]>
Commit | Line | Data |
---|---|---|
34dc7c2f BB |
1 | /* |
2 | * CDDL HEADER START | |
3 | * | |
4 | * The contents of this file are subject to the terms of the | |
5 | * Common Development and Distribution License (the "License"). | |
6 | * You may not use this file except in compliance with the License. | |
7 | * | |
8 | * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE | |
9 | * or http://www.opensolaris.org/os/licensing. | |
10 | * See the License for the specific language governing permissions | |
11 | * and limitations under the License. | |
12 | * | |
13 | * When distributing Covered Code, include this CDDL HEADER in each | |
14 | * file and include the License file at usr/src/OPENSOLARIS.LICENSE. | |
15 | * If applicable, add the following below this CDDL HEADER, with the | |
16 | * fields enclosed by brackets "[]" replaced with your own identifying | |
17 | * information: Portions Copyright [yyyy] [name of copyright owner] | |
18 | * | |
19 | * CDDL HEADER END | |
20 | */ | |
21 | /* | |
428870ff | 22 | * Copyright 2010 Sun Microsystems, Inc. All rights reserved. |
34dc7c2f BB |
23 | * Use is subject to license terms. |
24 | */ | |
25 | ||
1bd201e7 | 26 | /* |
3dfb57a3 | 27 | * Copyright (c) 2012, 2015 by Delphix. All rights reserved. |
1bd201e7 CS |
28 | */ |
29 | ||
34dc7c2f BB |
30 | #include <sys/zfs_context.h> |
31 | #include <sys/spa.h> | |
32 | #include <sys/vdev_impl.h> | |
33 | #include <sys/zio.h> | |
34 | #include <sys/fs/zfs.h> | |
35 | ||
36 | /* | |
37 | * Virtual device vector for mirroring. | |
38 | */ | |
39 | ||
40 | typedef struct mirror_child { | |
41 | vdev_t *mc_vd; | |
42 | uint64_t mc_offset; | |
43 | int mc_error; | |
9f500936 | 44 | int mc_load; |
b128c09f BB |
45 | uint8_t mc_tried; |
46 | uint8_t mc_skipped; | |
47 | uint8_t mc_speculative; | |
34dc7c2f BB |
48 | } mirror_child_t; |
49 | ||
50 | typedef struct mirror_map { | |
9f500936 | 51 | int *mm_preferred; |
52 | int mm_preferred_cnt; | |
34dc7c2f | 53 | int mm_children; |
9f500936 | 54 | boolean_t mm_replacing; |
55 | boolean_t mm_root; | |
56 | mirror_child_t mm_child[]; | |
34dc7c2f BB |
57 | } mirror_map_t; |
58 | ||
9f500936 | 59 | static int vdev_mirror_shift = 21; |
60 | ||
556011db | 61 | /* |
9f500936 | 62 | * The load configuration settings below are tuned by default for |
63 | * the case where all devices are of the same rotational type. | |
556011db | 64 | * |
9f500936 | 65 | * If there is a mixture of rotating and non-rotating media, setting |
66 | * zfs_vdev_mirror_non_rotating_seek_inc to 0 may well provide better results | |
67 | * as it will direct more reads to the non-rotating vdevs which are more likely | |
68 | * to have a higher performance. | |
556011db | 69 | */ |
9f500936 | 70 | |
71 | /* Rotating media load calculation configuration. */ | |
72 | static int zfs_vdev_mirror_rotating_inc = 0; | |
73 | static int zfs_vdev_mirror_rotating_seek_inc = 5; | |
74 | static int zfs_vdev_mirror_rotating_seek_offset = 1 * 1024 * 1024; | |
75 | ||
