?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
3329 spa_sync() spends 10-20% of its time in spa_free_sync_cb()
[unleashed.git] / usr / src / uts / common / fs / zfs / zio.c
blob79c634d94cb336507d71e815dcfdf6f3168816ff
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 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
67 "zio_ioctl"
68 };
69
70 /*
71 * ==========================================================================
72 * I/O kmem caches
73 * ==========================================================================
74 */
75 kmem_cache_t *zio_cache;
76 kmem_cache_t *zio_link_cache;
77 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
78 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
79
80 #ifdef _KERNEL
81 extern vmem_t *zio_alloc_arena;
82 #endif
83 extern int zfs_mg_alloc_failures;
84
85 /*
86 * The following actions directly effect the spa's sync-to-convergence logic.
87 * The values below define the sync pass when we start performing the action.
88 * Care should be taken when changing these values as they directly impact
89 * spa_sync() performance. Tuning these values may introduce subtle performance
90 * pathologies and should only be done in the context of performance analysis.
91 * These tunables will eventually be removed and replaced with #defines once
92 * enough analysis has been done to determine optimal values.
93 *
94 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
95 * regular blocks are not deferred.
96 */
97 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
98 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
99 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
100
101 /*
102 * An allocating zio is one that either currently has the DVA allocate
103 * stage set or will have it later in its lifetime.
104 */
105 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
106
107 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
108
109 #ifdef ZFS_DEBUG
110 int zio_buf_debug_limit = 16384;
111 #else
112 int zio_buf_debug_limit = 0;
113 #endif
114
115 void
116 zio_init(void)
117 {
118 size_t c;
119 vmem_t *data_alloc_arena = NULL;
120
121 #ifdef _KERNEL
122 data_alloc_arena = zio_alloc_arena;
123 #endif
124 zio_cache = kmem_cache_create("zio_cache",
125 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
126 zio_link_cache = kmem_cache_create("zio_link_cache",
127 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
128
129 /*
130 * For small buffers, we want a cache for each multiple of
131 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
132 * for each quarter-power of 2. For large buffers, we want
133 * a cache for each multiple of PAGESIZE.
134 */
135 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
136 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
137 size_t p2 = size;
138 size_t align = 0;
139 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
140
141 while (p2 & (p2 - 1))
142 p2 &= p2 - 1;
143
144 #ifndef _KERNEL
145 /*
146 * If we are using watchpoints, put each buffer on its own page,
147 * to eliminate the performance overhead of trapping to the
148 * kernel when modifying a non-watched buffer that shares the
149 * page with a watched buffer.
150 */
151 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
152 continue;
153 #endif
154 if (size <= 4 * SPA_MINBLOCKSIZE) {
155 align = SPA_MINBLOCKSIZE;
156 } else if (IS_P2ALIGNED(size, PAGESIZE)) {
157 align = PAGESIZE;
158 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
159 align = p2 >> 2;
160 }
161
162 if (align != 0) {
163 char name[36];
164 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
165 zio_buf_cache[c] = kmem_cache_create(name, size,
166 align, NULL, NULL, NULL, NULL, NULL, cflags);
167
168 /*
169 * Since zio_data bufs do not appear in crash dumps, we
170 * pass KMC_NOTOUCH so that no allocator metadata is
171 * stored with the buffers.
172 */
173 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
174 zio_data_buf_cache[c] = kmem_cache_create(name, size,
175 align, NULL, NULL, NULL, NULL, data_alloc_arena,
176 cflags | KMC_NOTOUCH);
177 }
178 }
179
180 while (--c != 0) {
181 ASSERT(zio_buf_cache[c] != NULL);
182 if (zio_buf_cache[c - 1] == NULL)
183 zio_buf_cache[c - 1] = zio_buf_cache[c];
184
185 ASSERT(zio_data_buf_cache[c] != NULL);
186 if (zio_data_buf_cache[c - 1] == NULL)
187 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
188 }
189
190 /*
191 * The zio write taskqs have 1 thread per cpu, allow 1/2 of the taskqs
192 * to fail 3 times per txg or 8 failures, whichever is greater.
193 */
194 zfs_mg_alloc_failures = MAX((3 * max_ncpus / 2), 8);
195
196 zio_inject_init();
197 }
198
199 void
200 zio_fini(void)
201 {
202 size_t c;
203 kmem_cache_t *last_cache = NULL;
204 kmem_cache_t *last_data_cache = NULL;
205
206 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
207 if (zio_buf_cache[c] != last_cache) {
208 last_cache = zio_buf_cache[c];
209 kmem_cache_destroy(zio_buf_cache[c]);
210 }
211 zio_buf_cache[c] = NULL;
212
213 if (zio_data_buf_cache[c] != last_data_cache) {
214 last_data_cache = zio_data_buf_cache[c];
215 kmem_cache_destroy(zio_data_buf_cache[c]);
216 }
217 zio_data_buf_cache[c] = NULL;
218 }
219
220 kmem_cache_destroy(zio_link_cache);
221 kmem_cache_destroy(zio_cache);
222
223 zio_inject_fini();
224 }
225
226 /*
227 * ==========================================================================
228 * Allocate and free I/O buffers
229 * ==========================================================================
230 */
231
232 /*
233 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
234 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
235 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
236 * excess / transient data in-core during a crashdump.
237 */
238 void *
239 zio_buf_alloc(size_t size)
240 {
241 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
242
243 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
244
245 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
246 }
247
248 /*
249 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
250 * crashdump if the kernel panics. This exists so that we will limit the amount
251 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
252 * of kernel heap dumped to disk when the kernel panics)
253 */
254 void *
255 zio_data_buf_alloc(size_t size)
256 {
257 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
258
259 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
260
261 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
262 }
263
264 void
265 zio_buf_free(void *buf, size_t size)
266 {
267 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
268
269 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
270
271 kmem_cache_free(zio_buf_cache[c], buf);
272 }
273
274 void
275 zio_data_buf_free(void *buf, size_t size)
276 {
277 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
278
279 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
280
281 kmem_cache_free(zio_data_buf_cache[c], buf);
282 }
283
284 /*
285 * ==========================================================================
286 * Push and pop I/O transform buffers
287 * ==========================================================================
288 */
289 static void
290 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
291 zio_transform_func_t *transform)
292 {
293 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
294
295 zt->zt_orig_data = zio->io_data;
296 zt->zt_orig_size = zio->io_size;
297 zt->zt_bufsize = bufsize;
298 zt->zt_transform = transform;
299
300 zt->zt_next = zio->io_transform_stack;
301 zio->io_transform_stack = zt;
302
303 zio->io_data = data;
304 zio->io_size = size;
305 }
306
307 static void
308 zio_pop_transforms(zio_t *zio)
309 {
310 zio_transform_t *zt;
311
312 while ((zt = zio->io_transform_stack) != NULL) {
313 if (zt->zt_transform != NULL)
314 zt->zt_transform(zio,
315 zt->zt_orig_data, zt->zt_orig_size);
316
317 if (zt->zt_bufsize != 0)
318 zio_buf_free(zio->io_data, zt->zt_bufsize);
319
320 zio->io_data = zt->zt_orig_data;
321 zio->io_size = zt->zt_orig_size;
322 zio->io_transform_stack = zt->zt_next;
323
324 kmem_free(zt, sizeof (zio_transform_t));
325 }
326 }
327
328 /*
329 * ==========================================================================
330 * I/O transform callbacks for subblocks and decompression
331 * ==========================================================================
332 */
333 static void
334 zio_subblock(zio_t *zio, void *data, uint64_t size)
335 {
336 ASSERT(zio->io_size > size);
337
338 if (zio->io_type == ZIO_TYPE_READ)
339 bcopy(zio->io_data, data, size);
340 }
341
342 static void
343 zio_decompress(zio_t *zio, void *data, uint64_t size)
344 {
345 if (zio->io_error == 0 &&
346 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
347 zio->io_data, data, zio->io_size, size) != 0)
348 zio->io_error = EIO;
349 }
350
351 /*
352 * ==========================================================================
353 * I/O parent/child relationships and pipeline interlocks
354 * ==========================================================================
355 */
356 /*
357 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
358 * continue calling these functions until they return NULL.
359 * Otherwise, the next caller will pick up the list walk in
360 * some indeterminate state. (Otherwise every caller would
361 * have to pass in a cookie to keep the state represented by
362 * io_walk_link, which gets annoying.)
363 */
364 zio_t *
365 zio_walk_parents(zio_t *cio)
366 {
367 zio_link_t *zl = cio->io_walk_link;
368 list_t *pl = &cio->io_parent_list;
369
370 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
371 cio->io_walk_link = zl;
372
373 if (zl == NULL)
374 return (NULL);
375
376 ASSERT(zl->zl_child == cio);
377 return (zl->zl_parent);
378 }
379
380 zio_t *
381 zio_walk_children(zio_t *pio)
382 {
383 zio_link_t *zl = pio->io_walk_link;
384 list_t *cl = &pio->io_child_list;
385
386 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
387 pio->io_walk_link = zl;
388
389 if (zl == NULL)
390 return (NULL);
391
392 ASSERT(zl->zl_parent == pio);
393 return (zl->zl_child);
394 }
395
396 zio_t *
397 zio_unique_parent(zio_t *cio)
398 {
399 zio_t *pio = zio_walk_parents(cio);
400
401 VERIFY(zio_walk_parents(cio) == NULL);
402 return (pio);
403 }
404
405 void
406 zio_add_child(zio_t *pio, zio_t *cio)
407 {
408 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
409
410 /*
411 * Logical I/Os can have logical, gang, or vdev children.
412 * Gang I/Os can have gang or vdev children.
413 * Vdev I/Os can only have vdev children.
414 * The following ASSERT captures all of these constraints.
