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) 2011, 2014 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25 * Copyright 2013 Joyent, Inc. All rights reserved.
26 */
27
28 #include <sys/sysmacros.h>
29 #include <sys/zfs_context.h>
30 #include <sys/fm/fs/zfs.h>
31 #include <sys/spa.h>
32 #include <sys/txg.h>
33 #include <sys/spa_impl.h>
34 #include <sys/vdev_impl.h>
35 #include <sys/zio_impl.h>
36 #include <sys/zio_compress.h>
37 #include <sys/zio_checksum.h>
38 #include <sys/dmu_objset.h>
39 #include <sys/arc.h>
40 #include <sys/ddt.h>
41 #include <sys/zfs_zone.h>
42 #include <sys/blkptr.h>
43 #include <sys/zfeature.h>
44
45 /*
46 * ==========================================================================
47 * I/O type descriptions
48 * ==========================================================================
49 */
50 const char *zio_type_name[ZIO_TYPES] = {
51 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
52 "zio_ioctl"
53 };
54
55 /*
56 * ==========================================================================
57 * I/O kmem caches
58 * ==========================================================================
59 */
60 kmem_cache_t *zio_cache;
61 kmem_cache_t *zio_link_cache;
62 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
63 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
64
65 #ifdef _KERNEL
66 extern vmem_t *zio_alloc_arena;
67 #endif
68
69 #define ZIO_PIPELINE_CONTINUE 0x100
70 #define ZIO_PIPELINE_STOP 0x101
71
72 /*
73 * The following actions directly effect the spa's sync-to-convergence logic.
74 * The values below define the sync pass when we start performing the action.
75 * Care should be taken when changing these values as they directly impact
76 * spa_sync() performance. Tuning these values may introduce subtle performance
77 * pathologies and should only be done in the context of performance analysis.
78 * These tunables will eventually be removed and replaced with #defines once
79 * enough analysis has been done to determine optimal values.
80 *
81 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
82 * regular blocks are not deferred.
83 */
84 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
85 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
86 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
87
88 /*
89 * An allocating zio is one that either currently has the DVA allocate
90 * stage set or will have it later in its lifetime.
91 */
92 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
93
94 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
95
96 #ifdef ZFS_DEBUG
97 int zio_buf_debug_limit = 16384;
98 #else
99 int zio_buf_debug_limit = 0;
100 #endif
101
102 void
103 zio_init(void)
104 {
105 size_t c;
106 vmem_t *data_alloc_arena = NULL;
107
108 #ifdef _KERNEL
109 data_alloc_arena = zio_alloc_arena;
110 #endif
111 zio_cache = kmem_cache_create("zio_cache",
112 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
113 zio_link_cache = kmem_cache_create("zio_link_cache",
114 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
115
116 /*
117 * For small buffers, we want a cache for each multiple of
118 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
119 * for each quarter-power of 2.
120 */
121 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
122 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
123 size_t p2 = size;
124 size_t align = 0;
125 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
126
127 while (!ISP2(p2))
128 p2 &= p2 - 1;
129
130 #ifndef _KERNEL
131 /*
132 * If we are using watchpoints, put each buffer on its own page,
133 * to eliminate the performance overhead of trapping to the
134 * kernel when modifying a non-watched buffer that shares the
135 * page with a watched buffer.
136 */
137 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
138 continue;
139 #endif
140 if (size <= 4 * SPA_MINBLOCKSIZE) {
141 align = SPA_MINBLOCKSIZE;
142 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
143 align = MIN(p2 >> 2, PAGESIZE);
144 }
145
146 if (align != 0) {
147 char name[36];
148 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
149 zio_buf_cache[c] = kmem_cache_create(name, size,
150 align, NULL, NULL, NULL, NULL, NULL, cflags);
151
152 /*
153 * Since zio_data bufs do not appear in crash dumps, we
154 * pass KMC_NOTOUCH so that no allocator metadata is
155 * stored with the buffers.
156 */
157 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
158 zio_data_buf_cache[c] = kmem_cache_create(name, size,
159 align, NULL, NULL, NULL, NULL, data_alloc_arena,
160 cflags | KMC_NOTOUCH);
161 }
162 }
163
164 while (--c != 0) {
165 ASSERT(zio_buf_cache[c] != NULL);
166 if (zio_buf_cache[c - 1] == NULL)
167 zio_buf_cache[c - 1] = zio_buf_cache[c];
168
169 ASSERT(zio_data_buf_cache[c] != NULL);
170 if (zio_data_buf_cache[c - 1] == NULL)
171 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
172 }
173
174 zio_inject_init();
175 }
176
177 void
178 zio_fini(void)
179 {
180 size_t c;
181 kmem_cache_t *last_cache = NULL;
182 kmem_cache_t *last_data_cache = NULL;
183
184 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
185 if (zio_buf_cache[c] != last_cache) {
186 last_cache = zio_buf_cache[c];
187 kmem_cache_destroy(zio_buf_cache[c]);
188 }
189 zio_buf_cache[c] = NULL;
190
191 if (zio_data_buf_cache[c] != last_data_cache) {
192 last_data_cache = zio_data_buf_cache[c];
193 kmem_cache_destroy(zio_data_buf_cache[c]);
194 }
195 zio_data_buf_cache[c] = NULL;
196 }
197
198 kmem_cache_destroy(zio_link_cache);
199 kmem_cache_destroy(zio_cache);
200
201 zio_inject_fini();
202 }
203
204 /*
205 * ==========================================================================
206 * Allocate and free I/O buffers
207 * ==========================================================================
208 */
209
210 /*
211 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
212 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
213 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
214 * excess / transient data in-core during a crashdump.
215 */
216 void *
217 zio_buf_alloc(size_t size)
218 {
219 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
220
221 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
222
223 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
224 }
225
226 /*
227 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
228 * crashdump if the kernel panics. This exists so that we will limit the amount
229 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
230 * of kernel heap dumped to disk when the kernel panics)
231 */
232 void *
233 zio_data_buf_alloc(size_t size)
234 {
235 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
236
237 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
238
239 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
240 }
241
242 void
243 zio_buf_free(void *buf, size_t size)
244 {
245 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
246
247 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
248
249 kmem_cache_free(zio_buf_cache[c], buf);
250 }
251
252 void
253 zio_data_buf_free(void *buf, size_t size)
254 {
255 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
256
257 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
258
259 kmem_cache_free(zio_data_buf_cache[c], buf);
260 }
261
262 /*
263 * ==========================================================================
264 * Push and pop I/O transform buffers
265 * ==========================================================================
266 */
267 static void
268 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
269 zio_transform_func_t *transform)
270 {
271 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
272
273 zt->zt_orig_data = zio->io_data;
274 zt->zt_orig_size = zio->io_size;
275 zt->zt_bufsize = bufsize;
276 zt->zt_transform = transform;
277
278 zt->zt_next = zio->io_transform_stack;
279 zio->io_transform_stack = zt;
280
281 zio->io_data = data;
282 zio->io_size = size;
283 }
284
285 static void
286 zio_pop_transforms(zio_t *zio)
287 {
288 zio_transform_t *zt;
289
290 while ((zt = zio->io_transform_stack) != NULL) {
291 if (zt->zt_transform != NULL)
292 zt->zt_transform(zio,
293 zt->zt_orig_data, zt->zt_orig_size);
294
295 if (zt->zt_bufsize != 0)
296 zio_buf_free(zio->io_data, zt->zt_bufsize);
297
298 zio->io_data = zt->zt_orig_data;
299 zio->io_size = zt->zt_orig_size;
300 zio->io_transform_stack = zt->zt_next;
301
302 kmem_free(zt, sizeof (zio_transform_t));
303 }
304 }
305
306 /*
307 * ==========================================================================
308 * I/O transform callbacks for subblocks and decompression
309 * ==========================================================================
310 */
311 static void
312 zio_subblock(zio_t *zio, void *data, uint64_t size)
313 {
314 ASSERT(zio->io_size > size);
315
316 if (zio->io_type == ZIO_TYPE_READ)
317 bcopy(zio->io_data, data, size);
318 }
319
320 static void
321 zio_decompress(zio_t *zio, void *data, uint64_t size)
322 {
323 if (zio->io_error == 0 &&
324 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
325 zio->io_data, data, zio->io_size, size) != 0)
326 zio->io_error = SET_ERROR(EIO);
327 }
328
329 /*
330 * ==========================================================================
331 * I/O parent/child relationships and pipeline interlocks
332 * ==========================================================================
333 */
334 /*
335 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
336 * continue calling these functions until they return NULL.
337 * Otherwise, the next caller will pick up the list walk in
338 * some indeterminate state. (Otherwise every caller would
339 * have to pass in a cookie to keep the state represented by
340 * io_walk_link, which gets annoying.)
341 */
342 zio_t *
343 zio_walk_parents(zio_t *cio)
344 {
345 zio_link_t *zl = cio->io_walk_link;
346 list_t *pl = &cio->io_parent_list;
347
348 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
349 cio->io_walk_link = zl;
350
351 if (zl == NULL)
352 return (NULL);
353
354 ASSERT(zl->zl_child == cio);
355 return (zl->zl_parent);
356 }
357
358 zio_t *
359 zio_walk_children(zio_t *pio)
360 {
361 zio_link_t *zl = pio->io_walk_link;
362 list_t *cl = &pio->io_child_list;
363
364 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
365 pio->io_walk_link = zl;
366
367 if (zl == NULL)
368 return (NULL);
369
370 ASSERT(zl->zl_parent == pio);
371 return (zl->zl_child);
372 }
373
374 zio_t *
375 zio_unique_parent(zio_t *cio)
376 {
377 zio_t *pio = zio_walk_parents(cio);
378
379 VERIFY(zio_walk_parents(cio) == NULL);
380 return (pio);
381 }
382
383 void
384 zio_add_child(zio_t *pio, zio_t *cio)
385 {
386 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
387
388 /*
389 * Logical I/Os can have logical, gang, or vdev children.
390 * Gang I/Os can have gang or vdev children.
391 * Vdev I/Os can only have vdev children.
392 * The following ASSERT captures all of these constraints.
