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 SYSCTL_UINT(_vfs_zfs_trim, OID_AUTO, txg_delay, CTLFLAG_RWTUN, &trim_txg_delay,
89 0, "Delay TRIMs by up to this many TXGs");
90 SYSCTL_UINT(_vfs_zfs_trim, OID_AUTO, timeout, CTLFLAG_RWTUN, &trim_timeout, 0,
91 "Delay TRIMs by up to this many seconds");
92 SYSCTL_UINT(_vfs_zfs_trim, OID_AUTO, max_interval, CTLFLAG_RWTUN,
93 &trim_max_interval, 0,
94 "Maximum interval between TRIM queue processing (seconds)");
96 SYSCTL_DECL(_vfs_zfs_vdev);
97 SYSCTL_QUAD(_vfs_zfs_vdev, OID_AUTO, trim_max_bytes, CTLFLAG_RWTUN,
98 &trim_vdev_max_bytes, 0,
99 "Maximum pending TRIM bytes for a vdev");
100 SYSCTL_UINT(_vfs_zfs_vdev, OID_AUTO, trim_max_pending, CTLFLAG_RWTUN,
101 &trim_vdev_max_pending, 0,
102 "Maximum pending TRIM segments for a vdev");
104 static void trim_map_vdev_commit_done(spa_t *spa, vdev_t *vd);
107 trim_map_seg_compare(const void *x1, const void *x2)
109 const trim_seg_t *s1 = x1;
110 const trim_seg_t *s2 = x2;
112 if (s1->ts_start < s2->ts_start) {
113 if (s1->ts_end > s2->ts_start)
117 if (s1->ts_start > s2->ts_start) {
118 if (s1->ts_start < s2->ts_end)
126 trim_map_zio_compare(const void *x1, const void *x2)
128 const zio_t *z1 = x1;
129 const zio_t *z2 = x2;
131 if (z1->io_offset < z2->io_offset) {
132 if (z1->io_offset + z1->io_size > z2->io_offset)
136 if (z1->io_offset > z2->io_offset) {
137 if (z1->io_offset < z2->io_offset + z2->io_size)
145 trim_map_create(vdev_t *vd)
149 ASSERT(zfs_trim_enabled && !vd->vdev_notrim &&
150 vd->vdev_ops->vdev_op_leaf);
152 tm = kmem_zalloc(sizeof (*tm), KM_SLEEP);
153 mutex_init(&tm->tm_lock, NULL, MUTEX_DEFAULT, NULL);
154 list_create(&tm->tm_head, sizeof (trim_seg_t),
155 offsetof(trim_seg_t, ts_next));
156 list_create(&tm->tm_pending_writes, sizeof (zio_t),
157 offsetof(zio_t, io_trim_link));
158 avl_create(&tm->tm_queued_frees, trim_map_seg_compare,
159 sizeof (trim_seg_t), offsetof(trim_seg_t, ts_node));
160 avl_create(&tm->tm_inflight_frees, trim_map_seg_compare,
161 sizeof (trim_seg_t), offsetof(trim_seg_t, ts_node));
162 avl_create(&tm->tm_inflight_writes, trim_map_zio_compare,
163 sizeof (zio_t), offsetof(zio_t, io_trim_node));
164 vd->vdev_trimmap = tm;
168 trim_map_destroy(vdev_t *vd)
173 ASSERT(vd->vdev_ops->vdev_op_leaf);
175 if (!zfs_trim_enabled)
178 tm = vd->vdev_trimmap;
183 * We may have been called before trim_map_vdev_commit_done()
184 * had a chance to run, so do it now to prune the remaining
187 trim_map_vdev_commit_done(vd->vdev_spa, vd);
189 mutex_enter(&tm->tm_lock);
190 while ((ts = list_head(&tm->tm_head)) != NULL) {
191 avl_remove(&tm->tm_queued_frees, ts);
192 list_remove(&tm->tm_head, ts);
193 kmem_free(ts, sizeof (*ts));
194 TRIM_MAP_SDEC(tm, ts->ts_end - ts->ts_start);
197 mutex_exit(&tm->tm_lock);
199 avl_destroy(&tm->tm_queued_frees);
200 avl_destroy(&tm->tm_inflight_frees);
201 avl_destroy(&tm->tm_inflight_writes);
202 list_destroy(&tm->tm_pending_writes);
203 list_destroy(&tm->tm_head);
204 mutex_destroy(&tm->tm_lock);
205 kmem_free(tm, sizeof (*tm));
206 vd->vdev_trimmap = NULL;
210 trim_map_segment_add(trim_map_t *tm, uint64_t start, uint64_t end, uint64_t txg)
213 trim_seg_t tsearch, *ts_before, *ts_after, *ts;
214 boolean_t merge_before, merge_after;
217 ASSERT(MUTEX_HELD(&tm->tm_lock));
221 tsearch.ts_start = start;
222 tsearch.ts_end = end;