76 | /* Non-rotating media load calculation configuration. */ | |
77 | static int zfs_vdev_mirror_non_rotating_inc = 0; | |
78 | static int zfs_vdev_mirror_non_rotating_seek_inc = 1; | |
79 | ||
80 | static inline size_t | |
81 | vdev_mirror_map_size(int children) | |
82 | { | |
83 | return (offsetof(mirror_map_t, mm_child[children]) + | |
84 | sizeof (int) * children); | |
85 | } | |
86 | ||
87 | static inline mirror_map_t * | |
88 | vdev_mirror_map_alloc(int children, boolean_t replacing, boolean_t root) | |
89 | { | |
90 | mirror_map_t *mm; | |
91 | ||
92 | mm = kmem_zalloc(vdev_mirror_map_size(children), KM_SLEEP); | |
93 | mm->mm_children = children; | |
94 | mm->mm_replacing = replacing; | |
95 | mm->mm_root = root; | |
96 | mm->mm_preferred = (int *)((uintptr_t)mm + | |
97 | offsetof(mirror_map_t, mm_child[children])); | |
98 | ||
99 | return (mm); | |
100 | } | |
34dc7c2f | 101 | |
b128c09f BB |
102 | static void |
103 | vdev_mirror_map_free(zio_t *zio) | |
104 | { | |
105 | mirror_map_t *mm = zio->io_vsd; | |
106 | ||
9f500936 | 107 | kmem_free(mm, vdev_mirror_map_size(mm->mm_children)); |
b128c09f BB |
108 | } |
109 | ||
428870ff BB |
110 | static const zio_vsd_ops_t vdev_mirror_vsd_ops = { |
111 | vdev_mirror_map_free, | |
112 | zio_vsd_default_cksum_report | |
113 | }; | |
114 | ||
556011db | 115 | static int |
9f500936 | 116 | vdev_mirror_load(mirror_map_t *mm, vdev_t *vd, uint64_t zio_offset) |
556011db | 117 | { |
9f500936 | 118 | uint64_t lastoffset; |
119 | int load; | |
120 | ||
121 | /* All DVAs have equal weight at the root. */ | |
122 | if (mm->mm_root) | |
123 | return (INT_MAX); | |
124 | ||
125 | /* | |
126 | * We don't return INT_MAX if the device is resilvering i.e. | |
127 | * vdev_resilver_txg != 0 as when tested performance was slightly | |
128 | * worse overall when resilvering with compared to without. | |
129 | */ | |
130 | ||
131 | /* Standard load based on pending queue length. */ | |
132 | load = vdev_queue_length(vd); | |
133 | lastoffset = vdev_queue_lastoffset(vd); | |
134 | ||
135 | if (vd->vdev_nonrot) { | |
136 | /* Non-rotating media. */ | |
137 | if (lastoffset == zio_offset) | |
138 | return (load + zfs_vdev_mirror_non_rotating_inc); | |
139 | ||
140 | /* | |
141 | * Apply a seek penalty even for non-rotating devices as | |
142 | * sequential I/O's can be aggregated into fewer operations on | |
143 | * the device, thus avoiding unnecessary per-command overhead | |
144 | * and boosting performance. | |
145 | */ | |
146 | return (load + zfs_vdev_mirror_non_rotating_seek_inc); | |
147 | } | |
148 | ||
149 | /* Rotating media I/O's which directly follow the last I/O. */ | |
150 | if (lastoffset == zio_offset) | |
151 | return (load + zfs_vdev_mirror_rotating_inc); | |
152 | ||
153 | /* | |
154 | * Apply half the seek increment to I/O's within seek offset | |
155 | * of the last I/O queued to this vdev as they should incure less | |
156 | * of a seek increment. | |
157 | */ | |
158 | if (ABS(lastoffset - zio_offset) < | |
159 | zfs_vdev_mirror_rotating_seek_offset) | |
