projects / GitHub / mt8127 / android_kernel_alcatel_ttab.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
drivers: power: report battery voltage in AOSP compatible format
[GitHub/mt8127/android_kernel_alcatel_ttab.git] / drivers / md / raid10.c
1 /*
2 * raid10.c : Multiple Devices driver for Linux
3 *
4 * Copyright (C) 2000-2004 Neil Brown
5 *
6 * RAID-10 support for md.
7 *
8 * Base on code in raid1.c. See raid1.c for further copyright information.
9 *
10 *
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
14 * any later version.
15 *
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19 */
20
21 #include <linux/slab.h>
22 #include <linux/delay.h>
23 #include <linux/blkdev.h>
24 #include <linux/module.h>
25 #include <linux/seq_file.h>
26 #include <linux/ratelimit.h>
27 #include <linux/kthread.h>
28 #include "md.h"
29 #include "raid10.h"
30 #include "raid0.h"
31 #include "bitmap.h"
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)
40 * far_offset (stored in bit 16 of layout )
41 * use_far_sets (stored in bit 17 of layout )
42 *
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.
51 *
52 * If far_offset is true, then the far_copies are handled a bit differently.
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]
71 */
72
73 /*
74 * Number of guaranteed r10bios in case of extreme VM load:
75 */
76 #define NR_RAID10_BIOS 256
77
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
93 * the raid10 thread, we become 'congested' to provide back-pressure
94 * for writeback.
95 */
96 static int max_queued_requests = 1024;
97
98 static void allow_barrier(struct r10conf *conf);
99 static void lower_barrier(struct r10conf *conf);
100 static int enough(struct r10conf *conf, int ignore);
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);
106
107 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
108 {
109 struct r10conf *conf = data;
110 int size = offsetof(struct r10bio, devs[conf->copies]);
111
112 /* allocate a r10bio with room for raid_disks entries in the
113 * bios array */
114 return kzalloc(size, gfp_flags);
115 }
116
117 static void r10bio_pool_free(void *r10_bio, void *data)
118 {
119 kfree(r10_bio);
120 }
121
122 /* Maximum size of each resync request */
123 #define RESYNC_BLOCK_SIZE (64*1024)
124 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
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)
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 */
137 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
138 {
139 struct r10conf *conf = data;
140 struct page *page;
141 struct r10bio *r10_bio;
142 struct bio *bio;
143 int i, j;
144 int nalloc;
145
146 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
147 if (!r10_bio)
148 return NULL;
149
150 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
151 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
152 nalloc = conf->copies; /* resync */
153 else
154 nalloc = 2; /* recovery */
155
156 /*
157 * Allocate bios.
158 */
159 for (j = nalloc ; j-- ; ) {
160 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
161 if (!bio)
162 goto out_free_bio;
163 r10_bio->devs[j].bio = bio;
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;
170 }
171 /*
172 * Allocate RESYNC_PAGES data pages and attach them
173 * where needed.
174 */
175 for (j = 0 ; j < nalloc; j++) {
176 struct bio *rbio = r10_bio->devs[j].repl_bio;
177 bio = r10_bio->devs[j].bio;
178 for (i = 0; i < RESYNC_PAGES; i++) {
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 */
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);
188 if (unlikely(!page))
189 goto out_free_pages;
190
191 bio->bi_io_vec[i].bv_page = page;
192 if (rbio)
193 rbio->bi_io_vec[i].bv_page = page;
194 }
195 }
196
197 return r10_bio;
198
199 out_free_pages:
200 for ( ; i > 0 ; i--)
201 safe_put_page(bio->bi_io_vec[i-1].bv_page);
202 while (j--)
203 for (i = 0; i < RESYNC_PAGES ; i++)
204 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
205 j = 0;
206 out_free_bio:
207 for ( ; j < nalloc; j++) {
208 if (r10_bio->devs[j].bio)
209 bio_put(r10_bio->devs[j].bio);
210 if (r10_bio->devs[j].repl_bio)
211 bio_put(r10_bio->devs[j].repl_bio);
212 }
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;
220 struct r10conf *conf = data;
221 struct r10bio *r10bio = __r10_bio;
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++) {
228 safe_put_page(bio->bi_io_vec[i].bv_page);
229 bio->bi_io_vec[i].bv_page = NULL;
230 }
231 bio_put(bio);
232 }
233 bio = r10bio->devs[j].repl_bio;
234 if (bio)
235 bio_put(bio);
236 }
237 r10bio_pool_free(r10bio, conf);
238 }
239
240 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
241 {
242 int i;
243
244 for (i = 0; i < conf->copies; i++) {
245 struct bio **bio = & r10_bio->devs[i].bio;
246 if (!BIO_SPECIAL(*bio))
247 bio_put(*bio);
248 *bio = NULL;
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;
253 }
254 }
255
256 static void free_r10bio(struct r10bio *r10_bio)
257 {
258 struct r10conf *conf = r10_bio->mddev->private;
259
260 put_all_bios(conf, r10_bio);
261 mempool_free(r10_bio, conf->r10bio_pool);
262 }
263
264 static void put_buf(struct r10bio *r10_bio)
265 {
266 struct r10conf *conf = r10_bio->mddev->private;
267
268 mempool_free(r10_bio, conf->r10buf_pool);
269
270 lower_barrier(conf);
271 }
272
273 static void reschedule_retry(struct r10bio *r10_bio)
274 {
275 unsigned long flags;
276 struct mddev *mddev = r10_bio->mddev;
277 struct r10conf *conf = mddev->private;
278
279 spin_lock_irqsave(&conf->device_lock, flags);
280 list_add(&r10_bio->retry_list, &conf->retry_list);
281 conf->nr_queued ++;
282 spin_unlock_irqrestore(&conf->device_lock, flags);
283
284 /* wake up frozen array... */
285 wake_up(&conf->wait_barrier);
286
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 */
295 static void raid_end_bio_io(struct r10bio *r10_bio)
296 {
297 struct bio *bio = r10_bio->master_bio;
298 int done;
299 struct r10conf *conf = r10_bio->mddev->private;
300
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 }
319 free_r10bio(r10_bio);
320 }
321
322 /*
323 * Update disk head position estimator based on IRQ completion info.
324 */
325 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
326 {
327 struct r10conf *conf = r10_bio->mddev->private;
328
329 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
330 r10_bio->devs[slot].addr + (r10_bio->sectors);
331 }
332
333 /*
334 * Find the disk number which triggered given bio
335 */
336 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
337 struct bio *bio, int *slotp, int *replp)
338 {
339 int slot;
340 int repl = 0;
341
342 for (slot = 0; slot < conf->copies; slot++) {
343 if (r10_bio->devs[slot].bio == bio)
344 break;
345 if (r10_bio->devs[slot].repl_bio == bio) {
346 repl = 1;
347 break;
348 }
349 }
350
351 BUG_ON(slot == conf->copies);
352 update_head_pos(slot, r10_bio);
353
354 if (slotp)
355 *slotp = slot;
356 if (replp)
357 *replp = repl;
358 return r10_bio->devs[slot].devnum;
359 }
360
361 static void raid10_end_read_request(struct bio *bio, int error)
362 {
363 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
364 struct r10bio *r10_bio = bio->bi_private;
365 int slot, dev;
366 struct md_rdev *rdev;
367 struct r10conf *conf = r10_bio->mddev->private;
368
369
370 slot = r10_bio->read_slot;
371 dev = r10_bio->devs[slot].devnum;
372 rdev = r10_bio->devs[slot].rdev;
373 /*
374 * this branch is our 'one mirror IO has finished' event handler:
375 */
376 update_head_pos(slot, r10_bio);
377
378 if (uptodate) {
379 /*
380 * Set R10BIO_Uptodate in our master bio, so that
381 * we will return a good error code to the higher
382 * levels even if IO on some other mirrored buffer fails.
383 *
384 * The 'master' represents the composite IO operation to
385 * user-side. So if something waits for IO, then it will
386 * wait for the 'master' bio.
387 */
388 set_bit(R10BIO_Uptodate, &r10_bio->state);
389 } else {
390 /* If all other devices that store this block have
391 * failed, we want to return the error upwards rather
392 * than fail the last device. Here we redefine
393 * "uptodate" to mean "Don't want to retry"
394 */
395 unsigned long flags;
396 spin_lock_irqsave(&conf->device_lock, flags);
397 if (!enough(conf, rdev->raid_disk))
398 uptodate = 1;
399 spin_unlock_irqrestore(&conf->device_lock, flags);
400 }
401 if (uptodate) {
402 raid_end_bio_io(r10_bio);
403 rdev_dec_pending(rdev, conf->mddev);
404 } else {
405 /*
406 * oops, read error - keep the refcount on the rdev
407 */
408 char b[BDEVNAME_SIZE];
409 printk_ratelimited(KERN_ERR
410 "md/raid10:%s: %s: rescheduling sector %llu\n",
411 mdname(conf->mddev),
412 bdevname(rdev->bdev, b),
413 (unsigned long long)r10_bio->sector);
414 set_bit(R10BIO_ReadError, &r10_bio->state);
415 reschedule_retry(r10_bio);
416 }
417 }
418
419 static void close_write(struct r10bio *r10_bio)
420 {
421 /* clear the bitmap if all writes complete successfully */
422 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
423 r10_bio->sectors,
424 !test_bit(R10BIO_Degraded, &r10_bio->state),
425 0);
426 md_write_end(r10_bio->mddev);
427 }
428
429 static void one_write_done(struct r10bio *r10_bio)
430 {
431 if (atomic_dec_and_test(&r10_bio->remaining)) {
432 if (test_bit(R10BIO_WriteError, &r10_bio->state))
433 reschedule_retry(r10_bio);
434 else {
435 close_write(r10_bio);
436 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
437 reschedule_retry(r10_bio);
438 else
439 raid_end_bio_io(r10_bio);
440 }
441 }
442 }
443
444 static void raid10_end_write_request(struct bio *bio, int error)
445 {
446 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
447 struct r10bio *r10_bio = bio->bi_private;
448 int dev;
449 int dec_rdev = 1;
450 struct r10conf *conf = r10_bio->mddev->private;
451 int slot, repl;
452 struct md_rdev *rdev = NULL;
453
454 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
455
456 if (repl)
457 rdev = conf->mirrors[dev].replacement;
458 if (!rdev) {
459 smp_rmb();
460 repl = 0;
461 rdev = conf->mirrors[dev].rdev;
462 }
463 /*
464 * this branch is our 'one mirror IO has finished' event handler:
465 */
466 if (!uptodate) {
467 if (repl)
468 /* Never record new bad blocks to replacement,
469 * just fail it.
470 */
471 md_error(rdev->mddev, rdev);
472 else {
473 set_bit(WriteErrorSeen, &rdev->flags);
474 if (!test_and_set_bit(WantReplacement, &rdev->flags))
475 set_bit(MD_RECOVERY_NEEDED,
476 &rdev->mddev->recovery);
477 set_bit(R10BIO_WriteError, &r10_bio->state);
478 dec_rdev = 0;
479 }
480 } else {
481 /*
482 * Set R10BIO_Uptodate in our master bio, so that
483 * we will return a good error code for to the higher
484 * levels even if IO on some other mirrored buffer fails.
485 *
486 * The 'master' represents the composite IO operation to
487 * user-side. So if something waits for IO, then it will
488 * wait for the 'master' bio.
489 */
490 sector_t first_bad;
491 int bad_sectors;
492
493 /*
494 * Do not set R10BIO_Uptodate if the current device is
495 * rebuilding or Faulty. This is because we cannot use
496 * such device for properly reading the data back (we could
497 * potentially use it, if the current write would have felt
498 * before rdev->recovery_offset, but for simplicity we don't
499 * check this here.
500 */
501 if (test_bit(In_sync, &rdev->flags) &&
502 !test_bit(Faulty, &rdev->flags))
503 set_bit(R10BIO_Uptodate, &r10_bio->state);
504
505 /* Maybe we can clear some bad blocks. */
506 if (is_badblock(rdev,
507 r10_bio->devs[slot].addr,
508 r10_bio->sectors,
509 &first_bad, &bad_sectors)) {
510 bio_put(bio);
511 if (repl)
512 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
513 else
514 r10_bio->devs[slot].bio = IO_MADE_GOOD;
515 dec_rdev = 0;
516 set_bit(R10BIO_MadeGood, &r10_bio->state);
517 }
518 }
519
520 /*
521 *
522 * Let's see if all mirrored write operations have finished
523 * already.
524 */
525 one_write_done(r10_bio);
526 if (dec_rdev)
527 rdev_dec_pending(rdev, conf->mddev);
528 }
529
530 /*
531 * RAID10 layout manager
532 * As well as the chunksize and raid_disks count, there are two
533 * parameters: near_copies and far_copies.
534 * near_copies * far_copies must be <= raid_disks.
535 * Normally one of these will be 1.
536 * If both are 1, we get raid0.
537 * If near_copies == raid_disks, we get raid1.
538 *
539 * Chunks are laid out in raid0 style with near_copies copies of the
540 * first chunk, followed by near_copies copies of the next chunk and
541 * so on.
542 * If far_copies > 1, then after 1/far_copies of the array has been assigned
543 * as described above, we start again with a device offset of near_copies.
544 * So we effectively have another copy of the whole array further down all
545 * the drives, but with blocks on different drives.
546 * With this layout, and block is never stored twice on the one device.
547 *
548 * raid10_find_phys finds the sector offset of a given virtual sector
549 * on each device that it is on.
550 *
551 * raid10_find_virt does the reverse mapping, from a device and a
552 * sector offset to a virtual address
553 */
554
555 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
556 {
557 int n,f;
558 sector_t sector;
559 sector_t chunk;
560 sector_t stripe;
561 int dev;
562 int slot = 0;
563 int last_far_set_start, last_far_set_size;
564
565 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
566 last_far_set_start *= geo->far_set_size;
567
568 last_far_set_size = geo->far_set_size;
569 last_far_set_size += (geo->raid_disks % geo->far_set_size);
570
571 /* now calculate first sector/dev */
572 chunk = r10bio->sector >> geo->chunk_shift;
573 sector = r10bio->sector & geo->chunk_mask;
574
575 chunk *= geo->near_copies;
576 stripe = chunk;
577 dev = sector_div(stripe, geo->raid_disks);
578 if (geo->far_offset)
579 stripe *= geo->far_copies;
580
581 sector += stripe << geo->chunk_shift;
582
583 /* and calculate all the others */
584 for (n = 0; n < geo->near_copies; n++) {
585 int d = dev;
586 int set;
587 sector_t s = sector;
588 r10bio->devs[slot].devnum = d;
589 r10bio->devs[slot].addr = s;
590 slot++;
591
592 for (f = 1; f < geo->far_copies; f++) {
593 set = d / geo->far_set_size;
594 d += geo->near_copies;
595
596 if ((geo->raid_disks % geo->far_set_size) &&
597 (d > last_far_set_start)) {
598 d -= last_far_set_start;
599 d %= last_far_set_size;
600 d += last_far_set_start;
601 } else {
602 d %= geo->far_set_size;
603 d += geo->far_set_size * set;
604 }
605 s += geo->stride;
606 r10bio->devs[slot].devnum = d;
607 r10bio->devs[slot].addr = s;
608 slot++;
609 }
610 dev++;
611 if (dev >= geo->raid_disks) {
612 dev = 0;
613 sector += (geo->chunk_mask + 1);
614 }
615 }
616 }
617
618 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
619 {
620 struct geom *geo = &conf->geo;
621
622 if (conf->reshape_progress != MaxSector &&
623 ((r10bio->sector >= conf->reshape_progress) !=
624 conf->mddev->reshape_backwards)) {
625 set_bit(R10BIO_Previous, &r10bio->state);
626 geo = &conf->prev;
627 } else
628 clear_bit(R10BIO_Previous, &r10bio->state);
629
630 __raid10_find_phys(geo, r10bio);
631 }
632
633 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
634 {
635 sector_t offset, chunk, vchunk;
636 /* Never use conf->prev as this is only called during resync
637 * or recovery, so reshape isn't happening
638 */
639 struct geom *geo = &conf->geo;
640 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
641 int far_set_size = geo->far_set_size;
642 int last_far_set_start;
643
644 if (geo->raid_disks % geo->far_set_size) {
645 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
646 last_far_set_start *= geo->far_set_size;
647
648 if (dev >= last_far_set_start) {
649 far_set_size = geo->far_set_size;
650 far_set_size += (geo->raid_disks % geo->far_set_size);
651 far_set_start = last_far_set_start;
652 }
653 }
654
655 offset = sector & geo->chunk_mask;
656 if (geo->far_offset) {
657 int fc;
658 chunk = sector >> geo->chunk_shift;
659 fc = sector_div(chunk, geo->far_copies);
660 dev -= fc * geo->near_copies;
661 if (dev < far_set_start)
662 dev += far_set_size;
663 } else {
664 while (sector >= geo->stride) {
665 sector -= geo->stride;
666 if (dev < (geo->near_copies + far_set_start))
667 dev += far_set_size - geo->near_copies;
668 else
669 dev -= geo->near_copies;
670 }
671 chunk = sector >> geo->chunk_shift;
672 }
673 vchunk = chunk * geo->raid_disks + dev;
674 sector_div(vchunk, geo->near_copies);
675 return (vchunk << geo->chunk_shift) + offset;
676 }
677
678 /**
679 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
680 * @q: request queue
681 * @bvm: properties of new bio
682 * @biovec: the request that could be merged to it.
683 *
684 * Return amount of bytes we can accept at this offset
685 * This requires checking for end-of-chunk if near_copies != raid_disks,
686 * and for subordinate merge_bvec_fns if merge_check_needed.
687 */
688 static int raid10_mergeable_bvec(struct request_queue *q,
689 struct bvec_merge_data *bvm,
690 struct bio_vec *biovec)
691 {
692 struct mddev *mddev = q->queuedata;
693 struct r10conf *conf = mddev->private;
694 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
695 int max;
696 unsigned int chunk_sectors;
697 unsigned int bio_sectors = bvm->bi_size >> 9;
698 struct geom *geo = &conf->geo;
699
700 chunk_sectors = (conf->geo.chunk_mask & conf->prev.chunk_mask) + 1;
701 if (conf->reshape_progress != MaxSector &&
702 ((sector >= conf->reshape_progress) !=
703 conf->mddev->reshape_backwards))
704 geo = &conf->prev;
705
706 if (geo->near_copies < geo->raid_disks) {
707 max = (chunk_sectors - ((sector & (chunk_sectors - 1))
708 + bio_sectors)) << 9;
709 if (max < 0)
710 /* bio_add cannot handle a negative return */
711 max = 0;
712 if (max <= biovec->bv_len && bio_sectors == 0)
713 return biovec->bv_len;
714 } else
715 max = biovec->bv_len;
716
717 if (mddev->merge_check_needed) {
718 struct {
719 struct r10bio r10_bio;
720 struct r10dev devs[conf->copies];
721 } on_stack;
722 struct r10bio *r10_bio = &on_stack.r10_bio;
723 int s;
724 if (conf->reshape_progress != MaxSector) {
725 /* Cannot give any guidance during reshape */
726 if (max <= biovec->bv_len && bio_sectors == 0)
727 return biovec->bv_len;
728 return 0;
729 }
730 r10_bio->sector = sector;
731 raid10_find_phys(conf, r10_bio);
732 rcu_read_lock();
733 for (s = 0; s < conf->copies; s++) {
734 int disk = r10_bio->devs[s].devnum;
735 struct md_rdev *rdev = rcu_dereference(
736 conf->mirrors[disk].rdev);
737 if (rdev && !test_bit(Faulty, &rdev->flags)) {
738 struct request_queue *q =
739 bdev_get_queue(rdev->bdev);
740 if (q->merge_bvec_fn) {
741 bvm->bi_sector = r10_bio->devs[s].addr
742 + rdev->data_offset;
743 bvm->bi_bdev = rdev->bdev;
744 max = min(max, q->merge_bvec_fn(
745 q, bvm, biovec));
746 }
747 }
748 rdev = rcu_dereference(conf->mirrors[disk].replacement);
749 if (rdev && !test_bit(Faulty, &rdev->flags)) {
750 struct request_queue *q =
751 bdev_get_queue(rdev->bdev);
752 if (q->merge_bvec_fn) {
753 bvm->bi_sector = r10_bio->devs[s].addr
754 + rdev->data_offset;
755 bvm->bi_bdev = rdev->bdev;
756 max = min(max, q->merge_bvec_fn(
757 q, bvm, biovec));
758 }
759 }
760 }
761 rcu_read_unlock();
762 }
763 return max;
764 }
765
766 /*
767 * This routine returns the disk from which the requested read should
768 * be done. There is a per-array 'next expected sequential IO' sector
769 * number - if this matches on the next IO then we use the last disk.
