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) 2013 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 = SET_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
653 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
654 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
655 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
656 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
657
658 zio->io_ready = ready;
659 zio->io_prop = *zp;
660
661 return (zio);
662 }
663
664 zio_t *
665 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
666 uint64_t size, zio_done_func_t *done, void *private, int priority,
667 enum zio_flag flags, zbookmark_t *zb)
668 {
669 zio_t *zio;
670
671 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
672 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
673 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
674
675 return (zio);
676 }
677
678 void
679 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
680 {
681 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
682 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
683 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
684 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
685
686 /*
687 * We must reset the io_prop to match the values that existed
688 * when the bp was first written by dmu_sync() keeping in mind
689 * that nopwrite and dedup are mutually exclusive.
690 */
691 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
692 zio->io_prop.zp_nopwrite = nopwrite;
693 zio->io_prop.zp_copies = copies;
694 zio->io_bp_override = bp;
695 }
696
697 void
698 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
699 {
700 metaslab_check_free(spa, bp);
701 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
702 }
703
704 zio_t *
705 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
706 enum zio_flag flags)
707 {
708 zio_t *zio;
709
710 dprintf_bp(bp, "freeing in txg %llu, pass %u",
711 (longlong_t)txg, spa->spa_sync_pass);
712
713 ASSERT(!BP_IS_HOLE(bp));
714 ASSERT(spa_syncing_txg(spa) == txg);
715 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
716
717 metaslab_check_free(spa, bp);
718
719 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
720 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
721 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
722
723 return (zio);
724 }
725
726 zio_t *
727 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
728 zio_done_func_t *done, void *private, enum zio_flag flags)
729 {
730 zio_t *zio;
731
732 /*
733 * A claim is an allocation of a specific block. Claims are needed
734 * to support immediate writes in the intent log. The issue is that
735 * immediate writes contain committed data, but in a txg that was
736 * *not* committed. Upon opening the pool after an unclean shutdown,
737 * the intent log claims all blocks that contain immediate write data
738 * so that the SPA knows they're in use.
739 *
740 * All claims *must* be resolved in the first txg -- before the SPA
741 * starts allocating blocks -- so that nothing is allocated twice.
742 * If txg == 0 we just verify that the block is claimable.
743 */
744 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
745 ASSERT(txg == spa_first_txg(spa) || txg == 0);
746 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
747
748 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
749 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
750 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
751
752 return (zio);
753 }
754
755 zio_t *
756 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
757 zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
758 {
759 zio_t *zio;
760 int c;
761
762 if (vd->vdev_children == 0) {
763 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
764 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
765 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
766
767 zio->io_cmd = cmd;
768 } else {
769 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
770
771 for (c = 0; c < vd->vdev_children; c++)
772 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
773 done, private, priority, flags));
774 }
775
776 return (zio);
777 }
778
779 zio_t *
780 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
781 void *data, int checksum, zio_done_func_t *done, void *private,
782 int priority, enum zio_flag flags, boolean_t labels)
783 {
784 zio_t *zio;
785
786 ASSERT(vd->vdev_children == 0);
787 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
788 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
789 ASSERT3U(offset + size, <=, vd->vdev_psize);
790
791 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
792 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
793 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
794
795 zio->io_prop.zp_checksum = checksum;
796
797 return (zio);
798 }
799
800 zio_t *
801 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
802 void *data, int checksum, zio_done_func_t *done, void *private,
803 int priority, enum zio_flag flags, boolean_t labels)
804 {
805 zio_t *zio;
806
807 ASSERT(vd->vdev_children == 0);
808 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
809 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
810 ASSERT3U(offset + size, <=, vd->vdev_psize);
811
812 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
813 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
814 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
815
816 zio->io_prop.zp_checksum = checksum;
817
818 if (zio_checksum_table[checksum].ci_eck) {
819 /*
820 * zec checksums are necessarily destructive -- they modify
821 * the end of the write buffer to hold the verifier/checksum.
822 * Therefore, we must make a local copy in case the data is
823 * being written to multiple places in parallel.
824 */
825 void *wbuf = zio_buf_alloc(size);
826 bcopy(data, wbuf, size);
827 zio_push_transform(zio, wbuf, size, size, NULL);
828 }
829
830 return (zio);
831 }
832
833 /*
834 * Create a child I/O to do some work for us.
835 */
836 zio_t *
837 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
838 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
839 zio_done_func_t *done, void *private)
840 {
841 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
842 zio_t *zio;
843
844 ASSERT(vd->vdev_parent ==
845 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
846
847 if (type == ZIO_TYPE_READ && bp != NULL) {
848 /*
849 * If we have the bp, then the child should perform the
850 * checksum and the parent need not. This pushes error
851 * detection as close to the leaves as possible and
852 * eliminates redundant checksums in the interior nodes.
853 */
854 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
855 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
856 }
857
858 if (vd->vdev_children == 0)
859 offset += VDEV_LABEL_START_SIZE;
860
861 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
862
863 /*
864 * If we've decided to do a repair, the write is not speculative --
865 * even if the original read was.
866 */
867 if (flags & ZIO_FLAG_IO_REPAIR)
868 flags &= ~ZIO_FLAG_SPECULATIVE;
869
870 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
871 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
872 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
873
874 return (zio);
875 }
876
877 zio_t *
878 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
879 int type, int priority, enum zio_flag flags,
880 zio_done_func_t *done, void *private)
881 {
882 zio_t *zio;
883
884 ASSERT(vd->vdev_ops->vdev_op_leaf);
885
886 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
887 data, size, done, private, type, priority,
888 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
889 vd, offset, NULL,
890 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
891
892 return (zio);
893 }
894
895 void
896 zio_flush(zio_t *zio, vdev_t *vd)
897 {
898 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
899 NULL, NULL, ZIO_PRIORITY_NOW,
900 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
901 }
902
903 void
904 zio_shrink(zio_t *zio, uint64_t size)
905 {
906 ASSERT(zio->io_executor == NULL);
907 ASSERT(zio->io_orig_size == zio->io_size);
908 ASSERT(size <= zio->io_size);
909
910 /*
911 * We don't shrink for raidz because of problems with the
912 * reconstruction when reading back less than the block size.
913 * Note, BP_IS_RAIDZ() assumes no compression.
914 */
915 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
916 if (!BP_IS_RAIDZ(zio->io_bp))
917 zio->io_orig_size = zio->io_size = size;
918 }
919
920 /*
921 * ==========================================================================
922 * Prepare to read and write logical blocks
923 * ==========================================================================
924 */
925
926 static int
927 zio_read_bp_init(zio_t *zio)
928 {
929 blkptr_t *bp = zio->io_bp;
930
931 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
932 zio->io_child_type == ZIO_CHILD_LOGICAL &&
933 !(zio->io_flags & ZIO_FLAG_RAW)) {
934 uint64_t psize = BP_GET_PSIZE(bp);
935 void *cbuf = zio_buf_alloc(psize);
936
937 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
938 }
939
940 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
941 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
942
943 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
944 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
945
946 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
947 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
948
949 return (ZIO_PIPELINE_CONTINUE);
950 }
951
952 static int
953 zio_write_bp_init(zio_t *zio)
954 {
955 spa_t *spa = zio->io_spa;
956 zio_prop_t *zp = &zio->io_prop;
957 enum zio_compress compress = zp->zp_compress;
958 blkptr_t *bp = zio->io_bp;
959 uint64_t lsize = zio->io_size;
960 uint64_t psize = lsize;
961 int pass = 1;
962
963 /*
964 * If our children haven't all reached the ready stage,
965 * wait for them and then repeat this pipeline stage.
966 */
967 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
968 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
969 return (ZIO_PIPELINE_STOP);
970
971 if (!IO_IS_ALLOCATING(zio))
972 return (ZIO_PIPELINE_CONTINUE);
973
974 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
975
976 if (zio->io_bp_override) {
977 ASSERT(bp->blk_birth != zio->io_txg);
978 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
979
980 *bp = *zio->io_bp_override;
981 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
982
983 /*
984 * If we've been overridden and nopwrite is set then
985 * set the flag accordingly to indicate that a nopwrite
986 * has already occurred.
987 */
988 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
989 ASSERT(!zp->zp_dedup);
990 zio->io_flags |= ZIO_FLAG_NOPWRITE;
991 return (ZIO_PIPELINE_CONTINUE);
992 }
993
994 ASSERT(!zp->zp_nopwrite);
995
996 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
997 return (ZIO_PIPELINE_CONTINUE);
998
999 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1000 zp->zp_dedup_verify);
1001
1002 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1003 BP_SET_DEDUP(bp, 1);
1004 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1005 return (ZIO_PIPELINE_CONTINUE);
1006 }
1007 zio->io_bp_override = NULL;
1008 BP_ZERO(bp);
1009 }
1010
1011 if (bp->blk_birth == zio->io_txg) {
1012 /*
1013 * We're rewriting an existing block, which means we're
1014 * working on behalf of spa_sync(). For spa_sync() to
1015 * converge, it must eventually be the case that we don't
1016 * have to allocate new blocks. But compression changes
1017 * the blocksize, which forces a reallocate, and makes
1018 * convergence take longer. Therefore, after the first
1019 * few passes, stop compressing to ensure convergence.
