]> git.proxmox.com Git - mirror_zfs.git/blob - module/zfs/zio.c
projects / mirror_zfs.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Rebase master to b108
[mirror_zfs.git] / module / zfs / zio.c
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 * Copyright 2009 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
24 */
25
26 #include <sys/zfs_context.h>
27 #include <sys/fm/fs/zfs.h>
28 #include <sys/spa.h>
29 #include <sys/txg.h>
30 #include <sys/spa_impl.h>
31 #include <sys/vdev_impl.h>
32 #include <sys/zio_impl.h>
33 #include <sys/zio_compress.h>
34 #include <sys/zio_checksum.h>
35
36 /*
37 * ==========================================================================
38 * I/O priority table
39 * ==========================================================================
40 */
41 uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = {
42 0, /* ZIO_PRIORITY_NOW */
43 0, /* ZIO_PRIORITY_SYNC_READ */
44 0, /* ZIO_PRIORITY_SYNC_WRITE */
45 6, /* ZIO_PRIORITY_ASYNC_READ */
46 4, /* ZIO_PRIORITY_ASYNC_WRITE */
47 4, /* ZIO_PRIORITY_FREE */
48 0, /* ZIO_PRIORITY_CACHE_FILL */
49 0, /* ZIO_PRIORITY_LOG_WRITE */
50 10, /* ZIO_PRIORITY_RESILVER */
51 20, /* ZIO_PRIORITY_SCRUB */
52 };
53
54 /*
55 * ==========================================================================
56 * I/O type descriptions
57 * ==========================================================================
58 */
59 char *zio_type_name[ZIO_TYPES] = {
60 "null", "read", "write", "free", "claim", "ioctl" };
61
62 #define SYNC_PASS_DEFERRED_FREE 1 /* defer frees after this pass */
63 #define SYNC_PASS_DONT_COMPRESS 4 /* don't compress after this pass */
64 #define SYNC_PASS_REWRITE 1 /* rewrite new bps after this pass */
65
66 /*
67 * ==========================================================================
68 * I/O kmem caches
69 * ==========================================================================
70 */
71 kmem_cache_t *zio_cache;
72 kmem_cache_t *zio_link_cache;
73 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
74 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
75
76 #ifdef _KERNEL
77 extern vmem_t *zio_alloc_arena;
78 #endif
79
80 /*
81 * An allocating zio is one that either currently has the DVA allocate
82 * stage set or will have it later in its lifetime.
83 */
84 #define IO_IS_ALLOCATING(zio) \
85 ((zio)->io_orig_pipeline & (1U << ZIO_STAGE_DVA_ALLOCATE))
86
87 void
88 zio_init(void)
89 {
90 size_t c;
91 vmem_t *data_alloc_arena = NULL;
92
93 #ifdef _KERNEL
94 data_alloc_arena = zio_alloc_arena;
95 #endif
96 zio_cache = kmem_cache_create("zio_cache",
97 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
98 zio_link_cache = kmem_cache_create("zio_link_cache",
99 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
100
101 /*
102 * For small buffers, we want a cache for each multiple of
103 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
104 * for each quarter-power of 2. For large buffers, we want
105 * a cache for each multiple of PAGESIZE.
106 */
107 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
108 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
109 size_t p2 = size;
110 size_t align = 0;
111
112 while (p2 & (p2 - 1))
113 p2 &= p2 - 1;
114
115 if (size <= 4 * SPA_MINBLOCKSIZE) {
116 align = SPA_MINBLOCKSIZE;
117 } else if (P2PHASE(size, PAGESIZE) == 0) {
118 align = PAGESIZE;
119 } else if (P2PHASE(size, p2 >> 2) == 0) {
120 align = p2 >> 2;
121 }
122
123 if (align != 0) {
124 char name[36];
125 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
126 zio_buf_cache[c] = kmem_cache_create(name, size,
127 align, NULL, NULL, NULL, NULL, NULL, KMC_NODEBUG);
128
129 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
130 zio_data_buf_cache[c] = kmem_cache_create(name, size,
131 align, NULL, NULL, NULL, NULL, data_alloc_arena,
132 KMC_NODEBUG);
133 }
134 }
135
136 while (--c != 0) {
137 ASSERT(zio_buf_cache[c] != NULL);
138 if (zio_buf_cache[c - 1] == NULL)
139 zio_buf_cache[c - 1] = zio_buf_cache[c];
140
141 ASSERT(zio_data_buf_cache[c] != NULL);
142 if (zio_data_buf_cache[c - 1] == NULL)
143 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
144 }
145
146 zio_inject_init();
147 }
148
149 void
150 zio_fini(void)
151 {
152 size_t c;
153 kmem_cache_t *last_cache = NULL;
154 kmem_cache_t *last_data_cache = NULL;
155
156 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
157 if (zio_buf_cache[c] != last_cache) {
158 last_cache = zio_buf_cache[c];
159 kmem_cache_destroy(zio_buf_cache[c]);
160 }
161 zio_buf_cache[c] = NULL;
162
163 if (zio_data_buf_cache[c] != last_data_cache) {
164 last_data_cache = zio_data_buf_cache[c];
165 kmem_cache_destroy(zio_data_buf_cache[c]);
166 }
167 zio_data_buf_cache[c] = NULL;
168 }
169
170 kmem_cache_destroy(zio_link_cache);
171 kmem_cache_destroy(zio_cache);
172
173 zio_inject_fini();
174 }
175
176 /*
177 * ==========================================================================
178 * Allocate and free I/O buffers
179 * ==========================================================================
180 */
181
182 /*
183 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
184 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
185 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
186 * excess / transient data in-core during a crashdump.
187 */
188 void *
189 zio_buf_alloc(size_t size)
190 {
191 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
192
193 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
194
195 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
196 }
197
198 /*
199 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
200 * crashdump if the kernel panics. This exists so that we will limit the amount
201 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
202 * of kernel heap dumped to disk when the kernel panics)
203 */
204 void *
205 zio_data_buf_alloc(size_t size)
206 {
207 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
208
209 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
210
211 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
212 }
213
214 void
215 zio_buf_free(void *buf, size_t size)
216 {
217 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
218
219 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
220
221 kmem_cache_free(zio_buf_cache[c], buf);
222 }
223
224 void
225 zio_data_buf_free(void *buf, size_t size)
226 {
227 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
228
229 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
230
231 kmem_cache_free(zio_data_buf_cache[c], buf);
232 }
233
234 /*
235 * ==========================================================================
236 * Push and pop I/O transform buffers
237 * ==========================================================================
238 */
239 static void
240 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
241 zio_transform_func_t *transform)
242 {
243 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
244
245 zt->zt_orig_data = zio->io_data;
246 zt->zt_orig_size = zio->io_size;
247 zt->zt_bufsize = bufsize;
248 zt->zt_transform = transform;
249
250 zt->zt_next = zio->io_transform_stack;
251 zio->io_transform_stack = zt;
252
253 zio->io_data = data;
254 zio->io_size = size;
255 }
256
257 static void
258 zio_pop_transforms(zio_t *zio)
259 {
260 zio_transform_t *zt;
261
262 while ((zt = zio->io_transform_stack) != NULL) {
263 if (zt->zt_transform != NULL)
264 zt->zt_transform(zio,
265 zt->zt_orig_data, zt->zt_orig_size);
266
267 zio_buf_free(zio->io_data, zt->zt_bufsize);
268
269 zio->io_data = zt->zt_orig_data;
270 zio->io_size = zt->zt_orig_size;
271 zio->io_transform_stack = zt->zt_next;
272
273 kmem_free(zt, sizeof (zio_transform_t));
274 }
275 }
276
277 /*
278 * ==========================================================================
279 * I/O transform callbacks for subblocks and decompression
280 * ==========================================================================
281 */
282 static void
283 zio_subblock(zio_t *zio, void *data, uint64_t size)
284 {
285 ASSERT(zio->io_size > size);
286
287 if (zio->io_type == ZIO_TYPE_READ)
288 bcopy(zio->io_data, data, size);
289 }
290
291 static void
292 zio_decompress(zio_t *zio, void *data, uint64_t size)
293 {
294 if (zio->io_error == 0 &&
295 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
296 zio->io_data, zio->io_size, data, size) != 0)
297 zio->io_error = EIO;
298 }
299
300 /*
301 * ==========================================================================
302 * I/O parent/child relationships and pipeline interlocks
303 * ==========================================================================
304 */
305 /*
306 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
307 * continue calling these functions until they return NULL.
308 * Otherwise, the next caller will pick up the list walk in
309 * some indeterminate state. (Otherwise every caller would
310 * have to pass in a cookie to keep the state represented by
311 * io_walk_link, which gets annoying.)