770 * There is also a per-disk 'last know head position' sector that is
771 * maintained from IRQ contexts, both the normal and the resync IO
772 * completion handlers update this position correctly. If there is no
773 * perfect sequential match then we pick the disk whose head is closest.
774 *
775 * If there are 2 mirrors in the same 2 devices, performance degrades
776 * because position is mirror, not device based.
777 *
778 * The rdev for the device selected will have nr_pending incremented.
779 */
780
781 /*
782 * FIXME: possibly should rethink readbalancing and do it differently
783 * depending on near_copies / far_copies geometry.
784 */
785 static struct md_rdev *read_balance(struct r10conf *conf,
786 struct r10bio *r10_bio,
787 int *max_sectors)
788 {
789 const sector_t this_sector = r10_bio->sector;
790 int disk, slot;
791 int sectors = r10_bio->sectors;
792 int best_good_sectors;
793 sector_t new_distance, best_dist;
794 struct md_rdev *best_rdev, *rdev = NULL;
795 int do_balance;
796 int best_slot;
797 struct geom *geo = &conf->geo;
798
799 raid10_find_phys(conf, r10_bio);
800 rcu_read_lock();
801 retry:
802 sectors = r10_bio->sectors;
803 best_slot = -1;
804 best_rdev = NULL;
805 best_dist = MaxSector;
806 best_good_sectors = 0;
807 do_balance = 1;
808 /*
809 * Check if we can balance. We can balance on the whole
810 * device if no resync is going on (recovery is ok), or below
811 * the resync window. We take the first readable disk when
812 * above the resync window.
813 */
814 if (conf->mddev->recovery_cp < MaxSector
815 && (this_sector + sectors >= conf->next_resync))
816 do_balance = 0;
817
818 for (slot = 0; slot < conf->copies ; slot++) {
819 sector_t first_bad;
820 int bad_sectors;
821 sector_t dev_sector;
822
823 if (r10_bio->devs[slot].bio == IO_BLOCKED)
824 continue;
825 disk = r10_bio->devs[slot].devnum;
826 rdev = rcu_dereference(conf->mirrors[disk].replacement);
827 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
828 test_bit(Unmerged, &rdev->flags) ||
829 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
830 rdev = rcu_dereference(conf->mirrors[disk].rdev);
831 if (rdev == NULL ||
832 test_bit(Faulty, &rdev->flags) ||
833 test_bit(Unmerged, &rdev->flags))
834 continue;
835 if (!test_bit(In_sync, &rdev->flags) &&
836 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
837 continue;
838
839 dev_sector = r10_bio->devs[slot].addr;
840 if (is_badblock(rdev, dev_sector, sectors,
841 &first_bad, &bad_sectors)) {
842 if (best_dist < MaxSector)
843 /* Already have a better slot */
844 continue;
845 if (first_bad <= dev_sector) {
846 /* Cannot read here. If this is the
847 * 'primary' device, then we must not read
848 * beyond 'bad_sectors' from another device.
849 */
850 bad_sectors -= (dev_sector - first_bad);
851 if (!do_balance && sectors > bad_sectors)
852 sectors = bad_sectors;
853 if (best_good_sectors > sectors)
854 best_good_sectors = sectors;
855 } else {
856 sector_t good_sectors =
857 first_bad - dev_sector;
858 if (good_sectors > best_good_sectors) {
859 best_good_sectors = good_sectors;
860 best_slot = slot;
861 best_rdev = rdev;
862 }
863 if (!do_balance)
864 /* Must read from here */
865 break;
866 }
867 continue;
868 } else
869 best_good_sectors = sectors;
870
871 if (!do_balance)
872 break;
873
874 /* This optimisation is debatable, and completely destroys
875 * sequential read speed for 'far copies' arrays. So only
876 * keep it for 'near' arrays, and review those later.
877 */
878 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
879 break;
880
881 /* for far > 1 always use the lowest address */
882 if (geo->far_copies > 1)
883 new_distance = r10_bio->devs[slot].addr;
884 else
885 new_distance = abs(r10_bio->devs[slot].addr -
886 conf->mirrors[disk].head_position);
887 if (new_distance < best_dist) {
888 best_dist = new_distance;
889 best_slot = slot;
890 best_rdev = rdev;
891 }
892 }
893 if (slot >= conf->copies) {
894 slot = best_slot;
895 rdev = best_rdev;
896 }
897
898 if (slot >= 0) {
899 atomic_inc(&rdev->nr_pending);
900 if (test_bit(Faulty, &rdev->flags)) {
901 /* Cannot risk returning a device that failed
902 * before we inc'ed nr_pending
903 */
904 rdev_dec_pending(rdev, conf->mddev);
905 goto retry;
906 }
907 r10_bio->read_slot = slot;
908 } else
909 rdev = NULL;
910 rcu_read_unlock();
911 *max_sectors = best_good_sectors;
912
913 return rdev;
914 }
915
916 int md_raid10_congested(struct mddev *mddev, int bits)
917 {
918 struct r10conf *conf = mddev->private;
919 int i, ret = 0;
920
921 if ((bits & (1 << BDI_async_congested)) &&
922 conf->pending_count >= max_queued_requests)
923 return 1;
924
925 rcu_read_lock();
926 for (i = 0;
927 (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
928 && ret == 0;
929 i++) {
930 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
931 if (rdev && !test_bit(Faulty, &rdev->flags)) {
932 struct request_queue *q = bdev_get_queue(rdev->bdev);
933
934 ret |= bdi_congested(&q->backing_dev_info, bits);
935 }
936 }
937 rcu_read_unlock();
938 return ret;
939 }
940 EXPORT_SYMBOL_GPL(md_raid10_congested);
941
942 static int raid10_congested(void *data, int bits)
943 {
944 struct mddev *mddev = data;
945
946 return mddev_congested(mddev, bits) ||
947 md_raid10_congested(mddev, bits);
948 }
949
950 static void flush_pending_writes(struct r10conf *conf)
951 {
952 /* Any writes that have been queued but are awaiting
953 * bitmap updates get flushed here.
954 */
955 spin_lock_irq(&conf->device_lock);
956
957 if (conf->pending_bio_list.head) {
958 struct bio *bio;
959 bio = bio_list_get(&conf->pending_bio_list);
960 conf->pending_count = 0;
961 spin_unlock_irq(&conf->device_lock);
962 /* flush any pending bitmap writes to disk
963 * before proceeding w/ I/O */
964 bitmap_unplug(conf->mddev->bitmap);
965 wake_up(&conf->wait_barrier);
966
967 while (bio) { /* submit pending writes */
968 struct bio *next = bio->bi_next;
969 bio->bi_next = NULL;
970 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
971 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
972 /* Just ignore it */
973 bio_endio(bio, 0);
974 else
975 generic_make_request(bio);
976 bio = next;
977 }
978 } else
979 spin_unlock_irq(&conf->device_lock);
980 }
981
982 /* Barriers....
983 * Sometimes we need to suspend IO while we do something else,
984 * either some resync/recovery, or reconfigure the array.
985 * To do this we raise a 'barrier'.
986 * The 'barrier' is a counter that can be raised multiple times
987 * to count how many activities are happening which preclude
988 * normal IO.
989 * We can only raise the barrier if there is no pending IO.
990 * i.e. if nr_pending == 0.
991 * We choose only to raise the barrier if no-one is waiting for the
992 * barrier to go down. This means that as soon as an IO request
993 * is ready, no other operations which require a barrier will start
994 * until the IO request has had a chance.
995 *
996 * So: regular IO calls 'wait_barrier'. When that returns there
997 * is no backgroup IO happening, It must arrange to call
998 * allow_barrier when it has finished its IO.
999 * backgroup IO calls must call raise_barrier. Once that returns
1000 * there is no normal IO happeing. It must arrange to call
1001 * lower_barrier when the particular background IO completes.
1002 */
1003
1004 static void raise_barrier(struct r10conf *conf, int force)
1005 {
1006 BUG_ON(force && !conf->barrier);
1007 spin_lock_irq(&conf->resync_lock);
1008
1009 /* Wait until no block IO is waiting (unless 'force') */
1010 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
1011 conf->resync_lock);
1012
1013 /* block any new IO from starting */
1014 conf->barrier++;
1015
1016 /* Now wait for all pending IO to complete */
1017 wait_event_lock_irq(conf->wait_barrier,
1018 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
1019 conf->resync_lock);
1020
1021 spin_unlock_irq(&conf->resync_lock);
1022 }
1023
1024 static void lower_barrier(struct r10conf *conf)
1025 {
1026 unsigned long flags;
1027 spin_lock_irqsave(&conf->resync_lock, flags);
1028 conf->barrier--;
1029 spin_unlock_irqrestore(&conf->resync_lock, flags);
1030 wake_up(&conf->wait_barrier);
1031 }
1032
1033 static void wait_barrier(struct r10conf *conf)
1034 {
1035 spin_lock_irq(&conf->resync_lock);
1036 if (conf->barrier) {
1037 conf->nr_waiting++;
1038 /* Wait for the barrier to drop.
1039 * However if there are already pending
1040 * requests (preventing the barrier from
1041 * rising completely), and the
1042 * pre-process bio queue isn't empty,
1043 * then don't wait, as we need to empty
1044 * that queue to get the nr_pending
1045 * count down.
1046 */
1047 wait_event_lock_irq(conf->wait_barrier,
1048 !conf->barrier ||
1049 (conf->nr_pending &&
1050 current->bio_list &&
1051 !bio_list_empty(current->bio_list)),
1052 conf->resync_lock);
1053 conf->nr_waiting--;
1054 }
1055 conf->nr_pending++;
1056 spin_unlock_irq(&conf->resync_lock);
1057 }
1058
1059 static void allow_barrier(struct r10conf *conf)
1060 {
1061 unsigned long flags;
1062 spin_lock_irqsave(&conf->resync_lock, flags);
1063 conf->nr_pending--;
1064 spin_unlock_irqrestore(&conf->resync_lock, flags);
1065 wake_up(&conf->wait_barrier);
1066 }
1067
1068 static void freeze_array(struct r10conf *conf, int extra)
1069 {
1070 /* stop syncio and normal IO and wait for everything to
1071 * go quiet.
1072 * We increment barrier and nr_waiting, and then
1073 * wait until nr_pending match nr_queued+extra
1074 * This is called in the context of one normal IO request
1075 * that has failed. Thus any sync request that might be pending
1076 * will be blocked by nr_pending, and we need to wait for
1077 * pending IO requests to complete or be queued for re-try.
1078 * Thus the number queued (nr_queued) plus this request (extra)
1079 * must match the number of pending IOs (nr_pending) before
1080 * we continue.
1081 */
1082 spin_lock_irq(&conf->resync_lock);
1083 conf->barrier++;
1084 conf->nr_waiting++;
1085 wait_event_lock_irq_cmd(conf->wait_barrier,
1086 conf->nr_pending == conf->nr_queued+extra,
1087 conf->resync_lock,
1088 flush_pending_writes(conf));
1089
1090 spin_unlock_irq(&conf->resync_lock);
1091 }
1092
1093 static void unfreeze_array(struct r10conf *conf)
1094 {
1095 /* reverse the effect of the freeze */
1096 spin_lock_irq(&conf->resync_lock);
1097 conf->barrier--;
1098 conf->nr_waiting--;
1099 wake_up(&conf->wait_barrier);
1100 spin_unlock_irq(&conf->resync_lock);
1101 }
1102
1103 static sector_t choose_data_offset(struct r10bio *r10_bio,
1104 struct md_rdev *rdev)
1105 {
1106 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1107 test_bit(R10BIO_Previous, &r10_bio->state))
1108 return rdev->data_offset;
1109 else
1110 return rdev->new_data_offset;
1111 }
1112
1113 struct raid10_plug_cb {
1114 struct blk_plug_cb cb;
1115 struct bio_list pending;
1116 int pending_cnt;
1117 };
1118
1119 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1120 {
1121 struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1122 cb);
1123 struct mddev *mddev = plug->cb.data;
1124 struct r10conf *conf = mddev->private;
1125 struct bio *bio;
1126
1127 if (from_schedule || current->bio_list) {
1128 spin_lock_irq(&conf->device_lock);
1129 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1130 conf->pending_count += plug->pending_cnt;
1131 spin_unlock_irq(&conf->device_lock);
1132 wake_up(&conf->wait_barrier);
1133 md_wakeup_thread(mddev->thread);
1134 kfree(plug);
1135 return;
1136 }
1137
1138 /* we aren't scheduling, so we can do the write-out directly. */
1139 bio = bio_list_get(&plug->pending);
1140 bitmap_unplug(mddev->bitmap);
1141 wake_up(&conf->wait_barrier);
1142
1143 while (bio) { /* submit pending writes */
1144 struct bio *next = bio->bi_next;
1145 bio->bi_next = NULL;
1146 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
1147 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1148 /* Just ignore it */
1149 bio_endio(bio, 0);
1150 else
1151 generic_make_request(bio);
1152 bio = next;
1153 }
1154 kfree(plug);
1155 }
1156
1157 static void make_request(struct mddev *mddev, struct bio * bio)
1158 {
1159 struct r10conf *conf = mddev->private;
1160 struct r10bio *r10_bio;
1161 struct bio *read_bio;
1162 int i;
1163 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1164 int chunk_sects = chunk_mask + 1;
1165 const int rw = bio_data_dir(bio);
1166 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1167 const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
1168 const unsigned long do_discard = (bio->bi_rw
1169 & (REQ_DISCARD | REQ_SECURE));
1170 const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1171 unsigned long flags;
1172 struct md_rdev *blocked_rdev;
1173 struct blk_plug_cb *cb;
1174 struct raid10_plug_cb *plug = NULL;
1175 int sectors_handled;
1176 int max_sectors;
1177 int sectors;
1178
1179 if (unlikely(bio->bi_rw & REQ_FLUSH)) {
1180 md_flush_request(mddev, bio);
1181 return;
1182 }
1183
1184 /* If this request crosses a chunk boundary, we need to
1185 * split it. This will only happen for 1 PAGE (or less) requests.
1186 */
1187 if (unlikely((bio->bi_sector & chunk_mask) + bio_sectors(bio)
1188 > chunk_sects
1189 && (conf->geo.near_copies < conf->geo.raid_disks
1190 || conf->prev.near_copies < conf->prev.raid_disks))) {
1191 struct bio_pair *bp;
1192 /* Sanity check -- queue functions should prevent this happening */
1193 if (bio_segments(bio) > 1)
1194 goto bad_map;
1195 /* This is a one page bio that upper layers
1196 * refuse to split for us, so we need to split it.
1197 */
1198 bp = bio_split(bio,
1199 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
1200
1201 /* Each of these 'make_request' calls will call 'wait_barrier'.
1202 * If the first succeeds but the second blocks due to the resync
1203 * thread raising the barrier, we will deadlock because the
1204 * IO to the underlying device will be queued in generic_make_request
1205 * and will never complete, so will never reduce nr_pending.
1206 * So increment nr_waiting here so no new raise_barriers will
1207 * succeed, and so the second wait_barrier cannot block.
1208 */
1209 spin_lock_irq(&conf->resync_lock);
1210 conf->nr_waiting++;
1211 spin_unlock_irq(&conf->resync_lock);
1212
1213 make_request(mddev, &bp->bio1);
1214 make_request(mddev, &bp->bio2);
1215
1216 spin_lock_irq(&conf->resync_lock);
1217 conf->nr_waiting--;
1218 wake_up(&conf->wait_barrier);
1219 spin_unlock_irq(&conf->resync_lock);
1220
1221 bio_pair_release(bp);
1222 return;
1223 bad_map:
1224 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
1225 " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
1226 (unsigned long long)bio->bi_sector, bio_sectors(bio) / 2);
1227
1228 bio_io_error(bio);
1229 return;
1230 }
1231
1232 md_write_start(mddev, bio);
1233
1234 /*
1235 * Register the new request and wait if the reconstruction
1236 * thread has put up a bar for new requests.
1237 * Continue immediately if no resync is active currently.
1238 */
1239 wait_barrier(conf);
1240
1241 sectors = bio_sectors(bio);
1242 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1243 bio->bi_sector < conf->reshape_progress &&
1244 bio->bi_sector + sectors > conf->reshape_progress) {
1245 /* IO spans the reshape position. Need to wait for
1246 * reshape to pass
1247 */
1248 allow_barrier(conf);
1249 wait_event(conf->wait_barrier,
1250 conf->reshape_progress <= bio->bi_sector ||
1251 conf->reshape_progress >= bio->bi_sector + sectors);
1252 wait_barrier(conf);
1253 }
1254 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1255 bio_data_dir(bio) == WRITE &&
1256 (mddev->reshape_backwards
1257 ? (bio->bi_sector < conf->reshape_safe &&
1258 bio->bi_sector + sectors > conf->reshape_progress)
1259 : (bio->bi_sector + sectors > conf->reshape_safe &&
1260 bio->bi_sector < conf->reshape_progress))) {
1261 /* Need to update reshape_position in metadata */
1262 mddev->reshape_position = conf->reshape_progress;
1263 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1264 set_bit(MD_CHANGE_PENDING, &mddev->flags);
1265 md_wakeup_thread(mddev->thread);
1266 wait_event(mddev->sb_wait,
1267 !test_bit(MD_CHANGE_PENDING, &mddev->flags));
1268
1269 conf->reshape_safe = mddev->reshape_position;
1270 }
1271
1272 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1273
1274 r10_bio->master_bio = bio;
1275 r10_bio->sectors = sectors;
1276
1277 r10_bio->mddev = mddev;
1278 r10_bio->sector = bio->bi_sector;
1279 r10_bio->state = 0;
1280
1281 /* We might need to issue multiple reads to different
1282 * devices if there are bad blocks around, so we keep
1283 * track of the number of reads in bio->bi_phys_segments.
1284 * If this is 0, there is only one r10_bio and no locking
1285 * will be needed when the request completes. If it is
1286 * non-zero, then it is the number of not-completed requests.
