xref: /illumos-gate/usr/src/uts/common/fs/zfs/vdev_initialize.c (revision 4d7988d6050abba5c1ff60e7fd196e95c22e20f4)
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, 2019 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/refcount.h>
31 #include <sys/metaslab_impl.h>
32 #include <sys/dsl_synctask.h>
33 #include <sys/zap.h>
34 #include <sys/dmu_tx.h>
35 
36 /*
37  * Value that is written to disk during initialization.
38  */
39 uint64_t zfs_initialize_value = 0xdeadbeefdeadbeefULL;
40 
41 /* maximum number of I/Os outstanding per leaf vdev */
42 int zfs_initialize_limit = 1;
43 
44 /* size of initializing writes; default 1MiB, see zfs_remove_max_segment */
45 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 }
53 
54 static void
vdev_initialize_zap_update_sync(void * arg,dmu_tx_t * tx)55 vdev_initialize_zap_update_sync(void *arg, dmu_tx_t *tx)
56 {
57 	/*
58 	 * We pass in the guid instead of the vdev_t since the vdev may
59 	 * have been freed prior to the sync task being processed. This
60 	 * happens when a vdev is detached as we call spa_config_vdev_exit(),
61 	 * stop the initializing thread, schedule the sync task, and free
62 	 * the vdev. Later when the scheduled sync task is invoked, it would
63 	 * find that the vdev has been freed.
64 	 */
65 	uint64_t guid = *(uint64_t *)arg;
66 	uint64_t txg = dmu_tx_get_txg(tx);
67 	kmem_free(arg, sizeof (uint64_t));
68 
69 	vdev_t *vd = spa_lookup_by_guid(tx->tx_pool->dp_spa, guid, B_FALSE);
70 	if (vd == NULL || vd->vdev_top->vdev_removing || !vdev_is_concrete(vd))
71 		return;
72 
73 	uint64_t last_offset = vd->vdev_initialize_offset[txg & TXG_MASK];
74 	vd->vdev_initialize_offset[txg & TXG_MASK] = 0;
75 
76 	VERIFY(vd->vdev_leaf_zap != 0);
77 
78 	objset_t *mos = vd->vdev_spa->spa_meta_objset;
79 
80 	if (last_offset > 0) {
81 		vd->vdev_initialize_last_offset = last_offset;
82 		VERIFY0(zap_update(mos, vd->vdev_leaf_zap,
83 		    VDEV_LEAF_ZAP_INITIALIZE_LAST_OFFSET,
84 		    sizeof (last_offset), 1, &last_offset, tx));
85 	}
86 	if (vd->vdev_initialize_action_time > 0) {
87 		uint64_t val = (uint64_t)vd->vdev_initialize_action_time;
88 		VERIFY0(zap_update(mos, vd->vdev_leaf_zap,
89 		    VDEV_LEAF_ZAP_INITIALIZE_ACTION_TIME, sizeof (val),
90 		    1, &val, tx));
91 	}
92 
93 	uint64_t initialize_state = vd->vdev_initialize_state;
94 	VERIFY0(zap_update(mos, vd->vdev_leaf_zap,
95 	    VDEV_LEAF_ZAP_INITIALIZE_STATE, sizeof (initialize_state), 1,
96 	    &initialize_state, tx));
97 }
98 
99 static void
vdev_initialize_change_state(vdev_t * vd,vdev_initializing_state_t new_state)100 vdev_initialize_change_state(vdev_t *vd, vdev_initializing_state_t new_state)
101 {
102 	ASSERT(MUTEX_HELD(&vd->vdev_initialize_lock));
103 	spa_t *spa = vd->vdev_spa;
104 
105 	if (new_state == vd->vdev_initialize_state)
106 		return;
107 
108 	/*
109 	 * Copy the vd's guid, this will be freed by the sync task.
110 	 */
111 	uint64_t *guid = kmem_zalloc(sizeof (uint64_t), KM_SLEEP);
112 	*guid = vd->vdev_guid;
113 
114 	/*
115 	 * If we're suspending, then preserving the original start time.
