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 https://opensource.org/licenses/CDDL-1.0.
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, 2024 by Delphix. All rights reserved.
24 */
25
26 #include <sys/spa.h>
27 #include <sys/spa_impl.h>
28 #include <sys/txg.h>
29 #include <sys/vdev_impl.h>
30 #include <sys/metaslab_impl.h>
31 #include <sys/dsl_synctask.h>
32 #include <sys/zap.h>
33 #include <sys/dmu_tx.h>
34 #include <sys/vdev_initialize.h>
35
36 /*
37 * Value that is written to disk during initialization.
38 */
39 static uint64_t zfs_initialize_value = 0xdeadbeefdeadbeeeULL;
40
41 /* maximum number of I/Os outstanding per leaf vdev */
42 static const int zfs_initialize_limit = 1;
43
44 /* size of initializing writes; default 1MiB, see zfs_remove_max_segment */
45 static uint64_t zfs_initialize_chunk_size = 1024 * 1024;
46
47 static boolean_t
vdev_initialize_should_stop(vdev_t * vd)48 vdev_initialize_should_stop(vdev_t *vd)
49 {
50 return (vd->vdev_initialize_exit_wanted || !vdev_writeable(vd) ||
51 vd->vdev_detached || vd->vdev_top->vdev_removing ||
52 vd->vdev_top->vdev_rz_expanding);
53 }
54
55 static void
vdev_initialize_zap_update_sync(void * arg,dmu_tx_t * tx)56 vdev_initialize_zap_update_sync(void *arg, dmu_tx_t *tx)
57 {
58 /*
59 * We pass in the guid instead of the vdev_t since the vdev may
60 * have been freed prior to the sync task being processed. This
61 * happens when a vdev is detached as we call spa_config_vdev_exit(),
62 * stop the initializing thread, schedule the sync task, and free
63 * the vdev. Later when the scheduled sync task is invoked, it would
64 * find that the vdev has been freed.
65 */
66 uint64_t guid = *(uint64_t *)arg;
67 uint64_t txg = dmu_tx_get_txg(tx);
68 kmem_free(arg, sizeof (uint64_t));
69
70 vdev_t *vd = spa_lookup_by_guid(tx->tx_pool->dp_spa, guid, B_FALSE);
71 if (vd == NULL || vd->vdev_top->vdev_removing ||
72 !vdev_is_concrete(vd) || vd->vdev_top->vdev_rz_expanding)
73 return;
74
75 uint64_t last_offset = vd->vdev_initialize_offset[txg & TXG_MASK];
76 vd->vdev_initialize_offset[txg & TXG_MASK] = 0;
77
78 VERIFY(vd->vdev_leaf_zap != 0);
79
80 objset_t *mos = vd->vdev_spa->spa_meta_objset;
81
82 if (last_offset > 0) {
83 vd->vdev_initialize_last_offset = last_offset;
84 VERIFY0(zap_update(mos, vd->vdev_leaf_zap,
85 VDEV_LEAF_ZAP_INITIALIZE_LAST_OFFSET,
86 sizeof (last_offset), 1, &last_offset, tx));
87 }
88 if (vd->vdev_initialize_action_time > 0) {
89 uint64_t val = (uint64_t)vd->vdev_initialize_action_time;
90 VERIFY0(zap_update(mos, vd->vdev_leaf_zap,
91 VDEV_LEAF_ZAP_INITIALIZE_ACTION_TIME, sizeof (val),
92 1, &val, tx));
93 }
94
95 uint64_t initialize_state = vd->vdev_initialize_state;
96 VERIFY0(zap_update(mos, vd->vdev_leaf_zap,
97 VDEV_LEAF_ZAP_INITIALIZE_STATE, sizeof (initialize_state), 1,
98 &initialize_state, tx));
99 }
100
101 static void
vdev_initialize_zap_remove_sync(void * arg,dmu_tx_t * tx)102 vdev_initialize_zap_remove_sync(void *arg, dmu_tx_t *tx)
103 {
104 uint64_t guid = *(uint64_t *)arg;
105
106 kmem_free(arg, sizeof (uint64_t));
107
108 vdev_t *vd = spa_lookup_by_guid(tx->tx_pool->dp_spa, guid, B_FALSE);
109 if (vd == NULL || vd->vdev_top->vdev_removing || !vdev_is_concrete(vd))
110 return;
111
112 ASSERT3S(vd->vdev_initialize_state, ==, VDEV_INITIALIZE_NONE);
113 ASSERT3U(vd->vdev_leaf_zap, !=, 0);
114
115 vd->vdev_initialize_last_offset = 0;
116 vd->vdev_initialize_action_time = 0;
117
118 objset_t *mos = vd->vdev_spa->spa_meta_objset;
119 int error;
120
121 error = zap_remove(mos, vd->vdev_leaf_zap,
122 VDEV_LEAF_ZAP_INITIALIZE_LAST_OFFSET, tx);
123 VERIFY(error == 0 || error == ENOENT);
124
125 error = zap_remove(mos, vd->vdev_leaf_zap,
126 VDEV_LEAF_ZAP_INITIALIZE_STATE, tx);
127 VERIFY(error == 0 || error == ENOENT);
128
129 error = zap_remove(mos, vd->vdev_leaf_zap,
130 VDEV_LEAF_ZAP_INITIALIZE_ACTION_TIME, tx);
131 VERIFY(error == 0 || error == ENOENT);
132 }
133
134 static void
vdev_initialize_change_state(vdev_t * vd,vdev_initializing_state_t new_state)135 vdev_initialize_change_state(vdev_t *vd, vdev_initializing_state_t new_state)
136 {
137 ASSERT(MUTEX_HELD(&vd->vdev_initialize_lock));
138 spa_t *spa = vd->vdev_spa;
139
140 if (new_state == vd->vdev_initialize_state)
141 return;
142
143 /*
144 * Copy the vd's guid, this will be freed by the sync task.
