author | brendan |
Fri, 09 Nov 2007 21:33:30 -0800 | |
changeset 5450 | b25030891c44 |
parent 5403 | 0bfd0977c989 |
child 5530 | 4ed96167d864 |
permissions | -rw-r--r-- |
789 | 1 |
/* |
2 |
* CDDL HEADER START |
|
3 |
* |
|
4 |
* The contents of this file are subject to the terms of the |
|
1544 | 5 |
* Common Development and Distribution License (the "License"). |
6 |
* You may not use this file except in compliance with the License. |
|
789 | 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 |
/* |
|
3459
358b6bdb15b6
6512311 zio_done() is chunky and needs to go on a diet
ek110237
parents:
3290
diff
changeset
|
22 |
* Copyright 2007 Sun Microsystems, Inc. All rights reserved. |
789 | 23 |
* Use is subject to license terms. |
24 |
*/ |
|
25 |
||
26 |
#pragma ident "%Z%%M% %I% %E% SMI" |
|
27 |
||
28 |
#include <sys/zfs_context.h> |
|
1544 | 29 |
#include <sys/fm/fs/zfs.h> |
789 | 30 |
#include <sys/spa.h> |
31 |
#include <sys/txg.h> |
|
32 |
#include <sys/spa_impl.h> |
|
33 |
#include <sys/vdev_impl.h> |
|
34 |
#include <sys/zio_impl.h> |
|
35 |
#include <sys/zio_compress.h> |
|
36 |
#include <sys/zio_checksum.h> |
|
37 |
||
38 |
/* |
|
39 |
* ========================================================================== |
|
40 |
* I/O priority table |
|
41 |
* ========================================================================== |
|
42 |
*/ |
|
43 |
uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = { |
|
44 |
0, /* ZIO_PRIORITY_NOW */ |
|
45 |
0, /* ZIO_PRIORITY_SYNC_READ */ |
|
46 |
0, /* ZIO_PRIORITY_SYNC_WRITE */ |
|
47 |
6, /* ZIO_PRIORITY_ASYNC_READ */ |
|
48 |
4, /* ZIO_PRIORITY_ASYNC_WRITE */ |
|
49 |
4, /* ZIO_PRIORITY_FREE */ |
|
50 |
0, /* ZIO_PRIORITY_CACHE_FILL */ |
|
51 |
0, /* ZIO_PRIORITY_LOG_WRITE */ |
|
52 |
10, /* ZIO_PRIORITY_RESILVER */ |
|
53 |
20, /* ZIO_PRIORITY_SCRUB */ |
|
54 |
}; |
|
55 |
||
56 |
/* |
|
57 |
* ========================================================================== |
|
58 |
* I/O type descriptions |
|
59 |
* ========================================================================== |
|
60 |
*/ |
|
61 |
char *zio_type_name[ZIO_TYPES] = { |
|
62 |
"null", "read", "write", "free", "claim", "ioctl" }; |
|
63 |
||
64 |
/* At or above this size, force gang blocking - for testing */ |
|
65 |
uint64_t zio_gang_bang = SPA_MAXBLOCKSIZE + 1; |
|
66 |
||
3668 | 67 |
/* Force an allocation failure when non-zero */ |
68 |
uint16_t zio_zil_fail_shift = 0; |
|
5329 | 69 |
uint16_t zio_io_fail_shift = 0; |
70 |
||
71 |
/* Enable/disable the write-retry logic */ |
|
72 |
int zio_write_retry = 1; |
|
73 |
||
74 |
/* Taskq to handle reissuing of I/Os */ |
|
75 |
taskq_t *zio_taskq; |
|
76 |
int zio_resume_threads = 4; |
|
3668 | 77 |
|
789 | 78 |
typedef struct zio_sync_pass { |
79 |
int zp_defer_free; /* defer frees after this pass */ |
|
80 |
int zp_dontcompress; /* don't compress after this pass */ |
|
81 |
int zp_rewrite; /* rewrite new bps after this pass */ |
|
82 |
} zio_sync_pass_t; |
|
83 |
||
84 |
zio_sync_pass_t zio_sync_pass = { |
|
85 |
1, /* zp_defer_free */ |
|
86 |
4, /* zp_dontcompress */ |
|
87 |
1, /* zp_rewrite */ |
|
88 |
}; |
|
89 |
||
5329 | 90 |
static boolean_t zio_io_should_fail(uint16_t); |
91 |
||
789 | 92 |
/* |
93 |
* ========================================================================== |
|
94 |
* I/O kmem caches |
|
95 |
* ========================================================================== |
|
96 |
*/ |
|
4055 | 97 |
kmem_cache_t *zio_cache; |
789 | 98 |
kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; |
3290 | 99 |
kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; |
100 |
||
101 |
#ifdef _KERNEL |
|
102 |
extern vmem_t *zio_alloc_arena; |
|
103 |
#endif |
|
789 | 104 |
|
5329 | 105 |
/* |
106 |
* Determine if we are allowed to issue the IO based on the |
|
107 |
* pool state. If we must wait then block until we are told |
|
108 |
* that we may continue. |
|
109 |
*/ |
|
110 |
#define ZIO_ENTER(spa) { \ |
|
111 |
if (spa->spa_state == POOL_STATE_IO_FAILURE) { \ |
|
112 |
mutex_enter(&spa->spa_zio_lock); \ |
|
113 |
while (spa->spa_state == POOL_STATE_IO_FAILURE) \ |
|
114 |
cv_wait(&spa->spa_zio_cv, &spa->spa_zio_lock); \ |
|
115 |
mutex_exit(&spa->spa_zio_lock); \ |
|
116 |
} \ |
|
117 |
} |
|
118 |
||
119 |
/* |
|
120 |
* An allocation zio is one that either currently has the DVA allocate |
|
121 |
* stage set or will have it later in it's lifetime. |
|
122 |
*/ |
|
123 |
#define IO_IS_ALLOCATING(zio) \ |
|
124 |
((zio)->io_orig_pipeline == ZIO_WRITE_PIPELINE || \ |
|
125 |
(zio)->io_pipeline & (1U << ZIO_STAGE_DVA_ALLOCATE)) |
|
126 |
||
789 | 127 |
void |
128 |
zio_init(void) |
|
129 |
{ |
|
130 |
size_t c; |
|
3290 | 131 |
vmem_t *data_alloc_arena = NULL; |
132 |
||
133 |
#ifdef _KERNEL |
|
134 |
data_alloc_arena = zio_alloc_arena; |
|
135 |
#endif |
|
789 | 136 |
|
4055 | 137 |
zio_cache = kmem_cache_create("zio_cache", sizeof (zio_t), 0, |
138 |
NULL, NULL, NULL, NULL, NULL, 0); |
|
139 |
||
789 | 140 |
/* |
141 |
* For small buffers, we want a cache for each multiple of |
|
142 |
* SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache |
|
143 |
* for each quarter-power of 2. For large buffers, we want |
|
144 |
* a cache for each multiple of PAGESIZE. |
|
145 |
*/ |
|
146 |
for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { |
|
147 |
size_t size = (c + 1) << SPA_MINBLOCKSHIFT; |
|
148 |
size_t p2 = size; |
|
149 |
size_t align = 0; |
|
150 |
||
151 |
while (p2 & (p2 - 1)) |
|
152 |
p2 &= p2 - 1; |
|
153 |
||
154 |
if (size <= 4 * SPA_MINBLOCKSIZE) { |
|
155 |
align = SPA_MINBLOCKSIZE; |
|
156 |
} else if (P2PHASE(size, PAGESIZE) == 0) { |
|
157 |
align = PAGESIZE; |
|
158 |
} else if (P2PHASE(size, p2 >> 2) == 0) { |
|
159 |
align = p2 >> 2; |
|
160 |
} |
|
161 |
||
162 |
if (align != 0) { |
|
3290 | 163 |
char name[36]; |
2856 | 164 |
(void) sprintf(name, "zio_buf_%lu", (ulong_t)size); |
789 | 165 |
zio_buf_cache[c] = kmem_cache_create(name, size, |
849
8d799fd81a9b
6345023 /dev/zfs fails to open once ZFS module is unloaded
bonwick
parents:
789
diff
changeset
|
166 |
align, NULL, NULL, NULL, NULL, NULL, KMC_NODEBUG); |
3290 | 167 |
|
168 |
(void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size); |
|
169 |
zio_data_buf_cache[c] = kmem_cache_create(name, size, |
|
170 |
align, NULL, NULL, NULL, NULL, data_alloc_arena, |
|
171 |
KMC_NODEBUG); |
|
172 |
||
789 | 173 |
dprintf("creating cache for size %5lx align %5lx\n", |
174 |
size, align); |
|
175 |
} |
|
176 |
} |
|
177 |
||
178 |
while (--c != 0) { |
|
179 |
ASSERT(zio_buf_cache[c] != NULL); |
|
180 |
if (zio_buf_cache[c - 1] == NULL) |
|
181 |
zio_buf_cache[c - 1] = zio_buf_cache[c]; |
|
3290 | 182 |
|
183 |
ASSERT(zio_data_buf_cache[c] != NULL); |
|
184 |
if (zio_data_buf_cache[c - 1] == NULL) |
|
185 |
zio_data_buf_cache[c - 1] = zio_data_buf_cache[c]; |
|
789 | 186 |
} |
1544 | 187 |
|
5329 | 188 |
zio_taskq = taskq_create("zio_taskq", zio_resume_threads, |
189 |
maxclsyspri, 50, INT_MAX, TASKQ_PREPOPULATE); |
|
190 |
||
1544 | 191 |
zio_inject_init(); |
789 | 192 |
} |
193 |
||
194 |
void |
|
195 |
zio_fini(void) |
|
196 |
{ |
|
197 |
size_t c; |
|
198 |
kmem_cache_t *last_cache = NULL; |
|
3290 | 199 |
kmem_cache_t *last_data_cache = NULL; |
789 | 200 |
|
201 |
for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { |
|
202 |
if (zio_buf_cache[c] != last_cache) { |
|
203 |
last_cache = zio_buf_cache[c]; |
|
204 |
kmem_cache_destroy(zio_buf_cache[c]); |
|
205 |
} |
|
206 |
zio_buf_cache[c] = NULL; |
|
3290 | 207 |
|
208 |
if (zio_data_buf_cache[c] != last_data_cache) { |
|
209 |
last_data_cache = zio_data_buf_cache[c]; |
|
210 |
kmem_cache_destroy(zio_data_buf_cache[c]); |
|
211 |
} |
|
212 |
zio_data_buf_cache[c] = NULL; |
|
789 | 213 |
} |
1544 | 214 |
|
5329 | 215 |
taskq_destroy(zio_taskq); |
216 |
||
4055 | 217 |
kmem_cache_destroy(zio_cache); |
218 |
||
1544 | 219 |
zio_inject_fini(); |
789 | 220 |
} |
221 |
||
222 |
/* |
|
223 |
* ========================================================================== |
|
224 |
* Allocate and free I/O buffers |
|
225 |
* ========================================================================== |
|
226 |
*/ |
|
3290 | 227 |
|
228 |
/* |
|
229 |
* Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a |
|
230 |
* crashdump if the kernel panics, so use it judiciously. Obviously, it's |
|
231 |
* useful to inspect ZFS metadata, but if possible, we should avoid keeping |
|
232 |
* excess / transient data in-core during a crashdump. |
|
233 |
*/ |
|
789 | 234 |
void * |
235 |
zio_buf_alloc(size_t size) |
|
236 |
{ |
|
237 |
size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; |
|
238 |
||
239 |
ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
|
240 |
||
241 |
return (kmem_cache_alloc(zio_buf_cache[c], KM_SLEEP)); |
|
242 |
} |
|
243 |
||
3290 | 244 |
/* |
245 |
* Use zio_data_buf_alloc to allocate data. The data will not appear in a |
|
246 |
* crashdump if the kernel panics. This exists so that we will limit the amount |
|
247 |
* of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount |
|
248 |
* of kernel heap dumped to disk when the kernel panics) |
|
249 |
*/ |
|
250 |
void * |
|
251 |
zio_data_buf_alloc(size_t size) |
|
252 |
{ |
|
253 |
size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; |
|
254 |
||
255 |
ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
|
256 |
||
257 |
return (kmem_cache_alloc(zio_data_buf_cache[c], KM_SLEEP)); |
|
258 |
} |
|
259 |
||
789 | 260 |
void |
261 |
zio_buf_free(void *buf, size_t size) |
|
262 |
{ |
|
263 |
size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; |
|
264 |
||
265 |
ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
|
266 |
||
267 |
kmem_cache_free(zio_buf_cache[c], buf); |
|
268 |
} |
|
269 |
||
3290 | 270 |
void |
271 |
zio_data_buf_free(void *buf, size_t size) |
|
272 |
{ |
|
273 |
size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; |
|
274 |
||
275 |
ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
|
276 |
||
277 |
kmem_cache_free(zio_data_buf_cache[c], buf); |
|
278 |
} |
|
3463 | 279 |
|
789 | 280 |
/* |
281 |
* ========================================================================== |
|
282 |
* Push and pop I/O transform buffers |
|
283 |
* ========================================================================== |
|
284 |
*/ |
|
285 |
static void |
|
286 |
zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize) |
|
287 |
{ |
|
288 |
zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP); |
|
289 |
||
290 |
zt->zt_data = data; |
|
291 |
zt->zt_size = size; |
|
292 |
zt->zt_bufsize = bufsize; |
|
293 |
||
294 |
zt->zt_next = zio->io_transform_stack; |
|
295 |
zio->io_transform_stack = zt; |
|
296 |
||
297 |
zio->io_data = data; |
|
298 |
zio->io_size = size; |
|
299 |
} |
|
300 |
||
301 |
static void |
|
302 |
zio_pop_transform(zio_t *zio, void **data, uint64_t *size, uint64_t *bufsize) |
|
303 |
{ |
|
304 |
zio_transform_t *zt = zio->io_transform_stack; |
|
305 |
||
306 |
*data = zt->zt_data; |
|
307 |
*size = zt->zt_size; |
|
308 |
*bufsize = zt->zt_bufsize; |
|
309 |
||
310 |
zio->io_transform_stack = zt->zt_next; |
|
311 |
kmem_free(zt, sizeof (zio_transform_t)); |