160 | return (load + (zfs_vdev_mirror_rotating_seek_inc / 2)); | |
161 | ||
162 | /* Apply the full seek increment to all other I/O's. */ | |
163 | return (load + zfs_vdev_mirror_rotating_seek_inc); | |
556011db BB |
164 | } |
165 | ||
a1687880 BB |
166 | /* |
167 | * Avoid inlining the function to keep vdev_mirror_io_start(), which | |
168 | * is this functions only caller, as small as possible on the stack. | |
169 | */ | |
170 | noinline static mirror_map_t * | |
9f500936 | 171 | vdev_mirror_map_init(zio_t *zio) |
34dc7c2f BB |
172 | { |
173 | mirror_map_t *mm = NULL; | |
174 | mirror_child_t *mc; | |
175 | vdev_t *vd = zio->io_vd; | |
9f500936 | 176 | int c; |
34dc7c2f BB |
177 | |
178 | if (vd == NULL) { | |
179 | dva_t *dva = zio->io_bp->blk_dva; | |
180 | spa_t *spa = zio->io_spa; | |
181 | ||
9f500936 | 182 | mm = vdev_mirror_map_alloc(BP_GET_NDVAS(zio->io_bp), B_FALSE, |
183 | B_TRUE); | |
34dc7c2f BB |
184 | for (c = 0; c < mm->mm_children; c++) { |
185 | mc = &mm->mm_child[c]; | |
186 | ||
187 | mc->mc_vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[c])); | |
188 | mc->mc_offset = DVA_GET_OFFSET(&dva[c]); | |
189 | } | |
190 | } else { | |
9f500936 | 191 | mm = vdev_mirror_map_alloc(vd->vdev_children, |
192 | (vd->vdev_ops == &vdev_replacing_ops || | |
193 | vd->vdev_ops == &vdev_spare_ops), B_FALSE); | |
34dc7c2f BB |
194 | for (c = 0; c < mm->mm_children; c++) { |
195 | mc = &mm->mm_child[c]; | |
196 | mc->mc_vd = vd->vdev_child[c]; | |
197 | mc->mc_offset = zio->io_offset; | |
198 | } | |
199 | } | |
200 | ||
201 | zio->io_vsd = mm; | |
428870ff | 202 | zio->io_vsd_ops = &vdev_mirror_vsd_ops; |
34dc7c2f BB |
203 | return (mm); |
204 | } | |
205 | ||
34dc7c2f | 206 | static int |
1bd201e7 CS |
207 | vdev_mirror_open(vdev_t *vd, uint64_t *asize, uint64_t *max_asize, |
208 | uint64_t *ashift) | |
34dc7c2f | 209 | { |
34dc7c2f | 210 | int numerrors = 0; |
45d1cae3 | 211 | int lasterror = 0; |
d6320ddb | 212 | int c; |
34dc7c2f BB |
213 | |
214 | if (vd->vdev_children == 0) { | |
215 | vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL; | |
2e528b49 | 216 | return (SET_ERROR(EINVAL)); |
34dc7c2f BB |
217 | } |
218 | ||
45d1cae3 | 219 | vdev_open_children(vd); |
34dc7c2f | 220 | |
d6320ddb | 221 | for (c = 0; c < vd->vdev_children; c++) { |
45d1cae3 BB |
222 | vdev_t *cvd = vd->vdev_child[c]; |
223 | ||
224 | if (cvd->vdev_open_error) { | |
225 | lasterror = cvd->vdev_open_error; | |
34dc7c2f BB |
226 | numerrors++; |
227 | continue; | |
228 | } | |
229 | ||
230 | *asize = MIN(*asize - 1, cvd->vdev_asize - 1) + 1; | |
1bd201e7 | 231 | *max_asize = MIN(*max_asize - 1, cvd->vdev_max_asize - 1) + 1; |
34dc7c2f BB |
232 | *ashift = MAX(*ashift, cvd->vdev_ashift); |
233 | } | |
234 | ||
235 | if (numerrors == vd->vdev_children) { | |
236 | vd->vdev_stat.vs_aux = VDEV_AUX_NO_REPLICAS; | |
237 | return (lasterror); | |
238 | } | |
239 | ||
240 | return (0); | |
241 | } | |
242 | ||
243 | static void | |
244 | vdev_mirror_close(vdev_t *vd) | |
245 | { | |
d6320ddb BB |