393 */
394 ASSERT(cio->io_child_type <= pio->io_child_type);
395
396 zl->zl_parent = pio;
397 zl->zl_child = cio;
398
399 mutex_enter(&cio->io_lock);
400 mutex_enter(&pio->io_lock);
401
402 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
403
404 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
405 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
406
407 list_insert_head(&pio->io_child_list, zl);
408 list_insert_head(&cio->io_parent_list, zl);
409
410 pio->io_child_count++;
411 cio->io_parent_count++;
412
413 mutex_exit(&pio->io_lock);
414 mutex_exit(&cio->io_lock);
415 }
416
417 static void
418 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
419 {
420 ASSERT(zl->zl_parent == pio);
421 ASSERT(zl->zl_child == cio);
422
423 mutex_enter(&cio->io_lock);
424 mutex_enter(&pio->io_lock);
425
426 list_remove(&pio->io_child_list, zl);
427 list_remove(&cio->io_parent_list, zl);
428
429 pio->io_child_count--;
430 cio->io_parent_count--;
431
432 mutex_exit(&pio->io_lock);
433 mutex_exit(&cio->io_lock);
434
435 kmem_cache_free(zio_link_cache, zl);
436 }
437
438 static boolean_t
439 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
440 {
441 uint64_t *countp = &zio->io_children[child][wait];
442 boolean_t waiting = B_FALSE;
443
444 mutex_enter(&zio->io_lock);
445 ASSERT(zio->io_stall == NULL);
446 if (*countp != 0) {
447 zio->io_stage >>= 1;
448 zio->io_stall = countp;
449 waiting = B_TRUE;
450 }
451 mutex_exit(&zio->io_lock);
452
453 return (waiting);
454 }
455
456 static void
457 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
458 {
459 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
460 int *errorp = &pio->io_child_error[zio->io_child_type];
461
462 mutex_enter(&pio->io_lock);
463 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
464 *errorp = zio_worst_error(*errorp, zio->io_error);
465 pio->io_reexecute |= zio->io_reexecute;
466 ASSERT3U(*countp, >, 0);
467
468 (*countp)--;
469
470 if (*countp == 0 && pio->io_stall == countp) {
471 pio->io_stall = NULL;
472 mutex_exit(&pio->io_lock);
473 zio_execute(pio);
474 } else {
475 mutex_exit(&pio->io_lock);
476 }
477 }
478
479 static void
480 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
481 {
482 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
483 zio->io_error = zio->io_child_error[c];
484 }
485
486 /*
487 * ==========================================================================
488 * Create the various types of I/O (read, write, free, etc)
489 * ==========================================================================
490 */
491 static zio_t *
492 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
493 void *data, uint64_t size, zio_done_func_t *done, void *private,
494 zio_type_t type, zio_priority_t priority, enum zio_flag flags,
495 vdev_t *vd, uint64_t offset, const zbookmark_phys_t *zb,
496 enum zio_stage stage, enum zio_stage pipeline)
497 {
498 zio_t *zio;
499
500 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
501 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
502 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
503
504 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
505 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
506 ASSERT(vd || stage == ZIO_STAGE_OPEN);
507
508 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
509 bzero(zio, sizeof (zio_t));
510
511 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
512 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
513
514 list_create(&zio->io_parent_list, sizeof (zio_link_t),
515 offsetof(zio_link_t, zl_parent_node));
516 list_create(&zio->io_child_list, sizeof (zio_link_t),
517 offsetof(zio_link_t, zl_child_node));
518
519 if (vd != NULL)
520 zio->io_child_type = ZIO_CHILD_VDEV;
521 else if (flags & ZIO_FLAG_GANG_CHILD)
522 zio->io_child_type = ZIO_CHILD_GANG;
523 else if (flags & ZIO_FLAG_DDT_CHILD)
524 zio->io_child_type = ZIO_CHILD_DDT;
525 else
526 zio->io_child_type = ZIO_CHILD_LOGICAL;
527
528 if (bp != NULL) {
529 zio->io_bp = (blkptr_t *)bp;
530 zio->io_bp_copy = *bp;
531 zio->io_bp_orig = *bp;
532 if (type != ZIO_TYPE_WRITE ||
533 zio->io_child_type == ZIO_CHILD_DDT)
534 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
535 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
536 zio->io_logical = zio;
537 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
538 pipeline |= ZIO_GANG_STAGES;
539 }
540
541 zio->io_spa = spa;
542 zio->io_txg = txg;
543 zio->io_done = done;
544 zio->io_private = private;
545 zio->io_type = type;
546 zio->io_priority = priority;
547 zio->io_vd = vd;
548 zio->io_offset = offset;
549 zio->io_orig_data = zio->io_data = data;
550 zio->io_orig_size = zio->io_size = size;
551 zio->io_orig_flags = zio->io_flags = flags;
552 zio->io_orig_stage = zio->io_stage = stage;
553 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
554
555 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
556 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
557
558 if (zb != NULL)
559 zio->io_bookmark = *zb;
560
561 if (pio != NULL) {
562 zio->io_zoneid = pio->io_zoneid;
563 if (zio->io_logical == NULL)
564 zio->io_logical = pio->io_logical;
565 if (zio->io_child_type == ZIO_CHILD_GANG)
566 zio->io_gang_leader = pio->io_gang_leader;
567 zio_add_child(pio, zio);
568 } else {
569 zfs_zone_zio_init(zio);
570 }
571
572 return (zio);
573 }
574
575 static void
576 zio_destroy(zio_t *zio)
577 {
578 list_destroy(&zio->io_parent_list);
579 list_destroy(&zio->io_child_list);
580 mutex_destroy(&zio->io_lock);
581 cv_destroy(&zio->io_cv);
582 kmem_cache_free(zio_cache, zio);
583 }
584
585 zio_t *
586 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
587 void *private, enum zio_flag flags)
588 {
589 zio_t *zio;
590
591 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
592 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
593 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
594
595 return (zio);
596 }
597
598 zio_t *
599 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
600 {
601 return (zio_null(NULL, spa, NULL, done, private, flags));
602 }
603
604 void
605 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
606 {
607 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
608 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
609 bp, (longlong_t)BP_GET_TYPE(bp));
610 }
611 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
612 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
613 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
614 bp, (longlong_t)BP_GET_CHECKSUM(bp));
615 }
616 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
617 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
618 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
619 bp, (longlong_t)BP_GET_COMPRESS(bp));
620 }
621 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
622 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
623 bp, (longlong_t)BP_GET_LSIZE(bp));
624 }
625 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
626 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
627 bp, (longlong_t)BP_GET_PSIZE(bp));
628 }
629
630 if (BP_IS_EMBEDDED(bp)) {
631 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
632 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
633 bp, (longlong_t)BPE_GET_ETYPE(bp));
634 }
635 }
636
637 /*
638 * Pool-specific checks.
639 *
640 * Note: it would be nice to verify that the blk_birth and
641 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
642 * allows the birth time of log blocks (and dmu_sync()-ed blocks
643 * that are in the log) to be arbitrarily large.
644 */
645 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
646 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
647 if (vdevid >= spa->spa_root_vdev->vdev_children) {
648 zfs_panic_recover("blkptr at %p DVA %u has invalid "
649 "VDEV %llu",
650 bp, i, (longlong_t)vdevid);
651 }
652 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
653 if (vd == NULL) {
654 zfs_panic_recover("blkptr at %p DVA %u has invalid "
655 "VDEV %llu",
656 bp, i, (longlong_t)vdevid);
657 }
658 if (vd->vdev_ops == &vdev_hole_ops) {
659 zfs_panic_recover("blkptr at %p DVA %u has hole "
660 "VDEV %llu",
661 bp, i, (longlong_t)vdevid);
662
663 }
664 if (vd->vdev_ops == &vdev_missing_ops) {
665 /*
666 * "missing" vdevs are valid during import, but we
667 * don't have their detailed info (e.g. asize), so
668 * we can't perform any more checks on them.
669 */
670 continue;
671 }
672 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
673 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
674 if (BP_IS_GANG(bp))
675 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
676 if (offset + asize > vd->vdev_asize) {
677 zfs_panic_recover("blkptr at %p DVA %u has invalid "
678 "OFFSET %llu",
679 bp, i, (longlong_t)offset);
680 }
681 }
682 }
683
684 zio_t *
685 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
686 void *data, uint64_t size, zio_done_func_t *done, void *private,
687 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
688 {
689 zio_t *zio;
690
691 zfs_blkptr_verify(spa, bp);
692
693 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
694 data, size, done, private,
695 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
696 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
697 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
698
699 return (zio);
700 }
701
702 zio_t *
703 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
704 void *data, uint64_t size, const zio_prop_t *zp,
705 zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done,
706 void *private,
707 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
708 {
709 zio_t *zio;
710
711 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
712 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
713 zp->zp_compress >= ZIO_COMPRESS_OFF &&
714 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
715 DMU_OT_IS_VALID(zp->zp_type) &&
716 zp->zp_level < 32 &&
717 zp->zp_copies > 0 &&
718 zp->zp_copies <= spa_max_replication(spa));
719
720 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
721 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
722 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
723 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
724
725 zio->io_ready = ready;
726 zio->io_physdone = physdone;
727 zio->io_prop = *zp;
728
729 /*
730 * Data can be NULL if we are going to call zio_write_override() to
731 * provide the already-allocated BP. But we may need the data to
732 * verify a dedup hit (if requested). In this case, don't try to
733 * dedup (just take the already-allocated BP verbatim).
734 */
735 if (data == NULL && zio->io_prop.zp_dedup_verify) {
736 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
737 }
738
739 return (zio);
740 }
741
742 zio_t *
743 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
744 uint64_t size, zio_done_func_t *done, void *private,
745 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
746 {
747 zio_t *zio;
748
749 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
750 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
751 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
752
753 return (zio);
754 }
755
756 void
757 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
758 {
759 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
760 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
761 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
762 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
763
764 /*
765 * We must reset the io_prop to match the values that existed
766 * when the bp was first written by dmu_sync() keeping in mind
767 * that nopwrite and dedup are mutually exclusive.
768 */
769 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
770 zio->io_prop.zp_nopwrite = nopwrite;
771 zio->io_prop.zp_copies = copies;
772 zio->io_bp_override = bp;
773 }
774
775 void
776 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
777 {
778
779 /*
780 * The check for EMBEDDED is a performance optimization. We
781 * process the free here (by ignoring it) rather than
782 * putting it on the list and then processing it in zio_free_sync().
783 */
784 if (BP_IS_EMBEDDED(bp))
785 return;
786 metaslab_check_free(spa, bp);
787
788 /*
789 * Frees that are for the currently-syncing txg, are not going to be
790 * deferred, and which will not need to do a read (i.e. not GANG or
791 * DEDUP), can be processed immediately. Otherwise, put them on the
792 * in-memory list for later processing.
793 */
794 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
795 txg != spa->spa_syncing_txg ||
796 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
797 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
798 } else {
799 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0)));
800 }
801 }
802
803 zio_t *
804 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
805 enum zio_flag flags)
806 {
807 zio_t *zio;
808 enum zio_stage stage = ZIO_FREE_PIPELINE;
809
810 ASSERT(!BP_IS_HOLE(bp));
811 ASSERT(spa_syncing_txg(spa) == txg);
812 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
813
814 if (BP_IS_EMBEDDED(bp))
815 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
816
817 metaslab_check_free(spa, bp);
818 arc_freed(spa, bp);
819
820 /*
821 * GANG and DEDUP blocks can induce a read (for the gang block header,
822 * or the DDT), so issue them asynchronously so that this thread is
823 * not tied up.
824 */
825 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
826 stage |= ZIO_STAGE_ISSUE_ASYNC;
827
828 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
829 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags,
830 NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
831
832 return (zio);
833 }
834
835 zio_t *
836 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
837 zio_done_func_t *done, void *private, enum zio_flag flags)
838 {
839 zio_t *zio;
840
841 dprintf_bp(bp, "claiming in txg %llu", txg);
842
843 if (BP_IS_EMBEDDED(bp))
844 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
845
846 /*
847 * A claim is an allocation of a specific block. Claims are needed
848 * to support immediate writes in the intent log. The issue is that
849 * immediate writes contain committed data, but in a txg that was
850 * *not* committed. Upon opening the pool after an unclean shutdown,
851 * the intent log claims all blocks that contain immediate write data
852 * so that the SPA knows they're in use.
853 *
854 * All claims *must* be resolved in the first txg -- before the SPA
855 * starts allocating blocks -- so that nothing is allocated twice.
856 * If txg == 0 we just verify that the block is claimable.
857 */
858 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
859 ASSERT(txg == spa_first_txg(spa) || txg == 0);
860 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
861
862 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
863 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
864 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
865
866 return (zio);
867 }
868
869 zio_t *
870 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
871 zio_done_func_t *done, void *private, enum zio_flag flags)
872 {
873 zio_t *zio;
874 int c;
875
876 if (vd->vdev_children == 0) {
877 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
878 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
879 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
880
881 zio->io_cmd = cmd;
882 } else {
883 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
884
885 for (c = 0; c < vd->vdev_children; c++)
886 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
887 done, private, flags));
888 }
889
890 return (zio);
891 }
892
893 zio_t *
894 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
895 void *data, int checksum, zio_done_func_t *done, void *private,
896 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
897 {
898 zio_t *zio;
899
900 ASSERT(vd->vdev_children == 0);
901 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
902 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
903 ASSERT3U(offset + size, <=, vd->vdev_psize);
904
905 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
906 ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
907 NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
908
909 zio->io_prop.zp_checksum = checksum;
910
911 return (zio);
912 }
913
914 zio_t *
915 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
916 void *data, int checksum, zio_done_func_t *done, void *private,
917 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
918 {
919 zio_t *zio;
920
921 ASSERT(vd->vdev_children == 0);
922 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
923 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
924 ASSERT3U(offset + size, <=, vd->vdev_psize);
925
926 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
927 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
928 NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
929
930 zio->io_prop.zp_checksum = checksum;
931
932 if (zio_checksum_table[checksum].ci_eck) {
933 /*
934 * zec checksums are necessarily destructive -- they modify
935 * the end of the write buffer to hold the verifier/checksum.
