); + } + + strfree(name); +} + +static void +spa_iostats_destroy(spa_t *spa) +{ + spa_history_kstat_t *shk = &spa->spa_stats.iostats; + kstat_t *ksp = shk->kstat; + if (ksp) { + kmem_free(ksp->ks_data, sizeof (spa_iostats_t)); + kstat_delete(ksp); + } + + mutex_destroy(&shk->lock); +} + void spa_stats_init(spa_t *spa) { @@ -896,11 +995,13 @@ spa_stats_init(spa_t *spa) spa_io_history_init(spa); spa_mmp_history_init(spa); spa_state_init(spa); + spa_iostats_init(spa); } void spa_stats_destroy(spa_t *spa) { + spa_iostats_destroy(spa); spa_health_destroy(spa); spa_tx_assign_destroy(spa); spa_txg_history_destroy(spa); diff --git a/module/zfs/txg.c b/module/zfs/txg.c index db0f60cd1..b3f895302 100644 --- a/module/zfs/txg.c +++ b/module/zfs/txg.c @@ -694,8 +694,12 @@ txg_wait_synced(dsl_pool_t *dp, uint64_t txg) mutex_exit(&tx->tx_sync_lock); } +/* + * Wait for the specified open transaction group. Set should_quiesce + * when the current open txg should be quiesced immediately. + */ void -txg_wait_open(dsl_pool_t *dp, uint64_t txg) +txg_wait_open(dsl_pool_t *dp, uint64_t txg, boolean_t should_quiesce) { tx_state_t *tx = &dp->dp_tx; @@ -705,7 +709,7 @@ txg_wait_open(dsl_pool_t *dp, uint64_t txg) ASSERT3U(tx->tx_threads, ==, 2); if (txg == 0) txg = tx->tx_open_txg + 1; - if (tx->tx_quiesce_txg_waiting < txg) + if (tx->tx_quiesce_txg_waiting < txg && should_quiesce) tx->tx_quiesce_txg_waiting = txg; dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n", txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting); diff --git a/module/zfs/vdev.c b/module/zfs/vdev.c index ae1c2bcec..085ae6873 100644 --- a/module/zfs/vdev.c +++ b/module/zfs/vdev.c @@ -51,6 +51,7 @@ #include <sys/dsl_scan.h> #include <sys/abd.h> #include <sys/vdev_initialize.h> +#include <sys/vdev_trim.h> #include <sys/zvol.h> #include <sys/zfs_ratelimit.h> @@ -543,6 +544,7 @@ vdev_alloc_common(spa_t *spa, uint_t id, uint64_t guid, vdev_ops_t *ops) list_link_init(&vd->vdev_state_dirty_node); list_link_init(&vd->vdev_initialize_node); list_link_init(&vd->vdev_leaf_node); + list_link_init(&vd->vdev_trim_node); mutex_init(&vd->vdev_dtl_lock, NULL, MUTEX_NOLOCKDEP, NULL); mutex_init(&vd->vdev_stat_lock, NULL, MUTEX_DEFAULT, NULL); mutex_init(&vd->vdev_probe_lock, NULL, MUTEX_DEFAULT, NULL); @@ -551,6 +553,12 @@ vdev_alloc_common(spa_t *spa, uint_t id, uint64_t guid, vdev_ops_t *ops) mutex_init(&vd->vdev_initialize_io_lock, NULL, MUTEX_DEFAULT, NULL); cv_init(&vd->vdev_initialize_cv, NULL, CV_DEFAULT, NULL); cv_init(&vd->vdev_initialize_io_cv, NULL, CV_DEFAULT, NULL); + mutex_init(&vd->vdev_trim_lock, NULL, MUTEX_DEFAULT, NULL); + mutex_init(&vd->vdev_autotrim_lock, NULL, MUTEX_DEFAULT, NULL); + mutex_init(&vd->vdev_trim_io_lock, NULL, MUTEX_DEFAULT, NULL); + cv_init(&vd->vdev_trim_cv, NULL, CV_DEFAULT, NULL); + cv_init(&vd->vdev_autotrim_cv, NULL, CV_DEFAULT, NULL); + cv_init(&vd->vdev_trim_io_cv, NULL, CV_DEFAULT, NULL); for (int t = 0; t < DTL_TYPES; t++) { vd->vdev_dtl[t] = range_tree_create(NULL, NULL); @@ -875,7 +883,10 @@ void vdev_free(vdev_t *vd) { spa_t *spa = vd->vdev_spa; + ASSERT3P(vd->vdev_initialize_thread, ==, NULL); + ASSERT3P(vd->vdev_trim_thread, ==, NULL); + ASSERT3P(vd->vdev_autotrim_thread, ==, NULL); /* * Scan queues are normally destroyed at the end of a scan. If the @@ -906,7 +917,6 @@ vdev_free(vdev_t *vd) ASSERT(vd->vdev_child == NULL); ASSERT(vd->vdev_guid_sum == vd->vdev_guid); - ASSERT(vd->vdev_initialize_thread == NULL); /* * Discard allocation state. @@ -988,6 +998,12 @@ vdev_free(vdev_t *vd) mutex_destroy(&vd->vdev_initialize_io_lock); cv_destroy(&vd->vdev_initialize_io_cv); cv_destroy(&vd->vdev_initialize_cv); + mutex_destroy(&vd->vdev_trim_lock); + mutex_destroy(&vd->vdev_autotrim_lock); + mutex_destroy(&vd->vdev_trim_io_lock); + cv_destroy(&vd->vdev_trim_cv); + cv_destroy(&vd->vdev_autotrim_cv); + cv_destroy(&vd->vdev_trim_io_cv); zfs_ratelimit_fini(&vd->vdev_delay_rl); zfs_ratelimit_fini(&vd->vdev_checksum_rl); @@ -3475,6 +3491,16 @@ vdev_online(spa_t *spa, uint64_t guid, uint64_t flags, vdev_state_t *newstate) } mutex_exit(&vd->vdev_initialize_lock); + /* Restart trimming if necessary */ + mutex_enter(&vd->vdev_trim_lock); + if (vdev_writeable(vd) && + vd->vdev_trim_thread == NULL && + vd->vdev_trim_state == VDEV_TRIM_ACTIVE) { + (void) vdev_trim(vd, vd->vdev_trim_rate, vd->vdev_trim_partial, + vd->vdev_trim_secure); + } + mutex_exit(&vd->vdev_trim_lock); + if (wasoffline || (oldstate < VDEV_STATE_DEGRADED && vd->vdev_state >= VDEV_STATE_DEGRADED)) @@ -3745,8 +3771,7 @@ vdev_accessible(vdev_t *vd, zio_t *zio) static void vdev_get_child_stat(vdev_t *cvd, vdev_stat_t *vs, vdev_stat_t *cvs) { - int t; - for (t = 0; t < ZIO_TYPES; t++) { + for (int t = 0; t < VS_ZIO_TYPES; t++) { vs->vs_ops[t] += cvs->vs_ops[t]; vs->vs_bytes[t] += cvs->vs_bytes[t]; } @@ -3873,7 +3898,7 @@ vdev_get_stats_ex(vdev_t *vd, vdev_stat_t *vs, vdev_stat_ex_t *vsx) vs->vs_rsize += VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE; /* - * Report intializing progress. Since we don't + * Report initializing progress. Since we don't * have the initializing locks held, this is only * an estimate (although a fairly accurate one). */ @@ -3884,9 +3909,20 @@ vdev_get_stats_ex(vdev_t *vd, vdev_stat_t *vs, vdev_stat_ex_t *vsx) vs->vs_initialize_state = vd->vdev_initialize_state; vs->vs_initialize_action_time = vd->vdev_initialize_action_time; + + /* + * Report manual TRIM progress. Since we don't have + * the manual TRIM locks held, this is only an + * estimate (although fairly accurate one). + */ + vs->vs_trim_notsup = !vd->vdev_has_trim; + vs->vs_trim_bytes_done = vd->vdev_trim_bytes_done; + vs->vs_trim_bytes_est = vd->vdev_trim_bytes_est; + vs->vs_trim_state = vd->vdev_trim_state; + vs->vs_trim_action_time = vd->vdev_trim_action_time; } /* - * Report expandable space on top-level, non-auxillary devices + * Report expandable space on top-level, non-auxiliary devices * only. The expandable space is reported in terms of metaslab * sized units since that determines how much space the pool * can expand. @@ -4004,9 +4040,18 @@ vdev_stat_update(zio_t *zio, uint64_t psize) */ if (vd->vdev_ops->vdev_op_leaf && (zio->io_priority < ZIO_PRIORITY_NUM_QUEUEABLE)) { + zio_type_t vs_type = type; + + /* + * TRIM ops and bytes are reported to user space as + * ZIO_TYPE_IOCTL. This is done to preserve the + * vdev_stat_t structure layout for user space. + */ + if (type == ZIO_TYPE_TRIM) + vs_type = ZIO_TYPE_IOCTL; - vs->vs_ops[type]++; - vs->vs_bytes[type] += psize; + vs->vs_ops[vs_type]++; + vs->vs_bytes[vs_type] += psize; if (flags & ZIO_FLAG_DELEGATED) { vsx->vsx_agg_histo[zio->io_priority] @@ -4104,7 +4149,8 @@ vdev_deflated_space(vdev_t *vd, int64_t space) } /* - * Update the in-core space usage stats for this vdev and the root vdev. + * Update the in-core space usage stats for this vdev, its metaslab class, + * and the root vdev. */ void vdev_space_update(vdev_t *vd, int64_t alloc_delta, int64_t defer_delta, @@ -4650,12 +4696,56 @@ vdev_set_deferred_resilver(spa_t *spa, vdev_t *vd) spa->spa_resilver_deferred = B_TRUE; } +/* + * Translate a logical range to the physical range for the specified vdev_t. + * This function is initially called with a leaf vdev and will walk each + * parent vdev until it reaches a top-level vdev. Once the top-level is + * reached the physical range is initialized and the recursive function + * begins to unwind. As it unwinds it calls the parent's vdev specific + * translation function to do the real conversion. + */ +void +vdev_xlate(vdev_t *vd, const range_seg_t *logical_rs, range_seg_t *physical_rs) +{ + /* + * Walk up the vdev tree + */ + if (vd != vd->vdev_top) { + vdev_xlate(vd->vdev_parent, logical_rs, physical_rs); + } else { + /* + * We've reached the top-level vdev, initialize the + * physical range to the logical range and start to + * unwind. + */ + physical_rs->rs_start = logical_rs->rs_start; + physical_rs->rs_end = logical_rs->rs_end; + return; + } + + vdev_t *pvd = vd->vdev_parent; + ASSERT3P(pvd, !=, NULL); + ASSERT3P(pvd->vdev_ops->vdev_op_xlate, !=, NULL); + + /* + * As this recursive function unwinds, translate the logical + * range into its physical components by calling the + * vdev specific translate function. + */ + range_seg_t intermediate = { { { 0, 0 } } }; + pvd->vdev_ops->vdev_op_xlate(vd, physical_rs, &intermediate); + + physical_rs->rs_start = intermediate.rs_start; + physical_rs->rs_end = intermediate.rs_end; +} + #if defined(_KERNEL) EXPORT_SYMBOL(vdev_fault); EXPORT_SYMBOL(vdev_degrade); EXPORT_SYMBOL(vdev_online); EXPORT_SYMBOL(vdev_offline); EXPORT_SYMBOL(vdev_clear); + /* BEGIN CSTYLED */ module_param(zfs_vdev_default_ms_count, int, 0644); MODULE_PARM_DESC(zfs_vdev_default_ms_count, diff --git a/module/zfs/vdev_disk.c b/module/zfs/vdev_disk.c index 4ac08c861..c2312e6fa 100644 --- a/module/zfs/vdev_disk.c +++ b/module/zfs/vdev_disk.c @@ -30,6 +30,7 @@ #include <sys/spa_impl.h> #include <sys/vdev_disk.h> #include <sys/vdev_impl.h> +#include <sys/vdev_trim.h> #include <sys/abd.h> #include <sys/fs/zfs.h> #include <sys/zio.h> @@ -223,7 +224,7 @@ vdev_elevator_switch(vdev_t *v, char *elevator) strfree(argv[2]); #endif /* HAVE_ELEVATOR_CHANGE */ if (error) { - zfs_dbgmsg("Unable to set \"%s\" scheduler for %s (%s): %d\n", + zfs_dbgmsg("Unable to set \"%s\" scheduler for %s (%s): %d", elevator, v->vdev_path, device, error); } } @@ -322,7 +323,7 @@ vdev_disk_open(vdev_t *v, uint64_t *psize, uint64_t *max_psize, if (IS_ERR(bdev)) { int error = -PTR_ERR(bdev); - vdev_dbgmsg(v, "open error=%d count=%d\n", error, count); + vdev_dbgmsg(v, "open error=%d count=%d", error, count); vd->vd_bdev = NULL; v->vdev_tsd = vd; rw_exit(&vd->vd_lock); @@ -333,14 +334,22 @@ vdev_disk_open(vdev_t *v, uint64_t *psize, uint64_t *max_psize, rw_exit(&vd->vd_lock); } + struct request_queue *q = bdev_get_queue(vd->vd_bdev); + /* Determine the physical block size */ block_size = vdev_bdev_block_size(vd->vd_bdev); /* Clear the nowritecache bit, causes vdev_reopen() to try again. */ v->vdev_nowritecache = B_FALSE; + /* Set when device reports it supports TRIM. */ + v->vdev_has_trim = !!blk_queue_discard(q); + + /* Set when device reports it supports secure TRIM. */ + v->vdev_has_securetrim = !!blk_queue_discard_secure(q); + /* Inform the ZIO pipeline that we are non-rotational */ - v->vdev_nonrot = blk_queue_nonrot(bdev_get_queue(vd->vd_bdev)); + v->vdev_nonrot = blk_queue_nonrot(q); /* Physical volume size in bytes for the partition */ *psize = bdev_capacity(vd->vd_bdev); @@ -728,6 +737,7 @@ vdev_disk_io_start(zio_t *zio) { vdev_t *v = zio->io_vd; vdev_disk_t *vd = v->vdev_tsd; + unsigned long trim_flags = 0; int rw, flags, error; /* @@ -813,6 +823,19 @@ vdev_disk_io_start(zio_t *zio) #endif break; + case ZIO_TYPE_TRIM: +#if defined(BLKDEV_DISCARD_SECURE) + if (zio->io_trim_flags & ZIO_TRIM_SECURE) + trim_flags |= BLKDEV_DISCARD_SECURE; +#endif + zio->io_error = -blkdev_issue_discard(vd->vd_bdev, + zio->io_offset >> 9, zio->io_size >> 9, GFP_NOFS, + trim_flags); + + rw_exit(&vd->vd_lock); + zio_interrupt(zio); + return; + default: rw_exit(&vd->vd_lock); zio->io_error = SET_ERROR(ENOTSUP); diff --git a/module/zfs/vdev_file.c b/module/zfs/vdev_file.c index 3551898e0..c04f40ca4 100644 --- a/module/zfs/vdev_file.c +++ b/module/zfs/vdev_file.c @@ -28,10 +28,13 @@ #include <sys/spa_impl.h> #include <sys/vdev_file.h> #include <sys/vdev_impl.h> +#include <sys/vdev_trim.h> #include <sys/zio.h> #include <sys/fs/zfs.h> #include <sys/fm/fs/zfs.h> #include <sys/abd.h> +#include <sys/fcntl.h> +#include <sys/vnode.h> /* * Virtual device vector for files. @@ -60,10 +63,25 @@ vdev_file_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize, vattr_t vattr; int error; - /* Rotational optimizations only make sense on block devices */ + /* + * Rotational optimizations only make sense on block devices. + */ vd->vdev_nonrot = B_TRUE; /* + * Allow TRIM on file based vdevs. This may not always be supported, + * since it depends on your kernel version and underlying filesystem + * type but it is always safe to attempt. + */ + vd->vdev_has_trim = B_TRUE; + + /* + * Disable secure TRIM on file based vdevs. There is no way to + * request this behavior from the underlying filesystem. + */ + vd->vdev_has_securetrim = B_FALSE; + + /* * We must have a pathname, and it must be absolute. */ if (vd->vdev_path == NULL || vd->vdev_path[0] != '/') { @@ -229,6 +247,21 @@ vdev_file_io_start(zio_t *zio) zio_execute(zio); return; + } else if (zio->io_type == ZIO_TYPE_TRIM) { + struct flock flck; + + ASSERT3U(zio->io_size, !=, 0); + bzero(&flck, sizeof (flck)); + flck.l_type = F_FREESP; + flck.l_start = zio->io_offset; + flck.l_len = zio->io_size; + flck.l_whence = 0; + + zio->io_error = VOP_SPACE(vf->vf_vnode, F_FREESP, &flck, + 0, 0, kcred, NULL); + + zio_execute(zio); + return; } zio->io_target_timestamp = zio_handle_io_delay(zio); diff --git a/module/zfs/vdev_initialize.c b/module/zfs/vdev_initialize.c index bca2db7a4..b15901326 100644 --- a/module/zfs/vdev_initialize.c +++ b/module/zfs/vdev_initialize.c @@ -34,12 +34,6 @@ #include <sys/dmu_tx.h> /* - * Maximum number of metaslabs per group that can be initialized - * simultaneously. - */ -int max_initialize_ms = 3; - -/* * Value that is written to disk during initialization. */ #ifdef _ILP32 @@ -132,7 +126,7 @@ vdev_initialize_change_state(vdev_t *vd, vdev_initializing_state_t new_state) dmu_tx_t *tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir); VERIFY0(dmu_tx_assign(tx, TXG_WAIT)); dsl_sync_task_nowait(spa_get_dsl(spa), vdev_initialize_zap_update_sync, - guid, 2, ZFS_SPACE_CHECK_RESERVED, tx); + guid, 2, ZFS_SPACE_CHECK_NONE, tx); switch (new_state) { case VDEV_INITIALIZE_ACTIVE: @@ -251,49 +245,6 @@ vdev_initialize_write(vdev_t *vd, uint64_t start, uint64_t size, abd_t *data) } /* - * Translate a logical range to the physical range for the specified vdev_t. - * This function is initially called with a leaf vdev and will walk each - * parent vdev until it reaches a top-level vdev. Once the top-level is - * reached the physical range is initialized and the recursive function - * begins to unwind. As it unwinds it calls the parent's vdev specific - * translation function to do the real conversion. - */ -void -vdev_xlate(vdev_t *vd, const range_seg_t *logical_rs, range_seg_t *physical_rs) -{ - /* - * Walk up the vdev tree - */ - if (vd != vd->vdev_top) { - vdev_xlate(vd->vdev_parent, logical_rs, physical_rs); - } else { - /* - * We've reached the top-level vdev, initialize the - * physical range to the logical range and start to - * unwind. - */ - physical_rs->rs_start = logical_rs->rs_start; - physical_rs->rs_end = logical_rs->rs_end; - return; - } - - vdev_t *pvd = vd->vdev_parent; - ASSERT3P(pvd, !=, NULL); - ASSERT3P(pvd->vdev_ops->vdev_op_xlate, !=, NULL); - - /* - * As this recursive function unwinds, translate the logical - * range into its physical components by calling the - * vdev specific translate function. - */ - range_seg_t intermediate = { { { 0, 0 } } }; - pvd->vdev_ops->vdev_op_xlate(vd, physical_rs, &intermediate); - - physical_rs->rs_start = intermediate.rs_start; - physical_rs->rs_end = intermediate.rs_end; -} - -/* * Callback to fill each ABD chunk with zfs_initialize_value. len must be * divisible by sizeof (uint64_t), and buf must be 8-byte aligned. The ABD * allocation will guarantee these for us. @@ -363,81 +314,6 @@ vdev_initialize_ranges(vdev_t *vd, abd_t *data) } static void -vdev_initialize_mg_wait(metaslab_group_t *mg) -{ - ASSERT(MUTEX_HELD(&mg->mg_ms_initialize_lock)); - while (mg->mg_initialize_updating) { - cv_wait(&mg->mg_ms_initialize_cv, &mg->mg_ms_initialize_lock); - } -} - -static void -vdev_initialize_mg_mark(metaslab_group_t *mg) -{ - ASSERT(MUTEX_HELD(&mg->mg_ms_initialize_lock)); - ASSERT(mg->mg_initialize_updating); - - while (mg->mg_ms_initializing >= max_initialize_ms) { - cv_wait(&mg->mg_ms_initialize_cv, &mg->mg_ms_initialize_lock); - } - mg->mg_ms_initializing++; - ASSERT3U(mg->mg_ms_initializing, <=, max_initialize_ms); -} - -/* - * Mark the metaslab as being initialized to prevent any allocations - * on this metaslab. We must also track how many metaslabs are currently - * being initialized within a metaslab group and limit them to prevent - * allocation failures from occurring because all metaslabs are being - * initialized. - */ -static void -vdev_initialize_ms_mark(metaslab_t *msp) -{ - ASSERT(!MUTEX_HELD(&msp->ms_lock)); - metaslab_group_t *mg = msp->ms_group; - - mutex_enter(&mg->mg_ms_initialize_lock); - - /* - * To keep an accurate count of how many threads are initializing - * a specific metaslab group, we only allow one thread to mark - * the metaslab group at a time. This ensures that the value of - * ms_initializing will be accurate when we decide to mark a metaslab - * group as being initialized. To do this we force all other threads - * to wait till the metaslab's mg_initialize_updating flag is no - * longer set. - */ - vdev_initialize_mg_wait(mg); - mg->mg_initialize_updating = B_TRUE; - if (msp->ms_initializing == 0) { - vdev_initialize_mg_mark(mg); - } - mutex_enter(&msp->ms_lock); - msp->ms_initializing++; - mutex_exit(&msp->ms_lock); - - mg->mg_initialize_updating = B_FALSE; - cv_broadcast(&mg->mg_ms_initialize_cv); - mutex_exit(&mg->mg_ms_initialize_lock); -} - -static void -vdev_initialize_ms_unmark(metaslab_t *msp) -{ - ASSERT(!MUTEX_HELD(&msp->ms_lock)); - metaslab_group_t *mg = msp->ms_group; - mutex_enter(&mg->mg_ms_initialize_lock); - mutex_enter(&msp->ms_lock); - if (--msp->ms_initializing == 0) { - mg->mg_ms_initializing--; - cv_broadcast(&mg->mg_ms_initialize_cv); - } - mutex_exit(&msp->ms_lock); - mutex_exit(&mg->mg_ms_initialize_lock); -} - -static void vdev_initialize_calculate_progress(vdev_t *vd) { ASSERT(spa_config_held(vd->vdev_spa, SCL_CONFIG, RW_READER) || @@ -535,9 +411,8 @@ vdev_initialize_load(vdev_t *vd) return (err); } - /* - * Convert the logical range into a physcial range and add it to our + * Convert the logical range into a physical range and add it to our * avl tree. */ void @@ -618,7 +493,8 @@ vdev_initialize_thread(void *arg) ms_count = vd->vdev_top->vdev_ms_count; } - vdev_initialize_ms_mark(msp); + spa_config_exit(spa, SCL_CONFIG, FTAG); + metaslab_disable(msp); mutex_enter(&msp->ms_lock); VERIFY0(metaslab_load(msp)); @@ -626,16 +502,8 @@ vdev_initialize_thread(void *arg) vd); mutex_exit(&msp->ms_lock); - spa_config_exit(spa, SCL_CONFIG, FTAG); error = vdev_initialize_ranges(vd, deadbeef); - - /* - * Wait for the outstanding IO to be synced to prevent - * newly allocated blocks from being overwritten. - */ - txg_wait_synced(spa_get_dsl(spa), 0); - - vdev_initialize_ms_unmark(msp); + metaslab_enable(msp, B_TRUE); spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); range_tree_vacate(vd->vdev_initialize_tree, NULL, NULL); @@ -853,12 +721,11 @@ vdev_initialize_restart(vdev_t *vd) } #if defined(_KERNEL) -EXPORT_SYMBOL(vdev_initialize_restart); -EXPORT_SYMBOL(vdev_xlate); EXPORT_SYMBOL(vdev_initialize); EXPORT_SYMBOL(vdev_initialize_stop); EXPORT_SYMBOL(vdev_initialize_stop_all); EXPORT_SYMBOL(vdev_initialize_stop_wait); +EXPORT_SYMBOL(vdev_initialize_restart); /* CSTYLED */ module_param(zfs_initialize_value, ulong, 0644); diff --git a/module/zfs/vdev_label.c b/module/zfs/vdev_label.c index a03722d05..a0e373b3d 100644 --- a/module/zfs/vdev_label.c +++ b/module/zfs/vdev_label.c @@ -251,6 +251,9 @@ vdev_config_generate_stats(vdev_t *vd, nvlist_t *nv) fnvlist_add_uint64(nvx, ZPOOL_CONFIG_VDEV_SCRUB_ACTIVE_QUEUE, vsx->vsx_active_queue[ZIO_PRIORITY_SCRUB]); + fnvlist_add_uint64(nvx, ZPOOL_CONFIG_VDEV_TRIM_ACTIVE_QUEUE, + vsx->vsx_active_queue[ZIO_PRIORITY_TRIM]); + /* ZIOs pending */ fnvlist_add_uint64(nvx, ZPOOL_CONFIG_VDEV_SYNC_R_PEND_QUEUE, vsx->vsx_pend_queue[ZIO_PRIORITY_SYNC_READ]); @@ -267,6 +270,9 @@ vdev_config_generate_stats(vdev_t *vd, nvlist_t *nv) fnvlist_add_uint64(nvx, ZPOOL_CONFIG_VDEV_SCRUB_PEND_QUEUE, vsx->vsx_pend_queue[ZIO_PRIORITY_SCRUB]); + fnvlist_add_uint64(nvx, ZPOOL_CONFIG_VDEV_TRIM_PEND_QUEUE, + vsx->vsx_pend_queue[ZIO_PRIORITY_TRIM]); + /* Histograms */ fnvlist_add_uint64_array(nvx, ZPOOL_CONFIG_VDEV_TOT_R_LAT_HISTO, vsx->vsx_total_histo[ZIO_TYPE_READ], @@ -304,6 +310,10 @@ vdev_config_generate_stats(vdev_t *vd, nvlist_t *nv) vsx->vsx_queue_histo[ZIO_PRIORITY_SCRUB], ARRAY_SIZE(vsx->vsx_queue_histo[ZIO_PRIORITY_SCRUB])); + fnvlist_add_uint64_array(nvx, ZPOOL_CONFIG_VDEV_TRIM_LAT_HISTO, + vsx->vsx_queue_histo[ZIO_PRIORITY_TRIM], + ARRAY_SIZE(vsx->vsx_queue_histo[ZIO_PRIORITY_TRIM])); + /* Request sizes */ fnvlist_add_uint64_array(nvx, ZPOOL_CONFIG_VDEV_SYNC_IND_R_HISTO, vsx->vsx_ind_histo[ZIO_PRIORITY_SYNC_READ], @@ -325,6 +335,10 @@ vdev_config_generate_stats(vdev_t *vd, nvlist_t *nv) vsx->vsx_ind_histo[ZIO_PRIORITY_SCRUB], ARRAY_SIZE(vsx->vsx_ind_histo[ZIO_PRIORITY_SCRUB])); + fnvlist_add_uint64_array(nvx, ZPOOL_CONFIG_VDEV_IND_TRIM_HISTO, + vsx->vsx_ind_histo[ZIO_PRIORITY_TRIM], + ARRAY_SIZE(vsx->vsx_ind_histo[ZIO_PRIORITY_TRIM])); + fnvlist_add_uint64_array(nvx, ZPOOL_CONFIG_VDEV_SYNC_AGG_R_HISTO, vsx->vsx_agg_histo[ZIO_PRIORITY_SYNC_READ], ARRAY_SIZE(vsx->vsx_agg_histo[ZIO_PRIORITY_SYNC_READ])); @@ -345,6 +359,10 @@ vdev_config_generate_stats(vdev_t *vd, nvlist_t *nv) vsx->vsx_agg_histo[ZIO_PRIORITY_SCRUB], ARRAY_SIZE(vsx->vsx_agg_histo[ZIO_PRIORITY_SCRUB])); + fnvlist_add_uint64_array(nvx, ZPOOL_CONFIG_VDEV_AGG_TRIM_HISTO, + vsx->vsx_agg_histo[ZIO_PRIORITY_TRIM], + ARRAY_SIZE(vsx->vsx_agg_histo[ZIO_PRIORITY_TRIM])); + /* IO delays */ fnvlist_add_uint64(nvx, ZPOOL_CONFIG_VDEV_SLOW_IOS, vs->vs_slow_ios); diff --git a/module/zfs/vdev_queue.c b/module/zfs/vdev_queue.c index a1861d5f0..e74df76b7 100644 --- a/module/zfs/vdev_queue.c +++ b/module/zfs/vdev_queue.c @@ -156,6 +156,8 @@ uint32_t zfs_vdev_removal_min_active = 1; uint32_t zfs_vdev_removal_max_active = 2; uint32_t zfs_vdev_initializing_min_active = 1; uint32_t zfs_vdev_initializing_max_active = 1; +uint32_t zfs_vdev_trim_min_active = 1; +uint32_t zfs_vdev_trim_max_active = 2; /* * When the pool has less than zfs_vdev_async_write_active_min_dirty_percent @@ -203,6 +205,12 @@ int zfs_vdev_queue_depth_pct = 300; */ int zfs_vdev_def_queue_depth = 32; +/* + * Allow TRIM I/Os to be aggregated. This should normally not be needed since + * TRIM I/O for extents up to zfs_trim_extent_bytes_max (128M) can be submitted + * by the TRIM code in zfs_trim.c. + */ +int zfs_vdev_aggregate_trim = 0; int vdev_queue_offset_compare(const void *x1, const void *x2) @@ -227,11 +235,13 @@ vdev_queue_class_tree(vdev_queue_t *vq, zio_priority_t p) static inline avl_tree_t * vdev_queue_type_tree(vdev_queue_t *vq, zio_type_t t) { - ASSERT(t == ZIO_TYPE_READ || t == ZIO_TYPE_WRITE); + ASSERT(t == ZIO_TYPE_READ || t == ZIO_TYPE_WRITE || t == ZIO_TYPE_TRIM); if (t == ZIO_TYPE_READ) return (&vq->vq_read_offset_tree); - else + else if (t == ZIO_TYPE_WRITE) return (&vq->vq_write_offset_tree); + else + return (&vq->vq_trim_offset_tree); } int @@ -266,6 +276,8 @@ vdev_queue_class_min_active(zio_priority_t p) return (zfs_vdev_removal_min_active); case ZIO_PRIORITY_INITIALIZING: return (zfs_vdev_initializing_min_active); + case ZIO_PRIORITY_TRIM: + return (zfs_vdev_trim_min_active); default: panic("invalid priority %u", p); return (0); @@ -338,6 +350,8 @@ vdev_queue_class_max_active(spa_t *spa, zio_priority_t p) return (zfs_vdev_removal_max_active); case ZIO_PRIORITY_INITIALIZING: return (zfs_vdev_initializing_max_active); + case ZIO_PRIORITY_TRIM: + return (zfs_vdev_trim_max_active); default: panic("invalid priority %u", p); return (0); @@ -398,19 +412,25 @@ vdev_queue_init(vdev_t *vd) avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE), vdev_queue_offset_compare, sizeof (zio_t), offsetof(struct zio, io_offset_node)); + avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_TRIM), + vdev_queue_offset_compare, sizeof (zio_t), + offsetof(struct zio, io_offset_node)); for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) { int (*compfn) (const void *, const void *); /* - * The synchronous i/o queues are dispatched in FIFO rather + * The synchronous/trim i/o queues are dispatched in FIFO rather * than LBA order. This provides more consistent latency for * these i/os. */ - if (p == ZIO_PRIORITY_SYNC_READ || p == ZIO_PRIORITY_SYNC_WRITE) + if (p == ZIO_PRIORITY_SYNC_READ || + p == ZIO_PRIORITY_SYNC_WRITE || + p == ZIO_PRIORITY_TRIM) { compfn = vdev_queue_timestamp_compare; - else + } else { compfn = vdev_queue_offset_compare; + } avl_create(vdev_queue_class_tree(vq, p), compfn, sizeof (zio_t), offsetof(struct zio, io_queue_node)); } @@ -428,6 +448,7 @@ vdev_queue_fini(vdev_t *vd) avl_destroy(&vq->vq_active_tree); avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_READ)); avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE)); + avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_TRIM)); mutex_destroy(&vq->vq_lock); } @@ -559,6 +580,13 @@ vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio) if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE || limit == 0) return (NULL); + /* + * While TRIM commands could be aggregated based on offset this + * behavior is disabled until it's determined to be beneficial. + */ + if (zio->io_type == ZIO_TYPE_TRIM && !zfs_vdev_aggregate_trim) + return (NULL); + first = last = zio; if (zio->io_type == ZIO_TYPE_READ) @@ -732,7 +760,7 @@ again: * For LBA-ordered queues (async / scrub / initializing), issue the * i/o which follows the most recently issued i/o in LBA (offset) order. * - * For FIFO queues (sync), issue the i/o with the lowest timestamp. + * For FIFO queues (sync/trim), issue the i/o with the lowest timestamp. */ tree = vdev_queue_class_tree(vq, p); vq->vq_io_search.io_timestamp = 0; @@ -783,19 +811,27 @@ vdev_queue_io(zio_t *zio) * not match the child's i/o type. Fix it up here. */ if (zio->io_type == ZIO_TYPE_READ) { + ASSERT(zio->io_priority != ZIO_PRIORITY_TRIM); + if (zio->io_priority != ZIO_PRIORITY_SYNC_READ && zio->io_priority != ZIO_PRIORITY_ASYNC_READ && zio->io_priority != ZIO_PRIORITY_SCRUB && zio->io_priority != ZIO_PRIORITY_REMOVAL && - zio->io_priority != ZIO_PRIORITY_INITIALIZING) + zio->io_priority != ZIO_PRIORITY_INITIALIZING) { zio->io_priority = ZIO_PRIORITY_ASYNC_READ; - } else { - ASSERT(zio->io_type == ZIO_TYPE_WRITE); + } + } else if (zio->io_type == ZIO_TYPE_WRITE) { + ASSERT(zio->io_priority != ZIO_PRIORITY_TRIM); + if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE && zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE && zio->io_priority != ZIO_PRIORITY_REMOVAL && - zio->io_priority != ZIO_PRIORITY_INITIALIZING) + zio->io_priority != ZIO_PRIORITY_INITIALIZING) { zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE; + } + } else { + ASSERT(zio->io_type == ZIO_TYPE_TRIM); + ASSERT(zio->io_priority == ZIO_PRIORITY_TRIM); } zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE; @@ -922,6 +958,9 @@ module_param(zfs_vdev_aggregation_limit_non_rotating, int, 0644); MODULE_PARM_DESC(zfs_vdev_aggregation_limit_non_rotating, "Max vdev I/O aggregation size for non-rotating media"); +module_param(zfs_vdev_aggregate_trim, int, 0644); +MODULE_PARM_DESC(zfs_vdev_aggregate_trim, "Allow TRIM I/O to be aggregated"); + module_param(zfs_vdev_read_gap_limit, int, 0644); MODULE_PARM_DESC(zfs_vdev_read_gap_limit, "Aggregate read I/O over gap"); @@ -995,6 +1034,14 @@ module_param(zfs_vdev_sync_write_min_active, int, 0644); MODULE_PARM_DESC(zfs_vdev_sync_write_min_active, "Min active sync write I/Os per vdev"); +module_param(zfs_vdev_trim_max_active, int, 0644); +MODULE_PARM_DESC(zfs_vdev_trim_max_active, + "Max active trim/discard I/Os per vdev"); + +module_param(zfs_vdev_trim_min_active, int, 0644); +MODULE_PARM_DESC(zfs_vdev_trim_min_active, + "Min active trim/discard I/Os per vdev"); + module_param(zfs_vdev_queue_depth_pct, int, 0644); MODULE_PARM_DESC(zfs_vdev_queue_depth_pct, "Queue depth percentage for each top-level vdev"); diff --git a/module/zfs/vdev_raidz.c b/module/zfs/vdev_raidz.c index d11287bdc..215cd1c12 100644 --- a/module/zfs/vdev_raidz.c +++ b/module/zfs/vdev_raidz.c @@ -37,7 +37,7 @@ #include <sys/vdev_raidz_impl.h> #ifdef ZFS_DEBUG -#include <sys/vdev_initialize.h> /* vdev_xlate testing */ +#include <sys/vdev.h> /* For vdev_xlate() in vdev_raidz_io_verify() */ #endif /* diff --git a/module/zfs/vdev_removal.c b/module/zfs/vdev_removal.c index 98bf2194a..99d67b7be 100644 --- a/module/zfs/vdev_removal.c +++ b/module/zfs/vdev_removal.c @@ -45,6 +45,7 @@ #include <sys/vdev_indirect_mapping.h> #include <sys/abd.h> #include <sys/vdev_initialize.h> +#include <sys/vdev_trim.h> #include <sys/trace_vdev.h> /* @@ -1181,6 +1182,8 @@ vdev_remove_complete(spa_t *spa) txg = spa_vdev_enter(spa); vdev_t *vd = vdev_lookup_top(spa, spa->spa_vdev_removal->svr_vdev_id); ASSERT3P(vd->vdev_initialize_thread, ==, NULL); + ASSERT3P(vd->vdev_trim_thread, ==, NULL); + ASSERT3P(vd->vdev_autotrim_thread, ==, NULL); sysevent_t *ev = spa_event_create(spa, vd, NULL, ESC_ZFS_VDEV_REMOVE_DEV); @@ -1869,8 +1872,10 @@ spa_vdev_remove_log(vdev_t *vd, uint64_t *txg) spa_vdev_config_exit(spa, NULL, *txg, 0, FTAG); - /* Stop initializing */ + /* Stop initializing and TRIM */ vdev_initialize_stop_all(vd, VDEV_INITIALIZE_CANCELED); + vdev_trim_stop_all(vd, VDEV_TRIM_CANCELED); + vdev_autotrim_stop_wait(vd); *txg = spa_vdev_config_enter(spa); @@ -2051,11 +2056,13 @@ spa_vdev_remove_top(vdev_t *vd, uint64_t *txg) error = spa_reset_logs(spa); /* - * We stop any initializing that is currently in progress but leave - * the state as "active". This will allow the initializing to resume - * if the removal is canceled sometime later. + * We stop any initializing and TRIM that is currently in progress + * but leave the state as "active". This will allow the process to + * resume if the removal is canceled sometime later. */ vdev_initialize_stop_all(vd, VDEV_INITIALIZE_ACTIVE); + vdev_trim_stop_all(vd, VDEV_TRIM_ACTIVE); + vdev_autotrim_stop_wait(vd); *txg = spa_vdev_config_enter(spa); @@ -2069,6 +2076,8 @@ spa_vdev_remove_top(vdev_t *vd, uint64_t *txg) if (error != 0) { metaslab_group_activate(mg); spa_async_request(spa, SPA_ASYNC_INITIALIZE_RESTART); + spa_async_request(spa, SPA_ASYNC_TRIM_RESTART); + spa_async_request(spa, SPA_ASYNC_AUTOTRIM_RESTART); return (error); } diff --git a/module/zfs/vdev_trim.c b/module/zfs/vdev_trim.c new file mode 100644 index 000000000..5ad47cccd --- /dev/null +++ b/module/zfs/vdev_trim.c @@ -0,0 +1,1460 @@ +/* + * CDDL HEADER START + * + * The contents of this file are subject to the terms of the + * Common Development and Distribution License (the "License"). + * You may not use this file except in compliance with the License. + * + * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE + * or http://www.opensolaris.org/os/licensing. + * See the License for the specific language governing permissions + * and limitations under the License. + * + * When distributing Covered Code, include this CDDL HEADER in each + * file and include the License file at usr/src/OPENSOLARIS.LICENSE. + * If applicable, add the following below this CDDL HEADER, with the + * fields enclosed by brackets "[]" replaced