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 zio->io_bp_override = NULL;
1137 BP_ZERO(bp);
1138 }
1139
1140 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1141 /*
1142 * We're rewriting an existing block, which means we're
1143 * working on behalf of spa_sync(). For spa_sync() to
1144 * converge, it must eventually be the case that we don't
1145 * have to allocate new blocks. But compression changes
1146 * the blocksize, which forces a reallocate, and makes
1147 * convergence take longer. Therefore, after the first
1148 * few passes, stop compressing to ensure convergence.
1149 */
1150 pass = spa_sync_pass(spa);
1151
1152 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1153 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1154 ASSERT(!BP_GET_DEDUP(bp));
1155
1156 if (pass >= zfs_sync_pass_dont_compress)
1157 compress = ZIO_COMPRESS_OFF;
1158
1159 /* Make sure someone doesn't change their mind on overwrites */
1160 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1161 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1162 }
1163
1164 if (compress != ZIO_COMPRESS_OFF) {
1165 void *cbuf = zio_buf_alloc(lsize);
1166 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1167 if (psize == 0 || psize == lsize) {
1168 compress = ZIO_COMPRESS_OFF;
1169 zio_buf_free(cbuf, lsize);
1170 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1171 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1172 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1173 encode_embedded_bp_compressed(bp,
1174 cbuf, compress, lsize, psize);
1175 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1176 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1177 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1178 zio_buf_free(cbuf, lsize);
1179 bp->blk_birth = zio->io_txg;
1180 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1181 ASSERT(spa_feature_is_active(spa,
1182 SPA_FEATURE_EMBEDDED_DATA));
1183 return (ZIO_PIPELINE_CONTINUE);
1184 } else {
1185 /*
1186 * Round up compressed size to MINBLOCKSIZE and
1187 * zero the tail.
1188 */
1189 size_t rounded =
1190 P2ROUNDUP(psize, (size_t)SPA_MINBLOCKSIZE);
1191 if (rounded > psize) {
1192 bzero((char *)cbuf + psize, rounded - psize);
1193 psize = rounded;
1194 }
1195 if (psize == lsize) {
1196 compress = ZIO_COMPRESS_OFF;
1197 zio_buf_free(cbuf, lsize);
1198 } else {
1199 zio_push_transform(zio, cbuf,
1200 psize, lsize, NULL);
1201 }
1202 }
1203 }
1204
1205 /*
1206 * The final pass of spa_sync() must be all rewrites, but the first
1207 * few passes offer a trade-off: allocating blocks defers convergence,
1208 * but newly allocated blocks are sequential, so they can be written
1209 * to disk faster. Therefore, we allow the first few passes of
1210 * spa_sync() to allocate new blocks, but force rewrites after that.
1211 * There should only be a handful of blocks after pass 1 in any case.
1212 */
1213 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1214 BP_GET_PSIZE(bp) == psize &&
1215 pass >= zfs_sync_pass_rewrite) {
1216 ASSERT(psize != 0);
1217 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1218 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1219 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1220 } else {
1221 BP_ZERO(bp);
1222 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1223 }
1224
1225 if (psize == 0) {
1226 if (zio->io_bp_orig.blk_birth != 0 &&
1227 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1228 BP_SET_LSIZE(bp, lsize);
1229 BP_SET_TYPE(bp, zp->zp_type);
1230 BP_SET_LEVEL(bp, zp->zp_level);
1231 BP_SET_BIRTH(bp, zio->io_txg, 0);
1232 }
1233 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1234 } else {
1235 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1236 BP_SET_LSIZE(bp, lsize);
1237 BP_SET_TYPE(bp, zp->zp_type);
1238 BP_SET_LEVEL(bp, zp->zp_level);
1239 BP_SET_PSIZE(bp, psize);
1240 BP_SET_COMPRESS(bp, compress);
1241 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1242 BP_SET_DEDUP(bp, zp->zp_dedup);
1243 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1244 if (zp->zp_dedup) {
1245 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1246 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1247 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1248 }
1249 if (zp->zp_nopwrite) {
1250 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1251 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1252 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1253 }
1254 }
1255
1256 return (ZIO_PIPELINE_CONTINUE);
1257 }
1258
1259 static int
1260 zio_free_bp_init(zio_t *zio)
1261 {
1262 blkptr_t *bp = zio->io_bp;
1263
1264 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1265 if (BP_GET_DEDUP(bp))
1266 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1267 }
1268
1269 return (ZIO_PIPELINE_CONTINUE);
1270 }
1271
1272 /*
1273 * ==========================================================================
1274 * Execute the I/O pipeline
1275 * ==========================================================================
1276 */
1277
1278 static void
1279 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1280 {
1281 spa_t *spa = zio->io_spa;
1282 zio_type_t t = zio->io_type;
1283 int flags = (cutinline ? TQ_FRONT : 0);
1284
1285 /*
1286 * If we're a config writer or a probe, the normal issue and
1287 * interrupt threads may all be blocked waiting for the config lock.
1288 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1289 */
1290 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1291 t = ZIO_TYPE_NULL;
1292
1293 /*
1294 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1295 */
1296 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1297 t = ZIO_TYPE_NULL;
1298
1299 /*
1300 * If this is a high priority I/O, then use the high priority taskq if
1301 * available.
1302 */
1303 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1304 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1305 q++;
1306
1307 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1308
1309 /*
1310 * NB: We are assuming that the zio can only be dispatched
1311 * to a single taskq at a time. It would be a grievous error
1312 * to dispatch the zio to another taskq at the same time.
1313 */
1314 ASSERT(zio->io_tqent.tqent_next == NULL);
1315 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1316 flags, &zio->io_tqent);
1317 }
1318
1319 static boolean_t
1320 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1321 {
1322 kthread_t *executor = zio->io_executor;
1323 spa_t *spa = zio->io_spa;
1324
1325 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1326 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1327 uint_t i;
1328 for (i = 0; i < tqs->stqs_count; i++) {
1329 if (taskq_member(tqs->stqs_taskq[i], executor))
1330 return (B_TRUE);
1331 }
1332 }
1333
1334 return (B_FALSE);
1335 }
1336
1337 static int
1338 zio_issue_async(zio_t *zio)
1339 {
1340 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1341
1342 return (ZIO_PIPELINE_STOP);
1343 }
1344
1345 void
1346 zio_interrupt(zio_t *zio)
1347 {
1348 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1349 }
1350
1351 /*
1352 * Execute the I/O pipeline until one of the following occurs:
1353 *
1354 * (1) the I/O completes
1355 * (2) the pipeline stalls waiting for dependent child I/Os
1356 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1357 * (4) the I/O is delegated by vdev-level caching or aggregation
1358 * (5) the I/O is deferred due to vdev-level queueing
1359 * (6) the I/O is handed off to another thread.
1360 *
1361 * In all cases, the pipeline stops whenever there's no CPU work; it never
1362 * burns a thread in cv_wait().
1363 *
1364 * There's no locking on io_stage because there's no legitimate way
1365 * for multiple threads to be attempting to process the same I/O.
1366 */
1367 static zio_pipe_stage_t *zio_pipeline[];
1368
1369 void
1370 zio_execute(zio_t *zio)
1371 {
1372 zio->io_executor = curthread;
1373
1374 while (zio->io_stage < ZIO_STAGE_DONE) {
1375 enum zio_stage pipeline = zio->io_pipeline;
1376 enum zio_stage stage = zio->io_stage;
1377 int rv;
1378
1379 ASSERT(!MUTEX_HELD(&zio->io_lock));
1380 ASSERT(ISP2(stage));
1381 ASSERT(zio->io_stall == NULL);
1382
1383 do {
1384 stage <<= 1;
1385 } while ((stage & pipeline) == 0);
1386
1387 ASSERT(stage <= ZIO_STAGE_DONE);
1388
1389 /*
1390 * If we are in interrupt context and this pipeline stage
1391 * will grab a config lock that is held across I/O,
1392 * or may wait for an I/O that needs an interrupt thread
1393 * to complete, issue async to avoid deadlock.
1394 *
1395 * For VDEV_IO_START, we cut in line so that the io will
1396 * be sent to disk promptly.
