Loading...
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 */
96static int max_queued_requests = 1024;
97
98static void allow_barrier(struct r10conf *conf);
99static void lower_barrier(struct r10conf *conf);
100static int _enough(struct r10conf *conf, int previous, int ignore);
101static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
102 int *skipped);
103static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
104static void end_reshape_write(struct bio *bio, int error);
105static void end_reshape(struct r10conf *conf);
106
107static 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
117static 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 */
137static 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
199out_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;
206out_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
217static 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
240static 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
256static 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
264static 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
273static 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 */
295static 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 */
325static 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 */
336static 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
361static 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 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
396 rdev->raid_disk))
397 uptodate = 1;
398 }
399 if (uptodate) {
400 raid_end_bio_io(r10_bio);
401 rdev_dec_pending(rdev, conf->mddev);
402 } else {
403 /*
404 * oops, read error - keep the refcount on the rdev
405 */
406 char b[BDEVNAME_SIZE];
407 printk_ratelimited(KERN_ERR
408 "md/raid10:%s: %s: rescheduling sector %llu\n",
409 mdname(conf->mddev),
410 bdevname(rdev->bdev, b),
411 (unsigned long long)r10_bio->sector);
412 set_bit(R10BIO_ReadError, &r10_bio->state);
413 reschedule_retry(r10_bio);
414 }
415}
416
417static void close_write(struct r10bio *r10_bio)
418{
419 /* clear the bitmap if all writes complete successfully */
420 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
421 r10_bio->sectors,
422 !test_bit(R10BIO_Degraded, &r10_bio->state),
423 0);
424 md_write_end(r10_bio->mddev);
425}
426
427static void one_write_done(struct r10bio *r10_bio)
428{
429 if (atomic_dec_and_test(&r10_bio->remaining)) {
430 if (test_bit(R10BIO_WriteError, &r10_bio->state))
431 reschedule_retry(r10_bio);
432 else {
433 close_write(r10_bio);
434 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
435 reschedule_retry(r10_bio);
436 else
437 raid_end_bio_io(r10_bio);
438 }
439 }
440}
441
442static void raid10_end_write_request(struct bio *bio, int error)
443{
444 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
445 struct r10bio *r10_bio = bio->bi_private;
446 int dev;
447 int dec_rdev = 1;
448 struct r10conf *conf = r10_bio->mddev->private;
449 int slot, repl;
450 struct md_rdev *rdev = NULL;
451
452 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
453
454 if (repl)
455 rdev = conf->mirrors[dev].replacement;
456 if (!rdev) {
457 smp_rmb();
458 repl = 0;
459 rdev = conf->mirrors[dev].rdev;
460 }
461 /*
462 * this branch is our 'one mirror IO has finished' event handler:
463 */
464 if (!uptodate) {
465 if (repl)
466 /* Never record new bad blocks to replacement,
467 * just fail it.
468 */
469 md_error(rdev->mddev, rdev);
470 else {
471 set_bit(WriteErrorSeen, &rdev->flags);
472 if (!test_and_set_bit(WantReplacement, &rdev->flags))
473 set_bit(MD_RECOVERY_NEEDED,
474 &rdev->mddev->recovery);
475 set_bit(R10BIO_WriteError, &r10_bio->state);
476 dec_rdev = 0;
477 }
478 } else {
479 /*
480 * Set R10BIO_Uptodate in our master bio, so that
481 * we will return a good error code for to the higher
482 * levels even if IO on some other mirrored buffer fails.
483 *
484 * The 'master' represents the composite IO operation to
485 * user-side. So if something waits for IO, then it will
486 * wait for the 'master' bio.
487 */
488 sector_t first_bad;
489 int bad_sectors;
490
491 /*
492 * Do not set R10BIO_Uptodate if the current device is
493 * rebuilding or Faulty. This is because we cannot use
494 * such device for properly reading the data back (we could
495 * potentially use it, if the current write would have felt
496 * before rdev->recovery_offset, but for simplicity we don't
497 * check this here.
498 */
499 if (test_bit(In_sync, &rdev->flags) &&
500 !test_bit(Faulty, &rdev->flags))
501 set_bit(R10BIO_Uptodate, &r10_bio->state);
502
503 /* Maybe we can clear some bad blocks. */
504 if (is_badblock(rdev,
505 r10_bio->devs[slot].addr,
506 r10_bio->sectors,
507 &first_bad, &bad_sectors)) {
508 bio_put(bio);
509 if (repl)
510 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
511 else
512 r10_bio->devs[slot].bio = IO_MADE_GOOD;
513 dec_rdev = 0;
514 set_bit(R10BIO_MadeGood, &r10_bio->state);
515 }
516 }
517
518 /*
519 *
520 * Let's see if all mirrored write operations have finished
521 * already.
522 */
523 one_write_done(r10_bio);
524 if (dec_rdev)
525 rdev_dec_pending(rdev, conf->mddev);
526}
527
528/*
529 * RAID10 layout manager
530 * As well as the chunksize and raid_disks count, there are two
531 * parameters: near_copies and far_copies.
532 * near_copies * far_copies must be <= raid_disks.
533 * Normally one of these will be 1.
534 * If both are 1, we get raid0.
535 * If near_copies == raid_disks, we get raid1.
536 *
537 * Chunks are laid out in raid0 style with near_copies copies of the
538 * first chunk, followed by near_copies copies of the next chunk and
539 * so on.
540 * If far_copies > 1, then after 1/far_copies of the array has been assigned
541 * as described above, we start again with a device offset of near_copies.
542 * So we effectively have another copy of the whole array further down all
543 * the drives, but with blocks on different drives.
544 * With this layout, and block is never stored twice on the one device.
545 *
546 * raid10_find_phys finds the sector offset of a given virtual sector
547 * on each device that it is on.
548 *
549 * raid10_find_virt does the reverse mapping, from a device and a
550 * sector offset to a virtual address
551 */
552
553static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
554{
555 int n,f;
556 sector_t sector;
557 sector_t chunk;
558 sector_t stripe;
559 int dev;
560 int slot = 0;
561 int last_far_set_start, last_far_set_size;
562
563 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
564 last_far_set_start *= geo->far_set_size;
565
566 last_far_set_size = geo->far_set_size;
567 last_far_set_size += (geo->raid_disks % geo->far_set_size);
568
569 /* now calculate first sector/dev */
570 chunk = r10bio->sector >> geo->chunk_shift;
571 sector = r10bio->sector & geo->chunk_mask;
572
573 chunk *= geo->near_copies;
574 stripe = chunk;
575 dev = sector_div(stripe, geo->raid_disks);
576 if (geo->far_offset)
577 stripe *= geo->far_copies;
578
579 sector += stripe << geo->chunk_shift;
580
581 /* and calculate all the others */
582 for (n = 0; n < geo->near_copies; n++) {
583 int d = dev;
584 int set;
585 sector_t s = sector;
586 r10bio->devs[slot].devnum = d;
587 r10bio->devs[slot].addr = s;
588 slot++;
589
590 for (f = 1; f < geo->far_copies; f++) {
591 set = d / geo->far_set_size;
592 d += geo->near_copies;
593
594 if ((geo->raid_disks % geo->far_set_size) &&
595 (d > last_far_set_start)) {
596 d -= last_far_set_start;
597 d %= last_far_set_size;
598 d += last_far_set_start;
599 } else {
600 d %= geo->far_set_size;
601 d += geo->far_set_size * set;
602 }
603 s += geo->stride;
604 r10bio->devs[slot].devnum = d;
605 r10bio->devs[slot].addr = s;
606 slot++;
607 }
608 dev++;
609 if (dev >= geo->raid_disks) {
610 dev = 0;
611 sector += (geo->chunk_mask + 1);
612 }
613 }
614}
615
616static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
617{
618 struct geom *geo = &conf->geo;
619
620 if (conf->reshape_progress != MaxSector &&
621 ((r10bio->sector >= conf->reshape_progress) !=
622 conf->mddev->reshape_backwards)) {
623 set_bit(R10BIO_Previous, &r10bio->state);
624 geo = &conf->prev;
625 } else
626 clear_bit(R10BIO_Previous, &r10bio->state);
627
628 __raid10_find_phys(geo, r10bio);
629}
630
631static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
632{
633 sector_t offset, chunk, vchunk;
634 /* Never use conf->prev as this is only called during resync
635 * or recovery, so reshape isn't happening
636 */
637 struct geom *geo = &conf->geo;
638 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
639 int far_set_size = geo->far_set_size;
640 int last_far_set_start;
641
642 if (geo->raid_disks % geo->far_set_size) {
643 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
644 last_far_set_start *= geo->far_set_size;
645
646 if (dev >= last_far_set_start) {
647 far_set_size = geo->far_set_size;
648 far_set_size += (geo->raid_disks % geo->far_set_size);
649 far_set_start = last_far_set_start;
650 }
651 }
652
653 offset = sector & geo->chunk_mask;
654 if (geo->far_offset) {
655 int fc;
656 chunk = sector >> geo->chunk_shift;
657 fc = sector_div(chunk, geo->far_copies);
658 dev -= fc * geo->near_copies;
659 if (dev < far_set_start)
660 dev += far_set_size;
661 } else {
662 while (sector >= geo->stride) {
663 sector -= geo->stride;
664 if (dev < (geo->near_copies + far_set_start))
665 dev += far_set_size - geo->near_copies;
666 else
667 dev -= geo->near_copies;
668 }
669 chunk = sector >> geo->chunk_shift;
670 }
671 vchunk = chunk * geo->raid_disks + dev;
672 sector_div(vchunk, geo->near_copies);
673 return (vchunk << geo->chunk_shift) + offset;
674}
675
676/**
677 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
678 * @q: request queue
679 * @bvm: properties of new bio
680 * @biovec: the request that could be merged to it.
681 *
682 * Return amount of bytes we can accept at this offset
683 * This requires checking for end-of-chunk if near_copies != raid_disks,
684 * and for subordinate merge_bvec_fns if merge_check_needed.
685 */
686static int raid10_mergeable_bvec(struct request_queue *q,
687 struct bvec_merge_data *bvm,
688 struct bio_vec *biovec)
689{
690 struct mddev *mddev = q->queuedata;
691 struct r10conf *conf = mddev->private;
692 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
693 int max;
694 unsigned int chunk_sectors;
695 unsigned int bio_sectors = bvm->bi_size >> 9;
696 struct geom *geo = &conf->geo;
697
698 chunk_sectors = (conf->geo.chunk_mask & conf->prev.chunk_mask) + 1;
699 if (conf->reshape_progress != MaxSector &&
700 ((sector >= conf->reshape_progress) !=
701 conf->mddev->reshape_backwards))
702 geo = &conf->prev;
703
704 if (geo->near_copies < geo->raid_disks) {
705 max = (chunk_sectors - ((sector & (chunk_sectors - 1))
706 + bio_sectors)) << 9;
707 if (max < 0)
708 /* bio_add cannot handle a negative return */
709 max = 0;
710 if (max <= biovec->bv_len && bio_sectors == 0)
711 return biovec->bv_len;
712 } else
713 max = biovec->bv_len;
714
715 if (mddev->merge_check_needed) {
716 struct {
717 struct r10bio r10_bio;
718 struct r10dev devs[conf->copies];
719 } on_stack;
720 struct r10bio *r10_bio = &on_stack.r10_bio;
721 int s;
722 if (conf->reshape_progress != MaxSector) {
723 /* Cannot give any guidance during reshape */
724 if (max <= biovec->bv_len && bio_sectors == 0)
725 return biovec->bv_len;
726 return 0;
727 }
728 r10_bio->sector = sector;
729 raid10_find_phys(conf, r10_bio);
730 rcu_read_lock();
731 for (s = 0; s < conf->copies; s++) {
732 int disk = r10_bio->devs[s].devnum;
733 struct md_rdev *rdev = rcu_dereference(
734 conf->mirrors[disk].rdev);
735 if (rdev && !test_bit(Faulty, &rdev->flags)) {
736 struct request_queue *q =
737 bdev_get_queue(rdev->bdev);
738 if (q->merge_bvec_fn) {
739 bvm->bi_sector = r10_bio->devs[s].addr
740 + rdev->data_offset;
741 bvm->bi_bdev = rdev->bdev;
742 max = min(max, q->merge_bvec_fn(
743 q, bvm, biovec));
744 }
745 }
746 rdev = rcu_dereference(conf->mirrors[disk].replacement);
747 if (rdev && !test_bit(Faulty, &rdev->flags)) {
748 struct request_queue *q =
749 bdev_get_queue(rdev->bdev);
750 if (q->merge_bvec_fn) {
751 bvm->bi_sector = r10_bio->devs[s].addr
752 + rdev->data_offset;
753 bvm->bi_bdev = rdev->bdev;
754 max = min(max, q->merge_bvec_fn(
755 q, bvm, biovec));
756 }
757 }
758 }
759 rcu_read_unlock();
760 }
761 return max;
762}
763
764/*
765 * This routine returns the disk from which the requested read should
766 * be done. There is a per-array 'next expected sequential IO' sector
767 * number - if this matches on the next IO then we use the last disk.
768 * There is also a per-disk 'last know head position' sector that is
769 * maintained from IRQ contexts, both the normal and the resync IO
770 * completion handlers update this position correctly. If there is no
771 * perfect sequential match then we pick the disk whose head is closest.
772 *
773 * If there are 2 mirrors in the same 2 devices, performance degrades
774 * because position is mirror, not device based.
775 *
776 * The rdev for the device selected will have nr_pending incremented.
777 */
778
779/*
780 * FIXME: possibly should rethink readbalancing and do it differently
781 * depending on near_copies / far_copies geometry.
782 */
783static struct md_rdev *read_balance(struct r10conf *conf,
784 struct r10bio *r10_bio,
785 int *max_sectors)
786{
787 const sector_t this_sector = r10_bio->sector;
788 int disk, slot;
789 int sectors = r10_bio->sectors;
790 int best_good_sectors;
791 sector_t new_distance, best_dist;
792 struct md_rdev *best_rdev, *rdev = NULL;
793 int do_balance;
794 int best_slot;
795 struct geom *geo = &conf->geo;
796
797 raid10_find_phys(conf, r10_bio);
798 rcu_read_lock();
799retry:
800 sectors = r10_bio->sectors;
801 best_slot = -1;
802 best_rdev = NULL;
803 best_dist = MaxSector;
804 best_good_sectors = 0;
805 do_balance = 1;
806 /*
807 * Check if we can balance. We can balance on the whole
808 * device if no resync is going on (recovery is ok), or below
809 * the resync window. We take the first readable disk when
810 * above the resync window.
811 */
812 if (conf->mddev->recovery_cp < MaxSector
813 && (this_sector + sectors >= conf->next_resync))
814 do_balance = 0;
815
816 for (slot = 0; slot < conf->copies ; slot++) {
817 sector_t first_bad;
818 int bad_sectors;
819 sector_t dev_sector;
820
821 if (r10_bio->devs[slot].bio == IO_BLOCKED)
822 continue;
823 disk = r10_bio->devs[slot].devnum;
824 rdev = rcu_dereference(conf->mirrors[disk].replacement);
825 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
826 test_bit(Unmerged, &rdev->flags) ||
827 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
828 rdev = rcu_dereference(conf->mirrors[disk].rdev);
829 if (rdev == NULL ||
830 test_bit(Faulty, &rdev->flags) ||
831 test_bit(Unmerged, &rdev->flags))
832 continue;
833 if (!test_bit(In_sync, &rdev->flags) &&
834 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
835 continue;
836
837 dev_sector = r10_bio->devs[slot].addr;
838 if (is_badblock(rdev, dev_sector, sectors,
839 &first_bad, &bad_sectors)) {
840 if (best_dist < MaxSector)
841 /* Already have a better slot */
842 continue;
843 if (first_bad <= dev_sector) {
844 /* Cannot read here. If this is the
845 * 'primary' device, then we must not read
846 * beyond 'bad_sectors' from another device.
847 */
848 bad_sectors -= (dev_sector - first_bad);
849 if (!do_balance && sectors > bad_sectors)
850 sectors = bad_sectors;
851 if (best_good_sectors > sectors)
852 best_good_sectors = sectors;
853 } else {
854 sector_t good_sectors =
855 first_bad - dev_sector;
856 if (good_sectors > best_good_sectors) {
857 best_good_sectors = good_sectors;
858 best_slot = slot;
859 best_rdev = rdev;
860 }
861 if (!do_balance)
862 /* Must read from here */
863 break;
864 }
865 continue;
866 } else
867 best_good_sectors = sectors;
868
869 if (!do_balance)
870 break;
871
872 /* This optimisation is debatable, and completely destroys
873 * sequential read speed for 'far copies' arrays. So only
874 * keep it for 'near' arrays, and review those later.
875 */
876 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
877 break;
878
879 /* for far > 1 always use the lowest address */
880 if (geo->far_copies > 1)
881 new_distance = r10_bio->devs[slot].addr;
882 else
883 new_distance = abs(r10_bio->devs[slot].addr -
884 conf->mirrors[disk].head_position);
885 if (new_distance < best_dist) {
886 best_dist = new_distance;
887 best_slot = slot;
888 best_rdev = rdev;
889 }
890 }
891 if (slot >= conf->copies) {
892 slot = best_slot;
893 rdev = best_rdev;
894 }
895
896 if (slot >= 0) {
897 atomic_inc(&rdev->nr_pending);
898 if (test_bit(Faulty, &rdev->flags)) {
899 /* Cannot risk returning a device that failed
900 * before we inc'ed nr_pending
901 */
902 rdev_dec_pending(rdev, conf->mddev);
903 goto retry;
904 }
905 r10_bio->read_slot = slot;
906 } else
907 rdev = NULL;
908 rcu_read_unlock();
909 *max_sectors = best_good_sectors;
910
911 return rdev;
912}
913
914int md_raid10_congested(struct mddev *mddev, int bits)
915{
916 struct r10conf *conf = mddev->private;
917 int i, ret = 0;
918
919 if ((bits & (1 << BDI_async_congested)) &&
920 conf->pending_count >= max_queued_requests)
921 return 1;
922
923 rcu_read_lock();
924 for (i = 0;
925 (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
926 && ret == 0;
927 i++) {
928 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
929 if (rdev && !test_bit(Faulty, &rdev->flags)) {
930 struct request_queue *q = bdev_get_queue(rdev->bdev);
931
932 ret |= bdi_congested(&q->backing_dev_info, bits);
933 }
934 }
935 rcu_read_unlock();
936 return ret;
937}
938EXPORT_SYMBOL_GPL(md_raid10_congested);
939
940static int raid10_congested(void *data, int bits)
941{
942 struct mddev *mddev = data;
943
944 return mddev_congested(mddev, bits) ||
945 md_raid10_congested(mddev, bits);
946}
947
948static void flush_pending_writes(struct r10conf *conf)
949{
950 /* Any writes that have been queued but are awaiting
951 * bitmap updates get flushed here.
952 */
953 spin_lock_irq(&conf->device_lock);
954
955 if (conf->pending_bio_list.head) {
956 struct bio *bio;
957 bio = bio_list_get(&conf->pending_bio_list);
958 conf->pending_count = 0;
959 spin_unlock_irq(&conf->device_lock);
960 /* flush any pending bitmap writes to disk
961 * before proceeding w/ I/O */
962 bitmap_unplug(conf->mddev->bitmap);
963 wake_up(&conf->wait_barrier);
964
965 while (bio) { /* submit pending writes */
966 struct bio *next = bio->bi_next;
967 bio->bi_next = NULL;
968 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
969 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
970 /* Just ignore it */
971 bio_endio(bio, 0);
972 else
973 generic_make_request(bio);
974 bio = next;
975 }
976 } else
977 spin_unlock_irq(&conf->device_lock);
978}
979
980/* Barriers....
981 * Sometimes we need to suspend IO while we do something else,
982 * either some resync/recovery, or reconfigure the array.
983 * To do this we raise a 'barrier'.
984 * The 'barrier' is a counter that can be raised multiple times
985 * to count how many activities are happening which preclude
986 * normal IO.
987 * We can only raise the barrier if there is no pending IO.
988 * i.e. if nr_pending == 0.
989 * We choose only to raise the barrier if no-one is waiting for the
990 * barrier to go down. This means that as soon as an IO request
991 * is ready, no other operations which require a barrier will start
992 * until the IO request has had a chance.
993 *
994 * So: regular IO calls 'wait_barrier'. When that returns there
995 * is no backgroup IO happening, It must arrange to call
996 * allow_barrier when it has finished its IO.
997 * backgroup IO calls must call raise_barrier. Once that returns
998 * there is no normal IO happeing. It must arrange to call
999 * lower_barrier when the particular background IO completes.
1000 */
1001
1002static void raise_barrier(struct r10conf *conf, int force)
1003{
1004 BUG_ON(force && !conf->barrier);
1005 spin_lock_irq(&conf->resync_lock);
1006
1007 /* Wait until no block IO is waiting (unless 'force') */
1008 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
1009 conf->resync_lock);
1010
1011 /* block any new IO from starting */
1012 conf->barrier++;
1013
1014 /* Now wait for all pending IO to complete */
1015 wait_event_lock_irq(conf->wait_barrier,
1016 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
1017 conf->resync_lock);
1018
1019 spin_unlock_irq(&conf->resync_lock);
1020}
1021
1022static void lower_barrier(struct r10conf *conf)
1023{
1024 unsigned long flags;
1025 spin_lock_irqsave(&conf->resync_lock, flags);
1026 conf->barrier--;
1027 spin_unlock_irqrestore(&conf->resync_lock, flags);
1028 wake_up(&conf->wait_barrier);
1029}
1030
1031static void wait_barrier(struct r10conf *conf)
1032{
1033 spin_lock_irq(&conf->resync_lock);
1034 if (conf->barrier) {
1035 conf->nr_waiting++;
1036 /* Wait for the barrier to drop.
1037 * However if there are already pending
1038 * requests (preventing the barrier from
1039 * rising completely), and the
1040 * pre-process bio queue isn't empty,
1041 * then don't wait, as we need to empty
1042 * that queue to get the nr_pending
1043 * count down.
1044 */
1045 wait_event_lock_irq(conf->wait_barrier,
1046 !conf->barrier ||
1047 (conf->nr_pending &&
1048 current->bio_list &&
1049 !bio_list_empty(current->bio_list)),
1050 conf->resync_lock);
1051 conf->nr_waiting--;
1052 }
1053 conf->nr_pending++;
1054 spin_unlock_irq(&conf->resync_lock);
1055}
1056
1057static void allow_barrier(struct r10conf *conf)
1058{
1059 unsigned long flags;
1060 spin_lock_irqsave(&conf->resync_lock, flags);
1061 conf->nr_pending--;
1062 spin_unlock_irqrestore(&conf->resync_lock, flags);
1063 wake_up(&conf->wait_barrier);
1064}
1065
1066static void freeze_array(struct r10conf *conf, int extra)
1067{
1068 /* stop syncio and normal IO and wait for everything to
1069 * go quiet.
1070 * We increment barrier and nr_waiting, and then
1071 * wait until nr_pending match nr_queued+extra
1072 * This is called in the context of one normal IO request
1073 * that has failed. Thus any sync request that might be pending
1074 * will be blocked by nr_pending, and we need to wait for
1075 * pending IO requests to complete or be queued for re-try.
1076 * Thus the number queued (nr_queued) plus this request (extra)
1077 * must match the number of pending IOs (nr_pending) before
1078 * we continue.
1079 */
1080 spin_lock_irq(&conf->resync_lock);
1081 conf->barrier++;
1082 conf->nr_waiting++;
1083 wait_event_lock_irq_cmd(conf->wait_barrier,
1084 conf->nr_pending == conf->nr_queued+extra,
1085 conf->resync_lock,
1086 flush_pending_writes(conf));
1087
1088 spin_unlock_irq(&conf->resync_lock);
1089}
1090
1091static void unfreeze_array(struct r10conf *conf)
1092{
1093 /* reverse the effect of the freeze */
1094 spin_lock_irq(&conf->resync_lock);
1095 conf->barrier--;
1096 conf->nr_waiting--;
1097 wake_up(&conf->wait_barrier);
1098 spin_unlock_irq(&conf->resync_lock);
1099}
1100
1101static sector_t choose_data_offset(struct r10bio *r10_bio,
1102 struct md_rdev *rdev)
1103{
1104 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1105 test_bit(R10BIO_Previous, &r10_bio->state))
1106 return rdev->data_offset;
1107 else
1108 return rdev->new_data_offset;
1109}
1110
1111struct raid10_plug_cb {
1112 struct blk_plug_cb cb;
1113 struct bio_list pending;
1114 int pending_cnt;
1115};
1116
1117static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1118{
1119 struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1120 cb);
1121 struct mddev *mddev = plug->cb.data;
1122 struct r10conf *conf = mddev->private;
1123 struct bio *bio;
1124
1125 if (from_schedule || current->bio_list) {
1126 spin_lock_irq(&conf->device_lock);
1127 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1128 conf->pending_count += plug->pending_cnt;
1129 spin_unlock_irq(&conf->device_lock);
1130 wake_up(&conf->wait_barrier);
1131 md_wakeup_thread(mddev->thread);
1132 kfree(plug);
1133 return;
1134 }
1135
1136 /* we aren't scheduling, so we can do the write-out directly. */
1137 bio = bio_list_get(&plug->pending);
1138 bitmap_unplug(mddev->bitmap);
1139 wake_up(&conf->wait_barrier);
1140
1141 while (bio) { /* submit pending writes */
1142 struct bio *next = bio->bi_next;
1143 bio->bi_next = NULL;
1144 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
1145 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1146 /* Just ignore it */
1147 bio_endio(bio, 0);
1148 else
1149 generic_make_request(bio);
1150 bio = next;
1151 }
1152 kfree(plug);
1153}
1154
1155static void __make_request(struct mddev *mddev, struct bio *bio)
1156{
1157 struct r10conf *conf = mddev->private;
1158 struct r10bio *r10_bio;
1159 struct bio *read_bio;
1160 int i;
1161 const int rw = bio_data_dir(bio);
1162 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1163 const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
1164 const unsigned long do_discard = (bio->bi_rw
1165 & (REQ_DISCARD | REQ_SECURE));
1166 const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1167 unsigned long flags;
1168 struct md_rdev *blocked_rdev;
1169 struct blk_plug_cb *cb;
1170 struct raid10_plug_cb *plug = NULL;
1171 int sectors_handled;
1172 int max_sectors;
1173 int sectors;
1174
1175 /*
1176 * Register the new request and wait if the reconstruction
1177 * thread has put up a bar for new requests.
1178 * Continue immediately if no resync is active currently.
1179 */
1180 wait_barrier(conf);
1181
1182 sectors = bio_sectors(bio);
1183 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1184 bio->bi_iter.bi_sector < conf->reshape_progress &&
1185 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1186 /* IO spans the reshape position. Need to wait for
1187 * reshape to pass
1188 */
1189 allow_barrier(conf);
1190 wait_event(conf->wait_barrier,
1191 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1192 conf->reshape_progress >= bio->bi_iter.bi_sector +
1193 sectors);
1194 wait_barrier(conf);
1195 }
1196 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1197 bio_data_dir(bio) == WRITE &&
1198 (mddev->reshape_backwards
1199 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1200 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1201 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1202 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1203 /* Need to update reshape_position in metadata */
1204 mddev->reshape_position = conf->reshape_progress;
1205 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1206 set_bit(MD_CHANGE_PENDING, &mddev->flags);
1207 md_wakeup_thread(mddev->thread);
1208 wait_event(mddev->sb_wait,
1209 !test_bit(MD_CHANGE_PENDING, &mddev->flags));
1210
1211 conf->reshape_safe = mddev->reshape_position;
1212 }
1213
1214 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1215
1216 r10_bio->master_bio = bio;
1217 r10_bio->sectors = sectors;
1218
1219 r10_bio->mddev = mddev;
1220 r10_bio->sector = bio->bi_iter.bi_sector;
1221 r10_bio->state = 0;
1222
1223 /* We might need to issue multiple reads to different
1224 * devices if there are bad blocks around, so we keep
1225 * track of the number of reads in bio->bi_phys_segments.
1226 * If this is 0, there is only one r10_bio and no locking
1227 * will be needed when the request completes. If it is
1228 * non-zero, then it is the number of not-completed requests.
1229 */
1230 bio->bi_phys_segments = 0;
1231 clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1232
1233 if (rw == READ) {
1234 /*
1235 * read balancing logic:
1236 */
1237 struct md_rdev *rdev;
1238 int slot;
1239
1240read_again:
1241 rdev = read_balance(conf, r10_bio, &max_sectors);
1242 if (!rdev) {
1243 raid_end_bio_io(r10_bio);
1244 return;
1245 }
1246 slot = r10_bio->read_slot;
1247
1248 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1249 bio_trim(read_bio, r10_bio->sector - bio->bi_iter.bi_sector,
1250 max_sectors);
1251
1252 r10_bio->devs[slot].bio = read_bio;
1253 r10_bio->devs[slot].rdev = rdev;
1254
1255 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1256 choose_data_offset(r10_bio, rdev);
1257 read_bio->bi_bdev = rdev->bdev;
1258 read_bio->bi_end_io = raid10_end_read_request;
1259 read_bio->bi_rw = READ | do_sync;
1260 read_bio->bi_private = r10_bio;
1261
1262 if (max_sectors < r10_bio->sectors) {
1263 /* Could not read all from this device, so we will
1264 * need another r10_bio.
1265 */
1266 sectors_handled = (r10_bio->sector + max_sectors
1267 - bio->bi_iter.bi_sector);
1268 r10_bio->sectors = max_sectors;
1269 spin_lock_irq(&conf->device_lock);
1270 if (bio->bi_phys_segments == 0)
1271 bio->bi_phys_segments = 2;
1272 else
1273 bio->bi_phys_segments++;
1274 spin_unlock_irq(&conf->device_lock);
1275 /* Cannot call generic_make_request directly
1276 * as that will be queued in __generic_make_request
1277 * and subsequent mempool_alloc might block
1278 * waiting for it. so hand bio over to raid10d.
1279 */
1280 reschedule_retry(r10_bio);
1281
1282 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1283
1284 r10_bio->master_bio = bio;
1285 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1286 r10_bio->state = 0;
1287 r10_bio->mddev = mddev;
1288 r10_bio->sector = bio->bi_iter.bi_sector +
1289 sectors_handled;
1290 goto read_again;
1291 } else
1292 generic_make_request(read_bio);
1293 return;
1294 }
1295
1296 /*
1297 * WRITE:
1298 */
1299 if (conf->pending_count >= max_queued_requests) {
1300 md_wakeup_thread(mddev->thread);
1301 wait_event(conf->wait_barrier,
1302 conf->pending_count < max_queued_requests);
1303 }
1304 /* first select target devices under rcu_lock and
1305 * inc refcount on their rdev. Record them by setting
1306 * bios[x] to bio
1307 * If there are known/acknowledged bad blocks on any device
1308 * on which we have seen a write error, we want to avoid
1309 * writing to those blocks. This potentially requires several
1310 * writes to write around the bad blocks. Each set of writes
1311 * gets its own r10_bio with a set of bios attached. The number
1312 * of r10_bios is recored in bio->bi_phys_segments just as with
1313 * the read case.
