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, 2017 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25 * Copyright (c) 2014 Integros [integros.com]
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/blkptr.h>
42 #include <sys/zfeature.h>
43 #include <sys/metaslab_impl.h>
44 #include <sys/abd.h>
45
46 /*
47 * ==========================================================================
48 * I/O type descriptions
49 * ==========================================================================
50 */
51 const char *zio_type_name[ZIO_TYPES] = {
52 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
53 "zio_ioctl"
54 };
55
56 boolean_t zio_dva_throttle_enabled = B_TRUE;
57
58 /*
59 * ==========================================================================
60 * I/O kmem caches
61 * ==========================================================================
62 */
63 kmem_cache_t *zio_cache;
64 kmem_cache_t *zio_link_cache;
65 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
66 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
67
68 #ifdef _KERNEL
69 extern vmem_t *zio_alloc_arena;
70 #endif
71
72 #define ZIO_PIPELINE_CONTINUE 0x100
73 #define ZIO_PIPELINE_STOP 0x101
74
75 #define BP_SPANB(indblkshift, level) \
76 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
77 #define COMPARE_META_LEVEL 0x80000000ul
78 /*
79 * The following actions directly effect the spa's sync-to-convergence logic.
80 * The values below define the sync pass when we start performing the action.
81 * Care should be taken when changing these values as they directly impact
82 * spa_sync() performance. Tuning these values may introduce subtle performance
83 * pathologies and should only be done in the context of performance analysis.
84 * These tunables will eventually be removed and replaced with #defines once
85 * enough analysis has been done to determine optimal values.
86 *
87 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
88 * regular blocks are not deferred.
89 */
90 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
91 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
92 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
93
94 /*
95 * An allocating zio is one that either currently has the DVA allocate
96 * stage set or will have it later in its lifetime.
97 */
98 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
99
100 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
101
102 #ifdef ZFS_DEBUG
103 int zio_buf_debug_limit = 16384;
104 #else
105 int zio_buf_debug_limit = 0;
106 #endif
107
108 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
109
110 void
111 zio_init(void)
112 {
113 size_t c;
114 vmem_t *data_alloc_arena = NULL;
115
116 #ifdef _KERNEL
117 data_alloc_arena = zio_alloc_arena;
118 #endif
119 zio_cache = kmem_cache_create("zio_cache",
120 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
121 zio_link_cache = kmem_cache_create("zio_link_cache",
122 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
123
124 /*
125 * For small buffers, we want a cache for each multiple of
126 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
127 * for each quarter-power of 2.
128 */
129 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
130 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
131 size_t p2 = size;
132 size_t align = 0;
133 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
134
135 while (!ISP2(p2))
136 p2 &= p2 - 1;
137
138 #ifndef _KERNEL
139 /*
140 * If we are using watchpoints, put each buffer on its own page,
141 * to eliminate the performance overhead of trapping to the
142 * kernel when modifying a non-watched buffer that shares the
143 * page with a watched buffer.
144 */
145 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
146 continue;
147 #endif
148 if (size <= 4 * SPA_MINBLOCKSIZE) {
149 align = SPA_MINBLOCKSIZE;
150 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
151 align = MIN(p2 >> 2, PAGESIZE);
152 }
153
154 if (align != 0) {
155 char name[36];
156 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
157 zio_buf_cache[c] = kmem_cache_create(name, size,
158 align, NULL, NULL, NULL, NULL, NULL, cflags);
159
160 /*
161 * Since zio_data bufs do not appear in crash dumps, we
162 * pass KMC_NOTOUCH so that no allocator metadata is
163 * stored with the buffers.
164 */
165 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
166 zio_data_buf_cache[c] = kmem_cache_create(name, size,
167 align, NULL, NULL, NULL, NULL, data_alloc_arena,
168 cflags | KMC_NOTOUCH);
169 }
170 }
171
172 while (--c != 0) {
173 ASSERT(zio_buf_cache[c] != NULL);
174 if (zio_buf_cache[c - 1] == NULL)
175 zio_buf_cache[c - 1] = zio_buf_cache[c];
176
177 ASSERT(zio_data_buf_cache[c] != NULL);
178 if (zio_data_buf_cache[c - 1] == NULL)
179 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
180 }
181
182 zio_inject_init();
183 }
184
185 void
186 zio_fini(void)
187 {
188 size_t c;
189 kmem_cache_t *last_cache = NULL;
190 kmem_cache_t *last_data_cache = NULL;
191
192 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
193 if (zio_buf_cache[c] != last_cache) {
194 last_cache = zio_buf_cache[c];
195 kmem_cache_destroy(zio_buf_cache[c]);
196 }
197 zio_buf_cache[c] = NULL;
198
199 if (zio_data_buf_cache[c] != last_data_cache) {
200 last_data_cache = zio_data_buf_cache[c];
201 kmem_cache_destroy(zio_data_buf_cache[c]);
202 }
203 zio_data_buf_cache[c] = NULL;
204 }
205
206 kmem_cache_destroy(zio_link_cache);
207 kmem_cache_destroy(zio_cache);
208
209 zio_inject_fini();
210 }
211
212 /*
213 * ==========================================================================
214 * Allocate and free I/O buffers
215 * ==========================================================================
216 */
217
218 /*
219 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
220 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
221 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
222 * excess / transient data in-core during a crashdump.
223 */
224 void *
225 zio_buf_alloc(size_t size)
226 {
227 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
228
229 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
230
231 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
232 }
233
234 /*
235 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
236 * crashdump if the kernel panics. This exists so that we will limit the amount
237 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
238 * of kernel heap dumped to disk when the kernel panics)
239 */
240 void *
241 zio_data_buf_alloc(size_t size)
242 {
243 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
244
245 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
246
247 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
248 }
249
250 void
251 zio_buf_free(void *buf, size_t size)
252 {
253 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
254
255 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
256
257 kmem_cache_free(zio_buf_cache[c], buf);
258 }
259
260 void
261 zio_data_buf_free(void *buf, size_t size)
262 {
263 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
264
265 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
266
267 kmem_cache_free(zio_data_buf_cache[c], buf);
268 }
269
270 /*
271 * ==========================================================================
272 * Push and pop I/O transform buffers
273 * ==========================================================================
274 */
275 void
276 zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize,
277 zio_transform_func_t *transform)
278 {
279 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
280
281 /*
282 * Ensure that anyone expecting this zio to contain a linear ABD isn't
283 * going to get a nasty surprise when they try to access the data.
284 */
285 IMPLY(abd_is_linear(zio->io_abd), abd_is_linear(data));
286
287 zt->zt_orig_abd = zio->io_abd;
288 zt->zt_orig_size = zio->io_size;
289 zt->zt_bufsize = bufsize;
290 zt->zt_transform = transform;
291
292 zt->zt_next = zio->io_transform_stack;
293 zio->io_transform_stack = zt;
294
295 zio->io_abd = data;
296 zio->io_size = size;
297 }
298
299 void
300 zio_pop_transforms(zio_t *zio)
301 {
302 zio_transform_t *zt;
303
304 while ((zt = zio->io_transform_stack) != NULL) {
305 if (zt->zt_transform != NULL)
306 zt->zt_transform(zio,
307 zt->zt_orig_abd, zt->zt_orig_size);
308
309 if (zt->zt_bufsize != 0)
310 abd_free(zio->io_abd);
311
312 zio->io_abd = zt->zt_orig_abd;
313 zio->io_size = zt->zt_orig_size;
314 zio->io_transform_stack = zt->zt_next;
315
316 kmem_free(zt, sizeof (zio_transform_t));
317 }
318 }
319
320 /*
321 * ==========================================================================
322 * I/O transform callbacks for subblocks and decompression
323 * ==========================================================================
324 */
325 static void
326 zio_subblock(zio_t *zio, abd_t *data, uint64_t size)
327 {
328 ASSERT(zio->io_size > size);
329
330 if (zio->io_type == ZIO_TYPE_READ)
331 abd_copy(data, zio->io_abd, size);
332 }
333
334 static void
335 zio_decompress(zio_t *zio, abd_t *data, uint64_t size)
336 {
337 if (zio->io_error == 0) {
338 void *tmp = abd_borrow_buf(data, size);
339 int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
340 zio->io_abd, tmp, zio->io_size, size);
341 abd_return_buf_copy(data, tmp, size);
342
343 if (ret != 0)
344 zio->io_error = SET_ERROR(EIO);
345 }
346 }
347
348 /*
349 * ==========================================================================
350 * I/O parent/child relationships and pipeline interlocks
351 * ==========================================================================
352 */
353 zio_t *
354 zio_walk_parents(zio_t *cio, zio_link_t **zl)
355 {
356 list_t *pl = &cio->io_parent_list;
357
358 *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
359 if (*zl == NULL)
360 return (NULL);
361
362 ASSERT((*zl)->zl_child == cio);
363 return ((*zl)->zl_parent);
364 }
365
366 zio_t *
367 zio_walk_children(zio_t *pio, zio_link_t **zl)
368 {
369 list_t *cl = &pio->io_child_list;
370
371 *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
372 if (*zl == NULL)
373 return (NULL);
374
375 ASSERT((*zl)->zl_parent == pio);
376 return ((*zl)->zl_child);
377 }
378
379 zio_t *
380 zio_unique_parent(zio_t *cio)
381 {
382 zio_link_t *zl = NULL;
383 zio_t *pio = zio_walk_parents(cio, &zl);
384
385 VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
386 return (pio);
387 }
388
389 void
390 zio_add_child(zio_t *pio, zio_t *cio)
391 {
392 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
393
394 /*
395 * Logical I/Os can have logical, gang, or vdev children.
396 * Gang I/Os can have gang or vdev children.
397 * Vdev I/Os can only have vdev children.
398 * The following ASSERT captures all of these constraints.
399 */
400 ASSERT3S(cio->io_child_type, <=, pio->io_child_type);
401
402 zl->zl_parent = pio;
403 zl->zl_child = cio;
404
405 mutex_enter(&cio->io_lock);
406 mutex_enter(&pio->io_lock);
407
408 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
409
410 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
411 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
412
413 list_insert_head(&pio->io_child_list, zl);
414 list_insert_head(&cio->io_parent_list, zl);
415
416 pio->io_child_count++;
417 cio->io_parent_count++;
418
419 mutex_exit(&pio->io_lock);
420 mutex_exit(&cio->io_lock);
421 }
422
423 static void
424 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
425 {
426 ASSERT(zl->zl_parent == pio);
427 ASSERT(zl->zl_child == cio);
428
429 mutex_enter(&cio->io_lock);
430 mutex_enter(&pio->io_lock);
431
432 list_remove(&pio->io_child_list, zl);
433 list_remove(&cio->io_parent_list, zl);
434
435 pio->io_child_count--;
436 cio->io_parent_count--;
437
438 mutex_exit(&pio->io_lock);
439 mutex_exit(&cio->io_lock);
440
441 kmem_cache_free(zio_link_cache, zl);
442 }
443
444 static boolean_t
445 zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait)
446 {
447 boolean_t waiting = B_FALSE;
448
449 mutex_enter(&zio->io_lock);
450 ASSERT(zio->io_stall == NULL);
451 for (int c = 0; c < ZIO_CHILD_TYPES; c++) {
452 if (!(ZIO_CHILD_BIT_IS_SET(childbits, c)))
453 continue;
454
455 uint64_t *countp = &zio->io_children[c][wait];
456 if (*countp != 0) {
457 zio->io_stage >>= 1;
458 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
459 zio->io_stall = countp;
460 waiting = B_TRUE;
461 break;
462 }
463 }
464 mutex_exit(&zio->io_lock);
465 return (waiting);
466 }
467
468 static void
469 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
470 {
471 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
472 int *errorp = &pio->io_child_error[zio->io_child_type];
473
474 mutex_enter(&pio->io_lock);
475 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
476 *errorp = zio_worst_error(*errorp, zio->io_error);
477 pio->io_reexecute |= zio->io_reexecute;
478 ASSERT3U(*countp, >, 0);
479
480 (*countp)--;
481
482 if (*countp == 0 && pio->io_stall == countp) {
483 zio_taskq_type_t type =
484 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
485 ZIO_TASKQ_INTERRUPT;
486 pio->io_stall = NULL;
487 mutex_exit(&pio->io_lock);
488 /*
489 * Dispatch the parent zio in its own taskq so that
490 * the child can continue to make progress. This also
491 * prevents overflowing the stack when we have deeply nested
492 * parent-child relationships.
