xref: /illumos-gate/usr/src/uts/common/fs/zfs/vdev_trim.c (revision a385ae06552e0d045ef45ac2ff1dc08a334d955c)
1 /*
2  * CDDL HEADER START
3  *
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.
7  *
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.
12  *
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]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright (c) 2016 by Delphix. All rights reserved.
24  * Copyright (c) 2019 by Lawrence Livermore National Security, LLC.
25  * Copyright 2019 Joyent, Inc.
26  * Copyright 2023 RackTop Systems, Inc.
27  */
28 
29 #include <sys/spa.h>
30 #include <sys/spa_impl.h>
31 #include <sys/txg.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/vdev_trim.h>
34 #include <sys/refcount.h>
35 #include <sys/metaslab_impl.h>
36 #include <sys/dsl_synctask.h>
37 #include <sys/zap.h>
38 #include <sys/dmu_tx.h>
39 
40 /*
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.
46  *
47  * There are two supported TRIM methods; manual and automatic.
48  *
49  * Manual TRIM:
50  *
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
55  * required TRIM I/Os.
56  *
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.
62  *
63  * Automatic TRIM:
64  *
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.
69  *
70  * 1) Automatic TRIM happens continuously in the background and operates
71  *    solely on recently freed blocks (ms_trim not ms_allocatable).
72  *
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
78  *    children.
79  *
80  * 3) There is no automatic TRIM progress information stored on disk, nor
81  *    is it reported by 'zpool status'.
82  *
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.
88  *
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.
92  *
93  * For this reason it may be beneficial to occasionally manually TRIM a pool
94  * even when automatic TRIM is enabled.
95  */
96 
97 /*
98  * Maximum size of TRIM I/O, ranges will be chunked in to 128MiB lengths.
99  */
100 unsigned int zfs_trim_extent_bytes_max = 128 * 1024 * 1024;
101 
102 /*
103  * Minimum size of TRIM I/O, extents smaller than 32Kib will be skipped.
104  */
105 unsigned int zfs_trim_extent_bytes_min = 32 * 1024;
106 
107 /*
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.
114  */
115 unsigned int zfs_trim_metaslab_skip = 0;
116 
117 /*
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.
121  */
122 unsigned int zfs_trim_queue_limit = 10;
123 
124 /*
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.
131  *
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.
137  */
138 unsigned int zfs_trim_txg_batch = 32;
139 
140 /*
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.
145  */
146 typedef struct trim_args {
147 	/*
148 	 * These fields are set by the caller of vdev_trim_ranges().
149 	 */
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) */
157 
158 	/*
159 	 * These fields are updated by vdev_trim_ranges().
160 	 */
161 	hrtime_t	trim_start_time;	/* Start time */
162 	uint64_t	trim_bytes_done;	/* Bytes trimmed */
163 } trim_args_t;
164 
165 /*
166  * Determines whether a vdev_trim_thread() should be stopped.
167  */
168 static boolean_t
vdev_trim_should_stop(vdev_t * vd)169 vdev_trim_should_stop(vdev_t *vd)
170 {
171 	return (vd->vdev_trim_exit_wanted || !vdev_writeable(vd) ||
172 	    vd->vdev_detached || vd->vdev_top->vdev_removing);
173 }
174 
175 /*
176  * Determines whether a vdev_autotrim_thread() should be stopped.
177  */
178 static boolean_t
vdev_autotrim_should_stop(vdev_t * tvd)179 vdev_autotrim_should_stop(vdev_t *tvd)
180 {
181 	return (tvd->vdev_autotrim_exit_wanted ||
182 	    !vdev_writeable(tvd) || tvd->vdev_removing ||
183 	    spa_get_autotrim(tvd->vdev_spa) == SPA_AUTOTRIM_OFF);
184 }
185 
186 /*
187  * The sync task for updating the on-disk state of a manual TRIM.  This
188  * is scheduled by vdev_trim_change_state().
189  */
190 static void
vdev_trim_zap_update_sync(void * arg,dmu_tx_t * tx)191 vdev_trim_zap_update_sync(void *arg, dmu_tx_t *tx)
192 {
193 	/*
194 	 * We pass in the guid instead of the vdev_t since the vdev may
195 	 * have been freed prior to the sync task being processed.  This
196 	 * happens when a vdev is detached as we call spa_config_vdev_exit(),
197 	 * stop the trimming thread, schedule the sync task, and free
198 	 * the vdev. Later when the scheduled sync task is invoked, it would
199 	 * find that the vdev has been freed.
200 	 */
201 	uint64_t guid = *(uint64_t *)arg;
202 	uint64_t txg = dmu_tx_get_txg(tx);
203 	kmem_free(arg, sizeof (uint64_t));
204 
205 	vdev_t *vd = spa_lookup_by_guid(tx->tx_pool->dp_spa, guid, B_FALSE);
206 	if (vd == NULL || vd->vdev_top->vdev_removing || !vdev_is_concrete(vd))
207 		return;
208 
209 	uint64_t last_offset = vd->vdev_trim_offset[txg & TXG_MASK];
210 	vd->vdev_trim_offset[txg & TXG_MASK] = 0;
211 
212 	VERIFY3U(vd->vdev_leaf_zap, !=, 0);
213 
214 	objset_t *mos = vd->vdev_spa->spa_meta_objset;
215 
216 	if (last_offset > 0 || vd->vdev_trim_last_offset == UINT64_MAX) {
217 
218 		if (vd->vdev_trim_last_offset == UINT64_MAX)
219 			last_offset = 0;
220 
221 		vd->vdev_trim_last_offset = last_offset;
222 		VERIFY0(zap_update(mos, vd->vdev_leaf_zap,
223 		    VDEV_LEAF_ZAP_TRIM_LAST_OFFSET,
224 		    sizeof (last_offset), 1, &last_offset, tx));
225 	}
226 
227 	if (vd->vdev_trim_action_time > 0) {
228 		uint64_t val = (uint64_t)vd->vdev_trim_action_time;
229 		VERIFY0(zap_update(mos, vd->vdev_leaf_zap,
230 		    VDEV_LEAF_ZAP_TRIM_ACTION_TIME, sizeof (val),
231 		    1, &val, tx));
232 	}
233 
234 	if (vd->vdev_trim_rate > 0) {
235 		uint64_t rate = (uint64_t)vd->vdev_trim_rate;
236 
237 		if (rate == UINT64_MAX)
238 			rate = 0;
239 
240 		VERIFY0(zap_update(mos, vd->vdev_leaf_zap,
241 		    VDEV_LEAF_ZAP_TRIM_RATE, sizeof (rate), 1, &rate, tx));
242 	}
243 
244 	uint64_t partial = vd->vdev_trim_partial;
245 	if (partial == UINT64_MAX)
246 		partial = 0;
247 
248 	VERIFY0(zap_update(mos, vd->vdev_leaf_zap, VDEV_LEAF_ZAP_TRIM_PARTIAL,
249 	    sizeof (partial), 1, &partial, tx));
250 
251 	uint64_t secure = vd->vdev_trim_secure;
252 	if (secure == UINT64_MAX)
253 		secure = 0;
254 
255 	VERIFY0(zap_update(mos, vd->vdev_leaf_zap, VDEV_LEAF_ZAP_TRIM_SECURE,
256 	    sizeof (secure), 1, &secure, tx));
257 
258 
259 	uint64_t trim_state = vd->vdev_trim_state;
260 	VERIFY0(zap_update(mos, vd->vdev_leaf_zap, VDEV_LEAF_ZAP_TRIM_STATE,
261 	    sizeof (trim_state), 1, &trim_state, tx));
262 }
263 
264 /*
265  * Update the on-disk state of a manual TRIM.  This is called to request
266  * that a TRIM be started/suspended/canceled, or to change one of the
267  * TRIM options (partial, secure, rate).
