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 /*
23 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
25 * Copyright (c) 2014 Integros [integros.com]
26 * Copyright 2017 Nexenta Systems, Inc. All rights reserved.
27 */
28
29 #include <sys/sysmacros.h>
30 #include <sys/zfs_context.h>
31 #include <sys/fm/fs/zfs.h>
32 #include <sys/spa.h>
33 #include <sys/txg.h>
34 #include <sys/spa_impl.h>
35 #include <sys/vdev_impl.h>
36 #include <sys/zio_impl.h>
37 #include <sys/zio_compress.h>
38 #include <sys/zio_checksum.h>
39 #include <sys/dmu_objset.h>
40 #include <sys/arc.h>
41 #include <sys/ddt.h>
42 #include <sys/blkptr.h>
43 #include <sys/special.h>
44 #include <sys/blkptr.h>
45 #include <sys/zfeature.h>
46 #include <sys/dkioc_free_util.h>
47 #include <sys/dsl_scan.h>
48
49 #include <sys/metaslab_impl.h>
50 #include <sys/abd.h>
51
52 extern int zfs_txg_timeout;
53
54 /*
55 * ==========================================================================
56 * I/O type descriptions
57 * ==========================================================================
58 */
59 const char *zio_type_name[ZIO_TYPES] = {
60 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
61 "zio_ioctl"
62 };
63
64 boolean_t zio_dva_throttle_enabled = B_TRUE;
65
66 /*
67 * ==========================================================================
68 * I/O kmem caches
69 * ==========================================================================
70 */
71 kmem_cache_t *zio_cache;
72 kmem_cache_t *zio_link_cache;
73 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
74 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
75
76 #ifdef _KERNEL
77 extern vmem_t *zio_alloc_arena;
78 #endif
79
80 #define BP_SPANB(indblkshift, level) \
81 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
82 #define COMPARE_META_LEVEL 0x80000000ul
83
84 /*
85 * The following actions directly effect the spa's sync-to-convergence logic.
86 * The values below define the sync pass when we start performing the action.
87 * Care should be taken when changing these values as they directly impact
88 * spa_sync() performance. Tuning these values may introduce subtle performance
89 * pathologies and should only be done in the context of performance analysis.
90 * These tunables will eventually be removed and replaced with #defines once
91 * enough analysis has been done to determine optimal values.
92 *
93 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
94 * regular blocks are not deferred.
95 */
96 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
97 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
98 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
99
100 /*
101 * An allocating zio is one that either currently has the DVA allocate
102 * stage set or will have it later in its lifetime.
103 */
104 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
105
106 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
107
108 #ifdef ZFS_DEBUG
109 int zio_buf_debug_limit = 16384;
110 #else
111 int zio_buf_debug_limit = 0;
112 #endif
113
114 /*
115 * Fault insertion for stress testing
116 */
117 int zio_faulty_vdev_enabled = 0;
118 uint64_t zio_faulty_vdev_guid;
119 uint64_t zio_faulty_vdev_delay_us = 1000000; /* 1 second */
120
121 /*
122 * Tunable to allow for debugging SCSI UNMAP/SATA TRIM calls. Disabling
123 * it will prevent ZFS from attempting to issue DKIOCFREE ioctls to the
124 * underlying storage.
125 */
126 boolean_t zfs_trim = B_TRUE;
127 uint64_t zfs_trim_min_ext_sz = 1 << 20; /* 1 MB */
128
129 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
130
131 void
132 zio_init(void)
133 {
134 size_t c;
135 vmem_t *data_alloc_arena = NULL;
136
137 #ifdef _KERNEL
138 data_alloc_arena = zio_alloc_arena;
139 #endif
140 zio_cache = kmem_cache_create("zio_cache",
141 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
142 zio_link_cache = kmem_cache_create("zio_link_cache",
143 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
144
145 /*
146 * For small buffers, we want a cache for each multiple of
147 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
148 * for each quarter-power of 2.
149 */
150 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
151 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
152 size_t p2 = size;
153 size_t align = 0;
154 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
155
156 while (!ISP2(p2))
157 p2 &= p2 - 1;
158
159 #ifndef _KERNEL
160 /*
161 * If we are using watchpoints, put each buffer on its own page,
162 * to eliminate the performance overhead of trapping to the
163 * kernel when modifying a non-watched buffer that shares the
164 * page with a watched buffer.
165 */
166 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
167 continue;
168 #endif
169 if (size <= 4 * SPA_MINBLOCKSIZE) {
170 align = SPA_MINBLOCKSIZE;
171 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
172 align = MIN(p2 >> 2, PAGESIZE);
173 }
174
175 if (align != 0) {
176 char name[36];
177 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
178 zio_buf_cache[c] = kmem_cache_create(name, size,
179 align, NULL, NULL, NULL, NULL, NULL, cflags);
180
181 /*
182 * Since zio_data bufs do not appear in crash dumps, we
183 * pass KMC_NOTOUCH so that no allocator metadata is
184 * stored with the buffers.
185 */
186 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
187 zio_data_buf_cache[c] = kmem_cache_create(name, size,
188 align, NULL, NULL, NULL, NULL, data_alloc_arena,
189 cflags | KMC_NOTOUCH);
190 }
191 }
192
193 while (--c != 0) {
194 ASSERT(zio_buf_cache[c] != NULL);
195 if (zio_buf_cache[c - 1] == NULL)
196 zio_buf_cache[c - 1] = zio_buf_cache[c];
197
198 ASSERT(zio_data_buf_cache[c] != NULL);
199 if (zio_data_buf_cache[c - 1] == NULL)
200 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
201 }
202
203 zio_inject_init();
204
205 }
206
207 void
208 zio_fini(void)
209 {
210 size_t c;
211 kmem_cache_t *last_cache = NULL;
212 kmem_cache_t *last_data_cache = NULL;
213
214 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
215 if (zio_buf_cache[c] != last_cache) {
216 last_cache = zio_buf_cache[c];
217 kmem_cache_destroy(zio_buf_cache[c]);
218 }
219 zio_buf_cache[c] = NULL;
220
221 if (zio_data_buf_cache[c] != last_data_cache) {
222 last_data_cache = zio_data_buf_cache[c];
223 kmem_cache_destroy(zio_data_buf_cache[c]);
224 }
225 zio_data_buf_cache[c] = NULL;
226 }
227
228 kmem_cache_destroy(zio_link_cache);
229 kmem_cache_destroy(zio_cache);
230
231 zio_inject_fini();
232 }
233
234 /*
235 * ==========================================================================
236 * Allocate and free I/O buffers
237 * ==========================================================================
238 */
239
240 /*
241 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
242 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
243 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
244 * excess / transient data in-core during a crashdump.
245 */
246 void *
247 zio_buf_alloc(size_t size)
248 {
249 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
250
251 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
252
253 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
254 }
255
256 /*
257 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
258 * crashdump if the kernel panics. This exists so that we will limit the amount
259 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
260 * of kernel heap dumped to disk when the kernel panics)
261 */
262 void *
263 zio_data_buf_alloc(size_t size)
264 {
265 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
266
267 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
268
269 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
270 }
271
272 void
273 zio_buf_free(void *buf, size_t size)
274 {
275 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
276
277 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
278
279 kmem_cache_free(zio_buf_cache[c], buf);
280 }
281
282 void
283 zio_data_buf_free(void *buf, size_t size)
284 {
285 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
286
287 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
288
289 kmem_cache_free(zio_data_buf_cache[c], buf);
290 }
291
292 /*
293 * ==========================================================================
294 * Push and pop I/O transform buffers
295 * ==========================================================================
296 */
297 void
298 zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize,
299 zio_transform_func_t *transform)
300 {
301 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
302
303 /*
304 * Ensure that anyone expecting this zio to contain a linear ABD isn't
305 * going to get a nasty surprise when they try to access the data.
306 */
307 IMPLY(abd_is_linear(zio->io_abd), abd_is_linear(data));
308
309 zt->zt_orig_abd = zio->io_abd;
310 zt->zt_orig_size = zio->io_size;
311 zt->zt_bufsize = bufsize;
312 zt->zt_transform = transform;
313
314 zt->zt_next = zio->io_transform_stack;
315 zio->io_transform_stack = zt;
316
317 zio->io_abd = data;
318 zio->io_size = size;
319 }
320
321 void
322 zio_pop_transforms(zio_t *zio)
323 {
324 zio_transform_t *zt;
325
326 while ((zt = zio->io_transform_stack) != NULL) {
327 if (zt->zt_transform != NULL)
328 zt->zt_transform(zio,
329 zt->zt_orig_abd, zt->zt_orig_size);
330
331 if (zt->zt_bufsize != 0)
332 abd_free(zio->io_abd);
333
334 zio->io_abd = zt->zt_orig_abd;
335 zio->io_size = zt->zt_orig_size;
336 zio->io_transform_stack = zt->zt_next;
337
338 kmem_free(zt, sizeof (zio_transform_t));
339 }
340 }
341
342 /*
343 * ==========================================================================
344 * I/O transform callbacks for subblocks and decompression
345 * ==========================================================================
346 */
347 static void
348 zio_subblock(zio_t *zio, abd_t *data, uint64_t size)
349 {
350 ASSERT(zio->io_size > size);
351
352 if (zio->io_type == ZIO_TYPE_READ)
353 abd_copy(data, zio->io_abd, size);
354 }
355
356 static void
357 zio_decompress(zio_t *zio, abd_t *data, uint64_t size)
358 {
359 if (zio->io_error == 0) {
360 void *tmp = abd_borrow_buf(data, size);
361 int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
362 zio->io_abd, tmp, zio->io_size, size);
363 abd_return_buf_copy(data, tmp, size);
364
365 if (ret != 0)
366 zio->io_error = SET_ERROR(EIO);
367 }
368 }
369
370 /*
371 * ==========================================================================
372 * I/O parent/child relationships and pipeline interlocks
373 * ==========================================================================
374 */
375 zio_t *
376 zio_walk_parents(zio_t *cio, zio_link_t **zl)
377 {
378 list_t *pl = &cio->io_parent_list;
379
380 *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
381 if (*zl == NULL)
382 return (NULL);
383
384 ASSERT((*zl)->zl_child == cio);
385 return ((*zl)->zl_parent);
386 }
387
388 zio_t *
389 zio_walk_children(zio_t *pio, zio_link_t **zl)
390 {
391 list_t *cl = &pio->io_child_list;
392
393 *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
394 if (*zl == NULL)
395 return (NULL);
396
397 ASSERT((*zl)->zl_parent == pio);
398 return ((*zl)->zl_child);
399 }
400
401 zio_t *
402 zio_unique_parent(zio_t *cio)
403 {
404 zio_link_t *zl = NULL;
405 zio_t *pio = zio_walk_parents(cio, &zl);
406
407 VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
408 return (pio);
409 }
410
411 void
412 zio_add_child(zio_t *pio, zio_t *cio)
413 {
414 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
415
416 /*
417 * Logical I/Os can have logical, gang, or vdev children.
418 * Gang I/Os can have gang or vdev children.
419 * Vdev I/Os can only have vdev children.
420 * The following ASSERT captures all of these constraints.
421 */
422 ASSERT3S(cio->io_child_type, <=, pio->io_child_type);
423
424 zl->zl_parent = pio;
425 zl->zl_child = cio;
426
427 mutex_enter(&cio->io_lock);
428 mutex_enter(&pio->io_lock);
429
430 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
431
432 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
433 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
434
435 list_insert_head(&pio->io_child_list, zl);
436 list_insert_head(&cio->io_parent_list, zl);
437
438 pio->io_child_count++;
439 cio->io_parent_count++;
440
441 mutex_exit(&pio->io_lock);
442 mutex_exit(&cio->io_lock);
443 }
444
445 static void
446 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
447 {
448 ASSERT(zl->zl_parent == pio);
449 ASSERT(zl->zl_child == cio);
450
451 mutex_enter(&cio->io_lock);
452 mutex_enter(&pio->io_lock);
453
454 list_remove(&pio->io_child_list, zl);
455 list_remove(&cio->io_parent_list, zl);
456
457 pio->io_child_count--;
458 cio->io_parent_count--;
459
460 mutex_exit(&pio->io_lock);
461 mutex_exit(&cio->io_lock);
462
463 kmem_cache_free(zio_link_cache, zl);
464 }
465
466 static boolean_t
467 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
468 {
469 uint64_t *countp = &zio->io_children[child][wait];
470 boolean_t waiting = B_FALSE;
471
472 mutex_enter(&zio->io_lock);
473 ASSERT(zio->io_stall == NULL);
474 if (*countp != 0) {
475 zio->io_stage >>= 1;
476 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
477 zio->io_stall = countp;
478 waiting = B_TRUE;
479 }
480 mutex_exit(&zio->io_lock);
481
482 return (waiting);
483 }
484
485 static void
486 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
487 {
488 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
489 int *errorp = &pio->io_child_error[zio->io_child_type];
490
491 mutex_enter(&pio->io_lock);
492 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
493 *errorp = zio_worst_error(*errorp, zio->io_error);
494 pio->io_reexecute |= zio->io_reexecute;
495 ASSERT3U(*countp, >, 0);
496
497 (*countp)--;
498
499 if (*countp == 0 && pio->io_stall == countp) {
500 zio_taskq_type_t type =
501 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
502 ZIO_TASKQ_INTERRUPT;
503 pio->io_stall = NULL;
504 mutex_exit(&pio->io_lock);
505 /*
506 * Dispatch the parent zio in its own taskq so that
507 * the child can continue to make progress. This also
508 * prevents overflowing the stack when we have deeply nested
509 * parent-child relationships.
510 */
511 zio_taskq_dispatch(pio, type, B_FALSE);
512 } else {
513 mutex_exit(&pio->io_lock);
514 }
515 }
516
517 static void
518 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
519 {
520 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
521 zio->io_error = zio->io_child_error[c];
522 }
523
524 int
525 zio_bookmark_compare(const void *x1, const void *x2)
526 {
527 const zio_t *z1 = x1;
528 const zio_t *z2 = x2;
529
530 if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
531 return (-1);
532 if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
533 return (1);
534
535 if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
536 return (-1);
537 if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
538 return (1);
539
540 if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
541 return (-1);
542 if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
543 return (1);
544
545 if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
546 return (-1);
547 if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
548 return (1);
549
550 if (z1 < z2)
551 return (-1);
552 if (z1 > z2)
553 return (1);
554
555 return (0);
556 }
557
558 /*
559 * ==========================================================================
560 * Create the various types of I/O (read, write, free, etc)
561 * ==========================================================================
562 */
563 static zio_t *
564 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
565 abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done,
566 void *private, zio_type_t type, zio_priority_t priority,
567 enum zio_flag flags, vdev_t *vd, uint64_t offset,
568 const zbookmark_phys_t *zb, enum zio_stage stage, enum zio_stage pipeline)
569 {
570 zio_t *zio;
571
572 ASSERT3U(psize, <=, SPA_MAXBLOCKSIZE);
573 ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0);
574 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
575
576 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
577 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
578 ASSERT(vd || stage == ZIO_STAGE_OPEN);
579
580 IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW) != 0);
581
582 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
583 bzero(zio, sizeof (zio_t));
584
585 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
586 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
587
588 list_create(&zio->io_parent_list, sizeof (zio_link_t),
589 offsetof(zio_link_t, zl_parent_node));
590 list_create(&zio->io_child_list, sizeof (zio_link_t),
591 offsetof(zio_link_t, zl_child_node));
592 metaslab_trace_init(&zio->io_alloc_list);
593
594 if (vd != NULL)
595 zio->io_child_type = ZIO_CHILD_VDEV;
596 else if (flags & ZIO_FLAG_GANG_CHILD)
597 zio->io_child_type = ZIO_CHILD_GANG;
598 else if (flags & ZIO_FLAG_DDT_CHILD)
599 zio->io_child_type = ZIO_CHILD_DDT;
600 else
601 zio->io_child_type = ZIO_CHILD_LOGICAL;
602
603 if (bp != NULL) {
604 zio->io_bp = (blkptr_t *)bp;
605 zio->io_bp_copy = *bp;
606 zio->io_bp_orig = *bp;
607 if (type != ZIO_TYPE_WRITE ||
608 zio->io_child_type == ZIO_CHILD_DDT)
609 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
610 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
611 zio->io_logical = zio;
612 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
613 pipeline |= ZIO_GANG_STAGES;
614 }
615
616 zio->io_spa = spa;
617 zio->io_txg = txg;
618 zio->io_done = done;
619 zio->io_private = private;
620 zio->io_type = type;
621 zio->io_priority = priority;
622 zio->io_vd = vd;
623 zio->io_offset = offset;
624 zio->io_orig_abd = zio->io_abd = data;
625 zio->io_orig_size = zio->io_size = psize;
626 zio->io_lsize = lsize;
627 zio->io_orig_flags = zio->io_flags = flags;
628 zio->io_orig_stage = zio->io_stage = stage;
629 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
630 zio->io_pipeline_trace = ZIO_STAGE_OPEN;
631
632 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
633 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
634
635 if (zb != NULL)
636 zio->io_bookmark = *zb;
637
638 if (pio != NULL) {
639 zio->io_mc = pio->io_mc;
640 if (zio->io_logical == NULL)
641 zio->io_logical = pio->io_logical;
642 if (zio->io_child_type == ZIO_CHILD_GANG)
643 zio->io_gang_leader = pio->io_gang_leader;
644 zio_add_child(pio, zio);
645
646 /* copy the smartcomp setting when creating child zio's */
647 bcopy(&pio->io_smartcomp, &zio->io_smartcomp,
648 sizeof (zio->io_smartcomp));
649 }
650
651 return (zio);
652 }
653
654 static void
655 zio_destroy(zio_t *zio)
656 {
657 metaslab_trace_fini(&zio->io_alloc_list);
658 list_destroy(&zio->io_parent_list);
659 list_destroy(&zio->io_child_list);
660 mutex_destroy(&zio->io_lock);
661 cv_destroy(&zio->io_cv);
662 kmem_cache_free(zio_cache, zio);
663 }
664
665 zio_t *
666 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
667 void *private, enum zio_flag flags)
668 {
669 zio_t *zio;
670
671 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
672 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
673 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
674
675 return (zio);
676 }
677
678 zio_t *
679 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
680 {
681 return (zio_null(NULL, spa, NULL, done, private, flags));
682 }
683
684 void
685 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
686 {
687 /*
688 * SPECIAL-BP has two DVAs, but DVA[0] in this case is a
689 * temporary DVA, and after migration only the DVA[1]
690 * contains valid data. Therefore, we start walking for
691 * these BPs from DVA[1].