493 */
494 zio_taskq_dispatch(pio, type, B_FALSE);
495 } else {
496 mutex_exit(&pio->io_lock);
497 }
498 }
499
500 static void
501 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
502 {
503 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
504 zio->io_error = zio->io_child_error[c];
505 }
506
507 int
508 zio_bookmark_compare(const void *x1, const void *x2)
509 {
510 const zio_t *z1 = x1;
511 const zio_t *z2 = x2;
512
513 if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
514 return (-1);
515 if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
516 return (1);
517
518 if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
519 return (-1);
520 if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
521 return (1);
522
523 if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
524 return (-1);
525 if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
526 return (1);
527
528 if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
529 return (-1);
530 if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
531 return (1);
532
533 if (z1 < z2)
534 return (-1);
535 if (z1 > z2)
536 return (1);
537
538 return (0);
539 }
540
541 /*
542 * ==========================================================================
543 * Create the various types of I/O (read, write, free, etc)
544 * ==========================================================================
545 */
546 static zio_t *
547 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
548 abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done,
549 void *private, zio_type_t type, zio_priority_t priority,
550 enum zio_flag flags, vdev_t *vd, uint64_t offset,
551 const zbookmark_phys_t *zb, enum zio_stage stage, enum zio_stage pipeline)
552 {
553 zio_t *zio;
554
555 ASSERT3U(psize, <=, SPA_MAXBLOCKSIZE);
556 ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0);
557 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
558
559 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
560 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
561 ASSERT(vd || stage == ZIO_STAGE_OPEN);
562
563 IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW) != 0);
564
565 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
566 bzero(zio, sizeof (zio_t));
567
568 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
569 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
570
571 list_create(&zio->io_parent_list, sizeof (zio_link_t),
572 offsetof(zio_link_t, zl_parent_node));
573 list_create(&zio->io_child_list, sizeof (zio_link_t),
574 offsetof(zio_link_t, zl_child_node));
575 metaslab_trace_init(&zio->io_alloc_list);
576
577 if (vd != NULL)
578 zio->io_child_type = ZIO_CHILD_VDEV;
579 else if (flags & ZIO_FLAG_GANG_CHILD)
580 zio->io_child_type = ZIO_CHILD_GANG;
581 else if (flags & ZIO_FLAG_DDT_CHILD)
582 zio->io_child_type = ZIO_CHILD_DDT;
583 else
584 zio->io_child_type = ZIO_CHILD_LOGICAL;
585
586 if (bp != NULL) {
587 zio->io_bp = (blkptr_t *)bp;
588 zio->io_bp_copy = *bp;
589 zio->io_bp_orig = *bp;
590 if (type != ZIO_TYPE_WRITE ||
591 zio->io_child_type == ZIO_CHILD_DDT)
592 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
593 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
594 zio->io_logical = zio;
595 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
596 pipeline |= ZIO_GANG_STAGES;
597 }
598
599 zio->io_spa = spa;
600 zio->io_txg = txg;
601 zio->io_done = done;
602 zio->io_private = private;
603 zio->io_type = type;
604 zio->io_priority = priority;
605 zio->io_vd = vd;
606 zio->io_offset = offset;
607 zio->io_orig_abd = zio->io_abd = data;
608 zio->io_orig_size = zio->io_size = psize;
609 zio->io_lsize = lsize;
610 zio->io_orig_flags = zio->io_flags = flags;
611 zio->io_orig_stage = zio->io_stage = stage;
612 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
613 zio->io_pipeline_trace = ZIO_STAGE_OPEN;
614
615 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
616 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
617
618 if (zb != NULL)
619 zio->io_bookmark = *zb;
620
621 if (pio != NULL) {
622 if (zio->io_logical == NULL)
623 zio->io_logical = pio->io_logical;
624 if (zio->io_child_type == ZIO_CHILD_GANG)
625 zio->io_gang_leader = pio->io_gang_leader;
626 zio_add_child(pio, zio);
627 }
628
629 return (zio);
630 }
631
632 static void
633 zio_destroy(zio_t *zio)
634 {
635 metaslab_trace_fini(&zio->io_alloc_list);
636 list_destroy(&zio->io_parent_list);
637 list_destroy(&zio->io_child_list);
638 mutex_destroy(&zio->io_lock);
639 cv_destroy(&zio->io_cv);
640 kmem_cache_free(zio_cache, zio);
641 }
642
643 zio_t *
644 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
645 void *private, enum zio_flag flags)
646 {
647 zio_t *zio;
648
649 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
650 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
651 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
652
653 return (zio);
654 }
655
656 zio_t *
657 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
658 {
659 return (zio_null(NULL, spa, NULL, done, private, flags));
660 }
661
662 void
663 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
664 {
665 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
666 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
667 bp, (longlong_t)BP_GET_TYPE(bp));
668 }
669 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
670 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
671 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
672 bp, (longlong_t)BP_GET_CHECKSUM(bp));
673 }
674 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
675 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
676 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
677 bp, (longlong_t)BP_GET_COMPRESS(bp));
678 }
679 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
680 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
681 bp, (longlong_t)BP_GET_LSIZE(bp));
682 }
683 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
684 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
685 bp, (longlong_t)BP_GET_PSIZE(bp));
686 }
687
688 if (BP_IS_EMBEDDED(bp)) {
689 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
690 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
691 bp, (longlong_t)BPE_GET_ETYPE(bp));
692 }
693 }
694
695 /*
696 * Do not verify individual DVAs if the config is not trusted. This
697 * will be done once the zio is executed in vdev_mirror_map_alloc.
698 */
699 if (!spa->spa_trust_config)
700 return;
701
702 /*
703 * Pool-specific checks.
704 *
705 * Note: it would be nice to verify that the blk_birth and
706 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
707 * allows the birth time of log blocks (and dmu_sync()-ed blocks
708 * that are in the log) to be arbitrarily large.
709 */
710 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
711 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
712 if (vdevid >= spa->spa_root_vdev->vdev_children) {
713 zfs_panic_recover("blkptr at %p DVA %u has invalid "
714 "VDEV %llu",
715 bp, i, (longlong_t)vdevid);
716 continue;
717 }
718 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
719 if (vd == NULL) {
720 zfs_panic_recover("blkptr at %p DVA %u has invalid "
721 "VDEV %llu",
722 bp, i, (longlong_t)vdevid);
723 continue;
724 }
725 if (vd->vdev_ops == &vdev_hole_ops) {
726 zfs_panic_recover("blkptr at %p DVA %u has hole "
727 "VDEV %llu",
728 bp, i, (longlong_t)vdevid);
729 continue;
730 }
731 if (vd->vdev_ops == &vdev_missing_ops) {
732 /*
733 * "missing" vdevs are valid during import, but we
734 * don't have their detailed info (e.g. asize), so
735 * we can't perform any more checks on them.
736 */
737 continue;
738 }
739 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
740 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
741 if (BP_IS_GANG(bp))
742 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
743 if (offset + asize > vd->vdev_asize) {
744 zfs_panic_recover("blkptr at %p DVA %u has invalid "
745 "OFFSET %llu",
746 bp, i, (longlong_t)offset);
747 }
748 }
749 }
750
751 boolean_t
752 zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp)
753 {
754 uint64_t vdevid = DVA_GET_VDEV(dva);
755
756 if (vdevid >= spa->spa_root_vdev->vdev_children)
757 return (B_FALSE);
758
759 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
760 if (vd == NULL)
761 return (B_FALSE);
762
763 if (vd->vdev_ops == &vdev_hole_ops)
764 return (B_FALSE);
765
766 if (vd->vdev_ops == &vdev_missing_ops) {
767 return (B_FALSE);
768 }
769
770 uint64_t offset = DVA_GET_OFFSET(dva);
771 uint64_t asize = DVA_GET_ASIZE(dva);
772
773 if (BP_IS_GANG(bp))
774 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
775 if (offset + asize > vd->vdev_asize)
776 return (B_FALSE);
777
778 return (B_TRUE);
779 }
780
781 zio_t *
782 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
783 abd_t *data, uint64_t size, zio_done_func_t *done, void *private,
784 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
785 {
786 zio_t *zio;
787
788 zfs_blkptr_verify(spa, bp);
789
790 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
791 data, size, size, done, private,
792 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
793 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
794 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
795
796 return (zio);
797 }
798
799 zio_t *
800 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
801 abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
802 zio_done_func_t *ready, zio_done_func_t *children_ready,
803 zio_done_func_t *physdone, zio_done_func_t *done,
804 void *private, zio_priority_t priority, enum zio_flag flags,
805 const zbookmark_phys_t *zb)
806 {
807 zio_t *zio;
808
809 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
810 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
811 zp->zp_compress >= ZIO_COMPRESS_OFF &&
812 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
813 DMU_OT_IS_VALID(zp->zp_type) &&
814 zp->zp_level < 32 &&
815 zp->zp_copies > 0 &&
816 zp->zp_copies <= spa_max_replication(spa));
817
818 zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
819 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
820 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
821 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
822
823 zio->io_ready = ready;
824 zio->io_children_ready = children_ready;
825 zio->io_physdone = physdone;
826 zio->io_prop = *zp;
827
828 /*
829 * Data can be NULL if we are going to call zio_write_override() to
830 * provide the already-allocated BP. But we may need the data to
831 * verify a dedup hit (if requested). In this case, don't try to
832 * dedup (just take the already-allocated BP verbatim).
833 */
834 if (data == NULL && zio->io_prop.zp_dedup_verify) {
835 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
836 }
837
838 return (zio);
839 }
840
841 zio_t *
842 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data,
843 uint64_t size, zio_done_func_t *done, void *private,
844 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
845 {
846 zio_t *zio;
847
848 zio = zio_create(pio, spa, txg, bp, data, size, size, done, private,
849 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
850 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
851
852 return (zio);
853 }
854
855 void
856 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
857 {
858 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
859 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
860 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
861 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
862
863 /*
864 * We must reset the io_prop to match the values that existed
865 * when the bp was first written by dmu_sync() keeping in mind
866 * that nopwrite and dedup are mutually exclusive.
867 */
868 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
869 zio->io_prop.zp_nopwrite = nopwrite;
870 zio->io_prop.zp_copies = copies;
871 zio->io_bp_override = bp;
872 }
873
874 void
875 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
876 {
877
878 zfs_blkptr_verify(spa, bp);
879
880 /*
881 * The check for EMBEDDED is a performance optimization. We
882 * process the free here (by ignoring it) rather than
883 * putting it on the list and then processing it in zio_free_sync().
884 */
885 if (BP_IS_EMBEDDED(bp))
886 return;
887 metaslab_check_free(spa, bp);
888
889 /*
890 * Frees that are for the currently-syncing txg, are not going to be
891 * deferred, and which will not need to do a read (i.e. not GANG or
892 * DEDUP), can be processed immediately. Otherwise, put them on the
893 * in-memory list for later processing.
894 */
895 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
896 txg != spa->spa_syncing_txg ||
897 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
898 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
899 } else {
900 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0)));
901 }
902 }
903
904 zio_t *
905 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
906 enum zio_flag flags)
907 {
908 zio_t *zio;
909 enum zio_stage stage = ZIO_FREE_PIPELINE;
910
911 ASSERT(!BP_IS_HOLE(bp));
912 ASSERT(spa_syncing_txg(spa) == txg);
913 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
914
915 if (BP_IS_EMBEDDED(bp))
916 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
917
918 metaslab_check_free(spa, bp);
919 arc_freed(spa, bp);
920
921 /*
922 * GANG and DEDUP blocks can induce a read (for the gang block header,
923 * or the DDT), so issue them asynchronously so that this thread is
924 * not tied up.
925 */
926 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
927 stage |= ZIO_STAGE_ISSUE_ASYNC;
928
929 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
930 BP_GET_PSIZE(bp), NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
931 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
932
933 return (zio);
934 }
935
936 zio_t *
937 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
938 zio_done_func_t *done, void *private, enum zio_flag flags)
939 {
940 zio_t *zio;
941
942 zfs_blkptr_verify(spa, bp);
943
944 if (BP_IS_EMBEDDED(bp))
945 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
946
947 /*
948 * A claim is an allocation of a specific block. Claims are needed
949 * to support immediate writes in the intent log. The issue is that
950 * immediate writes contain committed data, but in a txg that was
951 * *not* committed. Upon opening the pool after an unclean shutdown,
952 * the intent log claims all blocks that contain immediate write data
953 * so that the SPA knows they're in use.
954 *
955 * All claims *must* be resolved in the first txg -- before the SPA
956 * starts allocating blocks -- so that nothing is allocated twice.
957 * If txg == 0 we just verify that the block is claimable.
958 */
959 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
960 ASSERT(txg == spa_first_txg(spa) || txg == 0);
961 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
962
963 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
964 BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
965 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
966 ASSERT0(zio->io_queued_timestamp);
967
968 return (zio);
969 }
970
971 zio_t *
972 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
973 zio_done_func_t *done, void *private, enum zio_flag flags)
974 {
975 zio_t *zio;
976 int c;
977
978 if (vd->vdev_children == 0) {
979 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
980 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
981 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
982
983 zio->io_cmd = cmd;
984 } else {
985 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
986
987 for (c = 0; c < vd->vdev_children; c++)
988 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
989 done, private, flags));
990 }
991
992 return (zio);
993 }
994
995 zio_t *
996 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
997 abd_t *data, int checksum, zio_done_func_t *done, void *private,
998 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
999 {
1000 zio_t *zio;
1001
1002 ASSERT(vd->vdev_children == 0);
1003 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1004 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1005 ASSERT3U(offset + size, <=, vd->vdev_psize);
1006
1007 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1008 private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1009 offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1010
1011 zio->io_prop.zp_checksum = checksum;
1012
1013 return (zio);
1014 }
1015
1016 zio_t *
1017 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1018 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1019 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1020 {
1021 zio_t *zio;
1022
1023 ASSERT(vd->vdev_children == 0);
1024 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1025 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1026 ASSERT3U(offset + size, <=, vd->vdev_psize);
1027
1028 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1029 private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1030 offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1031
1032 zio->io_prop.zp_checksum = checksum;
1033
1034 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1035 /*
1036 * zec checksums are necessarily destructive -- they modify
1037 * the end of the write buffer to hold the verifier/checksum.
1038 * Therefore, we must make a local copy in case the data is
1039 * being written to multiple places in parallel.
1040 */
1041 abd_t *wbuf = abd_alloc_sametype(data, size);
1042 abd_copy(wbuf, data, size);
1043
1044 zio_push_transform(zio, wbuf, size, size, NULL);
1045 }
1046
1047 return (zio);
1048 }
1049
1050 /*
1051 * Create a child I/O to do some work for us.
1052 */
1053 zio_t *
1054 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1055 abd_t *data, uint64_t size, int type, zio_priority_t priority,
1056 enum zio_flag flags, zio_done_func_t *done, void *private)
1057 {
1058 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1059 zio_t *zio;
1060
1061 /*
1062 * vdev child I/Os do not propagate their error to the parent.
1063 * Therefore, for correct operation the caller *must* check for
1064 * and handle the error in the child i/o's done callback.
1065 * The only exceptions are i/os that we don't care about
1066 * (OPTIONAL or REPAIR).
1067 */
1068 ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) ||
1069 done != NULL);
1070
1071 /*
1072 * In the common case, where the parent zio was to a normal vdev,
1073 * the child zio must be to a child vdev of that vdev. Otherwise,
1074 * the child zio must be to a top-level vdev.
1075 */
1076 if (pio->io_vd != NULL && pio->io_vd->vdev_ops != &vdev_indirect_ops) {
1077 ASSERT3P(vd->vdev_parent, ==, pio->io_vd);
1078 } else {
1079 ASSERT3P(vd, ==, vd->vdev_top);
1080 }
1081
1082 if (type == ZIO_TYPE_READ && bp != NULL) {
1083 /*
1084 * If we have the bp, then the child should perform the
1085 * checksum and the parent need not. This pushes error
1086 * detection as close to the leaves as possible and
1087 * eliminates redundant checksums in the interior nodes.
1088 */
1089 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1090 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1091 }
1092
1093 if (vd->vdev_ops->vdev_op_leaf) {
1094 ASSERT0(vd->vdev_children);
1095 offset += VDEV_LABEL_START_SIZE;
1096 }
1097
1098 flags |= ZIO_VDEV_CHILD_FLAGS(pio);
1099
1100 /*
1101 * If we've decided to do a repair, the write is not speculative --
1102 * even if the original read was.
1103 */
1104 if (flags & ZIO_FLAG_IO_REPAIR)
1105 flags &= ~ZIO_FLAG_SPECULATIVE;
1106
1107 /*
1108 * If we're creating a child I/O that is not associated with a
1109 * top-level vdev, then the child zio is not an allocating I/O.
1110 * If this is a retried I/O then we ignore it since we will
1111 * have already processed the original allocating I/O.