|
312 |
||
313 |
if ((zt = zio->io_transform_stack) != NULL) { |
|
314 |
zio->io_data = zt->zt_data; |
|
315 |
zio->io_size = zt->zt_size; |
|
316 |
} |
|
317 |
} |
|
318 |
||
319 |
static void |
|
320 |
zio_clear_transform_stack(zio_t *zio) |
|
321 |
{ |
|
322 |
void *data; |
|
323 |
uint64_t size, bufsize; |
|
324 |
||
325 |
ASSERT(zio->io_transform_stack != NULL); |
|
326 |
||
327 |
zio_pop_transform(zio, &data, &size, &bufsize); |
|
328 |
while (zio->io_transform_stack != NULL) { |
|
329 |
zio_buf_free(data, bufsize); |
|
330 |
zio_pop_transform(zio, &data, &size, &bufsize); |
|
331 |
} |
|
332 |
} |
|
333 |
||
334 |
/* |
|
335 |
* ========================================================================== |
|
336 |
* Create the various types of I/O (read, write, free) |
|
337 |
* ========================================================================== |
|
338 |
*/ |
|
339 |
static zio_t * |
|
340 |
zio_create(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, |
|
341 |
void *data, uint64_t size, zio_done_func_t *done, void *private, |
|
342 |
zio_type_t type, int priority, int flags, uint8_t stage, uint32_t pipeline) |
|
343 |
{ |
|
344 |
zio_t *zio; |
|
345 |
||
346 |
ASSERT3U(size, <=, SPA_MAXBLOCKSIZE); |
|
347 |
ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0); |
|
348 |
||
4055 | 349 |
zio = kmem_cache_alloc(zio_cache, KM_SLEEP); |
350 |
bzero(zio, sizeof (zio_t)); |
|
789 | 351 |
zio->io_parent = pio; |
352 |
zio->io_spa = spa; |
|
353 |
zio->io_txg = txg; |
|
4634
39bfb9e90d34
6437054 vdev_cache wises up: increase DB performance by 16%
ek110237
parents:
4527
diff
changeset
|
354 |
zio->io_flags = flags; |
789 | 355 |
if (bp != NULL) { |
356 |
zio->io_bp = bp; |
|
357 |
zio->io_bp_copy = *bp; |
|
358 |
zio->io_bp_orig = *bp; |
|
4634
39bfb9e90d34
6437054 vdev_cache wises up: increase DB performance by 16%
ek110237
parents:
4527
diff
changeset
|
359 |
if (dmu_ot[BP_GET_TYPE(bp)].ot_metadata || |
39bfb9e90d34
6437054 vdev_cache wises up: increase DB performance by 16%
ek110237
parents:
4527
diff
changeset
|
360 |
BP_GET_LEVEL(bp) != 0) |
39bfb9e90d34
6437054 vdev_cache wises up: increase DB performance by 16%
ek110237
parents:
4527
diff
changeset
|
361 |
zio->io_flags |= ZIO_FLAG_METADATA; |
789 | 362 |
} |
363 |
zio->io_done = done; |
|
364 |
zio->io_private = private; |
|
365 |
zio->io_type = type; |
|
366 |
zio->io_priority = priority; |
|
367 |
zio->io_stage = stage; |
|
368 |
zio->io_pipeline = pipeline; |
|
369 |
zio->io_async_stages = ZIO_ASYNC_PIPELINE_STAGES; |
|
370 |
zio->io_timestamp = lbolt64; |
|
4634
39bfb9e90d34
6437054 vdev_cache wises up: increase DB performance by 16%
ek110237
parents:
4527
diff
changeset
|
371 |
if (pio != NULL) |
39bfb9e90d34
6437054 vdev_cache wises up: increase DB performance by 16%
ek110237
parents:
4527
diff
changeset
|
372 |
zio->io_flags |= (pio->io_flags & ZIO_FLAG_METADATA); |
2856 | 373 |
mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL); |
4831
41ec732c6d9f
6584470 zdb needs to initialize the bpl_lock mutex
gw25295
parents:
4634
diff
changeset
|
374 |
cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL); |
789 | 375 |
zio_push_transform(zio, data, size, size); |
376 |
||
3463 | 377 |
/* |
378 |
* Note on config lock: |
|
379 |
* |
|
380 |
* If CONFIG_HELD is set, then the caller already has the config |
|
381 |
* lock, so we don't need it for this io. |
|
382 |
* |
|
383 |
* We set CONFIG_GRABBED to indicate that we have grabbed the |
|
384 |
* config lock on behalf of this io, so it should be released |
|
385 |
* in zio_done. |
|
386 |
* |
|
387 |
* Unless CONFIG_HELD is set, we will grab the config lock for |
|
388 |
* any top-level (parent-less) io, *except* NULL top-level ios. |
|
389 |
* The NULL top-level ios rarely have any children, so we delay |
|
390 |
* grabbing the lock until the first child is added (but it is |
|
391 |
* still grabbed on behalf of the top-level i/o, so additional |
|
392 |
* children don't need to also grab it). This greatly reduces |
|
393 |
* contention on the config lock. |
|
394 |
*/ |
|
789 | 395 |
if (pio == NULL) { |
3463 | 396 |
if (type != ZIO_TYPE_NULL && |
397 |
!(flags & ZIO_FLAG_CONFIG_HELD)) { |
|
1544 | 398 |
spa_config_enter(zio->io_spa, RW_READER, zio); |
3463 | 399 |
zio->io_flags |= ZIO_FLAG_CONFIG_GRABBED; |
400 |
} |
|
789 | 401 |
zio->io_root = zio; |
402 |
} else { |
|
403 |
zio->io_root = pio->io_root; |
|
1544 | 404 |
if (!(flags & ZIO_FLAG_NOBOOKMARK)) |
405 |
zio->io_logical = pio->io_logical; |
|
789 | 406 |
mutex_enter(&pio->io_lock); |
3463 | 407 |
if (pio->io_parent == NULL && |
408 |
pio->io_type == ZIO_TYPE_NULL && |
|
409 |
!(pio->io_flags & ZIO_FLAG_CONFIG_GRABBED) && |
|
410 |
!(pio->io_flags & ZIO_FLAG_CONFIG_HELD)) { |
|
411 |
pio->io_flags |= ZIO_FLAG_CONFIG_GRABBED; |
|
412 |
spa_config_enter(zio->io_spa, RW_READER, pio); |
|
413 |
} |
|
789 | 414 |
if (stage < ZIO_STAGE_READY) |
415 |
pio->io_children_notready++; |
|
416 |
pio->io_children_notdone++; |
|
417 |
zio->io_sibling_next = pio->io_child; |
|
418 |
zio->io_sibling_prev = NULL; |
|
419 |
if (pio->io_child != NULL) |
|
420 |
pio->io_child->io_sibling_prev = zio; |
|
421 |
pio->io_child = zio; |
|
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
422 |
zio->io_ndvas = pio->io_ndvas; |
789 | 423 |
mutex_exit(&pio->io_lock); |
424 |
} |
|
425 |
||
5329 | 426 |
/* |
427 |
* Save off the original state incase we need to retry later. |
|
428 |
*/ |
|
429 |
zio->io_orig_stage = zio->io_stage; |
|
430 |
zio->io_orig_pipeline = zio->io_pipeline; |
|
431 |
zio->io_orig_flags = zio->io_flags; |
|
432 |
||
789 | 433 |
return (zio); |
434 |
} |
|
435 |
||
5329 | 436 |
static void |
437 |
zio_reset(zio_t *zio) |
|
438 |
{ |
|
439 |
zio_clear_transform_stack(zio); |
|
440 |
||
441 |
zio->io_flags = zio->io_orig_flags; |
|
442 |
zio->io_stage = zio->io_orig_stage; |
|
443 |
zio->io_pipeline = zio->io_orig_pipeline; |
|
444 |
zio_push_transform(zio, zio->io_data, zio->io_size, zio->io_size); |
|
445 |
} |
|
446 |
||
789 | 447 |
zio_t * |
448 |
zio_null(zio_t *pio, spa_t *spa, zio_done_func_t *done, void *private, |
|
449 |
int flags) |
|
450 |
{ |
|
451 |
zio_t *zio; |
|
452 |
||
453 |
zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private, |
|
454 |
ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, ZIO_STAGE_OPEN, |
|
455 |
ZIO_WAIT_FOR_CHILDREN_PIPELINE); |
|
456 |
||
457 |
return (zio); |
|
458 |
} |
|
459 |
||
460 |
zio_t * |
|
461 |
zio_root(spa_t *spa, zio_done_func_t *done, void *private, int flags) |
|
462 |
{ |
|
463 |
return (zio_null(NULL, spa, done, private, flags)); |
|
464 |
} |
|
465 |
||
466 |
zio_t * |
|
467 |
zio_read(zio_t *pio, spa_t *spa, blkptr_t *bp, void *data, |
|
468 |
uint64_t size, zio_done_func_t *done, void *private, |
|
1544 | 469 |
int priority, int flags, zbookmark_t *zb) |
789 | 470 |
{ |
471 |
zio_t *zio; |
|
472 |
||
473 |
ASSERT3U(size, ==, BP_GET_LSIZE(bp)); |
|
474 |
||
5329 | 475 |
/* |
476 |
* If the user has specified that we allow I/Os to continue |
|
477 |
* then attempt to satisfy the read. |
|
478 |
*/ |
|
479 |
if (spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE) |
|
480 |
ZIO_ENTER(spa); |
|
481 |
||
789 | 482 |
zio = zio_create(pio, spa, bp->blk_birth, bp, data, size, done, private, |
2981 | 483 |
ZIO_TYPE_READ, priority, flags | ZIO_FLAG_USER, |
484 |
ZIO_STAGE_OPEN, ZIO_READ_PIPELINE); |
|
1544 | 485 |
zio->io_bookmark = *zb; |
486 |
||
487 |
zio->io_logical = zio; |
|
789 | 488 |
|
489 |
/* |
|
490 |
* Work off our copy of the bp so the caller can free it. |
|
491 |
*/ |
|
492 |
zio->io_bp = &zio->io_bp_copy; |
|
493 |
||
494 |
return (zio); |
|
495 |
} |
|
496 |
||
497 |
zio_t * |
|
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
498 |
zio_write(zio_t *pio, spa_t *spa, int checksum, int compress, int ncopies, |
789 | 499 |
uint64_t txg, blkptr_t *bp, void *data, uint64_t size, |
3547
e396e0a440b1
6512391 DMU should leverage ZIO dependencies to achieve greater parallelism
maybee
parents:
3463
diff
changeset
|
500 |
zio_done_func_t *ready, zio_done_func_t *done, void *private, int priority, |
e396e0a440b1
6512391 DMU should leverage ZIO dependencies to achieve greater parallelism
maybee
parents:
3463
diff
changeset
|
501 |
int flags, zbookmark_t *zb) |
789 | 502 |
{ |
503 |
zio_t *zio; |
|
504 |
||
505 |
ASSERT(checksum >= ZIO_CHECKSUM_OFF && |
|
506 |
checksum < ZIO_CHECKSUM_FUNCTIONS); |
|
507 |
||
508 |
ASSERT(compress >= ZIO_COMPRESS_OFF && |
|
509 |
compress < ZIO_COMPRESS_FUNCTIONS); |
|
510 |
||
5329 | 511 |
ZIO_ENTER(spa); |
512 |
||
789 | 513 |
zio = zio_create(pio, spa, txg, bp, data, size, done, private, |
2981 | 514 |
ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_USER, |
789 | 515 |
ZIO_STAGE_OPEN, ZIO_WRITE_PIPELINE); |
516 |
||
3547
e396e0a440b1
6512391 DMU should leverage ZIO dependencies to achieve greater parallelism
maybee
parents:
3463
diff
changeset
|
517 |
zio->io_ready = ready; |
e396e0a440b1
6512391 DMU should leverage ZIO dependencies to achieve greater parallelism
maybee
parents:
3463
diff
changeset
|
518 |
|
1544 | 519 |
zio->io_bookmark = *zb; |
520 |
||
521 |
zio->io_logical = zio; |
|
522 |
||
789 | 523 |
zio->io_checksum = checksum; |
524 |
zio->io_compress = compress; |
|
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
525 |
zio->io_ndvas = ncopies; |
789 | 526 |
|
527 |
if (compress != ZIO_COMPRESS_OFF) |
|
528 |
zio->io_async_stages |= 1U << ZIO_STAGE_WRITE_COMPRESS; |
|
529 |
||
530 |
if (bp->blk_birth != txg) { |
|
531 |
/* XXX the bp usually (always?) gets re-zeroed later */ |
|
532 |
BP_ZERO(bp); |
|
533 |
BP_SET_LSIZE(bp, size); |
|
534 |
BP_SET_PSIZE(bp, size); |
|
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
535 |
} else { |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
536 |
/* Make sure someone doesn't change their mind on overwrites */ |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
537 |
ASSERT(MIN(zio->io_ndvas + BP_IS_GANG(bp), |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
538 |
spa_max_replication(spa)) == BP_GET_NDVAS(bp)); |
789 | 539 |
} |
540 |
||
541 |
return (zio); |
|
542 |
} |
|
543 |
||
544 |
zio_t * |
|
545 |
zio_rewrite(zio_t *pio, spa_t *spa, int checksum, |
|
546 |
uint64_t txg, blkptr_t *bp, void *data, uint64_t size, |
|
1544 | 547 |
zio_done_func_t *done, void *private, int priority, int flags, |
548 |
zbookmark_t *zb) |
|
789 | 549 |
{ |
550 |
zio_t *zio; |
|
551 |
||
552 |
zio = zio_create(pio, spa, txg, bp, data, size, done, private, |
|
2981 | 553 |
ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_USER, |
789 | 554 |
ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE); |
555 |
||
1544 | 556 |
zio->io_bookmark = *zb; |
789 | 557 |
zio->io_checksum = checksum; |
558 |
zio->io_compress = ZIO_COMPRESS_OFF; |
|
559 |
||
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
560 |
if (pio != NULL) |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
561 |
ASSERT3U(zio->io_ndvas, <=, BP_GET_NDVAS(bp)); |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
562 |
|
789 | 563 |
return (zio); |
564 |
} |
|
565 |
||
5329 | 566 |
static void |
567 |
zio_write_allocate_ready(zio_t *zio) |
