1/*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21/*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2014 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25 */
26
27#include <sys/zfs_context.h>
28#include <sys/fm/fs/zfs.h>
29#include <sys/spa.h>
30#include <sys/txg.h>
31#include <sys/spa_impl.h>
32#include <sys/vdev_impl.h>
33#include <sys/zio_impl.h>
34#include <sys/zio_compress.h>
35#include <sys/zio_checksum.h>
36#include <sys/dmu_objset.h>
37#include <sys/arc.h>
38#include <sys/ddt.h>
39#include <sys/trim_map.h>
40#include <sys/blkptr.h>
41#include <sys/zfeature.h>
42
43SYSCTL_DECL(_vfs_zfs);
44SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO");
45#if defined(__amd64__)
46static int zio_use_uma = 1;
47#else
48static int zio_use_uma = 0;
49#endif
50SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
51 "Use uma(9) for ZIO allocations");
52static int zio_exclude_metadata = 0;
53SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, exclude_metadata, CTLFLAG_RDTUN, &zio_exclude_metadata, 0,
54 "Exclude metadata buffers from dumps as well");
55
56zio_trim_stats_t zio_trim_stats = {
57 { "bytes", KSTAT_DATA_UINT64,
58 "Number of bytes successfully TRIMmed" },
59 { "success", KSTAT_DATA_UINT64,
60 "Number of successful TRIM requests" },
61 { "unsupported", KSTAT_DATA_UINT64,
62 "Number of TRIM requests that failed because TRIM is not supported" },
63 { "failed", KSTAT_DATA_UINT64,
64 "Number of TRIM requests that failed for reasons other than not supported" },
65};
66
67static kstat_t *zio_trim_ksp;
68
69/*
70 * ==========================================================================
71 * I/O type descriptions
72 * ==========================================================================
73 */
74const char *zio_type_name[ZIO_TYPES] = {
75 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
76 "zio_ioctl"
77};
78
79/*
80 * ==========================================================================
81 * I/O kmem caches
82 * ==========================================================================
83 */
84kmem_cache_t *zio_cache;
85kmem_cache_t *zio_link_cache;
86kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
87kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
88
89#ifdef _KERNEL
90extern vmem_t *zio_alloc_arena;
91#endif
92
93/*
94 * The following actions directly effect the spa's sync-to-convergence logic.
95 * The values below define the sync pass when we start performing the action.
96 * Care should be taken when changing these values as they directly impact
97 * spa_sync() performance. Tuning these values may introduce subtle performance
98 * pathologies and should only be done in the context of performance analysis.
99 * These tunables will eventually be removed and replaced with #defines once
100 * enough analysis has been done to determine optimal values.
101 *
102 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
103 * regular blocks are not deferred.
104 */
105int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
106SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN,
107 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass");
108int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
109SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN,
110 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass");
111int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
112SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN,
113 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass");
114
115/*
116 * An allocating zio is one that either currently has the DVA allocate
117 * stage set or will have it later in its lifetime.
118 */
119#define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
120
121boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
122
123#ifdef ZFS_DEBUG
124int zio_buf_debug_limit = 16384;
125#else
126int zio_buf_debug_limit = 0;
127#endif
128
129void
130zio_init(void)
131{
132 size_t c;
133 zio_cache = kmem_cache_create("zio_cache",
134 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
135 zio_link_cache = kmem_cache_create("zio_link_cache",
136 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
137 if (!zio_use_uma)
138 goto out;
139
140 /*
141 * For small buffers, we want a cache for each multiple of
142 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
143 * for each quarter-power of 2. For large buffers, we want
144 * a cache for each multiple of PAGESIZE.
145 */
146 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
147 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
148 size_t p2 = size;
149 size_t align = 0;
150 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
151
152 while (p2 & (p2 - 1))
153 p2 &= p2 - 1;
154
155#ifdef illumos
156#ifndef _KERNEL
157 /*
158 * If we are using watchpoints, put each buffer on its own page,
159 * to eliminate the performance overhead of trapping to the
160 * kernel when modifying a non-watched buffer that shares the
161 * page with a watched buffer.
162 */
163 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
164 continue;
165#endif
166#endif /* illumos */
167 if (size <= 4 * SPA_MINBLOCKSIZE) {
168 align = SPA_MINBLOCKSIZE;
169 } else if (IS_P2ALIGNED(size, PAGESIZE)) {
170 align = PAGESIZE;
171 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
172 align = p2 >> 2;
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, NULL,
189 cflags | KMC_NOTOUCH | KMC_NODEBUG);
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 }
202out:
203
204 zio_inject_init();
205
206 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc",
207 KSTAT_TYPE_NAMED,
208 sizeof(zio_trim_stats) / sizeof(kstat_named_t),
209 KSTAT_FLAG_VIRTUAL);
210
211 if (zio_trim_ksp != NULL) {
212 zio_trim_ksp->ks_data = &zio_trim_stats;
213 kstat_install(zio_trim_ksp);
214 }
215}
216
217void
218zio_fini(void)
219{
220 size_t c;
221 kmem_cache_t *last_cache = NULL;
222 kmem_cache_t *last_data_cache = NULL;
223
224 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
225 if (zio_buf_cache[c] != last_cache) {
226 last_cache = zio_buf_cache[c];
227 kmem_cache_destroy(zio_buf_cache[c]);
228 }
229 zio_buf_cache[c] = NULL;
230
231 if (zio_data_buf_cache[c] != last_data_cache) {
232 last_data_cache = zio_data_buf_cache[c];
233 kmem_cache_destroy(zio_data_buf_cache[c]);
234 }
235 zio_data_buf_cache[c] = NULL;
236 }
237
238 kmem_cache_destroy(zio_link_cache);
239 kmem_cache_destroy(zio_cache);
240
241 zio_inject_fini();
242
243 if (zio_trim_ksp != NULL) {
244 kstat_delete(zio_trim_ksp);
245 zio_trim_ksp = NULL;
246 }
247}
248
249/*
250 * ==========================================================================
251 * Allocate and free I/O buffers
252 * ==========================================================================
253 */
254
255/*
256 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
257 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
258 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
259 * excess / transient data in-core during a crashdump.
260 */
261void *
262zio_buf_alloc(size_t size)
263{
264 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
265 int flags = zio_exclude_metadata ? KM_NODEBUG : 0;
266
267 ASSERT3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
268
269 if (zio_use_uma)
270 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
271 else
272 return (kmem_alloc(size, KM_SLEEP|flags));
273}
274
275/*
276 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
277 * crashdump if the kernel panics. This exists so that we will limit the amount
278 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
279 * of kernel heap dumped to disk when the kernel panics)
280 */
281void *
282zio_data_buf_alloc(size_t size)
283{
284 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
285
286 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
287
288 if (zio_use_uma)
289 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
290 else
291 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
292}
293
294void
295zio_buf_free(void *buf, size_t size)
296{
297 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
298
299 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
300
301 if (zio_use_uma)
302 kmem_cache_free(zio_buf_cache[c], buf);
303 else
304 kmem_free(buf, size);
305}
306
307void
308zio_data_buf_free(void *buf, size_t size)
309{
310 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
311
312 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
313
314 if (zio_use_uma)
315 kmem_cache_free(zio_data_buf_cache[c], buf);
316 else
317 kmem_free(buf, size);
318}
319
320/*
321 * ==========================================================================
322 * Push and pop I/O transform buffers
323 * ==========================================================================
324 */
325static void
326zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
327 zio_transform_func_t *transform)
328{
329 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
330
331 zt->zt_orig_data = zio->io_data;
332 zt->zt_orig_size = zio->io_size;
333 zt->zt_bufsize = bufsize;
334 zt->zt_transform = transform;
335
336 zt->zt_next = zio->io_transform_stack;
337 zio->io_transform_stack = zt;
338
339 zio->io_data = data;
340 zio->io_size = size;
341}
342
343static void
344zio_pop_transforms(zio_t *zio)
345{
346 zio_transform_t *zt;
347
348 while ((zt = zio->io_transform_stack) != NULL) {
349 if (zt->zt_transform != NULL)
350 zt->zt_transform(zio,
351 zt->zt_orig_data, zt->zt_orig_size);
352
353 if (zt->zt_bufsize != 0)
354 zio_buf_free(zio->io_data, zt->zt_bufsize);
355
356 zio->io_data = zt->zt_orig_data;
357 zio->io_size = zt->zt_orig_size;
358 zio->io_transform_stack = zt->zt_next;
359
360 kmem_free(zt, sizeof (zio_transform_t));
361 }
362}
363
364/*
365 * ==========================================================================
366 * I/O transform callbacks for subblocks and decompression
367 * ==========================================================================
368 */
369static void
370zio_subblock(zio_t *zio, void *data, uint64_t size)
371{
372 ASSERT(zio->io_size > size);
373
374 if (zio->io_type == ZIO_TYPE_READ)
375 bcopy(zio->io_data, data, size);
376}
377
378static void
379zio_decompress(zio_t *zio, void *data, uint64_t size)
380{
381 if (zio->io_error == 0 &&
382 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
383 zio->io_data, data, zio->io_size, size) != 0)
384 zio->io_error = SET_ERROR(EIO);
385}
386
387/*
388 * ==========================================================================
389 * I/O parent/child relationships and pipeline interlocks
390 * ==========================================================================
391 */
392/*
393 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
394 * continue calling these functions until they return NULL.
395 * Otherwise, the next caller will pick up the list walk in
396 * some indeterminate state. (Otherwise every caller would
397 * have to pass in a cookie to keep the state represented by
398 * io_walk_link, which gets annoying.)
399 */
400zio_t *
401zio_walk_parents(zio_t *cio)
402{
403 zio_link_t *zl = cio->io_walk_link;
404 list_t *pl = &cio->io_parent_list;
405
406 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
407 cio->io_walk_link = zl;
408
409 if (zl == NULL)
410 return (NULL);
411
412 ASSERT(zl->zl_child == cio);
413 return (zl->zl_parent);
414}
415
416zio_t *
417zio_walk_children(zio_t *pio)
418{
419 zio_link_t *zl = pio->io_walk_link;
420 list_t *cl = &pio->io_child_list;
421
422 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
423 pio->io_walk_link = zl;
424
425 if (zl == NULL)
426 return (NULL);
427
428 ASSERT(zl->zl_parent == pio);
429 return (zl->zl_child);
430}
431
432zio_t *
433zio_unique_parent(zio_t *cio)
434{
435 zio_t *pio = zio_walk_parents(cio);
436
437 VERIFY(zio_walk_parents(cio) == NULL);
438 return (pio);
439}
440
441void
442zio_add_child(zio_t *pio, zio_t *cio)
443{
444 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
445
446 /*
447 * Logical I/Os can have logical, gang, or vdev children.
448 * Gang I/Os can have gang or vdev children.
449 * Vdev I/Os can only have vdev children.
450 * The following ASSERT captures all of these constraints.
451 */
452 ASSERT(cio->io_child_type <= pio->io_child_type);
453
454 zl->zl_parent = pio;
455 zl->zl_child = cio;
456
457 mutex_enter(&cio->io_lock);
458 mutex_enter(&pio->io_lock);
459
460 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
461
462 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
463 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
464
465 list_insert_head(&pio->io_child_list, zl);
466 list_insert_head(&cio->io_parent_list, zl);
467
468 pio->io_child_count++;
469 cio->io_parent_count++;
470
471 mutex_exit(&pio->io_lock);
472 mutex_exit(&cio->io_lock);
473}
474
475static void
476zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
477{
478 ASSERT(zl->zl_parent == pio);
479 ASSERT(zl->zl_child == cio);
480
481 mutex_enter(&cio->io_lock);
482 mutex_enter(&pio->io_lock);
483
484 list_remove(&pio->io_child_list, zl);
485 list_remove(&cio->io_parent_list, zl);
486
487 pio->io_child_count--;
488 cio->io_parent_count--;
489
490 mutex_exit(&pio->io_lock);
491 mutex_exit(&cio->io_lock);
492
493 kmem_cache_free(zio_link_cache, zl);
494}
495
496static boolean_t
497zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
498{
499 uint64_t *countp = &zio->io_children[child][wait];
500 boolean_t waiting = B_FALSE;
501
502 mutex_enter(&zio->io_lock);
503 ASSERT(zio->io_stall == NULL);
504 if (*countp != 0) {
505 zio->io_stage >>= 1;
506 zio->io_stall = countp;
507 waiting = B_TRUE;
508 }
509 mutex_exit(&zio->io_lock);
510
511 return (waiting);
512}
513
514static void
515zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
516{
517 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
518 int *errorp = &pio->io_child_error[zio->io_child_type];
519
520 mutex_enter(&pio->io_lock);
521 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
522 *errorp = zio_worst_error(*errorp, zio->io_error);
523 pio->io_reexecute |= zio->io_reexecute;
524 ASSERT3U(*countp, >, 0);
525
526 (*countp)--;
527
528 if (*countp == 0 && pio->io_stall == countp) {
529 pio->io_stall = NULL;
530 mutex_exit(&pio->io_lock);
531 zio_execute(pio);
532 } else {
533 mutex_exit(&pio->io_lock);
534 }
535}
536
537static void
538zio_inherit_child_errors(zio_t *zio, enum zio_child c)
539{
540 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
541 zio->io_error = zio->io_child_error[c];
542}
543
544/*
545 * ==========================================================================
546 * Create the various types of I/O (read, write, free, etc)
547 * ==========================================================================
548 */
549static zio_t *
550zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
551 void *data, uint64_t size, zio_done_func_t *done, void *private,
552 zio_type_t type, zio_priority_t priority, enum zio_flag flags,
553 vdev_t *vd, uint64_t offset, const zbookmark_phys_t *zb,
554 enum zio_stage stage, enum zio_stage pipeline)
555{
556 zio_t *zio;
557
558 ASSERT3U(type == ZIO_TYPE_FREE || size, <=, SPA_MAXBLOCKSIZE);
559 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
560 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
561
562 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
563 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
564 ASSERT(vd || stage == ZIO_STAGE_OPEN);
565
566 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
567 bzero(zio, sizeof (zio_t));
568
569 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
570 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
571
572 list_create(&zio->io_parent_list, sizeof (zio_link_t),
573 offsetof(zio_link_t, zl_parent_node));
574 list_create(&zio->io_child_list, sizeof (zio_link_t),
575 offsetof(zio_link_t, zl_child_node));
576
577 if (vd != NULL)
578 zio->io_child_type = ZIO_CHILD_VDEV;
579 else if (flags & ZIO_FLAG_GANG_CHILD)
580 zio->io_child_type = ZIO_CHILD_GANG;
581 else if (flags & ZIO_FLAG_DDT_CHILD)
582 zio->io_child_type = ZIO_CHILD_DDT;
583 else
584 zio->io_child_type = ZIO_CHILD_LOGICAL;
585
586 if (bp != NULL) {
587 zio->io_bp = (blkptr_t *)bp;
588 zio->io_bp_copy = *bp;
589 zio->io_bp_orig = *bp;
590 if (type != ZIO_TYPE_WRITE ||
591 zio->io_child_type == ZIO_CHILD_DDT)
592 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
593 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
594 zio->io_logical = zio;
595 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
596 pipeline |= ZIO_GANG_STAGES;
597 }
598
599 zio->io_spa = spa;
600 zio->io_txg = txg;
601 zio->io_done = done;
602 zio->io_private = private;
603 zio->io_type = type;
604 zio->io_priority = priority;
605 zio->io_vd = vd;
606 zio->io_offset = offset;
607 zio->io_orig_data = zio->io_data = data;
608 zio->io_orig_size = zio->io_size = size;
609 zio->io_orig_flags = zio->io_flags = flags;
610 zio->io_orig_stage = zio->io_stage = stage;
611 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
612
613 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
614 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
615
616 if (zb != NULL)
617 zio->io_bookmark = *zb;
618
619 if (pio != NULL) {
620 if (zio->io_logical == NULL)
621 zio->io_logical = pio->io_logical;
622 if (zio->io_child_type == ZIO_CHILD_GANG)
623 zio->io_gang_leader = pio->io_gang_leader;
624 zio_add_child(pio, zio);
625 }
626
627 return (zio);
628}
629
630static void
631zio_destroy(zio_t *zio)
632{
633 list_destroy(&zio->io_parent_list);
634 list_destroy(&zio->io_child_list);
635 mutex_destroy(&zio->io_lock);
636 cv_destroy(&zio->io_cv);
637 kmem_cache_free(zio_cache, zio);
638}
639
640zio_t *
641zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
642 void *private, enum zio_flag flags)
643{
644 zio_t *zio;
645
646 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
647 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
648 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
649
650 return (zio);
651}
652
653zio_t *
654zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
655{
656 return (zio_null(NULL, spa, NULL, done, private, flags));
657}
658
659zio_t *
660zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
661 void *data, uint64_t size, zio_done_func_t *done, void *private,
662 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
663{
664 zio_t *zio;
665
666 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
667 data, size, done, private,
668 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
669 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
670 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
671
672 return (zio);
673}
674
675zio_t *
676zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
677 void *data, uint64_t size, const zio_prop_t *zp,
678 zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done,
679 void *private,
680 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
681{
682 zio_t *zio;
683
684 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
685 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
686 zp->zp_compress >= ZIO_COMPRESS_OFF &&
687 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
688 DMU_OT_IS_VALID(zp->zp_type) &&
689 zp->zp_level < 32 &&
690 zp->zp_copies > 0 &&
691 zp->zp_copies <= spa_max_replication(spa));
692
693 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
694 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
695 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
696 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
697
698 zio->io_ready = ready;
699 zio->io_physdone = physdone;
700 zio->io_prop = *zp;
701
702 /*
703 * Data can be NULL if we are going to call zio_write_override() to
704 * provide the already-allocated BP. But we may need the data to
705 * verify a dedup hit (if requested). In this case, don't try to
706 * dedup (just take the already-allocated BP verbatim).
707 */
708 if (data == NULL && zio->io_prop.zp_dedup_verify) {
709 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
710 }
711
712 return (zio);
713}
714
715zio_t *
716zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
717 uint64_t size, zio_done_func_t *done, void *private,
718 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
719{
720 zio_t *zio;
721
722 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
723 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
724 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
725
726 return (zio);
727}
728
729void
730zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
731{
732 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
733 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
734 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
735 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
736
737 /*
738 * We must reset the io_prop to match the values that existed
739 * when the bp was first written by dmu_sync() keeping in mind
740 * that nopwrite and dedup are mutually exclusive.
741 */
742 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
743 zio->io_prop.zp_nopwrite = nopwrite;
744 zio->io_prop.zp_copies = copies;
745 zio->io_bp_override = bp;
746}
747
748void
749zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
750{
751
752 /*
753 * The check for EMBEDDED is a performance optimization. We
754 * process the free here (by ignoring it) rather than
755 * putting it on the list and then processing it in zio_free_sync().
756 */
757 if (BP_IS_EMBEDDED(bp))
758 return;
759 metaslab_check_free(spa, bp);
760
761 /*
762 * Frees that are for the currently-syncing txg, are not going to be
763 * deferred, and which will not need to do a read (i.e. not GANG or
764 * DEDUP), can be processed immediately. Otherwise, put them on the
765 * in-memory list for later processing.
766 */
767 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
768 txg != spa->spa_syncing_txg ||
769 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
770 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
771 } else {
772 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
773 BP_GET_PSIZE(bp), 0)));
774 }
775}
776
777zio_t *
778zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
779 uint64_t size, enum zio_flag flags)
780{
781 zio_t *zio;
782 enum zio_stage stage = ZIO_FREE_PIPELINE;
783
784 ASSERT(!BP_IS_HOLE(bp));
785 ASSERT(spa_syncing_txg(spa) == txg);
786 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
787
788 if (BP_IS_EMBEDDED(bp))
789 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
790
791 metaslab_check_free(spa, bp);
792 arc_freed(spa, bp);
793
794 if (zfs_trim_enabled)
795 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
796 ZIO_STAGE_VDEV_IO_ASSESS;
797 /*
798 * GANG and DEDUP blocks can induce a read (for the gang block header,
799 * or the DDT), so issue them asynchronously so that this thread is
800 * not tied up.
801 */
802 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
803 stage |= ZIO_STAGE_ISSUE_ASYNC;
804
805 flags |= ZIO_FLAG_DONT_QUEUE;
806
807 zio = zio_create(pio, spa, txg, bp, NULL, size,
808 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags,
809 NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
810
811 return (zio);
812}
813
814zio_t *
815zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
816 zio_done_func_t *done, void *private, enum zio_flag flags)
817{
818 zio_t *zio;
819
820 dprintf_bp(bp, "claiming in txg %llu", txg);
821
822 if (BP_IS_EMBEDDED(bp))
823 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
824
825 /*
826 * A claim is an allocation of a specific block. Claims are needed
827 * to support immediate writes in the intent log. The issue is that
828 * immediate writes contain committed data, but in a txg that was
829 * *not* committed. Upon opening the pool after an unclean shutdown,
830 * the intent log claims all blocks that contain immediate write data
831 * so that the SPA knows they're in use.
832 *
833 * All claims *must* be resolved in the first txg -- before the SPA
834 * starts allocating blocks -- so that nothing is allocated twice.
835 * If txg == 0 we just verify that the block is claimable.
836 */
837 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
838 ASSERT(txg == spa_first_txg(spa) || txg == 0);
839 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
840
841 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
842 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
843 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
844
845 return (zio);
846}
847
848zio_t *
849zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
850 uint64_t size, zio_done_func_t *done, void *private,
851 zio_priority_t priority, enum zio_flag flags)
852{
853 zio_t *zio;
854 int c;
855
856 if (vd->vdev_children == 0) {
857 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private,
858 ZIO_TYPE_IOCTL, priority, flags, vd, offset, NULL,
859 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
860
861 zio->io_cmd = cmd;
862 } else {
863 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
864
865 for (c = 0; c < vd->vdev_children; c++)
866 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
867 offset, size, done, private, priority, flags));
868 }
869
870 return (zio);
871}
872
873zio_t *
874zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
875 void *data, int checksum, zio_done_func_t *done, void *private,
876 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
877{
878 zio_t *zio;
879
880 ASSERT(vd->vdev_children == 0);
881 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
882 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
883 ASSERT3U(offset + size, <=, vd->vdev_psize);
884
885 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
886 ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
887 NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
888
889 zio->io_prop.zp_checksum = checksum;
890
891 return (zio);
892}
893
894zio_t *
895zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
896 void *data, int checksum, zio_done_func_t *done, void *private,
897 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
898{
899 zio_t *zio;
900
901 ASSERT(vd->vdev_children == 0);
902 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
903 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
904 ASSERT3U(offset + size, <=, vd->vdev_psize);
905
906 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
907 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
908 NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
909
910 zio->io_prop.zp_checksum = checksum;
911
912 if (zio_checksum_table[checksum].ci_eck) {
913 /*
914 * zec checksums are necessarily destructive -- they modify
915 * the end of the write buffer to hold the verifier/checksum.
916 * Therefore, we must make a local copy in case the data is
917 * being written to multiple places in parallel.
918 */
919 void *wbuf = zio_buf_alloc(size);
920 bcopy(data, wbuf, size);
921 zio_push_transform(zio, wbuf, size, size, NULL);
922 }
923
924 return (zio);
925}
926
927/*
928 * Create a child I/O to do some work for us.
929 */
930zio_t *
931zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
932 void *data, uint64_t size, int type, zio_priority_t priority,
933 enum zio_flag flags, zio_done_func_t *done, void *private)
934{
935 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
936 zio_t *zio;
937
938 ASSERT(vd->vdev_parent ==
939 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
940
941 if (type == ZIO_TYPE_READ && bp != NULL) {
942 /*
943 * If we have the bp, then the child should perform the
944 * checksum and the parent need not. This pushes error
945 * detection as close to the leaves as possible and
946 * eliminates redundant checksums in the interior nodes.
