4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or https://opensource.org/licenses/CDDL-1.0.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
23 * Copyright (c) 2016 by Delphix. All rights reserved.
24 * Copyright (c) 2019 by Lawrence Livermore National Security, LLC.
25 * Copyright (c) 2021 Hewlett Packard Enterprise Development LP
26 * Copyright 2023 RackTop Systems, Inc.
30 #include <sys/spa_impl.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/vdev_trim.h>
34 #include <sys/metaslab_impl.h>
35 #include <sys/dsl_synctask.h>
37 #include <sys/dmu_tx.h>
38 #include <sys/arc_impl.h>
41 * TRIM is a feature which is used to notify a SSD that some previously
42 * written space is no longer allocated by the pool. This is useful because
43 * writes to a SSD must be performed to blocks which have first been erased.
44 * Ensuring the SSD always has a supply of erased blocks for new writes
45 * helps prevent the performance from deteriorating.
47 * There are two supported TRIM methods; manual and automatic.
51 * A manual TRIM is initiated by running the 'zpool trim' command. A single
52 * 'vdev_trim' thread is created for each leaf vdev, and it is responsible for
53 * managing that vdev TRIM process. This involves iterating over all the
54 * metaslabs, calculating the unallocated space ranges, and then issuing the
57 * While a metaslab is being actively trimmed it is not eligible to perform
58 * new allocations. After traversing all of the metaslabs the thread is
59 * terminated. Finally, both the requested options and current progress of
60 * the TRIM are regularly written to the pool. This allows the TRIM to be
61 * suspended and resumed as needed.
65 * An automatic TRIM is enabled by setting the 'autotrim' pool property
66 * to 'on'. When enabled, a `vdev_autotrim' thread is created for each
67 * top-level (not leaf) vdev in the pool. These threads perform the same
68 * core TRIM process as a manual TRIM, but with a few key differences.
70 * 1) Automatic TRIM happens continuously in the background and operates
71 * solely on recently freed blocks (ms_trim not ms_allocatable).
73 * 2) Each thread is associated with a top-level (not leaf) vdev. This has
74 * the benefit of simplifying the threading model, it makes it easier
75 * to coordinate administrative commands, and it ensures only a single
76 * metaslab is disabled at a time. Unlike manual TRIM, this means each
77 * 'vdev_autotrim' thread is responsible for issuing TRIM I/Os for its
80 * 3) There is no automatic TRIM progress information stored on disk, nor
81 * is it reported by 'zpool status'.
83 * While the automatic TRIM process is highly effective it is more likely
84 * than a manual TRIM to encounter tiny ranges. Ranges less than or equal to
85 * 'zfs_trim_extent_bytes_min' (32k) are considered too small to efficiently
86 * TRIM and are skipped. This means small amounts of freed space may not
87 * be automatically trimmed.
89 * Furthermore, devices with attached hot spares and devices being actively
90 * replaced are skipped. This is done to avoid adding additional stress to
91 * a potentially unhealthy device and to minimize the required rebuild time.
93 * For this reason it may be beneficial to occasionally manually TRIM a pool
94 * even when automatic TRIM is enabled.
98 * Maximum size of TRIM I/O, ranges will be chunked in to 128MiB lengths.
100 static unsigned int zfs_trim_extent_bytes_max = 128 * 1024 * 1024;
103 * Minimum size of TRIM I/O, extents smaller than 32Kib will be skipped.
105 static unsigned int zfs_trim_extent_bytes_min = 32 * 1024;
108 * Skip uninitialized metaslabs during the TRIM process. This option is
109 * useful for pools constructed from large thinly-provisioned devices where
110 * TRIM operations are slow. As a pool ages an increasing fraction of
111 * the pools metaslabs will be initialized progressively degrading the
112 * usefulness of this option. This setting is stored when starting a
113 * manual TRIM and will persist for the duration of the requested TRIM.
115 unsigned int zfs_trim_metaslab_skip = 0;
118 * Maximum number of queued TRIM I/Os per leaf vdev. The number of
119 * concurrent TRIM I/Os issued to the device is controlled by the
120 * zfs_vdev_trim_min_active and zfs_vdev_trim_max_active module options.
122 static unsigned int zfs_trim_queue_limit = 10;
125 * The minimum number of transaction groups between automatic trims of a
126 * metaslab. This setting represents a trade-off between issuing more
127 * efficient TRIM operations, by allowing them to be aggregated longer,
128 * and issuing them promptly so the trimmed space is available. Note
129 * that this value is a minimum; metaslabs can be trimmed less frequently
130 * when there are a large number of ranges which need to be trimmed.
132 * Increasing this value will allow frees to be aggregated for a longer
133 * time. This can result is larger TRIM operations, and increased memory
134 * usage in order to track the ranges to be trimmed. Decreasing this value
135 * has the opposite effect. The default value of 32 was determined though
136 * testing to be a reasonable compromise.
138 static unsigned int zfs_trim_txg_batch = 32;
141 * The trim_args are a control structure which describe how a leaf vdev
142 * should be trimmed. The core elements are the vdev, the metaslab being
143 * trimmed and a range tree containing the extents to TRIM. All provided
144 * ranges must be within the metaslab.
146 typedef struct trim_args {
148 * These fields are set by the caller of vdev_trim_ranges().
150 vdev_t *trim_vdev; /* Leaf vdev to TRIM */
151 metaslab_t *trim_msp; /* Disabled metaslab */
152 range_tree_t *trim_tree; /* TRIM ranges (in metaslab) */
153 trim_type_t trim_type; /* Manual or auto TRIM */
154 uint64_t trim_extent_bytes_max; /* Maximum TRIM I/O size */
155 uint64_t trim_extent_bytes_min; /* Minimum TRIM I/O size */
156 enum trim_flag trim_flags; /* TRIM flags (secure) */
159 * These fields are updated by vdev_trim_ranges().
161 hrtime_t trim_start_time; /* Start time */
162 uint64_t trim_bytes_done; /* Bytes trimmed */
166 * Determines whether a vdev_trim_thread() should be stopped.
169 vdev_trim_should_stop(vdev_t *vd)
171 return (vd->vdev_trim_exit_wanted || !vdev_writeable(vd) ||
172 vd->vdev_detached || vd->vdev_top->vdev_removing);
176 * Determines whether a vdev_autotrim_thread() should be stopped.
179 vdev_autotrim_should_stop(vdev_t *tvd)
181 return (tvd->vdev_autotrim_exit_wanted ||
182 !vdev_writeable(tvd) || tvd->vdev_removing ||
183 spa_get_autotrim(tvd->vdev_spa) == SPA_AUTOTRIM_OFF);
187 * Wait for given number of kicks, return true if the wait is aborted due to
188 * vdev_autotrim_exit_wanted.
191 vdev_autotrim_wait_kick(vdev_t *vd, int num_of_kick)
193 mutex_enter(&vd->vdev_autotrim_lock);
194 for (int i = 0; i < num_of_kick; i++) {
195 if (vd->vdev_autotrim_exit_wanted)
197 cv_wait(&vd->vdev_autotrim_kick_cv, &vd->vdev_autotrim_lock);
199 boolean_t exit_wanted = vd->vdev_autotrim_exit_wanted;
200 mutex_exit(&vd->vdev_autotrim_lock);
202 return (exit_wanted);
206 * The sync task for updating the on-disk state of a manual TRIM. This
207 * is scheduled by vdev_trim_change_state().
210 vdev_trim_zap_update_sync(void *arg, dmu_tx_t *tx)
213 * We pass in the guid instead of the vdev_t since the vdev may
214 * have been freed prior to the sync task being processed. This
215 * happens when a vdev is detached as we call spa_config_vdev_exit(),
216 * stop the trimming thread, schedule the sync task, and free
217 * the vdev. Later when the scheduled sync task is invoked, it would
218 * find that the vdev has been freed.
