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 *
1188 * (1) the I/O completes
1189 * (2) the pipeline stalls waiting for dependent child I/Os
1190 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1191 * (4) the I/O is delegated by vdev-level caching or aggregation
1192 * (5) the I/O is deferred due to vdev-level queueing
1193 * (6) the I/O is handed off to another thread.
1194 *
1195 * In all cases, the pipeline stops whenever there's no CPU work; it never
1196 * burns a thread in cv_wait().
1197 *
1198 * There's no locking on io_stage because there's no legitimate way
1199 * for multiple threads to be attempting to process the same I/O.
1200 */
1201 static zio_pipe_stage_t *zio_pipeline[];
1202
1203 void
1204 zio_execute(zio_t *zio)
1205 {
1206 zio->io_executor = curthread;
1207
1208 while (zio->io_stage < ZIO_STAGE_DONE) {
1209 enum zio_stage pipeline = zio->io_pipeline;
1210 enum zio_stage stage = zio->io_stage;
1211 int rv;
1212
1213 ASSERT(!MUTEX_HELD(&zio->io_lock));
1214 ASSERT(ISP2(stage));
1215 ASSERT(zio->io_stall == NULL);
1216
1217 do {
1218 stage <<= 1;
1219 } while ((stage & pipeline) == 0);
1220
1221 ASSERT(stage <= ZIO_STAGE_DONE);
1222
1223 /*
1224 * If we are in interrupt context and this pipeline stage
1225 * will grab a config lock that is held across I/O,
1226 * or may wait for an I/O that needs an interrupt thread
1227 * to complete, issue async to avoid deadlock.
1228 *
1229 * For VDEV_IO_START, we cut in line so that the io will
1230 * be sent to disk promptly.
1231 */
1232 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1233 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1234 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1235 zio_requeue_io_start_cut_in_line : B_FALSE;
1236 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1237 return;
1238 }
1239
1240 zio->io_stage = stage;
1241 rv = zio_pipeline[highbit(stage) - 1](zio);
1242
1243 if (rv == ZIO_PIPELINE_STOP)
1244 return;
1245
1246 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1247 }
1248 }
1249
1250 /*
1251 * ==========================================================================
1252 * Initiate I/O, either sync or async
1253 * ==========================================================================
1254 */
1255 int
1256 zio_wait(zio_t *zio)
1257 {
1258 int error;
1259
1260 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1261 ASSERT(zio->io_executor == NULL);
1262
1263 zio->io_waiter = curthread;
1264
1265 zio_execute(zio);
1266
1267 mutex_enter(&zio->io_lock);
1268 while (zio->io_executor != NULL)
1269 cv_wait(&zio->io_cv, &zio->io_lock);
1270 mutex_exit(&zio->io_lock);
1271
1272 error = zio->io_error;
1273 zio_destroy(zio);
1274
1275 return (error);
1276 }
1277
1278 void
1279 zio_nowait(zio_t *zio)
1280 {
1281 ASSERT(zio->io_executor == NULL);
1282
1283 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1284 zio_unique_parent(zio) == NULL) {
1285 /*
1286 * This is a logical async I/O with no parent to wait for it.
1287 * We add it to the spa_async_root_zio "Godfather" I/O which
1288 * will ensure they complete prior to unloading the pool.
1289 */
1290 spa_t *spa = zio->io_spa;
1291
1292 zio_add_child(spa->spa_async_zio_root, zio);
1293 }
1294
1295 zio_execute(zio);
1296 }
1297
1298 /*
1299 * ==========================================================================
1300 * Reexecute or suspend/resume failed I/O
1301 * ==========================================================================
1302 */
1303
1304 static void
1305 zio_reexecute(zio_t *pio)
1306 {
1307 zio_t *cio, *cio_next;
1308
1309 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1310 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1311 ASSERT(pio->io_gang_leader == NULL);
1312 ASSERT(pio->io_gang_tree == NULL);
1313
1314 pio->io_flags = pio->io_orig_flags;
1315 pio->io_stage = pio->io_orig_stage;
1316 pio->io_pipeline = pio->io_orig_pipeline;
1317 pio->io_reexecute = 0;
1318 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1319 pio->io_error = 0;
1320 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1321 pio->io_state[w] = 0;
1322 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1323 pio->io_child_error[c] = 0;
1324
1325 if (IO_IS_ALLOCATING(pio))
1326 BP_ZERO(pio->io_bp);
1327
1328 /*
1329 * As we reexecute pio's children, new children could be created.
1330 * New children go to the head of pio's io_child_list, however,
1331 * so we will (correctly) not reexecute them. The key is that
1332 * the remainder of pio's io_child_list, from 'cio_next' onward,
1333 * cannot be affected by any side effects of reexecuting 'cio'.
1334 */
1335 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1336 cio_next = zio_walk_children(pio);
1337 mutex_enter(&pio->io_lock);
1338 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1339 pio->io_children[cio->io_child_type][w]++;
1340 mutex_exit(&pio->io_lock);
1341 zio_reexecute(cio);
1342 }
1343
1344 /*
1345 * Now that all children have been reexecuted, execute the parent.
1346 * We don't reexecute "The Godfather" I/O here as it's the
1347 * responsibility of the caller to wait on him.
1348 */
1349 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1350 zio_execute(pio);
1351 }
1352
1353 void
1354 zio_suspend(spa_t *spa, zio_t *zio)
1355 {
1356 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1357 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1358 "failure and the failure mode property for this pool "
1359 "is set to panic.", spa_name(spa));
1360
1361 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1362
1363 mutex_enter(&spa->spa_suspend_lock);
1364
1365 if (spa->spa_suspend_zio_root == NULL)
1366 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1367 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1368 ZIO_FLAG_GODFATHER);
1369
1370 spa->spa_suspended = B_TRUE;
1371
1372 if (zio != NULL) {
1373 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1374 ASSERT(zio != spa->spa_suspend_zio_root);
1375 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1376 ASSERT(zio_unique_parent(zio) == NULL);
1377 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1378 zio_add_child(spa->spa_suspend_zio_root, zio);
1379 }
1380
1381 mutex_exit(&spa->spa_suspend_lock);
1382 }
1383
1384 int
1385 zio_resume(spa_t *spa)
1386 {
1387 zio_t *pio;
1388
1389 /*
1390 * Reexecute all previously suspended i/o.
1391 */
1392 mutex_enter(&spa->spa_suspend_lock);
1393 spa->spa_suspended = B_FALSE;
1394 cv_broadcast(&spa->spa_suspend_cv);
1395 pio = spa->spa_suspend_zio_root;
1396 spa->spa_suspend_zio_root = NULL;
1397 mutex_exit(&spa->spa_suspend_lock);
1398
1399 if (pio == NULL)
1400 return (0);
1401
1402 zio_reexecute(pio);
1403 return (zio_wait(pio));
1404 }
1405
1406 void
1407 zio_resume_wait(spa_t *spa)
1408 {
1409 mutex_enter(&spa->spa_suspend_lock);
1410 while (spa_suspended(spa))
1411 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1412 mutex_exit(&spa->spa_suspend_lock);
1413 }
1414
1415 /*
1416 * ==========================================================================
1417 * Gang blocks.