415 */
416 ASSERT(cio->io_child_type <= pio->io_child_type);
417
418 zl->zl_parent = pio;
419 zl->zl_child = cio;
420
421 mutex_enter(&cio->io_lock);
422 mutex_enter(&pio->io_lock);
423
424 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
425
426 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
427 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
428
429 list_insert_head(&pio->io_child_list, zl);
430 list_insert_head(&cio->io_parent_list, zl);
431
432 pio->io_child_count++;
433 cio->io_parent_count++;
434
435 mutex_exit(&pio->io_lock);
436 mutex_exit(&cio->io_lock);
437 }
438
439 static void
440 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
441 {
442 ASSERT(zl->zl_parent == pio);
443 ASSERT(zl->zl_child == cio);
444
445 mutex_enter(&cio->io_lock);
446 mutex_enter(&pio->io_lock);
447
448 list_remove(&pio->io_child_list, zl);
449 list_remove(&cio->io_parent_list, zl);
450
451 pio->io_child_count--;
452 cio->io_parent_count--;
453
454 mutex_exit(&pio->io_lock);
455 mutex_exit(&cio->io_lock);
456
457 kmem_cache_free(zio_link_cache, zl);
458 }
459
460 static boolean_t
461 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
462 {
463 uint64_t *countp = &zio->io_children[child][wait];
464 boolean_t waiting = B_FALSE;
465
466 mutex_enter(&zio->io_lock);
467 ASSERT(zio->io_stall == NULL);
468 if (*countp != 0) {
469 zio->io_stage >>= 1;
470 zio->io_stall = countp;
471 waiting = B_TRUE;
472 }
473 mutex_exit(&zio->io_lock);
474
475 return (waiting);
476 }
477
478 static void
479 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
480 {
481 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
482 int *errorp = &pio->io_child_error[zio->io_child_type];
483
484 mutex_enter(&pio->io_lock);
485 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
486 *errorp = zio_worst_error(*errorp, zio->io_error);
487 pio->io_reexecute |= zio->io_reexecute;
488 ASSERT3U(*countp, >, 0);
489 if (--*countp == 0 && pio->io_stall == countp) {
490 pio->io_stall = NULL;
491 mutex_exit(&pio->io_lock);
492 zio_execute(pio);
493 } else {
494 mutex_exit(&pio->io_lock);
495 }
496 }
497
498 static void
499 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
500 {
501 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
502 zio->io_error = zio->io_child_error[c];
503 }
504
505 /*
506 * ==========================================================================
507 * Create the various types of I/O (read, write, free, etc)
508 * ==========================================================================
509 */
510 static zio_t *
511 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
512 void *data, uint64_t size, zio_done_func_t *done, void *private,
513 zio_type_t type, int priority, enum zio_flag flags,
514 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
515 enum zio_stage stage, enum zio_stage pipeline)
516 {
517 zio_t *zio;
518
519 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
520 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
521 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
522
523 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
524 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
525 ASSERT(vd || stage == ZIO_STAGE_OPEN);
526
527 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
528 bzero(zio, sizeof (zio_t));
529
530 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
531 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
532
533 list_create(&zio->io_parent_list, sizeof (zio_link_t),
534 offsetof(zio_link_t, zl_parent_node));
535 list_create(&zio->io_child_list, sizeof (zio_link_t),
536 offsetof(zio_link_t, zl_child_node));
537
538 if (vd != NULL)
539 zio->io_child_type = ZIO_CHILD_VDEV;
540 else if (flags & ZIO_FLAG_GANG_CHILD)
541 zio->io_child_type = ZIO_CHILD_GANG;
542 else if (flags & ZIO_FLAG_DDT_CHILD)
543 zio->io_child_type = ZIO_CHILD_DDT;
544 else
545 zio->io_child_type = ZIO_CHILD_LOGICAL;
546
547 if (bp != NULL) {
548 zio->io_bp = (blkptr_t *)bp;
549 zio->io_bp_copy = *bp;
550 zio->io_bp_orig = *bp;
551 if (type != ZIO_TYPE_WRITE ||
552 zio->io_child_type == ZIO_CHILD_DDT)
553 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
554 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
555 zio->io_logical = zio;
556 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
557 pipeline |= ZIO_GANG_STAGES;
558 }
559
560 zio->io_spa = spa;
561 zio->io_txg = txg;
562 zio->io_done = done;
563 zio->io_private = private;
564 zio->io_type = type;
565 zio->io_priority = priority;
566 zio->io_vd = vd;
567 zio->io_offset = offset;
568 zio->io_orig_data = zio->io_data = data;
569 zio->io_orig_size = zio->io_size = size;
570 zio->io_orig_flags = zio->io_flags = flags;
571 zio->io_orig_stage = zio->io_stage = stage;
572 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
573
574 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
575 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
576
577 if (zb != NULL)
578 zio->io_bookmark = *zb;
579
580 if (pio != NULL) {
581 if (zio->io_logical == NULL)
582 zio->io_logical = pio->io_logical;
583 if (zio->io_child_type == ZIO_CHILD_GANG)
584 zio->io_gang_leader = pio->io_gang_leader;
585 zio_add_child(pio, zio);
586 }
587
588 return (zio);
589 }
590
591 static void
592 zio_destroy(zio_t *zio)
593 {
594 list_destroy(&zio->io_parent_list);
595 list_destroy(&zio->io_child_list);
596 mutex_destroy(&zio->io_lock);
597 cv_destroy(&zio->io_cv);
598 kmem_cache_free(zio_cache, zio);
599 }
600
601 zio_t *
602 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
603 void *private, enum zio_flag flags)
604 {
605 zio_t *zio;
606
607 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
608 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
609 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
610
611 return (zio);
612 }
613
614 zio_t *
615 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
616 {
617 return (zio_null(NULL, spa, NULL, done, private, flags));
618 }
619
620 zio_t *
621 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
622 void *data, uint64_t size, zio_done_func_t *done, void *private,
623 int priority, enum zio_flag flags, const zbookmark_t *zb)
624 {
625 zio_t *zio;
626
627 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
628 data, size, done, private,
629 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
630 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
631 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
632
633 return (zio);
634 }
635
636 zio_t *
637 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
638 void *data, uint64_t size, const zio_prop_t *zp,
639 zio_done_func_t *ready, zio_done_func_t *done, void *private,
640 int priority, enum zio_flag flags, const zbookmark_t *zb)
641 {
642 zio_t *zio;
643
644 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
645 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
646 zp->zp_compress >= ZIO_COMPRESS_OFF &&
647 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
648 DMU_OT_IS_VALID(zp->zp_type) &&
649 zp->zp_level < 32 &&
650 zp->zp_copies > 0 &&
651 zp->zp_copies <= spa_max_replication(spa) &&
652 zp->zp_dedup <= 1 &&
653 zp->zp_dedup_verify <= 1);
654
655 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
656 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
657 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
658 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
659
660 zio->io_ready = ready;
661 zio->io_prop = *zp;
662
663 return (zio);
664 }
665
666 zio_t *
667 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
668 uint64_t size, zio_done_func_t *done, void *private, int priority,
669 enum zio_flag flags, zbookmark_t *zb)
670 {
671 zio_t *zio;
672
673 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
674 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
675 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
676
677 return (zio);
678 }
679
680 void
681 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
682 {
683 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
684 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
685 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
686 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
687
688 zio->io_prop.zp_copies = copies;
689 zio->io_bp_override = bp;
690 }
691
692 void
693 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
694 {
695 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
696 }
697
698 zio_t *
699 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
700 enum zio_flag flags)
701 {
702 zio_t *zio;
703
704 dprintf_bp(bp, "freeing in txg %llu, pass %u",
705 (longlong_t)txg, spa->spa_sync_pass);
706
707 ASSERT(!BP_IS_HOLE(bp));
708 ASSERT(spa_syncing_txg(spa) == txg);
709 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
710
711 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
712 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
713 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
714
715 return (zio);
716 }
717
718 zio_t *
719 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
720 zio_done_func_t *done, void *private, enum zio_flag flags)
721 {
722 zio_t *zio;
723
724 /*
725 * A claim is an allocation of a specific block. Claims are needed
726 * to support immediate writes in the intent log. The issue is that
727 * immediate writes contain committed data, but in a txg that was
728 * *not* committed. Upon opening the pool after an unclean shutdown,
729 * the intent log claims all blocks that contain immediate write data
730 * so that the SPA knows they're in use.
731 *
732 * All claims *must* be resolved in the first txg -- before the SPA
733 * starts allocating blocks -- so that nothing is allocated twice.
734 * If txg == 0 we just verify that the block is claimable.
735 */
736 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
737 ASSERT(txg == spa_first_txg(spa) || txg == 0);
738 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
739
740 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
741 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
742 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
743
744 return (zio);
745 }
746
747 zio_t *
748 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
749 zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
750 {
751 zio_t *zio;
752 int c;
753
754 if (vd->vdev_children == 0) {
755 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
756 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
757 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
758
759 zio->io_cmd = cmd;
760 } else {
761 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
762
763 for (c = 0; c < vd->vdev_children; c++)
764 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
765 done, private, priority, flags));
766 }
767
768 return (zio);
769 }
770
771 zio_t *
772 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
773 void *data, int checksum, zio_done_func_t *done, void *private,
774 int priority, enum zio_flag flags, boolean_t labels)
775 {
776 zio_t *zio;
777
778 ASSERT(vd->vdev_children == 0);
779 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
780 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
781 ASSERT3U(offset + size, <=, vd->vdev_psize);
782
783 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
784 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
785 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
786
787 zio->io_prop.zp_checksum = checksum;
788
789 return (zio);
790 }
791
792 zio_t *
793 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
794 void *data, int checksum, zio_done_func_t *done, void *private,
795 int priority, enum zio_flag flags, boolean_t labels)
796 {
797 zio_t *zio;
798
799 ASSERT(vd->vdev_children == 0);
800 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
801 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
802 ASSERT3U(offset + size, <=, vd->vdev_psize);
803
804 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
805 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
806 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
807
808 zio->io_prop.zp_checksum = checksum;
809
810 if (zio_checksum_table[checksum].ci_eck) {
811 /*
812 * zec checksums are necessarily destructive -- they modify
813 * the end of the write buffer to hold the verifier/checksum.
814 * Therefore, we must make a local copy in case the data is
815 * being written to multiple places in parallel.
816 */
817 void *wbuf = zio_buf_alloc(size);
818 bcopy(data, wbuf, size);
819 zio_push_transform(zio, wbuf, size, size, NULL);
820 }
821
822 return (zio);
823 }
824
825 /*
826 * Create a child I/O to do some work for us.
827 */
828 zio_t *
829 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
830 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
831 zio_done_func_t *done, void *private)
832 {
833 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
834 zio_t *zio;
835
836 ASSERT(vd->vdev_parent ==
837 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
838
839 if (type == ZIO_TYPE_READ && bp != NULL) {
840 /*
841 * If we have the bp, then the child should perform the
842 * checksum and the parent need not. This pushes error
843 * detection as close to the leaves as possible and
844 * eliminates redundant checksums in the interior nodes.
845 */
846 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
847 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
848 }
849
850 if (vd->vdev_children == 0)
851 offset += VDEV_LABEL_START_SIZE;
852
853 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
854
855 /*
856 * If we've decided to do a repair, the write is not speculative --
857 * even if the original read was.
