]> 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
Illumos #1909: disk sync write perf regression when slog is used post oi_148
[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 (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25 */
26
27 #include <sys/zfs_context.h>
28 #include <sys/fm/fs/zfs.h>
29 #include <sys/spa.h>
30 #include <sys/txg.h>
31 #include <sys/spa_impl.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/zio_impl.h>
34 #include <sys/zio_compress.h>
35 #include <sys/zio_checksum.h>
36 #include <sys/dmu_objset.h>
37 #include <sys/arc.h>
38 #include <sys/ddt.h>
39
40 /*
41 * ==========================================================================
42 * I/O priority table
43 * ==========================================================================
44 */
45 uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = {
46 0, /* ZIO_PRIORITY_NOW */
47 0, /* ZIO_PRIORITY_SYNC_READ */
48 0, /* ZIO_PRIORITY_SYNC_WRITE */
49 0, /* ZIO_PRIORITY_LOG_WRITE */
50 1, /* ZIO_PRIORITY_CACHE_FILL */
51 1, /* ZIO_PRIORITY_AGG */
52 4, /* ZIO_PRIORITY_FREE */
53 4, /* ZIO_PRIORITY_ASYNC_WRITE */
54 6, /* ZIO_PRIORITY_ASYNC_READ */
55 10, /* ZIO_PRIORITY_RESILVER */
56 20, /* ZIO_PRIORITY_SCRUB */
57 2, /* ZIO_PRIORITY_DDT_PREFETCH */
58 };
59
60 /*
61 * ==========================================================================
62 * I/O type descriptions
63 * ==========================================================================
64 */
65 char *zio_type_name[ZIO_TYPES] = {
66 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl"
67 };
68
69 /*
70 * ==========================================================================
71 * I/O kmem caches
72 * ==========================================================================
73 */
74 kmem_cache_t *zio_cache;
75 kmem_cache_t *zio_link_cache;
76 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
77 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
78 int zio_bulk_flags = 0;
79 int zio_delay_max = ZIO_DELAY_MAX;
80
81 #ifdef _KERNEL
82 extern vmem_t *zio_alloc_arena;
83 #endif
84 extern int zfs_mg_alloc_failures;
85
86 /*
87 * An allocating zio is one that either currently has the DVA allocate
88 * stage set or will have it later in its lifetime.
89 */
90 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
91
92 int zio_requeue_io_start_cut_in_line = 1;
93
94 #ifdef ZFS_DEBUG
95 int zio_buf_debug_limit = 16384;
96 #else
97 int zio_buf_debug_limit = 0;
98 #endif
99
100 static inline void __zio_execute(zio_t *zio);
101
102 static int
103 zio_cons(void *arg, void *unused, int kmflag)
104 {
105 zio_t *zio = arg;
106
107 bzero(zio, sizeof (zio_t));
108
109 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
110 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
111
112 list_create(&zio->io_parent_list, sizeof (zio_link_t),
113 offsetof(zio_link_t, zl_parent_node));
114 list_create(&zio->io_child_list, sizeof (zio_link_t),
115 offsetof(zio_link_t, zl_child_node));
116
117 return (0);
118 }
119
120 static void
121 zio_dest(void *arg, void *unused)
122 {
123 zio_t *zio = arg;
124
125 mutex_destroy(&zio->io_lock);
126 cv_destroy(&zio->io_cv);
127 list_destroy(&zio->io_parent_list);
128 list_destroy(&zio->io_child_list);
129 }
130
131 void
132 zio_init(void)
133 {
134 size_t c;
135 vmem_t *data_alloc_arena = NULL;
136
137 #ifdef _KERNEL
138 data_alloc_arena = zio_alloc_arena;
139 #endif
140 zio_cache = kmem_cache_create("zio_cache", sizeof (zio_t), 0,
141 zio_cons, zio_dest, NULL, NULL, NULL, KMC_KMEM);
142 zio_link_cache = kmem_cache_create("zio_link_cache",
143 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, KMC_KMEM);
144
145 /*
146 * For small buffers, we want a cache for each multiple of
147 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
148 * for each quarter-power of 2. For large buffers, we want
149 * a cache for each multiple of PAGESIZE.
150 */
151 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
152 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
153 size_t p2 = size;
154 size_t align = 0;
155
156 while (p2 & (p2 - 1))
157 p2 &= p2 - 1;
158
159 if (size <= 4 * SPA_MINBLOCKSIZE) {
160 align = SPA_MINBLOCKSIZE;
161 } else if (P2PHASE(size, PAGESIZE) == 0) {
162 align = PAGESIZE;
163 } else if (P2PHASE(size, p2 >> 2) == 0) {
164 align = p2 >> 2;
165 }
166
167 if (align != 0) {
168 char name[36];
169 int flags = zio_bulk_flags;
170
171 /*
172 * The smallest buffers (512b) are heavily used and
173 * experience a lot of churn. The slabs allocated
174 * for them are also relatively small (32K). Thus
175 * in over to avoid expensive calls to vmalloc() we
176 * make an exception to the usual slab allocation
177 * policy and force these buffers to be kmem backed.
178 */
179 if (size == (1 << SPA_MINBLOCKSHIFT))
180 flags |= KMC_KMEM;
181
182 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
183 zio_buf_cache[c] = kmem_cache_create(name, size,
184 align, NULL, NULL, NULL, NULL, NULL, flags);
185
186 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
187 zio_data_buf_cache[c] = kmem_cache_create(name, size,
188 align, NULL, NULL, NULL, NULL,
189 data_alloc_arena, flags);
190 }
191 }
192
193 while (--c != 0) {
194 ASSERT(zio_buf_cache[c] != NULL);
195 if (zio_buf_cache[c - 1] == NULL)
196 zio_buf_cache[c - 1] = zio_buf_cache[c];
197
198 ASSERT(zio_data_buf_cache[c] != NULL);
199 if (zio_data_buf_cache[c - 1] == NULL)
200 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
201 }
202
203 /*
204 * The zio write taskqs have 1 thread per cpu, allow 1/2 of the taskqs
205 * to fail 3 times per txg or 8 failures, whichever is greater.
206 */
207 zfs_mg_alloc_failures = MAX((3 * max_ncpus / 2), 8);
208
209 zio_inject_init();
210 }
211
212 void
213 zio_fini(void)
214 {
215 size_t c;
216 kmem_cache_t *last_cache = NULL;
217 kmem_cache_t *last_data_cache = NULL;
218
219 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
220 if (zio_buf_cache[c] != last_cache) {
221 last_cache = zio_buf_cache[c];
222 kmem_cache_destroy(zio_buf_cache[c]);
223 }
224 zio_buf_cache[c] = NULL;
225
226 if (zio_data_buf_cache[c] != last_data_cache) {
227 last_data_cache = zio_data_buf_cache[c];
228 kmem_cache_destroy(zio_data_buf_cache[c]);
229 }
230 zio_data_buf_cache[c] = NULL;
231 }
232
233 kmem_cache_destroy(zio_link_cache);
234 kmem_cache_destroy(zio_cache);
235
236 zio_inject_fini();
237 }
238
239 /*
240 * ==========================================================================
241 * Allocate and free I/O buffers
242 * ==========================================================================
243 */
244
245 /*
246 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
247 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
248 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
249 * excess / transient data in-core during a crashdump.
250 */
251 void *
252 zio_buf_alloc(size_t size)
253 {
254 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
255
256 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
257
258 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
259 }
260
261 /*
262 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
263 * crashdump if the kernel panics. This exists so that we will limit the amount
264 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
265 * of kernel heap dumped to disk when the kernel panics)
266 */
267 void *
268 zio_data_buf_alloc(size_t size)
269 {
270 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
271
272 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
273
274 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
275 }
276
277 void
278 zio_buf_free(void *buf, size_t size)
279 {
280 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
281
282 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
283
284 kmem_cache_free(zio_buf_cache[c], buf);
285 }
286
287 void
288 zio_data_buf_free(void *buf, size_t size)
289 {
290 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
291
292 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
293
294 kmem_cache_free(zio_data_buf_cache[c], buf);
295 }
296
297 /*
298 * ==========================================================================
299 * Push and pop I/O transform buffers
300 * ==========================================================================
301 */
302 static void
303 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
304 zio_transform_func_t *transform)
305 {
306 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_PUSHPAGE);
307
308 zt->zt_orig_data = zio->io_data;
309 zt->zt_orig_size = zio->io_size;
310 zt->zt_bufsize = bufsize;
311 zt->zt_transform = transform;
312
313 zt->zt_next = zio->io_transform_stack;
314 zio->io_transform_stack = zt;
315
316 zio->io_data = data;
317 zio->io_size = size;
318 }
319
320 static void
321 zio_pop_transforms(zio_t *zio)
322 {
323 zio_transform_t *zt;
324
325 while ((zt = zio->io_transform_stack) != NULL) {
326 if (zt->zt_transform != NULL)
327 zt->zt_transform(zio,
328 zt->zt_orig_data, zt->zt_orig_size);
329
330 if (zt->zt_bufsize != 0)
331 zio_buf_free(zio->io_data, zt->zt_bufsize);
332
333 zio->io_data = zt->zt_orig_data;
334 zio->io_size = zt->zt_orig_size;
335 zio->io_transform_stack = zt->zt_next;
336
337 kmem_free(zt, sizeof (zio_transform_t));
338 }
339 }
340
341 /*
342 * ==========================================================================
343 * I/O transform callbacks for subblocks and decompression
344 * ==========================================================================
345 */
346 static void
347 zio_subblock(zio_t *zio, void *data, uint64_t size)
348 {
349 ASSERT(zio->io_size > size);
350
351 if (zio->io_type == ZIO_TYPE_READ)
352 bcopy(zio->io_data, data, size);
353 }
354
355 static void
356 zio_decompress(zio_t *zio, void *data, uint64_t size)
357 {
358 if (zio->io_error == 0 &&
359 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
360 zio->io_data, data, zio->io_size, size) != 0)
361 zio->io_error = EIO;
362 }
363
364 /*
365 * ==========================================================================
366 * I/O parent/child relationships and pipeline interlocks
367 * ==========================================================================
368 */
369 /*
370 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
371 * continue calling these functions until they return NULL.
372 * Otherwise, the next caller will pick up the list walk in
373 * some indeterminate state. (Otherwise every caller would
374 * have to pass in a cookie to keep the state represented by
375 * io_walk_link, which gets annoying.)
376 */
377 zio_t *
378 zio_walk_parents(zio_t *cio)
379 {
380 zio_link_t *zl = cio->io_walk_link;
381 list_t *pl = &cio->io_parent_list;
382
383 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
384 cio->io_walk_link = zl;
385
386 if (zl == NULL)
387 return (NULL);
388
389 ASSERT(zl->zl_child == cio);
390 return (zl->zl_parent);
391 }
392
393 zio_t *
394 zio_walk_children(zio_t *pio)
395 {
396 zio_link_t *zl = pio->io_walk_link;
397 list_t *cl = &pio->io_child_list;
398
399 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
400 pio->io_walk_link = zl;
401
402 if (zl == NULL)
403 return (NULL);
404
405 ASSERT(zl->zl_parent == pio);
406 return (zl->zl_child);
407 }
408
409 zio_t *
410 zio_unique_parent(zio_t *cio)
411 {
412 zio_t *pio = zio_walk_parents(cio);
413
414 VERIFY(zio_walk_parents(cio) == NULL);
415 return (pio);
416 }
417
418 void
419 zio_add_child(zio_t *pio, zio_t *cio)
420 {
421 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_PUSHPAGE);
422 int w;
423
424 /*
425 * Logical I/Os can have logical, gang, or vdev children.