1418 *
1419 * A gang block is a collection of small blocks that looks to the DMU
1420 * like one large block. When zio_dva_allocate() cannot find a block
1421 * of the requested size, due to either severe fragmentation or the pool
1422 * being nearly full, it calls zio_write_gang_block() to construct the
1423 * block from smaller fragments.
1424 *
1425 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1426 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1427 * an indirect block: it's an array of block pointers. It consumes
1428 * only one sector and hence is allocatable regardless of fragmentation.
1429 * The gang header's bps point to its gang members, which hold the data.
1430 *
1431 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1432 * as the verifier to ensure uniqueness of the SHA256 checksum.
1433 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1434 * not the gang header. This ensures that data block signatures (needed for
1435 * deduplication) are independent of how the block is physically stored.
1436 *
1437 * Gang blocks can be nested: a gang member may itself be a gang block.
1438 * Thus every gang block is a tree in which root and all interior nodes are
1439 * gang headers, and the leaves are normal blocks that contain user data.
1440 * The root of the gang tree is called the gang leader.
1441 *
1442 * To perform any operation (read, rewrite, free, claim) on a gang block,
1443 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1444 * in the io_gang_tree field of the original logical i/o by recursively
1445 * reading the gang leader and all gang headers below it. This yields
1446 * an in-core tree containing the contents of every gang header and the
1447 * bps for every constituent of the gang block.
1448 *
1449 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1450 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1451 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1452 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1453 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1454 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1455 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1456 * of the gang header plus zio_checksum_compute() of the data to update the
1457 * gang header's blk_cksum as described above.
1458 *
1459 * The two-phase assemble/issue model solves the problem of partial failure --
1460 * what if you'd freed part of a gang block but then couldn't read the
1461 * gang header for another part? Assembling the entire gang tree first
1462 * ensures that all the necessary gang header I/O has succeeded before
1463 * starting the actual work of free, claim, or write. Once the gang tree
1464 * is assembled, free and claim are in-memory operations that cannot fail.
1465 *
1466 * In the event that a gang write fails, zio_dva_unallocate() walks the
1467 * gang tree to immediately free (i.e. insert back into the space map)
1468 * everything we've allocated. This ensures that we don't get ENOSPC
1469 * errors during repeated suspend/resume cycles due to a flaky device.
1470 *
1471 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1472 * the gang tree, we won't modify the block, so we can safely defer the free
1473 * (knowing that the block is still intact). If we *can* assemble the gang
1474 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1475 * each constituent bp and we can allocate a new block on the next sync pass.
1476 *
1477 * In all cases, the gang tree allows complete recovery from partial failure.
1478 * ==========================================================================
1479 */
1480
1481 static zio_t *
1482 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1483 {
1484 if (gn != NULL)
1485 return (pio);
1486
1487 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1488 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1489 &pio->io_bookmark));
1490 }
1491
1492 zio_t *
1493 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1494 {
1495 zio_t *zio;
1496
1497 if (gn != NULL) {
1498 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1499 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1500 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1501 /*
1502 * As we rewrite each gang header, the pipeline will compute
1503 * a new gang block header checksum for it; but no one will
1504 * compute a new data checksum, so we do that here. The one
1505 * exception is the gang leader: the pipeline already computed
1506 * its data checksum because that stage precedes gang assembly.
1507 * (Presently, nothing actually uses interior data checksums;
1508 * this is just good hygiene.)
1509 */
1510 if (gn != pio->io_gang_leader->io_gang_tree) {
1511 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1512 data, BP_GET_PSIZE(bp));
1513 }
1514 /*
1515 * If we are here to damage data for testing purposes,
1516 * leave the GBH alone so that we can detect the damage.
1517 */
1518 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1519 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1520 } else {
1521 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1522 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1523 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1524 }
1525
1526 return (zio);
1527 }
1528
1529 /* ARGSUSED */
1530 zio_t *
1531 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1532 {
1533 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1534 ZIO_GANG_CHILD_FLAGS(pio)));
1535 }
1536
1537 /* ARGSUSED */
1538 zio_t *
1539 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1540 {
1541 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1542 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1543 }
1544
1545 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1546 NULL,
1547 zio_read_gang,
1548 zio_rewrite_gang,
1549 zio_free_gang,
1550 zio_claim_gang,
1551 NULL
1552 };
1553
1554 static void zio_gang_tree_assemble_done(zio_t *zio);
1555
1556 static zio_gang_node_t *
1557 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1558 {
1559 zio_gang_node_t *gn;
1560
1561 ASSERT(*gnpp == NULL);
1562
1563 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1564 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1565 *gnpp = gn;
1566
1567 return (gn);
1568 }
1569
1570 static void
1571 zio_gang_node_free(zio_gang_node_t **gnpp)
1572 {
1573 zio_gang_node_t *gn = *gnpp;
1574
1575 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1576 ASSERT(gn->gn_child[g] == NULL);
1577
1578 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1579 kmem_free(gn, sizeof (*gn));
1580 *gnpp = NULL;
1581 }
1582
1583 static void
1584 zio_gang_tree_free(zio_gang_node_t **gnpp)
1585 {
1586 zio_gang_node_t *gn = *gnpp;
1587
1588 if (gn == NULL)
1589 return;
1590
1591 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1592 zio_gang_tree_free(&gn->gn_child[g]);
1593
1594 zio_gang_node_free(gnpp);
1595 }
1596
1597 static void
1598 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1599 {
1600 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1601
1602 ASSERT(gio->io_gang_leader == gio);
1603 ASSERT(BP_IS_GANG(bp));
1604
1605 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1606 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1607 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1608 }
1609
1610 static void
1611 zio_gang_tree_assemble_done(zio_t *zio)
1612 {
1613 zio_t *gio = zio->io_gang_leader;
1614 zio_gang_node_t *gn = zio->io_private;
1615 blkptr_t *bp = zio->io_bp;
1616
1617 ASSERT(gio == zio_unique_parent(zio));
1618 ASSERT(zio->io_child_count == 0);
1619
1620 if (zio->io_error)
1621 return;
1622
1623 if (BP_SHOULD_BYTESWAP(bp))
1624 byteswap_uint64_array(zio->io_data, zio->io_size);
1625
1626 ASSERT(zio->io_data == gn->gn_gbh);
1627 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1628 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1629
1630 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1631 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1632 if (!BP_IS_GANG(gbp))
1633 continue;
1634 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1635 }
1636 }
1637
1638 static void
1639 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1640 {
1641 zio_t *gio = pio->io_gang_leader;
1642 zio_t *zio;
1643
1644 ASSERT(BP_IS_GANG(bp) == !!gn);
1645 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1646 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1647
1648 /*
1649 * If you're a gang header, your data is in gn->gn_gbh.
1650 * If you're a gang member, your data is in 'data' and gn == NULL.