1397 */
1398 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1399 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1400 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1401 zio_requeue_io_start_cut_in_line : B_FALSE;
1402 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1403 return;
1404 }
1405
1406 zio->io_stage = stage;
1407 rv = zio_pipeline[highbit64(stage) - 1](zio);
1408
1409 if (rv == ZIO_PIPELINE_STOP)
1410 return;
1411
1412 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1413 }
1414 }
1415
1416 /*
1417 * ==========================================================================
1418 * Initiate I/O, either sync or async
1419 * ==========================================================================
1420 */
1421 int
1422 zio_wait(zio_t *zio)
1423 {
1424 int error;
1425
1426 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1427 ASSERT(zio->io_executor == NULL);
1428
1429 zio->io_waiter = curthread;
1430
1431 zio_execute(zio);
1432
1433 mutex_enter(&zio->io_lock);
1434 while (zio->io_executor != NULL)
1435 cv_wait(&zio->io_cv, &zio->io_lock);
1436 mutex_exit(&zio->io_lock);
1437
1438 error = zio->io_error;
1439 zio_destroy(zio);
1440
1441 return (error);
1442 }
1443
1444 void
1445 zio_nowait(zio_t *zio)
1446 {
1447 ASSERT(zio->io_executor == NULL);
1448
1449 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1450 zio_unique_parent(zio) == NULL) {
1451 /*
1452 * This is a logical async I/O with no parent to wait for it.
1453 * We add it to the spa_async_root_zio "Godfather" I/O which
1454 * will ensure they complete prior to unloading the pool.
1455 */
1456 spa_t *spa = zio->io_spa;
1457
1458 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1459 }
1460
1461 zio_execute(zio);
1462 }
1463
1464 /*
1465 * ==========================================================================
1466 * Reexecute or suspend/resume failed I/O
1467 * ==========================================================================
1468 */
1469
1470 static void
1471 zio_reexecute(zio_t *pio)
1472 {
1473 zio_t *cio, *cio_next;
1474
1475 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1476 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1477 ASSERT(pio->io_gang_leader == NULL);
1478 ASSERT(pio->io_gang_tree == NULL);
1479
1480 pio->io_flags = pio->io_orig_flags;
1481 pio->io_stage = pio->io_orig_stage;
1482 pio->io_pipeline = pio->io_orig_pipeline;
1483 pio->io_reexecute = 0;
1484 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1485 pio->io_error = 0;
1486 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1487 pio->io_state[w] = 0;
1488 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1489 pio->io_child_error[c] = 0;
1490
1491 if (IO_IS_ALLOCATING(pio))
1492 BP_ZERO(pio->io_bp);
1493
1494 /*
1495 * As we reexecute pio's children, new children could be created.
1496 * New children go to the head of pio's io_child_list, however,
1497 * so we will (correctly) not reexecute them. The key is that
1498 * the remainder of pio's io_child_list, from 'cio_next' onward,
1499 * cannot be affected by any side effects of reexecuting 'cio'.
1500 */
1501 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1502 cio_next = zio_walk_children(pio);
1503 mutex_enter(&pio->io_lock);
1504 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1505 pio->io_children[cio->io_child_type][w]++;
1506 mutex_exit(&pio->io_lock);
1507 zio_reexecute(cio);
1508 }
1509
1510 /*
1511 * Now that all children have been reexecuted, execute the parent.
1512 * We don't reexecute "The Godfather" I/O here as it's the
1513 * responsibility of the caller to wait on him.
1514 */
1515 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1516 zio_execute(pio);
1517 }
1518
1519 void
1520 zio_suspend(spa_t *spa, zio_t *zio)
1521 {
1522 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1523 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1524 "failure and the failure mode property for this pool "
1525 "is set to panic.", spa_name(spa));
1526
1527 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1528
1529 mutex_enter(&spa->spa_suspend_lock);
1530
1531 if (spa->spa_suspend_zio_root == NULL)
1532 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1533 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1534 ZIO_FLAG_GODFATHER);
1535
1536 spa->spa_suspended = B_TRUE;
1537
1538 if (zio != NULL) {
1539 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1540 ASSERT(zio != spa->spa_suspend_zio_root);
1541 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1542 ASSERT(zio_unique_parent(zio) == NULL);
1543 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1544 zio_add_child(spa->spa_suspend_zio_root, zio);
1545 }
1546
1547 mutex_exit(&spa->spa_suspend_lock);
1548 }
1549
1550 int
1551 zio_resume(spa_t *spa)
1552 {
1553 zio_t *pio;
1554
1555 /*
1556 * Reexecute all previously suspended i/o.
1557 */
1558 mutex_enter(&spa->spa_suspend_lock);
1559 spa->spa_suspended = B_FALSE;
1560 cv_broadcast(&spa->spa_suspend_cv);
1561 pio = spa->spa_suspend_zio_root;
1562 spa->spa_suspend_zio_root = NULL;
1563 mutex_exit(&spa->spa_suspend_lock);
1564
1565 if (pio == NULL)
1566 return (0);
1567
1568 zio_reexecute(pio);
1569 return (zio_wait(pio));
1570 }
1571
1572 void
1573 zio_resume_wait(spa_t *spa)
1574 {
1575 mutex_enter(&spa->spa_suspend_lock);
1576 while (spa_suspended(spa))
1577 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1578 mutex_exit(&spa->spa_suspend_lock);
1579 }
1580
1581 /*
1582 * ==========================================================================
1583 * Gang blocks.
1584 *
1585 * A gang block is a collection of small blocks that looks to the DMU
1586 * like one large block. When zio_dva_allocate() cannot find a block
1587 * of the requested size, due to either severe fragmentation or the pool
1588 * being nearly full, it calls zio_write_gang_block() to construct the
1589 * block from smaller fragments.
1590 *
1591 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1592 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1593 * an indirect block: it's an array of block pointers. It consumes
1594 * only one sector and hence is allocatable regardless of fragmentation.
1595 * The gang header's bps point to its gang members, which hold the data.
1596 *
1597 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1598 * as the verifier to ensure uniqueness of the SHA256 checksum.
1599 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1600 * not the gang header. This ensures that data block signatures (needed for
1601 * deduplication) are independent of how the block is physically stored.
1602 *
1603 * Gang blocks can be nested: a gang member may itself be a gang block.
1604 * Thus every gang block is a tree in which root and all interior nodes are
1605 * gang headers, and the leaves are normal blocks that contain user data.
1606 * The root of the gang tree is called the gang leader.
1607 *
1608 * To perform any operation (read, rewrite, free, claim) on a gang block,
1609 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1610 * in the io_gang_tree field of the original logical i/o by recursively
1611 * reading the gang leader and all gang headers below it. This yields
1612 * an in-core tree containing the contents of every gang header and the
1613 * bps for every constituent of the gang block.
1614 *
1615 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1616 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1617 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1618 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1619 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1620 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1621 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1622 * of the gang header plus zio_checksum_compute() of the data to update the
1623 * gang header's blk_cksum as described above.
1624 *
1625 * The two-phase assemble/issue model solves the problem of partial failure --
1626 * what if you'd freed part of a gang block but then couldn't read the
1627 * gang header for another part? Assembling the entire gang tree first
1628 * ensures that all the necessary gang header I/O has succeeded before
1629 * starting the actual work of free, claim, or write. Once the gang tree
1630 * is assembled, free and claim are in-memory operations that cannot fail.
1631 *
1632 * In the event that a gang write fails, zio_dva_unallocate() walks the
1633 * gang tree to immediately free (i.e. insert back into the space map)
1634 * everything we've allocated. This ensures that we don't get ENOSPC
1635 * errors during repeated suspend/resume cycles due to a flaky device.
1636 *
1637 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1638 * the gang tree, we won't modify the block, so we can safely defer the free
1639 * (knowing that the block is still intact). If we *can* assemble the gang
1640 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1641 * each constituent bp and we can allocate a new block on the next sync pass.
1642 *
1643 * In all cases, the gang tree allows complete recovery from partial failure.