268  */
269 static void
vdev_trim_change_state(vdev_t * vd,vdev_trim_state_t new_state,uint64_t rate,boolean_t partial,boolean_t secure)270 vdev_trim_change_state(vdev_t *vd, vdev_trim_state_t new_state,
271     uint64_t rate, boolean_t partial, boolean_t secure)
272 {
273 	ASSERT(MUTEX_HELD(&vd->vdev_trim_lock));
274 	spa_t *spa = vd->vdev_spa;
275 
276 	if (new_state == vd->vdev_trim_state)
277 		return;
278 
279 	/*
280 	 * Copy the vd's guid, this will be freed by the sync task.
281 	 */
282 	uint64_t *guid = kmem_zalloc(sizeof (uint64_t), KM_SLEEP);
283 	*guid = vd->vdev_guid;
284 
285 	/*
286 	 * If we're suspending, then preserve the original start time.
287 	 */
288 	if (vd->vdev_trim_state != VDEV_TRIM_SUSPENDED) {
289 		vd->vdev_trim_action_time = gethrestime_sec();
290 	}
291 
292 	/*
293 	 * If we're activating, then preserve the requested rate and trim
294 	 * method.  Setting the last offset and rate to UINT64_MAX is used
295 	 * as a sentinel to indicate they should be reset to default values.
296 	 */
297 	if (new_state == VDEV_TRIM_ACTIVE) {
298 		if (vd->vdev_trim_state == VDEV_TRIM_COMPLETE ||
299 		    vd->vdev_trim_state == VDEV_TRIM_CANCELED) {
300 			vd->vdev_trim_last_offset = UINT64_MAX;
301 			vd->vdev_trim_rate = UINT64_MAX;
302 			vd->vdev_trim_partial = UINT64_MAX;
303 			vd->vdev_trim_secure = UINT64_MAX;
304 		}
305 
306 		if (rate != 0)
307 			vd->vdev_trim_rate = rate;
308 
309 		if (partial != 0)
310 			vd->vdev_trim_partial = partial;
311 
312 		if (secure != 0)
313 			vd->vdev_trim_secure = secure;
314 	}
315 
316 	boolean_t resumed = !!(vd->vdev_trim_state == VDEV_TRIM_SUSPENDED);
317 	vd->vdev_trim_state = new_state;
318 
319 	dmu_tx_t *tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir);
320 	VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
321 	dsl_sync_task_nowait(spa_get_dsl(spa), vdev_trim_zap_update_sync,
322 	    guid, 2, ZFS_SPACE_CHECK_NONE, tx);
323 
324 	switch (new_state) {
325 	case VDEV_TRIM_ACTIVE:
326 		spa_event_notify(spa, vd, NULL,
327 		    resumed ? ESC_ZFS_TRIM_RESUME : ESC_ZFS_TRIM_START);
328 		spa_history_log_internal(spa, "trim", tx,
329 		    "vdev=%s activated", vd->vdev_path);
330 		break;
331 	case VDEV_TRIM_SUSPENDED:
332 		spa_event_notify(spa, vd, NULL, ESC_ZFS_TRIM_SUSPEND);
333 		spa_history_log_internal(spa, "trim", tx,
334 		    "vdev=%s suspended", vd->vdev_path);
335 		break;
336 	case VDEV_TRIM_CANCELED:
337 		spa_event_notify(spa, vd, NULL, ESC_ZFS_TRIM_CANCEL);
338 		spa_history_log_internal(spa, "trim", tx,
339 		    "vdev=%s canceled", vd->vdev_path);
340 		break;
341 	case VDEV_TRIM_COMPLETE:
342 		spa_event_notify(spa, vd, NULL, ESC_ZFS_TRIM_FINISH);
343 		spa_history_log_internal(spa, "trim", tx,
344 		    "vdev=%s complete", vd->vdev_path);
345 		break;
346 	default:
347 		panic("invalid state %llu", (unsigned long long)new_state);
348 	}
349 
350 	dmu_tx_commit(tx);
351 }
352 
353 /*
354  * The zio_done_func_t done callback for each manual TRIM issued.  It is
355  * responsible for updating the TRIM stats, reissuing failed TRIM I/Os,
356  * and limiting the number of in-flight TRIM I/Os.
357  */
358 static void
vdev_trim_cb(zio_t * zio)359 vdev_trim_cb(zio_t *zio)
360 {
361 	vdev_t *vd = zio->io_vd;
362 
363 	mutex_enter(&vd->vdev_trim_io_lock);
364 	if (zio->io_error == ENXIO && !vdev_writeable(vd)) {
365 		/*
366 		 * The I/O failed because the vdev was unavailable; roll the
367 		 * last offset back. (This works because spa_sync waits on
368 		 * spa_txg_zio before it runs sync tasks.)
369 		 */
370 		uint64_t *offset =
371 		    &vd->vdev_trim_offset[zio->io_txg & TXG_MASK];
372 		*offset = MIN(*offset, zio->io_offset);
373 	} else {
374 		if (zio->io_error != 0) {
375 			vd->vdev_stat.vs_trim_errors++;
376 			/*
377 			 * spa_iostats_trim_add(vd->vdev_spa, TRIM_TYPE_MANUAL,
378 			 *  0, 0, 0, 0, 1, zio->io_orig_size);
379 			 */
380 		} else {
381 			/*
382 			 * spa_iostats_trim_add(vd->vdev_spa, TRIM_TYPE_MANUAL,
383 			 *  1, zio->io_orig_size, 0, 0, 0, 0);
384 			 */
385 		}
386 
387 		vd->vdev_trim_bytes_done += zio->io_orig_size;
388 	}
389 
390 	ASSERT3U(vd->vdev_trim_inflight[TRIM_TYPE_MANUAL], >, 0);
391 	vd->vdev_trim_inflight[TRIM_TYPE_MANUAL]--;
392 	cv_broadcast(&vd->vdev_trim_io_cv);
393 	mutex_exit(&vd->vdev_trim_io_lock);
394 
395 	spa_config_exit(vd->vdev_spa, SCL_STATE_ALL, vd);
396 }
397 
398 /*
399  * The zio_done_func_t done callback for each automatic TRIM issued.  It
400  * is responsible for updating the TRIM stats and limiting the number of
401  * in-flight TRIM I/Os.  Automatic TRIM I/Os are best effort and are
402  * never reissued on failure.
403  */
404 static void
vdev_autotrim_cb(zio_t * zio)405 vdev_autotrim_cb(zio_t *zio)
406 {
407 	vdev_t *vd = zio->io_vd;
408 
409 	mutex_enter(&vd->vdev_trim_io_lock);
410 
411 	if (zio->io_error != 0) {
412 		vd->vdev_stat.vs_trim_errors++;
413 		/*
414 		 * spa_iostats_trim_add(vd->vdev_spa, TRIM_TYPE_AUTO,
415 		 *  0, 0, 0, 0, 1, zio->io_orig_size);
416 		 */
417 	} else {
418 		/*
419 		 * spa_iostats_trim_add(vd->vdev_spa, TRIM_TYPE_AUTO,
420 		 *  1, zio->io_orig_size, 0, 0, 0, 0);
421 		 */
422 
423 		vd->vdev_autotrim_bytes_done += zio->io_orig_size;
424 	}
425 
426 	ASSERT3U(vd->vdev_trim_inflight[TRIM_TYPE_AUTO], >, 0);
427 	vd->vdev_trim_inflight[TRIM_TYPE_AUTO]--;
428 	cv_broadcast(&vd->vdev_trim_io_cv);
429 	mutex_exit(&vd->vdev_trim_io_lock);
430 
431 	spa_config_exit(vd->vdev_spa, SCL_STATE_ALL, vd);
432 }
433 
434 /*
435  * Returns the average trim rate in bytes/sec for the ta->trim_vdev.