|
568 |
{ |
|
569 |
/* Free up the previous block */ |
|
570 |
if (!BP_IS_HOLE(&zio->io_bp_orig)) { |
|
571 |
zio_nowait(zio_free(zio, zio->io_spa, zio->io_txg, |
|
572 |
&zio->io_bp_orig, NULL, NULL)); |
|
573 |
} |
|
574 |
} |
|
575 |
||
789 | 576 |
static zio_t * |
577 |
zio_write_allocate(zio_t *pio, spa_t *spa, int checksum, |
|
578 |
uint64_t txg, blkptr_t *bp, void *data, uint64_t size, |
|
579 |
zio_done_func_t *done, void *private, int priority, int flags) |
|
580 |
{ |
|
581 |
zio_t *zio; |
|
582 |
||
583 |
BP_ZERO(bp); |
|
584 |
BP_SET_LSIZE(bp, size); |
|
585 |
BP_SET_PSIZE(bp, size); |
|
586 |
BP_SET_COMPRESS(bp, ZIO_COMPRESS_OFF); |
|
587 |
||
588 |
zio = zio_create(pio, spa, txg, bp, data, size, done, private, |
|
589 |
ZIO_TYPE_WRITE, priority, flags, |
|
590 |
ZIO_STAGE_OPEN, ZIO_WRITE_ALLOCATE_PIPELINE); |
|
591 |
||
592 |
zio->io_checksum = checksum; |
|
593 |
zio->io_compress = ZIO_COMPRESS_OFF; |
|
5329 | 594 |
zio->io_ready = zio_write_allocate_ready; |
789 | 595 |
|
596 |
return (zio); |
|
597 |
} |
|
598 |
||
599 |
zio_t * |
|
600 |
zio_free(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, |
|
601 |
zio_done_func_t *done, void *private) |
|
602 |
{ |
|
603 |
zio_t *zio; |
|
604 |
||
605 |
ASSERT(!BP_IS_HOLE(bp)); |
|
606 |
||
607 |
if (txg == spa->spa_syncing_txg && |
|
608 |
spa->spa_sync_pass > zio_sync_pass.zp_defer_free) { |
|
609 |
bplist_enqueue_deferred(&spa->spa_sync_bplist, bp); |
|
610 |
return (zio_null(pio, spa, NULL, NULL, 0)); |
|
611 |
} |
|
612 |
||
613 |
zio = zio_create(pio, spa, txg, bp, NULL, 0, done, private, |
|
2981 | 614 |
ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, ZIO_FLAG_USER, |
789 | 615 |
ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE); |
616 |
||
617 |
zio->io_bp = &zio->io_bp_copy; |
|
618 |
||
619 |
return (zio); |
|
620 |
} |
|
621 |
||
622 |
zio_t * |
|
623 |
zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, |
|
624 |
zio_done_func_t *done, void *private) |
|
625 |
{ |
|
626 |
zio_t *zio; |
|
627 |
||
628 |
/* |
|
629 |
* A claim is an allocation of a specific block. Claims are needed |
|
630 |
* to support immediate writes in the intent log. The issue is that |
|
631 |
* immediate writes contain committed data, but in a txg that was |
|
632 |
* *not* committed. Upon opening the pool after an unclean shutdown, |
|
633 |
* the intent log claims all blocks that contain immediate write data |
|
634 |
* so that the SPA knows they're in use. |
|
635 |
* |
|
636 |
* All claims *must* be resolved in the first txg -- before the SPA |
|
637 |
* starts allocating blocks -- so that nothing is allocated twice. |
|
638 |
*/ |
|
639 |
ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa)); |
|
640 |
ASSERT3U(spa_first_txg(spa), <=, txg); |
|
641 |
||
642 |
zio = zio_create(pio, spa, txg, bp, NULL, 0, done, private, |
|
643 |
ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, 0, |
|
644 |
ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE); |
|
645 |
||
646 |
zio->io_bp = &zio->io_bp_copy; |
|
647 |
||
648 |
return (zio); |
|
649 |
} |
|
650 |
||
651 |
zio_t * |
|
652 |
zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, |
|
653 |
zio_done_func_t *done, void *private, int priority, int flags) |
|
654 |
{ |
|
655 |
zio_t *zio; |
|
656 |
int c; |
|
657 |
||
658 |
if (vd->vdev_children == 0) { |
|
659 |
zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private, |
|
660 |
ZIO_TYPE_IOCTL, priority, flags, |
|
661 |
ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE); |
|
662 |
||
663 |
zio->io_vd = vd; |
|
664 |
zio->io_cmd = cmd; |
|
665 |
} else { |
|
666 |
zio = zio_null(pio, spa, NULL, NULL, flags); |
|
667 |
||
668 |
for (c = 0; c < vd->vdev_children; c++) |
|
669 |
zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd, |
|
670 |
done, private, priority, flags)); |
|
671 |
} |
|
672 |
||
673 |
return (zio); |
|
674 |
} |
|
675 |
||
676 |
static void |
|
677 |
zio_phys_bp_init(vdev_t *vd, blkptr_t *bp, uint64_t offset, uint64_t size, |
|
5450 | 678 |
int checksum, boolean_t labels) |
789 | 679 |
{ |
680 |
ASSERT(vd->vdev_children == 0); |
|
681 |
||
682 |
ASSERT(size <= SPA_MAXBLOCKSIZE); |
|
683 |
ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0); |
|
684 |
ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); |
|
685 |
||
5450 | 686 |
#ifdef ZFS_DEBUG |
687 |
if (labels) { |
|
688 |
ASSERT(offset + size <= VDEV_LABEL_START_SIZE || |
|
689 |
offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); |
|
690 |
} |
|
691 |
#endif |
|
789 | 692 |
ASSERT3U(offset + size, <=, vd->vdev_psize); |
693 |
||
694 |
BP_ZERO(bp); |
|
695 |
||
696 |
BP_SET_LSIZE(bp, size); |
|
697 |
BP_SET_PSIZE(bp, size); |
|
698 |
||
699 |
BP_SET_CHECKSUM(bp, checksum); |
|
700 |
BP_SET_COMPRESS(bp, ZIO_COMPRESS_OFF); |
|
701 |
BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); |
|
702 |
||
703 |
if (checksum != ZIO_CHECKSUM_OFF) |
|
704 |
ZIO_SET_CHECKSUM(&bp->blk_cksum, offset, 0, 0, 0); |
|
705 |
} |
|
706 |
||
707 |
zio_t * |
|
708 |
zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, |
|
709 |
void *data, int checksum, zio_done_func_t *done, void *private, |
|
5450 | 710 |
int priority, int flags, boolean_t labels) |
789 | 711 |
{ |
712 |
zio_t *zio; |
|
713 |
blkptr_t blk; |
|
714 |
||
5329 | 715 |
ZIO_ENTER(vd->vdev_spa); |
716 |
||
5450 | 717 |
zio_phys_bp_init(vd, &blk, offset, size, checksum, labels); |
789 | 718 |
|
719 |
zio = zio_create(pio, vd->vdev_spa, 0, &blk, data, size, done, private, |
|
720 |
ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, |
|
721 |
ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE); |
|
722 |
||
723 |
zio->io_vd = vd; |
|
724 |
zio->io_offset = offset; |
|
725 |
||
726 |
/* |
|
727 |
* Work off our copy of the bp so the caller can free it. |
|
728 |
*/ |
|
729 |
zio->io_bp = &zio->io_bp_copy; |
|
730 |
||
731 |
return (zio); |
|
732 |
} |
|
733 |
||
734 |
zio_t * |
|
735 |
zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, |
|
736 |
void *data, int checksum, zio_done_func_t *done, void *private, |
|
5450 | 737 |
int priority, int flags, boolean_t labels) |
789 | 738 |
{ |
739 |
zio_block_tail_t *zbt; |
|
740 |
void *wbuf; |
|
741 |
zio_t *zio; |
|
742 |
blkptr_t blk; |
|
743 |
||
5329 | 744 |
ZIO_ENTER(vd->vdev_spa); |
745 |
||
5450 | 746 |
zio_phys_bp_init(vd, &blk, offset, size, checksum, labels); |
789 | 747 |
|
748 |
zio = zio_create(pio, vd->vdev_spa, 0, &blk, data, size, done, private, |
|
749 |
ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, |
|
750 |
ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE); |
|
751 |
||
752 |
zio->io_vd = vd; |
|
753 |
zio->io_offset = offset; |
|
754 |
||
755 |
zio->io_bp = &zio->io_bp_copy; |
|
756 |
zio->io_checksum = checksum; |
|
757 |
||
758 |
if (zio_checksum_table[checksum].ci_zbt) { |
|
759 |
/* |
|
760 |
* zbt checksums are necessarily destructive -- they modify |
|
761 |
* one word of the write buffer to hold the verifier/checksum. |
|
762 |
* Therefore, we must make a local copy in case the data is |
|
763 |
* being written to multiple places. |
|
764 |
*/ |
|
765 |
wbuf = zio_buf_alloc(size); |
|
766 |
bcopy(data, wbuf, size); |
|
767 |
zio_push_transform(zio, wbuf, size, size); |
|
768 |
||
769 |
zbt = (zio_block_tail_t *)((char *)wbuf + size) - 1; |
|
770 |
zbt->zbt_cksum = blk.blk_cksum; |
|
771 |
} |
|
772 |
||
773 |
return (zio); |
|
774 |
} |
|
775 |
||
776 |
/* |
|
777 |
* Create a child I/O to do some work for us. It has no associated bp. |
|
778 |
*/ |
|
779 |
zio_t * |
|
780 |
zio_vdev_child_io(zio_t *zio, blkptr_t *bp, vdev_t *vd, uint64_t offset, |
|
781 |
void *data, uint64_t size, int type, int priority, int flags, |
|
782 |
zio_done_func_t *done, void *private) |
|
783 |
{ |
|
784 |
uint32_t pipeline = ZIO_VDEV_CHILD_PIPELINE; |
|
785 |
zio_t *cio; |
|
786 |
||
787 |
if (type == ZIO_TYPE_READ && bp != NULL) { |
|
788 |
/* |
|
789 |
* If we have the bp, then the child should perform the |
|
790 |
* checksum and the parent need not. This pushes error |
|
791 |
* detection as close to the leaves as possible and |
|
792 |
* eliminates redundant checksums in the interior nodes. |
|
793 |
*/ |
|
794 |
pipeline |= 1U << ZIO_STAGE_CHECKSUM_VERIFY; |
|
795 |
zio->io_pipeline &= ~(1U << ZIO_STAGE_CHECKSUM_VERIFY); |
|
796 |
} |
|
797 |
||
798 |
cio = zio_create(zio, zio->io_spa, zio->io_txg, bp, data, size, |
|
799 |
done, private, type, priority, |
|
800 |
(zio->io_flags & ZIO_FLAG_VDEV_INHERIT) | ZIO_FLAG_CANFAIL | flags, |
|
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
801 |
ZIO_STAGE_VDEV_IO_START - 1, pipeline); |
789 | 802 |
|
803 |
cio->io_vd = vd; |
|
804 |
cio->io_offset = offset; |
|
805 |
||
806 |
return (cio); |
|
807 |
} |
|
808 |
||
809 |
/* |
|
810 |
* ========================================================================== |
|
811 |
* Initiate I/O, either sync or async |
|
812 |
* ========================================================================== |
|
813 |
*/ |
|
814 |
int |
|
815 |
zio_wait(zio_t *zio) |
|
816 |
{ |
|
817 |
int error; |
|
818 |
||
819 |
ASSERT(zio->io_stage == ZIO_STAGE_OPEN); |
|
820 |
||
821 |
zio->io_waiter = curthread; |
|
822 |
||
823 |
zio_next_stage_async(zio); |
|
824 |
||
825 |
mutex_enter(&zio->io_lock); |
|
826 |
while (zio->io_stalled != ZIO_STAGE_DONE) |
|
827 |
cv_wait(&zio->io_cv, &zio->io_lock); |
|
828 |
mutex_exit(&zio->io_lock); |
|
829 |
||
830 |
error = zio->io_error; |
|
2856 | 831 |
mutex_destroy(&zio->io_lock); |
4831
41ec732c6d9f
6584470 zdb needs to initialize the bpl_lock mutex
gw25295
parents:
4634
diff
changeset
|
832 |
cv_destroy(&zio->io_cv); |
4055 | 833 |
kmem_cache_free(zio_cache, zio); |
789 | 834 |
|
835 |
return (error); |
|
836 |
} |
|
837 |
||
838 |
void |
|
839 |
zio_nowait(zio_t *zio) |
|
840 |
{ |
|
841 |
zio_next_stage_async(zio); |
|
842 |
} |
|
843 |
||
844 |
/* |
|
845 |
* ========================================================================== |
|
846 |
* I/O pipeline interlocks: parent/child dependency scoreboarding |
|
847 |
* ========================================================================== |
|
848 |
*/ |
|
849 |
static void |
|
850 |
zio_wait_for_children(zio_t *zio, uint32_t stage, uint64_t *countp) |
|
851 |
{ |
|
852 |
mutex_enter(&zio->io_lock); |
|
853 |
if (*countp == 0) { |
|
854 |
ASSERT(zio->io_stalled == 0); |
|
855 |
mutex_exit(&zio->io_lock); |
|
856 |
zio_next_stage(zio); |
|
857 |
} else { |
|
858 |
zio->io_stalled = stage; |
|
859 |
mutex_exit(&zio->io_lock); |
|
860 |
} |
|
861 |
} |
|
862 |
||
863 |
static void |
|
864 |
zio_notify_parent(zio_t *zio, uint32_t stage, uint64_t *countp) |
|
865 |
{ |
|
866 |
zio_t *pio = zio->io_parent; |
|
867 |
||
868 |
mutex_enter(&pio->io_lock); |
|
869 |
if (pio->io_error == 0 && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE)) |
|
870 |
pio->io_error = zio->io_error; |
|
5329 | 871 |
ASSERT3U(*countp, >, 0); |
789 | 872 |
if (--*countp == 0 && pio->io_stalled == stage) { |
873 |
pio->io_stalled = 0; |
|
874 |
mutex_exit(&pio->io_lock); |
|
875 |
zio_next_stage_async(pio); |
|
876 |
} else { |
|
877 |
mutex_exit(&pio->io_lock); |
|
878 |
} |
|
879 |
} |
|
880 |
||
881 |
static void |
