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 <trace/events/block.h>
29#include "md.h"
30#include "raid10.h"
31#include "raid0.h"
32#include "bitmap.h"
33
34/*
35 * RAID10 provides a combination of RAID0 and RAID1 functionality.
36 * The layout of data is defined by
37 * chunk_size
38 * raid_disks
39 * near_copies (stored in low byte of layout)
40 * far_copies (stored in second byte of layout)
41 * far_offset (stored in bit 16 of layout )
42 * use_far_sets (stored in bit 17 of layout )
43 * use_far_sets_bugfixed (stored in bit 18 of layout )
44 *
45 * The data to be stored is divided into chunks using chunksize. Each device
46 * is divided into far_copies sections. In each section, chunks are laid out
47 * in a style similar to raid0, but near_copies copies of each chunk is stored
48 * (each on a different drive). The starting device for each section is offset
49 * near_copies from the starting device of the previous section. Thus there
50 * are (near_copies * far_copies) of each chunk, and each is on a different
51 * drive. near_copies and far_copies must be at least one, and their product
52 * is at most raid_disks.
53 *
54 * If far_offset is true, then the far_copies are handled a bit differently.
55 * The copies are still in different stripes, but instead of being very far
56 * apart on disk, there are adjacent stripes.
57 *
58 * The far and offset algorithms are handled slightly differently if
59 * 'use_far_sets' is true. In this case, the array's devices are grouped into
60 * sets that are (near_copies * far_copies) in size. The far copied stripes
61 * are still shifted by 'near_copies' devices, but this shifting stays confined
62 * to the set rather than the entire array. This is done to improve the number
63 * of device combinations that can fail without causing the array to fail.
64 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
65 * on a device):
66 * A B C D A B C D E
67 * ... ...
68 * D A B C E A B C D
69 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
70 * [A B] [C D] [A B] [C D E]
71 * |...| |...| |...| | ... |
72 * [B A] [D C] [B A] [E C D]
73 */
74
75/*
76 * Number of guaranteed r10bios in case of extreme VM load:
77 */
78#define NR_RAID10_BIOS 256
79
80/* when we get a read error on a read-only array, we redirect to another
81 * device without failing the first device, or trying to over-write to
82 * correct the read error. To keep track of bad blocks on a per-bio
83 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
84 */
85#define IO_BLOCKED ((struct bio *)1)
86/* When we successfully write to a known bad-block, we need to remove the
87 * bad-block marking which must be done from process context. So we record
88 * the success by setting devs[n].bio to IO_MADE_GOOD
89 */
90#define IO_MADE_GOOD ((struct bio *)2)
91
92#define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
93
94/* When there are this many requests queued to be written by
95 * the raid10 thread, we become 'congested' to provide back-pressure
96 * for writeback.
97 */
98static int max_queued_requests = 1024;
99
100static void allow_barrier(struct r10conf *conf);
101static void lower_barrier(struct r10conf *conf);
102static int _enough(struct r10conf *conf, int previous, int ignore);
103static int enough(struct r10conf *conf, int ignore);
104static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
105 int *skipped);
106static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
107static void end_reshape_write(struct bio *bio);
108static void end_reshape(struct r10conf *conf);
109
110#define raid10_log(md, fmt, args...) \
111 do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid10 " fmt, ##args); } while (0)
112
113static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
114{
115 struct r10conf *conf = data;
116 int size = offsetof(struct r10bio, devs[conf->copies]);
117
118 /* allocate a r10bio with room for raid_disks entries in the
119 * bios array */
120 return kzalloc(size, gfp_flags);
121}
122
123static void r10bio_pool_free(void *r10_bio, void *data)
124{
125 kfree(r10_bio);
126}
127
128/* Maximum size of each resync request */
129#define RESYNC_BLOCK_SIZE (64*1024)
130#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
131/* amount of memory to reserve for resync requests */
132#define RESYNC_WINDOW (1024*1024)
133/* maximum number of concurrent requests, memory permitting */
134#define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
135
136/*
137 * When performing a resync, we need to read and compare, so
138 * we need as many pages are there are copies.
139 * When performing a recovery, we need 2 bios, one for read,
140 * one for write (we recover only one drive per r10buf)
141 *
142 */
143static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
144{
145 struct r10conf *conf = data;
146 struct page *page;
147 struct r10bio *r10_bio;
148 struct bio *bio;
149 int i, j;
150 int nalloc;
151
152 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
153 if (!r10_bio)
154 return NULL;
155
156 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
157 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
158 nalloc = conf->copies; /* resync */
159 else
160 nalloc = 2; /* recovery */
161
162 /*
163 * Allocate bios.
164 */
165 for (j = nalloc ; j-- ; ) {
166 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
167 if (!bio)
168 goto out_free_bio;
169 r10_bio->devs[j].bio = bio;
170 if (!conf->have_replacement)
171 continue;
172 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
173 if (!bio)
174 goto out_free_bio;
175 r10_bio->devs[j].repl_bio = bio;
176 }
177 /*
178 * Allocate RESYNC_PAGES data pages and attach them
179 * where needed.
180 */
181 for (j = 0 ; j < nalloc; j++) {
182 struct bio *rbio = r10_bio->devs[j].repl_bio;
183 bio = r10_bio->devs[j].bio;
184 for (i = 0; i < RESYNC_PAGES; i++) {
185 if (j > 0 && !test_bit(MD_RECOVERY_SYNC,
186 &conf->mddev->recovery)) {
187 /* we can share bv_page's during recovery
188 * and reshape */
189 struct bio *rbio = r10_bio->devs[0].bio;
190 page = rbio->bi_io_vec[i].bv_page;
191 get_page(page);
192 } else
193 page = alloc_page(gfp_flags);
194 if (unlikely(!page))
195 goto out_free_pages;
196
197 bio->bi_io_vec[i].bv_page = page;
198 if (rbio)
199 rbio->bi_io_vec[i].bv_page = page;
200 }
201 }
202
203 return r10_bio;
204
205out_free_pages:
206 for ( ; i > 0 ; i--)
207 safe_put_page(bio->bi_io_vec[i-1].bv_page);
208 while (j--)
209 for (i = 0; i < RESYNC_PAGES ; i++)
210 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
211 j = 0;
212out_free_bio:
213 for ( ; j < nalloc; j++) {
214 if (r10_bio->devs[j].bio)
215 bio_put(r10_bio->devs[j].bio);
216 if (r10_bio->devs[j].repl_bio)
217 bio_put(r10_bio->devs[j].repl_bio);
218 }
219 r10bio_pool_free(r10_bio, conf);
220 return NULL;
221}
222
223static void r10buf_pool_free(void *__r10_bio, void *data)
224{
225 int i;
226 struct r10conf *conf = data;
227 struct r10bio *r10bio = __r10_bio;
228 int j;
229
230 for (j=0; j < conf->copies; j++) {
231 struct bio *bio = r10bio->devs[j].bio;
232 if (bio) {
233 for (i = 0; i < RESYNC_PAGES; i++) {
234 safe_put_page(bio->bi_io_vec[i].bv_page);
235 bio->bi_io_vec[i].bv_page = NULL;
236 }
237 bio_put(bio);
238 }
239 bio = r10bio->devs[j].repl_bio;
240 if (bio)
241 bio_put(bio);
242 }
243 r10bio_pool_free(r10bio, conf);
244}
245
246static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
247{
248 int i;
249
250 for (i = 0; i < conf->copies; i++) {
251 struct bio **bio = & r10_bio->devs[i].bio;
252 if (!BIO_SPECIAL(*bio))
253 bio_put(*bio);
254 *bio = NULL;
255 bio = &r10_bio->devs[i].repl_bio;
256 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
257 bio_put(*bio);
258 *bio = NULL;
259 }
260}
261
262static void free_r10bio(struct r10bio *r10_bio)
263{
264 struct r10conf *conf = r10_bio->mddev->private;
265
266 put_all_bios(conf, r10_bio);
267 mempool_free(r10_bio, conf->r10bio_pool);
268}
269
270static void put_buf(struct r10bio *r10_bio)
271{
272 struct r10conf *conf = r10_bio->mddev->private;
273
274 mempool_free(r10_bio, conf->r10buf_pool);
275
276 lower_barrier(conf);
277}
278
279static void reschedule_retry(struct r10bio *r10_bio)
280{
281 unsigned long flags;
282 struct mddev *mddev = r10_bio->mddev;
283 struct r10conf *conf = mddev->private;
284
285 spin_lock_irqsave(&conf->device_lock, flags);
286 list_add(&r10_bio->retry_list, &conf->retry_list);
287 conf->nr_queued ++;
288 spin_unlock_irqrestore(&conf->device_lock, flags);
289
290 /* wake up frozen array... */
291 wake_up(&conf->wait_barrier);
292
293 md_wakeup_thread(mddev->thread);
294}
295
296/*
297 * raid_end_bio_io() is called when we have finished servicing a mirrored
298 * operation and are ready to return a success/failure code to the buffer
299 * cache layer.
300 */
301static void raid_end_bio_io(struct r10bio *r10_bio)
302{
303 struct bio *bio = r10_bio->master_bio;
304 int done;
305 struct r10conf *conf = r10_bio->mddev->private;
306
307 if (bio->bi_phys_segments) {
308 unsigned long flags;
309 spin_lock_irqsave(&conf->device_lock, flags);
310 bio->bi_phys_segments--;
311 done = (bio->bi_phys_segments == 0);
312 spin_unlock_irqrestore(&conf->device_lock, flags);
313 } else
314 done = 1;
315 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
316 bio->bi_error = -EIO;
317 if (done) {
318 bio_endio(bio);
319 /*
320 * Wake up any possible resync thread that waits for the device
321 * to go idle.
322 */
323 allow_barrier(conf);
324 }
325 free_r10bio(r10_bio);
326}
327
328/*
329 * Update disk head position estimator based on IRQ completion info.
330 */
331static inline void update_head_pos(int slot, struct r10bio *r10_bio)
332{
333 struct r10conf *conf = r10_bio->mddev->private;
334
335 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
336 r10_bio->devs[slot].addr + (r10_bio->sectors);
337}
338
339/*
340 * Find the disk number which triggered given bio
341 */
342static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
343 struct bio *bio, int *slotp, int *replp)
344{
345 int slot;
346 int repl = 0;
347
348 for (slot = 0; slot < conf->copies; slot++) {
349 if (r10_bio->devs[slot].bio == bio)
350 break;
351 if (r10_bio->devs[slot].repl_bio == bio) {
352 repl = 1;
353 break;
354 }
355 }
356
357 BUG_ON(slot == conf->copies);
358 update_head_pos(slot, r10_bio);
359
360 if (slotp)
361 *slotp = slot;
362 if (replp)
363 *replp = repl;
364 return r10_bio->devs[slot].devnum;
365}
366
367static void raid10_end_read_request(struct bio *bio)
368{
369 int uptodate = !bio->bi_error;
370 struct r10bio *r10_bio = bio->bi_private;
371 int slot, dev;
372 struct md_rdev *rdev;
373 struct r10conf *conf = r10_bio->mddev->private;
374
375 slot = r10_bio->read_slot;
376 dev = r10_bio->devs[slot].devnum;
377 rdev = r10_bio->devs[slot].rdev;
378 /*
379 * this branch is our 'one mirror IO has finished' event handler:
380 */
381 update_head_pos(slot, r10_bio);
382
383 if (uptodate) {
384 /*
385 * Set R10BIO_Uptodate in our master bio, so that
386 * we will return a good error code to the higher
387 * levels even if IO on some other mirrored buffer fails.
388 *
389 * The 'master' represents the composite IO operation to
390 * user-side. So if something waits for IO, then it will
391 * wait for the 'master' bio.
392 */
393 set_bit(R10BIO_Uptodate, &r10_bio->state);
394 } else {
395 /* If all other devices that store this block have
396 * failed, we want to return the error upwards rather
397 * than fail the last device. Here we redefine
398 * "uptodate" to mean "Don't want to retry"
399 */
400 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
401 rdev->raid_disk))
402 uptodate = 1;
403 }
404 if (uptodate) {
405 raid_end_bio_io(r10_bio);
406 rdev_dec_pending(rdev, conf->mddev);
407 } else {
408 /*
409 * oops, read error - keep the refcount on the rdev
410 */
411 char b[BDEVNAME_SIZE];
412 pr_err_ratelimited("md/raid10:%s: %s: rescheduling sector %llu\n",
413 mdname(conf->mddev),
414 bdevname(rdev->bdev, b),
415 (unsigned long long)r10_bio->sector);
416 set_bit(R10BIO_ReadError, &r10_bio->state);
417 reschedule_retry(r10_bio);
418 }
419}
420
421static void close_write(struct r10bio *r10_bio)
422{
423 /* clear the bitmap if all writes complete successfully */
424 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
425 r10_bio->sectors,
426 !test_bit(R10BIO_Degraded, &r10_bio->state),
427 0);
428 md_write_end(r10_bio->mddev);
429}
430
431static void one_write_done(struct r10bio *r10_bio)
432{
433 if (atomic_dec_and_test(&r10_bio->remaining)) {
434 if (test_bit(R10BIO_WriteError, &r10_bio->state))
435 reschedule_retry(r10_bio);
436 else {
437 close_write(r10_bio);
438 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
439 reschedule_retry(r10_bio);
440 else
441 raid_end_bio_io(r10_bio);
442 }
443 }
444}
445
446static void raid10_end_write_request(struct bio *bio)
447{
448 struct r10bio *r10_bio = bio->bi_private;
449 int dev;
450 int dec_rdev = 1;
451 struct r10conf *conf = r10_bio->mddev->private;
452 int slot, repl;
453 struct md_rdev *rdev = NULL;
454 struct bio *to_put = NULL;
455 bool discard_error;
456
457 discard_error = bio->bi_error && bio_op(bio) == REQ_OP_DISCARD;
458
459 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
460
461 if (repl)
462 rdev = conf->mirrors[dev].replacement;
463 if (!rdev) {
464 smp_rmb();
465 repl = 0;
466 rdev = conf->mirrors[dev].rdev;
467 }
468 /*
469 * this branch is our 'one mirror IO has finished' event handler:
470 */
471 if (bio->bi_error && !discard_error) {
472 if (repl)
473 /* Never record new bad blocks to replacement,
474 * just fail it.
475 */
476 md_error(rdev->mddev, rdev);
477 else {
478 set_bit(WriteErrorSeen, &rdev->flags);
479 if (!test_and_set_bit(WantReplacement, &rdev->flags))
480 set_bit(MD_RECOVERY_NEEDED,
481 &rdev->mddev->recovery);
482
483 dec_rdev = 0;
484 if (test_bit(FailFast, &rdev->flags) &&
485 (bio->bi_opf & MD_FAILFAST)) {
486 md_error(rdev->mddev, rdev);
487 if (!test_bit(Faulty, &rdev->flags))
488 /* This is the only remaining device,
489 * We need to retry the write without
490 * FailFast
491 */
492 set_bit(R10BIO_WriteError, &r10_bio->state);
493 else {
494 r10_bio->devs[slot].bio = NULL;
495 to_put = bio;
496 dec_rdev = 1;
497 }
498 } else
499 set_bit(R10BIO_WriteError, &r10_bio->state);
500 }
501 } else {
502 /*
503 * Set R10BIO_Uptodate in our master bio, so that
504 * we will return a good error code for to the higher
505 * levels even if IO on some other mirrored buffer fails.
506 *
507 * The 'master' represents the composite IO operation to
508 * user-side. So if something waits for IO, then it will
509 * wait for the 'master' bio.
510 */
511 sector_t first_bad;
512 int bad_sectors;
513
514 /*
515 * Do not set R10BIO_Uptodate if the current device is
516 * rebuilding or Faulty. This is because we cannot use
517 * such device for properly reading the data back (we could
518 * potentially use it, if the current write would have felt
519 * before rdev->recovery_offset, but for simplicity we don't
520 * check this here.
521 */
522 if (test_bit(In_sync, &rdev->flags) &&
523 !test_bit(Faulty, &rdev->flags))
524 set_bit(R10BIO_Uptodate, &r10_bio->state);
525
526 /* Maybe we can clear some bad blocks. */
527 if (is_badblock(rdev,
528 r10_bio->devs[slot].addr,
529 r10_bio->sectors,
530 &first_bad, &bad_sectors) && !discard_error) {
531 bio_put(bio);
532 if (repl)
533 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
534 else
535 r10_bio->devs[slot].bio = IO_MADE_GOOD;
536 dec_rdev = 0;
537 set_bit(R10BIO_MadeGood, &r10_bio->state);
538 }
539 }
540
541 /*
542 *
543 * Let's see if all mirrored write operations have finished
544 * already.
545 */
546 one_write_done(r10_bio);
547 if (dec_rdev)
548 rdev_dec_pending(rdev, conf->mddev);
549 if (to_put)
550 bio_put(to_put);
551}
552
553/*
554 * RAID10 layout manager
555 * As well as the chunksize and raid_disks count, there are two
556 * parameters: near_copies and far_copies.
557 * near_copies * far_copies must be <= raid_disks.
558 * Normally one of these will be 1.
559 * If both are 1, we get raid0.
560 * If near_copies == raid_disks, we get raid1.
561 *
562 * Chunks are laid out in raid0 style with near_copies copies of the
563 * first chunk, followed by near_copies copies of the next chunk and
564 * so on.
565 * If far_copies > 1, then after 1/far_copies of the array has been assigned
566 * as described above, we start again with a device offset of near_copies.
567 * So we effectively have another copy of the whole array further down all
568 * the drives, but with blocks on different drives.
569 * With this layout, and block is never stored twice on the one device.
570 *
571 * raid10_find_phys finds the sector offset of a given virtual sector
572 * on each device that it is on.
573 *
574 * raid10_find_virt does the reverse mapping, from a device and a
575 * sector offset to a virtual address
576 */
577
578static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
579{
580 int n,f;
581 sector_t sector;
582 sector_t chunk;
583 sector_t stripe;
584 int dev;
585 int slot = 0;
586 int last_far_set_start, last_far_set_size;
587
588 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
589 last_far_set_start *= geo->far_set_size;
590
591 last_far_set_size = geo->far_set_size;
592 last_far_set_size += (geo->raid_disks % geo->far_set_size);
593
594 /* now calculate first sector/dev */
595 chunk = r10bio->sector >> geo->chunk_shift;
596 sector = r10bio->sector & geo->chunk_mask;
597
598 chunk *= geo->near_copies;
599 stripe = chunk;
600 dev = sector_div(stripe, geo->raid_disks);
601 if (geo->far_offset)
602 stripe *= geo->far_copies;
603
604 sector += stripe << geo->chunk_shift;
605
606 /* and calculate all the others */
607 for (n = 0; n < geo->near_copies; n++) {
608 int d = dev;
609 int set;
610 sector_t s = sector;
611 r10bio->devs[slot].devnum = d;
612 r10bio->devs[slot].addr = s;
613 slot++;
614
615 for (f = 1; f < geo->far_copies; f++) {
616 set = d / geo->far_set_size;
617 d += geo->near_copies;
618
619 if ((geo->raid_disks % geo->far_set_size) &&
620 (d > last_far_set_start)) {
621 d -= last_far_set_start;
622 d %= last_far_set_size;
623 d += last_far_set_start;
624 } else {
625 d %= geo->far_set_size;
626 d += geo->far_set_size * set;
627 }
628 s += geo->stride;
629 r10bio->devs[slot].devnum = d;
630 r10bio->devs[slot].addr = s;
631 slot++;
632 }
633 dev++;
634 if (dev >= geo->raid_disks) {
635 dev = 0;
636 sector += (geo->chunk_mask + 1);
637 }
638 }
639}
640
641static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
642{
643 struct geom *geo = &conf->geo;
644
645 if (conf->reshape_progress != MaxSector &&
646 ((r10bio->sector >= conf->reshape_progress) !=
647 conf->mddev->reshape_backwards)) {
648 set_bit(R10BIO_Previous, &r10bio->state);
649 geo = &conf->prev;
650 } else
651 clear_bit(R10BIO_Previous, &r10bio->state);
652
653 __raid10_find_phys(geo, r10bio);
654}
655
656static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
657{
658 sector_t offset, chunk, vchunk;
659 /* Never use conf->prev as this is only called during resync
660 * or recovery, so reshape isn't happening
661 */
662 struct geom *geo = &conf->geo;
663 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
664 int far_set_size = geo->far_set_size;
665 int last_far_set_start;
666
667 if (geo->raid_disks % geo->far_set_size) {
668 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
669 last_far_set_start *= geo->far_set_size;
670
671 if (dev >= last_far_set_start) {
672 far_set_size = geo->far_set_size;
673 far_set_size += (geo->raid_disks % geo->far_set_size);
674 far_set_start = last_far_set_start;
675 }
676 }
677
678 offset = sector & geo->chunk_mask;
679 if (geo->far_offset) {
680 int fc;
681 chunk = sector >> geo->chunk_shift;
682 fc = sector_div(chunk, geo->far_copies);
683 dev -= fc * geo->near_copies;
684 if (dev < far_set_start)
685 dev += far_set_size;
686 } else {
687 while (sector >= geo->stride) {
688 sector -= geo->stride;
689 if (dev < (geo->near_copies + far_set_start))
690 dev += far_set_size - geo->near_copies;
691 else
692 dev -= geo->near_copies;
693 }
694 chunk = sector >> geo->chunk_shift;
695 }
696 vchunk = chunk * geo->raid_disks + dev;
697 sector_div(vchunk, geo->near_copies);
698 return (vchunk << geo->chunk_shift) + offset;
699}
700
701/*
702 * This routine returns the disk from which the requested read should
703 * be done. There is a per-array 'next expected sequential IO' sector
704 * number - if this matches on the next IO then we use the last disk.