1287 */
1288 bio->bi_phys_segments = 0;
1289 clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1290
1291 if (rw == READ) {
1292 /*
1293 * read balancing logic:
1294 */
1295 struct md_rdev *rdev;
1296 int slot;
1297
1298 read_again:
1299 rdev = read_balance(conf, r10_bio, &max_sectors);
1300 if (!rdev) {
1301 raid_end_bio_io(r10_bio);
1302 return;
1303 }
1304 slot = r10_bio->read_slot;
1305
1306 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1307 md_trim_bio(read_bio, r10_bio->sector - bio->bi_sector,
1308 max_sectors);
1309
1310 r10_bio->devs[slot].bio = read_bio;
1311 r10_bio->devs[slot].rdev = rdev;
1312
1313 read_bio->bi_sector = r10_bio->devs[slot].addr +
1314 choose_data_offset(r10_bio, rdev);
1315 read_bio->bi_bdev = rdev->bdev;
1316 read_bio->bi_end_io = raid10_end_read_request;
1317 read_bio->bi_rw = READ | do_sync;
1318 read_bio->bi_private = r10_bio;
1319
1320 if (max_sectors < r10_bio->sectors) {
1321 /* Could not read all from this device, so we will
1322 * need another r10_bio.
1323 */
1324 sectors_handled = (r10_bio->sector + max_sectors
1325 - bio->bi_sector);
1326 r10_bio->sectors = max_sectors;
1327 spin_lock_irq(&conf->device_lock);
1328 if (bio->bi_phys_segments == 0)
1329 bio->bi_phys_segments = 2;
1330 else
1331 bio->bi_phys_segments++;
1332 spin_unlock_irq(&conf->device_lock);
1333 /* Cannot call generic_make_request directly
1334 * as that will be queued in __generic_make_request
1335 * and subsequent mempool_alloc might block
1336 * waiting for it. so hand bio over to raid10d.
1337 */
1338 reschedule_retry(r10_bio);
1339
1340 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1341
1342 r10_bio->master_bio = bio;
1343 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1344 r10_bio->state = 0;
1345 r10_bio->mddev = mddev;
1346 r10_bio->sector = bio->bi_sector + sectors_handled;
1347 goto read_again;
1348 } else
1349 generic_make_request(read_bio);
1350 return;
1351 }
1352
1353 /*
1354 * WRITE:
1355 */
1356 if (conf->pending_count >= max_queued_requests) {
1357 md_wakeup_thread(mddev->thread);
1358 wait_event(conf->wait_barrier,
1359 conf->pending_count < max_queued_requests);
1360 }
1361 /* first select target devices under rcu_lock and
1362 * inc refcount on their rdev. Record them by setting
1363 * bios[x] to bio
1364 * If there are known/acknowledged bad blocks on any device
1365 * on which we have seen a write error, we want to avoid
1366 * writing to those blocks. This potentially requires several
1367 * writes to write around the bad blocks. Each set of writes
1368 * gets its own r10_bio with a set of bios attached. The number
1369 * of r10_bios is recored in bio->bi_phys_segments just as with
1370 * the read case.
1371 */
1372
1373 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1374 raid10_find_phys(conf, r10_bio);
1375 retry_write:
1376 blocked_rdev = NULL;
1377 rcu_read_lock();
1378 max_sectors = r10_bio->sectors;
1379
1380 for (i = 0; i < conf->copies; i++) {
1381 int d = r10_bio->devs[i].devnum;
1382 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1383 struct md_rdev *rrdev = rcu_dereference(
1384 conf->mirrors[d].replacement);
1385 if (rdev == rrdev)
1386 rrdev = NULL;
1387 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1388 atomic_inc(&rdev->nr_pending);
1389 blocked_rdev = rdev;
1390 break;
1391 }
1392 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1393 atomic_inc(&rrdev->nr_pending);
1394 blocked_rdev = rrdev;
1395 break;
1396 }
1397 if (rdev && (test_bit(Faulty, &rdev->flags)
1398 || test_bit(Unmerged, &rdev->flags)))
1399 rdev = NULL;
1400 if (rrdev && (test_bit(Faulty, &rrdev->flags)
1401 || test_bit(Unmerged, &rrdev->flags)))
1402 rrdev = NULL;
1403
1404 r10_bio->devs[i].bio = NULL;
1405 r10_bio->devs[i].repl_bio = NULL;
1406
1407 if (!rdev && !rrdev) {
1408 set_bit(R10BIO_Degraded, &r10_bio->state);
1409 continue;
1410 }
1411 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1412 sector_t first_bad;
1413 sector_t dev_sector = r10_bio->devs[i].addr;
1414 int bad_sectors;
1415 int is_bad;
1416
1417 is_bad = is_badblock(rdev, dev_sector,
1418 max_sectors,
1419 &first_bad, &bad_sectors);
1420 if (is_bad < 0) {
1421 /* Mustn't write here until the bad block
1422 * is acknowledged
1423 */
1424 atomic_inc(&rdev->nr_pending);
1425 set_bit(BlockedBadBlocks, &rdev->flags);
1426 blocked_rdev = rdev;
1427 break;
1428 }
1429 if (is_bad && first_bad <= dev_sector) {
1430 /* Cannot write here at all */
1431 bad_sectors -= (dev_sector - first_bad);
1432 if (bad_sectors < max_sectors)
1433 /* Mustn't write more than bad_sectors
1434 * to other devices yet
1435 */
1436 max_sectors = bad_sectors;
1437 /* We don't set R10BIO_Degraded as that
1438 * only applies if the disk is missing,
1439 * so it might be re-added, and we want to
1440 * know to recover this chunk.
1441 * In this case the device is here, and the
1442 * fact that this chunk is not in-sync is
1443 * recorded in the bad block log.
1444 */
1445 continue;
1446 }
1447 if (is_bad) {
1448 int good_sectors = first_bad - dev_sector;
1449 if (good_sectors < max_sectors)
1450 max_sectors = good_sectors;
1451 }
1452 }
1453 if (rdev) {
1454 r10_bio->devs[i].bio = bio;
1455 atomic_inc(&rdev->nr_pending);
1456 }
1457 if (rrdev) {
1458 r10_bio->devs[i].repl_bio = bio;
1459 atomic_inc(&rrdev->nr_pending);
1460 }
1461 }
1462 rcu_read_unlock();
1463
1464 if (unlikely(blocked_rdev)) {
1465 /* Have to wait for this device to get unblocked, then retry */
1466 int j;
1467 int d;
1468
1469 for (j = 0; j < i; j++) {
1470 if (r10_bio->devs[j].bio) {
1471 d = r10_bio->devs[j].devnum;
1472 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1473 }
1474 if (r10_bio->devs[j].repl_bio) {
1475 struct md_rdev *rdev;
1476 d = r10_bio->devs[j].devnum;
1477 rdev = conf->mirrors[d].replacement;
1478 if (!rdev) {
1479 /* Race with remove_disk */
1480 smp_mb();
1481 rdev = conf->mirrors[d].rdev;
1482 }
1483 rdev_dec_pending(rdev, mddev);
1484 }
1485 }
1486 allow_barrier(conf);
1487 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1488 wait_barrier(conf);
1489 goto retry_write;
1490 }
1491
1492 if (max_sectors < r10_bio->sectors) {
1493 /* We are splitting this into multiple parts, so
1494 * we need to prepare for allocating another r10_bio.
1495 */
1496 r10_bio->sectors = max_sectors;
1497 spin_lock_irq(&conf->device_lock);
1498 if (bio->bi_phys_segments == 0)
1499 bio->bi_phys_segments = 2;
1500 else
1501 bio->bi_phys_segments++;
1502 spin_unlock_irq(&conf->device_lock);
1503 }
1504 sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector;
1505
1506 atomic_set(&r10_bio->remaining, 1);
1507 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1508
1509 for (i = 0; i < conf->copies; i++) {
1510 struct bio *mbio;
1511 int d = r10_bio->devs[i].devnum;
1512 if (r10_bio->devs[i].bio) {
1513 struct md_rdev *rdev = conf->mirrors[d].rdev;
1514 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1515 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1516 max_sectors);
1517 r10_bio->devs[i].bio = mbio;
1518
1519 mbio->bi_sector = (r10_bio->devs[i].addr+
1520 choose_data_offset(r10_bio,
1521 rdev));
1522 mbio->bi_bdev = rdev->bdev;
1523 mbio->bi_end_io = raid10_end_write_request;
1524 mbio->bi_rw =
1525 WRITE | do_sync | do_fua | do_discard | do_same;
1526 mbio->bi_private = r10_bio;
1527
1528 atomic_inc(&r10_bio->remaining);
1529
1530 cb = blk_check_plugged(raid10_unplug, mddev,
1531 sizeof(*plug));
1532 if (cb)
1533 plug = container_of(cb, struct raid10_plug_cb,
1534 cb);
1535 else
1536 plug = NULL;
1537 spin_lock_irqsave(&conf->device_lock, flags);
1538 if (plug) {
1539 bio_list_add(&plug->pending, mbio);
1540 plug->pending_cnt++;
1541 } else {
1542 bio_list_add(&conf->pending_bio_list, mbio);
1543 conf->pending_count++;
1544 }
1545 spin_unlock_irqrestore(&conf->device_lock, flags);
1546 if (!plug)
1547 md_wakeup_thread(mddev->thread);
1548 }
1549
1550 if (r10_bio->devs[i].repl_bio) {
1551 struct md_rdev *rdev = conf->mirrors[d].replacement;
1552 if (rdev == NULL) {
1553 /* Replacement just got moved to main 'rdev' */
1554 smp_mb();
1555 rdev = conf->mirrors[d].rdev;
1556 }
1557 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1558 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1559 max_sectors);
1560 r10_bio->devs[i].repl_bio = mbio;
1561
1562 mbio->bi_sector = (r10_bio->devs[i].addr +
1563 choose_data_offset(
1564 r10_bio, rdev));
1565 mbio->bi_bdev = rdev->bdev;
1566 mbio->bi_end_io = raid10_end_write_request;
1567 mbio->bi_rw =
1568 WRITE | do_sync | do_fua | do_discard | do_same;
1569 mbio->bi_private = r10_bio;
1570
1571 atomic_inc(&r10_bio->remaining);
1572
1573 cb = blk_check_plugged(raid10_unplug, mddev,
1574 sizeof(*plug));
1575 if (cb)
1576 plug = container_of(cb, struct raid10_plug_cb,
1577 cb);
1578 else
1579 plug = NULL;
1580 spin_lock_irqsave(&conf->device_lock, flags);
1581 if (plug) {
1582 bio_list_add(&plug->pending, mbio);
1583 plug->pending_cnt++;
1584 } else {
1585 bio_list_add(&conf->pending_bio_list, mbio);
1586 conf->pending_count++;
1587 }
1588 spin_unlock_irqrestore(&conf->device_lock, flags);
1589 if (!plug)
1590 md_wakeup_thread(mddev->thread);
1591 }
1592 }
1593
1594 /* Don't remove the bias on 'remaining' (one_write_done) until
1595 * after checking if we need to go around again.
1596 */
1597
1598 if (sectors_handled < bio_sectors(bio)) {
1599 one_write_done(r10_bio);
1600 /* We need another r10_bio. It has already been counted
1601 * in bio->bi_phys_segments.
1602 */
1603 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1604
1605 r10_bio->master_bio = bio;
1606 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1607
1608 r10_bio->mddev = mddev;
1609 r10_bio->sector = bio->bi_sector + sectors_handled;
1610 r10_bio->state = 0;
1611 goto retry_write;
1612 }
1613 one_write_done(r10_bio);
1614
1615 /* In case raid10d snuck in to freeze_array */
1616 wake_up(&conf->wait_barrier);
1617 }
1618
1619 static void status(struct seq_file *seq, struct mddev *mddev)
1620 {
1621 struct r10conf *conf = mddev->private;
1622 int i;
1623
1624 if (conf->geo.near_copies < conf->geo.raid_disks)
1625 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1626 if (conf->geo.near_copies > 1)
1627 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1628 if (conf->geo.far_copies > 1) {
1629 if (conf->geo.far_offset)
1630 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1631 else
1632 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1633 }
1634 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1635 conf->geo.raid_disks - mddev->degraded);
1636 for (i = 0; i < conf->geo.raid_disks; i++)
1637 seq_printf(seq, "%s",
1638 conf->mirrors[i].rdev &&
1639 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1640 seq_printf(seq, "]");
1641 }
1642
1643 /* check if there are enough drives for
1644 * every block to appear on atleast one.
1645 * Don't consider the device numbered 'ignore'
1646 * as we might be about to remove it.
1647 */
1648 static int _enough(struct r10conf *conf, struct geom *geo, int ignore)
1649 {
1650 int first = 0;
1651
1652 do {
1653 int n = conf->copies;
1654 int cnt = 0;
1655 int this = first;
1656 while (n--) {
1657 if (conf->mirrors[this].rdev &&
1658 this != ignore)
1659 cnt++;
1660 this = (this+1) % geo->raid_disks;
1661 }
1662 if (cnt == 0)
1663 return 0;
1664 first = (first + geo->near_copies) % geo->raid_disks;
1665 } while (first != 0);
1666 return 1;
1667 }
1668
1669 static int enough(struct r10conf *conf, int ignore)
1670 {
1671 return _enough(conf, &conf->geo, ignore) &&
1672 _enough(conf, &conf->prev, ignore);
1673 }
1674
1675 static void error(struct mddev *mddev, struct md_rdev *rdev)
1676 {
1677 char b[BDEVNAME_SIZE];
1678 struct r10conf *conf = mddev->private;
1679
1680 /*
1681 * If it is not operational, then we have already marked it as dead
1682 * else if it is the last working disks, ignore the error, let the
1683 * next level up know.
1684 * else mark the drive as failed
1685 */
1686 if (test_bit(In_sync, &rdev->flags)
1687 && !enough(conf, rdev->raid_disk))
1688 /*
1689 * Don't fail the drive, just return an IO error.
1690 */
1691 return;
1692 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1693 unsigned long flags;
1694 spin_lock_irqsave(&conf->device_lock, flags);
1695 mddev->degraded++;
1696 spin_unlock_irqrestore(&conf->device_lock, flags);
1697 }
1698 /*
1699 * If recovery is running, make sure it aborts.
1700 */
1701 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1702 set_bit(Blocked, &rdev->flags);
1703 set_bit(Faulty, &rdev->flags);
1704 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1705 printk(KERN_ALERT
1706 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1707 "md/raid10:%s: Operation continuing on %d devices.\n",
1708 mdname(mddev), bdevname(rdev->bdev, b),
1709 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1710 }
1711
1712 static void print_conf(struct r10conf *conf)
1713 {
1714 int i;
1715 struct raid10_info *tmp;
1716
1717 printk(KERN_DEBUG "RAID10 conf printout:\n");
1718 if (!conf) {
1719 printk(KERN_DEBUG "(!conf)\n");
1720 return;
1721 }
1722 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1723 conf->geo.raid_disks);
1724
1725 for (i = 0; i < conf->geo.raid_disks; i++) {
1726 char b[BDEVNAME_SIZE];
1727 tmp = conf->mirrors + i;
1728 if (tmp->rdev)
1729 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1730 i, !test_bit(In_sync, &tmp->rdev->flags),
1731 !test_bit(Faulty, &tmp->rdev->flags),
1732 bdevname(tmp->rdev->bdev,b));
1733 }
1734 }
1735
1736 static void close_sync(struct r10conf *conf)
1737 {
1738 wait_barrier(conf);
1739 allow_barrier(conf);
1740
1741 mempool_destroy(conf->r10buf_pool);
1742 conf->r10buf_pool = NULL;
1743 }
1744
1745 static int raid10_spare_active(struct mddev *mddev)
1746 {
1747 int i;
1748 struct r10conf *conf = mddev->private;
1749 struct raid10_info *tmp;
1750 int count = 0;
1751 unsigned long flags;
1752
1753 /*
1754 * Find all non-in_sync disks within the RAID10 configuration
1755 * and mark them in_sync
1756 */
1757 for (i = 0; i < conf->geo.raid_disks; i++) {
1758 tmp = conf->mirrors + i;
1759 if (tmp->replacement
1760 && tmp->replacement->recovery_offset == MaxSector
1761 && !test_bit(Faulty, &tmp->replacement->flags)
1762 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1763 /* Replacement has just become active */
1764 if (!tmp->rdev
1765 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1766 count++;
1767 if (tmp->rdev) {
1768 /* Replaced device not technically faulty,
1769 * but we need to be sure it gets removed
1770 * and never re-added.
1771 */
1772 set_bit(Faulty, &tmp->rdev->flags);
1773 sysfs_notify_dirent_safe(
1774 tmp->rdev->sysfs_state);
1775 }
1776 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1777 } else if (tmp->rdev
1778 && tmp->rdev->recovery_offset == MaxSector
1779 && !test_bit(Faulty, &tmp->rdev->flags)
1780 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1781 count++;
1782 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1783 }
1784 }
1785 spin_lock_irqsave(&conf->device_lock, flags);
1786 mddev->degraded -= count;
1787 spin_unlock_irqrestore(&conf->device_lock, flags);
1788
1789 print_conf(conf);
1790 return count;
1791 }
1792
1793
1794 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1795 {
1796 struct r10conf *conf = mddev->private;
1797 int err = -EEXIST;
1798 int mirror;
1799 int first = 0;
1800 int last = conf->geo.raid_disks - 1;
1801 struct request_queue *q = bdev_get_queue(rdev->bdev);
1802
1803 if (mddev->recovery_cp < MaxSector)
1804 /* only hot-add to in-sync arrays, as recovery is
1805 * very different from resync
1806 */
1807 return -EBUSY;
1808 if (rdev->saved_raid_disk < 0 && !_enough(conf, &conf->prev, -1))
1809 return -EINVAL;
1810
1811 if (rdev->raid_disk >= 0)
1812 first = last = rdev->raid_disk;
1813
1814 if (q->merge_bvec_fn) {
1815 set_bit(Unmerged, &rdev->flags);
1816 mddev->merge_check_needed = 1;
1817 }
1818
1819 if (rdev->saved_raid_disk >= first &&
1820 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1821 mirror = rdev->saved_raid_disk;
1822 else
1823 mirror = first;
1824 for ( ; mirror <= last ; mirror++) {
1825 struct raid10_info *p = &conf->mirrors[mirror];
1826 if (p->recovery_disabled == mddev->recovery_disabled)
1827 continue;
1828 if (p->rdev) {
1829 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1830 p->replacement != NULL)
1831 continue;
1832 clear_bit(In_sync, &rdev->flags);
1833 set_bit(Replacement, &rdev->flags);
1834 rdev->raid_disk = mirror;
1835 err = 0;
1836 disk_stack_limits(mddev->gendisk, rdev->bdev,
1837 rdev->data_offset << 9);
1838 conf->fullsync = 1;
1839 rcu_assign_pointer(p->replacement, rdev);
1840 break;
1841 }
1842
1843 disk_stack_limits(mddev->gendisk, rdev->bdev,
1844 rdev->data_offset << 9);
1845
1846 p->head_position = 0;
1847 p->recovery_disabled = mddev->recovery_disabled - 1;
1848 rdev->raid_disk = mirror;
1849 err = 0;
1850 if (rdev->saved_raid_disk != mirror)
1851 conf->fullsync = 1;
1852 rcu_assign_pointer(p->rdev, rdev);
1853 break;
1854 }
1855 if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1856 /* Some requests might not have seen this new
1857 * merge_bvec_fn. We must wait for them to complete
1858 * before merging the device fully.
1859 * First we make sure any code which has tested
1860 * our function has submitted the request, then
1861 * we wait for all outstanding requests to complete.