|
882 |
zio_wait_children_ready(zio_t *zio) |
|
883 |
{ |
|
884 |
zio_wait_for_children(zio, ZIO_STAGE_WAIT_CHILDREN_READY, |
|
885 |
&zio->io_children_notready); |
|
886 |
} |
|
887 |
||
888 |
void |
|
889 |
zio_wait_children_done(zio_t *zio) |
|
890 |
{ |
|
891 |
zio_wait_for_children(zio, ZIO_STAGE_WAIT_CHILDREN_DONE, |
|
892 |
&zio->io_children_notdone); |
|
893 |
} |
|
894 |
||
895 |
static void |
|
5329 | 896 |
zio_read_init(zio_t *zio) |
897 |
{ |
|
898 |
if (BP_GET_COMPRESS(zio->io_bp) != ZIO_COMPRESS_OFF) { |
|
899 |
uint64_t csize = BP_GET_PSIZE(zio->io_bp); |
|
900 |
void *cbuf = zio_buf_alloc(csize); |
|
901 |
||
902 |
zio_push_transform(zio, cbuf, csize, csize); |
|
903 |
zio->io_pipeline |= 1U << ZIO_STAGE_READ_DECOMPRESS; |
|
904 |
} |
|
905 |
||
906 |
if (BP_IS_GANG(zio->io_bp)) { |
|
907 |
uint64_t gsize = SPA_GANGBLOCKSIZE; |
|
908 |
void *gbuf = zio_buf_alloc(gsize); |
|
909 |
||
910 |
zio_push_transform(zio, gbuf, gsize, gsize); |
|
911 |
zio->io_pipeline |= 1U << ZIO_STAGE_READ_GANG_MEMBERS; |
|
912 |
} |
|
913 |
zio_next_stage(zio); |
|
914 |
} |
|
915 |
||
916 |
static void |
|
789 | 917 |
zio_ready(zio_t *zio) |
918 |
{ |
|
919 |
zio_t *pio = zio->io_parent; |
|
920 |
||
3547
e396e0a440b1
6512391 DMU should leverage ZIO dependencies to achieve greater parallelism
maybee
parents:
3463
diff
changeset
|
921 |
if (zio->io_ready) |
e396e0a440b1
6512391 DMU should leverage ZIO dependencies to achieve greater parallelism
maybee
parents:
3463
diff
changeset
|
922 |
zio->io_ready(zio); |
e396e0a440b1
6512391 DMU should leverage ZIO dependencies to achieve greater parallelism
maybee
parents:
3463
diff
changeset
|
923 |
|
789 | 924 |
if (pio != NULL) |
925 |
zio_notify_parent(zio, ZIO_STAGE_WAIT_CHILDREN_READY, |
|
926 |
&pio->io_children_notready); |
|
927 |
||
928 |
if (zio->io_bp) |
|
929 |
zio->io_bp_copy = *zio->io_bp; |
|
930 |
||
931 |
zio_next_stage(zio); |
|
932 |
} |
|
933 |
||
934 |
static void |
|
5329 | 935 |
zio_vdev_retry_io(zio_t *zio) |
789 | 936 |
{ |
937 |
zio_t *pio = zio->io_parent; |
|
5329 | 938 |
|
939 |
/* |
|
940 |
* Preserve the failed bp so that the io_ready() callback can |
|
941 |
* update the accounting accordingly. The callback will also be |
|
942 |
* responsible for freeing the previously allocated block, if one |
|
943 |
* exists. |
|
944 |
*/ |
|
945 |
zio->io_bp_orig = *zio->io_bp; |
|
946 |
||
947 |
/* |
|
948 |
* We must zero out the old DVA and blk_birth before reallocating |
|
5403 | 949 |
* the bp. |
5329 | 950 |
*/ |
5403 | 951 |
BP_ZERO_DVAS(zio->io_bp); |
5329 | 952 |
zio_reset(zio); |
953 |
||
954 |
if (pio) { |
|
955 |
/* |
|
956 |
* Let the parent know that we will |
|
957 |
* re-alloc the write (=> new bp info). |
|
958 |
*/ |
|
959 |
mutex_enter(&pio->io_lock); |
|
960 |
pio->io_children_notready++; |
|
961 |
||
962 |
/* |
|
963 |
* If the parent I/O is still in the open stage, then |
|
964 |
* don't bother telling it to retry since it hasn't |
|
965 |
* progressed far enough for it to care. |
|
966 |
*/ |
|
967 |
if (pio->io_stage > ZIO_STAGE_OPEN && IO_IS_ALLOCATING(pio)) |
|
968 |
pio->io_flags |= ZIO_FLAG_WRITE_RETRY; |
|
969 |
||
970 |
ASSERT(pio->io_stage <= ZIO_STAGE_WAIT_CHILDREN_DONE); |
|
971 |
mutex_exit(&pio->io_lock); |
|
972 |
} |
|
973 |
||
974 |
/* |
|
975 |
* We are getting ready to process the retry request so clear |
|
976 |
* the flag and the zio's current error status. |
|
977 |
*/ |
|
978 |
zio->io_flags &= ~ZIO_FLAG_WRITE_RETRY; |
|
979 |
zio->io_error = 0; |
|
980 |
zio_next_stage_async(zio); |
|
981 |
} |
|
982 |
||
983 |
int |
|
984 |
zio_vdev_resume_io(spa_t *spa) |
|
985 |
{ |
|
986 |
zio_t *zio; |
|
987 |
||
988 |
mutex_enter(&spa->spa_zio_lock); |
|
989 |
||
990 |
/* |
|
991 |
* Probe all of vdevs that have experienced an I/O error. |
|
992 |
* If we are still unable to verify the integrity of the vdev |
|
993 |
* then we prevent the resume from proceeeding. |
|
994 |
*/ |
|
995 |
for (zio = list_head(&spa->spa_zio_list); zio != NULL; |
|
996 |
zio = list_next(&spa->spa_zio_list, zio)) { |
|
997 |
int error = 0; |
|
998 |
||
999 |
/* We only care about I/Os that must succeed */ |
|
1000 |
if (zio->io_vd == NULL || zio->io_flags & ZIO_FLAG_CANFAIL) |
|
1001 |
continue; |
|
1002 |
error = vdev_probe(zio->io_vd); |
|
1003 |
if (error) { |
|
1004 |
mutex_exit(&spa->spa_zio_lock); |
|
1005 |
return (error); |
|
1006 |
} |
|
1007 |
} |
|
1008 |
||
1009 |
/* |
|
1010 |
* Clear the vdev stats so that I/O can flow. |
|
1011 |
*/ |
|
1012 |
vdev_clear(spa, NULL, B_FALSE); |
|
1013 |
||
1014 |
spa->spa_state = POOL_STATE_ACTIVE; |
|
1015 |
while ((zio = list_head(&spa->spa_zio_list)) != NULL) { |
|
1016 |
list_remove(&spa->spa_zio_list, zio); |
|
1017 |
zio->io_error = 0; |
|
1018 |
||
1019 |
/* |
|
1020 |
* If we are resuming an allocating I/O then we force it |
|
1021 |
* to retry and let it resume operation where it left off. |
|
1022 |
* Otherwise, go back to the ready stage and pick up from |
|
1023 |
* there. |
|
1024 |
*/ |
|
1025 |
if (zio_write_retry && IO_IS_ALLOCATING(zio)) { |
|
1026 |
zio->io_flags |= ZIO_FLAG_WRITE_RETRY; |
|
1027 |
zio->io_stage--; |
|
1028 |
} else { |
|
1029 |
zio->io_stage = ZIO_STAGE_READY; |
|
1030 |
} |
|
1031 |
||
1032 |
(void) taskq_dispatch(zio_taskq, zio_resubmit_stage_async, |
|
1033 |
zio, TQ_SLEEP); |
|
1034 |
} |
|
1035 |
mutex_exit(&spa->spa_zio_lock); |
|
1036 |
||
1037 |
/* |
|
1038 |
* Wait for the taskqs to finish and recheck the pool state since |
|
1039 |
* it's possible that a resumed I/O has failed again. |
|
1040 |
*/ |
|
1041 |
taskq_wait(zio_taskq); |
|
1042 |
if (spa_state(spa) == POOL_STATE_IO_FAILURE) |
|
1043 |
return (EIO); |
|
1044 |
||
1045 |
mutex_enter(&spa->spa_zio_lock); |
|
1046 |
cv_broadcast(&spa->spa_zio_cv); |
|
1047 |
mutex_exit(&spa->spa_zio_lock); |
|
1048 |
||
1049 |
return (0); |
|
1050 |
} |
|
1051 |
||
1052 |
static void |
|
1053 |
zio_vdev_suspend_io(zio_t *zio) |
|
1054 |
{ |
|
1055 |
spa_t *spa = zio->io_spa; |
|
1056 |
||
1057 |
/* |
|
1058 |
* We've experienced an unrecoverable failure so |
|
1059 |
* set the pool state accordingly and queue all |
|
1060 |
* failed IOs. |
|
1061 |
*/ |
|
1062 |
spa->spa_state = POOL_STATE_IO_FAILURE; |
|
1063 |
||
1064 |
mutex_enter(&spa->spa_zio_lock); |
|
1065 |
list_insert_tail(&spa->spa_zio_list, zio); |
|
1066 |
||
1067 |
#ifndef _KERNEL |
|
1068 |
/* Used to notify ztest that the pool has suspended */ |
|
1069 |
cv_broadcast(&spa->spa_zio_cv); |
|
1070 |
#endif |
|
1071 |
mutex_exit(&spa->spa_zio_lock); |
|
1072 |
} |
|
1073 |
||
1074 |
static void |
|
1075 |
zio_assess(zio_t *zio) |
|
1076 |
{ |
|
789 | 1077 |
spa_t *spa = zio->io_spa; |
1078 |
blkptr_t *bp = zio->io_bp; |
|
1079 |
vdev_t *vd = zio->io_vd; |
|
1080 |
||
1081 |
ASSERT(zio->io_children_notready == 0); |
|
1082 |
ASSERT(zio->io_children_notdone == 0); |
|
1083 |
||
1084 |
if (bp != NULL) { |
|
1085 |
ASSERT(bp->blk_pad[0] == 0); |
|
1086 |
ASSERT(bp->blk_pad[1] == 0); |
|
1087 |
ASSERT(bp->blk_pad[2] == 0); |
|
1088 |
ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0); |
|
1089 |
if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) && |
|
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1090 |
!(zio->io_flags & ZIO_FLAG_IO_REPAIR)) { |
789 | 1091 |
ASSERT(!BP_SHOULD_BYTESWAP(bp)); |
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1092 |
if (zio->io_ndvas != 0) |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1093 |
ASSERT3U(zio->io_ndvas, <=, BP_GET_NDVAS(bp)); |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1094 |
ASSERT(BP_COUNT_GANG(bp) == 0 || |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1095 |
(BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp))); |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1096 |
} |
789 | 1097 |
} |
1098 |
||
5329 | 1099 |
/* |
1100 |
* Some child I/O has indicated that a retry is necessary, so |
|
1101 |
* we set an error on the I/O and let the logic below do the |
|
1102 |
* rest. |
|
1103 |
*/ |
|
1104 |
if (zio->io_flags & ZIO_FLAG_WRITE_RETRY) |
|
1105 |
zio->io_error = ERESTART; |
|
1106 |
||
789 | 1107 |
if (vd != NULL) |
1108 |
vdev_stat_update(zio); |
|
1109 |
||
1110 |
if (zio->io_error) { |
|
1544 | 1111 |
/* |
1112 |
* If this I/O is attached to a particular vdev, |
|
1113 |
* generate an error message describing the I/O failure |
|
1114 |
* at the block level. We ignore these errors if the |
|
1115 |
* device is currently unavailable. |
|
1116 |
*/ |
|
1732 | 1117 |
if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd)) |
5329 | 1118 |
zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0); |
789 | 1119 |
|
1544 | 1120 |
if ((zio->io_error == EIO || |
1121 |
!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) && |
|
1122 |
zio->io_logical == zio) { |
|
1123 |
/* |
|
1124 |
* For root I/O requests, tell the SPA to log the error |
|
1125 |
* appropriately. Also, generate a logical data |
|
1126 |
* ereport. |
|
1127 |
*/ |
|
5329 | 1128 |
spa_log_error(spa, zio); |
1544 | 1129 |
|
5329 | 1130 |
zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio, |
1131 |
0, 0); |
|
1544 | 1132 |
} |
789 | 1133 |
|
1544 | 1134 |
/* |
5403 | 1135 |
* If we are an allocating I/O then we attempt to reissue |
1136 |
* the I/O on another vdev unless the pool is out of space. |
|
1137 |
* We handle this condition based on the spa's failmode |
|
1138 |
* property. |
|
5329 | 1139 |
*/ |
1140 |
if (zio_write_retry && zio->io_error != ENOSPC && |
|
5403 | 1141 |
IO_IS_ALLOCATING(zio)) { |
5329 | 1142 |
zio_vdev_retry_io(zio); |
1143 |
return; |
|
1144 |
} |
|
1145 |
ASSERT(!(zio->io_flags & ZIO_FLAG_WRITE_RETRY)); |
|
1146 |
||
1147 |
/* |
|
1148 |
* For I/O requests that cannot fail, we carry out |
|
1149 |
* the requested behavior based on the failmode pool |
|
1150 |
* property. |
|
1151 |
* |
|
1152 |
* XXX - Need to differentiate between an ENOSPC as |
|
1153 |
* a result of vdev failures vs. a full pool. |
|
1544 | 1154 |
*/ |
1155 |
if (!(zio->io_flags & ZIO_FLAG_CANFAIL)) { |
|
3459
358b6bdb15b6
6512311 zio_done() is chunky and needs to go on a diet
ek110237
parents:
3290
diff
changeset
|
1156 |
char *blkbuf; |
358b6bdb15b6
6512311 zio_done() is chunky and needs to go on a diet
ek110237
parents:
3290
diff
changeset
|
1157 |
|
5329 | 1158 |
#ifdef ZFS_DEBUG |
3459
358b6bdb15b6
6512311 zio_done() is chunky and needs to go on a diet
ek110237
parents:
3290
diff
changeset
|
1159 |
blkbuf = kmem_alloc(BP_SPRINTF_LEN, KM_NOSLEEP); |
358b6bdb15b6
6512311 zio_done() is chunky and needs to go on a diet
ek110237
parents:
3290
diff
changeset
|
1160 |
if (blkbuf) { |
358b6bdb15b6
6512311 zio_done() is chunky and needs to go on a diet
ek110237
parents:
3290
diff
changeset
|
1161 |
sprintf_blkptr(blkbuf, BP_SPRINTF_LEN, |
358b6bdb15b6
6512311 zio_done() is chunky and needs to go on a diet
ek110237
parents:
3290
diff
changeset
|
1162 |
bp ? bp : &zio->io_bp_copy); |
358b6bdb15b6
6512311 zio_done() is chunky and needs to go on a diet
ek110237
parents:
3290
diff
changeset
|
1163 |
} |
5329 | 1164 |
cmn_err(CE_WARN, "ZFS: %s (%s on %s off %llx: zio %p " |
1165 |
"%s): error %d", zio->io_error == ECKSUM ? |
|
1544 | 1166 |