1112 */
1113 if (flags & ZIO_FLAG_IO_ALLOCATING &&
1114 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1115 metaslab_class_t *mc = spa_normal_class(pio->io_spa);
1116
1117 ASSERT(mc->mc_alloc_throttle_enabled);
1118 ASSERT(type == ZIO_TYPE_WRITE);
1119 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1120 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1121 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1122 pio->io_child_type == ZIO_CHILD_GANG);
1123
1124 flags &= ~ZIO_FLAG_IO_ALLOCATING;
1125 }
1126
1127 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1128 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1129 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1130 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1131
1132 zio->io_physdone = pio->io_physdone;
1133 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1134 zio->io_logical->io_phys_children++;
1135
1136 return (zio);
1137 }
1138
1139 zio_t *
1140 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size,
1141 int type, zio_priority_t priority, enum zio_flag flags,
1142 zio_done_func_t *done, void *private)
1143 {
1144 zio_t *zio;
1145
1146 ASSERT(vd->vdev_ops->vdev_op_leaf);
1147
1148 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1149 data, size, size, done, private, type, priority,
1150 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1151 vd, offset, NULL,
1152 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1153
1154 return (zio);
1155 }
1156
1157 void
1158 zio_flush(zio_t *zio, vdev_t *vd)
1159 {
1160 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
1161 NULL, NULL,
1162 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1163 }
1164
1165 void
1166 zio_shrink(zio_t *zio, uint64_t size)
1167 {
1168 ASSERT3P(zio->io_executor, ==, NULL);
1169 ASSERT3P(zio->io_orig_size, ==, zio->io_size);
1170 ASSERT3U(size, <=, zio->io_size);
1171
1172 /*
1173 * We don't shrink for raidz because of problems with the
1174 * reconstruction when reading back less than the block size.
1175 * Note, BP_IS_RAIDZ() assumes no compression.
1176 */
1177 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1178 if (!BP_IS_RAIDZ(zio->io_bp)) {
1179 /* we are not doing a raw write */
1180 ASSERT3U(zio->io_size, ==, zio->io_lsize);
1181 zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1182 }
1183 }
1184
1185 /*
1186 * ==========================================================================
1187 * Prepare to read and write logical blocks
1188 * ==========================================================================
1189 */
1190
1191 static int
1192 zio_read_bp_init(zio_t *zio)
1193 {
1194 blkptr_t *bp = zio->io_bp;
1195
1196 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1197
1198 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1199 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1200 !(zio->io_flags & ZIO_FLAG_RAW)) {
1201 uint64_t psize =
1202 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1203 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1204 psize, psize, zio_decompress);
1205 }
1206
1207 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1208 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1209
1210 int psize = BPE_GET_PSIZE(bp);
1211 void *data = abd_borrow_buf(zio->io_abd, psize);
1212 decode_embedded_bp_compressed(bp, data);
1213 abd_return_buf_copy(zio->io_abd, data, psize);
1214 } else {
1215 ASSERT(!BP_IS_EMBEDDED(bp));
1216 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1217 }
1218
1219 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1220 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1221
1222 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1223 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1224
1225 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1226 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1227
1228 return (ZIO_PIPELINE_CONTINUE);
1229 }
1230
1231 static int
1232 zio_write_bp_init(zio_t *zio)
1233 {
1234 if (!IO_IS_ALLOCATING(zio))
1235 return (ZIO_PIPELINE_CONTINUE);
1236
1237 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1238
1239 if (zio->io_bp_override) {
1240 blkptr_t *bp = zio->io_bp;
1241 zio_prop_t *zp = &zio->io_prop;
1242
1243 ASSERT(bp->blk_birth != zio->io_txg);
1244 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1245
1246 *bp = *zio->io_bp_override;
1247 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1248
1249 if (BP_IS_EMBEDDED(bp))
1250 return (ZIO_PIPELINE_CONTINUE);
1251
1252 /*
1253 * If we've been overridden and nopwrite is set then
1254 * set the flag accordingly to indicate that a nopwrite
1255 * has already occurred.
1256 */
1257 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1258 ASSERT(!zp->zp_dedup);
1259 ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1260 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1261 return (ZIO_PIPELINE_CONTINUE);
1262 }
1263
1264 ASSERT(!zp->zp_nopwrite);
1265
1266 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1267 return (ZIO_PIPELINE_CONTINUE);
1268
1269 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1270 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1271
1272 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1273 BP_SET_DEDUP(bp, 1);
1274 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1275 return (ZIO_PIPELINE_CONTINUE);
1276 }
1277
1278 /*
1279 * We were unable to handle this as an override bp, treat
1280 * it as a regular write I/O.
1281 */
1282 zio->io_bp_override = NULL;
1283 *bp = zio->io_bp_orig;
1284 zio->io_pipeline = zio->io_orig_pipeline;
1285 }
1286
1287 return (ZIO_PIPELINE_CONTINUE);
1288 }
1289
1290 static int
1291 zio_write_compress(zio_t *zio)
1292 {
1293 spa_t *spa = zio->io_spa;
1294 zio_prop_t *zp = &zio->io_prop;
1295 enum zio_compress compress = zp->zp_compress;
1296 blkptr_t *bp = zio->io_bp;
1297 uint64_t lsize = zio->io_lsize;
1298 uint64_t psize = zio->io_size;
1299 int pass = 1;
1300
1301 EQUIV(lsize != psize, (zio->io_flags & ZIO_FLAG_RAW) != 0);
1302
1303 /*
1304 * If our children haven't all reached the ready stage,
1305 * wait for them and then repeat this pipeline stage.
1306 */
1307 if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT |
1308 ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) {
1309 return (ZIO_PIPELINE_STOP);
1310 }
1311
1312 if (!IO_IS_ALLOCATING(zio))
1313 return (ZIO_PIPELINE_CONTINUE);
1314
1315 if (zio->io_children_ready != NULL) {
1316 /*
1317 * Now that all our children are ready, run the callback
1318 * associated with this zio in case it wants to modify the
1319 * data to be written.
1320 */
1321 ASSERT3U(zp->zp_level, >, 0);
1322 zio->io_children_ready(zio);
1323 }
1324
1325 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1326 ASSERT(zio->io_bp_override == NULL);
1327
1328 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1329 /*
1330 * We're rewriting an existing block, which means we're
1331 * working on behalf of spa_sync(). For spa_sync() to
1332 * converge, it must eventually be the case that we don't
1333 * have to allocate new blocks. But compression changes
1334 * the blocksize, which forces a reallocate, and makes
1335 * convergence take longer. Therefore, after the first
1336 * few passes, stop compressing to ensure convergence.
1337 */
1338 pass = spa_sync_pass(spa);
1339
1340 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1341 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1342 ASSERT(!BP_GET_DEDUP(bp));
1343
1344 if (pass >= zfs_sync_pass_dont_compress)
1345 compress = ZIO_COMPRESS_OFF;
1346
1347 /* Make sure someone doesn't change their mind on overwrites */
1348 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1349 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1350 }
1351
1352 /* If it's a compressed write that is not raw, compress the buffer. */
1353 if (compress != ZIO_COMPRESS_OFF && psize == lsize) {
1354 void *cbuf = zio_buf_alloc(lsize);
1355 psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize);
1356 if (psize == 0 || psize == lsize) {
1357 compress = ZIO_COMPRESS_OFF;
1358 zio_buf_free(cbuf, lsize);
1359 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1360 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1361 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1362 encode_embedded_bp_compressed(bp,
1363 cbuf, compress, lsize, psize);
1364 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1365 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1366 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1367 zio_buf_free(cbuf, lsize);
1368 bp->blk_birth = zio->io_txg;
1369 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1370 ASSERT(spa_feature_is_active(spa,
1371 SPA_FEATURE_EMBEDDED_DATA));
1372 return (ZIO_PIPELINE_CONTINUE);
1373 } else {
1374 /*
1375 * Round up compressed size up to the ashift
1376 * of the smallest-ashift device, and zero the tail.
1377 * This ensures that the compressed size of the BP
1378 * (and thus compressratio property) are correct,
1379 * in that we charge for the padding used to fill out
1380 * the last sector.
1381 */
1382 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1383 size_t rounded = (size_t)P2ROUNDUP(psize,
1384 1ULL << spa->spa_min_ashift);
1385 if (rounded >= lsize) {
1386 compress = ZIO_COMPRESS_OFF;
1387 zio_buf_free(cbuf, lsize);
1388 psize = lsize;
1389 } else {
1390 abd_t *cdata = abd_get_from_buf(cbuf, lsize);
1391 abd_take_ownership_of_buf(cdata, B_TRUE);
1392 abd_zero_off(cdata, psize, rounded - psize);
1393 psize = rounded;
1394 zio_push_transform(zio, cdata,
1395 psize, lsize, NULL);
1396 }
1397 }
1398
1399 /*
1400 * We were unable to handle this as an override bp, treat
1401 * it as a regular write I/O.
1402 */
1403 zio->io_bp_override = NULL;
1404 *bp = zio->io_bp_orig;
1405 zio->io_pipeline = zio->io_orig_pipeline;
1406 } else {
1407 ASSERT3U(psize, !=, 0);
1408 }
1409
1410 /*
1411 * The final pass of spa_sync() must be all rewrites, but the first
1412 * few passes offer a trade-off: allocating blocks defers convergence,
1413 * but newly allocated blocks are sequential, so they can be written
1414 * to disk faster. Therefore, we allow the first few passes of
1415 * spa_sync() to allocate new blocks, but force rewrites after that.
1416 * There should only be a handful of blocks after pass 1 in any case.
1417 */
1418 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1419 BP_GET_PSIZE(bp) == psize &&
1420 pass >= zfs_sync_pass_rewrite) {
1421 ASSERT(psize != 0);
1422 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1423 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1424 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1425 } else {
1426 BP_ZERO(bp);
1427 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1428 }
1429
1430 if (psize == 0) {
1431 if (zio->io_bp_orig.blk_birth != 0 &&
1432 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1433 BP_SET_LSIZE(bp, lsize);
1434 BP_SET_TYPE(bp, zp->zp_type);
1435 BP_SET_LEVEL(bp, zp->zp_level);
1436 BP_SET_BIRTH(bp, zio->io_txg, 0);
1437 }
1438 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1439 } else {
1440 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1441 BP_SET_LSIZE(bp, lsize);
1442 BP_SET_TYPE(bp, zp->zp_type);
1443 BP_SET_LEVEL(bp, zp->zp_level);
1444 BP_SET_PSIZE(bp, psize);
1445 BP_SET_COMPRESS(bp, compress);
1446 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1447 BP_SET_DEDUP(bp, zp->zp_dedup);
1448 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1449 if (zp->zp_dedup) {
1450 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1451 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1452 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1453 }
1454 if (zp->zp_nopwrite) {
1455 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1456 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1457 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1458 }
1459 }
1460 return (ZIO_PIPELINE_CONTINUE);
1461 }
1462
1463 static int
1464 zio_free_bp_init(zio_t *zio)
1465 {
1466 blkptr_t *bp = zio->io_bp;
1467
1468 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1469 if (BP_GET_DEDUP(bp))
1470 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1471 }
1472
1473 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1474
1475 return (ZIO_PIPELINE_CONTINUE);
1476 }
1477
1478 /*
1479 * ==========================================================================
1480 * Execute the I/O pipeline
1481 * ==========================================================================
1482 */
1483
1484 static void
1485 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1486 {
1487 spa_t *spa = zio->io_spa;
1488 zio_type_t t = zio->io_type;
1489 int flags = (cutinline ? TQ_FRONT : 0);
1490
1491 /*
1492 * If we're a config writer or a probe, the normal issue and
1493 * interrupt threads may all be blocked waiting for the config lock.
1494 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1495 */
1496 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1497 t = ZIO_TYPE_NULL;
1498
1499 /*
1500 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1501 */
1502 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1503 t = ZIO_TYPE_NULL;
1504
1505 /*
1506 * If this is a high priority I/O, then use the high priority taskq if
1507 * available.
1508 */
1509 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1510 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1511 q++;
1512
1513 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1514
1515 /*
1516 * NB: We are assuming that the zio can only be dispatched
1517 * to a single taskq at a time. It would be a grievous error
1518 * to dispatch the zio to another taskq at the same time.
1519 */
1520 ASSERT(zio->io_tqent.tqent_next == NULL);
1521 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1522 flags, &zio->io_tqent);
1523 }
1524
1525 static boolean_t
1526 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1527 {
1528 kthread_t *executor = zio->io_executor;
1529 spa_t *spa = zio->io_spa;
1530
1531 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1532 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1533 uint_t i;
1534 for (i = 0; i < tqs->stqs_count; i++) {
1535 if (taskq_member(tqs->stqs_taskq[i], executor))
1536 return (B_TRUE);
1537 }
1538 }
1539
1540 return (B_FALSE);
1541 }
1542
1543 static int
1544 zio_issue_async(zio_t *zio)
1545 {
1546 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1547
1548 return (ZIO_PIPELINE_STOP);
1549 }
1550
1551 void
1552 zio_interrupt(zio_t *zio)
1553 {
1554 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1555 }
1556
1557 void
1558 zio_delay_interrupt(zio_t *zio)
1559 {
1560 /*
1561 * The timeout_generic() function isn't defined in userspace, so
1562 * rather than trying to implement the function, the zio delay
1563 * functionality has been disabled for userspace builds.
1564 */
1565
1566 #ifdef _KERNEL
1567 /*
1568 * If io_target_timestamp is zero, then no delay has been registered
1569 * for this IO, thus jump to the end of this function and "skip" the
1570 * delay; issuing it directly to the zio layer.
1571 */
1572 if (zio->io_target_timestamp != 0) {
1573 hrtime_t now = gethrtime();
1574
1575 if (now >= zio->io_target_timestamp) {
1576 /*
1577 * This IO has already taken longer than the target
1578 * delay to complete, so we don't want to delay it
1579 * any longer; we "miss" the delay and issue it
1580 * directly to the zio layer. This is likely due to
1581 * the target latency being set to a value less than
1582 * the underlying hardware can satisfy (e.g. delay
1583 * set to 1ms, but the disks take 10ms to complete an
1584 * IO request).
1585 */
1586
1587 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1588 hrtime_t, now);
1589
1590 zio_interrupt(zio);
1591 } else {
1592 hrtime_t diff = zio->io_target_timestamp - now;
1593
1594 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1595 hrtime_t, now, hrtime_t, diff);
1596
1597 (void) timeout_generic(CALLOUT_NORMAL,
1598 (void (*)(void *))zio_interrupt, zio, diff, 1, 0);
1599 }
1600
1601 return;
1602 }
1603 #endif
1604
1605 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
1606 zio_interrupt(zio);
1607 }
1608
1609 /*
1610 * Execute the I/O pipeline until one of the following occurs:
1611 *
1612 * (1) the I/O completes
1613 * (2) the pipeline stalls waiting for dependent child I/Os
1614 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1615 * (4) the I/O is delegated by vdev-level caching or aggregation
1616 * (5) the I/O is deferred due to vdev-level queueing
1617 * (6) the I/O is handed off to another thread.
1618 *
1619 * In all cases, the pipeline stops whenever there's no CPU work; it never
1620 * burns a thread in cv_wait().
1621 *
1622 * There's no locking on io_stage because there's no legitimate way
1623 * for multiple threads to be attempting to process the same I/O.
1624 */
1625 static zio_pipe_stage_t *zio_pipeline[];
1626
1627 void
1628 zio_execute(zio_t *zio)
1629 {
1630 zio->io_executor = curthread;
1631
1632 ASSERT3U(zio->io_queued_timestamp, >, 0);
1633
1634 while (zio->io_stage < ZIO_STAGE_DONE) {
1635 enum zio_stage pipeline = zio->io_pipeline;
1636 enum zio_stage stage = zio->io_stage;
1637 int rv;
1638
1639 ASSERT(!MUTEX_HELD(&zio->io_lock));
1640 ASSERT(ISP2(stage));
1641 ASSERT(zio->io_stall == NULL);
1642
1643 do {
1644 stage <<= 1;
1645 } while ((stage & pipeline) == 0);
1646
1647 ASSERT(stage <= ZIO_STAGE_DONE);
1648
1649 /*
1650 * If we are in interrupt context and this pipeline stage
1651 * will grab a config lock that is held across I/O,
1652 * or may wait for an I/O that needs an interrupt thread
1653 * to complete, issue async to avoid deadlock.
