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> |
056075c7 | 24 | #include <linux/module.h> |
bff61975 | 25 | #include <linux/seq_file.h> |
8bda470e | 26 | #include <linux/ratelimit.h> |
3ea7daa5 | 27 | #include <linux/kthread.h> |
43b2e5d8 | 28 | #include "md.h" |
ef740c37 | 29 | #include "raid10.h" |
dab8b292 | 30 | #include "raid0.h" |
ef740c37 | 31 | #include "bitmap.h" |
1da177e4 LT |
32 | |
33 | /* | |
34 | * RAID10 provides a combination of RAID0 and RAID1 functionality. | |
35 | * The layout of data is defined by | |
36 | * chunk_size | |
37 | * raid_disks | |
38 | * near_copies (stored in low byte of layout) | |
39 | * far_copies (stored in second byte of layout) | |
c93983bf | 40 | * far_offset (stored in bit 16 of layout ) |
475901af | 41 | * use_far_sets (stored in bit 17 of layout ) |
1da177e4 | 42 | * |
475901af JB |
43 | * The data to be stored is divided into chunks using chunksize. Each device |
44 | * is divided into far_copies sections. In each section, chunks are laid out | |
45 | * in a style similar to raid0, but near_copies copies of each chunk is stored | |
46 | * (each on a different drive). The starting device for each section is offset | |
47 | * near_copies from the starting device of the previous section. Thus there | |
48 | * are (near_copies * far_copies) of each chunk, and each is on a different | |
49 | * drive. near_copies and far_copies must be at least one, and their product | |
50 | * is at most raid_disks. | |
c93983bf N |
51 | * |
52 | * If far_offset is true, then the far_copies are handled a bit differently. | |
475901af JB |
53 | * The copies are still in different stripes, but instead of being very far |
54 | * apart on disk, there are adjacent stripes. | |
55 | * | |
56 | * The far and offset algorithms are handled slightly differently if | |
57 | * 'use_far_sets' is true. In this case, the array's devices are grouped into | |
58 | * sets that are (near_copies * far_copies) in size. The far copied stripes | |
59 | * are still shifted by 'near_copies' devices, but this shifting stays confined | |
60 | * to the set rather than the entire array. This is done to improve the number | |
61 | * of device combinations that can fail without causing the array to fail. | |
62 | * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk | |
63 | * on a device): | |
64 | * A B C D A B C D E | |
65 | * ... ... | |
66 | * D A B C E A B C D | |
67 | * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s): | |
68 | * [A B] [C D] [A B] [C D E] | |
69 | * |...| |...| |...| | ... | | |
70 | * [B A] [D C] [B A] [E C D] | |
1da177e4 LT |
71 | */ |
72 | ||
73 | /* | |
74 | * Number of guaranteed r10bios in case of extreme VM load: | |
75 | */ | |
76 | #define NR_RAID10_BIOS 256 | |
77 | ||
473e87ce JB |
78 | /* when we get a read error on a read-only array, we redirect to another |
79 | * device without failing the first device, or trying to over-write to | |
80 | * correct the read error. To keep track of bad blocks on a per-bio | |
81 | * level, we store IO_BLOCKED in the appropriate 'bios' pointer | |
82 | */ | |
83 | #define IO_BLOCKED ((struct bio *)1) | |
84 | /* When we successfully write to a known bad-block, we need to remove the | |
85 | * bad-block marking which must be done from process context. So we record | |
86 | * the success by setting devs[n].bio to IO_MADE_GOOD | |
87 | */ | |
88 | #define IO_MADE_GOOD ((struct bio *)2) | |
89 | ||
90 | #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2) | |
91 | ||
92 | /* When there are this many requests queued to be written by | |
34db0cd6 N |
93 | * the raid10 thread, we become 'congested' to provide back-pressure |
94 | * for writeback. | |
95 | */ | |
96 | static int max_queued_requests = 1024; | |
97 | ||
e879a879 N |
98 | static void allow_barrier(struct r10conf *conf); |
99 | static void lower_barrier(struct r10conf *conf); | |
635f6416 | 100 | static int _enough(struct r10conf *conf, int previous, int ignore); |
3ea7daa5 N |
101 | static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, |
102 | int *skipped); | |
103 | static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio); | |
104 | static void end_reshape_write(struct bio *bio, int error); | |
105 | static void end_reshape(struct r10conf *conf); | |
0a27ec96 | 106 | |
dd0fc66f | 107 | static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data) |
1da177e4 | 108 | { |
e879a879 | 109 | struct r10conf *conf = data; |
9f2c9d12 | 110 | int size = offsetof(struct r10bio, devs[conf->copies]); |
1da177e4 | 111 | |
69335ef3 N |
112 | /* allocate a r10bio with room for raid_disks entries in the |
113 | * bios array */ | |
7eaceacc | 114 | return kzalloc(size, gfp_flags); |
1da177e4 LT |
115 | } |
116 | ||
117 | static void r10bio_pool_free(void *r10_bio, void *data) | |
118 | { | |
119 | kfree(r10_bio); | |
120 | } | |
121 | ||
0310fa21 | 122 | /* Maximum size of each resync request */ |
1da177e4 | 123 | #define RESYNC_BLOCK_SIZE (64*1024) |
1da177e4 | 124 | #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE) |
0310fa21 N |
125 | /* amount of memory to reserve for resync requests */ |
126 | #define RESYNC_WINDOW (1024*1024) | |
127 | /* maximum number of concurrent requests, memory permitting */ | |
128 | #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE) | |
1da177e4 LT |
129 | |
130 | /* | |
131 | * When performing a resync, we need to read and compare, so | |
132 | * we need as many pages are there are copies. | |
133 | * When performing a recovery, we need 2 bios, one for read, | |
134 | * one for write (we recover only one drive per r10buf) | |
135 | * | |
136 | */ | |
dd0fc66f | 137 | static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data) |
1da177e4 | 138 | { |
e879a879 | 139 | struct r10conf *conf = data; |
1da177e4 | 140 | struct page *page; |
9f2c9d12 | 141 | struct r10bio *r10_bio; |
1da177e4 LT |
142 | struct bio *bio; |
143 | int i, j; | |
144 | int nalloc; | |
145 | ||
146 | r10_bio = r10bio_pool_alloc(gfp_flags, conf); | |
7eaceacc | 147 | if (!r10_bio) |
1da177e4 | 148 | return NULL; |
1da177e4 | 149 | |
3ea7daa5 N |
150 | if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) || |
151 | test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery)) | |
1da177e4 LT |
152 | nalloc = conf->copies; /* resync */ |
153 | else | |
154 | nalloc = 2; /* recovery */ | |
155 | ||
156 | /* | |
157 | * Allocate bios. | |
158 | */ | |
159 | for (j = nalloc ; j-- ; ) { | |
6746557f | 160 | bio = bio_kmalloc(gfp_flags, RESYNC_PAGES); |
1da177e4 LT |
161 | if (!bio) |
162 | goto out_free_bio; | |
163 | r10_bio->devs[j].bio = bio; | |
69335ef3 N |
164 | if (!conf->have_replacement) |
165 | continue; | |
166 | bio = bio_kmalloc(gfp_flags, RESYNC_PAGES); | |
167 | if (!bio) | |
168 | goto out_free_bio; | |
169 | r10_bio->devs[j].repl_bio = bio; | |
1da177e4 LT |
170 | } |
171 | /* | |
172 | * Allocate RESYNC_PAGES data pages and attach them | |
173 | * where needed. | |
174 | */ | |
175 | for (j = 0 ; j < nalloc; j++) { | |
69335ef3 | 176 | struct bio *rbio = r10_bio->devs[j].repl_bio; |
1da177e4 LT |
177 | bio = r10_bio->devs[j].bio; |
178 | for (i = 0; i < RESYNC_PAGES; i++) { | |
3ea7daa5 N |
179 | if (j > 0 && !test_bit(MD_RECOVERY_SYNC, |
180 | &conf->mddev->recovery)) { | |
181 | /* we can share bv_page's during recovery | |
182 | * and reshape */ | |
c65060ad NK |
183 | struct bio *rbio = r10_bio->devs[0].bio; |
184 | page = rbio->bi_io_vec[i].bv_page; | |
185 | get_page(page); | |
186 | } else | |
187 | page = alloc_page(gfp_flags); | |
1da177e4 LT |
188 | if (unlikely(!page)) |
189 | goto out_free_pages; | |
190 | ||
191 | bio->bi_io_vec[i].bv_page = page; | |
69335ef3 N |
192 | if (rbio) |
193 | rbio->bi_io_vec[i].bv_page = page; | |
1da177e4 LT |
194 | } |
195 | } | |
196 | ||
197 | return r10_bio; | |
198 | ||
199 | out_free_pages: | |
200 | for ( ; i > 0 ; i--) | |
1345b1d8 | 201 | safe_put_page(bio->bi_io_vec[i-1].bv_page); |
1da177e4 LT |
202 | while (j--) |
203 | for (i = 0; i < RESYNC_PAGES ; i++) | |
1345b1d8 | 204 | safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page); |
5fdd2cf8 | 205 | j = 0; |
1da177e4 | 206 | out_free_bio: |
5fdd2cf8 | 207 | for ( ; j < nalloc; j++) { |
208 | if (r10_bio->devs[j].bio) | |
209 | bio_put(r10_bio->devs[j].bio); | |
69335ef3 N |
210 | if (r10_bio->devs[j].repl_bio) |
211 | bio_put(r10_bio->devs[j].repl_bio); | |
212 | } | |
1da177e4 LT |
213 | r10bio_pool_free(r10_bio, conf); |
214 | return NULL; | |
215 | } | |
216 | ||
217 | static void r10buf_pool_free(void *__r10_bio, void *data) | |
218 | { | |
219 | int i; | |
e879a879 | 220 | struct r10conf *conf = data; |
9f2c9d12 | 221 | struct r10bio *r10bio = __r10_bio; |
1da177e4 LT |
222 | int j; |
223 | ||
224 | for (j=0; j < conf->copies; j++) { | |
225 | struct bio *bio = r10bio->devs[j].bio; | |
226 | if (bio) { | |
227 | for (i = 0; i < RESYNC_PAGES; i++) { | |
1345b1d8 | 228 | safe_put_page(bio->bi_io_vec[i].bv_page); |
1da177e4 LT |
229 | bio->bi_io_vec[i].bv_page = NULL; |
230 | } | |
231 | bio_put(bio); | |
232 | } | |
69335ef3 N |
233 | bio = r10bio->devs[j].repl_bio; |
234 | if (bio) | |
235 | bio_put(bio); | |
1da177e4 LT |
236 | } |
237 | r10bio_pool_free(r10bio, conf); | |
238 | } | |
239 | ||
e879a879 | 240 | static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio) |
1da177e4 LT |
241 | { |
242 | int i; | |
243 | ||
244 | for (i = 0; i < conf->copies; i++) { | |
245 | struct bio **bio = & r10_bio->devs[i].bio; | |
749c55e9 | 246 | if (!BIO_SPECIAL(*bio)) |
1da177e4 LT |
247 | bio_put(*bio); |
248 | *bio = NULL; | |
69335ef3 N |
249 | bio = &r10_bio->devs[i].repl_bio; |
250 | if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio)) | |
251 | bio_put(*bio); | |
252 | *bio = NULL; | |
1da177e4 LT |
253 | } |
254 | } | |
255 | ||
9f2c9d12 | 256 | static void free_r10bio(struct r10bio *r10_bio) |
1da177e4 | 257 | { |
e879a879 | 258 | struct r10conf *conf = r10_bio->mddev->private; |
1da177e4 | 259 | |
1da177e4 LT |
260 | put_all_bios(conf, r10_bio); |
261 | mempool_free(r10_bio, conf->r10bio_pool); | |
262 | } | |
263 | ||
9f2c9d12 | 264 | static void put_buf(struct r10bio *r10_bio) |
1da177e4 | 265 | { |
e879a879 | 266 | struct r10conf *conf = r10_bio->mddev->private; |
1da177e4 LT |
267 | |
268 | mempool_free(r10_bio, conf->r10buf_pool); | |
269 | ||
0a27ec96 | 270 | lower_barrier(conf); |
1da177e4 LT |
271 | } |
272 | ||
9f2c9d12 | 273 | static void reschedule_retry(struct r10bio *r10_bio) |
1da177e4 LT |
274 | { |
275 | unsigned long flags; | |
fd01b88c | 276 | struct mddev *mddev = r10_bio->mddev; |
e879a879 | 277 | struct r10conf *conf = mddev->private; |
1da177e4 LT |
278 | |
279 | spin_lock_irqsave(&conf->device_lock, flags); | |
280 | list_add(&r10_bio->retry_list, &conf->retry_list); | |
4443ae10 | 281 | conf->nr_queued ++; |
1da177e4 LT |
282 | spin_unlock_irqrestore(&conf->device_lock, flags); |
283 | ||
388667be AJ |
284 | /* wake up frozen array... */ |
285 | wake_up(&conf->wait_barrier); | |
286 | ||
1da177e4 LT |
287 | md_wakeup_thread(mddev->thread); |
288 | } | |
289 | ||
290 | /* | |
291 | * raid_end_bio_io() is called when we have finished servicing a mirrored | |
292 | * operation and are ready to return a success/failure code to the buffer | |
293 | * cache layer. | |
294 | */ | |
9f2c9d12 | 295 | static void raid_end_bio_io(struct r10bio *r10_bio) |
1da177e4 LT |
296 | { |
297 | struct bio *bio = r10_bio->master_bio; | |
856e08e2 | 298 | int done; |
e879a879 | 299 | struct r10conf *conf = r10_bio->mddev->private; |
1da177e4 | 300 | |
856e08e2 N |
301 | if (bio->bi_phys_segments) { |
302 | unsigned long flags; | |
303 | spin_lock_irqsave(&conf->device_lock, flags); | |
304 | bio->bi_phys_segments--; | |
305 | done = (bio->bi_phys_segments == 0); | |
306 | spin_unlock_irqrestore(&conf->device_lock, flags); | |
307 | } else | |
308 | done = 1; | |
309 | if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) | |
310 | clear_bit(BIO_UPTODATE, &bio->bi_flags); | |
311 | if (done) { | |
312 | bio_endio(bio, 0); | |
313 | /* | |
314 | * Wake up any possible resync thread that waits for the device | |
315 | * to go idle. | |
316 | */ | |
317 | allow_barrier(conf); | |
318 | } | |
1da177e4 LT |
319 | free_r10bio(r10_bio); |
320 | } | |
321 | ||
322 | /* | |
323 | * Update disk head position estimator based on IRQ completion info. | |
324 | */ | |
9f2c9d12 | 325 | static inline void update_head_pos(int slot, struct r10bio *r10_bio) |
1da177e4 | 326 | { |
e879a879 | 327 | struct r10conf *conf = r10_bio->mddev->private; |
1da177e4 LT |
328 | |
329 | conf->mirrors[r10_bio->devs[slot].devnum].head_position = | |
330 | r10_bio->devs[slot].addr + (r10_bio->sectors); | |
331 | } | |
332 | ||
778ca018 NK |
333 | /* |
334 | * Find the disk number which triggered given bio | |
335 | */ | |
e879a879 | 336 | static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio, |
69335ef3 | 337 | struct bio *bio, int *slotp, int *replp) |
778ca018 NK |
338 | { |
339 | int slot; | |
69335ef3 | 340 | int repl = 0; |
778ca018 | 341 | |
69335ef3 | 342 | for (slot = 0; slot < conf->copies; slot++) { |
778ca018 NK |
343 | if (r10_bio->devs[slot].bio == bio) |
344 | break; | |
69335ef3 N |
345 | if (r10_bio->devs[slot].repl_bio == bio) { |
346 | repl = 1; | |
347 | break; | |
348 | } | |
349 | } | |
778ca018 NK |
350 | |
351 | BUG_ON(slot == conf->copies); | |
352 | update_head_pos(slot, r10_bio); | |
353 | ||
749c55e9 N |
354 | if (slotp) |
355 | *slotp = slot; | |
69335ef3 N |
356 | if (replp) |
357 | *replp = repl; | |
778ca018 NK |
358 | return r10_bio->devs[slot].devnum; |
359 | } | |
360 | ||
6712ecf8 | 361 | static void raid10_end_read_request(struct bio *bio, int error) |
1da177e4 LT |
362 | { |
363 | int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags); | |
9f2c9d12 | 364 | struct r10bio *r10_bio = bio->bi_private; |
1da177e4 | 365 | int slot, dev; |
abbf098e | 366 | struct md_rdev *rdev; |
e879a879 | 367 | struct r10conf *conf = r10_bio->mddev->private; |
1da177e4 | 368 | |
1da177e4 LT |
369 | slot = r10_bio->read_slot; |
370 | dev = r10_bio->devs[slot].devnum; | |
abbf098e | 371 | rdev = r10_bio->devs[slot].rdev; |
1da177e4 LT |
372 | /* |
373 | * this branch is our 'one mirror IO has finished' event handler: | |
374 | */ | |
4443ae10 N |
375 | update_head_pos(slot, r10_bio); |
376 | ||
377 | if (uptodate) { | |
1da177e4 LT |
378 | /* |
379 | * Set R10BIO_Uptodate in our master bio, so that | |
380 | * we will return a good error code to the higher | |
381 | * levels even if IO on some other mirrored buffer fails. | |
382 | * | |
383 | * The 'master' represents the composite IO operation to | |
384 | * user-side. So if something waits for IO, then it will | |
385 | * wait for the 'master' bio. | |
386 | */ | |
387 | set_bit(R10BIO_Uptodate, &r10_bio->state); | |
fae8cc5e N |
388 | } else { |
389 | /* If all other devices that store this block have | |
390 | * failed, we want to return the error upwards rather | |
391 | * than fail the last device. Here we redefine | |
392 | * "uptodate" to mean "Don't want to retry" | |
393 | */ | |
635f6416 N |
394 | if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state), |
395 | rdev->raid_disk)) | |
fae8cc5e | 396 | uptodate = 1; |
fae8cc5e N |
397 | } |
398 | if (uptodate) { | |
1da177e4 | 399 | raid_end_bio_io(r10_bio); |
abbf098e | 400 | rdev_dec_pending(rdev, conf->mddev); |
4443ae10 | 401 | } else { |
1da177e4 | 402 | /* |
7c4e06ff | 403 | * oops, read error - keep the refcount on the rdev |
1da177e4 LT |
404 | */ |
405 | char b[BDEVNAME_SIZE]; | |
8bda470e CD |
406 | printk_ratelimited(KERN_ERR |
407 | "md/raid10:%s: %s: rescheduling sector %llu\n", | |
408 | mdname(conf->mddev), | |
abbf098e | 409 | bdevname(rdev->bdev, b), |
8bda470e | 410 | (unsigned long long)r10_bio->sector); |
856e08e2 | 411 | set_bit(R10BIO_ReadError, &r10_bio->state); |
1da177e4 LT |
412 | reschedule_retry(r10_bio); |
413 | } | |
1da177e4 LT |
414 | } |
415 | ||
9f2c9d12 | 416 | static void close_write(struct r10bio *r10_bio) |
bd870a16 N |
417 | { |
418 | /* clear the bitmap if all writes complete successfully */ | |
419 | bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector, | |
420 | r10_bio->sectors, | |
421 | !test_bit(R10BIO_Degraded, &r10_bio->state), | |
422 | 0); | |
423 | md_write_end(r10_bio->mddev); | |
424 | } | |
425 | ||
9f2c9d12 | 426 | static void one_write_done(struct r10bio *r10_bio) |
19d5f834 N |
427 | { |
428 | if (atomic_dec_and_test(&r10_bio->remaining)) { | |
429 | if (test_bit(R10BIO_WriteError, &r10_bio->state)) | |
430 | reschedule_retry(r10_bio); | |
431 | else { | |
432 | close_write(r10_bio); | |
433 | if (test_bit(R10BIO_MadeGood, &r10_bio->state)) | |
434 | reschedule_retry(r10_bio); | |
435 | else | |
436 | raid_end_bio_io(r10_bio); | |
437 | } | |
438 | } | |
439 | } | |
440 | ||
6712ecf8 | 441 | static void raid10_end_write_request(struct bio *bio, int error) |
1da177e4 LT |
442 | { |
443 | int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags); | |
9f2c9d12 | 444 | struct r10bio *r10_bio = bio->bi_private; |
778ca018 | 445 | int dev; |
749c55e9 | 446 | int dec_rdev = 1; |
e879a879 | 447 | struct r10conf *conf = r10_bio->mddev->private; |
475b0321 | 448 | int slot, repl; |
4ca40c2c | 449 | struct md_rdev *rdev = NULL; |
1da177e4 | 450 | |
475b0321 | 451 | dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl); |
1da177e4 | 452 | |
475b0321 N |
453 | if (repl) |
454 | rdev = conf->mirrors[dev].replacement; | |
4ca40c2c N |
455 | if (!rdev) { |
456 | smp_rmb(); | |
457 | repl = 0; | |
475b0321 | 458 | rdev = conf->mirrors[dev].rdev; |
4ca40c2c | 459 | } |
1da177e4 LT |
460 | /* |
461 | * this branch is our 'one mirror IO has finished' event handler: | |
462 | */ | |
6cce3b23 | 463 | if (!uptodate) { |
475b0321 N |
464 | if (repl) |
465 | /* Never record new bad blocks to replacement, | |
466 | * just fail it. | |
467 | */ | |
468 | md_error(rdev->mddev, rdev); | |
469 | else { | |
470 | set_bit(WriteErrorSeen, &rdev->flags); | |
b7044d41 N |
471 | if (!test_and_set_bit(WantReplacement, &rdev->flags)) |
472 | set_bit(MD_RECOVERY_NEEDED, | |
473 | &rdev->mddev->recovery); | |
475b0321 N |
474 | set_bit(R10BIO_WriteError, &r10_bio->state); |
475 | dec_rdev = 0; | |
476 | } | |
749c55e9 | 477 | } else { |
1da177e4 LT |
478 | /* |
479 | * Set R10BIO_Uptodate in our master bio, so that | |
480 | * we will return a good error code for to the higher | |
481 | * levels even if IO on some other mirrored buffer fails. | |
482 | * | |
483 | * The 'master' represents the composite IO operation to | |
484 | * user-side. So if something waits for IO, then it will | |
485 | * wait for the 'master' bio. | |
486 | */ | |
749c55e9 N |
487 | sector_t first_bad; |
488 | int bad_sectors; | |
489 | ||
3056e3ae AL |
490 | /* |
491 | * Do not set R10BIO_Uptodate if the current device is | |
492 | * rebuilding or Faulty. This is because we cannot use | |
493 | * such device for properly reading the data back (we could | |
494 | * potentially use it, if the current write would have felt | |
495 | * before rdev->recovery_offset, but for simplicity we don't | |
496 | * check this here. | |
497 | */ | |
498 | if (test_bit(In_sync, &rdev->flags) && | |
499 | !test_bit(Faulty, &rdev->flags)) | |
500 | set_bit(R10BIO_Uptodate, &r10_bio->state); | |
1da177e4 | 501 | |
749c55e9 | 502 | /* Maybe we can clear some bad blocks. */ |
475b0321 | 503 | if (is_badblock(rdev, |
749c55e9 N |
504 | r10_bio->devs[slot].addr, |
505 | r10_bio->sectors, | |
506 | &first_bad, &bad_sectors)) { | |
507 | bio_put(bio); | |
475b0321 N |
508 | if (repl) |
509 | r10_bio->devs[slot].repl_bio = IO_MADE_GOOD; | |
510 | else | |
511 | r10_bio->devs[slot].bio = IO_MADE_GOOD; | |
749c55e9 N |
512 | dec_rdev = 0; |
513 | set_bit(R10BIO_MadeGood, &r10_bio->state); | |
514 | } | |
515 | } | |
516 | ||
1da177e4 LT |
517 | /* |
518 | * | |
519 | * Let's see if all mirrored write operations have finished | |
520 | * already. | |
521 | */ | |
19d5f834 | 522 | one_write_done(r10_bio); |
749c55e9 | 523 | if (dec_rdev) |
884162df | 524 | rdev_dec_pending(rdev, conf->mddev); |
1da177e4 LT |
525 | } |
526 | ||
1da177e4 LT |
527 | /* |
528 | * RAID10 layout manager | |
25985edc | 529 | * As well as the chunksize and raid_disks count, there are two |
1da177e4 LT |
530 | * parameters: near_copies and far_copies. |
531 | * near_copies * far_copies must be <= raid_disks. | |
532 | * Normally one of these will be 1. | |
533 | * If both are 1, we get raid0. | |
534 | * If near_copies == raid_disks, we get raid1. | |
535 | * | |
25985edc | 536 | * Chunks are laid out in raid0 style with near_copies copies of the |
1da177e4 LT |
537 | * first chunk, followed by near_copies copies of the next chunk and |
538 | * so on. | |
539 | * If far_copies > 1, then after 1/far_copies of the array has been assigned | |
540 | * as described above, we start again with a device offset of near_copies. | |
541 | * So we effectively have another copy of the whole array further down all | |
542 | * the drives, but with blocks on different drives. | |
543 | * With this layout, and block is never stored twice on the one device. | |
544 | * | |
545 | * raid10_find_phys finds the sector offset of a given virtual sector | |
c93983bf | 546 | * on each device that it is on. |
1da177e4 LT |
547 | * |
548 | * raid10_find_virt does the reverse mapping, from a device and a | |
549 | * sector offset to a virtual address | |
550 | */ | |
551 | ||
f8c9e74f | 552 | static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio) |
1da177e4 LT |
553 | { |
554 | int n,f; | |
555 | sector_t sector; | |
556 | sector_t chunk; | |
557 | sector_t stripe; | |
558 | int dev; | |
1da177e4 | 559 | int slot = 0; |
9a3152ab JB |
560 | int last_far_set_start, last_far_set_size; |
561 | ||
562 | last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1; | |
563 | last_far_set_start *= geo->far_set_size; | |
564 | ||
565 | last_far_set_size = geo->far_set_size; | |
566 | last_far_set_size += (geo->raid_disks % geo->far_set_size); | |
1da177e4 LT |
567 | |
568 | /* now calculate first sector/dev */ | |
5cf00fcd N |
569 | chunk = r10bio->sector >> geo->chunk_shift; |
570 | sector = r10bio->sector & geo->chunk_mask; | |
1da177e4 | 571 | |
5cf00fcd | 572 | chunk *= geo->near_copies; |
1da177e4 | 573 | stripe = chunk; |
5cf00fcd N |
574 | dev = sector_div(stripe, geo->raid_disks); |
575 | if (geo->far_offset) | |
576 | stripe *= geo->far_copies; | |
1da177e4 | 577 | |
5cf00fcd | 578 | sector += stripe << geo->chunk_shift; |
1da177e4 LT |
579 | |
580 | /* and calculate all the others */ | |
5cf00fcd | 581 | for (n = 0; n < geo->near_copies; n++) { |
1da177e4 | 582 | int d = dev; |
475901af | 583 | int set; |
1da177e4 | 584 | sector_t s = sector; |
1da177e4 | 585 | r10bio->devs[slot].devnum = d; |
4c0ca26b | 586 | r10bio->devs[slot].addr = s; |
1da177e4 LT |
587 | slot++; |
588 | ||
5cf00fcd | 589 | for (f = 1; f < geo->far_copies; f++) { |
475901af | 590 | set = d / geo->far_set_size; |
5cf00fcd | 591 | d += geo->near_copies; |
475901af | 592 | |
9a3152ab JB |
593 | if ((geo->raid_disks % geo->far_set_size) && |
594 | (d > last_far_set_start)) { | |
595 | d -= last_far_set_start; | |
596 | d %= last_far_set_size; | |
597 | d += last_far_set_start; | |
598 | } else { | |
599 | d %= geo->far_set_size; | |
600 | d += geo->far_set_size * set; | |
601 | } | |