1020 */
1021 pass = spa_sync_pass(spa);
1022
1023 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1024 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1025 ASSERT(!BP_GET_DEDUP(bp));
1026
1027 if (pass >= zfs_sync_pass_dont_compress)
1028 compress = ZIO_COMPRESS_OFF;
1029
1030 /* Make sure someone doesn't change their mind on overwrites */
1031 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
1032 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1033 }
1034
1035 if (compress != ZIO_COMPRESS_OFF) {
1036 void *cbuf = zio_buf_alloc(lsize);
1037 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1038 if (psize == 0 || psize == lsize) {
1039 compress = ZIO_COMPRESS_OFF;
1040 zio_buf_free(cbuf, lsize);
1041 } else {
1042 ASSERT(psize < lsize);
1043 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1044 }
1045 }
1046
1047 /*
1048 * The final pass of spa_sync() must be all rewrites, but the first
1049 * few passes offer a trade-off: allocating blocks defers convergence,
1050 * but newly allocated blocks are sequential, so they can be written
1051 * to disk faster. Therefore, we allow the first few passes of
1052 * spa_sync() to allocate new blocks, but force rewrites after that.
1053 * There should only be a handful of blocks after pass 1 in any case.
1054 */
1055 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1056 pass >= zfs_sync_pass_rewrite) {
1057 ASSERT(psize != 0);
1058 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1059 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1060 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1061 } else {
1062 BP_ZERO(bp);
1063 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1064 }
1065
1066 if (psize == 0) {
1067 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1068 } else {
1069 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1070 BP_SET_LSIZE(bp, lsize);
1071 BP_SET_PSIZE(bp, psize);
1072 BP_SET_COMPRESS(bp, compress);
1073 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1074 BP_SET_TYPE(bp, zp->zp_type);
1075 BP_SET_LEVEL(bp, zp->zp_level);
1076 BP_SET_DEDUP(bp, zp->zp_dedup);
1077 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1078 if (zp->zp_dedup) {
1079 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1080 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1081 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1082 }
1083 if (zp->zp_nopwrite) {
1084 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1085 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1086 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1087 }
1088 }
1089
1090 return (ZIO_PIPELINE_CONTINUE);
1091 }
1092
1093 static int
1094 zio_free_bp_init(zio_t *zio)
1095 {
1096 blkptr_t *bp = zio->io_bp;
1097
1098 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1099 if (BP_GET_DEDUP(bp))
1100 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1101 }
1102
1103 return (ZIO_PIPELINE_CONTINUE);
1104 }
1105
1106 /*
1107 * ==========================================================================
1108 * Execute the I/O pipeline
1109 * ==========================================================================
1110 */
1111
1112 static void
1113 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1114 {
1115 spa_t *spa = zio->io_spa;
1116 zio_type_t t = zio->io_type;
1117 int flags = (cutinline ? TQ_FRONT : 0);
1118
1119 /*
1120 * If we're a config writer or a probe, the normal issue and
1121 * interrupt threads may all be blocked waiting for the config lock.
1122 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1123 */
1124 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1125 t = ZIO_TYPE_NULL;
1126
1127 /*
1128 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1129 */
1130 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1131 t = ZIO_TYPE_NULL;
1132
1133 /*
1134 * If this is a high priority I/O, then use the high priority taskq if
1135 * available.
1136 */
1137 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1138 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1139 q++;
1140
1141 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1142
1143 /*
1144 * NB: We are assuming that the zio can only be dispatched
1145 * to a single taskq at a time. It would be a grievous error
1146 * to dispatch the zio to another taskq at the same time.
1147 */
1148 ASSERT(zio->io_tqent.tqent_next == NULL);
1149 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1150 flags, &zio->io_tqent);
1151 }
1152
1153 static boolean_t
1154 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1155 {
1156 kthread_t *executor = zio->io_executor;
1157 spa_t *spa = zio->io_spa;
1158
1159 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1160 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1161 uint_t i;
1162 for (i = 0; i < tqs->stqs_count; i++) {
1163 if (taskq_member(tqs->stqs_taskq[i], executor))
1164 return (B_TRUE);
1165 }
1166 }
1167
1168 return (B_FALSE);
1169 }
1170
1171 static int
1172 zio_issue_async(zio_t *zio)
1173 {
1174 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1175
1176 return (ZIO_PIPELINE_STOP);
1177 }
1178
1179 void
1180 zio_interrupt(zio_t *zio)
1181 {
1182 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1183 }
1184
1185 /*
1186 * Execute the I/O pipeline until one of the following occurs:
1187 * (1) the I/O completes; (2) the pipeline stalls waiting for
1188 * dependent child I/Os; (3) the I/O issues, so we're waiting
1189 * for an I/O completion interrupt; (4) the I/O is delegated by
1190 * vdev-level caching or aggregation; (5) the I/O is deferred
1191 * due to vdev-level queueing; (6) the I/O is handed off to
1192 * another thread. In all cases, the pipeline stops whenever
1193 * there's no CPU work; it never burns a thread in cv_wait().
1194 *
1195 * There's no locking on io_stage because there's no legitimate way
1196 * for multiple threads to be attempting to process the same I/O.
1197 */
1198 static zio_pipe_stage_t *zio_pipeline[];
1199
1200 void
1201 zio_execute(zio_t *zio)
1202 {
1203 zio->io_executor = curthread;
1204
1205 while (zio->io_stage < ZIO_STAGE_DONE) {
1206 enum zio_stage pipeline = zio->io_pipeline;
1207 enum zio_stage stage = zio->io_stage;
1208 int rv;
1209
1210 ASSERT(!MUTEX_HELD(&zio->io_lock));
1211 ASSERT(ISP2(stage));
1212 ASSERT(zio->io_stall == NULL);
1213
1214 do {
1215 stage <<= 1;
1216 } while ((stage & pipeline) == 0);
1217
1218 ASSERT(stage <= ZIO_STAGE_DONE);
1219
1220 /*
1221 * If we are in interrupt context and this pipeline stage
1222 * will grab a config lock that is held across I/O,
1223 * or may wait for an I/O that needs an interrupt thread
1224 * to complete, issue async to avoid deadlock.
1225 *
1226 * For VDEV_IO_START, we cut in line so that the io will
1227 * be sent to disk promptly.
1228 */
1229 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1230 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1231 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1232 zio_requeue_io_start_cut_in_line : B_FALSE;
1233 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1234 return;
1235 }
1236
1237 zio->io_stage = stage;
1238 rv = zio_pipeline[highbit(stage) - 1](zio);
1239
1240 if (rv == ZIO_PIPELINE_STOP)
1241 return;
1242
1243 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1244 }
1245 }
1246
1247 /*
1248 * ==========================================================================
1249 * Initiate I/O, either sync or async
1250 * ==========================================================================
1251 */
1252 int
1253 zio_wait(zio_t *zio)
1254 {
1255 int error;
1256
1257 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1258 ASSERT(zio->io_executor == NULL);
1259
1260 zio->io_waiter = curthread;
1261
1262 zio_execute(zio);
1263
1264 mutex_enter(&zio->io_lock);
1265 while (zio->io_executor != NULL)
1266 cv_wait(&zio->io_cv, &zio->io_lock);
1267 mutex_exit(&zio->io_lock);
1268
1269 error = zio->io_error;
1270 zio_destroy(zio);
1271
1272 return (error);
1273 }
1274
1275 void
1276 zio_nowait(zio_t *zio)
1277 {
1278 ASSERT(zio->io_executor == NULL);
1279
1280 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1281 zio_unique_parent(zio) == NULL) {
1282 /*
1283 * This is a logical async I/O with no parent to wait for it.
1284 * We add it to the spa_async_root_zio "Godfather" I/O which
1285 * will ensure they complete prior to unloading the pool.
1286 */
1287 spa_t *spa = zio->io_spa;
1288
1289 zio_add_child(spa->spa_async_zio_root, zio);
1290 }
1291
1292 zio_execute(zio);
1293 }
1294
1295 /*
1296 * ==========================================================================
1297 * Reexecute or suspend/resume failed I/O
1298 * ==========================================================================
1299 */
1300
1301 static void
1302 zio_reexecute(zio_t *pio)
1303 {
1304 zio_t *cio, *cio_next;
1305
1306 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1307 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1308 ASSERT(pio->io_gang_leader == NULL);
1309 ASSERT(pio->io_gang_tree == NULL);
1310
1311 pio->io_flags = pio->io_orig_flags;
1312 pio->io_stage = pio->io_orig_stage;
1313 pio->io_pipeline = pio->io_orig_pipeline;
1314 pio->io_reexecute = 0;
1315 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1316 pio->io_error = 0;
1317 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1318 pio->io_state[w] = 0;
1319 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1320 pio->io_child_error[c] = 0;
1321
1322 if (IO_IS_ALLOCATING(pio))
1323 BP_ZERO(pio->io_bp);
1324
1325 /*
1326 * As we reexecute pio's children, new children could be created.