312 */
313 zio_t *
314 zio_walk_parents(zio_t *cio)
315 {
316 zio_link_t *zl = cio->io_walk_link;
317 list_t *pl = &cio->io_parent_list;
318
319 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
320 cio->io_walk_link = zl;
321
322 if (zl == NULL)
323 return (NULL);
324
325 ASSERT(zl->zl_child == cio);
326 return (zl->zl_parent);
327 }
328
329 zio_t *
330 zio_walk_children(zio_t *pio)
331 {
332 zio_link_t *zl = pio->io_walk_link;
333 list_t *cl = &pio->io_child_list;
334
335 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
336 pio->io_walk_link = zl;
337
338 if (zl == NULL)
339 return (NULL);
340
341 ASSERT(zl->zl_parent == pio);
342 return (zl->zl_child);
343 }
344
345 zio_t *
346 zio_unique_parent(zio_t *cio)
347 {
348 zio_t *pio = zio_walk_parents(cio);
349
350 VERIFY(zio_walk_parents(cio) == NULL);
351 return (pio);
352 }
353
354 void
355 zio_add_child(zio_t *pio, zio_t *cio)
356 {
357 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
358
359 /*
360 * Logical I/Os can have logical, gang, or vdev children.
361 * Gang I/Os can have gang or vdev children.
362 * Vdev I/Os can only have vdev children.
363 * The following ASSERT captures all of these constraints.
364 */
365 ASSERT(cio->io_child_type <= pio->io_child_type);
366
367 zl->zl_parent = pio;
368 zl->zl_child = cio;
369
370 mutex_enter(&cio->io_lock);
371 mutex_enter(&pio->io_lock);
372
373 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
374
375 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
376 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
377
378 list_insert_head(&pio->io_child_list, zl);
379 list_insert_head(&cio->io_parent_list, zl);
380
381 mutex_exit(&pio->io_lock);
382 mutex_exit(&cio->io_lock);
383 }
384
385 static void
386 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
387 {
388 ASSERT(zl->zl_parent == pio);
389 ASSERT(zl->zl_child == cio);
390
391 mutex_enter(&cio->io_lock);
392 mutex_enter(&pio->io_lock);
393
394 list_remove(&pio->io_child_list, zl);
395 list_remove(&cio->io_parent_list, zl);
396
397 mutex_exit(&pio->io_lock);
398 mutex_exit(&cio->io_lock);
399
400 kmem_cache_free(zio_link_cache, zl);
401 }
402
403 static boolean_t
404 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
405 {
406 uint64_t *countp = &zio->io_children[child][wait];
407 boolean_t waiting = B_FALSE;
408
409 mutex_enter(&zio->io_lock);
410 ASSERT(zio->io_stall == NULL);
411 if (*countp != 0) {
412 zio->io_stage--;
413 zio->io_stall = countp;
414 waiting = B_TRUE;
415 }
416 mutex_exit(&zio->io_lock);
417
418 return (waiting);
419 }
420
421 static void
422 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
423 {
424 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
425 int *errorp = &pio->io_child_error[zio->io_child_type];
426
427 mutex_enter(&pio->io_lock);
428 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
429 *errorp = zio_worst_error(*errorp, zio->io_error);
430 pio->io_reexecute |= zio->io_reexecute;
431 ASSERT3U(*countp, >, 0);
432 if (--*countp == 0 && pio->io_stall == countp) {
433 pio->io_stall = NULL;
434 mutex_exit(&pio->io_lock);
435 zio_execute(pio);
436 } else {
437 mutex_exit(&pio->io_lock);
438 }
439 }
440
441 static void
442 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
443 {
444 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
445 zio->io_error = zio->io_child_error[c];
446 }
447
448 /*
449 * ==========================================================================
450 * Create the various types of I/O (read, write, free, etc)
451 * ==========================================================================
452 */
453 static zio_t *
454 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
455 void *data, uint64_t size, zio_done_func_t *done, void *private,
456 zio_type_t type, int priority, int flags, vdev_t *vd, uint64_t offset,
457 const zbookmark_t *zb, uint8_t stage, uint32_t pipeline)
458 {
459 zio_t *zio;
460
461 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
462 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
463 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
464
465 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
466 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
467 ASSERT(vd || stage == ZIO_STAGE_OPEN);
468
469 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
470 bzero(zio, sizeof (zio_t));
471
472 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
473 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
474
475 list_create(&zio->io_parent_list, sizeof (zio_link_t),
476 offsetof(zio_link_t, zl_parent_node));
477 list_create(&zio->io_child_list, sizeof (zio_link_t),
478 offsetof(zio_link_t, zl_child_node));
479
480 if (vd != NULL)
481 zio->io_child_type = ZIO_CHILD_VDEV;
482 else if (flags & ZIO_FLAG_GANG_CHILD)
483 zio->io_child_type = ZIO_CHILD_GANG;
484 else
485 zio->io_child_type = ZIO_CHILD_LOGICAL;
486
487 if (bp != NULL) {
488 zio->io_bp = bp;
489 zio->io_bp_copy = *bp;
490 zio->io_bp_orig = *bp;
491 if (type != ZIO_TYPE_WRITE)
492 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
493 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
494 if (BP_IS_GANG(bp))
495 pipeline |= ZIO_GANG_STAGES;
496 zio->io_logical = zio;
497 }
498 }
499
500 zio->io_spa = spa;
501 zio->io_txg = txg;
502 zio->io_data = data;
503 zio->io_size = size;
504 zio->io_done = done;
505 zio->io_private = private;
506 zio->io_type = type;
507 zio->io_priority = priority;
508 zio->io_vd = vd;
509 zio->io_offset = offset;
510 zio->io_orig_flags = zio->io_flags = flags;
511 zio->io_orig_stage = zio->io_stage = stage;
512 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
513
514 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
515 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
516
517 if (zb != NULL)
518 zio->io_bookmark = *zb;
519
520 if (pio != NULL) {
521 if (zio->io_logical == NULL)
522 zio->io_logical = pio->io_logical;
523 zio_add_child(pio, zio);
524 }
525
526 return (zio);
527 }
528
529 static void
530 zio_destroy(zio_t *zio)
531 {
532 spa_t *spa = zio->io_spa;
533 uint8_t async_root = zio->io_async_root;
534
535 list_destroy(&zio->io_parent_list);
536 list_destroy(&zio->io_child_list);
537 mutex_destroy(&zio->io_lock);
538 cv_destroy(&zio->io_cv);
539 kmem_cache_free(zio_cache, zio);
540
541 if (async_root) {
542 mutex_enter(&spa->spa_async_root_lock);
543 if (--spa->spa_async_root_count == 0)
544 cv_broadcast(&spa->spa_async_root_cv);
545 mutex_exit(&spa->spa_async_root_lock);
546 }
547 }
548
549 zio_t *
550 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
551 void *private, int flags)
552 {
553 zio_t *zio;
554
555 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
556 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
557 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
558
559 return (zio);
560 }
561
562 zio_t *
563 zio_root(spa_t *spa, zio_done_func_t *done, void *private, int flags)
564 {
565 return (zio_null(NULL, spa, NULL, done, private, flags));
566 }
567
568 zio_t *
569 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
570 void *data, uint64_t size, zio_done_func_t *done, void *private,
571 int priority, int flags, const zbookmark_t *zb)
572 {
573 zio_t *zio;
574
575 zio = zio_create(pio, spa, bp->blk_birth, (blkptr_t *)bp,
576 data, size, done, private,
577 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
578 ZIO_STAGE_OPEN, ZIO_READ_PIPELINE);
579
580 return (zio);
581 }
582
583 void
584 zio_skip_write(zio_t *zio)
585 {
586 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
587 ASSERT(zio->io_stage == ZIO_STAGE_READY);
588 ASSERT(!BP_IS_GANG(zio->io_bp));
589
590 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
591 }
592
593 zio_t *
594 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
595 void *data, uint64_t size, zio_prop_t *zp,
596 zio_done_func_t *ready, zio_done_func_t *done, void *private,
597 int priority, int flags, const zbookmark_t *zb)
598 {
599 zio_t *zio;
600
601 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
602 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
603 zp->zp_compress >= ZIO_COMPRESS_OFF &&
604 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
605 zp->zp_type < DMU_OT_NUMTYPES &&
606 zp->zp_level < 32 &&
607 zp->zp_ndvas > 0 &&
608 zp->zp_ndvas <= spa_max_replication(spa));
609 ASSERT(ready != NULL);
610
611 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
612 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
613 ZIO_STAGE_OPEN, ZIO_WRITE_PIPELINE);
614
615 zio->io_ready = ready;
616 zio->io_prop = *zp;
617
618 return (zio);
619 }
620
621 zio_t *
622 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
623 uint64_t size, zio_done_func_t *done, void *private, int priority,
624 int flags, zbookmark_t *zb)
625 {
626 zio_t *zio;
627
628 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
629 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
630 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
631
632 return (zio);
633 }
634
635 zio_t *
636 zio_free(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
637 zio_done_func_t *done, void *private, int flags)
638 {
639 zio_t *zio;
640
641 ASSERT(!BP_IS_HOLE(bp));
642
643 if (bp->blk_fill == BLK_FILL_ALREADY_FREED)
644 return (zio_null(pio, spa, NULL, NULL, NULL, flags));
645
646 if (txg == spa->spa_syncing_txg &&
647 spa_sync_pass(spa) > SYNC_PASS_DEFERRED_FREE) {
648 bplist_enqueue_deferred(&spa->spa_sync_bplist, bp);
649 return (zio_null(pio, spa, NULL, NULL, NULL, flags));
650 }
651
652 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
653 done, private, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
654 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
655
656 return (zio);
657 }
658
659 zio_t *
660 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
661 zio_done_func_t *done, void *private, int flags)
662 {
663 zio_t *zio;
664
665 /*
666 * A claim is an allocation of a specific block. Claims are needed
667 * to support immediate writes in the intent log. The issue is that
668 * immediate writes contain committed data, but in a txg that was
669 * *not* committed. Upon opening the pool after an unclean shutdown,
670 * the intent log claims all blocks that contain immediate write data
671 * so that the SPA knows they're in use.