1654 *
1655 * For VDEV_IO_START, we cut in line so that the io will
1656 * be sent to disk promptly.
1657 */
1658 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1659 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1660 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1661 zio_requeue_io_start_cut_in_line : B_FALSE;
1662 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1663 return;
1664 }
1665
1666 zio->io_stage = stage;
1667 zio->io_pipeline_trace |= zio->io_stage;
1668 rv = zio_pipeline[highbit64(stage) - 1](zio);
1669
1670 if (rv == ZIO_PIPELINE_STOP)
1671 return;
1672
1673 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1674 }
1675 }
1676
1677 /*
1678 * ==========================================================================
1679 * Initiate I/O, either sync or async
1680 * ==========================================================================
1681 */
1682 int
1683 zio_wait(zio_t *zio)
1684 {
1685 int error;
1686
1687 ASSERT3P(zio->io_stage, ==, ZIO_STAGE_OPEN);
1688 ASSERT3P(zio->io_executor, ==, NULL);
1689
1690 zio->io_waiter = curthread;
1691 ASSERT0(zio->io_queued_timestamp);
1692 zio->io_queued_timestamp = gethrtime();
1693
1694 zio_execute(zio);
1695
1696 mutex_enter(&zio->io_lock);
1697 while (zio->io_executor != NULL)
1698 cv_wait(&zio->io_cv, &zio->io_lock);
1699 mutex_exit(&zio->io_lock);
1700
1701 error = zio->io_error;
1702 zio_destroy(zio);
1703
1704 return (error);
1705 }
1706
1707 void
1708 zio_nowait(zio_t *zio)
1709 {
1710 ASSERT3P(zio->io_executor, ==, NULL);
1711
1712 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1713 zio_unique_parent(zio) == NULL) {
1714 /*
1715 * This is a logical async I/O with no parent to wait for it.
1716 * We add it to the spa_async_root_zio "Godfather" I/O which
1717 * will ensure they complete prior to unloading the pool.
1718 */
1719 spa_t *spa = zio->io_spa;
1720
1721 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1722 }
1723
1724 ASSERT0(zio->io_queued_timestamp);
1725 zio->io_queued_timestamp = gethrtime();
1726 zio_execute(zio);
1727 }
1728
1729 /*
1730 * ==========================================================================
1731 * Reexecute, cancel, or suspend/resume failed I/O
1732 * ==========================================================================
1733 */
1734
1735 static void
1736 zio_reexecute(zio_t *pio)
1737 {
1738 zio_t *cio, *cio_next;
1739
1740 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1741 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1742 ASSERT(pio->io_gang_leader == NULL);
1743 ASSERT(pio->io_gang_tree == NULL);
1744
1745 pio->io_flags = pio->io_orig_flags;
1746 pio->io_stage = pio->io_orig_stage;
1747 pio->io_pipeline = pio->io_orig_pipeline;
1748 pio->io_reexecute = 0;
1749 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1750 pio->io_pipeline_trace = 0;
1751 pio->io_error = 0;
1752 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1753 pio->io_state[w] = 0;
1754 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1755 pio->io_child_error[c] = 0;
1756
1757 if (IO_IS_ALLOCATING(pio))
1758 BP_ZERO(pio->io_bp);
1759
1760 /*
1761 * As we reexecute pio's children, new children could be created.
1762 * New children go to the head of pio's io_child_list, however,
1763 * so we will (correctly) not reexecute them. The key is that
1764 * the remainder of pio's io_child_list, from 'cio_next' onward,
1765 * cannot be affected by any side effects of reexecuting 'cio'.
1766 */
1767 zio_link_t *zl = NULL;
1768 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
1769 cio_next = zio_walk_children(pio, &zl);
1770 mutex_enter(&pio->io_lock);
1771 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1772 pio->io_children[cio->io_child_type][w]++;
1773 mutex_exit(&pio->io_lock);
1774 zio_reexecute(cio);
1775 }
1776
1777 /*
1778 * Now that all children have been reexecuted, execute the parent.
1779 * We don't reexecute "The Godfather" I/O here as it's the
1780 * responsibility of the caller to wait on it.
1781 */
1782 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
1783 pio->io_queued_timestamp = gethrtime();
1784 zio_execute(pio);
1785 }
1786 }
1787
1788 void
1789 zio_suspend(spa_t *spa, zio_t *zio)
1790 {
1791 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1792 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1793 "failure and the failure mode property for this pool "
1794 "is set to panic.", spa_name(spa));
1795
1796 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1797
1798 mutex_enter(&spa->spa_suspend_lock);
1799
1800 if (spa->spa_suspend_zio_root == NULL)
1801 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1802 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1803 ZIO_FLAG_GODFATHER);
1804
1805 spa->spa_suspended = B_TRUE;
1806
1807 if (zio != NULL) {
1808 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1809 ASSERT(zio != spa->spa_suspend_zio_root);
1810 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1811 ASSERT(zio_unique_parent(zio) == NULL);
1812 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1813 zio_add_child(spa->spa_suspend_zio_root, zio);
1814 }
1815
1816 mutex_exit(&spa->spa_suspend_lock);
1817 }
1818
1819 int
1820 zio_resume(spa_t *spa)
1821 {
1822 zio_t *pio;
1823
1824 /*
1825 * Reexecute all previously suspended i/o.
1826 */
1827 mutex_enter(&spa->spa_suspend_lock);
1828 spa->spa_suspended = B_FALSE;
1829 cv_broadcast(&spa->spa_suspend_cv);
1830 pio = spa->spa_suspend_zio_root;
1831 spa->spa_suspend_zio_root = NULL;
1832 mutex_exit(&spa->spa_suspend_lock);
1833
1834 if (pio == NULL)
1835 return (0);
1836
1837 zio_reexecute(pio);
1838 return (zio_wait(pio));
1839 }
1840
1841 void
1842 zio_resume_wait(spa_t *spa)
1843 {
1844 mutex_enter(&spa->spa_suspend_lock);
1845 while (spa_suspended(spa))
1846 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1847 mutex_exit(&spa->spa_suspend_lock);
1848 }
1849
1850 /*
1851 * ==========================================================================
1852 * Gang blocks.
1853 *
1854 * A gang block is a collection of small blocks that looks to the DMU
1855 * like one large block. When zio_dva_allocate() cannot find a block
1856 * of the requested size, due to either severe fragmentation or the pool
1857 * being nearly full, it calls zio_write_gang_block() to construct the
1858 * block from smaller fragments.
1859 *
1860 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1861 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1862 * an indirect block: it's an array of block pointers. It consumes
1863 * only one sector and hence is allocatable regardless of fragmentation.
1864 * The gang header's bps point to its gang members, which hold the data.
1865 *
1866 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1867 * as the verifier to ensure uniqueness of the SHA256 checksum.
1868 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1869 * not the gang header. This ensures that data block signatures (needed for
1870 * deduplication) are independent of how the block is physically stored.
1871 *
1872 * Gang blocks can be nested: a gang member may itself be a gang block.
1873 * Thus every gang block is a tree in which root and all interior nodes are
1874 * gang headers, and the leaves are normal blocks that contain user data.
1875 * The root of the gang tree is called the gang leader.
1876 *
1877 * To perform any operation (read, rewrite, free, claim) on a gang block,
1878 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1879 * in the io_gang_tree field of the original logical i/o by recursively
1880 * reading the gang leader and all gang headers below it. This yields
1881 * an in-core tree containing the contents of every gang header and the
1882 * bps for every constituent of the gang block.
1883 *
1884 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1885 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1886 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1887 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1888 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1889 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1890 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1891 * of the gang header plus zio_checksum_compute() of the data to update the
1892 * gang header's blk_cksum as described above.
1893 *
1894 * The two-phase assemble/issue model solves the problem of partial failure --
1895 * what if you'd freed part of a gang block but then couldn't read the
1896 * gang header for another part? Assembling the entire gang tree first
1897 * ensures that all the necessary gang header I/O has succeeded before
1898 * starting the actual work of free, claim, or write. Once the gang tree
1899 * is assembled, free and claim are in-memory operations that cannot fail.
1900 *
1901 * In the event that a gang write fails, zio_dva_unallocate() walks the
1902 * gang tree to immediately free (i.e. insert back into the space map)
1903 * everything we've allocated. This ensures that we don't get ENOSPC
1904 * errors during repeated suspend/resume cycles due to a flaky device.
1905 *
1906 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1907 * the gang tree, we won't modify the block, so we can safely defer the free
1908 * (knowing that the block is still intact). If we *can* assemble the gang
1909 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1910 * each constituent bp and we can allocate a new block on the next sync pass.
1911 *
1912 * In all cases, the gang tree allows complete recovery from partial failure.
1913 * ==========================================================================
1914 */
1915
1916 static void
1917 zio_gang_issue_func_done(zio_t *zio)
1918 {
1919 abd_put(zio->io_abd);
1920 }
1921
1922 static zio_t *
1923 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
1924 uint64_t offset)
1925 {
1926 if (gn != NULL)
1927 return (pio);
1928
1929 return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset),
1930 BP_GET_PSIZE(bp), zio_gang_issue_func_done,
1931 NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1932 &pio->io_bookmark));
1933 }
1934
1935 static zio_t *
1936 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
1937 uint64_t offset)
1938 {
1939 zio_t *zio;
1940
1941 if (gn != NULL) {
1942 abd_t *gbh_abd =
1943 abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1944 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1945 gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL,
1946 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1947 &pio->io_bookmark);
1948 /*
1949 * As we rewrite each gang header, the pipeline will compute
1950 * a new gang block header checksum for it; but no one will
1951 * compute a new data checksum, so we do that here. The one
1952 * exception is the gang leader: the pipeline already computed
1953 * its data checksum because that stage precedes gang assembly.
1954 * (Presently, nothing actually uses interior data checksums;
1955 * this is just good hygiene.)
1956 */
1957 if (gn != pio->io_gang_leader->io_gang_tree) {
1958 abd_t *buf = abd_get_offset(data, offset);
1959
1960 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1961 buf, BP_GET_PSIZE(bp));
1962
1963 abd_put(buf);
1964 }
1965 /*
1966 * If we are here to damage data for testing purposes,
1967 * leave the GBH alone so that we can detect the damage.
1968 */
1969 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1970 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1971 } else {
1972 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1973 abd_get_offset(data, offset), BP_GET_PSIZE(bp),
1974 zio_gang_issue_func_done, NULL, pio->io_priority,
1975 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1976 }
1977
1978 return (zio);
1979 }
1980
1981 /* ARGSUSED */
1982 static zio_t *
1983 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
1984 uint64_t offset)
1985 {
1986 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1987 ZIO_GANG_CHILD_FLAGS(pio)));
1988 }
1989
1990 /* ARGSUSED */
1991 static zio_t *
1992 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
1993 uint64_t offset)
1994 {
1995 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1996 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1997 }
1998
1999 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2000 NULL,
2001 zio_read_gang,
2002 zio_rewrite_gang,
2003 zio_free_gang,
2004 zio_claim_gang,
2005 NULL
2006 };
2007
2008 static void zio_gang_tree_assemble_done(zio_t *zio);
2009
2010 static zio_gang_node_t *
2011 zio_gang_node_alloc(zio_gang_node_t **gnpp)
2012 {
2013 zio_gang_node_t *gn;
2014
2015 ASSERT(*gnpp == NULL);
2016
2017 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2018 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2019 *gnpp = gn;
2020
2021 return (gn);
2022 }
2023
2024 static void
2025 zio_gang_node_free(zio_gang_node_t **gnpp)
2026 {
2027 zio_gang_node_t *gn = *gnpp;
2028
2029 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2030 ASSERT(gn->gn_child[g] == NULL);
2031
2032 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2033 kmem_free(gn, sizeof (*gn));
2034 *gnpp = NULL;
2035 }
2036
2037 static void
2038 zio_gang_tree_free(zio_gang_node_t **gnpp)
2039 {
2040 zio_gang_node_t *gn = *gnpp;
2041
2042 if (gn == NULL)
2043 return;
2044
2045 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2046 zio_gang_tree_free(&gn->gn_child[g]);
2047
2048 zio_gang_node_free(gnpp);
2049 }
2050
2051 static void
2052 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2053 {
2054 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2055 abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2056
2057 ASSERT(gio->io_gang_leader == gio);
2058 ASSERT(BP_IS_GANG(bp));
2059
2060 zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2061 zio_gang_tree_assemble_done, gn, gio->io_priority,
2062 ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2063 }
2064
2065 static void
2066 zio_gang_tree_assemble_done(zio_t *zio)
2067 {
2068 zio_t *gio = zio->io_gang_leader;
2069 zio_gang_node_t *gn = zio->io_private;
2070 blkptr_t *bp = zio->io_bp;
2071
2072 ASSERT(gio == zio_unique_parent(zio));
2073 ASSERT(zio->io_child_count == 0);
2074
2075 if (zio->io_error)
2076 return;
2077
2078 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2079 if (BP_SHOULD_BYTESWAP(bp))
2080 byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size);
2081
2082 ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh);
2083 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2084 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2085
2086 abd_put(zio->io_abd);
2087
2088 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2089 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2090 if (!BP_IS_GANG(gbp))
2091 continue;
2092 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2093 }
2094 }
2095
2096 static void
2097 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data,
2098 uint64_t offset)
2099 {
2100 zio_t *gio = pio->io_gang_leader;
2101 zio_t *zio;
2102
2103 ASSERT(BP_IS_GANG(bp) == !!gn);
2104 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2105 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2106
2107 /*
2108 * If you're a gang header, your data is in gn->gn_gbh.
2109 * If you're a gang member, your data is in 'data' and gn == NULL.
