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 http://www.opensolaris.org/os/licensing.
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]
22 * Copyright (c) 2012 Pawel Jakub Dawidek <pawel@dawidek.net>.
23 * All rights reserved.
26 #include <sys/zfs_context.h>
27 #include <sys/spa_impl.h>
28 #include <sys/vdev_impl.h>
29 #include <sys/trim_map.h>
33 * Calculate the zio end, upgrading based on ashift which would be
34 * done by zio_vdev_io_start.
36 * This makes free range consolidation much more effective
37 * than it would otherwise be as well as ensuring that entire
38 * blocks are invalidated by writes.
40 #define TRIM_ZIO_END(vd, offset, size) (offset + \
41 P2ROUNDUP(size, 1ULL << vd->vdev_top->vdev_ashift))
43 /* Maximal segment size for ATA TRIM. */
44 #define TRIM_MAP_SIZE_FACTOR (512 << 16)
46 #define TRIM_MAP_SEGS(size) (1 + (size) / TRIM_MAP_SIZE_FACTOR)
48 #define TRIM_MAP_ADD(tm, ts) do { \
49 list_insert_tail(&(tm)->tm_head, (ts)); \
50 (tm)->tm_pending += TRIM_MAP_SEGS((ts)->ts_end - (ts)->ts_start); \
53 #define TRIM_MAP_REM(tm, ts) do { \
54 list_remove(&(tm)->tm_head, (ts)); \
55 (tm)->tm_pending -= TRIM_MAP_SEGS((ts)->ts_end - (ts)->ts_start); \
58 typedef struct trim_map {
59 list_t tm_head; /* List of segments sorted by txg. */
60 avl_tree_t tm_queued_frees; /* AVL tree of segments waiting for TRIM. */
61 avl_tree_t tm_inflight_frees; /* AVL tree of in-flight TRIMs. */
62 avl_tree_t tm_inflight_writes; /* AVL tree of in-flight writes. */
63 list_t tm_pending_writes; /* Writes blocked on in-flight frees. */
65 uint64_t tm_pending; /* Count of pending TRIMs. */
68 typedef struct trim_seg {
69 avl_node_t ts_node; /* AVL node. */
70 list_node_t ts_next; /* List element. */
71 uint64_t ts_start; /* Starting offset of this segment. */
72 uint64_t ts_end; /* Ending offset (non-inclusive). */
73 uint64_t ts_txg; /* Segment creation txg. */
74 hrtime_t ts_time; /* Segment creation time. */
77 extern boolean_t zfs_trim_enabled;
79 static u_int trim_txg_delay = 32; /* Keep deleted data up to 32 TXG */
80 static u_int trim_timeout = 30; /* Keep deleted data up to 30s */
81 static u_int trim_max_interval = 1; /* 1s delays between TRIMs */
82 static u_int trim_vdev_max_pending = 10000; /* Keep up to 10K segments */
84 SYSCTL_DECL(_vfs_zfs);
85 SYSCTL_NODE(_vfs_zfs, OID_AUTO, trim, CTLFLAG_RD, 0, "ZFS TRIM");
87 TUNABLE_INT("vfs.zfs.trim.txg_delay", &trim_txg_delay);
88 SYSCTL_UINT(_vfs_zfs_trim, OID_AUTO, txg_delay, CTLFLAG_RWTUN, &trim_txg_delay,
89 0, "Delay TRIMs by up to this many TXGs");
91 TUNABLE_INT("vfs.zfs.trim.timeout", &trim_timeout);
92 SYSCTL_UINT(_vfs_zfs_trim, OID_AUTO, timeout, CTLFLAG_RWTUN, &trim_timeout, 0,
93 "Delay TRIMs by up to this many seconds");
95 TUNABLE_INT("vfs.zfs.trim.max_interval", &trim_max_interval);
96 SYSCTL_UINT(_vfs_zfs_trim, OID_AUTO, max_interval, CTLFLAG_RWTUN,
97 &trim_max_interval, 0,
98 "Maximum interval between TRIM queue processing (seconds)");
100 SYSCTL_DECL(_vfs_zfs_vdev);
