]> 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 b117
[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 zio->io_logical = zio;
495 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
496 pipeline |= ZIO_GANG_STAGES;
497 }
498
499 zio->io_spa = spa;
500 zio->io_txg = txg;
501 zio->io_data = data;
502 zio->io_size = size;
503 zio->io_done = done;
504 zio->io_private = private;
505 zio->io_type = type;
506 zio->io_priority = priority;
507 zio->io_vd = vd;
508 zio->io_offset = offset;
509 zio->io_orig_flags = zio->io_flags = flags;
510 zio->io_orig_stage = zio->io_stage = stage;
511 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
512
513 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
514 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
515
516 if (zb != NULL)
517 zio->io_bookmark = *zb;
518
519 if (pio != NULL) {
520 if (zio->io_logical == NULL)
521 zio->io_logical = pio->io_logical;
522 if (zio->io_child_type == ZIO_CHILD_GANG)
523 zio->io_gang_leader = pio->io_gang_leader;
524 zio_add_child(pio, zio);
525 }
526
527 return (zio);
528 }
529
530 static void
531 zio_destroy(zio_t *zio)
532 {
533 list_destroy(&zio->io_parent_list);
534 list_destroy(&zio->io_child_list);
535 mutex_destroy(&zio->io_lock);
536 cv_destroy(&zio->io_cv);
537 kmem_cache_free(zio_cache, zio);
538 }
539
540 zio_t *
541 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
542 void *private, int flags)
543 {
544 zio_t *zio;
545
546 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
547 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
548 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
549
550 return (zio);
551 }
552
553 zio_t *
554 zio_root(spa_t *spa, zio_done_func_t *done, void *private, int flags)
555 {
556 return (zio_null(NULL, spa, NULL, done, private, flags));
557 }
558
559 zio_t *
560 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
561 void *data, uint64_t size, zio_done_func_t *done, void *private,
562 int priority, int flags, const zbookmark_t *zb)
563 {
564 zio_t *zio;
565
566 zio = zio_create(pio, spa, bp->blk_birth, (blkptr_t *)bp,
567 data, size, done, private,
568 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
569 ZIO_STAGE_OPEN, ZIO_READ_PIPELINE);
570
571 return (zio);
572 }
573
574 void
575 zio_skip_write(zio_t *zio)
576 {
577 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
578 ASSERT(zio->io_stage == ZIO_STAGE_READY);
579 ASSERT(!BP_IS_GANG(zio->io_bp));
580
581 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
582 }
583
584 zio_t *
585 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
586 void *data, uint64_t size, zio_prop_t *zp,
587 zio_done_func_t *ready, zio_done_func_t *done, void *private,
588 int priority, int flags, const zbookmark_t *zb)
589 {
590 zio_t *zio;
591
592 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
593 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
594 zp->zp_compress >= ZIO_COMPRESS_OFF &&
595 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
596 zp->zp_type < DMU_OT_NUMTYPES &&
597 zp->zp_level < 32 &&
598 zp->zp_ndvas > 0 &&
599 zp->zp_ndvas <= spa_max_replication(spa));
600 ASSERT(ready != NULL);
601
602 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
603 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
604 ZIO_STAGE_OPEN, ZIO_WRITE_PIPELINE);
605
606 zio->io_ready = ready;
607 zio->io_prop = *zp;
608
609 return (zio);
610 }
611
612 zio_t *
613 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
614 uint64_t size, zio_done_func_t *done, void *private, int priority,
615 int flags, zbookmark_t *zb)
616 {
617 zio_t *zio;
618
619 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
620 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
621 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
622
623 return (zio);
624 }
625
626 zio_t *
627 zio_free(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
628 zio_done_func_t *done, void *private, int flags)
629 {
630 zio_t *zio;
631
632 ASSERT(!BP_IS_HOLE(bp));
633
634 if (bp->blk_fill == BLK_FILL_ALREADY_FREED)
635 return (zio_null(pio, spa, NULL, NULL, NULL, flags));
636
637 if (txg == spa->spa_syncing_txg &&
638 spa_sync_pass(spa) > SYNC_PASS_DEFERRED_FREE) {
639 bplist_enqueue_deferred(&spa->spa_sync_bplist, bp);
640 return (zio_null(pio, spa, NULL, NULL, NULL, flags));
641 }
642
643 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
644 done, private, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
645 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
646
647 return (zio);
648 }
649
650 zio_t *
651 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
652 zio_done_func_t *done, void *private, int flags)
653 {
654 zio_t *zio;
655
656 /*
657 * A claim is an allocation of a specific block. Claims are needed
658 * to support immediate writes in the intent log. The issue is that
659 * immediate writes contain committed data, but in a txg that was
660 * *not* committed. Upon opening the pool after an unclean shutdown,
661 * the intent log claims all blocks that contain immediate write data
662 * so that the SPA knows they're in use.
663 *
664 * All claims *must* be resolved in the first txg -- before the SPA
665 * starts allocating blocks -- so that nothing is allocated twice.
666 */
667 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
668 ASSERT3U(spa_first_txg(spa), <=, txg);
669
670 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
671 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
672 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
673
674 return (zio);
675 }
676
677 zio_t *
678 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
679 zio_done_func_t *done, void *private, int priority, int flags)
680 {
681 zio_t *zio;
682 int c;
683
684 if (vd->vdev_children == 0) {
685 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
686 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
687 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
688
689 zio->io_cmd = cmd;
690 } else {
691 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
692
693 for (c = 0; c < vd->vdev_children; c++)
694 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
695 done, private, priority, flags));
696 }
697
698 return (zio);
699 }
700
701 zio_t *
702 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
703 void *data, int checksum, zio_done_func_t *done, void *private,
704 int priority, int flags, boolean_t labels)
705 {
706 zio_t *zio;
707
708 ASSERT(vd->vdev_children == 0);
709 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
710 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
711 ASSERT3U(offset + size, <=, vd->vdev_psize);
712
713 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
714 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
715 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
716
717 zio->io_prop.zp_checksum = checksum;
718
719 return (zio);
720 }
721
722 zio_t *
723 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
724 void *data, int checksum, zio_done_func_t *done, void *private,
725 int priority, int flags, boolean_t labels)
726 {
727 zio_t *zio;
728
729 ASSERT(vd->vdev_children == 0);
730 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
731 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
732 ASSERT3U(offset + size, <=, vd->vdev_psize);
733
734 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
735 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
736 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
737
738 zio->io_prop.zp_checksum = checksum;
739
740 if (zio_checksum_table[checksum].ci_zbt) {
741 /*
742 * zbt checksums are necessarily destructive -- they modify
743 * the end of the write buffer to hold the verifier/checksum.
744 * Therefore, we must make a local copy in case the data is
745 * being written to multiple places in parallel.
746 */
747 void *wbuf = zio_buf_alloc(size);
748 bcopy(data, wbuf, size);
749 zio_push_transform(zio, wbuf, size, size, NULL);
750 }
751
752 return (zio);
753 }
754
755 /*
756 * Create a child I/O to do some work for us.
