]> git.proxmox.com Git - mirror_zfs-debian.git/blob - module/zfs/zio.c
projects / mirror_zfs-debian.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Add FASTWRITE algorithm for synchronous writes.
[mirror_zfs-debian.git] / module / zfs / zio.c
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) 2012 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
40 /*
41 * ==========================================================================
42 * I/O priority table
43 * ==========================================================================
44 */
45 uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = {
46 0, /* ZIO_PRIORITY_NOW */
47 0, /* ZIO_PRIORITY_SYNC_READ */
48 0, /* ZIO_PRIORITY_SYNC_WRITE */
49 0, /* ZIO_PRIORITY_LOG_WRITE */
50 1, /* ZIO_PRIORITY_CACHE_FILL */
51 1, /* ZIO_PRIORITY_AGG */
52 4, /* ZIO_PRIORITY_FREE */
53 4, /* ZIO_PRIORITY_ASYNC_WRITE */
54 6, /* ZIO_PRIORITY_ASYNC_READ */
55 10, /* ZIO_PRIORITY_RESILVER */
56 20, /* ZIO_PRIORITY_SCRUB */
57 2, /* ZIO_PRIORITY_DDT_PREFETCH */
58 };
59
60 /*
61 * ==========================================================================
62 * I/O type descriptions
63 * ==========================================================================
64 */
65 char *zio_type_name[ZIO_TYPES] = {
66 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl"
67 };
68
69 /*
70 * ==========================================================================
71 * I/O kmem caches
72 * ==========================================================================
73 */
74 kmem_cache_t *zio_cache;
75 kmem_cache_t *zio_link_cache;
76 kmem_cache_t *zio_vdev_cache;
77 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
78 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
79 int zio_bulk_flags = 0;
80 int zio_delay_max = ZIO_DELAY_MAX;
81
82 #ifdef _KERNEL
83 extern vmem_t *zio_alloc_arena;
84 #endif
85 extern int zfs_mg_alloc_failures;
86
87 /*
88 * An allocating zio is one that either currently has the DVA allocate
89 * stage set or will have it later in its lifetime.
90 */
91 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
92
93 int zio_requeue_io_start_cut_in_line = 1;
94
95 #ifdef ZFS_DEBUG
96 int zio_buf_debug_limit = 16384;
97 #else
98 int zio_buf_debug_limit = 0;
99 #endif
100
101 static inline void __zio_execute(zio_t *zio);
102
103 static int
104 zio_cons(void *arg, void *unused, int kmflag)
105 {
106 zio_t *zio = arg;
107
108 bzero(zio, sizeof (zio_t));
109
110 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
111 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
112
113 list_create(&zio->io_parent_list, sizeof (zio_link_t),
114 offsetof(zio_link_t, zl_parent_node));
115 list_create(&zio->io_child_list, sizeof (zio_link_t),
116 offsetof(zio_link_t, zl_child_node));
117
118 return (0);
119 }
120
121 static void
122 zio_dest(void *arg, void *unused)
123 {
124 zio_t *zio = arg;
125
126 mutex_destroy(&zio->io_lock);
127 cv_destroy(&zio->io_cv);
128 list_destroy(&zio->io_parent_list);
129 list_destroy(&zio->io_child_list);
130 }
131
132 void
133 zio_init(void)
134 {
135 size_t c;
136 vmem_t *data_alloc_arena = NULL;
137
138 #ifdef _KERNEL
139 data_alloc_arena = zio_alloc_arena;
140 #endif
141 zio_cache = kmem_cache_create("zio_cache", sizeof (zio_t), 0,
142 zio_cons, zio_dest, NULL, NULL, NULL, KMC_KMEM);
143 zio_link_cache = kmem_cache_create("zio_link_cache",
144 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, KMC_KMEM);
145 zio_vdev_cache = kmem_cache_create("zio_vdev_cache", sizeof(vdev_io_t),
146 PAGESIZE, NULL, NULL, NULL, NULL, NULL, KMC_VMEM);
147
148 /*
149 * For small buffers, we want a cache for each multiple of
150 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
151 * for each quarter-power of 2. For large buffers, we want
152 * a cache for each multiple of PAGESIZE.
153 */
154 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
155 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
156 size_t p2 = size;
157 size_t align = 0;
158
159 while (p2 & (p2 - 1))
160 p2 &= p2 - 1;
161
162 if (size <= 4 * SPA_MINBLOCKSIZE) {
163 align = SPA_MINBLOCKSIZE;
164 } else if (P2PHASE(size, PAGESIZE) == 0) {
165 align = PAGESIZE;
166 } else if (P2PHASE(size, p2 >> 2) == 0) {
167 align = p2 >> 2;
168 }
169
170 if (align != 0) {
171 char name[36];
172 int flags = zio_bulk_flags;
173
174 /*
175 * The smallest buffers (512b) are heavily used and
176 * experience a lot of churn. The slabs allocated
177 * for them are also relatively small (32K). Thus
178 * in over to avoid expensive calls to vmalloc() we
179 * make an exception to the usual slab allocation
180 * policy and force these buffers to be kmem backed.
181 */
182 if (size == (1 << SPA_MINBLOCKSHIFT))
183 flags |= KMC_KMEM;
184
185 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
186 zio_buf_cache[c] = kmem_cache_create(name, size,
187 align, NULL, NULL, NULL, NULL, NULL, flags);
188
189 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
190 zio_data_buf_cache[c] = kmem_cache_create(name, size,
191 align, NULL, NULL, NULL, NULL,
192 data_alloc_arena, flags);
193 }
194 }
195
196 while (--c != 0) {
197 ASSERT(zio_buf_cache[c] != NULL);
198 if (zio_buf_cache[c - 1] == NULL)
199 zio_buf_cache[c - 1] = zio_buf_cache[c];
200
201 ASSERT(zio_data_buf_cache[c] != NULL);
202 if (zio_data_buf_cache[c - 1] == NULL)
203 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
204 }
205
206 /*
207 * The zio write taskqs have 1 thread per cpu, allow 1/2 of the taskqs
208 * to fail 3 times per txg or 8 failures, whichever is greater.
209 */
210 zfs_mg_alloc_failures = MAX((3 * max_ncpus / 2), 8);
211
212 zio_inject_init();
213 }
214
215 void
216 zio_fini(void)
217 {
218 size_t c;
219 kmem_cache_t *last_cache = NULL;
220 kmem_cache_t *last_data_cache = NULL;
221
222 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
223 if (zio_buf_cache[c] != last_cache) {
224 last_cache = zio_buf_cache[c];
225 kmem_cache_destroy(zio_buf_cache[c]);
226 }
227 zio_buf_cache[c] = NULL;
228
229 if (zio_data_buf_cache[c] != last_data_cache) {
230 last_data_cache = zio_data_buf_cache[c];
231 kmem_cache_destroy(zio_data_buf_cache[c]);
232 }
233 zio_data_buf_cache[c] = NULL;
234 }
235
236 kmem_cache_destroy(zio_vdev_cache);
237 kmem_cache_destroy(zio_link_cache);
238 kmem_cache_destroy(zio_cache);
239
240 zio_inject_fini();
241 }
242
243 /*
244 * ==========================================================================
245 * Allocate and free I/O buffers
246 * ==========================================================================
247 */
248
249 /*
250 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
251 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
252 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
253 * excess / transient data in-core during a crashdump.
254 */
255 void *
256 zio_buf_alloc(size_t size)
257 {
258 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
259
260 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
261
262 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE | KM_NODEBUG));
263 }
264
265 /*
266 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
267 * crashdump if the kernel panics. This exists so that we will limit the amount
268 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
269 * of kernel heap dumped to disk when the kernel panics)
270 */
271 void *
272 zio_data_buf_alloc(size_t size)
273 {
274 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
275
276 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
277
278 return (kmem_cache_alloc(zio_data_buf_cache[c],
279 KM_PUSHPAGE | KM_NODEBUG));
280 }
281
282 void
283 zio_buf_free(void *buf, size_t size)
284 {
285 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
286
287 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
288
289 kmem_cache_free(zio_buf_cache[c], buf);
290 }
291
292 void
293 zio_data_buf_free(void *buf, size_t size)
294 {
295 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
296
297 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
298
299 kmem_cache_free(zio_data_buf_cache[c], buf);
300 }
301
302 /*
303 * Dedicated I/O buffers to ensure that memory fragmentation never prevents
304 * or significantly delays the issuing of a zio. These buffers are used
305 * to aggregate I/O and could be used for raidz stripes.
306 */
307 void *
308 zio_vdev_alloc(void)
309 {
310 return (kmem_cache_alloc(zio_vdev_cache, KM_PUSHPAGE));
311 }
312
313 void
314 zio_vdev_free(void *buf)
315 {
316 kmem_cache_free(zio_vdev_cache, buf);
317
318 }
319
320 /*
321 * ==========================================================================
322 * Push and pop I/O transform buffers
323 * ==========================================================================
324 */
325 static void
326 zio_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_PUSHPAGE);
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
343 static void
344 zio_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 */
369 static void
370 zio_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
378 static void
379 zio_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 = 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 */
400 zio_t *
401 zio_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
416 zio_t *
417 zio_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
432 zio_t *
433 zio_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
441 void
442 zio_add_child(zio_t *pio, zio_t *cio)
443 {
444 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_PUSHPAGE);
445 int w;
446
447 /*
448 * Logical I/Os can have logical, gang, or vdev children.
449 * Gang I/Os can have gang or vdev children.
450 * Vdev I/Os can only have vdev children.
451 * The following ASSERT captures all of these constraints.