220 uint64_t guid = *(uint64_t *)arg;
221 uint64_t txg = dmu_tx_get_txg(tx);
222 kmem_free(arg, sizeof (uint64_t));
224 vdev_t *vd = spa_lookup_by_guid(tx->tx_pool->dp_spa, guid, B_FALSE);
225 if (vd == NULL || vd->vdev_top->vdev_removing || !vdev_is_concrete(vd))
228 uint64_t last_offset = vd->vdev_trim_offset[txg & TXG_MASK];
229 vd->vdev_trim_offset[txg & TXG_MASK] = 0;
231 VERIFY3U(vd->vdev_leaf_zap, !=, 0);
233 objset_t *mos = vd->vdev_spa->spa_meta_objset;
235 if (last_offset > 0 || vd->vdev_trim_last_offset == UINT64_MAX) {
237 if (vd->vdev_trim_last_offset == UINT64_MAX)
240 vd->vdev_trim_last_offset = last_offset;
241 VERIFY0(zap_update(mos, vd->vdev_leaf_zap,
242 VDEV_LEAF_ZAP_TRIM_LAST_OFFSET,
243 sizeof (last_offset), 1, &last_offset, tx));
246 if (vd->vdev_trim_action_time > 0) {
247 uint64_t val = (uint64_t)vd->vdev_trim_action_time;
248 VERIFY0(zap_update(mos, vd->vdev_leaf_zap,
249 VDEV_LEAF_ZAP_TRIM_ACTION_TIME, sizeof (val),
253 if (vd->vdev_trim_rate > 0) {
254 uint64_t rate = (uint64_t)vd->vdev_trim_rate;
256 if (rate == UINT64_MAX)
259 VERIFY0(zap_update(mos, vd->vdev_leaf_zap,
260 VDEV_LEAF_ZAP_TRIM_RATE, sizeof (rate), 1, &rate, tx));
263 uint64_t partial = vd->vdev_trim_partial;
264 if (partial == UINT64_MAX)
267 VERIFY0(zap_update(mos, vd->vdev_leaf_zap, VDEV_LEAF_ZAP_TRIM_PARTIAL,
268 sizeof (partial), 1, &partial, tx));
270 uint64_t secure = vd->vdev_trim_secure;
271 if (secure == UINT64_MAX)
274 VERIFY0(zap_update(mos, vd->vdev_leaf_zap, VDEV_LEAF_ZAP_TRIM_SECURE,
275 sizeof (secure), 1, &secure, tx));
278 uint64_t trim_state = vd->vdev_trim_state;
279 VERIFY0(zap_update(mos, vd->vdev_leaf_zap, VDEV_LEAF_ZAP_TRIM_STATE,
280 sizeof (trim_state), 1, &trim_state, tx));
284 * Update the on-disk state of a manual TRIM. This is called to request
285 * that a TRIM be started/suspended/canceled, or to change one of the
286 * TRIM options (partial, secure, rate).
289 vdev_trim_change_state(vdev_t *vd, vdev_trim_state_t new_state,
290 uint64_t rate, boolean_t partial, boolean_t secure)
292 ASSERT(MUTEX_HELD(&vd->vdev_trim_lock));
293 spa_t *spa = vd->vdev_spa;
295 if (new_state == vd->vdev_trim_state)
299 * Copy the vd's guid, this will be freed by the sync task.
301 uint64_t *guid = kmem_zalloc(sizeof (uint64_t), KM_SLEEP);
302 *guid = vd->vdev_guid;
305 * If we're suspending, then preserve the original start time.
307 if (vd->vdev_trim_state != VDEV_TRIM_SUSPENDED) {
308 vd->vdev_trim_action_time = gethrestime_sec();
312 * If we're activating, then preserve the requested rate and trim
313 * method. Setting the last offset and rate to UINT64_MAX is used
314 * as a sentinel to indicate they should be reset to default values.
316 if (new_state == VDEV_TRIM_ACTIVE) {
317 if (vd->vdev_trim_state == VDEV_TRIM_COMPLETE ||
318 vd->vdev_trim_state == VDEV_TRIM_CANCELED) {
319 vd->vdev_trim_last_offset = UINT64_MAX;
320 vd->vdev_trim_rate = UINT64_MAX;
321 vd->vdev_trim_partial = UINT64_MAX;
322 vd->vdev_trim_secure = UINT64_MAX;
326 vd->vdev_trim_rate = rate;
329 vd->vdev_trim_partial = partial;
332 vd->vdev_trim_secure = secure;
335 vdev_trim_state_t old_state = vd->vdev_trim_state;
336 boolean_t resumed = (old_state == VDEV_TRIM_SUSPENDED);
337 vd->vdev_trim_state = new_state;
339 dmu_tx_t *tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir);
340 VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
341 dsl_sync_task_nowait(spa_get_dsl(spa), vdev_trim_zap_update_sync,
345 case VDEV_TRIM_ACTIVE:
346 spa_event_notify(spa, vd, NULL,
347 resumed ? ESC_ZFS_TRIM_RESUME : ESC_ZFS_TRIM_START);
348 spa_history_log_internal(spa, "trim", tx,
349 "vdev=%s activated", vd->vdev_path);
351 case VDEV_TRIM_SUSPENDED:
352 spa_event_notify(spa, vd, NULL, ESC_ZFS_TRIM_SUSPEND);
353 spa_history_log_internal(spa, "trim", tx,
354 "vdev=%s suspended", vd->vdev_path);
356 case VDEV_TRIM_CANCELED:
357 if (old_state == VDEV_TRIM_ACTIVE ||
358 old_state == VDEV_TRIM_SUSPENDED) {
359 spa_event_notify(spa, vd, NULL, ESC_ZFS_TRIM_CANCEL);
360 spa_history_log_internal(spa, "trim", tx,
361 "vdev=%s canceled", vd->vdev_path);
364 case VDEV_TRIM_COMPLETE:
365 spa_event_notify(spa, vd, NULL, ESC_ZFS_TRIM_FINISH);
366 spa_history_log_internal(spa, "trim", tx,
367 "vdev=%s complete", vd->vdev_path);
370 panic("invalid state %llu", (unsigned long long)new_state);
375 if (new_state != VDEV_TRIM_ACTIVE)
376 spa_notify_waiters(spa);
380 * The zio_done_func_t done callback for each manual TRIM issued. It is
381 * responsible for updating the TRIM stats, reissuing failed TRIM I/Os,
382 * and limiting the number of in flight TRIM I/Os.
385 vdev_trim_cb(zio_t *zio)
387 vdev_t *vd = zio->io_vd;
389 mutex_enter(&vd->vdev_trim_io_lock);
390 if (zio->io_error == ENXIO && !vdev_writeable(vd)) {
392 * The I/O failed because the vdev was unavailable; roll the
393 * last offset back. (This works because spa_sync waits on
394 * spa_txg_zio before it runs sync tasks.)
397 &vd->vdev_trim_offset[zio->io_txg & TXG_MASK];
398 *offset = MIN(*offset, zio->io_offset);
400 if (zio->io_error != 0) {
401 vd->vdev_stat.vs_trim_errors++;
402 spa_iostats_trim_add(vd->vdev_spa, TRIM_TYPE_MANUAL,
403 0, 0, 0, 0, 1, zio->io_orig_size);
405 spa_iostats_trim_add(vd->vdev_spa, TRIM_TYPE_MANUAL,
406 1, zio->io_orig_size, 0, 0, 0, 0);
409 vd->vdev_trim_bytes_done += zio->io_orig_size;
412 ASSERT3U(vd->vdev_trim_inflight[TRIM_TYPE_MANUAL], >, 0);
413 vd->vdev_trim_inflight[TRIM_TYPE_MANUAL]--;
414 cv_broadcast(&vd->vdev_trim_io_cv);
415 mutex_exit(&vd->vdev_trim_io_lock);
417 spa_config_exit(vd->vdev_spa, SCL_STATE_ALL, vd);
421 * The zio_done_func_t done callback for each automatic TRIM issued. It
422 * is responsible for updating the TRIM stats and limiting the number of
423 * in flight TRIM I/Os. Automatic TRIM I/Os are best effort and are
424 * never reissued on failure.
427 vdev_autotrim_cb(zio_t *zio)
429 vdev_t *vd = zio->io_vd;
431 mutex_enter(&vd->vdev_trim_io_lock);
433 if (zio->io_error != 0) {
434 vd->vdev_stat.vs_trim_errors++;
435 spa_iostats_trim_add(vd->vdev_spa, TRIM_TYPE_AUTO,
436 0, 0, 0, 0, 1, zio->io_orig_size);
438 spa_iostats_trim_add(vd->vdev_spa, TRIM_TYPE_AUTO,
439 1, zio->io_orig_size, 0, 0, 0, 0);
442 ASSERT3U(vd->vdev_trim_inflight[TRIM_TYPE_AUTO], >, 0);
443 vd->vdev_trim_inflight[TRIM_TYPE_AUTO]--;
444 cv_broadcast(&vd->vdev_trim_io_cv);
445 mutex_exit(&vd->vdev_trim_io_lock);
447 spa_config_exit(vd->vdev_spa, SCL_STATE_ALL, vd);
451 * The zio_done_func_t done callback for each TRIM issued via
452 * vdev_trim_simple(). It is responsible for updating the TRIM stats and
453 * limiting the number of in flight TRIM I/Os. Simple TRIM I/Os are best
454 * effort and are never reissued on failure.