224 ts = avl_find(&tm->tm_queued_frees, &tsearch, &where);
226 if (start < ts->ts_start)
227 trim_map_segment_add(tm, start, ts->ts_start, txg);
228 if (end > ts->ts_end)
229 trim_map_segment_add(tm, ts->ts_end, end, txg);
233 ts_before = avl_nearest(&tm->tm_queued_frees, where, AVL_BEFORE);
234 ts_after = avl_nearest(&tm->tm_queued_frees, where, AVL_AFTER);
236 merge_before = (ts_before != NULL && ts_before->ts_end == start);
237 merge_after = (ts_after != NULL && ts_after->ts_start == end);
239 if (merge_before && merge_after) {
240 TRIM_MAP_SINC(tm, ts_after->ts_start - ts_before->ts_end);
242 avl_remove(&tm->tm_queued_frees, ts_before);
243 list_remove(&tm->tm_head, ts_before);
244 ts_after->ts_start = ts_before->ts_start;
245 ts_after->ts_txg = txg;
246 ts_after->ts_time = time;
247 kmem_free(ts_before, sizeof (*ts_before));
248 } else if (merge_before) {
249 TRIM_MAP_SINC(tm, end - ts_before->ts_end);
250 ts_before->ts_end = end;
251 ts_before->ts_txg = txg;
252 ts_before->ts_time = time;
253 } else if (merge_after) {
254 TRIM_MAP_SINC(tm, ts_after->ts_start - start);
255 ts_after->ts_start = start;
256 ts_after->ts_txg = txg;
257 ts_after->ts_time = time;
259 TRIM_MAP_SINC(tm, end - start);
261 ts = kmem_alloc(sizeof (*ts), KM_SLEEP);
262 ts->ts_start = start;
266 avl_insert(&tm->tm_queued_frees, ts, where);
267 list_insert_tail(&tm->tm_head, ts);
272 trim_map_segment_remove(trim_map_t *tm, trim_seg_t *ts, uint64_t start,
276 boolean_t left_over, right_over;
278 ASSERT(MUTEX_HELD(&tm->tm_lock));
280 left_over = (ts->ts_start < start);
281 right_over = (ts->ts_end > end);
283 TRIM_MAP_SDEC(tm, end - start);
284 if (left_over && right_over) {
285 nts = kmem_alloc(sizeof (*nts), KM_SLEEP);
287 nts->ts_end = ts->ts_end;
288 nts->ts_txg = ts->ts_txg;
289 nts->ts_time = ts->ts_time;
291 avl_insert_here(&tm->tm_queued_frees, nts, ts, AVL_AFTER);
292 list_insert_after(&tm->tm_head, ts, nts);
294 } else if (left_over) {
296 } else if (right_over) {
299 avl_remove(&tm->tm_queued_frees, ts);
300 list_remove(&tm->tm_head, ts);
302 kmem_free(ts, sizeof (*ts));
307 trim_map_free_locked(trim_map_t *tm, uint64_t start, uint64_t end, uint64_t txg)
311 ASSERT(MUTEX_HELD(&tm->tm_lock));
313 zsearch.io_offset = start;
314 zsearch.io_size = end - start;
316 zs = avl_find(&tm->tm_inflight_writes, &zsearch, NULL);
318 trim_map_segment_add(tm, start, end, txg);
321 if (start < zs->io_offset)
322 trim_map_free_locked(tm, start, zs->io_offset, txg);
323 if (zs->io_offset + zs->io_size < end)
324 trim_map_free_locked(tm, zs->io_offset + zs->io_size, end, txg);
328 trim_map_free(vdev_t *vd, uint64_t offset, uint64_t size, uint64_t txg)
330 trim_map_t *tm = vd->vdev_trimmap;
332 if (!zfs_trim_enabled || vd->vdev_notrim || tm == NULL)
335 mutex_enter(&tm->tm_lock);
336 trim_map_free_locked(tm, offset, TRIM_ZIO_END(vd, offset, size), txg);
337 mutex_exit(&tm->tm_lock);
341 trim_map_write_start(zio_t *zio)
343 vdev_t *vd = zio->io_vd;
344 trim_map_t *tm = vd->vdev_trimmap;
345 trim_seg_t tsearch, *ts;
346 boolean_t left_over, right_over;
349 if (!zfs_trim_enabled || vd->vdev_notrim || tm == NULL)
352 start = zio->io_offset;
353 end = TRIM_ZIO_END(zio->io_vd, start, zio->io_size);
354 tsearch.ts_start = start;
355 tsearch.ts_end = end;
357 mutex_enter(&tm->tm_lock);
360 * Checking for colliding in-flight frees.