452 */
453 ASSERT(cio->io_child_type <= pio->io_child_type);
454
455 zl->zl_parent = pio;
456 zl->zl_child = cio;
457
458 mutex_enter(&cio->io_lock);
459 mutex_enter(&pio->io_lock);
460
461 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
462
463 for (w = 0; w < ZIO_WAIT_TYPES; w++)
464 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
465
466 list_insert_head(&pio->io_child_list, zl);
467 list_insert_head(&cio->io_parent_list, zl);
468
469 pio->io_child_count++;
470 cio->io_parent_count++;
471
472 mutex_exit(&pio->io_lock);
473 mutex_exit(&cio->io_lock);
474 }
475
476 static void
477 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
478 {
479 ASSERT(zl->zl_parent == pio);
480 ASSERT(zl->zl_child == cio);
481
482 mutex_enter(&cio->io_lock);
483 mutex_enter(&pio->io_lock);
484
485 list_remove(&pio->io_child_list, zl);
486 list_remove(&cio->io_parent_list, zl);
487
488 pio->io_child_count--;
489 cio->io_parent_count--;
490
491 mutex_exit(&pio->io_lock);
492 mutex_exit(&cio->io_lock);
493
494 kmem_cache_free(zio_link_cache, zl);
495 }
496
497 static boolean_t
498 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
499 {
500 uint64_t *countp = &zio->io_children[child][wait];
501 boolean_t waiting = B_FALSE;
502
503 mutex_enter(&zio->io_lock);
504 ASSERT(zio->io_stall == NULL);
505 if (*countp != 0) {
506 zio->io_stage >>= 1;
507 zio->io_stall = countp;
508 waiting = B_TRUE;
509 }
510 mutex_exit(&zio->io_lock);
511
512 return (waiting);
513 }
514
515 __attribute__((always_inline))
516 static inline void
517 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
518 {
519 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
520 int *errorp = &pio->io_child_error[zio->io_child_type];
521
522 mutex_enter(&pio->io_lock);
523 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
524 *errorp = zio_worst_error(*errorp, zio->io_error);
525 pio->io_reexecute |= zio->io_reexecute;
526 ASSERT3U(*countp, >, 0);
527 if (--*countp == 0 && pio->io_stall == countp) {
528 pio->io_stall = NULL;
529 mutex_exit(&pio->io_lock);
530 __zio_execute(pio);
531 } else {
532 mutex_exit(&pio->io_lock);
533 }
534 }
535
536 static void
537 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
538 {
539 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
540 zio->io_error = zio->io_child_error[c];
541 }
542
543 /*
544 * ==========================================================================
545 * Create the various types of I/O (read, write, free, etc)
546 * ==========================================================================
547 */
548 static zio_t *
549 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
550 void *data, uint64_t size, zio_done_func_t *done, void *private,
551 zio_type_t type, int priority, enum zio_flag flags,
552 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
553 enum zio_stage stage, enum zio_stage pipeline)
554 {
555 zio_t *zio;
556
557 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
558 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
559 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
560
561 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
562 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
563 ASSERT(vd || stage == ZIO_STAGE_OPEN);
564
565 zio = kmem_cache_alloc(zio_cache, KM_PUSHPAGE);
566
567 if (vd != NULL)
568 zio->io_child_type = ZIO_CHILD_VDEV;
569 else if (flags & ZIO_FLAG_GANG_CHILD)
570 zio->io_child_type = ZIO_CHILD_GANG;
571 else if (flags & ZIO_FLAG_DDT_CHILD)
572 zio->io_child_type = ZIO_CHILD_DDT;
573 else
574 zio->io_child_type = ZIO_CHILD_LOGICAL;
575
576 if (bp != NULL) {
577 zio->io_logical = NULL;
578 zio->io_bp = (blkptr_t *)bp;
579 zio->io_bp_copy = *bp;
580 zio->io_bp_orig = *bp;
581 if (type != ZIO_TYPE_WRITE ||
582 zio->io_child_type == ZIO_CHILD_DDT)
583 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
584 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
585 zio->io_logical = zio;
586 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
587 pipeline |= ZIO_GANG_STAGES;
588 } else {
589 zio->io_logical = NULL;
590 zio->io_bp = NULL;
591 bzero(&zio->io_bp_copy, sizeof (blkptr_t));
592 bzero(&zio->io_bp_orig, sizeof (blkptr_t));
593 }
594
595 zio->io_spa = spa;
596 zio->io_txg = txg;
597 zio->io_ready = NULL;
598 zio->io_done = done;
599 zio->io_private = private;
600 zio->io_prev_space_delta = 0;
601 zio->io_type = type;
602 zio->io_priority = priority;
603 zio->io_vd = vd;
604 zio->io_vsd = NULL;
605 zio->io_vsd_ops = NULL;
606 zio->io_offset = offset;
607 zio->io_deadline = 0;
608 zio->io_orig_data = zio->io_data = data;
609 zio->io_orig_size = zio->io_size = size;
610 zio->io_orig_flags = zio->io_flags = flags;
611 zio->io_orig_stage = zio->io_stage = stage;
612 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
613 bzero(&zio->io_prop, sizeof (zio_prop_t));
614 zio->io_cmd = 0;
615 zio->io_reexecute = 0;
616 zio->io_bp_override = NULL;
617 zio->io_walk_link = NULL;
618 zio->io_transform_stack = NULL;
619 zio->io_delay = 0;
620 zio->io_error = 0;
621 zio->io_child_count = 0;
622 zio->io_parent_count = 0;
623 zio->io_stall = NULL;
624 zio->io_gang_leader = NULL;
625 zio->io_gang_tree = NULL;
626 zio->io_executor = NULL;
627 zio->io_waiter = NULL;
628 zio->io_cksum_report = NULL;
629 zio->io_ena = 0;
630 bzero(zio->io_child_error, sizeof (int) * ZIO_CHILD_TYPES);
631 bzero(zio->io_children,
632 sizeof (uint64_t) * ZIO_CHILD_TYPES * ZIO_WAIT_TYPES);
633 bzero(&zio->io_bookmark, sizeof (zbookmark_t));
634
635 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
636 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
637
638 if (zb != NULL)
639 zio->io_bookmark = *zb;
640
641 if (pio != NULL) {
642 if (zio->io_logical == NULL)
643 zio->io_logical = pio->io_logical;
644 if (zio->io_child_type == ZIO_CHILD_GANG)
645 zio->io_gang_leader = pio->io_gang_leader;
646 zio_add_child(pio, zio);
647 }
648
649 taskq_init_ent(&zio->io_tqent);
650
651 return (zio);
652 }
653
654 static void
655 zio_destroy(zio_t *zio)
656 {
657 kmem_cache_free(zio_cache, zio);
658 }
659
660 zio_t *
661 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
662 void *private, enum zio_flag flags)
663 {
664 zio_t *zio;
665
666 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
667 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
668 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
669
670 return (zio);
671 }
672
673 zio_t *
674 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
675 {
676 return (zio_null(NULL, spa, NULL, done, private, flags));
677 }
678
679 zio_t *
680 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
681 void *data, uint64_t size, zio_done_func_t *done, void *private,
682 int priority, enum zio_flag flags, const zbookmark_t *zb)
683 {
684 zio_t *zio;
685
686 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
687 data, size, done, private,
688 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
689 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
690 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
691
692 return (zio);
693 }
694
695 zio_t *
696 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
697 void *data, uint64_t size, const zio_prop_t *zp,
698 zio_done_func_t *ready, zio_done_func_t *done, void *private,
699 int priority, enum zio_flag flags, const zbookmark_t *zb)
700 {
701 zio_t *zio;
702
703 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
704 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
705 zp->zp_compress >= ZIO_COMPRESS_OFF &&
706 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
707 zp->zp_type < DMU_OT_NUMTYPES &&
708 zp->zp_level < 32 &&
709 zp->zp_copies > 0 &&
710 zp->zp_copies <= spa_max_replication(spa) &&
711 zp->zp_dedup <= 1 &&
712 zp->zp_dedup_verify <= 1);
713
714 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
715 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
716 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
717 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
718
719 zio->io_ready = ready;
720 zio->io_prop = *zp;
721
722 return (zio);
723 }
724
725 zio_t *
726 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
727 uint64_t size, zio_done_func_t *done, void *private, int priority,
728 enum zio_flag flags, zbookmark_t *zb)
729 {
730 zio_t *zio;
731
732 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
733 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
734 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
735
736 return (zio);
737 }
738
739 void
740 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
741 {
742 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
743 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
744 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
745 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
746
747 zio->io_prop.zp_copies = copies;
748 zio->io_bp_override = bp;
749 }
750
751 void
752 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
753 {
754 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
755 }
756
757 zio_t *
758 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
759 enum zio_flag flags)
760 {
761 zio_t *zio;
762
763 dprintf_bp(bp, "freeing in txg %llu, pass %u",
764 (longlong_t)txg, spa->spa_sync_pass);
765
766 ASSERT(!BP_IS_HOLE(bp));
767 ASSERT(spa_syncing_txg(spa) == txg);
768 ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
769
770 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
771 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
772 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
773
774 return (zio);
775 }
776
777 zio_t *
778 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
779 zio_done_func_t *done, void *private, enum zio_flag flags)
780 {
781 zio_t *zio;
782
783 /*
784 * A claim is an allocation of a specific block. Claims are needed
785 * to support immediate writes in the intent log. The issue is that
786 * immediate writes contain committed data, but in a txg that was
787 * *not* committed. Upon opening the pool after an unclean shutdown,
788 * the intent log claims all blocks that contain immediate write data
789 * so that the SPA knows they're in use.
790 *
791 * All claims *must* be resolved in the first txg -- before the SPA
792 * starts allocating blocks -- so that nothing is allocated twice.
793 * If txg == 0 we just verify that the block is claimable.
794 */
795 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
796 ASSERT(txg == spa_first_txg(spa) || txg == 0);
797 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
798
799 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
800 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
801 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
802
803 return (zio);
804 }
805
806 zio_t *
807 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
808 zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
809 {
810 zio_t *zio;
811 int c;
812
813 if (vd->vdev_children == 0) {
814 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
815 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
816 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
817
818 zio->io_cmd = cmd;
819 } else {
820 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
821
822 for (c = 0; c < vd->vdev_children; c++)
823 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
824 done, private, priority, flags));
825 }
826
827 return (zio);
828 }
829
830 zio_t *
831 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
832 void *data, int checksum, zio_done_func_t *done, void *private,
833 int priority, enum zio_flag flags, boolean_t labels)
834 {
835 zio_t *zio;
836
837 ASSERT(vd->vdev_children == 0);
838 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
839 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
840 ASSERT3U(offset + size, <=, vd->vdev_psize);
841
842 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
843 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
844 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
845
846 zio->io_prop.zp_checksum = checksum;
847
848 return (zio);
849 }
850
851 zio_t *
852 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
853 void *data, int checksum, zio_done_func_t *done, void *private,
854 int priority, enum zio_flag flags, boolean_t labels)
855 {
856 zio_t *zio;
857
858 ASSERT(vd->vdev_children == 0);
859 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
860 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
861 ASSERT3U(offset + size, <=, vd->vdev_psize);
862
863 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
864 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
865 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
866
867 zio->io_prop.zp_checksum = checksum;
868
869 if (zio_checksum_table[checksum].ci_eck) {
870 /*
871 * zec checksums are necessarily destructive -- they modify
872 * the end of the write buffer to hold the verifier/checksum.
873 * Therefore, we must make a local copy in case the data is
874 * being written to multiple places in parallel.
875 */
876 void *wbuf = zio_buf_alloc(size);
877 bcopy(data, wbuf, size);
878 zio_push_transform(zio, wbuf, size, size, NULL);
879 }
880
881 return (zio);
882 }
883
884 /*
885 * Create a child I/O to do some work for us.
886 */
887 zio_t *
888 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
889 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
890 zio_done_func_t *done, void *private)
891 {
892 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
893 zio_t *zio;
894
895 ASSERT(vd->vdev_parent ==
896 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
897
898 if (type == ZIO_TYPE_READ && bp != NULL) {
899 /*
900 * If we have the bp, then the child should perform the
901 * checksum and the parent need not. This pushes error
902 * detection as close to the leaves as possible and
903 * eliminates redundant checksums in the interior nodes.
904 */
905 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
906 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
907 }
908
909 if (vd->vdev_children == 0)
910 offset += VDEV_LABEL_START_SIZE;
911
912 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
913
914 /*
915 * If we've decided to do a repair, the write is not speculative --
916 * even if the original read was.
