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 #define TRIM_MAP_SINC(tm, size) \
44 atomic_add_64(&(tm)->tm_bytes, (size))
46 #define TRIM_MAP_SDEC(tm, size) \
47 atomic_add_64(&(tm)->tm_bytes, -(size))
49 #define TRIM_MAP_QINC(tm) \
50 atomic_inc_64(&(tm)->tm_pending); \
52 #define TRIM_MAP_QDEC(tm) \
53 atomic_dec_64(&(tm)->tm_pending);
55 typedef struct trim_map {
56 list_t tm_head; /* List of segments sorted by txg. */
57 avl_tree_t tm_queued_frees; /* AVL tree of segments waiting for TRIM. */
58 avl_tree_t tm_inflight_frees; /* AVL tree of in-flight TRIMs. */
59 avl_tree_t tm_inflight_writes; /* AVL tree of in-flight writes. */
60 list_t tm_pending_writes; /* Writes blocked on in-flight frees. */
62 uint64_t tm_pending; /* Count of pending TRIMs. */
63 uint64_t tm_bytes; /* Total size in bytes of queued TRIMs. */
66 typedef struct trim_seg {
67 avl_node_t ts_node; /* AVL node. */
68 list_node_t ts_next; /* List element. */
69 uint64_t ts_start; /* Starting offset of this segment. */
70 uint64_t ts_end; /* Ending offset (non-inclusive). */
71 uint64_t ts_txg; /* Segment creation txg. */
72 hrtime_t ts_time; /* Segment creation time. */
75 extern boolean_t zfs_trim_enabled;
77 static u_int trim_txg_delay = 32;
78 static u_int trim_timeout = 30;
79 static u_int trim_max_interval = 1;
80 /* Limit outstanding TRIMs to 2G (max size for a single TRIM request) */
81 static uint64_t trim_vdev_max_bytes = 2147483648;
82 /* Limit outstanding TRIMs to 64 (max ranges for a single TRIM request) */
83 static u_int trim_vdev_max_pending = 64;
85 SYSCTL_DECL(_vfs_zfs);
86 SYSCTL_NODE(_vfs_zfs, OID_AUTO, trim, CTLFLAG_RD, 0, "ZFS TRIM");
88 TUNABLE_INT("vfs.zfs.trim.txg_delay", &trim_txg_delay);
89 SYSCTL_UINT(_vfs_zfs_trim, OID_AUTO, txg_delay, CTLFLAG_RWTUN, &trim_txg_delay,
90 0, "Delay TRIMs by up to this many TXGs");
92 TUNABLE_INT("vfs.zfs.trim.timeout", &trim_timeout);
93 SYSCTL_UINT(_vfs_zfs_trim, OID_AUTO, timeout, CTLFLAG_RWTUN, &trim_timeout, 0,
94 "Delay TRIMs by up to this many seconds");
96 TUNABLE_INT("vfs.zfs.trim.max_interval", &trim_max_interval);
97 SYSCTL_UINT(_vfs_zfs_trim, OID_AUTO, max_interval, CTLFLAG_RWTUN,
98 &trim_max_interval, 0,
99 "Maximum interval between TRIM queue processing (seconds)");
101 SYSCTL_DECL(_vfs_zfs_vdev);
102 TUNABLE_QUAD("vfs.zfs.vdev.trim_max_bytes", &trim_vdev_max_bytes);
103 SYSCTL_QUAD(_vfs_zfs_vdev, OID_AUTO, trim_max_bytes, CTLFLAG_RWTUN,
104 &trim_vdev_max_bytes, 0,
105 "Maximum pending TRIM bytes for a vdev");
107 TUNABLE_INT("vfs.zfs.vdev.trim_max_pending", &trim_vdev_max_pending);
108 SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, trim_max_pending, CTLFLAG_RWTUN,
109 &trim_vdev_max_pending, 0,
110 "Maximum pending TRIM segments for a vdev");
113 static void trim_map_vdev_commit_done(spa_t *spa, vdev_t *vd);
116 trim_map_seg_compare(const void *x1, const void *x2)