1327 * New children go to the head of pio's io_child_list, however,
1328 * so we will (correctly) not reexecute them. The key is that
1329 * the remainder of pio's io_child_list, from 'cio_next' onward,
1330 * cannot be affected by any side effects of reexecuting 'cio'.
1331 */
1332 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1333 cio_next = zio_walk_children(pio);
1334 mutex_enter(&pio->io_lock);
1335 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1336 pio->io_children[cio->io_child_type][w]++;
1337 mutex_exit(&pio->io_lock);
1338 zio_reexecute(cio);
1339 }
1340
1341 /*
1342 * Now that all children have been reexecuted, execute the parent.
1343 * We don't reexecute "The Godfather" I/O here as it's the
1344 * responsibility of the caller to wait on him.
1345 */
1346 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1347 zio_execute(pio);
1348 }
1349
1350 void
1351 zio_suspend(spa_t *spa, zio_t *zio)
1352 {
1353 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1354 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1355 "failure and the failure mode property for this pool "
1356 "is set to panic.", spa_name(spa));
1357
1358 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1359
1360 mutex_enter(&spa->spa_suspend_lock);
1361
1362 if (spa->spa_suspend_zio_root == NULL)
1363 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1364 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1365 ZIO_FLAG_GODFATHER);
1366
1367 spa->spa_suspended = B_TRUE;
1368
1369 if (zio != NULL) {
1370 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1371 ASSERT(zio != spa->spa_suspend_zio_root);
1372 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1373 ASSERT(zio_unique_parent(zio) == NULL);
1374 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1375 zio_add_child(spa->spa_suspend_zio_root, zio);
1376 }
1377
1378 mutex_exit(&spa->spa_suspend_lock);
1379 }
1380
1381 int
1382 zio_resume(spa_t *spa)
1383 {
1384 zio_t *pio;
1385
1386 /*
1387 * Reexecute all previously suspended i/o.
1388 */
1389 mutex_enter(&spa->spa_suspend_lock);
1390 spa->spa_suspended = B_FALSE;
1391 cv_broadcast(&spa->spa_suspend_cv);
1392 pio = spa->spa_suspend_zio_root;
1393 spa->spa_suspend_zio_root = NULL;
1394 mutex_exit(&spa->spa_suspend_lock);
1395
1396 if (pio == NULL)
1397 return (0);
1398
1399 zio_reexecute(pio);
1400 return (zio_wait(pio));
1401 }
1402
1403 void
1404 zio_resume_wait(spa_t *spa)
1405 {
1406 mutex_enter(&spa->spa_suspend_lock);
1407 while (spa_suspended(spa))
1408 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1409 mutex_exit(&spa->spa_suspend_lock);
1410 }
1411
1412 /*
1413 * ==========================================================================
1414 * Gang blocks.
1415 *
1416 * A gang block is a collection of small blocks that looks to the DMU
1417 * like one large block. When zio_dva_allocate() cannot find a block
1418 * of the requested size, due to either severe fragmentation or the pool
1419 * being nearly full, it calls zio_write_gang_block() to construct the
1420 * block from smaller fragments.
1421 *
1422 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1423 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1424 * an indirect block: it's an array of block pointers. It consumes
1425 * only one sector and hence is allocatable regardless of fragmentation.
1426 * The gang header's bps point to its gang members, which hold the data.
1427 *
1428 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1429 * as the verifier to ensure uniqueness of the SHA256 checksum.
1430 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1431 * not the gang header. This ensures that data block signatures (needed for
1432 * deduplication) are independent of how the block is physically stored.
1433 *
1434 * Gang blocks can be nested: a gang member may itself be a gang block.
1435 * Thus every gang block is a tree in which root and all interior nodes are
1436 * gang headers, and the leaves are normal blocks that contain user data.
1437 * The root of the gang tree is called the gang leader.
1438 *
1439 * To perform any operation (read, rewrite, free, claim) on a gang block,
1440 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1441 * in the io_gang_tree field of the original logical i/o by recursively
1442 * reading the gang leader and all gang headers below it. This yields
1443 * an in-core tree containing the contents of every gang header and the
1444 * bps for every constituent of the gang block.
1445 *
1446 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1447 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1448 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1449 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1450 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1451 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1452 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1453 * of the gang header plus zio_checksum_compute() of the data to update the
1454 * gang header's blk_cksum as described above.
1455 *
1456 * The two-phase assemble/issue model solves the problem of partial failure --
1457 * what if you'd freed part of a gang block but then couldn't read the
1458 * gang header for another part? Assembling the entire gang tree first
1459 * ensures that all the necessary gang header I/O has succeeded before
1460 * starting the actual work of free, claim, or write. Once the gang tree
1461 * is assembled, free and claim are in-memory operations that cannot fail.
1462 *
1463 * In the event that a gang write fails, zio_dva_unallocate() walks the
1464 * gang tree to immediately free (i.e. insert back into the space map)
1465 * everything we've allocated. This ensures that we don't get ENOSPC
1466 * errors during repeated suspend/resume cycles due to a flaky device.
1467 *
1468 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1469 * the gang tree, we won't modify the block, so we can safely defer the free
1470 * (knowing that the block is still intact). If we *can* assemble the gang
1471 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1472 * each constituent bp and we can allocate a new block on the next sync pass.
1473 *
1474 * In all cases, the gang tree allows complete recovery from partial failure.
1475 * ==========================================================================
1476 */
1477
1478 static zio_t *
1479 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1480 {
1481 if (gn != NULL)
1482 return (pio);
1483
1484 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1485 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1486 &pio->io_bookmark));
1487 }
1488
1489 zio_t *
1490 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1491 {
1492 zio_t *zio;
1493
1494 if (gn != NULL) {
1495 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1496 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1497 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1498 /*
1499 * As we rewrite each gang header, the pipeline will compute
1500 * a new gang block header checksum for it; but no one will
1501 * compute a new data checksum, so we do that here. The one
1502 * exception is the gang leader: the pipeline already computed
1503 * its data checksum because that stage precedes gang assembly.
1504 * (Presently, nothing actually uses interior data checksums;
1505 * this is just good hygiene.)
1506 */
1507 if (gn != pio->io_gang_leader->io_gang_tree) {
1508 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1509 data, BP_GET_PSIZE(bp));
1510 }
1511 /*
1512 * If we are here to damage data for testing purposes,
1513 * leave the GBH alone so that we can detect the damage.
1514 */
1515 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1516 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1517 } else {
1518 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1519 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1520 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1521 }
1522
1523 return (zio);
1524 }
1525
1526 /* ARGSUSED */
1527 zio_t *
1528 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1529 {
1530 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1531 ZIO_GANG_CHILD_FLAGS(pio)));
1532 }
1533
1534 /* ARGSUSED */
1535 zio_t *
1536 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1537 {
1538 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1539 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1540 }
1541
1542 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1543 NULL,
1544 zio_read_gang,
1545 zio_rewrite_gang,
1546 zio_free_gang,
1547 zio_claim_gang,
1548 NULL
1549 };
1550
1551 static void zio_gang_tree_assemble_done(zio_t *zio);
1552
1553 static zio_gang_node_t *
1554 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1555 {
1556 zio_gang_node_t *gn;
1557
1558 ASSERT(*gnpp == NULL);
1559
1560 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1561 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1562 *gnpp = gn;
1563
1564 return (gn);
1565 }
1566
1567 static void
1568 zio_gang_node_free(zio_gang_node_t **gnpp)
1569 {
1570 zio_gang_node_t *gn = *gnpp;
1571
1572 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1573 ASSERT(gn->gn_child[g] == NULL);
1574
1575 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1576 kmem_free(gn, sizeof (*gn));
1577 *gnpp = NULL;
1578 }
1579
1580 static void
1581 zio_gang_tree_free(zio_gang_node_t **gnpp)
1582 {
1583 zio_gang_node_t *gn = *gnpp;
1584
1585 if (gn == NULL)
1586 return;
1587
1588 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1589 zio_gang_tree_free(&gn->gn_child[g]);
1590
1591 zio_gang_node_free(gnpp);
1592 }
1593
1594 static void
1595 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1596 {
1597 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1598
1599 ASSERT(gio->io_gang_leader == gio);
1600 ASSERT(BP_IS_GANG(bp));
1601
1602 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1603 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1604 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1605 }
1606
1607 static void
1608 zio_gang_tree_assemble_done(zio_t *zio)
1609 {
1610 zio_t *gio = zio->io_gang_leader;
1611 zio_gang_node_t *gn = zio->io_private;
1612 blkptr_t *bp = zio->io_bp;
1613
1614 ASSERT(gio == zio_unique_parent(zio));
1615 ASSERT(zio->io_child_count == 0);
1616
1617 if (zio->io_error)
1618 return;
1619
1620 if (BP_SHOULD_BYTESWAP(bp))
1621 byteswap_uint64_array(zio->io_data, zio->io_size);
1622
1623 ASSERT(zio->io_data == gn->gn_gbh);
1624 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1625 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1626
1627 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1628 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1629 if (!BP_IS_GANG(gbp))
1630 continue;
1631 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1632 }
1633 }
1634
1635 static void
1636 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1637 {
1638 zio_t *gio = pio->io_gang_leader;
1639 zio_t *zio;
1640
1641 ASSERT(BP_IS_GANG(bp) == !!gn);
1642 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1643 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1644
1645 /*
1646 * If you're a gang header, your data is in gn->gn_gbh.