917 */
918 if (flags & ZIO_FLAG_IO_REPAIR)
919 flags &= ~ZIO_FLAG_SPECULATIVE;
920
921 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
922 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
923 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
924
925 return (zio);
926 }
927
928 zio_t *
929 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
930 int type, int priority, enum zio_flag flags,
931 zio_done_func_t *done, void *private)
932 {
933 zio_t *zio;
934
935 ASSERT(vd->vdev_ops->vdev_op_leaf);
936
937 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
938 data, size, done, private, type, priority,
939 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
940 vd, offset, NULL,
941 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
942
943 return (zio);
944 }
945
946 void
947 zio_flush(zio_t *zio, vdev_t *vd)
948 {
949 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
950 NULL, NULL, ZIO_PRIORITY_NOW,
951 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
952 }
953
954 void
955 zio_shrink(zio_t *zio, uint64_t size)
956 {
957 ASSERT(zio->io_executor == NULL);
958 ASSERT(zio->io_orig_size == zio->io_size);
959 ASSERT(size <= zio->io_size);
960
961 /*
962 * We don't shrink for raidz because of problems with the
963 * reconstruction when reading back less than the block size.
964 * Note, BP_IS_RAIDZ() assumes no compression.
965 */
966 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
967 if (!BP_IS_RAIDZ(zio->io_bp))
968 zio->io_orig_size = zio->io_size = size;
969 }
970
971 /*
972 * ==========================================================================
973 * Prepare to read and write logical blocks
974 * ==========================================================================
975 */
976
977 static int
978 zio_read_bp_init(zio_t *zio)
979 {
980 blkptr_t *bp = zio->io_bp;
981
982 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
983 zio->io_child_type == ZIO_CHILD_LOGICAL &&
984 !(zio->io_flags & ZIO_FLAG_RAW)) {
985 uint64_t psize = BP_GET_PSIZE(bp);
986 void *cbuf = zio_buf_alloc(psize);
987
988 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
989 }
990
991 if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
992 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
993
994 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
995 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
996
997 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
998 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
999
1000 return (ZIO_PIPELINE_CONTINUE);
1001 }
1002
1003 static int
1004 zio_write_bp_init(zio_t *zio)
1005 {
1006 spa_t *spa = zio->io_spa;
1007 zio_prop_t *zp = &zio->io_prop;
1008 enum zio_compress compress = zp->zp_compress;
1009 blkptr_t *bp = zio->io_bp;
1010 uint64_t lsize = zio->io_size;
1011 uint64_t psize = lsize;
1012 int pass = 1;
1013
1014 /*
1015 * If our children haven't all reached the ready stage,
1016 * wait for them and then repeat this pipeline stage.
1017 */
1018 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1019 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1020 return (ZIO_PIPELINE_STOP);
1021
1022 if (!IO_IS_ALLOCATING(zio))
1023 return (ZIO_PIPELINE_CONTINUE);
1024
1025 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1026
1027 if (zio->io_bp_override) {
1028 ASSERT(bp->blk_birth != zio->io_txg);
1029 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1030
1031 *bp = *zio->io_bp_override;
1032 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1033
1034 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1035 return (ZIO_PIPELINE_CONTINUE);
1036
1037 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1038 zp->zp_dedup_verify);
1039
1040 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1041 BP_SET_DEDUP(bp, 1);
1042 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1043 return (ZIO_PIPELINE_CONTINUE);
1044 }
1045 zio->io_bp_override = NULL;
1046 BP_ZERO(bp);
1047 }
1048
1049 if (bp->blk_birth == zio->io_txg) {
1050 /*
1051 * We're rewriting an existing block, which means we're
1052 * working on behalf of spa_sync(). For spa_sync() to
1053 * converge, it must eventually be the case that we don't
1054 * have to allocate new blocks. But compression changes
1055 * the blocksize, which forces a reallocate, and makes
1056 * convergence take longer. Therefore, after the first
1057 * few passes, stop compressing to ensure convergence.
1058 */
1059 pass = spa_sync_pass(spa);
1060
1061 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1062 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1063 ASSERT(!BP_GET_DEDUP(bp));
1064
1065 if (pass > SYNC_PASS_DONT_COMPRESS)
1066 compress = ZIO_COMPRESS_OFF;
1067
1068 /* Make sure someone doesn't change their mind on overwrites */
1069 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
1070 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1071 }
1072
1073 if (compress != ZIO_COMPRESS_OFF) {
1074 void *cbuf = zio_buf_alloc(lsize);
1075 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1076 if (psize == 0 || psize == lsize) {
1077 compress = ZIO_COMPRESS_OFF;
1078 zio_buf_free(cbuf, lsize);
1079 } else {
1080 ASSERT(psize < lsize);
1081 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1082 }
1083 }
1084
1085 /*
1086 * The final pass of spa_sync() must be all rewrites, but the first
1087 * few passes offer a trade-off: allocating blocks defers convergence,
1088 * but newly allocated blocks are sequential, so they can be written
1089 * to disk faster. Therefore, we allow the first few passes of
1090 * spa_sync() to allocate new blocks, but force rewrites after that.
1091 * There should only be a handful of blocks after pass 1 in any case.
1092 */
1093 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1094 pass > SYNC_PASS_REWRITE) {
1095 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1096 ASSERT(psize != 0);
1097 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1098 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1099 } else {
1100 BP_ZERO(bp);
1101 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1102 }
1103
1104 if (psize == 0) {
1105 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1106 } else {
1107 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1108 BP_SET_LSIZE(bp, lsize);
1109 BP_SET_PSIZE(bp, psize);
1110 BP_SET_COMPRESS(bp, compress);
1111 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1112 BP_SET_TYPE(bp, zp->zp_type);
1113 BP_SET_LEVEL(bp, zp->zp_level);
1114 BP_SET_DEDUP(bp, zp->zp_dedup);
1115 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1116 if (zp->zp_dedup) {
1117 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1118 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1119 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1120 }
1121 }
1122
1123 return (ZIO_PIPELINE_CONTINUE);
1124 }
1125
1126 static int
1127 zio_free_bp_init(zio_t *zio)
1128 {
1129 blkptr_t *bp = zio->io_bp;
1130
1131 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1132 if (BP_GET_DEDUP(bp))
1133 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1134 }
1135
1136 return (ZIO_PIPELINE_CONTINUE);
1137 }
1138
1139 /*
1140 * ==========================================================================
1141 * Execute the I/O pipeline
1142 * ==========================================================================
1143 */
1144
1145 static void
1146 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1147 {
1148 spa_t *spa = zio->io_spa;
1149 zio_type_t t = zio->io_type;
1150 int flags = (cutinline ? TQ_FRONT : 0);
1151
1152 /*
1153 * If we're a config writer or a probe, the normal issue and
1154 * interrupt threads may all be blocked waiting for the config lock.
1155 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1156 */
1157 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1158 t = ZIO_TYPE_NULL;
1159
1160 /*
1161 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1162 */
1163 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1164 t = ZIO_TYPE_NULL;
1165
1166 /*
1167 * If this is a high priority I/O, then use the high priority taskq.
1168 */
1169 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1170 spa->spa_zio_taskq[t][q + 1] != NULL)
1171 q++;
1172
1173 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1174
1175 /*
1176 * NB: We are assuming that the zio can only be dispatched
1177 * to a single taskq at a time. It would be a grievous error
1178 * to dispatch the zio to another taskq at the same time.
1179 */
1180 ASSERT(taskq_empty_ent(&zio->io_tqent));
1181 taskq_dispatch_ent(spa->spa_zio_taskq[t][q],
1182 (task_func_t *)zio_execute, zio, flags, &zio->io_tqent);
1183 }
1184
1185 static boolean_t
1186 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1187 {
1188 kthread_t *executor = zio->io_executor;
1189 spa_t *spa = zio->io_spa;
1190 zio_type_t t;
1191
1192 for (t = 0; t < ZIO_TYPES; t++)
1193 if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1194 return (B_TRUE);
1195
1196 return (B_FALSE);
1197 }
1198
1199 static int
1200 zio_issue_async(zio_t *zio)
1201 {
1202 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1203
1204 return (ZIO_PIPELINE_STOP);
1205 }
1206
1207 void
1208 zio_interrupt(zio_t *zio)
1209 {
1210 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1211 }
1212
1213 /*
1214 * Execute the I/O pipeline until one of the following occurs:
1215 * (1) the I/O completes; (2) the pipeline stalls waiting for
1216 * dependent child I/Os; (3) the I/O issues, so we're waiting
1217 * for an I/O completion interrupt; (4) the I/O is delegated by
1218 * vdev-level caching or aggregation; (5) the I/O is deferred
1219 * due to vdev-level queueing; (6) the I/O is handed off to
1220 * another thread. In all cases, the pipeline stops whenever
1221 * there's no CPU work; it never burns a thread in cv_wait().
1222 *
1223 * There's no locking on io_stage because there's no legitimate way
1224 * for multiple threads to be attempting to process the same I/O.
1225 */
1226 static zio_pipe_stage_t *zio_pipeline[];
1227
1228 /*
1229 * zio_execute() is a wrapper around the static function
1230 * __zio_execute() so that we can force __zio_execute() to be
1231 * inlined. This reduces stack overhead which is important
1232 * because __zio_execute() is called recursively in several zio
1233 * code paths. zio_execute() itself cannot be inlined because
1234 * it is externally visible.
1235 */
1236 void
1237 zio_execute(zio_t *zio)
1238 {
1239 __zio_execute(zio);
1240 }
1241
1242 __attribute__((always_inline))
1243 static inline void
1244 __zio_execute(zio_t *zio)
1245 {
1246 zio->io_executor = curthread;
1247
1248 while (zio->io_stage < ZIO_STAGE_DONE) {
1249 enum zio_stage pipeline = zio->io_pipeline;
1250 enum zio_stage stage = zio->io_stage;
1251 dsl_pool_t *dsl;
1252 boolean_t cut;
1253 int rv;
1254
1255 ASSERT(!MUTEX_HELD(&zio->io_lock));
1256 ASSERT(ISP2(stage));
1257 ASSERT(zio->io_stall == NULL);
1258
1259 do {
1260 stage <<= 1;
1261 } while ((stage & pipeline) == 0);
1262
1263 ASSERT(stage <= ZIO_STAGE_DONE);
1264
1265 dsl = spa_get_dsl(zio->io_spa);
1266 cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1267 zio_requeue_io_start_cut_in_line : B_FALSE;
1268
1269 /*
1270 * If we are in interrupt context and this pipeline stage
1271 * will grab a config lock that is held across I/O,
1272 * or may wait for an I/O that needs an interrupt thread
1273 * to complete, issue async to avoid deadlock.
1274 *
1275 * If we are in the txg_sync_thread or being called
1276 * during pool init issue async to minimize stack depth.
1277 * Both of these call paths may be recursively called.
1278 *
1279 * For VDEV_IO_START, we cut in line so that the io will
1280 * be sent to disk promptly.