457 vdev_trim_simple_cb(zio_t *zio)
459 vdev_t *vd = zio->io_vd;
461 mutex_enter(&vd->vdev_trim_io_lock);
463 if (zio->io_error != 0) {
464 vd->vdev_stat.vs_trim_errors++;
465 spa_iostats_trim_add(vd->vdev_spa, TRIM_TYPE_SIMPLE,
466 0, 0, 0, 0, 1, zio->io_orig_size);
468 spa_iostats_trim_add(vd->vdev_spa, TRIM_TYPE_SIMPLE,
469 1, zio->io_orig_size, 0, 0, 0, 0);
472 ASSERT3U(vd->vdev_trim_inflight[TRIM_TYPE_SIMPLE], >, 0);
473 vd->vdev_trim_inflight[TRIM_TYPE_SIMPLE]--;
474 cv_broadcast(&vd->vdev_trim_io_cv);
475 mutex_exit(&vd->vdev_trim_io_lock);
477 spa_config_exit(vd->vdev_spa, SCL_STATE_ALL, vd);
480 * Returns the average trim rate in bytes/sec for the ta->trim_vdev.
483 vdev_trim_calculate_rate(trim_args_t *ta)
485 return (ta->trim_bytes_done * 1000 /
486 (NSEC2MSEC(gethrtime() - ta->trim_start_time) + 1));
490 * Issues a physical TRIM and takes care of rate limiting (bytes/sec)
491 * and number of concurrent TRIM I/Os.
494 vdev_trim_range(trim_args_t *ta, uint64_t start, uint64_t size)
496 vdev_t *vd = ta->trim_vdev;
497 spa_t *spa = vd->vdev_spa;
500 mutex_enter(&vd->vdev_trim_io_lock);
503 * Limit manual TRIM I/Os to the requested rate. This does not
504 * apply to automatic TRIM since no per vdev rate can be specified.
506 if (ta->trim_type == TRIM_TYPE_MANUAL) {
507 while (vd->vdev_trim_rate != 0 && !vdev_trim_should_stop(vd) &&
508 vdev_trim_calculate_rate(ta) > vd->vdev_trim_rate) {
509 cv_timedwait_idle(&vd->vdev_trim_io_cv,
510 &vd->vdev_trim_io_lock, ddi_get_lbolt() +
514 ta->trim_bytes_done += size;
516 /* Limit in flight trimming I/Os */
517 while (vd->vdev_trim_inflight[0] + vd->vdev_trim_inflight[1] +
518 vd->vdev_trim_inflight[2] >= zfs_trim_queue_limit) {
519 cv_wait(&vd->vdev_trim_io_cv, &vd->vdev_trim_io_lock);
521 vd->vdev_trim_inflight[ta->trim_type]++;
522 mutex_exit(&vd->vdev_trim_io_lock);
524 dmu_tx_t *tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir);
525 VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
526 uint64_t txg = dmu_tx_get_txg(tx);
528 spa_config_enter(spa, SCL_STATE_ALL, vd, RW_READER);
529 mutex_enter(&vd->vdev_trim_lock);
531 if (ta->trim_type == TRIM_TYPE_MANUAL &&
532 vd->vdev_trim_offset[txg & TXG_MASK] == 0) {
533 uint64_t *guid = kmem_zalloc(sizeof (uint64_t), KM_SLEEP);
534 *guid = vd->vdev_guid;
536 /* This is the first write of this txg. */
537 dsl_sync_task_nowait(spa_get_dsl(spa),
538 vdev_trim_zap_update_sync, guid, tx);
542 * We know the vdev_t will still be around since all consumers of
543 * vdev_free must stop the trimming first.
545 if ((ta->trim_type == TRIM_TYPE_MANUAL &&
546 vdev_trim_should_stop(vd)) ||
547 (ta->trim_type == TRIM_TYPE_AUTO &&
548 vdev_autotrim_should_stop(vd->vdev_top))) {
549 mutex_enter(&vd->vdev_trim_io_lock);
550 vd->vdev_trim_inflight[ta->trim_type]--;
551 mutex_exit(&vd->vdev_trim_io_lock);
552 spa_config_exit(vd->vdev_spa, SCL_STATE_ALL, vd);
553 mutex_exit(&vd->vdev_trim_lock);
555 return (SET_ERROR(EINTR));
557 mutex_exit(&vd->vdev_trim_lock);
559 if (ta->trim_type == TRIM_TYPE_MANUAL)
560 vd->vdev_trim_offset[txg & TXG_MASK] = start + size;
562 if (ta->trim_type == TRIM_TYPE_MANUAL) {
564 } else if (ta->trim_type == TRIM_TYPE_AUTO) {
565 cb = vdev_autotrim_cb;
567 cb = vdev_trim_simple_cb;
570 zio_nowait(zio_trim(spa->spa_txg_zio[txg & TXG_MASK], vd,
571 start, size, cb, NULL, ZIO_PRIORITY_TRIM, ZIO_FLAG_CANFAIL,
573 /* vdev_trim_cb and vdev_autotrim_cb release SCL_STATE_ALL */
581 * Issues TRIM I/Os for all ranges in the provided ta->trim_tree range tree.
582 * Additional parameters describing how the TRIM should be performed must
583 * be set in the trim_args structure. See the trim_args definition for
584 * additional information.
587 vdev_trim_ranges(trim_args_t *ta)
589 vdev_t *vd = ta->trim_vdev;
590 zfs_btree_t *t = &ta->trim_tree->rt_root;
591 zfs_btree_index_t idx;
592 uint64_t extent_bytes_max = ta->trim_extent_bytes_max;
593 uint64_t extent_bytes_min = ta->trim_extent_bytes_min;
594 spa_t *spa = vd->vdev_spa;
597 ta->trim_start_time = gethrtime();
598 ta->trim_bytes_done = 0;
600 for (range_seg_t *rs = zfs_btree_first(t, &idx); rs != NULL;
601 rs = zfs_btree_next(t, &idx, &idx)) {
602 uint64_t size = rs_get_end(rs, ta->trim_tree) - rs_get_start(rs,
605 if (extent_bytes_min && size < extent_bytes_min) {
606 spa_iostats_trim_add(spa, ta->trim_type,
607 0, 0, 1, size, 0, 0);
611 /* Split range into legally-sized physical chunks */
612 uint64_t writes_required = ((size - 1) / extent_bytes_max) + 1;
614 for (uint64_t w = 0; w < writes_required; w++) {
615 error = vdev_trim_range(ta, VDEV_LABEL_START_SIZE +
616 rs_get_start(rs, ta->trim_tree) +
617 (w *extent_bytes_max), MIN(size -
618 (w * extent_bytes_max), extent_bytes_max));
627 * Make sure all TRIMs for this metaslab have completed before
628 * returning. TRIM zios have lower priority over regular or syncing
629 * zios, so all TRIM zios for this metaslab must complete before the
630 * metaslab is re-enabled. Otherwise it's possible write zios to
631 * this metaslab could cut ahead of still queued TRIM zios for this
632 * metaslab causing corruption if the ranges overlap.
634 mutex_enter(&vd->vdev_trim_io_lock);
635 while (vd->vdev_trim_inflight[0] > 0) {
636 cv_wait(&vd->vdev_trim_io_cv, &vd->vdev_trim_io_lock);
638 mutex_exit(&vd->vdev_trim_io_lock);
644 vdev_trim_xlate_last_rs_end(void *arg, range_seg64_t *physical_rs)
646 uint64_t *last_rs_end = (uint64_t *)arg;
648 if (physical_rs->rs_end > *last_rs_end)
649 *last_rs_end = physical_rs->rs_end;
653 vdev_trim_xlate_progress(void *arg, range_seg64_t *physical_rs)
655 vdev_t *vd = (vdev_t *)arg;
657 uint64_t size = physical_rs->rs_end - physical_rs->rs_start;
658 vd->vdev_trim_bytes_est += size;
660 if (vd->vdev_trim_last_offset >= physical_rs->rs_end) {
661 vd->vdev_trim_bytes_done += size;
662 } else if (vd->vdev_trim_last_offset > physical_rs->rs_start &&
663 vd->vdev_trim_last_offset <= physical_rs->rs_end) {
664 vd->vdev_trim_bytes_done +=
665 vd->vdev_trim_last_offset - physical_rs->rs_start;
670 * Calculates the completion percentage of a manual TRIM.
