]>
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> | |
f384c045 | 32 | #include <sys/spa_impl.h> |
33 | #include <sys/dsl_pool.h> | |
34 | #include <sys/dsl_scan.h> | |
34dc7c2f BB |
35 | #include <sys/vdev_impl.h> |
36 | #include <sys/zio.h> | |
a6255b7f | 37 | #include <sys/abd.h> |
34dc7c2f BB |
38 | #include <sys/fs/zfs.h> |
39 | ||
551905dd GN |
40 | /* |
41 | * Vdev mirror kstats | |
42 | */ | |
43 | static kstat_t *mirror_ksp = NULL; | |
44 | ||
45 | typedef struct mirror_stats { | |
46 | kstat_named_t vdev_mirror_stat_rotating_linear; | |
47 | kstat_named_t vdev_mirror_stat_rotating_offset; | |
48 | kstat_named_t vdev_mirror_stat_rotating_seek; | |
49 | kstat_named_t vdev_mirror_stat_non_rotating_linear; | |
50 | kstat_named_t vdev_mirror_stat_non_rotating_seek; | |
51 | ||
52 | kstat_named_t vdev_mirror_stat_preferred_found; | |
53 | kstat_named_t vdev_mirror_stat_preferred_not_found; | |
54 | } mirror_stats_t; | |
55 | ||
56 | static mirror_stats_t mirror_stats = { | |
57 | /* New I/O follows directly the last I/O */ | |
58 | { "rotating_linear", KSTAT_DATA_UINT64 }, | |
59 | /* New I/O is within zfs_vdev_mirror_rotating_seek_offset of the last */ | |
60 | { "rotating_offset", KSTAT_DATA_UINT64 }, | |
61 | /* New I/O requires random seek */ | |
62 | { "rotating_seek", KSTAT_DATA_UINT64 }, | |
63 | /* New I/O follows directly the last I/O (nonrot) */ | |
64 | { "non_rotating_linear", KSTAT_DATA_UINT64 }, | |
65 | /* New I/O requires random seek (nonrot) */ | |
66 | { "non_rotating_seek", KSTAT_DATA_UINT64 }, | |
67 | /* Preferred child vdev found */ | |
68 | { "preferred_found", KSTAT_DATA_UINT64 }, | |
69 | /* Preferred child vdev not found or equal load */ | |
70 | { "preferred_not_found", KSTAT_DATA_UINT64 }, | |
71 | ||
72 | }; | |
73 | ||
74 | #define MIRROR_STAT(stat) (mirror_stats.stat.value.ui64) | |
75 | #define MIRROR_INCR(stat, val) atomic_add_64(&MIRROR_STAT(stat), val) | |
76 | #define MIRROR_BUMP(stat) MIRROR_INCR(stat, 1) | |
77 | ||
78 | void | |
79 | vdev_mirror_stat_init(void) | |
80 | { | |
81 | mirror_ksp = kstat_create("zfs", 0, "vdev_mirror_stats", | |
82 | "misc", KSTAT_TYPE_NAMED, | |
83 | sizeof (mirror_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); | |
84 | if (mirror_ksp != NULL) { | |
85 | mirror_ksp->ks_data = &mirror_stats; | |
86 | kstat_install(mirror_ksp); | |
87 | } | |
88 | } | |
89 | ||
90 | void | |
91 | vdev_mirror_stat_fini(void) | |
92 | { | |
93 | if (mirror_ksp != NULL) { | |
94 | kstat_delete(mirror_ksp); | |
95 | mirror_ksp = NULL; | |
96 | } | |
97 | } | |
98 | ||
34dc7c2f BB |
99 | /* |
100 | * Virtual device vector for mirroring. | |
101 | */ | |
102 | ||
103 | typedef struct mirror_child { | |
104 | vdev_t *mc_vd; | |
105 | uint64_t mc_offset; | |
106 | int mc_error; | |
9f500936 | 107 | int mc_load; |
b128c09f BB |
108 | uint8_t mc_tried; |
109 | uint8_t mc_skipped; | |
110 | uint8_t mc_speculative; | |
34dc7c2f BB |
111 | } mirror_child_t; |
112 | ||
113 | typedef struct mirror_map { | |
9f500936 | 114 | int *mm_preferred; |
115 | int mm_preferred_cnt; | |
34dc7c2f | 116 | int mm_children; |
f384c045 | 117 | boolean_t mm_resilvering; |
9f500936 | 118 | boolean_t mm_root; |
119 | mirror_child_t mm_child[]; | |
34dc7c2f BB |
120 | } mirror_map_t; |
121 | ||
9f500936 | 122 | static int vdev_mirror_shift = 21; |
123 | ||
556011db | 124 | /* |
9f500936 | 125 | * The load configuration settings below are tuned by default for |
126 | * the case where all devices are of the same rotational type. | |
556011db | 127 | * |
9f500936 | 128 | * If there is a mixture of rotating and non-rotating media, setting |
129 | * zfs_vdev_mirror_non_rotating_seek_inc to 0 may well provide better results | |
130 | * as it will direct more reads to the non-rotating vdevs which are more likely | |
