]> git.proxmox.com Git - mirror_zfs-debian.git/blob - module/zfs/zio.c
projects / mirror_zfs-debian.git / blob
? search:


d347920ea6bb79a183df87a997da0b03220909d7
[mirror_zfs-debian.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 2008 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_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
73 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
74
75 #ifdef _KERNEL
76 extern vmem_t *zio_alloc_arena;
77 #endif
78
79 /*
80 * An allocating zio is one that either currently has the DVA allocate
81 * stage set or will have it later in its lifetime.
82 */
83 #define IO_IS_ALLOCATING(zio) \
84 ((zio)->io_orig_pipeline & (1U << ZIO_STAGE_DVA_ALLOCATE))
85
86 void
87 zio_init(void)
88 {
89 size_t c;
90 vmem_t *data_alloc_arena = NULL;
91
92 #ifdef _KERNEL
93 data_alloc_arena = zio_alloc_arena;
94 #endif
95 zio_cache = kmem_cache_create("zio_cache", sizeof (zio_t), 0,
96 NULL, NULL, NULL, NULL, NULL, 0);
97
98 /*
99 * For small buffers, we want a cache for each multiple of
100 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
101 * for each quarter-power of 2. For large buffers, we want
102 * a cache for each multiple of PAGESIZE.
103 */
104 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
105 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
106 size_t p2 = size;
107 size_t align = 0;
108
109 while (p2 & (p2 - 1))
110 p2 &= p2 - 1;
111
112 if (size <= 4 * SPA_MINBLOCKSIZE) {
113 align = SPA_MINBLOCKSIZE;
114 } else if (P2PHASE(size, PAGESIZE) == 0) {
115 align = PAGESIZE;
116 } else if (P2PHASE(size, p2 >> 2) == 0) {
117 align = p2 >> 2;
118 }
119
120 if (align != 0) {
121 char name[36];
122 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
123 zio_buf_cache[c] = kmem_cache_create(name, size,
124 align, NULL, NULL, NULL, NULL, NULL, KMC_NODEBUG);
125
126 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
127 zio_data_buf_cache[c] = kmem_cache_create(name, size,
128 align, NULL, NULL, NULL, NULL, data_alloc_arena,
129 KMC_NODEBUG);
130 }
131 }
132
133 while (--c != 0) {
134 ASSERT(zio_buf_cache[c] != NULL);
135 if (zio_buf_cache[c - 1] == NULL)
136 zio_buf_cache[c - 1] = zio_buf_cache[c];
137
138 ASSERT(zio_data_buf_cache[c] != NULL);
139 if (zio_data_buf_cache[c - 1] == NULL)
140 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
141 }
142
143 zio_inject_init();
144 }
145
146 void
147 zio_fini(void)
148 {
149 size_t c;
150 kmem_cache_t *last_cache = NULL;
151 kmem_cache_t *last_data_cache = NULL;
152
153 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
154 if (zio_buf_cache[c] != last_cache) {
155 last_cache = zio_buf_cache[c];
156 kmem_cache_destroy(zio_buf_cache[c]);
157 }
158 zio_buf_cache[c] = NULL;
159
160 if (zio_data_buf_cache[c] != last_data_cache) {
161 last_data_cache = zio_data_buf_cache[c];
162 kmem_cache_destroy(zio_data_buf_cache[c]);
163 }
164 zio_data_buf_cache[c] = NULL;
165 }
166
167 kmem_cache_destroy(zio_cache);
168
169 zio_inject_fini();
170 }
171
172 /*
173 * ==========================================================================
174 * Allocate and free I/O buffers
175 * ==========================================================================
176 */
177
178 /*
179 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
180 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
181 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
182 * excess / transient data in-core during a crashdump.
183 */
184 void *
185 zio_buf_alloc(size_t size)
186 {
187 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
188
189 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
190
191 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
192 }
193
194 /*
195 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
196 * crashdump if the kernel panics. This exists so that we will limit the amount
197 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
198 * of kernel heap dumped to disk when the kernel panics)
199 */
200 void *
201 zio_data_buf_alloc(size_t size)
202 {
203 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
204
205 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
206
207 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
208 }
209
210 void
211 zio_buf_free(void *buf, size_t size)
212 {
213 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
214
215 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
216
217 kmem_cache_free(zio_buf_cache[c], buf);
218 }
219
220 void
221 zio_data_buf_free(void *buf, size_t size)
222 {
223 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
224
225 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
226
227 kmem_cache_free(zio_data_buf_cache[c], buf);
228 }
229
230 /*
231 * ==========================================================================
232 * Push and pop I/O transform buffers
233 * ==========================================================================
234 */
235 static void
236 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
237 zio_transform_func_t *transform)
238 {
239 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
240
241 zt->zt_orig_data = zio->io_data;
242 zt->zt_orig_size = zio->io_size;
243 zt->zt_bufsize = bufsize;
244 zt->zt_transform = transform;
245
246 zt->zt_next = zio->io_transform_stack;
247 zio->io_transform_stack = zt;
248
249 zio->io_data = data;
250 zio->io_size = size;
251 }
252
253 static void
254 zio_pop_transforms(zio_t *zio)
255 {
256 zio_transform_t *zt;
257
258 while ((zt = zio->io_transform_stack) != NULL) {
259 if (zt->zt_transform != NULL)
260 zt->zt_transform(zio,
261 zt->zt_orig_data, zt->zt_orig_size);
262
263 zio_buf_free(zio->io_data, zt->zt_bufsize);
264
265 zio->io_data = zt->zt_orig_data;
266 zio->io_size = zt->zt_orig_size;
267 zio->io_transform_stack = zt->zt_next;
268
269 kmem_free(zt, sizeof (zio_transform_t));
270 }
271 }
272
273 /*
274 * ==========================================================================
275 * I/O transform callbacks for subblocks and decompression
276 * ==========================================================================
277 */
278 static void
279 zio_subblock(zio_t *zio, void *data, uint64_t size)
280 {
281 ASSERT(zio->io_size > size);
282
283 if (zio->io_type == ZIO_TYPE_READ)
284 bcopy(zio->io_data, data, size);
285 }
286
287 static void
288 zio_decompress(zio_t *zio, void *data, uint64_t size)
289 {
290 if (zio->io_error == 0 &&
291 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
292 zio->io_data, zio->io_size, data, size) != 0)
293 zio->io_error = EIO;
294 }
295
296 /*
297 * ==========================================================================
298 * I/O parent/child relationships and pipeline interlocks
299 * ==========================================================================
300 */
301
302 static void
303 zio_add_child(zio_t *pio, zio_t *zio)
304 {
305 mutex_enter(&pio->io_lock);
306 if (zio->io_stage < ZIO_STAGE_READY)
307 pio->io_children[zio->io_child_type][ZIO_WAIT_READY]++;
308 if (zio->io_stage < ZIO_STAGE_DONE)
309 pio->io_children[zio->io_child_type][ZIO_WAIT_DONE]++;
310 zio->io_sibling_prev = NULL;
311 zio->io_sibling_next = pio->io_child;
312 if (pio->io_child != NULL)
313 pio->io_child->io_sibling_prev = zio;
314 pio->io_child = zio;
315 zio->io_parent = pio;
316 mutex_exit(&pio->io_lock);
317 }
318
319 static void
320 zio_remove_child(zio_t *pio, zio_t *zio)
321 {
322 zio_t *next, *prev;
323
324 ASSERT(zio->io_parent == pio);
325
326 mutex_enter(&pio->io_lock);
327 next = zio->io_sibling_next;
328 prev = zio->io_sibling_prev;
329 if (next != NULL)
330 next->io_sibling_prev = prev;
331 if (prev != NULL)
332 prev->io_sibling_next = next;
333 if (pio->io_child == zio)
334 pio->io_child = next;
335 mutex_exit(&pio->io_lock);
336 }
337
338 static boolean_t
339 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
340 {
341 uint64_t *countp = &zio->io_children[child][wait];
342 boolean_t waiting = B_FALSE;
343
344 mutex_enter(&zio->io_lock);
345 ASSERT(zio->io_stall == NULL);
346 if (*countp != 0) {
347 zio->io_stage--;
348 zio->io_stall = countp;
349 waiting = B_TRUE;
350 }
351 mutex_exit(&zio->io_lock);
352
353 return (waiting);
354 }
355
356 static void
357 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
358 {
359 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
360 int *errorp = &pio->io_child_error[zio->io_child_type];
361
362 mutex_enter(&pio->io_lock);
363 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
364 *errorp = zio_worst_error(*errorp, zio->io_error);
365 pio->io_reexecute |= zio->io_reexecute;
366 ASSERT3U(*countp, >, 0);
367 if (--*countp == 0 && pio->io_stall == countp) {
368 pio->io_stall = NULL;
369 mutex_exit(&pio->io_lock);
370 zio_execute(pio);
371 } else {
372 mutex_exit(&pio->io_lock);
373 }
374 }
375
376 static void
377 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
378 {
379 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
380 zio->io_error = zio->io_child_error[c];
381 }
382
383 /*
384 * ==========================================================================
385 * Create the various types of I/O (read, write, free, etc)
386 * ==========================================================================
387 */
388 static zio_t *
389 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
390 void *data, uint64_t size, zio_done_func_t *done, void *private,
391 zio_type_t type, int priority, int flags, vdev_t *vd, uint64_t offset,
392 const zbookmark_t *zb, uint8_t stage, uint32_t pipeline)
393 {
394 zio_t *zio;
395
396 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
397 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
398 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
399
400 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
401 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
402 ASSERT(vd || stage == ZIO_STAGE_OPEN);
403
404 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
405 bzero(zio, sizeof (zio_t));
406
407 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
408 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
409
410 if (vd != NULL)
411 zio->io_child_type = ZIO_CHILD_VDEV;
412 else if (flags & ZIO_FLAG_GANG_CHILD)
413 zio->io_child_type = ZIO_CHILD_GANG;
414 else
415 zio->io_child_type = ZIO_CHILD_LOGICAL;
416
417 if (bp != NULL) {
418 zio->io_bp = bp;
419 zio->io_bp_copy = *bp;
420 zio->io_bp_orig = *bp;
421 if (type != ZIO_TYPE_WRITE)
422 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
423 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
424 if (BP_IS_GANG(bp))
425 pipeline |= ZIO_GANG_STAGES;
426 zio->io_logical = zio;
427 }
428 }
429
430 zio->io_spa = spa;
431 zio->io_txg = txg;
432 zio->io_data = data;
433 zio->io_size = size;
434 zio->io_done = done;
435 zio->io_private = private;
436 zio->io_type = type;
437 zio->io_priority = priority;
438 zio->io_vd = vd;
439 zio->io_offset = offset;
440 zio->io_orig_flags = zio->io_flags = flags;
441 zio->io_orig_stage = zio->io_stage = stage;
442 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
443
444 if (zb != NULL)
445 zio->io_bookmark = *zb;
446
447 if (pio != NULL) {
448 /*
449 * Logical I/Os can have logical, gang, or vdev children.