5cf00fcd | 602 | s += geo->stride; |
1da177e4 LT |
603 | r10bio->devs[slot].devnum = d; |
604 | r10bio->devs[slot].addr = s; | |
605 | slot++; | |
606 | } | |
607 | dev++; | |
5cf00fcd | 608 | if (dev >= geo->raid_disks) { |
1da177e4 | 609 | dev = 0; |
5cf00fcd | 610 | sector += (geo->chunk_mask + 1); |
1da177e4 LT |
611 | } |
612 | } | |
f8c9e74f N |
613 | } |
614 | ||
615 | static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio) | |
616 | { | |
617 | struct geom *geo = &conf->geo; | |
618 | ||
619 | if (conf->reshape_progress != MaxSector && | |
620 | ((r10bio->sector >= conf->reshape_progress) != | |
621 | conf->mddev->reshape_backwards)) { | |
622 | set_bit(R10BIO_Previous, &r10bio->state); | |
623 | geo = &conf->prev; | |
624 | } else | |
625 | clear_bit(R10BIO_Previous, &r10bio->state); | |
626 | ||
627 | __raid10_find_phys(geo, r10bio); | |
1da177e4 LT |
628 | } |
629 | ||
e879a879 | 630 | static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev) |
1da177e4 LT |
631 | { |
632 | sector_t offset, chunk, vchunk; | |
f8c9e74f N |
633 | /* Never use conf->prev as this is only called during resync |
634 | * or recovery, so reshape isn't happening | |
635 | */ | |
5cf00fcd | 636 | struct geom *geo = &conf->geo; |
475901af JB |
637 | int far_set_start = (dev / geo->far_set_size) * geo->far_set_size; |
638 | int far_set_size = geo->far_set_size; | |
9a3152ab JB |
639 | int last_far_set_start; |
640 | ||
641 | if (geo->raid_disks % geo->far_set_size) { | |
642 | last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1; | |
643 | last_far_set_start *= geo->far_set_size; | |
644 | ||
645 | if (dev >= last_far_set_start) { | |
646 | far_set_size = geo->far_set_size; | |
647 | far_set_size += (geo->raid_disks % geo->far_set_size); | |
648 | far_set_start = last_far_set_start; | |
649 | } | |
650 | } | |
1da177e4 | 651 | |
5cf00fcd N |
652 | offset = sector & geo->chunk_mask; |
653 | if (geo->far_offset) { | |
c93983bf | 654 | int fc; |
5cf00fcd N |
655 | chunk = sector >> geo->chunk_shift; |
656 | fc = sector_div(chunk, geo->far_copies); | |
657 | dev -= fc * geo->near_copies; | |
475901af JB |
658 | if (dev < far_set_start) |
659 | dev += far_set_size; | |
c93983bf | 660 | } else { |
5cf00fcd N |
661 | while (sector >= geo->stride) { |
662 | sector -= geo->stride; | |
475901af JB |
663 | if (dev < (geo->near_copies + far_set_start)) |
664 | dev += far_set_size - geo->near_copies; | |
c93983bf | 665 | else |
5cf00fcd | 666 | dev -= geo->near_copies; |
c93983bf | 667 | } |
5cf00fcd | 668 | chunk = sector >> geo->chunk_shift; |
c93983bf | 669 | } |
5cf00fcd N |
670 | vchunk = chunk * geo->raid_disks + dev; |
671 | sector_div(vchunk, geo->near_copies); | |
672 | return (vchunk << geo->chunk_shift) + offset; | |
1da177e4 LT |
673 | } |
674 | ||
675 | /** | |
676 | * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged | |
64590f45 | 677 | * @mddev: the md device |
cc371e66 | 678 | * @bvm: properties of new bio |
1da177e4 LT |
679 | * @biovec: the request that could be merged to it. |
680 | * | |
681 | * Return amount of bytes we can accept at this offset | |
050b6615 N |
682 | * This requires checking for end-of-chunk if near_copies != raid_disks, |
683 | * and for subordinate merge_bvec_fns if merge_check_needed. | |
1da177e4 | 684 | */ |
64590f45 | 685 | static int raid10_mergeable_bvec(struct mddev *mddev, |
cc371e66 AK |
686 | struct bvec_merge_data *bvm, |
687 | struct bio_vec *biovec) | |
1da177e4 | 688 | { |
050b6615 | 689 | struct r10conf *conf = mddev->private; |
cc371e66 | 690 | sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev); |
1da177e4 | 691 | int max; |
3ea7daa5 | 692 | unsigned int chunk_sectors; |
cc371e66 | 693 | unsigned int bio_sectors = bvm->bi_size >> 9; |
5cf00fcd | 694 | struct geom *geo = &conf->geo; |
1da177e4 | 695 | |
3ea7daa5 | 696 | chunk_sectors = (conf->geo.chunk_mask & conf->prev.chunk_mask) + 1; |
f8c9e74f N |
697 | if (conf->reshape_progress != MaxSector && |
698 | ((sector >= conf->reshape_progress) != | |
699 | conf->mddev->reshape_backwards)) | |
700 | geo = &conf->prev; | |
701 | ||
5cf00fcd | 702 | if (geo->near_copies < geo->raid_disks) { |
050b6615 N |
703 | max = (chunk_sectors - ((sector & (chunk_sectors - 1)) |
704 | + bio_sectors)) << 9; | |
705 | if (max < 0) | |
706 | /* bio_add cannot handle a negative return */ | |
707 | max = 0; | |
708 | if (max <= biovec->bv_len && bio_sectors == 0) | |
709 | return biovec->bv_len; | |
710 | } else | |
711 | max = biovec->bv_len; | |
712 | ||
713 | if (mddev->merge_check_needed) { | |
e0ee7785 N |
714 | struct { |
715 | struct r10bio r10_bio; | |
716 | struct r10dev devs[conf->copies]; | |
717 | } on_stack; | |
718 | struct r10bio *r10_bio = &on_stack.r10_bio; | |
050b6615 | 719 | int s; |
f8c9e74f N |
720 | if (conf->reshape_progress != MaxSector) { |
721 | /* Cannot give any guidance during reshape */ | |
722 | if (max <= biovec->bv_len && bio_sectors == 0) | |
723 | return biovec->bv_len; | |
724 | return 0; | |
725 | } | |
e0ee7785 N |
726 | r10_bio->sector = sector; |
727 | raid10_find_phys(conf, r10_bio); | |
050b6615 N |
728 | rcu_read_lock(); |
729 | for (s = 0; s < conf->copies; s++) { | |
e0ee7785 | 730 | int disk = r10_bio->devs[s].devnum; |
050b6615 N |
731 | struct md_rdev *rdev = rcu_dereference( |
732 | conf->mirrors[disk].rdev); | |
733 | if (rdev && !test_bit(Faulty, &rdev->flags)) { | |
734 | struct request_queue *q = | |
735 | bdev_get_queue(rdev->bdev); | |
736 | if (q->merge_bvec_fn) { | |
e0ee7785 | 737 | bvm->bi_sector = r10_bio->devs[s].addr |
050b6615 N |
738 | + rdev->data_offset; |
739 | bvm->bi_bdev = rdev->bdev; | |
740 | max = min(max, q->merge_bvec_fn( | |
741 | q, bvm, biovec)); | |
742 | } | |
743 | } | |
744 | rdev = rcu_dereference(conf->mirrors[disk].replacement); | |
745 | if (rdev && !test_bit(Faulty, &rdev->flags)) { | |
746 | struct request_queue *q = | |
747 | bdev_get_queue(rdev->bdev); | |
748 | if (q->merge_bvec_fn) { | |
e0ee7785 | 749 | bvm->bi_sector = r10_bio->devs[s].addr |
050b6615 N |
750 | + rdev->data_offset; |
751 | bvm->bi_bdev = rdev->bdev; | |
752 | max = min(max, q->merge_bvec_fn( | |
753 | q, bvm, biovec)); | |
754 | } | |
755 | } | |
756 | } | |
757 | rcu_read_unlock(); | |
758 | } | |
759 | return max; | |
1da177e4 LT |
760 | } |
761 | ||
762 | /* | |
763 | * This routine returns the disk from which the requested read should | |
764 | * be done. There is a per-array 'next expected sequential IO' sector | |
765 | * number - if this matches on the next IO then we use the last disk. | |
766 | * There is also a per-disk 'last know head position' sector that is | |
767 | * maintained from IRQ contexts, both the normal and the resync IO | |
768 | * completion handlers update this position correctly. If there is no | |
769 | * perfect sequential match then we pick the disk whose head is closest. | |
770 | * | |
771 | * If there are 2 mirrors in the same 2 devices, performance degrades | |
772 | * because position is mirror, not device based. | |
773 | * | |
774 | * The rdev for the device selected will have nr_pending incremented. | |
775 | */ | |
776 | ||
777 | /* | |
778 | * FIXME: possibly should rethink readbalancing and do it differently | |
779 | * depending on near_copies / far_copies geometry. | |
780 | */ | |
96c3fd1f N |
781 | static struct md_rdev *read_balance(struct r10conf *conf, |
782 | struct r10bio *r10_bio, | |
783 | int *max_sectors) | |
1da177e4 | 784 | { |
af3a2cd6 | 785 | const sector_t this_sector = r10_bio->sector; |
56d99121 | 786 | int disk, slot; |
856e08e2 N |
787 | int sectors = r10_bio->sectors; |
788 | int best_good_sectors; | |
56d99121 | 789 | sector_t new_distance, best_dist; |
3bbae04b | 790 | struct md_rdev *best_rdev, *rdev = NULL; |
56d99121 N |
791 | int do_balance; |
792 | int best_slot; | |
5cf00fcd | 793 | struct geom *geo = &conf->geo; |
1da177e4 LT |
794 | |
795 | raid10_find_phys(conf, r10_bio); | |
796 | rcu_read_lock(); | |
56d99121 | 797 | retry: |
856e08e2 | 798 | sectors = r10_bio->sectors; |
56d99121 | 799 | best_slot = -1; |
abbf098e | 800 | best_rdev = NULL; |
56d99121 | 801 | best_dist = MaxSector; |
856e08e2 | 802 | best_good_sectors = 0; |
56d99121 | 803 | do_balance = 1; |
1da177e4 LT |
804 | /* |
805 | * Check if we can balance. We can balance on the whole | |
6cce3b23 N |
806 | * device if no resync is going on (recovery is ok), or below |
807 | * the resync window. We take the first readable disk when | |
808 | * above the resync window. | |
1da177e4 LT |
809 | */ |
810 | if (conf->mddev->recovery_cp < MaxSector | |
56d99121 N |
811 | && (this_sector + sectors >= conf->next_resync)) |
812 | do_balance = 0; | |
1da177e4 | 813 | |
56d99121 | 814 | for (slot = 0; slot < conf->copies ; slot++) { |
856e08e2 N |
815 | sector_t first_bad; |
816 | int bad_sectors; | |
817 | sector_t dev_sector; | |
818 | ||
56d99121 N |
819 | if (r10_bio->devs[slot].bio == IO_BLOCKED) |
820 | continue; | |
1da177e4 | 821 | disk = r10_bio->devs[slot].devnum; |
abbf098e N |
822 | rdev = rcu_dereference(conf->mirrors[disk].replacement); |
823 | if (rdev == NULL || test_bit(Faulty, &rdev->flags) || | |
050b6615 | 824 | test_bit(Unmerged, &rdev->flags) || |
abbf098e N |
825 | r10_bio->devs[slot].addr + sectors > rdev->recovery_offset) |
826 | rdev = rcu_dereference(conf->mirrors[disk].rdev); | |
050b6615 N |
827 | if (rdev == NULL || |
828 | test_bit(Faulty, &rdev->flags) || | |
829 | test_bit(Unmerged, &rdev->flags)) | |
abbf098e N |
830 | continue; |
831 | if (!test_bit(In_sync, &rdev->flags) && | |
832 | r10_bio->devs[slot].addr + sectors > rdev->recovery_offset) | |
56d99121 N |
833 | continue; |
834 | ||
856e08e2 N |
835 | dev_sector = r10_bio->devs[slot].addr; |
836 | if (is_badblock(rdev, dev_sector, sectors, | |
837 | &first_bad, &bad_sectors)) { | |
838 | if (best_dist < MaxSector) | |
839 | /* Already have a better slot */ | |
840 | continue; | |
841 | if (first_bad <= dev_sector) { | |
842 | /* Cannot read here. If this is the | |
843 | * 'primary' device, then we must not read | |
844 | * beyond 'bad_sectors' from another device. | |
845 | */ | |
846 | bad_sectors -= (dev_sector - first_bad); | |
847 | if (!do_balance && sectors > bad_sectors) | |
848 | sectors = bad_sectors; | |
849 | if (best_good_sectors > sectors) | |
850 | best_good_sectors = sectors; | |
851 | } else { | |
852 | sector_t good_sectors = | |
853 | first_bad - dev_sector; | |
854 | if (good_sectors > best_good_sectors) { | |
855 | best_good_sectors = good_sectors; | |
856 | best_slot = slot; | |
abbf098e | 857 | best_rdev = rdev; |
856e08e2 N |
858 | } |
859 | if (!do_balance) | |
860 | /* Must read from here */ | |
861 | break; | |
862 | } | |
863 | continue; | |
864 | } else | |
865 | best_good_sectors = sectors; | |
866 | ||
56d99121 N |
867 | if (!do_balance) |
868 | break; | |
1da177e4 | 869 | |
22dfdf52 N |
870 | /* This optimisation is debatable, and completely destroys |
871 | * sequential read speed for 'far copies' arrays. So only | |
872 | * keep it for 'near' arrays, and review those later. | |
873 | */ | |
5cf00fcd | 874 | if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending)) |
1da177e4 | 875 | break; |
8ed3a195 KS |
876 | |
877 | /* for far > 1 always use the lowest address */ | |
5cf00fcd | 878 | if (geo->far_copies > 1) |
56d99121 | 879 | new_distance = r10_bio->devs[slot].addr; |
8ed3a195 | 880 | else |
56d99121 N |
881 | new_distance = abs(r10_bio->devs[slot].addr - |
882 | conf->mirrors[disk].head_position); | |
883 | if (new_distance < best_dist) { | |
884 | best_dist = new_distance; | |
885 | best_slot = slot; | |
abbf098e | 886 | best_rdev = rdev; |
1da177e4 LT |
887 | } |
888 | } | |
abbf098e | 889 | if (slot >= conf->copies) { |
56d99121 | 890 | slot = best_slot; |
abbf098e N |
891 | rdev = best_rdev; |
892 | } | |
1da177e4 | 893 | |
56d99121 | 894 | if (slot >= 0) { |
56d99121 N |
895 | atomic_inc(&rdev->nr_pending); |
896 | if (test_bit(Faulty, &rdev->flags)) { | |
897 | /* Cannot risk returning a device that failed | |
898 | * before we inc'ed nr_pending | |
899 | */ | |
900 | rdev_dec_pending(rdev, conf->mddev); | |
901 | goto retry; | |
902 | } | |
903 | r10_bio->read_slot = slot; | |
904 | } else | |
96c3fd1f | 905 | rdev = NULL; |
1da177e4 | 906 | rcu_read_unlock(); |
856e08e2 | 907 | *max_sectors = best_good_sectors; |
1da177e4 | 908 | |
96c3fd1f | 909 | return rdev; |
1da177e4 LT |
910 | } |
911 | ||
5c675f83 | 912 | static int raid10_congested(struct mddev *mddev, int bits) |
0d129228 | 913 | { |
e879a879 | 914 | struct r10conf *conf = mddev->private; |
0d129228 N |
915 | int i, ret = 0; |
916 | ||
34db0cd6 N |
917 | if ((bits & (1 << BDI_async_congested)) && |
918 | conf->pending_count >= max_queued_requests) | |
919 | return 1; | |
920 | ||
0d129228 | 921 | rcu_read_lock(); |
f8c9e74f N |
922 | for (i = 0; |
923 | (i < conf->geo.raid_disks || i < conf->prev.raid_disks) | |
924 | && ret == 0; | |
925 | i++) { | |
3cb03002 | 926 | struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); |
0d129228 | 927 | if (rdev && !test_bit(Faulty, &rdev->flags)) { |
165125e1 | 928 | struct request_queue *q = bdev_get_queue(rdev->bdev); |
0d129228 N |
929 | |
930 | ret |= bdi_congested(&q->backing_dev_info, bits); | |
931 | } | |
932 | } | |
933 | rcu_read_unlock(); | |
934 | return ret; | |
935 | } | |
936 | ||
e879a879 | 937 | static void flush_pending_writes(struct r10conf *conf) |
a35e63ef N |
938 | { |
939 | /* Any writes that have been queued but are awaiting | |
940 | * bitmap updates get flushed here. | |
a35e63ef | 941 | */ |
a35e63ef N |
942 | spin_lock_irq(&conf->device_lock); |
943 | ||
944 | if (conf->pending_bio_list.head) { | |
945 | struct bio *bio; | |
946 | bio = bio_list_get(&conf->pending_bio_list); | |
34db0cd6 | 947 | conf->pending_count = 0; |
a35e63ef N |
948 | spin_unlock_irq(&conf->device_lock); |
949 | /* flush any pending bitmap writes to disk | |
950 | * before proceeding w/ I/O */ | |
951 | bitmap_unplug(conf->mddev->bitmap); | |
34db0cd6 | 952 | wake_up(&conf->wait_barrier); |
a35e63ef N |
953 | |
954 | while (bio) { /* submit pending writes */ | |
955 | struct bio *next = bio->bi_next; | |
956 | bio->bi_next = NULL; | |
532a2a3f SL |
957 | if (unlikely((bio->bi_rw & REQ_DISCARD) && |
958 | !blk_queue_discard(bdev_get_queue(bio->bi_bdev)))) | |
959 | /* Just ignore it */ | |
960 | bio_endio(bio, 0); | |
961 | else | |
962 | generic_make_request(bio); | |
a35e63ef N |
963 | bio = next; |
964 | } | |
a35e63ef N |
965 | } else |
966 | spin_unlock_irq(&conf->device_lock); | |
a35e63ef | 967 | } |
7eaceacc | 968 | |
0a27ec96 N |
969 | /* Barriers.... |
970 | * Sometimes we need to suspend IO while we do something else, | |
971 | * either some resync/recovery, or reconfigure the array. | |
972 | * To do this we raise a 'barrier'. | |
973 | * The 'barrier' is a counter that can be raised multiple times | |
974 | * to count how many activities are happening which preclude | |
975 | * normal IO. | |
976 | * We can only raise the barrier if there is no pending IO. | |
977 | * i.e. if nr_pending == 0. | |
978 | * We choose only to raise the barrier if no-one is waiting for the | |
979 | * barrier to go down. This means that as soon as an IO request | |
980 | * is ready, no other operations which require a barrier will start | |
981 | * until the IO request has had a chance. | |
982 | * | |
983 | * So: regular IO calls 'wait_barrier'. When that returns there | |
984 | * is no backgroup IO happening, It must arrange to call | |
985 | * allow_barrier when it has finished its IO. | |
986 | * backgroup IO calls must call raise_barrier. Once that returns | |
987 | * there is no normal IO happeing. It must arrange to call | |
988 | * lower_barrier when the particular background IO completes. | |
1da177e4 | 989 | */ |
1da177e4 | 990 | |
e879a879 | 991 | static void raise_barrier(struct r10conf *conf, int force) |
1da177e4 | 992 | { |
6cce3b23 | 993 | BUG_ON(force && !conf->barrier); |
1da177e4 | 994 | spin_lock_irq(&conf->resync_lock); |
0a27ec96 | 995 | |
6cce3b23 N |
996 | /* Wait until no block IO is waiting (unless 'force') */ |
997 | wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting, | |
eed8c02e | 998 | conf->resync_lock); |
0a27ec96 N |
999 | |
1000 | /* block any new IO from starting */ | |
1001 | conf->barrier++; | |
1002 | ||
c3b328ac | 1003 | /* Now wait for all pending IO to complete */ |
0a27ec96 N |
1004 | wait_event_lock_irq(conf->wait_barrier, |
1005 | !conf->nr_pending && conf->barrier < RESYNC_DEPTH, | |
eed8c02e | 1006 | conf->resync_lock); |
0a27ec96 N |
1007 | |
1008 | spin_unlock_irq(&conf->resync_lock); | |
1009 | } | |
1010 | ||
e879a879 | 1011 | static void lower_barrier(struct r10conf *conf) |
0a27ec96 N |
1012 | { |
1013 | unsigned long flags; | |
1014 | spin_lock_irqsave(&conf->resync_lock, flags); | |
1015 | conf->barrier--; | |
1016 | spin_unlock_irqrestore(&conf->resync_lock, flags); | |
1017 | wake_up(&conf->wait_barrier); | |
1018 | } | |
1019 | ||
e879a879 | 1020 | static void wait_barrier(struct r10conf *conf) |
0a27ec96 N |
1021 | { |
1022 | spin_lock_irq(&conf->resync_lock); | |
1023 | if (conf->barrier) { | |
1024 | conf->nr_waiting++; | |
d6b42dcb N |
1025 | /* Wait for the barrier to drop. |
1026 | * However if there are already pending | |
1027 | * requests (preventing the barrier from | |
1028 | * rising completely), and the | |
1029 | * pre-process bio queue isn't empty, | |
1030 | * then don't wait, as we need to empty | |
1031 | * that queue to get the nr_pending | |
1032 | * count down. | |
1033 | */ | |
1034 | wait_event_lock_irq(conf->wait_barrier, | |
1035 | !conf->barrier || | |
1036 | (conf->nr_pending && | |
1037 | current->bio_list && | |
1038 | !bio_list_empty(current->bio_list)), | |
eed8c02e | 1039 | conf->resync_lock); |
0a27ec96 | 1040 | conf->nr_waiting--; |
1da177e4 | 1041 | } |
0a27ec96 | 1042 | conf->nr_pending++; |
1da177e4 LT |
1043 | spin_unlock_irq(&conf->resync_lock); |
1044 | } | |
1045 | ||
e879a879 | 1046 | static void allow_barrier(struct r10conf *conf) |
0a27ec96 N |
1047 | { |
1048 | unsigned long flags; | |
1049 | spin_lock_irqsave(&conf->resync_lock, flags); | |
1050 | conf->nr_pending--; | |
1051 | spin_unlock_irqrestore(&conf->resync_lock, flags); | |
1052 | wake_up(&conf->wait_barrier); | |
1053 | } | |
1054 | ||
e2d59925 | 1055 | static void freeze_array(struct r10conf *conf, int extra) |
4443ae10 N |
1056 | { |
1057 | /* stop syncio and normal IO and wait for everything to | |
f188593e | 1058 | * go quiet. |
4443ae10 | 1059 | * We increment barrier and nr_waiting, and then |
e2d59925 | 1060 | * wait until nr_pending match nr_queued+extra |
1c830532 N |
1061 | * This is called in the context of one normal IO request |
1062 | * that has failed. Thus any sync request that might be pending | |
1063 | * will be blocked by nr_pending, and we need to wait for | |
1064 | * pending IO requests to complete or be queued for re-try. | |
e2d59925 | 1065 | * Thus the number queued (nr_queued) plus this request (extra) |
1c830532 N |
1066 | * must match the number of pending IOs (nr_pending) before |
1067 | * we continue. | |
4443ae10 N |
1068 | */ |
1069 | spin_lock_irq(&conf->resync_lock); | |
1070 | conf->barrier++; | |
1071 | conf->nr_waiting++; | |
eed8c02e | 1072 | wait_event_lock_irq_cmd(conf->wait_barrier, |
e2d59925 | 1073 | conf->nr_pending == conf->nr_queued+extra, |
eed8c02e LC |
1074 | conf->resync_lock, |
1075 | flush_pending_writes(conf)); | |
c3b328ac | 1076 | |
4443ae10 N |
1077 | spin_unlock_irq(&conf->resync_lock); |
1078 | } | |
1079 | ||
e879a879 | 1080 | static void unfreeze_array(struct r10conf *conf) |
4443ae10 N |
1081 | { |
1082 | /* reverse the effect of the freeze */ | |
1083 | spin_lock_irq(&conf->resync_lock); | |
1084 | conf->barrier--; | |
1085 | conf->nr_waiting--; | |
1086 | wake_up(&conf->wait_barrier); | |
1087 | spin_unlock_irq(&conf->resync_lock); | |
1088 | } | |
1089 | ||
f8c9e74f N |
1090 | static sector_t choose_data_offset(struct r10bio *r10_bio, |
1091 | struct md_rdev *rdev) | |
1092 | { | |
1093 | if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) || | |
1094 | test_bit(R10BIO_Previous, &r10_bio->state)) | |
1095 | return rdev->data_offset; | |
1096 | else | |
1097 | return rdev->new_data_offset; | |
1098 | } | |
1099 | ||
57c67df4 N |
1100 | struct raid10_plug_cb { |
1101 | struct blk_plug_cb cb; | |
1102 | struct bio_list pending; | |
1103 | int pending_cnt; | |
1104 | }; | |
1105 | ||
1106 | static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule) | |
1107 | { | |
1108 | struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb, | |
1109 | cb); | |
1110 | struct mddev *mddev = plug->cb.data; | |
1111 | struct r10conf *conf = mddev->private; | |
1112 | struct bio *bio; | |
1113 | ||
874807a8 | 1114 | if (from_schedule || current->bio_list) { |
57c67df4 N |
1115 | spin_lock_irq(&conf->device_lock); |
1116 | bio_list_merge(&conf->pending_bio_list, &plug->pending); | |
1117 | conf->pending_count += plug->pending_cnt; | |
1118 | spin_unlock_irq(&conf->device_lock); | |
ee0b0244 | 1119 | wake_up(&conf->wait_barrier); |
57c67df4 N |
1120 | md_wakeup_thread(mddev->thread); |
1121 | kfree(plug); | |
1122 | return; | |
1123 | } | |
1124 | ||
1125 | /* we aren't scheduling, so we can do the write-out directly. */ | |
1126 | bio = bio_list_get(&plug->pending); | |
1127 | bitmap_unplug(mddev->bitmap); | |
1128 | wake_up(&conf->wait_barrier); | |
1129 | ||
1130 | while (bio) { /* submit pending writes */ | |
1131 | struct bio *next = bio->bi_next; | |
1132 | bio->bi_next = NULL; | |
32f9f570 SL |
1133 | if (unlikely((bio->bi_rw & REQ_DISCARD) && |
1134 | !blk_queue_discard(bdev_get_queue(bio->bi_bdev)))) | |
1135 | /* Just ignore it */ | |
1136 | bio_endio(bio, 0); | |
1137 | else | |
1138 | generic_make_request(bio); | |
57c67df4 N |
1139 | bio = next; |
1140 | } | |
1141 | kfree(plug); | |
1142 | } | |
1143 | ||
20d0189b | 1144 | static void __make_request(struct mddev *mddev, struct bio *bio) |
1da177e4 | 1145 | { |
e879a879 | 1146 | struct r10conf *conf = mddev->private; |
9f2c9d12 | 1147 | struct r10bio *r10_bio; |
1da177e4 LT |
1148 | struct bio *read_bio; |
1149 | int i; | |
a362357b | 1150 | const int rw = bio_data_dir(bio); |
2c7d46ec | 1151 | const unsigned long do_sync = (bio->bi_rw & REQ_SYNC); |
e9c7469b | 1152 | const unsigned long do_fua = (bio->bi_rw & REQ_FUA); |
532a2a3f SL |
1153 | const unsigned long do_discard = (bio->bi_rw |
1154 | & (REQ_DISCARD | REQ_SECURE)); | |
c8dc9c65 | 1155 | const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME); |
6cce3b23 | 1156 | unsigned long flags; |
3cb03002 | 1157 | struct md_rdev *blocked_rdev; |
57c67df4 N |
1158 | struct blk_plug_cb *cb; |
1159 | struct raid10_plug_cb *plug = NULL; | |
d4432c23 N |
1160 | int sectors_handled; |
1161 | int max_sectors; | |
3ea7daa5 | 1162 | int sectors; |
1da177e4 | 1163 | |
cc13b1d1 N |
1164 | /* |
1165 | * Register the new request and wait if the reconstruction | |
1166 | * thread has put up a bar for new requests. | |
1167 | * Continue immediately if no resync is active currently. | |
1168 | */ | |
1169 | wait_barrier(conf); | |
1170 | ||
aa8b57aa | 1171 | sectors = bio_sectors(bio); |
3ea7daa5 | 1172 | while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) && |
4f024f37 KO |
1173 | bio->bi_iter.bi_sector < conf->reshape_progress && |
1174 | bio->bi_iter.bi_sector + sectors > conf->reshape_progress) { | |