705 * There is also a per-disk 'last know head position' sector that is
706 * maintained from IRQ contexts, both the normal and the resync IO
707 * completion handlers update this position correctly. If there is no
708 * perfect sequential match then we pick the disk whose head is closest.
709 *
710 * If there are 2 mirrors in the same 2 devices, performance degrades
711 * because position is mirror, not device based.
712 *
713 * The rdev for the device selected will have nr_pending incremented.
714 */
715
716/*
717 * FIXME: possibly should rethink readbalancing and do it differently
718 * depending on near_copies / far_copies geometry.
719 */
720static struct md_rdev *read_balance(struct r10conf *conf,
721 struct r10bio *r10_bio,
722 int *max_sectors)
723{
724 const sector_t this_sector = r10_bio->sector;
725 int disk, slot;
726 int sectors = r10_bio->sectors;
727 int best_good_sectors;
728 sector_t new_distance, best_dist;
729 struct md_rdev *best_rdev, *rdev = NULL;
730 int do_balance;
731 int best_slot;
732 struct geom *geo = &conf->geo;
733
734 raid10_find_phys(conf, r10_bio);
735 rcu_read_lock();
736 sectors = r10_bio->sectors;
737 best_slot = -1;
738 best_rdev = NULL;
739 best_dist = MaxSector;
740 best_good_sectors = 0;
741 do_balance = 1;
742 clear_bit(R10BIO_FailFast, &r10_bio->state);
743 /*
744 * Check if we can balance. We can balance on the whole
745 * device if no resync is going on (recovery is ok), or below
746 * the resync window. We take the first readable disk when
747 * above the resync window.
748 */
749 if (conf->mddev->recovery_cp < MaxSector
750 && (this_sector + sectors >= conf->next_resync))
751 do_balance = 0;
752
753 for (slot = 0; slot < conf->copies ; slot++) {
754 sector_t first_bad;
755 int bad_sectors;
756 sector_t dev_sector;
757
758 if (r10_bio->devs[slot].bio == IO_BLOCKED)
759 continue;
760 disk = r10_bio->devs[slot].devnum;
761 rdev = rcu_dereference(conf->mirrors[disk].replacement);
762 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
763 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
764 rdev = rcu_dereference(conf->mirrors[disk].rdev);
765 if (rdev == NULL ||
766 test_bit(Faulty, &rdev->flags))
767 continue;
768 if (!test_bit(In_sync, &rdev->flags) &&
769 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
770 continue;
771
772 dev_sector = r10_bio->devs[slot].addr;
773 if (is_badblock(rdev, dev_sector, sectors,
774 &first_bad, &bad_sectors)) {
775 if (best_dist < MaxSector)
776 /* Already have a better slot */
777 continue;
778 if (first_bad <= dev_sector) {
779 /* Cannot read here. If this is the
780 * 'primary' device, then we must not read
781 * beyond 'bad_sectors' from another device.
782 */
783 bad_sectors -= (dev_sector - first_bad);
784 if (!do_balance && sectors > bad_sectors)
785 sectors = bad_sectors;
786 if (best_good_sectors > sectors)
787 best_good_sectors = sectors;
788 } else {
789 sector_t good_sectors =
790 first_bad - dev_sector;
791 if (good_sectors > best_good_sectors) {
792 best_good_sectors = good_sectors;
793 best_slot = slot;
794 best_rdev = rdev;
795 }
796 if (!do_balance)
797 /* Must read from here */
798 break;
799 }
800 continue;
801 } else
802 best_good_sectors = sectors;
803
804 if (!do_balance)
805 break;
806
807 if (best_slot >= 0)
808 /* At least 2 disks to choose from so failfast is OK */
809 set_bit(R10BIO_FailFast, &r10_bio->state);
810 /* This optimisation is debatable, and completely destroys
811 * sequential read speed for 'far copies' arrays. So only
812 * keep it for 'near' arrays, and review those later.
813 */
814 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
815 new_distance = 0;
816
817 /* for far > 1 always use the lowest address */
818 else if (geo->far_copies > 1)
819 new_distance = r10_bio->devs[slot].addr;
820 else
821 new_distance = abs(r10_bio->devs[slot].addr -
822 conf->mirrors[disk].head_position);
823 if (new_distance < best_dist) {
824 best_dist = new_distance;
825 best_slot = slot;
826 best_rdev = rdev;
827 }
828 }
829 if (slot >= conf->copies) {
830 slot = best_slot;
831 rdev = best_rdev;
832 }
833
834 if (slot >= 0) {
835 atomic_inc(&rdev->nr_pending);
836 r10_bio->read_slot = slot;
837 } else
838 rdev = NULL;
839 rcu_read_unlock();
840 *max_sectors = best_good_sectors;
841
842 return rdev;
843}
844
845static int raid10_congested(struct mddev *mddev, int bits)
846{
847 struct r10conf *conf = mddev->private;
848 int i, ret = 0;
849
850 if ((bits & (1 << WB_async_congested)) &&
851 conf->pending_count >= max_queued_requests)
852 return 1;
853
854 rcu_read_lock();
855 for (i = 0;
856 (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
857 && ret == 0;
858 i++) {
859 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
860 if (rdev && !test_bit(Faulty, &rdev->flags)) {
861 struct request_queue *q = bdev_get_queue(rdev->bdev);
862
863 ret |= bdi_congested(&q->backing_dev_info, bits);
864 }
865 }
866 rcu_read_unlock();
867 return ret;
868}
869
870static void flush_pending_writes(struct r10conf *conf)
871{
872 /* Any writes that have been queued but are awaiting
873 * bitmap updates get flushed here.
874 */
875 spin_lock_irq(&conf->device_lock);
876
877 if (conf->pending_bio_list.head) {
878 struct bio *bio;
879 bio = bio_list_get(&conf->pending_bio_list);
880 conf->pending_count = 0;
881 spin_unlock_irq(&conf->device_lock);
882 /* flush any pending bitmap writes to disk
883 * before proceeding w/ I/O */
884 bitmap_unplug(conf->mddev->bitmap);
885 wake_up(&conf->wait_barrier);
886
887 while (bio) { /* submit pending writes */
888 struct bio *next = bio->bi_next;
889 struct md_rdev *rdev = (void*)bio->bi_bdev;
890 bio->bi_next = NULL;
891 bio->bi_bdev = rdev->bdev;
892 if (test_bit(Faulty, &rdev->flags)) {
893 bio->bi_error = -EIO;
894 bio_endio(bio);
895 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
896 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
897 /* Just ignore it */
898 bio_endio(bio);
899 else
900 generic_make_request(bio);
901 bio = next;
902 }
903 } else
904 spin_unlock_irq(&conf->device_lock);
905}
906
907/* Barriers....
908 * Sometimes we need to suspend IO while we do something else,
909 * either some resync/recovery, or reconfigure the array.
910 * To do this we raise a 'barrier'.
911 * The 'barrier' is a counter that can be raised multiple times
912 * to count how many activities are happening which preclude
913 * normal IO.
914 * We can only raise the barrier if there is no pending IO.
915 * i.e. if nr_pending == 0.
916 * We choose only to raise the barrier if no-one is waiting for the
917 * barrier to go down. This means that as soon as an IO request
918 * is ready, no other operations which require a barrier will start
919 * until the IO request has had a chance.
920 *
921 * So: regular IO calls 'wait_barrier'. When that returns there
922 * is no backgroup IO happening, It must arrange to call
923 * allow_barrier when it has finished its IO.
924 * backgroup IO calls must call raise_barrier. Once that returns
925 * there is no normal IO happeing. It must arrange to call
926 * lower_barrier when the particular background IO completes.
927 */
928
929static void raise_barrier(struct r10conf *conf, int force)
930{
931 BUG_ON(force && !conf->barrier);
932 spin_lock_irq(&conf->resync_lock);
933
934 /* Wait until no block IO is waiting (unless 'force') */
935 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
936 conf->resync_lock);
937
938 /* block any new IO from starting */
939 conf->barrier++;
940
941 /* Now wait for all pending IO to complete */
942 wait_event_lock_irq(conf->wait_barrier,
943 !atomic_read(&conf->nr_pending) && conf->barrier < RESYNC_DEPTH,
944 conf->resync_lock);
945
946 spin_unlock_irq(&conf->resync_lock);
947}
948
949static void lower_barrier(struct r10conf *conf)
950{
951 unsigned long flags;
952 spin_lock_irqsave(&conf->resync_lock, flags);
953 conf->barrier--;
954 spin_unlock_irqrestore(&conf->resync_lock, flags);
955 wake_up(&conf->wait_barrier);
956}
957
958static void wait_barrier(struct r10conf *conf)
959{
960 spin_lock_irq(&conf->resync_lock);
961 if (conf->barrier) {
962 conf->nr_waiting++;
963 /* Wait for the barrier to drop.
964 * However if there are already pending
965 * requests (preventing the barrier from
966 * rising completely), and the
967 * pre-process bio queue isn't empty,
968 * then don't wait, as we need to empty
969 * that queue to get the nr_pending
970 * count down.
971 */
972 raid10_log(conf->mddev, "wait barrier");
973 wait_event_lock_irq(conf->wait_barrier,
974 !conf->barrier ||
975 (atomic_read(&conf->nr_pending) &&
976 current->bio_list &&
977 (!bio_list_empty(¤t->bio_list[0]) ||
978 !bio_list_empty(¤t->bio_list[1]))),
979 conf->resync_lock);
980 conf->nr_waiting--;
981 if (!conf->nr_waiting)
982 wake_up(&conf->wait_barrier);
983 }
984 atomic_inc(&conf->nr_pending);
985 spin_unlock_irq(&conf->resync_lock);
986}
987
988static void allow_barrier(struct r10conf *conf)
989{
990 if ((atomic_dec_and_test(&conf->nr_pending)) ||
991 (conf->array_freeze_pending))
992 wake_up(&conf->wait_barrier);
993}
994
995static void freeze_array(struct r10conf *conf, int extra)
996{
997 /* stop syncio and normal IO and wait for everything to
998 * go quiet.
999 * We increment barrier and nr_waiting, and then
1000 * wait until nr_pending match nr_queued+extra
1001 * This is called in the context of one normal IO request
1002 * that has failed. Thus any sync request that might be pending
1003 * will be blocked by nr_pending, and we need to wait for
1004 * pending IO requests to complete or be queued for re-try.
1005 * Thus the number queued (nr_queued) plus this request (extra)
1006 * must match the number of pending IOs (nr_pending) before
1007 * we continue.
1008 */
1009 spin_lock_irq(&conf->resync_lock);
1010 conf->array_freeze_pending++;
1011 conf->barrier++;
1012 conf->nr_waiting++;
1013 wait_event_lock_irq_cmd(conf->wait_barrier,
1014 atomic_read(&conf->nr_pending) == conf->nr_queued+extra,
1015 conf->resync_lock,
1016 flush_pending_writes(conf));
1017
1018 conf->array_freeze_pending--;
1019 spin_unlock_irq(&conf->resync_lock);
1020}
1021
1022static void unfreeze_array(struct r10conf *conf)
1023{
1024 /* reverse the effect of the freeze */
1025 spin_lock_irq(&conf->resync_lock);
1026 conf->barrier--;
1027 conf->nr_waiting--;
1028 wake_up(&conf->wait_barrier);
1029 spin_unlock_irq(&conf->resync_lock);
1030}
1031
1032static sector_t choose_data_offset(struct r10bio *r10_bio,
1033 struct md_rdev *rdev)
1034{
1035 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1036 test_bit(R10BIO_Previous, &r10_bio->state))
1037 return rdev->data_offset;
1038 else
1039 return rdev->new_data_offset;
1040}
1041
1042struct raid10_plug_cb {
1043 struct blk_plug_cb cb;
1044 struct bio_list pending;
1045 int pending_cnt;
1046};
1047
1048static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1049{
1050 struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1051 cb);
1052 struct mddev *mddev = plug->cb.data;
1053 struct r10conf *conf = mddev->private;
1054 struct bio *bio;
1055
1056 if (from_schedule || current->bio_list) {
1057 spin_lock_irq(&conf->device_lock);
1058 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1059 conf->pending_count += plug->pending_cnt;
1060 spin_unlock_irq(&conf->device_lock);
1061 wake_up(&conf->wait_barrier);
1062 md_wakeup_thread(mddev->thread);
1063 kfree(plug);
1064 return;
1065 }
1066
1067 /* we aren't scheduling, so we can do the write-out directly. */
1068 bio = bio_list_get(&plug->pending);
1069 bitmap_unplug(mddev->bitmap);
1070 wake_up(&conf->wait_barrier);
1071
1072 while (bio) { /* submit pending writes */
1073 struct bio *next = bio->bi_next;
1074 struct md_rdev *rdev = (void*)bio->bi_bdev;
1075 bio->bi_next = NULL;
1076 bio->bi_bdev = rdev->bdev;
1077 if (test_bit(Faulty, &rdev->flags)) {
1078 bio->bi_error = -EIO;
1079 bio_endio(bio);
1080 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
1081 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1082 /* Just ignore it */
1083 bio_endio(bio);
1084 else
1085 generic_make_request(bio);
1086 bio = next;
1087 }
1088 kfree(plug);
1089}
1090
1091static void raid10_read_request(struct mddev *mddev, struct bio *bio,
1092 struct r10bio *r10_bio)
1093{
1094 struct r10conf *conf = mddev->private;
1095 struct bio *read_bio;
1096 const int op = bio_op(bio);
1097 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1098 int sectors_handled;
1099 int max_sectors;
1100 sector_t sectors;
1101 struct md_rdev *rdev;
1102 int slot;
1103
1104 /*
1105 * Register the new request and wait if the reconstruction
1106 * thread has put up a bar for new requests.
1107 * Continue immediately if no resync is active currently.
1108 */
1109 wait_barrier(conf);
1110
1111 sectors = bio_sectors(bio);
1112 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1113 bio->bi_iter.bi_sector < conf->reshape_progress &&
1114 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1115 /*
1116 * IO spans the reshape position. Need to wait for reshape to
1117 * pass
1118 */
1119 raid10_log(conf->mddev, "wait reshape");
1120 allow_barrier(conf);
1121 wait_event(conf->wait_barrier,
1122 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1123 conf->reshape_progress >= bio->bi_iter.bi_sector +
1124 sectors);
1125 wait_barrier(conf);
1126 }
1127
1128read_again:
1129 rdev = read_balance(conf, r10_bio, &max_sectors);
1130 if (!rdev) {
1131 raid_end_bio_io(r10_bio);
1132 return;
1133 }
1134 slot = r10_bio->read_slot;
1135
1136 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1137 bio_trim(read_bio, r10_bio->sector - bio->bi_iter.bi_sector,
1138 max_sectors);
1139
1140 r10_bio->devs[slot].bio = read_bio;
1141 r10_bio->devs[slot].rdev = rdev;
1142
1143 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1144 choose_data_offset(r10_bio, rdev);
1145 read_bio->bi_bdev = rdev->bdev;
1146 read_bio->bi_end_io = raid10_end_read_request;
1147 bio_set_op_attrs(read_bio, op, do_sync);
1148 if (test_bit(FailFast, &rdev->flags) &&
1149 test_bit(R10BIO_FailFast, &r10_bio->state))
1150 read_bio->bi_opf |= MD_FAILFAST;
1151 read_bio->bi_private = r10_bio;
1152
1153 if (mddev->gendisk)
1154 trace_block_bio_remap(bdev_get_queue(read_bio->bi_bdev),
1155 read_bio, disk_devt(mddev->gendisk),
1156 r10_bio->sector);
1157 if (max_sectors < r10_bio->sectors) {
1158 /*
1159 * Could not read all from this device, so we will need another
1160 * r10_bio.
1161 */
1162 sectors_handled = (r10_bio->sector + max_sectors
1163 - bio->bi_iter.bi_sector);
1164 r10_bio->sectors = max_sectors;
1165 spin_lock_irq(&conf->device_lock);
1166 if (bio->bi_phys_segments == 0)
1167 bio->bi_phys_segments = 2;
1168 else
1169 bio->bi_phys_segments++;
1170 spin_unlock_irq(&conf->device_lock);
1171 /*
1172 * Cannot call generic_make_request directly as that will be
1173 * queued in __generic_make_request and subsequent
1174 * mempool_alloc might block waiting for it. so hand bio over
1175 * to raid10d.
1176 */
1177 reschedule_retry(r10_bio);
1178
1179 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1180
1181 r10_bio->master_bio = bio;
1182 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1183 r10_bio->state = 0;
1184 r10_bio->mddev = mddev;
1185 r10_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1186 goto read_again;
1187 } else
1188 generic_make_request(read_bio);
1189 return;
1190}
1191
1192static void raid10_write_request(struct mddev *mddev, struct bio *bio,
1193 struct r10bio *r10_bio)
1194{
1195 struct r10conf *conf = mddev->private;
1196 int i;
1197 const int op = bio_op(bio);
1198 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1199 const unsigned long do_fua = (bio->bi_opf & REQ_FUA);
1200 unsigned long flags;
1201 struct md_rdev *blocked_rdev;
1202 struct blk_plug_cb *cb;
1203 struct raid10_plug_cb *plug = NULL;
1204 sector_t sectors;
1205 int sectors_handled;
1206 int max_sectors;
1207
1208 md_write_start(mddev, bio);
1209
1210 /*
1211 * Register the new request and wait if the reconstruction
1212 * thread has put up a bar for new requests.
1213 * Continue immediately if no resync is active currently.