2110 */
2111 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset);
2112
2113 if (gn != NULL) {
2114 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2115
2116 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2117 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2118 if (BP_IS_HOLE(gbp))
2119 continue;
2120 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data,
2121 offset);
2122 offset += BP_GET_PSIZE(gbp);
2123 }
2124 }
2125
2126 if (gn == gio->io_gang_tree)
2127 ASSERT3U(gio->io_size, ==, offset);
2128
2129 if (zio != pio)
2130 zio_nowait(zio);
2131 }
2132
2133 static int
2134 zio_gang_assemble(zio_t *zio)
2135 {
2136 blkptr_t *bp = zio->io_bp;
2137
2138 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2139 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2140
2141 zio->io_gang_leader = zio;
2142
2143 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2144
2145 return (ZIO_PIPELINE_CONTINUE);
2146 }
2147
2148 static int
2149 zio_gang_issue(zio_t *zio)
2150 {
2151 blkptr_t *bp = zio->io_bp;
2152
2153 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) {
2154 return (ZIO_PIPELINE_STOP);
2155 }
2156
2157 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2158 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2159
2160 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2161 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd,
2162 0);
2163 else
2164 zio_gang_tree_free(&zio->io_gang_tree);
2165
2166 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2167
2168 return (ZIO_PIPELINE_CONTINUE);
2169 }
2170
2171 static void
2172 zio_write_gang_member_ready(zio_t *zio)
2173 {
2174 zio_t *pio = zio_unique_parent(zio);
2175 zio_t *gio = zio->io_gang_leader;
2176 dva_t *cdva = zio->io_bp->blk_dva;
2177 dva_t *pdva = pio->io_bp->blk_dva;
2178 uint64_t asize;
2179
2180 if (BP_IS_HOLE(zio->io_bp))
2181 return;
2182
2183 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2184
2185 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2186 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2187 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2188 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2189 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2190
2191 mutex_enter(&pio->io_lock);
2192 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2193 ASSERT(DVA_GET_GANG(&pdva[d]));
2194 asize = DVA_GET_ASIZE(&pdva[d]);
2195 asize += DVA_GET_ASIZE(&cdva[d]);
2196 DVA_SET_ASIZE(&pdva[d], asize);
2197 }
2198 mutex_exit(&pio->io_lock);
2199 }
2200
2201 static void
2202 zio_write_gang_done(zio_t *zio)
2203 {
2204 abd_put(zio->io_abd);
2205 }
2206
2207 static int
2208 zio_write_gang_block(zio_t *pio)
2209 {
2210 spa_t *spa = pio->io_spa;
2211 metaslab_class_t *mc = spa_normal_class(spa);
2212 blkptr_t *bp = pio->io_bp;
2213 zio_t *gio = pio->io_gang_leader;
2214 zio_t *zio;
2215 zio_gang_node_t *gn, **gnpp;
2216 zio_gbh_phys_t *gbh;
2217 abd_t *gbh_abd;
2218 uint64_t txg = pio->io_txg;
2219 uint64_t resid = pio->io_size;
2220 uint64_t lsize;
2221 int copies = gio->io_prop.zp_copies;
2222 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2223 zio_prop_t zp;
2224 int error;
2225
2226 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2227 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2228 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2229 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2230
2231 flags |= METASLAB_ASYNC_ALLOC;
2232 VERIFY(refcount_held(&mc->mc_alloc_slots, pio));
2233
2234 /*
2235 * The logical zio has already placed a reservation for
2236 * 'copies' allocation slots but gang blocks may require
2237 * additional copies. These additional copies
2238 * (i.e. gbh_copies - copies) are guaranteed to succeed
2239 * since metaslab_class_throttle_reserve() always allows
2240 * additional reservations for gang blocks.
2241 */
2242 VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
2243 pio, flags));
2244 }
2245
2246 error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
2247 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags,
2248 &pio->io_alloc_list, pio);
2249 if (error) {
2250 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2251 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2252 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2253
2254 /*
2255 * If we failed to allocate the gang block header then
2256 * we remove any additional allocation reservations that
2257 * we placed here. The original reservation will
2258 * be removed when the logical I/O goes to the ready
2259 * stage.
2260 */
2261 metaslab_class_throttle_unreserve(mc,
2262 gbh_copies - copies, pio);
2263 }
2264 pio->io_error = error;
2265 return (ZIO_PIPELINE_CONTINUE);
2266 }
2267
2268 if (pio == gio) {
2269 gnpp = &gio->io_gang_tree;
2270 } else {
2271 gnpp = pio->io_private;
2272 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2273 }
2274
2275 gn = zio_gang_node_alloc(gnpp);
2276 gbh = gn->gn_gbh;
2277 bzero(gbh, SPA_GANGBLOCKSIZE);
2278 gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE);
2279
2280 /*
2281 * Create the gang header.
2282 */
2283 zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2284 zio_write_gang_done, NULL, pio->io_priority,
2285 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2286
2287 /*
2288 * Create and nowait the gang children.
2289 */
2290 for (int g = 0; resid != 0; resid -= lsize, g++) {
2291 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2292 SPA_MINBLOCKSIZE);
2293 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2294
2295 zp.zp_checksum = gio->io_prop.zp_checksum;
2296 zp.zp_compress = ZIO_COMPRESS_OFF;
2297 zp.zp_type = DMU_OT_NONE;
2298 zp.zp_level = 0;
2299 zp.zp_copies = gio->io_prop.zp_copies;
2300 zp.zp_dedup = B_FALSE;
2301 zp.zp_dedup_verify = B_FALSE;
2302 zp.zp_nopwrite = B_FALSE;
2303
2304 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2305 abd_get_offset(pio->io_abd, pio->io_size - resid), lsize,
2306 lsize, &zp, zio_write_gang_member_ready, NULL, NULL,
2307 zio_write_gang_done, &gn->gn_child[g], pio->io_priority,
2308 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2309
2310 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2311 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2312 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2313
2314 /*
2315 * Gang children won't throttle but we should
2316 * account for their work, so reserve an allocation
2317 * slot for them here.
2318 */
2319 VERIFY(metaslab_class_throttle_reserve(mc,
2320 zp.zp_copies, cio, flags));
2321 }
2322 zio_nowait(cio);
2323 }
2324
2325 /*
2326 * Set pio's pipeline to just wait for zio to finish.
2327 */
2328 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2329
2330 zio_nowait(zio);
2331
2332 return (ZIO_PIPELINE_CONTINUE);
2333 }
2334
2335 /*
2336 * The zio_nop_write stage in the pipeline determines if allocating a
2337 * new bp is necessary. The nopwrite feature can handle writes in
2338 * either syncing or open context (i.e. zil writes) and as a result is
2339 * mutually exclusive with dedup.
2340 *
2341 * By leveraging a cryptographically secure checksum, such as SHA256, we
2342 * can compare the checksums of the new data and the old to determine if
2343 * allocating a new block is required. Note that our requirements for
2344 * cryptographic strength are fairly weak: there can't be any accidental
2345 * hash collisions, but we don't need to be secure against intentional
2346 * (malicious) collisions. To trigger a nopwrite, you have to be able
2347 * to write the file to begin with, and triggering an incorrect (hash
2348 * collision) nopwrite is no worse than simply writing to the file.
2349 * That said, there are no known attacks against the checksum algorithms
2350 * used for nopwrite, assuming that the salt and the checksums
2351 * themselves remain secret.
2352 */
2353 static int
2354 zio_nop_write(zio_t *zio)
2355 {
2356 blkptr_t *bp = zio->io_bp;
2357 blkptr_t *bp_orig = &zio->io_bp_orig;
2358 zio_prop_t *zp = &zio->io_prop;
2359
2360 ASSERT(BP_GET_LEVEL(bp) == 0);
2361 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2362 ASSERT(zp->zp_nopwrite);
2363 ASSERT(!zp->zp_dedup);
2364 ASSERT(zio->io_bp_override == NULL);
2365 ASSERT(IO_IS_ALLOCATING(zio));
2366
2367 /*
2368 * Check to see if the original bp and the new bp have matching
2369 * characteristics (i.e. same checksum, compression algorithms, etc).
2370 * If they don't then just continue with the pipeline which will
2371 * allocate a new bp.
2372 */
2373 if (BP_IS_HOLE(bp_orig) ||
2374 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2375 ZCHECKSUM_FLAG_NOPWRITE) ||
2376 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2377 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2378 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2379 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2380 return (ZIO_PIPELINE_CONTINUE);
2381
2382 /*
2383 * If the checksums match then reset the pipeline so that we
2384 * avoid allocating a new bp and issuing any I/O.
2385 */
2386 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2387 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2388 ZCHECKSUM_FLAG_NOPWRITE);
2389 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2390 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2391 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2392 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2393 sizeof (uint64_t)) == 0);
2394
2395 *bp = *bp_orig;
2396 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2397 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2398 }
2399
2400 return (ZIO_PIPELINE_CONTINUE);
2401 }
2402
2403 /*
2404 * ==========================================================================
2405 * Dedup
2406 * ==========================================================================
2407 */
2408 static void
2409 zio_ddt_child_read_done(zio_t *zio)
2410 {
2411 blkptr_t *bp = zio->io_bp;
2412 ddt_entry_t *dde = zio->io_private;
2413 ddt_phys_t *ddp;
2414 zio_t *pio = zio_unique_parent(zio);
2415
2416 mutex_enter(&pio->io_lock);
2417 ddp = ddt_phys_select(dde, bp);
2418 if (zio->io_error == 0)
2419 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2420
2421 if (zio->io_error == 0 && dde->dde_repair_abd == NULL)
2422 dde->dde_repair_abd = zio->io_abd;
2423 else
2424 abd_free(zio->io_abd);
2425 mutex_exit(&pio->io_lock);
2426 }
2427
2428 static int
2429 zio_ddt_read_start(zio_t *zio)
2430 {
2431 blkptr_t *bp = zio->io_bp;
2432
2433 ASSERT(BP_GET_DEDUP(bp));
2434 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2435 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2436
2437 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2438 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2439 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2440 ddt_phys_t *ddp = dde->dde_phys;
2441 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2442 blkptr_t blk;
2443
2444 ASSERT(zio->io_vsd == NULL);
2445 zio->io_vsd = dde;
2446
2447 if (ddp_self == NULL)
2448 return (ZIO_PIPELINE_CONTINUE);
2449
2450 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2451 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2452 continue;
2453 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2454 &blk);
2455 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2456 abd_alloc_for_io(zio->io_size, B_TRUE),
2457 zio->io_size, zio_ddt_child_read_done, dde,
2458 zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) |
2459 ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark));
2460 }
2461 return (ZIO_PIPELINE_CONTINUE);
2462 }
2463
2464 zio_nowait(zio_read(zio, zio->io_spa, bp,
2465 zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority,
2466 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2467
2468 return (ZIO_PIPELINE_CONTINUE);
2469 }
2470
2471 static int
2472 zio_ddt_read_done(zio_t *zio)
2473 {
2474 blkptr_t *bp = zio->io_bp;
2475
2476 if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) {
2477 return (ZIO_PIPELINE_STOP);
2478 }
2479
2480 ASSERT(BP_GET_DEDUP(bp));
2481 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2482 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2483
2484 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2485 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2486 ddt_entry_t *dde = zio->io_vsd;
2487 if (ddt == NULL) {
2488 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2489 return (ZIO_PIPELINE_CONTINUE);
2490 }
2491 if (dde == NULL) {
2492 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2493 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2494 return (ZIO_PIPELINE_STOP);
2495 }
2496 if (dde->dde_repair_abd != NULL) {
2497 abd_copy(zio->io_abd, dde->dde_repair_abd,
2498 zio->io_size);
2499 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2500 }
2501 ddt_repair_done(ddt, dde);
2502 zio->io_vsd = NULL;
2503 }
2504
2505 ASSERT(zio->io_vsd == NULL);
2506
2507 return (ZIO_PIPELINE_CONTINUE);
2508 }
2509
2510 static boolean_t
2511 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2512 {
2513 spa_t *spa = zio->io_spa;
2514 boolean_t do_raw = (zio->io_flags & ZIO_FLAG_RAW);
2515
2516 /* We should never get a raw, override zio */
2517 ASSERT(!(zio->io_bp_override && do_raw));
2518
2519 /*
2520 * Note: we compare the original data, not the transformed data,
2521 * because when zio->io_bp is an override bp, we will not have
2522 * pushed the I/O transforms. That's an important optimization
2523 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2524 */
2525 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2526 zio_t *lio = dde->dde_lead_zio[p];
2527
2528 if (lio != NULL) {
2529 return (lio->io_orig_size != zio->io_orig_size ||
2530 abd_cmp(zio->io_orig_abd, lio->io_orig_abd,
2531 zio->io_orig_size) != 0);
2532 }
2533 }
2534
2535 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2536 ddt_phys_t *ddp = &dde->dde_phys[p];
2537
2538 if (ddp->ddp_phys_birth != 0) {
2539 arc_buf_t *abuf = NULL;
2540 arc_flags_t aflags = ARC_FLAG_WAIT;
2541 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
2542 blkptr_t blk = *zio->io_bp;
2543 int error;
2544
2545 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2546
2547 ddt_exit(ddt);
2548
2549 /*
2550 * Intuitively, it would make more sense to compare
2551 * io_abd than io_orig_abd in the raw case since you
2552 * don't want to look at any transformations that have
2553 * happened to the data. However, for raw I/Os the
2554 * data will actually be the same in io_abd and
2555 * io_orig_abd, so all we have to do is issue this as
2556 * a raw ARC read.
2557 */
2558 if (do_raw) {
2559 zio_flags |= ZIO_FLAG_RAW;
2560 ASSERT3U(zio->io_size, ==, zio->io_orig_size);
2561 ASSERT0(abd_cmp(zio->io_abd, zio->io_orig_abd,
2562 zio->io_size));
2563 ASSERT3P(zio->io_transform_stack, ==, NULL);
2564 }
2565
2566 error = arc_read(NULL, spa, &blk,
2567 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2568 zio_flags, &aflags, &zio->io_bookmark);
2569
2570 if (error == 0) {
2571 if (arc_buf_size(abuf) != zio->io_orig_size ||
2572 abd_cmp_buf(zio->io_orig_abd, abuf->b_data,
2573 zio->io_orig_size) != 0)
2574 error = SET_ERROR(EEXIST);
2575 arc_buf_destroy(abuf, &abuf);
2576 }
2577
2578 ddt_enter(ddt);
2579 return (error != 0);
2580 }
2581 }
2582
2583 return (B_FALSE);
2584 }
2585
2586 static void
2587 zio_ddt_child_write_ready(zio_t *zio)
2588 {
2589 int p = zio->io_prop.zp_copies;
2590 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2591 ddt_entry_t *dde = zio->io_private;
2592 ddt_phys_t *ddp = &dde->dde_phys[p];
2593 zio_t *pio;
2594
2595 if (zio->io_error)
2596 return;
2597
2598 ddt_enter(ddt);
2599
2600 ASSERT(dde->dde_lead_zio[p] == zio);
2601
2602 ddt_phys_fill(ddp, zio->io_bp);
2603
2604 zio_link_t *zl = NULL;
2605 while ((pio = zio_walk_parents(zio, &zl)) != NULL)
2606 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2607
2608 ddt_exit(ddt);
2609 }
2610
2611 static void
2612 zio_ddt_child_write_done(zio_t *zio)
2613 {
2614 int p = zio->io_prop.zp_copies;
2615 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2616 ddt_entry_t *dde = zio->io_private;
2617 ddt_phys_t *ddp = &dde->dde_phys[p];
2618
2619 ddt_enter(ddt);
2620
2621 ASSERT(ddp->ddp_refcnt == 0);
2622 ASSERT(dde->dde_lead_zio[p] == zio);
2623 dde->dde_lead_zio[p] = NULL;
2624
2625 if (zio->io_error == 0) {
2626 zio_link_t *zl = NULL;
2627 while (zio_walk_parents(zio, &zl) != NULL)
2628 ddt_phys_addref(ddp);
2629 } else {
2630 ddt_phys_clear(ddp);
2631 }
2632
2633 ddt_exit(ddt);
2634 }
2635
2636 static void
2637 zio_ddt_ditto_write_done(zio_t *zio)
2638 {
2639 int p = DDT_PHYS_DITTO;
2640 zio_prop_t *zp = &zio->io_prop;
2641 blkptr_t *bp = zio->io_bp;
2642 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2643 ddt_entry_t *dde = zio->io_private;
2644 ddt_phys_t *ddp = &dde->dde_phys[p];
2645 ddt_key_t *ddk = &dde->dde_key;
2646
2647 ddt_enter(ddt);
2648
2649 ASSERT(ddp->ddp_refcnt == 0);
2650 ASSERT(dde->dde_lead_zio[p] == zio);
2651 dde->dde_lead_zio[p] = NULL;
2652
2653 if (zio->io_error == 0) {
2654 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2655 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2656 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2657 if (ddp->ddp_phys_birth != 0)
2658 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2659 ddt_phys_fill(ddp, bp);
2660 }
2661
2662 ddt_exit(ddt);
2663 }
2664
2665 static int
2666 zio_ddt_write(zio_t *zio)
2667 {
2668 spa_t *spa = zio->io_spa;
2669 blkptr_t *bp = zio->io_bp;
2670 uint64_t txg = zio->io_txg;
2671 zio_prop_t *zp = &zio->io_prop;
2672 int p = zp->zp_copies;
2673 int ditto_copies;
2674 zio_t *cio = NULL;
2675 zio_t *dio = NULL;
2676 ddt_t *ddt = ddt_select(spa, bp);
2677 ddt_entry_t *dde;
2678 ddt_phys_t *ddp;
2679
2680 ASSERT(BP_GET_DEDUP(bp));
2681 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2682 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2683 ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
2684
2685 ddt_enter(ddt);
2686 dde = ddt_lookup(ddt, bp, B_TRUE);
2687 ddp = &dde->dde_phys[p];
2688
2689 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2690 /*
2691 * If we're using a weak checksum, upgrade to a strong checksum
2692 * and try again. If we're already using a strong checksum,
2693 * we can't resolve it, so just convert to an ordinary write.