1651 */
1652 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1653
1654 if (gn != NULL) {
1655 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1656
1657 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1658 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1659 if (BP_IS_HOLE(gbp))
1660 continue;
1661 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1662 data = (char *)data + BP_GET_PSIZE(gbp);
1663 }
1664 }
1665
1666 if (gn == gio->io_gang_tree)
1667 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1668
1669 if (zio != pio)
1670 zio_nowait(zio);
1671 }
1672
1673 static int
1674 zio_gang_assemble(zio_t *zio)
1675 {
1676 blkptr_t *bp = zio->io_bp;
1677
1678 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1679 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1680
1681 zio->io_gang_leader = zio;
1682
1683 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1684
1685 return (ZIO_PIPELINE_CONTINUE);
1686 }
1687
1688 static int
1689 zio_gang_issue(zio_t *zio)
1690 {
1691 blkptr_t *bp = zio->io_bp;
1692
1693 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1694 return (ZIO_PIPELINE_STOP);
1695
1696 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1697 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1698
1699 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1700 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1701 else
1702 zio_gang_tree_free(&zio->io_gang_tree);
1703
1704 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1705
1706 return (ZIO_PIPELINE_CONTINUE);
1707 }
1708
1709 static void
1710 zio_write_gang_member_ready(zio_t *zio)
1711 {
1712 zio_t *pio = zio_unique_parent(zio);
1713 zio_t *gio = zio->io_gang_leader;
1714 dva_t *cdva = zio->io_bp->blk_dva;
1715 dva_t *pdva = pio->io_bp->blk_dva;
1716 uint64_t asize;
1717
1718 if (BP_IS_HOLE(zio->io_bp))
1719 return;
1720
1721 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1722
1723 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1724 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1725 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1726 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1727 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1728
1729 mutex_enter(&pio->io_lock);
1730 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1731 ASSERT(DVA_GET_GANG(&pdva[d]));
1732 asize = DVA_GET_ASIZE(&pdva[d]);
1733 asize += DVA_GET_ASIZE(&cdva[d]);
1734 DVA_SET_ASIZE(&pdva[d], asize);
1735 }
1736 mutex_exit(&pio->io_lock);
1737 }
1738
1739 static int
1740 zio_write_gang_block(zio_t *pio)
1741 {
1742 spa_t *spa = pio->io_spa;
1743 blkptr_t *bp = pio->io_bp;
1744 zio_t *gio = pio->io_gang_leader;
1745 zio_t *zio;
1746 zio_gang_node_t *gn, **gnpp;
1747 zio_gbh_phys_t *gbh;
1748 uint64_t txg = pio->io_txg;
1749 uint64_t resid = pio->io_size;
1750 uint64_t lsize;
1751 int copies = gio->io_prop.zp_copies;
1752 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1753 zio_prop_t zp;
1754 int error;
1755
1756 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1757 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1758 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1759 if (error) {
1760 pio->io_error = error;
1761 return (ZIO_PIPELINE_CONTINUE);
1762 }
1763
1764 if (pio == gio) {
1765 gnpp = &gio->io_gang_tree;
1766 } else {
1767 gnpp = pio->io_private;
1768 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1769 }
1770
1771 gn = zio_gang_node_alloc(gnpp);
1772 gbh = gn->gn_gbh;
1773 bzero(gbh, SPA_GANGBLOCKSIZE);
1774
1775 /*
1776 * Create the gang header.
1777 */
1778 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1779 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1780
1781 /*
1782 * Create and nowait the gang children.
1783 */
1784 for (int g = 0; resid != 0; resid -= lsize, g++) {
1785 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1786 SPA_MINBLOCKSIZE);
1787 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1788
1789 zp.zp_checksum = gio->io_prop.zp_checksum;
1790 zp.zp_compress = ZIO_COMPRESS_OFF;
1791 zp.zp_type = DMU_OT_NONE;
1792 zp.zp_level = 0;
1793 zp.zp_copies = gio->io_prop.zp_copies;
1794 zp.zp_dedup = B_FALSE;
1795 zp.zp_dedup_verify = B_FALSE;
1796 zp.zp_nopwrite = B_FALSE;
1797
1798 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1799 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1800 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1801 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1802 &pio->io_bookmark));
1803 }
1804
1805 /*
1806 * Set pio's pipeline to just wait for zio to finish.
1807 */
1808 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1809
1810 zio_nowait(zio);
1811
1812 return (ZIO_PIPELINE_CONTINUE);
1813 }
1814
1815 /*
1816 * The zio_nop_write stage in the pipeline determines if allocating
1817 * a new bp is necessary. By leveraging a cryptographically secure checksum,
1818 * such as SHA256, we can compare the checksums of the new data and the old
1819 * to determine if allocating a new block is required. The nopwrite
1820 * feature can handle writes in either syncing or open context (i.e. zil
1821 * writes) and as a result is mutually exclusive with dedup.
1822 */
1823 static int
1824 zio_nop_write(zio_t *zio)
1825 {
1826 blkptr_t *bp = zio->io_bp;
1827 blkptr_t *bp_orig = &zio->io_bp_orig;
1828 zio_prop_t *zp = &zio->io_prop;
1829
1830 ASSERT(BP_GET_LEVEL(bp) == 0);
1831 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1832 ASSERT(zp->zp_nopwrite);
1833 ASSERT(!zp->zp_dedup);
1834 ASSERT(zio->io_bp_override == NULL);
1835 ASSERT(IO_IS_ALLOCATING(zio));
1836
1837 /*
1838 * Check to see if the original bp and the new bp have matching
1839 * characteristics (i.e. same checksum, compression algorithms, etc).
1840 * If they don't then just continue with the pipeline which will
1841 * allocate a new bp.
1842 */
1843 if (BP_IS_HOLE(bp_orig) ||
1844 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
1845 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
1846 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
1847 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
1848 zp->zp_copies != BP_GET_NDVAS(bp_orig))
1849 return (ZIO_PIPELINE_CONTINUE);
1850
1851 /*
1852 * If the checksums match then reset the pipeline so that we
1853 * avoid allocating a new bp and issuing any I/O.