131 | * to have a higher performance. | |
556011db | 132 | */ |
9f500936 | 133 | |
134 | /* Rotating media load calculation configuration. */ | |
135 | static int zfs_vdev_mirror_rotating_inc = 0; | |
136 | static int zfs_vdev_mirror_rotating_seek_inc = 5; | |
137 | static int zfs_vdev_mirror_rotating_seek_offset = 1 * 1024 * 1024; | |
138 | ||
139 | /* Non-rotating media load calculation configuration. */ | |
140 | static int zfs_vdev_mirror_non_rotating_inc = 0; | |
141 | static int zfs_vdev_mirror_non_rotating_seek_inc = 1; | |
142 | ||
143 | static inline size_t | |
144 | vdev_mirror_map_size(int children) | |
145 | { | |
146 | return (offsetof(mirror_map_t, mm_child[children]) + | |
147 | sizeof (int) * children); | |
148 | } | |
149 | ||
150 | static inline mirror_map_t * | |
f384c045 | 151 | vdev_mirror_map_alloc(int children, boolean_t resilvering, boolean_t root) |
9f500936 | 152 | { |
153 | mirror_map_t *mm; | |
154 | ||
155 | mm = kmem_zalloc(vdev_mirror_map_size(children), KM_SLEEP); | |
156 | mm->mm_children = children; | |
f384c045 | 157 | mm->mm_resilvering = resilvering; |
9f500936 | 158 | mm->mm_root = root; |
159 | mm->mm_preferred = (int *)((uintptr_t)mm + | |
160 | offsetof(mirror_map_t, mm_child[children])); | |
161 | ||
162 | return (mm); | |
163 | } | |
34dc7c2f | 164 | |
b128c09f BB |
165 | static void |
166 | vdev_mirror_map_free(zio_t *zio) | |
167 | { | |
168 | mirror_map_t *mm = zio->io_vsd; | |
169 | ||
9f500936 | 170 | kmem_free(mm, vdev_mirror_map_size(mm->mm_children)); |
b128c09f BB |
171 | } |
172 | ||
428870ff | 173 | static const zio_vsd_ops_t vdev_mirror_vsd_ops = { |
56d8d8ac MW |
174 | .vsd_free = vdev_mirror_map_free, |
175 | .vsd_cksum_report = zio_vsd_default_cksum_report | |
428870ff BB |
176 | }; |
177 | ||
556011db | 178 | static int |
9f500936 | 179 | vdev_mirror_load(mirror_map_t *mm, vdev_t *vd, uint64_t zio_offset) |
556011db | 180 | { |
d6c6590c GN |
181 | uint64_t last_offset; |
182 | int64_t offset_diff; | |
9f500936 | 183 | int load; |
184 | ||
185 | /* All DVAs have equal weight at the root. */ | |
186 | if (mm->mm_root) | |
187 | return (INT_MAX); | |
188 | ||
189 | /* | |
190 | * We don't return INT_MAX if the device is resilvering i.e. | |
191 | * vdev_resilver_txg != 0 as when tested performance was slightly | |
192 | * worse overall when resilvering with compared to without. | |
193 | */ | |
194 | ||
d6c6590c GN |
195 | /* Fix zio_offset for leaf vdevs */ |
196 | if (vd->vdev_ops->vdev_op_leaf) | |
197 | zio_offset += VDEV_LABEL_START_SIZE; | |
198 | ||
9f500936 | 199 | /* Standard load based on pending queue length. */ |
200 | load = vdev_queue_length(vd); | |
d6c6590c | 201 | last_offset = vdev_queue_last_offset(vd); |
9f500936 | 202 | |
203 | if (vd->vdev_nonrot) { | |
204 | /* Non-rotating media. */ | |
551905dd GN |
205 | if (last_offset == zio_offset) { |
206 | MIRROR_BUMP(vdev_mirror_stat_non_rotating_linear); | |
9f500936 | 207 | return (load + zfs_vdev_mirror_non_rotating_inc); |
551905dd | 208 | } |
9f500936 | 209 | |
210 | /* | |
211 | * Apply a seek penalty even for non-rotating devices as | |
212 | * sequential I/O's can be aggregated into fewer operations on | |
213 | * the device, thus avoiding unnecessary per-command overhead | |
214 | * and boosting performance. | |
215 | */ | |
551905dd | 216 | MIRROR_BUMP(vdev_mirror_stat_non_rotating_seek); |
9f500936 | 217 | return (load + zfs_vdev_mirror_non_rotating_seek_inc); |
218 | } | |
219 | ||
220 | /* Rotating media I/O's which directly follow the last I/O. */ | |
551905dd GN |
221 | if (last_offset == zio_offset) { |
222 | MIRROR_BUMP(vdev_mirror_stat_rotating_linear); | |
9f500936 | 223 | return (load + zfs_vdev_mirror_rotating_inc); |
551905dd | 224 | } |
9f500936 | 225 | |