145 */
146 uint64_t *guid = kmem_zalloc(sizeof (uint64_t), KM_SLEEP);
147 *guid = vd->vdev_guid;
148
149 /*
150 * If we're suspending, then preserving the original start time.
151 */
152 if (vd->vdev_initialize_state != VDEV_INITIALIZE_SUSPENDED) {
153 vd->vdev_initialize_action_time = gethrestime_sec();
154 }
155
156 vdev_initializing_state_t old_state = vd->vdev_initialize_state;
157 vd->vdev_initialize_state = new_state;
158
159 dmu_tx_t *tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir);
160 VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
161
162 if (new_state != VDEV_INITIALIZE_NONE) {
163 dsl_sync_task_nowait(spa_get_dsl(spa),
164 vdev_initialize_zap_update_sync, guid, tx);
165 } else {
166 dsl_sync_task_nowait(spa_get_dsl(spa),
167 vdev_initialize_zap_remove_sync, guid, tx);
168 }
169
170 switch (new_state) {
171 case VDEV_INITIALIZE_ACTIVE:
172 spa_history_log_internal(spa, "initialize", tx,
173 "vdev=%s activated", vd->vdev_path);
174 break;
175 case VDEV_INITIALIZE_SUSPENDED:
176 spa_history_log_internal(spa, "initialize", tx,
177 "vdev=%s suspended", vd->vdev_path);
178 break;
179 case VDEV_INITIALIZE_CANCELED:
180 if (old_state == VDEV_INITIALIZE_ACTIVE ||
181 old_state == VDEV_INITIALIZE_SUSPENDED)
182 spa_history_log_internal(spa, "initialize", tx,
183 "vdev=%s canceled", vd->vdev_path);
184 break;
185 case VDEV_INITIALIZE_COMPLETE:
186 spa_history_log_internal(spa, "initialize", tx,
187 "vdev=%s complete", vd->vdev_path);
188 break;
189 case VDEV_INITIALIZE_NONE:
190 spa_history_log_internal(spa, "uninitialize", tx,
191 "vdev=%s", vd->vdev_path);
192 break;
193 default:
194 panic("invalid state %llu", (unsigned long long)new_state);
195 }
196
197 dmu_tx_commit(tx);
198
199 if (new_state != VDEV_INITIALIZE_ACTIVE)
200 spa_notify_waiters(spa);
201 }
202
203 static void
vdev_initialize_cb(zio_t * zio)204 vdev_initialize_cb(zio_t *zio)
205 {
206 vdev_t *vd = zio->io_vd;
207 mutex_enter(&vd->vdev_initialize_io_lock);
208 if (zio->io_error == ENXIO && !vdev_writeable(vd)) {
209 /*
210 * The I/O failed because the vdev was unavailable; roll the
211 * last offset back. (This works because spa_sync waits on
212 * spa_txg_zio before it runs sync tasks.)
213 */
214 uint64_t *off =
215 &vd->vdev_initialize_offset[zio->io_txg & TXG_MASK];
216 *off = MIN(*off, zio->io_offset);
217 } else {
218 /*
219 * Since initializing is best-effort, we ignore I/O errors and
220 * rely on vdev_probe to determine if the errors are more
221 * critical.
