1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2014 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25 */
26
27 #include <sys/zfs_context.h>
28 #include <sys/fm/fs/zfs.h>
29 #include <sys/spa.h>
30 #include <sys/txg.h>
31 #include <sys/spa_impl.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/zio_impl.h>
34 #include <sys/zio_compress.h>
35 #include <sys/zio_checksum.h>
36 #include <sys/dmu_objset.h>
37 #include <sys/arc.h>
38 #include <sys/ddt.h>
39 #include <sys/blkptr.h>
40 #include <sys/zfeature.h>
41
42 /*
43 * ==========================================================================
44 * I/O type descriptions
45 * ==========================================================================
46 */
47 const char *zio_type_name[ZIO_TYPES] = {
48 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
49 "zio_ioctl"
50 };
51
52 /*
53 * ==========================================================================
54 * I/O kmem caches
55 * ==========================================================================
56 */
57 kmem_cache_t *zio_cache;
58 kmem_cache_t *zio_link_cache;
59 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
60 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
61
62 #ifdef _KERNEL
63 extern vmem_t *zio_alloc_arena;
64 #endif
65
66 /*
67 * The following actions directly effect the spa's sync-to-convergence logic.
68 * The values below define the sync pass when we start performing the action.
69 * Care should be taken when changing these values as they directly impact
70 * spa_sync() performance. Tuning these values may introduce subtle performance
71 * pathologies and should only be done in the context of performance analysis.
72 * These tunables will eventually be removed and replaced with #defines once
73 * enough analysis has been done to determine optimal values.
74 *
75 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
76 * regular blocks are not deferred.
77 */
78 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
79 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
80 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
81
82 /*
83 * An allocating zio is one that either currently has the DVA allocate
84 * stage set or will have it later in its lifetime.
85 */
86 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
87
88 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
89
90 #ifdef ZFS_DEBUG
91 int zio_buf_debug_limit = 16384;
92 #else
93 int zio_buf_debug_limit = 0;
94 #endif
95
96 void
97 zio_init(void)
98 {
99 size_t c;
100 vmem_t *data_alloc_arena = NULL;
101
102 #ifdef _KERNEL
103 data_alloc_arena = zio_alloc_arena;
104 #endif
105 zio_cache = kmem_cache_create("zio_cache",
106 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
107 zio_link_cache = kmem_cache_create("zio_link_cache",
108 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
109
110 /*
111 * For small buffers, we want a cache for each multiple of
112 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
113 * for each quarter-power of 2. For large buffers, we want
114 * a cache for each multiple of PAGESIZE.
115 */
116 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
117 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
118 size_t p2 = size;
119 size_t align = 0;
120 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
121
122 while (p2 & (p2 - 1))
123 p2 &= p2 - 1;
124
125 #ifndef _KERNEL
126 /*
127 * If we are using watchpoints, put each buffer on its own page,
128 * to eliminate the performance overhead of trapping to the
129 * kernel when modifying a non-watched buffer that shares the
130 * page with a watched buffer.
131 */
132 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
133 continue;
134 #endif
135 if (size <= 4 * SPA_MINBLOCKSIZE) {
136 align = SPA_MINBLOCKSIZE;
137 } else if (IS_P2ALIGNED(size, PAGESIZE)) {
138 align = PAGESIZE;
139 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
140 align = p2 >> 2;
141 }
142
143 if (align != 0) {
144 char name[36];
145 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
146 zio_buf_cache[c] = kmem_cache_create(name, size,
147 align, NULL, NULL, NULL, NULL, NULL, cflags);
148
149 /*
150 * Since zio_data bufs do not appear in crash dumps, we
151 * pass KMC_NOTOUCH so that no allocator metadata is
152 * stored with the buffers.
153 */
154 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
155 zio_data_buf_cache[c] = kmem_cache_create(name, size,
156 align, NULL, NULL, NULL, NULL, data_alloc_arena,
157 cflags | KMC_NOTOUCH);
158 }
159 }
160
161 while (--c != 0) {
162 ASSERT(zio_buf_cache[c] != NULL);
163 if (zio_buf_cache[c - 1] == NULL)
164 zio_buf_cache[c - 1] = zio_buf_cache[c];
165
166 ASSERT(zio_data_buf_cache[c] != NULL);
167 if (zio_data_buf_cache[c - 1] == NULL)
168 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
169 }
170
171 zio_inject_init();
172 }
173
174 void
175 zio_fini(void)
176 {
177 size_t c;
178 kmem_cache_t *last_cache = NULL;
179 kmem_cache_t *last_data_cache = NULL;
180
181 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
182 if (zio_buf_cache[c] != last_cache) {
183 last_cache = zio_buf_cache[c];
184 kmem_cache_destroy(zio_buf_cache[c]);
185 }
186 zio_buf_cache[c] = NULL;
187
188 if (zio_data_buf_cache[c] != last_data_cache) {
189 last_data_cache = zio_data_buf_cache[c];
190 kmem_cache_destroy(zio_data_buf_cache[c]);
191 }
192 zio_data_buf_cache[c] = NULL;
193 }
194
195 kmem_cache_destroy(zio_link_cache);
196 kmem_cache_destroy(zio_cache);
197
198 zio_inject_fini();
199 }
200
201 /*
202 * ==========================================================================
203 * Allocate and free I/O buffers
204 * ==========================================================================
205 */
206
207 /*
208 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
209 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
210 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
211 * excess / transient data in-core during a crashdump.
212 */
213 void *
214 zio_buf_alloc(size_t size)
215 {
216 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
217
218 ASSERT3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
219
220 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
221 }
222
223 /*
224 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
225 * crashdump if the kernel panics. This exists so that we will limit the amount
226 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
227 * of kernel heap dumped to disk when the kernel panics)
228 */
229 void *
230 zio_data_buf_alloc(size_t size)
231 {
232 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
233
234 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
235
236 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
237 }
238
239 void
240 zio_buf_free(void *buf, size_t size)
241 {
242 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
243
244 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
245
246 kmem_cache_free(zio_buf_cache[c], buf);
247 }
248
249 void
250 zio_data_buf_free(void *buf, size_t size)
251 {
252 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
253
254 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
255
256 kmem_cache_free(zio_data_buf_cache[c], buf);
257 }
258
259 /*
260 * ==========================================================================
261 * Push and pop I/O transform buffers
262 * ==========================================================================
263 */
264 static void
265 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
266 zio_transform_func_t *transform)
267 {
268 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
269
270 zt->zt_orig_data = zio->io_data;
271 zt->zt_orig_size = zio->io_size;
272 zt->zt_bufsize = bufsize;
273 zt->zt_transform = transform;
274
275 zt->zt_next = zio->io_transform_stack;
276 zio->io_transform_stack = zt;
277
278 zio->io_data = data;
279 zio->io_size = size;
280 }
281
282 static void
283 zio_pop_transforms(zio_t *zio)
284 {
285 zio_transform_t *zt;
286
287 while ((zt = zio->io_transform_stack) != NULL) {
288 if (zt->zt_transform != NULL)
289 zt->zt_transform(zio,
290 zt->zt_orig_data, zt->zt_orig_size);
291
292 if (zt->zt_bufsize != 0)
293 zio_buf_free(zio->io_data, zt->zt_bufsize);
294
295 zio->io_data = zt->zt_orig_data;
296 zio->io_size = zt->zt_orig_size;
297 zio->io_transform_stack = zt->zt_next;
298
299 kmem_free(zt, sizeof (zio_transform_t));
300 }
301 }
302
303 /*
304 * ==========================================================================
305 * I/O transform callbacks for subblocks and decompression
306 * ==========================================================================
307 */
308 static void
309 zio_subblock(zio_t *zio, void *data, uint64_t size)
310 {
311 ASSERT(zio->io_size > size);
312
313 if (zio->io_type == ZIO_TYPE_READ)
314 bcopy(zio->io_data, data, size);
315 }
316
317 static void
318 zio_decompress(zio_t *zio, void *data, uint64_t size)
319 {
320 if (zio->io_error == 0 &&
321 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
322 zio->io_data, data, zio->io_size, size) != 0)
323 zio->io_error = SET_ERROR(EIO);
324 }
325
326 /*
327 * ==========================================================================
328 * I/O parent/child relationships and pipeline interlocks
329 * ==========================================================================
330 */
331 /*
332 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
333 * continue calling these functions until they return NULL.
334 * Otherwise, the next caller will pick up the list walk in
335 * some indeterminate state. (Otherwise every caller would
336 * have to pass in a cookie to keep the state represented by
337 * io_walk_link, which gets annoying.)
338 */
339 zio_t *
340 zio_walk_parents(zio_t *cio)
341 {
342 zio_link_t *zl = cio->io_walk_link;
343 list_t *pl = &cio->io_parent_list;
344
345 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
346 cio->io_walk_link = zl;
347
348 if (zl == NULL)
349 return (NULL);
350
351 ASSERT(zl->zl_child == cio);
352 return (zl->zl_parent);
353 }
354
355 zio_t *
356 zio_walk_children(zio_t *pio)
357 {
358 zio_link_t *zl = pio->io_walk_link;
359 list_t *cl = &pio->io_child_list;
360
361 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
362 pio->io_walk_link = zl;
363
364 if (zl == NULL)
365 return (NULL);
366
367 ASSERT(zl->zl_parent == pio);
368 return (zl->zl_child);
369 }
370
371 zio_t *
372 zio_unique_parent(zio_t *cio)
373 {
374 zio_t *pio = zio_walk_parents(cio);
375
376 VERIFY(zio_walk_parents(cio) == NULL);
377 return (pio);
378 }
379
380 void
381 zio_add_child(zio_t *pio, zio_t *cio)
382 {
383 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
384
385 /*
386 * Logical I/Os can have logical, gang, or vdev children.
387 * Gang I/Os can have gang or vdev children.
388 * Vdev I/Os can only have vdev children.
389 * The following ASSERT captures all of these constraints.