1214 */
1215 wait_barrier(conf);
1216
1217 sectors = bio_sectors(bio);
1218 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1219 bio->bi_iter.bi_sector < conf->reshape_progress &&
1220 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1221 /*
1222 * IO spans the reshape position. Need to wait for reshape to
1223 * pass
1224 */
1225 raid10_log(conf->mddev, "wait reshape");
1226 allow_barrier(conf);
1227 wait_event(conf->wait_barrier,
1228 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1229 conf->reshape_progress >= bio->bi_iter.bi_sector +
1230 sectors);
1231 wait_barrier(conf);
1232 }
1233
1234 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1235 (mddev->reshape_backwards
1236 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1237 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1238 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1239 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1240 /* Need to update reshape_position in metadata */
1241 mddev->reshape_position = conf->reshape_progress;
1242 set_mask_bits(&mddev->sb_flags, 0,
1243 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1244 md_wakeup_thread(mddev->thread);
1245 raid10_log(conf->mddev, "wait reshape metadata");
1246 wait_event(mddev->sb_wait,
1247 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
1248
1249 conf->reshape_safe = mddev->reshape_position;
1250 }
1251
1252 if (conf->pending_count >= max_queued_requests) {
1253 md_wakeup_thread(mddev->thread);
1254 raid10_log(mddev, "wait queued");
1255 wait_event(conf->wait_barrier,
1256 conf->pending_count < max_queued_requests);
1257 }
1258 /* first select target devices under rcu_lock and
1259 * inc refcount on their rdev. Record them by setting
1260 * bios[x] to bio
1261 * If there are known/acknowledged bad blocks on any device
1262 * on which we have seen a write error, we want to avoid
1263 * writing to those blocks. This potentially requires several
1264 * writes to write around the bad blocks. Each set of writes
1265 * gets its own r10_bio with a set of bios attached. The number
1266 * of r10_bios is recored in bio->bi_phys_segments just as with
1267 * the read case.
1268 */
1269
1270 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1271 raid10_find_phys(conf, r10_bio);
1272retry_write:
1273 blocked_rdev = NULL;
1274 rcu_read_lock();
1275 max_sectors = r10_bio->sectors;
1276
1277 for (i = 0; i < conf->copies; i++) {
1278 int d = r10_bio->devs[i].devnum;
1279 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1280 struct md_rdev *rrdev = rcu_dereference(
1281 conf->mirrors[d].replacement);
1282 if (rdev == rrdev)
1283 rrdev = NULL;
1284 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1285 atomic_inc(&rdev->nr_pending);
1286 blocked_rdev = rdev;
1287 break;
1288 }
1289 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1290 atomic_inc(&rrdev->nr_pending);
1291 blocked_rdev = rrdev;
1292 break;
1293 }
1294 if (rdev && (test_bit(Faulty, &rdev->flags)))
1295 rdev = NULL;
1296 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1297 rrdev = NULL;
1298
1299 r10_bio->devs[i].bio = NULL;
1300 r10_bio->devs[i].repl_bio = NULL;
1301
1302 if (!rdev && !rrdev) {
1303 set_bit(R10BIO_Degraded, &r10_bio->state);
1304 continue;
1305 }
1306 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1307 sector_t first_bad;
1308 sector_t dev_sector = r10_bio->devs[i].addr;
1309 int bad_sectors;
1310 int is_bad;
1311
1312 is_bad = is_badblock(rdev, dev_sector, max_sectors,
1313 &first_bad, &bad_sectors);
1314 if (is_bad < 0) {
1315 /* Mustn't write here until the bad block
1316 * is acknowledged
1317 */
1318 atomic_inc(&rdev->nr_pending);
1319 set_bit(BlockedBadBlocks, &rdev->flags);
1320 blocked_rdev = rdev;
1321 break;
1322 }
1323 if (is_bad && first_bad <= dev_sector) {
1324 /* Cannot write here at all */
1325 bad_sectors -= (dev_sector - first_bad);
1326 if (bad_sectors < max_sectors)
1327 /* Mustn't write more than bad_sectors
1328 * to other devices yet
1329 */
1330 max_sectors = bad_sectors;
1331 /* We don't set R10BIO_Degraded as that
1332 * only applies if the disk is missing,
1333 * so it might be re-added, and we want to
1334 * know to recover this chunk.
1335 * In this case the device is here, and the
1336 * fact that this chunk is not in-sync is
1337 * recorded in the bad block log.
1338 */
1339 continue;
1340 }
1341 if (is_bad) {
1342 int good_sectors = first_bad - dev_sector;
1343 if (good_sectors < max_sectors)
1344 max_sectors = good_sectors;
1345 }
1346 }
1347 if (rdev) {
1348 r10_bio->devs[i].bio = bio;
1349 atomic_inc(&rdev->nr_pending);
1350 }
1351 if (rrdev) {
1352 r10_bio->devs[i].repl_bio = bio;
1353 atomic_inc(&rrdev->nr_pending);
1354 }
1355 }
1356 rcu_read_unlock();
1357
1358 if (unlikely(blocked_rdev)) {
1359 /* Have to wait for this device to get unblocked, then retry */
1360 int j;
1361 int d;
1362
1363 for (j = 0; j < i; j++) {
1364 if (r10_bio->devs[j].bio) {
1365 d = r10_bio->devs[j].devnum;
1366 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1367 }
1368 if (r10_bio->devs[j].repl_bio) {
1369 struct md_rdev *rdev;
1370 d = r10_bio->devs[j].devnum;
1371 rdev = conf->mirrors[d].replacement;
1372 if (!rdev) {
1373 /* Race with remove_disk */
1374 smp_mb();
1375 rdev = conf->mirrors[d].rdev;
1376 }
1377 rdev_dec_pending(rdev, mddev);
1378 }
1379 }
1380 allow_barrier(conf);
1381 raid10_log(conf->mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1382 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1383 wait_barrier(conf);
1384 goto retry_write;
1385 }
1386
1387 if (max_sectors < r10_bio->sectors) {
1388 /* We are splitting this into multiple parts, so
1389 * we need to prepare for allocating another r10_bio.
1390 */
1391 r10_bio->sectors = max_sectors;
1392 spin_lock_irq(&conf->device_lock);
1393 if (bio->bi_phys_segments == 0)
1394 bio->bi_phys_segments = 2;
1395 else
1396 bio->bi_phys_segments++;
1397 spin_unlock_irq(&conf->device_lock);
1398 }
1399 sectors_handled = r10_bio->sector + max_sectors -
1400 bio->bi_iter.bi_sector;
1401
1402 atomic_set(&r10_bio->remaining, 1);
1403 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1404
1405 for (i = 0; i < conf->copies; i++) {
1406 struct bio *mbio;
1407 int d = r10_bio->devs[i].devnum;
1408 if (r10_bio->devs[i].bio) {
1409 struct md_rdev *rdev = conf->mirrors[d].rdev;
1410 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1411 bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1412 max_sectors);
1413 r10_bio->devs[i].bio = mbio;
1414
1415 mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+
1416 choose_data_offset(r10_bio, rdev));
1417 mbio->bi_bdev = rdev->bdev;
1418 mbio->bi_end_io = raid10_end_write_request;
1419 bio_set_op_attrs(mbio, op, do_sync | do_fua);
1420 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags) &&
1421 enough(conf, d))
1422 mbio->bi_opf |= MD_FAILFAST;
1423 mbio->bi_private = r10_bio;
1424
1425 if (conf->mddev->gendisk)
1426 trace_block_bio_remap(bdev_get_queue(mbio->bi_bdev),
1427 mbio, disk_devt(conf->mddev->gendisk),
1428 r10_bio->sector);
1429 /* flush_pending_writes() needs access to the rdev so...*/
1430 mbio->bi_bdev = (void*)rdev;
1431
1432 atomic_inc(&r10_bio->remaining);
1433
1434 cb = blk_check_plugged(raid10_unplug, mddev,
1435 sizeof(*plug));
1436 if (cb)
1437 plug = container_of(cb, struct raid10_plug_cb,
1438 cb);
1439 else
1440 plug = NULL;
1441 spin_lock_irqsave(&conf->device_lock, flags);
1442 if (plug) {
1443 bio_list_add(&plug->pending, mbio);
1444 plug->pending_cnt++;
1445 } else {
1446 bio_list_add(&conf->pending_bio_list, mbio);
1447 conf->pending_count++;
1448 }
1449 spin_unlock_irqrestore(&conf->device_lock, flags);
1450 if (!plug)
1451 md_wakeup_thread(mddev->thread);
1452 }
1453
1454 if (r10_bio->devs[i].repl_bio) {
1455 struct md_rdev *rdev = conf->mirrors[d].replacement;
1456 if (rdev == NULL) {
1457 /* Replacement just got moved to main 'rdev' */
1458 smp_mb();
1459 rdev = conf->mirrors[d].rdev;
1460 }
1461 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1462 bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1463 max_sectors);
1464 r10_bio->devs[i].repl_bio = mbio;
1465
1466 mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr +
1467 choose_data_offset(r10_bio, rdev));
1468 mbio->bi_bdev = rdev->bdev;
1469 mbio->bi_end_io = raid10_end_write_request;
1470 bio_set_op_attrs(mbio, op, do_sync | do_fua);
1471 mbio->bi_private = r10_bio;
1472
1473 if (conf->mddev->gendisk)
1474 trace_block_bio_remap(bdev_get_queue(mbio->bi_bdev),
1475 mbio, disk_devt(conf->mddev->gendisk),
1476 r10_bio->sector);
1477 /* flush_pending_writes() needs access to the rdev so...*/
1478 mbio->bi_bdev = (void*)rdev;
1479
1480 atomic_inc(&r10_bio->remaining);
1481 spin_lock_irqsave(&conf->device_lock, flags);
1482 bio_list_add(&conf->pending_bio_list, mbio);
1483 conf->pending_count++;
1484 spin_unlock_irqrestore(&conf->device_lock, flags);
1485 if (!mddev_check_plugged(mddev))
1486 md_wakeup_thread(mddev->thread);
1487 }
1488 }
1489
1490 /* Don't remove the bias on 'remaining' (one_write_done) until
1491 * after checking if we need to go around again.
1492 */
1493
1494 if (sectors_handled < bio_sectors(bio)) {
1495 one_write_done(r10_bio);
1496 /* We need another r10_bio. It has already been counted
1497 * in bio->bi_phys_segments.
1498 */
1499 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1500
1501 r10_bio->master_bio = bio;
1502 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1503
1504 r10_bio->mddev = mddev;
1505 r10_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1506 r10_bio->state = 0;
1507 goto retry_write;
1508 }
1509 one_write_done(r10_bio);
1510}
1511
1512static void __make_request(struct mddev *mddev, struct bio *bio)
1513{
1514 struct r10conf *conf = mddev->private;
1515 struct r10bio *r10_bio;
1516
1517 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1518
1519 r10_bio->master_bio = bio;
1520 r10_bio->sectors = bio_sectors(bio);
1521
1522 r10_bio->mddev = mddev;
1523 r10_bio->sector = bio->bi_iter.bi_sector;
1524 r10_bio->state = 0;
1525
1526 /*
1527 * We might need to issue multiple reads to different devices if there
1528 * are bad blocks around, so we keep track of the number of reads in
1529 * bio->bi_phys_segments. If this is 0, there is only one r10_bio and
1530 * no locking will be needed when the request completes. If it is
1531 * non-zero, then it is the number of not-completed requests.
1532 */
1533 bio->bi_phys_segments = 0;
1534 bio_clear_flag(bio, BIO_SEG_VALID);
1535
1536 if (bio_data_dir(bio) == READ)
1537 raid10_read_request(mddev, bio, r10_bio);
1538 else
1539 raid10_write_request(mddev, bio, r10_bio);
1540}
1541
1542static void raid10_make_request(struct mddev *mddev, struct bio *bio)
1543{
1544 struct r10conf *conf = mddev->private;
1545 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1546 int chunk_sects = chunk_mask + 1;
1547
1548 struct bio *split;
1549
1550 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1551 md_flush_request(mddev, bio);
1552 return;
1553 }
1554
1555 do {
1556
1557 /*
1558 * If this request crosses a chunk boundary, we need to split
1559 * it.
1560 */
1561 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1562 bio_sectors(bio) > chunk_sects
1563 && (conf->geo.near_copies < conf->geo.raid_disks
1564 || conf->prev.near_copies <
1565 conf->prev.raid_disks))) {
1566 split = bio_split(bio, chunk_sects -
1567 (bio->bi_iter.bi_sector &
1568 (chunk_sects - 1)),
1569 GFP_NOIO, fs_bio_set);
1570 bio_chain(split, bio);
1571 } else {
1572 split = bio;
1573 }
1574
1575 /*
1576 * If a bio is splitted, the first part of bio will pass
1577 * barrier but the bio is queued in current->bio_list (see
1578 * generic_make_request). If there is a raise_barrier() called
1579 * here, the second part of bio can't pass barrier. But since
1580 * the first part bio isn't dispatched to underlaying disks
1581 * yet, the barrier is never released, hence raise_barrier will
1582 * alays wait. We have a deadlock.
1583 * Note, this only happens in read path. For write path, the
1584 * first part of bio is dispatched in a schedule() call
1585 * (because of blk plug) or offloaded to raid10d.
1586 * Quitting from the function immediately can change the bio
1587 * order queued in bio_list and avoid the deadlock.
1588 */
1589 __make_request(mddev, split);
1590 if (split != bio && bio_data_dir(bio) == READ) {
1591 generic_make_request(bio);
1592 break;
1593 }
1594 } while (split != bio);
1595
1596 /* In case raid10d snuck in to freeze_array */
1597 wake_up(&conf->wait_barrier);
1598}
1599
1600static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1601{
1602 struct r10conf *conf = mddev->private;
1603 int i;
1604
1605 if (conf->geo.near_copies < conf->geo.raid_disks)
1606 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1607 if (conf->geo.near_copies > 1)
1608 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1609 if (conf->geo.far_copies > 1) {
1610 if (conf->geo.far_offset)
1611 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1612 else
1613 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1614 if (conf->geo.far_set_size != conf->geo.raid_disks)
1615 seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1616 }
1617 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1618 conf->geo.raid_disks - mddev->degraded);
1619 rcu_read_lock();
1620 for (i = 0; i < conf->geo.raid_disks; i++) {
1621 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1622 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1623 }
1624 rcu_read_unlock();
1625 seq_printf(seq, "]");
1626}
1627
1628/* check if there are enough drives for
1629 * every block to appear on atleast one.
1630 * Don't consider the device numbered 'ignore'
1631 * as we might be about to remove it.
1632 */
1633static int _enough(struct r10conf *conf, int previous, int ignore)
1634{
1635 int first = 0;
1636 int has_enough = 0;
1637 int disks, ncopies;
1638 if (previous) {
1639 disks = conf->prev.raid_disks;
1640 ncopies = conf->prev.near_copies;
1641 } else {
1642 disks = conf->geo.raid_disks;
1643 ncopies = conf->geo.near_copies;
1644 }
1645
1646 rcu_read_lock();
1647 do {
1648 int n = conf->copies;
1649 int cnt = 0;
1650 int this = first;
1651 while (n--) {
1652 struct md_rdev *rdev;
1653 if (this != ignore &&
1654 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1655 test_bit(In_sync, &rdev->flags))
1656 cnt++;
1657 this = (this+1) % disks;
1658 }
1659 if (cnt == 0)
1660 goto out;
1661 first = (first + ncopies) % disks;
1662 } while (first != 0);
1663 has_enough = 1;
1664out:
1665 rcu_read_unlock();
1666 return has_enough;
1667}
1668
1669static int enough(struct r10conf *conf, int ignore)
1670{
1671 /* when calling 'enough', both 'prev' and 'geo' must
1672 * be stable.
1673 * This is ensured if ->reconfig_mutex or ->device_lock
1674 * is held.
1675 */
1676 return _enough(conf, 0, ignore) &&
1677 _enough(conf, 1, ignore);
1678}
1679
1680static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1681{
1682 char b[BDEVNAME_SIZE];
1683 struct r10conf *conf = mddev->private;
1684 unsigned long flags;
1685
1686 /*
1687 * If it is not operational, then we have already marked it as dead
1688 * else if it is the last working disks, ignore the error, let the
1689 * next level up know.
1690 * else mark the drive as failed
1691 */
1692 spin_lock_irqsave(&conf->device_lock, flags);
1693 if (test_bit(In_sync, &rdev->flags)
1694 && !enough(conf, rdev->raid_disk)) {
1695 /*
1696 * Don't fail the drive, just return an IO error.
1697 */
1698 spin_unlock_irqrestore(&conf->device_lock, flags);
1699 return;
1700 }
1701 if (test_and_clear_bit(In_sync, &rdev->flags))
1702 mddev->degraded++;
1703 /*
1704 * If recovery is running, make sure it aborts.
1705 */
1706 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1707 set_bit(Blocked, &rdev->flags);
1708 set_bit(Faulty, &rdev->flags);
1709 set_mask_bits(&mddev->sb_flags, 0,
1710 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1711 spin_unlock_irqrestore(&conf->device_lock, flags);
1712 pr_crit("md/raid10:%s: Disk failure on %s, disabling device.\n"
1713 "md/raid10:%s: Operation continuing on %d devices.\n",
1714 mdname(mddev), bdevname(rdev->bdev, b),
1715 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1716}
1717
1718static void print_conf(struct r10conf *conf)
1719{
1720 int i;
1721 struct md_rdev *rdev;
1722
1723 pr_debug("RAID10 conf printout:\n");
1724 if (!conf) {
1725 pr_debug("(!conf)\n");
1726 return;
1727 }
1728 pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1729 conf->geo.raid_disks);
1730
1731 /* This is only called with ->reconfix_mutex held, so
1732 * rcu protection of rdev is not needed */
1733 for (i = 0; i < conf->geo.raid_disks; i++) {
1734 char b[BDEVNAME_SIZE];
1735 rdev = conf->mirrors[i].rdev;
1736 if (rdev)
1737 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1738 i, !test_bit(In_sync, &rdev->flags),
1739 !test_bit(Faulty, &rdev->flags),
1740 bdevname(rdev->bdev,b));
1741 }
1742}
1743
1744static void close_sync(struct r10conf *conf)
1745{
1746 wait_barrier(conf);
1747 allow_barrier(conf);
1748
1749 mempool_destroy(conf->r10buf_pool);
1750 conf->r10buf_pool = NULL;
1751}
1752
1753static int raid10_spare_active(struct mddev *mddev)
1754{
1755 int i;
1756 struct r10conf *conf = mddev->private;
1757 struct raid10_info *tmp;
1758 int count = 0;
1759 unsigned long flags;
1760
1761 /*
1762 * Find all non-in_sync disks within the RAID10 configuration
1763 * and mark them in_sync
1764 */
1765 for (i = 0; i < conf->geo.raid_disks; i++) {
1766 tmp = conf->mirrors + i;
1767 if (tmp->replacement
1768 && tmp->replacement->recovery_offset == MaxSector
1769 && !test_bit(Faulty, &tmp->replacement->flags)
1770 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1771 /* Replacement has just become active */
1772 if (!tmp->rdev
1773 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1774 count++;
1775 if (tmp->rdev) {
1776 /* Replaced device not technically faulty,
1777 * but we need to be sure it gets removed
1778 * and never re-added.