1854 */
1855 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
1856 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
1857 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
1858 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
1859 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
1860 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
1861 sizeof (uint64_t)) == 0);
1862
1863 *bp = *bp_orig;
1864 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1865 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1866 }
1867
1868 return (ZIO_PIPELINE_CONTINUE);
1869 }
1870
1871 /*
1872 * ==========================================================================
1873 * Dedup
1874 * ==========================================================================
1875 */
1876 static void
1877 zio_ddt_child_read_done(zio_t *zio)
1878 {
1879 blkptr_t *bp = zio->io_bp;
1880 ddt_entry_t *dde = zio->io_private;
1881 ddt_phys_t *ddp;
1882 zio_t *pio = zio_unique_parent(zio);
1883
1884 mutex_enter(&pio->io_lock);
1885 ddp = ddt_phys_select(dde, bp);
1886 if (zio->io_error == 0)
1887 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1888 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1889 dde->dde_repair_data = zio->io_data;
1890 else
1891 zio_buf_free(zio->io_data, zio->io_size);
1892 mutex_exit(&pio->io_lock);
1893 }
1894
1895 static int
1896 zio_ddt_read_start(zio_t *zio)
1897 {
1898 blkptr_t *bp = zio->io_bp;
1899
1900 ASSERT(BP_GET_DEDUP(bp));
1901 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1902 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1903
1904 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1905 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1906 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1907 ddt_phys_t *ddp = dde->dde_phys;
1908 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1909 blkptr_t blk;
1910
1911 ASSERT(zio->io_vsd == NULL);
1912 zio->io_vsd = dde;
1913
1914 if (ddp_self == NULL)
1915 return (ZIO_PIPELINE_CONTINUE);
1916
1917 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1918 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1919 continue;
1920 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1921 &blk);
1922 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1923 zio_buf_alloc(zio->io_size), zio->io_size,
1924 zio_ddt_child_read_done, dde, zio->io_priority,
1925 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1926 &zio->io_bookmark));
1927 }
1928 return (ZIO_PIPELINE_CONTINUE);
1929 }
1930
1931 zio_nowait(zio_read(zio, zio->io_spa, bp,
1932 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1933 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1934
1935 return (ZIO_PIPELINE_CONTINUE);
1936 }
1937
1938 static int
1939 zio_ddt_read_done(zio_t *zio)
1940 {
1941 blkptr_t *bp = zio->io_bp;
1942
1943 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1944 return (ZIO_PIPELINE_STOP);
1945
1946 ASSERT(BP_GET_DEDUP(bp));
1947 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1948 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1949
1950 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1951 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1952 ddt_entry_t *dde = zio->io_vsd;
1953 if (ddt == NULL) {
1954 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1955 return (ZIO_PIPELINE_CONTINUE);
1956 }
1957 if (dde == NULL) {
1958 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1959 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1960 return (ZIO_PIPELINE_STOP);
1961 }
1962 if (dde->dde_repair_data != NULL) {
1963 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1964 zio->io_child_error[ZIO_CHILD_DDT] = 0;
1965 }
1966 ddt_repair_done(ddt, dde);
1967 zio->io_vsd = NULL;
1968 }
1969
1970 ASSERT(zio->io_vsd == NULL);
1971
1972 return (ZIO_PIPELINE_CONTINUE);
1973 }
1974
1975 static boolean_t
1976 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1977 {
1978 spa_t *spa = zio->io_spa;
1979
1980 /*
1981 * Note: we compare the original data, not the transformed data,
1982 * because when zio->io_bp is an override bp, we will not have
1983 * pushed the I/O transforms. That's an important optimization
1984 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1985 */
1986 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1987 zio_t *lio = dde->dde_lead_zio[p];
1988
1989 if (lio != NULL) {
1990 return (lio->io_orig_size != zio->io_orig_size ||
1991 bcmp(zio->io_orig_data, lio->io_orig_data,
1992 zio->io_orig_size) != 0);
1993 }
1994 }
1995
1996 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1997 ddt_phys_t *ddp = &dde->dde_phys[p];
1998
1999 if (ddp->ddp_phys_birth != 0) {
2000 arc_buf_t *abuf = NULL;
2001 uint32_t aflags = ARC_WAIT;
2002 blkptr_t blk = *zio->io_bp;
2003 int error;
2004
2005 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2006
2007 ddt_exit(ddt);
2008
2009 error = arc_read(NULL, spa, &blk,
2010 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2011 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2012 &aflags, &zio->io_bookmark);
2013
2014 if (error == 0) {
2015 if (arc_buf_size(abuf) != zio->io_orig_size ||
2016 bcmp(abuf->b_data, zio->io_orig_data,
2017 zio->io_orig_size) != 0)
2018 error = SET_ERROR(EEXIST);
2019 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2020 }
2021
2022 ddt_enter(ddt);
2023 return (error != 0);
2024 }
2025 }
2026
2027 return (B_FALSE);
2028 }
2029
2030 static void
2031 zio_ddt_child_write_ready(zio_t *zio)
2032 {
2033 int p = zio->io_prop.zp_copies;
2034 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2035 ddt_entry_t *dde = zio->io_private;
2036 ddt_phys_t *ddp = &dde->dde_phys[p];
2037 zio_t *pio;
2038
2039 if (zio->io_error)
2040 return;
2041
2042 ddt_enter(ddt);
2043
2044 ASSERT(dde->dde_lead_zio[p] == zio);
2045
2046 ddt_phys_fill(ddp, zio->io_bp);
2047
2048 while ((pio = zio_walk_parents(zio)) != NULL)
2049 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2050
2051 ddt_exit(ddt);
2052 }
2053
2054 static void
2055 zio_ddt_child_write_done(zio_t *zio)
2056 {
2057 int p = zio->io_prop.zp_copies;
2058 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2059 ddt_entry_t *dde = zio->io_private;
2060 ddt_phys_t *ddp = &dde->dde_phys[p];
2061
2062 ddt_enter(ddt);
2063
2064 ASSERT(ddp->ddp_refcnt == 0);
2065 ASSERT(dde->dde_lead_zio[p] == zio);
2066 dde->dde_lead_zio[p] = NULL;
2067
2068 if (zio->io_error == 0) {
2069 while (zio_walk_parents(zio) != NULL)
2070 ddt_phys_addref(ddp);
2071 } else {
2072 ddt_phys_clear(ddp);
2073 }
2074
2075 ddt_exit(ddt);
2076 }
2077
2078 static void
2079 zio_ddt_ditto_write_done(zio_t *zio)
2080 {
2081 int p = DDT_PHYS_DITTO;
2082 zio_prop_t *zp = &zio->io_prop;
2083 blkptr_t *bp = zio->io_bp;
2084 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2085 ddt_entry_t *dde = zio->io_private;
2086 ddt_phys_t *ddp = &dde->dde_phys[p];
2087 ddt_key_t *ddk = &dde->dde_key;
2088
2089 ddt_enter(ddt);
2090
2091 ASSERT(ddp->ddp_refcnt == 0);
2092 ASSERT(dde->dde_lead_zio[p] == zio);
2093 dde->dde_lead_zio[p] = NULL;
2094
2095 if (zio->io_error == 0) {
2096 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2097 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2098 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2099 if (ddp->ddp_phys_birth != 0)
2100 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2101 ddt_phys_fill(ddp, bp);
2102 }
2103
2104 ddt_exit(ddt);
2105 }
2106
2107 static int
2108 zio_ddt_write(zio_t *zio)
2109 {
2110 spa_t *spa = zio->io_spa;
2111 blkptr_t *bp = zio->io_bp;
2112 uint64_t txg = zio->io_txg;
2113 zio_prop_t *zp = &zio->io_prop;
2114 int p = zp->zp_copies;
2115 int ditto_copies;
2116 zio_t *cio = NULL;
2117 zio_t *dio = NULL;
2118 ddt_t *ddt = ddt_select(spa, bp);
2119 ddt_entry_t *dde;
2120 ddt_phys_t *ddp;
2121
2122 ASSERT(BP_GET_DEDUP(bp));
2123 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2124 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2125
2126 ddt_enter(ddt);
2127 dde = ddt_lookup(ddt, bp, B_TRUE);
2128 ddp = &dde->dde_phys[p];
2129
2130 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2131 /*
2132 * If we're using a weak checksum, upgrade to a strong checksum
2133 * and try again. If we're already using a strong checksum,
2134 * we can't resolve it, so just convert to an ordinary write.
2135 * (And automatically e-mail a paper to Nature?)
2136 */
2137 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2138 zp->zp_checksum = spa_dedup_checksum(spa);
2139 zio_pop_transforms(zio);
2140 zio->io_stage = ZIO_STAGE_OPEN;
2141 BP_ZERO(bp);
2142 } else {
2143 zp->zp_dedup = B_FALSE;
2144 }
2145 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2146 ddt_exit(ddt);
2147 return (ZIO_PIPELINE_CONTINUE);
2148 }
2149
2150 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2151 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2152
2153 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2154 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2155 zio_prop_t czp = *zp;
2156
2157 czp.zp_copies = ditto_copies;
2158
2159 /*
2160 * If we arrived here with an override bp, we won't have run
2161 * the transform stack, so we won't have the data we need to
2162 * generate a child i/o. So, toss the override bp and restart.