936 * Therefore, we must make a local copy in case the data is
937 * being written to multiple places in parallel.
938 */
939 void *wbuf = zio_buf_alloc(size);
940 bcopy(data, wbuf, size);
941 zio_push_transform(zio, wbuf, size, size, NULL);
942 }
943
944 return (zio);
945 }
946
947 /*
948 * Create a child I/O to do some work for us.
949 */
950 zio_t *
951 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
952 void *data, uint64_t size, int type, zio_priority_t priority,
953 enum zio_flag flags, zio_done_func_t *done, void *private)
954 {
955 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
956 zio_t *zio;
957
958 ASSERT(vd->vdev_parent ==
959 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
960
961 if (type == ZIO_TYPE_READ && bp != NULL) {
962 /*
963 * If we have the bp, then the child should perform the
964 * checksum and the parent need not. This pushes error
965 * detection as close to the leaves as possible and
966 * eliminates redundant checksums in the interior nodes.
967 */
968 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
969 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
970 }
971
972 if (vd->vdev_children == 0)
973 offset += VDEV_LABEL_START_SIZE;
974
975 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
976
977 /*
978 * If we've decided to do a repair, the write is not speculative --
979 * even if the original read was.
980 */
981 if (flags & ZIO_FLAG_IO_REPAIR)
982 flags &= ~ZIO_FLAG_SPECULATIVE;
983
984 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
985 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
986 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
987
988 zio->io_physdone = pio->io_physdone;
989 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
990 zio->io_logical->io_phys_children++;
991
992 return (zio);
993 }
994
995 zio_t *
996 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
997 int type, zio_priority_t priority, enum zio_flag flags,
998 zio_done_func_t *done, void *private)
999 {
1000 zio_t *zio;
1001
1002 ASSERT(vd->vdev_ops->vdev_op_leaf);
1003
1004 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1005 data, size, done, private, type, priority,
1006 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1007 vd, offset, NULL,
1008 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1009
1010 return (zio);
1011 }
1012
1013 void
1014 zio_flush(zio_t *zio, vdev_t *vd)
1015 {
1016 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
1017 NULL, NULL,
1018 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1019 }
1020
1021 void
1022 zio_shrink(zio_t *zio, uint64_t size)
1023 {
1024 ASSERT(zio->io_executor == NULL);
1025 ASSERT(zio->io_orig_size == zio->io_size);
1026 ASSERT(size <= zio->io_size);
1027
1028 /*
1029 * We don't shrink for raidz because of problems with the
1030 * reconstruction when reading back less than the block size.
1031 * Note, BP_IS_RAIDZ() assumes no compression.
1032 */
1033 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1034 if (!BP_IS_RAIDZ(zio->io_bp))
1035 zio->io_orig_size = zio->io_size = size;
1036 }
1037
1038 /*
1039 * ==========================================================================
1040 * Prepare to read and write logical blocks
1041 * ==========================================================================
1042 */
1043
1044 static int
1045 zio_read_bp_init(zio_t *zio)
1046 {
1047 blkptr_t *bp = zio->io_bp;
1048
1049 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1050 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1051 !(zio->io_flags & ZIO_FLAG_RAW)) {
1052 uint64_t psize =
1053 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1054 void *cbuf = zio_buf_alloc(psize);
1055
1056 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1057 }
1058
1059 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1060 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1061 decode_embedded_bp_compressed(bp, zio->io_data);
1062 } else {
1063 ASSERT(!BP_IS_EMBEDDED(bp));
1064 }
1065
1066 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1067 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1068
1069 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1070 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1071
1072 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1073 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1074
1075 return (ZIO_PIPELINE_CONTINUE);
1076 }
1077
1078 static int
1079 zio_write_bp_init(zio_t *zio)
1080 {
1081 spa_t *spa = zio->io_spa;
1082 zio_prop_t *zp = &zio->io_prop;
1083 enum zio_compress compress = zp->zp_compress;
1084 blkptr_t *bp = zio->io_bp;
1085 uint64_t lsize = zio->io_size;
1086 uint64_t psize = lsize;
1087 int pass = 1;
1088
1089 /*
1090 * If our children haven't all reached the ready stage,
1091 * wait for them and then repeat this pipeline stage.
1092 */
1093 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1094 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1095 return (ZIO_PIPELINE_STOP);
1096
1097 if (!IO_IS_ALLOCATING(zio))
1098 return (ZIO_PIPELINE_CONTINUE);
1099
1100 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1101
1102 if (zio->io_bp_override) {
1103 ASSERT(bp->blk_birth != zio->io_txg);
1104 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1105
1106 *bp = *zio->io_bp_override;
1107 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1108
1109 if (BP_IS_EMBEDDED(bp))
1110 return (ZIO_PIPELINE_CONTINUE);
1111
1112 /*
1113 * If we've been overridden and nopwrite is set then
1114 * set the flag accordingly to indicate that a nopwrite
1115 * has already occurred.
1116 */
1117 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1118 ASSERT(!zp->zp_dedup);
1119 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1120 return (ZIO_PIPELINE_CONTINUE);
1121 }
1122
1123 ASSERT(!zp->zp_nopwrite);
1124
1125 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1126 return (ZIO_PIPELINE_CONTINUE);
1127
1128 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1129 zp->zp_dedup_verify);
1130
1131 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1132 BP_SET_DEDUP(bp, 1);
1133 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1134 return (ZIO_PIPELINE_CONTINUE);
1135 }
1136 }
1137
1138 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1139 /*
1140 * We're rewriting an existing block, which means we're
1141 * working on behalf of spa_sync(). For spa_sync() to
1142 * converge, it must eventually be the case that we don't
1143 * have to allocate new blocks. But compression changes
1144 * the blocksize, which forces a reallocate, and makes
1145 * convergence take longer. Therefore, after the first
1146 * few passes, stop compressing to ensure convergence.
1147 */
1148 pass = spa_sync_pass(spa);
1149
1150 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1151 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1152 ASSERT(!BP_GET_DEDUP(bp));
1153
1154 if (pass >= zfs_sync_pass_dont_compress)
1155 compress = ZIO_COMPRESS_OFF;
1156
1157 /* Make sure someone doesn't change their mind on overwrites */
1158 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1159 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1160 }
1161
1162 if (compress != ZIO_COMPRESS_OFF) {
1163 void *cbuf = zio_buf_alloc(lsize);
1164 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1165 if (psize == 0 || psize == lsize) {
1166 compress = ZIO_COMPRESS_OFF;
1167 zio_buf_free(cbuf, lsize);
1168 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1169 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1170 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1171 encode_embedded_bp_compressed(bp,
1172 cbuf, compress, lsize, psize);
1173 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1174 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1175 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1176 zio_buf_free(cbuf, lsize);
1177 bp->blk_birth = zio->io_txg;
1178 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1179 ASSERT(spa_feature_is_active(spa,
1180 SPA_FEATURE_EMBEDDED_DATA));
1181 return (ZIO_PIPELINE_CONTINUE);
1182 } else {
1183 /*
1184 * Round up compressed size to MINBLOCKSIZE and
1185 * zero the tail.
1186 */
1187 size_t rounded =
1188 P2ROUNDUP(psize, (size_t)SPA_MINBLOCKSIZE);
1189 if (rounded > psize) {
1190 bzero((char *)cbuf + psize, rounded - psize);
1191 psize = rounded;
1192 }
1193 if (psize == lsize) {
1194 compress = ZIO_COMPRESS_OFF;
1195 zio_buf_free(cbuf, lsize);
1196 } else {
1197 zio_push_transform(zio, cbuf,
1198 psize, lsize, NULL);
1199 }
1200 }
1201 }
1202
1203 /*
1204 * The final pass of spa_sync() must be all rewrites, but the first
1205 * few passes offer a trade-off: allocating blocks defers convergence,
1206 * but newly allocated blocks are sequential, so they can be written
1207 * to disk faster. Therefore, we allow the first few passes of
1208 * spa_sync() to allocate new blocks, but force rewrites after that.
1209 * There should only be a handful of blocks after pass 1 in any case.
1210 */
1211 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1212 BP_GET_PSIZE(bp) == psize &&
1213 pass >= zfs_sync_pass_rewrite) {
1214 ASSERT(psize != 0);
1215 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1216 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1217 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1218 } else {
1219 BP_ZERO(bp);
1220 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1221 }
1222
1223 if (psize == 0) {
1224 if (zio->io_bp_orig.blk_birth != 0 &&
1225 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1226 BP_SET_LSIZE(bp, lsize);
1227 BP_SET_TYPE(bp, zp->zp_type);
1228 BP_SET_LEVEL(bp, zp->zp_level);
1229 BP_SET_BIRTH(bp, zio->io_txg, 0);
1230 }
1231 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1232 } else {
1233 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1234 BP_SET_LSIZE(bp, lsize);
1235 BP_SET_TYPE(bp, zp->zp_type);
1236 BP_SET_LEVEL(bp, zp->zp_level);
1237 BP_SET_PSIZE(bp, psize);
1238 BP_SET_COMPRESS(bp, compress);
1239 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1240 BP_SET_DEDUP(bp, zp->zp_dedup);
1241 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1242 if (zp->zp_dedup) {
1243 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1244 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1245 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1246 }
1247 if (zp->zp_nopwrite) {
1248 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1249 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1250 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1251 }
1252 }
1253
1254 return (ZIO_PIPELINE_CONTINUE);
1255 }
1256
1257 static int
1258 zio_free_bp_init(zio_t *zio)
1259 {
1260 blkptr_t *bp = zio->io_bp;
1261
1262 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1263 if (BP_GET_DEDUP(bp))
1264 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1265 }
1266
1267 return (ZIO_PIPELINE_CONTINUE);
1268 }
1269
1270 /*
1271 * ==========================================================================
1272 * Execute the I/O pipeline
1273 * ==========================================================================
1274 */
1275
1276 static void
1277 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1278 {
1279 spa_t *spa = zio->io_spa;
1280 zio_type_t t = zio->io_type;
1281 int flags = (cutinline ? TQ_FRONT : 0);
1282
1283 /*
1284 * If we're a config writer or a probe, the normal issue and
1285 * interrupt threads may all be blocked waiting for the config lock.
1286 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1287 */
1288 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1289 t = ZIO_TYPE_NULL;
1290
1291 /*
1292 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1293 */
1294 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1295 t = ZIO_TYPE_NULL;
1296
1297 /*
1298 * If this is a high priority I/O, then use the high priority taskq if
1299 * available.
1300 */
1301 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1302 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1303 q++;
1304
1305 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1306
1307 /*
1308 * NB: We are assuming that the zio can only be dispatched
1309 * to a single taskq at a time. It would be a grievous error
1310 * to dispatch the zio to another taskq at the same time.