1862 */
1863 synchronize_sched();
1864 freeze_array(conf, 0);
1865 unfreeze_array(conf);
1866 clear_bit(Unmerged, &rdev->flags);
1867 }
1868 md_integrity_add_rdev(rdev, mddev);
1869 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1870 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1871
1872 print_conf(conf);
1873 return err;
1874 }
1875
1876 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1877 {
1878 struct r10conf *conf = mddev->private;
1879 int err = 0;
1880 int number = rdev->raid_disk;
1881 struct md_rdev **rdevp;
1882 struct raid10_info *p = conf->mirrors + number;
1883
1884 print_conf(conf);
1885 if (rdev == p->rdev)
1886 rdevp = &p->rdev;
1887 else if (rdev == p->replacement)
1888 rdevp = &p->replacement;
1889 else
1890 return 0;
1891
1892 if (test_bit(In_sync, &rdev->flags) ||
1893 atomic_read(&rdev->nr_pending)) {
1894 err = -EBUSY;
1895 goto abort;
1896 }
1897 /* Only remove faulty devices if recovery
1898 * is not possible.
1899 */
1900 if (!test_bit(Faulty, &rdev->flags) &&
1901 mddev->recovery_disabled != p->recovery_disabled &&
1902 (!p->replacement || p->replacement == rdev) &&
1903 number < conf->geo.raid_disks &&
1904 enough(conf, -1)) {
1905 err = -EBUSY;
1906 goto abort;
1907 }
1908 *rdevp = NULL;
1909 synchronize_rcu();
1910 if (atomic_read(&rdev->nr_pending)) {
1911 /* lost the race, try later */
1912 err = -EBUSY;
1913 *rdevp = rdev;
1914 goto abort;
1915 } else if (p->replacement) {
1916 /* We must have just cleared 'rdev' */
1917 p->rdev = p->replacement;
1918 clear_bit(Replacement, &p->replacement->flags);
1919 smp_mb(); /* Make sure other CPUs may see both as identical
1920 * but will never see neither -- if they are careful.
1921 */
1922 p->replacement = NULL;
1923 clear_bit(WantReplacement, &rdev->flags);
1924 } else
1925 /* We might have just remove the Replacement as faulty
1926 * Clear the flag just in case
1927 */
1928 clear_bit(WantReplacement, &rdev->flags);
1929
1930 err = md_integrity_register(mddev);
1931
1932 abort:
1933
1934 print_conf(conf);
1935 return err;
1936 }
1937
1938
1939 static void end_sync_read(struct bio *bio, int error)
1940 {
1941 struct r10bio *r10_bio = bio->bi_private;
1942 struct r10conf *conf = r10_bio->mddev->private;
1943 int d;
1944
1945 if (bio == r10_bio->master_bio) {
1946 /* this is a reshape read */
1947 d = r10_bio->read_slot; /* really the read dev */
1948 } else
1949 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1950
1951 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1952 set_bit(R10BIO_Uptodate, &r10_bio->state);
1953 else
1954 /* The write handler will notice the lack of
1955 * R10BIO_Uptodate and record any errors etc
1956 */
1957 atomic_add(r10_bio->sectors,
1958 &conf->mirrors[d].rdev->corrected_errors);
1959
1960 /* for reconstruct, we always reschedule after a read.
1961 * for resync, only after all reads
1962 */
1963 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1964 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1965 atomic_dec_and_test(&r10_bio->remaining)) {
1966 /* we have read all the blocks,
1967 * do the comparison in process context in raid10d
1968 */
1969 reschedule_retry(r10_bio);
1970 }
1971 }
1972
1973 static void end_sync_request(struct r10bio *r10_bio)
1974 {
1975 struct mddev *mddev = r10_bio->mddev;
1976
1977 while (atomic_dec_and_test(&r10_bio->remaining)) {
1978 if (r10_bio->master_bio == NULL) {
1979 /* the primary of several recovery bios */
1980 sector_t s = r10_bio->sectors;
1981 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1982 test_bit(R10BIO_WriteError, &r10_bio->state))
1983 reschedule_retry(r10_bio);
1984 else
1985 put_buf(r10_bio);
1986 md_done_sync(mddev, s, 1);
1987 break;
1988 } else {
1989 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1990 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1991 test_bit(R10BIO_WriteError, &r10_bio->state))
1992 reschedule_retry(r10_bio);
1993 else
1994 put_buf(r10_bio);
1995 r10_bio = r10_bio2;
1996 }
1997 }
1998 }
1999
2000 static void end_sync_write(struct bio *bio, int error)
2001 {
2002 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
2003 struct r10bio *r10_bio = bio->bi_private;
2004 struct mddev *mddev = r10_bio->mddev;
2005 struct r10conf *conf = mddev->private;
2006 int d;
2007 sector_t first_bad;
2008 int bad_sectors;
2009 int slot;
2010 int repl;
2011 struct md_rdev *rdev = NULL;
2012
2013 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
2014 if (repl)
2015 rdev = conf->mirrors[d].replacement;
2016 else
2017 rdev = conf->mirrors[d].rdev;
2018
2019 if (!uptodate) {
2020 if (repl)
2021 md_error(mddev, rdev);
2022 else {
2023 set_bit(WriteErrorSeen, &rdev->flags);
2024 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2025 set_bit(MD_RECOVERY_NEEDED,
2026 &rdev->mddev->recovery);
2027 set_bit(R10BIO_WriteError, &r10_bio->state);
2028 }
2029 } else if (is_badblock(rdev,
2030 r10_bio->devs[slot].addr,
2031 r10_bio->sectors,
2032 &first_bad, &bad_sectors))
2033 set_bit(R10BIO_MadeGood, &r10_bio->state);
2034
2035 rdev_dec_pending(rdev, mddev);
2036
2037 end_sync_request(r10_bio);
2038 }
2039
2040 /*
2041 * Note: sync and recover and handled very differently for raid10
2042 * This code is for resync.
2043 * For resync, we read through virtual addresses and read all blocks.
2044 * If there is any error, we schedule a write. The lowest numbered
2045 * drive is authoritative.
2046 * However requests come for physical address, so we need to map.
2047 * For every physical address there are raid_disks/copies virtual addresses,
2048 * which is always are least one, but is not necessarly an integer.
2049 * This means that a physical address can span multiple chunks, so we may
2050 * have to submit multiple io requests for a single sync request.
2051 */
2052 /*
2053 * We check if all blocks are in-sync and only write to blocks that
2054 * aren't in sync
2055 */
2056 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2057 {
2058 struct r10conf *conf = mddev->private;
2059 int i, first;
2060 struct bio *tbio, *fbio;
2061 int vcnt;
2062
2063 atomic_set(&r10_bio->remaining, 1);
2064
2065 /* find the first device with a block */
2066 for (i=0; i<conf->copies; i++)
2067 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
2068 break;
2069
2070 if (i == conf->copies)
2071 goto done;
2072
2073 first = i;
2074 fbio = r10_bio->devs[i].bio;
2075
2076 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2077 /* now find blocks with errors */
2078 for (i=0 ; i < conf->copies ; i++) {
2079 int j, d;
2080
2081 tbio = r10_bio->devs[i].bio;
2082
2083 if (tbio->bi_end_io != end_sync_read)
2084 continue;
2085 if (i == first)
2086 continue;
2087 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
2088 /* We know that the bi_io_vec layout is the same for
2089 * both 'first' and 'i', so we just compare them.
2090 * All vec entries are PAGE_SIZE;
2091 */
2092 int sectors = r10_bio->sectors;
2093 for (j = 0; j < vcnt; j++) {
2094 int len = PAGE_SIZE;
2095 if (sectors < (len / 512))
2096 len = sectors * 512;
2097 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
2098 page_address(tbio->bi_io_vec[j].bv_page),
2099 len))
2100 break;
2101 sectors -= len/512;
2102 }
2103 if (j == vcnt)
2104 continue;
2105 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2106 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2107 /* Don't fix anything. */
2108 continue;
2109 }
2110 /* Ok, we need to write this bio, either to correct an
2111 * inconsistency or to correct an unreadable block.
2112 * First we need to fixup bv_offset, bv_len and
2113 * bi_vecs, as the read request might have corrupted these
2114 */
2115 bio_reset(tbio);
2116
2117 tbio->bi_vcnt = vcnt;
2118 tbio->bi_size = r10_bio->sectors << 9;
2119 tbio->bi_rw = WRITE;
2120 tbio->bi_private = r10_bio;
2121 tbio->bi_sector = r10_bio->devs[i].addr;
2122
2123 for (j=0; j < vcnt ; j++) {
2124 tbio->bi_io_vec[j].bv_offset = 0;
2125 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
2126
2127 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2128 page_address(fbio->bi_io_vec[j].bv_page),
2129 PAGE_SIZE);
2130 }
2131 tbio->bi_end_io = end_sync_write;
2132
2133 d = r10_bio->devs[i].devnum;
2134 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2135 atomic_inc(&r10_bio->remaining);
2136 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2137
2138 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
2139 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2140 generic_make_request(tbio);
2141 }
2142
2143 /* Now write out to any replacement devices
2144 * that are active
2145 */
2146 for (i = 0; i < conf->copies; i++) {
2147 int j, d;
2148
2149 tbio = r10_bio->devs[i].repl_bio;
2150 if (!tbio || !tbio->bi_end_io)
2151 continue;
2152 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2153 && r10_bio->devs[i].bio != fbio)
2154 for (j = 0; j < vcnt; j++)
2155 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2156 page_address(fbio->bi_io_vec[j].bv_page),
2157 PAGE_SIZE);
2158 d = r10_bio->devs[i].devnum;
2159 atomic_inc(&r10_bio->remaining);
2160 md_sync_acct(conf->mirrors[d].replacement->bdev,
2161 bio_sectors(tbio));
2162 generic_make_request(tbio);
2163 }
2164
2165 done:
2166 if (atomic_dec_and_test(&r10_bio->remaining)) {
2167 md_done_sync(mddev, r10_bio->sectors, 1);
2168 put_buf(r10_bio);
2169 }
2170 }
2171
2172 /*
2173 * Now for the recovery code.
2174 * Recovery happens across physical sectors.
2175 * We recover all non-is_sync drives by finding the virtual address of
2176 * each, and then choose a working drive that also has that virt address.
2177 * There is a separate r10_bio for each non-in_sync drive.
2178 * Only the first two slots are in use. The first for reading,
2179 * The second for writing.
2180 *
2181 */
2182 static void fix_recovery_read_error(struct r10bio *r10_bio)
2183 {
2184 /* We got a read error during recovery.
2185 * We repeat the read in smaller page-sized sections.
2186 * If a read succeeds, write it to the new device or record
2187 * a bad block if we cannot.
2188 * If a read fails, record a bad block on both old and
2189 * new devices.
2190 */
2191 struct mddev *mddev = r10_bio->mddev;
2192 struct r10conf *conf = mddev->private;
2193 struct bio *bio = r10_bio->devs[0].bio;
2194 sector_t sect = 0;
2195 int sectors = r10_bio->sectors;
2196 int idx = 0;
2197 int dr = r10_bio->devs[0].devnum;
2198 int dw = r10_bio->devs[1].devnum;
2199
2200 while (sectors) {
2201 int s = sectors;
2202 struct md_rdev *rdev;
2203 sector_t addr;
2204 int ok;
2205
2206 if (s > (PAGE_SIZE>>9))
2207 s = PAGE_SIZE >> 9;
2208
2209 rdev = conf->mirrors[dr].rdev;
2210 addr = r10_bio->devs[0].addr + sect,
2211 ok = sync_page_io(rdev,
2212 addr,
2213 s << 9,
2214 bio->bi_io_vec[idx].bv_page,
2215 READ, false);
2216 if (ok) {
2217 rdev = conf->mirrors[dw].rdev;
2218 addr = r10_bio->devs[1].addr + sect;
2219 ok = sync_page_io(rdev,
2220 addr,
2221 s << 9,
2222 bio->bi_io_vec[idx].bv_page,
2223 WRITE, false);
2224 if (!ok) {
2225 set_bit(WriteErrorSeen, &rdev->flags);
2226 if (!test_and_set_bit(WantReplacement,
2227 &rdev->flags))
2228 set_bit(MD_RECOVERY_NEEDED,
2229 &rdev->mddev->recovery);
2230 }
2231 }
2232 if (!ok) {
2233 /* We don't worry if we cannot set a bad block -
2234 * it really is bad so there is no loss in not
2235 * recording it yet
2236 */
2237 rdev_set_badblocks(rdev, addr, s, 0);
2238
2239 if (rdev != conf->mirrors[dw].rdev) {
2240 /* need bad block on destination too */
2241 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2242 addr = r10_bio->devs[1].addr + sect;
2243 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2244 if (!ok) {
2245 /* just abort the recovery */
2246 printk(KERN_NOTICE
2247 "md/raid10:%s: recovery aborted"
2248 " due to read error\n",
2249 mdname(mddev));
2250
2251 conf->mirrors[dw].recovery_disabled
2252 = mddev->recovery_disabled;
2253 set_bit(MD_RECOVERY_INTR,
2254 &mddev->recovery);
2255 break;
2256 }
2257 }
2258 }
2259
2260 sectors -= s;
2261 sect += s;
2262 idx++;
2263 }
2264 }
2265
2266 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2267 {
2268 struct r10conf *conf = mddev->private;
2269 int d;
2270 struct bio *wbio, *wbio2;
2271
2272 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2273 fix_recovery_read_error(r10_bio);
2274 end_sync_request(r10_bio);
2275 return;
2276 }
2277
2278 /*
2279 * share the pages with the first bio
2280 * and submit the write request
2281 */
2282 d = r10_bio->devs[1].devnum;
2283 wbio = r10_bio->devs[1].bio;
2284 wbio2 = r10_bio->devs[1].repl_bio;
2285 /* Need to test wbio2->bi_end_io before we call
2286 * generic_make_request as if the former is NULL,
2287 * the latter is free to free wbio2.
2288 */
2289 if (wbio2 && !wbio2->bi_end_io)
2290 wbio2 = NULL;
2291 if (wbio->bi_end_io) {
2292 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2293 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2294 generic_make_request(wbio);
2295 }
2296 if (wbio2) {
2297 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2298 md_sync_acct(conf->mirrors[d].replacement->bdev,
2299 bio_sectors(wbio2));
2300 generic_make_request(wbio2);
2301 }
2302 }
2303
2304
2305 /*
2306 * Used by fix_read_error() to decay the per rdev read_errors.
2307 * We halve the read error count for every hour that has elapsed
2308 * since the last recorded read error.
2309 *
2310 */
2311 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2312 {
2313 struct timespec cur_time_mon;
2314 unsigned long hours_since_last;
2315 unsigned int read_errors = atomic_read(&rdev->read_errors);
2316
2317 ktime_get_ts(&cur_time_mon);
2318
2319 if (rdev->last_read_error.tv_sec == 0 &&
2320 rdev->last_read_error.tv_nsec == 0) {
2321 /* first time we've seen a read error */
2322 rdev->last_read_error = cur_time_mon;
2323 return;
2324 }
2325
2326 hours_since_last = (cur_time_mon.tv_sec -
2327 rdev->last_read_error.tv_sec) / 3600;
2328
2329 rdev->last_read_error = cur_time_mon;
2330
2331 /*
2332 * if hours_since_last is > the number of bits in read_errors
2333 * just set read errors to 0. We do this to avoid
2334 * overflowing the shift of read_errors by hours_since_last.
2335 */
2336 if (hours_since_last >= 8 * sizeof(read_errors))
2337 atomic_set(&rdev->read_errors, 0);
2338 else
2339 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2340 }
2341
2342 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2343 int sectors, struct page *page, int rw)
2344 {
2345 sector_t first_bad;
2346 int bad_sectors;
2347
2348 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2349 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2350 return -1;
2351 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2352 /* success */
2353 return 1;
2354 if (rw == WRITE) {
2355 set_bit(WriteErrorSeen, &rdev->flags);
2356 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2357 set_bit(MD_RECOVERY_NEEDED,
2358 &rdev->mddev->recovery);
2359 }
2360 /* need to record an error - either for the block or the device */
2361 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2362 md_error(rdev->mddev, rdev);
2363 return 0;
2364 }
2365
2366 /*
2367 * This is a kernel thread which:
2368 *
2369 * 1. Retries failed read operations on working mirrors.
2370 * 2. Updates the raid superblock when problems encounter.
2371 * 3. Performs writes following reads for array synchronising.
2372 */
2373
2374 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2375 {
2376 int sect = 0; /* Offset from r10_bio->sector */
2377 int sectors = r10_bio->sectors;
2378 struct md_rdev*rdev;
2379 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2380 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2381
2382 /* still own a reference to this rdev, so it cannot
2383 * have been cleared recently.
2384 */
2385 rdev = conf->mirrors[d].rdev;
2386
2387 if (test_bit(Faulty, &rdev->flags))
2388 /* drive has already been failed, just ignore any
2389 more fix_read_error() attempts */
2390 return;
2391
2392 check_decay_read_errors(mddev, rdev);
2393 atomic_inc(&rdev->read_errors);
2394 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2395 char b[BDEVNAME_SIZE];
2396 bdevname(rdev->bdev, b);
2397
2398 printk(KERN_NOTICE
2399 "md/raid10:%s: %s: Raid device exceeded "
2400 "read_error threshold [cur %d:max %d]\n",
2401 mdname(mddev), b,
2402 atomic_read(&rdev->read_errors), max_read_errors);
2403 printk(KERN_NOTICE
2404 "md/raid10:%s: %s: Failing raid device\n",
2405 mdname(mddev), b);
2406 md_error(mddev, conf->mirrors[d].rdev);
2407 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2408 return;
2409 }
2410
2411 while(sectors) {
2412 int s = sectors;
2413 int sl = r10_bio->read_slot;
2414 int success = 0;
2415 int start;
2416
2417 if (s > (PAGE_SIZE>>9))
2418 s = PAGE_SIZE >> 9;
2419
2420 rcu_read_lock();
2421 do {
2422 sector_t first_bad;
2423 int bad_sectors;
2424
2425 d = r10_bio->devs[sl].devnum;
2426 rdev = rcu_dereference(conf->mirrors[d].rdev);
2427 if (rdev &&
2428 !test_bit(Unmerged, &rdev->flags) &&
2429 test_bit(In_sync, &rdev->flags) &&
2430 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2431 &first_bad, &bad_sectors) == 0) {
2432 atomic_inc(&rdev->nr_pending);
2433 rcu_read_unlock();
2434 success = sync_page_io(rdev,
2435 r10_bio->devs[sl].addr +
2436 sect,
2437 s<<9,
2438 conf->tmppage, READ, false);
2439 rdev_dec_pending(rdev, mddev);
2440 rcu_read_lock();
2441 if (success)
2442 break;
2443 }
2444 sl++;
2445 if (sl == conf->copies)
2446 sl = 0;
2447 } while (!success && sl != r10_bio->read_slot);
2448 rcu_read_unlock();
2449
2450 if (!success) {
2451 /* Cannot read from anywhere, just mark the block
2452 * as bad on the first device to discourage future
2453 * reads.