390 */
391 ASSERT(cio->io_child_type <= pio->io_child_type);
392
393 zl->zl_parent = pio;
394 zl->zl_child = cio;
395
396 mutex_enter(&cio->io_lock);
397 mutex_enter(&pio->io_lock);
398
399 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
400
401 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
402 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
403
404 list_insert_head(&pio->io_child_list, zl);
405 list_insert_head(&cio->io_parent_list, zl);
406
407 pio->io_child_count++;
408 cio->io_parent_count++;
409
410 mutex_exit(&pio->io_lock);
411 mutex_exit(&cio->io_lock);
412 }
413
414 static void
415 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
416 {
417 ASSERT(zl->zl_parent == pio);
418 ASSERT(zl->zl_child == cio);
419
420 mutex_enter(&cio->io_lock);
421 mutex_enter(&pio->io_lock);
422
423 list_remove(&pio->io_child_list, zl);
424 list_remove(&cio->io_parent_list, zl);
425
426 pio->io_child_count--;
427 cio->io_parent_count--;
428
429 mutex_exit(&pio->io_lock);
430 mutex_exit(&cio->io_lock);
431
432 kmem_cache_free(zio_link_cache, zl);
433 }
434
435 static boolean_t
436 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
437 {
438 uint64_t *countp = &zio->io_children[child][wait];
439 boolean_t waiting = B_FALSE;
440
441 mutex_enter(&zio->io_lock);
442 ASSERT(zio->io_stall == NULL);
443 if (*countp != 0) {
444 zio->io_stage >>= 1;
445 zio->io_stall = countp;
446 waiting = B_TRUE;
447 }
448 mutex_exit(&zio->io_lock);
449
450 return (waiting);
451 }
452
453 static void
454 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
455 {
456 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
457 int *errorp = &pio->io_child_error[zio->io_child_type];
458
459 mutex_enter(&pio->io_lock);
460 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
461 *errorp = zio_worst_error(*errorp, zio->io_error);
462 pio->io_reexecute |= zio->io_reexecute;
463 ASSERT3U(*countp, >, 0);
464
465 (*countp)--;
466
467 if (*countp == 0 && pio->io_stall == countp) {
468 pio->io_stall = NULL;
469 mutex_exit(&pio->io_lock);
470 zio_execute(pio);
471 } else {
472 mutex_exit(&pio->io_lock);
473 }
474 }
475
476 static void
477 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
478 {
479 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
480 zio->io_error = zio->io_child_error[c];
481 }
482
483 /*
484 * ==========================================================================
485 * Create the various types of I/O (read, write, free, etc)
486 * ==========================================================================
487 */
488 static zio_t *
489 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
490 void *data, uint64_t size, zio_done_func_t *done, void *private,
491 zio_type_t type, zio_priority_t priority, enum zio_flag flags,
492 vdev_t *vd, uint64_t offset, const zbookmark_phys_t *zb,
493 enum zio_stage stage, enum zio_stage pipeline)
494 {
495 zio_t *zio;
496
497 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
498 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
499 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
500
501 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
502 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
503 ASSERT(vd || stage == ZIO_STAGE_OPEN);
504
505 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
506 bzero(zio, sizeof (zio_t));
507
508 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
509 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
510
511 list_create(&zio->io_parent_list, sizeof (zio_link_t),
512 offsetof(zio_link_t, zl_parent_node));
513 list_create(&zio->io_child_list, sizeof (zio_link_t),
514 offsetof(zio_link_t, zl_child_node));
515
516 if (vd != NULL)
517 zio->io_child_type = ZIO_CHILD_VDEV;
518 else if (flags & ZIO_FLAG_GANG_CHILD)
519 zio->io_child_type = ZIO_CHILD_GANG;
520 else if (flags & ZIO_FLAG_DDT_CHILD)
521 zio->io_child_type = ZIO_CHILD_DDT;
522 else
523 zio->io_child_type = ZIO_CHILD_LOGICAL;
524
525 if (bp != NULL) {
526 zio->io_bp = (blkptr_t *)bp;
527 zio->io_bp_copy = *bp;
528 zio->io_bp_orig = *bp;
529 if (type != ZIO_TYPE_WRITE ||
530 zio->io_child_type == ZIO_CHILD_DDT)
531 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
532 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
533 zio->io_logical = zio;
534 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
535 pipeline |= ZIO_GANG_STAGES;
536 }
537
538 zio->io_spa = spa;
539 zio->io_txg = txg;
540 zio->io_done = done;
541 zio->io_private = private;
542 zio->io_type = type;
543 zio->io_priority = priority;
544 zio->io_vd = vd;
545 zio->io_offset = offset;
546 zio->io_orig_data = zio->io_data = data;
547 zio->io_orig_size = zio->io_size = size;
548 zio->io_orig_flags = zio->io_flags = flags;
549 zio->io_orig_stage = zio->io_stage = stage;
550 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
551
552 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
553 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
554
555 if (zb != NULL)
556 zio->io_bookmark = *zb;
557
558 if (pio != NULL) {
559 if (zio->io_logical == NULL)
560 zio->io_logical = pio->io_logical;
561 if (zio->io_child_type == ZIO_CHILD_GANG)
562 zio->io_gang_leader = pio->io_gang_leader;
563 zio_add_child(pio, zio);
564 }
565
566 return (zio);
567 }
568
569 static void
570 zio_destroy(zio_t *zio)
571 {
572 list_destroy(&zio->io_parent_list);
573 list_destroy(&zio->io_child_list);
574 mutex_destroy(&zio->io_lock);
575 cv_destroy(&zio->io_cv);
576 kmem_cache_free(zio_cache, zio);
577 }
578
579 zio_t *
580 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
581 void *private, enum zio_flag flags)
582 {
583 zio_t *zio;
584
585 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
586 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
587 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
588
589 return (zio);
590 }
591
592 zio_t *
593 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
594 {
595 return (zio_null(NULL, spa, NULL, done, private, flags));
596 }
597
598 zio_t *
599 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
600 void *data, uint64_t size, zio_done_func_t *done, void *private,
601 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
602 {
603 zio_t *zio;
604
605 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
606 data, size, done, private,
607 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
608 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
609 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
610
611 return (zio);
612 }
613
614 zio_t *
615 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
616 void *data, uint64_t size, const zio_prop_t *zp,
617 zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done,
618 void *private,
619 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
620 {
621 zio_t *zio;
622
623 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
624 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
625 zp->zp_compress >= ZIO_COMPRESS_OFF &&
626 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
627 DMU_OT_IS_VALID(zp->zp_type) &&
628 zp->zp_level < 32 &&
629 zp->zp_copies > 0 &&
630 zp->zp_copies <= spa_max_replication(spa));
631
632 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
633 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
634 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
635 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
636
637 zio->io_ready = ready;
638 zio->io_physdone = physdone;
639 zio->io_prop = *zp;
640
641 /*
642 * Data can be NULL if we are going to call zio_write_override() to
643 * provide the already-allocated BP. But we may need the data to
644 * verify a dedup hit (if requested). In this case, don't try to
645 * dedup (just take the already-allocated BP verbatim).
646 */
647 if (data == NULL && zio->io_prop.zp_dedup_verify) {
648 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
649 }
650
651 return (zio);
652 }
653
654 zio_t *
655 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
656 uint64_t size, zio_done_func_t *done, void *private,
657 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
658 {
659 zio_t *zio;
660
661 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
662 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
663 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
664
665 return (zio);
666 }
667
668 void
669 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
670 {
671 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
672 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
673 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
674 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
675
676 /*
677 * We must reset the io_prop to match the values that existed
678 * when the bp was first written by dmu_sync() keeping in mind
679 * that nopwrite and dedup are mutually exclusive.
680 */
681 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
682 zio->io_prop.zp_nopwrite = nopwrite;
683 zio->io_prop.zp_copies = copies;
684 zio->io_bp_override = bp;
685 }
686
687 void
688 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
689 {
690
691 /*
692 * The check for EMBEDDED is a performance optimization. We
693 * process the free here (by ignoring it) rather than
694 * putting it on the list and then processing it in zio_free_sync().
695 */
696 if (BP_IS_EMBEDDED(bp))
697 return;
698 metaslab_check_free(spa, bp);
699
700 /*
701 * Frees that are for the currently-syncing txg, are not going to be
702 * deferred, and which will not need to do a read (i.e. not GANG or
703 * DEDUP), can be processed immediately. Otherwise, put them on the
704 * in-memory list for later processing.
705 */
706 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
707 txg != spa->spa_syncing_txg ||
708 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
709 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
710 } else {
711 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0)));
712 }
713 }
714
715 zio_t *
716 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
717 enum zio_flag flags)
718 {
719 zio_t *zio;
720 enum zio_stage stage = ZIO_FREE_PIPELINE;
721
722 ASSERT(!BP_IS_HOLE(bp));
723 ASSERT(spa_syncing_txg(spa) == txg);
724 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
725
726 if (BP_IS_EMBEDDED(bp))
727 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
728
729 metaslab_check_free(spa, bp);
730 arc_freed(spa, bp);
731
732 /*
733 * GANG and DEDUP blocks can induce a read (for the gang block header,
734 * or the DDT), so issue them asynchronously so that this thread is
735 * not tied up.
736 */
737 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
738 stage |= ZIO_STAGE_ISSUE_ASYNC;
739
740 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
741 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags,
742 NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
743
744 return (zio);
745 }
746
747 zio_t *
748 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
749 zio_done_func_t *done, void *private, enum zio_flag flags)
750 {
751 zio_t *zio;
752
753 dprintf_bp(bp, "claiming in txg %llu", txg);
754
755 if (BP_IS_EMBEDDED(bp))
756 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
757
758 /*
759 * A claim is an allocation of a specific block. Claims are needed
760 * to support immediate writes in the intent log. The issue is that
761 * immediate writes contain committed data, but in a txg that was
762 * *not* committed. Upon opening the pool after an unclean shutdown,
763 * the intent log claims all blocks that contain immediate write data
764 * so that the SPA knows they're in use.
765 *
766 * All claims *must* be resolved in the first txg -- before the SPA
767 * starts allocating blocks -- so that nothing is allocated twice.
768 * If txg == 0 we just verify that the block is claimable.
769 */
770 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
771 ASSERT(txg == spa_first_txg(spa) || txg == 0);
772 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
773
774 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
775 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
776 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
777
778 return (zio);
779 }
780
781 zio_t *
782 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
783 zio_done_func_t *done, void *private, enum zio_flag flags)
784 {
785 zio_t *zio;
786 int c;
787
788 if (vd->vdev_children == 0) {
789 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
790 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
791 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
792
793 zio->io_cmd = cmd;
794 } else {
795 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
796
797 for (c = 0; c < vd->vdev_children; c++)
798 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
799 done, private, flags));
800 }
801
802 return (zio);
803 }
804
805 zio_t *
806 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
807 void *data, int checksum, zio_done_func_t *done, void *private,
808 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
809 {
810 zio_t *zio;
811
812 ASSERT(vd->vdev_children == 0);
813 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
814 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
815 ASSERT3U(offset + size, <=, vd->vdev_psize);
816
817 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
818 ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
819 NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
820
821 zio->io_prop.zp_checksum = checksum;
822
823 return (zio);
824 }
825
826 zio_t *
827 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
828 void *data, int checksum, zio_done_func_t *done, void *private,
829 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
830 {
831 zio_t *zio;
832
833 ASSERT(vd->vdev_children == 0);
834 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
835 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
836 ASSERT3U(offset + size, <=, vd->vdev_psize);
837
838 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
839 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
840 NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
841
842 zio->io_prop.zp_checksum = checksum;
843
844 if (zio_checksum_table[checksum].ci_eck) {
845 /*
846 * zec checksums are necessarily destructive -- they modify
847 * the end of the write buffer to hold the verifier/checksum.
848 * Therefore, we must make a local copy in case the data is
849 * being written to multiple places in parallel.
850 */
851 void *wbuf = zio_buf_alloc(size);
852 bcopy(data, wbuf, size);
853 zio_push_transform(zio, wbuf, size, size, NULL);
854 }
855
856 return (zio);
857 }
858
859 /*
860 * Create a child I/O to do some work for us.
861 */
862 zio_t *
863 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
864 void *data, uint64_t size, int type, zio_priority_t priority,
865 enum zio_flag flags, zio_done_func_t *done, void *private)
866 {
867 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
868 zio_t *zio;
869
870 ASSERT(vd->vdev_parent ==
871 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
872
873 if (type == ZIO_TYPE_READ && bp != NULL) {
874 /*
875 * If we have the bp, then the child should perform the
876 * checksum and the parent need not. This pushes error
877 * detection as close to the leaves as possible and
878 * eliminates redundant checksums in the interior nodes.
879 */
880 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
881 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
882 }
883
884 if (vd->vdev_children == 0)
885 offset += VDEV_LABEL_START_SIZE;
886
887 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
888
889 /*
890 * If we've decided to do a repair, the write is not speculative --
891 * even if the original read was.