858 */
859 if (flags & ZIO_FLAG_IO_REPAIR)
860 flags &= ~ZIO_FLAG_SPECULATIVE;
861
862 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
863 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
864 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
865
866 return (zio);
867 }
868
869 zio_t *
870 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
871 int type, int priority, enum zio_flag flags,
872 zio_done_func_t *done, void *private)
873 {
874 zio_t *zio;
875
876 ASSERT(vd->vdev_ops->vdev_op_leaf);
877
878 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
879 data, size, done, private, type, priority,
880 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
881 vd, offset, NULL,
882 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
883
884 return (zio);
885 }
886
887 void
888 zio_flush(zio_t *zio, vdev_t *vd)
889 {
890 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
891 NULL, NULL, ZIO_PRIORITY_NOW,
892 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
893 }
894
895 void
896 zio_shrink(zio_t *zio, uint64_t size)
897 {
898 ASSERT(zio->io_executor == NULL);
899 ASSERT(zio->io_orig_size == zio->io_size);
900 ASSERT(size <= zio->io_size);
901
902 /*
903 * We don't shrink for raidz because of problems with the
904 * reconstruction when reading back less than the block size.
905 * Note, BP_IS_RAIDZ() assumes no compression.
906 */
907 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
908 if (!BP_IS_RAIDZ(zio->io_bp))
909 zio->io_orig_size = zio->io_size = size;
910 }
911
912 /*
913 * ==========================================================================
914 * Prepare to read and write logical blocks
915 * ==========================================================================
916 */
917
918 static int
919 zio_read_bp_init(zio_t *zio)
920 {
921 blkptr_t *bp = zio->io_bp;
922
923 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
924 zio->io_child_type == ZIO_CHILD_LOGICAL &&
925 !(zio->io_flags & ZIO_FLAG_RAW)) {
926 uint64_t psize = BP_GET_PSIZE(bp);
927 void *cbuf = zio_buf_alloc(psize);
928
929 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
930 }
931
932 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
933 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
934
935 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
936 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
937
938 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
939 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
940
941 return (ZIO_PIPELINE_CONTINUE);
942 }
943
944 static int
945 zio_write_bp_init(zio_t *zio)
946 {
947 spa_t *spa = zio->io_spa;
948 zio_prop_t *zp = &zio->io_prop;
949 enum zio_compress compress = zp->zp_compress;
950 blkptr_t *bp = zio->io_bp;
951 uint64_t lsize = zio->io_size;
952 uint64_t psize = lsize;
953 int pass = 1;
954
955 /*
956 * If our children haven't all reached the ready stage,
957 * wait for them and then repeat this pipeline stage.
958 */
959 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
960 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
961 return (ZIO_PIPELINE_STOP);
962
963 if (!IO_IS_ALLOCATING(zio))
964 return (ZIO_PIPELINE_CONTINUE);
965
966 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
967
968 if (zio->io_bp_override) {
969 ASSERT(bp->blk_birth != zio->io_txg);
970 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
971
972 *bp = *zio->io_bp_override;
973 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
974
975 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
976 return (ZIO_PIPELINE_CONTINUE);
977
978 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
979 zp->zp_dedup_verify);
980
981 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
982 BP_SET_DEDUP(bp, 1);
983 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
984 return (ZIO_PIPELINE_CONTINUE);
985 }
986 zio->io_bp_override = NULL;
987 BP_ZERO(bp);
988 }
989
990 if (bp->blk_birth == zio->io_txg) {
991 /*
992 * We're rewriting an existing block, which means we're
993 * working on behalf of spa_sync(). For spa_sync() to
994 * converge, it must eventually be the case that we don't
995 * have to allocate new blocks. But compression changes
996 * the blocksize, which forces a reallocate, and makes
997 * convergence take longer. Therefore, after the first
998 * few passes, stop compressing to ensure convergence.
999 */
1000 pass = spa_sync_pass(spa);
1001
1002 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1003 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1004 ASSERT(!BP_GET_DEDUP(bp));
1005
1006 if (pass >= zfs_sync_pass_dont_compress)
1007 compress = ZIO_COMPRESS_OFF;
1008
1009 /* Make sure someone doesn't change their mind on overwrites */
1010 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
1011 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1012 }
1013
1014 if (compress != ZIO_COMPRESS_OFF) {
1015 void *cbuf = zio_buf_alloc(lsize);
1016 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1017 if (psize == 0 || psize == lsize) {
1018 compress = ZIO_COMPRESS_OFF;
1019 zio_buf_free(cbuf, lsize);
1020 } else {
1021 ASSERT(psize < lsize);
1022 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1023 }
1024 }
1025
1026 /*
1027 * The final pass of spa_sync() must be all rewrites, but the first
1028 * few passes offer a trade-off: allocating blocks defers convergence,
1029 * but newly allocated blocks are sequential, so they can be written
1030 * to disk faster. Therefore, we allow the first few passes of
1031 * spa_sync() to allocate new blocks, but force rewrites after that.
1032 * There should only be a handful of blocks after pass 1 in any case.
1033 */
1034 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1035 pass >= zfs_sync_pass_rewrite) {
1036 ASSERT(psize != 0);
1037 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1038 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1039 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1040 } else {
1041 BP_ZERO(bp);
1042 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1043 }
1044
1045 if (psize == 0) {
1046 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1047 } else {
1048 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1049 BP_SET_LSIZE(bp, lsize);
1050 BP_SET_PSIZE(bp, psize);
1051 BP_SET_COMPRESS(bp, compress);
1052 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1053 BP_SET_TYPE(bp, zp->zp_type);
1054 BP_SET_LEVEL(bp, zp->zp_level);
1055 BP_SET_DEDUP(bp, zp->zp_dedup);
1056 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1057 if (zp->zp_dedup) {
1058 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1059 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1060 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1061 }
1062 }
1063
1064 return (ZIO_PIPELINE_CONTINUE);
1065 }
1066
1067 static int
1068 zio_free_bp_init(zio_t *zio)
1069 {
1070 blkptr_t *bp = zio->io_bp;
1071
1072 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1073 if (BP_GET_DEDUP(bp))
1074 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1075 }
1076
1077 return (ZIO_PIPELINE_CONTINUE);
1078 }
1079
1080 /*
1081 * ==========================================================================
1082 * Execute the I/O pipeline
1083 * ==========================================================================
1084 */
1085
1086 static void
1087 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1088 {
1089 spa_t *spa = zio->io_spa;
1090 zio_type_t t = zio->io_type;
1091 int flags = (cutinline ? TQ_FRONT : 0);
1092
1093 /*
1094 * If we're a config writer or a probe, the normal issue and
1095 * interrupt threads may all be blocked waiting for the config lock.
1096 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1097 */
1098 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1099 t = ZIO_TYPE_NULL;
1100
1101 /*
1102 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1103 */
1104 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1105 t = ZIO_TYPE_NULL;
1106
1107 /*
1108 * If this is a high priority I/O, then use the high priority taskq.
1109 */
1110 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1111 spa->spa_zio_taskq[t][q + 1] != NULL)
1112 q++;
1113
1114 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1115
1116 /*
1117 * NB: We are assuming that the zio can only be dispatched
1118 * to a single taskq at a time. It would be a grievous error
1119 * to dispatch the zio to another taskq at the same time.
1120 */
1121 ASSERT(zio->io_tqent.tqent_next == NULL);
1122 taskq_dispatch_ent(spa->spa_zio_taskq[t][q],
1123 (task_func_t *)zio_execute, zio, flags, &zio->io_tqent);
1124 }
1125
1126 static boolean_t
1127 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1128 {
1129 kthread_t *executor = zio->io_executor;
1130 spa_t *spa = zio->io_spa;
1131
1132 for (zio_type_t t = 0; t < ZIO_TYPES; t++)
1133 if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1134 return (B_TRUE);
1135
1136 return (B_FALSE);
1137 }
1138
1139 static int
1140 zio_issue_async(zio_t *zio)
1141 {
1142 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1143
1144 return (ZIO_PIPELINE_STOP);
1145 }
1146
1147 void
1148 zio_interrupt(zio_t *zio)
1149 {
1150 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1151 }
1152
1153 /*
1154 * Execute the I/O pipeline until one of the following occurs:
1155 * (1) the I/O completes; (2) the pipeline stalls waiting for
1156 * dependent child I/Os; (3) the I/O issues, so we're waiting
1157 * for an I/O completion interrupt; (4) the I/O is delegated by
1158 * vdev-level caching or aggregation; (5) the I/O is deferred
1159 * due to vdev-level queueing; (6) the I/O is handed off to
1160 * another thread. In all cases, the pipeline stops whenever
1161 * there's no CPU work; it never burns a thread in cv_wait().
1162 *
1163 * There's no locking on io_stage because there's no legitimate way
1164 * for multiple threads to be attempting to process the same I/O.
1165 */
1166 static zio_pipe_stage_t *zio_pipeline[];
1167
1168 void
1169 zio_execute(zio_t *zio)
1170 {
1171 zio->io_executor = curthread;
1172
1173 while (zio->io_stage < ZIO_STAGE_DONE) {
1174 enum zio_stage pipeline = zio->io_pipeline;
1175 enum zio_stage stage = zio->io_stage;
1176 int rv;
1177
1178 ASSERT(!MUTEX_HELD(&zio->io_lock));
1179 ASSERT(ISP2(stage));
1180 ASSERT(zio->io_stall == NULL);
1181
1182 do {
1183 stage <<= 1;
1184 } while ((stage & pipeline) == 0);
1185
1186 ASSERT(stage <= ZIO_STAGE_DONE);
1187
1188 /*
1189 * If we are in interrupt context and this pipeline stage
1190 * will grab a config lock that is held across I/O,
1191 * or may wait for an I/O that needs an interrupt thread
1192 * to complete, issue async to avoid deadlock.
1193 *
1194 * For VDEV_IO_START, we cut in line so that the io will
1195 * be sent to disk promptly.
1196 */
1197 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1198 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1199 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1200 zio_requeue_io_start_cut_in_line : B_FALSE;
1201 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1202 return;
1203 }
1204
1205 zio->io_stage = stage;
1206 rv = zio_pipeline[highbit(stage) - 1](zio);
1207
1208 if (rv == ZIO_PIPELINE_STOP)
1209 return;
1210
1211 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1212 }
1213 }
1214
1215 /*
1216 * ==========================================================================
1217 * Initiate I/O, either sync or async
1218 * ==========================================================================
1219 */
1220 int
1221 zio_wait(zio_t *zio)
1222 {
1223 int error;
1224
1225 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1226 ASSERT(zio->io_executor == NULL);
1227
1228 zio->io_waiter = curthread;
1229
1230 zio_execute(zio);
1231
1232 mutex_enter(&zio->io_lock);
1233 while (zio->io_executor != NULL)
1234 cv_wait(&zio->io_cv, &zio->io_lock);
1235 mutex_exit(&zio->io_lock);
1236
1237 error = zio->io_error;
1238 zio_destroy(zio);
1239
1240 return (error);
1241 }
1242
1243 void
1244 zio_nowait(zio_t *zio)
1245 {
1246 ASSERT(zio->io_executor == NULL);
1247
1248 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1249 zio_unique_parent(zio) == NULL) {
1250 /*
1251 * This is a logical async I/O with no parent to wait for it.
