]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * raid10.c : Multiple Devices driver for Linux | |
3 | * | |
4 | * Copyright (C) 2000-2004 Neil Brown | |
5 | * | |
6 | * RAID-10 support for md. | |
7 | * | |
25985edc | 8 | * Base on code in raid1.c. See raid1.c for further copyright information. |
1da177e4 LT |
9 | * |
10 | * | |
11 | * This program is free software; you can redistribute it and/or modify | |
12 | * it under the terms of the GNU General Public License as published by | |
13 | * the Free Software Foundation; either version 2, or (at your option) | |
14 | * any later version. | |
15 | * | |
16 | * You should have received a copy of the GNU General Public License | |
17 | * (for example /usr/src/linux/COPYING); if not, write to the Free | |
18 | * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. | |
19 | */ | |
20 | ||
5a0e3ad6 | 21 | #include <linux/slab.h> |
25570727 | 22 | #include <linux/delay.h> |
bff61975 | 23 | #include <linux/blkdev.h> |
bff61975 | 24 | #include <linux/seq_file.h> |
8bda470e | 25 | #include <linux/ratelimit.h> |
43b2e5d8 | 26 | #include "md.h" |
ef740c37 | 27 | #include "raid10.h" |
dab8b292 | 28 | #include "raid0.h" |
ef740c37 | 29 | #include "bitmap.h" |
1da177e4 LT |
30 | |
31 | /* | |
32 | * RAID10 provides a combination of RAID0 and RAID1 functionality. | |
33 | * The layout of data is defined by | |
34 | * chunk_size | |
35 | * raid_disks | |
36 | * near_copies (stored in low byte of layout) | |
37 | * far_copies (stored in second byte of layout) | |
c93983bf | 38 | * far_offset (stored in bit 16 of layout ) |
1da177e4 LT |
39 | * |
40 | * The data to be stored is divided into chunks using chunksize. | |
41 | * Each device is divided into far_copies sections. | |
42 | * In each section, chunks are laid out in a style similar to raid0, but | |
43 | * near_copies copies of each chunk is stored (each on a different drive). | |
44 | * The starting device for each section is offset near_copies from the starting | |
45 | * device of the previous section. | |
c93983bf | 46 | * Thus they are (near_copies*far_copies) of each chunk, and each is on a different |
1da177e4 LT |
47 | * drive. |
48 | * near_copies and far_copies must be at least one, and their product is at most | |
49 | * raid_disks. | |
c93983bf N |
50 | * |
51 | * If far_offset is true, then the far_copies are handled a bit differently. | |
52 | * The copies are still in different stripes, but instead of be very far apart | |
53 | * on disk, there are adjacent stripes. | |
1da177e4 LT |
54 | */ |
55 | ||
56 | /* | |
57 | * Number of guaranteed r10bios in case of extreme VM load: | |
58 | */ | |
59 | #define NR_RAID10_BIOS 256 | |
60 | ||
0a27ec96 N |
61 | static void allow_barrier(conf_t *conf); |
62 | static void lower_barrier(conf_t *conf); | |
63 | ||
dd0fc66f | 64 | static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data) |
1da177e4 LT |
65 | { |
66 | conf_t *conf = data; | |
1da177e4 LT |
67 | int size = offsetof(struct r10bio_s, devs[conf->copies]); |
68 | ||
69 | /* allocate a r10bio with room for raid_disks entries in the bios array */ | |
7eaceacc | 70 | return kzalloc(size, gfp_flags); |
1da177e4 LT |
71 | } |
72 | ||
73 | static void r10bio_pool_free(void *r10_bio, void *data) | |
74 | { | |
75 | kfree(r10_bio); | |
76 | } | |
77 | ||
0310fa21 | 78 | /* Maximum size of each resync request */ |
1da177e4 | 79 | #define RESYNC_BLOCK_SIZE (64*1024) |
1da177e4 | 80 | #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE) |
0310fa21 N |
81 | /* amount of memory to reserve for resync requests */ |
82 | #define RESYNC_WINDOW (1024*1024) | |
83 | /* maximum number of concurrent requests, memory permitting */ | |
84 | #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE) | |
1da177e4 LT |
85 | |
86 | /* | |
87 | * When performing a resync, we need to read and compare, so | |
88 | * we need as many pages are there are copies. | |
89 | * When performing a recovery, we need 2 bios, one for read, | |
90 | * one for write (we recover only one drive per r10buf) | |
91 | * | |
92 | */ | |
dd0fc66f | 93 | static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data) |
1da177e4 LT |
94 | { |
95 | conf_t *conf = data; | |
96 | struct page *page; | |
97 | r10bio_t *r10_bio; | |
98 | struct bio *bio; | |
99 | int i, j; | |
100 | int nalloc; | |
101 | ||
102 | r10_bio = r10bio_pool_alloc(gfp_flags, conf); | |
7eaceacc | 103 | if (!r10_bio) |
1da177e4 | 104 | return NULL; |
1da177e4 LT |
105 | |
106 | if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery)) | |
107 | nalloc = conf->copies; /* resync */ | |
108 | else | |
109 | nalloc = 2; /* recovery */ | |
110 | ||
111 | /* | |
112 | * Allocate bios. | |
113 | */ | |
114 | for (j = nalloc ; j-- ; ) { | |
6746557f | 115 | bio = bio_kmalloc(gfp_flags, RESYNC_PAGES); |
1da177e4 LT |
116 | if (!bio) |
117 | goto out_free_bio; | |
118 | r10_bio->devs[j].bio = bio; | |
119 | } | |
120 | /* | |
121 | * Allocate RESYNC_PAGES data pages and attach them | |
122 | * where needed. | |
123 | */ | |
124 | for (j = 0 ; j < nalloc; j++) { | |
125 | bio = r10_bio->devs[j].bio; | |
126 | for (i = 0; i < RESYNC_PAGES; i++) { | |
c65060ad NK |
127 | if (j == 1 && !test_bit(MD_RECOVERY_SYNC, |
128 | &conf->mddev->recovery)) { | |
129 | /* we can share bv_page's during recovery */ | |
130 | struct bio *rbio = r10_bio->devs[0].bio; | |
131 | page = rbio->bi_io_vec[i].bv_page; | |
132 | get_page(page); | |
133 | } else | |
134 | page = alloc_page(gfp_flags); | |
1da177e4 LT |
135 | if (unlikely(!page)) |
136 | goto out_free_pages; | |
137 | ||
138 | bio->bi_io_vec[i].bv_page = page; | |
139 | } | |
140 | } | |
141 | ||
142 | return r10_bio; | |
143 | ||
144 | out_free_pages: | |
145 | for ( ; i > 0 ; i--) | |
1345b1d8 | 146 | safe_put_page(bio->bi_io_vec[i-1].bv_page); |
1da177e4 LT |
147 | while (j--) |
148 | for (i = 0; i < RESYNC_PAGES ; i++) | |
1345b1d8 | 149 | safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page); |
1da177e4 LT |
150 | j = -1; |
151 | out_free_bio: | |
152 | while ( ++j < nalloc ) | |
153 | bio_put(r10_bio->devs[j].bio); | |
154 | r10bio_pool_free(r10_bio, conf); | |
155 | return NULL; | |
156 | } | |
157 | ||
158 | static void r10buf_pool_free(void *__r10_bio, void *data) | |
159 | { | |
160 | int i; | |
161 | conf_t *conf = data; | |
162 | r10bio_t *r10bio = __r10_bio; | |
163 | int j; | |
164 | ||
165 | for (j=0; j < conf->copies; j++) { | |
166 | struct bio *bio = r10bio->devs[j].bio; | |
167 | if (bio) { | |
168 | for (i = 0; i < RESYNC_PAGES; i++) { | |
1345b1d8 | 169 | safe_put_page(bio->bi_io_vec[i].bv_page); |
1da177e4 LT |
170 | bio->bi_io_vec[i].bv_page = NULL; |
171 | } | |
172 | bio_put(bio); | |
173 | } | |
174 | } | |
175 | r10bio_pool_free(r10bio, conf); | |
176 | } | |
177 | ||
178 | static void put_all_bios(conf_t *conf, r10bio_t *r10_bio) | |
179 | { | |
180 | int i; | |
181 | ||
182 | for (i = 0; i < conf->copies; i++) { | |
183 | struct bio **bio = & r10_bio->devs[i].bio; | |
749c55e9 | 184 | if (!BIO_SPECIAL(*bio)) |
1da177e4 LT |
185 | bio_put(*bio); |
186 | *bio = NULL; | |
187 | } | |
188 | } | |
189 | ||
858119e1 | 190 | static void free_r10bio(r10bio_t *r10_bio) |
1da177e4 | 191 | { |
070ec55d | 192 | conf_t *conf = r10_bio->mddev->private; |
1da177e4 | 193 | |
1da177e4 LT |
194 | put_all_bios(conf, r10_bio); |
195 | mempool_free(r10_bio, conf->r10bio_pool); | |
196 | } | |
197 | ||
858119e1 | 198 | static void put_buf(r10bio_t *r10_bio) |
1da177e4 | 199 | { |
070ec55d | 200 | conf_t *conf = r10_bio->mddev->private; |
1da177e4 LT |
201 | |
202 | mempool_free(r10_bio, conf->r10buf_pool); | |
203 | ||
0a27ec96 | 204 | lower_barrier(conf); |
1da177e4 LT |
205 | } |
206 | ||
207 | static void reschedule_retry(r10bio_t *r10_bio) | |
208 | { | |
209 | unsigned long flags; | |
210 | mddev_t *mddev = r10_bio->mddev; | |
070ec55d | 211 | conf_t *conf = mddev->private; |
1da177e4 LT |
212 | |
213 | spin_lock_irqsave(&conf->device_lock, flags); | |
214 | list_add(&r10_bio->retry_list, &conf->retry_list); | |
4443ae10 | 215 | conf->nr_queued ++; |
1da177e4 LT |
216 | spin_unlock_irqrestore(&conf->device_lock, flags); |
217 | ||
388667be AJ |
218 | /* wake up frozen array... */ |
219 | wake_up(&conf->wait_barrier); | |
220 | ||
1da177e4 LT |
221 | md_wakeup_thread(mddev->thread); |
222 | } | |
223 | ||
224 | /* | |
225 | * raid_end_bio_io() is called when we have finished servicing a mirrored | |
226 | * operation and are ready to return a success/failure code to the buffer | |
227 | * cache layer. | |
228 | */ | |
229 | static void raid_end_bio_io(r10bio_t *r10_bio) | |
230 | { | |
231 | struct bio *bio = r10_bio->master_bio; | |
856e08e2 N |
232 | int done; |
233 | conf_t *conf = r10_bio->mddev->private; | |
1da177e4 | 234 | |
856e08e2 N |
235 | if (bio->bi_phys_segments) { |
236 | unsigned long flags; | |
237 | spin_lock_irqsave(&conf->device_lock, flags); | |
238 | bio->bi_phys_segments--; | |
239 | done = (bio->bi_phys_segments == 0); | |
240 | spin_unlock_irqrestore(&conf->device_lock, flags); | |
241 | } else | |
242 | done = 1; | |
243 | if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) | |
244 | clear_bit(BIO_UPTODATE, &bio->bi_flags); | |
245 | if (done) { | |
246 | bio_endio(bio, 0); | |
247 | /* | |
248 | * Wake up any possible resync thread that waits for the device | |
249 | * to go idle. | |
250 | */ | |
251 | allow_barrier(conf); | |
252 | } | |
1da177e4 LT |
253 | free_r10bio(r10_bio); |
254 | } | |
255 | ||
256 | /* | |
257 | * Update disk head position estimator based on IRQ completion info. | |
258 | */ | |
259 | static inline void update_head_pos(int slot, r10bio_t *r10_bio) | |
260 | { | |
070ec55d | 261 | conf_t *conf = r10_bio->mddev->private; |
1da177e4 LT |
262 | |
263 | conf->mirrors[r10_bio->devs[slot].devnum].head_position = | |
264 | r10_bio->devs[slot].addr + (r10_bio->sectors); | |
265 | } | |
266 | ||
778ca018 NK |
267 | /* |
268 | * Find the disk number which triggered given bio | |
269 | */ | |
749c55e9 N |
270 | static int find_bio_disk(conf_t *conf, r10bio_t *r10_bio, |
271 | struct bio *bio, int *slotp) | |
778ca018 NK |
272 | { |
273 | int slot; | |
274 | ||
275 | for (slot = 0; slot < conf->copies; slot++) | |
276 | if (r10_bio->devs[slot].bio == bio) | |
277 | break; | |
278 | ||
279 | BUG_ON(slot == conf->copies); | |
280 | update_head_pos(slot, r10_bio); | |
281 | ||
749c55e9 N |
282 | if (slotp) |
283 | *slotp = slot; | |
778ca018 NK |
284 | return r10_bio->devs[slot].devnum; |
285 | } | |
286 | ||
6712ecf8 | 287 | static void raid10_end_read_request(struct bio *bio, int error) |
1da177e4 LT |
288 | { |
289 | int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags); | |
7b92813c | 290 | r10bio_t *r10_bio = bio->bi_private; |
1da177e4 | 291 | int slot, dev; |
070ec55d | 292 | conf_t *conf = r10_bio->mddev->private; |
1da177e4 | 293 | |
1da177e4 LT |
294 | |
295 | slot = r10_bio->read_slot; | |
296 | dev = r10_bio->devs[slot].devnum; | |
297 | /* | |
298 | * this branch is our 'one mirror IO has finished' event handler: | |
299 | */ | |
4443ae10 N |
300 | update_head_pos(slot, r10_bio); |
301 | ||
302 | if (uptodate) { | |
1da177e4 LT |
303 | /* |
304 | * Set R10BIO_Uptodate in our master bio, so that | |
305 | * we will return a good error code to the higher | |
306 | * levels even if IO on some other mirrored buffer fails. | |
307 | * | |
308 | * The 'master' represents the composite IO operation to | |
309 | * user-side. So if something waits for IO, then it will | |
310 | * wait for the 'master' bio. | |
311 | */ | |
312 | set_bit(R10BIO_Uptodate, &r10_bio->state); | |
1da177e4 | 313 | raid_end_bio_io(r10_bio); |
7c4e06ff | 314 | rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev); |
4443ae10 | 315 | } else { |
1da177e4 | 316 | /* |
7c4e06ff | 317 | * oops, read error - keep the refcount on the rdev |
1da177e4 LT |
318 | */ |
319 | char b[BDEVNAME_SIZE]; | |
8bda470e CD |
320 | printk_ratelimited(KERN_ERR |
321 | "md/raid10:%s: %s: rescheduling sector %llu\n", | |
322 | mdname(conf->mddev), | |
323 | bdevname(conf->mirrors[dev].rdev->bdev, b), | |
324 | (unsigned long long)r10_bio->sector); | |
856e08e2 | 325 | set_bit(R10BIO_ReadError, &r10_bio->state); |
1da177e4 LT |
326 | reschedule_retry(r10_bio); |
327 | } | |
1da177e4 LT |
328 | } |
329 | ||
bd870a16 N |
330 | static void close_write(r10bio_t *r10_bio) |
331 | { | |
332 | /* clear the bitmap if all writes complete successfully */ | |
333 | bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector, | |
334 | r10_bio->sectors, | |
335 | !test_bit(R10BIO_Degraded, &r10_bio->state), | |
336 | 0); | |
337 | md_write_end(r10_bio->mddev); | |
338 | } | |
339 | ||
6712ecf8 | 340 | static void raid10_end_write_request(struct bio *bio, int error) |
1da177e4 LT |
341 | { |
342 | int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags); | |
7b92813c | 343 | r10bio_t *r10_bio = bio->bi_private; |
778ca018 | 344 | int dev; |
749c55e9 | 345 | int dec_rdev = 1; |
070ec55d | 346 | conf_t *conf = r10_bio->mddev->private; |
749c55e9 | 347 | int slot; |
1da177e4 | 348 | |
749c55e9 | 349 | dev = find_bio_disk(conf, r10_bio, bio, &slot); |
1da177e4 LT |
350 | |
351 | /* | |
352 | * this branch is our 'one mirror IO has finished' event handler: | |
353 | */ | |
6cce3b23 | 354 | if (!uptodate) { |
bd870a16 N |
355 | set_bit(WriteErrorSeen, &conf->mirrors[dev].rdev->flags); |
356 | set_bit(R10BIO_WriteError, &r10_bio->state); | |
357 | dec_rdev = 0; | |
749c55e9 | 358 | } else { |
1da177e4 LT |
359 | /* |
360 | * Set R10BIO_Uptodate in our master bio, so that | |
361 | * we will return a good error code for to the higher | |
362 | * levels even if IO on some other mirrored buffer fails. | |
363 | * | |
364 | * The 'master' represents the composite IO operation to | |
365 | * user-side. So if something waits for IO, then it will | |
366 | * wait for the 'master' bio. | |
367 | */ | |
749c55e9 N |
368 | sector_t first_bad; |
369 | int bad_sectors; | |
370 | ||
1da177e4 LT |
371 | set_bit(R10BIO_Uptodate, &r10_bio->state); |
372 | ||
749c55e9 N |
373 | /* Maybe we can clear some bad blocks. */ |