673 vdev_trim_calculate_progress(vdev_t *vd)
675 ASSERT(spa_config_held(vd->vdev_spa, SCL_CONFIG, RW_READER) ||
676 spa_config_held(vd->vdev_spa, SCL_CONFIG, RW_WRITER));
677 ASSERT(vd->vdev_leaf_zap != 0);
679 vd->vdev_trim_bytes_est = 0;
680 vd->vdev_trim_bytes_done = 0;
682 for (uint64_t i = 0; i < vd->vdev_top->vdev_ms_count; i++) {
683 metaslab_t *msp = vd->vdev_top->vdev_ms[i];
684 mutex_enter(&msp->ms_lock);
686 uint64_t ms_free = (msp->ms_size -
687 metaslab_allocated_space(msp)) /
688 vdev_get_ndisks(vd->vdev_top);
691 * Convert the metaslab range to a physical range
692 * on our vdev. We use this to determine if we are
693 * in the middle of this metaslab range.
695 range_seg64_t logical_rs, physical_rs, remain_rs;
696 logical_rs.rs_start = msp->ms_start;
697 logical_rs.rs_end = msp->ms_start + msp->ms_size;
699 /* Metaslab space after this offset has not been trimmed. */
700 vdev_xlate(vd, &logical_rs, &physical_rs, &remain_rs);
701 if (vd->vdev_trim_last_offset <= physical_rs.rs_start) {
702 vd->vdev_trim_bytes_est += ms_free;
703 mutex_exit(&msp->ms_lock);
707 /* Metaslab space before this offset has been trimmed */
708 uint64_t last_rs_end = physical_rs.rs_end;
709 if (!vdev_xlate_is_empty(&remain_rs)) {
710 vdev_xlate_walk(vd, &remain_rs,
711 vdev_trim_xlate_last_rs_end, &last_rs_end);
714 if (vd->vdev_trim_last_offset > last_rs_end) {
715 vd->vdev_trim_bytes_done += ms_free;
716 vd->vdev_trim_bytes_est += ms_free;
717 mutex_exit(&msp->ms_lock);
722 * If we get here, we're in the middle of trimming this
723 * metaslab. Load it and walk the free tree for more
724 * accurate progress estimation.
726 VERIFY0(metaslab_load(msp));
728 range_tree_t *rt = msp->ms_allocatable;
729 zfs_btree_t *bt = &rt->rt_root;
730 zfs_btree_index_t idx;
731 for (range_seg_t *rs = zfs_btree_first(bt, &idx);
732 rs != NULL; rs = zfs_btree_next(bt, &idx, &idx)) {
733 logical_rs.rs_start = rs_get_start(rs, rt);
734 logical_rs.rs_end = rs_get_end(rs, rt);
736 vdev_xlate_walk(vd, &logical_rs,
737 vdev_trim_xlate_progress, vd);
739 mutex_exit(&msp->ms_lock);
744 * Load from disk the vdev's manual TRIM information. This includes the
745 * state, progress, and options provided when initiating the manual TRIM.
748 vdev_trim_load(vdev_t *vd)
751 ASSERT(spa_config_held(vd->vdev_spa, SCL_CONFIG, RW_READER) ||
752 spa_config_held(vd->vdev_spa, SCL_CONFIG, RW_WRITER));
753 ASSERT(vd->vdev_leaf_zap != 0);
755 if (vd->vdev_trim_state == VDEV_TRIM_ACTIVE ||
756 vd->vdev_trim_state == VDEV_TRIM_SUSPENDED) {
757 err = zap_lookup(vd->vdev_spa->spa_meta_objset,
758 vd->vdev_leaf_zap, VDEV_LEAF_ZAP_TRIM_LAST_OFFSET,
759 sizeof (vd->vdev_trim_last_offset), 1,
760 &vd->vdev_trim_last_offset);
762 vd->vdev_trim_last_offset = 0;
767 err = zap_lookup(vd->vdev_spa->spa_meta_objset,
768 vd->vdev_leaf_zap, VDEV_LEAF_ZAP_TRIM_RATE,
769 sizeof (vd->vdev_trim_rate), 1,
770 &vd->vdev_trim_rate);
772 vd->vdev_trim_rate = 0;
778 err = zap_lookup(vd->vdev_spa->spa_meta_objset,
779 vd->vdev_leaf_zap, VDEV_LEAF_ZAP_TRIM_PARTIAL,
780 sizeof (vd->vdev_trim_partial), 1,
781 &vd->vdev_trim_partial);
783 vd->vdev_trim_partial = 0;
789 err = zap_lookup(vd->vdev_spa->spa_meta_objset,
790 vd->vdev_leaf_zap, VDEV_LEAF_ZAP_TRIM_SECURE,
791 sizeof (vd->vdev_trim_secure), 1,
792 &vd->vdev_trim_secure);
794 vd->vdev_trim_secure = 0;
800 vdev_trim_calculate_progress(vd);
806 vdev_trim_xlate_range_add(void *arg, range_seg64_t *physical_rs)
808 trim_args_t *ta = arg;
809 vdev_t *vd = ta->trim_vdev;
812 * Only a manual trim will be traversing the vdev sequentially.
813 * For an auto trim all valid ranges should be added.
815 if (ta->trim_type == TRIM_TYPE_MANUAL) {
817 /* Only add segments that we have not visited yet */
818 if (physical_rs->rs_end <= vd->vdev_trim_last_offset)
821 /* Pick up where we left off mid-range. */
822 if (vd->vdev_trim_last_offset > physical_rs->rs_start) {
823 ASSERT3U(physical_rs->rs_end, >,
824 vd->vdev_trim_last_offset);
825 physical_rs->rs_start = vd->vdev_trim_last_offset;
829 ASSERT3U(physical_rs->rs_end, >, physical_rs->rs_start);
831 range_tree_add(ta->trim_tree, physical_rs->rs_start,
832 physical_rs->rs_end - physical_rs->rs_start);
836 * Convert the logical range into physical ranges and add them to the
837 * range tree passed in the trim_args_t.
840 vdev_trim_range_add(void *arg, uint64_t start, uint64_t size)
842 trim_args_t *ta = arg;
843 vdev_t *vd = ta->trim_vdev;
844 range_seg64_t logical_rs;
845 logical_rs.rs_start = start;
846 logical_rs.rs_end = start + size;
849 * Every range to be trimmed must be part of ms_allocatable.
850 * When ZFS_DEBUG_TRIM is set load the metaslab to verify this
851 * is always the case.
853 if (zfs_flags & ZFS_DEBUG_TRIM) {
854 metaslab_t *msp = ta->trim_msp;
855 VERIFY0(metaslab_load(msp));
856 VERIFY3B(msp->ms_loaded, ==, B_TRUE);
857 VERIFY(range_tree_contains(msp->ms_allocatable, start, size));
860 ASSERT(vd->vdev_ops->vdev_op_leaf);
861 vdev_xlate_walk(vd, &logical_rs, vdev_trim_xlate_range_add, arg);
865 * Each manual TRIM thread is responsible for trimming the unallocated
866 * space for each leaf vdev. This is accomplished by sequentially iterating
867 * over its top-level metaslabs and issuing TRIM I/O for the space described
868 * by its ms_allocatable. While a metaslab is undergoing trimming it is
869 * not eligible for new allocations.
871 static __attribute__((noreturn)) void
872 vdev_trim_thread(void *arg)
875 spa_t *spa = vd->vdev_spa;
880 * The VDEV_LEAF_ZAP_TRIM_* entries may have been updated by
881 * vdev_trim(). Wait for the updated values to be reflected
882 * in the zap in order to start with the requested settings.
884 txg_wait_synced(spa_get_dsl(vd->vdev_spa), 0);
886 ASSERT(vdev_is_concrete(vd));
887 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
889 vd->vdev_trim_last_offset = 0;
890 vd->vdev_trim_rate = 0;
891 vd->vdev_trim_partial = 0;
892 vd->vdev_trim_secure = 0;
894 VERIFY0(vdev_trim_load(vd));
897 ta.trim_extent_bytes_max = zfs_trim_extent_bytes_max;
898 ta.trim_extent_bytes_min = zfs_trim_extent_bytes_min;
899 ta.trim_tree = range_tree_create(NULL, RANGE_SEG64, NULL, 0, 0);
900 ta.trim_type = TRIM_TYPE_MANUAL;
904 * When a secure TRIM has been requested infer that the intent
905 * is that everything must be trimmed. Override the default
906 * minimum TRIM size to prevent ranges from being skipped.