426 * Gang I/Os can have gang or vdev children.
427 * Vdev I/Os can only have vdev children.
428 * The following ASSERT captures all of these constraints.
429 */
430 ASSERT(cio->io_child_type <= pio->io_child_type);
431
432 zl->zl_parent = pio;
433 zl->zl_child = cio;
434
435 mutex_enter(&cio->io_lock);
436 mutex_enter(&pio->io_lock);
437
438 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
439
440 for (w = 0; w < ZIO_WAIT_TYPES; w++)
441 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
442
443 list_insert_head(&pio->io_child_list, zl);
444 list_insert_head(&cio->io_parent_list, zl);
445
446 pio->io_child_count++;
447 cio->io_parent_count++;
448
449 mutex_exit(&pio->io_lock);
450 mutex_exit(&cio->io_lock);
451 }
452
453 static void
454 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
455 {
456 ASSERT(zl->zl_parent == pio);
457 ASSERT(zl->zl_child == cio);
458
459 mutex_enter(&cio->io_lock);
460 mutex_enter(&pio->io_lock);
461
462 list_remove(&pio->io_child_list, zl);
463 list_remove(&cio->io_parent_list, zl);
464
465 pio->io_child_count--;
466 cio->io_parent_count--;
467
468 mutex_exit(&pio->io_lock);
469 mutex_exit(&cio->io_lock);
470
471 kmem_cache_free(zio_link_cache, zl);
472 }
473
474 static boolean_t
475 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
476 {
477 uint64_t *countp = &zio->io_children[child][wait];
478 boolean_t waiting = B_FALSE;
479
480 mutex_enter(&zio->io_lock);
481 ASSERT(zio->io_stall == NULL);
482 if (*countp != 0) {
483 zio->io_stage >>= 1;
484 zio->io_stall = countp;
485 waiting = B_TRUE;
486 }
487 mutex_exit(&zio->io_lock);
488
489 return (waiting);
490 }
491
492 __attribute__((always_inline))
493 static inline void
494 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
495 {
496 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
497 int *errorp = &pio->io_child_error[zio->io_child_type];
498
499 mutex_enter(&pio->io_lock);
500 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
501 *errorp = zio_worst_error(*errorp, zio->io_error);
502 pio->io_reexecute |= zio->io_reexecute;
503 ASSERT3U(*countp, >, 0);
504 if (--*countp == 0 && pio->io_stall == countp) {
505 pio->io_stall = NULL;
506 mutex_exit(&pio->io_lock);
507 __zio_execute(pio);
508 } else {
509 mutex_exit(&pio->io_lock);
510 }
511 }
512
513 static void
514 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
515 {
516 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
517 zio->io_error = zio->io_child_error[c];
518 }
519
520 /*
521 * ==========================================================================
522 * Create the various types of I/O (read, write, free, etc)
523 * ==========================================================================
524 */
525 static zio_t *
526 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
527 void *data, uint64_t size, zio_done_func_t *done, void *private,
528 zio_type_t type, int priority, enum zio_flag flags,
529 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
530 enum zio_stage stage, enum zio_stage pipeline)
531 {
532 zio_t *zio;
533
534 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
535 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
536 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
537
538 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
539 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
540 ASSERT(vd || stage == ZIO_STAGE_OPEN);
541
542 zio = kmem_cache_alloc(zio_cache, KM_PUSHPAGE);
543
544 if (vd != NULL)
545 zio->io_child_type = ZIO_CHILD_VDEV;
546 else if (flags & ZIO_FLAG_GANG_CHILD)
547 zio->io_child_type = ZIO_CHILD_GANG;
548 else if (flags & ZIO_FLAG_DDT_CHILD)
549 zio->io_child_type = ZIO_CHILD_DDT;
550 else
551 zio->io_child_type = ZIO_CHILD_LOGICAL;
552
553 if (bp != NULL) {
554 zio->io_logical = NULL;
555 zio->io_bp = (blkptr_t *)bp;
556 zio->io_bp_copy = *bp;
557 zio->io_bp_orig = *bp;
558 if (type != ZIO_TYPE_WRITE ||
559 zio->io_child_type == ZIO_CHILD_DDT)
560 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
561 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
562 zio->io_logical = zio;
563 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
564 pipeline |= ZIO_GANG_STAGES;
565 } else {
566 zio->io_logical = NULL;
567 zio->io_bp = NULL;
568 bzero(&zio->io_bp_copy, sizeof (blkptr_t));
569 bzero(&zio->io_bp_orig, sizeof (blkptr_t));
570 }
571
572 zio->io_spa = spa;
573 zio->io_txg = txg;
574 zio->io_ready = NULL;
575 zio->io_done = done;
576 zio->io_private = private;
577 zio->io_prev_space_delta = 0;
578 zio->io_type = type;
579 zio->io_priority = priority;
580 zio->io_vd = vd;
581 zio->io_vsd = NULL;
582 zio->io_vsd_ops = NULL;
583 zio->io_offset = offset;
584 zio->io_deadline = 0;
585 zio->io_orig_data = zio->io_data = data;
586 zio->io_orig_size = zio->io_size = size;
587 zio->io_orig_flags = zio->io_flags = flags;
588 zio->io_orig_stage = zio->io_stage = stage;
589 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
590 bzero(&zio->io_prop, sizeof (zio_prop_t));
591 zio->io_cmd = 0;
592 zio->io_reexecute = 0;
593 zio->io_bp_override = NULL;
594 zio->io_walk_link = NULL;
595 zio->io_transform_stack = NULL;
596 zio->io_delay = 0;
597 zio->io_error = 0;
598 zio->io_child_count = 0;
599 zio->io_parent_count = 0;
600 zio->io_stall = NULL;
601 zio->io_gang_leader = NULL;
602 zio->io_gang_tree = NULL;
603 zio->io_executor = NULL;
604 zio->io_waiter = NULL;
605 zio->io_cksum_report = NULL;
606 zio->io_ena = 0;
607 bzero(zio->io_child_error, sizeof (int) * ZIO_CHILD_TYPES);
608 bzero(zio->io_children,
609 sizeof (uint64_t) * ZIO_CHILD_TYPES * ZIO_WAIT_TYPES);
610 bzero(&zio->io_bookmark, sizeof (zbookmark_t));
611
612 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
613 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
614
615 if (zb != NULL)
616 zio->io_bookmark = *zb;
617
618 if (pio != NULL) {
619 if (zio->io_logical == NULL)
620 zio->io_logical = pio->io_logical;
621 if (zio->io_child_type == ZIO_CHILD_GANG)
622 zio->io_gang_leader = pio->io_gang_leader;
623 zio_add_child(pio, zio);
624 }
625
626 taskq_init_ent(&zio->io_tqent);
627
628 return (zio);
629 }
630
631 static void
632 zio_destroy(zio_t *zio)
633 {
634 kmem_cache_free(zio_cache, zio);
635 }
636
637 zio_t *
638 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
639 void *private, enum zio_flag flags)
640 {
641 zio_t *zio;
642
643 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
644 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
645 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
646
647 return (zio);
648 }
649
650 zio_t *
651 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
652 {
653 return (zio_null(NULL, spa, NULL, done, private, flags));
654 }
655
656 zio_t *
657 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
658 void *data, uint64_t size, zio_done_func_t *done, void *private,
659 int priority, enum zio_flag flags, const zbookmark_t *zb)
660 {
661 zio_t *zio;
662
663 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
664 data, size, done, private,
665 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
666 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
667 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
668
669 return (zio);
670 }
671
672 zio_t *
673 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
674 void *data, uint64_t size, const zio_prop_t *zp,
675 zio_done_func_t *ready, zio_done_func_t *done, void *private,
676 int priority, enum zio_flag flags, const zbookmark_t *zb)
677 {
678 zio_t *zio;
679
680 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
681 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
682 zp->zp_compress >= ZIO_COMPRESS_OFF &&
683 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
684 zp->zp_type < DMU_OT_NUMTYPES &&
685 zp->zp_level < 32 &&
686 zp->zp_copies > 0 &&
687 zp->zp_copies <= spa_max_replication(spa) &&
688 zp->zp_dedup <= 1 &&
689 zp->zp_dedup_verify <= 1);
690
691 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
692 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
693 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
694 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
695
696 zio->io_ready = ready;
697 zio->io_prop = *zp;
698
699 return (zio);
700 }
701
702 zio_t *
703 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
704 uint64_t size, zio_done_func_t *done, void *private, int priority,
705 enum zio_flag flags, zbookmark_t *zb)
706 {
707 zio_t *zio;
708
709 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
710 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
711 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
712
713 return (zio);
714 }
715
716 void
717 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
718 {
719 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
720 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
721 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
722 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
723
724 zio->io_prop.zp_copies = copies;
725 zio->io_bp_override = bp;
726 }
727
728 void
729 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
730 {
731 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
732 }
733
734 zio_t *
735 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
736 enum zio_flag flags)
737 {
738 zio_t *zio;
739
740 dprintf_bp(bp, "freeing in txg %llu, pass %u",
741 (longlong_t)txg, spa->spa_sync_pass);
742
743 ASSERT(!BP_IS_HOLE(bp));
744 ASSERT(spa_syncing_txg(spa) == txg);
745 ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
746
747 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
748 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
749 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
750
751 return (zio);
752 }
753
754 zio_t *
755 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
756 zio_done_func_t *done, void *private, enum zio_flag flags)
757 {
758 zio_t *zio;
759
760 /*
761 * A claim is an allocation of a specific block. Claims are needed
762 * to support immediate writes in the intent log. The issue is that
763 * immediate writes contain committed data, but in a txg that was
764 * *not* committed. Upon opening the pool after an unclean shutdown,
765 * the intent log claims all blocks that contain immediate write data
766 * so that the SPA knows they're in use.
767 *
768 * All claims *must* be resolved in the first txg -- before the SPA
769 * starts allocating blocks -- so that nothing is allocated twice.
770 * If txg == 0 we just verify that the block is claimable.
771 */
772 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
773 ASSERT(txg == spa_first_txg(spa) || txg == 0);
774 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
775
776 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
777 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
778 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
779
780 return (zio);
781 }
782
783 zio_t *
784 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
785 zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
786 {
787 zio_t *zio;
788 int c;
789
790 if (vd->vdev_children == 0) {
791 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
792 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
793 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
794
795 zio->io_cmd = cmd;
796 } else {
797 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
798
799 for (c = 0; c < vd->vdev_children; c++)
800 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
801 done, private, priority, flags));
802 }
803
804 return (zio);
805 }
806
807 zio_t *
808 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
809 void *data, int checksum, zio_done_func_t *done, void *private,
810 int priority, enum zio_flag flags, boolean_t labels)
811 {
812 zio_t *zio;
813
814 ASSERT(vd->vdev_children == 0);
815 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
816 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
817 ASSERT3U(offset + size, <=, vd->vdev_psize);
818
819 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
820 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
821 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
822
823 zio->io_prop.zp_checksum = checksum;
824
825 return (zio);
826 }
827
828 zio_t *
829 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
830 void *data, int checksum, zio_done_func_t *done, void *private,
831 int priority, enum zio_flag flags, boolean_t labels)
832 {
833 zio_t *zio;
834
835 ASSERT(vd->vdev_children == 0);
836 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
837 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
838 ASSERT3U(offset + size, <=, vd->vdev_psize);
839
840 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
841 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
842 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
843
844 zio->io_prop.zp_checksum = checksum;
845
846 if (zio_checksum_table[checksum].ci_eck) {
847 /*
848 * zec checksums are necessarily destructive -- they modify
849 * the end of the write buffer to hold the verifier/checksum.
