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 | * | |
8 | * Base on code in raid1.c. See raid1.c for futher copyright information. | |
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 | ||
25570727 | 21 | #include <linux/delay.h> |
bff61975 | 22 | #include <linux/blkdev.h> |
bff61975 | 23 | #include <linux/seq_file.h> |
43b2e5d8 | 24 | #include "md.h" |
bff61975 | 25 | #include "dm-bio-list.h" |
ef740c37 CH |
26 | #include "raid10.h" |
27 | #include "bitmap.h" | |
1da177e4 LT |
28 | |
29 | /* | |
30 | * RAID10 provides a combination of RAID0 and RAID1 functionality. | |
31 | * The layout of data is defined by | |
32 | * chunk_size | |
33 | * raid_disks | |
34 | * near_copies (stored in low byte of layout) | |
35 | * far_copies (stored in second byte of layout) | |
c93983bf | 36 | * far_offset (stored in bit 16 of layout ) |
1da177e4 LT |
37 | * |
38 | * The data to be stored is divided into chunks using chunksize. | |
39 | * Each device is divided into far_copies sections. | |
40 | * In each section, chunks are laid out in a style similar to raid0, but | |
41 | * near_copies copies of each chunk is stored (each on a different drive). | |
42 | * The starting device for each section is offset near_copies from the starting | |
43 | * device of the previous section. | |
c93983bf | 44 | * Thus they are (near_copies*far_copies) of each chunk, and each is on a different |
1da177e4 LT |
45 | * drive. |
46 | * near_copies and far_copies must be at least one, and their product is at most | |
47 | * raid_disks. | |
c93983bf N |
48 | * |
49 | * If far_offset is true, then the far_copies are handled a bit differently. | |
50 | * The copies are still in different stripes, but instead of be very far apart | |
51 | * on disk, there are adjacent stripes. | |
1da177e4 LT |
52 | */ |
53 | ||
54 | /* | |
55 | * Number of guaranteed r10bios in case of extreme VM load: | |
56 | */ | |
57 | #define NR_RAID10_BIOS 256 | |
58 | ||
59 | static void unplug_slaves(mddev_t *mddev); | |
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; | |
67 | r10bio_t *r10_bio; | |
68 | int size = offsetof(struct r10bio_s, devs[conf->copies]); | |
69 | ||
70 | /* allocate a r10bio with room for raid_disks entries in the bios array */ | |
9ffae0cf N |
71 | r10_bio = kzalloc(size, gfp_flags); |
72 | if (!r10_bio) | |
1da177e4 LT |
73 | unplug_slaves(conf->mddev); |
74 | ||
75 | return r10_bio; | |
76 | } | |
77 | ||
78 | static void r10bio_pool_free(void *r10_bio, void *data) | |
79 | { | |
80 | kfree(r10_bio); | |
81 | } | |
82 | ||
0310fa21 | 83 | /* Maximum size of each resync request */ |
1da177e4 | 84 | #define RESYNC_BLOCK_SIZE (64*1024) |
1da177e4 | 85 | #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE) |
0310fa21 N |
86 | /* amount of memory to reserve for resync requests */ |
87 | #define RESYNC_WINDOW (1024*1024) | |
88 | /* maximum number of concurrent requests, memory permitting */ | |
89 | #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE) | |
1da177e4 LT |
90 | |
91 | /* | |
92 | * When performing a resync, we need to read and compare, so | |
93 | * we need as many pages are there are copies. | |
94 | * When performing a recovery, we need 2 bios, one for read, | |
95 | * one for write (we recover only one drive per r10buf) | |
96 | * | |
97 | */ | |
dd0fc66f | 98 | static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data) |
1da177e4 LT |
99 | { |
100 | conf_t *conf = data; | |
101 | struct page *page; | |
102 | r10bio_t *r10_bio; | |
103 | struct bio *bio; | |
104 | int i, j; | |
105 | int nalloc; | |
106 | ||
107 | r10_bio = r10bio_pool_alloc(gfp_flags, conf); | |
108 | if (!r10_bio) { | |
109 | unplug_slaves(conf->mddev); | |
110 | return NULL; | |
111 | } | |
112 | ||
113 | if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery)) | |
114 | nalloc = conf->copies; /* resync */ | |
115 | else | |
116 | nalloc = 2; /* recovery */ | |
117 | ||
118 | /* | |
119 | * Allocate bios. | |
120 | */ | |
121 | for (j = nalloc ; j-- ; ) { | |
122 | bio = bio_alloc(gfp_flags, RESYNC_PAGES); | |
123 | if (!bio) | |
124 | goto out_free_bio; | |
125 | r10_bio->devs[j].bio = bio; | |
126 | } | |
127 | /* | |
128 | * Allocate RESYNC_PAGES data pages and attach them | |
129 | * where needed. | |
130 | */ | |
131 | for (j = 0 ; j < nalloc; j++) { | |
132 | bio = r10_bio->devs[j].bio; | |
133 | for (i = 0; i < RESYNC_PAGES; i++) { | |
134 | page = alloc_page(gfp_flags); | |
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; | |
0eb3ff12 | 184 | if (*bio && *bio != IO_BLOCKED) |
1da177e4 LT |
185 | bio_put(*bio); |
186 | *bio = NULL; | |
187 | } | |
188 | } | |
189 | ||
858119e1 | 190 | static void free_r10bio(r10bio_t *r10_bio) |
1da177e4 | 191 | { |
1da177e4 LT |
192 | conf_t *conf = mddev_to_conf(r10_bio->mddev); |
193 | ||
194 | /* | |
195 | * Wake up any possible resync thread that waits for the device | |
196 | * to go idle. | |
197 | */ | |
0a27ec96 | 198 | allow_barrier(conf); |
1da177e4 LT |
199 | |
200 | put_all_bios(conf, r10_bio); | |
201 | mempool_free(r10_bio, conf->r10bio_pool); | |
202 | } | |
203 | ||
858119e1 | 204 | static void put_buf(r10bio_t *r10_bio) |
1da177e4 LT |
205 | { |
206 | conf_t *conf = mddev_to_conf(r10_bio->mddev); | |
1da177e4 LT |
207 | |
208 | mempool_free(r10_bio, conf->r10buf_pool); | |
209 | ||
0a27ec96 | 210 | lower_barrier(conf); |
1da177e4 LT |
211 | } |
212 | ||
213 | static void reschedule_retry(r10bio_t *r10_bio) | |
214 | { | |
215 | unsigned long flags; | |
216 | mddev_t *mddev = r10_bio->mddev; | |
217 | conf_t *conf = mddev_to_conf(mddev); | |
218 | ||
219 | spin_lock_irqsave(&conf->device_lock, flags); | |
220 | list_add(&r10_bio->retry_list, &conf->retry_list); | |
4443ae10 | 221 | conf->nr_queued ++; |
1da177e4 LT |
222 | spin_unlock_irqrestore(&conf->device_lock, flags); |
223 | ||
388667be AJ |
224 | /* wake up frozen array... */ |
225 | wake_up(&conf->wait_barrier); | |
226 | ||
1da177e4 LT |
227 | md_wakeup_thread(mddev->thread); |
228 | } | |
229 | ||
230 | /* | |
231 | * raid_end_bio_io() is called when we have finished servicing a mirrored | |
232 | * operation and are ready to return a success/failure code to the buffer | |
233 | * cache layer. | |
234 | */ | |
235 | static void raid_end_bio_io(r10bio_t *r10_bio) | |
236 | { | |
237 | struct bio *bio = r10_bio->master_bio; | |
238 | ||
6712ecf8 | 239 | bio_endio(bio, |
1da177e4 LT |
240 | test_bit(R10BIO_Uptodate, &r10_bio->state) ? 0 : -EIO); |
241 | free_r10bio(r10_bio); | |
242 | } | |
243 | ||
244 | /* | |
245 | * Update disk head position estimator based on IRQ completion info. | |
246 | */ | |
247 | static inline void update_head_pos(int slot, r10bio_t *r10_bio) | |
248 | { | |
249 | conf_t *conf = mddev_to_conf(r10_bio->mddev); | |
250 | ||
251 | conf->mirrors[r10_bio->devs[slot].devnum].head_position = | |
252 | r10_bio->devs[slot].addr + (r10_bio->sectors); | |
253 | } | |
254 | ||
6712ecf8 | 255 | static void raid10_end_read_request(struct bio *bio, int error) |
1da177e4 LT |
256 | { |
257 | int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags); | |
258 | r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private); | |
259 | int slot, dev; | |
260 | conf_t *conf = mddev_to_conf(r10_bio->mddev); | |
261 | ||
1da177e4 LT |
262 | |
263 | slot = r10_bio->read_slot; | |
264 | dev = r10_bio->devs[slot].devnum; | |
265 | /* | |
266 | * this branch is our 'one mirror IO has finished' event handler: | |
267 | */ | |
4443ae10 N |
268 | update_head_pos(slot, r10_bio); |
269 | ||
270 | if (uptodate) { | |
1da177e4 LT |
271 | /* |
272 | * Set R10BIO_Uptodate in our master bio, so that | |
273 | * we will return a good error code to the higher | |
274 | * levels even if IO on some other mirrored buffer fails. | |
275 | * | |
276 | * The 'master' represents the composite IO operation to | |
277 | * user-side. So if something waits for IO, then it will | |
278 | * wait for the 'master' bio. | |
279 | */ | |
280 | set_bit(R10BIO_Uptodate, &r10_bio->state); | |
1da177e4 | 281 | raid_end_bio_io(r10_bio); |
4443ae10 | 282 | } else { |
1da177e4 LT |
283 | /* |
284 | * oops, read error: | |
285 | */ | |
286 | char b[BDEVNAME_SIZE]; | |
287 | if (printk_ratelimit()) | |
288 | printk(KERN_ERR "raid10: %s: rescheduling sector %llu\n", | |
289 | bdevname(conf->mirrors[dev].rdev->bdev,b), (unsigned long long)r10_bio->sector); | |
290 | reschedule_retry(r10_bio); | |
291 | } | |
292 | ||
293 | rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev); | |
1da177e4 LT |
294 | } |
295 | ||
6712ecf8 | 296 | static void raid10_end_write_request(struct bio *bio, int error) |
1da177e4 LT |
297 | { |
298 | int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags); | |
299 | r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private); | |
300 | int slot, dev; | |
301 | conf_t *conf = mddev_to_conf(r10_bio->mddev); | |
302 | ||
1da177e4 LT |
303 | for (slot = 0; slot < conf->copies; slot++) |
304 | if (r10_bio->devs[slot].bio == bio) | |
305 | break; | |
306 | dev = r10_bio->devs[slot].devnum; | |
307 | ||
308 | /* | |
309 | * this branch is our 'one mirror IO has finished' event handler: | |
310 | */ | |
6cce3b23 | 311 | if (!uptodate) { |
1da177e4 | 312 | md_error(r10_bio->mddev, conf->mirrors[dev].rdev); |
6cce3b23 N |
313 | /* an I/O failed, we can't clear the bitmap */ |
314 | set_bit(R10BIO_Degraded, &r10_bio->state); | |
315 | } else | |
1da177e4 LT |
316 | /* |
317 | * Set R10BIO_Uptodate in our master bio, so that | |
318 | * we will return a good error code for to the higher | |
319 | * levels even if IO on some other mirrored buffer fails. | |
320 | * | |
321 | * The 'master' represents the composite IO operation to | |
322 | * user-side. So if something waits for IO, then it will | |
323 | * wait for the 'master' bio. | |
324 | */ | |
325 | set_bit(R10BIO_Uptodate, &r10_bio->state); | |
326 | ||
327 | update_head_pos(slot, r10_bio); | |
328 | ||
329 | /* | |
330 | * | |
331 | * Let's see if all mirrored write operations have finished | |
332 | * already. | |
333 | */ | |
334 | if (atomic_dec_and_test(&r10_bio->remaining)) { | |
6cce3b23 N |
335 | /* clear the bitmap if all writes complete successfully */ |
336 | bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector, | |
337 | r10_bio->sectors, | |
338 | !test_bit(R10BIO_Degraded, &r10_bio->state), | |
339 | 0); | |
1da177e4 LT |
340 | md_write_end(r10_bio->mddev); |
341 | raid_end_bio_io(r10_bio); | |
342 | } | |
343 | ||
344 | rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev); | |