908 if (vd->vdev_trim_secure) {
909 ta.trim_flags |= ZIO_TRIM_SECURE;
910 ta.trim_extent_bytes_min = SPA_MINBLOCKSIZE;
913 uint64_t ms_count = 0;
914 for (uint64_t i = 0; !vd->vdev_detached &&
915 i < vd->vdev_top->vdev_ms_count; i++) {
916 metaslab_t *msp = vd->vdev_top->vdev_ms[i];
919 * If we've expanded the top-level vdev or it's our
920 * first pass, calculate our progress.
922 if (vd->vdev_top->vdev_ms_count != ms_count) {
923 vdev_trim_calculate_progress(vd);
924 ms_count = vd->vdev_top->vdev_ms_count;
927 spa_config_exit(spa, SCL_CONFIG, FTAG);
928 metaslab_disable(msp);
929 mutex_enter(&msp->ms_lock);
930 VERIFY0(metaslab_load(msp));
933 * If a partial TRIM was requested skip metaslabs which have
934 * never been initialized and thus have never been written.
936 if (msp->ms_sm == NULL && vd->vdev_trim_partial) {
937 mutex_exit(&msp->ms_lock);
938 metaslab_enable(msp, B_FALSE, B_FALSE);
939 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
940 vdev_trim_calculate_progress(vd);
945 range_tree_walk(msp->ms_allocatable, vdev_trim_range_add, &ta);
946 range_tree_vacate(msp->ms_trim, NULL, NULL);
947 mutex_exit(&msp->ms_lock);
949 error = vdev_trim_ranges(&ta);
950 metaslab_enable(msp, B_TRUE, B_FALSE);
951 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
953 range_tree_vacate(ta.trim_tree, NULL, NULL);
958 spa_config_exit(spa, SCL_CONFIG, FTAG);
960 range_tree_destroy(ta.trim_tree);
962 mutex_enter(&vd->vdev_trim_lock);
963 if (!vd->vdev_trim_exit_wanted) {
964 if (vdev_writeable(vd)) {
965 vdev_trim_change_state(vd, VDEV_TRIM_COMPLETE,
966 vd->vdev_trim_rate, vd->vdev_trim_partial,
967 vd->vdev_trim_secure);
968 } else if (vd->vdev_faulted) {
969 vdev_trim_change_state(vd, VDEV_TRIM_CANCELED,
970 vd->vdev_trim_rate, vd->vdev_trim_partial,
971 vd->vdev_trim_secure);
974 ASSERT(vd->vdev_trim_thread != NULL || vd->vdev_trim_inflight[0] == 0);
977 * Drop the vdev_trim_lock while we sync out the txg since it's
978 * possible that a device might be trying to come online and must
979 * check to see if it needs to restart a trim. That thread will be
980 * holding the spa_config_lock which would prevent the txg_wait_synced
983 mutex_exit(&vd->vdev_trim_lock);
984 txg_wait_synced(spa_get_dsl(spa), 0);
985 mutex_enter(&vd->vdev_trim_lock);
987 vd->vdev_trim_thread = NULL;
988 cv_broadcast(&vd->vdev_trim_cv);
989 mutex_exit(&vd->vdev_trim_lock);
995 * Initiates a manual TRIM for the vdev_t. Callers must hold vdev_trim_lock,
996 * the vdev_t must be a leaf and cannot already be manually trimming.
999 vdev_trim(vdev_t *vd, uint64_t rate, boolean_t partial, boolean_t secure)
1001 ASSERT(MUTEX_HELD(&vd->vdev_trim_lock));
1002 ASSERT(vd->vdev_ops->vdev_op_leaf);
1003 ASSERT(vdev_is_concrete(vd));
1004 ASSERT3P(vd->vdev_trim_thread, ==, NULL);
1005 ASSERT(!vd->vdev_detached);
1006 ASSERT(!vd->vdev_trim_exit_wanted);
1007 ASSERT(!vd->vdev_top->vdev_removing);
1009 vdev_trim_change_state(vd, VDEV_TRIM_ACTIVE, rate, partial, secure);
1010 vd->vdev_trim_thread = thread_create(NULL, 0,
1011 vdev_trim_thread, vd, 0, &p0, TS_RUN, maxclsyspri);
1015 * Wait for the trimming thread to be terminated (canceled or stopped).
1018 vdev_trim_stop_wait_impl(vdev_t *vd)
1020 ASSERT(MUTEX_HELD(&vd->vdev_trim_lock));
1022 while (vd->vdev_trim_thread != NULL)
1023 cv_wait(&vd->vdev_trim_cv, &vd->vdev_trim_lock);
1025 ASSERT3P(vd->vdev_trim_thread, ==, NULL);
1026 vd->vdev_trim_exit_wanted = B_FALSE;
1030 * Wait for vdev trim threads which were listed to cleanly exit.
1033 vdev_trim_stop_wait(spa_t *spa, list_t *vd_list)
1038 ASSERT(MUTEX_HELD(&spa_namespace_lock));
1040 while ((vd = list_remove_head(vd_list)) != NULL) {
1041 mutex_enter(&vd->vdev_trim_lock);
1042 vdev_trim_stop_wait_impl(vd);
1043 mutex_exit(&vd->vdev_trim_lock);
1048 * Stop trimming a device, with the resultant trimming state being tgt_state.
1049 * For blocking behavior pass NULL for vd_list. Otherwise, when a list_t is
1050 * provided the stopping vdev is inserted in to the list. Callers are then
1051 * required to call vdev_trim_stop_wait() to block for all the trim threads
1052 * to exit. The caller must hold vdev_trim_lock and must not be writing to
1053 * the spa config, as the trimming thread may try to enter the config as a
1054 * reader before exiting.
1057 vdev_trim_stop(vdev_t *vd, vdev_trim_state_t tgt_state, list_t *vd_list)
1059 ASSERT(!spa_config_held(vd->vdev_spa, SCL_CONFIG|SCL_STATE, RW_WRITER));
1060 ASSERT(MUTEX_HELD(&vd->vdev_trim_lock));
1061 ASSERT(vd->vdev_ops->vdev_op_leaf);
1062 ASSERT(vdev_is_concrete(vd));
1065 * Allow cancel requests to proceed even if the trim thread has
1068 if (vd->vdev_trim_thread == NULL && tgt_state != VDEV_TRIM_CANCELED)
1071 vdev_trim_change_state(vd, tgt_state, 0, 0, 0);
1072 vd->vdev_trim_exit_wanted = B_TRUE;
1074 if (vd_list == NULL) {
1075 vdev_trim_stop_wait_impl(vd);
1077 ASSERT(MUTEX_HELD(&spa_namespace_lock));
1078 list_insert_tail(vd_list, vd);
1083 * Requests that all listed vdevs stop trimming.
1086 vdev_trim_stop_all_impl(vdev_t *vd, vdev_trim_state_t tgt_state,
1089 if (vd->vdev_ops->vdev_op_leaf && vdev_is_concrete(vd)) {
1090 mutex_enter(&vd->vdev_trim_lock);
1091 vdev_trim_stop(vd, tgt_state, vd_list);
1092 mutex_exit(&vd->vdev_trim_lock);
1096 for (uint64_t i = 0; i < vd->vdev_children; i++) {
1097 vdev_trim_stop_all_impl(vd->vdev_child[i], tgt_state,
1103 * Convenience function to stop trimming of a vdev tree and set all trim
1104 * thread pointers to NULL.
1107 vdev_trim_stop_all(vdev_t *vd, vdev_trim_state_t tgt_state)
1109 spa_t *spa = vd->vdev_spa;
1113 ASSERT(MUTEX_HELD(&spa_namespace_lock));
1115 list_create(&vd_list, sizeof (vdev_t),
1116 offsetof(vdev_t, vdev_trim_node));
1118 vdev_trim_stop_all_impl(vd, tgt_state, &vd_list);
1121 * Iterate over cache devices and request stop trimming the
1122 * whole device in case we export the pool or remove the cache
1123 * device prematurely.