1644 * ==========================================================================
1645 */
1646
1647 static zio_t *
1648 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1649 {
1650 if (gn != NULL)
1651 return (pio);
1652
1653 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1654 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1655 &pio->io_bookmark));
1656 }
1657
1658 zio_t *
1659 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1660 {
1661 zio_t *zio;
1662
1663 if (gn != NULL) {
1664 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1665 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1666 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1667 /*
1668 * As we rewrite each gang header, the pipeline will compute
1669 * a new gang block header checksum for it; but no one will
1670 * compute a new data checksum, so we do that here. The one
1671 * exception is the gang leader: the pipeline already computed
1672 * its data checksum because that stage precedes gang assembly.
1673 * (Presently, nothing actually uses interior data checksums;
1674 * this is just good hygiene.)
1675 */
1676 if (gn != pio->io_gang_leader->io_gang_tree) {
1677 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1678 data, BP_GET_PSIZE(bp));
1679 }
1680 /*
1681 * If we are here to damage data for testing purposes,
1682 * leave the GBH alone so that we can detect the damage.
1683 */
1684 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1685 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1686 } else {
1687 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1688 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1689 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1690 }
1691
1692 return (zio);
1693 }
1694
1695 /* ARGSUSED */
1696 zio_t *
1697 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1698 {
1699 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1700 ZIO_GANG_CHILD_FLAGS(pio)));
1701 }
1702
1703 /* ARGSUSED */
1704 zio_t *
1705 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1706 {
1707 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1708 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1709 }
1710
1711 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1712 NULL,
1713 zio_read_gang,
1714 zio_rewrite_gang,
1715 zio_free_gang,
1716 zio_claim_gang,
1717 NULL
1718 };
1719
1720 static void zio_gang_tree_assemble_done(zio_t *zio);
1721
1722 static zio_gang_node_t *
1723 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1724 {
1725 zio_gang_node_t *gn;
1726
1727 ASSERT(*gnpp == NULL);
1728
1729 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1730 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1731 *gnpp = gn;
1732
1733 return (gn);
1734 }
1735
1736 static void
1737 zio_gang_node_free(zio_gang_node_t **gnpp)
1738 {
1739 zio_gang_node_t *gn = *gnpp;
1740
1741 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1742 ASSERT(gn->gn_child[g] == NULL);
1743
1744 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1745 kmem_free(gn, sizeof (*gn));
1746 *gnpp = NULL;
1747 }
1748
1749 static void
1750 zio_gang_tree_free(zio_gang_node_t **gnpp)
1751 {
1752 zio_gang_node_t *gn = *gnpp;
1753
1754 if (gn == NULL)
1755 return;
1756
1757 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1758 zio_gang_tree_free(&gn->gn_child[g]);
1759
1760 zio_gang_node_free(gnpp);
1761 }
1762
1763 static void
1764 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1765 {
1766 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1767
1768 ASSERT(gio->io_gang_leader == gio);
1769 ASSERT(BP_IS_GANG(bp));
1770
1771 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1772 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1773 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1774 }
1775
1776 static void
1777 zio_gang_tree_assemble_done(zio_t *zio)
1778 {
1779 zio_t *gio = zio->io_gang_leader;
1780 zio_gang_node_t *gn = zio->io_private;
1781 blkptr_t *bp = zio->io_bp;
1782
1783 ASSERT(gio == zio_unique_parent(zio));
1784 ASSERT(zio->io_child_count == 0);
1785
1786 if (zio->io_error)
1787 return;
1788
1789 if (BP_SHOULD_BYTESWAP(bp))
1790 byteswap_uint64_array(zio->io_data, zio->io_size);
1791
1792 ASSERT(zio->io_data == gn->gn_gbh);
1793 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1794 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1795
1796 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1797 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1798 if (!BP_IS_GANG(gbp))
1799 continue;
1800 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1801 }
1802 }
1803
1804 static void
1805 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1806 {
1807 zio_t *gio = pio->io_gang_leader;
1808 zio_t *zio;
1809
1810 ASSERT(BP_IS_GANG(bp) == !!gn);
1811 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1812 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1813
1814 /*
1815 * If you're a gang header, your data is in gn->gn_gbh.
1816 * If you're a gang member, your data is in 'data' and gn == NULL.
1817 */
1818 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1819
1820 if (gn != NULL) {
1821 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1822
1823 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1824 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1825 if (BP_IS_HOLE(gbp))
1826 continue;
1827 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1828 data = (char *)data + BP_GET_PSIZE(gbp);
1829 }
1830 }
1831
1832 if (gn == gio->io_gang_tree)
1833 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1834
1835 if (zio != pio)
1836 zio_nowait(zio);
1837 }
1838
1839 static int
1840 zio_gang_assemble(zio_t *zio)
1841 {
1842 blkptr_t *bp = zio->io_bp;
1843
1844 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1845 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1846
1847 zio->io_gang_leader = zio;
1848
1849 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1850
1851 return (ZIO_PIPELINE_CONTINUE);
1852 }
1853
1854 static int
1855 zio_gang_issue(zio_t *zio)
1856 {
1857 blkptr_t *bp = zio->io_bp;
1858
1859 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1860 return (ZIO_PIPELINE_STOP);
1861
1862 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1863 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1864
1865 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1866 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1867 else
1868 zio_gang_tree_free(&zio->io_gang_tree);
1869
1870 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1871
1872 return (ZIO_PIPELINE_CONTINUE);
1873 }
1874
1875 static void
1876 zio_write_gang_member_ready(zio_t *zio)
1877 {
1878 zio_t *pio = zio_unique_parent(zio);
1879 zio_t *gio = zio->io_gang_leader;
1880 dva_t *cdva = zio->io_bp->blk_dva;
1881 dva_t *pdva = pio->io_bp->blk_dva;
1882 uint64_t asize;
1883
1884 if (BP_IS_HOLE(zio->io_bp))
1885 return;
1886
1887 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1888
1889 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1890 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1891 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1892 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1893 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1894
1895 mutex_enter(&pio->io_lock);
1896 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1897 ASSERT(DVA_GET_GANG(&pdva[d]));
1898 asize = DVA_GET_ASIZE(&pdva[d]);
1899 asize += DVA_GET_ASIZE(&cdva[d]);
1900 DVA_SET_ASIZE(&pdva[d], asize);
1901 }
1902 mutex_exit(&pio->io_lock);
1903 }
1904
1905 static int
1906 zio_write_gang_block(zio_t *pio)
1907 {
1908 spa_t *spa = pio->io_spa;
1909 blkptr_t *bp = pio->io_bp;
1910 zio_t *gio = pio->io_gang_leader;
1911 zio_t *zio;
1912 zio_gang_node_t *gn, **gnpp;
1913 zio_gbh_phys_t *gbh;
1914 uint64_t txg = pio->io_txg;
1915 uint64_t resid = pio->io_size;
1916 uint64_t lsize;
1917 int copies = gio->io_prop.zp_copies;
1918 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1919 zio_prop_t zp;
1920 int error;
1921
1922 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1923 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1924 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1925 if (error) {
1926 pio->io_error = error;
1927 return (ZIO_PIPELINE_CONTINUE);
1928 }
1929
1930 if (pio == gio) {
1931 gnpp = &gio->io_gang_tree;
1932 } else {
1933 gnpp = pio->io_private;
1934 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1935 }
1936
1937 gn = zio_gang_node_alloc(gnpp);
1938 gbh = gn->gn_gbh;
1939 bzero(gbh, SPA_GANGBLOCKSIZE);
1940
1941 /*
1942 * Create the gang header.
1943 */
1944 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1945 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1946
1947 /*
1948 * Create and nowait the gang children.
1949 */
1950 for (int g = 0; resid != 0; resid -= lsize, g++) {
1951 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1952 SPA_MINBLOCKSIZE);
1953 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1954
1955 zp.zp_checksum = gio->io_prop.zp_checksum;
1956 zp.zp_compress = ZIO_COMPRESS_OFF;
1957 zp.zp_type = DMU_OT_NONE;
1958 zp.zp_level = 0;
1959 zp.zp_copies = gio->io_prop.zp_copies;
1960 zp.zp_dedup = B_FALSE;
1961 zp.zp_dedup_verify = B_FALSE;
1962 zp.zp_nopwrite = B_FALSE;
1963
1964 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1965 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1966 zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g],
1967 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1968 &pio->io_bookmark));
1969 }
1970
1971 /*
1972 * Set pio's pipeline to just wait for zio to finish.