1314 */
1315
1316 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1317 raid10_find_phys(conf, r10_bio);
1318retry_write:
1319 blocked_rdev = NULL;
1320 rcu_read_lock();
1321 max_sectors = r10_bio->sectors;
1322
1323 for (i = 0; i < conf->copies; i++) {
1324 int d = r10_bio->devs[i].devnum;
1325 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1326 struct md_rdev *rrdev = rcu_dereference(
1327 conf->mirrors[d].replacement);
1328 if (rdev == rrdev)
1329 rrdev = NULL;
1330 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1331 atomic_inc(&rdev->nr_pending);
1332 blocked_rdev = rdev;
1333 break;
1334 }
1335 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1336 atomic_inc(&rrdev->nr_pending);
1337 blocked_rdev = rrdev;
1338 break;
1339 }
1340 if (rdev && (test_bit(Faulty, &rdev->flags)
1341 || test_bit(Unmerged, &rdev->flags)))
1342 rdev = NULL;
1343 if (rrdev && (test_bit(Faulty, &rrdev->flags)
1344 || test_bit(Unmerged, &rrdev->flags)))
1345 rrdev = NULL;
1346
1347 r10_bio->devs[i].bio = NULL;
1348 r10_bio->devs[i].repl_bio = NULL;
1349
1350 if (!rdev && !rrdev) {
1351 set_bit(R10BIO_Degraded, &r10_bio->state);
1352 continue;
1353 }
1354 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1355 sector_t first_bad;
1356 sector_t dev_sector = r10_bio->devs[i].addr;
1357 int bad_sectors;
1358 int is_bad;
1359
1360 is_bad = is_badblock(rdev, dev_sector,
1361 max_sectors,
1362 &first_bad, &bad_sectors);
1363 if (is_bad < 0) {
1364 /* Mustn't write here until the bad block
1365 * is acknowledged
1366 */
1367 atomic_inc(&rdev->nr_pending);
1368 set_bit(BlockedBadBlocks, &rdev->flags);
1369 blocked_rdev = rdev;
1370 break;
1371 }
1372 if (is_bad && first_bad <= dev_sector) {
1373 /* Cannot write here at all */
1374 bad_sectors -= (dev_sector - first_bad);
1375 if (bad_sectors < max_sectors)
1376 /* Mustn't write more than bad_sectors
1377 * to other devices yet
1378 */
1379 max_sectors = bad_sectors;
1380 /* We don't set R10BIO_Degraded as that
1381 * only applies if the disk is missing,
1382 * so it might be re-added, and we want to
1383 * know to recover this chunk.
1384 * In this case the device is here, and the
1385 * fact that this chunk is not in-sync is
1386 * recorded in the bad block log.
1387 */
1388 continue;
1389 }
1390 if (is_bad) {
1391 int good_sectors = first_bad - dev_sector;
1392 if (good_sectors < max_sectors)
1393 max_sectors = good_sectors;
1394 }
1395 }
1396 if (rdev) {
1397 r10_bio->devs[i].bio = bio;
1398 atomic_inc(&rdev->nr_pending);
1399 }
1400 if (rrdev) {
1401 r10_bio->devs[i].repl_bio = bio;
1402 atomic_inc(&rrdev->nr_pending);
1403 }
1404 }
1405 rcu_read_unlock();
1406
1407 if (unlikely(blocked_rdev)) {
1408 /* Have to wait for this device to get unblocked, then retry */
1409 int j;
1410 int d;
1411
1412 for (j = 0; j < i; j++) {
1413 if (r10_bio->devs[j].bio) {
1414 d = r10_bio->devs[j].devnum;
1415 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1416 }
1417 if (r10_bio->devs[j].repl_bio) {
1418 struct md_rdev *rdev;
1419 d = r10_bio->devs[j].devnum;
1420 rdev = conf->mirrors[d].replacement;
1421 if (!rdev) {
1422 /* Race with remove_disk */
1423 smp_mb();
1424 rdev = conf->mirrors[d].rdev;
1425 }
1426 rdev_dec_pending(rdev, mddev);
1427 }
1428 }
1429 allow_barrier(conf);
1430 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1431 wait_barrier(conf);
1432 goto retry_write;
1433 }
1434
1435 if (max_sectors < r10_bio->sectors) {
1436 /* We are splitting this into multiple parts, so
1437 * we need to prepare for allocating another r10_bio.
1438 */
1439 r10_bio->sectors = max_sectors;
1440 spin_lock_irq(&conf->device_lock);
1441 if (bio->bi_phys_segments == 0)
1442 bio->bi_phys_segments = 2;
1443 else
1444 bio->bi_phys_segments++;
1445 spin_unlock_irq(&conf->device_lock);
1446 }
1447 sectors_handled = r10_bio->sector + max_sectors -
1448 bio->bi_iter.bi_sector;
1449
1450 atomic_set(&r10_bio->remaining, 1);
1451 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1452
1453 for (i = 0; i < conf->copies; i++) {
1454 struct bio *mbio;
1455 int d = r10_bio->devs[i].devnum;
1456 if (r10_bio->devs[i].bio) {
1457 struct md_rdev *rdev = conf->mirrors[d].rdev;
1458 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1459 bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1460 max_sectors);
1461 r10_bio->devs[i].bio = mbio;
1462
1463 mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+
1464 choose_data_offset(r10_bio,
1465 rdev));
1466 mbio->bi_bdev = rdev->bdev;
1467 mbio->bi_end_io = raid10_end_write_request;
1468 mbio->bi_rw =
1469 WRITE | do_sync | do_fua | do_discard | do_same;
1470 mbio->bi_private = r10_bio;
1471
1472 atomic_inc(&r10_bio->remaining);
1473
1474 cb = blk_check_plugged(raid10_unplug, mddev,
1475 sizeof(*plug));
1476 if (cb)
1477 plug = container_of(cb, struct raid10_plug_cb,
1478 cb);
1479 else
1480 plug = NULL;
1481 spin_lock_irqsave(&conf->device_lock, flags);
1482 if (plug) {
1483 bio_list_add(&plug->pending, mbio);
1484 plug->pending_cnt++;
1485 } else {
1486 bio_list_add(&conf->pending_bio_list, mbio);
1487 conf->pending_count++;
1488 }
1489 spin_unlock_irqrestore(&conf->device_lock, flags);
1490 if (!plug)
1491 md_wakeup_thread(mddev->thread);
1492 }
1493
1494 if (r10_bio->devs[i].repl_bio) {
1495 struct md_rdev *rdev = conf->mirrors[d].replacement;
1496 if (rdev == NULL) {
1497 /* Replacement just got moved to main 'rdev' */
1498 smp_mb();
1499 rdev = conf->mirrors[d].rdev;
1500 }
1501 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1502 bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1503 max_sectors);
1504 r10_bio->devs[i].repl_bio = mbio;
1505
1506 mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr +
1507 choose_data_offset(
1508 r10_bio, rdev));
1509 mbio->bi_bdev = rdev->bdev;
1510 mbio->bi_end_io = raid10_end_write_request;
1511 mbio->bi_rw =
1512 WRITE | do_sync | do_fua | do_discard | do_same;
1513 mbio->bi_private = r10_bio;
1514
1515 atomic_inc(&r10_bio->remaining);
1516 spin_lock_irqsave(&conf->device_lock, flags);
1517 bio_list_add(&conf->pending_bio_list, mbio);
1518 conf->pending_count++;
1519 spin_unlock_irqrestore(&conf->device_lock, flags);
1520 if (!mddev_check_plugged(mddev))
1521 md_wakeup_thread(mddev->thread);
1522 }
1523 }
1524
1525 /* Don't remove the bias on 'remaining' (one_write_done) until
1526 * after checking if we need to go around again.
1527 */
1528
1529 if (sectors_handled < bio_sectors(bio)) {
1530 one_write_done(r10_bio);
1531 /* We need another r10_bio. It has already been counted
1532 * in bio->bi_phys_segments.
1533 */
1534 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1535
1536 r10_bio->master_bio = bio;
1537 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1538
1539 r10_bio->mddev = mddev;
1540 r10_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1541 r10_bio->state = 0;
1542 goto retry_write;
1543 }
1544 one_write_done(r10_bio);
1545}
1546
1547static void make_request(struct mddev *mddev, struct bio *bio)
1548{
1549 struct r10conf *conf = mddev->private;
1550 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1551 int chunk_sects = chunk_mask + 1;
1552
1553 struct bio *split;
1554
1555 if (unlikely(bio->bi_rw & REQ_FLUSH)) {
1556 md_flush_request(mddev, bio);
1557 return;
1558 }
1559
1560 md_write_start(mddev, bio);
1561
1562
1563 do {
1564
1565 /*
1566 * If this request crosses a chunk boundary, we need to split
1567 * it.
1568 */
1569 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1570 bio_sectors(bio) > chunk_sects
1571 && (conf->geo.near_copies < conf->geo.raid_disks
1572 || conf->prev.near_copies <
1573 conf->prev.raid_disks))) {
1574 split = bio_split(bio, chunk_sects -
1575 (bio->bi_iter.bi_sector &
1576 (chunk_sects - 1)),
1577 GFP_NOIO, fs_bio_set);
1578 bio_chain(split, bio);
1579 } else {
1580 split = bio;
1581 }
1582
1583 __make_request(mddev, split);
1584 } while (split != bio);
1585
1586 /* In case raid10d snuck in to freeze_array */
1587 wake_up(&conf->wait_barrier);
1588}
1589
1590static void status(struct seq_file *seq, struct mddev *mddev)
1591{
1592 struct r10conf *conf = mddev->private;
1593 int i;
1594
1595 if (conf->geo.near_copies < conf->geo.raid_disks)
1596 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1597 if (conf->geo.near_copies > 1)
1598 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1599 if (conf->geo.far_copies > 1) {
1600 if (conf->geo.far_offset)
1601 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1602 else
1603 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1604 }
1605 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1606 conf->geo.raid_disks - mddev->degraded);
1607 for (i = 0; i < conf->geo.raid_disks; i++)
1608 seq_printf(seq, "%s",
1609 conf->mirrors[i].rdev &&
1610 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1611 seq_printf(seq, "]");
1612}
1613
1614/* check if there are enough drives for
1615 * every block to appear on atleast one.
1616 * Don't consider the device numbered 'ignore'
1617 * as we might be about to remove it.
1618 */
1619static int _enough(struct r10conf *conf, int previous, int ignore)
1620{
1621 int first = 0;
1622 int has_enough = 0;
1623 int disks, ncopies;
1624 if (previous) {
1625 disks = conf->prev.raid_disks;
1626 ncopies = conf->prev.near_copies;
1627 } else {
1628 disks = conf->geo.raid_disks;
1629 ncopies = conf->geo.near_copies;
1630 }
1631
1632 rcu_read_lock();
1633 do {
1634 int n = conf->copies;
1635 int cnt = 0;
1636 int this = first;
1637 while (n--) {
1638 struct md_rdev *rdev;
1639 if (this != ignore &&
1640 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1641 test_bit(In_sync, &rdev->flags))
1642 cnt++;
1643 this = (this+1) % disks;
1644 }
1645 if (cnt == 0)
1646 goto out;
1647 first = (first + ncopies) % disks;
1648 } while (first != 0);
1649 has_enough = 1;
1650out:
1651 rcu_read_unlock();
1652 return has_enough;
1653}
1654
1655static int enough(struct r10conf *conf, int ignore)
1656{
1657 /* when calling 'enough', both 'prev' and 'geo' must
1658 * be stable.
1659 * This is ensured if ->reconfig_mutex or ->device_lock
1660 * is held.
1661 */
1662 return _enough(conf, 0, ignore) &&
1663 _enough(conf, 1, ignore);
1664}
1665
1666static void error(struct mddev *mddev, struct md_rdev *rdev)
1667{
1668 char b[BDEVNAME_SIZE];
1669 struct r10conf *conf = mddev->private;
1670 unsigned long flags;
1671
1672 /*
1673 * If it is not operational, then we have already marked it as dead
1674 * else if it is the last working disks, ignore the error, let the
1675 * next level up know.
1676 * else mark the drive as failed
1677 */
1678 spin_lock_irqsave(&conf->device_lock, flags);
1679 if (test_bit(In_sync, &rdev->flags)
1680 && !enough(conf, rdev->raid_disk)) {
1681 /*
1682 * Don't fail the drive, just return an IO error.
1683 */
1684 spin_unlock_irqrestore(&conf->device_lock, flags);
1685 return;
1686 }
1687 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1688 mddev->degraded++;
1689 /*
1690 * if recovery is running, make sure it aborts.
1691 */
1692 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1693 }
1694 set_bit(Blocked, &rdev->flags);
1695 set_bit(Faulty, &rdev->flags);
1696 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1697 spin_unlock_irqrestore(&conf->device_lock, flags);
1698 printk(KERN_ALERT
1699 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1700 "md/raid10:%s: Operation continuing on %d devices.\n",
1701 mdname(mddev), bdevname(rdev->bdev, b),
1702 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1703}
1704
1705static void print_conf(struct r10conf *conf)
1706{
1707 int i;
1708 struct raid10_info *tmp;
1709
1710 printk(KERN_DEBUG "RAID10 conf printout:\n");
1711 if (!conf) {
1712 printk(KERN_DEBUG "(!conf)\n");
1713 return;
1714 }
1715 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1716 conf->geo.raid_disks);
1717
1718 for (i = 0; i < conf->geo.raid_disks; i++) {
1719 char b[BDEVNAME_SIZE];
1720 tmp = conf->mirrors + i;
1721 if (tmp->rdev)
1722 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1723 i, !test_bit(In_sync, &tmp->rdev->flags),
1724 !test_bit(Faulty, &tmp->rdev->flags),
1725 bdevname(tmp->rdev->bdev,b));
1726 }
1727}
1728
1729static void close_sync(struct r10conf *conf)
1730{
1731 wait_barrier(conf);
1732 allow_barrier(conf);
1733
1734 mempool_destroy(conf->r10buf_pool);
1735 conf->r10buf_pool = NULL;
1736}
1737
1738static int raid10_spare_active(struct mddev *mddev)
1739{
1740 int i;
1741 struct r10conf *conf = mddev->private;
1742 struct raid10_info *tmp;
1743 int count = 0;
1744 unsigned long flags;
1745
1746 /*
1747 * Find all non-in_sync disks within the RAID10 configuration
1748 * and mark them in_sync
1749 */
1750 for (i = 0; i < conf->geo.raid_disks; i++) {
1751 tmp = conf->mirrors + i;
1752 if (tmp->replacement
1753 && tmp->replacement->recovery_offset == MaxSector
1754 && !test_bit(Faulty, &tmp->replacement->flags)
1755 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1756 /* Replacement has just become active */
1757 if (!tmp->rdev
1758 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1759 count++;
1760 if (tmp->rdev) {
1761 /* Replaced device not technically faulty,
1762 * but we need to be sure it gets removed
1763 * and never re-added.
1764 */
1765 set_bit(Faulty, &tmp->rdev->flags);
1766 sysfs_notify_dirent_safe(
1767 tmp->rdev->sysfs_state);
1768 }
1769 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1770 } else if (tmp->rdev
1771 && tmp->rdev->recovery_offset == MaxSector
1772 && !test_bit(Faulty, &tmp->rdev->flags)
1773 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1774 count++;
1775 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1776 }
1777 }
1778 spin_lock_irqsave(&conf->device_lock, flags);
1779 mddev->degraded -= count;
1780 spin_unlock_irqrestore(&conf->device_lock, flags);
1781
1782 print_conf(conf);
1783 return count;
1784}
1785
1786
1787static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1788{
1789 struct r10conf *conf = mddev->private;
1790 int err = -EEXIST;
1791 int mirror;
1792 int first = 0;
1793 int last = conf->geo.raid_disks - 1;
1794 struct request_queue *q = bdev_get_queue(rdev->bdev);
1795
1796 if (mddev->recovery_cp < MaxSector)
1797 /* only hot-add to in-sync arrays, as recovery is
1798 * very different from resync
1799 */
1800 return -EBUSY;
1801 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1802 return -EINVAL;
1803
1804 if (rdev->raid_disk >= 0)
1805 first = last = rdev->raid_disk;
1806
1807 if (q->merge_bvec_fn) {
1808 set_bit(Unmerged, &rdev->flags);
1809 mddev->merge_check_needed = 1;
1810 }
1811
1812 if (rdev->saved_raid_disk >= first &&
1813 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1814 mirror = rdev->saved_raid_disk;
1815 else
1816 mirror = first;
1817 for ( ; mirror <= last ; mirror++) {
1818 struct raid10_info *p = &conf->mirrors[mirror];
1819 if (p->recovery_disabled == mddev->recovery_disabled)
1820 continue;
1821 if (p->rdev) {
1822 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1823 p->replacement != NULL)
1824 continue;
1825 clear_bit(In_sync, &rdev->flags);
1826 set_bit(Replacement, &rdev->flags);
1827 rdev->raid_disk = mirror;
1828 err = 0;
1829 if (mddev->gendisk)
1830 disk_stack_limits(mddev->gendisk, rdev->bdev,
1831 rdev->data_offset << 9);
1832 conf->fullsync = 1;
1833 rcu_assign_pointer(p->replacement, rdev);
1834 break;
1835 }
1836
1837 if (mddev->gendisk)
1838 disk_stack_limits(mddev->gendisk, rdev->bdev,
1839 rdev->data_offset << 9);
1840
1841 p->head_position = 0;
1842 p->recovery_disabled = mddev->recovery_disabled - 1;
1843 rdev->raid_disk = mirror;
1844 err = 0;
1845 if (rdev->saved_raid_disk != mirror)
1846 conf->fullsync = 1;
1847 rcu_assign_pointer(p->rdev, rdev);
1848 break;
1849 }
1850 if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1851 /* Some requests might not have seen this new
1852 * merge_bvec_fn. We must wait for them to complete
1853 * before merging the device fully.
1854 * First we make sure any code which has tested
1855 * our function has submitted the request, then
1856 * we wait for all outstanding requests to complete.
1857 */
1858 synchronize_sched();
1859 freeze_array(conf, 0);
1860 unfreeze_array(conf);
1861 clear_bit(Unmerged, &rdev->flags);
1862 }
1863 md_integrity_add_rdev(rdev, mddev);
1864 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1865 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1866
1867 print_conf(conf);
1868 return err;
1869}
1870
1871static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1872{
1873 struct r10conf *conf = mddev->private;
1874 int err = 0;
1875 int number = rdev->raid_disk;
1876 struct md_rdev **rdevp;
1877 struct raid10_info *p = conf->mirrors + number;
1878
1879 print_conf(conf);
1880 if (rdev == p->rdev)
1881 rdevp = &p->rdev;
1882 else if (rdev == p->replacement)
1883 rdevp = &p->replacement;
1884 else
1885 return 0;
1886
1887 if (test_bit(In_sync, &rdev->flags) ||
1888 atomic_read(&rdev->nr_pending)) {
1889 err = -EBUSY;
1890 goto abort;
1891 }
1892 /* Only remove faulty devices if recovery
1893 * is not possible.
1894 */
1895 if (!test_bit(Faulty, &rdev->flags) &&
1896 mddev->recovery_disabled != p->recovery_disabled &&
1897 (!p->replacement || p->replacement == rdev) &&
1898 number < conf->geo.raid_disks &&
1899 enough(conf, -1)) {
1900 err = -EBUSY;
1901 goto abort;
1902 }
1903 *rdevp = NULL;
1904 synchronize_rcu();
1905 if (atomic_read(&rdev->nr_pending)) {
1906 /* lost the race, try later */
1907 err = -EBUSY;
1908 *rdevp = rdev;
1909 goto abort;
1910 } else if (p->replacement) {
1911 /* We must have just cleared 'rdev' */
1912 p->rdev = p->replacement;
1913 clear_bit(Replacement, &p->replacement->flags);
1914 smp_mb(); /* Make sure other CPUs may see both as identical
1915 * but will never see neither -- if they are careful.
1916 */
1917 p->replacement = NULL;
1918 clear_bit(WantReplacement, &rdev->flags);
1919 } else
1920 /* We might have just remove the Replacement as faulty
1921 * Clear the flag just in case
1922 */
1923 clear_bit(WantReplacement, &rdev->flags);
1924
1925 err = md_integrity_register(mddev);
1926
1927abort:
1928
1929 print_conf(conf);
1930 return err;
1931}
1932
1933
1934static void end_sync_read(struct bio *bio, int error)
1935{
1936 struct r10bio *r10_bio = bio->bi_private;
1937 struct r10conf *conf = r10_bio->mddev->private;
1938 int d;
1939
1940 if (bio == r10_bio->master_bio) {
1941 /* this is a reshape read */
1942 d = r10_bio->read_slot; /* really the read dev */
1943 } else
1944 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1945
1946 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1947 set_bit(R10BIO_Uptodate, &r10_bio->state);
1948 else
1949 /* The write handler will notice the lack of
1950 * R10BIO_Uptodate and record any errors etc
1951 */
1952 atomic_add(r10_bio->sectors,
1953 &conf->mirrors[d].rdev->corrected_errors);
1954
1955 /* for reconstruct, we always reschedule after a read.
1956 * for resync, only after all reads
1957 */
1958 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1959 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1960 atomic_dec_and_test(&r10_bio->remaining)) {
1961 /* we have read all the blocks,
1962 * do the comparison in process context in raid10d
1963 */
1964 reschedule_retry(r10_bio);
1965 }
1966}
1967
1968static void end_sync_request(struct r10bio *r10_bio)
1969{
1970 struct mddev *mddev = r10_bio->mddev;
1971
1972 while (atomic_dec_and_test(&r10_bio->remaining)) {
1973 if (r10_bio->master_bio == NULL) {
1974 /* the primary of several recovery bios */
1975 sector_t s = r10_bio->sectors;
1976 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1977 test_bit(R10BIO_WriteError, &r10_bio->state))
1978 reschedule_retry(r10_bio);
1979 else
1980 put_buf(r10_bio);
1981 md_done_sync(mddev, s, 1);
1982 break;
1983 } else {
1984 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1985 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1986 test_bit(R10BIO_WriteError, &r10_bio->state))
1987 reschedule_retry(r10_bio);
1988 else
1989 put_buf(r10_bio);
1990 r10_bio = r10_bio2;
1991 }
1992 }
1993}
1994
1995static void end_sync_write(struct bio *bio, int error)
1996{
1997 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1998 struct r10bio *r10_bio = bio->bi_private;
1999 struct mddev *mddev = r10_bio->mddev;
2000 struct r10conf *conf = mddev->private;
2001 int d;
2002 sector_t first_bad;
2003 int bad_sectors;
2004 int slot;
2005 int repl;
2006 struct md_rdev *rdev = NULL;
2007
2008 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
2009 if (repl)
2010 rdev = conf->mirrors[d].replacement;
2011 else
2012 rdev = conf->mirrors[d].rdev;
2013
2014 if (!uptodate) {
2015 if (repl)
2016 md_error(mddev, rdev);
2017 else {
2018 set_bit(WriteErrorSeen, &rdev->flags);
2019 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2020 set_bit(MD_RECOVERY_NEEDED,
2021 &rdev->mddev->recovery);
2022 set_bit(R10BIO_WriteError, &r10_bio->state);
2023 }
2024 } else if (is_badblock(rdev,
2025 r10_bio->devs[slot].addr,
2026 r10_bio->sectors,
2027 &first_bad, &bad_sectors))
2028 set_bit(R10BIO_MadeGood, &r10_bio->state);
2029
2030 rdev_dec_pending(rdev, mddev);
2031
2032 end_sync_request(r10_bio);
2033}
2034
2035/*
2036 * Note: sync and recover and handled very differently for raid10
2037 * This code is for resync.
2038 * For resync, we read through virtual addresses and read all blocks.
2039 * If there is any error, we schedule a write. The lowest numbered
2040 * drive is authoritative.
2041 * However requests come for physical address, so we need to map.
2042 * For every physical address there are raid_disks/copies virtual addresses,
2043 * which is always are least one, but is not necessarly an integer.
2044 * This means that a physical address can span multiple chunks, so we may
2045 * have to submit multiple io requests for a single sync request.
2046 */
2047/*
2048 * We check if all blocks are in-sync and only write to blocks that
2049 * aren't in sync
2050 */
2051static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2052{
2053 struct r10conf *conf = mddev->private;
2054 int i, first;
2055 struct bio *tbio, *fbio;
2056 int vcnt;
2057
2058 atomic_set(&r10_bio->remaining, 1);
2059
2060 /* find the first device with a block */
2061 for (i=0; i<conf->copies; i++)
2062 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
2063 break;
2064
2065 if (i == conf->copies)
2066 goto done;
2067
2068 first = i;
2069 fbio = r10_bio->devs[i].bio;
2070
2071 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2072 /* now find blocks with errors */
2073 for (i=0 ; i < conf->copies ; i++) {
2074 int j, d;
2075
2076 tbio = r10_bio->devs[i].bio;
2077
2078 if (tbio->bi_end_io != end_sync_read)
2079 continue;
2080 if (i == first)
2081 continue;
2082 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
2083 /* We know that the bi_io_vec layout is the same for
2084 * both 'first' and 'i', so we just compare them.
2085 * All vec entries are PAGE_SIZE;
2086 */
2087 int sectors = r10_bio->sectors;
2088 for (j = 0; j < vcnt; j++) {
2089 int len = PAGE_SIZE;
2090 if (sectors < (len / 512))
2091 len = sectors * 512;
2092 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
2093 page_address(tbio->bi_io_vec[j].bv_page),
2094 len))
2095 break;
2096 sectors -= len/512;
2097 }
2098 if (j == vcnt)
2099 continue;
2100 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2101 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2102 /* Don't fix anything. */
2103 continue;
2104 }
2105 /* Ok, we need to write this bio, either to correct an
2106 * inconsistency or to correct an unreadable block.
2107 * First we need to fixup bv_offset, bv_len and
2108 * bi_vecs, as the read request might have corrupted these
2109 */
2110 bio_reset(tbio);
2111
2112 tbio->bi_vcnt = vcnt;
2113 tbio->bi_iter.bi_size = r10_bio->sectors << 9;
2114 tbio->bi_rw = WRITE;
2115 tbio->bi_private = r10_bio;
2116 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2117
2118 for (j=0; j < vcnt ; j++) {
2119 tbio->bi_io_vec[j].bv_offset = 0;
2120 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
2121
2122 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2123 page_address(fbio->bi_io_vec[j].bv_page),
2124 PAGE_SIZE);
2125 }
2126 tbio->bi_end_io = end_sync_write;
2127
2128 d = r10_bio->devs[i].devnum;
2129 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2130 atomic_inc(&r10_bio->remaining);
2131 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2132
2133 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2134 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2135 generic_make_request(tbio);
2136 }
2137
2138 /* Now write out to any replacement devices
2139 * that are active
2140 */
2141 for (i = 0; i < conf->copies; i++) {
2142 int j, d;
2143
2144 tbio = r10_bio->devs[i].repl_bio;
2145 if (!tbio || !tbio->bi_end_io)
2146 continue;
2147 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2148 && r10_bio->devs[i].bio != fbio)
2149 for (j = 0; j < vcnt; j++)
2150 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2151 page_address(fbio->bi_io_vec[j].bv_page),
2152 PAGE_SIZE);
2153 d = r10_bio->devs[i].devnum;
2154 atomic_inc(&r10_bio->remaining);
2155 md_sync_acct(conf->mirrors[d].replacement->bdev,
2156 bio_sectors(tbio));
2157 generic_make_request(tbio);
2158 }
2159
2160done:
2161 if (atomic_dec_and_test(&r10_bio->remaining)) {
2162 md_done_sync(mddev, r10_bio->sectors, 1);
2163 put_buf(r10_bio);
2164 }
2165}
2166
2167/*
2168 * Now for the recovery code.
2169 * Recovery happens across physical sectors.
2170 * We recover all non-is_sync drives by finding the virtual address of
2171 * each, and then choose a working drive that also has that virt address.
2172 * There is a separate r10_bio for each non-in_sync drive.
2173 * Only the first two slots are in use. The first for reading,
2174 * The second for writing.
2175 *
2176 */
2177static void fix_recovery_read_error(struct r10bio *r10_bio)
2178{
2179 /* We got a read error during recovery.
2180 * We repeat the read in smaller page-sized sections.
2181 * If a read succeeds, write it to the new device or record
2182 * a bad block if we cannot.
2183 * If a read fails, record a bad block on both old and
2184 * new devices.
2185 */
2186 struct mddev *mddev = r10_bio->mddev;
2187 struct r10conf *conf = mddev->private;
2188 struct bio *bio = r10_bio->devs[0].bio;
2189 sector_t sect = 0;
2190 int sectors = r10_bio->sectors;
2191 int idx = 0;
2192 int dr = r10_bio->devs[0].devnum;
2193 int dw = r10_bio->devs[1].devnum;
2194
2195 while (sectors) {
2196 int s = sectors;
2197 struct md_rdev *rdev;
2198 sector_t addr;
2199 int ok;
2200
2201 if (s > (PAGE_SIZE>>9))
2202 s = PAGE_SIZE >> 9;
2203
2204 rdev = conf->mirrors[dr].rdev;
2205 addr = r10_bio->devs[0].addr + sect,
2206 ok = sync_page_io(rdev,
2207 addr,
2208 s << 9,
2209 bio->bi_io_vec[idx].bv_page,
2210 READ, false);
2211 if (ok) {
2212 rdev = conf->mirrors[dw].rdev;
2213 addr = r10_bio->devs[1].addr + sect;
2214 ok = sync_page_io(rdev,
2215 addr,
2216 s << 9,
2217 bio->bi_io_vec[idx].bv_page,
2218 WRITE, false);
2219 if (!ok) {
2220 set_bit(WriteErrorSeen, &rdev->flags);
2221 if (!test_and_set_bit(WantReplacement,
2222 &rdev->flags))
2223 set_bit(MD_RECOVERY_NEEDED,
2224 &rdev->mddev->recovery);
2225 }
2226 }
2227 if (!ok) {
2228 /* We don't worry if we cannot set a bad block -
2229 * it really is bad so there is no loss in not
2230 * recording it yet
2231 */
2232 rdev_set_badblocks(rdev, addr, s, 0);
2233
2234 if (rdev != conf->mirrors[dw].rdev) {
2235 /* need bad block on destination too */
2236 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2237 addr = r10_bio->devs[1].addr + sect;
2238 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2239 if (!ok) {
2240 /* just abort the recovery */
2241 printk(KERN_NOTICE
2242 "md/raid10:%s: recovery aborted"
2243 " due to read error\n",
2244 mdname(mddev));
2245
2246 conf->mirrors[dw].recovery_disabled
2247 = mddev->recovery_disabled;
2248 set_bit(MD_RECOVERY_INTR,
2249 &mddev->recovery);
2250 break;
2251 }
2252 }
2253 }
2254
2255 sectors -= s;
2256 sect += s;
2257 idx++;
2258 }
2259}
2260
2261static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2262{
2263 struct r10conf *conf = mddev->private;
2264 int d;
2265 struct bio *wbio, *wbio2;
2266
2267 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2268 fix_recovery_read_error(r10_bio);
2269 end_sync_request(r10_bio);
2270 return;
2271 }
2272
2273 /*
2274 * share the pages with the first bio
2275 * and submit the write request
2276 */
2277 d = r10_bio->devs[1].devnum;
2278 wbio = r10_bio->devs[1].bio;
2279 wbio2 = r10_bio->devs[1].repl_bio;
2280 /* Need to test wbio2->bi_end_io before we call
2281 * generic_make_request as if the former is NULL,
2282 * the latter is free to free wbio2.