1125 for (int i = 0; i < spa->spa_l2cache.sav_count; i++) {
1126 vd_l2cache = spa->spa_l2cache.sav_vdevs[i];
1127 vdev_trim_stop_all_impl(vd_l2cache, tgt_state, &vd_list);
1130 vdev_trim_stop_wait(spa, &vd_list);
1132 if (vd->vdev_spa->spa_sync_on) {
1133 /* Make sure that our state has been synced to disk */
1134 txg_wait_synced(spa_get_dsl(vd->vdev_spa), 0);
1137 list_destroy(&vd_list);
1141 * Conditionally restarts a manual TRIM given its on-disk state.
1144 vdev_trim_restart(vdev_t *vd)
1146 ASSERT(MUTEX_HELD(&spa_namespace_lock));
1147 ASSERT(!spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER));
1149 if (vd->vdev_leaf_zap != 0) {
1150 mutex_enter(&vd->vdev_trim_lock);
1151 uint64_t trim_state = VDEV_TRIM_NONE;
1152 int err = zap_lookup(vd->vdev_spa->spa_meta_objset,
1153 vd->vdev_leaf_zap, VDEV_LEAF_ZAP_TRIM_STATE,
1154 sizeof (trim_state), 1, &trim_state);
1155 ASSERT(err == 0 || err == ENOENT);
1156 vd->vdev_trim_state = trim_state;
1158 uint64_t timestamp = 0;
1159 err = zap_lookup(vd->vdev_spa->spa_meta_objset,
1160 vd->vdev_leaf_zap, VDEV_LEAF_ZAP_TRIM_ACTION_TIME,
1161 sizeof (timestamp), 1, ×tamp);
1162 ASSERT(err == 0 || err == ENOENT);
1163 vd->vdev_trim_action_time = timestamp;
1165 if (vd->vdev_trim_state == VDEV_TRIM_SUSPENDED ||
1167 /* load progress for reporting, but don't resume */
1168 VERIFY0(vdev_trim_load(vd));
1169 } else if (vd->vdev_trim_state == VDEV_TRIM_ACTIVE &&
1170 vdev_writeable(vd) && !vd->vdev_top->vdev_removing &&
1171 vd->vdev_trim_thread == NULL) {
1172 VERIFY0(vdev_trim_load(vd));
1173 vdev_trim(vd, vd->vdev_trim_rate,
1174 vd->vdev_trim_partial, vd->vdev_trim_secure);
1177 mutex_exit(&vd->vdev_trim_lock);
1180 for (uint64_t i = 0; i < vd->vdev_children; i++) {
1181 vdev_trim_restart(vd->vdev_child[i]);
1186 * Used by the automatic TRIM when ZFS_DEBUG_TRIM is set to verify that
1187 * every TRIM range is contained within ms_allocatable.
1190 vdev_trim_range_verify(void *arg, uint64_t start, uint64_t size)
1192 trim_args_t *ta = arg;
1193 metaslab_t *msp = ta->trim_msp;
1195 VERIFY3B(msp->ms_loaded, ==, B_TRUE);
1196 VERIFY3U(msp->ms_disabled, >, 0);
1197 VERIFY(range_tree_contains(msp->ms_allocatable, start, size));
1201 * Each automatic TRIM thread is responsible for managing the trimming of a
1202 * top-level vdev in the pool. No automatic TRIM state is maintained on-disk.
1204 * N.B. This behavior is different from a manual TRIM where a thread
1205 * is created for each leaf vdev, instead of each top-level vdev.
1207 static __attribute__((noreturn)) void
1208 vdev_autotrim_thread(void *arg)
1211 spa_t *spa = vd->vdev_spa;
1214 mutex_enter(&vd->vdev_autotrim_lock);
1215 ASSERT3P(vd->vdev_top, ==, vd);
1216 ASSERT3P(vd->vdev_autotrim_thread, !=, NULL);
1217 mutex_exit(&vd->vdev_autotrim_lock);
1218 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
1220 while (!vdev_autotrim_should_stop(vd)) {
1221 int txgs_per_trim = MAX(zfs_trim_txg_batch, 1);
1222 uint64_t extent_bytes_max = zfs_trim_extent_bytes_max;
1223 uint64_t extent_bytes_min = zfs_trim_extent_bytes_min;
1226 * All of the metaslabs are divided in to groups of size
1227 * num_metaslabs / zfs_trim_txg_batch. Each of these groups
1228 * is composed of metaslabs which are spread evenly over the
1231 * For example, when zfs_trim_txg_batch = 32 (default) then
1232 * group 0 will contain metaslabs 0, 32, 64, ...;
1233 * group 1 will contain metaslabs 1, 33, 65, ...;
1234 * group 2 will contain metaslabs 2, 34, 66, ...; and so on.
1236 * On each pass through the while() loop one of these groups
1237 * is selected. This is accomplished by using a shift value
1238 * to select the starting metaslab, then striding over the
1239 * metaslabs using the zfs_trim_txg_batch size. This is
1240 * done to accomplish two things.
1242 * 1) By dividing the metaslabs in to groups, and making sure
1243 * that each group takes a minimum of one txg to process.
1244 * Then zfs_trim_txg_batch controls the minimum number of
1245 * txgs which must occur before a metaslab is revisited.
1247 * 2) Selecting non-consecutive metaslabs distributes the
1248 * TRIM commands for a group evenly over the entire device.
1249 * This can be advantageous for certain types of devices.
1251 for (uint64_t i = shift % txgs_per_trim; i < vd->vdev_ms_count;
1252 i += txgs_per_trim) {
1253 metaslab_t *msp = vd->vdev_ms[i];
1254 range_tree_t *trim_tree;
1255 boolean_t issued_trim = B_FALSE;
1256 boolean_t wait_aborted = B_FALSE;
1258 spa_config_exit(spa, SCL_CONFIG, FTAG);
1259 metaslab_disable(msp);
1260 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
1262 mutex_enter(&msp->ms_lock);
1265 * Skip the metaslab when it has never been allocated
1266 * or when there are no recent frees to trim.
1268 if (msp->ms_sm == NULL ||
1269 range_tree_is_empty(msp->ms_trim)) {
1270 mutex_exit(&msp->ms_lock);
1271 metaslab_enable(msp, B_FALSE, B_FALSE);
1276 * Skip the metaslab when it has already been disabled.
1277 * This may happen when a manual TRIM or initialize
1278 * operation is running concurrently. In the case
1279 * of a manual TRIM, the ms_trim tree will have been
1280 * vacated. Only ranges added after the manual TRIM
1281 * disabled the metaslab will be included in the tree.
1282 * These will be processed when the automatic TRIM
1283 * next revisits this metaslab.
1285 if (msp->ms_disabled > 1) {
1286 mutex_exit(&msp->ms_lock);
1287 metaslab_enable(msp, B_FALSE, B_FALSE);
1292 * Allocate an empty range tree which is swapped in
1293 * for the existing ms_trim tree while it is processed.
1295 trim_tree = range_tree_create(NULL, RANGE_SEG64, NULL,
1297 range_tree_swap(&msp->ms_trim, &trim_tree);
1298 ASSERT(range_tree_is_empty(msp->ms_trim));
1301 * There are two cases when constructing the per-vdev
1302 * trim trees for a metaslab. If the top-level vdev
1303 * has no children then it is also a leaf and should
1304 * be trimmed. Otherwise our children are the leaves
1305 * and a trim tree should be constructed for each.
1308 uint64_t children = vd->vdev_children;
1309 if (children == 0) {
1311 tap = kmem_zalloc(sizeof (trim_args_t) *
1312 children, KM_SLEEP);
1313 tap[0].trim_vdev = vd;
1315 tap = kmem_zalloc(sizeof (trim_args_t) *
1316 children, KM_SLEEP);
1318 for (uint64_t c = 0; c < children; c++) {
1319 tap[c].trim_vdev = vd->vdev_child[c];
1323 for (uint64_t c = 0; c < children; c++) {
1324 trim_args_t *ta = &tap[c];
1325 vdev_t *cvd = ta->trim_vdev;
1328 ta->trim_extent_bytes_max = extent_bytes_max;
1329 ta->trim_extent_bytes_min = extent_bytes_min;
1330 ta->trim_type = TRIM_TYPE_AUTO;
1333 if (cvd->vdev_detached ||
1334 !vdev_writeable(cvd) ||
1335 !cvd->vdev_has_trim ||
1336 cvd->vdev_trim_thread != NULL) {
1341 * When a device has an attached hot spare, or
1342 * is being replaced it will not be trimmed.
1343 * This is done to avoid adding additional
1344 * stress to a potentially unhealthy device,
1345 * and to minimize the required rebuild time.