226 | /* | |
227 | * Apply half the seek increment to I/O's within seek offset | |
d6c6590c | 228 | * of the last I/O issued to this vdev as they should incur less |
9f500936 | 229 | * of a seek increment. |
230 | */ | |
d6c6590c | 231 | offset_diff = (int64_t)(last_offset - zio_offset); |
551905dd GN |
232 | if (ABS(offset_diff) < zfs_vdev_mirror_rotating_seek_offset) { |
233 | MIRROR_BUMP(vdev_mirror_stat_rotating_offset); | |
9f500936 | 234 | return (load + (zfs_vdev_mirror_rotating_seek_inc / 2)); |
551905dd | 235 | } |
9f500936 | 236 | |
237 | /* Apply the full seek increment to all other I/O's. */ | |
551905dd | 238 | MIRROR_BUMP(vdev_mirror_stat_rotating_seek); |
9f500936 | 239 | return (load + zfs_vdev_mirror_rotating_seek_inc); |
556011db BB |
240 | } |
241 | ||
a1687880 BB |
242 | /* |
243 | * Avoid inlining the function to keep vdev_mirror_io_start(), which | |
244 | * is this functions only caller, as small as possible on the stack. | |
245 | */ | |
246 | noinline static mirror_map_t * | |
9f500936 | 247 | vdev_mirror_map_init(zio_t *zio) |
34dc7c2f BB |
248 | { |
249 | mirror_map_t *mm = NULL; | |
250 | mirror_child_t *mc; | |
251 | vdev_t *vd = zio->io_vd; | |
9f500936 | 252 | int c; |
34dc7c2f BB |
253 | |
254 | if (vd == NULL) { | |
255 | dva_t *dva = zio->io_bp->blk_dva; | |
256 | spa_t *spa = zio->io_spa; | |
6cb8e530 | 257 | dva_t dva_copy[SPA_DVAS_PER_BP]; |
34dc7c2f | 258 | |
6cb8e530 PZ |
259 | c = BP_GET_NDVAS(zio->io_bp); |
260 | ||
261 | /* | |
262 | * If we do not trust the pool config, some DVAs might be | |
263 | * invalid or point to vdevs that do not exist. We skip them. | |
264 | */ | |
265 | if (!spa_trust_config(spa)) { | |
266 | ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ); | |
267 | int j = 0; | |
268 | for (int i = 0; i < c; i++) { | |
269 | if (zfs_dva_valid(spa, &dva[i], zio->io_bp)) | |
270 | dva_copy[j++] = dva[i]; | |
271 | } | |
272 | if (j == 0) { | |
273 | zio->io_vsd = NULL; | |
274 | zio->io_error = ENXIO; | |
275 | return (NULL); | |
276 | } | |
277 | if (j < c) { | |
278 | dva = dva_copy; | |
279 | c = j; | |
280 | } | |
281 | } | |
282 | ||
283 | mm = vdev_mirror_map_alloc(c, B_FALSE, B_TRUE); | |
34dc7c2f BB |
284 | for (c = 0; c < mm->mm_children; c++) { |
285 | mc = &mm->mm_child[c]; | |
286 | ||
287 | mc->mc_vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[c])); | |
288 | mc->mc_offset = DVA_GET_OFFSET(&dva[c]); | |
289 | } | |
290 | } else { | |
f384c045 | 291 | /* |
292 | * If we are resilvering, then we should handle scrub reads | |
293 | * differently; we shouldn't issue them to the resilvering | |
294 | * device because it might not have those blocks. | |
295 | * | |
296 | * We are resilvering iff: | |
297 | * 1) We are a replacing vdev (ie our name is "replacing-1" or | |
298 | * "spare-1" or something like that), and | |
299 | * 2) The pool is currently being resilvered. | |
300 | * | |
301 | * We cannot simply check vd->vdev_resilver_txg, because it's | |
302 | * not set in this path. | |
303 | * | |
304 | * Nor can we just check our vdev_ops; there are cases (such as | |
305 | * when a user types "zpool replace pool odev spare_dev" and | |
306 | * spare_dev is in the spare list, or when a spare device is | |
307 | * automatically used to replace a DEGRADED device) when | |
308 | * resilvering is complete but both the original vdev and the | |
309 | * spare vdev remain in the pool. That behavior is intentional. | |
310 | * It helps implement the policy that a spare should be | |
311 | * automatically removed from the pool after the user replaces | |
312 | * the device that originally failed. | |
313 | * | |
314 | * If a spa load is in progress, then spa_dsl_pool may be | |
315 | * uninitialized. But we shouldn't be resilvering during a spa | |