947 */
948 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
949 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
950 }
951
952 /* Not all IO types require vdev io done stage e.g. free */
953 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE))
954 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE;
955
956 if (vd->vdev_children == 0)
957 offset += VDEV_LABEL_START_SIZE;
958
959 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
960
961 /*
962 * If we've decided to do a repair, the write is not speculative --
963 * even if the original read was.
964 */
965 if (flags & ZIO_FLAG_IO_REPAIR)
966 flags &= ~ZIO_FLAG_SPECULATIVE;
967
968 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
969 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
970 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
971
972 zio->io_physdone = pio->io_physdone;
973 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
974 zio->io_logical->io_phys_children++;
975
976 return (zio);
977}
978
979zio_t *
980zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
981 int type, zio_priority_t priority, enum zio_flag flags,
982 zio_done_func_t *done, void *private)
983{
984 zio_t *zio;
985
986 ASSERT(vd->vdev_ops->vdev_op_leaf);
987
988 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
989 data, size, done, private, type, priority,
990 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
991 vd, offset, NULL,
992 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
993
994 return (zio);
995}
996
997void
998zio_flush(zio_t *zio, vdev_t *vd)
999{
1000 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
1001 NULL, NULL, ZIO_PRIORITY_NOW,
1002 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1003}
1004
1005zio_t *
1006zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
1007{
1008
1009 ASSERT(vd->vdev_ops->vdev_op_leaf);
1010
1011 return (zio_create(zio, spa, 0, NULL, NULL, size, NULL, NULL,
1012 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE |
1013 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY,
1014 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE));
1015}
1016
1017void
1018zio_shrink(zio_t *zio, uint64_t size)
1019{
1020 ASSERT(zio->io_executor == NULL);
1021 ASSERT(zio->io_orig_size == zio->io_size);
1022 ASSERT(size <= zio->io_size);
1023
1024 /*
1025 * We don't shrink for raidz because of problems with the
1026 * reconstruction when reading back less than the block size.
1027 * Note, BP_IS_RAIDZ() assumes no compression.
1028 */
1029 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1030 if (!BP_IS_RAIDZ(zio->io_bp))
1031 zio->io_orig_size = zio->io_size = size;
1032}
1033
1034/*
1035 * ==========================================================================
1036 * Prepare to read and write logical blocks
1037 * ==========================================================================
1038 */
1039
1040static int
1041zio_read_bp_init(zio_t *zio)
1042{
1043 blkptr_t *bp = zio->io_bp;
1044
1045 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1046 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1047 !(zio->io_flags & ZIO_FLAG_RAW)) {
1048 uint64_t psize =
1049 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1050 void *cbuf = zio_buf_alloc(psize);
1051
1052 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1053 }
1054
1055 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1056 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1057 decode_embedded_bp_compressed(bp, zio->io_data);
1058 } else {
1059 ASSERT(!BP_IS_EMBEDDED(bp));
1060 }
1061
1062 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1063 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1064
1065 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1066 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1067
1068 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1069 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1070
1071 return (ZIO_PIPELINE_CONTINUE);
1072}
1073
1074static int
1075zio_write_bp_init(zio_t *zio)
1076{
1077 spa_t *spa = zio->io_spa;
1078 zio_prop_t *zp = &zio->io_prop;
1079 enum zio_compress compress = zp->zp_compress;
1080 blkptr_t *bp = zio->io_bp;
1081 uint64_t lsize = zio->io_size;
1082 uint64_t psize = lsize;
1083 int pass = 1;
1084
1085 /*
1086 * If our children haven't all reached the ready stage,
1087 * wait for them and then repeat this pipeline stage.
1088 */
1089 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1090 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1091 return (ZIO_PIPELINE_STOP);
1092
1093 if (!IO_IS_ALLOCATING(zio))
1094 return (ZIO_PIPELINE_CONTINUE);
1095
1096 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1097
1098 if (zio->io_bp_override) {
1099 ASSERT(bp->blk_birth != zio->io_txg);
1100 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1101
1102 *bp = *zio->io_bp_override;
1103 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1104
1105 if (BP_IS_EMBEDDED(bp))
1106 return (ZIO_PIPELINE_CONTINUE);
1107
1108 /*
1109 * If we've been overridden and nopwrite is set then
1110 * set the flag accordingly to indicate that a nopwrite
1111 * has already occurred.
1112 */
1113 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1114 ASSERT(!zp->zp_dedup);
1115 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1116 return (ZIO_PIPELINE_CONTINUE);
1117 }
1118
1119 ASSERT(!zp->zp_nopwrite);
1120
1121 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1122 return (ZIO_PIPELINE_CONTINUE);
1123
1124 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1125 zp->zp_dedup_verify);
1126
1127 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1128 BP_SET_DEDUP(bp, 1);
1129 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1130 return (ZIO_PIPELINE_CONTINUE);
1131 }
1132 zio->io_bp_override = NULL;
1133 BP_ZERO(bp);
1134 }
1135
1136 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1137 /*
1138 * We're rewriting an existing block, which means we're
1139 * working on behalf of spa_sync(). For spa_sync() to
1140 * converge, it must eventually be the case that we don't
1141 * have to allocate new blocks. But compression changes
1142 * the blocksize, which forces a reallocate, and makes
1143 * convergence take longer. Therefore, after the first
1144 * few passes, stop compressing to ensure convergence.
1145 */
1146 pass = spa_sync_pass(spa);
1147
1148 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1149 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1150 ASSERT(!BP_GET_DEDUP(bp));
1151
1152 if (pass >= zfs_sync_pass_dont_compress)
1153 compress = ZIO_COMPRESS_OFF;
1154
1155 /* Make sure someone doesn't change their mind on overwrites */
1156 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1157 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1158 }
1159
1160 if (compress != ZIO_COMPRESS_OFF) {
1161 void *cbuf = zio_buf_alloc(lsize);
1162 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1163 if (psize == 0 || psize == lsize) {
1164 compress = ZIO_COMPRESS_OFF;
1165 zio_buf_free(cbuf, lsize);
1166 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1167 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1168 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1169 encode_embedded_bp_compressed(bp,
1170 cbuf, compress, lsize, psize);
1171 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1172 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1173 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1174 zio_buf_free(cbuf, lsize);
1175 bp->blk_birth = zio->io_txg;
1176 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1177 ASSERT(spa_feature_is_active(spa,
1178 SPA_FEATURE_EMBEDDED_DATA));
1179 return (ZIO_PIPELINE_CONTINUE);
1180 } else {
1181 /*
1182 * Round up compressed size to MINBLOCKSIZE and
1183 * zero the tail.
1184 */
1185 size_t rounded =
1186 P2ROUNDUP(psize, (size_t)SPA_MINBLOCKSIZE);
1187 if (rounded > psize) {
1188 bzero((char *)cbuf + psize, rounded - psize);
1189 psize = rounded;
1190 }
1191 if (psize == lsize) {
1192 compress = ZIO_COMPRESS_OFF;
1193 zio_buf_free(cbuf, lsize);
1194 } else {
1195 zio_push_transform(zio, cbuf,
1196 psize, lsize, NULL);
1197 }
1198 }
1199 }
1200
1201 /*
1202 * The final pass of spa_sync() must be all rewrites, but the first
1203 * few passes offer a trade-off: allocating blocks defers convergence,
1204 * but newly allocated blocks are sequential, so they can be written
1205 * to disk faster. Therefore, we allow the first few passes of
1206 * spa_sync() to allocate new blocks, but force rewrites after that.
1207 * There should only be a handful of blocks after pass 1 in any case.
1208 */
1209 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1210 BP_GET_PSIZE(bp) == psize &&
1211 pass >= zfs_sync_pass_rewrite) {
1212 ASSERT(psize != 0);
1213 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1214 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1215 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1216 } else {
1217 BP_ZERO(bp);
1218 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1219 }
1220
1221 if (psize == 0) {
1222 if (zio->io_bp_orig.blk_birth != 0 &&
1223 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1224 BP_SET_LSIZE(bp, lsize);
1225 BP_SET_TYPE(bp, zp->zp_type);
1226 BP_SET_LEVEL(bp, zp->zp_level);
1227 BP_SET_BIRTH(bp, zio->io_txg, 0);
1228 }
1229 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1230 } else {
1231 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1232 BP_SET_LSIZE(bp, lsize);
1233 BP_SET_TYPE(bp, zp->zp_type);
1234 BP_SET_LEVEL(bp, zp->zp_level);
1235 BP_SET_PSIZE(bp, psize);
1236 BP_SET_COMPRESS(bp, compress);
1237 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1238 BP_SET_DEDUP(bp, zp->zp_dedup);
1239 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1240 if (zp->zp_dedup) {
1241 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1242 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1243 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1244 }
1245 if (zp->zp_nopwrite) {
1246 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1247 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1248 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1249 }
1250 }
1251
1252 return (ZIO_PIPELINE_CONTINUE);
1253}
1254
1255static int
1256zio_free_bp_init(zio_t *zio)
1257{
1258 blkptr_t *bp = zio->io_bp;
1259
1260 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1261 if (BP_GET_DEDUP(bp))
1262 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1263 }
1264
1265 return (ZIO_PIPELINE_CONTINUE);
1266}
1267
1268/*
1269 * ==========================================================================
1270 * Execute the I/O pipeline
1271 * ==========================================================================
1272 */
1273
1274static void
1275zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1276{
1277 spa_t *spa = zio->io_spa;
1278 zio_type_t t = zio->io_type;
1279 int flags = (cutinline ? TQ_FRONT : 0);
1280
1281 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1282
1283 /*
1284 * If we're a config writer or a probe, the normal issue and
1285 * interrupt threads may all be blocked waiting for the config lock.
1286 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1287 */
1288 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1289 t = ZIO_TYPE_NULL;
1290
1291 /*
1292 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1293 */
1294 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1295 t = ZIO_TYPE_NULL;
1296
1297 /*
1298 * If this is a high priority I/O, then use the high priority taskq if
1299 * available.
1300 */
1301 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1302 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1303 q++;
1304
1305 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1306
1307 /*
1308 * NB: We are assuming that the zio can only be dispatched
1309 * to a single taskq at a time. It would be a grievous error
1310 * to dispatch the zio to another taskq at the same time.
1311 */
1312#if defined(illumos) || !defined(_KERNEL)
1313 ASSERT(zio->io_tqent.tqent_next == NULL);
1314#else
1315 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1316#endif
1317 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1318 flags, &zio->io_tqent);
1319}
1320
1321static boolean_t
1322zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1323{
1324 kthread_t *executor = zio->io_executor;
1325 spa_t *spa = zio->io_spa;
1326
1327 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1328 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1329 uint_t i;
1330 for (i = 0; i < tqs->stqs_count; i++) {
1331 if (taskq_member(tqs->stqs_taskq[i], executor))
1332 return (B_TRUE);
1333 }
1334 }
1335
1336 return (B_FALSE);
1337}
1338
1339static int
1340zio_issue_async(zio_t *zio)
1341{
1342 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1343
1344 return (ZIO_PIPELINE_STOP);
1345}
1346
1347void
1348zio_interrupt(zio_t *zio)
1349{
1350 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1351}
1352
1353/*
1354 * Execute the I/O pipeline until one of the following occurs:
1355 *
1356 * (1) the I/O completes
1357 * (2) the pipeline stalls waiting for dependent child I/Os
1358 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1359 * (4) the I/O is delegated by vdev-level caching or aggregation
1360 * (5) the I/O is deferred due to vdev-level queueing
1361 * (6) the I/O is handed off to another thread.
1362 *
1363 * In all cases, the pipeline stops whenever there's no CPU work; it never
1364 * burns a thread in cv_wait().
1365 *
1366 * There's no locking on io_stage because there's no legitimate way
1367 * for multiple threads to be attempting to process the same I/O.
1368 */
1369static zio_pipe_stage_t *zio_pipeline[];
1370
1371void
1372zio_execute(zio_t *zio)
1373{
1374 zio->io_executor = curthread;
1375
1376 while (zio->io_stage < ZIO_STAGE_DONE) {
1377 enum zio_stage pipeline = zio->io_pipeline;
1378 enum zio_stage stage = zio->io_stage;
1379 int rv;
1380
1381 ASSERT(!MUTEX_HELD(&zio->io_lock));
1382 ASSERT(ISP2(stage));
1383 ASSERT(zio->io_stall == NULL);
1384
1385 do {
1386 stage <<= 1;
1387 } while ((stage & pipeline) == 0);
1388
1389 ASSERT(stage <= ZIO_STAGE_DONE);
1390
1391 /*
1392 * If we are in interrupt context and this pipeline stage
1393 * will grab a config lock that is held across I/O,
1394 * or may wait for an I/O that needs an interrupt thread
1395 * to complete, issue async to avoid deadlock.
1396 *
1397 * For VDEV_IO_START, we cut in line so that the io will
1398 * be sent to disk promptly.
1399 */
1400 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1401 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1402 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1403 zio_requeue_io_start_cut_in_line : B_FALSE;
1404 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1405 return;
1406 }
1407
1408 zio->io_stage = stage;
1409 rv = zio_pipeline[highbit64(stage) - 1](zio);
1410
1411 if (rv == ZIO_PIPELINE_STOP)
1412 return;
1413
1414 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1415 }
1416}
1417
1418/*
1419 * ==========================================================================
1420 * Initiate I/O, either sync or async
1421 * ==========================================================================
1422 */
1423int
1424zio_wait(zio_t *zio)
1425{
1426 int error;
1427
1428 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1429 ASSERT(zio->io_executor == NULL);
1430
1431 zio->io_waiter = curthread;
1432
1433 zio_execute(zio);
1434
1435 mutex_enter(&zio->io_lock);
1436 while (zio->io_executor != NULL)
1437 cv_wait(&zio->io_cv, &zio->io_lock);
1438 mutex_exit(&zio->io_lock);
1439
1440 error = zio->io_error;
1441 zio_destroy(zio);
1442
1443 return (error);
1444}
1445
1446void
1447zio_nowait(zio_t *zio)
1448{
1449 ASSERT(zio->io_executor == NULL);
1450
1451 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1452 zio_unique_parent(zio) == NULL) {
1453 /*
1454 * This is a logical async I/O with no parent to wait for it.
1455 * We add it to the spa_async_root_zio "Godfather" I/O which
1456 * will ensure they complete prior to unloading the pool.
1457 */
1458 spa_t *spa = zio->io_spa;
1459
| 1/*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21/*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2014 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25 */
26
27#include <sys/zfs_context.h>
28#include <sys/fm/fs/zfs.h>
29#include <sys/spa.h>
30#include <sys/txg.h>
31#include <sys/spa_impl.h>
32#include <sys/vdev_impl.h>
33#include <sys/zio_impl.h>
34#include <sys/zio_compress.h>
35#include <sys/zio_checksum.h>
36#include <sys/dmu_objset.h>
37#include <sys/arc.h>
38#include <sys/ddt.h>
39#include <sys/trim_map.h>
40#include <sys/blkptr.h>
41#include <sys/zfeature.h>
42
43SYSCTL_DECL(_vfs_zfs);
44SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO");
45#if defined(__amd64__)
46static int zio_use_uma = 1;
47#else
48static int zio_use_uma = 0;
49#endif
50SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
51 "Use uma(9) for ZIO allocations");
52static int zio_exclude_metadata = 0;
53SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, exclude_metadata, CTLFLAG_RDTUN, &zio_exclude_metadata, 0,
54 "Exclude metadata buffers from dumps as well");
55
56zio_trim_stats_t zio_trim_stats = {
57 { "bytes", KSTAT_DATA_UINT64,
58 "Number of bytes successfully TRIMmed" },
59 { "success", KSTAT_DATA_UINT64,
60 "Number of successful TRIM requests" },
61 { "unsupported", KSTAT_DATA_UINT64,
62 "Number of TRIM requests that failed because TRIM is not supported" },
63 { "failed", KSTAT_DATA_UINT64,
64 "Number of TRIM requests that failed for reasons other than not supported" },
65};
66
67static kstat_t *zio_trim_ksp;
68
69/*
70 * ==========================================================================
71 * I/O type descriptions
72 * ==========================================================================
73 */
74const char *zio_type_name[ZIO_TYPES] = {
75 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
76 "zio_ioctl"
77};
78
79/*
80 * ==========================================================================
81 * I/O kmem caches
82 * ==========================================================================
83 */
84kmem_cache_t *zio_cache;
85kmem_cache_t *zio_link_cache;
86kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
87kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
88
89#ifdef _KERNEL
90extern vmem_t *zio_alloc_arena;
91#endif
92
93/*
94 * The following actions directly effect the spa's sync-to-convergence logic.
95 * The values below define the sync pass when we start performing the action.
96 * Care should be taken when changing these values as they directly impact
97 * spa_sync() performance. Tuning these values may introduce subtle performance
98 * pathologies and should only be done in the context of performance analysis.
99 * These tunables will eventually be removed and replaced with #defines once
100 * enough analysis has been done to determine optimal values.
101 *
102 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
103 * regular blocks are not deferred.
104 */
105int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
106SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN,
107 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass");
108int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
109SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN,
110 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass");
111int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
112SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN,
113 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass");
114
115/*
116 * An allocating zio is one that either currently has the DVA allocate
117 * stage set or will have it later in its lifetime.
118 */
119#define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
120
121boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
122
123#ifdef ZFS_DEBUG
124int zio_buf_debug_limit = 16384;
125#else
126int zio_buf_debug_limit = 0;
127#endif
128
129void
130zio_init(void)
131{
132 size_t c;
133 zio_cache = kmem_cache_create("zio_cache",
134 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
135 zio_link_cache = kmem_cache_create("zio_link_cache",
136 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
137 if (!zio_use_uma)
138 goto out;
139
140 /*
141 * For small buffers, we want a cache for each multiple of
142 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
143 * for each quarter-power of 2. For large buffers, we want
144 * a cache for each multiple of PAGESIZE.
145 */
146 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
147 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
148 size_t p2 = size;
149 size_t align = 0;
150 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
151
152 while (p2 & (p2 - 1))
153 p2 &= p2 - 1;
154
155#ifdef illumos
156#ifndef _KERNEL
157 /*
158 * If we are using watchpoints, put each buffer on its own page,
159 * to eliminate the performance overhead of trapping to the
160 * kernel when modifying a non-watched buffer that shares the
161 * page with a watched buffer.
162 */
163 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
164 continue;
165#endif
166#endif /* illumos */
167 if (size <= 4 * SPA_MINBLOCKSIZE) {
168 align = SPA_MINBLOCKSIZE;
169 } else if (IS_P2ALIGNED(size, PAGESIZE)) {
170 align = PAGESIZE;
171 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
172 align = p2 >> 2;
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, NULL,
189 cflags | KMC_NOTOUCH | KMC_NODEBUG);
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 }
202out:
203
204 zio_inject_init();
205
206 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc",
207 KSTAT_TYPE_NAMED,
208 sizeof(zio_trim_stats) / sizeof(kstat_named_t),
209 KSTAT_FLAG_VIRTUAL);
210
211 if (zio_trim_ksp != NULL) {
212 zio_trim_ksp->ks_data = &zio_trim_stats;
213 kstat_install(zio_trim_ksp);
214 }
215}
216
217void
218zio_fini(void)
219{
220 size_t c;
221 kmem_cache_t *last_cache = NULL;
222 kmem_cache_t *last_data_cache = NULL;
223
224 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
225 if (zio_buf_cache[c] != last_cache) {
226 last_cache = zio_buf_cache[c];
227 kmem_cache_destroy(zio_buf_cache[c]);
228 }
229 zio_buf_cache[c] = NULL;
230
231 if (zio_data_buf_cache[c] != last_data_cache) {
232 last_data_cache = zio_data_buf_cache[c];
233 kmem_cache_destroy(zio_data_buf_cache[c]);
234 }
235 zio_data_buf_cache[c] = NULL;
236 }
237
238 kmem_cache_destroy(zio_link_cache);
239 kmem_cache_destroy(zio_cache);
240
241 zio_inject_fini();
242
243 if (zio_trim_ksp != NULL) {
244 kstat_delete(zio_trim_ksp);
245 zio_trim_ksp = NULL;
246 }
247}
248
249/*
250 * ==========================================================================
251 * Allocate and free I/O buffers
252 * ==========================================================================
253 */
254
255/*
256 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
257 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
258 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
259 * excess / transient data in-core during a crashdump.
260 */
261void *
262zio_buf_alloc(size_t size)
263{
264 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
265 int flags = zio_exclude_metadata ? KM_NODEBUG : 0;
266
267 ASSERT3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
268
269 if (zio_use_uma)
270 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
271 else
272 return (kmem_alloc(size, KM_SLEEP|flags));
273}
274
275/*
276 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
277 * crashdump if the kernel panics. This exists so that we will limit the amount
278 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
279 * of kernel heap dumped to disk when the kernel panics)
280 */
281void *
282zio_data_buf_alloc(size_t size)
283{
284 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
285
286 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
287
288 if (zio_use_uma)
289 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
290 else
291 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
292}
293
294void
295zio_buf_free(void *buf, size_t size)
296{
297 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
298
299 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
300
301 if (zio_use_uma)
302 kmem_cache_free(zio_buf_cache[c], buf);
303 else
304 kmem_free(buf, size);
305}
306
307void
308zio_data_buf_free(void *buf, size_t size)
309{
310 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
311
312 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
313
314 if (zio_use_uma)
315 kmem_cache_free(zio_data_buf_cache[c], buf);
316 else
317 kmem_free(buf, size);
318}
319
320/*
321 * ==========================================================================
322 * Push and pop I/O transform buffers
323 * ==========================================================================
324 */
325static void
326zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
327 zio_transform_func_t *transform)
328{
329 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
330
331 zt->zt_orig_data = zio->io_data;
332 zt->zt_orig_size = zio->io_size;
333 zt->zt_bufsize = bufsize;
334 zt->zt_transform = transform;
335
336 zt->zt_next = zio->io_transform_stack;
337 zio->io_transform_stack = zt;
338
339 zio->io_data = data;
340 zio->io_size = size;
341}
342
343static void
344zio_pop_transforms(zio_t *zio)
345{
346 zio_transform_t *zt;
347
348 while ((zt = zio->io_transform_stack) != NULL) {
349 if (zt->zt_transform != NULL)
350 zt->zt_transform(zio,
351 zt->zt_orig_data, zt->zt_orig_size);
352
353 if (zt->zt_bufsize != 0)
354 zio_buf_free(zio->io_data, zt->zt_bufsize);
355
356 zio->io_data = zt->zt_orig_data;
357 zio->io_size = zt->zt_orig_size;
358 zio->io_transform_stack = zt->zt_next;
359
360 kmem_free(zt, sizeof (zio_transform_t));
361 }
362}
363
364/*
365 * ==========================================================================
366 * I/O transform callbacks for subblocks and decompression
367 * ==========================================================================
368 */
369static void
370zio_subblock(zio_t *zio, void *data, uint64_t size)
371{
372 ASSERT(zio->io_size > size);
373
374 if (zio->io_type == ZIO_TYPE_READ)
375 bcopy(zio->io_data, data, size);
376}
377
378static void
379zio_decompress(zio_t *zio, void *data, uint64_t size)
380{
381 if (zio->io_error == 0 &&
382 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
383 zio->io_data, data, zio->io_size, size) != 0)
384 zio->io_error = SET_ERROR(EIO);
385}
386
387/*
388 * ==========================================================================
389 * I/O parent/child relationships and pipeline interlocks
390 * ==========================================================================
391 */
392/*
393 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
394 * continue calling these functions until they return NULL.