"bad checksum" : "I/O failure", |
1167 |
zio_type_name[zio->io_type], |
|
1168 |
vdev_description(vd), |
|
1169 |
(u_longlong_t)zio->io_offset, |
|
5329 | 1170 |
(void *)zio, blkbuf ? blkbuf : "", zio->io_error); |
1171 |
#endif |
|
1172 |
||
1173 |
if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC) { |
|
1174 |
fm_panic("Pool '%s' has encountered an " |
|
1175 |
"uncorrectable I/O failure and the " |
|
1176 |
"failure mode property for this pool " |
|
1177 |
"is set to panic.", spa_name(spa)); |
|
1178 |
} else { |
|
1179 |
cmn_err(CE_WARN, "Pool '%s' has encountered " |
|
1180 |
"an uncorrectable I/O error. Manual " |
|
1181 |
"intervention is required.", |
|
1182 |
spa_name(spa)); |
|
1183 |
zio_vdev_suspend_io(zio); |
|
1184 |
} |
|
1185 |
return; |
|
1544 | 1186 |
} |
789 | 1187 |
} |
5329 | 1188 |
ASSERT(!(zio->io_flags & ZIO_FLAG_WRITE_RETRY)); |
1189 |
ASSERT(zio->io_children_notready == 0); |
|
1190 |
zio_next_stage(zio); |
|
1191 |
} |
|
1192 |
||
1193 |
static void |
|
1194 |
zio_done(zio_t *zio) |
|
1195 |
{ |
|
1196 |
zio_t *pio = zio->io_parent; |
|
1197 |
spa_t *spa = zio->io_spa; |
|
1198 |
||
1199 |
ASSERT(zio->io_children_notready == 0); |
|
1200 |
ASSERT(zio->io_children_notdone == 0); |
|
1201 |
||
789 | 1202 |
zio_clear_transform_stack(zio); |
1203 |
||
1204 |
if (zio->io_done) |
|
1205 |
zio->io_done(zio); |
|
1206 |
||
1207 |
ASSERT(zio->io_delegate_list == NULL); |
|
1208 |
ASSERT(zio->io_delegate_next == NULL); |
|
1209 |
||
1210 |
if (pio != NULL) { |
|
1211 |
zio_t *next, *prev; |
|
1212 |
||
1213 |
mutex_enter(&pio->io_lock); |
|
1214 |
next = zio->io_sibling_next; |
|
1215 |
prev = zio->io_sibling_prev; |
|
1216 |
if (next != NULL) |
|
1217 |
next->io_sibling_prev = prev; |
|
1218 |
if (prev != NULL) |
|
1219 |
prev->io_sibling_next = next; |
|
1220 |
if (pio->io_child == zio) |
|
1221 |
pio->io_child = next; |
|
1222 |
mutex_exit(&pio->io_lock); |
|
1223 |
||
1224 |
zio_notify_parent(zio, ZIO_STAGE_WAIT_CHILDREN_DONE, |
|
1225 |
&pio->io_children_notdone); |
|
1226 |
} |
|
1227 |
||
3463 | 1228 |
/* |
4055 | 1229 |
* Note: this I/O is now done, and will shortly be freed, so there is no |
1230 |
* need to clear this (or any other) flag. |
|
3463 | 1231 |
*/ |
1232 |
if (zio->io_flags & ZIO_FLAG_CONFIG_GRABBED) |
|
1544 | 1233 |
spa_config_exit(spa, zio); |
789 | 1234 |
|
1235 |
if (zio->io_waiter != NULL) { |
|
1236 |
mutex_enter(&zio->io_lock); |
|
1237 |
ASSERT(zio->io_stage == ZIO_STAGE_DONE); |
|
1238 |
zio->io_stalled = zio->io_stage; |
|
1239 |
cv_broadcast(&zio->io_cv); |
|
1240 |
mutex_exit(&zio->io_lock); |
|
1241 |
} else { |
|
4831
41ec732c6d9f
6584470 zdb needs to initialize the bpl_lock mutex
gw25295
parents:
4634
diff
changeset
|
1242 |
mutex_destroy(&zio->io_lock); |
41ec732c6d9f
6584470 zdb needs to initialize the bpl_lock mutex
gw25295
parents:
4634
diff
changeset
|
1243 |
cv_destroy(&zio->io_cv); |
4055 | 1244 |
kmem_cache_free(zio_cache, zio); |
789 | 1245 |
} |
1246 |
} |
|
1247 |
||
1248 |
/* |
|
1249 |
* ========================================================================== |
|
1250 |
* Compression support |
|
1251 |
* ========================================================================== |
|
1252 |
*/ |
|
1253 |
static void |
|
1254 |
zio_write_compress(zio_t *zio) |
|
1255 |
{ |
|
1256 |
int compress = zio->io_compress; |
|
1257 |
blkptr_t *bp = zio->io_bp; |
|
1258 |
void *cbuf; |
|
1259 |
uint64_t lsize = zio->io_size; |
|
1260 |
uint64_t csize = lsize; |
|
1261 |
uint64_t cbufsize = 0; |
|
1262 |
int pass; |
|
1263 |
||
1264 |
if (bp->blk_birth == zio->io_txg) { |
|
1265 |
/* |
|
1266 |
* We're rewriting an existing block, which means we're |
|
1267 |
* working on behalf of spa_sync(). For spa_sync() to |
|
1268 |
* converge, it must eventually be the case that we don't |
|
1269 |
* have to allocate new blocks. But compression changes |
|
1270 |
* the blocksize, which forces a reallocate, and makes |
|
1271 |
* convergence take longer. Therefore, after the first |
|
1272 |
* few passes, stop compressing to ensure convergence. |
|
1273 |
*/ |
|
1274 |
pass = spa_sync_pass(zio->io_spa); |
|
1275 |
if (pass > zio_sync_pass.zp_dontcompress) |
|
1276 |
compress = ZIO_COMPRESS_OFF; |
|
1277 |
} else { |
|
1278 |
ASSERT(BP_IS_HOLE(bp)); |
|
1279 |
pass = 1; |
|
1280 |
} |
|
1281 |
||
1282 |
if (compress != ZIO_COMPRESS_OFF) |
|
1283 |
if (!zio_compress_data(compress, zio->io_data, zio->io_size, |
|
1284 |
&cbuf, &csize, &cbufsize)) |
|
1285 |
compress = ZIO_COMPRESS_OFF; |
|
1286 |
||
1287 |
if (compress != ZIO_COMPRESS_OFF && csize != 0) |
|
1288 |
zio_push_transform(zio, cbuf, csize, cbufsize); |
|
1289 |
||
1290 |
/* |
|
1291 |
* The final pass of spa_sync() must be all rewrites, but the first |
|
1292 |
* few passes offer a trade-off: allocating blocks defers convergence, |
|
1293 |
* but newly allocated blocks are sequential, so they can be written |
|
1294 |
* to disk faster. Therefore, we allow the first few passes of |
|
1295 |
* spa_sync() to reallocate new blocks, but force rewrites after that. |
|
1296 |
* There should only be a handful of blocks after pass 1 in any case. |
|
1297 |
*/ |
|
1298 |
if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == csize && |
|
1299 |
pass > zio_sync_pass.zp_rewrite) { |
|
1300 |
ASSERT(csize != 0); |
|
2885 | 1301 |
BP_SET_LSIZE(bp, lsize); |
1302 |
BP_SET_COMPRESS(bp, compress); |
|
789 | 1303 |
zio->io_pipeline = ZIO_REWRITE_PIPELINE; |
1304 |
} else { |
|
3882 | 1305 |
if (bp->blk_birth == zio->io_txg) |
1306 |
BP_ZERO(bp); |
|
789 | 1307 |
if (csize == 0) { |
1308 |
BP_ZERO(bp); |
|
1309 |
zio->io_pipeline = ZIO_WAIT_FOR_CHILDREN_PIPELINE; |
|
1310 |
} else { |
|
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1311 |
ASSERT3U(BP_GET_NDVAS(bp), ==, 0); |
789 | 1312 |
BP_SET_LSIZE(bp, lsize); |
1313 |
BP_SET_PSIZE(bp, csize); |
|
1314 |
BP_SET_COMPRESS(bp, compress); |
|
1315 |
zio->io_pipeline = ZIO_WRITE_ALLOCATE_PIPELINE; |
|
1316 |
} |
|
1317 |
} |
|
1318 |
||
1319 |
zio_next_stage(zio); |
|
1320 |
} |
|
1321 |
||
1322 |
static void |
|
1323 |
zio_read_decompress(zio_t *zio) |
|
1324 |
{ |
|
1325 |
blkptr_t *bp = zio->io_bp; |
|
1326 |
void *data; |
|
1327 |
uint64_t size; |
|
1328 |
uint64_t bufsize; |
|
1329 |
int compress = BP_GET_COMPRESS(bp); |
|
1330 |
||
1331 |
ASSERT(compress != ZIO_COMPRESS_OFF); |
|
1332 |
||
1333 |
zio_pop_transform(zio, &data, &size, &bufsize); |
|
1334 |
||
1335 |
if (zio_decompress_data(compress, data, size, |
|
1336 |
zio->io_data, zio->io_size)) |
|
1337 |
zio->io_error = EIO; |
|
1338 |
||
1339 |
zio_buf_free(data, bufsize); |
|
1340 |
||
1341 |
zio_next_stage(zio); |
|
1342 |
} |
|
1343 |
||
1344 |
/* |
|
1345 |
* ========================================================================== |
|
1346 |
* Gang block support |
|
1347 |
* ========================================================================== |
|
1348 |
*/ |
|
1349 |
static void |
|
1350 |
zio_gang_pipeline(zio_t *zio) |
|
1351 |
{ |
|
1352 |
/* |
|
1353 |
* By default, the pipeline assumes that we're dealing with a gang |
|
1354 |
* block. If we're not, strip out any gang-specific stages. |
|
1355 |
*/ |
|
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1356 |
if (!BP_IS_GANG(zio->io_bp)) |
789 | 1357 |
zio->io_pipeline &= ~ZIO_GANG_STAGES; |
1358 |
||
1359 |
zio_next_stage(zio); |
|
1360 |
} |
|
1361 |
||
1362 |
static void |
|
1363 |
zio_gang_byteswap(zio_t *zio) |
|
1364 |
{ |
|
1365 |
ASSERT(zio->io_size == SPA_GANGBLOCKSIZE); |
|
1366 |
||
1367 |
if (BP_SHOULD_BYTESWAP(zio->io_bp)) |
|
1368 |
byteswap_uint64_array(zio->io_data, zio->io_size); |
|
1369 |
} |
|
1370 |
||
1371 |
static void |
|
1372 |
zio_get_gang_header(zio_t *zio) |
|
1373 |
{ |
|
1374 |
blkptr_t *bp = zio->io_bp; |
|
1375 |
uint64_t gsize = SPA_GANGBLOCKSIZE; |
|
1376 |
void *gbuf = zio_buf_alloc(gsize); |
|
1377 |
||
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1378 |
ASSERT(BP_IS_GANG(bp)); |
789 | 1379 |
|
1380 |
zio_push_transform(zio, gbuf, gsize, gsize); |
|
1381 |
||
1382 |
zio_nowait(zio_create(zio, zio->io_spa, bp->blk_birth, bp, gbuf, gsize, |
|
1383 |
NULL, NULL, ZIO_TYPE_READ, zio->io_priority, |
|
1384 |
zio->io_flags & ZIO_FLAG_GANG_INHERIT, |
|
5329 | 1385 |
ZIO_STAGE_OPEN, ZIO_READ_GANG_PIPELINE)); |
789 | 1386 |
|
1387 |
zio_wait_children_done(zio); |
|
1388 |
} |
|
1389 |
||
1390 |
static void |
|
1391 |
zio_read_gang_members(zio_t *zio) |
|
1392 |
{ |
|
1393 |
zio_gbh_phys_t *gbh; |
|
1394 |
uint64_t gsize, gbufsize, loff, lsize; |
|
1395 |
int i; |
|
1396 |
||
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1397 |
ASSERT(BP_IS_GANG(zio->io_bp)); |
789 | 1398 |
|
1399 |
zio_gang_byteswap(zio); |
|
1400 |
zio_pop_transform(zio, (void **)&gbh, &gsize, &gbufsize); |
|
1401 |
||
1402 |
for (loff = 0, i = 0; loff != zio->io_size; loff += lsize, i++) { |
|
1403 |
blkptr_t *gbp = &gbh->zg_blkptr[i]; |
|
1404 |
lsize = BP_GET_PSIZE(gbp); |
|
1405 |
||
1406 |
ASSERT(BP_GET_COMPRESS(gbp) == ZIO_COMPRESS_OFF); |
|
1407 |
ASSERT3U(lsize, ==, BP_GET_LSIZE(gbp)); |
|
1408 |
ASSERT3U(loff + lsize, <=, zio->io_size); |
|
1409 |
ASSERT(i < SPA_GBH_NBLKPTRS); |
|
1410 |
ASSERT(!BP_IS_HOLE(gbp)); |
|
1411 |
||
1412 |
zio_nowait(zio_read(zio, zio->io_spa, gbp, |
|
1413 |
(char *)zio->io_data + loff, lsize, NULL, NULL, |
|
1544 | 1414 |
zio->io_priority, zio->io_flags & ZIO_FLAG_GANG_INHERIT, |
1415 |
&zio->io_bookmark)); |
|
789 | 1416 |
} |
1417 |
||
1418 |
zio_buf_free(gbh, gbufsize); |
|
1419 |
zio_wait_children_done(zio); |
|
1420 |
} |
|
1421 |
||
1422 |
static void |
|
1423 |
zio_rewrite_gang_members(zio_t *zio) |
|
1424 |
{ |
|
1425 |
zio_gbh_phys_t *gbh; |
|
1426 |
uint64_t gsize, gbufsize, loff, lsize; |
|
1427 |
int i; |
|
1428 |
||
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1429 |
ASSERT(BP_IS_GANG(zio->io_bp)); |
789 | 1430 |
ASSERT3U(zio->io_size, ==, SPA_GANGBLOCKSIZE); |
1431 |
||
1432 |
zio_gang_byteswap(zio); |
|
1433 |
zio_pop_transform(zio, (void **)&gbh, &gsize, &gbufsize); |
|
1434 |
||
1435 |
ASSERT(gsize == gbufsize); |
|
1436 |
||
1437 |
for (loff = 0, i = 0; loff != zio->io_size; loff += lsize, i++) { |
|
1438 |
blkptr_t *gbp = &gbh->zg_blkptr[i]; |
|
1439 |
lsize = BP_GET_PSIZE(gbp); |
|
1440 |
||
1441 |
ASSERT(BP_GET_COMPRESS(gbp) == ZIO_COMPRESS_OFF); |
|
1442 |
ASSERT3U(lsize, ==, BP_GET_LSIZE(gbp)); |
|
1443 |
ASSERT3U(loff + lsize, <=, zio->io_size); |
|
1444 |
ASSERT(i < SPA_GBH_NBLKPTRS); |
|
1445 |
ASSERT(!BP_IS_HOLE(gbp)); |
|
1446 |
||
1447 |
zio_nowait(zio_rewrite(zio, zio->io_spa, zio->io_checksum, |
|
1448 |
zio->io_txg, gbp, (char *)zio->io_data + loff, lsize, |
|
1544 | 1449 |
NULL, NULL, zio->io_priority, zio->io_flags, |
1450 |
&zio->io_bookmark)); |
|
789 | 1451 |
} |
1452 |
||
1453 |
zio_push_transform(zio, gbh, gsize, gbufsize); |
|
1454 |
zio_wait_children_ready(zio); |
|
1455 |
} |
|
1456 |
||
1457 |
static void |
|
1458 |
zio_free_gang_members(zio_t *zio) |
|
1459 |
{ |
|
1460 |
zio_gbh_phys_t *gbh; |
|
1461 |
uint64_t gsize, gbufsize; |
|
1462 |
int i; |
|
1463 |
||
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1464 |
ASSERT(BP_IS_GANG(zio->io_bp)); |
789 | 1465 |
|
1466 |
zio_gang_byteswap(zio); |
|
1467 |