316 | * load anyway. | |
317 | */ | |
318 | boolean_t replacing = (vd->vdev_ops == &vdev_replacing_ops || | |
319 | vd->vdev_ops == &vdev_spare_ops) && | |
320 | spa_load_state(vd->vdev_spa) == SPA_LOAD_NONE && | |
321 | dsl_scan_resilvering(vd->vdev_spa->spa_dsl_pool); | |
322 | mm = vdev_mirror_map_alloc(vd->vdev_children, replacing, | |
323 | B_FALSE); | |
34dc7c2f BB |
324 | for (c = 0; c < mm->mm_children; c++) { |
325 | mc = &mm->mm_child[c]; | |
326 | mc->mc_vd = vd->vdev_child[c]; | |
327 | mc->mc_offset = zio->io_offset; | |
328 | } | |
329 | } | |
330 | ||
331 | zio->io_vsd = mm; | |
428870ff | 332 | zio->io_vsd_ops = &vdev_mirror_vsd_ops; |
34dc7c2f BB |
333 | return (mm); |
334 | } | |
335 | ||
34dc7c2f | 336 | static int |
1bd201e7 CS |
337 | vdev_mirror_open(vdev_t *vd, uint64_t *asize, uint64_t *max_asize, |
338 | uint64_t *ashift) | |
34dc7c2f | 339 | { |
34dc7c2f | 340 | int numerrors = 0; |
45d1cae3 | 341 | int lasterror = 0; |
34dc7c2f BB |
342 | |
343 | if (vd->vdev_children == 0) { | |
344 | vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL; | |
2e528b49 | 345 | return (SET_ERROR(EINVAL)); |
34dc7c2f BB |
346 | } |
347 | ||
45d1cae3 | 348 | vdev_open_children(vd); |
34dc7c2f | 349 | |
1c27024e | 350 | for (int c = 0; c < vd->vdev_children; c++) { |
45d1cae3 BB |
351 | vdev_t *cvd = vd->vdev_child[c]; |
352 | ||
353 | if (cvd->vdev_open_error) { | |
354 | lasterror = cvd->vdev_open_error; | |
34dc7c2f BB |
355 | numerrors++; |
356 | continue; | |
357 | } | |
358 | ||
359 | *asize = MIN(*asize - 1, cvd->vdev_asize - 1) + 1; | |
1bd201e7 | 360 | *max_asize = MIN(*max_asize - 1, cvd->vdev_max_asize - 1) + 1; |
34dc7c2f BB |
361 | *ashift = MAX(*ashift, cvd->vdev_ashift); |
362 | } | |
363 | ||
364 | if (numerrors == vd->vdev_children) { | |
6cb8e530 PZ |
365 | if (vdev_children_are_offline(vd)) |
366 | vd->vdev_stat.vs_aux = VDEV_AUX_CHILDREN_OFFLINE; | |
367 | else | |
368 | vd->vdev_stat.vs_aux = VDEV_AUX_NO_REPLICAS; | |
34dc7c2f BB |
369 | return (lasterror); |
370 | } | |
371 | ||
372 | return (0); | |
373 | } | |
374 | ||
375 | static void | |
376 | vdev_mirror_close(vdev_t *vd) | |
377 | { | |
1c27024e | 378 | for (int c = 0; c < vd->vdev_children; c++) |
34dc7c2f BB |
379 | vdev_close(vd->vdev_child[c]); |
380 | } | |
381 | ||
382 | static void | |
383 | vdev_mirror_child_done(zio_t *zio) | |
384 | { | |
385 | mirror_child_t *mc = zio->io_private; | |
386 | ||
387 | mc->mc_error = zio->io_error; | |
388 | mc->mc_tried = 1; | |
389 | mc->mc_skipped = 0; | |
390 | } | |
391 | ||
392 | static void | |
393 | vdev_mirror_scrub_done(zio_t *zio) | |
394 | { | |
395 | mirror_child_t *mc = zio->io_private; | |
396 | ||
397 | if (zio->io_error == 0) { | |
d164b209 | 398 | zio_t *pio; |
3dfb57a3 | 399 | zio_link_t *zl = NULL; |
d164b209 BB |
400 | |
401 | mutex_enter(&zio->io_lock); | |
3dfb57a3 | 402 | while ((pio = zio_walk_parents(zio, &zl)) != NULL) { |
d164b209 BB |
403 | mutex_enter(&pio->io_lock); |
404 | ASSERT3U(zio->io_size, >=, pio->io_size); | |
a6255b7f | 405 | abd_copy(pio->io_abd, zio->io_abd, pio->io_size); |
d164b209 BB |
406 | mutex_exit(&pio->io_lock); |
407 | } | |
408 | mutex_exit(&zio->io_lock); | |
34dc7c2f BB |
409 | } |
410 | ||
a6255b7f | 411 | abd_free(zio->io_abd); |
34dc7c2f BB |
412 | |
413 | mc->mc_error = zio->io_error; | |
414 | mc->mc_tried = 1; | |
415 | mc->mc_skipped = 0; | |
416 | } | |
417 | ||
34dc7c2f | 418 | /* |
9f500936 | 419 | * Check the other, lower-index DVAs to see if they're on the same |
420 | * vdev as the child we picked. If they are, use them since they | |
421 | * are likely to have been allocated from the primary metaslab in | |
422 | * use at the time, and hence are more likely to have locality with | |
423 | * single-copy data. | |
424 | */ | |