1779 */
1780 set_bit(Faulty, &tmp->rdev->flags);
1781 sysfs_notify_dirent_safe(
1782 tmp->rdev->sysfs_state);
1783 }
1784 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1785 } else if (tmp->rdev
1786 && tmp->rdev->recovery_offset == MaxSector
1787 && !test_bit(Faulty, &tmp->rdev->flags)
1788 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1789 count++;
1790 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1791 }
1792 }
1793 spin_lock_irqsave(&conf->device_lock, flags);
1794 mddev->degraded -= count;
1795 spin_unlock_irqrestore(&conf->device_lock, flags);
1796
1797 print_conf(conf);
1798 return count;
1799}
1800
1801static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1802{
1803 struct r10conf *conf = mddev->private;
1804 int err = -EEXIST;
1805 int mirror;
1806 int first = 0;
1807 int last = conf->geo.raid_disks - 1;
1808
1809 if (mddev->recovery_cp < MaxSector)
1810 /* only hot-add to in-sync arrays, as recovery is
1811 * very different from resync
1812 */
1813 return -EBUSY;
1814 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1815 return -EINVAL;
1816
1817 if (md_integrity_add_rdev(rdev, mddev))
1818 return -ENXIO;
1819
1820 if (rdev->raid_disk >= 0)
1821 first = last = rdev->raid_disk;
1822
1823 if (rdev->saved_raid_disk >= first &&
1824 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1825 mirror = rdev->saved_raid_disk;
1826 else
1827 mirror = first;
1828 for ( ; mirror <= last ; mirror++) {
1829 struct raid10_info *p = &conf->mirrors[mirror];
1830 if (p->recovery_disabled == mddev->recovery_disabled)
1831 continue;
1832 if (p->rdev) {
1833 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1834 p->replacement != NULL)
1835 continue;
1836 clear_bit(In_sync, &rdev->flags);
1837 set_bit(Replacement, &rdev->flags);
1838 rdev->raid_disk = mirror;
1839 err = 0;
1840 if (mddev->gendisk)
1841 disk_stack_limits(mddev->gendisk, rdev->bdev,
1842 rdev->data_offset << 9);
1843 conf->fullsync = 1;
1844 rcu_assign_pointer(p->replacement, rdev);
1845 break;
1846 }
1847
1848 if (mddev->gendisk)
1849 disk_stack_limits(mddev->gendisk, rdev->bdev,
1850 rdev->data_offset << 9);
1851
1852 p->head_position = 0;
1853 p->recovery_disabled = mddev->recovery_disabled - 1;
1854 rdev->raid_disk = mirror;
1855 err = 0;
1856 if (rdev->saved_raid_disk != mirror)
1857 conf->fullsync = 1;
1858 rcu_assign_pointer(p->rdev, rdev);
1859 break;
1860 }
1861 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1862 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1863
1864 print_conf(conf);
1865 return err;
1866}
1867
1868static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1869{
1870 struct r10conf *conf = mddev->private;
1871 int err = 0;
1872 int number = rdev->raid_disk;
1873 struct md_rdev **rdevp;
1874 struct raid10_info *p = conf->mirrors + number;
1875
1876 print_conf(conf);
1877 if (rdev == p->rdev)
1878 rdevp = &p->rdev;
1879 else if (rdev == p->replacement)
1880 rdevp = &p->replacement;
1881 else
1882 return 0;
1883
1884 if (test_bit(In_sync, &rdev->flags) ||
1885 atomic_read(&rdev->nr_pending)) {
1886 err = -EBUSY;
1887 goto abort;
1888 }
1889 /* Only remove non-faulty devices if recovery
1890 * is not possible.
1891 */
1892 if (!test_bit(Faulty, &rdev->flags) &&
1893 mddev->recovery_disabled != p->recovery_disabled &&
1894 (!p->replacement || p->replacement == rdev) &&
1895 number < conf->geo.raid_disks &&
1896 enough(conf, -1)) {
1897 err = -EBUSY;
1898 goto abort;
1899 }
1900 *rdevp = NULL;
1901 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1902 synchronize_rcu();
1903 if (atomic_read(&rdev->nr_pending)) {
1904 /* lost the race, try later */
1905 err = -EBUSY;
1906 *rdevp = rdev;
1907 goto abort;
1908 }
1909 }
1910 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
1933static void end_sync_read(struct bio *bio)
1934{
1935 struct r10bio *r10_bio = bio->bi_private;
1936 struct r10conf *conf = r10_bio->mddev->private;
1937 int d;
1938
1939 if (bio == r10_bio->master_bio) {
1940 /* this is a reshape read */
1941 d = r10_bio->read_slot; /* really the read dev */
1942 } else
1943 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1944
1945 if (!bio->bi_error)
1946 set_bit(R10BIO_Uptodate, &r10_bio->state);
1947 else
1948 /* The write handler will notice the lack of
1949 * R10BIO_Uptodate and record any errors etc
1950 */
1951 atomic_add(r10_bio->sectors,
1952 &conf->mirrors[d].rdev->corrected_errors);
1953
1954 /* for reconstruct, we always reschedule after a read.
1955 * for resync, only after all reads
1956 */
1957 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1958 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1959 atomic_dec_and_test(&r10_bio->remaining)) {
1960 /* we have read all the blocks,
1961 * do the comparison in process context in raid10d
1962 */
1963 reschedule_retry(r10_bio);
1964 }
1965}
1966
1967static void end_sync_request(struct r10bio *r10_bio)
1968{
1969 struct mddev *mddev = r10_bio->mddev;
1970
1971 while (atomic_dec_and_test(&r10_bio->remaining)) {
1972 if (r10_bio->master_bio == NULL) {
1973 /* the primary of several recovery bios */
1974 sector_t s = r10_bio->sectors;
1975 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1976 test_bit(R10BIO_WriteError, &r10_bio->state))
1977 reschedule_retry(r10_bio);
1978 else
1979 put_buf(r10_bio);
1980 md_done_sync(mddev, s, 1);
1981 break;
1982 } else {
1983 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1984 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1985 test_bit(R10BIO_WriteError, &r10_bio->state))
1986 reschedule_retry(r10_bio);
1987 else
1988 put_buf(r10_bio);
1989 r10_bio = r10_bio2;
1990 }
1991 }
1992}
1993
1994static void end_sync_write(struct bio *bio)
1995{
1996 struct r10bio *r10_bio = bio->bi_private;
1997 struct mddev *mddev = r10_bio->mddev;
1998 struct r10conf *conf = mddev->private;
1999 int d;
2000 sector_t first_bad;
2001 int bad_sectors;
2002 int slot;
2003 int repl;
2004 struct md_rdev *rdev = NULL;
2005
2006 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
2007 if (repl)
2008 rdev = conf->mirrors[d].replacement;
2009 else
2010 rdev = conf->mirrors[d].rdev;
2011
2012 if (bio->bi_error) {
2013 if (repl)
2014 md_error(mddev, rdev);
2015 else {
2016 set_bit(WriteErrorSeen, &rdev->flags);
2017 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2018 set_bit(MD_RECOVERY_NEEDED,
2019 &rdev->mddev->recovery);
2020 set_bit(R10BIO_WriteError, &r10_bio->state);
2021 }
2022 } else if (is_badblock(rdev,
2023 r10_bio->devs[slot].addr,
2024 r10_bio->sectors,
2025 &first_bad, &bad_sectors))
2026 set_bit(R10BIO_MadeGood, &r10_bio->state);
2027
2028 rdev_dec_pending(rdev, mddev);
2029
2030 end_sync_request(r10_bio);
2031}
2032
2033/*
2034 * Note: sync and recover and handled very differently for raid10
2035 * This code is for resync.
2036 * For resync, we read through virtual addresses and read all blocks.
2037 * If there is any error, we schedule a write. The lowest numbered
2038 * drive is authoritative.
2039 * However requests come for physical address, so we need to map.
2040 * For every physical address there are raid_disks/copies virtual addresses,
2041 * which is always are least one, but is not necessarly an integer.
2042 * This means that a physical address can span multiple chunks, so we may
2043 * have to submit multiple io requests for a single sync request.
2044 */
2045/*
2046 * We check if all blocks are in-sync and only write to blocks that
2047 * aren't in sync
2048 */
2049static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2050{
2051 struct r10conf *conf = mddev->private;
2052 int i, first;
2053 struct bio *tbio, *fbio;
2054 int vcnt;
2055
2056 atomic_set(&r10_bio->remaining, 1);
2057
2058 /* find the first device with a block */
2059 for (i=0; i<conf->copies; i++)
2060 if (!r10_bio->devs[i].bio->bi_error)
2061 break;
2062
2063 if (i == conf->copies)
2064 goto done;
2065
2066 first = i;
2067 fbio = r10_bio->devs[i].bio;
2068 fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2069 fbio->bi_iter.bi_idx = 0;
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 struct md_rdev *rdev;
2076
2077 tbio = r10_bio->devs[i].bio;
2078
2079 if (tbio->bi_end_io != end_sync_read)
2080 continue;
2081 if (i == first)
2082 continue;
2083 d = r10_bio->devs[i].devnum;
2084 rdev = conf->mirrors[d].rdev;
2085 if (!r10_bio->devs[i].bio->bi_error) {
2086 /* We know that the bi_io_vec layout is the same for
2087 * both 'first' and 'i', so we just compare them.
2088 * All vec entries are PAGE_SIZE;
2089 */
2090 int sectors = r10_bio->sectors;
2091 for (j = 0; j < vcnt; j++) {
2092 int len = PAGE_SIZE;
2093 if (sectors < (len / 512))
2094 len = sectors * 512;
2095 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
2096 page_address(tbio->bi_io_vec[j].bv_page),
2097 len))
2098 break;
2099 sectors -= len/512;
2100 }
2101 if (j == vcnt)
2102 continue;
2103 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2104 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2105 /* Don't fix anything. */
2106 continue;
2107 } else if (test_bit(FailFast, &rdev->flags)) {
2108 /* Just give up on this device */
2109 md_error(rdev->mddev, rdev);
2110 continue;
2111 }
2112 /* Ok, we need to write this bio, either to correct an
2113 * inconsistency or to correct an unreadable block.
2114 * First we need to fixup bv_offset, bv_len and
2115 * bi_vecs, as the read request might have corrupted these
2116 */
2117 bio_reset(tbio);
2118
2119 tbio->bi_vcnt = vcnt;
2120 tbio->bi_iter.bi_size = fbio->bi_iter.bi_size;
2121 tbio->bi_private = r10_bio;
2122 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2123 tbio->bi_end_io = end_sync_write;
2124 bio_set_op_attrs(tbio, REQ_OP_WRITE, 0);
2125
2126 bio_copy_data(tbio, fbio);
2127
2128 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2129 atomic_inc(&r10_bio->remaining);
2130 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2131
2132 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2133 tbio->bi_opf |= MD_FAILFAST;
2134 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2135 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2136 generic_make_request(tbio);
2137 }
2138
2139 /* Now write out to any replacement devices
2140 * that are active
2141 */
2142 for (i = 0; i < conf->copies; i++) {
2143 int d;
2144
2145 tbio = r10_bio->devs[i].repl_bio;
2146 if (!tbio || !tbio->bi_end_io)
2147 continue;
2148 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2149 && r10_bio->devs[i].bio != fbio)
2150 bio_copy_data(tbio, fbio);
2151 d = r10_bio->devs[i].devnum;
2152 atomic_inc(&r10_bio->remaining);
2153 md_sync_acct(conf->mirrors[d].replacement->bdev,
2154 bio_sectors(tbio));
2155 generic_make_request(tbio);
2156 }
2157
2158done:
2159 if (atomic_dec_and_test(&r10_bio->remaining)) {
2160 md_done_sync(mddev, r10_bio->sectors, 1);
2161 put_buf(r10_bio);
2162 }
2163}
2164
2165/*
2166 * Now for the recovery code.
2167 * Recovery happens across physical sectors.
2168 * We recover all non-is_sync drives by finding the virtual address of
2169 * each, and then choose a working drive that also has that virt address.
2170 * There is a separate r10_bio for each non-in_sync drive.
2171 * Only the first two slots are in use. The first for reading,
2172 * The second for writing.
2173 *
2174 */
2175static void fix_recovery_read_error(struct r10bio *r10_bio)
2176{
2177 /* We got a read error during recovery.
2178 * We repeat the read in smaller page-sized sections.
2179 * If a read succeeds, write it to the new device or record
2180 * a bad block if we cannot.
2181 * If a read fails, record a bad block on both old and
2182 * new devices.
2183 */
2184 struct mddev *mddev = r10_bio->mddev;
2185 struct r10conf *conf = mddev->private;
2186 struct bio *bio = r10_bio->devs[0].bio;
2187 sector_t sect = 0;
2188 int sectors = r10_bio->sectors;
2189 int idx = 0;
2190 int dr = r10_bio->devs[0].devnum;
2191 int dw = r10_bio->devs[1].devnum;
2192
2193 while (sectors) {
2194 int s = sectors;
2195 struct md_rdev *rdev;
2196 sector_t addr;
2197 int ok;
2198
2199 if (s > (PAGE_SIZE>>9))
2200 s = PAGE_SIZE >> 9;
2201
2202 rdev = conf->mirrors[dr].rdev;
2203 addr = r10_bio->devs[0].addr + sect,
2204 ok = sync_page_io(rdev,
2205 addr,
2206 s << 9,
2207 bio->bi_io_vec[idx].bv_page,
2208 REQ_OP_READ, 0, false);
2209 if (ok) {
2210 rdev = conf->mirrors[dw].rdev;
2211 addr = r10_bio->devs[1].addr + sect;
2212 ok = sync_page_io(rdev,
2213 addr,
2214 s << 9,
2215 bio->bi_io_vec[idx].bv_page,
2216 REQ_OP_WRITE, 0, false);
2217 if (!ok) {
2218 set_bit(WriteErrorSeen, &rdev->flags);
2219 if (!test_and_set_bit(WantReplacement,
2220 &rdev->flags))
2221 set_bit(MD_RECOVERY_NEEDED,
2222 &rdev->mddev->recovery);
2223 }
2224 }
2225 if (!ok) {
2226 /* We don't worry if we cannot set a bad block -
2227 * it really is bad so there is no loss in not
2228 * recording it yet
2229 */
2230 rdev_set_badblocks(rdev, addr, s, 0);
2231
2232 if (rdev != conf->mirrors[dw].rdev) {
2233 /* need bad block on destination too */
2234 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2235 addr = r10_bio->devs[1].addr + sect;
2236 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2237 if (!ok) {
2238 /* just abort the recovery */
2239 pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2240 mdname(mddev));
2241
2242 conf->mirrors[dw].recovery_disabled
2243 = mddev->recovery_disabled;
2244 set_bit(MD_RECOVERY_INTR,
2245 &mddev->recovery);
2246 break;
2247 }
2248 }
2249 }
2250
2251 sectors -= s;
2252 sect += s;
2253 idx++;
2254 }
2255}
2256
2257static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2258{
2259 struct r10conf *conf = mddev->private;
2260 int d;
2261 struct bio *wbio, *wbio2;
2262
2263 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2264 fix_recovery_read_error(r10_bio);
2265 end_sync_request(r10_bio);
2266 return;
2267 }
2268
2269 /*
2270 * share the pages with the first bio
2271 * and submit the write request
2272 */
2273 d = r10_bio->devs[1].devnum;
2274 wbio = r10_bio->devs[1].bio;
2275 wbio2 = r10_bio->devs[1].repl_bio;
2276 /* Need to test wbio2->bi_end_io before we call
2277 * generic_make_request as if the former is NULL,
2278 * the latter is free to free wbio2.
2279 */
2280 if (wbio2 && !wbio2->bi_end_io)
2281 wbio2 = NULL;
2282 if (wbio->bi_end_io) {
2283 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2284 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2285 generic_make_request(wbio);
2286 }
2287 if (wbio2) {
2288 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2289 md_sync_acct(conf->mirrors[d].replacement->bdev,
2290 bio_sectors(wbio2));
2291 generic_make_request(wbio2);
2292 }
2293}
2294
2295/*
2296 * Used by fix_read_error() to decay the per rdev read_errors.
2297 * We halve the read error count for every hour that has elapsed
2298 * since the last recorded read error.
2299 *
2300 */
2301static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2302{
2303 long cur_time_mon;
2304 unsigned long hours_since_last;
2305 unsigned int read_errors = atomic_read(&rdev->read_errors);
2306
2307 cur_time_mon = ktime_get_seconds();
2308
2309 if (rdev->last_read_error == 0) {
2310 /* first time we've seen a read error */
2311 rdev->last_read_error = cur_time_mon;
2312 return;
2313 }
2314
2315 hours_since_last = (long)(cur_time_mon -
2316 rdev->last_read_error) / 3600;
2317
2318 rdev->last_read_error = cur_time_mon;
2319
2320 /*
2321 * if hours_since_last is > the number of bits in read_errors
2322 * just set read errors to 0. We do this to avoid
2323 * overflowing the shift of read_errors by hours_since_last.
2324 */
2325 if (hours_since_last >= 8 * sizeof(read_errors))
2326 atomic_set(&rdev->read_errors, 0);
2327 else
2328 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2329}
2330
2331static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2332 int sectors, struct page *page, int rw)
2333{
2334 sector_t first_bad;
2335 int bad_sectors;
2336
2337 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2338 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2339 return -1;
2340 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
2341 /* success */
2342 return 1;
2343 if (rw == WRITE) {
2344 set_bit(WriteErrorSeen, &rdev->flags);
2345 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2346 set_bit(MD_RECOVERY_NEEDED,
2347 &rdev->mddev->recovery);
2348 }
2349 /* need to record an error - either for the block or the device */
2350 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2351 md_error(rdev->mddev, rdev);
2352 return 0;
2353}
2354
2355/*
2356 * This is a kernel thread which:
2357 *
2358 * 1. Retries failed read operations on working mirrors.
2359 * 2. Updates the raid superblock when problems encounter.
2360 * 3. Performs writes following reads for array synchronising.
2361 */
2362
2363static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2364{
2365 int sect = 0; /* Offset from r10_bio->sector */
2366 int sectors = r10_bio->sectors;
2367 struct md_rdev*rdev;
2368 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2369 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2370
2371 /* still own a reference to this rdev, so it cannot
2372 * have been cleared recently.
2373 */
2374 rdev = conf->mirrors[d].rdev;
2375
2376 if (test_bit(Faulty, &rdev->flags))
2377 /* drive has already been failed, just ignore any
2378 more fix_read_error() attempts */
2379 return;
2380
2381 check_decay_read_errors(mddev, rdev);
2382 atomic_inc(&rdev->read_errors);
2383 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2384 char b[BDEVNAME_SIZE];
2385 bdevname(rdev->bdev, b);
2386
2387 pr_notice("md/raid10:%s: %s: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2388 mdname(mddev), b,
2389 atomic_read(&rdev->read_errors), max_read_errors);
2390 pr_notice("md/raid10:%s: %s: Failing raid device\n",
2391 mdname(mddev), b);
2392 md_error(mddev, rdev);
2393 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2394 return;
2395 }
2396
2397 while(sectors) {
2398 int s = sectors;
2399 int sl = r10_bio->read_slot;
2400 int success = 0;
2401 int start;
2402
2403 if (s > (PAGE_SIZE>>9))
2404 s = PAGE_SIZE >> 9;
2405
2406 rcu_read_lock();
2407 do {
2408 sector_t first_bad;
2409 int bad_sectors;
2410
2411 d = r10_bio->devs[sl].devnum;
2412 rdev = rcu_dereference(conf->mirrors[d].rdev);
2413 if (rdev &&
2414 test_bit(In_sync, &rdev->flags) &&
2415 !test_bit(Faulty, &rdev->flags) &&
2416 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2417 &first_bad, &bad_sectors) == 0) {
2418 atomic_inc(&rdev->nr_pending);
2419 rcu_read_unlock();
2420 success = sync_page_io(rdev,
2421 r10_bio->devs[sl].addr +
2422 sect,
2423 s<<9,
2424 conf->tmppage,
2425 REQ_OP_READ, 0, false);
2426 rdev_dec_pending(rdev, mddev);
2427 rcu_read_lock();
2428 if (success)
2429 break;
2430 }
2431 sl++;
2432 if (sl == conf->copies)
2433 sl = 0;
2434 } while (!success && sl != r10_bio->read_slot);
2435 rcu_read_unlock();
2436
2437 if (!success) {
2438 /* Cannot read from anywhere, just mark the block
2439 * as bad on the first device to discourage future
2440 * reads.