436  */
437 static uint64_t
vdev_trim_calculate_rate(trim_args_t * ta)438 vdev_trim_calculate_rate(trim_args_t *ta)
439 {
440 	return (ta->trim_bytes_done * 1000 /
441 	    (NSEC2MSEC(gethrtime() - ta->trim_start_time) + 1));
442 }
443 
444 /*
445  * Issues a physical TRIM and takes care of rate limiting (bytes/sec)
446  * and number of concurrent TRIM I/Os.
447  */
448 static int
vdev_trim_range(trim_args_t * ta,uint64_t start,uint64_t size)449 vdev_trim_range(trim_args_t *ta, uint64_t start, uint64_t size)
450 {
451 	vdev_t *vd = ta->trim_vdev;
452 	spa_t *spa = vd->vdev_spa;
453 
454 	mutex_enter(&vd->vdev_trim_io_lock);
455 
456 	/*
457 	 * Limit manual TRIM I/Os to the requested rate.  This does not
458 	 * apply to automatic TRIM since no per vdev rate can be specified.
459 	 */
460 	if (ta->trim_type == TRIM_TYPE_MANUAL) {
461 		while (vd->vdev_trim_rate != 0 && !vdev_trim_should_stop(vd) &&
462 		    vdev_trim_calculate_rate(ta) > vd->vdev_trim_rate) {
463 			cv_timedwait_sig(&vd->vdev_trim_io_cv,
464 			    &vd->vdev_trim_io_lock, ddi_get_lbolt() +
465 			    MSEC_TO_TICK(10));
466 		}
467 	}
468 	ta->trim_bytes_done += size;
469 
470 	/* Limit in-flight trimming I/Os */
471 	while (vd->vdev_trim_inflight[0] + vd->vdev_trim_inflight[1] >=
472 	    zfs_trim_queue_limit) {
473 		cv_wait(&vd->vdev_trim_io_cv, &vd->vdev_trim_io_lock);
474 	}
475 	vd->vdev_trim_inflight[ta->trim_type]++;
476 	mutex_exit(&vd->vdev_trim_io_lock);
477 
478 	dmu_tx_t *tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir);
479 	VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
480 	uint64_t txg = dmu_tx_get_txg(tx);
481 
482 	spa_config_enter(spa, SCL_STATE_ALL, vd, RW_READER);
483 	mutex_enter(&vd->vdev_trim_lock);
484 
485 	if (ta->trim_type == TRIM_TYPE_MANUAL &&
486 	    vd->vdev_trim_offset[txg & TXG_MASK] == 0) {
487 		uint64_t *guid = kmem_zalloc(sizeof (uint64_t), KM_SLEEP);
488 		*guid = vd->vdev_guid;
489 
490 		/* This is the first write of this txg. */
491 		dsl_sync_task_nowait(spa_get_dsl(spa),
492 		    vdev_trim_zap_update_sync, guid, 2,
493 		    ZFS_SPACE_CHECK_RESERVED, tx);
494 	}
495 
496 	/*
497 	 * We know the vdev_t will still be around since all consumers of
498 	 * vdev_free must stop the trimming first.
499 	 */
500 	if ((ta->trim_type == TRIM_TYPE_MANUAL &&
501 	    vdev_trim_should_stop(vd)) ||
502 	    (ta->trim_type == TRIM_TYPE_AUTO &&
503 	    vdev_autotrim_should_stop(vd->vdev_top))) {
504 		mutex_enter(&vd->vdev_trim_io_lock);
505 		vd->vdev_trim_inflight[ta->trim_type]--;
506 		mutex_exit(&vd->vdev_trim_io_lock);
507 		spa_config_exit(vd->vdev_spa, SCL_STATE_ALL, vd);
508 		mutex_exit(&vd->vdev_trim_lock);
509 		dmu_tx_commit(tx);
510 		return (SET_ERROR(EINTR));
511 	}
512 	mutex_exit(&vd->vdev_trim_lock);
513 
514 	if (ta->trim_type == TRIM_TYPE_MANUAL)
515 		vd->vdev_trim_offset[txg & TXG_MASK] = start + size;
516 
517 	zio_nowait(zio_trim(spa->spa_txg_zio[txg & TXG_MASK], vd,
518 	    start, size, ta->trim_type == TRIM_TYPE_MANUAL ?
519 	    vdev_trim_cb : vdev_autotrim_cb, NULL,
520 	    ZIO_PRIORITY_TRIM, ZIO_FLAG_CANFAIL, ta->trim_flags));
521 	/* vdev_trim_cb and vdev_autotrim_cb release SCL_STATE_ALL */
522 
523 	dmu_tx_commit(tx);
524 
525 	return (0);
526 }
527 
528 /*
529  * Issues TRIM I/Os for all ranges in the provided ta->trim_tree range tree.
530  * Additional parameters describing how the TRIM should be performed must
531  * be set in the trim_args structure.  See the trim_args definition for
532  * additional information.
533  */
534 static int
vdev_trim_ranges(trim_args_t * ta)535 vdev_trim_ranges(trim_args_t *ta)
536 {
537 	vdev_t *vd = ta->trim_vdev;
538 	zfs_btree_t *t = &ta->trim_tree->rt_root;
539 	zfs_btree_index_t idx;
540 	uint64_t extent_bytes_max = ta->trim_extent_bytes_max;
541 	uint64_t extent_bytes_min = ta->trim_extent_bytes_min;
542 	spa_t *spa = vd->vdev_spa;
543 	int error = 0;
544 
545 	ta->trim_start_time = gethrtime();
546 	ta->trim_bytes_done = 0;
547 
548 	for (range_seg_t *rs = zfs_btree_first(t, &idx); rs != NULL;
549 	    rs = zfs_btree_next(t, &idx, &idx)) {
550 		uint64_t size = rs_get_end(rs, ta->trim_tree) - rs_get_start(rs,
551 		    ta->trim_tree);
552 
553 		if (extent_bytes_min && size < extent_bytes_min) {
554 			/*
555 			 * spa_iostats_trim_add(spa, ta->trim_type,
556 			 *  0, 0, 1, size, 0, 0);
557 			 */
558 			continue;
559 		}
560 
561 		/* Split range into legally-sized physical chunks */
562 		uint64_t writes_required = ((size - 1) / extent_bytes_max) + 1;
563 
564 		for (uint64_t w = 0; w < writes_required; w++) {
565 			error = vdev_trim_range(ta, VDEV_LABEL_START_SIZE +
566 			    rs_get_start(rs, ta->trim_tree) +
567 			    (w *extent_bytes_max), MIN(size -
568 			    (w * extent_bytes_max), extent_bytes_max));
569 			if (error != 0) {
570 				goto done;
571 			}
572 		}
573 	}
574 
575 done:
576 	/*
577 	 * Make sure all TRIMs for this metaslab have completed before
578 	 * returning. TRIM zios have lower priority over regular or syncing
579 	 * zios, so all TRIM zios for this metaslab must complete before the
580 	 * metaslab is re-enabled. Otherwise it's possible write zios to
581 	 * this metaslab could cut ahead of still queued TRIM zios for this
582 	 * metaslab causing corruption if the ranges overlap.
583 	 */
584 	mutex_enter(&vd->vdev_trim_io_lock);
585 	while (vd->vdev_trim_inflight[TRIM_TYPE_MANUAL] > 0) {
586 		cv_wait(&vd->vdev_trim_io_cv, &vd->vdev_trim_io_lock);
587 	}
588 	mutex_exit(&vd->vdev_trim_io_lock);
589 
590 	return (error);
591 }
592 
593 /*
594  * Calculates the completion percentage of a manual TRIM.