1da177e4 LT |
345 | } |
346 | ||
347 | ||
348 | /* | |
349 | * RAID10 layout manager | |
350 | * Aswell as the chunksize and raid_disks count, there are two | |
351 | * parameters: near_copies and far_copies. | |
352 | * near_copies * far_copies must be <= raid_disks. | |
353 | * Normally one of these will be 1. | |
354 | * If both are 1, we get raid0. | |
355 | * If near_copies == raid_disks, we get raid1. | |
356 | * | |
357 | * Chunks are layed out in raid0 style with near_copies copies of the | |
358 | * first chunk, followed by near_copies copies of the next chunk and | |
359 | * so on. | |
360 | * If far_copies > 1, then after 1/far_copies of the array has been assigned | |
361 | * as described above, we start again with a device offset of near_copies. | |
362 | * So we effectively have another copy of the whole array further down all | |
363 | * the drives, but with blocks on different drives. | |
364 | * With this layout, and block is never stored twice on the one device. | |
365 | * | |
366 | * raid10_find_phys finds the sector offset of a given virtual sector | |
c93983bf | 367 | * on each device that it is on. |
1da177e4 LT |
368 | * |
369 | * raid10_find_virt does the reverse mapping, from a device and a | |
370 | * sector offset to a virtual address | |
371 | */ | |
372 | ||
373 | static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio) | |
374 | { | |
375 | int n,f; | |
376 | sector_t sector; | |
377 | sector_t chunk; | |
378 | sector_t stripe; | |
379 | int dev; | |
380 | ||
381 | int slot = 0; | |
382 | ||
383 | /* now calculate first sector/dev */ | |
384 | chunk = r10bio->sector >> conf->chunk_shift; | |
385 | sector = r10bio->sector & conf->chunk_mask; | |
386 | ||
387 | chunk *= conf->near_copies; | |
388 | stripe = chunk; | |
389 | dev = sector_div(stripe, conf->raid_disks); | |
c93983bf N |
390 | if (conf->far_offset) |
391 | stripe *= conf->far_copies; | |
1da177e4 LT |
392 | |
393 | sector += stripe << conf->chunk_shift; | |
394 | ||
395 | /* and calculate all the others */ | |
396 | for (n=0; n < conf->near_copies; n++) { | |
397 | int d = dev; | |
398 | sector_t s = sector; | |
399 | r10bio->devs[slot].addr = sector; | |
400 | r10bio->devs[slot].devnum = d; | |
401 | slot++; | |
402 | ||
403 | for (f = 1; f < conf->far_copies; f++) { | |
404 | d += conf->near_copies; | |
405 | if (d >= conf->raid_disks) | |
406 | d -= conf->raid_disks; | |
407 | s += conf->stride; | |
408 | r10bio->devs[slot].devnum = d; | |
409 | r10bio->devs[slot].addr = s; | |
410 | slot++; | |
411 | } | |
412 | dev++; | |
413 | if (dev >= conf->raid_disks) { | |
414 | dev = 0; | |
415 | sector += (conf->chunk_mask + 1); | |
416 | } | |
417 | } | |
418 | BUG_ON(slot != conf->copies); | |
419 | } | |
420 | ||
421 | static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev) | |
422 | { | |
423 | sector_t offset, chunk, vchunk; | |
424 | ||
1da177e4 | 425 | offset = sector & conf->chunk_mask; |
c93983bf N |
426 | if (conf->far_offset) { |
427 | int fc; | |
428 | chunk = sector >> conf->chunk_shift; | |
429 | fc = sector_div(chunk, conf->far_copies); | |
430 | dev -= fc * conf->near_copies; | |
431 | if (dev < 0) | |
432 | dev += conf->raid_disks; | |
433 | } else { | |
64a742bc | 434 | while (sector >= conf->stride) { |
c93983bf N |
435 | sector -= conf->stride; |
436 | if (dev < conf->near_copies) | |
437 | dev += conf->raid_disks - conf->near_copies; | |
438 | else | |
439 | dev -= conf->near_copies; | |
440 | } | |
441 | chunk = sector >> conf->chunk_shift; | |
442 | } | |
1da177e4 LT |
443 | vchunk = chunk * conf->raid_disks + dev; |
444 | sector_div(vchunk, conf->near_copies); | |
445 | return (vchunk << conf->chunk_shift) + offset; | |
446 | } | |
447 | ||
448 | /** | |
449 | * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged | |
450 | * @q: request queue | |
cc371e66 | 451 | * @bvm: properties of new bio |
1da177e4 LT |
452 | * @biovec: the request that could be merged to it. |
453 | * | |
454 | * Return amount of bytes we can accept at this offset | |
455 | * If near_copies == raid_disk, there are no striping issues, | |
456 | * but in that case, the function isn't called at all. | |
457 | */ | |
cc371e66 AK |
458 | static int raid10_mergeable_bvec(struct request_queue *q, |
459 | struct bvec_merge_data *bvm, | |
460 | struct bio_vec *biovec) | |
1da177e4 LT |
461 | { |
462 | mddev_t *mddev = q->queuedata; | |
cc371e66 | 463 | sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev); |
1da177e4 LT |
464 | int max; |
465 | unsigned int chunk_sectors = mddev->chunk_size >> 9; | |
cc371e66 | 466 | unsigned int bio_sectors = bvm->bi_size >> 9; |
1da177e4 LT |
467 | |
468 | max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9; | |
469 | if (max < 0) max = 0; /* bio_add cannot handle a negative return */ | |
cc371e66 AK |
470 | if (max <= biovec->bv_len && bio_sectors == 0) |
471 | return biovec->bv_len; | |
1da177e4 LT |
472 | else |
473 | return max; | |
474 | } | |
475 | ||
476 | /* | |
477 | * This routine returns the disk from which the requested read should | |
478 | * be done. There is a per-array 'next expected sequential IO' sector | |
479 | * number - if this matches on the next IO then we use the last disk. | |
480 | * There is also a per-disk 'last know head position' sector that is | |
481 | * maintained from IRQ contexts, both the normal and the resync IO | |
482 | * completion handlers update this position correctly. If there is no | |
483 | * perfect sequential match then we pick the disk whose head is closest. | |
484 | * | |
485 | * If there are 2 mirrors in the same 2 devices, performance degrades | |
486 | * because position is mirror, not device based. | |
487 | * | |
488 | * The rdev for the device selected will have nr_pending incremented. | |
489 | */ | |
490 | ||
491 | /* | |
492 | * FIXME: possibly should rethink readbalancing and do it differently | |
493 | * depending on near_copies / far_copies geometry. | |
494 | */ | |
495 | static int read_balance(conf_t *conf, r10bio_t *r10_bio) | |
496 | { | |
497 | const unsigned long this_sector = r10_bio->sector; | |
498 | int disk, slot, nslot; | |
499 | const int sectors = r10_bio->sectors; | |
500 | sector_t new_distance, current_distance; | |
d6065f7b | 501 | mdk_rdev_t *rdev; |
1da177e4 LT |
502 | |
503 | raid10_find_phys(conf, r10_bio); | |
504 | rcu_read_lock(); | |
505 | /* | |
506 | * Check if we can balance. We can balance on the whole | |
6cce3b23 N |
507 | * device if no resync is going on (recovery is ok), or below |
508 | * the resync window. We take the first readable disk when | |
509 | * above the resync window. | |
1da177e4 LT |
510 | */ |
511 | if (conf->mddev->recovery_cp < MaxSector | |
512 | && (this_sector + sectors >= conf->next_resync)) { | |
513 | /* make sure that disk is operational */ | |
514 | slot = 0; | |
515 | disk = r10_bio->devs[slot].devnum; | |
516 | ||
d6065f7b | 517 | while ((rdev = rcu_dereference(conf->mirrors[disk].rdev)) == NULL || |
0eb3ff12 | 518 | r10_bio->devs[slot].bio == IO_BLOCKED || |
b2d444d7 | 519 | !test_bit(In_sync, &rdev->flags)) { |
1da177e4 LT |
520 | slot++; |
521 | if (slot == conf->copies) { | |
522 | slot = 0; | |
523 | disk = -1; | |
524 | break; | |
525 | } | |
526 | disk = r10_bio->devs[slot].devnum; | |
527 | } | |
528 | goto rb_out; | |
529 | } | |
530 | ||
531 | ||
532 | /* make sure the disk is operational */ | |
533 | slot = 0; | |
534 | disk = r10_bio->devs[slot].devnum; | |
d6065f7b | 535 | while ((rdev=rcu_dereference(conf->mirrors[disk].rdev)) == NULL || |
0eb3ff12 | 536 | r10_bio->devs[slot].bio == IO_BLOCKED || |
b2d444d7 | 537 | !test_bit(In_sync, &rdev->flags)) { |
1da177e4 LT |
538 | slot ++; |
539 | if (slot == conf->copies) { | |
540 | disk = -1; | |
541 | goto rb_out; | |
542 | } | |
543 | disk = r10_bio->devs[slot].devnum; | |
544 | } | |
545 | ||
546 | ||
3ec67ac1 N |
547 | current_distance = abs(r10_bio->devs[slot].addr - |
548 | conf->mirrors[disk].head_position); | |
1da177e4 | 549 | |
8ed3a195 KS |
550 | /* Find the disk whose head is closest, |
551 | * or - for far > 1 - find the closest to partition beginning */ | |
1da177e4 LT |
552 | |
553 | for (nslot = slot; nslot < conf->copies; nslot++) { | |
554 | int ndisk = r10_bio->devs[nslot].devnum; | |
555 | ||
556 | ||
d6065f7b | 557 | if ((rdev=rcu_dereference(conf->mirrors[ndisk].rdev)) == NULL || |
0eb3ff12 | 558 | r10_bio->devs[nslot].bio == IO_BLOCKED || |
b2d444d7 | 559 | !test_bit(In_sync, &rdev->flags)) |
1da177e4 LT |
560 | continue; |
561 | ||
22dfdf52 N |
562 | /* This optimisation is debatable, and completely destroys |
563 | * sequential read speed for 'far copies' arrays. So only | |
564 | * keep it for 'near' arrays, and review those later. | |
565 | */ | |
566 | if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending)) { | |
1da177e4 LT |
567 | disk = ndisk; |
568 | slot = nslot; | |
569 | break; | |
570 | } | |
8ed3a195 KS |
571 | |
572 | /* for far > 1 always use the lowest address */ | |
573 | if (conf->far_copies > 1) | |
574 | new_distance = r10_bio->devs[nslot].addr; | |
575 | else | |
576 | new_distance = abs(r10_bio->devs[nslot].addr - | |
577 | conf->mirrors[ndisk].head_position); | |
1da177e4 LT |
578 | if (new_distance < current_distance) { |
579 | current_distance = new_distance; | |
580 | disk = ndisk; | |
581 | slot = nslot; | |
582 | } | |
583 | } | |
584 | ||
585 | rb_out: | |
586 | r10_bio->read_slot = slot; | |
587 | /* conf->next_seq_sect = this_sector + sectors;*/ | |
588 | ||
d6065f7b | 589 | if (disk >= 0 && (rdev=rcu_dereference(conf->mirrors[disk].rdev))!= NULL) |
1da177e4 | 590 | atomic_inc(&conf->mirrors[disk].rdev->nr_pending); |
29fc7e3e N |
591 | else |
592 | disk = -1; | |
1da177e4 LT |
593 | rcu_read_unlock(); |
594 | ||
595 | return disk; | |
596 | } | |
597 | ||
598 | static void unplug_slaves(mddev_t *mddev) | |
599 | { | |
600 | conf_t *conf = mddev_to_conf(mddev); | |
601 | int i; | |
602 | ||
603 | rcu_read_lock(); | |
604 | for (i=0; i<mddev->raid_disks; i++) { | |
d6065f7b | 605 | mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev); |
b2d444d7 | 606 | if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) { |
165125e1 | 607 | struct request_queue *r_queue = bdev_get_queue(rdev->bdev); |
1da177e4 LT |
608 | |
609 | atomic_inc(&rdev->nr_pending); | |
610 | rcu_read_unlock(); | |
611 | ||
2ad8b1ef | 612 | blk_unplug(r_queue); |
1da177e4 LT |
613 | |
614 | rdev_dec_pending(rdev, mddev); | |
615 | rcu_read_lock(); | |
616 | } | |