1281 */
1282 if (((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1283 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) ||
1284 (dsl != NULL && dsl_pool_sync_context(dsl))) {
1285 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1286 return;
1287 }
1288
1289 zio->io_stage = stage;
1290 rv = zio_pipeline[highbit(stage) - 1](zio);
1291
1292 if (rv == ZIO_PIPELINE_STOP)
1293 return;
1294
1295 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1296 }
1297 }
1298
1299
1300 /*
1301 * ==========================================================================
1302 * Initiate I/O, either sync or async
1303 * ==========================================================================
1304 */
1305 int
1306 zio_wait(zio_t *zio)
1307 {
1308 int error;
1309
1310 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1311 ASSERT(zio->io_executor == NULL);
1312
1313 zio->io_waiter = curthread;
1314
1315 __zio_execute(zio);
1316
1317 mutex_enter(&zio->io_lock);
1318 while (zio->io_executor != NULL)
1319 cv_wait(&zio->io_cv, &zio->io_lock);
1320 mutex_exit(&zio->io_lock);
1321
1322 error = zio->io_error;
1323 zio_destroy(zio);
1324
1325 return (error);
1326 }
1327
1328 void
1329 zio_nowait(zio_t *zio)
1330 {
1331 ASSERT(zio->io_executor == NULL);
1332
1333 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1334 zio_unique_parent(zio) == NULL) {
1335 /*
1336 * This is a logical async I/O with no parent to wait for it.
1337 * We add it to the spa_async_root_zio "Godfather" I/O which
1338 * will ensure they complete prior to unloading the pool.
1339 */
1340 spa_t *spa = zio->io_spa;
1341
1342 zio_add_child(spa->spa_async_zio_root, zio);
1343 }
1344
1345 __zio_execute(zio);
1346 }
1347
1348 /*
1349 * ==========================================================================
1350 * Reexecute or suspend/resume failed I/O
1351 * ==========================================================================
1352 */
1353
1354 static void
1355 zio_reexecute(zio_t *pio)
1356 {
1357 zio_t *cio, *cio_next;
1358 int c, w;
1359
1360 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1361 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1362 ASSERT(pio->io_gang_leader == NULL);
1363 ASSERT(pio->io_gang_tree == NULL);
1364
1365 pio->io_flags = pio->io_orig_flags;
1366 pio->io_stage = pio->io_orig_stage;
1367 pio->io_pipeline = pio->io_orig_pipeline;
1368 pio->io_reexecute = 0;
1369 pio->io_error = 0;
1370 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1371 pio->io_state[w] = 0;
1372 for (c = 0; c < ZIO_CHILD_TYPES; c++)
1373 pio->io_child_error[c] = 0;
1374
1375 if (IO_IS_ALLOCATING(pio))
1376 BP_ZERO(pio->io_bp);
1377
1378 /*
1379 * As we reexecute pio's children, new children could be created.
1380 * New children go to the head of pio's io_child_list, however,
1381 * so we will (correctly) not reexecute them. The key is that
1382 * the remainder of pio's io_child_list, from 'cio_next' onward,
1383 * cannot be affected by any side effects of reexecuting 'cio'.
1384 */
1385 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1386 cio_next = zio_walk_children(pio);
1387 mutex_enter(&pio->io_lock);
1388 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1389 pio->io_children[cio->io_child_type][w]++;
1390 mutex_exit(&pio->io_lock);
1391 zio_reexecute(cio);
1392 }
1393
1394 /*
1395 * Now that all children have been reexecuted, execute the parent.
1396 * We don't reexecute "The Godfather" I/O here as it's the
1397 * responsibility of the caller to wait on him.
1398 */
1399 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1400 __zio_execute(pio);
1401 }
1402
1403 void
1404 zio_suspend(spa_t *spa, zio_t *zio)
1405 {
1406 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1407 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1408 "failure and the failure mode property for this pool "
1409 "is set to panic.", spa_name(spa));
1410
1411 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1412
1413 mutex_enter(&spa->spa_suspend_lock);
1414
1415 if (spa->spa_suspend_zio_root == NULL)
1416 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1417 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1418 ZIO_FLAG_GODFATHER);
1419
1420 spa->spa_suspended = B_TRUE;
1421
1422 if (zio != NULL) {
1423 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1424 ASSERT(zio != spa->spa_suspend_zio_root);
1425 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1426 ASSERT(zio_unique_parent(zio) == NULL);
1427 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1428 zio_add_child(spa->spa_suspend_zio_root, zio);
1429 }
1430
1431 mutex_exit(&spa->spa_suspend_lock);
1432 }
1433
1434 int
1435 zio_resume(spa_t *spa)
1436 {
1437 zio_t *pio;
1438
1439 /*
1440 * Reexecute all previously suspended i/o.
1441 */
1442 mutex_enter(&spa->spa_suspend_lock);
1443 spa->spa_suspended = B_FALSE;
1444 cv_broadcast(&spa->spa_suspend_cv);
1445 pio = spa->spa_suspend_zio_root;
1446 spa->spa_suspend_zio_root = NULL;
1447 mutex_exit(&spa->spa_suspend_lock);
1448
1449 if (pio == NULL)
1450 return (0);
1451
1452 zio_reexecute(pio);
1453 return (zio_wait(pio));
1454 }
1455
1456 void
1457 zio_resume_wait(spa_t *spa)
1458 {
1459 mutex_enter(&spa->spa_suspend_lock);
1460 while (spa_suspended(spa))
1461 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1462 mutex_exit(&spa->spa_suspend_lock);
1463 }
1464
1465 /*
1466 * ==========================================================================
1467 * Gang blocks.
1468 *
1469 * A gang block is a collection of small blocks that looks to the DMU
1470 * like one large block. When zio_dva_allocate() cannot find a block
1471 * of the requested size, due to either severe fragmentation or the pool
1472 * being nearly full, it calls zio_write_gang_block() to construct the
1473 * block from smaller fragments.
1474 *
1475 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1476 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1477 * an indirect block: it's an array of block pointers. It consumes
1478 * only one sector and hence is allocatable regardless of fragmentation.
1479 * The gang header's bps point to its gang members, which hold the data.
1480 *
1481 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1482 * as the verifier to ensure uniqueness of the SHA256 checksum.
1483 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1484 * not the gang header. This ensures that data block signatures (needed for
1485 * deduplication) are independent of how the block is physically stored.
1486 *
1487 * Gang blocks can be nested: a gang member may itself be a gang block.
1488 * Thus every gang block is a tree in which root and all interior nodes are
1489 * gang headers, and the leaves are normal blocks that contain user data.
1490 * The root of the gang tree is called the gang leader.
1491 *
1492 * To perform any operation (read, rewrite, free, claim) on a gang block,
1493 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1494 * in the io_gang_tree field of the original logical i/o by recursively
1495 * reading the gang leader and all gang headers below it. This yields
1496 * an in-core tree containing the contents of every gang header and the
1497 * bps for every constituent of the gang block.
1498 *
1499 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1500 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1501 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1502 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1503 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1504 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1505 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1506 * of the gang header plus zio_checksum_compute() of the data to update the
1507 * gang header's blk_cksum as described above.
1508 *
1509 * The two-phase assemble/issue model solves the problem of partial failure --
1510 * what if you'd freed part of a gang block but then couldn't read the
1511 * gang header for another part? Assembling the entire gang tree first
1512 * ensures that all the necessary gang header I/O has succeeded before
1513 * starting the actual work of free, claim, or write. Once the gang tree
1514 * is assembled, free and claim are in-memory operations that cannot fail.
1515 *
1516 * In the event that a gang write fails, zio_dva_unallocate() walks the
1517 * gang tree to immediately free (i.e. insert back into the space map)
1518 * everything we've allocated. This ensures that we don't get ENOSPC
1519 * errors during repeated suspend/resume cycles due to a flaky device.
1520 *
1521 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1522 * the gang tree, we won't modify the block, so we can safely defer the free
1523 * (knowing that the block is still intact). If we *can* assemble the gang
1524 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1525 * each constituent bp and we can allocate a new block on the next sync pass.
1526 *
1527 * In all cases, the gang tree allows complete recovery from partial failure.
1528 * ==========================================================================
1529 */
1530
1531 static zio_t *
1532 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1533 {
1534 if (gn != NULL)
1535 return (pio);
1536
1537 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1538 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1539 &pio->io_bookmark));
1540 }
1541
1542 zio_t *
1543 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1544 {
1545 zio_t *zio;
1546
1547 if (gn != NULL) {
1548 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1549 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1550 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1551 /*
1552 * As we rewrite each gang header, the pipeline will compute
1553 * a new gang block header checksum for it; but no one will
1554 * compute a new data checksum, so we do that here. The one
1555 * exception is the gang leader: the pipeline already computed
1556 * its data checksum because that stage precedes gang assembly.
1557 * (Presently, nothing actually uses interior data checksums;
1558 * this is just good hygiene.)
1559 */
1560 if (gn != pio->io_gang_leader->io_gang_tree) {
1561 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1562 data, BP_GET_PSIZE(bp));
1563 }
1564 /*
1565 * If we are here to damage data for testing purposes,
1566 * leave the GBH alone so that we can detect the damage.
1567 */
1568 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1569 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1570 } else {
1571 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1572 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1573 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1574 }
1575
1576 return (zio);
1577 }
1578
1579 /* ARGSUSED */
1580 zio_t *
1581 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1582 {
1583 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1584 ZIO_GANG_CHILD_FLAGS(pio)));
1585 }
1586
1587 /* ARGSUSED */
1588 zio_t *
1589 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1590 {
1591 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1592 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1593 }
1594
1595 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1596 NULL,
1597 zio_read_gang,
1598 zio_rewrite_gang,
1599 zio_free_gang,
1600 zio_claim_gang,
1601 NULL
1602 };
1603
1604 static void zio_gang_tree_assemble_done(zio_t *zio);
1605
1606 static zio_gang_node_t *
1607 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1608 {
1609 zio_gang_node_t *gn;
1610
1611 ASSERT(*gnpp == NULL);
1612
1613 gn = kmem_zalloc(sizeof (*gn), KM_PUSHPAGE);
1614 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1615 *gnpp = gn;
1616
1617 return (gn);
1618 }
1619
1620 static void
1621 zio_gang_node_free(zio_gang_node_t **gnpp)
1622 {
1623 zio_gang_node_t *gn = *gnpp;
1624 int g;
1625
1626 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1627 ASSERT(gn->gn_child[g] == NULL);
1628
1629 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1630 kmem_free(gn, sizeof (*gn));
1631 *gnpp = NULL;
1632 }
1633
1634 static void
1635 zio_gang_tree_free(zio_gang_node_t **gnpp)
1636 {
1637 zio_gang_node_t *gn = *gnpp;
1638 int g;
1639
1640 if (gn == NULL)
1641 return;
1642
1643 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1644 zio_gang_tree_free(&gn->gn_child[g]);
1645
1646 zio_gang_node_free(gnpp);
1647 }
1648
1649 static void
1650 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1651 {
1652 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1653
1654 ASSERT(gio->io_gang_leader == gio);
1655 ASSERT(BP_IS_GANG(bp));
1656
1657 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1658 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1659 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1660 }
1661
1662 static void
1663 zio_gang_tree_assemble_done(zio_t *zio)
1664 {
1665 zio_t *gio = zio->io_gang_leader;
1666 zio_gang_node_t *gn = zio->io_private;
1667 blkptr_t *bp = zio->io_bp;
1668 int g;
1669
1670 ASSERT(gio == zio_unique_parent(zio));
1671 ASSERT(zio->io_child_count == 0);
1672
1673 if (zio->io_error)
1674 return;
1675
1676 if (BP_SHOULD_BYTESWAP(bp))
1677 byteswap_uint64_array(zio->io_data, zio->io_size);
1678
1679 ASSERT(zio->io_data == gn->gn_gbh);
1680 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1681 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1682
1683 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1684 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1685 if (!BP_IS_GANG(gbp))
1686 continue;
1687 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1688 }
1689 }
1690
1691 static void
1692 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1693 {
1694 zio_t *gio = pio->io_gang_leader;
1695 zio_t *zio;
1696 int g;
1697
1698 ASSERT(BP_IS_GANG(bp) == !!gn);
1699 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1700 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1701
1702 /*
1703 * If you're a gang header, your data is in gn->gn_gbh.