3ea7daa5 N |
1175 | /* IO spans the reshape position. Need to wait for |
1176 | * reshape to pass | |
1177 | */ | |
1178 | allow_barrier(conf); | |
1179 | wait_event(conf->wait_barrier, | |
4f024f37 KO |
1180 | conf->reshape_progress <= bio->bi_iter.bi_sector || |
1181 | conf->reshape_progress >= bio->bi_iter.bi_sector + | |
1182 | sectors); | |
3ea7daa5 N |
1183 | wait_barrier(conf); |
1184 | } | |
1185 | if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) && | |
1186 | bio_data_dir(bio) == WRITE && | |
1187 | (mddev->reshape_backwards | |
4f024f37 KO |
1188 | ? (bio->bi_iter.bi_sector < conf->reshape_safe && |
1189 | bio->bi_iter.bi_sector + sectors > conf->reshape_progress) | |
1190 | : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe && | |
1191 | bio->bi_iter.bi_sector < conf->reshape_progress))) { | |
3ea7daa5 N |
1192 | /* Need to update reshape_position in metadata */ |
1193 | mddev->reshape_position = conf->reshape_progress; | |
1194 | set_bit(MD_CHANGE_DEVS, &mddev->flags); | |
1195 | set_bit(MD_CHANGE_PENDING, &mddev->flags); | |
1196 | md_wakeup_thread(mddev->thread); | |
1197 | wait_event(mddev->sb_wait, | |
1198 | !test_bit(MD_CHANGE_PENDING, &mddev->flags)); | |
1199 | ||
1200 | conf->reshape_safe = mddev->reshape_position; | |
1201 | } | |
1202 | ||
1da177e4 LT |
1203 | r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO); |
1204 | ||
1205 | r10_bio->master_bio = bio; | |
3ea7daa5 | 1206 | r10_bio->sectors = sectors; |
1da177e4 LT |
1207 | |
1208 | r10_bio->mddev = mddev; | |
4f024f37 | 1209 | r10_bio->sector = bio->bi_iter.bi_sector; |
6cce3b23 | 1210 | r10_bio->state = 0; |
1da177e4 | 1211 | |
856e08e2 N |
1212 | /* We might need to issue multiple reads to different |
1213 | * devices if there are bad blocks around, so we keep | |
1214 | * track of the number of reads in bio->bi_phys_segments. | |
1215 | * If this is 0, there is only one r10_bio and no locking | |
1216 | * will be needed when the request completes. If it is | |
1217 | * non-zero, then it is the number of not-completed requests. | |
1218 | */ | |
1219 | bio->bi_phys_segments = 0; | |
1220 | clear_bit(BIO_SEG_VALID, &bio->bi_flags); | |
1221 | ||
a362357b | 1222 | if (rw == READ) { |
1da177e4 LT |
1223 | /* |
1224 | * read balancing logic: | |
1225 | */ | |
96c3fd1f | 1226 | struct md_rdev *rdev; |
856e08e2 N |
1227 | int slot; |
1228 | ||
1229 | read_again: | |
96c3fd1f N |
1230 | rdev = read_balance(conf, r10_bio, &max_sectors); |
1231 | if (!rdev) { | |
1da177e4 | 1232 | raid_end_bio_io(r10_bio); |
5a7bbad2 | 1233 | return; |
1da177e4 | 1234 | } |
96c3fd1f | 1235 | slot = r10_bio->read_slot; |
1da177e4 | 1236 | |
a167f663 | 1237 | read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev); |
4f024f37 | 1238 | bio_trim(read_bio, r10_bio->sector - bio->bi_iter.bi_sector, |
6678d83f | 1239 | max_sectors); |
1da177e4 LT |
1240 | |
1241 | r10_bio->devs[slot].bio = read_bio; | |
abbf098e | 1242 | r10_bio->devs[slot].rdev = rdev; |
1da177e4 | 1243 | |
4f024f37 | 1244 | read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr + |
f8c9e74f | 1245 | choose_data_offset(r10_bio, rdev); |
96c3fd1f | 1246 | read_bio->bi_bdev = rdev->bdev; |
1da177e4 | 1247 | read_bio->bi_end_io = raid10_end_read_request; |
7b6d91da | 1248 | read_bio->bi_rw = READ | do_sync; |
1da177e4 LT |
1249 | read_bio->bi_private = r10_bio; |
1250 | ||
856e08e2 N |
1251 | if (max_sectors < r10_bio->sectors) { |
1252 | /* Could not read all from this device, so we will | |
1253 | * need another r10_bio. | |
1254 | */ | |
b50c259e | 1255 | sectors_handled = (r10_bio->sector + max_sectors |
4f024f37 | 1256 | - bio->bi_iter.bi_sector); |
856e08e2 N |
1257 | r10_bio->sectors = max_sectors; |
1258 | spin_lock_irq(&conf->device_lock); | |
1259 | if (bio->bi_phys_segments == 0) | |
1260 | bio->bi_phys_segments = 2; | |
1261 | else | |
1262 | bio->bi_phys_segments++; | |
b50c259e | 1263 | spin_unlock_irq(&conf->device_lock); |
856e08e2 N |
1264 | /* Cannot call generic_make_request directly |
1265 | * as that will be queued in __generic_make_request | |
1266 | * and subsequent mempool_alloc might block | |
1267 | * waiting for it. so hand bio over to raid10d. | |
1268 | */ | |
1269 | reschedule_retry(r10_bio); | |
1270 | ||
1271 | r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO); | |
1272 | ||
1273 | r10_bio->master_bio = bio; | |
aa8b57aa | 1274 | r10_bio->sectors = bio_sectors(bio) - sectors_handled; |
856e08e2 N |
1275 | r10_bio->state = 0; |
1276 | r10_bio->mddev = mddev; | |
4f024f37 KO |
1277 | r10_bio->sector = bio->bi_iter.bi_sector + |
1278 | sectors_handled; | |
856e08e2 N |
1279 | goto read_again; |
1280 | } else | |
1281 | generic_make_request(read_bio); | |
5a7bbad2 | 1282 | return; |
1da177e4 LT |
1283 | } |
1284 | ||
1285 | /* | |
1286 | * WRITE: | |
1287 | */ | |
34db0cd6 N |
1288 | if (conf->pending_count >= max_queued_requests) { |
1289 | md_wakeup_thread(mddev->thread); | |
1290 | wait_event(conf->wait_barrier, | |
1291 | conf->pending_count < max_queued_requests); | |
1292 | } | |
6bfe0b49 | 1293 | /* first select target devices under rcu_lock and |
1da177e4 LT |
1294 | * inc refcount on their rdev. Record them by setting |
1295 | * bios[x] to bio | |
d4432c23 N |
1296 | * If there are known/acknowledged bad blocks on any device |
1297 | * on which we have seen a write error, we want to avoid | |
1298 | * writing to those blocks. This potentially requires several | |
1299 | * writes to write around the bad blocks. Each set of writes | |
1300 | * gets its own r10_bio with a set of bios attached. The number | |
1301 | * of r10_bios is recored in bio->bi_phys_segments just as with | |
1302 | * the read case. | |
1da177e4 | 1303 | */ |
c3b328ac | 1304 | |
69335ef3 | 1305 | r10_bio->read_slot = -1; /* make sure repl_bio gets freed */ |
1da177e4 | 1306 | raid10_find_phys(conf, r10_bio); |
d4432c23 | 1307 | retry_write: |
cb6969e8 | 1308 | blocked_rdev = NULL; |
1da177e4 | 1309 | rcu_read_lock(); |
d4432c23 N |
1310 | max_sectors = r10_bio->sectors; |
1311 | ||
1da177e4 LT |
1312 | for (i = 0; i < conf->copies; i++) { |
1313 | int d = r10_bio->devs[i].devnum; | |
3cb03002 | 1314 | struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev); |
475b0321 N |
1315 | struct md_rdev *rrdev = rcu_dereference( |
1316 | conf->mirrors[d].replacement); | |
4ca40c2c N |
1317 | if (rdev == rrdev) |
1318 | rrdev = NULL; | |
6bfe0b49 DW |
1319 | if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) { |
1320 | atomic_inc(&rdev->nr_pending); | |
1321 | blocked_rdev = rdev; | |
1322 | break; | |
1323 | } | |
475b0321 N |
1324 | if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) { |
1325 | atomic_inc(&rrdev->nr_pending); | |
1326 | blocked_rdev = rrdev; | |
1327 | break; | |
1328 | } | |
e7c0c3fa N |
1329 | if (rdev && (test_bit(Faulty, &rdev->flags) |
1330 | || test_bit(Unmerged, &rdev->flags))) | |
1331 | rdev = NULL; | |
050b6615 N |
1332 | if (rrdev && (test_bit(Faulty, &rrdev->flags) |
1333 | || test_bit(Unmerged, &rrdev->flags))) | |
475b0321 N |
1334 | rrdev = NULL; |
1335 | ||
d4432c23 | 1336 | r10_bio->devs[i].bio = NULL; |
475b0321 | 1337 | r10_bio->devs[i].repl_bio = NULL; |
e7c0c3fa N |
1338 | |
1339 | if (!rdev && !rrdev) { | |
6cce3b23 | 1340 | set_bit(R10BIO_Degraded, &r10_bio->state); |
d4432c23 N |
1341 | continue; |
1342 | } | |
e7c0c3fa | 1343 | if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) { |
d4432c23 N |
1344 | sector_t first_bad; |
1345 | sector_t dev_sector = r10_bio->devs[i].addr; | |
1346 | int bad_sectors; | |
1347 | int is_bad; | |
1348 | ||
1349 | is_bad = is_badblock(rdev, dev_sector, | |
1350 | max_sectors, | |
1351 | &first_bad, &bad_sectors); | |
1352 | if (is_bad < 0) { | |
1353 | /* Mustn't write here until the bad block | |
1354 | * is acknowledged | |
1355 | */ | |
1356 | atomic_inc(&rdev->nr_pending); | |
1357 | set_bit(BlockedBadBlocks, &rdev->flags); | |
1358 | blocked_rdev = rdev; | |
1359 | break; | |
1360 | } | |
1361 | if (is_bad && first_bad <= dev_sector) { | |
1362 | /* Cannot write here at all */ | |
1363 | bad_sectors -= (dev_sector - first_bad); | |
1364 | if (bad_sectors < max_sectors) | |
1365 | /* Mustn't write more than bad_sectors | |
1366 | * to other devices yet | |
1367 | */ | |
1368 | max_sectors = bad_sectors; | |
1369 | /* We don't set R10BIO_Degraded as that | |
1370 | * only applies if the disk is missing, | |
1371 | * so it might be re-added, and we want to | |
1372 | * know to recover this chunk. | |
1373 | * In this case the device is here, and the | |
1374 | * fact that this chunk is not in-sync is | |
1375 | * recorded in the bad block log. | |
1376 | */ | |
1377 | continue; | |
1378 | } | |
1379 | if (is_bad) { | |
1380 | int good_sectors = first_bad - dev_sector; | |
1381 | if (good_sectors < max_sectors) | |
1382 | max_sectors = good_sectors; | |
1383 | } | |
6cce3b23 | 1384 | } |
e7c0c3fa N |
1385 | if (rdev) { |
1386 | r10_bio->devs[i].bio = bio; | |
1387 | atomic_inc(&rdev->nr_pending); | |
1388 | } | |
475b0321 N |
1389 | if (rrdev) { |
1390 | r10_bio->devs[i].repl_bio = bio; | |
1391 | atomic_inc(&rrdev->nr_pending); | |
1392 | } | |
1da177e4 LT |
1393 | } |
1394 | rcu_read_unlock(); | |
1395 | ||
6bfe0b49 DW |
1396 | if (unlikely(blocked_rdev)) { |
1397 | /* Have to wait for this device to get unblocked, then retry */ | |
1398 | int j; | |
1399 | int d; | |
1400 | ||
475b0321 | 1401 | for (j = 0; j < i; j++) { |
6bfe0b49 DW |
1402 | if (r10_bio->devs[j].bio) { |
1403 | d = r10_bio->devs[j].devnum; | |
1404 | rdev_dec_pending(conf->mirrors[d].rdev, mddev); | |
1405 | } | |
475b0321 | 1406 | if (r10_bio->devs[j].repl_bio) { |
4ca40c2c | 1407 | struct md_rdev *rdev; |
475b0321 | 1408 | d = r10_bio->devs[j].devnum; |
4ca40c2c N |
1409 | rdev = conf->mirrors[d].replacement; |
1410 | if (!rdev) { | |
1411 | /* Race with remove_disk */ | |
1412 | smp_mb(); | |
1413 | rdev = conf->mirrors[d].rdev; | |
1414 | } | |
1415 | rdev_dec_pending(rdev, mddev); | |
475b0321 N |
1416 | } |
1417 | } | |
6bfe0b49 DW |
1418 | allow_barrier(conf); |
1419 | md_wait_for_blocked_rdev(blocked_rdev, mddev); | |
1420 | wait_barrier(conf); | |
1421 | goto retry_write; | |
1422 | } | |
1423 | ||
d4432c23 N |
1424 | if (max_sectors < r10_bio->sectors) { |
1425 | /* We are splitting this into multiple parts, so | |
1426 | * we need to prepare for allocating another r10_bio. | |
1427 | */ | |
1428 | r10_bio->sectors = max_sectors; | |
1429 | spin_lock_irq(&conf->device_lock); | |
1430 | if (bio->bi_phys_segments == 0) | |
1431 | bio->bi_phys_segments = 2; | |
1432 | else | |
1433 | bio->bi_phys_segments++; | |
1434 | spin_unlock_irq(&conf->device_lock); | |
1435 | } | |
4f024f37 KO |
1436 | sectors_handled = r10_bio->sector + max_sectors - |
1437 | bio->bi_iter.bi_sector; | |
d4432c23 | 1438 | |
4e78064f | 1439 | atomic_set(&r10_bio->remaining, 1); |
d4432c23 | 1440 | bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0); |
06d91a5f | 1441 | |
1da177e4 LT |
1442 | for (i = 0; i < conf->copies; i++) { |
1443 | struct bio *mbio; | |
1444 | int d = r10_bio->devs[i].devnum; | |
e7c0c3fa N |
1445 | if (r10_bio->devs[i].bio) { |
1446 | struct md_rdev *rdev = conf->mirrors[d].rdev; | |
1447 | mbio = bio_clone_mddev(bio, GFP_NOIO, mddev); | |
4f024f37 | 1448 | bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector, |
6678d83f | 1449 | max_sectors); |
e7c0c3fa N |
1450 | r10_bio->devs[i].bio = mbio; |
1451 | ||
4f024f37 | 1452 | mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+ |
e7c0c3fa N |
1453 | choose_data_offset(r10_bio, |
1454 | rdev)); | |
1455 | mbio->bi_bdev = rdev->bdev; | |
1456 | mbio->bi_end_io = raid10_end_write_request; | |
c8dc9c65 JL |
1457 | mbio->bi_rw = |
1458 | WRITE | do_sync | do_fua | do_discard | do_same; | |
e7c0c3fa N |
1459 | mbio->bi_private = r10_bio; |
1460 | ||
1461 | atomic_inc(&r10_bio->remaining); | |
1462 | ||
1463 | cb = blk_check_plugged(raid10_unplug, mddev, | |
1464 | sizeof(*plug)); | |
1465 | if (cb) | |
1466 | plug = container_of(cb, struct raid10_plug_cb, | |
1467 | cb); | |
1468 | else | |
1469 | plug = NULL; | |
1470 | spin_lock_irqsave(&conf->device_lock, flags); | |
1471 | if (plug) { | |
1472 | bio_list_add(&plug->pending, mbio); | |
1473 | plug->pending_cnt++; | |
1474 | } else { | |
1475 | bio_list_add(&conf->pending_bio_list, mbio); | |
1476 | conf->pending_count++; | |
1477 | } | |
1478 | spin_unlock_irqrestore(&conf->device_lock, flags); | |
1479 | if (!plug) | |
1480 | md_wakeup_thread(mddev->thread); | |
1481 | } | |
57c67df4 | 1482 | |
e7c0c3fa N |
1483 | if (r10_bio->devs[i].repl_bio) { |
1484 | struct md_rdev *rdev = conf->mirrors[d].replacement; | |
1485 | if (rdev == NULL) { | |
1486 | /* Replacement just got moved to main 'rdev' */ | |
1487 | smp_mb(); | |
1488 | rdev = conf->mirrors[d].rdev; | |
1489 | } | |
1490 | mbio = bio_clone_mddev(bio, GFP_NOIO, mddev); | |
4f024f37 | 1491 | bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector, |
6678d83f | 1492 | max_sectors); |
e7c0c3fa N |
1493 | r10_bio->devs[i].repl_bio = mbio; |
1494 | ||
4f024f37 | 1495 | mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr + |
e7c0c3fa N |
1496 | choose_data_offset( |
1497 | r10_bio, rdev)); | |
1498 | mbio->bi_bdev = rdev->bdev; | |
1499 | mbio->bi_end_io = raid10_end_write_request; | |
c8dc9c65 JL |
1500 | mbio->bi_rw = |
1501 | WRITE | do_sync | do_fua | do_discard | do_same; | |
e7c0c3fa N |
1502 | mbio->bi_private = r10_bio; |
1503 | ||
1504 | atomic_inc(&r10_bio->remaining); | |
1505 | spin_lock_irqsave(&conf->device_lock, flags); | |
57c67df4 N |
1506 | bio_list_add(&conf->pending_bio_list, mbio); |
1507 | conf->pending_count++; | |
e7c0c3fa N |
1508 | spin_unlock_irqrestore(&conf->device_lock, flags); |
1509 | if (!mddev_check_plugged(mddev)) | |
1510 | md_wakeup_thread(mddev->thread); | |
57c67df4 | 1511 | } |
1da177e4 LT |
1512 | } |
1513 | ||
079fa166 N |
1514 | /* Don't remove the bias on 'remaining' (one_write_done) until |
1515 | * after checking if we need to go around again. | |
1516 | */ | |
a35e63ef | 1517 | |
aa8b57aa | 1518 | if (sectors_handled < bio_sectors(bio)) { |
079fa166 | 1519 | one_write_done(r10_bio); |
5e570289 | 1520 | /* We need another r10_bio. It has already been counted |
d4432c23 N |
1521 | * in bio->bi_phys_segments. |
1522 | */ | |
1523 | r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO); | |
1524 | ||
1525 | r10_bio->master_bio = bio; | |
aa8b57aa | 1526 | r10_bio->sectors = bio_sectors(bio) - sectors_handled; |
d4432c23 N |
1527 | |
1528 | r10_bio->mddev = mddev; | |
4f024f37 | 1529 | r10_bio->sector = bio->bi_iter.bi_sector + sectors_handled; |
d4432c23 N |
1530 | r10_bio->state = 0; |
1531 | goto retry_write; | |
1532 | } | |
079fa166 | 1533 | one_write_done(r10_bio); |
20d0189b KO |
1534 | } |
1535 | ||
1536 | static void make_request(struct mddev *mddev, struct bio *bio) | |
1537 | { | |
1538 | struct r10conf *conf = mddev->private; | |
1539 | sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask); | |
1540 | int chunk_sects = chunk_mask + 1; | |
1541 | ||
1542 | struct bio *split; | |
1543 | ||
1544 | if (unlikely(bio->bi_rw & REQ_FLUSH)) { | |
1545 | md_flush_request(mddev, bio); | |
1546 | return; | |
1547 | } | |
1548 | ||
1549 | md_write_start(mddev, bio); | |
1550 | ||
20d0189b KO |
1551 | do { |
1552 | ||
1553 | /* | |
1554 | * If this request crosses a chunk boundary, we need to split | |
1555 | * it. | |
1556 | */ | |
1557 | if (unlikely((bio->bi_iter.bi_sector & chunk_mask) + | |
1558 | bio_sectors(bio) > chunk_sects | |
1559 | && (conf->geo.near_copies < conf->geo.raid_disks | |
1560 | || conf->prev.near_copies < | |
1561 | conf->prev.raid_disks))) { | |
1562 | split = bio_split(bio, chunk_sects - | |
1563 | (bio->bi_iter.bi_sector & | |
1564 | (chunk_sects - 1)), | |
1565 | GFP_NOIO, fs_bio_set); | |
1566 | bio_chain(split, bio); | |
1567 | } else { | |
1568 | split = bio; | |
1569 | } | |
1570 | ||
1571 | __make_request(mddev, split); | |
1572 | } while (split != bio); | |
079fa166 N |
1573 | |
1574 | /* In case raid10d snuck in to freeze_array */ | |
1575 | wake_up(&conf->wait_barrier); | |
1da177e4 LT |
1576 | } |
1577 | ||
fd01b88c | 1578 | static void status(struct seq_file *seq, struct mddev *mddev) |
1da177e4 | 1579 | { |
e879a879 | 1580 | struct r10conf *conf = mddev->private; |
1da177e4 LT |
1581 | int i; |
1582 | ||
5cf00fcd | 1583 | if (conf->geo.near_copies < conf->geo.raid_disks) |
9d8f0363 | 1584 | seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2); |
5cf00fcd N |
1585 | if (conf->geo.near_copies > 1) |
1586 | seq_printf(seq, " %d near-copies", conf->geo.near_copies); | |
1587 | if (conf->geo.far_copies > 1) { | |
1588 | if (conf->geo.far_offset) | |
1589 | seq_printf(seq, " %d offset-copies", conf->geo.far_copies); | |
c93983bf | 1590 | else |
5cf00fcd | 1591 | seq_printf(seq, " %d far-copies", conf->geo.far_copies); |
c93983bf | 1592 | } |
5cf00fcd N |
1593 | seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks, |
1594 | conf->geo.raid_disks - mddev->degraded); | |
1595 | for (i = 0; i < conf->geo.raid_disks; i++) | |
1da177e4 LT |
1596 | seq_printf(seq, "%s", |
1597 | conf->mirrors[i].rdev && | |
b2d444d7 | 1598 | test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_"); |
1da177e4 LT |
1599 | seq_printf(seq, "]"); |
1600 | } | |
1601 | ||
700c7213 N |
1602 | /* check if there are enough drives for |
1603 | * every block to appear on atleast one. | |
1604 | * Don't consider the device numbered 'ignore' | |
1605 | * as we might be about to remove it. | |
1606 | */ | |
635f6416 | 1607 | static int _enough(struct r10conf *conf, int previous, int ignore) |
700c7213 N |
1608 | { |
1609 | int first = 0; | |
725d6e57 | 1610 | int has_enough = 0; |
635f6416 N |
1611 | int disks, ncopies; |
1612 | if (previous) { | |
1613 | disks = conf->prev.raid_disks; | |
1614 | ncopies = conf->prev.near_copies; | |
1615 | } else { | |
1616 | disks = conf->geo.raid_disks; | |
1617 | ncopies = conf->geo.near_copies; | |
1618 | } | |
700c7213 | 1619 | |
725d6e57 | 1620 | rcu_read_lock(); |
700c7213 N |
1621 | do { |
1622 | int n = conf->copies; | |
1623 | int cnt = 0; | |
80b48124 | 1624 | int this = first; |
700c7213 | 1625 | while (n--) { |
725d6e57 N |
1626 | struct md_rdev *rdev; |
1627 | if (this != ignore && | |
1628 | (rdev = rcu_dereference(conf->mirrors[this].rdev)) && | |
1629 | test_bit(In_sync, &rdev->flags)) | |
700c7213 | 1630 | cnt++; |
635f6416 | 1631 | this = (this+1) % disks; |
700c7213 N |
1632 | } |
1633 | if (cnt == 0) | |
725d6e57 | 1634 | goto out; |
635f6416 | 1635 | first = (first + ncopies) % disks; |
700c7213 | 1636 | } while (first != 0); |
725d6e57 N |
1637 | has_enough = 1; |
1638 | out: | |
1639 | rcu_read_unlock(); | |
1640 | return has_enough; | |
700c7213 N |
1641 | } |
1642 | ||
f8c9e74f N |
1643 | static int enough(struct r10conf *conf, int ignore) |
1644 | { | |
635f6416 N |
1645 | /* when calling 'enough', both 'prev' and 'geo' must |
1646 | * be stable. | |
1647 | * This is ensured if ->reconfig_mutex or ->device_lock | |
1648 | * is held. | |
1649 | */ | |
1650 | return _enough(conf, 0, ignore) && | |
1651 | _enough(conf, 1, ignore); | |
f8c9e74f N |
1652 | } |
1653 | ||
fd01b88c | 1654 | static void error(struct mddev *mddev, struct md_rdev *rdev) |
1da177e4 LT |
1655 | { |
1656 | char b[BDEVNAME_SIZE]; | |
e879a879 | 1657 | struct r10conf *conf = mddev->private; |
635f6416 | 1658 | unsigned long flags; |
1da177e4 LT |
1659 | |
1660 | /* | |
1661 | * If it is not operational, then we have already marked it as dead | |
1662 | * else if it is the last working disks, ignore the error, let the | |
1663 | * next level up know. | |
1664 | * else mark the drive as failed | |
1665 | */ | |
635f6416 | 1666 | spin_lock_irqsave(&conf->device_lock, flags); |
b2d444d7 | 1667 | if (test_bit(In_sync, &rdev->flags) |
635f6416 | 1668 | && !enough(conf, rdev->raid_disk)) { |
1da177e4 LT |
1669 | /* |
1670 | * Don't fail the drive, just return an IO error. | |
1da177e4 | 1671 | */ |
635f6416 | 1672 | spin_unlock_irqrestore(&conf->device_lock, flags); |
1da177e4 | 1673 | return; |
635f6416 | 1674 | } |
2446dba0 | 1675 | if (test_and_clear_bit(In_sync, &rdev->flags)) |
1da177e4 | 1676 | mddev->degraded++; |
2446dba0 N |
1677 | /* |
1678 | * If recovery is running, make sure it aborts. | |
1679 | */ | |
1680 | set_bit(MD_RECOVERY_INTR, &mddev->recovery); | |
de393cde | 1681 | set_bit(Blocked, &rdev->flags); |
b2d444d7 | 1682 | set_bit(Faulty, &rdev->flags); |
850b2b42 | 1683 | set_bit(MD_CHANGE_DEVS, &mddev->flags); |
635f6416 | 1684 | spin_unlock_irqrestore(&conf->device_lock, flags); |
067032bc JP |
1685 | printk(KERN_ALERT |
1686 | "md/raid10:%s: Disk failure on %s, disabling device.\n" | |
1687 | "md/raid10:%s: Operation continuing on %d devices.\n", | |
128595ed | 1688 | mdname(mddev), bdevname(rdev->bdev, b), |
5cf00fcd | 1689 | mdname(mddev), conf->geo.raid_disks - mddev->degraded); |
1da177e4 LT |
1690 | } |
1691 | ||
e879a879 | 1692 | static void print_conf(struct r10conf *conf) |
1da177e4 LT |
1693 | { |
1694 | int i; | |
dc280d98 | 1695 | struct raid10_info *tmp; |
1da177e4 | 1696 | |
128595ed | 1697 | printk(KERN_DEBUG "RAID10 conf printout:\n"); |
1da177e4 | 1698 | if (!conf) { |
128595ed | 1699 | printk(KERN_DEBUG "(!conf)\n"); |
1da177e4 LT |
1700 | return; |
1701 | } | |
5cf00fcd N |
1702 | printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded, |
1703 | conf->geo.raid_disks); | |
1da177e4 | 1704 | |
5cf00fcd | 1705 | for (i = 0; i < conf->geo.raid_disks; i++) { |
1da177e4 LT |
1706 | char b[BDEVNAME_SIZE]; |
1707 | tmp = conf->mirrors + i; | |
1708 | if (tmp->rdev) | |
128595ed | 1709 | printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n", |
b2d444d7 N |
1710 | i, !test_bit(In_sync, &tmp->rdev->flags), |
1711 | !test_bit(Faulty, &tmp->rdev->flags), | |
1da177e4 LT |
1712 | bdevname(tmp->rdev->bdev,b)); |
1713 | } | |
1714 | } | |
1715 | ||
e879a879 | 1716 | static void close_sync(struct r10conf *conf) |
1da177e4 | 1717 | { |
0a27ec96 N |
1718 | wait_barrier(conf); |
1719 | allow_barrier(conf); | |
1da177e4 LT |
1720 | |
1721 | mempool_destroy(conf->r10buf_pool); | |
1722 | conf->r10buf_pool = NULL; | |
1723 | } | |
1724 | ||
fd01b88c | 1725 | static int raid10_spare_active(struct mddev *mddev) |
1da177e4 LT |
1726 | { |
1727 | int i; | |
e879a879 | 1728 | struct r10conf *conf = mddev->private; |
dc280d98 | 1729 | struct raid10_info *tmp; |
6b965620 N |
1730 | int count = 0; |
1731 | unsigned long flags; | |
1da177e4 LT |
1732 | |
1733 | /* | |
1734 | * Find all non-in_sync disks within the RAID10 configuration | |
1735 | * and mark them in_sync | |
1736 | */ | |
5cf00fcd | 1737 | for (i = 0; i < conf->geo.raid_disks; i++) { |
1da177e4 | 1738 | tmp = conf->mirrors + i; |
4ca40c2c N |
1739 | if (tmp->replacement |
1740 | && tmp->replacement->recovery_offset == MaxSector | |
1741 | && !test_bit(Faulty, &tmp->replacement->flags) | |
1742 | && !test_and_set_bit(In_sync, &tmp->replacement->flags)) { | |
1743 | /* Replacement has just become active */ | |
1744 | if (!tmp->rdev | |
1745 | || !test_and_clear_bit(In_sync, &tmp->rdev->flags)) | |
1746 | count++; | |
1747 | if (tmp->rdev) { | |
1748 | /* Replaced device not technically faulty, | |
1749 | * but we need to be sure it gets removed | |
1750 | * and never re-added. | |
1751 | */ | |
1752 | set_bit(Faulty, &tmp->rdev->flags); | |