222 */
223 if (zio->io_error != 0)
224 vd->vdev_stat.vs_initialize_errors++;
225
226 vd->vdev_initialize_bytes_done += zio->io_orig_size;
227 }
228 ASSERT3U(vd->vdev_initialize_inflight, >, 0);
229 vd->vdev_initialize_inflight--;
230 cv_broadcast(&vd->vdev_initialize_io_cv);
231 mutex_exit(&vd->vdev_initialize_io_lock);
232
233 spa_config_exit(vd->vdev_spa, SCL_STATE_ALL, vd);
234 }
235
236 /* Takes care of physical writing and limiting # of concurrent ZIOs. */
237 static int
vdev_initialize_write(vdev_t * vd,uint64_t start,uint64_t size,abd_t * data)238 vdev_initialize_write(vdev_t *vd, uint64_t start, uint64_t size, abd_t *data)
239 {
240 spa_t *spa = vd->vdev_spa;
241
242 /* Limit inflight initializing I/Os */
243 mutex_enter(&vd->vdev_initialize_io_lock);
244 while (vd->vdev_initialize_inflight >= zfs_initialize_limit) {
245 cv_wait(&vd->vdev_initialize_io_cv,
246 &vd->vdev_initialize_io_lock);
247 }
248 vd->vdev_initialize_inflight++;
249 mutex_exit(&vd->vdev_initialize_io_lock);
250
251 dmu_tx_t *tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir);
252 VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
253 uint64_t txg = dmu_tx_get_txg(tx);
254
255 spa_config_enter(spa, SCL_STATE_ALL, vd, RW_READER);
256 mutex_enter(&vd->vdev_initialize_lock);
257
258 if (vd->vdev_initialize_offset[txg & TXG_MASK] == 0) {
259 uint64_t *guid = kmem_zalloc(sizeof (uint64_t), KM_SLEEP);
260 *guid = vd->vdev_guid;
261
262 /* This is the first write of this txg. */
263 dsl_sync_task_nowait(spa_get_dsl(spa),
264 vdev_initialize_zap_update_sync, guid, tx);
265 }
266
267 /*
268 * We know the vdev struct will still be around since all
269 * consumers of vdev_free must stop the initialization first.
270 */
271 if (vdev_initialize_should_stop(vd)) {
272 mutex_enter(&vd->vdev_initialize_io_lock);
273 ASSERT3U(vd->vdev_initialize_inflight, >, 0);
274 vd->vdev_initialize_inflight--;
275 mutex_exit(&vd->vdev_initialize_io_lock);
276 spa_config_exit(vd->vdev_spa, SCL_STATE_ALL, vd);
277 mutex_exit(&vd->vdev_initialize_lock);
278 dmu_tx_commit(tx);
279 return (SET_ERROR(EINTR));
280 }
281 mutex_exit(&vd->vdev_initialize_lock);
282
283 vd->vdev_initialize_offset[txg & TXG_MASK] = start + size;
284 zio_nowait(zio_write_phys(spa->spa_txg_zio[txg & TXG_MASK], vd, start,
285 size, data, ZIO_CHECKSUM_OFF, vdev_initialize_cb, NULL,
286 ZIO_PRIORITY_INITIALIZING, ZIO_FLAG_CANFAIL, B_FALSE));
287 /* vdev_initialize_cb releases SCL_STATE_ALL */
288
289 dmu_tx_commit(tx);
290
291 return (0);
292 }
293
294 /*
295 * Callback to fill each ABD chunk with zfs_initialize_value. len must be
296 * divisible by sizeof (uint64_t), and buf must be 8-byte aligned. The ABD
297 * allocation will guarantee these for us.