757 */
758 zio_t *
759 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
760 void *data, uint64_t size, int type, int priority, int flags,
761 zio_done_func_t *done, void *private)
762 {
763 uint32_t pipeline = ZIO_VDEV_CHILD_PIPELINE;
764 zio_t *zio;
765
766 ASSERT(vd->vdev_parent ==
767 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
768
769 if (type == ZIO_TYPE_READ && bp != NULL) {
770 /*
771 * If we have the bp, then the child should perform the
772 * checksum and the parent need not. This pushes error
773 * detection as close to the leaves as possible and
774 * eliminates redundant checksums in the interior nodes.
775 */
776 pipeline |= 1U << ZIO_STAGE_CHECKSUM_VERIFY;
777 pio->io_pipeline &= ~(1U << ZIO_STAGE_CHECKSUM_VERIFY);
778 }
779
780 if (vd->vdev_children == 0)
781 offset += VDEV_LABEL_START_SIZE;
782
783 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
784 done, private, type, priority,
785 (pio->io_flags & ZIO_FLAG_VDEV_INHERIT) |
786 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | flags,
787 vd, offset, &pio->io_bookmark,
788 ZIO_STAGE_VDEV_IO_START - 1, pipeline);
789
790 return (zio);
791 }
792
793 zio_t *
794 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
795 int type, int priority, int flags, zio_done_func_t *done, void *private)
796 {
797 zio_t *zio;
798
799 ASSERT(vd->vdev_ops->vdev_op_leaf);
800
801 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
802 data, size, done, private, type, priority,
803 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
804 vd, offset, NULL,
805 ZIO_STAGE_VDEV_IO_START - 1, ZIO_VDEV_CHILD_PIPELINE);
806
807 return (zio);
808 }
809
810 void
811 zio_flush(zio_t *zio, vdev_t *vd)
812 {
813 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
814 NULL, NULL, ZIO_PRIORITY_NOW,
815 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
816 }
817
818 /*
819 * ==========================================================================
820 * Prepare to read and write logical blocks
821 * ==========================================================================
822 */
823
824 static int
825 zio_read_bp_init(zio_t *zio)
826 {
827 blkptr_t *bp = zio->io_bp;
828
829 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
830 zio->io_child_type == ZIO_CHILD_LOGICAL &&
831 !(zio->io_flags & ZIO_FLAG_RAW)) {
832 uint64_t csize = BP_GET_PSIZE(bp);
833 void *cbuf = zio_buf_alloc(csize);
834
835 zio_push_transform(zio, cbuf, csize, csize, zio_decompress);
836 }
837
838 if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
839 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
840
841 return (ZIO_PIPELINE_CONTINUE);
842 }
843
844 static int
845 zio_write_bp_init(zio_t *zio)
846 {
847 zio_prop_t *zp = &zio->io_prop;
848 int compress = zp->zp_compress;
849 blkptr_t *bp = zio->io_bp;
850 void *cbuf;
851 uint64_t lsize = zio->io_size;
852 uint64_t csize = lsize;
853 uint64_t cbufsize = 0;
854 int pass = 1;
855
856 /*
857 * If our children haven't all reached the ready stage,
858 * wait for them and then repeat this pipeline stage.
859 */
860 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
861 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
862 return (ZIO_PIPELINE_STOP);
863
864 if (!IO_IS_ALLOCATING(zio))
865 return (ZIO_PIPELINE_CONTINUE);
866
867 ASSERT(compress != ZIO_COMPRESS_INHERIT);
868
869 if (bp->blk_birth == zio->io_txg) {
870 /*
871 * We're rewriting an existing block, which means we're
872 * working on behalf of spa_sync(). For spa_sync() to
873 * converge, it must eventually be the case that we don't
874 * have to allocate new blocks. But compression changes
875 * the blocksize, which forces a reallocate, and makes
876 * convergence take longer. Therefore, after the first
877 * few passes, stop compressing to ensure convergence.
878 */
879 pass = spa_sync_pass(zio->io_spa);
880
881 if (pass > SYNC_PASS_DONT_COMPRESS)
882 compress = ZIO_COMPRESS_OFF;
883
884 /* Make sure someone doesn't change their mind on overwrites */
885 ASSERT(MIN(zp->zp_ndvas + BP_IS_GANG(bp),
886 spa_max_replication(zio->io_spa)) == BP_GET_NDVAS(bp));
887 }
888
889 if (compress != ZIO_COMPRESS_OFF) {
890 if (!zio_compress_data(compress, zio->io_data, zio->io_size,
891 &cbuf, &csize, &cbufsize)) {
892 compress = ZIO_COMPRESS_OFF;
893 } else if (csize != 0) {
894 zio_push_transform(zio, cbuf, csize, cbufsize, NULL);
895 }
896 }
897
898 /*
899 * The final pass of spa_sync() must be all rewrites, but the first
900 * few passes offer a trade-off: allocating blocks defers convergence,
901 * but newly allocated blocks are sequential, so they can be written
902 * to disk faster. Therefore, we allow the first few passes of
903 * spa_sync() to allocate new blocks, but force rewrites after that.
904 * There should only be a handful of blocks after pass 1 in any case.
905 */
906 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == csize &&
907 pass > SYNC_PASS_REWRITE) {
908 ASSERT(csize != 0);
909 uint32_t gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
910 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
911 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
912 } else {
913 BP_ZERO(bp);
914 zio->io_pipeline = ZIO_WRITE_PIPELINE;
915 }
916
917 if (csize == 0) {
918 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
919 } else {
920 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
921 BP_SET_LSIZE(bp, lsize);
922 BP_SET_PSIZE(bp, csize);
923 BP_SET_COMPRESS(bp, compress);
924 BP_SET_CHECKSUM(bp, zp->zp_checksum);
925 BP_SET_TYPE(bp, zp->zp_type);
926 BP_SET_LEVEL(bp, zp->zp_level);
927 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
928 }
929
930 return (ZIO_PIPELINE_CONTINUE);
931 }
932
933 /*
934 * ==========================================================================
935 * Execute the I/O pipeline
936 * ==========================================================================
937 */
938
939 static void
940 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q)
941 {
942 zio_type_t t = zio->io_type;
943
944 /*
945 * If we're a config writer or a probe, the normal issue and
946 * interrupt threads may all be blocked waiting for the config lock.
947 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
948 */
949 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
950 t = ZIO_TYPE_NULL;
951
952 /*
953 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
954 */
955 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
956 t = ZIO_TYPE_NULL;
957
958 (void) taskq_dispatch(zio->io_spa->spa_zio_taskq[t][q],
959 (task_func_t *)zio_execute, zio, TQ_SLEEP);
960 }
961
962 static boolean_t
963 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
964 {
965 kthread_t *executor = zio->io_executor;
966 spa_t *spa = zio->io_spa;
967
968 for (zio_type_t t = 0; t < ZIO_TYPES; t++)
969 if (taskq_member(spa->spa_zio_taskq[t][q], executor))
970 return (B_TRUE);
971
972 return (B_FALSE);
973 }
974
975 static int
976 zio_issue_async(zio_t *zio)
977 {
978 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE);
979
980 return (ZIO_PIPELINE_STOP);
981 }
982
983 void
984 zio_interrupt(zio_t *zio)
985 {
986 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT);
987 }
988
989 /*
990 * Execute the I/O pipeline until one of the following occurs:
991 * (1) the I/O completes; (2) the pipeline stalls waiting for
992 * dependent child I/Os; (3) the I/O issues, so we're waiting
993 * for an I/O completion interrupt; (4) the I/O is delegated by
994 * vdev-level caching or aggregation; (5) the I/O is deferred
995 * due to vdev-level queueing; (6) the I/O is handed off to
996 * another thread. In all cases, the pipeline stops whenever
997 * there's no CPU work; it never burns a thread in cv_wait().