246 | int c; |
247 | ||
248 | for (c = 0; c < vd->vdev_children; c++) | |
34dc7c2f BB |
249 | vdev_close(vd->vdev_child[c]); |
250 | } | |
251 | ||
252 | static void | |
253 | vdev_mirror_child_done(zio_t *zio) | |
254 | { | |
255 | mirror_child_t *mc = zio->io_private; | |
256 | ||
257 | mc->mc_error = zio->io_error; | |
258 | mc->mc_tried = 1; | |
259 | mc->mc_skipped = 0; | |
260 | } | |
261 | ||
262 | static void | |
263 | vdev_mirror_scrub_done(zio_t *zio) | |
264 | { | |
265 | mirror_child_t *mc = zio->io_private; | |
266 | ||
267 | if (zio->io_error == 0) { | |
d164b209 | 268 | zio_t *pio; |
3dfb57a3 | 269 | zio_link_t *zl = NULL; |
d164b209 BB |
270 | |
271 | mutex_enter(&zio->io_lock); | |
3dfb57a3 | 272 | while ((pio = zio_walk_parents(zio, &zl)) != NULL) { |
d164b209 BB |
273 | mutex_enter(&pio->io_lock); |
274 | ASSERT3U(zio->io_size, >=, pio->io_size); | |
275 | bcopy(zio->io_data, pio->io_data, pio->io_size); | |
276 | mutex_exit(&pio->io_lock); | |
277 | } | |
278 | mutex_exit(&zio->io_lock); | |
34dc7c2f BB |
279 | } |
280 | ||
281 | zio_buf_free(zio->io_data, zio->io_size); | |
282 | ||
283 | mc->mc_error = zio->io_error; | |
284 | mc->mc_tried = 1; | |
285 | mc->mc_skipped = 0; | |
286 | } | |
287 | ||
34dc7c2f | 288 | /* |
9f500936 | 289 | * Check the other, lower-index DVAs to see if they're on the same |
290 | * vdev as the child we picked. If they are, use them since they | |
291 | * are likely to have been allocated from the primary metaslab in | |
292 | * use at the time, and hence are more likely to have locality with | |
293 | * single-copy data. | |
294 | */ | |
295 | static int | |
296 | vdev_mirror_dva_select(zio_t *zio, int p) | |
297 | { | |
298 | dva_t *dva = zio->io_bp->blk_dva; | |
299 | mirror_map_t *mm = zio->io_vsd; | |
300 | int preferred; | |
301 | int c; | |
302 | ||
303 | preferred = mm->mm_preferred[p]; | |
304 | for (p--; p >= 0; p--) { | |
305 | c = mm->mm_preferred[p]; | |
306 | if (DVA_GET_VDEV(&dva[c]) == DVA_GET_VDEV(&dva[preferred])) | |
307 | preferred = c; | |
308 | } | |
309 | return (preferred); | |
310 | } | |
311 | ||
312 | static int | |
313 | vdev_mirror_preferred_child_randomize(zio_t *zio) | |
314 | { | |
315 | mirror_map_t *mm = zio->io_vsd; | |
316 | int p; | |
317 | ||
318 | if (mm->mm_root) { | |
319 | p = spa_get_random(mm->mm_preferred_cnt); | |
320 | return (vdev_mirror_dva_select(zio, p)); | |
321 | } | |
322 | ||
323 | /* | |
324 | * To ensure we don't always favour the first matching vdev, | |
325 | * which could lead to wear leveling issues on SSD's, we | |
326 | * use the I/O offset as a pseudo random seed into the vdevs | |
327 | * which have the lowest load. | |
328 | */ | |
329 | p = (zio->io_offset >> vdev_mirror_shift) % mm->mm_preferred_cnt; | |
330 | return (mm->mm_preferred[p]); | |
331 | } | |
332 | ||
333 | /* | |
334 | * Try to find a vdev whose DTL doesn't contain the block we want to read | |
335 | * prefering vdevs based on determined load. | |
336 | * | |
34dc7c2f BB |