zio_pop_transform(zio, (void **)&gbh, &gsize, &gbufsize); |
|
1468 |
||
1469 |
for (i = 0; i < SPA_GBH_NBLKPTRS; i++) { |
|
1470 |
blkptr_t *gbp = &gbh->zg_blkptr[i]; |
|
1471 |
||
1472 |
if (BP_IS_HOLE(gbp)) |
|
1473 |
continue; |
|
1474 |
zio_nowait(zio_free(zio, zio->io_spa, zio->io_txg, |
|
1475 |
gbp, NULL, NULL)); |
|
1476 |
} |
|
1477 |
||
1478 |
zio_buf_free(gbh, gbufsize); |
|
1479 |
zio_next_stage(zio); |
|
1480 |
} |
|
1481 |
||
1482 |
static void |
|
1483 |
zio_claim_gang_members(zio_t *zio) |
|
1484 |
{ |
|
1485 |
zio_gbh_phys_t *gbh; |
|
1486 |
uint64_t gsize, gbufsize; |
|
1487 |
int i; |
|
1488 |
||
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1489 |
ASSERT(BP_IS_GANG(zio->io_bp)); |
789 | 1490 |
|
1491 |
zio_gang_byteswap(zio); |
|
1492 |
zio_pop_transform(zio, (void **)&gbh, &gsize, &gbufsize); |
|
1493 |
||
1494 |
for (i = 0; i < SPA_GBH_NBLKPTRS; i++) { |
|
1495 |
blkptr_t *gbp = &gbh->zg_blkptr[i]; |
|
1496 |
if (BP_IS_HOLE(gbp)) |
|
1497 |
continue; |
|
1498 |
zio_nowait(zio_claim(zio, zio->io_spa, zio->io_txg, |
|
1499 |
gbp, NULL, NULL)); |
|
1500 |
} |
|
1501 |
||
1502 |
zio_buf_free(gbh, gbufsize); |
|
1503 |
zio_next_stage(zio); |
|
1504 |
} |
|
1505 |
||
1506 |
static void |
|
1507 |
zio_write_allocate_gang_member_done(zio_t *zio) |
|
1508 |
{ |
|
1509 |
zio_t *pio = zio->io_parent; |
|
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1510 |
dva_t *cdva = zio->io_bp->blk_dva; |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1511 |
dva_t *pdva = pio->io_bp->blk_dva; |
789 | 1512 |
uint64_t asize; |
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1513 |
int d; |
789 | 1514 |
|
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1515 |
ASSERT3U(pio->io_ndvas, ==, zio->io_ndvas); |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1516 |
ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp)); |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1517 |
ASSERT3U(zio->io_ndvas, <=, BP_GET_NDVAS(zio->io_bp)); |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1518 |
ASSERT3U(pio->io_ndvas, <=, BP_GET_NDVAS(pio->io_bp)); |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1519 |
|
789 | 1520 |
mutex_enter(&pio->io_lock); |
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1521 |
for (d = 0; d < BP_GET_NDVAS(pio->io_bp); d++) { |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1522 |
ASSERT(DVA_GET_GANG(&pdva[d])); |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1523 |
asize = DVA_GET_ASIZE(&pdva[d]); |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1524 |
asize += DVA_GET_ASIZE(&cdva[d]); |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1525 |
DVA_SET_ASIZE(&pdva[d], asize); |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1526 |
} |
789 | 1527 |
mutex_exit(&pio->io_lock); |
1528 |
} |
|
1529 |
||
5329 | 1530 |
static int |
4527 | 1531 |
zio_write_allocate_gang_members(zio_t *zio, metaslab_class_t *mc) |
789 | 1532 |
{ |
1533 |
blkptr_t *bp = zio->io_bp; |
|
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1534 |
dva_t *dva = bp->blk_dva; |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1535 |
spa_t *spa = zio->io_spa; |
789 | 1536 |
zio_gbh_phys_t *gbh; |
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1537 |
uint64_t txg = zio->io_txg; |
789 | 1538 |
uint64_t resid = zio->io_size; |
1539 |
uint64_t maxalloc = P2ROUNDUP(zio->io_size >> 1, SPA_MINBLOCKSIZE); |
|
1540 |
uint64_t gsize, loff, lsize; |
|
1541 |
uint32_t gbps_left; |
|
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1542 |
int ndvas = zio->io_ndvas; |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1543 |
int gbh_ndvas = MIN(ndvas + 1, spa_max_replication(spa)); |
789 | 1544 |
int error; |
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1545 |
int i, d; |
789 | 1546 |
|
1547 |
gsize = SPA_GANGBLOCKSIZE; |
|
1548 |
gbps_left = SPA_GBH_NBLKPTRS; |
|
1549 |
||
4527 | 1550 |
error = metaslab_alloc(spa, mc, gsize, bp, gbh_ndvas, txg, NULL, |
1551 |
B_FALSE); |
|
5329 | 1552 |
if (error) |
1553 |
return (error); |
|
789 | 1554 |
|
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1555 |
for (d = 0; d < gbh_ndvas; d++) |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1556 |
DVA_SET_GANG(&dva[d], 1); |
789 | 1557 |
|
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1558 |
bp->blk_birth = txg; |
789 | 1559 |
|
1560 |
gbh = zio_buf_alloc(gsize); |
|
1561 |
bzero(gbh, gsize); |
|
1562 |
||
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1563 |
/* We need to test multi-level gang blocks */ |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1564 |
if (maxalloc >= zio_gang_bang && (lbolt & 0x1) == 0) |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1565 |
maxalloc = MAX(maxalloc >> 2, SPA_MINBLOCKSIZE); |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1566 |
|
789 | 1567 |
for (loff = 0, i = 0; loff != zio->io_size; |
1568 |
loff += lsize, resid -= lsize, gbps_left--, i++) { |
|
1569 |
blkptr_t *gbp = &gbh->zg_blkptr[i]; |
|
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1570 |
dva = gbp->blk_dva; |
789 | 1571 |
|
1572 |
ASSERT(gbps_left != 0); |
|
1573 |
maxalloc = MIN(maxalloc, resid); |
|
1574 |
||
1575 |
while (resid <= maxalloc * gbps_left) { |
|
4527 | 1576 |
error = metaslab_alloc(spa, mc, maxalloc, gbp, ndvas, |
3063
b252896b372b
6341569 zio_alloc_blk() vdev distribution performs badly
perrin
parents:
2981
diff
changeset
|
1577 |
txg, bp, B_FALSE); |
789 | 1578 |
if (error == 0) |
1579 |
break; |
|
1580 |
ASSERT3U(error, ==, ENOSPC); |
|
5329 | 1581 |
/* XXX - free up previous allocations? */ |
789 | 1582 |
if (maxalloc == SPA_MINBLOCKSIZE) |
5329 | 1583 |
return (error); |
789 | 1584 |
maxalloc = P2ROUNDUP(maxalloc >> 1, SPA_MINBLOCKSIZE); |
1585 |
} |
|
1586 |
||
1587 |
if (resid <= maxalloc * gbps_left) { |
|
1588 |
lsize = maxalloc; |
|
1589 |
BP_SET_LSIZE(gbp, lsize); |
|
1590 |
BP_SET_PSIZE(gbp, lsize); |
|
1591 |
BP_SET_COMPRESS(gbp, ZIO_COMPRESS_OFF); |
|
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1592 |
gbp->blk_birth = txg; |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1593 |
zio_nowait(zio_rewrite(zio, spa, |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1594 |
zio->io_checksum, txg, gbp, |
789 | 1595 |
(char *)zio->io_data + loff, lsize, |
1596 |
zio_write_allocate_gang_member_done, NULL, |
|
5403 | 1597 |
zio->io_priority, |
1598 |
zio->io_flags & ZIO_FLAG_GANG_INHERIT, |
|
1544 | 1599 |
&zio->io_bookmark)); |
789 | 1600 |
} else { |
1601 |
lsize = P2ROUNDUP(resid / gbps_left, SPA_MINBLOCKSIZE); |
|
1602 |
ASSERT(lsize != SPA_MINBLOCKSIZE); |
|
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1603 |
zio_nowait(zio_write_allocate(zio, spa, |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1604 |
zio->io_checksum, txg, gbp, |
789 | 1605 |
(char *)zio->io_data + loff, lsize, |
1606 |
zio_write_allocate_gang_member_done, NULL, |
|
5403 | 1607 |
zio->io_priority, |
1608 |
zio->io_flags & ZIO_FLAG_GANG_INHERIT)); |
|
789 | 1609 |
} |
1610 |
} |
|
1611 |
||
1612 |
ASSERT(resid == 0 && loff == zio->io_size); |
|
1613 |
||
1614 |
zio->io_pipeline |= 1U << ZIO_STAGE_GANG_CHECKSUM_GENERATE; |
|
1615 |
||
1616 |
zio_push_transform(zio, gbh, gsize, gsize); |
|
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1617 |
/* |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1618 |
* As much as we'd like this to be zio_wait_children_ready(), |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1619 |
* updating our ASIZE doesn't happen until the io_done callback, |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1620 |
* so we have to wait for that to finish in order for our BP |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1621 |
* to be stable. |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1622 |
*/ |
789 | 1623 |
zio_wait_children_done(zio); |
5329 | 1624 |
return (0); |
789 | 1625 |
} |
1626 |
||
1627 |
/* |
|
1628 |
* ========================================================================== |
|
1629 |
* Allocate and free blocks |
|
1630 |
* ========================================================================== |
|
1631 |
*/ |
|
1632 |
static void |
|
1633 |
zio_dva_allocate(zio_t *zio) |
|
1634 |
{ |
|
4527 | 1635 |
spa_t *spa = zio->io_spa; |
1636 |
metaslab_class_t *mc = spa->spa_normal_class; |
|
789 | 1637 |
blkptr_t *bp = zio->io_bp; |
1638 |
int error; |
|
1639 |
||
1640 |
ASSERT(BP_IS_HOLE(bp)); |
|
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1641 |
ASSERT3U(BP_GET_NDVAS(bp), ==, 0); |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1642 |
ASSERT3U(zio->io_ndvas, >, 0); |
4527 | 1643 |
ASSERT3U(zio->io_ndvas, <=, spa_max_replication(spa)); |
789 | 1644 |
|
1645 |
/* For testing, make some blocks above a certain size be gang blocks */ |
|
1646 |
if (zio->io_size >= zio_gang_bang && (lbolt & 0x3) == 0) { |
|
5329 | 1647 |
error = zio_write_allocate_gang_members(zio, mc); |
1648 |
if (error) |
|
1649 |
zio->io_error = error; |
|
789 | 1650 |
return; |
1651 |
} |
|
1652 |
||
5329 | 1653 |
/* |
1654 |
* For testing purposes, we force I/Os to retry. We don't allow |
|
1655 |
* retries beyond the first pass since those I/Os are non-allocating |
|
5403 | 1656 |
* writes. |
5329 | 1657 |
*/ |
1658 |
if (zio_io_fail_shift && |
|
1659 |
spa_sync_pass(zio->io_spa) <= zio_sync_pass.zp_rewrite && |
|
1660 |
zio_io_should_fail(zio_io_fail_shift)) |
|
1661 |
zio->io_flags |= ZIO_FLAG_WRITE_RETRY; |
|
1662 |
||
789 | 1663 |
ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); |
1664 |
||
4527 | 1665 |
error = metaslab_alloc(spa, mc, zio->io_size, bp, zio->io_ndvas, |
3063
b252896b372b
6341569 zio_alloc_blk() vdev distribution performs badly
perrin
parents:
2981
diff
changeset
|
1666 |
zio->io_txg, NULL, B_FALSE); |
789 | 1667 |
|
1668 |
if (error == 0) { |
|
1669 |
bp->blk_birth = zio->io_txg; |
|
5329 | 1670 |
} else if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) { |
1671 |
error = zio_write_allocate_gang_members(zio, mc); |
|
1672 |
if (error == 0) |
|
1673 |
return; |
|
1674 |
zio->io_error = error; |
|
789 | 1675 |
} else { |
1676 |
zio->io_error = error; |
|
1677 |
} |
|
1678 |
zio_next_stage(zio); |
|
1679 |
} |
|
1680 |
||
1681 |
static void |
|
1682 |
zio_dva_free(zio_t *zio) |
|
1683 |
{ |
|
1684 |
blkptr_t *bp = zio->io_bp; |
|
1685 |
||
1807
35c8b566d7af
6410711 intent log blocks don't get invited to pool parties
bonwick
parents:
1775
diff
changeset
|
1686 |
metaslab_free(zio->io_spa, bp, zio->io_txg, B_FALSE); |
789 | 1687 |
|
1688 |
BP_ZERO(bp); |
|
1689 |
||
1690 |
zio_next_stage(zio); |
|
1691 |
} |
|
1692 |
||
1693 |
static void |
|
1694 |
zio_dva_claim(zio_t *zio) |
|
1695 |
{ |
|
1807
35c8b566d7af
6410711 intent log blocks don't get invited to pool parties
bonwick
parents:
1775
diff
changeset
|
1696 |
zio->io_error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg); |
789 | 1697 |
|
1698 |
zio_next_stage(zio); |
|
1699 |
} |
|
1700 |
||
1701 |
/* |
|
1702 |
* ========================================================================== |
|
1703 |
* Read and write to physical devices |
|
1704 |
* ========================================================================== |
|
1705 |