2441 */
2442 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2443 rdev = conf->mirrors[dn].rdev;
2444
2445 if (!rdev_set_badblocks(
2446 rdev,
2447 r10_bio->devs[r10_bio->read_slot].addr
2448 + sect,
2449 s, 0)) {
2450 md_error(mddev, rdev);
2451 r10_bio->devs[r10_bio->read_slot].bio
2452 = IO_BLOCKED;
2453 }
2454 break;
2455 }
2456
2457 start = sl;
2458 /* write it back and re-read */
2459 rcu_read_lock();
2460 while (sl != r10_bio->read_slot) {
2461 char b[BDEVNAME_SIZE];
2462
2463 if (sl==0)
2464 sl = conf->copies;
2465 sl--;
2466 d = r10_bio->devs[sl].devnum;
2467 rdev = rcu_dereference(conf->mirrors[d].rdev);
2468 if (!rdev ||
2469 test_bit(Faulty, &rdev->flags) ||
2470 !test_bit(In_sync, &rdev->flags))
2471 continue;
2472
2473 atomic_inc(&rdev->nr_pending);
2474 rcu_read_unlock();
2475 if (r10_sync_page_io(rdev,
2476 r10_bio->devs[sl].addr +
2477 sect,
2478 s, conf->tmppage, WRITE)
2479 == 0) {
2480 /* Well, this device is dead */
2481 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %s)\n",
2482 mdname(mddev), s,
2483 (unsigned long long)(
2484 sect +
2485 choose_data_offset(r10_bio,
2486 rdev)),
2487 bdevname(rdev->bdev, b));
2488 pr_notice("md/raid10:%s: %s: failing drive\n",
2489 mdname(mddev),
2490 bdevname(rdev->bdev, b));
2491 }
2492 rdev_dec_pending(rdev, mddev);
2493 rcu_read_lock();
2494 }
2495 sl = start;
2496 while (sl != r10_bio->read_slot) {
2497 char b[BDEVNAME_SIZE];
2498
2499 if (sl==0)
2500 sl = conf->copies;
2501 sl--;
2502 d = r10_bio->devs[sl].devnum;
2503 rdev = rcu_dereference(conf->mirrors[d].rdev);
2504 if (!rdev ||
2505 test_bit(Faulty, &rdev->flags) ||
2506 !test_bit(In_sync, &rdev->flags))
2507 continue;
2508
2509 atomic_inc(&rdev->nr_pending);
2510 rcu_read_unlock();
2511 switch (r10_sync_page_io(rdev,
2512 r10_bio->devs[sl].addr +
2513 sect,
2514 s, conf->tmppage,
2515 READ)) {
2516 case 0:
2517 /* Well, this device is dead */
2518 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %s)\n",
2519 mdname(mddev), s,
2520 (unsigned long long)(
2521 sect +
2522 choose_data_offset(r10_bio, rdev)),
2523 bdevname(rdev->bdev, b));
2524 pr_notice("md/raid10:%s: %s: failing drive\n",
2525 mdname(mddev),
2526 bdevname(rdev->bdev, b));
2527 break;
2528 case 1:
2529 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %s)\n",
2530 mdname(mddev), s,
2531 (unsigned long long)(
2532 sect +
2533 choose_data_offset(r10_bio, rdev)),
2534 bdevname(rdev->bdev, b));
2535 atomic_add(s, &rdev->corrected_errors);
2536 }
2537
2538 rdev_dec_pending(rdev, mddev);
2539 rcu_read_lock();
2540 }
2541 rcu_read_unlock();
2542
2543 sectors -= s;
2544 sect += s;
2545 }
2546}
2547
2548static int narrow_write_error(struct r10bio *r10_bio, int i)
2549{
2550 struct bio *bio = r10_bio->master_bio;
2551 struct mddev *mddev = r10_bio->mddev;
2552 struct r10conf *conf = mddev->private;
2553 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2554 /* bio has the data to be written to slot 'i' where
2555 * we just recently had a write error.
2556 * We repeatedly clone the bio and trim down to one block,
2557 * then try the write. Where the write fails we record
2558 * a bad block.
2559 * It is conceivable that the bio doesn't exactly align with
2560 * blocks. We must handle this.
2561 *
2562 * We currently own a reference to the rdev.
2563 */
2564
2565 int block_sectors;
2566 sector_t sector;
2567 int sectors;
2568 int sect_to_write = r10_bio->sectors;
2569 int ok = 1;
2570
2571 if (rdev->badblocks.shift < 0)
2572 return 0;
2573
2574 block_sectors = roundup(1 << rdev->badblocks.shift,
2575 bdev_logical_block_size(rdev->bdev) >> 9);
2576 sector = r10_bio->sector;
2577 sectors = ((r10_bio->sector + block_sectors)
2578 & ~(sector_t)(block_sectors - 1))
2579 - sector;
2580
2581 while (sect_to_write) {
2582 struct bio *wbio;
2583 sector_t wsector;
2584 if (sectors > sect_to_write)
2585 sectors = sect_to_write;
2586 /* Write at 'sector' for 'sectors' */
2587 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2588 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2589 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2590 wbio->bi_iter.bi_sector = wsector +
2591 choose_data_offset(r10_bio, rdev);
2592 wbio->bi_bdev = rdev->bdev;
2593 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2594
2595 if (submit_bio_wait(wbio) < 0)
2596 /* Failure! */
2597 ok = rdev_set_badblocks(rdev, wsector,
2598 sectors, 0)
2599 && ok;
2600
2601 bio_put(wbio);
2602 sect_to_write -= sectors;
2603 sector += sectors;
2604 sectors = block_sectors;
2605 }
2606 return ok;
2607}
2608
2609static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2610{
2611 int slot = r10_bio->read_slot;
2612 struct bio *bio;
2613 struct r10conf *conf = mddev->private;
2614 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2615 char b[BDEVNAME_SIZE];
2616 unsigned long do_sync;
2617 int max_sectors;
2618 dev_t bio_dev;
2619 sector_t bio_last_sector;
2620
2621 /* we got a read error. Maybe the drive is bad. Maybe just
2622 * the block and we can fix it.
2623 * We freeze all other IO, and try reading the block from
2624 * other devices. When we find one, we re-write
2625 * and check it that fixes the read error.
2626 * This is all done synchronously while the array is
2627 * frozen.
2628 */
2629 bio = r10_bio->devs[slot].bio;
2630 bdevname(bio->bi_bdev, b);
2631 bio_dev = bio->bi_bdev->bd_dev;
2632 bio_last_sector = r10_bio->devs[slot].addr + rdev->data_offset + r10_bio->sectors;
2633 bio_put(bio);
2634 r10_bio->devs[slot].bio = NULL;
2635
2636 if (mddev->ro)
2637 r10_bio->devs[slot].bio = IO_BLOCKED;
2638 else if (!test_bit(FailFast, &rdev->flags)) {
2639 freeze_array(conf, 1);
2640 fix_read_error(conf, mddev, r10_bio);
2641 unfreeze_array(conf);
2642 } else
2643 md_error(mddev, rdev);
2644
2645 rdev_dec_pending(rdev, mddev);
2646
2647read_more:
2648 rdev = read_balance(conf, r10_bio, &max_sectors);
2649 if (rdev == NULL) {
2650 pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
2651 mdname(mddev), b,
2652 (unsigned long long)r10_bio->sector);
2653 raid_end_bio_io(r10_bio);
2654 return;
2655 }
2656
2657 do_sync = (r10_bio->master_bio->bi_opf & REQ_SYNC);
2658 slot = r10_bio->read_slot;
2659 pr_err_ratelimited("md/raid10:%s: %s: redirecting sector %llu to another mirror\n",
2660 mdname(mddev),
2661 bdevname(rdev->bdev, b),
2662 (unsigned long long)r10_bio->sector);
2663 bio = bio_clone_mddev(r10_bio->master_bio,
2664 GFP_NOIO, mddev);
2665 bio_trim(bio, r10_bio->sector - bio->bi_iter.bi_sector, max_sectors);
2666 r10_bio->devs[slot].bio = bio;
2667 r10_bio->devs[slot].rdev = rdev;
2668 bio->bi_iter.bi_sector = r10_bio->devs[slot].addr
2669 + choose_data_offset(r10_bio, rdev);
2670 bio->bi_bdev = rdev->bdev;
2671 bio_set_op_attrs(bio, REQ_OP_READ, do_sync);
2672 if (test_bit(FailFast, &rdev->flags) &&
2673 test_bit(R10BIO_FailFast, &r10_bio->state))
2674 bio->bi_opf |= MD_FAILFAST;
2675 bio->bi_private = r10_bio;
2676 bio->bi_end_io = raid10_end_read_request;
2677 trace_block_bio_remap(bdev_get_queue(bio->bi_bdev),
2678 bio, bio_dev,
2679 bio_last_sector - r10_bio->sectors);
2680
2681 if (max_sectors < r10_bio->sectors) {
2682 /* Drat - have to split this up more */
2683 struct bio *mbio = r10_bio->master_bio;
2684 int sectors_handled =
2685 r10_bio->sector + max_sectors
2686 - mbio->bi_iter.bi_sector;
2687 r10_bio->sectors = max_sectors;
2688 spin_lock_irq(&conf->device_lock);
2689 if (mbio->bi_phys_segments == 0)
2690 mbio->bi_phys_segments = 2;
2691 else
2692 mbio->bi_phys_segments++;
2693 spin_unlock_irq(&conf->device_lock);
2694 generic_make_request(bio);
2695
2696 r10_bio = mempool_alloc(conf->r10bio_pool,
2697 GFP_NOIO);
2698 r10_bio->master_bio = mbio;
2699 r10_bio->sectors = bio_sectors(mbio) - sectors_handled;
2700 r10_bio->state = 0;
2701 set_bit(R10BIO_ReadError,
2702 &r10_bio->state);
2703 r10_bio->mddev = mddev;
2704 r10_bio->sector = mbio->bi_iter.bi_sector
2705 + sectors_handled;
2706
2707 goto read_more;
2708 } else
2709 generic_make_request(bio);
2710}
2711
2712static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2713{
2714 /* Some sort of write request has finished and it
2715 * succeeded in writing where we thought there was a
2716 * bad block. So forget the bad block.
2717 * Or possibly if failed and we need to record
2718 * a bad block.
2719 */
2720 int m;
2721 struct md_rdev *rdev;
2722
2723 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2724 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2725 for (m = 0; m < conf->copies; m++) {
2726 int dev = r10_bio->devs[m].devnum;
2727 rdev = conf->mirrors[dev].rdev;
2728 if (r10_bio->devs[m].bio == NULL)
2729 continue;
2730 if (!r10_bio->devs[m].bio->bi_error) {
2731 rdev_clear_badblocks(
2732 rdev,
2733 r10_bio->devs[m].addr,
2734 r10_bio->sectors, 0);
2735 } else {
2736 if (!rdev_set_badblocks(
2737 rdev,
2738 r10_bio->devs[m].addr,
2739 r10_bio->sectors, 0))
2740 md_error(conf->mddev, rdev);
2741 }
2742 rdev = conf->mirrors[dev].replacement;
2743 if (r10_bio->devs[m].repl_bio == NULL)
2744 continue;
2745
2746 if (!r10_bio->devs[m].repl_bio->bi_error) {
2747 rdev_clear_badblocks(
2748 rdev,
2749 r10_bio->devs[m].addr,
2750 r10_bio->sectors, 0);
2751 } else {
2752 if (!rdev_set_badblocks(
2753 rdev,
2754 r10_bio->devs[m].addr,
2755 r10_bio->sectors, 0))
2756 md_error(conf->mddev, rdev);
2757 }
2758 }
2759 put_buf(r10_bio);
2760 } else {
2761 bool fail = false;
2762 for (m = 0; m < conf->copies; m++) {
2763 int dev = r10_bio->devs[m].devnum;
2764 struct bio *bio = r10_bio->devs[m].bio;
2765 rdev = conf->mirrors[dev].rdev;
2766 if (bio == IO_MADE_GOOD) {
2767 rdev_clear_badblocks(
2768 rdev,
2769 r10_bio->devs[m].addr,
2770 r10_bio->sectors, 0);
2771 rdev_dec_pending(rdev, conf->mddev);
2772 } else if (bio != NULL && bio->bi_error) {
2773 fail = true;
2774 if (!narrow_write_error(r10_bio, m)) {
2775 md_error(conf->mddev, rdev);
2776 set_bit(R10BIO_Degraded,
2777 &r10_bio->state);
2778 }
2779 rdev_dec_pending(rdev, conf->mddev);
2780 }
2781 bio = r10_bio->devs[m].repl_bio;
2782 rdev = conf->mirrors[dev].replacement;
2783 if (rdev && bio == IO_MADE_GOOD) {
2784 rdev_clear_badblocks(
2785 rdev,
2786 r10_bio->devs[m].addr,
2787 r10_bio->sectors, 0);
2788 rdev_dec_pending(rdev, conf->mddev);
2789 }
2790 }
2791 if (fail) {
2792 spin_lock_irq(&conf->device_lock);
2793 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2794 conf->nr_queued++;
2795 spin_unlock_irq(&conf->device_lock);
2796 md_wakeup_thread(conf->mddev->thread);
2797 } else {
2798 if (test_bit(R10BIO_WriteError,
2799 &r10_bio->state))
2800 close_write(r10_bio);
2801 raid_end_bio_io(r10_bio);
2802 }
2803 }
2804}
2805
2806static void raid10d(struct md_thread *thread)
2807{
2808 struct mddev *mddev = thread->mddev;
2809 struct r10bio *r10_bio;
2810 unsigned long flags;
2811 struct r10conf *conf = mddev->private;
2812 struct list_head *head = &conf->retry_list;
2813 struct blk_plug plug;
2814
2815 md_check_recovery(mddev);
2816
2817 if (!list_empty_careful(&conf->bio_end_io_list) &&
2818 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2819 LIST_HEAD(tmp);
2820 spin_lock_irqsave(&conf->device_lock, flags);
2821 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2822 while (!list_empty(&conf->bio_end_io_list)) {
2823 list_move(conf->bio_end_io_list.prev, &tmp);
2824 conf->nr_queued--;
2825 }
2826 }
2827 spin_unlock_irqrestore(&conf->device_lock, flags);
2828 while (!list_empty(&tmp)) {
2829 r10_bio = list_first_entry(&tmp, struct r10bio,
2830 retry_list);
2831 list_del(&r10_bio->retry_list);
2832 if (mddev->degraded)
2833 set_bit(R10BIO_Degraded, &r10_bio->state);
2834
2835 if (test_bit(R10BIO_WriteError,
2836 &r10_bio->state))
2837 close_write(r10_bio);
2838 raid_end_bio_io(r10_bio);
2839 }
2840 }
2841
2842 blk_start_plug(&plug);
2843 for (;;) {
2844
2845 flush_pending_writes(conf);
2846
2847 spin_lock_irqsave(&conf->device_lock, flags);
2848 if (list_empty(head)) {
2849 spin_unlock_irqrestore(&conf->device_lock, flags);
2850 break;
2851 }
2852 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2853 list_del(head->prev);
2854 conf->nr_queued--;
2855 spin_unlock_irqrestore(&conf->device_lock, flags);
2856
2857 mddev = r10_bio->mddev;
2858 conf = mddev->private;
2859 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2860 test_bit(R10BIO_WriteError, &r10_bio->state))
2861 handle_write_completed(conf, r10_bio);
2862 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2863 reshape_request_write(mddev, r10_bio);
2864 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2865 sync_request_write(mddev, r10_bio);
2866 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2867 recovery_request_write(mddev, r10_bio);
2868 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2869 handle_read_error(mddev, r10_bio);
2870 else {
2871 /* just a partial read to be scheduled from a
2872 * separate context
2873 */
2874 int slot = r10_bio->read_slot;
2875 generic_make_request(r10_bio->devs[slot].bio);
2876 }
2877
2878 cond_resched();
2879 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2880 md_check_recovery(mddev);
2881 }
2882 blk_finish_plug(&plug);
2883}
2884
2885static int init_resync(struct r10conf *conf)
2886{
2887 int buffs;
2888 int i;
2889
2890 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2891 BUG_ON(conf->r10buf_pool);
2892 conf->have_replacement = 0;
2893 for (i = 0; i < conf->geo.raid_disks; i++)
2894 if (conf->mirrors[i].replacement)
2895 conf->have_replacement = 1;
2896 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2897 if (!conf->r10buf_pool)
2898 return -ENOMEM;
2899 conf->next_resync = 0;
2900 return 0;
2901}
2902
2903/*
2904 * perform a "sync" on one "block"
2905 *
2906 * We need to make sure that no normal I/O request - particularly write
2907 * requests - conflict with active sync requests.
2908 *
2909 * This is achieved by tracking pending requests and a 'barrier' concept
2910 * that can be installed to exclude normal IO requests.
2911 *
2912 * Resync and recovery are handled very differently.
2913 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2914 *
2915 * For resync, we iterate over virtual addresses, read all copies,
2916 * and update if there are differences. If only one copy is live,
2917 * skip it.
2918 * For recovery, we iterate over physical addresses, read a good
2919 * value for each non-in_sync drive, and over-write.
2920 *
2921 * So, for recovery we may have several outstanding complex requests for a
2922 * given address, one for each out-of-sync device. We model this by allocating
2923 * a number of r10_bio structures, one for each out-of-sync device.
2924 * As we setup these structures, we collect all bio's together into a list
2925 * which we then process collectively to add pages, and then process again
2926 * to pass to generic_make_request.
2927 *
2928 * The r10_bio structures are linked using a borrowed master_bio pointer.
2929 * This link is counted in ->remaining. When the r10_bio that points to NULL
2930 * has its remaining count decremented to 0, the whole complex operation
2931 * is complete.
2932 *
2933 */
2934
2935static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2936 int *skipped)
2937{
2938 struct r10conf *conf = mddev->private;
2939 struct r10bio *r10_bio;
2940 struct bio *biolist = NULL, *bio;
2941 sector_t max_sector, nr_sectors;
2942 int i;
2943 int max_sync;
2944 sector_t sync_blocks;
2945 sector_t sectors_skipped = 0;
2946 int chunks_skipped = 0;
2947 sector_t chunk_mask = conf->geo.chunk_mask;
2948
2949 if (!conf->r10buf_pool)
2950 if (init_resync(conf))
2951 return 0;
2952
2953 /*
2954 * Allow skipping a full rebuild for incremental assembly
2955 * of a clean array, like RAID1 does.
2956 */
2957 if (mddev->bitmap == NULL &&
2958 mddev->recovery_cp == MaxSector &&
2959 mddev->reshape_position == MaxSector &&
2960 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2961 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2962 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2963 conf->fullsync == 0) {
2964 *skipped = 1;
2965 return mddev->dev_sectors - sector_nr;
2966 }
2967
2968 skipped:
2969 max_sector = mddev->dev_sectors;
2970 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2971 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2972 max_sector = mddev->resync_max_sectors;
2973 if (sector_nr >= max_sector) {
2974 /* If we aborted, we need to abort the
2975 * sync on the 'current' bitmap chucks (there can
2976 * be several when recovering multiple devices).
2977 * as we may have started syncing it but not finished.
2978 * We can find the current address in
2979 * mddev->curr_resync, but for recovery,
2980 * we need to convert that to several
2981 * virtual addresses.
2982 */
2983 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2984 end_reshape(conf);
2985 close_sync(conf);
2986 return 0;
2987 }
2988
2989 if (mddev->curr_resync < max_sector) { /* aborted */
2990 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2991 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2992 &sync_blocks, 1);
2993 else for (i = 0; i < conf->geo.raid_disks; i++) {
2994 sector_t sect =
2995 raid10_find_virt(conf, mddev->curr_resync, i);
2996 bitmap_end_sync(mddev->bitmap, sect,
2997 &sync_blocks, 1);
2998 }
2999 } else {
3000 /* completed sync */
3001 if ((!mddev->bitmap || conf->fullsync)
3002 && conf->have_replacement
3003 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3004 /* Completed a full sync so the replacements
3005 * are now fully recovered.
3006 */
3007 rcu_read_lock();
3008 for (i = 0; i < conf->geo.raid_disks; i++) {
3009 struct md_rdev *rdev =
3010 rcu_dereference(conf->mirrors[i].replacement);
3011 if (rdev)
3012 rdev->recovery_offset = MaxSector;
3013 }
3014 rcu_read_unlock();
3015 }
3016 conf->fullsync = 0;
3017 }
3018 bitmap_close_sync(mddev->bitmap);
3019 close_sync(conf);
3020 *skipped = 1;
3021 return sectors_skipped;
3022 }
3023
3024 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3025 return reshape_request(mddev, sector_nr, skipped);
3026
3027 if (chunks_skipped >= conf->geo.raid_disks) {
3028 /* if there has been nothing to do on any drive,
3029 * then there is nothing to do at all..
