1/*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21/*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25 * Copyright (c) 2014 Integros [integros.com]
26 */
27
28#include <sys/sysmacros.h>
29#include <sys/zfs_context.h>
30#include <sys/fm/fs/zfs.h>
31#include <sys/spa.h>
32#include <sys/txg.h>
33#include <sys/spa_impl.h>
34#include <sys/vdev_impl.h>
35#include <sys/zio_impl.h>
36#include <sys/zio_compress.h>
37#include <sys/zio_checksum.h>
38#include <sys/dmu_objset.h>
39#include <sys/arc.h>
40#include <sys/ddt.h>
41#include <sys/trim_map.h>
42#include <sys/blkptr.h>
43#include <sys/zfeature.h>
| 1/*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21/*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25 * Copyright (c) 2014 Integros [integros.com]
26 */
27
28#include <sys/sysmacros.h>
29#include <sys/zfs_context.h>
30#include <sys/fm/fs/zfs.h>
31#include <sys/spa.h>
32#include <sys/txg.h>
33#include <sys/spa_impl.h>
34#include <sys/vdev_impl.h>
35#include <sys/zio_impl.h>
36#include <sys/zio_compress.h>
37#include <sys/zio_checksum.h>
38#include <sys/dmu_objset.h>
39#include <sys/arc.h>
40#include <sys/ddt.h>
41#include <sys/trim_map.h>
42#include <sys/blkptr.h>
43#include <sys/zfeature.h>
|
| 44#include <sys/dsl_scan.h>
|
44#include <sys/metaslab_impl.h>
45#include <sys/abd.h>
46
47SYSCTL_DECL(_vfs_zfs);
48SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO");
49#if defined(__amd64__)
50static int zio_use_uma = 1;
51#else
52static int zio_use_uma = 0;
53#endif
54SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
55 "Use uma(9) for ZIO allocations");
56static int zio_exclude_metadata = 0;
57SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, exclude_metadata, CTLFLAG_RDTUN, &zio_exclude_metadata, 0,
58 "Exclude metadata buffers from dumps as well");
59
60zio_trim_stats_t zio_trim_stats = {
61 { "bytes", KSTAT_DATA_UINT64,
62 "Number of bytes successfully TRIMmed" },
63 { "success", KSTAT_DATA_UINT64,
64 "Number of successful TRIM requests" },
65 { "unsupported", KSTAT_DATA_UINT64,
66 "Number of TRIM requests that failed because TRIM is not supported" },
67 { "failed", KSTAT_DATA_UINT64,
68 "Number of TRIM requests that failed for reasons other than not supported" },
69};
70
71static kstat_t *zio_trim_ksp;
72
73/*
74 * ==========================================================================
75 * I/O type descriptions
76 * ==========================================================================
77 */
78const char *zio_type_name[ZIO_TYPES] = {
79 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
80 "zio_ioctl"
81};
82
83boolean_t zio_dva_throttle_enabled = B_TRUE;
84SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, dva_throttle_enabled, CTLFLAG_RDTUN,
85 &zio_dva_throttle_enabled, 0, "");
86
87/*
88 * ==========================================================================
89 * I/O kmem caches
90 * ==========================================================================
91 */
92kmem_cache_t *zio_cache;
93kmem_cache_t *zio_link_cache;
94kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
95kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
96
97#ifdef _KERNEL
98extern vmem_t *zio_alloc_arena;
99#endif
100
101#define ZIO_PIPELINE_CONTINUE 0x100
102#define ZIO_PIPELINE_STOP 0x101
103
104#define BP_SPANB(indblkshift, level) \
105 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
106#define COMPARE_META_LEVEL 0x80000000ul
107/*
108 * The following actions directly effect the spa's sync-to-convergence logic.
109 * The values below define the sync pass when we start performing the action.
110 * Care should be taken when changing these values as they directly impact
111 * spa_sync() performance. Tuning these values may introduce subtle performance
112 * pathologies and should only be done in the context of performance analysis.
113 * These tunables will eventually be removed and replaced with #defines once
114 * enough analysis has been done to determine optimal values.
115 *
116 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
117 * regular blocks are not deferred.
118 */
119int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
120SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN,
121 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass");
122int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
123SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN,
124 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass");
125int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
126SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN,
127 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass");
128
129/*
130 * An allocating zio is one that either currently has the DVA allocate
131 * stage set or will have it later in its lifetime.
132 */
133#define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
134
135boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
136
137#ifdef illumos
138#ifdef ZFS_DEBUG
139int zio_buf_debug_limit = 16384;
140#else
141int zio_buf_debug_limit = 0;
142#endif
143#endif
144
145static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
146
147void
148zio_init(void)
149{
150 size_t c;
151 zio_cache = kmem_cache_create("zio_cache",
152 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
153 zio_link_cache = kmem_cache_create("zio_link_cache",
154 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
155 if (!zio_use_uma)
156 goto out;
157
158 /*
159 * For small buffers, we want a cache for each multiple of
160 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
161 * for each quarter-power of 2.
162 */
163 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
164 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
165 size_t p2 = size;
166 size_t align = 0;
167 int cflags = zio_exclude_metadata ? KMC_NODEBUG : 0;
168
169 while (!ISP2(p2))
170 p2 &= p2 - 1;
171
172#ifdef illumos
173#ifndef _KERNEL
174 /*
175 * If we are using watchpoints, put each buffer on its own page,
176 * to eliminate the performance overhead of trapping to the
177 * kernel when modifying a non-watched buffer that shares the
178 * page with a watched buffer.
179 */
180 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
181 continue;
182#endif
183#endif /* illumos */
184 if (size <= 4 * SPA_MINBLOCKSIZE) {
185 align = SPA_MINBLOCKSIZE;
186 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
187 align = MIN(p2 >> 2, PAGESIZE);
188 }
189
190 if (align != 0) {
191 char name[36];
192 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
193 zio_buf_cache[c] = kmem_cache_create(name, size,
194 align, NULL, NULL, NULL, NULL, NULL, cflags);
195
196 /*
197 * Since zio_data bufs do not appear in crash dumps, we
198 * pass KMC_NOTOUCH so that no allocator metadata is
199 * stored with the buffers.
200 */
201 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
202 zio_data_buf_cache[c] = kmem_cache_create(name, size,
203 align, NULL, NULL, NULL, NULL, NULL,
204 cflags | KMC_NOTOUCH | KMC_NODEBUG);
205 }
206 }
207
208 while (--c != 0) {
209 ASSERT(zio_buf_cache[c] != NULL);
210 if (zio_buf_cache[c - 1] == NULL)
211 zio_buf_cache[c - 1] = zio_buf_cache[c];
212
213 ASSERT(zio_data_buf_cache[c] != NULL);
214 if (zio_data_buf_cache[c - 1] == NULL)
215 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
216 }
217out:
218
219 zio_inject_init();
220
221 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc",
222 KSTAT_TYPE_NAMED,
223 sizeof(zio_trim_stats) / sizeof(kstat_named_t),
224 KSTAT_FLAG_VIRTUAL);
225
226 if (zio_trim_ksp != NULL) {
227 zio_trim_ksp->ks_data = &zio_trim_stats;
228 kstat_install(zio_trim_ksp);
229 }
230}
231
232void
233zio_fini(void)
234{
235 size_t c;
236 kmem_cache_t *last_cache = NULL;
237 kmem_cache_t *last_data_cache = NULL;
238
239 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
240 if (zio_buf_cache[c] != last_cache) {
241 last_cache = zio_buf_cache[c];
242 kmem_cache_destroy(zio_buf_cache[c]);
243 }
244 zio_buf_cache[c] = NULL;
245
246 if (zio_data_buf_cache[c] != last_data_cache) {
247 last_data_cache = zio_data_buf_cache[c];
248 kmem_cache_destroy(zio_data_buf_cache[c]);
249 }
250 zio_data_buf_cache[c] = NULL;
251 }
252
253 kmem_cache_destroy(zio_link_cache);
254 kmem_cache_destroy(zio_cache);
255
256 zio_inject_fini();
257
258 if (zio_trim_ksp != NULL) {
259 kstat_delete(zio_trim_ksp);
260 zio_trim_ksp = NULL;
261 }
262}
263
264/*
265 * ==========================================================================
266 * Allocate and free I/O buffers
267 * ==========================================================================
268 */
269
270/*
271 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
272 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
273 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
274 * excess / transient data in-core during a crashdump.
275 */
276void *
277zio_buf_alloc(size_t size)
278{
279 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
280 int flags = zio_exclude_metadata ? KM_NODEBUG : 0;
281
282 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
283
284 if (zio_use_uma)
285 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
286 else
287 return (kmem_alloc(size, KM_SLEEP|flags));
288}
289
290/*
291 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
292 * crashdump if the kernel panics. This exists so that we will limit the amount
293 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
294 * of kernel heap dumped to disk when the kernel panics)
295 */
296void *
297zio_data_buf_alloc(size_t size)
298{
299 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
300
301 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
302
303 if (zio_use_uma)
304 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
305 else
306 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
307}
308
309void
310zio_buf_free(void *buf, size_t size)
311{
312 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
313
314 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
315
316 if (zio_use_uma)
317 kmem_cache_free(zio_buf_cache[c], buf);
318 else
319 kmem_free(buf, size);
320}
321
322void
323zio_data_buf_free(void *buf, size_t size)
324{
325 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
326
327 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
328
329 if (zio_use_uma)
330 kmem_cache_free(zio_data_buf_cache[c], buf);
331 else
332 kmem_free(buf, size);
333}
334
335/*
336 * ==========================================================================
337 * Push and pop I/O transform buffers
338 * ==========================================================================
339 */
340void
341zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize,
342 zio_transform_func_t *transform)
343{
344 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
345
346 /*
347 * Ensure that anyone expecting this zio to contain a linear ABD isn't
348 * going to get a nasty surprise when they try to access the data.
349 */
350#ifdef illumos
351 IMPLY(abd_is_linear(zio->io_abd), abd_is_linear(data));
352#else
353 IMPLY(zio->io_abd != NULL && abd_is_linear(zio->io_abd),
354 abd_is_linear(data));
355#endif
356
357 zt->zt_orig_abd = zio->io_abd;
358 zt->zt_orig_size = zio->io_size;
359 zt->zt_bufsize = bufsize;
360 zt->zt_transform = transform;
361
362 zt->zt_next = zio->io_transform_stack;
363 zio->io_transform_stack = zt;
364
365 zio->io_abd = data;
366 zio->io_size = size;
367}
368
369void
370zio_pop_transforms(zio_t *zio)
371{
372 zio_transform_t *zt;
373
374 while ((zt = zio->io_transform_stack) != NULL) {
375 if (zt->zt_transform != NULL)
376 zt->zt_transform(zio,
377 zt->zt_orig_abd, zt->zt_orig_size);
378
379 if (zt->zt_bufsize != 0)
380 abd_free(zio->io_abd);
381
382 zio->io_abd = zt->zt_orig_abd;
383 zio->io_size = zt->zt_orig_size;
384 zio->io_transform_stack = zt->zt_next;
385
386 kmem_free(zt, sizeof (zio_transform_t));
387 }
388}
389
390/*
391 * ==========================================================================
392 * I/O transform callbacks for subblocks and decompression
393 * ==========================================================================
394 */
395static void
396zio_subblock(zio_t *zio, abd_t *data, uint64_t size)
397{
398 ASSERT(zio->io_size > size);
399
400 if (zio->io_type == ZIO_TYPE_READ)
401 abd_copy(data, zio->io_abd, size);
402}
403
404static void
405zio_decompress(zio_t *zio, abd_t *data, uint64_t size)
406{
407 if (zio->io_error == 0) {
408 void *tmp = abd_borrow_buf(data, size);
409 int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
410 zio->io_abd, tmp, zio->io_size, size);
411 abd_return_buf_copy(data, tmp, size);
412
413 if (ret != 0)
414 zio->io_error = SET_ERROR(EIO);
415 }
416}
417
418/*
419 * ==========================================================================
420 * I/O parent/child relationships and pipeline interlocks
421 * ==========================================================================
422 */
423zio_t *
424zio_walk_parents(zio_t *cio, zio_link_t **zl)
425{
426 list_t *pl = &cio->io_parent_list;
427
428 *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
429 if (*zl == NULL)
430 return (NULL);
431
432 ASSERT((*zl)->zl_child == cio);
433 return ((*zl)->zl_parent);
434}
435
436zio_t *
437zio_walk_children(zio_t *pio, zio_link_t **zl)
438{
439 list_t *cl = &pio->io_child_list;
440
| 45#include <sys/metaslab_impl.h>
46#include <sys/abd.h>
47
48SYSCTL_DECL(_vfs_zfs);
49SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO");
50#if defined(__amd64__)
51static int zio_use_uma = 1;
52#else
53static int zio_use_uma = 0;
54#endif
55SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
56 "Use uma(9) for ZIO allocations");
57static int zio_exclude_metadata = 0;
58SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, exclude_metadata, CTLFLAG_RDTUN, &zio_exclude_metadata, 0,
59 "Exclude metadata buffers from dumps as well");
60
61zio_trim_stats_t zio_trim_stats = {
62 { "bytes", KSTAT_DATA_UINT64,
63 "Number of bytes successfully TRIMmed" },
64 { "success", KSTAT_DATA_UINT64,
65 "Number of successful TRIM requests" },
66 { "unsupported", KSTAT_DATA_UINT64,
67 "Number of TRIM requests that failed because TRIM is not supported" },
68 { "failed", KSTAT_DATA_UINT64,
69 "Number of TRIM requests that failed for reasons other than not supported" },
70};
71
72static kstat_t *zio_trim_ksp;
73
74/*
75 * ==========================================================================
76 * I/O type descriptions
77 * ==========================================================================
78 */
79const char *zio_type_name[ZIO_TYPES] = {
80 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
81 "zio_ioctl"
82};
83
84boolean_t zio_dva_throttle_enabled = B_TRUE;
85SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, dva_throttle_enabled, CTLFLAG_RDTUN,
86 &zio_dva_throttle_enabled, 0, "");
87
88/*
89 * ==========================================================================
90 * I/O kmem caches
91 * ==========================================================================
92 */
93kmem_cache_t *zio_cache;
94kmem_cache_t *zio_link_cache;
95kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
96kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
97
98#ifdef _KERNEL
99extern vmem_t *zio_alloc_arena;
100#endif
101
102#define ZIO_PIPELINE_CONTINUE 0x100
103#define ZIO_PIPELINE_STOP 0x101
104
105#define BP_SPANB(indblkshift, level) \
106 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
107#define COMPARE_META_LEVEL 0x80000000ul
108/*
109 * The following actions directly effect the spa's sync-to-convergence logic.
110 * The values below define the sync pass when we start performing the action.
111 * Care should be taken when changing these values as they directly impact
112 * spa_sync() performance. Tuning these values may introduce subtle performance
113 * pathologies and should only be done in the context of performance analysis.
114 * These tunables will eventually be removed and replaced with #defines once
115 * enough analysis has been done to determine optimal values.
116 *
117 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
118 * regular blocks are not deferred.
119 */
120int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
121SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN,
122 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass");
123int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
124SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN,
125 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass");
126int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
127SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN,
128 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass");
129
130/*
131 * An allocating zio is one that either currently has the DVA allocate
132 * stage set or will have it later in its lifetime.
133 */
134#define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
135
136boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
137
138#ifdef illumos
139#ifdef ZFS_DEBUG
140int zio_buf_debug_limit = 16384;
141#else
142int zio_buf_debug_limit = 0;
143#endif
144#endif
145
146static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
147
148void
149zio_init(void)
150{
151 size_t c;
152 zio_cache = kmem_cache_create("zio_cache",
153 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
154 zio_link_cache = kmem_cache_create("zio_link_cache",
155 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
156 if (!zio_use_uma)
157 goto out;
158
159 /*
160 * For small buffers, we want a cache for each multiple of
161 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
162 * for each quarter-power of 2.
163 */
164 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
165 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
166 size_t p2 = size;
167 size_t align = 0;
168 int cflags = zio_exclude_metadata ? KMC_NODEBUG : 0;
169
170 while (!ISP2(p2))
171 p2 &= p2 - 1;
172
173#ifdef illumos
174#ifndef _KERNEL
175 /*
176 * If we are using watchpoints, put each buffer on its own page,
177 * to eliminate the performance overhead of trapping to the
178 * kernel when modifying a non-watched buffer that shares the
179 * page with a watched buffer.
180 */
181 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
182 continue;
183#endif
184#endif /* illumos */
185 if (size <= 4 * SPA_MINBLOCKSIZE) {
186 align = SPA_MINBLOCKSIZE;
187 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
188 align = MIN(p2 >> 2, PAGESIZE);
189 }
190
191 if (align != 0) {
192 char name[36];
193 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
194 zio_buf_cache[c] = kmem_cache_create(name, size,
195 align, NULL, NULL, NULL, NULL, NULL, cflags);
196
197 /*
198 * Since zio_data bufs do not appear in crash dumps, we
199 * pass KMC_NOTOUCH so that no allocator metadata is
200 * stored with the buffers.
201 */
202 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
203 zio_data_buf_cache[c] = kmem_cache_create(name, size,
204 align, NULL, NULL, NULL, NULL, NULL,
205 cflags | KMC_NOTOUCH | KMC_NODEBUG);
206 }
207 }
208
209 while (--c != 0) {
210 ASSERT(zio_buf_cache[c] != NULL);
211 if (zio_buf_cache[c - 1] == NULL)
212 zio_buf_cache[c - 1] = zio_buf_cache[c];
213
214 ASSERT(zio_data_buf_cache[c] != NULL);
215 if (zio_data_buf_cache[c - 1] == NULL)
216 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
217 }
218out:
219
220 zio_inject_init();
221
222 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc",
223 KSTAT_TYPE_NAMED,
224 sizeof(zio_trim_stats) / sizeof(kstat_named_t),
225 KSTAT_FLAG_VIRTUAL);
226
227 if (zio_trim_ksp != NULL) {
228 zio_trim_ksp->ks_data = &zio_trim_stats;
229 kstat_install(zio_trim_ksp);
230 }
231}
232
233void
234zio_fini(void)
235{
236 size_t c;
237 kmem_cache_t *last_cache = NULL;
238 kmem_cache_t *last_data_cache = NULL;
239
240 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
241 if (zio_buf_cache[c] != last_cache) {
242 last_cache = zio_buf_cache[c];
243 kmem_cache_destroy(zio_buf_cache[c]);
244 }
245 zio_buf_cache[c] = NULL;
246
247 if (zio_data_buf_cache[c] != last_data_cache) {
248 last_data_cache = zio_data_buf_cache[c];
249 kmem_cache_destroy(zio_data_buf_cache[c]);
250 }
251 zio_data_buf_cache[c] = NULL;
252 }
253
254 kmem_cache_destroy(zio_link_cache);
255 kmem_cache_destroy(zio_cache);
256
257 zio_inject_fini();
258
259 if (zio_trim_ksp != NULL) {
260 kstat_delete(zio_trim_ksp);
261 zio_trim_ksp = NULL;
262 }
263}
264
265/*
266 * ==========================================================================
267 * Allocate and free I/O buffers
268 * ==========================================================================
269 */
270
271/*
272 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
273 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
274 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
275 * excess / transient data in-core during a crashdump.
276 */
277void *
278zio_buf_alloc(size_t size)
279{
280 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
281 int flags = zio_exclude_metadata ? KM_NODEBUG : 0;
282
283 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
284
285 if (zio_use_uma)
286 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
287 else
288 return (kmem_alloc(size, KM_SLEEP|flags));
289}
290
291/*
292 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
293 * crashdump if the kernel panics. This exists so that we will limit the amount
294 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
295 * of kernel heap dumped to disk when the kernel panics)
296 */
297void *
298zio_data_buf_alloc(size_t size)
299{
300 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
301
302 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
303
304 if (zio_use_uma)
305 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
306 else
307 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
308}
309
310void
311zio_buf_free(void *buf, size_t size)
312{
313 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
314
315 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
316
317 if (zio_use_uma)
318 kmem_cache_free(zio_buf_cache[c], buf);
319 else
320 kmem_free(buf, size);
321}
322
323void
324zio_data_buf_free(void *buf, size_t size)
325{
326 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
327
328 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
329
330 if (zio_use_uma)
331 kmem_cache_free(zio_data_buf_cache[c], buf);
332 else
333 kmem_free(buf, size);
334}
335
336/*
337 * ==========================================================================
338 * Push and pop I/O transform buffers
339 * ==========================================================================
340 */
341void
342zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize,
343 zio_transform_func_t *transform)
344{
345 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
346
347 /*
348 * Ensure that anyone expecting this zio to contain a linear ABD isn't
349 * going to get a nasty surprise when they try to access the data.
350 */
351#ifdef illumos
352 IMPLY(abd_is_linear(zio->io_abd), abd_is_linear(data));
353#else
354 IMPLY(zio->io_abd != NULL && abd_is_linear(zio->io_abd),
355 abd_is_linear(data));
356#endif
357
358 zt->zt_orig_abd = zio->io_abd;
359 zt->zt_orig_size = zio->io_size;
360 zt->zt_bufsize = bufsize;
361 zt->zt_transform = transform;
362
363 zt->zt_next = zio->io_transform_stack;
364 zio->io_transform_stack = zt;
365
366 zio->io_abd = data;
367 zio->io_size = size;
368}
369
370void
371zio_pop_transforms(zio_t *zio)
372{
373 zio_transform_t *zt;
374
375 while ((zt = zio->io_transform_stack) != NULL) {
376 if (zt->zt_transform != NULL)
377 zt->zt_transform(zio,
378 zt->zt_orig_abd, zt->zt_orig_size);
379
380 if (zt->zt_bufsize != 0)
381 abd_free(zio->io_abd);
382
383 zio->io_abd = zt->zt_orig_abd;
384 zio->io_size = zt->zt_orig_size;
385 zio->io_transform_stack = zt->zt_next;
386
387 kmem_free(zt, sizeof (zio_transform_t));
388 }
389}
390
391/*
392 * ==========================================================================
393 * I/O transform callbacks for subblocks and decompression
394 * ==========================================================================
395 */
396static void
397zio_subblock(zio_t *zio, abd_t *data, uint64_t size)
398{
399 ASSERT(zio->io_size > size);
400
401 if (zio->io_type == ZIO_TYPE_READ)
402 abd_copy(data, zio->io_abd, size);
403}
404
405static void
406zio_decompress(zio_t *zio, abd_t *data, uint64_t size)
407{
408 if (zio->io_error == 0) {
409 void *tmp = abd_borrow_buf(data, size);
410 int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
411 zio->io_abd, tmp, zio->io_size, size);
412 abd_return_buf_copy(data, tmp, size);
413
414 if (ret != 0)
415 zio->io_error = SET_ERROR(EIO);
416 }
417}
418
419/*
420 * ==========================================================================
421 * I/O parent/child relationships and pipeline interlocks
422 * ==========================================================================
423 */
424zio_t *
425zio_walk_parents(zio_t *cio, zio_link_t **zl)
426{
427 list_t *pl = &cio->io_parent_list;
428
429 *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
430 if (*zl == NULL)
431 return (NULL);
432
433 ASSERT((*zl)->zl_child == cio);
434 return ((*zl)->zl_parent);
435}
436
437zio_t *
438zio_walk_children(zio_t *pio, zio_link_t **zl)
439{
440 list_t *cl = &pio->io_child_list;
441
|
| 442 ASSERT(MUTEX_HELD(&pio->io_lock));
443
|
441 *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
442 if (*zl == NULL)
443 return (NULL);
444
445 ASSERT((*zl)->zl_parent == pio);
446 return ((*zl)->zl_child);
447}
448
449zio_t *
450zio_unique_parent(zio_t *cio)
451{
452 zio_link_t *zl = NULL;
453 zio_t *pio = zio_walk_parents(cio, &zl);
454
455 VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
456 return (pio);
457}
458
459void
460zio_add_child(zio_t *pio, zio_t *cio)
461{
462 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
463
464 /*
465 * Logical I/Os can have logical, gang, or vdev children.
466 * Gang I/Os can have gang or vdev children.
467 * Vdev I/Os can only have vdev children.
468 * The following ASSERT captures all of these constraints.
469 */
470 ASSERT3S(cio->io_child_type, <=, pio->io_child_type);
471
472 zl->zl_parent = pio;
473 zl->zl_child = cio;
474
| 444 *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
445 if (*zl == NULL)
446 return (NULL);
447
448 ASSERT((*zl)->zl_parent == pio);
449 return ((*zl)->zl_child);
450}
451
452zio_t *
453zio_unique_parent(zio_t *cio)
454{
455 zio_link_t *zl = NULL;
456 zio_t *pio = zio_walk_parents(cio, &zl);
457
458 VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
459 return (pio);
460}
461
462void
463zio_add_child(zio_t *pio, zio_t *cio)
464{
465 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
466
467 /*
468 * Logical I/Os can have logical, gang, or vdev children.
469 * Gang I/Os can have gang or vdev children.
470 * Vdev I/Os can only have vdev children.
471 * The following ASSERT captures all of these constraints.
472 */
473 ASSERT3S(cio->io_child_type, <=, pio->io_child_type);
474
475 zl->zl_parent = pio;
476 zl->zl_child = cio;
477
|
475 mutex_enter(&cio->io_lock);
| |
476 mutex_enter(&pio->io_lock);
| 478 mutex_enter(&pio->io_lock);
|
| 479 mutex_enter(&cio->io_lock);
|
477
478 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
479
480 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
481 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
482
483 list_insert_head(&pio->io_child_list, zl);
484 list_insert_head(&cio->io_parent_list, zl);
485
486 pio->io_child_count++;
487 cio->io_parent_count++;
488
| 480
481 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
482
483 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
484 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
485
486 list_insert_head(&pio->io_child_list, zl);
487 list_insert_head(&cio->io_parent_list, zl);
488
489 pio->io_child_count++;
490 cio->io_parent_count++;
491
|
489 mutex_exit(&pio->io_lock);
| |
490 mutex_exit(&cio->io_lock);
| 492 mutex_exit(&cio->io_lock);
|
| 493 mutex_exit(&pio->io_lock);
|
491}
492
493static void
494zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
495{
496 ASSERT(zl->zl_parent == pio);
497 ASSERT(zl->zl_child == cio);
498
| 494}
495
496static void
497zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
498{
499 ASSERT(zl->zl_parent == pio);
500 ASSERT(zl->zl_child == cio);
501
|
499 mutex_enter(&cio->io_lock);
| |
500 mutex_enter(&pio->io_lock);
| 502 mutex_enter(&pio->io_lock);
|
| 503 mutex_enter(&cio->io_lock);
|
501
502 list_remove(&pio->io_child_list, zl);
503 list_remove(&cio->io_parent_list, zl);
504
505 pio->io_child_count--;
506 cio->io_parent_count--;
507
| 504
505 list_remove(&pio->io_child_list, zl);
506 list_remove(&cio->io_parent_list, zl);
507
508 pio->io_child_count--;
509 cio->io_parent_count--;
510
|
508 mutex_exit(&pio->io_lock);
| |
509 mutex_exit(&cio->io_lock);
| 511 mutex_exit(&cio->io_lock);
|
510
| 512 mutex_exit(&pio->io_lock);
|
511 kmem_cache_free(zio_link_cache, zl);
512}
513
514static boolean_t
515zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait)
516{
517 boolean_t waiting = B_FALSE;
518
519 mutex_enter(&zio->io_lock);
520 ASSERT(zio->io_stall == NULL);
521 for (int c = 0; c < ZIO_CHILD_TYPES; c++) {
522 if (!(ZIO_CHILD_BIT_IS_SET(childbits, c)))
523 continue;
524
525 uint64_t *countp = &zio->io_children[c][wait];
526 if (*countp != 0) {
527 zio->io_stage >>= 1;
528 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
529 zio->io_stall = countp;
530 waiting = B_TRUE;
531 break;
532 }
533 }
534 mutex_exit(&zio->io_lock);
535 return (waiting);
536}
537
538static void
539zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
540{
541 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
542 int *errorp = &pio->io_child_error[zio->io_child_type];
543
544 mutex_enter(&pio->io_lock);
545 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
546 *errorp = zio_worst_error(*errorp, zio->io_error);
547 pio->io_reexecute |= zio->io_reexecute;
548 ASSERT3U(*countp, >, 0);
549
550 (*countp)--;
551
552 if (*countp == 0 && pio->io_stall == countp) {
553 zio_taskq_type_t type =
554 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
555 ZIO_TASKQ_INTERRUPT;
556 pio->io_stall = NULL;
557 mutex_exit(&pio->io_lock);
558 /*
559 * Dispatch the parent zio in its own taskq so that
560 * the child can continue to make progress. This also
561 * prevents overflowing the stack when we have deeply nested
562 * parent-child relationships.