617 | } | |
618 | rcu_read_unlock(); | |
619 | } | |
620 | ||
165125e1 | 621 | static void raid10_unplug(struct request_queue *q) |
1da177e4 | 622 | { |
6cce3b23 N |
623 | mddev_t *mddev = q->queuedata; |
624 | ||
1da177e4 | 625 | unplug_slaves(q->queuedata); |
6cce3b23 | 626 | md_wakeup_thread(mddev->thread); |
1da177e4 LT |
627 | } |
628 | ||
0d129228 N |
629 | static int raid10_congested(void *data, int bits) |
630 | { | |
631 | mddev_t *mddev = data; | |
632 | conf_t *conf = mddev_to_conf(mddev); | |
633 | int i, ret = 0; | |
634 | ||
635 | rcu_read_lock(); | |
636 | for (i = 0; i < mddev->raid_disks && ret == 0; i++) { | |
637 | mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev); | |
638 | if (rdev && !test_bit(Faulty, &rdev->flags)) { | |
165125e1 | 639 | struct request_queue *q = bdev_get_queue(rdev->bdev); |
0d129228 N |
640 | |
641 | ret |= bdi_congested(&q->backing_dev_info, bits); | |
642 | } | |
643 | } | |
644 | rcu_read_unlock(); | |
645 | return ret; | |
646 | } | |
647 | ||
a35e63ef N |
648 | static int flush_pending_writes(conf_t *conf) |
649 | { | |
650 | /* Any writes that have been queued but are awaiting | |
651 | * bitmap updates get flushed here. | |
652 | * We return 1 if any requests were actually submitted. | |
653 | */ | |
654 | int rv = 0; | |
655 | ||
656 | spin_lock_irq(&conf->device_lock); | |
657 | ||
658 | if (conf->pending_bio_list.head) { | |
659 | struct bio *bio; | |
660 | bio = bio_list_get(&conf->pending_bio_list); | |
661 | blk_remove_plug(conf->mddev->queue); | |
662 | spin_unlock_irq(&conf->device_lock); | |
663 | /* flush any pending bitmap writes to disk | |
664 | * before proceeding w/ I/O */ | |
665 | bitmap_unplug(conf->mddev->bitmap); | |
666 | ||
667 | while (bio) { /* submit pending writes */ | |
668 | struct bio *next = bio->bi_next; | |
669 | bio->bi_next = NULL; | |
670 | generic_make_request(bio); | |
671 | bio = next; | |
672 | } | |
673 | rv = 1; | |
674 | } else | |
675 | spin_unlock_irq(&conf->device_lock); | |
676 | return rv; | |
677 | } | |
0a27ec96 N |
678 | /* Barriers.... |
679 | * Sometimes we need to suspend IO while we do something else, | |
680 | * either some resync/recovery, or reconfigure the array. | |
681 | * To do this we raise a 'barrier'. | |
682 | * The 'barrier' is a counter that can be raised multiple times | |
683 | * to count how many activities are happening which preclude | |
684 | * normal IO. | |
685 | * We can only raise the barrier if there is no pending IO. | |
686 | * i.e. if nr_pending == 0. | |
687 | * We choose only to raise the barrier if no-one is waiting for the | |
688 | * barrier to go down. This means that as soon as an IO request | |
689 | * is ready, no other operations which require a barrier will start | |
690 | * until the IO request has had a chance. | |
691 | * | |
692 | * So: regular IO calls 'wait_barrier'. When that returns there | |
693 | * is no backgroup IO happening, It must arrange to call | |
694 | * allow_barrier when it has finished its IO. | |
695 | * backgroup IO calls must call raise_barrier. Once that returns | |
696 | * there is no normal IO happeing. It must arrange to call | |
697 | * lower_barrier when the particular background IO completes. | |
1da177e4 | 698 | */ |
1da177e4 | 699 | |
6cce3b23 | 700 | static void raise_barrier(conf_t *conf, int force) |
1da177e4 | 701 | { |
6cce3b23 | 702 | BUG_ON(force && !conf->barrier); |
1da177e4 | 703 | spin_lock_irq(&conf->resync_lock); |
0a27ec96 | 704 | |
6cce3b23 N |
705 | /* Wait until no block IO is waiting (unless 'force') */ |
706 | wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting, | |
0a27ec96 N |
707 | conf->resync_lock, |
708 | raid10_unplug(conf->mddev->queue)); | |
709 | ||
710 | /* block any new IO from starting */ | |
711 | conf->barrier++; | |
712 | ||
713 | /* No wait for all pending IO to complete */ | |
714 | wait_event_lock_irq(conf->wait_barrier, | |
715 | !conf->nr_pending && conf->barrier < RESYNC_DEPTH, | |
716 | conf->resync_lock, | |
717 | raid10_unplug(conf->mddev->queue)); | |
718 | ||
719 | spin_unlock_irq(&conf->resync_lock); | |
720 | } | |
721 | ||
722 | static void lower_barrier(conf_t *conf) | |
723 | { | |
724 | unsigned long flags; | |
725 | spin_lock_irqsave(&conf->resync_lock, flags); | |
726 | conf->barrier--; | |
727 | spin_unlock_irqrestore(&conf->resync_lock, flags); | |
728 | wake_up(&conf->wait_barrier); | |
729 | } | |
730 | ||
731 | static void wait_barrier(conf_t *conf) | |
732 | { | |
733 | spin_lock_irq(&conf->resync_lock); | |
734 | if (conf->barrier) { | |
735 | conf->nr_waiting++; | |
736 | wait_event_lock_irq(conf->wait_barrier, !conf->barrier, | |
737 | conf->resync_lock, | |
738 | raid10_unplug(conf->mddev->queue)); | |
739 | conf->nr_waiting--; | |
1da177e4 | 740 | } |
0a27ec96 | 741 | conf->nr_pending++; |
1da177e4 LT |
742 | spin_unlock_irq(&conf->resync_lock); |
743 | } | |
744 | ||
0a27ec96 N |
745 | static void allow_barrier(conf_t *conf) |
746 | { | |
747 | unsigned long flags; | |
748 | spin_lock_irqsave(&conf->resync_lock, flags); | |
749 | conf->nr_pending--; | |
750 | spin_unlock_irqrestore(&conf->resync_lock, flags); | |
751 | wake_up(&conf->wait_barrier); | |
752 | } | |
753 | ||
4443ae10 N |
754 | static void freeze_array(conf_t *conf) |
755 | { | |
756 | /* stop syncio and normal IO and wait for everything to | |
f188593e | 757 | * go quiet. |
4443ae10 | 758 | * We increment barrier and nr_waiting, and then |
1c830532 N |
759 | * wait until nr_pending match nr_queued+1 |
760 | * This is called in the context of one normal IO request | |
761 | * that has failed. Thus any sync request that might be pending | |
762 | * will be blocked by nr_pending, and we need to wait for | |
763 | * pending IO requests to complete or be queued for re-try. | |
764 | * Thus the number queued (nr_queued) plus this request (1) | |
765 | * must match the number of pending IOs (nr_pending) before | |
766 | * we continue. | |
4443ae10 N |
767 | */ |
768 | spin_lock_irq(&conf->resync_lock); | |
769 | conf->barrier++; | |
770 | conf->nr_waiting++; | |
771 | wait_event_lock_irq(conf->wait_barrier, | |
1c830532 | 772 | conf->nr_pending == conf->nr_queued+1, |
4443ae10 | 773 | conf->resync_lock, |
a35e63ef N |
774 | ({ flush_pending_writes(conf); |
775 | raid10_unplug(conf->mddev->queue); })); | |
4443ae10 N |
776 | spin_unlock_irq(&conf->resync_lock); |
777 | } | |
778 | ||
779 | static void unfreeze_array(conf_t *conf) | |
780 | { | |
781 | /* reverse the effect of the freeze */ | |
782 | spin_lock_irq(&conf->resync_lock); | |
783 | conf->barrier--; | |
784 | conf->nr_waiting--; | |
785 | wake_up(&conf->wait_barrier); | |
786 | spin_unlock_irq(&conf->resync_lock); | |
787 | } | |
788 | ||
165125e1 | 789 | static int make_request(struct request_queue *q, struct bio * bio) |
1da177e4 LT |
790 | { |
791 | mddev_t *mddev = q->queuedata; | |
792 | conf_t *conf = mddev_to_conf(mddev); | |
793 | mirror_info_t *mirror; | |
794 | r10bio_t *r10_bio; | |
795 | struct bio *read_bio; | |
c9959059 | 796 | int cpu; |
1da177e4 LT |
797 | int i; |
798 | int chunk_sects = conf->chunk_mask + 1; | |
a362357b | 799 | const int rw = bio_data_dir(bio); |
e3881a68 | 800 | const int do_sync = bio_sync(bio); |
6cce3b23 N |
801 | struct bio_list bl; |
802 | unsigned long flags; | |
6bfe0b49 | 803 | mdk_rdev_t *blocked_rdev; |
1da177e4 | 804 | |
e5dcdd80 | 805 | if (unlikely(bio_barrier(bio))) { |
6712ecf8 | 806 | bio_endio(bio, -EOPNOTSUPP); |
e5dcdd80 N |
807 | return 0; |
808 | } | |
809 | ||
1da177e4 LT |
810 | /* If this request crosses a chunk boundary, we need to |
811 | * split it. This will only happen for 1 PAGE (or less) requests. | |
812 | */ | |
813 | if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9) | |
814 | > chunk_sects && | |
815 | conf->near_copies < conf->raid_disks)) { | |
816 | struct bio_pair *bp; | |
817 | /* Sanity check -- queue functions should prevent this happening */ | |
818 | if (bio->bi_vcnt != 1 || | |
819 | bio->bi_idx != 0) | |
820 | goto bad_map; | |
821 | /* This is a one page bio that upper layers | |
822 | * refuse to split for us, so we need to split it. | |
823 | */ | |
6feef531 | 824 | bp = bio_split(bio, |
1da177e4 LT |
825 | chunk_sects - (bio->bi_sector & (chunk_sects - 1)) ); |
826 | if (make_request(q, &bp->bio1)) | |
827 | generic_make_request(&bp->bio1); | |
828 | if (make_request(q, &bp->bio2)) | |
829 | generic_make_request(&bp->bio2); | |
830 | ||
831 | bio_pair_release(bp); | |
832 | return 0; | |
833 | bad_map: | |
834 | printk("raid10_make_request bug: can't convert block across chunks" | |
835 | " or bigger than %dk %llu %d\n", chunk_sects/2, | |
836 | (unsigned long long)bio->bi_sector, bio->bi_size >> 10); | |
837 | ||
6712ecf8 | 838 | bio_io_error(bio); |
1da177e4 LT |
839 | return 0; |
840 | } | |
841 | ||
3d310eb7 | 842 | md_write_start(mddev, bio); |
06d91a5f | 843 | |
1da177e4 LT |
844 | /* |
845 | * Register the new request and wait if the reconstruction | |
846 | * thread has put up a bar for new requests. | |
847 | * Continue immediately if no resync is active currently. | |
848 | */ | |
0a27ec96 | 849 | wait_barrier(conf); |
1da177e4 | 850 | |
074a7aca TH |
851 | cpu = part_stat_lock(); |
852 | part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]); | |
853 | part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw], | |
854 | bio_sectors(bio)); | |
855 | part_stat_unlock(); | |
1da177e4 LT |
856 | |
857 | r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO); | |
858 | ||
859 | r10_bio->master_bio = bio; | |
860 | r10_bio->sectors = bio->bi_size >> 9; | |
861 | ||
862 | r10_bio->mddev = mddev; | |
863 | r10_bio->sector = bio->bi_sector; | |
6cce3b23 | 864 | r10_bio->state = 0; |
1da177e4 | 865 | |
a362357b | 866 | if (rw == READ) { |
1da177e4 LT |
867 | /* |
868 | * read balancing logic: | |
869 | */ | |
870 | int disk = read_balance(conf, r10_bio); | |
871 | int slot = r10_bio->read_slot; | |
872 | if (disk < 0) { | |
873 | raid_end_bio_io(r10_bio); | |
874 | return 0; | |
875 | } | |
876 | mirror = conf->mirrors + disk; | |
877 | ||
878 | read_bio = bio_clone(bio, GFP_NOIO); | |
879 | ||
880 | r10_bio->devs[slot].bio = read_bio; | |
881 | ||
882 | read_bio->bi_sector = r10_bio->devs[slot].addr + | |
883 | mirror->rdev->data_offset; | |
884 | read_bio->bi_bdev = mirror->rdev->bdev; | |
885 | read_bio->bi_end_io = raid10_end_read_request; | |
e3881a68 | 886 | read_bio->bi_rw = READ | do_sync; |
1da177e4 LT |
887 | read_bio->bi_private = r10_bio; |
888 | ||