1704 * If you're a gang member, your data is in 'data' and gn == NULL.
1705 */
1706 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1707
1708 if (gn != NULL) {
1709 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1710
1711 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1712 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1713 if (BP_IS_HOLE(gbp))
1714 continue;
1715 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1716 data = (char *)data + BP_GET_PSIZE(gbp);
1717 }
1718 }
1719
1720 if (gn == gio->io_gang_tree)
1721 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1722
1723 if (zio != pio)
1724 zio_nowait(zio);
1725 }
1726
1727 static int
1728 zio_gang_assemble(zio_t *zio)
1729 {
1730 blkptr_t *bp = zio->io_bp;
1731
1732 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1733 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1734
1735 zio->io_gang_leader = zio;
1736
1737 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1738
1739 return (ZIO_PIPELINE_CONTINUE);
1740 }
1741
1742 static int
1743 zio_gang_issue(zio_t *zio)
1744 {
1745 blkptr_t *bp = zio->io_bp;
1746
1747 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1748 return (ZIO_PIPELINE_STOP);
1749
1750 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1751 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1752
1753 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1754 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1755 else
1756 zio_gang_tree_free(&zio->io_gang_tree);
1757
1758 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1759
1760 return (ZIO_PIPELINE_CONTINUE);
1761 }
1762
1763 static void
1764 zio_write_gang_member_ready(zio_t *zio)
1765 {
1766 zio_t *pio = zio_unique_parent(zio);
1767 ASSERTV(zio_t *gio = zio->io_gang_leader;)
1768 dva_t *cdva = zio->io_bp->blk_dva;
1769 dva_t *pdva = pio->io_bp->blk_dva;
1770 uint64_t asize;
1771 int d;
1772
1773 if (BP_IS_HOLE(zio->io_bp))
1774 return;
1775
1776 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1777
1778 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1779 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1780 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1781 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1782 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1783
1784 mutex_enter(&pio->io_lock);
1785 for (d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1786 ASSERT(DVA_GET_GANG(&pdva[d]));
1787 asize = DVA_GET_ASIZE(&pdva[d]);
1788 asize += DVA_GET_ASIZE(&cdva[d]);
1789 DVA_SET_ASIZE(&pdva[d], asize);
1790 }
1791 mutex_exit(&pio->io_lock);
1792 }
1793
1794 static int
1795 zio_write_gang_block(zio_t *pio)
1796 {
1797 spa_t *spa = pio->io_spa;
1798 blkptr_t *bp = pio->io_bp;
1799 zio_t *gio = pio->io_gang_leader;
1800 zio_t *zio;
1801 zio_gang_node_t *gn, **gnpp;
1802 zio_gbh_phys_t *gbh;
1803 uint64_t txg = pio->io_txg;
1804 uint64_t resid = pio->io_size;
1805 uint64_t lsize;
1806 int copies = gio->io_prop.zp_copies;
1807 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1808 zio_prop_t zp;
1809 int g, error;
1810
1811 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1812 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1813 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1814 if (error) {
1815 pio->io_error = error;
1816 return (ZIO_PIPELINE_CONTINUE);
1817 }
1818
1819 if (pio == gio) {
1820 gnpp = &gio->io_gang_tree;
1821 } else {
1822 gnpp = pio->io_private;
1823 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1824 }
1825
1826 gn = zio_gang_node_alloc(gnpp);
1827 gbh = gn->gn_gbh;
1828 bzero(gbh, SPA_GANGBLOCKSIZE);
1829
1830 /*
1831 * Create the gang header.
1832 */
1833 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1834 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1835
1836 /*
1837 * Create and nowait the gang children.
1838 */
1839 for (g = 0; resid != 0; resid -= lsize, g++) {
1840 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1841 SPA_MINBLOCKSIZE);
1842 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1843
1844 zp.zp_checksum = gio->io_prop.zp_checksum;
1845 zp.zp_compress = ZIO_COMPRESS_OFF;
1846 zp.zp_type = DMU_OT_NONE;
1847 zp.zp_level = 0;
1848 zp.zp_copies = gio->io_prop.zp_copies;
1849 zp.zp_dedup = 0;
1850 zp.zp_dedup_verify = 0;
1851
1852 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1853 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1854 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1855 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1856 &pio->io_bookmark));
1857 }
1858
1859 /*
1860 * Set pio's pipeline to just wait for zio to finish.
1861 */
1862 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1863
1864 /*
1865 * We didn't allocate this bp, so make sure it doesn't get unmarked.
1866 */
1867 pio->io_flags &= ~ZIO_FLAG_FASTWRITE;
1868
1869 zio_nowait(zio);
1870
1871 return (ZIO_PIPELINE_CONTINUE);
1872 }
1873
1874 /*
1875 * ==========================================================================
1876 * Dedup
1877 * ==========================================================================
1878 */
1879 static void
1880 zio_ddt_child_read_done(zio_t *zio)
1881 {
1882 blkptr_t *bp = zio->io_bp;
1883 ddt_entry_t *dde = zio->io_private;
1884 ddt_phys_t *ddp;
1885 zio_t *pio = zio_unique_parent(zio);
1886
1887 mutex_enter(&pio->io_lock);
1888 ddp = ddt_phys_select(dde, bp);
1889 if (zio->io_error == 0)
1890 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1891 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1892 dde->dde_repair_data = zio->io_data;
1893 else
1894 zio_buf_free(zio->io_data, zio->io_size);
1895 mutex_exit(&pio->io_lock);
1896 }
1897
1898 static int
1899 zio_ddt_read_start(zio_t *zio)
1900 {
1901 blkptr_t *bp = zio->io_bp;
1902 int p;
1903
1904 ASSERT(BP_GET_DEDUP(bp));
1905 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1906 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1907
1908 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1909 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1910 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1911 ddt_phys_t *ddp = dde->dde_phys;
1912 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1913 blkptr_t blk;
1914
1915 ASSERT(zio->io_vsd == NULL);
1916 zio->io_vsd = dde;
1917
1918 if (ddp_self == NULL)
1919 return (ZIO_PIPELINE_CONTINUE);
1920
1921 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1922 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1923 continue;
1924 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1925 &blk);
1926 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1927 zio_buf_alloc(zio->io_size), zio->io_size,
1928 zio_ddt_child_read_done, dde, zio->io_priority,
1929 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1930 &zio->io_bookmark));
1931 }
1932 return (ZIO_PIPELINE_CONTINUE);
1933 }
1934
1935 zio_nowait(zio_read(zio, zio->io_spa, bp,
1936 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1937 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1938
1939 return (ZIO_PIPELINE_CONTINUE);
1940 }
1941
1942 static int
1943 zio_ddt_read_done(zio_t *zio)
1944 {
1945 blkptr_t *bp = zio->io_bp;
1946
1947 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1948 return (ZIO_PIPELINE_STOP);
1949
1950 ASSERT(BP_GET_DEDUP(bp));
1951 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1952 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1953
1954 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1955 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1956 ddt_entry_t *dde = zio->io_vsd;
1957 if (ddt == NULL) {
1958 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1959 return (ZIO_PIPELINE_CONTINUE);
1960 }
1961 if (dde == NULL) {
1962 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1963 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1964 return (ZIO_PIPELINE_STOP);
1965 }
1966 if (dde->dde_repair_data != NULL) {
1967 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1968 zio->io_child_error[ZIO_CHILD_DDT] = 0;
1969 }
1970 ddt_repair_done(ddt, dde);
1971 zio->io_vsd = NULL;
1972 }
1973
1974 ASSERT(zio->io_vsd == NULL);
1975
1976 return (ZIO_PIPELINE_CONTINUE);
1977 }
1978
1979 static boolean_t
1980 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1981 {
1982 spa_t *spa = zio->io_spa;
1983 int p;
1984
1985 /*
1986 * Note: we compare the original data, not the transformed data,
1987 * because when zio->io_bp is an override bp, we will not have
1988 * pushed the I/O transforms. That's an important optimization
1989 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1990 */
1991 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1992 zio_t *lio = dde->dde_lead_zio[p];
1993
1994 if (lio != NULL) {
1995 return (lio->io_orig_size != zio->io_orig_size ||
1996 bcmp(zio->io_orig_data, lio->io_orig_data,
1997 zio->io_orig_size) != 0);
1998 }
1999 }
2000
2001 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2002 ddt_phys_t *ddp = &dde->dde_phys[p];
2003
2004 if (ddp->ddp_phys_birth != 0) {
2005 arc_buf_t *abuf = NULL;
2006 uint32_t aflags = ARC_WAIT;
2007 blkptr_t blk = *zio->io_bp;
2008 int error;
2009
2010 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2011
2012 ddt_exit(ddt);
2013
2014 error = arc_read_nolock(NULL, spa, &blk,
2015 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2016 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2017 &aflags, &zio->io_bookmark);
2018
2019 if (error == 0) {
2020 if (arc_buf_size(abuf) != zio->io_orig_size ||
2021 bcmp(abuf->b_data, zio->io_orig_data,
2022 zio->io_orig_size) != 0)
2023 error = EEXIST;
2024 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
2025 }
2026
2027 ddt_enter(ddt);
2028 return (error != 0);
2029 }
2030 }
2031
2032 return (B_FALSE);
2033 }
2034
2035 static void
2036 zio_ddt_child_write_ready(zio_t *zio)
2037 {
2038 int p = zio->io_prop.zp_copies;
2039 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2040 ddt_entry_t *dde = zio->io_private;
2041 ddt_phys_t *ddp = &dde->dde_phys[p];
2042 zio_t *pio;
2043
2044 if (zio->io_error)
2045 return;
2046
2047 ddt_enter(ddt);
2048
2049 ASSERT(dde->dde_lead_zio[p] == zio);
2050
2051 ddt_phys_fill(ddp, zio->io_bp);
2052
2053 while ((pio = zio_walk_parents(zio)) != NULL)
2054 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2055
2056 ddt_exit(ddt);
2057 }
2058
2059 static void
2060 zio_ddt_child_write_done(zio_t *zio)
2061 {
2062 int p = zio->io_prop.zp_copies;
2063 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2064 ddt_entry_t *dde = zio->io_private;
2065 ddt_phys_t *ddp = &dde->dde_phys[p];
2066
2067 ddt_enter(ddt);
2068
2069 ASSERT(ddp->ddp_refcnt == 0);
2070 ASSERT(dde->dde_lead_zio[p] == zio);
2071 dde->dde_lead_zio[p] = NULL;
2072
2073 if (zio->io_error == 0) {
2074 while (zio_walk_parents(zio) != NULL)
2075 ddt_phys_addref(ddp);
2076 } else {
2077 ddt_phys_clear(ddp);
2078 }
2079
2080 ddt_exit(ddt);
2081 }
2082
2083 static void
2084 zio_ddt_ditto_write_done(zio_t *zio)
2085 {
2086 int p = DDT_PHYS_DITTO;
2087 blkptr_t *bp = zio->io_bp;
2088 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2089 ddt_entry_t *dde = zio->io_private;
2090 ddt_phys_t *ddp = &dde->dde_phys[p];
2091 ddt_key_t *ddk = &dde->dde_key;
2092 ASSERTV(zio_prop_t *zp = &zio->io_prop);
2093
2094 ddt_enter(ddt);
2095
2096 ASSERT(ddp->ddp_refcnt == 0);
2097 ASSERT(dde->dde_lead_zio[p] == zio);
2098 dde->dde_lead_zio[p] = NULL;
2099
2100 if (zio->io_error == 0) {
2101 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2102 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2103 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2104 if (ddp->ddp_phys_birth != 0)
2105 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2106 ddt_phys_fill(ddp, bp);
2107 }
2108
2109 ddt_exit(ddt);
2110 }
2111
2112 static int
2113 zio_ddt_write(zio_t *zio)
2114 {
2115 spa_t *spa = zio->io_spa;
2116 blkptr_t *bp = zio->io_bp;
2117 uint64_t txg = zio->io_txg;
2118 zio_prop_t *zp = &zio->io_prop;
2119 int p = zp->zp_copies;
2120 int ditto_copies;
2121 zio_t *cio = NULL;
2122 zio_t *dio = NULL;
2123 ddt_t *ddt = ddt_select(spa, bp);
2124 ddt_entry_t *dde;
2125 ddt_phys_t *ddp;
2126
2127 ASSERT(BP_GET_DEDUP(bp));
2128 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2129 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2130
2131 ddt_enter(ddt);
2132 dde = ddt_lookup(ddt, bp, B_TRUE);
2133 ddp = &dde->dde_phys[p];
2134
2135 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2136 /*
2137 * If we're using a weak checksum, upgrade to a strong checksum
2138 * and try again. If we're already using a strong checksum,
2139 * we can't resolve it, so just convert to an ordinary write.