563 */
564 zio_taskq_dispatch(pio, type, B_FALSE);
565 } else {
566 mutex_exit(&pio->io_lock);
567 }
568}
569
570static void
571zio_inherit_child_errors(zio_t *zio, enum zio_child c)
572{
573 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
574 zio->io_error = zio->io_child_error[c];
575}
576
577int
578zio_bookmark_compare(const void *x1, const void *x2)
579{
580 const zio_t *z1 = x1;
581 const zio_t *z2 = x2;
582
583 if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
584 return (-1);
585 if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
586 return (1);
587
588 if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
589 return (-1);
590 if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
591 return (1);
592
593 if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
594 return (-1);
595 if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
596 return (1);
597
598 if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
599 return (-1);
600 if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
601 return (1);
602
603 if (z1 < z2)
604 return (-1);
605 if (z1 > z2)
606 return (1);
607
608 return (0);
609}
610
611/*
612 * ==========================================================================
613 * Create the various types of I/O (read, write, free, etc)
614 * ==========================================================================
615 */
616static zio_t *
617zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
618 abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done,
619 void *private, zio_type_t type, zio_priority_t priority,
620 enum zio_flag flags, vdev_t *vd, uint64_t offset,
621 const zbookmark_phys_t *zb, enum zio_stage stage, enum zio_stage pipeline)
622{
623 zio_t *zio;
624
625 ASSERT3U(type == ZIO_TYPE_FREE || psize, <=, SPA_MAXBLOCKSIZE);
626 ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0);
627 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
628
629 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
630 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
631 ASSERT(vd || stage == ZIO_STAGE_OPEN);
632
633 IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW) != 0);
634
635 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
636 bzero(zio, sizeof (zio_t));
637
638 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
639 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
640
641 list_create(&zio->io_parent_list, sizeof (zio_link_t),
642 offsetof(zio_link_t, zl_parent_node));
643 list_create(&zio->io_child_list, sizeof (zio_link_t),
644 offsetof(zio_link_t, zl_child_node));
645 metaslab_trace_init(&zio->io_alloc_list);
646
647 if (vd != NULL)
648 zio->io_child_type = ZIO_CHILD_VDEV;
649 else if (flags & ZIO_FLAG_GANG_CHILD)
650 zio->io_child_type = ZIO_CHILD_GANG;
651 else if (flags & ZIO_FLAG_DDT_CHILD)
652 zio->io_child_type = ZIO_CHILD_DDT;
653 else
654 zio->io_child_type = ZIO_CHILD_LOGICAL;
655
656 if (bp != NULL) {
657 zio->io_bp = (blkptr_t *)bp;
658 zio->io_bp_copy = *bp;
659 zio->io_bp_orig = *bp;
660 if (type != ZIO_TYPE_WRITE ||
661 zio->io_child_type == ZIO_CHILD_DDT)
662 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
663 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
664 zio->io_logical = zio;
665 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
666 pipeline |= ZIO_GANG_STAGES;
667 }
668
669 zio->io_spa = spa;
670 zio->io_txg = txg;
671 zio->io_done = done;
672 zio->io_private = private;
673 zio->io_type = type;
674 zio->io_priority = priority;
675 zio->io_vd = vd;
676 zio->io_offset = offset;
677 zio->io_orig_abd = zio->io_abd = data;
678 zio->io_orig_size = zio->io_size = psize;
679 zio->io_lsize = lsize;
680 zio->io_orig_flags = zio->io_flags = flags;
681 zio->io_orig_stage = zio->io_stage = stage;
682 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
683 zio->io_pipeline_trace = ZIO_STAGE_OPEN;
684
685 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
686 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
687
688 if (zb != NULL)
689 zio->io_bookmark = *zb;
690
691 if (pio != NULL) {
692 if (zio->io_logical == NULL)
693 zio->io_logical = pio->io_logical;
694 if (zio->io_child_type == ZIO_CHILD_GANG)
695 zio->io_gang_leader = pio->io_gang_leader;
696 zio_add_child(pio, zio);
697 }
698
699 return (zio);
700}
701
702static void
703zio_destroy(zio_t *zio)
704{
705 metaslab_trace_fini(&zio->io_alloc_list);
706 list_destroy(&zio->io_parent_list);
707 list_destroy(&zio->io_child_list);
708 mutex_destroy(&zio->io_lock);
709 cv_destroy(&zio->io_cv);
710 kmem_cache_free(zio_cache, zio);
711}
712
713zio_t *
714zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
715 void *private, enum zio_flag flags)
716{
717 zio_t *zio;
718
719 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
720 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
721 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
722
723 return (zio);
724}
725
726zio_t *
727zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
728{
729 return (zio_null(NULL, spa, NULL, done, private, flags));
730}
731
732void
733zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
734{
735 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
736 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
737 bp, (longlong_t)BP_GET_TYPE(bp));
738 }
739 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
740 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
741 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
742 bp, (longlong_t)BP_GET_CHECKSUM(bp));
743 }
744 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
745 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
746 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
747 bp, (longlong_t)BP_GET_COMPRESS(bp));
748 }
749 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
750 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
751 bp, (longlong_t)BP_GET_LSIZE(bp));
752 }
753 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
754 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
755 bp, (longlong_t)BP_GET_PSIZE(bp));
756 }
757
758 if (BP_IS_EMBEDDED(bp)) {
759 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
760 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
761 bp, (longlong_t)BPE_GET_ETYPE(bp));
762 }
763 }
764
765 /*
766 * Do not verify individual DVAs if the config is not trusted. This
767 * will be done once the zio is executed in vdev_mirror_map_alloc.
768 */
769 if (!spa->spa_trust_config)
770 return;
771
772 /*
773 * Pool-specific checks.
774 *
775 * Note: it would be nice to verify that the blk_birth and
776 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
777 * allows the birth time of log blocks (and dmu_sync()-ed blocks
778 * that are in the log) to be arbitrarily large.
779 */
780 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
781 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
782 if (vdevid >= spa->spa_root_vdev->vdev_children) {
783 zfs_panic_recover("blkptr at %p DVA %u has invalid "
784 "VDEV %llu",
785 bp, i, (longlong_t)vdevid);
786 continue;
787 }
788 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
789 if (vd == NULL) {
790 zfs_panic_recover("blkptr at %p DVA %u has invalid "
791 "VDEV %llu",
792 bp, i, (longlong_t)vdevid);
793 continue;
794 }
795 if (vd->vdev_ops == &vdev_hole_ops) {
796 zfs_panic_recover("blkptr at %p DVA %u has hole "
797 "VDEV %llu",
798 bp, i, (longlong_t)vdevid);
799 continue;
800 }
801 if (vd->vdev_ops == &vdev_missing_ops) {
802 /*
803 * "missing" vdevs are valid during import, but we
804 * don't have their detailed info (e.g. asize), so
805 * we can't perform any more checks on them.
806 */
807 continue;
808 }
809 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
810 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
811 if (BP_IS_GANG(bp))
812 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
813 if (offset + asize > vd->vdev_asize) {
814 zfs_panic_recover("blkptr at %p DVA %u has invalid "
815 "OFFSET %llu",
816 bp, i, (longlong_t)offset);
817 }
818 }
819}
820
821boolean_t
822zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp)
823{
824 uint64_t vdevid = DVA_GET_VDEV(dva);
825
826 if (vdevid >= spa->spa_root_vdev->vdev_children)
827 return (B_FALSE);
828
829 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
830 if (vd == NULL)
831 return (B_FALSE);
832
833 if (vd->vdev_ops == &vdev_hole_ops)
834 return (B_FALSE);
835
836 if (vd->vdev_ops == &vdev_missing_ops) {
837 return (B_FALSE);
838 }
839
840 uint64_t offset = DVA_GET_OFFSET(dva);
841 uint64_t asize = DVA_GET_ASIZE(dva);
842
843 if (BP_IS_GANG(bp))
844 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
845 if (offset + asize > vd->vdev_asize)
846 return (B_FALSE);
847
848 return (B_TRUE);
849}
850
851zio_t *
852zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
853 abd_t *data, uint64_t size, zio_done_func_t *done, void *private,
854 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
855{
856 zio_t *zio;
857
858 zfs_blkptr_verify(spa, bp);
859
860 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
861 data, size, size, done, private,
862 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
863 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
864 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
865
866 return (zio);
867}
868
869zio_t *
870zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
871 abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
872 zio_done_func_t *ready, zio_done_func_t *children_ready,
873 zio_done_func_t *physdone, zio_done_func_t *done,
874 void *private, zio_priority_t priority, enum zio_flag flags,
875 const zbookmark_phys_t *zb)
876{
877 zio_t *zio;
878
879 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
880 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
881 zp->zp_compress >= ZIO_COMPRESS_OFF &&
882 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
883 DMU_OT_IS_VALID(zp->zp_type) &&
884 zp->zp_level < 32 &&
885 zp->zp_copies > 0 &&
886 zp->zp_copies <= spa_max_replication(spa));
887
888 zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
889 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
890 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
891 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
892
893 zio->io_ready = ready;
894 zio->io_children_ready = children_ready;
895 zio->io_physdone = physdone;
896 zio->io_prop = *zp;
897
898 /*
899 * Data can be NULL if we are going to call zio_write_override() to
900 * provide the already-allocated BP. But we may need the data to
901 * verify a dedup hit (if requested). In this case, don't try to
902 * dedup (just take the already-allocated BP verbatim).
903 */
904 if (data == NULL && zio->io_prop.zp_dedup_verify) {
905 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
906 }
907
908 return (zio);
909}
910
911zio_t *
912zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data,
913 uint64_t size, zio_done_func_t *done, void *private,
914 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
915{
916 zio_t *zio;
917
918 zio = zio_create(pio, spa, txg, bp, data, size, size, done, private,
919 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
920 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
921
922 return (zio);
923}
924
925void
926zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
927{
928 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
929 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
930 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
931 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
932
933 /*
934 * We must reset the io_prop to match the values that existed
935 * when the bp was first written by dmu_sync() keeping in mind
936 * that nopwrite and dedup are mutually exclusive.
937 */
938 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
939 zio->io_prop.zp_nopwrite = nopwrite;
940 zio->io_prop.zp_copies = copies;
941 zio->io_bp_override = bp;
942}
943
944void
945zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
946{
947
948 zfs_blkptr_verify(spa, bp);
949
950 /*
951 * The check for EMBEDDED is a performance optimization. We
952 * process the free here (by ignoring it) rather than
953 * putting it on the list and then processing it in zio_free_sync().
954 */
955 if (BP_IS_EMBEDDED(bp))
956 return;
957 metaslab_check_free(spa, bp);
958
959 /*
960 * Frees that are for the currently-syncing txg, are not going to be
961 * deferred, and which will not need to do a read (i.e. not GANG or
962 * DEDUP), can be processed immediately. Otherwise, put them on the
963 * in-memory list for later processing.
964 */
965 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
966 txg != spa->spa_syncing_txg ||
967 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
968 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
969 } else {
970 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
971 BP_GET_PSIZE(bp), 0)));
972 }
973}
974
975zio_t *
976zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
977 uint64_t size, enum zio_flag flags)
978{
979 zio_t *zio;
980 enum zio_stage stage = ZIO_FREE_PIPELINE;
981
982 ASSERT(!BP_IS_HOLE(bp));
983 ASSERT(spa_syncing_txg(spa) == txg);
984 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
985
986 if (BP_IS_EMBEDDED(bp))
987 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
988
989 metaslab_check_free(spa, bp);
990 arc_freed(spa, bp);
| 513 kmem_cache_free(zio_link_cache, zl);
514}
515
516static boolean_t
517zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait)
518{
519 boolean_t waiting = B_FALSE;
520
521 mutex_enter(&zio->io_lock);
522 ASSERT(zio->io_stall == NULL);
523 for (int c = 0; c < ZIO_CHILD_TYPES; c++) {
524 if (!(ZIO_CHILD_BIT_IS_SET(childbits, c)))
525 continue;
526
527 uint64_t *countp = &zio->io_children[c][wait];
528 if (*countp != 0) {
529 zio->io_stage >>= 1;
530 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
531 zio->io_stall = countp;
532 waiting = B_TRUE;
533 break;
534 }
535 }
536 mutex_exit(&zio->io_lock);
537 return (waiting);
538}
539
540static void
541zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
542{
543 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
544 int *errorp = &pio->io_child_error[zio->io_child_type];
545
546 mutex_enter(&pio->io_lock);
547 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
548 *errorp = zio_worst_error(*errorp, zio->io_error);
549 pio->io_reexecute |= zio->io_reexecute;
550 ASSERT3U(*countp, >, 0);
551
552 (*countp)--;
553
554 if (*countp == 0 && pio->io_stall == countp) {
555 zio_taskq_type_t type =
556 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
557 ZIO_TASKQ_INTERRUPT;
558 pio->io_stall = NULL;
559 mutex_exit(&pio->io_lock);
560 /*
561 * Dispatch the parent zio in its own taskq so that
562 * the child can continue to make progress. This also
563 * prevents overflowing the stack when we have deeply nested
564 * parent-child relationships.
565 */
566 zio_taskq_dispatch(pio, type, B_FALSE);
567 } else {
568 mutex_exit(&pio->io_lock);
569 }
570}
571
572static void
573zio_inherit_child_errors(zio_t *zio, enum zio_child c)
574{
575 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
576 zio->io_error = zio->io_child_error[c];
577}
578
579int
580zio_bookmark_compare(const void *x1, const void *x2)
581{
582 const zio_t *z1 = x1;
583 const zio_t *z2 = x2;
584
585 if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
586 return (-1);
587 if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
588 return (1);
589
590 if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
591 return (-1);
592 if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
593 return (1);
594
595 if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
596 return (-1);
597 if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
598 return (1);
599
600 if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
601 return (-1);
602 if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
603 return (1);
604
605 if (z1 < z2)
606 return (-1);
607 if (z1 > z2)
608 return (1);
609
610 return (0);
611}
612
613/*
614 * ==========================================================================
615 * Create the various types of I/O (read, write, free, etc)
616 * ==========================================================================
617 */
618static zio_t *
619zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
620 abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done,
621 void *private, zio_type_t type, zio_priority_t priority,
622 enum zio_flag flags, vdev_t *vd, uint64_t offset,
623 const zbookmark_phys_t *zb, enum zio_stage stage, enum zio_stage pipeline)
624{
625 zio_t *zio;
626
627 ASSERT3U(type == ZIO_TYPE_FREE || psize, <=, SPA_MAXBLOCKSIZE);
628 ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0);
629 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
630
631 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
632 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
633 ASSERT(vd || stage == ZIO_STAGE_OPEN);
634
635 IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW) != 0);
636
637 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
638 bzero(zio, sizeof (zio_t));
639
640 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
641 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
642
643 list_create(&zio->io_parent_list, sizeof (zio_link_t),
644 offsetof(zio_link_t, zl_parent_node));
645 list_create(&zio->io_child_list, sizeof (zio_link_t),
646 offsetof(zio_link_t, zl_child_node));
647 metaslab_trace_init(&zio->io_alloc_list);
648
649 if (vd != NULL)
650 zio->io_child_type = ZIO_CHILD_VDEV;
651 else if (flags & ZIO_FLAG_GANG_CHILD)
652 zio->io_child_type = ZIO_CHILD_GANG;
653 else if (flags & ZIO_FLAG_DDT_CHILD)
654 zio->io_child_type = ZIO_CHILD_DDT;
655 else
656 zio->io_child_type = ZIO_CHILD_LOGICAL;
657
658 if (bp != NULL) {
659 zio->io_bp = (blkptr_t *)bp;
660 zio->io_bp_copy = *bp;
661 zio->io_bp_orig = *bp;
662 if (type != ZIO_TYPE_WRITE ||
663 zio->io_child_type == ZIO_CHILD_DDT)
664 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
665 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
666 zio->io_logical = zio;
667 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
668 pipeline |= ZIO_GANG_STAGES;
669 }
670
671 zio->io_spa = spa;
672 zio->io_txg = txg;
673 zio->io_done = done;
674 zio->io_private = private;
675 zio->io_type = type;
676 zio->io_priority = priority;
677 zio->io_vd = vd;
678 zio->io_offset = offset;
679 zio->io_orig_abd = zio->io_abd = data;
680 zio->io_orig_size = zio->io_size = psize;
681 zio->io_lsize = lsize;
682 zio->io_orig_flags = zio->io_flags = flags;
683 zio->io_orig_stage = zio->io_stage = stage;
684 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
685 zio->io_pipeline_trace = ZIO_STAGE_OPEN;
686
687 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
688 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
689
690 if (zb != NULL)
691 zio->io_bookmark = *zb;
692
693 if (pio != NULL) {
694 if (zio->io_logical == NULL)
695 zio->io_logical = pio->io_logical;
696 if (zio->io_child_type == ZIO_CHILD_GANG)
697 zio->io_gang_leader = pio->io_gang_leader;
698 zio_add_child(pio, zio);
699 }
700
701 return (zio);
702}
703
704static void
705zio_destroy(zio_t *zio)
706{
707 metaslab_trace_fini(&zio->io_alloc_list);
708 list_destroy(&zio->io_parent_list);
709 list_destroy(&zio->io_child_list);
710 mutex_destroy(&zio->io_lock);
711 cv_destroy(&zio->io_cv);
712 kmem_cache_free(zio_cache, zio);
713}
714
715zio_t *
716zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
717 void *private, enum zio_flag flags)
718{
719 zio_t *zio;
720
721 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
722 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
723 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
724
725 return (zio);
726}
727
728zio_t *
729zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
730{
731 return (zio_null(NULL, spa, NULL, done, private, flags));
732}
733
734void
735zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
736{
737 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
738 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
739 bp, (longlong_t)BP_GET_TYPE(bp));
740 }
741 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
742 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
743 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
744 bp, (longlong_t)BP_GET_CHECKSUM(bp));
745 }
746 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
747 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
748 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
749 bp, (longlong_t)BP_GET_COMPRESS(bp));
750 }
751 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
752 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
753 bp, (longlong_t)BP_GET_LSIZE(bp));
754 }
755 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
756 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
757 bp, (longlong_t)BP_GET_PSIZE(bp));
758 }
759
760 if (BP_IS_EMBEDDED(bp)) {
761 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
762 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
763 bp, (longlong_t)BPE_GET_ETYPE(bp));
764 }
765 }
766
767 /*
768 * Do not verify individual DVAs if the config is not trusted. This
769 * will be done once the zio is executed in vdev_mirror_map_alloc.
770 */
771 if (!spa->spa_trust_config)
772 return;
773
774 /*
775 * Pool-specific checks.
776 *
777 * Note: it would be nice to verify that the blk_birth and
778 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
779 * allows the birth time of log blocks (and dmu_sync()-ed blocks
780 * that are in the log) to be arbitrarily large.
781 */
782 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
783 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
784 if (vdevid >= spa->spa_root_vdev->vdev_children) {
785 zfs_panic_recover("blkptr at %p DVA %u has invalid "
786 "VDEV %llu",
787 bp, i, (longlong_t)vdevid);
788 continue;
789 }
790 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
791 if (vd == NULL) {
792 zfs_panic_recover("blkptr at %p DVA %u has invalid "
793 "VDEV %llu",
794 bp, i, (longlong_t)vdevid);
795 continue;
796 }
797 if (vd->vdev_ops == &vdev_hole_ops) {
798 zfs_panic_recover("blkptr at %p DVA %u has hole "
799 "VDEV %llu",
800 bp, i, (longlong_t)vdevid);
801 continue;
802 }
803 if (vd->vdev_ops == &vdev_missing_ops) {
804 /*
805 * "missing" vdevs are valid during import, but we
806 * don't have their detailed info (e.g. asize), so
807 * we can't perform any more checks on them.
808 */
809 continue;
810 }
811 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
812 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
813 if (BP_IS_GANG(bp))
814 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
815 if (offset + asize > vd->vdev_asize) {
816 zfs_panic_recover("blkptr at %p DVA %u has invalid "
817 "OFFSET %llu",
818 bp, i, (longlong_t)offset);
819 }
820 }
821}
822
823boolean_t
824zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp)
825{
826 uint64_t vdevid = DVA_GET_VDEV(dva);
827
828 if (vdevid >= spa->spa_root_vdev->vdev_children)
829 return (B_FALSE);
830
831 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
832 if (vd == NULL)
833 return (B_FALSE);
834
835 if (vd->vdev_ops == &vdev_hole_ops)
836 return (B_FALSE);
837
838 if (vd->vdev_ops == &vdev_missing_ops) {
839 return (B_FALSE);
840 }
841
842 uint64_t offset = DVA_GET_OFFSET(dva);
843 uint64_t asize = DVA_GET_ASIZE(dva);
844
845 if (BP_IS_GANG(bp))
846 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
847 if (offset + asize > vd->vdev_asize)
848 return (B_FALSE);
849
850 return (B_TRUE);
851}
852
853zio_t *
854zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
855 abd_t *data, uint64_t size, zio_done_func_t *done, void *private,
856 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
857{
858 zio_t *zio;
859
860 zfs_blkptr_verify(spa, bp);
861
862 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
863 data, size, size, done, private,
864 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
865 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
866 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
867
868 return (zio);
869}
870
871zio_t *
872zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
873 abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
874 zio_done_func_t *ready, zio_done_func_t *children_ready,
875 zio_done_func_t *physdone, zio_done_func_t *done,
876 void *private, zio_priority_t priority, enum zio_flag flags,
877 const zbookmark_phys_t *zb)
878{
879 zio_t *zio;
880
881 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
882 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
883 zp->zp_compress >= ZIO_COMPRESS_OFF &&
884 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
885 DMU_OT_IS_VALID(zp->zp_type) &&
886 zp->zp_level < 32 &&
887 zp->zp_copies > 0 &&
888 zp->zp_copies <= spa_max_replication(spa));
889
890 zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
891 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
892 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
893 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
894
895 zio->io_ready = ready;
896 zio->io_children_ready = children_ready;
897 zio->io_physdone = physdone;
898 zio->io_prop = *zp;
899
900 /*
901 * Data can be NULL if we are going to call zio_write_override() to
902 * provide the already-allocated BP. But we may need the data to
903 * verify a dedup hit (if requested). In this case, don't try to
904 * dedup (just take the already-allocated BP verbatim).
905 */
906 if (data == NULL && zio->io_prop.zp_dedup_verify) {
907 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
908 }
909
910 return (zio);
911}
912
913zio_t *
914zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data,
915 uint64_t size, zio_done_func_t *done, void *private,
916 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
917{
918 zio_t *zio;
919
920 zio = zio_create(pio, spa, txg, bp, data, size, size, done, private,
921 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
922 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
923
924 return (zio);
925}
926
927void
928zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
929{
930 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
931 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
932 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
933 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
934
935 /*
936 * We must reset the io_prop to match the values that existed
937 * when the bp was first written by dmu_sync() keeping in mind
938 * that nopwrite and dedup are mutually exclusive.
939 */
940 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
941 zio->io_prop.zp_nopwrite = nopwrite;
942 zio->io_prop.zp_copies = copies;
943 zio->io_bp_override = bp;
944}
945
946void
947zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
948{
949
950 zfs_blkptr_verify(spa, bp);
951
952 /*
953 * The check for EMBEDDED is a performance optimization. We
954 * process the free here (by ignoring it) rather than
955 * putting it on the list and then processing it in zio_free_sync().
956 */
957 if (BP_IS_EMBEDDED(bp))
958 return;
959 metaslab_check_free(spa, bp);
960
961 /*
962 * Frees that are for the currently-syncing txg, are not going to be
963 * deferred, and which will not need to do a read (i.e. not GANG or
964 * DEDUP), can be processed immediately. Otherwise, put them on the
965 * in-memory list for later processing.
966 */
967 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
968 txg != spa->spa_syncing_txg ||
969 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
970 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
971 } else {
972 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
973 BP_GET_PSIZE(bp), 0)));
974 }
975}
976
977zio_t *
978zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
979 uint64_t size, enum zio_flag flags)
980{
981 zio_t *zio;
982 enum zio_stage stage = ZIO_FREE_PIPELINE;
983
984 ASSERT(!BP_IS_HOLE(bp));
985 ASSERT(spa_syncing_txg(spa) == txg);
986 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
987
988 if (BP_IS_EMBEDDED(bp))
989 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
990
991 metaslab_check_free(spa, bp);
992 arc_freed(spa, bp);
|
| 993 dsl_scan_freed(spa, bp);
|
991
992 if (zfs_trim_enabled)
993 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
994 ZIO_STAGE_VDEV_IO_ASSESS;
995 /*
996 * GANG and DEDUP blocks can induce a read (for the gang block header,
997 * or the DDT), so issue them asynchronously so that this thread is
998 * not tied up.
999 */
1000 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
1001 stage |= ZIO_STAGE_ISSUE_ASYNC;
1002
1003 flags |= ZIO_FLAG_DONT_QUEUE;
1004
1005 zio = zio_create(pio, spa, txg, bp, NULL, size,
1006 size, NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
1007 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
1008
1009 return (zio);
1010}
1011
1012zio_t *
1013zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1014 zio_done_func_t *done, void *private, enum zio_flag flags)
1015{
1016 zio_t *zio;
1017
1018 zfs_blkptr_verify(spa, bp);
1019
1020 if (BP_IS_EMBEDDED(bp))
1021 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1022
1023 /*
1024 * A claim is an allocation of a specific block. Claims are needed
1025 * to support immediate writes in the intent log. The issue is that
1026 * immediate writes contain committed data, but in a txg that was
1027 * *not* committed. Upon opening the pool after an unclean shutdown,
1028 * the intent log claims all blocks that contain immediate write data
1029 * so that the SPA knows they're in use.
1030 *
1031 * All claims *must* be resolved in the first txg -- before the SPA
1032 * starts allocating blocks -- so that nothing is allocated twice.
1033 * If txg == 0 we just verify that the block is claimable.
1034 */
1035 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <,
1036 spa_min_claim_txg(spa));
1037 ASSERT(txg == spa_min_claim_txg(spa) || txg == 0);
1038 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
1039
1040 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1041 BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
1042 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
1043 ASSERT0(zio->io_queued_timestamp);
1044
1045 return (zio);
1046}
1047
1048zio_t *
1049zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
1050 uint64_t size, zio_done_func_t *done, void *private,
1051 zio_priority_t priority, enum zio_flag flags)
1052{
1053 zio_t *zio;
1054 int c;
1055
1056 if (vd->vdev_children == 0) {
1057 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
1058 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
1059 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
1060
1061 zio->io_cmd = cmd;
1062 } else {
1063 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
1064
1065 for (c = 0; c < vd->vdev_children; c++)
1066 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
1067 offset, size, done, private, priority, flags));
1068 }
1069
1070 return (zio);
1071}
1072
1073zio_t *
1074zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1075 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1076 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1077{
1078 zio_t *zio;
1079
1080 ASSERT(vd->vdev_children == 0);
1081 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1082 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1083 ASSERT3U(offset + size, <=, vd->vdev_psize);
1084
1085 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1086 private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1087 offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1088
1089 zio->io_prop.zp_checksum = checksum;
1090
1091 return (zio);
1092}
1093
1094zio_t *
1095zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1096 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1097 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1098{
1099 zio_t *zio;
1100
1101 ASSERT(vd->vdev_children == 0);
1102 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1103 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1104 ASSERT3U(offset + size, <=, vd->vdev_psize);
1105
1106 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1107 private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1108 offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1109
1110 zio->io_prop.zp_checksum = checksum;
1111
1112 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1113 /*
1114 * zec checksums are necessarily destructive -- they modify
1115 * the end of the write buffer to hold the verifier/checksum.
1116 * Therefore, we must make a local copy in case the data is
1117 * being written to multiple places in parallel.
1118 */
1119 abd_t *wbuf = abd_alloc_sametype(data, size);
1120 abd_copy(wbuf, data, size);
1121
1122 zio_push_transform(zio, wbuf, size, size, NULL);
1123 }
1124
1125 return (zio);
1126}
1127
1128/*
1129 * Create a child I/O to do some work for us.
1130 */
1131zio_t *
1132zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1133 abd_t *data, uint64_t size, int type, zio_priority_t priority,
1134 enum zio_flag flags, zio_done_func_t *done, void *private)
1135{
1136 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1137 zio_t *zio;
1138
1139 /*
1140 * vdev child I/Os do not propagate their error to the parent.
1141 * Therefore, for correct operation the caller *must* check for
1142 * and handle the error in the child i/o's done callback.
1143 * The only exceptions are i/os that we don't care about
1144 * (OPTIONAL or REPAIR).
1145 */
1146 ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) ||
1147 done != NULL);
1148
1149 /*
1150 * In the common case, where the parent zio was to a normal vdev,
1151 * the child zio must be to a child vdev of that vdev. Otherwise,
1152 * the child zio must be to a top-level vdev.
1153 */
1154 if (pio->io_vd != NULL && pio->io_vd->vdev_ops != &vdev_indirect_ops) {
1155 ASSERT3P(vd->vdev_parent, ==, pio->io_vd);
1156 } else {
1157 ASSERT3P(vd, ==, vd->vdev_top);
1158 }
1159
1160 if (type == ZIO_TYPE_READ && bp != NULL) {
1161 /*
1162 * If we have the bp, then the child should perform the
1163 * checksum and the parent need not. This pushes error
1164 * detection as close to the leaves as possible and
1165 * eliminates redundant checksums in the interior nodes.
1166 */
1167 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1168 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1169 }
1170
1171 /* Not all IO types require vdev io done stage e.g. free */
1172 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE))
1173 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE;
1174
1175 if (vd->vdev_ops->vdev_op_leaf) {
1176 ASSERT0(vd->vdev_children);
1177 offset += VDEV_LABEL_START_SIZE;
1178 }
1179
1180 flags |= ZIO_VDEV_CHILD_FLAGS(pio);
1181
1182 /*
1183 * If we've decided to do a repair, the write is not speculative --
1184 * even if the original read was.
1185 */
1186 if (flags & ZIO_FLAG_IO_REPAIR)
1187 flags &= ~ZIO_FLAG_SPECULATIVE;
1188
1189 /*
1190 * If we're creating a child I/O that is not associated with a
1191 * top-level vdev, then the child zio is not an allocating I/O.
1192 * If this is a retried I/O then we ignore it since we will
1193 * have already processed the original allocating I/O.
1194 */
1195 if (flags & ZIO_FLAG_IO_ALLOCATING &&
1196 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1197 metaslab_class_t *mc = spa_normal_class(pio->io_spa);
1198
1199 ASSERT(mc->mc_alloc_throttle_enabled);
1200 ASSERT(type == ZIO_TYPE_WRITE);
1201 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1202 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1203 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1204 pio->io_child_type == ZIO_CHILD_GANG);
1205
1206 flags &= ~ZIO_FLAG_IO_ALLOCATING;
1207 }
1208
1209 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1210 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1211 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1212 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1213
1214 zio->io_physdone = pio->io_physdone;
1215 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1216 zio->io_logical->io_phys_children++;
1217
1218 return (zio);
1219}
1220
1221zio_t *
1222zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size,
1223 int type, zio_priority_t priority, enum zio_flag flags,
1224 zio_done_func_t *done, void *private)
1225{
1226 zio_t *zio;
1227
1228 ASSERT(vd->vdev_ops->vdev_op_leaf);
1229
1230 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1231 data, size, size, done, private, type, priority,
1232 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1233 vd, offset, NULL,
1234 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1235
1236 return (zio);
1237}
1238
1239void
1240zio_flush(zio_t *zio, vdev_t *vd)
1241{
1242 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
1243 NULL, NULL, ZIO_PRIORITY_NOW,
1244 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1245}
1246
1247zio_t *
1248zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
1249{
1250
1251 ASSERT(vd->vdev_ops->vdev_op_leaf);
1252
1253 return (zio_create(zio, spa, 0, NULL, NULL, size, size, NULL, NULL,
1254 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE |
1255 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY,
1256 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE));
1257}
1258
1259void
1260zio_shrink(zio_t *zio, uint64_t size)
1261{
1262 ASSERT3P(zio->io_executor, ==, NULL);
1263 ASSERT3P(zio->io_orig_size, ==, zio->io_size);
1264 ASSERT3U(size, <=, zio->io_size);
1265
1266 /*
1267 * We don't shrink for raidz because of problems with the
1268 * reconstruction when reading back less than the block size.
1269 * Note, BP_IS_RAIDZ() assumes no compression.
1270 */
1271 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1272 if (!BP_IS_RAIDZ(zio->io_bp)) {
1273 /* we are not doing a raw write */
1274 ASSERT3U(zio->io_size, ==, zio->io_lsize);
1275 zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1276 }
1277}
1278
1279/*
1280 * ==========================================================================
1281 * Prepare to read and write logical blocks
1282 * ==========================================================================
1283 */
1284
1285static int
1286zio_read_bp_init(zio_t *zio)
1287{
1288 blkptr_t *bp = zio->io_bp;
1289
1290 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1291
1292 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1293 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1294 !(zio->io_flags & ZIO_FLAG_RAW)) {
1295 uint64_t psize =
1296 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1297 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1298 psize, psize, zio_decompress);
1299 }
1300
1301 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1302 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1303
1304 int psize = BPE_GET_PSIZE(bp);
1305 void *data = abd_borrow_buf(zio->io_abd, psize);
1306 decode_embedded_bp_compressed(bp, data);
1307 abd_return_buf_copy(zio->io_abd, data, psize);
1308 } else {
1309 ASSERT(!BP_IS_EMBEDDED(bp));
1310 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1311 }
1312
1313 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1314 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1315
1316 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1317 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1318
1319 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1320 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1321
1322 return (ZIO_PIPELINE_CONTINUE);
1323}
1324
1325static int
1326zio_write_bp_init(zio_t *zio)
1327{
1328 if (!IO_IS_ALLOCATING(zio))
1329 return (ZIO_PIPELINE_CONTINUE);
1330
1331 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1332
1333 if (zio->io_bp_override) {
1334 blkptr_t *bp = zio->io_bp;
1335 zio_prop_t *zp = &zio->io_prop;
1336
1337 ASSERT(bp->blk_birth != zio->io_txg);
1338 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1339
1340 *bp = *zio->io_bp_override;
1341 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1342
1343 if (BP_IS_EMBEDDED(bp))
1344 return (ZIO_PIPELINE_CONTINUE);
1345
1346 /*
1347 * If we've been overridden and nopwrite is set then
1348 * set the flag accordingly to indicate that a nopwrite
1349 * has already occurred.