425 | static int | |
426 | vdev_mirror_dva_select(zio_t *zio, int p) | |
427 | { | |
428 | dva_t *dva = zio->io_bp->blk_dva; | |
429 | mirror_map_t *mm = zio->io_vsd; | |
430 | int preferred; | |
431 | int c; | |
432 | ||
433 | preferred = mm->mm_preferred[p]; | |
434 | for (p--; p >= 0; p--) { | |
435 | c = mm->mm_preferred[p]; | |
436 | if (DVA_GET_VDEV(&dva[c]) == DVA_GET_VDEV(&dva[preferred])) | |
437 | preferred = c; | |
438 | } | |
439 | return (preferred); | |
440 | } | |
441 | ||
442 | static int | |
443 | vdev_mirror_preferred_child_randomize(zio_t *zio) | |
444 | { | |
445 | mirror_map_t *mm = zio->io_vsd; | |
446 | int p; | |
447 | ||
448 | if (mm->mm_root) { | |
449 | p = spa_get_random(mm->mm_preferred_cnt); | |
450 | return (vdev_mirror_dva_select(zio, p)); | |
451 | } | |
452 | ||
453 | /* | |
454 | * To ensure we don't always favour the first matching vdev, | |
455 | * which could lead to wear leveling issues on SSD's, we | |
456 | * use the I/O offset as a pseudo random seed into the vdevs | |
457 | * which have the lowest load. | |
458 | */ | |
459 | p = (zio->io_offset >> vdev_mirror_shift) % mm->mm_preferred_cnt; | |
460 | return (mm->mm_preferred[p]); | |
461 | } | |
462 | ||
463 | /* | |
464 | * Try to find a vdev whose DTL doesn't contain the block we want to read | |
465 | * prefering vdevs based on determined load. | |
466 | * | |
34dc7c2f BB |
467 | * Try to find a child whose DTL doesn't contain the block we want to read. |
468 | * If we can't, try the read on any vdev we haven't already tried. | |
469 | */ | |
470 | static int | |
471 | vdev_mirror_child_select(zio_t *zio) | |
472 | { | |
473 | mirror_map_t *mm = zio->io_vsd; | |
34dc7c2f | 474 | uint64_t txg = zio->io_txg; |
9f500936 | 475 | int c, lowest_load; |
34dc7c2f | 476 | |
428870ff | 477 | ASSERT(zio->io_bp == NULL || BP_PHYSICAL_BIRTH(zio->io_bp) == txg); |
34dc7c2f | 478 | |
9f500936 | 479 | lowest_load = INT_MAX; |
480 | mm->mm_preferred_cnt = 0; | |
481 | for (c = 0; c < mm->mm_children; c++) { | |
482 | mirror_child_t *mc; | |
483 | ||
34dc7c2f BB |
484 | mc = &mm->mm_child[c]; |
485 | if (mc->mc_tried || mc->mc_skipped) | |
486 | continue; | |
9f500936 | 487 | |
33074f22 | 488 | if (mc->mc_vd == NULL || !vdev_readable(mc->mc_vd)) { |
2e528b49 | 489 | mc->mc_error = SET_ERROR(ENXIO); |
34dc7c2f BB |
490 | mc->mc_tried = 1; /* don't even try */ |
491 | mc->mc_skipped = 1; | |
492 | continue; | |
493 | } | |
9f500936 | 494 | |
495 | if (vdev_dtl_contains(mc->mc_vd, DTL_MISSING, txg, 1)) { | |
496 | mc->mc_error = SET_ERROR(ESTALE); | |
497 | mc->mc_skipped = 1; | |
498 | mc->mc_speculative = 1; | |
499 | continue; | |
500 | } | |
501 | ||
502 | mc->mc_load = vdev_mirror_load(mm, mc->mc_vd, mc->mc_offset); | |
503 | if (mc->mc_load > lowest_load) | |
504 | continue; | |
505 | ||
506 | if (mc->mc_load < lowest_load) { | |
507 | lowest_load = mc->mc_load; | |
508 | mm->mm_preferred_cnt = 0; | |
509 | } | |
510 | mm->mm_preferred[mm->mm_preferred_cnt] = c; | |
511 | mm->mm_preferred_cnt++; | |
512 | } | |
513 | ||
551905dd GN |
514 | if (mm->mm_preferred_cnt == 1) { |
515 | MIRROR_BUMP(vdev_mirror_stat_preferred_found); | |
9f500936 | 516 | return (mm->mm_preferred[0]); |
551905dd | 517 | } |
9f500936 | 518 | |
551905dd GN |
519 | if (mm->mm_preferred_cnt > 1) { |
520 | MIRROR_BUMP(vdev_mirror_stat_preferred_not_found); | |
d6c6590c | 521 | return (vdev_mirror_preferred_child_randomize(zio)); |
551905dd | 522 | } |
34dc7c2f BB |
523 | |
524 | /* | |
525 | * Every device is either missing or has this txg in its DTL. | |
526 | * Look for any child we haven't already tried before giving up. | |
527 | */ | |
9f500936 | 528 | for (c = 0; c < mm->mm_children; c++) { |
d6c6590c | 529 | if (!mm->mm_child[c].mc_tried) |
34dc7c2f | 530 | return (c); |