1973 */
1974 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1975
1976 zio_nowait(zio);
1977
1978 return (ZIO_PIPELINE_CONTINUE);
1979 }
1980
1981 /*
1982 * The zio_nop_write stage in the pipeline determines if allocating
1983 * a new bp is necessary. By leveraging a cryptographically secure checksum,
1984 * such as SHA256, we can compare the checksums of the new data and the old
1985 * to determine if allocating a new block is required. The nopwrite
1986 * feature can handle writes in either syncing or open context (i.e. zil
1987 * writes) and as a result is mutually exclusive with dedup.
1988 */
1989 static int
1990 zio_nop_write(zio_t *zio)
1991 {
1992 blkptr_t *bp = zio->io_bp;
1993 blkptr_t *bp_orig = &zio->io_bp_orig;
1994 zio_prop_t *zp = &zio->io_prop;
1995
1996 ASSERT(BP_GET_LEVEL(bp) == 0);
1997 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1998 ASSERT(zp->zp_nopwrite);
1999 ASSERT(!zp->zp_dedup);
2000 ASSERT(zio->io_bp_override == NULL);
2001 ASSERT(IO_IS_ALLOCATING(zio));
2002
2003 /*
2004 * Check to see if the original bp and the new bp have matching
2005 * characteristics (i.e. same checksum, compression algorithms, etc).
2006 * If they don't then just continue with the pipeline which will
2007 * allocate a new bp.
2008 */
2009 if (BP_IS_HOLE(bp_orig) ||
2010 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
2011 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2012 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2013 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2014 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2015 return (ZIO_PIPELINE_CONTINUE);
2016
2017 /*
2018 * If the checksums match then reset the pipeline so that we
2019 * avoid allocating a new bp and issuing any I/O.
2020 */
2021 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2022 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
2023 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2024 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2025 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2026 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2027 sizeof (uint64_t)) == 0);
2028
2029 *bp = *bp_orig;
2030 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2031 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2032 }
2033
2034 return (ZIO_PIPELINE_CONTINUE);
2035 }
2036
2037 /*
2038 * ==========================================================================
2039 * Dedup
2040 * ==========================================================================
2041 */
2042 static void
2043 zio_ddt_child_read_done(zio_t *zio)
2044 {
2045 blkptr_t *bp = zio->io_bp;
2046 ddt_entry_t *dde = zio->io_private;
2047 ddt_phys_t *ddp;
2048 zio_t *pio = zio_unique_parent(zio);
2049
2050 mutex_enter(&pio->io_lock);
2051 ddp = ddt_phys_select(dde, bp);
2052 if (zio->io_error == 0)
2053 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2054 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
2055 dde->dde_repair_data = zio->io_data;
2056 else
2057 zio_buf_free(zio->io_data, zio->io_size);
2058 mutex_exit(&pio->io_lock);
2059 }
2060
2061 static int
2062 zio_ddt_read_start(zio_t *zio)
2063 {
2064 blkptr_t *bp = zio->io_bp;
2065
2066 ASSERT(BP_GET_DEDUP(bp));
2067 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2068 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2069
2070 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2071 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2072 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2073 ddt_phys_t *ddp = dde->dde_phys;
2074 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2075 blkptr_t blk;
2076
2077 ASSERT(zio->io_vsd == NULL);
2078 zio->io_vsd = dde;
2079
2080 if (ddp_self == NULL)
2081 return (ZIO_PIPELINE_CONTINUE);
2082
2083 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2084 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2085 continue;
2086 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2087 &blk);
2088 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2089 zio_buf_alloc(zio->io_size), zio->io_size,
2090 zio_ddt_child_read_done, dde, zio->io_priority,
2091 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2092 &zio->io_bookmark));
2093 }
2094 return (ZIO_PIPELINE_CONTINUE);
2095 }
2096
2097 zio_nowait(zio_read(zio, zio->io_spa, bp,
2098 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2099 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2100
2101 return (ZIO_PIPELINE_CONTINUE);
2102 }
2103
2104 static int
2105 zio_ddt_read_done(zio_t *zio)
2106 {
2107 blkptr_t *bp = zio->io_bp;
2108
2109 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2110 return (ZIO_PIPELINE_STOP);
2111
2112 ASSERT(BP_GET_DEDUP(bp));
2113 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2114 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2115
2116 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2117 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2118 ddt_entry_t *dde = zio->io_vsd;
2119 if (ddt == NULL) {
2120 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2121 return (ZIO_PIPELINE_CONTINUE);
2122 }
2123 if (dde == NULL) {
2124 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2125 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2126 return (ZIO_PIPELINE_STOP);
2127 }
2128 if (dde->dde_repair_data != NULL) {
2129 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2130 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2131 }
2132 ddt_repair_done(ddt, dde);
2133 zio->io_vsd = NULL;
2134 }
2135
2136 ASSERT(zio->io_vsd == NULL);
2137
2138 return (ZIO_PIPELINE_CONTINUE);
2139 }
2140
2141 static boolean_t
2142 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2143 {
2144 spa_t *spa = zio->io_spa;
2145
2146 /*
2147 * Note: we compare the original data, not the transformed data,
2148 * because when zio->io_bp is an override bp, we will not have
2149 * pushed the I/O transforms. That's an important optimization
2150 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2151 */
2152 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2153 zio_t *lio = dde->dde_lead_zio[p];
2154
2155 if (lio != NULL) {
2156 return (lio->io_orig_size != zio->io_orig_size ||
2157 bcmp(zio->io_orig_data, lio->io_orig_data,
2158 zio->io_orig_size) != 0);
2159 }
2160 }
2161
2162 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2163 ddt_phys_t *ddp = &dde->dde_phys[p];
2164
2165 if (ddp->ddp_phys_birth != 0) {
2166 arc_buf_t *abuf = NULL;
2167 arc_flags_t aflags = ARC_FLAG_WAIT;
2168 blkptr_t blk = *zio->io_bp;
2169 int error;
2170
2171 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2172
2173 ddt_exit(ddt);
2174
2175 error = arc_read(NULL, spa, &blk,
2176 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2177 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2178 &aflags, &zio->io_bookmark);
2179
2180 if (error == 0) {
2181 if (arc_buf_size(abuf) != zio->io_orig_size ||
2182 bcmp(abuf->b_data, zio->io_orig_data,
2183 zio->io_orig_size) != 0)
2184 error = SET_ERROR(EEXIST);
2185 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2186 }
2187
2188 ddt_enter(ddt);
2189 return (error != 0);
2190 }
2191 }
2192
2193 return (B_FALSE);
2194 }
2195
2196 static void
2197 zio_ddt_child_write_ready(zio_t *zio)
2198 {
2199 int p = zio->io_prop.zp_copies;
2200 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2201 ddt_entry_t *dde = zio->io_private;
2202 ddt_phys_t *ddp = &dde->dde_phys[p];
2203 zio_t *pio;
2204
2205 if (zio->io_error)
2206 return;
2207
2208 ddt_enter(ddt);
2209
2210 ASSERT(dde->dde_lead_zio[p] == zio);
2211
2212 ddt_phys_fill(ddp, zio->io_bp);
2213
2214 while ((pio = zio_walk_parents(zio)) != NULL)
2215 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2216
2217 ddt_exit(ddt);
2218 }
2219
2220 static void
2221 zio_ddt_child_write_done(zio_t *zio)
2222 {
2223 int p = zio->io_prop.zp_copies;
2224 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2225 ddt_entry_t *dde = zio->io_private;
2226 ddt_phys_t *ddp = &dde->dde_phys[p];
2227
2228 ddt_enter(ddt);
2229
2230 ASSERT(ddp->ddp_refcnt == 0);
2231 ASSERT(dde->dde_lead_zio[p] == zio);
2232 dde->dde_lead_zio[p] = NULL;
2233
2234 if (zio->io_error == 0) {
2235 while (zio_walk_parents(zio) != NULL)
2236 ddt_phys_addref(ddp);
2237 } else {
2238 ddt_phys_clear(ddp);
2239 }
2240
2241 ddt_exit(ddt);
2242 }
2243
2244 static void
2245 zio_ddt_ditto_write_done(zio_t *zio)
2246 {
2247 int p = DDT_PHYS_DITTO;
2248 zio_prop_t *zp = &zio->io_prop;
2249 blkptr_t *bp = zio->io_bp;
2250 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2251 ddt_entry_t *dde = zio->io_private;
2252 ddt_phys_t *ddp = &dde->dde_phys[p];
2253 ddt_key_t *ddk = &dde->dde_key;
2254
2255 ddt_enter(ddt);
2256
2257 ASSERT(ddp->ddp_refcnt == 0);
2258 ASSERT(dde->dde_lead_zio[p] == zio);
2259 dde->dde_lead_zio[p] = NULL;
2260
2261 if (zio->io_error == 0) {
2262 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2263 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2264 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2265 if (ddp->ddp_phys_birth != 0)
2266 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2267 ddt_phys_fill(ddp, bp);
2268 }
2269
2270 ddt_exit(ddt);
2271 }
2272
2273 static int
2274 zio_ddt_write(zio_t *zio)
2275 {
2276 spa_t *spa = zio->io_spa;
2277 blkptr_t *bp = zio->io_bp;
2278 uint64_t txg = zio->io_txg;
2279 zio_prop_t *zp = &zio->io_prop;
2280 int p = zp->zp_copies;
2281 int ditto_copies;
2282 zio_t *cio = NULL;
2283 zio_t *dio = NULL;
2284 ddt_t *ddt = ddt_select(spa, bp);
2285 ddt_entry_t *dde;
2286 ddt_phys_t *ddp;
2287
2288 ASSERT(BP_GET_DEDUP(bp));
2289 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2290 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2291
2292 ddt_enter(ddt);
2293 dde = ddt_lookup(ddt, bp, B_TRUE);
2294 ddp = &dde->dde_phys[p];
2295
2296 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2297 /*
2298 * If we're using a weak checksum, upgrade to a strong checksum
2299 * and try again. If we're already using a strong checksum,
2300 * we can't resolve it, so just convert to an ordinary write.