1347 if (!cvd->vdev_ops->vdev_op_leaf)
1350 ta->trim_tree = range_tree_create(NULL,
1351 RANGE_SEG64, NULL, 0, 0);
1352 range_tree_walk(trim_tree,
1353 vdev_trim_range_add, ta);
1356 mutex_exit(&msp->ms_lock);
1357 spa_config_exit(spa, SCL_CONFIG, FTAG);
1360 * Issue the TRIM I/Os for all ranges covered by the
1361 * TRIM trees. These ranges are safe to TRIM because
1362 * no new allocations will be performed until the call
1363 * to metaslab_enabled() below.
1365 for (uint64_t c = 0; c < children; c++) {
1366 trim_args_t *ta = &tap[c];
1369 * Always yield to a manual TRIM if one has
1370 * been started for the child vdev.
1372 if (ta->trim_tree == NULL ||
1373 ta->trim_vdev->vdev_trim_thread != NULL) {
1378 * After this point metaslab_enable() must be
1379 * called with the sync flag set. This is done
1380 * here because vdev_trim_ranges() is allowed
1381 * to be interrupted (EINTR) before issuing all
1382 * of the required TRIM I/Os.
1384 issued_trim = B_TRUE;
1386 int error = vdev_trim_ranges(ta);
1392 * Verify every range which was trimmed is still
1393 * contained within the ms_allocatable tree.
1395 if (zfs_flags & ZFS_DEBUG_TRIM) {
1396 mutex_enter(&msp->ms_lock);
1397 VERIFY0(metaslab_load(msp));
1398 VERIFY3P(tap[0].trim_msp, ==, msp);
1399 range_tree_walk(trim_tree,
1400 vdev_trim_range_verify, &tap[0]);
1401 mutex_exit(&msp->ms_lock);
1404 range_tree_vacate(trim_tree, NULL, NULL);
1405 range_tree_destroy(trim_tree);
1408 * Wait for couples of kicks, to ensure the trim io is
1409 * synced. If the wait is aborted due to
1410 * vdev_autotrim_exit_wanted, we need to signal
1411 * metaslab_enable() to wait for sync.
1414 wait_aborted = vdev_autotrim_wait_kick(vd,
1415 TXG_CONCURRENT_STATES + TXG_DEFER_SIZE);
1418 metaslab_enable(msp, wait_aborted, B_FALSE);
1419 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
1421 for (uint64_t c = 0; c < children; c++) {
1422 trim_args_t *ta = &tap[c];
1424 if (ta->trim_tree == NULL)
1427 range_tree_vacate(ta->trim_tree, NULL, NULL);
1428 range_tree_destroy(ta->trim_tree);
1431 kmem_free(tap, sizeof (trim_args_t) * children);
1433 if (vdev_autotrim_should_stop(vd))
1437 spa_config_exit(spa, SCL_CONFIG, FTAG);
1439 vdev_autotrim_wait_kick(vd, 1);
1442 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
1445 for (uint64_t c = 0; c < vd->vdev_children; c++) {
1446 vdev_t *cvd = vd->vdev_child[c];
1447 mutex_enter(&cvd->vdev_trim_io_lock);
1449 while (cvd->vdev_trim_inflight[1] > 0) {
1450 cv_wait(&cvd->vdev_trim_io_cv,
1451 &cvd->vdev_trim_io_lock);
1453 mutex_exit(&cvd->vdev_trim_io_lock);
1456 spa_config_exit(spa, SCL_CONFIG, FTAG);
1459 * When exiting because the autotrim property was set to off, then
1460 * abandon any unprocessed ms_trim ranges to reclaim the memory.
1462 if (spa_get_autotrim(spa) == SPA_AUTOTRIM_OFF) {
1463 for (uint64_t i = 0; i < vd->vdev_ms_count; i++) {
1464 metaslab_t *msp = vd->vdev_ms[i];
1466 mutex_enter(&msp->ms_lock);
1467 range_tree_vacate(msp->ms_trim, NULL, NULL);
1468 mutex_exit(&msp->ms_lock);
1472 mutex_enter(&vd->vdev_autotrim_lock);
1473 ASSERT(vd->vdev_autotrim_thread != NULL);
1474 vd->vdev_autotrim_thread = NULL;
1475 cv_broadcast(&vd->vdev_autotrim_cv);
1476 mutex_exit(&vd->vdev_autotrim_lock);
1482 * Starts an autotrim thread, if needed, for each top-level vdev which can be
1483 * trimmed. A top-level vdev which has been evacuated will never be trimmed.
1486 vdev_autotrim(spa_t *spa)
1488 vdev_t *root_vd = spa->spa_root_vdev;
1490 for (uint64_t i = 0; i < root_vd->vdev_children; i++) {
1491 vdev_t *tvd = root_vd->vdev_child[i];
1493 mutex_enter(&tvd->vdev_autotrim_lock);
1494 if (vdev_writeable(tvd) && !tvd->vdev_removing &&
1495 tvd->vdev_autotrim_thread == NULL) {
1496 ASSERT3P(tvd->vdev_top, ==, tvd);
1498 tvd->vdev_autotrim_thread = thread_create(NULL, 0,
1499 vdev_autotrim_thread, tvd, 0, &p0, TS_RUN,
1501 ASSERT(tvd->vdev_autotrim_thread != NULL);
1503 mutex_exit(&tvd->vdev_autotrim_lock);
1508 * Wait for the vdev_autotrim_thread associated with the passed top-level
1509 * vdev to be terminated (canceled or stopped).
1512 vdev_autotrim_stop_wait(vdev_t *tvd)
1514 mutex_enter(&tvd->vdev_autotrim_lock);
1515 if (tvd->vdev_autotrim_thread != NULL) {
1516 tvd->vdev_autotrim_exit_wanted = B_TRUE;
1517 cv_broadcast(&tvd->vdev_autotrim_kick_cv);
1518 cv_wait(&tvd->vdev_autotrim_cv,
1519 &tvd->vdev_autotrim_lock);
1521 ASSERT3P(tvd->vdev_autotrim_thread, ==, NULL);
1522 tvd->vdev_autotrim_exit_wanted = B_FALSE;
1524 mutex_exit(&tvd->vdev_autotrim_lock);
1528 vdev_autotrim_kick(spa_t *spa)
1530 ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER));
1532 vdev_t *root_vd = spa->spa_root_vdev;
1535 for (uint64_t i = 0; i < root_vd->vdev_children; i++) {
1536 tvd = root_vd->vdev_child[i];
1538 mutex_enter(&tvd->vdev_autotrim_lock);
1539 if (tvd->vdev_autotrim_thread != NULL)
1540 cv_broadcast(&tvd->vdev_autotrim_kick_cv);
1541 mutex_exit(&tvd->vdev_autotrim_lock);
1546 * Wait for all of the vdev_autotrim_thread associated with the pool to
1547 * be terminated (canceled or stopped).
1550 vdev_autotrim_stop_all(spa_t *spa)
1552 vdev_t *root_vd = spa->spa_root_vdev;
1554 for (uint64_t i = 0; i < root_vd->vdev_children; i++)
1555 vdev_autotrim_stop_wait(root_vd->vdev_child[i]);
1559 * Conditionally restart all of the vdev_autotrim_thread's for the pool.