395 * Otherwise, the next caller will pick up the list walk in
396 * some indeterminate state. (Otherwise every caller would
397 * have to pass in a cookie to keep the state represented by
398 * io_walk_link, which gets annoying.)
399 */
400zio_t *
401zio_walk_parents(zio_t *cio)
402{
403 zio_link_t *zl = cio->io_walk_link;
404 list_t *pl = &cio->io_parent_list;
405
406 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
407 cio->io_walk_link = zl;
408
409 if (zl == NULL)
410 return (NULL);
411
412 ASSERT(zl->zl_child == cio);
413 return (zl->zl_parent);
414}
415
416zio_t *
417zio_walk_children(zio_t *pio)
418{
419 zio_link_t *zl = pio->io_walk_link;
420 list_t *cl = &pio->io_child_list;
421
422 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
423 pio->io_walk_link = zl;
424
425 if (zl == NULL)
426 return (NULL);
427
428 ASSERT(zl->zl_parent == pio);
429 return (zl->zl_child);
430}
431
432zio_t *
433zio_unique_parent(zio_t *cio)
434{
435 zio_t *pio = zio_walk_parents(cio);
436
437 VERIFY(zio_walk_parents(cio) == NULL);
438 return (pio);
439}
440
441void
442zio_add_child(zio_t *pio, zio_t *cio)
443{
444 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
445
446 /*
447 * Logical I/Os can have logical, gang, or vdev children.
448 * Gang I/Os can have gang or vdev children.
449 * Vdev I/Os can only have vdev children.
450 * The following ASSERT captures all of these constraints.
451 */
452 ASSERT(cio->io_child_type <= pio->io_child_type);
453
454 zl->zl_parent = pio;
455 zl->zl_child = cio;
456
457 mutex_enter(&cio->io_lock);
458 mutex_enter(&pio->io_lock);
459
460 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
461
462 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
463 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
464
465 list_insert_head(&pio->io_child_list, zl);
466 list_insert_head(&cio->io_parent_list, zl);
467
468 pio->io_child_count++;
469 cio->io_parent_count++;
470
471 mutex_exit(&pio->io_lock);
472 mutex_exit(&cio->io_lock);
473}
474
475static void
476zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
477{
478 ASSERT(zl->zl_parent == pio);
479 ASSERT(zl->zl_child == cio);
480
481 mutex_enter(&cio->io_lock);
482 mutex_enter(&pio->io_lock);
483
484 list_remove(&pio->io_child_list, zl);
485 list_remove(&cio->io_parent_list, zl);
486
487 pio->io_child_count--;
488 cio->io_parent_count--;
489
490 mutex_exit(&pio->io_lock);
491 mutex_exit(&cio->io_lock);
492
493 kmem_cache_free(zio_link_cache, zl);
494}
495
496static boolean_t
497zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
498{
499 uint64_t *countp = &zio->io_children[child][wait];
500 boolean_t waiting = B_FALSE;
501
502 mutex_enter(&zio->io_lock);
503 ASSERT(zio->io_stall == NULL);
504 if (*countp != 0) {
505 zio->io_stage >>= 1;
506 zio->io_stall = countp;
507 waiting = B_TRUE;
508 }
509 mutex_exit(&zio->io_lock);
510
511 return (waiting);
512}
513
514static void
515zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
516{
517 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
518 int *errorp = &pio->io_child_error[zio->io_child_type];
519
520 mutex_enter(&pio->io_lock);
521 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
522 *errorp = zio_worst_error(*errorp, zio->io_error);
523 pio->io_reexecute |= zio->io_reexecute;
524 ASSERT3U(*countp, >, 0);
525
526 (*countp)--;
527
528 if (*countp == 0 && pio->io_stall == countp) {
529 pio->io_stall = NULL;
530 mutex_exit(&pio->io_lock);
531 zio_execute(pio);
532 } else {
533 mutex_exit(&pio->io_lock);
534 }
535}
536
537static void
538zio_inherit_child_errors(zio_t *zio, enum zio_child c)
539{
540 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
541 zio->io_error = zio->io_child_error[c];
542}
543
544/*
545 * ==========================================================================
546 * Create the various types of I/O (read, write, free, etc)
547 * ==========================================================================
548 */
549static zio_t *
550zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
551 void *data, uint64_t size, zio_done_func_t *done, void *private,
552 zio_type_t type, zio_priority_t priority, enum zio_flag flags,
553 vdev_t *vd, uint64_t offset, const zbookmark_phys_t *zb,
554 enum zio_stage stage, enum zio_stage pipeline)
555{
556 zio_t *zio;
557
558 ASSERT3U(type == ZIO_TYPE_FREE || size, <=, SPA_MAXBLOCKSIZE);
559 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
560 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
561
562 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
563 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
564 ASSERT(vd || stage == ZIO_STAGE_OPEN);
565
566 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
567 bzero(zio, sizeof (zio_t));
568
569 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
570 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
571
572 list_create(&zio->io_parent_list, sizeof (zio_link_t),
573 offsetof(zio_link_t, zl_parent_node));
574 list_create(&zio->io_child_list, sizeof (zio_link_t),
575 offsetof(zio_link_t, zl_child_node));
576
577 if (vd != NULL)
578 zio->io_child_type = ZIO_CHILD_VDEV;
579 else if (flags & ZIO_FLAG_GANG_CHILD)
580 zio->io_child_type = ZIO_CHILD_GANG;
581 else if (flags & ZIO_FLAG_DDT_CHILD)
582 zio->io_child_type = ZIO_CHILD_DDT;
583 else
584 zio->io_child_type = ZIO_CHILD_LOGICAL;
585
586 if (bp != NULL) {
587 zio->io_bp = (blkptr_t *)bp;
588 zio->io_bp_copy = *bp;
589 zio->io_bp_orig = *bp;
590 if (type != ZIO_TYPE_WRITE ||
591 zio->io_child_type == ZIO_CHILD_DDT)
592 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
593 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
594 zio->io_logical = zio;
595 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
596 pipeline |= ZIO_GANG_STAGES;
597 }
598
599 zio->io_spa = spa;
600 zio->io_txg = txg;
601 zio->io_done = done;
602 zio->io_private = private;
603 zio->io_type = type;
604 zio->io_priority = priority;
605 zio->io_vd = vd;
606 zio->io_offset = offset;
607 zio->io_orig_data = zio->io_data = data;
608 zio->io_orig_size = zio->io_size = size;
609 zio->io_orig_flags = zio->io_flags = flags;
610 zio->io_orig_stage = zio->io_stage = stage;
611 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
612
613 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
614 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
615
616 if (zb != NULL)
617 zio->io_bookmark = *zb;
618
619 if (pio != NULL) {
620 if (zio->io_logical == NULL)
621 zio->io_logical = pio->io_logical;
622 if (zio->io_child_type == ZIO_CHILD_GANG)
623 zio->io_gang_leader = pio->io_gang_leader;
624 zio_add_child(pio, zio);
625 }
626
627 return (zio);
628}
629
630static void
631zio_destroy(zio_t *zio)
632{
633 list_destroy(&zio->io_parent_list);
634 list_destroy(&zio->io_child_list);
635 mutex_destroy(&zio->io_lock);
636 cv_destroy(&zio->io_cv);
637 kmem_cache_free(zio_cache, zio);
638}
639
640zio_t *
641zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
642 void *private, enum zio_flag flags)
643{
644 zio_t *zio;
645
646 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
647 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
648 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
649
650 return (zio);
651}
652
653zio_t *
654zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
655{
656 return (zio_null(NULL, spa, NULL, done, private, flags));
657}
658
659zio_t *
660zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
661 void *data, uint64_t size, zio_done_func_t *done, void *private,
662 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
663{
664 zio_t *zio;
665
666 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
667 data, size, done, private,
668 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
669 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
670 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
671
672 return (zio);
673}
674
675zio_t *
676zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
677 void *data, uint64_t size, const zio_prop_t *zp,
678 zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done,
679 void *private,
680 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
681{
682 zio_t *zio;
683
684 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
685 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
686 zp->zp_compress >= ZIO_COMPRESS_OFF &&
687 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
688 DMU_OT_IS_VALID(zp->zp_type) &&
689 zp->zp_level < 32 &&
690 zp->zp_copies > 0 &&
691 zp->zp_copies <= spa_max_replication(spa));
692
693 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
694 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
695 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
696 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
697
698 zio->io_ready = ready;
699 zio->io_physdone = physdone;
700 zio->io_prop = *zp;
701
702 /*
703 * Data can be NULL if we are going to call zio_write_override() to
704 * provide the already-allocated BP. But we may need the data to
705 * verify a dedup hit (if requested). In this case, don't try to
706 * dedup (just take the already-allocated BP verbatim).
707 */
708 if (data == NULL && zio->io_prop.zp_dedup_verify) {
709 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
710 }
711
712 return (zio);
713}
714
715zio_t *
716zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
717 uint64_t size, zio_done_func_t *done, void *private,
718 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
719{
720 zio_t *zio;
721
722 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
723 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
724 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
725
726 return (zio);
727}
728
729void
730zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
731{
732 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
733 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
734 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
735 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
736
737 /*
738 * We must reset the io_prop to match the values that existed
739 * when the bp was first written by dmu_sync() keeping in mind
740 * that nopwrite and dedup are mutually exclusive.
741 */
742 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
743 zio->io_prop.zp_nopwrite = nopwrite;
744 zio->io_prop.zp_copies = copies;
745 zio->io_bp_override = bp;
746}
747
748void
749zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
750{
751
752 /*
753 * The check for EMBEDDED is a performance optimization. We
754 * process the free here (by ignoring it) rather than
755 * putting it on the list and then processing it in zio_free_sync().
756 */
757 if (BP_IS_EMBEDDED(bp))
758 return;
759 metaslab_check_free(spa, bp);
760
761 /*
762 * Frees that are for the currently-syncing txg, are not going to be
763 * deferred, and which will not need to do a read (i.e. not GANG or
764 * DEDUP), can be processed immediately. Otherwise, put them on the
765 * in-memory list for later processing.
766 */
767 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
768 txg != spa->spa_syncing_txg ||
769 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
770 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
771 } else {
772 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
773 BP_GET_PSIZE(bp), 0)));
774 }
775}
776
777zio_t *
778zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
779 uint64_t size, enum zio_flag flags)
780{
781 zio_t *zio;
782 enum zio_stage stage = ZIO_FREE_PIPELINE;
783
784 ASSERT(!BP_IS_HOLE(bp));
785 ASSERT(spa_syncing_txg(spa) == txg);
786 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
787
788 if (BP_IS_EMBEDDED(bp))
789 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
790
791 metaslab_check_free(spa, bp);
792 arc_freed(spa, bp);
793
794 if (zfs_trim_enabled)
795 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
796 ZIO_STAGE_VDEV_IO_ASSESS;
797 /*
798 * GANG and DEDUP blocks can induce a read (for the gang block header,
799 * or the DDT), so issue them asynchronously so that this thread is
800 * not tied up.
801 */
802 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
803 stage |= ZIO_STAGE_ISSUE_ASYNC;
804
805 flags |= ZIO_FLAG_DONT_QUEUE;
806
807 zio = zio_create(pio, spa, txg, bp, NULL, size,
808 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags,
809 NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
810
811 return (zio);
812}
813
814zio_t *
815zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
816 zio_done_func_t *done, void *private, enum zio_flag flags)
817{
818 zio_t *zio;
819
820 dprintf_bp(bp, "claiming in txg %llu", txg);
821
822 if (BP_IS_EMBEDDED(bp))
823 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
824
825 /*
826 * A claim is an allocation of a specific block. Claims are needed
827 * to support immediate writes in the intent log. The issue is that
828 * immediate writes contain committed data, but in a txg that was
829 * *not* committed. Upon opening the pool after an unclean shutdown,
830 * the intent log claims all blocks that contain immediate write data
831 * so that the SPA knows they're in use.
832 *
833 * All claims *must* be resolved in the first txg -- before the SPA
834 * starts allocating blocks -- so that nothing is allocated twice.
835 * If txg == 0 we just verify that the block is claimable.
836 */
837 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
838 ASSERT(txg == spa_first_txg(spa) || txg == 0);
839 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
840
841 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
842 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
843 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
844
845 return (zio);
846}
847
848zio_t *
849zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
850 uint64_t size, zio_done_func_t *done, void *private,
851 zio_priority_t priority, enum zio_flag flags)
852{
853 zio_t *zio;
854 int c;
855
856 if (vd->vdev_children == 0) {
857 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private,
858 ZIO_TYPE_IOCTL, priority, flags, vd, offset, NULL,
859 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
860
861 zio->io_cmd = cmd;
862 } else {
863 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
864
865 for (c = 0; c < vd->vdev_children; c++)
866 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
867 offset, size, done, private, priority, flags));
868 }
869
870 return (zio);
871}
872
873zio_t *
874zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
875 void *data, int checksum, zio_done_func_t *done, void *private,
876 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
877{
878 zio_t *zio;
879
880 ASSERT(vd->vdev_children == 0);
881 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
882 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
883 ASSERT3U(offset + size, <=, vd->vdev_psize);
884
885 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
886 ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
887 NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
888
889 zio->io_prop.zp_checksum = checksum;
890
891 return (zio);
892}
893
894zio_t *
895zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
896 void *data, int checksum, zio_done_func_t *done, void *private,
897 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
898{
899 zio_t *zio;
900
901 ASSERT(vd->vdev_children == 0);
902 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
903 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
904 ASSERT3U(offset + size, <=, vd->vdev_psize);
905
906 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
907 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
908 NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
909
910 zio->io_prop.zp_checksum = checksum;
911
912 if (zio_checksum_table[checksum].ci_eck) {
913 /*
914 * zec checksums are necessarily destructive -- they modify
915 * the end of the write buffer to hold the verifier/checksum.
916 * Therefore, we must make a local copy in case the data is
917 * being written to multiple places in parallel.
918 */
919 void *wbuf = zio_buf_alloc(size);
920 bcopy(data, wbuf, size);
921 zio_push_transform(zio, wbuf, size, size, NULL);
922 }
923
924 return (zio);
925}
926
927/*
928 * Create a child I/O to do some work for us.
929 */
930zio_t *
931zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
932 void *data, uint64_t size, int type, zio_priority_t priority,
933 enum zio_flag flags, zio_done_func_t *done, void *private)
934{
935 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
936 zio_t *zio;
937
938 ASSERT(vd->vdev_parent ==
939 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
940
941 if (type == ZIO_TYPE_READ && bp != NULL) {
942 /*
943 * If we have the bp, then the child should perform the
944 * checksum and the parent need not. This pushes error
945 * detection as close to the leaves as possible and
946 * eliminates redundant checksums in the interior nodes.
947 */
948 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
949 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
950 }
951
952 /* Not all IO types require vdev io done stage e.g. free */
953 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE))
954 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE;
955
956 if (vd->vdev_children == 0)
957 offset += VDEV_LABEL_START_SIZE;
958
959 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
960
961 /*
962 * If we've decided to do a repair, the write is not speculative --
963 * even if the original read was.
964 */
965 if (flags & ZIO_FLAG_IO_REPAIR)
966 flags &= ~ZIO_FLAG_SPECULATIVE;
967
968 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
969 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
970 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
971
972 zio->io_physdone = pio->io_physdone;
973 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
974 zio->io_logical->io_phys_children++;
975
976 return (zio);
977}
978
979zio_t *
980zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
981 int type, zio_priority_t priority, enum zio_flag flags,
982 zio_done_func_t *done, void *private)
983{
984 zio_t *zio;
985
986 ASSERT(vd->vdev_ops->vdev_op_leaf);
987
988 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
989 data, size, done, private, type, priority,
990 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
991 vd, offset, NULL,
992 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
993
994 return (zio);
995}
996
997void
998zio_flush(zio_t *zio, vdev_t *vd)
999{
1000 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
1001 NULL, NULL, ZIO_PRIORITY_NOW,
1002 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1003}
1004
1005zio_t *
1006zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
1007{
1008
1009 ASSERT(vd->vdev_ops->vdev_op_leaf);
1010
1011 return (zio_create(zio, spa, 0, NULL, NULL, size, NULL, NULL,
1012 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE |
1013 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY,
1014 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE));
1015}
1016
1017void
1018zio_shrink(zio_t *zio, uint64_t size)
1019{
1020 ASSERT(zio->io_executor == NULL);
1021 ASSERT(zio->io_orig_size == zio->io_size);
1022 ASSERT(size <= zio->io_size);
1023
1024 /*
1025 * We don't shrink for raidz because of problems with the
1026 * reconstruction when reading back less than the block size.
1027 * Note, BP_IS_RAIDZ() assumes no compression.
1028 */
1029 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1030 if (!BP_IS_RAIDZ(zio->io_bp))
1031 zio->io_orig_size = zio->io_size = size;
1032}
1033
1034/*
1035 * ==========================================================================
1036 * Prepare to read and write logical blocks
1037 * ==========================================================================
1038 */
1039
1040static int
1041zio_read_bp_init(zio_t *zio)
1042{
1043 blkptr_t *bp = zio->io_bp;
1044
1045 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1046 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1047 !(zio->io_flags & ZIO_FLAG_RAW)) {
1048 uint64_t psize =
1049 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1050 void *cbuf = zio_buf_alloc(psize);
1051
1052 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1053 }
1054
1055 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1056 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1057 decode_embedded_bp_compressed(bp, zio->io_data);
1058 } else {
1059 ASSERT(!BP_IS_EMBEDDED(bp));
1060 }
1061
1062 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1063 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1064
1065 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1066 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1067
1068 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1069 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1070
1071 return (ZIO_PIPELINE_CONTINUE);
1072}
1073
1074static int
1075zio_write_bp_init(zio_t *zio)
1076{
1077 spa_t *spa = zio->io_spa;
1078 zio_prop_t *zp = &zio->io_prop;
1079 enum zio_compress compress = zp->zp_compress;
1080 blkptr_t *bp = zio->io_bp;
1081 uint64_t lsize = zio->io_size;
1082 uint64_t psize = lsize;
1083 int pass = 1;
1084
1085 /*
1086 * If our children haven't all reached the ready stage,
1087 * wait for them and then repeat this pipeline stage.
1088 */
1089 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1090 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1091 return (ZIO_PIPELINE_STOP);
1092
1093 if (!IO_IS_ALLOCATING(zio))
1094 return (ZIO_PIPELINE_CONTINUE);
1095
1096 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1097
1098 if (zio->io_bp_override) {
1099 ASSERT(bp->blk_birth != zio->io_txg);
1100 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1101
1102 *bp = *zio->io_bp_override;
1103 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1104
1105 if (BP_IS_EMBEDDED(bp))
1106 return (ZIO_PIPELINE_CONTINUE);
1107
1108 /*
1109 * If we've been overridden and nopwrite is set then
1110 * set the flag accordingly to indicate that a nopwrite
1111 * has already occurred.
1112 */
1113 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1114 ASSERT(!zp->zp_dedup);
1115 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1116 return (ZIO_PIPELINE_CONTINUE);
1117 }
1118
1119 ASSERT(!zp->zp_nopwrite);
1120
1121 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1122 return (ZIO_PIPELINE_CONTINUE);
1123
1124 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1125 zp->zp_dedup_verify);
1126
1127 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1128 BP_SET_DEDUP(bp, 1);
1129 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1130 return (ZIO_PIPELINE_CONTINUE);
1131 }
1132 zio->io_bp_override = NULL;
1133 BP_ZERO(bp);
1134 }
1135
1136 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1137 /*
1138 * We're rewriting an existing block, which means we're
1139 * working on behalf of spa_sync(). For spa_sync() to
1140 * converge, it must eventually be the case that we don't
1141 * have to allocate new blocks. But compression changes
1142 * the blocksize, which forces a reallocate, and makes
1143 * convergence take longer. Therefore, after the first
1144 * few passes, stop compressing to ensure convergence.
1145 */
1146 pass = spa_sync_pass(spa);
1147
1148 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1149 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1150 ASSERT(!BP_GET_DEDUP(bp));
1151
1152 if (pass >= zfs_sync_pass_dont_compress)
1153 compress = ZIO_COMPRESS_OFF;
1154
1155 /* Make sure someone doesn't change their mind on overwrites */
1156 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1157 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1158 }
1159
1160 if (compress != ZIO_COMPRESS_OFF) {
1161 void *cbuf = zio_buf_alloc(lsize);
1162 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1163 if (psize == 0 || psize == lsize) {
1164 compress = ZIO_COMPRESS_OFF;
1165 zio_buf_free(cbuf, lsize);
1166 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1167 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1168 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1169 encode_embedded_bp_compressed(bp,
1170 cbuf, compress, lsize, psize);
1171 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1172 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1173 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1174 zio_buf_free(cbuf, lsize);
1175 bp->blk_birth = zio->io_txg;
1176 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1177 ASSERT(spa_feature_is_active(spa,
1178 SPA_FEATURE_EMBEDDED_DATA));
1179 return (ZIO_PIPELINE_CONTINUE);
1180 } else {
1181 /*
1182 * Round up compressed size to MINBLOCKSIZE and
1183 * zero the tail.
1184 */
1185 size_t rounded =
1186 P2ROUNDUP(psize, (size_t)SPA_MINBLOCKSIZE);
1187 if (rounded > psize) {
1188 bzero((char *)cbuf + psize, rounded - psize);
1189 psize = rounded;
1190 }
1191 if (psize == lsize) {
1192 compress = ZIO_COMPRESS_OFF;
1193 zio_buf_free(cbuf, lsize);
1194 } else {
1195 zio_push_transform(zio, cbuf,
1196 psize, lsize, NULL);
1197 }
1198 }
1199 }
1200
1201 /*
1202 * The final pass of spa_sync() must be all rewrites, but the first
1203 * few passes offer a trade-off: allocating blocks defers convergence,
1204 * but newly allocated blocks are sequential, so they can be written
1205 * to disk faster. Therefore, we allow the first few passes of
1206 * spa_sync() to allocate new blocks, but force rewrites after that.
1207 * There should only be a handful of blocks after pass 1 in any case.
1208 */
1209 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1210 BP_GET_PSIZE(bp) == psize &&
1211 pass >= zfs_sync_pass_rewrite) {
1212 ASSERT(psize != 0);
1213 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1214 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1215 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1216 } else {
1217 BP_ZERO(bp);
1218 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1219 }
1220
1221 if (psize == 0) {
1222 if (zio->io_bp_orig.blk_birth != 0 &&
1223 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1224 BP_SET_LSIZE(bp, lsize);
1225 BP_SET_TYPE(bp, zp->zp_type);
1226 BP_SET_LEVEL(bp, zp->zp_level);
1227 BP_SET_BIRTH(bp, zio->io_txg, 0);
1228 }
1229 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1230 } else {
1231 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1232 BP_SET_LSIZE(bp, lsize);
1233 BP_SET_TYPE(bp, zp->zp_type);
1234 BP_SET_LEVEL(bp, zp->zp_level);
1235 BP_SET_PSIZE(bp, psize);
1236 BP_SET_COMPRESS(bp, compress);
1237 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1238 BP_SET_DEDUP(bp, zp->zp_dedup);
1239 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1240 if (zp->zp_dedup) {
1241 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1242 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1243 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1244 }
1245 if (zp->zp_nopwrite) {
1246 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1247 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1248 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1249 }
1250 }
1251
1252 return (ZIO_PIPELINE_CONTINUE);
1253}
1254
1255static int
1256zio_free_bp_init(zio_t *zio)
1257{
1258 blkptr_t *bp = zio->io_bp;
1259
1260 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1261 if (BP_GET_DEDUP(bp))
1262 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1263 }
1264
1265 return (ZIO_PIPELINE_CONTINUE);
1266}
1267
1268/*
1269 * ==========================================================================
1270 * Execute the I/O pipeline
1271 * ==========================================================================
1272 */
1273
1274static void
1275zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1276{
1277 spa_t *spa = zio->io_spa;
1278 zio_type_t t = zio->io_type;
1279 int flags = (cutinline ? TQ_FRONT : 0);
1280
1281 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1282
1283 /*
1284 * If we're a config writer or a probe, the normal issue and
1285 * interrupt threads may all be blocked waiting for the config lock.