692 */
693 int start_dva = BP_IS_SPECIAL(bp) ? 1 : 0;
694
695 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
696 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
697 bp, (longlong_t)BP_GET_TYPE(bp));
698 }
699 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
700 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
701 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
702 bp, (longlong_t)BP_GET_CHECKSUM(bp));
703 }
704 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
705 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
706 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
707 bp, (longlong_t)BP_GET_COMPRESS(bp));
708 }
709 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
710 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
711 bp, (longlong_t)BP_GET_LSIZE(bp));
712 }
713 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
714 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
715 bp, (longlong_t)BP_GET_PSIZE(bp));
716 }
717
718 if (BP_IS_EMBEDDED(bp)) {
719 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
720 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
721 bp, (longlong_t)BPE_GET_ETYPE(bp));
722 }
723 }
724
725 /*
726 * Pool-specific checks.
727 *
728 * Note: it would be nice to verify that the blk_birth and
729 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
730 * allows the birth time of log blocks (and dmu_sync()-ed blocks
731 * that are in the log) to be arbitrarily large.
732 */
733 for (int i = start_dva; i < BP_GET_NDVAS(bp); i++) {
734 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
735 if (vdevid >= spa->spa_root_vdev->vdev_children) {
736 zfs_panic_recover("blkptr at %p DVA %u has invalid "
737 "VDEV %llu",
738 bp, i, (longlong_t)vdevid);
739 continue;
740 }
741 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
742 if (vd == NULL) {
743 zfs_panic_recover("blkptr at %p DVA %u has invalid "
744 "VDEV %llu",
745 bp, i, (longlong_t)vdevid);
746 continue;
747 }
748 if (vd->vdev_ops == &vdev_hole_ops) {
749 zfs_panic_recover("blkptr at %p DVA %u has hole "
750 "VDEV %llu",
751 bp, i, (longlong_t)vdevid);
752 continue;
753 }
754 if (vd->vdev_ops == &vdev_missing_ops) {
755 /*
756 * "missing" vdevs are valid during import, but we
757 * don't have their detailed info (e.g. asize), so
758 * we can't perform any more checks on them.
759 */
760 continue;
761 }
762 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
763 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
764 if (BP_IS_GANG(bp))
765 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
766 if (offset + asize > vd->vdev_asize) {
767 zfs_panic_recover("blkptr at %p DVA %u has invalid "
768 "OFFSET %llu",
769 bp, i, (longlong_t)offset);
770 }
771 }
772 }
773
774 zio_t *
775 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
776 abd_t *data, uint64_t size, zio_done_func_t *done, void *private,
777 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
778 {
779 zio_t *zio;
780
781 zfs_blkptr_verify(spa, bp);
782
783 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
784 data, size, size, done, private,
785 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
786 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
787 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
788
789 return (zio);
790 }
791
792 zio_t *
793 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
794 abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
795 zio_done_func_t *ready, zio_done_func_t *children_ready,
796 zio_done_func_t *physdone, zio_done_func_t *done,
797 void *private, zio_priority_t priority, enum zio_flag flags,
798 const zbookmark_phys_t *zb,
799 const zio_smartcomp_info_t *smartcomp)
800 {
801 zio_t *zio;
802
803 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
804 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
805 zp->zp_compress >= ZIO_COMPRESS_OFF &&
806 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
807 DMU_OT_IS_VALID(zp->zp_type) &&
808 zp->zp_level < 32 &&
809 zp->zp_copies > 0 &&
810 zp->zp_copies <= spa_max_replication(spa));
811
812 zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
813 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
814 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
815 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
816
817 zio->io_ready = ready;
818 zio->io_children_ready = children_ready;
819 zio->io_physdone = physdone;
820 zio->io_prop = *zp;
821 if (smartcomp != NULL)
822 bcopy(smartcomp, &zio->io_smartcomp, sizeof (*smartcomp));
823
824 /*
825 * Data can be NULL if we are going to call zio_write_override() to
826 * provide the already-allocated BP. But we may need the data to
827 * verify a dedup hit (if requested). In this case, don't try to
828 * dedup (just take the already-allocated BP verbatim).
829 */
830 if (data == NULL && zio->io_prop.zp_dedup_verify) {
831 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
832 }
833
834 return (zio);
835 }
836
837 zio_t *
838 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data,
839 uint64_t size, zio_done_func_t *done, void *private,
840 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
841 {
842 zio_t *zio;
843
844 zio = zio_create(pio, spa, txg, bp, data, size, size, done, private,
845 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
846 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
847
848 return (zio);
849 }
850
851 void
852 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
853 {
854 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
855 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
856 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
857 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
858
859 /*
860 * We must reset the io_prop to match the values that existed
861 * when the bp was first written by dmu_sync() keeping in mind
862 * that nopwrite and dedup are mutually exclusive.
863 */
864 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
865 zio->io_prop.zp_nopwrite = nopwrite;
866 zio->io_prop.zp_copies = copies;
867 zio->io_bp_override = bp;
868 }
869
870 void
871 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
872 {
873
874 /*
875 * The check for EMBEDDED is a performance optimization. We
876 * process the free here (by ignoring it) rather than
877 * putting it on the list and then processing it in zio_free_sync().
878 */
879 if (BP_IS_EMBEDDED(bp))
880 return;
881 metaslab_check_free(spa, bp);
882
883 /*
884 * Frees that are for the currently-syncing txg, are not going to be
885 * deferred, and which will not need to do a read (i.e. not GANG or
886 * DEDUP), can be processed immediately. Otherwise, put them on the
887 * in-memory list for later processing.
888 */
889 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
890 txg != spa->spa_syncing_txg ||
891 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
892 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
893 } else {
894 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0)));
895 }
896 }
897
898 zio_t *
899 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
900 enum zio_flag flags)
901 {
902 zio_t *zio;
903 enum zio_stage stage = ZIO_FREE_PIPELINE;
904
905 ASSERT(!BP_IS_HOLE(bp));
906 ASSERT(spa_syncing_txg(spa) == txg);
907 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
908
909 if (BP_IS_EMBEDDED(bp))
910 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
911
912 metaslab_check_free(spa, bp);
913 arc_freed(spa, bp);
914 dsl_scan_freed(spa, bp);
915
916 /*
917 * GANG and DEDUP blocks can induce a read (for the gang block header,
918 * or the DDT), so issue them asynchronously so that this thread is
919 * not tied up.
920 */
921 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
922 stage |= ZIO_STAGE_ISSUE_ASYNC;
923
924 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
925 BP_GET_PSIZE(bp), NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
926 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
927
928 return (zio);
929 }
930
931 zio_t *
932 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
933 zio_done_func_t *done, void *private, enum zio_flag flags)
934 {
935 zio_t *zio;
936
937 dprintf_bp(bp, "claiming in txg %llu", txg);
938
939 if (BP_IS_EMBEDDED(bp))
940 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
941
942 /*
943 * A claim is an allocation of a specific block. Claims are needed
944 * to support immediate writes in the intent log. The issue is that
945 * immediate writes contain committed data, but in a txg that was
946 * *not* committed. Upon opening the pool after an unclean shutdown,
947 * the intent log claims all blocks that contain immediate write data
948 * so that the SPA knows they're in use.
949 *
950 * All claims *must* be resolved in the first txg -- before the SPA
951 * starts allocating blocks -- so that nothing is allocated twice.
952 * If txg == 0 we just verify that the block is claimable.
953 */
954 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
955 ASSERT(txg == spa_first_txg(spa) || txg == 0);
956 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
957
958 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
959 BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
960 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
961 ASSERT0(zio->io_queued_timestamp);
962
963 return (zio);
964 }
965
966 static zio_t *
967 zio_ioctl_with_pipeline(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
968 zio_done_func_t *done, void *private, enum zio_flag flags,
969 enum zio_stage pipeline)
970 {
971 zio_t *zio;
972 int c;
973
974 if (vd->vdev_children == 0) {
975 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
976 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
977 ZIO_STAGE_OPEN, pipeline);
978
979 zio->io_cmd = cmd;
980 } else {
981 zio = zio_null(pio, spa, vd, done, private, flags);
982 /*
983 * DKIOCFREE ioctl's need some special handling on interior
984 * vdevs. If the device provides an ops function to handle
985 * recomputing dkioc_free extents, then we call it.
986 * Otherwise the default behavior applies, which simply fans
987 * out the ioctl to all component vdevs.
988 */
989 if (cmd == DKIOCFREE && vd->vdev_ops->vdev_op_trim != NULL) {
990 vd->vdev_ops->vdev_op_trim(vd, zio, private);
991 } else {
992 for (c = 0; c < vd->vdev_children; c++)
993 zio_nowait(zio_ioctl_with_pipeline(zio,
994 spa, vd->vdev_child[c], cmd, NULL,
995 private, flags, pipeline));
996 }
997 }
998
999 return (zio);
1000 }
1001
1002 zio_t *
1003 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
1004 zio_done_func_t *done, void *private, enum zio_flag flags)
1005 {
1006 return (zio_ioctl_with_pipeline(pio, spa, vd, cmd, done,
1007 private, flags, ZIO_IOCTL_PIPELINE));
1008 }
1009
1010 /*
1011 * Callback for when a trim zio has completed. This simply frees the
1012 * dkioc_free_list_t extent list of the DKIOCFREE ioctl.
1013 */
1014 static void
1015 zio_trim_done(zio_t *zio)
1016 {
1017 VERIFY(zio->io_private != NULL);
1018 dfl_free(zio->io_private);
1019 }
1020
1021 static void
1022 zio_trim_check(uint64_t start, uint64_t len, void *msp)
1023 {
1024 metaslab_t *ms = msp;
1025 boolean_t held = MUTEX_HELD(&ms->ms_lock);
1026 if (!held)
1027 mutex_enter(&ms->ms_lock);
1028 ASSERT(ms->ms_trimming_ts != NULL);
1029 ASSERT(range_tree_contains(ms->ms_trimming_ts->ts_tree,
1030 start - VDEV_LABEL_START_SIZE, len));
1031 if (!held)
1032 mutex_exit(&ms->ms_lock);
1033 }
1034
1035 /*
1036 * Takes a bunch of freed extents and tells the underlying vdevs that the
1037 * space associated with these extents can be released.
1038 * This is used by flash storage to pre-erase blocks for rapid reuse later
1039 * and thin-provisioned block storage to reclaim unused blocks.
1040 */
1041 zio_t *
1042 zio_trim(spa_t *spa, vdev_t *vd, struct range_tree *tree,
1043 zio_done_func_t *done, void *private, enum zio_flag flags,
1044 int trim_flags, metaslab_t *msp)
1045 {
1046 dkioc_free_list_t *dfl = NULL;
1047 range_seg_t *rs;
1048 uint64_t rs_idx;
1049 uint64_t num_exts;
1050 uint64_t bytes_issued = 0, bytes_skipped = 0, exts_skipped = 0;
1051 /*
1052 * We need this to invoke the caller's `done' callback with the
1053 * correct io_private (not the dkioc_free_list_t, which is needed
1054 * by the underlying DKIOCFREE ioctl).
1055 */
1056 zio_t *sub_pio = zio_root(spa, done, private, flags);
1057
1058 ASSERT(range_tree_space(tree) != 0);
1059
1060 if (!zfs_trim)
1061 return (sub_pio);
1062
1063 num_exts = avl_numnodes(&tree->rt_root);
1064 dfl = kmem_zalloc(DFL_SZ(num_exts), KM_SLEEP);
1065 dfl->dfl_flags = trim_flags;
1066 dfl->dfl_num_exts = num_exts;
1067 dfl->dfl_offset = VDEV_LABEL_START_SIZE;
1068 if (msp) {
1069 dfl->dfl_ck_func = zio_trim_check;
1070 dfl->dfl_ck_arg = msp;
1071 }
1072
1073 for (rs = avl_first(&tree->rt_root), rs_idx = 0; rs != NULL;
1074 rs = AVL_NEXT(&tree->rt_root, rs)) {
1075 uint64_t len = rs->rs_end - rs->rs_start;
1076
1077 if (len < zfs_trim_min_ext_sz) {
1078 bytes_skipped += len;
1079 exts_skipped++;
1080 continue;
1081 }
1082
1083 dfl->dfl_exts[rs_idx].dfle_start = rs->rs_start;
1084 dfl->dfl_exts[rs_idx].dfle_length = len;
1085
1086 // check we're a multiple of the vdev ashift
1087 ASSERT0(dfl->dfl_exts[rs_idx].dfle_start &
1088 ((1 << vd->vdev_ashift) - 1));
1089 ASSERT0(dfl->dfl_exts[rs_idx].dfle_length &
1090 ((1 << vd->vdev_ashift) - 1));
1091
1092 rs_idx++;
1093 bytes_issued += len;
1094 }
1095
1096 spa_trimstats_update(spa, rs_idx, bytes_issued, exts_skipped,
1097 bytes_skipped);
1098
1099 /* the zfs_trim_min_ext_sz filter may have shortened the list */
1100 if (dfl->dfl_num_exts != rs_idx) {
1101 dkioc_free_list_t *dfl2 = kmem_zalloc(DFL_SZ(rs_idx), KM_SLEEP);
1102 bcopy(dfl, dfl2, DFL_SZ(rs_idx));
1103 dfl2->dfl_num_exts = rs_idx;
1104 dfl_free(dfl);
1105 dfl = dfl2;
1106 }
1107
1108 zio_nowait(zio_ioctl_with_pipeline(sub_pio, spa, vd, DKIOCFREE,
1109 zio_trim_done, dfl, ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE |
1110 ZIO_FLAG_DONT_RETRY, ZIO_TRIM_PIPELINE));
1111 return (sub_pio);
1112 }
1113
1114 zio_t *
1115 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1116 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1117 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1118 {
1119 zio_t *zio;
1120
1121 ASSERT(vd->vdev_children == 0);
1122 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1123 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1124 ASSERT3U(offset + size, <=, vd->vdev_psize);
1125
1126 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1127 private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1128 offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1129
1130 zio->io_prop.zp_checksum = checksum;
1131
1132 return (zio);
1133 }
1134
1135 zio_t *
1136 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1137 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1138 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1139 {
1140 zio_t *zio;
1141
1142 ASSERT(vd->vdev_children == 0);
1143 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1144 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1145 ASSERT3U(offset + size, <=, vd->vdev_psize);
1146
1147 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1148 private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1149 offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1150
1151 zio->io_prop.zp_checksum = checksum;
1152
1153 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1154 /*
1155 * zec checksums are necessarily destructive -- they modify
1156 * the end of the write buffer to hold the verifier/checksum.