998 *
999 * There's no locking on io_stage because there's no legitimate way
1000 * for multiple threads to be attempting to process the same I/O.
1001 */
1002 static zio_pipe_stage_t *zio_pipeline[ZIO_STAGES];
1003
1004 void
1005 zio_execute(zio_t *zio)
1006 {
1007 zio->io_executor = curthread;
1008
1009 while (zio->io_stage < ZIO_STAGE_DONE) {
1010 uint32_t pipeline = zio->io_pipeline;
1011 zio_stage_t stage = zio->io_stage;
1012 int rv;
1013
1014 ASSERT(!MUTEX_HELD(&zio->io_lock));
1015
1016 while (((1U << ++stage) & pipeline) == 0)
1017 continue;
1018
1019 ASSERT(stage <= ZIO_STAGE_DONE);
1020 ASSERT(zio->io_stall == NULL);
1021
1022 /*
1023 * If we are in interrupt context and this pipeline stage
1024 * will grab a config lock that is held across I/O,
1025 * issue async to avoid deadlock.
1026 */
1027 if (((1U << stage) & ZIO_CONFIG_LOCK_BLOCKING_STAGES) &&
1028 zio->io_vd == NULL &&
1029 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1030 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE);
1031 return;
1032 }
1033
1034 zio->io_stage = stage;
1035 rv = zio_pipeline[stage](zio);
1036
1037 if (rv == ZIO_PIPELINE_STOP)
1038 return;
1039
1040 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1041 }
1042 }
1043
1044 /*
1045 * ==========================================================================
1046 * Initiate I/O, either sync or async
1047 * ==========================================================================
1048 */
1049 int
1050 zio_wait(zio_t *zio)
1051 {
1052 int error;
1053
1054 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1055 ASSERT(zio->io_executor == NULL);
1056
1057 zio->io_waiter = curthread;
1058
1059 zio_execute(zio);
1060
1061 mutex_enter(&zio->io_lock);
1062 while (zio->io_executor != NULL)
1063 cv_wait(&zio->io_cv, &zio->io_lock);
1064 mutex_exit(&zio->io_lock);
1065
1066 error = zio->io_error;
1067 zio_destroy(zio);
1068
1069 return (error);
1070 }
1071
1072 void
1073 zio_nowait(zio_t *zio)
1074 {
1075 ASSERT(zio->io_executor == NULL);
1076
1077 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1078 zio_unique_parent(zio) == NULL) {
1079 /*
1080 * This is a logical async I/O with no parent to wait for it.
1081 * We add it to the spa_async_root_zio "Godfather" I/O which
1082 * will ensure they complete prior to unloading the pool.
1083 */
1084 spa_t *spa = zio->io_spa;
1085
1086 zio_add_child(spa->spa_async_zio_root, zio);
1087 }
1088
1089 zio_execute(zio);
1090 }
1091
1092 /*
1093 * ==========================================================================
1094 * Reexecute or suspend/resume failed I/O
1095 * ==========================================================================
1096 */
1097
1098 static void
1099 zio_reexecute(zio_t *pio)
1100 {
1101 zio_t *cio, *cio_next;
1102
1103 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1104 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1105 ASSERT(pio->io_gang_leader == NULL);
1106 ASSERT(pio->io_gang_tree == NULL);
1107
1108 pio->io_flags = pio->io_orig_flags;
1109 pio->io_stage = pio->io_orig_stage;
1110 pio->io_pipeline = pio->io_orig_pipeline;
1111 pio->io_reexecute = 0;
1112 pio->io_error = 0;
1113 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1114 pio->io_state[w] = 0;
1115 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1116 pio->io_child_error[c] = 0;
1117
1118 if (IO_IS_ALLOCATING(pio)) {
1119 /*
1120 * Remember the failed bp so that the io_ready() callback
1121 * can update its accounting upon reexecution. The block
1122 * was already freed in zio_done(); we indicate this with
1123 * a fill count of -1 so that zio_free() knows to skip it.
1124 */
1125 blkptr_t *bp = pio->io_bp;
1126 ASSERT(bp->blk_birth == 0 || bp->blk_birth == pio->io_txg);
1127 bp->blk_fill = BLK_FILL_ALREADY_FREED;
1128 pio->io_bp_orig = *bp;
1129 BP_ZERO(bp);
1130 }
1131
1132 /*
1133 * As we reexecute pio's children, new children could be created.
1134 * New children go to the head of pio's io_child_list, however,
1135 * so we will (correctly) not reexecute them. The key is that
1136 * the remainder of pio's io_child_list, from 'cio_next' onward,
1137 * cannot be affected by any side effects of reexecuting 'cio'.
1138 */
1139 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1140 cio_next = zio_walk_children(pio);
1141 mutex_enter(&pio->io_lock);
1142 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1143 pio->io_children[cio->io_child_type][w]++;
1144 mutex_exit(&pio->io_lock);
1145 zio_reexecute(cio);
1146 }
1147
1148 /*
1149 * Now that all children have been reexecuted, execute the parent.
1150 * We don't reexecute "The Godfather" I/O here as it's the
1151 * responsibility of the caller to wait on him.
1152 */
1153 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1154 zio_execute(pio);
1155 }
1156
1157 void
1158 zio_suspend(spa_t *spa, zio_t *zio)
1159 {
1160 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1161 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1162 "failure and the failure mode property for this pool "
1163 "is set to panic.", spa_name(spa));
1164
1165 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1166
1167 mutex_enter(&spa->spa_suspend_lock);
1168
1169 if (spa->spa_suspend_zio_root == NULL)
1170 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1171 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1172 ZIO_FLAG_GODFATHER);
1173
1174 spa->spa_suspended = B_TRUE;
1175
1176 if (zio != NULL) {
1177 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1178 ASSERT(zio != spa->spa_suspend_zio_root);
1179 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1180 ASSERT(zio_unique_parent(zio) == NULL);
1181 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1182 zio_add_child(spa->spa_suspend_zio_root, zio);
1183 }
1184
1185 mutex_exit(&spa->spa_suspend_lock);
1186 }
1187
1188 int
1189 zio_resume(spa_t *spa)
1190 {
1191 zio_t *pio;
1192
1193 /*
1194 * Reexecute all previously suspended i/o.