101 TUNABLE_INT("vfs.zfs.vdev.trim_max_pending", &trim_vdev_max_pending);
102 SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, trim_max_pending, CTLFLAG_RWTUN,
103 &trim_vdev_max_pending, 0,
104 "Maximum pending TRIM segments for a vdev");
107 static void trim_map_vdev_commit_done(spa_t *spa, vdev_t *vd);
110 trim_map_seg_compare(const void *x1, const void *x2)
112 const trim_seg_t *s1 = x1;
113 const trim_seg_t *s2 = x2;
115 if (s1->ts_start < s2->ts_start) {
116 if (s1->ts_end > s2->ts_start)
120 if (s1->ts_start > s2->ts_start) {
121 if (s1->ts_start < s2->ts_end)
129 trim_map_zio_compare(const void *x1, const void *x2)
131 const zio_t *z1 = x1;
132 const zio_t *z2 = x2;
134 if (z1->io_offset < z2->io_offset) {
135 if (z1->io_offset + z1->io_size > z2->io_offset)
139 if (z1->io_offset > z2->io_offset) {
140 if (z1->io_offset < z2->io_offset + z2->io_size)
148 trim_map_create(vdev_t *vd)
152 ASSERT(zfs_trim_enabled && !vd->vdev_notrim &&
153 vd->vdev_ops->vdev_op_leaf);
155 tm = kmem_zalloc(sizeof (*tm), KM_SLEEP);
156 mutex_init(&tm->tm_lock, NULL, MUTEX_DEFAULT, NULL);
157 list_create(&tm->tm_head, sizeof (trim_seg_t),
158 offsetof(trim_seg_t, ts_next));
159 list_create(&tm->tm_pending_writes, sizeof (zio_t),
160 offsetof(zio_t, io_trim_link));
161 avl_create(&tm->tm_queued_frees, trim_map_seg_compare,
162 sizeof (trim_seg_t), offsetof(trim_seg_t, ts_node));
163 avl_create(&tm->tm_inflight_frees, trim_map_seg_compare,
164 sizeof (trim_seg_t), offsetof(trim_seg_t, ts_node));
165 avl_create(&tm->tm_inflight_writes, trim_map_zio_compare,
166 sizeof (zio_t), offsetof(zio_t, io_trim_node));
167 vd->vdev_trimmap = tm;
171 trim_map_destroy(vdev_t *vd)
176 ASSERT(vd->vdev_ops->vdev_op_leaf);
178 if (!zfs_trim_enabled)
181 tm = vd->vdev_trimmap;
186 * We may have been called before trim_map_vdev_commit_done()
187 * had a chance to run, so do it now to prune the remaining
190 trim_map_vdev_commit_done(vd->vdev_spa, vd);
192 mutex_enter(&tm->tm_lock);
193 while ((ts = list_head(&tm->tm_head)) != NULL) {
194 avl_remove(&tm->tm_queued_frees, ts);
195 TRIM_MAP_REM(tm, ts);
196 kmem_free(ts, sizeof (*ts));
198 mutex_exit(&tm->tm_lock);
200 avl_destroy(&tm->tm_queued_frees);
201 avl_destroy(&tm->tm_inflight_frees);
202 avl_destroy(&tm->tm_inflight_writes);
203 list_destroy(&tm->tm_pending_writes);
204 list_destroy(&tm->tm_head);
205 mutex_destroy(&tm->tm_lock);
206 kmem_free(tm, sizeof (*tm));
207 vd->vdev_trimmap = NULL;
211 trim_map_segment_add(trim_map_t *tm, uint64_t start, uint64_t end, uint64_t txg)
214 trim_seg_t tsearch, *ts_before, *ts_after, *ts;
215 boolean_t merge_before, merge_after;
218 ASSERT(MUTEX_HELD(&tm->tm_lock));
222 tsearch.ts_start = start;
223 tsearch.ts_end = end;
225 ts = avl_find(&tm->tm_queued_frees, &tsearch, &where);
227 if (start < ts->ts_start)
228 trim_map_segment_add(tm, start, ts->ts_start, txg);
229 if (end > ts->ts_end)
230 trim_map_segment_add(tm, ts->ts_end, end, txg);
234 ts_before = avl_nearest(&tm->tm_queued_frees, where, AVL_BEFORE);