850 * Therefore, we must make a local copy in case the data is
851 * being written to multiple places in parallel.
852 */
853 void *wbuf = zio_buf_alloc(size);
854 bcopy(data, wbuf, size);
855 zio_push_transform(zio, wbuf, size, size, NULL);
856 }
857
858 return (zio);
859 }
860
861 /*
862 * Create a child I/O to do some work for us.
863 */
864 zio_t *
865 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
866 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
867 zio_done_func_t *done, void *private)
868 {
869 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
870 zio_t *zio;
871
872 ASSERT(vd->vdev_parent ==
873 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
874
875 if (type == ZIO_TYPE_READ && bp != NULL) {
876 /*
877 * If we have the bp, then the child should perform the
878 * checksum and the parent need not. This pushes error
879 * detection as close to the leaves as possible and
880 * eliminates redundant checksums in the interior nodes.
881 */
882 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
883 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
884 }
885
886 if (vd->vdev_children == 0)
887 offset += VDEV_LABEL_START_SIZE;
888
889 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
890
891 /*
892 * If we've decided to do a repair, the write is not speculative --
893 * even if the original read was.
894 */
895 if (flags & ZIO_FLAG_IO_REPAIR)
896 flags &= ~ZIO_FLAG_SPECULATIVE;
897
898 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
899 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
900 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
901
902 return (zio);
903 }
904
905 zio_t *
906 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
907 int type, int priority, enum zio_flag flags,
908 zio_done_func_t *done, void *private)
909 {
910 zio_t *zio;
911
912 ASSERT(vd->vdev_ops->vdev_op_leaf);
913
914 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
915 data, size, done, private, type, priority,
916 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
917 vd, offset, NULL,
918 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
919
920 return (zio);
921 }
922
923 void
924 zio_flush(zio_t *zio, vdev_t *vd)
925 {
926 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
927 NULL, NULL, ZIO_PRIORITY_NOW,
928 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
929 }
930
931 void
932 zio_shrink(zio_t *zio, uint64_t size)
933 {
934 ASSERT(zio->io_executor == NULL);
935 ASSERT(zio->io_orig_size == zio->io_size);
936 ASSERT(size <= zio->io_size);
937
938 /*
939 * We don't shrink for raidz because of problems with the
940 * reconstruction when reading back less than the block size.
941 * Note, BP_IS_RAIDZ() assumes no compression.
942 */
943 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
944 if (!BP_IS_RAIDZ(zio->io_bp))
945 zio->io_orig_size = zio->io_size = size;
946 }
947
948 /*
949 * ==========================================================================
950 * Prepare to read and write logical blocks
951 * ==========================================================================
952 */
953
954 static int
955 zio_read_bp_init(zio_t *zio)
956 {
957 blkptr_t *bp = zio->io_bp;
958
959 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
960 zio->io_child_type == ZIO_CHILD_LOGICAL &&
961 !(zio->io_flags & ZIO_FLAG_RAW)) {
962 uint64_t psize = BP_GET_PSIZE(bp);
963 void *cbuf = zio_buf_alloc(psize);
964
965 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
966 }
967
968 if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
969 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
970
971 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
972 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
973
974 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
975 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
976
977 return (ZIO_PIPELINE_CONTINUE);
978 }
979
980 static int
981 zio_write_bp_init(zio_t *zio)
982 {
983 spa_t *spa = zio->io_spa;
984 zio_prop_t *zp = &zio->io_prop;
985 enum zio_compress compress = zp->zp_compress;
986 blkptr_t *bp = zio->io_bp;
987 uint64_t lsize = zio->io_size;
988 uint64_t psize = lsize;
989 int pass = 1;
990
991 /*
992 * If our children haven't all reached the ready stage,
993 * wait for them and then repeat this pipeline stage.
994 */
995 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
996 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
997 return (ZIO_PIPELINE_STOP);
998
999 if (!IO_IS_ALLOCATING(zio))
1000 return (ZIO_PIPELINE_CONTINUE);
1001
1002 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1003
1004 if (zio->io_bp_override) {
1005 ASSERT(bp->blk_birth != zio->io_txg);
1006 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1007
1008 *bp = *zio->io_bp_override;
1009 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1010
1011 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1012 return (ZIO_PIPELINE_CONTINUE);
1013
1014 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1015 zp->zp_dedup_verify);
1016
1017 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1018 BP_SET_DEDUP(bp, 1);
1019 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1020 return (ZIO_PIPELINE_CONTINUE);
1021 }
1022 zio->io_bp_override = NULL;
1023 BP_ZERO(bp);
1024 }
1025
1026 if (bp->blk_birth == zio->io_txg) {
1027 /*
1028 * We're rewriting an existing block, which means we're
1029 * working on behalf of spa_sync(). For spa_sync() to
1030 * converge, it must eventually be the case that we don't
1031 * have to allocate new blocks. But compression changes
1032 * the blocksize, which forces a reallocate, and makes
1033 * convergence take longer. Therefore, after the first
1034 * few passes, stop compressing to ensure convergence.
1035 */
1036 pass = spa_sync_pass(spa);
1037
1038 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1039 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1040 ASSERT(!BP_GET_DEDUP(bp));
1041
1042 if (pass > SYNC_PASS_DONT_COMPRESS)
1043 compress = ZIO_COMPRESS_OFF;
1044
1045 /* Make sure someone doesn't change their mind on overwrites */
1046 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
1047 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1048 }
1049
1050 if (compress != ZIO_COMPRESS_OFF) {
1051 void *cbuf = zio_buf_alloc(lsize);
1052 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1053 if (psize == 0 || psize == lsize) {
1054 compress = ZIO_COMPRESS_OFF;
1055 zio_buf_free(cbuf, lsize);
1056 } else {
1057 ASSERT(psize < lsize);
1058 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1059 }
1060 }
1061
1062 /*
1063 * The final pass of spa_sync() must be all rewrites, but the first
1064 * few passes offer a trade-off: allocating blocks defers convergence,
1065 * but newly allocated blocks are sequential, so they can be written
1066 * to disk faster. Therefore, we allow the first few passes of
1067 * spa_sync() to allocate new blocks, but force rewrites after that.
1068 * There should only be a handful of blocks after pass 1 in any case.
1069 */
1070 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1071 pass > SYNC_PASS_REWRITE) {
1072 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1073 ASSERT(psize != 0);
1074 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1075 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1076 } else {
1077 BP_ZERO(bp);
1078 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1079 }
1080
1081 if (psize == 0) {
1082 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1083 } else {
1084 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1085 BP_SET_LSIZE(bp, lsize);
1086 BP_SET_PSIZE(bp, psize);
1087 BP_SET_COMPRESS(bp, compress);
1088 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1089 BP_SET_TYPE(bp, zp->zp_type);
1090 BP_SET_LEVEL(bp, zp->zp_level);
1091 BP_SET_DEDUP(bp, zp->zp_dedup);
1092 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1093 if (zp->zp_dedup) {
1094 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1095 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1096 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1097 }
1098 }
1099
1100 return (ZIO_PIPELINE_CONTINUE);
1101 }
1102
1103 static int
1104 zio_free_bp_init(zio_t *zio)
1105 {
1106 blkptr_t *bp = zio->io_bp;
1107
1108 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1109 if (BP_GET_DEDUP(bp))
1110 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1111 }
1112
1113 return (ZIO_PIPELINE_CONTINUE);
1114 }
1115
1116 /*
1117 * ==========================================================================
1118 * Execute the I/O pipeline
1119 * ==========================================================================
1120 */
1121
1122 static void
1123 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1124 {
1125 spa_t *spa = zio->io_spa;
1126 zio_type_t t = zio->io_type;
1127 int flags = (cutinline ? TQ_FRONT : 0);
1128
1129 /*
1130 * If we're a config writer or a probe, the normal issue and
1131 * interrupt threads may all be blocked waiting for the config lock.
1132 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1133 */
1134 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1135 t = ZIO_TYPE_NULL;
1136
1137 /*
1138 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1139 */
1140 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1141 t = ZIO_TYPE_NULL;
1142
1143 /*
1144 * If this is a high priority I/O, then use the high priority taskq.
1145 */
1146 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1147 spa->spa_zio_taskq[t][q + 1] != NULL)
1148 q++;
1149
1150 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1151
1152 /*
1153 * NB: We are assuming that the zio can only be dispatched
1154 * to a single taskq at a time. It would be a grievous error
1155 * to dispatch the zio to another taskq at the same time.
1156 */
1157 ASSERT(taskq_empty_ent(&zio->io_tqent));
1158 taskq_dispatch_ent(spa->spa_zio_taskq[t][q],
1159 (task_func_t *)zio_execute, zio, flags, &zio->io_tqent);
1160 }
1161
1162 static boolean_t
1163 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1164 {
1165 kthread_t *executor = zio->io_executor;
1166 spa_t *spa = zio->io_spa;
1167 zio_type_t t;
1168
1169 for (t = 0; t < ZIO_TYPES; t++)
1170 if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1171 return (B_TRUE);
1172
1173 return (B_FALSE);
1174 }
1175
1176 static int
1177 zio_issue_async(zio_t *zio)
1178 {
1179 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1180
1181 return (ZIO_PIPELINE_STOP);
1182 }
1183
1184 void
1185 zio_interrupt(zio_t *zio)
1186 {
1187 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1188 }
1189
1190 /*
1191 * Execute the I/O pipeline until one of the following occurs:
1192 * (1) the I/O completes; (2) the pipeline stalls waiting for
1193 * dependent child I/Os; (3) the I/O issues, so we're waiting
1194 * for an I/O completion interrupt; (4) the I/O is delegated by
1195 * vdev-level caching or aggregation; (5) the I/O is deferred
1196 * due to vdev-level queueing; (6) the I/O is handed off to
1197 * another thread. In all cases, the pipeline stops whenever
1198 * there's no CPU work; it never burns a thread in cv_wait().
1199 *
1200 * There's no locking on io_stage because there's no legitimate way
1201 * for multiple threads to be attempting to process the same I/O.
1202 */
1203 static zio_pipe_stage_t *zio_pipeline[];
1204
1205 /*
1206 * zio_execute() is a wrapper around the static function
1207 * __zio_execute() so that we can force __zio_execute() to be
1208 * inlined. This reduces stack overhead which is important
1209 * because __zio_execute() is called recursively in several zio
1210 * code paths. zio_execute() itself cannot be inlined because
1211 * it is externally visible.