298 */
299 static int
vdev_initialize_block_fill(void * buf,size_t len,void * unused)300 vdev_initialize_block_fill(void *buf, size_t len, void *unused)
301 {
302 (void) unused;
303
304 ASSERT0(len % sizeof (uint64_t));
305 for (uint64_t i = 0; i < len; i += sizeof (uint64_t)) {
306 *(uint64_t *)((char *)(buf) + i) = zfs_initialize_value;
307 }
308 return (0);
309 }
310
311 static abd_t *
vdev_initialize_block_alloc(void)312 vdev_initialize_block_alloc(void)
313 {
314 /* Allocate ABD for filler data */
315 abd_t *data = abd_alloc_for_io(zfs_initialize_chunk_size, B_FALSE);
316
317 ASSERT0(zfs_initialize_chunk_size % sizeof (uint64_t));
318 (void) abd_iterate_func(data, 0, zfs_initialize_chunk_size,
319 vdev_initialize_block_fill, NULL);
320
321 return (data);
322 }
323
324 static void
vdev_initialize_block_free(abd_t * data)325 vdev_initialize_block_free(abd_t *data)
326 {
327 abd_free(data);
328 }
329
330 static int
vdev_initialize_ranges(vdev_t * vd,abd_t * data)331 vdev_initialize_ranges(vdev_t *vd, abd_t *data)
332 {
333 zfs_range_tree_t *rt = vd->vdev_initialize_tree;
334 zfs_btree_t *bt = &rt->rt_root;
335 zfs_btree_index_t where;
336
337 for (zfs_range_seg_t *rs = zfs_btree_first(bt, &where); rs != NULL;
338 rs = zfs_btree_next(bt, &where, &where)) {
339 uint64_t size = zfs_rs_get_end(rs, rt) -
340 zfs_rs_get_start(rs, rt);
341
342 /* Split range into legally-sized physical chunks */
343 uint64_t writes_required =
344 ((size - 1) / zfs_initialize_chunk_size) + 1;
345
346 for (uint64_t w = 0; w < writes_required; w++) {
347 int error;
348
349 error = vdev_initialize_write(vd,
350 VDEV_LABEL_START_SIZE + zfs_rs_get_start(rs, rt) +
351 (w * zfs_initialize_chunk_size),
352 MIN(size - (w * zfs_initialize_chunk_size),
353 zfs_initialize_chunk_size), data);
354 if (error != 0)
355 return (error);
356 }
357 }
358 return (0);
359 }
360
361 static void
vdev_initialize_xlate_last_rs_end(void * arg,zfs_range_seg64_t * physical_rs)362 vdev_initialize_xlate_last_rs_end(void *arg, zfs_range_seg64_t *physical_rs)
363 {
364 uint64_t *last_rs_end = (uint64_t *)arg;
365
366 if (physical_rs->rs_end > *last_rs_end)
367 *last_rs_end = physical_rs->rs_end;
368 }
369
370 static void
vdev_initialize_xlate_progress(void * arg,zfs_range_seg64_t * physical_rs)371 vdev_initialize_xlate_progress(void *arg, zfs_range_seg64_t *physical_rs)
372 {
373 vdev_t *vd = (vdev_t *)arg;
374
375 uint64_t size = physical_rs->rs_end - physical_rs->rs_start;
376 vd->vdev_initialize_bytes_est += size;
377
378 if (vd->vdev_initialize_last_offset > physical_rs->rs_end) {
379 vd->vdev_initialize_bytes_done += size;
380 } else if (vd->vdev_initialize_last_offset > physical_rs->rs_start &&
381 vd->vdev_initialize_last_offset < physical_rs->rs_end) {
382 vd->vdev_initialize_bytes_done +=
383 vd->vdev_initialize_last_offset - physical_rs->rs_start;
384 }
385 }
386
387 static void
vdev_initialize_calculate_progress(vdev_t * vd)388 vdev_initialize_calculate_progress(vdev_t *vd)
389 {
390 ASSERT(spa_config_held(vd->vdev_spa, SCL_CONFIG, RW_READER) ||
391 spa_config_held(vd->vdev_spa, SCL_CONFIG, RW_WRITER));
392 ASSERT(vd->vdev_leaf_zap != 0);
393
394 vd->vdev_initialize_bytes_est = 0;
395 vd->vdev_initialize_bytes_done = 0;
396
397 for (uint64_t i = 0; i < vd->vdev_top->vdev_ms_count; i++) {
398 metaslab_t *msp = vd->vdev_top->vdev_ms[i];
399 mutex_enter(&msp->ms_lock);
400
401 uint64_t ms_free = (msp->ms_size -
402 metaslab_allocated_space(msp)) /
403 vdev_get_ndisks(vd->vdev_top);
404
405 /*
406 * Convert the metaslab range to a physical range
407 * on our vdev. We use this to determine if we are
408 * in the middle of this metaslab range.
409 */
410 zfs_range_seg64_t logical_rs, physical_rs, remain_rs;
411 logical_rs.rs_start = msp->ms_start;
412 logical_rs.rs_end = msp->ms_start + msp->ms_size;
413
414 /* Metaslab space after this offset has not been initialized */
415 vdev_xlate(vd, &logical_rs, &physical_rs, &remain_rs);
416 if (vd->vdev_initialize_last_offset <= physical_rs.rs_start) {
417 vd->vdev_initialize_bytes_est += ms_free;
418 mutex_exit(&msp->ms_lock);
419 continue;
420 }
421
422 /* Metaslab space before this offset has been initialized */
423 uint64_t last_rs_end = physical_rs.rs_end;
424 if (!vdev_xlate_is_empty(&remain_rs)) {
425 vdev_xlate_walk(vd, &remain_rs,
426 vdev_initialize_xlate_last_rs_end, &last_rs_end);
427 }
428
429 if (vd->vdev_initialize_last_offset > last_rs_end) {
430 vd->vdev_initialize_bytes_done += ms_free;
431 vd->vdev_initialize_bytes_est += ms_free;
432 mutex_exit(&msp->ms_lock);
433 continue;
434 }
435
436 /*
437 * If we get here, we're in the middle of initializing this
438 * metaslab. Load it and walk the free tree for more accurate
439 * progress estimation.