1286 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1287 */
1288 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1289 t = ZIO_TYPE_NULL;
1290
1291 /*
1292 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1293 */
1294 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1295 t = ZIO_TYPE_NULL;
1296
1297 /*
1298 * If this is a high priority I/O, then use the high priority taskq if
1299 * available.
1300 */
1301 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1302 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1303 q++;
1304
1305 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1306
1307 /*
1308 * NB: We are assuming that the zio can only be dispatched
1309 * to a single taskq at a time. It would be a grievous error
1310 * to dispatch the zio to another taskq at the same time.
1311 */
1312#if defined(illumos) || !defined(_KERNEL)
1313 ASSERT(zio->io_tqent.tqent_next == NULL);
1314#else
1315 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1316#endif
1317 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1318 flags, &zio->io_tqent);
1319}
1320
1321static boolean_t
1322zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1323{
1324 kthread_t *executor = zio->io_executor;
1325 spa_t *spa = zio->io_spa;
1326
1327 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1328 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1329 uint_t i;
1330 for (i = 0; i < tqs->stqs_count; i++) {
1331 if (taskq_member(tqs->stqs_taskq[i], executor))
1332 return (B_TRUE);
1333 }
1334 }
1335
1336 return (B_FALSE);
1337}
1338
1339static int
1340zio_issue_async(zio_t *zio)
1341{
1342 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1343
1344 return (ZIO_PIPELINE_STOP);
1345}
1346
1347void
1348zio_interrupt(zio_t *zio)
1349{
1350 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1351}
1352
1353/*
1354 * Execute the I/O pipeline until one of the following occurs:
1355 *
1356 * (1) the I/O completes
1357 * (2) the pipeline stalls waiting for dependent child I/Os
1358 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1359 * (4) the I/O is delegated by vdev-level caching or aggregation
1360 * (5) the I/O is deferred due to vdev-level queueing
1361 * (6) the I/O is handed off to another thread.
1362 *
1363 * In all cases, the pipeline stops whenever there's no CPU work; it never
1364 * burns a thread in cv_wait().
1365 *
1366 * There's no locking on io_stage because there's no legitimate way
1367 * for multiple threads to be attempting to process the same I/O.
1368 */
1369static zio_pipe_stage_t *zio_pipeline[];
1370
1371void
1372zio_execute(zio_t *zio)
1373{
1374 zio->io_executor = curthread;
1375
1376 while (zio->io_stage < ZIO_STAGE_DONE) {
1377 enum zio_stage pipeline = zio->io_pipeline;
1378 enum zio_stage stage = zio->io_stage;
1379 int rv;
1380
1381 ASSERT(!MUTEX_HELD(&zio->io_lock));
1382 ASSERT(ISP2(stage));
1383 ASSERT(zio->io_stall == NULL);
1384
1385 do {
1386 stage <<= 1;
1387 } while ((stage & pipeline) == 0);
1388
1389 ASSERT(stage <= ZIO_STAGE_DONE);
1390
1391 /*
1392 * If we are in interrupt context and this pipeline stage
1393 * will grab a config lock that is held across I/O,
1394 * or may wait for an I/O that needs an interrupt thread
1395 * to complete, issue async to avoid deadlock.
1396 *
1397 * For VDEV_IO_START, we cut in line so that the io will
1398 * be sent to disk promptly.
1399 */
1400 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1401 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1402 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1403 zio_requeue_io_start_cut_in_line : B_FALSE;
1404 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1405 return;
1406 }
1407
1408 zio->io_stage = stage;
1409 rv = zio_pipeline[highbit64(stage) - 1](zio);
1410
1411 if (rv == ZIO_PIPELINE_STOP)
1412 return;
1413
1414 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1415 }
1416}
1417
1418/*
1419 * ==========================================================================
1420 * Initiate I/O, either sync or async
1421 * ==========================================================================
1422 */
1423int
1424zio_wait(zio_t *zio)
1425{
1426 int error;
1427
1428 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1429 ASSERT(zio->io_executor == NULL);
1430
1431 zio->io_waiter = curthread;
1432
1433 zio_execute(zio);
1434
1435 mutex_enter(&zio->io_lock);
1436 while (zio->io_executor != NULL)
1437 cv_wait(&zio->io_cv, &zio->io_lock);
1438 mutex_exit(&zio->io_lock);
1439
1440 error = zio->io_error;
1441 zio_destroy(zio);
1442
1443 return (error);
1444}
1445
1446void
1447zio_nowait(zio_t *zio)
1448{
1449 ASSERT(zio->io_executor == NULL);
1450
1451 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1452 zio_unique_parent(zio) == NULL) {
1453 /*
1454 * This is a logical async I/O with no parent to wait for it.
1455 * We add it to the spa_async_root_zio "Godfather" I/O which
1456 * will ensure they complete prior to unloading the pool.
1457 */
1458 spa_t *spa = zio->io_spa;
1459
|
1461 }
1462
1463 zio_execute(zio);
1464}
1465
1466/*
1467 * ==========================================================================
1468 * Reexecute or suspend/resume failed I/O
1469 * ==========================================================================
1470 */
1471
1472static void
1473zio_reexecute(zio_t *pio)
1474{
1475 zio_t *cio, *cio_next;
1476
1477 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1478 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1479 ASSERT(pio->io_gang_leader == NULL);
1480 ASSERT(pio->io_gang_tree == NULL);
1481
1482 pio->io_flags = pio->io_orig_flags;
1483 pio->io_stage = pio->io_orig_stage;
1484 pio->io_pipeline = pio->io_orig_pipeline;
1485 pio->io_reexecute = 0;
1486 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1487 pio->io_error = 0;
1488 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1489 pio->io_state[w] = 0;
1490 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1491 pio->io_child_error[c] = 0;
1492
1493 if (IO_IS_ALLOCATING(pio))
1494 BP_ZERO(pio->io_bp);
1495
1496 /*
1497 * As we reexecute pio's children, new children could be created.
1498 * New children go to the head of pio's io_child_list, however,
1499 * so we will (correctly) not reexecute them. The key is that
1500 * the remainder of pio's io_child_list, from 'cio_next' onward,
1501 * cannot be affected by any side effects of reexecuting 'cio'.
1502 */
1503 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1504 cio_next = zio_walk_children(pio);
1505 mutex_enter(&pio->io_lock);
1506 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1507 pio->io_children[cio->io_child_type][w]++;
1508 mutex_exit(&pio->io_lock);
1509 zio_reexecute(cio);
1510 }
1511
1512 /*
1513 * Now that all children have been reexecuted, execute the parent.
1514 * We don't reexecute "The Godfather" I/O here as it's the
1515 * responsibility of the caller to wait on him.
1516 */
1517 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1518 zio_execute(pio);
1519}
1520
1521void
1522zio_suspend(spa_t *spa, zio_t *zio)
1523{
1524 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1525 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1526 "failure and the failure mode property for this pool "
1527 "is set to panic.", spa_name(spa));
1528
1529 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1530
1531 mutex_enter(&spa->spa_suspend_lock);
1532
1533 if (spa->spa_suspend_zio_root == NULL)
1534 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1535 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1536 ZIO_FLAG_GODFATHER);
1537
1538 spa->spa_suspended = B_TRUE;
1539
1540 if (zio != NULL) {
1541 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1542 ASSERT(zio != spa->spa_suspend_zio_root);
1543 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1544 ASSERT(zio_unique_parent(zio) == NULL);
1545 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1546 zio_add_child(spa->spa_suspend_zio_root, zio);
1547 }
1548
1549 mutex_exit(&spa->spa_suspend_lock);
1550}
1551
1552int
1553zio_resume(spa_t *spa)
1554{
1555 zio_t *pio;
1556
1557 /*
1558 * Reexecute all previously suspended i/o.
1559 */
1560 mutex_enter(&spa->spa_suspend_lock);
1561 spa->spa_suspended = B_FALSE;
1562 cv_broadcast(&spa->spa_suspend_cv);
1563 pio = spa->spa_suspend_zio_root;
1564 spa->spa_suspend_zio_root = NULL;
1565 mutex_exit(&spa->spa_suspend_lock);
1566
1567 if (pio == NULL)
1568 return (0);
1569
1570 zio_reexecute(pio);
1571 return (zio_wait(pio));
1572}
1573
1574void
1575zio_resume_wait(spa_t *spa)
1576{
1577 mutex_enter(&spa->spa_suspend_lock);
1578 while (spa_suspended(spa))
1579 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1580 mutex_exit(&spa->spa_suspend_lock);
1581}
1582
1583/*
1584 * ==========================================================================
1585 * Gang blocks.
1586 *
1587 * A gang block is a collection of small blocks that looks to the DMU
1588 * like one large block. When zio_dva_allocate() cannot find a block
1589 * of the requested size, due to either severe fragmentation or the pool
1590 * being nearly full, it calls zio_write_gang_block() to construct the
1591 * block from smaller fragments.
1592 *
1593 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1594 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1595 * an indirect block: it's an array of block pointers. It consumes
1596 * only one sector and hence is allocatable regardless of fragmentation.
1597 * The gang header's bps point to its gang members, which hold the data.
1598 *
1599 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1600 * as the verifier to ensure uniqueness of the SHA256 checksum.
1601 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1602 * not the gang header. This ensures that data block signatures (needed for
1603 * deduplication) are independent of how the block is physically stored.
1604 *
1605 * Gang blocks can be nested: a gang member may itself be a gang block.
1606 * Thus every gang block is a tree in which root and all interior nodes are
1607 * gang headers, and the leaves are normal blocks that contain user data.
1608 * The root of the gang tree is called the gang leader.
1609 *
1610 * To perform any operation (read, rewrite, free, claim) on a gang block,
1611 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1612 * in the io_gang_tree field of the original logical i/o by recursively
1613 * reading the gang leader and all gang headers below it. This yields
1614 * an in-core tree containing the contents of every gang header and the
1615 * bps for every constituent of the gang block.
1616 *
1617 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1618 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1619 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1620 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1621 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1622 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1623 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1624 * of the gang header plus zio_checksum_compute() of the data to update the
1625 * gang header's blk_cksum as described above.
1626 *
1627 * The two-phase assemble/issue model solves the problem of partial failure --
1628 * what if you'd freed part of a gang block but then couldn't read the
1629 * gang header for another part? Assembling the entire gang tree first
1630 * ensures that all the necessary gang header I/O has succeeded before
1631 * starting the actual work of free, claim, or write. Once the gang tree
1632 * is assembled, free and claim are in-memory operations that cannot fail.
1633 *
1634 * In the event that a gang write fails, zio_dva_unallocate() walks the
1635 * gang tree to immediately free (i.e. insert back into the space map)
1636 * everything we've allocated. This ensures that we don't get ENOSPC
1637 * errors during repeated suspend/resume cycles due to a flaky device.
1638 *
1639 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1640 * the gang tree, we won't modify the block, so we can safely defer the free
1641 * (knowing that the block is still intact). If we *can* assemble the gang
1642 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1643 * each constituent bp and we can allocate a new block on the next sync pass.
1644 *
1645 * In all cases, the gang tree allows complete recovery from partial failure.
1646 * ==========================================================================
1647 */
1648
1649static zio_t *
1650zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1651{
1652 if (gn != NULL)
1653 return (pio);
1654
1655 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1656 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1657 &pio->io_bookmark));
1658}
1659
1660zio_t *
1661zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1662{
1663 zio_t *zio;
1664
1665 if (gn != NULL) {
1666 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1667 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1668 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1669 /*
1670 * As we rewrite each gang header, the pipeline will compute
1671 * a new gang block header checksum for it; but no one will
1672 * compute a new data checksum, so we do that here. The one
1673 * exception is the gang leader: the pipeline already computed
1674 * its data checksum because that stage precedes gang assembly.
1675 * (Presently, nothing actually uses interior data checksums;
1676 * this is just good hygiene.)
1677 */
1678 if (gn != pio->io_gang_leader->io_gang_tree) {
1679 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1680 data, BP_GET_PSIZE(bp));
1681 }
1682 /*
1683 * If we are here to damage data for testing purposes,
1684 * leave the GBH alone so that we can detect the damage.
1685 */
1686 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1687 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1688 } else {
1689 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1690 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1691 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1692 }
1693
1694 return (zio);
1695}
1696
1697/* ARGSUSED */
1698zio_t *
1699zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1700{
1701 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1702 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
1703 ZIO_GANG_CHILD_FLAGS(pio)));
1704}
1705
1706/* ARGSUSED */
1707zio_t *
1708zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1709{
1710 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1711 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1712}
1713
1714static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1715 NULL,
1716 zio_read_gang,
1717 zio_rewrite_gang,
1718 zio_free_gang,
1719 zio_claim_gang,
1720 NULL
1721};
1722
1723static void zio_gang_tree_assemble_done(zio_t *zio);
1724
1725static zio_gang_node_t *
1726zio_gang_node_alloc(zio_gang_node_t **gnpp)
1727{
1728 zio_gang_node_t *gn;
1729
1730 ASSERT(*gnpp == NULL);
1731
1732 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1733 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1734 *gnpp = gn;
1735
1736 return (gn);
1737}
1738
1739static void
1740zio_gang_node_free(zio_gang_node_t **gnpp)
1741{
1742 zio_gang_node_t *gn = *gnpp;
1743
1744 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1745 ASSERT(gn->gn_child[g] == NULL);
1746
1747 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1748 kmem_free(gn, sizeof (*gn));
1749 *gnpp = NULL;
1750}
1751
1752static void
1753zio_gang_tree_free(zio_gang_node_t **gnpp)
1754{
1755 zio_gang_node_t *gn = *gnpp;
1756
1757 if (gn == NULL)
1758 return;
1759
1760 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1761 zio_gang_tree_free(&gn->gn_child[g]);
1762
1763 zio_gang_node_free(gnpp);
1764}
1765
1766static void
1767zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1768{
1769 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1770
1771 ASSERT(gio->io_gang_leader == gio);
1772 ASSERT(BP_IS_GANG(bp));
1773
1774 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1775 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1776 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1777}
1778
1779static void
1780zio_gang_tree_assemble_done(zio_t *zio)
1781{
1782 zio_t *gio = zio->io_gang_leader;
1783 zio_gang_node_t *gn = zio->io_private;
1784 blkptr_t *bp = zio->io_bp;
1785
1786 ASSERT(gio == zio_unique_parent(zio));
1787 ASSERT(zio->io_child_count == 0);
1788
1789 if (zio->io_error)
1790 return;
1791
1792 if (BP_SHOULD_BYTESWAP(bp))
1793 byteswap_uint64_array(zio->io_data, zio->io_size);
1794
1795 ASSERT(zio->io_data == gn->gn_gbh);
1796 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1797 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1798
1799 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1800 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1801 if (!BP_IS_GANG(gbp))
1802 continue;
1803 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1804 }
1805}
1806
1807static void
1808zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1809{
1810 zio_t *gio = pio->io_gang_leader;
1811 zio_t *zio;
1812
1813 ASSERT(BP_IS_GANG(bp) == !!gn);
1814 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1815 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1816
1817 /*
1818 * If you're a gang header, your data is in gn->gn_gbh.
1819 * If you're a gang member, your data is in 'data' and gn == NULL.
1820 */
1821 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1822
1823 if (gn != NULL) {
1824 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1825
1826 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1827 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1828 if (BP_IS_HOLE(gbp))
1829 continue;
1830 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1831 data = (char *)data + BP_GET_PSIZE(gbp);
1832 }
1833 }
1834
1835 if (gn == gio->io_gang_tree && gio->io_data != NULL)
1836 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1837
1838 if (zio != pio)
1839 zio_nowait(zio);
1840}
1841
1842static int
1843zio_gang_assemble(zio_t *zio)
1844{
1845 blkptr_t *bp = zio->io_bp;
1846
1847 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1848 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1849
1850 zio->io_gang_leader = zio;
1851
1852 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1853
1854 return (ZIO_PIPELINE_CONTINUE);
1855}
1856
1857static int
1858zio_gang_issue(zio_t *zio)
1859{
1860 blkptr_t *bp = zio->io_bp;
1861
1862 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1863 return (ZIO_PIPELINE_STOP);
1864
1865 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1866 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1867
1868 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1869 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1870 else
1871 zio_gang_tree_free(&zio->io_gang_tree);
1872
1873 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1874
1875 return (ZIO_PIPELINE_CONTINUE);
1876}
1877
1878static void
1879zio_write_gang_member_ready(zio_t *zio)
1880{
1881 zio_t *pio = zio_unique_parent(zio);
1882 zio_t *gio = zio->io_gang_leader;
1883 dva_t *cdva = zio->io_bp->blk_dva;
1884 dva_t *pdva = pio->io_bp->blk_dva;
1885 uint64_t asize;
1886
1887 if (BP_IS_HOLE(zio->io_bp))
1888 return;
1889
1890 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1891
1892 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1893 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1894 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1895 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1896 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1897
1898 mutex_enter(&pio->io_lock);
1899 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1900 ASSERT(DVA_GET_GANG(&pdva[d]));
1901 asize = DVA_GET_ASIZE(&pdva[d]);
1902 asize += DVA_GET_ASIZE(&cdva[d]);
1903 DVA_SET_ASIZE(&pdva[d], asize);
1904 }
1905 mutex_exit(&pio->io_lock);
1906}
1907
1908static int
1909zio_write_gang_block(zio_t *pio)
1910{
1911 spa_t *spa = pio->io_spa;
1912 blkptr_t *bp = pio->io_bp;
1913 zio_t *gio = pio->io_gang_leader;
1914 zio_t *zio;
1915 zio_gang_node_t *gn, **gnpp;
1916 zio_gbh_phys_t *gbh;
1917 uint64_t txg = pio->io_txg;
1918 uint64_t resid = pio->io_size;
1919 uint64_t lsize;
1920 int copies = gio->io_prop.zp_copies;
1921 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1922 zio_prop_t zp;
1923 int error;
1924
1925 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1926 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1927 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1928 if (error) {
1929 pio->io_error = error;
1930 return (ZIO_PIPELINE_CONTINUE);
1931 }
1932
1933 if (pio == gio) {
1934 gnpp = &gio->io_gang_tree;
1935 } else {
1936 gnpp = pio->io_private;
1937 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1938 }
1939
1940 gn = zio_gang_node_alloc(gnpp);
1941 gbh = gn->gn_gbh;
1942 bzero(gbh, SPA_GANGBLOCKSIZE);
1943
1944 /*
1945 * Create the gang header.
1946 */
1947 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1948 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1949
1950 /*
1951 * Create and nowait the gang children.
1952 */
1953 for (int g = 0; resid != 0; resid -= lsize, g++) {
1954 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1955 SPA_MINBLOCKSIZE);
1956 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1957
1958 zp.zp_checksum = gio->io_prop.zp_checksum;
1959 zp.zp_compress = ZIO_COMPRESS_OFF;
1960 zp.zp_type = DMU_OT_NONE;
1961 zp.zp_level = 0;
1962 zp.zp_copies = gio->io_prop.zp_copies;
1963 zp.zp_dedup = B_FALSE;
1964 zp.zp_dedup_verify = B_FALSE;
1965 zp.zp_nopwrite = B_FALSE;
1966
1967 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1968 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1969 zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g],
1970 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1971 &pio->io_bookmark));
1972 }
1973
1974 /*
1975 * Set pio's pipeline to just wait for zio to finish.
1976 */
1977 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1978
1979 zio_nowait(zio);
1980
1981 return (ZIO_PIPELINE_CONTINUE);
1982}
1983
1984/*
1985 * The zio_nop_write stage in the pipeline determines if allocating
1986 * a new bp is necessary. By leveraging a cryptographically secure checksum,
1987 * such as SHA256, we can compare the checksums of the new data and the old
1988 * to determine if allocating a new block is required. The nopwrite
1989 * feature can handle writes in either syncing or open context (i.e. zil
1990 * writes) and as a result is mutually exclusive with dedup.
1991 */
1992static int
1993zio_nop_write(zio_t *zio)
1994{
1995 blkptr_t *bp = zio->io_bp;
1996 blkptr_t *bp_orig = &zio->io_bp_orig;
1997 zio_prop_t *zp = &zio->io_prop;
1998
1999 ASSERT(BP_GET_LEVEL(bp) == 0);
2000 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2001 ASSERT(zp->zp_nopwrite);
2002 ASSERT(!zp->zp_dedup);
2003 ASSERT(zio->io_bp_override == NULL);
2004 ASSERT(IO_IS_ALLOCATING(zio));
2005
2006 /*
2007 * Check to see if the original bp and the new bp have matching
2008 * characteristics (i.e. same checksum, compression algorithms, etc).
2009 * If they don't then just continue with the pipeline which will
2010 * allocate a new bp.
2011 */
2012 if (BP_IS_HOLE(bp_orig) ||
2013 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
2014 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2015 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2016 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2017 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2018 return (ZIO_PIPELINE_CONTINUE);
2019
2020 /*
2021 * If the checksums match then reset the pipeline so that we
2022 * avoid allocating a new bp and issuing any I/O.