118 const trim_seg_t *s1 = x1;
119 const trim_seg_t *s2 = x2;
121 if (s1->ts_start < s2->ts_start) {
122 if (s1->ts_end > s2->ts_start)
126 if (s1->ts_start > s2->ts_start) {
127 if (s1->ts_start < s2->ts_end)
135 trim_map_zio_compare(const void *x1, const void *x2)
137 const zio_t *z1 = x1;
138 const zio_t *z2 = x2;
140 if (z1->io_offset < z2->io_offset) {
141 if (z1->io_offset + z1->io_size > z2->io_offset)
145 if (z1->io_offset > z2->io_offset) {
146 if (z1->io_offset < z2->io_offset + z2->io_size)
154 trim_map_create(vdev_t *vd)
158 ASSERT(vd->vdev_ops->vdev_op_leaf);
160 if (!zfs_trim_enabled)
163 tm = kmem_zalloc(sizeof (*tm), KM_SLEEP);
164 mutex_init(&tm->tm_lock, NULL, MUTEX_DEFAULT, NULL);
165 list_create(&tm->tm_head, sizeof (trim_seg_t),
166 offsetof(trim_seg_t, ts_next));
167 list_create(&tm->tm_pending_writes, sizeof (zio_t),
168 offsetof(zio_t, io_trim_link));
169 avl_create(&tm->tm_queued_frees, trim_map_seg_compare,
170 sizeof (trim_seg_t), offsetof(trim_seg_t, ts_node));
171 avl_create(&tm->tm_inflight_frees, trim_map_seg_compare,
172 sizeof (trim_seg_t), offsetof(trim_seg_t, ts_node));
173 avl_create(&tm->tm_inflight_writes, trim_map_zio_compare,
174 sizeof (zio_t), offsetof(zio_t, io_trim_node));
175 vd->vdev_trimmap = tm;
179 trim_map_destroy(vdev_t *vd)
184 ASSERT(vd->vdev_ops->vdev_op_leaf);
186 if (!zfs_trim_enabled)
189 tm = vd->vdev_trimmap;
194 * We may have been called before trim_map_vdev_commit_done()
195 * had a chance to run, so do it now to prune the remaining
198 trim_map_vdev_commit_done(vd->vdev_spa, vd);
200 mutex_enter(&tm->tm_lock);
201 while ((ts = list_head(&tm->tm_head)) != NULL) {
202 avl_remove(&tm->tm_queued_frees, ts);
203 list_remove(&tm->tm_head, ts);
204 kmem_free(ts, sizeof (*ts));
205 TRIM_MAP_SDEC(tm, ts->ts_end - ts->ts_start);
208 mutex_exit(&tm->tm_lock);
210 avl_destroy(&tm->tm_queued_frees);
211 avl_destroy(&tm->tm_inflight_frees);
212 avl_destroy(&tm->tm_inflight_writes);
213 list_destroy(&tm->tm_pending_writes);
214 list_destroy(&tm->tm_head);
215 mutex_destroy(&tm->tm_lock);
216 kmem_free(tm, sizeof (*tm));
217 vd->vdev_trimmap = NULL;
221 trim_map_segment_add(trim_map_t *tm, uint64_t start, uint64_t end, uint64_t txg)
224 trim_seg_t tsearch, *ts_before, *ts_after, *ts;
225 boolean_t merge_before, merge_after;
228 ASSERT(MUTEX_HELD(&tm->tm_lock));
232 tsearch.ts_start = start;
233 tsearch.ts_end = end;
235 ts = avl_find(&tm->tm_queued_frees, &tsearch, &where);
237 if (start < ts->ts_start)
238 trim_map_segment_add(tm, start, ts->ts_start, txg);
239 if (end > ts->ts_end)
240 trim_map_segment_add(tm, ts->ts_end, end, txg);
244 ts_before = avl_nearest(&tm->tm_queued_frees, where, AVL_BEFORE);
245 ts_after = avl_nearest(&tm->tm_queued_frees, where, AVL_AFTER);
247 merge_before = (ts_before != NULL && ts_before->ts_end == start);
248 merge_after = (ts_after != NULL && ts_after->ts_start == end);