1753 | sysfs_notify_dirent_safe( | |
1754 | tmp->rdev->sysfs_state); | |
1755 | } | |
1756 | sysfs_notify_dirent_safe(tmp->replacement->sysfs_state); | |
1757 | } else if (tmp->rdev | |
61e4947c | 1758 | && tmp->rdev->recovery_offset == MaxSector |
4ca40c2c N |
1759 | && !test_bit(Faulty, &tmp->rdev->flags) |
1760 | && !test_and_set_bit(In_sync, &tmp->rdev->flags)) { | |
6b965620 | 1761 | count++; |
2863b9eb | 1762 | sysfs_notify_dirent_safe(tmp->rdev->sysfs_state); |
1da177e4 LT |
1763 | } |
1764 | } | |
6b965620 N |
1765 | spin_lock_irqsave(&conf->device_lock, flags); |
1766 | mddev->degraded -= count; | |
1767 | spin_unlock_irqrestore(&conf->device_lock, flags); | |
1da177e4 LT |
1768 | |
1769 | print_conf(conf); | |
6b965620 | 1770 | return count; |
1da177e4 LT |
1771 | } |
1772 | ||
fd01b88c | 1773 | static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev) |
1da177e4 | 1774 | { |
e879a879 | 1775 | struct r10conf *conf = mddev->private; |
199050ea | 1776 | int err = -EEXIST; |
1da177e4 | 1777 | int mirror; |
6c2fce2e | 1778 | int first = 0; |
5cf00fcd | 1779 | int last = conf->geo.raid_disks - 1; |
050b6615 | 1780 | struct request_queue *q = bdev_get_queue(rdev->bdev); |
1da177e4 LT |
1781 | |
1782 | if (mddev->recovery_cp < MaxSector) | |
1783 | /* only hot-add to in-sync arrays, as recovery is | |
1784 | * very different from resync | |
1785 | */ | |
199050ea | 1786 | return -EBUSY; |
635f6416 | 1787 | if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1)) |
199050ea | 1788 | return -EINVAL; |
1da177e4 | 1789 | |
a53a6c85 | 1790 | if (rdev->raid_disk >= 0) |
6c2fce2e | 1791 | first = last = rdev->raid_disk; |
1da177e4 | 1792 | |
050b6615 N |
1793 | if (q->merge_bvec_fn) { |
1794 | set_bit(Unmerged, &rdev->flags); | |
1795 | mddev->merge_check_needed = 1; | |
1796 | } | |
1797 | ||
2c4193df | 1798 | if (rdev->saved_raid_disk >= first && |
6cce3b23 N |
1799 | conf->mirrors[rdev->saved_raid_disk].rdev == NULL) |
1800 | mirror = rdev->saved_raid_disk; | |
1801 | else | |
6c2fce2e | 1802 | mirror = first; |
2bb77736 | 1803 | for ( ; mirror <= last ; mirror++) { |
dc280d98 | 1804 | struct raid10_info *p = &conf->mirrors[mirror]; |
2bb77736 N |
1805 | if (p->recovery_disabled == mddev->recovery_disabled) |
1806 | continue; | |
b7044d41 N |
1807 | if (p->rdev) { |
1808 | if (!test_bit(WantReplacement, &p->rdev->flags) || | |
1809 | p->replacement != NULL) | |
1810 | continue; | |
1811 | clear_bit(In_sync, &rdev->flags); | |
1812 | set_bit(Replacement, &rdev->flags); | |
1813 | rdev->raid_disk = mirror; | |
1814 | err = 0; | |
9092c02d JB |
1815 | if (mddev->gendisk) |
1816 | disk_stack_limits(mddev->gendisk, rdev->bdev, | |
1817 | rdev->data_offset << 9); | |
b7044d41 N |
1818 | conf->fullsync = 1; |
1819 | rcu_assign_pointer(p->replacement, rdev); | |
1820 | break; | |
1821 | } | |
1da177e4 | 1822 | |
9092c02d JB |
1823 | if (mddev->gendisk) |
1824 | disk_stack_limits(mddev->gendisk, rdev->bdev, | |
1825 | rdev->data_offset << 9); | |
1da177e4 | 1826 | |
2bb77736 | 1827 | p->head_position = 0; |
d890fa2b | 1828 | p->recovery_disabled = mddev->recovery_disabled - 1; |
2bb77736 N |
1829 | rdev->raid_disk = mirror; |
1830 | err = 0; | |
1831 | if (rdev->saved_raid_disk != mirror) | |
1832 | conf->fullsync = 1; | |
1833 | rcu_assign_pointer(p->rdev, rdev); | |
1834 | break; | |
1835 | } | |
050b6615 N |
1836 | if (err == 0 && test_bit(Unmerged, &rdev->flags)) { |
1837 | /* Some requests might not have seen this new | |
1838 | * merge_bvec_fn. We must wait for them to complete | |
1839 | * before merging the device fully. | |
1840 | * First we make sure any code which has tested | |
1841 | * our function has submitted the request, then | |
1842 | * we wait for all outstanding requests to complete. | |
1843 | */ | |
1844 | synchronize_sched(); | |
e2d59925 N |
1845 | freeze_array(conf, 0); |
1846 | unfreeze_array(conf); | |
050b6615 N |
1847 | clear_bit(Unmerged, &rdev->flags); |
1848 | } | |
ac5e7113 | 1849 | md_integrity_add_rdev(rdev, mddev); |
ed30be07 | 1850 | if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev))) |
532a2a3f SL |
1851 | queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue); |
1852 | ||
1da177e4 | 1853 | print_conf(conf); |
199050ea | 1854 | return err; |
1da177e4 LT |
1855 | } |
1856 | ||
b8321b68 | 1857 | static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev) |
1da177e4 | 1858 | { |
e879a879 | 1859 | struct r10conf *conf = mddev->private; |
1da177e4 | 1860 | int err = 0; |
b8321b68 | 1861 | int number = rdev->raid_disk; |
c8ab903e | 1862 | struct md_rdev **rdevp; |
dc280d98 | 1863 | struct raid10_info *p = conf->mirrors + number; |
1da177e4 LT |
1864 | |
1865 | print_conf(conf); | |
c8ab903e N |
1866 | if (rdev == p->rdev) |
1867 | rdevp = &p->rdev; | |
1868 | else if (rdev == p->replacement) | |
1869 | rdevp = &p->replacement; | |
1870 | else | |
1871 | return 0; | |
1872 | ||
1873 | if (test_bit(In_sync, &rdev->flags) || | |
1874 | atomic_read(&rdev->nr_pending)) { | |
1875 | err = -EBUSY; | |
1876 | goto abort; | |
1877 | } | |
1878 | /* Only remove faulty devices if recovery | |
1879 | * is not possible. | |
1880 | */ | |
1881 | if (!test_bit(Faulty, &rdev->flags) && | |
1882 | mddev->recovery_disabled != p->recovery_disabled && | |
4ca40c2c | 1883 | (!p->replacement || p->replacement == rdev) && |
63aced61 | 1884 | number < conf->geo.raid_disks && |
c8ab903e N |
1885 | enough(conf, -1)) { |
1886 | err = -EBUSY; | |
1887 | goto abort; | |
1da177e4 | 1888 | } |
c8ab903e N |
1889 | *rdevp = NULL; |
1890 | synchronize_rcu(); | |
1891 | if (atomic_read(&rdev->nr_pending)) { | |
1892 | /* lost the race, try later */ | |
1893 | err = -EBUSY; | |
1894 | *rdevp = rdev; | |
1895 | goto abort; | |
4ca40c2c N |
1896 | } else if (p->replacement) { |
1897 | /* We must have just cleared 'rdev' */ | |
1898 | p->rdev = p->replacement; | |
1899 | clear_bit(Replacement, &p->replacement->flags); | |
1900 | smp_mb(); /* Make sure other CPUs may see both as identical | |
1901 | * but will never see neither -- if they are careful. | |
1902 | */ | |
1903 | p->replacement = NULL; | |
1904 | clear_bit(WantReplacement, &rdev->flags); | |
1905 | } else | |
1906 | /* We might have just remove the Replacement as faulty | |
1907 | * Clear the flag just in case | |
1908 | */ | |
1909 | clear_bit(WantReplacement, &rdev->flags); | |
1910 | ||
c8ab903e N |
1911 | err = md_integrity_register(mddev); |
1912 | ||
1da177e4 LT |
1913 | abort: |
1914 | ||
1915 | print_conf(conf); | |
1916 | return err; | |
1917 | } | |
1918 | ||
6712ecf8 | 1919 | static void end_sync_read(struct bio *bio, int error) |
1da177e4 | 1920 | { |
9f2c9d12 | 1921 | struct r10bio *r10_bio = bio->bi_private; |
e879a879 | 1922 | struct r10conf *conf = r10_bio->mddev->private; |
778ca018 | 1923 | int d; |
1da177e4 | 1924 | |
3ea7daa5 N |
1925 | if (bio == r10_bio->master_bio) { |
1926 | /* this is a reshape read */ | |
1927 | d = r10_bio->read_slot; /* really the read dev */ | |
1928 | } else | |
1929 | d = find_bio_disk(conf, r10_bio, bio, NULL, NULL); | |
0eb3ff12 N |
1930 | |
1931 | if (test_bit(BIO_UPTODATE, &bio->bi_flags)) | |
1932 | set_bit(R10BIO_Uptodate, &r10_bio->state); | |
e684e41d N |
1933 | else |
1934 | /* The write handler will notice the lack of | |
1935 | * R10BIO_Uptodate and record any errors etc | |
1936 | */ | |
4dbcdc75 N |
1937 | atomic_add(r10_bio->sectors, |
1938 | &conf->mirrors[d].rdev->corrected_errors); | |
1da177e4 LT |
1939 | |
1940 | /* for reconstruct, we always reschedule after a read. | |
1941 | * for resync, only after all reads | |
1942 | */ | |
73d5c38a | 1943 | rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev); |
1da177e4 LT |
1944 | if (test_bit(R10BIO_IsRecover, &r10_bio->state) || |
1945 | atomic_dec_and_test(&r10_bio->remaining)) { | |
1946 | /* we have read all the blocks, | |
1947 | * do the comparison in process context in raid10d | |
1948 | */ | |
1949 | reschedule_retry(r10_bio); | |
1950 | } | |
1da177e4 LT |
1951 | } |
1952 | ||
9f2c9d12 | 1953 | static void end_sync_request(struct r10bio *r10_bio) |
1da177e4 | 1954 | { |
fd01b88c | 1955 | struct mddev *mddev = r10_bio->mddev; |
dfc70645 | 1956 | |
1da177e4 LT |
1957 | while (atomic_dec_and_test(&r10_bio->remaining)) { |
1958 | if (r10_bio->master_bio == NULL) { | |
1959 | /* the primary of several recovery bios */ | |
73d5c38a | 1960 | sector_t s = r10_bio->sectors; |
1a0b7cd8 N |
1961 | if (test_bit(R10BIO_MadeGood, &r10_bio->state) || |
1962 | test_bit(R10BIO_WriteError, &r10_bio->state)) | |
749c55e9 N |
1963 | reschedule_retry(r10_bio); |
1964 | else | |
1965 | put_buf(r10_bio); | |
73d5c38a | 1966 | md_done_sync(mddev, s, 1); |
1da177e4 LT |
1967 | break; |
1968 | } else { | |
9f2c9d12 | 1969 | struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio; |
1a0b7cd8 N |
1970 | if (test_bit(R10BIO_MadeGood, &r10_bio->state) || |
1971 | test_bit(R10BIO_WriteError, &r10_bio->state)) | |
749c55e9 N |
1972 | reschedule_retry(r10_bio); |
1973 | else | |
1974 | put_buf(r10_bio); | |
1da177e4 LT |
1975 | r10_bio = r10_bio2; |
1976 | } | |
1977 | } | |
1da177e4 LT |
1978 | } |
1979 | ||
5e570289 N |
1980 | static void end_sync_write(struct bio *bio, int error) |
1981 | { | |
1982 | int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags); | |
9f2c9d12 | 1983 | struct r10bio *r10_bio = bio->bi_private; |
fd01b88c | 1984 | struct mddev *mddev = r10_bio->mddev; |
e879a879 | 1985 | struct r10conf *conf = mddev->private; |
5e570289 N |
1986 | int d; |
1987 | sector_t first_bad; | |
1988 | int bad_sectors; | |
1989 | int slot; | |
9ad1aefc | 1990 | int repl; |
4ca40c2c | 1991 | struct md_rdev *rdev = NULL; |
5e570289 | 1992 | |
9ad1aefc N |
1993 | d = find_bio_disk(conf, r10_bio, bio, &slot, &repl); |
1994 | if (repl) | |
1995 | rdev = conf->mirrors[d].replacement; | |
547414d1 | 1996 | else |
9ad1aefc | 1997 | rdev = conf->mirrors[d].rdev; |
5e570289 N |
1998 | |
1999 | if (!uptodate) { | |
9ad1aefc N |
2000 | if (repl) |
2001 | md_error(mddev, rdev); | |
2002 | else { | |
2003 | set_bit(WriteErrorSeen, &rdev->flags); | |
b7044d41 N |
2004 | if (!test_and_set_bit(WantReplacement, &rdev->flags)) |
2005 | set_bit(MD_RECOVERY_NEEDED, | |
2006 | &rdev->mddev->recovery); | |
9ad1aefc N |
2007 | set_bit(R10BIO_WriteError, &r10_bio->state); |
2008 | } | |
2009 | } else if (is_badblock(rdev, | |
5e570289 N |
2010 | r10_bio->devs[slot].addr, |
2011 | r10_bio->sectors, | |
2012 | &first_bad, &bad_sectors)) | |
2013 | set_bit(R10BIO_MadeGood, &r10_bio->state); | |
2014 | ||
9ad1aefc | 2015 | rdev_dec_pending(rdev, mddev); |
5e570289 N |
2016 | |
2017 | end_sync_request(r10_bio); | |
2018 | } | |
2019 | ||
1da177e4 LT |
2020 | /* |
2021 | * Note: sync and recover and handled very differently for raid10 | |
2022 | * This code is for resync. | |
2023 | * For resync, we read through virtual addresses and read all blocks. | |
2024 | * If there is any error, we schedule a write. The lowest numbered | |
2025 | * drive is authoritative. | |
2026 | * However requests come for physical address, so we need to map. | |
2027 | * For every physical address there are raid_disks/copies virtual addresses, | |
2028 | * which is always are least one, but is not necessarly an integer. | |
2029 | * This means that a physical address can span multiple chunks, so we may | |
2030 | * have to submit multiple io requests for a single sync request. | |
2031 | */ | |
2032 | /* | |
2033 | * We check if all blocks are in-sync and only write to blocks that | |
2034 | * aren't in sync | |
2035 | */ | |
9f2c9d12 | 2036 | static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio) |
1da177e4 | 2037 | { |
e879a879 | 2038 | struct r10conf *conf = mddev->private; |
1da177e4 LT |
2039 | int i, first; |
2040 | struct bio *tbio, *fbio; | |
f4380a91 | 2041 | int vcnt; |
1da177e4 LT |
2042 | |
2043 | atomic_set(&r10_bio->remaining, 1); | |
2044 | ||
2045 | /* find the first device with a block */ | |
2046 | for (i=0; i<conf->copies; i++) | |
2047 | if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) | |
2048 | break; | |
2049 | ||
2050 | if (i == conf->copies) | |
2051 | goto done; | |
2052 | ||
2053 | first = i; | |
2054 | fbio = r10_bio->devs[i].bio; | |
2055 | ||
f4380a91 | 2056 | vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9); |
1da177e4 | 2057 | /* now find blocks with errors */ |
0eb3ff12 N |
2058 | for (i=0 ; i < conf->copies ; i++) { |
2059 | int j, d; | |
1da177e4 | 2060 | |
1da177e4 | 2061 | tbio = r10_bio->devs[i].bio; |
0eb3ff12 N |
2062 | |
2063 | if (tbio->bi_end_io != end_sync_read) | |
2064 | continue; | |
2065 | if (i == first) | |
1da177e4 | 2066 | continue; |
0eb3ff12 N |
2067 | if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) { |
2068 | /* We know that the bi_io_vec layout is the same for | |
2069 | * both 'first' and 'i', so we just compare them. | |
2070 | * All vec entries are PAGE_SIZE; | |
2071 | */ | |
7bb23c49 N |
2072 | int sectors = r10_bio->sectors; |
2073 | for (j = 0; j < vcnt; j++) { | |
2074 | int len = PAGE_SIZE; | |
2075 | if (sectors < (len / 512)) | |
2076 | len = sectors * 512; | |
0eb3ff12 N |
2077 | if (memcmp(page_address(fbio->bi_io_vec[j].bv_page), |
2078 | page_address(tbio->bi_io_vec[j].bv_page), | |
7bb23c49 | 2079 | len)) |
0eb3ff12 | 2080 | break; |
7bb23c49 N |
2081 | sectors -= len/512; |
2082 | } | |
0eb3ff12 N |
2083 | if (j == vcnt) |
2084 | continue; | |
7f7583d4 | 2085 | atomic64_add(r10_bio->sectors, &mddev->resync_mismatches); |
f84ee364 N |
2086 | if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) |
2087 | /* Don't fix anything. */ | |
2088 | continue; | |
0eb3ff12 | 2089 | } |
f84ee364 N |
2090 | /* Ok, we need to write this bio, either to correct an |
2091 | * inconsistency or to correct an unreadable block. | |
1da177e4 LT |
2092 | * First we need to fixup bv_offset, bv_len and |
2093 | * bi_vecs, as the read request might have corrupted these | |
2094 | */ | |
8be185f2 KO |
2095 | bio_reset(tbio); |
2096 | ||
1da177e4 | 2097 | tbio->bi_vcnt = vcnt; |
4f024f37 | 2098 | tbio->bi_iter.bi_size = r10_bio->sectors << 9; |
1da177e4 LT |
2099 | tbio->bi_rw = WRITE; |
2100 | tbio->bi_private = r10_bio; | |
4f024f37 | 2101 | tbio->bi_iter.bi_sector = r10_bio->devs[i].addr; |
1da177e4 LT |
2102 | |
2103 | for (j=0; j < vcnt ; j++) { | |
2104 | tbio->bi_io_vec[j].bv_offset = 0; | |
2105 | tbio->bi_io_vec[j].bv_len = PAGE_SIZE; | |
2106 | ||
2107 | memcpy(page_address(tbio->bi_io_vec[j].bv_page), | |
2108 | page_address(fbio->bi_io_vec[j].bv_page), | |
2109 | PAGE_SIZE); | |
2110 | } | |
2111 | tbio->bi_end_io = end_sync_write; | |
2112 | ||
2113 | d = r10_bio->devs[i].devnum; | |
2114 | atomic_inc(&conf->mirrors[d].rdev->nr_pending); | |
2115 | atomic_inc(&r10_bio->remaining); | |
aa8b57aa | 2116 | md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio)); |
1da177e4 | 2117 | |
4f024f37 | 2118 | tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset; |
1da177e4 LT |
2119 | tbio->bi_bdev = conf->mirrors[d].rdev->bdev; |
2120 | generic_make_request(tbio); | |
2121 | } | |
2122 | ||
9ad1aefc N |
2123 | /* Now write out to any replacement devices |
2124 | * that are active | |
2125 | */ | |
2126 | for (i = 0; i < conf->copies; i++) { | |
2127 | int j, d; | |
9ad1aefc N |
2128 | |
2129 | tbio = r10_bio->devs[i].repl_bio; | |
2130 | if (!tbio || !tbio->bi_end_io) | |
2131 | continue; | |
2132 | if (r10_bio->devs[i].bio->bi_end_io != end_sync_write | |
2133 | && r10_bio->devs[i].bio != fbio) | |
2134 | for (j = 0; j < vcnt; j++) | |
2135 | memcpy(page_address(tbio->bi_io_vec[j].bv_page), | |
2136 | page_address(fbio->bi_io_vec[j].bv_page), | |
2137 | PAGE_SIZE); | |
2138 | d = r10_bio->devs[i].devnum; | |
2139 | atomic_inc(&r10_bio->remaining); | |
2140 | md_sync_acct(conf->mirrors[d].replacement->bdev, | |
aa8b57aa | 2141 | bio_sectors(tbio)); |
9ad1aefc N |
2142 | generic_make_request(tbio); |
2143 | } | |
2144 | ||
1da177e4 LT |
2145 | done: |
2146 | if (atomic_dec_and_test(&r10_bio->remaining)) { | |
2147 | md_done_sync(mddev, r10_bio->sectors, 1); | |
2148 | put_buf(r10_bio); | |
2149 | } | |
2150 | } | |
2151 | ||
2152 | /* | |
2153 | * Now for the recovery code. | |
2154 | * Recovery happens across physical sectors. | |
2155 | * We recover all non-is_sync drives by finding the virtual address of | |
2156 | * each, and then choose a working drive that also has that virt address. | |
2157 | * There is a separate r10_bio for each non-in_sync drive. | |
2158 | * Only the first two slots are in use. The first for reading, | |
2159 | * The second for writing. | |
2160 | * | |
2161 | */ | |
9f2c9d12 | 2162 | static void fix_recovery_read_error(struct r10bio *r10_bio) |
5e570289 N |
2163 | { |
2164 | /* We got a read error during recovery. | |
2165 | * We repeat the read in smaller page-sized sections. | |
2166 | * If a read succeeds, write it to the new device or record | |
2167 | * a bad block if we cannot. | |
2168 | * If a read fails, record a bad block on both old and | |
2169 | * new devices. | |
2170 | */ | |
fd01b88c | 2171 | struct mddev *mddev = r10_bio->mddev; |
e879a879 | 2172 | struct r10conf *conf = mddev->private; |
5e570289 N |
2173 | struct bio *bio = r10_bio->devs[0].bio; |
2174 | sector_t sect = 0; | |
2175 | int sectors = r10_bio->sectors; | |
2176 | int idx = 0; | |
2177 | int dr = r10_bio->devs[0].devnum; | |
2178 | int dw = r10_bio->devs[1].devnum; | |
2179 | ||
2180 | while (sectors) { | |
2181 | int s = sectors; | |
3cb03002 | 2182 | struct md_rdev *rdev; |
5e570289 N |
2183 | sector_t addr; |
2184 | int ok; | |
2185 | ||
2186 | if (s > (PAGE_SIZE>>9)) | |
2187 | s = PAGE_SIZE >> 9; | |
2188 | ||
2189 | rdev = conf->mirrors[dr].rdev; | |
2190 | addr = r10_bio->devs[0].addr + sect, | |
2191 | ok = sync_page_io(rdev, | |
2192 | addr, | |
2193 | s << 9, | |
2194 | bio->bi_io_vec[idx].bv_page, | |
2195 | READ, false); | |
2196 | if (ok) { | |
2197 | rdev = conf->mirrors[dw].rdev; | |
2198 | addr = r10_bio->devs[1].addr + sect; | |
2199 | ok = sync_page_io(rdev, | |
2200 | addr, | |
2201 | s << 9, | |
2202 | bio->bi_io_vec[idx].bv_page, | |
2203 | WRITE, false); | |
b7044d41 | 2204 | if (!ok) { |
5e570289 | 2205 | set_bit(WriteErrorSeen, &rdev->flags); |
b7044d41 N |
2206 | if (!test_and_set_bit(WantReplacement, |
2207 | &rdev->flags)) | |
2208 | set_bit(MD_RECOVERY_NEEDED, | |
2209 | &rdev->mddev->recovery); | |
2210 | } | |
5e570289 N |
2211 | } |
2212 | if (!ok) { | |
2213 | /* We don't worry if we cannot set a bad block - | |
2214 | * it really is bad so there is no loss in not | |
2215 | * recording it yet | |
2216 | */ | |
2217 | rdev_set_badblocks(rdev, addr, s, 0); | |
2218 | ||
2219 | if (rdev != conf->mirrors[dw].rdev) { | |
2220 | /* need bad block on destination too */ | |
3cb03002 | 2221 | struct md_rdev *rdev2 = conf->mirrors[dw].rdev; |
5e570289 N |
2222 | addr = r10_bio->devs[1].addr + sect; |
2223 | ok = rdev_set_badblocks(rdev2, addr, s, 0); | |
2224 | if (!ok) { | |
2225 | /* just abort the recovery */ | |
2226 | printk(KERN_NOTICE | |
2227 | "md/raid10:%s: recovery aborted" | |
2228 | " due to read error\n", | |
2229 | mdname(mddev)); | |
2230 | ||
2231 | conf->mirrors[dw].recovery_disabled | |
2232 | = mddev->recovery_disabled; | |
2233 | set_bit(MD_RECOVERY_INTR, | |
2234 | &mddev->recovery); | |
2235 | break; | |
2236 | } | |
2237 | } | |
2238 | } | |
2239 | ||
2240 | sectors -= s; | |
2241 | sect += s; | |
2242 | idx++; | |
2243 | } | |
2244 | } | |
1da177e4 | 2245 | |
9f2c9d12 | 2246 | static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio) |
1da177e4 | 2247 | { |
e879a879 | 2248 | struct r10conf *conf = mddev->private; |
c65060ad | 2249 | int d; |
24afd80d | 2250 | struct bio *wbio, *wbio2; |
1da177e4 | 2251 | |
5e570289 N |
2252 | if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) { |
2253 | fix_recovery_read_error(r10_bio); | |
2254 | end_sync_request(r10_bio); | |
2255 | return; | |
2256 | } | |
2257 | ||
c65060ad NK |
2258 | /* |
2259 | * share the pages with the first bio | |
1da177e4 LT |
2260 | * and submit the write request |
2261 | */ | |
1da177e4 | 2262 | d = r10_bio->devs[1].devnum; |
24afd80d N |
2263 | wbio = r10_bio->devs[1].bio; |
2264 | wbio2 = r10_bio->devs[1].repl_bio; | |
0eb25bb0 N |
2265 | /* Need to test wbio2->bi_end_io before we call |
2266 | * generic_make_request as if the former is NULL, | |
2267 | * the latter is free to free wbio2. | |
2268 | */ | |
2269 | if (wbio2 && !wbio2->bi_end_io) | |
2270 | wbio2 = NULL; | |
24afd80d N |
2271 | if (wbio->bi_end_io) { |
2272 | atomic_inc(&conf->mirrors[d].rdev->nr_pending); | |
aa8b57aa | 2273 | md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio)); |
24afd80d N |
2274 | generic_make_request(wbio); |
2275 | } | |
0eb25bb0 | 2276 | if (wbio2) { |
24afd80d N |
2277 | atomic_inc(&conf->mirrors[d].replacement->nr_pending); |
2278 | md_sync_acct(conf->mirrors[d].replacement->bdev, | |
aa8b57aa | 2279 | bio_sectors(wbio2)); |
24afd80d N |
2280 | generic_make_request(wbio2); |
2281 | } | |
1da177e4 LT |
2282 | } |
2283 | ||
1e50915f RB |
2284 | /* |
2285 | * Used by fix_read_error() to decay the per rdev read_errors. | |
2286 | * We halve the read error count for every hour that has elapsed | |
2287 | * since the last recorded read error. | |
2288 | * | |
2289 | */ | |
fd01b88c | 2290 | static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev) |
1e50915f RB |
2291 | { |
2292 | struct timespec cur_time_mon; | |
2293 | unsigned long hours_since_last; | |
2294 | unsigned int read_errors = atomic_read(&rdev->read_errors); | |
2295 | ||
2296 | ktime_get_ts(&cur_time_mon); | |
2297 | ||
2298 | if (rdev->last_read_error.tv_sec == 0 && | |
2299 | rdev->last_read_error.tv_nsec == 0) { | |
2300 | /* first time we've seen a read error */ | |
2301 | rdev->last_read_error = cur_time_mon; | |
2302 | return; | |
2303 | } | |
2304 | ||
2305 | hours_since_last = (cur_time_mon.tv_sec - | |
2306 | rdev->last_read_error.tv_sec) / 3600; | |
2307 | ||
2308 | rdev->last_read_error = cur_time_mon; | |
2309 | ||
2310 | /* | |
2311 | * if hours_since_last is > the number of bits in read_errors | |
2312 | * just set read errors to 0. We do this to avoid | |
2313 | * overflowing the shift of read_errors by hours_since_last. | |
2314 | */ | |
2315 | if (hours_since_last >= 8 * sizeof(read_errors)) | |
2316 | atomic_set(&rdev->read_errors, 0); | |
2317 | else | |
2318 | atomic_set(&rdev->read_errors, read_errors >> hours_since_last); | |
2319 | } | |
2320 | ||
3cb03002 | 2321 | static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector, |
58c54fcc N |
2322 | int sectors, struct page *page, int rw) |
2323 | { | |
2324 | sector_t first_bad; | |
2325 | int bad_sectors; | |
2326 | ||
2327 | if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors) | |
2328 | && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags))) | |
2329 | return -1; | |
2330 | if (sync_page_io(rdev, sector, sectors << 9, page, rw, false)) | |
2331 | /* success */ | |
2332 | return 1; | |
b7044d41 | 2333 | if (rw == WRITE) { |
58c54fcc | 2334 | set_bit(WriteErrorSeen, &rdev->flags); |
b7044d41 N |
2335 | if (!test_and_set_bit(WantReplacement, &rdev->flags)) |
2336 | set_bit(MD_RECOVERY_NEEDED, | |
2337 | &rdev->mddev->recovery); | |
2338 | } | |
58c54fcc N |
2339 | /* need to record an error - either for the block or the device */ |
2340 | if (!rdev_set_badblocks(rdev, sector, sectors, 0)) | |