440 */
441 VERIFY0(metaslab_load(msp));
442
443 zfs_btree_index_t where;
444 zfs_range_tree_t *rt = msp->ms_allocatable;
445 for (zfs_range_seg_t *rs =
446 zfs_btree_first(&rt->rt_root, &where); rs;
447 rs = zfs_btree_next(&rt->rt_root, &where,
448 &where)) {
449 logical_rs.rs_start = zfs_rs_get_start(rs, rt);
450 logical_rs.rs_end = zfs_rs_get_end(rs, rt);
451
452 vdev_xlate_walk(vd, &logical_rs,
453 vdev_initialize_xlate_progress, vd);
454 }
455 mutex_exit(&msp->ms_lock);
456 }
457 }
458
459 static int
vdev_initialize_load(vdev_t * vd)460 vdev_initialize_load(vdev_t *vd)
461 {
462 int err = 0;
463 ASSERT(spa_config_held(vd->vdev_spa, SCL_CONFIG, RW_READER) ||
464 spa_config_held(vd->vdev_spa, SCL_CONFIG, RW_WRITER));
465 ASSERT(vd->vdev_leaf_zap != 0);
466
467 if (vd->vdev_initialize_state == VDEV_INITIALIZE_ACTIVE ||
468 vd->vdev_initialize_state == VDEV_INITIALIZE_SUSPENDED) {
469 err = zap_lookup(vd->vdev_spa->spa_meta_objset,
470 vd->vdev_leaf_zap, VDEV_LEAF_ZAP_INITIALIZE_LAST_OFFSET,
471 sizeof (vd->vdev_initialize_last_offset), 1,
472 &vd->vdev_initialize_last_offset);
473 if (err == ENOENT) {
474 vd->vdev_initialize_last_offset = 0;
475 err = 0;
476 }
477 }
478
479 vdev_initialize_calculate_progress(vd);
480 return (err);
481 }
482
483 static void
vdev_initialize_xlate_range_add(void * arg,zfs_range_seg64_t * physical_rs)484 vdev_initialize_xlate_range_add(void *arg, zfs_range_seg64_t *physical_rs)
485 {
486 vdev_t *vd = arg;
487
488 /* Only add segments that we have not visited yet */
489 if (physical_rs->rs_end <= vd->vdev_initialize_last_offset)
490 return;
491
492 /* Pick up where we left off mid-range. */
493 if (vd->vdev_initialize_last_offset > physical_rs->rs_start) {
494 zfs_dbgmsg("range write: vd %s changed (%llu, %llu) to "
495 "(%llu, %llu)", vd->vdev_path,
496 (u_longlong_t)physical_rs->rs_start,
497 (u_longlong_t)physical_rs->rs_end,
498 (u_longlong_t)vd->vdev_initialize_last_offset,
499 (u_longlong_t)physical_rs->rs_end);
500 ASSERT3U(physical_rs->rs_end, >,
501 vd->vdev_initialize_last_offset);
502 physical_rs->rs_start = vd->vdev_initialize_last_offset;
503 }
504
505 ASSERT3U(physical_rs->rs_end, >, physical_rs->rs_start);
506
507 zfs_range_tree_add(vd->vdev_initialize_tree, physical_rs->rs_start,
508 physical_rs->rs_end - physical_rs->rs_start);
509 }
510
511 /*
512 * Convert the logical range into a physical range and add it to our
513 * avl tree.