235 ts_after = avl_nearest(&tm->tm_queued_frees, where, AVL_AFTER);
237 merge_before = (ts_before != NULL && ts_before->ts_end == start);
238 merge_after = (ts_after != NULL && ts_after->ts_start == end);
240 if (merge_before && merge_after) {
241 avl_remove(&tm->tm_queued_frees, ts_before);
242 TRIM_MAP_REM(tm, ts_before);
243 TRIM_MAP_REM(tm, ts_after);
244 ts_after->ts_start = ts_before->ts_start;
245 ts_after->ts_txg = txg;
246 ts_after->ts_time = time;
247 TRIM_MAP_ADD(tm, ts_after);
248 kmem_free(ts_before, sizeof (*ts_before));
249 } else if (merge_before) {
250 TRIM_MAP_REM(tm, ts_before);
251 ts_before->ts_end = end;
252 ts_before->ts_txg = txg;
253 ts_before->ts_time = time;
254 TRIM_MAP_ADD(tm, ts_before);
255 } else if (merge_after) {
256 TRIM_MAP_REM(tm, ts_after);
257 ts_after->ts_start = start;
258 ts_after->ts_txg = txg;
259 ts_after->ts_time = time;
260 TRIM_MAP_ADD(tm, ts_after);
262 ts = kmem_alloc(sizeof (*ts), KM_SLEEP);
263 ts->ts_start = start;
267 avl_insert(&tm->tm_queued_frees, ts, where);
268 TRIM_MAP_ADD(tm, ts);
273 trim_map_segment_remove(trim_map_t *tm, trim_seg_t *ts, uint64_t start,
277 boolean_t left_over, right_over;
279 ASSERT(MUTEX_HELD(&tm->tm_lock));
281 left_over = (ts->ts_start < start);
282 right_over = (ts->ts_end > end);
284 TRIM_MAP_REM(tm, ts);
285 if (left_over && right_over) {
286 nts = kmem_alloc(sizeof (*nts), KM_SLEEP);
288 nts->ts_end = ts->ts_end;
289 nts->ts_txg = ts->ts_txg;
290 nts->ts_time = ts->ts_time;
292 avl_insert_here(&tm->tm_queued_frees, nts, ts, AVL_AFTER);
293 TRIM_MAP_ADD(tm, ts);
294 TRIM_MAP_ADD(tm, nts);
295 } else if (left_over) {
297 TRIM_MAP_ADD(tm, ts);
298 } else if (right_over) {
300 TRIM_MAP_ADD(tm, ts);
302 avl_remove(&tm->tm_queued_frees, ts);
303 kmem_free(ts, sizeof (*ts));
308 trim_map_free_locked(trim_map_t *tm, uint64_t start, uint64_t end, uint64_t txg)
312 ASSERT(MUTEX_HELD(&tm->tm_lock));
314 zsearch.io_offset = start;
315 zsearch.io_size = end - start;
317 zs = avl_find(&tm->tm_inflight_writes, &zsearch, NULL);
319 trim_map_segment_add(tm, start, end, txg);
322 if (start < zs->io_offset)
323 trim_map_free_locked(tm, start, zs->io_offset, txg);
324 if (zs->io_offset + zs->io_size < end)
325 trim_map_free_locked(tm, zs->io_offset + zs->io_size, end, txg);
329 trim_map_free(vdev_t *vd, uint64_t offset, uint64_t size, uint64_t txg)
331 trim_map_t *tm = vd->vdev_trimmap;
333 if (!zfs_trim_enabled || vd->vdev_notrim || tm == NULL)
336 mutex_enter(&tm->tm_lock);
337 trim_map_free_locked(tm, offset, TRIM_ZIO_END(vd, offset, size), txg);
338 mutex_exit(&tm->tm_lock);
342 trim_map_write_start(zio_t *zio)
344 vdev_t *vd = zio->io_vd;
345 trim_map_t *tm = vd->vdev_trimmap;
346 trim_seg_t tsearch, *ts;
347 boolean_t left_over, right_over;
350 if (!zfs_trim_enabled || vd->vdev_notrim || tm == NULL)
353 start = zio->io_offset;
354 end = TRIM_ZIO_END(zio->io_vd, start, zio->io_size);
355 tsearch.ts_start = start;
356 tsearch.ts_end = end;
358 mutex_enter(&tm->tm_lock);
361 * Checking for colliding in-flight frees.