450 * Gang I/Os can have gang or vdev children.
451 * Vdev I/Os can only have vdev children.
452 * The following ASSERT captures all of these constraints.
453 */
454 ASSERT(zio->io_child_type <= pio->io_child_type);
455 if (zio->io_logical == NULL)
456 zio->io_logical = pio->io_logical;
457 zio_add_child(pio, zio);
458 }
459
460 return (zio);
461 }
462
463 static void
464 zio_destroy(zio_t *zio)
465 {
466 spa_t *spa = zio->io_spa;
467 uint8_t async_root = zio->io_async_root;
468
469 mutex_destroy(&zio->io_lock);
470 cv_destroy(&zio->io_cv);
471 kmem_cache_free(zio_cache, zio);
472
473 if (async_root) {
474 mutex_enter(&spa->spa_async_root_lock);
475 if (--spa->spa_async_root_count == 0)
476 cv_broadcast(&spa->spa_async_root_cv);
477 mutex_exit(&spa->spa_async_root_lock);
478 }
479 }
480
481 zio_t *
482 zio_null(zio_t *pio, spa_t *spa, zio_done_func_t *done, void *private,
483 int flags)
484 {
485 zio_t *zio;
486
487 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
488 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, NULL, 0, NULL,
489 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
490
491 return (zio);
492 }
493
494 zio_t *
495 zio_root(spa_t *spa, zio_done_func_t *done, void *private, int flags)
496 {
497 return (zio_null(NULL, spa, done, private, flags));
498 }
499
500 zio_t *
501 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
502 void *data, uint64_t size, zio_done_func_t *done, void *private,
503 int priority, int flags, const zbookmark_t *zb)
504 {
505 zio_t *zio;
506
507 zio = zio_create(pio, spa, bp->blk_birth, (blkptr_t *)bp,
508 data, size, done, private,
509 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
510 ZIO_STAGE_OPEN, ZIO_READ_PIPELINE);
511
512 return (zio);
513 }
514
515 void
516 zio_skip_write(zio_t *zio)
517 {
518 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
519 ASSERT(zio->io_stage == ZIO_STAGE_READY);
520 ASSERT(!BP_IS_GANG(zio->io_bp));
521
522 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
523 }
524
525 zio_t *
526 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
527 void *data, uint64_t size, zio_prop_t *zp,
528 zio_done_func_t *ready, zio_done_func_t *done, void *private,
529 int priority, int flags, const zbookmark_t *zb)
530 {
531 zio_t *zio;
532
533 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
534 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
535 zp->zp_compress >= ZIO_COMPRESS_OFF &&
536 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
537 zp->zp_type < DMU_OT_NUMTYPES &&
538 zp->zp_level < 32 &&
539 zp->zp_ndvas > 0 &&
540 zp->zp_ndvas <= spa_max_replication(spa));
541 ASSERT(ready != NULL);
542
543 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
544 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
545 ZIO_STAGE_OPEN, ZIO_WRITE_PIPELINE);
546
547 zio->io_ready = ready;
548 zio->io_prop = *zp;
549
550 return (zio);
551 }
552
553 zio_t *
554 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
555 uint64_t size, zio_done_func_t *done, void *private, int priority,
556 int flags, zbookmark_t *zb)
557 {
558 zio_t *zio;
559
560 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
561 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
562 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
563
564 return (zio);
565 }
566
567 zio_t *
568 zio_free(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
569 zio_done_func_t *done, void *private, int flags)
570 {
571 zio_t *zio;
572
573 ASSERT(!BP_IS_HOLE(bp));
574
575 if (bp->blk_fill == BLK_FILL_ALREADY_FREED)
576 return (zio_null(pio, spa, NULL, NULL, flags));
577
578 if (txg == spa->spa_syncing_txg &&
579 spa_sync_pass(spa) > SYNC_PASS_DEFERRED_FREE) {
580 bplist_enqueue_deferred(&spa->spa_sync_bplist, bp);
581 return (zio_null(pio, spa, NULL, NULL, flags));
582 }
583
584 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
585 done, private, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
586 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
587
588 return (zio);
589 }
590
591 zio_t *
592 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
593 zio_done_func_t *done, void *private, int flags)
594 {
595 zio_t *zio;
596
597 /*
598 * A claim is an allocation of a specific block. Claims are needed
599 * to support immediate writes in the intent log. The issue is that
600 * immediate writes contain committed data, but in a txg that was
601 * *not* committed. Upon opening the pool after an unclean shutdown,
602 * the intent log claims all blocks that contain immediate write data
603 * so that the SPA knows they're in use.
604 *
605 * All claims *must* be resolved in the first txg -- before the SPA
606 * starts allocating blocks -- so that nothing is allocated twice.
607 */
608 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
609 ASSERT3U(spa_first_txg(spa), <=, txg);
610
611 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
612 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
613 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
614
615 return (zio);
616 }
617
618 zio_t *
619 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
620 zio_done_func_t *done, void *private, int priority, int flags)
621 {
622 zio_t *zio;
623 int c;
624
625 if (vd->vdev_children == 0) {
626 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
627 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
628 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
629
630 zio->io_cmd = cmd;
631 } else {
632 zio = zio_null(pio, spa, NULL, NULL, flags);
633
634 for (c = 0; c < vd->vdev_children; c++)
635 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
636 done, private, priority, flags));
637 }
638
639 return (zio);
640 }
641
642 zio_t *
643 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
644 void *data, int checksum, zio_done_func_t *done, void *private,
645 int priority, int flags, boolean_t labels)
646 {
647 zio_t *zio;
648
649 ASSERT(vd->vdev_children == 0);
650 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
651 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
652 ASSERT3U(offset + size, <=, vd->vdev_psize);
653
654 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
655 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
656 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
657
658 zio->io_prop.zp_checksum = checksum;
659
660 return (zio);
661 }
662
663 zio_t *
664 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
665 void *data, int checksum, zio_done_func_t *done, void *private,
666 int priority, int flags, boolean_t labels)
667 {
668 zio_t *zio;
669
670 ASSERT(vd->vdev_children == 0);
671 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
672 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
673 ASSERT3U(offset + size, <=, vd->vdev_psize);
674
675 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
676 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
677 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
678
679 zio->io_prop.zp_checksum = checksum;
680
681 if (zio_checksum_table[checksum].ci_zbt) {
682 /*
683 * zbt checksums are necessarily destructive -- they modify
684 * the end of the write buffer to hold the verifier/checksum.