with your own identifying + * information: Portions Copyright [yyyy] [name of copyright owner] + * + * CDDL HEADER END + */ + +/* + * Copyright (c) 2016 by Delphix. All rights reserved. + * Copyright (c) 2019 by Lawrence Livermore National Security, LLC. + */ + +#include <sys/spa.h> +#include <sys/spa_impl.h> +#include <sys/txg.h> +#include <sys/vdev_impl.h> +#include <sys/vdev_trim.h> +#include <sys/refcount.h> +#include <sys/metaslab_impl.h> +#include <sys/dsl_synctask.h> +#include <sys/zap.h> +#include <sys/dmu_tx.h> + +/* + * TRIM is a feature which is used to notify a SSD that some previously + * written space is no longer allocated by the pool. This is useful because + * writes to a SSD must be performed to blocks which have first been erased. + * Ensuring the SSD always has a supply of erased blocks for new writes + * helps prevent the performance from deteriorating. + * + * There are two supported TRIM methods; manual and automatic. + * + * Manual TRIM: + * + * A manual TRIM is initiated by running the 'zpool trim' command. A single + * 'vdev_trim' thread is created for each leaf vdev, and it is responsible for + * managing that vdev TRIM process. This involves iterating over all the + * metaslabs, calculating the unallocated space ranges, and then issuing the + * required TRIM I/Os. + * + * While a metaslab is being actively trimmed it is not eligible to perform + * new allocations. After traversing all of the metaslabs the thread is + * terminated. Finally, both the requested options and current progress of + * the TRIM are regularly written to the pool. This allows the TRIM to be + * suspended and resumed as needed. + * + * Automatic TRIM: + * + * An automatic TRIM is enabled by setting the 'autotrim' pool property + * to 'on'. When enabled, a `vdev_autotrim' thread is created for each + * top-level (not leaf) vdev in the pool. These threads perform the same + * core TRIM process as a manual TRIM, but with a few key differences. + * + * 1) Automatic TRIM happens continuously in the background and operates + * solely on recently freed blocks (ms_trim not ms_allocatable). + * + * 2) Each thread is associated with a top-level (not leaf) vdev. This has + * the benefit of simplifying the threading model, it makes it easier + * to coordinate administrative commands, and it ensures only a single + * metaslab is disabled at a time. Unlike manual TRIM, this means each + * 'vdev_autotrim' thread is responsible for issuing TRIM I/Os for its + * children. + * + * 3) There is no automatic TRIM progress information stored on disk, nor + * is it reported by 'zpool status'. + * + * While the automatic TRIM process is highly effective it is more likely + * than a manual TRIM to encounter tiny ranges. Ranges less than or equal to + * 'zfs_trim_extent_bytes_min' (32k) are considered too small to efficiently + * TRIM and are skipped. This means small amounts of freed space may not + * be automatically trimmed. + * + * Furthermore, devices with attached hot spares and devices being actively + * replaced are skipped. This is done to avoid adding additional stress to + * a potentially unhealthy device and to minimize the required rebuild time. + * + * For this reason it may be beneficial to occasionally manually TRIM a pool + * even when automatic TRIM is enabled. + */ + +/* + * Maximum size of TRIM I/O, ranges will be chunked in to 128MiB lengths. + */ +unsigned int zfs_trim_extent_bytes_max = 128 * 1024 * 1024; + +/* + * Minimum size of TRIM I/O, extents smaller than 32Kib will be skipped. + */ +unsigned int zfs_trim_extent_bytes_min = 32 * 1024; + +/* + * Skip uninitialized metaslabs during the TRIM process. This option is + * useful for pools constructed from large thinly-provisioned devices where + * TRIM operations are slow. As a pool ages an increasing fraction of + * the pools metaslabs will be initialized progressively degrading the + * usefulness of this option. This setting is stored when starting a + * manual TRIM and will persist for the duration of the requested TRIM. + */ +unsigned int zfs_trim_metaslab_skip = 0; + +/* + * Maximum number of queued TRIM I/Os per leaf vdev. The number of + * concurrent TRIM I/Os issued to the device is controlled by the + * zfs_vdev_trim_min_active and zfs_vdev_trim_max_active module options. + */ +unsigned int zfs_trim_queue_limit = 10; + +/* + * The minimum number of transaction groups between automatic trims of a + * metaslab. This setting represents a trade-off between issuing more + * efficient TRIM operations, by allowing them to be aggregated longer, + * and issuing them promptly so the trimmed space is available. Note + * that this value is a minimum; metaslabs can be trimmed less frequently + * when there are a large number of ranges which need to be trimmed. + * + * Increasing this value will allow frees to be aggregated for a longer + * time. This can result is larger TRIM operations, and increased memory + * usage in order to track the ranges to be trimmed. Decreasing this value + * has the opposite effect. The default value of 32 was determined though + * testing to be a reasonable compromise. + */ +unsigned int zfs_trim_txg_batch = 32; + +/* + * The trim_args are a control structure which describe how a leaf vdev + * should be trimmed. The core elements are the vdev, the metaslab being + * trimmed and a range tree containing the extents to TRIM. All provided + * ranges must be within the metaslab. + */ +typedef struct trim_args { + /* + * These fields are set by the caller of vdev_trim_ranges(). + */ + vdev_t *trim_vdev; /* Leaf vdev to TRIM */ + metaslab_t *trim_msp; /* Disabled metaslab */ + range_tree_t *trim_tree; /* TRIM ranges (in metaslab) */ + trim_type_t trim_type; /* Manual or auto TRIM */ + uint64_t trim_extent_bytes_max; /* Maximum TRIM I/O size */ + uint64_t trim_extent_bytes_min; /* Minimum TRIM I/O size */ + enum trim_flag trim_flags; /* TRIM flags (secure) */ + + /* + * These fields are updated by vdev_trim_ranges(). + */ + hrtime_t trim_start_time; /* Start time */ + uint64_t trim_bytes_done; /* Bytes trimmed */ +} trim_args_t; + +/* + * Determines whether a vdev_trim_thread() should be stopped. + */ +static boolean_t +vdev_trim_should_stop(vdev_t *vd) +{ + return (vd->vdev_trim_exit_wanted || !vdev_writeable(vd) || + vd->vdev_detached || vd->vdev_top->vdev_removing); +} + +/* + * Determines whether a vdev_autotrim_thread() should be stopped. + */ +static boolean_t +vdev_autotrim_should_stop(vdev_t *tvd) +{ + return (tvd->vdev_autotrim_exit_wanted || + !vdev_writeable(tvd) || tvd->vdev_removing || + spa_get_autotrim(tvd->vdev_spa) == SPA_AUTOTRIM_OFF); +} + +/* + * The sync task for updating the on-disk state of a manual TRIM. This + * is scheduled by vdev_trim_change_state(). + */ +static void +vdev_trim_zap_update_sync(void *arg, dmu_tx_t *tx) +{ + /* + * We pass in the guid instead of the vdev_t since the vdev may + * have been freed prior to the sync task being processed. This + * happens when a vdev is detached as we call spa_config_vdev_exit(), + * stop the trimming thread, schedule the sync task, and free + * the vdev. Later when the scheduled sync task is invoked, it would + * find that the vdev has been freed. + */ + uint64_t guid = *(uint64_t *)arg; + uint64_t txg = dmu_tx_get_txg(tx); + kmem_free(arg, sizeof (uint64_t)); + + vdev_t *vd = spa_lookup_by_guid(tx->tx_pool->dp_spa, guid, B_FALSE); + if (vd == NULL || vd->vdev_top->vdev_removing || !vdev_is_concrete(vd)) + return; + + uint64_t last_offset = vd->vdev_trim_offset[txg & TXG_MASK]; + vd->vdev_trim_offset[txg & TXG_MASK] = 0; + + VERIFY3U(vd->vdev_leaf_zap, !=, 0); + + objset_t *mos = vd->vdev_spa->spa_meta_objset; + + if (last_offset > 0 || vd->vdev_trim_last_offset == UINT64_MAX) { + + if (vd->vdev_trim_last_offset == UINT64_MAX) + last_offset = 0; + + vd->vdev_trim_last_offset = last_offset; + VERIFY0(zap_update(mos, vd->vdev_leaf_zap, + VDEV_LEAF_ZAP_TRIM_LAST_OFFSET, + sizeof (last_offset), 1, &last_offset, tx)); + } + + if (vd->vdev_trim_action_time > 0) { + uint64_t val = (uint64_t)vd->vdev_trim_action_time; + VERIFY0(zap_update(mos, vd->vdev_leaf_zap, + VDEV_LEAF_ZAP_TRIM_ACTION_TIME, sizeof (val), + 1, &val, tx)); + } + + if (vd->vdev_trim_rate > 0) { + uint64_t rate = (uint64_t)vd->vdev_trim_rate; + + if (rate == UINT64_MAX) + rate = 0; + + VERIFY0(zap_update(mos, vd->vdev_leaf_zap, + VDEV_LEAF_ZAP_TRIM_RATE, sizeof (rate), 1, &rate, tx)); + } + + uint64_t partial = vd->vdev_trim_partial; + if (partial == UINT64_MAX) + partial = 0; + + VERIFY0(zap_update(mos, vd->vdev_leaf_zap, VDEV_LEAF_ZAP_TRIM_PARTIAL, + sizeof (partial), 1, &partial, tx)); + + uint64_t secure = vd->vdev_trim_secure; + if (secure == UINT64_MAX) + secure = 0; + + VERIFY0(zap_update(mos, vd->vdev_leaf_zap, VDEV_LEAF_ZAP_TRIM_SECURE, + sizeof (secure), 1, &secure, tx)); + + + uint64_t trim_state = vd->vdev_trim_state; + VERIFY0(zap_update(mos, vd->vdev_leaf_zap, VDEV_LEAF_ZAP_TRIM_STATE, + sizeof (trim_state), 1, &trim_state, tx)); +} + +/* + * Update the on-disk state of a manual TRIM. This is called to request + * that a TRIM be started/suspended/canceled, or to change one of the + * TRIM options (partial, secure, rate). + */ +static void +vdev_trim_change_state(vdev_t *vd, vdev_trim_state_t new_state, + uint64_t rate, boolean_t partial, boolean_t secure) +{ + ASSERT(MUTEX_HELD(&vd->vdev_trim_lock)); + spa_t *spa = vd->vdev_spa; + + if (new_state == vd->vdev_trim_state) + return; + + /* + * Copy the vd's guid, this will be freed by the sync task. + */ + uint64_t *guid = kmem_zalloc(sizeof (uint64_t), KM_SLEEP); + *guid = vd->vdev_guid; + + /* + * If we're suspending, then preserve the original start time. + */ + if (vd->vdev_trim_state != VDEV_TRIM_SUSPENDED) { + vd->vdev_trim_action_time = gethrestime_sec(); + } + + /* + * If we're activating, then preserve the requested rate and trim + * method. Setting the last offset and rate to UINT64_MAX is used + * as a sentinel to indicate they should be reset to default values. + */ + if (new_state == VDEV_TRIM_ACTIVE) { + if (vd->vdev_trim_state == VDEV_TRIM_COMPLETE || + vd->vdev_trim_state == VDEV_TRIM_CANCELED) { + vd->vdev_trim_last_offset = UINT64_MAX; + vd->vdev_trim_rate = UINT64_MAX; + vd->vdev_trim_partial = UINT64_MAX; + vd->vdev_trim_secure = UINT64_MAX; + } + + if (rate != 0) + vd->vdev_trim_rate = rate; + + if (partial != 0) + vd->vdev_trim_partial = partial; + + if (secure != 0) + vd->vdev_trim_secure = secure; + } + + boolean_t resumed = !!