889 | generic_make_request(read_bio); | |
890 | return 0; | |
891 | } | |
892 | ||
893 | /* | |
894 | * WRITE: | |
895 | */ | |
6bfe0b49 | 896 | /* first select target devices under rcu_lock and |
1da177e4 LT |
897 | * inc refcount on their rdev. Record them by setting |
898 | * bios[x] to bio | |
899 | */ | |
900 | raid10_find_phys(conf, r10_bio); | |
6bfe0b49 | 901 | retry_write: |
cb6969e8 | 902 | blocked_rdev = NULL; |
1da177e4 LT |
903 | rcu_read_lock(); |
904 | for (i = 0; i < conf->copies; i++) { | |
905 | int d = r10_bio->devs[i].devnum; | |
d6065f7b | 906 | mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev); |
6bfe0b49 DW |
907 | if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) { |
908 | atomic_inc(&rdev->nr_pending); | |
909 | blocked_rdev = rdev; | |
910 | break; | |
911 | } | |
912 | if (rdev && !test_bit(Faulty, &rdev->flags)) { | |
d6065f7b | 913 | atomic_inc(&rdev->nr_pending); |
1da177e4 | 914 | r10_bio->devs[i].bio = bio; |
6cce3b23 | 915 | } else { |
1da177e4 | 916 | r10_bio->devs[i].bio = NULL; |
6cce3b23 N |
917 | set_bit(R10BIO_Degraded, &r10_bio->state); |
918 | } | |
1da177e4 LT |
919 | } |
920 | rcu_read_unlock(); | |
921 | ||
6bfe0b49 DW |
922 | if (unlikely(blocked_rdev)) { |
923 | /* Have to wait for this device to get unblocked, then retry */ | |
924 | int j; | |
925 | int d; | |
926 | ||
927 | for (j = 0; j < i; j++) | |
928 | if (r10_bio->devs[j].bio) { | |
929 | d = r10_bio->devs[j].devnum; | |
930 | rdev_dec_pending(conf->mirrors[d].rdev, mddev); | |
931 | } | |
932 | allow_barrier(conf); | |
933 | md_wait_for_blocked_rdev(blocked_rdev, mddev); | |
934 | wait_barrier(conf); | |
935 | goto retry_write; | |
936 | } | |
937 | ||
6cce3b23 | 938 | atomic_set(&r10_bio->remaining, 0); |
06d91a5f | 939 | |
6cce3b23 | 940 | bio_list_init(&bl); |
1da177e4 LT |
941 | for (i = 0; i < conf->copies; i++) { |
942 | struct bio *mbio; | |
943 | int d = r10_bio->devs[i].devnum; | |
944 | if (!r10_bio->devs[i].bio) | |
945 | continue; | |
946 | ||
947 | mbio = bio_clone(bio, GFP_NOIO); | |
948 | r10_bio->devs[i].bio = mbio; | |
949 | ||
950 | mbio->bi_sector = r10_bio->devs[i].addr+ | |
951 | conf->mirrors[d].rdev->data_offset; | |
952 | mbio->bi_bdev = conf->mirrors[d].rdev->bdev; | |
953 | mbio->bi_end_io = raid10_end_write_request; | |
e3881a68 | 954 | mbio->bi_rw = WRITE | do_sync; |
1da177e4 LT |
955 | mbio->bi_private = r10_bio; |
956 | ||
957 | atomic_inc(&r10_bio->remaining); | |
6cce3b23 | 958 | bio_list_add(&bl, mbio); |
1da177e4 LT |
959 | } |
960 | ||
f6f953aa AR |
961 | if (unlikely(!atomic_read(&r10_bio->remaining))) { |
962 | /* the array is dead */ | |
963 | md_write_end(mddev); | |
964 | raid_end_bio_io(r10_bio); | |
965 | return 0; | |
966 | } | |
967 | ||
6cce3b23 N |
968 | bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0); |
969 | spin_lock_irqsave(&conf->device_lock, flags); | |
970 | bio_list_merge(&conf->pending_bio_list, &bl); | |
971 | blk_plug_device(mddev->queue); | |
972 | spin_unlock_irqrestore(&conf->device_lock, flags); | |
1da177e4 | 973 | |
a35e63ef N |
974 | /* In case raid10d snuck in to freeze_array */ |
975 | wake_up(&conf->wait_barrier); | |
976 | ||
e3881a68 LE |
977 | if (do_sync) |
978 | md_wakeup_thread(mddev->thread); | |
979 | ||
1da177e4 LT |
980 | return 0; |
981 | } | |
982 | ||
983 | static void status(struct seq_file *seq, mddev_t *mddev) | |
984 | { | |
985 | conf_t *conf = mddev_to_conf(mddev); | |
986 | int i; | |
987 | ||
988 | if (conf->near_copies < conf->raid_disks) | |
989 | seq_printf(seq, " %dK chunks", mddev->chunk_size/1024); | |
990 | if (conf->near_copies > 1) | |
991 | seq_printf(seq, " %d near-copies", conf->near_copies); | |
c93983bf N |
992 | if (conf->far_copies > 1) { |
993 | if (conf->far_offset) | |
994 | seq_printf(seq, " %d offset-copies", conf->far_copies); | |
995 | else | |
996 | seq_printf(seq, " %d far-copies", conf->far_copies); | |
997 | } | |
1da177e4 | 998 | seq_printf(seq, " [%d/%d] [", conf->raid_disks, |
76186dd8 | 999 | conf->raid_disks - mddev->degraded); |
1da177e4 LT |
1000 | for (i = 0; i < conf->raid_disks; i++) |
1001 | seq_printf(seq, "%s", | |
1002 | conf->mirrors[i].rdev && | |
b2d444d7 | 1003 | test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_"); |
1da177e4 LT |
1004 | seq_printf(seq, "]"); |
1005 | } | |
1006 | ||
1007 | static void error(mddev_t *mddev, mdk_rdev_t *rdev) | |
1008 | { | |
1009 | char b[BDEVNAME_SIZE]; | |
1010 | conf_t *conf = mddev_to_conf(mddev); | |
1011 | ||
1012 | /* | |
1013 | * If it is not operational, then we have already marked it as dead | |
1014 | * else if it is the last working disks, ignore the error, let the | |
1015 | * next level up know. | |
1016 | * else mark the drive as failed | |
1017 | */ | |
b2d444d7 | 1018 | if (test_bit(In_sync, &rdev->flags) |
76186dd8 | 1019 | && conf->raid_disks-mddev->degraded == 1) |
1da177e4 LT |
1020 | /* |
1021 | * Don't fail the drive, just return an IO error. | |
1022 | * The test should really be more sophisticated than | |
1023 | * "working_disks == 1", but it isn't critical, and | |
1024 | * can wait until we do more sophisticated "is the drive | |
1025 | * really dead" tests... | |
1026 | */ | |
1027 | return; | |
c04be0aa N |
1028 | if (test_and_clear_bit(In_sync, &rdev->flags)) { |
1029 | unsigned long flags; | |
1030 | spin_lock_irqsave(&conf->device_lock, flags); | |
1da177e4 | 1031 | mddev->degraded++; |
c04be0aa | 1032 | spin_unlock_irqrestore(&conf->device_lock, flags); |
1da177e4 LT |
1033 | /* |
1034 | * if recovery is running, make sure it aborts. | |
1035 | */ | |
dfc70645 | 1036 | set_bit(MD_RECOVERY_INTR, &mddev->recovery); |
1da177e4 | 1037 | } |
b2d444d7 | 1038 | set_bit(Faulty, &rdev->flags); |
850b2b42 | 1039 | set_bit(MD_CHANGE_DEVS, &mddev->flags); |
d7a420c9 NA |
1040 | printk(KERN_ALERT "raid10: Disk failure on %s, disabling device.\n" |
1041 | "raid10: Operation continuing on %d devices.\n", | |
76186dd8 | 1042 | bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded); |
1da177e4 LT |
1043 | } |
1044 | ||
1045 | static void print_conf(conf_t *conf) | |
1046 | { | |
1047 | int i; | |
1048 | mirror_info_t *tmp; | |
1049 | ||
1050 | printk("RAID10 conf printout:\n"); | |
1051 | if (!conf) { | |
1052 | printk("(!conf)\n"); | |
1053 | return; | |
1054 | } | |
76186dd8 | 1055 | printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded, |
1da177e4 LT |
1056 | conf->raid_disks); |
1057 | ||
1058 | for (i = 0; i < conf->raid_disks; i++) { | |
1059 | char b[BDEVNAME_SIZE]; | |
1060 | tmp = conf->mirrors + i; | |
1061 | if (tmp->rdev) | |
1062 | printk(" disk %d, wo:%d, o:%d, dev:%s\n", | |
b2d444d7 N |
1063 | i, !test_bit(In_sync, &tmp->rdev->flags), |
1064 | !test_bit(Faulty, &tmp->rdev->flags), | |
1da177e4 LT |
1065 | bdevname(tmp->rdev->bdev,b)); |
1066 | } | |
1067 | } | |
1068 | ||
1069 | static void close_sync(conf_t *conf) | |
1070 | { | |
0a27ec96 N |
1071 | wait_barrier(conf); |
1072 | allow_barrier(conf); | |
1da177e4 LT |
1073 | |
1074 | mempool_destroy(conf->r10buf_pool); | |
1075 | conf->r10buf_pool = NULL; | |
1076 | } | |
1077 | ||
6d508242 N |
1078 | /* check if there are enough drives for |
1079 | * every block to appear on atleast one | |
1080 | */ | |
1081 | static int enough(conf_t *conf) | |
1082 | { | |
1083 | int first = 0; | |
1084 | ||
1085 | do { | |
1086 | int n = conf->copies; | |
1087 | int cnt = 0; | |
1088 | while (n--) { | |
1089 | if (conf->mirrors[first].rdev) | |
1090 | cnt++; | |
1091 | first = (first+1) % conf->raid_disks; | |
1092 | } | |
1093 | if (cnt == 0) | |
1094 | return 0; | |
1095 | } while (first != 0); | |
1096 | return 1; | |
1097 | } | |
1098 | ||
1da177e4 LT |
1099 | static int raid10_spare_active(mddev_t *mddev) |
1100 | { | |
1101 | int i; | |
1102 | conf_t *conf = mddev->private; | |
1103 | mirror_info_t *tmp; | |
1104 | ||
1105 | /* | |
1106 | * Find all non-in_sync disks within the RAID10 configuration | |
1107 | * and mark them in_sync | |
1108 | */ | |
1109 | for (i = 0; i < conf->raid_disks; i++) { | |
1110 | tmp = conf->mirrors + i; | |
1111 | if (tmp->rdev | |
b2d444d7 | 1112 | && !test_bit(Faulty, &tmp->rdev->flags) |
c04be0aa N |
1113 | && !test_and_set_bit(In_sync, &tmp->rdev->flags)) { |
1114 | unsigned long flags; | |
1115 | spin_lock_irqsave(&conf->device_lock, flags); | |
1da177e4 | 1116 | mddev->degraded--; |
c04be0aa | 1117 | spin_unlock_irqrestore(&conf->device_lock, flags); |
1da177e4 LT |
1118 | } |
1119 | } | |
1120 | ||
1121 | print_conf(conf); | |
1122 | return 0; | |
1123 | } | |
1124 | ||
1125 | ||
1126 | static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev) | |
1127 | { | |
1128 | conf_t *conf = mddev->private; | |
199050ea | 1129 | int err = -EEXIST; |
1da177e4 LT |
1130 | int mirror; |
1131 | mirror_info_t *p; | |
6c2fce2e NB |
1132 | int first = 0; |
1133 | int last = mddev->raid_disks - 1; | |
1da177e4 LT |
1134 | |
1135 | if (mddev->recovery_cp < MaxSector) | |
1136 | /* only hot-add to in-sync arrays, as recovery is | |
1137 | * very different from resync | |
1138 | */ | |
199050ea | 1139 | return -EBUSY; |
6d508242 | 1140 | if (!enough(conf)) |
199050ea | 1141 | return -EINVAL; |
1da177e4 | 1142 | |
a53a6c85 | 1143 | if (rdev->raid_disk >= 0) |
6c2fce2e | 1144 | first = last = rdev->raid_disk; |
1da177e4 | 1145 | |
6cce3b23 | 1146 | if (rdev->saved_raid_disk >= 0 && |
6c2fce2e | 1147 | rdev->saved_raid_disk >= first && |
6cce3b23 N |
1148 | conf->mirrors[rdev->saved_raid_disk].rdev == NULL) |
1149 | mirror = rdev->saved_raid_disk; | |
1150 | else | |
6c2fce2e NB |
1151 | mirror = first; |
1152 | for ( ; mirror <= last ; mirror++) | |
1da177e4 LT |
1153 | if ( !(p=conf->mirrors+mirror)->rdev) { |
1154 | ||
1155 | blk_queue_stack_limits(mddev->queue, | |
1156 | rdev->bdev->bd_disk->queue); | |
1157 | /* as we don't honour merge_bvec_fn, we must never risk | |
1158 | * violating it, so limit ->max_sector to one PAGE, as | |
1159 | * a one page request is never in violation. | |
1160 | */ | |
1161 | if (rdev->bdev->bd_disk->queue->merge_bvec_fn && | |
1162 | mddev->queue->max_sectors > (PAGE_SIZE>>9)) | |
1163 | mddev->queue->max_sectors = (PAGE_SIZE>>9); | |
1164 | ||
1165 | p->head_position = 0; | |
1166 | rdev->raid_disk = mirror; | |
199050ea | 1167 | err = 0; |
6cce3b23 N |
1168 | if (rdev->saved_raid_disk != mirror) |
1169 | conf->fullsync = 1; | |
d6065f7b | 1170 | rcu_assign_pointer(p->rdev, rdev); |
1da177e4 LT |
1171 | break; |
1172 | } | |
1173 | ||
1174 | print_conf(conf); | |
199050ea | 1175 | return err; |