2023 */
2024 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2025 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
2026 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2027 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2028 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2029 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2030 sizeof (uint64_t)) == 0);
2031
2032 *bp = *bp_orig;
2033 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2034 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2035 }
2036
2037 return (ZIO_PIPELINE_CONTINUE);
2038}
2039
2040/*
2041 * ==========================================================================
2042 * Dedup
2043 * ==========================================================================
2044 */
2045static void
2046zio_ddt_child_read_done(zio_t *zio)
2047{
2048 blkptr_t *bp = zio->io_bp;
2049 ddt_entry_t *dde = zio->io_private;
2050 ddt_phys_t *ddp;
2051 zio_t *pio = zio_unique_parent(zio);
2052
2053 mutex_enter(&pio->io_lock);
2054 ddp = ddt_phys_select(dde, bp);
2055 if (zio->io_error == 0)
2056 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2057 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
2058 dde->dde_repair_data = zio->io_data;
2059 else
2060 zio_buf_free(zio->io_data, zio->io_size);
2061 mutex_exit(&pio->io_lock);
2062}
2063
2064static int
2065zio_ddt_read_start(zio_t *zio)
2066{
2067 blkptr_t *bp = zio->io_bp;
2068
2069 ASSERT(BP_GET_DEDUP(bp));
2070 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2071 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2072
2073 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2074 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2075 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2076 ddt_phys_t *ddp = dde->dde_phys;
2077 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2078 blkptr_t blk;
2079
2080 ASSERT(zio->io_vsd == NULL);
2081 zio->io_vsd = dde;
2082
2083 if (ddp_self == NULL)
2084 return (ZIO_PIPELINE_CONTINUE);
2085
2086 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2087 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2088 continue;
2089 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2090 &blk);
2091 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2092 zio_buf_alloc(zio->io_size), zio->io_size,
2093 zio_ddt_child_read_done, dde, zio->io_priority,
2094 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2095 &zio->io_bookmark));
2096 }
2097 return (ZIO_PIPELINE_CONTINUE);
2098 }
2099
2100 zio_nowait(zio_read(zio, zio->io_spa, bp,
2101 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2102 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2103
2104 return (ZIO_PIPELINE_CONTINUE);
2105}
2106
2107static int
2108zio_ddt_read_done(zio_t *zio)
2109{
2110 blkptr_t *bp = zio->io_bp;
2111
2112 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2113 return (ZIO_PIPELINE_STOP);
2114
2115 ASSERT(BP_GET_DEDUP(bp));
2116 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2117 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2118
2119 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2120 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2121 ddt_entry_t *dde = zio->io_vsd;
2122 if (ddt == NULL) {
2123 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2124 return (ZIO_PIPELINE_CONTINUE);
2125 }
2126 if (dde == NULL) {
2127 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2128 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2129 return (ZIO_PIPELINE_STOP);
2130 }
2131 if (dde->dde_repair_data != NULL) {
2132 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2133 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2134 }
2135 ddt_repair_done(ddt, dde);
2136 zio->io_vsd = NULL;
2137 }
2138
2139 ASSERT(zio->io_vsd == NULL);
2140
2141 return (ZIO_PIPELINE_CONTINUE);
2142}
2143
2144static boolean_t
2145zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2146{
2147 spa_t *spa = zio->io_spa;
2148
2149 /*
2150 * Note: we compare the original data, not the transformed data,
2151 * because when zio->io_bp is an override bp, we will not have
2152 * pushed the I/O transforms. That's an important optimization
2153 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2154 */
2155 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2156 zio_t *lio = dde->dde_lead_zio[p];
2157
2158 if (lio != NULL) {
2159 return (lio->io_orig_size != zio->io_orig_size ||
2160 bcmp(zio->io_orig_data, lio->io_orig_data,
2161 zio->io_orig_size) != 0);
2162 }
2163 }
2164
2165 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2166 ddt_phys_t *ddp = &dde->dde_phys[p];
2167
2168 if (ddp->ddp_phys_birth != 0) {
2169 arc_buf_t *abuf = NULL;
2170 uint32_t aflags = ARC_WAIT;
2171 blkptr_t blk = *zio->io_bp;
2172 int error;
2173
2174 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2175
2176 ddt_exit(ddt);
2177
2178 error = arc_read(NULL, spa, &blk,
2179 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2180 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2181 &aflags, &zio->io_bookmark);
2182
2183 if (error == 0) {
2184 if (arc_buf_size(abuf) != zio->io_orig_size ||
2185 bcmp(abuf->b_data, zio->io_orig_data,
2186 zio->io_orig_size) != 0)
2187 error = SET_ERROR(EEXIST);
2188 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2189 }
2190
2191 ddt_enter(ddt);
2192 return (error != 0);
2193 }
2194 }
2195
2196 return (B_FALSE);
2197}
2198
2199static void
2200zio_ddt_child_write_ready(zio_t *zio)
2201{
2202 int p = zio->io_prop.zp_copies;
2203 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2204 ddt_entry_t *dde = zio->io_private;
2205 ddt_phys_t *ddp = &dde->dde_phys[p];
2206 zio_t *pio;
2207
2208 if (zio->io_error)
2209 return;
2210
2211 ddt_enter(ddt);
2212
2213 ASSERT(dde->dde_lead_zio[p] == zio);
2214
2215 ddt_phys_fill(ddp, zio->io_bp);
2216
2217 while ((pio = zio_walk_parents(zio)) != NULL)
2218 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2219
2220 ddt_exit(ddt);
2221}
2222
2223static void
2224zio_ddt_child_write_done(zio_t *zio)
2225{
2226 int p = zio->io_prop.zp_copies;
2227 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2228 ddt_entry_t *dde = zio->io_private;
2229 ddt_phys_t *ddp = &dde->dde_phys[p];
2230
2231 ddt_enter(ddt);
2232
2233 ASSERT(ddp->ddp_refcnt == 0);
2234 ASSERT(dde->dde_lead_zio[p] == zio);
2235 dde->dde_lead_zio[p] = NULL;
2236
2237 if (zio->io_error == 0) {
2238 while (zio_walk_parents(zio) != NULL)
2239 ddt_phys_addref(ddp);
2240 } else {
2241 ddt_phys_clear(ddp);
2242 }
2243
2244 ddt_exit(ddt);
2245}
2246
2247static void
2248zio_ddt_ditto_write_done(zio_t *zio)
2249{
2250 int p = DDT_PHYS_DITTO;
2251 zio_prop_t *zp = &zio->io_prop;
2252 blkptr_t *bp = zio->io_bp;
2253 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2254 ddt_entry_t *dde = zio->io_private;
2255 ddt_phys_t *ddp = &dde->dde_phys[p];
2256 ddt_key_t *ddk = &dde->dde_key;
2257
2258 ddt_enter(ddt);
2259
2260 ASSERT(ddp->ddp_refcnt == 0);
2261 ASSERT(dde->dde_lead_zio[p] == zio);
2262 dde->dde_lead_zio[p] = NULL;
2263
2264 if (zio->io_error == 0) {
2265 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2266 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2267 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2268 if (ddp->ddp_phys_birth != 0)
2269 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2270 ddt_phys_fill(ddp, bp);
2271 }
2272
2273 ddt_exit(ddt);
2274}
2275
2276static int
2277zio_ddt_write(zio_t *zio)
2278{
2279 spa_t *spa = zio->io_spa;
2280 blkptr_t *bp = zio->io_bp;
2281 uint64_t txg = zio->io_txg;
2282 zio_prop_t *zp = &zio->io_prop;
2283 int p = zp->zp_copies;
2284 int ditto_copies;
2285 zio_t *cio = NULL;
2286 zio_t *dio = NULL;
2287 ddt_t *ddt = ddt_select(spa, bp);
2288 ddt_entry_t *dde;
2289 ddt_phys_t *ddp;
2290
2291 ASSERT(BP_GET_DEDUP(bp));
2292 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2293 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2294
2295 ddt_enter(ddt);
2296 dde = ddt_lookup(ddt, bp, B_TRUE);
2297 ddp = &dde->dde_phys[p];
2298
2299 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2300 /*
2301 * If we're using a weak checksum, upgrade to a strong checksum
2302 * and try again. If we're already using a strong checksum,
2303 * we can't resolve it, so just convert to an ordinary write.
2304 * (And automatically e-mail a paper to Nature?)
2305 */
2306 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2307 zp->zp_checksum = spa_dedup_checksum(spa);
2308 zio_pop_transforms(zio);
2309 zio->io_stage = ZIO_STAGE_OPEN;
2310 BP_ZERO(bp);
2311 } else {
2312 zp->zp_dedup = B_FALSE;
2313 }
2314 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2315 ddt_exit(ddt);
2316 return (ZIO_PIPELINE_CONTINUE);
2317 }
2318
2319 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2320 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2321
2322 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2323 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2324 zio_prop_t czp = *zp;
2325
2326 czp.zp_copies = ditto_copies;
2327
2328 /*
2329 * If we arrived here with an override bp, we won't have run
2330 * the transform stack, so we won't have the data we need to
2331 * generate a child i/o. So, toss the override bp and restart.
2332 * This is safe, because using the override bp is just an
2333 * optimization; and it's rare, so the cost doesn't matter.
2334 */
2335 if (zio->io_bp_override) {
2336 zio_pop_transforms(zio);
2337 zio->io_stage = ZIO_STAGE_OPEN;
2338 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2339 zio->io_bp_override = NULL;
2340 BP_ZERO(bp);
2341 ddt_exit(ddt);
2342 return (ZIO_PIPELINE_CONTINUE);
2343 }
2344
2345 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2346 zio->io_orig_size, &czp, NULL, NULL,
2347 zio_ddt_ditto_write_done, dde, zio->io_priority,
2348 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2349
2350 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2351 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2352 }
2353
2354 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2355 if (ddp->ddp_phys_birth != 0)
2356 ddt_bp_fill(ddp, bp, txg);
2357 if (dde->dde_lead_zio[p] != NULL)
2358 zio_add_child(zio, dde->dde_lead_zio[p]);
2359 else
2360 ddt_phys_addref(ddp);
2361 } else if (zio->io_bp_override) {
2362 ASSERT(bp->blk_birth == txg);
2363 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2364 ddt_phys_fill(ddp, bp);
2365 ddt_phys_addref(ddp);
2366 } else {
2367 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2368 zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL,
2369 zio_ddt_child_write_done, dde, zio->io_priority,
2370 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2371
2372 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2373 dde->dde_lead_zio[p] = cio;
2374 }
2375
2376 ddt_exit(ddt);
2377
2378 if (cio)
2379 zio_nowait(cio);
2380 if (dio)
2381 zio_nowait(dio);
2382
2383 return (ZIO_PIPELINE_CONTINUE);
2384}
2385
2386ddt_entry_t *freedde; /* for debugging */
2387
2388static int
2389zio_ddt_free(zio_t *zio)
2390{
2391 spa_t *spa = zio->io_spa;
2392 blkptr_t *bp = zio->io_bp;
2393 ddt_t *ddt = ddt_select(spa, bp);
2394 ddt_entry_t *dde;
2395 ddt_phys_t *ddp;
2396
2397 ASSERT(BP_GET_DEDUP(bp));
2398 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2399
2400 ddt_enter(ddt);
2401 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2402 ddp = ddt_phys_select(dde, bp);
2403 ddt_phys_decref(ddp);
2404 ddt_exit(ddt);
2405
2406 return (ZIO_PIPELINE_CONTINUE);
2407}
2408
2409/*
2410 * ==========================================================================
2411 * Allocate and free blocks
2412 * ==========================================================================
2413 */
2414static int
2415zio_dva_allocate(zio_t *zio)
2416{
2417 spa_t *spa = zio->io_spa;
2418 metaslab_class_t *mc = spa_normal_class(spa);
2419 blkptr_t *bp = zio->io_bp;
2420 int error;
2421 int flags = 0;
2422
2423 if (zio->io_gang_leader == NULL) {
2424 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2425 zio->io_gang_leader = zio;
2426 }
2427
2428 ASSERT(BP_IS_HOLE(bp));
2429 ASSERT0(BP_GET_NDVAS(bp));
2430 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2431 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2432 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2433
2434 /*
2435 * The dump device does not support gang blocks so allocation on
2436 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2437 * the "fast" gang feature.
2438 */
2439 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2440 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2441 METASLAB_GANG_CHILD : 0;
2442 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2443 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2444
2445 if (error) {
2446 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2447 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2448 error);
2449 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2450 return (zio_write_gang_block(zio));
2451 zio->io_error = error;
2452 }
2453
2454 return (ZIO_PIPELINE_CONTINUE);
2455}
2456
2457static int
2458zio_dva_free(zio_t *zio)
2459{
2460 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2461
2462 return (ZIO_PIPELINE_CONTINUE);
2463}
2464
2465static int
2466zio_dva_claim(zio_t *zio)
2467{
2468 int error;
2469
2470 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2471 if (error)
2472 zio->io_error = error;
2473
2474 return (ZIO_PIPELINE_CONTINUE);
2475}
2476
2477/*
2478 * Undo an allocation. This is used by zio_done() when an I/O fails
2479 * and we want to give back the block we just allocated.
2480 * This handles both normal blocks and gang blocks.
2481 */
2482static void
2483zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2484{
2485 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2486 ASSERT(zio->io_bp_override == NULL);
2487
2488 if (!BP_IS_HOLE(bp))
2489 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2490
2491 if (gn != NULL) {
2492 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2493 zio_dva_unallocate(zio, gn->gn_child[g],
2494 &gn->gn_gbh->zg_blkptr[g]);
2495 }
2496 }
2497}
2498
2499/*
2500 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2501 */
2502int
2503zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2504 uint64_t size, boolean_t use_slog)
2505{
2506 int error = 1;
2507
2508 ASSERT(txg > spa_syncing_txg(spa));
2509
2510 /*
2511 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2512 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2513 * when allocating them.
2514 */
2515 if (use_slog) {
2516 error = metaslab_alloc(spa, spa_log_class(spa), size,
2517 new_bp, 1, txg, old_bp,
2518 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2519 }
2520
2521 if (error) {
2522 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2523 new_bp, 1, txg, old_bp,
2524 METASLAB_HINTBP_AVOID);
2525 }
2526
2527 if (error == 0) {
2528 BP_SET_LSIZE(new_bp, size);
2529 BP_SET_PSIZE(new_bp, size);
2530 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2531 BP_SET_CHECKSUM(new_bp,
2532 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2533 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2534 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2535 BP_SET_LEVEL(new_bp, 0);
2536 BP_SET_DEDUP(new_bp, 0);
2537 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2538 }
2539
2540 return (error);
2541}
2542
2543/*
2544 * Free an intent log block.
2545 */
2546void
2547zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2548{
2549 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2550 ASSERT(!BP_IS_GANG(bp));
2551
2552 zio_free(spa, txg, bp);
2553}
2554
2555/*
2556 * ==========================================================================
2557 * Read, write and delete to physical devices
2558 * ==========================================================================
2559 */
2560static int
2561zio_vdev_io_start(zio_t *zio)
2562{
2563 vdev_t *vd = zio->io_vd;
2564 uint64_t align;
2565 spa_t *spa = zio->io_spa;
2566 int ret;
2567
2568 ASSERT(zio->io_error == 0);
2569 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2570
2571 if (vd == NULL) {
2572 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2573 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2574
2575 /*
2576 * The mirror_ops handle multiple DVAs in a single BP.
2577 */
2578 return (vdev_mirror_ops.vdev_op_io_start(zio));
2579 }
2580
2581 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE &&
2582 zio->io_priority == ZIO_PRIORITY_NOW) {
2583 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
2584 return (ZIO_PIPELINE_CONTINUE);
2585 }
2586
2587 /*
2588 * We keep track of time-sensitive I/Os so that the scan thread
2589 * can quickly react to certain workloads. In particular, we care
2590 * about non-scrubbing, top-level reads and writes with the following
2591 * characteristics:
2592 * - synchronous writes of user data to non-slog devices
2593 * - any reads of user data
2594 * When these conditions are met, adjust the timestamp of spa_last_io
2595 * which allows the scan thread to adjust its workload accordingly.
2596 */
2597 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2598 vd == vd->vdev_top && !vd->vdev_islog &&
2599 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2600 zio->io_txg != spa_syncing_txg(spa)) {
2601 uint64_t old = spa->spa_last_io;
2602 uint64_t new = ddi_get_lbolt64();
2603 if (old != new)
2604 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2605 }
2606
2607 align = 1ULL << vd->vdev_top->vdev_ashift;
2608
2609 if ((!(zio->io_flags & ZIO_FLAG_PHYSICAL) ||
2610 (vd->vdev_top->vdev_physical_ashift > SPA_MINBLOCKSHIFT)) &&
2611 P2PHASE(zio->io_size, align) != 0) {
2612 /* Transform logical writes to be a full physical block size. */
2613 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2614 char *abuf = NULL;
2615 if (zio->io_type == ZIO_TYPE_READ ||
2616 zio->io_type == ZIO_TYPE_WRITE)
2617 abuf = zio_buf_alloc(asize);
2618 ASSERT(vd == vd->vdev_top);
2619 if (zio->io_type == ZIO_TYPE_WRITE) {
2620 bcopy(zio->io_data, abuf, zio->io_size);
2621 bzero(abuf + zio->io_size, asize - zio->io_size);
2622 }
2623 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
2624 zio_subblock);
2625 }
2626
2627 /*
2628 * If this is not a physical io, make sure that it is properly aligned
2629 * before proceeding.
2630 */
2631 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
2632 ASSERT0(P2PHASE(zio->io_offset, align));
2633 ASSERT0(P2PHASE(zio->io_size, align));
2634 } else {
2635 /*
2636 * For physical writes, we allow 512b aligned writes and assume
2637 * the device will perform a read-modify-write as necessary.
2638 */
2639 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
2640 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
2641 }
2642
2643 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
2644
2645 /*
2646 * If this is a repair I/O, and there's no self-healing involved --
2647 * that is, we're just resilvering what we expect to resilver --
2648 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2649 * This prevents spurious resilvering with nested replication.
2650 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2651 * A is out of date, we'll read from C+D, then use the data to
2652 * resilver A+B -- but we don't actually want to resilver B, just A.
2653 * The top-level mirror has no way to know this, so instead we just
2654 * discard unnecessary repairs as we work our way down the vdev tree.
2655 * The same logic applies to any form of nested replication:
2656 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2657 */
2658 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2659 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2660 zio->io_txg != 0 && /* not a delegated i/o */
2661 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2662 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2663 zio_vdev_io_bypass(zio);
2664 return (ZIO_PIPELINE_CONTINUE);
2665 }
2666
2667 if (vd->vdev_ops->vdev_op_leaf) {
2668 switch (zio->io_type) {
2669 case ZIO_TYPE_READ:
2670 if (vdev_cache_read(zio))
2671 return (ZIO_PIPELINE_CONTINUE);
2672 /* FALLTHROUGH */
2673 case ZIO_TYPE_WRITE:
2674 case ZIO_TYPE_FREE:
2675 if ((zio = vdev_queue_io(zio)) == NULL)
2676 return (ZIO_PIPELINE_STOP);
2677
2678 if (!vdev_accessible(vd, zio)) {
2679 zio->io_error = SET_ERROR(ENXIO);
2680 zio_interrupt(zio);
2681 return (ZIO_PIPELINE_STOP);
2682 }
2683 break;
2684 }
2685 /*
2686 * Note that we ignore repair writes for TRIM because they can
2687 * conflict with normal writes. This isn't an issue because, by
2688 * definition, we only repair blocks that aren't freed.
2689 */
2690 if (zio->io_type == ZIO_TYPE_WRITE &&
2691 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2692 !trim_map_write_start(zio))
2693 return (ZIO_PIPELINE_STOP);
2694 }
2695
2696 ret = vd->vdev_ops->vdev_op_io_start(zio);
2697 ASSERT(ret == ZIO_PIPELINE_STOP);
2698
2699 return (ret);
2700}
2701
2702static int
2703zio_vdev_io_done(zio_t *zio)
2704{
2705 vdev_t *vd = zio->io_vd;
2706 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2707 boolean_t unexpected_error = B_FALSE;
2708
2709 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2710 return (ZIO_PIPELINE_STOP);
2711
2712 ASSERT(zio->io_type == ZIO_TYPE_READ ||
2713 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
2714
2715 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2716 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE ||
2717 zio->io_type == ZIO_TYPE_FREE)) {
2718
2719 if (zio->io_type == ZIO_TYPE_WRITE &&
2720 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
2721 trim_map_write_done(zio);
2722
2723 vdev_queue_io_done(zio);
2724
2725 if (zio->io_type == ZIO_TYPE_WRITE)
2726 vdev_cache_write(zio);
2727
2728 if (zio_injection_enabled && zio->io_error == 0)
2729 zio->io_error = zio_handle_device_injection(vd,
2730 zio, EIO);
2731
2732 if (zio_injection_enabled && zio->io_error == 0)
2733 zio->io_error = zio_handle_label_injection(zio, EIO);
2734
2735 if (zio->io_error) {
2736 if (zio->io_error == ENOTSUP &&
2737 zio->io_type == ZIO_TYPE_FREE) {
2738 /* Not all devices support TRIM. */
2739 } else if (!vdev_accessible(vd, zio)) {
2740 zio->io_error = SET_ERROR(ENXIO);
2741 } else {
2742 unexpected_error = B_TRUE;
2743 }
2744 }
2745 }
2746
2747 ops->vdev_op_io_done(zio);
2748
2749 if (unexpected_error)
2750 VERIFY(vdev_probe(vd, zio) == NULL);
2751
2752 return (ZIO_PIPELINE_CONTINUE);
2753}
2754
2755/*
2756 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2757 * disk, and use that to finish the checksum ereport later.
2758 */
2759static void
2760zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2761 const void *good_buf)
2762{
2763 /* no processing needed */
2764 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2765}
2766
2767/*ARGSUSED*/
2768void
2769zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2770{
2771 void *buf = zio_buf_alloc(zio->io_size);
2772
2773 bcopy(zio->io_data, buf, zio->io_size);
2774
2775 zcr->zcr_cbinfo = zio->io_size;
2776 zcr->zcr_cbdata = buf;
2777 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2778 zcr->zcr_free = zio_buf_free;
2779}
2780
2781static int
2782zio_vdev_io_assess(zio_t *zio)
2783{
2784 vdev_t *vd = zio->io_vd;
2785
2786 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2787 return (ZIO_PIPELINE_STOP);
2788
2789 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2790 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2791
2792 if (zio->io_vsd != NULL) {
2793 zio->io_vsd_ops->vsd_free(zio);
2794 zio->io_vsd = NULL;
2795 }
2796
2797 if (zio_injection_enabled && zio->io_error == 0)
2798 zio->io_error = zio_handle_fault_injection(zio, EIO);
2799
2800 if (zio->io_type == ZIO_TYPE_FREE &&
2801 zio->io_priority != ZIO_PRIORITY_NOW) {
2802 switch (zio->io_error) {
2803 case 0:
2804 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
2805 ZIO_TRIM_STAT_BUMP(success);
2806 break;
2807 case EOPNOTSUPP:
2808 ZIO_TRIM_STAT_BUMP(unsupported);
2809 break;
2810 default:
2811 ZIO_TRIM_STAT_BUMP(failed);
2812 break;
2813 }
2814 }
2815
2816 /*
2817 * If the I/O failed, determine whether we should attempt to retry it.
2818 *
2819 * On retry, we cut in line in the issue queue, since we don't want
2820 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2821 */
2822 if (zio->io_error && vd == NULL &&
2823 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2824 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2825 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2826 zio->io_error = 0;
2827 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2828 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2829 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2830 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2831 zio_requeue_io_start_cut_in_line);
2832 return (ZIO_PIPELINE_STOP);
2833 }
2834
2835 /*
2836 * If we got an error on a leaf device, convert it to ENXIO
2837 * if the device is not accessible at all.
2838 */
2839 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2840 !vdev_accessible(vd, zio))
2841 zio->io_error = SET_ERROR(ENXIO);
2842
2843 /*
2844 * If we can't write to an interior vdev (mirror or RAID-Z),
2845 * set vdev_cant_write so that we stop trying to allocate from it.
2846 */
2847 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2848 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2849 vd->vdev_cant_write = B_TRUE;
2850 }
2851
2852 if (zio->io_error)
2853 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2854
2855 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2856 zio->io_physdone != NULL) {
2857 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
2858 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
2859 zio->io_physdone(zio->io_logical);
2860 }
2861
2862 return (ZIO_PIPELINE_CONTINUE);
2863}
2864
2865void
2866zio_vdev_io_reissue(zio_t *zio)
2867{
2868 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2869 ASSERT(zio->io_error == 0);
2870
2871 zio->io_stage >>= 1;
2872}
2873
2874void
2875zio_vdev_io_redone(zio_t *zio)
2876{
2877 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2878
2879 zio->io_stage >>= 1;
2880}
2881
2882void
2883zio_vdev_io_bypass(zio_t *zio)
2884{
2885 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2886 ASSERT(zio->io_error == 0);
2887
2888 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2889 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2890}
2891
2892/*
2893 * ==========================================================================
2894 * Generate and verify checksums
2895 * ==========================================================================
2896 */
2897static int
2898zio_checksum_generate(zio_t *zio)
2899{
2900 blkptr_t *bp = zio->io_bp;
2901 enum zio_checksum checksum;
2902
2903 if (bp == NULL) {
2904 /*
2905 * This is zio_write_phys().
2906 * We're either generating a label checksum, or none at all.