374 | if (is_badblock(conf->mirrors[dev].rdev, | |
375 | r10_bio->devs[slot].addr, | |
376 | r10_bio->sectors, | |
377 | &first_bad, &bad_sectors)) { | |
378 | bio_put(bio); | |
379 | r10_bio->devs[slot].bio = IO_MADE_GOOD; | |
380 | dec_rdev = 0; | |
381 | set_bit(R10BIO_MadeGood, &r10_bio->state); | |
382 | } | |
383 | } | |
384 | ||
1da177e4 LT |
385 | /* |
386 | * | |
387 | * Let's see if all mirrored write operations have finished | |
388 | * already. | |
389 | */ | |
390 | if (atomic_dec_and_test(&r10_bio->remaining)) { | |
bd870a16 | 391 | if (test_bit(R10BIO_WriteError, &r10_bio->state)) |
749c55e9 | 392 | reschedule_retry(r10_bio); |
bd870a16 N |
393 | else { |
394 | close_write(r10_bio); | |
395 | if (test_bit(R10BIO_MadeGood, &r10_bio->state)) | |
396 | reschedule_retry(r10_bio); | |
397 | else | |
398 | raid_end_bio_io(r10_bio); | |
399 | } | |
1da177e4 | 400 | } |
749c55e9 N |
401 | if (dec_rdev) |
402 | rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev); | |
1da177e4 LT |
403 | } |
404 | ||
405 | ||
406 | /* | |
407 | * RAID10 layout manager | |
25985edc | 408 | * As well as the chunksize and raid_disks count, there are two |
1da177e4 LT |
409 | * parameters: near_copies and far_copies. |
410 | * near_copies * far_copies must be <= raid_disks. | |
411 | * Normally one of these will be 1. | |
412 | * If both are 1, we get raid0. | |
413 | * If near_copies == raid_disks, we get raid1. | |
414 | * | |
25985edc | 415 | * Chunks are laid out in raid0 style with near_copies copies of the |
1da177e4 LT |
416 | * first chunk, followed by near_copies copies of the next chunk and |
417 | * so on. | |
418 | * If far_copies > 1, then after 1/far_copies of the array has been assigned | |
419 | * as described above, we start again with a device offset of near_copies. | |
420 | * So we effectively have another copy of the whole array further down all | |
421 | * the drives, but with blocks on different drives. | |
422 | * With this layout, and block is never stored twice on the one device. | |
423 | * | |
424 | * raid10_find_phys finds the sector offset of a given virtual sector | |
c93983bf | 425 | * on each device that it is on. |
1da177e4 LT |
426 | * |
427 | * raid10_find_virt does the reverse mapping, from a device and a | |
428 | * sector offset to a virtual address | |
429 | */ | |
430 | ||
431 | static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio) | |
432 | { | |
433 | int n,f; | |
434 | sector_t sector; | |
435 | sector_t chunk; | |
436 | sector_t stripe; | |
437 | int dev; | |
438 | ||
439 | int slot = 0; | |
440 | ||
441 | /* now calculate first sector/dev */ | |
442 | chunk = r10bio->sector >> conf->chunk_shift; | |
443 | sector = r10bio->sector & conf->chunk_mask; | |
444 | ||
445 | chunk *= conf->near_copies; | |
446 | stripe = chunk; | |
447 | dev = sector_div(stripe, conf->raid_disks); | |
c93983bf N |
448 | if (conf->far_offset) |
449 | stripe *= conf->far_copies; | |
1da177e4 LT |
450 | |
451 | sector += stripe << conf->chunk_shift; | |
452 | ||
453 | /* and calculate all the others */ | |
454 | for (n=0; n < conf->near_copies; n++) { | |
455 | int d = dev; | |
456 | sector_t s = sector; | |
457 | r10bio->devs[slot].addr = sector; | |
458 | r10bio->devs[slot].devnum = d; | |
459 | slot++; | |
460 | ||
461 | for (f = 1; f < conf->far_copies; f++) { | |
462 | d += conf->near_copies; | |
463 | if (d >= conf->raid_disks) | |
464 | d -= conf->raid_disks; | |
465 | s += conf->stride; | |
466 | r10bio->devs[slot].devnum = d; | |
467 | r10bio->devs[slot].addr = s; | |
468 | slot++; | |
469 | } | |
470 | dev++; | |
471 | if (dev >= conf->raid_disks) { | |
472 | dev = 0; | |
473 | sector += (conf->chunk_mask + 1); | |
474 | } | |
475 | } | |
476 | BUG_ON(slot != conf->copies); | |
477 | } | |
478 | ||
479 | static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev) | |
480 | { | |
481 | sector_t offset, chunk, vchunk; | |
482 | ||
1da177e4 | 483 | offset = sector & conf->chunk_mask; |
c93983bf N |
484 | if (conf->far_offset) { |
485 | int fc; | |
486 | chunk = sector >> conf->chunk_shift; | |
487 | fc = sector_div(chunk, conf->far_copies); | |
488 | dev -= fc * conf->near_copies; | |
489 | if (dev < 0) | |
490 | dev += conf->raid_disks; | |
491 | } else { | |
64a742bc | 492 | while (sector >= conf->stride) { |
c93983bf N |
493 | sector -= conf->stride; |
494 | if (dev < conf->near_copies) | |
495 | dev += conf->raid_disks - conf->near_copies; | |
496 | else | |
497 | dev -= conf->near_copies; | |
498 | } | |
499 | chunk = sector >> conf->chunk_shift; | |
500 | } | |
1da177e4 LT |
501 | vchunk = chunk * conf->raid_disks + dev; |
502 | sector_div(vchunk, conf->near_copies); | |
503 | return (vchunk << conf->chunk_shift) + offset; | |
504 | } | |
505 | ||
506 | /** | |
507 | * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged | |
508 | * @q: request queue | |
cc371e66 | 509 | * @bvm: properties of new bio |
1da177e4 LT |
510 | * @biovec: the request that could be merged to it. |
511 | * | |
512 | * Return amount of bytes we can accept at this offset | |
513 | * If near_copies == raid_disk, there are no striping issues, | |
514 | * but in that case, the function isn't called at all. | |
515 | */ | |
cc371e66 AK |
516 | static int raid10_mergeable_bvec(struct request_queue *q, |
517 | struct bvec_merge_data *bvm, | |
518 | struct bio_vec *biovec) | |
1da177e4 LT |
519 | { |
520 | mddev_t *mddev = q->queuedata; | |
cc371e66 | 521 | sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev); |
1da177e4 | 522 | int max; |
9d8f0363 | 523 | unsigned int chunk_sectors = mddev->chunk_sectors; |
cc371e66 | 524 | unsigned int bio_sectors = bvm->bi_size >> 9; |
1da177e4 LT |
525 | |
526 | max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9; | |
527 | if (max < 0) max = 0; /* bio_add cannot handle a negative return */ | |
cc371e66 AK |
528 | if (max <= biovec->bv_len && bio_sectors == 0) |
529 | return biovec->bv_len; | |
1da177e4 LT |
530 | else |
531 | return max; | |
532 | } | |
533 | ||
534 | /* | |
535 | * This routine returns the disk from which the requested read should | |
536 | * be done. There is a per-array 'next expected sequential IO' sector | |
537 | * number - if this matches on the next IO then we use the last disk. | |
538 | * There is also a per-disk 'last know head position' sector that is | |
539 | * maintained from IRQ contexts, both the normal and the resync IO | |
540 | * completion handlers update this position correctly. If there is no | |
541 | * perfect sequential match then we pick the disk whose head is closest. | |
542 | * | |
543 | * If there are 2 mirrors in the same 2 devices, performance degrades | |
544 | * because position is mirror, not device based. | |
545 | * | |
546 | * The rdev for the device selected will have nr_pending incremented. | |
547 | */ | |
548 | ||
549 | /* | |
550 | * FIXME: possibly should rethink readbalancing and do it differently | |
551 | * depending on near_copies / far_copies geometry. | |
552 | */ | |
856e08e2 | 553 | static int read_balance(conf_t *conf, r10bio_t *r10_bio, int *max_sectors) |
1da177e4 | 554 | { |
af3a2cd6 | 555 | const sector_t this_sector = r10_bio->sector; |
56d99121 | 556 | int disk, slot; |
856e08e2 N |
557 | int sectors = r10_bio->sectors; |
558 | int best_good_sectors; | |
56d99121 | 559 | sector_t new_distance, best_dist; |
d6065f7b | 560 | mdk_rdev_t *rdev; |
56d99121 N |
561 | int do_balance; |
562 | int best_slot; | |
1da177e4 LT |
563 | |
564 | raid10_find_phys(conf, r10_bio); | |
565 | rcu_read_lock(); | |
56d99121 | 566 | retry: |
856e08e2 | 567 | sectors = r10_bio->sectors; |
56d99121 N |
568 | best_slot = -1; |
569 | best_dist = MaxSector; | |
856e08e2 | 570 | best_good_sectors = 0; |
56d99121 | 571 | do_balance = 1; |
1da177e4 LT |
572 | /* |
573 | * Check if we can balance. We can balance on the whole | |
6cce3b23 N |
574 | * device if no resync is going on (recovery is ok), or below |
575 | * the resync window. We take the first readable disk when | |
576 | * above the resync window. | |
1da177e4 LT |
577 | */ |
578 | if (conf->mddev->recovery_cp < MaxSector | |
56d99121 N |
579 | && (this_sector + sectors >= conf->next_resync)) |
580 | do_balance = 0; | |
1da177e4 | 581 | |
56d99121 | 582 | for (slot = 0; slot < conf->copies ; slot++) { |
856e08e2 N |
583 | sector_t first_bad; |
584 | int bad_sectors; | |
585 | sector_t dev_sector; | |
586 | ||
56d99121 N |
587 | if (r10_bio->devs[slot].bio == IO_BLOCKED) |
588 | continue; | |
1da177e4 | 589 | disk = r10_bio->devs[slot].devnum; |
56d99121 N |
590 | rdev = rcu_dereference(conf->mirrors[disk].rdev); |
591 | if (rdev == NULL) | |
1da177e4 | 592 | continue; |
56d99121 N |
593 | if (!test_bit(In_sync, &rdev->flags)) |
594 | continue; | |
595 | ||
856e08e2 N |
596 | dev_sector = r10_bio->devs[slot].addr; |
597 | if (is_badblock(rdev, dev_sector, sectors, | |
598 | &first_bad, &bad_sectors)) { | |
599 | if (best_dist < MaxSector) | |
600 | /* Already have a better slot */ | |
601 | continue; | |
602 | if (first_bad <= dev_sector) { | |
603 | /* Cannot read here. If this is the | |
604 | * 'primary' device, then we must not read | |
605 | * beyond 'bad_sectors' from another device. | |
606 | */ | |
607 | bad_sectors -= (dev_sector - first_bad); | |
608 | if (!do_balance && sectors > bad_sectors) | |
609 | sectors = bad_sectors; | |
610 | if (best_good_sectors > sectors) | |
611 | best_good_sectors = sectors; | |
612 | } else { | |
613 | sector_t good_sectors = | |
614 | first_bad - dev_sector; | |
615 | if (good_sectors > best_good_sectors) { | |
616 | best_good_sectors = good_sectors; | |
617 | best_slot = slot; | |
618 | } | |
619 | if (!do_balance) | |
620 | /* Must read from here */ | |
621 | break; | |
622 | } | |
623 | continue; | |
624 | } else | |
625 | best_good_sectors = sectors; | |
626 | ||
56d99121 N |
627 | if (!do_balance) |
628 | break; | |
1da177e4 | 629 | |
22dfdf52 N |
630 | /* This optimisation is debatable, and completely destroys |
631 | * sequential read speed for 'far copies' arrays. So only | |
632 | * keep it for 'near' arrays, and review those later. | |
633 | */ | |
56d99121 | 634 | if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending)) |
1da177e4 | 635 | break; |
8ed3a195 KS |
636 | |
637 | /* for far > 1 always use the lowest address */ | |
638 | if (conf->far_copies > 1) | |
56d99121 | 639 | new_distance = r10_bio->devs[slot].addr; |
8ed3a195 | 640 | else |
56d99121 N |
641 | new_distance = abs(r10_bio->devs[slot].addr - |
642 | conf->mirrors[disk].head_position); | |
643 | if (new_distance < best_dist) { | |
644 | best_dist = new_distance; | |
645 | best_slot = slot; | |
1da177e4 LT |
646 | } |
647 | } | |
56d99121 N |
648 | if (slot == conf->copies) |
649 | slot = best_slot; | |
1da177e4 | 650 | |
56d99121 N |
651 | if (slot >= 0) { |
652 | disk = r10_bio->devs[slot].devnum; | |
653 | rdev = rcu_dereference(conf->mirrors[disk].rdev); | |
654 | if (!rdev) | |
655 | goto retry; | |
656 | atomic_inc(&rdev->nr_pending); | |
657 | if (test_bit(Faulty, &rdev->flags)) { | |
658 | /* Cannot risk returning a device that failed | |
659 | * before we inc'ed nr_pending | |
660 | */ | |
661 | rdev_dec_pending(rdev, conf->mddev); | |
662 | goto retry; | |
663 | } | |
664 | r10_bio->read_slot = slot; | |
665 | } else | |
29fc7e3e | 666 | disk = -1; |
1da177e4 | 667 | rcu_read_unlock(); |
856e08e2 | 668 | *max_sectors = best_good_sectors; |
1da177e4 LT |
669 | |
670 | return disk; | |
671 | } | |
672 | ||
0d129228 N |
673 | static int raid10_congested(void *data, int bits) |
674 | { | |
675 | mddev_t *mddev = data; | |
070ec55d | 676 | conf_t *conf = mddev->private; |
0d129228 N |
677 | int i, ret = 0; |
678 | ||
3fa841d7 N |
679 | if (mddev_congested(mddev, bits)) |
680 | return 1; | |
0d129228 | 681 | rcu_read_lock(); |
84707f38 | 682 | for (i = 0; i < conf->raid_disks && ret == 0; i++) { |
0d129228 N |
683 | mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev); |
684 | if (rdev && !test_bit(Faulty, &rdev->flags)) { | |
165125e1 | 685 | struct request_queue *q = bdev_get_queue(rdev->bdev); |
0d129228 N |
686 | |
687 | ret |= bdi_congested(&q->backing_dev_info, bits); | |
688 | } | |
689 | } | |
690 | rcu_read_unlock(); | |
691 | return ret; | |
692 | } | |
693 | ||
7eaceacc | 694 | static void flush_pending_writes(conf_t *conf) |
a35e63ef N |
695 | { |
696 | /* Any writes that have been queued but are awaiting | |
697 | * bitmap updates get flushed here. | |
a35e63ef | 698 | */ |
a35e63ef N |
699 | spin_lock_irq(&conf->device_lock); |
700 | ||
701 | if (conf->pending_bio_list.head) { | |
702 | struct bio *bio; | |
703 | bio = bio_list_get(&conf->pending_bio_list); | |
a35e63ef N |
704 | spin_unlock_irq(&conf->device_lock); |
705 | /* flush any pending bitmap writes to disk | |
706 | * before proceeding w/ I/O */ | |
707 | bitmap_unplug(conf->mddev->bitmap); | |
708 | ||
709 | while (bio) { /* submit pending writes */ | |
710 | struct bio *next = bio->bi_next; | |
711 | bio->bi_next = NULL; | |
712 | generic_make_request(bio); | |
713 | bio = next; | |
714 | } | |
a35e63ef N |
715 | } else |
716 | spin_unlock_irq(&conf->device_lock); | |
a35e63ef | 717 | } |
7eaceacc | 718 | |
0a27ec96 N |
719 | /* Barriers.... |
720 | * Sometimes we need to suspend IO while we do something else, | |
721 | * either some resync/recovery, or reconfigure the array. | |
722 | * To do this we raise a 'barrier'. | |
723 | * The 'barrier' is a counter that can be raised multiple times | |
724 | * to count how many activities are happening which preclude | |
725 | * normal IO. | |
726 | * We can only raise the barrier if there is no pending IO. | |
727 | * i.e. if nr_pending == 0. | |
728 | * We choose only to raise the barrier if no-one is waiting for the | |
729 | * barrier to go down. This means that as soon as an IO request | |
730 | * is ready, no other operations which require a barrier will start | |
731 | * until the IO request has had a chance. | |
732 | * | |
733 | * So: regular IO calls 'wait_barrier'. When that returns there | |
734 | * is no backgroup IO happening, It must arrange to call | |
735 | * allow_barrier when it has finished its IO. | |