892 */
893 if (flags & ZIO_FLAG_IO_REPAIR)
894 flags &= ~ZIO_FLAG_SPECULATIVE;
895
896 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
897 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
898 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
899
900 zio->io_physdone = pio->io_physdone;
901 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
902 zio->io_logical->io_phys_children++;
903
904 return (zio);
905 }
906
907 zio_t *
908 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
909 int type, zio_priority_t priority, enum zio_flag flags,
910 zio_done_func_t *done, void *private)
911 {
912 zio_t *zio;
913
914 ASSERT(vd->vdev_ops->vdev_op_leaf);
915
916 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
917 data, size, done, private, type, priority,
918 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
919 vd, offset, NULL,
920 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
921
922 return (zio);
923 }
924
925 void
926 zio_flush(zio_t *zio, vdev_t *vd)
927 {
928 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
929 NULL, NULL,
930 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
931 }
932
933 void
934 zio_shrink(zio_t *zio, uint64_t size)
935 {
936 ASSERT(zio->io_executor == NULL);
937 ASSERT(zio->io_orig_size == zio->io_size);
938 ASSERT(size <= zio->io_size);
939
940 /*
941 * We don't shrink for raidz because of problems with the
942 * reconstruction when reading back less than the block size.
943 * Note, BP_IS_RAIDZ() assumes no compression.
944 */
945 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
946 if (!BP_IS_RAIDZ(zio->io_bp))
947 zio->io_orig_size = zio->io_size = size;
948 }
949
950 /*
951 * ==========================================================================
952 * Prepare to read and write logical blocks
953 * ==========================================================================
954 */
955
956 static int
957 zio_read_bp_init(zio_t *zio)
958 {
959 blkptr_t *bp = zio->io_bp;
960
961 if (!BP_IS_EMBEDDED(bp) && BP_GET_PROP_RESERVATION(bp)) {
962 memset(zio->io_orig_data, 0, zio->io_orig_size);
963 zio->io_pipeline = ZIO_INTERLOCK_STAGES;
964 return (ZIO_PIPELINE_CONTINUE);
965 }
966
967 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
968 zio->io_child_type == ZIO_CHILD_LOGICAL &&
969 !(zio->io_flags & ZIO_FLAG_RAW)) {
970 uint64_t psize =
971 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
972 void *cbuf = zio_buf_alloc(psize);
973
974 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
975 }
976
977 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
978 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
979 decode_embedded_bp_compressed(bp, zio->io_data);
980 } else {
981 ASSERT(!BP_IS_EMBEDDED(bp));
982 }
983
984 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
985 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
986
987 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
988 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
989
990 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
991 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
992
993 return (ZIO_PIPELINE_CONTINUE);
994 }
995
996 static int
997 zio_write_bp_init(zio_t *zio)
998 {
999 spa_t *spa = zio->io_spa;
1000 zio_prop_t *zp = &zio->io_prop;
1001 enum zio_compress compress = zp->zp_compress;
1002 enum zio_checksum checksum = zp->zp_checksum;
1003 uint8_t dedup = zp->zp_dedup;
1004 blkptr_t *bp = zio->io_bp;
1005 uint64_t lsize = zio->io_size;
1006 uint64_t psize = lsize;
1007 int pass = 1;
1008
1009 /*
1010 * If our children haven't all reached the ready stage,
1011 * wait for them and then repeat this pipeline stage.
1012 */
1013 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1014 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1015 return (ZIO_PIPELINE_STOP);
1016
1017 if (!IO_IS_ALLOCATING(zio))
1018 return (ZIO_PIPELINE_CONTINUE);
1019
1020 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1021
1022 if (zp->zp_zero_write && !(zio->io_pipeline & ZIO_GANG_STAGES)) {
1023 dedup = B_FALSE;
1024 compress = ZIO_COMPRESS_OFF;
1025 checksum = ZIO_CHECKSUM_OFF;
1026 }
1027
1028 if (zio->io_bp_override) {
1029 ASSERT(bp->blk_birth != zio->io_txg);
1030 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1031
1032 *bp = *zio->io_bp_override;
1033 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1034
1035 if (BP_IS_EMBEDDED(bp))
1036 return (ZIO_PIPELINE_CONTINUE);
1037
1038 /*
1039 * If we've been overridden and nopwrite is set then
1040 * set the flag accordingly to indicate that a nopwrite
1041 * has already occurred.
1042 */
1043 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1044 ASSERT(!zp->zp_dedup);
1045 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1046 return (ZIO_PIPELINE_CONTINUE);
1047 }
1048
1049 ASSERT(!zp->zp_nopwrite);
1050
1051 if (BP_IS_HOLE(bp) || !dedup)
1052 return (ZIO_PIPELINE_CONTINUE);
1053
1054 ASSERT(zio_checksum_table[checksum].ci_dedup ||
1055 zp->zp_dedup_verify);
1056
1057 if (BP_GET_CHECKSUM(bp) == checksum) {
1058 BP_SET_DEDUP(bp, 1);
1059 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1060 return (ZIO_PIPELINE_CONTINUE);
1061 }
1062 zio->io_bp_override = NULL;
1063 BP_ZERO(bp);
1064 }
1065
1066 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1067 /*
1068 * We're rewriting an existing block, which means we're
1069 * working on behalf of spa_sync(). For spa_sync() to
1070 * converge, it must eventually be the case that we don't
1071 * have to allocate new blocks. But compression changes
1072 * the blocksize, which forces a reallocate, and makes
1073 * convergence take longer. Therefore, after the first
1074 * few passes, stop compressing to ensure convergence.
1075 */
1076 pass = spa_sync_pass(spa);
1077
1078 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1079 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1080 ASSERT(!BP_GET_DEDUP(bp));
1081
1082 if (pass >= zfs_sync_pass_dont_compress)
1083 compress = ZIO_COMPRESS_OFF;
1084
1085 /* Make sure someone doesn't change their mind on overwrites */
1086 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1087 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1088 }
1089
1090 if (compress != ZIO_COMPRESS_OFF) {
1091 void *cbuf = zio_buf_alloc(lsize);
1092 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1093 if (psize == 0 || psize == lsize) {
1094 compress = ZIO_COMPRESS_OFF;
1095 zio_buf_free(cbuf, lsize);
1096 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1097 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1098 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1099 encode_embedded_bp_compressed(bp,
1100 cbuf, compress, lsize, psize);
1101 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1102 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1103 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1104 zio_buf_free(cbuf, lsize);
1105 bp->blk_birth = zio->io_txg;
1106 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1107 ASSERT(spa_feature_is_active(spa,
1108 SPA_FEATURE_EMBEDDED_DATA));
1109 return (ZIO_PIPELINE_CONTINUE);
1110 } else {
1111 /*
1112 * Round up compressed size to MINBLOCKSIZE and
1113 * zero the tail.
1114 */
1115 size_t rounded =
1116 P2ROUNDUP(psize, (size_t)SPA_MINBLOCKSIZE);
1117 if (rounded > psize) {
1118 bzero((char *)cbuf + psize, rounded - psize);
1119 psize = rounded;
1120 }
1121 if (psize == lsize) {
1122 compress = ZIO_COMPRESS_OFF;
1123 zio_buf_free(cbuf, lsize);
1124 } else {
1125 zio_push_transform(zio, cbuf,
1126 psize, lsize, NULL);
1127 }
1128 }
1129 }
1130
1131 /*
1132 * The final pass of spa_sync() must be all rewrites, but the first
1133 * few passes offer a trade-off: allocating blocks defers convergence,
1134 * but newly allocated blocks are sequential, so they can be written
1135 * to disk faster. Therefore, we allow the first few passes of
1136 * spa_sync() to allocate new blocks, but force rewrites after that.
1137 * There should only be a handful of blocks after pass 1 in any case.
1138 */
1139 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1140 BP_GET_PSIZE(bp) == psize &&
1141 pass >= zfs_sync_pass_rewrite) {
1142 ASSERT(psize != 0);
1143 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1144 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1145 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1146 } else {
1147 BP_ZERO(bp);
1148 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1149 }
1150
1151 if (psize == 0) {
1152 if (zio->io_bp_orig.blk_birth != 0 &&
1153 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1154 BP_SET_LSIZE(bp, lsize);
1155 BP_SET_TYPE(bp, zp->zp_type);
1156 BP_SET_LEVEL(bp, zp->zp_level);
1157 BP_SET_BIRTH(bp, zio->io_txg, 0);
1158 }
1159 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1160 } else {
1161 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1162 BP_SET_LSIZE(bp, lsize);
1163 BP_SET_TYPE(bp, zp->zp_type);
1164 BP_SET_LEVEL(bp, zp->zp_level);
1165 BP_SET_PSIZE(bp, psize);
1166 BP_SET_COMPRESS(bp, compress);
1167 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1168 BP_SET_DEDUP(bp, zp->zp_dedup);
1169 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1170 if (zp->zp_zero_write && !(zio->io_pipeline & ZIO_GANG_STAGES)) {
1171 boolean_t need_allocate = B_FALSE;
1172 if (zio->io_pipeline & ZIO_STAGE_DVA_ALLOCATE)
1173 need_allocate = B_TRUE;
1174 zio->io_pipeline = ZIO_INTERLOCK_STAGES;
1175 if (need_allocate)
1176 zio->io_pipeline |= ZIO_STAGE_DVA_ALLOCATE;
1177 BP_SET_PROP_RESERVATION(bp, 1);
1178 } else {
1179 BP_SET_PROP_RESERVATION(bp, 0);
1180 }
1181 if (zp->zp_dedup) {
1182 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1183 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1184 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1185 }
1186 if (zp->zp_nopwrite) {
1187 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1188 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1189 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1190 }
1191 }
1192
1193 return (ZIO_PIPELINE_CONTINUE);
1194 }
1195
1196 static int
1197 zio_free_bp_init(zio_t *zio)
1198 {
1199 blkptr_t *bp = zio->io_bp;
1200
1201 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1202 if (BP_GET_DEDUP(bp))
1203 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1204 }
1205
1206 return (ZIO_PIPELINE_CONTINUE);
1207 }
1208
1209 /*
1210 * ==========================================================================
1211 * Execute the I/O pipeline
1212 * ==========================================================================
1213 */
1214
1215 static void
1216 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1217 {
1218 spa_t *spa = zio->io_spa;
1219 zio_type_t t = zio->io_type;
1220 int flags = (cutinline ? TQ_FRONT : 0);
1221
1222 /*
1223 * If we're a config writer or a probe, the normal issue and
1224 * interrupt threads may all be blocked waiting for the config lock.
1225 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1226 */
1227 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1228 t = ZIO_TYPE_NULL;
1229
1230 /*
1231 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1232 */
1233 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1234 t = ZIO_TYPE_NULL;
1235
1236 /*
1237 * If this is a high priority I/O, then use the high priority taskq if
1238 * available.
1239 */
1240 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1241 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1242 q++;
1243
1244 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1245
1246 /*
1247 * NB: We are assuming that the zio can only be dispatched
1248 * to a single taskq at a time. It would be a grievous error
1249 * to dispatch the zio to another taskq at the same time.
1250 */
1251 ASSERT(zio->io_tqent.tqent_next == NULL);
1252 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1253 flags, &zio->io_tqent);
1254 }
1255
1256 static boolean_t
1257 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1258 {
1259 kthread_t *executor = zio->io_executor;
1260 spa_t *spa = zio->io_spa;
1261
1262 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1263 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1264 uint_t i;
1265 for (i = 0; i < tqs->stqs_count; i++) {
1266 if (taskq_member(tqs->stqs_taskq[i], executor))
1267 return (B_TRUE);
1268 }
1269 }
1270
1271 return (B_FALSE);
1272 }
1273
1274 static int
1275 zio_issue_async(zio_t *zio)
1276 {
1277 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1278
1279 return (ZIO_PIPELINE_STOP);
1280 }
1281
1282 void
1283 zio_interrupt(zio_t *zio)
1284 {
1285 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1286 }
1287
1288 /*
1289 * Execute the I/O pipeline until one of the following occurs:
1290 *
1291 * (1) the I/O completes
1292 * (2) the pipeline stalls waiting for dependent child I/Os
1293 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1294 * (4) the I/O is delegated by vdev-level caching or aggregation
1295 * (5) the I/O is deferred due to vdev-level queueing
1296 * (6) the I/O is handed off to another thread.