1647 * If you're a gang member, your data is in 'data' and gn == NULL.
1648 */
1649 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1650
1651 if (gn != NULL) {
1652 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1653
1654 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1655 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1656 if (BP_IS_HOLE(gbp))
1657 continue;
1658 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1659 data = (char *)data + BP_GET_PSIZE(gbp);
1660 }
1661 }
1662
1663 if (gn == gio->io_gang_tree)
1664 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1665
1666 if (zio != pio)
1667 zio_nowait(zio);
1668 }
1669
1670 static int
1671 zio_gang_assemble(zio_t *zio)
1672 {
1673 blkptr_t *bp = zio->io_bp;
1674
1675 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1676 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1677
1678 zio->io_gang_leader = zio;
1679
1680 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1681
1682 return (ZIO_PIPELINE_CONTINUE);
1683 }
1684
1685 static int
1686 zio_gang_issue(zio_t *zio)
1687 {
1688 blkptr_t *bp = zio->io_bp;
1689
1690 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1691 return (ZIO_PIPELINE_STOP);
1692
1693 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1694 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1695
1696 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1697 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1698 else
1699 zio_gang_tree_free(&zio->io_gang_tree);
1700
1701 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1702
1703 return (ZIO_PIPELINE_CONTINUE);
1704 }
1705
1706 static void
1707 zio_write_gang_member_ready(zio_t *zio)
1708 {
1709 zio_t *pio = zio_unique_parent(zio);
1710 zio_t *gio = zio->io_gang_leader;
1711 dva_t *cdva = zio->io_bp->blk_dva;
1712 dva_t *pdva = pio->io_bp->blk_dva;
1713 uint64_t asize;
1714
1715 if (BP_IS_HOLE(zio->io_bp))
1716 return;
1717
1718 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1719
1720 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1721 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1722 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1723 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1724 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1725
1726 mutex_enter(&pio->io_lock);
1727 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1728 ASSERT(DVA_GET_GANG(&pdva[d]));
1729 asize = DVA_GET_ASIZE(&pdva[d]);
1730 asize += DVA_GET_ASIZE(&cdva[d]);
1731 DVA_SET_ASIZE(&pdva[d], asize);
1732 }
1733 mutex_exit(&pio->io_lock);
1734 }
1735
1736 static int
1737 zio_write_gang_block(zio_t *pio)
1738 {
1739 spa_t *spa = pio->io_spa;
1740 blkptr_t *bp = pio->io_bp;
1741 zio_t *gio = pio->io_gang_leader;
1742 zio_t *zio;
1743 zio_gang_node_t *gn, **gnpp;
1744 zio_gbh_phys_t *gbh;
1745 uint64_t txg = pio->io_txg;
1746 uint64_t resid = pio->io_size;
1747 uint64_t lsize;
1748 int copies = gio->io_prop.zp_copies;
1749 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1750 zio_prop_t zp;
1751 int error;
1752
1753 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1754 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1755 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1756 if (error) {
1757 pio->io_error = error;
1758 return (ZIO_PIPELINE_CONTINUE);
1759 }
1760
1761 if (pio == gio) {
1762 gnpp = &gio->io_gang_tree;
1763 } else {
1764 gnpp = pio->io_private;
1765 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1766 }
1767
1768 gn = zio_gang_node_alloc(gnpp);
1769 gbh = gn->gn_gbh;
1770 bzero(gbh, SPA_GANGBLOCKSIZE);
1771
1772 /*
1773 * Create the gang header.
1774 */
1775 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1776 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1777
1778 /*
1779 * Create and nowait the gang children.
1780 */
1781 for (int g = 0; resid != 0; resid -= lsize, g++) {
1782 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1783 SPA_MINBLOCKSIZE);
1784 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1785
1786 zp.zp_checksum = gio->io_prop.zp_checksum;
1787 zp.zp_compress = ZIO_COMPRESS_OFF;
1788 zp.zp_type = DMU_OT_NONE;
1789 zp.zp_level = 0;
1790 zp.zp_copies = gio->io_prop.zp_copies;
1791 zp.zp_dedup = B_FALSE;
1792 zp.zp_dedup_verify = B_FALSE;
1793 zp.zp_nopwrite = B_FALSE;
1794
1795 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1796 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1797 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1798 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1799 &pio->io_bookmark));
1800 }
1801
1802 /*
1803 * Set pio's pipeline to just wait for zio to finish.
1804 */
1805 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1806
1807 zio_nowait(zio);
1808
1809 return (ZIO_PIPELINE_CONTINUE);
1810 }
1811
1812 /*
1813 * The zio_nop_write stage in the pipeline determines if allocating
1814 * a new bp is necessary. By leveraging a cryptographically secure checksum,
1815 * such as SHA256, we can compare the checksums of the new data and the old
1816 * to determine if allocating a new block is required. The nopwrite
1817 * feature can handle writes in either syncing or open context (i.e. zil
1818 * writes) and as a result is mutually exclusive with dedup.
1819 */
1820 static int
1821 zio_nop_write(zio_t *zio)
1822 {
1823 blkptr_t *bp = zio->io_bp;
1824 blkptr_t *bp_orig = &zio->io_bp_orig;
1825 zio_prop_t *zp = &zio->io_prop;
1826
1827 ASSERT(BP_GET_LEVEL(bp) == 0);
1828 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1829 ASSERT(zp->zp_nopwrite);
1830 ASSERT(!zp->zp_dedup);
1831 ASSERT(zio->io_bp_override == NULL);
1832 ASSERT(IO_IS_ALLOCATING(zio));
1833
1834 /*
1835 * Check to see if the original bp and the new bp have matching
1836 * characteristics (i.e. same checksum, compression algorithms, etc).
1837 * If they don't then just continue with the pipeline which will
1838 * allocate a new bp.
1839 */
1840 if (BP_IS_HOLE(bp_orig) ||
1841 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
1842 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
1843 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
1844 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
1845 zp->zp_copies != BP_GET_NDVAS(bp_orig))
1846 return (ZIO_PIPELINE_CONTINUE);
1847
1848 /*
1849 * If the checksums match then reset the pipeline so that we
1850 * avoid allocating a new bp and issuing any I/O.