1157 * Therefore, we must make a local copy in case the data is
1158 * being written to multiple places in parallel.
1159 */
1160 abd_t *wbuf = abd_alloc_sametype(data, size);
1161 abd_copy(wbuf, data, size);
1162
1163 zio_push_transform(zio, wbuf, size, size, NULL);
1164 }
1165
1166 return (zio);
1167 }
1168
1169 /*
1170 * Create a child I/O to do some work for us.
1171 */
1172 zio_t *
1173 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1174 abd_t *data, uint64_t size, int type, zio_priority_t priority,
1175 enum zio_flag flags, zio_done_func_t *done, void *private)
1176 {
1177 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1178 zio_t *zio;
1179
1180 ASSERT(vd->vdev_parent ==
1181 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
1182
1183 if (type == ZIO_TYPE_READ && bp != NULL) {
1184 /*
1185 * If we have the bp, then the child should perform the
1186 * checksum and the parent need not. This pushes error
1187 * detection as close to the leaves as possible and
1188 * eliminates redundant checksums in the interior nodes.
1189 */
1190 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1191 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1192 }
1193
1194 if (vd->vdev_children == 0)
1195 offset += VDEV_LABEL_START_SIZE;
1196
1197 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
1198
1199 /*
1200 * If we've decided to do a repair, the write is not speculative --
1201 * even if the original read was.
1202 */
1203 if (flags & ZIO_FLAG_IO_REPAIR)
1204 flags &= ~ZIO_FLAG_SPECULATIVE;
1205
1206 /*
1207 * If we're creating a child I/O that is not associated with a
1208 * top-level vdev, then the child zio is not an allocating I/O.
1209 * If this is a retried I/O then we ignore it since we will
1210 * have already processed the original allocating I/O.
1211 */
1212 if (flags & ZIO_FLAG_IO_ALLOCATING &&
1213 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1214 metaslab_class_t *mc = pio->io_mc;
1215
1216 ASSERT(mc->mc_alloc_throttle_enabled);
1217 ASSERT(type == ZIO_TYPE_WRITE);
1218 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1219 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1220 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1221 pio->io_child_type == ZIO_CHILD_GANG);
1222
1223 flags &= ~ZIO_FLAG_IO_ALLOCATING;
1224 }
1225
1226 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1227 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1228 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1229 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1230
1231 zio->io_physdone = pio->io_physdone;
1232 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1233 zio->io_logical->io_phys_children++;
1234
1235 return (zio);
1236 }
1237
1238 zio_t *
1239 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size,
1240 int type, zio_priority_t priority, enum zio_flag flags,
1241 zio_done_func_t *done, void *private)
1242 {
1243 zio_t *zio;
1244
1245 ASSERT(vd->vdev_ops->vdev_op_leaf);
1246
1247 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1248 data, size, size, done, private, type, priority,
1249 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1250 vd, offset, NULL,
1251 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1252
1253 return (zio);
1254 }
1255
1256 void
1257 zio_flush(zio_t *zio, vdev_t *vd)
1258 {
1259 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
1260 NULL, NULL,
1261 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1262 }
1263
1264 void
1265 zio_shrink(zio_t *zio, uint64_t size)
1266 {
1267 ASSERT3P(zio->io_executor, ==, NULL);
1268 ASSERT3P(zio->io_orig_size, ==, zio->io_size);
1269 ASSERT3U(size, <=, zio->io_size);
1270
1271 /*
1272 * We don't shrink for raidz because of problems with the
1273 * reconstruction when reading back less than the block size.
1274 * Note, BP_IS_RAIDZ() assumes no compression.
1275 */
1276 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1277 if (!BP_IS_RAIDZ(zio->io_bp)) {
1278 /* we are not doing a raw write */
1279 ASSERT3U(zio->io_size, ==, zio->io_lsize);
1280 zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1281 }
1282 }
1283
1284 /*
1285 * ==========================================================================
1286 * Prepare to read and write logical blocks
1287 * ==========================================================================
1288 */
1289
1290 static int
1291 zio_read_bp_init(zio_t *zio)
1292 {
1293 blkptr_t *bp = zio->io_bp;
1294
1295 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1296 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1297 !(zio->io_flags & ZIO_FLAG_RAW)) {
1298 uint64_t psize =
1299 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1300 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1301 psize, psize, zio_decompress);
1302 }
1303
1304 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1305 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1306
1307 int psize = BPE_GET_PSIZE(bp);
1308 void *data = abd_borrow_buf(zio->io_abd, psize);
1309 decode_embedded_bp_compressed(bp, data);
1310 abd_return_buf_copy(zio->io_abd, data, psize);
1311 } else {
1312 ASSERT(!BP_IS_EMBEDDED(bp));
1313 }
1314
1315 if (!BP_IS_METADATA(bp))
1316 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1317
1318 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1319 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1320
1321 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1322 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1323
1324 return (ZIO_PIPELINE_CONTINUE);
1325 }
1326
1327 static int
1328 zio_write_bp_init(zio_t *zio)
1329 {
1330 if (!IO_IS_ALLOCATING(zio))
1331 return (ZIO_PIPELINE_CONTINUE);
1332
1333 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1334
1335 if (zio->io_bp_override) {
1336 blkptr_t *bp = zio->io_bp;
1337 zio_prop_t *zp = &zio->io_prop;
1338
1339 ASSERT(bp->blk_birth != zio->io_txg);
1340 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1341
1342 *bp = *zio->io_bp_override;
1343 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1344
1345 if (BP_IS_EMBEDDED(bp))
1346 return (ZIO_PIPELINE_CONTINUE);
1347
1348 /*
1349 * If we've been overridden and nopwrite is set then
1350 * set the flag accordingly to indicate that a nopwrite
1351 * has already occurred.
1352 */
1353 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1354 ASSERT(!zp->zp_dedup);
1355 ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1356 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1357 return (ZIO_PIPELINE_CONTINUE);
1358 }
1359
1360 ASSERT(!zp->zp_nopwrite);
1361
1362 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1363 return (ZIO_PIPELINE_CONTINUE);
1364
1365 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1366 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1367
1368 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1369 BP_SET_DEDUP(bp, 1);
1370 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1371 return (ZIO_PIPELINE_CONTINUE);
1372 }
1373
1374 /*
1375 * We were unable to handle this as an override bp, treat
1376 * it as a regular write I/O.
1377 */
1378 zio->io_bp_override = NULL;
1379 *bp = zio->io_bp_orig;
1380 zio->io_pipeline = zio->io_orig_pipeline;
1381 }
1382
1383 return (ZIO_PIPELINE_CONTINUE);
1384 }
1385
1386 static int
1387 zio_write_compress(zio_t *zio)
1388 {
1389 spa_t *spa = zio->io_spa;
1390 zio_prop_t *zp = &zio->io_prop;
1391 enum zio_compress compress = zp->zp_compress;
1392 blkptr_t *bp = zio->io_bp;
1393 uint64_t lsize = zio->io_lsize;
1394 uint64_t psize = zio->io_size;
1395 int pass = 1;
1396
1397 EQUIV(lsize != psize, (zio->io_flags & ZIO_FLAG_RAW) != 0);
1398
1399 /*
1400 * If our children haven't all reached the ready stage,
1401 * wait for them and then repeat this pipeline stage.
1402 */
1403 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1404 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1405 return (ZIO_PIPELINE_STOP);
1406
1407 if (!IO_IS_ALLOCATING(zio))
1408 return (ZIO_PIPELINE_CONTINUE);
1409
1410 if (zio->io_children_ready != NULL) {
1411 /*
1412 * Now that all our children are ready, run the callback
1413 * associated with this zio in case it wants to modify the
1414 * data to be written.
1415 */
1416 ASSERT3U(zp->zp_level, >, 0);
1417 zio->io_children_ready(zio);
1418 }
1419
1420 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1421 ASSERT(zio->io_bp_override == NULL);
1422
1423 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1424 /*
1425 * We're rewriting an existing block, which means we're
1426 * working on behalf of spa_sync(). For spa_sync() to
1427 * converge, it must eventually be the case that we don't
1428 * have to allocate new blocks. But compression changes
1429 * the blocksize, which forces a reallocate, and makes
1430 * convergence take longer. Therefore, after the first
1431 * few passes, stop compressing to ensure convergence.
1432 */
1433 pass = spa_sync_pass(spa);
1434
1435 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1436 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1437 ASSERT(!BP_GET_DEDUP(bp));
1438
1439 if (pass >= zfs_sync_pass_dont_compress)
1440 compress = ZIO_COMPRESS_OFF;
1441
1442 /* Make sure someone doesn't change their mind on overwrites */
1443 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1444 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1445 }
1446
1447 DTRACE_PROBE1(zio_compress_ready, zio_t *, zio);
1448 /* If it's a compressed write that is not raw, compress the buffer. */
1449 if (compress != ZIO_COMPRESS_OFF && psize == lsize &&
1450 ZIO_SHOULD_COMPRESS(zio)) {
1451 void *cbuf = zio_buf_alloc(lsize);
1452 psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize);
1453 if (psize == 0 || psize == lsize) {
1454 compress = ZIO_COMPRESS_OFF;
1455 zio_buf_free(cbuf, lsize);
1456 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1457 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1458 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1459 encode_embedded_bp_compressed(bp,
1460 cbuf, compress, lsize, psize);
1461 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1462 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1463 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1464 zio_buf_free(cbuf, lsize);
1465 bp->blk_birth = zio->io_txg;
1466 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1467 ASSERT(spa_feature_is_active(spa,
1468 SPA_FEATURE_EMBEDDED_DATA));
1469 if (zio->io_smartcomp.sc_result != NULL) {
1470 zio->io_smartcomp.sc_result(
1471 zio->io_smartcomp.sc_userinfo, zio);
1472 } else {
1473 ASSERT(zio->io_smartcomp.sc_ask == NULL);
1474 }
1475 return (ZIO_PIPELINE_CONTINUE);
1476 } else {
1477 /*
1478 * Round up compressed size up to the ashift
1479 * of the smallest-ashift device, and zero the tail.
1480 * This ensures that the compressed size of the BP
1481 * (and thus compressratio property) are correct,
1482 * in that we charge for the padding used to fill out
1483 * the last sector.
1484 */
1485 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1486 size_t rounded = (size_t)P2ROUNDUP(psize,
1487 1ULL << spa->spa_min_ashift);
1488 if (rounded >= lsize) {
1489 compress = ZIO_COMPRESS_OFF;
1490 zio_buf_free(cbuf, lsize);
1491 psize = lsize;
1492 } else {
1493 abd_t *cdata = abd_get_from_buf(cbuf, lsize);
1494 abd_take_ownership_of_buf(cdata, B_TRUE);
1495 abd_zero_off(cdata, psize, rounded - psize);
1496 psize = rounded;
1497 zio_push_transform(zio, cdata,
1498 psize, lsize, NULL);
1499 }
1500 }
1501
1502 if (zio->io_smartcomp.sc_result != NULL) {
1503 zio->io_smartcomp.sc_result(
1504 zio->io_smartcomp.sc_userinfo, zio);
1505 } else {
1506 ASSERT(zio->io_smartcomp.sc_ask == NULL);
1507 }
1508
1509 /*
1510 * We were unable to handle this as an override bp, treat
1511 * it as a regular write I/O.
1512 */
1513 zio->io_bp_override = NULL;
1514 *bp = zio->io_bp_orig;
1515 zio->io_pipeline = zio->io_orig_pipeline;
1516 } else {
1517 ASSERT3U(psize, !=, 0);
1518
1519 /*
1520 * We are here because of:
1521 * - compress == ZIO_COMPRESS_OFF
1522 * - SmartCompression decides don't compress this data
1523 * - this is a RAW-write
1524 *
1525 * In case of RAW-write we should not override "compress"
1526 */
1527 if ((zio->io_flags & ZIO_FLAG_RAW) == 0)
1528 compress = ZIO_COMPRESS_OFF;
1529 }
1530
1531 /*
1532 * The final pass of spa_sync() must be all rewrites, but the first
1533 * few passes offer a trade-off: allocating blocks defers convergence,
1534 * but newly allocated blocks are sequential, so they can be written
1535 * to disk faster. Therefore, we allow the first few passes of
1536 * spa_sync() to allocate new blocks, but force rewrites after that.
1537 * There should only be a handful of blocks after pass 1 in any case.
1538 */
1539 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1540 BP_GET_PSIZE(bp) == psize &&
1541 pass >= zfs_sync_pass_rewrite) {
1542 ASSERT(psize != 0);
1543 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1544 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1545 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1546 } else {
1547 BP_ZERO(bp);
1548 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1549 }
1550
1551 if (psize == 0) {
1552 if (zio->io_bp_orig.blk_birth != 0 &&
1553 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1554 BP_SET_LSIZE(bp, lsize);
1555 BP_SET_TYPE(bp, zp->zp_type);
1556 BP_SET_LEVEL(bp, zp->zp_level);
1557 BP_SET_BIRTH(bp, zio->io_txg, 0);
1558 }
1559 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1560 } else {
1561 if (zp->zp_dedup) {
1562 /* check the best-effort dedup setting */
1563 zio_best_effort_dedup(zio);
1564 }
1565 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1566 BP_SET_LSIZE(bp, lsize);
1567 BP_SET_TYPE(bp, zp->zp_type);
1568 BP_SET_LEVEL(bp, zp->zp_level);
1569 BP_SET_PSIZE(bp, psize);
1570 BP_SET_COMPRESS(bp, compress);
1571 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1572 BP_SET_DEDUP(bp, zp->zp_dedup);
1573 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1574 if (zp->zp_dedup) {
1575 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1576 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1577 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1578 }
1579 if (zp->zp_nopwrite) {
1580 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1581 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1582 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1583 }
1584 }
1585 return (ZIO_PIPELINE_CONTINUE);
1586 }
1587
1588 static int
1589 zio_free_bp_init(zio_t *zio)
1590 {
1591 blkptr_t *bp = zio->io_bp;
1592
1593 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1594 if (BP_GET_DEDUP(bp))
1595 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1596 }
1597
1598 return (ZIO_PIPELINE_CONTINUE);
1599 }
1600
1601 /*
1602 * ==========================================================================
1603 * Execute the I/O pipeline
1604 * ==========================================================================
1605 */
1606
1607 static void
1608 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1609 {
1610 spa_t *spa = zio->io_spa;
1611 zio_type_t t = zio->io_type;
1612 int flags = (cutinline ? TQ_FRONT : 0);
1613
1614 /*
1615 * If we're a config writer or a probe, the normal issue and
1616 * interrupt threads may all be blocked waiting for the config lock.
1617 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1618 */
1619 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1620 t = ZIO_TYPE_NULL;
1621
1622 /*
1623 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1624 */
1625 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1626 t = ZIO_TYPE_NULL;
1627
1628 /*
1629 * If this is a high priority I/O, then use the high priority taskq if
1630 * available.
1631 */
1632 if ((zio->io_priority == ZIO_PRIORITY_NOW ||
1633 zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) &&
1634 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1635 q++;
1636
1637 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1638
1639 /*
1640 * NB: We are assuming that the zio can only be dispatched
1641 * to a single taskq at a time. It would be a grievous error
1642 * to dispatch the zio to another taskq at the same time.
1643 */
1644 ASSERT(zio->io_tqent.tqent_next == NULL);
1645 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1646 flags, &zio->io_tqent);
1647 }
1648
1649 static boolean_t
1650 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1651 {
1652 kthread_t *executor = zio->io_executor;
1653 spa_t *spa = zio->io_spa;
1654
1655 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1656 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1657 uint_t i;
1658 for (i = 0; i < tqs->stqs_count; i++) {
1659 if (taskq_member(tqs->stqs_taskq[i], executor))
1660 return (B_TRUE);
1661 }
1662 }
1663
1664 return (B_FALSE);
1665 }
1666
1667 static int
1668 zio_issue_async(zio_t *zio)
1669 {
1670 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1671
1672 return (ZIO_PIPELINE_STOP);
1673 }
1674
1675 void
1676 zio_interrupt(zio_t *zio)
1677 {
1678 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1679 }
1680
1681 void
1682 zio_delay_interrupt(zio_t *zio)
1683 {
1684 /*
1685 * The timeout_generic() function isn't defined in userspace, so
1686 * rather than trying to implement the function, the zio delay
1687 * functionality has been disabled for userspace builds.
1688 */
1689
1690 #ifdef _KERNEL
1691 /*
1692 * If io_target_timestamp is zero, then no delay has been registered
1693 * for this IO, thus jump to the end of this function and "skip" the
1694 * delay; issuing it directly to the zio layer.