1311 */
1312 ASSERT(zio->io_tqent.tqent_next == NULL);
1313 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1314 flags, &zio->io_tqent);
1315 }
1316
1317 static boolean_t
1318 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1319 {
1320 kthread_t *executor = zio->io_executor;
1321 spa_t *spa = zio->io_spa;
1322
1323 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1324 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1325 uint_t i;
1326 for (i = 0; i < tqs->stqs_count; i++) {
1327 if (taskq_member(tqs->stqs_taskq[i], executor))
1328 return (B_TRUE);
1329 }
1330 }
1331
1332 return (B_FALSE);
1333 }
1334
1335 static int
1336 zio_issue_async(zio_t *zio)
1337 {
1338 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1339
1340 return (ZIO_PIPELINE_STOP);
1341 }
1342
1343 void
1344 zio_interrupt(zio_t *zio)
1345 {
1346 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1347 }
1348
1349 /*
1350 * Execute the I/O pipeline until one of the following occurs:
1351 *
1352 * (1) the I/O completes
1353 * (2) the pipeline stalls waiting for dependent child I/Os
1354 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1355 * (4) the I/O is delegated by vdev-level caching or aggregation
1356 * (5) the I/O is deferred due to vdev-level queueing
1357 * (6) the I/O is handed off to another thread.
1358 *
1359 * In all cases, the pipeline stops whenever there's no CPU work; it never
1360 * burns a thread in cv_wait().
1361 *
1362 * There's no locking on io_stage because there's no legitimate way
1363 * for multiple threads to be attempting to process the same I/O.
1364 */
1365 static zio_pipe_stage_t *zio_pipeline[];
1366
1367 void
1368 zio_execute(zio_t *zio)
1369 {
1370 zio->io_executor = curthread;
1371
1372 while (zio->io_stage < ZIO_STAGE_DONE) {
1373 enum zio_stage pipeline = zio->io_pipeline;
1374 enum zio_stage stage = zio->io_stage;
1375 int rv;
1376
1377 ASSERT(!MUTEX_HELD(&zio->io_lock));
1378 ASSERT(ISP2(stage));
1379 ASSERT(zio->io_stall == NULL);
1380
1381 do {
1382 stage <<= 1;
1383 } while ((stage & pipeline) == 0);
1384
1385 ASSERT(stage <= ZIO_STAGE_DONE);
1386
1387 /*
1388 * If we are in interrupt context and this pipeline stage
1389 * will grab a config lock that is held across I/O,
1390 * or may wait for an I/O that needs an interrupt thread
1391 * to complete, issue async to avoid deadlock.
1392 *
1393 * For VDEV_IO_START, we cut in line so that the io will
1394 * be sent to disk promptly.
1395 */
1396 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1397 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1398 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1399 zio_requeue_io_start_cut_in_line : B_FALSE;
1400 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1401 return;
1402 }
1403
1404 zio->io_stage = stage;
1405 rv = zio_pipeline[highbit64(stage) - 1](zio);
1406
1407 if (rv == ZIO_PIPELINE_STOP)
1408 return;
1409
1410 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1411 }
1412 }
1413
1414 /*
1415 * ==========================================================================
1416 * Initiate I/O, either sync or async
1417 * ==========================================================================
1418 */
1419 int
1420 zio_wait(zio_t *zio)
1421 {
1422 int error;
1423
1424 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1425 ASSERT(zio->io_executor == NULL);
1426
1427 zio->io_waiter = curthread;
1428
1429 zio_execute(zio);
1430
1431 mutex_enter(&zio->io_lock);
1432 while (zio->io_executor != NULL)
1433 cv_wait(&zio->io_cv, &zio->io_lock);
1434 mutex_exit(&zio->io_lock);
1435
1436 error = zio->io_error;
1437 zio_destroy(zio);
1438
1439 return (error);
1440 }
1441
1442 void
1443 zio_nowait(zio_t *zio)
1444 {
1445 ASSERT(zio->io_executor == NULL);
1446
1447 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1448 zio_unique_parent(zio) == NULL) {
1449 /*
1450 * This is a logical async I/O with no parent to wait for it.
1451 * We add it to the spa_async_root_zio "Godfather" I/O which
1452 * will ensure they complete prior to unloading the pool.
1453 */
1454 spa_t *spa = zio->io_spa;
1455
1456 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1457 }
1458
1459 zio_execute(zio);
1460 }
1461
1462 /*
1463 * ==========================================================================
1464 * Reexecute or suspend/resume failed I/O
1465 * ==========================================================================
1466 */
1467
1468 static void
1469 zio_reexecute(zio_t *pio)
1470 {
1471 zio_t *cio, *cio_next;
1472
1473 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1474 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1475 ASSERT(pio->io_gang_leader == NULL);
1476 ASSERT(pio->io_gang_tree == NULL);
1477
1478 pio->io_flags = pio->io_orig_flags;
1479 pio->io_stage = pio->io_orig_stage;
1480 pio->io_pipeline = pio->io_orig_pipeline;
1481 pio->io_reexecute = 0;
1482 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1483 pio->io_error = 0;
1484 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1485 pio->io_state[w] = 0;
1486 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1487 pio->io_child_error[c] = 0;
1488
1489 if (IO_IS_ALLOCATING(pio))
1490 BP_ZERO(pio->io_bp);
1491
1492 /*
1493 * As we reexecute pio's children, new children could be created.
1494 * New children go to the head of pio's io_child_list, however,
1495 * so we will (correctly) not reexecute them. The key is that
1496 * the remainder of pio's io_child_list, from 'cio_next' onward,
1497 * cannot be affected by any side effects of reexecuting 'cio'.
1498 */
1499 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1500 cio_next = zio_walk_children(pio);
1501 mutex_enter(&pio->io_lock);
1502 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1503 pio->io_children[cio->io_child_type][w]++;
1504 mutex_exit(&pio->io_lock);
1505 zio_reexecute(cio);
1506 }
1507
1508 /*
1509 * Now that all children have been reexecuted, execute the parent.
1510 * We don't reexecute "The Godfather" I/O here as it's the
1511 * responsibility of the caller to wait on him.
1512 */
1513 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1514 zio_execute(pio);
1515 }
1516
1517 void
1518 zio_suspend(spa_t *spa, zio_t *zio)
1519 {
1520 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1521 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1522 "failure and the failure mode property for this pool "
1523 "is set to panic.", spa_name(spa));
1524
1525 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1526
1527 mutex_enter(&spa->spa_suspend_lock);
1528
1529 if (spa->spa_suspend_zio_root == NULL)
1530 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1531 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1532 ZIO_FLAG_GODFATHER);
1533
1534 spa->spa_suspended = B_TRUE;
1535
1536 if (zio != NULL) {
1537 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1538 ASSERT(zio != spa->spa_suspend_zio_root);
1539 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1540 ASSERT(zio_unique_parent(zio) == NULL);
1541 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1542 zio_add_child(spa->spa_suspend_zio_root, zio);
1543 }
1544
1545 mutex_exit(&spa->spa_suspend_lock);
1546 }
1547
1548 int
1549 zio_resume(spa_t *spa)
1550 {
1551 zio_t *pio;
1552
1553 /*
1554 * Reexecute all previously suspended i/o.
1555 */
1556 mutex_enter(&spa->spa_suspend_lock);
1557 spa->spa_suspended = B_FALSE;
1558 cv_broadcast(&spa->spa_suspend_cv);
1559 pio = spa->spa_suspend_zio_root;
1560 spa->spa_suspend_zio_root = NULL;
1561 mutex_exit(&spa->spa_suspend_lock);
1562
1563 if (pio == NULL)
1564 return (0);
1565
1566 zio_reexecute(pio);
1567 return (zio_wait(pio));
1568 }
1569
1570 void
1571 zio_resume_wait(spa_t *spa)
1572 {
1573 mutex_enter(&spa->spa_suspend_lock);
1574 while (spa_suspended(spa))
1575 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1576 mutex_exit(&spa->spa_suspend_lock);
1577 }
1578
1579 /*
1580 * ==========================================================================
1581 * Gang blocks.
1582 *
1583 * A gang block is a collection of small blocks that looks to the DMU
1584 * like one large block. When zio_dva_allocate() cannot find a block
1585 * of the requested size, due to either severe fragmentation or the pool
1586 * being nearly full, it calls zio_write_gang_block() to construct the
1587 * block from smaller fragments.
1588 *
1589 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1590 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1591 * an indirect block: it's an array of block pointers. It consumes
1592 * only one sector and hence is allocatable regardless of fragmentation.
1593 * The gang header's bps point to its gang members, which hold the data.
1594 *
1595 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1596 * as the verifier to ensure uniqueness of the SHA256 checksum.
1597 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1598 * not the gang header. This ensures that data block signatures (needed for
1599 * deduplication) are independent of how the block is physically stored.
1600 *
1601 * Gang blocks can be nested: a gang member may itself be a gang block.
1602 * Thus every gang block is a tree in which root and all interior nodes are
1603 * gang headers, and the leaves are normal blocks that contain user data.
1604 * The root of the gang tree is called the gang leader.
1605 *
1606 * To perform any operation (read, rewrite, free, claim) on a gang block,
1607 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1608 * in the io_gang_tree field of the original logical i/o by recursively
1609 * reading the gang leader and all gang headers below it. This yields
1610 * an in-core tree containing the contents of every gang header and the
1611 * bps for every constituent of the gang block.
1612 *
1613 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1614 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1615 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1616 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1617 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1618 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1619 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1620 * of the gang header plus zio_checksum_compute() of the data to update the
1621 * gang header's blk_cksum as described above.
1622 *
1623 * The two-phase assemble/issue model solves the problem of partial failure --
1624 * what if you'd freed part of a gang block but then couldn't read the
1625 * gang header for another part? Assembling the entire gang tree first
1626 * ensures that all the necessary gang header I/O has succeeded before
1627 * starting the actual work of free, claim, or write. Once the gang tree
1628 * is assembled, free and claim are in-memory operations that cannot fail.
1629 *
1630 * In the event that a gang write fails, zio_dva_unallocate() walks the
1631 * gang tree to immediately free (i.e. insert back into the space map)
1632 * everything we've allocated. This ensures that we don't get ENOSPC
1633 * errors during repeated suspend/resume cycles due to a flaky device.
1634 *
1635 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1636 * the gang tree, we won't modify the block, so we can safely defer the free
1637 * (knowing that the block is still intact). If we *can* assemble the gang
1638 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1639 * each constituent bp and we can allocate a new block on the next sync pass.
1640 *
1641 * In all cases, the gang tree allows complete recovery from partial failure.
1642 * ==========================================================================
1643 */
1644
1645 static zio_t *
1646 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1647 {
1648 if (gn != NULL)
1649 return (pio);
1650
1651 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1652 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1653 &pio->io_bookmark));
1654 }
1655
1656 zio_t *
1657 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1658 {
1659 zio_t *zio;
1660
1661 if (gn != NULL) {
1662 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1663 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1664 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1665 /*
1666 * As we rewrite each gang header, the pipeline will compute
1667 * a new gang block header checksum for it; but no one will
1668 * compute a new data checksum, so we do that here. The one
1669 * exception is the gang leader: the pipeline already computed
1670 * its data checksum because that stage precedes gang assembly.
1671 * (Presently, nothing actually uses interior data checksums;
1672 * this is just good hygiene.)
1673 */
1674 if (gn != pio->io_gang_leader->io_gang_tree) {
1675 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1676 data, BP_GET_PSIZE(bp));
1677 }
1678 /*
1679 * If we are here to damage data for testing purposes,
1680 * leave the GBH alone so that we can detect the damage.