672 *
673 * All claims *must* be resolved in the first txg -- before the SPA
674 * starts allocating blocks -- so that nothing is allocated twice.
675 */
676 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
677 ASSERT3U(spa_first_txg(spa), <=, txg);
678
679 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
680 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
681 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
682
683 return (zio);
684 }
685
686 zio_t *
687 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
688 zio_done_func_t *done, void *private, int priority, int flags)
689 {
690 zio_t *zio;
691 int c;
692
693 if (vd->vdev_children == 0) {
694 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
695 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
696 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
697
698 zio->io_cmd = cmd;
699 } else {
700 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
701
702 for (c = 0; c < vd->vdev_children; c++)
703 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
704 done, private, priority, flags));
705 }
706
707 return (zio);
708 }
709
710 zio_t *
711 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
712 void *data, int checksum, zio_done_func_t *done, void *private,
713 int priority, int flags, boolean_t labels)
714 {
715 zio_t *zio;
716
717 ASSERT(vd->vdev_children == 0);
718 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
719 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
720 ASSERT3U(offset + size, <=, vd->vdev_psize);
721
722 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
723 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
724 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
725
726 zio->io_prop.zp_checksum = checksum;
727
728 return (zio);
729 }
730
731 zio_t *
732 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
733 void *data, int checksum, zio_done_func_t *done, void *private,
734 int priority, int flags, boolean_t labels)
735 {
736 zio_t *zio;
737
738 ASSERT(vd->vdev_children == 0);
739 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
740 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
741 ASSERT3U(offset + size, <=, vd->vdev_psize);
742
743 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
744 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
745 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
746
747 zio->io_prop.zp_checksum = checksum;
748
749 if (zio_checksum_table[checksum].ci_zbt) {
750 /*
751 * zbt checksums are necessarily destructive -- they modify
752 * the end of the write buffer to hold the verifier/checksum.
753 * Therefore, we must make a local copy in case the data is
754 * being written to multiple places in parallel.
755 */
756 void *wbuf = zio_buf_alloc(size);
757 bcopy(data, wbuf, size);
758 zio_push_transform(zio, wbuf, size, size, NULL);
759 }
760
761 return (zio);
762 }
763
764 /*
765 * Create a child I/O to do some work for us.
766 */
767 zio_t *
768 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
769 void *data, uint64_t size, int type, int priority, int flags,
770 zio_done_func_t *done, void *private)
771 {
772 uint32_t pipeline = ZIO_VDEV_CHILD_PIPELINE;
773 zio_t *zio;
774
775 ASSERT(vd->vdev_parent ==
776 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
777
778 if (type == ZIO_TYPE_READ && bp != NULL) {
779 /*
780 * If we have the bp, then the child should perform the
781 * checksum and the parent need not. This pushes error
782 * detection as close to the leaves as possible and
783 * eliminates redundant checksums in the interior nodes.
784 */
785 pipeline |= 1U << ZIO_STAGE_CHECKSUM_VERIFY;
786 pio->io_pipeline &= ~(1U << ZIO_STAGE_CHECKSUM_VERIFY);
787 }
788
789 if (vd->vdev_children == 0)
790 offset += VDEV_LABEL_START_SIZE;
791
792 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
793 done, private, type, priority,
794 (pio->io_flags & ZIO_FLAG_VDEV_INHERIT) |
795 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | flags,
796 vd, offset, &pio->io_bookmark,
797 ZIO_STAGE_VDEV_IO_START - 1, pipeline);
798
799 return (zio);
800 }
801
802 zio_t *
803 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
804 int type, int priority, int flags, zio_done_func_t *done, void *private)
805 {
806 zio_t *zio;
807
808 ASSERT(vd->vdev_ops->vdev_op_leaf);
809
810 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
811 data, size, done, private, type, priority,
812 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
813 vd, offset, NULL,
814 ZIO_STAGE_VDEV_IO_START - 1, ZIO_VDEV_CHILD_PIPELINE);
815
816 return (zio);
817 }
818
819 void
820 zio_flush(zio_t *zio, vdev_t *vd)
821 {
822 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
823 NULL, NULL, ZIO_PRIORITY_NOW,
824 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
825 }
826
827 /*
828 * ==========================================================================
829 * Prepare to read and write logical blocks
830 * ==========================================================================
831 */
832
833 static int
834 zio_read_bp_init(zio_t *zio)
835 {
836 blkptr_t *bp = zio->io_bp;
837
838 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
839 zio->io_logical == zio && !(zio->io_flags & ZIO_FLAG_RAW)) {
840 uint64_t csize = BP_GET_PSIZE(bp);
841 void *cbuf = zio_buf_alloc(csize);
842
843 zio_push_transform(zio, cbuf, csize, csize, zio_decompress);
844 }
845
846 if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
847 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
848
849 return (ZIO_PIPELINE_CONTINUE);
850 }
851
852 static int
853 zio_write_bp_init(zio_t *zio)
854 {
855 zio_prop_t *zp = &zio->io_prop;
856 int compress = zp->zp_compress;
857 blkptr_t *bp = zio->io_bp;
858 void *cbuf;
859 uint64_t lsize = zio->io_size;
860 uint64_t csize = lsize;
861 uint64_t cbufsize = 0;
862 int pass = 1;
863
864 /*
865 * If our children haven't all reached the ready stage,
866 * wait for them and then repeat this pipeline stage.
867 */
868 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
869 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
870 return (ZIO_PIPELINE_STOP);
871
872 if (!IO_IS_ALLOCATING(zio))
873 return (ZIO_PIPELINE_CONTINUE);
874
875 ASSERT(compress != ZIO_COMPRESS_INHERIT);
876
877 if (bp->blk_birth == zio->io_txg) {
878 /*
879 * We're rewriting an existing block, which means we're
880 * working on behalf of spa_sync(). For spa_sync() to
881 * converge, it must eventually be the case that we don't
882 * have to allocate new blocks. But compression changes
883 * the blocksize, which forces a reallocate, and makes
884 * convergence take longer. Therefore, after the first
885 * few passes, stop compressing to ensure convergence.
886 */
887 pass = spa_sync_pass(zio->io_spa);
888 ASSERT(pass > 1);
889
890 if (pass > SYNC_PASS_DONT_COMPRESS)
891 compress = ZIO_COMPRESS_OFF;
892
893 /*
894 * Only MOS (objset 0) data should need to be rewritten.
895 */
896 ASSERT(zio->io_logical->io_bookmark.zb_objset == 0);
897
898 /* Make sure someone doesn't change their mind on overwrites */
899 ASSERT(MIN(zp->zp_ndvas + BP_IS_GANG(bp),
900 spa_max_replication(zio->io_spa)) == BP_GET_NDVAS(bp));
901 }
902
903 if (compress != ZIO_COMPRESS_OFF) {
904 if (!zio_compress_data(compress, zio->io_data, zio->io_size,
905 &cbuf, &csize, &cbufsize)) {
906 compress = ZIO_COMPRESS_OFF;
907 } else if (csize != 0) {
908 zio_push_transform(zio, cbuf, csize, cbufsize, NULL);
909 }
910 }
911
912 /*
913 * The final pass of spa_sync() must be all rewrites, but the first
914 * few passes offer a trade-off: allocating blocks defers convergence,
915 * but newly allocated blocks are sequential, so they can be written
916 * to disk faster. Therefore, we allow the first few passes of
917 * spa_sync() to allocate new blocks, but force rewrites after that.
918 * There should only be a handful of blocks after pass 1 in any case.