250 if (merge_before && merge_after) {
251 TRIM_MAP_SINC(tm, ts_after->ts_start - ts_before->ts_end);
253 avl_remove(&tm->tm_queued_frees, ts_before);
254 list_remove(&tm->tm_head, ts_before);
255 ts_after->ts_start = ts_before->ts_start;
256 ts_after->ts_txg = txg;
257 ts_after->ts_time = time;
258 kmem_free(ts_before, sizeof (*ts_before));
259 } else if (merge_before) {
260 TRIM_MAP_SINC(tm, end - ts_before->ts_end);
261 ts_before->ts_end = end;
262 ts_before->ts_txg = txg;
263 ts_before->ts_time = time;
264 } else if (merge_after) {
265 TRIM_MAP_SINC(tm, ts_after->ts_start - start);
266 ts_after->ts_start = start;
267 ts_after->ts_txg = txg;
268 ts_after->ts_time = time;
270 TRIM_MAP_SINC(tm, end - start);
272 ts = kmem_alloc(sizeof (*ts), KM_SLEEP);
273 ts->ts_start = start;
277 avl_insert(&tm->tm_queued_frees, ts, where);
278 list_insert_tail(&tm->tm_head, ts);
283 trim_map_segment_remove(trim_map_t *tm, trim_seg_t *ts, uint64_t start,
287 boolean_t left_over, right_over;
289 ASSERT(MUTEX_HELD(&tm->tm_lock));
291 left_over = (ts->ts_start < start);
292 right_over = (ts->ts_end > end);
294 TRIM_MAP_SDEC(tm, end - start);
295 if (left_over && right_over) {
296 nts = kmem_alloc(sizeof (*nts), KM_SLEEP);
298 nts->ts_end = ts->ts_end;
299 nts->ts_txg = ts->ts_txg;
300 nts->ts_time = ts->ts_time;
302 avl_insert_here(&tm->tm_queued_frees, nts, ts, AVL_AFTER);
303 list_insert_after(&tm->tm_head, ts, nts);
305 } else if (left_over) {
307 } else if (right_over) {
310 avl_remove(&tm->tm_queued_frees, ts);
311 list_remove(&tm->tm_head, ts);
313 kmem_free(ts, sizeof (*ts));
318 trim_map_free_locked(trim_map_t *tm, uint64_t start, uint64_t end, uint64_t txg)
322 ASSERT(MUTEX_HELD(&tm->tm_lock));
324 zsearch.io_offset = start;
325 zsearch.io_size = end - start;
327 zs = avl_find(&tm->tm_inflight_writes, &zsearch, NULL);
329 trim_map_segment_add(tm, start, end, txg);
332 if (start < zs->io_offset)
333 trim_map_free_locked(tm, start, zs->io_offset, txg);
334 if (zs->io_offset + zs->io_size < end)
335 trim_map_free_locked(tm, zs->io_offset + zs->io_size, end, txg);
339 trim_map_free(vdev_t *vd, uint64_t offset, uint64_t size, uint64_t txg)
341 trim_map_t *tm = vd->vdev_trimmap;
343 if (!zfs_trim_enabled || vd->vdev_notrim || tm == NULL)
346 mutex_enter(&tm->tm_lock);
347 trim_map_free_locked(tm, offset, TRIM_ZIO_END(vd, offset, size), txg);
348 mutex_exit(&tm->tm_lock);
352 trim_map_write_start(zio_t *zio)
354 vdev_t *vd = zio->io_vd;
355 trim_map_t *tm = vd->vdev_trimmap;
356 trim_seg_t tsearch, *ts;
357 boolean_t left_over, right_over;
360 if (!zfs_trim_enabled || vd->vdev_notrim || tm == NULL)
363 start = zio->io_offset;
364 end = TRIM_ZIO_END(zio->io_vd, start, zio->io_size);
365 tsearch.ts_start = start;
366 tsearch.ts_end = end;
368 mutex_enter(&tm->tm_lock);
371 * Checking for colliding in-flight frees.