1350 */
1351 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1352 ASSERT(!zp->zp_dedup);
1353 ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1354 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1355 return (ZIO_PIPELINE_CONTINUE);
1356 }
1357
1358 ASSERT(!zp->zp_nopwrite);
1359
1360 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1361 return (ZIO_PIPELINE_CONTINUE);
1362
1363 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1364 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1365
1366 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1367 BP_SET_DEDUP(bp, 1);
1368 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1369 return (ZIO_PIPELINE_CONTINUE);
1370 }
1371
1372 /*
1373 * We were unable to handle this as an override bp, treat
1374 * it as a regular write I/O.
1375 */
1376 zio->io_bp_override = NULL;
1377 *bp = zio->io_bp_orig;
1378 zio->io_pipeline = zio->io_orig_pipeline;
1379 }
1380
1381 return (ZIO_PIPELINE_CONTINUE);
1382}
1383
1384static int
1385zio_write_compress(zio_t *zio)
1386{
1387 spa_t *spa = zio->io_spa;
1388 zio_prop_t *zp = &zio->io_prop;
1389 enum zio_compress compress = zp->zp_compress;
1390 blkptr_t *bp = zio->io_bp;
1391 uint64_t lsize = zio->io_lsize;
1392 uint64_t psize = zio->io_size;
1393 int pass = 1;
1394
1395 EQUIV(lsize != psize, (zio->io_flags & ZIO_FLAG_RAW) != 0);
1396
1397 /*
1398 * If our children haven't all reached the ready stage,
1399 * wait for them and then repeat this pipeline stage.
1400 */
1401 if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT |
1402 ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) {
1403 return (ZIO_PIPELINE_STOP);
1404 }
1405
1406 if (!IO_IS_ALLOCATING(zio))
1407 return (ZIO_PIPELINE_CONTINUE);
1408
1409 if (zio->io_children_ready != NULL) {
1410 /*
1411 * Now that all our children are ready, run the callback
1412 * associated with this zio in case it wants to modify the
1413 * data to be written.
1414 */
1415 ASSERT3U(zp->zp_level, >, 0);
1416 zio->io_children_ready(zio);
1417 }
1418
1419 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1420 ASSERT(zio->io_bp_override == NULL);
1421
1422 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1423 /*
1424 * We're rewriting an existing block, which means we're
1425 * working on behalf of spa_sync(). For spa_sync() to
1426 * converge, it must eventually be the case that we don't
1427 * have to allocate new blocks. But compression changes
1428 * the blocksize, which forces a reallocate, and makes
1429 * convergence take longer. Therefore, after the first
1430 * few passes, stop compressing to ensure convergence.
1431 */
1432 pass = spa_sync_pass(spa);
1433
1434 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1435 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1436 ASSERT(!BP_GET_DEDUP(bp));
1437
1438 if (pass >= zfs_sync_pass_dont_compress)
1439 compress = ZIO_COMPRESS_OFF;
1440
1441 /* Make sure someone doesn't change their mind on overwrites */
1442 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1443 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1444 }
1445
1446 /* If it's a compressed write that is not raw, compress the buffer. */
1447 if (compress != ZIO_COMPRESS_OFF && psize == lsize) {
1448 void *cbuf = zio_buf_alloc(lsize);
1449 psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize);
1450 if (psize == 0 || psize == lsize) {
1451 compress = ZIO_COMPRESS_OFF;
1452 zio_buf_free(cbuf, lsize);
1453 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1454 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1455 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1456 encode_embedded_bp_compressed(bp,
1457 cbuf, compress, lsize, psize);
1458 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1459 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1460 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1461 zio_buf_free(cbuf, lsize);
1462 bp->blk_birth = zio->io_txg;
1463 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1464 ASSERT(spa_feature_is_active(spa,
1465 SPA_FEATURE_EMBEDDED_DATA));
1466 return (ZIO_PIPELINE_CONTINUE);
1467 } else {
1468 /*
1469 * Round up compressed size up to the ashift
1470 * of the smallest-ashift device, and zero the tail.
1471 * This ensures that the compressed size of the BP
1472 * (and thus compressratio property) are correct,
1473 * in that we charge for the padding used to fill out
1474 * the last sector.
1475 */
1476 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1477 size_t rounded = (size_t)P2ROUNDUP(psize,
1478 1ULL << spa->spa_min_ashift);
1479 if (rounded >= lsize) {
1480 compress = ZIO_COMPRESS_OFF;
1481 zio_buf_free(cbuf, lsize);
1482 psize = lsize;
1483 } else {
1484 abd_t *cdata = abd_get_from_buf(cbuf, lsize);
1485 abd_take_ownership_of_buf(cdata, B_TRUE);
1486 abd_zero_off(cdata, psize, rounded - psize);
1487 psize = rounded;
1488 zio_push_transform(zio, cdata,
1489 psize, lsize, NULL);
1490 }
1491 }
1492
1493 /*
1494 * We were unable to handle this as an override bp, treat
1495 * it as a regular write I/O.
1496 */
1497 zio->io_bp_override = NULL;
1498 *bp = zio->io_bp_orig;
1499 zio->io_pipeline = zio->io_orig_pipeline;
1500 } else {
1501 ASSERT3U(psize, !=, 0);
1502 }
1503
1504 /*
1505 * The final pass of spa_sync() must be all rewrites, but the first
1506 * few passes offer a trade-off: allocating blocks defers convergence,
1507 * but newly allocated blocks are sequential, so they can be written
1508 * to disk faster. Therefore, we allow the first few passes of
1509 * spa_sync() to allocate new blocks, but force rewrites after that.
1510 * There should only be a handful of blocks after pass 1 in any case.
1511 */
1512 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1513 BP_GET_PSIZE(bp) == psize &&
1514 pass >= zfs_sync_pass_rewrite) {
1515 ASSERT(psize != 0);
1516 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1517 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1518 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1519 } else {
1520 BP_ZERO(bp);
1521 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1522 }
1523
1524 if (psize == 0) {
1525 if (zio->io_bp_orig.blk_birth != 0 &&
1526 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1527 BP_SET_LSIZE(bp, lsize);
1528 BP_SET_TYPE(bp, zp->zp_type);
1529 BP_SET_LEVEL(bp, zp->zp_level);
1530 BP_SET_BIRTH(bp, zio->io_txg, 0);
1531 }
1532 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1533 } else {
1534 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1535 BP_SET_LSIZE(bp, lsize);
1536 BP_SET_TYPE(bp, zp->zp_type);
1537 BP_SET_LEVEL(bp, zp->zp_level);
1538 BP_SET_PSIZE(bp, psize);
1539 BP_SET_COMPRESS(bp, compress);
1540 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1541 BP_SET_DEDUP(bp, zp->zp_dedup);
1542 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1543 if (zp->zp_dedup) {
1544 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1545 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1546 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1547 }
1548 if (zp->zp_nopwrite) {
1549 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1550 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1551 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1552 }
1553 }
1554 return (ZIO_PIPELINE_CONTINUE);
1555}
1556
1557static int
1558zio_free_bp_init(zio_t *zio)
1559{
1560 blkptr_t *bp = zio->io_bp;
1561
1562 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1563 if (BP_GET_DEDUP(bp))
1564 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1565 }
1566
1567 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1568
1569 return (ZIO_PIPELINE_CONTINUE);
1570}
1571
1572/*
1573 * ==========================================================================
1574 * Execute the I/O pipeline
1575 * ==========================================================================
1576 */
1577
1578static void
1579zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1580{
1581 spa_t *spa = zio->io_spa;
1582 zio_type_t t = zio->io_type;
1583 int flags = (cutinline ? TQ_FRONT : 0);
1584
1585 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1586
1587 /*
1588 * If we're a config writer or a probe, the normal issue and
1589 * interrupt threads may all be blocked waiting for the config lock.
1590 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1591 */
1592 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1593 t = ZIO_TYPE_NULL;
1594
1595 /*
1596 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1597 */
1598 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1599 t = ZIO_TYPE_NULL;
1600
1601 /*
1602 * If this is a high priority I/O, then use the high priority taskq if
1603 * available.
1604 */
1605 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1606 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1607 q++;
1608
1609 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1610
1611 /*
1612 * NB: We are assuming that the zio can only be dispatched
1613 * to a single taskq at a time. It would be a grievous error
1614 * to dispatch the zio to another taskq at the same time.
1615 */
1616#if defined(illumos) || !defined(_KERNEL)
1617 ASSERT(zio->io_tqent.tqent_next == NULL);
1618#else
1619 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1620#endif
1621 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1622 flags, &zio->io_tqent);
1623}
1624
1625static boolean_t
1626zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1627{
1628 kthread_t *executor = zio->io_executor;
1629 spa_t *spa = zio->io_spa;
1630
1631 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1632 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1633 uint_t i;
1634 for (i = 0; i < tqs->stqs_count; i++) {
1635 if (taskq_member(tqs->stqs_taskq[i], executor))
1636 return (B_TRUE);
1637 }
1638 }
1639
1640 return (B_FALSE);
1641}
1642
1643static int
1644zio_issue_async(zio_t *zio)
1645{
1646 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1647
1648 return (ZIO_PIPELINE_STOP);
1649}
1650
1651void
1652zio_interrupt(zio_t *zio)
1653{
1654 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1655}
1656
1657void
1658zio_delay_interrupt(zio_t *zio)
1659{
1660 /*
1661 * The timeout_generic() function isn't defined in userspace, so
1662 * rather than trying to implement the function, the zio delay
1663 * functionality has been disabled for userspace builds.
1664 */
1665
1666#ifdef _KERNEL
1667 /*
1668 * If io_target_timestamp is zero, then no delay has been registered
1669 * for this IO, thus jump to the end of this function and "skip" the
1670 * delay; issuing it directly to the zio layer.
1671 */
1672 if (zio->io_target_timestamp != 0) {
1673 hrtime_t now = gethrtime();
1674
1675 if (now >= zio->io_target_timestamp) {
1676 /*
1677 * This IO has already taken longer than the target
1678 * delay to complete, so we don't want to delay it
1679 * any longer; we "miss" the delay and issue it
1680 * directly to the zio layer. This is likely due to
1681 * the target latency being set to a value less than
1682 * the underlying hardware can satisfy (e.g. delay
1683 * set to 1ms, but the disks take 10ms to complete an
1684 * IO request).
1685 */
1686
1687 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1688 hrtime_t, now);
1689
1690 zio_interrupt(zio);
1691 } else {
1692 hrtime_t diff = zio->io_target_timestamp - now;
1693
1694 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1695 hrtime_t, now, hrtime_t, diff);
1696
1697 (void) timeout_generic(CALLOUT_NORMAL,
1698 (void (*)(void *))zio_interrupt, zio, diff, 1, 0);
1699 }
1700
1701 return;
1702 }
1703#endif
1704
1705 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
1706 zio_interrupt(zio);
1707}
1708
1709/*
1710 * Execute the I/O pipeline until one of the following occurs:
1711 *
1712 * (1) the I/O completes
1713 * (2) the pipeline stalls waiting for dependent child I/Os
1714 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1715 * (4) the I/O is delegated by vdev-level caching or aggregation
1716 * (5) the I/O is deferred due to vdev-level queueing
1717 * (6) the I/O is handed off to another thread.
1718 *
1719 * In all cases, the pipeline stops whenever there's no CPU work; it never
1720 * burns a thread in cv_wait().
1721 *
1722 * There's no locking on io_stage because there's no legitimate way
1723 * for multiple threads to be attempting to process the same I/O.
1724 */
1725static zio_pipe_stage_t *zio_pipeline[];
1726
1727void
1728zio_execute(zio_t *zio)
1729{
1730 zio->io_executor = curthread;
1731
1732 ASSERT3U(zio->io_queued_timestamp, >, 0);
1733
1734 while (zio->io_stage < ZIO_STAGE_DONE) {
1735 enum zio_stage pipeline = zio->io_pipeline;
1736 enum zio_stage stage = zio->io_stage;
1737 int rv;
1738
1739 ASSERT(!MUTEX_HELD(&zio->io_lock));
1740 ASSERT(ISP2(stage));
1741 ASSERT(zio->io_stall == NULL);
1742
1743 do {
1744 stage <<= 1;
1745 } while ((stage & pipeline) == 0);
1746
1747 ASSERT(stage <= ZIO_STAGE_DONE);
1748
1749 /*
1750 * If we are in interrupt context and this pipeline stage
1751 * will grab a config lock that is held across I/O,
1752 * or may wait for an I/O that needs an interrupt thread
1753 * to complete, issue async to avoid deadlock.
1754 *
1755 * For VDEV_IO_START, we cut in line so that the io will
1756 * be sent to disk promptly.
1757 */
1758 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1759 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1760 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1761 zio_requeue_io_start_cut_in_line : B_FALSE;
1762 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1763 return;
1764 }
1765
1766 zio->io_stage = stage;
1767 zio->io_pipeline_trace |= zio->io_stage;
1768 rv = zio_pipeline[highbit64(stage) - 1](zio);
1769
1770 if (rv == ZIO_PIPELINE_STOP)
1771 return;
1772
1773 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1774 }
1775}
1776
1777/*
1778 * ==========================================================================
1779 * Initiate I/O, either sync or async
1780 * ==========================================================================
1781 */
1782int
1783zio_wait(zio_t *zio)
1784{
1785 int error;
1786
1787 ASSERT3P(zio->io_stage, ==, ZIO_STAGE_OPEN);
1788 ASSERT3P(zio->io_executor, ==, NULL);
1789
1790 zio->io_waiter = curthread;
1791 ASSERT0(zio->io_queued_timestamp);
1792 zio->io_queued_timestamp = gethrtime();
1793
1794 zio_execute(zio);
1795
1796 mutex_enter(&zio->io_lock);
1797 while (zio->io_executor != NULL)
1798 cv_wait(&zio->io_cv, &zio->io_lock);
1799 mutex_exit(&zio->io_lock);
1800
1801 error = zio->io_error;
1802 zio_destroy(zio);
1803
1804 return (error);
1805}
1806
1807void
1808zio_nowait(zio_t *zio)
1809{
1810 ASSERT3P(zio->io_executor, ==, NULL);
1811
1812 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1813 zio_unique_parent(zio) == NULL) {
1814 /*
1815 * This is a logical async I/O with no parent to wait for it.
1816 * We add it to the spa_async_root_zio "Godfather" I/O which
1817 * will ensure they complete prior to unloading the pool.
1818 */
1819 spa_t *spa = zio->io_spa;
1820
1821 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1822 }
1823
1824 ASSERT0(zio->io_queued_timestamp);
1825 zio->io_queued_timestamp = gethrtime();
1826 zio_execute(zio);
1827}
1828
1829/*
1830 * ==========================================================================
1831 * Reexecute, cancel, or suspend/resume failed I/O
1832 * ==========================================================================
1833 */
1834
1835static void
1836zio_reexecute(zio_t *pio)
1837{
1838 zio_t *cio, *cio_next;
1839
1840 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1841 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1842 ASSERT(pio->io_gang_leader == NULL);
1843 ASSERT(pio->io_gang_tree == NULL);
1844
1845 pio->io_flags = pio->io_orig_flags;
1846 pio->io_stage = pio->io_orig_stage;
1847 pio->io_pipeline = pio->io_orig_pipeline;
1848 pio->io_reexecute = 0;
1849 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1850 pio->io_pipeline_trace = 0;
1851 pio->io_error = 0;
1852 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1853 pio->io_state[w] = 0;
1854 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1855 pio->io_child_error[c] = 0;
1856
1857 if (IO_IS_ALLOCATING(pio))
1858 BP_ZERO(pio->io_bp);
1859
1860 /*
1861 * As we reexecute pio's children, new children could be created.
1862 * New children go to the head of pio's io_child_list, however,
1863 * so we will (correctly) not reexecute them. The key is that
1864 * the remainder of pio's io_child_list, from 'cio_next' onward,
1865 * cannot be affected by any side effects of reexecuting 'cio'.
1866 */
1867 zio_link_t *zl = NULL;
| 994
995 if (zfs_trim_enabled)
996 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
997 ZIO_STAGE_VDEV_IO_ASSESS;
998 /*
999 * GANG and DEDUP blocks can induce a read (for the gang block header,
1000 * or the DDT), so issue them asynchronously so that this thread is
1001 * not tied up.
1002 */
1003 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
1004 stage |= ZIO_STAGE_ISSUE_ASYNC;
1005
1006 flags |= ZIO_FLAG_DONT_QUEUE;
1007
1008 zio = zio_create(pio, spa, txg, bp, NULL, size,
1009 size, NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
1010 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
1011
1012 return (zio);
1013}
1014
1015zio_t *
1016zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1017 zio_done_func_t *done, void *private, enum zio_flag flags)
1018{
1019 zio_t *zio;
1020
1021 zfs_blkptr_verify(spa, bp);
1022
1023 if (BP_IS_EMBEDDED(bp))
1024 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1025
1026 /*
1027 * A claim is an allocation of a specific block. Claims are needed
1028 * to support immediate writes in the intent log. The issue is that
1029 * immediate writes contain committed data, but in a txg that was
1030 * *not* committed. Upon opening the pool after an unclean shutdown,
1031 * the intent log claims all blocks that contain immediate write data
1032 * so that the SPA knows they're in use.
1033 *
1034 * All claims *must* be resolved in the first txg -- before the SPA
1035 * starts allocating blocks -- so that nothing is allocated twice.
1036 * If txg == 0 we just verify that the block is claimable.
1037 */
1038 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <,
1039 spa_min_claim_txg(spa));
1040 ASSERT(txg == spa_min_claim_txg(spa) || txg == 0);
1041 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
1042
1043 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1044 BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
1045 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
1046 ASSERT0(zio->io_queued_timestamp);
1047
1048 return (zio);
1049}
1050
1051zio_t *
1052zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
1053 uint64_t size, zio_done_func_t *done, void *private,
1054 zio_priority_t priority, enum zio_flag flags)
1055{
1056 zio_t *zio;
1057 int c;
1058
1059 if (vd->vdev_children == 0) {
1060 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
1061 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
1062 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
1063
1064 zio->io_cmd = cmd;
1065 } else {
1066 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
1067
1068 for (c = 0; c < vd->vdev_children; c++)
1069 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
1070 offset, size, done, private, priority, flags));
1071 }
1072
1073 return (zio);
1074}
1075
1076zio_t *
1077zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1078 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1079 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1080{
1081 zio_t *zio;
1082
1083 ASSERT(vd->vdev_children == 0);
1084 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1085 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1086 ASSERT3U(offset + size, <=, vd->vdev_psize);
1087
1088 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1089 private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1090 offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1091
1092 zio->io_prop.zp_checksum = checksum;
1093
1094 return (zio);
1095}
1096
1097zio_t *
1098zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1099 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1100 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1101{
1102 zio_t *zio;
1103
1104 ASSERT(vd->vdev_children == 0);
1105 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1106 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1107 ASSERT3U(offset + size, <=, vd->vdev_psize);
1108
1109 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1110 private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1111 offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1112
1113 zio->io_prop.zp_checksum = checksum;
1114
1115 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1116 /*
1117 * zec checksums are necessarily destructive -- they modify
1118 * the end of the write buffer to hold the verifier/checksum.
1119 * Therefore, we must make a local copy in case the data is
1120 * being written to multiple places in parallel.
1121 */
1122 abd_t *wbuf = abd_alloc_sametype(data, size);
1123 abd_copy(wbuf, data, size);
1124
1125 zio_push_transform(zio, wbuf, size, size, NULL);
1126 }
1127
1128 return (zio);
1129}
1130
1131/*
1132 * Create a child I/O to do some work for us.
1133 */
1134zio_t *
1135zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1136 abd_t *data, uint64_t size, int type, zio_priority_t priority,
1137 enum zio_flag flags, zio_done_func_t *done, void *private)
1138{
1139 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1140 zio_t *zio;
1141
1142 /*
1143 * vdev child I/Os do not propagate their error to the parent.
1144 * Therefore, for correct operation the caller *must* check for
1145 * and handle the error in the child i/o's done callback.
1146 * The only exceptions are i/os that we don't care about
1147 * (OPTIONAL or REPAIR).
1148 */
1149 ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) ||
1150 done != NULL);
1151
1152 /*
1153 * In the common case, where the parent zio was to a normal vdev,
1154 * the child zio must be to a child vdev of that vdev. Otherwise,
1155 * the child zio must be to a top-level vdev.
1156 */
1157 if (pio->io_vd != NULL && pio->io_vd->vdev_ops != &vdev_indirect_ops) {
1158 ASSERT3P(vd->vdev_parent, ==, pio->io_vd);
1159 } else {
1160 ASSERT3P(vd, ==, vd->vdev_top);
1161 }
1162
1163 if (type == ZIO_TYPE_READ && bp != NULL) {
1164 /*
1165 * If we have the bp, then the child should perform the
1166 * checksum and the parent need not. This pushes error
1167 * detection as close to the leaves as possible and
1168 * eliminates redundant checksums in the interior nodes.
1169 */
1170 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1171 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1172 }
1173
1174 /* Not all IO types require vdev io done stage e.g. free */
1175 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE))
1176 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE;
1177
1178 if (vd->vdev_ops->vdev_op_leaf) {
1179 ASSERT0(vd->vdev_children);
1180 offset += VDEV_LABEL_START_SIZE;
1181 }
1182
1183 flags |= ZIO_VDEV_CHILD_FLAGS(pio);
1184
1185 /*
1186 * If we've decided to do a repair, the write is not speculative --
1187 * even if the original read was.
1188 */
1189 if (flags & ZIO_FLAG_IO_REPAIR)
1190 flags &= ~ZIO_FLAG_SPECULATIVE;
1191
1192 /*
1193 * If we're creating a child I/O that is not associated with a
1194 * top-level vdev, then the child zio is not an allocating I/O.
1195 * If this is a retried I/O then we ignore it since we will
1196 * have already processed the original allocating I/O.
1197 */
1198 if (flags & ZIO_FLAG_IO_ALLOCATING &&
1199 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1200 metaslab_class_t *mc = spa_normal_class(pio->io_spa);
1201
1202 ASSERT(mc->mc_alloc_throttle_enabled);
1203 ASSERT(type == ZIO_TYPE_WRITE);
1204 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1205 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1206 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1207 pio->io_child_type == ZIO_CHILD_GANG);
1208
1209 flags &= ~ZIO_FLAG_IO_ALLOCATING;
1210 }
1211
1212 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1213 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1214 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1215 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1216
1217 zio->io_physdone = pio->io_physdone;
1218 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1219 zio->io_logical->io_phys_children++;
1220
1221 return (zio);
1222}
1223
1224zio_t *
1225zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size,
1226 int type, zio_priority_t priority, enum zio_flag flags,
1227 zio_done_func_t *done, void *private)
1228{
1229 zio_t *zio;
1230
1231 ASSERT(vd->vdev_ops->vdev_op_leaf);
1232
1233 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1234 data, size, size, done, private, type, priority,
1235 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1236 vd, offset, NULL,
1237 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1238
1239 return (zio);
1240}
1241
1242void
1243zio_flush(zio_t *zio, vdev_t *vd)
1244{
1245 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
1246 NULL, NULL, ZIO_PRIORITY_NOW,
1247 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1248}
1249
1250zio_t *
1251zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
1252{
1253
1254 ASSERT(vd->vdev_ops->vdev_op_leaf);
1255
1256 return (zio_create(zio, spa, 0, NULL, NULL, size, size, NULL, NULL,
1257 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE |
1258 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY,
1259 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE));
1260}
1261
1262void
1263zio_shrink(zio_t *zio, uint64_t size)
1264{
1265 ASSERT3P(zio->io_executor, ==, NULL);
1266 ASSERT3P(zio->io_orig_size, ==, zio->io_size);
1267 ASSERT3U(size, <=, zio->io_size);
1268
1269 /*
1270 * We don't shrink for raidz because of problems with the
1271 * reconstruction when reading back less than the block size.
1272 * Note, BP_IS_RAIDZ() assumes no compression.
1273 */
1274 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1275 if (!BP_IS_RAIDZ(zio->io_bp)) {
1276 /* we are not doing a raw write */
1277 ASSERT3U(zio->io_size, ==, zio->io_lsize);
1278 zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1279 }
1280}
1281
1282/*
1283 * ==========================================================================
1284 * Prepare to read and write logical blocks
1285 * ==========================================================================
1286 */
1287
1288static int
1289zio_read_bp_init(zio_t *zio)
1290{
1291 blkptr_t *bp = zio->io_bp;
1292
1293 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1294
1295 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1296 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1297 !(zio->io_flags & ZIO_FLAG_RAW)) {
1298 uint64_t psize =
1299 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1300 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1301 psize, psize, zio_decompress);
1302 }
1303
1304 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1305 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1306
1307 int psize = BPE_GET_PSIZE(bp);
1308 void *data = abd_borrow_buf(zio->io_abd, psize);
1309 decode_embedded_bp_compressed(bp, data);
1310 abd_return_buf_copy(zio->io_abd, data, psize);
1311 } else {
1312 ASSERT(!BP_IS_EMBEDDED(bp));
1313 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1314 }
1315
1316 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1317 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1318
1319 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1320 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1321
1322 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1323 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1324
1325 return (ZIO_PIPELINE_CONTINUE);
1326}
1327
1328static int
1329zio_write_bp_init(zio_t *zio)
1330{
1331 if (!IO_IS_ALLOCATING(zio))
1332 return (ZIO_PIPELINE_CONTINUE);
1333
1334 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1335
1336 if (zio->io_bp_override) {
1337 blkptr_t *bp = zio->io_bp;
1338 zio_prop_t *zp = &zio->io_prop;
1339
1340 ASSERT(bp->blk_birth != zio->io_txg);
1341 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1342
1343 *bp = *zio->io_bp_override;
1344 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1345
1346 if (BP_IS_EMBEDDED(bp))
1347 return (ZIO_PIPELINE_CONTINUE);
1348
1349 /*
1350 * If we've been overridden and nopwrite is set then
1351 * set the flag accordingly to indicate that a nopwrite
1352 * has already occurred.
1353 */
1354 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1355 ASSERT(!zp->zp_dedup);
1356 ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1357 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1358 return (ZIO_PIPELINE_CONTINUE);
1359 }
1360
1361 ASSERT(!zp->zp_nopwrite);
1362
1363 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1364 return (ZIO_PIPELINE_CONTINUE);
1365
1366 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1367 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1368
1369 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1370 BP_SET_DEDUP(bp, 1);
1371 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1372 return (ZIO_PIPELINE_CONTINUE);
1373 }
1374
1375 /*
1376 * We were unable to handle this as an override bp, treat
1377 * it as a regular write I/O.
1378 */
1379 zio->io_bp_override = NULL;
1380 *bp = zio->io_bp_orig;
1381 zio->io_pipeline = zio->io_orig_pipeline;
1382 }
1383
1384 return (ZIO_PIPELINE_CONTINUE);
1385}
1386
1387static int
1388zio_write_compress(zio_t *zio)
1389{
1390 spa_t *spa = zio->io_spa;
1391 zio_prop_t *zp = &zio->io_prop;
1392 enum zio_compress compress = zp->zp_compress;
1393 blkptr_t *bp = zio->io_bp;
1394 uint64_t lsize = zio->io_lsize;
1395 uint64_t psize = zio->io_size;
1396 int pass = 1;
1397
1398 EQUIV(lsize != psize, (zio->io_flags & ZIO_FLAG_RAW) != 0);
1399
1400 /*
1401 * If our children haven't all reached the ready stage,
1402 * wait for them and then repeat this pipeline stage.
1403 */
1404 if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT |
1405 ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) {
1406 return (ZIO_PIPELINE_STOP);
1407 }
1408
1409 if (!IO_IS_ALLOCATING(zio))
1410 return (ZIO_PIPELINE_CONTINUE);
1411
1412 if (zio->io_children_ready != NULL) {
1413 /*
1414 * Now that all our children are ready, run the callback
1415 * associated with this zio in case it wants to modify the
1416 * data to be written.
1417 */
1418 ASSERT3U(zp->zp_level, >, 0);
1419 zio->io_children_ready(zio);
1420 }
1421
1422 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1423 ASSERT(zio->io_bp_override == NULL);
1424
1425 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1426 /*
1427 * We're rewriting an existing block, which means we're
1428 * working on behalf of spa_sync(). For spa_sync() to
1429 * converge, it must eventually be the case that we don't
1430 * have to allocate new blocks. But compression changes
1431 * the blocksize, which forces a reallocate, and makes
1432 * convergence take longer. Therefore, after the first
1433 * few passes, stop compressing to ensure convergence.
1434 */
1435 pass = spa_sync_pass(spa);
1436
1437 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1438 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1439 ASSERT(!BP_GET_DEDUP(bp));
1440
1441 if (pass >= zfs_sync_pass_dont_compress)
1442 compress = ZIO_COMPRESS_OFF;
1443
1444 /* Make sure someone doesn't change their mind on overwrites */
1445 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1446 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1447 }
1448
1449 /* If it's a compressed write that is not raw, compress the buffer. */
1450 if (compress != ZIO_COMPRESS_OFF && psize == lsize) {
1451 void *cbuf = zio_buf_alloc(lsize);
1452 psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize);
1453 if (psize == 0 || psize == lsize) {
1454 compress = ZIO_COMPRESS_OFF;
1455 zio_buf_free(cbuf, lsize);
1456 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1457 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1458 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1459 encode_embedded_bp_compressed(bp,
1460 cbuf, compress, lsize, psize);
1461 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1462 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1463 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1464 zio_buf_free(cbuf, lsize);
1465 bp->blk_birth = zio->io_txg;
1466 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1467 ASSERT(spa_feature_is_active(spa,
1468 SPA_FEATURE_EMBEDDED_DATA));
1469 return (ZIO_PIPELINE_CONTINUE);
1470 } else {
1471 /*
1472 * Round up compressed size up to the ashift
1473 * of the smallest-ashift device, and zero the tail.
1474 * This ensures that the compressed size of the BP
1475 * (and thus compressratio property) are correct,
1476 * in that we charge for the padding used to fill out
1477 * the last sector.
1478 */
1479 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1480 size_t rounded = (size_t)P2ROUNDUP(psize,
1481 1ULL << spa->spa_min_ashift);
1482 if (rounded >= lsize) {
1483 compress = ZIO_COMPRESS_OFF;
1484 zio_buf_free(cbuf, lsize);
1485 psize = lsize;
1486 } else {
1487 abd_t *cdata = abd_get_from_buf(cbuf, lsize);
1488 abd_take_ownership_of_buf(cdata, B_TRUE);
1489 abd_zero_off(cdata, psize, rounded - psize);
1490 psize = rounded;
1491 zio_push_transform(zio, cdata,
1492 psize, lsize, NULL);
1493 }
1494 }
1495
1496 /*
1497 * We were unable to handle this as an override bp, treat
1498 * it as a regular write I/O.