2163 * This is safe, because using the override bp is just an
2164 * optimization; and it's rare, so the cost doesn't matter.
2165 */
2166 if (zio->io_bp_override) {
2167 zio_pop_transforms(zio);
2168 zio->io_stage = ZIO_STAGE_OPEN;
2169 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2170 zio->io_bp_override = NULL;
2171 BP_ZERO(bp);
2172 ddt_exit(ddt);
2173 return (ZIO_PIPELINE_CONTINUE);
2174 }
2175
2176 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2177 zio->io_orig_size, &czp, NULL,
2178 zio_ddt_ditto_write_done, dde, zio->io_priority,
2179 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2180
2181 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2182 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2183 }
2184
2185 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2186 if (ddp->ddp_phys_birth != 0)
2187 ddt_bp_fill(ddp, bp, txg);
2188 if (dde->dde_lead_zio[p] != NULL)
2189 zio_add_child(zio, dde->dde_lead_zio[p]);
2190 else
2191 ddt_phys_addref(ddp);
2192 } else if (zio->io_bp_override) {
2193 ASSERT(bp->blk_birth == txg);
2194 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2195 ddt_phys_fill(ddp, bp);
2196 ddt_phys_addref(ddp);
2197 } else {
2198 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2199 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2200 zio_ddt_child_write_done, dde, zio->io_priority,
2201 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2202
2203 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2204 dde->dde_lead_zio[p] = cio;
2205 }
2206
2207 ddt_exit(ddt);
2208
2209 if (cio)
2210 zio_nowait(cio);
2211 if (dio)
2212 zio_nowait(dio);
2213
2214 return (ZIO_PIPELINE_CONTINUE);
2215 }
2216
2217 ddt_entry_t *freedde; /* for debugging */
2218
2219 static int
2220 zio_ddt_free(zio_t *zio)
2221 {
2222 spa_t *spa = zio->io_spa;
2223 blkptr_t *bp = zio->io_bp;
2224 ddt_t *ddt = ddt_select(spa, bp);
2225 ddt_entry_t *dde;
2226 ddt_phys_t *ddp;
2227
2228 ASSERT(BP_GET_DEDUP(bp));
2229 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2230
2231 ddt_enter(ddt);
2232 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2233 ddp = ddt_phys_select(dde, bp);
2234 ddt_phys_decref(ddp);
2235 ddt_exit(ddt);
2236
2237 return (ZIO_PIPELINE_CONTINUE);
2238 }
2239
2240 /*
2241 * ==========================================================================
2242 * Allocate and free blocks
2243 * ==========================================================================
2244 */
2245 static int
2246 zio_dva_allocate(zio_t *zio)
2247 {
2248 spa_t *spa = zio->io_spa;
2249 metaslab_class_t *mc = spa_normal_class(spa);
2250 blkptr_t *bp = zio->io_bp;
2251 int error;
2252 int flags = 0;
2253
2254 if (zio->io_gang_leader == NULL) {
2255 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2256 zio->io_gang_leader = zio;
2257 }
2258
2259 ASSERT(BP_IS_HOLE(bp));
2260 ASSERT0(BP_GET_NDVAS(bp));
2261 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2262 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2263 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2264
2265 /*
2266 * The dump device does not support gang blocks so allocation on
2267 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2268 * the "fast" gang feature.
2269 */
2270 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2271 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2272 METASLAB_GANG_CHILD : 0;
2273 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2274 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2275
2276 if (error) {
2277 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2278 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2279 error);
2280 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2281 return (zio_write_gang_block(zio));
2282 zio->io_error = error;
2283 }
2284
2285 return (ZIO_PIPELINE_CONTINUE);
2286 }
2287
2288 static int
2289 zio_dva_free(zio_t *zio)
2290 {
2291 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2292
2293 return (ZIO_PIPELINE_CONTINUE);
2294 }
2295
2296 static int
2297 zio_dva_claim(zio_t *zio)
2298 {
2299 int error;
2300
2301 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2302 if (error)
2303 zio->io_error = error;
2304
2305 return (ZIO_PIPELINE_CONTINUE);
2306 }
2307
2308 /*
2309 * Undo an allocation. This is used by zio_done() when an I/O fails
2310 * and we want to give back the block we just allocated.
2311 * This handles both normal blocks and gang blocks.
2312 */
2313 static void
2314 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2315 {
2316 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2317 ASSERT(zio->io_bp_override == NULL);
2318
2319 if (!BP_IS_HOLE(bp))
2320 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2321
2322 if (gn != NULL) {
2323 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2324 zio_dva_unallocate(zio, gn->gn_child[g],
2325 &gn->gn_gbh->zg_blkptr[g]);
2326 }
2327 }
2328 }
2329
2330 /*
2331 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2332 */
2333 int
2334 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2335 uint64_t size, boolean_t use_slog)
2336 {
2337 int error = 1;
2338
2339 ASSERT(txg > spa_syncing_txg(spa));
2340
2341 /*
2342 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2343 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2344 * when allocating them.
2345 */
2346 if (use_slog) {
2347 error = metaslab_alloc(spa, spa_log_class(spa), size,
2348 new_bp, 1, txg, old_bp,
2349 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2350 }
2351
2352 if (error) {
2353 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2354 new_bp, 1, txg, old_bp,
2355 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2356 }
2357
2358 if (error == 0) {
2359 BP_SET_LSIZE(new_bp, size);
2360 BP_SET_PSIZE(new_bp, size);
2361 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2362 BP_SET_CHECKSUM(new_bp,
2363 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2364 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2365 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2366 BP_SET_LEVEL(new_bp, 0);
2367 BP_SET_DEDUP(new_bp, 0);
2368 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2369 }
2370
2371 return (error);
2372 }
2373
2374 /*
2375 * Free an intent log block.
2376 */
2377 void
2378 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2379 {
2380 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2381 ASSERT(!BP_IS_GANG(bp));
2382
2383 zio_free(spa, txg, bp);
2384 }
2385
2386 /*
2387 * ==========================================================================
2388 * Read and write to physical devices
2389 * ==========================================================================
2390 */
2391 static int
2392 zio_vdev_io_start(zio_t *zio)
2393 {
2394 vdev_t *vd = zio->io_vd;
2395 uint64_t align;
2396 spa_t *spa = zio->io_spa;
2397
2398 ASSERT(zio->io_error == 0);
2399 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2400
2401 if (vd == NULL) {
2402 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2403 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2404
2405 /*
2406 * The mirror_ops handle multiple DVAs in a single BP.
2407 */
2408 return (vdev_mirror_ops.vdev_op_io_start(zio));
2409 }
2410
2411 /*
2412 * We keep track of time-sensitive I/Os so that the scan thread
2413 * can quickly react to certain workloads. In particular, we care
2414 * about non-scrubbing, top-level reads and writes with the following
2415 * characteristics:
2416 * - synchronous writes of user data to non-slog devices
2417 * - any reads of user data
2418 * When these conditions are met, adjust the timestamp of spa_last_io
2419 * which allows the scan thread to adjust its workload accordingly.