1195 */
1196 mutex_enter(&spa->spa_suspend_lock);
1197 spa->spa_suspended = B_FALSE;
1198 cv_broadcast(&spa->spa_suspend_cv);
1199 pio = spa->spa_suspend_zio_root;
1200 spa->spa_suspend_zio_root = NULL;
1201 mutex_exit(&spa->spa_suspend_lock);
1202
1203 if (pio == NULL)
1204 return (0);
1205
1206 zio_reexecute(pio);
1207 return (zio_wait(pio));
1208 }
1209
1210 void
1211 zio_resume_wait(spa_t *spa)
1212 {
1213 mutex_enter(&spa->spa_suspend_lock);
1214 while (spa_suspended(spa))
1215 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1216 mutex_exit(&spa->spa_suspend_lock);
1217 }
1218
1219 /*
1220 * ==========================================================================
1221 * Gang blocks.
1222 *
1223 * A gang block is a collection of small blocks that looks to the DMU
1224 * like one large block. When zio_dva_allocate() cannot find a block
1225 * of the requested size, due to either severe fragmentation or the pool
1226 * being nearly full, it calls zio_write_gang_block() to construct the
1227 * block from smaller fragments.
1228 *
1229 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1230 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1231 * an indirect block: it's an array of block pointers. It consumes
1232 * only one sector and hence is allocatable regardless of fragmentation.
1233 * The gang header's bps point to its gang members, which hold the data.
1234 *
1235 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1236 * as the verifier to ensure uniqueness of the SHA256 checksum.
1237 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1238 * not the gang header. This ensures that data block signatures (needed for
1239 * deduplication) are independent of how the block is physically stored.
1240 *
1241 * Gang blocks can be nested: a gang member may itself be a gang block.
1242 * Thus every gang block is a tree in which root and all interior nodes are
1243 * gang headers, and the leaves are normal blocks that contain user data.
1244 * The root of the gang tree is called the gang leader.
1245 *
1246 * To perform any operation (read, rewrite, free, claim) on a gang block,
1247 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1248 * in the io_gang_tree field of the original logical i/o by recursively
1249 * reading the gang leader and all gang headers below it. This yields
1250 * an in-core tree containing the contents of every gang header and the
1251 * bps for every constituent of the gang block.
1252 *
1253 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1254 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1255 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1256 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1257 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1258 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1259 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1260 * of the gang header plus zio_checksum_compute() of the data to update the
1261 * gang header's blk_cksum as described above.
1262 *
1263 * The two-phase assemble/issue model solves the problem of partial failure --
1264 * what if you'd freed part of a gang block but then couldn't read the
1265 * gang header for another part? Assembling the entire gang tree first
1266 * ensures that all the necessary gang header I/O has succeeded before
1267 * starting the actual work of free, claim, or write. Once the gang tree
1268 * is assembled, free and claim are in-memory operations that cannot fail.
1269 *
1270 * In the event that a gang write fails, zio_dva_unallocate() walks the
1271 * gang tree to immediately free (i.e. insert back into the space map)
1272 * everything we've allocated. This ensures that we don't get ENOSPC
1273 * errors during repeated suspend/resume cycles due to a flaky device.
1274 *
1275 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1276 * the gang tree, we won't modify the block, so we can safely defer the free
1277 * (knowing that the block is still intact). If we *can* assemble the gang
1278 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1279 * each constituent bp and we can allocate a new block on the next sync pass.
1280 *
1281 * In all cases, the gang tree allows complete recovery from partial failure.
1282 * ==========================================================================
1283 */
1284
1285 static zio_t *
1286 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1287 {
1288 if (gn != NULL)
1289 return (pio);
1290
1291 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1292 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1293 &pio->io_bookmark));
1294 }
1295
1296 zio_t *
1297 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1298 {
1299 zio_t *zio;
1300
1301 if (gn != NULL) {
1302 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1303 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1304 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1305 /*
1306 * As we rewrite each gang header, the pipeline will compute
1307 * a new gang block header checksum for it; but no one will
1308 * compute a new data checksum, so we do that here. The one
1309 * exception is the gang leader: the pipeline already computed
1310 * its data checksum because that stage precedes gang assembly.
1311 * (Presently, nothing actually uses interior data checksums;
1312 * this is just good hygiene.)
1313 */
1314 if (gn != pio->io_gang_leader->io_gang_tree) {
1315 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1316 data, BP_GET_PSIZE(bp));
1317 }
1318 } else {
1319 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1320 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1321 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1322 }
1323
1324 return (zio);
1325 }
1326
1327 /* ARGSUSED */
1328 zio_t *
1329 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1330 {
1331 return (zio_free(pio, pio->io_spa, pio->io_txg, bp,
1332 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1333 }
1334
1335 /* ARGSUSED */
1336 zio_t *
1337 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1338 {
1339 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1340 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1341 }
1342
1343 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1344 NULL,
1345 zio_read_gang,
1346 zio_rewrite_gang,
1347 zio_free_gang,
1348 zio_claim_gang,
1349 NULL
1350 };
1351
1352 static void zio_gang_tree_assemble_done(zio_t *zio);
1353
1354 static zio_gang_node_t *
1355 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1356 {
1357 zio_gang_node_t *gn;
1358
1359 ASSERT(*gnpp == NULL);
1360
1361 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1362 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1363 *gnpp = gn;
1364
1365 return (gn);
1366 }
1367
1368 static void
1369 zio_gang_node_free(zio_gang_node_t **gnpp)
1370 {
1371 zio_gang_node_t *gn = *gnpp;
1372
1373 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1374 ASSERT(gn->gn_child[g] == NULL);
1375
1376 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1377 kmem_free(gn, sizeof (*gn));
1378 *gnpp = NULL;
1379 }
1380
1381 static void
1382 zio_gang_tree_free(zio_gang_node_t **gnpp)
1383 {
1384 zio_gang_node_t *gn = *gnpp;
1385
1386 if (gn == NULL)
1387 return;
1388
1389 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1390 zio_gang_tree_free(&gn->gn_child[g]);
1391
1392 zio_gang_node_free(gnpp);
1393 }
1394
1395 static void
1396 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1397 {
1398 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1399
1400 ASSERT(gio->io_gang_leader == gio);
1401 ASSERT(BP_IS_GANG(bp));
1402
1403 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1404 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1405 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1406 }
1407
1408 static void
1409 zio_gang_tree_assemble_done(zio_t *zio)
1410 {
1411 zio_t *gio = zio->io_gang_leader;
1412 zio_gang_node_t *gn = zio->io_private;
1413 blkptr_t *bp = zio->io_bp;
1414
1415 ASSERT(gio == zio_unique_parent(zio));
1416 ASSERT(zio_walk_children(zio) == NULL);
1417
1418 if (zio->io_error)
1419 return;
1420
1421 if (BP_SHOULD_BYTESWAP(bp))
1422 byteswap_uint64_array(zio->io_data, zio->io_size);
1423
1424 ASSERT(zio->io_data == gn->gn_gbh);
1425 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1426 ASSERT(gn->gn_gbh->zg_tail.zbt_magic == ZBT_MAGIC);
1427
1428 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1429 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1430 if (!BP_IS_GANG(gbp))
1431 continue;
1432 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1433 }
1434 }
1435
1436 static void
1437 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1438 {
1439 zio_t *gio = pio->io_gang_leader;
1440 zio_t *zio;
1441
1442 ASSERT(BP_IS_GANG(bp) == !!gn);
1443 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1444 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1445
1446 /*
1447 * If you're a gang header, your data is in gn->gn_gbh.