1695 */
1696 if (zio->io_target_timestamp != 0) {
1697 hrtime_t now = gethrtime();
1698
1699 if (now >= zio->io_target_timestamp) {
1700 /*
1701 * This IO has already taken longer than the target
1702 * delay to complete, so we don't want to delay it
1703 * any longer; we "miss" the delay and issue it
1704 * directly to the zio layer. This is likely due to
1705 * the target latency being set to a value less than
1706 * the underlying hardware can satisfy (e.g. delay
1707 * set to 1ms, but the disks take 10ms to complete an
1708 * IO request).
1709 */
1710
1711 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1712 hrtime_t, now);
1713
1714 zio_interrupt(zio);
1715 } else {
1716 hrtime_t diff = zio->io_target_timestamp - now;
1717
1718 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1719 hrtime_t, now, hrtime_t, diff);
1720
1721 (void) timeout_generic(CALLOUT_NORMAL,
1722 (void (*)(void *))zio_interrupt, zio, diff, 1, 0);
1723 }
1724
1725 return;
1726 }
1727 #endif
1728
1729 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
1730 zio_interrupt(zio);
1731 }
1732
1733 /*
1734 * Execute the I/O pipeline until one of the following occurs:
1735 *
1736 * (1) the I/O completes
1737 * (2) the pipeline stalls waiting for dependent child I/Os
1738 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1739 * (4) the I/O is delegated by vdev-level caching or aggregation
1740 * (5) the I/O is deferred due to vdev-level queueing
1741 * (6) the I/O is handed off to another thread.
1742 *
1743 * In all cases, the pipeline stops whenever there's no CPU work; it never
1744 * burns a thread in cv_wait().
1745 *
1746 * There's no locking on io_stage because there's no legitimate way
1747 * for multiple threads to be attempting to process the same I/O.
1748 */
1749 static zio_pipe_stage_t *zio_pipeline[];
1750
1751 void
1752 zio_execute(zio_t *zio)
1753 {
1754 zio->io_executor = curthread;
1755
1756 ASSERT3U(zio->io_queued_timestamp, >, 0);
1757
1758 while (zio->io_stage < ZIO_STAGE_DONE) {
1759 enum zio_stage pipeline = zio->io_pipeline;
1760 enum zio_stage old_stage = zio->io_stage;
1761 enum zio_stage stage = zio->io_stage;
1762 int rv;
1763
1764 ASSERT(!MUTEX_HELD(&zio->io_lock));
1765 ASSERT(ISP2(stage));
1766 ASSERT(zio->io_stall == NULL);
1767
1768 do {
1769 stage <<= 1;
1770 } while ((stage & pipeline) == 0);
1771
1772 ASSERT(stage <= ZIO_STAGE_DONE);
1773
1774 /*
1775 * If we are in interrupt context and this pipeline stage
1776 * will grab a config lock that is held across I/O,
1777 * or may wait for an I/O that needs an interrupt thread
1778 * to complete, issue async to avoid deadlock.
1779 *
1780 * For VDEV_IO_START, we cut in line so that the io will
1781 * be sent to disk promptly.
1782 */
1783 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1784 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1785 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1786 zio_requeue_io_start_cut_in_line : B_FALSE;
1787 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1788 return;
1789 }
1790
1791 zio->io_stage = stage;
1792 zio->io_pipeline_trace |= zio->io_stage;
1793 rv = zio_pipeline[highbit64(stage) - 1](zio);
1794
1795 if (rv == ZIO_PIPELINE_STOP)
1796 return;
1797
1798 if (rv == ZIO_PIPELINE_RESTART_STAGE) {
1799 zio->io_stage = old_stage;
1800 (void) zio_issue_async(zio);
1801 return;
1802 }
1803
1804 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1805 }
1806 }
1807
1808 /*
1809 * ==========================================================================
1810 * Initiate I/O, either sync or async
1811 * ==========================================================================
1812 */
1813 int
1814 zio_wait(zio_t *zio)
1815 {
1816 int error;
1817
1818 ASSERT3P(zio->io_stage, ==, ZIO_STAGE_OPEN);
1819 ASSERT3P(zio->io_executor, ==, NULL);
1820
1821 zio->io_waiter = curthread;
1822 ASSERT0(zio->io_queued_timestamp);
1823 zio->io_queued_timestamp = gethrtime();
1824
1825 zio_execute(zio);
1826
1827 mutex_enter(&zio->io_lock);
1828 while (zio->io_executor != NULL)
1829 cv_wait(&zio->io_cv, &zio->io_lock);
1830 mutex_exit(&zio->io_lock);
1831
1832 error = zio->io_error;
1833 zio_destroy(zio);
1834
1835 return (error);
1836 }
1837
1838 void
1839 zio_nowait(zio_t *zio)
1840 {
1841 ASSERT3P(zio->io_executor, ==, NULL);
1842
1843 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1844 zio_unique_parent(zio) == NULL) {
1845 /*
1846 * This is a logical async I/O with no parent to wait for it.
1847 * We add it to the spa_async_root_zio "Godfather" I/O which
1848 * will ensure they complete prior to unloading the pool.
1849 */
1850 spa_t *spa = zio->io_spa;
1851
1852 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1853 }
1854
1855 ASSERT0(zio->io_queued_timestamp);
1856 zio->io_queued_timestamp = gethrtime();
1857 zio_execute(zio);
1858 }
1859
1860 /*
1861 * ==========================================================================
1862 * Reexecute, cancel, or suspend/resume failed I/O
1863 * ==========================================================================
1864 */
1865
1866 static void
1867 zio_reexecute(zio_t *pio)
1868 {
1869 zio_t *cio, *cio_next;
1870
1871 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1872 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1873 ASSERT(pio->io_gang_leader == NULL);
1874 ASSERT(pio->io_gang_tree == NULL);
1875
1876 pio->io_flags = pio->io_orig_flags;
1877 pio->io_stage = pio->io_orig_stage;
1878 pio->io_pipeline = pio->io_orig_pipeline;
1879 pio->io_reexecute = 0;
1880 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1881 pio->io_pipeline_trace = 0;
1882 pio->io_error = 0;
1883 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1884 pio->io_state[w] = 0;
1885 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1886 pio->io_child_error[c] = 0;
1887
1888 if (IO_IS_ALLOCATING(pio))
1889 BP_ZERO(pio->io_bp);
1890
1891 /*
1892 * As we reexecute pio's children, new children could be created.
1893 * New children go to the head of pio's io_child_list, however,
1894 * so we will (correctly) not reexecute them. The key is that
1895 * the remainder of pio's io_child_list, from 'cio_next' onward,
1896 * cannot be affected by any side effects of reexecuting 'cio'.
1897 */
1898 zio_link_t *zl = NULL;
1899 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
1900 cio_next = zio_walk_children(pio, &zl);
1901 mutex_enter(&pio->io_lock);
1902 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1903 pio->io_children[cio->io_child_type][w]++;
1904 mutex_exit(&pio->io_lock);
1905 zio_reexecute(cio);
1906 }
1907
1908 /*
1909 * Now that all children have been reexecuted, execute the parent.
1910 * We don't reexecute "The Godfather" I/O here as it's the
1911 * responsibility of the caller to wait on it.
1912 */
1913 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
1914 pio->io_queued_timestamp = gethrtime();
1915 zio_execute(pio);
1916 }
1917 }
1918
1919 void
1920 zio_suspend(spa_t *spa, zio_t *zio)
1921 {
1922 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1923 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1924 "failure and the failure mode property for this pool "
1925 "is set to panic.", spa_name(spa));
1926
1927 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1928
1929 mutex_enter(&spa->spa_suspend_lock);
1930
1931 if (spa->spa_suspend_zio_root == NULL)
1932 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1933 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1934 ZIO_FLAG_GODFATHER);
1935
1936 spa->spa_suspended = B_TRUE;
1937
1938 if (zio != NULL) {
1939 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1940 ASSERT(zio != spa->spa_suspend_zio_root);
1941 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1942 ASSERT(zio_unique_parent(zio) == NULL);
1943 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1944 zio_add_child(spa->spa_suspend_zio_root, zio);
1945 }
1946
1947 mutex_exit(&spa->spa_suspend_lock);
1948 }
1949
1950 int
1951 zio_resume(spa_t *spa)
1952 {
1953 zio_t *pio;
1954
1955 /*
1956 * Reexecute all previously suspended i/o.
1957 */
1958 mutex_enter(&spa->spa_suspend_lock);
1959 spa->spa_suspended = B_FALSE;
1960 cv_broadcast(&spa->spa_suspend_cv);
1961 pio = spa->spa_suspend_zio_root;
1962 spa->spa_suspend_zio_root = NULL;
1963 mutex_exit(&spa->spa_suspend_lock);
1964
1965 if (pio == NULL)
1966 return (0);
1967
1968 zio_reexecute(pio);
1969 return (zio_wait(pio));
1970 }
1971
1972 void
1973 zio_resume_wait(spa_t *spa)
1974 {
1975 mutex_enter(&spa->spa_suspend_lock);
1976 while (spa_suspended(spa))
1977 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1978 mutex_exit(&spa->spa_suspend_lock);
1979 }
1980
1981 /*
1982 * ==========================================================================
1983 * Gang blocks.
1984 *
1985 * A gang block is a collection of small blocks that looks to the DMU
1986 * like one large block. When zio_dva_allocate() cannot find a block
1987 * of the requested size, due to either severe fragmentation or the pool
1988 * being nearly full, it calls zio_write_gang_block() to construct the
1989 * block from smaller fragments.
1990 *
1991 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1992 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1993 * an indirect block: it's an array of block pointers. It consumes
1994 * only one sector and hence is allocatable regardless of fragmentation.
1995 * The gang header's bps point to its gang members, which hold the data.
1996 *
1997 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1998 * as the verifier to ensure uniqueness of the SHA256 checksum.
1999 * Critically, the gang block bp's blk_cksum is the checksum of the data,
2000 * not the gang header. This ensures that data block signatures (needed for
2001 * deduplication) are independent of how the block is physically stored.
2002 *
2003 * Gang blocks can be nested: a gang member may itself be a gang block.
2004 * Thus every gang block is a tree in which root and all interior nodes are
2005 * gang headers, and the leaves are normal blocks that contain user data.
2006 * The root of the gang tree is called the gang leader.
2007 *
2008 * To perform any operation (read, rewrite, free, claim) on a gang block,
2009 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
2010 * in the io_gang_tree field of the original logical i/o by recursively
2011 * reading the gang leader and all gang headers below it. This yields
2012 * an in-core tree containing the contents of every gang header and the
2013 * bps for every constituent of the gang block.
2014 *
2015 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
2016 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
2017 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
2018 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
2019 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
2020 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
2021 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
2022 * of the gang header plus zio_checksum_compute() of the data to update the
2023 * gang header's blk_cksum as described above.
2024 *
2025 * The two-phase assemble/issue model solves the problem of partial failure --
2026 * what if you'd freed part of a gang block but then couldn't read the
2027 * gang header for another part? Assembling the entire gang tree first
2028 * ensures that all the necessary gang header I/O has succeeded before
2029 * starting the actual work of free, claim, or write. Once the gang tree
2030 * is assembled, free and claim are in-memory operations that cannot fail.
2031 *
2032 * In the event that a gang write fails, zio_dva_unallocate() walks the
2033 * gang tree to immediately free (i.e. insert back into the space map)
2034 * everything we've allocated. This ensures that we don't get ENOSPC
2035 * errors during repeated suspend/resume cycles due to a flaky device.
2036 *
2037 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
2038 * the gang tree, we won't modify the block, so we can safely defer the free
2039 * (knowing that the block is still intact). If we *can* assemble the gang
2040 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2041 * each constituent bp and we can allocate a new block on the next sync pass.
2042 *
2043 * In all cases, the gang tree allows complete recovery from partial failure.
2044 * ==========================================================================
2045 */
2046
2047 static void
2048 zio_gang_issue_func_done(zio_t *zio)
2049 {
2050 abd_put(zio->io_abd);
2051 }
2052
2053 static zio_t *
2054 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2055 uint64_t offset)
2056 {
2057 if (gn != NULL)
2058 return (pio);
2059
2060 return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset),
2061 BP_GET_PSIZE(bp), zio_gang_issue_func_done,
2062 NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2063 &pio->io_bookmark));
2064 }
2065
2066 static zio_t *
2067 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2068 uint64_t offset)
2069 {
2070 zio_t *zio;
2071
2072 if (gn != NULL) {
2073 abd_t *gbh_abd =
2074 abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2075 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2076 gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL,
2077 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2078 &pio->io_bookmark);
2079 /*
2080 * As we rewrite each gang header, the pipeline will compute
2081 * a new gang block header checksum for it; but no one will
2082 * compute a new data checksum, so we do that here. The one
2083 * exception is the gang leader: the pipeline already computed
2084 * its data checksum because that stage precedes gang assembly.
2085 * (Presently, nothing actually uses interior data checksums;
2086 * this is just good hygiene.)
2087 */
2088 if (gn != pio->io_gang_leader->io_gang_tree) {
2089 abd_t *buf = abd_get_offset(data, offset);
2090
2091 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
2092 buf, BP_GET_PSIZE(bp));
2093
2094 abd_put(buf);
2095 }
2096 /*
2097 * If we are here to damage data for testing purposes,
2098 * leave the GBH alone so that we can detect the damage.
2099 */
2100 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
2101 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2102 } else {
2103 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2104 abd_get_offset(data, offset), BP_GET_PSIZE(bp),
2105 zio_gang_issue_func_done, NULL, pio->io_priority,
2106 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2107 }
2108
2109 return (zio);
2110 }
2111
2112 /* ARGSUSED */
2113 static zio_t *
2114 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2115 uint64_t offset)
2116 {
2117 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
2118 ZIO_GANG_CHILD_FLAGS(pio)));
2119 }
2120
2121 /* ARGSUSED */
2122 static zio_t *
2123 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2124 uint64_t offset)
2125 {
2126 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
2127 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
2128 }
2129
2130 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2131 NULL,
2132 zio_read_gang,
2133 zio_rewrite_gang,
2134 zio_free_gang,
2135 zio_claim_gang,
2136 NULL
2137 };
2138
2139 static void zio_gang_tree_assemble_done(zio_t *zio);
2140
2141 static zio_gang_node_t *
2142 zio_gang_node_alloc(zio_gang_node_t **gnpp)
2143 {
2144 zio_gang_node_t *gn;
2145
2146 ASSERT(*gnpp == NULL);
2147
2148 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2149 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2150 *gnpp = gn;
2151
2152 return (gn);
2153 }
2154
2155 static void
2156 zio_gang_node_free(zio_gang_node_t **gnpp)
2157 {
2158 zio_gang_node_t *gn = *gnpp;
2159
2160 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2161 ASSERT(gn->gn_child[g] == NULL);
2162
2163 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2164 kmem_free(gn, sizeof (*gn));
2165 *gnpp = NULL;
2166 }
2167
2168 static void
2169 zio_gang_tree_free(zio_gang_node_t **gnpp)
2170 {
2171 zio_gang_node_t *gn = *gnpp;
2172
2173 if (gn == NULL)
2174 return;
2175
2176 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2177 zio_gang_tree_free(&gn->gn_child[g]);
2178
2179 zio_gang_node_free(gnpp);
2180 }
2181
2182 static void
2183 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2184 {
2185 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2186 abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2187
2188 ASSERT(gio->io_gang_leader == gio);
2189 ASSERT(BP_IS_GANG(bp));
2190
2191 zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2192 zio_gang_tree_assemble_done, gn, gio->io_priority,
2193 ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2194 }
2195
2196 static void
2197 zio_gang_tree_assemble_done(zio_t *zio)
2198 {
2199 zio_t *gio = zio->io_gang_leader;
2200 zio_gang_node_t *gn = zio->io_private;
2201 blkptr_t *bp = zio->io_bp;
2202
2203 ASSERT(gio == zio_unique_parent(zio));
2204 ASSERT(zio->io_child_count == 0);
2205
2206 if (zio->io_error)
2207 return;
2208
2209 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2210 if (BP_SHOULD_BYTESWAP(bp))
2211 byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size);
2212
2213 ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh);
2214 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2215 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2216
2217 abd_put(zio->io_abd);
2218
2219 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2220 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2221 if (!BP_IS_GANG(gbp))
2222 continue;
2223 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2224 }
2225 }
2226
2227 static void
2228 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data,
2229 uint64_t offset)
2230 {
2231 zio_t *gio = pio->io_gang_leader;
2232 zio_t *zio;
2233
2234 ASSERT(BP_IS_GANG(bp) == !!gn);
2235 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2236 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2237
2238 /*
2239 * If you're a gang header, your data is in gn->gn_gbh.