9f500936 | 531 | } |
34dc7c2f BB |
532 | |
533 | /* | |
534 | * Every child failed. There's no place left to look. | |
535 | */ | |
536 | return (-1); | |
537 | } | |
538 | ||
98b25418 | 539 | static void |
34dc7c2f BB |
540 | vdev_mirror_io_start(zio_t *zio) |
541 | { | |
542 | mirror_map_t *mm; | |
543 | mirror_child_t *mc; | |
544 | int c, children; | |
545 | ||
9f500936 | 546 | mm = vdev_mirror_map_init(zio); |
34dc7c2f | 547 | |
6cb8e530 PZ |
548 | if (mm == NULL) { |
549 | ASSERT(!spa_trust_config(zio->io_spa)); | |
550 | ASSERT(zio->io_type == ZIO_TYPE_READ); | |
551 | zio_execute(zio); | |
552 | return; | |
553 | } | |
554 | ||
34dc7c2f | 555 | if (zio->io_type == ZIO_TYPE_READ) { |
9e052db4 | 556 | if (zio->io_bp != NULL && |
f384c045 | 557 | (zio->io_flags & ZIO_FLAG_SCRUB) && !mm->mm_resilvering) { |
34dc7c2f | 558 | /* |
9e052db4 MA |
559 | * For scrubbing reads (if we can verify the |
560 | * checksum here, as indicated by io_bp being | |
561 | * non-NULL) we need to allocate a read buffer for | |
562 | * each child and issue reads to all children. If | |
563 | * any child succeeds, it will copy its data into | |
564 | * zio->io_data in vdev_mirror_scrub_done. | |
34dc7c2f BB |
565 | */ |
566 | for (c = 0; c < mm->mm_children; c++) { | |
567 | mc = &mm->mm_child[c]; | |
568 | zio_nowait(zio_vdev_child_io(zio, zio->io_bp, | |
569 | mc->mc_vd, mc->mc_offset, | |
a6255b7f DQ |
570 | abd_alloc_sametype(zio->io_abd, |
571 | zio->io_size), zio->io_size, | |
b128c09f | 572 | zio->io_type, zio->io_priority, 0, |
34dc7c2f BB |
573 | vdev_mirror_scrub_done, mc)); |
574 | } | |
98b25418 GW |
575 | zio_execute(zio); |
576 | return; | |
34dc7c2f BB |
577 | } |
578 | /* | |
579 | * For normal reads just pick one child. | |
580 | */ | |
581 | c = vdev_mirror_child_select(zio); | |
582 | children = (c >= 0); | |
583 | } else { | |
584 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
585 | ||
586 | /* | |
fb5f0bc8 | 587 | * Writes go to all children. |
34dc7c2f | 588 | */ |
fb5f0bc8 BB |
589 | c = 0; |
590 | children = mm->mm_children; | |
34dc7c2f BB |
591 | } |
592 | ||
593 | while (children--) { | |
594 | mc = &mm->mm_child[c]; | |
595 | zio_nowait(zio_vdev_child_io(zio, zio->io_bp, | |
a6255b7f | 596 | mc->mc_vd, mc->mc_offset, zio->io_abd, zio->io_size, |
b128c09f BB |
597 | zio->io_type, zio->io_priority, 0, |
598 | vdev_mirror_child_done, mc)); | |
34dc7c2f BB |
599 | c++; |
600 | } | |
601 | ||
98b25418 | 602 | zio_execute(zio); |
34dc7c2f BB |
603 | } |
604 | ||
605 | static int | |
b128c09f BB |
606 | vdev_mirror_worst_error(mirror_map_t *mm) |
607 | { | |
1c27024e | 608 | int error[2] = { 0, 0 }; |
b128c09f | 609 | |
1c27024e | 610 | for (int c = 0; c < mm->mm_children; c++) { |
b128c09f BB |
611 | mirror_child_t *mc = &mm->mm_child[c]; |
612 | int s = mc->mc_speculative; | |
613 | error[s] = zio_worst_error(error[s], mc->mc_error); | |
614 | } | |
615 | ||
616 | return (error[0] ? error[0] : error[1]); | |
617 | } | |
618 | ||
619 | static void | |
34dc7c2f BB |
620 | vdev_mirror_io_done(zio_t *zio) |
621 | { | |
622 | mirror_map_t *mm = zio->io_vsd; | |
623 | mirror_child_t *mc; | |
624 | int c; | |
625 | int good_copies = 0; | |
626 | int unexpected_errors = 0; | |
627 | ||
6cb8e530 PZ |
628 | if (mm == NULL) |
629 | return; | |
630 | ||
34dc7c2f BB |
631 | for (c = 0; c < mm->mm_children; c++) { |
632 | mc = &mm->mm_child[c]; | |
633 | ||
34dc7c2f | 634 | if (mc->mc_error) { |
34dc7c2f BB |
635 | if (!mc->mc_skipped) |
636 | unexpected_errors++; | |
b128c09f BB |
637 | } else if (mc->mc_tried) { |
638 | good_copies++; | |
34dc7c2f BB |
639 | } |
640 | } | |
641 | ||
642 | if (zio->io_type == ZIO_TYPE_WRITE) { | |
643 | /* | |
644 | * XXX -- for now, treat partial writes as success. | |
b128c09f BB |
645 | * |