2140 * (And automatically e-mail a paper to Nature?)
2141 */
2142 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2143 zp->zp_checksum = spa_dedup_checksum(spa);
2144 zio_pop_transforms(zio);
2145 zio->io_stage = ZIO_STAGE_OPEN;
2146 BP_ZERO(bp);
2147 } else {
2148 zp->zp_dedup = 0;
2149 }
2150 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2151 ddt_exit(ddt);
2152 return (ZIO_PIPELINE_CONTINUE);
2153 }
2154
2155 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2156 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2157
2158 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2159 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2160 zio_prop_t czp = *zp;
2161
2162 czp.zp_copies = ditto_copies;
2163
2164 /*
2165 * If we arrived here with an override bp, we won't have run
2166 * the transform stack, so we won't have the data we need to
2167 * generate a child i/o. So, toss the override bp and restart.
2168 * This is safe, because using the override bp is just an
2169 * optimization; and it's rare, so the cost doesn't matter.
2170 */
2171 if (zio->io_bp_override) {
2172 zio_pop_transforms(zio);
2173 zio->io_stage = ZIO_STAGE_OPEN;
2174 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2175 zio->io_bp_override = NULL;
2176 BP_ZERO(bp);
2177 ddt_exit(ddt);
2178 return (ZIO_PIPELINE_CONTINUE);
2179 }
2180
2181 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2182 zio->io_orig_size, &czp, NULL,
2183 zio_ddt_ditto_write_done, dde, zio->io_priority,
2184 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2185
2186 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2187 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2188 }
2189
2190 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2191 if (ddp->ddp_phys_birth != 0)
2192 ddt_bp_fill(ddp, bp, txg);
2193 if (dde->dde_lead_zio[p] != NULL)
2194 zio_add_child(zio, dde->dde_lead_zio[p]);
2195 else
2196 ddt_phys_addref(ddp);
2197 } else if (zio->io_bp_override) {
2198 ASSERT(bp->blk_birth == txg);
2199 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2200 ddt_phys_fill(ddp, bp);
2201 ddt_phys_addref(ddp);
2202 } else {
2203 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2204 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2205 zio_ddt_child_write_done, dde, zio->io_priority,
2206 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2207
2208 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2209 dde->dde_lead_zio[p] = cio;
2210 }
2211
2212 ddt_exit(ddt);
2213
2214 if (cio)
2215 zio_nowait(cio);
2216 if (dio)
2217 zio_nowait(dio);
2218
2219 return (ZIO_PIPELINE_CONTINUE);
2220 }
2221
2222 ddt_entry_t *freedde; /* for debugging */
2223
2224 static int
2225 zio_ddt_free(zio_t *zio)
2226 {
2227 spa_t *spa = zio->io_spa;
2228 blkptr_t *bp = zio->io_bp;
2229 ddt_t *ddt = ddt_select(spa, bp);
2230 ddt_entry_t *dde;
2231 ddt_phys_t *ddp;
2232
2233 ASSERT(BP_GET_DEDUP(bp));
2234 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2235
2236 ddt_enter(ddt);
2237 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2238 ddp = ddt_phys_select(dde, bp);
2239 ddt_phys_decref(ddp);
2240 ddt_exit(ddt);
2241
2242 return (ZIO_PIPELINE_CONTINUE);
2243 }
2244
2245 /*
2246 * ==========================================================================
2247 * Allocate and free blocks
2248 * ==========================================================================
2249 */
2250 static int
2251 zio_dva_allocate(zio_t *zio)
2252 {
2253 spa_t *spa = zio->io_spa;
2254 metaslab_class_t *mc = spa_normal_class(spa);
2255 blkptr_t *bp = zio->io_bp;
2256 int error;
2257 int flags = 0;
2258
2259 if (zio->io_gang_leader == NULL) {
2260 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2261 zio->io_gang_leader = zio;
2262 }
2263
2264 ASSERT(BP_IS_HOLE(bp));
2265 ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
2266 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2267 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2268 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2269
2270 /*
2271 * The dump device does not support gang blocks so allocation on
2272 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2273 * the "fast" gang feature.
2274 */
2275 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2276 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2277 METASLAB_GANG_CHILD : 0;
2278 flags |= (zio->io_flags & ZIO_FLAG_FASTWRITE) ? METASLAB_FASTWRITE : 0;
2279 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2280 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2281
2282 if (error) {
2283 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2284 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2285 error);
2286 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2287 return (zio_write_gang_block(zio));
2288 zio->io_error = error;
2289 }
2290
2291 return (ZIO_PIPELINE_CONTINUE);
2292 }
2293
2294 static int
2295 zio_dva_free(zio_t *zio)
2296 {
2297 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2298
2299 return (ZIO_PIPELINE_CONTINUE);
2300 }
2301
2302 static int
2303 zio_dva_claim(zio_t *zio)
2304 {
2305 int error;
2306
2307 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2308 if (error)
2309 zio->io_error = error;
2310
2311 return (ZIO_PIPELINE_CONTINUE);
2312 }
2313
2314 /*
2315 * Undo an allocation. This is used by zio_done() when an I/O fails
2316 * and we want to give back the block we just allocated.
2317 * This handles both normal blocks and gang blocks.
2318 */
2319 static void
2320 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2321 {
2322 int g;
2323
2324 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2325 ASSERT(zio->io_bp_override == NULL);
2326
2327 if (!BP_IS_HOLE(bp))
2328 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2329
2330 if (gn != NULL) {
2331 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2332 zio_dva_unallocate(zio, gn->gn_child[g],
2333 &gn->gn_gbh->zg_blkptr[g]);
2334 }
2335 }
2336 }
2337
2338 /*
2339 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2340 */
2341 int
2342 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, uint64_t size,
2343 boolean_t use_slog)
2344 {
2345 int error = 1;
2346
2347 ASSERT(txg > spa_syncing_txg(spa));
2348
2349 /*
2350 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2351 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2352 * when allocating them.
2353 */
2354 if (use_slog) {
2355 error = metaslab_alloc(spa, spa_log_class(spa), size,
2356 new_bp, 1, txg, NULL,
2357 METASLAB_FASTWRITE | METASLAB_GANG_AVOID);
2358 }
2359
2360 if (error) {
2361 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2362 new_bp, 1, txg, NULL,
2363 METASLAB_FASTWRITE | METASLAB_GANG_AVOID);
2364 }
2365
2366 if (error == 0) {
2367 BP_SET_LSIZE(new_bp, size);
2368 BP_SET_PSIZE(new_bp, size);
2369 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2370 BP_SET_CHECKSUM(new_bp,
2371 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2372 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2373 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2374 BP_SET_LEVEL(new_bp, 0);
2375 BP_SET_DEDUP(new_bp, 0);
2376 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2377 }
2378
2379 return (error);
2380 }
2381
2382 /*
2383 * Free an intent log block.
2384 */
2385 void
2386 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2387 {
2388 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2389 ASSERT(!BP_IS_GANG(bp));
2390
2391 zio_free(spa, txg, bp);
2392 }
2393
2394 /*
2395 * ==========================================================================
2396 * Read and write to physical devices
2397 * ==========================================================================
2398 */
2399 static int
2400 zio_vdev_io_start(zio_t *zio)
2401 {
2402 vdev_t *vd = zio->io_vd;
2403 uint64_t align;
2404 spa_t *spa = zio->io_spa;
2405
2406 ASSERT(zio->io_error == 0);
2407 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2408
2409 if (vd == NULL) {
2410 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2411 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2412
2413 /*
2414 * The mirror_ops handle multiple DVAs in a single BP.
2415 */
2416 return (vdev_mirror_ops.vdev_op_io_start(zio));
2417 }
2418
2419 /*
2420 * We keep track of time-sensitive I/Os so that the scan thread
2421 * can quickly react to certain workloads. In particular, we care
2422 * about non-scrubbing, top-level reads and writes with the following
2423 * characteristics:
2424 * - synchronous writes of user data to non-slog devices
2425 * - any reads of user data
2426 * When these conditions are met, adjust the timestamp of spa_last_io
2427 * which allows the scan thread to adjust its workload accordingly.
2428 */
2429 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2430 vd == vd->vdev_top && !vd->vdev_islog &&
2431 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2432 zio->io_txg != spa_syncing_txg(spa)) {
2433 uint64_t old = spa->spa_last_io;
2434 uint64_t new = ddi_get_lbolt64();
2435 if (old != new)
2436 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2437 }
2438
2439 align = 1ULL << vd->vdev_top->vdev_ashift;
2440
2441 /*
2442 * On Linux, we don't care about read alignment. The backing block
2443 * device driver will take care of that for us.
2444 * The only exception is raidz, which needs a full block for parity.