1212 */
1213 void
1214 zio_execute(zio_t *zio)
1215 {
1216 __zio_execute(zio);
1217 }
1218
1219 __attribute__((always_inline))
1220 static inline void
1221 __zio_execute(zio_t *zio)
1222 {
1223 zio->io_executor = curthread;
1224
1225 while (zio->io_stage < ZIO_STAGE_DONE) {
1226 enum zio_stage pipeline = zio->io_pipeline;
1227 enum zio_stage stage = zio->io_stage;
1228 dsl_pool_t *dsl;
1229 boolean_t cut;
1230 int rv;
1231
1232 ASSERT(!MUTEX_HELD(&zio->io_lock));
1233 ASSERT(ISP2(stage));
1234 ASSERT(zio->io_stall == NULL);
1235
1236 do {
1237 stage <<= 1;
1238 } while ((stage & pipeline) == 0);
1239
1240 ASSERT(stage <= ZIO_STAGE_DONE);
1241
1242 dsl = spa_get_dsl(zio->io_spa);
1243 cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1244 zio_requeue_io_start_cut_in_line : B_FALSE;
1245
1246 /*
1247 * If we are in interrupt context and this pipeline stage
1248 * will grab a config lock that is held across I/O,
1249 * or may wait for an I/O that needs an interrupt thread
1250 * to complete, issue async to avoid deadlock.
1251 *
1252 * If we are in the txg_sync_thread or being called
1253 * during pool init issue async to minimize stack depth.
1254 * Both of these call paths may be recursively called.
1255 *
1256 * For VDEV_IO_START, we cut in line so that the io will
1257 * be sent to disk promptly.
1258 */
1259 if (((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1260 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) ||
1261 (dsl != NULL && dsl_pool_sync_context(dsl))) {
1262 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1263 return;
1264 }
1265
1266 zio->io_stage = stage;
1267 rv = zio_pipeline[highbit(stage) - 1](zio);
1268
1269 if (rv == ZIO_PIPELINE_STOP)
1270 return;
1271
1272 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1273 }
1274 }
1275
1276
1277 /*
1278 * ==========================================================================
1279 * Initiate I/O, either sync or async
1280 * ==========================================================================
1281 */
1282 int
1283 zio_wait(zio_t *zio)
1284 {
1285 int error;
1286
1287 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1288 ASSERT(zio->io_executor == NULL);
1289
1290 zio->io_waiter = curthread;
1291
1292 __zio_execute(zio);
1293
1294 mutex_enter(&zio->io_lock);
1295 while (zio->io_executor != NULL)
1296 cv_wait(&zio->io_cv, &zio->io_lock);
1297 mutex_exit(&zio->io_lock);
1298
1299 error = zio->io_error;
1300 zio_destroy(zio);
1301
1302 return (error);
1303 }
1304
1305 void
1306 zio_nowait(zio_t *zio)
1307 {
1308 ASSERT(zio->io_executor == NULL);
1309
1310 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1311 zio_unique_parent(zio) == NULL) {
1312 /*
1313 * This is a logical async I/O with no parent to wait for it.
1314 * We add it to the spa_async_root_zio "Godfather" I/O which
1315 * will ensure they complete prior to unloading the pool.
1316 */
1317 spa_t *spa = zio->io_spa;
1318
1319 zio_add_child(spa->spa_async_zio_root, zio);
1320 }
1321
1322 __zio_execute(zio);
1323 }
1324
1325 /*
1326 * ==========================================================================
1327 * Reexecute or suspend/resume failed I/O
1328 * ==========================================================================
1329 */
1330
1331 static void
1332 zio_reexecute(zio_t *pio)
1333 {
1334 zio_t *cio, *cio_next;
1335 int c, w;
1336
1337 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1338 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1339 ASSERT(pio->io_gang_leader == NULL);
1340 ASSERT(pio->io_gang_tree == NULL);
1341
1342 pio->io_flags = pio->io_orig_flags;
1343 pio->io_stage = pio->io_orig_stage;
1344 pio->io_pipeline = pio->io_orig_pipeline;
1345 pio->io_reexecute = 0;
1346 pio->io_error = 0;
1347 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1348 pio->io_state[w] = 0;
1349 for (c = 0; c < ZIO_CHILD_TYPES; c++)
1350 pio->io_child_error[c] = 0;
1351
1352 if (IO_IS_ALLOCATING(pio))
1353 BP_ZERO(pio->io_bp);
1354
1355 /*
1356 * As we reexecute pio's children, new children could be created.
1357 * New children go to the head of pio's io_child_list, however,
1358 * so we will (correctly) not reexecute them. The key is that
1359 * the remainder of pio's io_child_list, from 'cio_next' onward,
1360 * cannot be affected by any side effects of reexecuting 'cio'.
1361 */
1362 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1363 cio_next = zio_walk_children(pio);
1364 mutex_enter(&pio->io_lock);
1365 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1366 pio->io_children[cio->io_child_type][w]++;
1367 mutex_exit(&pio->io_lock);
1368 zio_reexecute(cio);
1369 }
1370
1371 /*
1372 * Now that all children have been reexecuted, execute the parent.
1373 * We don't reexecute "The Godfather" I/O here as it's the
1374 * responsibility of the caller to wait on him.
1375 */
1376 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1377 __zio_execute(pio);
1378 }
1379
1380 void
1381 zio_suspend(spa_t *spa, zio_t *zio)
1382 {
1383 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1384 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1385 "failure and the failure mode property for this pool "
1386 "is set to panic.", spa_name(spa));
1387
1388 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1389
1390 mutex_enter(&spa->spa_suspend_lock);
1391
1392 if (spa->spa_suspend_zio_root == NULL)
1393 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1394 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1395 ZIO_FLAG_GODFATHER);
1396
1397 spa->spa_suspended = B_TRUE;
1398
1399 if (zio != NULL) {
1400 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1401 ASSERT(zio != spa->spa_suspend_zio_root);
1402 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1403 ASSERT(zio_unique_parent(zio) == NULL);
1404 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1405 zio_add_child(spa->spa_suspend_zio_root, zio);
1406 }
1407
1408 mutex_exit(&spa->spa_suspend_lock);
1409 }
1410
1411 int
1412 zio_resume(spa_t *spa)
1413 {
1414 zio_t *pio;
1415
1416 /*
1417 * Reexecute all previously suspended i/o.
1418 */
1419 mutex_enter(&spa->spa_suspend_lock);
1420 spa->spa_suspended = B_FALSE;
1421 cv_broadcast(&spa->spa_suspend_cv);
1422 pio = spa->spa_suspend_zio_root;
1423 spa->spa_suspend_zio_root = NULL;
1424 mutex_exit(&spa->spa_suspend_lock);
1425
1426 if (pio == NULL)
1427 return (0);
1428
1429 zio_reexecute(pio);
1430 return (zio_wait(pio));
1431 }
1432
1433 void
1434 zio_resume_wait(spa_t *spa)
1435 {
1436 mutex_enter(&spa->spa_suspend_lock);
1437 while (spa_suspended(spa))
1438 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1439 mutex_exit(&spa->spa_suspend_lock);
1440 }
1441
1442 /*
1443 * ==========================================================================
1444 * Gang blocks.
1445 *
1446 * A gang block is a collection of small blocks that looks to the DMU
1447 * like one large block. When zio_dva_allocate() cannot find a block
1448 * of the requested size, due to either severe fragmentation or the pool
1449 * being nearly full, it calls zio_write_gang_block() to construct the
1450 * block from smaller fragments.
1451 *
1452 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1453 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1454 * an indirect block: it's an array of block pointers. It consumes
1455 * only one sector and hence is allocatable regardless of fragmentation.
1456 * The gang header's bps point to its gang members, which hold the data.
1457 *
1458 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1459 * as the verifier to ensure uniqueness of the SHA256 checksum.
1460 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1461 * not the gang header. This ensures that data block signatures (needed for
1462 * deduplication) are independent of how the block is physically stored.
1463 *
1464 * Gang blocks can be nested: a gang member may itself be a gang block.
1465 * Thus every gang block is a tree in which root and all interior nodes are
1466 * gang headers, and the leaves are normal blocks that contain user data.
1467 * The root of the gang tree is called the gang leader.
1468 *
1469 * To perform any operation (read, rewrite, free, claim) on a gang block,
1470 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1471 * in the io_gang_tree field of the original logical i/o by recursively
1472 * reading the gang leader and all gang headers below it. This yields
1473 * an in-core tree containing the contents of every gang header and the
1474 * bps for every constituent of the gang block.
1475 *
1476 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1477 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1478 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1479 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1480 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1481 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1482 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1483 * of the gang header plus zio_checksum_compute() of the data to update the
1484 * gang header's blk_cksum as described above.
1485 *
1486 * The two-phase assemble/issue model solves the problem of partial failure --
1487 * what if you'd freed part of a gang block but then couldn't read the
1488 * gang header for another part? Assembling the entire gang tree first
1489 * ensures that all the necessary gang header I/O has succeeded before
1490 * starting the actual work of free, claim, or write. Once the gang tree
1491 * is assembled, free and claim are in-memory operations that cannot fail.
1492 *
1493 * In the event that a gang write fails, zio_dva_unallocate() walks the
1494 * gang tree to immediately free (i.e. insert back into the space map)
1495 * everything we've allocated. This ensures that we don't get ENOSPC
1496 * errors during repeated suspend/resume cycles due to a flaky device.
1497 *
1498 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1499 * the gang tree, we won't modify the block, so we can safely defer the free
1500 * (knowing that the block is still intact). If we *can* assemble the gang
1501 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1502 * each constituent bp and we can allocate a new block on the next sync pass.
1503 *
1504 * In all cases, the gang tree allows complete recovery from partial failure.
1505 * ==========================================================================
1506 */
1507
1508 static zio_t *
1509 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1510 {
1511 if (gn != NULL)
1512 return (pio);
1513
1514 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1515 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1516 &pio->io_bookmark));
1517 }
1518
1519 zio_t *
1520 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1521 {
1522 zio_t *zio;
1523
1524 if (gn != NULL) {
1525 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1526 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1527 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1528 /*
1529 * As we rewrite each gang header, the pipeline will compute
1530 * a new gang block header checksum for it; but no one will
1531 * compute a new data checksum, so we do that here. The one
1532 * exception is the gang leader: the pipeline already computed
1533 * its data checksum because that stage precedes gang assembly.
1534 * (Presently, nothing actually uses interior data checksums;
1535 * this is just good hygiene.)
1536 */
1537 if (gn != pio->io_gang_leader->io_gang_tree) {
1538 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1539 data, BP_GET_PSIZE(bp));
1540 }
1541 /*
1542 * If we are here to damage data for testing purposes,
1543 * leave the GBH alone so that we can detect the damage.