363 ts = avl_find(&tm->tm_inflight_frees, &tsearch, NULL);
365 list_insert_tail(&tm->tm_pending_writes, zio);
366 mutex_exit(&tm->tm_lock);
370 ts = avl_find(&tm->tm_queued_frees, &tsearch, NULL);
373 * Loop until all overlapping segments are removed.
376 trim_map_segment_remove(tm, ts, start, end);
377 ts = avl_find(&tm->tm_queued_frees, &tsearch, NULL);
378 } while (ts != NULL);
380 avl_add(&tm->tm_inflight_writes, zio);
382 mutex_exit(&tm->tm_lock);
388 trim_map_write_done(zio_t *zio)
390 vdev_t *vd = zio->io_vd;
391 trim_map_t *tm = vd->vdev_trimmap;
394 * Don't check for vdev_notrim, since the write could have
395 * started before vdev_notrim was set.
397 if (!zfs_trim_enabled || tm == NULL)
400 mutex_enter(&tm->tm_lock);
402 * Don't fail if the write isn't in the tree, since the write
403 * could have started after vdev_notrim was set.
405 if (zio->io_trim_node.avl_child[0] ||
406 zio->io_trim_node.avl_child[1] ||
407 AVL_XPARENT(&zio->io_trim_node) ||
408 tm->tm_inflight_writes.avl_root == &zio->io_trim_node)
409 avl_remove(&tm->tm_inflight_writes, zio);
410 mutex_exit(&tm->tm_lock);
414 * Return the oldest segment (the one with the lowest txg / time) or NULL if:
415 * 1. The list is empty
416 * 2. The first element's txg is greater than txgsafe
417 * 3. The first element's txg is not greater than the txg argument and the
418 * the first element's time is not greater than time argument
421 trim_map_first(trim_map_t *tm, uint64_t txg, uint64_t txgsafe, hrtime_t time,
426 ASSERT(MUTEX_HELD(&tm->tm_lock));
427 VERIFY(txgsafe >= txg);
429 ts = list_head(&tm->tm_head);
430 if (ts != NULL && ts->ts_txg <= txgsafe &&
431 (ts->ts_txg <= txg || ts->ts_time <= time || force))
437 trim_map_vdev_commit(spa_t *spa, zio_t *zio, vdev_t *vd)
439 trim_map_t *tm = vd->vdev_trimmap;
441 uint64_t size, offset, txgtarget, txgsafe;
445 ASSERT(vd->vdev_ops->vdev_op_leaf);
450 timelimit = gethrtime() - (hrtime_t)trim_timeout * NANOSEC;
451 if (vd->vdev_isl2cache) {
452 txgsafe = UINT64_MAX;
453 txgtarget = UINT64_MAX;
455 txgsafe = MIN(spa_last_synced_txg(spa), spa_freeze_txg(spa));
456 if (txgsafe > trim_txg_delay)
457 txgtarget = txgsafe - trim_txg_delay;
462 mutex_enter(&tm->tm_lock);
464 if (tm->tm_pending > trim_vdev_max_pending)
465 hard = (tm->tm_pending - trim_vdev_max_pending) / 4;
466 soft = P2ROUNDUP(hard + tm->tm_pending / trim_timeout + 1, 64);
467 /* Loop until we have sent all outstanding free's */
469 (ts = trim_map_first(tm, txgtarget, txgsafe, timelimit, hard > 0))
471 TRIM_MAP_REM(tm, ts);
472 avl_remove(&tm->tm_queued_frees, ts);
473 avl_add(&tm->tm_inflight_frees, ts);
474 size = ts->ts_end - ts->ts_start;
475 offset = ts->ts_start;
477 * We drop the lock while we call zio_nowait as the IO
478 * scheduler can result in a different IO being run e.g.
479 * a write which would result in a recursive lock.