2301 * (And automatically e-mail a paper to Nature?)
2302 */
2303 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2304 zp->zp_checksum = spa_dedup_checksum(spa);
2305 zio_pop_transforms(zio);
2306 zio->io_stage = ZIO_STAGE_OPEN;
2307 BP_ZERO(bp);
2308 } else {
2309 zp->zp_dedup = B_FALSE;
2310 }
2311 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2312 ddt_exit(ddt);
2313 return (ZIO_PIPELINE_CONTINUE);
2314 }
2315
2316 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2317 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2318
2319 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2320 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2321 zio_prop_t czp = *zp;
2322
2323 czp.zp_copies = ditto_copies;
2324
2325 /*
2326 * If we arrived here with an override bp, we won't have run
2327 * the transform stack, so we won't have the data we need to
2328 * generate a child i/o. So, toss the override bp and restart.
2329 * This is safe, because using the override bp is just an
2330 * optimization; and it's rare, so the cost doesn't matter.
2331 */
2332 if (zio->io_bp_override) {
2333 zio_pop_transforms(zio);
2334 zio->io_stage = ZIO_STAGE_OPEN;
2335 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2336 zio->io_bp_override = NULL;
2337 BP_ZERO(bp);
2338 ddt_exit(ddt);
2339 return (ZIO_PIPELINE_CONTINUE);
2340 }
2341
2342 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2343 zio->io_orig_size, &czp, NULL, NULL,
2344 zio_ddt_ditto_write_done, dde, zio->io_priority,
2345 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2346
2347 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2348 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2349 }
2350
2351 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2352 if (ddp->ddp_phys_birth != 0)
2353 ddt_bp_fill(ddp, bp, txg);
2354 if (dde->dde_lead_zio[p] != NULL)
2355 zio_add_child(zio, dde->dde_lead_zio[p]);
2356 else
2357 ddt_phys_addref(ddp);
2358 } else if (zio->io_bp_override) {
2359 ASSERT(bp->blk_birth == txg);
2360 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2361 ddt_phys_fill(ddp, bp);
2362 ddt_phys_addref(ddp);
2363 } else {
2364 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2365 zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL,
2366 zio_ddt_child_write_done, dde, zio->io_priority,
2367 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2368
2369 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2370 dde->dde_lead_zio[p] = cio;
2371 }
2372
2373 ddt_exit(ddt);
2374
2375 if (cio)
2376 zio_nowait(cio);
2377 if (dio)
2378 zio_nowait(dio);
2379
2380 return (ZIO_PIPELINE_CONTINUE);
2381 }
2382
2383 ddt_entry_t *freedde; /* for debugging */
2384
2385 static int
2386 zio_ddt_free(zio_t *zio)
2387 {
2388 spa_t *spa = zio->io_spa;
2389 blkptr_t *bp = zio->io_bp;
2390 ddt_t *ddt = ddt_select(spa, bp);
2391 ddt_entry_t *dde;
2392 ddt_phys_t *ddp;
2393
2394 ASSERT(BP_GET_DEDUP(bp));
2395 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2396
2397 ddt_enter(ddt);
2398 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2399 ddp = ddt_phys_select(dde, bp);
2400 ddt_phys_decref(ddp);
2401 ddt_exit(ddt);
2402
2403 return (ZIO_PIPELINE_CONTINUE);
2404 }
2405
2406 /*
2407 * ==========================================================================
2408 * Allocate and free blocks
2409 * ==========================================================================
2410 */
2411 static int
2412 zio_dva_allocate(zio_t *zio)
2413 {
2414 spa_t *spa = zio->io_spa;
2415 metaslab_class_t *mc = spa_normal_class(spa);
2416 blkptr_t *bp = zio->io_bp;
2417 int error;
2418 int flags = 0;
2419
2420 if (zio->io_gang_leader == NULL) {
2421 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2422 zio->io_gang_leader = zio;
2423 }
2424
2425 ASSERT(BP_IS_HOLE(bp));
2426 ASSERT0(BP_GET_NDVAS(bp));
2427 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2428 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2429 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2430
2431 /*
2432 * The dump device does not support gang blocks so allocation on
2433 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2434 * the "fast" gang feature.
2435 */
2436 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2437 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2438 METASLAB_GANG_CHILD : 0;
2439 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2440 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2441
2442 if (error) {
2443 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2444 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2445 error);
2446 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2447 return (zio_write_gang_block(zio));
2448 zio->io_error = error;
2449 }
2450
2451 return (ZIO_PIPELINE_CONTINUE);
2452 }
2453
2454 static int
2455 zio_dva_free(zio_t *zio)
2456 {
2457 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2458
2459 return (ZIO_PIPELINE_CONTINUE);
2460 }
2461
2462 static int
2463 zio_dva_claim(zio_t *zio)
2464 {
2465 int error;
2466
2467 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2468 if (error)
2469 zio->io_error = error;
2470
2471 return (ZIO_PIPELINE_CONTINUE);
2472 }
2473
2474 /*
2475 * Undo an allocation. This is used by zio_done() when an I/O fails
2476 * and we want to give back the block we just allocated.
2477 * This handles both normal blocks and gang blocks.
2478 */
2479 static void
2480 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2481 {
2482 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2483 ASSERT(zio->io_bp_override == NULL);
2484
2485 if (!BP_IS_HOLE(bp))
2486 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2487
2488 if (gn != NULL) {
2489 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2490 zio_dva_unallocate(zio, gn->gn_child[g],
2491 &gn->gn_gbh->zg_blkptr[g]);
2492 }
2493 }
2494 }
2495
2496 /*
2497 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2498 */
2499 int
2500 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2501 uint64_t size, boolean_t use_slog)
2502 {
2503 int error = 1;
2504
2505 ASSERT(txg > spa_syncing_txg(spa));
2506
2507 /*
2508 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2509 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2510 * when allocating them.
2511 */
2512 if (use_slog) {
2513 error = metaslab_alloc(spa, spa_log_class(spa), size,
2514 new_bp, 1, txg, old_bp,
2515 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2516 }
2517
2518 if (error) {
2519 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2520 new_bp, 1, txg, old_bp,
2521 METASLAB_HINTBP_AVOID);
2522 }
2523
2524 if (error == 0) {
2525 BP_SET_LSIZE(new_bp, size);
2526 BP_SET_PSIZE(new_bp, size);
2527 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2528 BP_SET_CHECKSUM(new_bp,
2529 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2530 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2531 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2532 BP_SET_LEVEL(new_bp, 0);
2533 BP_SET_DEDUP(new_bp, 0);
2534 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2535 }
2536
2537 return (error);
2538 }
2539
2540 /*
2541 * Free an intent log block.