1544 */
1545 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1546 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1547 } else {
1548 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1549 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1550 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1551 }
1552
1553 return (zio);
1554 }
1555
1556 /* ARGSUSED */
1557 zio_t *
1558 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1559 {
1560 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1561 ZIO_GANG_CHILD_FLAGS(pio)));
1562 }
1563
1564 /* ARGSUSED */
1565 zio_t *
1566 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1567 {
1568 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1569 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1570 }
1571
1572 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1573 NULL,
1574 zio_read_gang,
1575 zio_rewrite_gang,
1576 zio_free_gang,
1577 zio_claim_gang,
1578 NULL
1579 };
1580
1581 static void zio_gang_tree_assemble_done(zio_t *zio);
1582
1583 static zio_gang_node_t *
1584 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1585 {
1586 zio_gang_node_t *gn;
1587
1588 ASSERT(*gnpp == NULL);
1589
1590 gn = kmem_zalloc(sizeof (*gn), KM_PUSHPAGE);
1591 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1592 *gnpp = gn;
1593
1594 return (gn);
1595 }
1596
1597 static void
1598 zio_gang_node_free(zio_gang_node_t **gnpp)
1599 {
1600 zio_gang_node_t *gn = *gnpp;
1601 int g;
1602
1603 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1604 ASSERT(gn->gn_child[g] == NULL);
1605
1606 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1607 kmem_free(gn, sizeof (*gn));
1608 *gnpp = NULL;
1609 }
1610
1611 static void
1612 zio_gang_tree_free(zio_gang_node_t **gnpp)
1613 {
1614 zio_gang_node_t *gn = *gnpp;
1615 int g;
1616
1617 if (gn == NULL)
1618 return;
1619
1620 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1621 zio_gang_tree_free(&gn->gn_child[g]);
1622
1623 zio_gang_node_free(gnpp);
1624 }
1625
1626 static void
1627 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1628 {
1629 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1630
1631 ASSERT(gio->io_gang_leader == gio);
1632 ASSERT(BP_IS_GANG(bp));
1633
1634 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1635 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1636 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1637 }
1638
1639 static void
1640 zio_gang_tree_assemble_done(zio_t *zio)
1641 {
1642 zio_t *gio = zio->io_gang_leader;
1643 zio_gang_node_t *gn = zio->io_private;
1644 blkptr_t *bp = zio->io_bp;
1645 int g;
1646
1647 ASSERT(gio == zio_unique_parent(zio));
1648 ASSERT(zio->io_child_count == 0);
1649
1650 if (zio->io_error)
1651 return;
1652
1653 if (BP_SHOULD_BYTESWAP(bp))
1654 byteswap_uint64_array(zio->io_data, zio->io_size);
1655
1656 ASSERT(zio->io_data == gn->gn_gbh);
1657 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1658 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1659
1660 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1661 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1662 if (!BP_IS_GANG(gbp))
1663 continue;
1664 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1665 }
1666 }
1667
1668 static void
1669 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1670 {
1671 zio_t *gio = pio->io_gang_leader;
1672 zio_t *zio;
1673 int g;
1674
1675 ASSERT(BP_IS_GANG(bp) == !!gn);
1676 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1677 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1678
1679 /*
1680 * If you're a gang header, your data is in gn->gn_gbh.
1681 * If you're a gang member, your data is in 'data' and gn == NULL.
1682 */
1683 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1684
1685 if (gn != NULL) {
1686 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1687
1688 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1689 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1690 if (BP_IS_HOLE(gbp))
1691 continue;
1692 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1693 data = (char *)data + BP_GET_PSIZE(gbp);
1694 }
1695 }
1696
1697 if (gn == gio->io_gang_tree)
1698 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1699
1700 if (zio != pio)
1701 zio_nowait(zio);
1702 }
1703
1704 static int
1705 zio_gang_assemble(zio_t *zio)
1706 {
1707 blkptr_t *bp = zio->io_bp;
1708
1709 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1710 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1711
1712 zio->io_gang_leader = zio;
1713
1714 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1715
1716 return (ZIO_PIPELINE_CONTINUE);
1717 }
1718
1719 static int
1720 zio_gang_issue(zio_t *zio)
1721 {
1722 blkptr_t *bp = zio->io_bp;
1723
1724 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1725 return (ZIO_PIPELINE_STOP);
1726
1727 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1728 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1729
1730 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1731 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1732 else
1733 zio_gang_tree_free(&zio->io_gang_tree);
1734
1735 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1736
1737 return (ZIO_PIPELINE_CONTINUE);
1738 }
1739
1740 static void
1741 zio_write_gang_member_ready(zio_t *zio)
1742 {
1743 zio_t *pio = zio_unique_parent(zio);
1744 ASSERTV(zio_t *gio = zio->io_gang_leader;)
1745 dva_t *cdva = zio->io_bp->blk_dva;
1746 dva_t *pdva = pio->io_bp->blk_dva;
1747 uint64_t asize;
1748 int d;
1749
1750 if (BP_IS_HOLE(zio->io_bp))
1751 return;
1752
1753 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1754
1755 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1756 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1757 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1758 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1759 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1760
1761 mutex_enter(&pio->io_lock);
1762 for (d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1763 ASSERT(DVA_GET_GANG(&pdva[d]));
1764 asize = DVA_GET_ASIZE(&pdva[d]);
1765 asize += DVA_GET_ASIZE(&cdva[d]);
1766 DVA_SET_ASIZE(&pdva[d], asize);
1767 }
1768 mutex_exit(&pio->io_lock);
1769 }
1770
1771 static int
1772 zio_write_gang_block(zio_t *pio)
1773 {
1774 spa_t *spa = pio->io_spa;
1775 blkptr_t *bp = pio->io_bp;
1776 zio_t *gio = pio->io_gang_leader;
1777 zio_t *zio;
1778 zio_gang_node_t *gn, **gnpp;
1779 zio_gbh_phys_t *gbh;
1780 uint64_t txg = pio->io_txg;
1781 uint64_t resid = pio->io_size;
1782 uint64_t lsize;
1783 int copies = gio->io_prop.zp_copies;
1784 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1785 zio_prop_t zp;
1786 int g, error;
1787
1788 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1789 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1790 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1791 if (error) {
1792 pio->io_error = error;
1793 return (ZIO_PIPELINE_CONTINUE);
1794 }
1795
1796 if (pio == gio) {
1797 gnpp = &gio->io_gang_tree;
1798 } else {
1799 gnpp = pio->io_private;
1800 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1801 }
1802
1803 gn = zio_gang_node_alloc(gnpp);
1804 gbh = gn->gn_gbh;
1805 bzero(gbh, SPA_GANGBLOCKSIZE);
1806
1807 /*
1808 * Create the gang header.
1809 */
1810 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1811 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1812
1813 /*
1814 * Create and nowait the gang children.
1815 */
1816 for (g = 0; resid != 0; resid -= lsize, g++) {
1817 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1818 SPA_MINBLOCKSIZE);
1819 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1820
1821 zp.zp_checksum = gio->io_prop.zp_checksum;
1822 zp.zp_compress = ZIO_COMPRESS_OFF;
1823 zp.zp_type = DMU_OT_NONE;
1824 zp.zp_level = 0;
1825 zp.zp_copies = gio->io_prop.zp_copies;
1826 zp.zp_dedup = 0;
1827 zp.zp_dedup_verify = 0;
1828
1829 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1830 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1831 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1832 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1833 &pio->io_bookmark));
1834 }
1835
1836 /*
1837 * Set pio's pipeline to just wait for zio to finish.
1838 */
1839 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1840
1841 zio_nowait(zio);
1842
1843 return (ZIO_PIPELINE_CONTINUE);
1844 }
1845
1846 /*
1847 * ==========================================================================
1848 * Dedup
1849 * ==========================================================================
1850 */
1851 static void
1852 zio_ddt_child_read_done(zio_t *zio)
1853 {
1854 blkptr_t *bp = zio->io_bp;
1855 ddt_entry_t *dde = zio->io_private;
1856 ddt_phys_t *ddp;
1857 zio_t *pio = zio_unique_parent(zio);
1858
1859 mutex_enter(&pio->io_lock);
1860 ddp = ddt_phys_select(dde, bp);
1861 if (zio->io_error == 0)
1862 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1863 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1864 dde->dde_repair_data = zio->io_data;
1865 else
1866 zio_buf_free(zio->io_data, zio->io_size);
1867 mutex_exit(&pio->io_lock);
1868 }
1869
1870 static int
1871 zio_ddt_read_start(zio_t *zio)
1872 {
1873 blkptr_t *bp = zio->io_bp;
1874 int p;
1875
1876 ASSERT(BP_GET_DEDUP(bp));
1877 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1878 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1879
1880 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1881 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1882 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1883 ddt_phys_t *ddp = dde->dde_phys;
1884 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1885 blkptr_t blk;
1886
1887 ASSERT(zio->io_vsd == NULL);
1888 zio->io_vsd = dde;
1889
1890 if (ddp_self == NULL)
1891 return (ZIO_PIPELINE_CONTINUE);
1892
1893 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1894 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1895 continue;
1896 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1897 &blk);
1898 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1899 zio_buf_alloc(zio->io_size), zio->io_size,
1900 zio_ddt_child_read_done, dde, zio->io_priority,
1901 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1902 &zio->io_bookmark));
1903 }
1904 return (ZIO_PIPELINE_CONTINUE);
1905 }
1906
1907 zio_nowait(zio_read(zio, zio->io_spa, bp,
1908 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1909 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1910
1911 return (ZIO_PIPELINE_CONTINUE);
1912 }
1913
1914 static int
1915 zio_ddt_read_done(zio_t *zio)
1916 {
1917 blkptr_t *bp = zio->io_bp;
1918
1919 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1920 return (ZIO_PIPELINE_STOP);
1921
1922 ASSERT(BP_GET_DEDUP(bp));
1923 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1924 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1925
1926 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1927 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1928 ddt_entry_t *dde = zio->io_vsd;
1929 if (ddt == NULL) {
1930 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1931 return (ZIO_PIPELINE_CONTINUE);
1932 }
1933 if (dde == NULL) {
1934 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1935 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1936 return (ZIO_PIPELINE_STOP);
1937 }
1938 if (dde->dde_repair_data != NULL) {
1939 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1940 zio->io_child_error[ZIO_CHILD_DDT] = 0;
1941 }
1942 ddt_repair_done(ddt, dde);
1943 zio->io_vsd = NULL;
1944 }
1945
1946 ASSERT(zio->io_vsd == NULL);
1947
1948 return (ZIO_PIPELINE_CONTINUE);
1949 }
1950
1951 static boolean_t
1952 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1953 {
1954 spa_t *spa = zio->io_spa;
1955 int p;
1956
1957 /*
1958 * Note: we compare the original data, not the transformed data,
1959 * because when zio->io_bp is an override bp, we will not have
1960 * pushed the I/O transforms. That's an important optimization
1961 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1962 */
1963 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1964 zio_t *lio = dde->dde_lead_zio[p];
1965
1966 if (lio != NULL) {
1967 return (lio->io_orig_size != zio->io_orig_size ||
1968 bcmp(zio->io_orig_data, lio->io_orig_data,
1969 zio->io_orig_size) != 0);
1970 }
1971 }
1972
1973 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1974 ddt_phys_t *ddp = &dde->dde_phys[p];