1252 * We add it to the spa_async_root_zio "Godfather" I/O which
1253 * will ensure they complete prior to unloading the pool.
1254 */
1255 spa_t *spa = zio->io_spa;
1256
1257 zio_add_child(spa->spa_async_zio_root, zio);
1258 }
1259
1260 zio_execute(zio);
1261 }
1262
1263 /*
1264 * ==========================================================================
1265 * Reexecute or suspend/resume failed I/O
1266 * ==========================================================================
1267 */
1268
1269 static void
1270 zio_reexecute(zio_t *pio)
1271 {
1272 zio_t *cio, *cio_next;
1273
1274 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1275 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1276 ASSERT(pio->io_gang_leader == NULL);
1277 ASSERT(pio->io_gang_tree == NULL);
1278
1279 pio->io_flags = pio->io_orig_flags;
1280 pio->io_stage = pio->io_orig_stage;
1281 pio->io_pipeline = pio->io_orig_pipeline;
1282 pio->io_reexecute = 0;
1283 pio->io_error = 0;
1284 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1285 pio->io_state[w] = 0;
1286 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1287 pio->io_child_error[c] = 0;
1288
1289 if (IO_IS_ALLOCATING(pio))
1290 BP_ZERO(pio->io_bp);
1291
1292 /*
1293 * As we reexecute pio's children, new children could be created.
1294 * New children go to the head of pio's io_child_list, however,
1295 * so we will (correctly) not reexecute them. The key is that
1296 * the remainder of pio's io_child_list, from 'cio_next' onward,
1297 * cannot be affected by any side effects of reexecuting 'cio'.
1298 */
1299 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1300 cio_next = zio_walk_children(pio);
1301 mutex_enter(&pio->io_lock);
1302 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1303 pio->io_children[cio->io_child_type][w]++;
1304 mutex_exit(&pio->io_lock);
1305 zio_reexecute(cio);
1306 }
1307
1308 /*
1309 * Now that all children have been reexecuted, execute the parent.
1310 * We don't reexecute "The Godfather" I/O here as it's the
1311 * responsibility of the caller to wait on him.
1312 */
1313 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1314 zio_execute(pio);
1315 }
1316
1317 void
1318 zio_suspend(spa_t *spa, zio_t *zio)
1319 {
1320 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1321 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1322 "failure and the failure mode property for this pool "
1323 "is set to panic.", spa_name(spa));
1324
1325 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1326
1327 mutex_enter(&spa->spa_suspend_lock);
1328
1329 if (spa->spa_suspend_zio_root == NULL)
1330 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1331 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1332 ZIO_FLAG_GODFATHER);
1333
1334 spa->spa_suspended = B_TRUE;
1335
1336 if (zio != NULL) {
1337 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1338 ASSERT(zio != spa->spa_suspend_zio_root);
1339 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1340 ASSERT(zio_unique_parent(zio) == NULL);
1341 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1342 zio_add_child(spa->spa_suspend_zio_root, zio);
1343 }
1344
1345 mutex_exit(&spa->spa_suspend_lock);
1346 }
1347
1348 int
1349 zio_resume(spa_t *spa)
1350 {
1351 zio_t *pio;
1352
1353 /*
1354 * Reexecute all previously suspended i/o.
1355 */
1356 mutex_enter(&spa->spa_suspend_lock);
1357 spa->spa_suspended = B_FALSE;
1358 cv_broadcast(&spa->spa_suspend_cv);
1359 pio = spa->spa_suspend_zio_root;
1360 spa->spa_suspend_zio_root = NULL;
1361 mutex_exit(&spa->spa_suspend_lock);
1362
1363 if (pio == NULL)
1364 return (0);
1365
1366 zio_reexecute(pio);
1367 return (zio_wait(pio));
1368 }
1369
1370 void
1371 zio_resume_wait(spa_t *spa)
1372 {
1373 mutex_enter(&spa->spa_suspend_lock);
1374 while (spa_suspended(spa))
1375 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1376 mutex_exit(&spa->spa_suspend_lock);
1377 }
1378
1379 /*
1380 * ==========================================================================
1381 * Gang blocks.
1382 *
1383 * A gang block is a collection of small blocks that looks to the DMU
1384 * like one large block. When zio_dva_allocate() cannot find a block
1385 * of the requested size, due to either severe fragmentation or the pool
1386 * being nearly full, it calls zio_write_gang_block() to construct the
1387 * block from smaller fragments.
1388 *
1389 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1390 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1391 * an indirect block: it's an array of block pointers. It consumes
1392 * only one sector and hence is allocatable regardless of fragmentation.
1393 * The gang header's bps point to its gang members, which hold the data.
1394 *
1395 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1396 * as the verifier to ensure uniqueness of the SHA256 checksum.
1397 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1398 * not the gang header. This ensures that data block signatures (needed for
1399 * deduplication) are independent of how the block is physically stored.
1400 *
1401 * Gang blocks can be nested: a gang member may itself be a gang block.
1402 * Thus every gang block is a tree in which root and all interior nodes are
1403 * gang headers, and the leaves are normal blocks that contain user data.
1404 * The root of the gang tree is called the gang leader.
1405 *
1406 * To perform any operation (read, rewrite, free, claim) on a gang block,
1407 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1408 * in the io_gang_tree field of the original logical i/o by recursively
1409 * reading the gang leader and all gang headers below it. This yields
1410 * an in-core tree containing the contents of every gang header and the
1411 * bps for every constituent of the gang block.
1412 *
1413 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1414 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1415 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1416 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1417 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1418 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1419 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1420 * of the gang header plus zio_checksum_compute() of the data to update the
1421 * gang header's blk_cksum as described above.
1422 *
1423 * The two-phase assemble/issue model solves the problem of partial failure --
1424 * what if you'd freed part of a gang block but then couldn't read the
1425 * gang header for another part? Assembling the entire gang tree first
1426 * ensures that all the necessary gang header I/O has succeeded before
1427 * starting the actual work of free, claim, or write. Once the gang tree
1428 * is assembled, free and claim are in-memory operations that cannot fail.
1429 *
1430 * In the event that a gang write fails, zio_dva_unallocate() walks the
1431 * gang tree to immediately free (i.e. insert back into the space map)
1432 * everything we've allocated. This ensures that we don't get ENOSPC
1433 * errors during repeated suspend/resume cycles due to a flaky device.
1434 *
1435 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1436 * the gang tree, we won't modify the block, so we can safely defer the free
1437 * (knowing that the block is still intact). If we *can* assemble the gang
1438 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1439 * each constituent bp and we can allocate a new block on the next sync pass.
1440 *
1441 * In all cases, the gang tree allows complete recovery from partial failure.
1442 * ==========================================================================
1443 */
1444
1445 static zio_t *
1446 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1447 {
1448 if (gn != NULL)
1449 return (pio);
1450
1451 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1452 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1453 &pio->io_bookmark));
1454 }
1455
1456 zio_t *
1457 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1458 {
1459 zio_t *zio;
1460
1461 if (gn != NULL) {
1462 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1463 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1464 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1465 /*
1466 * As we rewrite each gang header, the pipeline will compute
1467 * a new gang block header checksum for it; but no one will
1468 * compute a new data checksum, so we do that here. The one
1469 * exception is the gang leader: the pipeline already computed
1470 * its data checksum because that stage precedes gang assembly.
1471 * (Presently, nothing actually uses interior data checksums;
1472 * this is just good hygiene.)
1473 */
1474 if (gn != pio->io_gang_leader->io_gang_tree) {
1475 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1476 data, BP_GET_PSIZE(bp));
1477 }
1478 /*
1479 * If we are here to damage data for testing purposes,
1480 * leave the GBH alone so that we can detect the damage.
1481 */
1482 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1483 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1484 } else {
1485 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1486 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1487 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1488 }
1489
1490 return (zio);
1491 }
1492
1493 /* ARGSUSED */
1494 zio_t *
1495 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1496 {
1497 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1498 ZIO_GANG_CHILD_FLAGS(pio)));
1499 }
1500
1501 /* ARGSUSED */
1502 zio_t *
1503 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1504 {
1505 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1506 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1507 }
1508
1509 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1510 NULL,
1511 zio_read_gang,
1512 zio_rewrite_gang,
1513 zio_free_gang,
1514 zio_claim_gang,
1515 NULL
1516 };
1517
1518 static void zio_gang_tree_assemble_done(zio_t *zio);
1519
1520 static zio_gang_node_t *
1521 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1522 {
1523 zio_gang_node_t *gn;
1524
1525 ASSERT(*gnpp == NULL);
1526
1527 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1528 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1529 *gnpp = gn;
1530
1531 return (gn);
1532 }
1533
1534 static void
1535 zio_gang_node_free(zio_gang_node_t **gnpp)
1536 {
1537 zio_gang_node_t *gn = *gnpp;
1538
1539 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1540 ASSERT(gn->gn_child[g] == NULL);
1541
1542 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1543 kmem_free(gn, sizeof (*gn));
1544 *gnpp = NULL;
1545 }
1546
1547 static void
1548 zio_gang_tree_free(zio_gang_node_t **gnpp)
1549 {
1550 zio_gang_node_t *gn = *gnpp;
1551
1552 if (gn == NULL)
1553 return;
1554
1555 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1556 zio_gang_tree_free(&gn->gn_child[g]);
1557
1558 zio_gang_node_free(gnpp);
1559 }
1560
1561 static void
1562 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1563 {
1564 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1565
1566 ASSERT(gio->io_gang_leader == gio);
1567 ASSERT(BP_IS_GANG(bp));
1568
1569 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1570 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1571 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1572 }
1573
1574 static void
1575 zio_gang_tree_assemble_done(zio_t *zio)
1576 {
1577 zio_t *gio = zio->io_gang_leader;
1578 zio_gang_node_t *gn = zio->io_private;
1579 blkptr_t *bp = zio->io_bp;
1580
1581 ASSERT(gio == zio_unique_parent(zio));
1582 ASSERT(zio->io_child_count == 0);
1583
1584 if (zio->io_error)
1585 return;
1586
1587 if (BP_SHOULD_BYTESWAP(bp))
1588 byteswap_uint64_array(zio->io_data, zio->io_size);
1589
1590 ASSERT(zio->io_data == gn->gn_gbh);
1591 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1592 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1593
1594 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1595 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1596 if (!BP_IS_GANG(gbp))
1597 continue;
1598 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1599 }
1600 }
1601
1602 static void
1603 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1604 {
1605 zio_t *gio = pio->io_gang_leader;
1606 zio_t *zio;
1607
1608 ASSERT(BP_IS_GANG(bp) == !!gn);
1609 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1610 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1611
1612 /*
1613 * If you're a gang header, your data is in gn->gn_gbh.