116 	 */
117 	if (vd->vdev_initialize_state != VDEV_INITIALIZE_SUSPENDED) {
118 		vd->vdev_initialize_action_time = gethrestime_sec();
119 	}
120 	vd->vdev_initialize_state = new_state;
121 
122 	dmu_tx_t *tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir);
123 	VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
124 	dsl_sync_task_nowait(spa_get_dsl(spa), vdev_initialize_zap_update_sync,
125 	    guid, 2, ZFS_SPACE_CHECK_NONE, tx);
126 
127 	switch (new_state) {
128 	case VDEV_INITIALIZE_ACTIVE:
129 		spa_history_log_internal(spa, "initialize", tx,
130 		    "vdev=%s activated", vd->vdev_path);
131 		break;
132 	case VDEV_INITIALIZE_SUSPENDED:
133 		spa_history_log_internal(spa, "initialize", tx,
134 		    "vdev=%s suspended", vd->vdev_path);
135 		break;
136 	case VDEV_INITIALIZE_CANCELED:
137 		spa_history_log_internal(spa, "initialize", tx,
138 		    "vdev=%s canceled", vd->vdev_path);
139 		break;
140 	case VDEV_INITIALIZE_COMPLETE:
141 		spa_history_log_internal(spa, "initialize", tx,
142 		    "vdev=%s complete", vd->vdev_path);
143 		break;
144 	default:
145 		panic("invalid state %llu", (unsigned long long)new_state);
146 	}
147 
148 	dmu_tx_commit(tx);
149 }
150 
151 static void
vdev_initialize_cb(zio_t * zio)152 vdev_initialize_cb(zio_t *zio)
153 {
154 	vdev_t *vd = zio->io_vd;
155 	mutex_enter(&vd->vdev_initialize_io_lock);
156 	if (zio->io_error == ENXIO && !vdev_writeable(vd)) {
157 		/*
158 		 * The I/O failed because the vdev was unavailable; roll the
159 		 * last offset back. (This works because spa_sync waits on
160 		 * spa_txg_zio before it runs sync tasks.)
161 		 */
162 		uint64_t *off =
163 		    &vd->vdev_initialize_offset[zio->io_txg & TXG_MASK];
164 		*off = MIN(*off, zio->io_offset);
165 	} else {
166 		/*
167 		 * Since initializing is best-effort, we ignore I/O errors and
168 		 * rely on vdev_probe to determine if the errors are more
169 		 * critical.
170 		 */
171 		if (zio->io_error != 0)
172 			vd->vdev_stat.vs_initialize_errors++;
173 
174 		vd->vdev_initialize_bytes_done += zio->io_orig_size;
175 	}
176 	ASSERT3U(vd->vdev_initialize_inflight, >, 0);
177 	vd->vdev_initialize_inflight--;
178 	cv_broadcast(&vd->vdev_initialize_io_cv);
179 	mutex_exit(&vd->vdev_initialize_io_lock);
180 
181 	spa_config_exit(vd->vdev_spa, SCL_STATE_ALL, vd);
182 }
183 
184 /* Takes care of physical writing and limiting # of concurrent ZIOs. */
185 static int
vdev_initialize_write(vdev_t * vd,uint64_t start,uint64_t size,abd_t * data)186 vdev_initialize_write(vdev_t *vd, uint64_t start, uint64_t size, abd_t *data)
187 {
188 	spa_t *spa = vd->vdev_spa;
189 
190 	/* Limit inflight initializing I/Os */
191 	mutex_enter(&vd->vdev_initialize_io_lock);
192 	while (vd->vdev_initialize_inflight >= zfs_initialize_limit) {
193 		cv_wait(&vd->vdev_initialize_io_cv,
194 		    &vd->vdev_initialize_io_lock);
195 	}
196 	vd->vdev_initialize_inflight++;
197 	mutex_exit(&vd->vdev_initialize_io_lock);
198 
199 	dmu_tx_t *tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir);
200 	VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
201 	uint64_t txg = dmu_tx_get_txg(tx);
202 
203 	spa_config_enter(spa, SCL_STATE_ALL, vd, RW_READER);
204 	mutex_enter(&vd->vdev_initialize_lock);
205 
206 	if (vd->vdev_initialize_offset[txg & TXG_MASK] == 0) {
207 		uint64_t *guid = kmem_zalloc(sizeof (uint64_t), KM_SLEEP);
208 		*guid = vd->vdev_guid;
209 
210 		/* This is the first write of this txg. */
211 		dsl_sync_task_nowait(spa_get_dsl(spa),
212 		    vdev_initialize_zap_update_sync, guid, 2,
213 		    ZFS_SPACE_CHECK_RESERVED, tx);
214 	}
215 
216 	/*
217 	 * We know the vdev struct will still be around since all
218 	 * consumers of vdev_free must stop the initialization first.