685 * Therefore, we must make a local copy in case the data is
686 * being written to multiple places in parallel.
687 */
688 void *wbuf = zio_buf_alloc(size);
689 bcopy(data, wbuf, size);
690 zio_push_transform(zio, wbuf, size, size, NULL);
691 }
692
693 return (zio);
694 }
695
696 /*
697 * Create a child I/O to do some work for us.
698 */
699 zio_t *
700 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
701 void *data, uint64_t size, int type, int priority, int flags,
702 zio_done_func_t *done, void *private)
703 {
704 uint32_t pipeline = ZIO_VDEV_CHILD_PIPELINE;
705 zio_t *zio;
706
707 ASSERT(vd->vdev_parent ==
708 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
709
710 if (type == ZIO_TYPE_READ && bp != NULL) {
711 /*
712 * If we have the bp, then the child should perform the
713 * checksum and the parent need not. This pushes error
714 * detection as close to the leaves as possible and
715 * eliminates redundant checksums in the interior nodes.
716 */
717 pipeline |= 1U << ZIO_STAGE_CHECKSUM_VERIFY;
718 pio->io_pipeline &= ~(1U << ZIO_STAGE_CHECKSUM_VERIFY);
719 }
720
721 if (vd->vdev_children == 0)
722 offset += VDEV_LABEL_START_SIZE;
723
724 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
725 done, private, type, priority,
726 (pio->io_flags & ZIO_FLAG_VDEV_INHERIT) |
727 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | flags,
728 vd, offset, &pio->io_bookmark,
729 ZIO_STAGE_VDEV_IO_START - 1, pipeline);
730
731 return (zio);
732 }
733
734 zio_t *
735 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
736 int type, int priority, int flags, zio_done_func_t *done, void *private)
737 {
738 zio_t *zio;
739
740 ASSERT(vd->vdev_ops->vdev_op_leaf);
741
742 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
743 data, size, done, private, type, priority,
744 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
745 vd, offset, NULL,
746 ZIO_STAGE_VDEV_IO_START - 1, ZIO_VDEV_CHILD_PIPELINE);
747
748 return (zio);
749 }
750
751 void
752 zio_flush(zio_t *zio, vdev_t *vd)
753 {
754 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
755 NULL, NULL, ZIO_PRIORITY_NOW,
756 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
757 }
758
759 /*
760 * ==========================================================================
761 * Prepare to read and write logical blocks
762 * ==========================================================================
763 */
764
765 static int
766 zio_read_bp_init(zio_t *zio)
767 {
768 blkptr_t *bp = zio->io_bp;
769
770 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF && zio->io_logical == zio) {
771 uint64_t csize = BP_GET_PSIZE(bp);
772 void *cbuf = zio_buf_alloc(csize);
773
774 zio_push_transform(zio, cbuf, csize, csize, zio_decompress);
775 }
776
777 if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
778 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
779
780 return (ZIO_PIPELINE_CONTINUE);
781 }
782
783 static int
784 zio_write_bp_init(zio_t *zio)
785 {
786 zio_prop_t *zp = &zio->io_prop;
787 int compress = zp->zp_compress;
788 blkptr_t *bp = zio->io_bp;
789 void *cbuf;
790 uint64_t lsize = zio->io_size;
791 uint64_t csize = lsize;
792 uint64_t cbufsize = 0;
793 int pass = 1;
794
795 /*
796 * If our children haven't all reached the ready stage,
797 * wait for them and then repeat this pipeline stage.
798 */
799 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
800 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
801 return (ZIO_PIPELINE_STOP);
802
803 if (!IO_IS_ALLOCATING(zio))
804 return (ZIO_PIPELINE_CONTINUE);
805
806 ASSERT(compress != ZIO_COMPRESS_INHERIT);
807
808 if (bp->blk_birth == zio->io_txg) {
809 /*
810 * We're rewriting an existing block, which means we're
811 * working on behalf of spa_sync(). For spa_sync() to
812 * converge, it must eventually be the case that we don't
813 * have to allocate new blocks. But compression changes
814 * the blocksize, which forces a reallocate, and makes
815 * convergence take longer. Therefore, after the first
816 * few passes, stop compressing to ensure convergence.
817 */
818 pass = spa_sync_pass(zio->io_spa);
819 ASSERT(pass > 1);
820
821 if (pass > SYNC_PASS_DONT_COMPRESS)
822 compress = ZIO_COMPRESS_OFF;
823
824 /*
825 * Only MOS (objset 0) data should need to be rewritten.
826 */
827 ASSERT(zio->io_logical->io_bookmark.zb_objset == 0);
828
829 /* Make sure someone doesn't change their mind on overwrites */
830 ASSERT(MIN(zp->zp_ndvas + BP_IS_GANG(bp),
831 spa_max_replication(zio->io_spa)) == BP_GET_NDVAS(bp));
832 }
833
834 if (compress != ZIO_COMPRESS_OFF) {
835 if (!zio_compress_data(compress, zio->io_data, zio->io_size,
836 &cbuf, &csize, &cbufsize)) {
837 compress = ZIO_COMPRESS_OFF;
838 } else if (csize != 0) {
839 zio_push_transform(zio, cbuf, csize, cbufsize, NULL);
840 }
841 }
842
843 /*
844 * The final pass of spa_sync() must be all rewrites, but the first
845 * few passes offer a trade-off: allocating blocks defers convergence,
846 * but newly allocated blocks are sequential, so they can be written
847 * to disk faster. Therefore, we allow the first few passes of
848 * spa_sync() to allocate new blocks, but force rewrites after that.
849 * There should only be a handful of blocks after pass 1 in any case.
850 */
851 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == csize &&
852 pass > SYNC_PASS_REWRITE) {
853 ASSERT(csize != 0);
854 uint32_t gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
855 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
856 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
857 } else {
858 BP_ZERO(bp);
859 zio->io_pipeline = ZIO_WRITE_PIPELINE;
860 }
861
862 if (csize == 0) {
863 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
864 } else {
865 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
866 BP_SET_LSIZE(bp, lsize);
867 BP_SET_PSIZE(bp, csize);
868 BP_SET_COMPRESS(bp, compress);
869 BP_SET_CHECKSUM(bp, zp->zp_checksum);
870 BP_SET_TYPE(bp, zp->zp_type);
871 BP_SET_LEVEL(bp, zp->zp_level);
872 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
873 }
874
875 return (ZIO_PIPELINE_CONTINUE);
876 }
877
878 /*
879 * ==========================================================================
880 * Execute the I/O pipeline
881 * ==========================================================================
882 */
883
884 static void
885 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q)
886 {
887 zio_type_t t = zio->io_type;
888
889 /*
890 * If we're a config writer, the normal issue and interrupt threads
891 * may all be blocked waiting for the config lock. In this case,
892 * select the otherwise-unused taskq for ZIO_TYPE_NULL.
893 */
894 if (zio->io_flags & ZIO_FLAG_CONFIG_WRITER)
895 t = ZIO_TYPE_NULL;
896
897 /*
898 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
899 */
900 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
901 t = ZIO_TYPE_NULL;
902
903 (void) taskq_dispatch(zio->io_spa->spa_zio_taskq[t][q],
904 (task_func_t *)zio_execute, zio, TQ_SLEEP);
905 }
906
907 static boolean_t
908 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
909 {
910 kthread_t *executor = zio->io_executor;
911 spa_t *spa = zio->io_spa;
912
913 for (zio_type_t t = 0; t < ZIO_TYPES; t++)
914 if (taskq_member(spa->spa_zio_taskq[t][q], executor))
915 return (B_TRUE);
916
917 return (B_FALSE);
918 }
919
920 static int
921 zio_issue_async(zio_t *zio)
922 {
923 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE);
924
925 return (ZIO_PIPELINE_STOP);
926 }
927
928 void
929 zio_interrupt(zio_t *zio)
930 {
931 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT);
932 }
933
934 /*
935 * Execute the I/O pipeline until one of the following occurs:
936 * (1) the I/O completes; (2) the pipeline stalls waiting for
937 * dependent child I/Os; (3) the I/O issues, so we're waiting
938 * for an I/O completion interrupt; (4) the I/O is delegated by
939 * vdev-level caching or aggregation; (5) the I/O is deferred
940 * due to vdev-level queueing; (6) the I/O is handed off to
941 * another thread. In all cases, the pipeline stops whenever
942 * there's no CPU work; it never burns a thread in cv_wait().