1da177e4 LT |
1176 | } |
1177 | ||
1178 | static int raid10_remove_disk(mddev_t *mddev, int number) | |
1179 | { | |
1180 | conf_t *conf = mddev->private; | |
1181 | int err = 0; | |
1182 | mdk_rdev_t *rdev; | |
1183 | mirror_info_t *p = conf->mirrors+ number; | |
1184 | ||
1185 | print_conf(conf); | |
1186 | rdev = p->rdev; | |
1187 | if (rdev) { | |
b2d444d7 | 1188 | if (test_bit(In_sync, &rdev->flags) || |
1da177e4 LT |
1189 | atomic_read(&rdev->nr_pending)) { |
1190 | err = -EBUSY; | |
1191 | goto abort; | |
1192 | } | |
dfc70645 N |
1193 | /* Only remove faulty devices in recovery |
1194 | * is not possible. | |
1195 | */ | |
1196 | if (!test_bit(Faulty, &rdev->flags) && | |
1197 | enough(conf)) { | |
1198 | err = -EBUSY; | |
1199 | goto abort; | |
1200 | } | |
1da177e4 | 1201 | p->rdev = NULL; |
fbd568a3 | 1202 | synchronize_rcu(); |
1da177e4 LT |
1203 | if (atomic_read(&rdev->nr_pending)) { |
1204 | /* lost the race, try later */ | |
1205 | err = -EBUSY; | |
1206 | p->rdev = rdev; | |
1207 | } | |
1208 | } | |
1209 | abort: | |
1210 | ||
1211 | print_conf(conf); | |
1212 | return err; | |
1213 | } | |
1214 | ||
1215 | ||
6712ecf8 | 1216 | static void end_sync_read(struct bio *bio, int error) |
1da177e4 | 1217 | { |
1da177e4 LT |
1218 | r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private); |
1219 | conf_t *conf = mddev_to_conf(r10_bio->mddev); | |
1220 | int i,d; | |
1221 | ||
1da177e4 LT |
1222 | for (i=0; i<conf->copies; i++) |
1223 | if (r10_bio->devs[i].bio == bio) | |
1224 | break; | |
b6385483 | 1225 | BUG_ON(i == conf->copies); |
1da177e4 LT |
1226 | update_head_pos(i, r10_bio); |
1227 | d = r10_bio->devs[i].devnum; | |
0eb3ff12 N |
1228 | |
1229 | if (test_bit(BIO_UPTODATE, &bio->bi_flags)) | |
1230 | set_bit(R10BIO_Uptodate, &r10_bio->state); | |
4dbcdc75 N |
1231 | else { |
1232 | atomic_add(r10_bio->sectors, | |
1233 | &conf->mirrors[d].rdev->corrected_errors); | |
1234 | if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery)) | |
1235 | md_error(r10_bio->mddev, | |
1236 | conf->mirrors[d].rdev); | |
1237 | } | |
1da177e4 LT |
1238 | |
1239 | /* for reconstruct, we always reschedule after a read. | |
1240 | * for resync, only after all reads | |
1241 | */ | |
73d5c38a | 1242 | rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev); |
1da177e4 LT |
1243 | if (test_bit(R10BIO_IsRecover, &r10_bio->state) || |
1244 | atomic_dec_and_test(&r10_bio->remaining)) { | |
1245 | /* we have read all the blocks, | |
1246 | * do the comparison in process context in raid10d | |
1247 | */ | |
1248 | reschedule_retry(r10_bio); | |
1249 | } | |
1da177e4 LT |
1250 | } |
1251 | ||
6712ecf8 | 1252 | static void end_sync_write(struct bio *bio, int error) |
1da177e4 LT |
1253 | { |
1254 | int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags); | |
1255 | r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private); | |
1256 | mddev_t *mddev = r10_bio->mddev; | |
1257 | conf_t *conf = mddev_to_conf(mddev); | |
1258 | int i,d; | |
1259 | ||
1da177e4 LT |
1260 | for (i = 0; i < conf->copies; i++) |
1261 | if (r10_bio->devs[i].bio == bio) | |
1262 | break; | |
1263 | d = r10_bio->devs[i].devnum; | |
1264 | ||
1265 | if (!uptodate) | |
1266 | md_error(mddev, conf->mirrors[d].rdev); | |
dfc70645 | 1267 | |
1da177e4 LT |
1268 | update_head_pos(i, r10_bio); |
1269 | ||
73d5c38a | 1270 | rdev_dec_pending(conf->mirrors[d].rdev, mddev); |
1da177e4 LT |
1271 | while (atomic_dec_and_test(&r10_bio->remaining)) { |
1272 | if (r10_bio->master_bio == NULL) { | |
1273 | /* the primary of several recovery bios */ | |
73d5c38a | 1274 | sector_t s = r10_bio->sectors; |
1da177e4 | 1275 | put_buf(r10_bio); |
73d5c38a | 1276 | md_done_sync(mddev, s, 1); |
1da177e4 LT |
1277 | break; |
1278 | } else { | |
1279 | r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio; | |
1280 | put_buf(r10_bio); | |
1281 | r10_bio = r10_bio2; | |
1282 | } | |
1283 | } | |
1da177e4 LT |
1284 | } |
1285 | ||
1286 | /* | |
1287 | * Note: sync and recover and handled very differently for raid10 | |
1288 | * This code is for resync. | |
1289 | * For resync, we read through virtual addresses and read all blocks. | |
1290 | * If there is any error, we schedule a write. The lowest numbered | |
1291 | * drive is authoritative. | |
1292 | * However requests come for physical address, so we need to map. | |
1293 | * For every physical address there are raid_disks/copies virtual addresses, | |
1294 | * which is always are least one, but is not necessarly an integer. | |
1295 | * This means that a physical address can span multiple chunks, so we may | |
1296 | * have to submit multiple io requests for a single sync request. | |
1297 | */ | |
1298 | /* | |
1299 | * We check if all blocks are in-sync and only write to blocks that | |
1300 | * aren't in sync | |
1301 | */ | |
1302 | static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio) | |
1303 | { | |
1304 | conf_t *conf = mddev_to_conf(mddev); | |
1305 | int i, first; | |
1306 | struct bio *tbio, *fbio; | |
1307 | ||
1308 | atomic_set(&r10_bio->remaining, 1); | |
1309 | ||
1310 | /* find the first device with a block */ | |
1311 | for (i=0; i<conf->copies; i++) | |
1312 | if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) | |
1313 | break; | |
1314 | ||
1315 | if (i == conf->copies) | |
1316 | goto done; | |
1317 | ||
1318 | first = i; | |
1319 | fbio = r10_bio->devs[i].bio; | |
1320 | ||
1321 | /* now find blocks with errors */ | |
0eb3ff12 N |
1322 | for (i=0 ; i < conf->copies ; i++) { |
1323 | int j, d; | |
1324 | int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9); | |
1da177e4 | 1325 | |
1da177e4 | 1326 | tbio = r10_bio->devs[i].bio; |
0eb3ff12 N |
1327 | |
1328 | if (tbio->bi_end_io != end_sync_read) | |
1329 | continue; | |
1330 | if (i == first) | |
1da177e4 | 1331 | continue; |
0eb3ff12 N |
1332 | if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) { |
1333 | /* We know that the bi_io_vec layout is the same for | |
1334 | * both 'first' and 'i', so we just compare them. | |
1335 | * All vec entries are PAGE_SIZE; | |
1336 | */ | |
1337 | for (j = 0; j < vcnt; j++) | |
1338 | if (memcmp(page_address(fbio->bi_io_vec[j].bv_page), | |
1339 | page_address(tbio->bi_io_vec[j].bv_page), | |
1340 | PAGE_SIZE)) | |
1341 | break; | |
1342 | if (j == vcnt) | |
1343 | continue; | |
1344 | mddev->resync_mismatches += r10_bio->sectors; | |
1345 | } | |
18f08819 N |
1346 | if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) |
1347 | /* Don't fix anything. */ | |
1348 | continue; | |
1da177e4 LT |
1349 | /* Ok, we need to write this bio |
1350 | * First we need to fixup bv_offset, bv_len and | |
1351 | * bi_vecs, as the read request might have corrupted these | |
1352 | */ | |
1353 | tbio->bi_vcnt = vcnt; | |
1354 | tbio->bi_size = r10_bio->sectors << 9; | |
1355 | tbio->bi_idx = 0; | |
1356 | tbio->bi_phys_segments = 0; | |
1da177e4 LT |
1357 | tbio->bi_flags &= ~(BIO_POOL_MASK - 1); |
1358 | tbio->bi_flags |= 1 << BIO_UPTODATE; | |
1359 | tbio->bi_next = NULL; | |
1360 | tbio->bi_rw = WRITE; | |
1361 | tbio->bi_private = r10_bio; | |
1362 | tbio->bi_sector = r10_bio->devs[i].addr; | |
1363 | ||
1364 | for (j=0; j < vcnt ; j++) { | |
1365 | tbio->bi_io_vec[j].bv_offset = 0; | |
1366 | tbio->bi_io_vec[j].bv_len = PAGE_SIZE; | |
1367 | ||
1368 | memcpy(page_address(tbio->bi_io_vec[j].bv_page), | |
1369 | page_address(fbio->bi_io_vec[j].bv_page), | |
1370 | PAGE_SIZE); | |
1371 | } | |
1372 | tbio->bi_end_io = end_sync_write; | |
1373 | ||
1374 | d = r10_bio->devs[i].devnum; | |
1375 | atomic_inc(&conf->mirrors[d].rdev->nr_pending); | |
1376 | atomic_inc(&r10_bio->remaining); | |
1377 | md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9); | |
1378 | ||
1379 | tbio->bi_sector += conf->mirrors[d].rdev->data_offset; | |
1380 | tbio->bi_bdev = conf->mirrors[d].rdev->bdev; | |
1381 | generic_make_request(tbio); | |
1382 | } | |
1383 | ||
1384 | done: | |
1385 | if (atomic_dec_and_test(&r10_bio->remaining)) { | |
1386 | md_done_sync(mddev, r10_bio->sectors, 1); | |
1387 | put_buf(r10_bio); | |
1388 | } | |
1389 | } | |
1390 | ||
1391 | /* | |
1392 | * Now for the recovery code. | |
1393 | * Recovery happens across physical sectors. | |
1394 | * We recover all non-is_sync drives by finding the virtual address of | |
1395 | * each, and then choose a working drive that also has that virt address. | |
1396 | * There is a separate r10_bio for each non-in_sync drive. | |
1397 | * Only the first two slots are in use. The first for reading, | |
1398 | * The second for writing. | |
1399 | * | |
1400 | */ | |
1401 | ||
1402 | static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio) | |
1403 | { | |
1404 | conf_t *conf = mddev_to_conf(mddev); | |
1405 | int i, d; | |
1406 | struct bio *bio, *wbio; | |
1407 | ||
1408 | ||
1409 | /* move the pages across to the second bio | |
1410 | * and submit the write request | |
1411 | */ | |
1412 | bio = r10_bio->devs[0].bio; | |
1413 | wbio = r10_bio->devs[1].bio; | |
1414 | for (i=0; i < wbio->bi_vcnt; i++) { | |
1415 | struct page *p = bio->bi_io_vec[i].bv_page; | |
1416 | bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page; | |
1417 | wbio->bi_io_vec[i].bv_page = p; | |
1418 | } | |
1419 | d = r10_bio->devs[1].devnum; | |
1420 | ||
1421 | atomic_inc(&conf->mirrors[d].rdev->nr_pending); | |
1422 | md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9); | |
0eb3ff12 N |
1423 | if (test_bit(R10BIO_Uptodate, &r10_bio->state)) |
1424 | generic_make_request(wbio); | |
1425 | else | |
6712ecf8 | 1426 | bio_endio(wbio, -EIO); |
1da177e4 LT |
1427 | } |
1428 | ||
1429 | ||
1430 | /* | |
1431 | * This is a kernel thread which: | |
1432 | * | |
1433 | * 1. Retries failed read operations on working mirrors. | |
1434 | * 2. Updates the raid superblock when problems encounter. | |
6814d536 | 1435 | * 3. Performs writes following reads for array synchronising. |
1da177e4 LT |
1436 | */ |
1437 | ||
6814d536 N |
1438 | static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio) |
1439 | { | |
1440 | int sect = 0; /* Offset from r10_bio->sector */ | |
1441 | int sectors = r10_bio->sectors; | |
1442 | mdk_rdev_t*rdev; | |
1443 | while(sectors) { | |
1444 | int s = sectors; | |
1445 | int sl = r10_bio->read_slot; | |
1446 | int success = 0; | |
1447 | int start; | |
1448 | ||
1449 | if (s > (PAGE_SIZE>>9)) | |
1450 | s = PAGE_SIZE >> 9; | |
1451 | ||
1452 | rcu_read_lock(); | |
1453 | do { | |
1454 | int d = r10_bio->devs[sl].devnum; | |