219 	 */
220 	if (vdev_initialize_should_stop(vd)) {
221 		mutex_enter(&vd->vdev_initialize_io_lock);
222 		ASSERT3U(vd->vdev_initialize_inflight, >, 0);
223 		vd->vdev_initialize_inflight--;
224 		mutex_exit(&vd->vdev_initialize_io_lock);
225 		spa_config_exit(vd->vdev_spa, SCL_STATE_ALL, vd);
226 		mutex_exit(&vd->vdev_initialize_lock);
227 		dmu_tx_commit(tx);
228 		return (SET_ERROR(EINTR));
229 	}
230 	mutex_exit(&vd->vdev_initialize_lock);
231 
232 	vd->vdev_initialize_offset[txg & TXG_MASK] = start + size;
233 	zio_nowait(zio_write_phys(spa->spa_txg_zio[txg & TXG_MASK], vd, start,
234 	    size, data, ZIO_CHECKSUM_OFF, vdev_initialize_cb, NULL,
235 	    ZIO_PRIORITY_INITIALIZING, ZIO_FLAG_CANFAIL, B_FALSE));
236 	/* vdev_initialize_cb releases SCL_STATE_ALL */
237 
238 	dmu_tx_commit(tx);
239 
240 	return (0);
241 }
242 
243 /*
244  * Callback to fill each ABD chunk with zfs_initialize_value. len must be
245  * divisible by sizeof (uint64_t), and buf must be 8-byte aligned. The ABD
246  * allocation will guarantee these for us.
247  */
248 /* ARGSUSED */
249 static int
vdev_initialize_block_fill(void * buf,size_t len,void * unused)250 vdev_initialize_block_fill(void *buf, size_t len, void *unused)
251 {
252 	ASSERT0(len % sizeof (uint64_t));
253 	for (uint64_t i = 0; i < len; i += sizeof (uint64_t)) {
254 		*(uint64_t *)((char *)(buf) + i) = zfs_initialize_value;
255 	}
256 	return (0);
257 }
258 
259 static abd_t *
vdev_initialize_block_alloc()260 vdev_initialize_block_alloc()
261 {
262 	/* Allocate ABD for filler data */
263 	abd_t *data = abd_alloc_for_io(zfs_initialize_chunk_size, B_FALSE);
264 
265 	ASSERT0(zfs_initialize_chunk_size % sizeof (uint64_t));
266 	(void) abd_iterate_func(data, 0, zfs_initialize_chunk_size,
267 	    vdev_initialize_block_fill, NULL);
268 
269 	return (data);
270 }
271 
272 static void
vdev_initialize_block_free(abd_t * data)273 vdev_initialize_block_free(abd_t *data)
274 {
275 	abd_free(data);
276 }
277 
278 static int
vdev_initialize_ranges(vdev_t * vd,abd_t * data)279 vdev_initialize_ranges(vdev_t *vd, abd_t *data)
280 {
281 	range_tree_t *rt = vd->vdev_initialize_tree;
282 	zfs_btree_t *bt = &rt->rt_root;
283 	zfs_btree_index_t where;
284 
285 	for (range_seg_t *rs = zfs_btree_first(bt, &where); rs != NULL;
286 	    rs = zfs_btree_next(bt, &where, &where)) {