2283 */
2284 if (wbio2 && !wbio2->bi_end_io)
2285 wbio2 = NULL;
2286 if (wbio->bi_end_io) {
2287 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2288 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2289 generic_make_request(wbio);
2290 }
2291 if (wbio2) {
2292 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2293 md_sync_acct(conf->mirrors[d].replacement->bdev,
2294 bio_sectors(wbio2));
2295 generic_make_request(wbio2);
2296 }
2297}
2298
2299
2300/*
2301 * Used by fix_read_error() to decay the per rdev read_errors.
2302 * We halve the read error count for every hour that has elapsed
2303 * since the last recorded read error.
2304 *
2305 */
2306static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2307{
2308 struct timespec cur_time_mon;
2309 unsigned long hours_since_last;
2310 unsigned int read_errors = atomic_read(&rdev->read_errors);
2311
2312 ktime_get_ts(&cur_time_mon);
2313
2314 if (rdev->last_read_error.tv_sec == 0 &&
2315 rdev->last_read_error.tv_nsec == 0) {
2316 /* first time we've seen a read error */
2317 rdev->last_read_error = cur_time_mon;
2318 return;
2319 }
2320
2321 hours_since_last = (cur_time_mon.tv_sec -
2322 rdev->last_read_error.tv_sec) / 3600;
2323
2324 rdev->last_read_error = cur_time_mon;
2325
2326 /*
2327 * if hours_since_last is > the number of bits in read_errors
2328 * just set read errors to 0. We do this to avoid
2329 * overflowing the shift of read_errors by hours_since_last.
2330 */
2331 if (hours_since_last >= 8 * sizeof(read_errors))
2332 atomic_set(&rdev->read_errors, 0);
2333 else
2334 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2335}
2336
2337static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2338 int sectors, struct page *page, int rw)
2339{
2340 sector_t first_bad;
2341 int bad_sectors;
2342
2343 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2344 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2345 return -1;
2346 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2347 /* success */
2348 return 1;
2349 if (rw == WRITE) {
2350 set_bit(WriteErrorSeen, &rdev->flags);
2351 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2352 set_bit(MD_RECOVERY_NEEDED,
2353 &rdev->mddev->recovery);
2354 }
2355 /* need to record an error - either for the block or the device */
2356 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2357 md_error(rdev->mddev, rdev);
2358 return 0;
2359}
2360
2361/*
2362 * This is a kernel thread which:
2363 *
2364 * 1. Retries failed read operations on working mirrors.
2365 * 2. Updates the raid superblock when problems encounter.
2366 * 3. Performs writes following reads for array synchronising.
2367 */
2368
2369static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2370{
2371 int sect = 0; /* Offset from r10_bio->sector */
2372 int sectors = r10_bio->sectors;
2373 struct md_rdev*rdev;
2374 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2375 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2376
2377 /* still own a reference to this rdev, so it cannot
2378 * have been cleared recently.
2379 */
2380 rdev = conf->mirrors[d].rdev;
2381
2382 if (test_bit(Faulty, &rdev->flags))
2383 /* drive has already been failed, just ignore any
2384 more fix_read_error() attempts */
2385 return;
2386
2387 check_decay_read_errors(mddev, rdev);
2388 atomic_inc(&rdev->read_errors);
2389 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2390 char b[BDEVNAME_SIZE];
2391 bdevname(rdev->bdev, b);
2392
2393 printk(KERN_NOTICE
2394 "md/raid10:%s: %s: Raid device exceeded "
2395 "read_error threshold [cur %d:max %d]\n",
2396 mdname(mddev), b,
2397 atomic_read(&rdev->read_errors), max_read_errors);
2398 printk(KERN_NOTICE
2399 "md/raid10:%s: %s: Failing raid device\n",
2400 mdname(mddev), b);
2401 md_error(mddev, conf->mirrors[d].rdev);
2402 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2403 return;
2404 }
2405
2406 while(sectors) {
2407 int s = sectors;
2408 int sl = r10_bio->read_slot;
2409 int success = 0;
2410 int start;
2411
2412 if (s > (PAGE_SIZE>>9))
2413 s = PAGE_SIZE >> 9;
2414
2415 rcu_read_lock();
2416 do {
2417 sector_t first_bad;
2418 int bad_sectors;
2419
2420 d = r10_bio->devs[sl].devnum;
2421 rdev = rcu_dereference(conf->mirrors[d].rdev);
2422 if (rdev &&
2423 !test_bit(Unmerged, &rdev->flags) &&
2424 test_bit(In_sync, &rdev->flags) &&
2425 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2426 &first_bad, &bad_sectors) == 0) {
2427 atomic_inc(&rdev->nr_pending);
2428 rcu_read_unlock();
2429 success = sync_page_io(rdev,
2430 r10_bio->devs[sl].addr +
2431 sect,
2432 s<<9,
2433 conf->tmppage, READ, false);
2434 rdev_dec_pending(rdev, mddev);
2435 rcu_read_lock();
2436 if (success)
2437 break;
2438 }
2439 sl++;
2440 if (sl == conf->copies)
2441 sl = 0;
2442 } while (!success && sl != r10_bio->read_slot);
2443 rcu_read_unlock();
2444
2445 if (!success) {
2446 /* Cannot read from anywhere, just mark the block
2447 * as bad on the first device to discourage future
2448 * reads.
2449 */
2450 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2451 rdev = conf->mirrors[dn].rdev;
2452
2453 if (!rdev_set_badblocks(
2454 rdev,
2455 r10_bio->devs[r10_bio->read_slot].addr
2456 + sect,
2457 s, 0)) {
2458 md_error(mddev, rdev);
2459 r10_bio->devs[r10_bio->read_slot].bio
2460 = IO_BLOCKED;
2461 }
2462 break;
2463 }
2464
2465 start = sl;
2466 /* write it back and re-read */
2467 rcu_read_lock();
2468 while (sl != r10_bio->read_slot) {
2469 char b[BDEVNAME_SIZE];
2470
2471 if (sl==0)
2472 sl = conf->copies;
2473 sl--;
2474 d = r10_bio->devs[sl].devnum;
2475 rdev = rcu_dereference(conf->mirrors[d].rdev);
2476 if (!rdev ||
2477 test_bit(Unmerged, &rdev->flags) ||
2478 !test_bit(In_sync, &rdev->flags))
2479 continue;
2480
2481 atomic_inc(&rdev->nr_pending);
2482 rcu_read_unlock();
2483 if (r10_sync_page_io(rdev,
2484 r10_bio->devs[sl].addr +
2485 sect,
2486 s, conf->tmppage, WRITE)
2487 == 0) {
2488 /* Well, this device is dead */
2489 printk(KERN_NOTICE
2490 "md/raid10:%s: read correction "
2491 "write failed"
2492 " (%d sectors at %llu on %s)\n",
2493 mdname(mddev), s,
2494 (unsigned long long)(
2495 sect +
2496 choose_data_offset(r10_bio,
2497 rdev)),
2498 bdevname(rdev->bdev, b));
2499 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2500 "drive\n",
2501 mdname(mddev),
2502 bdevname(rdev->bdev, b));
2503 }
2504 rdev_dec_pending(rdev, mddev);
2505 rcu_read_lock();
2506 }
2507 sl = start;
2508 while (sl != r10_bio->read_slot) {
2509 char b[BDEVNAME_SIZE];
2510
2511 if (sl==0)
2512 sl = conf->copies;
2513 sl--;
2514 d = r10_bio->devs[sl].devnum;
2515 rdev = rcu_dereference(conf->mirrors[d].rdev);
2516 if (!rdev ||
2517 !test_bit(In_sync, &rdev->flags))
2518 continue;
2519
2520 atomic_inc(&rdev->nr_pending);
2521 rcu_read_unlock();
2522 switch (r10_sync_page_io(rdev,
2523 r10_bio->devs[sl].addr +
2524 sect,
2525 s, conf->tmppage,
2526 READ)) {
2527 case 0:
2528 /* Well, this device is dead */
2529 printk(KERN_NOTICE
2530 "md/raid10:%s: unable to read back "
2531 "corrected sectors"
2532 " (%d sectors at %llu on %s)\n",
2533 mdname(mddev), s,
2534 (unsigned long long)(
2535 sect +
2536 choose_data_offset(r10_bio, rdev)),
2537 bdevname(rdev->bdev, b));
2538 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2539 "drive\n",
2540 mdname(mddev),
2541 bdevname(rdev->bdev, b));
2542 break;
2543 case 1:
2544 printk(KERN_INFO
2545 "md/raid10:%s: read error corrected"
2546 " (%d sectors at %llu on %s)\n",
2547 mdname(mddev), s,
2548 (unsigned long long)(
2549 sect +
2550 choose_data_offset(r10_bio, rdev)),
2551 bdevname(rdev->bdev, b));
2552 atomic_add(s, &rdev->corrected_errors);
2553 }
2554
2555 rdev_dec_pending(rdev, mddev);
2556 rcu_read_lock();
2557 }
2558 rcu_read_unlock();
2559
2560 sectors -= s;
2561 sect += s;
2562 }
2563}
2564
2565static int narrow_write_error(struct r10bio *r10_bio, int i)
2566{
2567 struct bio *bio = r10_bio->master_bio;
2568 struct mddev *mddev = r10_bio->mddev;
2569 struct r10conf *conf = mddev->private;
2570 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2571 /* bio has the data to be written to slot 'i' where
2572 * we just recently had a write error.
2573 * We repeatedly clone the bio and trim down to one block,
2574 * then try the write. Where the write fails we record
2575 * a bad block.
2576 * It is conceivable that the bio doesn't exactly align with
2577 * blocks. We must handle this.
2578 *
2579 * We currently own a reference to the rdev.
2580 */
2581
2582 int block_sectors;
2583 sector_t sector;
2584 int sectors;
2585 int sect_to_write = r10_bio->sectors;
2586 int ok = 1;
2587
2588 if (rdev->badblocks.shift < 0)
2589 return 0;
2590
2591 block_sectors = 1 << rdev->badblocks.shift;
2592 sector = r10_bio->sector;
2593 sectors = ((r10_bio->sector + block_sectors)
2594 & ~(sector_t)(block_sectors - 1))
2595 - sector;
2596
2597 while (sect_to_write) {
2598 struct bio *wbio;
2599 if (sectors > sect_to_write)
2600 sectors = sect_to_write;
2601 /* Write at 'sector' for 'sectors' */
2602 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2603 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2604 wbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+
2605 choose_data_offset(r10_bio, rdev) +
2606 (sector - r10_bio->sector));
2607 wbio->bi_bdev = rdev->bdev;
2608 if (submit_bio_wait(WRITE, wbio) == 0)
2609 /* Failure! */
2610 ok = rdev_set_badblocks(rdev, sector,
2611 sectors, 0)
2612 && ok;
2613
2614 bio_put(wbio);
2615 sect_to_write -= sectors;
2616 sector += sectors;
2617 sectors = block_sectors;
2618 }
2619 return ok;
2620}
2621
2622static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2623{
2624 int slot = r10_bio->read_slot;
2625 struct bio *bio;
2626 struct r10conf *conf = mddev->private;
2627 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2628 char b[BDEVNAME_SIZE];
2629 unsigned long do_sync;
2630 int max_sectors;
2631
2632 /* we got a read error. Maybe the drive is bad. Maybe just
2633 * the block and we can fix it.
2634 * We freeze all other IO, and try reading the block from
2635 * other devices. When we find one, we re-write
2636 * and check it that fixes the read error.
2637 * This is all done synchronously while the array is
2638 * frozen.
2639 */
2640 bio = r10_bio->devs[slot].bio;
2641 bdevname(bio->bi_bdev, b);
2642 bio_put(bio);
2643 r10_bio->devs[slot].bio = NULL;
2644
2645 if (mddev->ro == 0) {
2646 freeze_array(conf, 1);
2647 fix_read_error(conf, mddev, r10_bio);
2648 unfreeze_array(conf);
2649 } else
2650 r10_bio->devs[slot].bio = IO_BLOCKED;
2651
2652 rdev_dec_pending(rdev, mddev);
2653
2654read_more:
2655 rdev = read_balance(conf, r10_bio, &max_sectors);
2656 if (rdev == NULL) {
2657 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2658 " read error for block %llu\n",
2659 mdname(mddev), b,
2660 (unsigned long long)r10_bio->sector);
2661 raid_end_bio_io(r10_bio);
2662 return;
2663 }
2664
2665 do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2666 slot = r10_bio->read_slot;
2667 printk_ratelimited(
2668 KERN_ERR
2669 "md/raid10:%s: %s: redirecting "
2670 "sector %llu to another mirror\n",
2671 mdname(mddev),
2672 bdevname(rdev->bdev, b),
2673 (unsigned long long)r10_bio->sector);
2674 bio = bio_clone_mddev(r10_bio->master_bio,
2675 GFP_NOIO, mddev);
2676 bio_trim(bio, r10_bio->sector - bio->bi_iter.bi_sector, max_sectors);
2677 r10_bio->devs[slot].bio = bio;
2678 r10_bio->devs[slot].rdev = rdev;
2679 bio->bi_iter.bi_sector = r10_bio->devs[slot].addr
2680 + choose_data_offset(r10_bio, rdev);
2681 bio->bi_bdev = rdev->bdev;
2682 bio->bi_rw = READ | do_sync;
2683 bio->bi_private = r10_bio;
2684 bio->bi_end_io = raid10_end_read_request;
2685 if (max_sectors < r10_bio->sectors) {
2686 /* Drat - have to split this up more */
2687 struct bio *mbio = r10_bio->master_bio;
2688 int sectors_handled =
2689 r10_bio->sector + max_sectors
2690 - mbio->bi_iter.bi_sector;
2691 r10_bio->sectors = max_sectors;
2692 spin_lock_irq(&conf->device_lock);
2693 if (mbio->bi_phys_segments == 0)
2694 mbio->bi_phys_segments = 2;
2695 else
2696 mbio->bi_phys_segments++;
2697 spin_unlock_irq(&conf->device_lock);
2698 generic_make_request(bio);
2699
2700 r10_bio = mempool_alloc(conf->r10bio_pool,
2701 GFP_NOIO);
2702 r10_bio->master_bio = mbio;
2703 r10_bio->sectors = bio_sectors(mbio) - sectors_handled;
2704 r10_bio->state = 0;
2705 set_bit(R10BIO_ReadError,
2706 &r10_bio->state);
2707 r10_bio->mddev = mddev;
2708 r10_bio->sector = mbio->bi_iter.bi_sector
2709 + sectors_handled;
2710
2711 goto read_more;
2712 } else
2713 generic_make_request(bio);
2714}
2715
2716static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2717{
2718 /* Some sort of write request has finished and it
2719 * succeeded in writing where we thought there was a
2720 * bad block. So forget the bad block.
2721 * Or possibly if failed and we need to record
2722 * a bad block.
2723 */
2724 int m;
2725 struct md_rdev *rdev;
2726
2727 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2728 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2729 for (m = 0; m < conf->copies; m++) {
2730 int dev = r10_bio->devs[m].devnum;
2731 rdev = conf->mirrors[dev].rdev;
2732 if (r10_bio->devs[m].bio == NULL)
2733 continue;
2734 if (test_bit(BIO_UPTODATE,
2735 &r10_bio->devs[m].bio->bi_flags)) {
2736 rdev_clear_badblocks(
2737 rdev,
2738 r10_bio->devs[m].addr,
2739 r10_bio->sectors, 0);
2740 } else {
2741 if (!rdev_set_badblocks(
2742 rdev,
2743 r10_bio->devs[m].addr,
2744 r10_bio->sectors, 0))
2745 md_error(conf->mddev, rdev);
2746 }
2747 rdev = conf->mirrors[dev].replacement;
2748 if (r10_bio->devs[m].repl_bio == NULL)
2749 continue;
2750 if (test_bit(BIO_UPTODATE,
2751 &r10_bio->devs[m].repl_bio->bi_flags)) {
2752 rdev_clear_badblocks(
2753 rdev,
2754 r10_bio->devs[m].addr,
2755 r10_bio->sectors, 0);
2756 } else {
2757 if (!rdev_set_badblocks(
2758 rdev,
2759 r10_bio->devs[m].addr,
2760 r10_bio->sectors, 0))
2761 md_error(conf->mddev, rdev);
2762 }
2763 }
2764 put_buf(r10_bio);
2765 } else {
2766 for (m = 0; m < conf->copies; m++) {
2767 int dev = r10_bio->devs[m].devnum;
2768 struct bio *bio = r10_bio->devs[m].bio;
2769 rdev = conf->mirrors[dev].rdev;
2770 if (bio == IO_MADE_GOOD) {
2771 rdev_clear_badblocks(
2772 rdev,
2773 r10_bio->devs[m].addr,
2774 r10_bio->sectors, 0);
2775 rdev_dec_pending(rdev, conf->mddev);
2776 } else if (bio != NULL &&
2777 !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2778 if (!narrow_write_error(r10_bio, m)) {
2779 md_error(conf->mddev, rdev);
2780 set_bit(R10BIO_Degraded,
2781 &r10_bio->state);
2782 }
2783 rdev_dec_pending(rdev, conf->mddev);
2784 }
2785 bio = r10_bio->devs[m].repl_bio;
2786 rdev = conf->mirrors[dev].replacement;
2787 if (rdev && bio == IO_MADE_GOOD) {
2788 rdev_clear_badblocks(
2789 rdev,
2790 r10_bio->devs[m].addr,
2791 r10_bio->sectors, 0);
2792 rdev_dec_pending(rdev, conf->mddev);
2793 }
2794 }
2795 if (test_bit(R10BIO_WriteError,
2796 &r10_bio->state))
2797 close_write(r10_bio);
2798 raid_end_bio_io(r10_bio);
2799 }
2800}
2801
2802static void raid10d(struct md_thread *thread)
2803{
2804 struct mddev *mddev = thread->mddev;
2805 struct r10bio *r10_bio;
2806 unsigned long flags;
2807 struct r10conf *conf = mddev->private;
2808 struct list_head *head = &conf->retry_list;
2809 struct blk_plug plug;
2810
2811 md_check_recovery(mddev);
2812
2813 blk_start_plug(&plug);
2814 for (;;) {
2815
2816 flush_pending_writes(conf);
2817
2818 spin_lock_irqsave(&conf->device_lock, flags);
2819 if (list_empty(head)) {
2820 spin_unlock_irqrestore(&conf->device_lock, flags);
2821 break;
2822 }
2823 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2824 list_del(head->prev);
2825 conf->nr_queued--;
2826 spin_unlock_irqrestore(&conf->device_lock, flags);
2827
2828 mddev = r10_bio->mddev;
2829 conf = mddev->private;
2830 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2831 test_bit(R10BIO_WriteError, &r10_bio->state))
2832 handle_write_completed(conf, r10_bio);
2833 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2834 reshape_request_write(mddev, r10_bio);
2835 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2836 sync_request_write(mddev, r10_bio);
2837 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2838 recovery_request_write(mddev, r10_bio);
2839 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2840 handle_read_error(mddev, r10_bio);
2841 else {
2842 /* just a partial read to be scheduled from a
2843 * separate context
2844 */
2845 int slot = r10_bio->read_slot;
2846 generic_make_request(r10_bio->devs[slot].bio);
2847 }
2848
2849 cond_resched();
2850 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2851 md_check_recovery(mddev);
2852 }
2853 blk_finish_plug(&plug);
2854}
2855
2856
2857static int init_resync(struct r10conf *conf)
2858{
2859 int buffs;
2860 int i;
2861
2862 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2863 BUG_ON(conf->r10buf_pool);
2864 conf->have_replacement = 0;
2865 for (i = 0; i < conf->geo.raid_disks; i++)
2866 if (conf->mirrors[i].replacement)
2867 conf->have_replacement = 1;
2868 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2869 if (!conf->r10buf_pool)
2870 return -ENOMEM;
2871 conf->next_resync = 0;
2872 return 0;
2873}
2874
2875/*
2876 * perform a "sync" on one "block"
2877 *
2878 * We need to make sure that no normal I/O request - particularly write
2879 * requests - conflict with active sync requests.
2880 *
2881 * This is achieved by tracking pending requests and a 'barrier' concept
2882 * that can be installed to exclude normal IO requests.
2883 *
2884 * Resync and recovery are handled very differently.
2885 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2886 *
2887 * For resync, we iterate over virtual addresses, read all copies,
2888 * and update if there are differences. If only one copy is live,
2889 * skip it.
2890 * For recovery, we iterate over physical addresses, read a good
2891 * value for each non-in_sync drive, and over-write.
2892 *
2893 * So, for recovery we may have several outstanding complex requests for a
2894 * given address, one for each out-of-sync device. We model this by allocating
2895 * a number of r10_bio structures, one for each out-of-sync device.
2896 * As we setup these structures, we collect all bio's together into a list
2897 * which we then process collectively to add pages, and then process again
2898 * to pass to generic_make_request.
2899 *
2900 * The r10_bio structures are linked using a borrowed master_bio pointer.
2901 * This link is counted in ->remaining. When the r10_bio that points to NULL
2902 * has its remaining count decremented to 0, the whole complex operation
2903 * is complete.
2904 *
2905 */
2906
2907static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
2908 int *skipped, int go_faster)
2909{
2910 struct r10conf *conf = mddev->private;
2911 struct r10bio *r10_bio;
2912 struct bio *biolist = NULL, *bio;
2913 sector_t max_sector, nr_sectors;
2914 int i;
2915 int max_sync;
2916 sector_t sync_blocks;
2917 sector_t sectors_skipped = 0;
2918 int chunks_skipped = 0;
2919 sector_t chunk_mask = conf->geo.chunk_mask;
2920
2921 if (!conf->r10buf_pool)
2922 if (init_resync(conf))
2923 return 0;
2924
2925 /*
2926 * Allow skipping a full rebuild for incremental assembly
2927 * of a clean array, like RAID1 does.
2928 */
2929 if (mddev->bitmap == NULL &&
2930 mddev->recovery_cp == MaxSector &&
2931 mddev->reshape_position == MaxSector &&
2932 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2933 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2934 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2935 conf->fullsync == 0) {
2936 *skipped = 1;
2937 return mddev->dev_sectors - sector_nr;
2938 }
2939
2940 skipped:
2941 max_sector = mddev->dev_sectors;
2942 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2943 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2944 max_sector = mddev->resync_max_sectors;
2945 if (sector_nr >= max_sector) {
2946 /* If we aborted, we need to abort the
2947 * sync on the 'current' bitmap chucks (there can
2948 * be several when recovering multiple devices).
2949 * as we may have started syncing it but not finished.
2950 * We can find the current address in
2951 * mddev->curr_resync, but for recovery,
2952 * we need to convert that to several
2953 * virtual addresses.
2954 */
2955 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2956 end_reshape(conf);
2957 return 0;
2958 }
2959
2960 if (mddev->curr_resync < max_sector) { /* aborted */
2961 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2962 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2963 &sync_blocks, 1);
2964 else for (i = 0; i < conf->geo.raid_disks; i++) {
2965 sector_t sect =
2966 raid10_find_virt(conf, mddev->curr_resync, i);
2967 bitmap_end_sync(mddev->bitmap, sect,
2968 &sync_blocks, 1);
2969 }
2970 } else {
2971 /* completed sync */
2972 if ((!mddev->bitmap || conf->fullsync)
2973 && conf->have_replacement
2974 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2975 /* Completed a full sync so the replacements
2976 * are now fully recovered.
2977 */
2978 for (i = 0; i < conf->geo.raid_disks; i++)
2979 if (conf->mirrors[i].replacement)
2980 conf->mirrors[i].replacement
2981 ->recovery_offset
2982 = MaxSector;
2983 }
2984 conf->fullsync = 0;
2985 }
2986 bitmap_close_sync(mddev->bitmap);
2987 close_sync(conf);
2988 *skipped = 1;
2989 return sectors_skipped;
2990 }
2991
2992 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2993 return reshape_request(mddev, sector_nr, skipped);
2994
2995 if (chunks_skipped >= conf->geo.raid_disks) {
2996 /* if there has been nothing to do on any drive,
2997 * then there is nothing to do at all..
2998 */
2999 *skipped = 1;
3000 return (max_sector - sector_nr) + sectors_skipped;
3001 }
3002
3003 if (max_sector > mddev->resync_max)
3004 max_sector = mddev->resync_max; /* Don't do IO beyond here */
3005
3006 /* make sure whole request will fit in a chunk - if chunks
3007 * are meaningful
3008 */
3009 if (conf->geo.near_copies < conf->geo.raid_disks &&
3010 max_sector > (sector_nr | chunk_mask))
3011 max_sector = (sector_nr | chunk_mask) + 1;
3012 /*
3013 * If there is non-resync activity waiting for us then
3014 * put in a delay to throttle resync.
3015 */
3016 if (!go_faster && conf->nr_waiting)
3017 msleep_interruptible(1000);
3018
3019 /* Again, very different code for resync and recovery.
3020 * Both must result in an r10bio with a list of bios that
3021 * have bi_end_io, bi_sector, bi_bdev set,
3022 * and bi_private set to the r10bio.
3023 * For recovery, we may actually create several r10bios
3024 * with 2 bios in each, that correspond to the bios in the main one.
3025 * In this case, the subordinate r10bios link back through a
3026 * borrowed master_bio pointer, and the counter in the master
3027 * includes a ref from each subordinate.
3028 */
3029 /* First, we decide what to do and set ->bi_end_io
3030 * To end_sync_read if we want to read, and
3031 * end_sync_write if we will want to write.
3032 */
3033
3034 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3035 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3036 /* recovery... the complicated one */
3037 int j;
3038 r10_bio = NULL;
3039
3040 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3041 int still_degraded;
3042 struct r10bio *rb2;
3043 sector_t sect;
3044 int must_sync;
3045 int any_working;
3046 struct raid10_info *mirror = &conf->mirrors[i];
3047
3048 if ((mirror->rdev == NULL ||
3049 test_bit(In_sync, &mirror->rdev->flags))
3050 &&
3051 (mirror->replacement == NULL ||
3052 test_bit(Faulty,
3053 &mirror->replacement->flags)))
3054 continue;
3055
3056 still_degraded = 0;
3057 /* want to reconstruct this device */
3058 rb2 = r10_bio;
3059 sect = raid10_find_virt(conf, sector_nr, i);
3060 if (sect >= mddev->resync_max_sectors) {
3061 /* last stripe is not complete - don't
3062 * try to recover this sector.
3063 */
3064 continue;
3065 }
3066 /* Unless we are doing a full sync, or a replacement
3067 * we only need to recover the block if it is set in
3068 * the bitmap
3069 */
3070 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3071 &sync_blocks, 1);
3072 if (sync_blocks < max_sync)
3073 max_sync = sync_blocks;
3074 if (!must_sync &&
3075 mirror->replacement == NULL &&
3076 !conf->fullsync) {
3077 /* yep, skip the sync_blocks here, but don't assume
3078 * that there will never be anything to do here
3079 */
3080 chunks_skipped = -1;
3081 continue;
3082 }
3083
3084 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3085 raise_barrier(conf, rb2 != NULL);
3086 atomic_set(&r10_bio->remaining, 0);
3087
3088 r10_bio->master_bio = (struct bio*)rb2;
3089 if (rb2)
3090 atomic_inc(&rb2->remaining);
3091 r10_bio->mddev = mddev;
3092 set_bit(R10BIO_IsRecover, &r10_bio->state);
3093 r10_bio->sector = sect;
3094
3095 raid10_find_phys(conf, r10_bio);
3096
3097 /* Need to check if the array will still be
3098 * degraded
3099 */
3100 for (j = 0; j < conf->geo.raid_disks; j++)
3101 if (conf->mirrors[j].rdev == NULL ||
3102 test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
3103 still_degraded = 1;
3104 break;
3105 }
3106
3107 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3108 &sync_blocks, still_degraded);
3109
3110 any_working = 0;
3111 for (j=0; j<conf->copies;j++) {
3112 int k;
3113 int d = r10_bio->devs[j].devnum;
3114 sector_t from_addr, to_addr;
3115 struct md_rdev *rdev;
3116 sector_t sector, first_bad;
3117 int bad_sectors;
3118 if (!conf->mirrors[d].rdev ||
3119 !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
3120 continue;
3121 /* This is where we read from */
3122 any_working = 1;
3123 rdev = conf->mirrors[d].rdev;
3124 sector = r10_bio->devs[j].addr;
3125
3126 if (is_badblock(rdev, sector, max_sync,
3127 &first_bad, &bad_sectors)) {
3128 if (first_bad > sector)
3129 max_sync = first_bad - sector;
3130 else {
3131 bad_sectors -= (sector
3132 - first_bad);
3133 if (max_sync > bad_sectors)
3134 max_sync = bad_sectors;
3135 continue;
3136 }
3137 }
3138 bio = r10_bio->devs[0].bio;
3139 bio_reset(bio);
3140 bio->bi_next = biolist;
3141 biolist = bio;
3142 bio->bi_private = r10_bio;
3143 bio->bi_end_io = end_sync_read;
3144 bio->bi_rw = READ;
3145 from_addr = r10_bio->devs[j].addr;
3146 bio->bi_iter.bi_sector = from_addr +
3147 rdev->data_offset;
3148 bio->bi_bdev = rdev->bdev;
3149 atomic_inc(&rdev->nr_pending);
3150 /* and we write to 'i' (if not in_sync) */
3151
3152 for (k=0; k<conf->copies; k++)
3153 if (r10_bio->devs[k].devnum == i)
3154 break;
3155 BUG_ON(k == conf->copies);
3156 to_addr = r10_bio->devs[k].addr;
3157 r10_bio->devs[0].devnum = d;
3158 r10_bio->devs[0].addr = from_addr;
3159 r10_bio->devs[1].devnum = i;
3160 r10_bio->devs[1].addr = to_addr;
3161
3162 rdev = mirror->rdev;
3163 if (!test_bit(In_sync, &rdev->flags)) {
3164 bio = r10_bio->devs[1].bio;
3165 bio_reset(bio);
3166 bio->bi_next = biolist;
3167 biolist = bio;
3168 bio->bi_private = r10_bio;
3169 bio->bi_end_io = end_sync_write;
3170 bio->bi_rw = WRITE;
3171 bio->bi_iter.bi_sector = to_addr
3172 + rdev->data_offset;
3173 bio->bi_bdev = rdev->bdev;
3174 atomic_inc(&r10_bio->remaining);
3175 } else
3176 r10_bio->devs[1].bio->bi_end_io = NULL;
3177
3178 /* and maybe write to replacement */
3179 bio = r10_bio->devs[1].repl_bio;
3180 if (bio)
3181 bio->bi_end_io = NULL;
3182 rdev = mirror->replacement;
3183 /* Note: if rdev != NULL, then bio
3184 * cannot be NULL as r10buf_pool_alloc will
3185 * have allocated it.