736 | * backgroup IO calls must call raise_barrier. Once that returns | |
737 | * there is no normal IO happeing. It must arrange to call | |
738 | * lower_barrier when the particular background IO completes. | |
1da177e4 | 739 | */ |
1da177e4 | 740 | |
6cce3b23 | 741 | static void raise_barrier(conf_t *conf, int force) |
1da177e4 | 742 | { |
6cce3b23 | 743 | BUG_ON(force && !conf->barrier); |
1da177e4 | 744 | spin_lock_irq(&conf->resync_lock); |
0a27ec96 | 745 | |
6cce3b23 N |
746 | /* Wait until no block IO is waiting (unless 'force') */ |
747 | wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting, | |
c3b328ac | 748 | conf->resync_lock, ); |
0a27ec96 N |
749 | |
750 | /* block any new IO from starting */ | |
751 | conf->barrier++; | |
752 | ||
c3b328ac | 753 | /* Now wait for all pending IO to complete */ |
0a27ec96 N |
754 | wait_event_lock_irq(conf->wait_barrier, |
755 | !conf->nr_pending && conf->barrier < RESYNC_DEPTH, | |
c3b328ac | 756 | conf->resync_lock, ); |
0a27ec96 N |
757 | |
758 | spin_unlock_irq(&conf->resync_lock); | |
759 | } | |
760 | ||
761 | static void lower_barrier(conf_t *conf) | |
762 | { | |
763 | unsigned long flags; | |
764 | spin_lock_irqsave(&conf->resync_lock, flags); | |
765 | conf->barrier--; | |
766 | spin_unlock_irqrestore(&conf->resync_lock, flags); | |
767 | wake_up(&conf->wait_barrier); | |
768 | } | |
769 | ||
770 | static void wait_barrier(conf_t *conf) | |
771 | { | |
772 | spin_lock_irq(&conf->resync_lock); | |
773 | if (conf->barrier) { | |
774 | conf->nr_waiting++; | |
775 | wait_event_lock_irq(conf->wait_barrier, !conf->barrier, | |
776 | conf->resync_lock, | |
c3b328ac | 777 | ); |
0a27ec96 | 778 | conf->nr_waiting--; |
1da177e4 | 779 | } |
0a27ec96 | 780 | conf->nr_pending++; |
1da177e4 LT |
781 | spin_unlock_irq(&conf->resync_lock); |
782 | } | |
783 | ||
0a27ec96 N |
784 | static void allow_barrier(conf_t *conf) |
785 | { | |
786 | unsigned long flags; | |
787 | spin_lock_irqsave(&conf->resync_lock, flags); | |
788 | conf->nr_pending--; | |
789 | spin_unlock_irqrestore(&conf->resync_lock, flags); | |
790 | wake_up(&conf->wait_barrier); | |
791 | } | |
792 | ||
4443ae10 N |
793 | static void freeze_array(conf_t *conf) |
794 | { | |
795 | /* stop syncio and normal IO and wait for everything to | |
f188593e | 796 | * go quiet. |
4443ae10 | 797 | * We increment barrier and nr_waiting, and then |
1c830532 N |
798 | * wait until nr_pending match nr_queued+1 |
799 | * This is called in the context of one normal IO request | |
800 | * that has failed. Thus any sync request that might be pending | |
801 | * will be blocked by nr_pending, and we need to wait for | |
802 | * pending IO requests to complete or be queued for re-try. | |
803 | * Thus the number queued (nr_queued) plus this request (1) | |
804 | * must match the number of pending IOs (nr_pending) before | |
805 | * we continue. | |
4443ae10 N |
806 | */ |
807 | spin_lock_irq(&conf->resync_lock); | |
808 | conf->barrier++; | |
809 | conf->nr_waiting++; | |
810 | wait_event_lock_irq(conf->wait_barrier, | |
1c830532 | 811 | conf->nr_pending == conf->nr_queued+1, |
4443ae10 | 812 | conf->resync_lock, |
c3b328ac N |
813 | flush_pending_writes(conf)); |
814 | ||
4443ae10 N |
815 | spin_unlock_irq(&conf->resync_lock); |
816 | } | |
817 | ||
818 | static void unfreeze_array(conf_t *conf) | |
819 | { | |
820 | /* reverse the effect of the freeze */ | |
821 | spin_lock_irq(&conf->resync_lock); | |
822 | conf->barrier--; | |
823 | conf->nr_waiting--; | |
824 | wake_up(&conf->wait_barrier); | |
825 | spin_unlock_irq(&conf->resync_lock); | |
826 | } | |
827 | ||
21a52c6d | 828 | static int make_request(mddev_t *mddev, struct bio * bio) |
1da177e4 | 829 | { |
070ec55d | 830 | conf_t *conf = mddev->private; |
1da177e4 LT |
831 | mirror_info_t *mirror; |
832 | r10bio_t *r10_bio; | |
833 | struct bio *read_bio; | |
834 | int i; | |
835 | int chunk_sects = conf->chunk_mask + 1; | |
a362357b | 836 | const int rw = bio_data_dir(bio); |
2c7d46ec | 837 | const unsigned long do_sync = (bio->bi_rw & REQ_SYNC); |
e9c7469b | 838 | const unsigned long do_fua = (bio->bi_rw & REQ_FUA); |
6cce3b23 | 839 | unsigned long flags; |
6bfe0b49 | 840 | mdk_rdev_t *blocked_rdev; |
c3b328ac | 841 | int plugged; |
d4432c23 N |
842 | int sectors_handled; |
843 | int max_sectors; | |
1da177e4 | 844 | |
e9c7469b TH |
845 | if (unlikely(bio->bi_rw & REQ_FLUSH)) { |
846 | md_flush_request(mddev, bio); | |
e5dcdd80 N |
847 | return 0; |
848 | } | |
849 | ||
1da177e4 LT |
850 | /* If this request crosses a chunk boundary, we need to |
851 | * split it. This will only happen for 1 PAGE (or less) requests. | |
852 | */ | |
853 | if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9) | |
854 | > chunk_sects && | |
855 | conf->near_copies < conf->raid_disks)) { | |
856 | struct bio_pair *bp; | |
857 | /* Sanity check -- queue functions should prevent this happening */ | |
858 | if (bio->bi_vcnt != 1 || | |
859 | bio->bi_idx != 0) | |
860 | goto bad_map; | |
861 | /* This is a one page bio that upper layers | |
862 | * refuse to split for us, so we need to split it. | |
863 | */ | |
6feef531 | 864 | bp = bio_split(bio, |
1da177e4 | 865 | chunk_sects - (bio->bi_sector & (chunk_sects - 1)) ); |
51e9ac77 N |
866 | |
867 | /* Each of these 'make_request' calls will call 'wait_barrier'. | |
868 | * If the first succeeds but the second blocks due to the resync | |
869 | * thread raising the barrier, we will deadlock because the | |
870 | * IO to the underlying device will be queued in generic_make_request | |
871 | * and will never complete, so will never reduce nr_pending. | |
872 | * So increment nr_waiting here so no new raise_barriers will | |
873 | * succeed, and so the second wait_barrier cannot block. | |
874 | */ | |
875 | spin_lock_irq(&conf->resync_lock); | |
876 | conf->nr_waiting++; | |
877 | spin_unlock_irq(&conf->resync_lock); | |
878 | ||
21a52c6d | 879 | if (make_request(mddev, &bp->bio1)) |
1da177e4 | 880 | generic_make_request(&bp->bio1); |
21a52c6d | 881 | if (make_request(mddev, &bp->bio2)) |
1da177e4 LT |
882 | generic_make_request(&bp->bio2); |
883 | ||
51e9ac77 N |
884 | spin_lock_irq(&conf->resync_lock); |
885 | conf->nr_waiting--; | |
886 | wake_up(&conf->wait_barrier); | |
887 | spin_unlock_irq(&conf->resync_lock); | |
888 | ||
1da177e4 LT |
889 | bio_pair_release(bp); |
890 | return 0; | |
891 | bad_map: | |
128595ed N |
892 | printk("md/raid10:%s: make_request bug: can't convert block across chunks" |
893 | " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2, | |
1da177e4 LT |
894 | (unsigned long long)bio->bi_sector, bio->bi_size >> 10); |
895 | ||
6712ecf8 | 896 | bio_io_error(bio); |
1da177e4 LT |
897 | return 0; |
898 | } | |
899 | ||
3d310eb7 | 900 | md_write_start(mddev, bio); |
06d91a5f | 901 | |
1da177e4 LT |
902 | /* |
903 | * Register the new request and wait if the reconstruction | |
904 | * thread has put up a bar for new requests. | |
905 | * Continue immediately if no resync is active currently. | |
906 | */ | |
0a27ec96 | 907 | wait_barrier(conf); |
1da177e4 | 908 | |
1da177e4 LT |
909 | r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO); |
910 | ||
911 | r10_bio->master_bio = bio; | |
912 | r10_bio->sectors = bio->bi_size >> 9; | |
913 | ||
914 | r10_bio->mddev = mddev; | |
915 | r10_bio->sector = bio->bi_sector; | |
6cce3b23 | 916 | r10_bio->state = 0; |
1da177e4 | 917 | |
856e08e2 N |
918 | /* We might need to issue multiple reads to different |
919 | * devices if there are bad blocks around, so we keep | |
920 | * track of the number of reads in bio->bi_phys_segments. | |
921 | * If this is 0, there is only one r10_bio and no locking | |
922 | * will be needed when the request completes. If it is | |
923 | * non-zero, then it is the number of not-completed requests. | |
924 | */ | |
925 | bio->bi_phys_segments = 0; | |
926 | clear_bit(BIO_SEG_VALID, &bio->bi_flags); | |
927 | ||
a362357b | 928 | if (rw == READ) { |
1da177e4 LT |
929 | /* |
930 | * read balancing logic: | |
931 | */ | |
856e08e2 N |
932 | int disk; |
933 | int slot; | |
934 | ||
935 | read_again: | |
936 | disk = read_balance(conf, r10_bio, &max_sectors); | |
937 | slot = r10_bio->read_slot; | |
1da177e4 LT |
938 | if (disk < 0) { |
939 | raid_end_bio_io(r10_bio); | |
940 | return 0; | |
941 | } | |
942 | mirror = conf->mirrors + disk; | |
943 | ||
a167f663 | 944 | read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev); |
856e08e2 N |
945 | md_trim_bio(read_bio, r10_bio->sector - bio->bi_sector, |
946 | max_sectors); | |
1da177e4 LT |
947 | |
948 | r10_bio->devs[slot].bio = read_bio; | |
949 | ||
950 | read_bio->bi_sector = r10_bio->devs[slot].addr + | |
951 | mirror->rdev->data_offset; | |
952 | read_bio->bi_bdev = mirror->rdev->bdev; | |
953 | read_bio->bi_end_io = raid10_end_read_request; | |
7b6d91da | 954 | read_bio->bi_rw = READ | do_sync; |
1da177e4 LT |
955 | read_bio->bi_private = r10_bio; |
956 | ||
856e08e2 N |
957 | if (max_sectors < r10_bio->sectors) { |
958 | /* Could not read all from this device, so we will | |
959 | * need another r10_bio. | |
960 | */ | |
856e08e2 N |
961 | sectors_handled = (r10_bio->sectors + max_sectors |
962 | - bio->bi_sector); | |
963 | r10_bio->sectors = max_sectors; | |
964 | spin_lock_irq(&conf->device_lock); | |
965 | if (bio->bi_phys_segments == 0) | |
966 | bio->bi_phys_segments = 2; | |
967 | else | |
968 | bio->bi_phys_segments++; | |
969 | spin_unlock(&conf->device_lock); | |
970 | /* Cannot call generic_make_request directly | |
971 | * as that will be queued in __generic_make_request | |
972 | * and subsequent mempool_alloc might block | |
973 | * waiting for it. so hand bio over to raid10d. | |
974 | */ | |
975 | reschedule_retry(r10_bio); | |
976 | ||
977 | r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO); | |
978 | ||
979 | r10_bio->master_bio = bio; | |
980 | r10_bio->sectors = ((bio->bi_size >> 9) | |
981 | - sectors_handled); | |
982 | r10_bio->state = 0; | |
983 | r10_bio->mddev = mddev; | |
984 | r10_bio->sector = bio->bi_sector + sectors_handled; | |
985 | goto read_again; | |
986 | } else | |
987 | generic_make_request(read_bio); | |
1da177e4 LT |
988 | return 0; |
989 | } | |
990 | ||
991 | /* | |
992 | * WRITE: | |
993 | */ | |
6bfe0b49 | 994 | /* first select target devices under rcu_lock and |
1da177e4 LT |
995 | * inc refcount on their rdev. Record them by setting |
996 | * bios[x] to bio | |
d4432c23 N |
997 | * If there are known/acknowledged bad blocks on any device |
998 | * on which we have seen a write error, we want to avoid | |
999 | * writing to those blocks. This potentially requires several | |
1000 | * writes to write around the bad blocks. Each set of writes | |
1001 | * gets its own r10_bio with a set of bios attached. The number | |
1002 | * of r10_bios is recored in bio->bi_phys_segments just as with | |
1003 | * the read case. | |
1da177e4 | 1004 | */ |
c3b328ac N |
1005 | plugged = mddev_check_plugged(mddev); |
1006 | ||
1da177e4 | 1007 | raid10_find_phys(conf, r10_bio); |
d4432c23 | 1008 | retry_write: |
cb6969e8 | 1009 | blocked_rdev = NULL; |
1da177e4 | 1010 | rcu_read_lock(); |
d4432c23 N |
1011 | max_sectors = r10_bio->sectors; |
1012 | ||
1da177e4 LT |
1013 | for (i = 0; i < conf->copies; i++) { |
1014 | int d = r10_bio->devs[i].devnum; | |
d6065f7b | 1015 | mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev); |
6bfe0b49 DW |
1016 | if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) { |
1017 | atomic_inc(&rdev->nr_pending); | |
1018 | blocked_rdev = rdev; | |
1019 | break; | |
1020 | } | |
d4432c23 N |
1021 | r10_bio->devs[i].bio = NULL; |
1022 | if (!rdev || test_bit(Faulty, &rdev->flags)) { | |
6cce3b23 | 1023 | set_bit(R10BIO_Degraded, &r10_bio->state); |
d4432c23 N |
1024 | continue; |
1025 | } | |
1026 | if (test_bit(WriteErrorSeen, &rdev->flags)) { | |
1027 | sector_t first_bad; | |
1028 | sector_t dev_sector = r10_bio->devs[i].addr; | |
1029 | int bad_sectors; | |
1030 | int is_bad; | |
1031 | ||
1032 | is_bad = is_badblock(rdev, dev_sector, | |
1033 | max_sectors, | |
1034 | &first_bad, &bad_sectors); | |
1035 | if (is_bad < 0) { | |
1036 | /* Mustn't write here until the bad block | |
1037 | * is acknowledged | |
1038 | */ | |
1039 | atomic_inc(&rdev->nr_pending); | |
1040 | set_bit(BlockedBadBlocks, &rdev->flags); | |
1041 | blocked_rdev = rdev; | |
1042 | break; | |
1043 | } | |
1044 | if (is_bad && first_bad <= dev_sector) { | |
1045 | /* Cannot write here at all */ | |
1046 | bad_sectors -= (dev_sector - first_bad); | |
1047 | if (bad_sectors < max_sectors) | |
1048 | /* Mustn't write more than bad_sectors | |
1049 | * to other devices yet | |
1050 | */ | |
1051 | max_sectors = bad_sectors; | |
1052 | /* We don't set R10BIO_Degraded as that | |
1053 | * only applies if the disk is missing, | |
1054 | * so it might be re-added, and we want to | |
1055 | * know to recover this chunk. | |
1056 | * In this case the device is here, and the | |
1057 | * fact that this chunk is not in-sync is | |
1058 | * recorded in the bad block log. | |
1059 | */ | |
1060 | continue; | |
1061 | } | |
1062 | if (is_bad) { | |
1063 | int good_sectors = first_bad - dev_sector; | |
1064 | if (good_sectors < max_sectors) | |
1065 | max_sectors = good_sectors; | |
1066 | } | |
6cce3b23 | 1067 | } |
d4432c23 N |
1068 | r10_bio->devs[i].bio = bio; |
1069 | atomic_inc(&rdev->nr_pending); | |
1da177e4 LT |
1070 | } |
1071 | rcu_read_unlock(); | |
1072 | ||
6bfe0b49 DW |
1073 | if (unlikely(blocked_rdev)) { |
1074 | /* Have to wait for this device to get unblocked, then retry */ | |
1075 | int j; | |
1076 | int d; | |
1077 | ||
1078 | for (j = 0; j < i; j++) | |
1079 | if (r10_bio->devs[j].bio) { | |
1080 | d = r10_bio->devs[j].devnum; | |
1081 | rdev_dec_pending(conf->mirrors[d].rdev, mddev); | |
1082 | } | |
1083 | allow_barrier(conf); | |
1084 | md_wait_for_blocked_rdev(blocked_rdev, mddev); | |
1085 | wait_barrier(conf); | |
1086 | goto retry_write; | |
1087 | } | |
1088 | ||
d4432c23 N |
1089 | if (max_sectors < r10_bio->sectors) { |
1090 | /* We are splitting this into multiple parts, so | |