2341 | md_error(rdev->mddev, rdev); | |
2342 | return 0; | |
2343 | } | |
2344 | ||
1da177e4 LT |
2345 | /* |
2346 | * This is a kernel thread which: | |
2347 | * | |
2348 | * 1. Retries failed read operations on working mirrors. | |
2349 | * 2. Updates the raid superblock when problems encounter. | |
6814d536 | 2350 | * 3. Performs writes following reads for array synchronising. |
1da177e4 LT |
2351 | */ |
2352 | ||
e879a879 | 2353 | static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio) |
6814d536 N |
2354 | { |
2355 | int sect = 0; /* Offset from r10_bio->sector */ | |
2356 | int sectors = r10_bio->sectors; | |
3cb03002 | 2357 | struct md_rdev*rdev; |
1e50915f | 2358 | int max_read_errors = atomic_read(&mddev->max_corr_read_errors); |
0544a21d | 2359 | int d = r10_bio->devs[r10_bio->read_slot].devnum; |
1e50915f | 2360 | |
7c4e06ff N |
2361 | /* still own a reference to this rdev, so it cannot |
2362 | * have been cleared recently. | |
2363 | */ | |
2364 | rdev = conf->mirrors[d].rdev; | |
1e50915f | 2365 | |
7c4e06ff N |
2366 | if (test_bit(Faulty, &rdev->flags)) |
2367 | /* drive has already been failed, just ignore any | |
2368 | more fix_read_error() attempts */ | |
2369 | return; | |
1e50915f | 2370 | |
7c4e06ff N |
2371 | check_decay_read_errors(mddev, rdev); |
2372 | atomic_inc(&rdev->read_errors); | |
2373 | if (atomic_read(&rdev->read_errors) > max_read_errors) { | |
2374 | char b[BDEVNAME_SIZE]; | |
2375 | bdevname(rdev->bdev, b); | |
1e50915f | 2376 | |
7c4e06ff N |
2377 | printk(KERN_NOTICE |
2378 | "md/raid10:%s: %s: Raid device exceeded " | |
2379 | "read_error threshold [cur %d:max %d]\n", | |
2380 | mdname(mddev), b, | |
2381 | atomic_read(&rdev->read_errors), max_read_errors); | |
2382 | printk(KERN_NOTICE | |
2383 | "md/raid10:%s: %s: Failing raid device\n", | |
2384 | mdname(mddev), b); | |
2385 | md_error(mddev, conf->mirrors[d].rdev); | |
fae8cc5e | 2386 | r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED; |
7c4e06ff | 2387 | return; |
1e50915f | 2388 | } |
1e50915f | 2389 | |
6814d536 N |
2390 | while(sectors) { |
2391 | int s = sectors; | |
2392 | int sl = r10_bio->read_slot; | |
2393 | int success = 0; | |
2394 | int start; | |
2395 | ||
2396 | if (s > (PAGE_SIZE>>9)) | |
2397 | s = PAGE_SIZE >> 9; | |
2398 | ||
2399 | rcu_read_lock(); | |
2400 | do { | |
8dbed5ce N |
2401 | sector_t first_bad; |
2402 | int bad_sectors; | |
2403 | ||
0544a21d | 2404 | d = r10_bio->devs[sl].devnum; |
6814d536 N |
2405 | rdev = rcu_dereference(conf->mirrors[d].rdev); |
2406 | if (rdev && | |
050b6615 | 2407 | !test_bit(Unmerged, &rdev->flags) && |
8dbed5ce N |
2408 | test_bit(In_sync, &rdev->flags) && |
2409 | is_badblock(rdev, r10_bio->devs[sl].addr + sect, s, | |
2410 | &first_bad, &bad_sectors) == 0) { | |
6814d536 N |
2411 | atomic_inc(&rdev->nr_pending); |
2412 | rcu_read_unlock(); | |
2b193363 | 2413 | success = sync_page_io(rdev, |
6814d536 | 2414 | r10_bio->devs[sl].addr + |
ccebd4c4 | 2415 | sect, |
6814d536 | 2416 | s<<9, |
ccebd4c4 | 2417 | conf->tmppage, READ, false); |
6814d536 N |
2418 | rdev_dec_pending(rdev, mddev); |
2419 | rcu_read_lock(); | |
2420 | if (success) | |
2421 | break; | |
2422 | } | |
2423 | sl++; | |
2424 | if (sl == conf->copies) | |
2425 | sl = 0; | |
2426 | } while (!success && sl != r10_bio->read_slot); | |
2427 | rcu_read_unlock(); | |
2428 | ||
2429 | if (!success) { | |
58c54fcc N |
2430 | /* Cannot read from anywhere, just mark the block |
2431 | * as bad on the first device to discourage future | |
2432 | * reads. | |
2433 | */ | |
6814d536 | 2434 | int dn = r10_bio->devs[r10_bio->read_slot].devnum; |
58c54fcc N |
2435 | rdev = conf->mirrors[dn].rdev; |
2436 | ||
2437 | if (!rdev_set_badblocks( | |
2438 | rdev, | |
2439 | r10_bio->devs[r10_bio->read_slot].addr | |
2440 | + sect, | |
fae8cc5e | 2441 | s, 0)) { |
58c54fcc | 2442 | md_error(mddev, rdev); |
fae8cc5e N |
2443 | r10_bio->devs[r10_bio->read_slot].bio |
2444 | = IO_BLOCKED; | |
2445 | } | |
6814d536 N |
2446 | break; |
2447 | } | |
2448 | ||
2449 | start = sl; | |
2450 | /* write it back and re-read */ | |
2451 | rcu_read_lock(); | |
2452 | while (sl != r10_bio->read_slot) { | |
67b8dc4b | 2453 | char b[BDEVNAME_SIZE]; |
0544a21d | 2454 | |
6814d536 N |
2455 | if (sl==0) |
2456 | sl = conf->copies; | |
2457 | sl--; | |
2458 | d = r10_bio->devs[sl].devnum; | |
2459 | rdev = rcu_dereference(conf->mirrors[d].rdev); | |
1294b9c9 | 2460 | if (!rdev || |
050b6615 | 2461 | test_bit(Unmerged, &rdev->flags) || |
1294b9c9 N |
2462 | !test_bit(In_sync, &rdev->flags)) |
2463 | continue; | |
2464 | ||
2465 | atomic_inc(&rdev->nr_pending); | |
2466 | rcu_read_unlock(); | |
58c54fcc N |
2467 | if (r10_sync_page_io(rdev, |
2468 | r10_bio->devs[sl].addr + | |
2469 | sect, | |
055d3747 | 2470 | s, conf->tmppage, WRITE) |
1294b9c9 N |
2471 | == 0) { |
2472 | /* Well, this device is dead */ | |
2473 | printk(KERN_NOTICE | |
2474 | "md/raid10:%s: read correction " | |
2475 | "write failed" | |
2476 | " (%d sectors at %llu on %s)\n", | |
2477 | mdname(mddev), s, | |
2478 | (unsigned long long)( | |
f8c9e74f N |
2479 | sect + |
2480 | choose_data_offset(r10_bio, | |
2481 | rdev)), | |
1294b9c9 N |
2482 | bdevname(rdev->bdev, b)); |
2483 | printk(KERN_NOTICE "md/raid10:%s: %s: failing " | |
2484 | "drive\n", | |
2485 | mdname(mddev), | |
2486 | bdevname(rdev->bdev, b)); | |
6814d536 | 2487 | } |
1294b9c9 N |
2488 | rdev_dec_pending(rdev, mddev); |
2489 | rcu_read_lock(); | |
6814d536 N |
2490 | } |
2491 | sl = start; | |
2492 | while (sl != r10_bio->read_slot) { | |
1294b9c9 | 2493 | char b[BDEVNAME_SIZE]; |
0544a21d | 2494 | |
6814d536 N |
2495 | if (sl==0) |
2496 | sl = conf->copies; | |
2497 | sl--; | |
2498 | d = r10_bio->devs[sl].devnum; | |
2499 | rdev = rcu_dereference(conf->mirrors[d].rdev); | |
1294b9c9 N |
2500 | if (!rdev || |
2501 | !test_bit(In_sync, &rdev->flags)) | |
2502 | continue; | |
6814d536 | 2503 | |
1294b9c9 N |
2504 | atomic_inc(&rdev->nr_pending); |
2505 | rcu_read_unlock(); | |
58c54fcc N |
2506 | switch (r10_sync_page_io(rdev, |
2507 | r10_bio->devs[sl].addr + | |
2508 | sect, | |
055d3747 | 2509 | s, conf->tmppage, |
58c54fcc N |
2510 | READ)) { |
2511 | case 0: | |
1294b9c9 N |
2512 | /* Well, this device is dead */ |
2513 | printk(KERN_NOTICE | |
2514 | "md/raid10:%s: unable to read back " | |
2515 | "corrected sectors" | |
2516 | " (%d sectors at %llu on %s)\n", | |
2517 | mdname(mddev), s, | |
2518 | (unsigned long long)( | |
f8c9e74f N |
2519 | sect + |
2520 | choose_data_offset(r10_bio, rdev)), | |
1294b9c9 N |
2521 | bdevname(rdev->bdev, b)); |
2522 | printk(KERN_NOTICE "md/raid10:%s: %s: failing " | |
2523 | "drive\n", | |
2524 | mdname(mddev), | |
2525 | bdevname(rdev->bdev, b)); | |
58c54fcc N |
2526 | break; |
2527 | case 1: | |
1294b9c9 N |
2528 | printk(KERN_INFO |
2529 | "md/raid10:%s: read error corrected" | |
2530 | " (%d sectors at %llu on %s)\n", | |
2531 | mdname(mddev), s, | |
2532 | (unsigned long long)( | |
f8c9e74f N |
2533 | sect + |
2534 | choose_data_offset(r10_bio, rdev)), | |
1294b9c9 N |
2535 | bdevname(rdev->bdev, b)); |
2536 | atomic_add(s, &rdev->corrected_errors); | |
6814d536 | 2537 | } |
1294b9c9 N |
2538 | |
2539 | rdev_dec_pending(rdev, mddev); | |
2540 | rcu_read_lock(); | |
6814d536 N |
2541 | } |
2542 | rcu_read_unlock(); | |
2543 | ||
2544 | sectors -= s; | |
2545 | sect += s; | |
2546 | } | |
2547 | } | |
2548 | ||
9f2c9d12 | 2549 | static int narrow_write_error(struct r10bio *r10_bio, int i) |
bd870a16 N |
2550 | { |
2551 | struct bio *bio = r10_bio->master_bio; | |
fd01b88c | 2552 | struct mddev *mddev = r10_bio->mddev; |
e879a879 | 2553 | struct r10conf *conf = mddev->private; |
3cb03002 | 2554 | struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev; |
bd870a16 N |
2555 | /* bio has the data to be written to slot 'i' where |
2556 | * we just recently had a write error. | |
2557 | * We repeatedly clone the bio and trim down to one block, | |
2558 | * then try the write. Where the write fails we record | |
2559 | * a bad block. | |
2560 | * It is conceivable that the bio doesn't exactly align with | |
2561 | * blocks. We must handle this. | |
2562 | * | |
2563 | * We currently own a reference to the rdev. | |
2564 | */ | |
2565 | ||
2566 | int block_sectors; | |
2567 | sector_t sector; | |
2568 | int sectors; | |
2569 | int sect_to_write = r10_bio->sectors; | |
2570 | int ok = 1; | |
2571 | ||
2572 | if (rdev->badblocks.shift < 0) | |
2573 | return 0; | |
2574 | ||
f04ebb0b N |
2575 | block_sectors = roundup(1 << rdev->badblocks.shift, |
2576 | bdev_logical_block_size(rdev->bdev) >> 9); | |
bd870a16 N |
2577 | sector = r10_bio->sector; |
2578 | sectors = ((r10_bio->sector + block_sectors) | |
2579 | & ~(sector_t)(block_sectors - 1)) | |
2580 | - sector; | |
2581 | ||
2582 | while (sect_to_write) { | |
2583 | struct bio *wbio; | |
2584 | if (sectors > sect_to_write) | |
2585 | sectors = sect_to_write; | |
2586 | /* Write at 'sector' for 'sectors' */ | |
2587 | wbio = bio_clone_mddev(bio, GFP_NOIO, mddev); | |
4f024f37 KO |
2588 | bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors); |
2589 | wbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+ | |
f8c9e74f | 2590 | choose_data_offset(r10_bio, rdev) + |
bd870a16 N |
2591 | (sector - r10_bio->sector)); |
2592 | wbio->bi_bdev = rdev->bdev; | |
2593 | if (submit_bio_wait(WRITE, wbio) == 0) | |
2594 | /* Failure! */ | |
2595 | ok = rdev_set_badblocks(rdev, sector, | |
2596 | sectors, 0) | |
2597 | && ok; | |
2598 | ||
2599 | bio_put(wbio); | |
2600 | sect_to_write -= sectors; | |
2601 | sector += sectors; | |
2602 | sectors = block_sectors; | |
2603 | } | |
2604 | return ok; | |
2605 | } | |
2606 | ||
9f2c9d12 | 2607 | static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio) |
560f8e55 N |
2608 | { |
2609 | int slot = r10_bio->read_slot; | |
560f8e55 | 2610 | struct bio *bio; |
e879a879 | 2611 | struct r10conf *conf = mddev->private; |
abbf098e | 2612 | struct md_rdev *rdev = r10_bio->devs[slot].rdev; |
560f8e55 N |
2613 | char b[BDEVNAME_SIZE]; |
2614 | unsigned long do_sync; | |
856e08e2 | 2615 | int max_sectors; |
560f8e55 N |
2616 | |
2617 | /* we got a read error. Maybe the drive is bad. Maybe just | |
2618 | * the block and we can fix it. | |
2619 | * We freeze all other IO, and try reading the block from | |
2620 | * other devices. When we find one, we re-write | |
2621 | * and check it that fixes the read error. | |
2622 | * This is all done synchronously while the array is | |
2623 | * frozen. | |
2624 | */ | |
fae8cc5e N |
2625 | bio = r10_bio->devs[slot].bio; |
2626 | bdevname(bio->bi_bdev, b); | |
2627 | bio_put(bio); | |
2628 | r10_bio->devs[slot].bio = NULL; | |
2629 | ||
560f8e55 | 2630 | if (mddev->ro == 0) { |
e2d59925 | 2631 | freeze_array(conf, 1); |
560f8e55 N |
2632 | fix_read_error(conf, mddev, r10_bio); |
2633 | unfreeze_array(conf); | |
fae8cc5e N |
2634 | } else |
2635 | r10_bio->devs[slot].bio = IO_BLOCKED; | |
2636 | ||
abbf098e | 2637 | rdev_dec_pending(rdev, mddev); |
560f8e55 | 2638 | |
7399c31b | 2639 | read_more: |
96c3fd1f N |
2640 | rdev = read_balance(conf, r10_bio, &max_sectors); |
2641 | if (rdev == NULL) { | |
560f8e55 N |
2642 | printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O" |
2643 | " read error for block %llu\n", | |
7399c31b | 2644 | mdname(mddev), b, |
560f8e55 N |
2645 | (unsigned long long)r10_bio->sector); |
2646 | raid_end_bio_io(r10_bio); | |
560f8e55 N |
2647 | return; |
2648 | } | |
2649 | ||
2650 | do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC); | |
560f8e55 | 2651 | slot = r10_bio->read_slot; |
560f8e55 N |
2652 | printk_ratelimited( |
2653 | KERN_ERR | |
055d3747 | 2654 | "md/raid10:%s: %s: redirecting " |
560f8e55 N |
2655 | "sector %llu to another mirror\n", |
2656 | mdname(mddev), | |
2657 | bdevname(rdev->bdev, b), | |
2658 | (unsigned long long)r10_bio->sector); | |
2659 | bio = bio_clone_mddev(r10_bio->master_bio, | |
2660 | GFP_NOIO, mddev); | |
4f024f37 | 2661 | bio_trim(bio, r10_bio->sector - bio->bi_iter.bi_sector, max_sectors); |
560f8e55 | 2662 | r10_bio->devs[slot].bio = bio; |
abbf098e | 2663 | r10_bio->devs[slot].rdev = rdev; |
4f024f37 | 2664 | bio->bi_iter.bi_sector = r10_bio->devs[slot].addr |
f8c9e74f | 2665 | + choose_data_offset(r10_bio, rdev); |
560f8e55 N |
2666 | bio->bi_bdev = rdev->bdev; |
2667 | bio->bi_rw = READ | do_sync; | |
2668 | bio->bi_private = r10_bio; | |
2669 | bio->bi_end_io = raid10_end_read_request; | |
7399c31b N |
2670 | if (max_sectors < r10_bio->sectors) { |
2671 | /* Drat - have to split this up more */ | |
2672 | struct bio *mbio = r10_bio->master_bio; | |
2673 | int sectors_handled = | |
2674 | r10_bio->sector + max_sectors | |
4f024f37 | 2675 | - mbio->bi_iter.bi_sector; |
7399c31b N |
2676 | r10_bio->sectors = max_sectors; |
2677 | spin_lock_irq(&conf->device_lock); | |
2678 | if (mbio->bi_phys_segments == 0) | |
2679 | mbio->bi_phys_segments = 2; | |
2680 | else | |
2681 | mbio->bi_phys_segments++; | |
2682 | spin_unlock_irq(&conf->device_lock); | |
2683 | generic_make_request(bio); | |
7399c31b N |
2684 | |
2685 | r10_bio = mempool_alloc(conf->r10bio_pool, | |
2686 | GFP_NOIO); | |
2687 | r10_bio->master_bio = mbio; | |
aa8b57aa | 2688 | r10_bio->sectors = bio_sectors(mbio) - sectors_handled; |
7399c31b N |
2689 | r10_bio->state = 0; |
2690 | set_bit(R10BIO_ReadError, | |
2691 | &r10_bio->state); | |
2692 | r10_bio->mddev = mddev; | |
4f024f37 | 2693 | r10_bio->sector = mbio->bi_iter.bi_sector |
7399c31b N |
2694 | + sectors_handled; |
2695 | ||
2696 | goto read_more; | |
2697 | } else | |
2698 | generic_make_request(bio); | |
560f8e55 N |
2699 | } |
2700 | ||
e879a879 | 2701 | static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio) |
749c55e9 N |
2702 | { |
2703 | /* Some sort of write request has finished and it | |
2704 | * succeeded in writing where we thought there was a | |
2705 | * bad block. So forget the bad block. | |
1a0b7cd8 N |
2706 | * Or possibly if failed and we need to record |
2707 | * a bad block. | |
749c55e9 N |
2708 | */ |
2709 | int m; | |
3cb03002 | 2710 | struct md_rdev *rdev; |
749c55e9 N |
2711 | |
2712 | if (test_bit(R10BIO_IsSync, &r10_bio->state) || | |
2713 | test_bit(R10BIO_IsRecover, &r10_bio->state)) { | |
1a0b7cd8 N |
2714 | for (m = 0; m < conf->copies; m++) { |
2715 | int dev = r10_bio->devs[m].devnum; | |
2716 | rdev = conf->mirrors[dev].rdev; | |
2717 | if (r10_bio->devs[m].bio == NULL) | |
2718 | continue; | |
2719 | if (test_bit(BIO_UPTODATE, | |
749c55e9 | 2720 | &r10_bio->devs[m].bio->bi_flags)) { |
749c55e9 N |
2721 | rdev_clear_badblocks( |
2722 | rdev, | |
2723 | r10_bio->devs[m].addr, | |
c6563a8c | 2724 | r10_bio->sectors, 0); |
1a0b7cd8 N |
2725 | } else { |
2726 | if (!rdev_set_badblocks( | |
2727 | rdev, | |
2728 | r10_bio->devs[m].addr, | |
2729 | r10_bio->sectors, 0)) | |
2730 | md_error(conf->mddev, rdev); | |
749c55e9 | 2731 | } |
9ad1aefc N |
2732 | rdev = conf->mirrors[dev].replacement; |
2733 | if (r10_bio->devs[m].repl_bio == NULL) | |
2734 | continue; | |
2735 | if (test_bit(BIO_UPTODATE, | |
2736 | &r10_bio->devs[m].repl_bio->bi_flags)) { | |
2737 | rdev_clear_badblocks( | |
2738 | rdev, | |
2739 | r10_bio->devs[m].addr, | |
c6563a8c | 2740 | r10_bio->sectors, 0); |
9ad1aefc N |
2741 | } else { |
2742 | if (!rdev_set_badblocks( | |
2743 | rdev, | |
2744 | r10_bio->devs[m].addr, | |
2745 | r10_bio->sectors, 0)) | |
2746 | md_error(conf->mddev, rdev); | |
2747 | } | |
1a0b7cd8 | 2748 | } |
749c55e9 N |
2749 | put_buf(r10_bio); |
2750 | } else { | |
bd870a16 N |
2751 | for (m = 0; m < conf->copies; m++) { |
2752 | int dev = r10_bio->devs[m].devnum; | |
2753 | struct bio *bio = r10_bio->devs[m].bio; | |
2754 | rdev = conf->mirrors[dev].rdev; | |
2755 | if (bio == IO_MADE_GOOD) { | |
749c55e9 N |
2756 | rdev_clear_badblocks( |
2757 | rdev, | |
2758 | r10_bio->devs[m].addr, | |
c6563a8c | 2759 | r10_bio->sectors, 0); |
749c55e9 | 2760 | rdev_dec_pending(rdev, conf->mddev); |
bd870a16 N |
2761 | } else if (bio != NULL && |
2762 | !test_bit(BIO_UPTODATE, &bio->bi_flags)) { | |
2763 | if (!narrow_write_error(r10_bio, m)) { | |
2764 | md_error(conf->mddev, rdev); | |
2765 | set_bit(R10BIO_Degraded, | |
2766 | &r10_bio->state); | |
2767 | } | |
2768 | rdev_dec_pending(rdev, conf->mddev); | |
749c55e9 | 2769 | } |
475b0321 N |
2770 | bio = r10_bio->devs[m].repl_bio; |
2771 | rdev = conf->mirrors[dev].replacement; | |
4ca40c2c | 2772 | if (rdev && bio == IO_MADE_GOOD) { |
475b0321 N |
2773 | rdev_clear_badblocks( |
2774 | rdev, | |
2775 | r10_bio->devs[m].addr, | |
c6563a8c | 2776 | r10_bio->sectors, 0); |
475b0321 N |
2777 | rdev_dec_pending(rdev, conf->mddev); |
2778 | } | |
bd870a16 N |
2779 | } |
2780 | if (test_bit(R10BIO_WriteError, | |
2781 | &r10_bio->state)) | |
2782 | close_write(r10_bio); | |
749c55e9 N |
2783 | raid_end_bio_io(r10_bio); |
2784 | } | |
2785 | } | |
2786 | ||
4ed8731d | 2787 | static void raid10d(struct md_thread *thread) |
1da177e4 | 2788 | { |
4ed8731d | 2789 | struct mddev *mddev = thread->mddev; |
9f2c9d12 | 2790 | struct r10bio *r10_bio; |
1da177e4 | 2791 | unsigned long flags; |
e879a879 | 2792 | struct r10conf *conf = mddev->private; |
1da177e4 | 2793 | struct list_head *head = &conf->retry_list; |
e1dfa0a2 | 2794 | struct blk_plug plug; |
1da177e4 LT |
2795 | |
2796 | md_check_recovery(mddev); | |
1da177e4 | 2797 | |
e1dfa0a2 | 2798 | blk_start_plug(&plug); |
1da177e4 | 2799 | for (;;) { |
6cce3b23 | 2800 | |
0021b7bc | 2801 | flush_pending_writes(conf); |
6cce3b23 | 2802 | |
a35e63ef N |
2803 | spin_lock_irqsave(&conf->device_lock, flags); |
2804 | if (list_empty(head)) { | |
2805 | spin_unlock_irqrestore(&conf->device_lock, flags); | |
1da177e4 | 2806 | break; |
a35e63ef | 2807 | } |
9f2c9d12 | 2808 | r10_bio = list_entry(head->prev, struct r10bio, retry_list); |
1da177e4 | 2809 | list_del(head->prev); |
4443ae10 | 2810 | conf->nr_queued--; |
1da177e4 LT |
2811 | spin_unlock_irqrestore(&conf->device_lock, flags); |
2812 | ||
2813 | mddev = r10_bio->mddev; | |
070ec55d | 2814 | conf = mddev->private; |
bd870a16 N |
2815 | if (test_bit(R10BIO_MadeGood, &r10_bio->state) || |
2816 | test_bit(R10BIO_WriteError, &r10_bio->state)) | |
749c55e9 | 2817 | handle_write_completed(conf, r10_bio); |
3ea7daa5 N |
2818 | else if (test_bit(R10BIO_IsReshape, &r10_bio->state)) |
2819 | reshape_request_write(mddev, r10_bio); | |
749c55e9 | 2820 | else if (test_bit(R10BIO_IsSync, &r10_bio->state)) |
1da177e4 | 2821 | sync_request_write(mddev, r10_bio); |
7eaceacc | 2822 | else if (test_bit(R10BIO_IsRecover, &r10_bio->state)) |
1da177e4 | 2823 | recovery_request_write(mddev, r10_bio); |
856e08e2 | 2824 | else if (test_bit(R10BIO_ReadError, &r10_bio->state)) |
560f8e55 | 2825 | handle_read_error(mddev, r10_bio); |
856e08e2 N |
2826 | else { |
2827 | /* just a partial read to be scheduled from a | |
2828 | * separate context | |
2829 | */ | |
2830 | int slot = r10_bio->read_slot; | |
2831 | generic_make_request(r10_bio->devs[slot].bio); | |
2832 | } | |
560f8e55 | 2833 | |
1d9d5241 | 2834 | cond_resched(); |
de393cde N |
2835 | if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) |
2836 | md_check_recovery(mddev); | |
1da177e4 | 2837 | } |
e1dfa0a2 | 2838 | blk_finish_plug(&plug); |
1da177e4 LT |
2839 | } |
2840 | ||
e879a879 | 2841 | static int init_resync(struct r10conf *conf) |
1da177e4 LT |
2842 | { |
2843 | int buffs; | |
69335ef3 | 2844 | int i; |
1da177e4 LT |
2845 | |
2846 | buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE; | |
b6385483 | 2847 | BUG_ON(conf->r10buf_pool); |
69335ef3 | 2848 | conf->have_replacement = 0; |
5cf00fcd | 2849 | for (i = 0; i < conf->geo.raid_disks; i++) |
69335ef3 N |
2850 | if (conf->mirrors[i].replacement) |
2851 | conf->have_replacement = 1; | |
1da177e4 LT |
2852 | conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf); |
2853 | if (!conf->r10buf_pool) | |
2854 | return -ENOMEM; | |
2855 | conf->next_resync = 0; | |
2856 | return 0; | |
2857 | } | |
2858 | ||
2859 | /* | |
2860 | * perform a "sync" on one "block" | |
2861 | * | |
2862 | * We need to make sure that no normal I/O request - particularly write | |
2863 | * requests - conflict with active sync requests. | |
2864 | * | |
2865 | * This is achieved by tracking pending requests and a 'barrier' concept | |
2866 | * that can be installed to exclude normal IO requests. | |
2867 | * | |
2868 | * Resync and recovery are handled very differently. | |
2869 | * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery. | |
2870 | * | |
2871 | * For resync, we iterate over virtual addresses, read all copies, | |
2872 | * and update if there are differences. If only one copy is live, | |
2873 | * skip it. | |
2874 | * For recovery, we iterate over physical addresses, read a good | |
2875 | * value for each non-in_sync drive, and over-write. | |
2876 | * | |
2877 | * So, for recovery we may have several outstanding complex requests for a | |
2878 | * given address, one for each out-of-sync device. We model this by allocating | |
2879 | * a number of r10_bio structures, one for each out-of-sync device. | |
2880 | * As we setup these structures, we collect all bio's together into a list | |
2881 | * which we then process collectively to add pages, and then process again | |
2882 | * to pass to generic_make_request. | |
2883 | * | |
2884 | * The r10_bio structures are linked using a borrowed master_bio pointer. | |
2885 | * This link is counted in ->remaining. When the r10_bio that points to NULL | |
2886 | * has its remaining count decremented to 0, the whole complex operation | |
2887 | * is complete. | |
2888 | * | |
2889 | */ | |
2890 | ||
fd01b88c | 2891 | static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, |
ab9d47e9 | 2892 | int *skipped, int go_faster) |
1da177e4 | 2893 | { |
e879a879 | 2894 | struct r10conf *conf = mddev->private; |
9f2c9d12 | 2895 | struct r10bio *r10_bio; |
1da177e4 LT |
2896 | struct bio *biolist = NULL, *bio; |
2897 | sector_t max_sector, nr_sectors; | |
1da177e4 | 2898 | int i; |
6cce3b23 | 2899 | int max_sync; |
57dab0bd | 2900 | sector_t sync_blocks; |
1da177e4 LT |
2901 | sector_t sectors_skipped = 0; |
2902 | int chunks_skipped = 0; | |
5cf00fcd | 2903 | sector_t chunk_mask = conf->geo.chunk_mask; |
1da177e4 LT |
2904 | |
2905 | if (!conf->r10buf_pool) | |
2906 | if (init_resync(conf)) | |
57afd89f | 2907 | return 0; |
1da177e4 | 2908 | |
7e83ccbe MW |
2909 | /* |
2910 | * Allow skipping a full rebuild for incremental assembly | |
2911 | * of a clean array, like RAID1 does. | |
2912 | */ | |
2913 | if (mddev->bitmap == NULL && | |
2914 | mddev->recovery_cp == MaxSector && | |
13765120 N |
2915 | mddev->reshape_position == MaxSector && |
2916 | !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && | |
7e83ccbe | 2917 | !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) && |
13765120 | 2918 | !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) && |
7e83ccbe MW |
2919 | conf->fullsync == 0) { |
2920 | *skipped = 1; | |
13765120 | 2921 | return mddev->dev_sectors - sector_nr; |
7e83ccbe MW |
2922 | } |
2923 | ||
1da177e4 | 2924 | skipped: |
58c0fed4 | 2925 | max_sector = mddev->dev_sectors; |
3ea7daa5 N |
2926 | if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) || |
2927 | test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) | |
1da177e4 LT |
2928 | max_sector = mddev->resync_max_sectors; |