1614 * If you're a gang member, your data is in 'data' and gn == NULL.
1615 */
1616 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1617
1618 if (gn != NULL) {
1619 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1620
1621 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1622 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1623 if (BP_IS_HOLE(gbp))
1624 continue;
1625 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1626 data = (char *)data + BP_GET_PSIZE(gbp);
1627 }
1628 }
1629
1630 if (gn == gio->io_gang_tree)
1631 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1632
1633 if (zio != pio)
1634 zio_nowait(zio);
1635 }
1636
1637 static int
1638 zio_gang_assemble(zio_t *zio)
1639 {
1640 blkptr_t *bp = zio->io_bp;
1641
1642 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1643 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1644
1645 zio->io_gang_leader = zio;
1646
1647 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1648
1649 return (ZIO_PIPELINE_CONTINUE);
1650 }
1651
1652 static int
1653 zio_gang_issue(zio_t *zio)
1654 {
1655 blkptr_t *bp = zio->io_bp;
1656
1657 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1658 return (ZIO_PIPELINE_STOP);
1659
1660 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1661 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1662
1663 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1664 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1665 else
1666 zio_gang_tree_free(&zio->io_gang_tree);
1667
1668 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1669
1670 return (ZIO_PIPELINE_CONTINUE);
1671 }
1672
1673 static void
1674 zio_write_gang_member_ready(zio_t *zio)
1675 {
1676 zio_t *pio = zio_unique_parent(zio);
1677 zio_t *gio = zio->io_gang_leader;
1678 dva_t *cdva = zio->io_bp->blk_dva;
1679 dva_t *pdva = pio->io_bp->blk_dva;
1680 uint64_t asize;
1681
1682 if (BP_IS_HOLE(zio->io_bp))
1683 return;
1684
1685 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1686
1687 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1688 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1689 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1690 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1691 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1692
1693 mutex_enter(&pio->io_lock);
1694 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1695 ASSERT(DVA_GET_GANG(&pdva[d]));
1696 asize = DVA_GET_ASIZE(&pdva[d]);
1697 asize += DVA_GET_ASIZE(&cdva[d]);
1698 DVA_SET_ASIZE(&pdva[d], asize);
1699 }
1700 mutex_exit(&pio->io_lock);
1701 }
1702
1703 static int
1704 zio_write_gang_block(zio_t *pio)
1705 {
1706 spa_t *spa = pio->io_spa;
1707 blkptr_t *bp = pio->io_bp;
1708 zio_t *gio = pio->io_gang_leader;
1709 zio_t *zio;
1710 zio_gang_node_t *gn, **gnpp;
1711 zio_gbh_phys_t *gbh;
1712 uint64_t txg = pio->io_txg;
1713 uint64_t resid = pio->io_size;
1714 uint64_t lsize;
1715 int copies = gio->io_prop.zp_copies;
1716 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1717 zio_prop_t zp;
1718 int error;
1719
1720 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1721 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1722 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1723 if (error) {
1724 pio->io_error = error;
1725 return (ZIO_PIPELINE_CONTINUE);
1726 }
1727
1728 if (pio == gio) {
1729 gnpp = &gio->io_gang_tree;
1730 } else {
1731 gnpp = pio->io_private;
1732 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1733 }
1734
1735 gn = zio_gang_node_alloc(gnpp);
1736 gbh = gn->gn_gbh;
1737 bzero(gbh, SPA_GANGBLOCKSIZE);
1738
1739 /*
1740 * Create the gang header.
1741 */
1742 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1743 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1744
1745 /*
1746 * Create and nowait the gang children.
1747 */
1748 for (int g = 0; resid != 0; resid -= lsize, g++) {
1749 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1750 SPA_MINBLOCKSIZE);
1751 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1752
1753 zp.zp_checksum = gio->io_prop.zp_checksum;
1754 zp.zp_compress = ZIO_COMPRESS_OFF;
1755 zp.zp_type = DMU_OT_NONE;
1756 zp.zp_level = 0;
1757 zp.zp_copies = gio->io_prop.zp_copies;
1758 zp.zp_dedup = 0;
1759 zp.zp_dedup_verify = 0;
1760
1761 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1762 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1763 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1764 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1765 &pio->io_bookmark));
1766 }
1767
1768 /*
1769 * Set pio's pipeline to just wait for zio to finish.
1770 */
1771 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1772
1773 zio_nowait(zio);
1774
1775 return (ZIO_PIPELINE_CONTINUE);
1776 }
1777
1778 /*
1779 * ==========================================================================
1780 * Dedup
1781 * ==========================================================================
1782 */
1783 static void
1784 zio_ddt_child_read_done(zio_t *zio)
1785 {
1786 blkptr_t *bp = zio->io_bp;
1787 ddt_entry_t *dde = zio->io_private;
1788 ddt_phys_t *ddp;
1789 zio_t *pio = zio_unique_parent(zio);
1790
1791 mutex_enter(&pio->io_lock);
1792 ddp = ddt_phys_select(dde, bp);
1793 if (zio->io_error == 0)
1794 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1795 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1796 dde->dde_repair_data = zio->io_data;
1797 else
1798 zio_buf_free(zio->io_data, zio->io_size);
1799 mutex_exit(&pio->io_lock);
1800 }
1801
1802 static int
1803 zio_ddt_read_start(zio_t *zio)
1804 {
1805 blkptr_t *bp = zio->io_bp;
1806
1807 ASSERT(BP_GET_DEDUP(bp));
1808 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1809 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1810
1811 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1812 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1813 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1814 ddt_phys_t *ddp = dde->dde_phys;
1815 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1816 blkptr_t blk;
1817
1818 ASSERT(zio->io_vsd == NULL);
1819 zio->io_vsd = dde;
1820
1821 if (ddp_self == NULL)
1822 return (ZIO_PIPELINE_CONTINUE);
1823
1824 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1825 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1826 continue;
1827 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1828 &blk);
1829 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1830 zio_buf_alloc(zio->io_size), zio->io_size,
1831 zio_ddt_child_read_done, dde, zio->io_priority,
1832 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1833 &zio->io_bookmark));
1834 }
1835 return (ZIO_PIPELINE_CONTINUE);
1836 }
1837
1838 zio_nowait(zio_read(zio, zio->io_spa, bp,
1839 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1840 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1841
1842 return (ZIO_PIPELINE_CONTINUE);
1843 }
1844
1845 static int
1846 zio_ddt_read_done(zio_t *zio)
1847 {
1848 blkptr_t *bp = zio->io_bp;
1849
1850 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1851 return (ZIO_PIPELINE_STOP);
1852
1853 ASSERT(BP_GET_DEDUP(bp));
1854 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1855 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1856
1857 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1858 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1859 ddt_entry_t *dde = zio->io_vsd;
1860 if (ddt == NULL) {
1861 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1862 return (ZIO_PIPELINE_CONTINUE);
1863 }
1864 if (dde == NULL) {
1865 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1866 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1867 return (ZIO_PIPELINE_STOP);
1868 }
1869 if (dde->dde_repair_data != NULL) {
1870 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1871 zio->io_child_error[ZIO_CHILD_DDT] = 0;
1872 }
1873 ddt_repair_done(ddt, dde);
1874 zio->io_vsd = NULL;
1875 }
1876
1877 ASSERT(zio->io_vsd == NULL);
1878
1879 return (ZIO_PIPELINE_CONTINUE);
1880 }
1881
1882 static boolean_t
1883 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1884 {
1885 spa_t *spa = zio->io_spa;
1886
1887 /*
1888 * Note: we compare the original data, not the transformed data,
1889 * because when zio->io_bp is an override bp, we will not have
1890 * pushed the I/O transforms. That's an important optimization
1891 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1892 */
1893 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1894 zio_t *lio = dde->dde_lead_zio[p];
1895
1896 if (lio != NULL) {
1897 return (lio->io_orig_size != zio->io_orig_size ||
1898 bcmp(zio->io_orig_data, lio->io_orig_data,
1899 zio->io_orig_size) != 0);
1900 }
1901 }
1902
1903 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1904 ddt_phys_t *ddp = &dde->dde_phys[p];
1905
1906 if (ddp->ddp_phys_birth != 0) {
1907 arc_buf_t *abuf = NULL;
1908 uint32_t aflags = ARC_WAIT;
1909 blkptr_t blk = *zio->io_bp;
1910 int error;
1911
1912 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
1913
1914 ddt_exit(ddt);
1915
1916 error = arc_read_nolock(NULL, spa, &blk,
1917 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
1918 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1919 &aflags, &zio->io_bookmark);
1920
1921 if (error == 0) {
1922 if (arc_buf_size(abuf) != zio->io_orig_size ||
1923 bcmp(abuf->b_data, zio->io_orig_data,
1924 zio->io_orig_size) != 0)
1925 error = EEXIST;
1926 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
1927 }
1928
1929 ddt_enter(ddt);
1930 return (error != 0);
1931 }
1932 }
1933
1934 return (B_FALSE);
1935 }
1936
1937 static void
1938 zio_ddt_child_write_ready(zio_t *zio)
1939 {
1940 int p = zio->io_prop.zp_copies;
1941 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1942 ddt_entry_t *dde = zio->io_private;
1943 ddt_phys_t *ddp = &dde->dde_phys[p];
1944 zio_t *pio;
1945
1946 if (zio->io_error)
1947 return;
1948
1949 ddt_enter(ddt);
1950
1951 ASSERT(dde->dde_lead_zio[p] == zio);
1952
1953 ddt_phys_fill(ddp, zio->io_bp);
1954
1955 while ((pio = zio_walk_parents(zio)) != NULL)
1956 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
1957
1958 ddt_exit(ddt);
1959 }
1960
1961 static void
1962 zio_ddt_child_write_done(zio_t *zio)
1963 {
1964 int p = zio->io_prop.zp_copies;
1965 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1966 ddt_entry_t *dde = zio->io_private;
1967 ddt_phys_t *ddp = &dde->dde_phys[p];
1968
1969 ddt_enter(ddt);
1970
1971 ASSERT(ddp->ddp_refcnt == 0);
1972 ASSERT(dde->dde_lead_zio[p] == zio);
1973 dde->dde_lead_zio[p] = NULL;
1974
1975 if (zio->io_error == 0) {
1976 while (zio_walk_parents(zio) != NULL)
1977 ddt_phys_addref(ddp);
1978 } else {
1979 ddt_phys_clear(ddp);
1980 }
1981
1982 ddt_exit(ddt);
1983 }
1984
1985 static void
1986 zio_ddt_ditto_write_done(zio_t *zio)
1987 {
1988 int p = DDT_PHYS_DITTO;
1989 zio_prop_t *zp = &zio->io_prop;
1990 blkptr_t *bp = zio->io_bp;
1991 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1992 ddt_entry_t *dde = zio->io_private;
1993 ddt_phys_t *ddp = &dde->dde_phys[p];
1994 ddt_key_t *ddk = &dde->dde_key;
1995
1996 ddt_enter(ddt);
1997
1998 ASSERT(ddp->ddp_refcnt == 0);
1999 ASSERT(dde->dde_lead_zio[p] == zio);
2000 dde->dde_lead_zio[p] = NULL;
2001
2002 if (zio->io_error == 0) {
2003 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2004 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2005 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2006 if (ddp->ddp_phys_birth != 0)
2007 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2008 ddt_phys_fill(ddp, bp);
2009 }
2010
2011 ddt_exit(ddt);
2012 }
2013
2014 static int
2015 zio_ddt_write(zio_t *zio)
2016 {
2017 spa_t *spa = zio->io_spa;
2018 blkptr_t *bp = zio->io_bp;
2019 uint64_t txg = zio->io_txg;
2020 zio_prop_t *zp = &zio->io_prop;
2021 int p = zp->zp_copies;
2022 int ditto_copies;
2023 zio_t *cio = NULL;
2024 zio_t *dio = NULL;
2025 ddt_t *ddt = ddt_select(spa, bp);
2026 ddt_entry_t *dde;
2027 ddt_phys_t *ddp;
2028
2029 ASSERT(BP_GET_DEDUP(bp));
2030 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2031 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2032
2033 ddt_enter(ddt);
2034 dde = ddt_lookup(ddt, bp, B_TRUE);
2035 ddp = &dde->dde_phys[p];
2036
2037 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2038 /*
2039 * If we're using a weak checksum, upgrade to a strong checksum
2040 * and try again. If we're already using a strong checksum,
2041 * we can't resolve it, so just convert to an ordinary write.