2907 */
2908 checksum = zio->io_prop.zp_checksum;
2909
2910 if (checksum == ZIO_CHECKSUM_OFF)
2911 return (ZIO_PIPELINE_CONTINUE);
2912
2913 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2914 } else {
2915 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2916 ASSERT(!IO_IS_ALLOCATING(zio));
2917 checksum = ZIO_CHECKSUM_GANG_HEADER;
2918 } else {
2919 checksum = BP_GET_CHECKSUM(bp);
2920 }
2921 }
2922
2923 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2924
2925 return (ZIO_PIPELINE_CONTINUE);
2926}
2927
2928static int
2929zio_checksum_verify(zio_t *zio)
2930{
2931 zio_bad_cksum_t info;
2932 blkptr_t *bp = zio->io_bp;
2933 int error;
2934
2935 ASSERT(zio->io_vd != NULL);
2936
2937 if (bp == NULL) {
2938 /*
2939 * This is zio_read_phys().
2940 * We're either verifying a label checksum, or nothing at all.
2941 */
2942 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2943 return (ZIO_PIPELINE_CONTINUE);
2944
2945 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2946 }
2947
2948 if ((error = zio_checksum_error(zio, &info)) != 0) {
2949 zio->io_error = error;
2950 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2951 zfs_ereport_start_checksum(zio->io_spa,
2952 zio->io_vd, zio, zio->io_offset,
2953 zio->io_size, NULL, &info);
2954 }
2955 }
2956
2957 return (ZIO_PIPELINE_CONTINUE);
2958}
2959
2960/*
2961 * Called by RAID-Z to ensure we don't compute the checksum twice.
2962 */
2963void
2964zio_checksum_verified(zio_t *zio)
2965{
2966 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2967}
2968
2969/*
2970 * ==========================================================================
2971 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2972 * An error of 0 indicates success. ENXIO indicates whole-device failure,
2973 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2974 * indicate errors that are specific to one I/O, and most likely permanent.
2975 * Any other error is presumed to be worse because we weren't expecting it.
2976 * ==========================================================================
2977 */
2978int
2979zio_worst_error(int e1, int e2)
2980{
2981 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2982 int r1, r2;
2983
2984 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2985 if (e1 == zio_error_rank[r1])
2986 break;
2987
2988 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2989 if (e2 == zio_error_rank[r2])
2990 break;
2991
2992 return (r1 > r2 ? e1 : e2);
2993}
2994
2995/*
2996 * ==========================================================================
2997 * I/O completion
2998 * ==========================================================================
2999 */
3000static int
3001zio_ready(zio_t *zio)
3002{
3003 blkptr_t *bp = zio->io_bp;
3004 zio_t *pio, *pio_next;
3005
3006 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
3007 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
3008 return (ZIO_PIPELINE_STOP);
3009
3010 if (zio->io_ready) {
3011 ASSERT(IO_IS_ALLOCATING(zio));
3012 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3013 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3014 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3015
3016 zio->io_ready(zio);
3017 }
3018
3019 if (bp != NULL && bp != &zio->io_bp_copy)
3020 zio->io_bp_copy = *bp;
3021
3022 if (zio->io_error)
3023 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3024
3025 mutex_enter(&zio->io_lock);
3026 zio->io_state[ZIO_WAIT_READY] = 1;
3027 pio = zio_walk_parents(zio);
3028 mutex_exit(&zio->io_lock);
3029
3030 /*
3031 * As we notify zio's parents, new parents could be added.
3032 * New parents go to the head of zio's io_parent_list, however,
3033 * so we will (correctly) not notify them. The remainder of zio's
3034 * io_parent_list, from 'pio_next' onward, cannot change because
3035 * all parents must wait for us to be done before they can be done.
3036 */
3037 for (; pio != NULL; pio = pio_next) {
3038 pio_next = zio_walk_parents(zio);
3039 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3040 }
3041
3042 if (zio->io_flags & ZIO_FLAG_NODATA) {
3043 if (BP_IS_GANG(bp)) {
3044 zio->io_flags &= ~ZIO_FLAG_NODATA;
3045 } else {
3046 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
3047 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3048 }
3049 }
3050
3051 if (zio_injection_enabled &&
3052 zio->io_spa->spa_syncing_txg == zio->io_txg)
3053 zio_handle_ignored_writes(zio);
3054
3055 return (ZIO_PIPELINE_CONTINUE);
3056}
3057
3058static int
3059zio_done(zio_t *zio)
3060{
3061 spa_t *spa = zio->io_spa;
3062 zio_t *lio = zio->io_logical;
3063 blkptr_t *bp = zio->io_bp;
3064 vdev_t *vd = zio->io_vd;
3065 uint64_t psize = zio->io_size;
3066 zio_t *pio, *pio_next;
3067
3068 /*
3069 * If our children haven't all completed,
3070 * wait for them and then repeat this pipeline stage.
3071 */
3072 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
3073 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
3074 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
3075 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
3076 return (ZIO_PIPELINE_STOP);
3077
3078 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3079 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3080 ASSERT(zio->io_children[c][w] == 0);
3081
3082 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3083 ASSERT(bp->blk_pad[0] == 0);
3084 ASSERT(bp->blk_pad[1] == 0);
3085 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3086 (bp == zio_unique_parent(zio)->io_bp));
3087 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3088 zio->io_bp_override == NULL &&
3089 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3090 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3091 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3092 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3093 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3094 }
3095 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3096 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3097 }
3098
3099 /*
3100 * If there were child vdev/gang/ddt errors, they apply to us now.
3101 */
3102 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3103 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3104 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3105
3106 /*
3107 * If the I/O on the transformed data was successful, generate any
3108 * checksum reports now while we still have the transformed data.
3109 */
3110 if (zio->io_error == 0) {
3111 while (zio->io_cksum_report != NULL) {
3112 zio_cksum_report_t *zcr = zio->io_cksum_report;
3113 uint64_t align = zcr->zcr_align;
3114 uint64_t asize = P2ROUNDUP(psize, align);
3115 char *abuf = zio->io_data;
3116
3117 if (asize != psize) {
3118 abuf = zio_buf_alloc(asize);
3119 bcopy(zio->io_data, abuf, psize);
3120 bzero(abuf + psize, asize - psize);
3121 }
3122
3123 zio->io_cksum_report = zcr->zcr_next;
3124 zcr->zcr_next = NULL;
3125 zcr->zcr_finish(zcr, abuf);
3126 zfs_ereport_free_checksum(zcr);
3127
3128 if (asize != psize)
3129 zio_buf_free(abuf, asize);
3130 }
3131 }
3132
3133 zio_pop_transforms(zio); /* note: may set zio->io_error */
3134
3135 vdev_stat_update(zio, psize);
3136
3137 if (zio->io_error) {
3138 /*
3139 * If this I/O is attached to a particular vdev,
3140 * generate an error message describing the I/O failure
3141 * at the block level. We ignore these errors if the
3142 * device is currently unavailable.
3143 */
3144 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3145 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3146
3147 if ((zio->io_error == EIO || !(zio->io_flags &
3148 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3149 zio == lio) {
3150 /*
3151 * For logical I/O requests, tell the SPA to log the
3152 * error and generate a logical data ereport.
3153 */
3154 spa_log_error(spa, zio);
3155 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3156 0, 0);
3157 }
3158 }
3159
3160 if (zio->io_error && zio == lio) {
3161 /*
3162 * Determine whether zio should be reexecuted. This will
3163 * propagate all the way to the root via zio_notify_parent().
3164 */
3165 ASSERT(vd == NULL && bp != NULL);
3166 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3167
3168 if (IO_IS_ALLOCATING(zio) &&
3169 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3170 if (zio->io_error != ENOSPC)
3171 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3172 else
3173 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3174 }
3175
3176 if ((zio->io_type == ZIO_TYPE_READ ||
3177 zio->io_type == ZIO_TYPE_FREE) &&
3178 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3179 zio->io_error == ENXIO &&
3180 spa_load_state(spa) == SPA_LOAD_NONE &&
3181 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3182 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3183
3184 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3185 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3186
3187 /*
3188 * Here is a possibly good place to attempt to do
3189 * either combinatorial reconstruction or error correction
3190 * based on checksums. It also might be a good place
3191 * to send out preliminary ereports before we suspend
3192 * processing.
3193 */
3194 }
3195
3196 /*
3197 * If there were logical child errors, they apply to us now.
3198 * We defer this until now to avoid conflating logical child
3199 * errors with errors that happened to the zio itself when
3200 * updating vdev stats and reporting FMA events above.
3201 */
3202 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3203
3204 if ((zio->io_error || zio->io_reexecute) &&
3205 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3206 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3207 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3208
3209 zio_gang_tree_free(&zio->io_gang_tree);
3210
3211 /*
3212 * Godfather I/Os should never suspend.
3213 */
3214 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3215 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3216 zio->io_reexecute = 0;
3217
3218 if (zio->io_reexecute) {
3219 /*
3220 * This is a logical I/O that wants to reexecute.
3221 *
3222 * Reexecute is top-down. When an i/o fails, if it's not
3223 * the root, it simply notifies its parent and sticks around.
3224 * The parent, seeing that it still has children in zio_done(),
3225 * does the same. This percolates all the way up to the root.
3226 * The root i/o will reexecute or suspend the entire tree.
3227 *
3228 * This approach ensures that zio_reexecute() honors
3229 * all the original i/o dependency relationships, e.g.
3230 * parents not executing until children are ready.
3231 */
3232 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3233
3234 zio->io_gang_leader = NULL;
3235
3236 mutex_enter(&zio->io_lock);
3237 zio->io_state[ZIO_WAIT_DONE] = 1;
3238 mutex_exit(&zio->io_lock);
3239
3240 /*
3241 * "The Godfather" I/O monitors its children but is
3242 * not a true parent to them. It will track them through
3243 * the pipeline but severs its ties whenever they get into
3244 * trouble (e.g. suspended). This allows "The Godfather"
3245 * I/O to return status without blocking.
3246 */
3247 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3248 zio_link_t *zl = zio->io_walk_link;
3249 pio_next = zio_walk_parents(zio);
3250
3251 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3252 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3253 zio_remove_child(pio, zio, zl);
3254 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3255 }
3256 }
3257
3258 if ((pio = zio_unique_parent(zio)) != NULL) {
3259 /*
3260 * We're not a root i/o, so there's nothing to do
3261 * but notify our parent. Don't propagate errors
3262 * upward since we haven't permanently failed yet.
3263 */
3264 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3265 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3266 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3267 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3268 /*
3269 * We'd fail again if we reexecuted now, so suspend
3270 * until conditions improve (e.g. device comes online).
3271 */
3272 zio_suspend(spa, zio);
3273 } else {
3274 /*
3275 * Reexecution is potentially a huge amount of work.
3276 * Hand it off to the otherwise-unused claim taskq.
3277 */
3278#if defined(illumos) || !defined(_KERNEL)
3279 ASSERT(zio->io_tqent.tqent_next == NULL);
3280#else
3281 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
3282#endif
3283 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3284 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3285 0, &zio->io_tqent);
3286 }
3287 return (ZIO_PIPELINE_STOP);
3288 }
3289
3290 ASSERT(zio->io_child_count == 0);
3291 ASSERT(zio->io_reexecute == 0);
3292 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3293
3294 /*
3295 * Report any checksum errors, since the I/O is complete.
3296 */
3297 while (zio->io_cksum_report != NULL) {
3298 zio_cksum_report_t *zcr = zio->io_cksum_report;
3299 zio->io_cksum_report = zcr->zcr_next;
3300 zcr->zcr_next = NULL;
3301 zcr->zcr_finish(zcr, NULL);
3302 zfs_ereport_free_checksum(zcr);
3303 }
3304
3305 /*
3306 * It is the responsibility of the done callback to ensure that this
3307 * particular zio is no longer discoverable for adoption, and as
3308 * such, cannot acquire any new parents.
3309 */
3310 if (zio->io_done)
3311 zio->io_done(zio);
3312
3313 mutex_enter(&zio->io_lock);
3314 zio->io_state[ZIO_WAIT_DONE] = 1;
3315 mutex_exit(&zio->io_lock);
3316
3317 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3318 zio_link_t *zl = zio->io_walk_link;
3319 pio_next = zio_walk_parents(zio);
3320 zio_remove_child(pio, zio, zl);
3321 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3322 }
3323
3324 if (zio->io_waiter != NULL) {
3325 mutex_enter(&zio->io_lock);
3326 zio->io_executor = NULL;
3327 cv_broadcast(&zio->io_cv);
3328 mutex_exit(&zio->io_lock);
3329 } else {
3330 zio_destroy(zio);
3331 }
3332
3333 return (ZIO_PIPELINE_STOP);
3334}
3335
3336/*
3337 * ==========================================================================
3338 * I/O pipeline definition
3339 * ==========================================================================
3340 */
3341static zio_pipe_stage_t *zio_pipeline[] = {
3342 NULL,
3343 zio_read_bp_init,
3344 zio_free_bp_init,
3345 zio_issue_async,
3346 zio_write_bp_init,
3347 zio_checksum_generate,
3348 zio_nop_write,
3349 zio_ddt_read_start,
3350 zio_ddt_read_done,
3351 zio_ddt_write,
3352 zio_ddt_free,
3353 zio_gang_assemble,
3354 zio_gang_issue,
3355 zio_dva_allocate,
3356 zio_dva_free,
3357 zio_dva_claim,
3358 zio_ready,
3359 zio_vdev_io_start,
3360 zio_vdev_io_done,
3361 zio_vdev_io_assess,
3362 zio_checksum_verify,
3363 zio_done
3364};
3365
3366/* dnp is the dnode for zb1->zb_object */
3367boolean_t
3368zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_phys_t *zb1,
3369 const zbookmark_phys_t *zb2)
3370{
3371 uint64_t zb1nextL0, zb2thisobj;
3372
3373 ASSERT(zb1->zb_objset == zb2->zb_objset);
3374 ASSERT(zb2->zb_level == 0);
3375
3376 /* The objset_phys_t isn't before anything. */
3377 if (dnp == NULL)
3378 return (B_FALSE);
3379
3380 zb1nextL0 = (zb1->zb_blkid + 1) <<
3381 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3382
3383 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3384 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3385
3386 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3387 uint64_t nextobj = zb1nextL0 *
3388 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3389 return (nextobj <= zb2thisobj);
3390 }
3391
3392 if (zb1->zb_object < zb2thisobj)
3393 return (B_TRUE);
3394 if (zb1->zb_object > zb2thisobj)
3395 return (B_FALSE);
3396 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3397 return (B_FALSE);
3398 return (zb1nextL0 <= zb2->zb_blkid);
3399}
| 1461 }
1462
1463 zio_execute(zio);
1464}
1465
1466/*
1467 * ==========================================================================
1468 * Reexecute or suspend/resume failed I/O
1469 * ==========================================================================
1470 */
1471
1472static void
1473zio_reexecute(zio_t *pio)
1474{
1475 zio_t *cio, *cio_next;
1476
1477 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1478 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1479 ASSERT(pio->io_gang_leader == NULL);
1480 ASSERT(pio->io_gang_tree == NULL);
1481
1482 pio->io_flags = pio->io_orig_flags;
1483 pio->io_stage = pio->io_orig_stage;
1484 pio->io_pipeline = pio->io_orig_pipeline;
1485 pio->io_reexecute = 0;
1486 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1487 pio->io_error = 0;
1488 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1489 pio->io_state[w] = 0;
1490 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1491 pio->io_child_error[c] = 0;
1492
1493 if (IO_IS_ALLOCATING(pio))
1494 BP_ZERO(pio->io_bp);
1495
1496 /*
1497 * As we reexecute pio's children, new children could be created.
1498 * New children go to the head of pio's io_child_list, however,
1499 * so we will (correctly) not reexecute them. The key is that
1500 * the remainder of pio's io_child_list, from 'cio_next' onward,
1501 * cannot be affected by any side effects of reexecuting 'cio'.
1502 */
1503 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1504 cio_next = zio_walk_children(pio);
1505 mutex_enter(&pio->io_lock);
1506 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1507 pio->io_children[cio->io_child_type][w]++;
1508 mutex_exit(&pio->io_lock);
1509 zio_reexecute(cio);
1510 }
1511
1512 /*
1513 * Now that all children have been reexecuted, execute the parent.
1514 * We don't reexecute "The Godfather" I/O here as it's the
1515 * responsibility of the caller to wait on him.
1516 */
1517 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1518 zio_execute(pio);
1519}
1520
1521void
1522zio_suspend(spa_t *spa, zio_t *zio)
1523{
1524 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1525 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1526 "failure and the failure mode property for this pool "
1527 "is set to panic.", spa_name(spa));
1528
1529 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1530
1531 mutex_enter(&spa->spa_suspend_lock);
1532
1533 if (spa->spa_suspend_zio_root == NULL)
1534 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1535 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1536 ZIO_FLAG_GODFATHER);
1537
1538 spa->spa_suspended = B_TRUE;
1539
1540 if (zio != NULL) {
1541 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1542 ASSERT(zio != spa->spa_suspend_zio_root);
1543 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1544 ASSERT(zio_unique_parent(zio) == NULL);
1545 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1546 zio_add_child(spa->spa_suspend_zio_root, zio);
1547 }
1548
1549 mutex_exit(&spa->spa_suspend_lock);
1550}
1551
1552int
1553zio_resume(spa_t *spa)
1554{
1555 zio_t *pio;
1556
1557 /*
1558 * Reexecute all previously suspended i/o.
1559 */
1560 mutex_enter(&spa->spa_suspend_lock);
1561 spa->spa_suspended = B_FALSE;
1562 cv_broadcast(&spa->spa_suspend_cv);
1563 pio = spa->spa_suspend_zio_root;
1564 spa->spa_suspend_zio_root = NULL;
1565 mutex_exit(&spa->spa_suspend_lock);
1566
1567 if (pio == NULL)
1568 return (0);
1569
1570 zio_reexecute(pio);
1571 return (zio_wait(pio));
1572}
1573
1574void
1575zio_resume_wait(spa_t *spa)
1576{
1577 mutex_enter(&spa->spa_suspend_lock);
1578 while (spa_suspended(spa))
1579 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1580 mutex_exit(&spa->spa_suspend_lock);
1581}
1582
1583/*
1584 * ==========================================================================
1585 * Gang blocks.
1586 *
1587 * A gang block is a collection of small blocks that looks to the DMU
1588 * like one large block. When zio_dva_allocate() cannot find a block
1589 * of the requested size, due to either severe fragmentation or the pool
1590 * being nearly full, it calls zio_write_gang_block() to construct the
1591 * block from smaller fragments.
1592 *
1593 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1594 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1595 * an indirect block: it's an array of block pointers. It consumes
1596 * only one sector and hence is allocatable regardless of fragmentation.
1597 * The gang header's bps point to its gang members, which hold the data.
1598 *
1599 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1600 * as the verifier to ensure uniqueness of the SHA256 checksum.
1601 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1602 * not the gang header. This ensures that data block signatures (needed for
1603 * deduplication) are independent of how the block is physically stored.
1604 *
1605 * Gang blocks can be nested: a gang member may itself be a gang block.
1606 * Thus every gang block is a tree in which root and all interior nodes are
1607 * gang headers, and the leaves are normal blocks that contain user data.
1608 * The root of the gang tree is called the gang leader.
1609 *
1610 * To perform any operation (read, rewrite, free, claim) on a gang block,
1611 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1612 * in the io_gang_tree field of the original logical i/o by recursively
1613 * reading the gang leader and all gang headers below it. This yields
1614 * an in-core tree containing the contents of every gang header and the
1615 * bps for every constituent of the gang block.
1616 *
1617 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1618 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1619 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1620 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1621 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1622 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1623 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1624 * of the gang header plus zio_checksum_compute() of the data to update the
1625 * gang header's blk_cksum as described above.
1626 *
1627 * The two-phase assemble/issue model solves the problem of partial failure --
1628 * what if you'd freed part of a gang block but then couldn't read the
1629 * gang header for another part? Assembling the entire gang tree first
1630 * ensures that all the necessary gang header I/O has succeeded before
1631 * starting the actual work of free, claim, or write. Once the gang tree
1632 * is assembled, free and claim are in-memory operations that cannot fail.
1633 *
1634 * In the event that a gang write fails, zio_dva_unallocate() walks the
1635 * gang tree to immediately free (i.e. insert back into the space map)
1636 * everything we've allocated. This ensures that we don't get ENOSPC
1637 * errors during repeated suspend/resume cycles due to a flaky device.
1638 *
1639 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1640 * the gang tree, we won't modify the block, so we can safely defer the free
1641 * (knowing that the block is still intact). If we *can* assemble the gang
1642 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1643 * each constituent bp and we can allocate a new block on the next sync pass.
1644 *
1645 * In all cases, the gang tree allows complete recovery from partial failure.
1646 * ==========================================================================
1647 */
1648
1649static zio_t *
1650zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1651{
1652 if (gn != NULL)
1653 return (pio);
1654
1655 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1656 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1657 &pio->io_bookmark));
1658}
1659
1660zio_t *
1661zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1662{
1663 zio_t *zio;
1664
1665 if (gn != NULL) {
1666 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1667 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1668 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1669 /*
1670 * As we rewrite each gang header, the pipeline will compute
1671 * a new gang block header checksum for it; but no one will
1672 * compute a new data checksum, so we do that here. The one
1673 * exception is the gang leader: the pipeline already computed
1674 * its data checksum because that stage precedes gang assembly.
1675 * (Presently, nothing actually uses interior data checksums;
1676 * this is just good hygiene.)
1677 */
1678 if (gn != pio->io_gang_leader->io_gang_tree) {
1679 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1680 data, BP_GET_PSIZE(bp));
1681 }
1682 /*
1683 * If we are here to damage data for testing purposes,
1684 * leave the GBH alone so that we can detect the damage.
1685 */
1686 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1687 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1688 } else {
1689 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1690 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1691 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1692 }
1693
1694 return (zio);
1695}
1696
1697/* ARGSUSED */
1698zio_t *
1699zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1700{
1701 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1702 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
1703 ZIO_GANG_CHILD_FLAGS(pio)));
1704}
1705
1706/* ARGSUSED */
1707zio_t *
1708zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1709{
1710 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1711 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1712}
1713
1714static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1715 NULL,
1716 zio_read_gang,
1717 zio_rewrite_gang,
1718 zio_free_gang,
1719 zio_claim_gang,
1720 NULL
1721};
1722
1723static void zio_gang_tree_assemble_done(zio_t *zio);
1724
1725static zio_gang_node_t *
1726zio_gang_node_alloc(zio_gang_node_t **gnpp)
1727{
1728 zio_gang_node_t *gn;
1729
1730 ASSERT(*gnpp == NULL);
1731
1732 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1733 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1734 *gnpp = gn;
1735
1736 return (gn);
1737}
1738
1739static void
1740zio_gang_node_free(zio_gang_node_t **gnpp)
1741{
1742 zio_gang_node_t *gn = *gnpp;
1743
1744 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1745 ASSERT(gn->gn_child[g] == NULL);
1746
1747 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1748 kmem_free(gn, sizeof (*gn));
1749 *gnpp = NULL;
1750}
1751
1752static void
1753zio_gang_tree_free(zio_gang_node_t **gnpp)
1754{
1755 zio_gang_node_t *gn = *gnpp;
1756
1757 if (gn == NULL)
1758 return;
1759
1760 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1761 zio_gang_tree_free(&gn->gn_child[g]);
1762
1763 zio_gang_node_free(gnpp);
1764}
1765
1766static void
1767zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1768{
1769 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1770
1771 ASSERT(gio->io_gang_leader == gio);
1772 ASSERT(BP_IS_GANG(bp));
1773
1774 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1775 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1776 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1777}
1778
1779static void
1780zio_gang_tree_assemble_done(zio_t *zio)
1781{
1782 zio_t *gio = zio->io_gang_leader;
1783 zio_gang_node_t *gn = zio->io_private;
1784 blkptr_t *bp = zio->io_bp;
1785
1786 ASSERT(gio == zio_unique_parent(zio));
1787 ASSERT(zio->io_child_count == 0);
1788
1789 if (zio->io_error)
1790 return;
1791
1792 if (BP_SHOULD_BYTESWAP(bp))
1793 byteswap_uint64_array(zio->io_data, zio->io_size);
1794
1795 ASSERT(zio->io_data == gn->gn_gbh);
1796 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1797 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1798
1799 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1800 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1801 if (!BP_IS_GANG(gbp))
1802 continue;
1803 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1804 }
1805}
1806
1807static void
1808zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1809{
1810 zio_t *gio = pio->io_gang_leader;
1811 zio_t *zio;
1812
1813 ASSERT(BP_IS_GANG(bp) == !!gn);
1814 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1815 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1816
1817 /*
1818 * If you're a gang header, your data is in gn->gn_gbh.