373 ts = avl_find(&tm->tm_inflight_frees, &tsearch, NULL);
375 list_insert_tail(&tm->tm_pending_writes, zio);
376 mutex_exit(&tm->tm_lock);
380 ts = avl_find(&tm->tm_queued_frees, &tsearch, NULL);
383 * Loop until all overlapping segments are removed.
386 trim_map_segment_remove(tm, ts, start, end);
387 ts = avl_find(&tm->tm_queued_frees, &tsearch, NULL);
388 } while (ts != NULL);
390 avl_add(&tm->tm_inflight_writes, zio);
392 mutex_exit(&tm->tm_lock);
398 trim_map_write_done(zio_t *zio)
400 vdev_t *vd = zio->io_vd;
401 trim_map_t *tm = vd->vdev_trimmap;
404 * Don't check for vdev_notrim, since the write could have
405 * started before vdev_notrim was set.
407 if (!zfs_trim_enabled || tm == NULL)
410 mutex_enter(&tm->tm_lock);
412 * Don't fail if the write isn't in the tree, since the write
413 * could have started after vdev_notrim was set.
415 if (zio->io_trim_node.avl_child[0] ||
416 zio->io_trim_node.avl_child[1] ||
417 AVL_XPARENT(&zio->io_trim_node) ||
418 tm->tm_inflight_writes.avl_root == &zio->io_trim_node)
419 avl_remove(&tm->tm_inflight_writes, zio);
420 mutex_exit(&tm->tm_lock);
424 * Return the oldest segment (the one with the lowest txg / time) or NULL if:
425 * 1. The list is empty
426 * 2. The first element's txg is greater than txgsafe
427 * 3. The first element's txg is not greater than the txg argument and the
428 * the first element's time is not greater than time argument
431 trim_map_first(trim_map_t *tm, uint64_t txg, uint64_t txgsafe, hrtime_t time)
435 ASSERT(MUTEX_HELD(&tm->tm_lock));
436 VERIFY(txgsafe >= txg);
438 ts = list_head(&tm->tm_head);
439 if (ts != NULL && ts->ts_txg <= txgsafe &&
440 (ts->ts_txg <= txg || ts->ts_time <= time ||
441 tm->tm_bytes > trim_vdev_max_bytes ||
442 tm->tm_pending > trim_vdev_max_pending))
448 trim_map_vdev_commit(spa_t *spa, zio_t *zio, vdev_t *vd)
450 trim_map_t *tm = vd->vdev_trimmap;
452 uint64_t size, txgtarget, txgsafe;
455 ASSERT(vd->vdev_ops->vdev_op_leaf);
460 timelimit = gethrtime() - trim_timeout * NANOSEC;
461 if (vd->vdev_isl2cache) {
462 txgsafe = UINT64_MAX;
463 txgtarget = UINT64_MAX;
465 txgsafe = MIN(spa_last_synced_txg(spa), spa_freeze_txg(spa));
466 if (txgsafe > trim_txg_delay)
467 txgtarget = txgsafe - trim_txg_delay;
472 mutex_enter(&tm->tm_lock);
473 /* Loop until we have sent all outstanding free's */
474 while ((ts = trim_map_first(tm, txgtarget, txgsafe, timelimit))
476 list_remove(&tm->tm_head, ts);
477 avl_remove(&tm->tm_queued_frees, ts);
478 avl_add(&tm->tm_inflight_frees, ts);
479 size = ts->ts_end - ts->ts_start;
480 zio_nowait(zio_trim(zio, spa, vd, ts->ts_start, size));
481 TRIM_MAP_SDEC(tm, size);
484 mutex_exit(&tm->tm_lock);
488 trim_map_vdev_commit_done(spa_t *spa, vdev_t *vd)
490 trim_map_t *tm = vd->vdev_trimmap;
492 list_t pending_writes;
494 uint64_t start, size;