1499 */
1500 zio->io_bp_override = NULL;
1501 *bp = zio->io_bp_orig;
1502 zio->io_pipeline = zio->io_orig_pipeline;
1503 } else {
1504 ASSERT3U(psize, !=, 0);
1505 }
1506
1507 /*
1508 * The final pass of spa_sync() must be all rewrites, but the first
1509 * few passes offer a trade-off: allocating blocks defers convergence,
1510 * but newly allocated blocks are sequential, so they can be written
1511 * to disk faster. Therefore, we allow the first few passes of
1512 * spa_sync() to allocate new blocks, but force rewrites after that.
1513 * There should only be a handful of blocks after pass 1 in any case.
1514 */
1515 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1516 BP_GET_PSIZE(bp) == psize &&
1517 pass >= zfs_sync_pass_rewrite) {
1518 ASSERT(psize != 0);
1519 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1520 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1521 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1522 } else {
1523 BP_ZERO(bp);
1524 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1525 }
1526
1527 if (psize == 0) {
1528 if (zio->io_bp_orig.blk_birth != 0 &&
1529 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1530 BP_SET_LSIZE(bp, lsize);
1531 BP_SET_TYPE(bp, zp->zp_type);
1532 BP_SET_LEVEL(bp, zp->zp_level);
1533 BP_SET_BIRTH(bp, zio->io_txg, 0);
1534 }
1535 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1536 } else {
1537 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1538 BP_SET_LSIZE(bp, lsize);
1539 BP_SET_TYPE(bp, zp->zp_type);
1540 BP_SET_LEVEL(bp, zp->zp_level);
1541 BP_SET_PSIZE(bp, psize);
1542 BP_SET_COMPRESS(bp, compress);
1543 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1544 BP_SET_DEDUP(bp, zp->zp_dedup);
1545 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1546 if (zp->zp_dedup) {
1547 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1548 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1549 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1550 }
1551 if (zp->zp_nopwrite) {
1552 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1553 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1554 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1555 }
1556 }
1557 return (ZIO_PIPELINE_CONTINUE);
1558}
1559
1560static int
1561zio_free_bp_init(zio_t *zio)
1562{
1563 blkptr_t *bp = zio->io_bp;
1564
1565 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1566 if (BP_GET_DEDUP(bp))
1567 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1568 }
1569
1570 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1571
1572 return (ZIO_PIPELINE_CONTINUE);
1573}
1574
1575/*
1576 * ==========================================================================
1577 * Execute the I/O pipeline
1578 * ==========================================================================
1579 */
1580
1581static void
1582zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1583{
1584 spa_t *spa = zio->io_spa;
1585 zio_type_t t = zio->io_type;
1586 int flags = (cutinline ? TQ_FRONT : 0);
1587
1588 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1589
1590 /*
1591 * If we're a config writer or a probe, the normal issue and
1592 * interrupt threads may all be blocked waiting for the config lock.
1593 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1594 */
1595 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1596 t = ZIO_TYPE_NULL;
1597
1598 /*
1599 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1600 */
1601 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1602 t = ZIO_TYPE_NULL;
1603
1604 /*
1605 * If this is a high priority I/O, then use the high priority taskq if
1606 * available.
1607 */
1608 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1609 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1610 q++;
1611
1612 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1613
1614 /*
1615 * NB: We are assuming that the zio can only be dispatched
1616 * to a single taskq at a time. It would be a grievous error
1617 * to dispatch the zio to another taskq at the same time.
1618 */
1619#if defined(illumos) || !defined(_KERNEL)
1620 ASSERT(zio->io_tqent.tqent_next == NULL);
1621#else
1622 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1623#endif
1624 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1625 flags, &zio->io_tqent);
1626}
1627
1628static boolean_t
1629zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1630{
1631 kthread_t *executor = zio->io_executor;
1632 spa_t *spa = zio->io_spa;
1633
1634 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1635 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1636 uint_t i;
1637 for (i = 0; i < tqs->stqs_count; i++) {
1638 if (taskq_member(tqs->stqs_taskq[i], executor))
1639 return (B_TRUE);
1640 }
1641 }
1642
1643 return (B_FALSE);
1644}
1645
1646static int
1647zio_issue_async(zio_t *zio)
1648{
1649 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1650
1651 return (ZIO_PIPELINE_STOP);
1652}
1653
1654void
1655zio_interrupt(zio_t *zio)
1656{
1657 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1658}
1659
1660void
1661zio_delay_interrupt(zio_t *zio)
1662{
1663 /*
1664 * The timeout_generic() function isn't defined in userspace, so
1665 * rather than trying to implement the function, the zio delay
1666 * functionality has been disabled for userspace builds.
1667 */
1668
1669#ifdef _KERNEL
1670 /*
1671 * If io_target_timestamp is zero, then no delay has been registered
1672 * for this IO, thus jump to the end of this function and "skip" the
1673 * delay; issuing it directly to the zio layer.
1674 */
1675 if (zio->io_target_timestamp != 0) {
1676 hrtime_t now = gethrtime();
1677
1678 if (now >= zio->io_target_timestamp) {
1679 /*
1680 * This IO has already taken longer than the target
1681 * delay to complete, so we don't want to delay it
1682 * any longer; we "miss" the delay and issue it
1683 * directly to the zio layer. This is likely due to
1684 * the target latency being set to a value less than
1685 * the underlying hardware can satisfy (e.g. delay
1686 * set to 1ms, but the disks take 10ms to complete an
1687 * IO request).
1688 */
1689
1690 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1691 hrtime_t, now);
1692
1693 zio_interrupt(zio);
1694 } else {
1695 hrtime_t diff = zio->io_target_timestamp - now;
1696
1697 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1698 hrtime_t, now, hrtime_t, diff);
1699
1700 (void) timeout_generic(CALLOUT_NORMAL,
1701 (void (*)(void *))zio_interrupt, zio, diff, 1, 0);
1702 }
1703
1704 return;
1705 }
1706#endif
1707
1708 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
1709 zio_interrupt(zio);
1710}
1711
1712/*
1713 * Execute the I/O pipeline until one of the following occurs:
1714 *
1715 * (1) the I/O completes
1716 * (2) the pipeline stalls waiting for dependent child I/Os
1717 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1718 * (4) the I/O is delegated by vdev-level caching or aggregation
1719 * (5) the I/O is deferred due to vdev-level queueing
1720 * (6) the I/O is handed off to another thread.
1721 *
1722 * In all cases, the pipeline stops whenever there's no CPU work; it never
1723 * burns a thread in cv_wait().
1724 *
1725 * There's no locking on io_stage because there's no legitimate way
1726 * for multiple threads to be attempting to process the same I/O.
1727 */
1728static zio_pipe_stage_t *zio_pipeline[];
1729
1730void
1731zio_execute(zio_t *zio)
1732{
1733 zio->io_executor = curthread;
1734
1735 ASSERT3U(zio->io_queued_timestamp, >, 0);
1736
1737 while (zio->io_stage < ZIO_STAGE_DONE) {
1738 enum zio_stage pipeline = zio->io_pipeline;
1739 enum zio_stage stage = zio->io_stage;
1740 int rv;
1741
1742 ASSERT(!MUTEX_HELD(&zio->io_lock));
1743 ASSERT(ISP2(stage));
1744 ASSERT(zio->io_stall == NULL);
1745
1746 do {
1747 stage <<= 1;
1748 } while ((stage & pipeline) == 0);
1749
1750 ASSERT(stage <= ZIO_STAGE_DONE);
1751
1752 /*
1753 * If we are in interrupt context and this pipeline stage
1754 * will grab a config lock that is held across I/O,
1755 * or may wait for an I/O that needs an interrupt thread
1756 * to complete, issue async to avoid deadlock.
1757 *
1758 * For VDEV_IO_START, we cut in line so that the io will
1759 * be sent to disk promptly.
1760 */
1761 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1762 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1763 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1764 zio_requeue_io_start_cut_in_line : B_FALSE;
1765 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1766 return;
1767 }
1768
1769 zio->io_stage = stage;
1770 zio->io_pipeline_trace |= zio->io_stage;
1771 rv = zio_pipeline[highbit64(stage) - 1](zio);
1772
1773 if (rv == ZIO_PIPELINE_STOP)
1774 return;
1775
1776 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1777 }
1778}
1779
1780/*
1781 * ==========================================================================
1782 * Initiate I/O, either sync or async
1783 * ==========================================================================
1784 */
1785int
1786zio_wait(zio_t *zio)
1787{
1788 int error;
1789
1790 ASSERT3P(zio->io_stage, ==, ZIO_STAGE_OPEN);
1791 ASSERT3P(zio->io_executor, ==, NULL);
1792
1793 zio->io_waiter = curthread;
1794 ASSERT0(zio->io_queued_timestamp);
1795 zio->io_queued_timestamp = gethrtime();
1796
1797 zio_execute(zio);
1798
1799 mutex_enter(&zio->io_lock);
1800 while (zio->io_executor != NULL)
1801 cv_wait(&zio->io_cv, &zio->io_lock);
1802 mutex_exit(&zio->io_lock);
1803
1804 error = zio->io_error;
1805 zio_destroy(zio);
1806
1807 return (error);
1808}
1809
1810void
1811zio_nowait(zio_t *zio)
1812{
1813 ASSERT3P(zio->io_executor, ==, NULL);
1814
1815 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1816 zio_unique_parent(zio) == NULL) {
1817 /*
1818 * This is a logical async I/O with no parent to wait for it.
1819 * We add it to the spa_async_root_zio "Godfather" I/O which
1820 * will ensure they complete prior to unloading the pool.
1821 */
1822 spa_t *spa = zio->io_spa;
1823
1824 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1825 }
1826
1827 ASSERT0(zio->io_queued_timestamp);
1828 zio->io_queued_timestamp = gethrtime();
1829 zio_execute(zio);
1830}
1831
1832/*
1833 * ==========================================================================
1834 * Reexecute, cancel, or suspend/resume failed I/O
1835 * ==========================================================================
1836 */
1837
1838static void
1839zio_reexecute(zio_t *pio)
1840{
1841 zio_t *cio, *cio_next;
1842
1843 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1844 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1845 ASSERT(pio->io_gang_leader == NULL);
1846 ASSERT(pio->io_gang_tree == NULL);
1847
1848 pio->io_flags = pio->io_orig_flags;
1849 pio->io_stage = pio->io_orig_stage;
1850 pio->io_pipeline = pio->io_orig_pipeline;
1851 pio->io_reexecute = 0;
1852 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1853 pio->io_pipeline_trace = 0;
1854 pio->io_error = 0;
1855 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1856 pio->io_state[w] = 0;
1857 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1858 pio->io_child_error[c] = 0;
1859
1860 if (IO_IS_ALLOCATING(pio))
1861 BP_ZERO(pio->io_bp);
1862
1863 /*
1864 * As we reexecute pio's children, new children could be created.
1865 * New children go to the head of pio's io_child_list, however,
1866 * so we will (correctly) not reexecute them. The key is that
1867 * the remainder of pio's io_child_list, from 'cio_next' onward,
1868 * cannot be affected by any side effects of reexecuting 'cio'.
1869 */
1870 zio_link_t *zl = NULL;
|
| 1871 mutex_enter(&pio->io_lock);
|
1868 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
1869 cio_next = zio_walk_children(pio, &zl);
| 1872 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
1873 cio_next = zio_walk_children(pio, &zl);
|
1870 mutex_enter(&pio->io_lock);
| |
1871 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1872 pio->io_children[cio->io_child_type][w]++;
1873 mutex_exit(&pio->io_lock);
1874 zio_reexecute(cio);
| 1874 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1875 pio->io_children[cio->io_child_type][w]++;
1876 mutex_exit(&pio->io_lock);
1877 zio_reexecute(cio);
|
| 1878 mutex_enter(&pio->io_lock);
|
1875 }
| 1879 }
|
| 1880 mutex_exit(&pio->io_lock);
|
1876
1877 /*
1878 * Now that all children have been reexecuted, execute the parent.
1879 * We don't reexecute "The Godfather" I/O here as it's the
1880 * responsibility of the caller to wait on it.
1881 */
1882 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
1883 pio->io_queued_timestamp = gethrtime();
1884 zio_execute(pio);
1885 }
1886}
1887
1888void
1889zio_suspend(spa_t *spa, zio_t *zio)
1890{
1891 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1892 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1893 "failure and the failure mode property for this pool "
1894 "is set to panic.", spa_name(spa));
1895
1896 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1897
1898 mutex_enter(&spa->spa_suspend_lock);
1899
1900 if (spa->spa_suspend_zio_root == NULL)
1901 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1902 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1903 ZIO_FLAG_GODFATHER);
1904
1905 spa->spa_suspended = B_TRUE;
1906
1907 if (zio != NULL) {
1908 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1909 ASSERT(zio != spa->spa_suspend_zio_root);
1910 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1911 ASSERT(zio_unique_parent(zio) == NULL);
1912 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1913 zio_add_child(spa->spa_suspend_zio_root, zio);
1914 }
1915
1916 mutex_exit(&spa->spa_suspend_lock);
1917}
1918
1919int
1920zio_resume(spa_t *spa)
1921{
1922 zio_t *pio;
1923
1924 /*
1925 * Reexecute all previously suspended i/o.
1926 */
1927 mutex_enter(&spa->spa_suspend_lock);
1928 spa->spa_suspended = B_FALSE;
1929 cv_broadcast(&spa->spa_suspend_cv);
1930 pio = spa->spa_suspend_zio_root;
1931 spa->spa_suspend_zio_root = NULL;
1932 mutex_exit(&spa->spa_suspend_lock);
1933
1934 if (pio == NULL)
1935 return (0);
1936
1937 zio_reexecute(pio);
1938 return (zio_wait(pio));
1939}
1940
1941void
1942zio_resume_wait(spa_t *spa)
1943{
1944 mutex_enter(&spa->spa_suspend_lock);
1945 while (spa_suspended(spa))
1946 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1947 mutex_exit(&spa->spa_suspend_lock);
1948}
1949
1950/*
1951 * ==========================================================================
1952 * Gang blocks.
1953 *
1954 * A gang block is a collection of small blocks that looks to the DMU
1955 * like one large block. When zio_dva_allocate() cannot find a block
1956 * of the requested size, due to either severe fragmentation or the pool
1957 * being nearly full, it calls zio_write_gang_block() to construct the
1958 * block from smaller fragments.
1959 *
1960 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1961 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1962 * an indirect block: it's an array of block pointers. It consumes
1963 * only one sector and hence is allocatable regardless of fragmentation.
1964 * The gang header's bps point to its gang members, which hold the data.
1965 *
1966 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1967 * as the verifier to ensure uniqueness of the SHA256 checksum.
1968 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1969 * not the gang header. This ensures that data block signatures (needed for
1970 * deduplication) are independent of how the block is physically stored.
1971 *
1972 * Gang blocks can be nested: a gang member may itself be a gang block.
1973 * Thus every gang block is a tree in which root and all interior nodes are
1974 * gang headers, and the leaves are normal blocks that contain user data.
1975 * The root of the gang tree is called the gang leader.
1976 *
1977 * To perform any operation (read, rewrite, free, claim) on a gang block,
1978 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1979 * in the io_gang_tree field of the original logical i/o by recursively
1980 * reading the gang leader and all gang headers below it. This yields
1981 * an in-core tree containing the contents of every gang header and the
1982 * bps for every constituent of the gang block.
1983 *
1984 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1985 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1986 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1987 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1988 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1989 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1990 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1991 * of the gang header plus zio_checksum_compute() of the data to update the
1992 * gang header's blk_cksum as described above.
1993 *
1994 * The two-phase assemble/issue model solves the problem of partial failure --
1995 * what if you'd freed part of a gang block but then couldn't read the
1996 * gang header for another part? Assembling the entire gang tree first
1997 * ensures that all the necessary gang header I/O has succeeded before
1998 * starting the actual work of free, claim, or write. Once the gang tree
1999 * is assembled, free and claim are in-memory operations that cannot fail.
2000 *
2001 * In the event that a gang write fails, zio_dva_unallocate() walks the
2002 * gang tree to immediately free (i.e. insert back into the space map)
2003 * everything we've allocated. This ensures that we don't get ENOSPC
2004 * errors during repeated suspend/resume cycles due to a flaky device.
2005 *
2006 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
2007 * the gang tree, we won't modify the block, so we can safely defer the free
2008 * (knowing that the block is still intact). If we *can* assemble the gang
2009 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2010 * each constituent bp and we can allocate a new block on the next sync pass.
2011 *
2012 * In all cases, the gang tree allows complete recovery from partial failure.
2013 * ==========================================================================
2014 */
2015
2016static void
2017zio_gang_issue_func_done(zio_t *zio)
2018{
2019 abd_put(zio->io_abd);
2020}
2021
2022static zio_t *
2023zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2024 uint64_t offset)
2025{
2026 if (gn != NULL)
2027 return (pio);
2028
2029 return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset),
2030 BP_GET_PSIZE(bp), zio_gang_issue_func_done,
2031 NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2032 &pio->io_bookmark));
2033}
2034
2035static zio_t *
2036zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2037 uint64_t offset)
2038{
2039 zio_t *zio;
2040
2041 if (gn != NULL) {
2042 abd_t *gbh_abd =
2043 abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2044 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2045 gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL,
2046 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2047 &pio->io_bookmark);
2048 /*
2049 * As we rewrite each gang header, the pipeline will compute
2050 * a new gang block header checksum for it; but no one will
2051 * compute a new data checksum, so we do that here. The one
2052 * exception is the gang leader: the pipeline already computed
2053 * its data checksum because that stage precedes gang assembly.
2054 * (Presently, nothing actually uses interior data checksums;
2055 * this is just good hygiene.)
2056 */
2057 if (gn != pio->io_gang_leader->io_gang_tree) {
2058 abd_t *buf = abd_get_offset(data, offset);
2059
2060 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
2061 buf, BP_GET_PSIZE(bp));
2062
2063 abd_put(buf);
2064 }
2065 /*
2066 * If we are here to damage data for testing purposes,
2067 * leave the GBH alone so that we can detect the damage.
2068 */
2069 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
2070 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2071 } else {
2072 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2073 abd_get_offset(data, offset), BP_GET_PSIZE(bp),
2074 zio_gang_issue_func_done, NULL, pio->io_priority,
2075 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2076 }
2077
2078 return (zio);
2079}
2080
2081/* ARGSUSED */
2082static zio_t *
2083zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2084 uint64_t offset)
2085{
2086 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
2087 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
2088 ZIO_GANG_CHILD_FLAGS(pio)));
2089}
2090
2091/* ARGSUSED */
2092static zio_t *
2093zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2094 uint64_t offset)
2095{
2096 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
2097 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
2098}
2099
2100static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2101 NULL,
2102 zio_read_gang,
2103 zio_rewrite_gang,
2104 zio_free_gang,
2105 zio_claim_gang,
2106 NULL
2107};
2108
2109static void zio_gang_tree_assemble_done(zio_t *zio);
2110
2111static zio_gang_node_t *
2112zio_gang_node_alloc(zio_gang_node_t **gnpp)
2113{
2114 zio_gang_node_t *gn;
2115
2116 ASSERT(*gnpp == NULL);
2117
2118 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2119 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2120 *gnpp = gn;
2121
2122 return (gn);
2123}
2124
2125static void
2126zio_gang_node_free(zio_gang_node_t **gnpp)
2127{
2128 zio_gang_node_t *gn = *gnpp;
2129
2130 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2131 ASSERT(gn->gn_child[g] == NULL);
2132
2133 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2134 kmem_free(gn, sizeof (*gn));
2135 *gnpp = NULL;
2136}
2137
2138static void
2139zio_gang_tree_free(zio_gang_node_t **gnpp)
2140{
2141 zio_gang_node_t *gn = *gnpp;
2142
2143 if (gn == NULL)
2144 return;
2145
2146 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2147 zio_gang_tree_free(&gn->gn_child[g]);
2148
2149 zio_gang_node_free(gnpp);
2150}
2151
2152static void
2153zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2154{
2155 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2156 abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2157
2158 ASSERT(gio->io_gang_leader == gio);
2159 ASSERT(BP_IS_GANG(bp));
2160
2161 zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2162 zio_gang_tree_assemble_done, gn, gio->io_priority,
2163 ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2164}
2165
2166static void
2167zio_gang_tree_assemble_done(zio_t *zio)
2168{
2169 zio_t *gio = zio->io_gang_leader;
2170 zio_gang_node_t *gn = zio->io_private;
2171 blkptr_t *bp = zio->io_bp;
2172
2173 ASSERT(gio == zio_unique_parent(zio));
2174 ASSERT(zio->io_child_count == 0);
2175
2176 if (zio->io_error)
2177 return;
2178
2179 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2180 if (BP_SHOULD_BYTESWAP(bp))
2181 byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size);
2182
2183 ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh);
2184 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2185 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2186
2187 abd_put(zio->io_abd);
2188
2189 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2190 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2191 if (!BP_IS_GANG(gbp))
2192 continue;
2193 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2194 }
2195}
2196
2197static void
2198zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data,
2199 uint64_t offset)
2200{
2201 zio_t *gio = pio->io_gang_leader;
2202 zio_t *zio;
2203
2204 ASSERT(BP_IS_GANG(bp) == !!gn);
2205 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2206 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2207
2208 /*
2209 * If you're a gang header, your data is in gn->gn_gbh.
2210 * If you're a gang member, your data is in 'data' and gn == NULL.
2211 */
2212 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset);
2213
2214 if (gn != NULL) {
2215 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2216
2217 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2218 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2219 if (BP_IS_HOLE(gbp))
2220 continue;
2221 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data,
2222 offset);
2223 offset += BP_GET_PSIZE(gbp);
2224 }
2225 }
2226
2227 if (gn == gio->io_gang_tree && gio->io_abd != NULL)
2228 ASSERT3U(gio->io_size, ==, offset);
2229
2230 if (zio != pio)
2231 zio_nowait(zio);
2232}
2233
2234static int
2235zio_gang_assemble(zio_t *zio)
2236{
2237 blkptr_t *bp = zio->io_bp;
2238
2239 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2240 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2241
2242 zio->io_gang_leader = zio;
2243
2244 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2245
2246 return (ZIO_PIPELINE_CONTINUE);
2247}
2248
2249static int
2250zio_gang_issue(zio_t *zio)
2251{
2252 blkptr_t *bp = zio->io_bp;
2253
2254 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) {
2255 return (ZIO_PIPELINE_STOP);
2256 }
2257
2258 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2259 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2260
2261 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2262 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd,
2263 0);
2264 else
2265 zio_gang_tree_free(&zio->io_gang_tree);
2266
2267 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2268
2269 return (ZIO_PIPELINE_CONTINUE);
2270}
2271
2272static void
2273zio_write_gang_member_ready(zio_t *zio)
2274{
2275 zio_t *pio = zio_unique_parent(zio);
2276 zio_t *gio = zio->io_gang_leader;
2277 dva_t *cdva = zio->io_bp->blk_dva;
2278 dva_t *pdva = pio->io_bp->blk_dva;
2279 uint64_t asize;
2280
2281 if (BP_IS_HOLE(zio->io_bp))
2282 return;
2283
2284 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2285
2286 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2287 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2288 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2289 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2290 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2291
2292 mutex_enter(&pio->io_lock);
2293 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2294 ASSERT(DVA_GET_GANG(&pdva[d]));
2295 asize = DVA_GET_ASIZE(&pdva[d]);
2296 asize += DVA_GET_ASIZE(&cdva[d]);
2297 DVA_SET_ASIZE(&pdva[d], asize);
2298 }
2299 mutex_exit(&pio->io_lock);
2300}
2301
2302static void
2303zio_write_gang_done(zio_t *zio)
2304{
2305 abd_put(zio->io_abd);
2306}
2307
2308static int
2309zio_write_gang_block(zio_t *pio)
2310{
2311 spa_t *spa = pio->io_spa;
2312 metaslab_class_t *mc = spa_normal_class(spa);
2313 blkptr_t *bp = pio->io_bp;
2314 zio_t *gio = pio->io_gang_leader;
2315 zio_t *zio;
2316 zio_gang_node_t *gn, **gnpp;
2317 zio_gbh_phys_t *gbh;
2318 abd_t *gbh_abd;
2319 uint64_t txg = pio->io_txg;
2320 uint64_t resid = pio->io_size;
2321 uint64_t lsize;
2322 int copies = gio->io_prop.zp_copies;
2323 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2324 zio_prop_t zp;
2325 int error;
2326
2327 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2328 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2329 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2330 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2331
2332 flags |= METASLAB_ASYNC_ALLOC;
2333 VERIFY(refcount_held(&mc->mc_alloc_slots, pio));
2334
2335 /*
2336 * The logical zio has already placed a reservation for
2337 * 'copies' allocation slots but gang blocks may require
2338 * additional copies. These additional copies
2339 * (i.e. gbh_copies - copies) are guaranteed to succeed
2340 * since metaslab_class_throttle_reserve() always allows
2341 * additional reservations for gang blocks.
2342 */
2343 VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
2344 pio, flags));
2345 }
2346
2347 error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
2348 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags,
2349 &pio->io_alloc_list, pio);
2350 if (error) {
2351 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2352 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2353 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2354
2355 /*
2356 * If we failed to allocate the gang block header then
2357 * we remove any additional allocation reservations that
2358 * we placed here. The original reservation will
2359 * be removed when the logical I/O goes to the ready
2360 * stage.
2361 */
2362 metaslab_class_throttle_unreserve(mc,
2363 gbh_copies - copies, pio);
2364 }
2365 pio->io_error = error;
2366 return (ZIO_PIPELINE_CONTINUE);
2367 }
2368
2369 if (pio == gio) {
2370 gnpp = &gio->io_gang_tree;
2371 } else {
2372 gnpp = pio->io_private;
2373 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2374 }
2375
2376 gn = zio_gang_node_alloc(gnpp);
2377 gbh = gn->gn_gbh;
2378 bzero(gbh, SPA_GANGBLOCKSIZE);
2379 gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE);
2380
2381 /*
2382 * Create the gang header.
2383 */
2384 zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2385 zio_write_gang_done, NULL, pio->io_priority,
2386 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2387
2388 /*
2389 * Create and nowait the gang children.
2390 */
2391 for (int g = 0; resid != 0; resid -= lsize, g++) {
2392 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2393 SPA_MINBLOCKSIZE);
2394 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2395
2396 zp.zp_checksum = gio->io_prop.zp_checksum;
2397 zp.zp_compress = ZIO_COMPRESS_OFF;
2398 zp.zp_type = DMU_OT_NONE;
2399 zp.zp_level = 0;
2400 zp.zp_copies = gio->io_prop.zp_copies;
2401 zp.zp_dedup = B_FALSE;
2402 zp.zp_dedup_verify = B_FALSE;
2403 zp.zp_nopwrite = B_FALSE;
2404
2405 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2406 abd_get_offset(pio->io_abd, pio->io_size - resid), lsize,
2407 lsize, &zp, zio_write_gang_member_ready, NULL, NULL,
2408 zio_write_gang_done, &gn->gn_child[g], pio->io_priority,
2409 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2410
2411 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2412 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2413 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2414
2415 /*
2416 * Gang children won't throttle but we should
2417 * account for their work, so reserve an allocation
2418 * slot for them here.
2419 */
2420 VERIFY(metaslab_class_throttle_reserve(mc,
2421 zp.zp_copies, cio, flags));
2422 }
2423 zio_nowait(cio);
2424 }
2425
2426 /*
2427 * Set pio's pipeline to just wait for zio to finish.
2428 */
2429 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2430
2431 zio_nowait(zio);
2432
2433 return (ZIO_PIPELINE_CONTINUE);
2434}
2435
2436/*
2437 * The zio_nop_write stage in the pipeline determines if allocating a
2438 * new bp is necessary. The nopwrite feature can handle writes in
2439 * either syncing or open context (i.e. zil writes) and as a result is
2440 * mutually exclusive with dedup.
2441 *
2442 * By leveraging a cryptographically secure checksum, such as SHA256, we
2443 * can compare the checksums of the new data and the old to determine if
2444 * allocating a new block is required. Note that our requirements for
2445 * cryptographic strength are fairly weak: there can't be any accidental
2446 * hash collisions, but we don't need to be secure against intentional
2447 * (malicious) collisions. To trigger a nopwrite, you have to be able
2448 * to write the file to begin with, and triggering an incorrect (hash
2449 * collision) nopwrite is no worse than simply writing to the file.
2450 * That said, there are no known attacks against the checksum algorithms
2451 * used for nopwrite, assuming that the salt and the checksums
2452 * themselves remain secret.
2453 */
2454static int
2455zio_nop_write(zio_t *zio)
2456{
2457 blkptr_t *bp = zio->io_bp;
2458 blkptr_t *bp_orig = &zio->io_bp_orig;
2459 zio_prop_t *zp = &zio->io_prop;
2460
2461 ASSERT(BP_GET_LEVEL(bp) == 0);
2462 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2463 ASSERT(zp->zp_nopwrite);
2464 ASSERT(!zp->zp_dedup);
2465 ASSERT(zio->io_bp_override == NULL);
2466 ASSERT(IO_IS_ALLOCATING(zio));
2467
2468 /*
2469 * Check to see if the original bp and the new bp have matching
2470 * characteristics (i.e. same checksum, compression algorithms, etc).
2471 * If they don't then just continue with the pipeline which will
2472 * allocate a new bp.
2473 */
2474 if (BP_IS_HOLE(bp_orig) ||
2475 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2476 ZCHECKSUM_FLAG_NOPWRITE) ||
2477 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2478 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2479 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2480 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2481 return (ZIO_PIPELINE_CONTINUE);
2482
2483 /*
2484 * If the checksums match then reset the pipeline so that we
2485 * avoid allocating a new bp and issuing any I/O.