1091 | * we need to prepare for allocating another r10_bio. | |
1092 | */ | |
1093 | r10_bio->sectors = max_sectors; | |
1094 | spin_lock_irq(&conf->device_lock); | |
1095 | if (bio->bi_phys_segments == 0) | |
1096 | bio->bi_phys_segments = 2; | |
1097 | else | |
1098 | bio->bi_phys_segments++; | |
1099 | spin_unlock_irq(&conf->device_lock); | |
1100 | } | |
1101 | sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector; | |
1102 | ||
4e78064f | 1103 | atomic_set(&r10_bio->remaining, 1); |
d4432c23 | 1104 | bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0); |
06d91a5f | 1105 | |
1da177e4 LT |
1106 | for (i = 0; i < conf->copies; i++) { |
1107 | struct bio *mbio; | |
1108 | int d = r10_bio->devs[i].devnum; | |
1109 | if (!r10_bio->devs[i].bio) | |
1110 | continue; | |
1111 | ||
a167f663 | 1112 | mbio = bio_clone_mddev(bio, GFP_NOIO, mddev); |
d4432c23 N |
1113 | md_trim_bio(mbio, r10_bio->sector - bio->bi_sector, |
1114 | max_sectors); | |
1da177e4 LT |
1115 | r10_bio->devs[i].bio = mbio; |
1116 | ||
d4432c23 N |
1117 | mbio->bi_sector = (r10_bio->devs[i].addr+ |
1118 | conf->mirrors[d].rdev->data_offset); | |
1da177e4 LT |
1119 | mbio->bi_bdev = conf->mirrors[d].rdev->bdev; |
1120 | mbio->bi_end_io = raid10_end_write_request; | |
e9c7469b | 1121 | mbio->bi_rw = WRITE | do_sync | do_fua; |
1da177e4 LT |
1122 | mbio->bi_private = r10_bio; |
1123 | ||
1124 | atomic_inc(&r10_bio->remaining); | |
4e78064f N |
1125 | spin_lock_irqsave(&conf->device_lock, flags); |
1126 | bio_list_add(&conf->pending_bio_list, mbio); | |
4e78064f | 1127 | spin_unlock_irqrestore(&conf->device_lock, flags); |
1da177e4 LT |
1128 | } |
1129 | ||
4e78064f N |
1130 | if (atomic_dec_and_test(&r10_bio->remaining)) { |
1131 | /* This matches the end of raid10_end_write_request() */ | |
1132 | bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector, | |
1133 | r10_bio->sectors, | |
1134 | !test_bit(R10BIO_Degraded, &r10_bio->state), | |
1135 | 0); | |
f6f953aa AR |
1136 | md_write_end(mddev); |
1137 | raid_end_bio_io(r10_bio); | |
f6f953aa AR |
1138 | } |
1139 | ||
a35e63ef N |
1140 | /* In case raid10d snuck in to freeze_array */ |
1141 | wake_up(&conf->wait_barrier); | |
1142 | ||
d4432c23 N |
1143 | if (sectors_handled < (bio->bi_size >> 9)) { |
1144 | /* We need another r1_bio. It has already been counted | |
1145 | * in bio->bi_phys_segments. | |
1146 | */ | |
1147 | r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO); | |
1148 | ||
1149 | r10_bio->master_bio = bio; | |
1150 | r10_bio->sectors = (bio->bi_size >> 9) - sectors_handled; | |
1151 | ||
1152 | r10_bio->mddev = mddev; | |
1153 | r10_bio->sector = bio->bi_sector + sectors_handled; | |
1154 | r10_bio->state = 0; | |
1155 | goto retry_write; | |
1156 | } | |
1157 | ||
c3b328ac | 1158 | if (do_sync || !mddev->bitmap || !plugged) |
e3881a68 | 1159 | md_wakeup_thread(mddev->thread); |
1da177e4 LT |
1160 | return 0; |
1161 | } | |
1162 | ||
1163 | static void status(struct seq_file *seq, mddev_t *mddev) | |
1164 | { | |
070ec55d | 1165 | conf_t *conf = mddev->private; |
1da177e4 LT |
1166 | int i; |
1167 | ||
1168 | if (conf->near_copies < conf->raid_disks) | |
9d8f0363 | 1169 | seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2); |
1da177e4 LT |
1170 | if (conf->near_copies > 1) |
1171 | seq_printf(seq, " %d near-copies", conf->near_copies); | |
c93983bf N |
1172 | if (conf->far_copies > 1) { |
1173 | if (conf->far_offset) | |
1174 | seq_printf(seq, " %d offset-copies", conf->far_copies); | |
1175 | else | |
1176 | seq_printf(seq, " %d far-copies", conf->far_copies); | |
1177 | } | |
1da177e4 | 1178 | seq_printf(seq, " [%d/%d] [", conf->raid_disks, |
76186dd8 | 1179 | conf->raid_disks - mddev->degraded); |
1da177e4 LT |
1180 | for (i = 0; i < conf->raid_disks; i++) |
1181 | seq_printf(seq, "%s", | |
1182 | conf->mirrors[i].rdev && | |
b2d444d7 | 1183 | test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_"); |
1da177e4 LT |
1184 | seq_printf(seq, "]"); |
1185 | } | |
1186 | ||
700c7213 N |
1187 | /* check if there are enough drives for |
1188 | * every block to appear on atleast one. | |
1189 | * Don't consider the device numbered 'ignore' | |
1190 | * as we might be about to remove it. | |
1191 | */ | |
1192 | static int enough(conf_t *conf, int ignore) | |
1193 | { | |
1194 | int first = 0; | |
1195 | ||
1196 | do { | |
1197 | int n = conf->copies; | |
1198 | int cnt = 0; | |
1199 | while (n--) { | |
1200 | if (conf->mirrors[first].rdev && | |
1201 | first != ignore) | |
1202 | cnt++; | |
1203 | first = (first+1) % conf->raid_disks; | |
1204 | } | |
1205 | if (cnt == 0) | |
1206 | return 0; | |
1207 | } while (first != 0); | |
1208 | return 1; | |
1209 | } | |
1210 | ||
1da177e4 LT |
1211 | static void error(mddev_t *mddev, mdk_rdev_t *rdev) |
1212 | { | |
1213 | char b[BDEVNAME_SIZE]; | |
070ec55d | 1214 | conf_t *conf = mddev->private; |
1da177e4 LT |
1215 | |
1216 | /* | |
1217 | * If it is not operational, then we have already marked it as dead | |
1218 | * else if it is the last working disks, ignore the error, let the | |
1219 | * next level up know. | |
1220 | * else mark the drive as failed | |
1221 | */ | |
b2d444d7 | 1222 | if (test_bit(In_sync, &rdev->flags) |
700c7213 | 1223 | && !enough(conf, rdev->raid_disk)) |
1da177e4 LT |
1224 | /* |
1225 | * Don't fail the drive, just return an IO error. | |
1da177e4 LT |
1226 | */ |
1227 | return; | |
c04be0aa N |
1228 | if (test_and_clear_bit(In_sync, &rdev->flags)) { |
1229 | unsigned long flags; | |
1230 | spin_lock_irqsave(&conf->device_lock, flags); | |
1da177e4 | 1231 | mddev->degraded++; |
c04be0aa | 1232 | spin_unlock_irqrestore(&conf->device_lock, flags); |
1da177e4 LT |
1233 | /* |
1234 | * if recovery is running, make sure it aborts. | |
1235 | */ | |
dfc70645 | 1236 | set_bit(MD_RECOVERY_INTR, &mddev->recovery); |
1da177e4 | 1237 | } |
de393cde | 1238 | set_bit(Blocked, &rdev->flags); |
b2d444d7 | 1239 | set_bit(Faulty, &rdev->flags); |
850b2b42 | 1240 | set_bit(MD_CHANGE_DEVS, &mddev->flags); |
067032bc JP |
1241 | printk(KERN_ALERT |
1242 | "md/raid10:%s: Disk failure on %s, disabling device.\n" | |
1243 | "md/raid10:%s: Operation continuing on %d devices.\n", | |
128595ed N |
1244 | mdname(mddev), bdevname(rdev->bdev, b), |
1245 | mdname(mddev), conf->raid_disks - mddev->degraded); | |
1da177e4 LT |
1246 | } |
1247 | ||
1248 | static void print_conf(conf_t *conf) | |
1249 | { | |
1250 | int i; | |
1251 | mirror_info_t *tmp; | |
1252 | ||
128595ed | 1253 | printk(KERN_DEBUG "RAID10 conf printout:\n"); |
1da177e4 | 1254 | if (!conf) { |
128595ed | 1255 | printk(KERN_DEBUG "(!conf)\n"); |
1da177e4 LT |
1256 | return; |
1257 | } | |
128595ed | 1258 | printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded, |
1da177e4 LT |
1259 | conf->raid_disks); |
1260 | ||
1261 | for (i = 0; i < conf->raid_disks; i++) { | |
1262 | char b[BDEVNAME_SIZE]; | |
1263 | tmp = conf->mirrors + i; | |
1264 | if (tmp->rdev) | |
128595ed | 1265 | printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n", |
b2d444d7 N |
1266 | i, !test_bit(In_sync, &tmp->rdev->flags), |
1267 | !test_bit(Faulty, &tmp->rdev->flags), | |
1da177e4 LT |
1268 | bdevname(tmp->rdev->bdev,b)); |
1269 | } | |
1270 | } | |
1271 | ||
1272 | static void close_sync(conf_t *conf) | |
1273 | { | |
0a27ec96 N |
1274 | wait_barrier(conf); |
1275 | allow_barrier(conf); | |
1da177e4 LT |
1276 | |
1277 | mempool_destroy(conf->r10buf_pool); | |
1278 | conf->r10buf_pool = NULL; | |
1279 | } | |
1280 | ||
1281 | static int raid10_spare_active(mddev_t *mddev) | |
1282 | { | |
1283 | int i; | |
1284 | conf_t *conf = mddev->private; | |
1285 | mirror_info_t *tmp; | |
6b965620 N |
1286 | int count = 0; |
1287 | unsigned long flags; | |
1da177e4 LT |
1288 | |
1289 | /* | |
1290 | * Find all non-in_sync disks within the RAID10 configuration | |
1291 | * and mark them in_sync | |
1292 | */ | |
1293 | for (i = 0; i < conf->raid_disks; i++) { | |
1294 | tmp = conf->mirrors + i; | |
1295 | if (tmp->rdev | |
b2d444d7 | 1296 | && !test_bit(Faulty, &tmp->rdev->flags) |
c04be0aa | 1297 | && !test_and_set_bit(In_sync, &tmp->rdev->flags)) { |
6b965620 | 1298 | count++; |
e6ffbcb6 | 1299 | sysfs_notify_dirent(tmp->rdev->sysfs_state); |
1da177e4 LT |
1300 | } |
1301 | } | |
6b965620 N |
1302 | spin_lock_irqsave(&conf->device_lock, flags); |
1303 | mddev->degraded -= count; | |
1304 | spin_unlock_irqrestore(&conf->device_lock, flags); | |
1da177e4 LT |
1305 | |
1306 | print_conf(conf); | |
6b965620 | 1307 | return count; |
1da177e4 LT |
1308 | } |
1309 | ||
1310 | ||
1311 | static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev) | |
1312 | { | |
1313 | conf_t *conf = mddev->private; | |
199050ea | 1314 | int err = -EEXIST; |
1da177e4 | 1315 | int mirror; |
6c2fce2e | 1316 | int first = 0; |
84707f38 | 1317 | int last = conf->raid_disks - 1; |
1da177e4 LT |
1318 | |
1319 | if (mddev->recovery_cp < MaxSector) | |
1320 | /* only hot-add to in-sync arrays, as recovery is | |
1321 | * very different from resync | |
1322 | */ | |
199050ea | 1323 | return -EBUSY; |
700c7213 | 1324 | if (!enough(conf, -1)) |
199050ea | 1325 | return -EINVAL; |
1da177e4 | 1326 | |
a53a6c85 | 1327 | if (rdev->raid_disk >= 0) |
6c2fce2e | 1328 | first = last = rdev->raid_disk; |
1da177e4 | 1329 | |
2c4193df | 1330 | if (rdev->saved_raid_disk >= first && |
6cce3b23 N |
1331 | conf->mirrors[rdev->saved_raid_disk].rdev == NULL) |
1332 | mirror = rdev->saved_raid_disk; | |
1333 | else | |
6c2fce2e | 1334 | mirror = first; |
2bb77736 N |
1335 | for ( ; mirror <= last ; mirror++) { |
1336 | mirror_info_t *p = &conf->mirrors[mirror]; | |
1337 | if (p->recovery_disabled == mddev->recovery_disabled) | |
1338 | continue; | |
1339 | if (!p->rdev) | |
1340 | continue; | |
1da177e4 | 1341 | |
2bb77736 N |
1342 | disk_stack_limits(mddev->gendisk, rdev->bdev, |
1343 | rdev->data_offset << 9); | |
1344 | /* as we don't honour merge_bvec_fn, we must | |
1345 | * never risk violating it, so limit | |
1346 | * ->max_segments to one lying with a single | |
1347 | * page, as a one page request is never in | |
1348 | * violation. | |
1349 | */ | |
1350 | if (rdev->bdev->bd_disk->queue->merge_bvec_fn) { | |
1351 | blk_queue_max_segments(mddev->queue, 1); | |
1352 | blk_queue_segment_boundary(mddev->queue, | |
1353 | PAGE_CACHE_SIZE - 1); | |
1da177e4 LT |
1354 | } |
1355 | ||
2bb77736 N |
1356 | p->head_position = 0; |
1357 | rdev->raid_disk = mirror; | |
1358 | err = 0; | |
1359 | if (rdev->saved_raid_disk != mirror) | |
1360 | conf->fullsync = 1; | |
1361 | rcu_assign_pointer(p->rdev, rdev); | |
1362 | break; | |
1363 | } | |
1364 | ||
ac5e7113 | 1365 | md_integrity_add_rdev(rdev, mddev); |
1da177e4 | 1366 | print_conf(conf); |
199050ea | 1367 | return err; |
1da177e4 LT |
1368 | } |
1369 | ||
1370 | static int raid10_remove_disk(mddev_t *mddev, int number) | |
1371 | { | |
1372 | conf_t *conf = mddev->private; | |
1373 | int err = 0; | |
1374 | mdk_rdev_t *rdev; | |
1375 | mirror_info_t *p = conf->mirrors+ number; | |
1376 | ||
1377 | print_conf(conf); | |
1378 | rdev = p->rdev; | |
1379 | if (rdev) { | |
b2d444d7 | 1380 | if (test_bit(In_sync, &rdev->flags) || |
1da177e4 LT |
1381 | atomic_read(&rdev->nr_pending)) { |
1382 | err = -EBUSY; | |
1383 | goto abort; | |
1384 | } | |
dfc70645 N |
1385 | /* Only remove faulty devices in recovery |
1386 | * is not possible. | |
1387 | */ | |
1388 | if (!test_bit(Faulty, &rdev->flags) && | |
2bb77736 | 1389 | mddev->recovery_disabled != p->recovery_disabled && |
700c7213 | 1390 | enough(conf, -1)) { |
dfc70645 N |
1391 | err = -EBUSY; |
1392 | goto abort; | |
1393 | } | |
1da177e4 | 1394 | p->rdev = NULL; |
fbd568a3 | 1395 | synchronize_rcu(); |
1da177e4 LT |
1396 | if (atomic_read(&rdev->nr_pending)) { |
1397 | /* lost the race, try later */ | |
1398 | err = -EBUSY; | |
1399 | p->rdev = rdev; | |
ac5e7113 | 1400 | goto abort; |
1da177e4 | 1401 | } |
a91a2785 | 1402 | err = md_integrity_register(mddev); |
1da177e4 LT |
1403 | } |
1404 | abort: | |
1405 | ||
1406 | print_conf(conf); | |
1407 | return err; | |
1408 | } | |
1409 | ||
1410 | ||
6712ecf8 | 1411 | static void end_sync_read(struct bio *bio, int error) |
1da177e4 | 1412 | { |
7b92813c | 1413 | r10bio_t *r10_bio = bio->bi_private; |
070ec55d | 1414 | conf_t *conf = r10_bio->mddev->private; |
778ca018 | 1415 | int d; |
1da177e4 | 1416 | |
749c55e9 | 1417 | d = find_bio_disk(conf, r10_bio, bio, NULL); |
0eb3ff12 N |
1418 | |
1419 | if (test_bit(BIO_UPTODATE, &bio->bi_flags)) | |
1420 | set_bit(R10BIO_Uptodate, &r10_bio->state); | |
4dbcdc75 N |
1421 | else { |
1422 | atomic_add(r10_bio->sectors, | |
1423 | &conf->mirrors[d].rdev->corrected_errors); | |
1424 | if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery)) | |
1425 | md_error(r10_bio->mddev, | |
1426 | conf->mirrors[d].rdev); | |
1427 | } | |
1da177e4 LT |
1428 | |
1429 | /* for reconstruct, we always reschedule after a read. | |
1430 | * for resync, only after all reads | |
1431 | */ | |
73d5c38a | 1432 | rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev); |
1da177e4 LT |
1433 | if (test_bit(R10BIO_IsRecover, &r10_bio->state) || |
1434 | atomic_dec_and_test(&r10_bio->remaining)) { | |
1435 | /* we have read all the blocks, | |
1436 | * do the comparison in process context in raid10d | |
1437 | */ | |
1438 | reschedule_retry(r10_bio); | |
1439 | } | |
1da177e4 LT |
1440 | } |
1441 | ||
6712ecf8 | 1442 | static void end_sync_write(struct bio *bio, int error) |
1da177e4 LT |
1443 | { |
1444 | int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags); | |
7b92813c | 1445 | r10bio_t *r10_bio = bio->bi_private; |
1da177e4 | 1446 | mddev_t *mddev = r10_bio->mddev; |
070ec55d | 1447 | conf_t *conf = mddev->private; |
778ca018 | 1448 | int d; |
749c55e9 N |
1449 | sector_t first_bad; |
1450 | int bad_sectors; | |
1451 | int slot; | |
1da177e4 | 1452 | |
749c55e9 | 1453 | d = find_bio_disk(conf, r10_bio, bio, &slot); |
1da177e4 LT |
1454 | |
1455 | if (!uptodate) | |
1456 | md_error(mddev, conf->mirrors[d].rdev); | |
749c55e9 N |
1457 | else if (is_badblock(conf->mirrors[d].rdev, |
1458 | r10_bio->devs[slot].addr, | |
1459 | r10_bio->sectors, | |