595  */
596 static void
vdev_trim_calculate_progress(vdev_t * vd)597 vdev_trim_calculate_progress(vdev_t *vd)
598 {
599 	ASSERT(spa_config_held(vd->vdev_spa, SCL_CONFIG, RW_READER) ||
600 	    spa_config_held(vd->vdev_spa, SCL_CONFIG, RW_WRITER));
601 	ASSERT(vd->vdev_leaf_zap != 0);
602 
603 	vd->vdev_trim_bytes_est = 0;
604 	vd->vdev_trim_bytes_done = 0;
605 
606 	for (uint64_t i = 0; i < vd->vdev_top->vdev_ms_count; i++) {
607 		metaslab_t *msp = vd->vdev_top->vdev_ms[i];
608 		mutex_enter(&msp->ms_lock);
609 
610 		uint64_t ms_free = msp->ms_size -
611 		    metaslab_allocated_space(msp);
612 
613 		if (vd->vdev_top->vdev_ops == &vdev_raidz_ops)
614 			ms_free /= vd->vdev_top->vdev_children;
615 
616 		/*
617 		 * Convert the metaslab range to a physical range
618 		 * on our vdev. We use this to determine if we are
619 		 * in the middle of this metaslab range.
620 		 */
621 		range_seg64_t logical_rs, physical_rs;
622 		logical_rs.rs_start = msp->ms_start;
623 		logical_rs.rs_end = msp->ms_start + msp->ms_size;
624 		vdev_xlate(vd, &logical_rs, &physical_rs);
625 
626 		if (vd->vdev_trim_last_offset <= physical_rs.rs_start) {
627 			vd->vdev_trim_bytes_est += ms_free;
628 			mutex_exit(&msp->ms_lock);
629 			continue;
630 		} else if (vd->vdev_trim_last_offset > physical_rs.rs_end) {
631 			vd->vdev_trim_bytes_done += ms_free;
632 			vd->vdev_trim_bytes_est += ms_free;
633 			mutex_exit(&msp->ms_lock);
634 			continue;
635 		}
636 
637 		/*
638 		 * If we get here, we're in the middle of trimming this
639 		 * metaslab.  Load it and walk the free tree for more
640 		 * accurate progress estimation.
641 		 */
642 		VERIFY0(metaslab_load(msp));
643 
644 		range_tree_t *rt = msp->ms_allocatable;
645 		zfs_btree_t *bt = &rt->rt_root;
646 		zfs_btree_index_t idx;
647 		for (range_seg_t *rs = zfs_btree_first(bt, &idx);
648 		    rs != NULL; rs = zfs_btree_next(bt, &idx, &idx)) {
649 			logical_rs.rs_start = rs_get_start(rs, rt);
650 			logical_rs.rs_end = rs_get_end(rs, rt);
651 			vdev_xlate(vd, &logical_rs, &physical_rs);
652 
653 			uint64_t size = physical_rs.rs_end -
654 			    physical_rs.rs_start;
655 			vd->vdev_trim_bytes_est += size;
656 			if (vd->vdev_trim_last_offset >= physical_rs.rs_end) {
657 				vd->vdev_trim_bytes_done += size;
658 			} else if (vd->vdev_trim_last_offset >
659 			    physical_rs.rs_start &&
660 			    vd->vdev_trim_last_offset <=
661 			    physical_rs.rs_end) {
662 				vd->vdev_trim_bytes_done +=
663 				    vd->vdev_trim_last_offset -
664 				    physical_rs.rs_start;
665 			}
666 		}
667 		mutex_exit(&msp->ms_lock);
668 	}
669 }
670 
671 /*
672  * Load from disk the vdev's manual TRIM information.  This includes the
673  * state, progress, and options provided when initiating the manual TRIM.
674  */
675 static int
vdev_trim_load(vdev_t * vd)676 vdev_trim_load(vdev_t *vd)
677 {
678 	int err = 0;
679 	ASSERT(spa_config_held(vd->vdev_spa, SCL_CONFIG, RW_READER) ||
680 	    spa_config_held(vd->vdev_spa, SCL_CONFIG, RW_WRITER));
681 	ASSERT(vd->vdev_leaf_zap != 0);
682 
683 	if (vd->vdev_trim_state == VDEV_TRIM_ACTIVE ||
684 	    vd->vdev_trim_state == VDEV_TRIM_SUSPENDED) {
685 		err = zap_lookup(vd->vdev_spa->spa_meta_objset,
686 		    vd->vdev_leaf_zap, VDEV_LEAF_ZAP_TRIM_LAST_OFFSET,
687 		    sizeof (vd->vdev_trim_last_offset), 1,
688 		    &vd->vdev_trim_last_offset);
689 		if (err == ENOENT) {
690 			vd->vdev_trim_last_offset = 0;
691 			err = 0;
692 		}
693 
694 		if (err == 0) {
695 			err = zap_lookup(vd->vdev_spa->spa_meta_objset,
696 			    vd->vdev_leaf_zap, VDEV_LEAF_ZAP_TRIM_RATE,
697 			    sizeof (vd->vdev_trim_rate), 1,
698 			    &vd->vdev_trim_rate);
699 			if (err == ENOENT) {
700 				vd->vdev_trim_rate = 0;
701 				err = 0;
702 			}
703 		}
704 
705 		if (err == 0) {
706 			err = zap_lookup(vd->vdev_spa->spa_meta_objset,
707 			    vd->vdev_leaf_zap, VDEV_LEAF_ZAP_TRIM_PARTIAL,
708 			    sizeof (vd->vdev_trim_partial), 1,
709 			    &vd->vdev_trim_partial);
710 			if (err == ENOENT) {
711 				vd->vdev_trim_partial = 0;
712 				err = 0;
713 			}
714 		}
715 
716 		if (err == 0) {
717 			err = zap_lookup(vd->vdev_spa->spa_meta_objset,
718 			    vd->vdev_leaf_zap, VDEV_LEAF_ZAP_TRIM_SECURE,
719 			    sizeof (vd->vdev_trim_secure), 1,
720 			    &vd->vdev_trim_secure);
721 			if (err == ENOENT) {
722 				vd->vdev_trim_secure = 0;
723 				err = 0;
724 			}
725 		}
726 	}
727 
728 	vdev_trim_calculate_progress(vd);
729 
730 	return (err);
731 }
732 
733 /*
734  * Convert the logical range into a physical range and add it to the
735  * range tree passed in the trim_args_t.
736  */
737 static void
vdev_trim_range_add(void * arg,uint64_t start,uint64_t size)738 vdev_trim_range_add(void *arg, uint64_t start, uint64_t size)
739 {
740 	trim_args_t *ta = arg;
741 	vdev_t *vd = ta->trim_vdev;
742 	range_seg64_t logical_rs, physical_rs;
743 	logical_rs.rs_start = start;
744 	logical_rs.rs_end = start + size;
745 
746 	/*
747 	 * Every range to be trimmed must be part of ms_allocatable.
748 	 * When ZFS_DEBUG_TRIM is set load the metaslab to verify this
749 	 * is always the case.
750 	 */
751 	if (zfs_flags & ZFS_DEBUG_TRIM) {
752 		metaslab_t *msp = ta->trim_msp;
753 		VERIFY0(metaslab_load(msp));
754 		VERIFY3B(msp->ms_loaded, ==, B_TRUE);
755 		VERIFY(range_tree_contains(msp->ms_allocatable, start, size));
756 	}
757 
758 	ASSERT(vd->vdev_ops->vdev_op_leaf);
759 	vdev_xlate(vd, &logical_rs, &physical_rs);
760 
761 	IMPLY(vd->vdev_top == vd,
762 	    logical_rs.rs_start == physical_rs.rs_start);
763 	IMPLY(vd->vdev_top == vd,
764 	    logical_rs.rs_end == physical_rs.rs_end);
765 
766 	/*
767 	 * Only a manual trim will be traversing the vdev sequentially.