2240 * If you're a gang member, your data is in 'data' and gn == NULL.
2241 */
2242 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset);
2243
2244 if (gn != NULL) {
2245 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2246
2247 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2248 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2249 if (BP_IS_HOLE(gbp))
2250 continue;
2251 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data,
2252 offset);
2253 offset += BP_GET_PSIZE(gbp);
2254 }
2255 }
2256
2257 if (gn == gio->io_gang_tree)
2258 ASSERT3U(gio->io_size, ==, offset);
2259
2260 if (zio != pio)
2261 zio_nowait(zio);
2262 }
2263
2264 static int
2265 zio_gang_assemble(zio_t *zio)
2266 {
2267 blkptr_t *bp = zio->io_bp;
2268
2269 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2270 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2271
2272 zio->io_gang_leader = zio;
2273
2274 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2275
2276 return (ZIO_PIPELINE_CONTINUE);
2277 }
2278
2279 static int
2280 zio_gang_issue(zio_t *zio)
2281 {
2282 blkptr_t *bp = zio->io_bp;
2283
2284 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
2285 return (ZIO_PIPELINE_STOP);
2286
2287 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2288 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2289
2290 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2291 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd,
2292 0);
2293 else
2294 zio_gang_tree_free(&zio->io_gang_tree);
2295
2296 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2297
2298 return (ZIO_PIPELINE_CONTINUE);
2299 }
2300
2301 static void
2302 zio_write_gang_member_ready(zio_t *zio)
2303 {
2304 zio_t *pio = zio_unique_parent(zio);
2305 zio_t *gio = zio->io_gang_leader;
2306 dva_t *cdva = zio->io_bp->blk_dva;
2307 dva_t *pdva = pio->io_bp->blk_dva;
2308 uint64_t asize;
2309
2310 if (BP_IS_HOLE(zio->io_bp))
2311 return;
2312
2313 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2314
2315 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2316 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2317 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2318 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2319 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2320
2321 mutex_enter(&pio->io_lock);
2322 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2323 ASSERT(DVA_GET_GANG(&pdva[d]));
2324 asize = DVA_GET_ASIZE(&pdva[d]);
2325 asize += DVA_GET_ASIZE(&cdva[d]);
2326 DVA_SET_ASIZE(&pdva[d], asize);
2327 }
2328 mutex_exit(&pio->io_lock);
2329 }
2330
2331 static void
2332 zio_write_gang_done(zio_t *zio)
2333 {
2334 abd_put(zio->io_abd);
2335 }
2336
2337 static int
2338 zio_write_gang_block(zio_t *pio)
2339 {
2340 spa_t *spa = pio->io_spa;
2341 metaslab_class_t *mc = pio->io_mc;
2342 blkptr_t *bp = pio->io_bp;
2343 zio_t *gio = pio->io_gang_leader;
2344 zio_t *zio;
2345 zio_gang_node_t *gn, **gnpp;
2346 zio_gbh_phys_t *gbh;
2347 abd_t *gbh_abd;
2348 uint64_t txg = pio->io_txg;
2349 uint64_t resid = pio->io_size;
2350 uint64_t lsize;
2351 int copies = gio->io_prop.zp_copies;
2352 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2353 zio_prop_t zp;
2354 int error;
2355
2356 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2357 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2358 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2359 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2360
2361 flags |= METASLAB_ASYNC_ALLOC;
2362 VERIFY(refcount_held(&mc->mc_alloc_slots, pio));
2363
2364 /*
2365 * The logical zio has already placed a reservation for
2366 * 'copies' allocation slots but gang blocks may require
2367 * additional copies. These additional copies
2368 * (i.e. gbh_copies - copies) are guaranteed to succeed
2369 * since metaslab_class_throttle_reserve() always allows
2370 * additional reservations for gang blocks.
2371 */
2372 VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
2373 pio, flags));
2374 }
2375
2376 error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
2377 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags,
2378 &pio->io_alloc_list, pio);
2379 if (error) {
2380 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2381 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2382 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2383
2384 /*
2385 * If we failed to allocate the gang block header then
2386 * we remove any additional allocation reservations that
2387 * we placed here. The original reservation will
2388 * be removed when the logical I/O goes to the ready
2389 * stage.
2390 */
2391 metaslab_class_throttle_unreserve(mc,
2392 gbh_copies - copies, pio);
2393 }
2394 pio->io_error = error;
2395 return (ZIO_PIPELINE_CONTINUE);
2396 }
2397
2398 if (pio == gio) {
2399 gnpp = &gio->io_gang_tree;
2400 } else {
2401 gnpp = pio->io_private;
2402 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2403 }
2404
2405 gn = zio_gang_node_alloc(gnpp);
2406 gbh = gn->gn_gbh;
2407 bzero(gbh, SPA_GANGBLOCKSIZE);
2408 gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE);
2409
2410 /*
2411 * Create the gang header.
2412 */
2413 zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2414 zio_write_gang_done, NULL, pio->io_priority,
2415 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2416
2417 /*
2418 * Create and nowait the gang children.
2419 */
2420 for (int g = 0; resid != 0; resid -= lsize, g++) {
2421 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2422 SPA_MINBLOCKSIZE);
2423 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2424
2425 zp.zp_checksum = gio->io_prop.zp_checksum;
2426 zp.zp_compress = ZIO_COMPRESS_OFF;
2427 zp.zp_type = DMU_OT_NONE;
2428 zp.zp_level = 0;
2429 zp.zp_copies = gio->io_prop.zp_copies;
2430 zp.zp_dedup = B_FALSE;
2431 zp.zp_dedup_verify = B_FALSE;
2432 zp.zp_nopwrite = B_FALSE;
2433
2434 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2435 abd_get_offset(pio->io_abd, pio->io_size - resid), lsize,
2436 lsize, &zp, zio_write_gang_member_ready, NULL, NULL,
2437 zio_write_gang_done, &gn->gn_child[g], pio->io_priority,
2438 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark,
2439 &pio->io_smartcomp);
2440
2441 cio->io_mc = mc;
2442
2443 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2444 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2445 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2446
2447 /*
2448 * Gang children won't throttle but we should
2449 * account for their work, so reserve an allocation
2450 * slot for them here.
2451 */
2452 VERIFY(metaslab_class_throttle_reserve(mc,
2453 zp.zp_copies, cio, flags));
2454 }
2455 zio_nowait(cio);
2456 }
2457
2458 /*
2459 * Set pio's pipeline to just wait for zio to finish.
2460 */
2461 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2462
2463 zio_nowait(zio);
2464
2465 return (ZIO_PIPELINE_CONTINUE);
2466 }
2467
2468 /*
2469 * The zio_nop_write stage in the pipeline determines if allocating a
2470 * new bp is necessary. The nopwrite feature can handle writes in
2471 * either syncing or open context (i.e. zil writes) and as a result is
2472 * mutually exclusive with dedup.
2473 *
2474 * By leveraging a cryptographically secure checksum, such as SHA256, we
2475 * can compare the checksums of the new data and the old to determine if
2476 * allocating a new block is required. Note that our requirements for
2477 * cryptographic strength are fairly weak: there can't be any accidental
2478 * hash collisions, but we don't need to be secure against intentional
2479 * (malicious) collisions. To trigger a nopwrite, you have to be able
2480 * to write the file to begin with, and triggering an incorrect (hash
2481 * collision) nopwrite is no worse than simply writing to the file.
2482 * That said, there are no known attacks against the checksum algorithms
2483 * used for nopwrite, assuming that the salt and the checksums
2484 * themselves remain secret.
2485 */
2486 static int
2487 zio_nop_write(zio_t *zio)
2488 {
2489 blkptr_t *bp = zio->io_bp;
2490 blkptr_t *bp_orig = &zio->io_bp_orig;
2491 zio_prop_t *zp = &zio->io_prop;
2492
2493 ASSERT(BP_GET_LEVEL(bp) == 0);
2494 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2495 ASSERT(zp->zp_nopwrite);
2496 ASSERT(!zp->zp_dedup);
2497 ASSERT(zio->io_bp_override == NULL);
2498 ASSERT(IO_IS_ALLOCATING(zio));
2499
2500 /*
2501 * Check to see if the original bp and the new bp have matching
2502 * characteristics (i.e. same checksum, compression algorithms, etc).
2503 * If they don't then just continue with the pipeline which will
2504 * allocate a new bp.
2505 */
2506 if (BP_IS_HOLE(bp_orig) ||
2507 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2508 ZCHECKSUM_FLAG_NOPWRITE) ||
2509 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2510 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2511 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2512 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2513 return (ZIO_PIPELINE_CONTINUE);
2514
2515 /*
2516 * If the checksums match then reset the pipeline so that we
2517 * avoid allocating a new bp and issuing any I/O.
2518 */
2519 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2520 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2521 ZCHECKSUM_FLAG_NOPWRITE);
2522 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2523 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2524 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2525 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2526 sizeof (uint64_t)) == 0);
2527
2528 *bp = *bp_orig;
2529 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2530 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2531 }
2532
2533 return (ZIO_PIPELINE_CONTINUE);
2534 }
2535
2536 /*
2537 * ==========================================================================
2538 * Dedup
2539 * ==========================================================================
2540 */
2541 static void
2542 zio_ddt_child_read_done(zio_t *zio)
2543 {
2544 blkptr_t *bp = zio->io_bp;
2545 ddt_entry_t *dde = zio->io_private;
2546 ddt_phys_t *ddp;
2547 zio_t *pio = zio_unique_parent(zio);
2548
2549 mutex_enter(&pio->io_lock);
2550 ddp = ddt_phys_select(dde, bp);
2551 if (zio->io_error == 0)
2552 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2553
2554 if (zio->io_error == 0 && dde->dde_repair_abd == NULL)
2555 dde->dde_repair_abd = zio->io_abd;
2556 else
2557 abd_free(zio->io_abd);
2558 mutex_exit(&pio->io_lock);
2559 }
2560
2561 static int
2562 zio_ddt_read_start(zio_t *zio)
2563 {
2564 blkptr_t *bp = zio->io_bp;
2565
2566 ASSERT(BP_GET_DEDUP(bp));
2567 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2568 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2569
2570 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2571 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2572 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2573 ddt_phys_t *ddp = dde->dde_phys;
2574 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2575 blkptr_t blk;
2576
2577 ASSERT(zio->io_vsd == NULL);
2578 zio->io_vsd = dde;
2579
2580 if (ddp_self == NULL)
2581 return (ZIO_PIPELINE_CONTINUE);
2582
2583 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2584 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2585 continue;
2586 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2587 &blk);
2588 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2589 abd_alloc_for_io(zio->io_size, B_TRUE),
2590 zio->io_size, zio_ddt_child_read_done, dde,
2591 zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) |
2592 ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark));
2593 }
2594 return (ZIO_PIPELINE_CONTINUE);
2595 }
2596
2597 zio_nowait(zio_read(zio, zio->io_spa, bp,
2598 zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority,
2599 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2600
2601 return (ZIO_PIPELINE_CONTINUE);
2602 }
2603
2604 static int
2605 zio_ddt_read_done(zio_t *zio)
2606 {
2607 blkptr_t *bp = zio->io_bp;
2608
2609 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2610 return (ZIO_PIPELINE_STOP);
2611
2612 ASSERT(BP_GET_DEDUP(bp));
2613 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2614 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2615
2616 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2617 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2618 ddt_entry_t *dde = zio->io_vsd;
2619 if (ddt == NULL) {
2620 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2621 return (ZIO_PIPELINE_CONTINUE);
2622 }
2623 if (dde == NULL) {
2624 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2625 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2626 return (ZIO_PIPELINE_STOP);
2627 }
2628 if (dde->dde_repair_abd != NULL) {
2629 abd_copy(zio->io_abd, dde->dde_repair_abd,
2630 zio->io_size);
2631 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2632 }
2633 ddt_repair_done(ddt, dde);
2634 zio->io_vsd = NULL;
2635 }
2636
2637 ASSERT(zio->io_vsd == NULL);
2638
2639 return (ZIO_PIPELINE_CONTINUE);
2640 }
2641
2642 /* ARGSUSED */
2643 static boolean_t
2644 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2645 {
2646 spa_t *spa = zio->io_spa;
2647 boolean_t do_raw = (zio->io_flags & ZIO_FLAG_RAW);
2648
2649 /* We should never get a raw, override zio */
2650 ASSERT(!(zio->io_bp_override && do_raw));
2651
2652 /*
2653 * Note: we compare the original data, not the transformed data,
2654 * because when zio->io_bp is an override bp, we will not have
2655 * pushed the I/O transforms. That's an important optimization
2656 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2657 */
2658 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2659 zio_t *lio = dde->dde_lead_zio[p];
2660
2661 if (lio != NULL) {
2662 return (lio->io_orig_size != zio->io_orig_size ||
2663 abd_cmp(zio->io_orig_abd, lio->io_orig_abd,
2664 zio->io_orig_size) != 0);
2665 }
2666 }
2667
2668 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2669 ddt_phys_t *ddp = &dde->dde_phys[p];
2670
2671 if (ddp->ddp_phys_birth != 0) {
2672 arc_buf_t *abuf = NULL;
2673 arc_flags_t aflags = ARC_FLAG_WAIT;
2674 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
2675 blkptr_t blk = *zio->io_bp;
2676 int error;
2677
2678 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2679
2680 dde_exit(dde);
2681
2682 /*
2683 * Intuitively, it would make more sense to compare
2684 * io_abd than io_orig_abd in the raw case since you
2685 * don't want to look at any transformations that have
2686 * happened to the data. However, for raw I/Os the
2687 * data will actually be the same in io_abd and
2688 * io_orig_abd, so all we have to do is issue this as
2689 * a raw ARC read.