2486 */
2487 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2488 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2489 ZCHECKSUM_FLAG_NOPWRITE);
2490 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2491 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2492 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2493 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2494 sizeof (uint64_t)) == 0);
2495
2496 *bp = *bp_orig;
2497 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2498 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2499 }
2500
2501 return (ZIO_PIPELINE_CONTINUE);
2502}
2503
2504/*
2505 * ==========================================================================
2506 * Dedup
2507 * ==========================================================================
2508 */
2509static void
2510zio_ddt_child_read_done(zio_t *zio)
2511{
2512 blkptr_t *bp = zio->io_bp;
2513 ddt_entry_t *dde = zio->io_private;
2514 ddt_phys_t *ddp;
2515 zio_t *pio = zio_unique_parent(zio);
2516
2517 mutex_enter(&pio->io_lock);
2518 ddp = ddt_phys_select(dde, bp);
2519 if (zio->io_error == 0)
2520 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2521
2522 if (zio->io_error == 0 && dde->dde_repair_abd == NULL)
2523 dde->dde_repair_abd = zio->io_abd;
2524 else
2525 abd_free(zio->io_abd);
2526 mutex_exit(&pio->io_lock);
2527}
2528
2529static int
2530zio_ddt_read_start(zio_t *zio)
2531{
2532 blkptr_t *bp = zio->io_bp;
2533
2534 ASSERT(BP_GET_DEDUP(bp));
2535 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2536 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2537
2538 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2539 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2540 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2541 ddt_phys_t *ddp = dde->dde_phys;
2542 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2543 blkptr_t blk;
2544
2545 ASSERT(zio->io_vsd == NULL);
2546 zio->io_vsd = dde;
2547
2548 if (ddp_self == NULL)
2549 return (ZIO_PIPELINE_CONTINUE);
2550
2551 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2552 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2553 continue;
2554 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2555 &blk);
2556 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2557 abd_alloc_for_io(zio->io_size, B_TRUE),
2558 zio->io_size, zio_ddt_child_read_done, dde,
2559 zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) |
2560 ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark));
2561 }
2562 return (ZIO_PIPELINE_CONTINUE);
2563 }
2564
2565 zio_nowait(zio_read(zio, zio->io_spa, bp,
2566 zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority,
2567 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2568
2569 return (ZIO_PIPELINE_CONTINUE);
2570}
2571
2572static int
2573zio_ddt_read_done(zio_t *zio)
2574{
2575 blkptr_t *bp = zio->io_bp;
2576
2577 if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) {
2578 return (ZIO_PIPELINE_STOP);
2579 }
2580
2581 ASSERT(BP_GET_DEDUP(bp));
2582 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2583 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2584
2585 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2586 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2587 ddt_entry_t *dde = zio->io_vsd;
2588 if (ddt == NULL) {
2589 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2590 return (ZIO_PIPELINE_CONTINUE);
2591 }
2592 if (dde == NULL) {
2593 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2594 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2595 return (ZIO_PIPELINE_STOP);
2596 }
2597 if (dde->dde_repair_abd != NULL) {
2598 abd_copy(zio->io_abd, dde->dde_repair_abd,
2599 zio->io_size);
2600 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2601 }
2602 ddt_repair_done(ddt, dde);
2603 zio->io_vsd = NULL;
2604 }
2605
2606 ASSERT(zio->io_vsd == NULL);
2607
2608 return (ZIO_PIPELINE_CONTINUE);
2609}
2610
2611static boolean_t
2612zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2613{
2614 spa_t *spa = zio->io_spa;
2615 boolean_t do_raw = (zio->io_flags & ZIO_FLAG_RAW);
2616
2617 /* We should never get a raw, override zio */
2618 ASSERT(!(zio->io_bp_override && do_raw));
2619
2620 /*
2621 * Note: we compare the original data, not the transformed data,
2622 * because when zio->io_bp is an override bp, we will not have
2623 * pushed the I/O transforms. That's an important optimization
2624 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2625 */
2626 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2627 zio_t *lio = dde->dde_lead_zio[p];
2628
2629 if (lio != NULL) {
2630 return (lio->io_orig_size != zio->io_orig_size ||
2631 abd_cmp(zio->io_orig_abd, lio->io_orig_abd,
2632 zio->io_orig_size) != 0);
2633 }
2634 }
2635
2636 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2637 ddt_phys_t *ddp = &dde->dde_phys[p];
2638
2639 if (ddp->ddp_phys_birth != 0) {
2640 arc_buf_t *abuf = NULL;
2641 arc_flags_t aflags = ARC_FLAG_WAIT;
2642 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
2643 blkptr_t blk = *zio->io_bp;
2644 int error;
2645
2646 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2647
2648 ddt_exit(ddt);
2649
2650 /*
2651 * Intuitively, it would make more sense to compare
2652 * io_abd than io_orig_abd in the raw case since you
2653 * don't want to look at any transformations that have
2654 * happened to the data. However, for raw I/Os the
2655 * data will actually be the same in io_abd and
2656 * io_orig_abd, so all we have to do is issue this as
2657 * a raw ARC read.
2658 */
2659 if (do_raw) {
2660 zio_flags |= ZIO_FLAG_RAW;
2661 ASSERT3U(zio->io_size, ==, zio->io_orig_size);
2662 ASSERT0(abd_cmp(zio->io_abd, zio->io_orig_abd,
2663 zio->io_size));
2664 ASSERT3P(zio->io_transform_stack, ==, NULL);
2665 }
2666
2667 error = arc_read(NULL, spa, &blk,
2668 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2669 zio_flags, &aflags, &zio->io_bookmark);
2670
2671 if (error == 0) {
2672 if (arc_buf_size(abuf) != zio->io_orig_size ||
2673 abd_cmp_buf(zio->io_orig_abd, abuf->b_data,
2674 zio->io_orig_size) != 0)
2675 error = SET_ERROR(EEXIST);
2676 arc_buf_destroy(abuf, &abuf);
2677 }
2678
2679 ddt_enter(ddt);
2680 return (error != 0);
2681 }
2682 }
2683
2684 return (B_FALSE);
2685}
2686
2687static void
2688zio_ddt_child_write_ready(zio_t *zio)
2689{
2690 int p = zio->io_prop.zp_copies;
2691 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2692 ddt_entry_t *dde = zio->io_private;
2693 ddt_phys_t *ddp = &dde->dde_phys[p];
2694 zio_t *pio;
2695
2696 if (zio->io_error)
2697 return;
2698
2699 ddt_enter(ddt);
2700
2701 ASSERT(dde->dde_lead_zio[p] == zio);
2702
2703 ddt_phys_fill(ddp, zio->io_bp);
2704
2705 zio_link_t *zl = NULL;
2706 while ((pio = zio_walk_parents(zio, &zl)) != NULL)
2707 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2708
2709 ddt_exit(ddt);
2710}
2711
2712static void
2713zio_ddt_child_write_done(zio_t *zio)
2714{
2715 int p = zio->io_prop.zp_copies;
2716 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2717 ddt_entry_t *dde = zio->io_private;
2718 ddt_phys_t *ddp = &dde->dde_phys[p];
2719
2720 ddt_enter(ddt);
2721
2722 ASSERT(ddp->ddp_refcnt == 0);
2723 ASSERT(dde->dde_lead_zio[p] == zio);
2724 dde->dde_lead_zio[p] = NULL;
2725
2726 if (zio->io_error == 0) {
2727 zio_link_t *zl = NULL;
2728 while (zio_walk_parents(zio, &zl) != NULL)
2729 ddt_phys_addref(ddp);
2730 } else {
2731 ddt_phys_clear(ddp);
2732 }
2733
2734 ddt_exit(ddt);
2735}
2736
2737static void
2738zio_ddt_ditto_write_done(zio_t *zio)
2739{
2740 int p = DDT_PHYS_DITTO;
2741 zio_prop_t *zp = &zio->io_prop;
2742 blkptr_t *bp = zio->io_bp;
2743 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2744 ddt_entry_t *dde = zio->io_private;
2745 ddt_phys_t *ddp = &dde->dde_phys[p];
2746 ddt_key_t *ddk = &dde->dde_key;
2747
2748 ddt_enter(ddt);
2749
2750 ASSERT(ddp->ddp_refcnt == 0);
2751 ASSERT(dde->dde_lead_zio[p] == zio);
2752 dde->dde_lead_zio[p] = NULL;
2753
2754 if (zio->io_error == 0) {
2755 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2756 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2757 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2758 if (ddp->ddp_phys_birth != 0)
2759 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2760 ddt_phys_fill(ddp, bp);
2761 }
2762
2763 ddt_exit(ddt);
2764}
2765
2766static int
2767zio_ddt_write(zio_t *zio)
2768{
2769 spa_t *spa = zio->io_spa;
2770 blkptr_t *bp = zio->io_bp;
2771 uint64_t txg = zio->io_txg;
2772 zio_prop_t *zp = &zio->io_prop;
2773 int p = zp->zp_copies;
2774 int ditto_copies;
2775 zio_t *cio = NULL;
2776 zio_t *dio = NULL;
2777 ddt_t *ddt = ddt_select(spa, bp);
2778 ddt_entry_t *dde;
2779 ddt_phys_t *ddp;
2780
2781 ASSERT(BP_GET_DEDUP(bp));
2782 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2783 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2784 ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
2785
2786 ddt_enter(ddt);
2787 dde = ddt_lookup(ddt, bp, B_TRUE);
2788 ddp = &dde->dde_phys[p];
2789
2790 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2791 /*
2792 * If we're using a weak checksum, upgrade to a strong checksum
2793 * and try again. If we're already using a strong checksum,
2794 * we can't resolve it, so just convert to an ordinary write.
2795 * (And automatically e-mail a paper to Nature?)
2796 */
2797 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
2798 ZCHECKSUM_FLAG_DEDUP)) {
2799 zp->zp_checksum = spa_dedup_checksum(spa);
2800 zio_pop_transforms(zio);
2801 zio->io_stage = ZIO_STAGE_OPEN;
2802 BP_ZERO(bp);
2803 } else {
2804 zp->zp_dedup = B_FALSE;
2805 BP_SET_DEDUP(bp, B_FALSE);
2806 }
2807 ASSERT(!BP_GET_DEDUP(bp));
2808 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2809 ddt_exit(ddt);
2810 return (ZIO_PIPELINE_CONTINUE);
2811 }
2812
2813 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2814 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2815
2816 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2817 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2818 zio_prop_t czp = *zp;
2819
2820 czp.zp_copies = ditto_copies;
2821
2822 /*
2823 * If we arrived here with an override bp, we won't have run
2824 * the transform stack, so we won't have the data we need to
2825 * generate a child i/o. So, toss the override bp and restart.
2826 * This is safe, because using the override bp is just an
2827 * optimization; and it's rare, so the cost doesn't matter.
2828 */
2829 if (zio->io_bp_override) {
2830 zio_pop_transforms(zio);
2831 zio->io_stage = ZIO_STAGE_OPEN;
2832 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2833 zio->io_bp_override = NULL;
2834 BP_ZERO(bp);
2835 ddt_exit(ddt);
2836 return (ZIO_PIPELINE_CONTINUE);
2837 }
2838
2839 dio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2840 zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL,
2841 NULL, zio_ddt_ditto_write_done, dde, zio->io_priority,
2842 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2843
2844 zio_push_transform(dio, zio->io_abd, zio->io_size, 0, NULL);
2845 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2846 }
2847
2848 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2849 if (ddp->ddp_phys_birth != 0)
2850 ddt_bp_fill(ddp, bp, txg);
2851 if (dde->dde_lead_zio[p] != NULL)
2852 zio_add_child(zio, dde->dde_lead_zio[p]);
2853 else
2854 ddt_phys_addref(ddp);
2855 } else if (zio->io_bp_override) {
2856 ASSERT(bp->blk_birth == txg);
2857 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2858 ddt_phys_fill(ddp, bp);
2859 ddt_phys_addref(ddp);
2860 } else {
2861 cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2862 zio->io_orig_size, zio->io_orig_size, zp,
2863 zio_ddt_child_write_ready, NULL, NULL,
2864 zio_ddt_child_write_done, dde, zio->io_priority,
2865 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2866
2867 zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL);
2868 dde->dde_lead_zio[p] = cio;
2869 }
2870
2871 ddt_exit(ddt);
2872
2873 if (cio)
2874 zio_nowait(cio);
2875 if (dio)
2876 zio_nowait(dio);
2877
2878 return (ZIO_PIPELINE_CONTINUE);
2879}
2880
2881ddt_entry_t *freedde; /* for debugging */
2882
2883static int
2884zio_ddt_free(zio_t *zio)
2885{
2886 spa_t *spa = zio->io_spa;
2887 blkptr_t *bp = zio->io_bp;
2888 ddt_t *ddt = ddt_select(spa, bp);
2889 ddt_entry_t *dde;
2890 ddt_phys_t *ddp;
2891
2892 ASSERT(BP_GET_DEDUP(bp));
2893 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2894
2895 ddt_enter(ddt);
2896 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2897 ddp = ddt_phys_select(dde, bp);
2898 ddt_phys_decref(ddp);
2899 ddt_exit(ddt);
2900
2901 return (ZIO_PIPELINE_CONTINUE);
2902}
2903
2904/*
2905 * ==========================================================================
2906 * Allocate and free blocks
2907 * ==========================================================================
2908 */
2909
2910static zio_t *
2911zio_io_to_allocate(spa_t *spa)
2912{
2913 zio_t *zio;
2914
2915 ASSERT(MUTEX_HELD(&spa->spa_alloc_lock));
2916
2917 zio = avl_first(&spa->spa_alloc_tree);
2918 if (zio == NULL)
2919 return (NULL);
2920
2921 ASSERT(IO_IS_ALLOCATING(zio));
2922
2923 /*
2924 * Try to place a reservation for this zio. If we're unable to
2925 * reserve then we throttle.
2926 */
2927 if (!metaslab_class_throttle_reserve(spa_normal_class(spa),
2928 zio->io_prop.zp_copies, zio, 0)) {
2929 return (NULL);
2930 }
2931
2932 avl_remove(&spa->spa_alloc_tree, zio);
2933 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
2934
2935 return (zio);
2936}
2937
2938static int
2939zio_dva_throttle(zio_t *zio)
2940{
2941 spa_t *spa = zio->io_spa;
2942 zio_t *nio;
2943
2944 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
2945 !spa_normal_class(zio->io_spa)->mc_alloc_throttle_enabled ||
2946 zio->io_child_type == ZIO_CHILD_GANG ||
2947 zio->io_flags & ZIO_FLAG_NODATA) {
2948 return (ZIO_PIPELINE_CONTINUE);
2949 }
2950
2951 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2952
2953 ASSERT3U(zio->io_queued_timestamp, >, 0);
2954 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
2955
2956 mutex_enter(&spa->spa_alloc_lock);
2957
2958 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2959 avl_add(&spa->spa_alloc_tree, zio);
2960
2961 nio = zio_io_to_allocate(zio->io_spa);
2962 mutex_exit(&spa->spa_alloc_lock);
2963
2964 if (nio == zio)
2965 return (ZIO_PIPELINE_CONTINUE);
2966
2967 if (nio != NULL) {
2968 ASSERT(nio->io_stage == ZIO_STAGE_DVA_THROTTLE);
2969 /*
2970 * We are passing control to a new zio so make sure that
2971 * it is processed by a different thread. We do this to
2972 * avoid stack overflows that can occur when parents are
2973 * throttled and children are making progress. We allow
2974 * it to go to the head of the taskq since it's already
2975 * been waiting.
2976 */
2977 zio_taskq_dispatch(nio, ZIO_TASKQ_ISSUE, B_TRUE);
2978 }
2979 return (ZIO_PIPELINE_STOP);
2980}
2981
2982void
2983zio_allocate_dispatch(spa_t *spa)
2984{
2985 zio_t *zio;
2986
2987 mutex_enter(&spa->spa_alloc_lock);
2988 zio = zio_io_to_allocate(spa);
2989 mutex_exit(&spa->spa_alloc_lock);
2990 if (zio == NULL)
2991 return;
2992
2993 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
2994 ASSERT0(zio->io_error);
2995 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
2996}
2997
2998static int
2999zio_dva_allocate(zio_t *zio)
3000{
3001 spa_t *spa = zio->io_spa;
3002 metaslab_class_t *mc = spa_normal_class(spa);
3003 blkptr_t *bp = zio->io_bp;
3004 int error;
3005 int flags = 0;
3006
3007 if (zio->io_gang_leader == NULL) {
3008 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3009 zio->io_gang_leader = zio;
3010 }
3011
3012 ASSERT(BP_IS_HOLE(bp));
3013 ASSERT0(BP_GET_NDVAS(bp));
3014 ASSERT3U(zio->io_prop.zp_copies, >, 0);
3015 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
3016 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
3017
3018 if (zio->io_flags & ZIO_FLAG_NODATA) {
3019 flags |= METASLAB_DONT_THROTTLE;
3020 }
3021 if (zio->io_flags & ZIO_FLAG_GANG_CHILD) {
3022 flags |= METASLAB_GANG_CHILD;
3023 }
3024 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) {
3025 flags |= METASLAB_ASYNC_ALLOC;
3026 }
3027
3028 error = metaslab_alloc(spa, mc, zio->io_size, bp,
3029 zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3030 &zio->io_alloc_list, zio);
3031
3032 if (error != 0) {
3033 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
3034 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
3035 error);
3036 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
3037 return (zio_write_gang_block(zio));
3038 zio->io_error = error;
3039 }
3040
3041 return (ZIO_PIPELINE_CONTINUE);
3042}
3043
3044static int
3045zio_dva_free(zio_t *zio)
3046{
3047 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
3048
3049 return (ZIO_PIPELINE_CONTINUE);
3050}
3051
3052static int
3053zio_dva_claim(zio_t *zio)
3054{
3055 int error;
3056
3057 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
3058 if (error)
3059 zio->io_error = error;
3060
3061 return (ZIO_PIPELINE_CONTINUE);
3062}
3063
3064/*
3065 * Undo an allocation. This is used by zio_done() when an I/O fails
3066 * and we want to give back the block we just allocated.
3067 * This handles both normal blocks and gang blocks.
3068 */
3069static void
3070zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
3071{
3072 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
3073 ASSERT(zio->io_bp_override == NULL);
3074
3075 if (!BP_IS_HOLE(bp))
3076 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
3077
3078 if (gn != NULL) {
3079 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
3080 zio_dva_unallocate(zio, gn->gn_child[g],
3081 &gn->gn_gbh->zg_blkptr[g]);
3082 }
3083 }
3084}
3085
3086/*
3087 * Try to allocate an intent log block. Return 0 on success, errno on failure.
3088 */
3089int
3090zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
3091 uint64_t size, boolean_t *slog)
3092{
3093 int error = 1;
3094 zio_alloc_list_t io_alloc_list;
3095
3096 ASSERT(txg > spa_syncing_txg(spa));
3097
3098 metaslab_trace_init(&io_alloc_list);
3099 error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
3100 txg, old_bp, METASLAB_HINTBP_AVOID, &io_alloc_list, NULL);
3101 if (error == 0) {
3102 *slog = TRUE;
3103 } else {
3104 error = metaslab_alloc(spa, spa_normal_class(spa), size,
3105 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID,
3106 &io_alloc_list, NULL);
3107 if (error == 0)
3108 *slog = FALSE;
3109 }
3110 metaslab_trace_fini(&io_alloc_list);
3111
3112 if (error == 0) {
3113 BP_SET_LSIZE(new_bp, size);
3114 BP_SET_PSIZE(new_bp, size);
3115 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
3116 BP_SET_CHECKSUM(new_bp,
3117 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
3118 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
3119 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3120 BP_SET_LEVEL(new_bp, 0);
3121 BP_SET_DEDUP(new_bp, 0);
3122 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
3123 } else {
3124 zfs_dbgmsg("%s: zil block allocation failure: "
3125 "size %llu, error %d", spa_name(spa), size, error);
3126 }
3127
3128 return (error);
3129}
3130
3131/*
3132 * ==========================================================================
3133 * Read, write and delete to physical devices
3134 * ==========================================================================
3135 */
3136
3137
3138/*
3139 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3140 * stops after this stage and will resume upon I/O completion.
3141 * However, there are instances where the vdev layer may need to
3142 * continue the pipeline when an I/O was not issued. Since the I/O
3143 * that was sent to the vdev layer might be different than the one
3144 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3145 * force the underlying vdev layers to call either zio_execute() or
3146 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3147 */
3148static int
3149zio_vdev_io_start(zio_t *zio)
3150{
3151 vdev_t *vd = zio->io_vd;
3152 uint64_t align;
3153 spa_t *spa = zio->io_spa;
3154 int ret;
3155
3156 ASSERT(zio->io_error == 0);
3157 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3158
3159 if (vd == NULL) {
3160 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3161 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3162
3163 /*
3164 * The mirror_ops handle multiple DVAs in a single BP.
3165 */
3166 vdev_mirror_ops.vdev_op_io_start(zio);
3167 return (ZIO_PIPELINE_STOP);
3168 }
3169
3170 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE &&
3171 zio->io_priority == ZIO_PRIORITY_NOW) {
3172 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
3173 return (ZIO_PIPELINE_CONTINUE);
3174 }
3175
3176 ASSERT3P(zio->io_logical, !=, zio);
3177 if (zio->io_type == ZIO_TYPE_WRITE) {
3178 ASSERT(spa->spa_trust_config);
3179
3180 if (zio->io_vd->vdev_removing) {
3181 ASSERT(zio->io_flags &
3182 (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL |
3183 ZIO_FLAG_INDUCE_DAMAGE));
3184 }
3185 }
3186
| 1881
1882 /*
1883 * Now that all children have been reexecuted, execute the parent.
1884 * We don't reexecute "The Godfather" I/O here as it's the
1885 * responsibility of the caller to wait on it.
1886 */
1887 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
1888 pio->io_queued_timestamp = gethrtime();
1889 zio_execute(pio);
1890 }
1891}
1892
1893void
1894zio_suspend(spa_t *spa, zio_t *zio)
1895{
1896 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1897 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1898 "failure and the failure mode property for this pool "
1899 "is set to panic.", spa_name(spa));
1900
1901 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1902
1903 mutex_enter(&spa->spa_suspend_lock);
1904
1905 if (spa->spa_suspend_zio_root == NULL)
1906 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1907 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1908 ZIO_FLAG_GODFATHER);
1909
1910 spa->spa_suspended = B_TRUE;
1911
1912 if (zio != NULL) {
1913 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1914 ASSERT(zio != spa->spa_suspend_zio_root);
1915 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1916 ASSERT(zio_unique_parent(zio) == NULL);
1917 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1918 zio_add_child(spa->spa_suspend_zio_root, zio);
1919 }
1920
1921 mutex_exit(&spa->spa_suspend_lock);
1922}
1923
1924int
1925zio_resume(spa_t *spa)
1926{
1927 zio_t *pio;
1928
1929 /*
1930 * Reexecute all previously suspended i/o.
1931 */
1932 mutex_enter(&spa->spa_suspend_lock);
1933 spa->spa_suspended = B_FALSE;
1934 cv_broadcast(&spa->spa_suspend_cv);
1935 pio = spa->spa_suspend_zio_root;
1936 spa->spa_suspend_zio_root = NULL;
1937 mutex_exit(&spa->spa_suspend_lock);
1938
1939 if (pio == NULL)
1940 return (0);
1941
1942 zio_reexecute(pio);
1943 return (zio_wait(pio));
1944}
1945
1946void
1947zio_resume_wait(spa_t *spa)
1948{
1949 mutex_enter(&spa->spa_suspend_lock);
1950 while (spa_suspended(spa))
1951 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1952 mutex_exit(&spa->spa_suspend_lock);
1953}
1954
1955/*
1956 * ==========================================================================
1957 * Gang blocks.
1958 *
1959 * A gang block is a collection of small blocks that looks to the DMU
1960 * like one large block. When zio_dva_allocate() cannot find a block
1961 * of the requested size, due to either severe fragmentation or the pool
1962 * being nearly full, it calls zio_write_gang_block() to construct the
1963 * block from smaller fragments.
1964 *
1965 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1966 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1967 * an indirect block: it's an array of block pointers. It consumes
1968 * only one sector and hence is allocatable regardless of fragmentation.
1969 * The gang header's bps point to its gang members, which hold the data.
1970 *
1971 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1972 * as the verifier to ensure uniqueness of the SHA256 checksum.
1973 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1974 * not the gang header. This ensures that data block signatures (needed for
1975 * deduplication) are independent of how the block is physically stored.
1976 *
1977 * Gang blocks can be nested: a gang member may itself be a gang block.
1978 * Thus every gang block is a tree in which root and all interior nodes are
1979 * gang headers, and the leaves are normal blocks that contain user data.
1980 * The root of the gang tree is called the gang leader.
1981 *
1982 * To perform any operation (read, rewrite, free, claim) on a gang block,
1983 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1984 * in the io_gang_tree field of the original logical i/o by recursively
1985 * reading the gang leader and all gang headers below it. This yields
1986 * an in-core tree containing the contents of every gang header and the
1987 * bps for every constituent of the gang block.
1988 *
1989 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1990 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1991 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1992 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1993 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1994 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1995 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1996 * of the gang header plus zio_checksum_compute() of the data to update the
1997 * gang header's blk_cksum as described above.
1998 *
1999 * The two-phase assemble/issue model solves the problem of partial failure --
2000 * what if you'd freed part of a gang block but then couldn't read the
2001 * gang header for another part? Assembling the entire gang tree first
2002 * ensures that all the necessary gang header I/O has succeeded before
2003 * starting the actual work of free, claim, or write. Once the gang tree
2004 * is assembled, free and claim are in-memory operations that cannot fail.
2005 *
2006 * In the event that a gang write fails, zio_dva_unallocate() walks the
2007 * gang tree to immediately free (i.e. insert back into the space map)
2008 * everything we've allocated. This ensures that we don't get ENOSPC
2009 * errors during repeated suspend/resume cycles due to a flaky device.
2010 *
2011 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
2012 * the gang tree, we won't modify the block, so we can safely defer the free
2013 * (knowing that the block is still intact). If we *can* assemble the gang
2014 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2015 * each constituent bp and we can allocate a new block on the next sync pass.
2016 *
2017 * In all cases, the gang tree allows complete recovery from partial failure.
2018 * ==========================================================================
2019 */
2020
2021static void
2022zio_gang_issue_func_done(zio_t *zio)
2023{
2024 abd_put(zio->io_abd);
2025}
2026
2027static zio_t *
2028zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2029 uint64_t offset)
2030{
2031 if (gn != NULL)
2032 return (pio);
2033
2034 return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset),
2035 BP_GET_PSIZE(bp), zio_gang_issue_func_done,
2036 NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2037 &pio->io_bookmark));
2038}
2039
2040static zio_t *
2041zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2042 uint64_t offset)
2043{
2044 zio_t *zio;
2045
2046 if (gn != NULL) {
2047 abd_t *gbh_abd =
2048 abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2049 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2050 gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL,
2051 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2052 &pio->io_bookmark);
2053 /*
2054 * As we rewrite each gang header, the pipeline will compute
2055 * a new gang block header checksum for it; but no one will
2056 * compute a new data checksum, so we do that here. The one
2057 * exception is the gang leader: the pipeline already computed
2058 * its data checksum because that stage precedes gang assembly.
2059 * (Presently, nothing actually uses interior data checksums;
2060 * this is just good hygiene.)
2061 */
2062 if (gn != pio->io_gang_leader->io_gang_tree) {
2063 abd_t *buf = abd_get_offset(data, offset);
2064
2065 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
2066 buf, BP_GET_PSIZE(bp));
2067
2068 abd_put(buf);
2069 }
2070 /*
2071 * If we are here to damage data for testing purposes,
2072 * leave the GBH alone so that we can detect the damage.
2073 */
2074 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
2075 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2076 } else {
2077 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2078 abd_get_offset(data, offset), BP_GET_PSIZE(bp),
2079 zio_gang_issue_func_done, NULL, pio->io_priority,
2080 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2081 }
2082
2083 return (zio);
2084}
2085
2086/* ARGSUSED */
2087static zio_t *
2088zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2089 uint64_t offset)
2090{
2091 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
2092 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
2093 ZIO_GANG_CHILD_FLAGS(pio)));
2094}
2095
2096/* ARGSUSED */
2097static zio_t *
2098zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2099 uint64_t offset)
2100{
2101 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
2102 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
2103}
2104
2105static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2106 NULL,
2107 zio_read_gang,
2108 zio_rewrite_gang,
2109 zio_free_gang,
2110 zio_claim_gang,
2111 NULL
2112};
2113
2114static void zio_gang_tree_assemble_done(zio_t *zio);
2115
2116static zio_gang_node_t *
2117zio_gang_node_alloc(zio_gang_node_t **gnpp)
2118{
2119 zio_gang_node_t *gn;
2120
2121 ASSERT(*gnpp == NULL);
2122
2123 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2124 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2125 *gnpp = gn;
2126
2127 return (gn);
2128}
2129
2130static void
2131zio_gang_node_free(zio_gang_node_t **gnpp)
2132{
2133 zio_gang_node_t *gn = *gnpp;
2134
2135 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2136 ASSERT(gn->gn_child[g] == NULL);
2137
2138 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2139 kmem_free(gn, sizeof (*gn));
2140 *gnpp = NULL;
2141}
2142
2143static void
2144zio_gang_tree_free(zio_gang_node_t **gnpp)
2145{
2146 zio_gang_node_t *gn = *gnpp;
2147
2148 if (gn == NULL)
2149 return;
2150
2151 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2152 zio_gang_tree_free(&gn->gn_child[g]);
2153
2154 zio_gang_node_free(gnpp);
2155}
2156
2157static void
2158zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2159{
2160 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2161 abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2162
2163 ASSERT(gio->io_gang_leader == gio);
2164 ASSERT(BP_IS_GANG(bp));
2165
2166 zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2167 zio_gang_tree_assemble_done, gn, gio->io_priority,
2168 ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2169}
2170
2171static void
2172zio_gang_tree_assemble_done(zio_t *zio)
2173{
2174 zio_t *gio = zio->io_gang_leader;
2175 zio_gang_node_t *gn = zio->io_private;
2176 blkptr_t *bp = zio->io_bp;
2177
2178 ASSERT(gio == zio_unique_parent(zio));
2179 ASSERT(zio->io_child_count == 0);
2180
2181 if (zio->io_error)
2182 return;
2183
2184 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2185 if (BP_SHOULD_BYTESWAP(bp))
2186 byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size);
2187
2188 ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh);
2189 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2190 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2191
2192 abd_put(zio->io_abd);
2193
2194 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2195 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2196 if (!BP_IS_GANG(gbp))
2197 continue;
2198 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2199 }
2200}
2201
2202static void
2203zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data,
2204 uint64_t offset)
2205{
2206 zio_t *gio = pio->io_gang_leader;
2207 zio_t *zio;
2208
2209 ASSERT(BP_IS_GANG(bp) == !!gn);
2210 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2211 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2212
2213 /*
2214 * If you're a gang header, your data is in gn->gn_gbh.
2215 * If you're a gang member, your data is in 'data' and gn == NULL.
2216 */
2217 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset);
2218
2219 if (gn != NULL) {
2220 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2221
2222 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2223 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2224 if (BP_IS_HOLE(gbp))
2225 continue;
2226 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data,
2227 offset);
2228 offset += BP_GET_PSIZE(gbp);
2229 }
2230 }
2231
2232 if (gn == gio->io_gang_tree && gio->io_abd != NULL)
2233 ASSERT3U(gio->io_size, ==, offset);
2234
2235 if (zio != pio)
2236 zio_nowait(zio);
2237}
2238
2239static int
2240zio_gang_assemble(zio_t *zio)
2241{
2242 blkptr_t *bp = zio->io_bp;
2243
2244 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2245 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2246
2247 zio->io_gang_leader = zio;
2248
2249 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2250
2251 return (ZIO_PIPELINE_CONTINUE);
2252}
2253
2254static int
2255zio_gang_issue(zio_t *zio)
2256{
2257 blkptr_t *bp = zio->io_bp;
2258
2259 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) {
2260 return (ZIO_PIPELINE_STOP);
2261 }
2262
2263 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2264 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2265
2266 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2267 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd,
2268 0);
2269 else
2270 zio_gang_tree_free(&zio->io_gang_tree);
2271
2272 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2273
2274 return (ZIO_PIPELINE_CONTINUE);
2275}
2276
2277static void
2278zio_write_gang_member_ready(zio_t *zio)
2279{
2280 zio_t *pio = zio_unique_parent(zio);
2281 zio_t *gio = zio->io_gang_leader;
2282 dva_t *cdva = zio->io_bp->blk_dva;
2283 dva_t *pdva = pio->io_bp->blk_dva;
2284 uint64_t asize;
2285
2286 if (BP_IS_HOLE(zio->io_bp))
2287 return;
2288
2289 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2290
2291 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2292 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2293 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2294 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2295 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2296
2297 mutex_enter(&pio->io_lock);
2298 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2299 ASSERT(DVA_GET_GANG(&pdva[d]));
2300 asize = DVA_GET_ASIZE(&pdva[d]);
2301 asize += DVA_GET_ASIZE(&cdva[d]);
2302 DVA_SET_ASIZE(&pdva[d], asize);
2303 }
2304 mutex_exit(&pio->io_lock);
2305}
2306
2307static void
2308zio_write_gang_done(zio_t *zio)
2309{
2310 abd_put(zio->io_abd);
2311}
2312
2313static int
2314zio_write_gang_block(zio_t *pio)
2315{
2316 spa_t *spa = pio->io_spa;
2317 metaslab_class_t *mc = spa_normal_class(spa);
2318 blkptr_t *bp = pio->io_bp;
2319 zio_t *gio = pio->io_gang_leader;
2320 zio_t *zio;
2321 zio_gang_node_t *gn, **gnpp;
2322 zio_gbh_phys_t *gbh;
2323 abd_t *gbh_abd;
2324 uint64_t txg = pio->io_txg;
2325 uint64_t resid = pio->io_size;
2326 uint64_t lsize;
2327 int copies = gio->io_prop.zp_copies;
2328 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2329 zio_prop_t zp;
2330 int error;
2331
2332 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2333 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2334 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2335 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2336
2337 flags |= METASLAB_ASYNC_ALLOC;
2338 VERIFY(refcount_held(&mc->mc_alloc_slots, pio));
2339
2340 /*
2341 * The logical zio has already placed a reservation for
2342 * 'copies' allocation slots but gang blocks may require
2343 * additional copies. These additional copies
2344 * (i.e. gbh_copies - copies) are guaranteed to succeed
2345 * since metaslab_class_throttle_reserve() always allows
2346 * additional reservations for gang blocks.