1681 */
1682 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1683 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1684 } else {
1685 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1686 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1687 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1688 }
1689
1690 return (zio);
1691 }
1692
1693 /* ARGSUSED */
1694 zio_t *
1695 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1696 {
1697 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1698 ZIO_GANG_CHILD_FLAGS(pio)));
1699 }
1700
1701 /* ARGSUSED */
1702 zio_t *
1703 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1704 {
1705 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1706 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1707 }
1708
1709 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1710 NULL,
1711 zio_read_gang,
1712 zio_rewrite_gang,
1713 zio_free_gang,
1714 zio_claim_gang,
1715 NULL
1716 };
1717
1718 static void zio_gang_tree_assemble_done(zio_t *zio);
1719
1720 static zio_gang_node_t *
1721 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1722 {
1723 zio_gang_node_t *gn;
1724
1725 ASSERT(*gnpp == NULL);
1726
1727 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1728 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1729 *gnpp = gn;
1730
1731 return (gn);
1732 }
1733
1734 static void
1735 zio_gang_node_free(zio_gang_node_t **gnpp)
1736 {
1737 zio_gang_node_t *gn = *gnpp;
1738
1739 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1740 ASSERT(gn->gn_child[g] == NULL);
1741
1742 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1743 kmem_free(gn, sizeof (*gn));
1744 *gnpp = NULL;
1745 }
1746
1747 static void
1748 zio_gang_tree_free(zio_gang_node_t **gnpp)
1749 {
1750 zio_gang_node_t *gn = *gnpp;
1751
1752 if (gn == NULL)
1753 return;
1754
1755 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1756 zio_gang_tree_free(&gn->gn_child[g]);
1757
1758 zio_gang_node_free(gnpp);
1759 }
1760
1761 static void
1762 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1763 {
1764 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1765
1766 ASSERT(gio->io_gang_leader == gio);
1767 ASSERT(BP_IS_GANG(bp));
1768
1769 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1770 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1771 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1772 }
1773
1774 static void
1775 zio_gang_tree_assemble_done(zio_t *zio)
1776 {
1777 zio_t *gio = zio->io_gang_leader;
1778 zio_gang_node_t *gn = zio->io_private;
1779 blkptr_t *bp = zio->io_bp;
1780
1781 ASSERT(gio == zio_unique_parent(zio));
1782 ASSERT(zio->io_child_count == 0);
1783
1784 if (zio->io_error)
1785 return;
1786
1787 if (BP_SHOULD_BYTESWAP(bp))
1788 byteswap_uint64_array(zio->io_data, zio->io_size);
1789
1790 ASSERT(zio->io_data == gn->gn_gbh);
1791 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1792 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1793
1794 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1795 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1796 if (!BP_IS_GANG(gbp))
1797 continue;
1798 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1799 }
1800 }
1801
1802 static void
1803 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1804 {
1805 zio_t *gio = pio->io_gang_leader;
1806 zio_t *zio;
1807
1808 ASSERT(BP_IS_GANG(bp) == !!gn);
1809 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1810 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1811
1812 /*
1813 * If you're a gang header, your data is in gn->gn_gbh.
1814 * If you're a gang member, your data is in 'data' and gn == NULL.
1815 */
1816 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1817
1818 if (gn != NULL) {
1819 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1820
1821 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1822 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1823 if (BP_IS_HOLE(gbp))
1824 continue;
1825 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1826 data = (char *)data + BP_GET_PSIZE(gbp);
1827 }
1828 }
1829
1830 if (gn == gio->io_gang_tree)
1831 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1832
1833 if (zio != pio)
1834 zio_nowait(zio);
1835 }
1836
1837 static int
1838 zio_gang_assemble(zio_t *zio)
1839 {
1840 blkptr_t *bp = zio->io_bp;
1841
1842 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1843 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1844
1845 zio->io_gang_leader = zio;
1846
1847 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1848
1849 return (ZIO_PIPELINE_CONTINUE);
1850 }
1851
1852 static int
1853 zio_gang_issue(zio_t *zio)
1854 {
1855 blkptr_t *bp = zio->io_bp;
1856
1857 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1858 return (ZIO_PIPELINE_STOP);
1859
1860 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1861 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1862
1863 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1864 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1865 else
1866 zio_gang_tree_free(&zio->io_gang_tree);
1867
1868 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1869
1870 return (ZIO_PIPELINE_CONTINUE);
1871 }
1872
1873 static void
1874 zio_write_gang_member_ready(zio_t *zio)
1875 {
1876 zio_t *pio = zio_unique_parent(zio);
1877 zio_t *gio = zio->io_gang_leader;
1878 dva_t *cdva = zio->io_bp->blk_dva;
1879 dva_t *pdva = pio->io_bp->blk_dva;
1880 uint64_t asize;
1881
1882 if (BP_IS_HOLE(zio->io_bp))
1883 return;
1884
1885 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1886
1887 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1888 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1889 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1890 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1891 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1892
1893 mutex_enter(&pio->io_lock);
1894 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1895 ASSERT(DVA_GET_GANG(&pdva[d]));
1896 asize = DVA_GET_ASIZE(&pdva[d]);
1897 asize += DVA_GET_ASIZE(&cdva[d]);
1898 DVA_SET_ASIZE(&pdva[d], asize);
1899 }
1900 mutex_exit(&pio->io_lock);
1901 }
1902
1903 static int
1904 zio_write_gang_block(zio_t *pio)
1905 {
1906 spa_t *spa = pio->io_spa;
1907 blkptr_t *bp = pio->io_bp;
1908 zio_t *gio = pio->io_gang_leader;
1909 zio_t *zio;
1910 zio_gang_node_t *gn, **gnpp;
1911 zio_gbh_phys_t *gbh;
1912 uint64_t txg = pio->io_txg;
1913 uint64_t resid = pio->io_size;
1914 uint64_t lsize;
1915 int copies = gio->io_prop.zp_copies;
1916 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1917 zio_prop_t zp;
1918 int error;
1919
1920 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1921 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1922 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1923 if (error) {
1924 pio->io_error = error;
1925 return (ZIO_PIPELINE_CONTINUE);
1926 }
1927
1928 if (pio == gio) {
1929 gnpp = &gio->io_gang_tree;
1930 } else {
1931 gnpp = pio->io_private;
1932 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1933 }
1934
1935 gn = zio_gang_node_alloc(gnpp);
1936 gbh = gn->gn_gbh;
1937 bzero(gbh, SPA_GANGBLOCKSIZE);
1938
1939 /*
1940 * Create the gang header.
1941 */
1942 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1943 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1944
1945 /*
1946 * Create and nowait the gang children.
1947 */
1948 for (int g = 0; resid != 0; resid -= lsize, g++) {
1949 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1950 SPA_MINBLOCKSIZE);
1951 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1952
1953 zp.zp_checksum = gio->io_prop.zp_checksum;
1954 zp.zp_compress = ZIO_COMPRESS_OFF;
1955 zp.zp_type = DMU_OT_NONE;
1956 zp.zp_level = 0;
1957 zp.zp_copies = gio->io_prop.zp_copies;
1958 zp.zp_dedup = B_FALSE;
1959 zp.zp_dedup_verify = B_FALSE;
1960 zp.zp_nopwrite = B_FALSE;
1961
1962 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1963 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1964 zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g],
1965 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1966 &pio->io_bookmark));
1967 }
1968
1969 /*
1970 * Set pio's pipeline to just wait for zio to finish.
1971 */
1972 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1973
1974 zio_nowait(zio);
1975
1976 return (ZIO_PIPELINE_CONTINUE);
1977 }
1978
1979 /*
1980 * The zio_nop_write stage in the pipeline determines if allocating
1981 * a new bp is necessary. By leveraging a cryptographically secure checksum,
1982 * such as SHA256, we can compare the checksums of the new data and the old
1983 * to determine if allocating a new block is required. The nopwrite
1984 * feature can handle writes in either syncing or open context (i.e. zil
1985 * writes) and as a result is mutually exclusive with dedup.
1986 */
1987 static int
1988 zio_nop_write(zio_t *zio)
1989 {
1990 blkptr_t *bp = zio->io_bp;
1991 blkptr_t *bp_orig = &zio->io_bp_orig;
1992 zio_prop_t *zp = &zio->io_prop;
1993
1994 ASSERT(BP_GET_LEVEL(bp) == 0);
1995 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1996 ASSERT(zp->zp_nopwrite);
1997 ASSERT(!zp->zp_dedup);
1998 ASSERT(zio->io_bp_override == NULL);
1999 ASSERT(IO_IS_ALLOCATING(zio));
2000
2001 /*
2002 * Check to see if the original bp and the new bp have matching
2003 * characteristics (i.e. same checksum, compression algorithms, etc).
2004 * If they don't then just continue with the pipeline which will
2005 * allocate a new bp.
2006 */
2007 if (BP_IS_HOLE(bp_orig) ||
2008 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
2009 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2010 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2011 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2012 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2013 return (ZIO_PIPELINE_CONTINUE);
2014
2015 /*
2016 * If the checksums match then reset the pipeline so that we
2017 * avoid allocating a new bp and issuing any I/O.
2018 */
2019 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2020 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
2021 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2022 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2023 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2024 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2025 sizeof (uint64_t)) == 0);
2026
2027 *bp = *bp_orig;
2028 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2029 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2030 }
2031
2032 return (ZIO_PIPELINE_CONTINUE);
2033 }
2034
2035 /*
2036 * ==========================================================================
2037 * Dedup
2038 * ==========================================================================
2039 */
2040 static void
2041 zio_ddt_child_read_done(zio_t *zio)
2042 {
2043 blkptr_t *bp = zio->io_bp;
2044 ddt_entry_t *dde = zio->io_private;
2045 ddt_phys_t *ddp;
2046 zio_t *pio = zio_unique_parent(zio);
2047
2048 mutex_enter(&pio->io_lock);
2049 ddp = ddt_phys_select(dde, bp);
2050 if (zio->io_error == 0)
2051 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2052 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
2053 dde->dde_repair_data = zio->io_data;
2054 else
2055 zio_buf_free(zio->io_data, zio->io_size);
2056 mutex_exit(&pio->io_lock);
2057 }
2058
2059 static int
2060 zio_ddt_read_start(zio_t *zio)
2061 {
2062 blkptr_t *bp = zio->io_bp;
2063
2064 ASSERT(BP_GET_DEDUP(bp));
2065 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2066 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2067
2068 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2069 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2070 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2071 ddt_phys_t *ddp = dde->dde_phys;
2072 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2073 blkptr_t blk;
2074
2075 ASSERT(zio->io_vsd == NULL);
2076 zio->io_vsd = dde;
2077
2078 if (ddp_self == NULL)
2079 return (ZIO_PIPELINE_CONTINUE);
2080
2081 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2082 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2083 continue;
2084 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2085 &blk);
2086 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2087 zio_buf_alloc(zio->io_size), zio->io_size,
2088 zio_ddt_child_read_done, dde, zio->io_priority,
2089 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2090 &zio->io_bookmark));
2091 }
2092 return (ZIO_PIPELINE_CONTINUE);
2093 }
2094
2095 zio_nowait(zio_read(zio, zio->io_spa, bp,
2096 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2097 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2098
2099 return (ZIO_PIPELINE_CONTINUE);
2100 }
2101
2102 static int
2103 zio_ddt_read_done(zio_t *zio)
2104 {
2105 blkptr_t *bp = zio->io_bp;
2106
2107 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2108 return (ZIO_PIPELINE_STOP);
2109
2110 ASSERT(BP_GET_DEDUP(bp));
2111 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2112 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2113