1448 * If you're a gang member, your data is in 'data' and gn == NULL.
1449 */
1450 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1451
1452 if (gn != NULL) {
1453 ASSERT(gn->gn_gbh->zg_tail.zbt_magic == ZBT_MAGIC);
1454
1455 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1456 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1457 if (BP_IS_HOLE(gbp))
1458 continue;
1459 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1460 data = (char *)data + BP_GET_PSIZE(gbp);
1461 }
1462 }
1463
1464 if (gn == gio->io_gang_tree)
1465 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1466
1467 if (zio != pio)
1468 zio_nowait(zio);
1469 }
1470
1471 static int
1472 zio_gang_assemble(zio_t *zio)
1473 {
1474 blkptr_t *bp = zio->io_bp;
1475
1476 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1477 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1478
1479 zio->io_gang_leader = zio;
1480
1481 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1482
1483 return (ZIO_PIPELINE_CONTINUE);
1484 }
1485
1486 static int
1487 zio_gang_issue(zio_t *zio)
1488 {
1489 blkptr_t *bp = zio->io_bp;
1490
1491 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1492 return (ZIO_PIPELINE_STOP);
1493
1494 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1495 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1496
1497 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1498 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1499 else
1500 zio_gang_tree_free(&zio->io_gang_tree);
1501
1502 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1503
1504 return (ZIO_PIPELINE_CONTINUE);
1505 }
1506
1507 static void
1508 zio_write_gang_member_ready(zio_t *zio)
1509 {
1510 zio_t *pio = zio_unique_parent(zio);
1511 zio_t *gio = zio->io_gang_leader;
1512 dva_t *cdva = zio->io_bp->blk_dva;
1513 dva_t *pdva = pio->io_bp->blk_dva;
1514 uint64_t asize;
1515
1516 if (BP_IS_HOLE(zio->io_bp))
1517 return;
1518
1519 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1520
1521 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1522 ASSERT3U(zio->io_prop.zp_ndvas, ==, gio->io_prop.zp_ndvas);
1523 ASSERT3U(zio->io_prop.zp_ndvas, <=, BP_GET_NDVAS(zio->io_bp));
1524 ASSERT3U(pio->io_prop.zp_ndvas, <=, BP_GET_NDVAS(pio->io_bp));
1525 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1526
1527 mutex_enter(&pio->io_lock);
1528 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1529 ASSERT(DVA_GET_GANG(&pdva[d]));
1530 asize = DVA_GET_ASIZE(&pdva[d]);
1531 asize += DVA_GET_ASIZE(&cdva[d]);
1532 DVA_SET_ASIZE(&pdva[d], asize);
1533 }
1534 mutex_exit(&pio->io_lock);
1535 }
1536
1537 static int
1538 zio_write_gang_block(zio_t *pio)
1539 {
1540 spa_t *spa = pio->io_spa;
1541 blkptr_t *bp = pio->io_bp;
1542 zio_t *gio = pio->io_gang_leader;
1543 zio_t *zio;
1544 zio_gang_node_t *gn, **gnpp;
1545 zio_gbh_phys_t *gbh;
1546 uint64_t txg = pio->io_txg;
1547 uint64_t resid = pio->io_size;
1548 uint64_t lsize;
1549 int ndvas = gio->io_prop.zp_ndvas;
1550 int gbh_ndvas = MIN(ndvas + 1, spa_max_replication(spa));
1551 zio_prop_t zp;
1552 int error;
1553
1554 error = metaslab_alloc(spa, spa->spa_normal_class, SPA_GANGBLOCKSIZE,
1555 bp, gbh_ndvas, txg, pio == gio ? NULL : gio->io_bp,
1556 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1557 if (error) {
1558 pio->io_error = error;
1559 return (ZIO_PIPELINE_CONTINUE);
1560 }
1561
1562 if (pio == gio) {
1563 gnpp = &gio->io_gang_tree;
1564 } else {
1565 gnpp = pio->io_private;
1566 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1567 }
1568
1569 gn = zio_gang_node_alloc(gnpp);
1570 gbh = gn->gn_gbh;
1571 bzero(gbh, SPA_GANGBLOCKSIZE);
1572
1573 /*
1574 * Create the gang header.
1575 */
1576 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1577 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1578
1579 /*
1580 * Create and nowait the gang children.
1581 */
1582 for (int g = 0; resid != 0; resid -= lsize, g++) {
1583 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1584 SPA_MINBLOCKSIZE);
1585 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1586
1587 zp.zp_checksum = gio->io_prop.zp_checksum;
1588 zp.zp_compress = ZIO_COMPRESS_OFF;
1589 zp.zp_type = DMU_OT_NONE;
1590 zp.zp_level = 0;
1591 zp.zp_ndvas = gio->io_prop.zp_ndvas;
1592
1593 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1594 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1595 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1596 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1597 &pio->io_bookmark));
1598 }
1599
1600 /*
1601 * Set pio's pipeline to just wait for zio to finish.
1602 */
1603 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1604
1605 zio_nowait(zio);
1606
1607 return (ZIO_PIPELINE_CONTINUE);
1608 }
1609
1610 /*
1611 * ==========================================================================
1612 * Allocate and free blocks
1613 * ==========================================================================
1614 */
1615
1616 static int
1617 zio_dva_allocate(zio_t *zio)
1618 {
1619 spa_t *spa = zio->io_spa;
1620 metaslab_class_t *mc = spa->spa_normal_class;
1621 blkptr_t *bp = zio->io_bp;
1622 int error;
1623
1624 if (zio->io_gang_leader == NULL) {
1625 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1626 zio->io_gang_leader = zio;
1627 }
1628
1629 ASSERT(BP_IS_HOLE(bp));
1630 ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
1631 ASSERT3U(zio->io_prop.zp_ndvas, >, 0);
1632 ASSERT3U(zio->io_prop.zp_ndvas, <=, spa_max_replication(spa));
1633 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
1634
1635 error = metaslab_alloc(spa, mc, zio->io_size, bp,
1636 zio->io_prop.zp_ndvas, zio->io_txg, NULL, 0);
1637
1638 if (error) {
1639 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
1640 return (zio_write_gang_block(zio));
1641 zio->io_error = error;
1642 }
1643
1644 return (ZIO_PIPELINE_CONTINUE);
1645 }
1646
1647 static int
1648 zio_dva_free(zio_t *zio)
1649 {
1650 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
1651
1652 return (ZIO_PIPELINE_CONTINUE);
1653 }
1654
1655 static int
1656 zio_dva_claim(zio_t *zio)
1657 {
1658 int error;
1659
1660 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
1661 if (error)
1662 zio->io_error = error;
1663
1664 return (ZIO_PIPELINE_CONTINUE);
1665 }
1666
1667 /*
1668 * Undo an allocation. This is used by zio_done() when an I/O fails
1669 * and we want to give back the block we just allocated.
1670 * This handles both normal blocks and gang blocks.