1297 *
1298 * In all cases, the pipeline stops whenever there's no CPU work; it never
1299 * burns a thread in cv_wait().
1300 *
1301 * There's no locking on io_stage because there's no legitimate way
1302 * for multiple threads to be attempting to process the same I/O.
1303 */
1304 static zio_pipe_stage_t *zio_pipeline[];
1305
1306 void
1307 zio_execute(zio_t *zio)
1308 {
1309 zio->io_executor = curthread;
1310
1311 while (zio->io_stage < ZIO_STAGE_DONE) {
1312 enum zio_stage pipeline = zio->io_pipeline;
1313 enum zio_stage stage = zio->io_stage;
1314 int rv;
1315
1316 ASSERT(!MUTEX_HELD(&zio->io_lock));
1317 ASSERT(ISP2(stage));
1318 ASSERT(zio->io_stall == NULL);
1319
1320 do {
1321 stage <<= 1;
1322 } while ((stage & pipeline) == 0);
1323
1324 ASSERT(stage <= ZIO_STAGE_DONE);
1325
1326 /*
1327 * If we are in interrupt context and this pipeline stage
1328 * will grab a config lock that is held across I/O,
1329 * or may wait for an I/O that needs an interrupt thread
1330 * to complete, issue async to avoid deadlock.
1331 *
1332 * For VDEV_IO_START, we cut in line so that the io will
1333 * be sent to disk promptly.
1334 */
1335 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1336 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1337 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1338 zio_requeue_io_start_cut_in_line : B_FALSE;
1339 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1340 return;
1341 }
1342
1343 zio->io_stage = stage;
1344 rv = zio_pipeline[highbit64(stage) - 1](zio);
1345
1346 if (rv == ZIO_PIPELINE_STOP)
1347 return;
1348
1349 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1350 }
1351 }
1352
1353 /*
1354 * ==========================================================================
1355 * Initiate I/O, either sync or async
1356 * ==========================================================================
1357 */
1358 int
1359 zio_wait(zio_t *zio)
1360 {
1361 int error;
1362
1363 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1364 ASSERT(zio->io_executor == NULL);
1365
1366 zio->io_waiter = curthread;
1367
1368 zio_execute(zio);
1369
1370 mutex_enter(&zio->io_lock);
1371 while (zio->io_executor != NULL)
1372 cv_wait(&zio->io_cv, &zio->io_lock);
1373 mutex_exit(&zio->io_lock);
1374
1375 error = zio->io_error;
1376 zio_destroy(zio);
1377
1378 return (error);
1379 }
1380
1381 void
1382 zio_nowait(zio_t *zio)
1383 {
1384 ASSERT(zio->io_executor == NULL);
1385
1386 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1387 zio_unique_parent(zio) == NULL) {
1388 /*
1389 * This is a logical async I/O with no parent to wait for it.
1390 * We add it to the spa_async_root_zio "Godfather" I/O which
1391 * will ensure they complete prior to unloading the pool.
1392 */
1393 spa_t *spa = zio->io_spa;
1394
1395 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1396 }
1397
1398 zio_execute(zio);
1399 }
1400
1401 /*
1402 * ==========================================================================
1403 * Reexecute or suspend/resume failed I/O
1404 * ==========================================================================
1405 */
1406
1407 static void
1408 zio_reexecute(zio_t *pio)
1409 {
1410 zio_t *cio, *cio_next;
1411
1412 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1413 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1414 ASSERT(pio->io_gang_leader == NULL);
1415 ASSERT(pio->io_gang_tree == NULL);
1416
1417 pio->io_flags = pio->io_orig_flags;
1418 pio->io_stage = pio->io_orig_stage;
1419 pio->io_pipeline = pio->io_orig_pipeline;
1420 pio->io_reexecute = 0;
1421 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1422 pio->io_error = 0;
1423 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1424 pio->io_state[w] = 0;
1425 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1426 pio->io_child_error[c] = 0;
1427
1428 if (IO_IS_ALLOCATING(pio))
1429 BP_ZERO(pio->io_bp);
1430
1431 /*
1432 * As we reexecute pio's children, new children could be created.
1433 * New children go to the head of pio's io_child_list, however,
1434 * so we will (correctly) not reexecute them. The key is that
1435 * the remainder of pio's io_child_list, from 'cio_next' onward,
1436 * cannot be affected by any side effects of reexecuting 'cio'.
1437 */
1438 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1439 cio_next = zio_walk_children(pio);
1440 mutex_enter(&pio->io_lock);
1441 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1442 pio->io_children[cio->io_child_type][w]++;
1443 mutex_exit(&pio->io_lock);
1444 zio_reexecute(cio);
1445 }
1446
1447 /*
1448 * Now that all children have been reexecuted, execute the parent.
1449 * We don't reexecute "The Godfather" I/O here as it's the
1450 * responsibility of the caller to wait on him.
1451 */
1452 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1453 zio_execute(pio);
1454 }
1455
1456 void
1457 zio_suspend(spa_t *spa, zio_t *zio)
1458 {
1459 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1460 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1461 "failure and the failure mode property for this pool "
1462 "is set to panic.", spa_name(spa));
1463
1464 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1465
1466 mutex_enter(&spa->spa_suspend_lock);
1467
1468 if (spa->spa_suspend_zio_root == NULL)
1469 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1470 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1471 ZIO_FLAG_GODFATHER);
1472
1473 spa->spa_suspended = B_TRUE;
1474
1475 if (zio != NULL) {
1476 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1477 ASSERT(zio != spa->spa_suspend_zio_root);
1478 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1479 ASSERT(zio_unique_parent(zio) == NULL);
1480 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1481 zio_add_child(spa->spa_suspend_zio_root, zio);
1482 }
1483
1484 mutex_exit(&spa->spa_suspend_lock);
1485 }
1486
1487 int
1488 zio_resume(spa_t *spa)
1489 {
1490 zio_t *pio;
1491
1492 /*
1493 * Reexecute all previously suspended i/o.
1494 */
1495 mutex_enter(&spa->spa_suspend_lock);
1496 spa->spa_suspended = B_FALSE;
1497 cv_broadcast(&spa->spa_suspend_cv);
1498 pio = spa->spa_suspend_zio_root;
1499 spa->spa_suspend_zio_root = NULL;
1500 mutex_exit(&spa->spa_suspend_lock);
1501
1502 if (pio == NULL)
1503 return (0);
1504
1505 zio_reexecute(pio);
1506 return (zio_wait(pio));
1507 }
1508
1509 void
1510 zio_resume_wait(spa_t *spa)
1511 {
1512 mutex_enter(&spa->spa_suspend_lock);
1513 while (spa_suspended(spa))
1514 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1515 mutex_exit(&spa->spa_suspend_lock);
1516 }
1517
1518 /*
1519 * ==========================================================================
1520 * Gang blocks.
1521 *
1522 * A gang block is a collection of small blocks that looks to the DMU
1523 * like one large block. When zio_dva_allocate() cannot find a block
1524 * of the requested size, due to either severe fragmentation or the pool
1525 * being nearly full, it calls zio_write_gang_block() to construct the
1526 * block from smaller fragments.
1527 *
1528 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1529 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1530 * an indirect block: it's an array of block pointers. It consumes
1531 * only one sector and hence is allocatable regardless of fragmentation.
1532 * The gang header's bps point to its gang members, which hold the data.
1533 *
1534 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1535 * as the verifier to ensure uniqueness of the SHA256 checksum.
1536 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1537 * not the gang header. This ensures that data block signatures (needed for
1538 * deduplication) are independent of how the block is physically stored.
1539 *
1540 * Gang blocks can be nested: a gang member may itself be a gang block.
1541 * Thus every gang block is a tree in which root and all interior nodes are
1542 * gang headers, and the leaves are normal blocks that contain user data.
1543 * The root of the gang tree is called the gang leader.
1544 *
1545 * To perform any operation (read, rewrite, free, claim) on a gang block,
1546 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1547 * in the io_gang_tree field of the original logical i/o by recursively
1548 * reading the gang leader and all gang headers below it. This yields
1549 * an in-core tree containing the contents of every gang header and the
1550 * bps for every constituent of the gang block.
1551 *
1552 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1553 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1554 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1555 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1556 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1557 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1558 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1559 * of the gang header plus zio_checksum_compute() of the data to update the
1560 * gang header's blk_cksum as described above.
1561 *
1562 * The two-phase assemble/issue model solves the problem of partial failure --
1563 * what if you'd freed part of a gang block but then couldn't read the
1564 * gang header for another part? Assembling the entire gang tree first
1565 * ensures that all the necessary gang header I/O has succeeded before
1566 * starting the actual work of free, claim, or write. Once the gang tree
1567 * is assembled, free and claim are in-memory operations that cannot fail.
1568 *
1569 * In the event that a gang write fails, zio_dva_unallocate() walks the
1570 * gang tree to immediately free (i.e. insert back into the space map)
1571 * everything we've allocated. This ensures that we don't get ENOSPC
1572 * errors during repeated suspend/resume cycles due to a flaky device.
1573 *
1574 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1575 * the gang tree, we won't modify the block, so we can safely defer the free
1576 * (knowing that the block is still intact). If we *can* assemble the gang
1577 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1578 * each constituent bp and we can allocate a new block on the next sync pass.
1579 *
1580 * In all cases, the gang tree allows complete recovery from partial failure.
1581 * ==========================================================================
1582 */
1583
1584 static zio_t *
1585 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1586 {
1587 if (gn != NULL)
1588 return (pio);
1589
1590 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1591 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1592 &pio->io_bookmark));
1593 }
1594
1595 zio_t *
1596 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1597 {
1598 zio_t *zio;
1599
1600 if (gn != NULL) {
1601 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1602 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1603 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1604 /*
1605 * As we rewrite each gang header, the pipeline will compute
1606 * a new gang block header checksum for it; but no one will
1607 * compute a new data checksum, so we do that here. The one
1608 * exception is the gang leader: the pipeline already computed
1609 * its data checksum because that stage precedes gang assembly.
1610 * (Presently, nothing actually uses interior data checksums;
1611 * this is just good hygiene.)
1612 */
1613 if (gn != pio->io_gang_leader->io_gang_tree) {
1614 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1615 data, BP_GET_PSIZE(bp));
1616 }
1617 /*
1618 * If we are here to damage data for testing purposes,
1619 * leave the GBH alone so that we can detect the damage.