2114 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2115 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2116 ddt_entry_t *dde = zio->io_vsd;
2117 if (ddt == NULL) {
2118 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2119 return (ZIO_PIPELINE_CONTINUE);
2120 }
2121 if (dde == NULL) {
2122 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2123 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2124 return (ZIO_PIPELINE_STOP);
2125 }
2126 if (dde->dde_repair_data != NULL) {
2127 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2128 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2129 }
2130 ddt_repair_done(ddt, dde);
2131 zio->io_vsd = NULL;
2132 }
2133
2134 ASSERT(zio->io_vsd == NULL);
2135
2136 return (ZIO_PIPELINE_CONTINUE);
2137 }
2138
2139 static boolean_t
2140 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2141 {
2142 spa_t *spa = zio->io_spa;
2143
2144 /*
2145 * Note: we compare the original data, not the transformed data,
2146 * because when zio->io_bp is an override bp, we will not have
2147 * pushed the I/O transforms. That's an important optimization
2148 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2149 */
2150 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2151 zio_t *lio = dde->dde_lead_zio[p];
2152
2153 if (lio != NULL) {
2154 return (lio->io_orig_size != zio->io_orig_size ||
2155 bcmp(zio->io_orig_data, lio->io_orig_data,
2156 zio->io_orig_size) != 0);
2157 }
2158 }
2159
2160 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2161 ddt_phys_t *ddp = &dde->dde_phys[p];
2162
2163 if (ddp->ddp_phys_birth != 0) {
2164 arc_buf_t *abuf = NULL;
2165 arc_flags_t aflags = ARC_FLAG_WAIT;
2166 blkptr_t blk = *zio->io_bp;
2167 int error;
2168
2169 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2170
2171 ddt_exit(ddt);
2172
2173 error = arc_read(NULL, spa, &blk,
2174 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2175 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2176 &aflags, &zio->io_bookmark);
2177
2178 if (error == 0) {
2179 if (arc_buf_size(abuf) != zio->io_orig_size ||
2180 bcmp(abuf->b_data, zio->io_orig_data,
2181 zio->io_orig_size) != 0)
2182 error = SET_ERROR(EEXIST);
2183 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2184 }
2185
2186 ddt_enter(ddt);
2187 return (error != 0);
2188 }
2189 }
2190
2191 return (B_FALSE);
2192 }
2193
2194 static void
2195 zio_ddt_child_write_ready(zio_t *zio)
2196 {
2197 int p = zio->io_prop.zp_copies;
2198 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2199 ddt_entry_t *dde = zio->io_private;
2200 ddt_phys_t *ddp = &dde->dde_phys[p];
2201 zio_t *pio;
2202
2203 if (zio->io_error)
2204 return;
2205
2206 ddt_enter(ddt);
2207
2208 ASSERT(dde->dde_lead_zio[p] == zio);
2209
2210 ddt_phys_fill(ddp, zio->io_bp);
2211
2212 while ((pio = zio_walk_parents(zio)) != NULL)
2213 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2214
2215 ddt_exit(ddt);
2216 }
2217
2218 static void
2219 zio_ddt_child_write_done(zio_t *zio)
2220 {
2221 int p = zio->io_prop.zp_copies;
2222 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2223 ddt_entry_t *dde = zio->io_private;
2224 ddt_phys_t *ddp = &dde->dde_phys[p];
2225
2226 ddt_enter(ddt);
2227
2228 ASSERT(ddp->ddp_refcnt == 0);
2229 ASSERT(dde->dde_lead_zio[p] == zio);
2230 dde->dde_lead_zio[p] = NULL;
2231
2232 if (zio->io_error == 0) {
2233 while (zio_walk_parents(zio) != NULL)
2234 ddt_phys_addref(ddp);
2235 } else {
2236 ddt_phys_clear(ddp);
2237 }
2238
2239 ddt_exit(ddt);
2240 }
2241
2242 static void
2243 zio_ddt_ditto_write_done(zio_t *zio)
2244 {
2245 int p = DDT_PHYS_DITTO;
2246 zio_prop_t *zp = &zio->io_prop;
2247 blkptr_t *bp = zio->io_bp;
2248 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2249 ddt_entry_t *dde = zio->io_private;
2250 ddt_phys_t *ddp = &dde->dde_phys[p];
2251 ddt_key_t *ddk = &dde->dde_key;
2252
2253 ddt_enter(ddt);
2254
2255 ASSERT(ddp->ddp_refcnt == 0);
2256 ASSERT(dde->dde_lead_zio[p] == zio);
2257 dde->dde_lead_zio[p] = NULL;
2258
2259 if (zio->io_error == 0) {
2260 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2261 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2262 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2263 if (ddp->ddp_phys_birth != 0)
2264 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2265 ddt_phys_fill(ddp, bp);
2266 }
2267
2268 ddt_exit(ddt);
2269 }
2270
2271 static int
2272 zio_ddt_write(zio_t *zio)
2273 {
2274 spa_t *spa = zio->io_spa;
2275 blkptr_t *bp = zio->io_bp;
2276 uint64_t txg = zio->io_txg;
2277 zio_prop_t *zp = &zio->io_prop;
2278 int p = zp->zp_copies;
2279 int ditto_copies;
2280 zio_t *cio = NULL;
2281 zio_t *dio = NULL;
2282 ddt_t *ddt = ddt_select(spa, bp);
2283 ddt_entry_t *dde;
2284 ddt_phys_t *ddp;
2285
2286 ASSERT(BP_GET_DEDUP(bp));
2287 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2288 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2289
2290 ddt_enter(ddt);
2291 dde = ddt_lookup(ddt, bp, B_TRUE);
2292 ddp = &dde->dde_phys[p];
2293
2294 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2295 /*
2296 * If we're using a weak checksum, upgrade to a strong checksum
2297 * and try again. If we're already using a strong checksum,
2298 * we can't resolve it, so just convert to an ordinary write.
2299 * (And automatically e-mail a paper to Nature?)
2300 */
2301 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2302 zp->zp_checksum = spa_dedup_checksum(spa);
2303 zio_pop_transforms(zio);
2304 zio->io_stage = ZIO_STAGE_OPEN;
2305 BP_ZERO(bp);
2306 } else {
2307 zp->zp_dedup = B_FALSE;
2308 }
2309 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2310 ddt_exit(ddt);
2311 return (ZIO_PIPELINE_CONTINUE);
2312 }
2313
2314 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2315 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2316
2317 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2318 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2319 zio_prop_t czp = *zp;
2320
2321 czp.zp_copies = ditto_copies;
2322
2323 /*
2324 * If we arrived here with an override bp, we won't have run
2325 * the transform stack, so we won't have the data we need to
2326 * generate a child i/o. So, toss the override bp and restart.
2327 * This is safe, because using the override bp is just an
2328 * optimization; and it's rare, so the cost doesn't matter.
2329 */
2330 if (zio->io_bp_override) {
2331 zio_pop_transforms(zio);
2332 zio->io_stage = ZIO_STAGE_OPEN;
2333 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2334 zio->io_bp_override = NULL;
2335 BP_ZERO(bp);
2336 ddt_exit(ddt);
2337 return (ZIO_PIPELINE_CONTINUE);
2338 }
2339
2340 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2341 zio->io_orig_size, &czp, NULL, NULL,
2342 zio_ddt_ditto_write_done, dde, zio->io_priority,
2343 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2344
2345 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2346 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2347 }
2348
2349 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2350 if (ddp->ddp_phys_birth != 0)
2351 ddt_bp_fill(ddp, bp, txg);
2352 if (dde->dde_lead_zio[p] != NULL)
2353 zio_add_child(zio, dde->dde_lead_zio[p]);
2354 else
2355 ddt_phys_addref(ddp);
2356 } else if (zio->io_bp_override) {
2357 ASSERT(bp->blk_birth == txg);
2358 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2359 ddt_phys_fill(ddp, bp);
2360 ddt_phys_addref(ddp);
2361 } else {
2362 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2363 zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL,
2364 zio_ddt_child_write_done, dde, zio->io_priority,
2365 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2366
2367 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2368 dde->dde_lead_zio[p] = cio;
2369 }
2370
2371 ddt_exit(ddt);
2372
2373 if (cio)
2374 zio_nowait(cio);
2375 if (dio)
2376 zio_nowait(dio);
2377
2378 return (ZIO_PIPELINE_CONTINUE);
2379 }
2380
2381 ddt_entry_t *freedde; /* for debugging */
2382
2383 static int
2384 zio_ddt_free(zio_t *zio)
2385 {
2386 spa_t *spa = zio->io_spa;
2387 blkptr_t *bp = zio->io_bp;
2388 ddt_t *ddt = ddt_select(spa, bp);
2389 ddt_entry_t *dde;
2390 ddt_phys_t *ddp;
2391
2392 ASSERT(BP_GET_DEDUP(bp));
2393 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2394
2395 ddt_enter(ddt);
2396 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2397 ddp = ddt_phys_select(dde, bp);
2398 ddt_phys_decref(ddp);
2399 ddt_exit(ddt);
2400
2401 return (ZIO_PIPELINE_CONTINUE);
2402 }
2403
2404 /*
2405 * ==========================================================================
2406 * Allocate and free blocks
2407 * ==========================================================================
2408 */
2409 static int
2410 zio_dva_allocate(zio_t *zio)
2411 {
2412 spa_t *spa = zio->io_spa;
2413 metaslab_class_t *mc = spa_normal_class(spa);
2414 blkptr_t *bp = zio->io_bp;
2415 int error;
2416 int flags = 0;
2417
2418 if (zio->io_gang_leader == NULL) {
2419 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2420 zio->io_gang_leader = zio;
2421 }
2422
2423 ASSERT(BP_IS_HOLE(bp));
2424 ASSERT0(BP_GET_NDVAS(bp));
2425 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2426 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2427 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2428
2429 /*
2430 * The dump device does not support gang blocks so allocation on
2431 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2432 * the "fast" gang feature.
2433 */
2434 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2435 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2436 METASLAB_GANG_CHILD : 0;
2437 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2438 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2439
2440 if (error) {
2441 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2442 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2443 error);
2444 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2445 return (zio_write_gang_block(zio));
2446 zio->io_error = error;
2447 }
2448
2449 return (ZIO_PIPELINE_CONTINUE);
2450 }
2451
2452 static int
2453 zio_dva_free(zio_t *zio)
2454 {
2455 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2456
2457 return (ZIO_PIPELINE_CONTINUE);
2458 }
2459
2460 static int
2461 zio_dva_claim(zio_t *zio)
2462 {
2463 int error;
2464
2465 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2466 if (error)
2467 zio->io_error = error;
2468
2469 return (ZIO_PIPELINE_CONTINUE);
2470 }
2471
2472 /*
2473 * Undo an allocation. This is used by zio_done() when an I/O fails
2474 * and we want to give back the block we just allocated.
2475 * This handles both normal blocks and gang blocks.
2476 */
2477 static void
2478 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2479 {
2480 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2481 ASSERT(zio->io_bp_override == NULL);
2482
2483 if (!BP_IS_HOLE(bp))
2484 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2485
2486 if (gn != NULL) {
2487 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2488 zio_dva_unallocate(zio, gn->gn_child[g],
2489 &gn->gn_gbh->zg_blkptr[g]);
2490 }
2491 }
2492 }
2493
2494 /*
2495 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2496 */
2497 int
2498 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2499 uint64_t size, boolean_t use_slog)
2500 {
2501 int error = 1;
2502
2503 ASSERT(txg > spa_syncing_txg(spa));
2504
2505 /*
2506 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2507 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2508 * when allocating them.
2509 */
2510 if (use_slog) {
2511 error = metaslab_alloc(spa, spa_log_class(spa), size,
2512 new_bp, 1, txg, old_bp,
2513 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2514 }
2515
2516 if (error) {
2517 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2518 new_bp, 1, txg, old_bp,
2519 METASLAB_HINTBP_AVOID);
2520 }
2521
2522 if (error == 0) {
2523 BP_SET_LSIZE(new_bp, size);
2524 BP_SET_PSIZE(new_bp, size);
2525 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2526 BP_SET_CHECKSUM(new_bp,
2527 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2528 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2529 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2530 BP_SET_LEVEL(new_bp, 0);
2531 BP_SET_DEDUP(new_bp, 0);
2532 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2533 }
2534
2535 return (error);
2536 }
2537
2538 /*
2539 * Free an intent log block.
2540 */
2541 void
2542 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2543 {
2544 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2545 ASSERT(!BP_IS_GANG(bp));
2546
2547 zio_free(spa, txg, bp);
2548 }
2549
2550 /*
2551 * ==========================================================================
2552 * Read and write to physical devices
2553 * ==========================================================================
2554 */
2555
2556
2557 /*
2558 * Issue an I/O to the underlying vdev. Typically the issue pipeline
2559 * stops after this stage and will resume upon I/O completion.