1851 */
1852 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
1853 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
1854 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
1855 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
1856 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
1857 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
1858 sizeof (uint64_t)) == 0);
1859
1860 *bp = *bp_orig;
1861 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1862 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1863 }
1864
1865 return (ZIO_PIPELINE_CONTINUE);
1866 }
1867
1868 /*
1869 * ==========================================================================
1870 * Dedup
1871 * ==========================================================================
1872 */
1873 static void
1874 zio_ddt_child_read_done(zio_t *zio)
1875 {
1876 blkptr_t *bp = zio->io_bp;
1877 ddt_entry_t *dde = zio->io_private;
1878 ddt_phys_t *ddp;
1879 zio_t *pio = zio_unique_parent(zio);
1880
1881 mutex_enter(&pio->io_lock);
1882 ddp = ddt_phys_select(dde, bp);
1883 if (zio->io_error == 0)
1884 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1885 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1886 dde->dde_repair_data = zio->io_data;
1887 else
1888 zio_buf_free(zio->io_data, zio->io_size);
1889 mutex_exit(&pio->io_lock);
1890 }
1891
1892 static int
1893 zio_ddt_read_start(zio_t *zio)
1894 {
1895 blkptr_t *bp = zio->io_bp;
1896
1897 ASSERT(BP_GET_DEDUP(bp));
1898 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1899 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1900
1901 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1902 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1903 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1904 ddt_phys_t *ddp = dde->dde_phys;
1905 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1906 blkptr_t blk;
1907
1908 ASSERT(zio->io_vsd == NULL);
1909 zio->io_vsd = dde;
1910
1911 if (ddp_self == NULL)
1912 return (ZIO_PIPELINE_CONTINUE);
1913
1914 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1915 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1916 continue;
1917 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1918 &blk);
1919 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1920 zio_buf_alloc(zio->io_size), zio->io_size,
1921 zio_ddt_child_read_done, dde, zio->io_priority,
1922 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1923 &zio->io_bookmark));
1924 }
1925 return (ZIO_PIPELINE_CONTINUE);
1926 }
1927
1928 zio_nowait(zio_read(zio, zio->io_spa, bp,
1929 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1930 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1931
1932 return (ZIO_PIPELINE_CONTINUE);
1933 }
1934
1935 static int
1936 zio_ddt_read_done(zio_t *zio)
1937 {
1938 blkptr_t *bp = zio->io_bp;
1939
1940 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1941 return (ZIO_PIPELINE_STOP);
1942
1943 ASSERT(BP_GET_DEDUP(bp));
1944 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1945 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1946
1947 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1948 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1949 ddt_entry_t *dde = zio->io_vsd;
1950 if (ddt == NULL) {
1951 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1952 return (ZIO_PIPELINE_CONTINUE);
1953 }
1954 if (dde == NULL) {
1955 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1956 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1957 return (ZIO_PIPELINE_STOP);
1958 }
1959 if (dde->dde_repair_data != NULL) {
1960 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1961 zio->io_child_error[ZIO_CHILD_DDT] = 0;
1962 }
1963 ddt_repair_done(ddt, dde);
1964 zio->io_vsd = NULL;
1965 }
1966
1967 ASSERT(zio->io_vsd == NULL);
1968
1969 return (ZIO_PIPELINE_CONTINUE);
1970 }
1971
1972 static boolean_t
1973 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1974 {
1975 spa_t *spa = zio->io_spa;
1976
1977 /*
1978 * Note: we compare the original data, not the transformed data,
1979 * because when zio->io_bp is an override bp, we will not have
1980 * pushed the I/O transforms. That's an important optimization
1981 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1982 */
1983 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1984 zio_t *lio = dde->dde_lead_zio[p];
1985
1986 if (lio != NULL) {
1987 return (lio->io_orig_size != zio->io_orig_size ||
1988 bcmp(zio->io_orig_data, lio->io_orig_data,
1989 zio->io_orig_size) != 0);
1990 }
1991 }
1992
1993 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1994 ddt_phys_t *ddp = &dde->dde_phys[p];
1995
1996 if (ddp->ddp_phys_birth != 0) {
1997 arc_buf_t *abuf = NULL;
1998 uint32_t aflags = ARC_WAIT;
1999 blkptr_t blk = *zio->io_bp;
2000 int error;
2001
2002 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2003
2004 ddt_exit(ddt);
2005
2006 error = arc_read(NULL, spa, &blk,
2007 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2008 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2009 &aflags, &zio->io_bookmark);
2010
2011 if (error == 0) {
2012 if (arc_buf_size(abuf) != zio->io_orig_size ||
2013 bcmp(abuf->b_data, zio->io_orig_data,
2014 zio->io_orig_size) != 0)
2015 error = SET_ERROR(EEXIST);
2016 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2017 }
2018
2019 ddt_enter(ddt);
2020 return (error != 0);
2021 }
2022 }
2023
2024 return (B_FALSE);
2025 }
2026
2027 static void
2028 zio_ddt_child_write_ready(zio_t *zio)
2029 {
2030 int p = zio->io_prop.zp_copies;
2031 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2032 ddt_entry_t *dde = zio->io_private;
2033 ddt_phys_t *ddp = &dde->dde_phys[p];
2034 zio_t *pio;
2035
2036 if (zio->io_error)
2037 return;
2038
2039 ddt_enter(ddt);
2040
2041 ASSERT(dde->dde_lead_zio[p] == zio);
2042
2043 ddt_phys_fill(ddp, zio->io_bp);
2044
2045 while ((pio = zio_walk_parents(zio)) != NULL)
2046 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2047
2048 ddt_exit(ddt);
2049 }
2050
2051 static void
2052 zio_ddt_child_write_done(zio_t *zio)
2053 {
2054 int p = zio->io_prop.zp_copies;
2055 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2056 ddt_entry_t *dde = zio->io_private;
2057 ddt_phys_t *ddp = &dde->dde_phys[p];
2058
2059 ddt_enter(ddt);
2060
2061 ASSERT(ddp->ddp_refcnt == 0);
2062 ASSERT(dde->dde_lead_zio[p] == zio);
2063 dde->dde_lead_zio[p] = NULL;
2064
2065 if (zio->io_error == 0) {
2066 while (zio_walk_parents(zio) != NULL)
2067 ddt_phys_addref(ddp);
2068 } else {
2069 ddt_phys_clear(ddp);
2070 }
2071
2072 ddt_exit(ddt);
2073 }
2074
2075 static void
2076 zio_ddt_ditto_write_done(zio_t *zio)
2077 {
2078 int p = DDT_PHYS_DITTO;
2079 zio_prop_t *zp = &zio->io_prop;
2080 blkptr_t *bp = zio->io_bp;
2081 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2082 ddt_entry_t *dde = zio->io_private;
2083 ddt_phys_t *ddp = &dde->dde_phys[p];
2084 ddt_key_t *ddk = &dde->dde_key;
2085
2086 ddt_enter(ddt);
2087
2088 ASSERT(ddp->ddp_refcnt == 0);
2089 ASSERT(dde->dde_lead_zio[p] == zio);
2090 dde->dde_lead_zio[p] = NULL;
2091
2092 if (zio->io_error == 0) {
2093 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2094 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2095 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2096 if (ddp->ddp_phys_birth != 0)
2097 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2098 ddt_phys_fill(ddp, bp);
2099 }
2100
2101 ddt_exit(ddt);
2102 }
2103
2104 static int
2105 zio_ddt_write(zio_t *zio)
2106 {
2107 spa_t *spa = zio->io_spa;
2108 blkptr_t *bp = zio->io_bp;
2109 uint64_t txg = zio->io_txg;
2110 zio_prop_t *zp = &zio->io_prop;
2111 int p = zp->zp_copies;
2112 int ditto_copies;
2113 zio_t *cio = NULL;
2114 zio_t *dio = NULL;
2115 ddt_t *ddt = ddt_select(spa, bp);
2116 ddt_entry_t *dde;
2117 ddt_phys_t *ddp;
2118
2119 ASSERT(BP_GET_DEDUP(bp));
2120 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2121 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2122
2123 ddt_enter(ddt);
2124 dde = ddt_lookup(ddt, bp, B_TRUE);
2125 ddp = &dde->dde_phys[p];
2126
2127 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2128 /*
2129 * If we're using a weak checksum, upgrade to a strong checksum
2130 * and try again. If we're already using a strong checksum,
2131 * we can't resolve it, so just convert to an ordinary write.
2132 * (And automatically e-mail a paper to Nature?)
2133 */
2134 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2135 zp->zp_checksum = spa_dedup_checksum(spa);
2136 zio_pop_transforms(zio);
2137 zio->io_stage = ZIO_STAGE_OPEN;
2138 BP_ZERO(bp);
2139 } else {
2140 zp->zp_dedup = B_FALSE;
2141 }
2142 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2143 ddt_exit(ddt);
2144 return (ZIO_PIPELINE_CONTINUE);
2145 }
2146
2147 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2148 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2149
2150 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2151 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2152 zio_prop_t czp = *zp;
2153
2154 czp.zp_copies = ditto_copies;
2155
2156 /*
2157 * If we arrived here with an override bp, we won't have run
2158 * the transform stack, so we won't have the data we need to
2159 * generate a child i/o. So, toss the override bp and restart.
2160 * This is safe, because using the override bp is just an
2161 * optimization; and it's rare, so the cost doesn't matter.
2162 */
2163 if (zio->io_bp_override) {
2164 zio_pop_transforms(zio);
2165 zio->io_stage = ZIO_STAGE_OPEN;
2166 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2167 zio->io_bp_override = NULL;
2168 BP_ZERO(bp);
2169 ddt_exit(ddt);
2170 return (ZIO_PIPELINE_CONTINUE);
2171 }
2172
2173 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2174 zio->io_orig_size, &czp, NULL,
2175 zio_ddt_ditto_write_done, dde, zio->io_priority,
2176 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2177
2178 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2179 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2180 }
2181
2182 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2183 if (ddp->ddp_phys_birth != 0)
2184 ddt_bp_fill(ddp, bp, txg);
2185 if (dde->dde_lead_zio[p] != NULL)
2186 zio_add_child(zio, dde->dde_lead_zio[p]);
2187 else
2188 ddt_phys_addref(ddp);
2189 } else if (zio->io_bp_override) {
2190 ASSERT(bp->blk_birth == txg);
2191 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2192 ddt_phys_fill(ddp, bp);
2193 ddt_phys_addref(ddp);
2194 } else {
2195 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2196 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2197 zio_ddt_child_write_done, dde, zio->io_priority,
2198 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2199
2200 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2201 dde->dde_lead_zio[p] = cio;
2202 }
2203
2204 ddt_exit(ddt);
2205
2206 if (cio)
2207 zio_nowait(cio);
2208 if (dio)
2209 zio_nowait(dio);
2210
2211 return (ZIO_PIPELINE_CONTINUE);
2212 }
2213
2214 ddt_entry_t *freedde; /* for debugging */
2215
2216 static int
2217 zio_ddt_free(zio_t *zio)
2218 {
2219 spa_t *spa = zio->io_spa;
2220 blkptr_t *bp = zio->io_bp;
2221 ddt_t *ddt = ddt_select(spa, bp);
2222 ddt_entry_t *dde;
2223 ddt_phys_t *ddp;
2224
2225 ASSERT(BP_GET_DEDUP(bp));
2226 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2227
2228 ddt_enter(ddt);
2229 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2230 ddp = ddt_phys_select(dde, bp);
2231 ddt_phys_decref(ddp);
2232 ddt_exit(ddt);
2233
2234 return (ZIO_PIPELINE_CONTINUE);
2235 }
2236
2237 /*
2238 * ==========================================================================
2239 * Allocate and free blocks
2240 * ==========================================================================
2241 */
2242 static int
2243 zio_dva_allocate(zio_t *zio)
2244 {
2245 spa_t *spa = zio->io_spa;
2246 metaslab_class_t *mc = spa_normal_class(spa);
2247 blkptr_t *bp = zio->io_bp;
2248 int error;
2249 int flags = 0;
2250
2251 if (zio->io_gang_leader == NULL) {
2252 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2253 zio->io_gang_leader = zio;
2254 }
2255
2256 ASSERT(BP_IS_HOLE(bp));
2257 ASSERT0(BP_GET_NDVAS(bp));
2258 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2259 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2260 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2261
2262 /*
2263 * The dump device does not support gang blocks so allocation on
2264 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2265 * the "fast" gang feature.