3186 * So the second test here is pointless.
3187 * But it keeps semantic-checkers happy, and
3188 * this comment keeps human reviewers
3189 * happy.
3190 */
3191 if (rdev == NULL || bio == NULL ||
3192 test_bit(Faulty, &rdev->flags))
3193 break;
3194 bio_reset(bio);
3195 bio->bi_next = biolist;
3196 biolist = bio;
3197 bio->bi_private = r10_bio;
3198 bio->bi_end_io = end_sync_write;
3199 bio->bi_rw = WRITE;
3200 bio->bi_iter.bi_sector = to_addr +
3201 rdev->data_offset;
3202 bio->bi_bdev = rdev->bdev;
3203 atomic_inc(&r10_bio->remaining);
3204 break;
3205 }
3206 if (j == conf->copies) {
3207 /* Cannot recover, so abort the recovery or
3208 * record a bad block */
3209 if (any_working) {
3210 /* problem is that there are bad blocks
3211 * on other device(s)
3212 */
3213 int k;
3214 for (k = 0; k < conf->copies; k++)
3215 if (r10_bio->devs[k].devnum == i)
3216 break;
3217 if (!test_bit(In_sync,
3218 &mirror->rdev->flags)
3219 && !rdev_set_badblocks(
3220 mirror->rdev,
3221 r10_bio->devs[k].addr,
3222 max_sync, 0))
3223 any_working = 0;
3224 if (mirror->replacement &&
3225 !rdev_set_badblocks(
3226 mirror->replacement,
3227 r10_bio->devs[k].addr,
3228 max_sync, 0))
3229 any_working = 0;
3230 }
3231 if (!any_working) {
3232 if (!test_and_set_bit(MD_RECOVERY_INTR,
3233 &mddev->recovery))
3234 printk(KERN_INFO "md/raid10:%s: insufficient "
3235 "working devices for recovery.\n",
3236 mdname(mddev));
3237 mirror->recovery_disabled
3238 = mddev->recovery_disabled;
3239 }
3240 put_buf(r10_bio);
3241 if (rb2)
3242 atomic_dec(&rb2->remaining);
3243 r10_bio = rb2;
3244 break;
3245 }
3246 }
3247 if (biolist == NULL) {
3248 while (r10_bio) {
3249 struct r10bio *rb2 = r10_bio;
3250 r10_bio = (struct r10bio*) rb2->master_bio;
3251 rb2->master_bio = NULL;
3252 put_buf(rb2);
3253 }
3254 goto giveup;
3255 }
3256 } else {
3257 /* resync. Schedule a read for every block at this virt offset */
3258 int count = 0;
3259
3260 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3261
3262 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3263 &sync_blocks, mddev->degraded) &&
3264 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3265 &mddev->recovery)) {
3266 /* We can skip this block */
3267 *skipped = 1;
3268 return sync_blocks + sectors_skipped;
3269 }
3270 if (sync_blocks < max_sync)
3271 max_sync = sync_blocks;
3272 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3273
3274 r10_bio->mddev = mddev;
3275 atomic_set(&r10_bio->remaining, 0);
3276 raise_barrier(conf, 0);
3277 conf->next_resync = sector_nr;
3278
3279 r10_bio->master_bio = NULL;
3280 r10_bio->sector = sector_nr;
3281 set_bit(R10BIO_IsSync, &r10_bio->state);
3282 raid10_find_phys(conf, r10_bio);
3283 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3284
3285 for (i = 0; i < conf->copies; i++) {
3286 int d = r10_bio->devs[i].devnum;
3287 sector_t first_bad, sector;
3288 int bad_sectors;
3289
3290 if (r10_bio->devs[i].repl_bio)
3291 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3292
3293 bio = r10_bio->devs[i].bio;
3294 bio_reset(bio);
3295 clear_bit(BIO_UPTODATE, &bio->bi_flags);
3296 if (conf->mirrors[d].rdev == NULL ||
3297 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
3298 continue;
3299 sector = r10_bio->devs[i].addr;
3300 if (is_badblock(conf->mirrors[d].rdev,
3301 sector, max_sync,
3302 &first_bad, &bad_sectors)) {
3303 if (first_bad > sector)
3304 max_sync = first_bad - sector;
3305 else {
3306 bad_sectors -= (sector - first_bad);
3307 if (max_sync > bad_sectors)
3308 max_sync = bad_sectors;
3309 continue;
3310 }
3311 }
3312 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3313 atomic_inc(&r10_bio->remaining);
3314 bio->bi_next = biolist;
3315 biolist = bio;
3316 bio->bi_private = r10_bio;
3317 bio->bi_end_io = end_sync_read;
3318 bio->bi_rw = READ;
3319 bio->bi_iter.bi_sector = sector +
3320 conf->mirrors[d].rdev->data_offset;
3321 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
3322 count++;
3323
3324 if (conf->mirrors[d].replacement == NULL ||
3325 test_bit(Faulty,
3326 &conf->mirrors[d].replacement->flags))
3327 continue;
3328
3329 /* Need to set up for writing to the replacement */
3330 bio = r10_bio->devs[i].repl_bio;
3331 bio_reset(bio);
3332 clear_bit(BIO_UPTODATE, &bio->bi_flags);
3333
3334 sector = r10_bio->devs[i].addr;
3335 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3336 bio->bi_next = biolist;
3337 biolist = bio;
3338 bio->bi_private = r10_bio;
3339 bio->bi_end_io = end_sync_write;
3340 bio->bi_rw = WRITE;
3341 bio->bi_iter.bi_sector = sector +
3342 conf->mirrors[d].replacement->data_offset;
3343 bio->bi_bdev = conf->mirrors[d].replacement->bdev;
3344 count++;
3345 }
3346
3347 if (count < 2) {
3348 for (i=0; i<conf->copies; i++) {
3349 int d = r10_bio->devs[i].devnum;
3350 if (r10_bio->devs[i].bio->bi_end_io)
3351 rdev_dec_pending(conf->mirrors[d].rdev,
3352 mddev);
3353 if (r10_bio->devs[i].repl_bio &&
3354 r10_bio->devs[i].repl_bio->bi_end_io)
3355 rdev_dec_pending(
3356 conf->mirrors[d].replacement,
3357 mddev);
3358 }
3359 put_buf(r10_bio);
3360 biolist = NULL;
3361 goto giveup;
3362 }
3363 }
3364
3365 nr_sectors = 0;
3366 if (sector_nr + max_sync < max_sector)
3367 max_sector = sector_nr + max_sync;
3368 do {
3369 struct page *page;
3370 int len = PAGE_SIZE;
3371 if (sector_nr + (len>>9) > max_sector)
3372 len = (max_sector - sector_nr) << 9;
3373 if (len == 0)
3374 break;
3375 for (bio= biolist ; bio ; bio=bio->bi_next) {
3376 struct bio *bio2;
3377 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3378 if (bio_add_page(bio, page, len, 0))
3379 continue;
3380
3381 /* stop here */
3382 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3383 for (bio2 = biolist;
3384 bio2 && bio2 != bio;
3385 bio2 = bio2->bi_next) {
3386 /* remove last page from this bio */
3387 bio2->bi_vcnt--;
3388 bio2->bi_iter.bi_size -= len;
3389 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
3390 }
3391 goto bio_full;
3392 }
3393 nr_sectors += len>>9;
3394 sector_nr += len>>9;
3395 } while (biolist->bi_vcnt < RESYNC_PAGES);
3396 bio_full:
3397 r10_bio->sectors = nr_sectors;
3398
3399 while (biolist) {
3400 bio = biolist;
3401 biolist = biolist->bi_next;
3402
3403 bio->bi_next = NULL;
3404 r10_bio = bio->bi_private;
3405 r10_bio->sectors = nr_sectors;
3406
3407 if (bio->bi_end_io == end_sync_read) {
3408 md_sync_acct(bio->bi_bdev, nr_sectors);
3409 set_bit(BIO_UPTODATE, &bio->bi_flags);
3410 generic_make_request(bio);
3411 }
3412 }
3413
3414 if (sectors_skipped)
3415 /* pretend they weren't skipped, it makes
3416 * no important difference in this case
3417 */
3418 md_done_sync(mddev, sectors_skipped, 1);
3419
3420 return sectors_skipped + nr_sectors;
3421 giveup:
3422 /* There is nowhere to write, so all non-sync
3423 * drives must be failed or in resync, all drives
3424 * have a bad block, so try the next chunk...
3425 */
3426 if (sector_nr + max_sync < max_sector)
3427 max_sector = sector_nr + max_sync;
3428
3429 sectors_skipped += (max_sector - sector_nr);
3430 chunks_skipped ++;
3431 sector_nr = max_sector;
3432 goto skipped;
3433}
3434
3435static sector_t
3436raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3437{
3438 sector_t size;
3439 struct r10conf *conf = mddev->private;
3440
3441 if (!raid_disks)
3442 raid_disks = min(conf->geo.raid_disks,
3443 conf->prev.raid_disks);
3444 if (!sectors)
3445 sectors = conf->dev_sectors;
3446
3447 size = sectors >> conf->geo.chunk_shift;
3448 sector_div(size, conf->geo.far_copies);
3449 size = size * raid_disks;
3450 sector_div(size, conf->geo.near_copies);
3451
3452 return size << conf->geo.chunk_shift;
3453}
3454
3455static void calc_sectors(struct r10conf *conf, sector_t size)
3456{
3457 /* Calculate the number of sectors-per-device that will
3458 * actually be used, and set conf->dev_sectors and
3459 * conf->stride
3460 */
3461
3462 size = size >> conf->geo.chunk_shift;
3463 sector_div(size, conf->geo.far_copies);
3464 size = size * conf->geo.raid_disks;
3465 sector_div(size, conf->geo.near_copies);
3466 /* 'size' is now the number of chunks in the array */
3467 /* calculate "used chunks per device" */
3468 size = size * conf->copies;
3469
3470 /* We need to round up when dividing by raid_disks to
3471 * get the stride size.
3472 */
3473 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3474
3475 conf->dev_sectors = size << conf->geo.chunk_shift;
3476
3477 if (conf->geo.far_offset)
3478 conf->geo.stride = 1 << conf->geo.chunk_shift;
3479 else {
3480 sector_div(size, conf->geo.far_copies);
3481 conf->geo.stride = size << conf->geo.chunk_shift;
3482 }
3483}
3484
3485enum geo_type {geo_new, geo_old, geo_start};
3486static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3487{
3488 int nc, fc, fo;
3489 int layout, chunk, disks;
3490 switch (new) {
3491 case geo_old:
3492 layout = mddev->layout;
3493 chunk = mddev->chunk_sectors;
3494 disks = mddev->raid_disks - mddev->delta_disks;
3495 break;
3496 case geo_new:
3497 layout = mddev->new_layout;
3498 chunk = mddev->new_chunk_sectors;
3499 disks = mddev->raid_disks;
3500 break;
3501 default: /* avoid 'may be unused' warnings */
3502 case geo_start: /* new when starting reshape - raid_disks not
3503 * updated yet. */
3504 layout = mddev->new_layout;
3505 chunk = mddev->new_chunk_sectors;
3506 disks = mddev->raid_disks + mddev->delta_disks;
3507 break;
3508 }
3509 if (layout >> 18)
3510 return -1;
3511 if (chunk < (PAGE_SIZE >> 9) ||
3512 !is_power_of_2(chunk))
3513 return -2;
3514 nc = layout & 255;
3515 fc = (layout >> 8) & 255;
3516 fo = layout & (1<<16);
3517 geo->raid_disks = disks;
3518 geo->near_copies = nc;
3519 geo->far_copies = fc;
3520 geo->far_offset = fo;
3521 geo->far_set_size = (layout & (1<<17)) ? disks / fc : disks;
3522 geo->chunk_mask = chunk - 1;
3523 geo->chunk_shift = ffz(~chunk);
3524 return nc*fc;
3525}
3526
3527static struct r10conf *setup_conf(struct mddev *mddev)
3528{
3529 struct r10conf *conf = NULL;
3530 int err = -EINVAL;
3531 struct geom geo;
3532 int copies;
3533
3534 copies = setup_geo(&geo, mddev, geo_new);
3535
3536 if (copies == -2) {
3537 printk(KERN_ERR "md/raid10:%s: chunk size must be "
3538 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3539 mdname(mddev), PAGE_SIZE);
3540 goto out;
3541 }
3542
3543 if (copies < 2 || copies > mddev->raid_disks) {
3544 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3545 mdname(mddev), mddev->new_layout);
3546 goto out;
3547 }
3548
3549 err = -ENOMEM;
3550 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3551 if (!conf)
3552 goto out;
3553
3554 /* FIXME calc properly */
3555 conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3556 max(0,-mddev->delta_disks)),
3557 GFP_KERNEL);
3558 if (!conf->mirrors)
3559 goto out;
3560
3561 conf->tmppage = alloc_page(GFP_KERNEL);
3562 if (!conf->tmppage)
3563 goto out;
3564
3565 conf->geo = geo;
3566 conf->copies = copies;
3567 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3568 r10bio_pool_free, conf);
3569 if (!conf->r10bio_pool)
3570 goto out;
3571
3572 calc_sectors(conf, mddev->dev_sectors);
3573 if (mddev->reshape_position == MaxSector) {
3574 conf->prev = conf->geo;
3575 conf->reshape_progress = MaxSector;
3576 } else {
3577 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3578 err = -EINVAL;
3579 goto out;
3580 }
3581 conf->reshape_progress = mddev->reshape_position;
3582 if (conf->prev.far_offset)
3583 conf->prev.stride = 1 << conf->prev.chunk_shift;
3584 else
3585 /* far_copies must be 1 */
3586 conf->prev.stride = conf->dev_sectors;
3587 }
3588 spin_lock_init(&conf->device_lock);
3589 INIT_LIST_HEAD(&conf->retry_list);
3590
3591 spin_lock_init(&conf->resync_lock);
3592 init_waitqueue_head(&conf->wait_barrier);
3593
3594 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3595 if (!conf->thread)
3596 goto out;
3597
3598 conf->mddev = mddev;
3599 return conf;
3600
3601 out:
3602 if (err == -ENOMEM)
3603 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3604 mdname(mddev));
3605 if (conf) {
3606 if (conf->r10bio_pool)
3607 mempool_destroy(conf->r10bio_pool);
3608 kfree(conf->mirrors);
3609 safe_put_page(conf->tmppage);
3610 kfree(conf);
3611 }
3612 return ERR_PTR(err);
3613}
3614
3615static int run(struct mddev *mddev)
3616{
3617 struct r10conf *conf;
3618 int i, disk_idx, chunk_size;
3619 struct raid10_info *disk;
3620 struct md_rdev *rdev;
3621 sector_t size;
3622 sector_t min_offset_diff = 0;
3623 int first = 1;
3624 bool discard_supported = false;
3625
3626 if (mddev->private == NULL) {
3627 conf = setup_conf(mddev);
3628 if (IS_ERR(conf))
3629 return PTR_ERR(conf);
3630 mddev->private = conf;
3631 }
3632 conf = mddev->private;
3633 if (!conf)
3634 goto out;
3635
3636 mddev->thread = conf->thread;
3637 conf->thread = NULL;
3638
3639 chunk_size = mddev->chunk_sectors << 9;
3640 if (mddev->queue) {
3641 blk_queue_max_discard_sectors(mddev->queue,
3642 mddev->chunk_sectors);
3643 blk_queue_max_write_same_sectors(mddev->queue, 0);
3644 blk_queue_io_min(mddev->queue, chunk_size);
3645 if (conf->geo.raid_disks % conf->geo.near_copies)
3646 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3647 else
3648 blk_queue_io_opt(mddev->queue, chunk_size *
3649 (conf->geo.raid_disks / conf->geo.near_copies));
3650 }
3651
3652 rdev_for_each(rdev, mddev) {
3653 long long diff;
3654 struct request_queue *q;
3655
3656 disk_idx = rdev->raid_disk;
3657 if (disk_idx < 0)
3658 continue;
3659 if (disk_idx >= conf->geo.raid_disks &&
3660 disk_idx >= conf->prev.raid_disks)
3661 continue;
3662 disk = conf->mirrors + disk_idx;
3663
3664 if (test_bit(Replacement, &rdev->flags)) {
3665 if (disk->replacement)
3666 goto out_free_conf;
3667 disk->replacement = rdev;
3668 } else {
3669 if (disk->rdev)
3670 goto out_free_conf;
3671 disk->rdev = rdev;
3672 }
3673 q = bdev_get_queue(rdev->bdev);
3674 if (q->merge_bvec_fn)
3675 mddev->merge_check_needed = 1;
3676 diff = (rdev->new_data_offset - rdev->data_offset);
3677 if (!mddev->reshape_backwards)
3678 diff = -diff;
3679 if (diff < 0)
3680 diff = 0;
3681 if (first || diff < min_offset_diff)
3682 min_offset_diff = diff;
3683
3684 if (mddev->gendisk)
3685 disk_stack_limits(mddev->gendisk, rdev->bdev,
3686 rdev->data_offset << 9);
3687
3688 disk->head_position = 0;
3689
3690 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3691 discard_supported = true;
3692 }
3693
3694 if (mddev->queue) {
3695 if (discard_supported)
3696 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3697 mddev->queue);
3698 else
3699 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3700 mddev->queue);
3701 }
3702 /* need to check that every block has at least one working mirror */
3703 if (!enough(conf, -1)) {
3704 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
3705 mdname(mddev));
3706 goto out_free_conf;
3707 }
3708
3709 if (conf->reshape_progress != MaxSector) {
3710 /* must ensure that shape change is supported */
3711 if (conf->geo.far_copies != 1 &&
3712 conf->geo.far_offset == 0)
3713 goto out_free_conf;
3714 if (conf->prev.far_copies != 1 &&
3715 conf->prev.far_offset == 0)
3716 goto out_free_conf;
3717 }
3718
3719 mddev->degraded = 0;
3720 for (i = 0;
3721 i < conf->geo.raid_disks
3722 || i < conf->prev.raid_disks;
3723 i++) {
3724
3725 disk = conf->mirrors + i;
3726
3727 if (!disk->rdev && disk->replacement) {
3728 /* The replacement is all we have - use it */
3729 disk->rdev = disk->replacement;
3730 disk->replacement = NULL;
3731 clear_bit(Replacement, &disk->rdev->flags);
3732 }
3733
3734 if (!disk->rdev ||
3735 !test_bit(In_sync, &disk->rdev->flags)) {
3736 disk->head_position = 0;
3737 mddev->degraded++;
3738 if (disk->rdev &&
3739 disk->rdev->saved_raid_disk < 0)
3740 conf->fullsync = 1;
3741 }
3742 disk->recovery_disabled = mddev->recovery_disabled - 1;
3743 }
3744
3745 if (mddev->recovery_cp != MaxSector)
3746 printk(KERN_NOTICE "md/raid10:%s: not clean"
3747 " -- starting background reconstruction\n",
3748 mdname(mddev));
3749 printk(KERN_INFO
3750 "md/raid10:%s: active with %d out of %d devices\n",
3751 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3752 conf->geo.raid_disks);
3753 /*
3754 * Ok, everything is just fine now
3755 */
3756 mddev->dev_sectors = conf->dev_sectors;
3757 size = raid10_size(mddev, 0, 0);
3758 md_set_array_sectors(mddev, size);
3759 mddev->resync_max_sectors = size;
3760
3761 if (mddev->queue) {
3762 int stripe = conf->geo.raid_disks *
3763 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3764 mddev->queue->backing_dev_info.congested_fn = raid10_congested;
3765 mddev->queue->backing_dev_info.congested_data = mddev;
3766
3767 /* Calculate max read-ahead size.
3768 * We need to readahead at least twice a whole stripe....
3769 * maybe...
3770 */
3771 stripe /= conf->geo.near_copies;
3772 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3773 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3774 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
3775 }
3776
3777
3778 if (md_integrity_register(mddev))
3779 goto out_free_conf;
3780
3781 if (conf->reshape_progress != MaxSector) {
3782 unsigned long before_length, after_length;
3783
3784 before_length = ((1 << conf->prev.chunk_shift) *
3785 conf->prev.far_copies);
3786 after_length = ((1 << conf->geo.chunk_shift) *
3787 conf->geo.far_copies);
3788
3789 if (max(before_length, after_length) > min_offset_diff) {
3790 /* This cannot work */
3791 printk("md/raid10: offset difference not enough to continue reshape\n");
3792 goto out_free_conf;
3793 }
3794 conf->offset_diff = min_offset_diff;
3795
3796 conf->reshape_safe = conf->reshape_progress;
3797 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3798 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3799 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3800 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3801 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3802 "reshape");
3803 }
3804
3805 return 0;
3806
3807out_free_conf:
3808 md_unregister_thread(&mddev->thread);
3809 if (conf->r10bio_pool)
3810 mempool_destroy(conf->r10bio_pool);
3811 safe_put_page(conf->tmppage);
3812 kfree(conf->mirrors);
3813 kfree(conf);
3814 mddev->private = NULL;
3815out:
3816 return -EIO;
3817}
3818
3819static int stop(struct mddev *mddev)
3820{
3821 struct r10conf *conf = mddev->private;
3822
3823 raise_barrier(conf, 0);
3824 lower_barrier(conf);
3825
3826 md_unregister_thread(&mddev->thread);
3827 if (mddev->queue)
3828 /* the unplug fn references 'conf'*/
3829 blk_sync_queue(mddev->queue);
3830
3831 if (conf->r10bio_pool)
3832 mempool_destroy(conf->r10bio_pool);
3833 safe_put_page(conf->tmppage);
3834 kfree(conf->mirrors);
3835 kfree(conf);
3836 mddev->private = NULL;
3837 return 0;
3838}
3839
3840static void raid10_quiesce(struct mddev *mddev, int state)
3841{
3842 struct r10conf *conf = mddev->private;
3843
3844 switch(state) {
3845 case 1:
3846 raise_barrier(conf, 0);
3847 break;
3848 case 0:
3849 lower_barrier(conf);
3850 break;
3851 }
3852}
3853
3854static int raid10_resize(struct mddev *mddev, sector_t sectors)
3855{
3856 /* Resize of 'far' arrays is not supported.
3857 * For 'near' and 'offset' arrays we can set the
3858 * number of sectors used to be an appropriate multiple
3859 * of the chunk size.
3860 * For 'offset', this is far_copies*chunksize.
3861 * For 'near' the multiplier is the LCM of
3862 * near_copies and raid_disks.
3863 * So if far_copies > 1 && !far_offset, fail.
3864 * Else find LCM(raid_disks, near_copy)*far_copies and
3865 * multiply by chunk_size. Then round to this number.
3866 * This is mostly done by raid10_size()
3867 */
3868 struct r10conf *conf = mddev->private;
3869 sector_t oldsize, size;
3870
3871 if (mddev->reshape_position != MaxSector)
3872 return -EBUSY;
3873
3874 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3875 return -EINVAL;
3876
3877 oldsize = raid10_size(mddev, 0, 0);
3878 size = raid10_size(mddev, sectors, 0);
3879 if (mddev->external_size &&
3880 mddev->array_sectors > size)
3881 return -EINVAL;
3882 if (mddev->bitmap) {
3883 int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3884 if (ret)
3885 return ret;
3886 }
3887 md_set_array_sectors(mddev, size);
3888 set_capacity(mddev->gendisk, mddev->array_sectors);
3889 revalidate_disk(mddev->gendisk);
3890 if (sectors > mddev->dev_sectors &&
3891 mddev->recovery_cp > oldsize) {
3892 mddev->recovery_cp = oldsize;
3893 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3894 }
3895 calc_sectors(conf, sectors);
3896 mddev->dev_sectors = conf->dev_sectors;
3897 mddev->resync_max_sectors = size;
3898 return 0;
3899}
3900
3901static void *raid10_takeover_raid0(struct mddev *mddev)
3902{
3903 struct md_rdev *rdev;
3904 struct r10conf *conf;
3905
3906 if (mddev->degraded > 0) {
3907 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3908 mdname(mddev));
3909 return ERR_PTR(-EINVAL);
3910 }
3911
3912 /* Set new parameters */
3913 mddev->new_level = 10;
3914 /* new layout: far_copies = 1, near_copies = 2 */
3915 mddev->new_layout = (1<<8) + 2;
3916 mddev->new_chunk_sectors = mddev->chunk_sectors;
3917 mddev->delta_disks = mddev->raid_disks;
3918 mddev->raid_disks *= 2;
3919 /* make sure it will be not marked as dirty */
3920 mddev->recovery_cp = MaxSector;
3921
3922 conf = setup_conf(mddev);
3923 if (!IS_ERR(conf)) {
3924 rdev_for_each(rdev, mddev)
3925 if (rdev->raid_disk >= 0)
3926 rdev->new_raid_disk = rdev->raid_disk * 2;
3927 conf->barrier = 1;
3928 }
3929
3930 return conf;
3931}
3932
3933static void *raid10_takeover(struct mddev *mddev)
3934{
3935 struct r0conf *raid0_conf;
3936
3937 /* raid10 can take over:
3938 * raid0 - providing it has only two drives
3939 */
3940 if (mddev->level == 0) {
3941 /* for raid0 takeover only one zone is supported */
3942 raid0_conf = mddev->private;
3943 if (raid0_conf->nr_strip_zones > 1) {
3944 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3945 " with more than one zone.\n",
3946 mdname(mddev));
3947 return ERR_PTR(-EINVAL);
3948 }
3949 return raid10_takeover_raid0(mddev);
3950 }
3951 return ERR_PTR(-EINVAL);
3952}
3953
3954static int raid10_check_reshape(struct mddev *mddev)
3955{
3956 /* Called when there is a request to change
3957 * - layout (to ->new_layout)
3958 * - chunk size (to ->new_chunk_sectors)
3959 * - raid_disks (by delta_disks)
3960 * or when trying to restart a reshape that was ongoing.
3961 *
3962 * We need to validate the request and possibly allocate
3963 * space if that might be an issue later.
3964 *
3965 * Currently we reject any reshape of a 'far' mode array,
3966 * allow chunk size to change if new is generally acceptable,
3967 * allow raid_disks to increase, and allow
3968 * a switch between 'near' mode and 'offset' mode.
3969 */
3970 struct r10conf *conf = mddev->private;
3971 struct geom geo;
3972
3973 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
3974 return -EINVAL;
3975
3976 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
3977 /* mustn't change number of copies */
3978 return -EINVAL;
3979 if (geo.far_copies > 1 && !geo.far_offset)
3980 /* Cannot switch to 'far' mode */
3981 return -EINVAL;
3982
3983 if (mddev->array_sectors & geo.chunk_mask)
3984 /* not factor of array size */
3985 return -EINVAL;
3986
3987 if (!enough(conf, -1))
3988 return -EINVAL;
3989
3990 kfree(conf->mirrors_new);
3991 conf->mirrors_new = NULL;
3992 if (mddev->delta_disks > 0) {
3993 /* allocate new 'mirrors' list */
3994 conf->mirrors_new = kzalloc(
3995 sizeof(struct raid10_info)
3996 *(mddev->raid_disks +
3997 mddev->delta_disks),
3998 GFP_KERNEL);
3999 if (!conf->mirrors_new)
4000 return -ENOMEM;
4001 }
4002 return 0;
4003}
4004
4005/*
4006 * Need to check if array has failed when deciding whether to:
4007 * - start an array
4008 * - remove non-faulty devices
4009 * - add a spare
4010 * - allow a reshape
4011 * This determination is simple when no reshape is happening.
4012 * However if there is a reshape, we need to carefully check
4013 * both the before and after sections.
4014 * This is because some failed devices may only affect one
4015 * of the two sections, and some non-in_sync devices may
4016 * be insync in the section most affected by failed devices.
4017 */
4018static int calc_degraded(struct r10conf *conf)
4019{
4020 int degraded, degraded2;
4021 int i;
4022
4023 rcu_read_lock();
4024 degraded = 0;
4025 /* 'prev' section first */
4026 for (i = 0; i < conf->prev.raid_disks; i++) {
4027 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4028 if (!rdev || test_bit(Faulty, &rdev->flags))
4029 degraded++;
4030 else if (!test_bit(In_sync, &rdev->flags))
4031 /* When we can reduce the number of devices in
4032 * an array, this might not contribute to
4033 * 'degraded'. It does now.
4034 */
4035 degraded++;
4036 }
4037 rcu_read_unlock();
4038 if (conf->geo.raid_disks == conf->prev.raid_disks)
4039 return degraded;
4040 rcu_read_lock();
4041 degraded2 = 0;
4042 for (i = 0; i < conf->geo.raid_disks; i++) {
4043 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4044 if (!rdev || test_bit(Faulty, &rdev->flags))
4045 degraded2++;
4046 else if (!test_bit(In_sync, &rdev->flags)) {
4047 /* If reshape is increasing the number of devices,
4048 * this section has already been recovered, so
4049 * it doesn't contribute to degraded.
4050 * else it does.
4051 */
4052 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4053 degraded2++;
4054 }
4055 }
4056 rcu_read_unlock();
4057 if (degraded2 > degraded)
4058 return degraded2;
4059 return degraded;
4060}
4061
4062static int raid10_start_reshape(struct mddev *mddev)
4063{
4064 /* A 'reshape' has been requested. This commits
4065 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4066 * This also checks if there are enough spares and adds them
4067 * to the array.
4068 * We currently require enough spares to make the final
4069 * array non-degraded. We also require that the difference
4070 * between old and new data_offset - on each device - is
4071 * enough that we never risk over-writing.