3030 */
3031 *skipped = 1;
3032 return (max_sector - sector_nr) + sectors_skipped;
3033 }
3034
3035 if (max_sector > mddev->resync_max)
3036 max_sector = mddev->resync_max; /* Don't do IO beyond here */
3037
3038 /* make sure whole request will fit in a chunk - if chunks
3039 * are meaningful
3040 */
3041 if (conf->geo.near_copies < conf->geo.raid_disks &&
3042 max_sector > (sector_nr | chunk_mask))
3043 max_sector = (sector_nr | chunk_mask) + 1;
3044
3045 /*
3046 * If there is non-resync activity waiting for a turn, then let it
3047 * though before starting on this new sync request.
3048 */
3049 if (conf->nr_waiting)
3050 schedule_timeout_uninterruptible(1);
3051
3052 /* Again, very different code for resync and recovery.
3053 * Both must result in an r10bio with a list of bios that
3054 * have bi_end_io, bi_sector, bi_bdev set,
3055 * and bi_private set to the r10bio.
3056 * For recovery, we may actually create several r10bios
3057 * with 2 bios in each, that correspond to the bios in the main one.
3058 * In this case, the subordinate r10bios link back through a
3059 * borrowed master_bio pointer, and the counter in the master
3060 * includes a ref from each subordinate.
3061 */
3062 /* First, we decide what to do and set ->bi_end_io
3063 * To end_sync_read if we want to read, and
3064 * end_sync_write if we will want to write.
3065 */
3066
3067 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3068 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3069 /* recovery... the complicated one */
3070 int j;
3071 r10_bio = NULL;
3072
3073 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3074 int still_degraded;
3075 struct r10bio *rb2;
3076 sector_t sect;
3077 int must_sync;
3078 int any_working;
3079 struct raid10_info *mirror = &conf->mirrors[i];
3080 struct md_rdev *mrdev, *mreplace;
3081
3082 rcu_read_lock();
3083 mrdev = rcu_dereference(mirror->rdev);
3084 mreplace = rcu_dereference(mirror->replacement);
3085
3086 if ((mrdev == NULL ||
3087 test_bit(Faulty, &mrdev->flags) ||
3088 test_bit(In_sync, &mrdev->flags)) &&
3089 (mreplace == NULL ||
3090 test_bit(Faulty, &mreplace->flags))) {
3091 rcu_read_unlock();
3092 continue;
3093 }
3094
3095 still_degraded = 0;
3096 /* want to reconstruct this device */
3097 rb2 = r10_bio;
3098 sect = raid10_find_virt(conf, sector_nr, i);
3099 if (sect >= mddev->resync_max_sectors) {
3100 /* last stripe is not complete - don't
3101 * try to recover this sector.
3102 */
3103 rcu_read_unlock();
3104 continue;
3105 }
3106 if (mreplace && test_bit(Faulty, &mreplace->flags))
3107 mreplace = NULL;
3108 /* Unless we are doing a full sync, or a replacement
3109 * we only need to recover the block if it is set in
3110 * the bitmap
3111 */
3112 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3113 &sync_blocks, 1);
3114 if (sync_blocks < max_sync)
3115 max_sync = sync_blocks;
3116 if (!must_sync &&
3117 mreplace == NULL &&
3118 !conf->fullsync) {
3119 /* yep, skip the sync_blocks here, but don't assume
3120 * that there will never be anything to do here
3121 */
3122 chunks_skipped = -1;
3123 rcu_read_unlock();
3124 continue;
3125 }
3126 atomic_inc(&mrdev->nr_pending);
3127 if (mreplace)
3128 atomic_inc(&mreplace->nr_pending);
3129 rcu_read_unlock();
3130
3131 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3132 r10_bio->state = 0;
3133 raise_barrier(conf, rb2 != NULL);
3134 atomic_set(&r10_bio->remaining, 0);
3135
3136 r10_bio->master_bio = (struct bio*)rb2;
3137 if (rb2)
3138 atomic_inc(&rb2->remaining);
3139 r10_bio->mddev = mddev;
3140 set_bit(R10BIO_IsRecover, &r10_bio->state);
3141 r10_bio->sector = sect;
3142
3143 raid10_find_phys(conf, r10_bio);
3144
3145 /* Need to check if the array will still be
3146 * degraded
3147 */
3148 rcu_read_lock();
3149 for (j = 0; j < conf->geo.raid_disks; j++) {
3150 struct md_rdev *rdev = rcu_dereference(
3151 conf->mirrors[j].rdev);
3152 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3153 still_degraded = 1;
3154 break;
3155 }
3156 }
3157
3158 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3159 &sync_blocks, still_degraded);
3160
3161 any_working = 0;
3162 for (j=0; j<conf->copies;j++) {
3163 int k;
3164 int d = r10_bio->devs[j].devnum;
3165 sector_t from_addr, to_addr;
3166 struct md_rdev *rdev =
3167 rcu_dereference(conf->mirrors[d].rdev);
3168 sector_t sector, first_bad;
3169 int bad_sectors;
3170 if (!rdev ||
3171 !test_bit(In_sync, &rdev->flags))
3172 continue;
3173 /* This is where we read from */
3174 any_working = 1;
3175 sector = r10_bio->devs[j].addr;
3176
3177 if (is_badblock(rdev, sector, max_sync,
3178 &first_bad, &bad_sectors)) {
3179 if (first_bad > sector)
3180 max_sync = first_bad - sector;
3181 else {
3182 bad_sectors -= (sector
3183 - first_bad);
3184 if (max_sync > bad_sectors)
3185 max_sync = bad_sectors;
3186 continue;
3187 }
3188 }
3189 bio = r10_bio->devs[0].bio;
3190 bio_reset(bio);
3191 bio->bi_next = biolist;
3192 biolist = bio;
3193 bio->bi_private = r10_bio;
3194 bio->bi_end_io = end_sync_read;
3195 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3196 if (test_bit(FailFast, &rdev->flags))
3197 bio->bi_opf |= MD_FAILFAST;
3198 from_addr = r10_bio->devs[j].addr;
3199 bio->bi_iter.bi_sector = from_addr +
3200 rdev->data_offset;
3201 bio->bi_bdev = rdev->bdev;
3202 atomic_inc(&rdev->nr_pending);
3203 /* and we write to 'i' (if not in_sync) */
3204
3205 for (k=0; k<conf->copies; k++)
3206 if (r10_bio->devs[k].devnum == i)
3207 break;
3208 BUG_ON(k == conf->copies);
3209 to_addr = r10_bio->devs[k].addr;
3210 r10_bio->devs[0].devnum = d;
3211 r10_bio->devs[0].addr = from_addr;
3212 r10_bio->devs[1].devnum = i;
3213 r10_bio->devs[1].addr = to_addr;
3214
3215 if (!test_bit(In_sync, &mrdev->flags)) {
3216 bio = r10_bio->devs[1].bio;
3217 bio_reset(bio);
3218 bio->bi_next = biolist;
3219 biolist = bio;
3220 bio->bi_private = r10_bio;
3221 bio->bi_end_io = end_sync_write;
3222 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3223 bio->bi_iter.bi_sector = to_addr
3224 + mrdev->data_offset;
3225 bio->bi_bdev = mrdev->bdev;
3226 atomic_inc(&r10_bio->remaining);
3227 } else
3228 r10_bio->devs[1].bio->bi_end_io = NULL;
3229
3230 /* and maybe write to replacement */
3231 bio = r10_bio->devs[1].repl_bio;
3232 if (bio)
3233 bio->bi_end_io = NULL;
3234 /* Note: if mreplace != NULL, then bio
3235 * cannot be NULL as r10buf_pool_alloc will
3236 * have allocated it.
3237 * So the second test here is pointless.
3238 * But it keeps semantic-checkers happy, and
3239 * this comment keeps human reviewers
3240 * happy.
3241 */
3242 if (mreplace == NULL || bio == NULL ||
3243 test_bit(Faulty, &mreplace->flags))
3244 break;
3245 bio_reset(bio);
3246 bio->bi_next = biolist;
3247 biolist = bio;
3248 bio->bi_private = r10_bio;
3249 bio->bi_end_io = end_sync_write;
3250 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3251 bio->bi_iter.bi_sector = to_addr +
3252 mreplace->data_offset;
3253 bio->bi_bdev = mreplace->bdev;
3254 atomic_inc(&r10_bio->remaining);
3255 break;
3256 }
3257 rcu_read_unlock();
3258 if (j == conf->copies) {
3259 /* Cannot recover, so abort the recovery or
3260 * record a bad block */
3261 if (any_working) {
3262 /* problem is that there are bad blocks
3263 * on other device(s)
3264 */
3265 int k;
3266 for (k = 0; k < conf->copies; k++)
3267 if (r10_bio->devs[k].devnum == i)
3268 break;
3269 if (!test_bit(In_sync,
3270 &mrdev->flags)
3271 && !rdev_set_badblocks(
3272 mrdev,
3273 r10_bio->devs[k].addr,
3274 max_sync, 0))
3275 any_working = 0;
3276 if (mreplace &&
3277 !rdev_set_badblocks(
3278 mreplace,
3279 r10_bio->devs[k].addr,
3280 max_sync, 0))
3281 any_working = 0;
3282 }
3283 if (!any_working) {
3284 if (!test_and_set_bit(MD_RECOVERY_INTR,
3285 &mddev->recovery))
3286 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3287 mdname(mddev));
3288 mirror->recovery_disabled
3289 = mddev->recovery_disabled;
3290 }
3291 put_buf(r10_bio);
3292 if (rb2)
3293 atomic_dec(&rb2->remaining);
3294 r10_bio = rb2;
3295 rdev_dec_pending(mrdev, mddev);
3296 if (mreplace)
3297 rdev_dec_pending(mreplace, mddev);
3298 break;
3299 }
3300 rdev_dec_pending(mrdev, mddev);
3301 if (mreplace)
3302 rdev_dec_pending(mreplace, mddev);
3303 if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3304 /* Only want this if there is elsewhere to
3305 * read from. 'j' is currently the first
3306 * readable copy.
3307 */
3308 int targets = 1;
3309 for (; j < conf->copies; j++) {
3310 int d = r10_bio->devs[j].devnum;
3311 if (conf->mirrors[d].rdev &&
3312 test_bit(In_sync,
3313 &conf->mirrors[d].rdev->flags))
3314 targets++;
3315 }
3316 if (targets == 1)
3317 r10_bio->devs[0].bio->bi_opf
3318 &= ~MD_FAILFAST;
3319 }
3320 }
3321 if (biolist == NULL) {
3322 while (r10_bio) {
3323 struct r10bio *rb2 = r10_bio;
3324 r10_bio = (struct r10bio*) rb2->master_bio;
3325 rb2->master_bio = NULL;
3326 put_buf(rb2);
3327 }
3328 goto giveup;
3329 }
3330 } else {
3331 /* resync. Schedule a read for every block at this virt offset */
3332 int count = 0;
3333
3334 bitmap_cond_end_sync(mddev->bitmap, sector_nr, 0);
3335
3336 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3337 &sync_blocks, mddev->degraded) &&
3338 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3339 &mddev->recovery)) {
3340 /* We can skip this block */
3341 *skipped = 1;
3342 return sync_blocks + sectors_skipped;
3343 }
3344 if (sync_blocks < max_sync)
3345 max_sync = sync_blocks;
3346 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3347 r10_bio->state = 0;
3348
3349 r10_bio->mddev = mddev;
3350 atomic_set(&r10_bio->remaining, 0);
3351 raise_barrier(conf, 0);
3352 conf->next_resync = sector_nr;
3353
3354 r10_bio->master_bio = NULL;
3355 r10_bio->sector = sector_nr;
3356 set_bit(R10BIO_IsSync, &r10_bio->state);
3357 raid10_find_phys(conf, r10_bio);
3358 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3359
3360 for (i = 0; i < conf->copies; i++) {
3361 int d = r10_bio->devs[i].devnum;
3362 sector_t first_bad, sector;
3363 int bad_sectors;
3364 struct md_rdev *rdev;
3365
3366 if (r10_bio->devs[i].repl_bio)
3367 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3368
3369 bio = r10_bio->devs[i].bio;
3370 bio_reset(bio);
3371 bio->bi_error = -EIO;
3372 rcu_read_lock();
3373 rdev = rcu_dereference(conf->mirrors[d].rdev);
3374 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3375 rcu_read_unlock();
3376 continue;
3377 }
3378 sector = r10_bio->devs[i].addr;
3379 if (is_badblock(rdev, sector, max_sync,
3380 &first_bad, &bad_sectors)) {
3381 if (first_bad > sector)
3382 max_sync = first_bad - sector;
3383 else {
3384 bad_sectors -= (sector - first_bad);
3385 if (max_sync > bad_sectors)
3386 max_sync = bad_sectors;
3387 rcu_read_unlock();
3388 continue;
3389 }
3390 }
3391 atomic_inc(&rdev->nr_pending);
3392 atomic_inc(&r10_bio->remaining);
3393 bio->bi_next = biolist;
3394 biolist = bio;
3395 bio->bi_private = r10_bio;
3396 bio->bi_end_io = end_sync_read;
3397 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3398 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
3399 bio->bi_opf |= MD_FAILFAST;
3400 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3401 bio->bi_bdev = rdev->bdev;
3402 count++;
3403
3404 rdev = rcu_dereference(conf->mirrors[d].replacement);
3405 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3406 rcu_read_unlock();
3407 continue;
3408 }
3409 atomic_inc(&rdev->nr_pending);
3410 rcu_read_unlock();
3411
3412 /* Need to set up for writing to the replacement */
3413 bio = r10_bio->devs[i].repl_bio;
3414 bio_reset(bio);
3415 bio->bi_error = -EIO;
3416
3417 sector = r10_bio->devs[i].addr;
3418 bio->bi_next = biolist;
3419 biolist = bio;
3420 bio->bi_private = r10_bio;
3421 bio->bi_end_io = end_sync_write;
3422 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3423 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
3424 bio->bi_opf |= MD_FAILFAST;
3425 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3426 bio->bi_bdev = rdev->bdev;
3427 count++;
3428 }
3429
3430 if (count < 2) {
3431 for (i=0; i<conf->copies; i++) {
3432 int d = r10_bio->devs[i].devnum;
3433 if (r10_bio->devs[i].bio->bi_end_io)
3434 rdev_dec_pending(conf->mirrors[d].rdev,
3435 mddev);
3436 if (r10_bio->devs[i].repl_bio &&
3437 r10_bio->devs[i].repl_bio->bi_end_io)
3438 rdev_dec_pending(
3439 conf->mirrors[d].replacement,
3440 mddev);
3441 }
3442 put_buf(r10_bio);
3443 biolist = NULL;
3444 goto giveup;
3445 }
3446 }
3447
3448 nr_sectors = 0;
3449 if (sector_nr + max_sync < max_sector)
3450 max_sector = sector_nr + max_sync;
3451 do {
3452 struct page *page;
3453 int len = PAGE_SIZE;
3454 if (sector_nr + (len>>9) > max_sector)
3455 len = (max_sector - sector_nr) << 9;
3456 if (len == 0)
3457 break;
3458 for (bio= biolist ; bio ; bio=bio->bi_next) {
3459 struct bio *bio2;
3460 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3461 if (bio_add_page(bio, page, len, 0))
3462 continue;
3463
3464 /* stop here */
3465 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3466 for (bio2 = biolist;
3467 bio2 && bio2 != bio;
3468 bio2 = bio2->bi_next) {
3469 /* remove last page from this bio */
3470 bio2->bi_vcnt--;
3471 bio2->bi_iter.bi_size -= len;
3472 bio_clear_flag(bio2, BIO_SEG_VALID);
3473 }
3474 goto bio_full;
3475 }
3476 nr_sectors += len>>9;
3477 sector_nr += len>>9;
3478 } while (biolist->bi_vcnt < RESYNC_PAGES);
3479 bio_full:
3480 r10_bio->sectors = nr_sectors;
3481
3482 while (biolist) {
3483 bio = biolist;
3484 biolist = biolist->bi_next;
3485
3486 bio->bi_next = NULL;
3487 r10_bio = bio->bi_private;
3488 r10_bio->sectors = nr_sectors;
3489
3490 if (bio->bi_end_io == end_sync_read) {
3491 md_sync_acct(bio->bi_bdev, nr_sectors);
3492 bio->bi_error = 0;
3493 generic_make_request(bio);
3494 }
3495 }
3496
3497 if (sectors_skipped)
3498 /* pretend they weren't skipped, it makes
3499 * no important difference in this case
3500 */
3501 md_done_sync(mddev, sectors_skipped, 1);
3502
3503 return sectors_skipped + nr_sectors;
3504 giveup:
3505 /* There is nowhere to write, so all non-sync
3506 * drives must be failed or in resync, all drives
3507 * have a bad block, so try the next chunk...
3508 */
3509 if (sector_nr + max_sync < max_sector)
3510 max_sector = sector_nr + max_sync;
3511
3512 sectors_skipped += (max_sector - sector_nr);
3513 chunks_skipped ++;
3514 sector_nr = max_sector;
3515 goto skipped;
3516}
3517
3518static sector_t
3519raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3520{
3521 sector_t size;
3522 struct r10conf *conf = mddev->private;
3523
3524 if (!raid_disks)
3525 raid_disks = min(conf->geo.raid_disks,
3526 conf->prev.raid_disks);
3527 if (!sectors)
3528 sectors = conf->dev_sectors;
3529
3530 size = sectors >> conf->geo.chunk_shift;
3531 sector_div(size, conf->geo.far_copies);
3532 size = size * raid_disks;
3533 sector_div(size, conf->geo.near_copies);
3534
3535 return size << conf->geo.chunk_shift;
3536}
3537
3538static void calc_sectors(struct r10conf *conf, sector_t size)
3539{
3540 /* Calculate the number of sectors-per-device that will
3541 * actually be used, and set conf->dev_sectors and
3542 * conf->stride
3543 */
3544
3545 size = size >> conf->geo.chunk_shift;
3546 sector_div(size, conf->geo.far_copies);
3547 size = size * conf->geo.raid_disks;
3548 sector_div(size, conf->geo.near_copies);
3549 /* 'size' is now the number of chunks in the array */
3550 /* calculate "used chunks per device" */
3551 size = size * conf->copies;
3552
3553 /* We need to round up when dividing by raid_disks to
3554 * get the stride size.