1975
1976 if (ddp->ddp_phys_birth != 0) {
1977 arc_buf_t *abuf = NULL;
1978 uint32_t aflags = ARC_WAIT;
1979 blkptr_t blk = *zio->io_bp;
1980 int error;
1981
1982 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
1983
1984 ddt_exit(ddt);
1985
1986 error = arc_read_nolock(NULL, spa, &blk,
1987 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
1988 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1989 &aflags, &zio->io_bookmark);
1990
1991 if (error == 0) {
1992 if (arc_buf_size(abuf) != zio->io_orig_size ||
1993 bcmp(abuf->b_data, zio->io_orig_data,
1994 zio->io_orig_size) != 0)
1995 error = EEXIST;
1996 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
1997 }
1998
1999 ddt_enter(ddt);
2000 return (error != 0);
2001 }
2002 }
2003
2004 return (B_FALSE);
2005 }
2006
2007 static void
2008 zio_ddt_child_write_ready(zio_t *zio)
2009 {
2010 int p = zio->io_prop.zp_copies;
2011 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2012 ddt_entry_t *dde = zio->io_private;
2013 ddt_phys_t *ddp = &dde->dde_phys[p];
2014 zio_t *pio;
2015
2016 if (zio->io_error)
2017 return;
2018
2019 ddt_enter(ddt);
2020
2021 ASSERT(dde->dde_lead_zio[p] == zio);
2022
2023 ddt_phys_fill(ddp, zio->io_bp);
2024
2025 while ((pio = zio_walk_parents(zio)) != NULL)
2026 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2027
2028 ddt_exit(ddt);
2029 }
2030
2031 static void
2032 zio_ddt_child_write_done(zio_t *zio)
2033 {
2034 int p = zio->io_prop.zp_copies;
2035 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2036 ddt_entry_t *dde = zio->io_private;
2037 ddt_phys_t *ddp = &dde->dde_phys[p];
2038
2039 ddt_enter(ddt);
2040
2041 ASSERT(ddp->ddp_refcnt == 0);
2042 ASSERT(dde->dde_lead_zio[p] == zio);
2043 dde->dde_lead_zio[p] = NULL;
2044
2045 if (zio->io_error == 0) {
2046 while (zio_walk_parents(zio) != NULL)
2047 ddt_phys_addref(ddp);
2048 } else {
2049 ddt_phys_clear(ddp);
2050 }
2051
2052 ddt_exit(ddt);
2053 }
2054
2055 static void
2056 zio_ddt_ditto_write_done(zio_t *zio)
2057 {
2058 int p = DDT_PHYS_DITTO;
2059 blkptr_t *bp = zio->io_bp;
2060 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2061 ddt_entry_t *dde = zio->io_private;
2062 ddt_phys_t *ddp = &dde->dde_phys[p];
2063 ddt_key_t *ddk = &dde->dde_key;
2064 ASSERTV(zio_prop_t *zp = &zio->io_prop);
2065
2066 ddt_enter(ddt);
2067
2068 ASSERT(ddp->ddp_refcnt == 0);
2069 ASSERT(dde->dde_lead_zio[p] == zio);
2070 dde->dde_lead_zio[p] = NULL;
2071
2072 if (zio->io_error == 0) {
2073 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2074 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2075 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2076 if (ddp->ddp_phys_birth != 0)
2077 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2078 ddt_phys_fill(ddp, bp);
2079 }
2080
2081 ddt_exit(ddt);
2082 }
2083
2084 static int
2085 zio_ddt_write(zio_t *zio)
2086 {
2087 spa_t *spa = zio->io_spa;
2088 blkptr_t *bp = zio->io_bp;
2089 uint64_t txg = zio->io_txg;
2090 zio_prop_t *zp = &zio->io_prop;
2091 int p = zp->zp_copies;
2092 int ditto_copies;
2093 zio_t *cio = NULL;
2094 zio_t *dio = NULL;
2095 ddt_t *ddt = ddt_select(spa, bp);
2096 ddt_entry_t *dde;
2097 ddt_phys_t *ddp;
2098
2099 ASSERT(BP_GET_DEDUP(bp));
2100 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2101 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2102
2103 ddt_enter(ddt);
2104 dde = ddt_lookup(ddt, bp, B_TRUE);
2105 ddp = &dde->dde_phys[p];
2106
2107 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2108 /*
2109 * If we're using a weak checksum, upgrade to a strong checksum
2110 * and try again. If we're already using a strong checksum,
2111 * we can't resolve it, so just convert to an ordinary write.
2112 * (And automatically e-mail a paper to Nature?)
2113 */
2114 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2115 zp->zp_checksum = spa_dedup_checksum(spa);
2116 zio_pop_transforms(zio);
2117 zio->io_stage = ZIO_STAGE_OPEN;
2118 BP_ZERO(bp);
2119 } else {
2120 zp->zp_dedup = 0;
2121 }
2122 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2123 ddt_exit(ddt);
2124 return (ZIO_PIPELINE_CONTINUE);
2125 }
2126
2127 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2128 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2129
2130 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2131 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2132 zio_prop_t czp = *zp;
2133
2134 czp.zp_copies = ditto_copies;
2135
2136 /*
2137 * If we arrived here with an override bp, we won't have run
2138 * the transform stack, so we won't have the data we need to
2139 * generate a child i/o. So, toss the override bp and restart.
2140 * This is safe, because using the override bp is just an
2141 * optimization; and it's rare, so the cost doesn't matter.
2142 */
2143 if (zio->io_bp_override) {
2144 zio_pop_transforms(zio);
2145 zio->io_stage = ZIO_STAGE_OPEN;
2146 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2147 zio->io_bp_override = NULL;
2148 BP_ZERO(bp);
2149 ddt_exit(ddt);
2150 return (ZIO_PIPELINE_CONTINUE);
2151 }
2152
2153 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2154 zio->io_orig_size, &czp, NULL,
2155 zio_ddt_ditto_write_done, dde, zio->io_priority,
2156 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2157
2158 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2159 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2160 }
2161
2162 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2163 if (ddp->ddp_phys_birth != 0)
2164 ddt_bp_fill(ddp, bp, txg);
2165 if (dde->dde_lead_zio[p] != NULL)
2166 zio_add_child(zio, dde->dde_lead_zio[p]);
2167 else
2168 ddt_phys_addref(ddp);
2169 } else if (zio->io_bp_override) {
2170 ASSERT(bp->blk_birth == txg);
2171 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2172 ddt_phys_fill(ddp, bp);
2173 ddt_phys_addref(ddp);
2174 } else {
2175 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2176 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2177 zio_ddt_child_write_done, dde, zio->io_priority,
2178 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2179
2180 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2181 dde->dde_lead_zio[p] = cio;
2182 }
2183
2184 ddt_exit(ddt);
2185
2186 if (cio)
2187 zio_nowait(cio);
2188 if (dio)
2189 zio_nowait(dio);
2190
2191 return (ZIO_PIPELINE_CONTINUE);
2192 }
2193
2194 ddt_entry_t *freedde; /* for debugging */
2195
2196 static int
2197 zio_ddt_free(zio_t *zio)
2198 {
2199 spa_t *spa = zio->io_spa;
2200 blkptr_t *bp = zio->io_bp;
2201 ddt_t *ddt = ddt_select(spa, bp);
2202 ddt_entry_t *dde;
2203 ddt_phys_t *ddp;
2204
2205 ASSERT(BP_GET_DEDUP(bp));
2206 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2207
2208 ddt_enter(ddt);
2209 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2210 ddp = ddt_phys_select(dde, bp);
2211 ddt_phys_decref(ddp);
2212 ddt_exit(ddt);
2213
2214 return (ZIO_PIPELINE_CONTINUE);
2215 }
2216
2217 /*
2218 * ==========================================================================
2219 * Allocate and free blocks
2220 * ==========================================================================
2221 */
2222 static int
2223 zio_dva_allocate(zio_t *zio)
2224 {
2225 spa_t *spa = zio->io_spa;
2226 metaslab_class_t *mc = spa_normal_class(spa);
2227 blkptr_t *bp = zio->io_bp;
2228 int error;
2229 int flags = 0;
2230
2231 if (zio->io_gang_leader == NULL) {
2232 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2233 zio->io_gang_leader = zio;
2234 }
2235
2236 ASSERT(BP_IS_HOLE(bp));
2237 ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
2238 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2239 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2240 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2241
2242 /*
2243 * The dump device does not support gang blocks so allocation on
2244 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2245 * the "fast" gang feature.
2246 */
2247 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2248 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2249 METASLAB_GANG_CHILD : 0;
2250 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2251 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2252
2253 if (error) {
2254 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2255 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2256 error);
2257 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2258 return (zio_write_gang_block(zio));
2259 zio->io_error = error;
2260 }
2261
2262 return (ZIO_PIPELINE_CONTINUE);
2263 }
2264
2265 static int
2266 zio_dva_free(zio_t *zio)
2267 {
2268 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2269
2270 return (ZIO_PIPELINE_CONTINUE);
2271 }
2272
2273 static int
2274 zio_dva_claim(zio_t *zio)
2275 {
2276 int error;
2277
2278 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2279 if (error)
2280 zio->io_error = error;
2281
2282 return (ZIO_PIPELINE_CONTINUE);
2283 }
2284
2285 /*
2286 * Undo an allocation. This is used by zio_done() when an I/O fails
2287 * and we want to give back the block we just allocated.
2288 * This handles both normal blocks and gang blocks.
2289 */
2290 static void
2291 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2292 {
2293 int g;
2294
2295 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2296 ASSERT(zio->io_bp_override == NULL);
2297
2298 if (!BP_IS_HOLE(bp))
2299 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2300
2301 if (gn != NULL) {
2302 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2303 zio_dva_unallocate(zio, gn->gn_child[g],
2304 &gn->gn_gbh->zg_blkptr[g]);
2305 }
2306 }
2307 }
2308
2309 /*
2310 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2311 */
2312 int
2313 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2314 uint64_t size, boolean_t use_slog)
2315 {
2316 int error = 1;
2317
2318 ASSERT(txg > spa_syncing_txg(spa));
2319
2320 /*
2321 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2322 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2323 * when allocating them.
2324 */
2325 if (use_slog) {
2326 error = metaslab_alloc(spa, spa_log_class(spa), size,
2327 new_bp, 1, txg, old_bp,
2328 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2329 }
2330
2331 if (error) {
2332 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2333 new_bp, 1, txg, old_bp,
2334 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2335 }
2336
2337 if (error == 0) {
2338 BP_SET_LSIZE(new_bp, size);
2339 BP_SET_PSIZE(new_bp, size);
2340 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2341 BP_SET_CHECKSUM(new_bp,
2342 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2343 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2344 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2345 BP_SET_LEVEL(new_bp, 0);
2346 BP_SET_DEDUP(new_bp, 0);
2347 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2348 }
2349
2350 return (error);
2351 }
2352
2353 /*
2354 * Free an intent log block.
2355 */
2356 void
2357 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2358 {
2359 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2360 ASSERT(!BP_IS_GANG(bp));
2361
2362 zio_free(spa, txg, bp);
2363 }
2364
2365 /*
2366 * ==========================================================================
2367 * Read and write to physical devices
2368 * ==========================================================================
2369 */
2370 static int
2371 zio_vdev_io_start(zio_t *zio)
2372 {
2373 vdev_t *vd = zio->io_vd;
2374 uint64_t align;
2375 spa_t *spa = zio->io_spa;
2376
2377 ASSERT(zio->io_error == 0);
2378 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2379
2380 if (vd == NULL) {
2381 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2382 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2383
2384 /*
2385 * The mirror_ops handle multiple DVAs in a single BP.
2386 */
2387 return (vdev_mirror_ops.vdev_op_io_start(zio));
2388 }
2389
2390 /*
2391 * We keep track of time-sensitive I/Os so that the scan thread
2392 * can quickly react to certain workloads. In particular, we care
2393 * about non-scrubbing, top-level reads and writes with the following
2394 * characteristics:
2395 * - synchronous writes of user data to non-slog devices
2396 * - any reads of user data
2397 * When these conditions are met, adjust the timestamp of spa_last_io
2398 * which allows the scan thread to adjust its workload accordingly.