919 */
920 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == csize &&
921 pass > SYNC_PASS_REWRITE) {
922 ASSERT(csize != 0);
923 uint32_t gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
924 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
925 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
926 } else {
927 BP_ZERO(bp);
928 zio->io_pipeline = ZIO_WRITE_PIPELINE;
929 }
930
931 if (csize == 0) {
932 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
933 } else {
934 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
935 BP_SET_LSIZE(bp, lsize);
936 BP_SET_PSIZE(bp, csize);
937 BP_SET_COMPRESS(bp, compress);
938 BP_SET_CHECKSUM(bp, zp->zp_checksum);
939 BP_SET_TYPE(bp, zp->zp_type);
940 BP_SET_LEVEL(bp, zp->zp_level);
941 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
942 }
943
944 return (ZIO_PIPELINE_CONTINUE);
945 }
946
947 /*
948 * ==========================================================================
949 * Execute the I/O pipeline
950 * ==========================================================================
951 */
952
953 static void
954 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q)
955 {
956 zio_type_t t = zio->io_type;
957
958 /*
959 * If we're a config writer, the normal issue and interrupt threads
960 * may all be blocked waiting for the config lock. In this case,
961 * select the otherwise-unused taskq for ZIO_TYPE_NULL.
962 */
963 if (zio->io_flags & ZIO_FLAG_CONFIG_WRITER)
964 t = ZIO_TYPE_NULL;
965
966 /*
967 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
968 */
969 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
970 t = ZIO_TYPE_NULL;
971
972 (void) taskq_dispatch(zio->io_spa->spa_zio_taskq[t][q],
973 (task_func_t *)zio_execute, zio, TQ_SLEEP);
974 }
975
976 static boolean_t
977 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
978 {
979 kthread_t *executor = zio->io_executor;
980 spa_t *spa = zio->io_spa;
981
982 for (zio_type_t t = 0; t < ZIO_TYPES; t++)
983 if (taskq_member(spa->spa_zio_taskq[t][q], executor))
984 return (B_TRUE);
985
986 return (B_FALSE);
987 }
988
989 static int
990 zio_issue_async(zio_t *zio)
991 {
992 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE);
993
994 return (ZIO_PIPELINE_STOP);
995 }
996
997 void
998 zio_interrupt(zio_t *zio)
999 {
1000 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT);
1001 }
1002
1003 /*
1004 * Execute the I/O pipeline until one of the following occurs:
1005 * (1) the I/O completes; (2) the pipeline stalls waiting for
1006 * dependent child I/Os; (3) the I/O issues, so we're waiting
1007 * for an I/O completion interrupt; (4) the I/O is delegated by
1008 * vdev-level caching or aggregation; (5) the I/O is deferred
1009 * due to vdev-level queueing; (6) the I/O is handed off to
1010 * another thread. In all cases, the pipeline stops whenever
1011 * there's no CPU work; it never burns a thread in cv_wait().
1012 *
1013 * There's no locking on io_stage because there's no legitimate way
1014 * for multiple threads to be attempting to process the same I/O.
1015 */
1016 static zio_pipe_stage_t *zio_pipeline[ZIO_STAGES];
1017
1018 void
1019 zio_execute(zio_t *zio)
1020 {
1021 zio->io_executor = curthread;
1022
1023 while (zio->io_stage < ZIO_STAGE_DONE) {
1024 uint32_t pipeline = zio->io_pipeline;
1025 zio_stage_t stage = zio->io_stage;
1026 int rv;
1027
1028 ASSERT(!MUTEX_HELD(&zio->io_lock));
1029
1030 while (((1U << ++stage) & pipeline) == 0)
1031 continue;
1032
1033 ASSERT(stage <= ZIO_STAGE_DONE);
1034 ASSERT(zio->io_stall == NULL);
1035
1036 /*
1037 * If we are in interrupt context and this pipeline stage
1038 * will grab a config lock that is held across I/O,
1039 * issue async to avoid deadlock.
1040 */
1041 if (((1U << stage) & ZIO_CONFIG_LOCK_BLOCKING_STAGES) &&
1042 zio->io_vd == NULL &&
1043 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1044 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE);
1045 return;
1046 }
1047
1048 zio->io_stage = stage;
1049 rv = zio_pipeline[stage](zio);
1050
1051 if (rv == ZIO_PIPELINE_STOP)
1052 return;
1053
1054 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1055 }
1056 }
1057
1058 /*
1059 * ==========================================================================
1060 * Initiate I/O, either sync or async
1061 * ==========================================================================
1062 */
1063 int
1064 zio_wait(zio_t *zio)
1065 {
1066 int error;
1067
1068 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1069 ASSERT(zio->io_executor == NULL);
1070
1071 zio->io_waiter = curthread;
1072
1073 zio_execute(zio);
1074
1075 mutex_enter(&zio->io_lock);
1076 while (zio->io_executor != NULL)
1077 cv_wait(&zio->io_cv, &zio->io_lock);
1078 mutex_exit(&zio->io_lock);
1079
1080 error = zio->io_error;
1081 zio_destroy(zio);
1082
1083 return (error);
1084 }
1085
1086 void
1087 zio_nowait(zio_t *zio)
1088 {
1089 ASSERT(zio->io_executor == NULL);
1090
1091 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1092 zio_unique_parent(zio) == NULL) {
1093 /*
1094 * This is a logical async I/O with no parent to wait for it.
1095 * Track how many outstanding I/Os of this type exist so
1096 * that spa_unload() knows when they are all done.
1097 */
1098 spa_t *spa = zio->io_spa;
1099 zio->io_async_root = B_TRUE;
1100 mutex_enter(&spa->spa_async_root_lock);
1101 spa->spa_async_root_count++;
1102 mutex_exit(&spa->spa_async_root_lock);
1103 }
1104
1105 zio_execute(zio);
1106 }
1107
1108 /*
1109 * ==========================================================================
1110 * Reexecute or suspend/resume failed I/O
1111 * ==========================================================================
1112 */
1113
1114 static void
1115 zio_reexecute(zio_t *pio)
1116 {
1117 zio_t *cio, *cio_next;
1118
1119 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1120 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1121
1122 pio->io_flags = pio->io_orig_flags;
1123 pio->io_stage = pio->io_orig_stage;
1124 pio->io_pipeline = pio->io_orig_pipeline;
1125 pio->io_reexecute = 0;
1126 pio->io_error = 0;
1127 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1128 pio->io_state[w] = 0;
1129 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1130 pio->io_child_error[c] = 0;
1131
1132 if (IO_IS_ALLOCATING(pio)) {
1133 /*
1134 * Remember the failed bp so that the io_ready() callback
1135 * can update its accounting upon reexecution. The block
1136 * was already freed in zio_done(); we indicate this with
1137 * a fill count of -1 so that zio_free() knows to skip it.
1138 */
1139 blkptr_t *bp = pio->io_bp;
1140 ASSERT(bp->blk_birth == 0 || bp->blk_birth == pio->io_txg);
1141 bp->blk_fill = BLK_FILL_ALREADY_FREED;
1142 pio->io_bp_orig = *bp;
1143 BP_ZERO(bp);
1144 }
1145
1146 /*
1147 * As we reexecute pio's children, new children could be created.
1148 * New children go to the head of pio's io_child_list, however,
1149 * so we will (correctly) not reexecute them. The key is that
1150 * the remainder of pio's io_child_list, from 'cio_next' onward,
1151 * cannot be affected by any side effects of reexecuting 'cio'.
1152 */
1153 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1154 cio_next = zio_walk_children(pio);
1155 mutex_enter(&pio->io_lock);
1156 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1157 pio->io_children[cio->io_child_type][w]++;
1158 mutex_exit(&pio->io_lock);
1159 zio_reexecute(cio);
1160 }
1161
1162 /*
1163 * Now that all children have been reexecuted, execute the parent.
1164 */
1165 zio_execute(pio);
1166 }
1167
1168 void
1169 zio_suspend(spa_t *spa, zio_t *zio)
1170 {
1171 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1172 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1173 "failure and the failure mode property for this pool "
1174 "is set to panic.", spa_name(spa));
1175
1176 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1177
1178 mutex_enter(&spa->spa_suspend_lock);
1179
1180 if (spa->spa_suspend_zio_root == NULL)
1181 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL, 0);
1182
1183 spa->spa_suspended = B_TRUE;
1184
1185 if (zio != NULL) {
1186 ASSERT(zio != spa->spa_suspend_zio_root);
1187 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1188 ASSERT(zio_unique_parent(zio) == NULL);
1189 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1190 zio_add_child(spa->spa_suspend_zio_root, zio);
1191 }
1192
1193 mutex_exit(&spa->spa_suspend_lock);
1194 }
1195
1196 void
1197 zio_resume(spa_t *spa)
1198 {
1199 zio_t *pio, *cio, *cio_next;
1200
1201 /*
1202 * Reexecute all previously suspended i/o.