2929 | if (sector_nr >= max_sector) { | |
6cce3b23 N |
2930 | /* If we aborted, we need to abort the |
2931 | * sync on the 'current' bitmap chucks (there can | |
2932 | * be several when recovering multiple devices). | |
2933 | * as we may have started syncing it but not finished. | |
2934 | * We can find the current address in | |
2935 | * mddev->curr_resync, but for recovery, | |
2936 | * we need to convert that to several | |
2937 | * virtual addresses. | |
2938 | */ | |
3ea7daa5 N |
2939 | if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) { |
2940 | end_reshape(conf); | |
b3968552 | 2941 | close_sync(conf); |
3ea7daa5 N |
2942 | return 0; |
2943 | } | |
2944 | ||
6cce3b23 N |
2945 | if (mddev->curr_resync < max_sector) { /* aborted */ |
2946 | if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) | |
2947 | bitmap_end_sync(mddev->bitmap, mddev->curr_resync, | |
2948 | &sync_blocks, 1); | |
5cf00fcd | 2949 | else for (i = 0; i < conf->geo.raid_disks; i++) { |
6cce3b23 N |
2950 | sector_t sect = |
2951 | raid10_find_virt(conf, mddev->curr_resync, i); | |
2952 | bitmap_end_sync(mddev->bitmap, sect, | |
2953 | &sync_blocks, 1); | |
2954 | } | |
9ad1aefc N |
2955 | } else { |
2956 | /* completed sync */ | |
2957 | if ((!mddev->bitmap || conf->fullsync) | |
2958 | && conf->have_replacement | |
2959 | && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) { | |
2960 | /* Completed a full sync so the replacements | |
2961 | * are now fully recovered. | |
2962 | */ | |
5cf00fcd | 2963 | for (i = 0; i < conf->geo.raid_disks; i++) |
9ad1aefc N |
2964 | if (conf->mirrors[i].replacement) |
2965 | conf->mirrors[i].replacement | |
2966 | ->recovery_offset | |
2967 | = MaxSector; | |
2968 | } | |
6cce3b23 | 2969 | conf->fullsync = 0; |
9ad1aefc | 2970 | } |
6cce3b23 | 2971 | bitmap_close_sync(mddev->bitmap); |
1da177e4 | 2972 | close_sync(conf); |
57afd89f | 2973 | *skipped = 1; |
1da177e4 LT |
2974 | return sectors_skipped; |
2975 | } | |
3ea7daa5 N |
2976 | |
2977 | if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) | |
2978 | return reshape_request(mddev, sector_nr, skipped); | |
2979 | ||
5cf00fcd | 2980 | if (chunks_skipped >= conf->geo.raid_disks) { |
1da177e4 LT |
2981 | /* if there has been nothing to do on any drive, |
2982 | * then there is nothing to do at all.. | |
2983 | */ | |
57afd89f N |
2984 | *skipped = 1; |
2985 | return (max_sector - sector_nr) + sectors_skipped; | |
1da177e4 LT |
2986 | } |
2987 | ||
c6207277 N |
2988 | if (max_sector > mddev->resync_max) |
2989 | max_sector = mddev->resync_max; /* Don't do IO beyond here */ | |
2990 | ||
1da177e4 LT |
2991 | /* make sure whole request will fit in a chunk - if chunks |
2992 | * are meaningful | |
2993 | */ | |
5cf00fcd N |
2994 | if (conf->geo.near_copies < conf->geo.raid_disks && |
2995 | max_sector > (sector_nr | chunk_mask)) | |
2996 | max_sector = (sector_nr | chunk_mask) + 1; | |
1da177e4 LT |
2997 | /* |
2998 | * If there is non-resync activity waiting for us then | |
2999 | * put in a delay to throttle resync. | |
3000 | */ | |
0a27ec96 | 3001 | if (!go_faster && conf->nr_waiting) |
1da177e4 | 3002 | msleep_interruptible(1000); |
1da177e4 LT |
3003 | |
3004 | /* Again, very different code for resync and recovery. | |
3005 | * Both must result in an r10bio with a list of bios that | |
3006 | * have bi_end_io, bi_sector, bi_bdev set, | |
3007 | * and bi_private set to the r10bio. | |
3008 | * For recovery, we may actually create several r10bios | |
3009 | * with 2 bios in each, that correspond to the bios in the main one. | |
3010 | * In this case, the subordinate r10bios link back through a | |
3011 | * borrowed master_bio pointer, and the counter in the master | |
3012 | * includes a ref from each subordinate. | |
3013 | */ | |
3014 | /* First, we decide what to do and set ->bi_end_io | |
3015 | * To end_sync_read if we want to read, and | |
3016 | * end_sync_write if we will want to write. | |
3017 | */ | |
3018 | ||
6cce3b23 | 3019 | max_sync = RESYNC_PAGES << (PAGE_SHIFT-9); |
1da177e4 LT |
3020 | if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) { |
3021 | /* recovery... the complicated one */ | |
e875ecea | 3022 | int j; |
1da177e4 LT |
3023 | r10_bio = NULL; |
3024 | ||
5cf00fcd | 3025 | for (i = 0 ; i < conf->geo.raid_disks; i++) { |
ab9d47e9 | 3026 | int still_degraded; |
9f2c9d12 | 3027 | struct r10bio *rb2; |
ab9d47e9 N |
3028 | sector_t sect; |
3029 | int must_sync; | |
e875ecea | 3030 | int any_working; |
dc280d98 | 3031 | struct raid10_info *mirror = &conf->mirrors[i]; |
24afd80d N |
3032 | |
3033 | if ((mirror->rdev == NULL || | |
3034 | test_bit(In_sync, &mirror->rdev->flags)) | |
3035 | && | |
3036 | (mirror->replacement == NULL || | |
3037 | test_bit(Faulty, | |
3038 | &mirror->replacement->flags))) | |
ab9d47e9 | 3039 | continue; |
1da177e4 | 3040 | |
ab9d47e9 N |
3041 | still_degraded = 0; |
3042 | /* want to reconstruct this device */ | |
3043 | rb2 = r10_bio; | |
3044 | sect = raid10_find_virt(conf, sector_nr, i); | |
fc448a18 N |
3045 | if (sect >= mddev->resync_max_sectors) { |
3046 | /* last stripe is not complete - don't | |
3047 | * try to recover this sector. | |
3048 | */ | |
3049 | continue; | |
3050 | } | |
24afd80d N |
3051 | /* Unless we are doing a full sync, or a replacement |
3052 | * we only need to recover the block if it is set in | |
3053 | * the bitmap | |
ab9d47e9 N |
3054 | */ |
3055 | must_sync = bitmap_start_sync(mddev->bitmap, sect, | |
3056 | &sync_blocks, 1); | |
3057 | if (sync_blocks < max_sync) | |
3058 | max_sync = sync_blocks; | |
3059 | if (!must_sync && | |
24afd80d | 3060 | mirror->replacement == NULL && |
ab9d47e9 N |
3061 | !conf->fullsync) { |
3062 | /* yep, skip the sync_blocks here, but don't assume | |
3063 | * that there will never be anything to do here | |
3064 | */ | |
3065 | chunks_skipped = -1; | |
3066 | continue; | |
3067 | } | |
6cce3b23 | 3068 | |
ab9d47e9 | 3069 | r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO); |
cb8b12b5 | 3070 | r10_bio->state = 0; |
ab9d47e9 N |
3071 | raise_barrier(conf, rb2 != NULL); |
3072 | atomic_set(&r10_bio->remaining, 0); | |
18055569 | 3073 | |
ab9d47e9 N |
3074 | r10_bio->master_bio = (struct bio*)rb2; |
3075 | if (rb2) | |
3076 | atomic_inc(&rb2->remaining); | |
3077 | r10_bio->mddev = mddev; | |
3078 | set_bit(R10BIO_IsRecover, &r10_bio->state); | |
3079 | r10_bio->sector = sect; | |
1da177e4 | 3080 | |
ab9d47e9 N |
3081 | raid10_find_phys(conf, r10_bio); |
3082 | ||
3083 | /* Need to check if the array will still be | |
3084 | * degraded | |
3085 | */ | |
5cf00fcd | 3086 | for (j = 0; j < conf->geo.raid_disks; j++) |
ab9d47e9 N |
3087 | if (conf->mirrors[j].rdev == NULL || |
3088 | test_bit(Faulty, &conf->mirrors[j].rdev->flags)) { | |
3089 | still_degraded = 1; | |
87fc767b | 3090 | break; |
1da177e4 | 3091 | } |
ab9d47e9 N |
3092 | |
3093 | must_sync = bitmap_start_sync(mddev->bitmap, sect, | |
3094 | &sync_blocks, still_degraded); | |
3095 | ||
e875ecea | 3096 | any_working = 0; |
ab9d47e9 | 3097 | for (j=0; j<conf->copies;j++) { |
e875ecea | 3098 | int k; |
ab9d47e9 | 3099 | int d = r10_bio->devs[j].devnum; |
5e570289 | 3100 | sector_t from_addr, to_addr; |
3cb03002 | 3101 | struct md_rdev *rdev; |
40c356ce N |
3102 | sector_t sector, first_bad; |
3103 | int bad_sectors; | |
ab9d47e9 N |
3104 | if (!conf->mirrors[d].rdev || |
3105 | !test_bit(In_sync, &conf->mirrors[d].rdev->flags)) | |
3106 | continue; | |
3107 | /* This is where we read from */ | |
e875ecea | 3108 | any_working = 1; |
40c356ce N |
3109 | rdev = conf->mirrors[d].rdev; |
3110 | sector = r10_bio->devs[j].addr; | |
3111 | ||
3112 | if (is_badblock(rdev, sector, max_sync, | |
3113 | &first_bad, &bad_sectors)) { | |
3114 | if (first_bad > sector) | |
3115 | max_sync = first_bad - sector; | |
3116 | else { | |
3117 | bad_sectors -= (sector | |
3118 | - first_bad); | |
3119 | if (max_sync > bad_sectors) | |
3120 | max_sync = bad_sectors; | |
3121 | continue; | |
3122 | } | |
3123 | } | |
ab9d47e9 | 3124 | bio = r10_bio->devs[0].bio; |
8be185f2 | 3125 | bio_reset(bio); |
ab9d47e9 N |
3126 | bio->bi_next = biolist; |
3127 | biolist = bio; | |
3128 | bio->bi_private = r10_bio; | |
3129 | bio->bi_end_io = end_sync_read; | |
3130 | bio->bi_rw = READ; | |
5e570289 | 3131 | from_addr = r10_bio->devs[j].addr; |
4f024f37 KO |
3132 | bio->bi_iter.bi_sector = from_addr + |
3133 | rdev->data_offset; | |
24afd80d N |
3134 | bio->bi_bdev = rdev->bdev; |
3135 | atomic_inc(&rdev->nr_pending); | |
3136 | /* and we write to 'i' (if not in_sync) */ | |
ab9d47e9 N |
3137 | |
3138 | for (k=0; k<conf->copies; k++) | |
3139 | if (r10_bio->devs[k].devnum == i) | |
3140 | break; | |
3141 | BUG_ON(k == conf->copies); | |
5e570289 | 3142 | to_addr = r10_bio->devs[k].addr; |
ab9d47e9 | 3143 | r10_bio->devs[0].devnum = d; |
5e570289 | 3144 | r10_bio->devs[0].addr = from_addr; |
ab9d47e9 | 3145 | r10_bio->devs[1].devnum = i; |
5e570289 | 3146 | r10_bio->devs[1].addr = to_addr; |
ab9d47e9 | 3147 | |
24afd80d N |
3148 | rdev = mirror->rdev; |
3149 | if (!test_bit(In_sync, &rdev->flags)) { | |
3150 | bio = r10_bio->devs[1].bio; | |
8be185f2 | 3151 | bio_reset(bio); |
24afd80d N |
3152 | bio->bi_next = biolist; |
3153 | biolist = bio; | |
3154 | bio->bi_private = r10_bio; | |
3155 | bio->bi_end_io = end_sync_write; | |
3156 | bio->bi_rw = WRITE; | |
4f024f37 | 3157 | bio->bi_iter.bi_sector = to_addr |
24afd80d N |
3158 | + rdev->data_offset; |
3159 | bio->bi_bdev = rdev->bdev; | |
3160 | atomic_inc(&r10_bio->remaining); | |
3161 | } else | |
3162 | r10_bio->devs[1].bio->bi_end_io = NULL; | |
3163 | ||
3164 | /* and maybe write to replacement */ | |
3165 | bio = r10_bio->devs[1].repl_bio; | |
3166 | if (bio) | |
3167 | bio->bi_end_io = NULL; | |
3168 | rdev = mirror->replacement; | |
3169 | /* Note: if rdev != NULL, then bio | |
3170 | * cannot be NULL as r10buf_pool_alloc will | |
3171 | * have allocated it. | |
3172 | * So the second test here is pointless. | |
3173 | * But it keeps semantic-checkers happy, and | |
3174 | * this comment keeps human reviewers | |
3175 | * happy. | |
3176 | */ | |
3177 | if (rdev == NULL || bio == NULL || | |
3178 | test_bit(Faulty, &rdev->flags)) | |
3179 | break; | |
8be185f2 | 3180 | bio_reset(bio); |
24afd80d N |
3181 | bio->bi_next = biolist; |
3182 | biolist = bio; | |
3183 | bio->bi_private = r10_bio; | |
3184 | bio->bi_end_io = end_sync_write; | |
3185 | bio->bi_rw = WRITE; | |
4f024f37 KO |
3186 | bio->bi_iter.bi_sector = to_addr + |
3187 | rdev->data_offset; | |
24afd80d N |
3188 | bio->bi_bdev = rdev->bdev; |
3189 | atomic_inc(&r10_bio->remaining); | |
ab9d47e9 N |
3190 | break; |
3191 | } | |
3192 | if (j == conf->copies) { | |
e875ecea N |
3193 | /* Cannot recover, so abort the recovery or |
3194 | * record a bad block */ | |
e875ecea N |
3195 | if (any_working) { |
3196 | /* problem is that there are bad blocks | |
3197 | * on other device(s) | |
3198 | */ | |
3199 | int k; | |
3200 | for (k = 0; k < conf->copies; k++) | |
3201 | if (r10_bio->devs[k].devnum == i) | |
3202 | break; | |
24afd80d N |
3203 | if (!test_bit(In_sync, |
3204 | &mirror->rdev->flags) | |
3205 | && !rdev_set_badblocks( | |
3206 | mirror->rdev, | |
3207 | r10_bio->devs[k].addr, | |
3208 | max_sync, 0)) | |
3209 | any_working = 0; | |
3210 | if (mirror->replacement && | |
3211 | !rdev_set_badblocks( | |
3212 | mirror->replacement, | |
e875ecea N |
3213 | r10_bio->devs[k].addr, |
3214 | max_sync, 0)) | |
3215 | any_working = 0; | |
3216 | } | |
3217 | if (!any_working) { | |
3218 | if (!test_and_set_bit(MD_RECOVERY_INTR, | |
3219 | &mddev->recovery)) | |
3220 | printk(KERN_INFO "md/raid10:%s: insufficient " | |
3221 | "working devices for recovery.\n", | |
3222 | mdname(mddev)); | |
24afd80d | 3223 | mirror->recovery_disabled |
e875ecea N |
3224 | = mddev->recovery_disabled; |
3225 | } | |
e8b84915 N |
3226 | put_buf(r10_bio); |
3227 | if (rb2) | |
3228 | atomic_dec(&rb2->remaining); | |
3229 | r10_bio = rb2; | |
ab9d47e9 | 3230 | break; |
1da177e4 | 3231 | } |
ab9d47e9 | 3232 | } |
1da177e4 LT |
3233 | if (biolist == NULL) { |
3234 | while (r10_bio) { | |
9f2c9d12 N |
3235 | struct r10bio *rb2 = r10_bio; |
3236 | r10_bio = (struct r10bio*) rb2->master_bio; | |
1da177e4 LT |
3237 | rb2->master_bio = NULL; |
3238 | put_buf(rb2); | |
3239 | } | |
3240 | goto giveup; | |
3241 | } | |
3242 | } else { | |
3243 | /* resync. Schedule a read for every block at this virt offset */ | |
3244 | int count = 0; | |
6cce3b23 | 3245 | |
78200d45 N |
3246 | bitmap_cond_end_sync(mddev->bitmap, sector_nr); |
3247 | ||
6cce3b23 N |
3248 | if (!bitmap_start_sync(mddev->bitmap, sector_nr, |
3249 | &sync_blocks, mddev->degraded) && | |
ab9d47e9 N |
3250 | !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, |
3251 | &mddev->recovery)) { | |
6cce3b23 N |
3252 | /* We can skip this block */ |
3253 | *skipped = 1; | |
3254 | return sync_blocks + sectors_skipped; | |
3255 | } | |
3256 | if (sync_blocks < max_sync) | |
3257 | max_sync = sync_blocks; | |
1da177e4 | 3258 | r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO); |
cb8b12b5 | 3259 | r10_bio->state = 0; |
1da177e4 | 3260 | |
1da177e4 LT |
3261 | r10_bio->mddev = mddev; |
3262 | atomic_set(&r10_bio->remaining, 0); | |
6cce3b23 N |
3263 | raise_barrier(conf, 0); |
3264 | conf->next_resync = sector_nr; | |
1da177e4 LT |
3265 | |
3266 | r10_bio->master_bio = NULL; | |
3267 | r10_bio->sector = sector_nr; | |
3268 | set_bit(R10BIO_IsSync, &r10_bio->state); | |
3269 | raid10_find_phys(conf, r10_bio); | |
5cf00fcd | 3270 | r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1; |
1da177e4 | 3271 | |
5cf00fcd | 3272 | for (i = 0; i < conf->copies; i++) { |
1da177e4 | 3273 | int d = r10_bio->devs[i].devnum; |
40c356ce N |
3274 | sector_t first_bad, sector; |
3275 | int bad_sectors; | |
3276 | ||
9ad1aefc N |
3277 | if (r10_bio->devs[i].repl_bio) |
3278 | r10_bio->devs[i].repl_bio->bi_end_io = NULL; | |
3279 | ||
1da177e4 | 3280 | bio = r10_bio->devs[i].bio; |
8be185f2 | 3281 | bio_reset(bio); |
af03b8e4 | 3282 | clear_bit(BIO_UPTODATE, &bio->bi_flags); |
1da177e4 | 3283 | if (conf->mirrors[d].rdev == NULL || |
b2d444d7 | 3284 | test_bit(Faulty, &conf->mirrors[d].rdev->flags)) |
1da177e4 | 3285 | continue; |
40c356ce N |
3286 | sector = r10_bio->devs[i].addr; |
3287 | if (is_badblock(conf->mirrors[d].rdev, | |
3288 | sector, max_sync, | |
3289 | &first_bad, &bad_sectors)) { | |
3290 | if (first_bad > sector) | |
3291 | max_sync = first_bad - sector; | |
3292 | else { | |
3293 | bad_sectors -= (sector - first_bad); | |
3294 | if (max_sync > bad_sectors) | |
91502f09 | 3295 | max_sync = bad_sectors; |
40c356ce N |
3296 | continue; |
3297 | } | |
3298 | } | |
1da177e4 LT |
3299 | atomic_inc(&conf->mirrors[d].rdev->nr_pending); |
3300 | atomic_inc(&r10_bio->remaining); | |
3301 | bio->bi_next = biolist; | |
3302 | biolist = bio; | |
3303 | bio->bi_private = r10_bio; | |
3304 | bio->bi_end_io = end_sync_read; | |
802ba064 | 3305 | bio->bi_rw = READ; |
4f024f37 | 3306 | bio->bi_iter.bi_sector = sector + |
1da177e4 LT |
3307 | conf->mirrors[d].rdev->data_offset; |
3308 | bio->bi_bdev = conf->mirrors[d].rdev->bdev; | |
3309 | count++; | |
9ad1aefc N |
3310 | |
3311 | if (conf->mirrors[d].replacement == NULL || | |
3312 | test_bit(Faulty, | |
3313 | &conf->mirrors[d].replacement->flags)) | |
3314 | continue; | |
3315 | ||
3316 | /* Need to set up for writing to the replacement */ | |
3317 | bio = r10_bio->devs[i].repl_bio; | |
8be185f2 | 3318 | bio_reset(bio); |
9ad1aefc N |
3319 | clear_bit(BIO_UPTODATE, &bio->bi_flags); |
3320 | ||
3321 | sector = r10_bio->devs[i].addr; | |
3322 | atomic_inc(&conf->mirrors[d].rdev->nr_pending); | |
3323 | bio->bi_next = biolist; | |
3324 | biolist = bio; | |
3325 | bio->bi_private = r10_bio; | |
3326 | bio->bi_end_io = end_sync_write; | |
3327 | bio->bi_rw = WRITE; | |
4f024f37 | 3328 | bio->bi_iter.bi_sector = sector + |
9ad1aefc N |
3329 | conf->mirrors[d].replacement->data_offset; |
3330 | bio->bi_bdev = conf->mirrors[d].replacement->bdev; | |
3331 | count++; | |
1da177e4 LT |
3332 | } |
3333 | ||
3334 | if (count < 2) { | |
3335 | for (i=0; i<conf->copies; i++) { | |
3336 | int d = r10_bio->devs[i].devnum; | |
3337 | if (r10_bio->devs[i].bio->bi_end_io) | |
ab9d47e9 N |
3338 | rdev_dec_pending(conf->mirrors[d].rdev, |
3339 | mddev); | |
9ad1aefc N |
3340 | if (r10_bio->devs[i].repl_bio && |
3341 | r10_bio->devs[i].repl_bio->bi_end_io) | |
3342 | rdev_dec_pending( | |
3343 | conf->mirrors[d].replacement, | |
3344 | mddev); | |
1da177e4 LT |
3345 | } |
3346 | put_buf(r10_bio); | |
3347 | biolist = NULL; | |
3348 | goto giveup; | |
3349 | } | |
3350 | } | |
3351 | ||
1da177e4 | 3352 | nr_sectors = 0; |
6cce3b23 N |
3353 | if (sector_nr + max_sync < max_sector) |
3354 | max_sector = sector_nr + max_sync; | |
1da177e4 LT |
3355 | do { |
3356 | struct page *page; | |
3357 | int len = PAGE_SIZE; | |
1da177e4 LT |
3358 | if (sector_nr + (len>>9) > max_sector) |
3359 | len = (max_sector - sector_nr) << 9; | |
3360 | if (len == 0) | |
3361 | break; | |
3362 | for (bio= biolist ; bio ; bio=bio->bi_next) { | |
ab9d47e9 | 3363 | struct bio *bio2; |
1da177e4 | 3364 | page = bio->bi_io_vec[bio->bi_vcnt].bv_page; |
ab9d47e9 N |
3365 | if (bio_add_page(bio, page, len, 0)) |
3366 | continue; | |
3367 | ||
3368 | /* stop here */ | |
3369 | bio->bi_io_vec[bio->bi_vcnt].bv_page = page; | |
3370 | for (bio2 = biolist; | |
3371 | bio2 && bio2 != bio; | |
3372 | bio2 = bio2->bi_next) { | |
3373 | /* remove last page from this bio */ | |
3374 | bio2->bi_vcnt--; | |
4f024f37 | 3375 | bio2->bi_iter.bi_size -= len; |
3fd83717 | 3376 | __clear_bit(BIO_SEG_VALID, &bio2->bi_flags); |
1da177e4 | 3377 | } |
ab9d47e9 | 3378 | goto bio_full; |
1da177e4 LT |
3379 | } |
3380 | nr_sectors += len>>9; | |
3381 | sector_nr += len>>9; | |
3382 | } while (biolist->bi_vcnt < RESYNC_PAGES); | |
3383 | bio_full: | |
3384 | r10_bio->sectors = nr_sectors; | |
3385 | ||
3386 | while (biolist) { | |
3387 | bio = biolist; | |
3388 | biolist = biolist->bi_next; | |
3389 | ||
3390 | bio->bi_next = NULL; | |
3391 | r10_bio = bio->bi_private; | |
3392 | r10_bio->sectors = nr_sectors; | |
3393 | ||
3394 | if (bio->bi_end_io == end_sync_read) { | |
3395 | md_sync_acct(bio->bi_bdev, nr_sectors); | |
7bb23c49 | 3396 | set_bit(BIO_UPTODATE, &bio->bi_flags); |
1da177e4 LT |
3397 | generic_make_request(bio); |
3398 | } | |
3399 | } | |
3400 | ||
57afd89f N |
3401 | if (sectors_skipped) |
3402 | /* pretend they weren't skipped, it makes | |
3403 | * no important difference in this case | |
3404 | */ | |
3405 | md_done_sync(mddev, sectors_skipped, 1); | |
3406 | ||
1da177e4 LT |
3407 | return sectors_skipped + nr_sectors; |
3408 | giveup: | |
3409 | /* There is nowhere to write, so all non-sync | |
e875ecea N |
3410 | * drives must be failed or in resync, all drives |
3411 | * have a bad block, so try the next chunk... | |
1da177e4 | 3412 | */ |
09b4068a N |
3413 | if (sector_nr + max_sync < max_sector) |
3414 | max_sector = sector_nr + max_sync; | |
3415 | ||
3416 | sectors_skipped += (max_sector - sector_nr); | |
1da177e4 LT |
3417 | chunks_skipped ++; |
3418 | sector_nr = max_sector; | |
1da177e4 | 3419 | goto skipped; |
1da177e4 LT |
3420 | } |
3421 | ||
80c3a6ce | 3422 | static sector_t |
fd01b88c | 3423 | raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks) |
80c3a6ce DW |
3424 | { |
3425 | sector_t size; | |
e879a879 | 3426 | struct r10conf *conf = mddev->private; |
80c3a6ce DW |
3427 | |
3428 | if (!raid_disks) | |
3ea7daa5 N |
3429 | raid_disks = min(conf->geo.raid_disks, |
3430 | conf->prev.raid_disks); | |
80c3a6ce | 3431 | if (!sectors) |
dab8b292 | 3432 | sectors = conf->dev_sectors; |
80c3a6ce | 3433 | |
5cf00fcd N |
3434 | size = sectors >> conf->geo.chunk_shift; |
3435 | sector_div(size, conf->geo.far_copies); | |
80c3a6ce | 3436 | size = size * raid_disks; |
5cf00fcd | 3437 | sector_div(size, conf->geo.near_copies); |
80c3a6ce | 3438 | |
5cf00fcd | 3439 | return size << conf->geo.chunk_shift; |
80c3a6ce DW |
3440 | } |
3441 | ||
6508fdbf N |
3442 | static void calc_sectors(struct r10conf *conf, sector_t size) |
3443 | { | |
3444 | /* Calculate the number of sectors-per-device that will | |
3445 | * actually be used, and set conf->dev_sectors and | |
3446 | * conf->stride | |
3447 | */ | |
3448 | ||
5cf00fcd N |
3449 | size = size >> conf->geo.chunk_shift; |
3450 | sector_div(size, conf->geo.far_copies); | |
3451 | size = size * conf->geo.raid_disks; | |
3452 | sector_div(size, conf->geo.near_copies); | |
6508fdbf N |
3453 | /* 'size' is now the number of chunks in the array */ |
3454 | /* calculate "used chunks per device" */ | |
3455 | size = size * conf->copies; | |
3456 | ||
3457 | /* We need to round up when dividing by raid_disks to | |
3458 | * get the stride size. | |
3459 | */ | |
5cf00fcd | 3460 | size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks); |
6508fdbf | 3461 | |
5cf00fcd | 3462 | conf->dev_sectors = size << conf->geo.chunk_shift; |
6508fdbf | 3463 | |
5cf00fcd N |
3464 | if (conf->geo.far_offset) |
3465 | conf->geo.stride = 1 << conf->geo.chunk_shift; | |
6508fdbf | 3466 | else { |
5cf00fcd N |
3467 | sector_div(size, conf->geo.far_copies); |
3468 | conf->geo.stride = size << conf->geo.chunk_shift; | |
6508fdbf N |
3469 | } |
3470 | } | |
dab8b292 | 3471 | |
deb200d0 N |
3472 | enum geo_type {geo_new, geo_old, geo_start}; |
3473 | static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new) | |
3474 | { | |
3475 | int nc, fc, fo; | |
3476 | int layout, chunk, disks; | |
3477 | switch (new) { | |
3478 | case geo_old: | |
3479 | layout = mddev->layout; | |
3480 | chunk = mddev->chunk_sectors; | |
3481 | disks = mddev->raid_disks - mddev->delta_disks; | |
3482 | break; | |
3483 | case geo_new: | |
3484 | layout = mddev->new_layout; | |
3485 | chunk = mddev->new_chunk_sectors; | |
3486 | disks = mddev->raid_disks; | |
3487 | break; | |
3488 | default: /* avoid 'may be unused' warnings */ | |
3489 | case geo_start: /* new when starting reshape - raid_disks not | |
3490 | * updated yet. */ | |
3491 | layout = mddev->new_layout; | |
3492 | chunk = mddev->new_chunk_sectors; | |
3493 | disks = mddev->raid_disks + mddev->delta_disks; | |
3494 | break; | |
3495 | } | |
475901af | 3496 | if (layout >> 18) |
deb200d0 N |
3497 | return -1; |
3498 | if (chunk < (PAGE_SIZE >> 9) || | |
3499 | !is_power_of_2(chunk)) | |
3500 | return -2; | |
3501 | nc = layout & 255; | |
3502 | fc = (layout >> 8) & 255; | |
3503 | fo = layout & (1<<16); | |
3504 | geo->raid_disks = disks; | |
3505 | geo->near_copies = nc; | |
3506 | geo->far_copies = fc; | |
3507 | geo->far_offset = fo; | |
475901af | 3508 | geo->far_set_size = (layout & (1<<17)) ? disks / fc : disks; |
deb200d0 N |
3509 | geo->chunk_mask = chunk - 1; |
3510 | geo->chunk_shift = ffz(~chunk); | |
3511 | return nc*fc; | |
3512 | } | |
3513 | ||
e879a879 | 3514 | static struct r10conf *setup_conf(struct mddev *mddev) |
1da177e4 | 3515 | { |
e879a879 | 3516 | struct r10conf *conf = NULL; |
dab8b292 | 3517 | int err = -EINVAL; |
deb200d0 N |
3518 | struct geom geo; |
3519 | int copies; | |
3520 | ||
3521 | copies = setup_geo(&geo, mddev, geo_new); | |
1da177e4 | 3522 | |
deb200d0 | 3523 | if (copies == -2) { |
128595ed N |
3524 | printk(KERN_ERR "md/raid10:%s: chunk size must be " |
3525 | "at least PAGE_SIZE(%ld) and be a power of 2.\n", | |
3526 | mdname(mddev), PAGE_SIZE); | |