4072 */
4073
4074 unsigned long before_length, after_length;
4075 sector_t min_offset_diff = 0;
4076 int first = 1;
4077 struct geom new;
4078 struct r10conf *conf = mddev->private;
4079 struct md_rdev *rdev;
4080 int spares = 0;
4081 int ret;
4082
4083 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4084 return -EBUSY;
4085
4086 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4087 return -EINVAL;
4088
4089 before_length = ((1 << conf->prev.chunk_shift) *
4090 conf->prev.far_copies);
4091 after_length = ((1 << conf->geo.chunk_shift) *
4092 conf->geo.far_copies);
4093
4094 rdev_for_each(rdev, mddev) {
4095 if (!test_bit(In_sync, &rdev->flags)
4096 && !test_bit(Faulty, &rdev->flags))
4097 spares++;
4098 if (rdev->raid_disk >= 0) {
4099 long long diff = (rdev->new_data_offset
4100 - rdev->data_offset);
4101 if (!mddev->reshape_backwards)
4102 diff = -diff;
4103 if (diff < 0)
4104 diff = 0;
4105 if (first || diff < min_offset_diff)
4106 min_offset_diff = diff;
4107 }
4108 }
4109
4110 if (max(before_length, after_length) > min_offset_diff)
4111 return -EINVAL;
4112
4113 if (spares < mddev->delta_disks)
4114 return -EINVAL;
4115
4116 conf->offset_diff = min_offset_diff;
4117 spin_lock_irq(&conf->device_lock);
4118 if (conf->mirrors_new) {
4119 memcpy(conf->mirrors_new, conf->mirrors,
4120 sizeof(struct raid10_info)*conf->prev.raid_disks);
4121 smp_mb();
4122 kfree(conf->mirrors_old); /* FIXME and elsewhere */
4123 conf->mirrors_old = conf->mirrors;
4124 conf->mirrors = conf->mirrors_new;
4125 conf->mirrors_new = NULL;
4126 }
4127 setup_geo(&conf->geo, mddev, geo_start);
4128 smp_mb();
4129 if (mddev->reshape_backwards) {
4130 sector_t size = raid10_size(mddev, 0, 0);
4131 if (size < mddev->array_sectors) {
4132 spin_unlock_irq(&conf->device_lock);
4133 printk(KERN_ERR "md/raid10:%s: array size must be reduce before number of disks\n",
4134 mdname(mddev));
4135 return -EINVAL;
4136 }
4137 mddev->resync_max_sectors = size;
4138 conf->reshape_progress = size;
4139 } else
4140 conf->reshape_progress = 0;
4141 spin_unlock_irq(&conf->device_lock);
4142
4143 if (mddev->delta_disks && mddev->bitmap) {
4144 ret = bitmap_resize(mddev->bitmap,
4145 raid10_size(mddev, 0,
4146 conf->geo.raid_disks),
4147 0, 0);
4148 if (ret)
4149 goto abort;
4150 }
4151 if (mddev->delta_disks > 0) {
4152 rdev_for_each(rdev, mddev)
4153 if (rdev->raid_disk < 0 &&
4154 !test_bit(Faulty, &rdev->flags)) {
4155 if (raid10_add_disk(mddev, rdev) == 0) {
4156 if (rdev->raid_disk >=
4157 conf->prev.raid_disks)
4158 set_bit(In_sync, &rdev->flags);
4159 else
4160 rdev->recovery_offset = 0;
4161
4162 if (sysfs_link_rdev(mddev, rdev))
4163 /* Failure here is OK */;
4164 }
4165 } else if (rdev->raid_disk >= conf->prev.raid_disks
4166 && !test_bit(Faulty, &rdev->flags)) {
4167 /* This is a spare that was manually added */
4168 set_bit(In_sync, &rdev->flags);
4169 }
4170 }
4171 /* When a reshape changes the number of devices,
4172 * ->degraded is measured against the larger of the
4173 * pre and post numbers.
4174 */
4175 spin_lock_irq(&conf->device_lock);
4176 mddev->degraded = calc_degraded(conf);
4177 spin_unlock_irq(&conf->device_lock);
4178 mddev->raid_disks = conf->geo.raid_disks;
4179 mddev->reshape_position = conf->reshape_progress;
4180 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4181
4182 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4183 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4184 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4185 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4186
4187 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4188 "reshape");
4189 if (!mddev->sync_thread) {
4190 ret = -EAGAIN;
4191 goto abort;
4192 }
4193 conf->reshape_checkpoint = jiffies;
4194 md_wakeup_thread(mddev->sync_thread);
4195 md_new_event(mddev);
4196 return 0;
4197
4198abort:
4199 mddev->recovery = 0;
4200 spin_lock_irq(&conf->device_lock);
4201 conf->geo = conf->prev;
4202 mddev->raid_disks = conf->geo.raid_disks;
4203 rdev_for_each(rdev, mddev)
4204 rdev->new_data_offset = rdev->data_offset;
4205 smp_wmb();
4206 conf->reshape_progress = MaxSector;
4207 mddev->reshape_position = MaxSector;
4208 spin_unlock_irq(&conf->device_lock);
4209 return ret;
4210}
4211
4212/* Calculate the last device-address that could contain
4213 * any block from the chunk that includes the array-address 's'
4214 * and report the next address.
4215 * i.e. the address returned will be chunk-aligned and after
4216 * any data that is in the chunk containing 's'.
4217 */
4218static sector_t last_dev_address(sector_t s, struct geom *geo)
4219{
4220 s = (s | geo->chunk_mask) + 1;
4221 s >>= geo->chunk_shift;
4222 s *= geo->near_copies;
4223 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4224 s *= geo->far_copies;
4225 s <<= geo->chunk_shift;
4226 return s;
4227}
4228
4229/* Calculate the first device-address that could contain
4230 * any block from the chunk that includes the array-address 's'.
4231 * This too will be the start of a chunk
4232 */
4233static sector_t first_dev_address(sector_t s, struct geom *geo)
4234{
4235 s >>= geo->chunk_shift;
4236 s *= geo->near_copies;
4237 sector_div(s, geo->raid_disks);
4238 s *= geo->far_copies;
4239 s <<= geo->chunk_shift;
4240 return s;
4241}
4242
4243static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4244 int *skipped)
4245{
4246 /* We simply copy at most one chunk (smallest of old and new)
4247 * at a time, possibly less if that exceeds RESYNC_PAGES,
4248 * or we hit a bad block or something.
4249 * This might mean we pause for normal IO in the middle of
4250 * a chunk, but that is not a problem was mddev->reshape_position
4251 * can record any location.
4252 *
4253 * If we will want to write to a location that isn't
4254 * yet recorded as 'safe' (i.e. in metadata on disk) then
4255 * we need to flush all reshape requests and update the metadata.
4256 *
4257 * When reshaping forwards (e.g. to more devices), we interpret
4258 * 'safe' as the earliest block which might not have been copied
4259 * down yet. We divide this by previous stripe size and multiply
4260 * by previous stripe length to get lowest device offset that we
4261 * cannot write to yet.
4262 * We interpret 'sector_nr' as an address that we want to write to.
4263 * From this we use last_device_address() to find where we might
4264 * write to, and first_device_address on the 'safe' position.
4265 * If this 'next' write position is after the 'safe' position,
4266 * we must update the metadata to increase the 'safe' position.
4267 *
4268 * When reshaping backwards, we round in the opposite direction
4269 * and perform the reverse test: next write position must not be
4270 * less than current safe position.
4271 *
4272 * In all this the minimum difference in data offsets
4273 * (conf->offset_diff - always positive) allows a bit of slack,
4274 * so next can be after 'safe', but not by more than offset_disk
4275 *
4276 * We need to prepare all the bios here before we start any IO
4277 * to ensure the size we choose is acceptable to all devices.
4278 * The means one for each copy for write-out and an extra one for
4279 * read-in.
4280 * We store the read-in bio in ->master_bio and the others in
4281 * ->devs[x].bio and ->devs[x].repl_bio.
4282 */
4283 struct r10conf *conf = mddev->private;
4284 struct r10bio *r10_bio;
4285 sector_t next, safe, last;
4286 int max_sectors;
4287 int nr_sectors;
4288 int s;
4289 struct md_rdev *rdev;
4290 int need_flush = 0;
4291 struct bio *blist;
4292 struct bio *bio, *read_bio;
4293 int sectors_done = 0;
4294
4295 if (sector_nr == 0) {
4296 /* If restarting in the middle, skip the initial sectors */
4297 if (mddev->reshape_backwards &&
4298 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4299 sector_nr = (raid10_size(mddev, 0, 0)
4300 - conf->reshape_progress);
4301 } else if (!mddev->reshape_backwards &&
4302 conf->reshape_progress > 0)
4303 sector_nr = conf->reshape_progress;
4304 if (sector_nr) {
4305 mddev->curr_resync_completed = sector_nr;
4306 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4307 *skipped = 1;
4308 return sector_nr;
4309 }
4310 }
4311
4312 /* We don't use sector_nr to track where we are up to
4313 * as that doesn't work well for ->reshape_backwards.
4314 * So just use ->reshape_progress.
4315 */
4316 if (mddev->reshape_backwards) {
4317 /* 'next' is the earliest device address that we might
4318 * write to for this chunk in the new layout
4319 */
4320 next = first_dev_address(conf->reshape_progress - 1,
4321 &conf->geo);
4322
4323 /* 'safe' is the last device address that we might read from
4324 * in the old layout after a restart
4325 */
4326 safe = last_dev_address(conf->reshape_safe - 1,
4327 &conf->prev);
4328
4329 if (next + conf->offset_diff < safe)
4330 need_flush = 1;
4331
4332 last = conf->reshape_progress - 1;
4333 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4334 & conf->prev.chunk_mask);
4335 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4336 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4337 } else {
4338 /* 'next' is after the last device address that we
4339 * might write to for this chunk in the new layout
4340 */
4341 next = last_dev_address(conf->reshape_progress, &conf->geo);
4342
4343 /* 'safe' is the earliest device address that we might
4344 * read from in the old layout after a restart
4345 */
4346 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4347
4348 /* Need to update metadata if 'next' might be beyond 'safe'
4349 * as that would possibly corrupt data
4350 */
4351 if (next > safe + conf->offset_diff)
4352 need_flush = 1;
4353
4354 sector_nr = conf->reshape_progress;
4355 last = sector_nr | (conf->geo.chunk_mask
4356 & conf->prev.chunk_mask);
4357
4358 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4359 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4360 }
4361
4362 if (need_flush ||
4363 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4364 /* Need to update reshape_position in metadata */
4365 wait_barrier(conf);
4366 mddev->reshape_position = conf->reshape_progress;
4367 if (mddev->reshape_backwards)
4368 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4369 - conf->reshape_progress;
4370 else
4371 mddev->curr_resync_completed = conf->reshape_progress;
4372 conf->reshape_checkpoint = jiffies;
4373 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4374 md_wakeup_thread(mddev->thread);
4375 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4376 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4377 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4378 allow_barrier(conf);
4379 return sectors_done;
4380 }
4381 conf->reshape_safe = mddev->reshape_position;
4382 allow_barrier(conf);
4383 }
4384
4385read_more:
4386 /* Now schedule reads for blocks from sector_nr to last */
4387 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4388 raise_barrier(conf, sectors_done != 0);
4389 atomic_set(&r10_bio->remaining, 0);
4390 r10_bio->mddev = mddev;
4391 r10_bio->sector = sector_nr;
4392 set_bit(R10BIO_IsReshape, &r10_bio->state);
4393 r10_bio->sectors = last - sector_nr + 1;
4394 rdev = read_balance(conf, r10_bio, &max_sectors);
4395 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4396
4397 if (!rdev) {
4398 /* Cannot read from here, so need to record bad blocks
4399 * on all the target devices.
4400 */
4401 // FIXME
4402 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4403 return sectors_done;
4404 }
4405
4406 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4407
4408 read_bio->bi_bdev = rdev->bdev;
4409 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4410 + rdev->data_offset);
4411 read_bio->bi_private = r10_bio;
4412 read_bio->bi_end_io = end_sync_read;
4413 read_bio->bi_rw = READ;
4414 read_bio->bi_flags &= ~(BIO_POOL_MASK - 1);
4415 read_bio->bi_flags |= 1 << BIO_UPTODATE;
4416 read_bio->bi_vcnt = 0;
4417 read_bio->bi_iter.bi_size = 0;
4418 r10_bio->master_bio = read_bio;
4419 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4420
4421 /* Now find the locations in the new layout */
4422 __raid10_find_phys(&conf->geo, r10_bio);
4423
4424 blist = read_bio;
4425 read_bio->bi_next = NULL;
4426
4427 for (s = 0; s < conf->copies*2; s++) {
4428 struct bio *b;
4429 int d = r10_bio->devs[s/2].devnum;
4430 struct md_rdev *rdev2;
4431 if (s&1) {
4432 rdev2 = conf->mirrors[d].replacement;
4433 b = r10_bio->devs[s/2].repl_bio;
4434 } else {
4435 rdev2 = conf->mirrors[d].rdev;
4436 b = r10_bio->devs[s/2].bio;
4437 }
4438 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4439 continue;
4440
4441 bio_reset(b);
4442 b->bi_bdev = rdev2->bdev;
4443 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4444 rdev2->new_data_offset;
4445 b->bi_private = r10_bio;
4446 b->bi_end_io = end_reshape_write;
4447 b->bi_rw = WRITE;
4448 b->bi_next = blist;
4449 blist = b;
4450 }
4451
4452 /* Now add as many pages as possible to all of these bios. */
4453
4454 nr_sectors = 0;
4455 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4456 struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4457 int len = (max_sectors - s) << 9;
4458 if (len > PAGE_SIZE)
4459 len = PAGE_SIZE;
4460 for (bio = blist; bio ; bio = bio->bi_next) {
4461 struct bio *bio2;
4462 if (bio_add_page(bio, page, len, 0))
4463 continue;
4464
4465 /* Didn't fit, must stop */
4466 for (bio2 = blist;
4467 bio2 && bio2 != bio;
4468 bio2 = bio2->bi_next) {
4469 /* Remove last page from this bio */
4470 bio2->bi_vcnt--;
4471 bio2->bi_iter.bi_size -= len;
4472 bio2->bi_flags &= ~(1<<BIO_SEG_VALID);
4473 }
4474 goto bio_full;
4475 }
4476 sector_nr += len >> 9;
4477 nr_sectors += len >> 9;
4478 }
4479bio_full:
4480 r10_bio->sectors = nr_sectors;
4481
4482 /* Now submit the read */
4483 md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4484 atomic_inc(&r10_bio->remaining);
4485 read_bio->bi_next = NULL;
4486 generic_make_request(read_bio);
4487 sector_nr += nr_sectors;
4488 sectors_done += nr_sectors;
4489 if (sector_nr <= last)
4490 goto read_more;
4491
4492 /* Now that we have done the whole section we can
4493 * update reshape_progress
4494 */
4495 if (mddev->reshape_backwards)
4496 conf->reshape_progress -= sectors_done;
4497 else
4498 conf->reshape_progress += sectors_done;
4499
4500 return sectors_done;
4501}
4502
4503static void end_reshape_request(struct r10bio *r10_bio);
4504static int handle_reshape_read_error(struct mddev *mddev,
4505 struct r10bio *r10_bio);
4506static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4507{
4508 /* Reshape read completed. Hopefully we have a block
4509 * to write out.
4510 * If we got a read error then we do sync 1-page reads from
4511 * elsewhere until we find the data - or give up.
4512 */
4513 struct r10conf *conf = mddev->private;
4514 int s;
4515
4516 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4517 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4518 /* Reshape has been aborted */
4519 md_done_sync(mddev, r10_bio->sectors, 0);
4520 return;
4521 }
4522
4523 /* We definitely have the data in the pages, schedule the
4524 * writes.
4525 */
4526 atomic_set(&r10_bio->remaining, 1);
4527 for (s = 0; s < conf->copies*2; s++) {
4528 struct bio *b;
4529 int d = r10_bio->devs[s/2].devnum;
4530 struct md_rdev *rdev;
4531 if (s&1) {
4532 rdev = conf->mirrors[d].replacement;
4533 b = r10_bio->devs[s/2].repl_bio;
4534 } else {
4535 rdev = conf->mirrors[d].rdev;
4536 b = r10_bio->devs[s/2].bio;
4537 }
4538 if (!rdev || test_bit(Faulty, &rdev->flags))
4539 continue;
4540 atomic_inc(&rdev->nr_pending);
4541 md_sync_acct(b->bi_bdev, r10_bio->sectors);
4542 atomic_inc(&r10_bio->remaining);
4543 b->bi_next = NULL;
4544 generic_make_request(b);
4545 }
4546 end_reshape_request(r10_bio);
4547}
4548
4549static void end_reshape(struct r10conf *conf)
4550{
4551 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4552 return;
4553
4554 spin_lock_irq(&conf->device_lock);
4555 conf->prev = conf->geo;
4556 md_finish_reshape(conf->mddev);
4557 smp_wmb();
4558 conf->reshape_progress = MaxSector;
4559 spin_unlock_irq(&conf->device_lock);
4560
4561 /* read-ahead size must cover two whole stripes, which is
4562 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4563 */
4564 if (conf->mddev->queue) {
4565 int stripe = conf->geo.raid_disks *
4566 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4567 stripe /= conf->geo.near_copies;
4568 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4569 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4570 }
4571 conf->fullsync = 0;
4572}
4573
4574
4575static int handle_reshape_read_error(struct mddev *mddev,
4576 struct r10bio *r10_bio)
4577{
4578 /* Use sync reads to get the blocks from somewhere else */
4579 int sectors = r10_bio->sectors;
4580 struct r10conf *conf = mddev->private;
4581 struct {
4582 struct r10bio r10_bio;
4583 struct r10dev devs[conf->copies];
4584 } on_stack;
4585 struct r10bio *r10b = &on_stack.r10_bio;
4586 int slot = 0;
4587 int idx = 0;
4588 struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
4589
4590 r10b->sector = r10_bio->sector;
4591 __raid10_find_phys(&conf->prev, r10b);
4592
4593 while (sectors) {
4594 int s = sectors;
4595 int success = 0;
4596 int first_slot = slot;
4597
4598 if (s > (PAGE_SIZE >> 9))
4599 s = PAGE_SIZE >> 9;
4600
4601 while (!success) {
4602 int d = r10b->devs[slot].devnum;
4603 struct md_rdev *rdev = conf->mirrors[d].rdev;
4604 sector_t addr;
4605 if (rdev == NULL ||
4606 test_bit(Faulty, &rdev->flags) ||
4607 !test_bit(In_sync, &rdev->flags))
4608 goto failed;
4609
4610 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4611 success = sync_page_io(rdev,
4612 addr,
4613 s << 9,
4614 bvec[idx].bv_page,
4615 READ, false);
4616 if (success)
4617 break;
4618 failed:
4619 slot++;
4620 if (slot >= conf->copies)
4621 slot = 0;
4622 if (slot == first_slot)
4623 break;
4624 }
4625 if (!success) {
4626 /* couldn't read this block, must give up */
4627 set_bit(MD_RECOVERY_INTR,
4628 &mddev->recovery);
4629 return -EIO;
4630 }
4631 sectors -= s;
4632 idx++;
4633 }
4634 return 0;
4635}
4636
4637static void end_reshape_write(struct bio *bio, int error)
4638{
4639 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
4640 struct r10bio *r10_bio = bio->bi_private;
4641 struct mddev *mddev = r10_bio->mddev;
4642 struct r10conf *conf = mddev->private;
4643 int d;
4644 int slot;
4645 int repl;
4646 struct md_rdev *rdev = NULL;
4647
4648 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4649 if (repl)
4650 rdev = conf->mirrors[d].replacement;
4651 if (!rdev) {
4652 smp_mb();
4653 rdev = conf->mirrors[d].rdev;
4654 }
4655
4656 if (!uptodate) {
4657 /* FIXME should record badblock */
4658 md_error(mddev, rdev);
4659 }
4660
4661 rdev_dec_pending(rdev, mddev);
4662 end_reshape_request(r10_bio);
4663}
4664
4665static void end_reshape_request(struct r10bio *r10_bio)
4666{
4667 if (!atomic_dec_and_test(&r10_bio->remaining))
4668 return;
4669 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4670 bio_put(r10_bio->master_bio);
4671 put_buf(r10_bio);
4672}
4673
4674static void raid10_finish_reshape(struct mddev *mddev)
4675{
4676 struct r10conf *conf = mddev->private;
4677
4678 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4679 return;
4680
4681 if (mddev->delta_disks > 0) {
4682 sector_t size = raid10_size(mddev, 0, 0);
4683 md_set_array_sectors(mddev, size);
4684 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4685 mddev->recovery_cp = mddev->resync_max_sectors;
4686 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4687 }
4688 mddev->resync_max_sectors = size;
4689 set_capacity(mddev->gendisk, mddev->array_sectors);
4690 revalidate_disk(mddev->gendisk);
4691 } else {
4692 int d;
4693 for (d = conf->geo.raid_disks ;
4694 d < conf->geo.raid_disks - mddev->delta_disks;
4695 d++) {
4696 struct md_rdev *rdev = conf->mirrors[d].rdev;
4697 if (rdev)
4698 clear_bit(In_sync, &rdev->flags);
4699 rdev = conf->mirrors[d].replacement;
4700 if (rdev)
4701 clear_bit(In_sync, &rdev->flags);
4702 }
4703 }
4704 mddev->layout = mddev->new_layout;
4705 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4706 mddev->reshape_position = MaxSector;
4707 mddev->delta_disks = 0;
4708 mddev->reshape_backwards = 0;
4709}
4710
4711static struct md_personality raid10_personality =
4712{
4713 .name = "raid10",
4714 .level = 10,
4715 .owner = THIS_MODULE,
4716 .make_request = make_request,
4717 .run = run,
4718 .stop = stop,
4719 .status = status,
4720 .error_handler = error,
4721 .hot_add_disk = raid10_add_disk,
4722 .hot_remove_disk= raid10_remove_disk,
4723 .spare_active = raid10_spare_active,
4724 .sync_request = sync_request,
4725 .quiesce = raid10_quiesce,
4726 .size = raid10_size,
4727 .resize = raid10_resize,
4728 .takeover = raid10_takeover,
4729 .check_reshape = raid10_check_reshape,
4730 .start_reshape = raid10_start_reshape,
4731 .finish_reshape = raid10_finish_reshape,
4732};
4733
4734static int __init raid_init(void)
4735{
4736 return register_md_personality(&raid10_personality);
4737}
4738
4739static void raid_exit(void)
4740{
4741 unregister_md_personality(&raid10_personality);
4742}
4743
4744module_init(raid_init);
4745module_exit(raid_exit);
4746MODULE_LICENSE("GPL");
4747MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4748MODULE_ALIAS("md-personality-9"); /* RAID10 */
4749MODULE_ALIAS("md-raid10");
4750MODULE_ALIAS("md-level-10");
4751
4752module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);
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/seq_file.h>
25#include <linux/ratelimit.h>
26#include "md.h"
27#include "raid10.h"
28#include "raid0.h"
29#include "bitmap.h"
30
31/*
32 * RAID10 provides a combination of RAID0 and RAID1 functionality.
33 * The layout of data is defined by
34 * chunk_size
35 * raid_disks
36 * near_copies (stored in low byte of layout)
37 * far_copies (stored in second byte of layout)
38 * far_offset (stored in bit 16 of layout )
39 *
40 * The data to be stored is divided into chunks using chunksize.
41 * Each device is divided into far_copies sections.
42 * In each section, chunks are laid out in a style similar to raid0, but
43 * near_copies copies of each chunk is stored (each on a different drive).
44 * The starting device for each section is offset near_copies from the starting
45 * device of the previous section.
46 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
47 * drive.
48 * near_copies and far_copies must be at least one, and their product is at most
49 * raid_disks.
50 *
51 * If far_offset is true, then the far_copies are handled a bit differently.
52 * The copies are still in different stripes, but instead of be very far apart
53 * on disk, there are adjacent stripes.
54 */
55
56/*
57 * Number of guaranteed r10bios in case of extreme VM load:
58 */
59#define NR_RAID10_BIOS 256
60
61static void allow_barrier(conf_t *conf);
62static void lower_barrier(conf_t *conf);
63
64static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
65{
66 conf_t *conf = data;
67 int size = offsetof(struct r10bio_s, devs[conf->copies]);
68
69 /* allocate a r10bio with room for raid_disks entries in the bios array */
70 return kzalloc(size, gfp_flags);
71}
72
73static void r10bio_pool_free(void *r10_bio, void *data)
74{
75 kfree(r10_bio);
76}
77
78/* Maximum size of each resync request */
79#define RESYNC_BLOCK_SIZE (64*1024)
80#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
81/* amount of memory to reserve for resync requests */
82#define RESYNC_WINDOW (1024*1024)
83/* maximum number of concurrent requests, memory permitting */
84#define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
85
86/*
87 * When performing a resync, we need to read and compare, so
88 * we need as many pages are there are copies.
89 * When performing a recovery, we need 2 bios, one for read,
90 * one for write (we recover only one drive per r10buf)
91 *
92 */
93static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
94{
95 conf_t *conf = data;
96 struct page *page;
97 r10bio_t *r10_bio;
98 struct bio *bio;
99 int i, j;
100 int nalloc;
101
102 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
103 if (!r10_bio)
104 return NULL;
105
106 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
107 nalloc = conf->copies; /* resync */
108 else
109 nalloc = 2; /* recovery */
110
111 /*
112 * Allocate bios.
113 */
114 for (j = nalloc ; j-- ; ) {
115 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
116 if (!bio)
117 goto out_free_bio;
118 r10_bio->devs[j].bio = bio;
119 }
120 /*
121 * Allocate RESYNC_PAGES data pages and attach them
122 * where needed.
123 */
124 for (j = 0 ; j < nalloc; j++) {
125 bio = r10_bio->devs[j].bio;
126 for (i = 0; i < RESYNC_PAGES; i++) {
127 if (j == 1 && !test_bit(MD_RECOVERY_SYNC,
128 &conf->mddev->recovery)) {
129 /* we can share bv_page's during recovery */
130 struct bio *rbio = r10_bio->devs[0].bio;
131 page = rbio->bi_io_vec[i].bv_page;
132 get_page(page);
133 } else
134 page = alloc_page(gfp_flags);
135 if (unlikely(!page))
136 goto out_free_pages;
137
138 bio->bi_io_vec[i].bv_page = page;
139 }
140 }
141
142 return r10_bio;
143
144out_free_pages:
145 for ( ; i > 0 ; i--)
146 safe_put_page(bio->bi_io_vec[i-1].bv_page);
147 while (j--)
148 for (i = 0; i < RESYNC_PAGES ; i++)
149 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
150 j = -1;
151out_free_bio:
152 while ( ++j < nalloc )
153 bio_put(r10_bio->devs[j].bio);
154 r10bio_pool_free(r10_bio, conf);
155 return NULL;
156}
157
158static void r10buf_pool_free(void *__r10_bio, void *data)
159{
160 int i;
161 conf_t *conf = data;
162 r10bio_t *r10bio = __r10_bio;
163 int j;
164
165 for (j=0; j < conf->copies; j++) {
166 struct bio *bio = r10bio->devs[j].bio;
167 if (bio) {
168 for (i = 0; i < RESYNC_PAGES; i++) {
169 safe_put_page(bio->bi_io_vec[i].bv_page);
170 bio->bi_io_vec[i].bv_page = NULL;
171 }
172 bio_put(bio);
173 }
174 }
175 r10bio_pool_free(r10bio, conf);
176}
177
178static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
179{
180 int i;
181
182 for (i = 0; i < conf->copies; i++) {
183 struct bio **bio = & r10_bio->devs[i].bio;
184 if (!BIO_SPECIAL(*bio))
185 bio_put(*bio);
186 *bio = NULL;
187 }
188}
189
190static void free_r10bio(r10bio_t *r10_bio)
191{
192 conf_t *conf = r10_bio->mddev->private;
193
194 put_all_bios(conf, r10_bio);
195 mempool_free(r10_bio, conf->r10bio_pool);
196}
197
198static void put_buf(r10bio_t *r10_bio)
199{
200 conf_t *conf = r10_bio->mddev->private;
201
202 mempool_free(r10_bio, conf->r10buf_pool);
203
204 lower_barrier(conf);
205}
206
207static void reschedule_retry(r10bio_t *r10_bio)
208{
209 unsigned long flags;
210 mddev_t *mddev = r10_bio->mddev;
211 conf_t *conf = mddev->private;
212
213 spin_lock_irqsave(&conf->device_lock, flags);
214 list_add(&r10_bio->retry_list, &conf->retry_list);
215 conf->nr_queued ++;
216 spin_unlock_irqrestore(&conf->device_lock, flags);
217
218 /* wake up frozen array... */
219 wake_up(&conf->wait_barrier);
220
221 md_wakeup_thread(mddev->thread);
222}
223
224/*
225 * raid_end_bio_io() is called when we have finished servicing a mirrored
226 * operation and are ready to return a success/failure code to the buffer
227 * cache layer.
228 */
229static void raid_end_bio_io(r10bio_t *r10_bio)
230{
231 struct bio *bio = r10_bio->master_bio;
232 int done;
233 conf_t *conf = r10_bio->mddev->private;
234
235 if (bio->bi_phys_segments) {
236 unsigned long flags;
237 spin_lock_irqsave(&conf->device_lock, flags);
238 bio->bi_phys_segments--;
239 done = (bio->bi_phys_segments == 0);
240 spin_unlock_irqrestore(&conf->device_lock, flags);
241 } else
242 done = 1;
243 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
244 clear_bit(BIO_UPTODATE, &bio->bi_flags);
245 if (done) {
246 bio_endio(bio, 0);
247 /*
248 * Wake up any possible resync thread that waits for the device
249 * to go idle.
250 */
251 allow_barrier(conf);
252 }
253 free_r10bio(r10_bio);
254}
255
256/*
257 * Update disk head position estimator based on IRQ completion info.
258 */
259static inline void update_head_pos(int slot, r10bio_t *r10_bio)
260{
261 conf_t *conf = r10_bio->mddev->private;
262
263 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
264 r10_bio->devs[slot].addr + (r10_bio->sectors);
265}
266
267/*
268 * Find the disk number which triggered given bio
269 */
270static int find_bio_disk(conf_t *conf, r10bio_t *r10_bio,
271 struct bio *bio, int *slotp)
272{
273 int slot;
274
275 for (slot = 0; slot < conf->copies; slot++)
276 if (r10_bio->devs[slot].bio == bio)
277 break;
278
279 BUG_ON(slot == conf->copies);
280 update_head_pos(slot, r10_bio);
281
282 if (slotp)
283 *slotp = slot;
284 return r10_bio->devs[slot].devnum;
285}
286
287static void raid10_end_read_request(struct bio *bio, int error)
288{
289 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
290 r10bio_t *r10_bio = bio->bi_private;
291 int slot, dev;
292 conf_t *conf = r10_bio->mddev->private;
293
294
295 slot = r10_bio->read_slot;
296 dev = r10_bio->devs[slot].devnum;
297 /*
298 * this branch is our 'one mirror IO has finished' event handler:
299 */
300 update_head_pos(slot, r10_bio);
301
302 if (uptodate) {
303 /*
304 * Set R10BIO_Uptodate in our master bio, so that
305 * we will return a good error code to the higher
306 * levels even if IO on some other mirrored buffer fails.