*/ |
|
1706 |
||
1707 |
static void |
|
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1708 |
zio_vdev_io_start(zio_t *zio) |
789 | 1709 |
{ |
1710 |
vdev_t *vd = zio->io_vd; |
|
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1711 |
vdev_t *tvd = vd ? vd->vdev_top : NULL; |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1712 |
blkptr_t *bp = zio->io_bp; |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1713 |
uint64_t align; |
5329 | 1714 |
spa_t *spa = zio->io_spa; |
1715 |
||
1716 |
/* |
|
1717 |
* If the pool is already in a failure state then just suspend |
|
1718 |
* this IO until the problem is resolved. We will reissue them |
|
1719 |
* at that time. |
|
1720 |
*/ |
|
1721 |
if (spa_state(spa) == POOL_STATE_IO_FAILURE && |
|
1722 |
zio->io_type == ZIO_TYPE_WRITE) { |
|
1723 |
zio_vdev_suspend_io(zio); |
|
1724 |
return; |
|
1725 |
} |
|
789 | 1726 |
|
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1727 |
if (vd == NULL) { |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1728 |
/* The mirror_ops handle multiple DVAs in a single BP */ |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1729 |
vdev_mirror_ops.vdev_op_io_start(zio); |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1730 |
return; |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1731 |
} |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1732 |
|
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1733 |
align = 1ULL << tvd->vdev_ashift; |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1734 |
|
1732 | 1735 |
if (zio->io_retries == 0 && vd == tvd) |
789 | 1736 |
zio->io_flags |= ZIO_FLAG_FAILFAST; |
1737 |
||
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1738 |
if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) && |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1739 |
vd->vdev_children == 0) { |
789 | 1740 |
zio->io_flags |= ZIO_FLAG_PHYSICAL; |
1741 |
zio->io_offset += VDEV_LABEL_START_SIZE; |
|
1742 |
} |
|
1743 |
||
1732 | 1744 |
if (P2PHASE(zio->io_size, align) != 0) { |
1745 |
uint64_t asize = P2ROUNDUP(zio->io_size, align); |
|
1746 |
char *abuf = zio_buf_alloc(asize); |
|
1747 |
ASSERT(vd == tvd); |
|
1748 |
if (zio->io_type == ZIO_TYPE_WRITE) { |
|
1749 |
bcopy(zio->io_data, abuf, zio->io_size); |
|
1750 |
bzero(abuf + zio->io_size, asize - zio->io_size); |
|
1751 |
} |
|
1752 |
zio_push_transform(zio, abuf, asize, asize); |
|
1753 |
ASSERT(!(zio->io_flags & ZIO_FLAG_SUBBLOCK)); |
|
1754 |
zio->io_flags |= ZIO_FLAG_SUBBLOCK; |
|
1755 |
} |
|
1756 |
||
1757 |
ASSERT(P2PHASE(zio->io_offset, align) == 0); |
|
1758 |
ASSERT(P2PHASE(zio->io_size, align) == 0); |
|
1759 |
ASSERT(bp == NULL || |
|
1760 |
P2ROUNDUP(ZIO_GET_IOSIZE(zio), align) == zio->io_size); |
|
789 | 1761 |
ASSERT(zio->io_type != ZIO_TYPE_WRITE || (spa_mode & FWRITE)); |
1762 |
||
1763 |
vdev_io_start(zio); |
|
1764 |
||
1765 |
/* zio_next_stage_async() gets called from io completion interrupt */ |
|
1766 |
} |
|
1767 |
||
1768 |
static void |
|
1769 |
zio_vdev_io_done(zio_t *zio) |
|
1770 |
{ |
|
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1771 |
if (zio->io_vd == NULL) |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1772 |
/* The mirror_ops handle multiple DVAs in a single BP */ |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1773 |
vdev_mirror_ops.vdev_op_io_done(zio); |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1774 |
else |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1775 |
vdev_io_done(zio); |
789 | 1776 |
} |
1777 |
||
1778 |
/* XXPOLICY */ |
|
1544 | 1779 |
boolean_t |
789 | 1780 |
zio_should_retry(zio_t *zio) |
1781 |
{ |
|
1782 |
vdev_t *vd = zio->io_vd; |
|
1783 |
||
1784 |
if (zio->io_error == 0) |
|
1785 |
return (B_FALSE); |
|
1786 |
if (zio->io_delegate_list != NULL) |
|
1787 |
return (B_FALSE); |
|
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1788 |
if (vd && vd != vd->vdev_top) |
789 | 1789 |
return (B_FALSE); |
1790 |
if (zio->io_flags & ZIO_FLAG_DONT_RETRY) |
|
1791 |
return (B_FALSE); |
|
1544 | 1792 |
if (zio->io_retries > 0) |
789 | 1793 |
return (B_FALSE); |
1794 |
||
1795 |
return (B_TRUE); |
|
1796 |
} |
|
1797 |
||
1798 |
static void |
|
1799 |
zio_vdev_io_assess(zio_t *zio) |
|
1800 |
{ |
|
1801 |
vdev_t *vd = zio->io_vd; |
|
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1802 |
vdev_t *tvd = vd ? vd->vdev_top : NULL; |
789 | 1803 |
|
1544 | 1804 |
ASSERT(zio->io_vsd == NULL); |
789 | 1805 |
|
1732 | 1806 |
if (zio->io_flags & ZIO_FLAG_SUBBLOCK) { |
1807 |
void *abuf; |
|
1808 |
uint64_t asize; |
|
1809 |
ASSERT(vd == tvd); |
|
1810 |
zio_pop_transform(zio, &abuf, &asize, &asize); |
|
1811 |
if (zio->io_type == ZIO_TYPE_READ) |
|
1812 |
bcopy(abuf, zio->io_data, zio->io_size); |
|
1813 |
zio_buf_free(abuf, asize); |
|
1814 |
zio->io_flags &= ~ZIO_FLAG_SUBBLOCK; |
|
1815 |
} |
|
1816 |
||
1544 | 1817 |
if (zio_injection_enabled && !zio->io_error) |
1818 |
zio->io_error = zio_handle_fault_injection(zio, EIO); |
|
789 | 1819 |
|
1820 |
/* |
|
1821 |
* If the I/O failed, determine whether we should attempt to retry it. |
|
1822 |
*/ |
|
1823 |
/* XXPOLICY */ |
|
1824 |
if (zio_should_retry(zio)) { |
|
1825 |
ASSERT(tvd == vd); |
|
1826 |
||
1827 |
zio->io_retries++; |
|
1828 |
zio->io_error = 0; |
|
3463 | 1829 |
zio->io_flags &= ZIO_FLAG_VDEV_INHERIT | |
1830 |
ZIO_FLAG_CONFIG_GRABBED; |
|
789 | 1831 |
/* XXPOLICY */ |
1832 |
zio->io_flags &= ~ZIO_FLAG_FAILFAST; |
|
1833 |
zio->io_flags |= ZIO_FLAG_DONT_CACHE; |
|
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1834 |
zio->io_stage = ZIO_STAGE_VDEV_IO_START - 1; |
789 | 1835 |
|
1836 |
dprintf("retry #%d for %s to %s offset %llx\n", |
|
1837 |
zio->io_retries, zio_type_name[zio->io_type], |
|
1838 |
vdev_description(vd), zio->io_offset); |
|
1839 |
||
1544 | 1840 |
zio_next_stage_async(zio); |
1841 |
return; |
|
1842 |
} |
|
789 | 1843 |
|
1844 |
zio_next_stage(zio); |
|
1845 |
} |
|
1846 |
||
1847 |
void |
|
1848 |
zio_vdev_io_reissue(zio_t *zio) |
|
1849 |
{ |
|
1850 |
ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); |
|
1851 |
ASSERT(zio->io_error == 0); |
|
1852 |
||
1853 |
zio->io_stage--; |
|
1854 |
} |
|
1855 |
||
1856 |
void |
|
1857 |
zio_vdev_io_redone(zio_t *zio) |
|
1858 |
{ |
|
1859 |
ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE); |
|
1860 |
||
1861 |
zio->io_stage--; |
|
1862 |
} |
|
1863 |
||
1864 |
void |
|
1865 |
zio_vdev_io_bypass(zio_t *zio) |
|
1866 |
{ |
|
1867 |
ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); |
|
1868 |
ASSERT(zio->io_error == 0); |
|
1869 |
||
1870 |
zio->io_flags |= ZIO_FLAG_IO_BYPASS; |
|
1871 |
zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS - 1; |
|
1872 |
} |
|
1873 |
||
1874 |
/* |
|
1875 |
* ========================================================================== |
|
1876 |
* Generate and verify checksums |
|
1877 |
* ========================================================================== |
|
1878 |
*/ |
|
1879 |
static void |
|
1880 |
zio_checksum_generate(zio_t *zio) |
|
1881 |
{ |
|
1882 |
int checksum = zio->io_checksum; |
|
1883 |
blkptr_t *bp = zio->io_bp; |
|
1884 |
||
1885 |
ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); |
|
1886 |
||
1887 |
BP_SET_CHECKSUM(bp, checksum); |
|
1888 |
BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); |
|
1889 |
||
1890 |
zio_checksum(checksum, &bp->blk_cksum, zio->io_data, zio->io_size); |
|
1891 |
||
1892 |
zio_next_stage(zio); |
|
1893 |
} |
|
1894 |
||
1895 |
static void |
|
1896 |
zio_gang_checksum_generate(zio_t *zio) |
|
1897 |
{ |
|
1898 |
zio_cksum_t zc; |
|
1899 |
zio_gbh_phys_t *gbh = zio->io_data; |
|
1900 |
||
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1901 |
ASSERT(BP_IS_GANG(zio->io_bp)); |
789 | 1902 |
ASSERT3U(zio->io_size, ==, SPA_GANGBLOCKSIZE); |
1903 |
||
1904 |
zio_set_gang_verifier(zio, &gbh->zg_tail.zbt_cksum); |
|
1905 |
||
1906 |
zio_checksum(ZIO_CHECKSUM_GANG_HEADER, &zc, zio->io_data, zio->io_size); |
|
1907 |
||
1908 |
zio_next_stage(zio); |
|
1909 |
} |
|
1910 |
||
1911 |
static void |
|
1912 |
zio_checksum_verify(zio_t *zio) |
|
1913 |
{ |
|
1914 |
if (zio->io_bp != NULL) { |
|
1915 |
zio->io_error = zio_checksum_error(zio); |
|
1544 | 1916 |
if (zio->io_error && !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) |
1917 |
zfs_ereport_post(FM_EREPORT_ZFS_CHECKSUM, |
|
1918 |
zio->io_spa, zio->io_vd, zio, 0, 0); |
|
789 | 1919 |
} |
1920 |
||
1921 |
zio_next_stage(zio); |
|
1922 |
} |
|
1923 |
||
1924 |
/* |
|
1925 |
* Called by RAID-Z to ensure we don't compute the checksum twice. |
|
1926 |
*/ |
|
1927 |
void |
|
1928 |
zio_checksum_verified(zio_t *zio) |
|
1929 |
{ |
|
1930 |
zio->io_pipeline &= ~(1U << ZIO_STAGE_CHECKSUM_VERIFY); |
|
1931 |
} |
|
1932 |
||
1933 |
/* |
|
1934 |
* Set the external verifier for a gang block based on stuff in the bp |
|
1935 |
*/ |
|
1936 |
void |
|
1937 |
zio_set_gang_verifier(zio_t *zio, zio_cksum_t *zcp) |
|
1938 |
{ |
|
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1939 |
blkptr_t *bp = zio->io_bp; |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1940 |
|
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1941 |
zcp->zc_word[0] = DVA_GET_VDEV(BP_IDENTITY(bp)); |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1942 |
zcp->zc_word[1] = DVA_GET_OFFSET(BP_IDENTITY(bp)); |
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
1943 |
zcp->zc_word[2] = bp->blk_birth; |
789 | 1944 |
zcp->zc_word[3] = 0; |
1945 |
} |
|
1946 |
||
1947 |
/* |
|
1948 |
* ========================================================================== |
|
1949 |
* Define the pipeline |
|
1950 |
* ========================================================================== |
|
1951 |
*/ |
|
1952 |
typedef void zio_pipe_stage_t(zio_t *zio); |
|
1953 |
||
1954 |
static void |
|
1955 |
zio_badop(zio_t *zio) |
|
1956 |
{ |
|
1957 |
panic("Invalid I/O pipeline stage %u for zio %p", zio->io_stage, zio); |
|
1958 |
} |
|
1959 |
||
1960 |
zio_pipe_stage_t *zio_pipeline[ZIO_STAGE_DONE + 2] = { |
|
1961 |
zio_badop, |
|
1962 |
zio_wait_children_ready, |
|
1963 |
zio_write_compress, |
|
1964 |
zio_checksum_generate, |
|
1965 |
zio_gang_pipeline, |
|
1966 |
zio_get_gang_header, |
|
1967 |
zio_rewrite_gang_members, |
|
1968 |
zio_free_gang_members, |
|
1969 |
zio_claim_gang_members, |
|
1970 |
zio_dva_allocate, |
|
1971 |
zio_dva_free, |
|
1972 |
zio_dva_claim, |
|
1973 |
zio_gang_checksum_generate, |
|
1974 |
zio_ready, |
|
5329 | 1975 |
zio_read_init, |
789 | 1976 |
zio_vdev_io_start, |
1977 |
zio_vdev_io_done, |
|
1978 |
zio_vdev_io_assess, |
|
1979 |
zio_wait_children_done, |
|
1980 |
zio_checksum_verify, |
|
1981 |
zio_read_gang_members, |
|
1982 |
zio_read_decompress, |
|
5329 | 1983 |
zio_assess, |
789 | 1984 |
zio_done, |
1985 |
zio_badop |
|
1986 |
}; |
|
1987 |
||
1988 |
/* |
|
1989 |
* Move an I/O to the next stage of the pipeline and execute that stage. |
|
1990 |
* There's no locking on io_stage because there's no legitimate way for |
|
1991 |
* multiple threads to be attempting to process the same I/O. |
|
1992 |
*/ |
|
1993 |
void |
|
1994 |
zio_next_stage(zio_t *zio) |
|
1995 |
{ |
|
1996 |
uint32_t pipeline = zio->io_pipeline; |
|