287 		uint64_t size = rs_get_end(rs, rt) - rs_get_start(rs, rt);
288 
289 		/* Split range into legally-sized physical chunks */
290 		uint64_t writes_required =
291 		    ((size - 1) / zfs_initialize_chunk_size) + 1;
292 
293 		for (uint64_t w = 0; w < writes_required; w++) {
294 			int error;
295 
296 			error = vdev_initialize_write(vd,
297 			    VDEV_LABEL_START_SIZE + rs_get_start(rs, rt) +
298 			    (w * zfs_initialize_chunk_size),
299 			    MIN(size - (w * zfs_initialize_chunk_size),
300 			    zfs_initialize_chunk_size), data);
301 			if (error != 0)
302 				return (error);
303 		}
304 	}
305 	return (0);
306 }
307 
308 static void
vdev_initialize_calculate_progress(vdev_t * vd)309 vdev_initialize_calculate_progress(vdev_t *vd)
310 {
311 	ASSERT(spa_config_held(vd->vdev_spa, SCL_CONFIG, RW_READER) ||
312 	    spa_config_held(vd->vdev_spa, SCL_CONFIG, RW_WRITER));
313 	ASSERT(vd->vdev_leaf_zap != 0);
314 
315 	vd->vdev_initialize_bytes_est = 0;
316 	vd->vdev_initialize_bytes_done = 0;
317 
318 	for (uint64_t i = 0; i < vd->vdev_top->vdev_ms_count; i++) {
319 		metaslab_t *msp = vd->vdev_top->vdev_ms[i];
320 		mutex_enter(&msp->ms_lock);
321 
322 		uint64_t ms_free = msp->ms_size -
323 		    metaslab_allocated_space(msp);
324 
325 		if (vd->vdev_top->vdev_ops == &vdev_raidz_ops)
326 			ms_free /= vd->vdev_top->vdev_children;
327 
328 		/*
329 		 * Convert the metaslab range to a physical range
330 		 * on our vdev. We use this to determine if we are
331 		 * in the middle of this metaslab range.
332 		 */
333 		range_seg64_t logical_rs, physical_rs;
334 		logical_rs.rs_start = msp->ms_start;
335 		logical_rs.rs_end = msp->ms_start + msp->ms_size;
336 		vdev_xlate(vd, &logical_rs, &physical_rs);
337 
338 		if (vd->vdev_initialize_last_offset <= physical_rs.rs_start) {
339 			vd->vdev_initialize_bytes_est += ms_free;
340 			mutex_exit(&msp->ms_lock);
341 			continue;
342 		} else if (vd->vdev_initialize_last_offset >
343 		    physical_rs.rs_end) {
344 			vd->vdev_initialize_bytes_done += ms_free;
345 			vd->vdev_initialize_bytes_est += ms_free;
346 			mutex_exit(&msp->ms_lock);
347 			continue;
348 		}
349 
350 		/*
351 		 * If we get here, we're in the middle of initializing this
352 		 * metaslab. Load it and walk the free tree for more accurate
353 		 * progress estimation.