307 *
308 * The 'master' represents the composite IO operation to
309 * user-side. So if something waits for IO, then it will
310 * wait for the 'master' bio.
311 */
312 set_bit(R10BIO_Uptodate, &r10_bio->state);
313 raid_end_bio_io(r10_bio);
314 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
315 } else {
316 /*
317 * oops, read error - keep the refcount on the rdev
318 */
319 char b[BDEVNAME_SIZE];
320 printk_ratelimited(KERN_ERR
321 "md/raid10:%s: %s: rescheduling sector %llu\n",
322 mdname(conf->mddev),
323 bdevname(conf->mirrors[dev].rdev->bdev, b),
324 (unsigned long long)r10_bio->sector);
325 set_bit(R10BIO_ReadError, &r10_bio->state);
326 reschedule_retry(r10_bio);
327 }
328}
329
330static void close_write(r10bio_t *r10_bio)
331{
332 /* clear the bitmap if all writes complete successfully */
333 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
334 r10_bio->sectors,
335 !test_bit(R10BIO_Degraded, &r10_bio->state),
336 0);
337 md_write_end(r10_bio->mddev);
338}
339
340static void one_write_done(r10bio_t *r10_bio)
341{
342 if (atomic_dec_and_test(&r10_bio->remaining)) {
343 if (test_bit(R10BIO_WriteError, &r10_bio->state))
344 reschedule_retry(r10_bio);
345 else {
346 close_write(r10_bio);
347 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
348 reschedule_retry(r10_bio);
349 else
350 raid_end_bio_io(r10_bio);
351 }
352 }
353}
354
355static void raid10_end_write_request(struct bio *bio, int error)
356{
357 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
358 r10bio_t *r10_bio = bio->bi_private;
359 int dev;
360 int dec_rdev = 1;
361 conf_t *conf = r10_bio->mddev->private;
362 int slot;
363
364 dev = find_bio_disk(conf, r10_bio, bio, &slot);
365
366 /*
367 * this branch is our 'one mirror IO has finished' event handler:
368 */
369 if (!uptodate) {
370 set_bit(WriteErrorSeen, &conf->mirrors[dev].rdev->flags);
371 set_bit(R10BIO_WriteError, &r10_bio->state);
372 dec_rdev = 0;
373 } else {
374 /*
375 * Set R10BIO_Uptodate in our master bio, so that
376 * we will return a good error code for to the higher
377 * levels even if IO on some other mirrored buffer fails.
378 *
379 * The 'master' represents the composite IO operation to
380 * user-side. So if something waits for IO, then it will
381 * wait for the 'master' bio.
382 */
383 sector_t first_bad;
384 int bad_sectors;
385
386 set_bit(R10BIO_Uptodate, &r10_bio->state);
387
388 /* Maybe we can clear some bad blocks. */
389 if (is_badblock(conf->mirrors[dev].rdev,
390 r10_bio->devs[slot].addr,
391 r10_bio->sectors,
392 &first_bad, &bad_sectors)) {
393 bio_put(bio);
394 r10_bio->devs[slot].bio = IO_MADE_GOOD;
395 dec_rdev = 0;
396 set_bit(R10BIO_MadeGood, &r10_bio->state);
397 }
398 }
399
400 /*
401 *
402 * Let's see if all mirrored write operations have finished
403 * already.
404 */
405 one_write_done(r10_bio);
406 if (dec_rdev)
407 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
408}
409
410
411/*
412 * RAID10 layout manager
413 * As well as the chunksize and raid_disks count, there are two
414 * parameters: near_copies and far_copies.
415 * near_copies * far_copies must be <= raid_disks.
416 * Normally one of these will be 1.
417 * If both are 1, we get raid0.
418 * If near_copies == raid_disks, we get raid1.
419 *
420 * Chunks are laid out in raid0 style with near_copies copies of the
421 * first chunk, followed by near_copies copies of the next chunk and
422 * so on.
423 * If far_copies > 1, then after 1/far_copies of the array has been assigned
424 * as described above, we start again with a device offset of near_copies.
425 * So we effectively have another copy of the whole array further down all
426 * the drives, but with blocks on different drives.
427 * With this layout, and block is never stored twice on the one device.
428 *
429 * raid10_find_phys finds the sector offset of a given virtual sector
430 * on each device that it is on.
431 *
432 * raid10_find_virt does the reverse mapping, from a device and a
433 * sector offset to a virtual address
434 */
435
436static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
437{
438 int n,f;
439 sector_t sector;
440 sector_t chunk;
441 sector_t stripe;
442 int dev;
443
444 int slot = 0;
445
446 /* now calculate first sector/dev */
447 chunk = r10bio->sector >> conf->chunk_shift;
448 sector = r10bio->sector & conf->chunk_mask;
449
450 chunk *= conf->near_copies;
451 stripe = chunk;
452 dev = sector_div(stripe, conf->raid_disks);
453 if (conf->far_offset)
454 stripe *= conf->far_copies;
455
456 sector += stripe << conf->chunk_shift;
457
458 /* and calculate all the others */
459 for (n=0; n < conf->near_copies; n++) {
460 int d = dev;
461 sector_t s = sector;
462 r10bio->devs[slot].addr = sector;
463 r10bio->devs[slot].devnum = d;
464 slot++;
465
466 for (f = 1; f < conf->far_copies; f++) {
467 d += conf->near_copies;
468 if (d >= conf->raid_disks)
469 d -= conf->raid_disks;
470 s += conf->stride;
471 r10bio->devs[slot].devnum = d;
472 r10bio->devs[slot].addr = s;
473 slot++;
474 }
475 dev++;
476 if (dev >= conf->raid_disks) {
477 dev = 0;
478 sector += (conf->chunk_mask + 1);
479 }
480 }
481 BUG_ON(slot != conf->copies);
482}
483
484static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
485{
486 sector_t offset, chunk, vchunk;
487
488 offset = sector & conf->chunk_mask;
489 if (conf->far_offset) {
490 int fc;
491 chunk = sector >> conf->chunk_shift;
492 fc = sector_div(chunk, conf->far_copies);
493 dev -= fc * conf->near_copies;
494 if (dev < 0)
495 dev += conf->raid_disks;
496 } else {
497 while (sector >= conf->stride) {
498 sector -= conf->stride;
499 if (dev < conf->near_copies)
500 dev += conf->raid_disks - conf->near_copies;
501 else
502 dev -= conf->near_copies;
503 }
504 chunk = sector >> conf->chunk_shift;
505 }
506 vchunk = chunk * conf->raid_disks + dev;
507 sector_div(vchunk, conf->near_copies);
508 return (vchunk << conf->chunk_shift) + offset;
509}
510
511/**
512 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
513 * @q: request queue
514 * @bvm: properties of new bio
515 * @biovec: the request that could be merged to it.
516 *
517 * Return amount of bytes we can accept at this offset
518 * If near_copies == raid_disk, there are no striping issues,
519 * but in that case, the function isn't called at all.
520 */
521static int raid10_mergeable_bvec(struct request_queue *q,
522 struct bvec_merge_data *bvm,
523 struct bio_vec *biovec)
524{
525 mddev_t *mddev = q->queuedata;
526 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
527 int max;
528 unsigned int chunk_sectors = mddev->chunk_sectors;
529 unsigned int bio_sectors = bvm->bi_size >> 9;
530
531 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
532 if (max < 0) max = 0; /* bio_add cannot handle a negative return */
533 if (max <= biovec->bv_len && bio_sectors == 0)
534 return biovec->bv_len;
535 else
536 return max;
537}
538
539/*
540 * This routine returns the disk from which the requested read should
541 * be done. There is a per-array 'next expected sequential IO' sector
542 * number - if this matches on the next IO then we use the last disk.
543 * There is also a per-disk 'last know head position' sector that is
544 * maintained from IRQ contexts, both the normal and the resync IO
545 * completion handlers update this position correctly. If there is no
546 * perfect sequential match then we pick the disk whose head is closest.
547 *
548 * If there are 2 mirrors in the same 2 devices, performance degrades
549 * because position is mirror, not device based.
550 *
551 * The rdev for the device selected will have nr_pending incremented.
552 */
553
554/*
555 * FIXME: possibly should rethink readbalancing and do it differently
556 * depending on near_copies / far_copies geometry.
557 */
558static int read_balance(conf_t *conf, r10bio_t *r10_bio, int *max_sectors)
559{
560 const sector_t this_sector = r10_bio->sector;
561 int disk, slot;
562 int sectors = r10_bio->sectors;
563 int best_good_sectors;
564 sector_t new_distance, best_dist;
565 mdk_rdev_t *rdev;
566 int do_balance;
567 int best_slot;
568
569 raid10_find_phys(conf, r10_bio);
570 rcu_read_lock();
571retry:
572 sectors = r10_bio->sectors;
573 best_slot = -1;
574 best_dist = MaxSector;
575 best_good_sectors = 0;
576 do_balance = 1;
577 /*
578 * Check if we can balance. We can balance on the whole
579 * device if no resync is going on (recovery is ok), or below
580 * the resync window. We take the first readable disk when
581 * above the resync window.
582 */
583 if (conf->mddev->recovery_cp < MaxSector
584 && (this_sector + sectors >= conf->next_resync))
585 do_balance = 0;
586
587 for (slot = 0; slot < conf->copies ; slot++) {
588 sector_t first_bad;
589 int bad_sectors;
590 sector_t dev_sector;
591
592 if (r10_bio->devs[slot].bio == IO_BLOCKED)
593 continue;
594 disk = r10_bio->devs[slot].devnum;
595 rdev = rcu_dereference(conf->mirrors[disk].rdev);
596 if (rdev == NULL)
597 continue;
598 if (!test_bit(In_sync, &rdev->flags))
599 continue;
600
601 dev_sector = r10_bio->devs[slot].addr;
602 if (is_badblock(rdev, dev_sector, sectors,
603 &first_bad, &bad_sectors)) {
604 if (best_dist < MaxSector)
605 /* Already have a better slot */
606 continue;
607 if (first_bad <= dev_sector) {
608 /* Cannot read here. If this is the
609 * 'primary' device, then we must not read
610 * beyond 'bad_sectors' from another device.
611 */
612 bad_sectors -= (dev_sector - first_bad);
613 if (!do_balance && sectors > bad_sectors)
614 sectors = bad_sectors;
615 if (best_good_sectors > sectors)
616 best_good_sectors = sectors;
617 } else {
618 sector_t good_sectors =
619 first_bad - dev_sector;
620 if (good_sectors > best_good_sectors) {
621 best_good_sectors = good_sectors;
622 best_slot = slot;
623 }
624 if (!do_balance)
625 /* Must read from here */
626 break;
627 }
628 continue;
629 } else
630 best_good_sectors = sectors;
631
632 if (!do_balance)
633 break;
634
635 /* This optimisation is debatable, and completely destroys
636 * sequential read speed for 'far copies' arrays. So only
637 * keep it for 'near' arrays, and review those later.
638 */
639 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending))
640 break;
641
642 /* for far > 1 always use the lowest address */
643 if (conf->far_copies > 1)
644 new_distance = r10_bio->devs[slot].addr;
645 else
646 new_distance = abs(r10_bio->devs[slot].addr -
647 conf->mirrors[disk].head_position);
648 if (new_distance < best_dist) {
649 best_dist = new_distance;
650 best_slot = slot;
651 }
652 }
653 if (slot == conf->copies)
654 slot = best_slot;
655
656 if (slot >= 0) {
657 disk = r10_bio->devs[slot].devnum;
658 rdev = rcu_dereference(conf->mirrors[disk].rdev);
659 if (!rdev)
660 goto retry;
661 atomic_inc(&rdev->nr_pending);
662 if (test_bit(Faulty, &rdev->flags)) {
663 /* Cannot risk returning a device that failed
664 * before we inc'ed nr_pending
665 */
666 rdev_dec_pending(rdev, conf->mddev);
667 goto retry;
668 }
669 r10_bio->read_slot = slot;
670 } else
671 disk = -1;
672 rcu_read_unlock();
673 *max_sectors = best_good_sectors;
674
675 return disk;
676}
677
678static int raid10_congested(void *data, int bits)
679{
680 mddev_t *mddev = data;
681 conf_t *conf = mddev->private;
682 int i, ret = 0;
683
684 if (mddev_congested(mddev, bits))
685 return 1;
686 rcu_read_lock();
687 for (i = 0; i < conf->raid_disks && ret == 0; i++) {
688 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
689 if (rdev && !test_bit(Faulty, &rdev->flags)) {
690 struct request_queue *q = bdev_get_queue(rdev->bdev);
691
692 ret |= bdi_congested(&q->backing_dev_info, bits);
693 }
694 }
695 rcu_read_unlock();
696 return ret;
697}
698
699static void flush_pending_writes(conf_t *conf)
700{
701 /* Any writes that have been queued but are awaiting
702 * bitmap updates get flushed here.
703 */
704 spin_lock_irq(&conf->device_lock);
705
706 if (conf->pending_bio_list.head) {
707 struct bio *bio;
708 bio = bio_list_get(&conf->pending_bio_list);
709 spin_unlock_irq(&conf->device_lock);
710 /* flush any pending bitmap writes to disk
711 * before proceeding w/ I/O */
712 bitmap_unplug(conf->mddev->bitmap);
713
714 while (bio) { /* submit pending writes */
715 struct bio *next = bio->bi_next;
716 bio->bi_next = NULL;
717 generic_make_request(bio);
718 bio = next;
719 }
720 } else
721 spin_unlock_irq(&conf->device_lock);
722}
723
724/* Barriers....
725 * Sometimes we need to suspend IO while we do something else,
726 * either some resync/recovery, or reconfigure the array.
727 * To do this we raise a 'barrier'.
728 * The 'barrier' is a counter that can be raised multiple times
729 * to count how many activities are happening which preclude
730 * normal IO.
731 * We can only raise the barrier if there is no pending IO.
732 * i.e. if nr_pending == 0.
733 * We choose only to raise the barrier if no-one is waiting for the
734 * barrier to go down. This means that as soon as an IO request
735 * is ready, no other operations which require a barrier will start
736 * until the IO request has had a chance.
737 *
738 * So: regular IO calls 'wait_barrier'. When that returns there
739 * is no backgroup IO happening, It must arrange to call
740 * allow_barrier when it has finished its IO.
741 * backgroup IO calls must call raise_barrier. Once that returns
742 * there is no normal IO happeing. It must arrange to call
743 * lower_barrier when the particular background IO completes.
744 */
745
746static void raise_barrier(conf_t *conf, int force)
747{
748 BUG_ON(force && !conf->barrier);
749 spin_lock_irq(&conf->resync_lock);
750
751 /* Wait until no block IO is waiting (unless 'force') */
752 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
753 conf->resync_lock, );
754
755 /* block any new IO from starting */
756 conf->barrier++;
757
758 /* Now wait for all pending IO to complete */
759 wait_event_lock_irq(conf->wait_barrier,
760 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
761 conf->resync_lock, );
762
763 spin_unlock_irq(&conf->resync_lock);
764}
765
766static void lower_barrier(conf_t *conf)
767{
768 unsigned long flags;
769 spin_lock_irqsave(&conf->resync_lock, flags);
770 conf->barrier--;
771 spin_unlock_irqrestore(&conf->resync_lock, flags);
772 wake_up(&conf->wait_barrier);
773}
774
775static void wait_barrier(conf_t *conf)
776{
777 spin_lock_irq(&conf->resync_lock);
778 if (conf->barrier) {
779 conf->nr_waiting++;
780 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
781 conf->resync_lock,
782 );
783 conf->nr_waiting--;
784 }
785 conf->nr_pending++;
786 spin_unlock_irq(&conf->resync_lock);
787}
788
789static void allow_barrier(conf_t *conf)
790{
791 unsigned long flags;
792 spin_lock_irqsave(&conf->resync_lock, flags);
793 conf->nr_pending--;
794 spin_unlock_irqrestore(&conf->resync_lock, flags);
795 wake_up(&conf->wait_barrier);
796}
797
798static void freeze_array(conf_t *conf)
799{
800 /* stop syncio and normal IO and wait for everything to
801 * go quiet.
802 * We increment barrier and nr_waiting, and then
803 * wait until nr_pending match nr_queued+1
804 * This is called in the context of one normal IO request
805 * that has failed. Thus any sync request that might be pending
806 * will be blocked by nr_pending, and we need to wait for
807 * pending IO requests to complete or be queued for re-try.
808 * Thus the number queued (nr_queued) plus this request (1)
809 * must match the number of pending IOs (nr_pending) before
810 * we continue.
811 */
812 spin_lock_irq(&conf->resync_lock);
813 conf->barrier++;
814 conf->nr_waiting++;
815 wait_event_lock_irq(conf->wait_barrier,
816 conf->nr_pending == conf->nr_queued+1,
817 conf->resync_lock,
818 flush_pending_writes(conf));
819
820 spin_unlock_irq(&conf->resync_lock);
821}
822
823static void unfreeze_array(conf_t *conf)
824{
825 /* reverse the effect of the freeze */
826 spin_lock_irq(&conf->resync_lock);
827 conf->barrier--;
828 conf->nr_waiting--;
829 wake_up(&conf->wait_barrier);
830 spin_unlock_irq(&conf->resync_lock);
831}
832
833static int make_request(mddev_t *mddev, struct bio * bio)
834{
835 conf_t *conf = mddev->private;
836 mirror_info_t *mirror;
837 r10bio_t *r10_bio;
838 struct bio *read_bio;
839 int i;
840 int chunk_sects = conf->chunk_mask + 1;
841 const int rw = bio_data_dir(bio);
842 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
843 const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
844 unsigned long flags;
845 mdk_rdev_t *blocked_rdev;
846 int plugged;
847 int sectors_handled;
848 int max_sectors;
849
850 if (unlikely(bio->bi_rw & REQ_FLUSH)) {
851 md_flush_request(mddev, bio);
852 return 0;
853 }
854
855 /* If this request crosses a chunk boundary, we need to
856 * split it. This will only happen for 1 PAGE (or less) requests.
857 */
858 if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
859 > chunk_sects &&
860 conf->near_copies < conf->raid_disks)) {
861 struct bio_pair *bp;
862 /* Sanity check -- queue functions should prevent this happening */
863 if (bio->bi_vcnt != 1 ||
864 bio->bi_idx != 0)
865 goto bad_map;
866 /* This is a one page bio that upper layers
867 * refuse to split for us, so we need to split it.
868 */
869 bp = bio_split(bio,
870 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
871
872 /* Each of these 'make_request' calls will call 'wait_barrier'.
873 * If the first succeeds but the second blocks due to the resync
874 * thread raising the barrier, we will deadlock because the
875 * IO to the underlying device will be queued in generic_make_request
876 * and will never complete, so will never reduce nr_pending.
877 * So increment nr_waiting here so no new raise_barriers will
878 * succeed, and so the second wait_barrier cannot block.
879 */
880 spin_lock_irq(&conf->resync_lock);
881 conf->nr_waiting++;
882 spin_unlock_irq(&conf->resync_lock);
883
884 if (make_request(mddev, &bp->bio1))
885 generic_make_request(&bp->bio1);
886 if (make_request(mddev, &bp->bio2))
887 generic_make_request(&bp->bio2);
888
889 spin_lock_irq(&conf->resync_lock);
890 conf->nr_waiting--;
891 wake_up(&conf->wait_barrier);
892 spin_unlock_irq(&conf->resync_lock);
893
894 bio_pair_release(bp);
895 return 0;
896 bad_map:
897 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
898 " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
899 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
900
901 bio_io_error(bio);
902 return 0;
903 }
904
905 md_write_start(mddev, bio);
906
907 /*
908 * Register the new request and wait if the reconstruction
909 * thread has put up a bar for new requests.
910 * Continue immediately if no resync is active currently.
911 */
912 wait_barrier(conf);
913
914 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
915
916 r10_bio->master_bio = bio;
917 r10_bio->sectors = bio->bi_size >> 9;
918
919 r10_bio->mddev = mddev;
920 r10_bio->sector = bio->bi_sector;
921 r10_bio->state = 0;
922
923 /* We might need to issue multiple reads to different
924 * devices if there are bad blocks around, so we keep
925 * track of the number of reads in bio->bi_phys_segments.
926 * If this is 0, there is only one r10_bio and no locking
927 * will be needed when the request completes. If it is
928 * non-zero, then it is the number of not-completed requests.
929 */
930 bio->bi_phys_segments = 0;
931 clear_bit(BIO_SEG_VALID, &bio->bi_flags);
932
933 if (rw == READ) {
934 /*
935 * read balancing logic:
936 */
937 int disk;
938 int slot;
939
940read_again:
941 disk = read_balance(conf, r10_bio, &max_sectors);
942 slot = r10_bio->read_slot;
943 if (disk < 0) {
944 raid_end_bio_io(r10_bio);
945 return 0;
946 }
947 mirror = conf->mirrors + disk;
948
949 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
950 md_trim_bio(read_bio, r10_bio->sector - bio->bi_sector,
951 max_sectors);
952
953 r10_bio->devs[slot].bio = read_bio;
954
955 read_bio->bi_sector = r10_bio->devs[slot].addr +
956 mirror->rdev->data_offset;
957 read_bio->bi_bdev = mirror->rdev->bdev;
958 read_bio->bi_end_io = raid10_end_read_request;
959 read_bio->bi_rw = READ | do_sync;
960 read_bio->bi_private = r10_bio;
961
962 if (max_sectors < r10_bio->sectors) {
963 /* Could not read all from this device, so we will
964 * need another r10_bio.
965 */
966 sectors_handled = (r10_bio->sectors + max_sectors
967 - bio->bi_sector);
968 r10_bio->sectors = max_sectors;
969 spin_lock_irq(&conf->device_lock);
970 if (bio->bi_phys_segments == 0)
971 bio->bi_phys_segments = 2;
972 else
973 bio->bi_phys_segments++;
974 spin_unlock(&conf->device_lock);
975 /* Cannot call generic_make_request directly
976 * as that will be queued in __generic_make_request
977 * and subsequent mempool_alloc might block
978 * waiting for it. so hand bio over to raid10d.
979 */
980 reschedule_retry(r10_bio);
981
982 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
983
984 r10_bio->master_bio = bio;
985 r10_bio->sectors = ((bio->bi_size >> 9)
986 - sectors_handled);
987 r10_bio->state = 0;
988 r10_bio->mddev = mddev;
989 r10_bio->sector = bio->bi_sector + sectors_handled;
990 goto read_again;
991 } else
992 generic_make_request(read_bio);
993 return 0;
994 }
995
996 /*
997 * WRITE:
998 */
999 /* first select target devices under rcu_lock and
1000 * inc refcount on their rdev. Record them by setting
1001 * bios[x] to bio
1002 * If there are known/acknowledged bad blocks on any device
1003 * on which we have seen a write error, we want to avoid
1004 * writing to those blocks. This potentially requires several
1005 * writes to write around the bad blocks. Each set of writes
1006 * gets its own r10_bio with a set of bios attached. The number
1007 * of r10_bios is recored in bio->bi_phys_segments just as with
1008 * the read case.
1009 */
1010 plugged = mddev_check_plugged(mddev);
1011
1012 raid10_find_phys(conf, r10_bio);
1013retry_write:
1014 blocked_rdev = NULL;
1015 rcu_read_lock();
1016 max_sectors = r10_bio->sectors;
1017
1018 for (i = 0; i < conf->copies; i++) {
1019 int d = r10_bio->devs[i].devnum;
1020 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
1021 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1022 atomic_inc(&rdev->nr_pending);
1023 blocked_rdev = rdev;
1024 break;
1025 }
1026 r10_bio->devs[i].bio = NULL;
1027 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1028 set_bit(R10BIO_Degraded, &r10_bio->state);
1029 continue;
1030 }
1031 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1032 sector_t first_bad;
1033 sector_t dev_sector = r10_bio->devs[i].addr;
1034 int bad_sectors;
1035 int is_bad;
1036
1037 is_bad = is_badblock(rdev, dev_sector,
1038 max_sectors,
1039 &first_bad, &bad_sectors);
1040 if (is_bad < 0) {
1041 /* Mustn't write here until the bad block
1042 * is acknowledged
1043 */
1044 atomic_inc(&rdev->nr_pending);
1045 set_bit(BlockedBadBlocks, &rdev->flags);
1046 blocked_rdev = rdev;
1047 break;
1048 }
1049 if (is_bad && first_bad <= dev_sector) {
1050 /* Cannot write here at all */
1051 bad_sectors -= (dev_sector - first_bad);
1052 if (bad_sectors < max_sectors)
1053 /* Mustn't write more than bad_sectors
1054 * to other devices yet
1055 */
1056 max_sectors = bad_sectors;
1057 /* We don't set R10BIO_Degraded as that
1058 * only applies if the disk is missing,
1059 * so it might be re-added, and we want to
1060 * know to recover this chunk.
1061 * In this case the device is here, and the
1062 * fact that this chunk is not in-sync is
1063 * recorded in the bad block log.
1064 */
1065 continue;
1066 }
1067 if (is_bad) {
1068 int good_sectors = first_bad - dev_sector;
1069 if (good_sectors < max_sectors)
1070 max_sectors = good_sectors;
1071 }
1072 }
1073 r10_bio->devs[i].bio = bio;
1074 atomic_inc(&rdev->nr_pending);
1075 }
1076 rcu_read_unlock();
1077
1078 if (unlikely(blocked_rdev)) {
1079 /* Have to wait for this device to get unblocked, then retry */
1080 int j;
1081 int d;
1082
1083 for (j = 0; j < i; j++)
1084 if (r10_bio->devs[j].bio) {
1085 d = r10_bio->devs[j].devnum;
1086 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1087 }
1088 allow_barrier(conf);
1089 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1090 wait_barrier(conf);
1091 goto retry_write;
1092 }
1093
1094 if (max_sectors < r10_bio->sectors) {
1095 /* We are splitting this into multiple parts, so
1096 * we need to prepare for allocating another r10_bio.
1097 */
1098 r10_bio->sectors = max_sectors;
1099 spin_lock_irq(&conf->device_lock);
1100 if (bio->bi_phys_segments == 0)
1101 bio->bi_phys_segments = 2;
1102 else
1103 bio->bi_phys_segments++;
1104 spin_unlock_irq(&conf->device_lock);
1105 }
1106 sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector;
1107
1108 atomic_set(&r10_bio->remaining, 1);
1109 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1110
1111 for (i = 0; i < conf->copies; i++) {
1112 struct bio *mbio;
1113 int d = r10_bio->devs[i].devnum;
1114 if (!r10_bio->devs[i].bio)
1115 continue;
1116
1117 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1118 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1119 max_sectors);
1120 r10_bio->devs[i].bio = mbio;
1121
1122 mbio->bi_sector = (r10_bio->devs[i].addr+
1123 conf->mirrors[d].rdev->data_offset);
1124 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1125 mbio->bi_end_io = raid10_end_write_request;
1126 mbio->bi_rw = WRITE | do_sync | do_fua;
1127 mbio->bi_private = r10_bio;
1128
1129 atomic_inc(&r10_bio->remaining);
1130 spin_lock_irqsave(&conf->device_lock, flags);
1131 bio_list_add(&conf->pending_bio_list, mbio);
1132 spin_unlock_irqrestore(&conf->device_lock, flags);
1133 }
1134
1135 /* Don't remove the bias on 'remaining' (one_write_done) until
1136 * after checking if we need to go around again.
1137 */
1138
1139 if (sectors_handled < (bio->bi_size >> 9)) {
1140 one_write_done(r10_bio);
1141 /* We need another r10_bio. It has already been counted
1142 * in bio->bi_phys_segments.
1143 */
1144 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1145
1146 r10_bio->master_bio = bio;
1147 r10_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1148
1149 r10_bio->mddev = mddev;
1150 r10_bio->sector = bio->bi_sector + sectors_handled;
1151 r10_bio->state = 0;
1152 goto retry_write;
1153 }
1154 one_write_done(r10_bio);
1155
1156 /* In case raid10d snuck in to freeze_array */
1157 wake_up(&conf->wait_barrier);
1158
1159 if (do_sync || !mddev->bitmap || !plugged)
1160 md_wakeup_thread(mddev->thread);
1161 return 0;
1162}
1163
1164static void status(struct seq_file *seq, mddev_t *mddev)
1165{
1166 conf_t *conf = mddev->private;
1167 int i;
1168
1169 if (conf->near_copies < conf->raid_disks)
1170 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1171 if (conf->near_copies > 1)
1172 seq_printf(seq, " %d near-copies", conf->near_copies);
1173 if (conf->far_copies > 1) {
1174 if (conf->far_offset)
1175 seq_printf(seq, " %d offset-copies", conf->far_copies);
1176 else
1177 seq_printf(seq, " %d far-copies", conf->far_copies);
1178 }
1179 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1180 conf->raid_disks - mddev->degraded);
1181 for (i = 0; i < conf->raid_disks; i++)
1182 seq_printf(seq, "%s",
1183 conf->mirrors[i].rdev &&
1184 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1185 seq_printf(seq, "]");
1186}
1187
1188/* check if there are enough drives for
1189 * every block to appear on atleast one.
1190 * Don't consider the device numbered 'ignore'
1191 * as we might be about to remove it.
1192 */
1193static int enough(conf_t *conf, int ignore)
1194{
1195 int first = 0;
1196
1197 do {
1198 int n = conf->copies;
1199 int cnt = 0;
1200 while (n--) {
1201 if (conf->mirrors[first].rdev &&
1202 first != ignore)
1203 cnt++;
1204 first = (first+1) % conf->raid_disks;
1205 }
1206 if (cnt == 0)
1207 return 0;
1208 } while (first != 0);
1209 return 1;
1210}
1211
1212static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1213{
1214 char b[BDEVNAME_SIZE];
1215 conf_t *conf = mddev->private;
1216
1217 /*
1218 * If it is not operational, then we have already marked it as dead
1219 * else if it is the last working disks, ignore the error, let the
1220 * next level up know.
1221 * else mark the drive as failed
1222 */
1223 if (test_bit(In_sync, &rdev->flags)
1224 && !enough(conf, rdev->raid_disk))
1225 /*
1226 * Don't fail the drive, just return an IO error.