337 | * Try to find a child whose DTL doesn't contain the block we want to read. |
338 | * If we can't, try the read on any vdev we haven't already tried. | |
339 | */ | |
340 | static int | |
341 | vdev_mirror_child_select(zio_t *zio) | |
342 | { | |
343 | mirror_map_t *mm = zio->io_vsd; | |
34dc7c2f | 344 | uint64_t txg = zio->io_txg; |
9f500936 | 345 | int c, lowest_load; |
34dc7c2f | 346 | |
428870ff | 347 | ASSERT(zio->io_bp == NULL || BP_PHYSICAL_BIRTH(zio->io_bp) == txg); |
34dc7c2f | 348 | |
9f500936 | 349 | lowest_load = INT_MAX; |
350 | mm->mm_preferred_cnt = 0; | |
351 | for (c = 0; c < mm->mm_children; c++) { | |
352 | mirror_child_t *mc; | |
353 | ||
34dc7c2f BB |
354 | mc = &mm->mm_child[c]; |
355 | if (mc->mc_tried || mc->mc_skipped) | |
356 | continue; | |
9f500936 | 357 | |
33074f22 | 358 | if (mc->mc_vd == NULL || !vdev_readable(mc->mc_vd)) { |
2e528b49 | 359 | mc->mc_error = SET_ERROR(ENXIO); |
34dc7c2f BB |
360 | mc->mc_tried = 1; /* don't even try */ |
361 | mc->mc_skipped = 1; | |
362 | continue; | |
363 | } | |
9f500936 | 364 | |
365 | if (vdev_dtl_contains(mc->mc_vd, DTL_MISSING, txg, 1)) { | |
366 | mc->mc_error = SET_ERROR(ESTALE); | |
367 | mc->mc_skipped = 1; | |
368 | mc->mc_speculative = 1; | |
369 | continue; | |
370 | } | |
371 | ||
372 | mc->mc_load = vdev_mirror_load(mm, mc->mc_vd, mc->mc_offset); | |
373 | if (mc->mc_load > lowest_load) | |
374 | continue; | |
375 | ||
376 | if (mc->mc_load < lowest_load) { | |
377 | lowest_load = mc->mc_load; | |
378 | mm->mm_preferred_cnt = 0; | |
379 | } | |
380 | mm->mm_preferred[mm->mm_preferred_cnt] = c; | |
381 | mm->mm_preferred_cnt++; | |
382 | } | |
383 | ||
384 | if (mm->mm_preferred_cnt == 1) { | |
385 | vdev_queue_register_lastoffset( | |
386 | mm->mm_child[mm->mm_preferred[0]].mc_vd, zio); | |
387 | return (mm->mm_preferred[0]); | |
388 | } | |
389 | ||
390 | if (mm->mm_preferred_cnt > 1) { | |
391 | int c = vdev_mirror_preferred_child_randomize(zio); | |
392 | ||
393 | vdev_queue_register_lastoffset(mm->mm_child[c].mc_vd, zio); | |
394 | return (c); | |
34dc7c2f BB |
395 | } |
396 | ||
397 | /* | |
398 | * Every device is either missing or has this txg in its DTL. | |
399 | * Look for any child we haven't already tried before giving up. | |
400 | */ | |
9f500936 | 401 | for (c = 0; c < mm->mm_children; c++) { |
402 | if (!mm->mm_child[c].mc_tried) { | |
403 | vdev_queue_register_lastoffset(mm->mm_child[c].mc_vd, | |
404 | zio); | |
34dc7c2f | 405 | return (c); |
9f500936 | 406 | } |
407 | } | |
34dc7c2f BB |
408 | |
409 | /* | |
410 | * Every child failed. There's no place left to look. | |
411 | */ | |
412 | return (-1); | |
413 | } | |
414 | ||
98b25418 | 415 | static void |
34dc7c2f BB |
416 | vdev_mirror_io_start(zio_t *zio) |
417 | { | |
418 | mirror_map_t *mm; | |
419 | mirror_child_t *mc; | |
420 | int c, children; | |
421 | ||
9f500936 | 422 | mm = vdev_mirror_map_init(zio); |
34dc7c2f BB |
423 | |
424 | if (zio->io_type == ZIO_TYPE_READ) { | |
425 | if ((zio->io_flags & ZIO_FLAG_SCRUB) && !mm->mm_replacing) { | |
426 | /* | |