768 	 * For an auto trim all valid ranges should be added.
769 	 */
770 	if (ta->trim_type == TRIM_TYPE_MANUAL) {
771 
772 		/* Only add segments that we have not visited yet */
773 		if (physical_rs.rs_end <= vd->vdev_trim_last_offset)
774 			return;
775 
776 		/* Pick up where we left off mid-range. */
777 		if (vd->vdev_trim_last_offset > physical_rs.rs_start) {
778 			ASSERT3U(physical_rs.rs_end, >,
779 			    vd->vdev_trim_last_offset);
780 			physical_rs.rs_start = vd->vdev_trim_last_offset;
781 		}
782 	}
783 
784 	ASSERT3U(physical_rs.rs_end, >=, physical_rs.rs_start);
785 
786 	/*
787 	 * With raidz, it's possible that the logical range does not live on
788 	 * this leaf vdev. We only add the physical range to this vdev's if it
789 	 * has a length greater than 0.
790 	 */
791 	if (physical_rs.rs_end > physical_rs.rs_start) {
792 		range_tree_add(ta->trim_tree, physical_rs.rs_start,
793 		    physical_rs.rs_end - physical_rs.rs_start);
794 	} else {
795 		ASSERT3U(physical_rs.rs_end, ==, physical_rs.rs_start);
796 	}
797 }
798 
799 /*
800  * Each manual TRIM thread is responsible for trimming the unallocated
801  * space for each leaf vdev.  This is accomplished by sequentially iterating
802  * over its top-level metaslabs and issuing TRIM I/O for the space described
803  * by its ms_allocatable.  While a metaslab is undergoing trimming it is
804  * not eligible for new allocations.
805  */
806 static void
vdev_trim_thread(void * arg)807 vdev_trim_thread(void *arg)
808 {
809 	vdev_t *vd = arg;
810 	spa_t *spa = vd->vdev_spa;
811 	trim_args_t ta;
812 	int error = 0;
813 
814 	/*
815 	 * The VDEV_LEAF_ZAP_TRIM_* entries may have been updated by
816 	 * vdev_trim().  Wait for the updated values to be reflected
817 	 * in the zap in order to start with the requested settings.
818 	 */
819 	txg_wait_synced(spa_get_dsl(vd->vdev_spa), 0);
820 
821 	ASSERT(vdev_is_concrete(vd));
822 	spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
823 
824 	vd->vdev_trim_last_offset = 0;
825 	vd->vdev_trim_rate = 0;
826 	vd->vdev_trim_partial = 0;
827 	vd->vdev_trim_secure = 0;
828 
829 	VERIFY0(vdev_trim_load(vd));
830 
831 	ta.trim_vdev = vd;
832 	ta.trim_extent_bytes_max = zfs_trim_extent_bytes_max;
833 	ta.trim_extent_bytes_min = zfs_trim_extent_bytes_min;
834 	ta.trim_tree = range_tree_create(NULL, RANGE_SEG64, NULL, 0, 0);
835 	ta.trim_type = TRIM_TYPE_MANUAL;
836 	ta.trim_flags = 0;
837 
838 	/*
839 	 * When a secure TRIM has been requested infer that the intent
840 	 * is that everything must be trimmed.  Override the default
841 	 * minimum TRIM size to prevent ranges from being skipped.
842 	 */
843 	if (vd->vdev_trim_secure) {
844 		ta.trim_flags |= ZIO_TRIM_SECURE;
845 		ta.trim_extent_bytes_min = SPA_MINBLOCKSIZE;
846 	}
847 
848 	uint64_t ms_count = 0;
849 	for (uint64_t i = 0; !vd->vdev_detached &&
850 	    i < vd->vdev_top->vdev_ms_count; i++) {
851 		metaslab_t *msp = vd->vdev_top->vdev_ms[i];
852 
853 		/*
854 		 * If we've expanded the top-level vdev or it's our
855 		 * first pass, calculate our progress.
856 		 */
857 		if (vd->vdev_top->vdev_ms_count != ms_count) {
858 			vdev_trim_calculate_progress(vd);
859 			ms_count = vd->vdev_top->vdev_ms_count;
860 		}
861 
862 		spa_config_exit(spa, SCL_CONFIG, FTAG);
863 		metaslab_disable(msp);
864 		mutex_enter(&msp->ms_lock);
865 		VERIFY0(metaslab_load(msp));
866 
867 		/*
868 		 * If a partial TRIM was requested skip metaslabs which have
869 		 * never been initialized and thus have never been written.
870 		 */
871 		if (msp->ms_sm == NULL && vd->vdev_trim_partial) {
872 			mutex_exit(&msp->ms_lock);
873 			metaslab_enable(msp, B_FALSE, B_FALSE);
874 			spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
875 			vdev_trim_calculate_progress(vd);
876 			continue;
877 		}
878 
879 		ta.trim_msp = msp;
880 		range_tree_walk(msp->ms_allocatable, vdev_trim_range_add, &ta);
881 		range_tree_vacate(msp->ms_trim, NULL, NULL);
882 		mutex_exit(&msp->ms_lock);
883 
884 		error = vdev_trim_ranges(&ta);
885 		metaslab_enable(msp, B_TRUE, B_FALSE);
886 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
887 
888 		range_tree_vacate(ta.trim_tree, NULL, NULL);
889 		if (error != 0)
890 			break;
891 	}
892 
893 	spa_config_exit(spa, SCL_CONFIG, FTAG);
894 
895 	range_tree_destroy(ta.trim_tree);
896 
897 	mutex_enter(&vd->vdev_trim_lock);
898 	if (!vd->vdev_trim_exit_wanted && vdev_writeable(vd)) {
899 		vdev_trim_change_state(vd, VDEV_TRIM_COMPLETE,
900 		    vd->vdev_trim_rate, vd->vdev_trim_partial,
901 		    vd->vdev_trim_secure);
902 	}
903 	ASSERT(vd->vdev_trim_thread != NULL || vd->vdev_trim_inflight[0] == 0);
904 
905 	/*
906 	 * Drop the vdev_trim_lock while we sync out the txg since it's
907 	 * possible that a device might be trying to come online and must
908 	 * check to see if it needs to restart a trim. That thread will be
909 	 * holding the spa_config_lock which would prevent the txg_wait_synced
910 	 * from completing.
911 	 */
912 	mutex_exit(&vd->vdev_trim_lock);
913 	txg_wait_synced(spa_get_dsl(spa), 0);
914 	mutex_enter(&vd->vdev_trim_lock);
915 
916 	vd->vdev_trim_thread = NULL;
917 	cv_broadcast(&vd->vdev_trim_cv);
918 	mutex_exit(&vd->vdev_trim_lock);
919 }
920 
921 /*
922  * Initiates a manual TRIM for the vdev_t.  Callers must hold vdev_trim_lock,
923  * the vdev_t must be a leaf and cannot already be manually trimming.
924  */
925 void
vdev_trim(vdev_t * vd,uint64_t rate,boolean_t partial,boolean_t secure)926 vdev_trim(vdev_t *vd, uint64_t rate, boolean_t partial, boolean_t secure)
927 {
928 	ASSERT(MUTEX_HELD(&vd->vdev_trim_lock));
929 	ASSERT(vd->vdev_ops->vdev_op_leaf);
930 	ASSERT(vdev_is_concrete(vd));
931 	ASSERT3P(vd->vdev_trim_thread, ==, NULL);
932 	ASSERT(!vd->vdev_detached);
933 	ASSERT(!vd->vdev_trim_exit_wanted);
934 	ASSERT(!vd->vdev_top->vdev_removing);
935 
936 	vdev_trim_change_state(vd, VDEV_TRIM_ACTIVE, rate, partial, secure);
937 	vd->vdev_trim_thread = thread_create(NULL, 0,
938 	    vdev_trim_thread, vd, 0, &p0, TS_RUN, maxclsyspri);
939 }
940 
941 /*
942  * Wait for the trimming thread to be terminated (canceled or stopped).