943 *
944 * There's no locking on io_stage because there's no legitimate way
945 * for multiple threads to be attempting to process the same I/O.
946 */
947 static zio_pipe_stage_t *zio_pipeline[ZIO_STAGES];
948
949 void
950 zio_execute(zio_t *zio)
951 {
952 zio->io_executor = curthread;
953
954 while (zio->io_stage < ZIO_STAGE_DONE) {
955 uint32_t pipeline = zio->io_pipeline;
956 zio_stage_t stage = zio->io_stage;
957 int rv;
958
959 ASSERT(!MUTEX_HELD(&zio->io_lock));
960
961 while (((1U << ++stage) & pipeline) == 0)
962 continue;
963
964 ASSERT(stage <= ZIO_STAGE_DONE);
965 ASSERT(zio->io_stall == NULL);
966
967 /*
968 * If we are in interrupt context and this pipeline stage
969 * will grab a config lock that is held across I/O,
970 * issue async to avoid deadlock.
971 */
972 if (((1U << stage) & ZIO_CONFIG_LOCK_BLOCKING_STAGES) &&
973 zio->io_vd == NULL &&
974 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
975 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE);
976 return;
977 }
978
979 zio->io_stage = stage;
980 rv = zio_pipeline[stage](zio);
981
982 if (rv == ZIO_PIPELINE_STOP)
983 return;
984
985 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
986 }
987 }
988
989 /*
990 * ==========================================================================
991 * Initiate I/O, either sync or async
992 * ==========================================================================
993 */
994 int
995 zio_wait(zio_t *zio)
996 {
997 int error;
998
999 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1000 ASSERT(zio->io_executor == NULL);
1001
1002 zio->io_waiter = curthread;
1003
1004 zio_execute(zio);
1005
1006 mutex_enter(&zio->io_lock);
1007 while (zio->io_executor != NULL)
1008 cv_wait(&zio->io_cv, &zio->io_lock);
1009 mutex_exit(&zio->io_lock);
1010
1011 error = zio->io_error;
1012 zio_destroy(zio);
1013
1014 return (error);
1015 }
1016
1017 void
1018 zio_nowait(zio_t *zio)
1019 {
1020 ASSERT(zio->io_executor == NULL);
1021
1022 if (zio->io_parent == NULL && zio->io_child_type == ZIO_CHILD_LOGICAL) {
1023 /*
1024 * This is a logical async I/O with no parent to wait for it.
1025 * Attach it to the pool's global async root zio so that
1026 * spa_unload() has a way of waiting for async I/O to finish.
1027 */
1028 spa_t *spa = zio->io_spa;
1029 zio->io_async_root = B_TRUE;
1030 mutex_enter(&spa->spa_async_root_lock);
1031 spa->spa_async_root_count++;
1032 mutex_exit(&spa->spa_async_root_lock);
1033 }
1034
1035 zio_execute(zio);
1036 }
1037
1038 /*
1039 * ==========================================================================
1040 * Reexecute or suspend/resume failed I/O
1041 * ==========================================================================
1042 */
1043
1044 static void
1045 zio_reexecute(zio_t *pio)
1046 {
1047 zio_t *zio, *zio_next;
1048
1049 pio->io_flags = pio->io_orig_flags;
1050 pio->io_stage = pio->io_orig_stage;
1051 pio->io_pipeline = pio->io_orig_pipeline;
1052 pio->io_reexecute = 0;
1053 pio->io_error = 0;
1054 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1055 pio->io_child_error[c] = 0;
1056
1057 if (IO_IS_ALLOCATING(pio)) {
1058 /*
1059 * Remember the failed bp so that the io_ready() callback
1060 * can update its accounting upon reexecution. The block
1061 * was already freed in zio_done(); we indicate this with
1062 * a fill count of -1 so that zio_free() knows to skip it.
1063 */
1064 blkptr_t *bp = pio->io_bp;
1065 ASSERT(bp->blk_birth == 0 || bp->blk_birth == pio->io_txg);
1066 bp->blk_fill = BLK_FILL_ALREADY_FREED;
1067 pio->io_bp_orig = *bp;
1068 BP_ZERO(bp);
1069 }
1070
1071 /*
1072 * As we reexecute pio's children, new children could be created.
1073 * New children go to the head of the io_child list, however,
1074 * so we will (correctly) not reexecute them. The key is that
1075 * the remainder of the io_child list, from 'zio_next' onward,
1076 * cannot be affected by any side effects of reexecuting 'zio'.
1077 */
1078 for (zio = pio->io_child; zio != NULL; zio = zio_next) {
1079 zio_next = zio->io_sibling_next;
1080 mutex_enter(&pio->io_lock);
1081 pio->io_children[zio->io_child_type][ZIO_WAIT_READY]++;
1082 pio->io_children[zio->io_child_type][ZIO_WAIT_DONE]++;
1083 mutex_exit(&pio->io_lock);
1084 zio_reexecute(zio);
1085 }
1086
1087 /*
1088 * Now that all children have been reexecuted, execute the parent.
1089 */
1090 zio_execute(pio);
1091 }
1092
1093 void
1094 zio_suspend(spa_t *spa, zio_t *zio)
1095 {
1096 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1097 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1098 "failure and the failure mode property for this pool "
1099 "is set to panic.", spa_name(spa));
1100
1101 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1102
1103 mutex_enter(&spa->spa_suspend_lock);
1104
1105 if (spa->spa_suspend_zio_root == NULL)
1106 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL, 0);
1107
1108 spa->spa_suspended = B_TRUE;
1109
1110 if (zio != NULL) {
1111 ASSERT(zio != spa->spa_suspend_zio_root);
1112 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1113 ASSERT(zio->io_parent == NULL);
1114 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1115 zio_add_child(spa->spa_suspend_zio_root, zio);
1116 }
1117
1118 mutex_exit(&spa->spa_suspend_lock);
1119 }
1120
1121 void
1122 zio_resume(spa_t *spa)
1123 {
1124 zio_t *pio, *zio;
1125
1126 /*
1127 * Reexecute all previously suspended i/o.
1128 */
1129 mutex_enter(&spa->spa_suspend_lock);
1130 spa->spa_suspended = B_FALSE;
1131 cv_broadcast(&spa->spa_suspend_cv);
1132 pio = spa->spa_suspend_zio_root;
1133 spa->spa_suspend_zio_root = NULL;
1134 mutex_exit(&spa->spa_suspend_lock);
1135
1136 if (pio == NULL)
1137 return;
1138
1139 while ((zio = pio->io_child) != NULL) {
1140 zio_remove_child(pio, zio);
1141 zio->io_parent = NULL;
1142 zio_reexecute(zio);
1143 }
1144
1145 ASSERT(pio->io_children[ZIO_CHILD_LOGICAL][ZIO_WAIT_DONE] == 0);
1146
1147 (void) zio_wait(pio);
1148 }
1149
1150 void
1151 zio_resume_wait(spa_t *spa)
1152 {
1153 mutex_enter(&spa->spa_suspend_lock);
1154 while (spa_suspended(spa))
1155 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1156 mutex_exit(&spa->spa_suspend_lock);
1157 }
1158
1159 /*
1160 * ==========================================================================
1161 * Gang blocks.
1162 *
1163 * A gang block is a collection of small blocks that looks to the DMU
1164 * like one large block. When zio_dva_allocate() cannot find a block
1165 * of the requested size, due to either severe fragmentation or the pool
1166 * being nearly full, it calls zio_write_gang_block() to construct the
1167 * block from smaller fragments.
1168 *
1169 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1170 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1171 * an indirect block: it's an array of block pointers. It consumes
1172 * only one sector and hence is allocatable regardless of fragmentation.
1173 * The gang header's bps point to its gang members, which hold the data.
1174 *
1175 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1176 * as the verifier to ensure uniqueness of the SHA256 checksum.
1177 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1178 * not the gang header. This ensures that data block signatures (needed for
1179 * deduplication) are independent of how the block is physically stored.
1180 *
1181 * Gang blocks can be nested: a gang member may itself be a gang block.
1182 * Thus every gang block is a tree in which root and all interior nodes are
1183 * gang headers, and the leaves are normal blocks that contain user data.
1184 * The root of the gang tree is called the gang leader.