1227 */
1228 return;
1229 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1230 unsigned long flags;
1231 spin_lock_irqsave(&conf->device_lock, flags);
1232 mddev->degraded++;
1233 spin_unlock_irqrestore(&conf->device_lock, flags);
1234 /*
1235 * if recovery is running, make sure it aborts.
1236 */
1237 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1238 }
1239 set_bit(Blocked, &rdev->flags);
1240 set_bit(Faulty, &rdev->flags);
1241 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1242 printk(KERN_ALERT
1243 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1244 "md/raid10:%s: Operation continuing on %d devices.\n",
1245 mdname(mddev), bdevname(rdev->bdev, b),
1246 mdname(mddev), conf->raid_disks - mddev->degraded);
1247}
1248
1249static void print_conf(conf_t *conf)
1250{
1251 int i;
1252 mirror_info_t *tmp;
1253
1254 printk(KERN_DEBUG "RAID10 conf printout:\n");
1255 if (!conf) {
1256 printk(KERN_DEBUG "(!conf)\n");
1257 return;
1258 }
1259 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1260 conf->raid_disks);
1261
1262 for (i = 0; i < conf->raid_disks; i++) {
1263 char b[BDEVNAME_SIZE];
1264 tmp = conf->mirrors + i;
1265 if (tmp->rdev)
1266 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1267 i, !test_bit(In_sync, &tmp->rdev->flags),
1268 !test_bit(Faulty, &tmp->rdev->flags),
1269 bdevname(tmp->rdev->bdev,b));
1270 }
1271}
1272
1273static void close_sync(conf_t *conf)
1274{
1275 wait_barrier(conf);
1276 allow_barrier(conf);
1277
1278 mempool_destroy(conf->r10buf_pool);
1279 conf->r10buf_pool = NULL;
1280}
1281
1282static int raid10_spare_active(mddev_t *mddev)
1283{
1284 int i;
1285 conf_t *conf = mddev->private;
1286 mirror_info_t *tmp;
1287 int count = 0;
1288 unsigned long flags;
1289
1290 /*
1291 * Find all non-in_sync disks within the RAID10 configuration
1292 * and mark them in_sync
1293 */
1294 for (i = 0; i < conf->raid_disks; i++) {
1295 tmp = conf->mirrors + i;
1296 if (tmp->rdev
1297 && !test_bit(Faulty, &tmp->rdev->flags)
1298 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1299 count++;
1300 sysfs_notify_dirent(tmp->rdev->sysfs_state);
1301 }
1302 }
1303 spin_lock_irqsave(&conf->device_lock, flags);
1304 mddev->degraded -= count;
1305 spin_unlock_irqrestore(&conf->device_lock, flags);
1306
1307 print_conf(conf);
1308 return count;
1309}
1310
1311
1312static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1313{
1314 conf_t *conf = mddev->private;
1315 int err = -EEXIST;
1316 int mirror;
1317 int first = 0;
1318 int last = conf->raid_disks - 1;
1319
1320 if (mddev->recovery_cp < MaxSector)
1321 /* only hot-add to in-sync arrays, as recovery is
1322 * very different from resync
1323 */
1324 return -EBUSY;
1325 if (!enough(conf, -1))
1326 return -EINVAL;
1327
1328 if (rdev->raid_disk >= 0)
1329 first = last = rdev->raid_disk;
1330
1331 if (rdev->saved_raid_disk >= first &&
1332 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1333 mirror = rdev->saved_raid_disk;
1334 else
1335 mirror = first;
1336 for ( ; mirror <= last ; mirror++) {
1337 mirror_info_t *p = &conf->mirrors[mirror];
1338 if (p->recovery_disabled == mddev->recovery_disabled)
1339 continue;
1340 if (!p->rdev)
1341 continue;
1342
1343 disk_stack_limits(mddev->gendisk, rdev->bdev,
1344 rdev->data_offset << 9);
1345 /* as we don't honour merge_bvec_fn, we must
1346 * never risk violating it, so limit
1347 * ->max_segments to one lying with a single
1348 * page, as a one page request is never in
1349 * violation.
1350 */
1351 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1352 blk_queue_max_segments(mddev->queue, 1);
1353 blk_queue_segment_boundary(mddev->queue,
1354 PAGE_CACHE_SIZE - 1);
1355 }
1356
1357 p->head_position = 0;
1358 rdev->raid_disk = mirror;
1359 err = 0;
1360 if (rdev->saved_raid_disk != mirror)
1361 conf->fullsync = 1;
1362 rcu_assign_pointer(p->rdev, rdev);
1363 break;
1364 }
1365
1366 md_integrity_add_rdev(rdev, mddev);
1367 print_conf(conf);
1368 return err;
1369}
1370
1371static int raid10_remove_disk(mddev_t *mddev, int number)
1372{
1373 conf_t *conf = mddev->private;
1374 int err = 0;
1375 mdk_rdev_t *rdev;
1376 mirror_info_t *p = conf->mirrors+ number;
1377
1378 print_conf(conf);
1379 rdev = p->rdev;
1380 if (rdev) {
1381 if (test_bit(In_sync, &rdev->flags) ||
1382 atomic_read(&rdev->nr_pending)) {
1383 err = -EBUSY;
1384 goto abort;
1385 }
1386 /* Only remove faulty devices in recovery
1387 * is not possible.
1388 */
1389 if (!test_bit(Faulty, &rdev->flags) &&
1390 mddev->recovery_disabled != p->recovery_disabled &&
1391 enough(conf, -1)) {
1392 err = -EBUSY;
1393 goto abort;
1394 }
1395 p->rdev = NULL;
1396 synchronize_rcu();
1397 if (atomic_read(&rdev->nr_pending)) {
1398 /* lost the race, try later */
1399 err = -EBUSY;
1400 p->rdev = rdev;
1401 goto abort;
1402 }
1403 err = md_integrity_register(mddev);
1404 }
1405abort:
1406
1407 print_conf(conf);
1408 return err;
1409}
1410
1411
1412static void end_sync_read(struct bio *bio, int error)
1413{
1414 r10bio_t *r10_bio = bio->bi_private;
1415 conf_t *conf = r10_bio->mddev->private;
1416 int d;
1417
1418 d = find_bio_disk(conf, r10_bio, bio, NULL);
1419
1420 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1421 set_bit(R10BIO_Uptodate, &r10_bio->state);
1422 else
1423 /* The write handler will notice the lack of
1424 * R10BIO_Uptodate and record any errors etc
1425 */
1426 atomic_add(r10_bio->sectors,
1427 &conf->mirrors[d].rdev->corrected_errors);
1428
1429 /* for reconstruct, we always reschedule after a read.
1430 * for resync, only after all reads
1431 */
1432 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1433 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1434 atomic_dec_and_test(&r10_bio->remaining)) {
1435 /* we have read all the blocks,
1436 * do the comparison in process context in raid10d
1437 */
1438 reschedule_retry(r10_bio);
1439 }
1440}
1441
1442static void end_sync_request(r10bio_t *r10_bio)
1443{
1444 mddev_t *mddev = r10_bio->mddev;
1445
1446 while (atomic_dec_and_test(&r10_bio->remaining)) {
1447 if (r10_bio->master_bio == NULL) {
1448 /* the primary of several recovery bios */
1449 sector_t s = r10_bio->sectors;
1450 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1451 test_bit(R10BIO_WriteError, &r10_bio->state))
1452 reschedule_retry(r10_bio);
1453 else
1454 put_buf(r10_bio);
1455 md_done_sync(mddev, s, 1);
1456 break;
1457 } else {
1458 r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1459 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1460 test_bit(R10BIO_WriteError, &r10_bio->state))
1461 reschedule_retry(r10_bio);
1462 else
1463 put_buf(r10_bio);
1464 r10_bio = r10_bio2;
1465 }
1466 }
1467}
1468
1469static void end_sync_write(struct bio *bio, int error)
1470{
1471 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1472 r10bio_t *r10_bio = bio->bi_private;
1473 mddev_t *mddev = r10_bio->mddev;
1474 conf_t *conf = mddev->private;
1475 int d;
1476 sector_t first_bad;
1477 int bad_sectors;
1478 int slot;
1479
1480 d = find_bio_disk(conf, r10_bio, bio, &slot);
1481
1482 if (!uptodate) {
1483 set_bit(WriteErrorSeen, &conf->mirrors[d].rdev->flags);
1484 set_bit(R10BIO_WriteError, &r10_bio->state);
1485 } else if (is_badblock(conf->mirrors[d].rdev,
1486 r10_bio->devs[slot].addr,
1487 r10_bio->sectors,
1488 &first_bad, &bad_sectors))
1489 set_bit(R10BIO_MadeGood, &r10_bio->state);
1490
1491 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1492
1493 end_sync_request(r10_bio);
1494}
1495
1496/*
1497 * Note: sync and recover and handled very differently for raid10
1498 * This code is for resync.
1499 * For resync, we read through virtual addresses and read all blocks.
1500 * If there is any error, we schedule a write. The lowest numbered
1501 * drive is authoritative.
1502 * However requests come for physical address, so we need to map.
1503 * For every physical address there are raid_disks/copies virtual addresses,
1504 * which is always are least one, but is not necessarly an integer.
1505 * This means that a physical address can span multiple chunks, so we may
1506 * have to submit multiple io requests for a single sync request.
1507 */
1508/*
1509 * We check if all blocks are in-sync and only write to blocks that
1510 * aren't in sync
1511 */
1512static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1513{
1514 conf_t *conf = mddev->private;
1515 int i, first;
1516 struct bio *tbio, *fbio;
1517
1518 atomic_set(&r10_bio->remaining, 1);
1519
1520 /* find the first device with a block */
1521 for (i=0; i<conf->copies; i++)
1522 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1523 break;
1524
1525 if (i == conf->copies)
1526 goto done;
1527
1528 first = i;
1529 fbio = r10_bio->devs[i].bio;
1530
1531 /* now find blocks with errors */
1532 for (i=0 ; i < conf->copies ; i++) {
1533 int j, d;
1534 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1535
1536 tbio = r10_bio->devs[i].bio;
1537
1538 if (tbio->bi_end_io != end_sync_read)
1539 continue;
1540 if (i == first)
1541 continue;
1542 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1543 /* We know that the bi_io_vec layout is the same for
1544 * both 'first' and 'i', so we just compare them.
1545 * All vec entries are PAGE_SIZE;
1546 */
1547 for (j = 0; j < vcnt; j++)
1548 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1549 page_address(tbio->bi_io_vec[j].bv_page),
1550 PAGE_SIZE))
1551 break;
1552 if (j == vcnt)
1553 continue;
1554 mddev->resync_mismatches += r10_bio->sectors;
1555 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1556 /* Don't fix anything. */
1557 continue;
1558 }
1559 /* Ok, we need to write this bio, either to correct an
1560 * inconsistency or to correct an unreadable block.
1561 * First we need to fixup bv_offset, bv_len and
1562 * bi_vecs, as the read request might have corrupted these
1563 */
1564 tbio->bi_vcnt = vcnt;
1565 tbio->bi_size = r10_bio->sectors << 9;
1566 tbio->bi_idx = 0;
1567 tbio->bi_phys_segments = 0;
1568 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1569 tbio->bi_flags |= 1 << BIO_UPTODATE;
1570 tbio->bi_next = NULL;
1571 tbio->bi_rw = WRITE;
1572 tbio->bi_private = r10_bio;
1573 tbio->bi_sector = r10_bio->devs[i].addr;
1574
1575 for (j=0; j < vcnt ; j++) {
1576 tbio->bi_io_vec[j].bv_offset = 0;
1577 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1578
1579 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1580 page_address(fbio->bi_io_vec[j].bv_page),
1581 PAGE_SIZE);
1582 }
1583 tbio->bi_end_io = end_sync_write;
1584
1585 d = r10_bio->devs[i].devnum;
1586 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1587 atomic_inc(&r10_bio->remaining);
1588 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1589
1590 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1591 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1592 generic_make_request(tbio);
1593 }
1594
1595done:
1596 if (atomic_dec_and_test(&r10_bio->remaining)) {
1597 md_done_sync(mddev, r10_bio->sectors, 1);
1598 put_buf(r10_bio);
1599 }
1600}
1601
1602/*
1603 * Now for the recovery code.
1604 * Recovery happens across physical sectors.
1605 * We recover all non-is_sync drives by finding the virtual address of
1606 * each, and then choose a working drive that also has that virt address.
1607 * There is a separate r10_bio for each non-in_sync drive.
1608 * Only the first two slots are in use. The first for reading,
1609 * The second for writing.
1610 *
1611 */
1612static void fix_recovery_read_error(r10bio_t *r10_bio)
1613{
1614 /* We got a read error during recovery.
1615 * We repeat the read in smaller page-sized sections.
1616 * If a read succeeds, write it to the new device or record
1617 * a bad block if we cannot.
1618 * If a read fails, record a bad block on both old and
1619 * new devices.
1620 */
1621 mddev_t *mddev = r10_bio->mddev;
1622 conf_t *conf = mddev->private;
1623 struct bio *bio = r10_bio->devs[0].bio;
1624 sector_t sect = 0;
1625 int sectors = r10_bio->sectors;
1626 int idx = 0;
1627 int dr = r10_bio->devs[0].devnum;
1628 int dw = r10_bio->devs[1].devnum;
1629
1630 while (sectors) {
1631 int s = sectors;
1632 mdk_rdev_t *rdev;
1633 sector_t addr;
1634 int ok;
1635
1636 if (s > (PAGE_SIZE>>9))
1637 s = PAGE_SIZE >> 9;
1638
1639 rdev = conf->mirrors[dr].rdev;
1640 addr = r10_bio->devs[0].addr + sect,
1641 ok = sync_page_io(rdev,
1642 addr,
1643 s << 9,
1644 bio->bi_io_vec[idx].bv_page,
1645 READ, false);
1646 if (ok) {
1647 rdev = conf->mirrors[dw].rdev;
1648 addr = r10_bio->devs[1].addr + sect;
1649 ok = sync_page_io(rdev,
1650 addr,
1651 s << 9,
1652 bio->bi_io_vec[idx].bv_page,
1653 WRITE, false);
1654 if (!ok)
1655 set_bit(WriteErrorSeen, &rdev->flags);
1656 }
1657 if (!ok) {
1658 /* We don't worry if we cannot set a bad block -
1659 * it really is bad so there is no loss in not
1660 * recording it yet
1661 */
1662 rdev_set_badblocks(rdev, addr, s, 0);
1663
1664 if (rdev != conf->mirrors[dw].rdev) {
1665 /* need bad block on destination too */
1666 mdk_rdev_t *rdev2 = conf->mirrors[dw].rdev;
1667 addr = r10_bio->devs[1].addr + sect;
1668 ok = rdev_set_badblocks(rdev2, addr, s, 0);
1669 if (!ok) {
1670 /* just abort the recovery */
1671 printk(KERN_NOTICE
1672 "md/raid10:%s: recovery aborted"
1673 " due to read error\n",
1674 mdname(mddev));
1675
1676 conf->mirrors[dw].recovery_disabled
1677 = mddev->recovery_disabled;
1678 set_bit(MD_RECOVERY_INTR,
1679 &mddev->recovery);
1680 break;
1681 }
1682 }
1683 }
1684
1685 sectors -= s;
1686 sect += s;
1687 idx++;
1688 }
1689}
1690
1691static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1692{
1693 conf_t *conf = mddev->private;
1694 int d;
1695 struct bio *wbio;
1696
1697 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
1698 fix_recovery_read_error(r10_bio);
1699 end_sync_request(r10_bio);
1700 return;
1701 }
1702
1703 /*
1704 * share the pages with the first bio
1705 * and submit the write request
1706 */
1707 wbio = r10_bio->devs[1].bio;
1708 d = r10_bio->devs[1].devnum;
1709
1710 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1711 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1712 generic_make_request(wbio);
1713}
1714
1715
1716/*
1717 * Used by fix_read_error() to decay the per rdev read_errors.
1718 * We halve the read error count for every hour that has elapsed
1719 * since the last recorded read error.
1720 *
1721 */
1722static void check_decay_read_errors(mddev_t *mddev, mdk_rdev_t *rdev)
1723{
1724 struct timespec cur_time_mon;
1725 unsigned long hours_since_last;
1726 unsigned int read_errors = atomic_read(&rdev->read_errors);
1727
1728 ktime_get_ts(&cur_time_mon);
1729
1730 if (rdev->last_read_error.tv_sec == 0 &&
1731 rdev->last_read_error.tv_nsec == 0) {
1732 /* first time we've seen a read error */
1733 rdev->last_read_error = cur_time_mon;
1734 return;
1735 }
1736
1737 hours_since_last = (cur_time_mon.tv_sec -
1738 rdev->last_read_error.tv_sec) / 3600;
1739
1740 rdev->last_read_error = cur_time_mon;
1741
1742 /*
1743 * if hours_since_last is > the number of bits in read_errors
1744 * just set read errors to 0. We do this to avoid
1745 * overflowing the shift of read_errors by hours_since_last.
1746 */
1747 if (hours_since_last >= 8 * sizeof(read_errors))
1748 atomic_set(&rdev->read_errors, 0);
1749 else
1750 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
1751}
1752
1753static int r10_sync_page_io(mdk_rdev_t *rdev, sector_t sector,
1754 int sectors, struct page *page, int rw)
1755{
1756 sector_t first_bad;
1757 int bad_sectors;
1758
1759 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
1760 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
1761 return -1;
1762 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1763 /* success */
1764 return 1;
1765 if (rw == WRITE)
1766 set_bit(WriteErrorSeen, &rdev->flags);
1767 /* need to record an error - either for the block or the device */
1768 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1769 md_error(rdev->mddev, rdev);
1770 return 0;
1771}
1772
1773/*
1774 * This is a kernel thread which:
1775 *
1776 * 1. Retries failed read operations on working mirrors.
1777 * 2. Updates the raid superblock when problems encounter.
1778 * 3. Performs writes following reads for array synchronising.
1779 */
1780
1781static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1782{
1783 int sect = 0; /* Offset from r10_bio->sector */
1784 int sectors = r10_bio->sectors;
1785 mdk_rdev_t*rdev;
1786 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
1787 int d = r10_bio->devs[r10_bio->read_slot].devnum;
1788
1789 /* still own a reference to this rdev, so it cannot
1790 * have been cleared recently.
1791 */
1792 rdev = conf->mirrors[d].rdev;
1793
1794 if (test_bit(Faulty, &rdev->flags))
1795 /* drive has already been failed, just ignore any
1796 more fix_read_error() attempts */
1797 return;
1798
1799 check_decay_read_errors(mddev, rdev);
1800 atomic_inc(&rdev->read_errors);
1801 if (atomic_read(&rdev->read_errors) > max_read_errors) {
1802 char b[BDEVNAME_SIZE];
1803 bdevname(rdev->bdev, b);
1804
1805 printk(KERN_NOTICE
1806 "md/raid10:%s: %s: Raid device exceeded "
1807 "read_error threshold [cur %d:max %d]\n",
1808 mdname(mddev), b,
1809 atomic_read(&rdev->read_errors), max_read_errors);
1810 printk(KERN_NOTICE
1811 "md/raid10:%s: %s: Failing raid device\n",
1812 mdname(mddev), b);
1813 md_error(mddev, conf->mirrors[d].rdev);
1814 return;
1815 }
1816
1817 while(sectors) {
1818 int s = sectors;
1819 int sl = r10_bio->read_slot;
1820 int success = 0;
1821 int start;
1822
1823 if (s > (PAGE_SIZE>>9))
1824 s = PAGE_SIZE >> 9;
1825
1826 rcu_read_lock();
1827 do {
1828 sector_t first_bad;
1829 int bad_sectors;
1830
1831 d = r10_bio->devs[sl].devnum;
1832 rdev = rcu_dereference(conf->mirrors[d].rdev);
1833 if (rdev &&
1834 test_bit(In_sync, &rdev->flags) &&
1835 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
1836 &first_bad, &bad_sectors) == 0) {
1837 atomic_inc(&rdev->nr_pending);
1838 rcu_read_unlock();
1839 success = sync_page_io(rdev,
1840 r10_bio->devs[sl].addr +
1841 sect,
1842 s<<9,
1843 conf->tmppage, READ, false);
1844 rdev_dec_pending(rdev, mddev);
1845 rcu_read_lock();
1846 if (success)
1847 break;
1848 }
1849 sl++;
1850 if (sl == conf->copies)
1851 sl = 0;
1852 } while (!success && sl != r10_bio->read_slot);
1853 rcu_read_unlock();
1854
1855 if (!success) {
1856 /* Cannot read from anywhere, just mark the block
1857 * as bad on the first device to discourage future
1858 * reads.
1859 */
1860 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1861 rdev = conf->mirrors[dn].rdev;
1862
1863 if (!rdev_set_badblocks(
1864 rdev,
1865 r10_bio->devs[r10_bio->read_slot].addr
1866 + sect,
1867 s, 0))
1868 md_error(mddev, rdev);
1869 break;
1870 }
1871
1872 start = sl;
1873 /* write it back and re-read */
1874 rcu_read_lock();
1875 while (sl != r10_bio->read_slot) {
1876 char b[BDEVNAME_SIZE];
1877
1878 if (sl==0)
1879 sl = conf->copies;
1880 sl--;
1881 d = r10_bio->devs[sl].devnum;
1882 rdev = rcu_dereference(conf->mirrors[d].rdev);
1883 if (!rdev ||
1884 !test_bit(In_sync, &rdev->flags))
1885 continue;
1886
1887 atomic_inc(&rdev->nr_pending);
1888 rcu_read_unlock();
1889 if (r10_sync_page_io(rdev,
1890 r10_bio->devs[sl].addr +
1891 sect,
1892 s<<9, conf->tmppage, WRITE)
1893 == 0) {
1894 /* Well, this device is dead */
1895 printk(KERN_NOTICE
1896 "md/raid10:%s: read correction "
1897 "write failed"
1898 " (%d sectors at %llu on %s)\n",
1899 mdname(mddev), s,
1900 (unsigned long long)(
1901 sect + rdev->data_offset),
1902 bdevname(rdev->bdev, b));
1903 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
1904 "drive\n",
1905 mdname(mddev),
1906 bdevname(rdev->bdev, b));
1907 }
1908 rdev_dec_pending(rdev, mddev);
1909 rcu_read_lock();
1910 }
1911 sl = start;
1912 while (sl != r10_bio->read_slot) {
1913 char b[BDEVNAME_SIZE];
1914
1915 if (sl==0)
1916 sl = conf->copies;
1917 sl--;
1918 d = r10_bio->devs[sl].devnum;
1919 rdev = rcu_dereference(conf->mirrors[d].rdev);
1920 if (!rdev ||
1921 !test_bit(In_sync, &rdev->flags))
1922 continue;
1923
1924 atomic_inc(&rdev->nr_pending);
1925 rcu_read_unlock();
1926 switch (r10_sync_page_io(rdev,
1927 r10_bio->devs[sl].addr +
1928 sect,
1929 s<<9, conf->tmppage,
1930 READ)) {
1931 case 0:
1932 /* Well, this device is dead */
1933 printk(KERN_NOTICE
1934 "md/raid10:%s: unable to read back "
1935 "corrected sectors"
1936 " (%d sectors at %llu on %s)\n",
1937 mdname(mddev), s,
1938 (unsigned long long)(
1939 sect + rdev->data_offset),
1940 bdevname(rdev->bdev, b));
1941 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
1942 "drive\n",
1943 mdname(mddev),
1944 bdevname(rdev->bdev, b));
1945 break;
1946 case 1:
1947 printk(KERN_INFO
1948 "md/raid10:%s: read error corrected"
1949 " (%d sectors at %llu on %s)\n",
1950 mdname(mddev), s,
1951 (unsigned long long)(
1952 sect + rdev->data_offset),
1953 bdevname(rdev->bdev, b));
1954 atomic_add(s, &rdev->corrected_errors);
1955 }
1956
1957 rdev_dec_pending(rdev, mddev);
1958 rcu_read_lock();
1959 }
1960 rcu_read_unlock();
1961
1962 sectors -= s;
1963 sect += s;
1964 }
1965}
1966
1967static void bi_complete(struct bio *bio, int error)
1968{
1969 complete((struct completion *)bio->bi_private);
1970}
1971
1972static int submit_bio_wait(int rw, struct bio *bio)
1973{
1974 struct completion event;
1975 rw |= REQ_SYNC;
1976
1977 init_completion(&event);
1978 bio->bi_private = &event;
1979 bio->bi_end_io = bi_complete;
1980 submit_bio(rw, bio);
1981 wait_for_completion(&event);
1982
1983 return test_bit(BIO_UPTODATE, &bio->bi_flags);
1984}
1985
1986static int narrow_write_error(r10bio_t *r10_bio, int i)
1987{
1988 struct bio *bio = r10_bio->master_bio;
1989 mddev_t *mddev = r10_bio->mddev;
1990 conf_t *conf = mddev->private;
1991 mdk_rdev_t *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
1992 /* bio has the data to be written to slot 'i' where
1993 * we just recently had a write error.
1994 * We repeatedly clone the bio and trim down to one block,
1995 * then try the write. Where the write fails we record
1996 * a bad block.
1997 * It is conceivable that the bio doesn't exactly align with
1998 * blocks. We must handle this.
1999 *
2000 * We currently own a reference to the rdev.
2001 */
2002
2003 int block_sectors;
2004 sector_t sector;
2005 int sectors;
2006 int sect_to_write = r10_bio->sectors;
2007 int ok = 1;
2008
2009 if (rdev->badblocks.shift < 0)
2010 return 0;
2011
2012 block_sectors = 1 << rdev->badblocks.shift;
2013 sector = r10_bio->sector;
2014 sectors = ((r10_bio->sector + block_sectors)
2015 & ~(sector_t)(block_sectors - 1))
2016 - sector;
2017
2018 while (sect_to_write) {
2019 struct bio *wbio;
2020 if (sectors > sect_to_write)
2021 sectors = sect_to_write;
2022 /* Write at 'sector' for 'sectors' */
2023 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2024 md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2025 wbio->bi_sector = (r10_bio->devs[i].addr+
2026 rdev->data_offset+
2027 (sector - r10_bio->sector));
2028 wbio->bi_bdev = rdev->bdev;
2029 if (submit_bio_wait(WRITE, wbio) == 0)
2030 /* Failure! */
2031 ok = rdev_set_badblocks(rdev, sector,
2032 sectors, 0)
2033 && ok;
2034
2035 bio_put(wbio);
2036 sect_to_write -= sectors;
2037 sector += sectors;
2038 sectors = block_sectors;
2039 }
2040 return ok;
2041}
2042
2043static void handle_read_error(mddev_t *mddev, r10bio_t *r10_bio)
2044{
2045 int slot = r10_bio->read_slot;
2046 int mirror = r10_bio->devs[slot].devnum;
2047 struct bio *bio;
2048 conf_t *conf = mddev->private;
2049 mdk_rdev_t *rdev;
2050 char b[BDEVNAME_SIZE];
2051 unsigned long do_sync;
2052 int max_sectors;
2053
2054 /* we got a read error. Maybe the drive is bad. Maybe just
2055 * the block and we can fix it.
2056 * We freeze all other IO, and try reading the block from
2057 * other devices. When we find one, we re-write
2058 * and check it that fixes the read error.
2059 * This is all done synchronously while the array is
2060 * frozen.
2061 */
2062 if (mddev->ro == 0) {
2063 freeze_array(conf);
2064 fix_read_error(conf, mddev, r10_bio);
2065 unfreeze_array(conf);
2066 }
2067 rdev_dec_pending(conf->mirrors[mirror].rdev, mddev);
2068
2069 bio = r10_bio->devs[slot].bio;
2070 bdevname(bio->bi_bdev, b);
2071 r10_bio->devs[slot].bio =
2072 mddev->ro ? IO_BLOCKED : NULL;
2073read_more:
2074 mirror = read_balance(conf, r10_bio, &max_sectors);
2075 if (mirror == -1) {
2076 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2077 " read error for block %llu\n",
2078 mdname(mddev), b,
2079 (unsigned long long)r10_bio->sector);
2080 raid_end_bio_io(r10_bio);
2081 bio_put(bio);
2082 return;
2083 }
2084
2085 do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2086 if (bio)
2087 bio_put(bio);
2088 slot = r10_bio->read_slot;
2089 rdev = conf->mirrors[mirror].rdev;
2090 printk_ratelimited(
2091 KERN_ERR
2092 "md/raid10:%s: %s: redirecting"
2093 "sector %llu to another mirror\n",
2094 mdname(mddev),
2095 bdevname(rdev->bdev, b),
2096 (unsigned long long)r10_bio->sector);
2097 bio = bio_clone_mddev(r10_bio->master_bio,
2098 GFP_NOIO, mddev);
2099 md_trim_bio(bio,
2100 r10_bio->sector - bio->bi_sector,
2101 max_sectors);
2102 r10_bio->devs[slot].bio = bio;
2103 bio->bi_sector = r10_bio->devs[slot].addr
2104 + rdev->data_offset;
2105 bio->bi_bdev = rdev->bdev;
2106 bio->bi_rw = READ | do_sync;
2107 bio->bi_private = r10_bio;
2108 bio->bi_end_io = raid10_end_read_request;
2109 if (max_sectors < r10_bio->sectors) {
2110 /* Drat - have to split this up more */
2111 struct bio *mbio = r10_bio->master_bio;
2112 int sectors_handled =
2113 r10_bio->sector + max_sectors
2114 - mbio->bi_sector;
2115 r10_bio->sectors = max_sectors;
2116 spin_lock_irq(&conf->device_lock);
2117 if (mbio->bi_phys_segments == 0)
2118 mbio->bi_phys_segments = 2;
2119 else
2120 mbio->bi_phys_segments++;
2121 spin_unlock_irq(&conf->device_lock);
2122 generic_make_request(bio);
2123 bio = NULL;
2124
2125 r10_bio = mempool_alloc(conf->r10bio_pool,
2126 GFP_NOIO);
2127 r10_bio->master_bio = mbio;
2128 r10_bio->sectors = (mbio->bi_size >> 9)
2129 - sectors_handled;
2130 r10_bio->state = 0;
2131 set_bit(R10BIO_ReadError,
2132 &r10_bio->state);
2133 r10_bio->mddev = mddev;
2134 r10_bio->sector = mbio->bi_sector
2135 + sectors_handled;
2136
2137 goto read_more;
2138 } else
2139 generic_make_request(bio);
2140}
2141
2142static void handle_write_completed(conf_t *conf, r10bio_t *r10_bio)
2143{
2144 /* Some sort of write request has finished and it
2145 * succeeded in writing where we thought there was a
2146 * bad block. So forget the bad block.
2147 * Or possibly if failed and we need to record
2148 * a bad block.