2266 */
2267 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2268 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2269 METASLAB_GANG_CHILD : 0;
2270 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2271 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2272
2273 if (error) {
2274 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2275 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2276 error);
2277 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2278 return (zio_write_gang_block(zio));
2279 zio->io_error = error;
2280 }
2281
2282 return (ZIO_PIPELINE_CONTINUE);
2283 }
2284
2285 static int
2286 zio_dva_free(zio_t *zio)
2287 {
2288 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2289
2290 return (ZIO_PIPELINE_CONTINUE);
2291 }
2292
2293 static int
2294 zio_dva_claim(zio_t *zio)
2295 {
2296 int error;
2297
2298 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2299 if (error)
2300 zio->io_error = error;
2301
2302 return (ZIO_PIPELINE_CONTINUE);
2303 }
2304
2305 /*
2306 * Undo an allocation. This is used by zio_done() when an I/O fails
2307 * and we want to give back the block we just allocated.
2308 * This handles both normal blocks and gang blocks.
2309 */
2310 static void
2311 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2312 {
2313 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2314 ASSERT(zio->io_bp_override == NULL);
2315
2316 if (!BP_IS_HOLE(bp))
2317 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2318
2319 if (gn != NULL) {
2320 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2321 zio_dva_unallocate(zio, gn->gn_child[g],
2322 &gn->gn_gbh->zg_blkptr[g]);
2323 }
2324 }
2325 }
2326
2327 /*
2328 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2329 */
2330 int
2331 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2332 uint64_t size, boolean_t use_slog)
2333 {
2334 int error = 1;
2335
2336 ASSERT(txg > spa_syncing_txg(spa));
2337
2338 /*
2339 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2340 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2341 * when allocating them.
2342 */
2343 if (use_slog) {
2344 error = metaslab_alloc(spa, spa_log_class(spa), size,
2345 new_bp, 1, txg, old_bp,
2346 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2347 }
2348
2349 if (error) {
2350 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2351 new_bp, 1, txg, old_bp,
2352 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2353 }
2354
2355 if (error == 0) {
2356 BP_SET_LSIZE(new_bp, size);
2357 BP_SET_PSIZE(new_bp, size);
2358 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2359 BP_SET_CHECKSUM(new_bp,
2360 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2361 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2362 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2363 BP_SET_LEVEL(new_bp, 0);
2364 BP_SET_DEDUP(new_bp, 0);
2365 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2366 }
2367
2368 return (error);
2369 }
2370
2371 /*
2372 * Free an intent log block.
2373 */
2374 void
2375 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2376 {
2377 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2378 ASSERT(!BP_IS_GANG(bp));
2379
2380 zio_free(spa, txg, bp);
2381 }
2382
2383 /*
2384 * ==========================================================================
2385 * Read and write to physical devices
2386 * ==========================================================================
2387 */
2388 static int
2389 zio_vdev_io_start(zio_t *zio)
2390 {
2391 vdev_t *vd = zio->io_vd;
2392 uint64_t align;
2393 spa_t *spa = zio->io_spa;
2394
2395 ASSERT(zio->io_error == 0);
2396 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2397
2398 if (vd == NULL) {
2399 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2400 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2401
2402 /*
2403 * The mirror_ops handle multiple DVAs in a single BP.
2404 */
2405 return (vdev_mirror_ops.vdev_op_io_start(zio));
2406 }
2407
2408 /*
2409 * We keep track of time-sensitive I/Os so that the scan thread
2410 * can quickly react to certain workloads. In particular, we care
2411 * about non-scrubbing, top-level reads and writes with the following
2412 * characteristics:
2413 * - synchronous writes of user data to non-slog devices
2414 * - any reads of user data
2415 * When these conditions are met, adjust the timestamp of spa_last_io
2416 * which allows the scan thread to adjust its workload accordingly.
2417 */
2418 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2419 vd == vd->vdev_top && !vd->vdev_islog &&
2420 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2421 zio->io_txg != spa_syncing_txg(spa)) {
2422 uint64_t old = spa->spa_last_io;
2423 uint64_t new = ddi_get_lbolt64();
2424 if (old != new)
2425 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2426 }
2427
2428 align = 1ULL << vd->vdev_top->vdev_ashift;
2429
2430 if (P2PHASE(zio->io_size, align) != 0) {
2431 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2432 char *abuf = zio_buf_alloc(asize);
2433 ASSERT(vd == vd->vdev_top);
2434 if (zio->io_type == ZIO_TYPE_WRITE) {
2435 bcopy(zio->io_data, abuf, zio->io_size);
2436 bzero(abuf + zio->io_size, asize - zio->io_size);
2437 }
2438 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2439 }
2440
2441 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2442 ASSERT(P2PHASE(zio->io_size, align) == 0);
2443 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2444
2445 /*
2446 * If this is a repair I/O, and there's no self-healing involved --
2447 * that is, we're just resilvering what we expect to resilver --
2448 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2449 * This prevents spurious resilvering with nested replication.
2450 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2451 * A is out of date, we'll read from C+D, then use the data to
2452 * resilver A+B -- but we don't actually want to resilver B, just A.
2453 * The top-level mirror has no way to know this, so instead we just
2454 * discard unnecessary repairs as we work our way down the vdev tree.
2455 * The same logic applies to any form of nested replication:
2456 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2457 */
2458 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2459 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2460 zio->io_txg != 0 && /* not a delegated i/o */
2461 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2462 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2463 zio_vdev_io_bypass(zio);
2464 return (ZIO_PIPELINE_CONTINUE);
2465 }
2466
2467 if (vd->vdev_ops->vdev_op_leaf &&
2468 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2469
2470 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2471 return (ZIO_PIPELINE_CONTINUE);
2472
2473 if ((zio = vdev_queue_io(zio)) == NULL)
2474 return (ZIO_PIPELINE_STOP);
2475
2476 if (!vdev_accessible(vd, zio)) {
2477 zio->io_error = SET_ERROR(ENXIO);
2478 zio_interrupt(zio);
2479 return (ZIO_PIPELINE_STOP);
2480 }
2481 }
2482
2483 return (vd->vdev_ops->vdev_op_io_start(zio));
2484 }
2485
2486 static int
2487 zio_vdev_io_done(zio_t *zio)
2488 {
2489 vdev_t *vd = zio->io_vd;
2490 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2491 boolean_t unexpected_error = B_FALSE;
2492
2493 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2494 return (ZIO_PIPELINE_STOP);
2495
2496 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2497
2498 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2499
2500 vdev_queue_io_done(zio);
2501
2502 if (zio->io_type == ZIO_TYPE_WRITE)
2503 vdev_cache_write(zio);
2504
2505 if (zio_injection_enabled && zio->io_error == 0)
2506 zio->io_error = zio_handle_device_injection(vd,
2507 zio, EIO);
2508
2509 if (zio_injection_enabled && zio->io_error == 0)
2510 zio->io_error = zio_handle_label_injection(zio, EIO);
2511
2512 if (zio->io_error) {
2513 if (!vdev_accessible(vd, zio)) {
2514 zio->io_error = SET_ERROR(ENXIO);
2515 } else {
2516 unexpected_error = B_TRUE;
2517 }
2518 }
2519 }
2520
2521 ops->vdev_op_io_done(zio);
2522
2523 if (unexpected_error)
2524 VERIFY(vdev_probe(vd, zio) == NULL);
2525
2526 return (ZIO_PIPELINE_CONTINUE);
2527 }
2528
2529 /*
2530 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2531 * disk, and use that to finish the checksum ereport later.