1620 */
1621 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1622 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1623 } else {
1624 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1625 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1626 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1627 }
1628
1629 return (zio);
1630 }
1631
1632 /* ARGSUSED */
1633 zio_t *
1634 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1635 {
1636 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1637 ZIO_GANG_CHILD_FLAGS(pio)));
1638 }
1639
1640 /* ARGSUSED */
1641 zio_t *
1642 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1643 {
1644 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1645 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1646 }
1647
1648 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1649 NULL,
1650 zio_read_gang,
1651 zio_rewrite_gang,
1652 zio_free_gang,
1653 zio_claim_gang,
1654 NULL
1655 };
1656
1657 static void zio_gang_tree_assemble_done(zio_t *zio);
1658
1659 static zio_gang_node_t *
1660 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1661 {
1662 zio_gang_node_t *gn;
1663
1664 ASSERT(*gnpp == NULL);
1665
1666 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1667 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1668 *gnpp = gn;
1669
1670 return (gn);
1671 }
1672
1673 static void
1674 zio_gang_node_free(zio_gang_node_t **gnpp)
1675 {
1676 zio_gang_node_t *gn = *gnpp;
1677
1678 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1679 ASSERT(gn->gn_child[g] == NULL);
1680
1681 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1682 kmem_free(gn, sizeof (*gn));
1683 *gnpp = NULL;
1684 }
1685
1686 static void
1687 zio_gang_tree_free(zio_gang_node_t **gnpp)
1688 {
1689 zio_gang_node_t *gn = *gnpp;
1690
1691 if (gn == NULL)
1692 return;
1693
1694 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1695 zio_gang_tree_free(&gn->gn_child[g]);
1696
1697 zio_gang_node_free(gnpp);
1698 }
1699
1700 static void
1701 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1702 {
1703 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1704
1705 ASSERT(gio->io_gang_leader == gio);
1706 ASSERT(BP_IS_GANG(bp));
1707
1708 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1709 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1710 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1711 }
1712
1713 static void
1714 zio_gang_tree_assemble_done(zio_t *zio)
1715 {
1716 zio_t *gio = zio->io_gang_leader;
1717 zio_gang_node_t *gn = zio->io_private;
1718 blkptr_t *bp = zio->io_bp;
1719
1720 ASSERT(gio == zio_unique_parent(zio));
1721 ASSERT(zio->io_child_count == 0);
1722
1723 if (zio->io_error)
1724 return;
1725
1726 if (BP_SHOULD_BYTESWAP(bp))
1727 byteswap_uint64_array(zio->io_data, zio->io_size);
1728
1729 ASSERT(zio->io_data == gn->gn_gbh);
1730 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1731 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1732
1733 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1734 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1735 if (!BP_IS_GANG(gbp))
1736 continue;
1737 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1738 }
1739 }
1740
1741 static void
1742 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1743 {
1744 zio_t *gio = pio->io_gang_leader;
1745 zio_t *zio;
1746
1747 ASSERT(BP_IS_GANG(bp) == !!gn);
1748 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1749 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1750
1751 /*
1752 * If you're a gang header, your data is in gn->gn_gbh.
1753 * If you're a gang member, your data is in 'data' and gn == NULL.
1754 */
1755 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1756
1757 if (gn != NULL) {
1758 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1759
1760 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1761 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1762 if (BP_IS_HOLE(gbp))
1763 continue;
1764 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1765 data = (char *)data + BP_GET_PSIZE(gbp);
1766 }
1767 }
1768
1769 if (gn == gio->io_gang_tree)
1770 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1771
1772 if (zio != pio)
1773 zio_nowait(zio);
1774 }
1775
1776 static int
1777 zio_gang_assemble(zio_t *zio)
1778 {
1779 blkptr_t *bp = zio->io_bp;
1780
1781 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1782 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1783
1784 zio->io_gang_leader = zio;
1785
1786 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1787
1788 return (ZIO_PIPELINE_CONTINUE);
1789 }
1790
1791 static int
1792 zio_gang_issue(zio_t *zio)
1793 {
1794 blkptr_t *bp = zio->io_bp;
1795
1796 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1797 return (ZIO_PIPELINE_STOP);
1798
1799 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1800 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1801
1802 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1803 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1804 else
1805 zio_gang_tree_free(&zio->io_gang_tree);
1806
1807 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1808
1809 return (ZIO_PIPELINE_CONTINUE);
1810 }
1811
1812 static void
1813 zio_write_gang_member_ready(zio_t *zio)
1814 {
1815 zio_t *pio = zio_unique_parent(zio);
1816 zio_t *gio = zio->io_gang_leader;
1817 dva_t *cdva = zio->io_bp->blk_dva;
1818 dva_t *pdva = pio->io_bp->blk_dva;
1819 uint64_t asize;
1820
1821 if (BP_IS_HOLE(zio->io_bp))
1822 return;
1823
1824 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1825
1826 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1827 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1828 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1829 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1830 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1831
1832 mutex_enter(&pio->io_lock);
1833 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1834 ASSERT(DVA_GET_GANG(&pdva[d]));
1835 asize = DVA_GET_ASIZE(&pdva[d]);
1836 asize += DVA_GET_ASIZE(&cdva[d]);
1837 DVA_SET_ASIZE(&pdva[d], asize);
1838 }
1839 mutex_exit(&pio->io_lock);
1840 }
1841
1842 static int
1843 zio_write_gang_block(zio_t *pio)
1844 {
1845 spa_t *spa = pio->io_spa;
1846 blkptr_t *bp = pio->io_bp;
1847 zio_t *gio = pio->io_gang_leader;
1848 zio_t *zio;
1849 zio_gang_node_t *gn, **gnpp;
1850 zio_gbh_phys_t *gbh;
1851 uint64_t txg = pio->io_txg;
1852 uint64_t resid = pio->io_size;
1853 uint64_t lsize;
1854 int copies = gio->io_prop.zp_copies;
1855 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1856 zio_prop_t zp;
1857 int error;
1858
1859 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1860 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1861 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1862 if (error) {
1863 pio->io_error = error;
1864 return (ZIO_PIPELINE_CONTINUE);
1865 }
1866
1867 if (pio == gio) {
1868 gnpp = &gio->io_gang_tree;
1869 } else {
1870 gnpp = pio->io_private;
1871 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1872 }
1873
1874 gn = zio_gang_node_alloc(gnpp);
1875 gbh = gn->gn_gbh;
1876 bzero(gbh, SPA_GANGBLOCKSIZE);
1877
1878 /*
1879 * Create the gang header.
1880 */
1881 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1882 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1883
1884 /*
1885 * Create and nowait the gang children.
1886 */
1887 for (int g = 0; resid != 0; resid -= lsize, g++) {
1888 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1889 SPA_MINBLOCKSIZE);
1890 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1891
1892 zp.zp_checksum = gio->io_prop.zp_checksum;
1893 zp.zp_compress = ZIO_COMPRESS_OFF;
1894 zp.zp_type = DMU_OT_NONE;
1895 zp.zp_level = 0;
1896 zp.zp_copies = gio->io_prop.zp_copies;
1897 zp.zp_dedup = B_FALSE;
1898 zp.zp_dedup_verify = B_FALSE;
1899 zp.zp_zero_write = B_FALSE;
1900 zp.zp_nopwrite = B_FALSE;
1901
1902 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1903 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1904 zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g],
1905 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1906 &pio->io_bookmark));
1907 }
1908
1909 /*
1910 * Set pio's pipeline to just wait for zio to finish.
1911 */
1912 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1913
1914 zio_nowait(zio);
1915
1916 return (ZIO_PIPELINE_CONTINUE);
1917 }
1918
1919 /*
1920 * The zio_nop_write stage in the pipeline determines if allocating
1921 * a new bp is necessary. By leveraging a cryptographically secure checksum,
1922 * such as SHA256, we can compare the checksums of the new data and the old
1923 * to determine if allocating a new block is required. The nopwrite
1924 * feature can handle writes in either syncing or open context (i.e. zil
1925 * writes) and as a result is mutually exclusive with dedup.
1926 */
1927 static int
1928 zio_nop_write(zio_t *zio)
1929 {
1930 blkptr_t *bp = zio->io_bp;
1931 blkptr_t *bp_orig = &zio->io_bp_orig;
1932 zio_prop_t *zp = &zio->io_prop;
1933
1934 ASSERT(BP_GET_LEVEL(bp) == 0);
1935 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1936 ASSERT(zp->zp_nopwrite);
1937 ASSERT(!zp->zp_dedup);
1938 ASSERT(zio->io_bp_override == NULL);
1939 ASSERT(IO_IS_ALLOCATING(zio));
1940
1941 /*
1942 * Check to see if the original bp and the new bp have matching
1943 * characteristics (i.e. same checksum, compression algorithms, etc).
1944 * If they don't then just continue with the pipeline which will
1945 * allocate a new bp.
1946 */
1947 if (BP_IS_HOLE(bp_orig) ||
1948 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
1949 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
1950 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
1951 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
1952 zp->zp_copies != BP_GET_NDVAS(bp_orig))
1953 return (ZIO_PIPELINE_CONTINUE);
1954
1955 /*
1956 * If the checksums match then reset the pipeline so that we
1957 * avoid allocating a new bp and issuing any I/O.
1958 */
1959 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
1960 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
1961 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
1962 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
1963 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
1964 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
1965 sizeof (uint64_t)) == 0);
1966
1967 *bp = *bp_orig;
1968 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1969 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1970 }
1971
1972 return (ZIO_PIPELINE_CONTINUE);
1973 }
1974
1975 /*
1976 * ==========================================================================
1977 * Dedup
1978 * ==========================================================================
1979 */
1980 static void
1981 zio_ddt_child_read_done(zio_t *zio)
1982 {
1983 blkptr_t *bp = zio->io_bp;
1984 ddt_entry_t *dde = zio->io_private;
1985 ddt_phys_t *ddp;
1986 zio_t *pio = zio_unique_parent(zio);
1987
1988 mutex_enter(&pio->io_lock);
1989 ddp = ddt_phys_select(dde, bp);
1990 if (zio->io_error == 0)
1991 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1992 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1993 dde->dde_repair_data = zio->io_data;
1994 else
1995 zio_buf_free(zio->io_data, zio->io_size);
1996 mutex_exit(&pio->io_lock);
1997 }
1998
1999 static int
2000 zio_ddt_read_start(zio_t *zio)
2001 {
2002 blkptr_t *bp = zio->io_bp;
2003
2004 ASSERT(BP_GET_DEDUP(bp));
2005 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2006 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2007
2008 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2009 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2010 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2011 ddt_phys_t *ddp = dde->dde_phys;
2012 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2013 blkptr_t blk;
2014
2015 ASSERT(zio->io_vsd == NULL);
2016 zio->io_vsd = dde;
2017
2018 if (ddp_self == NULL)
2019 return (ZIO_PIPELINE_CONTINUE);
2020
2021 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2022 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2023 continue;
2024 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2025 &blk);
2026 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2027 zio_buf_alloc(zio->io_size), zio->io_size,
2028 zio_ddt_child_read_done, dde, zio->io_priority,
2029 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2030 &zio->io_bookmark));
2031 }
2032 return (ZIO_PIPELINE_CONTINUE);
2033 }
2034
2035 zio_nowait(zio_read(zio, zio->io_spa, bp,
2036 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2037 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2038
2039 return (ZIO_PIPELINE_CONTINUE);
2040 }
2041
2042 static int
2043 zio_ddt_read_done(zio_t *zio)
2044 {
2045 blkptr_t *bp = zio->io_bp;
2046
2047 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2048 return (ZIO_PIPELINE_STOP);
2049
2050 ASSERT(BP_GET_DEDUP(bp));
2051 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2052 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2053
2054 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2055 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2056 ddt_entry_t *dde = zio->io_vsd;