2542 */
2543 void
2544 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2545 {
2546 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2547 ASSERT(!BP_IS_GANG(bp));
2548
2549 zio_free(spa, txg, bp);
2550 }
2551
2552 /*
2553 * ==========================================================================
2554 * Read and write to physical devices
2555 * ==========================================================================
2556 */
2557
2558
2559 /*
2560 * Issue an I/O to the underlying vdev. Typically the issue pipeline
2561 * stops after this stage and will resume upon I/O completion.
2562 * However, there are instances where the vdev layer may need to
2563 * continue the pipeline when an I/O was not issued. Since the I/O
2564 * that was sent to the vdev layer might be different than the one
2565 * currently active in the pipeline (see vdev_queue_io()), we explicitly
2566 * force the underlying vdev layers to call either zio_execute() or
2567 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
2568 */
2569 static int
2570 zio_vdev_io_start(zio_t *zio)
2571 {
2572 vdev_t *vd = zio->io_vd;
2573 uint64_t align;
2574 spa_t *spa = zio->io_spa;
2575
2576 ASSERT(zio->io_error == 0);
2577 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2578
2579 if (vd == NULL) {
2580 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2581 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2582
2583 /*
2584 * The mirror_ops handle multiple DVAs in a single BP.
2585 */
2586 vdev_mirror_ops.vdev_op_io_start(zio);
2587 return (ZIO_PIPELINE_STOP);
2588 }
2589
2590 /*
2591 * We keep track of time-sensitive I/Os so that the scan thread
2592 * can quickly react to certain workloads. In particular, we care
2593 * about non-scrubbing, top-level reads and writes with the following
2594 * characteristics:
2595 * - synchronous writes of user data to non-slog devices
2596 * - any reads of user data
2597 * When these conditions are met, adjust the timestamp of spa_last_io
2598 * which allows the scan thread to adjust its workload accordingly.
2599 */
2600 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2601 vd == vd->vdev_top && !vd->vdev_islog &&
2602 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2603 zio->io_txg != spa_syncing_txg(spa)) {
2604 uint64_t old = spa->spa_last_io;
2605 uint64_t new = ddi_get_lbolt64();
2606 if (old != new)
2607 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2608 }
2609
2610 align = 1ULL << vd->vdev_top->vdev_ashift;
2611
2612 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
2613 P2PHASE(zio->io_size, align) != 0) {
2614 /* Transform logical writes to be a full physical block size. */
2615 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2616 char *abuf = zio_buf_alloc(asize);
2617 ASSERT(vd == vd->vdev_top);
2618 if (zio->io_type == ZIO_TYPE_WRITE) {
2619 bcopy(zio->io_data, abuf, zio->io_size);
2620 bzero(abuf + zio->io_size, asize - zio->io_size);
2621 }
2622 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2623 }
2624
2625 /*
2626 * If this is not a physical io, make sure that it is properly aligned
2627 * before proceeding.
2628 */
2629 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
2630 ASSERT0(P2PHASE(zio->io_offset, align));
2631 ASSERT0(P2PHASE(zio->io_size, align));
2632 } else {
2633 /*
2634 * For physical writes, we allow 512b aligned writes and assume
2635 * the device will perform a read-modify-write as necessary.
2636 */
2637 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
2638 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
2639 }
2640
2641 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2642
2643 /*
2644 * If this is a repair I/O, and there's no self-healing involved --
2645 * that is, we're just resilvering what we expect to resilver --
2646 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2647 * This prevents spurious resilvering with nested replication.
2648 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2649 * A is out of date, we'll read from C+D, then use the data to
2650 * resilver A+B -- but we don't actually want to resilver B, just A.
2651 * The top-level mirror has no way to know this, so instead we just
2652 * discard unnecessary repairs as we work our way down the vdev tree.
2653 * The same logic applies to any form of nested replication:
2654 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2655 */
2656 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2657 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2658 zio->io_txg != 0 && /* not a delegated i/o */
2659 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2660 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2661 zio_vdev_io_bypass(zio);
2662 return (ZIO_PIPELINE_CONTINUE);
2663 }
2664
2665 if (vd->vdev_ops->vdev_op_leaf &&
2666 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2667
2668 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio))
2669 return (ZIO_PIPELINE_CONTINUE);
2670
2671 if ((zio = vdev_queue_io(zio)) == NULL)
2672 return (ZIO_PIPELINE_STOP);
2673
2674 if (!vdev_accessible(vd, zio)) {
2675 zio->io_error = SET_ERROR(ENXIO);
2676 zio_interrupt(zio);
2677 return (ZIO_PIPELINE_STOP);
2678 }
2679 }
2680
2681 vd->vdev_ops->vdev_op_io_start(zio);
2682 return (ZIO_PIPELINE_STOP);
2683 }
2684
2685 static int
2686 zio_vdev_io_done(zio_t *zio)
2687 {
2688 vdev_t *vd = zio->io_vd;
2689 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2690 boolean_t unexpected_error = B_FALSE;
2691
2692 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2693 return (ZIO_PIPELINE_STOP);
2694
2695 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2696
2697 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2698
2699 vdev_queue_io_done(zio);
2700
2701 if (zio->io_type == ZIO_TYPE_WRITE)
2702 vdev_cache_write(zio);
2703
2704 if (zio_injection_enabled && zio->io_error == 0)
2705 zio->io_error = zio_handle_device_injection(vd,
2706 zio, EIO);
2707
2708 if (zio_injection_enabled && zio->io_error == 0)
2709 zio->io_error = zio_handle_label_injection(zio, EIO);
2710
2711 if (zio->io_error) {
2712 if (!vdev_accessible(vd, zio)) {
2713 zio->io_error = SET_ERROR(ENXIO);
2714 } else {
2715 unexpected_error = B_TRUE;
2716 }
2717 }
2718 }
2719
2720 ops->vdev_op_io_done(zio);
2721
2722 if (unexpected_error)
2723 VERIFY(vdev_probe(vd, zio) == NULL);
2724
2725 return (ZIO_PIPELINE_CONTINUE);
2726 }
2727
2728 /*
2729 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2730 * disk, and use that to finish the checksum ereport later.
2731 */
2732 static void
2733 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2734 const void *good_buf)
2735 {
2736 /* no processing needed */
2737 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2738 }
2739
2740 /*ARGSUSED*/
2741 void
2742 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2743 {
2744 void *buf = zio_buf_alloc(zio->io_size);
2745
2746 bcopy(zio->io_data, buf, zio->io_size);
2747
2748 zcr->zcr_cbinfo = zio->io_size;
2749 zcr->zcr_cbdata = buf;
2750 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2751 zcr->zcr_free = zio_buf_free;
2752 }
2753
2754 static int
2755 zio_vdev_io_assess(zio_t *zio)
2756 {
2757 vdev_t *vd = zio->io_vd;
2758
2759 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2760 return (ZIO_PIPELINE_STOP);
2761
2762 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2763 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2764
2765 if (zio->io_vsd != NULL) {
2766 zio->io_vsd_ops->vsd_free(zio);
2767 zio->io_vsd = NULL;
2768 }
2769
2770 if (zio_injection_enabled && zio->io_error == 0)
2771 zio->io_error = zio_handle_fault_injection(zio, EIO);
2772
2773 /*
2774 * If the I/O failed, determine whether we should attempt to retry it.
2775 *
2776 * On retry, we cut in line in the issue queue, since we don't want
2777 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2778 */
2779 if (zio->io_error && vd == NULL &&
2780 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2781 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2782 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2783 zio->io_error = 0;
2784 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2785 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2786 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2787 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2788 zio_requeue_io_start_cut_in_line);
2789 return (ZIO_PIPELINE_STOP);
2790 }
2791
2792 /*
2793 * If we got an error on a leaf device, convert it to ENXIO
2794 * if the device is not accessible at all.
2795 */
2796 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2797 !vdev_accessible(vd, zio))
2798 zio->io_error = SET_ERROR(ENXIO);
2799
2800 /*
2801 * If we can't write to an interior vdev (mirror or RAID-Z),
2802 * set vdev_cant_write so that we stop trying to allocate from it.