2690 */
2691 if (do_raw) {
2692 zio_flags |= ZIO_FLAG_RAW;
2693 ASSERT3U(zio->io_size, ==, zio->io_orig_size);
2694 ASSERT0(abd_cmp(zio->io_abd, zio->io_orig_abd,
2695 zio->io_size));
2696 ASSERT3P(zio->io_transform_stack, ==, NULL);
2697 }
2698
2699 error = arc_read(NULL, spa, &blk,
2700 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2701 zio_flags, &aflags, &zio->io_bookmark);
2702
2703 if (error == 0) {
2704 if (arc_buf_size(abuf) != zio->io_orig_size ||
2705 abd_cmp_buf(zio->io_orig_abd, abuf->b_data,
2706 zio->io_orig_size) != 0)
2707 error = SET_ERROR(EEXIST);
2708 arc_buf_destroy(abuf, &abuf);
2709 }
2710
2711 dde_enter(dde);
2712 return (error != 0);
2713 }
2714 }
2715
2716 return (B_FALSE);
2717 }
2718
2719 static void
2720 zio_ddt_child_write_ready(zio_t *zio)
2721 {
2722 int p = zio->io_prop.zp_copies;
2723 ddt_entry_t *dde = zio->io_private;
2724 ddt_phys_t *ddp = &dde->dde_phys[p];
2725 zio_t *pio;
2726
2727 if (zio->io_error)
2728 return;
2729
2730 dde_enter(dde);
2731
2732 ASSERT(dde->dde_lead_zio[p] == zio);
2733
2734 ddt_phys_fill(ddp, zio->io_bp);
2735
2736 zio_link_t *zl = NULL;
2737 while ((pio = zio_walk_parents(zio, &zl)) != NULL)
2738 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2739
2740 dde_exit(dde);
2741 }
2742
2743 static void
2744 zio_ddt_child_write_done(zio_t *zio)
2745 {
2746 int p = zio->io_prop.zp_copies;
2747 ddt_entry_t *dde = zio->io_private;
2748 ddt_phys_t *ddp = &dde->dde_phys[p];
2749
2750 dde_enter(dde);
2751
2752 ASSERT(ddp->ddp_refcnt == 0);
2753 ASSERT(dde->dde_lead_zio[p] == zio);
2754 dde->dde_lead_zio[p] = NULL;
2755
2756 if (zio->io_error == 0) {
2757 zio_link_t *zl = NULL;
2758 while (zio_walk_parents(zio, &zl) != NULL)
2759 ddt_phys_addref(ddp);
2760 } else {
2761 ddt_phys_clear(ddp);
2762 }
2763
2764 dde_exit(dde);
2765 }
2766
2767 static void
2768 zio_ddt_ditto_write_done(zio_t *zio)
2769 {
2770 int p = DDT_PHYS_DITTO;
2771 zio_prop_t *zp = &zio->io_prop;
2772 blkptr_t *bp = zio->io_bp;
2773 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2774 ddt_entry_t *dde = zio->io_private;
2775 ddt_phys_t *ddp = &dde->dde_phys[p];
2776 ddt_key_t *ddk = &dde->dde_key;
2777
2778 dde_enter(dde);
2779
2780 ASSERT(ddp->ddp_refcnt == 0);
2781 ASSERT(dde->dde_lead_zio[p] == zio);
2782 dde->dde_lead_zio[p] = NULL;
2783
2784 if (zio->io_error == 0) {
2785 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2786 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2787 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2788 if (ddp->ddp_phys_birth != 0)
2789 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2790 ddt_phys_fill(ddp, bp);
2791 }
2792
2793 dde_exit(dde);
2794 }
2795
2796 static int
2797 zio_ddt_write(zio_t *zio)
2798 {
2799 spa_t *spa = zio->io_spa;
2800 blkptr_t *bp = zio->io_bp;
2801 uint64_t txg = zio->io_txg;
2802 zio_prop_t *zp = &zio->io_prop;
2803 int p = zp->zp_copies;
2804 int ditto_copies;
2805 zio_t *cio = NULL;
2806 zio_t *dio = NULL;
2807 ddt_t *ddt = ddt_select(spa, bp);
2808 ddt_entry_t *dde;
2809 ddt_phys_t *ddp;
2810
2811 ASSERT(BP_GET_DEDUP(bp));
2812 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2813 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2814 ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
2815
2816 dde = ddt_lookup(ddt, bp, B_TRUE);
2817
2818 /*
2819 * If we're not using special tier, for each new DDE that's not on disk:
2820 * disable dedup if we have exhausted "allowed" DDT L2/ARC space
2821 */
2822 if ((dde->dde_state & DDE_NEW) && !spa->spa_usesc &&
2823 (zfs_ddt_limit_type != DDT_NO_LIMIT || zfs_ddt_byte_ceiling != 0)) {
2824 /* turn off dedup if we need to stop DDT growth */
2825 if (spa_enable_dedup_cap(spa)) {
2826 dde->dde_state |= DDE_DONT_SYNC;
2827
2828 /* disable dedup and use the ordinary write pipeline */
2829 zio_pop_transforms(zio);
2830 zp->zp_dedup = zp->zp_dedup_verify = B_FALSE;
2831 zio->io_stage = ZIO_STAGE_OPEN;
2832 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2833 zio->io_bp_override = NULL;
2834 BP_ZERO(bp);
2835 dde_exit(dde);
2836
2837 return (ZIO_PIPELINE_CONTINUE);
2838 }
2839 }
2840 ASSERT(!(dde->dde_state & DDE_DONT_SYNC));
2841
2842 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2843 /*
2844 * If we're using a weak checksum, upgrade to a strong checksum
2845 * and try again. If we're already using a strong checksum,
2846 * we can't resolve it, so just convert to an ordinary write.
2847 * (And automatically e-mail a paper to Nature?)
2848 */
2849 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
2850 ZCHECKSUM_FLAG_DEDUP)) {
2851 zp->zp_checksum = spa_dedup_checksum(spa);
2852 zio_pop_transforms(zio);
2853 zio->io_stage = ZIO_STAGE_OPEN;
2854 BP_ZERO(bp);
2855 } else {
2856 zp->zp_dedup = B_FALSE;
2857 BP_SET_DEDUP(bp, B_FALSE);
2858 }
2859 ASSERT(!BP_GET_DEDUP(bp));
2860 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2861 dde_exit(dde);
2862 return (ZIO_PIPELINE_CONTINUE);
2863 }
2864
2865 ddp = &dde->dde_phys[p];
2866 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2867 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2868
2869 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2870 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2871 zio_prop_t czp = *zp;
2872
2873 czp.zp_copies = ditto_copies;
2874
2875 /*
2876 * If we arrived here with an override bp, we won't have run
2877 * the transform stack, so we won't have the data we need to
2878 * generate a child i/o. So, toss the override bp and restart.
2879 * This is safe, because using the override bp is just an
2880 * optimization; and it's rare, so the cost doesn't matter.
2881 */
2882 if (zio->io_bp_override) {
2883 zio_pop_transforms(zio);
2884 zio->io_stage = ZIO_STAGE_OPEN;
2885 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2886 zio->io_bp_override = NULL;
2887 BP_ZERO(bp);
2888 dde_exit(dde);
2889 return (ZIO_PIPELINE_CONTINUE);
2890 }
2891
2892 dio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2893 zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL,
2894 NULL, zio_ddt_ditto_write_done, dde, zio->io_priority,
2895 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark, NULL);
2896
2897 zio_push_transform(dio, zio->io_abd, zio->io_size, 0, NULL);
2898 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2899 }
2900
2901 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2902 if (ddp->ddp_phys_birth != 0)
2903 ddt_bp_fill(ddp, bp, txg);
2904 if (dde->dde_lead_zio[p] != NULL)
2905 zio_add_child(zio, dde->dde_lead_zio[p]);
2906 else
2907 ddt_phys_addref(ddp);
2908 } else if (zio->io_bp_override) {
2909 ASSERT(bp->blk_birth == txg);
2910 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2911 ddt_phys_fill(ddp, bp);
2912 ddt_phys_addref(ddp);
2913 } else {
2914 cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2915 zio->io_orig_size, zio->io_orig_size, zp,
2916 zio_ddt_child_write_ready, NULL, NULL,
2917 zio_ddt_child_write_done, dde, zio->io_priority,
2918 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark, NULL);
2919
2920 zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL);
2921 dde->dde_lead_zio[p] = cio;
2922 }
2923
2924 dde_exit(dde);
2925
2926 if (cio)
2927 zio_nowait(cio);
2928 if (dio)
2929 zio_nowait(dio);
2930
2931 return (ZIO_PIPELINE_CONTINUE);
2932 }
2933
2934 ddt_entry_t *freedde; /* for debugging */
2935
2936 static int
2937 zio_ddt_free(zio_t *zio)
2938 {
2939 spa_t *spa = zio->io_spa;
2940 blkptr_t *bp = zio->io_bp;
2941 ddt_t *ddt = ddt_select(spa, bp);
2942 ddt_entry_t *dde;
2943 ddt_phys_t *ddp;
2944
2945 ASSERT(BP_GET_DEDUP(bp));
2946 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2947
2948 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2949 ddp = ddt_phys_select(dde, bp);
2950 if (ddp)
2951 ddt_phys_decref(ddp);
2952 dde_exit(dde);
2953
2954 return (ZIO_PIPELINE_CONTINUE);
2955 }
2956
2957 /*
2958 * ==========================================================================
2959 * Allocate and free blocks
2960 * ==========================================================================
2961 */
2962
2963 static zio_t *
2964 zio_io_to_allocate(metaslab_class_t *mc)
2965 {
2966 zio_t *zio;
2967
2968 ASSERT(MUTEX_HELD(&mc->mc_alloc_lock));
2969
2970 zio = avl_first(&mc->mc_alloc_tree);
2971 if (zio == NULL)
2972 return (NULL);
2973
2974 ASSERT(IO_IS_ALLOCATING(zio));
2975
2976 /*
2977 * Try to place a reservation for this zio. If we're unable to
2978 * reserve then we throttle.
2979 */
2980 if (!metaslab_class_throttle_reserve(mc,
2981 zio->io_prop.zp_copies, zio, 0)) {
2982 return (NULL);
2983 }
2984
2985 avl_remove(&mc->mc_alloc_tree, zio);
2986 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
2987
2988 return (zio);
2989 }
2990
2991 static int
2992 zio_dva_throttle(zio_t *zio)
2993 {
2994 spa_t *spa = zio->io_spa;
2995 zio_t *nio;
2996
2997 /* We need to use parent's MetaslabClass */
2998 if (zio->io_mc == NULL) {
2999 zio->io_mc = spa_select_class(spa, zio);
3000 if (zio->io_prop.zp_usewbc)
3001 return (ZIO_PIPELINE_CONTINUE);
3002 }
3003
3004 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
3005 !zio->io_mc->mc_alloc_throttle_enabled ||
3006 zio->io_child_type == ZIO_CHILD_GANG ||
3007 zio->io_flags & ZIO_FLAG_NODATA) {
3008 return (ZIO_PIPELINE_CONTINUE);
3009 }
3010
3011 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3012
3013 ASSERT3U(zio->io_queued_timestamp, >, 0);
3014 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
3015
3016 mutex_enter(&zio->io_mc->mc_alloc_lock);
3017
3018 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3019 avl_add(&zio->io_mc->mc_alloc_tree, zio);
3020
3021 nio = zio_io_to_allocate(zio->io_mc);
3022 mutex_exit(&zio->io_mc->mc_alloc_lock);
3023
3024 if (nio == zio)
3025 return (ZIO_PIPELINE_CONTINUE);
3026
3027 if (nio != NULL) {
3028 ASSERT(nio->io_stage == ZIO_STAGE_DVA_THROTTLE);
3029 /*
3030 * We are passing control to a new zio so make sure that
3031 * it is processed by a different thread. We do this to
3032 * avoid stack overflows that can occur when parents are
3033 * throttled and children are making progress. We allow
3034 * it to go to the head of the taskq since it's already
3035 * been waiting.
3036 */
3037 zio_taskq_dispatch(nio, ZIO_TASKQ_ISSUE, B_TRUE);
3038 }
3039 return (ZIO_PIPELINE_STOP);
3040 }
3041
3042 void
3043 zio_allocate_dispatch(metaslab_class_t *mc)
3044 {
3045 zio_t *zio;
3046
3047 mutex_enter(&mc->mc_alloc_lock);
3048 zio = zio_io_to_allocate(mc);
3049 mutex_exit(&mc->mc_alloc_lock);
3050 if (zio == NULL)
3051 return;
3052
3053 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
3054 ASSERT0(zio->io_error);
3055 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
3056 }
3057
3058 static int
3059 zio_dva_allocate(zio_t *zio)
3060 {
3061 spa_t *spa = zio->io_spa;
3062 metaslab_class_t *mc = zio->io_mc;
3063
3064 blkptr_t *bp = zio->io_bp;
3065 int error;
3066 int flags = 0;
3067
3068 if (zio->io_gang_leader == NULL) {
3069 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3070 zio->io_gang_leader = zio;
3071 }
3072
3073 ASSERT(BP_IS_HOLE(bp));
3074 ASSERT0(BP_GET_NDVAS(bp));
3075 ASSERT3U(zio->io_prop.zp_copies, >, 0);
3076 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
3077 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
3078
3079 if (zio->io_flags & ZIO_FLAG_NODATA || zio->io_prop.zp_usewbc) {
3080 flags |= METASLAB_DONT_THROTTLE;
3081 }
3082 if (zio->io_flags & ZIO_FLAG_GANG_CHILD) {
3083 flags |= METASLAB_GANG_CHILD;
3084 }
3085 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE &&
3086 zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3087 flags |= METASLAB_ASYNC_ALLOC;
3088 }
3089
3090 error = metaslab_alloc(spa, mc, zio->io_size, bp,
3091 zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3092 &zio->io_alloc_list, zio);
3093
3094 #ifdef _KERNEL
3095 DTRACE_PROBE6(zio_dva_allocate,
3096 uint64_t, DVA_GET_VDEV(&bp->blk_dva[0]),
3097 uint64_t, DVA_GET_VDEV(&bp->blk_dva[1]),
3098 uint64_t, BP_GET_LEVEL(bp),
3099 boolean_t, BP_IS_SPECIAL(bp),
3100 boolean_t, BP_IS_METADATA(bp),
3101 int, error);
3102 #endif
3103
3104 if (error != 0) {
3105 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
3106 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
3107 error);
3108 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) {
3109 if (zio->io_prop.zp_usewbc) {
3110 zio->io_prop.zp_usewbc = B_FALSE;
3111 zio->io_prop.zp_usesc = B_FALSE;
3112 zio->io_mc = spa_normal_class(spa);
3113 }
3114
3115 return (zio_write_gang_block(zio));
3116 }
3117
3118 zio->io_error = error;
3119 }
3120
3121 return (ZIO_PIPELINE_CONTINUE);
3122 }
3123
3124 static int
3125 zio_dva_free(zio_t *zio)
3126 {
3127 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
3128
3129 return (ZIO_PIPELINE_CONTINUE);
3130 }
3131
3132 static int
3133 zio_dva_claim(zio_t *zio)
3134 {
3135 int error;
3136
3137 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
3138 if (error)
3139 zio->io_error = error;
3140
3141 return (ZIO_PIPELINE_CONTINUE);
3142 }
3143
3144 /*
3145 * Undo an allocation. This is used by zio_done() when an I/O fails
3146 * and we want to give back the block we just allocated.
3147 * This handles both normal blocks and gang blocks.
3148 */
3149 static void
3150 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
3151 {
3152 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
3153 ASSERT(zio->io_bp_override == NULL);
3154
3155 if (!BP_IS_HOLE(bp))
3156 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
3157
3158 if (gn != NULL) {
3159 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
3160 zio_dva_unallocate(zio, gn->gn_child[g],
3161 &gn->gn_gbh->zg_blkptr[g]);
3162 }
3163 }
3164 }
3165
3166 /*
3167 * Try to allocate an intent log block. Return 0 on success, errno on failure.
3168 */
3169 int
3170 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
3171 uint64_t size, boolean_t *slog)
3172 {
3173 int error = 1;
3174 zio_alloc_list_t io_alloc_list;
3175 spa_meta_placement_t *mp = &spa->spa_meta_policy;
3176
3177 ASSERT(txg > spa_syncing_txg(spa));
3178
3179 metaslab_trace_init(&io_alloc_list);
3180
3181 /*
3182 * ZIL blocks are always contiguous (i.e. not gang blocks)
3183 * so we set the METASLAB_HINTBP_AVOID flag so that they
3184 * don't "fast gang" when allocating them.
3185 * If the caller indicates that slog is not to be used
3186 * (via use_slog)
3187 * separate allocation class will not indeed be used,
3188 * independently of whether this is log or special
3189 */
3190
3191 if (spa_has_slogs(spa)) {
3192 error = metaslab_alloc(spa, spa_log_class(spa),
3193 size, new_bp, 1, txg, old_bp,
3194 METASLAB_HINTBP_AVOID, &io_alloc_list, NULL);
3195
3196 DTRACE_PROBE2(zio_alloc_zil_log,
3197 spa_t *, spa, int, error);
3198
3199 if (error == 0)
3200 *slog = TRUE;
3201 }
3202
3203 /*
3204 * use special when failed to allocate from the regular
3205 * slog, but only if allowed and if the special used
3206 * space is below watermarks
3207 */
3208 if (error != 0 && spa_can_special_be_used(spa) &&
3209 mp->spa_sync_to_special != SYNC_TO_SPECIAL_DISABLED) {
3210 error = metaslab_alloc(spa, spa_special_class(spa),
3211 size, new_bp, 1, txg, old_bp,
3212 METASLAB_HINTBP_AVOID, &io_alloc_list, NULL);
3213
3214 DTRACE_PROBE2(zio_alloc_zil_special,
3215 spa_t *, spa, int, error);
3216
3217 if (error == 0)
3218 *slog = FALSE;
3219 }
3220
3221 if (error != 0) {
3222 error = metaslab_alloc(spa, spa_normal_class(spa), size,
3223 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID,
3224 &io_alloc_list, NULL);
3225
3226 DTRACE_PROBE2(zio_alloc_zil_normal,
3227 spa_t *, spa, int, error);
3228
3229 if (error == 0)
3230 *slog = FALSE;
3231 }
3232
3233 metaslab_trace_fini(&io_alloc_list);
3234
3235 if (error == 0) {
3236 BP_SET_LSIZE(new_bp, size);
3237 BP_SET_PSIZE(new_bp, size);
3238 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
3239 BP_SET_CHECKSUM(new_bp,
3240 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
3241 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
3242 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3243 BP_SET_LEVEL(new_bp, 0);
3244 BP_SET_DEDUP(new_bp, 0);
3245 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
3246 } else {
3247 zfs_dbgmsg("%s: zil block allocation failure: "
3248 "size %llu, error %d", spa_name(spa), size, error);
3249 }
3250
3251 return (error);
3252 }
3253
3254 /*
3255 * Free an intent log block.
3256 */
3257 void
3258 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
3259 {
3260 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
3261 ASSERT(!BP_IS_GANG(bp));
3262
3263 zio_free(spa, txg, bp);
3264 }
3265
3266 /*
3267 * ==========================================================================
3268 * Read and write to physical devices
3269 * ==========================================================================
3270 */
3271
3272
3273 /*
3274 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3275 * stops after this stage and will resume upon I/O completion.