2057 if (ddt == NULL) {
2058 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2059 return (ZIO_PIPELINE_CONTINUE);
2060 }
2061 if (dde == NULL) {
2062 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2063 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2064 return (ZIO_PIPELINE_STOP);
2065 }
2066 if (dde->dde_repair_data != NULL) {
2067 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2068 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2069 }
2070 ddt_repair_done(ddt, dde);
2071 zio->io_vsd = NULL;
2072 }
2073
2074 ASSERT(zio->io_vsd == NULL);
2075
2076 return (ZIO_PIPELINE_CONTINUE);
2077 }
2078
2079 static boolean_t
2080 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2081 {
2082 spa_t *spa = zio->io_spa;
2083
2084 /*
2085 * Note: we compare the original data, not the transformed data,
2086 * because when zio->io_bp is an override bp, we will not have
2087 * pushed the I/O transforms. That's an important optimization
2088 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2089 */
2090 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2091 zio_t *lio = dde->dde_lead_zio[p];
2092
2093 if (lio != NULL) {
2094 return (lio->io_orig_size != zio->io_orig_size ||
2095 bcmp(zio->io_orig_data, lio->io_orig_data,
2096 zio->io_orig_size) != 0);
2097 }
2098 }
2099
2100 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2101 ddt_phys_t *ddp = &dde->dde_phys[p];
2102
2103 if (ddp->ddp_phys_birth != 0) {
2104 arc_buf_t *abuf = NULL;
2105 uint32_t aflags = ARC_WAIT;
2106 blkptr_t blk = *zio->io_bp;
2107 int error;
2108
2109 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2110
2111 ddt_exit(ddt);
2112
2113 error = arc_read(NULL, spa, &blk,
2114 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2115 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2116 &aflags, &zio->io_bookmark);
2117
2118 if (error == 0) {
2119 if (arc_buf_size(abuf) != zio->io_orig_size ||
2120 bcmp(abuf->b_data, zio->io_orig_data,
2121 zio->io_orig_size) != 0)
2122 error = SET_ERROR(EEXIST);
2123 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2124 }
2125
2126 ddt_enter(ddt);
2127 return (error != 0);
2128 }
2129 }
2130
2131 return (B_FALSE);
2132 }
2133
2134 static void
2135 zio_ddt_child_write_ready(zio_t *zio)
2136 {
2137 int p = zio->io_prop.zp_copies;
2138 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2139 ddt_entry_t *dde = zio->io_private;
2140 ddt_phys_t *ddp = &dde->dde_phys[p];
2141 zio_t *pio;
2142
2143 if (zio->io_error)
2144 return;
2145
2146 ddt_enter(ddt);
2147
2148 ASSERT(dde->dde_lead_zio[p] == zio);
2149
2150 ddt_phys_fill(ddp, zio->io_bp);
2151
2152 while ((pio = zio_walk_parents(zio)) != NULL)
2153 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2154
2155 ddt_exit(ddt);
2156 }
2157
2158 static void
2159 zio_ddt_child_write_done(zio_t *zio)
2160 {
2161 int p = zio->io_prop.zp_copies;
2162 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2163 ddt_entry_t *dde = zio->io_private;
2164 ddt_phys_t *ddp = &dde->dde_phys[p];
2165
2166 ddt_enter(ddt);
2167
2168 ASSERT(ddp->ddp_refcnt == 0);
2169 ASSERT(dde->dde_lead_zio[p] == zio);
2170 dde->dde_lead_zio[p] = NULL;
2171
2172 if (zio->io_error == 0) {
2173 while (zio_walk_parents(zio) != NULL)
2174 ddt_phys_addref(ddp);
2175 } else {
2176 ddt_phys_clear(ddp);
2177 }
2178
2179 ddt_exit(ddt);
2180 }
2181
2182 static void
2183 zio_ddt_ditto_write_done(zio_t *zio)
2184 {
2185 int p = DDT_PHYS_DITTO;
2186 zio_prop_t *zp = &zio->io_prop;
2187 blkptr_t *bp = zio->io_bp;
2188 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2189 ddt_entry_t *dde = zio->io_private;
2190 ddt_phys_t *ddp = &dde->dde_phys[p];
2191 ddt_key_t *ddk = &dde->dde_key;
2192
2193 ddt_enter(ddt);
2194
2195 ASSERT(ddp->ddp_refcnt == 0);
2196 ASSERT(dde->dde_lead_zio[p] == zio);
2197 dde->dde_lead_zio[p] = NULL;
2198
2199 if (zio->io_error == 0) {
2200 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2201 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2202 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2203 if (ddp->ddp_phys_birth != 0)
2204 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2205 ddt_phys_fill(ddp, bp);
2206 }
2207
2208 ddt_exit(ddt);
2209 }
2210
2211 static int
2212 zio_ddt_write(zio_t *zio)
2213 {
2214 spa_t *spa = zio->io_spa;
2215 blkptr_t *bp = zio->io_bp;
2216 uint64_t txg = zio->io_txg;
2217 zio_prop_t *zp = &zio->io_prop;
2218 int p = zp->zp_copies;
2219 int ditto_copies;
2220 zio_t *cio = NULL;
2221 zio_t *dio = NULL;
2222 ddt_t *ddt = ddt_select(spa, bp);
2223 ddt_entry_t *dde;
2224 ddt_phys_t *ddp;
2225
2226 ASSERT(BP_GET_DEDUP(bp));
2227 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2228 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2229
2230 ddt_enter(ddt);
2231 dde = ddt_lookup(ddt, bp, B_TRUE);
2232 ddp = &dde->dde_phys[p];
2233
2234 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2235 /*
2236 * If we're using a weak checksum, upgrade to a strong checksum
2237 * and try again. If we're already using a strong checksum,
2238 * we can't resolve it, so just convert to an ordinary write.
2239 * (And automatically e-mail a paper to Nature?)
2240 */
2241 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2242 zp->zp_checksum = spa_dedup_checksum(spa);
2243 zio_pop_transforms(zio);
2244 zio->io_stage = ZIO_STAGE_OPEN;
2245 BP_ZERO(bp);
2246 } else {
2247 zp->zp_dedup = B_FALSE;
2248 }
2249 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2250 ddt_exit(ddt);
2251 return (ZIO_PIPELINE_CONTINUE);
2252 }
2253
2254 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2255 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2256
2257 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2258 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2259 zio_prop_t czp = *zp;
2260
2261 czp.zp_copies = ditto_copies;
2262
2263 /*
2264 * If we arrived here with an override bp, we won't have run
2265 * the transform stack, so we won't have the data we need to
2266 * generate a child i/o. So, toss the override bp and restart.
2267 * This is safe, because using the override bp is just an
2268 * optimization; and it's rare, so the cost doesn't matter.
2269 */
2270 if (zio->io_bp_override) {
2271 zio_pop_transforms(zio);
2272 zio->io_stage = ZIO_STAGE_OPEN;
2273 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2274 zio->io_bp_override = NULL;
2275 BP_ZERO(bp);
2276 ddt_exit(ddt);
2277 return (ZIO_PIPELINE_CONTINUE);
2278 }
2279
2280 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2281 zio->io_orig_size, &czp, NULL, NULL,
2282 zio_ddt_ditto_write_done, dde, zio->io_priority,
2283 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2284
2285 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2286 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2287 }
2288
2289 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2290 if (ddp->ddp_phys_birth != 0)
2291 ddt_bp_fill(ddp, bp, txg);
2292 if (dde->dde_lead_zio[p] != NULL)
2293 zio_add_child(zio, dde->dde_lead_zio[p]);
2294 else
2295 ddt_phys_addref(ddp);
2296 } else if (zio->io_bp_override) {
2297 ASSERT(bp->blk_birth == txg);
2298 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2299 ddt_phys_fill(ddp, bp);
2300 ddt_phys_addref(ddp);
2301 } else {
2302 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2303 zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL,
2304 zio_ddt_child_write_done, dde, zio->io_priority,
2305 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2306
2307 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2308 dde->dde_lead_zio[p] = cio;
2309 }
2310
2311 ddt_exit(ddt);
2312
2313 if (cio)
2314 zio_nowait(cio);
2315 if (dio)
2316 zio_nowait(dio);
2317
2318 return (ZIO_PIPELINE_CONTINUE);
2319 }
2320
2321 ddt_entry_t *freedde; /* for debugging */
2322
2323 static int
2324 zio_ddt_free(zio_t *zio)
2325 {
2326 spa_t *spa = zio->io_spa;
2327 blkptr_t *bp = zio->io_bp;
2328 ddt_t *ddt = ddt_select(spa, bp);
2329 ddt_entry_t *dde;
2330 ddt_phys_t *ddp;
2331
2332 ASSERT(BP_GET_DEDUP(bp));
2333 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2334
2335 ddt_enter(ddt);
2336 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2337 ddp = ddt_phys_select(dde, bp);
2338 ddt_phys_decref(ddp);
2339 ddt_exit(ddt);
2340
2341 return (ZIO_PIPELINE_CONTINUE);
2342 }
2343
2344 /*
2345 * ==========================================================================
2346 * Allocate and free blocks
2347 * ==========================================================================
2348 */
2349 static int
2350 zio_dva_allocate(zio_t *zio)
2351 {
2352 spa_t *spa = zio->io_spa;
2353 metaslab_class_t *mc = spa_normal_class(spa);
2354 blkptr_t *bp = zio->io_bp;
2355 int error;
2356 int flags = 0;
2357
2358 if (zio->io_gang_leader == NULL) {
2359 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2360 zio->io_gang_leader = zio;
2361 }
2362
2363 ASSERT(BP_IS_HOLE(bp));
2364 ASSERT0(BP_GET_NDVAS(bp));
2365 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2366 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2367 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2368
2369 /*
2370 * The dump device does not support gang blocks so allocation on
2371 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2372 * the "fast" gang feature.
2373 */
2374 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2375 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2376 METASLAB_GANG_CHILD : 0;
2377 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2378 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2379
2380 if (error) {
2381 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2382 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2383 error);
2384 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2385 return (zio_write_gang_block(zio));
2386 zio->io_error = error;
2387 }
2388
2389 return (ZIO_PIPELINE_CONTINUE);
2390 }
2391
2392 static int
2393 zio_dva_free(zio_t *zio)
2394 {
2395 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2396
2397 return (ZIO_PIPELINE_CONTINUE);
2398 }
2399
2400 static int
2401 zio_dva_claim(zio_t *zio)
2402 {
2403 int error;
2404
2405 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2406 if (error)
2407 zio->io_error = error;
2408
2409 return (ZIO_PIPELINE_CONTINUE);
2410 }
2411
2412 /*
2413 * Undo an allocation. This is used by zio_done() when an I/O fails
2414 * and we want to give back the block we just allocated.
2415 * This handles both normal blocks and gang blocks.
2416 */
2417 static void
2418 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2419 {
2420 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2421 ASSERT(zio->io_bp_override == NULL);
2422
2423 if (!BP_IS_HOLE(bp))
2424 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2425
2426 if (gn != NULL) {
2427 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2428 zio_dva_unallocate(zio, gn->gn_child[g],
2429 &gn->gn_gbh->zg_blkptr[g]);
2430 }
2431 }
2432 }
2433
2434 /*
2435 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2436 */
2437 int
2438 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2439 uint64_t size, boolean_t use_slog)
2440 {
2441 int error = 1;
2442
2443 ASSERT(txg > spa_syncing_txg(spa));
2444
2445 /*
2446 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2447 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2448 * when allocating them.
2449 */
2450 if (use_slog) {
2451 error = metaslab_alloc(spa, spa_log_class(spa), size,
2452 new_bp, 1, txg, old_bp,
2453 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2454 }
2455
2456 if (error) {
2457 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2458 new_bp, 1, txg, old_bp,
2459 METASLAB_HINTBP_AVOID);
2460 }
2461
2462 if (error == 0) {
2463 BP_SET_LSIZE(new_bp, size);
2464 BP_SET_PSIZE(new_bp, size);
2465 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2466 BP_SET_CHECKSUM(new_bp,
2467 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2468 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2469 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2470 BP_SET_LEVEL(new_bp, 0);
2471 BP_SET_DEDUP(new_bp, 0);
2472 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2473 }
2474
2475 return (error);
2476 }
2477
2478 /*
2479 * Free an intent log block.
2480 */
2481 void
2482 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2483 {
2484 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2485 ASSERT(!BP_IS_GANG(bp));
2486
2487 zio_free(spa, txg, bp);
2488 }
2489
2490 /*
2491 * ==========================================================================
2492 * Read and write to physical devices
2493 * ==========================================================================
2494 */
2495 static int
2496 zio_vdev_io_start(zio_t *zio)
2497 {
2498 vdev_t *vd = zio->io_vd;
2499 uint64_t align;
2500 spa_t *spa = zio->io_spa;
2501
2502 ASSERT(zio->io_error == 0);
2503 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2504
2505 if (vd == NULL) {
2506 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2507 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2508
2509 /*
2510 * The mirror_ops handle multiple DVAs in a single BP.