2347 */
2348 VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
2349 pio, flags));
2350 }
2351
2352 error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
2353 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags,
2354 &pio->io_alloc_list, pio);
2355 if (error) {
2356 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2357 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2358 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2359
2360 /*
2361 * If we failed to allocate the gang block header then
2362 * we remove any additional allocation reservations that
2363 * we placed here. The original reservation will
2364 * be removed when the logical I/O goes to the ready
2365 * stage.
2366 */
2367 metaslab_class_throttle_unreserve(mc,
2368 gbh_copies - copies, pio);
2369 }
2370 pio->io_error = error;
2371 return (ZIO_PIPELINE_CONTINUE);
2372 }
2373
2374 if (pio == gio) {
2375 gnpp = &gio->io_gang_tree;
2376 } else {
2377 gnpp = pio->io_private;
2378 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2379 }
2380
2381 gn = zio_gang_node_alloc(gnpp);
2382 gbh = gn->gn_gbh;
2383 bzero(gbh, SPA_GANGBLOCKSIZE);
2384 gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE);
2385
2386 /*
2387 * Create the gang header.
2388 */
2389 zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2390 zio_write_gang_done, NULL, pio->io_priority,
2391 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2392
2393 /*
2394 * Create and nowait the gang children.
2395 */
2396 for (int g = 0; resid != 0; resid -= lsize, g++) {
2397 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2398 SPA_MINBLOCKSIZE);
2399 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2400
2401 zp.zp_checksum = gio->io_prop.zp_checksum;
2402 zp.zp_compress = ZIO_COMPRESS_OFF;
2403 zp.zp_type = DMU_OT_NONE;
2404 zp.zp_level = 0;
2405 zp.zp_copies = gio->io_prop.zp_copies;
2406 zp.zp_dedup = B_FALSE;
2407 zp.zp_dedup_verify = B_FALSE;
2408 zp.zp_nopwrite = B_FALSE;
2409
2410 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2411 abd_get_offset(pio->io_abd, pio->io_size - resid), lsize,
2412 lsize, &zp, zio_write_gang_member_ready, NULL, NULL,
2413 zio_write_gang_done, &gn->gn_child[g], pio->io_priority,
2414 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2415
2416 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2417 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2418 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2419
2420 /*
2421 * Gang children won't throttle but we should
2422 * account for their work, so reserve an allocation
2423 * slot for them here.
2424 */
2425 VERIFY(metaslab_class_throttle_reserve(mc,
2426 zp.zp_copies, cio, flags));
2427 }
2428 zio_nowait(cio);
2429 }
2430
2431 /*
2432 * Set pio's pipeline to just wait for zio to finish.
2433 */
2434 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2435
2436 zio_nowait(zio);
2437
2438 return (ZIO_PIPELINE_CONTINUE);
2439}
2440
2441/*
2442 * The zio_nop_write stage in the pipeline determines if allocating a
2443 * new bp is necessary. The nopwrite feature can handle writes in
2444 * either syncing or open context (i.e. zil writes) and as a result is
2445 * mutually exclusive with dedup.
2446 *
2447 * By leveraging a cryptographically secure checksum, such as SHA256, we
2448 * can compare the checksums of the new data and the old to determine if
2449 * allocating a new block is required. Note that our requirements for
2450 * cryptographic strength are fairly weak: there can't be any accidental
2451 * hash collisions, but we don't need to be secure against intentional
2452 * (malicious) collisions. To trigger a nopwrite, you have to be able
2453 * to write the file to begin with, and triggering an incorrect (hash
2454 * collision) nopwrite is no worse than simply writing to the file.
2455 * That said, there are no known attacks against the checksum algorithms
2456 * used for nopwrite, assuming that the salt and the checksums
2457 * themselves remain secret.
2458 */
2459static int
2460zio_nop_write(zio_t *zio)
2461{
2462 blkptr_t *bp = zio->io_bp;
2463 blkptr_t *bp_orig = &zio->io_bp_orig;
2464 zio_prop_t *zp = &zio->io_prop;
2465
2466 ASSERT(BP_GET_LEVEL(bp) == 0);
2467 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2468 ASSERT(zp->zp_nopwrite);
2469 ASSERT(!zp->zp_dedup);
2470 ASSERT(zio->io_bp_override == NULL);
2471 ASSERT(IO_IS_ALLOCATING(zio));
2472
2473 /*
2474 * Check to see if the original bp and the new bp have matching
2475 * characteristics (i.e. same checksum, compression algorithms, etc).
2476 * If they don't then just continue with the pipeline which will
2477 * allocate a new bp.
2478 */
2479 if (BP_IS_HOLE(bp_orig) ||
2480 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2481 ZCHECKSUM_FLAG_NOPWRITE) ||
2482 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2483 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2484 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2485 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2486 return (ZIO_PIPELINE_CONTINUE);
2487
2488 /*
2489 * If the checksums match then reset the pipeline so that we
2490 * avoid allocating a new bp and issuing any I/O.
2491 */
2492 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2493 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2494 ZCHECKSUM_FLAG_NOPWRITE);
2495 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2496 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2497 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2498 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2499 sizeof (uint64_t)) == 0);
2500
2501 *bp = *bp_orig;
2502 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2503 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2504 }
2505
2506 return (ZIO_PIPELINE_CONTINUE);
2507}
2508
2509/*
2510 * ==========================================================================
2511 * Dedup
2512 * ==========================================================================
2513 */
2514static void
2515zio_ddt_child_read_done(zio_t *zio)
2516{
2517 blkptr_t *bp = zio->io_bp;
2518 ddt_entry_t *dde = zio->io_private;
2519 ddt_phys_t *ddp;
2520 zio_t *pio = zio_unique_parent(zio);
2521
2522 mutex_enter(&pio->io_lock);
2523 ddp = ddt_phys_select(dde, bp);
2524 if (zio->io_error == 0)
2525 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2526
2527 if (zio->io_error == 0 && dde->dde_repair_abd == NULL)
2528 dde->dde_repair_abd = zio->io_abd;
2529 else
2530 abd_free(zio->io_abd);
2531 mutex_exit(&pio->io_lock);
2532}
2533
2534static int
2535zio_ddt_read_start(zio_t *zio)
2536{
2537 blkptr_t *bp = zio->io_bp;
2538
2539 ASSERT(BP_GET_DEDUP(bp));
2540 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2541 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2542
2543 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2544 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2545 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2546 ddt_phys_t *ddp = dde->dde_phys;
2547 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2548 blkptr_t blk;
2549
2550 ASSERT(zio->io_vsd == NULL);
2551 zio->io_vsd = dde;
2552
2553 if (ddp_self == NULL)
2554 return (ZIO_PIPELINE_CONTINUE);
2555
2556 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2557 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2558 continue;
2559 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2560 &blk);
2561 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2562 abd_alloc_for_io(zio->io_size, B_TRUE),
2563 zio->io_size, zio_ddt_child_read_done, dde,
2564 zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) |
2565 ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark));
2566 }
2567 return (ZIO_PIPELINE_CONTINUE);
2568 }
2569
2570 zio_nowait(zio_read(zio, zio->io_spa, bp,
2571 zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority,
2572 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2573
2574 return (ZIO_PIPELINE_CONTINUE);
2575}
2576
2577static int
2578zio_ddt_read_done(zio_t *zio)
2579{
2580 blkptr_t *bp = zio->io_bp;
2581
2582 if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) {
2583 return (ZIO_PIPELINE_STOP);
2584 }
2585
2586 ASSERT(BP_GET_DEDUP(bp));
2587 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2588 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2589
2590 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2591 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2592 ddt_entry_t *dde = zio->io_vsd;
2593 if (ddt == NULL) {
2594 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2595 return (ZIO_PIPELINE_CONTINUE);
2596 }
2597 if (dde == NULL) {
2598 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2599 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2600 return (ZIO_PIPELINE_STOP);
2601 }
2602 if (dde->dde_repair_abd != NULL) {
2603 abd_copy(zio->io_abd, dde->dde_repair_abd,
2604 zio->io_size);
2605 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2606 }
2607 ddt_repair_done(ddt, dde);
2608 zio->io_vsd = NULL;
2609 }
2610
2611 ASSERT(zio->io_vsd == NULL);
2612
2613 return (ZIO_PIPELINE_CONTINUE);
2614}
2615
2616static boolean_t
2617zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2618{
2619 spa_t *spa = zio->io_spa;
2620 boolean_t do_raw = (zio->io_flags & ZIO_FLAG_RAW);
2621
2622 /* We should never get a raw, override zio */
2623 ASSERT(!(zio->io_bp_override && do_raw));
2624
2625 /*
2626 * Note: we compare the original data, not the transformed data,
2627 * because when zio->io_bp is an override bp, we will not have
2628 * pushed the I/O transforms. That's an important optimization
2629 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2630 */
2631 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2632 zio_t *lio = dde->dde_lead_zio[p];
2633
2634 if (lio != NULL) {
2635 return (lio->io_orig_size != zio->io_orig_size ||
2636 abd_cmp(zio->io_orig_abd, lio->io_orig_abd,
2637 zio->io_orig_size) != 0);
2638 }
2639 }
2640
2641 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2642 ddt_phys_t *ddp = &dde->dde_phys[p];
2643
2644 if (ddp->ddp_phys_birth != 0) {
2645 arc_buf_t *abuf = NULL;
2646 arc_flags_t aflags = ARC_FLAG_WAIT;
2647 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
2648 blkptr_t blk = *zio->io_bp;
2649 int error;
2650
2651 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2652
2653 ddt_exit(ddt);
2654
2655 /*
2656 * Intuitively, it would make more sense to compare
2657 * io_abd than io_orig_abd in the raw case since you
2658 * don't want to look at any transformations that have
2659 * happened to the data. However, for raw I/Os the
2660 * data will actually be the same in io_abd and
2661 * io_orig_abd, so all we have to do is issue this as
2662 * a raw ARC read.
2663 */
2664 if (do_raw) {
2665 zio_flags |= ZIO_FLAG_RAW;
2666 ASSERT3U(zio->io_size, ==, zio->io_orig_size);
2667 ASSERT0(abd_cmp(zio->io_abd, zio->io_orig_abd,
2668 zio->io_size));
2669 ASSERT3P(zio->io_transform_stack, ==, NULL);
2670 }
2671
2672 error = arc_read(NULL, spa, &blk,
2673 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2674 zio_flags, &aflags, &zio->io_bookmark);
2675
2676 if (error == 0) {
2677 if (arc_buf_size(abuf) != zio->io_orig_size ||
2678 abd_cmp_buf(zio->io_orig_abd, abuf->b_data,
2679 zio->io_orig_size) != 0)
2680 error = SET_ERROR(EEXIST);
2681 arc_buf_destroy(abuf, &abuf);
2682 }
2683
2684 ddt_enter(ddt);
2685 return (error != 0);
2686 }
2687 }
2688
2689 return (B_FALSE);
2690}
2691
2692static void
2693zio_ddt_child_write_ready(zio_t *zio)
2694{
2695 int p = zio->io_prop.zp_copies;
2696 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2697 ddt_entry_t *dde = zio->io_private;
2698 ddt_phys_t *ddp = &dde->dde_phys[p];
2699 zio_t *pio;
2700
2701 if (zio->io_error)
2702 return;
2703
2704 ddt_enter(ddt);
2705
2706 ASSERT(dde->dde_lead_zio[p] == zio);
2707
2708 ddt_phys_fill(ddp, zio->io_bp);
2709
2710 zio_link_t *zl = NULL;
2711 while ((pio = zio_walk_parents(zio, &zl)) != NULL)
2712 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2713
2714 ddt_exit(ddt);
2715}
2716
2717static void
2718zio_ddt_child_write_done(zio_t *zio)
2719{
2720 int p = zio->io_prop.zp_copies;
2721 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2722 ddt_entry_t *dde = zio->io_private;
2723 ddt_phys_t *ddp = &dde->dde_phys[p];
2724
2725 ddt_enter(ddt);
2726
2727 ASSERT(ddp->ddp_refcnt == 0);
2728 ASSERT(dde->dde_lead_zio[p] == zio);
2729 dde->dde_lead_zio[p] = NULL;
2730
2731 if (zio->io_error == 0) {
2732 zio_link_t *zl = NULL;
2733 while (zio_walk_parents(zio, &zl) != NULL)
2734 ddt_phys_addref(ddp);
2735 } else {
2736 ddt_phys_clear(ddp);
2737 }
2738
2739 ddt_exit(ddt);
2740}
2741
2742static void
2743zio_ddt_ditto_write_done(zio_t *zio)
2744{
2745 int p = DDT_PHYS_DITTO;
2746 zio_prop_t *zp = &zio->io_prop;
2747 blkptr_t *bp = zio->io_bp;
2748 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2749 ddt_entry_t *dde = zio->io_private;
2750 ddt_phys_t *ddp = &dde->dde_phys[p];
2751 ddt_key_t *ddk = &dde->dde_key;
2752
2753 ddt_enter(ddt);
2754
2755 ASSERT(ddp->ddp_refcnt == 0);
2756 ASSERT(dde->dde_lead_zio[p] == zio);
2757 dde->dde_lead_zio[p] = NULL;
2758
2759 if (zio->io_error == 0) {
2760 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2761 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2762 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2763 if (ddp->ddp_phys_birth != 0)
2764 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2765 ddt_phys_fill(ddp, bp);
2766 }
2767
2768 ddt_exit(ddt);
2769}
2770
2771static int
2772zio_ddt_write(zio_t *zio)
2773{
2774 spa_t *spa = zio->io_spa;
2775 blkptr_t *bp = zio->io_bp;
2776 uint64_t txg = zio->io_txg;
2777 zio_prop_t *zp = &zio->io_prop;
2778 int p = zp->zp_copies;
2779 int ditto_copies;
2780 zio_t *cio = NULL;
2781 zio_t *dio = NULL;
2782 ddt_t *ddt = ddt_select(spa, bp);
2783 ddt_entry_t *dde;
2784 ddt_phys_t *ddp;
2785
2786 ASSERT(BP_GET_DEDUP(bp));
2787 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2788 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2789 ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
2790
2791 ddt_enter(ddt);
2792 dde = ddt_lookup(ddt, bp, B_TRUE);
2793 ddp = &dde->dde_phys[p];
2794
2795 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2796 /*
2797 * If we're using a weak checksum, upgrade to a strong checksum
2798 * and try again. If we're already using a strong checksum,
2799 * we can't resolve it, so just convert to an ordinary write.
2800 * (And automatically e-mail a paper to Nature?)
2801 */
2802 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
2803 ZCHECKSUM_FLAG_DEDUP)) {
2804 zp->zp_checksum = spa_dedup_checksum(spa);
2805 zio_pop_transforms(zio);
2806 zio->io_stage = ZIO_STAGE_OPEN;
2807 BP_ZERO(bp);
2808 } else {
2809 zp->zp_dedup = B_FALSE;
2810 BP_SET_DEDUP(bp, B_FALSE);
2811 }
2812 ASSERT(!BP_GET_DEDUP(bp));
2813 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2814 ddt_exit(ddt);
2815 return (ZIO_PIPELINE_CONTINUE);
2816 }
2817
2818 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2819 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2820
2821 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2822 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2823 zio_prop_t czp = *zp;
2824
2825 czp.zp_copies = ditto_copies;
2826
2827 /*
2828 * If we arrived here with an override bp, we won't have run
2829 * the transform stack, so we won't have the data we need to
2830 * generate a child i/o. So, toss the override bp and restart.
2831 * This is safe, because using the override bp is just an
2832 * optimization; and it's rare, so the cost doesn't matter.
2833 */
2834 if (zio->io_bp_override) {
2835 zio_pop_transforms(zio);
2836 zio->io_stage = ZIO_STAGE_OPEN;
2837 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2838 zio->io_bp_override = NULL;
2839 BP_ZERO(bp);
2840 ddt_exit(ddt);
2841 return (ZIO_PIPELINE_CONTINUE);
2842 }
2843
2844 dio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2845 zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL,
2846 NULL, zio_ddt_ditto_write_done, dde, zio->io_priority,
2847 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2848
2849 zio_push_transform(dio, zio->io_abd, zio->io_size, 0, NULL);
2850 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2851 }
2852
2853 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2854 if (ddp->ddp_phys_birth != 0)
2855 ddt_bp_fill(ddp, bp, txg);
2856 if (dde->dde_lead_zio[p] != NULL)
2857 zio_add_child(zio, dde->dde_lead_zio[p]);
2858 else
2859 ddt_phys_addref(ddp);
2860 } else if (zio->io_bp_override) {
2861 ASSERT(bp->blk_birth == txg);
2862 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2863 ddt_phys_fill(ddp, bp);
2864 ddt_phys_addref(ddp);
2865 } else {
2866 cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2867 zio->io_orig_size, zio->io_orig_size, zp,
2868 zio_ddt_child_write_ready, NULL, NULL,
2869 zio_ddt_child_write_done, dde, zio->io_priority,
2870 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2871
2872 zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL);
2873 dde->dde_lead_zio[p] = cio;
2874 }
2875
2876 ddt_exit(ddt);
2877
2878 if (cio)
2879 zio_nowait(cio);
2880 if (dio)
2881 zio_nowait(dio);
2882
2883 return (ZIO_PIPELINE_CONTINUE);
2884}
2885
2886ddt_entry_t *freedde; /* for debugging */
2887
2888static int
2889zio_ddt_free(zio_t *zio)
2890{
2891 spa_t *spa = zio->io_spa;
2892 blkptr_t *bp = zio->io_bp;
2893 ddt_t *ddt = ddt_select(spa, bp);
2894 ddt_entry_t *dde;
2895 ddt_phys_t *ddp;
2896
2897 ASSERT(BP_GET_DEDUP(bp));
2898 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2899
2900 ddt_enter(ddt);
2901 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2902 ddp = ddt_phys_select(dde, bp);
2903 ddt_phys_decref(ddp);
2904 ddt_exit(ddt);
2905
2906 return (ZIO_PIPELINE_CONTINUE);
2907}
2908
2909/*
2910 * ==========================================================================
2911 * Allocate and free blocks
2912 * ==========================================================================
2913 */
2914
2915static zio_t *
2916zio_io_to_allocate(spa_t *spa)
2917{
2918 zio_t *zio;
2919
2920 ASSERT(MUTEX_HELD(&spa->spa_alloc_lock));
2921
2922 zio = avl_first(&spa->spa_alloc_tree);
2923 if (zio == NULL)
2924 return (NULL);
2925
2926 ASSERT(IO_IS_ALLOCATING(zio));
2927
2928 /*
2929 * Try to place a reservation for this zio. If we're unable to
2930 * reserve then we throttle.
2931 */
2932 if (!metaslab_class_throttle_reserve(spa_normal_class(spa),
2933 zio->io_prop.zp_copies, zio, 0)) {
2934 return (NULL);
2935 }
2936
2937 avl_remove(&spa->spa_alloc_tree, zio);
2938 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
2939
2940 return (zio);
2941}
2942
2943static int
2944zio_dva_throttle(zio_t *zio)
2945{
2946 spa_t *spa = zio->io_spa;
2947 zio_t *nio;
2948
2949 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
2950 !spa_normal_class(zio->io_spa)->mc_alloc_throttle_enabled ||
2951 zio->io_child_type == ZIO_CHILD_GANG ||
2952 zio->io_flags & ZIO_FLAG_NODATA) {
2953 return (ZIO_PIPELINE_CONTINUE);
2954 }
2955
2956 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2957
2958 ASSERT3U(zio->io_queued_timestamp, >, 0);
2959 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
2960
2961 mutex_enter(&spa->spa_alloc_lock);
2962
2963 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2964 avl_add(&spa->spa_alloc_tree, zio);
2965
2966 nio = zio_io_to_allocate(zio->io_spa);
2967 mutex_exit(&spa->spa_alloc_lock);
2968
2969 if (nio == zio)
2970 return (ZIO_PIPELINE_CONTINUE);
2971
2972 if (nio != NULL) {
2973 ASSERT(nio->io_stage == ZIO_STAGE_DVA_THROTTLE);
2974 /*
2975 * We are passing control to a new zio so make sure that
2976 * it is processed by a different thread. We do this to
2977 * avoid stack overflows that can occur when parents are
2978 * throttled and children are making progress. We allow
2979 * it to go to the head of the taskq since it's already
2980 * been waiting.
2981 */
2982 zio_taskq_dispatch(nio, ZIO_TASKQ_ISSUE, B_TRUE);
2983 }
2984 return (ZIO_PIPELINE_STOP);
2985}
2986
2987void
2988zio_allocate_dispatch(spa_t *spa)
2989{
2990 zio_t *zio;
2991
2992 mutex_enter(&spa->spa_alloc_lock);
2993 zio = zio_io_to_allocate(spa);
2994 mutex_exit(&spa->spa_alloc_lock);
2995 if (zio == NULL)
2996 return;
2997
2998 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
2999 ASSERT0(zio->io_error);
3000 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
3001}
3002
3003static int
3004zio_dva_allocate(zio_t *zio)
3005{
3006 spa_t *spa = zio->io_spa;
3007 metaslab_class_t *mc = spa_normal_class(spa);
3008 blkptr_t *bp = zio->io_bp;
3009 int error;
3010 int flags = 0;
3011
3012 if (zio->io_gang_leader == NULL) {
3013 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3014 zio->io_gang_leader = zio;
3015 }
3016
3017 ASSERT(BP_IS_HOLE(bp));
3018 ASSERT0(BP_GET_NDVAS(bp));
3019 ASSERT3U(zio->io_prop.zp_copies, >, 0);
3020 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
3021 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
3022
3023 if (zio->io_flags & ZIO_FLAG_NODATA) {
3024 flags |= METASLAB_DONT_THROTTLE;
3025 }
3026 if (zio->io_flags & ZIO_FLAG_GANG_CHILD) {
3027 flags |= METASLAB_GANG_CHILD;
3028 }
3029 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) {
3030 flags |= METASLAB_ASYNC_ALLOC;
3031 }
3032
3033 error = metaslab_alloc(spa, mc, zio->io_size, bp,
3034 zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3035 &zio->io_alloc_list, zio);
3036
3037 if (error != 0) {
3038 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
3039 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
3040 error);
3041 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
3042 return (zio_write_gang_block(zio));
3043 zio->io_error = error;
3044 }
3045
3046 return (ZIO_PIPELINE_CONTINUE);
3047}
3048
3049static int
3050zio_dva_free(zio_t *zio)
3051{
3052 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
3053
3054 return (ZIO_PIPELINE_CONTINUE);
3055}
3056
3057static int
3058zio_dva_claim(zio_t *zio)
3059{
3060 int error;
3061
3062 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
3063 if (error)
3064 zio->io_error = error;
3065
3066 return (ZIO_PIPELINE_CONTINUE);
3067}
3068
3069/*
3070 * Undo an allocation. This is used by zio_done() when an I/O fails
3071 * and we want to give back the block we just allocated.
3072 * This handles both normal blocks and gang blocks.
3073 */
3074static void
3075zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
3076{
3077 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
3078 ASSERT(zio->io_bp_override == NULL);
3079
3080 if (!BP_IS_HOLE(bp))
3081 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
3082
3083 if (gn != NULL) {
3084 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
3085 zio_dva_unallocate(zio, gn->gn_child[g],
3086 &gn->gn_gbh->zg_blkptr[g]);
3087 }
3088 }
3089}
3090
3091/*
3092 * Try to allocate an intent log block. Return 0 on success, errno on failure.
3093 */
3094int
3095zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
3096 uint64_t size, boolean_t *slog)
3097{
3098 int error = 1;
3099 zio_alloc_list_t io_alloc_list;
3100
3101 ASSERT(txg > spa_syncing_txg(spa));
3102
3103 metaslab_trace_init(&io_alloc_list);
3104 error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
3105 txg, old_bp, METASLAB_HINTBP_AVOID, &io_alloc_list, NULL);
3106 if (error == 0) {
3107 *slog = TRUE;
3108 } else {
3109 error = metaslab_alloc(spa, spa_normal_class(spa), size,
3110 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID,
3111 &io_alloc_list, NULL);
3112 if (error == 0)
3113 *slog = FALSE;
3114 }
3115 metaslab_trace_fini(&io_alloc_list);
3116
3117 if (error == 0) {
3118 BP_SET_LSIZE(new_bp, size);
3119 BP_SET_PSIZE(new_bp, size);
3120 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
3121 BP_SET_CHECKSUM(new_bp,
3122 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
3123 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
3124 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3125 BP_SET_LEVEL(new_bp, 0);
3126 BP_SET_DEDUP(new_bp, 0);
3127 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
3128 } else {
3129 zfs_dbgmsg("%s: zil block allocation failure: "
3130 "size %llu, error %d", spa_name(spa), size, error);
3131 }
3132
3133 return (error);
3134}
3135
3136/*
3137 * ==========================================================================
3138 * Read, write and delete to physical devices
3139 * ==========================================================================
3140 */
3141
3142
3143/*
3144 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3145 * stops after this stage and will resume upon I/O completion.
3146 * However, there are instances where the vdev layer may need to
3147 * continue the pipeline when an I/O was not issued. Since the I/O
3148 * that was sent to the vdev layer might be different than the one
3149 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3150 * force the underlying vdev layers to call either zio_execute() or
3151 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3152 */
3153static int
3154zio_vdev_io_start(zio_t *zio)
3155{
3156 vdev_t *vd = zio->io_vd;
3157 uint64_t align;
3158 spa_t *spa = zio->io_spa;
3159 int ret;
3160
3161 ASSERT(zio->io_error == 0);
3162 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3163
3164 if (vd == NULL) {
3165 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3166 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3167
3168 /*
3169 * The mirror_ops handle multiple DVAs in a single BP.
3170 */
3171 vdev_mirror_ops.vdev_op_io_start(zio);
3172 return (ZIO_PIPELINE_STOP);
3173 }
3174
3175 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE &&
3176 zio->io_priority == ZIO_PRIORITY_NOW) {
3177 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
3178 return (ZIO_PIPELINE_CONTINUE);
3179 }
3180
3181 ASSERT3P(zio->io_logical, !=, zio);
3182 if (zio->io_type == ZIO_TYPE_WRITE) {
3183 ASSERT(spa->spa_trust_config);
3184
3185 if (zio->io_vd->vdev_removing) {
3186 ASSERT(zio->io_flags &
3187 (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL |
3188 ZIO_FLAG_INDUCE_DAMAGE));
3189 }
3190 }
3191
|
3187 /*
3188 * We keep track of time-sensitive I/Os so that the scan thread
3189 * can quickly react to certain workloads. In particular, we care
3190 * about non-scrubbing, top-level reads and writes with the following
3191 * characteristics:
3192 * - synchronous writes of user data to non-slog devices
3193 * - any reads of user data
3194 * When these conditions are met, adjust the timestamp of spa_last_io
3195 * which allows the scan thread to adjust its workload accordingly.
3196 */
3197 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
3198 vd == vd->vdev_top && !vd->vdev_islog &&
3199 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
3200 zio->io_txg != spa_syncing_txg(spa)) {
3201 uint64_t old = spa->spa_last_io;
3202 uint64_t new = ddi_get_lbolt64();
3203 if (old != new)
3204 (void) atomic_cas_64(&spa->spa_last_io, old, new);
3205 }
3206
| 3192 /*
3193 * We keep track of time-sensitive I/Os so that the scan thread
3194 * can quickly react to certain workloads. In particular, we care
3195 * about non-scrubbing, top-level reads and writes with the following
3196 * characteristics:
3197 * - synchronous writes of user data to non-slog devices
3198 * - any reads of user data
3199 * When these conditions are met, adjust the timestamp of spa_last_io
3200 * which allows the scan thread to adjust its workload accordingly.
3201 */
3202 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
3203 vd == vd->vdev_top && !vd->vdev_islog &&
3204 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
3205 zio->io_txg != spa_syncing_txg(spa)) {
3206 uint64_t old = spa->spa_last_io;
3207 uint64_t new = ddi_get_lbolt64();
3208 if (old != new)
3209 (void) atomic_cas_64(&spa->spa_last_io, old, new);
3210 }
|
3207 align = 1ULL << vd->vdev_top->vdev_ashift;
3208
3209 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3210 P2PHASE(zio->io_size, align) != 0) {
3211 /* Transform logical writes to be a full physical block size. */
3212 uint64_t asize = P2ROUNDUP(zio->io_size, align);
3213 abd_t *abuf = NULL;
3214 if (zio->io_type == ZIO_TYPE_READ ||
3215 zio->io_type == ZIO_TYPE_WRITE)
3216 abuf = abd_alloc_sametype(zio->io_abd, asize);
3217 ASSERT(vd == vd->vdev_top);
3218 if (zio->io_type == ZIO_TYPE_WRITE) {
3219 abd_copy(abuf, zio->io_abd, zio->io_size);
3220 abd_zero_off(abuf, zio->io_size, asize - zio->io_size);
3221 }
3222 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
3223 zio_subblock);
3224 }
3225
3226 /*
3227 * If this is not a physical io, make sure that it is properly aligned
3228 * before proceeding.
3229 */
3230 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3231 ASSERT0(P2PHASE(zio->io_offset, align));
3232 ASSERT0(P2PHASE(zio->io_size, align));
3233 } else {
3234 /*
3235 * For the physical io we allow alignment
3236 * to a logical block size.
3237 */
3238 uint64_t log_align =
3239 1ULL << vd->vdev_top->vdev_logical_ashift;
3240 ASSERT0(P2PHASE(zio->io_offset, log_align));
3241 ASSERT0(P2PHASE(zio->io_size, log_align));
3242 }
3243
3244 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
3245
3246 /*
3247 * If this is a repair I/O, and there's no self-healing involved --
3248 * that is, we're just resilvering what we expect to resilver --
3249 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3250 * This prevents spurious resilvering with nested replication.
3251 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
3252 * A is out of date, we'll read from C+D, then use the data to
3253 * resilver A+B -- but we don't actually want to resilver B, just A.