1819 * If you're a gang member, your data is in 'data' and gn == NULL.
1820 */
1821 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1822
1823 if (gn != NULL) {
1824 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1825
1826 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1827 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1828 if (BP_IS_HOLE(gbp))
1829 continue;
1830 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1831 data = (char *)data + BP_GET_PSIZE(gbp);
1832 }
1833 }
1834
1835 if (gn == gio->io_gang_tree && gio->io_data != NULL)
1836 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1837
1838 if (zio != pio)
1839 zio_nowait(zio);
1840}
1841
1842static int
1843zio_gang_assemble(zio_t *zio)
1844{
1845 blkptr_t *bp = zio->io_bp;
1846
1847 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1848 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1849
1850 zio->io_gang_leader = zio;
1851
1852 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1853
1854 return (ZIO_PIPELINE_CONTINUE);
1855}
1856
1857static int
1858zio_gang_issue(zio_t *zio)
1859{
1860 blkptr_t *bp = zio->io_bp;
1861
1862 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1863 return (ZIO_PIPELINE_STOP);
1864
1865 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1866 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1867
1868 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1869 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1870 else
1871 zio_gang_tree_free(&zio->io_gang_tree);
1872
1873 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1874
1875 return (ZIO_PIPELINE_CONTINUE);
1876}
1877
1878static void
1879zio_write_gang_member_ready(zio_t *zio)
1880{
1881 zio_t *pio = zio_unique_parent(zio);
1882 zio_t *gio = zio->io_gang_leader;
1883 dva_t *cdva = zio->io_bp->blk_dva;
1884 dva_t *pdva = pio->io_bp->blk_dva;
1885 uint64_t asize;
1886
1887 if (BP_IS_HOLE(zio->io_bp))
1888 return;
1889
1890 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1891
1892 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1893 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1894 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1895 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1896 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1897
1898 mutex_enter(&pio->io_lock);
1899 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1900 ASSERT(DVA_GET_GANG(&pdva[d]));
1901 asize = DVA_GET_ASIZE(&pdva[d]);
1902 asize += DVA_GET_ASIZE(&cdva[d]);
1903 DVA_SET_ASIZE(&pdva[d], asize);
1904 }
1905 mutex_exit(&pio->io_lock);
1906}
1907
1908static int
1909zio_write_gang_block(zio_t *pio)
1910{
1911 spa_t *spa = pio->io_spa;
1912 blkptr_t *bp = pio->io_bp;
1913 zio_t *gio = pio->io_gang_leader;
1914 zio_t *zio;
1915 zio_gang_node_t *gn, **gnpp;
1916 zio_gbh_phys_t *gbh;
1917 uint64_t txg = pio->io_txg;
1918 uint64_t resid = pio->io_size;
1919 uint64_t lsize;
1920 int copies = gio->io_prop.zp_copies;
1921 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1922 zio_prop_t zp;
1923 int error;
1924
1925 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1926 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1927 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1928 if (error) {
1929 pio->io_error = error;
1930 return (ZIO_PIPELINE_CONTINUE);
1931 }
1932
1933 if (pio == gio) {
1934 gnpp = &gio->io_gang_tree;
1935 } else {
1936 gnpp = pio->io_private;
1937 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1938 }
1939
1940 gn = zio_gang_node_alloc(gnpp);
1941 gbh = gn->gn_gbh;
1942 bzero(gbh, SPA_GANGBLOCKSIZE);
1943
1944 /*
1945 * Create the gang header.
1946 */
1947 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1948 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1949
1950 /*
1951 * Create and nowait the gang children.
1952 */
1953 for (int g = 0; resid != 0; resid -= lsize, g++) {
1954 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1955 SPA_MINBLOCKSIZE);
1956 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1957
1958 zp.zp_checksum = gio->io_prop.zp_checksum;
1959 zp.zp_compress = ZIO_COMPRESS_OFF;
1960 zp.zp_type = DMU_OT_NONE;
1961 zp.zp_level = 0;
1962 zp.zp_copies = gio->io_prop.zp_copies;
1963 zp.zp_dedup = B_FALSE;
1964 zp.zp_dedup_verify = B_FALSE;
1965 zp.zp_nopwrite = B_FALSE;
1966
1967 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1968 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1969 zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g],
1970 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1971 &pio->io_bookmark));
1972 }
1973
1974 /*
1975 * Set pio's pipeline to just wait for zio to finish.
1976 */
1977 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1978
1979 zio_nowait(zio);
1980
1981 return (ZIO_PIPELINE_CONTINUE);
1982}
1983
1984/*
1985 * The zio_nop_write stage in the pipeline determines if allocating
1986 * a new bp is necessary. By leveraging a cryptographically secure checksum,
1987 * such as SHA256, we can compare the checksums of the new data and the old
1988 * to determine if allocating a new block is required. The nopwrite
1989 * feature can handle writes in either syncing or open context (i.e. zil
1990 * writes) and as a result is mutually exclusive with dedup.
1991 */
1992static int
1993zio_nop_write(zio_t *zio)
1994{
1995 blkptr_t *bp = zio->io_bp;
1996 blkptr_t *bp_orig = &zio->io_bp_orig;
1997 zio_prop_t *zp = &zio->io_prop;
1998
1999 ASSERT(BP_GET_LEVEL(bp) == 0);
2000 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2001 ASSERT(zp->zp_nopwrite);
2002 ASSERT(!zp->zp_dedup);
2003 ASSERT(zio->io_bp_override == NULL);
2004 ASSERT(IO_IS_ALLOCATING(zio));
2005
2006 /*
2007 * Check to see if the original bp and the new bp have matching
2008 * characteristics (i.e. same checksum, compression algorithms, etc).
2009 * If they don't then just continue with the pipeline which will
2010 * allocate a new bp.
2011 */
2012 if (BP_IS_HOLE(bp_orig) ||
2013 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
2014 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2015 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2016 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2017 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2018 return (ZIO_PIPELINE_CONTINUE);
2019
2020 /*
2021 * If the checksums match then reset the pipeline so that we
2022 * avoid allocating a new bp and issuing any I/O.
2023 */
2024 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2025 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
2026 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2027 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2028 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2029 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2030 sizeof (uint64_t)) == 0);
2031
2032 *bp = *bp_orig;
2033 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2034 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2035 }
2036
2037 return (ZIO_PIPELINE_CONTINUE);
2038}
2039
2040/*
2041 * ==========================================================================
2042 * Dedup
2043 * ==========================================================================
2044 */
2045static void
2046zio_ddt_child_read_done(zio_t *zio)
2047{
2048 blkptr_t *bp = zio->io_bp;
2049 ddt_entry_t *dde = zio->io_private;
2050 ddt_phys_t *ddp;
2051 zio_t *pio = zio_unique_parent(zio);
2052
2053 mutex_enter(&pio->io_lock);
2054 ddp = ddt_phys_select(dde, bp);
2055 if (zio->io_error == 0)
2056 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2057 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
2058 dde->dde_repair_data = zio->io_data;
2059 else
2060 zio_buf_free(zio->io_data, zio->io_size);
2061 mutex_exit(&pio->io_lock);
2062}
2063
2064static int
2065zio_ddt_read_start(zio_t *zio)
2066{
2067 blkptr_t *bp = zio->io_bp;
2068
2069 ASSERT(BP_GET_DEDUP(bp));
2070 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2071 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2072
2073 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2074 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2075 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2076 ddt_phys_t *ddp = dde->dde_phys;
2077 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2078 blkptr_t blk;
2079
2080 ASSERT(zio->io_vsd == NULL);
2081 zio->io_vsd = dde;
2082
2083 if (ddp_self == NULL)
2084 return (ZIO_PIPELINE_CONTINUE);
2085
2086 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2087 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2088 continue;
2089 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2090 &blk);
2091 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2092 zio_buf_alloc(zio->io_size), zio->io_size,
2093 zio_ddt_child_read_done, dde, zio->io_priority,
2094 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2095 &zio->io_bookmark));
2096 }
2097 return (ZIO_PIPELINE_CONTINUE);
2098 }
2099
2100 zio_nowait(zio_read(zio, zio->io_spa, bp,
2101 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2102 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2103
2104 return (ZIO_PIPELINE_CONTINUE);
2105}
2106
2107static int
2108zio_ddt_read_done(zio_t *zio)
2109{
2110 blkptr_t *bp = zio->io_bp;
2111
2112 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2113 return (ZIO_PIPELINE_STOP);
2114
2115 ASSERT(BP_GET_DEDUP(bp));
2116 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2117 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2118
2119 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2120 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2121 ddt_entry_t *dde = zio->io_vsd;
2122 if (ddt == NULL) {
2123 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2124 return (ZIO_PIPELINE_CONTINUE);
2125 }
2126 if (dde == NULL) {
2127 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2128 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2129 return (ZIO_PIPELINE_STOP);
2130 }
2131 if (dde->dde_repair_data != NULL) {
2132 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2133 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2134 }
2135 ddt_repair_done(ddt, dde);
2136 zio->io_vsd = NULL;
2137 }
2138
2139 ASSERT(zio->io_vsd == NULL);
2140
2141 return (ZIO_PIPELINE_CONTINUE);
2142}
2143
2144static boolean_t
2145zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2146{
2147 spa_t *spa = zio->io_spa;
2148
2149 /*
2150 * Note: we compare the original data, not the transformed data,
2151 * because when zio->io_bp is an override bp, we will not have
2152 * pushed the I/O transforms. That's an important optimization
2153 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2154 */
2155 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2156 zio_t *lio = dde->dde_lead_zio[p];
2157
2158 if (lio != NULL) {
2159 return (lio->io_orig_size != zio->io_orig_size ||
2160 bcmp(zio->io_orig_data, lio->io_orig_data,
2161 zio->io_orig_size) != 0);
2162 }
2163 }
2164
2165 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2166 ddt_phys_t *ddp = &dde->dde_phys[p];
2167
2168 if (ddp->ddp_phys_birth != 0) {
2169 arc_buf_t *abuf = NULL;
2170 uint32_t aflags = ARC_WAIT;
2171 blkptr_t blk = *zio->io_bp;
2172 int error;
2173
2174 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2175
2176 ddt_exit(ddt);
2177
2178 error = arc_read(NULL, spa, &blk,
2179 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2180 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2181 &aflags, &zio->io_bookmark);
2182
2183 if (error == 0) {
2184 if (arc_buf_size(abuf) != zio->io_orig_size ||
2185 bcmp(abuf->b_data, zio->io_orig_data,
2186 zio->io_orig_size) != 0)
2187 error = SET_ERROR(EEXIST);
2188 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2189 }
2190
2191 ddt_enter(ddt);
2192 return (error != 0);
2193 }
2194 }
2195
2196 return (B_FALSE);
2197}
2198
2199static void
2200zio_ddt_child_write_ready(zio_t *zio)
2201{
2202 int p = zio->io_prop.zp_copies;
2203 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2204 ddt_entry_t *dde = zio->io_private;
2205 ddt_phys_t *ddp = &dde->dde_phys[p];
2206 zio_t *pio;
2207
2208 if (zio->io_error)
2209 return;
2210
2211 ddt_enter(ddt);
2212
2213 ASSERT(dde->dde_lead_zio[p] == zio);
2214
2215 ddt_phys_fill(ddp, zio->io_bp);
2216
2217 while ((pio = zio_walk_parents(zio)) != NULL)
2218 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2219
2220 ddt_exit(ddt);
2221}
2222
2223static void
2224zio_ddt_child_write_done(zio_t *zio)
2225{
2226 int p = zio->io_prop.zp_copies;
2227 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2228 ddt_entry_t *dde = zio->io_private;
2229 ddt_phys_t *ddp = &dde->dde_phys[p];
2230
2231 ddt_enter(ddt);
2232
2233 ASSERT(ddp->ddp_refcnt == 0);
2234 ASSERT(dde->dde_lead_zio[p] == zio);
2235 dde->dde_lead_zio[p] = NULL;
2236
2237 if (zio->io_error == 0) {
2238 while (zio_walk_parents(zio) != NULL)
2239 ddt_phys_addref(ddp);
2240 } else {
2241 ddt_phys_clear(ddp);
2242 }
2243
2244 ddt_exit(ddt);
2245}
2246
2247static void
2248zio_ddt_ditto_write_done(zio_t *zio)
2249{
2250 int p = DDT_PHYS_DITTO;
2251 zio_prop_t *zp = &zio->io_prop;
2252 blkptr_t *bp = zio->io_bp;
2253 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2254 ddt_entry_t *dde = zio->io_private;
2255 ddt_phys_t *ddp = &dde->dde_phys[p];
2256 ddt_key_t *ddk = &dde->dde_key;
2257
2258 ddt_enter(ddt);
2259
2260 ASSERT(ddp->ddp_refcnt == 0);
2261 ASSERT(dde->dde_lead_zio[p] == zio);
2262 dde->dde_lead_zio[p] = NULL;
2263
2264 if (zio->io_error == 0) {
2265 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2266 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2267 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2268 if (ddp->ddp_phys_birth != 0)
2269 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2270 ddt_phys_fill(ddp, bp);
2271 }
2272
2273 ddt_exit(ddt);
2274}
2275
2276static int
2277zio_ddt_write(zio_t *zio)
2278{
2279 spa_t *spa = zio->io_spa;
2280 blkptr_t *bp = zio->io_bp;
2281 uint64_t txg = zio->io_txg;
2282 zio_prop_t *zp = &zio->io_prop;
2283 int p = zp->zp_copies;
2284 int ditto_copies;
2285 zio_t *cio = NULL;
2286 zio_t *dio = NULL;
2287 ddt_t *ddt = ddt_select(spa, bp);
2288 ddt_entry_t *dde;
2289 ddt_phys_t *ddp;
2290
2291 ASSERT(BP_GET_DEDUP(bp));
2292 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2293 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2294
2295 ddt_enter(ddt);
2296 dde = ddt_lookup(ddt, bp, B_TRUE);
2297 ddp = &dde->dde_phys[p];
2298
2299 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2300 /*
2301 * If we're using a weak checksum, upgrade to a strong checksum
2302 * and try again. If we're already using a strong checksum,
2303 * we can't resolve it, so just convert to an ordinary write.
2304 * (And automatically e-mail a paper to Nature?)
2305 */
2306 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2307 zp->zp_checksum = spa_dedup_checksum(spa);
2308 zio_pop_transforms(zio);
2309 zio->io_stage = ZIO_STAGE_OPEN;
2310 BP_ZERO(bp);
2311 } else {
2312 zp->zp_dedup = B_FALSE;
2313 }
2314 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2315 ddt_exit(ddt);
2316 return (ZIO_PIPELINE_CONTINUE);
2317 }
2318
2319 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2320 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2321
2322 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2323 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2324 zio_prop_t czp = *zp;
2325
2326 czp.zp_copies = ditto_copies;
2327
2328 /*
2329 * If we arrived here with an override bp, we won't have run
2330 * the transform stack, so we won't have the data we need to
2331 * generate a child i/o. So, toss the override bp and restart.
2332 * This is safe, because using the override bp is just an
2333 * optimization; and it's rare, so the cost doesn't matter.
2334 */
2335 if (zio->io_bp_override) {
2336 zio_pop_transforms(zio);
2337 zio->io_stage = ZIO_STAGE_OPEN;
2338 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2339 zio->io_bp_override = NULL;
2340 BP_ZERO(bp);
2341 ddt_exit(ddt);
2342 return (ZIO_PIPELINE_CONTINUE);
2343 }
2344
2345 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2346 zio->io_orig_size, &czp, NULL, NULL,
2347 zio_ddt_ditto_write_done, dde, zio->io_priority,
2348 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2349
2350 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2351 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2352 }
2353
2354 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2355 if (ddp->ddp_phys_birth != 0)
2356 ddt_bp_fill(ddp, bp, txg);
2357 if (dde->dde_lead_zio[p] != NULL)
2358 zio_add_child(zio, dde->dde_lead_zio[p]);
2359 else
2360 ddt_phys_addref(ddp);
2361 } else if (zio->io_bp_override) {
2362 ASSERT(bp->blk_birth == txg);
2363 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2364 ddt_phys_fill(ddp, bp);
2365 ddt_phys_addref(ddp);
2366 } else {
2367 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2368 zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL,
2369 zio_ddt_child_write_done, dde, zio->io_priority,
2370 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2371
2372 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2373 dde->dde_lead_zio[p] = cio;
2374 }
2375
2376 ddt_exit(ddt);
2377
2378 if (cio)
2379 zio_nowait(cio);
2380 if (dio)
2381 zio_nowait(dio);
2382
2383 return (ZIO_PIPELINE_CONTINUE);
2384}
2385
2386ddt_entry_t *freedde; /* for debugging */
2387
2388static int
2389zio_ddt_free(zio_t *zio)
2390{
2391 spa_t *spa = zio->io_spa;
2392 blkptr_t *bp = zio->io_bp;
2393 ddt_t *ddt = ddt_select(spa, bp);
2394 ddt_entry_t *dde;
2395 ddt_phys_t *ddp;
2396
2397 ASSERT(BP_GET_DEDUP(bp));
2398 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2399
2400 ddt_enter(ddt);
2401 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2402 ddp = ddt_phys_select(dde, bp);
2403 ddt_phys_decref(ddp);
2404 ddt_exit(ddt);
2405
2406 return (ZIO_PIPELINE_CONTINUE);
2407}
2408
2409/*
2410 * ==========================================================================
2411 * Allocate and free blocks
2412 * ==========================================================================
2413 */
2414static int
2415zio_dva_allocate(zio_t *zio)
2416{
2417 spa_t *spa = zio->io_spa;
2418 metaslab_class_t *mc = spa_normal_class(spa);
2419 blkptr_t *bp = zio->io_bp;
2420 int error;
2421 int flags = 0;
2422
2423 if (zio->io_gang_leader == NULL) {
2424 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2425 zio->io_gang_leader = zio;
2426 }
2427
2428 ASSERT(BP_IS_HOLE(bp));
2429 ASSERT0(BP_GET_NDVAS(bp));
2430 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2431 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2432 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2433
2434 /*
2435 * The dump device does not support gang blocks so allocation on
2436 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2437 * the "fast" gang feature.
2438 */
2439 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2440 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2441 METASLAB_GANG_CHILD : 0;
2442 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2443 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2444
2445 if (error) {
2446 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2447 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2448 error);
2449 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2450 return (zio_write_gang_block(zio));
2451 zio->io_error = error;
2452 }
2453
2454 return (ZIO_PIPELINE_CONTINUE);
2455}
2456
2457static int
2458zio_dva_free(zio_t *zio)
2459{
2460 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2461
2462 return (ZIO_PIPELINE_CONTINUE);
2463}
2464
2465static int
2466zio_dva_claim(zio_t *zio)
2467{
2468 int error;
2469
2470 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2471 if (error)
2472 zio->io_error = error;
2473
2474 return (ZIO_PIPELINE_CONTINUE);
2475}
2476
2477/*
2478 * Undo an allocation. This is used by zio_done() when an I/O fails
2479 * and we want to give back the block we just allocated.
2480 * This handles both normal blocks and gang blocks.
2481 */
2482static void
2483zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2484{
2485 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2486 ASSERT(zio->io_bp_override == NULL);
2487
2488 if (!BP_IS_HOLE(bp))
2489 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2490
2491 if (gn != NULL) {
2492 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2493 zio_dva_unallocate(zio, gn->gn_child[g],
2494 &gn->gn_gbh->zg_blkptr[g]);
2495 }
2496 }
2497}
2498
2499/*
2500 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2501 */
2502int
2503zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2504 uint64_t size, boolean_t use_slog)
2505{
2506 int error = 1;
2507
2508 ASSERT(txg > spa_syncing_txg(spa));
2509
2510 /*
2511 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2512 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2513 * when allocating them.
2514 */
2515 if (use_slog) {
2516 error = metaslab_alloc(spa, spa_log_class(spa), size,
2517 new_bp, 1, txg, old_bp,
2518 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2519 }
2520
2521 if (error) {
2522 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2523 new_bp, 1, txg, old_bp,
2524 METASLAB_HINTBP_AVOID);
2525 }
2526
2527 if (error == 0) {
2528 BP_SET_LSIZE(new_bp, size);
2529 BP_SET_PSIZE(new_bp, size);
2530 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2531 BP_SET_CHECKSUM(new_bp,
2532 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2533 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2534 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2535 BP_SET_LEVEL(new_bp, 0);
2536 BP_SET_DEDUP(new_bp, 0);
2537 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2538 }
2539
2540 return (error);
2541}
2542
2543/*
2544 * Free an intent log block.
2545 */
2546void
2547zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2548{
2549 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2550 ASSERT(!BP_IS_GANG(bp));
2551
2552 zio_free(spa, txg, bp);
2553}
2554
2555/*
2556 * ==========================================================================
2557 * Read, write and delete to physical devices
2558 * ==========================================================================
2559 */
2560static int
2561zio_vdev_io_start(zio_t *zio)
2562{
2563 vdev_t *vd = zio->io_vd;
2564 uint64_t align;
2565 spa_t *spa = zio->io_spa;
2566 int ret;
2567
2568 ASSERT(zio->io_error == 0);
2569 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2570
2571 if (vd == NULL) {
2572 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2573 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2574
2575 /*
2576 * The mirror_ops handle multiple DVAs in a single BP.
2577 */
2578 return (vdev_mirror_ops.vdev_op_io_start(zio));
2579 }
2580
2581 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE &&
2582 zio->io_priority == ZIO_PRIORITY_NOW) {
2583 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
2584 return (ZIO_PIPELINE_CONTINUE);
2585 }
2586
2587 /*
2588 * We keep track of time-sensitive I/Os so that the scan thread
2589 * can quickly react to certain workloads. In particular, we care
2590 * about non-scrubbing, top-level reads and writes with the following
2591 * characteristics:
2592 * - synchronous writes of user data to non-slog devices
2593 * - any reads of user data
2594 * When these conditions are met, adjust the timestamp of spa_last_io
2595 * which allows the scan thread to adjust its workload accordingly.