1671 */
1672 static void
1673 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
1674 {
1675 spa_t *spa = zio->io_spa;
1676 boolean_t now = !(zio->io_flags & ZIO_FLAG_IO_REWRITE);
1677
1678 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
1679
1680 if (zio->io_bp == bp && !now) {
1681 /*
1682 * This is a rewrite for sync-to-convergence.
1683 * We can't do a metaslab_free(NOW) because bp wasn't allocated
1684 * during this sync pass, which means that metaslab_sync()
1685 * already committed the allocation.
1686 */
1687 ASSERT(DVA_EQUAL(BP_IDENTITY(bp),
1688 BP_IDENTITY(&zio->io_bp_orig)));
1689 ASSERT(spa_sync_pass(spa) > 1);
1690
1691 if (BP_IS_GANG(bp) && gn == NULL) {
1692 /*
1693 * This is a gang leader whose gang header(s) we
1694 * couldn't read now, so defer the free until later.
1695 * The block should still be intact because without
1696 * the headers, we'd never even start the rewrite.
1697 */
1698 bplist_enqueue_deferred(&spa->spa_sync_bplist, bp);
1699 return;
1700 }
1701 }
1702
1703 if (!BP_IS_HOLE(bp))
1704 metaslab_free(spa, bp, bp->blk_birth, now);
1705
1706 if (gn != NULL) {
1707 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1708 zio_dva_unallocate(zio, gn->gn_child[g],
1709 &gn->gn_gbh->zg_blkptr[g]);
1710 }
1711 }
1712 }
1713
1714 /*
1715 * Try to allocate an intent log block. Return 0 on success, errno on failure.
1716 */
1717 int
1718 zio_alloc_blk(spa_t *spa, uint64_t size, blkptr_t *new_bp, blkptr_t *old_bp,
1719 uint64_t txg)
1720 {
1721 int error;
1722
1723 error = metaslab_alloc(spa, spa->spa_log_class, size,
1724 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
1725
1726 if (error)
1727 error = metaslab_alloc(spa, spa->spa_normal_class, size,
1728 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
1729
1730 if (error == 0) {
1731 BP_SET_LSIZE(new_bp, size);
1732 BP_SET_PSIZE(new_bp, size);
1733 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
1734 BP_SET_CHECKSUM(new_bp, ZIO_CHECKSUM_ZILOG);
1735 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
1736 BP_SET_LEVEL(new_bp, 0);
1737 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
1738 }
1739
1740 return (error);
1741 }
1742
1743 /*
1744 * Free an intent log block. We know it can't be a gang block, so there's
1745 * nothing to do except metaslab_free() it.
1746 */
1747 void
1748 zio_free_blk(spa_t *spa, blkptr_t *bp, uint64_t txg)
1749 {
1750 ASSERT(!BP_IS_GANG(bp));
1751
1752 metaslab_free(spa, bp, txg, B_FALSE);
1753 }
1754
1755 /*
1756 * ==========================================================================
1757 * Read and write to physical devices
1758 * ==========================================================================
1759 */
1760 static int
1761 zio_vdev_io_start(zio_t *zio)
1762 {
1763 vdev_t *vd = zio->io_vd;
1764 uint64_t align;
1765 spa_t *spa = zio->io_spa;
1766
1767 ASSERT(zio->io_error == 0);
1768 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
1769
1770 if (vd == NULL) {
1771 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
1772 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
1773
1774 /*
1775 * The mirror_ops handle multiple DVAs in a single BP.
1776 */
1777 return (vdev_mirror_ops.vdev_op_io_start(zio));
1778 }
1779
1780 align = 1ULL << vd->vdev_top->vdev_ashift;
1781
1782 if (P2PHASE(zio->io_size, align) != 0) {
1783 uint64_t asize = P2ROUNDUP(zio->io_size, align);
1784 char *abuf = zio_buf_alloc(asize);
1785 ASSERT(vd == vd->vdev_top);
1786 if (zio->io_type == ZIO_TYPE_WRITE) {
1787 bcopy(zio->io_data, abuf, zio->io_size);
1788 bzero(abuf + zio->io_size, asize - zio->io_size);
1789 }
1790 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
1791 }
1792
1793 ASSERT(P2PHASE(zio->io_offset, align) == 0);
1794 ASSERT(P2PHASE(zio->io_size, align) == 0);
1795 ASSERT(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
1796
1797 /*
1798 * If this is a repair I/O, and there's no self-healing involved --
1799 * that is, we're just resilvering what we expect to resilver --
1800 * then don't do the I/O unless zio's txg is actually in vd's DTL.
1801 * This prevents spurious resilvering with nested replication.
1802 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
1803 * A is out of date, we'll read from C+D, then use the data to
1804 * resilver A+B -- but we don't actually want to resilver B, just A.
1805 * The top-level mirror has no way to know this, so instead we just
1806 * discard unnecessary repairs as we work our way down the vdev tree.
1807 * The same logic applies to any form of nested replication:
1808 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
1809 */
1810 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
1811 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
1812 zio->io_txg != 0 && /* not a delegated i/o */
1813 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
1814 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
1815 zio_vdev_io_bypass(zio);
1816 return (ZIO_PIPELINE_CONTINUE);
1817 }
1818
1819 if (vd->vdev_ops->vdev_op_leaf &&
1820 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
1821
1822 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
1823 return (ZIO_PIPELINE_CONTINUE);
1824
1825 if ((zio = vdev_queue_io(zio)) == NULL)
1826 return (ZIO_PIPELINE_STOP);
1827
1828 if (!vdev_accessible(vd, zio)) {
1829 zio->io_error = ENXIO;
1830 zio_interrupt(zio);
1831 return (ZIO_PIPELINE_STOP);
1832 }
1833 }
1834
1835 return (vd->vdev_ops->vdev_op_io_start(zio));
1836 }
1837
1838 static int
1839 zio_vdev_io_done(zio_t *zio)
1840 {
1841 vdev_t *vd = zio->io_vd;
1842 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
1843 boolean_t unexpected_error = B_FALSE;
1844
1845 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
1846 return (ZIO_PIPELINE_STOP);
1847
1848 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
1849
1850 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
1851
1852 vdev_queue_io_done(zio);
1853
1854 if (zio->io_type == ZIO_TYPE_WRITE)
1855 vdev_cache_write(zio);
1856
1857 if (zio_injection_enabled && zio->io_error == 0)
1858 zio->io_error = zio_handle_device_injection(vd,
1859 zio, EIO);
1860
1861 if (zio_injection_enabled && zio->io_error == 0)
1862 zio->io_error = zio_handle_label_injection(zio, EIO);
1863
1864 if (zio->io_error) {
1865 if (!vdev_accessible(vd, zio)) {
1866 zio->io_error = ENXIO;
1867 } else {
1868 unexpected_error = B_TRUE;
1869 }
1870 }
1871 }
1872
1873 ops->vdev_op_io_done(zio);
1874
1875 if (unexpected_error)
1876 VERIFY(vdev_probe(vd, zio) == NULL);
1877
1878 return (ZIO_PIPELINE_CONTINUE);
1879 }
1880
1881 static int
1882 zio_vdev_io_assess(zio_t *zio)
1883 {
1884 vdev_t *vd = zio->io_vd;
1885
1886 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
1887 return (ZIO_PIPELINE_STOP);
1888
1889 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
1890 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
1891
1892 if (zio->io_vsd != NULL) {
1893 zio->io_vsd_free(zio);
1894 zio->io_vsd = NULL;
1895 }
1896
1897 if (zio_injection_enabled && zio->io_error == 0)
1898 zio->io_error = zio_handle_fault_injection(zio, EIO);
1899
1900 /*
1901 * If the I/O failed, determine whether we should attempt to retry it.