2445 */
2446 if (P2PHASE(zio->io_size, align) != 0 &&
2447 (zio->io_type != ZIO_TYPE_READ ||
2448 vd->vdev_ops == &vdev_raidz_ops)) {
2449 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2450 char *abuf = zio_buf_alloc(asize);
2451 if (zio->io_type == ZIO_TYPE_WRITE) {
2452 bcopy(zio->io_data, abuf, zio->io_size);
2453 bzero(abuf + zio->io_size, asize - zio->io_size);
2454 }
2455 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2456 ASSERT(P2PHASE(zio->io_size, align) == 0);
2457 }
2458
2459 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2460 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2461
2462 /*
2463 * If this is a repair I/O, and there's no self-healing involved --
2464 * that is, we're just resilvering what we expect to resilver --
2465 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2466 * This prevents spurious resilvering with nested replication.
2467 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2468 * A is out of date, we'll read from C+D, then use the data to
2469 * resilver A+B -- but we don't actually want to resilver B, just A.
2470 * The top-level mirror has no way to know this, so instead we just
2471 * discard unnecessary repairs as we work our way down the vdev tree.
2472 * The same logic applies to any form of nested replication:
2473 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2474 */
2475 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2476 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2477 zio->io_txg != 0 && /* not a delegated i/o */
2478 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2479 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2480 zio_vdev_io_bypass(zio);
2481 return (ZIO_PIPELINE_CONTINUE);
2482 }
2483
2484 if (vd->vdev_ops->vdev_op_leaf &&
2485 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2486
2487 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2488 return (ZIO_PIPELINE_CONTINUE);
2489
2490 if ((zio = vdev_queue_io(zio)) == NULL)
2491 return (ZIO_PIPELINE_STOP);
2492
2493 if (!vdev_accessible(vd, zio)) {
2494 zio->io_error = ENXIO;
2495 zio_interrupt(zio);
2496 return (ZIO_PIPELINE_STOP);
2497 }
2498 }
2499
2500 return (vd->vdev_ops->vdev_op_io_start(zio));
2501 }
2502
2503 static int
2504 zio_vdev_io_done(zio_t *zio)
2505 {
2506 vdev_t *vd = zio->io_vd;
2507 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2508 boolean_t unexpected_error = B_FALSE;
2509
2510 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2511 return (ZIO_PIPELINE_STOP);
2512
2513 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2514
2515 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2516
2517 vdev_queue_io_done(zio);
2518
2519 if (zio->io_type == ZIO_TYPE_WRITE)
2520 vdev_cache_write(zio);
2521
2522 if (zio_injection_enabled && zio->io_error == 0)
2523 zio->io_error = zio_handle_device_injection(vd,
2524 zio, EIO);
2525
2526 if (zio_injection_enabled && zio->io_error == 0)
2527 zio->io_error = zio_handle_label_injection(zio, EIO);
2528
2529 if (zio->io_error) {
2530 if (!vdev_accessible(vd, zio)) {
2531 zio->io_error = ENXIO;
2532 } else {
2533 unexpected_error = B_TRUE;
2534 }
2535 }
2536 }
2537
2538 ops->vdev_op_io_done(zio);
2539
2540 if (unexpected_error)
2541 VERIFY(vdev_probe(vd, zio) == NULL);
2542
2543 return (ZIO_PIPELINE_CONTINUE);
2544 }
2545
2546 /*
2547 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2548 * disk, and use that to finish the checksum ereport later.
2549 */
2550 static void
2551 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2552 const void *good_buf)
2553 {
2554 /* no processing needed */
2555 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2556 }
2557
2558 /*ARGSUSED*/
2559 void
2560 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2561 {
2562 void *buf = zio_buf_alloc(zio->io_size);
2563
2564 bcopy(zio->io_data, buf, zio->io_size);
2565
2566 zcr->zcr_cbinfo = zio->io_size;
2567 zcr->zcr_cbdata = buf;
2568 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2569 zcr->zcr_free = zio_buf_free;
2570 }
2571
2572 static int
2573 zio_vdev_io_assess(zio_t *zio)
2574 {
2575 vdev_t *vd = zio->io_vd;
2576
2577 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2578 return (ZIO_PIPELINE_STOP);
2579
2580 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2581 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2582
2583 if (zio->io_vsd != NULL) {
2584 zio->io_vsd_ops->vsd_free(zio);
2585 zio->io_vsd = NULL;
2586 }
2587
2588 if (zio_injection_enabled && zio->io_error == 0)
2589 zio->io_error = zio_handle_fault_injection(zio, EIO);
2590
2591 /*
2592 * If the I/O failed, determine whether we should attempt to retry it.
2593 *
2594 * On retry, we cut in line in the issue queue, since we don't want
2595 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2596 */
2597 if (zio->io_error && vd == NULL &&
2598 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2599 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2600 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2601 zio->io_error = 0;
2602 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2603 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2604 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2605 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2606 zio_requeue_io_start_cut_in_line);
2607 return (ZIO_PIPELINE_STOP);
2608 }
2609
2610 /*
2611 * If we got an error on a leaf device, convert it to ENXIO
2612 * if the device is not accessible at all.
2613 */
2614 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2615 !vdev_accessible(vd, zio))
2616 zio->io_error = ENXIO;
2617
2618 /*
2619 * If we can't write to an interior vdev (mirror or RAID-Z),
2620 * set vdev_cant_write so that we stop trying to allocate from it.
2621 */
2622 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2623 vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2624 vd->vdev_cant_write = B_TRUE;
2625
2626 if (zio->io_error)
2627 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2628
2629 return (ZIO_PIPELINE_CONTINUE);
2630 }
2631
2632 void
2633 zio_vdev_io_reissue(zio_t *zio)
2634 {
2635 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2636 ASSERT(zio->io_error == 0);
2637
2638 zio->io_stage >>= 1;
2639 }
2640
2641 void
2642 zio_vdev_io_redone(zio_t *zio)
2643 {
2644 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2645
2646 zio->io_stage >>= 1;
2647 }
2648
2649 void
2650 zio_vdev_io_bypass(zio_t *zio)
2651 {
2652 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2653 ASSERT(zio->io_error == 0);
2654
2655 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2656 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2657 }
2658
2659 /*
2660 * ==========================================================================
2661 * Generate and verify checksums
2662 * ==========================================================================
2663 */
2664 static int
2665 zio_checksum_generate(zio_t *zio)
2666 {
2667 blkptr_t *bp = zio->io_bp;
2668 enum zio_checksum checksum;
2669
2670 if (bp == NULL) {
2671 /*
2672 * This is zio_write_phys().
2673 * We're either generating a label checksum, or none at all.
2674 */
2675 checksum = zio->io_prop.zp_checksum;
2676
2677 if (checksum == ZIO_CHECKSUM_OFF)
2678 return (ZIO_PIPELINE_CONTINUE);
2679
2680 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2681 } else {
2682 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2683 ASSERT(!IO_IS_ALLOCATING(zio));
2684 checksum = ZIO_CHECKSUM_GANG_HEADER;
2685 } else {
2686 checksum = BP_GET_CHECKSUM(bp);
2687 }
2688 }
2689
2690 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2691
2692 return (ZIO_PIPELINE_CONTINUE);
2693 }
2694
2695 static int
2696 zio_checksum_verify(zio_t *zio)
2697 {
2698 zio_bad_cksum_t info;
2699 blkptr_t *bp = zio->io_bp;
2700 int error;
2701
2702 ASSERT(zio->io_vd != NULL);
2703
2704 if (bp == NULL) {
2705 /*
2706 * This is zio_read_phys().
2707 * We're either verifying a label checksum, or nothing at all.
2708 */
2709 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2710 return (ZIO_PIPELINE_CONTINUE);
2711
2712 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2713 }
2714
2715 if ((error = zio_checksum_error(zio, &info)) != 0) {
2716 zio->io_error = error;
2717 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2718 zfs_ereport_start_checksum(zio->io_spa,
2719 zio->io_vd, zio, zio->io_offset,
2720 zio->io_size, NULL, &info);
2721 }
2722 }
2723
2724 return (ZIO_PIPELINE_CONTINUE);
2725 }
2726
2727 /*
2728 * Called by RAID-Z to ensure we don't compute the checksum twice.
2729 */
2730 void
2731 zio_checksum_verified(zio_t *zio)
2732 {
2733 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2734 }
2735
2736 /*
2737 * ==========================================================================
2738 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2739 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2740 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2741 * indicate errors that are specific to one I/O, and most likely permanent.
2742 * Any other error is presumed to be worse because we weren't expecting it.
2743 * ==========================================================================
2744 */
2745 int
2746 zio_worst_error(int e1, int e2)
2747 {
2748 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2749 int r1, r2;
2750
2751 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2752 if (e1 == zio_error_rank[r1])
2753 break;
2754
2755 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2756 if (e2 == zio_error_rank[r2])
2757 break;
2758
2759 return (r1 > r2 ? e1 : e2);
2760 }
2761
2762 /*
2763 * ==========================================================================
2764 * I/O completion
2765 * ==========================================================================
2766 */
2767 static int
2768 zio_ready(zio_t *zio)
2769 {
2770 blkptr_t *bp = zio->io_bp;
2771 zio_t *pio, *pio_next;
2772
2773 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2774 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2775 return (ZIO_PIPELINE_STOP);
2776
2777 if (zio->io_ready) {
2778 ASSERT(IO_IS_ALLOCATING(zio));
2779 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2780 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2781
2782 zio->io_ready(zio);
2783 }
2784
2785 if (bp != NULL && bp != &zio->io_bp_copy)
2786 zio->io_bp_copy = *bp;
2787
2788 if (zio->io_error)
2789 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2790
2791 mutex_enter(&zio->io_lock);
2792 zio->io_state[ZIO_WAIT_READY] = 1;
2793 pio = zio_walk_parents(zio);
2794 mutex_exit(&zio->io_lock);
2795
2796 /*
2797 * As we notify zio's parents, new parents could be added.
2798 * New parents go to the head of zio's io_parent_list, however,
2799 * so we will (correctly) not notify them. The remainder of zio's
2800 * io_parent_list, from 'pio_next' onward, cannot change because
2801 * all parents must wait for us to be done before they can be done.
2802 */
2803 for (; pio != NULL; pio = pio_next) {
2804 pio_next = zio_walk_parents(zio);
2805 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2806 }
2807
2808 if (zio->io_flags & ZIO_FLAG_NODATA) {
2809 if (BP_IS_GANG(bp)) {
2810 zio->io_flags &= ~ZIO_FLAG_NODATA;
2811 } else {
2812 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2813 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2814 }
2815 }
2816
2817 if (zio_injection_enabled &&
2818 zio->io_spa->spa_syncing_txg == zio->io_txg)
2819 zio_handle_ignored_writes(zio);
2820
2821 return (ZIO_PIPELINE_CONTINUE);
2822 }
2823
2824 static int
2825 zio_done(zio_t *zio)
2826 {
2827 zio_t *pio, *pio_next;
2828 int c, w;
2829
2830 /*
2831 * If our children haven't all completed,
2832 * wait for them and then repeat this pipeline stage.