1185 *
1186 * To perform any operation (read, rewrite, free, claim) on a gang block,
1187 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1188 * in the io_gang_tree field of the original logical i/o by recursively
1189 * reading the gang leader and all gang headers below it. This yields
1190 * an in-core tree containing the contents of every gang header and the
1191 * bps for every constituent of the gang block.
1192 *
1193 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1194 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1195 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1196 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1197 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1198 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1199 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1200 * of the gang header plus zio_checksum_compute() of the data to update the
1201 * gang header's blk_cksum as described above.
1202 *
1203 * The two-phase assemble/issue model solves the problem of partial failure --
1204 * what if you'd freed part of a gang block but then couldn't read the
1205 * gang header for another part? Assembling the entire gang tree first
1206 * ensures that all the necessary gang header I/O has succeeded before
1207 * starting the actual work of free, claim, or write. Once the gang tree
1208 * is assembled, free and claim are in-memory operations that cannot fail.
1209 *
1210 * In the event that a gang write fails, zio_dva_unallocate() walks the
1211 * gang tree to immediately free (i.e. insert back into the space map)
1212 * everything we've allocated. This ensures that we don't get ENOSPC
1213 * errors during repeated suspend/resume cycles due to a flaky device.
1214 *
1215 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1216 * the gang tree, we won't modify the block, so we can safely defer the free
1217 * (knowing that the block is still intact). If we *can* assemble the gang
1218 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1219 * each constituent bp and we can allocate a new block on the next sync pass.
1220 *
1221 * In all cases, the gang tree allows complete recovery from partial failure.
1222 * ==========================================================================
1223 */
1224
1225 static zio_t *
1226 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1227 {
1228 if (gn != NULL)
1229 return (pio);
1230
1231 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1232 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1233 &pio->io_bookmark));
1234 }
1235
1236 zio_t *
1237 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1238 {
1239 zio_t *zio;
1240
1241 if (gn != NULL) {
1242 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1243 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1244 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1245 /*
1246 * As we rewrite each gang header, the pipeline will compute
1247 * a new gang block header checksum for it; but no one will
1248 * compute a new data checksum, so we do that here. The one
1249 * exception is the gang leader: the pipeline already computed
1250 * its data checksum because that stage precedes gang assembly.
1251 * (Presently, nothing actually uses interior data checksums;
1252 * this is just good hygiene.)
1253 */
1254 if (gn != pio->io_logical->io_gang_tree) {
1255 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1256 data, BP_GET_PSIZE(bp));
1257 }
1258 } else {
1259 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1260 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1261 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1262 }
1263
1264 return (zio);
1265 }
1266
1267 /* ARGSUSED */
1268 zio_t *
1269 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1270 {
1271 return (zio_free(pio, pio->io_spa, pio->io_txg, bp,
1272 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1273 }
1274
1275 /* ARGSUSED */
1276 zio_t *
1277 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1278 {
1279 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1280 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1281 }
1282
1283 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1284 NULL,
1285 zio_read_gang,
1286 zio_rewrite_gang,
1287 zio_free_gang,
1288 zio_claim_gang,
1289 NULL
1290 };
1291
1292 static void zio_gang_tree_assemble_done(zio_t *zio);
1293
1294 static zio_gang_node_t *
1295 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1296 {
1297 zio_gang_node_t *gn;
1298
1299 ASSERT(*gnpp == NULL);
1300
1301 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1302 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1303 *gnpp = gn;
1304
1305 return (gn);
1306 }
1307
1308 static void
1309 zio_gang_node_free(zio_gang_node_t **gnpp)
1310 {
1311 zio_gang_node_t *gn = *gnpp;
1312
1313 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1314 ASSERT(gn->gn_child[g] == NULL);
1315
1316 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1317 kmem_free(gn, sizeof (*gn));
1318 *gnpp = NULL;
1319 }
1320
1321 static void
1322 zio_gang_tree_free(zio_gang_node_t **gnpp)
1323 {
1324 zio_gang_node_t *gn = *gnpp;
1325
1326 if (gn == NULL)
1327 return;
1328
1329 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1330 zio_gang_tree_free(&gn->gn_child[g]);
1331
1332 zio_gang_node_free(gnpp);
1333 }
1334
1335 static void
1336 zio_gang_tree_assemble(zio_t *lio, blkptr_t *bp, zio_gang_node_t **gnpp)
1337 {
1338 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1339
1340 ASSERT(lio->io_logical == lio);
1341 ASSERT(BP_IS_GANG(bp));
1342
1343 zio_nowait(zio_read(lio, lio->io_spa, bp, gn->gn_gbh,
1344 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1345 lio->io_priority, ZIO_GANG_CHILD_FLAGS(lio), &lio->io_bookmark));
1346 }
1347
1348 static void
1349 zio_gang_tree_assemble_done(zio_t *zio)
1350 {
1351 zio_t *lio = zio->io_logical;
1352 zio_gang_node_t *gn = zio->io_private;
1353 blkptr_t *bp = zio->io_bp;
1354
1355 ASSERT(zio->io_parent == lio);
1356 ASSERT(zio->io_child == NULL);
1357
1358 if (zio->io_error)
1359 return;
1360
1361 if (BP_SHOULD_BYTESWAP(bp))
1362 byteswap_uint64_array(zio->io_data, zio->io_size);
1363
1364 ASSERT(zio->io_data == gn->gn_gbh);
1365 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1366 ASSERT(gn->gn_gbh->zg_tail.zbt_magic == ZBT_MAGIC);
1367
1368 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1369 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1370 if (!BP_IS_GANG(gbp))
1371 continue;
1372 zio_gang_tree_assemble(lio, gbp, &gn->gn_child[g]);
1373 }
1374 }
1375
1376 static void
1377 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1378 {
1379 zio_t *lio = pio->io_logical;
1380 zio_t *zio;
1381
1382 ASSERT(BP_IS_GANG(bp) == !!gn);
1383 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(lio->io_bp));
1384 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == lio->io_gang_tree);
1385
1386 /*
1387 * If you're a gang header, your data is in gn->gn_gbh.
1388 * If you're a gang member, your data is in 'data' and gn == NULL.
1389 */
1390 zio = zio_gang_issue_func[lio->io_type](pio, bp, gn, data);
1391
1392 if (gn != NULL) {
1393 ASSERT(gn->gn_gbh->zg_tail.zbt_magic == ZBT_MAGIC);
1394
1395 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1396 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1397 if (BP_IS_HOLE(gbp))
1398 continue;
1399 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1400 data = (char *)data + BP_GET_PSIZE(gbp);
1401 }
1402 }
1403
1404 if (gn == lio->io_gang_tree)
1405 ASSERT3P((char *)lio->io_data + lio->io_size, ==, data);
1406
1407 if (zio != pio)
1408 zio_nowait(zio);
1409 }
1410
1411 static int
1412 zio_gang_assemble(zio_t *zio)
1413 {
1414 blkptr_t *bp = zio->io_bp;
1415
1416 ASSERT(BP_IS_GANG(bp) && zio == zio->io_logical);
1417
1418 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1419
1420 return (ZIO_PIPELINE_CONTINUE);
1421 }
1422
1423 static int
1424 zio_gang_issue(zio_t *zio)
1425 {
1426 zio_t *lio = zio->io_logical;
1427 blkptr_t *bp = zio->io_bp;
1428
1429 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1430 return (ZIO_PIPELINE_STOP);
1431
1432 ASSERT(BP_IS_GANG(bp) && zio == lio);
1433
1434 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1435 zio_gang_tree_issue(lio, lio->io_gang_tree, bp, lio->io_data);
1436 else
1437 zio_gang_tree_free(&lio->io_gang_tree);
1438
1439 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1440
1441 return (ZIO_PIPELINE_CONTINUE);
1442 }
1443
1444 static void
1445 zio_write_gang_member_ready(zio_t *zio)
1446 {
1447 zio_t *pio = zio->io_parent;
1448 zio_t *lio = zio->io_logical;
1449 dva_t *cdva = zio->io_bp->blk_dva;
1450 dva_t *pdva = pio->io_bp->blk_dva;
1451 uint64_t asize;
1452
1453 if (BP_IS_HOLE(zio->io_bp))
1454 return;
1455
1456 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1457
1458 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1459 ASSERT3U(zio->io_prop.zp_ndvas, ==, lio->io_prop.zp_ndvas);
1460 ASSERT3U(zio->io_prop.zp_ndvas, <=, BP_GET_NDVAS(zio->io_bp));
1461 ASSERT3U(pio->io_prop.zp_ndvas, <=, BP_GET_NDVAS(pio->io_bp));
1462 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1463
1464 mutex_enter(&pio->io_lock);
1465 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1466 ASSERT(DVA_GET_GANG(&pdva[d]));
1467 asize = DVA_GET_ASIZE(&pdva[d]);
1468 asize += DVA_GET_ASIZE(&cdva[d]);
1469 DVA_SET_ASIZE(&pdva[d], asize);
1470 }
1471 mutex_exit(&pio->io_lock);
1472 }
1473
1474 static int
1475 zio_write_gang_block(zio_t *pio)
1476 {
1477 spa_t *spa = pio->io_spa;
1478 blkptr_t *bp = pio->io_bp;
1479 zio_t *lio = pio->io_logical;
1480 zio_t *zio;
1481 zio_gang_node_t *gn, **gnpp;
1482 zio_gbh_phys_t *gbh;
1483 uint64_t txg = pio->io_txg;
1484 uint64_t resid = pio->io_size;
1485 uint64_t lsize;
1486 int ndvas = lio->io_prop.zp_ndvas;
1487 int gbh_ndvas = MIN(ndvas + 1, spa_max_replication(spa));
1488 zio_prop_t zp;
1489 int error;
1490
1491 error = metaslab_alloc(spa, spa->spa_normal_class, SPA_GANGBLOCKSIZE,
1492 bp, gbh_ndvas, txg, pio == lio ? NULL : lio->io_bp,
1493 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1494 if (error) {
1495 pio->io_error = error;
1496 return (ZIO_PIPELINE_CONTINUE);
1497 }
1498
1499 if (pio == lio) {
1500 gnpp = &lio->io_gang_tree;
1501 } else {
1502 gnpp = pio->io_private;
1503 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1504 }
1505
1506 gn = zio_gang_node_alloc(gnpp);
1507 gbh = gn->gn_gbh;
1508 bzero(gbh, SPA_GANGBLOCKSIZE);
1509
1510 /*
1511 * Create the gang header.