1460 | &first_bad, &bad_sectors)) | |
1461 | set_bit(R10BIO_MadeGood, &r10_bio->state); | |
dfc70645 | 1462 | |
73d5c38a | 1463 | rdev_dec_pending(conf->mirrors[d].rdev, mddev); |
1da177e4 LT |
1464 | while (atomic_dec_and_test(&r10_bio->remaining)) { |
1465 | if (r10_bio->master_bio == NULL) { | |
1466 | /* the primary of several recovery bios */ | |
73d5c38a | 1467 | sector_t s = r10_bio->sectors; |
749c55e9 N |
1468 | if (test_bit(R10BIO_MadeGood, &r10_bio->state)) |
1469 | reschedule_retry(r10_bio); | |
1470 | else | |
1471 | put_buf(r10_bio); | |
73d5c38a | 1472 | md_done_sync(mddev, s, 1); |
1da177e4 LT |
1473 | break; |
1474 | } else { | |
1475 | r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio; | |
749c55e9 N |
1476 | if (test_bit(R10BIO_MadeGood, &r10_bio->state)) |
1477 | reschedule_retry(r10_bio); | |
1478 | else | |
1479 | put_buf(r10_bio); | |
1da177e4 LT |
1480 | r10_bio = r10_bio2; |
1481 | } | |
1482 | } | |
1da177e4 LT |
1483 | } |
1484 | ||
1485 | /* | |
1486 | * Note: sync and recover and handled very differently for raid10 | |
1487 | * This code is for resync. | |
1488 | * For resync, we read through virtual addresses and read all blocks. | |
1489 | * If there is any error, we schedule a write. The lowest numbered | |
1490 | * drive is authoritative. | |
1491 | * However requests come for physical address, so we need to map. | |
1492 | * For every physical address there are raid_disks/copies virtual addresses, | |
1493 | * which is always are least one, but is not necessarly an integer. | |
1494 | * This means that a physical address can span multiple chunks, so we may | |
1495 | * have to submit multiple io requests for a single sync request. | |
1496 | */ | |
1497 | /* | |
1498 | * We check if all blocks are in-sync and only write to blocks that | |
1499 | * aren't in sync | |
1500 | */ | |
1501 | static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio) | |
1502 | { | |
070ec55d | 1503 | conf_t *conf = mddev->private; |
1da177e4 LT |
1504 | int i, first; |
1505 | struct bio *tbio, *fbio; | |
1506 | ||
1507 | atomic_set(&r10_bio->remaining, 1); | |
1508 | ||
1509 | /* find the first device with a block */ | |
1510 | for (i=0; i<conf->copies; i++) | |
1511 | if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) | |
1512 | break; | |
1513 | ||
1514 | if (i == conf->copies) | |
1515 | goto done; | |
1516 | ||
1517 | first = i; | |
1518 | fbio = r10_bio->devs[i].bio; | |
1519 | ||
1520 | /* now find blocks with errors */ | |
0eb3ff12 N |
1521 | for (i=0 ; i < conf->copies ; i++) { |
1522 | int j, d; | |
1523 | int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9); | |
1da177e4 | 1524 | |
1da177e4 | 1525 | tbio = r10_bio->devs[i].bio; |
0eb3ff12 N |
1526 | |
1527 | if (tbio->bi_end_io != end_sync_read) | |
1528 | continue; | |
1529 | if (i == first) | |
1da177e4 | 1530 | continue; |
0eb3ff12 N |
1531 | if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) { |
1532 | /* We know that the bi_io_vec layout is the same for | |
1533 | * both 'first' and 'i', so we just compare them. | |
1534 | * All vec entries are PAGE_SIZE; | |
1535 | */ | |
1536 | for (j = 0; j < vcnt; j++) | |
1537 | if (memcmp(page_address(fbio->bi_io_vec[j].bv_page), | |
1538 | page_address(tbio->bi_io_vec[j].bv_page), | |
1539 | PAGE_SIZE)) | |
1540 | break; | |
1541 | if (j == vcnt) | |
1542 | continue; | |
1543 | mddev->resync_mismatches += r10_bio->sectors; | |
1544 | } | |
18f08819 N |
1545 | if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) |
1546 | /* Don't fix anything. */ | |
1547 | continue; | |
1da177e4 LT |
1548 | /* Ok, we need to write this bio |
1549 | * First we need to fixup bv_offset, bv_len and | |
1550 | * bi_vecs, as the read request might have corrupted these | |
1551 | */ | |
1552 | tbio->bi_vcnt = vcnt; | |
1553 | tbio->bi_size = r10_bio->sectors << 9; | |
1554 | tbio->bi_idx = 0; | |
1555 | tbio->bi_phys_segments = 0; | |
1da177e4 LT |
1556 | tbio->bi_flags &= ~(BIO_POOL_MASK - 1); |
1557 | tbio->bi_flags |= 1 << BIO_UPTODATE; | |
1558 | tbio->bi_next = NULL; | |
1559 | tbio->bi_rw = WRITE; | |
1560 | tbio->bi_private = r10_bio; | |
1561 | tbio->bi_sector = r10_bio->devs[i].addr; | |
1562 | ||
1563 | for (j=0; j < vcnt ; j++) { | |
1564 | tbio->bi_io_vec[j].bv_offset = 0; | |
1565 | tbio->bi_io_vec[j].bv_len = PAGE_SIZE; | |
1566 | ||
1567 | memcpy(page_address(tbio->bi_io_vec[j].bv_page), | |
1568 | page_address(fbio->bi_io_vec[j].bv_page), | |
1569 | PAGE_SIZE); | |
1570 | } | |
1571 | tbio->bi_end_io = end_sync_write; | |
1572 | ||
1573 | d = r10_bio->devs[i].devnum; | |
1574 | atomic_inc(&conf->mirrors[d].rdev->nr_pending); | |
1575 | atomic_inc(&r10_bio->remaining); | |
1576 | md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9); | |
1577 | ||
1578 | tbio->bi_sector += conf->mirrors[d].rdev->data_offset; | |
1579 | tbio->bi_bdev = conf->mirrors[d].rdev->bdev; | |
1580 | generic_make_request(tbio); | |
1581 | } | |
1582 | ||
1583 | done: | |
1584 | if (atomic_dec_and_test(&r10_bio->remaining)) { | |
1585 | md_done_sync(mddev, r10_bio->sectors, 1); | |
1586 | put_buf(r10_bio); | |
1587 | } | |
1588 | } | |
1589 | ||
1590 | /* | |
1591 | * Now for the recovery code. | |
1592 | * Recovery happens across physical sectors. | |
1593 | * We recover all non-is_sync drives by finding the virtual address of | |
1594 | * each, and then choose a working drive that also has that virt address. | |
1595 | * There is a separate r10_bio for each non-in_sync drive. | |
1596 | * Only the first two slots are in use. The first for reading, | |
1597 | * The second for writing. | |
1598 | * | |
1599 | */ | |
1600 | ||
1601 | static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio) | |
1602 | { | |
070ec55d | 1603 | conf_t *conf = mddev->private; |
c65060ad NK |
1604 | int d; |
1605 | struct bio *wbio; | |
1da177e4 | 1606 | |
c65060ad NK |
1607 | /* |
1608 | * share the pages with the first bio | |
1da177e4 LT |
1609 | * and submit the write request |
1610 | */ | |
1da177e4 | 1611 | wbio = r10_bio->devs[1].bio; |
1da177e4 LT |
1612 | d = r10_bio->devs[1].devnum; |
1613 | ||
1614 | atomic_inc(&conf->mirrors[d].rdev->nr_pending); | |
1615 | md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9); | |
0eb3ff12 N |
1616 | if (test_bit(R10BIO_Uptodate, &r10_bio->state)) |
1617 | generic_make_request(wbio); | |
2bb77736 N |
1618 | else { |
1619 | printk(KERN_NOTICE | |
1620 | "md/raid10:%s: recovery aborted due to read error\n", | |
1621 | mdname(mddev)); | |
1622 | conf->mirrors[d].recovery_disabled = mddev->recovery_disabled; | |
1623 | set_bit(MD_RECOVERY_INTR, &mddev->recovery); | |
1624 | bio_endio(wbio, 0); | |
1625 | } | |
1da177e4 LT |
1626 | } |
1627 | ||
1628 | ||
1e50915f RB |
1629 | /* |
1630 | * Used by fix_read_error() to decay the per rdev read_errors. | |
1631 | * We halve the read error count for every hour that has elapsed | |
1632 | * since the last recorded read error. | |
1633 | * | |
1634 | */ | |
1635 | static void check_decay_read_errors(mddev_t *mddev, mdk_rdev_t *rdev) | |
1636 | { | |
1637 | struct timespec cur_time_mon; | |
1638 | unsigned long hours_since_last; | |
1639 | unsigned int read_errors = atomic_read(&rdev->read_errors); | |
1640 | ||
1641 | ktime_get_ts(&cur_time_mon); | |
1642 | ||
1643 | if (rdev->last_read_error.tv_sec == 0 && | |
1644 | rdev->last_read_error.tv_nsec == 0) { | |
1645 | /* first time we've seen a read error */ | |
1646 | rdev->last_read_error = cur_time_mon; | |
1647 | return; | |
1648 | } | |
1649 | ||
1650 | hours_since_last = (cur_time_mon.tv_sec - | |
1651 | rdev->last_read_error.tv_sec) / 3600; | |
1652 | ||
1653 | rdev->last_read_error = cur_time_mon; | |
1654 | ||
1655 | /* | |
1656 | * if hours_since_last is > the number of bits in read_errors | |
1657 | * just set read errors to 0. We do this to avoid | |
1658 | * overflowing the shift of read_errors by hours_since_last. | |
1659 | */ | |
1660 | if (hours_since_last >= 8 * sizeof(read_errors)) | |
1661 | atomic_set(&rdev->read_errors, 0); | |
1662 | else | |
1663 | atomic_set(&rdev->read_errors, read_errors >> hours_since_last); | |
1664 | } | |
1665 | ||
1da177e4 LT |
1666 | /* |
1667 | * This is a kernel thread which: | |
1668 | * | |
1669 | * 1. Retries failed read operations on working mirrors. | |
1670 | * 2. Updates the raid superblock when problems encounter. | |
6814d536 | 1671 | * 3. Performs writes following reads for array synchronising. |
1da177e4 LT |
1672 | */ |
1673 | ||
6814d536 N |
1674 | static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio) |
1675 | { | |
1676 | int sect = 0; /* Offset from r10_bio->sector */ | |
1677 | int sectors = r10_bio->sectors; | |
1678 | mdk_rdev_t*rdev; | |
1e50915f | 1679 | int max_read_errors = atomic_read(&mddev->max_corr_read_errors); |
0544a21d | 1680 | int d = r10_bio->devs[r10_bio->read_slot].devnum; |
1e50915f | 1681 | |
7c4e06ff N |
1682 | /* still own a reference to this rdev, so it cannot |
1683 | * have been cleared recently. | |
1684 | */ | |
1685 | rdev = conf->mirrors[d].rdev; | |
1e50915f | 1686 | |
7c4e06ff N |
1687 | if (test_bit(Faulty, &rdev->flags)) |
1688 | /* drive has already been failed, just ignore any | |
1689 | more fix_read_error() attempts */ | |
1690 | return; | |
1e50915f | 1691 | |
7c4e06ff N |
1692 | check_decay_read_errors(mddev, rdev); |
1693 | atomic_inc(&rdev->read_errors); | |
1694 | if (atomic_read(&rdev->read_errors) > max_read_errors) { | |
1695 | char b[BDEVNAME_SIZE]; | |
1696 | bdevname(rdev->bdev, b); | |
1e50915f | 1697 | |
7c4e06ff N |
1698 | printk(KERN_NOTICE |
1699 | "md/raid10:%s: %s: Raid device exceeded " | |
1700 | "read_error threshold [cur %d:max %d]\n", | |
1701 | mdname(mddev), b, | |
1702 | atomic_read(&rdev->read_errors), max_read_errors); | |
1703 | printk(KERN_NOTICE | |
1704 | "md/raid10:%s: %s: Failing raid device\n", | |
1705 | mdname(mddev), b); | |
1706 | md_error(mddev, conf->mirrors[d].rdev); | |
1707 | return; | |
1e50915f | 1708 | } |
1e50915f | 1709 | |
6814d536 N |
1710 | while(sectors) { |
1711 | int s = sectors; | |
1712 | int sl = r10_bio->read_slot; | |
1713 | int success = 0; | |
1714 | int start; | |
1715 | ||
1716 | if (s > (PAGE_SIZE>>9)) | |
1717 | s = PAGE_SIZE >> 9; | |
1718 | ||
1719 | rcu_read_lock(); | |
1720 | do { | |
8dbed5ce N |
1721 | sector_t first_bad; |
1722 | int bad_sectors; | |
1723 | ||
0544a21d | 1724 | d = r10_bio->devs[sl].devnum; |
6814d536 N |
1725 | rdev = rcu_dereference(conf->mirrors[d].rdev); |
1726 | if (rdev && | |
8dbed5ce N |
1727 | test_bit(In_sync, &rdev->flags) && |
1728 | is_badblock(rdev, r10_bio->devs[sl].addr + sect, s, | |
1729 | &first_bad, &bad_sectors) == 0) { | |
6814d536 N |
1730 | atomic_inc(&rdev->nr_pending); |
1731 | rcu_read_unlock(); | |
2b193363 | 1732 | success = sync_page_io(rdev, |
6814d536 | 1733 | r10_bio->devs[sl].addr + |
ccebd4c4 | 1734 | sect, |
6814d536 | 1735 | s<<9, |
ccebd4c4 | 1736 | conf->tmppage, READ, false); |
6814d536 N |
1737 | rdev_dec_pending(rdev, mddev); |
1738 | rcu_read_lock(); | |
1739 | if (success) | |
1740 | break; | |
1741 | } | |
1742 | sl++; | |
1743 | if (sl == conf->copies) | |
1744 | sl = 0; | |
1745 | } while (!success && sl != r10_bio->read_slot); | |
1746 | rcu_read_unlock(); | |
1747 | ||
1748 | if (!success) { | |
1749 | /* Cannot read from anywhere -- bye bye array */ | |
1750 | int dn = r10_bio->devs[r10_bio->read_slot].devnum; | |
1751 | md_error(mddev, conf->mirrors[dn].rdev); | |
1752 | break; | |
1753 | } | |
1754 | ||
1755 | start = sl; | |
1756 | /* write it back and re-read */ | |
1757 | rcu_read_lock(); | |
1758 | while (sl != r10_bio->read_slot) { | |
67b8dc4b | 1759 | char b[BDEVNAME_SIZE]; |
0544a21d | 1760 | |
6814d536 N |
1761 | if (sl==0) |
1762 | sl = conf->copies; | |
1763 | sl--; | |
1764 | d = r10_bio->devs[sl].devnum; | |
1765 | rdev = rcu_dereference(conf->mirrors[d].rdev); | |
1294b9c9 N |
1766 | if (!rdev || |
1767 | !test_bit(In_sync, &rdev->flags)) | |
1768 | continue; | |
1769 | ||
1770 | atomic_inc(&rdev->nr_pending); | |
1771 | rcu_read_unlock(); | |
1772 | if (sync_page_io(rdev, | |
1773 | r10_bio->devs[sl].addr + | |
1774 | sect, | |
1775 | s<<9, conf->tmppage, WRITE, false) | |
1776 | == 0) { | |
1777 | /* Well, this device is dead */ | |
1778 | printk(KERN_NOTICE | |
1779 | "md/raid10:%s: read correction " | |
1780 | "write failed" | |
1781 | " (%d sectors at %llu on %s)\n", | |
1782 | mdname(mddev), s, | |
1783 | (unsigned long long)( | |
1784 | sect + rdev->data_offset), | |
1785 | bdevname(rdev->bdev, b)); | |
1786 | printk(KERN_NOTICE "md/raid10:%s: %s: failing " | |
1787 | "drive\n", | |
1788 | mdname(mddev), | |
1789 | bdevname(rdev->bdev, b)); | |
1790 | md_error(mddev, rdev); | |
6814d536 | 1791 | } |
1294b9c9 N |
1792 | rdev_dec_pending(rdev, mddev); |
1793 | rcu_read_lock(); | |
6814d536 N |
1794 | } |
1795 | sl = start; | |
1796 | while (sl != r10_bio->read_slot) { | |
1294b9c9 | 1797 | char b[BDEVNAME_SIZE]; |
0544a21d | 1798 | |
6814d536 N |
1799 | if (sl==0) |
1800 | sl = conf->copies; | |
1801 | sl--; | |
1802 | d = r10_bio->devs[sl].devnum; | |
1803 | rdev = rcu_dereference(conf->mirrors[d].rdev); | |
1294b9c9 N |
1804 | if (!rdev || |
1805 | !test_bit(In_sync, &rdev->flags)) | |
1806 | continue; | |
6814d536 | 1807 | |
1294b9c9 N |
1808 | atomic_inc(&rdev->nr_pending); |
1809 | rcu_read_unlock(); | |
1810 | if (sync_page_io(rdev, | |
1811 | r10_bio->devs[sl].addr + | |
1812 | sect, | |
1813 | s<<9, conf->tmppage, | |
1814 | READ, false) == 0) { | |
1815 | /* Well, this device is dead */ | |
1816 | printk(KERN_NOTICE | |
1817 | "md/raid10:%s: unable to read back " | |
1818 | "corrected sectors" | |
1819 | " (%d sectors at %llu on %s)\n", | |
1820 | mdname(mddev), s, | |
1821 | (unsigned long long)( | |
1822 | sect + rdev->data_offset), | |
1823 | bdevname(rdev->bdev, b)); | |
1824 | printk(KERN_NOTICE "md/raid10:%s: %s: failing " | |
1825 | "drive\n", | |
1826 | mdname(mddev), | |
1827 | bdevname(rdev->bdev, b)); | |
1828 | ||
1829 | md_error(mddev, rdev); | |
1830 | } else { | |
1831 | printk(KERN_INFO | |
1832 | "md/raid10:%s: read error corrected" | |
1833 | " (%d sectors at %llu on %s)\n", | |