2454 */
2455 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2456 rdev = conf->mirrors[dn].rdev;
2457
2458 if (!rdev_set_badblocks(
2459 rdev,
2460 r10_bio->devs[r10_bio->read_slot].addr
2461 + sect,
2462 s, 0)) {
2463 md_error(mddev, rdev);
2464 r10_bio->devs[r10_bio->read_slot].bio
2465 = IO_BLOCKED;
2466 }
2467 break;
2468 }
2469
2470 start = sl;
2471 /* write it back and re-read */
2472 rcu_read_lock();
2473 while (sl != r10_bio->read_slot) {
2474 char b[BDEVNAME_SIZE];
2475
2476 if (sl==0)
2477 sl = conf->copies;
2478 sl--;
2479 d = r10_bio->devs[sl].devnum;
2480 rdev = rcu_dereference(conf->mirrors[d].rdev);
2481 if (!rdev ||
2482 test_bit(Unmerged, &rdev->flags) ||
2483 !test_bit(In_sync, &rdev->flags))
2484 continue;
2485
2486 atomic_inc(&rdev->nr_pending);
2487 rcu_read_unlock();
2488 if (r10_sync_page_io(rdev,
2489 r10_bio->devs[sl].addr +
2490 sect,
2491 s, conf->tmppage, WRITE)
2492 == 0) {
2493 /* Well, this device is dead */
2494 printk(KERN_NOTICE
2495 "md/raid10:%s: read correction "
2496 "write failed"
2497 " (%d sectors at %llu on %s)\n",
2498 mdname(mddev), s,
2499 (unsigned long long)(
2500 sect +
2501 choose_data_offset(r10_bio,
2502 rdev)),
2503 bdevname(rdev->bdev, b));
2504 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2505 "drive\n",
2506 mdname(mddev),
2507 bdevname(rdev->bdev, b));
2508 }
2509 rdev_dec_pending(rdev, mddev);
2510 rcu_read_lock();
2511 }
2512 sl = start;
2513 while (sl != r10_bio->read_slot) {
2514 char b[BDEVNAME_SIZE];
2515
2516 if (sl==0)
2517 sl = conf->copies;
2518 sl--;
2519 d = r10_bio->devs[sl].devnum;
2520 rdev = rcu_dereference(conf->mirrors[d].rdev);
2521 if (!rdev ||
2522 !test_bit(In_sync, &rdev->flags))
2523 continue;
2524
2525 atomic_inc(&rdev->nr_pending);
2526 rcu_read_unlock();
2527 switch (r10_sync_page_io(rdev,
2528 r10_bio->devs[sl].addr +
2529 sect,
2530 s, conf->tmppage,
2531 READ)) {
2532 case 0:
2533 /* Well, this device is dead */
2534 printk(KERN_NOTICE
2535 "md/raid10:%s: unable to read back "
2536 "corrected sectors"
2537 " (%d sectors at %llu on %s)\n",
2538 mdname(mddev), s,
2539 (unsigned long long)(
2540 sect +
2541 choose_data_offset(r10_bio, rdev)),
2542 bdevname(rdev->bdev, b));
2543 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2544 "drive\n",
2545 mdname(mddev),
2546 bdevname(rdev->bdev, b));
2547 break;
2548 case 1:
2549 printk(KERN_INFO
2550 "md/raid10:%s: read error corrected"
2551 " (%d sectors at %llu on %s)\n",
2552 mdname(mddev), s,
2553 (unsigned long long)(
2554 sect +
2555 choose_data_offset(r10_bio, rdev)),
2556 bdevname(rdev->bdev, b));
2557 atomic_add(s, &rdev->corrected_errors);
2558 }
2559
2560 rdev_dec_pending(rdev, mddev);
2561 rcu_read_lock();
2562 }
2563 rcu_read_unlock();
2564
2565 sectors -= s;
2566 sect += s;
2567 }
2568 }
2569
2570 static int narrow_write_error(struct r10bio *r10_bio, int i)
2571 {
2572 struct bio *bio = r10_bio->master_bio;
2573 struct mddev *mddev = r10_bio->mddev;
2574 struct r10conf *conf = mddev->private;
2575 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2576 /* bio has the data to be written to slot 'i' where
2577 * we just recently had a write error.
2578 * We repeatedly clone the bio and trim down to one block,
2579 * then try the write. Where the write fails we record
2580 * a bad block.
2581 * It is conceivable that the bio doesn't exactly align with
2582 * blocks. We must handle this.
2583 *
2584 * We currently own a reference to the rdev.
2585 */
2586
2587 int block_sectors;
2588 sector_t sector;
2589 int sectors;
2590 int sect_to_write = r10_bio->sectors;
2591 int ok = 1;
2592
2593 if (rdev->badblocks.shift < 0)
2594 return 0;
2595
2596 block_sectors = 1 << rdev->badblocks.shift;
2597 sector = r10_bio->sector;
2598 sectors = ((r10_bio->sector + block_sectors)
2599 & ~(sector_t)(block_sectors - 1))
2600 - sector;
2601
2602 while (sect_to_write) {
2603 struct bio *wbio;
2604 if (sectors > sect_to_write)
2605 sectors = sect_to_write;
2606 /* Write at 'sector' for 'sectors' */
2607 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2608 md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2609 wbio->bi_sector = (r10_bio->devs[i].addr+
2610 choose_data_offset(r10_bio, rdev) +
2611 (sector - r10_bio->sector));
2612 wbio->bi_bdev = rdev->bdev;
2613 if (submit_bio_wait(WRITE, wbio) < 0)
2614 /* Failure! */
2615 ok = rdev_set_badblocks(rdev, sector,
2616 sectors, 0)
2617 && ok;
2618
2619 bio_put(wbio);
2620 sect_to_write -= sectors;
2621 sector += sectors;
2622 sectors = block_sectors;
2623 }
2624 return ok;
2625 }
2626
2627 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2628 {
2629 int slot = r10_bio->read_slot;
2630 struct bio *bio;
2631 struct r10conf *conf = mddev->private;
2632 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2633 char b[BDEVNAME_SIZE];
2634 unsigned long do_sync;
2635 int max_sectors;
2636
2637 /* we got a read error. Maybe the drive is bad. Maybe just
2638 * the block and we can fix it.
2639 * We freeze all other IO, and try reading the block from
2640 * other devices. When we find one, we re-write
2641 * and check it that fixes the read error.
2642 * This is all done synchronously while the array is
2643 * frozen.
2644 */
2645 bio = r10_bio->devs[slot].bio;
2646 bdevname(bio->bi_bdev, b);
2647 bio_put(bio);
2648 r10_bio->devs[slot].bio = NULL;
2649
2650 if (mddev->ro == 0) {
2651 freeze_array(conf, 1);
2652 fix_read_error(conf, mddev, r10_bio);
2653 unfreeze_array(conf);
2654 } else
2655 r10_bio->devs[slot].bio = IO_BLOCKED;
2656
2657 rdev_dec_pending(rdev, mddev);
2658
2659 read_more:
2660 rdev = read_balance(conf, r10_bio, &max_sectors);
2661 if (rdev == NULL) {
2662 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2663 " read error for block %llu\n",
2664 mdname(mddev), b,
2665 (unsigned long long)r10_bio->sector);
2666 raid_end_bio_io(r10_bio);
2667 return;
2668 }
2669
2670 do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2671 slot = r10_bio->read_slot;
2672 printk_ratelimited(
2673 KERN_ERR
2674 "md/raid10:%s: %s: redirecting "
2675 "sector %llu to another mirror\n",
2676 mdname(mddev),
2677 bdevname(rdev->bdev, b),
2678 (unsigned long long)r10_bio->sector);
2679 bio = bio_clone_mddev(r10_bio->master_bio,
2680 GFP_NOIO, mddev);
2681 md_trim_bio(bio,
2682 r10_bio->sector - bio->bi_sector,
2683 max_sectors);
2684 r10_bio->devs[slot].bio = bio;
2685 r10_bio->devs[slot].rdev = rdev;
2686 bio->bi_sector = r10_bio->devs[slot].addr
2687 + choose_data_offset(r10_bio, rdev);
2688 bio->bi_bdev = rdev->bdev;
2689 bio->bi_rw = READ | do_sync;
2690 bio->bi_private = r10_bio;
2691 bio->bi_end_io = raid10_end_read_request;
2692 if (max_sectors < r10_bio->sectors) {
2693 /* Drat - have to split this up more */
2694 struct bio *mbio = r10_bio->master_bio;
2695 int sectors_handled =
2696 r10_bio->sector + max_sectors
2697 - mbio->bi_sector;
2698 r10_bio->sectors = max_sectors;
2699 spin_lock_irq(&conf->device_lock);
2700 if (mbio->bi_phys_segments == 0)
2701 mbio->bi_phys_segments = 2;
2702 else
2703 mbio->bi_phys_segments++;
2704 spin_unlock_irq(&conf->device_lock);
2705 generic_make_request(bio);
2706
2707 r10_bio = mempool_alloc(conf->r10bio_pool,
2708 GFP_NOIO);
2709 r10_bio->master_bio = mbio;
2710 r10_bio->sectors = bio_sectors(mbio) - sectors_handled;
2711 r10_bio->state = 0;
2712 set_bit(R10BIO_ReadError,
2713 &r10_bio->state);
2714 r10_bio->mddev = mddev;
2715 r10_bio->sector = mbio->bi_sector
2716 + sectors_handled;
2717
2718 goto read_more;
2719 } else
2720 generic_make_request(bio);
2721 }
2722
2723 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2724 {
2725 /* Some sort of write request has finished and it
2726 * succeeded in writing where we thought there was a
2727 * bad block. So forget the bad block.
2728 * Or possibly if failed and we need to record
2729 * a bad block.
2730 */
2731 int m;
2732 struct md_rdev *rdev;
2733
2734 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2735 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2736 for (m = 0; m < conf->copies; m++) {
2737 int dev = r10_bio->devs[m].devnum;
2738 rdev = conf->mirrors[dev].rdev;
2739 if (r10_bio->devs[m].bio == NULL)
2740 continue;
2741 if (test_bit(BIO_UPTODATE,
2742 &r10_bio->devs[m].bio->bi_flags)) {
2743 rdev_clear_badblocks(
2744 rdev,
2745 r10_bio->devs[m].addr,
2746 r10_bio->sectors, 0);
2747 } else {
2748 if (!rdev_set_badblocks(
2749 rdev,
2750 r10_bio->devs[m].addr,
2751 r10_bio->sectors, 0))
2752 md_error(conf->mddev, rdev);
2753 }
2754 rdev = conf->mirrors[dev].replacement;
2755 if (r10_bio->devs[m].repl_bio == NULL)
2756 continue;
2757 if (test_bit(BIO_UPTODATE,
2758 &r10_bio->devs[m].repl_bio->bi_flags)) {
2759 rdev_clear_badblocks(
2760 rdev,
2761 r10_bio->devs[m].addr,
2762 r10_bio->sectors, 0);
2763 } else {
2764 if (!rdev_set_badblocks(
2765 rdev,
2766 r10_bio->devs[m].addr,
2767 r10_bio->sectors, 0))
2768 md_error(conf->mddev, rdev);
2769 }
2770 }
2771 put_buf(r10_bio);
2772 } else {
2773 for (m = 0; m < conf->copies; m++) {
2774 int dev = r10_bio->devs[m].devnum;
2775 struct bio *bio = r10_bio->devs[m].bio;
2776 rdev = conf->mirrors[dev].rdev;
2777 if (bio == IO_MADE_GOOD) {
2778 rdev_clear_badblocks(
2779 rdev,
2780 r10_bio->devs[m].addr,
2781 r10_bio->sectors, 0);
2782 rdev_dec_pending(rdev, conf->mddev);
2783 } else if (bio != NULL &&
2784 !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2785 if (!narrow_write_error(r10_bio, m)) {
2786 md_error(conf->mddev, rdev);
2787 set_bit(R10BIO_Degraded,
2788 &r10_bio->state);
2789 }
2790 rdev_dec_pending(rdev, conf->mddev);
2791 }
2792 bio = r10_bio->devs[m].repl_bio;
2793 rdev = conf->mirrors[dev].replacement;
2794 if (rdev && bio == IO_MADE_GOOD) {
2795 rdev_clear_badblocks(
2796 rdev,
2797 r10_bio->devs[m].addr,
2798 r10_bio->sectors, 0);
2799 rdev_dec_pending(rdev, conf->mddev);
2800 }
2801 }
2802 if (test_bit(R10BIO_WriteError,
2803 &r10_bio->state))
2804 close_write(r10_bio);
2805 raid_end_bio_io(r10_bio);
2806 }
2807 }
2808
2809 static void raid10d(struct md_thread *thread)
2810 {
2811 struct mddev *mddev = thread->mddev;
2812 struct r10bio *r10_bio;
2813 unsigned long flags;
2814 struct r10conf *conf = mddev->private;
2815 struct list_head *head = &conf->retry_list;
2816 struct blk_plug plug;
2817
2818 md_check_recovery(mddev);
2819
2820 blk_start_plug(&plug);
2821 for (;;) {
2822
2823 flush_pending_writes(conf);
2824
2825 spin_lock_irqsave(&conf->device_lock, flags);
2826 if (list_empty(head)) {
2827 spin_unlock_irqrestore(&conf->device_lock, flags);
2828 break;
2829 }
2830 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2831 list_del(head->prev);
2832 conf->nr_queued--;
2833 spin_unlock_irqrestore(&conf->device_lock, flags);
2834
2835 mddev = r10_bio->mddev;
2836 conf = mddev->private;
2837 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2838 test_bit(R10BIO_WriteError, &r10_bio->state))
2839 handle_write_completed(conf, r10_bio);
2840 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2841 reshape_request_write(mddev, r10_bio);
2842 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2843 sync_request_write(mddev, r10_bio);
2844 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2845 recovery_request_write(mddev, r10_bio);
2846 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2847 handle_read_error(mddev, r10_bio);
2848 else {
2849 /* just a partial read to be scheduled from a
2850 * separate context
2851 */
2852 int slot = r10_bio->read_slot;
2853 generic_make_request(r10_bio->devs[slot].bio);
2854 }
2855
2856 cond_resched();
2857 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2858 md_check_recovery(mddev);
2859 }
2860 blk_finish_plug(&plug);
2861 }
2862
2863
2864 static int init_resync(struct r10conf *conf)
2865 {
2866 int buffs;
2867 int i;
2868
2869 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2870 BUG_ON(conf->r10buf_pool);
2871 conf->have_replacement = 0;
2872 for (i = 0; i < conf->geo.raid_disks; i++)
2873 if (conf->mirrors[i].replacement)
2874 conf->have_replacement = 1;
2875 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2876 if (!conf->r10buf_pool)
2877 return -ENOMEM;
2878 conf->next_resync = 0;
2879 return 0;
2880 }
2881
2882 /*
2883 * perform a "sync" on one "block"
2884 *
2885 * We need to make sure that no normal I/O request - particularly write
2886 * requests - conflict with active sync requests.
2887 *
2888 * This is achieved by tracking pending requests and a 'barrier' concept
2889 * that can be installed to exclude normal IO requests.
2890 *
2891 * Resync and recovery are handled very differently.
2892 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2893 *
2894 * For resync, we iterate over virtual addresses, read all copies,
2895 * and update if there are differences. If only one copy is live,
2896 * skip it.
2897 * For recovery, we iterate over physical addresses, read a good
2898 * value for each non-in_sync drive, and over-write.
2899 *
2900 * So, for recovery we may have several outstanding complex requests for a
2901 * given address, one for each out-of-sync device. We model this by allocating
2902 * a number of r10_bio structures, one for each out-of-sync device.
2903 * As we setup these structures, we collect all bio's together into a list
2904 * which we then process collectively to add pages, and then process again
2905 * to pass to generic_make_request.
2906 *
2907 * The r10_bio structures are linked using a borrowed master_bio pointer.
2908 * This link is counted in ->remaining. When the r10_bio that points to NULL
2909 * has its remaining count decremented to 0, the whole complex operation
2910 * is complete.
2911 *
2912 */
2913
2914 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
2915 int *skipped, int go_faster)
2916 {
2917 struct r10conf *conf = mddev->private;
2918 struct r10bio *r10_bio;
2919 struct bio *biolist = NULL, *bio;
2920 sector_t max_sector, nr_sectors;
2921 int i;
2922 int max_sync;
2923 sector_t sync_blocks;
2924 sector_t sectors_skipped = 0;
2925 int chunks_skipped = 0;
2926 sector_t chunk_mask = conf->geo.chunk_mask;
2927
2928 if (!conf->r10buf_pool)
2929 if (init_resync(conf))
2930 return 0;
2931
2932 /*
2933 * Allow skipping a full rebuild for incremental assembly
2934 * of a clean array, like RAID1 does.
2935 */
2936 if (mddev->bitmap == NULL &&
2937 mddev->recovery_cp == MaxSector &&
2938 mddev->reshape_position == MaxSector &&
2939 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2940 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2941 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2942 conf->fullsync == 0) {
2943 *skipped = 1;
2944 return mddev->dev_sectors - sector_nr;
2945 }
2946
2947 skipped:
2948 max_sector = mddev->dev_sectors;
2949 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2950 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2951 max_sector = mddev->resync_max_sectors;
2952 if (sector_nr >= max_sector) {
2953 /* If we aborted, we need to abort the
2954 * sync on the 'current' bitmap chucks (there can
2955 * be several when recovering multiple devices).
2956 * as we may have started syncing it but not finished.
2957 * We can find the current address in
2958 * mddev->curr_resync, but for recovery,
2959 * we need to convert that to several
2960 * virtual addresses.
2961 */
2962 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2963 end_reshape(conf);
2964 close_sync(conf);
2965 return 0;
2966 }
2967
2968 if (mddev->curr_resync < max_sector) { /* aborted */
2969 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2970 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2971 &sync_blocks, 1);
2972 else for (i = 0; i < conf->geo.raid_disks; i++) {
2973 sector_t sect =
2974 raid10_find_virt(conf, mddev->curr_resync, i);
2975 bitmap_end_sync(mddev->bitmap, sect,
2976 &sync_blocks, 1);
2977 }
2978 } else {
2979 /* completed sync */
2980 if ((!mddev->bitmap || conf->fullsync)
2981 && conf->have_replacement
2982 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2983 /* Completed a full sync so the replacements
2984 * are now fully recovered.
2985 */
2986 for (i = 0; i < conf->geo.raid_disks; i++)
2987 if (conf->mirrors[i].replacement)
2988 conf->mirrors[i].replacement
2989 ->recovery_offset
2990 = MaxSector;
2991 }
2992 conf->fullsync = 0;
2993 }
2994 bitmap_close_sync(mddev->bitmap);
2995 close_sync(conf);
2996 *skipped = 1;
2997 return sectors_skipped;
2998 }
2999
3000 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3001 return reshape_request(mddev, sector_nr, skipped);
3002
3003 if (chunks_skipped >= conf->geo.raid_disks) {
3004 /* if there has been nothing to do on any drive,
3005 * then there is nothing to do at all..
3006 */
3007 *skipped = 1;
3008 return (max_sector - sector_nr) + sectors_skipped;
3009 }
3010
3011 if (max_sector > mddev->resync_max)
3012 max_sector = mddev->resync_max; /* Don't do IO beyond here */
3013
3014 /* make sure whole request will fit in a chunk - if chunks
3015 * are meaningful
3016 */
3017 if (conf->geo.near_copies < conf->geo.raid_disks &&
3018 max_sector > (sector_nr | chunk_mask))
3019 max_sector = (sector_nr | chunk_mask) + 1;
3020 /*
3021 * If there is non-resync activity waiting for us then
3022 * put in a delay to throttle resync.
3023 */
3024 if (!go_faster && conf->nr_waiting)
3025 msleep_interruptible(1000);
3026
3027 /* Again, very different code for resync and recovery.
3028 * Both must result in an r10bio with a list of bios that
3029 * have bi_end_io, bi_sector, bi_bdev set,
3030 * and bi_private set to the r10bio.