427 | * For scrubbing reads we need to allocate a read | |
428 | * buffer for each child and issue reads to all | |
429 | * children. If any child succeeds, it will copy its | |
430 | * data into zio->io_data in vdev_mirror_scrub_done. | |
431 | */ | |
432 | for (c = 0; c < mm->mm_children; c++) { | |
433 | mc = &mm->mm_child[c]; | |
434 | zio_nowait(zio_vdev_child_io(zio, zio->io_bp, | |
435 | mc->mc_vd, mc->mc_offset, | |
436 | zio_buf_alloc(zio->io_size), zio->io_size, | |
b128c09f | 437 | zio->io_type, zio->io_priority, 0, |
34dc7c2f BB |
438 | vdev_mirror_scrub_done, mc)); |
439 | } | |
98b25418 GW |
440 | zio_execute(zio); |
441 | return; | |
34dc7c2f BB |
442 | } |
443 | /* | |
444 | * For normal reads just pick one child. | |
445 | */ | |
446 | c = vdev_mirror_child_select(zio); | |
447 | children = (c >= 0); | |
448 | } else { | |
449 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
450 | ||
451 | /* | |
fb5f0bc8 | 452 | * Writes go to all children. |
34dc7c2f | 453 | */ |
fb5f0bc8 BB |
454 | c = 0; |
455 | children = mm->mm_children; | |
34dc7c2f BB |
456 | } |
457 | ||
458 | while (children--) { | |
459 | mc = &mm->mm_child[c]; | |
460 | zio_nowait(zio_vdev_child_io(zio, zio->io_bp, | |
b128c09f BB |
461 | mc->mc_vd, mc->mc_offset, zio->io_data, zio->io_size, |
462 | zio->io_type, zio->io_priority, 0, | |
463 | vdev_mirror_child_done, mc)); | |
34dc7c2f BB |
464 | c++; |
465 | } | |
466 | ||
98b25418 | 467 | zio_execute(zio); |
34dc7c2f BB |
468 | } |
469 | ||
470 | static int | |
b128c09f BB |
471 | vdev_mirror_worst_error(mirror_map_t *mm) |
472 | { | |
d6320ddb | 473 | int c, error[2] = { 0, 0 }; |
b128c09f | 474 | |
d6320ddb | 475 | for (c = 0; c < mm->mm_children; c++) { |
b128c09f BB |
476 | mirror_child_t *mc = &mm->mm_child[c]; |
477 | int s = mc->mc_speculative; | |
478 | error[s] = zio_worst_error(error[s], mc->mc_error); | |
479 | } | |
480 | ||
481 | return (error[0] ? error[0] : error[1]); | |
482 | } | |
483 | ||
484 | static void | |
34dc7c2f BB |
485 | vdev_mirror_io_done(zio_t *zio) |
486 | { | |
487 | mirror_map_t *mm = zio->io_vsd; | |
488 | mirror_child_t *mc; | |
489 | int c; | |
490 | int good_copies = 0; | |
491 | int unexpected_errors = 0; | |
492 | ||
34dc7c2f BB |
493 | for (c = 0; c < mm->mm_children; c++) { |
494 | mc = &mm->mm_child[c]; | |
495 | ||
34dc7c2f | 496 | if (mc->mc_error) { |
34dc7c2f BB |
497 | if (!mc->mc_skipped) |
498 | unexpected_errors++; | |
b128c09f BB |
499 | } else if (mc->mc_tried) { |
500 | good_copies++; | |
34dc7c2f BB |
501 | } |
502 | } | |
503 | ||
504 | if (zio->io_type == ZIO_TYPE_WRITE) { | |
505 | /* | |
506 | * XXX -- for now, treat partial writes as success. | |
b128c09f BB |
507 | * |
508 | * Now that we support write reallocation, it would be better | |
509 | * to treat partial failure as real failure unless there are | |
510 | * no non-degraded top-level vdevs left, and not update DTLs | |
511 | * if we intend to reallocate. | |
34dc7c2f BB |
512 | */ |
513 | /* XXPOLICY */ | |
b128c09f BB |