2399 */
2400 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2401 vd == vd->vdev_top && !vd->vdev_islog &&
2402 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2403 zio->io_txg != spa_syncing_txg(spa)) {
2404 uint64_t old = spa->spa_last_io;
2405 uint64_t new = ddi_get_lbolt64();
2406 if (old != new)
2407 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2408 }
2409
2410 align = 1ULL << vd->vdev_top->vdev_ashift;
2411
2412 if (P2PHASE(zio->io_size, align) != 0) {
2413 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2414 char *abuf = zio_buf_alloc(asize);
2415 ASSERT(vd == vd->vdev_top);
2416 if (zio->io_type == ZIO_TYPE_WRITE) {
2417 bcopy(zio->io_data, abuf, zio->io_size);
2418 bzero(abuf + zio->io_size, asize - zio->io_size);
2419 }
2420 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2421 }
2422
2423 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2424 ASSERT(P2PHASE(zio->io_size, align) == 0);
2425 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2426
2427 /*
2428 * If this is a repair I/O, and there's no self-healing involved --
2429 * that is, we're just resilvering what we expect to resilver --
2430 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2431 * This prevents spurious resilvering with nested replication.
2432 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2433 * A is out of date, we'll read from C+D, then use the data to
2434 * resilver A+B -- but we don't actually want to resilver B, just A.
2435 * The top-level mirror has no way to know this, so instead we just
2436 * discard unnecessary repairs as we work our way down the vdev tree.
2437 * The same logic applies to any form of nested replication:
2438 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2439 */
2440 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2441 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2442 zio->io_txg != 0 && /* not a delegated i/o */
2443 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2444 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2445 zio_vdev_io_bypass(zio);
2446 return (ZIO_PIPELINE_CONTINUE);
2447 }
2448
2449 if (vd->vdev_ops->vdev_op_leaf &&
2450 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2451
2452 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2453 return (ZIO_PIPELINE_CONTINUE);
2454
2455 if ((zio = vdev_queue_io(zio)) == NULL)
2456 return (ZIO_PIPELINE_STOP);
2457
2458 if (!vdev_accessible(vd, zio)) {
2459 zio->io_error = ENXIO;
2460 zio_interrupt(zio);
2461 return (ZIO_PIPELINE_STOP);
2462 }
2463 }
2464
2465 return (vd->vdev_ops->vdev_op_io_start(zio));
2466 }
2467
2468 static int
2469 zio_vdev_io_done(zio_t *zio)
2470 {
2471 vdev_t *vd = zio->io_vd;
2472 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2473 boolean_t unexpected_error = B_FALSE;
2474
2475 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2476 return (ZIO_PIPELINE_STOP);
2477
2478 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2479
2480 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2481
2482 vdev_queue_io_done(zio);
2483
2484 if (zio->io_type == ZIO_TYPE_WRITE)
2485 vdev_cache_write(zio);
2486
2487 if (zio_injection_enabled && zio->io_error == 0)
2488 zio->io_error = zio_handle_device_injection(vd,
2489 zio, EIO);
2490
2491 if (zio_injection_enabled && zio->io_error == 0)
2492 zio->io_error = zio_handle_label_injection(zio, EIO);
2493
2494 if (zio->io_error) {
2495 if (!vdev_accessible(vd, zio)) {
2496 zio->io_error = ENXIO;
2497 } else {
2498 unexpected_error = B_TRUE;
2499 }
2500 }
2501 }
2502
2503 ops->vdev_op_io_done(zio);
2504
2505 if (unexpected_error)
2506 VERIFY(vdev_probe(vd, zio) == NULL);
2507
2508 return (ZIO_PIPELINE_CONTINUE);
2509 }
2510
2511 /*
2512 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2513 * disk, and use that to finish the checksum ereport later.
2514 */
2515 static void
2516 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2517 const void *good_buf)
2518 {
2519 /* no processing needed */
2520 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2521 }
2522
2523 /*ARGSUSED*/
2524 void
2525 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2526 {
2527 void *buf = zio_buf_alloc(zio->io_size);
2528
2529 bcopy(zio->io_data, buf, zio->io_size);
2530
2531 zcr->zcr_cbinfo = zio->io_size;
2532 zcr->zcr_cbdata = buf;
2533 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2534 zcr->zcr_free = zio_buf_free;
2535 }
2536
2537 static int
2538 zio_vdev_io_assess(zio_t *zio)
2539 {
2540 vdev_t *vd = zio->io_vd;
2541
2542 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2543 return (ZIO_PIPELINE_STOP);
2544
2545 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2546 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2547
2548 if (zio->io_vsd != NULL) {
2549 zio->io_vsd_ops->vsd_free(zio);
2550 zio->io_vsd = NULL;
2551 }
2552
2553 if (zio_injection_enabled && zio->io_error == 0)
2554 zio->io_error = zio_handle_fault_injection(zio, EIO);
2555
2556 /*
2557 * If the I/O failed, determine whether we should attempt to retry it.
2558 *
2559 * On retry, we cut in line in the issue queue, since we don't want
2560 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2561 */
2562 if (zio->io_error && vd == NULL &&
2563 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2564 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2565 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2566 zio->io_error = 0;
2567 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2568 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2569 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2570 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2571 zio_requeue_io_start_cut_in_line);
2572 return (ZIO_PIPELINE_STOP);
2573 }
2574
2575 /*
2576 * If we got an error on a leaf device, convert it to ENXIO
2577 * if the device is not accessible at all.
2578 */
2579 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2580 !vdev_accessible(vd, zio))
2581 zio->io_error = ENXIO;
2582
2583 /*
2584 * If we can't write to an interior vdev (mirror or RAID-Z),
2585 * set vdev_cant_write so that we stop trying to allocate from it.
2586 */
2587 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2588 vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2589 vd->vdev_cant_write = B_TRUE;
2590
2591 if (zio->io_error)
2592 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2593
2594 return (ZIO_PIPELINE_CONTINUE);
2595 }
2596
2597 void
2598 zio_vdev_io_reissue(zio_t *zio)
2599 {
2600 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2601 ASSERT(zio->io_error == 0);
2602
2603 zio->io_stage >>= 1;
2604 }
2605
2606 void
2607 zio_vdev_io_redone(zio_t *zio)
2608 {
2609 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2610
2611 zio->io_stage >>= 1;
2612 }
2613
2614 void
2615 zio_vdev_io_bypass(zio_t *zio)
2616 {
2617 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2618 ASSERT(zio->io_error == 0);
2619
2620 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2621 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2622 }
2623
2624 /*
2625 * ==========================================================================
2626 * Generate and verify checksums
2627 * ==========================================================================
2628 */
2629 static int
2630 zio_checksum_generate(zio_t *zio)
2631 {
2632 blkptr_t *bp = zio->io_bp;
2633 enum zio_checksum checksum;
2634
2635 if (bp == NULL) {
2636 /*
2637 * This is zio_write_phys().
2638 * We're either generating a label checksum, or none at all.
2639 */
2640 checksum = zio->io_prop.zp_checksum;
2641
2642 if (checksum == ZIO_CHECKSUM_OFF)
2643 return (ZIO_PIPELINE_CONTINUE);
2644
2645 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2646 } else {
2647 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2648 ASSERT(!IO_IS_ALLOCATING(zio));
2649 checksum = ZIO_CHECKSUM_GANG_HEADER;
2650 } else {
2651 checksum = BP_GET_CHECKSUM(bp);
2652 }
2653 }
2654
2655 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2656
2657 return (ZIO_PIPELINE_CONTINUE);
2658 }
2659
2660 static int
2661 zio_checksum_verify(zio_t *zio)
2662 {
2663 zio_bad_cksum_t info;
2664 blkptr_t *bp = zio->io_bp;
2665 int error;
2666
2667 ASSERT(zio->io_vd != NULL);
2668
2669 if (bp == NULL) {
2670 /*
2671 * This is zio_read_phys().
2672 * We're either verifying a label checksum, or nothing at all.
2673 */
2674 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2675 return (ZIO_PIPELINE_CONTINUE);
2676
2677 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2678 }
2679
2680 if ((error = zio_checksum_error(zio, &info)) != 0) {
2681 zio->io_error = error;
2682 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2683 zfs_ereport_start_checksum(zio->io_spa,
2684 zio->io_vd, zio, zio->io_offset,
2685 zio->io_size, NULL, &info);
2686 }
2687 }
2688
2689 return (ZIO_PIPELINE_CONTINUE);
2690 }
2691
2692 /*
2693 * Called by RAID-Z to ensure we don't compute the checksum twice.
2694 */
2695 void
2696 zio_checksum_verified(zio_t *zio)
2697 {
2698 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2699 }
2700
2701 /*
2702 * ==========================================================================
2703 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2704 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2705 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2706 * indicate errors that are specific to one I/O, and most likely permanent.
2707 * Any other error is presumed to be worse because we weren't expecting it.
2708 * ==========================================================================
2709 */
2710 int
2711 zio_worst_error(int e1, int e2)
2712 {
2713 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2714 int r1, r2;
2715
2716 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2717 if (e1 == zio_error_rank[r1])
2718 break;
2719
2720 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2721 if (e2 == zio_error_rank[r2])
2722 break;
2723
2724 return (r1 > r2 ? e1 : e2);
2725 }
2726
2727 /*
2728 * ==========================================================================
2729 * I/O completion
2730 * ==========================================================================
2731 */
2732 static int
2733 zio_ready(zio_t *zio)
2734 {
2735 blkptr_t *bp = zio->io_bp;
2736 zio_t *pio, *pio_next;
2737
2738 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2739 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2740 return (ZIO_PIPELINE_STOP);
2741
2742 if (zio->io_ready) {
2743 ASSERT(IO_IS_ALLOCATING(zio));
2744 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2745 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2746
2747 zio->io_ready(zio);
2748 }
2749
2750 if (bp != NULL && bp != &zio->io_bp_copy)
2751 zio->io_bp_copy = *bp;
2752
2753 if (zio->io_error)
2754 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2755
2756 mutex_enter(&zio->io_lock);
2757 zio->io_state[ZIO_WAIT_READY] = 1;
2758 pio = zio_walk_parents(zio);
2759 mutex_exit(&zio->io_lock);
2760
2761 /*
2762 * As we notify zio's parents, new parents could be added.
2763 * New parents go to the head of zio's io_parent_list, however,
2764 * so we will (correctly) not notify them. The remainder of zio's
2765 * io_parent_list, from 'pio_next' onward, cannot change because
2766 * all parents must wait for us to be done before they can be done.
2767 */
2768 for (; pio != NULL; pio = pio_next) {
2769 pio_next = zio_walk_parents(zio);
2770 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2771 }
2772
2773 if (zio->io_flags & ZIO_FLAG_NODATA) {
2774 if (BP_IS_GANG(bp)) {
2775 zio->io_flags &= ~ZIO_FLAG_NODATA;
2776 } else {
2777 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2778 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2779 }
2780 }
2781
2782 if (zio_injection_enabled &&
2783 zio->io_spa->spa_syncing_txg == zio->io_txg)
2784 zio_handle_ignored_writes(zio);
2785
2786 return (ZIO_PIPELINE_CONTINUE);
2787 }
2788
2789 static int
2790 zio_done(zio_t *zio)
2791 {
2792 zio_t *pio, *pio_next;
2793 int c, w;
2794
2795 /*
2796 * If our children haven't all completed,
2797 * wait for them and then repeat this pipeline stage.