2694 * (And automatically e-mail a paper to Nature?)
2695 */
2696 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
2697 ZCHECKSUM_FLAG_DEDUP)) {
2698 zp->zp_checksum = spa_dedup_checksum(spa);
2699 zio_pop_transforms(zio);
2700 zio->io_stage = ZIO_STAGE_OPEN;
2701 BP_ZERO(bp);
2702 } else {
2703 zp->zp_dedup = B_FALSE;
2704 BP_SET_DEDUP(bp, B_FALSE);
2705 }
2706 ASSERT(!BP_GET_DEDUP(bp));
2707 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2708 ddt_exit(ddt);
2709 return (ZIO_PIPELINE_CONTINUE);
2710 }
2711
2712 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2713 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2714
2715 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2716 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2717 zio_prop_t czp = *zp;
2718
2719 czp.zp_copies = ditto_copies;
2720
2721 /*
2722 * If we arrived here with an override bp, we won't have run
2723 * the transform stack, so we won't have the data we need to
2724 * generate a child i/o. So, toss the override bp and restart.
2725 * This is safe, because using the override bp is just an
2726 * optimization; and it's rare, so the cost doesn't matter.
2727 */
2728 if (zio->io_bp_override) {
2729 zio_pop_transforms(zio);
2730 zio->io_stage = ZIO_STAGE_OPEN;
2731 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2732 zio->io_bp_override = NULL;
2733 BP_ZERO(bp);
2734 ddt_exit(ddt);
2735 return (ZIO_PIPELINE_CONTINUE);
2736 }
2737
2738 dio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2739 zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL,
2740 NULL, zio_ddt_ditto_write_done, dde, zio->io_priority,
2741 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2742
2743 zio_push_transform(dio, zio->io_abd, zio->io_size, 0, NULL);
2744 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2745 }
2746
2747 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2748 if (ddp->ddp_phys_birth != 0)
2749 ddt_bp_fill(ddp, bp, txg);
2750 if (dde->dde_lead_zio[p] != NULL)
2751 zio_add_child(zio, dde->dde_lead_zio[p]);
2752 else
2753 ddt_phys_addref(ddp);
2754 } else if (zio->io_bp_override) {
2755 ASSERT(bp->blk_birth == txg);
2756 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2757 ddt_phys_fill(ddp, bp);
2758 ddt_phys_addref(ddp);
2759 } else {
2760 cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2761 zio->io_orig_size, zio->io_orig_size, zp,
2762 zio_ddt_child_write_ready, NULL, NULL,
2763 zio_ddt_child_write_done, dde, zio->io_priority,
2764 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2765
2766 zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL);
2767 dde->dde_lead_zio[p] = cio;
2768 }
2769
2770 ddt_exit(ddt);
2771
2772 if (cio)
2773 zio_nowait(cio);
2774 if (dio)
2775 zio_nowait(dio);
2776
2777 return (ZIO_PIPELINE_CONTINUE);
2778 }
2779
2780 ddt_entry_t *freedde; /* for debugging */
2781
2782 static int
2783 zio_ddt_free(zio_t *zio)
2784 {
2785 spa_t *spa = zio->io_spa;
2786 blkptr_t *bp = zio->io_bp;
2787 ddt_t *ddt = ddt_select(spa, bp);
2788 ddt_entry_t *dde;
2789 ddt_phys_t *ddp;
2790
2791 ASSERT(BP_GET_DEDUP(bp));
2792 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2793
2794 ddt_enter(ddt);
2795 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2796 ddp = ddt_phys_select(dde, bp);
2797 ddt_phys_decref(ddp);
2798 ddt_exit(ddt);
2799
2800 return (ZIO_PIPELINE_CONTINUE);
2801 }
2802
2803 /*
2804 * ==========================================================================
2805 * Allocate and free blocks
2806 * ==========================================================================
2807 */
2808
2809 static zio_t *
2810 zio_io_to_allocate(spa_t *spa)
2811 {
2812 zio_t *zio;
2813
2814 ASSERT(MUTEX_HELD(&spa->spa_alloc_lock));
2815
2816 zio = avl_first(&spa->spa_alloc_tree);
2817 if (zio == NULL)
2818 return (NULL);
2819
2820 ASSERT(IO_IS_ALLOCATING(zio));
2821
2822 /*
2823 * Try to place a reservation for this zio. If we're unable to
2824 * reserve then we throttle.
2825 */
2826 if (!metaslab_class_throttle_reserve(spa_normal_class(spa),
2827 zio->io_prop.zp_copies, zio, 0)) {
2828 return (NULL);
2829 }
2830
2831 avl_remove(&spa->spa_alloc_tree, zio);
2832 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
2833
2834 return (zio);
2835 }
2836
2837 static int
2838 zio_dva_throttle(zio_t *zio)
2839 {
2840 spa_t *spa = zio->io_spa;
2841 zio_t *nio;
2842
2843 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
2844 !spa_normal_class(zio->io_spa)->mc_alloc_throttle_enabled ||
2845 zio->io_child_type == ZIO_CHILD_GANG ||
2846 zio->io_flags & ZIO_FLAG_NODATA) {
2847 return (ZIO_PIPELINE_CONTINUE);
2848 }
2849
2850 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2851
2852 ASSERT3U(zio->io_queued_timestamp, >, 0);
2853 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
2854
2855 mutex_enter(&spa->spa_alloc_lock);
2856
2857 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2858 avl_add(&spa->spa_alloc_tree, zio);
2859
2860 nio = zio_io_to_allocate(zio->io_spa);
2861 mutex_exit(&spa->spa_alloc_lock);
2862
2863 if (nio == zio)
2864 return (ZIO_PIPELINE_CONTINUE);
2865
2866 if (nio != NULL) {
2867 ASSERT(nio->io_stage == ZIO_STAGE_DVA_THROTTLE);
2868 /*
2869 * We are passing control to a new zio so make sure that
2870 * it is processed by a different thread. We do this to
2871 * avoid stack overflows that can occur when parents are
2872 * throttled and children are making progress. We allow
2873 * it to go to the head of the taskq since it's already
2874 * been waiting.
2875 */
2876 zio_taskq_dispatch(nio, ZIO_TASKQ_ISSUE, B_TRUE);
2877 }
2878 return (ZIO_PIPELINE_STOP);
2879 }
2880
2881 void
2882 zio_allocate_dispatch(spa_t *spa)
2883 {
2884 zio_t *zio;
2885
2886 mutex_enter(&spa->spa_alloc_lock);
2887 zio = zio_io_to_allocate(spa);
2888 mutex_exit(&spa->spa_alloc_lock);
2889 if (zio == NULL)
2890 return;
2891
2892 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
2893 ASSERT0(zio->io_error);
2894 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
2895 }
2896
2897 static int
2898 zio_dva_allocate(zio_t *zio)
2899 {
2900 spa_t *spa = zio->io_spa;
2901 metaslab_class_t *mc = spa_normal_class(spa);
2902 blkptr_t *bp = zio->io_bp;
2903 int error;
2904 int flags = 0;
2905
2906 if (zio->io_gang_leader == NULL) {
2907 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2908 zio->io_gang_leader = zio;
2909 }
2910
2911 ASSERT(BP_IS_HOLE(bp));
2912 ASSERT0(BP_GET_NDVAS(bp));
2913 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2914 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2915 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2916
2917 if (zio->io_flags & ZIO_FLAG_NODATA) {
2918 flags |= METASLAB_DONT_THROTTLE;
2919 }
2920 if (zio->io_flags & ZIO_FLAG_GANG_CHILD) {
2921 flags |= METASLAB_GANG_CHILD;
2922 }
2923 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) {
2924 flags |= METASLAB_ASYNC_ALLOC;
2925 }
2926
2927 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2928 zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
2929 &zio->io_alloc_list, zio);
2930
2931 if (error != 0) {
2932 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2933 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2934 error);
2935 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2936 return (zio_write_gang_block(zio));
2937 zio->io_error = error;
2938 }
2939
2940 return (ZIO_PIPELINE_CONTINUE);
2941 }
2942
2943 static int
2944 zio_dva_free(zio_t *zio)
2945 {
2946 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2947
2948 return (ZIO_PIPELINE_CONTINUE);
2949 }
2950
2951 static int
2952 zio_dva_claim(zio_t *zio)
2953 {
2954 int error;
2955
2956 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2957 if (error)
2958 zio->io_error = error;
2959
2960 return (ZIO_PIPELINE_CONTINUE);
2961 }
2962
2963 /*
2964 * Undo an allocation. This is used by zio_done() when an I/O fails
2965 * and we want to give back the block we just allocated.
2966 * This handles both normal blocks and gang blocks.
2967 */
2968 static void
2969 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2970 {
2971 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2972 ASSERT(zio->io_bp_override == NULL);
2973
2974 if (!BP_IS_HOLE(bp))
2975 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2976
2977 if (gn != NULL) {
2978 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2979 zio_dva_unallocate(zio, gn->gn_child[g],
2980 &gn->gn_gbh->zg_blkptr[g]);
2981 }
2982 }
2983 }
2984
2985 /*
2986 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2987 */
2988 int
2989 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2990 uint64_t size, boolean_t *slog)
2991 {
2992 int error = 1;
2993 zio_alloc_list_t io_alloc_list;
2994
2995 ASSERT(txg > spa_syncing_txg(spa));
2996
2997 metaslab_trace_init(&io_alloc_list);
2998 error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
2999 txg, old_bp, METASLAB_HINTBP_AVOID, &io_alloc_list, NULL);
3000 if (error == 0) {
3001 *slog = TRUE;
3002 } else {
3003 error = metaslab_alloc(spa, spa_normal_class(spa), size,
3004 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID,
3005 &io_alloc_list, NULL);
3006 if (error == 0)
3007 *slog = FALSE;
3008 }
3009 metaslab_trace_fini(&io_alloc_list);
3010
3011 if (error == 0) {
3012 BP_SET_LSIZE(new_bp, size);
3013 BP_SET_PSIZE(new_bp, size);
3014 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
3015 BP_SET_CHECKSUM(new_bp,
3016 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
3017 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
3018 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3019 BP_SET_LEVEL(new_bp, 0);
3020 BP_SET_DEDUP(new_bp, 0);
3021 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
3022 } else {
3023 zfs_dbgmsg("%s: zil block allocation failure: "
3024 "size %llu, error %d", spa_name(spa), size, error);
3025 }
3026
3027 return (error);
3028 }
3029
3030 /*
3031 * Free an intent log block.
3032 */
3033 void
3034 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
3035 {
3036 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
3037 ASSERT(!BP_IS_GANG(bp));
3038
3039 zio_free(spa, txg, bp);
3040 }
3041
3042 /*
3043 * ==========================================================================
3044 * Read and write to physical devices
3045 * ==========================================================================
3046 */
3047
3048
3049 /*
3050 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3051 * stops after this stage and will resume upon I/O completion.
3052 * However, there are instances where the vdev layer may need to
3053 * continue the pipeline when an I/O was not issued. Since the I/O
3054 * that was sent to the vdev layer might be different than the one
3055 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3056 * force the underlying vdev layers to call either zio_execute() or
3057 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3058 */
3059 static int
3060 zio_vdev_io_start(zio_t *zio)
3061 {
3062 vdev_t *vd = zio->io_vd;
3063 uint64_t align;
3064 spa_t *spa = zio->io_spa;
3065
3066 ASSERT(zio->io_error == 0);
3067 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3068
3069 if (vd == NULL) {
3070 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3071 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3072
3073 /*
3074 * The mirror_ops handle multiple DVAs in a single BP.
3075 */
3076 vdev_mirror_ops.vdev_op_io_start(zio);
3077 return (ZIO_PIPELINE_STOP);
3078 }
3079
3080 ASSERT3P(zio->io_logical, !=, zio);
3081 if (zio->io_type == ZIO_TYPE_WRITE) {
3082 ASSERT(spa->spa_trust_config);
3083
3084 if (zio->io_vd->vdev_removing) {
3085 ASSERT(zio->io_flags &
3086 (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL |
3087 ZIO_FLAG_INDUCE_DAMAGE));
3088 }
3089 }
3090
3091 /*
3092 * We keep track of time-sensitive I/Os so that the scan thread
3093 * can quickly react to certain workloads. In particular, we care
3094 * about non-scrubbing, top-level reads and writes with the following
3095 * characteristics:
3096 * - synchronous writes of user data to non-slog devices
3097 * - any reads of user data
3098 * When these conditions are met, adjust the timestamp of spa_last_io
3099 * which allows the scan thread to adjust its workload accordingly.