1203 */
1204 mutex_enter(&spa->spa_suspend_lock);
1205 spa->spa_suspended = B_FALSE;
1206 cv_broadcast(&spa->spa_suspend_cv);
1207 pio = spa->spa_suspend_zio_root;
1208 spa->spa_suspend_zio_root = NULL;
1209 mutex_exit(&spa->spa_suspend_lock);
1210
1211 if (pio == NULL)
1212 return;
1213
1214 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1215 zio_link_t *zl = pio->io_walk_link;
1216 cio_next = zio_walk_children(pio);
1217 zio_remove_child(pio, cio, zl);
1218 zio_reexecute(cio);
1219 }
1220
1221 ASSERT(pio->io_children[ZIO_CHILD_LOGICAL][ZIO_WAIT_DONE] == 0);
1222
1223 (void) zio_wait(pio);
1224 }
1225
1226 void
1227 zio_resume_wait(spa_t *spa)
1228 {
1229 mutex_enter(&spa->spa_suspend_lock);
1230 while (spa_suspended(spa))
1231 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1232 mutex_exit(&spa->spa_suspend_lock);
1233 }
1234
1235 /*
1236 * ==========================================================================
1237 * Gang blocks.
1238 *
1239 * A gang block is a collection of small blocks that looks to the DMU
1240 * like one large block. When zio_dva_allocate() cannot find a block
1241 * of the requested size, due to either severe fragmentation or the pool
1242 * being nearly full, it calls zio_write_gang_block() to construct the
1243 * block from smaller fragments.
1244 *
1245 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1246 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1247 * an indirect block: it's an array of block pointers. It consumes
1248 * only one sector and hence is allocatable regardless of fragmentation.
1249 * The gang header's bps point to its gang members, which hold the data.
1250 *
1251 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1252 * as the verifier to ensure uniqueness of the SHA256 checksum.
1253 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1254 * not the gang header. This ensures that data block signatures (needed for
1255 * deduplication) are independent of how the block is physically stored.
1256 *
1257 * Gang blocks can be nested: a gang member may itself be a gang block.
1258 * Thus every gang block is a tree in which root and all interior nodes are
1259 * gang headers, and the leaves are normal blocks that contain user data.
1260 * The root of the gang tree is called the gang leader.
1261 *
1262 * To perform any operation (read, rewrite, free, claim) on a gang block,
1263 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1264 * in the io_gang_tree field of the original logical i/o by recursively
1265 * reading the gang leader and all gang headers below it. This yields
1266 * an in-core tree containing the contents of every gang header and the
1267 * bps for every constituent of the gang block.
1268 *
1269 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1270 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1271 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1272 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1273 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1274 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1275 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1276 * of the gang header plus zio_checksum_compute() of the data to update the
1277 * gang header's blk_cksum as described above.
1278 *
1279 * The two-phase assemble/issue model solves the problem of partial failure --
1280 * what if you'd freed part of a gang block but then couldn't read the
1281 * gang header for another part? Assembling the entire gang tree first
1282 * ensures that all the necessary gang header I/O has succeeded before
1283 * starting the actual work of free, claim, or write. Once the gang tree
1284 * is assembled, free and claim are in-memory operations that cannot fail.
1285 *
1286 * In the event that a gang write fails, zio_dva_unallocate() walks the
1287 * gang tree to immediately free (i.e. insert back into the space map)
1288 * everything we've allocated. This ensures that we don't get ENOSPC
1289 * errors during repeated suspend/resume cycles due to a flaky device.
1290 *
1291 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1292 * the gang tree, we won't modify the block, so we can safely defer the free
1293 * (knowing that the block is still intact). If we *can* assemble the gang
1294 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1295 * each constituent bp and we can allocate a new block on the next sync pass.
1296 *
1297 * In all cases, the gang tree allows complete recovery from partial failure.
1298 * ==========================================================================
1299 */
1300
1301 static zio_t *
1302 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1303 {
1304 if (gn != NULL)
1305 return (pio);
1306
1307 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1308 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1309 &pio->io_bookmark));
1310 }
1311
1312 zio_t *
1313 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1314 {
1315 zio_t *zio;
1316
1317 if (gn != NULL) {
1318 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1319 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1320 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1321 /*
1322 * As we rewrite each gang header, the pipeline will compute
1323 * a new gang block header checksum for it; but no one will
1324 * compute a new data checksum, so we do that here. The one
1325 * exception is the gang leader: the pipeline already computed
1326 * its data checksum because that stage precedes gang assembly.
1327 * (Presently, nothing actually uses interior data checksums;
1328 * this is just good hygiene.)
1329 */
1330 if (gn != pio->io_logical->io_gang_tree) {
1331 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1332 data, BP_GET_PSIZE(bp));
1333 }
1334 } else {
1335 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1336 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1337 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1338 }
1339
1340 return (zio);
1341 }
1342
1343 /* ARGSUSED */
1344 zio_t *
1345 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1346 {
1347 return (zio_free(pio, pio->io_spa, pio->io_txg, bp,
1348 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1349 }
1350
1351 /* ARGSUSED */
1352 zio_t *
1353 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1354 {
1355 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1356 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1357 }
1358
1359 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1360 NULL,
1361 zio_read_gang,
1362 zio_rewrite_gang,
1363 zio_free_gang,
1364 zio_claim_gang,
1365 NULL
1366 };
1367
1368 static void zio_gang_tree_assemble_done(zio_t *zio);
1369
1370 static zio_gang_node_t *
1371 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1372 {
1373 zio_gang_node_t *gn;
1374
1375 ASSERT(*gnpp == NULL);
1376
1377 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1378 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1379 *gnpp = gn;
1380
1381 return (gn);
1382 }
1383
1384 static void
1385 zio_gang_node_free(zio_gang_node_t **gnpp)
1386 {
1387 zio_gang_node_t *gn = *gnpp;
1388
1389 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1390 ASSERT(gn->gn_child[g] == NULL);
1391
1392 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1393 kmem_free(gn, sizeof (*gn));
1394 *gnpp = NULL;
1395 }
1396
1397 static void
1398 zio_gang_tree_free(zio_gang_node_t **gnpp)
1399 {
1400 zio_gang_node_t *gn = *gnpp;
1401
1402 if (gn == NULL)
1403 return;
1404
1405 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1406 zio_gang_tree_free(&gn->gn_child[g]);
1407
1408 zio_gang_node_free(gnpp);
1409 }
1410
1411 static void
1412 zio_gang_tree_assemble(zio_t *lio, blkptr_t *bp, zio_gang_node_t **gnpp)
1413 {
1414 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1415
1416 ASSERT(lio->io_logical == lio);
1417 ASSERT(BP_IS_GANG(bp));
1418
1419 zio_nowait(zio_read(lio, lio->io_spa, bp, gn->gn_gbh,
1420 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1421 lio->io_priority, ZIO_GANG_CHILD_FLAGS(lio), &lio->io_bookmark));
1422 }
1423
1424 static void
1425 zio_gang_tree_assemble_done(zio_t *zio)
1426 {
1427 zio_t *lio = zio->io_logical;
1428 zio_gang_node_t *gn = zio->io_private;
1429 blkptr_t *bp = zio->io_bp;
1430 zio_t *pio = zio_unique_parent(zio);
1431
1432 ASSERT(pio == lio);
1433 ASSERT(zio_walk_children(zio) == NULL);
1434
1435 if (zio->io_error)
1436 return;
1437
1438 if (BP_SHOULD_BYTESWAP(bp))
1439 byteswap_uint64_array(zio->io_data, zio->io_size);
1440
1441 ASSERT(zio->io_data == gn->gn_gbh);
1442 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1443 ASSERT(gn->gn_gbh->zg_tail.zbt_magic == ZBT_MAGIC);
1444
1445 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1446 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1447 if (!BP_IS_GANG(gbp))
1448 continue;
1449 zio_gang_tree_assemble(lio, gbp, &gn->gn_child[g]);
1450 }
1451 }
1452
1453 static void
1454 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1455 {
1456 zio_t *lio = pio->io_logical;
1457 zio_t *zio;
1458
1459 ASSERT(BP_IS_GANG(bp) == !!gn);
1460 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(lio->io_bp));
1461 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == lio->io_gang_tree);
1462
1463 /*
1464 * If you're a gang header, your data is in gn->gn_gbh.
1465 * If you're a gang member, your data is in 'data' and gn == NULL.