514 */
515 static void
vdev_initialize_range_add(void * arg,uint64_t start,uint64_t size)516 vdev_initialize_range_add(void *arg, uint64_t start, uint64_t size)
517 {
518 vdev_t *vd = arg;
519 zfs_range_seg64_t logical_rs;
520 logical_rs.rs_start = start;
521 logical_rs.rs_end = start + size;
522
523 ASSERT(vd->vdev_ops->vdev_op_leaf);
524 vdev_xlate_walk(vd, &logical_rs, vdev_initialize_xlate_range_add, arg);
525 }
526
527 static __attribute__((noreturn)) void
vdev_initialize_thread(void * arg)528 vdev_initialize_thread(void *arg)
529 {
530 vdev_t *vd = arg;
531 spa_t *spa = vd->vdev_spa;
532 int error = 0;
533 uint64_t ms_count = 0;
534
535 ASSERT(vdev_is_concrete(vd));
536 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
537
538 vd->vdev_initialize_last_offset = 0;
539 VERIFY0(vdev_initialize_load(vd));
540
541 abd_t *deadbeef = vdev_initialize_block_alloc();
542
543 vd->vdev_initialize_tree = zfs_range_tree_create(NULL, ZFS_RANGE_SEG64,
544 NULL, 0, 0);
545
546 for (uint64_t i = 0; !vd->vdev_detached &&
547 i < vd->vdev_top->vdev_ms_count; i++) {
548 metaslab_t *msp = vd->vdev_top->vdev_ms[i];
549 boolean_t unload_when_done = B_FALSE;
550
551 /*
552 * If we've expanded the top-level vdev or it's our
553 * first pass, calculate our progress.
554 */
555 if (vd->vdev_top->vdev_ms_count != ms_count) {
556 vdev_initialize_calculate_progress(vd);
557 ms_count = vd->vdev_top->vdev_ms_count;
558 }
559
560 spa_config_exit(spa, SCL_CONFIG, FTAG);
561 metaslab_disable(msp);
562 mutex_enter(&msp->ms_lock);
563 if (!msp->ms_loaded && !msp->ms_loading)
564 unload_when_done = B_TRUE;
565 VERIFY0(metaslab_load(msp));
566
567 zfs_range_tree_walk(msp->ms_allocatable,
568 vdev_initialize_range_add, vd);
569 mutex_exit(&msp->ms_lock);
570
571 error = vdev_initialize_ranges(vd, deadbeef);
572 metaslab_enable(msp, B_TRUE, unload_when_done);
573 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
574
575 zfs_range_tree_vacate(vd->vdev_initialize_tree, NULL, NULL);
576 if (error != 0)
577 break;
578 }
579
580 spa_config_exit(spa, SCL_CONFIG, FTAG);
581 mutex_enter(&vd->vdev_initialize_io_lock);
582 while (vd->vdev_initialize_inflight > 0) {
583 cv_wait(&vd->vdev_initialize_io_cv,
584 &vd->vdev_initialize_io_lock);
585 }
586 mutex_exit(&vd->vdev_initialize_io_lock);
587
588 zfs_range_tree_destroy(vd->vdev_initialize_tree);
589 vdev_initialize_block_free(deadbeef);
590 vd->vdev_initialize_tree = NULL;
591
592 mutex_enter(&vd->vdev_initialize_lock);
593 if (!vd->vdev_initialize_exit_wanted) {
594 if (vdev_writeable(vd)) {
595 vdev_initialize_change_state(vd,
596 VDEV_INITIALIZE_COMPLETE);
597 } else if (vd->vdev_faulted) {
598 vdev_initialize_change_state(vd,
599 VDEV_INITIALIZE_CANCELED);
600 }
601 }
602 ASSERT(vd->vdev_initialize_thread != NULL ||
603 vd->vdev_initialize_inflight == 0);
604
605 /*
606 * Drop the vdev_initialize_lock while we sync out the
607 * txg since it's possible that a device might be trying to
608 * come online and must check to see if it needs to restart an
609 * initialization. That thread will be holding the spa_config_lock
610 * which would prevent the txg_wait_synced from completing.
611 */
612 mutex_exit(&vd->vdev_initialize_lock);
613 txg_wait_synced(spa_get_dsl(spa), 0);
614 mutex_enter(&vd->vdev_initialize_lock);
615
616 vd->vdev_initialize_thread = NULL;
617 cv_broadcast(&vd->vdev_initialize_cv);
618 mutex_exit(&vd->vdev_initialize_lock);
619
620 thread_exit();
621 }
622
623 /*
624 * Initiates a device. Caller must hold vdev_initialize_lock.
625 * Device must be a leaf and not already be initializing.
626 */
627 void
vdev_initialize(vdev_t * vd)628 vdev_initialize(vdev_t *vd)
629 {
630 ASSERT(MUTEX_HELD(&vd->vdev_initialize_lock));
631 ASSERT(vd->vdev_ops->vdev_op_leaf);
632 ASSERT(vdev_is_concrete(vd));
633 ASSERT3P(vd->vdev_initialize_thread, ==, NULL);
634 ASSERT(!vd->vdev_detached);
635 ASSERT(!vd->vdev_initialize_exit_wanted);
636 ASSERT(!vd->vdev_top->vdev_removing);
637 ASSERT(!vd->vdev_top->vdev_rz_expanding);
638
639 vdev_initialize_change_state(vd, VDEV_INITIALIZE_ACTIVE);
640 vd->vdev_initialize_thread = thread_create(NULL, 0,
641 vdev_initialize_thread, vd, 0, &p0, TS_RUN, maxclsyspri);
642 }
643
644 /*
645 * Uninitializes a device. Caller must hold vdev_initialize_lock.