1512 */
1513 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1514 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1515
1516 /*
1517 * Create and nowait the gang children.
1518 */
1519 for (int g = 0; resid != 0; resid -= lsize, g++) {
1520 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1521 SPA_MINBLOCKSIZE);
1522 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1523
1524 zp.zp_checksum = lio->io_prop.zp_checksum;
1525 zp.zp_compress = ZIO_COMPRESS_OFF;
1526 zp.zp_type = DMU_OT_NONE;
1527 zp.zp_level = 0;
1528 zp.zp_ndvas = lio->io_prop.zp_ndvas;
1529
1530 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1531 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1532 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1533 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1534 &pio->io_bookmark));
1535 }
1536
1537 /*
1538 * Set pio's pipeline to just wait for zio to finish.
1539 */
1540 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1541
1542 zio_nowait(zio);
1543
1544 return (ZIO_PIPELINE_CONTINUE);
1545 }
1546
1547 /*
1548 * ==========================================================================
1549 * Allocate and free blocks
1550 * ==========================================================================
1551 */
1552
1553 static int
1554 zio_dva_allocate(zio_t *zio)
1555 {
1556 spa_t *spa = zio->io_spa;
1557 metaslab_class_t *mc = spa->spa_normal_class;
1558 blkptr_t *bp = zio->io_bp;
1559 int error;
1560
1561 ASSERT(BP_IS_HOLE(bp));
1562 ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
1563 ASSERT3U(zio->io_prop.zp_ndvas, >, 0);
1564 ASSERT3U(zio->io_prop.zp_ndvas, <=, spa_max_replication(spa));
1565 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
1566
1567 error = metaslab_alloc(spa, mc, zio->io_size, bp,
1568 zio->io_prop.zp_ndvas, zio->io_txg, NULL, 0);
1569
1570 if (error) {
1571 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
1572 return (zio_write_gang_block(zio));
1573 zio->io_error = error;
1574 }
1575
1576 return (ZIO_PIPELINE_CONTINUE);
1577 }
1578
1579 static int
1580 zio_dva_free(zio_t *zio)
1581 {
1582 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
1583
1584 return (ZIO_PIPELINE_CONTINUE);
1585 }
1586
1587 static int
1588 zio_dva_claim(zio_t *zio)
1589 {
1590 int error;
1591
1592 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
1593 if (error)
1594 zio->io_error = error;
1595
1596 return (ZIO_PIPELINE_CONTINUE);
1597 }
1598
1599 /*
1600 * Undo an allocation. This is used by zio_done() when an I/O fails
1601 * and we want to give back the block we just allocated.
1602 * This handles both normal blocks and gang blocks.
1603 */
1604 static void
1605 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
1606 {
1607 spa_t *spa = zio->io_spa;
1608 boolean_t now = !(zio->io_flags & ZIO_FLAG_IO_REWRITE);
1609
1610 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
1611
1612 if (zio->io_bp == bp && !now) {
1613 /*
1614 * This is a rewrite for sync-to-convergence.
1615 * We can't do a metaslab_free(NOW) because bp wasn't allocated
1616 * during this sync pass, which means that metaslab_sync()
1617 * already committed the allocation.
1618 */
1619 ASSERT(DVA_EQUAL(BP_IDENTITY(bp),
1620 BP_IDENTITY(&zio->io_bp_orig)));
1621 ASSERT(spa_sync_pass(spa) > 1);
1622
1623 if (BP_IS_GANG(bp) && gn == NULL) {
1624 /*
1625 * This is a gang leader whose gang header(s) we
1626 * couldn't read now, so defer the free until later.
1627 * The block should still be intact because without
1628 * the headers, we'd never even start the rewrite.
1629 */
1630 bplist_enqueue_deferred(&spa->spa_sync_bplist, bp);
1631 return;
1632 }
1633 }
1634
1635 if (!BP_IS_HOLE(bp))
1636 metaslab_free(spa, bp, bp->blk_birth, now);
1637
1638 if (gn != NULL) {
1639 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1640 zio_dva_unallocate(zio, gn->gn_child[g],
1641 &gn->gn_gbh->zg_blkptr[g]);
1642 }
1643 }
1644 }
1645
1646 /*
1647 * Try to allocate an intent log block. Return 0 on success, errno on failure.
1648 */
1649 int
1650 zio_alloc_blk(spa_t *spa, uint64_t size, blkptr_t *new_bp, blkptr_t *old_bp,
1651 uint64_t txg)
1652 {
1653 int error;
1654
1655 error = metaslab_alloc(spa, spa->spa_log_class, size,
1656 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
1657
1658 if (error)
1659 error = metaslab_alloc(spa, spa->spa_normal_class, size,
1660 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
1661
1662 if (error == 0) {
1663 BP_SET_LSIZE(new_bp, size);
1664 BP_SET_PSIZE(new_bp, size);
1665 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
1666 BP_SET_CHECKSUM(new_bp, ZIO_CHECKSUM_ZILOG);
1667 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
1668 BP_SET_LEVEL(new_bp, 0);
1669 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
1670 }
1671
1672 return (error);
1673 }
1674
1675 /*
1676 * Free an intent log block. We know it can't be a gang block, so there's
1677 * nothing to do except metaslab_free() it.
1678 */
1679 void
1680 zio_free_blk(spa_t *spa, blkptr_t *bp, uint64_t txg)
1681 {
1682 ASSERT(!BP_IS_GANG(bp));
1683
1684 metaslab_free(spa, bp, txg, B_FALSE);
1685 }
1686
1687 /*
1688 * ==========================================================================
1689 * Read and write to physical devices
1690 * ==========================================================================
1691 */
1692
1693 static void
1694 zio_vdev_io_probe_done(zio_t *zio)
1695 {
1696 zio_t *dio;
1697 vdev_t *vd = zio->io_private;
1698
1699 mutex_enter(&vd->vdev_probe_lock);
1700 ASSERT(vd->vdev_probe_zio == zio);
1701 vd->vdev_probe_zio = NULL;
1702 mutex_exit(&vd->vdev_probe_lock);
1703
1704 while ((dio = zio->io_delegate_list) != NULL) {
1705 zio->io_delegate_list = dio->io_delegate_next;
1706 dio->io_delegate_next = NULL;
1707 if (!vdev_accessible(vd, dio))
1708 dio->io_error = ENXIO;
1709 zio_execute(dio);
1710 }
1711 }
1712
1713 /*
1714 * Probe the device to determine whether I/O failure is specific to this
1715 * zio (e.g. a bad sector) or affects the entire vdev (e.g. unplugged).