3031 * For recovery, we may actually create several r10bios
3032 * with 2 bios in each, that correspond to the bios in the main one.
3033 * In this case, the subordinate r10bios link back through a
3034 * borrowed master_bio pointer, and the counter in the master
3035 * includes a ref from each subordinate.
3036 */
3037 /* First, we decide what to do and set ->bi_end_io
3038 * To end_sync_read if we want to read, and
3039 * end_sync_write if we will want to write.
3040 */
3041
3042 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3043 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3044 /* recovery... the complicated one */
3045 int j;
3046 r10_bio = NULL;
3047
3048 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3049 int still_degraded;
3050 struct r10bio *rb2;
3051 sector_t sect;
3052 int must_sync;
3053 int any_working;
3054 struct raid10_info *mirror = &conf->mirrors[i];
3055
3056 if ((mirror->rdev == NULL ||
3057 test_bit(In_sync, &mirror->rdev->flags))
3058 &&
3059 (mirror->replacement == NULL ||
3060 test_bit(Faulty,
3061 &mirror->replacement->flags)))
3062 continue;
3063
3064 still_degraded = 0;
3065 /* want to reconstruct this device */
3066 rb2 = r10_bio;
3067 sect = raid10_find_virt(conf, sector_nr, i);
3068 if (sect >= mddev->resync_max_sectors) {
3069 /* last stripe is not complete - don't
3070 * try to recover this sector.
3071 */
3072 continue;
3073 }
3074 /* Unless we are doing a full sync, or a replacement
3075 * we only need to recover the block if it is set in
3076 * the bitmap
3077 */
3078 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3079 &sync_blocks, 1);
3080 if (sync_blocks < max_sync)
3081 max_sync = sync_blocks;
3082 if (!must_sync &&
3083 mirror->replacement == NULL &&
3084 !conf->fullsync) {
3085 /* yep, skip the sync_blocks here, but don't assume
3086 * that there will never be anything to do here
3087 */
3088 chunks_skipped = -1;
3089 continue;
3090 }
3091
3092 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3093 raise_barrier(conf, rb2 != NULL);
3094 atomic_set(&r10_bio->remaining, 0);
3095
3096 r10_bio->master_bio = (struct bio*)rb2;
3097 if (rb2)
3098 atomic_inc(&rb2->remaining);
3099 r10_bio->mddev = mddev;
3100 set_bit(R10BIO_IsRecover, &r10_bio->state);
3101 r10_bio->sector = sect;
3102
3103 raid10_find_phys(conf, r10_bio);
3104
3105 /* Need to check if the array will still be
3106 * degraded
3107 */
3108 for (j = 0; j < conf->geo.raid_disks; j++)
3109 if (conf->mirrors[j].rdev == NULL ||
3110 test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
3111 still_degraded = 1;
3112 break;
3113 }
3114
3115 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3116 &sync_blocks, still_degraded);
3117
3118 any_working = 0;
3119 for (j=0; j<conf->copies;j++) {
3120 int k;
3121 int d = r10_bio->devs[j].devnum;
3122 sector_t from_addr, to_addr;
3123 struct md_rdev *rdev;
3124 sector_t sector, first_bad;
3125 int bad_sectors;
3126 if (!conf->mirrors[d].rdev ||
3127 !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
3128 continue;
3129 /* This is where we read from */
3130 any_working = 1;
3131 rdev = conf->mirrors[d].rdev;
3132 sector = r10_bio->devs[j].addr;
3133
3134 if (is_badblock(rdev, sector, max_sync,
3135 &first_bad, &bad_sectors)) {
3136 if (first_bad > sector)
3137 max_sync = first_bad - sector;
3138 else {
3139 bad_sectors -= (sector
3140 - first_bad);
3141 if (max_sync > bad_sectors)
3142 max_sync = bad_sectors;
3143 continue;
3144 }
3145 }
3146 bio = r10_bio->devs[0].bio;
3147 bio_reset(bio);
3148 bio->bi_next = biolist;
3149 biolist = bio;
3150 bio->bi_private = r10_bio;
3151 bio->bi_end_io = end_sync_read;
3152 bio->bi_rw = READ;
3153 from_addr = r10_bio->devs[j].addr;
3154 bio->bi_sector = from_addr + rdev->data_offset;
3155 bio->bi_bdev = rdev->bdev;
3156 atomic_inc(&rdev->nr_pending);
3157 /* and we write to 'i' (if not in_sync) */
3158
3159 for (k=0; k<conf->copies; k++)
3160 if (r10_bio->devs[k].devnum == i)
3161 break;
3162 BUG_ON(k == conf->copies);
3163 to_addr = r10_bio->devs[k].addr;
3164 r10_bio->devs[0].devnum = d;
3165 r10_bio->devs[0].addr = from_addr;
3166 r10_bio->devs[1].devnum = i;
3167 r10_bio->devs[1].addr = to_addr;
3168
3169 rdev = mirror->rdev;
3170 if (!test_bit(In_sync, &rdev->flags)) {
3171 bio = r10_bio->devs[1].bio;
3172 bio_reset(bio);
3173 bio->bi_next = biolist;
3174 biolist = bio;
3175 bio->bi_private = r10_bio;
3176 bio->bi_end_io = end_sync_write;
3177 bio->bi_rw = WRITE;
3178 bio->bi_sector = to_addr
3179 + rdev->data_offset;
3180 bio->bi_bdev = rdev->bdev;
3181 atomic_inc(&r10_bio->remaining);
3182 } else
3183 r10_bio->devs[1].bio->bi_end_io = NULL;
3184
3185 /* and maybe write to replacement */
3186 bio = r10_bio->devs[1].repl_bio;
3187 if (bio)
3188 bio->bi_end_io = NULL;
3189 rdev = mirror->replacement;
3190 /* Note: if rdev != NULL, then bio
3191 * cannot be NULL as r10buf_pool_alloc will
3192 * have allocated it.
3193 * So the second test here is pointless.
3194 * But it keeps semantic-checkers happy, and
3195 * this comment keeps human reviewers
3196 * happy.
3197 */
3198 if (rdev == NULL || bio == NULL ||
3199 test_bit(Faulty, &rdev->flags))
3200 break;
3201 bio_reset(bio);
3202 bio->bi_next = biolist;
3203 biolist = bio;
3204 bio->bi_private = r10_bio;
3205 bio->bi_end_io = end_sync_write;
3206 bio->bi_rw = WRITE;
3207 bio->bi_sector = to_addr + rdev->data_offset;
3208 bio->bi_bdev = rdev->bdev;
3209 atomic_inc(&r10_bio->remaining);
3210 break;
3211 }
3212 if (j == conf->copies) {
3213 /* Cannot recover, so abort the recovery or
3214 * record a bad block */
3215 if (any_working) {
3216 /* problem is that there are bad blocks
3217 * on other device(s)
3218 */
3219 int k;
3220 for (k = 0; k < conf->copies; k++)
3221 if (r10_bio->devs[k].devnum == i)
3222 break;
3223 if (!test_bit(In_sync,
3224 &mirror->rdev->flags)
3225 && !rdev_set_badblocks(
3226 mirror->rdev,
3227 r10_bio->devs[k].addr,
3228 max_sync, 0))
3229 any_working = 0;
3230 if (mirror->replacement &&
3231 !rdev_set_badblocks(
3232 mirror->replacement,
3233 r10_bio->devs[k].addr,
3234 max_sync, 0))
3235 any_working = 0;
3236 }
3237 if (!any_working) {
3238 if (!test_and_set_bit(MD_RECOVERY_INTR,
3239 &mddev->recovery))
3240 printk(KERN_INFO "md/raid10:%s: insufficient "
3241 "working devices for recovery.\n",
3242 mdname(mddev));
3243 mirror->recovery_disabled
3244 = mddev->recovery_disabled;
3245 }
3246 put_buf(r10_bio);
3247 if (rb2)
3248 atomic_dec(&rb2->remaining);
3249 r10_bio = rb2;
3250 break;
3251 }
3252 }
3253 if (biolist == NULL) {
3254 while (r10_bio) {
3255 struct r10bio *rb2 = r10_bio;
3256 r10_bio = (struct r10bio*) rb2->master_bio;
3257 rb2->master_bio = NULL;
3258 put_buf(rb2);
3259 }
3260 goto giveup;
3261 }
3262 } else {
3263 /* resync. Schedule a read for every block at this virt offset */
3264 int count = 0;
3265
3266 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3267
3268 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3269 &sync_blocks, mddev->degraded) &&
3270 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3271 &mddev->recovery)) {
3272 /* We can skip this block */
3273 *skipped = 1;
3274 return sync_blocks + sectors_skipped;
3275 }
3276 if (sync_blocks < max_sync)
3277 max_sync = sync_blocks;
3278 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3279
3280 r10_bio->mddev = mddev;
3281 atomic_set(&r10_bio->remaining, 0);
3282 raise_barrier(conf, 0);
3283 conf->next_resync = sector_nr;
3284
3285 r10_bio->master_bio = NULL;
3286 r10_bio->sector = sector_nr;
3287 set_bit(R10BIO_IsSync, &r10_bio->state);
3288 raid10_find_phys(conf, r10_bio);
3289 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3290
3291 for (i = 0; i < conf->copies; i++) {
3292 int d = r10_bio->devs[i].devnum;
3293 sector_t first_bad, sector;
3294 int bad_sectors;
3295
3296 if (r10_bio->devs[i].repl_bio)
3297 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3298
3299 bio = r10_bio->devs[i].bio;
3300 bio_reset(bio);
3301 clear_bit(BIO_UPTODATE, &bio->bi_flags);
3302 if (conf->mirrors[d].rdev == NULL ||
3303 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
3304 continue;
3305 sector = r10_bio->devs[i].addr;
3306 if (is_badblock(conf->mirrors[d].rdev,
3307 sector, max_sync,
3308 &first_bad, &bad_sectors)) {
3309 if (first_bad > sector)
3310 max_sync = first_bad - sector;
3311 else {
3312 bad_sectors -= (sector - first_bad);
3313 if (max_sync > bad_sectors)
3314 max_sync = bad_sectors;
3315 continue;
3316 }
3317 }
3318 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3319 atomic_inc(&r10_bio->remaining);
3320 bio->bi_next = biolist;
3321 biolist = bio;
3322 bio->bi_private = r10_bio;
3323 bio->bi_end_io = end_sync_read;
3324 bio->bi_rw = READ;
3325 bio->bi_sector = sector +
3326 conf->mirrors[d].rdev->data_offset;
3327 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
3328 count++;
3329
3330 if (conf->mirrors[d].replacement == NULL ||
3331 test_bit(Faulty,
3332 &conf->mirrors[d].replacement->flags))
3333 continue;
3334
3335 /* Need to set up for writing to the replacement */
3336 bio = r10_bio->devs[i].repl_bio;
3337 bio_reset(bio);
3338 clear_bit(BIO_UPTODATE, &bio->bi_flags);
3339
3340 sector = r10_bio->devs[i].addr;
3341 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3342 bio->bi_next = biolist;
3343 biolist = bio;
3344 bio->bi_private = r10_bio;
3345 bio->bi_end_io = end_sync_write;
3346 bio->bi_rw = WRITE;
3347 bio->bi_sector = sector +
3348 conf->mirrors[d].replacement->data_offset;
3349 bio->bi_bdev = conf->mirrors[d].replacement->bdev;
3350 count++;
3351 }
3352
3353 if (count < 2) {
3354 for (i=0; i<conf->copies; i++) {
3355 int d = r10_bio->devs[i].devnum;
3356 if (r10_bio->devs[i].bio->bi_end_io)
3357 rdev_dec_pending(conf->mirrors[d].rdev,
3358 mddev);
3359 if (r10_bio->devs[i].repl_bio &&
3360 r10_bio->devs[i].repl_bio->bi_end_io)
3361 rdev_dec_pending(
3362 conf->mirrors[d].replacement,
3363 mddev);
3364 }
3365 put_buf(r10_bio);
3366 biolist = NULL;
3367 goto giveup;
3368 }
3369 }
3370
3371 nr_sectors = 0;
3372 if (sector_nr + max_sync < max_sector)
3373 max_sector = sector_nr + max_sync;
3374 do {
3375 struct page *page;
3376 int len = PAGE_SIZE;
3377 if (sector_nr + (len>>9) > max_sector)
3378 len = (max_sector - sector_nr) << 9;
3379 if (len == 0)
3380 break;
3381 for (bio= biolist ; bio ; bio=bio->bi_next) {
3382 struct bio *bio2;
3383 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3384 if (bio_add_page(bio, page, len, 0))
3385 continue;
3386
3387 /* stop here */
3388 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3389 for (bio2 = biolist;
3390 bio2 && bio2 != bio;
3391 bio2 = bio2->bi_next) {
3392 /* remove last page from this bio */
3393 bio2->bi_vcnt--;
3394 bio2->bi_size -= len;
3395 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
3396 }
3397 goto bio_full;
3398 }
3399 nr_sectors += len>>9;
3400 sector_nr += len>>9;
3401 } while (biolist->bi_vcnt < RESYNC_PAGES);
3402 bio_full:
3403 r10_bio->sectors = nr_sectors;
3404
3405 while (biolist) {
3406 bio = biolist;
3407 biolist = biolist->bi_next;
3408
3409 bio->bi_next = NULL;
3410 r10_bio = bio->bi_private;
3411 r10_bio->sectors = nr_sectors;
3412
3413 if (bio->bi_end_io == end_sync_read) {
3414 md_sync_acct(bio->bi_bdev, nr_sectors);
3415 set_bit(BIO_UPTODATE, &bio->bi_flags);
3416 generic_make_request(bio);
3417 }
3418 }
3419
3420 if (sectors_skipped)
3421 /* pretend they weren't skipped, it makes
3422 * no important difference in this case
3423 */
3424 md_done_sync(mddev, sectors_skipped, 1);
3425
3426 return sectors_skipped + nr_sectors;
3427 giveup:
3428 /* There is nowhere to write, so all non-sync
3429 * drives must be failed or in resync, all drives
3430 * have a bad block, so try the next chunk...
3431 */
3432 if (sector_nr + max_sync < max_sector)
3433 max_sector = sector_nr + max_sync;
3434
3435 sectors_skipped += (max_sector - sector_nr);
3436 chunks_skipped ++;
3437 sector_nr = max_sector;
3438 goto skipped;
3439 }
3440
3441 static sector_t
3442 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3443 {
3444 sector_t size;
3445 struct r10conf *conf = mddev->private;
3446
3447 if (!raid_disks)
3448 raid_disks = min(conf->geo.raid_disks,
3449 conf->prev.raid_disks);
3450 if (!sectors)
3451 sectors = conf->dev_sectors;
3452
3453 size = sectors >> conf->geo.chunk_shift;
3454 sector_div(size, conf->geo.far_copies);
3455 size = size * raid_disks;
3456 sector_div(size, conf->geo.near_copies);
3457
3458 return size << conf->geo.chunk_shift;
3459 }
3460
3461 static void calc_sectors(struct r10conf *conf, sector_t size)
3462 {
3463 /* Calculate the number of sectors-per-device that will
3464 * actually be used, and set conf->dev_sectors and
3465 * conf->stride
3466 */
3467
3468 size = size >> conf->geo.chunk_shift;
3469 sector_div(size, conf->geo.far_copies);
3470 size = size * conf->geo.raid_disks;
3471 sector_div(size, conf->geo.near_copies);
3472 /* 'size' is now the number of chunks in the array */
3473 /* calculate "used chunks per device" */
3474 size = size * conf->copies;
3475
3476 /* We need to round up when dividing by raid_disks to
3477 * get the stride size.