1834 | mdname(mddev), s, | |
1835 | (unsigned long long)( | |
1836 | sect + rdev->data_offset), | |
1837 | bdevname(rdev->bdev, b)); | |
1838 | atomic_add(s, &rdev->corrected_errors); | |
6814d536 | 1839 | } |
1294b9c9 N |
1840 | |
1841 | rdev_dec_pending(rdev, mddev); | |
1842 | rcu_read_lock(); | |
6814d536 N |
1843 | } |
1844 | rcu_read_unlock(); | |
1845 | ||
1846 | sectors -= s; | |
1847 | sect += s; | |
1848 | } | |
1849 | } | |
1850 | ||
bd870a16 N |
1851 | static void bi_complete(struct bio *bio, int error) |
1852 | { | |
1853 | complete((struct completion *)bio->bi_private); | |
1854 | } | |
1855 | ||
1856 | static int submit_bio_wait(int rw, struct bio *bio) | |
1857 | { | |
1858 | struct completion event; | |
1859 | rw |= REQ_SYNC; | |
1860 | ||
1861 | init_completion(&event); | |
1862 | bio->bi_private = &event; | |
1863 | bio->bi_end_io = bi_complete; | |
1864 | submit_bio(rw, bio); | |
1865 | wait_for_completion(&event); | |
1866 | ||
1867 | return test_bit(BIO_UPTODATE, &bio->bi_flags); | |
1868 | } | |
1869 | ||
1870 | static int narrow_write_error(r10bio_t *r10_bio, int i) | |
1871 | { | |
1872 | struct bio *bio = r10_bio->master_bio; | |
1873 | mddev_t *mddev = r10_bio->mddev; | |
1874 | conf_t *conf = mddev->private; | |
1875 | mdk_rdev_t *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev; | |
1876 | /* bio has the data to be written to slot 'i' where | |
1877 | * we just recently had a write error. | |
1878 | * We repeatedly clone the bio and trim down to one block, | |
1879 | * then try the write. Where the write fails we record | |
1880 | * a bad block. | |
1881 | * It is conceivable that the bio doesn't exactly align with | |
1882 | * blocks. We must handle this. | |
1883 | * | |
1884 | * We currently own a reference to the rdev. | |
1885 | */ | |
1886 | ||
1887 | int block_sectors; | |
1888 | sector_t sector; | |
1889 | int sectors; | |
1890 | int sect_to_write = r10_bio->sectors; | |
1891 | int ok = 1; | |
1892 | ||
1893 | if (rdev->badblocks.shift < 0) | |
1894 | return 0; | |
1895 | ||
1896 | block_sectors = 1 << rdev->badblocks.shift; | |
1897 | sector = r10_bio->sector; | |
1898 | sectors = ((r10_bio->sector + block_sectors) | |
1899 | & ~(sector_t)(block_sectors - 1)) | |
1900 | - sector; | |
1901 | ||
1902 | while (sect_to_write) { | |
1903 | struct bio *wbio; | |
1904 | if (sectors > sect_to_write) | |
1905 | sectors = sect_to_write; | |
1906 | /* Write at 'sector' for 'sectors' */ | |
1907 | wbio = bio_clone_mddev(bio, GFP_NOIO, mddev); | |
1908 | md_trim_bio(wbio, sector - bio->bi_sector, sectors); | |
1909 | wbio->bi_sector = (r10_bio->devs[i].addr+ | |
1910 | rdev->data_offset+ | |
1911 | (sector - r10_bio->sector)); | |
1912 | wbio->bi_bdev = rdev->bdev; | |
1913 | if (submit_bio_wait(WRITE, wbio) == 0) | |
1914 | /* Failure! */ | |
1915 | ok = rdev_set_badblocks(rdev, sector, | |
1916 | sectors, 0) | |
1917 | && ok; | |
1918 | ||
1919 | bio_put(wbio); | |
1920 | sect_to_write -= sectors; | |
1921 | sector += sectors; | |
1922 | sectors = block_sectors; | |
1923 | } | |
1924 | return ok; | |
1925 | } | |
1926 | ||
560f8e55 N |
1927 | static void handle_read_error(mddev_t *mddev, r10bio_t *r10_bio) |
1928 | { | |
1929 | int slot = r10_bio->read_slot; | |
1930 | int mirror = r10_bio->devs[slot].devnum; | |
1931 | struct bio *bio; | |
1932 | conf_t *conf = mddev->private; | |
1933 | mdk_rdev_t *rdev; | |
1934 | char b[BDEVNAME_SIZE]; | |
1935 | unsigned long do_sync; | |
856e08e2 | 1936 | int max_sectors; |
560f8e55 N |
1937 | |
1938 | /* we got a read error. Maybe the drive is bad. Maybe just | |
1939 | * the block and we can fix it. | |
1940 | * We freeze all other IO, and try reading the block from | |
1941 | * other devices. When we find one, we re-write | |
1942 | * and check it that fixes the read error. | |
1943 | * This is all done synchronously while the array is | |
1944 | * frozen. | |
1945 | */ | |
1946 | if (mddev->ro == 0) { | |
1947 | freeze_array(conf); | |
1948 | fix_read_error(conf, mddev, r10_bio); | |
1949 | unfreeze_array(conf); | |
1950 | } | |
1951 | rdev_dec_pending(conf->mirrors[mirror].rdev, mddev); | |
1952 | ||
1953 | bio = r10_bio->devs[slot].bio; | |
7399c31b | 1954 | bdevname(bio->bi_bdev, b); |
560f8e55 N |
1955 | r10_bio->devs[slot].bio = |
1956 | mddev->ro ? IO_BLOCKED : NULL; | |
7399c31b | 1957 | read_more: |
856e08e2 | 1958 | mirror = read_balance(conf, r10_bio, &max_sectors); |
7399c31b | 1959 | if (mirror == -1) { |
560f8e55 N |
1960 | printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O" |
1961 | " read error for block %llu\n", | |
7399c31b | 1962 | mdname(mddev), b, |
560f8e55 N |
1963 | (unsigned long long)r10_bio->sector); |
1964 | raid_end_bio_io(r10_bio); | |
1965 | bio_put(bio); | |
1966 | return; | |
1967 | } | |
1968 | ||
1969 | do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC); | |
7399c31b N |
1970 | if (bio) |
1971 | bio_put(bio); | |
560f8e55 N |
1972 | slot = r10_bio->read_slot; |
1973 | rdev = conf->mirrors[mirror].rdev; | |
1974 | printk_ratelimited( | |
1975 | KERN_ERR | |
1976 | "md/raid10:%s: %s: redirecting" | |
1977 | "sector %llu to another mirror\n", | |
1978 | mdname(mddev), | |
1979 | bdevname(rdev->bdev, b), | |
1980 | (unsigned long long)r10_bio->sector); | |
1981 | bio = bio_clone_mddev(r10_bio->master_bio, | |
1982 | GFP_NOIO, mddev); | |
7399c31b N |
1983 | md_trim_bio(bio, |
1984 | r10_bio->sector - bio->bi_sector, | |
1985 | max_sectors); | |
560f8e55 N |
1986 | r10_bio->devs[slot].bio = bio; |
1987 | bio->bi_sector = r10_bio->devs[slot].addr | |
1988 | + rdev->data_offset; | |
1989 | bio->bi_bdev = rdev->bdev; | |
1990 | bio->bi_rw = READ | do_sync; | |
1991 | bio->bi_private = r10_bio; | |
1992 | bio->bi_end_io = raid10_end_read_request; | |
7399c31b N |
1993 | if (max_sectors < r10_bio->sectors) { |
1994 | /* Drat - have to split this up more */ | |
1995 | struct bio *mbio = r10_bio->master_bio; | |
1996 | int sectors_handled = | |
1997 | r10_bio->sector + max_sectors | |
1998 | - mbio->bi_sector; | |
1999 | r10_bio->sectors = max_sectors; | |
2000 | spin_lock_irq(&conf->device_lock); | |
2001 | if (mbio->bi_phys_segments == 0) | |
2002 | mbio->bi_phys_segments = 2; | |
2003 | else | |
2004 | mbio->bi_phys_segments++; | |
2005 | spin_unlock_irq(&conf->device_lock); | |
2006 | generic_make_request(bio); | |
2007 | bio = NULL; | |
2008 | ||
2009 | r10_bio = mempool_alloc(conf->r10bio_pool, | |
2010 | GFP_NOIO); | |
2011 | r10_bio->master_bio = mbio; | |
2012 | r10_bio->sectors = (mbio->bi_size >> 9) | |
2013 | - sectors_handled; | |
2014 | r10_bio->state = 0; | |
2015 | set_bit(R10BIO_ReadError, | |
2016 | &r10_bio->state); | |
2017 | r10_bio->mddev = mddev; | |
2018 | r10_bio->sector = mbio->bi_sector | |
2019 | + sectors_handled; | |
2020 | ||
2021 | goto read_more; | |
2022 | } else | |
2023 | generic_make_request(bio); | |
560f8e55 N |
2024 | } |
2025 | ||
749c55e9 N |
2026 | static void handle_write_completed(conf_t *conf, r10bio_t *r10_bio) |
2027 | { | |
2028 | /* Some sort of write request has finished and it | |
2029 | * succeeded in writing where we thought there was a | |
2030 | * bad block. So forget the bad block. | |
2031 | */ | |
2032 | int m; | |
2033 | mdk_rdev_t *rdev; | |
2034 | ||
2035 | if (test_bit(R10BIO_IsSync, &r10_bio->state) || | |
2036 | test_bit(R10BIO_IsRecover, &r10_bio->state)) { | |
2037 | for (m = 0; m < conf->copies; m++) | |
2038 | if (r10_bio->devs[m].bio && | |
2039 | test_bit(BIO_UPTODATE, | |
2040 | &r10_bio->devs[m].bio->bi_flags)) { | |
2041 | int dev = r10_bio->devs[m].devnum; | |
2042 | rdev = conf->mirrors[dev].rdev; | |
2043 | rdev_clear_badblocks( | |
2044 | rdev, | |
2045 | r10_bio->devs[m].addr, | |
2046 | r10_bio->sectors); | |
2047 | } | |
2048 | put_buf(r10_bio); | |
2049 | } else { | |
bd870a16 N |
2050 | for (m = 0; m < conf->copies; m++) { |
2051 | int dev = r10_bio->devs[m].devnum; | |
2052 | struct bio *bio = r10_bio->devs[m].bio; | |
2053 | rdev = conf->mirrors[dev].rdev; | |
2054 | if (bio == IO_MADE_GOOD) { | |
749c55e9 N |
2055 | rdev_clear_badblocks( |
2056 | rdev, | |
2057 | r10_bio->devs[m].addr, | |
2058 | r10_bio->sectors); | |
2059 | rdev_dec_pending(rdev, conf->mddev); | |
bd870a16 N |
2060 | } else if (bio != NULL && |
2061 | !test_bit(BIO_UPTODATE, &bio->bi_flags)) { | |
2062 | if (!narrow_write_error(r10_bio, m)) { | |
2063 | md_error(conf->mddev, rdev); | |
2064 | set_bit(R10BIO_Degraded, | |
2065 | &r10_bio->state); | |
2066 | } | |
2067 | rdev_dec_pending(rdev, conf->mddev); | |
749c55e9 | 2068 | } |
bd870a16 N |
2069 | } |
2070 | if (test_bit(R10BIO_WriteError, | |
2071 | &r10_bio->state)) | |
2072 | close_write(r10_bio); | |
749c55e9 N |
2073 | raid_end_bio_io(r10_bio); |
2074 | } | |
2075 | } | |
2076 | ||
1da177e4 LT |
2077 | static void raid10d(mddev_t *mddev) |
2078 | { | |
2079 | r10bio_t *r10_bio; | |
1da177e4 | 2080 | unsigned long flags; |
070ec55d | 2081 | conf_t *conf = mddev->private; |
1da177e4 | 2082 | struct list_head *head = &conf->retry_list; |
e1dfa0a2 | 2083 | struct blk_plug plug; |
1da177e4 LT |
2084 | |
2085 | md_check_recovery(mddev); | |
1da177e4 | 2086 | |
e1dfa0a2 | 2087 | blk_start_plug(&plug); |
1da177e4 | 2088 | for (;;) { |
6cce3b23 | 2089 | |
7eaceacc | 2090 | flush_pending_writes(conf); |
6cce3b23 | 2091 | |
a35e63ef N |
2092 | spin_lock_irqsave(&conf->device_lock, flags); |
2093 | if (list_empty(head)) { | |
2094 | spin_unlock_irqrestore(&conf->device_lock, flags); | |
1da177e4 | 2095 | break; |
a35e63ef | 2096 | } |
1da177e4 LT |
2097 | r10_bio = list_entry(head->prev, r10bio_t, retry_list); |
2098 | list_del(head->prev); | |
4443ae10 | 2099 | conf->nr_queued--; |
1da177e4 LT |
2100 | spin_unlock_irqrestore(&conf->device_lock, flags); |
2101 | ||
2102 | mddev = r10_bio->mddev; | |
070ec55d | 2103 | conf = mddev->private; |
bd870a16 N |
2104 | if (test_bit(R10BIO_MadeGood, &r10_bio->state) || |
2105 | test_bit(R10BIO_WriteError, &r10_bio->state)) | |
749c55e9 N |
2106 | handle_write_completed(conf, r10_bio); |
2107 | else if (test_bit(R10BIO_IsSync, &r10_bio->state)) | |
1da177e4 | 2108 | sync_request_write(mddev, r10_bio); |
7eaceacc | 2109 | else if (test_bit(R10BIO_IsRecover, &r10_bio->state)) |
1da177e4 | 2110 | recovery_request_write(mddev, r10_bio); |
856e08e2 | 2111 | else if (test_bit(R10BIO_ReadError, &r10_bio->state)) |
560f8e55 | 2112 | handle_read_error(mddev, r10_bio); |
856e08e2 N |
2113 | else { |
2114 | /* just a partial read to be scheduled from a | |
2115 | * separate context | |
2116 | */ | |
2117 | int slot = r10_bio->read_slot; | |
2118 | generic_make_request(r10_bio->devs[slot].bio); | |
2119 | } | |
560f8e55 | 2120 | |
1d9d5241 | 2121 | cond_resched(); |
de393cde N |
2122 | if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) |
2123 | md_check_recovery(mddev); | |
1da177e4 | 2124 | } |
e1dfa0a2 | 2125 | blk_finish_plug(&plug); |
1da177e4 LT |
2126 | } |
2127 | ||
2128 | ||
2129 | static int init_resync(conf_t *conf) | |
2130 | { | |
2131 | int buffs; | |
2132 | ||
2133 | buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE; | |
b6385483 | 2134 | BUG_ON(conf->r10buf_pool); |
1da177e4 LT |
2135 | conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf); |
2136 | if (!conf->r10buf_pool) | |
2137 | return -ENOMEM; | |
2138 | conf->next_resync = 0; | |
2139 | return 0; | |
2140 | } | |
2141 | ||
2142 | /* | |
2143 | * perform a "sync" on one "block" | |
2144 | * | |
2145 | * We need to make sure that no normal I/O request - particularly write | |
2146 | * requests - conflict with active sync requests. | |
2147 | * | |
2148 | * This is achieved by tracking pending requests and a 'barrier' concept | |
2149 | * that can be installed to exclude normal IO requests. | |
2150 | * | |
2151 | * Resync and recovery are handled very differently. | |
2152 | * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery. | |
2153 | * | |
2154 | * For resync, we iterate over virtual addresses, read all copies, | |
2155 | * and update if there are differences. If only one copy is live, | |
2156 | * skip it. | |
2157 | * For recovery, we iterate over physical addresses, read a good | |
2158 | * value for each non-in_sync drive, and over-write. | |
2159 | * | |
2160 | * So, for recovery we may have several outstanding complex requests for a | |
2161 | * given address, one for each out-of-sync device. We model this by allocating | |
2162 | * a number of r10_bio structures, one for each out-of-sync device. | |
2163 | * As we setup these structures, we collect all bio's together into a list | |
2164 | * which we then process collectively to add pages, and then process again | |
2165 | * to pass to generic_make_request. | |
2166 | * | |
2167 | * The r10_bio structures are linked using a borrowed master_bio pointer. | |
2168 | * This link is counted in ->remaining. When the r10_bio that points to NULL | |
2169 | * has its remaining count decremented to 0, the whole complex operation | |
2170 | * is complete. | |
2171 | * | |
2172 | */ | |
2173 | ||
ab9d47e9 N |
2174 | static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, |
2175 | int *skipped, int go_faster) | |
1da177e4 | 2176 | { |
070ec55d | 2177 | conf_t *conf = mddev->private; |
1da177e4 LT |
2178 | r10bio_t *r10_bio; |
2179 | struct bio *biolist = NULL, *bio; | |
2180 | sector_t max_sector, nr_sectors; | |
1da177e4 | 2181 | int i; |
6cce3b23 | 2182 | int max_sync; |
57dab0bd | 2183 | sector_t sync_blocks; |
1da177e4 LT |
2184 | sector_t sectors_skipped = 0; |
2185 | int chunks_skipped = 0; | |
2186 | ||
2187 | if (!conf->r10buf_pool) | |
2188 | if (init_resync(conf)) | |
57afd89f | 2189 | return 0; |
1da177e4 LT |
2190 | |
2191 | skipped: | |
58c0fed4 | 2192 | max_sector = mddev->dev_sectors; |
1da177e4 LT |
2193 | if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) |
2194 | max_sector = mddev->resync_max_sectors; | |
2195 | if (sector_nr >= max_sector) { | |
6cce3b23 N |
2196 | /* If we aborted, we need to abort the |
2197 | * sync on the 'current' bitmap chucks (there can | |
2198 | * be several when recovering multiple devices). | |
2199 | * as we may have started syncing it but not finished. | |