2803 */
2804 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2805 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2806 vd->vdev_cant_write = B_TRUE;
2807 }
2808
2809 if (zio->io_error)
2810 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2811
2812 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2813 zio->io_physdone != NULL) {
2814 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
2815 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
2816 zio->io_physdone(zio->io_logical);
2817 }
2818
2819 return (ZIO_PIPELINE_CONTINUE);
2820 }
2821
2822 void
2823 zio_vdev_io_reissue(zio_t *zio)
2824 {
2825 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2826 ASSERT(zio->io_error == 0);
2827
2828 zio->io_stage >>= 1;
2829 }
2830
2831 void
2832 zio_vdev_io_redone(zio_t *zio)
2833 {
2834 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2835
2836 zio->io_stage >>= 1;
2837 }
2838
2839 void
2840 zio_vdev_io_bypass(zio_t *zio)
2841 {
2842 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2843 ASSERT(zio->io_error == 0);
2844
2845 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2846 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2847 }
2848
2849 /*
2850 * ==========================================================================
2851 * Generate and verify checksums
2852 * ==========================================================================
2853 */
2854 static int
2855 zio_checksum_generate(zio_t *zio)
2856 {
2857 blkptr_t *bp = zio->io_bp;
2858 enum zio_checksum checksum;
2859
2860 if (bp == NULL) {
2861 /*
2862 * This is zio_write_phys().
2863 * We're either generating a label checksum, or none at all.
2864 */
2865 checksum = zio->io_prop.zp_checksum;
2866
2867 if (checksum == ZIO_CHECKSUM_OFF)
2868 return (ZIO_PIPELINE_CONTINUE);
2869
2870 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2871 } else {
2872 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2873 ASSERT(!IO_IS_ALLOCATING(zio));
2874 checksum = ZIO_CHECKSUM_GANG_HEADER;
2875 } else {
2876 checksum = BP_GET_CHECKSUM(bp);
2877 }
2878 }
2879
2880 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2881
2882 return (ZIO_PIPELINE_CONTINUE);
2883 }
2884
2885 static int
2886 zio_checksum_verify(zio_t *zio)
2887 {
2888 zio_bad_cksum_t info;
2889 blkptr_t *bp = zio->io_bp;
2890 int error;
2891
2892 ASSERT(zio->io_vd != NULL);
2893
2894 if (bp == NULL) {
2895 /*
2896 * This is zio_read_phys().
2897 * We're either verifying a label checksum, or nothing at all.
2898 */
2899 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2900 return (ZIO_PIPELINE_CONTINUE);
2901
2902 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2903 }
2904
2905 if ((error = zio_checksum_error(zio, &info)) != 0) {
2906 zio->io_error = error;
2907 if (error == ECKSUM &&
2908 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2909 zfs_ereport_start_checksum(zio->io_spa,
2910 zio->io_vd, zio, zio->io_offset,
2911 zio->io_size, NULL, &info);
2912 }
2913 }
2914
2915 return (ZIO_PIPELINE_CONTINUE);
2916 }
2917
2918 /*
2919 * Called by RAID-Z to ensure we don't compute the checksum twice.
2920 */
2921 void
2922 zio_checksum_verified(zio_t *zio)
2923 {
2924 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2925 }
2926
2927 /*
2928 * ==========================================================================
2929 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2930 * An error of 0 indicates success. ENXIO indicates whole-device failure,
2931 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2932 * indicate errors that are specific to one I/O, and most likely permanent.
2933 * Any other error is presumed to be worse because we weren't expecting it.
2934 * ==========================================================================
2935 */
2936 int
2937 zio_worst_error(int e1, int e2)
2938 {
2939 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2940 int r1, r2;
2941
2942 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2943 if (e1 == zio_error_rank[r1])
2944 break;
2945
2946 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2947 if (e2 == zio_error_rank[r2])
2948 break;
2949
2950 return (r1 > r2 ? e1 : e2);
2951 }
2952
2953 /*
2954 * ==========================================================================
2955 * I/O completion
2956 * ==========================================================================
2957 */
2958 static int
2959 zio_ready(zio_t *zio)
2960 {
2961 blkptr_t *bp = zio->io_bp;
2962 zio_t *pio, *pio_next;
2963
2964 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2965 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2966 return (ZIO_PIPELINE_STOP);
2967
2968 if (zio->io_ready) {
2969 ASSERT(IO_IS_ALLOCATING(zio));
2970 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
2971 (zio->io_flags & ZIO_FLAG_NOPWRITE));
2972 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2973
2974 zio->io_ready(zio);
2975 }
2976
2977 if (bp != NULL && bp != &zio->io_bp_copy)
2978 zio->io_bp_copy = *bp;
2979
2980 if (zio->io_error)
2981 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2982
2983 mutex_enter(&zio->io_lock);
2984 zio->io_state[ZIO_WAIT_READY] = 1;
2985 pio = zio_walk_parents(zio);
2986 mutex_exit(&zio->io_lock);
2987
2988 /*
2989 * As we notify zio's parents, new parents could be added.
2990 * New parents go to the head of zio's io_parent_list, however,
2991 * so we will (correctly) not notify them. The remainder of zio's
2992 * io_parent_list, from 'pio_next' onward, cannot change because
2993 * all parents must wait for us to be done before they can be done.
2994 */
2995 for (; pio != NULL; pio = pio_next) {
2996 pio_next = zio_walk_parents(zio);
2997 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2998 }
2999
3000 if (zio->io_flags & ZIO_FLAG_NODATA) {
3001 if (BP_IS_GANG(bp)) {
3002 zio->io_flags &= ~ZIO_FLAG_NODATA;
3003 } else {
3004 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
3005 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3006 }
3007 }
3008
3009 if (zio_injection_enabled &&
3010 zio->io_spa->spa_syncing_txg == zio->io_txg)
3011 zio_handle_ignored_writes(zio);
3012
3013 return (ZIO_PIPELINE_CONTINUE);
3014 }
3015
3016 static int
3017 zio_done(zio_t *zio)
3018 {
3019 spa_t *spa = zio->io_spa;
3020 zio_t *lio = zio->io_logical;
3021 blkptr_t *bp = zio->io_bp;
3022 vdev_t *vd = zio->io_vd;
3023 uint64_t psize = zio->io_size;
3024 zio_t *pio, *pio_next;
3025
3026 /*
3027 * If our children haven't all completed,
3028 * wait for them and then repeat this pipeline stage.
3029 */
3030 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
3031 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
3032 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
3033 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
3034 return (ZIO_PIPELINE_STOP);
3035
3036 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3037 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3038 ASSERT(zio->io_children[c][w] == 0);
3039
3040 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3041 ASSERT(bp->blk_pad[0] == 0);
3042 ASSERT(bp->blk_pad[1] == 0);
3043 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3044 (bp == zio_unique_parent(zio)->io_bp));
3045 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3046 zio->io_bp_override == NULL &&
3047 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3048 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3049 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3050 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3051 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3052 }
3053 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3054 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3055 }
3056
3057 /*
3058 * If there were child vdev/gang/ddt errors, they apply to us now.
3059 */
3060 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3061 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3062 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3063
3064 /*
3065 * If the I/O on the transformed data was successful, generate any
3066 * checksum reports now while we still have the transformed data.
3067 */
3068 if (zio->io_error == 0) {
3069 while (zio->io_cksum_report != NULL) {
3070 zio_cksum_report_t *zcr = zio->io_cksum_report;
3071 uint64_t align = zcr->zcr_align;
3072 uint64_t asize = P2ROUNDUP(psize, align);
3073 char *abuf = zio->io_data;
3074
3075 if (asize != psize) {
3076 abuf = zio_buf_alloc(asize);
3077 bcopy(zio->io_data, abuf, psize);
3078 bzero(abuf + psize, asize - psize);
3079 }
3080
3081 zio->io_cksum_report = zcr->zcr_next;
3082 zcr->zcr_next = NULL;
3083 zcr->zcr_finish(zcr, abuf);
3084 zfs_ereport_free_checksum(zcr);
3085
3086 if (asize != psize)
3087 zio_buf_free(abuf, asize);
3088 }
3089 }
3090
3091 zio_pop_transforms(zio); /* note: may set zio->io_error */
3092
3093 vdev_stat_update(zio, psize);
3094
3095 if (zio->io_error) {
3096 /*
3097 * If this I/O is attached to a particular vdev,
3098 * generate an error message describing the I/O failure
3099 * at the block level. We ignore these errors if the
3100 * device is currently unavailable.