354 		 */
355 		VERIFY0(metaslab_load(msp));
356 
357 		zfs_btree_index_t where;
358 		range_tree_t *rt = msp->ms_allocatable;
359 		for (range_seg_t *rs =
360 		    zfs_btree_first(&rt->rt_root, &where); rs;
361 		    rs = zfs_btree_next(&rt->rt_root, &where,
362 		    &where)) {
363 			logical_rs.rs_start = rs_get_start(rs, rt);
364 			logical_rs.rs_end = rs_get_end(rs, rt);
365 			vdev_xlate(vd, &logical_rs, &physical_rs);
366 
367 			uint64_t size = physical_rs.rs_end -
368 			    physical_rs.rs_start;
369 			vd->vdev_initialize_bytes_est += size;
370 			if (vd->vdev_initialize_last_offset >
371 			    physical_rs.rs_end) {
372 				vd->vdev_initialize_bytes_done += size;
373 			} else if (vd->vdev_initialize_last_offset >
374 			    physical_rs.rs_start &&
375 			    vd->vdev_initialize_last_offset <
376 			    physical_rs.rs_end) {
377 				vd->vdev_initialize_bytes_done +=
378 				    vd->vdev_initialize_last_offset -
379 				    physical_rs.rs_start;
380 			}
381 		}
382 		mutex_exit(&msp->ms_lock);
383 	}
384 }
385 
386 static int
vdev_initialize_load(vdev_t * vd)387 vdev_initialize_load(vdev_t *vd)
388 {
389 	int err = 0;
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 	if (vd->vdev_initialize_state == VDEV_INITIALIZE_ACTIVE ||
395 	    vd->vdev_initialize_state == VDEV_INITIALIZE_SUSPENDED) {
396 		err = zap_lookup(vd->vdev_spa->spa_meta_objset,
397 		    vd->vdev_leaf_zap, VDEV_LEAF_ZAP_INITIALIZE_LAST_OFFSET,
398 		    sizeof (vd->vdev_initialize_last_offset), 1,
399 		    &vd->vdev_initialize_last_offset);
400 		if (err == ENOENT) {
401 			vd->vdev_initialize_last_offset = 0;
402 			err = 0;
403 		}
404 	}
405 
406 	vdev_initialize_calculate_progress(vd);
407 	return (err);
408 }
409 
410 
411 /*
412  * Convert the logical range into a physical range and add it to our
413  * avl tree.
414  */
415 void
vdev_initialize_range_add(void * arg,uint64_t start,uint64_t size)416 vdev_initialize_range_add(void *arg, uint64_t start, uint64_t size)
417 {
418 	vdev_t *vd = arg;
419 	range_seg64_t logical_rs, physical_rs;
420 	logical_rs.rs_start = start;
421 	logical_rs.rs_end = start + size;
422 
423 	ASSERT(vd->vdev_ops->vdev_op_leaf);
424 	vdev_xlate(vd, &logical_rs, &physical_rs);
425 
426 	IMPLY(vd->vdev_top == vd,
427 	    logical_rs.rs_start == physical_rs.rs_start);
428 	IMPLY(vd->vdev_top == vd,
429 	    logical_rs.rs_end == physical_rs.rs_end);
430 
431 	/* Only add segments that we have not visited yet */
432 	if (physical_rs.rs_end <= vd->vdev_initialize_last_offset)
433 		return;
434 
435 	/* Pick up where we left off mid-range. */
436 	if (vd->vdev_initialize_last_offset > physical_rs.rs_start) {
437 		zfs_dbgmsg("range write: vd %s changed (%llu, %llu) to "
438 		    "(%llu, %llu)", vd->vdev_path,
439 		    (u_longlong_t)physical_rs.rs_start,
440 		    (u_longlong_t)physical_rs.rs_end,
441 		    (u_longlong_t)vd->vdev_initialize_last_offset,
442 		    (u_longlong_t)physical_rs.rs_end);
443 		ASSERT3U(physical_rs.rs_end, >,
444 		    vd->vdev_initialize_last_offset);
445 		physical_rs.rs_start = vd->vdev_initialize_last_offset;
446 	}
447 	ASSERT3U(physical_rs.rs_end, >=, physical_rs.rs_start);
448 
449 	/*
450 	 * With raidz, it's possible that the logical range does not live on
451 	 * this leaf vdev. We only add the physical range to this vdev's if it
452 	 * has a length greater than 0.