2420 */
2421 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2422 vd == vd->vdev_top && !vd->vdev_islog &&
2423 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2424 zio->io_txg != spa_syncing_txg(spa)) {
2425 uint64_t old = spa->spa_last_io;
2426 uint64_t new = ddi_get_lbolt64();
2427 if (old != new)
2428 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2429 }
2430
2431 align = 1ULL << vd->vdev_top->vdev_ashift;
2432
2433 if (P2PHASE(zio->io_size, align) != 0) {
2434 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2435 char *abuf = zio_buf_alloc(asize);
2436 ASSERT(vd == vd->vdev_top);
2437 if (zio->io_type == ZIO_TYPE_WRITE) {
2438 bcopy(zio->io_data, abuf, zio->io_size);
2439 bzero(abuf + zio->io_size, asize - zio->io_size);
2440 }
2441 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2442 }
2443
2444 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2445 ASSERT(P2PHASE(zio->io_size, align) == 0);
2446 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2447
2448 /*
2449 * If this is a repair I/O, and there's no self-healing involved --
2450 * that is, we're just resilvering what we expect to resilver --
2451 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2452 * This prevents spurious resilvering with nested replication.
2453 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2454 * A is out of date, we'll read from C+D, then use the data to
2455 * resilver A+B -- but we don't actually want to resilver B, just A.
2456 * The top-level mirror has no way to know this, so instead we just
2457 * discard unnecessary repairs as we work our way down the vdev tree.
2458 * The same logic applies to any form of nested replication:
2459 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2460 */
2461 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2462 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2463 zio->io_txg != 0 && /* not a delegated i/o */
2464 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2465 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2466 zio_vdev_io_bypass(zio);
2467 return (ZIO_PIPELINE_CONTINUE);
2468 }
2469
2470 if (vd->vdev_ops->vdev_op_leaf &&
2471 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2472
2473 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2474 return (ZIO_PIPELINE_CONTINUE);
2475
2476 if ((zio = vdev_queue_io(zio)) == NULL)
2477 return (ZIO_PIPELINE_STOP);
2478
2479 if (!vdev_accessible(vd, zio)) {
2480 zio->io_error = SET_ERROR(ENXIO);
2481 zio_interrupt(zio);
2482 return (ZIO_PIPELINE_STOP);
2483 }
2484 }
2485
2486 return (vd->vdev_ops->vdev_op_io_start(zio));
2487 }
2488
2489 static int
2490 zio_vdev_io_done(zio_t *zio)
2491 {
2492 vdev_t *vd = zio->io_vd;
2493 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2494 boolean_t unexpected_error = B_FALSE;
2495
2496 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2497 return (ZIO_PIPELINE_STOP);
2498
2499 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2500
2501 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2502
2503 vdev_queue_io_done(zio);
2504
2505 if (zio->io_type == ZIO_TYPE_WRITE)
2506 vdev_cache_write(zio);
2507
2508 if (zio_injection_enabled && zio->io_error == 0)
2509 zio->io_error = zio_handle_device_injection(vd,
2510 zio, EIO);
2511
2512 if (zio_injection_enabled && zio->io_error == 0)
2513 zio->io_error = zio_handle_label_injection(zio, EIO);
2514
2515 if (zio->io_error) {
2516 if (!vdev_accessible(vd, zio)) {
2517 zio->io_error = SET_ERROR(ENXIO);
2518 } else {
2519 unexpected_error = B_TRUE;
2520 }
2521 }
2522 }
2523
2524 ops->vdev_op_io_done(zio);
2525
2526 if (unexpected_error)
2527 VERIFY(vdev_probe(vd, zio) == NULL);
2528
2529 return (ZIO_PIPELINE_CONTINUE);
2530 }
2531
2532 /*
2533 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2534 * disk, and use that to finish the checksum ereport later.
2535 */
2536 static void
2537 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2538 const void *good_buf)
2539 {
2540 /* no processing needed */
2541 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2542 }
2543
2544 /*ARGSUSED*/
2545 void
2546 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2547 {
2548 void *buf = zio_buf_alloc(zio->io_size);
2549
2550 bcopy(zio->io_data, buf, zio->io_size);
2551
2552 zcr->zcr_cbinfo = zio->io_size;
2553 zcr->zcr_cbdata = buf;
2554 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2555 zcr->zcr_free = zio_buf_free;
2556 }
2557
2558 static int
2559 zio_vdev_io_assess(zio_t *zio)
2560 {
2561 vdev_t *vd = zio->io_vd;
2562
2563 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2564 return (ZIO_PIPELINE_STOP);
2565
2566 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2567 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2568
2569 if (zio->io_vsd != NULL) {
2570 zio->io_vsd_ops->vsd_free(zio);
2571 zio->io_vsd = NULL;
2572 }
2573
2574 if (zio_injection_enabled && zio->io_error == 0)
2575 zio->io_error = zio_handle_fault_injection(zio, EIO);
2576
2577 /*
2578 * If the I/O failed, determine whether we should attempt to retry it.
2579 *
2580 * On retry, we cut in line in the issue queue, since we don't want
2581 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2582 */
2583 if (zio->io_error && vd == NULL &&
2584 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2585 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2586 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2587 zio->io_error = 0;
2588 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2589 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2590 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2591 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2592 zio_requeue_io_start_cut_in_line);
2593 return (ZIO_PIPELINE_STOP);
2594 }
2595
2596 /*
2597 * If we got an error on a leaf device, convert it to ENXIO
2598 * if the device is not accessible at all.
2599 */
2600 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2601 !vdev_accessible(vd, zio))
2602 zio->io_error = SET_ERROR(ENXIO);
2603
2604 /*
2605 * If we can't write to an interior vdev (mirror or RAID-Z),
2606 * set vdev_cant_write so that we stop trying to allocate from it.
2607 */
2608 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2609 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2610 vd->vdev_cant_write = B_TRUE;
2611 }
2612
2613 if (zio->io_error)
2614 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2615
2616 return (ZIO_PIPELINE_CONTINUE);
2617 }
2618
2619 void
2620 zio_vdev_io_reissue(zio_t *zio)
2621 {
2622 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2623 ASSERT(zio->io_error == 0);
2624
2625 zio->io_stage >>= 1;
2626 }
2627
2628 void
2629 zio_vdev_io_redone(zio_t *zio)
2630 {
2631 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2632
2633 zio->io_stage >>= 1;
2634 }
2635
2636 void
2637 zio_vdev_io_bypass(zio_t *zio)
2638 {
2639 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2640 ASSERT(zio->io_error == 0);
2641
2642 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2643 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2644 }
2645
2646 /*
2647 * ==========================================================================
2648 * Generate and verify checksums
2649 * ==========================================================================
2650 */
2651 static int
2652 zio_checksum_generate(zio_t *zio)
2653 {
2654 blkptr_t *bp = zio->io_bp;
2655 enum zio_checksum checksum;
2656
2657 if (bp == NULL) {
2658 /*
2659 * This is zio_write_phys().
2660 * We're either generating a label checksum, or none at all.
2661 */
2662 checksum = zio->io_prop.zp_checksum;
2663
2664 if (checksum == ZIO_CHECKSUM_OFF)
2665 return (ZIO_PIPELINE_CONTINUE);
2666
2667 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2668 } else {
2669 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2670 ASSERT(!IO_IS_ALLOCATING(zio));
2671 checksum = ZIO_CHECKSUM_GANG_HEADER;
2672 } else {
2673 checksum = BP_GET_CHECKSUM(bp);
2674 }
2675 }
2676
2677 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2678
2679 return (ZIO_PIPELINE_CONTINUE);
2680 }
2681
2682 static int
2683 zio_checksum_verify(zio_t *zio)
2684 {
2685 zio_bad_cksum_t info;
2686 blkptr_t *bp = zio->io_bp;
2687 int error;
2688
2689 ASSERT(zio->io_vd != NULL);
2690
2691 if (bp == NULL) {
2692 /*
2693 * This is zio_read_phys().