2042 * (And automatically e-mail a paper to Nature?)
2043 */
2044 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2045 zp->zp_checksum = spa_dedup_checksum(spa);
2046 zio_pop_transforms(zio);
2047 zio->io_stage = ZIO_STAGE_OPEN;
2048 BP_ZERO(bp);
2049 } else {
2050 zp->zp_dedup = 0;
2051 }
2052 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2053 ddt_exit(ddt);
2054 return (ZIO_PIPELINE_CONTINUE);
2055 }
2056
2057 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2058 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2059
2060 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2061 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2062 zio_prop_t czp = *zp;
2063
2064 czp.zp_copies = ditto_copies;
2065
2066 /*
2067 * If we arrived here with an override bp, we won't have run
2068 * the transform stack, so we won't have the data we need to
2069 * generate a child i/o. So, toss the override bp and restart.
2070 * This is safe, because using the override bp is just an
2071 * optimization; and it's rare, so the cost doesn't matter.
2072 */
2073 if (zio->io_bp_override) {
2074 zio_pop_transforms(zio);
2075 zio->io_stage = ZIO_STAGE_OPEN;
2076 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2077 zio->io_bp_override = NULL;
2078 BP_ZERO(bp);
2079 ddt_exit(ddt);
2080 return (ZIO_PIPELINE_CONTINUE);
2081 }
2082
2083 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2084 zio->io_orig_size, &czp, NULL,
2085 zio_ddt_ditto_write_done, dde, zio->io_priority,
2086 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2087
2088 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2089 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2090 }
2091
2092 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2093 if (ddp->ddp_phys_birth != 0)
2094 ddt_bp_fill(ddp, bp, txg);
2095 if (dde->dde_lead_zio[p] != NULL)
2096 zio_add_child(zio, dde->dde_lead_zio[p]);
2097 else
2098 ddt_phys_addref(ddp);
2099 } else if (zio->io_bp_override) {
2100 ASSERT(bp->blk_birth == txg);
2101 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2102 ddt_phys_fill(ddp, bp);
2103 ddt_phys_addref(ddp);
2104 } else {
2105 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2106 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2107 zio_ddt_child_write_done, dde, zio->io_priority,
2108 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2109
2110 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2111 dde->dde_lead_zio[p] = cio;
2112 }
2113
2114 ddt_exit(ddt);
2115
2116 if (cio)
2117 zio_nowait(cio);
2118 if (dio)
2119 zio_nowait(dio);
2120
2121 return (ZIO_PIPELINE_CONTINUE);
2122 }
2123
2124 ddt_entry_t *freedde; /* for debugging */
2125
2126 static int
2127 zio_ddt_free(zio_t *zio)
2128 {
2129 spa_t *spa = zio->io_spa;
2130 blkptr_t *bp = zio->io_bp;
2131 ddt_t *ddt = ddt_select(spa, bp);
2132 ddt_entry_t *dde;
2133 ddt_phys_t *ddp;
2134
2135 ASSERT(BP_GET_DEDUP(bp));
2136 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2137
2138 ddt_enter(ddt);
2139 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2140 ddp = ddt_phys_select(dde, bp);
2141 ddt_phys_decref(ddp);
2142 ddt_exit(ddt);
2143
2144 return (ZIO_PIPELINE_CONTINUE);
2145 }
2146
2147 /*
2148 * ==========================================================================
2149 * Allocate and free blocks
2150 * ==========================================================================
2151 */
2152 static int
2153 zio_dva_allocate(zio_t *zio)
2154 {
2155 spa_t *spa = zio->io_spa;
2156 metaslab_class_t *mc = spa_normal_class(spa);
2157 blkptr_t *bp = zio->io_bp;
2158 int error;
2159 int flags = 0;
2160
2161 if (zio->io_gang_leader == NULL) {
2162 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2163 zio->io_gang_leader = zio;
2164 }
2165
2166 ASSERT(BP_IS_HOLE(bp));
2167 ASSERT0(BP_GET_NDVAS(bp));
2168 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2169 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2170 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2171
2172 /*
2173 * The dump device does not support gang blocks so allocation on
2174 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2175 * the "fast" gang feature.
2176 */
2177 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2178 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2179 METASLAB_GANG_CHILD : 0;
2180 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2181 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2182
2183 if (error) {
2184 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2185 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2186 error);
2187 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2188 return (zio_write_gang_block(zio));
2189 zio->io_error = error;
2190 }
2191
2192 return (ZIO_PIPELINE_CONTINUE);
2193 }
2194
2195 static int
2196 zio_dva_free(zio_t *zio)
2197 {
2198 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2199
2200 return (ZIO_PIPELINE_CONTINUE);
2201 }
2202
2203 static int
2204 zio_dva_claim(zio_t *zio)
2205 {
2206 int error;
2207
2208 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2209 if (error)
2210 zio->io_error = error;
2211
2212 return (ZIO_PIPELINE_CONTINUE);
2213 }
2214
2215 /*
2216 * Undo an allocation. This is used by zio_done() when an I/O fails
2217 * and we want to give back the block we just allocated.
2218 * This handles both normal blocks and gang blocks.
2219 */
2220 static void
2221 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2222 {
2223 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2224 ASSERT(zio->io_bp_override == NULL);
2225
2226 if (!BP_IS_HOLE(bp))
2227 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2228
2229 if (gn != NULL) {
2230 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2231 zio_dva_unallocate(zio, gn->gn_child[g],
2232 &gn->gn_gbh->zg_blkptr[g]);
2233 }
2234 }
2235 }
2236
2237 /*
2238 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2239 */
2240 int
2241 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2242 uint64_t size, boolean_t use_slog)
2243 {
2244 int error = 1;
2245
2246 ASSERT(txg > spa_syncing_txg(spa));
2247
2248 /*
2249 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2250 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2251 * when allocating them.
2252 */
2253 if (use_slog) {
2254 error = metaslab_alloc(spa, spa_log_class(spa), size,
2255 new_bp, 1, txg, old_bp,
2256 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2257 }
2258
2259 if (error) {
2260 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2261 new_bp, 1, txg, old_bp,
2262 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2263 }
2264
2265 if (error == 0) {
2266 BP_SET_LSIZE(new_bp, size);
2267 BP_SET_PSIZE(new_bp, size);
2268 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2269 BP_SET_CHECKSUM(new_bp,
2270 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2271 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2272 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2273 BP_SET_LEVEL(new_bp, 0);
2274 BP_SET_DEDUP(new_bp, 0);
2275 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2276 }
2277
2278 return (error);
2279 }
2280
2281 /*
2282 * Free an intent log block.
2283 */
2284 void
2285 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2286 {
2287 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2288 ASSERT(!BP_IS_GANG(bp));
2289
2290 zio_free(spa, txg, bp);
2291 }
2292
2293 /*
2294 * ==========================================================================
2295 * Read and write to physical devices
2296 * ==========================================================================
2297 */
2298 static int
2299 zio_vdev_io_start(zio_t *zio)
2300 {
2301 vdev_t *vd = zio->io_vd;
2302 uint64_t align;
2303 spa_t *spa = zio->io_spa;
2304
2305 ASSERT(zio->io_error == 0);
2306 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2307
2308 if (vd == NULL) {
2309 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2310 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2311
2312 /*
2313 * The mirror_ops handle multiple DVAs in a single BP.
2314 */
2315 return (vdev_mirror_ops.vdev_op_io_start(zio));
2316 }
2317
2318 /*
2319 * We keep track of time-sensitive I/Os so that the scan thread
2320 * can quickly react to certain workloads. In particular, we care
2321 * about non-scrubbing, top-level reads and writes with the following
2322 * characteristics:
2323 * - synchronous writes of user data to non-slog devices
2324 * - any reads of user data
2325 * When these conditions are met, adjust the timestamp of spa_last_io
2326 * which allows the scan thread to adjust its workload accordingly.
2327 */
2328 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2329 vd == vd->vdev_top && !vd->vdev_islog &&
2330 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2331 zio->io_txg != spa_syncing_txg(spa)) {
2332 uint64_t old = spa->spa_last_io;
2333 uint64_t new = ddi_get_lbolt64();
2334 if (old != new)
2335 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2336 }
2337
2338 align = 1ULL << vd->vdev_top->vdev_ashift;
2339
2340 if (P2PHASE(zio->io_size, align) != 0) {
2341 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2342 char *abuf = zio_buf_alloc(asize);
2343 ASSERT(vd == vd->vdev_top);
2344 if (zio->io_type == ZIO_TYPE_WRITE) {
2345 bcopy(zio->io_data, abuf, zio->io_size);
2346 bzero(abuf + zio->io_size, asize - zio->io_size);
2347 }
2348 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2349 }
2350
2351 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2352 ASSERT(P2PHASE(zio->io_size, align) == 0);
2353 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2354
2355 /*
2356 * If this is a repair I/O, and there's no self-healing involved --
2357 * that is, we're just resilvering what we expect to resilver --
2358 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2359 * This prevents spurious resilvering with nested replication.
