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