943  */
944 static void
vdev_trim_stop_wait_impl(vdev_t * vd)945 vdev_trim_stop_wait_impl(vdev_t *vd)
946 {
947 	ASSERT(MUTEX_HELD(&vd->vdev_trim_lock));
948 
949 	while (vd->vdev_trim_thread != NULL)
950 		cv_wait(&vd->vdev_trim_cv, &vd->vdev_trim_lock);
951 
952 	ASSERT3P(vd->vdev_trim_thread, ==, NULL);
953 	vd->vdev_trim_exit_wanted = B_FALSE;
954 }
955 
956 /*
957  * Wait for vdev trim threads which were listed to cleanly exit.
958  */
959 void
vdev_trim_stop_wait(spa_t * spa,list_t * vd_list)960 vdev_trim_stop_wait(spa_t *spa, list_t *vd_list)
961 {
962 	vdev_t *vd;
963 
964 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
965 
966 	while ((vd = list_remove_head(vd_list)) != NULL) {
967 		mutex_enter(&vd->vdev_trim_lock);
968 		vdev_trim_stop_wait_impl(vd);
969 		mutex_exit(&vd->vdev_trim_lock);
970 	}
971 }
972 
973 /*
974  * Stop trimming a device, with the resultant trimming state being tgt_state.
975  * For blocking behavior pass NULL for vd_list.  Otherwise, when a list_t is
976  * provided the stopping vdev is inserted in to the list.  Callers are then
977  * required to call vdev_trim_stop_wait() to block for all the trim threads
978  * to exit.  The caller must hold vdev_trim_lock and must not be writing to
979  * the spa config, as the trimming thread may try to enter the config as a
980  * reader before exiting.
981  */
982 void
vdev_trim_stop(vdev_t * vd,vdev_trim_state_t tgt_state,list_t * vd_list)983 vdev_trim_stop(vdev_t *vd, vdev_trim_state_t tgt_state, list_t *vd_list)
984 {
985 	ASSERT(!spa_config_held(vd->vdev_spa, SCL_CONFIG|SCL_STATE, RW_WRITER));
986 	ASSERT(MUTEX_HELD(&vd->vdev_trim_lock));
987 	ASSERT(vd->vdev_ops->vdev_op_leaf);
988 	ASSERT(vdev_is_concrete(vd));
989 
990 	/*
991 	 * Allow cancel requests to proceed even if the trim thread has
992 	 * stopped.
993 	 */
994 	if (vd->vdev_trim_thread == NULL && tgt_state != VDEV_TRIM_CANCELED)
995 		return;
996 
997 	vdev_trim_change_state(vd, tgt_state, 0, 0, 0);
998 	vd->vdev_trim_exit_wanted = B_TRUE;
999 
1000 	if (vd_list == NULL) {
1001 		vdev_trim_stop_wait_impl(vd);
1002 	} else {
1003 		ASSERT(MUTEX_HELD(&spa_namespace_lock));
1004 		list_insert_tail(vd_list, vd);
1005 	}
1006 }
1007 
1008 /*
1009  * Requests that all listed vdevs stop trimming.
1010  */
1011 static void
vdev_trim_stop_all_impl(vdev_t * vd,vdev_trim_state_t tgt_state,list_t * vd_list)1012 vdev_trim_stop_all_impl(vdev_t *vd, vdev_trim_state_t tgt_state,
1013     list_t *vd_list)
1014 {
1015 	if (vd->vdev_ops->vdev_op_leaf && vdev_is_concrete(vd)) {
1016 		mutex_enter(&vd->vdev_trim_lock);
1017 		vdev_trim_stop(vd, tgt_state, vd_list);
1018 		mutex_exit(&vd->vdev_trim_lock);
1019 		return;
1020 	}
1021 
1022 	for (uint64_t i = 0; i < vd->vdev_children; i++) {
1023 		vdev_trim_stop_all_impl(vd->vdev_child[i], tgt_state,
1024 		    vd_list);
1025 	}
1026 }
1027 
1028 /*
1029  * Convenience function to stop trimming of a vdev tree and set all trim
1030  * thread pointers to NULL.
1031  */
1032 void
vdev_trim_stop_all(vdev_t * vd,vdev_trim_state_t tgt_state)1033 vdev_trim_stop_all(vdev_t *vd, vdev_trim_state_t tgt_state)
1034 {
1035 	spa_t *spa = vd->vdev_spa;
1036 	list_t vd_list;
1037 
1038 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
1039 
1040 	list_create(&vd_list, sizeof (vdev_t),
1041 	    offsetof(vdev_t, vdev_trim_node));
1042 
1043 	vdev_trim_stop_all_impl(vd, tgt_state, &vd_list);
1044 	vdev_trim_stop_wait(spa, &vd_list);
1045 
1046 	if (vd->vdev_spa->spa_sync_on) {
1047 		/* Make sure that our state has been synced to disk */
1048 		txg_wait_synced(spa_get_dsl(vd->vdev_spa), 0);
1049 	}
1050 
1051 	list_destroy(&vd_list);
1052 }
1053 
1054 /*
1055  * Conditionally restarts a manual TRIM given its on-disk state.
1056  */
1057 void
vdev_trim_restart(vdev_t * vd)1058 vdev_trim_restart(vdev_t *vd)
1059 {
1060 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
1061 	ASSERT(!spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER));
1062 
1063 	if (vd->vdev_leaf_zap != 0) {
1064 		mutex_enter(&vd->vdev_trim_lock);
1065 		uint64_t trim_state = VDEV_TRIM_NONE;
1066 		int err = zap_lookup(vd->vdev_spa->spa_meta_objset,
1067 		    vd->vdev_leaf_zap, VDEV_LEAF_ZAP_TRIM_STATE,
1068 		    sizeof (trim_state), 1, &trim_state);
1069 		ASSERT(err == 0 || err == ENOENT);
1070 		vd->vdev_trim_state = trim_state;
1071 
1072 		uint64_t timestamp = 0;
1073 		err = zap_lookup(vd->vdev_spa->spa_meta_objset,
1074 		    vd->vdev_leaf_zap, VDEV_LEAF_ZAP_TRIM_ACTION_TIME,
1075 		    sizeof (timestamp), 1, &timestamp);
1076 		ASSERT(err == 0 || err == ENOENT);
1077 		vd->vdev_trim_action_time = (time_t)timestamp;
1078 
1079 		if (vd->vdev_trim_state == VDEV_TRIM_SUSPENDED ||
1080 		    vd->vdev_offline) {
1081 			/* load progress for reporting, but don't resume */
1082 			VERIFY0(vdev_trim_load(vd));
1083 		} else if (vd->vdev_trim_state == VDEV_TRIM_ACTIVE &&
1084 		    vdev_writeable(vd) && !vd->vdev_top->vdev_removing &&
1085 		    vd->vdev_trim_thread == NULL) {
1086 			VERIFY0(vdev_trim_load(vd));
1087 			vdev_trim(vd, vd->vdev_trim_rate,
1088 			    vd->vdev_trim_partial, vd->vdev_trim_secure);
1089 		}
1090 
1091 		mutex_exit(&vd->vdev_trim_lock);
1092 	}
1093 
1094 	for (uint64_t i = 0; i < vd->vdev_children; i++) {
1095 		vdev_trim_restart(vd->vdev_child[i]);
1096 	}
1097 }
1098 
1099 /*
1100  * Used by the automatic TRIM when ZFS_DEBUG_TRIM is set to verify that
1101  * every TRIM range is contained within ms_allocatable.