2360 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2361 * A is out of date, we'll read from C+D, then use the data to
2362 * resilver A+B -- but we don't actually want to resilver B, just A.
2363 * The top-level mirror has no way to know this, so instead we just
2364 * discard unnecessary repairs as we work our way down the vdev tree.
2365 * The same logic applies to any form of nested replication:
2366 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2367 */
2368 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2369 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2370 zio->io_txg != 0 && /* not a delegated i/o */
2371 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2372 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2373 zio_vdev_io_bypass(zio);
2374 return (ZIO_PIPELINE_CONTINUE);
2375 }
2376
2377 if (vd->vdev_ops->vdev_op_leaf &&
2378 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2379
2380 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2381 return (ZIO_PIPELINE_CONTINUE);
2382
2383 if ((zio = vdev_queue_io(zio)) == NULL)
2384 return (ZIO_PIPELINE_STOP);
2385
2386 if (!vdev_accessible(vd, zio)) {
2387 zio->io_error = ENXIO;
2388 zio_interrupt(zio);
2389 return (ZIO_PIPELINE_STOP);
2390 }
2391 }
2392
2393 return (vd->vdev_ops->vdev_op_io_start(zio));
2394 }
2395
2396 static int
2397 zio_vdev_io_done(zio_t *zio)
2398 {
2399 vdev_t *vd = zio->io_vd;
2400 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2401 boolean_t unexpected_error = B_FALSE;
2402
2403 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2404 return (ZIO_PIPELINE_STOP);
2405
2406 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2407
2408 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2409
2410 vdev_queue_io_done(zio);
2411
2412 if (zio->io_type == ZIO_TYPE_WRITE)
2413 vdev_cache_write(zio);
2414
2415 if (zio_injection_enabled && zio->io_error == 0)
2416 zio->io_error = zio_handle_device_injection(vd,
2417 zio, EIO);
2418
2419 if (zio_injection_enabled && zio->io_error == 0)
2420 zio->io_error = zio_handle_label_injection(zio, EIO);
2421
2422 if (zio->io_error) {
2423 if (!vdev_accessible(vd, zio)) {
2424 zio->io_error = ENXIO;
2425 } else {
2426 unexpected_error = B_TRUE;
2427 }
2428 }
2429 }
2430
2431 ops->vdev_op_io_done(zio);
2432
2433 if (unexpected_error)
2434 VERIFY(vdev_probe(vd, zio) == NULL);
2435
2436 return (ZIO_PIPELINE_CONTINUE);
2437 }
2438
2439 /*
2440 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2441 * disk, and use that to finish the checksum ereport later.
2442 */
2443 static void
2444 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2445 const void *good_buf)
2446 {
2447 /* no processing needed */
2448 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2449 }
2450
2451 /*ARGSUSED*/
2452 void
2453 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2454 {
2455 void *buf = zio_buf_alloc(zio->io_size);
2456
2457 bcopy(zio->io_data, buf, zio->io_size);
2458
2459 zcr->zcr_cbinfo = zio->io_size;
2460 zcr->zcr_cbdata = buf;
2461 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2462 zcr->zcr_free = zio_buf_free;
2463 }
2464
2465 static int
2466 zio_vdev_io_assess(zio_t *zio)
2467 {
2468 vdev_t *vd = zio->io_vd;
2469
2470 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2471 return (ZIO_PIPELINE_STOP);
2472
2473 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2474 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2475
2476 if (zio->io_vsd != NULL) {
2477 zio->io_vsd_ops->vsd_free(zio);
2478 zio->io_vsd = NULL;
2479 }
2480
2481 if (zio_injection_enabled && zio->io_error == 0)
2482 zio->io_error = zio_handle_fault_injection(zio, EIO);
2483
2484 /*
2485 * If the I/O failed, determine whether we should attempt to retry it.
2486 *
2487 * On retry, we cut in line in the issue queue, since we don't want
2488 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2489 */
2490 if (zio->io_error && vd == NULL &&
2491 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2492 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2493 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2494 zio->io_error = 0;
2495 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2496 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2497 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2498 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2499 zio_requeue_io_start_cut_in_line);
2500 return (ZIO_PIPELINE_STOP);
2501 }
2502
2503 /*
2504 * If we got an error on a leaf device, convert it to ENXIO
2505 * if the device is not accessible at all.
2506 */
2507 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2508 !vdev_accessible(vd, zio))
2509 zio->io_error = ENXIO;
2510
2511 /*
2512 * If we can't write to an interior vdev (mirror or RAID-Z),
2513 * set vdev_cant_write so that we stop trying to allocate from it.
2514 */
2515 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2516 vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2517 vd->vdev_cant_write = B_TRUE;
2518
2519 if (zio->io_error)
2520 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2521
2522 return (ZIO_PIPELINE_CONTINUE);
2523 }
2524
2525 void
2526 zio_vdev_io_reissue(zio_t *zio)
2527 {
2528 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2529 ASSERT(zio->io_error == 0);
2530
2531 zio->io_stage >>= 1;
2532 }
2533
2534 void
2535 zio_vdev_io_redone(zio_t *zio)
2536 {
2537 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2538
2539 zio->io_stage >>= 1;
2540 }
2541
2542 void
2543 zio_vdev_io_bypass(zio_t *zio)
2544 {
2545 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2546 ASSERT(zio->io_error == 0);
2547
2548 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2549 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2550 }
2551
2552 /*
2553 * ==========================================================================
2554 * Generate and verify checksums
2555 * ==========================================================================
2556 */
2557 static int
2558 zio_checksum_generate(zio_t *zio)
2559 {
2560 blkptr_t *bp = zio->io_bp;
2561 enum zio_checksum checksum;
2562
2563 if (bp == NULL) {
2564 /*
2565 * This is zio_write_phys().
2566 * We're either generating a label checksum, or none at all.
2567 */
2568 checksum = zio->io_prop.zp_checksum;
2569
2570 if (checksum == ZIO_CHECKSUM_OFF)
2571 return (ZIO_PIPELINE_CONTINUE);
2572
2573 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2574 } else {
2575 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2576 ASSERT(!IO_IS_ALLOCATING(zio));
2577 checksum = ZIO_CHECKSUM_GANG_HEADER;
2578 } else {
2579 checksum = BP_GET_CHECKSUM(bp);
2580 }
2581 }
2582
2583 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2584
2585 return (ZIO_PIPELINE_CONTINUE);
2586 }
2587
2588 static int
2589 zio_checksum_verify(zio_t *zio)
2590 {
2591 zio_bad_cksum_t info;
2592 blkptr_t *bp = zio->io_bp;
2593 int error;
2594
2595 ASSERT(zio->io_vd != NULL);
2596
2597 if (bp == NULL) {
2598 /*
2599 * This is zio_read_phys().
2600 * We're either verifying a label checksum, or nothing at all.
2601 */
2602 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2603 return (ZIO_PIPELINE_CONTINUE);
2604
2605 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2606 }
2607
2608 if ((error = zio_checksum_error(zio, &info)) != 0) {
2609 zio->io_error = error;
2610 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2611 zfs_ereport_start_checksum(zio->io_spa,
2612 zio->io_vd, zio, zio->io_offset,
2613 zio->io_size, NULL, &info);
2614 }
2615 }
2616
2617 return (ZIO_PIPELINE_CONTINUE);
2618 }
2619
2620 /*
2621 * Called by RAID-Z to ensure we don't compute the checksum twice.
2622 */
2623 void
2624 zio_checksum_verified(zio_t *zio)
2625 {
2626 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2627 }
2628
2629 /*
2630 * ==========================================================================
2631 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2632 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2633 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2634 * indicate errors that are specific to one I/O, and most likely permanent.
2635 * Any other error is presumed to be worse because we weren't expecting it.
2636 * ==========================================================================
2637 */
2638 int
2639 zio_worst_error(int e1, int e2)
2640 {
2641 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2642 int r1, r2;
2643
2644 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2645 if (e1 == zio_error_rank[r1])
2646 break;
2647
2648 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2649 if (e2 == zio_error_rank[r2])
2650 break;
2651
2652 return (r1 > r2 ? e1 : e2);
2653 }
2654
2655 /*
2656 * ==========================================================================
2657 * I/O completion
2658 * ==========================================================================
2659 */
2660 static int
2661 zio_ready(zio_t *zio)
2662 {
2663 blkptr_t *bp = zio->io_bp;
2664 zio_t *pio, *pio_next;
2665
2666 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2667 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2668 return (ZIO_PIPELINE_STOP);
2669
2670 if (zio->io_ready) {
2671 ASSERT(IO_IS_ALLOCATING(zio));
2672 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2673 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2674
2675 zio->io_ready(zio);
2676 }
2677
2678 if (bp != NULL && bp != &zio->io_bp_copy)
2679 zio->io_bp_copy = *bp;
2680
2681 if (zio->io_error)
2682 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2683
2684 mutex_enter(&zio->io_lock);
2685 zio->io_state[ZIO_WAIT_READY] = 1;
2686 pio = zio_walk_parents(zio);
2687 mutex_exit(&zio->io_lock);
2688
2689 /*
2690 * As we notify zio's parents, new parents could be added.
2691 * New parents go to the head of zio's io_parent_list, however,
2692 * so we will (correctly) not notify them. The remainder of zio's
2693 * io_parent_list, from 'pio_next' onward, cannot change because
2694 * all parents must wait for us to be done before they can be done.
2695 */
2696 for (; pio != NULL; pio = pio_next) {
2697 pio_next = zio_walk_parents(zio);
2698 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2699 }
2700
2701 if (zio->io_flags & ZIO_FLAG_NODATA) {
2702 if (BP_IS_GANG(bp)) {
2703 zio->io_flags &= ~ZIO_FLAG_NODATA;
2704 } else {
2705 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2706 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2707 }
2708 }
2709
2710 if (zio_injection_enabled &&
2711 zio->io_spa->spa_syncing_txg == zio->io_txg)
2712 zio_handle_ignored_writes(zio);
2713
2714 return (ZIO_PIPELINE_CONTINUE);
2715 }
2716
2717 static int
2718 zio_done(zio_t *zio)
2719 {
2720 spa_t *spa = zio->io_spa;
2721 zio_t *lio = zio->io_logical;
2722 blkptr_t *bp = zio->io_bp;
2723 vdev_t *vd = zio->io_vd;
2724 uint64_t psize = zio->io_size;
2725 zio_t *pio, *pio_next;
2726
2727 /*
2728 * If our children haven't all completed,
2729 * wait for them and then repeat this pipeline stage.