1455 | rdev = rcu_dereference(conf->mirrors[d].rdev); | |
1456 | if (rdev && | |
1457 | test_bit(In_sync, &rdev->flags)) { | |
1458 | atomic_inc(&rdev->nr_pending); | |
1459 | rcu_read_unlock(); | |
1460 | success = sync_page_io(rdev->bdev, | |
1461 | r10_bio->devs[sl].addr + | |
1462 | sect + rdev->data_offset, | |
1463 | s<<9, | |
1464 | conf->tmppage, READ); | |
1465 | rdev_dec_pending(rdev, mddev); | |
1466 | rcu_read_lock(); | |
1467 | if (success) | |
1468 | break; | |
1469 | } | |
1470 | sl++; | |
1471 | if (sl == conf->copies) | |
1472 | sl = 0; | |
1473 | } while (!success && sl != r10_bio->read_slot); | |
1474 | rcu_read_unlock(); | |
1475 | ||
1476 | if (!success) { | |
1477 | /* Cannot read from anywhere -- bye bye array */ | |
1478 | int dn = r10_bio->devs[r10_bio->read_slot].devnum; | |
1479 | md_error(mddev, conf->mirrors[dn].rdev); | |
1480 | break; | |
1481 | } | |
1482 | ||
1483 | start = sl; | |
1484 | /* write it back and re-read */ | |
1485 | rcu_read_lock(); | |
1486 | while (sl != r10_bio->read_slot) { | |
1487 | int d; | |
1488 | if (sl==0) | |
1489 | sl = conf->copies; | |
1490 | sl--; | |
1491 | d = r10_bio->devs[sl].devnum; | |
1492 | rdev = rcu_dereference(conf->mirrors[d].rdev); | |
1493 | if (rdev && | |
1494 | test_bit(In_sync, &rdev->flags)) { | |
1495 | atomic_inc(&rdev->nr_pending); | |
1496 | rcu_read_unlock(); | |
1497 | atomic_add(s, &rdev->corrected_errors); | |
1498 | if (sync_page_io(rdev->bdev, | |
1499 | r10_bio->devs[sl].addr + | |
1500 | sect + rdev->data_offset, | |
1501 | s<<9, conf->tmppage, WRITE) | |
1502 | == 0) | |
1503 | /* Well, this device is dead */ | |
1504 | md_error(mddev, rdev); | |
1505 | rdev_dec_pending(rdev, mddev); | |
1506 | rcu_read_lock(); | |
1507 | } | |
1508 | } | |
1509 | sl = start; | |
1510 | while (sl != r10_bio->read_slot) { | |
1511 | int d; | |
1512 | if (sl==0) | |
1513 | sl = conf->copies; | |
1514 | sl--; | |
1515 | d = r10_bio->devs[sl].devnum; | |
1516 | rdev = rcu_dereference(conf->mirrors[d].rdev); | |
1517 | if (rdev && | |
1518 | test_bit(In_sync, &rdev->flags)) { | |
1519 | char b[BDEVNAME_SIZE]; | |
1520 | atomic_inc(&rdev->nr_pending); | |
1521 | rcu_read_unlock(); | |
1522 | if (sync_page_io(rdev->bdev, | |
1523 | r10_bio->devs[sl].addr + | |
1524 | sect + rdev->data_offset, | |
1525 | s<<9, conf->tmppage, READ) == 0) | |
1526 | /* Well, this device is dead */ | |
1527 | md_error(mddev, rdev); | |
1528 | else | |
1529 | printk(KERN_INFO | |
1530 | "raid10:%s: read error corrected" | |
1531 | " (%d sectors at %llu on %s)\n", | |
1532 | mdname(mddev), s, | |
969b755a RD |
1533 | (unsigned long long)(sect+ |
1534 | rdev->data_offset), | |
6814d536 N |
1535 | bdevname(rdev->bdev, b)); |
1536 | ||
1537 | rdev_dec_pending(rdev, mddev); | |
1538 | rcu_read_lock(); | |
1539 | } | |
1540 | } | |
1541 | rcu_read_unlock(); | |
1542 | ||
1543 | sectors -= s; | |
1544 | sect += s; | |
1545 | } | |
1546 | } | |
1547 | ||
1da177e4 LT |
1548 | static void raid10d(mddev_t *mddev) |
1549 | { | |
1550 | r10bio_t *r10_bio; | |
1551 | struct bio *bio; | |
1552 | unsigned long flags; | |
1553 | conf_t *conf = mddev_to_conf(mddev); | |
1554 | struct list_head *head = &conf->retry_list; | |
1555 | int unplug=0; | |
1556 | mdk_rdev_t *rdev; | |
1557 | ||
1558 | md_check_recovery(mddev); | |
1da177e4 LT |
1559 | |
1560 | for (;;) { | |
1561 | char b[BDEVNAME_SIZE]; | |
6cce3b23 | 1562 | |
a35e63ef | 1563 | unplug += flush_pending_writes(conf); |
6cce3b23 | 1564 | |
a35e63ef N |
1565 | spin_lock_irqsave(&conf->device_lock, flags); |
1566 | if (list_empty(head)) { | |
1567 | spin_unlock_irqrestore(&conf->device_lock, flags); | |
1da177e4 | 1568 | break; |
a35e63ef | 1569 | } |
1da177e4 LT |
1570 | r10_bio = list_entry(head->prev, r10bio_t, retry_list); |
1571 | list_del(head->prev); | |
4443ae10 | 1572 | conf->nr_queued--; |
1da177e4 LT |
1573 | spin_unlock_irqrestore(&conf->device_lock, flags); |
1574 | ||
1575 | mddev = r10_bio->mddev; | |
1576 | conf = mddev_to_conf(mddev); | |
1577 | if (test_bit(R10BIO_IsSync, &r10_bio->state)) { | |
1578 | sync_request_write(mddev, r10_bio); | |
1579 | unplug = 1; | |
1580 | } else if (test_bit(R10BIO_IsRecover, &r10_bio->state)) { | |
1581 | recovery_request_write(mddev, r10_bio); | |
1582 | unplug = 1; | |
1583 | } else { | |
1584 | int mirror; | |
4443ae10 N |
1585 | /* we got a read error. Maybe the drive is bad. Maybe just |
1586 | * the block and we can fix it. | |
1587 | * We freeze all other IO, and try reading the block from | |
1588 | * other devices. When we find one, we re-write | |
1589 | * and check it that fixes the read error. | |
1590 | * This is all done synchronously while the array is | |
1591 | * frozen. | |
1592 | */ | |
6814d536 N |
1593 | if (mddev->ro == 0) { |
1594 | freeze_array(conf); | |
1595 | fix_read_error(conf, mddev, r10_bio); | |
1596 | unfreeze_array(conf); | |
4443ae10 N |
1597 | } |
1598 | ||
1da177e4 | 1599 | bio = r10_bio->devs[r10_bio->read_slot].bio; |
0eb3ff12 N |
1600 | r10_bio->devs[r10_bio->read_slot].bio = |
1601 | mddev->ro ? IO_BLOCKED : NULL; | |
1da177e4 LT |
1602 | mirror = read_balance(conf, r10_bio); |
1603 | if (mirror == -1) { | |
1604 | printk(KERN_ALERT "raid10: %s: unrecoverable I/O" | |
1605 | " read error for block %llu\n", | |
1606 | bdevname(bio->bi_bdev,b), | |
1607 | (unsigned long long)r10_bio->sector); | |
1608 | raid_end_bio_io(r10_bio); | |
14e71344 | 1609 | bio_put(bio); |
1da177e4 | 1610 | } else { |
e3881a68 | 1611 | const int do_sync = bio_sync(r10_bio->master_bio); |
14e71344 | 1612 | bio_put(bio); |
1da177e4 LT |
1613 | rdev = conf->mirrors[mirror].rdev; |
1614 | if (printk_ratelimit()) | |
1615 | printk(KERN_ERR "raid10: %s: redirecting sector %llu to" | |
1616 | " another mirror\n", | |
1617 | bdevname(rdev->bdev,b), | |
1618 | (unsigned long long)r10_bio->sector); | |
1619 | bio = bio_clone(r10_bio->master_bio, GFP_NOIO); | |
1620 | r10_bio->devs[r10_bio->read_slot].bio = bio; | |
1621 | bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr | |
1622 | + rdev->data_offset; | |
1623 | bio->bi_bdev = rdev->bdev; | |
e3881a68 | 1624 | bio->bi_rw = READ | do_sync; |
1da177e4 LT |
1625 | bio->bi_private = r10_bio; |
1626 | bio->bi_end_io = raid10_end_read_request; | |
1627 | unplug = 1; | |
1628 | generic_make_request(bio); | |
1629 | } | |
1630 | } | |
1631 | } | |
1da177e4 LT |
1632 | if (unplug) |
1633 | unplug_slaves(mddev); | |
1634 | } | |
1635 | ||
1636 | ||
1637 | static int init_resync(conf_t *conf) | |
1638 | { | |
1639 | int buffs; | |
1640 | ||
1641 | buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE; | |
b6385483 | 1642 | BUG_ON(conf->r10buf_pool); |
1da177e4 LT |
1643 | conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf); |
1644 | if (!conf->r10buf_pool) | |
1645 | return -ENOMEM; | |
1646 | conf->next_resync = 0; | |
1647 | return 0; | |
1648 | } | |
1649 | ||
1650 | /* | |
1651 | * perform a "sync" on one "block" | |
1652 | * | |
1653 | * We need to make sure that no normal I/O request - particularly write | |
1654 | * requests - conflict with active sync requests. | |
1655 | * | |
1656 | * This is achieved by tracking pending requests and a 'barrier' concept | |
1657 | * that can be installed to exclude normal IO requests. | |
1658 | * | |
1659 | * Resync and recovery are handled very differently. | |
1660 | * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery. | |
1661 | * | |
1662 | * For resync, we iterate over virtual addresses, read all copies, | |
1663 | * and update if there are differences. If only one copy is live, | |
1664 | * skip it. | |
1665 | * For recovery, we iterate over physical addresses, read a good | |
1666 | * value for each non-in_sync drive, and over-write. | |
1667 | * | |
1668 | * So, for recovery we may have several outstanding complex requests for a | |
1669 | * given address, one for each out-of-sync device. We model this by allocating | |
1670 | * a number of r10_bio structures, one for each out-of-sync device. | |
1671 | * As we setup these structures, we collect all bio's together into a list | |
1672 | * which we then process collectively to add pages, and then process again | |
1673 | * to pass to generic_make_request. | |
1674 | * | |
1675 | * The r10_bio structures are linked using a borrowed master_bio pointer. | |
1676 | * This link is counted in ->remaining. When the r10_bio that points to NULL | |
1677 | * has its remaining count decremented to 0, the whole complex operation | |
1678 | * is complete. | |
1679 | * | |
1680 | */ | |
1681 | ||
57afd89f | 1682 | static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster) |
1da177e4 LT |
1683 | { |
1684 | conf_t *conf = mddev_to_conf(mddev); | |
1685 | r10bio_t *r10_bio; | |
1686 | struct bio *biolist = NULL, *bio; | |
1687 | sector_t max_sector, nr_sectors; | |
1688 | int disk; | |
1689 | int i; | |
6cce3b23 N |
1690 | int max_sync; |
1691 | int sync_blocks; | |
1da177e4 LT |
1692 | |
1693 | sector_t sectors_skipped = 0; | |
1694 | int chunks_skipped = 0; | |
1695 | ||
1696 | if (!conf->r10buf_pool) | |
1697 | if (init_resync(conf)) | |
57afd89f | 1698 | return 0; |
1da177e4 LT |
1699 | |
1700 | skipped: | |
58c0fed4 | 1701 | max_sector = mddev->dev_sectors; |
1da177e4 LT |
1702 | if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) |
1703 | max_sector = mddev->resync_max_sectors; | |
1704 | if (sector_nr >= max_sector) { | |
6cce3b23 N |
1705 | /* If we aborted, we need to abort the |
1706 | * sync on the 'current' bitmap chucks (there can | |
1707 | * be several when recovering multiple devices). | |
1708 | * as we may have started syncing it but not finished. | |
1709 | * We can find the current address in | |
1710 | * mddev->curr_resync, but for recovery, | |
1711 | * we need to convert that to several | |
1712 | * virtual addresses. | |
1713 | */ | |
1714 | if (mddev->curr_resync < max_sector) { /* aborted */ | |
1715 | if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) | |
1716 | bitmap_end_sync(mddev->bitmap, mddev->curr_resync, | |
1717 | &sync_blocks, 1); | |
1718 | else for (i=0; i<conf->raid_disks; i++) { | |
1719 | sector_t sect = | |
1720 | raid10_find_virt(conf, mddev->curr_resync, i); | |
1721 | bitmap_end_sync(mddev->bitmap, sect, | |
1722 | &sync_blocks, 1); | |
1723 | } | |
1724 | } else /* completed sync */ | |
1725 | conf->fullsync = 0; | |
1726 | ||
1727 | bitmap_close_sync(mddev->bitmap); | |
1da177e4 | 1728 | close_sync(conf); |
57afd89f | 1729 | *skipped = 1; |
1da177e4 LT |
1730 | return sectors_skipped; |
1731 | } | |