497 ASSERT(vd->vdev_ops->vdev_op_leaf);
502 mutex_enter(&tm->tm_lock);
503 if (!avl_is_empty(&tm->tm_inflight_frees)) {
505 while ((ts = avl_destroy_nodes(&tm->tm_inflight_frees,
507 kmem_free(ts, sizeof (*ts));
510 list_create(&pending_writes, sizeof (zio_t), offsetof(zio_t,
512 list_move_tail(&pending_writes, &tm->tm_pending_writes);
513 mutex_exit(&tm->tm_lock);
515 while ((zio = list_remove_head(&pending_writes)) != NULL) {
516 zio_vdev_io_reissue(zio);
519 list_destroy(&pending_writes);
523 trim_map_commit(spa_t *spa, zio_t *zio, vdev_t *vd)
530 if (vd->vdev_ops->vdev_op_leaf) {
531 trim_map_vdev_commit(spa, zio, vd);
533 for (c = 0; c < vd->vdev_children; c++)
534 trim_map_commit(spa, zio, vd->vdev_child[c]);
539 trim_map_commit_done(spa_t *spa, vdev_t *vd)
546 if (vd->vdev_ops->vdev_op_leaf) {
547 trim_map_vdev_commit_done(spa, vd);
549 for (c = 0; c < vd->vdev_children; c++)
550 trim_map_commit_done(spa, vd->vdev_child[c]);
555 trim_thread(void *arg)
561 (void) snprintf(curthread->td_name, sizeof(curthread->td_name),
562 "trim %s", spa_name(spa));
566 mutex_enter(&spa->spa_trim_lock);
567 if (spa->spa_trim_thread == NULL) {
568 spa->spa_trim_thread = curthread;
569 cv_signal(&spa->spa_trim_cv);
570 mutex_exit(&spa->spa_trim_lock);
574 (void) cv_timedwait(&spa->spa_trim_cv, &spa->spa_trim_lock,
575 hz * trim_max_interval);
576 mutex_exit(&spa->spa_trim_lock);
578 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
580 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
581 trim_map_commit(spa, zio, spa->spa_root_vdev);
582 (void) zio_wait(zio);
583 trim_map_commit_done(spa, spa->spa_root_vdev);
584 spa_config_exit(spa, SCL_STATE, FTAG);
589 trim_thread_create(spa_t *spa)
592 if (!zfs_trim_enabled)
595 mutex_init(&spa->spa_trim_lock, NULL, MUTEX_DEFAULT, NULL);
596 cv_init(&spa->spa_trim_cv, NULL, CV_DEFAULT, NULL);
597 mutex_enter(&spa->spa_trim_lock);
598 spa->spa_trim_thread = thread_create(NULL, 0, trim_thread, spa, 0, &p0,
599 TS_RUN, minclsyspri);
600 mutex_exit(&spa->spa_trim_lock);
604 trim_thread_destroy(spa_t *spa)
607 if (!zfs_trim_enabled)
609 if (spa->spa_trim_thread == NULL)
612 mutex_enter(&spa->spa_trim_lock);
613 /* Setting spa_trim_thread to NULL tells the thread to stop. */
614 spa->spa_trim_thread = NULL;
615 cv_signal(&spa->spa_trim_cv);
616 /* The thread will set it back to != NULL on exit. */
617 while (spa->spa_trim_thread == NULL)
618 cv_wait(&spa->spa_trim_cv, &spa->spa_trim_lock);
619 spa->spa_trim_thread = NULL;
620 mutex_exit(&spa->spa_trim_lock);
622 cv_destroy(&spa->spa_trim_cv);
623 mutex_destroy(&spa->spa_trim_lock);
627 trim_thread_wakeup(spa_t *spa)
630 if (!zfs_trim_enabled)
632 if (spa->spa_trim_thread == NULL)
635 mutex_enter(&spa->spa_trim_lock);
636 cv_signal(&spa->spa_trim_cv);
637 mutex_exit(&spa->spa_trim_lock);