1716 */
1717 static int
1718 zio_vdev_io_probe(zio_t *zio)
1719 {
1720 vdev_t *vd = zio->io_vd;
1721 zio_t *pio = NULL;
1722 boolean_t created_pio = B_FALSE;
1723
1724 /*
1725 * Don't probe the probe.
1726 */
1727 if (zio->io_flags & ZIO_FLAG_PROBE)
1728 return (ZIO_PIPELINE_CONTINUE);
1729
1730 /*
1731 * To prevent 'probe storms' when a device fails, we create
1732 * just one probe i/o at a time. All zios that want to probe
1733 * this vdev will join the probe zio's io_delegate_list.
1734 */
1735 mutex_enter(&vd->vdev_probe_lock);
1736
1737 if ((pio = vd->vdev_probe_zio) == NULL) {
1738 vd->vdev_probe_zio = pio = zio_root(zio->io_spa,
1739 zio_vdev_io_probe_done, vd, ZIO_FLAG_CANFAIL);
1740 created_pio = B_TRUE;
1741 vd->vdev_probe_wanted = B_TRUE;
1742 spa_async_request(zio->io_spa, SPA_ASYNC_PROBE);
1743 }
1744
1745 zio->io_delegate_next = pio->io_delegate_list;
1746 pio->io_delegate_list = zio;
1747
1748 mutex_exit(&vd->vdev_probe_lock);
1749
1750 if (created_pio) {
1751 zio_nowait(vdev_probe(vd, pio));
1752 zio_nowait(pio);
1753 }
1754
1755 return (ZIO_PIPELINE_STOP);
1756 }
1757
1758 static int
1759 zio_vdev_io_start(zio_t *zio)
1760 {
1761 vdev_t *vd = zio->io_vd;
1762 uint64_t align;
1763 spa_t *spa = zio->io_spa;
1764
1765 ASSERT(zio->io_error == 0);
1766 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
1767
1768 if (vd == NULL) {
1769 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
1770 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
1771
1772 /*
1773 * The mirror_ops handle multiple DVAs in a single BP.
1774 */
1775 return (vdev_mirror_ops.vdev_op_io_start(zio));
1776 }
1777
1778 align = 1ULL << vd->vdev_top->vdev_ashift;
1779
1780 if (P2PHASE(zio->io_size, align) != 0) {
1781 uint64_t asize = P2ROUNDUP(zio->io_size, align);
1782 char *abuf = zio_buf_alloc(asize);
1783 ASSERT(vd == vd->vdev_top);
1784 if (zio->io_type == ZIO_TYPE_WRITE) {
1785 bcopy(zio->io_data, abuf, zio->io_size);
1786 bzero(abuf + zio->io_size, asize - zio->io_size);
1787 }
1788 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
1789 }
1790
1791 ASSERT(P2PHASE(zio->io_offset, align) == 0);
1792 ASSERT(P2PHASE(zio->io_size, align) == 0);
1793 ASSERT(zio->io_type != ZIO_TYPE_WRITE || (spa_mode & FWRITE));
1794
1795 if (vd->vdev_ops->vdev_op_leaf &&
1796 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
1797
1798 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
1799 return (ZIO_PIPELINE_STOP);
1800
1801 if ((zio = vdev_queue_io(zio)) == NULL)
1802 return (ZIO_PIPELINE_STOP);
1803
1804 if (!vdev_accessible(vd, zio)) {
1805 zio->io_error = ENXIO;
1806 zio_interrupt(zio);
1807 return (ZIO_PIPELINE_STOP);
1808 }
1809
1810 }
1811
1812 return (vd->vdev_ops->vdev_op_io_start(zio));
1813 }
1814
1815 static int
1816 zio_vdev_io_done(zio_t *zio)
1817 {
1818 vdev_t *vd = zio->io_vd;
1819 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
1820 boolean_t unexpected_error = B_FALSE;
1821
1822 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
1823 return (ZIO_PIPELINE_STOP);
1824
1825 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
1826
1827 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
1828
1829 vdev_queue_io_done(zio);
1830
1831 if (zio->io_type == ZIO_TYPE_WRITE)
1832 vdev_cache_write(zio);
1833
1834 if (zio_injection_enabled && zio->io_error == 0)
1835 zio->io_error = zio_handle_device_injection(vd, EIO);
1836
1837 if (zio_injection_enabled && zio->io_error == 0)
1838 zio->io_error = zio_handle_label_injection(zio, EIO);
1839
1840 if (zio->io_error) {
1841 if (!vdev_accessible(vd, zio)) {
1842 zio->io_error = ENXIO;
1843 } else {
1844 unexpected_error = B_TRUE;
1845 }
1846 }
1847 }
1848
1849 ops->vdev_op_io_done(zio);
1850
1851 if (unexpected_error)
1852 return (zio_vdev_io_probe(zio));
1853
1854 return (ZIO_PIPELINE_CONTINUE);
1855 }
1856
1857 static int
1858 zio_vdev_io_assess(zio_t *zio)
1859 {
1860 vdev_t *vd = zio->io_vd;
1861
1862 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
1863 return (ZIO_PIPELINE_STOP);
1864
1865 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
1866 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
1867
1868 if (zio->io_vsd != NULL) {
1869 zio->io_vsd_free(zio);
1870 zio->io_vsd = NULL;
1871 }
1872
1873 if (zio_injection_enabled && zio->io_error == 0)
1874 zio->io_error = zio_handle_fault_injection(zio, EIO);
1875
1876 /*
1877 * If the I/O failed, determine whether we should attempt to retry it.
1878 */
1879 if (zio->io_error && vd == NULL &&
1880 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
1881 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
1882 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
1883 zio->io_error = 0;
1884 zio->io_flags |= ZIO_FLAG_IO_RETRY |
1885 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
1886 zio->io_stage = ZIO_STAGE_VDEV_IO_START - 1;
1887 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE);
1888 return (ZIO_PIPELINE_STOP);
1889 }
1890
1891 /*
1892 * If we got an error on a leaf device, convert it to ENXIO
1893 * if the device is not accessible at all.
1894 */
1895 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
1896 !vdev_accessible(vd, zio))
1897 zio->io_error = ENXIO;
1898
1899 /*
1900 * If we can't write to an interior vdev (mirror or RAID-Z),
1901 * set vdev_cant_write so that we stop trying to allocate from it.
1902 */
1903 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
1904 vd != NULL && !vd->vdev_ops->vdev_op_leaf)
1905 vd->vdev_cant_write = B_TRUE;
1906
1907 if (zio->io_error)
1908 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1909
1910 return (ZIO_PIPELINE_CONTINUE);
1911 }
1912
1913 void
1914 zio_vdev_io_reissue(zio_t *zio)
1915 {
1916 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
1917 ASSERT(zio->io_error == 0);
1918
1919 zio->io_stage--;
1920 }
1921
1922 void
1923 zio_vdev_io_redone(zio_t *zio)
1924 {
1925 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
1926
1927 zio->io_stage--;
1928 }
1929
1930 void
1931 zio_vdev_io_bypass(zio_t *zio)
1932 {
1933 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
1934 ASSERT(zio->io_error == 0);
1935
1936 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
1937 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS - 1;
1938 }
1939
1940 /*
1941 * ==========================================================================
1942 * Generate and verify checksums
1943 * ==========================================================================
1944 */
1945 static int
1946 zio_checksum_generate(zio_t *zio)
1947 {
1948 blkptr_t *bp = zio->io_bp;
1949 enum zio_checksum checksum;
1950
1951 if (bp == NULL) {
1952 /*
1953 * This is zio_write_phys().
1954 * We're either generating a label checksum, or none at all.
1955 */
1956 checksum = zio->io_prop.zp_checksum;
1957
1958 if (checksum == ZIO_CHECKSUM_OFF)
1959 return (ZIO_PIPELINE_CONTINUE);
1960
1961 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
1962 } else {
1963 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
1964 ASSERT(!IO_IS_ALLOCATING(zio));
1965 checksum = ZIO_CHECKSUM_GANG_HEADER;
1966 } else {
1967 checksum = BP_GET_CHECKSUM(bp);
1968 }
1969 }
1970
1971 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
1972
1973 return (ZIO_PIPELINE_CONTINUE);
1974 }
1975
1976 static int
1977 zio_checksum_verify(zio_t *zio)
1978 {
1979 blkptr_t *bp = zio->io_bp;
1980 int error;
1981
1982 if (bp == NULL) {
1983 /*
1984 * This is zio_read_phys().
1985 * We're either verifying a label checksum, or nothing at all.