1102  */
1103 static void
vdev_trim_range_verify(void * arg,uint64_t start,uint64_t size)1104 vdev_trim_range_verify(void *arg, uint64_t start, uint64_t size)
1105 {
1106 	trim_args_t *ta = arg;
1107 	metaslab_t *msp = ta->trim_msp;
1108 
1109 	VERIFY3B(msp->ms_loaded, ==, B_TRUE);
1110 	VERIFY3U(msp->ms_disabled, >, 0);
1111 	VERIFY(range_tree_contains(msp->ms_allocatable, start, size));
1112 }
1113 
1114 /*
1115  * Each automatic TRIM thread is responsible for managing the trimming of a
1116  * top-level vdev in the pool.  No automatic TRIM state is maintained on-disk.
1117  *
1118  * N.B. This behavior is different from a manual TRIM where a thread
1119  * is created for each leaf vdev, instead of each top-level vdev.
1120  */
1121 static void
vdev_autotrim_thread(void * arg)1122 vdev_autotrim_thread(void *arg)
1123 {
1124 	vdev_t *vd = arg;
1125 	spa_t *spa = vd->vdev_spa;
1126 	int shift = 0;
1127 
1128 	mutex_enter(&vd->vdev_autotrim_lock);
1129 	ASSERT3P(vd->vdev_top, ==, vd);
1130 	ASSERT3P(vd->vdev_autotrim_thread, !=, NULL);
1131 	mutex_exit(&vd->vdev_autotrim_lock);
1132 	spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
1133 
1134 	uint64_t extent_bytes_max = zfs_trim_extent_bytes_max;
1135 	uint64_t extent_bytes_min = zfs_trim_extent_bytes_min;
1136 
1137 	while (!vdev_autotrim_should_stop(vd)) {
1138 		int txgs_per_trim = MAX(zfs_trim_txg_batch, 1);
1139 		boolean_t issued_trim = B_FALSE;
1140 
1141 		/*
1142 		 * All of the metaslabs are divided in to groups of size
1143 		 * num_metaslabs / zfs_trim_txg_batch.  Each of these groups
1144 		 * is composed of metaslabs which are spread evenly over the
1145 		 * device.
1146 		 *
1147 		 * For example, when zfs_trim_txg_batch = 32 (default) then
1148 		 * group 0 will contain metaslabs 0, 32, 64, ...;
1149 		 * group 1 will contain metaslabs 1, 33, 65, ...;
1150 		 * group 2 will contain metaslabs 2, 34, 66, ...; and so on.
1151 		 *
1152 		 * On each pass through the while() loop one of these groups
1153 		 * is selected.  This is accomplished by using a shift value
1154 		 * to select the starting metaslab, then striding over the
1155 		 * metaslabs using the zfs_trim_txg_batch size.  This is
1156 		 * done to accomplish two things.
1157 		 *
1158 		 * 1) By dividing the metaslabs into groups, and making sure
1159 		 *    that each group takes a minimum of one txg to process.
1160 		 *    Then zfs_trim_txg_batch controls the minimum number of
1161 		 *    txgs which must occur before a metaslab is revisited.
1162 		 *
1163 		 * 2) Selecting non-consecutive metaslabs distributes the
1164 		 *    TRIM commands for a group evenly over the entire device.
1165 		 *    This can be advantageous for certain types of devices.
1166 		 */
1167 		for (uint64_t i = shift % txgs_per_trim; i < vd->vdev_ms_count;
1168 		    i += txgs_per_trim) {
1169 			metaslab_t *msp = vd->vdev_ms[i];
1170 			range_tree_t *trim_tree;
1171 
1172 			spa_config_exit(spa, SCL_CONFIG, FTAG);
1173 			metaslab_disable(msp);
1174 			spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
1175 
1176 			mutex_enter(&msp->ms_lock);
1177 
1178 			/*
1179 			 * Skip the metaslab when it has never been allocated
1180 			 * or when there are no recent frees to trim.
1181 			 */
1182 			if (msp->ms_sm == NULL ||
1183 			    range_tree_is_empty(msp->ms_trim)) {
1184 				mutex_exit(&msp->ms_lock);
1185 				metaslab_enable(msp, B_FALSE, B_FALSE);
1186 				continue;
1187 			}
1188 
1189 			/*
1190 			 * Skip the metaslab when it has already been disabled.
1191 			 * This may happen when a manual TRIM or initialize
1192 			 * operation is running concurrently.  In the case
1193 			 * of a manual TRIM, the ms_trim tree will have been
1194 			 * vacated.  Only ranges added after the manual TRIM
1195 			 * disabled the metaslab will be included in the tree.
1196 			 * These will be processed when the automatic TRIM
1197 			 * next revisits this metaslab.
1198 			 */
1199 			if (msp->ms_disabled > 1) {
1200 				mutex_exit(&msp->ms_lock);
1201 				metaslab_enable(msp, B_FALSE, B_FALSE);
1202 				continue;
1203 			}
1204 
1205 			/*
1206 			 * Allocate an empty range tree which is swapped in
1207 			 * for the existing ms_trim tree while it is processed.
1208 			 */
1209 			trim_tree = range_tree_create(NULL, RANGE_SEG64, NULL,
1210 			    0, 0);
1211 			range_tree_swap(&msp->ms_trim, &trim_tree);
1212 			ASSERT(range_tree_is_empty(msp->ms_trim));
1213 
1214 			/*
1215 			 * There are two cases when constructing the per-vdev
1216 			 * trim trees for a metaslab.  If the top-level vdev
1217 			 * has no children then it is also a leaf and should
1218 			 * be trimmed.  Otherwise our children are the leaves
1219 			 * and a trim tree should be constructed for each.
1220 			 */
1221 			trim_args_t *tap;
1222 			uint64_t children = vd->vdev_children;
1223 			if (children == 0) {
1224 				children = 1;
1225 				tap = kmem_zalloc(sizeof (trim_args_t) *
1226 				    children, KM_SLEEP);
1227 				tap[0].trim_vdev = vd;
1228 			} else {
1229 				tap = kmem_zalloc(sizeof (trim_args_t) *
1230 				    children, KM_SLEEP);
1231 
1232 				for (uint64_t c = 0; c < children; c++) {
1233 					tap[c].trim_vdev = vd->vdev_child[c];
1234 				}
1235 			}
1236 
1237 			for (uint64_t c = 0; c < children; c++) {
1238 				trim_args_t *ta = &tap[c];
1239 				vdev_t *cvd = ta->trim_vdev;
1240 
1241 				ta->trim_msp = msp;
1242 				ta->trim_extent_bytes_max = extent_bytes_max;
1243 				ta->trim_extent_bytes_min = extent_bytes_min;
1244 				ta->trim_type = TRIM_TYPE_AUTO;
1245 				ta->trim_flags = 0;
1246 
1247 				if (cvd->vdev_detached ||
1248 				    !vdev_writeable(cvd) ||
1249 				    !cvd->vdev_has_trim ||
1250 				    cvd->vdev_trim_thread != NULL) {
1251 					continue;
1252 				}
1253 
1254 				/*
1255 				 * When a device has an attached hot spare, or
1256 				 * is being replaced it will not be trimmed.
1257 				 * This is done to avoid adding additional
1258 				 * stress to a potentially unhealthy device,
1259 				 * and to minimize the required rebuild time.
1260 				 */
1261 				if (!cvd->vdev_ops->vdev_op_leaf)
1262 					continue;
1263 
1264 				ta->trim_tree = range_tree_create(NULL,
1265 				    RANGE_SEG64, NULL, 0, 0);
1266 				range_tree_walk(trim_tree,
1267 				    vdev_trim_range_add, ta);
1268 			}
1269 
1270 			mutex_exit(&msp->ms_lock);
1271 			spa_config_exit(spa, SCL_CONFIG, FTAG);
1272 
1273 			/*
1274 			 * Issue the TRIM I/Os for all ranges covered by the
1275 			 * TRIM trees.  These ranges are safe to TRIM because
1276 			 * no new allocations will be performed until the call
1277 			 * to metaslab_enabled() below.