1732 | if (chunks_skipped >= conf->raid_disks) { | |
1733 | /* if there has been nothing to do on any drive, | |
1734 | * then there is nothing to do at all.. | |
1735 | */ | |
57afd89f N |
1736 | *skipped = 1; |
1737 | return (max_sector - sector_nr) + sectors_skipped; | |
1da177e4 LT |
1738 | } |
1739 | ||
c6207277 N |
1740 | if (max_sector > mddev->resync_max) |
1741 | max_sector = mddev->resync_max; /* Don't do IO beyond here */ | |
1742 | ||
1da177e4 LT |
1743 | /* make sure whole request will fit in a chunk - if chunks |
1744 | * are meaningful | |
1745 | */ | |
1746 | if (conf->near_copies < conf->raid_disks && | |
1747 | max_sector > (sector_nr | conf->chunk_mask)) | |
1748 | max_sector = (sector_nr | conf->chunk_mask) + 1; | |
1749 | /* | |
1750 | * If there is non-resync activity waiting for us then | |
1751 | * put in a delay to throttle resync. | |
1752 | */ | |
0a27ec96 | 1753 | if (!go_faster && conf->nr_waiting) |
1da177e4 | 1754 | msleep_interruptible(1000); |
1da177e4 LT |
1755 | |
1756 | /* Again, very different code for resync and recovery. | |
1757 | * Both must result in an r10bio with a list of bios that | |
1758 | * have bi_end_io, bi_sector, bi_bdev set, | |
1759 | * and bi_private set to the r10bio. | |
1760 | * For recovery, we may actually create several r10bios | |
1761 | * with 2 bios in each, that correspond to the bios in the main one. | |
1762 | * In this case, the subordinate r10bios link back through a | |
1763 | * borrowed master_bio pointer, and the counter in the master | |
1764 | * includes a ref from each subordinate. | |
1765 | */ | |
1766 | /* First, we decide what to do and set ->bi_end_io | |
1767 | * To end_sync_read if we want to read, and | |
1768 | * end_sync_write if we will want to write. | |
1769 | */ | |
1770 | ||
6cce3b23 | 1771 | max_sync = RESYNC_PAGES << (PAGE_SHIFT-9); |
1da177e4 LT |
1772 | if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) { |
1773 | /* recovery... the complicated one */ | |
1774 | int i, j, k; | |
1775 | r10_bio = NULL; | |
1776 | ||
1777 | for (i=0 ; i<conf->raid_disks; i++) | |
1778 | if (conf->mirrors[i].rdev && | |
b2d444d7 | 1779 | !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) { |
6cce3b23 | 1780 | int still_degraded = 0; |
1da177e4 LT |
1781 | /* want to reconstruct this device */ |
1782 | r10bio_t *rb2 = r10_bio; | |
6cce3b23 N |
1783 | sector_t sect = raid10_find_virt(conf, sector_nr, i); |
1784 | int must_sync; | |
1785 | /* Unless we are doing a full sync, we only need | |
1786 | * to recover the block if it is set in the bitmap | |
1787 | */ | |
1788 | must_sync = bitmap_start_sync(mddev->bitmap, sect, | |
1789 | &sync_blocks, 1); | |
1790 | if (sync_blocks < max_sync) | |
1791 | max_sync = sync_blocks; | |
1792 | if (!must_sync && | |
1793 | !conf->fullsync) { | |
1794 | /* yep, skip the sync_blocks here, but don't assume | |
1795 | * that there will never be anything to do here | |
1796 | */ | |
1797 | chunks_skipped = -1; | |
1798 | continue; | |
1799 | } | |
1da177e4 LT |
1800 | |
1801 | r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO); | |
6cce3b23 | 1802 | raise_barrier(conf, rb2 != NULL); |
1da177e4 LT |
1803 | atomic_set(&r10_bio->remaining, 0); |
1804 | ||
1805 | r10_bio->master_bio = (struct bio*)rb2; | |
1806 | if (rb2) | |
1807 | atomic_inc(&rb2->remaining); | |
1808 | r10_bio->mddev = mddev; | |
1809 | set_bit(R10BIO_IsRecover, &r10_bio->state); | |
6cce3b23 N |
1810 | r10_bio->sector = sect; |
1811 | ||
1da177e4 | 1812 | raid10_find_phys(conf, r10_bio); |
6cce3b23 N |
1813 | /* Need to check if this section will still be |
1814 | * degraded | |
1815 | */ | |
1816 | for (j=0; j<conf->copies;j++) { | |
1817 | int d = r10_bio->devs[j].devnum; | |
1818 | if (conf->mirrors[d].rdev == NULL || | |
a24a8dd8 | 1819 | test_bit(Faulty, &conf->mirrors[d].rdev->flags)) { |
6cce3b23 | 1820 | still_degraded = 1; |
a24a8dd8 N |
1821 | break; |
1822 | } | |
6cce3b23 N |
1823 | } |
1824 | must_sync = bitmap_start_sync(mddev->bitmap, sect, | |
1825 | &sync_blocks, still_degraded); | |
1826 | ||
1da177e4 LT |
1827 | for (j=0; j<conf->copies;j++) { |
1828 | int d = r10_bio->devs[j].devnum; | |
1829 | if (conf->mirrors[d].rdev && | |
b2d444d7 | 1830 | test_bit(In_sync, &conf->mirrors[d].rdev->flags)) { |
1da177e4 LT |
1831 | /* This is where we read from */ |
1832 | bio = r10_bio->devs[0].bio; | |
1833 | bio->bi_next = biolist; | |
1834 | biolist = bio; | |
1835 | bio->bi_private = r10_bio; | |
1836 | bio->bi_end_io = end_sync_read; | |
802ba064 | 1837 | bio->bi_rw = READ; |
1da177e4 LT |
1838 | bio->bi_sector = r10_bio->devs[j].addr + |
1839 | conf->mirrors[d].rdev->data_offset; | |
1840 | bio->bi_bdev = conf->mirrors[d].rdev->bdev; | |
1841 | atomic_inc(&conf->mirrors[d].rdev->nr_pending); | |
1842 | atomic_inc(&r10_bio->remaining); | |
1843 | /* and we write to 'i' */ | |
1844 | ||
1845 | for (k=0; k<conf->copies; k++) | |
1846 | if (r10_bio->devs[k].devnum == i) | |
1847 | break; | |
64a742bc | 1848 | BUG_ON(k == conf->copies); |
1da177e4 LT |
1849 | bio = r10_bio->devs[1].bio; |
1850 | bio->bi_next = biolist; | |
1851 | biolist = bio; | |
1852 | bio->bi_private = r10_bio; | |
1853 | bio->bi_end_io = end_sync_write; | |
802ba064 | 1854 | bio->bi_rw = WRITE; |
1da177e4 LT |
1855 | bio->bi_sector = r10_bio->devs[k].addr + |
1856 | conf->mirrors[i].rdev->data_offset; | |
1857 | bio->bi_bdev = conf->mirrors[i].rdev->bdev; | |
1858 | ||
1859 | r10_bio->devs[0].devnum = d; | |
1860 | r10_bio->devs[1].devnum = i; | |
1861 | ||
1862 | break; | |
1863 | } | |
1864 | } | |
1865 | if (j == conf->copies) { | |
87fc767b N |
1866 | /* Cannot recover, so abort the recovery */ |
1867 | put_buf(r10_bio); | |
a07e6ab4 T |
1868 | if (rb2) |
1869 | atomic_dec(&rb2->remaining); | |
87fc767b | 1870 | r10_bio = rb2; |
dfc70645 N |
1871 | if (!test_and_set_bit(MD_RECOVERY_INTR, |
1872 | &mddev->recovery)) | |
87fc767b N |
1873 | printk(KERN_INFO "raid10: %s: insufficient working devices for recovery.\n", |
1874 | mdname(mddev)); | |
1875 | break; | |
1da177e4 LT |
1876 | } |
1877 | } | |
1878 | if (biolist == NULL) { | |
1879 | while (r10_bio) { | |
1880 | r10bio_t *rb2 = r10_bio; | |
1881 | r10_bio = (r10bio_t*) rb2->master_bio; | |
1882 | rb2->master_bio = NULL; | |
1883 | put_buf(rb2); | |
1884 | } | |
1885 | goto giveup; | |
1886 | } | |
1887 | } else { | |
1888 | /* resync. Schedule a read for every block at this virt offset */ | |
1889 | int count = 0; | |
6cce3b23 | 1890 | |
78200d45 N |
1891 | bitmap_cond_end_sync(mddev->bitmap, sector_nr); |
1892 | ||
6cce3b23 N |
1893 | if (!bitmap_start_sync(mddev->bitmap, sector_nr, |
1894 | &sync_blocks, mddev->degraded) && | |
1895 | !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) { | |
1896 | /* We can skip this block */ | |
1897 | *skipped = 1; | |
1898 | return sync_blocks + sectors_skipped; | |
1899 | } | |
1900 | if (sync_blocks < max_sync) | |
1901 | max_sync = sync_blocks; | |
1da177e4 LT |
1902 | r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO); |
1903 | ||
1da177e4 LT |
1904 | r10_bio->mddev = mddev; |
1905 | atomic_set(&r10_bio->remaining, 0); | |
6cce3b23 N |
1906 | raise_barrier(conf, 0); |
1907 | conf->next_resync = sector_nr; | |
1da177e4 LT |
1908 | |
1909 | r10_bio->master_bio = NULL; | |
1910 | r10_bio->sector = sector_nr; | |
1911 | set_bit(R10BIO_IsSync, &r10_bio->state); | |
1912 | raid10_find_phys(conf, r10_bio); | |
1913 | r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1; | |
1914 | ||
1915 | for (i=0; i<conf->copies; i++) { | |
1916 | int d = r10_bio->devs[i].devnum; | |
1917 | bio = r10_bio->devs[i].bio; | |
1918 | bio->bi_end_io = NULL; | |
af03b8e4 | 1919 | clear_bit(BIO_UPTODATE, &bio->bi_flags); |
1da177e4 | 1920 | if (conf->mirrors[d].rdev == NULL || |
b2d444d7 | 1921 | test_bit(Faulty, &conf->mirrors[d].rdev->flags)) |
1da177e4 LT |
1922 | continue; |
1923 | atomic_inc(&conf->mirrors[d].rdev->nr_pending); | |
1924 | atomic_inc(&r10_bio->remaining); | |
1925 | bio->bi_next = biolist; | |
1926 | biolist = bio; | |
1927 | bio->bi_private = r10_bio; | |
1928 | bio->bi_end_io = end_sync_read; | |
802ba064 | 1929 | bio->bi_rw = READ; |
1da177e4 LT |
1930 | bio->bi_sector = r10_bio->devs[i].addr + |
1931 | conf->mirrors[d].rdev->data_offset; | |
1932 | bio->bi_bdev = conf->mirrors[d].rdev->bdev; | |
1933 | count++; | |
1934 | } | |
1935 | ||
1936 | if (count < 2) { | |
1937 | for (i=0; i<conf->copies; i++) { | |
1938 | int d = r10_bio->devs[i].devnum; | |
1939 | if (r10_bio->devs[i].bio->bi_end_io) | |
1940 | rdev_dec_pending(conf->mirrors[d].rdev, mddev); | |
1941 | } | |
1942 | put_buf(r10_bio); | |
1943 | biolist = NULL; | |
1944 | goto giveup; | |
1945 | } | |
1946 | } | |
1947 | ||
1948 | for (bio = biolist; bio ; bio=bio->bi_next) { | |
1949 | ||
1950 | bio->bi_flags &= ~(BIO_POOL_MASK - 1); | |
1951 | if (bio->bi_end_io) | |
1952 | bio->bi_flags |= 1 << BIO_UPTODATE; | |
1953 | bio->bi_vcnt = 0; | |
1954 | bio->bi_idx = 0; | |
1955 | bio->bi_phys_segments = 0; | |
1da177e4 LT |
1956 | bio->bi_size = 0; |
1957 | } | |
1958 | ||
1959 | nr_sectors = 0; | |
6cce3b23 N |
1960 | if (sector_nr + max_sync < max_sector) |
1961 | max_sector = sector_nr + max_sync; | |
1da177e4 LT |
1962 | do { |
1963 | struct page *page; | |
1964 | int len = PAGE_SIZE; | |
1965 | disk = 0; | |
1966 | if (sector_nr + (len>>9) > max_sector) | |
1967 | len = (max_sector - sector_nr) << 9; | |
1968 | if (len == 0) | |
1969 | break; | |
1970 | for (bio= biolist ; bio ; bio=bio->bi_next) { | |
1971 | page = bio->bi_io_vec[bio->bi_vcnt].bv_page; | |
1972 | if (bio_add_page(bio, page, len, 0) == 0) { | |
1973 | /* stop here */ | |
1974 | struct bio *bio2; | |
1975 | bio->bi_io_vec[bio->bi_vcnt].bv_page = page; | |
1976 | for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) { | |
1977 | /* remove last page from this bio */ | |
1978 | bio2->bi_vcnt--; | |
1979 | bio2->bi_size -= len; | |
1980 | bio2->bi_flags &= ~(1<< BIO_SEG_VALID); | |
1981 | } | |
1982 | goto bio_full; | |
1983 | } | |
1984 | disk = i; | |
1985 | } | |
1986 | nr_sectors += len>>9; | |
1987 | sector_nr += len>>9; | |
1988 | } while (biolist->bi_vcnt < RESYNC_PAGES); | |
1989 | bio_full: | |
1990 | r10_bio->sectors = nr_sectors; | |
1991 | ||
1992 | while (biolist) { | |
1993 | bio = biolist; | |
1994 | biolist = biolist->bi_next; | |
1995 | ||
1996 | bio->bi_next = NULL; | |
1997 | r10_bio = bio->bi_private; | |
1998 | r10_bio->sectors = nr_sectors; | |
1999 | ||
2000 | if (bio->bi_end_io == end_sync_read) { | |