dab8b292 | 3527 | goto out; |
1da177e4 | 3528 | } |
2604b703 | 3529 | |
deb200d0 | 3530 | if (copies < 2 || copies > mddev->raid_disks) { |
128595ed | 3531 | printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n", |
f73ea873 | 3532 | mdname(mddev), mddev->new_layout); |
1da177e4 LT |
3533 | goto out; |
3534 | } | |
dab8b292 TM |
3535 | |
3536 | err = -ENOMEM; | |
e879a879 | 3537 | conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL); |
dab8b292 | 3538 | if (!conf) |
1da177e4 | 3539 | goto out; |
dab8b292 | 3540 | |
3ea7daa5 | 3541 | /* FIXME calc properly */ |
dc280d98 | 3542 | conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks + |
78eaa0d4 | 3543 | max(0,-mddev->delta_disks)), |
dab8b292 TM |
3544 | GFP_KERNEL); |
3545 | if (!conf->mirrors) | |
3546 | goto out; | |
4443ae10 N |
3547 | |
3548 | conf->tmppage = alloc_page(GFP_KERNEL); | |
3549 | if (!conf->tmppage) | |
dab8b292 TM |
3550 | goto out; |
3551 | ||
deb200d0 N |
3552 | conf->geo = geo; |
3553 | conf->copies = copies; | |
dab8b292 TM |
3554 | conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc, |
3555 | r10bio_pool_free, conf); | |
3556 | if (!conf->r10bio_pool) | |
3557 | goto out; | |
3558 | ||
6508fdbf | 3559 | calc_sectors(conf, mddev->dev_sectors); |
3ea7daa5 N |
3560 | if (mddev->reshape_position == MaxSector) { |
3561 | conf->prev = conf->geo; | |
3562 | conf->reshape_progress = MaxSector; | |
3563 | } else { | |
3564 | if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) { | |
3565 | err = -EINVAL; | |
3566 | goto out; | |
3567 | } | |
3568 | conf->reshape_progress = mddev->reshape_position; | |
3569 | if (conf->prev.far_offset) | |
3570 | conf->prev.stride = 1 << conf->prev.chunk_shift; | |
3571 | else | |
3572 | /* far_copies must be 1 */ | |
3573 | conf->prev.stride = conf->dev_sectors; | |
3574 | } | |
e7e72bf6 | 3575 | spin_lock_init(&conf->device_lock); |
dab8b292 TM |
3576 | INIT_LIST_HEAD(&conf->retry_list); |
3577 | ||
3578 | spin_lock_init(&conf->resync_lock); | |
3579 | init_waitqueue_head(&conf->wait_barrier); | |
3580 | ||
0232605d | 3581 | conf->thread = md_register_thread(raid10d, mddev, "raid10"); |
dab8b292 TM |
3582 | if (!conf->thread) |
3583 | goto out; | |
3584 | ||
dab8b292 TM |
3585 | conf->mddev = mddev; |
3586 | return conf; | |
3587 | ||
3588 | out: | |
3ea7daa5 N |
3589 | if (err == -ENOMEM) |
3590 | printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n", | |
3591 | mdname(mddev)); | |
dab8b292 TM |
3592 | if (conf) { |
3593 | if (conf->r10bio_pool) | |
3594 | mempool_destroy(conf->r10bio_pool); | |
3595 | kfree(conf->mirrors); | |
3596 | safe_put_page(conf->tmppage); | |
3597 | kfree(conf); | |
3598 | } | |
3599 | return ERR_PTR(err); | |
3600 | } | |
3601 | ||
fd01b88c | 3602 | static int run(struct mddev *mddev) |
dab8b292 | 3603 | { |
e879a879 | 3604 | struct r10conf *conf; |
dab8b292 | 3605 | int i, disk_idx, chunk_size; |
dc280d98 | 3606 | struct raid10_info *disk; |
3cb03002 | 3607 | struct md_rdev *rdev; |
dab8b292 | 3608 | sector_t size; |
3ea7daa5 N |
3609 | sector_t min_offset_diff = 0; |
3610 | int first = 1; | |
532a2a3f | 3611 | bool discard_supported = false; |
dab8b292 TM |
3612 | |
3613 | if (mddev->private == NULL) { | |
3614 | conf = setup_conf(mddev); | |
3615 | if (IS_ERR(conf)) | |
3616 | return PTR_ERR(conf); | |
3617 | mddev->private = conf; | |
3618 | } | |
3619 | conf = mddev->private; | |
3620 | if (!conf) | |
3621 | goto out; | |
3622 | ||
dab8b292 TM |
3623 | mddev->thread = conf->thread; |
3624 | conf->thread = NULL; | |
3625 | ||
8f6c2e4b | 3626 | chunk_size = mddev->chunk_sectors << 9; |
cc4d1efd | 3627 | if (mddev->queue) { |
532a2a3f SL |
3628 | blk_queue_max_discard_sectors(mddev->queue, |
3629 | mddev->chunk_sectors); | |
5026d7a9 | 3630 | blk_queue_max_write_same_sectors(mddev->queue, 0); |
cc4d1efd JB |
3631 | blk_queue_io_min(mddev->queue, chunk_size); |
3632 | if (conf->geo.raid_disks % conf->geo.near_copies) | |
3633 | blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks); | |
3634 | else | |
3635 | blk_queue_io_opt(mddev->queue, chunk_size * | |
3636 | (conf->geo.raid_disks / conf->geo.near_copies)); | |
3637 | } | |
8f6c2e4b | 3638 | |
dafb20fa | 3639 | rdev_for_each(rdev, mddev) { |
3ea7daa5 | 3640 | long long diff; |
aba336bd | 3641 | struct request_queue *q; |
34b343cf | 3642 | |
1da177e4 | 3643 | disk_idx = rdev->raid_disk; |
f8c9e74f N |
3644 | if (disk_idx < 0) |
3645 | continue; | |
3646 | if (disk_idx >= conf->geo.raid_disks && | |
3647 | disk_idx >= conf->prev.raid_disks) | |
1da177e4 LT |
3648 | continue; |
3649 | disk = conf->mirrors + disk_idx; | |
3650 | ||
56a2559b N |
3651 | if (test_bit(Replacement, &rdev->flags)) { |
3652 | if (disk->replacement) | |
3653 | goto out_free_conf; | |
3654 | disk->replacement = rdev; | |
3655 | } else { | |
3656 | if (disk->rdev) | |
3657 | goto out_free_conf; | |
3658 | disk->rdev = rdev; | |
3659 | } | |
aba336bd N |
3660 | q = bdev_get_queue(rdev->bdev); |
3661 | if (q->merge_bvec_fn) | |
3662 | mddev->merge_check_needed = 1; | |
3ea7daa5 N |
3663 | diff = (rdev->new_data_offset - rdev->data_offset); |
3664 | if (!mddev->reshape_backwards) | |
3665 | diff = -diff; | |
3666 | if (diff < 0) | |
3667 | diff = 0; | |
3668 | if (first || diff < min_offset_diff) | |
3669 | min_offset_diff = diff; | |
56a2559b | 3670 | |
cc4d1efd JB |
3671 | if (mddev->gendisk) |
3672 | disk_stack_limits(mddev->gendisk, rdev->bdev, | |
3673 | rdev->data_offset << 9); | |
1da177e4 LT |
3674 | |
3675 | disk->head_position = 0; | |
532a2a3f SL |
3676 | |
3677 | if (blk_queue_discard(bdev_get_queue(rdev->bdev))) | |
3678 | discard_supported = true; | |
1da177e4 | 3679 | } |
3ea7daa5 | 3680 | |
ed30be07 JB |
3681 | if (mddev->queue) { |
3682 | if (discard_supported) | |
3683 | queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, | |
3684 | mddev->queue); | |
3685 | else | |
3686 | queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD, | |
3687 | mddev->queue); | |
3688 | } | |
6d508242 | 3689 | /* need to check that every block has at least one working mirror */ |
700c7213 | 3690 | if (!enough(conf, -1)) { |
128595ed | 3691 | printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n", |
6d508242 | 3692 | mdname(mddev)); |
1da177e4 LT |
3693 | goto out_free_conf; |
3694 | } | |
3695 | ||
3ea7daa5 N |
3696 | if (conf->reshape_progress != MaxSector) { |
3697 | /* must ensure that shape change is supported */ | |
3698 | if (conf->geo.far_copies != 1 && | |
3699 | conf->geo.far_offset == 0) | |
3700 | goto out_free_conf; | |
3701 | if (conf->prev.far_copies != 1 && | |
78eaa0d4 | 3702 | conf->prev.far_offset == 0) |
3ea7daa5 N |
3703 | goto out_free_conf; |
3704 | } | |
3705 | ||
1da177e4 | 3706 | mddev->degraded = 0; |
f8c9e74f N |
3707 | for (i = 0; |
3708 | i < conf->geo.raid_disks | |
3709 | || i < conf->prev.raid_disks; | |
3710 | i++) { | |
1da177e4 LT |
3711 | |
3712 | disk = conf->mirrors + i; | |
3713 | ||
56a2559b N |
3714 | if (!disk->rdev && disk->replacement) { |
3715 | /* The replacement is all we have - use it */ | |
3716 | disk->rdev = disk->replacement; | |
3717 | disk->replacement = NULL; | |
3718 | clear_bit(Replacement, &disk->rdev->flags); | |
3719 | } | |
3720 | ||
5fd6c1dc | 3721 | if (!disk->rdev || |
2e333e89 | 3722 | !test_bit(In_sync, &disk->rdev->flags)) { |
1da177e4 LT |
3723 | disk->head_position = 0; |
3724 | mddev->degraded++; | |
0b59bb64 N |
3725 | if (disk->rdev && |
3726 | disk->rdev->saved_raid_disk < 0) | |
8c2e870a | 3727 | conf->fullsync = 1; |
1da177e4 | 3728 | } |
d890fa2b | 3729 | disk->recovery_disabled = mddev->recovery_disabled - 1; |
1da177e4 LT |
3730 | } |
3731 | ||
8c6ac868 | 3732 | if (mddev->recovery_cp != MaxSector) |
128595ed | 3733 | printk(KERN_NOTICE "md/raid10:%s: not clean" |
8c6ac868 AN |
3734 | " -- starting background reconstruction\n", |
3735 | mdname(mddev)); | |
1da177e4 | 3736 | printk(KERN_INFO |
128595ed | 3737 | "md/raid10:%s: active with %d out of %d devices\n", |
5cf00fcd N |
3738 | mdname(mddev), conf->geo.raid_disks - mddev->degraded, |
3739 | conf->geo.raid_disks); | |
1da177e4 LT |
3740 | /* |
3741 | * Ok, everything is just fine now | |
3742 | */ | |
dab8b292 TM |
3743 | mddev->dev_sectors = conf->dev_sectors; |
3744 | size = raid10_size(mddev, 0, 0); | |
3745 | md_set_array_sectors(mddev, size); | |
3746 | mddev->resync_max_sectors = size; | |
1da177e4 | 3747 | |
cc4d1efd | 3748 | if (mddev->queue) { |
5cf00fcd | 3749 | int stripe = conf->geo.raid_disks * |
9d8f0363 | 3750 | ((mddev->chunk_sectors << 9) / PAGE_SIZE); |
cc4d1efd JB |
3751 | |
3752 | /* Calculate max read-ahead size. | |
3753 | * We need to readahead at least twice a whole stripe.... | |
3754 | * maybe... | |
3755 | */ | |
5cf00fcd | 3756 | stripe /= conf->geo.near_copies; |
3ea7daa5 N |
3757 | if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe) |
3758 | mddev->queue->backing_dev_info.ra_pages = 2 * stripe; | |
1da177e4 LT |
3759 | } |
3760 | ||
a91a2785 MP |
3761 | if (md_integrity_register(mddev)) |
3762 | goto out_free_conf; | |
3763 | ||
3ea7daa5 N |
3764 | if (conf->reshape_progress != MaxSector) { |
3765 | unsigned long before_length, after_length; | |
3766 | ||
3767 | before_length = ((1 << conf->prev.chunk_shift) * | |
3768 | conf->prev.far_copies); | |
3769 | after_length = ((1 << conf->geo.chunk_shift) * | |
3770 | conf->geo.far_copies); | |
3771 | ||
3772 | if (max(before_length, after_length) > min_offset_diff) { | |
3773 | /* This cannot work */ | |
3774 | printk("md/raid10: offset difference not enough to continue reshape\n"); | |
3775 | goto out_free_conf; | |
3776 | } | |
3777 | conf->offset_diff = min_offset_diff; | |
3778 | ||
3779 | conf->reshape_safe = conf->reshape_progress; | |
3780 | clear_bit(MD_RECOVERY_SYNC, &mddev->recovery); | |
3781 | clear_bit(MD_RECOVERY_CHECK, &mddev->recovery); | |
3782 | set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery); | |
3783 | set_bit(MD_RECOVERY_RUNNING, &mddev->recovery); | |
3784 | mddev->sync_thread = md_register_thread(md_do_sync, mddev, | |
3785 | "reshape"); | |
3786 | } | |
3787 | ||
1da177e4 LT |
3788 | return 0; |
3789 | ||
3790 | out_free_conf: | |
01f96c0a | 3791 | md_unregister_thread(&mddev->thread); |
1da177e4 LT |
3792 | if (conf->r10bio_pool) |
3793 | mempool_destroy(conf->r10bio_pool); | |
1345b1d8 | 3794 | safe_put_page(conf->tmppage); |
990a8baf | 3795 | kfree(conf->mirrors); |
1da177e4 LT |
3796 | kfree(conf); |
3797 | mddev->private = NULL; | |
3798 | out: | |
3799 | return -EIO; | |
3800 | } | |
3801 | ||
afa0f557 | 3802 | static void raid10_free(struct mddev *mddev, void *priv) |
1da177e4 | 3803 | { |
afa0f557 | 3804 | struct r10conf *conf = priv; |
1da177e4 | 3805 | |
1da177e4 LT |
3806 | if (conf->r10bio_pool) |
3807 | mempool_destroy(conf->r10bio_pool); | |
0fea7ed8 | 3808 | safe_put_page(conf->tmppage); |
990a8baf | 3809 | kfree(conf->mirrors); |
c4796e21 N |
3810 | kfree(conf->mirrors_old); |
3811 | kfree(conf->mirrors_new); | |
1da177e4 | 3812 | kfree(conf); |
1da177e4 LT |
3813 | } |
3814 | ||
fd01b88c | 3815 | static void raid10_quiesce(struct mddev *mddev, int state) |
6cce3b23 | 3816 | { |
e879a879 | 3817 | struct r10conf *conf = mddev->private; |
6cce3b23 N |
3818 | |
3819 | switch(state) { | |
3820 | case 1: | |
3821 | raise_barrier(conf, 0); | |
3822 | break; | |
3823 | case 0: | |
3824 | lower_barrier(conf); | |
3825 | break; | |
3826 | } | |
6cce3b23 | 3827 | } |
1da177e4 | 3828 | |
006a09a0 N |
3829 | static int raid10_resize(struct mddev *mddev, sector_t sectors) |
3830 | { | |
3831 | /* Resize of 'far' arrays is not supported. | |
3832 | * For 'near' and 'offset' arrays we can set the | |
3833 | * number of sectors used to be an appropriate multiple | |
3834 | * of the chunk size. | |
3835 | * For 'offset', this is far_copies*chunksize. | |
3836 | * For 'near' the multiplier is the LCM of | |
3837 | * near_copies and raid_disks. | |
3838 | * So if far_copies > 1 && !far_offset, fail. | |
3839 | * Else find LCM(raid_disks, near_copy)*far_copies and | |
3840 | * multiply by chunk_size. Then round to this number. | |
3841 | * This is mostly done by raid10_size() | |
3842 | */ | |
3843 | struct r10conf *conf = mddev->private; | |
3844 | sector_t oldsize, size; | |
3845 | ||
f8c9e74f N |
3846 | if (mddev->reshape_position != MaxSector) |
3847 | return -EBUSY; | |
3848 | ||
5cf00fcd | 3849 | if (conf->geo.far_copies > 1 && !conf->geo.far_offset) |
006a09a0 N |
3850 | return -EINVAL; |
3851 | ||
3852 | oldsize = raid10_size(mddev, 0, 0); | |
3853 | size = raid10_size(mddev, sectors, 0); | |
a4a6125a N |
3854 | if (mddev->external_size && |
3855 | mddev->array_sectors > size) | |
006a09a0 | 3856 | return -EINVAL; |
a4a6125a N |
3857 | if (mddev->bitmap) { |
3858 | int ret = bitmap_resize(mddev->bitmap, size, 0, 0); | |
3859 | if (ret) | |
3860 | return ret; | |
3861 | } | |
3862 | md_set_array_sectors(mddev, size); | |
006a09a0 N |
3863 | set_capacity(mddev->gendisk, mddev->array_sectors); |
3864 | revalidate_disk(mddev->gendisk); | |
3865 | if (sectors > mddev->dev_sectors && | |
3866 | mddev->recovery_cp > oldsize) { | |
3867 | mddev->recovery_cp = oldsize; | |
3868 | set_bit(MD_RECOVERY_NEEDED, &mddev->recovery); | |
3869 | } | |
6508fdbf N |
3870 | calc_sectors(conf, sectors); |
3871 | mddev->dev_sectors = conf->dev_sectors; | |
006a09a0 N |
3872 | mddev->resync_max_sectors = size; |
3873 | return 0; | |
3874 | } | |
3875 | ||
53a6ab4d | 3876 | static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs) |
dab8b292 | 3877 | { |
3cb03002 | 3878 | struct md_rdev *rdev; |
e879a879 | 3879 | struct r10conf *conf; |
dab8b292 TM |
3880 | |
3881 | if (mddev->degraded > 0) { | |
128595ed N |
3882 | printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n", |
3883 | mdname(mddev)); | |
dab8b292 TM |
3884 | return ERR_PTR(-EINVAL); |
3885 | } | |
53a6ab4d | 3886 | sector_div(size, devs); |
dab8b292 | 3887 | |
dab8b292 TM |
3888 | /* Set new parameters */ |
3889 | mddev->new_level = 10; | |
3890 | /* new layout: far_copies = 1, near_copies = 2 */ | |
3891 | mddev->new_layout = (1<<8) + 2; | |
3892 | mddev->new_chunk_sectors = mddev->chunk_sectors; | |
3893 | mddev->delta_disks = mddev->raid_disks; | |
dab8b292 TM |
3894 | mddev->raid_disks *= 2; |
3895 | /* make sure it will be not marked as dirty */ | |
3896 | mddev->recovery_cp = MaxSector; | |
53a6ab4d | 3897 | mddev->dev_sectors = size; |
dab8b292 TM |
3898 | |
3899 | conf = setup_conf(mddev); | |
02214dc5 | 3900 | if (!IS_ERR(conf)) { |
dafb20fa | 3901 | rdev_for_each(rdev, mddev) |
53a6ab4d | 3902 | if (rdev->raid_disk >= 0) { |
e93f68a1 | 3903 | rdev->new_raid_disk = rdev->raid_disk * 2; |
53a6ab4d N |
3904 | rdev->sectors = size; |
3905 | } | |
02214dc5 KW |
3906 | conf->barrier = 1; |
3907 | } | |
3908 | ||
dab8b292 TM |
3909 | return conf; |
3910 | } | |
3911 | ||
fd01b88c | 3912 | static void *raid10_takeover(struct mddev *mddev) |
dab8b292 | 3913 | { |
e373ab10 | 3914 | struct r0conf *raid0_conf; |
dab8b292 TM |
3915 | |
3916 | /* raid10 can take over: | |
3917 | * raid0 - providing it has only two drives | |
3918 | */ | |
3919 | if (mddev->level == 0) { | |
3920 | /* for raid0 takeover only one zone is supported */ | |
e373ab10 N |
3921 | raid0_conf = mddev->private; |
3922 | if (raid0_conf->nr_strip_zones > 1) { | |
128595ed N |
3923 | printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0" |
3924 | " with more than one zone.\n", | |
3925 | mdname(mddev)); | |
dab8b292 TM |
3926 | return ERR_PTR(-EINVAL); |
3927 | } | |
53a6ab4d N |
3928 | return raid10_takeover_raid0(mddev, |
3929 | raid0_conf->strip_zone->zone_end, | |
3930 | raid0_conf->strip_zone->nb_dev); | |
dab8b292 TM |
3931 | } |
3932 | return ERR_PTR(-EINVAL); | |
3933 | } | |
3934 | ||
3ea7daa5 N |
3935 | static int raid10_check_reshape(struct mddev *mddev) |
3936 | { | |
3937 | /* Called when there is a request to change | |
3938 | * - layout (to ->new_layout) | |
3939 | * - chunk size (to ->new_chunk_sectors) | |
3940 | * - raid_disks (by delta_disks) | |
3941 | * or when trying to restart a reshape that was ongoing. | |
3942 | * | |
3943 | * We need to validate the request and possibly allocate | |
3944 | * space if that might be an issue later. | |
3945 | * | |
3946 | * Currently we reject any reshape of a 'far' mode array, | |
3947 | * allow chunk size to change if new is generally acceptable, | |
3948 | * allow raid_disks to increase, and allow | |
3949 | * a switch between 'near' mode and 'offset' mode. | |
3950 | */ | |
3951 | struct r10conf *conf = mddev->private; | |
3952 | struct geom geo; | |
3953 | ||
3954 | if (conf->geo.far_copies != 1 && !conf->geo.far_offset) | |
3955 | return -EINVAL; | |
3956 | ||
3957 | if (setup_geo(&geo, mddev, geo_start) != conf->copies) | |
3958 | /* mustn't change number of copies */ | |
3959 | return -EINVAL; | |
3960 | if (geo.far_copies > 1 && !geo.far_offset) | |
3961 | /* Cannot switch to 'far' mode */ | |
3962 | return -EINVAL; | |
3963 | ||
3964 | if (mddev->array_sectors & geo.chunk_mask) | |
3965 | /* not factor of array size */ | |
3966 | return -EINVAL; | |
3967 | ||
3ea7daa5 N |
3968 | if (!enough(conf, -1)) |
3969 | return -EINVAL; | |
3970 | ||
3971 | kfree(conf->mirrors_new); | |
3972 | conf->mirrors_new = NULL; | |
3973 | if (mddev->delta_disks > 0) { | |
3974 | /* allocate new 'mirrors' list */ | |
3975 | conf->mirrors_new = kzalloc( | |
dc280d98 | 3976 | sizeof(struct raid10_info) |
3ea7daa5 N |
3977 | *(mddev->raid_disks + |
3978 | mddev->delta_disks), | |
3979 | GFP_KERNEL); | |
3980 | if (!conf->mirrors_new) | |
3981 | return -ENOMEM; | |
3982 | } | |
3983 | return 0; | |
3984 | } | |
3985 | ||
3986 | /* | |
3987 | * Need to check if array has failed when deciding whether to: | |
3988 | * - start an array | |
3989 | * - remove non-faulty devices | |
3990 | * - add a spare | |
3991 | * - allow a reshape | |
3992 | * This determination is simple when no reshape is happening. | |
3993 | * However if there is a reshape, we need to carefully check | |
3994 | * both the before and after sections. | |
3995 | * This is because some failed devices may only affect one | |
3996 | * of the two sections, and some non-in_sync devices may | |
3997 | * be insync in the section most affected by failed devices. | |
3998 | */ | |
3999 | static int calc_degraded(struct r10conf *conf) | |
4000 | { | |
4001 | int degraded, degraded2; | |
4002 | int i; | |
4003 | ||
4004 | rcu_read_lock(); | |
4005 | degraded = 0; | |
4006 | /* 'prev' section first */ | |
4007 | for (i = 0; i < conf->prev.raid_disks; i++) { | |
4008 | struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); | |
4009 | if (!rdev || test_bit(Faulty, &rdev->flags)) | |
4010 | degraded++; | |
4011 | else if (!test_bit(In_sync, &rdev->flags)) | |
4012 | /* When we can reduce the number of devices in | |
4013 | * an array, this might not contribute to | |
4014 | * 'degraded'. It does now. | |
4015 | */ | |
4016 | degraded++; | |
4017 | } | |
4018 | rcu_read_unlock(); | |
4019 | if (conf->geo.raid_disks == conf->prev.raid_disks) | |
4020 | return degraded; | |
4021 | rcu_read_lock(); | |
4022 | degraded2 = 0; | |
4023 | for (i = 0; i < conf->geo.raid_disks; i++) { | |
4024 | struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); | |
4025 | if (!rdev || test_bit(Faulty, &rdev->flags)) | |
4026 | degraded2++; | |
4027 | else if (!test_bit(In_sync, &rdev->flags)) { | |
4028 | /* If reshape is increasing the number of devices, | |
4029 | * this section has already been recovered, so | |
4030 | * it doesn't contribute to degraded. | |
4031 | * else it does. | |
4032 | */ | |
4033 | if (conf->geo.raid_disks <= conf->prev.raid_disks) | |
4034 | degraded2++; | |
4035 | } | |
4036 | } | |
4037 | rcu_read_unlock(); | |
4038 | if (degraded2 > degraded) | |
4039 | return degraded2; | |
4040 | return degraded; | |
4041 | } | |
4042 | ||
4043 | static int raid10_start_reshape(struct mddev *mddev) | |
4044 | { | |
4045 | /* A 'reshape' has been requested. This commits | |
4046 | * the various 'new' fields and sets MD_RECOVER_RESHAPE | |
4047 | * This also checks if there are enough spares and adds them | |
4048 | * to the array. | |
4049 | * We currently require enough spares to make the final | |
4050 | * array non-degraded. We also require that the difference | |
4051 | * between old and new data_offset - on each device - is | |
4052 | * enough that we never risk over-writing. | |
4053 | */ | |
4054 | ||
4055 | unsigned long before_length, after_length; | |
4056 | sector_t min_offset_diff = 0; | |
4057 | int first = 1; | |
4058 | struct geom new; | |
4059 | struct r10conf *conf = mddev->private; | |
4060 | struct md_rdev *rdev; | |
4061 | int spares = 0; | |
bb63a701 | 4062 | int ret; |
3ea7daa5 N |
4063 | |
4064 | if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery)) | |
4065 | return -EBUSY; | |
4066 | ||
4067 | if (setup_geo(&new, mddev, geo_start) != conf->copies) | |
4068 | return -EINVAL; | |
4069 | ||
4070 | before_length = ((1 << conf->prev.chunk_shift) * | |
4071 | conf->prev.far_copies); | |
4072 | after_length = ((1 << conf->geo.chunk_shift) * | |
4073 | conf->geo.far_copies); | |
4074 | ||
4075 | rdev_for_each(rdev, mddev) { | |
4076 | if (!test_bit(In_sync, &rdev->flags) | |
4077 | && !test_bit(Faulty, &rdev->flags)) | |
4078 | spares++; | |
4079 | if (rdev->raid_disk >= 0) { | |
4080 | long long diff = (rdev->new_data_offset | |
4081 | - rdev->data_offset); | |
4082 | if (!mddev->reshape_backwards) | |
4083 | diff = -diff; | |
4084 | if (diff < 0) | |
4085 | diff = 0; | |
4086 | if (first || diff < min_offset_diff) | |
4087 | min_offset_diff = diff; | |
4088 | } | |
4089 | } | |
4090 | ||
4091 | if (max(before_length, after_length) > min_offset_diff) | |
4092 | return -EINVAL; | |
4093 | ||
4094 | if (spares < mddev->delta_disks) | |
4095 | return -EINVAL; | |
4096 | ||
4097 | conf->offset_diff = min_offset_diff; | |
4098 | spin_lock_irq(&conf->device_lock); | |
4099 | if (conf->mirrors_new) { | |
4100 | memcpy(conf->mirrors_new, conf->mirrors, | |
dc280d98 | 4101 | sizeof(struct raid10_info)*conf->prev.raid_disks); |
3ea7daa5 | 4102 | smp_mb(); |
c4796e21 | 4103 | kfree(conf->mirrors_old); |
3ea7daa5 N |
4104 | conf->mirrors_old = conf->mirrors; |
4105 | conf->mirrors = conf->mirrors_new; | |
4106 | conf->mirrors_new = NULL; | |
4107 | } | |
4108 | setup_geo(&conf->geo, mddev, geo_start); | |
4109 | smp_mb(); | |
4110 | if (mddev->reshape_backwards) { | |
4111 | sector_t size = raid10_size(mddev, 0, 0); | |
4112 | if (size < mddev->array_sectors) { | |
4113 | spin_unlock_irq(&conf->device_lock); | |
4114 | printk(KERN_ERR "md/raid10:%s: array size must be reduce before number of disks\n", | |
4115 | mdname(mddev)); | |
4116 | return -EINVAL; | |
4117 | } | |
4118 | mddev->resync_max_sectors = size; | |
4119 | conf->reshape_progress = size; | |
4120 | } else | |
4121 | conf->reshape_progress = 0; | |
4122 | spin_unlock_irq(&conf->device_lock); | |
4123 | ||
bb63a701 N |
4124 | if (mddev->delta_disks && mddev->bitmap) { |
4125 | ret = bitmap_resize(mddev->bitmap, | |
4126 | raid10_size(mddev, 0, | |
4127 | conf->geo.raid_disks), | |
4128 | 0, 0); | |
4129 | if (ret) | |
4130 | goto abort; | |
4131 | } | |
3ea7daa5 N |
4132 | if (mddev->delta_disks > 0) { |
4133 | rdev_for_each(rdev, mddev) | |
4134 | if (rdev->raid_disk < 0 && | |
4135 | !test_bit(Faulty, &rdev->flags)) { | |
4136 | if (raid10_add_disk(mddev, rdev) == 0) { | |