2149 */
2150 int m;
2151 mdk_rdev_t *rdev;
2152
2153 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2154 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2155 for (m = 0; m < conf->copies; m++) {
2156 int dev = r10_bio->devs[m].devnum;
2157 rdev = conf->mirrors[dev].rdev;
2158 if (r10_bio->devs[m].bio == NULL)
2159 continue;
2160 if (test_bit(BIO_UPTODATE,
2161 &r10_bio->devs[m].bio->bi_flags)) {
2162 rdev_clear_badblocks(
2163 rdev,
2164 r10_bio->devs[m].addr,
2165 r10_bio->sectors);
2166 } else {
2167 if (!rdev_set_badblocks(
2168 rdev,
2169 r10_bio->devs[m].addr,
2170 r10_bio->sectors, 0))
2171 md_error(conf->mddev, rdev);
2172 }
2173 }
2174 put_buf(r10_bio);
2175 } else {
2176 for (m = 0; m < conf->copies; m++) {
2177 int dev = r10_bio->devs[m].devnum;
2178 struct bio *bio = r10_bio->devs[m].bio;
2179 rdev = conf->mirrors[dev].rdev;
2180 if (bio == IO_MADE_GOOD) {
2181 rdev_clear_badblocks(
2182 rdev,
2183 r10_bio->devs[m].addr,
2184 r10_bio->sectors);
2185 rdev_dec_pending(rdev, conf->mddev);
2186 } else if (bio != NULL &&
2187 !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2188 if (!narrow_write_error(r10_bio, m)) {
2189 md_error(conf->mddev, rdev);
2190 set_bit(R10BIO_Degraded,
2191 &r10_bio->state);
2192 }
2193 rdev_dec_pending(rdev, conf->mddev);
2194 }
2195 }
2196 if (test_bit(R10BIO_WriteError,
2197 &r10_bio->state))
2198 close_write(r10_bio);
2199 raid_end_bio_io(r10_bio);
2200 }
2201}
2202
2203static void raid10d(mddev_t *mddev)
2204{
2205 r10bio_t *r10_bio;
2206 unsigned long flags;
2207 conf_t *conf = mddev->private;
2208 struct list_head *head = &conf->retry_list;
2209 struct blk_plug plug;
2210
2211 md_check_recovery(mddev);
2212
2213 blk_start_plug(&plug);
2214 for (;;) {
2215
2216 flush_pending_writes(conf);
2217
2218 spin_lock_irqsave(&conf->device_lock, flags);
2219 if (list_empty(head)) {
2220 spin_unlock_irqrestore(&conf->device_lock, flags);
2221 break;
2222 }
2223 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
2224 list_del(head->prev);
2225 conf->nr_queued--;
2226 spin_unlock_irqrestore(&conf->device_lock, flags);
2227
2228 mddev = r10_bio->mddev;
2229 conf = mddev->private;
2230 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2231 test_bit(R10BIO_WriteError, &r10_bio->state))
2232 handle_write_completed(conf, r10_bio);
2233 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2234 sync_request_write(mddev, r10_bio);
2235 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2236 recovery_request_write(mddev, r10_bio);
2237 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2238 handle_read_error(mddev, r10_bio);
2239 else {
2240 /* just a partial read to be scheduled from a
2241 * separate context
2242 */
2243 int slot = r10_bio->read_slot;
2244 generic_make_request(r10_bio->devs[slot].bio);
2245 }
2246
2247 cond_resched();
2248 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2249 md_check_recovery(mddev);
2250 }
2251 blk_finish_plug(&plug);
2252}
2253
2254
2255static int init_resync(conf_t *conf)
2256{
2257 int buffs;
2258
2259 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2260 BUG_ON(conf->r10buf_pool);
2261 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2262 if (!conf->r10buf_pool)
2263 return -ENOMEM;
2264 conf->next_resync = 0;
2265 return 0;
2266}
2267
2268/*
2269 * perform a "sync" on one "block"
2270 *
2271 * We need to make sure that no normal I/O request - particularly write
2272 * requests - conflict with active sync requests.
2273 *
2274 * This is achieved by tracking pending requests and a 'barrier' concept
2275 * that can be installed to exclude normal IO requests.
2276 *
2277 * Resync and recovery are handled very differently.
2278 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2279 *
2280 * For resync, we iterate over virtual addresses, read all copies,
2281 * and update if there are differences. If only one copy is live,
2282 * skip it.
2283 * For recovery, we iterate over physical addresses, read a good
2284 * value for each non-in_sync drive, and over-write.
2285 *
2286 * So, for recovery we may have several outstanding complex requests for a
2287 * given address, one for each out-of-sync device. We model this by allocating
2288 * a number of r10_bio structures, one for each out-of-sync device.
2289 * As we setup these structures, we collect all bio's together into a list
2290 * which we then process collectively to add pages, and then process again
2291 * to pass to generic_make_request.
2292 *
2293 * The r10_bio structures are linked using a borrowed master_bio pointer.
2294 * This link is counted in ->remaining. When the r10_bio that points to NULL
2295 * has its remaining count decremented to 0, the whole complex operation
2296 * is complete.
2297 *
2298 */
2299
2300static sector_t sync_request(mddev_t *mddev, sector_t sector_nr,
2301 int *skipped, int go_faster)
2302{
2303 conf_t *conf = mddev->private;
2304 r10bio_t *r10_bio;
2305 struct bio *biolist = NULL, *bio;
2306 sector_t max_sector, nr_sectors;
2307 int i;
2308 int max_sync;
2309 sector_t sync_blocks;
2310 sector_t sectors_skipped = 0;
2311 int chunks_skipped = 0;
2312
2313 if (!conf->r10buf_pool)
2314 if (init_resync(conf))
2315 return 0;
2316
2317 skipped:
2318 max_sector = mddev->dev_sectors;
2319 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2320 max_sector = mddev->resync_max_sectors;
2321 if (sector_nr >= max_sector) {
2322 /* If we aborted, we need to abort the
2323 * sync on the 'current' bitmap chucks (there can
2324 * be several when recovering multiple devices).
2325 * as we may have started syncing it but not finished.
2326 * We can find the current address in
2327 * mddev->curr_resync, but for recovery,
2328 * we need to convert that to several
2329 * virtual addresses.
2330 */
2331 if (mddev->curr_resync < max_sector) { /* aborted */
2332 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2333 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2334 &sync_blocks, 1);
2335 else for (i=0; i<conf->raid_disks; i++) {
2336 sector_t sect =
2337 raid10_find_virt(conf, mddev->curr_resync, i);
2338 bitmap_end_sync(mddev->bitmap, sect,
2339 &sync_blocks, 1);
2340 }
2341 } else /* completed sync */
2342 conf->fullsync = 0;
2343
2344 bitmap_close_sync(mddev->bitmap);
2345 close_sync(conf);
2346 *skipped = 1;
2347 return sectors_skipped;
2348 }
2349 if (chunks_skipped >= conf->raid_disks) {
2350 /* if there has been nothing to do on any drive,
2351 * then there is nothing to do at all..
2352 */
2353 *skipped = 1;
2354 return (max_sector - sector_nr) + sectors_skipped;
2355 }
2356
2357 if (max_sector > mddev->resync_max)
2358 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2359
2360 /* make sure whole request will fit in a chunk - if chunks
2361 * are meaningful
2362 */
2363 if (conf->near_copies < conf->raid_disks &&
2364 max_sector > (sector_nr | conf->chunk_mask))
2365 max_sector = (sector_nr | conf->chunk_mask) + 1;
2366 /*
2367 * If there is non-resync activity waiting for us then
2368 * put in a delay to throttle resync.
2369 */
2370 if (!go_faster && conf->nr_waiting)
2371 msleep_interruptible(1000);
2372
2373 /* Again, very different code for resync and recovery.
2374 * Both must result in an r10bio with a list of bios that
2375 * have bi_end_io, bi_sector, bi_bdev set,
2376 * and bi_private set to the r10bio.
2377 * For recovery, we may actually create several r10bios
2378 * with 2 bios in each, that correspond to the bios in the main one.
2379 * In this case, the subordinate r10bios link back through a
2380 * borrowed master_bio pointer, and the counter in the master
2381 * includes a ref from each subordinate.
2382 */
2383 /* First, we decide what to do and set ->bi_end_io
2384 * To end_sync_read if we want to read, and
2385 * end_sync_write if we will want to write.
2386 */
2387
2388 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2389 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2390 /* recovery... the complicated one */
2391 int j;
2392 r10_bio = NULL;
2393
2394 for (i=0 ; i<conf->raid_disks; i++) {
2395 int still_degraded;
2396 r10bio_t *rb2;
2397 sector_t sect;
2398 int must_sync;
2399 int any_working;
2400
2401 if (conf->mirrors[i].rdev == NULL ||
2402 test_bit(In_sync, &conf->mirrors[i].rdev->flags))
2403 continue;
2404
2405 still_degraded = 0;
2406 /* want to reconstruct this device */
2407 rb2 = r10_bio;
2408 sect = raid10_find_virt(conf, sector_nr, i);
2409 /* Unless we are doing a full sync, we only need
2410 * to recover the block if it is set in the bitmap
2411 */
2412 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2413 &sync_blocks, 1);
2414 if (sync_blocks < max_sync)
2415 max_sync = sync_blocks;
2416 if (!must_sync &&
2417 !conf->fullsync) {
2418 /* yep, skip the sync_blocks here, but don't assume
2419 * that there will never be anything to do here
2420 */
2421 chunks_skipped = -1;
2422 continue;
2423 }
2424
2425 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2426 raise_barrier(conf, rb2 != NULL);
2427 atomic_set(&r10_bio->remaining, 0);
2428
2429 r10_bio->master_bio = (struct bio*)rb2;
2430 if (rb2)
2431 atomic_inc(&rb2->remaining);
2432 r10_bio->mddev = mddev;
2433 set_bit(R10BIO_IsRecover, &r10_bio->state);
2434 r10_bio->sector = sect;
2435
2436 raid10_find_phys(conf, r10_bio);
2437
2438 /* Need to check if the array will still be
2439 * degraded
2440 */
2441 for (j=0; j<conf->raid_disks; j++)
2442 if (conf->mirrors[j].rdev == NULL ||
2443 test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
2444 still_degraded = 1;
2445 break;
2446 }
2447
2448 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2449 &sync_blocks, still_degraded);
2450
2451 any_working = 0;
2452 for (j=0; j<conf->copies;j++) {
2453 int k;
2454 int d = r10_bio->devs[j].devnum;
2455 sector_t from_addr, to_addr;
2456 mdk_rdev_t *rdev;
2457 sector_t sector, first_bad;
2458 int bad_sectors;
2459 if (!conf->mirrors[d].rdev ||
2460 !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
2461 continue;
2462 /* This is where we read from */
2463 any_working = 1;
2464 rdev = conf->mirrors[d].rdev;
2465 sector = r10_bio->devs[j].addr;
2466
2467 if (is_badblock(rdev, sector, max_sync,
2468 &first_bad, &bad_sectors)) {
2469 if (first_bad > sector)
2470 max_sync = first_bad - sector;
2471 else {
2472 bad_sectors -= (sector
2473 - first_bad);
2474 if (max_sync > bad_sectors)
2475 max_sync = bad_sectors;
2476 continue;
2477 }
2478 }
2479 bio = r10_bio->devs[0].bio;
2480 bio->bi_next = biolist;
2481 biolist = bio;
2482 bio->bi_private = r10_bio;
2483 bio->bi_end_io = end_sync_read;
2484 bio->bi_rw = READ;
2485 from_addr = r10_bio->devs[j].addr;
2486 bio->bi_sector = from_addr +
2487 conf->mirrors[d].rdev->data_offset;
2488 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2489 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2490 atomic_inc(&r10_bio->remaining);
2491 /* and we write to 'i' */
2492
2493 for (k=0; k<conf->copies; k++)
2494 if (r10_bio->devs[k].devnum == i)
2495 break;
2496 BUG_ON(k == conf->copies);
2497 bio = r10_bio->devs[1].bio;
2498 bio->bi_next = biolist;
2499 biolist = bio;
2500 bio->bi_private = r10_bio;
2501 bio->bi_end_io = end_sync_write;
2502 bio->bi_rw = WRITE;
2503 to_addr = r10_bio->devs[k].addr;
2504 bio->bi_sector = to_addr +
2505 conf->mirrors[i].rdev->data_offset;
2506 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
2507
2508 r10_bio->devs[0].devnum = d;
2509 r10_bio->devs[0].addr = from_addr;
2510 r10_bio->devs[1].devnum = i;
2511 r10_bio->devs[1].addr = to_addr;
2512
2513 break;
2514 }
2515 if (j == conf->copies) {
2516 /* Cannot recover, so abort the recovery or
2517 * record a bad block */
2518 put_buf(r10_bio);
2519 if (rb2)
2520 atomic_dec(&rb2->remaining);
2521 r10_bio = rb2;
2522 if (any_working) {
2523 /* problem is that there are bad blocks
2524 * on other device(s)
2525 */
2526 int k;
2527 for (k = 0; k < conf->copies; k++)
2528 if (r10_bio->devs[k].devnum == i)
2529 break;
2530 if (!rdev_set_badblocks(
2531 conf->mirrors[i].rdev,
2532 r10_bio->devs[k].addr,
2533 max_sync, 0))
2534 any_working = 0;
2535 }
2536 if (!any_working) {
2537 if (!test_and_set_bit(MD_RECOVERY_INTR,
2538 &mddev->recovery))
2539 printk(KERN_INFO "md/raid10:%s: insufficient "
2540 "working devices for recovery.\n",
2541 mdname(mddev));
2542 conf->mirrors[i].recovery_disabled
2543 = mddev->recovery_disabled;
2544 }
2545 break;
2546 }
2547 }
2548 if (biolist == NULL) {
2549 while (r10_bio) {
2550 r10bio_t *rb2 = r10_bio;
2551 r10_bio = (r10bio_t*) rb2->master_bio;
2552 rb2->master_bio = NULL;
2553 put_buf(rb2);
2554 }
2555 goto giveup;
2556 }
2557 } else {
2558 /* resync. Schedule a read for every block at this virt offset */
2559 int count = 0;
2560
2561 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2562
2563 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2564 &sync_blocks, mddev->degraded) &&
2565 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
2566 &mddev->recovery)) {
2567 /* We can skip this block */
2568 *skipped = 1;
2569 return sync_blocks + sectors_skipped;
2570 }
2571 if (sync_blocks < max_sync)
2572 max_sync = sync_blocks;
2573 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2574
2575 r10_bio->mddev = mddev;
2576 atomic_set(&r10_bio->remaining, 0);
2577 raise_barrier(conf, 0);
2578 conf->next_resync = sector_nr;
2579
2580 r10_bio->master_bio = NULL;
2581 r10_bio->sector = sector_nr;
2582 set_bit(R10BIO_IsSync, &r10_bio->state);
2583 raid10_find_phys(conf, r10_bio);
2584 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
2585
2586 for (i=0; i<conf->copies; i++) {
2587 int d = r10_bio->devs[i].devnum;
2588 sector_t first_bad, sector;
2589 int bad_sectors;
2590
2591 bio = r10_bio->devs[i].bio;
2592 bio->bi_end_io = NULL;
2593 clear_bit(BIO_UPTODATE, &bio->bi_flags);
2594 if (conf->mirrors[d].rdev == NULL ||
2595 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
2596 continue;
2597 sector = r10_bio->devs[i].addr;
2598 if (is_badblock(conf->mirrors[d].rdev,
2599 sector, max_sync,
2600 &first_bad, &bad_sectors)) {
2601 if (first_bad > sector)
2602 max_sync = first_bad - sector;
2603 else {
2604 bad_sectors -= (sector - first_bad);
2605 if (max_sync > bad_sectors)
2606 max_sync = max_sync;
2607 continue;
2608 }
2609 }
2610 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2611 atomic_inc(&r10_bio->remaining);
2612 bio->bi_next = biolist;
2613 biolist = bio;
2614 bio->bi_private = r10_bio;
2615 bio->bi_end_io = end_sync_read;
2616 bio->bi_rw = READ;
2617 bio->bi_sector = sector +
2618 conf->mirrors[d].rdev->data_offset;
2619 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2620 count++;
2621 }
2622
2623 if (count < 2) {
2624 for (i=0; i<conf->copies; i++) {
2625 int d = r10_bio->devs[i].devnum;
2626 if (r10_bio->devs[i].bio->bi_end_io)
2627 rdev_dec_pending(conf->mirrors[d].rdev,
2628 mddev);
2629 }
2630 put_buf(r10_bio);
2631 biolist = NULL;
2632 goto giveup;
2633 }
2634 }
2635
2636 for (bio = biolist; bio ; bio=bio->bi_next) {
2637
2638 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
2639 if (bio->bi_end_io)
2640 bio->bi_flags |= 1 << BIO_UPTODATE;
2641 bio->bi_vcnt = 0;
2642 bio->bi_idx = 0;
2643 bio->bi_phys_segments = 0;
2644 bio->bi_size = 0;
2645 }
2646
2647 nr_sectors = 0;
2648 if (sector_nr + max_sync < max_sector)
2649 max_sector = sector_nr + max_sync;
2650 do {
2651 struct page *page;
2652 int len = PAGE_SIZE;
2653 if (sector_nr + (len>>9) > max_sector)
2654 len = (max_sector - sector_nr) << 9;
2655 if (len == 0)
2656 break;
2657 for (bio= biolist ; bio ; bio=bio->bi_next) {
2658 struct bio *bio2;
2659 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2660 if (bio_add_page(bio, page, len, 0))
2661 continue;
2662
2663 /* stop here */
2664 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2665 for (bio2 = biolist;
2666 bio2 && bio2 != bio;
2667 bio2 = bio2->bi_next) {
2668 /* remove last page from this bio */
2669 bio2->bi_vcnt--;
2670 bio2->bi_size -= len;
2671 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
2672 }
2673 goto bio_full;
2674 }
2675 nr_sectors += len>>9;
2676 sector_nr += len>>9;
2677 } while (biolist->bi_vcnt < RESYNC_PAGES);
2678 bio_full:
2679 r10_bio->sectors = nr_sectors;
2680
2681 while (biolist) {
2682 bio = biolist;
2683 biolist = biolist->bi_next;
2684
2685 bio->bi_next = NULL;
2686 r10_bio = bio->bi_private;
2687 r10_bio->sectors = nr_sectors;
2688
2689 if (bio->bi_end_io == end_sync_read) {
2690 md_sync_acct(bio->bi_bdev, nr_sectors);
2691 generic_make_request(bio);
2692 }
2693 }
2694
2695 if (sectors_skipped)
2696 /* pretend they weren't skipped, it makes
2697 * no important difference in this case
2698 */
2699 md_done_sync(mddev, sectors_skipped, 1);
2700
2701 return sectors_skipped + nr_sectors;
2702 giveup:
2703 /* There is nowhere to write, so all non-sync
2704 * drives must be failed or in resync, all drives
2705 * have a bad block, so try the next chunk...
2706 */
2707 if (sector_nr + max_sync < max_sector)
2708 max_sector = sector_nr + max_sync;
2709
2710 sectors_skipped += (max_sector - sector_nr);
2711 chunks_skipped ++;
2712 sector_nr = max_sector;
2713 goto skipped;
2714}
2715
2716static sector_t
2717raid10_size(mddev_t *mddev, sector_t sectors, int raid_disks)
2718{
2719 sector_t size;
2720 conf_t *conf = mddev->private;
2721
2722 if (!raid_disks)
2723 raid_disks = conf->raid_disks;
2724 if (!sectors)
2725 sectors = conf->dev_sectors;
2726
2727 size = sectors >> conf->chunk_shift;
2728 sector_div(size, conf->far_copies);
2729 size = size * raid_disks;
2730 sector_div(size, conf->near_copies);
2731
2732 return size << conf->chunk_shift;
2733}
2734
2735
2736static conf_t *setup_conf(mddev_t *mddev)
2737{
2738 conf_t *conf = NULL;
2739 int nc, fc, fo;
2740 sector_t stride, size;
2741 int err = -EINVAL;
2742
2743 if (mddev->new_chunk_sectors < (PAGE_SIZE >> 9) ||
2744 !is_power_of_2(mddev->new_chunk_sectors)) {
2745 printk(KERN_ERR "md/raid10:%s: chunk size must be "
2746 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
2747 mdname(mddev), PAGE_SIZE);
2748 goto out;
2749 }
2750
2751 nc = mddev->new_layout & 255;
2752 fc = (mddev->new_layout >> 8) & 255;
2753 fo = mddev->new_layout & (1<<16);
2754
2755 if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
2756 (mddev->new_layout >> 17)) {
2757 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
2758 mdname(mddev), mddev->new_layout);
2759 goto out;
2760 }
2761
2762 err = -ENOMEM;
2763 conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2764 if (!conf)
2765 goto out;
2766
2767 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2768 GFP_KERNEL);
2769 if (!conf->mirrors)
2770 goto out;
2771
2772 conf->tmppage = alloc_page(GFP_KERNEL);
2773 if (!conf->tmppage)
2774 goto out;
2775
2776
2777 conf->raid_disks = mddev->raid_disks;
2778 conf->near_copies = nc;
2779 conf->far_copies = fc;
2780 conf->copies = nc*fc;
2781 conf->far_offset = fo;
2782 conf->chunk_mask = mddev->new_chunk_sectors - 1;
2783 conf->chunk_shift = ffz(~mddev->new_chunk_sectors);
2784
2785 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2786 r10bio_pool_free, conf);
2787 if (!conf->r10bio_pool)
2788 goto out;
2789
2790 size = mddev->dev_sectors >> conf->chunk_shift;
2791 sector_div(size, fc);
2792 size = size * conf->raid_disks;
2793 sector_div(size, nc);
2794 /* 'size' is now the number of chunks in the array */
2795 /* calculate "used chunks per device" in 'stride' */
2796 stride = size * conf->copies;
2797
2798 /* We need to round up when dividing by raid_disks to
2799 * get the stride size.
2800 */
2801 stride += conf->raid_disks - 1;
2802 sector_div(stride, conf->raid_disks);
2803
2804 conf->dev_sectors = stride << conf->chunk_shift;
2805
2806 if (fo)
2807 stride = 1;
2808 else
2809 sector_div(stride, fc);
2810 conf->stride = stride << conf->chunk_shift;
2811
2812
2813 spin_lock_init(&conf->device_lock);
2814 INIT_LIST_HEAD(&conf->retry_list);
2815
2816 spin_lock_init(&conf->resync_lock);
2817 init_waitqueue_head(&conf->wait_barrier);
2818
2819 conf->thread = md_register_thread(raid10d, mddev, NULL);
2820 if (!conf->thread)
2821 goto out;
2822
2823 conf->mddev = mddev;
2824 return conf;
2825
2826 out:
2827 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
2828 mdname(mddev));
2829 if (conf) {
2830 if (conf->r10bio_pool)
2831 mempool_destroy(conf->r10bio_pool);
2832 kfree(conf->mirrors);
2833 safe_put_page(conf->tmppage);
2834 kfree(conf);
2835 }
2836 return ERR_PTR(err);
2837}
2838
2839static int run(mddev_t *mddev)
2840{
2841 conf_t *conf;
2842 int i, disk_idx, chunk_size;
2843 mirror_info_t *disk;
2844 mdk_rdev_t *rdev;
2845 sector_t size;
2846
2847 /*
2848 * copy the already verified devices into our private RAID10
2849 * bookkeeping area. [whatever we allocate in run(),
2850 * should be freed in stop()]
2851 */
2852
2853 if (mddev->private == NULL) {
2854 conf = setup_conf(mddev);
2855 if (IS_ERR(conf))
2856 return PTR_ERR(conf);
2857 mddev->private = conf;
2858 }
2859 conf = mddev->private;
2860 if (!conf)
2861 goto out;
2862
2863 mddev->thread = conf->thread;
2864 conf->thread = NULL;
2865
2866 chunk_size = mddev->chunk_sectors << 9;
2867 blk_queue_io_min(mddev->queue, chunk_size);
2868 if (conf->raid_disks % conf->near_copies)
2869 blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks);
2870 else
2871 blk_queue_io_opt(mddev->queue, chunk_size *
2872 (conf->raid_disks / conf->near_copies));
2873
2874 list_for_each_entry(rdev, &mddev->disks, same_set) {
2875
2876 disk_idx = rdev->raid_disk;
2877 if (disk_idx >= conf->raid_disks
2878 || disk_idx < 0)
2879 continue;
2880 disk = conf->mirrors + disk_idx;
2881
2882 disk->rdev = rdev;
2883 disk_stack_limits(mddev->gendisk, rdev->bdev,
2884 rdev->data_offset << 9);
2885 /* as we don't honour merge_bvec_fn, we must never risk
2886 * violating it, so limit max_segments to 1 lying
2887 * within a single page.
2888 */
2889 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2890 blk_queue_max_segments(mddev->queue, 1);
2891 blk_queue_segment_boundary(mddev->queue,
2892 PAGE_CACHE_SIZE - 1);
2893 }
2894
2895 disk->head_position = 0;
2896 }
2897 /* need to check that every block has at least one working mirror */
2898 if (!enough(conf, -1)) {
2899 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
2900 mdname(mddev));
2901 goto out_free_conf;
2902 }
2903
2904 mddev->degraded = 0;
2905 for (i = 0; i < conf->raid_disks; i++) {
2906
2907 disk = conf->mirrors + i;
2908
2909 if (!disk->rdev ||
2910 !test_bit(In_sync, &disk->rdev->flags)) {
2911 disk->head_position = 0;
2912 mddev->degraded++;
2913 if (disk->rdev)
2914 conf->fullsync = 1;
2915 }
2916 }
2917
2918 if (mddev->recovery_cp != MaxSector)
2919 printk(KERN_NOTICE "md/raid10:%s: not clean"
2920 " -- starting background reconstruction\n",
2921 mdname(mddev));
2922 printk(KERN_INFO
2923 "md/raid10:%s: active with %d out of %d devices\n",
2924 mdname(mddev), conf->raid_disks - mddev->degraded,
2925 conf->raid_disks);
2926 /*
2927 * Ok, everything is just fine now
2928 */
2929 mddev->dev_sectors = conf->dev_sectors;
2930 size = raid10_size(mddev, 0, 0);
2931 md_set_array_sectors(mddev, size);
2932 mddev->resync_max_sectors = size;
2933
2934 mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2935 mddev->queue->backing_dev_info.congested_data = mddev;
2936
2937 /* Calculate max read-ahead size.
2938 * We need to readahead at least twice a whole stripe....
2939 * maybe...
2940 */
2941 {
2942 int stripe = conf->raid_disks *
2943 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
2944 stripe /= conf->near_copies;
2945 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2946 mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2947 }
2948
2949 if (conf->near_copies < conf->raid_disks)
2950 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2951
2952 if (md_integrity_register(mddev))
2953 goto out_free_conf;
2954
2955 return 0;
2956
2957out_free_conf:
2958 md_unregister_thread(&mddev->thread);
2959 if (conf->r10bio_pool)
2960 mempool_destroy(conf->r10bio_pool);
2961 safe_put_page(conf->tmppage);
2962 kfree(conf->mirrors);
2963 kfree(conf);
2964 mddev->private = NULL;
2965out:
2966 return -EIO;
2967}
2968
2969static int stop(mddev_t *mddev)
2970{
2971 conf_t *conf = mddev->private;
2972
2973 raise_barrier(conf, 0);
2974 lower_barrier(conf);
2975
2976 md_unregister_thread(&mddev->thread);
2977 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2978 if (conf->r10bio_pool)
2979 mempool_destroy(conf->r10bio_pool);
2980 kfree(conf->mirrors);
2981 kfree(conf);
2982 mddev->private = NULL;
2983 return 0;
2984}
2985
2986static void raid10_quiesce(mddev_t *mddev, int state)
2987{
2988 conf_t *conf = mddev->private;
2989
2990 switch(state) {
2991 case 1:
2992 raise_barrier(conf, 0);
2993 break;
2994 case 0:
2995 lower_barrier(conf);
2996 break;
2997 }
2998}
2999
3000static void *raid10_takeover_raid0(mddev_t *mddev)
3001{
3002 mdk_rdev_t *rdev;
3003 conf_t *conf;
3004
3005 if (mddev->degraded > 0) {
3006 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3007 mdname(mddev));
3008 return ERR_PTR(-EINVAL);
3009 }
3010
3011 /* Set new parameters */
3012 mddev->new_level = 10;
3013 /* new layout: far_copies = 1, near_copies = 2 */
3014 mddev->new_layout = (1<<8) + 2;
3015 mddev->new_chunk_sectors = mddev->chunk_sectors;
3016 mddev->delta_disks = mddev->raid_disks;
3017 mddev->raid_disks *= 2;
3018 /* make sure it will be not marked as dirty */
3019 mddev->recovery_cp = MaxSector;
3020
3021 conf = setup_conf(mddev);
3022 if (!IS_ERR(conf)) {
3023 list_for_each_entry(rdev, &mddev->disks, same_set)
3024 if (rdev->raid_disk >= 0)
3025 rdev->new_raid_disk = rdev->raid_disk * 2;
3026 conf->barrier = 1;
3027 }
3028
3029 return conf;
3030}
3031
3032static void *raid10_takeover(mddev_t *mddev)
3033{
3034 struct raid0_private_data *raid0_priv;
3035
3036 /* raid10 can take over:
3037 * raid0 - providing it has only two drives
3038 */
3039 if (mddev->level == 0) {
3040 /* for raid0 takeover only one zone is supported */
3041 raid0_priv = mddev->private;
3042 if (raid0_priv->nr_strip_zones > 1) {
3043 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3044 " with more than one zone.\n",
3045 mdname(mddev));
3046 return ERR_PTR(-EINVAL);
3047 }
3048 return raid10_takeover_raid0(mddev);
3049 }
3050 return ERR_PTR(-EINVAL);
3051}
3052
3053static struct mdk_personality raid10_personality =
3054{
3055 .name = "raid10",
3056 .level = 10,
3057 .owner = THIS_MODULE,
3058 .make_request = make_request,
3059 .run = run,
3060 .stop = stop,
3061 .status = status,
3062 .error_handler = error,
3063 .hot_add_disk = raid10_add_disk,
3064 .hot_remove_disk= raid10_remove_disk,
3065 .spare_active = raid10_spare_active,
3066 .sync_request = sync_request,
3067 .quiesce = raid10_quiesce,
3068 .size = raid10_size,
3069 .takeover = raid10_takeover,
3070};
3071
3072static int __init raid_init(void)
3073{
3074 return register_md_personality(&raid10_personality);
3075}
3076
3077static void raid_exit(void)
3078{
3079 unregister_md_personality(&raid10_personality);
3080}
3081
3082module_init(raid_init);
3083module_exit(raid_exit);
3084MODULE_LICENSE("GPL");
3085MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
3086MODULE_ALIAS("md-personality-9"); /* RAID10 */
3087MODULE_ALIAS("md-raid10");
3088MODULE_ALIAS("md-level-10");