514 | if (good_copies != mm->mm_children) { |
515 | /* | |
516 | * Always require at least one good copy. | |
517 | * | |
518 | * For ditto blocks (io_vd == NULL), require | |
519 | * all copies to be good. | |
520 | * | |
521 | * XXX -- for replacing vdevs, there's no great answer. | |
522 | * If the old device is really dead, we may not even | |
523 | * be able to access it -- so we only want to | |
524 | * require good writes to the new device. But if | |
525 | * the new device turns out to be flaky, we want | |
526 | * to be able to detach it -- which requires all | |
527 | * writes to the old device to have succeeded. | |
528 | */ | |
529 | if (good_copies == 0 || zio->io_vd == NULL) | |
530 | zio->io_error = vdev_mirror_worst_error(mm); | |
531 | } | |
532 | return; | |
34dc7c2f BB |
533 | } |
534 | ||
535 | ASSERT(zio->io_type == ZIO_TYPE_READ); | |
536 | ||
537 | /* | |
538 | * If we don't have a good copy yet, keep trying other children. | |
539 | */ | |
540 | /* XXPOLICY */ | |
541 | if (good_copies == 0 && (c = vdev_mirror_child_select(zio)) != -1) { | |
542 | ASSERT(c >= 0 && c < mm->mm_children); | |
543 | mc = &mm->mm_child[c]; | |
34dc7c2f BB |
544 | zio_vdev_io_redone(zio); |
545 | zio_nowait(zio_vdev_child_io(zio, zio->io_bp, | |
546 | mc->mc_vd, mc->mc_offset, zio->io_data, zio->io_size, | |
b128c09f | 547 | ZIO_TYPE_READ, zio->io_priority, 0, |
34dc7c2f | 548 | vdev_mirror_child_done, mc)); |
b128c09f | 549 | return; |
34dc7c2f BB |
550 | } |
551 | ||
552 | /* XXPOLICY */ | |
b128c09f BB |
553 | if (good_copies == 0) { |
554 | zio->io_error = vdev_mirror_worst_error(mm); | |
34dc7c2f | 555 | ASSERT(zio->io_error != 0); |
b128c09f | 556 | } |
34dc7c2f | 557 | |
fb5f0bc8 | 558 | if (good_copies && spa_writeable(zio->io_spa) && |
34dc7c2f BB |
559 | (unexpected_errors || |
560 | (zio->io_flags & ZIO_FLAG_RESILVER) || | |
561 | ((zio->io_flags & ZIO_FLAG_SCRUB) && mm->mm_replacing))) { | |
34dc7c2f BB |
562 | /* |
563 | * Use the good data we have in hand to repair damaged children. | |
34dc7c2f | 564 | */ |
34dc7c2f BB |
565 | for (c = 0; c < mm->mm_children; c++) { |
566 | /* | |
567 | * Don't rewrite known good children. | |
568 | * Not only is it unnecessary, it could | |
569 | * actually be harmful: if the system lost | |
570 | * power while rewriting the only good copy, | |
571 | * there would be no good copies left! | |
572 | */ | |
573 | mc = &mm->mm_child[c]; | |
574 | ||
575 | if (mc->mc_error == 0) { | |
576 | if (mc->mc_tried) | |
577 | continue; | |
578 | if (!(zio->io_flags & ZIO_FLAG_SCRUB) && | |
fb5f0bc8 | 579 | !vdev_dtl_contains(mc->mc_vd, DTL_PARTIAL, |
34dc7c2f BB |
580 | zio->io_txg, 1)) |
581 | continue; | |
2e528b49 | 582 | mc->mc_error = SET_ERROR(ESTALE); |
34dc7c2f BB |
583 | } |
584 | ||
b128c09f BB |
585 | zio_nowait(zio_vdev_child_io(zio, zio->io_bp, |
586 | mc->mc_vd, mc->mc_offset, | |
587 | zio->io_data, zio->io_size, | |
e8b96c60 | 588 | ZIO_TYPE_WRITE, ZIO_PRIORITY_ASYNC_WRITE, |
fb5f0bc8 BB |
589 | ZIO_FLAG_IO_REPAIR | (unexpected_errors ? |