2511 */
2512 return (vdev_mirror_ops.vdev_op_io_start(zio));
2513 }
2514
2515 /*
2516 * We keep track of time-sensitive I/Os so that the scan thread
2517 * can quickly react to certain workloads. In particular, we care
2518 * about non-scrubbing, top-level reads and writes with the following
2519 * characteristics:
2520 * - synchronous writes of user data to non-slog devices
2521 * - any reads of user data
2522 * When these conditions are met, adjust the timestamp of spa_last_io
2523 * which allows the scan thread to adjust its workload accordingly.
2524 */
2525 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2526 vd == vd->vdev_top && !vd->vdev_islog &&
2527 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2528 zio->io_txg != spa_syncing_txg(spa)) {
2529 uint64_t old = spa->spa_last_io;
2530 uint64_t new = ddi_get_lbolt64();
2531 if (old != new)
2532 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2533 }
2534
2535 align = 1ULL << vd->vdev_top->vdev_ashift;
2536
2537 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
2538 P2PHASE(zio->io_size, align) != 0) {
2539 /* Transform logical writes to be a full physical block size. */
2540 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2541 char *abuf = zio_buf_alloc(asize);
2542 ASSERT(vd == vd->vdev_top);
2543 if (zio->io_type == ZIO_TYPE_WRITE) {
2544 bcopy(zio->io_data, abuf, zio->io_size);
2545 bzero(abuf + zio->io_size, asize - zio->io_size);
2546 }
2547 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2548 }
2549
2550 /*
2551 * If this is not a physical io, make sure that it is properly aligned
2552 * before proceeding.
2553 */
2554 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
2555 ASSERT0(P2PHASE(zio->io_offset, align));
2556 ASSERT0(P2PHASE(zio->io_size, align));
2557 } else {
2558 /*
2559 * For physical writes, we allow 512b aligned writes and assume
2560 * the device will perform a read-modify-write as necessary.
2561 */
2562 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
2563 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
2564 }
2565
2566 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2567
2568 /*
2569 * If this is a repair I/O, and there's no self-healing involved --
2570 * that is, we're just resilvering what we expect to resilver --
2571 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2572 * This prevents spurious resilvering with nested replication.
2573 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2574 * A is out of date, we'll read from C+D, then use the data to
2575 * resilver A+B -- but we don't actually want to resilver B, just A.
2576 * The top-level mirror has no way to know this, so instead we just
2577 * discard unnecessary repairs as we work our way down the vdev tree.
2578 * The same logic applies to any form of nested replication:
2579 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2580 */
2581 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2582 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2583 zio->io_txg != 0 && /* not a delegated i/o */
2584 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2585 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2586 zio_vdev_io_bypass(zio);
2587 return (ZIO_PIPELINE_CONTINUE);
2588 }
2589
2590 if (vd->vdev_ops->vdev_op_leaf &&
2591 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2592
2593 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio))
2594 return (ZIO_PIPELINE_CONTINUE);
2595
2596 if ((zio = vdev_queue_io(zio)) == NULL)
2597 return (ZIO_PIPELINE_STOP);
2598
2599 if (!vdev_accessible(vd, zio)) {
2600 zio->io_error = SET_ERROR(ENXIO);
2601 zio_interrupt(zio);
2602 return (ZIO_PIPELINE_STOP);
2603 }
2604 }
2605
2606 return (vd->vdev_ops->vdev_op_io_start(zio));
2607 }
2608
2609 static int
2610 zio_vdev_io_done(zio_t *zio)
2611 {
2612 vdev_t *vd = zio->io_vd;
2613 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2614 boolean_t unexpected_error = B_FALSE;
2615
2616 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2617 return (ZIO_PIPELINE_STOP);
2618
2619 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2620
2621 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2622
2623 vdev_queue_io_done(zio);
2624
2625 if (zio->io_type == ZIO_TYPE_WRITE)
2626 vdev_cache_write(zio);
2627
2628 if (zio_injection_enabled && zio->io_error == 0)
2629 zio->io_error = zio_handle_device_injection(vd,
2630 zio, EIO);
2631
2632 if (zio_injection_enabled && zio->io_error == 0)
2633 zio->io_error = zio_handle_label_injection(zio, EIO);
2634
2635 if (zio->io_error) {
2636 if (!vdev_accessible(vd, zio)) {
2637 zio->io_error = SET_ERROR(ENXIO);
2638 } else {
2639 unexpected_error = B_TRUE;
2640 }
2641 }
2642 }
2643
2644 ops->vdev_op_io_done(zio);
2645
2646 if (unexpected_error)
2647 VERIFY(vdev_probe(vd, zio) == NULL);
2648
2649 return (ZIO_PIPELINE_CONTINUE);
2650 }
2651
2652 /*
2653 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2654 * disk, and use that to finish the checksum ereport later.
2655 */
2656 static void
2657 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2658 const void *good_buf)
2659 {
2660 /* no processing needed */
2661 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2662 }
2663
2664 /*ARGSUSED*/
2665 void
2666 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2667 {
2668 void *buf = zio_buf_alloc(zio->io_size);
2669
2670 bcopy(zio->io_data, buf, zio->io_size);
2671
2672 zcr->zcr_cbinfo = zio->io_size;
2673 zcr->zcr_cbdata = buf;
2674 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2675 zcr->zcr_free = zio_buf_free;
2676 }
2677
2678 static int
2679 zio_vdev_io_assess(zio_t *zio)
2680 {
2681 vdev_t *vd = zio->io_vd;
2682
2683 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2684 return (ZIO_PIPELINE_STOP);
2685
2686 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2687 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2688
2689 if (zio->io_vsd != NULL) {
2690 zio->io_vsd_ops->vsd_free(zio);
2691 zio->io_vsd = NULL;
2692 }
2693
2694 if (zio_injection_enabled && zio->io_error == 0)
2695 zio->io_error = zio_handle_fault_injection(zio, EIO);
2696
2697 /*
2698 * If the I/O failed, determine whether we should attempt to retry it.
2699 *
2700 * On retry, we cut in line in the issue queue, since we don't want
2701 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2702 */
2703 if (zio->io_error && vd == NULL &&
2704 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2705 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2706 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2707 zio->io_error = 0;
2708 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2709 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2710 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2711 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2712 zio_requeue_io_start_cut_in_line);
2713 return (ZIO_PIPELINE_STOP);
2714 }
2715
2716 /*
2717 * If we got an error on a leaf device, convert it to ENXIO
2718 * if the device is not accessible at all.
2719 */
2720 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2721 !vdev_accessible(vd, zio))
2722 zio->io_error = SET_ERROR(ENXIO);
2723
2724 /*
2725 * If we can't write to an interior vdev (mirror or RAID-Z),
2726 * set vdev_cant_write so that we stop trying to allocate from it.
2727 */
2728 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2729 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2730 vd->vdev_cant_write = B_TRUE;
2731 }
2732
2733 if (zio->io_error)
2734 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2735
2736 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2737 zio->io_physdone != NULL) {
2738 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
2739 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
2740 zio->io_physdone(zio->io_logical);
2741 }
2742
2743 return (ZIO_PIPELINE_CONTINUE);
2744 }
2745
2746 void
2747 zio_vdev_io_reissue(zio_t *zio)
2748 {
2749 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2750 ASSERT(zio->io_error == 0);
2751
2752 zio->io_stage >>= 1;
2753 }
2754
2755 void
2756 zio_vdev_io_redone(zio_t *zio)
2757 {
2758 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2759
2760 zio->io_stage >>= 1;
2761 }
2762
2763 void
2764 zio_vdev_io_bypass(zio_t *zio)
2765 {
2766 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2767 ASSERT(zio->io_error == 0);
2768
2769 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2770 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2771 }
2772
2773 /*
2774 * ==========================================================================
2775 * Generate and verify checksums
2776 * ==========================================================================
2777 */
2778 static int
2779 zio_checksum_generate(zio_t *zio)
2780 {
2781 blkptr_t *bp = zio->io_bp;
2782 enum zio_checksum checksum;
2783
2784 if (bp == NULL) {
2785 /*
2786 * This is zio_write_phys().
2787 * We're either generating a label checksum, or none at all.
2788 */
2789 checksum = zio->io_prop.zp_checksum;
2790
2791 if (checksum == ZIO_CHECKSUM_OFF)
2792 return (ZIO_PIPELINE_CONTINUE);
2793
2794 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2795 } else {
2796 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2797 ASSERT(!IO_IS_ALLOCATING(zio));
2798 checksum = ZIO_CHECKSUM_GANG_HEADER;
2799 } else {
2800 checksum = BP_GET_CHECKSUM(bp);
2801 }
2802 }
2803
2804 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2805
2806 return (ZIO_PIPELINE_CONTINUE);
2807 }
2808
2809 static int
2810 zio_checksum_verify(zio_t *zio)
2811 {
2812 zio_bad_cksum_t info;
2813 blkptr_t *bp = zio->io_bp;
2814 int error;
2815
2816 ASSERT(zio->io_vd != NULL);
2817
2818 if (bp == NULL) {
2819 /*
2820 * This is zio_read_phys().
2821 * We're either verifying a label checksum, or nothing at all.
2822 */
2823 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2824 return (ZIO_PIPELINE_CONTINUE);
2825
2826 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2827 }
2828
2829 if ((error = zio_checksum_error(zio, &info)) != 0) {
2830 zio->io_error = error;
2831 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2832 zfs_ereport_start_checksum(zio->io_spa,
2833 zio->io_vd, zio, zio->io_offset,
2834 zio->io_size, NULL, &info);
2835 }
2836 }
2837
2838 return (ZIO_PIPELINE_CONTINUE);
2839 }
2840
2841 /*
2842 * Called by RAID-Z to ensure we don't compute the checksum twice.
2843 */
2844 void
2845 zio_checksum_verified(zio_t *zio)
2846 {
2847 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2848 }
2849
2850 /*
2851 * ==========================================================================
2852 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2853 * An error of 0 indicates success. ENXIO indicates whole-device failure,
2854 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2855 * indicate errors that are specific to one I/O, and most likely permanent.
2856 * Any other error is presumed to be worse because we weren't expecting it.
2857 * ==========================================================================
2858 */
2859 int
2860 zio_worst_error(int e1, int e2)
2861 {
2862 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2863 int r1, r2;
2864
2865 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2866 if (e1 == zio_error_rank[r1])
2867 break;
2868
2869 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2870 if (e2 == zio_error_rank[r2])
2871 break;
2872
2873 return (r1 > r2 ? e1 : e2);
2874 }
2875
2876 /*
2877 * ==========================================================================
2878 * I/O completion
2879 * ==========================================================================
2880 */
2881 static int
2882 zio_ready(zio_t *zio)
2883 {
2884 blkptr_t *bp = zio->io_bp;
2885 zio_t *pio, *pio_next;
2886
2887 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2888 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2889 return (ZIO_PIPELINE_STOP);
2890
2891 if (zio->io_ready) {
2892 ASSERT(IO_IS_ALLOCATING(zio));
2893 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
2894 (zio->io_flags & ZIO_FLAG_NOPWRITE));
2895 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2896
2897 zio->io_ready(zio);
2898 }
2899
2900 if (bp != NULL && bp != &zio->io_bp_copy)
2901 zio->io_bp_copy = *bp;
2902
2903 if (zio->io_error)
2904 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2905
2906 mutex_enter(&zio->io_lock);
2907 zio->io_state[ZIO_WAIT_READY] = 1;
2908 pio = zio_walk_parents(zio);
2909 mutex_exit(&zio->io_lock);
2910
2911 /*
2912 * As we notify zio's parents, new parents could be added.