3555 */
3556 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3557
3558 conf->dev_sectors = size << conf->geo.chunk_shift;
3559
3560 if (conf->geo.far_offset)
3561 conf->geo.stride = 1 << conf->geo.chunk_shift;
3562 else {
3563 sector_div(size, conf->geo.far_copies);
3564 conf->geo.stride = size << conf->geo.chunk_shift;
3565 }
3566}
3567
3568enum geo_type {geo_new, geo_old, geo_start};
3569static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3570{
3571 int nc, fc, fo;
3572 int layout, chunk, disks;
3573 switch (new) {
3574 case geo_old:
3575 layout = mddev->layout;
3576 chunk = mddev->chunk_sectors;
3577 disks = mddev->raid_disks - mddev->delta_disks;
3578 break;
3579 case geo_new:
3580 layout = mddev->new_layout;
3581 chunk = mddev->new_chunk_sectors;
3582 disks = mddev->raid_disks;
3583 break;
3584 default: /* avoid 'may be unused' warnings */
3585 case geo_start: /* new when starting reshape - raid_disks not
3586 * updated yet. */
3587 layout = mddev->new_layout;
3588 chunk = mddev->new_chunk_sectors;
3589 disks = mddev->raid_disks + mddev->delta_disks;
3590 break;
3591 }
3592 if (layout >> 19)
3593 return -1;
3594 if (chunk < (PAGE_SIZE >> 9) ||
3595 !is_power_of_2(chunk))
3596 return -2;
3597 nc = layout & 255;
3598 fc = (layout >> 8) & 255;
3599 fo = layout & (1<<16);
3600 geo->raid_disks = disks;
3601 geo->near_copies = nc;
3602 geo->far_copies = fc;
3603 geo->far_offset = fo;
3604 switch (layout >> 17) {
3605 case 0: /* original layout. simple but not always optimal */
3606 geo->far_set_size = disks;
3607 break;
3608 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3609 * actually using this, but leave code here just in case.*/
3610 geo->far_set_size = disks/fc;
3611 WARN(geo->far_set_size < fc,
3612 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3613 break;
3614 case 2: /* "improved" layout fixed to match documentation */
3615 geo->far_set_size = fc * nc;
3616 break;
3617 default: /* Not a valid layout */
3618 return -1;
3619 }
3620 geo->chunk_mask = chunk - 1;
3621 geo->chunk_shift = ffz(~chunk);
3622 return nc*fc;
3623}
3624
3625static struct r10conf *setup_conf(struct mddev *mddev)
3626{
3627 struct r10conf *conf = NULL;
3628 int err = -EINVAL;
3629 struct geom geo;
3630 int copies;
3631
3632 copies = setup_geo(&geo, mddev, geo_new);
3633
3634 if (copies == -2) {
3635 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3636 mdname(mddev), PAGE_SIZE);
3637 goto out;
3638 }
3639
3640 if (copies < 2 || copies > mddev->raid_disks) {
3641 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3642 mdname(mddev), mddev->new_layout);
3643 goto out;
3644 }
3645
3646 err = -ENOMEM;
3647 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3648 if (!conf)
3649 goto out;
3650
3651 /* FIXME calc properly */
3652 conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3653 max(0,-mddev->delta_disks)),
3654 GFP_KERNEL);
3655 if (!conf->mirrors)
3656 goto out;
3657
3658 conf->tmppage = alloc_page(GFP_KERNEL);
3659 if (!conf->tmppage)
3660 goto out;
3661
3662 conf->geo = geo;
3663 conf->copies = copies;
3664 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3665 r10bio_pool_free, conf);
3666 if (!conf->r10bio_pool)
3667 goto out;
3668
3669 calc_sectors(conf, mddev->dev_sectors);
3670 if (mddev->reshape_position == MaxSector) {
3671 conf->prev = conf->geo;
3672 conf->reshape_progress = MaxSector;
3673 } else {
3674 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3675 err = -EINVAL;
3676 goto out;
3677 }
3678 conf->reshape_progress = mddev->reshape_position;
3679 if (conf->prev.far_offset)
3680 conf->prev.stride = 1 << conf->prev.chunk_shift;
3681 else
3682 /* far_copies must be 1 */
3683 conf->prev.stride = conf->dev_sectors;
3684 }
3685 conf->reshape_safe = conf->reshape_progress;
3686 spin_lock_init(&conf->device_lock);
3687 INIT_LIST_HEAD(&conf->retry_list);
3688 INIT_LIST_HEAD(&conf->bio_end_io_list);
3689
3690 spin_lock_init(&conf->resync_lock);
3691 init_waitqueue_head(&conf->wait_barrier);
3692 atomic_set(&conf->nr_pending, 0);
3693
3694 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3695 if (!conf->thread)
3696 goto out;
3697
3698 conf->mddev = mddev;
3699 return conf;
3700
3701 out:
3702 if (conf) {
3703 mempool_destroy(conf->r10bio_pool);
3704 kfree(conf->mirrors);
3705 safe_put_page(conf->tmppage);
3706 kfree(conf);
3707 }
3708 return ERR_PTR(err);
3709}
3710
3711static int raid10_run(struct mddev *mddev)
3712{
3713 struct r10conf *conf;
3714 int i, disk_idx, chunk_size;
3715 struct raid10_info *disk;
3716 struct md_rdev *rdev;
3717 sector_t size;
3718 sector_t min_offset_diff = 0;
3719 int first = 1;
3720 bool discard_supported = false;
3721
3722 if (mddev->private == NULL) {
3723 conf = setup_conf(mddev);
3724 if (IS_ERR(conf))
3725 return PTR_ERR(conf);
3726 mddev->private = conf;
3727 }
3728 conf = mddev->private;
3729 if (!conf)
3730 goto out;
3731
3732 mddev->thread = conf->thread;
3733 conf->thread = NULL;
3734
3735 chunk_size = mddev->chunk_sectors << 9;
3736 if (mddev->queue) {
3737 blk_queue_max_discard_sectors(mddev->queue,
3738 mddev->chunk_sectors);
3739 blk_queue_max_write_same_sectors(mddev->queue, 0);
3740 blk_queue_io_min(mddev->queue, chunk_size);
3741 if (conf->geo.raid_disks % conf->geo.near_copies)
3742 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3743 else
3744 blk_queue_io_opt(mddev->queue, chunk_size *
3745 (conf->geo.raid_disks / conf->geo.near_copies));
3746 }
3747
3748 rdev_for_each(rdev, mddev) {
3749 long long diff;
3750 struct request_queue *q;
3751
3752 disk_idx = rdev->raid_disk;
3753 if (disk_idx < 0)
3754 continue;
3755 if (disk_idx >= conf->geo.raid_disks &&
3756 disk_idx >= conf->prev.raid_disks)
3757 continue;
3758 disk = conf->mirrors + disk_idx;
3759
3760 if (test_bit(Replacement, &rdev->flags)) {
3761 if (disk->replacement)
3762 goto out_free_conf;
3763 disk->replacement = rdev;
3764 } else {
3765 if (disk->rdev)
3766 goto out_free_conf;
3767 disk->rdev = rdev;
3768 }
3769 q = bdev_get_queue(rdev->bdev);
3770 diff = (rdev->new_data_offset - rdev->data_offset);
3771 if (!mddev->reshape_backwards)
3772 diff = -diff;
3773 if (diff < 0)
3774 diff = 0;
3775 if (first || diff < min_offset_diff)
3776 min_offset_diff = diff;
3777
3778 if (mddev->gendisk)
3779 disk_stack_limits(mddev->gendisk, rdev->bdev,
3780 rdev->data_offset << 9);
3781
3782 disk->head_position = 0;
3783
3784 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3785 discard_supported = true;
3786 }
3787
3788 if (mddev->queue) {
3789 if (discard_supported)
3790 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3791 mddev->queue);
3792 else
3793 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3794 mddev->queue);
3795 }
3796 /* need to check that every block has at least one working mirror */
3797 if (!enough(conf, -1)) {
3798 pr_err("md/raid10:%s: not enough operational mirrors.\n",
3799 mdname(mddev));
3800 goto out_free_conf;
3801 }
3802
3803 if (conf->reshape_progress != MaxSector) {
3804 /* must ensure that shape change is supported */
3805 if (conf->geo.far_copies != 1 &&
3806 conf->geo.far_offset == 0)
3807 goto out_free_conf;
3808 if (conf->prev.far_copies != 1 &&
3809 conf->prev.far_offset == 0)
3810 goto out_free_conf;
3811 }
3812
3813 mddev->degraded = 0;
3814 for (i = 0;
3815 i < conf->geo.raid_disks
3816 || i < conf->prev.raid_disks;
3817 i++) {
3818
3819 disk = conf->mirrors + i;
3820
3821 if (!disk->rdev && disk->replacement) {
3822 /* The replacement is all we have - use it */
3823 disk->rdev = disk->replacement;
3824 disk->replacement = NULL;
3825 clear_bit(Replacement, &disk->rdev->flags);
3826 }
3827
3828 if (!disk->rdev ||
3829 !test_bit(In_sync, &disk->rdev->flags)) {
3830 disk->head_position = 0;
3831 mddev->degraded++;
3832 if (disk->rdev &&
3833 disk->rdev->saved_raid_disk < 0)
3834 conf->fullsync = 1;
3835 }
3836 disk->recovery_disabled = mddev->recovery_disabled - 1;
3837 }
3838
3839 if (mddev->recovery_cp != MaxSector)
3840 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
3841 mdname(mddev));
3842 pr_info("md/raid10:%s: active with %d out of %d devices\n",
3843 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3844 conf->geo.raid_disks);
3845 /*
3846 * Ok, everything is just fine now
3847 */
3848 mddev->dev_sectors = conf->dev_sectors;
3849 size = raid10_size(mddev, 0, 0);
3850 md_set_array_sectors(mddev, size);
3851 mddev->resync_max_sectors = size;
3852 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3853
3854 if (mddev->queue) {
3855 int stripe = conf->geo.raid_disks *
3856 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3857
3858 /* Calculate max read-ahead size.
3859 * We need to readahead at least twice a whole stripe....
3860 * maybe...
3861 */
3862 stripe /= conf->geo.near_copies;
3863 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3864 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3865 }
3866
3867 if (md_integrity_register(mddev))
3868 goto out_free_conf;
3869
3870 if (conf->reshape_progress != MaxSector) {
3871 unsigned long before_length, after_length;
3872
3873 before_length = ((1 << conf->prev.chunk_shift) *
3874 conf->prev.far_copies);
3875 after_length = ((1 << conf->geo.chunk_shift) *
3876 conf->geo.far_copies);
3877
3878 if (max(before_length, after_length) > min_offset_diff) {
3879 /* This cannot work */
3880 pr_warn("md/raid10: offset difference not enough to continue reshape\n");
3881 goto out_free_conf;
3882 }
3883 conf->offset_diff = min_offset_diff;
3884
3885 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3886 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3887 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3888 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3889 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3890 "reshape");
3891 }
3892
3893 return 0;
3894
3895out_free_conf:
3896 md_unregister_thread(&mddev->thread);
3897 mempool_destroy(conf->r10bio_pool);
3898 safe_put_page(conf->tmppage);
3899 kfree(conf->mirrors);
3900 kfree(conf);
3901 mddev->private = NULL;
3902out:
3903 return -EIO;
3904}
3905
3906static void raid10_free(struct mddev *mddev, void *priv)
3907{
3908 struct r10conf *conf = priv;
3909
3910 mempool_destroy(conf->r10bio_pool);
3911 safe_put_page(conf->tmppage);
3912 kfree(conf->mirrors);
3913 kfree(conf->mirrors_old);
3914 kfree(conf->mirrors_new);
3915 kfree(conf);
3916}
3917
3918static void raid10_quiesce(struct mddev *mddev, int state)
3919{
3920 struct r10conf *conf = mddev->private;
3921
3922 switch(state) {
3923 case 1:
3924 raise_barrier(conf, 0);
3925 break;
3926 case 0:
3927 lower_barrier(conf);
3928 break;
3929 }
3930}
3931
3932static int raid10_resize(struct mddev *mddev, sector_t sectors)
3933{
3934 /* Resize of 'far' arrays is not supported.
3935 * For 'near' and 'offset' arrays we can set the
3936 * number of sectors used to be an appropriate multiple
3937 * of the chunk size.
3938 * For 'offset', this is far_copies*chunksize.
3939 * For 'near' the multiplier is the LCM of
3940 * near_copies and raid_disks.
3941 * So if far_copies > 1 && !far_offset, fail.
3942 * Else find LCM(raid_disks, near_copy)*far_copies and
3943 * multiply by chunk_size. Then round to this number.
3944 * This is mostly done by raid10_size()
3945 */
3946 struct r10conf *conf = mddev->private;
3947 sector_t oldsize, size;
3948
3949 if (mddev->reshape_position != MaxSector)
3950 return -EBUSY;
3951
3952 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3953 return -EINVAL;
3954
3955 oldsize = raid10_size(mddev, 0, 0);
3956 size = raid10_size(mddev, sectors, 0);
3957 if (mddev->external_size &&
3958 mddev->array_sectors > size)
3959 return -EINVAL;
3960 if (mddev->bitmap) {
3961 int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3962 if (ret)
3963 return ret;
3964 }
3965 md_set_array_sectors(mddev, size);
3966 if (mddev->queue) {
3967 set_capacity(mddev->gendisk, mddev->array_sectors);
3968 revalidate_disk(mddev->gendisk);
3969 }
3970 if (sectors > mddev->dev_sectors &&
3971 mddev->recovery_cp > oldsize) {
3972 mddev->recovery_cp = oldsize;
3973 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3974 }
3975 calc_sectors(conf, sectors);
3976 mddev->dev_sectors = conf->dev_sectors;
3977 mddev->resync_max_sectors = size;
3978 return 0;
3979}
3980
3981static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
3982{
3983 struct md_rdev *rdev;
3984 struct r10conf *conf;
3985
3986 if (mddev->degraded > 0) {
3987 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
3988 mdname(mddev));
3989 return ERR_PTR(-EINVAL);
3990 }
3991 sector_div(size, devs);
3992
3993 /* Set new parameters */
3994 mddev->new_level = 10;
3995 /* new layout: far_copies = 1, near_copies = 2 */
3996 mddev->new_layout = (1<<8) + 2;
3997 mddev->new_chunk_sectors = mddev->chunk_sectors;
3998 mddev->delta_disks = mddev->raid_disks;
3999 mddev->raid_disks *= 2;
4000 /* make sure it will be not marked as dirty */
4001 mddev->recovery_cp = MaxSector;
4002 mddev->dev_sectors = size;
4003
4004 conf = setup_conf(mddev);
4005 if (!IS_ERR(conf)) {
4006 rdev_for_each(rdev, mddev)
4007 if (rdev->raid_disk >= 0) {
4008 rdev->new_raid_disk = rdev->raid_disk * 2;
4009 rdev->sectors = size;
4010 }
4011 conf->barrier = 1;
4012 }
4013
4014 return conf;
4015}
4016
4017static void *raid10_takeover(struct mddev *mddev)
4018{
4019 struct r0conf *raid0_conf;
4020
4021 /* raid10 can take over:
4022 * raid0 - providing it has only two drives
4023 */
4024 if (mddev->level == 0) {
4025 /* for raid0 takeover only one zone is supported */
4026 raid0_conf = mddev->private;
4027 if (raid0_conf->nr_strip_zones > 1) {
4028 pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4029 mdname(mddev));
4030 return ERR_PTR(-EINVAL);
4031 }
4032 return raid10_takeover_raid0(mddev,
4033 raid0_conf->strip_zone->zone_end,
4034 raid0_conf->strip_zone->nb_dev);
4035 }
4036 return ERR_PTR(-EINVAL);
4037}
4038
4039static int raid10_check_reshape(struct mddev *mddev)
4040{
4041 /* Called when there is a request to change
4042 * - layout (to ->new_layout)
4043 * - chunk size (to ->new_chunk_sectors)
4044 * - raid_disks (by delta_disks)
4045 * or when trying to restart a reshape that was ongoing.
4046 *
4047 * We need to validate the request and possibly allocate
4048 * space if that might be an issue later.
4049 *
4050 * Currently we reject any reshape of a 'far' mode array,
4051 * allow chunk size to change if new is generally acceptable,
4052 * allow raid_disks to increase, and allow
4053 * a switch between 'near' mode and 'offset' mode.
4054 */
4055 struct r10conf *conf = mddev->private;
4056 struct geom geo;
4057
4058 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
4059 return -EINVAL;
4060
4061 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
4062 /* mustn't change number of copies */
4063 return -EINVAL;
4064 if (geo.far_copies > 1 && !geo.far_offset)
4065 /* Cannot switch to 'far' mode */
4066 return -EINVAL;
4067
4068 if (mddev->array_sectors & geo.chunk_mask)
4069 /* not factor of array size */
4070 return -EINVAL;
4071
4072 if (!enough(conf, -1))
4073 return -EINVAL;
4074
4075 kfree(conf->mirrors_new);
4076 conf->mirrors_new = NULL;
4077 if (mddev->delta_disks > 0) {
4078 /* allocate new 'mirrors' list */
4079 conf->mirrors_new = kzalloc(
4080 sizeof(struct raid10_info)
4081 *(mddev->raid_disks +
4082 mddev->delta_disks),
4083 GFP_KERNEL);
4084 if (!conf->mirrors_new)
4085 return -ENOMEM;
4086 }
4087 return 0;
4088}
4089
4090/*
4091 * Need to check if array has failed when deciding whether to:
4092 * - start an array
4093 * - remove non-faulty devices
4094 * - add a spare
4095 * - allow a reshape
4096 * This determination is simple when no reshape is happening.
4097 * However if there is a reshape, we need to carefully check
4098 * both the before and after sections.
4099 * This is because some failed devices may only affect one
4100 * of the two sections, and some non-in_sync devices may
4101 * be insync in the section most affected by failed devices.
4102 */
4103static int calc_degraded(struct r10conf *conf)
4104{
4105 int degraded, degraded2;
4106 int i;
4107
4108 rcu_read_lock();
4109 degraded = 0;
4110 /* 'prev' section first */
4111 for (i = 0; i < conf->prev.raid_disks; i++) {
4112 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4113 if (!rdev || test_bit(Faulty, &rdev->flags))
4114 degraded++;
4115 else if (!test_bit(In_sync, &rdev->flags))
4116 /* When we can reduce the number of devices in
4117 * an array, this might not contribute to
4118 * 'degraded'. It does now.
4119 */
4120 degraded++;
4121 }
4122 rcu_read_unlock();
4123 if (conf->geo.raid_disks == conf->prev.raid_disks)
4124 return degraded;
4125 rcu_read_lock();
4126 degraded2 = 0;
4127 for (i = 0; i < conf->geo.raid_disks; i++) {
4128 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4129 if (!rdev || test_bit(Faulty, &rdev->flags))
4130 degraded2++;
4131 else if (!test_bit(In_sync, &rdev->flags)) {
4132 /* If reshape is increasing the number of devices,
4133 * this section has already been recovered, so
4134 * it doesn't contribute to degraded.
4135 * else it does.
4136 */
4137 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4138 degraded2++;
4139 }
4140 }
4141 rcu_read_unlock();
4142 if (degraded2 > degraded)
4143 return degraded2;
4144 return degraded;
4145}
4146
4147static int raid10_start_reshape(struct mddev *mddev)
4148{
4149 /* A 'reshape' has been requested. This commits
4150 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4151 * This also checks if there are enough spares and adds them
4152 * to the array.
4153 * We currently require enough spares to make the final
4154 * array non-degraded. We also require that the difference
4155 * between old and new data_offset - on each device - is
4156 * enough that we never risk over-writing.