590 | ZIO_FLAG_SELF_HEAL : 0), NULL, NULL)); | |
34dc7c2f | 591 | } |
34dc7c2f | 592 | } |
34dc7c2f BB |
593 | } |
594 | ||
595 | static void | |
596 | vdev_mirror_state_change(vdev_t *vd, int faulted, int degraded) | |
597 | { | |
598 | if (faulted == vd->vdev_children) | |
599 | vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN, | |
600 | VDEV_AUX_NO_REPLICAS); | |
601 | else if (degraded + faulted != 0) | |
602 | vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, VDEV_AUX_NONE); | |
603 | else | |
604 | vdev_set_state(vd, B_FALSE, VDEV_STATE_HEALTHY, VDEV_AUX_NONE); | |
605 | } | |
606 | ||
607 | vdev_ops_t vdev_mirror_ops = { | |
608 | vdev_mirror_open, | |
609 | vdev_mirror_close, | |
34dc7c2f BB |
610 | vdev_default_asize, |
611 | vdev_mirror_io_start, | |
612 | vdev_mirror_io_done, | |
613 | vdev_mirror_state_change, | |
428870ff BB |
614 | NULL, |
615 | NULL, | |
34dc7c2f BB |
616 | VDEV_TYPE_MIRROR, /* name of this vdev type */ |
617 | B_FALSE /* not a leaf vdev */ | |
618 | }; | |
619 | ||
620 | vdev_ops_t vdev_replacing_ops = { | |
621 | vdev_mirror_open, | |
622 | vdev_mirror_close, | |
34dc7c2f BB |
623 | vdev_default_asize, |
624 | vdev_mirror_io_start, | |
625 | vdev_mirror_io_done, | |
626 | vdev_mirror_state_change, | |
428870ff BB |
627 | NULL, |
628 | NULL, | |
34dc7c2f BB |
629 | VDEV_TYPE_REPLACING, /* name of this vdev type */ |
630 | B_FALSE /* not a leaf vdev */ | |
631 | }; | |
632 | ||
633 | vdev_ops_t vdev_spare_ops = { | |
634 | vdev_mirror_open, | |
635 | vdev_mirror_close, | |
34dc7c2f BB |
636 | vdev_default_asize, |
637 | vdev_mirror_io_start, | |
638 | vdev_mirror_io_done, | |
639 | vdev_mirror_state_change, | |
428870ff BB |
640 | NULL, |
641 | NULL, | |
34dc7c2f BB |
642 | VDEV_TYPE_SPARE, /* name of this vdev type */ |
643 | B_FALSE /* not a leaf vdev */ | |
644 | }; | |
556011db BB |
645 | |
646 | #if defined(_KERNEL) && defined(HAVE_SPL) | |
9f500936 | 647 | module_param(zfs_vdev_mirror_rotating_inc, int, 0644); |
648 | MODULE_PARM_DESC(zfs_vdev_mirror_rotating_inc, | |
649 | "Rotating media load increment for non-seeking I/O's"); | |
650 | ||
651 | module_param(zfs_vdev_mirror_rotating_seek_inc, int, 0644); | |
652 | MODULE_PARM_DESC(zfs_vdev_mirror_rotating_seek_inc, | |
653 | "Rotating media load increment for seeking I/O's"); | |
654 | ||
655 | module_param(zfs_vdev_mirror_rotating_seek_offset, int, 0644); | |
656 | MODULE_PARM_DESC(zfs_vdev_mirror_rotating_seek_offset, | |
657 | "Offset in bytes from the last I/O which " | |
658 | "triggers a reduced rotating media seek increment"); | |
659 | ||
660 | module_param(zfs_vdev_mirror_non_rotating_inc, int, 0644); | |
661 | MODULE_PARM_DESC(zfs_vdev_mirror_non_rotating_inc, | |
662 | "Non-rotating media load increment for non-seeking I/O's"); | |
663 | ||
664 | module_param(zfs_vdev_mirror_non_rotating_seek_inc, int, 0644); | |
665 | MODULE_PARM_DESC(zfs_vdev_mirror_non_rotating_seek_inc, | |
666 | "Non-rotating media load increment for seeking I/O's"); | |
667 | ||
556011db | 668 | #endif |