2200 | * We can find the current address in | |
2201 | * mddev->curr_resync, but for recovery, | |
2202 | * we need to convert that to several | |
2203 | * virtual addresses. | |
2204 | */ | |
2205 | if (mddev->curr_resync < max_sector) { /* aborted */ | |
2206 | if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) | |
2207 | bitmap_end_sync(mddev->bitmap, mddev->curr_resync, | |
2208 | &sync_blocks, 1); | |
2209 | else for (i=0; i<conf->raid_disks; i++) { | |
2210 | sector_t sect = | |
2211 | raid10_find_virt(conf, mddev->curr_resync, i); | |
2212 | bitmap_end_sync(mddev->bitmap, sect, | |
2213 | &sync_blocks, 1); | |
2214 | } | |
2215 | } else /* completed sync */ | |
2216 | conf->fullsync = 0; | |
2217 | ||
2218 | bitmap_close_sync(mddev->bitmap); | |
1da177e4 | 2219 | close_sync(conf); |
57afd89f | 2220 | *skipped = 1; |
1da177e4 LT |
2221 | return sectors_skipped; |
2222 | } | |
2223 | if (chunks_skipped >= conf->raid_disks) { | |
2224 | /* if there has been nothing to do on any drive, | |
2225 | * then there is nothing to do at all.. | |
2226 | */ | |
57afd89f N |
2227 | *skipped = 1; |
2228 | return (max_sector - sector_nr) + sectors_skipped; | |
1da177e4 LT |
2229 | } |
2230 | ||
c6207277 N |
2231 | if (max_sector > mddev->resync_max) |
2232 | max_sector = mddev->resync_max; /* Don't do IO beyond here */ | |
2233 | ||
1da177e4 LT |
2234 | /* make sure whole request will fit in a chunk - if chunks |
2235 | * are meaningful | |
2236 | */ | |
2237 | if (conf->near_copies < conf->raid_disks && | |
2238 | max_sector > (sector_nr | conf->chunk_mask)) | |
2239 | max_sector = (sector_nr | conf->chunk_mask) + 1; | |
2240 | /* | |
2241 | * If there is non-resync activity waiting for us then | |
2242 | * put in a delay to throttle resync. | |
2243 | */ | |
0a27ec96 | 2244 | if (!go_faster && conf->nr_waiting) |
1da177e4 | 2245 | msleep_interruptible(1000); |
1da177e4 LT |
2246 | |
2247 | /* Again, very different code for resync and recovery. | |
2248 | * Both must result in an r10bio with a list of bios that | |
2249 | * have bi_end_io, bi_sector, bi_bdev set, | |
2250 | * and bi_private set to the r10bio. | |
2251 | * For recovery, we may actually create several r10bios | |
2252 | * with 2 bios in each, that correspond to the bios in the main one. | |
2253 | * In this case, the subordinate r10bios link back through a | |
2254 | * borrowed master_bio pointer, and the counter in the master | |
2255 | * includes a ref from each subordinate. | |
2256 | */ | |
2257 | /* First, we decide what to do and set ->bi_end_io | |
2258 | * To end_sync_read if we want to read, and | |
2259 | * end_sync_write if we will want to write. | |
2260 | */ | |
2261 | ||
6cce3b23 | 2262 | max_sync = RESYNC_PAGES << (PAGE_SHIFT-9); |
1da177e4 LT |
2263 | if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) { |
2264 | /* recovery... the complicated one */ | |
e875ecea | 2265 | int j; |
1da177e4 LT |
2266 | r10_bio = NULL; |
2267 | ||
ab9d47e9 N |
2268 | for (i=0 ; i<conf->raid_disks; i++) { |
2269 | int still_degraded; | |
2270 | r10bio_t *rb2; | |
2271 | sector_t sect; | |
2272 | int must_sync; | |
e875ecea | 2273 | int any_working; |
1da177e4 | 2274 | |
ab9d47e9 N |
2275 | if (conf->mirrors[i].rdev == NULL || |
2276 | test_bit(In_sync, &conf->mirrors[i].rdev->flags)) | |
2277 | continue; | |
1da177e4 | 2278 | |
ab9d47e9 N |
2279 | still_degraded = 0; |
2280 | /* want to reconstruct this device */ | |
2281 | rb2 = r10_bio; | |
2282 | sect = raid10_find_virt(conf, sector_nr, i); | |
2283 | /* Unless we are doing a full sync, we only need | |
2284 | * to recover the block if it is set in the bitmap | |
2285 | */ | |
2286 | must_sync = bitmap_start_sync(mddev->bitmap, sect, | |
2287 | &sync_blocks, 1); | |
2288 | if (sync_blocks < max_sync) | |
2289 | max_sync = sync_blocks; | |
2290 | if (!must_sync && | |
2291 | !conf->fullsync) { | |
2292 | /* yep, skip the sync_blocks here, but don't assume | |
2293 | * that there will never be anything to do here | |
2294 | */ | |
2295 | chunks_skipped = -1; | |
2296 | continue; | |
2297 | } | |
6cce3b23 | 2298 | |
ab9d47e9 N |
2299 | r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO); |
2300 | raise_barrier(conf, rb2 != NULL); | |
2301 | atomic_set(&r10_bio->remaining, 0); | |
18055569 | 2302 | |
ab9d47e9 N |
2303 | r10_bio->master_bio = (struct bio*)rb2; |
2304 | if (rb2) | |
2305 | atomic_inc(&rb2->remaining); | |
2306 | r10_bio->mddev = mddev; | |
2307 | set_bit(R10BIO_IsRecover, &r10_bio->state); | |
2308 | r10_bio->sector = sect; | |
1da177e4 | 2309 | |
ab9d47e9 N |
2310 | raid10_find_phys(conf, r10_bio); |
2311 | ||
2312 | /* Need to check if the array will still be | |
2313 | * degraded | |
2314 | */ | |
2315 | for (j=0; j<conf->raid_disks; j++) | |
2316 | if (conf->mirrors[j].rdev == NULL || | |
2317 | test_bit(Faulty, &conf->mirrors[j].rdev->flags)) { | |
2318 | still_degraded = 1; | |
87fc767b | 2319 | break; |
1da177e4 | 2320 | } |
ab9d47e9 N |
2321 | |
2322 | must_sync = bitmap_start_sync(mddev->bitmap, sect, | |
2323 | &sync_blocks, still_degraded); | |
2324 | ||
e875ecea | 2325 | any_working = 0; |
ab9d47e9 | 2326 | for (j=0; j<conf->copies;j++) { |
e875ecea | 2327 | int k; |
ab9d47e9 | 2328 | int d = r10_bio->devs[j].devnum; |
40c356ce N |
2329 | mdk_rdev_t *rdev; |
2330 | sector_t sector, first_bad; | |
2331 | int bad_sectors; | |
ab9d47e9 N |
2332 | if (!conf->mirrors[d].rdev || |
2333 | !test_bit(In_sync, &conf->mirrors[d].rdev->flags)) | |
2334 | continue; | |
2335 | /* This is where we read from */ | |
e875ecea | 2336 | any_working = 1; |
40c356ce N |
2337 | rdev = conf->mirrors[d].rdev; |
2338 | sector = r10_bio->devs[j].addr; | |
2339 | ||
2340 | if (is_badblock(rdev, sector, max_sync, | |
2341 | &first_bad, &bad_sectors)) { | |
2342 | if (first_bad > sector) | |
2343 | max_sync = first_bad - sector; | |
2344 | else { | |
2345 | bad_sectors -= (sector | |
2346 | - first_bad); | |
2347 | if (max_sync > bad_sectors) | |
2348 | max_sync = bad_sectors; | |
2349 | continue; | |
2350 | } | |
2351 | } | |
ab9d47e9 N |
2352 | bio = r10_bio->devs[0].bio; |
2353 | bio->bi_next = biolist; | |
2354 | biolist = bio; | |
2355 | bio->bi_private = r10_bio; | |
2356 | bio->bi_end_io = end_sync_read; | |
2357 | bio->bi_rw = READ; | |
2358 | bio->bi_sector = r10_bio->devs[j].addr + | |
2359 | conf->mirrors[d].rdev->data_offset; | |
2360 | bio->bi_bdev = conf->mirrors[d].rdev->bdev; | |
2361 | atomic_inc(&conf->mirrors[d].rdev->nr_pending); | |
2362 | atomic_inc(&r10_bio->remaining); | |
2363 | /* and we write to 'i' */ | |
2364 | ||
2365 | for (k=0; k<conf->copies; k++) | |
2366 | if (r10_bio->devs[k].devnum == i) | |
2367 | break; | |
2368 | BUG_ON(k == conf->copies); | |
2369 | bio = r10_bio->devs[1].bio; | |
2370 | bio->bi_next = biolist; | |
2371 | biolist = bio; | |
2372 | bio->bi_private = r10_bio; | |
2373 | bio->bi_end_io = end_sync_write; | |
2374 | bio->bi_rw = WRITE; | |
2375 | bio->bi_sector = r10_bio->devs[k].addr + | |
2376 | conf->mirrors[i].rdev->data_offset; | |
2377 | bio->bi_bdev = conf->mirrors[i].rdev->bdev; | |
2378 | ||
2379 | r10_bio->devs[0].devnum = d; | |
2380 | r10_bio->devs[1].devnum = i; | |
2381 | ||
2382 | break; | |
2383 | } | |
2384 | if (j == conf->copies) { | |
e875ecea N |
2385 | /* Cannot recover, so abort the recovery or |
2386 | * record a bad block */ | |
ab9d47e9 N |
2387 | put_buf(r10_bio); |
2388 | if (rb2) | |
2389 | atomic_dec(&rb2->remaining); | |
2390 | r10_bio = rb2; | |
e875ecea N |
2391 | if (any_working) { |
2392 | /* problem is that there are bad blocks | |
2393 | * on other device(s) | |
2394 | */ | |
2395 | int k; | |
2396 | for (k = 0; k < conf->copies; k++) | |
2397 | if (r10_bio->devs[k].devnum == i) | |
2398 | break; | |
2399 | if (!rdev_set_badblocks( | |
2400 | conf->mirrors[i].rdev, | |
2401 | r10_bio->devs[k].addr, | |
2402 | max_sync, 0)) | |
2403 | any_working = 0; | |
2404 | } | |
2405 | if (!any_working) { | |
2406 | if (!test_and_set_bit(MD_RECOVERY_INTR, | |
2407 | &mddev->recovery)) | |
2408 | printk(KERN_INFO "md/raid10:%s: insufficient " | |
2409 | "working devices for recovery.\n", | |
2410 | mdname(mddev)); | |
2411 | conf->mirrors[i].recovery_disabled | |
2412 | = mddev->recovery_disabled; | |
2413 | } | |
ab9d47e9 | 2414 | break; |
1da177e4 | 2415 | } |
ab9d47e9 | 2416 | } |
1da177e4 LT |
2417 | if (biolist == NULL) { |
2418 | while (r10_bio) { | |
2419 | r10bio_t *rb2 = r10_bio; | |
2420 | r10_bio = (r10bio_t*) rb2->master_bio; | |
2421 | rb2->master_bio = NULL; | |
2422 | put_buf(rb2); | |
2423 | } | |
2424 | goto giveup; | |
2425 | } | |
2426 | } else { | |
2427 | /* resync. Schedule a read for every block at this virt offset */ | |
2428 | int count = 0; | |
6cce3b23 | 2429 | |
78200d45 N |
2430 | bitmap_cond_end_sync(mddev->bitmap, sector_nr); |
2431 | ||
6cce3b23 N |
2432 | if (!bitmap_start_sync(mddev->bitmap, sector_nr, |
2433 | &sync_blocks, mddev->degraded) && | |
ab9d47e9 N |
2434 | !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, |
2435 | &mddev->recovery)) { | |
6cce3b23 N |
2436 | /* We can skip this block */ |
2437 | *skipped = 1; | |
2438 | return sync_blocks + sectors_skipped; | |
2439 | } | |
2440 | if (sync_blocks < max_sync) | |
2441 | max_sync = sync_blocks; | |
1da177e4 LT |
2442 | r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO); |
2443 | ||
1da177e4 LT |
2444 | r10_bio->mddev = mddev; |
2445 | atomic_set(&r10_bio->remaining, 0); | |
6cce3b23 N |
2446 | raise_barrier(conf, 0); |
2447 | conf->next_resync = sector_nr; | |
1da177e4 LT |
2448 | |
2449 | r10_bio->master_bio = NULL; | |
2450 | r10_bio->sector = sector_nr; | |
2451 | set_bit(R10BIO_IsSync, &r10_bio->state); | |
2452 | raid10_find_phys(conf, r10_bio); | |
2453 | r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1; | |
2454 | ||
2455 | for (i=0; i<conf->copies; i++) { | |
2456 | int d = r10_bio->devs[i].devnum; | |
40c356ce N |
2457 | sector_t first_bad, sector; |
2458 | int bad_sectors; | |
2459 | ||
1da177e4 LT |
2460 | bio = r10_bio->devs[i].bio; |
2461 | bio->bi_end_io = NULL; | |
af03b8e4 | 2462 | clear_bit(BIO_UPTODATE, &bio->bi_flags); |
1da177e4 | 2463 | if (conf->mirrors[d].rdev == NULL || |
b2d444d7 | 2464 | test_bit(Faulty, &conf->mirrors[d].rdev->flags)) |
1da177e4 | 2465 | continue; |
40c356ce N |
2466 | sector = r10_bio->devs[i].addr; |
2467 | if (is_badblock(conf->mirrors[d].rdev, | |
2468 | sector, max_sync, | |
2469 | &first_bad, &bad_sectors)) { | |
2470 | if (first_bad > sector) | |
2471 | max_sync = first_bad - sector; | |
2472 | else { | |
2473 | bad_sectors -= (sector - first_bad); | |
2474 | if (max_sync > bad_sectors) | |
2475 | max_sync = max_sync; | |
2476 | continue; | |
2477 | } | |
2478 | } | |
1da177e4 LT |
2479 | atomic_inc(&conf->mirrors[d].rdev->nr_pending); |
2480 | atomic_inc(&r10_bio->remaining); | |
2481 | bio->bi_next = biolist; | |
2482 | biolist = bio; | |
2483 | bio->bi_private = r10_bio; | |
2484 | bio->bi_end_io = end_sync_read; | |
802ba064 | 2485 | bio->bi_rw = READ; |
40c356ce | 2486 | bio->bi_sector = sector + |
1da177e4 LT |
2487 | conf->mirrors[d].rdev->data_offset; |
2488 | bio->bi_bdev = conf->mirrors[d].rdev->bdev; | |
2489 | count++; | |
2490 | } | |
2491 | ||
2492 | if (count < 2) { | |
2493 | for (i=0; i<conf->copies; i++) { | |
2494 | int d = r10_bio->devs[i].devnum; | |
2495 | if (r10_bio->devs[i].bio->bi_end_io) | |
ab9d47e9 N |
2496 | rdev_dec_pending(conf->mirrors[d].rdev, |
2497 | mddev); | |
1da177e4 LT |
2498 | } |
2499 | put_buf(r10_bio); | |
2500 | biolist = NULL; | |
2501 | goto giveup; | |
2502 | } | |
2503 | } | |
2504 | ||
2505 | for (bio = biolist; bio ; bio=bio->bi_next) { | |
2506 | ||
2507 | bio->bi_flags &= ~(BIO_POOL_MASK - 1); | |
2508 | if (bio->bi_end_io) | |
2509 | bio->bi_flags |= 1 << BIO_UPTODATE; | |
2510 | bio->bi_vcnt = 0; | |
2511 | bio->bi_idx = 0; | |
2512 | bio->bi_phys_segments = 0; | |
1da177e4 LT |
2513 | bio->bi_size = 0; |
2514 | } | |
2515 | ||
2516 | nr_sectors = 0; | |
6cce3b23 N |
2517 | if (sector_nr + max_sync < max_sector) |
2518 | max_sector = sector_nr + max_sync; | |
1da177e4 LT |
2519 | do { |
2520 | struct page *page; | |
2521 | int len = PAGE_SIZE; | |
1da177e4 LT |
2522 | if (sector_nr + (len>>9) > max_sector) |
2523 | len = (max_sector - sector_nr) << 9; | |
2524 | if (len == 0) | |
2525 | break; | |
2526 | for (bio= biolist ; bio ; bio=bio->bi_next) { | |
ab9d47e9 | 2527 | struct bio *bio2; |
1da177e4 | 2528 | page = bio->bi_io_vec[bio->bi_vcnt].bv_page; |
ab9d47e9 N |
2529 | if (bio_add_page(bio, page, len, 0)) |
2530 | continue; | |
2531 | ||
2532 | /* stop here */ | |
2533 | bio->bi_io_vec[bio->bi_vcnt].bv_page = page; | |
2534 | for (bio2 = biolist; | |
2535 | bio2 && bio2 != bio; | |
2536 | bio2 = bio2->bi_next) { | |
2537 | /* remove last page from this bio */ | |
2538 | bio2->bi_vcnt--; | |
2539 | bio2->bi_size -= len; | |
2540 | bio2->bi_flags &= ~(1<< BIO_SEG_VALID); | |
1da177e4 | 2541 | } |
ab9d47e9 | 2542 | goto bio_full; |
1da177e4 LT |
2543 | } |
2544 | nr_sectors += len>>9; | |
2545 | sector_nr += len>>9; | |
2546 | } while (biolist->bi_vcnt < RESYNC_PAGES); | |
2547 | bio_full: | |
2548 | r10_bio->sectors = nr_sectors; | |
2549 | ||
2550 | while (biolist) { | |
2551 | bio = biolist; | |
2552 | biolist = biolist->bi_next; | |
2553 | ||
2554 | bio->bi_next = NULL; | |
2555 | r10_bio = bio->bi_private; | |
2556 | r10_bio->sectors = nr_sectors; | |
2557 | ||
2558 | if (bio->bi_end_io == end_sync_read) { | |
2559 | md_sync_acct(bio->bi_bdev, nr_sectors); | |
2560 | generic_make_request(bio); | |
2561 | } | |
2562 | } | |
2563 | ||
57afd89f N |
2564 | if (sectors_skipped) |
2565 | /* pretend they weren't skipped, it makes | |
2566 | * no important difference in this case | |
2567 | */ | |
2568 | md_done_sync(mddev, sectors_skipped, 1); | |
2569 | ||
1da177e4 LT |
2570 | return sectors_skipped + nr_sectors; |
2571 | giveup: | |
2572 | /* There is nowhere to write, so all non-sync | |
e875ecea N |
2573 | * drives must be failed or in resync, all drives |
2574 | * have a bad block, so try the next chunk... | |
1da177e4 | 2575 | */ |
09b4068a N |
2576 | if (sector_nr + max_sync < max_sector) |