3101 */
3102 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3103 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3104
3105 if ((zio->io_error == EIO || !(zio->io_flags &
3106 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3107 zio == lio) {
3108 /*
3109 * For logical I/O requests, tell the SPA to log the
3110 * error and generate a logical data ereport.
3111 */
3112 spa_log_error(spa, zio);
3113 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3114 0, 0);
3115 }
3116 }
3117
3118 if (zio->io_error && zio == lio) {
3119 /*
3120 * Determine whether zio should be reexecuted. This will
3121 * propagate all the way to the root via zio_notify_parent().
3122 */
3123 ASSERT(vd == NULL && bp != NULL);
3124 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3125
3126 if (IO_IS_ALLOCATING(zio) &&
3127 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3128 if (zio->io_error != ENOSPC)
3129 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3130 else
3131 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3132 }
3133
3134 if ((zio->io_type == ZIO_TYPE_READ ||
3135 zio->io_type == ZIO_TYPE_FREE) &&
3136 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3137 zio->io_error == ENXIO &&
3138 spa_load_state(spa) == SPA_LOAD_NONE &&
3139 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3140 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3141
3142 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3143 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3144
3145 /*
3146 * Here is a possibly good place to attempt to do
3147 * either combinatorial reconstruction or error correction
3148 * based on checksums. It also might be a good place
3149 * to send out preliminary ereports before we suspend
3150 * processing.
3151 */
3152 }
3153
3154 /*
3155 * If there were logical child errors, they apply to us now.
3156 * We defer this until now to avoid conflating logical child
3157 * errors with errors that happened to the zio itself when
3158 * updating vdev stats and reporting FMA events above.
3159 */
3160 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3161
3162 if ((zio->io_error || zio->io_reexecute) &&
3163 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3164 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3165 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3166
3167 zio_gang_tree_free(&zio->io_gang_tree);
3168
3169 /*
3170 * Godfather I/Os should never suspend.
3171 */
3172 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3173 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3174 zio->io_reexecute = 0;
3175
3176 if (zio->io_reexecute) {
3177 /*
3178 * This is a logical I/O that wants to reexecute.
3179 *
3180 * Reexecute is top-down. When an i/o fails, if it's not
3181 * the root, it simply notifies its parent and sticks around.
3182 * The parent, seeing that it still has children in zio_done(),
3183 * does the same. This percolates all the way up to the root.
3184 * The root i/o will reexecute or suspend the entire tree.
3185 *
3186 * This approach ensures that zio_reexecute() honors
3187 * all the original i/o dependency relationships, e.g.
3188 * parents not executing until children are ready.
3189 */
3190 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3191
3192 zio->io_gang_leader = NULL;
3193
3194 mutex_enter(&zio->io_lock);
3195 zio->io_state[ZIO_WAIT_DONE] = 1;
3196 mutex_exit(&zio->io_lock);
3197
3198 /*
3199 * "The Godfather" I/O monitors its children but is
3200 * not a true parent to them. It will track them through
3201 * the pipeline but severs its ties whenever they get into
3202 * trouble (e.g. suspended). This allows "The Godfather"
3203 * I/O to return status without blocking.
3204 */
3205 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3206 zio_link_t *zl = zio->io_walk_link;
3207 pio_next = zio_walk_parents(zio);
3208
3209 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3210 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3211 zio_remove_child(pio, zio, zl);
3212 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3213 }
3214 }
3215
3216 if ((pio = zio_unique_parent(zio)) != NULL) {
3217 /*
3218 * We're not a root i/o, so there's nothing to do
3219 * but notify our parent. Don't propagate errors
3220 * upward since we haven't permanently failed yet.
3221 */
3222 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3223 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3224 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3225 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3226 /*
3227 * We'd fail again if we reexecuted now, so suspend
3228 * until conditions improve (e.g. device comes online).
3229 */
3230 zio_suspend(spa, zio);
3231 } else {
3232 /*
3233 * Reexecution is potentially a huge amount of work.
3234 * Hand it off to the otherwise-unused claim taskq.
3235 */
3236 ASSERT(zio->io_tqent.tqent_next == NULL);
3237 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3238 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3239 0, &zio->io_tqent);
3240 }
3241 return (ZIO_PIPELINE_STOP);
3242 }
3243
3244 ASSERT(zio->io_child_count == 0);
3245 ASSERT(zio->io_reexecute == 0);
3246 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3247
3248 /*
3249 * Report any checksum errors, since the I/O is complete.
3250 */
3251 while (zio->io_cksum_report != NULL) {
3252 zio_cksum_report_t *zcr = zio->io_cksum_report;
3253 zio->io_cksum_report = zcr->zcr_next;
3254 zcr->zcr_next = NULL;
3255 zcr->zcr_finish(zcr, NULL);
3256 zfs_ereport_free_checksum(zcr);
3257 }
3258
3259 /*
3260 * It is the responsibility of the done callback to ensure that this
3261 * particular zio is no longer discoverable for adoption, and as
3262 * such, cannot acquire any new parents.
3263 */
3264 if (zio->io_done)
3265 zio->io_done(zio);
3266
3267 mutex_enter(&zio->io_lock);
3268 zio->io_state[ZIO_WAIT_DONE] = 1;
3269 mutex_exit(&zio->io_lock);
3270
3271 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3272 zio_link_t *zl = zio->io_walk_link;
3273 pio_next = zio_walk_parents(zio);
3274 zio_remove_child(pio, zio, zl);
3275 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3276 }
3277
3278 if (zio->io_waiter != NULL) {
3279 mutex_enter(&zio->io_lock);
3280 zio->io_executor = NULL;
3281 cv_broadcast(&zio->io_cv);
3282 mutex_exit(&zio->io_lock);
3283 } else {
3284 zio_destroy(zio);
3285 }
3286
3287 return (ZIO_PIPELINE_STOP);
3288 }
3289
3290 /*
3291 * ==========================================================================
3292 * I/O pipeline definition
3293 * ==========================================================================
3294 */
3295 static zio_pipe_stage_t *zio_pipeline[] = {
3296 NULL,
3297 zio_read_bp_init,
3298 zio_free_bp_init,
3299 zio_issue_async,
3300 zio_write_bp_init,
3301 zio_checksum_generate,
3302 zio_nop_write,
3303 zio_ddt_read_start,
3304 zio_ddt_read_done,
3305 zio_ddt_write,
3306 zio_ddt_free,
3307 zio_gang_assemble,
3308 zio_gang_issue,
3309 zio_dva_allocate,
3310 zio_dva_free,
3311 zio_dva_claim,
3312 zio_ready,
3313 zio_vdev_io_start,
3314 zio_vdev_io_done,
3315 zio_vdev_io_assess,
3316 zio_checksum_verify,
3317 zio_done
3318 };
3319
3320 /* dnp is the dnode for zb1->zb_object */
3321 boolean_t
3322 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_phys_t *zb1,
3323 const zbookmark_phys_t *zb2)
3324 {
3325 uint64_t zb1nextL0, zb2thisobj;
3326
3327 ASSERT(zb1->zb_objset == zb2->zb_objset);
3328 ASSERT(zb2->zb_level == 0);
3329
3330 /* The objset_phys_t isn't before anything. */
3331 if (dnp == NULL)
3332 return (B_FALSE);
3333
3334 zb1nextL0 = (zb1->zb_blkid + 1) <<
3335 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3336
3337 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3338 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3339
3340 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3341 uint64_t nextobj = zb1nextL0 *
3342 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3343 return (nextobj <= zb2thisobj);
3344 }
3345
3346 if (zb1->zb_object < zb2thisobj)
3347 return (B_TRUE);
3348 if (zb1->zb_object > zb2thisobj)
3349 return (B_FALSE);
3350 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3351 return (B_FALSE);
3352 return (zb1nextL0 <= zb2->zb_blkid);
3353 }