2532 */
2533 static void
2534 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2535 const void *good_buf)
2536 {
2537 /* no processing needed */
2538 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2539 }
2540
2541 /*ARGSUSED*/
2542 void
2543 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2544 {
2545 void *buf = zio_buf_alloc(zio->io_size);
2546
2547 bcopy(zio->io_data, buf, zio->io_size);
2548
2549 zcr->zcr_cbinfo = zio->io_size;
2550 zcr->zcr_cbdata = buf;
2551 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2552 zcr->zcr_free = zio_buf_free;
2553 }
2554
2555 static int
2556 zio_vdev_io_assess(zio_t *zio)
2557 {
2558 vdev_t *vd = zio->io_vd;
2559
2560 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2561 return (ZIO_PIPELINE_STOP);
2562
2563 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2564 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2565
2566 if (zio->io_vsd != NULL) {
2567 zio->io_vsd_ops->vsd_free(zio);
2568 zio->io_vsd = NULL;
2569 }
2570
2571 if (zio_injection_enabled && zio->io_error == 0)
2572 zio->io_error = zio_handle_fault_injection(zio, EIO);
2573
2574 /*
2575 * If the I/O failed, determine whether we should attempt to retry it.
2576 *
2577 * On retry, we cut in line in the issue queue, since we don't want
2578 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2579 */
2580 if (zio->io_error && vd == NULL &&
2581 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2582 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2583 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2584 zio->io_error = 0;
2585 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2586 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2587 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2588 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2589 zio_requeue_io_start_cut_in_line);
2590 return (ZIO_PIPELINE_STOP);
2591 }
2592
2593 /*
2594 * If we got an error on a leaf device, convert it to ENXIO
2595 * if the device is not accessible at all.
2596 */
2597 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2598 !vdev_accessible(vd, zio))
2599 zio->io_error = SET_ERROR(ENXIO);
2600
2601 /*
2602 * If we can't write to an interior vdev (mirror or RAID-Z),
2603 * set vdev_cant_write so that we stop trying to allocate from it.
2604 */
2605 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2606 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2607 vd->vdev_cant_write = B_TRUE;
2608 }
2609
2610 if (zio->io_error)
2611 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2612
2613 return (ZIO_PIPELINE_CONTINUE);
2614 }
2615
2616 void
2617 zio_vdev_io_reissue(zio_t *zio)
2618 {
2619 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2620 ASSERT(zio->io_error == 0);
2621
2622 zio->io_stage >>= 1;
2623 }
2624
2625 void
2626 zio_vdev_io_redone(zio_t *zio)
2627 {
2628 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2629
2630 zio->io_stage >>= 1;
2631 }
2632
2633 void
2634 zio_vdev_io_bypass(zio_t *zio)
2635 {
2636 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2637 ASSERT(zio->io_error == 0);
2638
2639 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2640 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2641 }
2642
2643 /*
2644 * ==========================================================================
2645 * Generate and verify checksums
2646 * ==========================================================================
2647 */
2648 static int
2649 zio_checksum_generate(zio_t *zio)
2650 {
2651 blkptr_t *bp = zio->io_bp;
2652 enum zio_checksum checksum;
2653
2654 if (bp == NULL) {
2655 /*
2656 * This is zio_write_phys().
2657 * We're either generating a label checksum, or none at all.
2658 */
2659 checksum = zio->io_prop.zp_checksum;
2660
2661 if (checksum == ZIO_CHECKSUM_OFF)
2662 return (ZIO_PIPELINE_CONTINUE);
2663
2664 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2665 } else {
2666 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2667 ASSERT(!IO_IS_ALLOCATING(zio));
2668 checksum = ZIO_CHECKSUM_GANG_HEADER;
2669 } else {
2670 checksum = BP_GET_CHECKSUM(bp);
2671 }
2672 }
2673
2674 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2675
2676 return (ZIO_PIPELINE_CONTINUE);
2677 }
2678
2679 static int
2680 zio_checksum_verify(zio_t *zio)
2681 {
2682 zio_bad_cksum_t info;
2683 blkptr_t *bp = zio->io_bp;
2684 int error;
2685
2686 ASSERT(zio->io_vd != NULL);
2687
2688 if (bp == NULL) {
2689 /*
2690 * This is zio_read_phys().
2691 * We're either verifying a label checksum, or nothing at all.
2692 */
2693 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2694 return (ZIO_PIPELINE_CONTINUE);
2695
2696 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2697 }
2698
2699 if ((error = zio_checksum_error(zio, &info)) != 0) {
2700 zio->io_error = error;
2701 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2702 zfs_ereport_start_checksum(zio->io_spa,
2703 zio->io_vd, zio, zio->io_offset,
2704 zio->io_size, NULL, &info);
2705 }
2706 }
2707
2708 return (ZIO_PIPELINE_CONTINUE);
2709 }
2710
2711 /*
2712 * Called by RAID-Z to ensure we don't compute the checksum twice.
2713 */
2714 void
2715 zio_checksum_verified(zio_t *zio)
2716 {
2717 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2718 }
2719
2720 /*
2721 * ==========================================================================
2722 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2723 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2724 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2725 * indicate errors that are specific to one I/O, and most likely permanent.
2726 * Any other error is presumed to be worse because we weren't expecting it.
2727 * ==========================================================================
2728 */
2729 int
2730 zio_worst_error(int e1, int e2)
2731 {
2732 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2733 int r1, r2;
2734
2735 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2736 if (e1 == zio_error_rank[r1])
2737 break;
2738
2739 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2740 if (e2 == zio_error_rank[r2])
2741 break;
2742
2743 return (r1 > r2 ? e1 : e2);
2744 }
2745
2746 /*
2747 * ==========================================================================
2748 * I/O completion
2749 * ==========================================================================
2750 */
2751 static int
2752 zio_ready(zio_t *zio)
2753 {
2754 blkptr_t *bp = zio->io_bp;
2755 zio_t *pio, *pio_next;
2756
2757 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2758 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2759 return (ZIO_PIPELINE_STOP);
2760
2761 if (zio->io_ready) {
2762 ASSERT(IO_IS_ALLOCATING(zio));
2763 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
2764 (zio->io_flags & ZIO_FLAG_NOPWRITE));
2765 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2766
2767 zio->io_ready(zio);
2768 }
2769
2770 if (bp != NULL && bp != &zio->io_bp_copy)
2771 zio->io_bp_copy = *bp;
2772
2773 if (zio->io_error)
2774 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2775
2776 mutex_enter(&zio->io_lock);
2777 zio->io_state[ZIO_WAIT_READY] = 1;
2778 pio = zio_walk_parents(zio);
2779 mutex_exit(&zio->io_lock);
2780
2781 /*
2782 * As we notify zio's parents, new parents could be added.
2783 * New parents go to the head of zio's io_parent_list, however,
2784 * so we will (correctly) not notify them. The remainder of zio's
2785 * io_parent_list, from 'pio_next' onward, cannot change because
2786 * all parents must wait for us to be done before they can be done.
2787 */
2788 for (; pio != NULL; pio = pio_next) {
2789 pio_next = zio_walk_parents(zio);
2790 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2791 }
2792
2793 if (zio->io_flags & ZIO_FLAG_NODATA) {
2794 if (BP_IS_GANG(bp)) {
2795 zio->io_flags &= ~ZIO_FLAG_NODATA;
2796 } else {
2797 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2798 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2799 }
2800 }
2801
2802 if (zio_injection_enabled &&
2803 zio->io_spa->spa_syncing_txg == zio->io_txg)
2804 zio_handle_ignored_writes(zio);
2805
2806 return (ZIO_PIPELINE_CONTINUE);
2807 }
2808
2809 static int
2810 zio_done(zio_t *zio)
2811 {
2812 spa_t *spa = zio->io_spa;
2813 zio_t *lio = zio->io_logical;
2814 blkptr_t *bp = zio->io_bp;
2815 vdev_t *vd = zio->io_vd;
2816 uint64_t psize = zio->io_size;
2817 zio_t *pio, *pio_next;
2818
2819 /*
2820 * If our children haven't all completed,
2821 * wait for them and then repeat this pipeline stage.
2822 */
2823 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2824 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2825 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2826 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2827 return (ZIO_PIPELINE_STOP);
2828
2829 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2830 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2831 ASSERT(zio->io_children[c][w] == 0);
2832
2833 if (bp != NULL) {
2834 ASSERT(bp->blk_pad[0] == 0);
2835 ASSERT(bp->blk_pad[1] == 0);
2836 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2837 (bp == zio_unique_parent(zio)->io_bp));
2838 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2839 zio->io_bp_override == NULL &&
2840 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2841 ASSERT(!BP_SHOULD_BYTESWAP(bp));
2842 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2843 ASSERT(BP_COUNT_GANG(bp) == 0 ||
2844 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2845 }
2846 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
2847 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
2848 }
2849
2850 /*
2851 * If there were child vdev/gang/ddt errors, they apply to us now.
2852 */
2853 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2854 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2855 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2856
2857 /*
2858 * If the I/O on the transformed data was successful, generate any
2859 * checksum reports now while we still have the transformed data.