2913 * New parents go to the head of zio's io_parent_list, however,
2914 * so we will (correctly) not notify them. The remainder of zio's
2915 * io_parent_list, from 'pio_next' onward, cannot change because
2916 * all parents must wait for us to be done before they can be done.
2917 */
2918 for (; pio != NULL; pio = pio_next) {
2919 pio_next = zio_walk_parents(zio);
2920 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2921 }
2922
2923 if (zio->io_flags & ZIO_FLAG_NODATA) {
2924 if (BP_IS_GANG(bp)) {
2925 zio->io_flags &= ~ZIO_FLAG_NODATA;
2926 } else {
2927 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2928 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2929 }
2930 }
2931
2932 if (zio_injection_enabled &&
2933 zio->io_spa->spa_syncing_txg == zio->io_txg)
2934 zio_handle_ignored_writes(zio);
2935
2936 return (ZIO_PIPELINE_CONTINUE);
2937 }
2938
2939 static int
2940 zio_done(zio_t *zio)
2941 {
2942 spa_t *spa = zio->io_spa;
2943 zio_t *lio = zio->io_logical;
2944 blkptr_t *bp = zio->io_bp;
2945 vdev_t *vd = zio->io_vd;
2946 uint64_t psize = zio->io_size;
2947 zio_t *pio, *pio_next;
2948
2949 /*
2950 * If our children haven't all completed,
2951 * wait for them and then repeat this pipeline stage.
2952 */
2953 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2954 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2955 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2956 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2957 return (ZIO_PIPELINE_STOP);
2958
2959 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2960 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2961 ASSERT(zio->io_children[c][w] == 0);
2962
2963 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
2964 ASSERT(bp->blk_pad[0] == 0);
2965 ASSERT(bp->blk_pad[1] == 0);
2966 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2967 (bp == zio_unique_parent(zio)->io_bp));
2968 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2969 zio->io_bp_override == NULL &&
2970 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2971 ASSERT(!BP_SHOULD_BYTESWAP(bp));
2972 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2973 ASSERT(BP_COUNT_GANG(bp) == 0 ||
2974 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2975 }
2976 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
2977 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
2978 }
2979
2980 /*
2981 * If there were child vdev/gang/ddt errors, they apply to us now.
2982 */
2983 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2984 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2985 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2986
2987 /*
2988 * If the I/O on the transformed data was successful, generate any
2989 * checksum reports now while we still have the transformed data.
2990 */
2991 if (zio->io_error == 0) {
2992 while (zio->io_cksum_report != NULL) {
2993 zio_cksum_report_t *zcr = zio->io_cksum_report;
2994 uint64_t align = zcr->zcr_align;
2995 uint64_t asize = P2ROUNDUP(psize, align);
2996 char *abuf = zio->io_data;
2997
2998 if (asize != psize) {
2999 abuf = zio_buf_alloc(asize);
3000 bcopy(zio->io_data, abuf, psize);
3001 bzero(abuf + psize, asize - psize);
3002 }
3003
3004 zio->io_cksum_report = zcr->zcr_next;
3005 zcr->zcr_next = NULL;
3006 zcr->zcr_finish(zcr, abuf);
3007 zfs_ereport_free_checksum(zcr);
3008
3009 if (asize != psize)
3010 zio_buf_free(abuf, asize);
3011 }
3012 }
3013
3014 zio_pop_transforms(zio); /* note: may set zio->io_error */
3015
3016 vdev_stat_update(zio, psize);
3017
3018 if (zio->io_error) {
3019 /*
3020 * If this I/O is attached to a particular vdev,
3021 * generate an error message describing the I/O failure
3022 * at the block level. We ignore these errors if the
3023 * device is currently unavailable.
3024 */
3025 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3026 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3027
3028 if ((zio->io_error == EIO || !(zio->io_flags &
3029 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3030 zio == lio) {
3031 /*
3032 * For logical I/O requests, tell the SPA to log the
3033 * error and generate a logical data ereport.
3034 */
3035 spa_log_error(spa, zio);
3036 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3037 0, 0);
3038 }
3039 }
3040
3041 if (zio->io_error && zio == lio) {
3042 /*
3043 * Determine whether zio should be reexecuted. This will
3044 * propagate all the way to the root via zio_notify_parent().
3045 */
3046 ASSERT(vd == NULL && bp != NULL);
3047 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3048
3049 if (IO_IS_ALLOCATING(zio) &&
3050 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3051 if (zio->io_error != ENOSPC)
3052 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3053 else
3054 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3055 }
3056
3057 if ((zio->io_type == ZIO_TYPE_READ ||
3058 zio->io_type == ZIO_TYPE_FREE) &&
3059 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3060 zio->io_error == ENXIO &&
3061 spa_load_state(spa) == SPA_LOAD_NONE &&
3062 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3063 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3064
3065 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3066 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3067
3068 /*
3069 * Here is a possibly good place to attempt to do
3070 * either combinatorial reconstruction or error correction
3071 * based on checksums. It also might be a good place
3072 * to send out preliminary ereports before we suspend
3073 * processing.
3074 */
3075 }
3076
3077 /*
3078 * If there were logical child errors, they apply to us now.
3079 * We defer this until now to avoid conflating logical child
3080 * errors with errors that happened to the zio itself when
3081 * updating vdev stats and reporting FMA events above.
3082 */
3083 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3084
3085 if ((zio->io_error || zio->io_reexecute) &&
3086 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3087 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3088 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3089
3090 zio_gang_tree_free(&zio->io_gang_tree);
3091
3092 /*
3093 * Godfather I/Os should never suspend.
3094 */
3095 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3096 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3097 zio->io_reexecute = 0;
3098
3099 if (zio->io_reexecute) {
3100 /*
3101 * This is a logical I/O that wants to reexecute.
3102 *
3103 * Reexecute is top-down. When an i/o fails, if it's not
3104 * the root, it simply notifies its parent and sticks around.
3105 * The parent, seeing that it still has children in zio_done(),
3106 * does the same. This percolates all the way up to the root.
3107 * The root i/o will reexecute or suspend the entire tree.
3108 *
3109 * This approach ensures that zio_reexecute() honors
3110 * all the original i/o dependency relationships, e.g.
3111 * parents not executing until children are ready.
3112 */
3113 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3114
3115 zio->io_gang_leader = NULL;
3116
3117 mutex_enter(&zio->io_lock);
3118 zio->io_state[ZIO_WAIT_DONE] = 1;
3119 mutex_exit(&zio->io_lock);
3120
3121 /*
3122 * "The Godfather" I/O monitors its children but is
3123 * not a true parent to them. It will track them through
3124 * the pipeline but severs its ties whenever they get into
3125 * trouble (e.g. suspended). This allows "The Godfather"
3126 * I/O to return status without blocking.
3127 */
3128 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3129 zio_link_t *zl = zio->io_walk_link;
3130 pio_next = zio_walk_parents(zio);
3131
3132 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3133 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3134 zio_remove_child(pio, zio, zl);
3135 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3136 }
3137 }
3138
3139 if ((pio = zio_unique_parent(zio)) != NULL) {
3140 /*
3141 * We're not a root i/o, so there's nothing to do
3142 * but notify our parent. Don't propagate errors
3143 * upward since we haven't permanently failed yet.
3144 */
3145 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3146 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3147 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3148 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3149 /*
3150 * We'd fail again if we reexecuted now, so suspend
3151 * until conditions improve (e.g. device comes online).
3152 */
3153 zio_suspend(spa, zio);
3154 } else {
3155 /*
3156 * Reexecution is potentially a huge amount of work.
3157 * Hand it off to the otherwise-unused claim taskq.
3158 */
3159 ASSERT(zio->io_tqent.tqent_next == NULL);
3160 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3161 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3162 0, &zio->io_tqent);
3163 }
3164 return (ZIO_PIPELINE_STOP);
3165 }
3166
3167 ASSERT(zio->io_child_count == 0);
3168 ASSERT(zio->io_reexecute == 0);
3169 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3170
3171 /*
3172 * Report any checksum errors, since the I/O is complete.
3173 */
3174 while (zio->io_cksum_report != NULL) {
3175 zio_cksum_report_t *zcr = zio->io_cksum_report;
3176 zio->io_cksum_report = zcr->zcr_next;
3177 zcr->zcr_next = NULL;
3178 zcr->zcr_finish(zcr, NULL);
3179 zfs_ereport_free_checksum(zcr);
3180 }
3181
3182 /*
3183 * It is the responsibility of the done callback to ensure that this
3184 * particular zio is no longer discoverable for adoption, and as
3185 * such, cannot acquire any new parents.
3186 */
3187 if (zio->io_done)
3188 zio->io_done(zio);
3189
3190 mutex_enter(&zio->io_lock);
3191 zio->io_state[ZIO_WAIT_DONE] = 1;
3192 mutex_exit(&zio->io_lock);
3193
3194 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3195 zio_link_t *zl = zio->io_walk_link;
3196 pio_next = zio_walk_parents(zio);
3197 zio_remove_child(pio, zio, zl);
3198 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3199 }
3200
3201 if (zio->io_waiter != NULL) {
3202 mutex_enter(&zio->io_lock);
3203 zio->io_executor = NULL;
3204 cv_broadcast(&zio->io_cv);
3205 mutex_exit(&zio->io_lock);
3206 } else {
3207 zio_destroy(zio);
3208 }
3209
3210 return (ZIO_PIPELINE_STOP);
3211 }
3212
3213 /*
3214 * ==========================================================================
3215 * I/O pipeline definition
3216 * ==========================================================================
3217 */
3218 static zio_pipe_stage_t *zio_pipeline[] = {
3219 NULL,
3220 zio_read_bp_init,
3221 zio_free_bp_init,
3222 zio_issue_async,
3223 zio_write_bp_init,
3224 zio_checksum_generate,
3225 zio_nop_write,
3226 zio_ddt_read_start,
3227 zio_ddt_read_done,
3228 zio_ddt_write,
3229 zio_ddt_free,
3230 zio_gang_assemble,
3231 zio_gang_issue,
3232 zio_dva_allocate,
3233 zio_dva_free,
3234 zio_dva_claim,
3235 zio_ready,
3236 zio_vdev_io_start,
3237 zio_vdev_io_done,
3238 zio_vdev_io_assess,
3239 zio_checksum_verify,
3240 zio_done
3241 };
3242
3243 /* dnp is the dnode for zb1->zb_object */
3244 boolean_t
3245 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_phys_t *zb1,
3246 const zbookmark_phys_t *zb2)
3247 {
3248 uint64_t zb1nextL0, zb2thisobj;
3249
3250 ASSERT(zb1->zb_objset == zb2->zb_objset);
3251 ASSERT(zb2->zb_level == 0);
3252
3253 /* The objset_phys_t isn't before anything. */
3254 if (dnp == NULL)
3255 return (B_FALSE);
3256
3257 zb1nextL0 = (zb1->zb_blkid + 1) <<
3258 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3259
3260 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3261 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3262
3263 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3264 uint64_t nextobj = zb1nextL0 *
3265 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3266 return (nextobj <= zb2thisobj);
3267 }
3268
3269 if (zb1->zb_object < zb2thisobj)
3270 return (B_TRUE);
3271 if (zb1->zb_object > zb2thisobj)
3272 return (B_FALSE);
3273 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3274 return (B_FALSE);
3275 return (zb1nextL0 <= zb2->zb_blkid);
3276 }