4137 | if (rdev->raid_disk >= | |
4138 | conf->prev.raid_disks) | |
4139 | set_bit(In_sync, &rdev->flags); | |
4140 | else | |
4141 | rdev->recovery_offset = 0; | |
4142 | ||
4143 | if (sysfs_link_rdev(mddev, rdev)) | |
4144 | /* Failure here is OK */; | |
4145 | } | |
4146 | } else if (rdev->raid_disk >= conf->prev.raid_disks | |
4147 | && !test_bit(Faulty, &rdev->flags)) { | |
4148 | /* This is a spare that was manually added */ | |
4149 | set_bit(In_sync, &rdev->flags); | |
4150 | } | |
4151 | } | |
4152 | /* When a reshape changes the number of devices, | |
4153 | * ->degraded is measured against the larger of the | |
4154 | * pre and post numbers. | |
4155 | */ | |
4156 | spin_lock_irq(&conf->device_lock); | |
4157 | mddev->degraded = calc_degraded(conf); | |
4158 | spin_unlock_irq(&conf->device_lock); | |
4159 | mddev->raid_disks = conf->geo.raid_disks; | |
4160 | mddev->reshape_position = conf->reshape_progress; | |
4161 | set_bit(MD_CHANGE_DEVS, &mddev->flags); | |
4162 | ||
4163 | clear_bit(MD_RECOVERY_SYNC, &mddev->recovery); | |
4164 | clear_bit(MD_RECOVERY_CHECK, &mddev->recovery); | |
4165 | set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery); | |
4166 | set_bit(MD_RECOVERY_RUNNING, &mddev->recovery); | |
4167 | ||
4168 | mddev->sync_thread = md_register_thread(md_do_sync, mddev, | |
4169 | "reshape"); | |
4170 | if (!mddev->sync_thread) { | |
bb63a701 N |
4171 | ret = -EAGAIN; |
4172 | goto abort; | |
3ea7daa5 N |
4173 | } |
4174 | conf->reshape_checkpoint = jiffies; | |
4175 | md_wakeup_thread(mddev->sync_thread); | |
4176 | md_new_event(mddev); | |
4177 | return 0; | |
bb63a701 N |
4178 | |
4179 | abort: | |
4180 | mddev->recovery = 0; | |
4181 | spin_lock_irq(&conf->device_lock); | |
4182 | conf->geo = conf->prev; | |
4183 | mddev->raid_disks = conf->geo.raid_disks; | |
4184 | rdev_for_each(rdev, mddev) | |
4185 | rdev->new_data_offset = rdev->data_offset; | |
4186 | smp_wmb(); | |
4187 | conf->reshape_progress = MaxSector; | |
4188 | mddev->reshape_position = MaxSector; | |
4189 | spin_unlock_irq(&conf->device_lock); | |
4190 | return ret; | |
3ea7daa5 N |
4191 | } |
4192 | ||
4193 | /* Calculate the last device-address that could contain | |
4194 | * any block from the chunk that includes the array-address 's' | |
4195 | * and report the next address. | |
4196 | * i.e. the address returned will be chunk-aligned and after | |
4197 | * any data that is in the chunk containing 's'. | |
4198 | */ | |
4199 | static sector_t last_dev_address(sector_t s, struct geom *geo) | |
4200 | { | |
4201 | s = (s | geo->chunk_mask) + 1; | |
4202 | s >>= geo->chunk_shift; | |
4203 | s *= geo->near_copies; | |
4204 | s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks); | |
4205 | s *= geo->far_copies; | |
4206 | s <<= geo->chunk_shift; | |
4207 | return s; | |
4208 | } | |
4209 | ||
4210 | /* Calculate the first device-address that could contain | |
4211 | * any block from the chunk that includes the array-address 's'. | |
4212 | * This too will be the start of a chunk | |
4213 | */ | |
4214 | static sector_t first_dev_address(sector_t s, struct geom *geo) | |
4215 | { | |
4216 | s >>= geo->chunk_shift; | |
4217 | s *= geo->near_copies; | |
4218 | sector_div(s, geo->raid_disks); | |
4219 | s *= geo->far_copies; | |
4220 | s <<= geo->chunk_shift; | |
4221 | return s; | |
4222 | } | |
4223 | ||
4224 | static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, | |
4225 | int *skipped) | |
4226 | { | |
4227 | /* We simply copy at most one chunk (smallest of old and new) | |
4228 | * at a time, possibly less if that exceeds RESYNC_PAGES, | |
4229 | * or we hit a bad block or something. | |
4230 | * This might mean we pause for normal IO in the middle of | |
4231 | * a chunk, but that is not a problem was mddev->reshape_position | |
4232 | * can record any location. | |
4233 | * | |
4234 | * If we will want to write to a location that isn't | |
4235 | * yet recorded as 'safe' (i.e. in metadata on disk) then | |
4236 | * we need to flush all reshape requests and update the metadata. | |
4237 | * | |
4238 | * When reshaping forwards (e.g. to more devices), we interpret | |
4239 | * 'safe' as the earliest block which might not have been copied | |
4240 | * down yet. We divide this by previous stripe size and multiply | |
4241 | * by previous stripe length to get lowest device offset that we | |
4242 | * cannot write to yet. | |
4243 | * We interpret 'sector_nr' as an address that we want to write to. | |
4244 | * From this we use last_device_address() to find where we might | |
4245 | * write to, and first_device_address on the 'safe' position. | |
4246 | * If this 'next' write position is after the 'safe' position, | |
4247 | * we must update the metadata to increase the 'safe' position. | |
4248 | * | |
4249 | * When reshaping backwards, we round in the opposite direction | |
4250 | * and perform the reverse test: next write position must not be | |
4251 | * less than current safe position. | |
4252 | * | |
4253 | * In all this the minimum difference in data offsets | |
4254 | * (conf->offset_diff - always positive) allows a bit of slack, | |
4255 | * so next can be after 'safe', but not by more than offset_disk | |
4256 | * | |
4257 | * We need to prepare all the bios here before we start any IO | |
4258 | * to ensure the size we choose is acceptable to all devices. | |
4259 | * The means one for each copy for write-out and an extra one for | |
4260 | * read-in. | |
4261 | * We store the read-in bio in ->master_bio and the others in | |
4262 | * ->devs[x].bio and ->devs[x].repl_bio. | |
4263 | */ | |
4264 | struct r10conf *conf = mddev->private; | |
4265 | struct r10bio *r10_bio; | |
4266 | sector_t next, safe, last; | |
4267 | int max_sectors; | |
4268 | int nr_sectors; | |
4269 | int s; | |
4270 | struct md_rdev *rdev; | |
4271 | int need_flush = 0; | |
4272 | struct bio *blist; | |
4273 | struct bio *bio, *read_bio; | |
4274 | int sectors_done = 0; | |
4275 | ||
4276 | if (sector_nr == 0) { | |
4277 | /* If restarting in the middle, skip the initial sectors */ | |
4278 | if (mddev->reshape_backwards && | |
4279 | conf->reshape_progress < raid10_size(mddev, 0, 0)) { | |
4280 | sector_nr = (raid10_size(mddev, 0, 0) | |
4281 | - conf->reshape_progress); | |
4282 | } else if (!mddev->reshape_backwards && | |
4283 | conf->reshape_progress > 0) | |
4284 | sector_nr = conf->reshape_progress; | |
4285 | if (sector_nr) { | |
4286 | mddev->curr_resync_completed = sector_nr; | |
4287 | sysfs_notify(&mddev->kobj, NULL, "sync_completed"); | |
4288 | *skipped = 1; | |
4289 | return sector_nr; | |
4290 | } | |
4291 | } | |
4292 | ||
4293 | /* We don't use sector_nr to track where we are up to | |
4294 | * as that doesn't work well for ->reshape_backwards. | |
4295 | * So just use ->reshape_progress. | |
4296 | */ | |
4297 | if (mddev->reshape_backwards) { | |
4298 | /* 'next' is the earliest device address that we might | |
4299 | * write to for this chunk in the new layout | |
4300 | */ | |
4301 | next = first_dev_address(conf->reshape_progress - 1, | |
4302 | &conf->geo); | |
4303 | ||
4304 | /* 'safe' is the last device address that we might read from | |
4305 | * in the old layout after a restart | |
4306 | */ | |
4307 | safe = last_dev_address(conf->reshape_safe - 1, | |
4308 | &conf->prev); | |
4309 | ||
4310 | if (next + conf->offset_diff < safe) | |
4311 | need_flush = 1; | |
4312 | ||
4313 | last = conf->reshape_progress - 1; | |
4314 | sector_nr = last & ~(sector_t)(conf->geo.chunk_mask | |
4315 | & conf->prev.chunk_mask); | |
4316 | if (sector_nr + RESYNC_BLOCK_SIZE/512 < last) | |
4317 | sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512; | |
4318 | } else { | |
4319 | /* 'next' is after the last device address that we | |
4320 | * might write to for this chunk in the new layout | |
4321 | */ | |
4322 | next = last_dev_address(conf->reshape_progress, &conf->geo); | |
4323 | ||
4324 | /* 'safe' is the earliest device address that we might | |
4325 | * read from in the old layout after a restart | |
4326 | */ | |
4327 | safe = first_dev_address(conf->reshape_safe, &conf->prev); | |
4328 | ||
4329 | /* Need to update metadata if 'next' might be beyond 'safe' | |
4330 | * as that would possibly corrupt data | |
4331 | */ | |
4332 | if (next > safe + conf->offset_diff) | |
4333 | need_flush = 1; | |
4334 | ||
4335 | sector_nr = conf->reshape_progress; | |
4336 | last = sector_nr | (conf->geo.chunk_mask | |
4337 | & conf->prev.chunk_mask); | |
4338 | ||
4339 | if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last) | |
4340 | last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1; | |
4341 | } | |
4342 | ||
4343 | if (need_flush || | |
4344 | time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) { | |
4345 | /* Need to update reshape_position in metadata */ | |
4346 | wait_barrier(conf); | |
4347 | mddev->reshape_position = conf->reshape_progress; | |
4348 | if (mddev->reshape_backwards) | |
4349 | mddev->curr_resync_completed = raid10_size(mddev, 0, 0) | |
4350 | - conf->reshape_progress; | |
4351 | else | |
4352 | mddev->curr_resync_completed = conf->reshape_progress; | |
4353 | conf->reshape_checkpoint = jiffies; | |
4354 | set_bit(MD_CHANGE_DEVS, &mddev->flags); | |
4355 | md_wakeup_thread(mddev->thread); | |
4356 | wait_event(mddev->sb_wait, mddev->flags == 0 || | |
c91abf5a N |
4357 | test_bit(MD_RECOVERY_INTR, &mddev->recovery)); |
4358 | if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) { | |
4359 | allow_barrier(conf); | |
4360 | return sectors_done; | |
4361 | } | |
3ea7daa5 N |
4362 | conf->reshape_safe = mddev->reshape_position; |
4363 | allow_barrier(conf); | |
4364 | } | |
4365 | ||
4366 | read_more: | |
4367 | /* Now schedule reads for blocks from sector_nr to last */ | |
4368 | r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO); | |
cb8b12b5 | 4369 | r10_bio->state = 0; |
3ea7daa5 N |
4370 | raise_barrier(conf, sectors_done != 0); |
4371 | atomic_set(&r10_bio->remaining, 0); | |
4372 | r10_bio->mddev = mddev; | |
4373 | r10_bio->sector = sector_nr; | |
4374 | set_bit(R10BIO_IsReshape, &r10_bio->state); | |
4375 | r10_bio->sectors = last - sector_nr + 1; | |
4376 | rdev = read_balance(conf, r10_bio, &max_sectors); | |
4377 | BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state)); | |
4378 | ||
4379 | if (!rdev) { | |
4380 | /* Cannot read from here, so need to record bad blocks | |
4381 | * on all the target devices. | |
4382 | */ | |
4383 | // FIXME | |
e337aead | 4384 | mempool_free(r10_bio, conf->r10buf_pool); |
3ea7daa5 N |
4385 | set_bit(MD_RECOVERY_INTR, &mddev->recovery); |
4386 | return sectors_done; | |
4387 | } | |
4388 | ||
4389 | read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev); | |
4390 | ||
4391 | read_bio->bi_bdev = rdev->bdev; | |
4f024f37 | 4392 | read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr |
3ea7daa5 N |
4393 | + rdev->data_offset); |
4394 | read_bio->bi_private = r10_bio; | |
4395 | read_bio->bi_end_io = end_sync_read; | |
4396 | read_bio->bi_rw = READ; | |
ce0b0a46 | 4397 | read_bio->bi_flags &= (~0UL << BIO_RESET_BITS); |
3fd83717 | 4398 | __set_bit(BIO_UPTODATE, &read_bio->bi_flags); |
3ea7daa5 | 4399 | read_bio->bi_vcnt = 0; |
4f024f37 | 4400 | read_bio->bi_iter.bi_size = 0; |
3ea7daa5 N |
4401 | r10_bio->master_bio = read_bio; |
4402 | r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum; | |
4403 | ||
4404 | /* Now find the locations in the new layout */ | |
4405 | __raid10_find_phys(&conf->geo, r10_bio); | |
4406 | ||
4407 | blist = read_bio; | |
4408 | read_bio->bi_next = NULL; | |
4409 | ||
4410 | for (s = 0; s < conf->copies*2; s++) { | |
4411 | struct bio *b; | |
4412 | int d = r10_bio->devs[s/2].devnum; | |
4413 | struct md_rdev *rdev2; | |
4414 | if (s&1) { | |
4415 | rdev2 = conf->mirrors[d].replacement; | |
4416 | b = r10_bio->devs[s/2].repl_bio; | |
4417 | } else { | |
4418 | rdev2 = conf->mirrors[d].rdev; | |
4419 | b = r10_bio->devs[s/2].bio; | |
4420 | } | |
4421 | if (!rdev2 || test_bit(Faulty, &rdev2->flags)) | |
4422 | continue; | |
8be185f2 KO |
4423 | |
4424 | bio_reset(b); | |
3ea7daa5 | 4425 | b->bi_bdev = rdev2->bdev; |
4f024f37 KO |
4426 | b->bi_iter.bi_sector = r10_bio->devs[s/2].addr + |
4427 | rdev2->new_data_offset; | |
3ea7daa5 N |
4428 | b->bi_private = r10_bio; |
4429 | b->bi_end_io = end_reshape_write; | |
4430 | b->bi_rw = WRITE; | |
3ea7daa5 | 4431 | b->bi_next = blist; |
3ea7daa5 N |
4432 | blist = b; |
4433 | } | |
4434 | ||
4435 | /* Now add as many pages as possible to all of these bios. */ | |
4436 | ||
4437 | nr_sectors = 0; | |
4438 | for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) { | |
4439 | struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page; | |
4440 | int len = (max_sectors - s) << 9; | |
4441 | if (len > PAGE_SIZE) | |
4442 | len = PAGE_SIZE; | |
4443 | for (bio = blist; bio ; bio = bio->bi_next) { | |
4444 | struct bio *bio2; | |
4445 | if (bio_add_page(bio, page, len, 0)) | |
4446 | continue; | |
4447 | ||
4448 | /* Didn't fit, must stop */ | |
4449 | for (bio2 = blist; | |
4450 | bio2 && bio2 != bio; | |
4451 | bio2 = bio2->bi_next) { | |
4452 | /* Remove last page from this bio */ | |
4453 | bio2->bi_vcnt--; | |
4f024f37 | 4454 | bio2->bi_iter.bi_size -= len; |
3fd83717 | 4455 | __clear_bit(BIO_SEG_VALID, &bio2->bi_flags); |
3ea7daa5 N |
4456 | } |
4457 | goto bio_full; | |
4458 | } | |
4459 | sector_nr += len >> 9; | |
4460 | nr_sectors += len >> 9; | |
4461 | } | |
4462 | bio_full: | |
4463 | r10_bio->sectors = nr_sectors; | |
4464 | ||
4465 | /* Now submit the read */ | |
4466 | md_sync_acct(read_bio->bi_bdev, r10_bio->sectors); | |
4467 | atomic_inc(&r10_bio->remaining); | |
4468 | read_bio->bi_next = NULL; | |
4469 | generic_make_request(read_bio); | |
4470 | sector_nr += nr_sectors; | |
4471 | sectors_done += nr_sectors; | |
4472 | if (sector_nr <= last) | |
4473 | goto read_more; | |
4474 | ||
4475 | /* Now that we have done the whole section we can | |
4476 | * update reshape_progress | |
4477 | */ | |
4478 | if (mddev->reshape_backwards) | |
4479 | conf->reshape_progress -= sectors_done; | |
4480 | else | |
4481 | conf->reshape_progress += sectors_done; | |
4482 | ||
4483 | return sectors_done; | |
4484 | } | |
4485 | ||
4486 | static void end_reshape_request(struct r10bio *r10_bio); | |
4487 | static int handle_reshape_read_error(struct mddev *mddev, | |
4488 | struct r10bio *r10_bio); | |
4489 | static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio) | |
4490 | { | |
4491 | /* Reshape read completed. Hopefully we have a block | |
4492 | * to write out. | |
4493 | * If we got a read error then we do sync 1-page reads from | |
4494 | * elsewhere until we find the data - or give up. | |
4495 | */ | |
4496 | struct r10conf *conf = mddev->private; | |
4497 | int s; | |
4498 | ||
4499 | if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) | |
4500 | if (handle_reshape_read_error(mddev, r10_bio) < 0) { | |
4501 | /* Reshape has been aborted */ | |
4502 | md_done_sync(mddev, r10_bio->sectors, 0); | |
4503 | return; | |
4504 | } | |
4505 | ||
4506 | /* We definitely have the data in the pages, schedule the | |
4507 | * writes. | |
4508 | */ | |
4509 | atomic_set(&r10_bio->remaining, 1); | |
4510 | for (s = 0; s < conf->copies*2; s++) { | |
4511 | struct bio *b; | |
4512 | int d = r10_bio->devs[s/2].devnum; | |
4513 | struct md_rdev *rdev; | |
4514 | if (s&1) { | |
4515 | rdev = conf->mirrors[d].replacement; | |
4516 | b = r10_bio->devs[s/2].repl_bio; | |
4517 | } else { | |
4518 | rdev = conf->mirrors[d].rdev; | |
4519 | b = r10_bio->devs[s/2].bio; | |
4520 | } | |
4521 | if (!rdev || test_bit(Faulty, &rdev->flags)) | |
4522 | continue; | |
4523 | atomic_inc(&rdev->nr_pending); | |
4524 | md_sync_acct(b->bi_bdev, r10_bio->sectors); | |
4525 | atomic_inc(&r10_bio->remaining); | |
4526 | b->bi_next = NULL; | |
4527 | generic_make_request(b); | |
4528 | } | |
4529 | end_reshape_request(r10_bio); | |
4530 | } | |
4531 | ||
4532 | static void end_reshape(struct r10conf *conf) | |
4533 | { | |
4534 | if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) | |
4535 | return; | |
4536 | ||
4537 | spin_lock_irq(&conf->device_lock); | |
4538 | conf->prev = conf->geo; | |
4539 | md_finish_reshape(conf->mddev); | |
4540 | smp_wmb(); | |
4541 | conf->reshape_progress = MaxSector; | |
4542 | spin_unlock_irq(&conf->device_lock); | |
4543 | ||
4544 | /* read-ahead size must cover two whole stripes, which is | |
4545 | * 2 * (datadisks) * chunksize where 'n' is the number of raid devices | |
4546 | */ | |
4547 | if (conf->mddev->queue) { | |
4548 | int stripe = conf->geo.raid_disks * | |
4549 | ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE); | |
4550 | stripe /= conf->geo.near_copies; | |
4551 | if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe) | |
4552 | conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe; | |
4553 | } | |
4554 | conf->fullsync = 0; | |
4555 | } | |
4556 | ||
3ea7daa5 N |
4557 | static int handle_reshape_read_error(struct mddev *mddev, |
4558 | struct r10bio *r10_bio) | |
4559 | { | |
4560 | /* Use sync reads to get the blocks from somewhere else */ | |
4561 | int sectors = r10_bio->sectors; | |
3ea7daa5 | 4562 | struct r10conf *conf = mddev->private; |
e0ee7785 N |
4563 | struct { |
4564 | struct r10bio r10_bio; | |
4565 | struct r10dev devs[conf->copies]; | |
4566 | } on_stack; | |
4567 | struct r10bio *r10b = &on_stack.r10_bio; | |
3ea7daa5 N |
4568 | int slot = 0; |
4569 | int idx = 0; | |
4570 | struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec; | |
4571 | ||
e0ee7785 N |
4572 | r10b->sector = r10_bio->sector; |
4573 | __raid10_find_phys(&conf->prev, r10b); | |
3ea7daa5 N |
4574 | |
4575 | while (sectors) { | |
4576 | int s = sectors; | |
4577 | int success = 0; | |
4578 | int first_slot = slot; | |
4579 | ||
4580 | if (s > (PAGE_SIZE >> 9)) | |
4581 | s = PAGE_SIZE >> 9; | |
4582 | ||
4583 | while (!success) { | |
e0ee7785 | 4584 | int d = r10b->devs[slot].devnum; |
3ea7daa5 N |
4585 | struct md_rdev *rdev = conf->mirrors[d].rdev; |
4586 | sector_t addr; | |
4587 | if (rdev == NULL || | |
4588 | test_bit(Faulty, &rdev->flags) || | |
4589 | !test_bit(In_sync, &rdev->flags)) | |
4590 | goto failed; | |
4591 | ||
e0ee7785 | 4592 | addr = r10b->devs[slot].addr + idx * PAGE_SIZE; |
3ea7daa5 N |
4593 | success = sync_page_io(rdev, |
4594 | addr, | |
4595 | s << 9, | |
4596 | bvec[idx].bv_page, | |
4597 | READ, false); | |
4598 | if (success) | |
4599 | break; | |
4600 | failed: | |
4601 | slot++; | |
4602 | if (slot >= conf->copies) | |
4603 | slot = 0; | |
4604 | if (slot == first_slot) | |
4605 | break; | |
4606 | } | |
4607 | if (!success) { | |
4608 | /* couldn't read this block, must give up */ | |
4609 | set_bit(MD_RECOVERY_INTR, | |
4610 | &mddev->recovery); | |
4611 | return -EIO; | |
4612 | } | |
4613 | sectors -= s; | |
4614 | idx++; | |
4615 | } | |
4616 | return 0; | |
4617 | } | |
4618 | ||
4619 | static void end_reshape_write(struct bio *bio, int error) | |
4620 | { | |
4621 | int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags); | |
4622 | struct r10bio *r10_bio = bio->bi_private; | |
4623 | struct mddev *mddev = r10_bio->mddev; | |
4624 | struct r10conf *conf = mddev->private; | |
4625 | int d; | |
4626 | int slot; | |
4627 | int repl; | |
4628 | struct md_rdev *rdev = NULL; | |
4629 | ||
4630 | d = find_bio_disk(conf, r10_bio, bio, &slot, &repl); | |
4631 | if (repl) | |
4632 | rdev = conf->mirrors[d].replacement; | |
4633 | if (!rdev) { | |
4634 | smp_mb(); | |
4635 | rdev = conf->mirrors[d].rdev; | |
4636 | } | |
4637 | ||
4638 | if (!uptodate) { | |
4639 | /* FIXME should record badblock */ | |
4640 | md_error(mddev, rdev); | |
4641 | } | |
4642 | ||
4643 | rdev_dec_pending(rdev, mddev); | |
4644 | end_reshape_request(r10_bio); | |
4645 | } | |
4646 | ||
4647 | static void end_reshape_request(struct r10bio *r10_bio) | |
4648 | { | |
4649 | if (!atomic_dec_and_test(&r10_bio->remaining)) | |
4650 | return; | |
4651 | md_done_sync(r10_bio->mddev, r10_bio->sectors, 1); | |
4652 | bio_put(r10_bio->master_bio); | |
4653 | put_buf(r10_bio); | |
4654 | } | |
4655 | ||
4656 | static void raid10_finish_reshape(struct mddev *mddev) | |
4657 | { | |
4658 | struct r10conf *conf = mddev->private; | |
4659 | ||
4660 | if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) | |
4661 | return; | |
4662 | ||
4663 | if (mddev->delta_disks > 0) { | |
4664 | sector_t size = raid10_size(mddev, 0, 0); | |
4665 | md_set_array_sectors(mddev, size); | |
4666 | if (mddev->recovery_cp > mddev->resync_max_sectors) { | |
4667 | mddev->recovery_cp = mddev->resync_max_sectors; | |
4668 | set_bit(MD_RECOVERY_NEEDED, &mddev->recovery); | |
4669 | } | |
4670 | mddev->resync_max_sectors = size; | |
4671 | set_capacity(mddev->gendisk, mddev->array_sectors); | |
4672 | revalidate_disk(mddev->gendisk); | |
63aced61 N |
4673 | } else { |
4674 | int d; | |
4675 | for (d = conf->geo.raid_disks ; | |
4676 | d < conf->geo.raid_disks - mddev->delta_disks; | |
4677 | d++) { | |
4678 | struct md_rdev *rdev = conf->mirrors[d].rdev; | |
4679 | if (rdev) | |
4680 | clear_bit(In_sync, &rdev->flags); | |
4681 | rdev = conf->mirrors[d].replacement; | |
4682 | if (rdev) | |
4683 | clear_bit(In_sync, &rdev->flags); | |
4684 | } | |
3ea7daa5 N |
4685 | } |
4686 | mddev->layout = mddev->new_layout; | |
4687 | mddev->chunk_sectors = 1 << conf->geo.chunk_shift; | |
4688 | mddev->reshape_position = MaxSector; | |
4689 | mddev->delta_disks = 0; | |
4690 | mddev->reshape_backwards = 0; | |
4691 | } | |
4692 | ||
84fc4b56 | 4693 | static struct md_personality raid10_personality = |
1da177e4 LT |
4694 | { |
4695 | .name = "raid10", | |
2604b703 | 4696 | .level = 10, |
1da177e4 LT |
4697 | .owner = THIS_MODULE, |
4698 | .make_request = make_request, | |
4699 | .run = run, | |
afa0f557 | 4700 | .free = raid10_free, |
1da177e4 LT |
4701 | .status = status, |
4702 | .error_handler = error, | |
4703 | .hot_add_disk = raid10_add_disk, | |
4704 | .hot_remove_disk= raid10_remove_disk, | |
4705 | .spare_active = raid10_spare_active, | |
4706 | .sync_request = sync_request, | |
6cce3b23 | 4707 | .quiesce = raid10_quiesce, |
80c3a6ce | 4708 | .size = raid10_size, |
006a09a0 | 4709 | .resize = raid10_resize, |
dab8b292 | 4710 | .takeover = raid10_takeover, |
3ea7daa5 N |
4711 | .check_reshape = raid10_check_reshape, |
4712 | .start_reshape = raid10_start_reshape, | |
4713 | .finish_reshape = raid10_finish_reshape, | |
5c675f83 | 4714 | .congested = raid10_congested, |
64590f45 | 4715 | .mergeable_bvec = raid10_mergeable_bvec, |
1da177e4 LT |
4716 | }; |
4717 | ||
4718 | static int __init raid_init(void) | |
4719 | { | |
2604b703 | 4720 | return register_md_personality(&raid10_personality); |
1da177e4 LT |
4721 | } |
4722 | ||
4723 | static void raid_exit(void) | |
4724 | { | |
2604b703 | 4725 | unregister_md_personality(&raid10_personality); |
1da177e4 LT |
4726 | } |
4727 | ||
4728 | module_init(raid_init); | |
4729 | module_exit(raid_exit); | |
4730 | MODULE_LICENSE("GPL"); | |
0efb9e61 | 4731 | MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD"); |
1da177e4 | 4732 | MODULE_ALIAS("md-personality-9"); /* RAID10 */ |
d9d166c2 | 4733 | MODULE_ALIAS("md-raid10"); |
2604b703 | 4734 | MODULE_ALIAS("md-level-10"); |
34db0cd6 N |
4735 | |
4736 | module_param(max_queued_requests, int, S_IRUGO|S_IWUSR); |