2860 */
2861 if (zio->io_error == 0) {
2862 while (zio->io_cksum_report != NULL) {
2863 zio_cksum_report_t *zcr = zio->io_cksum_report;
2864 uint64_t align = zcr->zcr_align;
2865 uint64_t asize = P2ROUNDUP(psize, align);
2866 char *abuf = zio->io_data;
2867
2868 if (asize != psize) {
2869 abuf = zio_buf_alloc(asize);
2870 bcopy(zio->io_data, abuf, psize);
2871 bzero(abuf + psize, asize - psize);
2872 }
2873
2874 zio->io_cksum_report = zcr->zcr_next;
2875 zcr->zcr_next = NULL;
2876 zcr->zcr_finish(zcr, abuf);
2877 zfs_ereport_free_checksum(zcr);
2878
2879 if (asize != psize)
2880 zio_buf_free(abuf, asize);
2881 }
2882 }
2883
2884 zio_pop_transforms(zio); /* note: may set zio->io_error */
2885
2886 vdev_stat_update(zio, psize);
2887
2888 if (zio->io_error) {
2889 /*
2890 * If this I/O is attached to a particular vdev,
2891 * generate an error message describing the I/O failure
2892 * at the block level. We ignore these errors if the
2893 * device is currently unavailable.
2894 */
2895 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
2896 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
2897
2898 if ((zio->io_error == EIO || !(zio->io_flags &
2899 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2900 zio == lio) {
2901 /*
2902 * For logical I/O requests, tell the SPA to log the
2903 * error and generate a logical data ereport.
2904 */
2905 spa_log_error(spa, zio);
2906 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
2907 0, 0);
2908 }
2909 }
2910
2911 if (zio->io_error && zio == lio) {
2912 /*
2913 * Determine whether zio should be reexecuted. This will
2914 * propagate all the way to the root via zio_notify_parent().
2915 */
2916 ASSERT(vd == NULL && bp != NULL);
2917 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2918
2919 if (IO_IS_ALLOCATING(zio) &&
2920 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2921 if (zio->io_error != ENOSPC)
2922 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2923 else
2924 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2925 }
2926
2927 if ((zio->io_type == ZIO_TYPE_READ ||
2928 zio->io_type == ZIO_TYPE_FREE) &&
2929 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2930 zio->io_error == ENXIO &&
2931 spa_load_state(spa) == SPA_LOAD_NONE &&
2932 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
2933 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2934
2935 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2936 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2937
2938 /*
2939 * Here is a possibly good place to attempt to do
2940 * either combinatorial reconstruction or error correction
2941 * based on checksums. It also might be a good place
2942 * to send out preliminary ereports before we suspend
2943 * processing.
2944 */
2945 }
2946
2947 /*
2948 * If there were logical child errors, they apply to us now.
2949 * We defer this until now to avoid conflating logical child
2950 * errors with errors that happened to the zio itself when
2951 * updating vdev stats and reporting FMA events above.
2952 */
2953 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2954
2955 if ((zio->io_error || zio->io_reexecute) &&
2956 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2957 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
2958 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
2959
2960 zio_gang_tree_free(&zio->io_gang_tree);
2961
2962 /*
2963 * Godfather I/Os should never suspend.
2964 */
2965 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2966 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2967 zio->io_reexecute = 0;
2968
2969 if (zio->io_reexecute) {
2970 /*
2971 * This is a logical I/O that wants to reexecute.
2972 *
2973 * Reexecute is top-down. When an i/o fails, if it's not
2974 * the root, it simply notifies its parent and sticks around.
2975 * The parent, seeing that it still has children in zio_done(),
2976 * does the same. This percolates all the way up to the root.
2977 * The root i/o will reexecute or suspend the entire tree.
2978 *
2979 * This approach ensures that zio_reexecute() honors
2980 * all the original i/o dependency relationships, e.g.
2981 * parents not executing until children are ready.
2982 */
2983 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2984
2985 zio->io_gang_leader = NULL;
2986
2987 mutex_enter(&zio->io_lock);
2988 zio->io_state[ZIO_WAIT_DONE] = 1;
2989 mutex_exit(&zio->io_lock);
2990
2991 /*
2992 * "The Godfather" I/O monitors its children but is
2993 * not a true parent to them. It will track them through
2994 * the pipeline but severs its ties whenever they get into
2995 * trouble (e.g. suspended). This allows "The Godfather"
2996 * I/O to return status without blocking.
2997 */
2998 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2999 zio_link_t *zl = zio->io_walk_link;
3000 pio_next = zio_walk_parents(zio);
3001
3002 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3003 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3004 zio_remove_child(pio, zio, zl);
3005 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3006 }
3007 }
3008
3009 if ((pio = zio_unique_parent(zio)) != NULL) {
3010 /*
3011 * We're not a root i/o, so there's nothing to do
3012 * but notify our parent. Don't propagate errors
3013 * upward since we haven't permanently failed yet.
3014 */
3015 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3016 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3017 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3018 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3019 /*
3020 * We'd fail again if we reexecuted now, so suspend
3021 * until conditions improve (e.g. device comes online).
3022 */
3023 zio_suspend(spa, zio);
3024 } else {
3025 /*
3026 * Reexecution is potentially a huge amount of work.
3027 * Hand it off to the otherwise-unused claim taskq.
3028 */
3029 ASSERT(zio->io_tqent.tqent_next == NULL);
3030 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3031 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3032 0, &zio->io_tqent);
3033 }
3034 return (ZIO_PIPELINE_STOP);
3035 }
3036
3037 ASSERT(zio->io_child_count == 0);
3038 ASSERT(zio->io_reexecute == 0);
3039 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3040
3041 /*
3042 * Report any checksum errors, since the I/O is complete.
3043 */
3044 while (zio->io_cksum_report != NULL) {
3045 zio_cksum_report_t *zcr = zio->io_cksum_report;
3046 zio->io_cksum_report = zcr->zcr_next;
3047 zcr->zcr_next = NULL;
3048 zcr->zcr_finish(zcr, NULL);
3049 zfs_ereport_free_checksum(zcr);
3050 }
3051
3052 /*
3053 * It is the responsibility of the done callback to ensure that this
3054 * particular zio is no longer discoverable for adoption, and as
3055 * such, cannot acquire any new parents.
3056 */
3057 if (zio->io_done)
3058 zio->io_done(zio);
3059
3060 mutex_enter(&zio->io_lock);
3061 zio->io_state[ZIO_WAIT_DONE] = 1;
3062 mutex_exit(&zio->io_lock);
3063
3064 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3065 zio_link_t *zl = zio->io_walk_link;
3066 pio_next = zio_walk_parents(zio);
3067 zio_remove_child(pio, zio, zl);
3068 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3069 }
3070
3071 if (zio->io_waiter != NULL) {
3072 mutex_enter(&zio->io_lock);
3073 zio->io_executor = NULL;
3074 cv_broadcast(&zio->io_cv);
3075 mutex_exit(&zio->io_lock);
3076 } else {
3077 zio_destroy(zio);
3078 }
3079
3080 return (ZIO_PIPELINE_STOP);
3081 }
3082
3083 /*
3084 * ==========================================================================
3085 * I/O pipeline definition
3086 * ==========================================================================
3087 */
3088 static zio_pipe_stage_t *zio_pipeline[] = {
3089 NULL,
3090 zio_read_bp_init,
3091 zio_free_bp_init,
3092 zio_issue_async,
3093 zio_write_bp_init,
3094 zio_checksum_generate,
3095 zio_nop_write,
3096 zio_ddt_read_start,
3097 zio_ddt_read_done,
3098 zio_ddt_write,
3099 zio_ddt_free,
3100 zio_gang_assemble,
3101 zio_gang_issue,
3102 zio_dva_allocate,
3103 zio_dva_free,
3104 zio_dva_claim,
3105 zio_ready,
3106 zio_vdev_io_start,
3107 zio_vdev_io_done,
3108 zio_vdev_io_assess,
3109 zio_checksum_verify,
3110 zio_done
3111 };
3112
3113 /* dnp is the dnode for zb1->zb_object */
3114 boolean_t
3115 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_t *zb1,
3116 const zbookmark_t *zb2)
3117 {
3118 uint64_t zb1nextL0, zb2thisobj;
3119
3120 ASSERT(zb1->zb_objset == zb2->zb_objset);
3121 ASSERT(zb2->zb_level == 0);
3122
3123 /*
3124 * A bookmark in the deadlist is considered to be after
3125 * everything else.
3126 */
3127 if (zb2->zb_object == DMU_DEADLIST_OBJECT)
3128 return (B_TRUE);
3129
3130 /* The objset_phys_t isn't before anything. */
3131 if (dnp == NULL)
3132 return (B_FALSE);
3133
3134 zb1nextL0 = (zb1->zb_blkid + 1) <<
3135 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3136
3137 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3138 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3139
3140 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3141 uint64_t nextobj = zb1nextL0 *
3142 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3143 return (nextobj <= zb2thisobj);
3144 }
3145
3146 if (zb1->zb_object < zb2thisobj)
3147 return (B_TRUE);
3148 if (zb1->zb_object > zb2thisobj)
3149 return (B_FALSE);
3150 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3151 return (B_FALSE);
3152 return (zb1nextL0 <= zb2->zb_blkid);
3153 }