1562 vdev_autotrim_restart(spa_t *spa)
1564 ASSERT(MUTEX_HELD(&spa_namespace_lock));
1566 if (spa->spa_autotrim)
1570 static __attribute__((noreturn)) void
1571 vdev_trim_l2arc_thread(void *arg)
1574 spa_t *spa = vd->vdev_spa;
1575 l2arc_dev_t *dev = l2arc_vdev_get(vd);
1576 trim_args_t ta = {0};
1577 range_seg64_t physical_rs;
1579 ASSERT(vdev_is_concrete(vd));
1580 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
1582 vd->vdev_trim_last_offset = 0;
1583 vd->vdev_trim_rate = 0;
1584 vd->vdev_trim_partial = 0;
1585 vd->vdev_trim_secure = 0;
1588 ta.trim_tree = range_tree_create(NULL, RANGE_SEG64, NULL, 0, 0);
1589 ta.trim_type = TRIM_TYPE_MANUAL;
1590 ta.trim_extent_bytes_max = zfs_trim_extent_bytes_max;
1591 ta.trim_extent_bytes_min = SPA_MINBLOCKSIZE;
1594 physical_rs.rs_start = vd->vdev_trim_bytes_done = 0;
1595 physical_rs.rs_end = vd->vdev_trim_bytes_est =
1596 vdev_get_min_asize(vd);
1598 range_tree_add(ta.trim_tree, physical_rs.rs_start,
1599 physical_rs.rs_end - physical_rs.rs_start);
1601 mutex_enter(&vd->vdev_trim_lock);
1602 vdev_trim_change_state(vd, VDEV_TRIM_ACTIVE, 0, 0, 0);
1603 mutex_exit(&vd->vdev_trim_lock);
1605 (void) vdev_trim_ranges(&ta);
1607 spa_config_exit(spa, SCL_CONFIG, FTAG);
1608 mutex_enter(&vd->vdev_trim_io_lock);
1609 while (vd->vdev_trim_inflight[TRIM_TYPE_MANUAL] > 0) {
1610 cv_wait(&vd->vdev_trim_io_cv, &vd->vdev_trim_io_lock);
1612 mutex_exit(&vd->vdev_trim_io_lock);
1614 range_tree_vacate(ta.trim_tree, NULL, NULL);
1615 range_tree_destroy(ta.trim_tree);
1617 mutex_enter(&vd->vdev_trim_lock);
1618 if (!vd->vdev_trim_exit_wanted && vdev_writeable(vd)) {
1619 vdev_trim_change_state(vd, VDEV_TRIM_COMPLETE,
1620 vd->vdev_trim_rate, vd->vdev_trim_partial,
1621 vd->vdev_trim_secure);
1623 ASSERT(vd->vdev_trim_thread != NULL ||
1624 vd->vdev_trim_inflight[TRIM_TYPE_MANUAL] == 0);
1627 * Drop the vdev_trim_lock while we sync out the txg since it's
1628 * possible that a device might be trying to come online and
1629 * must check to see if it needs to restart a trim. That thread
1630 * will be holding the spa_config_lock which would prevent the
1631 * txg_wait_synced from completing. Same strategy as in
1632 * vdev_trim_thread().
1634 mutex_exit(&vd->vdev_trim_lock);
1635 txg_wait_synced(spa_get_dsl(vd->vdev_spa), 0);
1636 mutex_enter(&vd->vdev_trim_lock);
1639 * Update the header of the cache device here, before
1640 * broadcasting vdev_trim_cv which may lead to the removal
1641 * of the device. The same applies for setting l2ad_trim_all to
1644 spa_config_enter(vd->vdev_spa, SCL_L2ARC, vd,
1646 memset(dev->l2ad_dev_hdr, 0, dev->l2ad_dev_hdr_asize);
1647 l2arc_dev_hdr_update(dev);
1648 spa_config_exit(vd->vdev_spa, SCL_L2ARC, vd);
1650 vd->vdev_trim_thread = NULL;
1651 if (vd->vdev_trim_state == VDEV_TRIM_COMPLETE)
1652 dev->l2ad_trim_all = B_FALSE;
1654 cv_broadcast(&vd->vdev_trim_cv);
1655 mutex_exit(&vd->vdev_trim_lock);
1661 * Punches out TRIM threads for the L2ARC devices in a spa and assigns them
1662 * to vd->vdev_trim_thread variable. This facilitates the management of
1663 * trimming the whole cache device using TRIM_TYPE_MANUAL upon addition
1664 * to a pool or pool creation or when the header of the device is invalid.
1667 vdev_trim_l2arc(spa_t *spa)
1669 ASSERT(MUTEX_HELD(&spa_namespace_lock));
1672 * Locate the spa's l2arc devices and kick off TRIM threads.
1674 for (int i = 0; i < spa->spa_l2cache.sav_count; i++) {
1675 vdev_t *vd = spa->spa_l2cache.sav_vdevs[i];
1676 l2arc_dev_t *dev = l2arc_vdev_get(vd);
1678 if (dev == NULL || !dev->l2ad_trim_all) {
1680 * Don't attempt TRIM if the vdev is UNAVAIL or if the
1681 * cache device was not marked for whole device TRIM
1682 * (ie l2arc_trim_ahead = 0, or the L2ARC device header
1683 * is valid with trim_state = VDEV_TRIM_COMPLETE and
1684 * l2ad_log_entries > 0).
1689 mutex_enter(&vd->vdev_trim_lock);
1690 ASSERT(vd->vdev_ops->vdev_op_leaf);
1691 ASSERT(vdev_is_concrete(vd));
1692 ASSERT3P(vd->vdev_trim_thread, ==, NULL);
1693 ASSERT(!vd->vdev_detached);
1694 ASSERT(!vd->vdev_trim_exit_wanted);
1695 ASSERT(!vd->vdev_top->vdev_removing);
1696 vdev_trim_change_state(vd, VDEV_TRIM_ACTIVE, 0, 0, 0);
1697 vd->vdev_trim_thread = thread_create(NULL, 0,
1698 vdev_trim_l2arc_thread, vd, 0, &p0, TS_RUN, maxclsyspri);
1699 mutex_exit(&vd->vdev_trim_lock);
1704 * A wrapper which calls vdev_trim_ranges(). It is intended to be called
1708 vdev_trim_simple(vdev_t *vd, uint64_t start, uint64_t size)
1710 trim_args_t ta = {0};
1711 range_seg64_t physical_rs;
1713 physical_rs.rs_start = start;
1714 physical_rs.rs_end = start + size;
1716 ASSERT(vdev_is_concrete(vd));
1717 ASSERT(vd->vdev_ops->vdev_op_leaf);
1718 ASSERT(!vd->vdev_detached);
1719 ASSERT(!vd->vdev_top->vdev_removing);
1722 ta.trim_tree = range_tree_create(NULL, RANGE_SEG64, NULL, 0, 0);
1723 ta.trim_type = TRIM_TYPE_SIMPLE;
1724 ta.trim_extent_bytes_max = zfs_trim_extent_bytes_max;
1725 ta.trim_extent_bytes_min = SPA_MINBLOCKSIZE;
1728 ASSERT3U(physical_rs.rs_end, >=, physical_rs.rs_start);
1730 if (physical_rs.rs_end > physical_rs.rs_start) {
1731 range_tree_add(ta.trim_tree, physical_rs.rs_start,
1732 physical_rs.rs_end - physical_rs.rs_start);
1734 ASSERT3U(physical_rs.rs_end, ==, physical_rs.rs_start);
1737 error = vdev_trim_ranges(&ta);
1739 mutex_enter(&vd->vdev_trim_io_lock);
1740 while (vd->vdev_trim_inflight[TRIM_TYPE_SIMPLE] > 0) {
1741 cv_wait(&vd->vdev_trim_io_cv, &vd->vdev_trim_io_lock);
1743 mutex_exit(&vd->vdev_trim_io_lock);
1745 range_tree_vacate(ta.trim_tree, NULL, NULL);
1746 range_tree_destroy(ta.trim_tree);
1751 EXPORT_SYMBOL(vdev_trim);
1752 EXPORT_SYMBOL(vdev_trim_stop);
1753 EXPORT_SYMBOL(vdev_trim_stop_all);
1754 EXPORT_SYMBOL(vdev_trim_stop_wait);
1755 EXPORT_SYMBOL(vdev_trim_restart);
1756 EXPORT_SYMBOL(vdev_autotrim);
1757 EXPORT_SYMBOL(vdev_autotrim_stop_all);
1758 EXPORT_SYMBOL(vdev_autotrim_stop_wait);
1759 EXPORT_SYMBOL(vdev_autotrim_restart);
1760 EXPORT_SYMBOL(vdev_trim_l2arc);
1761 EXPORT_SYMBOL(vdev_trim_simple);
1763 ZFS_MODULE_PARAM(zfs_trim, zfs_trim_, extent_bytes_max, UINT, ZMOD_RW,
1764 "Max size of TRIM commands, larger will be split");
1766 ZFS_MODULE_PARAM(zfs_trim, zfs_trim_, extent_bytes_min, UINT, ZMOD_RW,
1767 "Min size of TRIM commands, smaller will be skipped");
1769 ZFS_MODULE_PARAM(zfs_trim, zfs_trim_, metaslab_skip, UINT, ZMOD_RW,
1770 "Skip metaslabs which have never been initialized");
1772 ZFS_MODULE_PARAM(zfs_trim, zfs_trim_, txg_batch, UINT, ZMOD_RW,
1773 "Min number of txgs to aggregate frees before issuing TRIM");
1775 ZFS_MODULE_PARAM(zfs_trim, zfs_trim_, queue_limit, UINT, ZMOD_RW,
1776 "Max queued TRIMs outstanding per leaf vdev");