362 ts = avl_find(&tm->tm_inflight_frees, &tsearch, NULL);
364 list_insert_tail(&tm->tm_pending_writes, zio);
365 mutex_exit(&tm->tm_lock);
369 ts = avl_find(&tm->tm_queued_frees, &tsearch, NULL);
372 * Loop until all overlapping segments are removed.
375 trim_map_segment_remove(tm, ts, start, end);
376 ts = avl_find(&tm->tm_queued_frees, &tsearch, NULL);
377 } while (ts != NULL);
379 avl_add(&tm->tm_inflight_writes, zio);
381 mutex_exit(&tm->tm_lock);
387 trim_map_write_done(zio_t *zio)
389 vdev_t *vd = zio->io_vd;
390 trim_map_t *tm = vd->vdev_trimmap;
393 * Don't check for vdev_notrim, since the write could have
394 * started before vdev_notrim was set.
396 if (!zfs_trim_enabled || tm == NULL)
399 mutex_enter(&tm->tm_lock);
401 * Don't fail if the write isn't in the tree, since the write
402 * could have started after vdev_notrim was set.
404 if (zio->io_trim_node.avl_child[0] ||
405 zio->io_trim_node.avl_child[1] ||
406 AVL_XPARENT(&zio->io_trim_node) ||
407 tm->tm_inflight_writes.avl_root == &zio->io_trim_node)
408 avl_remove(&tm->tm_inflight_writes, zio);
409 mutex_exit(&tm->tm_lock);
413 * Return the oldest segment (the one with the lowest txg / time) or NULL if:
414 * 1. The list is empty
415 * 2. The first element's txg is greater than txgsafe
416 * 3. The first element's txg is not greater than the txg argument and the
417 * the first element's time is not greater than time argument
420 trim_map_first(trim_map_t *tm, uint64_t txg, uint64_t txgsafe, hrtime_t time)
424 ASSERT(MUTEX_HELD(&tm->tm_lock));
425 VERIFY(txgsafe >= txg);
427 ts = list_head(&tm->tm_head);
428 if (ts != NULL && ts->ts_txg <= txgsafe &&
429 (ts->ts_txg <= txg || ts->ts_time <= time ||
430 tm->tm_bytes > trim_vdev_max_bytes ||
431 tm->tm_pending > trim_vdev_max_pending))
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;
444 ASSERT(vd->vdev_ops->vdev_op_leaf);
449 timelimit = gethrtime() - trim_timeout * NANOSEC;
450 if (vd->vdev_isl2cache) {
451 txgsafe = UINT64_MAX;
452 txgtarget = UINT64_MAX;
454 txgsafe = MIN(spa_last_synced_txg(spa), spa_freeze_txg(spa));
455 if (txgsafe > trim_txg_delay)
456 txgtarget = txgsafe - trim_txg_delay;
461 mutex_enter(&tm->tm_lock);
462 /* Loop until we have sent all outstanding free's */
463 while ((ts = trim_map_first(tm, txgtarget, txgsafe, timelimit))
465 list_remove(&tm->tm_head, ts);
466 avl_remove(&tm->tm_queued_frees, ts);
467 avl_add(&tm->tm_inflight_frees, ts);
468 size = ts->ts_end - ts->ts_start;
469 offset = ts->ts_start;
470 TRIM_MAP_SDEC(tm, size);
473 * We drop the lock while we call zio_nowait as the IO
474 * scheduler can result in a different IO being run e.g.
475 * a write which would result in a recursive lock.