1466 */
1467 zio = zio_gang_issue_func[lio->io_type](pio, bp, gn, data);
1468
1469 if (gn != NULL) {
1470 ASSERT(gn->gn_gbh->zg_tail.zbt_magic == ZBT_MAGIC);
1471
1472 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1473 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1474 if (BP_IS_HOLE(gbp))
1475 continue;
1476 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1477 data = (char *)data + BP_GET_PSIZE(gbp);
1478 }
1479 }
1480
1481 if (gn == lio->io_gang_tree)
1482 ASSERT3P((char *)lio->io_data + lio->io_size, ==, data);
1483
1484 if (zio != pio)
1485 zio_nowait(zio);
1486 }
1487
1488 static int
1489 zio_gang_assemble(zio_t *zio)
1490 {
1491 blkptr_t *bp = zio->io_bp;
1492
1493 ASSERT(BP_IS_GANG(bp) && zio == zio->io_logical);
1494
1495 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1496
1497 return (ZIO_PIPELINE_CONTINUE);
1498 }
1499
1500 static int
1501 zio_gang_issue(zio_t *zio)
1502 {
1503 zio_t *lio = zio->io_logical;
1504 blkptr_t *bp = zio->io_bp;
1505
1506 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1507 return (ZIO_PIPELINE_STOP);
1508
1509 ASSERT(BP_IS_GANG(bp) && zio == lio);
1510
1511 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1512 zio_gang_tree_issue(lio, lio->io_gang_tree, bp, lio->io_data);
1513 else
1514 zio_gang_tree_free(&lio->io_gang_tree);
1515
1516 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1517
1518 return (ZIO_PIPELINE_CONTINUE);
1519 }
1520
1521 static void
1522 zio_write_gang_member_ready(zio_t *zio)
1523 {
1524 zio_t *pio = zio_unique_parent(zio);
1525 zio_t *lio = zio->io_logical;
1526 dva_t *cdva = zio->io_bp->blk_dva;
1527 dva_t *pdva = pio->io_bp->blk_dva;
1528 uint64_t asize;
1529
1530 if (BP_IS_HOLE(zio->io_bp))
1531 return;
1532
1533 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1534
1535 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1536 ASSERT3U(zio->io_prop.zp_ndvas, ==, lio->io_prop.zp_ndvas);
1537 ASSERT3U(zio->io_prop.zp_ndvas, <=, BP_GET_NDVAS(zio->io_bp));
1538 ASSERT3U(pio->io_prop.zp_ndvas, <=, BP_GET_NDVAS(pio->io_bp));
1539 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1540
1541 mutex_enter(&pio->io_lock);
1542 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1543 ASSERT(DVA_GET_GANG(&pdva[d]));
1544 asize = DVA_GET_ASIZE(&pdva[d]);
1545 asize += DVA_GET_ASIZE(&cdva[d]);
1546 DVA_SET_ASIZE(&pdva[d], asize);
1547 }
1548 mutex_exit(&pio->io_lock);
1549 }
1550
1551 static int
1552 zio_write_gang_block(zio_t *pio)
1553 {
1554 spa_t *spa = pio->io_spa;
1555 blkptr_t *bp = pio->io_bp;
1556 zio_t *lio = pio->io_logical;
1557 zio_t *zio;
1558 zio_gang_node_t *gn, **gnpp;
1559 zio_gbh_phys_t *gbh;
1560 uint64_t txg = pio->io_txg;
1561 uint64_t resid = pio->io_size;
1562 uint64_t lsize;
1563 int ndvas = lio->io_prop.zp_ndvas;
1564 int gbh_ndvas = MIN(ndvas + 1, spa_max_replication(spa));
1565 zio_prop_t zp;
1566 int error;
1567
1568 error = metaslab_alloc(spa, spa->spa_normal_class, SPA_GANGBLOCKSIZE,
1569 bp, gbh_ndvas, txg, pio == lio ? NULL : lio->io_bp,
1570 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1571 if (error) {
1572 pio->io_error = error;
1573 return (ZIO_PIPELINE_CONTINUE);
1574 }
1575
1576 if (pio == lio) {
1577 gnpp = &lio->io_gang_tree;
1578 } else {
1579 gnpp = pio->io_private;
1580 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1581 }
1582
1583 gn = zio_gang_node_alloc(gnpp);
1584 gbh = gn->gn_gbh;
1585 bzero(gbh, SPA_GANGBLOCKSIZE);
1586
1587 /*
1588 * Create the gang header.
1589 */
1590 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1591 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1592
1593 /*
1594 * Create and nowait the gang children.
1595 */
1596 for (int g = 0; resid != 0; resid -= lsize, g++) {
1597 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1598 SPA_MINBLOCKSIZE);
1599 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1600
1601 zp.zp_checksum = lio->io_prop.zp_checksum;
1602 zp.zp_compress = ZIO_COMPRESS_OFF;
1603 zp.zp_type = DMU_OT_NONE;
1604 zp.zp_level = 0;
1605 zp.zp_ndvas = lio->io_prop.zp_ndvas;
1606
1607 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1608 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1609 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1610 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1611 &pio->io_bookmark));
1612 }
1613
1614 /*
1615 * Set pio's pipeline to just wait for zio to finish.
1616 */
1617 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1618
1619 zio_nowait(zio);
1620
1621 return (ZIO_PIPELINE_CONTINUE);
1622 }
1623
1624 /*
1625 * ==========================================================================
1626 * Allocate and free blocks
1627 * ==========================================================================
1628 */
1629
1630 static int
1631 zio_dva_allocate(zio_t *zio)
1632 {
1633 spa_t *spa = zio->io_spa;
1634 metaslab_class_t *mc = spa->spa_normal_class;
1635 blkptr_t *bp = zio->io_bp;
1636 int error;
1637
1638 ASSERT(BP_IS_HOLE(bp));
1639 ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
1640 ASSERT3U(zio->io_prop.zp_ndvas, >, 0);
1641 ASSERT3U(zio->io_prop.zp_ndvas, <=, spa_max_replication(spa));
1642 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
1643
1644 error = metaslab_alloc(spa, mc, zio->io_size, bp,
1645 zio->io_prop.zp_ndvas, zio->io_txg, NULL, 0);
1646
1647 if (error) {
1648 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
1649 return (zio_write_gang_block(zio));
1650 zio->io_error = error;
1651 }
1652
1653 return (ZIO_PIPELINE_CONTINUE);
1654 }
1655
1656 static int
1657 zio_dva_free(zio_t *zio)
1658 {
1659 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
1660
1661 return (ZIO_PIPELINE_CONTINUE);
1662 }
1663
1664 static int
1665 zio_dva_claim(zio_t *zio)
1666 {
1667 int error;
1668
1669 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
1670 if (error)
1671 zio->io_error = error;
1672
1673 return (ZIO_PIPELINE_CONTINUE);
1674 }
1675
1676 /*
1677 * Undo an allocation. This is used by zio_done() when an I/O fails
1678 * and we want to give back the block we just allocated.
1679 * This handles both normal blocks and gang blocks.
1680 */
1681 static void
1682 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
1683 {
1684 spa_t *spa = zio->io_spa;
1685 boolean_t now = !(zio->io_flags & ZIO_FLAG_IO_REWRITE);
1686
1687 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
1688
1689 if (zio->io_bp == bp && !now) {
1690 /*
1691 * This is a rewrite for sync-to-convergence.
1692 * We can't do a metaslab_free(NOW) because bp wasn't allocated
1693 * during this sync pass, which means that metaslab_sync()
1694 * already committed the allocation.
1695 */
1696 ASSERT(DVA_EQUAL(BP_IDENTITY(bp),
1697 BP_IDENTITY(&zio->io_bp_orig)));
1698 ASSERT(spa_sync_pass(spa) > 1);
1699
1700 if (BP_IS_GANG(bp) && gn == NULL) {
1701 /*
1702 * This is a gang leader whose gang header(s) we
1703 * couldn't read now, so defer the free until later.
1704 * The block should still be intact because without
1705 * the headers, we'd never even start the rewrite.
1706 */
1707 bplist_enqueue_deferred(&spa->spa_sync_bplist, bp);
1708 return;
1709 }
1710 }
1711
1712 if (!BP_IS_HOLE(bp))
1713 metaslab_free(spa, bp, bp->blk_birth, now);
1714
1715 if (gn != NULL) {
1716 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1717 zio_dva_unallocate(zio, gn->gn_child[g],
1718 &gn->gn_gbh->zg_blkptr[g]);
1719 }
1720 }
1721 }
1722
1723 /*
1724 * Try to allocate an intent log block. Return 0 on success, errno on failure.
1725 */
1726 int
1727 zio_alloc_blk(spa_t *spa, uint64_t size, blkptr_t *new_bp, blkptr_t *old_bp,
1728 uint64_t txg)
1729 {
1730 int error;
1731
1732 error = metaslab_alloc(spa, spa->spa_log_class, size,
1733 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
1734
1735 if (error)
1736 error = metaslab_alloc(spa, spa->spa_normal_class, size,
1737 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
1738
1739 if (error == 0) {
1740 BP_SET_LSIZE(new_bp, size);
1741 BP_SET_PSIZE(new_bp, size);
1742 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
1743 BP_SET_CHECKSUM(new_bp, ZIO_CHECKSUM_ZILOG);
1744 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
1745 BP_SET_LEVEL(new_bp, 0);
1746 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
1747 }
1748
1749 return (error);
1750 }
1751
1752 /*
1753 * Free an intent log block. We know it can't be a gang block, so there's
1754 * nothing to do except metaslab_free() it.