646 | * Now that we support write reallocation, it would be better | |
647 | * to treat partial failure as real failure unless there are | |
648 | * no non-degraded top-level vdevs left, and not update DTLs | |
649 | * if we intend to reallocate. | |
34dc7c2f BB |
650 | */ |
651 | /* XXPOLICY */ | |
b128c09f BB |
652 | if (good_copies != mm->mm_children) { |
653 | /* | |
654 | * Always require at least one good copy. | |
655 | * | |
656 | * For ditto blocks (io_vd == NULL), require | |
657 | * all copies to be good. | |
658 | * | |
659 | * XXX -- for replacing vdevs, there's no great answer. | |
660 | * If the old device is really dead, we may not even | |
661 | * be able to access it -- so we only want to | |
662 | * require good writes to the new device. But if | |
663 | * the new device turns out to be flaky, we want | |
664 | * to be able to detach it -- which requires all | |
665 | * writes to the old device to have succeeded. | |
666 | */ | |
667 | if (good_copies == 0 || zio->io_vd == NULL) | |
668 | zio->io_error = vdev_mirror_worst_error(mm); | |
669 | } | |
670 | return; | |
34dc7c2f BB |
671 | } |
672 | ||
673 | ASSERT(zio->io_type == ZIO_TYPE_READ); | |
674 | ||
675 | /* | |
676 | * If we don't have a good copy yet, keep trying other children. | |
677 | */ | |
678 | /* XXPOLICY */ | |
679 | if (good_copies == 0 && (c = vdev_mirror_child_select(zio)) != -1) { | |
680 | ASSERT(c >= 0 && c < mm->mm_children); | |
681 | mc = &mm->mm_child[c]; | |
34dc7c2f BB |
682 | zio_vdev_io_redone(zio); |
683 | zio_nowait(zio_vdev_child_io(zio, zio->io_bp, | |
a6255b7f | 684 | mc->mc_vd, mc->mc_offset, zio->io_abd, zio->io_size, |
b128c09f | 685 | ZIO_TYPE_READ, zio->io_priority, 0, |
34dc7c2f | 686 | vdev_mirror_child_done, mc)); |
b128c09f | 687 | return; |
34dc7c2f BB |
688 | } |
689 | ||
690 | /* XXPOLICY */ | |
b128c09f BB |
691 | if (good_copies == 0) { |
692 | zio->io_error = vdev_mirror_worst_error(mm); | |
34dc7c2f | 693 | ASSERT(zio->io_error != 0); |
b128c09f | 694 | } |
34dc7c2f | 695 | |
fb5f0bc8 | 696 | if (good_copies && spa_writeable(zio->io_spa) && |
34dc7c2f BB |
697 | (unexpected_errors || |
698 | (zio->io_flags & ZIO_FLAG_RESILVER) || | |
f384c045 | 699 | ((zio->io_flags & ZIO_FLAG_SCRUB) && mm->mm_resilvering))) { |
34dc7c2f BB |
700 | /* |
701 | * Use the good data we have in hand to repair damaged children. | |
34dc7c2f | 702 | */ |
34dc7c2f BB |
703 | for (c = 0; c < mm->mm_children; c++) { |
704 | /* | |
705 | * Don't rewrite known good children. | |
706 | * Not only is it unnecessary, it could | |
707 | * actually be harmful: if the system lost | |
708 | * power while rewriting the only good copy, | |
709 | * there would be no good copies left! | |
710 | */ | |
711 | mc = &mm->mm_child[c]; | |
712 | ||
713 | if (mc->mc_error == 0) { | |
714 | if (mc->mc_tried) | |
715 | continue; | |
9e052db4 MA |
716 | /* |
717 | * We didn't try this child. We need to | |
718 | * repair it if: | |
719 | * 1. it's a scrub (in which case we have | |
720 | * tried everything that was healthy) | |
721 | * - or - | |
722 | * 2. it's an indirect vdev (in which case | |
723 | * it could point to any other vdev, which | |
724 | * might have a bad DTL) | |
725 | * - or - | |
726 | * 3. the DTL indicates that this data is | |
727 | * missing from this vdev | |
728 | */ | |
34dc7c2f | 729 | if (!(zio->io_flags & ZIO_FLAG_SCRUB) && |
9e052db4 | 730 | mc->mc_vd->vdev_ops != &vdev_indirect_ops && |
fb5f0bc8 | 731 | !vdev_dtl_contains(mc->mc_vd, DTL_PARTIAL, |
34dc7c2f BB |
732 | zio->io_txg, 1)) |
733 | continue; | |
2e528b49 | 734 | mc->mc_error = SET_ERROR(ESTALE); |
34dc7c2f BB |
735 | } |
736 | ||
b128c09f BB |
737 | zio_nowait(zio_vdev_child_io(zio, zio->io_bp, |
738 | mc->mc_vd, mc->mc_offset, | |
a6255b7f | 739 | zio->io_abd, zio->io_size, |