646 * Device must be a leaf and not already be initializing.
647 */
648 void
vdev_uninitialize(vdev_t * vd)649 vdev_uninitialize(vdev_t *vd)
650 {
651 ASSERT(MUTEX_HELD(&vd->vdev_initialize_lock));
652 ASSERT(vd->vdev_ops->vdev_op_leaf);
653 ASSERT(vdev_is_concrete(vd));
654 ASSERT3P(vd->vdev_initialize_thread, ==, NULL);
655 ASSERT(!vd->vdev_detached);
656 ASSERT(!vd->vdev_initialize_exit_wanted);
657 ASSERT(!vd->vdev_top->vdev_removing);
658
659 vdev_initialize_change_state(vd, VDEV_INITIALIZE_NONE);
660 }
661
662 /*
663 * Wait for the initialize thread to be terminated (cancelled or stopped).
664 */
665 static void
vdev_initialize_stop_wait_impl(vdev_t * vd)666 vdev_initialize_stop_wait_impl(vdev_t *vd)
667 {
668 ASSERT(MUTEX_HELD(&vd->vdev_initialize_lock));
669
670 while (vd->vdev_initialize_thread != NULL)
671 cv_wait(&vd->vdev_initialize_cv, &vd->vdev_initialize_lock);
672
673 ASSERT3P(vd->vdev_initialize_thread, ==, NULL);
674 vd->vdev_initialize_exit_wanted = B_FALSE;
675 }
676
677 /*
678 * Wait for vdev initialize threads which were either to cleanly exit.
679 */
680 void
vdev_initialize_stop_wait(spa_t * spa,list_t * vd_list)681 vdev_initialize_stop_wait(spa_t *spa, list_t *vd_list)
682 {
683 (void) spa;
684 vdev_t *vd;
685
686 ASSERT(MUTEX_HELD(&spa_namespace_lock) ||
687 spa->spa_export_thread == curthread);
688
689 while ((vd = list_remove_head(vd_list)) != NULL) {
690 mutex_enter(&vd->vdev_initialize_lock);
691 vdev_initialize_stop_wait_impl(vd);
692 mutex_exit(&vd->vdev_initialize_lock);
693 }
694 }
695
696 /*
697 * Stop initializing a device, with the resultant initializing state being
698 * tgt_state. For blocking behavior pass NULL for vd_list. Otherwise, when
699 * a list_t is provided the stopping vdev is inserted in to the list. Callers
700 * are then required to call vdev_initialize_stop_wait() to block for all the
701 * initialization threads to exit. The caller must hold vdev_initialize_lock
702 * and must not be writing to the spa config, as the initializing thread may
703 * try to enter the config as a reader before exiting.
704 */
705 void
vdev_initialize_stop(vdev_t * vd,vdev_initializing_state_t tgt_state,list_t * vd_list)706 vdev_initialize_stop(vdev_t *vd, vdev_initializing_state_t tgt_state,
707 list_t *vd_list)
708 {
709 ASSERT(!spa_config_held(vd->vdev_spa, SCL_CONFIG|SCL_STATE, RW_WRITER));
710 ASSERT(MUTEX_HELD(&vd->vdev_initialize_lock));
711 ASSERT(vd->vdev_ops->vdev_op_leaf);
712 ASSERT(vdev_is_concrete(vd));
713
714 /*
715 * Allow cancel requests to proceed even if the initialize thread
716 * has stopped.
717 */
718 if (vd->vdev_initialize_thread == NULL &&
719 tgt_state != VDEV_INITIALIZE_CANCELED) {
720 return;
721 }
722
723 vdev_initialize_change_state(vd, tgt_state);
724 vd->vdev_initialize_exit_wanted = B_TRUE;
725
726 if (vd_list == NULL) {
727 vdev_initialize_stop_wait_impl(vd);
728 } else {
729 ASSERT(MUTEX_HELD(&spa_namespace_lock) ||
730 vd->vdev_spa->spa_export_thread == curthread);
731 list_insert_tail(vd_list, vd);
732 }
733 }
734
735 static void
vdev_initialize_stop_all_impl(vdev_t * vd,vdev_initializing_state_t tgt_state,list_t * vd_list)736 vdev_initialize_stop_all_impl(vdev_t *vd, vdev_initializing_state_t tgt_state,
737 list_t *vd_list)
738 {
739 if (vd->vdev_ops->vdev_op_leaf && vdev_is_concrete(vd)) {
740 mutex_enter(&vd->vdev_initialize_lock);
741 vdev_initialize_stop(vd, tgt_state, vd_list);
742 mutex_exit(&vd->vdev_initialize_lock);
743 return;
744 }
745
746 for (uint64_t i = 0; i < vd->vdev_children; i++) {
747 vdev_initialize_stop_all_impl(vd->vdev_child[i], tgt_state,
748 vd_list);
749 }
750 }
751
752 /*
753 * Convenience function to stop initializing of a vdev tree and set all
754 * initialize thread pointers to NULL.