1755 */
1756 void
1757 zio_free_blk(spa_t *spa, blkptr_t *bp, uint64_t txg)
1758 {
1759 ASSERT(!BP_IS_GANG(bp));
1760
1761 metaslab_free(spa, bp, txg, B_FALSE);
1762 }
1763
1764 /*
1765 * ==========================================================================
1766 * Read and write to physical devices
1767 * ==========================================================================
1768 */
1769 static int
1770 zio_vdev_io_start(zio_t *zio)
1771 {
1772 vdev_t *vd = zio->io_vd;
1773 uint64_t align;
1774 spa_t *spa = zio->io_spa;
1775
1776 ASSERT(zio->io_error == 0);
1777 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
1778
1779 if (vd == NULL) {
1780 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
1781 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
1782
1783 /*
1784 * The mirror_ops handle multiple DVAs in a single BP.
1785 */
1786 return (vdev_mirror_ops.vdev_op_io_start(zio));
1787 }
1788
1789 align = 1ULL << vd->vdev_top->vdev_ashift;
1790
1791 if (P2PHASE(zio->io_size, align) != 0) {
1792 uint64_t asize = P2ROUNDUP(zio->io_size, align);
1793 char *abuf = zio_buf_alloc(asize);
1794 ASSERT(vd == vd->vdev_top);
1795 if (zio->io_type == ZIO_TYPE_WRITE) {
1796 bcopy(zio->io_data, abuf, zio->io_size);
1797 bzero(abuf + zio->io_size, asize - zio->io_size);
1798 }
1799 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
1800 }
1801
1802 ASSERT(P2PHASE(zio->io_offset, align) == 0);
1803 ASSERT(P2PHASE(zio->io_size, align) == 0);
1804 ASSERT(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
1805
1806 /*
1807 * If this is a repair I/O, and there's no self-healing involved --
1808 * that is, we're just resilvering what we expect to resilver --
1809 * then don't do the I/O unless zio's txg is actually in vd's DTL.
1810 * This prevents spurious resilvering with nested replication.
1811 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
1812 * A is out of date, we'll read from C+D, then use the data to
1813 * resilver A+B -- but we don't actually want to resilver B, just A.
1814 * The top-level mirror has no way to know this, so instead we just
1815 * discard unnecessary repairs as we work our way down the vdev tree.
1816 * The same logic applies to any form of nested replication:
1817 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
1818 */
1819 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
1820 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
1821 zio->io_txg != 0 && /* not a delegated i/o */
1822 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
1823 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
1824 zio_vdev_io_bypass(zio);
1825 return (ZIO_PIPELINE_CONTINUE);
1826 }
1827
1828 if (vd->vdev_ops->vdev_op_leaf &&
1829 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
1830
1831 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
1832 return (ZIO_PIPELINE_CONTINUE);
1833
1834 if ((zio = vdev_queue_io(zio)) == NULL)
1835 return (ZIO_PIPELINE_STOP);
1836
1837 if (!vdev_accessible(vd, zio)) {
1838 zio->io_error = ENXIO;
1839 zio_interrupt(zio);
1840 return (ZIO_PIPELINE_STOP);
1841 }
1842 }
1843
1844 return (vd->vdev_ops->vdev_op_io_start(zio));
1845 }
1846
1847 static int
1848 zio_vdev_io_done(zio_t *zio)
1849 {
1850 vdev_t *vd = zio->io_vd;
1851 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
1852 boolean_t unexpected_error = B_FALSE;
1853
1854 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
1855 return (ZIO_PIPELINE_STOP);
1856
1857 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
1858
1859 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
1860
1861 vdev_queue_io_done(zio);
1862
1863 if (zio->io_type == ZIO_TYPE_WRITE)
1864 vdev_cache_write(zio);
1865
1866 if (zio_injection_enabled && zio->io_error == 0)
1867 zio->io_error = zio_handle_device_injection(vd, EIO);
1868
1869 if (zio_injection_enabled && zio->io_error == 0)
1870 zio->io_error = zio_handle_label_injection(zio, EIO);
1871
1872 if (zio->io_error) {
1873 if (!vdev_accessible(vd, zio)) {
1874 zio->io_error = ENXIO;
1875 } else {
1876 unexpected_error = B_TRUE;
1877 }
1878 }
1879 }
1880
1881 ops->vdev_op_io_done(zio);
1882
1883 if (unexpected_error)
1884 VERIFY(vdev_probe(vd, zio) == NULL);
1885
1886 return (ZIO_PIPELINE_CONTINUE);
1887 }
1888
1889 static int
1890 zio_vdev_io_assess(zio_t *zio)
1891 {
1892 vdev_t *vd = zio->io_vd;
1893
1894 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
1895 return (ZIO_PIPELINE_STOP);
1896
1897 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
1898 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
1899
1900 if (zio->io_vsd != NULL) {
1901 zio->io_vsd_free(zio);
1902 zio->io_vsd = NULL;
1903 }
1904
1905 if (zio_injection_enabled && zio->io_error == 0)
1906 zio->io_error = zio_handle_fault_injection(zio, EIO);
1907
1908 /*
1909 * If the I/O failed, determine whether we should attempt to retry it.
1910 */
1911 if (zio->io_error && vd == NULL &&
1912 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
1913 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
1914 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
1915 zio->io_error = 0;
1916 zio->io_flags |= ZIO_FLAG_IO_RETRY |
1917 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
1918 zio->io_stage = ZIO_STAGE_VDEV_IO_START - 1;
1919 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE);
1920 return (ZIO_PIPELINE_STOP);
1921 }
1922
1923 /*
1924 * If we got an error on a leaf device, convert it to ENXIO
1925 * if the device is not accessible at all.
1926 */
1927 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
1928 !vdev_accessible(vd, zio))
1929 zio->io_error = ENXIO;
1930
1931 /*
1932 * If we can't write to an interior vdev (mirror or RAID-Z),
1933 * set vdev_cant_write so that we stop trying to allocate from it.
1934 */
1935 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
1936 vd != NULL && !vd->vdev_ops->vdev_op_leaf)
1937 vd->vdev_cant_write = B_TRUE;
1938
1939 if (zio->io_error)
1940 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1941
1942 return (ZIO_PIPELINE_CONTINUE);
1943 }
1944
1945 void
1946 zio_vdev_io_reissue(zio_t *zio)
1947 {
1948 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
1949 ASSERT(zio->io_error == 0);
1950
1951 zio->io_stage--;
1952 }
1953
1954 void
1955 zio_vdev_io_redone(zio_t *zio)
1956 {
1957 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
1958
1959 zio->io_stage--;
1960 }
1961
1962 void
1963 zio_vdev_io_bypass(zio_t *zio)
1964 {
1965 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
1966 ASSERT(zio->io_error == 0);
1967
1968 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
1969 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS - 1;
1970 }
1971
1972 /*
1973 * ==========================================================================
1974 * Generate and verify checksums
1975 * ==========================================================================
1976 */
1977 static int
1978 zio_checksum_generate(zio_t *zio)
1979 {
1980 blkptr_t *bp = zio->io_bp;
1981 enum zio_checksum checksum;
1982
1983 if (bp == NULL) {
1984 /*
1985 * This is zio_write_phys().
1986 * We're either generating a label checksum, or none at all.
1987 */
1988 checksum = zio->io_prop.zp_checksum;
1989
1990 if (checksum == ZIO_CHECKSUM_OFF)
1991 return (ZIO_PIPELINE_CONTINUE);
1992
1993 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
1994 } else {
1995 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
1996 ASSERT(!IO_IS_ALLOCATING(zio));
1997 checksum = ZIO_CHECKSUM_GANG_HEADER;
1998 } else {
1999 checksum = BP_GET_CHECKSUM(bp);
2000 }
2001 }
2002
2003 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2004
2005 return (ZIO_PIPELINE_CONTINUE);
2006 }
2007
2008 static int
2009 zio_checksum_verify(zio_t *zio)
2010 {
2011 blkptr_t *bp = zio->io_bp;
2012 int error;
2013
2014 if (bp == NULL) {
2015 /*
2016 * This is zio_read_phys().
2017 * We're either verifying a label checksum, or nothing at all.
2018 */
2019 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2020 return (ZIO_PIPELINE_CONTINUE);
2021
2022 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2023 }
2024
2025 if ((error = zio_checksum_error(zio)) != 0) {
2026 zio->io_error = error;
2027 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2028 zfs_ereport_post(FM_EREPORT_ZFS_CHECKSUM,
2029 zio->io_spa, zio->io_vd, zio, 0, 0);
2030 }
2031 }
2032
2033 return (ZIO_PIPELINE_CONTINUE);
2034 }
2035
2036 /*
2037 * Called by RAID-Z to ensure we don't compute the checksum twice.
2038 */
2039 void
2040 zio_checksum_verified(zio_t *zio)
2041 {
2042 zio->io_pipeline &= ~(1U << ZIO_STAGE_CHECKSUM_VERIFY);
2043 }
2044
2045 /*
2046 * ==========================================================================
2047 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2048 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2049 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2050 * indicate errors that are specific to one I/O, and most likely permanent.
2051 * Any other error is presumed to be worse because we weren't expecting it.