e8b96c60 | 740 | ZIO_TYPE_WRITE, ZIO_PRIORITY_ASYNC_WRITE, |
fb5f0bc8 BB |
741 | ZIO_FLAG_IO_REPAIR | (unexpected_errors ? |
742 | ZIO_FLAG_SELF_HEAL : 0), NULL, NULL)); | |
34dc7c2f | 743 | } |
34dc7c2f | 744 | } |
34dc7c2f BB |
745 | } |
746 | ||
747 | static void | |
748 | vdev_mirror_state_change(vdev_t *vd, int faulted, int degraded) | |
749 | { | |
6cb8e530 PZ |
750 | if (faulted == vd->vdev_children) { |
751 | if (vdev_children_are_offline(vd)) { | |
752 | vdev_set_state(vd, B_FALSE, VDEV_STATE_OFFLINE, | |
753 | VDEV_AUX_CHILDREN_OFFLINE); | |
754 | } else { | |
755 | vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN, | |
756 | VDEV_AUX_NO_REPLICAS); | |
757 | } | |
758 | } else if (degraded + faulted != 0) { | |
34dc7c2f | 759 | vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, VDEV_AUX_NONE); |
6cb8e530 | 760 | } else { |
34dc7c2f | 761 | vdev_set_state(vd, B_FALSE, VDEV_STATE_HEALTHY, VDEV_AUX_NONE); |
6cb8e530 | 762 | } |
34dc7c2f BB |
763 | } |
764 | ||
765 | vdev_ops_t vdev_mirror_ops = { | |
766 | vdev_mirror_open, | |
767 | vdev_mirror_close, | |
34dc7c2f BB |
768 | vdev_default_asize, |
769 | vdev_mirror_io_start, | |
770 | vdev_mirror_io_done, | |
771 | vdev_mirror_state_change, | |
428870ff BB |
772 | NULL, |
773 | NULL, | |
3d6da72d | 774 | NULL, |
a1d477c2 | 775 | NULL, |
619f0976 | 776 | vdev_default_xlate, |
34dc7c2f BB |
777 | VDEV_TYPE_MIRROR, /* name of this vdev type */ |
778 | B_FALSE /* not a leaf vdev */ | |
779 | }; | |
780 | ||
781 | vdev_ops_t vdev_replacing_ops = { | |
782 | vdev_mirror_open, | |
783 | vdev_mirror_close, | |
34dc7c2f BB |
784 | vdev_default_asize, |
785 | vdev_mirror_io_start, | |
786 | vdev_mirror_io_done, | |
787 | vdev_mirror_state_change, | |
428870ff BB |
788 | NULL, |
789 | NULL, | |
3d6da72d | 790 | NULL, |
a1d477c2 | 791 | NULL, |
619f0976 | 792 | vdev_default_xlate, |
34dc7c2f BB |
793 | VDEV_TYPE_REPLACING, /* name of this vdev type */ |
794 | B_FALSE /* not a leaf vdev */ | |
795 | }; | |
796 | ||
797 | vdev_ops_t vdev_spare_ops = { | |
798 | vdev_mirror_open, | |
799 | vdev_mirror_close, | |
34dc7c2f BB |
800 | vdev_default_asize, |
801 | vdev_mirror_io_start, | |
802 | vdev_mirror_io_done, | |
803 | vdev_mirror_state_change, | |
428870ff BB |
804 | NULL, |
805 | NULL, | |
3d6da72d | 806 | NULL, |
a1d477c2 | 807 | NULL, |
619f0976 | 808 | vdev_default_xlate, |
34dc7c2f BB |
809 | VDEV_TYPE_SPARE, /* name of this vdev type */ |
810 | B_FALSE /* not a leaf vdev */ | |
811 | }; | |
556011db | 812 | |
93ce2b4c | 813 | #if defined(_KERNEL) |
4ea3f864 | 814 | /* BEGIN CSTYLED */ |
9f500936 | 815 | module_param(zfs_vdev_mirror_rotating_inc, int, 0644); |
816 | MODULE_PARM_DESC(zfs_vdev_mirror_rotating_inc, | |
817 | "Rotating media load increment for non-seeking I/O's"); | |
818 | ||
819 | module_param(zfs_vdev_mirror_rotating_seek_inc, int, 0644); | |
820 | MODULE_PARM_DESC(zfs_vdev_mirror_rotating_seek_inc, | |
821 | "Rotating media load increment for seeking I/O's"); | |
822 | ||
823 | module_param(zfs_vdev_mirror_rotating_seek_offset, int, 0644); | |
4ea3f864 | 824 | |
9f500936 | 825 | MODULE_PARM_DESC(zfs_vdev_mirror_rotating_seek_offset, |
826 | "Offset in bytes from the last I/O which " | |
827 | "triggers a reduced rotating media seek increment"); | |
828 | ||
829 | module_param(zfs_vdev_mirror_non_rotating_inc, int, 0644); | |
830 | MODULE_PARM_DESC(zfs_vdev_mirror_non_rotating_inc, | |
831 | "Non-rotating media load increment for non-seeking I/O's"); | |
832 | ||
833 | module_param(zfs_vdev_mirror_non_rotating_seek_inc, int, 0644); | |
834 | MODULE_PARM_DESC(zfs_vdev_mirror_non_rotating_seek_inc, | |
835 | "Non-rotating media load increment for seeking I/O's"); | |
4ea3f864 | 836 | /* END CSTYLED */ |
556011db | 837 | #endif |