2560 * However, there are instances where the vdev layer may need to
2561 * continue the pipeline when an I/O was not issued. Since the I/O
2562 * that was sent to the vdev layer might be different than the one
2563 * currently active in the pipeline (see vdev_queue_io()), we explicitly
2564 * force the underlying vdev layers to call either zio_execute() or
2565 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
2566 */
2567 static int
2568 zio_vdev_io_start(zio_t *zio)
2569 {
2570 vdev_t *vd = zio->io_vd;
2571 uint64_t align;
2572 spa_t *spa = zio->io_spa;
2573
2574 ASSERT(zio->io_error == 0);
2575 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2576
2577 if (vd == NULL) {
2578 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2579 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2580
2581 /*
2582 * The mirror_ops handle multiple DVAs in a single BP.
2583 */
2584 vdev_mirror_ops.vdev_op_io_start(zio);
2585 return (ZIO_PIPELINE_STOP);
2586 }
2587
2588 /*
2589 * We keep track of time-sensitive I/Os so that the scan thread
2590 * can quickly react to certain workloads. In particular, we care
2591 * about non-scrubbing, top-level reads and writes with the following
2592 * characteristics:
2593 * - synchronous writes of user data to non-slog devices
2594 * - any reads of user data
2595 * When these conditions are met, adjust the timestamp of spa_last_io
2596 * which allows the scan thread to adjust its workload accordingly.
2597 */
2598 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2599 vd == vd->vdev_top && !vd->vdev_islog &&
2600 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2601 zio->io_txg != spa_syncing_txg(spa)) {
2602 uint64_t old = spa->spa_last_io;
2603 uint64_t new = ddi_get_lbolt64();
2604 if (old != new)
2605 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2606 }
2607
2608 align = 1ULL << vd->vdev_top->vdev_ashift;
2609
2610 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
2611 P2PHASE(zio->io_size, align) != 0) {
2612 /* Transform logical writes to be a full physical block size. */
2613 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2614 char *abuf = zio_buf_alloc(asize);
2615 ASSERT(vd == vd->vdev_top);
2616 if (zio->io_type == ZIO_TYPE_WRITE) {
2617 bcopy(zio->io_data, abuf, zio->io_size);
2618 bzero(abuf + zio->io_size, asize - zio->io_size);
2619 }
2620 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2621 }
2622
2623 /*
2624 * If this is not a physical io, make sure that it is properly aligned
2625 * before proceeding.
2626 */
2627 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
2628 ASSERT0(P2PHASE(zio->io_offset, align));
2629 ASSERT0(P2PHASE(zio->io_size, align));
2630 } else {
2631 /*
2632 * For physical writes, we allow 512b aligned writes and assume
2633 * the device will perform a read-modify-write as necessary.
2634 */
2635 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
2636 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
2637 }
2638
2639 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2640
2641 /*
2642 * If this is a repair I/O, and there's no self-healing involved --
2643 * that is, we're just resilvering what we expect to resilver --
2644 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2645 * This prevents spurious resilvering with nested replication.
2646 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2647 * A is out of date, we'll read from C+D, then use the data to
2648 * resilver A+B -- but we don't actually want to resilver B, just A.
2649 * The top-level mirror has no way to know this, so instead we just
2650 * discard unnecessary repairs as we work our way down the vdev tree.
2651 * The same logic applies to any form of nested replication:
2652 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2653 */
2654 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2655 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2656 zio->io_txg != 0 && /* not a delegated i/o */
2657 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2658 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2659 zio_vdev_io_bypass(zio);
2660 return (ZIO_PIPELINE_CONTINUE);
2661 }
2662
2663 if (vd->vdev_ops->vdev_op_leaf &&
2664 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2665
2666 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio))
2667 return (ZIO_PIPELINE_CONTINUE);
2668
2669 if ((zio = vdev_queue_io(zio)) == NULL)
2670 return (ZIO_PIPELINE_STOP);
2671
2672 if (!vdev_accessible(vd, zio)) {
2673 zio->io_error = SET_ERROR(ENXIO);
2674 zio_interrupt(zio);
2675 return (ZIO_PIPELINE_STOP);
2676 }
2677 }
2678
2679 vd->vdev_ops->vdev_op_io_start(zio);
2680 return (ZIO_PIPELINE_STOP);
2681 }
2682
2683 static int
2684 zio_vdev_io_done(zio_t *zio)
2685 {
2686 vdev_t *vd = zio->io_vd;
2687 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2688 boolean_t unexpected_error = B_FALSE;
2689
2690 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2691 return (ZIO_PIPELINE_STOP);
2692
2693 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2694
2695 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2696
2697 vdev_queue_io_done(zio);
2698
2699 if (zio->io_type == ZIO_TYPE_WRITE)
2700 vdev_cache_write(zio);
2701
2702 if (zio_injection_enabled && zio->io_error == 0)
2703 zio->io_error = zio_handle_device_injection(vd,
2704 zio, EIO);
2705
2706 if (zio_injection_enabled && zio->io_error == 0)
2707 zio->io_error = zio_handle_label_injection(zio, EIO);
2708
2709 if (zio->io_error) {
2710 if (!vdev_accessible(vd, zio)) {
2711 zio->io_error = SET_ERROR(ENXIO);
2712 } else {
2713 unexpected_error = B_TRUE;
2714 }
2715 }
2716 }
2717
2718 ops->vdev_op_io_done(zio);
2719
2720 if (unexpected_error)
2721 VERIFY(vdev_probe(vd, zio) == NULL);
2722
2723 return (ZIO_PIPELINE_CONTINUE);
2724 }
2725
2726 /*
2727 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2728 * disk, and use that to finish the checksum ereport later.
2729 */
2730 static void
2731 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2732 const void *good_buf)
2733 {
2734 /* no processing needed */
2735 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2736 }
2737
2738 /*ARGSUSED*/
2739 void
2740 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2741 {
2742 void *buf = zio_buf_alloc(zio->io_size);
2743
2744 bcopy(zio->io_data, buf, zio->io_size);
2745
2746 zcr->zcr_cbinfo = zio->io_size;
2747 zcr->zcr_cbdata = buf;
2748 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2749 zcr->zcr_free = zio_buf_free;
2750 }
2751
2752 static int
2753 zio_vdev_io_assess(zio_t *zio)
2754 {
2755 vdev_t *vd = zio->io_vd;
2756
2757 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2758 return (ZIO_PIPELINE_STOP);
2759
2760 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2761 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2762
2763 if (zio->io_vsd != NULL) {
2764 zio->io_vsd_ops->vsd_free(zio);
2765 zio->io_vsd = NULL;
2766 }
2767
2768 if (zio_injection_enabled && zio->io_error == 0)
2769 zio->io_error = zio_handle_fault_injection(zio, EIO);
2770
2771 /*
2772 * If the I/O failed, determine whether we should attempt to retry it.
2773 *
2774 * On retry, we cut in line in the issue queue, since we don't want
2775 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2776 */
2777 if (zio->io_error && vd == NULL &&
2778 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2779 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2780 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2781 zio->io_error = 0;
2782 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2783 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2784 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2785 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2786 zio_requeue_io_start_cut_in_line);
2787 return (ZIO_PIPELINE_STOP);
2788 }
2789
2790 /*
2791 * If we got an error on a leaf device, convert it to ENXIO
2792 * if the device is not accessible at all.
2793 */
2794 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2795 !vdev_accessible(vd, zio))
2796 zio->io_error = SET_ERROR(ENXIO);
2797
2798 /*
2799 * If we can't write to an interior vdev (mirror or RAID-Z),
2800 * set vdev_cant_write so that we stop trying to allocate from it.
2801 */
2802 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2803 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2804 vd->vdev_cant_write = B_TRUE;
2805 }
2806
2807 if (zio->io_error)
2808 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2809
2810 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2811 zio->io_physdone != NULL) {
2812 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
2813 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
2814 zio->io_physdone(zio->io_logical);
2815 }
2816
2817 return (ZIO_PIPELINE_CONTINUE);
2818 }
2819
2820 void
2821 zio_vdev_io_reissue(zio_t *zio)
2822 {
2823 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2824 ASSERT(zio->io_error == 0);
2825
2826 zio->io_stage >>= 1;
2827 }
2828
2829 void
2830 zio_vdev_io_redone(zio_t *zio)
2831 {
2832 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2833
2834 zio->io_stage >>= 1;
2835 }
2836
2837 void
2838 zio_vdev_io_bypass(zio_t *zio)
2839 {
2840 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2841 ASSERT(zio->io_error == 0);
2842
2843 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2844 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2845 }
2846
2847 /*
2848 * ==========================================================================
2849 * Generate and verify checksums
2850 * ==========================================================================
2851 */
2852 static int
2853 zio_checksum_generate(zio_t *zio)
2854 {
2855 blkptr_t *bp = zio->io_bp;
2856 enum zio_checksum checksum;
2857
2858 if (bp == NULL) {
2859 /*
2860 * This is zio_write_phys().
2861 * We're either generating a label checksum, or none at all.
2862 */
2863 checksum = zio->io_prop.zp_checksum;
2864
2865 if (checksum == ZIO_CHECKSUM_OFF)
2866 return (ZIO_PIPELINE_CONTINUE);
2867
2868 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2869 } else {
2870 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2871 ASSERT(!IO_IS_ALLOCATING(zio));
2872 checksum = ZIO_CHECKSUM_GANG_HEADER;
2873 } else {
2874 checksum = BP_GET_CHECKSUM(bp);
2875 }
2876 }
2877
2878 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2879
2880 return (ZIO_PIPELINE_CONTINUE);
2881 }
2882
2883 static int
2884 zio_checksum_verify(zio_t *zio)
2885 {
2886 zio_bad_cksum_t info;
2887 blkptr_t *bp = zio->io_bp;
2888 int error;
2889
2890 ASSERT(zio->io_vd != NULL);
2891
2892 if (bp == NULL) {
2893 /*
2894 * This is zio_read_phys().
2895 * We're either verifying a label checksum, or nothing at all.
2896 */
2897 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2898 return (ZIO_PIPELINE_CONTINUE);
2899
2900 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2901 }
2902
2903 if ((error = zio_checksum_error(zio, &info)) != 0) {
2904 zio->io_error = error;
2905 if (error == ECKSUM &&
2906 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2907 zfs_ereport_start_checksum(zio->io_spa,
2908 zio->io_vd, zio, zio->io_offset,
2909 zio->io_size, NULL, &info);
2910 }
2911 }
2912
2913 return (ZIO_PIPELINE_CONTINUE);
2914 }
2915
2916 /*
2917 * Called by RAID-Z to ensure we don't compute the checksum twice.
2918 */
2919 void
2920 zio_checksum_verified(zio_t *zio)
2921 {
2922 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2923 }
2924
2925 /*
2926 * ==========================================================================
2927 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2928 * An error of 0 indicates success. ENXIO indicates whole-device failure,
2929 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2930 * indicate errors that are specific to one I/O, and most likely permanent.
2931 * Any other error is presumed to be worse because we weren't expecting it.
2932 * ==========================================================================
2933 */
2934 int
2935 zio_worst_error(int e1, int e2)
2936 {
2937 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2938 int r1, r2;
2939
2940 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2941 if (e1 == zio_error_rank[r1])
2942 break;
2943
2944 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2945 if (e2 == zio_error_rank[r2])
2946 break;
2947
2948 return (r1 > r2 ? e1 : e2);
2949 }
2950
2951 /*
2952 * ==========================================================================
2953 * I/O completion
2954 * ==========================================================================
2955 */
2956 static int
2957 zio_ready(zio_t *zio)
2958 {
2959 blkptr_t *bp = zio->io_bp;
2960 zio_t *pio, *pio_next;
2961
2962 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2963 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2964 return (ZIO_PIPELINE_STOP);
2965
2966 if (zio->io_ready) {
2967 ASSERT(IO_IS_ALLOCATING(zio));
2968 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
2969 (zio->io_flags & ZIO_FLAG_NOPWRITE));
2970 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2971
2972 zio->io_ready(zio);
2973 }
2974
2975 if (bp != NULL && bp != &zio->io_bp_copy)
2976 zio->io_bp_copy = *bp;
2977
2978 if (zio->io_error)
2979 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2980
2981 mutex_enter(&zio->io_lock);
2982 zio->io_state[ZIO_WAIT_READY] = 1;
2983 pio = zio_walk_parents(zio);
2984 mutex_exit(&zio->io_lock);
2985
2986 /*
2987 * As we notify zio's parents, new parents could be added.