1902 */
1903 if (zio->io_error && vd == NULL &&
1904 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
1905 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
1906 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
1907 zio->io_error = 0;
1908 zio->io_flags |= ZIO_FLAG_IO_RETRY |
1909 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
1910 zio->io_stage = ZIO_STAGE_VDEV_IO_START - 1;
1911 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE);
1912 return (ZIO_PIPELINE_STOP);
1913 }
1914
1915 /*
1916 * If we got an error on a leaf device, convert it to ENXIO
1917 * if the device is not accessible at all.
1918 */
1919 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
1920 !vdev_accessible(vd, zio))
1921 zio->io_error = ENXIO;
1922
1923 /*
1924 * If we can't write to an interior vdev (mirror or RAID-Z),
1925 * set vdev_cant_write so that we stop trying to allocate from it.
1926 */
1927 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
1928 vd != NULL && !vd->vdev_ops->vdev_op_leaf)
1929 vd->vdev_cant_write = B_TRUE;
1930
1931 if (zio->io_error)
1932 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1933
1934 return (ZIO_PIPELINE_CONTINUE);
1935 }
1936
1937 void
1938 zio_vdev_io_reissue(zio_t *zio)
1939 {
1940 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
1941 ASSERT(zio->io_error == 0);
1942
1943 zio->io_stage--;
1944 }
1945
1946 void
1947 zio_vdev_io_redone(zio_t *zio)
1948 {
1949 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
1950
1951 zio->io_stage--;
1952 }
1953
1954 void
1955 zio_vdev_io_bypass(zio_t *zio)
1956 {
1957 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
1958 ASSERT(zio->io_error == 0);
1959
1960 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
1961 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS - 1;
1962 }
1963
1964 /*
1965 * ==========================================================================
1966 * Generate and verify checksums
1967 * ==========================================================================
1968 */
1969 static int
1970 zio_checksum_generate(zio_t *zio)
1971 {
1972 blkptr_t *bp = zio->io_bp;
1973 enum zio_checksum checksum;
1974
1975 if (bp == NULL) {
1976 /*
1977 * This is zio_write_phys().
1978 * We're either generating a label checksum, or none at all.
1979 */
1980 checksum = zio->io_prop.zp_checksum;
1981
1982 if (checksum == ZIO_CHECKSUM_OFF)
1983 return (ZIO_PIPELINE_CONTINUE);
1984
1985 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
1986 } else {
1987 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
1988 ASSERT(!IO_IS_ALLOCATING(zio));
1989 checksum = ZIO_CHECKSUM_GANG_HEADER;
1990 } else {
1991 checksum = BP_GET_CHECKSUM(bp);
1992 }
1993 }
1994
1995 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
1996
1997 return (ZIO_PIPELINE_CONTINUE);
1998 }
1999
2000 static int
2001 zio_checksum_verify(zio_t *zio)
2002 {
2003 blkptr_t *bp = zio->io_bp;
2004 int error;
2005
2006 if (bp == NULL) {
2007 /*
2008 * This is zio_read_phys().
2009 * We're either verifying a label checksum, or nothing at all.
2010 */
2011 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2012 return (ZIO_PIPELINE_CONTINUE);
2013
2014 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2015 }
2016
2017 if ((error = zio_checksum_error(zio)) != 0) {
2018 zio->io_error = error;
2019 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2020 zfs_ereport_post(FM_EREPORT_ZFS_CHECKSUM,
2021 zio->io_spa, zio->io_vd, zio, 0, 0);
2022 }
2023 }
2024
2025 return (ZIO_PIPELINE_CONTINUE);
2026 }
2027
2028 /*
2029 * Called by RAID-Z to ensure we don't compute the checksum twice.
2030 */
2031 void
2032 zio_checksum_verified(zio_t *zio)
2033 {
2034 zio->io_pipeline &= ~(1U << ZIO_STAGE_CHECKSUM_VERIFY);
2035 }
2036
2037 /*
2038 * ==========================================================================
2039 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2040 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2041 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2042 * indicate errors that are specific to one I/O, and most likely permanent.
2043 * Any other error is presumed to be worse because we weren't expecting it.
2044 * ==========================================================================
2045 */
2046 int
2047 zio_worst_error(int e1, int e2)
2048 {
2049 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2050 int r1, r2;
2051
2052 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2053 if (e1 == zio_error_rank[r1])
2054 break;
2055
2056 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2057 if (e2 == zio_error_rank[r2])
2058 break;
2059
2060 return (r1 > r2 ? e1 : e2);
2061 }
2062
2063 /*
2064 * ==========================================================================
2065 * I/O completion
2066 * ==========================================================================
2067 */
2068 static int
2069 zio_ready(zio_t *zio)
2070 {
2071 blkptr_t *bp = zio->io_bp;
2072 zio_t *pio, *pio_next;
2073
2074 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY))
2075 return (ZIO_PIPELINE_STOP);
2076
2077 if (zio->io_ready) {
2078 ASSERT(IO_IS_ALLOCATING(zio));
2079 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2080 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2081
2082 zio->io_ready(zio);
2083 }
2084
2085 if (bp != NULL && bp != &zio->io_bp_copy)
2086 zio->io_bp_copy = *bp;
2087
2088 if (zio->io_error)
2089 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2090
2091 mutex_enter(&zio->io_lock);
2092 zio->io_state[ZIO_WAIT_READY] = 1;
2093 pio = zio_walk_parents(zio);
2094 mutex_exit(&zio->io_lock);
2095
2096 /*
2097 * As we notify zio's parents, new parents could be added.
2098 * New parents go to the head of zio's io_parent_list, however,
2099 * so we will (correctly) not notify them. The remainder of zio's
2100 * io_parent_list, from 'pio_next' onward, cannot change because
2101 * all parents must wait for us to be done before they can be done.
2102 */
2103 for (; pio != NULL; pio = pio_next) {
2104 pio_next = zio_walk_parents(zio);
2105 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2106 }
2107
2108 return (ZIO_PIPELINE_CONTINUE);
2109 }
2110
2111 static int
2112 zio_done(zio_t *zio)
2113 {
2114 spa_t *spa = zio->io_spa;
2115 zio_t *lio = zio->io_logical;
2116 blkptr_t *bp = zio->io_bp;
2117 vdev_t *vd = zio->io_vd;
2118 uint64_t psize = zio->io_size;
2119 zio_t *pio, *pio_next;
2120
2121 /*
2122 * If our children haven't all completed,
2123 * wait for them and then repeat this pipeline stage.