2596 */
2597 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2598 vd == vd->vdev_top && !vd->vdev_islog &&
2599 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2600 zio->io_txg != spa_syncing_txg(spa)) {
2601 uint64_t old = spa->spa_last_io;
2602 uint64_t new = ddi_get_lbolt64();
2603 if (old != new)
2604 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2605 }
2606
2607 align = 1ULL << vd->vdev_top->vdev_ashift;
2608
2609 if ((!(zio->io_flags & ZIO_FLAG_PHYSICAL) ||
2610 (vd->vdev_top->vdev_physical_ashift > SPA_MINBLOCKSHIFT)) &&
2611 P2PHASE(zio->io_size, align) != 0) {
2612 /* Transform logical writes to be a full physical block size. */
2613 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2614 char *abuf = NULL;
2615 if (zio->io_type == ZIO_TYPE_READ ||
2616 zio->io_type == ZIO_TYPE_WRITE)
2617 abuf = zio_buf_alloc(asize);
2618 ASSERT(vd == vd->vdev_top);
2619 if (zio->io_type == ZIO_TYPE_WRITE) {
2620 bcopy(zio->io_data, abuf, zio->io_size);
2621 bzero(abuf + zio->io_size, asize - zio->io_size);
2622 }
2623 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
2624 zio_subblock);
2625 }
2626
2627 /*
2628 * If this is not a physical io, make sure that it is properly aligned
2629 * before proceeding.
2630 */
2631 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
2632 ASSERT0(P2PHASE(zio->io_offset, align));
2633 ASSERT0(P2PHASE(zio->io_size, align));
2634 } else {
2635 /*
2636 * For physical writes, we allow 512b aligned writes and assume
2637 * the device will perform a read-modify-write as necessary.
2638 */
2639 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
2640 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
2641 }
2642
2643 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
2644
2645 /*
2646 * If this is a repair I/O, and there's no self-healing involved --
2647 * that is, we're just resilvering what we expect to resilver --
2648 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2649 * This prevents spurious resilvering with nested replication.
2650 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2651 * A is out of date, we'll read from C+D, then use the data to
2652 * resilver A+B -- but we don't actually want to resilver B, just A.
2653 * The top-level mirror has no way to know this, so instead we just
2654 * discard unnecessary repairs as we work our way down the vdev tree.
2655 * The same logic applies to any form of nested replication:
2656 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2657 */
2658 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2659 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2660 zio->io_txg != 0 && /* not a delegated i/o */
2661 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2662 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2663 zio_vdev_io_bypass(zio);
2664 return (ZIO_PIPELINE_CONTINUE);
2665 }
2666
2667 if (vd->vdev_ops->vdev_op_leaf) {
2668 switch (zio->io_type) {
2669 case ZIO_TYPE_READ:
2670 if (vdev_cache_read(zio))
2671 return (ZIO_PIPELINE_CONTINUE);
2672 /* FALLTHROUGH */
2673 case ZIO_TYPE_WRITE:
2674 case ZIO_TYPE_FREE:
2675 if ((zio = vdev_queue_io(zio)) == NULL)
2676 return (ZIO_PIPELINE_STOP);
2677
2678 if (!vdev_accessible(vd, zio)) {
2679 zio->io_error = SET_ERROR(ENXIO);
2680 zio_interrupt(zio);
2681 return (ZIO_PIPELINE_STOP);
2682 }
2683 break;
2684 }
2685 /*
2686 * Note that we ignore repair writes for TRIM because they can
2687 * conflict with normal writes. This isn't an issue because, by
2688 * definition, we only repair blocks that aren't freed.
2689 */
2690 if (zio->io_type == ZIO_TYPE_WRITE &&
2691 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2692 !trim_map_write_start(zio))
2693 return (ZIO_PIPELINE_STOP);
2694 }
2695
2696 ret = vd->vdev_ops->vdev_op_io_start(zio);
2697 ASSERT(ret == ZIO_PIPELINE_STOP);
2698
2699 return (ret);
2700}
2701
2702static int
2703zio_vdev_io_done(zio_t *zio)
2704{
2705 vdev_t *vd = zio->io_vd;
2706 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2707 boolean_t unexpected_error = B_FALSE;
2708
2709 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2710 return (ZIO_PIPELINE_STOP);
2711
2712 ASSERT(zio->io_type == ZIO_TYPE_READ ||
2713 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
2714
2715 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2716 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE ||
2717 zio->io_type == ZIO_TYPE_FREE)) {
2718
2719 if (zio->io_type == ZIO_TYPE_WRITE &&
2720 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
2721 trim_map_write_done(zio);
2722
2723 vdev_queue_io_done(zio);
2724
2725 if (zio->io_type == ZIO_TYPE_WRITE)
2726 vdev_cache_write(zio);
2727
2728 if (zio_injection_enabled && zio->io_error == 0)
2729 zio->io_error = zio_handle_device_injection(vd,
2730 zio, EIO);
2731
2732 if (zio_injection_enabled && zio->io_error == 0)
2733 zio->io_error = zio_handle_label_injection(zio, EIO);
2734
2735 if (zio->io_error) {
2736 if (zio->io_error == ENOTSUP &&
2737 zio->io_type == ZIO_TYPE_FREE) {
2738 /* Not all devices support TRIM. */
2739 } else if (!vdev_accessible(vd, zio)) {
2740 zio->io_error = SET_ERROR(ENXIO);
2741 } else {
2742 unexpected_error = B_TRUE;
2743 }
2744 }
2745 }
2746
2747 ops->vdev_op_io_done(zio);
2748
2749 if (unexpected_error)
2750 VERIFY(vdev_probe(vd, zio) == NULL);
2751
2752 return (ZIO_PIPELINE_CONTINUE);
2753}
2754
2755/*
2756 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2757 * disk, and use that to finish the checksum ereport later.
2758 */
2759static void
2760zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2761 const void *good_buf)
2762{
2763 /* no processing needed */
2764 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2765}
2766
2767/*ARGSUSED*/
2768void
2769zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2770{
2771 void *buf = zio_buf_alloc(zio->io_size);
2772
2773 bcopy(zio->io_data, buf, zio->io_size);
2774
2775 zcr->zcr_cbinfo = zio->io_size;
2776 zcr->zcr_cbdata = buf;
2777 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2778 zcr->zcr_free = zio_buf_free;
2779}
2780
2781static int
2782zio_vdev_io_assess(zio_t *zio)
2783{
2784 vdev_t *vd = zio->io_vd;
2785
2786 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2787 return (ZIO_PIPELINE_STOP);
2788
2789 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2790 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2791
2792 if (zio->io_vsd != NULL) {
2793 zio->io_vsd_ops->vsd_free(zio);
2794 zio->io_vsd = NULL;
2795 }
2796
2797 if (zio_injection_enabled && zio->io_error == 0)
2798 zio->io_error = zio_handle_fault_injection(zio, EIO);
2799
2800 if (zio->io_type == ZIO_TYPE_FREE &&
2801 zio->io_priority != ZIO_PRIORITY_NOW) {
2802 switch (zio->io_error) {
2803 case 0:
2804 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
2805 ZIO_TRIM_STAT_BUMP(success);
2806 break;
2807 case EOPNOTSUPP:
2808 ZIO_TRIM_STAT_BUMP(unsupported);
2809 break;
2810 default:
2811 ZIO_TRIM_STAT_BUMP(failed);
2812 break;
2813 }
2814 }
2815
2816 /*
2817 * If the I/O failed, determine whether we should attempt to retry it.
2818 *
2819 * On retry, we cut in line in the issue queue, since we don't want
2820 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2821 */
2822 if (zio->io_error && vd == NULL &&
2823 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2824 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2825 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2826 zio->io_error = 0;
2827 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2828 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2829 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2830 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2831 zio_requeue_io_start_cut_in_line);
2832 return (ZIO_PIPELINE_STOP);
2833 }
2834
2835 /*
2836 * If we got an error on a leaf device, convert it to ENXIO
2837 * if the device is not accessible at all.
2838 */
2839 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2840 !vdev_accessible(vd, zio))
2841 zio->io_error = SET_ERROR(ENXIO);
2842
2843 /*
2844 * If we can't write to an interior vdev (mirror or RAID-Z),
2845 * set vdev_cant_write so that we stop trying to allocate from it.
2846 */
2847 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2848 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2849 vd->vdev_cant_write = B_TRUE;
2850 }
2851
2852 if (zio->io_error)
2853 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2854
2855 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2856 zio->io_physdone != NULL) {
2857 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
2858 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
2859 zio->io_physdone(zio->io_logical);
2860 }
2861
2862 return (ZIO_PIPELINE_CONTINUE);
2863}
2864
2865void
2866zio_vdev_io_reissue(zio_t *zio)
2867{
2868 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2869 ASSERT(zio->io_error == 0);
2870
2871 zio->io_stage >>= 1;
2872}
2873
2874void
2875zio_vdev_io_redone(zio_t *zio)
2876{
2877 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2878
2879 zio->io_stage >>= 1;
2880}
2881
2882void
2883zio_vdev_io_bypass(zio_t *zio)
2884{
2885 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2886 ASSERT(zio->io_error == 0);
2887
2888 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2889 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2890}
2891
2892/*
2893 * ==========================================================================
2894 * Generate and verify checksums
2895 * ==========================================================================
2896 */
2897static int
2898zio_checksum_generate(zio_t *zio)
2899{
2900 blkptr_t *bp = zio->io_bp;
2901 enum zio_checksum checksum;
2902
2903 if (bp == NULL) {
2904 /*
2905 * This is zio_write_phys().
2906 * We're either generating a label checksum, or none at all.
2907 */
2908 checksum = zio->io_prop.zp_checksum;
2909
2910 if (checksum == ZIO_CHECKSUM_OFF)
2911 return (ZIO_PIPELINE_CONTINUE);
2912
2913 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2914 } else {
2915 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2916 ASSERT(!IO_IS_ALLOCATING(zio));
2917 checksum = ZIO_CHECKSUM_GANG_HEADER;
2918 } else {
2919 checksum = BP_GET_CHECKSUM(bp);
2920 }
2921 }
2922
2923 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2924
2925 return (ZIO_PIPELINE_CONTINUE);
2926}
2927
2928static int
2929zio_checksum_verify(zio_t *zio)
2930{
2931 zio_bad_cksum_t info;
2932 blkptr_t *bp = zio->io_bp;
2933 int error;
2934
2935 ASSERT(zio->io_vd != NULL);
2936
2937 if (bp == NULL) {
2938 /*
2939 * This is zio_read_phys().
2940 * We're either verifying a label checksum, or nothing at all.
2941 */
2942 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2943 return (ZIO_PIPELINE_CONTINUE);
2944
2945 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2946 }
2947
2948 if ((error = zio_checksum_error(zio, &info)) != 0) {
2949 zio->io_error = error;
2950 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2951 zfs_ereport_start_checksum(zio->io_spa,
2952 zio->io_vd, zio, zio->io_offset,
2953 zio->io_size, NULL, &info);
2954 }
2955 }
2956
2957 return (ZIO_PIPELINE_CONTINUE);
2958}
2959
2960/*
2961 * Called by RAID-Z to ensure we don't compute the checksum twice.
2962 */
2963void
2964zio_checksum_verified(zio_t *zio)
2965{
2966 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2967}
2968
2969/*
2970 * ==========================================================================
2971 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2972 * An error of 0 indicates success. ENXIO indicates whole-device failure,
2973 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2974 * indicate errors that are specific to one I/O, and most likely permanent.
2975 * Any other error is presumed to be worse because we weren't expecting it.
2976 * ==========================================================================
2977 */
2978int
2979zio_worst_error(int e1, int e2)
2980{
2981 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2982 int r1, r2;
2983
2984 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2985 if (e1 == zio_error_rank[r1])
2986 break;
2987
2988 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2989 if (e2 == zio_error_rank[r2])
2990 break;
2991
2992 return (r1 > r2 ? e1 : e2);
2993}
2994
2995/*
2996 * ==========================================================================
2997 * I/O completion
2998 * ==========================================================================
2999 */
3000static int
3001zio_ready(zio_t *zio)
3002{
3003 blkptr_t *bp = zio->io_bp;
3004 zio_t *pio, *pio_next;
3005
3006 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
3007 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
3008 return (ZIO_PIPELINE_STOP);
3009
3010 if (zio->io_ready) {
3011 ASSERT(IO_IS_ALLOCATING(zio));
3012 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3013 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3014 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3015
3016 zio->io_ready(zio);
3017 }
3018
3019 if (bp != NULL && bp != &zio->io_bp_copy)
3020 zio->io_bp_copy = *bp;
3021
3022 if (zio->io_error)
3023 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3024
3025 mutex_enter(&zio->io_lock);
3026 zio->io_state[ZIO_WAIT_READY] = 1;
3027 pio = zio_walk_parents(zio);
3028 mutex_exit(&zio->io_lock);
3029
3030 /*
3031 * As we notify zio's parents, new parents could be added.
3032 * New parents go to the head of zio's io_parent_list, however,
3033 * so we will (correctly) not notify them. The remainder of zio's
3034 * io_parent_list, from 'pio_next' onward, cannot change because
3035 * all parents must wait for us to be done before they can be done.
3036 */
3037 for (; pio != NULL; pio = pio_next) {
3038 pio_next = zio_walk_parents(zio);
3039 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3040 }
3041
3042 if (zio->io_flags & ZIO_FLAG_NODATA) {
3043 if (BP_IS_GANG(bp)) {
3044 zio->io_flags &= ~ZIO_FLAG_NODATA;
3045 } else {
3046 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
3047 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3048 }
3049 }
3050
3051 if (zio_injection_enabled &&
3052 zio->io_spa->spa_syncing_txg == zio->io_txg)
3053 zio_handle_ignored_writes(zio);
3054
3055 return (ZIO_PIPELINE_CONTINUE);
3056}
3057
3058static int
3059zio_done(zio_t *zio)
3060{
3061 spa_t *spa = zio->io_spa;
3062 zio_t *lio = zio->io_logical;
3063 blkptr_t *bp = zio->io_bp;
3064 vdev_t *vd = zio->io_vd;
3065 uint64_t psize = zio->io_size;
3066 zio_t *pio, *pio_next;
3067
3068 /*
3069 * If our children haven't all completed,
3070 * wait for them and then repeat this pipeline stage.
3071 */
3072 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
3073 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
3074 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
3075 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
3076 return (ZIO_PIPELINE_STOP);
3077
3078 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3079 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3080 ASSERT(zio->io_children[c][w] == 0);
3081
3082 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3083 ASSERT(bp->blk_pad[0] == 0);
3084 ASSERT(bp->blk_pad[1] == 0);
3085 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3086 (bp == zio_unique_parent(zio)->io_bp));
3087 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3088 zio->io_bp_override == NULL &&
3089 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3090 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3091 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3092 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3093 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3094 }
3095 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3096 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3097 }
3098
3099 /*
3100 * If there were child vdev/gang/ddt errors, they apply to us now.
3101 */
3102 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3103 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3104 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3105
3106 /*
3107 * If the I/O on the transformed data was successful, generate any
3108 * checksum reports now while we still have the transformed data.
3109 */
3110 if (zio->io_error == 0) {
3111 while (zio->io_cksum_report != NULL) {
3112 zio_cksum_report_t *zcr = zio->io_cksum_report;
3113 uint64_t align = zcr->zcr_align;
3114 uint64_t asize = P2ROUNDUP(psize, align);
3115 char *abuf = zio->io_data;
3116
3117 if (asize != psize) {
3118 abuf = zio_buf_alloc(asize);
3119 bcopy(zio->io_data, abuf, psize);
3120 bzero(abuf + psize, asize - psize);
3121 }
3122
3123 zio->io_cksum_report = zcr->zcr_next;
3124 zcr->zcr_next = NULL;
3125 zcr->zcr_finish(zcr, abuf);
3126 zfs_ereport_free_checksum(zcr);
3127
3128 if (asize != psize)
3129 zio_buf_free(abuf, asize);
3130 }
3131 }
3132
3133 zio_pop_transforms(zio); /* note: may set zio->io_error */
3134
3135 vdev_stat_update(zio, psize);
3136
3137 if (zio->io_error) {
3138 /*
3139 * If this I/O is attached to a particular vdev,
3140 * generate an error message describing the I/O failure
3141 * at the block level. We ignore these errors if the
3142 * device is currently unavailable.
3143 */
3144 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3145 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3146
3147 if ((zio->io_error == EIO || !(zio->io_flags &
3148 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3149 zio == lio) {
3150 /*
3151 * For logical I/O requests, tell the SPA to log the
3152 * error and generate a logical data ereport.
3153 */
3154 spa_log_error(spa, zio);
3155 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3156 0, 0);
3157 }
3158 }
3159
3160 if (zio->io_error && zio == lio) {
3161 /*
3162 * Determine whether zio should be reexecuted. This will
3163 * propagate all the way to the root via zio_notify_parent().
3164 */
3165 ASSERT(vd == NULL && bp != NULL);
3166 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3167
3168 if (IO_IS_ALLOCATING(zio) &&
3169 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3170 if (zio->io_error != ENOSPC)
3171 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3172 else
3173 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3174 }
3175
3176 if ((zio->io_type == ZIO_TYPE_READ ||
3177 zio->io_type == ZIO_TYPE_FREE) &&
3178 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3179 zio->io_error == ENXIO &&
3180 spa_load_state(spa) == SPA_LOAD_NONE &&
3181 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3182 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3183
3184 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3185 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3186
3187 /*
3188 * Here is a possibly good place to attempt to do
3189 * either combinatorial reconstruction or error correction
3190 * based on checksums. It also might be a good place
3191 * to send out preliminary ereports before we suspend
3192 * processing.
3193 */
3194 }
3195
3196 /*
3197 * If there were logical child errors, they apply to us now.
3198 * We defer this until now to avoid conflating logical child
3199 * errors with errors that happened to the zio itself when
3200 * updating vdev stats and reporting FMA events above.
3201 */
3202 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3203
3204 if ((zio->io_error || zio->io_reexecute) &&
3205 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3206 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3207 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3208
3209 zio_gang_tree_free(&zio->io_gang_tree);
3210
3211 /*
3212 * Godfather I/Os should never suspend.
3213 */
3214 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3215 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3216 zio->io_reexecute = 0;
3217
3218 if (zio->io_reexecute) {
3219 /*
3220 * This is a logical I/O that wants to reexecute.
3221 *
3222 * Reexecute is top-down. When an i/o fails, if it's not
3223 * the root, it simply notifies its parent and sticks around.
3224 * The parent, seeing that it still has children in zio_done(),
3225 * does the same. This percolates all the way up to the root.
3226 * The root i/o will reexecute or suspend the entire tree.
3227 *
3228 * This approach ensures that zio_reexecute() honors
3229 * all the original i/o dependency relationships, e.g.
3230 * parents not executing until children are ready.
3231 */
3232 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3233
3234 zio->io_gang_leader = NULL;
3235
3236 mutex_enter(&zio->io_lock);
3237 zio->io_state[ZIO_WAIT_DONE] = 1;
3238 mutex_exit(&zio->io_lock);
3239
3240 /*
3241 * "The Godfather" I/O monitors its children but is
3242 * not a true parent to them. It will track them through
3243 * the pipeline but severs its ties whenever they get into
3244 * trouble (e.g. suspended). This allows "The Godfather"
3245 * I/O to return status without blocking.
3246 */
3247 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3248 zio_link_t *zl = zio->io_walk_link;
3249 pio_next = zio_walk_parents(zio);
3250
3251 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3252 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3253 zio_remove_child(pio, zio, zl);
3254 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3255 }
3256 }
3257
3258 if ((pio = zio_unique_parent(zio)) != NULL) {
3259 /*
3260 * We're not a root i/o, so there's nothing to do
3261 * but notify our parent. Don't propagate errors
3262 * upward since we haven't permanently failed yet.
3263 */
3264 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3265 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3266 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3267 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3268 /*
3269 * We'd fail again if we reexecuted now, so suspend
3270 * until conditions improve (e.g. device comes online).
3271 */
3272 zio_suspend(spa, zio);
3273 } else {
3274 /*
3275 * Reexecution is potentially a huge amount of work.
3276 * Hand it off to the otherwise-unused claim taskq.
3277 */
3278#if defined(illumos) || !defined(_KERNEL)
3279 ASSERT(zio->io_tqent.tqent_next == NULL);
3280#else
3281 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
3282#endif
3283 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3284 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3285 0, &zio->io_tqent);
3286 }
3287 return (ZIO_PIPELINE_STOP);
3288 }
3289
3290 ASSERT(zio->io_child_count == 0);
3291 ASSERT(zio->io_reexecute == 0);
3292 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3293
3294 /*
3295 * Report any checksum errors, since the I/O is complete.
3296 */
3297 while (zio->io_cksum_report != NULL) {
3298 zio_cksum_report_t *zcr = zio->io_cksum_report;
3299 zio->io_cksum_report = zcr->zcr_next;
3300 zcr->zcr_next = NULL;
3301 zcr->zcr_finish(zcr, NULL);
3302 zfs_ereport_free_checksum(zcr);
3303 }
3304
3305 /*
3306 * It is the responsibility of the done callback to ensure that this
3307 * particular zio is no longer discoverable for adoption, and as
3308 * such, cannot acquire any new parents.
3309 */
3310 if (zio->io_done)
3311 zio->io_done(zio);
3312
3313 mutex_enter(&zio->io_lock);
3314 zio->io_state[ZIO_WAIT_DONE] = 1;
3315 mutex_exit(&zio->io_lock);
3316
3317 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3318 zio_link_t *zl = zio->io_walk_link;
3319 pio_next = zio_walk_parents(zio);
3320 zio_remove_child(pio, zio, zl);
3321 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3322 }
3323
3324 if (zio->io_waiter != NULL) {
3325 mutex_enter(&zio->io_lock);
3326 zio->io_executor = NULL;
3327 cv_broadcast(&zio->io_cv);
3328 mutex_exit(&zio->io_lock);
3329 } else {
3330 zio_destroy(zio);
3331 }
3332
3333 return (ZIO_PIPELINE_STOP);
3334}
3335
3336/*
3337 * ==========================================================================
3338 * I/O pipeline definition
3339 * ==========================================================================
3340 */
3341static zio_pipe_stage_t *zio_pipeline[] = {
3342 NULL,
3343 zio_read_bp_init,
3344 zio_free_bp_init,
3345 zio_issue_async,
3346 zio_write_bp_init,
3347 zio_checksum_generate,
3348 zio_nop_write,
3349 zio_ddt_read_start,
3350 zio_ddt_read_done,
3351 zio_ddt_write,
3352 zio_ddt_free,
3353 zio_gang_assemble,
3354 zio_gang_issue,
3355 zio_dva_allocate,
3356 zio_dva_free,
3357 zio_dva_claim,
3358 zio_ready,
3359 zio_vdev_io_start,
3360 zio_vdev_io_done,
3361 zio_vdev_io_assess,
3362 zio_checksum_verify,
3363 zio_done
3364};
3365
3366/* dnp is the dnode for zb1->zb_object */
3367boolean_t
3368zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_phys_t *zb1,
3369 const zbookmark_phys_t *zb2)
3370{
3371 uint64_t zb1nextL0, zb2thisobj;
3372
3373 ASSERT(zb1->zb_objset == zb2->zb_objset);
3374 ASSERT(zb2->zb_level == 0);
3375
3376 /* The objset_phys_t isn't before anything. */
3377 if (dnp == NULL)
3378 return (B_FALSE);
3379
3380 zb1nextL0 = (zb1->zb_blkid + 1) <<
3381 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3382
3383 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3384 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3385
3386 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3387 uint64_t nextobj = zb1nextL0 *
3388 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3389 return (nextobj <= zb2thisobj);
3390 }
3391
3392 if (zb1->zb_object < zb2thisobj)
3393 return (B_TRUE);
3394 if (zb1->zb_object > zb2thisobj)
3395 return (B_FALSE);
3396 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3397 return (B_FALSE);
3398 return (zb1nextL0 <= zb2->zb_blkid);
3399}
|