1986 */
1987 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
1988 return (ZIO_PIPELINE_CONTINUE);
1989
1990 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
1991 }
1992
1993 if ((error = zio_checksum_error(zio)) != 0) {
1994 zio->io_error = error;
1995 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
1996 zfs_ereport_post(FM_EREPORT_ZFS_CHECKSUM,
1997 zio->io_spa, zio->io_vd, zio, 0, 0);
1998 }
1999 }
2000
2001 return (ZIO_PIPELINE_CONTINUE);
2002 }
2003
2004 /*
2005 * Called by RAID-Z to ensure we don't compute the checksum twice.
2006 */
2007 void
2008 zio_checksum_verified(zio_t *zio)
2009 {
2010 zio->io_pipeline &= ~(1U << ZIO_STAGE_CHECKSUM_VERIFY);
2011 }
2012
2013 /*
2014 * ==========================================================================
2015 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2016 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2017 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2018 * indicate errors that are specific to one I/O, and most likely permanent.
2019 * Any other error is presumed to be worse because we weren't expecting it.
2020 * ==========================================================================
2021 */
2022 int
2023 zio_worst_error(int e1, int e2)
2024 {
2025 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2026 int r1, r2;
2027
2028 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2029 if (e1 == zio_error_rank[r1])
2030 break;
2031
2032 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2033 if (e2 == zio_error_rank[r2])
2034 break;
2035
2036 return (r1 > r2 ? e1 : e2);
2037 }
2038
2039 /*
2040 * ==========================================================================
2041 * I/O completion
2042 * ==========================================================================
2043 */
2044 static int
2045 zio_ready(zio_t *zio)
2046 {
2047 blkptr_t *bp = zio->io_bp;
2048 zio_t *pio = zio->io_parent;
2049
2050 if (zio->io_ready) {
2051 if (BP_IS_GANG(bp) &&
2052 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY))
2053 return (ZIO_PIPELINE_STOP);
2054
2055 ASSERT(IO_IS_ALLOCATING(zio));
2056 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2057 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2058
2059 zio->io_ready(zio);
2060 }
2061
2062 if (bp != NULL && bp != &zio->io_bp_copy)
2063 zio->io_bp_copy = *bp;
2064
2065 if (zio->io_error)
2066 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2067
2068 if (pio != NULL)
2069 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2070
2071 return (ZIO_PIPELINE_CONTINUE);
2072 }
2073
2074 static int
2075 zio_done(zio_t *zio)
2076 {
2077 spa_t *spa = zio->io_spa;
2078 zio_t *pio = zio->io_parent;
2079 zio_t *lio = zio->io_logical;
2080 blkptr_t *bp = zio->io_bp;
2081 vdev_t *vd = zio->io_vd;
2082 uint64_t psize = zio->io_size;
2083
2084 /*
2085 * If our of children haven't all completed,
2086 * wait for them and then repeat this pipeline stage.
2087 */
2088 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2089 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2090 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2091 return (ZIO_PIPELINE_STOP);
2092
2093 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2094 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2095 ASSERT(zio->io_children[c][w] == 0);
2096
2097 if (bp != NULL) {
2098 ASSERT(bp->blk_pad[0] == 0);
2099 ASSERT(bp->blk_pad[1] == 0);
2100 ASSERT(bp->blk_pad[2] == 0);
2101 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2102 (pio != NULL && bp == pio->io_bp));
2103 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2104 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2105 ASSERT(!BP_SHOULD_BYTESWAP(bp));
2106 ASSERT3U(zio->io_prop.zp_ndvas, <=, BP_GET_NDVAS(bp));
2107 ASSERT(BP_COUNT_GANG(bp) == 0 ||
2108 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2109 }
2110 }
2111
2112 /*
2113 * If there were child vdev or gang errors, they apply to us now.
2114 */
2115 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2116 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2117
2118 zio_pop_transforms(zio); /* note: may set zio->io_error */
2119
2120 vdev_stat_update(zio, psize);
2121
2122 if (zio->io_error) {
2123 /*
2124 * If this I/O is attached to a particular vdev,
2125 * generate an error message describing the I/O failure
2126 * at the block level. We ignore these errors if the
2127 * device is currently unavailable.
2128 */
2129 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
2130 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
2131
2132 if ((zio->io_error == EIO ||
2133 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) && zio == lio) {
2134 /*
2135 * For logical I/O requests, tell the SPA to log the
2136 * error and generate a logical data ereport.
2137 */
2138 spa_log_error(spa, zio);
2139 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
2140 0, 0);
2141 }
2142 }
2143
2144 if (zio->io_error && zio == lio) {
2145 /*
2146 * Determine whether zio should be reexecuted. This will
2147 * propagate all the way to the root via zio_notify_parent().
2148 */
2149 ASSERT(vd == NULL && bp != NULL);
2150
2151 if (IO_IS_ALLOCATING(zio))
2152 if (zio->io_error != ENOSPC)
2153 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2154 else
2155 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2156
2157 if ((zio->io_type == ZIO_TYPE_READ ||
2158 zio->io_type == ZIO_TYPE_FREE) &&
2159 zio->io_error == ENXIO &&
2160 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
2161 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2162
2163 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2164 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2165 }
2166
2167 /*
2168 * If there were logical child errors, they apply to us now.
2169 * We defer this until now to avoid conflating logical child
2170 * errors with errors that happened to the zio itself when
2171 * updating vdev stats and reporting FMA events above.
2172 */
2173 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2174
2175 if (zio->io_reexecute) {
2176 /*
2177 * This is a logical I/O that wants to reexecute.
2178 *
2179 * Reexecute is top-down. When an i/o fails, if it's not
2180 * the root, it simply notifies its parent and sticks around.
2181 * The parent, seeing that it still has children in zio_done(),
2182 * does the same. This percolates all the way up to the root.
2183 * The root i/o will reexecute or suspend the entire tree.
2184 *
2185 * This approach ensures that zio_reexecute() honors
2186 * all the original i/o dependency relationships, e.g.
2187 * parents not executing until children are ready.
2188 */
2189 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2190
2191 if (IO_IS_ALLOCATING(zio))
2192 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
2193
2194 zio_gang_tree_free(&zio->io_gang_tree);
2195
2196 if (pio != NULL) {
2197 /*
2198 * We're not a root i/o, so there's nothing to do
2199 * but notify our parent. Don't propagate errors
2200 * upward since we haven't permanently failed yet.
2201 */
2202 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
2203 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2204 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
2205 /*
2206 * We'd fail again if we reexecuted now, so suspend
2207 * until conditions improve (e.g. device comes online).
2208 */
2209 zio_suspend(spa, zio);
2210 } else {
2211 /*
2212 * Reexecution is potentially a huge amount of work.
2213 * Hand it off to the otherwise-unused claim taskq.
2214 */
2215 (void) taskq_dispatch(
2216 spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2217 (task_func_t *)zio_reexecute, zio, TQ_SLEEP);
2218 }
2219 return (ZIO_PIPELINE_STOP);
2220 }
2221
2222 ASSERT(zio->io_child == NULL);
2223 ASSERT(zio->io_reexecute == 0);
2224 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
2225
2226 if (zio->io_done)
2227 zio->io_done(zio);
2228
2229 zio_gang_tree_free(&zio->io_gang_tree);
2230
2231 ASSERT(zio->io_delegate_list == NULL);
2232 ASSERT(zio->io_delegate_next == NULL);
2233
2234 if (pio != NULL) {
2235 zio_remove_child(pio, zio);
2236 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2237 }
2238
2239 if (zio->io_waiter != NULL) {
2240 mutex_enter(&zio->io_lock);
2241 zio->io_executor = NULL;
2242 cv_broadcast(&zio->io_cv);
2243 mutex_exit(&zio->io_lock);
2244 } else {
2245 zio_destroy(zio);
2246 }
2247
2248 return (ZIO_PIPELINE_STOP);
2249 }
2250
2251 /*
2252 * ==========================================================================
2253 * I/O pipeline definition
2254 * ==========================================================================
2255 */
2256 static zio_pipe_stage_t *zio_pipeline[ZIO_STAGES] = {
2257 NULL,
2258 zio_issue_async,
2259 zio_read_bp_init,
2260 zio_write_bp_init,
2261 zio_checksum_generate,
2262 zio_gang_assemble,
2263 zio_gang_issue,
2264 zio_dva_allocate,
2265 zio_dva_free,
2266 zio_dva_claim,
2267 zio_ready,
2268 zio_vdev_io_start,
2269 zio_vdev_io_done,
2270 zio_vdev_io_assess,
2271 zio_checksum_verify,
2272 zio_done
2273 };
Mirror of ZFSOnLinux for Debian