2798 */
2799 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2800 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2801 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2802 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2803 return (ZIO_PIPELINE_STOP);
2804
2805 for (c = 0; c < ZIO_CHILD_TYPES; c++)
2806 for (w = 0; w < ZIO_WAIT_TYPES; w++)
2807 ASSERT(zio->io_children[c][w] == 0);
2808
2809 if (zio->io_bp != NULL) {
2810 ASSERT(zio->io_bp->blk_pad[0] == 0);
2811 ASSERT(zio->io_bp->blk_pad[1] == 0);
2812 ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2813 (zio->io_bp == zio_unique_parent(zio)->io_bp));
2814 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) &&
2815 zio->io_bp_override == NULL &&
2816 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2817 ASSERT(!BP_SHOULD_BYTESWAP(zio->io_bp));
2818 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2819 ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 ||
2820 (BP_COUNT_GANG(zio->io_bp) == BP_GET_NDVAS(zio->io_bp)));
2821 }
2822 }
2823
2824 /*
2825 * If there were child vdev/gang/ddt errors, they apply to us now.
2826 */
2827 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2828 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2829 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2830
2831 /*
2832 * If the I/O on the transformed data was successful, generate any
2833 * checksum reports now while we still have the transformed data.
2834 */
2835 if (zio->io_error == 0) {
2836 while (zio->io_cksum_report != NULL) {
2837 zio_cksum_report_t *zcr = zio->io_cksum_report;
2838 uint64_t align = zcr->zcr_align;
2839 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2840 char *abuf = zio->io_data;
2841
2842 if (asize != zio->io_size) {
2843 abuf = zio_buf_alloc(asize);
2844 bcopy(zio->io_data, abuf, zio->io_size);
2845 bzero(abuf + zio->io_size, asize - zio->io_size);
2846 }
2847
2848 zio->io_cksum_report = zcr->zcr_next;
2849 zcr->zcr_next = NULL;
2850 zcr->zcr_finish(zcr, abuf);
2851 zfs_ereport_free_checksum(zcr);
2852
2853 if (asize != zio->io_size)
2854 zio_buf_free(abuf, asize);
2855 }
2856 }
2857
2858 zio_pop_transforms(zio); /* note: may set zio->io_error */
2859
2860 vdev_stat_update(zio, zio->io_size);
2861
2862 /*
2863 * If this I/O is attached to a particular vdev is slow, exeeding
2864 * 30 seconds to complete, post an error described the I/O delay.
2865 * We ignore these errors if the device is currently unavailable.
2866 */
2867 if (zio->io_delay >= zio_delay_max) {
2868 if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd))
2869 zfs_ereport_post(FM_EREPORT_ZFS_DELAY, zio->io_spa,
2870 zio->io_vd, zio, 0, 0);
2871 }
2872
2873 if (zio->io_error) {
2874 /*
2875 * If this I/O is attached to a particular vdev,
2876 * generate an error message describing the I/O failure
2877 * at the block level. We ignore these errors if the
2878 * device is currently unavailable.
2879 */
2880 if (zio->io_error != ECKSUM && zio->io_vd != NULL &&
2881 !vdev_is_dead(zio->io_vd))
2882 zfs_ereport_post(FM_EREPORT_ZFS_IO, zio->io_spa,
2883 zio->io_vd, zio, 0, 0);
2884
2885 if ((zio->io_error == EIO || !(zio->io_flags &
2886 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2887 zio == zio->io_logical) {
2888 /*
2889 * For logical I/O requests, tell the SPA to log the
2890 * error and generate a logical data ereport.
2891 */
2892 spa_log_error(zio->io_spa, zio);
2893 zfs_ereport_post(FM_EREPORT_ZFS_DATA, zio->io_spa, NULL, zio,
2894 0, 0);
2895 }
2896 }
2897
2898 if (zio->io_error && zio == zio->io_logical) {
2899 /*
2900 * Determine whether zio should be reexecuted. This will
2901 * propagate all the way to the root via zio_notify_parent().
2902 */
2903 ASSERT(zio->io_vd == NULL && zio->io_bp != NULL);
2904 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2905
2906 if (IO_IS_ALLOCATING(zio) &&
2907 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2908 if (zio->io_error != ENOSPC)
2909 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2910 else
2911 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2912 }
2913
2914 if ((zio->io_type == ZIO_TYPE_READ ||
2915 zio->io_type == ZIO_TYPE_FREE) &&
2916 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2917 zio->io_error == ENXIO &&
2918 spa_load_state(zio->io_spa) == SPA_LOAD_NONE &&
2919 spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE)
2920 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2921
2922 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2923 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2924
2925 /*
2926 * Here is a possibly good place to attempt to do
2927 * either combinatorial reconstruction or error correction
2928 * based on checksums. It also might be a good place
2929 * to send out preliminary ereports before we suspend
2930 * processing.
2931 */
2932 }
2933
2934 /*
2935 * If there were logical child errors, they apply to us now.
2936 * We defer this until now to avoid conflating logical child
2937 * errors with errors that happened to the zio itself when
2938 * updating vdev stats and reporting FMA events above.
2939 */
2940 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2941
2942 if ((zio->io_error || zio->io_reexecute) &&
2943 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2944 !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2945 zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp);
2946
2947 zio_gang_tree_free(&zio->io_gang_tree);
2948
2949 /*
2950 * Godfather I/Os should never suspend.
2951 */
2952 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2953 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2954 zio->io_reexecute = 0;
2955
2956 if (zio->io_reexecute) {
2957 /*
2958 * This is a logical I/O that wants to reexecute.
2959 *
2960 * Reexecute is top-down. When an i/o fails, if it's not
2961 * the root, it simply notifies its parent and sticks around.
2962 * The parent, seeing that it still has children in zio_done(),
2963 * does the same. This percolates all the way up to the root.
2964 * The root i/o will reexecute or suspend the entire tree.
2965 *
2966 * This approach ensures that zio_reexecute() honors
2967 * all the original i/o dependency relationships, e.g.
2968 * parents not executing until children are ready.
2969 */
2970 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2971
2972 zio->io_gang_leader = NULL;
2973
2974 mutex_enter(&zio->io_lock);
2975 zio->io_state[ZIO_WAIT_DONE] = 1;
2976 mutex_exit(&zio->io_lock);
2977
2978 /*
2979 * "The Godfather" I/O monitors its children but is
2980 * not a true parent to them. It will track them through
2981 * the pipeline but severs its ties whenever they get into
2982 * trouble (e.g. suspended). This allows "The Godfather"
2983 * I/O to return status without blocking.
2984 */
2985 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2986 zio_link_t *zl = zio->io_walk_link;
2987 pio_next = zio_walk_parents(zio);
2988
2989 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
2990 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
2991 zio_remove_child(pio, zio, zl);
2992 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2993 }
2994 }
2995
2996 if ((pio = zio_unique_parent(zio)) != NULL) {
2997 /*
2998 * We're not a root i/o, so there's nothing to do
2999 * but notify our parent. Don't propagate errors
3000 * upward since we haven't permanently failed yet.
3001 */
3002 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3003 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3004 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3005 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3006 /*
3007 * We'd fail again if we reexecuted now, so suspend
3008 * until conditions improve (e.g. device comes online).
3009 */
3010 zio_suspend(zio->io_spa, zio);
3011 } else {
3012 /*
3013 * Reexecution is potentially a huge amount of work.
3014 * Hand it off to the otherwise-unused claim taskq.
3015 */
3016 ASSERT(taskq_empty_ent(&zio->io_tqent));
3017 (void) taskq_dispatch_ent(
3018 zio->io_spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
3019 (task_func_t *)zio_reexecute, zio, 0,
3020 &zio->io_tqent);
3021 }
3022 return (ZIO_PIPELINE_STOP);
3023 }
3024
3025 ASSERT(zio->io_child_count == 0);
3026 ASSERT(zio->io_reexecute == 0);
3027 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3028
3029 /*
3030 * Report any checksum errors, since the I/O is complete.
3031 */
3032 while (zio->io_cksum_report != NULL) {
3033 zio_cksum_report_t *zcr = zio->io_cksum_report;
3034 zio->io_cksum_report = zcr->zcr_next;
3035 zcr->zcr_next = NULL;
3036 zcr->zcr_finish(zcr, NULL);
3037 zfs_ereport_free_checksum(zcr);
3038 }
3039
3040 /*
3041 * It is the responsibility of the done callback to ensure that this
3042 * particular zio is no longer discoverable for adoption, and as
3043 * such, cannot acquire any new parents.
3044 */
3045 if (zio->io_done)
3046 zio->io_done(zio);
3047
3048 mutex_enter(&zio->io_lock);
3049 zio->io_state[ZIO_WAIT_DONE] = 1;
3050 mutex_exit(&zio->io_lock);
3051
3052 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3053 zio_link_t *zl = zio->io_walk_link;
3054 pio_next = zio_walk_parents(zio);
3055 zio_remove_child(pio, zio, zl);
3056 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3057 }
3058
3059 if (zio->io_waiter != NULL) {
3060 mutex_enter(&zio->io_lock);
3061 zio->io_executor = NULL;
3062 cv_broadcast(&zio->io_cv);
3063 mutex_exit(&zio->io_lock);
3064 } else {
3065 zio_destroy(zio);
3066 }
3067
3068 return (ZIO_PIPELINE_STOP);
3069 }
3070
3071 /*
3072 * ==========================================================================
3073 * I/O pipeline definition
3074 * ==========================================================================
3075 */
3076 static zio_pipe_stage_t *zio_pipeline[] = {
3077 NULL,
3078 zio_read_bp_init,
3079 zio_free_bp_init,
3080 zio_issue_async,
3081 zio_write_bp_init,
3082 zio_checksum_generate,
3083 zio_ddt_read_start,
3084 zio_ddt_read_done,
3085 zio_ddt_write,
3086 zio_ddt_free,
3087 zio_gang_assemble,
3088 zio_gang_issue,
3089 zio_dva_allocate,
3090 zio_dva_free,
3091 zio_dva_claim,
3092 zio_ready,
3093 zio_vdev_io_start,
3094 zio_vdev_io_done,
3095 zio_vdev_io_assess,
3096 zio_checksum_verify,
3097 zio_done
3098 };
3099
3100 #if defined(_KERNEL) && defined(HAVE_SPL)
3101 /* Fault injection */
3102 EXPORT_SYMBOL(zio_injection_enabled);
3103 EXPORT_SYMBOL(zio_inject_fault);
3104 EXPORT_SYMBOL(zio_inject_list_next);
3105 EXPORT_SYMBOL(zio_clear_fault);
3106 EXPORT_SYMBOL(zio_handle_fault_injection);
3107 EXPORT_SYMBOL(zio_handle_device_injection);
3108 EXPORT_SYMBOL(zio_handle_label_injection);
3109 EXPORT_SYMBOL(zio_priority_table);
3110 EXPORT_SYMBOL(zio_type_name);
3111
3112 module_param(zio_bulk_flags, int, 0644);
3113 MODULE_PARM_DESC(zio_bulk_flags, "Additional flags to pass to bulk buffers");
3114
3115 module_param(zio_delay_max, int, 0644);
3116 MODULE_PARM_DESC(zio_delay_max, "Max zio millisec delay before posting event");
3117
3118 module_param(zio_requeue_io_start_cut_in_line, int, 0644);
3119 MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line, "Prioritize requeued I/O");
3120 #endif
ZFS On Linux mirror for PVE