2001 | md_sync_acct(bio->bi_bdev, nr_sectors); | |
2002 | generic_make_request(bio); | |
2003 | } | |
2004 | } | |
2005 | ||
57afd89f N |
2006 | if (sectors_skipped) |
2007 | /* pretend they weren't skipped, it makes | |
2008 | * no important difference in this case | |
2009 | */ | |
2010 | md_done_sync(mddev, sectors_skipped, 1); | |
2011 | ||
1da177e4 LT |
2012 | return sectors_skipped + nr_sectors; |
2013 | giveup: | |
2014 | /* There is nowhere to write, so all non-sync | |
2015 | * drives must be failed, so try the next chunk... | |
2016 | */ | |
09b4068a N |
2017 | if (sector_nr + max_sync < max_sector) |
2018 | max_sector = sector_nr + max_sync; | |
2019 | ||
2020 | sectors_skipped += (max_sector - sector_nr); | |
1da177e4 LT |
2021 | chunks_skipped ++; |
2022 | sector_nr = max_sector; | |
1da177e4 | 2023 | goto skipped; |
1da177e4 LT |
2024 | } |
2025 | ||
2026 | static int run(mddev_t *mddev) | |
2027 | { | |
2028 | conf_t *conf; | |
2029 | int i, disk_idx; | |
2030 | mirror_info_t *disk; | |
2031 | mdk_rdev_t *rdev; | |
c93983bf | 2032 | int nc, fc, fo; |
1da177e4 LT |
2033 | sector_t stride, size; |
2034 | ||
4bbf3771 N |
2035 | if (mddev->chunk_size < PAGE_SIZE) { |
2036 | printk(KERN_ERR "md/raid10: chunk size must be " | |
2037 | "at least PAGE_SIZE(%ld).\n", PAGE_SIZE); | |
2604b703 | 2038 | return -EINVAL; |
1da177e4 | 2039 | } |
2604b703 | 2040 | |
1da177e4 LT |
2041 | nc = mddev->layout & 255; |
2042 | fc = (mddev->layout >> 8) & 255; | |
c93983bf | 2043 | fo = mddev->layout & (1<<16); |
1da177e4 | 2044 | if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks || |
c93983bf | 2045 | (mddev->layout >> 17)) { |
1da177e4 LT |
2046 | printk(KERN_ERR "raid10: %s: unsupported raid10 layout: 0x%8x\n", |
2047 | mdname(mddev), mddev->layout); | |
2048 | goto out; | |
2049 | } | |
2050 | /* | |
2051 | * copy the already verified devices into our private RAID10 | |
2052 | * bookkeeping area. [whatever we allocate in run(), | |
2053 | * should be freed in stop()] | |
2054 | */ | |
4443ae10 | 2055 | conf = kzalloc(sizeof(conf_t), GFP_KERNEL); |
1da177e4 LT |
2056 | mddev->private = conf; |
2057 | if (!conf) { | |
2058 | printk(KERN_ERR "raid10: couldn't allocate memory for %s\n", | |
2059 | mdname(mddev)); | |
2060 | goto out; | |
2061 | } | |
4443ae10 | 2062 | conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks, |
1da177e4 LT |
2063 | GFP_KERNEL); |
2064 | if (!conf->mirrors) { | |
2065 | printk(KERN_ERR "raid10: couldn't allocate memory for %s\n", | |
2066 | mdname(mddev)); | |
2067 | goto out_free_conf; | |
2068 | } | |
4443ae10 N |
2069 | |
2070 | conf->tmppage = alloc_page(GFP_KERNEL); | |
2071 | if (!conf->tmppage) | |
2072 | goto out_free_conf; | |
1da177e4 | 2073 | |
64a742bc N |
2074 | conf->mddev = mddev; |
2075 | conf->raid_disks = mddev->raid_disks; | |
1da177e4 LT |
2076 | conf->near_copies = nc; |
2077 | conf->far_copies = fc; | |
2078 | conf->copies = nc*fc; | |
c93983bf | 2079 | conf->far_offset = fo; |
1da177e4 LT |
2080 | conf->chunk_mask = (sector_t)(mddev->chunk_size>>9)-1; |
2081 | conf->chunk_shift = ffz(~mddev->chunk_size) - 9; | |
58c0fed4 | 2082 | size = mddev->dev_sectors >> conf->chunk_shift; |
64a742bc N |
2083 | sector_div(size, fc); |
2084 | size = size * conf->raid_disks; | |
2085 | sector_div(size, nc); | |
2086 | /* 'size' is now the number of chunks in the array */ | |
2087 | /* calculate "used chunks per device" in 'stride' */ | |
2088 | stride = size * conf->copies; | |
af03b8e4 N |
2089 | |
2090 | /* We need to round up when dividing by raid_disks to | |
2091 | * get the stride size. | |
2092 | */ | |
2093 | stride += conf->raid_disks - 1; | |
64a742bc | 2094 | sector_div(stride, conf->raid_disks); |
58c0fed4 | 2095 | mddev->dev_sectors = stride << conf->chunk_shift; |
64a742bc | 2096 | |
c93983bf | 2097 | if (fo) |
64a742bc N |
2098 | stride = 1; |
2099 | else | |
c93983bf | 2100 | sector_div(stride, fc); |
64a742bc N |
2101 | conf->stride = stride << conf->chunk_shift; |
2102 | ||
1da177e4 LT |
2103 | conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc, |
2104 | r10bio_pool_free, conf); | |
2105 | if (!conf->r10bio_pool) { | |
2106 | printk(KERN_ERR "raid10: couldn't allocate memory for %s\n", | |
2107 | mdname(mddev)); | |
2108 | goto out_free_conf; | |
2109 | } | |
1da177e4 | 2110 | |
e7e72bf6 NB |
2111 | spin_lock_init(&conf->device_lock); |
2112 | mddev->queue->queue_lock = &conf->device_lock; | |
2113 | ||
159ec1fc | 2114 | list_for_each_entry(rdev, &mddev->disks, same_set) { |
1da177e4 LT |
2115 | disk_idx = rdev->raid_disk; |
2116 | if (disk_idx >= mddev->raid_disks | |
2117 | || disk_idx < 0) | |
2118 | continue; | |
2119 | disk = conf->mirrors + disk_idx; | |
2120 | ||
2121 | disk->rdev = rdev; | |
2122 | ||
2123 | blk_queue_stack_limits(mddev->queue, | |
2124 | rdev->bdev->bd_disk->queue); | |
2125 | /* as we don't honour merge_bvec_fn, we must never risk | |
2126 | * violating it, so limit ->max_sector to one PAGE, as | |
2127 | * a one page request is never in violation. | |
2128 | */ | |
2129 | if (rdev->bdev->bd_disk->queue->merge_bvec_fn && | |
2130 | mddev->queue->max_sectors > (PAGE_SIZE>>9)) | |
2131 | mddev->queue->max_sectors = (PAGE_SIZE>>9); | |
2132 | ||
2133 | disk->head_position = 0; | |
1da177e4 | 2134 | } |
1da177e4 LT |
2135 | INIT_LIST_HEAD(&conf->retry_list); |
2136 | ||
2137 | spin_lock_init(&conf->resync_lock); | |
0a27ec96 | 2138 | init_waitqueue_head(&conf->wait_barrier); |
1da177e4 | 2139 | |
6d508242 N |
2140 | /* need to check that every block has at least one working mirror */ |
2141 | if (!enough(conf)) { | |
2142 | printk(KERN_ERR "raid10: not enough operational mirrors for %s\n", | |
2143 | mdname(mddev)); | |
1da177e4 LT |
2144 | goto out_free_conf; |
2145 | } | |
2146 | ||
2147 | mddev->degraded = 0; | |
2148 | for (i = 0; i < conf->raid_disks; i++) { | |
2149 | ||
2150 | disk = conf->mirrors + i; | |
2151 | ||
5fd6c1dc | 2152 | if (!disk->rdev || |
2e333e89 | 2153 | !test_bit(In_sync, &disk->rdev->flags)) { |
1da177e4 LT |
2154 | disk->head_position = 0; |
2155 | mddev->degraded++; | |
8c2e870a NB |
2156 | if (disk->rdev) |
2157 | conf->fullsync = 1; | |
1da177e4 LT |
2158 | } |
2159 | } | |
2160 | ||
2161 | ||
2162 | mddev->thread = md_register_thread(raid10d, mddev, "%s_raid10"); | |
2163 | if (!mddev->thread) { | |
2164 | printk(KERN_ERR | |
2165 | "raid10: couldn't allocate thread for %s\n", | |
2166 | mdname(mddev)); | |
2167 | goto out_free_conf; | |
2168 | } | |
2169 | ||
2170 | printk(KERN_INFO | |
2171 | "raid10: raid set %s active with %d out of %d devices\n", | |
2172 | mdname(mddev), mddev->raid_disks - mddev->degraded, | |
2173 | mddev->raid_disks); | |
2174 | /* | |
2175 | * Ok, everything is just fine now | |
2176 | */ | |
f233ea5c | 2177 | mddev->array_sectors = size << conf->chunk_shift; |
64a742bc | 2178 | mddev->resync_max_sectors = size << conf->chunk_shift; |
1da177e4 | 2179 | |
7a5febe9 | 2180 | mddev->queue->unplug_fn = raid10_unplug; |
0d129228 N |
2181 | mddev->queue->backing_dev_info.congested_fn = raid10_congested; |
2182 | mddev->queue->backing_dev_info.congested_data = mddev; | |
7a5febe9 | 2183 | |
1da177e4 LT |
2184 | /* Calculate max read-ahead size. |
2185 | * We need to readahead at least twice a whole stripe.... | |
2186 | * maybe... | |
2187 | */ | |
2188 | { | |
8932c2e0 | 2189 | int stripe = conf->raid_disks * (mddev->chunk_size / PAGE_SIZE); |
1da177e4 LT |
2190 | stripe /= conf->near_copies; |
2191 | if (mddev->queue->backing_dev_info.ra_pages < 2* stripe) | |
2192 | mddev->queue->backing_dev_info.ra_pages = 2* stripe; | |
2193 | } | |
2194 | ||
2195 | if (conf->near_copies < mddev->raid_disks) | |
2196 | blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec); | |
2197 | return 0; | |
2198 | ||
2199 | out_free_conf: | |
2200 | if (conf->r10bio_pool) | |
2201 | mempool_destroy(conf->r10bio_pool); | |
1345b1d8 | 2202 | safe_put_page(conf->tmppage); |
990a8baf | 2203 | kfree(conf->mirrors); |
1da177e4 LT |
2204 | kfree(conf); |
2205 | mddev->private = NULL; | |
2206 | out: | |
2207 | return -EIO; | |
2208 | } | |
2209 | ||
2210 | static int stop(mddev_t *mddev) | |
2211 | { | |
2212 | conf_t *conf = mddev_to_conf(mddev); | |
2213 | ||
409c57f3 N |
2214 | raise_barrier(conf, 0); |
2215 | lower_barrier(conf); | |
2216 | ||
1da177e4 LT |
2217 | md_unregister_thread(mddev->thread); |
2218 | mddev->thread = NULL; | |
2219 | blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/ | |
2220 | if (conf->r10bio_pool) | |
2221 | mempool_destroy(conf->r10bio_pool); | |
990a8baf | 2222 | kfree(conf->mirrors); |
1da177e4 LT |
2223 | kfree(conf); |
2224 | mddev->private = NULL; | |
2225 | return 0; | |
2226 | } | |
2227 | ||
6cce3b23 N |
2228 | static void raid10_quiesce(mddev_t *mddev, int state) |
2229 | { | |
2230 | conf_t *conf = mddev_to_conf(mddev); | |
2231 | ||
2232 | switch(state) { | |
2233 | case 1: | |
2234 | raise_barrier(conf, 0); | |
2235 | break; | |
2236 | case 0: | |
2237 | lower_barrier(conf); | |
2238 | break; | |
2239 | } | |
2240 | if (mddev->thread) { | |
2241 | if (mddev->bitmap) | |
2242 | mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ; | |
2243 | else | |
2244 | mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT; | |
2245 | md_wakeup_thread(mddev->thread); | |
2246 | } | |
2247 | } | |
1da177e4 | 2248 | |
2604b703 | 2249 | static struct mdk_personality raid10_personality = |
1da177e4 LT |
2250 | { |
2251 | .name = "raid10", | |
2604b703 | 2252 | .level = 10, |
1da177e4 LT |
2253 | .owner = THIS_MODULE, |
2254 | .make_request = make_request, | |
2255 | .run = run, | |
2256 | .stop = stop, | |
2257 | .status = status, | |
2258 | .error_handler = error, | |
2259 | .hot_add_disk = raid10_add_disk, | |
2260 | .hot_remove_disk= raid10_remove_disk, | |
2261 | .spare_active = raid10_spare_active, | |
2262 | .sync_request = sync_request, | |
6cce3b23 | 2263 | .quiesce = raid10_quiesce, |
1da177e4 LT |
2264 | }; |
2265 | ||
2266 | static int __init raid_init(void) | |
2267 | { | |
2604b703 | 2268 | return register_md_personality(&raid10_personality); |
1da177e4 LT |
2269 | } |
2270 | ||
2271 | static void raid_exit(void) | |
2272 | { | |
2604b703 | 2273 | unregister_md_personality(&raid10_personality); |
1da177e4 LT |
2274 | } |
2275 | ||
2276 | module_init(raid_init); | |
2277 | module_exit(raid_exit); | |
2278 | MODULE_LICENSE("GPL"); | |
2279 | MODULE_ALIAS("md-personality-9"); /* RAID10 */ | |
d9d166c2 | 2280 | MODULE_ALIAS("md-raid10"); |
2604b703 | 2281 | MODULE_ALIAS("md-level-10"); |