2052 * ==========================================================================
2053 */
2054 int
2055 zio_worst_error(int e1, int e2)
2056 {
2057 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2058 int r1, r2;
2059
2060 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2061 if (e1 == zio_error_rank[r1])
2062 break;
2063
2064 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2065 if (e2 == zio_error_rank[r2])
2066 break;
2067
2068 return (r1 > r2 ? e1 : e2);
2069 }
2070
2071 /*
2072 * ==========================================================================
2073 * I/O completion
2074 * ==========================================================================
2075 */
2076 static int
2077 zio_ready(zio_t *zio)
2078 {
2079 blkptr_t *bp = zio->io_bp;
2080 zio_t *pio, *pio_next;
2081
2082 if (zio->io_ready) {
2083 if (BP_IS_GANG(bp) &&
2084 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY))
2085 return (ZIO_PIPELINE_STOP);
2086
2087 ASSERT(IO_IS_ALLOCATING(zio));
2088 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2089 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2090
2091 zio->io_ready(zio);
2092 }
2093
2094 if (bp != NULL && bp != &zio->io_bp_copy)
2095 zio->io_bp_copy = *bp;
2096
2097 if (zio->io_error)
2098 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2099
2100 mutex_enter(&zio->io_lock);
2101 zio->io_state[ZIO_WAIT_READY] = 1;
2102 pio = zio_walk_parents(zio);
2103 mutex_exit(&zio->io_lock);
2104
2105 /*
2106 * As we notify zio's parents, new parents could be added.
2107 * New parents go to the head of zio's io_parent_list, however,
2108 * so we will (correctly) not notify them. The remainder of zio's
2109 * io_parent_list, from 'pio_next' onward, cannot change because
2110 * all parents must wait for us to be done before they can be done.
2111 */
2112 for (; pio != NULL; pio = pio_next) {
2113 pio_next = zio_walk_parents(zio);
2114 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2115 }
2116
2117 return (ZIO_PIPELINE_CONTINUE);
2118 }
2119
2120 static int
2121 zio_done(zio_t *zio)
2122 {
2123 spa_t *spa = zio->io_spa;
2124 zio_t *lio = zio->io_logical;
2125 blkptr_t *bp = zio->io_bp;
2126 vdev_t *vd = zio->io_vd;
2127 uint64_t psize = zio->io_size;
2128 zio_t *pio, *pio_next;
2129
2130 /*
2131 * If our of children haven't all completed,
2132 * wait for them and then repeat this pipeline stage.
2133 */
2134 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2135 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2136 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2137 return (ZIO_PIPELINE_STOP);
2138
2139 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2140 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2141 ASSERT(zio->io_children[c][w] == 0);
2142
2143 if (bp != NULL) {
2144 ASSERT(bp->blk_pad[0] == 0);
2145 ASSERT(bp->blk_pad[1] == 0);
2146 ASSERT(bp->blk_pad[2] == 0);
2147 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2148 (bp == zio_unique_parent(zio)->io_bp));
2149 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2150 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2151 ASSERT(!BP_SHOULD_BYTESWAP(bp));
2152 ASSERT3U(zio->io_prop.zp_ndvas, <=, BP_GET_NDVAS(bp));
2153 ASSERT(BP_COUNT_GANG(bp) == 0 ||
2154 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2155 }
2156 }
2157
2158 /*
2159 * If there were child vdev or gang errors, they apply to us now.
2160 */
2161 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2162 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2163
2164 zio_pop_transforms(zio); /* note: may set zio->io_error */
2165
2166 vdev_stat_update(zio, psize);
2167
2168 if (zio->io_error) {
2169 /*
2170 * If this I/O is attached to a particular vdev,
2171 * generate an error message describing the I/O failure
2172 * at the block level. We ignore these errors if the
2173 * device is currently unavailable.
2174 */
2175 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
2176 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
2177
2178 if ((zio->io_error == EIO ||
2179 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) && zio == lio) {
2180 /*
2181 * For logical I/O requests, tell the SPA to log the
2182 * error and generate a logical data ereport.
2183 */
2184 spa_log_error(spa, zio);
2185 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
2186 0, 0);
2187 }
2188 }
2189
2190 if (zio->io_error && zio == lio) {
2191 /*
2192 * Determine whether zio should be reexecuted. This will
2193 * propagate all the way to the root via zio_notify_parent().
2194 */
2195 ASSERT(vd == NULL && bp != NULL);
2196
2197 if (IO_IS_ALLOCATING(zio))
2198 if (zio->io_error != ENOSPC)
2199 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2200 else
2201 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2202
2203 if ((zio->io_type == ZIO_TYPE_READ ||
2204 zio->io_type == ZIO_TYPE_FREE) &&
2205 zio->io_error == ENXIO &&
2206 spa->spa_load_state == SPA_LOAD_NONE &&
2207 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
2208 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2209
2210 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2211 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2212 }
2213
2214 /*
2215 * If there were logical child errors, they apply to us now.
2216 * We defer this until now to avoid conflating logical child
2217 * errors with errors that happened to the zio itself when
2218 * updating vdev stats and reporting FMA events above.
2219 */
2220 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2221
2222 if (zio->io_reexecute) {
2223 /*
2224 * This is a logical I/O that wants to reexecute.
2225 *
2226 * Reexecute is top-down. When an i/o fails, if it's not
2227 * the root, it simply notifies its parent and sticks around.
2228 * The parent, seeing that it still has children in zio_done(),
2229 * does the same. This percolates all the way up to the root.
2230 * The root i/o will reexecute or suspend the entire tree.
2231 *
2232 * This approach ensures that zio_reexecute() honors
2233 * all the original i/o dependency relationships, e.g.
2234 * parents not executing until children are ready.
2235 */
2236 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2237
2238 if (IO_IS_ALLOCATING(zio))
2239 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
2240
2241 zio_gang_tree_free(&zio->io_gang_tree);
2242
2243 mutex_enter(&zio->io_lock);
2244 zio->io_state[ZIO_WAIT_DONE] = 1;
2245 mutex_exit(&zio->io_lock);
2246
2247 if ((pio = zio_unique_parent(zio)) != NULL) {
2248 /*
2249 * We're not a root i/o, so there's nothing to do
2250 * but notify our parent. Don't propagate errors
2251 * upward since we haven't permanently failed yet.
2252 */
2253 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
2254 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2255 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
2256 /*
2257 * We'd fail again if we reexecuted now, so suspend
2258 * until conditions improve (e.g. device comes online).
2259 */
2260 zio_suspend(spa, zio);
2261 } else {
2262 /*
2263 * Reexecution is potentially a huge amount of work.
2264 * Hand it off to the otherwise-unused claim taskq.
2265 */
2266 (void) taskq_dispatch(
2267 spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2268 (task_func_t *)zio_reexecute, zio, TQ_SLEEP);
2269 }
2270 return (ZIO_PIPELINE_STOP);
2271 }
2272
2273 ASSERT(zio_walk_children(zio) == NULL);
2274 ASSERT(zio->io_reexecute == 0);
2275 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
2276
2277 /*
2278 * It is the responsibility of the done callback to ensure that this
2279 * particular zio is no longer discoverable for adoption, and as
2280 * such, cannot acquire any new parents.
2281 */
2282 if (zio->io_done)
2283 zio->io_done(zio);
2284
2285 zio_gang_tree_free(&zio->io_gang_tree);
2286
2287 mutex_enter(&zio->io_lock);
2288 zio->io_state[ZIO_WAIT_DONE] = 1;
2289 mutex_exit(&zio->io_lock);
2290
2291 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2292 zio_link_t *zl = zio->io_walk_link;
2293 pio_next = zio_walk_parents(zio);
2294 zio_remove_child(pio, zio, zl);
2295 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2296 }
2297
2298 if (zio->io_waiter != NULL) {
2299 mutex_enter(&zio->io_lock);
2300 zio->io_executor = NULL;
2301 cv_broadcast(&zio->io_cv);
2302 mutex_exit(&zio->io_lock);
2303 } else {
2304 zio_destroy(zio);
2305 }
2306
2307 return (ZIO_PIPELINE_STOP);
2308 }
2309
2310 /*
2311 * ==========================================================================
2312 * I/O pipeline definition
2313 * ==========================================================================
2314 */
2315 static zio_pipe_stage_t *zio_pipeline[ZIO_STAGES] = {
2316 NULL,
2317 zio_issue_async,
2318 zio_read_bp_init,
2319 zio_write_bp_init,
2320 zio_checksum_generate,
2321 zio_gang_assemble,
2322 zio_gang_issue,
2323 zio_dva_allocate,
2324 zio_dva_free,
2325 zio_dva_claim,
2326 zio_ready,
2327 zio_vdev_io_start,
2328 zio_vdev_io_done,
2329 zio_vdev_io_assess,
2330 zio_checksum_verify,
2331 zio_done
2332 };
ZFS On Linux mirror for PVE