3100 */
3101 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
3102 vd == vd->vdev_top && !vd->vdev_islog &&
3103 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
3104 zio->io_txg != spa_syncing_txg(spa)) {
3105 uint64_t old = spa->spa_last_io;
3106 uint64_t new = ddi_get_lbolt64();
3107 if (old != new)
3108 (void) atomic_cas_64(&spa->spa_last_io, old, new);
3109 }
3110
3111 align = 1ULL << vd->vdev_top->vdev_ashift;
3112
3113 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3114 P2PHASE(zio->io_size, align) != 0) {
3115 /* Transform logical writes to be a full physical block size. */
3116 uint64_t asize = P2ROUNDUP(zio->io_size, align);
3117 abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize);
3118 ASSERT(vd == vd->vdev_top);
3119 if (zio->io_type == ZIO_TYPE_WRITE) {
3120 abd_copy(abuf, zio->io_abd, zio->io_size);
3121 abd_zero_off(abuf, zio->io_size, asize - zio->io_size);
3122 }
3123 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
3124 }
3125
3126 /*
3127 * If this is not a physical io, make sure that it is properly aligned
3128 * before proceeding.
3129 */
3130 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3131 ASSERT0(P2PHASE(zio->io_offset, align));
3132 ASSERT0(P2PHASE(zio->io_size, align));
3133 } else {
3134 /*
3135 * For physical writes, we allow 512b aligned writes and assume
3136 * the device will perform a read-modify-write as necessary.
3137 */
3138 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
3139 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
3140 }
3141
3142 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
3143
3144 /*
3145 * If this is a repair I/O, and there's no self-healing involved --
3146 * that is, we're just resilvering what we expect to resilver --
3147 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3148 * This prevents spurious resilvering with nested replication.
3149 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
3150 * A is out of date, we'll read from C+D, then use the data to
3151 * resilver A+B -- but we don't actually want to resilver B, just A.
3152 * The top-level mirror has no way to know this, so instead we just
3153 * discard unnecessary repairs as we work our way down the vdev tree.
3154 * The same logic applies to any form of nested replication:
3155 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
3156 */
3157 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3158 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3159 zio->io_txg != 0 && /* not a delegated i/o */
3160 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3161 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3162 zio_vdev_io_bypass(zio);
3163 return (ZIO_PIPELINE_CONTINUE);
3164 }
3165
3166 if (vd->vdev_ops->vdev_op_leaf &&
3167 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
3168
3169 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio))
3170 return (ZIO_PIPELINE_CONTINUE);
3171
3172 if ((zio = vdev_queue_io(zio)) == NULL)
3173 return (ZIO_PIPELINE_STOP);
3174
3175 if (!vdev_accessible(vd, zio)) {
3176 zio->io_error = SET_ERROR(ENXIO);
3177 zio_interrupt(zio);
3178 return (ZIO_PIPELINE_STOP);
3179 }
3180 }
3181
3182 vd->vdev_ops->vdev_op_io_start(zio);
3183 return (ZIO_PIPELINE_STOP);
3184 }
3185
3186 static int
3187 zio_vdev_io_done(zio_t *zio)
3188 {
3189 vdev_t *vd = zio->io_vd;
3190 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3191 boolean_t unexpected_error = B_FALSE;
3192
3193 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3194 return (ZIO_PIPELINE_STOP);
3195 }
3196
3197 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
3198
3199 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
3200
3201 vdev_queue_io_done(zio);
3202
3203 if (zio->io_type == ZIO_TYPE_WRITE)
3204 vdev_cache_write(zio);
3205
3206 if (zio_injection_enabled && zio->io_error == 0)
3207 zio->io_error = zio_handle_device_injection(vd,
3208 zio, EIO);
3209
3210 if (zio_injection_enabled && zio->io_error == 0)
3211 zio->io_error = zio_handle_label_injection(zio, EIO);
3212
3213 if (zio->io_error) {
3214 if (!vdev_accessible(vd, zio)) {
3215 zio->io_error = SET_ERROR(ENXIO);
3216 } else {
3217 unexpected_error = B_TRUE;
3218 }
3219 }
3220 }
3221
3222 ops->vdev_op_io_done(zio);
3223
3224 if (unexpected_error)
3225 VERIFY(vdev_probe(vd, zio) == NULL);
3226
3227 return (ZIO_PIPELINE_CONTINUE);
3228 }
3229
3230 /*
3231 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3232 * disk, and use that to finish the checksum ereport later.
3233 */
3234 static void
3235 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
3236 const void *good_buf)
3237 {
3238 /* no processing needed */
3239 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
3240 }
3241
3242 /*ARGSUSED*/
3243 void
3244 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
3245 {
3246 void *buf = zio_buf_alloc(zio->io_size);
3247
3248 abd_copy_to_buf(buf, zio->io_abd, zio->io_size);
3249
3250 zcr->zcr_cbinfo = zio->io_size;
3251 zcr->zcr_cbdata = buf;
3252 zcr->zcr_finish = zio_vsd_default_cksum_finish;
3253 zcr->zcr_free = zio_buf_free;
3254 }
3255
3256 static int
3257 zio_vdev_io_assess(zio_t *zio)
3258 {
3259 vdev_t *vd = zio->io_vd;
3260
3261 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3262 return (ZIO_PIPELINE_STOP);
3263 }
3264
3265 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3266 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
3267
3268 if (zio->io_vsd != NULL) {
3269 zio->io_vsd_ops->vsd_free(zio);
3270 zio->io_vsd = NULL;
3271 }
3272
3273 if (zio_injection_enabled && zio->io_error == 0)
3274 zio->io_error = zio_handle_fault_injection(zio, EIO);
3275
3276 /*
3277 * If the I/O failed, determine whether we should attempt to retry it.
3278 *
3279 * On retry, we cut in line in the issue queue, since we don't want
3280 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3281 */
3282 if (zio->io_error && vd == NULL &&
3283 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
3284 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
3285 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
3286 zio->io_error = 0;
3287 zio->io_flags |= ZIO_FLAG_IO_RETRY |
3288 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
3289 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
3290 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
3291 zio_requeue_io_start_cut_in_line);
3292 return (ZIO_PIPELINE_STOP);
3293 }
3294
3295 /*
3296 * If we got an error on a leaf device, convert it to ENXIO
3297 * if the device is not accessible at all.
3298 */
3299 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3300 !vdev_accessible(vd, zio))
3301 zio->io_error = SET_ERROR(ENXIO);
3302
3303 /*
3304 * If we can't write to an interior vdev (mirror or RAID-Z),
3305 * set vdev_cant_write so that we stop trying to allocate from it.
3306 */
3307 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
3308 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
3309 vd->vdev_cant_write = B_TRUE;
3310 }
3311
3312 /*
3313 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3314 * attempts will ever succeed. In this case we set a persistent bit so
3315 * that we don't bother with it in the future.
3316 */
3317 if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) &&
3318 zio->io_type == ZIO_TYPE_IOCTL &&
3319 zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL)
3320 vd->vdev_nowritecache = B_TRUE;
3321
3322 if (zio->io_error)
3323 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3324
3325 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3326 zio->io_physdone != NULL) {
3327 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
3328 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
3329 zio->io_physdone(zio->io_logical);
3330 }
3331
3332 return (ZIO_PIPELINE_CONTINUE);
3333 }
3334
3335 void
3336 zio_vdev_io_reissue(zio_t *zio)
3337 {
3338 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3339 ASSERT(zio->io_error == 0);
3340
3341 zio->io_stage >>= 1;
3342 }
3343
3344 void
3345 zio_vdev_io_redone(zio_t *zio)
3346 {
3347 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
3348
3349 zio->io_stage >>= 1;
3350 }
3351
3352 void
3353 zio_vdev_io_bypass(zio_t *zio)
3354 {
3355 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3356 ASSERT(zio->io_error == 0);
3357
3358 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
3359 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
3360 }
3361
3362 /*
3363 * ==========================================================================
3364 * Generate and verify checksums
3365 * ==========================================================================
3366 */
3367 static int
3368 zio_checksum_generate(zio_t *zio)
3369 {
3370 blkptr_t *bp = zio->io_bp;
3371 enum zio_checksum checksum;
3372
3373 if (bp == NULL) {
3374 /*
3375 * This is zio_write_phys().
3376 * We're either generating a label checksum, or none at all.
3377 */
3378 checksum = zio->io_prop.zp_checksum;
3379
3380 if (checksum == ZIO_CHECKSUM_OFF)
3381 return (ZIO_PIPELINE_CONTINUE);
3382
3383 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
3384 } else {
3385 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
3386 ASSERT(!IO_IS_ALLOCATING(zio));
3387 checksum = ZIO_CHECKSUM_GANG_HEADER;
3388 } else {
3389 checksum = BP_GET_CHECKSUM(bp);
3390 }
3391 }
3392
3393 zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size);
3394
3395 return (ZIO_PIPELINE_CONTINUE);
3396 }
3397
3398 static int
3399 zio_checksum_verify(zio_t *zio)
3400 {
3401 zio_bad_cksum_t info;
3402 blkptr_t *bp = zio->io_bp;
3403 int error;
3404
3405 ASSERT(zio->io_vd != NULL);
3406
3407 if (bp == NULL) {
3408 /*
3409 * This is zio_read_phys().
3410 * We're either verifying a label checksum, or nothing at all.
3411 */
3412 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
3413 return (ZIO_PIPELINE_CONTINUE);
3414
3415 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
3416 }
3417
3418 if ((error = zio_checksum_error(zio, &info)) != 0) {
3419 zio->io_error = error;
3420 if (error == ECKSUM &&
3421 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
3422 zfs_ereport_start_checksum(zio->io_spa,
3423 zio->io_vd, zio, zio->io_offset,
3424 zio->io_size, NULL, &info);
3425 }
3426 }
3427
3428 return (ZIO_PIPELINE_CONTINUE);
3429 }
3430
3431 /*
3432 * Called by RAID-Z to ensure we don't compute the checksum twice.
3433 */
3434 void
3435 zio_checksum_verified(zio_t *zio)
3436 {
3437 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
3438 }
3439
3440 /*
3441 * ==========================================================================
3442 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3443 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3444 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3445 * indicate errors that are specific to one I/O, and most likely permanent.
3446 * Any other error is presumed to be worse because we weren't expecting it.
3447 * ==========================================================================
3448 */
3449 int
3450 zio_worst_error(int e1, int e2)
3451 {
3452 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
3453 int r1, r2;
3454
3455 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
3456 if (e1 == zio_error_rank[r1])
3457 break;
3458
3459 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3460 if (e2 == zio_error_rank[r2])
3461 break;
3462
3463 return (r1 > r2 ? e1 : e2);
3464 }
3465
3466 /*
3467 * ==========================================================================
3468 * I/O completion
3469 * ==========================================================================
3470 */
3471 static int
3472 zio_ready(zio_t *zio)
3473 {
3474 blkptr_t *bp = zio->io_bp;
3475 zio_t *pio, *pio_next;
3476 zio_link_t *zl = NULL;
3477
3478 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT,
3479 ZIO_WAIT_READY)) {
3480 return (ZIO_PIPELINE_STOP);
3481 }
3482
3483 if (zio->io_ready) {
3484 ASSERT(IO_IS_ALLOCATING(zio));
3485 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3486 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3487 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3488
3489 zio->io_ready(zio);
3490 }
3491
3492 if (bp != NULL && bp != &zio->io_bp_copy)
3493 zio->io_bp_copy = *bp;
3494
3495 if (zio->io_error != 0) {
3496 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3497
3498 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3499 ASSERT(IO_IS_ALLOCATING(zio));
3500 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3501 /*
3502 * We were unable to allocate anything, unreserve and
3503 * issue the next I/O to allocate.
3504 */
3505 metaslab_class_throttle_unreserve(
3506 spa_normal_class(zio->io_spa),
3507 zio->io_prop.zp_copies, zio);
3508 zio_allocate_dispatch(zio->io_spa);
3509 }
3510 }
3511
3512 mutex_enter(&zio->io_lock);
3513 zio->io_state[ZIO_WAIT_READY] = 1;
3514 pio = zio_walk_parents(zio, &zl);
3515 mutex_exit(&zio->io_lock);
3516
3517 /*
3518 * As we notify zio's parents, new parents could be added.
3519 * New parents go to the head of zio's io_parent_list, however,
3520 * so we will (correctly) not notify them. The remainder of zio's
3521 * io_parent_list, from 'pio_next' onward, cannot change because
3522 * all parents must wait for us to be done before they can be done.
3523 */
3524 for (; pio != NULL; pio = pio_next) {
3525 pio_next = zio_walk_parents(zio, &zl);
3526 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3527 }
3528
3529 if (zio->io_flags & ZIO_FLAG_NODATA) {
3530 if (BP_IS_GANG(bp)) {
3531 zio->io_flags &= ~ZIO_FLAG_NODATA;
3532 } else {
3533 ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE);
3534 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3535 }
3536 }
3537
3538 if (zio_injection_enabled &&
3539 zio->io_spa->spa_syncing_txg == zio->io_txg)
3540 zio_handle_ignored_writes(zio);
3541
3542 return (ZIO_PIPELINE_CONTINUE);
3543 }
3544
3545 /*
3546 * Update the allocation throttle accounting.
3547 */
3548 static void
3549 zio_dva_throttle_done(zio_t *zio)
3550 {
3551 zio_t *lio = zio->io_logical;
3552 zio_t *pio = zio_unique_parent(zio);
3553 vdev_t *vd = zio->io_vd;
3554 int flags = METASLAB_ASYNC_ALLOC;
3555
3556 ASSERT3P(zio->io_bp, !=, NULL);
3557 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
3558 ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
3559 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
3560 ASSERT(vd != NULL);
3561 ASSERT3P(vd, ==, vd->vdev_top);
3562 ASSERT(!(zio->io_flags & (ZIO_FLAG_IO_REPAIR | ZIO_FLAG_IO_RETRY)));
3563 ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
3564 ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
3565 ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
3566
3567 /*
3568 * Parents of gang children can have two flavors -- ones that
3569 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
3570 * and ones that allocated the constituent blocks. The allocation
3571 * throttle needs to know the allocating parent zio so we must find
3572 * it here.
3573 */
3574 if (pio->io_child_type == ZIO_CHILD_GANG) {
3575 /*
3576 * If our parent is a rewrite gang child then our grandparent
3577 * would have been the one that performed the allocation.
3578 */
3579 if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
3580 pio = zio_unique_parent(pio);
3581 flags |= METASLAB_GANG_CHILD;
3582 }
3583
3584 ASSERT(IO_IS_ALLOCATING(pio));
3585 ASSERT3P(zio, !=, zio->io_logical);
3586 ASSERT(zio->io_logical != NULL);
3587 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
3588 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
3589
3590 mutex_enter(&pio->io_lock);
3591 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags);
3592 mutex_exit(&pio->io_lock);
3593
3594 metaslab_class_throttle_unreserve(spa_normal_class(zio->io_spa),
3595 1, pio);
3596
3597 /*
3598 * Call into the pipeline to see if there is more work that
3599 * needs to be done. If there is work to be done it will be
3600 * dispatched to another taskq thread.
3601 */
3602 zio_allocate_dispatch(zio->io_spa);
3603 }
3604
3605 static int
3606 zio_done(zio_t *zio)
3607 {
3608 spa_t *spa = zio->io_spa;
3609 zio_t *lio = zio->io_logical;
3610 blkptr_t *bp = zio->io_bp;
3611 vdev_t *vd = zio->io_vd;
3612 uint64_t psize = zio->io_size;
3613 zio_t *pio, *pio_next;
3614 metaslab_class_t *mc = spa_normal_class(spa);
3615 zio_link_t *zl = NULL;
3616
3617 /*
3618 * If our children haven't all completed,
3619 * wait for them and then repeat this pipeline stage.