1997 |
||
1998 |
ASSERT(!MUTEX_HELD(&zio->io_lock)); |
|
1999 |
||
2000 |
if (zio->io_error) { |
|
2001 |
dprintf("zio %p vdev %s offset %llx stage %d error %d\n", |
|
2002 |
zio, vdev_description(zio->io_vd), |
|
2003 |
zio->io_offset, zio->io_stage, zio->io_error); |
|
2004 |
if (((1U << zio->io_stage) & ZIO_VDEV_IO_PIPELINE) == 0) |
|
2005 |
pipeline &= ZIO_ERROR_PIPELINE_MASK; |
|
2006 |
} |
|
2007 |
||
2008 |
while (((1U << ++zio->io_stage) & pipeline) == 0) |
|
2009 |
continue; |
|
2010 |
||
2011 |
ASSERT(zio->io_stage <= ZIO_STAGE_DONE); |
|
2012 |
ASSERT(zio->io_stalled == 0); |
|
2013 |
||
3689 | 2014 |
/* |
2015 |
* See the comment in zio_next_stage_async() about per-CPU taskqs. |
|
2016 |
*/ |
|
2017 |
if (((1U << zio->io_stage) & zio->io_async_stages) && |
|
2018 |
(zio->io_stage == ZIO_STAGE_WRITE_COMPRESS) && |
|
2019 |
!(zio->io_flags & ZIO_FLAG_METADATA)) { |
|
2020 |
taskq_t *tq = zio->io_spa->spa_zio_issue_taskq[zio->io_type]; |
|
2021 |
(void) taskq_dispatch(tq, |
|
2022 |
(task_func_t *)zio_pipeline[zio->io_stage], zio, TQ_SLEEP); |
|
2023 |
} else { |
|
2024 |
zio_pipeline[zio->io_stage](zio); |
|
2025 |
} |
|
789 | 2026 |
} |
2027 |
||
2028 |
void |
|
2029 |
zio_next_stage_async(zio_t *zio) |
|
2030 |
{ |
|
2031 |
taskq_t *tq; |
|
2032 |
uint32_t pipeline = zio->io_pipeline; |
|
2033 |
||
2034 |
ASSERT(!MUTEX_HELD(&zio->io_lock)); |
|
2035 |
||
2036 |
if (zio->io_error) { |
|
2037 |
dprintf("zio %p vdev %s offset %llx stage %d error %d\n", |
|
2038 |
zio, vdev_description(zio->io_vd), |
|
2039 |
zio->io_offset, zio->io_stage, zio->io_error); |
|
2040 |
if (((1U << zio->io_stage) & ZIO_VDEV_IO_PIPELINE) == 0) |
|
2041 |
pipeline &= ZIO_ERROR_PIPELINE_MASK; |
|
2042 |
} |
|
2043 |
||
2044 |
while (((1U << ++zio->io_stage) & pipeline) == 0) |
|
2045 |
continue; |
|
2046 |
||
2047 |
ASSERT(zio->io_stage <= ZIO_STAGE_DONE); |
|
2048 |
ASSERT(zio->io_stalled == 0); |
|
2049 |
||
2050 |
/* |
|
2051 |
* For performance, we'll probably want two sets of task queues: |
|
2052 |
* per-CPU issue taskqs and per-CPU completion taskqs. The per-CPU |
|
2053 |
* part is for read performance: since we have to make a pass over |
|
2054 |
* the data to checksum it anyway, we want to do this on the same CPU |
|
2055 |
* that issued the read, because (assuming CPU scheduling affinity) |
|
2056 |
* that thread is probably still there. Getting this optimization |
|
2057 |
* right avoids performance-hostile cache-to-cache transfers. |
|
2058 |
* |
|
2059 |
* Note that having two sets of task queues is also necessary for |
|
2060 |
* correctness: if all of the issue threads get bogged down waiting |
|
2061 |
* for dependent reads (e.g. metaslab freelist) to complete, then |
|
2062 |
* there won't be any threads available to service I/O completion |
|
2063 |
* interrupts. |
|
2064 |
*/ |
|
2065 |
if ((1U << zio->io_stage) & zio->io_async_stages) { |
|
2066 |
if (zio->io_stage < ZIO_STAGE_VDEV_IO_DONE) |
|
2067 |
tq = zio->io_spa->spa_zio_issue_taskq[zio->io_type]; |
|
2068 |
else |
|
2069 |
tq = zio->io_spa->spa_zio_intr_taskq[zio->io_type]; |
|
2070 |
(void) taskq_dispatch(tq, |
|
2071 |
(task_func_t *)zio_pipeline[zio->io_stage], zio, TQ_SLEEP); |
|
2072 |
} else { |
|
2073 |
zio_pipeline[zio->io_stage](zio); |
|
2074 |
} |
|
2075 |
} |
|
2076 |
||
5329 | 2077 |
void |
2078 |
zio_resubmit_stage_async(void *arg) |
|
2079 |
{ |
|
2080 |
zio_t *zio = (zio_t *)(uintptr_t)arg; |
|
2081 |
||
2082 |
zio_next_stage_async(zio); |
|
2083 |
} |
|
2084 |
||
3668 | 2085 |
static boolean_t |
5329 | 2086 |
zio_io_should_fail(uint16_t range) |
3668 | 2087 |
{ |
2088 |
static uint16_t allocs = 0; |
|
2089 |
||
5329 | 2090 |
return (P2PHASE(allocs++, 1U<<range) == 0); |
3668 | 2091 |
} |
2092 |
||
789 | 2093 |
/* |
2094 |
* Try to allocate an intent log block. Return 0 on success, errno on failure. |
|
2095 |
*/ |
|
2096 |
int |
|
3063
b252896b372b
6341569 zio_alloc_blk() vdev distribution performs badly
perrin
parents:
2981
diff
changeset
|
2097 |
zio_alloc_blk(spa_t *spa, uint64_t size, blkptr_t *new_bp, blkptr_t *old_bp, |
b252896b372b
6341569 zio_alloc_blk() vdev distribution performs badly
perrin
parents:
2981
diff
changeset
|
2098 |
uint64_t txg) |
789 | 2099 |
{ |
2100 |
int error; |
|
2101 |
||
1544 | 2102 |
spa_config_enter(spa, RW_READER, FTAG); |
789 | 2103 |
|
5329 | 2104 |
if (zio_zil_fail_shift && zio_io_should_fail(zio_zil_fail_shift)) { |
3668 | 2105 |
spa_config_exit(spa, FTAG); |
2106 |
return (ENOSPC); |
|
2107 |
} |
|
2108 |
||
3063
b252896b372b
6341569 zio_alloc_blk() vdev distribution performs badly
perrin
parents:
2981
diff
changeset
|
2109 |
/* |
4527 | 2110 |
* We were passed the previous log block's DVA in bp->blk_dva[0]. |
2111 |
* We use that as a hint for which vdev to allocate from next. |
|
3063
b252896b372b
6341569 zio_alloc_blk() vdev distribution performs badly
perrin
parents:
2981
diff
changeset
|
2112 |
*/ |
4527 | 2113 |
error = metaslab_alloc(spa, spa->spa_log_class, size, |
2114 |
new_bp, 1, txg, old_bp, B_TRUE); |
|
2115 |
||
2116 |
if (error) |
|
2117 |
error = metaslab_alloc(spa, spa->spa_normal_class, size, |
|
2118 |
new_bp, 1, txg, old_bp, B_TRUE); |
|
789 | 2119 |
|
2120 |
if (error == 0) { |
|
3063
b252896b372b
6341569 zio_alloc_blk() vdev distribution performs badly
perrin
parents:
2981
diff
changeset
|
2121 |
BP_SET_LSIZE(new_bp, size); |
b252896b372b
6341569 zio_alloc_blk() vdev distribution performs badly
perrin
parents:
2981
diff
changeset
|
2122 |
BP_SET_PSIZE(new_bp, size); |
b252896b372b
6341569 zio_alloc_blk() vdev distribution performs badly
perrin
parents:
2981
diff
changeset
|
2123 |
BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF); |
b252896b372b
6341569 zio_alloc_blk() vdev distribution performs badly
perrin
parents:
2981
diff
changeset
|
2124 |
BP_SET_CHECKSUM(new_bp, ZIO_CHECKSUM_ZILOG); |
b252896b372b
6341569 zio_alloc_blk() vdev distribution performs badly
perrin
parents:
2981
diff
changeset
|
2125 |
BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); |
b252896b372b
6341569 zio_alloc_blk() vdev distribution performs badly
perrin
parents:
2981
diff
changeset
|
2126 |
BP_SET_LEVEL(new_bp, 0); |
b252896b372b
6341569 zio_alloc_blk() vdev distribution performs badly
perrin
parents:
2981
diff
changeset
|
2127 |
BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER); |
b252896b372b
6341569 zio_alloc_blk() vdev distribution performs badly
perrin
parents:
2981
diff
changeset
|
2128 |
new_bp->blk_birth = txg; |
789 | 2129 |
} |
2130 |
||
1544 | 2131 |
spa_config_exit(spa, FTAG); |
789 | 2132 |
|
2133 |
return (error); |
|
2134 |
} |
|
2135 |
||
2136 |
/* |
|
2137 |
* Free an intent log block. We know it can't be a gang block, so there's |
|
2138 |
* nothing to do except metaslab_free() it. |
|
2139 |
*/ |
|
2140 |
void |
|
2141 |
zio_free_blk(spa_t *spa, blkptr_t *bp, uint64_t txg) |
|
2142 |
{ |
|
1775
e51e26b432c0
6410698 ZFS metadata needs to be more highly replicated (ditto blocks)
billm
parents:
1732
diff
changeset
|
2143 |
ASSERT(!BP_IS_GANG(bp)); |
789 | 2144 |
|
1544 | 2145 |
spa_config_enter(spa, RW_READER, FTAG); |
789 | 2146 |
|
1807
35c8b566d7af
6410711 intent log blocks don't get invited to pool parties
bonwick
parents:
1775
diff
changeset
|
2147 |
metaslab_free(spa, bp, txg, B_FALSE); |
789 | 2148 |
|
1544 | 2149 |
spa_config_exit(spa, FTAG); |
789 | 2150 |
} |
4469
7d15cd6c3854
6566482 panic replacing a pool device whilst intent log being used
perrin
parents:
4451
diff
changeset
|
2151 |
|
7d15cd6c3854
6566482 panic replacing a pool device whilst intent log being used
perrin
parents:
4451
diff
changeset
|
2152 |
/* |
7d15cd6c3854
6566482 panic replacing a pool device whilst intent log being used
perrin
parents:
4451
diff
changeset
|
2153 |
* start an async flush of the write cache for this vdev |
7d15cd6c3854
6566482 panic replacing a pool device whilst intent log being used
perrin
parents:
4451
diff
changeset
|
2154 |
*/ |
7d15cd6c3854
6566482 panic replacing a pool device whilst intent log being used
perrin
parents:
4451
diff
changeset
|
2155 |
void |
7d15cd6c3854
6566482 panic replacing a pool device whilst intent log being used
perrin
parents:
4451
diff
changeset
|
2156 |
zio_flush_vdev(spa_t *spa, uint64_t vdev, zio_t **zio) |
7d15cd6c3854
6566482 panic replacing a pool device whilst intent log being used
perrin
parents:
4451
diff
changeset
|
2157 |
{ |
7d15cd6c3854
6566482 panic replacing a pool device whilst intent log being used
perrin
parents:
4451
diff
changeset
|
2158 |
vdev_t *vd; |
7d15cd6c3854
6566482 panic replacing a pool device whilst intent log being used
perrin
parents:
4451
diff
changeset
|
2159 |
|
7d15cd6c3854
6566482 panic replacing a pool device whilst intent log being used
perrin
parents:
4451
diff
changeset
|
2160 |
/* |
7d15cd6c3854
6566482 panic replacing a pool device whilst intent log being used
perrin
parents:
4451
diff
changeset
|
2161 |
* Lock out configuration changes. |
7d15cd6c3854
6566482 panic replacing a pool device whilst intent log being used
perrin
parents:
4451
diff
changeset
|
2162 |
*/ |
7d15cd6c3854
6566482 panic replacing a pool device whilst intent log being used
perrin
parents:
4451
diff
changeset
|
2163 |
spa_config_enter(spa, RW_READER, FTAG); |
7d15cd6c3854
6566482 panic replacing a pool device whilst intent log being used
perrin
parents:
4451
diff
changeset
|
2164 |
|
7d15cd6c3854
6566482 panic replacing a pool device whilst intent log being used
perrin
parents:
4451
diff
changeset
|
2165 |
if (*zio == NULL) |
7d15cd6c3854
6566482 panic replacing a pool device whilst intent log being used
perrin
parents:
4451
diff
changeset
|
2166 |
*zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); |
7d15cd6c3854
6566482 panic replacing a pool device whilst intent log being used
perrin
parents:
4451
diff
changeset
|
2167 |
|
7d15cd6c3854
6566482 panic replacing a pool device whilst intent log being used
perrin
parents:
4451
diff
changeset
|
2168 |
vd = vdev_lookup_top(spa, vdev); |
7d15cd6c3854
6566482 panic replacing a pool device whilst intent log being used
perrin
parents:
4451
diff
changeset
|
2169 |
ASSERT(vd); |
7d15cd6c3854
6566482 panic replacing a pool device whilst intent log being used
perrin
parents:
4451
diff
changeset
|
2170 |
|
7d15cd6c3854
6566482 panic replacing a pool device whilst intent log being used
perrin
parents:
4451
diff
changeset
|
2171 |
(void) zio_nowait(zio_ioctl(*zio, spa, vd, DKIOCFLUSHWRITECACHE, |
7d15cd6c3854
6566482 panic replacing a pool device whilst intent log being used
perrin
parents:
4451
diff
changeset
|
2172 |
NULL, NULL, ZIO_PRIORITY_NOW, |
7d15cd6c3854
6566482 panic replacing a pool device whilst intent log being used
perrin
parents:
4451
diff
changeset
|
2173 |
ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY)); |
7d15cd6c3854
6566482 panic replacing a pool device whilst intent log being used
perrin
parents:
4451
diff
changeset
|
2174 |
|
7d15cd6c3854
6566482 panic replacing a pool device whilst intent log being used
perrin
parents:
4451
diff
changeset
|
2175 |
spa_config_exit(spa, FTAG); |
7d15cd6c3854
6566482 panic replacing a pool device whilst intent log being used
perrin
parents:
4451
diff
changeset
|
2176 |
} |