453 	 */
454 	if (physical_rs.rs_end > physical_rs.rs_start) {
455 		range_tree_add(vd->vdev_initialize_tree, physical_rs.rs_start,
456 		    physical_rs.rs_end - physical_rs.rs_start);
457 	} else {
458 		ASSERT3U(physical_rs.rs_end, ==, physical_rs.rs_start);
459 	}
460 }
461 
462 static void
vdev_initialize_thread(void * arg)463 vdev_initialize_thread(void *arg)
464 {
465 	vdev_t *vd = arg;
466 	spa_t *spa = vd->vdev_spa;
467 	int error = 0;
468 	uint64_t ms_count = 0;
469 
470 	ASSERT(vdev_is_concrete(vd));
471 	spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
472 
473 	vd->vdev_initialize_last_offset = 0;
474 	VERIFY0(vdev_initialize_load(vd));
475 
476 	abd_t *deadbeef = vdev_initialize_block_alloc();
477 
478 	vd->vdev_initialize_tree = range_tree_create(NULL, RANGE_SEG64, NULL,
479 	    0, 0);
480 
481 	for (uint64_t i = 0; !vd->vdev_detached &&
482 	    i < vd->vdev_top->vdev_ms_count; i++) {
483 		metaslab_t *msp = vd->vdev_top->vdev_ms[i];
484 		boolean_t unload_when_done = B_FALSE;
485 
486 		/*
487 		 * If we've expanded the top-level vdev or it's our
488 		 * first pass, calculate our progress.
489 		 */
490 		if (vd->vdev_top->vdev_ms_count != ms_count) {
491 			vdev_initialize_calculate_progress(vd);
492 			ms_count = vd->vdev_top->vdev_ms_count;
493 		}
494 
495 		spa_config_exit(spa, SCL_CONFIG, FTAG);
496 		metaslab_disable(msp);
497 		mutex_enter(&msp->ms_lock);
498 		if (!msp->ms_loaded && !msp->ms_loading)
499 			unload_when_done = B_TRUE;
500 		VERIFY0(metaslab_load(msp));
501 
502 		range_tree_walk(msp->ms_allocatable, vdev_initialize_range_add,
503 		    vd);
504 		mutex_exit(&msp->ms_lock);
505 
506 		error = vdev_initialize_ranges(vd, deadbeef);
507 		metaslab_enable(msp, B_TRUE, unload_when_done);
508 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
509 
510 		range_tree_vacate(vd->vdev_initialize_tree, NULL, NULL);
511 		if (error != 0)
512 			break;
513 	}
514 
515 	spa_config_exit(spa, SCL_CONFIG, FTAG);
516 	mutex_enter(&vd->vdev_initialize_io_lock);
517 	while (vd->vdev_initialize_inflight > 0) {
518 		cv_wait(&vd->vdev_initialize_io_cv,
519 		    &vd->vdev_initialize_io_lock);
520 	}
521 	mutex_exit(&vd->vdev_initialize_io_lock);
522 
523 	range_tree_destroy(vd->vdev_initialize_tree);
524 	vdev_initialize_block_free(deadbeef);
525 	vd->vdev_initialize_tree = NULL;
526 
527 	mutex_enter(&vd->vdev_initialize_lock);
528 	if (!vd->vdev_initialize_exit_wanted && vdev_writeable(vd)) {
529 		vdev_initialize_change_state(vd, VDEV_INITIALIZE_COMPLETE);
530 	}
531 	ASSERT(vd->vdev_initialize_thread != NULL ||
532 	    vd->vdev_initialize_inflight == 0);
533 
534 	/*
535 	 * Drop the vdev_initialize_lock while we sync out the
536 	 * txg since it's possible that a device might be trying to
537 	 * come online and must check to see if it needs to restart an
538 	 * initialization. That thread will be holding the spa_config_lock
539 	 * which would prevent the txg_wait_synced from completing.
540 	 */
541 	mutex_exit(&vd->vdev_initialize_lock);
542 	txg_wait_synced(spa_get_dsl(spa), 0);
543 	mutex_enter(&vd->vdev_initialize_lock);
544 
545 	vd->vdev_initialize_thread = NULL;
546 	cv_broadcast(&vd->vdev_initialize_cv);
547 	mutex_exit(&vd->vdev_initialize_lock);
548 }
549 
550 /*
551  * Initiates a device. Caller must hold vdev_initialize_lock.
552  * Device must be a leaf and not already be initializing.