3620 */
3621 if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) {
3622 return (ZIO_PIPELINE_STOP);
3623 }
3624
3625 /*
3626 * If the allocation throttle is enabled, then update the accounting.
3627 * We only track child I/Os that are part of an allocating async
3628 * write. We must do this since the allocation is performed
3629 * by the logical I/O but the actual write is done by child I/Os.
3630 */
3631 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
3632 zio->io_child_type == ZIO_CHILD_VDEV) {
3633 ASSERT(mc->mc_alloc_throttle_enabled);
3634 zio_dva_throttle_done(zio);
3635 }
3636
3637 /*
3638 * If the allocation throttle is enabled, verify that
3639 * we have decremented the refcounts for every I/O that was throttled.
3640 */
3641 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3642 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3643 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3644 ASSERT(bp != NULL);
3645 metaslab_group_alloc_verify(spa, zio->io_bp, zio);
3646 VERIFY(refcount_not_held(&mc->mc_alloc_slots, zio));
3647 }
3648
3649 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3650 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3651 ASSERT(zio->io_children[c][w] == 0);
3652
3653 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3654 ASSERT(bp->blk_pad[0] == 0);
3655 ASSERT(bp->blk_pad[1] == 0);
3656 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3657 (bp == zio_unique_parent(zio)->io_bp));
3658 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3659 zio->io_bp_override == NULL &&
3660 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3661 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3662 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3663 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3664 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3665 }
3666 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3667 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3668 }
3669
3670 /*
3671 * If there were child vdev/gang/ddt errors, they apply to us now.
3672 */
3673 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3674 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3675 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3676
3677 /*
3678 * If the I/O on the transformed data was successful, generate any
3679 * checksum reports now while we still have the transformed data.
3680 */
3681 if (zio->io_error == 0) {
3682 while (zio->io_cksum_report != NULL) {
3683 zio_cksum_report_t *zcr = zio->io_cksum_report;
3684 uint64_t align = zcr->zcr_align;
3685 uint64_t asize = P2ROUNDUP(psize, align);
3686 char *abuf = NULL;
3687 abd_t *adata = zio->io_abd;
3688
3689 if (asize != psize) {
3690 adata = abd_alloc_linear(asize, B_TRUE);
3691 abd_copy(adata, zio->io_abd, psize);
3692 abd_zero_off(adata, psize, asize - psize);
3693 }
3694
3695 if (adata != NULL)
3696 abuf = abd_borrow_buf_copy(adata, asize);
3697
3698 zio->io_cksum_report = zcr->zcr_next;
3699 zcr->zcr_next = NULL;
3700 zcr->zcr_finish(zcr, abuf);
3701 zfs_ereport_free_checksum(zcr);
3702
3703 if (adata != NULL)
3704 abd_return_buf(adata, abuf, asize);
3705
3706 if (asize != psize)
3707 abd_free(adata);
3708 }
3709 }
3710
3711 zio_pop_transforms(zio); /* note: may set zio->io_error */
3712
3713 vdev_stat_update(zio, psize);
3714
3715 if (zio->io_error) {
3716 /*
3717 * If this I/O is attached to a particular vdev,
3718 * generate an error message describing the I/O failure
3719 * at the block level. We ignore these errors if the
3720 * device is currently unavailable.
3721 */
3722 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3723 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3724
3725 if ((zio->io_error == EIO || !(zio->io_flags &
3726 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3727 zio == lio) {
3728 /*
3729 * For logical I/O requests, tell the SPA to log the
3730 * error and generate a logical data ereport.
3731 */
3732 spa_log_error(spa, zio);
3733 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3734 0, 0);
3735 }
3736 }
3737
3738 if (zio->io_error && zio == lio) {
3739 /*
3740 * Determine whether zio should be reexecuted. This will
3741 * propagate all the way to the root via zio_notify_parent().
3742 */
3743 ASSERT(vd == NULL && bp != NULL);
3744 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3745
3746 if (IO_IS_ALLOCATING(zio) &&
3747 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3748 if (zio->io_error != ENOSPC)
3749 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3750 else
3751 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3752 }
3753
3754 if ((zio->io_type == ZIO_TYPE_READ ||
3755 zio->io_type == ZIO_TYPE_FREE) &&
3756 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3757 zio->io_error == ENXIO &&
3758 spa_load_state(spa) == SPA_LOAD_NONE &&
3759 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3760 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3761
3762 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3763 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3764
3765 /*
3766 * Here is a possibly good place to attempt to do
3767 * either combinatorial reconstruction or error correction
3768 * based on checksums. It also might be a good place
3769 * to send out preliminary ereports before we suspend
3770 * processing.
3771 */
3772 }
3773
3774 /*
3775 * If there were logical child errors, they apply to us now.
3776 * We defer this until now to avoid conflating logical child
3777 * errors with errors that happened to the zio itself when
3778 * updating vdev stats and reporting FMA events above.
3779 */
3780 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3781
3782 if ((zio->io_error || zio->io_reexecute) &&
3783 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3784 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3785 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3786
3787 zio_gang_tree_free(&zio->io_gang_tree);
3788
3789 /*
3790 * Godfather I/Os should never suspend.
3791 */
3792 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3793 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3794 zio->io_reexecute = 0;
3795
3796 if (zio->io_reexecute) {
3797 /*
3798 * This is a logical I/O that wants to reexecute.
3799 *
3800 * Reexecute is top-down. When an i/o fails, if it's not
3801 * the root, it simply notifies its parent and sticks around.
3802 * The parent, seeing that it still has children in zio_done(),
3803 * does the same. This percolates all the way up to the root.
3804 * The root i/o will reexecute or suspend the entire tree.
3805 *
3806 * This approach ensures that zio_reexecute() honors
3807 * all the original i/o dependency relationships, e.g.
3808 * parents not executing until children are ready.
3809 */
3810 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3811
3812 zio->io_gang_leader = NULL;
3813
3814 mutex_enter(&zio->io_lock);
3815 zio->io_state[ZIO_WAIT_DONE] = 1;
3816 mutex_exit(&zio->io_lock);
3817
3818 /*
3819 * "The Godfather" I/O monitors its children but is
3820 * not a true parent to them. It will track them through
3821 * the pipeline but severs its ties whenever they get into
3822 * trouble (e.g. suspended). This allows "The Godfather"
3823 * I/O to return status without blocking.
3824 */
3825 zl = NULL;
3826 for (pio = zio_walk_parents(zio, &zl); pio != NULL;
3827 pio = pio_next) {
3828 zio_link_t *remove_zl = zl;
3829 pio_next = zio_walk_parents(zio, &zl);
3830
3831 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3832 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3833 zio_remove_child(pio, zio, remove_zl);
3834 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3835 }
3836 }
3837
3838 if ((pio = zio_unique_parent(zio)) != NULL) {
3839 /*
3840 * We're not a root i/o, so there's nothing to do
3841 * but notify our parent. Don't propagate errors
3842 * upward since we haven't permanently failed yet.
3843 */
3844 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3845 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3846 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3847 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3848 /*
3849 * We'd fail again if we reexecuted now, so suspend
3850 * until conditions improve (e.g. device comes online).
3851 */
3852 zio_suspend(spa, zio);
3853 } else {
3854 /*
3855 * Reexecution is potentially a huge amount of work.
3856 * Hand it off to the otherwise-unused claim taskq.
3857 */
3858 ASSERT(zio->io_tqent.tqent_next == NULL);
3859 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3860 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3861 0, &zio->io_tqent);
3862 }
3863 return (ZIO_PIPELINE_STOP);
3864 }
3865
3866 ASSERT(zio->io_child_count == 0);
3867 ASSERT(zio->io_reexecute == 0);
3868 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3869
3870 /*
3871 * Report any checksum errors, since the I/O is complete.
3872 */
3873 while (zio->io_cksum_report != NULL) {
3874 zio_cksum_report_t *zcr = zio->io_cksum_report;
3875 zio->io_cksum_report = zcr->zcr_next;
3876 zcr->zcr_next = NULL;
3877 zcr->zcr_finish(zcr, NULL);
3878 zfs_ereport_free_checksum(zcr);
3879 }
3880
3881 /*
3882 * It is the responsibility of the done callback to ensure that this
3883 * particular zio is no longer discoverable for adoption, and as
3884 * such, cannot acquire any new parents.
3885 */
3886 if (zio->io_done)
3887 zio->io_done(zio);
3888
3889 mutex_enter(&zio->io_lock);
3890 zio->io_state[ZIO_WAIT_DONE] = 1;
3891 mutex_exit(&zio->io_lock);
3892
3893 zl = NULL;
3894 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
3895 zio_link_t *remove_zl = zl;
3896 pio_next = zio_walk_parents(zio, &zl);
3897 zio_remove_child(pio, zio, remove_zl);
3898 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3899 }
3900
3901 if (zio->io_waiter != NULL) {
3902 mutex_enter(&zio->io_lock);
3903 zio->io_executor = NULL;
3904 cv_broadcast(&zio->io_cv);
3905 mutex_exit(&zio->io_lock);
3906 } else {
3907 zio_destroy(zio);
3908 }
3909
3910 return (ZIO_PIPELINE_STOP);
3911 }
3912
3913 /*
3914 * ==========================================================================
3915 * I/O pipeline definition
3916 * ==========================================================================
3917 */
3918 static zio_pipe_stage_t *zio_pipeline[] = {
3919 NULL,
3920 zio_read_bp_init,
3921 zio_write_bp_init,
3922 zio_free_bp_init,
3923 zio_issue_async,
3924 zio_write_compress,
3925 zio_checksum_generate,
3926 zio_nop_write,
3927 zio_ddt_read_start,
3928 zio_ddt_read_done,
3929 zio_ddt_write,
3930 zio_ddt_free,
3931 zio_gang_assemble,
3932 zio_gang_issue,
3933 zio_dva_throttle,
3934 zio_dva_allocate,
3935 zio_dva_free,
3936 zio_dva_claim,
3937 zio_ready,
3938 zio_vdev_io_start,
3939 zio_vdev_io_done,
3940 zio_vdev_io_assess,
3941 zio_checksum_verify,
3942 zio_done
3943 };
3944
3945
3946
3947
3948 /*
3949 * Compare two zbookmark_phys_t's to see which we would reach first in a
3950 * pre-order traversal of the object tree.
3951 *
3952 * This is simple in every case aside from the meta-dnode object. For all other
3953 * objects, we traverse them in order (object 1 before object 2, and so on).
3954 * However, all of these objects are traversed while traversing object 0, since
3955 * the data it points to is the list of objects. Thus, we need to convert to a
3956 * canonical representation so we can compare meta-dnode bookmarks to
3957 * non-meta-dnode bookmarks.
3958 *
3959 * We do this by calculating "equivalents" for each field of the zbookmark.
3960 * zbookmarks outside of the meta-dnode use their own object and level, and
3961 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
3962 * blocks this bookmark refers to) by multiplying their blkid by their span
3963 * (the number of L0 blocks contained within one block at their level).
3964 * zbookmarks inside the meta-dnode calculate their object equivalent
3965 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
3966 * level + 1<<31 (any value larger than a level could ever be) for their level.
3967 * This causes them to always compare before a bookmark in their object
3968 * equivalent, compare appropriately to bookmarks in other objects, and to
3969 * compare appropriately to other bookmarks in the meta-dnode.
3970 */
3971 int
3972 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
3973 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
3974 {
3975 /*
3976 * These variables represent the "equivalent" values for the zbookmark,
3977 * after converting zbookmarks inside the meta dnode to their
3978 * normal-object equivalents.
3979 */
3980 uint64_t zb1obj, zb2obj;
3981 uint64_t zb1L0, zb2L0;
3982 uint64_t zb1level, zb2level;
3983
3984 if (zb1->zb_object == zb2->zb_object &&
3985 zb1->zb_level == zb2->zb_level &&
3986 zb1->zb_blkid == zb2->zb_blkid)
3987 return (0);
3988
3989 /*
3990 * BP_SPANB calculates the span in blocks.
3991 */
3992 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
3993 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
3994
3995 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3996 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
3997 zb1L0 = 0;
3998 zb1level = zb1->zb_level + COMPARE_META_LEVEL;
3999 } else {
4000 zb1obj = zb1->zb_object;
4001 zb1level = zb1->zb_level;
4002 }
4003
4004 if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
4005 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4006 zb2L0 = 0;
4007 zb2level = zb2->zb_level + COMPARE_META_LEVEL;
4008 } else {
4009 zb2obj = zb2->zb_object;
4010 zb2level = zb2->zb_level;
4011 }
4012
4013 /* Now that we have a canonical representation, do the comparison. */
4014 if (zb1obj != zb2obj)
4015 return (zb1obj < zb2obj ? -1 : 1);
4016 else if (zb1L0 != zb2L0)
4017 return (zb1L0 < zb2L0 ? -1 : 1);
4018 else if (zb1level != zb2level)
4019 return (zb1level > zb2level ? -1 : 1);
4020 /*
4021 * This can (theoretically) happen if the bookmarks have the same object
4022 * and level, but different blkids, if the block sizes are not the same.
4023 * There is presently no way to change the indirect block sizes
4024 */
4025 return (0);
4026 }
4027
4028 /*
4029 * This function checks the following: given that last_block is the place that
4030 * our traversal stopped last time, does that guarantee that we've visited
4031 * every node under subtree_root? Therefore, we can't just use the raw output
4032 * of zbookmark_compare. We have to pass in a modified version of
4033 * subtree_root; by incrementing the block id, and then checking whether
4034 * last_block is before or equal to that, we can tell whether or not having
4035 * visited last_block implies that all of subtree_root's children have been
4036 * visited.
4037 */
4038 boolean_t
4039 zbookmark_subtree_completed(const dnode_phys_t *dnp,
4040 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
4041 {
4042 zbookmark_phys_t mod_zb = *subtree_root;
4043 mod_zb.zb_blkid++;
4044 ASSERT(last_block->zb_level == 0);
4045
4046 /* The objset_phys_t isn't before anything. */
4047 if (dnp == NULL)
4048 return (B_FALSE);
4049
4050 /*
4051 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4052 * data block size in sectors, because that variable is only used if
4053 * the bookmark refers to a block in the meta-dnode. Since we don't
4054 * know without examining it what object it refers to, and there's no
4055 * harm in passing in this value in other cases, we always pass it in.
4056 *
4057 * We pass in 0 for the indirect block size shift because zb2 must be
4058 * level 0. The indirect block size is only used to calculate the span
4059 * of the bookmark, but since the bookmark must be level 0, the span is
4060 * always 1, so the math works out.
4061 *
4062 * If you make changes to how the zbookmark_compare code works, be sure
4063 * to make sure that this code still works afterwards.
4064 */
4065 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
4066 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,
4067 last_block) <= 0);
4068 }