2577 | max_sector = sector_nr + max_sync; | |
2578 | ||
2579 | sectors_skipped += (max_sector - sector_nr); | |
1da177e4 LT |
2580 | chunks_skipped ++; |
2581 | sector_nr = max_sector; | |
1da177e4 | 2582 | goto skipped; |
1da177e4 LT |
2583 | } |
2584 | ||
80c3a6ce DW |
2585 | static sector_t |
2586 | raid10_size(mddev_t *mddev, sector_t sectors, int raid_disks) | |
2587 | { | |
2588 | sector_t size; | |
070ec55d | 2589 | conf_t *conf = mddev->private; |
80c3a6ce DW |
2590 | |
2591 | if (!raid_disks) | |
84707f38 | 2592 | raid_disks = conf->raid_disks; |
80c3a6ce | 2593 | if (!sectors) |
dab8b292 | 2594 | sectors = conf->dev_sectors; |
80c3a6ce DW |
2595 | |
2596 | size = sectors >> conf->chunk_shift; | |
2597 | sector_div(size, conf->far_copies); | |
2598 | size = size * raid_disks; | |
2599 | sector_div(size, conf->near_copies); | |
2600 | ||
2601 | return size << conf->chunk_shift; | |
2602 | } | |
2603 | ||
dab8b292 TM |
2604 | |
2605 | static conf_t *setup_conf(mddev_t *mddev) | |
1da177e4 | 2606 | { |
dab8b292 | 2607 | conf_t *conf = NULL; |
c93983bf | 2608 | int nc, fc, fo; |
1da177e4 | 2609 | sector_t stride, size; |
dab8b292 | 2610 | int err = -EINVAL; |
1da177e4 | 2611 | |
f73ea873 MT |
2612 | if (mddev->new_chunk_sectors < (PAGE_SIZE >> 9) || |
2613 | !is_power_of_2(mddev->new_chunk_sectors)) { | |
128595ed N |
2614 | printk(KERN_ERR "md/raid10:%s: chunk size must be " |
2615 | "at least PAGE_SIZE(%ld) and be a power of 2.\n", | |
2616 | mdname(mddev), PAGE_SIZE); | |
dab8b292 | 2617 | goto out; |
1da177e4 | 2618 | } |
2604b703 | 2619 | |
f73ea873 MT |
2620 | nc = mddev->new_layout & 255; |
2621 | fc = (mddev->new_layout >> 8) & 255; | |
2622 | fo = mddev->new_layout & (1<<16); | |
dab8b292 | 2623 | |
1da177e4 | 2624 | if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks || |
f73ea873 | 2625 | (mddev->new_layout >> 17)) { |
128595ed | 2626 | printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n", |
f73ea873 | 2627 | mdname(mddev), mddev->new_layout); |
1da177e4 LT |
2628 | goto out; |
2629 | } | |
dab8b292 TM |
2630 | |
2631 | err = -ENOMEM; | |
4443ae10 | 2632 | conf = kzalloc(sizeof(conf_t), GFP_KERNEL); |
dab8b292 | 2633 | if (!conf) |
1da177e4 | 2634 | goto out; |
dab8b292 | 2635 | |
4443ae10 | 2636 | conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks, |
dab8b292 TM |
2637 | GFP_KERNEL); |
2638 | if (!conf->mirrors) | |
2639 | goto out; | |
4443ae10 N |
2640 | |
2641 | conf->tmppage = alloc_page(GFP_KERNEL); | |
2642 | if (!conf->tmppage) | |
dab8b292 TM |
2643 | goto out; |
2644 | ||
1da177e4 | 2645 | |
64a742bc | 2646 | conf->raid_disks = mddev->raid_disks; |
1da177e4 LT |
2647 | conf->near_copies = nc; |
2648 | conf->far_copies = fc; | |
2649 | conf->copies = nc*fc; | |
c93983bf | 2650 | conf->far_offset = fo; |
dab8b292 TM |
2651 | conf->chunk_mask = mddev->new_chunk_sectors - 1; |
2652 | conf->chunk_shift = ffz(~mddev->new_chunk_sectors); | |
2653 | ||
2654 | conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc, | |
2655 | r10bio_pool_free, conf); | |
2656 | if (!conf->r10bio_pool) | |
2657 | goto out; | |
2658 | ||
58c0fed4 | 2659 | size = mddev->dev_sectors >> conf->chunk_shift; |
64a742bc N |
2660 | sector_div(size, fc); |
2661 | size = size * conf->raid_disks; | |
2662 | sector_div(size, nc); | |
2663 | /* 'size' is now the number of chunks in the array */ | |
2664 | /* calculate "used chunks per device" in 'stride' */ | |
2665 | stride = size * conf->copies; | |
af03b8e4 N |
2666 | |
2667 | /* We need to round up when dividing by raid_disks to | |
2668 | * get the stride size. | |
2669 | */ | |
2670 | stride += conf->raid_disks - 1; | |
64a742bc | 2671 | sector_div(stride, conf->raid_disks); |
dab8b292 TM |
2672 | |
2673 | conf->dev_sectors = stride << conf->chunk_shift; | |
64a742bc | 2674 | |
c93983bf | 2675 | if (fo) |
64a742bc N |
2676 | stride = 1; |
2677 | else | |
c93983bf | 2678 | sector_div(stride, fc); |
64a742bc N |
2679 | conf->stride = stride << conf->chunk_shift; |
2680 | ||
1da177e4 | 2681 | |
e7e72bf6 | 2682 | spin_lock_init(&conf->device_lock); |
dab8b292 TM |
2683 | INIT_LIST_HEAD(&conf->retry_list); |
2684 | ||
2685 | spin_lock_init(&conf->resync_lock); | |
2686 | init_waitqueue_head(&conf->wait_barrier); | |
2687 | ||
2688 | conf->thread = md_register_thread(raid10d, mddev, NULL); | |
2689 | if (!conf->thread) | |
2690 | goto out; | |
2691 | ||
dab8b292 TM |
2692 | conf->mddev = mddev; |
2693 | return conf; | |
2694 | ||
2695 | out: | |
128595ed | 2696 | printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n", |
dab8b292 TM |
2697 | mdname(mddev)); |
2698 | if (conf) { | |
2699 | if (conf->r10bio_pool) | |
2700 | mempool_destroy(conf->r10bio_pool); | |
2701 | kfree(conf->mirrors); | |
2702 | safe_put_page(conf->tmppage); | |
2703 | kfree(conf); | |
2704 | } | |
2705 | return ERR_PTR(err); | |
2706 | } | |
2707 | ||
2708 | static int run(mddev_t *mddev) | |
2709 | { | |
2710 | conf_t *conf; | |
2711 | int i, disk_idx, chunk_size; | |
2712 | mirror_info_t *disk; | |
2713 | mdk_rdev_t *rdev; | |
2714 | sector_t size; | |
2715 | ||
2716 | /* | |
2717 | * copy the already verified devices into our private RAID10 | |
2718 | * bookkeeping area. [whatever we allocate in run(), | |
2719 | * should be freed in stop()] | |
2720 | */ | |
2721 | ||
2722 | if (mddev->private == NULL) { | |
2723 | conf = setup_conf(mddev); | |
2724 | if (IS_ERR(conf)) | |
2725 | return PTR_ERR(conf); | |
2726 | mddev->private = conf; | |
2727 | } | |
2728 | conf = mddev->private; | |
2729 | if (!conf) | |
2730 | goto out; | |
2731 | ||
dab8b292 TM |
2732 | mddev->thread = conf->thread; |
2733 | conf->thread = NULL; | |
2734 | ||
8f6c2e4b MP |
2735 | chunk_size = mddev->chunk_sectors << 9; |
2736 | blk_queue_io_min(mddev->queue, chunk_size); | |
2737 | if (conf->raid_disks % conf->near_copies) | |
2738 | blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks); | |
2739 | else | |
2740 | blk_queue_io_opt(mddev->queue, chunk_size * | |
2741 | (conf->raid_disks / conf->near_copies)); | |
2742 | ||
159ec1fc | 2743 | list_for_each_entry(rdev, &mddev->disks, same_set) { |
34b343cf | 2744 | |
1da177e4 | 2745 | disk_idx = rdev->raid_disk; |
84707f38 | 2746 | if (disk_idx >= conf->raid_disks |
1da177e4 LT |
2747 | || disk_idx < 0) |
2748 | continue; | |
2749 | disk = conf->mirrors + disk_idx; | |
2750 | ||
2751 | disk->rdev = rdev; | |
8f6c2e4b MP |
2752 | disk_stack_limits(mddev->gendisk, rdev->bdev, |
2753 | rdev->data_offset << 9); | |
1da177e4 | 2754 | /* as we don't honour merge_bvec_fn, we must never risk |
627a2d3c N |
2755 | * violating it, so limit max_segments to 1 lying |
2756 | * within a single page. | |
1da177e4 | 2757 | */ |
627a2d3c N |
2758 | if (rdev->bdev->bd_disk->queue->merge_bvec_fn) { |
2759 | blk_queue_max_segments(mddev->queue, 1); | |
2760 | blk_queue_segment_boundary(mddev->queue, | |
2761 | PAGE_CACHE_SIZE - 1); | |
2762 | } | |
1da177e4 LT |
2763 | |
2764 | disk->head_position = 0; | |
1da177e4 | 2765 | } |
6d508242 | 2766 | /* need to check that every block has at least one working mirror */ |
700c7213 | 2767 | if (!enough(conf, -1)) { |
128595ed | 2768 | printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n", |
6d508242 | 2769 | mdname(mddev)); |
1da177e4 LT |
2770 | goto out_free_conf; |
2771 | } | |
2772 | ||
2773 | mddev->degraded = 0; | |
2774 | for (i = 0; i < conf->raid_disks; i++) { | |
2775 | ||
2776 | disk = conf->mirrors + i; | |
2777 | ||
5fd6c1dc | 2778 | if (!disk->rdev || |
2e333e89 | 2779 | !test_bit(In_sync, &disk->rdev->flags)) { |
1da177e4 LT |
2780 | disk->head_position = 0; |
2781 | mddev->degraded++; | |
8c2e870a NB |
2782 | if (disk->rdev) |
2783 | conf->fullsync = 1; | |
1da177e4 LT |
2784 | } |
2785 | } | |
2786 | ||
8c6ac868 | 2787 | if (mddev->recovery_cp != MaxSector) |
128595ed | 2788 | printk(KERN_NOTICE "md/raid10:%s: not clean" |
8c6ac868 AN |
2789 | " -- starting background reconstruction\n", |
2790 | mdname(mddev)); | |
1da177e4 | 2791 | printk(KERN_INFO |
128595ed | 2792 | "md/raid10:%s: active with %d out of %d devices\n", |
84707f38 N |
2793 | mdname(mddev), conf->raid_disks - mddev->degraded, |
2794 | conf->raid_disks); | |
1da177e4 LT |
2795 | /* |
2796 | * Ok, everything is just fine now | |
2797 | */ | |
dab8b292 TM |
2798 | mddev->dev_sectors = conf->dev_sectors; |
2799 | size = raid10_size(mddev, 0, 0); | |
2800 | md_set_array_sectors(mddev, size); | |
2801 | mddev->resync_max_sectors = size; | |
1da177e4 | 2802 | |
0d129228 N |
2803 | mddev->queue->backing_dev_info.congested_fn = raid10_congested; |
2804 | mddev->queue->backing_dev_info.congested_data = mddev; | |
7a5febe9 | 2805 | |
1da177e4 LT |
2806 | /* Calculate max read-ahead size. |
2807 | * We need to readahead at least twice a whole stripe.... | |
2808 | * maybe... | |
2809 | */ | |
2810 | { | |
9d8f0363 AN |
2811 | int stripe = conf->raid_disks * |
2812 | ((mddev->chunk_sectors << 9) / PAGE_SIZE); | |
1da177e4 LT |
2813 | stripe /= conf->near_copies; |
2814 | if (mddev->queue->backing_dev_info.ra_pages < 2* stripe) | |
2815 | mddev->queue->backing_dev_info.ra_pages = 2* stripe; | |
2816 | } | |
2817 | ||
84707f38 | 2818 | if (conf->near_copies < conf->raid_disks) |
1da177e4 | 2819 | blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec); |
a91a2785 MP |
2820 | |
2821 | if (md_integrity_register(mddev)) | |
2822 | goto out_free_conf; | |
2823 | ||
1da177e4 LT |
2824 | return 0; |
2825 | ||
2826 | out_free_conf: | |
589a594b | 2827 | md_unregister_thread(mddev->thread); |
1da177e4 LT |
2828 | if (conf->r10bio_pool) |
2829 | mempool_destroy(conf->r10bio_pool); | |
1345b1d8 | 2830 | safe_put_page(conf->tmppage); |
990a8baf | 2831 | kfree(conf->mirrors); |
1da177e4 LT |
2832 | kfree(conf); |
2833 | mddev->private = NULL; | |
2834 | out: | |
2835 | return -EIO; | |
2836 | } | |
2837 | ||
2838 | static int stop(mddev_t *mddev) | |
2839 | { | |
070ec55d | 2840 | conf_t *conf = mddev->private; |
1da177e4 | 2841 | |
409c57f3 N |
2842 | raise_barrier(conf, 0); |
2843 | lower_barrier(conf); | |
2844 | ||
1da177e4 LT |
2845 | md_unregister_thread(mddev->thread); |
2846 | mddev->thread = NULL; | |
2847 | blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/ | |
2848 | if (conf->r10bio_pool) | |
2849 | mempool_destroy(conf->r10bio_pool); | |
990a8baf | 2850 | kfree(conf->mirrors); |
1da177e4 LT |
2851 | kfree(conf); |
2852 | mddev->private = NULL; | |
2853 | return 0; | |
2854 | } | |
2855 | ||
6cce3b23 N |
2856 | static void raid10_quiesce(mddev_t *mddev, int state) |
2857 | { | |
070ec55d | 2858 | conf_t *conf = mddev->private; |
6cce3b23 N |
2859 | |
2860 | switch(state) { | |
2861 | case 1: | |
2862 | raise_barrier(conf, 0); | |
2863 | break; | |
2864 | case 0: | |
2865 | lower_barrier(conf); | |
2866 | break; | |
2867 | } | |
6cce3b23 | 2868 | } |
1da177e4 | 2869 | |
dab8b292 TM |
2870 | static void *raid10_takeover_raid0(mddev_t *mddev) |
2871 | { | |
2872 | mdk_rdev_t *rdev; | |
2873 | conf_t *conf; | |
2874 | ||
2875 | if (mddev->degraded > 0) { | |
128595ed N |
2876 | printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n", |
2877 | mdname(mddev)); | |
dab8b292 TM |
2878 | return ERR_PTR(-EINVAL); |
2879 | } | |
2880 | ||
dab8b292 TM |
2881 | /* Set new parameters */ |
2882 | mddev->new_level = 10; | |
2883 | /* new layout: far_copies = 1, near_copies = 2 */ | |
2884 | mddev->new_layout = (1<<8) + 2; | |
2885 | mddev->new_chunk_sectors = mddev->chunk_sectors; | |
2886 | mddev->delta_disks = mddev->raid_disks; | |
dab8b292 TM |
2887 | mddev->raid_disks *= 2; |
2888 | /* make sure it will be not marked as dirty */ | |
2889 | mddev->recovery_cp = MaxSector; | |
2890 | ||
2891 | conf = setup_conf(mddev); | |
02214dc5 | 2892 | if (!IS_ERR(conf)) { |
e93f68a1 N |
2893 | list_for_each_entry(rdev, &mddev->disks, same_set) |
2894 | if (rdev->raid_disk >= 0) | |
2895 | rdev->new_raid_disk = rdev->raid_disk * 2; | |
02214dc5 KW |
2896 | conf->barrier = 1; |
2897 | } | |
2898 | ||
dab8b292 TM |
2899 | return conf; |
2900 | } | |
2901 | ||
2902 | static void *raid10_takeover(mddev_t *mddev) | |
2903 | { | |
2904 | struct raid0_private_data *raid0_priv; | |
2905 | ||
2906 | /* raid10 can take over: | |
2907 | * raid0 - providing it has only two drives | |
2908 | */ | |
2909 | if (mddev->level == 0) { | |
2910 | /* for raid0 takeover only one zone is supported */ | |
2911 | raid0_priv = mddev->private; | |
2912 | if (raid0_priv->nr_strip_zones > 1) { | |
128595ed N |
2913 | printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0" |
2914 | " with more than one zone.\n", | |
2915 | mdname(mddev)); | |
dab8b292 TM |
2916 | return ERR_PTR(-EINVAL); |
2917 | } | |
2918 | return raid10_takeover_raid0(mddev); | |
2919 | } | |
2920 | return ERR_PTR(-EINVAL); | |
2921 | } | |
2922 | ||
2604b703 | 2923 | static struct mdk_personality raid10_personality = |
1da177e4 LT |
2924 | { |
2925 | .name = "raid10", | |
2604b703 | 2926 | .level = 10, |
1da177e4 LT |
2927 | .owner = THIS_MODULE, |
2928 | .make_request = make_request, | |
2929 | .run = run, | |
2930 | .stop = stop, | |
2931 | .status = status, | |
2932 | .error_handler = error, | |
2933 | .hot_add_disk = raid10_add_disk, | |
2934 | .hot_remove_disk= raid10_remove_disk, | |
2935 | .spare_active = raid10_spare_active, | |
2936 | .sync_request = sync_request, | |
6cce3b23 | 2937 | .quiesce = raid10_quiesce, |
80c3a6ce | 2938 | .size = raid10_size, |
dab8b292 | 2939 | .takeover = raid10_takeover, |
1da177e4 LT |
2940 | }; |
2941 | ||
2942 | static int __init raid_init(void) | |
2943 | { | |
2604b703 | 2944 | return register_md_personality(&raid10_personality); |
1da177e4 LT |
2945 | } |
2946 | ||
2947 | static void raid_exit(void) | |
2948 | { | |
2604b703 | 2949 | unregister_md_personality(&raid10_personality); |
1da177e4 LT |
2950 | } |
2951 | ||
2952 | module_init(raid_init); | |
2953 | module_exit(raid_exit); | |
2954 | MODULE_LICENSE("GPL"); | |
0efb9e61 | 2955 | MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD"); |
1da177e4 | 2956 | MODULE_ALIAS("md-personality-9"); /* RAID10 */ |
d9d166c2 | 2957 | MODULE_ALIAS("md-raid10"); |
2604b703 | 2958 | MODULE_ALIAS("md-level-10"); |