(vd->vdev_trim_state == VDEV_TRIM_SUSPENDED); + vd->vdev_trim_state = new_state; + + dmu_tx_t *tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir); + VERIFY0(dmu_tx_assign(tx, TXG_WAIT)); + dsl_sync_task_nowait(spa_get_dsl(spa), vdev_trim_zap_update_sync, + guid, 2, ZFS_SPACE_CHECK_NONE, tx); + + switch (new_state) { + case VDEV_TRIM_ACTIVE: + spa_event_notify(spa, vd, NULL, + resumed ? ESC_ZFS_TRIM_RESUME : ESC_ZFS_TRIM_START); + spa_history_log_internal(spa, "trim", tx, + "vdev=%s activated", vd->vdev_path); + break; + case VDEV_TRIM_SUSPENDED: + spa_event_notify(spa, vd, NULL, ESC_ZFS_TRIM_SUSPEND); + spa_history_log_internal(spa, "trim", tx, + "vdev=%s suspended", vd->vdev_path); + break; + case VDEV_TRIM_CANCELED: + spa_event_notify(spa, vd, NULL, ESC_ZFS_TRIM_CANCEL); + spa_history_log_internal(spa, "trim", tx, + "vdev=%s canceled", vd->vdev_path); + break; + case VDEV_TRIM_COMPLETE: + spa_event_notify(spa, vd, NULL, ESC_ZFS_TRIM_FINISH); + spa_history_log_internal(spa, "trim", tx, + "vdev=%s complete", vd->vdev_path); + break; + default: + panic("invalid state %llu", (unsigned long long)new_state); + } + + dmu_tx_commit(tx); +} + +/* + * The zio_done_func_t done callback for each manual TRIM issued. It is + * responsible for updating the TRIM stats, reissuing failed TRIM I/Os, + * and limiting the number of in flight TRIM I/Os. + */ +static void +vdev_trim_cb(zio_t *zio) +{ + vdev_t *vd = zio->io_vd; + + mutex_enter(&vd->vdev_trim_io_lock); + if (zio->io_error == ENXIO && !vdev_writeable(vd)) { + /* + * The I/O failed because the vdev was unavailable; roll the + * last offset back. (This works because spa_sync waits on + * spa_txg_zio before it runs sync tasks.) + */ + uint64_t *offset = + &vd->vdev_trim_offset[zio->io_txg & TXG_MASK]; + *offset = MIN(*offset, zio->io_offset); + } else { + if (zio->io_error != 0) { + vd->vdev_stat.vs_trim_errors++; + spa_iostats_trim_add(vd->vdev_spa, TRIM_TYPE_MANUAL, + 0, 0, 0, 0, 1, zio->io_orig_size); + } else { + spa_iostats_trim_add(vd->vdev_spa, TRIM_TYPE_MANUAL, + 1, zio->io_orig_size, 0, 0, 0, 0); + } + + vd->vdev_trim_bytes_done += zio->io_orig_size; + } + + ASSERT3U(vd->vdev_trim_inflight[TRIM_TYPE_MANUAL], >, 0); + vd->vdev_trim_inflight[TRIM_TYPE_MANUAL]--; + cv_broadcast(&vd->vdev_trim_io_cv); + mutex_exit(&vd->vdev_trim_io_lock); + + spa_config_exit(vd->vdev_spa, SCL_STATE_ALL, vd); +} + +/* + * The zio_done_func_t done callback for each automatic TRIM issued. It + * is responsible for updating the TRIM stats and limiting the number of + * in flight TRIM I/Os. Automatic TRIM I/Os are best effort and are + * never reissued on failure. + */ +static void +vdev_autotrim_cb(zio_t *zio) +{ + vdev_t *vd = zio->io_vd; + + mutex_enter(&vd->vdev_trim_io_lock); + + if (zio->io_error != 0) { + vd->vdev_stat.vs_trim_errors++; + spa_iostats_trim_add(vd->vdev_spa, TRIM_TYPE_AUTO, + 0, 0, 0, 0, 1, zio->io_orig_size); + } else { + spa_iostats_trim_add(vd->vdev_spa, TRIM_TYPE_AUTO, + 1, zio->io_orig_size, 0, 0, 0, 0); + } + + ASSERT3U(vd->vdev_trim_inflight[TRIM_TYPE_AUTO], >, 0); + vd->vdev_trim_inflight[TRIM_TYPE_AUTO]--; + cv_broadcast(&vd->vdev_trim_io_cv); + mutex_exit(&vd->vdev_trim_io_lock); + + spa_config_exit(vd->vdev_spa, SCL_STATE_ALL, vd); +} + +/* + * Returns the average trim rate in bytes/sec for the ta->trim_vdev. + */ +static uint64_t +vdev_trim_calculate_rate(trim_args_t *ta) +{ + return (ta->trim_bytes_done * 1000 / + (NSEC2MSEC(gethrtime() - ta->trim_start_time) + 1)); +} + +/* + * Issues a physical TRIM and takes care of rate limiting (bytes/sec) + * and number of concurrent TRIM I/Os. + */ +static int +vdev_trim_range(trim_args_t *ta, uint64_t start, uint64_t size) +{ + vdev_t *vd = ta->trim_vdev; + spa_t *spa = vd->vdev_spa; + + mutex_enter(&vd->vdev_trim_io_lock); + + /* + * Limit manual TRIM I/Os to the requested rate. This does not + * apply to automatic TRIM since no per vdev rate can be specified. + */ + if (ta->trim_type == TRIM_TYPE_MANUAL) { + while (vd->vdev_trim_rate != 0 && !vdev_trim_should_stop(vd) && + vdev_trim_calculate_rate(ta) > vd->vdev_trim_rate) { + cv_timedwait_sig(&vd->vdev_trim_io_cv, + &vd->vdev_trim_io_lock, ddi_get_lbolt() + + MSEC_TO_TICK(10)); + } + } + ta->trim_bytes_done += size; + + /* Limit in flight trimming I/Os */ + while (vd->vdev_trim_inflight[0] + vd->vdev_trim_inflight[1] >= + zfs_trim_queue_limit) { + cv_wait(&vd->vdev_trim_io_cv, &vd->vdev_trim_io_lock); + } + vd->vdev_trim_inflight[ta->trim_type]++; + mutex_exit(&vd->vdev_trim_io_lock); + + dmu_tx_t *tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir); + VERIFY0(dmu_tx_assign(tx, TXG_WAIT)); + uint64_t txg = dmu_tx_get_txg(tx); + + spa_config_enter(spa, SCL_STATE_ALL, vd, RW_READER); + mutex_enter(&vd->vdev_trim_lock); + + if (ta->trim_type == TRIM_TYPE_MANUAL && + vd->vdev_trim_offset[txg & TXG_MASK] == 0) { + uint64_t *guid = kmem_zalloc(sizeof (uint64_t), KM_SLEEP); + *guid = vd->vdev_guid; + + /* This is the first write of this txg. */ + dsl_sync_task_nowait(spa_get_dsl(spa), + vdev_trim_zap_update_sync, guid, 2, + ZFS_SPACE_CHECK_RESERVED, tx); + } + + /* + * We know the vdev_t will still be around since all consumers of + * vdev_free must stop the trimming first. + */ + if ((ta->trim_type == TRIM_TYPE_MANUAL && + vdev_trim_should_stop(vd)) || + (ta->trim_type == TRIM_TYPE_AUTO && + vdev_autotrim_should_stop(vd->vdev_top))) { + mutex_enter(&vd->vdev_trim_io_lock); + vd->vdev_trim_inflight[ta->trim_type]--; + mutex_exit(&vd->vdev_trim_io_lock); + spa_config_exit(vd->vdev_spa, SCL_STATE_ALL, vd); + mutex_exit(&vd->vdev_trim_lock); + dmu_tx_commit(tx); + return (SET_ERROR(EINTR)); + } + mutex_exit(&vd->vdev_trim_lock); + + if (ta->trim_type == TRIM_TYPE_MANUAL) + vd->vdev_trim_offset[txg & TXG_MASK] = start + size; + + zio_nowait(zio_trim(spa->spa_txg_zio[txg & TXG_MASK], vd, + start, size, ta->trim_type == TRIM_TYPE_MANUAL ? + vdev_trim_cb : vdev_autotrim_cb, NULL, + ZIO_PRIORITY_TRIM, ZIO_FLAG_CANFAIL, ta->trim_flags)); + /* vdev_trim_cb and vdev_autotrim_cb release SCL_STATE_ALL */ + + dmu_tx_commit(tx); + + return (0); +} + +/* + * Issues TRIM I/Os for all ranges in the provided ta->trim_tree range tree. + * Additional parameters describing how the TRIM should be performed must + * be set in the trim_args structure. See the trim_args definition for + * additional information. + */ +static int +vdev_trim_ranges(trim_args_t *ta) +{ + vdev_t *vd = ta->trim_vdev; + avl_tree_t *rt = &ta->trim_tree->rt_root; + uint64_t extent_bytes_max = ta->trim_extent_bytes_max; + uint64_t extent_bytes_min = ta->trim_extent_bytes_min; + spa_t *spa = vd->vdev_spa; + + ta->trim_start_time = gethrtime(); + ta->trim_bytes_done = 0; + + for (range_seg_t *rs = avl_first(rt); rs != NULL; + rs = AVL_NEXT(rt, rs)) { + uint64_t size = rs->rs_end - rs->rs_start; + + if (extent_bytes_min && size < extent_bytes_min) { + spa_iostats_trim_add(spa, ta->trim_type, + 0, 0, 1, size, 0, 0); + continue; + } + + /* Split range into legally-sized physical chunks */ + uint64_t writes_required = ((size - 1) / extent_bytes_max) + 1; + + for (uint64_t w = 0; w < writes_required; w++) { + int error; + + error = vdev_trim_range(ta, VDEV_LABEL_START_SIZE + + rs->rs_start + (w * extent_bytes_max), + MIN(size - (w * extent_bytes_max), + extent_bytes_max)); + if (error != 0) { + return (error); + } + } + } + + return (0); +} + +/* + * Calculates the completion percentage of a manual TRIM. + */ +static void +vdev_trim_calculate_progress(vdev_t *vd) +{ + ASSERT(spa_config_held(vd->vdev_spa, SCL_CONFIG, RW_READER) || + spa_config_held(vd->vdev_spa, SCL_CONFIG, RW_WRITER)); + ASSERT(vd->vdev_leaf_zap != 0); + + vd->vdev_trim_bytes_est = 0; + vd->vdev_trim_bytes_done = 0; + + for (uint64_t i = 0; i < vd->vdev_top->vdev_ms_count; i++) { + metaslab_t *msp = vd->vdev_top->vdev_ms[i]; + mutex_enter(&msp->ms_lock); + + uint64_t ms_free = msp->ms_size - + metaslab_allocated_space(msp); + + if (vd->vdev_top->vdev_ops == &vdev_raidz_ops) + ms_free /= vd->vdev_top->vdev_children; + + /* + * Convert the metaslab range to a physical range + * on our vdev. We use this to determine if we are + * in the middle of this metaslab range. + */ + range_seg_t logical_rs, physical_rs; + logical_rs.rs_start = msp->ms_start; + logical_rs.rs_end = msp->ms_start + msp->ms_size; + vdev_xlate(vd, &logical_rs, &physical_rs); + + if (vd->vdev_trim_last_offset <= physical_rs.rs_start) { + vd->vdev_trim_bytes_est += ms_free; + mutex_exit(&msp->ms_lock); + continue; + } else if (vd->vdev_trim_last_offset > physical_rs.rs_end) { + vd->vdev_trim_bytes_done += ms_free; + vd->vdev_trim_bytes_est += ms_free; + mutex_exit(&msp->ms_lock); + continue; + } + + /* + * If we get here, we're in the middle of trimming this + * metaslab. Load it and walk the free tree for more + * accurate progress estimation. + */ + VERIFY0(metaslab_load(msp)); + + for (range_seg_t *rs = avl_first(&msp->ms_allocatable->rt_root); + rs; rs = AVL_NEXT(&msp->ms_allocatable->rt_root, rs)) { + logical_rs.rs_start = rs->rs_start; + logical_rs.rs_end = rs->rs_end; + vdev_xlate(vd, &logical_rs, &physical_rs); + + uint64_t size = physical_rs.rs_end - + physical_rs.rs_start; + vd->vdev_trim_bytes_est += size; + if (vd->vdev_trim_last_offset >= physical_rs.rs_end) { + vd->vdev_trim_bytes_done += size; + } else if (vd->vdev_trim_last_offset > + physical_rs.rs_start && + vd->vdev_trim_last_offset <= + physical_rs.rs_end) { + vd->vdev_trim_bytes_done += + vd->vdev_trim_last_offset - + physical_rs.rs_start; + } + } + mutex_exit(&msp->ms_lock); + } +} + +/* + * Load from disk the vdev's manual TRIM information. This includes the + * state, progress, and options provided when initiating the manual TRIM. + */ +static int +vdev_trim_load(vdev_t *vd) +{ + int err = 0; + ASSERT(spa_config_held(vd->vdev_spa, SCL_CONFIG, RW_READER) || + spa_config_held(vd->vdev_spa, SCL_CONFIG, RW_WRITER)); + ASSERT(vd->vdev_leaf_zap != 0); + + if (vd->vdev_trim_state == VDEV_TRIM_ACTIVE || + vd->vdev_trim_state == VDEV_TRIM_SUSPENDED) { + err = zap_lookup(vd->vdev_spa->spa_meta_objset, + vd->vdev_leaf_zap, VDEV_LEAF_ZAP_TRIM_LAST_OFFSET, + sizeof (vd->vdev_trim_last_offset), 1, + &vd->vdev_trim_last_offset); + if (err == ENOENT) { + vd->vdev_trim_last_offset = 0; + err = 0; + } + + if (err == 0) { + err = zap_lookup(vd->vdev_spa->spa_meta_objset, + vd->vdev_leaf_zap, VDEV_LEAF_ZAP_TRIM_RATE, + sizeof (vd->vdev_trim_rate), 1, + &vd->vdev_trim_rate); + if (err == ENOENT) { + vd->vdev_trim_rate = 0; + err = 0; + } + } + + if (err == 0) { + err = zap_lookup(vd->vdev_spa->spa_meta_objset, + vd->vdev_leaf_zap, VDEV_LEAF_ZAP_TRIM_PARTIAL, + sizeof (vd->vdev_trim_partial), 1, + &vd->vdev_trim_partial); + if (err == ENOENT) { + vd->vdev_trim_partial = 0; + err = 0; + } + } + + if (err == 0) { + err = zap_lookup(vd->vdev_spa->spa_meta_objset, + vd->vdev_leaf_zap, VDEV_LEAF_ZAP_TRIM_SECURE, + sizeof (vd->vdev_trim_secure), 1, + &vd->vdev_trim_secure); + if (err == ENOENT) { + vd->vdev_trim_secure = 0; + err = 0; + } + } + } + + vdev_trim_calculate_progress(vd); + + return (err); +} + +/* + * Convert the logical range into a physical range and add it to the + * range tree passed in the trim_args_t. + */ +static void +vdev_trim_range_add(void *arg, uint64_t start, uint64_t size) +{ + trim_args_t *ta = arg; + vdev_t *vd = ta->trim_vdev; + range_seg_t logical_rs, physical_rs; + logical_rs.rs_start = start; + logical_rs.rs_end = start + size; + + /* + * Every range to be trimmed must be part of ms_allocatable. + * When ZFS_DEBUG_TRIM is set load the metaslab to verify this + * is always the case. + */ + if (zfs_flags & ZFS_DEBUG_TRIM) { + metaslab_t *msp = ta->trim_msp; + VERIFY0(metaslab_load(msp)); + VERIFY3B(msp->ms_loaded, ==, B_TRUE); + VERIFY(range_tree_find(msp->ms_allocatable, start, size)); + } + + ASSERT(vd->vdev_ops->vdev_op_leaf); + vdev_xlate(vd, &logical_rs, &physical_rs); + + IMPLY(vd->vdev_top == vd, + logical_rs.rs_start == physical_rs.rs_start); + IMPLY(vd->vdev_top == vd, + logical_rs.rs_end == physical_rs.rs_end); + + /* + * Only a manual trim will be traversing the vdev sequentially. + * For an auto trim all valid ranges should be added. + */ + if (ta->trim_type == TRIM_TYPE_MANUAL) { + + /* Only add segments that we have not visited yet */ + if (physical_rs.rs_end <= vd->vdev_trim_last_offset) + return; + + /* Pick up where we left off mid-range. */ + if (vd->vdev_trim_last_offset > physical_rs.rs_start) { + ASSERT3U(physical_rs.rs_end, >, + vd->vdev_trim_last_offset); + physical_rs.rs_start = vd->vdev_trim_last_offset; + } + } + + ASSERT3U(physical_rs.rs_end, >=, physical_rs.rs_start); + + /* + * With raidz, it's possible that the logical range does not live on + * this leaf vdev. We only add the physical range to this vdev's if it + * has a length greater than 0. + */ + if (physical_rs.rs_end > physical_rs.rs_start) { + range_tree_add(ta->trim_tree, physical_rs.rs_start, + physical_rs.rs_end - physical_rs.rs_start); + } else { + ASSERT3U(physical_rs.rs_end, ==, physical_rs.rs_start); + } +} + +/* + * Each manual TRIM thread is responsible for trimming the unallocated + * space for each leaf vdev. This is accomplished by sequentially iterating + * over its top-level metaslabs and issuing TRIM I/O for the space described + * by its ms_allocatable. While a metaslab is undergoing trimming it is + * not eligible for new allocations. + */ +static void +vdev_trim_thread(void *arg) +{ + vdev_t *vd = arg; + spa_t *spa = vd->vdev_spa; + trim_args_t ta; + int error = 0; + + /* + * The VDEV_LEAF_ZAP_TRIM_* entries may have been updated by + * vdev_trim(). Wait for the updated values to be reflected + * in the zap in order to start with the requested settings. + */ + txg_wait_synced(spa_get_dsl(vd->vdev_spa), 0); + + ASSERT(vdev_is_concrete(vd)); + spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); + + vd->vdev_trim_last_offset = 0; + vd->vdev_trim_rate = 0; + vd->vdev_trim_partial = 0; + vd->vdev_trim_secure = 0; + + VERIFY0(vdev_trim_load(vd)); + + ta.trim_vdev = vd; + ta.trim_extent_bytes_max = zfs_trim_extent_bytes_max; + ta.trim_extent_bytes_min = zfs_trim_extent_bytes_min; + ta.trim_tree = range_tree_create(NULL, NULL); + ta.trim_type = TRIM_TYPE_MANUAL; + ta.trim_flags = 0; + + /* + * When a secure TRIM has been requested infer that the intent + * is that everything must be trimmed. Override the default + * minimum TRIM size to prevent ranges from being skipped. + */ + if (vd->vdev_trim_secure) { + ta.trim_flags |= ZIO_TRIM_SECURE; + ta.trim_extent_bytes_min = SPA_MINBLOCKSIZE; + } + + uint64_t ms_count = 0; + for (uint64_t i = 0; !vd->vdev_detached && + i < vd->vdev_top->vdev_ms_count; i++) { + metaslab_t *msp = vd->vdev_top->vdev_ms[i]; + + /* + * If we've expanded the top-level vdev or it's our + * first pass, calculate our progress. + */ + if (vd->vdev_top->vdev_ms_count != ms_count) { + vdev_trim_calculate_progress(vd); + ms_count = vd->vdev_top->vdev_ms_count; + } + + spa_config_exit(spa, SCL_CONFIG, FTAG); + metaslab_disable(msp); + mutex_enter(&msp->ms_lock); + VERIFY0(metaslab_load(msp)); + + /* + * If a partial TRIM was requested skip metaslabs which have + * never been initialized and thus have never been written. + */ + if (msp->ms_sm == NULL && vd->vdev_trim_partial) { + mutex_exit(&msp->ms_lock); + metaslab_enable(msp, B_FALSE); + spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); + vdev_trim_calculate_progress(vd); + continue; + } + + ta.trim_msp = msp; + range_tree_walk(msp->ms_allocatable, vdev_trim_range_add, &ta); + range_tree_vacate(msp->ms_trim, NULL, NULL); + mutex_exit(&msp->ms_lock); + + error = vdev_trim_ranges(&ta); + metaslab_enable(msp, B_TRUE); + spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); + + range_tree_vacate(ta.trim_tree, NULL, NULL); + if (error != 0) + break; + } + + spa_config_exit(spa, SCL_CONFIG, FTAG); + mutex_enter(&vd->vdev_trim_io_lock); + while (vd->vdev_trim_inflight[0] > 0) { + cv_wait(&vd->vdev_trim_io_cv, &vd->vdev_trim_io_lock); + } + mutex_exit(&vd->vdev_trim_io_lock); + + range_tree_destroy(ta.trim_tree); + + mutex_enter(&vd->vdev_trim_lock); + if (!vd->vdev_trim_exit_wanted && vdev_writeable(vd)) { + vdev_trim_change_state(vd, VDEV_TRIM_COMPLETE, + vd->vdev_trim_rate, vd->vdev_trim_partial, + vd->vdev_trim_secure); + } + ASSERT(vd->vdev_trim_thread != NULL || vd->vdev_trim_inflight[0] == 0); + + /* + * Drop the vdev_trim_lock while we sync out the txg since it's + * possible that a device might be trying to come online and must + * check to see if it needs to restart a trim. That thread will be + * holding the spa_config_lock which would prevent the txg_wait_synced + * from completing. + */ + mutex_exit(&vd->vdev_trim_lock); + txg_wait_synced(spa_get_dsl(spa), 0); + mutex_enter(&vd->vdev_trim_lock); + + vd->vdev_trim_thread = NULL; + cv_broadcast(&vd->vdev_trim_cv); + mutex_exit(&vd->vdev_trim_lock); +} + +/* + * Initiates a manual TRIM for the vdev_t. Callers must hold vdev_trim_lock, + * the vdev_t must be a leaf and cannot already be manually trimming. + */ +void +vdev_trim(vdev_t *vd, uint64_t rate, boolean_t partial, boolean_t secure) +{ + ASSERT(MUTEX_HELD(&vd->vdev_trim_lock)); + ASSERT(vd->vdev_ops->vdev_op_leaf); + ASSERT(vdev_is_concrete(vd)); + ASSERT3P(vd->vdev_trim_thread, ==, NULL); + ASSERT(!vd->vdev_detached); + ASSERT(!vd->vdev_trim_exit_wanted); + ASSERT(!vd->vdev_top->vdev_removing); + + vdev_trim_change_state(vd, VDEV_TRIM_ACTIVE, rate, partial, secure); + vd->vdev_trim_thread = thread_create(NULL, 0, + vdev_trim_thread, vd, 0, &p0, TS_RUN, maxclsyspri); +} + +/* + * Wait for the trimming thread to be terminated (canceled or stopped). + */ +static void +vdev_trim_stop_wait_impl(vdev_t *vd) +{ + ASSERT(MUTEX_HELD(&vd->vdev_trim_lock)); + + while (vd->vdev_trim_thread != NULL) + cv_wait(&vd->vdev_trim_cv, &vd->vdev_trim_lock); + + ASSERT3P(vd->vdev_trim_thread, ==, NULL); + vd->vdev_trim_exit_wanted = B_FALSE; +} + +/* + * Wait for vdev trim threads which were listed to cleanly exit. + */ +void +vdev_trim_stop_wait(spa_t *spa, list_t *vd_list) +{ + vdev_t *vd; + + ASSERT(MUTEX_HELD(&spa_namespace_lock)); + + while ((vd = list_remove_head(vd_list)) != NULL) { + mutex_enter(&vd->vdev_trim_lock); + vdev_trim_stop_wait_impl(vd); + mutex_exit(&vd->vdev_trim_lock); + } +} + +/* + * Stop trimming a device, with the resultant trimming state being tgt_state. + * For blocking behavior pass NULL for vd_list. Otherwise, when a list_t is + * provided the stopping vdev is inserted in to the list. Callers are then + * required to call vdev_trim_stop_wait() to block for all the trim threads + * to exit. The caller must hold vdev_trim_lock and must not be writing to + * the spa config, as the trimming thread may try to enter the config as a + * reader before exiting. + */ +void +vdev_trim_stop(vdev_t *vd, vdev_trim_state_t tgt_state, list_t *vd_list) +{ + ASSERT(!spa_config_held(vd->vdev_spa, SCL_CONFIG|SCL_STATE, RW_WRITER)); + ASSERT(MUTEX_HELD(&vd->vdev_trim_lock)); + ASSERT(vd->vdev_ops->vdev_op_leaf); + ASSERT(vdev_is_concrete(vd)); + + /* + * Allow cancel requests to proceed even if the trim thread has + * stopped. + */ + if (vd->vdev_trim_thread == NULL && tgt_state != VDEV_TRIM_CANCELED) + return; + + vdev_trim_change_state(vd, tgt_state, 0, 0, 0); + vd->vdev_trim_exit_wanted = B_TRUE; + + if (vd_list == NULL) { + vdev_trim_stop_wait_impl(vd); + } else { + ASSERT(MUTEX_HELD(&spa_namespace_lock)); + list_insert_tail(vd_list, vd); + } +} + +/* + * Requests that all listed vdevs stop trimming. + */ +static void +vdev_trim_stop_all_impl(vdev_t *vd, vdev_trim_state_t tgt_state, + list_t *vd_list) +{ + if (vd->vdev_ops->vdev_op_leaf && vdev_is_concrete(vd)) { + mutex_enter(&vd->vdev_trim_lock); + vdev_trim_stop(vd, tgt_state, vd_list); + mutex_exit(&vd->vdev_trim_lock); + return; + } + + for (uint64_t i = 0; i < vd->vdev_children; i++) { + vdev_trim_stop_all_impl(vd->vdev_child[i], tgt_state, + vd_list); + } +} + +/* + * Convenience function to stop trimming of a vdev tree and set all trim + * thread pointers to NULL. + */ +void +vdev_trim_stop_all(vdev_t *vd, vdev_trim_state_t tgt_state) +{ + spa_t *spa = vd->vdev_spa; + list_t vd_list; + + ASSERT(MUTEX_HELD(&spa_namespace_lock)); + + list_create(&vd_list, sizeof (vdev_t), + offsetof(vdev_t, vdev_trim_node)); + + vdev_trim_stop_all_impl(vd, tgt_state, &vd_list); + vdev_trim_stop_wait(spa, &vd_list); + + if (vd->vdev_spa->spa_sync_on) { + /* Make sure that our state has been synced to disk */ + txg_wait_synced(spa_get_dsl(vd->vdev_spa), 0); + } + + list_destroy(&vd_list); +} + +/* + * Conditionally restarts a manual TRIM given its on-disk state. + */ +void +vdev_trim_restart(vdev_t *vd) +{ + ASSERT(MUTEX_HELD(&spa_namespace_lock)); + ASSERT(!spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER)); + + if (vd->vdev_leaf_zap != 0) { + mutex_enter(&vd->vdev_trim_lock); + uint64_t trim_state = VDEV_TRIM_NONE; + int err = zap_lookup(vd->vdev_spa->spa_meta_objset, + vd->vdev_leaf_zap, VDEV_LEAF_ZAP_TRIM_STATE, + sizeof (trim_state), 1, &trim_state); + ASSERT(err == 0 || err == ENOENT); + vd->vdev_trim_state = trim_state; + + uint64_t timestamp = 0; + err = zap_lookup(vd->vdev_spa->spa_meta_objset, + vd->vdev_leaf_zap, VDEV_LEAF_ZAP_TRIM_ACTION_TIME, + sizeof (timestamp), 1, ×tamp); + ASSERT(err == 0 || err == ENOENT); + vd->vdev_trim_action_time = (time_t)timestamp; + + if (vd->vdev_trim_state == VDEV_TRIM_SUSPENDED || + vd->vdev_offline) { + /* load progress for reporting, but don't resume */ + VERIFY0(vdev_trim_load(vd)); + } else if (vd->vdev_trim_state == VDEV_TRIM_ACTIVE && + vdev_writeable(vd) && !vd->vdev_top->vdev_removing && + vd->vdev_trim_thread == NULL) { + VERIFY0(vdev_trim_load(vd)); + vdev_trim(vd, vd->vdev_trim_rate, + vd->vdev_trim_partial, vd->vdev_trim_secure); + } + + mutex_exit(&vd->vdev_trim_lock); + } + + for (uint64_t i = 0; i < vd->vdev_children; i++) { + vdev_trim_restart(vd->vdev_child[i]); + } +} + +/* + * Used by the automatic TRIM when ZFS_DEBUG_TRIM is set to verify that + * every TRIM range is contained within ms_allocatable. + */ +static void +vdev_trim_range_verify(void *arg, uint64_t start, uint64_t size) +{ + trim_args_t *ta = arg; + metaslab_t *msp = ta->trim_msp; + + VERIFY3B(msp->ms_loaded, ==, B_TRUE); + VERIFY3U(msp->ms_disabled, >, 0); + VERIFY(range_tree_find(msp->ms_allocatable, start, size) != NULL); +} + +/* + * Each automatic TRIM thread is responsible for managing the trimming of a + * top-level vdev in the pool. No automatic TRIM state is maintained on-disk. + * + * N.B. This behavior is different from a manual TRIM where a thread + * is created for each leaf vdev, instead of each top-level vdev. + */ +static void +vdev_autotrim_thread(void *arg) +{ + vdev_t *vd = arg; + spa_t *spa = vd->vdev_spa; + int shift = 0; + + mutex_enter(&vd->vdev_autotrim_lock); + ASSERT3P(vd->vdev_top, ==, vd); + ASSERT3P(vd->vdev_autotrim_thread, !=, NULL); + mutex_exit(&vd->vdev_autotrim_lock); + spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); + + uint64_t extent_bytes_max = zfs_trim_extent_bytes_max; + uint64_t extent_bytes_min = zfs_trim_extent_bytes_min; + + while (!vdev_autotrim_should_stop(vd)) { + int txgs_per_trim = MAX(zfs_trim_txg_batch, 1); + boolean_t issued_trim = B_FALSE; + + /* + * All of the metaslabs are divided in to groups of size + * num_metaslabs / zfs_trim_txg_batch. Each of these groups + * is composed of metaslabs which are spread evenly over the + * device. + * + * For example, when zfs_trim_txg_batch = 32 (default) then + * group 0 will contain metaslabs 0, 32, 64, ...; + * group 1 will contain metaslabs 1, 33, 65, ...; + * group 2 will contain metaslabs 2, 34, 66, ...; and so on. + * + * On each pass through the while() loop one of these groups + * is selected. This is accomplished by using a shift value + * to select the starting metaslab, then striding over the + * metaslabs using the zfs_trim_txg_batch size. This is + * done to accomplish two things. + * + * 1) By dividing the metaslabs in to groups, and making sure + * that each group takes a minimum of one txg to process. + * Then zfs_trim_txg_batch controls the minimum number of + * txgs which must occur before a metaslab is revisited. + * + * 2) Selecting non-consecutive metaslabs distributes the + * TRIM commands for a group evenly over the entire device. + * This can be advantageous for certain types of devices. + */ + for (uint64_t i = shift % txgs_per_trim; i < vd->vdev_ms_count; + i += txgs_per_trim) { + metaslab_t *msp = vd->vdev_ms[i]; + range_tree_t *trim_tree; + + spa_config_exit(spa, SCL_CONFIG, FTAG); + metaslab_disable(msp); + spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); + + mutex_enter(&msp->ms_lock); + + /* + * Skip the metaslab when it has never been allocated + * or when there are no recent frees to trim. + */ + if (msp->ms_sm == NULL || + range_tree_is_empty(msp->ms_trim)) { + mutex_exit(&msp->ms_lock); + metaslab_enable(msp, B_FALSE); + continue; + } + + /* + * Skip the metaslab when it has already been disabled. + * This may happen when a manual TRIM or initialize + * operation is running concurrently. In the case + * of a manual TRIM, the ms_trim tree will have been + * vacated. Only ranges added after the manual TRIM + * disabled the metaslab will be included in the tree. + * These will be processed when the automatic TRIM + * next revisits this metaslab. + */ + if (msp->ms_disabled > 1) { + mutex_exit(&msp->ms_lock); + metaslab_enable(msp, B_FALSE); + continue; + } + + /* + * Allocate an empty range tree which is swapped in + * for the existing ms_trim tree while it is processed. + */ + trim_tree = range_tree_create(NULL, NULL); + range_tree_swap(&msp->ms_trim, &trim_tree); + ASSERT(range_tree_is_empty(msp->ms_trim)); + + /* + * There are two cases when constructing the per-vdev + * trim trees for a metaslab. If the top-level vdev + * has no children then it is also a leaf and should + * be trimmed. Otherwise our children are the leaves + * and a trim tree should be constructed for each. + */ + trim_args_t *tap; + uint64_t children = vd->vdev_children; + if (children == 0) { + children = 1; + tap = kmem_zalloc(sizeof (trim_args_t) * + children, KM_SLEEP); + tap[0].trim_vdev = vd; + } else { + tap = kmem_zalloc(sizeof (trim_args_t) * + children, KM_SLEEP); + + for (uint64_t c = 0; c < children; c++) { + tap[c].trim_vdev = vd->vdev_child[c]; + } + } + + for (uint64_t c = 0; c < children; c++) { + trim_args_t *ta = &tap[c]; + vdev_t *cvd = ta->trim_vdev; + + ta->trim_msp = msp; + ta->trim_extent_bytes_max = extent_bytes_max; + ta->trim_extent_bytes_min = extent_bytes_min; + ta->trim_type = TRIM_TYPE_AUTO; + ta->trim_flags = 0; + + if (cvd->vdev_detached || + !vdev_writeable(cvd) || + !cvd->vdev_has_trim || + cvd->vdev_trim_thread != NULL) { + continue; + } + + /* + * When a device has an attached hot spare, or + * is being replaced it will not be trimmed. + * This is done to avoid adding additional + * stress to a potentially unhealthy device, + * and to minimize the required rebuild time. + */ + if (!cvd->vdev_ops->vdev_op_leaf) + continue; + + ta->trim_tree = range_tree_create(NULL, NULL); + range_tree_walk(trim_tree, + vdev_trim_range_add, ta); + } + + mutex_exit(&msp->ms_lock); + spa_config_exit(spa, SCL_CONFIG, FTAG); + + /* + * Issue the TRIM I/Os for all ranges covered by the + * TRIM trees. These ranges are safe to TRIM because + * no new allocations will be performed until the call + * to metaslab_enabled() below. + */ + for (uint64_t c = 0; c < children; c++) { + trim_args_t *ta = &tap[c]; + + /* + * Always yield to a manual TRIM if one has + * been started for the child vdev. + */ + if (ta->trim_tree == NULL || + ta->trim_vdev->vdev_trim_thread != NULL) { + continue; + } + + /* + * After this point metaslab_enable() must be + * called with the sync flag set. This is done + * here because vdev_trim_ranges() is allowed + * to be interrupted (EINTR) before issuing all + * of the required TRIM I/Os. + */ + issued_trim = B_TRUE; + + int error = vdev_trim_ranges(ta); + if (error) + break; + } + + /* + * Verify every range which was trimmed is still + * contained within the ms_allocatable tree. + */ + if (zfs_flags & ZFS_DEBUG_TRIM) { + mutex_enter(&msp->ms_lock); + VERIFY0(metaslab_load(msp)); + VERIFY3P(tap[0].trim_msp, ==, msp); + range_tree_walk(trim_tree, + vdev_trim_range_verify, &tap[0]); + mutex_exit(&msp->ms_lock); + } + + range_tree_vacate(trim_tree, NULL, NULL); + range_tree_destroy(trim_tree); + + metaslab_enable(msp, issued_trim); + spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); + + for (uint64_t c = 0; c < children; c++) { + trim_args_t *ta = &tap[c]; + + if (ta->trim_tree == NULL) + continue; + + range_tree_vacate(ta->trim_tree, NULL, NULL); + range_tree_destroy(ta->trim_tree); + } + + kmem_free(tap, sizeof (trim_args_t) * children); + } + + spa_config_exit(spa, SCL_CONFIG, FTAG); + + /* + * After completing the group of metaslabs wait for the next + * open txg. This is done to make sure that a minimum of + * zfs_trim_txg_batch txgs will occur before these metaslabs + * are trimmed again. + */ + txg_wait_open(spa_get_dsl(spa), 0, issued_trim); + + shift++; + spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); + } + + for (uint64_t c = 0; c < vd->vdev_children; c++) { + vdev_t *cvd = vd->vdev_child[c]; + mutex_enter(&cvd->vdev_trim_io_lock); + + while (cvd->vdev_trim_inflight[1] > 0) { + cv_wait(&cvd->vdev_trim_io_cv, + &cvd->vdev_trim_io_lock); + } + mutex_exit(&cvd->vdev_trim_io_lock); + } + + spa_config_exit(spa, SCL_CONFIG, FTAG); + + /* + * When exiting because the autotrim property was set to off, then + * abandon any unprocessed ms_trim ranges to reclaim the memory. + */ + if (spa_get_autotrim(spa) == SPA_AUTOTRIM_OFF) { + for (uint64_t i = 0; i < vd->vdev_ms_count; i++) { + metaslab_t *msp = vd->vdev_ms[i]; + + mutex_enter(&msp->ms_lock); + range_tree_vacate(msp->ms_trim, NULL, NULL); + mutex_exit(&msp->ms_lock); + } + } + + mutex_enter(&vd->vdev_autotrim_lock); + ASSERT(vd->vdev_autotrim_thread != NULL); + vd->vdev_autotrim_thread = NULL; + cv_broadcast(&vd->vdev_autotrim_cv); + mutex_exit(&vd->vdev_autotrim_lock); +} + +/* + * Starts an autotrim thread, if needed, for each top-level vdev which can be + * trimmed. A top-level vdev which has been evacuated will never be trimmed. + */ +void +vdev_autotrim(spa_t *spa) +{ + vdev_t *root_vd = spa->spa_root_vdev; + + for (uint64_t i = 0; i < root_vd->vdev_children; i++) { + vdev_t *tvd = root_vd->vdev_child[i]; + + mutex_enter(&tvd->vdev_autotrim_lock); + if (vdev_writeable(tvd) && !tvd->vdev_removing && + tvd->vdev_autotrim_thread == NULL) { + ASSERT3P(tvd->vdev_top, ==, tvd); + + tvd->vdev_autotrim_thread = thread_create(NULL, 0, + vdev_autotrim_thread, tvd, 0, &p0, TS_RUN, + maxclsyspri); + ASSERT(tvd->vdev_autotrim_thread != NULL); + } + mutex_exit(&tvd->vdev_autotrim_lock); + } +} + +/* + * Wait for the vdev_autotrim_thread associated with the passed top-level + * vdev to be terminated (canceled or stopped). + */ +void +vdev_autotrim_stop_wait(vdev_t *tvd) +{ + mutex_enter(&tvd->vdev_autotrim_lock); + if (tvd->vdev_autotrim_thread != NULL) { + tvd->vdev_autotrim_exit_wanted = B_TRUE; + + while (tvd->vdev_autotrim_thread != NULL) { + cv_wait(&tvd->vdev_autotrim_cv, + &tvd->vdev_autotrim_lock); + } + + ASSERT3P(tvd->vdev_autotrim_thread, ==, NULL); + tvd->vdev_autotrim_exit_wanted = B_FALSE; + } + mutex_exit(&tvd->vdev_autotrim_lock); +} + +/* + * Wait for all of the vdev_autotrim_thread associated with the pool to + * be terminated (canceled or stopped). + */ +void +vdev_autotrim_stop_all(spa_t *spa) +{ + vdev_t *root_vd = spa->spa_root_vdev; + + for (uint64_t i = 0; i < root_vd->vdev_children; i++) + vdev_autotrim_stop_wait(root_vd->vdev_child[i]); +} + +/* + * Conditionally restart all of the vdev_autotrim_thread's for the pool. + */ +void +vdev_autotrim_restart(spa_t *spa) +{ + ASSERT(MUTEX_HELD(&spa_namespace_lock)); + + if (spa->spa_autotrim) + vdev_autotrim(spa); +} + +#if defined(_KERNEL) +EXPORT_SYMBOL(vdev_trim); +EXPORT_SYMBOL(vdev_trim_stop); +EXPORT_SYMBOL(vdev_trim_stop_all); +EXPORT_SYMBOL(vdev_trim_stop_wait); +EXPORT_SYMBOL(vdev_trim_restart); +EXPORT_SYMBOL(vdev_autotrim); +EXPORT_SYMBOL(vdev_autotrim_stop_all); +EXPORT_SYMBOL(vdev_autotrim_stop_wait); +EXPORT_SYMBOL(vdev_autotrim_restart); + +/* BEGIN CSTYLED */ +module_param(zfs_trim_extent_bytes_max, uint, 0644); +MODULE_PARM_DESC(zfs_trim_extent_bytes_max, + "Max size of TRIM commands, larger will be split"); + +module_param(zfs_trim_extent_bytes_min, uint, 0644); +MODULE_PARM_DESC(zfs_trim_extent_bytes_min, + "Min size of TRIM commands, smaller will be skipped"); + +module_param(zfs_trim_metaslab_skip, uint, 0644); +MODULE_PARM_DESC(zfs_trim_metaslab_skip, + "Skip metaslabs which have never been initialized"); + +module_param(zfs_trim_txg_batch, uint, 0644); +MODULE_PARM_DESC(zfs_trim_txg_batch, + "Min number of txgs to aggregate frees before issuing TRIM"); + +module_param(zfs_trim_queue_limit, uint, 0644); +MODULE_PARM_DESC(zfs_trim_queue_limit, + "Max queued TRIMs outstanding per leaf vdev"); +/* END CSTYLED */ +#endif diff --git a/module/zfs/zfs_ioctl.c b/module/zfs/zfs_ioctl.c index 047193c61..debe733da 100644 --- a/module/zfs/zfs_ioctl.c +++ b/module/zfs/zfs_ioctl.c @@ -204,6 +204,7 @@ #include <sys/zfs_sysfs.h> #include <sys/vdev_impl.h> #include <sys/vdev_initialize.h> +#include <sys/vdev_trim.h> #include <linux/miscdevice.h> #include <linux/slab.h> @@ -3885,7 +3886,7 @@ zfs_ioc_destroy(zfs_cmd_t *zc) /* * innvl: { - * "initialize_command" -> POOL_INITIALIZE_{CANCEL|DO|SUSPEND} (uint64) + * "initialize_command" -> POOL_INITIALIZE_{CANCEL|START|SUSPEND} (uint64) * "initialize_vdevs": { -> guids to initialize (nvlist) * "vdev_path_1": vdev_guid_1, (uint64), * "vdev_path_2": vdev_guid_2, (uint64), @@ -3919,7 +3920,7 @@ zfs_ioc_pool_initialize(const char *poolname, nvlist_t *innvl, nvlist_t *outnvl) } if (!(cmd_type == POOL_INITIALIZE_CANCEL || - cmd_type == POOL_INITIALIZE_DO || + cmd_type == POOL_INITIALIZE_START || cmd_type == POOL_INITIALIZE_SUSPEND)) { return (SET_ERROR(EINVAL)); } @@ -3958,6 +3959,91 @@ zfs_ioc_pool_initialize(const char *poolname, nvlist_t *innvl, nvlist_t *outnvl) } /* + * innvl: { + * "trim_command" -> POOL_TRIM_{CANCEL|START|SUSPEND} (uint64) + * "trim_vdevs": { -> guids to TRIM (nvlist) + * "vdev_path_1": vdev_guid_1, (uint64), + * "vdev_path_2": vdev_guid_2, (uint64), + * ... + * }, + * "trim_rate" -> Target TRIM rate in bytes/sec. + * "trim_secure" -> Set to request a secure TRIM. + * } + * + * outnvl: { + * "trim_vdevs": { -> TRIM errors (nvlist) + * "vdev_path_1": errno, see function body for possible errnos (uint64) + * "vdev_path_2": errno, ... (uint64) + * ... + * } + * } + * + * EINVAL is returned for an unknown commands or if any of the provided vdev + * guids have be specified with a type other than uint64. + */ +static const zfs_ioc_key_t zfs_keys_pool_trim[] = { + {ZPOOL_TRIM_COMMAND, DATA_TYPE_UINT64, 0}, + {ZPOOL_TRIM_VDEVS, DATA_TYPE_NVLIST, 0}, + {ZPOOL_TRIM_RATE, DATA_TYPE_UINT64, ZK_OPTIONAL}, + {ZPOOL_TRIM_SECURE, DATA_TYPE_BOOLEAN_VALUE, ZK_OPTIONAL}, +}; + +static int +zfs_ioc_pool_trim(const char *poolname, nvlist_t *innvl, nvlist_t *outnvl) +{ + uint64_t cmd_type; + if (nvlist_lookup_uint64(innvl, ZPOOL_TRIM_COMMAND, &cmd_type) != 0) + return (SET_ERROR(EINVAL)); + + if (!(cmd_type == POOL_TRIM_CANCEL || + cmd_type == POOL_TRIM_START || + cmd_type == POOL_TRIM_SUSPEND)) { + return (SET_ERROR(EINVAL)); + } + + nvlist_t *vdev_guids; + if (nvlist_lookup_nvlist(innvl, ZPOOL_TRIM_VDEVS, &vdev_guids) != 0) + return (SET_ERROR(EINVAL)); + + for (nvpair_t *pair = nvlist_next_nvpair(vdev_guids, NULL); + pair != NULL; pair = nvlist_next_nvpair(vdev_guids, pair)) { + uint64_t vdev_guid; + if (nvpair_value_uint64(pair, &vdev_guid) != 0) { + return (SET_ERROR(EINVAL)); + } + } + + /* Optional, defaults to maximum rate when not provided */ + uint64_t rate; + if (nvlist_lookup_uint64(innvl, ZPOOL_TRIM_RATE, &rate) != 0) + rate = 0; + + /* Optional, defaults to standard TRIM when not provided */ + boolean_t secure; + if (nvlist_lookup_boolean_value(innvl, ZPOOL_TRIM_SECURE, + &secure) != 0) { + secure = B_FALSE; + } + + spa_t *spa; + int error = spa_open(poolname, &spa, FTAG); + if (error != 0) + return (error); + + nvlist_t *vdev_errlist = fnvlist_alloc(); + int total_errors = spa_vdev_trim(spa, vdev_guids, cmd_type, + rate, !!zfs_trim_metaslab_skip, secure, vdev_errlist); + + if (fnvlist_size(vdev_errlist) > 0) + fnvlist_add_nvlist(outnvl, ZPOOL_TRIM_VDEVS, vdev_errlist); + + fnvlist_free(vdev_errlist); + + spa_close(spa, FTAG); + return (total_errors > 0 ? EINVAL : 0); +} + +/* * fsname is name of dataset to rollback (to most recent snapshot) * * innvl may contain name of expected target snapshot @@ -6580,6 +6666,11 @@ zfs_ioctl_init(void) POOL_CHECK_SUSPENDED | POOL_CHECK_READONLY, B_TRUE, B_TRUE, zfs_keys_pool_initialize, ARRAY_SIZE(zfs_keys_pool_initialize)); + zfs_ioctl_register("trim", ZFS_IOC_POOL_TRIM, + zfs_ioc_pool_trim, zfs_secpolicy_config, POOL_NAME, + POOL_CHECK_SUSPENDED | POOL_CHECK_READONLY, B_TRUE, B_TRUE, + zfs_keys_pool_trim, ARRAY_SIZE(zfs_keys_pool_trim)); + /* IOCTLS that use the legacy function signature */ zfs_ioctl_register_legacy(ZFS_IOC_POOL_FREEZE, zfs_ioc_pool_freeze, diff --git a/module/zfs/zfs_sysfs.c b/module/zfs/zfs_sysfs.c index 87c4ac117..ec8ae4216 100644 --- a/module/zfs/zfs_sysfs.c +++ b/module/zfs/zfs_sysfs.c @@ -358,7 +358,8 @@ pool_property_show(struct kobject *kobj, struct attribute *attr, char *buf) */ static const char *zfs_features[] = { /* --> Add new kernel features here (post ZoL 0.8.0) */ - "vdev_initialize" + "initialize", + "trim", }; #define ZFS_FEATURE_COUNT ARRAY_SIZE(zfs_features) diff --git a/module/zfs/zio.c b/module/zfs/zio.c index 0912f607f..1915de417 100644 --- a/module/zfs/zio.c +++ b/module/zfs/zio.c @@ -32,6 +32,7 @@ #include <sys/txg.h> #include <sys/spa_impl.h> #include <sys/vdev_impl.h> +#include <sys/vdev_trim.h> #include <sys/zio_impl.h> #include <sys/zio_compress.h> #include <sys/zio_checksum.h> @@ -58,7 +59,7 @@ const char *zio_type_name[ZIO_TYPES] = { * Note: Linux kernel thread name length is limited * so these names will differ from upstream open zfs. */ - "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl" + "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl", "z_trim" }; int zio_dva_throttle_enabled = B_TRUE; @@ -761,7 +762,7 @@ zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, { zio_t *zio; - ASSERT3U(psize, <=, SPA_MAXBLOCKSIZE); + IMPLY(type != ZIO_TYPE_TRIM, psize <= SPA_MAXBLOCKSIZE); ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0); ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); @@ -1212,6 +1213,26 @@ zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, } zio_t * +zio_trim(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, + zio_done_func_t *done, void *private, zio_priority_t priority, + enum zio_flag flags, enum trim_flag trim_flags) +{ + zio_t *zio; + + ASSERT0(vd->vdev_children); + ASSERT0(P2PHASE(offset, 1ULL << vd->vdev_ashift)); + ASSERT0(P2PHASE(size, 1ULL << vd->vdev_ashift)); + ASSERT3U(size, !=, 0); + + zio = zio_create(pio, vd->vdev_spa, 0, NULL, NULL, size, size, done, + private, ZIO_TYPE_TRIM, priority, flags | ZIO_FLAG_PHYSICAL, + vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_TRIM_PIPELINE); + zio->io_trim_flags = trim_flags; + + return (zio); +} + +zio_t * zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, abd_t *data, int checksum, zio_done_func_t *done, void *private, zio_priority_t priority, enum zio_flag flags, boolean_t labels) @@ -3562,7 +3583,6 @@ zio_alloc_zil(spa_t *spa, objset_t *os, uint64_t txg, blkptr_t *new_bp, * ========================================================================== */ - /* * Issue an I/O to the underlying vdev. Typically the issue pipeline * stops after this stage and will resume upon I/O completion. @@ -3685,8 +3705,8 @@ zio_vdev_io_start(zio_t *zio) return (zio); } - if (vd->vdev_ops->vdev_op_leaf && - (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) { + if (vd->vdev_ops->vdev_op_leaf && (zio->io_type == ZIO_TYPE_READ || + zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_TRIM)) { if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio)) return (zio); @@ -3717,7 +3737,8 @@ zio_vdev_io_done(zio_t *zio) return (NULL); } - ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE); + ASSERT(zio->io_type == ZIO_TYPE_READ || + zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_TRIM); if (zio->io_delay) zio->io_delay = gethrtime() - zio->io_delay; @@ -3736,7 +3757,7 @@ zio_vdev_io_done(zio_t *zio) if (zio_injection_enabled && zio->io_error == 0) zio->io_error = zio_handle_label_injection(zio, EIO); - if (zio->io_error) { + if (zio->io_error && zio->io_type != ZIO_TYPE_TRIM) { if (!vdev_accessible(vd, zio)) { zio->io_error = SET_ERROR(ENXIO); } else { @@ -3866,8 +3887,8 @@ zio_vdev_io_assess(zio_t *zio) /* * If a cache flush returns ENOTSUP or ENOTTY, we know that no future - * attempts will ever succeed. In this case we set a persistent bit so - * that we don't bother with it in the future. + * attempts will ever succeed. In this case we set a persistent + * boolean flag so that we don't bother with it in the future. */ if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) && zio->io_type == ZIO_TYPE_IOCTL && |