553  */
554 void
vdev_initialize(vdev_t * vd)555 vdev_initialize(vdev_t *vd)
556 {
557 	ASSERT(MUTEX_HELD(&vd->vdev_initialize_lock));
558 	ASSERT(vd->vdev_ops->vdev_op_leaf);
559 	ASSERT(vdev_is_concrete(vd));
560 	ASSERT3P(vd->vdev_initialize_thread, ==, NULL);
561 	ASSERT(!vd->vdev_detached);
562 	ASSERT(!vd->vdev_initialize_exit_wanted);
563 	ASSERT(!vd->vdev_top->vdev_removing);
564 
565 	vdev_initialize_change_state(vd, VDEV_INITIALIZE_ACTIVE);
566 	vd->vdev_initialize_thread = thread_create(NULL, 0,
567 	    vdev_initialize_thread, vd, 0, &p0, TS_RUN, maxclsyspri);
568 }
569 
570 /*
571  * Wait for the initialize thread to be terminated (cancelled or stopped).
572  */
573 static void
vdev_initialize_stop_wait_impl(vdev_t * vd)574 vdev_initialize_stop_wait_impl(vdev_t *vd)
575 {
576 	ASSERT(MUTEX_HELD(&vd->vdev_initialize_lock));
577 
578 	while (vd->vdev_initialize_thread != NULL)
579 		cv_wait(&vd->vdev_initialize_cv, &vd->vdev_initialize_lock);
580 
581 	ASSERT3P(vd->vdev_initialize_thread, ==, NULL);
582 	vd->vdev_initialize_exit_wanted = B_FALSE;
583 }
584 
585 /*
586  * Wait for vdev initialize threads which were either to cleanly exit.
587  */
588 void
vdev_initialize_stop_wait(spa_t * spa,list_t * vd_list)589 vdev_initialize_stop_wait(spa_t *spa, list_t *vd_list)
590 {
591 	vdev_t *vd;
592 
593 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
594 
595 	while ((vd = list_remove_head(vd_list)) != NULL) {
596 		mutex_enter(&vd->vdev_initialize_lock);
597 		vdev_initialize_stop_wait_impl(vd);
598 		mutex_exit(&vd->vdev_initialize_lock);
599 	}
600 }
601 
602 /*
603  * Stop initializing a device, with the resultant initializing state being
604  * tgt_state.  For blocking behavior pass NULL for vd_list.  Otherwise, when
605  * a list_t is provided the stopping vdev is inserted in to the list.  Callers
606  * are then required to call vdev_initialize_stop_wait() to block for all the
607  * initialization threads to exit.  The caller must hold vdev_initialize_lock
608  * and must not be writing to the spa config, as the initializing thread may
609  * try to enter the config as a reader before exiting.
610  */
611 void
vdev_initialize_stop(vdev_t * vd,vdev_initializing_state_t tgt_state,list_t * vd_list)612 vdev_initialize_stop(vdev_t *vd, vdev_initializing_state_t tgt_state,
613     list_t *vd_list)
614 {
615 	ASSERT(!spa_config_held(vd->vdev_spa, SCL_CONFIG|SCL_STATE, RW_WRITER));
616 	ASSERT(MUTEX_HELD(&vd->vdev_initialize_lock));
617 	ASSERT(vd->vdev_ops->vdev_op_leaf);
618 	ASSERT(vdev_is_concrete(vd));
619 
620 	/*
621 	 * Allow cancel requests to proceed even if the initialize thread
622 	 * has stopped.