2730 */
2731 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2732 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2733 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2734 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2735 return (ZIO_PIPELINE_STOP);
2736
2737 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2738 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2739 ASSERT(zio->io_children[c][w] == 0);
2740
2741 if (bp != NULL) {
2742 ASSERT(bp->blk_pad[0] == 0);
2743 ASSERT(bp->blk_pad[1] == 0);
2744 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2745 (bp == zio_unique_parent(zio)->io_bp));
2746 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2747 zio->io_bp_override == NULL &&
2748 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2749 ASSERT(!BP_SHOULD_BYTESWAP(bp));
2750 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2751 ASSERT(BP_COUNT_GANG(bp) == 0 ||
2752 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2753 }
2754 }
2755
2756 /*
2757 * If there were child vdev/gang/ddt errors, they apply to us now.
2758 */
2759 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2760 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2761 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2762
2763 /*
2764 * If the I/O on the transformed data was successful, generate any
2765 * checksum reports now while we still have the transformed data.
2766 */
2767 if (zio->io_error == 0) {
2768 while (zio->io_cksum_report != NULL) {
2769 zio_cksum_report_t *zcr = zio->io_cksum_report;
2770 uint64_t align = zcr->zcr_align;
2771 uint64_t asize = P2ROUNDUP(psize, align);
2772 char *abuf = zio->io_data;
2773
2774 if (asize != psize) {
2775 abuf = zio_buf_alloc(asize);
2776 bcopy(zio->io_data, abuf, psize);
2777 bzero(abuf + psize, asize - psize);
2778 }
2779
2780 zio->io_cksum_report = zcr->zcr_next;
2781 zcr->zcr_next = NULL;
2782 zcr->zcr_finish(zcr, abuf);
2783 zfs_ereport_free_checksum(zcr);
2784
2785 if (asize != psize)
2786 zio_buf_free(abuf, asize);
2787 }
2788 }
2789
2790 zio_pop_transforms(zio); /* note: may set zio->io_error */
2791
2792 vdev_stat_update(zio, psize);
2793
2794 if (zio->io_error) {
2795 /*
2796 * If this I/O is attached to a particular vdev,
2797 * generate an error message describing the I/O failure
2798 * at the block level. We ignore these errors if the
2799 * device is currently unavailable.
2800 */
2801 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
2802 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
2803
2804 if ((zio->io_error == EIO || !(zio->io_flags &
2805 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2806 zio == lio) {
2807 /*
2808 * For logical I/O requests, tell the SPA to log the
2809 * error and generate a logical data ereport.
2810 */
2811 spa_log_error(spa, zio);
2812 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
2813 0, 0);
2814 }
2815 }
2816
2817 if (zio->io_error && zio == lio) {
2818 /*
2819 * Determine whether zio should be reexecuted. This will
2820 * propagate all the way to the root via zio_notify_parent().
2821 */
2822 ASSERT(vd == NULL && bp != NULL);
2823 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2824
2825 if (IO_IS_ALLOCATING(zio) &&
2826 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2827 if (zio->io_error != ENOSPC)
2828 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2829 else
2830 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2831 }
2832
2833 if ((zio->io_type == ZIO_TYPE_READ ||
2834 zio->io_type == ZIO_TYPE_FREE) &&
2835 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2836 zio->io_error == ENXIO &&
2837 spa_load_state(spa) == SPA_LOAD_NONE &&
2838 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
2839 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2840
2841 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2842 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2843
2844 /*
2845 * Here is a possibly good place to attempt to do
2846 * either combinatorial reconstruction or error correction
2847 * based on checksums. It also might be a good place
2848 * to send out preliminary ereports before we suspend
2849 * processing.
2850 */
2851 }
2852
2853 /*
2854 * If there were logical child errors, they apply to us now.
2855 * We defer this until now to avoid conflating logical child
2856 * errors with errors that happened to the zio itself when
2857 * updating vdev stats and reporting FMA events above.
2858 */
2859 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2860
2861 if ((zio->io_error || zio->io_reexecute) &&
2862 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2863 !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2864 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
2865
2866 zio_gang_tree_free(&zio->io_gang_tree);
2867
2868 /*
2869 * Godfather I/Os should never suspend.
2870 */
2871 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2872 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2873 zio->io_reexecute = 0;
2874
2875 if (zio->io_reexecute) {
2876 /*
2877 * This is a logical I/O that wants to reexecute.
2878 *
2879 * Reexecute is top-down. When an i/o fails, if it's not
2880 * the root, it simply notifies its parent and sticks around.
2881 * The parent, seeing that it still has children in zio_done(),
2882 * does the same. This percolates all the way up to the root.
2883 * The root i/o will reexecute or suspend the entire tree.
2884 *
2885 * This approach ensures that zio_reexecute() honors
2886 * all the original i/o dependency relationships, e.g.
2887 * parents not executing until children are ready.
2888 */
2889 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2890
2891 zio->io_gang_leader = NULL;
2892
2893 mutex_enter(&zio->io_lock);
2894 zio->io_state[ZIO_WAIT_DONE] = 1;
2895 mutex_exit(&zio->io_lock);
2896
2897 /*
2898 * "The Godfather" I/O monitors its children but is
2899 * not a true parent to them. It will track them through
2900 * the pipeline but severs its ties whenever they get into
2901 * trouble (e.g. suspended). This allows "The Godfather"
2902 * I/O to return status without blocking.
2903 */
2904 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2905 zio_link_t *zl = zio->io_walk_link;
2906 pio_next = zio_walk_parents(zio);
2907
2908 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
2909 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
2910 zio_remove_child(pio, zio, zl);
2911 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2912 }
2913 }
2914
2915 if ((pio = zio_unique_parent(zio)) != NULL) {
2916 /*
2917 * We're not a root i/o, so there's nothing to do
2918 * but notify our parent. Don't propagate errors
2919 * upward since we haven't permanently failed yet.
2920 */
2921 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2922 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
2923 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2924 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
2925 /*
2926 * We'd fail again if we reexecuted now, so suspend
2927 * until conditions improve (e.g. device comes online).
2928 */
2929 zio_suspend(spa, zio);
2930 } else {
2931 /*
2932 * Reexecution is potentially a huge amount of work.
2933 * Hand it off to the otherwise-unused claim taskq.
2934 */
2935 ASSERT(zio->io_tqent.tqent_next == NULL);
2936 taskq_dispatch_ent(
2937 spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2938 (task_func_t *)zio_reexecute, zio, 0,
2939 &zio->io_tqent);
2940 }
2941 return (ZIO_PIPELINE_STOP);
2942 }
2943
2944 ASSERT(zio->io_child_count == 0);
2945 ASSERT(zio->io_reexecute == 0);
2946 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
2947
2948 /*
2949 * Report any checksum errors, since the I/O is complete.
2950 */
2951 while (zio->io_cksum_report != NULL) {
2952 zio_cksum_report_t *zcr = zio->io_cksum_report;
2953 zio->io_cksum_report = zcr->zcr_next;
2954 zcr->zcr_next = NULL;
2955 zcr->zcr_finish(zcr, NULL);
2956 zfs_ereport_free_checksum(zcr);
2957 }
2958
2959 /*
2960 * It is the responsibility of the done callback to ensure that this
2961 * particular zio is no longer discoverable for adoption, and as
2962 * such, cannot acquire any new parents.
2963 */
2964 if (zio->io_done)
2965 zio->io_done(zio);
2966
2967 mutex_enter(&zio->io_lock);
2968 zio->io_state[ZIO_WAIT_DONE] = 1;
2969 mutex_exit(&zio->io_lock);
2970
2971 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2972 zio_link_t *zl = zio->io_walk_link;
2973 pio_next = zio_walk_parents(zio);
2974 zio_remove_child(pio, zio, zl);
2975 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2976 }
2977
2978 if (zio->io_waiter != NULL) {
2979 mutex_enter(&zio->io_lock);
2980 zio->io_executor = NULL;
2981 cv_broadcast(&zio->io_cv);
2982 mutex_exit(&zio->io_lock);
2983 } else {
2984 zio_destroy(zio);
2985 }
2986
2987 return (ZIO_PIPELINE_STOP);
2988 }
2989
2990 /*
2991 * ==========================================================================
2992 * I/O pipeline definition
2993 * ==========================================================================
2994 */
2995 static zio_pipe_stage_t *zio_pipeline[] = {
2996 NULL,
2997 zio_read_bp_init,
2998 zio_free_bp_init,
2999 zio_issue_async,
3000 zio_write_bp_init,
3001 zio_checksum_generate,
3002 zio_ddt_read_start,
3003 zio_ddt_read_done,
3004 zio_ddt_write,
3005 zio_ddt_free,
3006 zio_gang_assemble,
3007 zio_gang_issue,
3008 zio_dva_allocate,
3009 zio_dva_free,
3010 zio_dva_claim,
3011 zio_ready,
3012 zio_vdev_io_start,
3013 zio_vdev_io_done,
3014 zio_vdev_io_assess,
3015 zio_checksum_verify,
3016 zio_done
3017 };
3018
3019 /* dnp is the dnode for zb1->zb_object */
3020 boolean_t
3021 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_t *zb1,
3022 const zbookmark_t *zb2)
3023 {
3024 uint64_t zb1nextL0, zb2thisobj;
3025
3026 ASSERT(zb1->zb_objset == zb2->zb_objset);
3027 ASSERT(zb2->zb_level == 0);
3028
3029 /*
3030 * A bookmark in the deadlist is considered to be after
3031 * everything else.
3032 */
3033 if (zb2->zb_object == DMU_DEADLIST_OBJECT)
3034 return (B_TRUE);
3035
3036 /* The objset_phys_t isn't before anything. */
3037 if (dnp == NULL)
3038 return (B_FALSE);
3039
3040 zb1nextL0 = (zb1->zb_blkid + 1) <<
3041 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3042
3043 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3044 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3045
3046 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3047 uint64_t nextobj = zb1nextL0 *
3048 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3049 return (nextobj <= zb2thisobj);
3050 }
3051
3052 if (zb1->zb_object < zb2thisobj)
3053 return (B_TRUE);
3054 if (zb1->zb_object > zb2thisobj)
3055 return (B_FALSE);
3056 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3057 return (B_FALSE);
3058 return (zb1nextL0 <= zb2->zb_blkid);
3059 }
Unleashed OS