2833 */
2834 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2835 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2836 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2837 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2838 return (ZIO_PIPELINE_STOP);
2839
2840 for (c = 0; c < ZIO_CHILD_TYPES; c++)
2841 for (w = 0; w < ZIO_WAIT_TYPES; w++)
2842 ASSERT(zio->io_children[c][w] == 0);
2843
2844 if (zio->io_bp != NULL) {
2845 ASSERT(zio->io_bp->blk_pad[0] == 0);
2846 ASSERT(zio->io_bp->blk_pad[1] == 0);
2847 ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2848 (zio->io_bp == zio_unique_parent(zio)->io_bp));
2849 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) &&
2850 zio->io_bp_override == NULL &&
2851 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2852 ASSERT(!BP_SHOULD_BYTESWAP(zio->io_bp));
2853 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2854 ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 ||
2855 (BP_COUNT_GANG(zio->io_bp) == BP_GET_NDVAS(zio->io_bp)));
2856 }
2857 }
2858
2859 /*
2860 * If there were child vdev/gang/ddt errors, they apply to us now.
2861 */
2862 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2863 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2864 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2865
2866 /*
2867 * If the I/O on the transformed data was successful, generate any
2868 * checksum reports now while we still have the transformed data.
2869 */
2870 if (zio->io_error == 0) {
2871 while (zio->io_cksum_report != NULL) {
2872 zio_cksum_report_t *zcr = zio->io_cksum_report;
2873 uint64_t align = zcr->zcr_align;
2874 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2875 char *abuf = zio->io_data;
2876
2877 if (asize != zio->io_size) {
2878 abuf = zio_buf_alloc(asize);
2879 bcopy(zio->io_data, abuf, zio->io_size);
2880 bzero(abuf + zio->io_size, asize - zio->io_size);
2881 }
2882
2883 zio->io_cksum_report = zcr->zcr_next;
2884 zcr->zcr_next = NULL;
2885 zcr->zcr_finish(zcr, abuf);
2886 zfs_ereport_free_checksum(zcr);
2887
2888 if (asize != zio->io_size)
2889 zio_buf_free(abuf, asize);
2890 }
2891 }
2892
2893 zio_pop_transforms(zio); /* note: may set zio->io_error */
2894
2895 vdev_stat_update(zio, zio->io_size);
2896
2897 /*
2898 * If this I/O is attached to a particular vdev is slow, exeeding
2899 * 30 seconds to complete, post an error described the I/O delay.
2900 * We ignore these errors if the device is currently unavailable.
2901 */
2902 if (zio->io_delay >= zio_delay_max) {
2903 if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd))
2904 zfs_ereport_post(FM_EREPORT_ZFS_DELAY, zio->io_spa,
2905 zio->io_vd, zio, 0, 0);
2906 }
2907
2908 if (zio->io_error) {
2909 /*
2910 * If this I/O is attached to a particular vdev,
2911 * generate an error message describing the I/O failure
2912 * at the block level. We ignore these errors if the
2913 * device is currently unavailable.
2914 */
2915 if (zio->io_error != ECKSUM && zio->io_vd != NULL &&
2916 !vdev_is_dead(zio->io_vd))
2917 zfs_ereport_post(FM_EREPORT_ZFS_IO, zio->io_spa,
2918 zio->io_vd, zio, 0, 0);
2919
2920 if ((zio->io_error == EIO || !(zio->io_flags &
2921 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2922 zio == zio->io_logical) {
2923 /*
2924 * For logical I/O requests, tell the SPA to log the
2925 * error and generate a logical data ereport.
2926 */
2927 spa_log_error(zio->io_spa, zio);
2928 zfs_ereport_post(FM_EREPORT_ZFS_DATA, zio->io_spa, NULL, zio,
2929 0, 0);
2930 }
2931 }
2932
2933 if (zio->io_error && zio == zio->io_logical) {
2934 /*
2935 * Determine whether zio should be reexecuted. This will
2936 * propagate all the way to the root via zio_notify_parent().
2937 */
2938 ASSERT(zio->io_vd == NULL && zio->io_bp != NULL);
2939 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2940
2941 if (IO_IS_ALLOCATING(zio) &&
2942 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2943 if (zio->io_error != ENOSPC)
2944 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2945 else
2946 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2947 }
2948
2949 if ((zio->io_type == ZIO_TYPE_READ ||
2950 zio->io_type == ZIO_TYPE_FREE) &&
2951 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2952 zio->io_error == ENXIO &&
2953 spa_load_state(zio->io_spa) == SPA_LOAD_NONE &&
2954 spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE)
2955 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2956
2957 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2958 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2959
2960 /*
2961 * Here is a possibly good place to attempt to do
2962 * either combinatorial reconstruction or error correction
2963 * based on checksums. It also might be a good place
2964 * to send out preliminary ereports before we suspend
2965 * processing.
2966 */
2967 }
2968
2969 /*
2970 * If there were logical child errors, they apply to us now.
2971 * We defer this until now to avoid conflating logical child
2972 * errors with errors that happened to the zio itself when
2973 * updating vdev stats and reporting FMA events above.
2974 */
2975 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2976
2977 if ((zio->io_error || zio->io_reexecute) &&
2978 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2979 !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2980 zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp);
2981
2982 zio_gang_tree_free(&zio->io_gang_tree);
2983
2984 /*
2985 * Godfather I/Os should never suspend.
2986 */
2987 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2988 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2989 zio->io_reexecute = 0;
2990
2991 if (zio->io_reexecute) {
2992 /*
2993 * This is a logical I/O that wants to reexecute.
2994 *
2995 * Reexecute is top-down. When an i/o fails, if it's not
2996 * the root, it simply notifies its parent and sticks around.
2997 * The parent, seeing that it still has children in zio_done(),
2998 * does the same. This percolates all the way up to the root.
2999 * The root i/o will reexecute or suspend the entire tree.
3000 *
3001 * This approach ensures that zio_reexecute() honors
3002 * all the original i/o dependency relationships, e.g.
3003 * parents not executing until children are ready.
3004 */
3005 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3006
3007 zio->io_gang_leader = NULL;
3008
3009 mutex_enter(&zio->io_lock);
3010 zio->io_state[ZIO_WAIT_DONE] = 1;
3011 mutex_exit(&zio->io_lock);
3012
3013 /*
3014 * "The Godfather" I/O monitors its children but is
3015 * not a true parent to them. It will track them through
3016 * the pipeline but severs its ties whenever they get into
3017 * trouble (e.g. suspended). This allows "The Godfather"
3018 * I/O to return status without blocking.
3019 */
3020 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3021 zio_link_t *zl = zio->io_walk_link;
3022 pio_next = zio_walk_parents(zio);
3023
3024 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3025 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3026 zio_remove_child(pio, zio, zl);
3027 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3028 }
3029 }
3030
3031 if ((pio = zio_unique_parent(zio)) != NULL) {
3032 /*
3033 * We're not a root i/o, so there's nothing to do
3034 * but notify our parent. Don't propagate errors
3035 * upward since we haven't permanently failed yet.
3036 */
3037 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3038 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3039 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3040 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3041 /*
3042 * We'd fail again if we reexecuted now, so suspend
3043 * until conditions improve (e.g. device comes online).
3044 */
3045 zio_suspend(zio->io_spa, zio);
3046 } else {
3047 /*
3048 * Reexecution is potentially a huge amount of work.
3049 * Hand it off to the otherwise-unused claim taskq.
3050 */
3051 ASSERT(taskq_empty_ent(&zio->io_tqent));
3052 (void) taskq_dispatch_ent(
3053 zio->io_spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
3054 (task_func_t *)zio_reexecute, zio, 0,
3055 &zio->io_tqent);
3056 }
3057 return (ZIO_PIPELINE_STOP);
3058 }
3059
3060 ASSERT(zio->io_child_count == 0);
3061 ASSERT(zio->io_reexecute == 0);
3062 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3063
3064 /*
3065 * Report any checksum errors, since the I/O is complete.
3066 */
3067 while (zio->io_cksum_report != NULL) {
3068 zio_cksum_report_t *zcr = zio->io_cksum_report;
3069 zio->io_cksum_report = zcr->zcr_next;
3070 zcr->zcr_next = NULL;
3071 zcr->zcr_finish(zcr, NULL);
3072 zfs_ereport_free_checksum(zcr);
3073 }
3074
3075 if (zio->io_flags & ZIO_FLAG_FASTWRITE && zio->io_bp &&
3076 !BP_IS_HOLE(zio->io_bp)) {
3077 metaslab_fastwrite_unmark(zio->io_spa, zio->io_bp);
3078 }
3079
3080 /*
3081 * It is the responsibility of the done callback to ensure that this
3082 * particular zio is no longer discoverable for adoption, and as
3083 * such, cannot acquire any new parents.
3084 */
3085 if (zio->io_done)
3086 zio->io_done(zio);
3087
3088 mutex_enter(&zio->io_lock);
3089 zio->io_state[ZIO_WAIT_DONE] = 1;
3090 mutex_exit(&zio->io_lock);
3091
3092 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3093 zio_link_t *zl = zio->io_walk_link;
3094 pio_next = zio_walk_parents(zio);
3095 zio_remove_child(pio, zio, zl);
3096 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3097 }
3098
3099 if (zio->io_waiter != NULL) {
3100 mutex_enter(&zio->io_lock);
3101 zio->io_executor = NULL;
3102 cv_broadcast(&zio->io_cv);
3103 mutex_exit(&zio->io_lock);
3104 } else {
3105 zio_destroy(zio);
3106 }
3107
3108 return (ZIO_PIPELINE_STOP);
3109 }
3110
3111 /*
3112 * ==========================================================================
3113 * I/O pipeline definition
3114 * ==========================================================================
3115 */
3116 static zio_pipe_stage_t *zio_pipeline[] = {
3117 NULL,
3118 zio_read_bp_init,
3119 zio_free_bp_init,
3120 zio_issue_async,
3121 zio_write_bp_init,
3122 zio_checksum_generate,
3123 zio_ddt_read_start,
3124 zio_ddt_read_done,
3125 zio_ddt_write,
3126 zio_ddt_free,
3127 zio_gang_assemble,
3128 zio_gang_issue,
3129 zio_dva_allocate,
3130 zio_dva_free,
3131 zio_dva_claim,
3132 zio_ready,
3133 zio_vdev_io_start,
3134 zio_vdev_io_done,
3135 zio_vdev_io_assess,
3136 zio_checksum_verify,
3137 zio_done
3138 };
3139
3140 #if defined(_KERNEL) && defined(HAVE_SPL)
3141 /* Fault injection */
3142 EXPORT_SYMBOL(zio_injection_enabled);
3143 EXPORT_SYMBOL(zio_inject_fault);
3144 EXPORT_SYMBOL(zio_inject_list_next);
3145 EXPORT_SYMBOL(zio_clear_fault);
3146 EXPORT_SYMBOL(zio_handle_fault_injection);
3147 EXPORT_SYMBOL(zio_handle_device_injection);
3148 EXPORT_SYMBOL(zio_handle_label_injection);
3149 EXPORT_SYMBOL(zio_priority_table);
3150 EXPORT_SYMBOL(zio_type_name);
3151
3152 module_param(zio_bulk_flags, int, 0644);
3153 MODULE_PARM_DESC(zio_bulk_flags, "Additional flags to pass to bulk buffers");
3154
3155 module_param(zio_delay_max, int, 0644);
3156 MODULE_PARM_DESC(zio_delay_max, "Max zio millisec delay before posting event");
3157
3158 module_param(zio_requeue_io_start_cut_in_line, int, 0644);
3159 MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line, "Prioritize requeued I/O");
3160 #endif
Mirror of ZFSOnLinux for Debian