481 mutex_exit(&tm->tm_lock);
483 zio_nowait(zio_trim(zio, spa, vd, offset, size));
485 soft -= TRIM_MAP_SEGS(size);
486 hard -= TRIM_MAP_SEGS(size);
487 mutex_enter(&tm->tm_lock);
489 mutex_exit(&tm->tm_lock);
493 trim_map_vdev_commit_done(spa_t *spa, vdev_t *vd)
495 trim_map_t *tm = vd->vdev_trimmap;
497 list_t pending_writes;
499 uint64_t start, size;
502 ASSERT(vd->vdev_ops->vdev_op_leaf);
507 mutex_enter(&tm->tm_lock);
508 if (!avl_is_empty(&tm->tm_inflight_frees)) {
510 while ((ts = avl_destroy_nodes(&tm->tm_inflight_frees,
512 kmem_free(ts, sizeof (*ts));
515 list_create(&pending_writes, sizeof (zio_t), offsetof(zio_t,
517 list_move_tail(&pending_writes, &tm->tm_pending_writes);
518 mutex_exit(&tm->tm_lock);
520 while ((zio = list_remove_head(&pending_writes)) != NULL) {
521 zio_vdev_io_reissue(zio);
524 list_destroy(&pending_writes);
528 trim_map_commit(spa_t *spa, zio_t *zio, vdev_t *vd)
535 if (vd->vdev_ops->vdev_op_leaf) {
536 trim_map_vdev_commit(spa, zio, vd);
538 for (c = 0; c < vd->vdev_children; c++)
539 trim_map_commit(spa, zio, vd->vdev_child[c]);
544 trim_map_commit_done(spa_t *spa, vdev_t *vd)
551 if (vd->vdev_ops->vdev_op_leaf) {
552 trim_map_vdev_commit_done(spa, vd);
554 for (c = 0; c < vd->vdev_children; c++)
555 trim_map_commit_done(spa, vd->vdev_child[c]);
560 trim_thread(void *arg)
566 (void) snprintf(curthread->td_name, sizeof(curthread->td_name),
567 "trim %s", spa_name(spa));
571 mutex_enter(&spa->spa_trim_lock);
572 if (spa->spa_trim_thread == NULL) {
573 spa->spa_trim_thread = curthread;
574 cv_signal(&spa->spa_trim_cv);
575 mutex_exit(&spa->spa_trim_lock);
579 (void) cv_timedwait(&spa->spa_trim_cv, &spa->spa_trim_lock,
580 hz * trim_max_interval);
581 mutex_exit(&spa->spa_trim_lock);
583 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
585 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
586 trim_map_commit(spa, zio, spa->spa_root_vdev);
587 (void) zio_wait(zio);
588 trim_map_commit_done(spa, spa->spa_root_vdev);
589 spa_config_exit(spa, SCL_STATE, FTAG);
594 trim_thread_create(spa_t *spa)
597 if (!zfs_trim_enabled)
600 mutex_init(&spa->spa_trim_lock, NULL, MUTEX_DEFAULT, NULL);
601 cv_init(&spa->spa_trim_cv, NULL, CV_DEFAULT, NULL);
602 mutex_enter(&spa->spa_trim_lock);
603 spa->spa_trim_thread = thread_create(NULL, 0, trim_thread, spa, 0, &p0,
604 TS_RUN, minclsyspri);
605 mutex_exit(&spa->spa_trim_lock);
609 trim_thread_destroy(spa_t *spa)
612 if (!zfs_trim_enabled)
614 if (spa->spa_trim_thread == NULL)
617 mutex_enter(&spa->spa_trim_lock);
618 /* Setting spa_trim_thread to NULL tells the thread to stop. */
619 spa->spa_trim_thread = NULL;
620 cv_signal(&spa->spa_trim_cv);
621 /* The thread will set it back to != NULL on exit. */
622 while (spa->spa_trim_thread == NULL)
623 cv_wait(&spa->spa_trim_cv, &spa->spa_trim_lock);
624 spa->spa_trim_thread = NULL;
625 mutex_exit(&spa->spa_trim_lock);
627 cv_destroy(&spa->spa_trim_cv);
628 mutex_destroy(&spa->spa_trim_lock);
632 trim_thread_wakeup(spa_t *spa)
635 if (!zfs_trim_enabled)
637 if (spa->spa_trim_thread == NULL)
640 mutex_enter(&spa->spa_trim_lock);
641 cv_signal(&spa->spa_trim_cv);
642 mutex_exit(&spa->spa_trim_lock);