3478 */
3479 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3480
3481 conf->dev_sectors = size << conf->geo.chunk_shift;
3482
3483 if (conf->geo.far_offset)
3484 conf->geo.stride = 1 << conf->geo.chunk_shift;
3485 else {
3486 sector_div(size, conf->geo.far_copies);
3487 conf->geo.stride = size << conf->geo.chunk_shift;
3488 }
3489 }
3490
3491 enum geo_type {geo_new, geo_old, geo_start};
3492 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3493 {
3494 int nc, fc, fo;
3495 int layout, chunk, disks;
3496 switch (new) {
3497 case geo_old:
3498 layout = mddev->layout;
3499 chunk = mddev->chunk_sectors;
3500 disks = mddev->raid_disks - mddev->delta_disks;
3501 break;
3502 case geo_new:
3503 layout = mddev->new_layout;
3504 chunk = mddev->new_chunk_sectors;
3505 disks = mddev->raid_disks;
3506 break;
3507 default: /* avoid 'may be unused' warnings */
3508 case geo_start: /* new when starting reshape - raid_disks not
3509 * updated yet. */
3510 layout = mddev->new_layout;
3511 chunk = mddev->new_chunk_sectors;
3512 disks = mddev->raid_disks + mddev->delta_disks;
3513 break;
3514 }
3515 if (layout >> 18)
3516 return -1;
3517 if (chunk < (PAGE_SIZE >> 9) ||
3518 !is_power_of_2(chunk))
3519 return -2;
3520 nc = layout & 255;
3521 fc = (layout >> 8) & 255;
3522 fo = layout & (1<<16);
3523 geo->raid_disks = disks;
3524 geo->near_copies = nc;
3525 geo->far_copies = fc;
3526 geo->far_offset = fo;
3527 geo->far_set_size = (layout & (1<<17)) ? disks / fc : disks;
3528 geo->chunk_mask = chunk - 1;
3529 geo->chunk_shift = ffz(~chunk);
3530 return nc*fc;
3531 }
3532
3533 static struct r10conf *setup_conf(struct mddev *mddev)
3534 {
3535 struct r10conf *conf = NULL;
3536 int err = -EINVAL;
3537 struct geom geo;
3538 int copies;
3539
3540 copies = setup_geo(&geo, mddev, geo_new);
3541
3542 if (copies == -2) {
3543 printk(KERN_ERR "md/raid10:%s: chunk size must be "
3544 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3545 mdname(mddev), PAGE_SIZE);
3546 goto out;
3547 }
3548
3549 if (copies < 2 || copies > mddev->raid_disks) {
3550 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3551 mdname(mddev), mddev->new_layout);
3552 goto out;
3553 }
3554
3555 err = -ENOMEM;
3556 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3557 if (!conf)
3558 goto out;
3559
3560 /* FIXME calc properly */
3561 conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3562 max(0,-mddev->delta_disks)),
3563 GFP_KERNEL);
3564 if (!conf->mirrors)
3565 goto out;
3566
3567 conf->tmppage = alloc_page(GFP_KERNEL);
3568 if (!conf->tmppage)
3569 goto out;
3570
3571 conf->geo = geo;
3572 conf->copies = copies;
3573 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3574 r10bio_pool_free, conf);
3575 if (!conf->r10bio_pool)
3576 goto out;
3577
3578 calc_sectors(conf, mddev->dev_sectors);
3579 if (mddev->reshape_position == MaxSector) {
3580 conf->prev = conf->geo;
3581 conf->reshape_progress = MaxSector;
3582 } else {
3583 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3584 err = -EINVAL;
3585 goto out;
3586 }
3587 conf->reshape_progress = mddev->reshape_position;
3588 if (conf->prev.far_offset)
3589 conf->prev.stride = 1 << conf->prev.chunk_shift;
3590 else
3591 /* far_copies must be 1 */
3592 conf->prev.stride = conf->dev_sectors;
3593 }
3594 conf->reshape_safe = conf->reshape_progress;
3595 spin_lock_init(&conf->device_lock);
3596 INIT_LIST_HEAD(&conf->retry_list);
3597
3598 spin_lock_init(&conf->resync_lock);
3599 init_waitqueue_head(&conf->wait_barrier);
3600
3601 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3602 if (!conf->thread)
3603 goto out;
3604
3605 conf->mddev = mddev;
3606 return conf;
3607
3608 out:
3609 if (err == -ENOMEM)
3610 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3611 mdname(mddev));
3612 if (conf) {
3613 if (conf->r10bio_pool)
3614 mempool_destroy(conf->r10bio_pool);
3615 kfree(conf->mirrors);
3616 safe_put_page(conf->tmppage);
3617 kfree(conf);
3618 }
3619 return ERR_PTR(err);
3620 }
3621
3622 static int run(struct mddev *mddev)
3623 {
3624 struct r10conf *conf;
3625 int i, disk_idx, chunk_size;
3626 struct raid10_info *disk;
3627 struct md_rdev *rdev;
3628 sector_t size;
3629 sector_t min_offset_diff = 0;
3630 int first = 1;
3631 bool discard_supported = false;
3632
3633 if (mddev->private == NULL) {
3634 conf = setup_conf(mddev);
3635 if (IS_ERR(conf))
3636 return PTR_ERR(conf);
3637 mddev->private = conf;
3638 }
3639 conf = mddev->private;
3640 if (!conf)
3641 goto out;
3642
3643 mddev->thread = conf->thread;
3644 conf->thread = NULL;
3645
3646 chunk_size = mddev->chunk_sectors << 9;
3647 if (mddev->queue) {
3648 blk_queue_max_discard_sectors(mddev->queue,
3649 mddev->chunk_sectors);
3650 blk_queue_max_write_same_sectors(mddev->queue, 0);
3651 blk_queue_io_min(mddev->queue, chunk_size);
3652 if (conf->geo.raid_disks % conf->geo.near_copies)
3653 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3654 else
3655 blk_queue_io_opt(mddev->queue, chunk_size *
3656 (conf->geo.raid_disks / conf->geo.near_copies));
3657 }
3658
3659 rdev_for_each(rdev, mddev) {
3660 long long diff;
3661 struct request_queue *q;
3662
3663 disk_idx = rdev->raid_disk;
3664 if (disk_idx < 0)
3665 continue;
3666 if (disk_idx >= conf->geo.raid_disks &&
3667 disk_idx >= conf->prev.raid_disks)
3668 continue;
3669 disk = conf->mirrors + disk_idx;
3670
3671 if (test_bit(Replacement, &rdev->flags)) {
3672 if (disk->replacement)
3673 goto out_free_conf;
3674 disk->replacement = rdev;
3675 } else {
3676 if (disk->rdev)
3677 goto out_free_conf;
3678 disk->rdev = rdev;
3679 }
3680 q = bdev_get_queue(rdev->bdev);
3681 if (q->merge_bvec_fn)
3682 mddev->merge_check_needed = 1;
3683 diff = (rdev->new_data_offset - rdev->data_offset);
3684 if (!mddev->reshape_backwards)
3685 diff = -diff;
3686 if (diff < 0)
3687 diff = 0;
3688 if (first || diff < min_offset_diff)
3689 min_offset_diff = diff;
3690
3691 if (mddev->gendisk)
3692 disk_stack_limits(mddev->gendisk, rdev->bdev,
3693 rdev->data_offset << 9);
3694
3695 disk->head_position = 0;
3696
3697 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3698 discard_supported = true;
3699 }
3700
3701 if (mddev->queue) {
3702 if (discard_supported)
3703 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3704 mddev->queue);
3705 else
3706 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3707 mddev->queue);
3708 }
3709 /* need to check that every block has at least one working mirror */
3710 if (!enough(conf, -1)) {
3711 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
3712 mdname(mddev));
3713 goto out_free_conf;
3714 }
3715
3716 if (conf->reshape_progress != MaxSector) {
3717 /* must ensure that shape change is supported */
3718 if (conf->geo.far_copies != 1 &&
3719 conf->geo.far_offset == 0)
3720 goto out_free_conf;
3721 if (conf->prev.far_copies != 1 &&
3722 conf->prev.far_offset == 0)
3723 goto out_free_conf;
3724 }
3725
3726 mddev->degraded = 0;
3727 for (i = 0;
3728 i < conf->geo.raid_disks
3729 || i < conf->prev.raid_disks;
3730 i++) {
3731
3732 disk = conf->mirrors + i;
3733
3734 if (!disk->rdev && disk->replacement) {
3735 /* The replacement is all we have - use it */
3736 disk->rdev = disk->replacement;
3737 disk->replacement = NULL;
3738 clear_bit(Replacement, &disk->rdev->flags);
3739 }
3740
3741 if (!disk->rdev ||
3742 !test_bit(In_sync, &disk->rdev->flags)) {
3743 disk->head_position = 0;
3744 mddev->degraded++;
3745 if (disk->rdev)
3746 conf->fullsync = 1;
3747 }
3748 disk->recovery_disabled = mddev->recovery_disabled - 1;
3749 }
3750
3751 if (mddev->recovery_cp != MaxSector)
3752 printk(KERN_NOTICE "md/raid10:%s: not clean"
3753 " -- starting background reconstruction\n",
3754 mdname(mddev));
3755 printk(KERN_INFO
3756 "md/raid10:%s: active with %d out of %d devices\n",
3757 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3758 conf->geo.raid_disks);
3759 /*
3760 * Ok, everything is just fine now
3761 */
3762 mddev->dev_sectors = conf->dev_sectors;
3763 size = raid10_size(mddev, 0, 0);
3764 md_set_array_sectors(mddev, size);
3765 mddev->resync_max_sectors = size;
3766
3767 if (mddev->queue) {
3768 int stripe = conf->geo.raid_disks *
3769 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3770 mddev->queue->backing_dev_info.congested_fn = raid10_congested;
3771 mddev->queue->backing_dev_info.congested_data = mddev;
3772
3773 /* Calculate max read-ahead size.
3774 * We need to readahead at least twice a whole stripe....
3775 * maybe...
3776 */
3777 stripe /= conf->geo.near_copies;
3778 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3779 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3780 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
3781 }
3782
3783
3784 if (md_integrity_register(mddev))
3785 goto out_free_conf;
3786
3787 if (conf->reshape_progress != MaxSector) {
3788 unsigned long before_length, after_length;
3789
3790 before_length = ((1 << conf->prev.chunk_shift) *
3791 conf->prev.far_copies);
3792 after_length = ((1 << conf->geo.chunk_shift) *
3793 conf->geo.far_copies);
3794
3795 if (max(before_length, after_length) > min_offset_diff) {
3796 /* This cannot work */
3797 printk("md/raid10: offset difference not enough to continue reshape\n");
3798 goto out_free_conf;
3799 }
3800 conf->offset_diff = min_offset_diff;
3801
3802 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3803 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3804 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3805 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3806 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3807 "reshape");
3808 }
3809
3810 return 0;
3811
3812 out_free_conf:
3813 md_unregister_thread(&mddev->thread);
3814 if (conf->r10bio_pool)
3815 mempool_destroy(conf->r10bio_pool);
3816 safe_put_page(conf->tmppage);
3817 kfree(conf->mirrors);
3818 kfree(conf);
3819 mddev->private = NULL;
3820 out:
3821 return -EIO;
3822 }
3823
3824 static int stop(struct mddev *mddev)
3825 {
3826 struct r10conf *conf = mddev->private;
3827
3828 raise_barrier(conf, 0);
3829 lower_barrier(conf);
3830
3831 md_unregister_thread(&mddev->thread);
3832 if (mddev->queue)
3833 /* the unplug fn references 'conf'*/
3834 blk_sync_queue(mddev->queue);
3835
3836 if (conf->r10bio_pool)
3837 mempool_destroy(conf->r10bio_pool);
3838 safe_put_page(conf->tmppage);
3839 kfree(conf->mirrors);
3840 kfree(conf);
3841 mddev->private = NULL;
3842 return 0;
3843 }
3844
3845 static void raid10_quiesce(struct mddev *mddev, int state)
3846 {
3847 struct r10conf *conf = mddev->private;
3848
3849 switch(state) {
3850 case 1:
3851 raise_barrier(conf, 0);
3852 break;
3853 case 0:
3854 lower_barrier(conf);
3855 break;
3856 }
3857 }
3858
3859 static int raid10_resize(struct mddev *mddev, sector_t sectors)
3860 {
3861 /* Resize of 'far' arrays is not supported.
3862 * For 'near' and 'offset' arrays we can set the
3863 * number of sectors used to be an appropriate multiple
3864 * of the chunk size.
3865 * For 'offset', this is far_copies*chunksize.
3866 * For 'near' the multiplier is the LCM of
3867 * near_copies and raid_disks.
3868 * So if far_copies > 1 && !far_offset, fail.
3869 * Else find LCM(raid_disks, near_copy)*far_copies and
3870 * multiply by chunk_size. Then round to this number.
3871 * This is mostly done by raid10_size()
3872 */
3873 struct r10conf *conf = mddev->private;
3874 sector_t oldsize, size;
3875
3876 if (mddev->reshape_position != MaxSector)
3877 return -EBUSY;
3878
3879 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3880 return -EINVAL;
3881
3882 oldsize = raid10_size(mddev, 0, 0);
3883 size = raid10_size(mddev, sectors, 0);
3884 if (mddev->external_size &&
3885 mddev->array_sectors > size)
3886 return -EINVAL;
3887 if (mddev->bitmap) {
3888 int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3889 if (ret)
3890 return ret;
3891 }
3892 md_set_array_sectors(mddev, size);
3893 set_capacity(mddev->gendisk, mddev->array_sectors);
3894 revalidate_disk(mddev->gendisk);
3895 if (sectors > mddev->dev_sectors &&
3896 mddev->recovery_cp > oldsize) {
3897 mddev->recovery_cp = oldsize;
3898 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3899 }
3900 calc_sectors(conf, sectors);
3901 mddev->dev_sectors = conf->dev_sectors;
3902 mddev->resync_max_sectors = size;
3903 return 0;
3904 }
3905
3906 static void *raid10_takeover_raid0(struct mddev *mddev)
3907 {
3908 struct md_rdev *rdev;
3909 struct r10conf *conf;
3910
3911 if (mddev->degraded > 0) {
3912 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3913 mdname(mddev));
3914 return ERR_PTR(-EINVAL);
3915 }
3916
3917 /* Set new parameters */
3918 mddev->new_level = 10;
3919 /* new layout: far_copies = 1, near_copies = 2 */
3920 mddev->new_layout = (1<<8) + 2;
3921 mddev->new_chunk_sectors = mddev->chunk_sectors;
3922 mddev->delta_disks = mddev->raid_disks;
3923 mddev->raid_disks *= 2;
3924 /* make sure it will be not marked as dirty */
3925 mddev->recovery_cp = MaxSector;
3926
3927 conf = setup_conf(mddev);
3928 if (!IS_ERR(conf)) {
3929 rdev_for_each(rdev, mddev)
3930 if (rdev->raid_disk >= 0)
3931 rdev->new_raid_disk = rdev->raid_disk * 2;
3932 conf->barrier = 1;
3933 }
3934
3935 return conf;
3936 }
3937
3938 static void *raid10_takeover(struct mddev *mddev)
3939 {
3940 struct r0conf *raid0_conf;
3941
3942 /* raid10 can take over:
3943 * raid0 - providing it has only two drives
3944 */
3945 if (mddev->level == 0) {
3946 /* for raid0 takeover only one zone is supported */
3947 raid0_conf = mddev->private;
3948 if (raid0_conf->nr_strip_zones > 1) {
3949 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3950 " with more than one zone.\n",
3951 mdname(mddev));
3952 return ERR_PTR(-EINVAL);
3953 }
3954 return raid10_takeover_raid0(mddev);
3955 }
3956 return ERR_PTR(-EINVAL);
3957 }
3958
3959 static int raid10_check_reshape(struct mddev *mddev)
3960 {
3961 /* Called when there is a request to change
3962 * - layout (to ->new_layout)
3963 * - chunk size (to ->new_chunk_sectors)
3964 * - raid_disks (by delta_disks)
3965 * or when trying to restart a reshape that was ongoing.
3966 *
3967 * We need to validate the request and possibly allocate
3968 * space if that might be an issue later.
3969 *
3970 * Currently we reject any reshape of a 'far' mode array,
3971 * allow chunk size to change if new is generally acceptable,
3972 * allow raid_disks to increase, and allow
3973 * a switch between 'near' mode and 'offset' mode.
3974 */
3975 struct r10conf *conf = mddev->private;
3976 struct geom geo;
3977
3978 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
3979 return -EINVAL;
3980
3981 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
3982 /* mustn't change number of copies */
3983 return -EINVAL;
3984 if (geo.far_copies > 1 && !geo.far_offset)
3985 /* Cannot switch to 'far' mode */
3986 return -EINVAL;
3987
3988 if (mddev->array_sectors & geo.chunk_mask)
3989 /* not factor of array size */
3990 return -EINVAL;
3991
3992 if (!enough(conf, -1))
3993 return -EINVAL;
3994
3995 kfree(conf->mirrors_new);
3996 conf->mirrors_new = NULL;
3997 if (mddev->delta_disks > 0) {
3998 /* allocate new 'mirrors' list */
3999 conf->mirrors_new = kzalloc(
4000 sizeof(struct raid10_info)
4001 *(mddev->raid_disks +
4002 mddev->delta_disks),
4003 GFP_KERNEL);
4004 if (!conf->mirrors_new)
4005 return -ENOMEM;
4006 }
4007 return 0;
4008 }
4009
4010 /*
4011 * Need to check if array has failed when deciding whether to:
4012 * - start an array
4013 * - remove non-faulty devices
4014 * - add a spare
4015 * - allow a reshape
4016 * This determination is simple when no reshape is happening.
4017 * However if there is a reshape, we need to carefully check
4018 * both the before and after sections.
4019 * This is because some failed devices may only affect one
4020 * of the two sections, and some non-in_sync devices may
4021 * be insync in the section most affected by failed devices.
4022 */
4023 static int calc_degraded(struct r10conf *conf)
4024 {
4025 int degraded, degraded2;
4026 int i;
4027
4028 rcu_read_lock();
4029 degraded = 0;
4030 /* 'prev' section first */
4031 for (i = 0; i < conf->prev.raid_disks; i++) {
4032 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4033 if (!rdev || test_bit(Faulty, &rdev->flags))
4034 degraded++;
4035 else if (!test_bit(In_sync, &rdev->flags))
4036 /* When we can reduce the number of devices in
4037 * an array, this might not contribute to
4038 * 'degraded'. It does now.
4039 */
4040 degraded++;
4041 }
4042 rcu_read_unlock();
4043 if (conf->geo.raid_disks == conf->prev.raid_disks)
4044 return degraded;
4045 rcu_read_lock();
4046 degraded2 = 0;
4047 for (i = 0; i < conf->geo.raid_disks; i++) {
4048 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4049 if (!rdev || test_bit(Faulty, &rdev->flags))
4050 degraded2++;
4051 else if (!test_bit(In_sync, &rdev->flags)) {
4052 /* If reshape is increasing the number of devices,
4053 * this section has already been recovered, so
4054 * it doesn't contribute to degraded.
4055 * else it does.
4056 */
4057 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4058 degraded2++;
4059 }
4060 }
4061 rcu_read_unlock();
4062 if (degraded2 > degraded)
4063 return degraded2;
4064 return degraded;
4065 }
4066
4067 static int raid10_start_reshape(struct mddev *mddev)
4068 {
4069 /* A 'reshape' has been requested. This commits
4070 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4071 * This also checks if there are enough spares and adds them
4072 * to the array.
4073 * We currently require enough spares to make the final
4074 * array non-degraded. We also require that the difference
4075 * between old and new data_offset - on each device - is
4076 * enough that we never risk over-writing.
4077 */
4078
4079 unsigned long before_length, after_length;
4080 sector_t min_offset_diff = 0;
4081 int first = 1;
4082 struct geom new;
4083 struct r10conf *conf = mddev->private;
4084 struct md_rdev *rdev;
4085 int spares = 0;
4086 int ret;
4087
4088 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4089 return -EBUSY;
4090
4091 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4092 return -EINVAL;
4093
4094 before_length = ((1 << conf->prev.chunk_shift) *
4095 conf->prev.far_copies);
4096 after_length = ((1 << conf->geo.chunk_shift) *
4097 conf->geo.far_copies);
4098
4099 rdev_for_each(rdev, mddev) {
4100 if (!test_bit(In_sync, &rdev->flags)
4101 && !test_bit(Faulty, &rdev->flags))
4102 spares++;
4103 if (rdev->raid_disk >= 0) {
4104 long long diff = (rdev->new_data_offset
4105 - rdev->data_offset);
4106 if (!mddev->reshape_backwards)
4107 diff = -diff;
4108 if (diff < 0)
4109 diff = 0;
4110 if (first || diff < min_offset_diff)
4111 min_offset_diff = diff;
4112 }
4113 }
4114
4115 if (max(before_length, after_length) > min_offset_diff)
4116 return -EINVAL;
4117
4118 if (spares < mddev->delta_disks)
4119 return -EINVAL;
4120
4121 conf->offset_diff = min_offset_diff;
4122 spin_lock_irq(&conf->device_lock);
4123 if (conf->mirrors_new) {
4124 memcpy(conf->mirrors_new, conf->mirrors,
4125 sizeof(struct raid10_info)*conf->prev.raid_disks);
4126 smp_mb();
4127 kfree(conf->mirrors_old); /* FIXME and elsewhere */
4128 conf->mirrors_old = conf->mirrors;
4129 conf->mirrors = conf->mirrors_new;
4130 conf->mirrors_new = NULL;
4131 }
4132 setup_geo(&conf->geo, mddev, geo_start);
4133 smp_mb();
4134 if (mddev->reshape_backwards) {
4135 sector_t size = raid10_size(mddev, 0, 0);
4136 if (size < mddev->array_sectors) {
4137 spin_unlock_irq(&conf->device_lock);
4138 printk(KERN_ERR "md/raid10:%s: array size must be reduce before number of disks\n",
4139 mdname(mddev));
4140 return -EINVAL;
4141 }
4142 mddev->resync_max_sectors = size;
4143 conf->reshape_progress = size;
4144 } else
4145 conf->reshape_progress = 0;
4146 conf->reshape_safe = conf->reshape_progress;
4147 spin_unlock_irq(&conf->device_lock);
4148
4149 if (mddev->delta_disks && mddev->bitmap) {
4150 ret = bitmap_resize(mddev->bitmap,
4151 raid10_size(mddev, 0,
4152 conf->geo.raid_disks),
4153 0, 0);
4154 if (ret)
4155 goto abort;
4156 }
4157 if (mddev->delta_disks > 0) {
4158 rdev_for_each(rdev, mddev)
4159 if (rdev->raid_disk < 0 &&
4160 !test_bit(Faulty, &rdev->flags)) {
4161 if (raid10_add_disk(mddev, rdev) == 0) {
4162 if (rdev->raid_disk >=
4163 conf->prev.raid_disks)
4164 set_bit(In_sync, &rdev->flags);
4165 else
4166 rdev->recovery_offset = 0;
4167
4168 if (sysfs_link_rdev(mddev, rdev))
4169 /* Failure here is OK */;
4170 }
4171 } else if (rdev->raid_disk >= conf->prev.raid_disks
4172 && !test_bit(Faulty, &rdev->flags)) {
4173 /* This is a spare that was manually added */
4174 set_bit(In_sync, &rdev->flags);
4175 }
4176 }
4177 /* When a reshape changes the number of devices,
4178 * ->degraded is measured against the larger of the
4179 * pre and post numbers.