1278 			 */
1279 			for (uint64_t c = 0; c < children; c++) {
1280 				trim_args_t *ta = &tap[c];
1281 
1282 				/*
1283 				 * Always yield to a manual TRIM if one has
1284 				 * been started for the child vdev.
1285 				 */
1286 				if (ta->trim_tree == NULL ||
1287 				    ta->trim_vdev->vdev_trim_thread != NULL) {
1288 					continue;
1289 				}
1290 
1291 				/*
1292 				 * After this point metaslab_enable() must be
1293 				 * called with the sync flag set.  This is done
1294 				 * here because vdev_trim_ranges() is allowed
1295 				 * to be interrupted (EINTR) before issuing all
1296 				 * of the required TRIM I/Os.
1297 				 */
1298 				issued_trim = B_TRUE;
1299 
1300 				int error = vdev_trim_ranges(ta);
1301 				if (error)
1302 					break;
1303 			}
1304 
1305 			/*
1306 			 * Verify every range which was trimmed is still
1307 			 * contained within the ms_allocatable tree.
1308 			 */
1309 			if (zfs_flags & ZFS_DEBUG_TRIM) {
1310 				mutex_enter(&msp->ms_lock);
1311 				VERIFY0(metaslab_load(msp));
1312 				VERIFY3P(tap[0].trim_msp, ==, msp);
1313 				range_tree_walk(trim_tree,
1314 				    vdev_trim_range_verify, &tap[0]);
1315 				mutex_exit(&msp->ms_lock);
1316 			}
1317 
1318 			range_tree_vacate(trim_tree, NULL, NULL);
1319 			range_tree_destroy(trim_tree);
1320 
1321 			metaslab_enable(msp, issued_trim, B_FALSE);
1322 			spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
1323 
1324 			for (uint64_t c = 0; c < children; c++) {
1325 				trim_args_t *ta = &tap[c];
1326 
1327 				if (ta->trim_tree == NULL)
1328 					continue;
1329 
1330 				range_tree_vacate(ta->trim_tree, NULL, NULL);
1331 				range_tree_destroy(ta->trim_tree);
1332 			}
1333 
1334 			kmem_free(tap, sizeof (trim_args_t) * children);
1335 		}
1336 
1337 		spa_config_exit(spa, SCL_CONFIG, FTAG);
1338 
1339 		/*
1340 		 * After completing the group of metaslabs wait for the next
1341 		 * open txg.  This is done to make sure that a minimum of
1342 		 * zfs_trim_txg_batch txgs will occur before these metaslabs
1343 		 * are trimmed again.
1344 		 */
1345 		txg_wait_open(spa_get_dsl(spa), 0, issued_trim);
1346 
1347 		shift++;
1348 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
1349 	}
1350 
1351 	for (uint64_t c = 0; c < vd->vdev_children; c++) {
1352 		vdev_t *cvd = vd->vdev_child[c];
1353 		mutex_enter(&cvd->vdev_trim_io_lock);
1354 
1355 		while (cvd->vdev_trim_inflight[1] > 0) {
1356 			cv_wait(&cvd->vdev_trim_io_cv,
1357 			    &cvd->vdev_trim_io_lock);
1358 		}
1359 		mutex_exit(&cvd->vdev_trim_io_lock);
1360 	}
1361 
1362 	spa_config_exit(spa, SCL_CONFIG, FTAG);
1363 
1364 	/*
1365 	 * When exiting because the autotrim property was set to off, then
1366 	 * abandon any unprocessed ms_trim ranges to reclaim the memory.
1367 	 */
1368 	if (spa_get_autotrim(spa) == SPA_AUTOTRIM_OFF) {
1369 		for (uint64_t i = 0; i < vd->vdev_ms_count; i++) {
1370 			metaslab_t *msp = vd->vdev_ms[i];
1371 
1372 			mutex_enter(&msp->ms_lock);
1373 			range_tree_vacate(msp->ms_trim, NULL, NULL);
1374 			mutex_exit(&msp->ms_lock);
1375 		}
1376 	}
1377 
1378 	mutex_enter(&vd->vdev_autotrim_lock);
1379 	ASSERT(vd->vdev_autotrim_thread != NULL);
1380 	vd->vdev_autotrim_thread = NULL;
1381 	cv_broadcast(&vd->vdev_autotrim_cv);
1382 	mutex_exit(&vd->vdev_autotrim_lock);
1383 }
1384 
1385 /*
1386  * Starts an autotrim thread, if needed, for each top-level vdev which can be
1387  * trimmed.  A top-level vdev which has been evacuated will never be trimmed.
1388  */
1389 void
vdev_autotrim(spa_t * spa)1390 vdev_autotrim(spa_t *spa)
1391 {
1392 	vdev_t *root_vd = spa->spa_root_vdev;
1393 
1394 	for (uint64_t i = 0; i < root_vd->vdev_children; i++) {
1395 		vdev_t *tvd = root_vd->vdev_child[i];
1396 
1397 		mutex_enter(&tvd->vdev_autotrim_lock);
1398 		if (vdev_writeable(tvd) && !tvd->vdev_removing &&
1399 		    tvd->vdev_autotrim_thread == NULL) {
1400 			ASSERT3P(tvd->vdev_top, ==, tvd);
1401 
1402 			tvd->vdev_autotrim_thread = thread_create(NULL, 0,
1403 			    vdev_autotrim_thread, tvd, 0, &p0, TS_RUN,
1404 			    maxclsyspri);
1405 			ASSERT(tvd->vdev_autotrim_thread != NULL);
1406 		}
1407 		mutex_exit(&tvd->vdev_autotrim_lock);
1408 	}
1409 }
1410 
1411 /*
1412  * Wait for the vdev_autotrim_thread associated with the passed top-level
1413  * vdev to be terminated (canceled or stopped).
1414  */
1415 void
vdev_autotrim_stop_wait(vdev_t * tvd)1416 vdev_autotrim_stop_wait(vdev_t *tvd)
1417 {
1418 	mutex_enter(&tvd->vdev_autotrim_lock);
1419 	if (tvd->vdev_autotrim_thread != NULL) {
1420 		tvd->vdev_autotrim_exit_wanted = B_TRUE;
1421 
1422 		while (tvd->vdev_autotrim_thread != NULL) {
1423 			cv_wait(&tvd->vdev_autotrim_cv,
1424 			    &tvd->vdev_autotrim_lock);
1425 		}
1426 
1427 		ASSERT3P(tvd->vdev_autotrim_thread, ==, NULL);
1428 		tvd->vdev_autotrim_exit_wanted = B_FALSE;
1429 	}
1430 	mutex_exit(&tvd->vdev_autotrim_lock);
1431 }
1432 
1433 /*
1434  * Wait for all of the vdev_autotrim_thread associated with the pool to
1435  * be terminated (canceled or stopped).
1436  */
1437 void
vdev_autotrim_stop_all(spa_t * spa)1438 vdev_autotrim_stop_all(spa_t *spa)
1439 {
1440 	vdev_t *root_vd = spa->spa_root_vdev;
1441 
1442 	for (uint64_t i = 0; i < root_vd->vdev_children; i++)
1443 		vdev_autotrim_stop_wait(root_vd->vdev_child[i]);
1444 }
1445 
1446 /*
1447  * Conditionally restart all of the vdev_autotrim_thread's for the pool.
1448  */
1449 void
vdev_autotrim_restart(spa_t * spa)1450 vdev_autotrim_restart(spa_t *spa)
1451 {
1452 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
1453 
1454 	if (spa->spa_autotrim)
1455 		vdev_autotrim(spa);
1456 }
1457