3254 * The top-level mirror has no way to know this, so instead we just
3255 * discard unnecessary repairs as we work our way down the vdev tree.
3256 * The same logic applies to any form of nested replication:
3257 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
3258 */
3259 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3260 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3261 zio->io_txg != 0 && /* not a delegated i/o */
3262 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3263 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3264 zio_vdev_io_bypass(zio);
3265 return (ZIO_PIPELINE_CONTINUE);
3266 }
3267
3268 if (vd->vdev_ops->vdev_op_leaf) {
3269 switch (zio->io_type) {
3270 case ZIO_TYPE_READ:
3271 if (vdev_cache_read(zio))
3272 return (ZIO_PIPELINE_CONTINUE);
3273 /* FALLTHROUGH */
3274 case ZIO_TYPE_WRITE:
3275 case ZIO_TYPE_FREE:
3276 if ((zio = vdev_queue_io(zio)) == NULL)
3277 return (ZIO_PIPELINE_STOP);
3278
3279 if (!vdev_accessible(vd, zio)) {
3280 zio->io_error = SET_ERROR(ENXIO);
3281 zio_interrupt(zio);
3282 return (ZIO_PIPELINE_STOP);
3283 }
3284 break;
3285 }
3286 /*
3287 * Note that we ignore repair writes for TRIM because they can
3288 * conflict with normal writes. This isn't an issue because, by
3289 * definition, we only repair blocks that aren't freed.
3290 */
3291 if (zio->io_type == ZIO_TYPE_WRITE &&
3292 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3293 !trim_map_write_start(zio))
3294 return (ZIO_PIPELINE_STOP);
3295 }
3296
3297 vd->vdev_ops->vdev_op_io_start(zio);
3298 return (ZIO_PIPELINE_STOP);
3299}
3300
3301static int
3302zio_vdev_io_done(zio_t *zio)
3303{
3304 vdev_t *vd = zio->io_vd;
3305 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3306 boolean_t unexpected_error = B_FALSE;
3307
3308 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3309 return (ZIO_PIPELINE_STOP);
3310 }
3311
3312 ASSERT(zio->io_type == ZIO_TYPE_READ ||
3313 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
3314
3315 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3316 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE ||
3317 zio->io_type == ZIO_TYPE_FREE)) {
3318
3319 if (zio->io_type == ZIO_TYPE_WRITE &&
3320 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
3321 trim_map_write_done(zio);
3322
3323 vdev_queue_io_done(zio);
3324
3325 if (zio->io_type == ZIO_TYPE_WRITE)
3326 vdev_cache_write(zio);
3327
3328 if (zio_injection_enabled && zio->io_error == 0)
3329 zio->io_error = zio_handle_device_injection(vd,
3330 zio, EIO);
3331
3332 if (zio_injection_enabled && zio->io_error == 0)
3333 zio->io_error = zio_handle_label_injection(zio, EIO);
3334
3335 if (zio->io_error) {
3336 if (zio->io_error == ENOTSUP &&
3337 zio->io_type == ZIO_TYPE_FREE) {
3338 /* Not all devices support TRIM. */
3339 } else if (!vdev_accessible(vd, zio)) {
3340 zio->io_error = SET_ERROR(ENXIO);
3341 } else {
3342 unexpected_error = B_TRUE;
3343 }
3344 }
3345 }
3346
3347 ops->vdev_op_io_done(zio);
3348
3349 if (unexpected_error)
3350 VERIFY(vdev_probe(vd, zio) == NULL);
3351
3352 return (ZIO_PIPELINE_CONTINUE);
3353}
3354
3355/*
| 3211 align = 1ULL << vd->vdev_top->vdev_ashift;
3212
3213 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3214 P2PHASE(zio->io_size, align) != 0) {
3215 /* Transform logical writes to be a full physical block size. */
3216 uint64_t asize = P2ROUNDUP(zio->io_size, align);
3217 abd_t *abuf = NULL;
3218 if (zio->io_type == ZIO_TYPE_READ ||
3219 zio->io_type == ZIO_TYPE_WRITE)
3220 abuf = abd_alloc_sametype(zio->io_abd, asize);
3221 ASSERT(vd == vd->vdev_top);
3222 if (zio->io_type == ZIO_TYPE_WRITE) {
3223 abd_copy(abuf, zio->io_abd, zio->io_size);
3224 abd_zero_off(abuf, zio->io_size, asize - zio->io_size);
3225 }
3226 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
3227 zio_subblock);
3228 }
3229
3230 /*
3231 * If this is not a physical io, make sure that it is properly aligned
3232 * before proceeding.
3233 */
3234 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3235 ASSERT0(P2PHASE(zio->io_offset, align));
3236 ASSERT0(P2PHASE(zio->io_size, align));
3237 } else {
3238 /*
3239 * For the physical io we allow alignment
3240 * to a logical block size.
3241 */
3242 uint64_t log_align =
3243 1ULL << vd->vdev_top->vdev_logical_ashift;
3244 ASSERT0(P2PHASE(zio->io_offset, log_align));
3245 ASSERT0(P2PHASE(zio->io_size, log_align));
3246 }
3247
3248 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
3249
3250 /*
3251 * If this is a repair I/O, and there's no self-healing involved --
3252 * that is, we're just resilvering what we expect to resilver --
3253 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3254 * This prevents spurious resilvering with nested replication.
3255 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
3256 * A is out of date, we'll read from C+D, then use the data to
3257 * resilver A+B -- but we don't actually want to resilver B, just A.
3258 * The top-level mirror has no way to know this, so instead we just
3259 * discard unnecessary repairs as we work our way down the vdev tree.
3260 * The same logic applies to any form of nested replication:
3261 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
3262 */
3263 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3264 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3265 zio->io_txg != 0 && /* not a delegated i/o */
3266 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3267 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3268 zio_vdev_io_bypass(zio);
3269 return (ZIO_PIPELINE_CONTINUE);
3270 }
3271
3272 if (vd->vdev_ops->vdev_op_leaf) {
3273 switch (zio->io_type) {
3274 case ZIO_TYPE_READ:
3275 if (vdev_cache_read(zio))
3276 return (ZIO_PIPELINE_CONTINUE);
3277 /* FALLTHROUGH */
3278 case ZIO_TYPE_WRITE:
3279 case ZIO_TYPE_FREE:
3280 if ((zio = vdev_queue_io(zio)) == NULL)
3281 return (ZIO_PIPELINE_STOP);
3282
3283 if (!vdev_accessible(vd, zio)) {
3284 zio->io_error = SET_ERROR(ENXIO);
3285 zio_interrupt(zio);
3286 return (ZIO_PIPELINE_STOP);
3287 }
3288 break;
3289 }
3290 /*
3291 * Note that we ignore repair writes for TRIM because they can
3292 * conflict with normal writes. This isn't an issue because, by
3293 * definition, we only repair blocks that aren't freed.
3294 */
3295 if (zio->io_type == ZIO_TYPE_WRITE &&
3296 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3297 !trim_map_write_start(zio))
3298 return (ZIO_PIPELINE_STOP);
3299 }
3300
3301 vd->vdev_ops->vdev_op_io_start(zio);
3302 return (ZIO_PIPELINE_STOP);
3303}
3304
3305static int
3306zio_vdev_io_done(zio_t *zio)
3307{
3308 vdev_t *vd = zio->io_vd;
3309 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3310 boolean_t unexpected_error = B_FALSE;
3311
3312 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3313 return (ZIO_PIPELINE_STOP);
3314 }
3315
3316 ASSERT(zio->io_type == ZIO_TYPE_READ ||
3317 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
3318
3319 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3320 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE ||
3321 zio->io_type == ZIO_TYPE_FREE)) {
3322
3323 if (zio->io_type == ZIO_TYPE_WRITE &&
3324 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
3325 trim_map_write_done(zio);
3326
3327 vdev_queue_io_done(zio);
3328
3329 if (zio->io_type == ZIO_TYPE_WRITE)
3330 vdev_cache_write(zio);
3331
3332 if (zio_injection_enabled && zio->io_error == 0)
3333 zio->io_error = zio_handle_device_injection(vd,
3334 zio, EIO);
3335
3336 if (zio_injection_enabled && zio->io_error == 0)
3337 zio->io_error = zio_handle_label_injection(zio, EIO);
3338
3339 if (zio->io_error) {
3340 if (zio->io_error == ENOTSUP &&
3341 zio->io_type == ZIO_TYPE_FREE) {
3342 /* Not all devices support TRIM. */
3343 } else if (!vdev_accessible(vd, zio)) {
3344 zio->io_error = SET_ERROR(ENXIO);
3345 } else {
3346 unexpected_error = B_TRUE;
3347 }
3348 }
3349 }
3350
3351 ops->vdev_op_io_done(zio);
3352
3353 if (unexpected_error)
3354 VERIFY(vdev_probe(vd, zio) == NULL);
3355
3356 return (ZIO_PIPELINE_CONTINUE);
3357}
3358
3359/*
|
| 3360 * This function is used to change the priority of an existing zio that is
3361 * currently in-flight. This is used by the arc to upgrade priority in the
3362 * event that a demand read is made for a block that is currently queued
3363 * as a scrub or async read IO. Otherwise, the high priority read request
3364 * would end up having to wait for the lower priority IO.
3365 */
3366void
3367zio_change_priority(zio_t *pio, zio_priority_t priority)
3368{
3369 zio_t *cio, *cio_next;
3370 zio_link_t *zl = NULL;
3371
3372 ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
3373
3374 if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) {
3375 vdev_queue_change_io_priority(pio, priority);
3376 } else {
3377 pio->io_priority = priority;
3378 }
3379
3380 mutex_enter(&pio->io_lock);
3381 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
3382 cio_next = zio_walk_children(pio, &zl);
3383 zio_change_priority(cio, priority);
3384 }
3385 mutex_exit(&pio->io_lock);
3386}
3387
3388/*
|
3356 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3357 * disk, and use that to finish the checksum ereport later.
3358 */
3359static void
3360zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
3361 const void *good_buf)
3362{
3363 /* no processing needed */
3364 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
3365}
3366
3367/*ARGSUSED*/
3368void
3369zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
3370{
3371 void *buf = zio_buf_alloc(zio->io_size);
3372
3373 abd_copy_to_buf(buf, zio->io_abd, zio->io_size);
3374
3375 zcr->zcr_cbinfo = zio->io_size;
3376 zcr->zcr_cbdata = buf;
3377 zcr->zcr_finish = zio_vsd_default_cksum_finish;
3378 zcr->zcr_free = zio_buf_free;
3379}
3380
3381static int
3382zio_vdev_io_assess(zio_t *zio)
3383{
3384 vdev_t *vd = zio->io_vd;
3385
3386 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3387 return (ZIO_PIPELINE_STOP);
3388 }
3389
3390 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3391 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
3392
3393 if (zio->io_vsd != NULL) {
3394 zio->io_vsd_ops->vsd_free(zio);
3395 zio->io_vsd = NULL;
3396 }
3397
3398 if (zio_injection_enabled && zio->io_error == 0)
3399 zio->io_error = zio_handle_fault_injection(zio, EIO);
3400
3401 if (zio->io_type == ZIO_TYPE_FREE &&
3402 zio->io_priority != ZIO_PRIORITY_NOW) {
3403 switch (zio->io_error) {
3404 case 0:
3405 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
3406 ZIO_TRIM_STAT_BUMP(success);
3407 break;
3408 case EOPNOTSUPP:
3409 ZIO_TRIM_STAT_BUMP(unsupported);
3410 break;
3411 default:
3412 ZIO_TRIM_STAT_BUMP(failed);
3413 break;
3414 }
3415 }
3416
3417 /*
3418 * If the I/O failed, determine whether we should attempt to retry it.
3419 *
3420 * On retry, we cut in line in the issue queue, since we don't want
3421 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3422 */
3423 if (zio->io_error && vd == NULL &&
3424 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
3425 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
3426 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
3427 zio->io_error = 0;
3428 zio->io_flags |= ZIO_FLAG_IO_RETRY |
3429 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
3430 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
3431 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
3432 zio_requeue_io_start_cut_in_line);
3433 return (ZIO_PIPELINE_STOP);
3434 }
3435
3436 /*
3437 * If we got an error on a leaf device, convert it to ENXIO
3438 * if the device is not accessible at all.
3439 */
3440 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3441 !vdev_accessible(vd, zio))
3442 zio->io_error = SET_ERROR(ENXIO);
3443
3444 /*
3445 * If we can't write to an interior vdev (mirror or RAID-Z),
3446 * set vdev_cant_write so that we stop trying to allocate from it.
3447 */
3448 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
3449 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
3450 vd->vdev_cant_write = B_TRUE;
3451 }
3452
3453 /*
3454 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3455 * attempts will ever succeed. In this case we set a persistent bit so
3456 * that we don't bother with it in the future.
3457 */
3458 if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) &&
3459 zio->io_type == ZIO_TYPE_IOCTL &&
3460 zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL)
3461 vd->vdev_nowritecache = B_TRUE;
3462
3463 if (zio->io_error)
3464 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3465
3466 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3467 zio->io_physdone != NULL) {
3468 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
3469 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
3470 zio->io_physdone(zio->io_logical);
3471 }
3472
3473 return (ZIO_PIPELINE_CONTINUE);
3474}
3475
3476void
3477zio_vdev_io_reissue(zio_t *zio)
3478{
3479 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3480 ASSERT(zio->io_error == 0);
3481
3482 zio->io_stage >>= 1;
3483}
3484
3485void
3486zio_vdev_io_redone(zio_t *zio)
3487{
3488 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
3489
3490 zio->io_stage >>= 1;
3491}
3492
3493void
3494zio_vdev_io_bypass(zio_t *zio)
3495{
3496 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3497 ASSERT(zio->io_error == 0);
3498
3499 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
3500 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
3501}
3502
3503/*
3504 * ==========================================================================
3505 * Generate and verify checksums
3506 * ==========================================================================
3507 */
3508static int
3509zio_checksum_generate(zio_t *zio)
3510{
3511 blkptr_t *bp = zio->io_bp;
3512 enum zio_checksum checksum;
3513
3514 if (bp == NULL) {
3515 /*
3516 * This is zio_write_phys().
3517 * We're either generating a label checksum, or none at all.
3518 */
3519 checksum = zio->io_prop.zp_checksum;
3520
3521 if (checksum == ZIO_CHECKSUM_OFF)
3522 return (ZIO_PIPELINE_CONTINUE);
3523
3524 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
3525 } else {
3526 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
3527 ASSERT(!IO_IS_ALLOCATING(zio));
3528 checksum = ZIO_CHECKSUM_GANG_HEADER;
3529 } else {
3530 checksum = BP_GET_CHECKSUM(bp);
3531 }
3532 }
3533
3534 zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size);
3535
3536 return (ZIO_PIPELINE_CONTINUE);
3537}
3538
3539static int
3540zio_checksum_verify(zio_t *zio)
3541{
3542 zio_bad_cksum_t info;
3543 blkptr_t *bp = zio->io_bp;
3544 int error;
3545
3546 ASSERT(zio->io_vd != NULL);
3547
3548 if (bp == NULL) {
3549 /*
3550 * This is zio_read_phys().
3551 * We're either verifying a label checksum, or nothing at all.
3552 */
3553 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
3554 return (ZIO_PIPELINE_CONTINUE);
3555
3556 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
3557 }
3558
3559 if ((error = zio_checksum_error(zio, &info)) != 0) {
3560 zio->io_error = error;
3561 if (error == ECKSUM &&
3562 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
3563 zfs_ereport_start_checksum(zio->io_spa,
3564 zio->io_vd, zio, zio->io_offset,
3565 zio->io_size, NULL, &info);
3566 }
3567 }
3568
3569 return (ZIO_PIPELINE_CONTINUE);
3570}
3571
3572/*
3573 * Called by RAID-Z to ensure we don't compute the checksum twice.
3574 */
3575void
3576zio_checksum_verified(zio_t *zio)
3577{
3578 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
3579}
3580
3581/*
3582 * ==========================================================================
3583 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3584 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3585 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3586 * indicate errors that are specific to one I/O, and most likely permanent.
3587 * Any other error is presumed to be worse because we weren't expecting it.
3588 * ==========================================================================
3589 */
3590int
3591zio_worst_error(int e1, int e2)
3592{
3593 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
3594 int r1, r2;
3595
3596 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
3597 if (e1 == zio_error_rank[r1])
3598 break;
3599
3600 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3601 if (e2 == zio_error_rank[r2])
3602 break;
3603
3604 return (r1 > r2 ? e1 : e2);
3605}
3606
3607/*
3608 * ==========================================================================
3609 * I/O completion
3610 * ==========================================================================
3611 */
3612static int
3613zio_ready(zio_t *zio)
3614{
3615 blkptr_t *bp = zio->io_bp;
3616 zio_t *pio, *pio_next;
3617 zio_link_t *zl = NULL;
3618
3619 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT,
3620 ZIO_WAIT_READY)) {
3621 return (ZIO_PIPELINE_STOP);
3622 }
3623
3624 if (zio->io_ready) {
3625 ASSERT(IO_IS_ALLOCATING(zio));
3626 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3627 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3628 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3629
3630 zio->io_ready(zio);
3631 }
3632
3633 if (bp != NULL && bp != &zio->io_bp_copy)
3634 zio->io_bp_copy = *bp;
3635
3636 if (zio->io_error != 0) {
3637 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3638
3639 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3640 ASSERT(IO_IS_ALLOCATING(zio));
3641 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3642 /*
3643 * We were unable to allocate anything, unreserve and
3644 * issue the next I/O to allocate.
3645 */
3646 metaslab_class_throttle_unreserve(
3647 spa_normal_class(zio->io_spa),
3648 zio->io_prop.zp_copies, zio);
3649 zio_allocate_dispatch(zio->io_spa);
3650 }
3651 }
3652
3653 mutex_enter(&zio->io_lock);
3654 zio->io_state[ZIO_WAIT_READY] = 1;
3655 pio = zio_walk_parents(zio, &zl);
3656 mutex_exit(&zio->io_lock);
3657
3658 /*
3659 * As we notify zio's parents, new parents could be added.
3660 * New parents go to the head of zio's io_parent_list, however,
3661 * so we will (correctly) not notify them. The remainder of zio's
3662 * io_parent_list, from 'pio_next' onward, cannot change because
3663 * all parents must wait for us to be done before they can be done.
3664 */
3665 for (; pio != NULL; pio = pio_next) {
3666 pio_next = zio_walk_parents(zio, &zl);
3667 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3668 }
3669
3670 if (zio->io_flags & ZIO_FLAG_NODATA) {
3671 if (BP_IS_GANG(bp)) {
3672 zio->io_flags &= ~ZIO_FLAG_NODATA;
3673 } else {
3674 ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE);
3675 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3676 }
3677 }
3678
3679 if (zio_injection_enabled &&
3680 zio->io_spa->spa_syncing_txg == zio->io_txg)
3681 zio_handle_ignored_writes(zio);
3682
3683 return (ZIO_PIPELINE_CONTINUE);
3684}
3685
3686/*
3687 * Update the allocation throttle accounting.
3688 */
3689static void
3690zio_dva_throttle_done(zio_t *zio)
3691{
3692 zio_t *lio = zio->io_logical;
3693 zio_t *pio = zio_unique_parent(zio);
3694 vdev_t *vd = zio->io_vd;
3695 int flags = METASLAB_ASYNC_ALLOC;
3696
3697 ASSERT3P(zio->io_bp, !=, NULL);
3698 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
3699 ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
3700 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
3701 ASSERT(vd != NULL);
3702 ASSERT3P(vd, ==, vd->vdev_top);
3703 ASSERT(!(zio->io_flags & (ZIO_FLAG_IO_REPAIR | ZIO_FLAG_IO_RETRY)));
3704 ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
3705 ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
3706 ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
3707
3708 /*
3709 * Parents of gang children can have two flavors -- ones that
3710 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
3711 * and ones that allocated the constituent blocks. The allocation
3712 * throttle needs to know the allocating parent zio so we must find
3713 * it here.
3714 */
3715 if (pio->io_child_type == ZIO_CHILD_GANG) {
3716 /*
3717 * If our parent is a rewrite gang child then our grandparent
3718 * would have been the one that performed the allocation.
3719 */
3720 if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
3721 pio = zio_unique_parent(pio);
3722 flags |= METASLAB_GANG_CHILD;
3723 }
3724
3725 ASSERT(IO_IS_ALLOCATING(pio));
3726 ASSERT3P(zio, !=, zio->io_logical);
3727 ASSERT(zio->io_logical != NULL);
3728 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
3729 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
3730
3731 mutex_enter(&pio->io_lock);
3732 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags);
3733 mutex_exit(&pio->io_lock);
3734
3735 metaslab_class_throttle_unreserve(spa_normal_class(zio->io_spa),
3736 1, pio);
3737
3738 /*
3739 * Call into the pipeline to see if there is more work that
3740 * needs to be done. If there is work to be done it will be
3741 * dispatched to another taskq thread.
3742 */
3743 zio_allocate_dispatch(zio->io_spa);
3744}
3745
3746static int
3747zio_done(zio_t *zio)
3748{
3749 spa_t *spa = zio->io_spa;
3750 zio_t *lio = zio->io_logical;
3751 blkptr_t *bp = zio->io_bp;
3752 vdev_t *vd = zio->io_vd;
3753 uint64_t psize = zio->io_size;
3754 zio_t *pio, *pio_next;
3755 metaslab_class_t *mc = spa_normal_class(spa);
3756 zio_link_t *zl = NULL;
3757
3758 /*
3759 * If our children haven't all completed,
3760 * wait for them and then repeat this pipeline stage.
3761 */
3762 if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) {
3763 return (ZIO_PIPELINE_STOP);
3764 }
3765
3766 /*
3767 * If the allocation throttle is enabled, then update the accounting.
3768 * We only track child I/Os that are part of an allocating async
3769 * write. We must do this since the allocation is performed
3770 * by the logical I/O but the actual write is done by child I/Os.
3771 */
3772 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
3773 zio->io_child_type == ZIO_CHILD_VDEV) {
3774 ASSERT(mc->mc_alloc_throttle_enabled);
3775 zio_dva_throttle_done(zio);
3776 }
3777
3778 /*
3779 * If the allocation throttle is enabled, verify that
3780 * we have decremented the refcounts for every I/O that was throttled.
3781 */
3782 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3783 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3784 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3785 ASSERT(bp != NULL);
3786 metaslab_group_alloc_verify(spa, zio->io_bp, zio);
3787 VERIFY(refcount_not_held(&mc->mc_alloc_slots, zio));
3788 }
3789
3790 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3791 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3792 ASSERT(zio->io_children[c][w] == 0);
3793
3794 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3795 ASSERT(bp->blk_pad[0] == 0);
3796 ASSERT(bp->blk_pad[1] == 0);
3797 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3798 (bp == zio_unique_parent(zio)->io_bp));
3799 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3800 zio->io_bp_override == NULL &&
3801 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3802 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3803 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3804 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3805 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3806 }
3807 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3808 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3809 }
3810
3811 /*
3812 * If there were child vdev/gang/ddt errors, they apply to us now.
3813 */
3814 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3815 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3816 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3817
3818 /*
3819 * If the I/O on the transformed data was successful, generate any
3820 * checksum reports now while we still have the transformed data.
3821 */
3822 if (zio->io_error == 0) {
3823 while (zio->io_cksum_report != NULL) {
3824 zio_cksum_report_t *zcr = zio->io_cksum_report;
3825 uint64_t align = zcr->zcr_align;
3826 uint64_t asize = P2ROUNDUP(psize, align);
3827 char *abuf = NULL;
3828 abd_t *adata = zio->io_abd;
3829
3830 if (asize != psize) {
3831 adata = abd_alloc_linear(asize, B_TRUE);
3832 abd_copy(adata, zio->io_abd, psize);
3833 abd_zero_off(adata, psize, asize - psize);
3834 }
3835
3836 if (adata != NULL)
3837 abuf = abd_borrow_buf_copy(adata, asize);
3838
3839 zio->io_cksum_report = zcr->zcr_next;
3840 zcr->zcr_next = NULL;
3841 zcr->zcr_finish(zcr, abuf);
3842 zfs_ereport_free_checksum(zcr);
3843
3844 if (adata != NULL)
3845 abd_return_buf(adata, abuf, asize);
3846
3847 if (asize != psize)
3848 abd_free(adata);
3849 }
3850 }
3851
3852 zio_pop_transforms(zio); /* note: may set zio->io_error */
3853
3854 vdev_stat_update(zio, psize);
3855
3856 if (zio->io_error) {
3857 /*
3858 * If this I/O is attached to a particular vdev,
3859 * generate an error message describing the I/O failure
3860 * at the block level. We ignore these errors if the
3861 * device is currently unavailable.
3862 */
3863 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3864 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3865
3866 if ((zio->io_error == EIO || !(zio->io_flags &
3867 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3868 zio == lio) {
3869 /*
3870 * For logical I/O requests, tell the SPA to log the
3871 * error and generate a logical data ereport.
3872 */
3873 spa_log_error(spa, zio);
3874 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3875 0, 0);
3876 }
3877 }
3878
3879 if (zio->io_error && zio == lio) {
3880 /*
3881 * Determine whether zio should be reexecuted. This will
3882 * propagate all the way to the root via zio_notify_parent().
3883 */
3884 ASSERT(vd == NULL && bp != NULL);
3885 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3886
3887 if (IO_IS_ALLOCATING(zio) &&
3888 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3889 if (zio->io_error != ENOSPC)
3890 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3891 else
3892 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3893 }
3894
3895 if ((zio->io_type == ZIO_TYPE_READ ||
3896 zio->io_type == ZIO_TYPE_FREE) &&
3897 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3898 zio->io_error == ENXIO &&
3899 spa_load_state(spa) == SPA_LOAD_NONE &&
3900 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3901 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3902
3903 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3904 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3905
3906 /*
3907 * Here is a possibly good place to attempt to do
3908 * either combinatorial reconstruction or error correction
3909 * based on checksums. It also might be a good place
3910 * to send out preliminary ereports before we suspend
3911 * processing.
3912 */
3913 }
3914
3915 /*
3916 * If there were logical child errors, they apply to us now.
3917 * We defer this until now to avoid conflating logical child
3918 * errors with errors that happened to the zio itself when
3919 * updating vdev stats and reporting FMA events above.
3920 */
3921 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3922
3923 if ((zio->io_error || zio->io_reexecute) &&
3924 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3925 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3926 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3927
3928 zio_gang_tree_free(&zio->io_gang_tree);
3929
3930 /*
3931 * Godfather I/Os should never suspend.
3932 */
3933 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3934 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3935 zio->io_reexecute = 0;
3936
3937 if (zio->io_reexecute) {
3938 /*
3939 * This is a logical I/O that wants to reexecute.
3940 *
3941 * Reexecute is top-down. When an i/o fails, if it's not
3942 * the root, it simply notifies its parent and sticks around.
3943 * The parent, seeing that it still has children in zio_done(),
3944 * does the same. This percolates all the way up to the root.
3945 * The root i/o will reexecute or suspend the entire tree.
3946 *
3947 * This approach ensures that zio_reexecute() honors
3948 * all the original i/o dependency relationships, e.g.
3949 * parents not executing until children are ready.
3950 */
3951 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3952
3953 zio->io_gang_leader = NULL;
3954
3955 mutex_enter(&zio->io_lock);
3956 zio->io_state[ZIO_WAIT_DONE] = 1;
3957 mutex_exit(&zio->io_lock);
3958
3959 /*
3960 * "The Godfather" I/O monitors its children but is
3961 * not a true parent to them. It will track them through
3962 * the pipeline but severs its ties whenever they get into
3963 * trouble (e.g. suspended). This allows "The Godfather"
3964 * I/O to return status without blocking.
3965 */
3966 zl = NULL;
3967 for (pio = zio_walk_parents(zio, &zl); pio != NULL;
3968 pio = pio_next) {
3969 zio_link_t *remove_zl = zl;
3970 pio_next = zio_walk_parents(zio, &zl);
3971
3972 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3973 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3974 zio_remove_child(pio, zio, remove_zl);
3975 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3976 }
3977 }
3978
3979 if ((pio = zio_unique_parent(zio)) != NULL) {
3980 /*
3981 * We're not a root i/o, so there's nothing to do
3982 * but notify our parent. Don't propagate errors
3983 * upward since we haven't permanently failed yet.
3984 */
3985 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3986 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3987 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3988 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3989 /*
3990 * We'd fail again if we reexecuted now, so suspend
3991 * until conditions improve (e.g. device comes online).
3992 */
3993 zio_suspend(spa, zio);
3994 } else {
3995 /*
3996 * Reexecution is potentially a huge amount of work.
3997 * Hand it off to the otherwise-unused claim taskq.
3998 */
3999#if defined(illumos) || !defined(_KERNEL)
4000 ASSERT(zio->io_tqent.tqent_next == NULL);
4001#else
4002 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
4003#endif
4004 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
4005 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
4006 0, &zio->io_tqent);
4007 }
4008 return (ZIO_PIPELINE_STOP);
4009 }
4010
4011 ASSERT(zio->io_child_count == 0);
4012 ASSERT(zio->io_reexecute == 0);
4013 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
4014
4015 /*
4016 * Report any checksum errors, since the I/O is complete.
4017 */
4018 while (zio->io_cksum_report != NULL) {
4019 zio_cksum_report_t *zcr = zio->io_cksum_report;
4020 zio->io_cksum_report = zcr->zcr_next;
4021 zcr->zcr_next = NULL;
4022 zcr->zcr_finish(zcr, NULL);
4023 zfs_ereport_free_checksum(zcr);
4024 }
4025
4026 /*
4027 * It is the responsibility of the done callback to ensure that this
4028 * particular zio is no longer discoverable for adoption, and as
4029 * such, cannot acquire any new parents.
4030 */
4031 if (zio->io_done)
4032 zio->io_done(zio);
4033
4034 mutex_enter(&zio->io_lock);
4035 zio->io_state[ZIO_WAIT_DONE] = 1;
4036 mutex_exit(&zio->io_lock);
4037
4038 zl = NULL;
4039 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
4040 zio_link_t *remove_zl = zl;
4041 pio_next = zio_walk_parents(zio, &zl);
4042 zio_remove_child(pio, zio, remove_zl);
4043 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4044 }
4045
4046 if (zio->io_waiter != NULL) {
4047 mutex_enter(&zio->io_lock);
4048 zio->io_executor = NULL;
4049 cv_broadcast(&zio->io_cv);
4050 mutex_exit(&zio->io_lock);
4051 } else {
4052 zio_destroy(zio);
4053 }
4054
4055 return (ZIO_PIPELINE_STOP);
4056}
4057
4058/*
4059 * ==========================================================================
4060 * I/O pipeline definition
4061 * ==========================================================================
4062 */
4063static zio_pipe_stage_t *zio_pipeline[] = {
4064 NULL,
4065 zio_read_bp_init,
4066 zio_write_bp_init,
4067 zio_free_bp_init,
4068 zio_issue_async,
4069 zio_write_compress,
4070 zio_checksum_generate,
4071 zio_nop_write,
4072 zio_ddt_read_start,
4073 zio_ddt_read_done,
4074 zio_ddt_write,
4075 zio_ddt_free,
4076 zio_gang_assemble,
4077 zio_gang_issue,
4078 zio_dva_throttle,
4079 zio_dva_allocate,
4080 zio_dva_free,
4081 zio_dva_claim,
4082 zio_ready,
4083 zio_vdev_io_start,
4084 zio_vdev_io_done,
4085 zio_vdev_io_assess,
4086 zio_checksum_verify,
4087 zio_done
4088};
4089
4090
4091
4092
4093/*
4094 * Compare two zbookmark_phys_t's to see which we would reach first in a
4095 * pre-order traversal of the object tree.