4180 */
4181 spin_lock_irq(&conf->device_lock);
4182 mddev->degraded = calc_degraded(conf);
4183 spin_unlock_irq(&conf->device_lock);
4184 mddev->raid_disks = conf->geo.raid_disks;
4185 mddev->reshape_position = conf->reshape_progress;
4186 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4187
4188 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4189 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4190 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4191 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4192
4193 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4194 "reshape");
4195 if (!mddev->sync_thread) {
4196 ret = -EAGAIN;
4197 goto abort;
4198 }
4199 conf->reshape_checkpoint = jiffies;
4200 md_wakeup_thread(mddev->sync_thread);
4201 md_new_event(mddev);
4202 return 0;
4203
4204 abort:
4205 mddev->recovery = 0;
4206 spin_lock_irq(&conf->device_lock);
4207 conf->geo = conf->prev;
4208 mddev->raid_disks = conf->geo.raid_disks;
4209 rdev_for_each(rdev, mddev)
4210 rdev->new_data_offset = rdev->data_offset;
4211 smp_wmb();
4212 conf->reshape_progress = MaxSector;
4213 conf->reshape_safe = MaxSector;
4214 mddev->reshape_position = MaxSector;
4215 spin_unlock_irq(&conf->device_lock);
4216 return ret;
4217 }
4218
4219 /* Calculate the last device-address that could contain
4220 * any block from the chunk that includes the array-address 's'
4221 * and report the next address.
4222 * i.e. the address returned will be chunk-aligned and after
4223 * any data that is in the chunk containing 's'.
4224 */
4225 static sector_t last_dev_address(sector_t s, struct geom *geo)
4226 {
4227 s = (s | geo->chunk_mask) + 1;
4228 s >>= geo->chunk_shift;
4229 s *= geo->near_copies;
4230 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4231 s *= geo->far_copies;
4232 s <<= geo->chunk_shift;
4233 return s;
4234 }
4235
4236 /* Calculate the first device-address that could contain
4237 * any block from the chunk that includes the array-address 's'.
4238 * This too will be the start of a chunk
4239 */
4240 static sector_t first_dev_address(sector_t s, struct geom *geo)
4241 {
4242 s >>= geo->chunk_shift;
4243 s *= geo->near_copies;
4244 sector_div(s, geo->raid_disks);
4245 s *= geo->far_copies;
4246 s <<= geo->chunk_shift;
4247 return s;
4248 }
4249
4250 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4251 int *skipped)
4252 {
4253 /* We simply copy at most one chunk (smallest of old and new)
4254 * at a time, possibly less if that exceeds RESYNC_PAGES,
4255 * or we hit a bad block or something.
4256 * This might mean we pause for normal IO in the middle of
4257 * a chunk, but that is not a problem was mddev->reshape_position
4258 * can record any location.
4259 *
4260 * If we will want to write to a location that isn't
4261 * yet recorded as 'safe' (i.e. in metadata on disk) then
4262 * we need to flush all reshape requests and update the metadata.
4263 *
4264 * When reshaping forwards (e.g. to more devices), we interpret
4265 * 'safe' as the earliest block which might not have been copied
4266 * down yet. We divide this by previous stripe size and multiply
4267 * by previous stripe length to get lowest device offset that we
4268 * cannot write to yet.
4269 * We interpret 'sector_nr' as an address that we want to write to.
4270 * From this we use last_device_address() to find where we might
4271 * write to, and first_device_address on the 'safe' position.
4272 * If this 'next' write position is after the 'safe' position,
4273 * we must update the metadata to increase the 'safe' position.
4274 *
4275 * When reshaping backwards, we round in the opposite direction
4276 * and perform the reverse test: next write position must not be
4277 * less than current safe position.
4278 *
4279 * In all this the minimum difference in data offsets
4280 * (conf->offset_diff - always positive) allows a bit of slack,
4281 * so next can be after 'safe', but not by more than offset_disk
4282 *
4283 * We need to prepare all the bios here before we start any IO
4284 * to ensure the size we choose is acceptable to all devices.
4285 * The means one for each copy for write-out and an extra one for
4286 * read-in.
4287 * We store the read-in bio in ->master_bio and the others in
4288 * ->devs[x].bio and ->devs[x].repl_bio.
4289 */
4290 struct r10conf *conf = mddev->private;
4291 struct r10bio *r10_bio;
4292 sector_t next, safe, last;
4293 int max_sectors;
4294 int nr_sectors;
4295 int s;
4296 struct md_rdev *rdev;
4297 int need_flush = 0;
4298 struct bio *blist;
4299 struct bio *bio, *read_bio;
4300 int sectors_done = 0;
4301
4302 if (sector_nr == 0) {
4303 /* If restarting in the middle, skip the initial sectors */
4304 if (mddev->reshape_backwards &&
4305 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4306 sector_nr = (raid10_size(mddev, 0, 0)
4307 - conf->reshape_progress);
4308 } else if (!mddev->reshape_backwards &&
4309 conf->reshape_progress > 0)
4310 sector_nr = conf->reshape_progress;
4311 if (sector_nr) {
4312 mddev->curr_resync_completed = sector_nr;
4313 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4314 *skipped = 1;
4315 return sector_nr;
4316 }
4317 }
4318
4319 /* We don't use sector_nr to track where we are up to
4320 * as that doesn't work well for ->reshape_backwards.
4321 * So just use ->reshape_progress.
4322 */
4323 if (mddev->reshape_backwards) {
4324 /* 'next' is the earliest device address that we might
4325 * write to for this chunk in the new layout
4326 */
4327 next = first_dev_address(conf->reshape_progress - 1,
4328 &conf->geo);
4329
4330 /* 'safe' is the last device address that we might read from
4331 * in the old layout after a restart
4332 */
4333 safe = last_dev_address(conf->reshape_safe - 1,
4334 &conf->prev);
4335
4336 if (next + conf->offset_diff < safe)
4337 need_flush = 1;
4338
4339 last = conf->reshape_progress - 1;
4340 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4341 & conf->prev.chunk_mask);
4342 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4343 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4344 } else {
4345 /* 'next' is after the last device address that we
4346 * might write to for this chunk in the new layout
4347 */
4348 next = last_dev_address(conf->reshape_progress, &conf->geo);
4349
4350 /* 'safe' is the earliest device address that we might
4351 * read from in the old layout after a restart
4352 */
4353 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4354
4355 /* Need to update metadata if 'next' might be beyond 'safe'
4356 * as that would possibly corrupt data
4357 */
4358 if (next > safe + conf->offset_diff)
4359 need_flush = 1;
4360
4361 sector_nr = conf->reshape_progress;
4362 last = sector_nr | (conf->geo.chunk_mask
4363 & conf->prev.chunk_mask);
4364
4365 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4366 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4367 }
4368
4369 if (need_flush ||
4370 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4371 /* Need to update reshape_position in metadata */
4372 wait_barrier(conf);
4373 mddev->reshape_position = conf->reshape_progress;
4374 if (mddev->reshape_backwards)
4375 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4376 - conf->reshape_progress;
4377 else
4378 mddev->curr_resync_completed = conf->reshape_progress;
4379 conf->reshape_checkpoint = jiffies;
4380 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4381 md_wakeup_thread(mddev->thread);
4382 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4383 kthread_should_stop());
4384 conf->reshape_safe = mddev->reshape_position;
4385 allow_barrier(conf);
4386 }
4387
4388 read_more:
4389 /* Now schedule reads for blocks from sector_nr to last */
4390 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4391 raise_barrier(conf, sectors_done != 0);
4392 atomic_set(&r10_bio->remaining, 0);
4393 r10_bio->mddev = mddev;
4394 r10_bio->sector = sector_nr;
4395 set_bit(R10BIO_IsReshape, &r10_bio->state);
4396 r10_bio->sectors = last - sector_nr + 1;
4397 rdev = read_balance(conf, r10_bio, &max_sectors);
4398 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4399
4400 if (!rdev) {
4401 /* Cannot read from here, so need to record bad blocks
4402 * on all the target devices.
4403 */
4404 // FIXME
4405 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4406 return sectors_done;
4407 }
4408
4409 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4410
4411 read_bio->bi_bdev = rdev->bdev;
4412 read_bio->bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4413 + rdev->data_offset);
4414 read_bio->bi_private = r10_bio;
4415 read_bio->bi_end_io = end_sync_read;
4416 read_bio->bi_rw = READ;
4417 read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4418 read_bio->bi_flags |= 1 << BIO_UPTODATE;
4419 read_bio->bi_vcnt = 0;
4420 read_bio->bi_size = 0;
4421 r10_bio->master_bio = read_bio;
4422 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4423
4424 /* Now find the locations in the new layout */
4425 __raid10_find_phys(&conf->geo, r10_bio);
4426
4427 blist = read_bio;
4428 read_bio->bi_next = NULL;
4429
4430 for (s = 0; s < conf->copies*2; s++) {
4431 struct bio *b;
4432 int d = r10_bio->devs[s/2].devnum;
4433 struct md_rdev *rdev2;
4434 if (s&1) {
4435 rdev2 = conf->mirrors[d].replacement;
4436 b = r10_bio->devs[s/2].repl_bio;
4437 } else {
4438 rdev2 = conf->mirrors[d].rdev;
4439 b = r10_bio->devs[s/2].bio;
4440 }
4441 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4442 continue;
4443
4444 bio_reset(b);
4445 b->bi_bdev = rdev2->bdev;
4446 b->bi_sector = r10_bio->devs[s/2].addr + rdev2->new_data_offset;
4447 b->bi_private = r10_bio;
4448 b->bi_end_io = end_reshape_write;
4449 b->bi_rw = WRITE;
4450 b->bi_next = blist;
4451 blist = b;
4452 }
4453
4454 /* Now add as many pages as possible to all of these bios. */
4455
4456 nr_sectors = 0;
4457 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4458 struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4459 int len = (max_sectors - s) << 9;
4460 if (len > PAGE_SIZE)
4461 len = PAGE_SIZE;
4462 for (bio = blist; bio ; bio = bio->bi_next) {
4463 struct bio *bio2;
4464 if (bio_add_page(bio, page, len, 0))
4465 continue;
4466
4467 /* Didn't fit, must stop */
4468 for (bio2 = blist;
4469 bio2 && bio2 != bio;
4470 bio2 = bio2->bi_next) {
4471 /* Remove last page from this bio */
4472 bio2->bi_vcnt--;
4473 bio2->bi_size -= len;
4474 bio2->bi_flags &= ~(1<<BIO_SEG_VALID);
4475 }
4476 goto bio_full;
4477 }
4478 sector_nr += len >> 9;
4479 nr_sectors += len >> 9;
4480 }
4481 bio_full:
4482 r10_bio->sectors = nr_sectors;
4483
4484 /* Now submit the read */
4485 md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4486 atomic_inc(&r10_bio->remaining);
4487 read_bio->bi_next = NULL;
4488 generic_make_request(read_bio);
4489 sector_nr += nr_sectors;
4490 sectors_done += nr_sectors;
4491 if (sector_nr <= last)
4492 goto read_more;
4493
4494 /* Now that we have done the whole section we can
4495 * update reshape_progress
4496 */
4497 if (mddev->reshape_backwards)
4498 conf->reshape_progress -= sectors_done;
4499 else
4500 conf->reshape_progress += sectors_done;
4501
4502 return sectors_done;
4503 }
4504
4505 static void end_reshape_request(struct r10bio *r10_bio);
4506 static int handle_reshape_read_error(struct mddev *mddev,
4507 struct r10bio *r10_bio);
4508 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4509 {
4510 /* Reshape read completed. Hopefully we have a block
4511 * to write out.
4512 * If we got a read error then we do sync 1-page reads from
4513 * elsewhere until we find the data - or give up.
4514 */
4515 struct r10conf *conf = mddev->private;
4516 int s;
4517
4518 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4519 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4520 /* Reshape has been aborted */
4521 md_done_sync(mddev, r10_bio->sectors, 0);
4522 return;
4523 }
4524
4525 /* We definitely have the data in the pages, schedule the
4526 * writes.
4527 */
4528 atomic_set(&r10_bio->remaining, 1);
4529 for (s = 0; s < conf->copies*2; s++) {
4530 struct bio *b;
4531 int d = r10_bio->devs[s/2].devnum;
4532 struct md_rdev *rdev;
4533 if (s&1) {
4534 rdev = conf->mirrors[d].replacement;
4535 b = r10_bio->devs[s/2].repl_bio;
4536 } else {
4537 rdev = conf->mirrors[d].rdev;
4538 b = r10_bio->devs[s/2].bio;
4539 }
4540 if (!rdev || test_bit(Faulty, &rdev->flags))
4541 continue;
4542 atomic_inc(&rdev->nr_pending);
4543 md_sync_acct(b->bi_bdev, r10_bio->sectors);
4544 atomic_inc(&r10_bio->remaining);
4545 b->bi_next = NULL;
4546 generic_make_request(b);
4547 }
4548 end_reshape_request(r10_bio);
4549 }
4550
4551 static void end_reshape(struct r10conf *conf)
4552 {
4553 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4554 return;
4555
4556 spin_lock_irq(&conf->device_lock);
4557 conf->prev = conf->geo;
4558 md_finish_reshape(conf->mddev);
4559 smp_wmb();
4560 conf->reshape_progress = MaxSector;
4561 conf->reshape_safe = MaxSector;
4562 spin_unlock_irq(&conf->device_lock);
4563
4564 /* read-ahead size must cover two whole stripes, which is
4565 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4566 */
4567 if (conf->mddev->queue) {
4568 int stripe = conf->geo.raid_disks *
4569 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4570 stripe /= conf->geo.near_copies;
4571 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4572 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4573 }
4574 conf->fullsync = 0;
4575 }
4576
4577
4578 static int handle_reshape_read_error(struct mddev *mddev,
4579 struct r10bio *r10_bio)
4580 {
4581 /* Use sync reads to get the blocks from somewhere else */
4582 int sectors = r10_bio->sectors;
4583 struct r10conf *conf = mddev->private;
4584 struct {
4585 struct r10bio r10_bio;
4586 struct r10dev devs[conf->copies];
4587 } on_stack;
4588 struct r10bio *r10b = &on_stack.r10_bio;
4589 int slot = 0;
4590 int idx = 0;
4591 struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
4592
4593 r10b->sector = r10_bio->sector;
4594 __raid10_find_phys(&conf->prev, r10b);
4595
4596 while (sectors) {
4597 int s = sectors;
4598 int success = 0;
4599 int first_slot = slot;
4600
4601 if (s > (PAGE_SIZE >> 9))
4602 s = PAGE_SIZE >> 9;
4603
4604 while (!success) {
4605 int d = r10b->devs[slot].devnum;
4606 struct md_rdev *rdev = conf->mirrors[d].rdev;
4607 sector_t addr;
4608 if (rdev == NULL ||
4609 test_bit(Faulty, &rdev->flags) ||
4610 !test_bit(In_sync, &rdev->flags))
4611 goto failed;
4612
4613 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4614 success = sync_page_io(rdev,
4615 addr,
4616 s << 9,
4617 bvec[idx].bv_page,
4618 READ, false);
4619 if (success)
4620 break;
4621 failed:
4622 slot++;
4623 if (slot >= conf->copies)
4624 slot = 0;
4625 if (slot == first_slot)
4626 break;
4627 }
4628 if (!success) {
4629 /* couldn't read this block, must give up */
4630 set_bit(MD_RECOVERY_INTR,
4631 &mddev->recovery);
4632 return -EIO;
4633 }
4634 sectors -= s;
4635 idx++;
4636 }
4637 return 0;
4638 }
4639
4640 static void end_reshape_write(struct bio *bio, int error)
4641 {
4642 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
4643 struct r10bio *r10_bio = bio->bi_private;
4644 struct mddev *mddev = r10_bio->mddev;
4645 struct r10conf *conf = mddev->private;
4646 int d;
4647 int slot;
4648 int repl;
4649 struct md_rdev *rdev = NULL;
4650
4651 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4652 if (repl)
4653 rdev = conf->mirrors[d].replacement;
4654 if (!rdev) {
4655 smp_mb();
4656 rdev = conf->mirrors[d].rdev;
4657 }
4658
4659 if (!uptodate) {
4660 /* FIXME should record badblock */
4661 md_error(mddev, rdev);
4662 }
4663
4664 rdev_dec_pending(rdev, mddev);
4665 end_reshape_request(r10_bio);
4666 }
4667
4668 static void end_reshape_request(struct r10bio *r10_bio)
4669 {
4670 if (!atomic_dec_and_test(&r10_bio->remaining))
4671 return;
4672 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4673 bio_put(r10_bio->master_bio);
4674 put_buf(r10_bio);
4675 }
4676
4677 static void raid10_finish_reshape(struct mddev *mddev)
4678 {
4679 struct r10conf *conf = mddev->private;
4680
4681 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4682 return;
4683
4684 if (mddev->delta_disks > 0) {
4685 sector_t size = raid10_size(mddev, 0, 0);
4686 md_set_array_sectors(mddev, size);
4687 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4688 mddev->recovery_cp = mddev->resync_max_sectors;
4689 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4690 }
4691 mddev->resync_max_sectors = size;
4692 set_capacity(mddev->gendisk, mddev->array_sectors);
4693 revalidate_disk(mddev->gendisk);
4694 } else {
4695 int d;
4696 for (d = conf->geo.raid_disks ;
4697 d < conf->geo.raid_disks - mddev->delta_disks;
4698 d++) {
4699 struct md_rdev *rdev = conf->mirrors[d].rdev;
4700 if (rdev)
4701 clear_bit(In_sync, &rdev->flags);
4702 rdev = conf->mirrors[d].replacement;
4703 if (rdev)
4704 clear_bit(In_sync, &rdev->flags);
4705 }
4706 }
4707 mddev->layout = mddev->new_layout;
4708 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4709 mddev->reshape_position = MaxSector;
4710 mddev->delta_disks = 0;
4711 mddev->reshape_backwards = 0;
4712 }
4713
4714 static struct md_personality raid10_personality =
4715 {
4716 .name = "raid10",
4717 .level = 10,
4718 .owner = THIS_MODULE,
4719 .make_request = make_request,
4720 .run = run,
4721 .stop = stop,
4722 .status = status,
4723 .error_handler = error,
4724 .hot_add_disk = raid10_add_disk,
4725 .hot_remove_disk= raid10_remove_disk,
4726 .spare_active = raid10_spare_active,
4727 .sync_request = sync_request,
4728 .quiesce = raid10_quiesce,
4729 .size = raid10_size,
4730 .resize = raid10_resize,
4731 .takeover = raid10_takeover,
4732 .check_reshape = raid10_check_reshape,
4733 .start_reshape = raid10_start_reshape,
4734 .finish_reshape = raid10_finish_reshape,
4735 };
4736
4737 static int __init raid_init(void)
4738 {
4739 return register_md_personality(&raid10_personality);
4740 }
4741
4742 static void raid_exit(void)
4743 {
4744 unregister_md_personality(&raid10_personality);
4745 }
4746
4747 module_init(raid_init);
4748 module_exit(raid_exit);
4749 MODULE_LICENSE("GPL");
4750 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4751 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4752 MODULE_ALIAS("md-raid10");
4753 MODULE_ALIAS("md-level-10");
4754
4755 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);
kernel source for telekom puls (alcatel/mediatek)