2988 * New parents go to the head of zio's io_parent_list, however,
2989 * so we will (correctly) not notify them. The remainder of zio's
2990 * io_parent_list, from 'pio_next' onward, cannot change because
2991 * all parents must wait for us to be done before they can be done.
2992 */
2993 for (; pio != NULL; pio = pio_next) {
2994 pio_next = zio_walk_parents(zio);
2995 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2996 }
2997
2998 if (zio->io_flags & ZIO_FLAG_NODATA) {
2999 if (BP_IS_GANG(bp)) {
3000 zio->io_flags &= ~ZIO_FLAG_NODATA;
3001 } else {
3002 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
3003 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3004 }
3005 }
3006
3007 if (zio_injection_enabled &&
3008 zio->io_spa->spa_syncing_txg == zio->io_txg)
3009 zio_handle_ignored_writes(zio);
3010
3011 return (ZIO_PIPELINE_CONTINUE);
3012 }
3013
3014 static int
3015 zio_done(zio_t *zio)
3016 {
3017 spa_t *spa = zio->io_spa;
3018 zio_t *lio = zio->io_logical;
3019 blkptr_t *bp = zio->io_bp;
3020 vdev_t *vd = zio->io_vd;
3021 uint64_t psize = zio->io_size;
3022 zio_t *pio, *pio_next;
3023
3024 /*
3025 * If our children haven't all completed,
3026 * wait for them and then repeat this pipeline stage.
3027 */
3028 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
3029 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
3030 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
3031 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
3032 return (ZIO_PIPELINE_STOP);
3033
3034 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3035 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3036 ASSERT(zio->io_children[c][w] == 0);
3037
3038 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3039 ASSERT(bp->blk_pad[0] == 0);
3040 ASSERT(bp->blk_pad[1] == 0);
3041 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3042 (bp == zio_unique_parent(zio)->io_bp));
3043 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3044 zio->io_bp_override == NULL &&
3045 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3046 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3047 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3048 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3049 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3050 }
3051 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3052 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3053 }
3054
3055 /*
3056 * If there were child vdev/gang/ddt errors, they apply to us now.
3057 */
3058 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3059 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3060 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3061
3062 /*
3063 * If the I/O on the transformed data was successful, generate any
3064 * checksum reports now while we still have the transformed data.
3065 */
3066 if (zio->io_error == 0) {
3067 while (zio->io_cksum_report != NULL) {
3068 zio_cksum_report_t *zcr = zio->io_cksum_report;
3069 uint64_t align = zcr->zcr_align;
3070 uint64_t asize = P2ROUNDUP(psize, align);
3071 char *abuf = zio->io_data;
3072
3073 if (asize != psize) {
3074 abuf = zio_buf_alloc(asize);
3075 bcopy(zio->io_data, abuf, psize);
3076 bzero(abuf + psize, asize - psize);
3077 }
3078
3079 zio->io_cksum_report = zcr->zcr_next;
3080 zcr->zcr_next = NULL;
3081 zcr->zcr_finish(zcr, abuf);
3082 zfs_ereport_free_checksum(zcr);
3083
3084 if (asize != psize)
3085 zio_buf_free(abuf, asize);
3086 }
3087 }
3088
3089 zio_pop_transforms(zio); /* note: may set zio->io_error */
3090
3091 vdev_stat_update(zio, psize);
3092
3093 if (zio->io_error) {
3094 /*
3095 * If this I/O is attached to a particular vdev,
3096 * generate an error message describing the I/O failure
3097 * at the block level. We ignore these errors if the
3098 * device is currently unavailable.
3099 */
3100 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3101 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3102
3103 if ((zio->io_error == EIO || !(zio->io_flags &
3104 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3105 zio == lio) {
3106 /*
3107 * For logical I/O requests, tell the SPA to log the
3108 * error and generate a logical data ereport.
3109 */
3110 spa_log_error(spa, zio);
3111 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3112 0, 0);
3113 }
3114 }
3115
3116 if (zio->io_error && zio == lio) {
3117 /*
3118 * Determine whether zio should be reexecuted. This will
3119 * propagate all the way to the root via zio_notify_parent().
3120 */
3121 ASSERT(vd == NULL && bp != NULL);
3122 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3123
3124 if (IO_IS_ALLOCATING(zio) &&
3125 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3126 if (zio->io_error != ENOSPC)
3127 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3128 else
3129 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3130 }
3131
3132 if ((zio->io_type == ZIO_TYPE_READ ||
3133 zio->io_type == ZIO_TYPE_FREE) &&
3134 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3135 zio->io_error == ENXIO &&
3136 spa_load_state(spa) == SPA_LOAD_NONE &&
3137 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3138 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3139
3140 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3141 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3142
3143 /*
3144 * Here is a possibly good place to attempt to do
3145 * either combinatorial reconstruction or error correction
3146 * based on checksums. It also might be a good place
3147 * to send out preliminary ereports before we suspend
3148 * processing.
3149 */
3150 }
3151
3152 /*
3153 * If there were logical child errors, they apply to us now.
3154 * We defer this until now to avoid conflating logical child
3155 * errors with errors that happened to the zio itself when
3156 * updating vdev stats and reporting FMA events above.
3157 */
3158 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3159
3160 if ((zio->io_error || zio->io_reexecute) &&
3161 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3162 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3163 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3164
3165 zio_gang_tree_free(&zio->io_gang_tree);
3166
3167 /*
3168 * Godfather I/Os should never suspend.
3169 */
3170 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3171 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3172 zio->io_reexecute = 0;
3173
3174 if (zio->io_reexecute) {
3175 /*
3176 * This is a logical I/O that wants to reexecute.
3177 *
3178 * Reexecute is top-down. When an i/o fails, if it's not
3179 * the root, it simply notifies its parent and sticks around.
3180 * The parent, seeing that it still has children in zio_done(),
3181 * does the same. This percolates all the way up to the root.
3182 * The root i/o will reexecute or suspend the entire tree.
3183 *
3184 * This approach ensures that zio_reexecute() honors
3185 * all the original i/o dependency relationships, e.g.
3186 * parents not executing until children are ready.
3187 */
3188 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3189
3190 zio->io_gang_leader = NULL;
3191
3192 mutex_enter(&zio->io_lock);
3193 zio->io_state[ZIO_WAIT_DONE] = 1;
3194 mutex_exit(&zio->io_lock);
3195
3196 /*
3197 * "The Godfather" I/O monitors its children but is
3198 * not a true parent to them. It will track them through
3199 * the pipeline but severs its ties whenever they get into
3200 * trouble (e.g. suspended). This allows "The Godfather"
3201 * I/O to return status without blocking.
3202 */
3203 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3204 zio_link_t *zl = zio->io_walk_link;
3205 pio_next = zio_walk_parents(zio);
3206
3207 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3208 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3209 zio_remove_child(pio, zio, zl);
3210 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3211 }
3212 }
3213
3214 if ((pio = zio_unique_parent(zio)) != NULL) {
3215 /*
3216 * We're not a root i/o, so there's nothing to do
3217 * but notify our parent. Don't propagate errors
3218 * upward since we haven't permanently failed yet.
3219 */
3220 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3221 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3222 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3223 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3224 /*
3225 * We'd fail again if we reexecuted now, so suspend
3226 * until conditions improve (e.g. device comes online).
3227 */
3228 zio_suspend(spa, zio);
3229 } else {
3230 /*
3231 * Reexecution is potentially a huge amount of work.
3232 * Hand it off to the otherwise-unused claim taskq.
3233 */
3234 ASSERT(zio->io_tqent.tqent_next == NULL);
3235 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3236 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3237 0, &zio->io_tqent);
3238 }
3239 return (ZIO_PIPELINE_STOP);
3240 }
3241
3242 ASSERT(zio->io_child_count == 0);
3243 ASSERT(zio->io_reexecute == 0);
3244 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3245
3246 /*
3247 * Report any checksum errors, since the I/O is complete.
3248 */
3249 while (zio->io_cksum_report != NULL) {
3250 zio_cksum_report_t *zcr = zio->io_cksum_report;
3251 zio->io_cksum_report = zcr->zcr_next;
3252 zcr->zcr_next = NULL;
3253 zcr->zcr_finish(zcr, NULL);
3254 zfs_ereport_free_checksum(zcr);
3255 }
3256
3257 /*
3258 * It is the responsibility of the done callback to ensure that this
3259 * particular zio is no longer discoverable for adoption, and as
3260 * such, cannot acquire any new parents.
3261 */
3262 if (zio->io_done)
3263 zio->io_done(zio);
3264
3265 mutex_enter(&zio->io_lock);
3266 zio->io_state[ZIO_WAIT_DONE] = 1;
3267 mutex_exit(&zio->io_lock);
3268
3269 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3270 zio_link_t *zl = zio->io_walk_link;
3271 pio_next = zio_walk_parents(zio);
3272 zio_remove_child(pio, zio, zl);
3273 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3274 }
3275
3276 if (zio->io_waiter != NULL) {
3277 mutex_enter(&zio->io_lock);
3278 zio->io_executor = NULL;
3279 cv_broadcast(&zio->io_cv);
3280 mutex_exit(&zio->io_lock);
3281 } else {
3282 zio_destroy(zio);
3283 }
3284
3285 return (ZIO_PIPELINE_STOP);
3286 }
3287
3288 /*
3289 * ==========================================================================
3290 * I/O pipeline definition
3291 * ==========================================================================
3292 */
3293 static zio_pipe_stage_t *zio_pipeline[] = {
3294 NULL,
3295 zio_read_bp_init,
3296 zio_free_bp_init,
3297 zio_issue_async,
3298 zio_write_bp_init,
3299 zio_checksum_generate,
3300 zio_nop_write,
3301 zio_ddt_read_start,
3302 zio_ddt_read_done,
3303 zio_ddt_write,
3304 zio_ddt_free,
3305 zio_gang_assemble,
3306 zio_gang_issue,
3307 zio_dva_allocate,
3308 zio_dva_free,
3309 zio_dva_claim,
3310 zio_ready,
3311 zio_vdev_io_start,
3312 zio_vdev_io_done,
3313 zio_vdev_io_assess,
3314 zio_checksum_verify,
3315 zio_done
3316 };
3317
3318 /* dnp is the dnode for zb1->zb_object */
3319 boolean_t
3320 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_phys_t *zb1,
3321 const zbookmark_phys_t *zb2)
3322 {
3323 uint64_t zb1nextL0, zb2thisobj;
3324
3325 ASSERT(zb1->zb_objset == zb2->zb_objset);
3326 ASSERT(zb2->zb_level == 0);
3327
3328 /* The objset_phys_t isn't before anything. */
3329 if (dnp == NULL)
3330 return (B_FALSE);
3331
3332 zb1nextL0 = (zb1->zb_blkid + 1) <<
3333 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3334
3335 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3336 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3337
3338 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3339 uint64_t nextobj = zb1nextL0 *
3340 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3341 return (nextobj <= zb2thisobj);
3342 }
3343
3344 if (zb1->zb_object < zb2thisobj)
3345 return (B_TRUE);
3346 if (zb1->zb_object > zb2thisobj)
3347 return (B_FALSE);
3348 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3349 return (B_FALSE);
3350 return (zb1nextL0 <= zb2->zb_blkid);
3351 }