3276 * However, there are instances where the vdev layer may need to
3277 * continue the pipeline when an I/O was not issued. Since the I/O
3278 * that was sent to the vdev layer might be different than the one
3279 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3280 * force the underlying vdev layers to call either zio_execute() or
3281 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3282 */
3283 static int
3284 zio_vdev_io_start(zio_t *zio)
3285 {
3286 vdev_t *vd = zio->io_vd;
3287 uint64_t align;
3288 spa_t *spa = zio->io_spa;
3289 zio_type_t type = zio->io_type;
3290 zio->io_vd_timestamp = gethrtime();
3291
3292 ASSERT(zio->io_error == 0);
3293 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3294
3295 if (vd == NULL) {
3296 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3297 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3298
3299 /*
3300 * The mirror_ops handle multiple DVAs in a single BP.
3301 */
3302 vdev_mirror_ops.vdev_op_io_start(zio);
3303 return (ZIO_PIPELINE_STOP);
3304 }
3305
3306 ASSERT3P(zio->io_logical, !=, zio);
3307
3308 align = 1ULL << vd->vdev_top->vdev_ashift;
3309
3310 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3311 P2PHASE(zio->io_size, align) != 0) {
3312 /* Transform logical writes to be a full physical block size. */
3313 uint64_t asize = P2ROUNDUP(zio->io_size, align);
3314 abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize);
3315 ASSERT(vd == vd->vdev_top);
3316 if (type == ZIO_TYPE_WRITE) {
3317 abd_copy(abuf, zio->io_abd, zio->io_size);
3318 abd_zero_off(abuf, zio->io_size, asize - zio->io_size);
3319 }
3320 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
3321 }
3322
3323 /*
3324 * If this is not a physical io, make sure that it is properly aligned
3325 * before proceeding.
3326 */
3327 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3328 ASSERT0(P2PHASE(zio->io_offset, align));
3329 ASSERT0(P2PHASE(zio->io_size, align));
3330 } else {
3331 /*
3332 * For physical writes, we allow 512b aligned writes and assume
3333 * the device will perform a read-modify-write as necessary.
3334 */
3335 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
3336 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
3337 }
3338
3339 VERIFY(type != ZIO_TYPE_WRITE || spa_writeable(spa));
3340
3341 /*
3342 * If this is a repair I/O, and there's no self-healing involved --
3343 * that is, we're just resilvering what we expect to resilver --
3344 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3345 * This prevents spurious resilvering with nested replication.
3346 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
3347 * A is out of date, we'll read from C+D, then use the data to
3348 * resilver A+B -- but we don't actually want to resilver B, just A.
3349 * The top-level mirror has no way to know this, so instead we just
3350 * discard unnecessary repairs as we work our way down the vdev tree.
3351 * The same logic applies to any form of nested replication:
3352 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
3353 */
3354 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3355 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3356 zio->io_txg != 0 && /* not a delegated i/o */
3357 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3358 ASSERT(type == ZIO_TYPE_WRITE);
3359 zio_vdev_io_bypass(zio);
3360 return (ZIO_PIPELINE_CONTINUE);
3361 }
3362
3363 if (vd->vdev_ops->vdev_op_leaf &&
3364 (type == ZIO_TYPE_READ || type == ZIO_TYPE_WRITE)) {
3365 if (type == ZIO_TYPE_READ && vdev_cache_read(zio))
3366 return (ZIO_PIPELINE_CONTINUE);
3367
3368 if ((zio = vdev_queue_io(zio)) == NULL)
3369 return (ZIO_PIPELINE_STOP);
3370
3371 if (!vdev_accessible(vd, zio)) {
3372 zio->io_error = SET_ERROR(ENXIO);
3373 zio_interrupt(zio);
3374 return (ZIO_PIPELINE_STOP);
3375 }
3376
3377 /*
3378 * Insert a fault simulation delay for a particular vdev.
3379 */
3380 if (zio_faulty_vdev_enabled &&
3381 (zio->io_vd->vdev_guid == zio_faulty_vdev_guid)) {
3382 delay(NSEC_TO_TICK(zio_faulty_vdev_delay_us *
3383 (NANOSEC / MICROSEC)));
3384 }
3385 }
3386
3387 vd->vdev_ops->vdev_op_io_start(zio);
3388 return (ZIO_PIPELINE_STOP);
3389 }
3390
3391 static int
3392 zio_vdev_io_done(zio_t *zio)
3393 {
3394 vdev_t *vd = zio->io_vd;
3395 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3396 boolean_t unexpected_error = B_FALSE;
3397
3398 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
3399 return (ZIO_PIPELINE_STOP);
3400
3401 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
3402
3403 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
3404 vdev_queue_io_done(zio);
3405
3406 if (zio->io_type == ZIO_TYPE_WRITE)
3407 vdev_cache_write(zio);
3408
3409 if (zio_injection_enabled && zio->io_error == 0)
3410 zio->io_error = zio_handle_device_injection(vd,
3411 zio, EIO);
3412
3413 if (zio_injection_enabled && zio->io_error == 0)
3414 zio->io_error = zio_handle_label_injection(zio, EIO);
3415
3416 if (zio->io_error) {
3417 if (!vdev_accessible(vd, zio)) {
3418 zio->io_error = SET_ERROR(ENXIO);
3419 } else {
3420 unexpected_error = B_TRUE;
3421 }
3422 }
3423 }
3424
3425 ops->vdev_op_io_done(zio);
3426
3427 if (unexpected_error)
3428 VERIFY(vdev_probe(vd, zio) == NULL);
3429
3430 /*
3431 * Measure delta between start and end of the I/O in nanoseconds.
3432 * XXX: Handle overflow.
3433 */
3434 zio->io_vd_timestamp = gethrtime() - zio->io_vd_timestamp;
3435
3436 return (ZIO_PIPELINE_CONTINUE);
3437 }
3438
3439 /*
3440 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3441 * disk, and use that to finish the checksum ereport later.
3442 */
3443 static void
3444 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
3445 const void *good_buf)
3446 {
3447 /* no processing needed */
3448 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
3449 }
3450
3451 /*ARGSUSED*/
3452 void
3453 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
3454 {
3455 void *buf = zio_buf_alloc(zio->io_size);
3456
3457 abd_copy_to_buf(buf, zio->io_abd, zio->io_size);
3458
3459 zcr->zcr_cbinfo = zio->io_size;
3460 zcr->zcr_cbdata = buf;
3461 zcr->zcr_finish = zio_vsd_default_cksum_finish;
3462 zcr->zcr_free = zio_buf_free;
3463 }
3464
3465 static int
3466 zio_vdev_io_assess(zio_t *zio)
3467 {
3468 vdev_t *vd = zio->io_vd;
3469
3470 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
3471 return (ZIO_PIPELINE_STOP);
3472
3473 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3474 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
3475
3476 if (zio->io_vsd != NULL) {
3477 zio->io_vsd_ops->vsd_free(zio);
3478 zio->io_vsd = NULL;
3479 }
3480
3481 if (zio_injection_enabled && zio->io_error == 0)
3482 zio->io_error = zio_handle_fault_injection(zio, EIO);
3483
3484 /*
3485 * If the I/O failed, determine whether we should attempt to retry it.
3486 *
3487 * On retry, we cut in line in the issue queue, since we don't want
3488 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3489 */
3490 if (zio->io_error && vd == NULL &&
3491 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
3492 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
3493 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
3494 zio->io_error = 0;
3495 zio->io_flags |= ZIO_FLAG_IO_RETRY |
3496 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
3497 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
3498 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
3499 zio_requeue_io_start_cut_in_line);
3500 return (ZIO_PIPELINE_STOP);
3501 }
3502
3503 /*
3504 * If we got an error on a leaf device, convert it to ENXIO
3505 * if the device is not accessible at all.
3506 */
3507 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3508 !vdev_accessible(vd, zio))
3509 zio->io_error = SET_ERROR(ENXIO);
3510
3511 /*
3512 * If we can't write to an interior vdev (mirror or RAID-Z),
3513 * set vdev_cant_write so that we stop trying to allocate from it.
3514 */
3515 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
3516 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
3517 vd->vdev_cant_write = B_TRUE;
3518 }
3519
3520 /*
3521 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3522 * attempts will ever succeed. In this case we set a persistent bit so
3523 * that we don't bother with it in the future.
3524 */
3525 if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) &&
3526 zio->io_type == ZIO_TYPE_IOCTL &&
3527 zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL)
3528 vd->vdev_nowritecache = B_TRUE;
3529
3530 if (zio->io_error)
3531 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3532
3533 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3534 zio->io_physdone != NULL) {
3535 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
3536 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
3537 zio->io_physdone(zio->io_logical);
3538 }
3539
3540 return (ZIO_PIPELINE_CONTINUE);
3541 }
3542
3543 void
3544 zio_vdev_io_reissue(zio_t *zio)
3545 {
3546 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3547 ASSERT(zio->io_error == 0);
3548
3549 zio->io_stage >>= 1;
3550 }
3551
3552 void
3553 zio_vdev_io_redone(zio_t *zio)
3554 {
3555 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
3556
3557 zio->io_stage >>= 1;
3558 }
3559
3560 void
3561 zio_vdev_io_bypass(zio_t *zio)
3562 {
3563 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3564 ASSERT(zio->io_error == 0);
3565
3566 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
3567 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
3568 }
3569
3570 /*
3571 * ==========================================================================
3572 * Generate and verify checksums
3573 * ==========================================================================
3574 */
3575 static int
3576 zio_checksum_generate(zio_t *zio)
3577 {
3578 blkptr_t *bp = zio->io_bp;
3579 enum zio_checksum checksum;
3580
3581 if (bp == NULL) {
3582 /*
3583 * This is zio_write_phys().
3584 * We're either generating a label checksum, or none at all.
3585 */
3586 checksum = zio->io_prop.zp_checksum;
3587
3588 if (checksum == ZIO_CHECKSUM_OFF)
3589 return (ZIO_PIPELINE_CONTINUE);
3590
3591 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
3592 } else {
3593 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
3594 ASSERT(!IO_IS_ALLOCATING(zio));
3595 checksum = ZIO_CHECKSUM_GANG_HEADER;
3596 } else {
3597 checksum = BP_GET_CHECKSUM(bp);
3598 }
3599 }
3600
3601 zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size);
3602
3603 return (ZIO_PIPELINE_CONTINUE);
3604 }
3605
3606 static int
3607 zio_checksum_verify(zio_t *zio)
3608 {
3609 zio_bad_cksum_t info;
3610 blkptr_t *bp = zio->io_bp;
3611 int error;
3612
3613 ASSERT(zio->io_vd != NULL);
3614
3615 if (bp == NULL) {
3616 /*
3617 * This is zio_read_phys().
3618 * We're either verifying a label checksum, or nothing at all.
3619 */
3620 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
3621 return (ZIO_PIPELINE_CONTINUE);
3622
3623 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
3624 }
3625
3626 if ((error = zio_checksum_error(zio, &info)) != 0) {
3627 zio->io_error = error;
3628 if (error == ECKSUM &&
3629 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
3630 zfs_ereport_start_checksum(zio->io_spa,
3631 zio->io_vd, zio, zio->io_offset,
3632 zio->io_size, NULL, &info);
3633 }
3634 }
3635
3636 return (ZIO_PIPELINE_CONTINUE);
3637 }
3638
3639 /*
3640 * Called by RAID-Z to ensure we don't compute the checksum twice.
3641 */
3642 void
3643 zio_checksum_verified(zio_t *zio)
3644 {
3645 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
3646 }
3647
3648 /*
3649 * ==========================================================================
3650 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3651 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3652 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3653 * indicate errors that are specific to one I/O, and most likely permanent.
3654 * Any other error is presumed to be worse because we weren't expecting it.
3655 * ==========================================================================
3656 */
3657 int
3658 zio_worst_error(int e1, int e2)
3659 {
3660 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
3661 int r1, r2;
3662
3663 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
3664 if (e1 == zio_error_rank[r1])
3665 break;
3666
3667 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3668 if (e2 == zio_error_rank[r2])
3669 break;
3670
3671 return (r1 > r2 ? e1 : e2);
3672 }
3673
3674 /*
3675 * ==========================================================================
3676 * I/O completion
3677 * ==========================================================================
3678 */
3679 static int
3680 zio_ready(zio_t *zio)
3681 {
3682 blkptr_t *bp = zio->io_bp;
3683 zio_t *pio, *pio_next;
3684 zio_link_t *zl = NULL;
3685
3686 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
3687 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
3688 return (ZIO_PIPELINE_STOP);
3689
3690 if (zio->io_ready) {
3691 ASSERT(IO_IS_ALLOCATING(zio));
3692 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3693 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3694 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3695
3696 zio->io_ready(zio);
3697 }
3698
3699 if (bp != NULL && bp != &zio->io_bp_copy)
3700 zio->io_bp_copy = *bp;
3701
3702 if (zio->io_error != 0) {
3703 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3704
3705 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3706 ASSERT(IO_IS_ALLOCATING(zio));
3707 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3708 /*
3709 * We were unable to allocate anything, unreserve and
3710 * issue the next I/O to allocate.
3711 */
3712 metaslab_class_throttle_unreserve(zio->io_mc,
3713 zio->io_prop.zp_copies, zio);
3714 zio_allocate_dispatch(zio->io_mc);
3715 }
3716 }
3717
3718 mutex_enter(&zio->io_lock);
3719 zio->io_state[ZIO_WAIT_READY] = 1;
3720 pio = zio_walk_parents(zio, &zl);
3721 mutex_exit(&zio->io_lock);
3722
3723 /*
3724 * As we notify zio's parents, new parents could be added.
3725 * New parents go to the head of zio's io_parent_list, however,
3726 * so we will (correctly) not notify them. The remainder of zio's
3727 * io_parent_list, from 'pio_next' onward, cannot change because
3728 * all parents must wait for us to be done before they can be done.
3729 */
3730 for (; pio != NULL; pio = pio_next) {
3731 pio_next = zio_walk_parents(zio, &zl);
3732 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3733 }
3734
3735 if (zio->io_flags & ZIO_FLAG_NODATA) {
3736 if (BP_IS_GANG(bp)) {
3737 zio->io_flags &= ~ZIO_FLAG_NODATA;
3738 } else {
3739 ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE);
3740 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3741 }
3742 }
3743
3744 if (zio_injection_enabled &&
3745 zio->io_spa->spa_syncing_txg == zio->io_txg)
3746 zio_handle_ignored_writes(zio);
3747
3748 return (ZIO_PIPELINE_CONTINUE);
3749 }
3750
3751 /*
3752 * Update the allocation throttle accounting.
3753 */
3754 static void
3755 zio_dva_throttle_done(zio_t *zio)
3756 {
3757 zio_t *lio = zio->io_logical;
3758 zio_t *pio = zio_unique_parent(zio);
3759 vdev_t *vd = zio->io_vd;
3760 int flags = METASLAB_ASYNC_ALLOC;
3761
3762 ASSERT3P(zio->io_bp, !=, NULL);
3763 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
3764 ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
3765 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
3766 ASSERT(vd != NULL);
3767 ASSERT3P(vd, ==, vd->vdev_top);
3768 ASSERT(!(zio->io_flags & (ZIO_FLAG_IO_REPAIR | ZIO_FLAG_IO_RETRY)));
3769 ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
3770 ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
3771 ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
3772
3773 /*
3774 * Parents of gang children can have two flavors -- ones that
3775 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
3776 * and ones that allocated the constituent blocks. The allocation
3777 * throttle needs to know the allocating parent zio so we must find
3778 * it here.
3779 */
3780 if (pio->io_child_type == ZIO_CHILD_GANG) {
3781 /*
3782 * If our parent is a rewrite gang child then our grandparent
3783 * would have been the one that performed the allocation.
3784 */
3785 if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
3786 pio = zio_unique_parent(pio);
3787 flags |= METASLAB_GANG_CHILD;
3788 }
3789
3790 ASSERT(IO_IS_ALLOCATING(pio));
3791 ASSERT3P(zio, !=, zio->io_logical);
3792 ASSERT(zio->io_logical != NULL);
3793 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
3794 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
3795
3796 mutex_enter(&pio->io_lock);
3797 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags);
3798 mutex_exit(&pio->io_lock);
3799
3800 metaslab_class_throttle_unreserve(pio->io_mc, 1, pio);
3801
3802 /*
3803 * Call into the pipeline to see if there is more work that
3804 * needs to be done. If there is work to be done it will be
3805 * dispatched to another taskq thread.