623 	 */
624 	if (vd->vdev_initialize_thread == NULL &&
625 	    tgt_state != VDEV_INITIALIZE_CANCELED) {
626 		return;
627 	}
628 
629 	vdev_initialize_change_state(vd, tgt_state);
630 	vd->vdev_initialize_exit_wanted = B_TRUE;
631 
632 	if (vd_list == NULL) {
633 		vdev_initialize_stop_wait_impl(vd);
634 	} else {
635 		ASSERT(MUTEX_HELD(&spa_namespace_lock));
636 		list_insert_tail(vd_list, vd);
637 	}
638 }
639 
640 static void
vdev_initialize_stop_all_impl(vdev_t * vd,vdev_initializing_state_t tgt_state,list_t * vd_list)641 vdev_initialize_stop_all_impl(vdev_t *vd, vdev_initializing_state_t tgt_state,
642     list_t *vd_list)
643 {
644 	if (vd->vdev_ops->vdev_op_leaf && vdev_is_concrete(vd)) {
645 		mutex_enter(&vd->vdev_initialize_lock);
646 		vdev_initialize_stop(vd, tgt_state, vd_list);
647 		mutex_exit(&vd->vdev_initialize_lock);
648 		return;
649 	}
650 
651 	for (uint64_t i = 0; i < vd->vdev_children; i++) {
652 		vdev_initialize_stop_all_impl(vd->vdev_child[i], tgt_state,
653 		    vd_list);
654 	}
655 }
656 
657 /*
658  * Convenience function to stop initializing of a vdev tree and set all
659  * initialize thread pointers to NULL.
660  */
661 void
vdev_initialize_stop_all(vdev_t * vd,vdev_initializing_state_t tgt_state)662 vdev_initialize_stop_all(vdev_t *vd, vdev_initializing_state_t tgt_state)
663 {
664 	spa_t *spa = vd->vdev_spa;
665 	list_t vd_list;
666 
667 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
668 
669 	list_create(&vd_list, sizeof (vdev_t),
670 	    offsetof(vdev_t, vdev_initialize_node));
671 
672 	vdev_initialize_stop_all_impl(vd, tgt_state, &vd_list);
673 	vdev_initialize_stop_wait(spa, &vd_list);
674 
675 	if (vd->vdev_spa->spa_sync_on) {
676 		/* Make sure that our state has been synced to disk */
677 		txg_wait_synced(spa_get_dsl(vd->vdev_spa), 0);
678 	}
679 
680 	list_destroy(&vd_list);
681 }
682 
683 void
vdev_initialize_restart(vdev_t * vd)684 vdev_initialize_restart(vdev_t *vd)
685 {
686 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
687 	ASSERT(!spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER));
688 
689 	if (vd->vdev_leaf_zap != 0) {
690 		mutex_enter(&vd->vdev_initialize_lock);
691 		uint64_t initialize_state = VDEV_INITIALIZE_NONE;
692 		int err = zap_lookup(vd->vdev_spa->spa_meta_objset,
693 		    vd->vdev_leaf_zap, VDEV_LEAF_ZAP_INITIALIZE_STATE,
694 		    sizeof (initialize_state), 1, &initialize_state);
695 		ASSERT(err == 0 || err == ENOENT);
696 		vd->vdev_initialize_state = initialize_state;
697 
698 		uint64_t timestamp = 0;
699 		err = zap_lookup(vd->vdev_spa->spa_meta_objset,
700 		    vd->vdev_leaf_zap, VDEV_LEAF_ZAP_INITIALIZE_ACTION_TIME,
701 		    sizeof (timestamp), 1, &timestamp);
702 		ASSERT(err == 0 || err == ENOENT);
703 		vd->vdev_initialize_action_time = (time_t)timestamp;
704 
705 		if (vd->vdev_initialize_state == VDEV_INITIALIZE_SUSPENDED ||
706 		    vd->vdev_offline) {
707 			/* load progress for reporting, but don't resume */
708 			VERIFY0(vdev_initialize_load(vd));
709 		} else if (vd->vdev_initialize_state ==
710 		    VDEV_INITIALIZE_ACTIVE && vdev_writeable(vd) &&
711 		    !vd->vdev_top->vdev_removing &&
712 		    vd->vdev_initialize_thread == NULL) {
713 			vdev_initialize(vd);
714 		}
715 
716 		mutex_exit(&vd->vdev_initialize_lock);
717 	}
718 
719 	for (uint64_t i = 0; i < vd->vdev_children; i++) {
720 		vdev_initialize_restart(vd->vdev_child[i]);
721 	}
722 }
723