2694 * We're either verifying a label checksum, or nothing at all.
2695 */
2696 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2697 return (ZIO_PIPELINE_CONTINUE);
2698
2699 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2700 }
2701
2702 if ((error = zio_checksum_error(zio, &info)) != 0) {
2703 zio->io_error = error;
2704 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2705 zfs_ereport_start_checksum(zio->io_spa,
2706 zio->io_vd, zio, zio->io_offset,
2707 zio->io_size, NULL, &info);
2708 }
2709 }
2710
2711 return (ZIO_PIPELINE_CONTINUE);
2712 }
2713
2714 /*
2715 * Called by RAID-Z to ensure we don't compute the checksum twice.
2716 */
2717 void
2718 zio_checksum_verified(zio_t *zio)
2719 {
2720 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2721 }
2722
2723 /*
2724 * ==========================================================================
2725 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2726 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2727 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2728 * indicate errors that are specific to one I/O, and most likely permanent.
2729 * Any other error is presumed to be worse because we weren't expecting it.
2730 * ==========================================================================
2731 */
2732 int
2733 zio_worst_error(int e1, int e2)
2734 {
2735 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2736 int r1, r2;
2737
2738 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2739 if (e1 == zio_error_rank[r1])
2740 break;
2741
2742 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2743 if (e2 == zio_error_rank[r2])
2744 break;
2745
2746 return (r1 > r2 ? e1 : e2);
2747 }
2748
2749 /*
2750 * ==========================================================================
2751 * I/O completion
2752 * ==========================================================================
2753 */
2754 static int
2755 zio_ready(zio_t *zio)
2756 {
2757 blkptr_t *bp = zio->io_bp;
2758 zio_t *pio, *pio_next;
2759
2760 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2761 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2762 return (ZIO_PIPELINE_STOP);
2763
2764 if (zio->io_ready) {
2765 ASSERT(IO_IS_ALLOCATING(zio));
2766 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
2767 (zio->io_flags & ZIO_FLAG_NOPWRITE));
2768 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2769
2770 zio->io_ready(zio);
2771 }
2772
2773 if (bp != NULL && bp != &zio->io_bp_copy)
2774 zio->io_bp_copy = *bp;
2775
2776 if (zio->io_error)
2777 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2778
2779 mutex_enter(&zio->io_lock);
2780 zio->io_state[ZIO_WAIT_READY] = 1;
2781 pio = zio_walk_parents(zio);
2782 mutex_exit(&zio->io_lock);
2783
2784 /*
2785 * As we notify zio's parents, new parents could be added.
2786 * New parents go to the head of zio's io_parent_list, however,
2787 * so we will (correctly) not notify them. The remainder of zio's
2788 * io_parent_list, from 'pio_next' onward, cannot change because
2789 * all parents must wait for us to be done before they can be done.
2790 */
2791 for (; pio != NULL; pio = pio_next) {
2792 pio_next = zio_walk_parents(zio);
2793 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2794 }
2795
2796 if (zio->io_flags & ZIO_FLAG_NODATA) {
2797 if (BP_IS_GANG(bp)) {
2798 zio->io_flags &= ~ZIO_FLAG_NODATA;
2799 } else {
2800 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2801 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2802 }
2803 }
2804
2805 if (zio_injection_enabled &&
2806 zio->io_spa->spa_syncing_txg == zio->io_txg)
2807 zio_handle_ignored_writes(zio);
2808
2809 return (ZIO_PIPELINE_CONTINUE);
2810 }
2811
2812 static int
2813 zio_done(zio_t *zio)
2814 {
2815 spa_t *spa = zio->io_spa;
2816 zio_t *lio = zio->io_logical;
2817 blkptr_t *bp = zio->io_bp;
2818 vdev_t *vd = zio->io_vd;
2819 uint64_t psize = zio->io_size;
2820 zio_t *pio, *pio_next;
2821
2822 /*
2823 * If our children haven't all completed,
2824 * wait for them and then repeat this pipeline stage.
2825 */
2826 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2827 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2828 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2829 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2830 return (ZIO_PIPELINE_STOP);
2831
2832 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2833 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2834 ASSERT(zio->io_children[c][w] == 0);
2835
2836 if (bp != NULL) {
2837 ASSERT(bp->blk_pad[0] == 0);
2838 ASSERT(bp->blk_pad[1] == 0);
2839 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2840 (bp == zio_unique_parent(zio)->io_bp));
2841 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2842 zio->io_bp_override == NULL &&
2843 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2844 ASSERT(!BP_SHOULD_BYTESWAP(bp));
2845 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2846 ASSERT(BP_COUNT_GANG(bp) == 0 ||
2847 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2848 }
2849 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
2850 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
2851 }
2852
2853 /*
2854 * If there were child vdev/gang/ddt errors, they apply to us now.
2855 */
2856 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2857 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2858 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2859
2860 /*
2861 * If the I/O on the transformed data was successful, generate any
2862 * checksum reports now while we still have the transformed data.
2863 */
2864 if (zio->io_error == 0) {
2865 while (zio->io_cksum_report != NULL) {
2866 zio_cksum_report_t *zcr = zio->io_cksum_report;
2867 uint64_t align = zcr->zcr_align;
2868 uint64_t asize = P2ROUNDUP(psize, align);
2869 char *abuf = zio->io_data;
2870
2871 if (asize != psize) {
2872 abuf = zio_buf_alloc(asize);
2873 bcopy(zio->io_data, abuf, psize);
2874 bzero(abuf + psize, asize - psize);
2875 }
2876
2877 zio->io_cksum_report = zcr->zcr_next;
2878 zcr->zcr_next = NULL;
2879 zcr->zcr_finish(zcr, abuf);
2880 zfs_ereport_free_checksum(zcr);
2881
2882 if (asize != psize)
2883 zio_buf_free(abuf, asize);
2884 }
2885 }
2886
2887 zio_pop_transforms(zio); /* note: may set zio->io_error */
2888
2889 vdev_stat_update(zio, psize);
2890
2891 if (zio->io_error) {
2892 /*
2893 * If this I/O is attached to a particular vdev,
2894 * generate an error message describing the I/O failure
2895 * at the block level. We ignore these errors if the
2896 * device is currently unavailable.
2897 */
2898 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
2899 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
2900
2901 if ((zio->io_error == EIO || !(zio->io_flags &
2902 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2903 zio == lio) {
2904 /*
2905 * For logical I/O requests, tell the SPA to log the
2906 * error and generate a logical data ereport.
2907 */
2908 spa_log_error(spa, zio);
2909 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
2910 0, 0);
2911 }
2912 }
2913
2914 if (zio->io_error && zio == lio) {
2915 /*
2916 * Determine whether zio should be reexecuted. This will
2917 * propagate all the way to the root via zio_notify_parent().