3806 */
3807 zio_allocate_dispatch(pio->io_mc);
3808 }
3809
3810 static int
3811 zio_done(zio_t *zio)
3812 {
3813 spa_t *spa = zio->io_spa;
3814 zio_t *lio = zio->io_logical;
3815 blkptr_t *bp = zio->io_bp;
3816 vdev_t *vd = zio->io_vd;
3817 uint64_t psize = zio->io_size;
3818 zio_t *pio, *pio_next;
3819 metaslab_class_t *mc = zio->io_mc;
3820 zio_link_t *zl = NULL;
3821
3822 /*
3823 * If our children haven't all completed,
3824 * wait for them and then repeat this pipeline stage.
3825 */
3826 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
3827 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
3828 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
3829 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
3830 return (ZIO_PIPELINE_STOP);
3831
3832 /*
3833 * If the allocation throttle is enabled, then update the accounting.
3834 * We only track child I/Os that are part of an allocating async
3835 * write. We must do this since the allocation is performed
3836 * by the logical I/O but the actual write is done by child I/Os.
3837 */
3838 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
3839 zio->io_child_type == ZIO_CHILD_VDEV) {
3840 ASSERT(mc->mc_alloc_throttle_enabled);
3841 zio_dva_throttle_done(zio);
3842 }
3843
3844 /*
3845 * If the allocation throttle is enabled, verify that
3846 * we have decremented the refcounts for every I/O that was throttled.
3847 */
3848 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3849 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3850 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3851 ASSERT(bp != NULL);
3852 metaslab_group_alloc_verify(spa, zio->io_bp, zio);
3853 VERIFY(refcount_not_held(&mc->mc_alloc_slots, zio));
3854 }
3855
3856 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3857 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3858 ASSERT(zio->io_children[c][w] == 0);
3859
3860 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3861 ASSERT(bp->blk_pad[0] == 0);
3862 ASSERT(bp->blk_pad[1] == 0);
3863 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3864 (bp == zio_unique_parent(zio)->io_bp));
3865 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3866 zio->io_bp_override == NULL &&
3867 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3868 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3869 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3870 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3871 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3872 }
3873 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3874 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3875 }
3876
3877 /*
3878 * If there were child vdev/gang/ddt errors, they apply to us now.
3879 */
3880 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3881 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3882 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3883
3884 /*
3885 * If the I/O on the transformed data was successful, generate any
3886 * checksum reports now while we still have the transformed data.
3887 */
3888 if (zio->io_error == 0) {
3889 while (zio->io_cksum_report != NULL) {
3890 zio_cksum_report_t *zcr = zio->io_cksum_report;
3891 uint64_t align = zcr->zcr_align;
3892 uint64_t asize = P2ROUNDUP(psize, align);
3893 char *abuf = NULL;
3894 abd_t *adata = zio->io_abd;
3895
3896 if (asize != psize) {
3897 adata = abd_alloc_linear(asize, B_TRUE);
3898 abd_copy(adata, zio->io_abd, psize);
3899 abd_zero_off(adata, psize, asize - psize);
3900 }
3901
3902 if (adata != NULL)
3903 abuf = abd_borrow_buf_copy(adata, asize);
3904
3905 zio->io_cksum_report = zcr->zcr_next;
3906 zcr->zcr_next = NULL;
3907 zcr->zcr_finish(zcr, abuf);
3908 zfs_ereport_free_checksum(zcr);
3909
3910 if (adata != NULL)
3911 abd_return_buf(adata, abuf, asize);
3912
3913 if (asize != psize)
3914 abd_free(adata);
3915 }
3916 }
3917
3918 zio_pop_transforms(zio); /* note: may set zio->io_error */
3919
3920 vdev_stat_update(zio, psize);
3921
3922 if (zio->io_error) {
3923 /*
3924 * If this I/O is attached to a particular vdev,
3925 * generate an error message describing the I/O failure
3926 * at the block level. We ignore these errors if the
3927 * device is currently unavailable.
3928 */
3929 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3930 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3931
3932 if ((zio->io_error == EIO || !(zio->io_flags &
3933 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3934 zio == lio) {
3935 /*
3936 * For logical I/O requests, tell the SPA to log the
3937 * error and generate a logical data ereport.
3938 */
3939 spa_log_error(spa, zio);
3940 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3941 0, 0);
3942 }
3943 }
3944
3945 if (zio->io_error && zio == lio) {
3946 /*
3947 * Determine whether zio should be reexecuted. This will
3948 * propagate all the way to the root via zio_notify_parent().
3949 */
3950 ASSERT(vd == NULL && bp != NULL);
3951 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3952
3953 if (IO_IS_ALLOCATING(zio) &&
3954 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3955 if (zio->io_error != ENOSPC)
3956 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3957 else
3958 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3959 }
3960
3961 if ((zio->io_type == ZIO_TYPE_READ ||
3962 zio->io_type == ZIO_TYPE_FREE) &&
3963 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3964 zio->io_error == ENXIO &&
3965 spa_load_state(spa) == SPA_LOAD_NONE &&
3966 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3967 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3968
3969 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3970 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3971
3972 /*
3973 * Here is a possibly good place to attempt to do
3974 * either combinatorial reconstruction or error correction
3975 * based on checksums. It also might be a good place
3976 * to send out preliminary ereports before we suspend
3977 * processing.
3978 */
3979 }
3980
3981 /*
3982 * If there were logical child errors, they apply to us now.
3983 * We defer this until now to avoid conflating logical child
3984 * errors with errors that happened to the zio itself when
3985 * updating vdev stats and reporting FMA events above.
3986 */
3987 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3988
3989 if ((zio->io_error || zio->io_reexecute) &&
3990 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3991 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3992 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3993
3994 zio_gang_tree_free(&zio->io_gang_tree);
3995
3996 /*
3997 * Godfather I/Os should never suspend.
3998 */
3999 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
4000 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
4001 zio->io_reexecute = 0;
4002
4003 if (zio->io_reexecute) {
4004 /*
4005 * This is a logical I/O that wants to reexecute.
4006 *
4007 * Reexecute is top-down. When an i/o fails, if it's not
4008 * the root, it simply notifies its parent and sticks around.
4009 * The parent, seeing that it still has children in zio_done(),
4010 * does the same. This percolates all the way up to the root.
4011 * The root i/o will reexecute or suspend the entire tree.
4012 *
4013 * This approach ensures that zio_reexecute() honors
4014 * all the original i/o dependency relationships, e.g.
4015 * parents not executing until children are ready.
4016 */
4017 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4018
4019 zio->io_gang_leader = NULL;
4020
4021 mutex_enter(&zio->io_lock);
4022 zio->io_state[ZIO_WAIT_DONE] = 1;
4023 mutex_exit(&zio->io_lock);
4024
4025 /*
4026 * "The Godfather" I/O monitors its children but is
4027 * not a true parent to them. It will track them through
4028 * the pipeline but severs its ties whenever they get into
4029 * trouble (e.g. suspended). This allows "The Godfather"
4030 * I/O to return status without blocking.
4031 */
4032 zl = NULL;
4033 for (pio = zio_walk_parents(zio, &zl); pio != NULL;
4034 pio = pio_next) {
4035 zio_link_t *remove_zl = zl;
4036 pio_next = zio_walk_parents(zio, &zl);
4037
4038 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
4039 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
4040 zio_remove_child(pio, zio, remove_zl);
4041 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4042 }
4043 }
4044
4045 if ((pio = zio_unique_parent(zio)) != NULL) {
4046 /*
4047 * We're not a root i/o, so there's nothing to do
4048 * but notify our parent. Don't propagate errors
4049 * upward since we haven't permanently failed yet.
4050 */
4051 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
4052 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
4053 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4054 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
4055 /*
4056 * We'd fail again if we reexecuted now, so suspend
4057 * until conditions improve (e.g. device comes online).
4058 */
4059 zio_suspend(spa, zio);
4060 } else {
4061 /*
4062 * Reexecution is potentially a huge amount of work.
4063 * Hand it off to the otherwise-unused claim taskq.
4064 */
4065 ASSERT(zio->io_tqent.tqent_next == NULL);
4066 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
4067 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
4068 0, &zio->io_tqent);
4069 }
4070 return (ZIO_PIPELINE_STOP);
4071 }
4072
4073 ASSERT(zio->io_child_count == 0);
4074 ASSERT(zio->io_reexecute == 0);
4075 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
4076
4077 /*
4078 * Report any checksum errors, since the I/O is complete.
4079 */
4080 while (zio->io_cksum_report != NULL) {
4081 zio_cksum_report_t *zcr = zio->io_cksum_report;
4082 zio->io_cksum_report = zcr->zcr_next;
4083 zcr->zcr_next = NULL;
4084 zcr->zcr_finish(zcr, NULL);
4085 zfs_ereport_free_checksum(zcr);
4086 }
4087
4088 /*
4089 * It is the responsibility of the done callback to ensure that this
4090 * particular zio is no longer discoverable for adoption, and as
4091 * such, cannot acquire any new parents.
4092 */
4093 if (zio->io_done)
4094 zio->io_done(zio);
4095
4096 mutex_enter(&zio->io_lock);
4097 zio->io_state[ZIO_WAIT_DONE] = 1;
4098 mutex_exit(&zio->io_lock);
4099
4100 zl = NULL;
4101 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
4102 zio_link_t *remove_zl = zl;
4103 pio_next = zio_walk_parents(zio, &zl);
4104 zio_remove_child(pio, zio, remove_zl);
4105 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4106 }
4107
4108 if (zio->io_waiter != NULL) {
4109 mutex_enter(&zio->io_lock);
4110 zio->io_executor = NULL;
4111 cv_broadcast(&zio->io_cv);
4112 mutex_exit(&zio->io_lock);
4113 } else {
4114 zio_destroy(zio);
4115 }
4116
4117 return (ZIO_PIPELINE_STOP);
4118 }
4119
4120 zio_t *
4121 zio_wbc(zio_type_t type, vdev_t *vd, abd_t *data,
4122 uint64_t size, uint64_t offset)
4123 {
4124 zio_t *zio = NULL;
4125
4126 switch (type) {
4127 case ZIO_TYPE_WRITE:
4128 zio = zio_create(NULL, vd->vdev_spa, 0, NULL, data, size,
4129 size, NULL, NULL, ZIO_TYPE_WRITE, ZIO_PRIORITY_ASYNC_WRITE,
4130 ZIO_FLAG_PHYSICAL, vd, offset,
4131 NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
4132 break;
4133 case ZIO_TYPE_READ:
4134 zio = zio_create(NULL, vd->vdev_spa, 0, NULL, data, size,
4135 size, NULL, NULL, ZIO_TYPE_READ, ZIO_PRIORITY_ASYNC_READ,
4136 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_PHYSICAL, vd, offset,
4137 NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
4138 break;
4139 default:
4140 ASSERT(0);
4141 }
4142
4143 zio->io_prop.zp_checksum = ZIO_CHECKSUM_OFF;
4144
4145 return (zio);
4146 }
4147
4148 /*
4149 * ==========================================================================
4150 * I/O pipeline definition
4151 * ==========================================================================
4152 */
4153 static zio_pipe_stage_t *zio_pipeline[] = {
4154 NULL,
4155 zio_read_bp_init,
4156 zio_write_bp_init,
4157 zio_free_bp_init,
4158 zio_issue_async,
4159 zio_write_compress,
4160 zio_checksum_generate,
4161 zio_nop_write,
4162 zio_ddt_read_start,
4163 zio_ddt_read_done,
4164 zio_ddt_write,
4165 zio_ddt_free,
4166 zio_gang_assemble,
4167 zio_gang_issue,
4168 zio_dva_throttle,
4169 zio_dva_allocate,
4170 zio_dva_free,
4171 zio_dva_claim,
4172 zio_ready,
4173 zio_vdev_io_start,
4174 zio_vdev_io_done,
4175 zio_vdev_io_assess,
4176 zio_checksum_verify,
4177 zio_done
4178 };
4179
4180
4181
4182
4183 /*
4184 * Compare two zbookmark_phys_t's to see which we would reach first in a
4185 * pre-order traversal of the object tree.
4186 *
4187 * This is simple in every case aside from the meta-dnode object. For all other
4188 * objects, we traverse them in order (object 1 before object 2, and so on).
4189 * However, all of these objects are traversed while traversing object 0, since
4190 * the data it points to is the list of objects. Thus, we need to convert to a
4191 * canonical representation so we can compare meta-dnode bookmarks to
4192 * non-meta-dnode bookmarks.
4193 *
4194 * We do this by calculating "equivalents" for each field of the zbookmark.
4195 * zbookmarks outside of the meta-dnode use their own object and level, and
4196 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4197 * blocks this bookmark refers to) by multiplying their blkid by their span
4198 * (the number of L0 blocks contained within one block at their level).
4199 * zbookmarks inside the meta-dnode calculate their object equivalent
4200 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4201 * level + 1<<31 (any value larger than a level could ever be) for their level.
4202 * This causes them to always compare before a bookmark in their object
4203 * equivalent, compare appropriately to bookmarks in other objects, and to
4204 * compare appropriately to other bookmarks in the meta-dnode.
4205 */
4206 int
4207 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
4208 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
4209 {
4210 /*
4211 * These variables represent the "equivalent" values for the zbookmark,
4212 * after converting zbookmarks inside the meta dnode to their
4213 * normal-object equivalents.
4214 */
4215 uint64_t zb1obj, zb2obj;
4216 uint64_t zb1L0, zb2L0;
4217 uint64_t zb1level, zb2level;
4218
4219 if (zb1->zb_object == zb2->zb_object &&
4220 zb1->zb_level == zb2->zb_level &&
4221 zb1->zb_blkid == zb2->zb_blkid)
4222 return (0);
4223
4224 /*
4225 * BP_SPANB calculates the span in blocks.
4226 */
4227 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
4228 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
4229
4230 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
4231 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4232 zb1L0 = 0;
4233 zb1level = zb1->zb_level + COMPARE_META_LEVEL;
4234 } else {
4235 zb1obj = zb1->zb_object;
4236 zb1level = zb1->zb_level;
4237 }
4238
4239 if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
4240 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4241 zb2L0 = 0;
4242 zb2level = zb2->zb_level + COMPARE_META_LEVEL;
4243 } else {
4244 zb2obj = zb2->zb_object;
4245 zb2level = zb2->zb_level;
4246 }
4247
4248 /* Now that we have a canonical representation, do the comparison. */
4249 if (zb1obj != zb2obj)
4250 return (zb1obj < zb2obj ? -1 : 1);
4251 else if (zb1L0 != zb2L0)
4252 return (zb1L0 < zb2L0 ? -1 : 1);
4253 else if (zb1level != zb2level)
4254 return (zb1level > zb2level ? -1 : 1);
4255 /*
4256 * This can (theoretically) happen if the bookmarks have the same object
4257 * and level, but different blkids, if the block sizes are not the same.
4258 * There is presently no way to change the indirect block sizes
4259 */
4260 return (0);
4261 }
4262
4263 /*
4264 * This function checks the following: given that last_block is the place that
4265 * our traversal stopped last time, does that guarantee that we've visited
4266 * every node under subtree_root? Therefore, we can't just use the raw output
4267 * of zbookmark_compare. We have to pass in a modified version of
4268 * subtree_root; by incrementing the block id, and then checking whether
4269 * last_block is before or equal to that, we can tell whether or not having
4270 * visited last_block implies that all of subtree_root's children have been
4271 * visited.
4272 */
4273 boolean_t
4274 zbookmark_subtree_completed(const dnode_phys_t *dnp,
4275 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
4276 {
4277 zbookmark_phys_t mod_zb = *subtree_root;
4278 mod_zb.zb_blkid++;
4279 ASSERT(last_block->zb_level == 0);
4280
4281 /* The objset_phys_t isn't before anything. */
4282 if (dnp == NULL)
4283 return (B_FALSE);
4284
4285 /*
4286 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4287 * data block size in sectors, because that variable is only used if
4288 * the bookmark refers to a block in the meta-dnode. Since we don't
4289 * know without examining it what object it refers to, and there's no
4290 * harm in passing in this value in other cases, we always pass it in.
4291 *
4292 * We pass in 0 for the indirect block size shift because zb2 must be
4293 * level 0. The indirect block size is only used to calculate the span
4294 * of the bookmark, but since the bookmark must be level 0, the span is
4295 * always 1, so the math works out.
4296 *
4297 * If you make changes to how the zbookmark_compare code works, be sure
4298 * to make sure that this code still works afterwards.
4299 */
4300 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
4301 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,
4302 last_block) <= 0);
4303 }