477 mutex_exit(&tm->tm_lock);
479 zio_nowait(zio_trim(zio, spa, vd, offset, size));
481 mutex_enter(&tm->tm_lock);
483 mutex_exit(&tm->tm_lock);
487 trim_map_vdev_commit_done(spa_t *spa, vdev_t *vd)
489 trim_map_t *tm = vd->vdev_trimmap;
491 list_t pending_writes;
493 uint64_t start, size;
496 ASSERT(vd->vdev_ops->vdev_op_leaf);
501 mutex_enter(&tm->tm_lock);
502 if (!avl_is_empty(&tm->tm_inflight_frees)) {
504 while ((ts = avl_destroy_nodes(&tm->tm_inflight_frees,
506 kmem_free(ts, sizeof (*ts));
509 list_create(&pending_writes, sizeof (zio_t), offsetof(zio_t,
511 list_move_tail(&pending_writes, &tm->tm_pending_writes);
512 mutex_exit(&tm->tm_lock);
514 while ((zio = list_remove_head(&pending_writes)) != NULL) {
515 zio_vdev_io_reissue(zio);
518 list_destroy(&pending_writes);
522 trim_map_commit(spa_t *spa, zio_t *zio, vdev_t *vd)
529 if (vd->vdev_ops->vdev_op_leaf) {
530 trim_map_vdev_commit(spa, zio, vd);
532 for (c = 0; c < vd->vdev_children; c++)
533 trim_map_commit(spa, zio, vd->vdev_child[c]);
538 trim_map_commit_done(spa_t *spa, vdev_t *vd)
545 if (vd->vdev_ops->vdev_op_leaf) {
546 trim_map_vdev_commit_done(spa, vd);
548 for (c = 0; c < vd->vdev_children; c++)
549 trim_map_commit_done(spa, vd->vdev_child[c]);
554 trim_thread(void *arg)
560 (void) snprintf(curthread->td_name, sizeof(curthread->td_name),
561 "trim %s", spa_name(spa));
565 mutex_enter(&spa->spa_trim_lock);
566 if (spa->spa_trim_thread == NULL) {
567 spa->spa_trim_thread = curthread;
568 cv_signal(&spa->spa_trim_cv);
569 mutex_exit(&spa->spa_trim_lock);
573 (void) cv_timedwait(&spa->spa_trim_cv, &spa->spa_trim_lock,
574 hz * trim_max_interval);
575 mutex_exit(&spa->spa_trim_lock);
577 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
579 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
580 trim_map_commit(spa, zio, spa->spa_root_vdev);
581 (void) zio_wait(zio);
582 trim_map_commit_done(spa, spa->spa_root_vdev);
583 spa_config_exit(spa, SCL_STATE, FTAG);
588 trim_thread_create(spa_t *spa)
591 if (!zfs_trim_enabled)
594 mutex_init(&spa->spa_trim_lock, NULL, MUTEX_DEFAULT, NULL);
595 cv_init(&spa->spa_trim_cv, NULL, CV_DEFAULT, NULL);
596 mutex_enter(&spa->spa_trim_lock);
597 spa->spa_trim_thread = thread_create(NULL, 0, trim_thread, spa, 0, &p0,
598 TS_RUN, minclsyspri);
599 mutex_exit(&spa->spa_trim_lock);
603 trim_thread_destroy(spa_t *spa)
606 if (!zfs_trim_enabled)
608 if (spa->spa_trim_thread == NULL)
611 mutex_enter(&spa->spa_trim_lock);
612 /* Setting spa_trim_thread to NULL tells the thread to stop. */
613 spa->spa_trim_thread = NULL;
614 cv_signal(&spa->spa_trim_cv);
615 /* The thread will set it back to != NULL on exit. */
616 while (spa->spa_trim_thread == NULL)
617 cv_wait(&spa->spa_trim_cv, &spa->spa_trim_lock);
618 spa->spa_trim_thread = NULL;
619 mutex_exit(&spa->spa_trim_lock);
621 cv_destroy(&spa->spa_trim_cv);
622 mutex_destroy(&spa->spa_trim_lock);
626 trim_thread_wakeup(spa_t *spa)
629 if (!zfs_trim_enabled)
631 if (spa->spa_trim_thread == NULL)
634 mutex_enter(&spa->spa_trim_lock);
635 cv_signal(&spa->spa_trim_cv);
636 mutex_exit(&spa->spa_trim_lock);