755 */
756 void
vdev_initialize_stop_all(vdev_t * vd,vdev_initializing_state_t tgt_state)757 vdev_initialize_stop_all(vdev_t *vd, vdev_initializing_state_t tgt_state)
758 {
759 spa_t *spa = vd->vdev_spa;
760 list_t vd_list;
761
762 ASSERT(MUTEX_HELD(&spa_namespace_lock) ||
763 spa->spa_export_thread == curthread);
764
765 list_create(&vd_list, sizeof (vdev_t),
766 offsetof(vdev_t, vdev_initialize_node));
767
768 vdev_initialize_stop_all_impl(vd, tgt_state, &vd_list);
769 vdev_initialize_stop_wait(spa, &vd_list);
770
771 if (vd->vdev_spa->spa_sync_on) {
772 /* Make sure that our state has been synced to disk */
773 txg_wait_synced(spa_get_dsl(vd->vdev_spa), 0);
774 }
775
776 list_destroy(&vd_list);
777 }
778
779 void
vdev_initialize_restart(vdev_t * vd)780 vdev_initialize_restart(vdev_t *vd)
781 {
782 ASSERT(MUTEX_HELD(&spa_namespace_lock) ||
783 vd->vdev_spa->spa_load_thread == curthread);
784 ASSERT(!spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER));
785
786 if (vd->vdev_leaf_zap != 0) {
787 mutex_enter(&vd->vdev_initialize_lock);
788 uint64_t initialize_state = VDEV_INITIALIZE_NONE;
789 int err = zap_lookup(vd->vdev_spa->spa_meta_objset,
790 vd->vdev_leaf_zap, VDEV_LEAF_ZAP_INITIALIZE_STATE,
791 sizeof (initialize_state), 1, &initialize_state);
792 ASSERT(err == 0 || err == ENOENT);
793 vd->vdev_initialize_state = initialize_state;
794
795 uint64_t timestamp = 0;
796 err = zap_lookup(vd->vdev_spa->spa_meta_objset,
797 vd->vdev_leaf_zap, VDEV_LEAF_ZAP_INITIALIZE_ACTION_TIME,
798 sizeof (timestamp), 1, ×tamp);
799 ASSERT(err == 0 || err == ENOENT);
800 vd->vdev_initialize_action_time = timestamp;
801
802 if ((vd->vdev_initialize_state == VDEV_INITIALIZE_SUSPENDED ||
803 vd->vdev_offline) && !vd->vdev_top->vdev_rz_expanding) {
804 /* load progress for reporting, but don't resume */
805 VERIFY0(vdev_initialize_load(vd));
806 } else if (vd->vdev_initialize_state ==
807 VDEV_INITIALIZE_ACTIVE && vdev_writeable(vd) &&
808 !vd->vdev_top->vdev_removing &&
809 !vd->vdev_top->vdev_rz_expanding &&
810 vd->vdev_initialize_thread == NULL) {
811 vdev_initialize(vd);
812 }
813
814 mutex_exit(&vd->vdev_initialize_lock);
815 }
816
817 for (uint64_t i = 0; i < vd->vdev_children; i++) {
818 vdev_initialize_restart(vd->vdev_child[i]);
819 }
820 }
821
822 EXPORT_SYMBOL(vdev_initialize);
823 EXPORT_SYMBOL(vdev_uninitialize);
824 EXPORT_SYMBOL(vdev_initialize_stop);
825 EXPORT_SYMBOL(vdev_initialize_stop_all);
826 EXPORT_SYMBOL(vdev_initialize_stop_wait);
827 EXPORT_SYMBOL(vdev_initialize_restart);
828
829 ZFS_MODULE_PARAM(zfs, zfs_, initialize_value, U64, ZMOD_RW,
830 "Value written during zpool initialize");
831
832 ZFS_MODULE_PARAM(zfs, zfs_, initialize_chunk_size, U64, ZMOD_RW,
833 "Size in bytes of writes by zpool initialize");
834