2124 */
2125 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2126 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2127 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2128 return (ZIO_PIPELINE_STOP);
2129
2130 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2131 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2132 ASSERT(zio->io_children[c][w] == 0);
2133
2134 if (bp != NULL) {
2135 ASSERT(bp->blk_pad[0] == 0);
2136 ASSERT(bp->blk_pad[1] == 0);
2137 ASSERT(bp->blk_pad[2] == 0);
2138 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2139 (bp == zio_unique_parent(zio)->io_bp));
2140 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2141 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2142 ASSERT(!BP_SHOULD_BYTESWAP(bp));
2143 ASSERT3U(zio->io_prop.zp_ndvas, <=, BP_GET_NDVAS(bp));
2144 ASSERT(BP_COUNT_GANG(bp) == 0 ||
2145 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2146 }
2147 }
2148
2149 /*
2150 * If there were child vdev or gang errors, they apply to us now.
2151 */
2152 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2153 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2154
2155 zio_pop_transforms(zio); /* note: may set zio->io_error */
2156
2157 vdev_stat_update(zio, psize);
2158
2159 if (zio->io_error) {
2160 /*
2161 * If this I/O is attached to a particular vdev,
2162 * generate an error message describing the I/O failure
2163 * at the block level. We ignore these errors if the
2164 * device is currently unavailable.
2165 */
2166 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
2167 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
2168
2169 if ((zio->io_error == EIO ||
2170 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) && zio == lio) {
2171 /*
2172 * For logical I/O requests, tell the SPA to log the
2173 * error and generate a logical data ereport.
2174 */
2175 spa_log_error(spa, zio);
2176 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
2177 0, 0);
2178 }
2179 }
2180
2181 if (zio->io_error && zio == lio) {
2182 /*
2183 * Determine whether zio should be reexecuted. This will
2184 * propagate all the way to the root via zio_notify_parent().
2185 */
2186 ASSERT(vd == NULL && bp != NULL);
2187
2188 if (IO_IS_ALLOCATING(zio))
2189 if (zio->io_error != ENOSPC)
2190 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2191 else
2192 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2193
2194 if ((zio->io_type == ZIO_TYPE_READ ||
2195 zio->io_type == ZIO_TYPE_FREE) &&
2196 zio->io_error == ENXIO &&
2197 spa->spa_load_state == SPA_LOAD_NONE &&
2198 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
2199 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2200
2201 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2202 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2203 }
2204
2205 /*
2206 * If there were logical child errors, they apply to us now.
2207 * We defer this until now to avoid conflating logical child
2208 * errors with errors that happened to the zio itself when
2209 * updating vdev stats and reporting FMA events above.
2210 */
2211 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2212
2213 if ((zio->io_error || zio->io_reexecute) && IO_IS_ALLOCATING(zio) &&
2214 zio->io_child_type == ZIO_CHILD_LOGICAL) {
2215 ASSERT(zio->io_child_type != ZIO_CHILD_GANG);
2216 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
2217 }
2218
2219 zio_gang_tree_free(&zio->io_gang_tree);
2220
2221 /*
2222 * Godfather I/Os should never suspend.
2223 */
2224 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2225 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2226 zio->io_reexecute = 0;
2227
2228 if (zio->io_reexecute) {
2229 /*
2230 * This is a logical I/O that wants to reexecute.
2231 *
2232 * Reexecute is top-down. When an i/o fails, if it's not
2233 * the root, it simply notifies its parent and sticks around.
2234 * The parent, seeing that it still has children in zio_done(),
2235 * does the same. This percolates all the way up to the root.
2236 * The root i/o will reexecute or suspend the entire tree.
2237 *
2238 * This approach ensures that zio_reexecute() honors
2239 * all the original i/o dependency relationships, e.g.
2240 * parents not executing until children are ready.
2241 */
2242 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2243
2244 zio->io_gang_leader = NULL;
2245
2246 mutex_enter(&zio->io_lock);
2247 zio->io_state[ZIO_WAIT_DONE] = 1;
2248 mutex_exit(&zio->io_lock);
2249
2250 /*
2251 * "The Godfather" I/O monitors its children but is
2252 * not a true parent to them. It will track them through
2253 * the pipeline but severs its ties whenever they get into
2254 * trouble (e.g. suspended). This allows "The Godfather"
2255 * I/O to return status without blocking.
2256 */
2257 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2258 zio_link_t *zl = zio->io_walk_link;
2259 pio_next = zio_walk_parents(zio);
2260
2261 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
2262 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
2263 zio_remove_child(pio, zio, zl);
2264 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2265 }
2266 }
2267
2268 if ((pio = zio_unique_parent(zio)) != NULL) {
2269 /*
2270 * We're not a root i/o, so there's nothing to do
2271 * but notify our parent. Don't propagate errors
2272 * upward since we haven't permanently failed yet.
2273 */
2274 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2275 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
2276 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2277 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
2278 /*
2279 * We'd fail again if we reexecuted now, so suspend
2280 * until conditions improve (e.g. device comes online).
2281 */
2282 zio_suspend(spa, zio);
2283 } else {
2284 /*
2285 * Reexecution is potentially a huge amount of work.
2286 * Hand it off to the otherwise-unused claim taskq.
2287 */
2288 (void) taskq_dispatch(
2289 spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2290 (task_func_t *)zio_reexecute, zio, TQ_SLEEP);
2291 }
2292 return (ZIO_PIPELINE_STOP);
2293 }
2294
2295 ASSERT(zio_walk_children(zio) == NULL);
2296 ASSERT(zio->io_reexecute == 0);
2297 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
2298
2299 /*
2300 * It is the responsibility of the done callback to ensure that this
2301 * particular zio is no longer discoverable for adoption, and as
2302 * such, cannot acquire any new parents.
2303 */
2304 if (zio->io_done)
2305 zio->io_done(zio);
2306
2307 mutex_enter(&zio->io_lock);
2308 zio->io_state[ZIO_WAIT_DONE] = 1;
2309 mutex_exit(&zio->io_lock);
2310
2311 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2312 zio_link_t *zl = zio->io_walk_link;
2313 pio_next = zio_walk_parents(zio);
2314 zio_remove_child(pio, zio, zl);
2315 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2316 }
2317
2318 if (zio->io_waiter != NULL) {
2319 mutex_enter(&zio->io_lock);
2320 zio->io_executor = NULL;
2321 cv_broadcast(&zio->io_cv);
2322 mutex_exit(&zio->io_lock);
2323 } else {
2324 zio_destroy(zio);
2325 }
2326
2327 return (ZIO_PIPELINE_STOP);
2328 }
2329
2330 /*
2331 * ==========================================================================
2332 * I/O pipeline definition
2333 * ==========================================================================
2334 */
2335 static zio_pipe_stage_t *zio_pipeline[ZIO_STAGES] = {
2336 NULL,
2337 zio_issue_async,
2338 zio_read_bp_init,
2339 zio_write_bp_init,
2340 zio_checksum_generate,
2341 zio_gang_assemble,
2342 zio_gang_issue,
2343 zio_dva_allocate,
2344 zio_dva_free,
2345 zio_dva_claim,
2346 zio_ready,
2347 zio_vdev_io_start,
2348 zio_vdev_io_done,
2349 zio_vdev_io_assess,
2350 zio_checksum_verify,
2351 zio_done
2352 };
ZFS On Linux mirror for PVE