2918 */
2919 ASSERT(vd == NULL && bp != NULL);
2920 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2921
2922 if (IO_IS_ALLOCATING(zio) &&
2923 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2924 if (zio->io_error != ENOSPC)
2925 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2926 else
2927 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2928 }
2929
2930 if ((zio->io_type == ZIO_TYPE_READ ||
2931 zio->io_type == ZIO_TYPE_FREE) &&
2932 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2933 zio->io_error == ENXIO &&
2934 spa_load_state(spa) == SPA_LOAD_NONE &&
2935 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
2936 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2937
2938 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2939 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2940
2941 /*
2942 * Here is a possibly good place to attempt to do
2943 * either combinatorial reconstruction or error correction
2944 * based on checksums. It also might be a good place
2945 * to send out preliminary ereports before we suspend
2946 * processing.
2947 */
2948 }
2949
2950 /*
2951 * If there were logical child errors, they apply to us now.
2952 * We defer this until now to avoid conflating logical child
2953 * errors with errors that happened to the zio itself when
2954 * updating vdev stats and reporting FMA events above.
2955 */
2956 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2957
2958 if ((zio->io_error || zio->io_reexecute) &&
2959 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2960 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
2961 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
2962
2963 zio_gang_tree_free(&zio->io_gang_tree);
2964
2965 /*
2966 * Godfather I/Os should never suspend.
2967 */
2968 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2969 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2970 zio->io_reexecute = 0;
2971
2972 if (zio->io_reexecute) {
2973 /*
2974 * This is a logical I/O that wants to reexecute.
2975 *
2976 * Reexecute is top-down. When an i/o fails, if it's not
2977 * the root, it simply notifies its parent and sticks around.
2978 * The parent, seeing that it still has children in zio_done(),
2979 * does the same. This percolates all the way up to the root.
2980 * The root i/o will reexecute or suspend the entire tree.
2981 *
2982 * This approach ensures that zio_reexecute() honors
2983 * all the original i/o dependency relationships, e.g.
2984 * parents not executing until children are ready.
2985 */
2986 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2987
2988 zio->io_gang_leader = NULL;
2989
2990 mutex_enter(&zio->io_lock);
2991 zio->io_state[ZIO_WAIT_DONE] = 1;
2992 mutex_exit(&zio->io_lock);
2993
2994 /*
2995 * "The Godfather" I/O monitors its children but is
2996 * not a true parent to them. It will track them through
2997 * the pipeline but severs its ties whenever they get into
2998 * trouble (e.g. suspended). This allows "The Godfather"
2999 * I/O to return status without blocking.
3000 */
3001 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3002 zio_link_t *zl = zio->io_walk_link;
3003 pio_next = zio_walk_parents(zio);
3004
3005 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3006 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3007 zio_remove_child(pio, zio, zl);
3008 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3009 }
3010 }
3011
3012 if ((pio = zio_unique_parent(zio)) != NULL) {
3013 /*
3014 * We're not a root i/o, so there's nothing to do
3015 * but notify our parent. Don't propagate errors
3016 * upward since we haven't permanently failed yet.
3017 */
3018 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3019 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3020 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3021 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3022 /*
3023 * We'd fail again if we reexecuted now, so suspend
3024 * until conditions improve (e.g. device comes online).
3025 */
3026 zio_suspend(spa, zio);
3027 } else {
3028 /*
3029 * Reexecution is potentially a huge amount of work.
3030 * Hand it off to the otherwise-unused claim taskq.
3031 */
3032 ASSERT(zio->io_tqent.tqent_next == NULL);
3033 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3034 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3035 0, &zio->io_tqent);
3036 }
3037 return (ZIO_PIPELINE_STOP);
3038 }
3039
3040 ASSERT(zio->io_child_count == 0);
3041 ASSERT(zio->io_reexecute == 0);
3042 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3043
3044 /*
3045 * Report any checksum errors, since the I/O is complete.
3046 */
3047 while (zio->io_cksum_report != NULL) {
3048 zio_cksum_report_t *zcr = zio->io_cksum_report;
3049 zio->io_cksum_report = zcr->zcr_next;
3050 zcr->zcr_next = NULL;
3051 zcr->zcr_finish(zcr, NULL);
3052 zfs_ereport_free_checksum(zcr);
3053 }
3054
3055 /*
3056 * It is the responsibility of the done callback to ensure that this
3057 * particular zio is no longer discoverable for adoption, and as
3058 * such, cannot acquire any new parents.
3059 */
3060 if (zio->io_done)
3061 zio->io_done(zio);
3062
3063 mutex_enter(&zio->io_lock);
3064 zio->io_state[ZIO_WAIT_DONE] = 1;
3065 mutex_exit(&zio->io_lock);
3066
3067 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3068 zio_link_t *zl = zio->io_walk_link;
3069 pio_next = zio_walk_parents(zio);
3070 zio_remove_child(pio, zio, zl);
3071 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3072 }
3073
3074 if (zio->io_waiter != NULL) {
3075 mutex_enter(&zio->io_lock);
3076 zio->io_executor = NULL;
3077 cv_broadcast(&zio->io_cv);
3078 mutex_exit(&zio->io_lock);
3079 } else {
3080 zio_destroy(zio);
3081 }
3082
3083 return (ZIO_PIPELINE_STOP);
3084 }
3085
3086 /*
3087 * ==========================================================================
3088 * I/O pipeline definition
3089 * ==========================================================================
3090 */
3091 static zio_pipe_stage_t *zio_pipeline[] = {
3092 NULL,
3093 zio_read_bp_init,
3094 zio_free_bp_init,
3095 zio_issue_async,
3096 zio_write_bp_init,
3097 zio_checksum_generate,
3098 zio_nop_write,
3099 zio_ddt_read_start,
3100 zio_ddt_read_done,
3101 zio_ddt_write,
3102 zio_ddt_free,
3103 zio_gang_assemble,
3104 zio_gang_issue,
3105 zio_dva_allocate,
3106 zio_dva_free,
3107 zio_dva_claim,
3108 zio_ready,
3109 zio_vdev_io_start,
3110 zio_vdev_io_done,
3111 zio_vdev_io_assess,
3112 zio_checksum_verify,
3113 zio_done
3114 };
3115
3116 /* dnp is the dnode for zb1->zb_object */
3117 boolean_t
3118 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_t *zb1,
3119 const zbookmark_t *zb2)
3120 {
3121 uint64_t zb1nextL0, zb2thisobj;
3122
3123 ASSERT(zb1->zb_objset == zb2->zb_objset);
3124 ASSERT(zb2->zb_level == 0);
3125
3126 /*
3127 * A bookmark in the deadlist is considered to be after
3128 * everything else.
3129 */
3130 if (zb2->zb_object == DMU_DEADLIST_OBJECT)
3131 return (B_TRUE);
3132
3133 /* The objset_phys_t isn't before anything. */
3134 if (dnp == NULL)
3135 return (B_FALSE);
3136
3137 zb1nextL0 = (zb1->zb_blkid + 1) <<
3138 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3139
3140 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3141 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3142
3143 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3144 uint64_t nextobj = zb1nextL0 *
3145 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3146 return (nextobj <= zb2thisobj);
3147 }
3148
3149 if (zb1->zb_object < zb2thisobj)
3150 return (B_TRUE);
3151 if (zb1->zb_object > zb2thisobj)
3152 return (B_FALSE);
3153 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3154 return (B_FALSE);
3155 return (zb1nextL0 <= zb2->zb_blkid);
3156 }