4096 *
4097 * This is simple in every case aside from the meta-dnode object. For all other
4098 * objects, we traverse them in order (object 1 before object 2, and so on).
4099 * However, all of these objects are traversed while traversing object 0, since
4100 * the data it points to is the list of objects. Thus, we need to convert to a
4101 * canonical representation so we can compare meta-dnode bookmarks to
4102 * non-meta-dnode bookmarks.
4103 *
4104 * We do this by calculating "equivalents" for each field of the zbookmark.
4105 * zbookmarks outside of the meta-dnode use their own object and level, and
4106 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4107 * blocks this bookmark refers to) by multiplying their blkid by their span
4108 * (the number of L0 blocks contained within one block at their level).
4109 * zbookmarks inside the meta-dnode calculate their object equivalent
4110 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4111 * level + 1<<31 (any value larger than a level could ever be) for their level.
4112 * This causes them to always compare before a bookmark in their object
4113 * equivalent, compare appropriately to bookmarks in other objects, and to
4114 * compare appropriately to other bookmarks in the meta-dnode.
4115 */
4116int
4117zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
4118 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
4119{
4120 /*
4121 * These variables represent the "equivalent" values for the zbookmark,
4122 * after converting zbookmarks inside the meta dnode to their
4123 * normal-object equivalents.
4124 */
4125 uint64_t zb1obj, zb2obj;
4126 uint64_t zb1L0, zb2L0;
4127 uint64_t zb1level, zb2level;
4128
4129 if (zb1->zb_object == zb2->zb_object &&
4130 zb1->zb_level == zb2->zb_level &&
4131 zb1->zb_blkid == zb2->zb_blkid)
4132 return (0);
4133
4134 /*
4135 * BP_SPANB calculates the span in blocks.
4136 */
4137 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
4138 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
4139
4140 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
4141 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4142 zb1L0 = 0;
4143 zb1level = zb1->zb_level + COMPARE_META_LEVEL;
4144 } else {
4145 zb1obj = zb1->zb_object;
4146 zb1level = zb1->zb_level;
4147 }
4148
4149 if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
4150 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4151 zb2L0 = 0;
4152 zb2level = zb2->zb_level + COMPARE_META_LEVEL;
4153 } else {
4154 zb2obj = zb2->zb_object;
4155 zb2level = zb2->zb_level;
4156 }
4157
4158 /* Now that we have a canonical representation, do the comparison. */
4159 if (zb1obj != zb2obj)
4160 return (zb1obj < zb2obj ? -1 : 1);
4161 else if (zb1L0 != zb2L0)
4162 return (zb1L0 < zb2L0 ? -1 : 1);
4163 else if (zb1level != zb2level)
4164 return (zb1level > zb2level ? -1 : 1);
4165 /*
4166 * This can (theoretically) happen if the bookmarks have the same object
4167 * and level, but different blkids, if the block sizes are not the same.
4168 * There is presently no way to change the indirect block sizes
4169 */
4170 return (0);
4171}
4172
4173/*
4174 * This function checks the following: given that last_block is the place that
4175 * our traversal stopped last time, does that guarantee that we've visited
4176 * every node under subtree_root? Therefore, we can't just use the raw output
4177 * of zbookmark_compare. We have to pass in a modified version of
4178 * subtree_root; by incrementing the block id, and then checking whether
4179 * last_block is before or equal to that, we can tell whether or not having
4180 * visited last_block implies that all of subtree_root's children have been
4181 * visited.
4182 */
4183boolean_t
4184zbookmark_subtree_completed(const dnode_phys_t *dnp,
4185 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
4186{
4187 zbookmark_phys_t mod_zb = *subtree_root;
4188 mod_zb.zb_blkid++;
4189 ASSERT(last_block->zb_level == 0);
4190
4191 /* The objset_phys_t isn't before anything. */
4192 if (dnp == NULL)
4193 return (B_FALSE);
4194
4195 /*
4196 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4197 * data block size in sectors, because that variable is only used if
4198 * the bookmark refers to a block in the meta-dnode. Since we don't
4199 * know without examining it what object it refers to, and there's no
4200 * harm in passing in this value in other cases, we always pass it in.
4201 *
4202 * We pass in 0 for the indirect block size shift because zb2 must be
4203 * level 0. The indirect block size is only used to calculate the span
4204 * of the bookmark, but since the bookmark must be level 0, the span is
4205 * always 1, so the math works out.
4206 *
4207 * If you make changes to how the zbookmark_compare code works, be sure
4208 * to make sure that this code still works afterwards.
4209 */
4210 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
4211 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,
4212 last_block) <= 0);
4213}
| 3389 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3390 * disk, and use that to finish the checksum ereport later.
3391 */
3392static void
3393zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
3394 const void *good_buf)
3395{
3396 /* no processing needed */
3397 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
3398}
3399
3400/*ARGSUSED*/
3401void
3402zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
3403{
3404 void *buf = zio_buf_alloc(zio->io_size);
3405
3406 abd_copy_to_buf(buf, zio->io_abd, zio->io_size);
3407
3408 zcr->zcr_cbinfo = zio->io_size;
3409 zcr->zcr_cbdata = buf;
3410 zcr->zcr_finish = zio_vsd_default_cksum_finish;
3411 zcr->zcr_free = zio_buf_free;
3412}
3413
3414static int
3415zio_vdev_io_assess(zio_t *zio)
3416{
3417 vdev_t *vd = zio->io_vd;
3418
3419 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3420 return (ZIO_PIPELINE_STOP);
3421 }
3422
3423 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3424 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
3425
3426 if (zio->io_vsd != NULL) {
3427 zio->io_vsd_ops->vsd_free(zio);
3428 zio->io_vsd = NULL;
3429 }
3430
3431 if (zio_injection_enabled && zio->io_error == 0)
3432 zio->io_error = zio_handle_fault_injection(zio, EIO);
3433
3434 if (zio->io_type == ZIO_TYPE_FREE &&
3435 zio->io_priority != ZIO_PRIORITY_NOW) {
3436 switch (zio->io_error) {
3437 case 0:
3438 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
3439 ZIO_TRIM_STAT_BUMP(success);
3440 break;
3441 case EOPNOTSUPP:
3442 ZIO_TRIM_STAT_BUMP(unsupported);
3443 break;
3444 default:
3445 ZIO_TRIM_STAT_BUMP(failed);
3446 break;
3447 }
3448 }
3449
3450 /*
3451 * If the I/O failed, determine whether we should attempt to retry it.
3452 *
3453 * On retry, we cut in line in the issue queue, since we don't want
3454 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3455 */
3456 if (zio->io_error && vd == NULL &&
3457 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
3458 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
3459 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
3460 zio->io_error = 0;
3461 zio->io_flags |= ZIO_FLAG_IO_RETRY |
3462 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
3463 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
3464 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
3465 zio_requeue_io_start_cut_in_line);
3466 return (ZIO_PIPELINE_STOP);
3467 }
3468
3469 /*
3470 * If we got an error on a leaf device, convert it to ENXIO
3471 * if the device is not accessible at all.
3472 */
3473 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3474 !vdev_accessible(vd, zio))
3475 zio->io_error = SET_ERROR(ENXIO);
3476
3477 /*
3478 * If we can't write to an interior vdev (mirror or RAID-Z),
3479 * set vdev_cant_write so that we stop trying to allocate from it.
3480 */
3481 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
3482 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
3483 vd->vdev_cant_write = B_TRUE;
3484 }
3485
3486 /*
3487 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3488 * attempts will ever succeed. In this case we set a persistent bit so
3489 * that we don't bother with it in the future.
3490 */
3491 if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) &&
3492 zio->io_type == ZIO_TYPE_IOCTL &&
3493 zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL)
3494 vd->vdev_nowritecache = B_TRUE;
3495
3496 if (zio->io_error)
3497 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3498
3499 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3500 zio->io_physdone != NULL) {
3501 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
3502 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
3503 zio->io_physdone(zio->io_logical);
3504 }
3505
3506 return (ZIO_PIPELINE_CONTINUE);
3507}
3508
3509void
3510zio_vdev_io_reissue(zio_t *zio)
3511{
3512 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3513 ASSERT(zio->io_error == 0);
3514
3515 zio->io_stage >>= 1;
3516}
3517
3518void
3519zio_vdev_io_redone(zio_t *zio)
3520{
3521 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
3522
3523 zio->io_stage >>= 1;
3524}
3525
3526void
3527zio_vdev_io_bypass(zio_t *zio)
3528{
3529 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3530 ASSERT(zio->io_error == 0);
3531
3532 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
3533 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
3534}
3535
3536/*
3537 * ==========================================================================
3538 * Generate and verify checksums
3539 * ==========================================================================
3540 */
3541static int
3542zio_checksum_generate(zio_t *zio)
3543{
3544 blkptr_t *bp = zio->io_bp;
3545 enum zio_checksum checksum;
3546
3547 if (bp == NULL) {
3548 /*
3549 * This is zio_write_phys().
3550 * We're either generating a label checksum, or none at all.
3551 */
3552 checksum = zio->io_prop.zp_checksum;
3553
3554 if (checksum == ZIO_CHECKSUM_OFF)
3555 return (ZIO_PIPELINE_CONTINUE);
3556
3557 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
3558 } else {
3559 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
3560 ASSERT(!IO_IS_ALLOCATING(zio));
3561 checksum = ZIO_CHECKSUM_GANG_HEADER;
3562 } else {
3563 checksum = BP_GET_CHECKSUM(bp);
3564 }
3565 }
3566
3567 zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size);
3568
3569 return (ZIO_PIPELINE_CONTINUE);
3570}
3571
3572static int
3573zio_checksum_verify(zio_t *zio)
3574{
3575 zio_bad_cksum_t info;
3576 blkptr_t *bp = zio->io_bp;
3577 int error;
3578
3579 ASSERT(zio->io_vd != NULL);
3580
3581 if (bp == NULL) {
3582 /*
3583 * This is zio_read_phys().
3584 * We're either verifying a label checksum, or nothing at all.
3585 */
3586 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
3587 return (ZIO_PIPELINE_CONTINUE);
3588
3589 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
3590 }
3591
3592 if ((error = zio_checksum_error(zio, &info)) != 0) {
3593 zio->io_error = error;
3594 if (error == ECKSUM &&
3595 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
3596 zfs_ereport_start_checksum(zio->io_spa,
3597 zio->io_vd, zio, zio->io_offset,
3598 zio->io_size, NULL, &info);
3599 }
3600 }
3601
3602 return (ZIO_PIPELINE_CONTINUE);
3603}
3604
3605/*
3606 * Called by RAID-Z to ensure we don't compute the checksum twice.
3607 */
3608void
3609zio_checksum_verified(zio_t *zio)
3610{
3611 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
3612}
3613
3614/*
3615 * ==========================================================================
3616 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3617 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3618 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3619 * indicate errors that are specific to one I/O, and most likely permanent.
3620 * Any other error is presumed to be worse because we weren't expecting it.
3621 * ==========================================================================
3622 */
3623int
3624zio_worst_error(int e1, int e2)
3625{
3626 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
3627 int r1, r2;
3628
3629 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
3630 if (e1 == zio_error_rank[r1])
3631 break;
3632
3633 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3634 if (e2 == zio_error_rank[r2])
3635 break;
3636
3637 return (r1 > r2 ? e1 : e2);
3638}
3639
3640/*
3641 * ==========================================================================
3642 * I/O completion
3643 * ==========================================================================
3644 */
3645static int
3646zio_ready(zio_t *zio)
3647{
3648 blkptr_t *bp = zio->io_bp;
3649 zio_t *pio, *pio_next;
3650 zio_link_t *zl = NULL;
3651
3652 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT,
3653 ZIO_WAIT_READY)) {
3654 return (ZIO_PIPELINE_STOP);
3655 }
3656
3657 if (zio->io_ready) {
3658 ASSERT(IO_IS_ALLOCATING(zio));
3659 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3660 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3661 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3662
3663 zio->io_ready(zio);
3664 }
3665
3666 if (bp != NULL && bp != &zio->io_bp_copy)
3667 zio->io_bp_copy = *bp;
3668
3669 if (zio->io_error != 0) {
3670 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3671
3672 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3673 ASSERT(IO_IS_ALLOCATING(zio));
3674 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3675 /*
3676 * We were unable to allocate anything, unreserve and
3677 * issue the next I/O to allocate.
3678 */
3679 metaslab_class_throttle_unreserve(
3680 spa_normal_class(zio->io_spa),
3681 zio->io_prop.zp_copies, zio);
3682 zio_allocate_dispatch(zio->io_spa);
3683 }
3684 }
3685
3686 mutex_enter(&zio->io_lock);
3687 zio->io_state[ZIO_WAIT_READY] = 1;
3688 pio = zio_walk_parents(zio, &zl);
3689 mutex_exit(&zio->io_lock);
3690
3691 /*
3692 * As we notify zio's parents, new parents could be added.
3693 * New parents go to the head of zio's io_parent_list, however,
3694 * so we will (correctly) not notify them. The remainder of zio's
3695 * io_parent_list, from 'pio_next' onward, cannot change because
3696 * all parents must wait for us to be done before they can be done.
3697 */
3698 for (; pio != NULL; pio = pio_next) {
3699 pio_next = zio_walk_parents(zio, &zl);
3700 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3701 }
3702
3703 if (zio->io_flags & ZIO_FLAG_NODATA) {
3704 if (BP_IS_GANG(bp)) {
3705 zio->io_flags &= ~ZIO_FLAG_NODATA;
3706 } else {
3707 ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE);
3708 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3709 }
3710 }
3711
3712 if (zio_injection_enabled &&
3713 zio->io_spa->spa_syncing_txg == zio->io_txg)
3714 zio_handle_ignored_writes(zio);
3715
3716 return (ZIO_PIPELINE_CONTINUE);
3717}
3718
3719/*
3720 * Update the allocation throttle accounting.
3721 */
3722static void
3723zio_dva_throttle_done(zio_t *zio)
3724{
3725 zio_t *lio = zio->io_logical;
3726 zio_t *pio = zio_unique_parent(zio);
3727 vdev_t *vd = zio->io_vd;
3728 int flags = METASLAB_ASYNC_ALLOC;
3729
3730 ASSERT3P(zio->io_bp, !=, NULL);
3731 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
3732 ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
3733 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
3734 ASSERT(vd != NULL);
3735 ASSERT3P(vd, ==, vd->vdev_top);
3736 ASSERT(!(zio->io_flags & (ZIO_FLAG_IO_REPAIR | ZIO_FLAG_IO_RETRY)));
3737 ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
3738 ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
3739 ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
3740
3741 /*
3742 * Parents of gang children can have two flavors -- ones that
3743 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
3744 * and ones that allocated the constituent blocks. The allocation
3745 * throttle needs to know the allocating parent zio so we must find
3746 * it here.
3747 */
3748 if (pio->io_child_type == ZIO_CHILD_GANG) {
3749 /*
3750 * If our parent is a rewrite gang child then our grandparent
3751 * would have been the one that performed the allocation.
3752 */
3753 if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
3754 pio = zio_unique_parent(pio);
3755 flags |= METASLAB_GANG_CHILD;
3756 }
3757
3758 ASSERT(IO_IS_ALLOCATING(pio));
3759 ASSERT3P(zio, !=, zio->io_logical);
3760 ASSERT(zio->io_logical != NULL);
3761 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
3762 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
3763
3764 mutex_enter(&pio->io_lock);
3765 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags);
3766 mutex_exit(&pio->io_lock);
3767
3768 metaslab_class_throttle_unreserve(spa_normal_class(zio->io_spa),
3769 1, pio);
3770
3771 /*
3772 * Call into the pipeline to see if there is more work that
3773 * needs to be done. If there is work to be done it will be
3774 * dispatched to another taskq thread.
3775 */
3776 zio_allocate_dispatch(zio->io_spa);
3777}
3778
3779static int
3780zio_done(zio_t *zio)
3781{
3782 spa_t *spa = zio->io_spa;
3783 zio_t *lio = zio->io_logical;
3784 blkptr_t *bp = zio->io_bp;
3785 vdev_t *vd = zio->io_vd;
3786 uint64_t psize = zio->io_size;
3787 zio_t *pio, *pio_next;
3788 metaslab_class_t *mc = spa_normal_class(spa);
3789 zio_link_t *zl = NULL;
3790
3791 /*
3792 * If our children haven't all completed,
3793 * wait for them and then repeat this pipeline stage.
3794 */
3795 if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) {
3796 return (ZIO_PIPELINE_STOP);
3797 }
3798
3799 /*
3800 * If the allocation throttle is enabled, then update the accounting.
3801 * We only track child I/Os that are part of an allocating async
3802 * write. We must do this since the allocation is performed
3803 * by the logical I/O but the actual write is done by child I/Os.
3804 */
3805 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
3806 zio->io_child_type == ZIO_CHILD_VDEV) {
3807 ASSERT(mc->mc_alloc_throttle_enabled);
3808 zio_dva_throttle_done(zio);
3809 }
3810
3811 /*
3812 * If the allocation throttle is enabled, verify that
3813 * we have decremented the refcounts for every I/O that was throttled.
3814 */
3815 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3816 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3817 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3818 ASSERT(bp != NULL);
3819 metaslab_group_alloc_verify(spa, zio->io_bp, zio);
3820 VERIFY(refcount_not_held(&mc->mc_alloc_slots, zio));
3821 }
3822
3823 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3824 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3825 ASSERT(zio->io_children[c][w] == 0);
3826
3827 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3828 ASSERT(bp->blk_pad[0] == 0);
3829 ASSERT(bp->blk_pad[1] == 0);
3830 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3831 (bp == zio_unique_parent(zio)->io_bp));
3832 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3833 zio->io_bp_override == NULL &&
3834 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3835 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3836 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3837 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3838 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3839 }
3840 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3841 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3842 }
3843
3844 /*
3845 * If there were child vdev/gang/ddt errors, they apply to us now.
3846 */
3847 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3848 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3849 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3850
3851 /*
3852 * If the I/O on the transformed data was successful, generate any
3853 * checksum reports now while we still have the transformed data.
3854 */
3855 if (zio->io_error == 0) {
3856 while (zio->io_cksum_report != NULL) {
3857 zio_cksum_report_t *zcr = zio->io_cksum_report;
3858 uint64_t align = zcr->zcr_align;
3859 uint64_t asize = P2ROUNDUP(psize, align);
3860 char *abuf = NULL;
3861 abd_t *adata = zio->io_abd;
3862
3863 if (asize != psize) {
3864 adata = abd_alloc_linear(asize, B_TRUE);
3865 abd_copy(adata, zio->io_abd, psize);
3866 abd_zero_off(adata, psize, asize - psize);
3867 }
3868
3869 if (adata != NULL)
3870 abuf = abd_borrow_buf_copy(adata, asize);
3871
3872 zio->io_cksum_report = zcr->zcr_next;
3873 zcr->zcr_next = NULL;
3874 zcr->zcr_finish(zcr, abuf);
3875 zfs_ereport_free_checksum(zcr);
3876
3877 if (adata != NULL)
3878 abd_return_buf(adata, abuf, asize);
3879
3880 if (asize != psize)
3881 abd_free(adata);
3882 }
3883 }
3884
3885 zio_pop_transforms(zio); /* note: may set zio->io_error */
3886
3887 vdev_stat_update(zio, psize);
3888
3889 if (zio->io_error) {
3890 /*
3891 * If this I/O is attached to a particular vdev,
3892 * generate an error message describing the I/O failure
3893 * at the block level. We ignore these errors if the
3894 * device is currently unavailable.
3895 */
3896 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3897 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3898
3899 if ((zio->io_error == EIO || !(zio->io_flags &
3900 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3901 zio == lio) {
3902 /*
3903 * For logical I/O requests, tell the SPA to log the
3904 * error and generate a logical data ereport.
3905 */
3906 spa_log_error(spa, zio);
3907 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3908 0, 0);
3909 }
3910 }
3911
3912 if (zio->io_error && zio == lio) {
3913 /*
3914 * Determine whether zio should be reexecuted. This will
3915 * propagate all the way to the root via zio_notify_parent().
3916 */
3917 ASSERT(vd == NULL && bp != NULL);
3918 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3919
3920 if (IO_IS_ALLOCATING(zio) &&
3921 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3922 if (zio->io_error != ENOSPC)
3923 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3924 else
3925 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3926 }
3927
3928 if ((zio->io_type == ZIO_TYPE_READ ||
3929 zio->io_type == ZIO_TYPE_FREE) &&
3930 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3931 zio->io_error == ENXIO &&
3932 spa_load_state(spa) == SPA_LOAD_NONE &&
3933 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3934 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3935
3936 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3937 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3938
3939 /*
3940 * Here is a possibly good place to attempt to do
3941 * either combinatorial reconstruction or error correction
3942 * based on checksums. It also might be a good place
3943 * to send out preliminary ereports before we suspend
3944 * processing.
3945 */
3946 }
3947
3948 /*
3949 * If there were logical child errors, they apply to us now.
3950 * We defer this until now to avoid conflating logical child
3951 * errors with errors that happened to the zio itself when
3952 * updating vdev stats and reporting FMA events above.
3953 */
3954 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3955
3956 if ((zio->io_error || zio->io_reexecute) &&
3957 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3958 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3959 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3960
3961 zio_gang_tree_free(&zio->io_gang_tree);
3962
3963 /*
3964 * Godfather I/Os should never suspend.
3965 */
3966 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3967 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3968 zio->io_reexecute = 0;
3969
3970 if (zio->io_reexecute) {
3971 /*
3972 * This is a logical I/O that wants to reexecute.
3973 *
3974 * Reexecute is top-down. When an i/o fails, if it's not
3975 * the root, it simply notifies its parent and sticks around.
3976 * The parent, seeing that it still has children in zio_done(),
3977 * does the same. This percolates all the way up to the root.
3978 * The root i/o will reexecute or suspend the entire tree.
3979 *
3980 * This approach ensures that zio_reexecute() honors
3981 * all the original i/o dependency relationships, e.g.
3982 * parents not executing until children are ready.
3983 */
3984 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3985
3986 zio->io_gang_leader = NULL;
3987
3988 mutex_enter(&zio->io_lock);
3989 zio->io_state[ZIO_WAIT_DONE] = 1;
3990 mutex_exit(&zio->io_lock);
3991
3992 /*
3993 * "The Godfather" I/O monitors its children but is
3994 * not a true parent to them. It will track them through
3995 * the pipeline but severs its ties whenever they get into
3996 * trouble (e.g. suspended). This allows "The Godfather"
3997 * I/O to return status without blocking.
3998 */
3999 zl = NULL;
4000 for (pio = zio_walk_parents(zio, &zl); pio != NULL;
4001 pio = pio_next) {
4002 zio_link_t *remove_zl = zl;
4003 pio_next = zio_walk_parents(zio, &zl);
4004
4005 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
4006 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
4007 zio_remove_child(pio, zio, remove_zl);
4008 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4009 }
4010 }
4011
4012 if ((pio = zio_unique_parent(zio)) != NULL) {
4013 /*
4014 * We're not a root i/o, so there's nothing to do
4015 * but notify our parent. Don't propagate errors
4016 * upward since we haven't permanently failed yet.
4017 */
4018 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
4019 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
4020 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4021 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
4022 /*
4023 * We'd fail again if we reexecuted now, so suspend
4024 * until conditions improve (e.g. device comes online).
4025 */
4026 zio_suspend(spa, zio);
4027 } else {
4028 /*
4029 * Reexecution is potentially a huge amount of work.
4030 * Hand it off to the otherwise-unused claim taskq.
4031 */
4032#if defined(illumos) || !defined(_KERNEL)
4033 ASSERT(zio->io_tqent.tqent_next == NULL);
4034#else
4035 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
4036#endif
4037 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
4038 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
4039 0, &zio->io_tqent);
4040 }
4041 return (ZIO_PIPELINE_STOP);
4042 }
4043
4044 ASSERT(zio->io_child_count == 0);
4045 ASSERT(zio->io_reexecute == 0);
4046 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
4047
4048 /*
4049 * Report any checksum errors, since the I/O is complete.
4050 */
4051 while (zio->io_cksum_report != NULL) {
4052 zio_cksum_report_t *zcr = zio->io_cksum_report;
4053 zio->io_cksum_report = zcr->zcr_next;
4054 zcr->zcr_next = NULL;
4055 zcr->zcr_finish(zcr, NULL);
4056 zfs_ereport_free_checksum(zcr);
4057 }
4058
4059 /*
4060 * It is the responsibility of the done callback to ensure that this
4061 * particular zio is no longer discoverable for adoption, and as
4062 * such, cannot acquire any new parents.
4063 */
4064 if (zio->io_done)
4065 zio->io_done(zio);
4066
4067 mutex_enter(&zio->io_lock);
4068 zio->io_state[ZIO_WAIT_DONE] = 1;
4069 mutex_exit(&zio->io_lock);
4070
4071 zl = NULL;
4072 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
4073 zio_link_t *remove_zl = zl;
4074 pio_next = zio_walk_parents(zio, &zl);
4075 zio_remove_child(pio, zio, remove_zl);
4076 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4077 }
4078
4079 if (zio->io_waiter != NULL) {
4080 mutex_enter(&zio->io_lock);
4081 zio->io_executor = NULL;
4082 cv_broadcast(&zio->io_cv);
4083 mutex_exit(&zio->io_lock);
4084 } else {
4085 zio_destroy(zio);
4086 }
4087
4088 return (ZIO_PIPELINE_STOP);
4089}
4090
4091/*
4092 * ==========================================================================
4093 * I/O pipeline definition
4094 * ==========================================================================
4095 */
4096static zio_pipe_stage_t *zio_pipeline[] = {
4097 NULL,
4098 zio_read_bp_init,
4099 zio_write_bp_init,
4100 zio_free_bp_init,
4101 zio_issue_async,
4102 zio_write_compress,
4103 zio_checksum_generate,
4104 zio_nop_write,
4105 zio_ddt_read_start,
4106 zio_ddt_read_done,
4107 zio_ddt_write,
4108 zio_ddt_free,
4109 zio_gang_assemble,
4110 zio_gang_issue,
4111 zio_dva_throttle,
4112 zio_dva_allocate,
4113 zio_dva_free,
4114 zio_dva_claim,
4115 zio_ready,
4116 zio_vdev_io_start,
4117 zio_vdev_io_done,
4118 zio_vdev_io_assess,
4119 zio_checksum_verify,
4120 zio_done
4121};
4122
4123
4124
4125
4126/*
4127 * Compare two zbookmark_phys_t's to see which we would reach first in a
4128 * pre-order traversal of the object tree.
4129 *
4130 * This is simple in every case aside from the meta-dnode object. For all other
4131 * objects, we traverse them in order (object 1 before object 2, and so on).
4132 * However, all of these objects are traversed while traversing object 0, since
4133 * the data it points to is the list of objects. Thus, we need to convert to a
4134 * canonical representation so we can compare meta-dnode bookmarks to
4135 * non-meta-dnode bookmarks.
4136 *
4137 * We do this by calculating "equivalents" for each field of the zbookmark.
4138 * zbookmarks outside of the meta-dnode use their own object and level, and
4139 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4140 * blocks this bookmark refers to) by multiplying their blkid by their span
4141 * (the number of L0 blocks contained within one block at their level).
4142 * zbookmarks inside the meta-dnode calculate their object equivalent
4143 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4144 * level + 1<<31 (any value larger than a level could ever be) for their level.
4145 * This causes them to always compare before a bookmark in their object
4146 * equivalent, compare appropriately to bookmarks in other objects, and to
4147 * compare appropriately to other bookmarks in the meta-dnode.
4148 */
4149int
4150zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
4151 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
4152{
4153 /*
4154 * These variables represent the "equivalent" values for the zbookmark,
4155 * after converting zbookmarks inside the meta dnode to their
4156 * normal-object equivalents.
4157 */
4158 uint64_t zb1obj, zb2obj;
4159 uint64_t zb1L0, zb2L0;
4160 uint64_t zb1level, zb2level;
4161
4162 if (zb1->zb_object == zb2->zb_object &&
4163 zb1->zb_level == zb2->zb_level &&
4164 zb1->zb_blkid == zb2->zb_blkid)
4165 return (0);
4166
4167 /*
4168 * BP_SPANB calculates the span in blocks.
4169 */
4170 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
4171 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
4172
4173 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
4174 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4175 zb1L0 = 0;
4176 zb1level = zb1->zb_level + COMPARE_META_LEVEL;
4177 } else {
4178 zb1obj = zb1->zb_object;
4179 zb1level = zb1->zb_level;
4180 }
4181
4182 if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
4183 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4184 zb2L0 = 0;
4185 zb2level = zb2->zb_level + COMPARE_META_LEVEL;
4186 } else {
4187 zb2obj = zb2->zb_object;
4188 zb2level = zb2->zb_level;
4189 }
4190
4191 /* Now that we have a canonical representation, do the comparison. */
4192 if (zb1obj != zb2obj)
4193 return (zb1obj < zb2obj ? -1 : 1);
4194 else if (zb1L0 != zb2L0)
4195 return (zb1L0 < zb2L0 ? -1 : 1);
4196 else if (zb1level != zb2level)
4197 return (zb1level > zb2level ? -1 : 1);
4198 /*
4199 * This can (theoretically) happen if the bookmarks have the same object
4200 * and level, but different blkids, if the block sizes are not the same.
4201 * There is presently no way to change the indirect block sizes
4202 */
4203 return (0);
4204}
4205
4206/*
4207 * This function checks the following: given that last_block is the place that
4208 * our traversal stopped last time, does that guarantee that we've visited
4209 * every node under subtree_root? Therefore, we can't just use the raw output
4210 * of zbookmark_compare. We have to pass in a modified version of
4211 * subtree_root; by incrementing the block id, and then checking whether
4212 * last_block is before or equal to that, we can tell whether or not having
4213 * visited last_block implies that all of subtree_root's children have been
4214 * visited.
4215 */
4216boolean_t
4217zbookmark_subtree_completed(const dnode_phys_t *dnp,
4218 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
4219{
4220 zbookmark_phys_t mod_zb = *subtree_root;
4221 mod_zb.zb_blkid++;
4222 ASSERT(last_block->zb_level == 0);
4223
4224 /* The objset_phys_t isn't before anything. */
4225 if (dnp == NULL)
4226 return (B_FALSE);
4227
4228 /*
4229 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4230 * data block size in sectors, because that variable is only used if
4231 * the bookmark refers to a block in the meta-dnode. Since we don't
4232 * know without examining it what object it refers to, and there's no
4233 * harm in passing in this value in other cases, we always pass it in.
4234 *
4235 * We pass in 0 for the indirect block size shift because zb2 must be
4236 * level 0. The indirect block size is only used to calculate the span
4237 * of the bookmark, but since the bookmark must be level 0, the span is
4238 * always 1, so the math works out.
4239 *
4240 * If you make changes to how the zbookmark_compare code works, be sure
4241 * to make sure that this code still works afterwards.
4242 */
4243 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
4244 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,
4245 last_block) <= 0);
4246}
|