4157 */
4158
4159 unsigned long before_length, after_length;
4160 sector_t min_offset_diff = 0;
4161 int first = 1;
4162 struct geom new;
4163 struct r10conf *conf = mddev->private;
4164 struct md_rdev *rdev;
4165 int spares = 0;
4166 int ret;
4167
4168 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4169 return -EBUSY;
4170
4171 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4172 return -EINVAL;
4173
4174 before_length = ((1 << conf->prev.chunk_shift) *
4175 conf->prev.far_copies);
4176 after_length = ((1 << conf->geo.chunk_shift) *
4177 conf->geo.far_copies);
4178
4179 rdev_for_each(rdev, mddev) {
4180 if (!test_bit(In_sync, &rdev->flags)
4181 && !test_bit(Faulty, &rdev->flags))
4182 spares++;
4183 if (rdev->raid_disk >= 0) {
4184 long long diff = (rdev->new_data_offset
4185 - rdev->data_offset);
4186 if (!mddev->reshape_backwards)
4187 diff = -diff;
4188 if (diff < 0)
4189 diff = 0;
4190 if (first || diff < min_offset_diff)
4191 min_offset_diff = diff;
4192 }
4193 }
4194
4195 if (max(before_length, after_length) > min_offset_diff)
4196 return -EINVAL;
4197
4198 if (spares < mddev->delta_disks)
4199 return -EINVAL;
4200
4201 conf->offset_diff = min_offset_diff;
4202 spin_lock_irq(&conf->device_lock);
4203 if (conf->mirrors_new) {
4204 memcpy(conf->mirrors_new, conf->mirrors,
4205 sizeof(struct raid10_info)*conf->prev.raid_disks);
4206 smp_mb();
4207 kfree(conf->mirrors_old);
4208 conf->mirrors_old = conf->mirrors;
4209 conf->mirrors = conf->mirrors_new;
4210 conf->mirrors_new = NULL;
4211 }
4212 setup_geo(&conf->geo, mddev, geo_start);
4213 smp_mb();
4214 if (mddev->reshape_backwards) {
4215 sector_t size = raid10_size(mddev, 0, 0);
4216 if (size < mddev->array_sectors) {
4217 spin_unlock_irq(&conf->device_lock);
4218 pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4219 mdname(mddev));
4220 return -EINVAL;
4221 }
4222 mddev->resync_max_sectors = size;
4223 conf->reshape_progress = size;
4224 } else
4225 conf->reshape_progress = 0;
4226 conf->reshape_safe = conf->reshape_progress;
4227 spin_unlock_irq(&conf->device_lock);
4228
4229 if (mddev->delta_disks && mddev->bitmap) {
4230 ret = bitmap_resize(mddev->bitmap,
4231 raid10_size(mddev, 0,
4232 conf->geo.raid_disks),
4233 0, 0);
4234 if (ret)
4235 goto abort;
4236 }
4237 if (mddev->delta_disks > 0) {
4238 rdev_for_each(rdev, mddev)
4239 if (rdev->raid_disk < 0 &&
4240 !test_bit(Faulty, &rdev->flags)) {
4241 if (raid10_add_disk(mddev, rdev) == 0) {
4242 if (rdev->raid_disk >=
4243 conf->prev.raid_disks)
4244 set_bit(In_sync, &rdev->flags);
4245 else
4246 rdev->recovery_offset = 0;
4247
4248 if (sysfs_link_rdev(mddev, rdev))
4249 /* Failure here is OK */;
4250 }
4251 } else if (rdev->raid_disk >= conf->prev.raid_disks
4252 && !test_bit(Faulty, &rdev->flags)) {
4253 /* This is a spare that was manually added */
4254 set_bit(In_sync, &rdev->flags);
4255 }
4256 }
4257 /* When a reshape changes the number of devices,
4258 * ->degraded is measured against the larger of the
4259 * pre and post numbers.
4260 */
4261 spin_lock_irq(&conf->device_lock);
4262 mddev->degraded = calc_degraded(conf);
4263 spin_unlock_irq(&conf->device_lock);
4264 mddev->raid_disks = conf->geo.raid_disks;
4265 mddev->reshape_position = conf->reshape_progress;
4266 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4267
4268 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4269 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4270 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4271 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4272 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4273
4274 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4275 "reshape");
4276 if (!mddev->sync_thread) {
4277 ret = -EAGAIN;
4278 goto abort;
4279 }
4280 conf->reshape_checkpoint = jiffies;
4281 md_wakeup_thread(mddev->sync_thread);
4282 md_new_event(mddev);
4283 return 0;
4284
4285abort:
4286 mddev->recovery = 0;
4287 spin_lock_irq(&conf->device_lock);
4288 conf->geo = conf->prev;
4289 mddev->raid_disks = conf->geo.raid_disks;
4290 rdev_for_each(rdev, mddev)
4291 rdev->new_data_offset = rdev->data_offset;
4292 smp_wmb();
4293 conf->reshape_progress = MaxSector;
4294 conf->reshape_safe = MaxSector;
4295 mddev->reshape_position = MaxSector;
4296 spin_unlock_irq(&conf->device_lock);
4297 return ret;
4298}
4299
4300/* Calculate the last device-address that could contain
4301 * any block from the chunk that includes the array-address 's'
4302 * and report the next address.
4303 * i.e. the address returned will be chunk-aligned and after
4304 * any data that is in the chunk containing 's'.
4305 */
4306static sector_t last_dev_address(sector_t s, struct geom *geo)
4307{
4308 s = (s | geo->chunk_mask) + 1;
4309 s >>= geo->chunk_shift;
4310 s *= geo->near_copies;
4311 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4312 s *= geo->far_copies;
4313 s <<= geo->chunk_shift;
4314 return s;
4315}
4316
4317/* Calculate the first device-address that could contain
4318 * any block from the chunk that includes the array-address 's'.
4319 * This too will be the start of a chunk
4320 */
4321static sector_t first_dev_address(sector_t s, struct geom *geo)
4322{
4323 s >>= geo->chunk_shift;
4324 s *= geo->near_copies;
4325 sector_div(s, geo->raid_disks);
4326 s *= geo->far_copies;
4327 s <<= geo->chunk_shift;
4328 return s;
4329}
4330
4331static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4332 int *skipped)
4333{
4334 /* We simply copy at most one chunk (smallest of old and new)
4335 * at a time, possibly less if that exceeds RESYNC_PAGES,
4336 * or we hit a bad block or something.
4337 * This might mean we pause for normal IO in the middle of
4338 * a chunk, but that is not a problem as mddev->reshape_position
4339 * can record any location.
4340 *
4341 * If we will want to write to a location that isn't
4342 * yet recorded as 'safe' (i.e. in metadata on disk) then
4343 * we need to flush all reshape requests and update the metadata.
4344 *
4345 * When reshaping forwards (e.g. to more devices), we interpret
4346 * 'safe' as the earliest block which might not have been copied
4347 * down yet. We divide this by previous stripe size and multiply
4348 * by previous stripe length to get lowest device offset that we
4349 * cannot write to yet.
4350 * We interpret 'sector_nr' as an address that we want to write to.
4351 * From this we use last_device_address() to find where we might
4352 * write to, and first_device_address on the 'safe' position.
4353 * If this 'next' write position is after the 'safe' position,
4354 * we must update the metadata to increase the 'safe' position.
4355 *
4356 * When reshaping backwards, we round in the opposite direction
4357 * and perform the reverse test: next write position must not be
4358 * less than current safe position.
4359 *
4360 * In all this the minimum difference in data offsets
4361 * (conf->offset_diff - always positive) allows a bit of slack,
4362 * so next can be after 'safe', but not by more than offset_diff
4363 *
4364 * We need to prepare all the bios here before we start any IO
4365 * to ensure the size we choose is acceptable to all devices.
4366 * The means one for each copy for write-out and an extra one for
4367 * read-in.
4368 * We store the read-in bio in ->master_bio and the others in
4369 * ->devs[x].bio and ->devs[x].repl_bio.
4370 */
4371 struct r10conf *conf = mddev->private;
4372 struct r10bio *r10_bio;
4373 sector_t next, safe, last;
4374 int max_sectors;
4375 int nr_sectors;
4376 int s;
4377 struct md_rdev *rdev;
4378 int need_flush = 0;
4379 struct bio *blist;
4380 struct bio *bio, *read_bio;
4381 int sectors_done = 0;
4382
4383 if (sector_nr == 0) {
4384 /* If restarting in the middle, skip the initial sectors */
4385 if (mddev->reshape_backwards &&
4386 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4387 sector_nr = (raid10_size(mddev, 0, 0)
4388 - conf->reshape_progress);
4389 } else if (!mddev->reshape_backwards &&
4390 conf->reshape_progress > 0)
4391 sector_nr = conf->reshape_progress;
4392 if (sector_nr) {
4393 mddev->curr_resync_completed = sector_nr;
4394 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4395 *skipped = 1;
4396 return sector_nr;
4397 }
4398 }
4399
4400 /* We don't use sector_nr to track where we are up to
4401 * as that doesn't work well for ->reshape_backwards.
4402 * So just use ->reshape_progress.
4403 */
4404 if (mddev->reshape_backwards) {
4405 /* 'next' is the earliest device address that we might
4406 * write to for this chunk in the new layout
4407 */
4408 next = first_dev_address(conf->reshape_progress - 1,
4409 &conf->geo);
4410
4411 /* 'safe' is the last device address that we might read from
4412 * in the old layout after a restart
4413 */
4414 safe = last_dev_address(conf->reshape_safe - 1,
4415 &conf->prev);
4416
4417 if (next + conf->offset_diff < safe)
4418 need_flush = 1;
4419
4420 last = conf->reshape_progress - 1;
4421 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4422 & conf->prev.chunk_mask);
4423 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4424 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4425 } else {
4426 /* 'next' is after the last device address that we
4427 * might write to for this chunk in the new layout
4428 */
4429 next = last_dev_address(conf->reshape_progress, &conf->geo);
4430
4431 /* 'safe' is the earliest device address that we might
4432 * read from in the old layout after a restart
4433 */
4434 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4435
4436 /* Need to update metadata if 'next' might be beyond 'safe'
4437 * as that would possibly corrupt data
4438 */
4439 if (next > safe + conf->offset_diff)
4440 need_flush = 1;
4441
4442 sector_nr = conf->reshape_progress;
4443 last = sector_nr | (conf->geo.chunk_mask
4444 & conf->prev.chunk_mask);
4445
4446 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4447 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4448 }
4449
4450 if (need_flush ||
4451 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4452 /* Need to update reshape_position in metadata */
4453 wait_barrier(conf);
4454 mddev->reshape_position = conf->reshape_progress;
4455 if (mddev->reshape_backwards)
4456 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4457 - conf->reshape_progress;
4458 else
4459 mddev->curr_resync_completed = conf->reshape_progress;
4460 conf->reshape_checkpoint = jiffies;
4461 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4462 md_wakeup_thread(mddev->thread);
4463 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4464 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4465 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4466 allow_barrier(conf);
4467 return sectors_done;
4468 }
4469 conf->reshape_safe = mddev->reshape_position;
4470 allow_barrier(conf);
4471 }
4472
4473read_more:
4474 /* Now schedule reads for blocks from sector_nr to last */
4475 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4476 r10_bio->state = 0;
4477 raise_barrier(conf, sectors_done != 0);
4478 atomic_set(&r10_bio->remaining, 0);
4479 r10_bio->mddev = mddev;
4480 r10_bio->sector = sector_nr;
4481 set_bit(R10BIO_IsReshape, &r10_bio->state);
4482 r10_bio->sectors = last - sector_nr + 1;
4483 rdev = read_balance(conf, r10_bio, &max_sectors);
4484 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4485
4486 if (!rdev) {
4487 /* Cannot read from here, so need to record bad blocks
4488 * on all the target devices.
4489 */
4490 // FIXME
4491 mempool_free(r10_bio, conf->r10buf_pool);
4492 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4493 return sectors_done;
4494 }
4495
4496 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4497
4498 read_bio->bi_bdev = rdev->bdev;
4499 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4500 + rdev->data_offset);
4501 read_bio->bi_private = r10_bio;
4502 read_bio->bi_end_io = end_sync_read;
4503 bio_set_op_attrs(read_bio, REQ_OP_READ, 0);
4504 read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4505 read_bio->bi_error = 0;
4506 read_bio->bi_vcnt = 0;
4507 read_bio->bi_iter.bi_size = 0;
4508 r10_bio->master_bio = read_bio;
4509 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4510
4511 /* Now find the locations in the new layout */
4512 __raid10_find_phys(&conf->geo, r10_bio);
4513
4514 blist = read_bio;
4515 read_bio->bi_next = NULL;
4516
4517 rcu_read_lock();
4518 for (s = 0; s < conf->copies*2; s++) {
4519 struct bio *b;
4520 int d = r10_bio->devs[s/2].devnum;
4521 struct md_rdev *rdev2;
4522 if (s&1) {
4523 rdev2 = rcu_dereference(conf->mirrors[d].replacement);
4524 b = r10_bio->devs[s/2].repl_bio;
4525 } else {
4526 rdev2 = rcu_dereference(conf->mirrors[d].rdev);
4527 b = r10_bio->devs[s/2].bio;
4528 }
4529 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4530 continue;
4531
4532 bio_reset(b);
4533 b->bi_bdev = rdev2->bdev;
4534 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4535 rdev2->new_data_offset;
4536 b->bi_private = r10_bio;
4537 b->bi_end_io = end_reshape_write;
4538 bio_set_op_attrs(b, REQ_OP_WRITE, 0);
4539 b->bi_next = blist;
4540 blist = b;
4541 }
4542
4543 /* Now add as many pages as possible to all of these bios. */
4544
4545 nr_sectors = 0;
4546 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4547 struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4548 int len = (max_sectors - s) << 9;
4549 if (len > PAGE_SIZE)
4550 len = PAGE_SIZE;
4551 for (bio = blist; bio ; bio = bio->bi_next) {
4552 struct bio *bio2;
4553 if (bio_add_page(bio, page, len, 0))
4554 continue;
4555
4556 /* Didn't fit, must stop */
4557 for (bio2 = blist;
4558 bio2 && bio2 != bio;
4559 bio2 = bio2->bi_next) {
4560 /* Remove last page from this bio */
4561 bio2->bi_vcnt--;
4562 bio2->bi_iter.bi_size -= len;
4563 bio_clear_flag(bio2, BIO_SEG_VALID);
4564 }
4565 goto bio_full;
4566 }
4567 sector_nr += len >> 9;
4568 nr_sectors += len >> 9;
4569 }
4570bio_full:
4571 rcu_read_unlock();
4572 r10_bio->sectors = nr_sectors;
4573
4574 /* Now submit the read */
4575 md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4576 atomic_inc(&r10_bio->remaining);
4577 read_bio->bi_next = NULL;
4578 generic_make_request(read_bio);
4579 sector_nr += nr_sectors;
4580 sectors_done += nr_sectors;
4581 if (sector_nr <= last)
4582 goto read_more;
4583
4584 /* Now that we have done the whole section we can
4585 * update reshape_progress
4586 */
4587 if (mddev->reshape_backwards)
4588 conf->reshape_progress -= sectors_done;
4589 else
4590 conf->reshape_progress += sectors_done;
4591
4592 return sectors_done;
4593}
4594
4595static void end_reshape_request(struct r10bio *r10_bio);
4596static int handle_reshape_read_error(struct mddev *mddev,
4597 struct r10bio *r10_bio);
4598static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4599{
4600 /* Reshape read completed. Hopefully we have a block
4601 * to write out.
4602 * If we got a read error then we do sync 1-page reads from
4603 * elsewhere until we find the data - or give up.
4604 */
4605 struct r10conf *conf = mddev->private;
4606 int s;
4607
4608 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4609 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4610 /* Reshape has been aborted */
4611 md_done_sync(mddev, r10_bio->sectors, 0);
4612 return;
4613 }
4614
4615 /* We definitely have the data in the pages, schedule the
4616 * writes.
4617 */
4618 atomic_set(&r10_bio->remaining, 1);
4619 for (s = 0; s < conf->copies*2; s++) {
4620 struct bio *b;
4621 int d = r10_bio->devs[s/2].devnum;
4622 struct md_rdev *rdev;
4623 rcu_read_lock();
4624 if (s&1) {
4625 rdev = rcu_dereference(conf->mirrors[d].replacement);
4626 b = r10_bio->devs[s/2].repl_bio;
4627 } else {
4628 rdev = rcu_dereference(conf->mirrors[d].rdev);
4629 b = r10_bio->devs[s/2].bio;
4630 }
4631 if (!rdev || test_bit(Faulty, &rdev->flags)) {
4632 rcu_read_unlock();
4633 continue;
4634 }
4635 atomic_inc(&rdev->nr_pending);
4636 rcu_read_unlock();
4637 md_sync_acct(b->bi_bdev, r10_bio->sectors);
4638 atomic_inc(&r10_bio->remaining);
4639 b->bi_next = NULL;
4640 generic_make_request(b);
4641 }
4642 end_reshape_request(r10_bio);
4643}
4644
4645static void end_reshape(struct r10conf *conf)
4646{
4647 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4648 return;
4649
4650 spin_lock_irq(&conf->device_lock);
4651 conf->prev = conf->geo;
4652 md_finish_reshape(conf->mddev);
4653 smp_wmb();
4654 conf->reshape_progress = MaxSector;
4655 conf->reshape_safe = MaxSector;
4656 spin_unlock_irq(&conf->device_lock);
4657
4658 /* read-ahead size must cover two whole stripes, which is
4659 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4660 */
4661 if (conf->mddev->queue) {
4662 int stripe = conf->geo.raid_disks *
4663 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4664 stripe /= conf->geo.near_copies;
4665 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4666 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4667 }
4668 conf->fullsync = 0;
4669}
4670
4671static int handle_reshape_read_error(struct mddev *mddev,
4672 struct r10bio *r10_bio)
4673{
4674 /* Use sync reads to get the blocks from somewhere else */
4675 int sectors = r10_bio->sectors;
4676 struct r10conf *conf = mddev->private;
4677 struct {
4678 struct r10bio r10_bio;
4679 struct r10dev devs[conf->copies];
4680 } on_stack;
4681 struct r10bio *r10b = &on_stack.r10_bio;
4682 int slot = 0;
4683 int idx = 0;
4684 struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
4685
4686 r10b->sector = r10_bio->sector;
4687 __raid10_find_phys(&conf->prev, r10b);
4688
4689 while (sectors) {
4690 int s = sectors;
4691 int success = 0;
4692 int first_slot = slot;
4693
4694 if (s > (PAGE_SIZE >> 9))
4695 s = PAGE_SIZE >> 9;
4696
4697 rcu_read_lock();
4698 while (!success) {
4699 int d = r10b->devs[slot].devnum;
4700 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4701 sector_t addr;
4702 if (rdev == NULL ||
4703 test_bit(Faulty, &rdev->flags) ||
4704 !test_bit(In_sync, &rdev->flags))
4705 goto failed;
4706
4707 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4708 atomic_inc(&rdev->nr_pending);
4709 rcu_read_unlock();
4710 success = sync_page_io(rdev,
4711 addr,
4712 s << 9,
4713 bvec[idx].bv_page,
4714 REQ_OP_READ, 0, false);
4715 rdev_dec_pending(rdev, mddev);
4716 rcu_read_lock();
4717 if (success)
4718 break;
4719 failed:
4720 slot++;
4721 if (slot >= conf->copies)
4722 slot = 0;
4723 if (slot == first_slot)
4724 break;
4725 }
4726 rcu_read_unlock();
4727 if (!success) {
4728 /* couldn't read this block, must give up */
4729 set_bit(MD_RECOVERY_INTR,
4730 &mddev->recovery);
4731 return -EIO;
4732 }
4733 sectors -= s;
4734 idx++;
4735 }
4736 return 0;
4737}
4738
4739static void end_reshape_write(struct bio *bio)
4740{
4741 struct r10bio *r10_bio = bio->bi_private;
4742 struct mddev *mddev = r10_bio->mddev;
4743 struct r10conf *conf = mddev->private;
4744 int d;
4745 int slot;
4746 int repl;
4747 struct md_rdev *rdev = NULL;
4748
4749 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4750 if (repl)
4751 rdev = conf->mirrors[d].replacement;
4752 if (!rdev) {
4753 smp_mb();
4754 rdev = conf->mirrors[d].rdev;
4755 }
4756
4757 if (bio->bi_error) {
4758 /* FIXME should record badblock */
4759 md_error(mddev, rdev);
4760 }
4761
4762 rdev_dec_pending(rdev, mddev);
4763 end_reshape_request(r10_bio);
4764}
4765
4766static void end_reshape_request(struct r10bio *r10_bio)
4767{
4768 if (!atomic_dec_and_test(&r10_bio->remaining))
4769 return;
4770 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4771 bio_put(r10_bio->master_bio);
4772 put_buf(r10_bio);
4773}
4774
4775static void raid10_finish_reshape(struct mddev *mddev)
4776{
4777 struct r10conf *conf = mddev->private;
4778
4779 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4780 return;
4781
4782 if (mddev->delta_disks > 0) {
4783 sector_t size = raid10_size(mddev, 0, 0);
4784 md_set_array_sectors(mddev, size);
4785 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4786 mddev->recovery_cp = mddev->resync_max_sectors;
4787 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4788 }
4789 mddev->resync_max_sectors = size;
4790 if (mddev->queue) {
4791 set_capacity(mddev->gendisk, mddev->array_sectors);
4792 revalidate_disk(mddev->gendisk);
4793 }
4794 } else {
4795 int d;
4796 rcu_read_lock();
4797 for (d = conf->geo.raid_disks ;
4798 d < conf->geo.raid_disks - mddev->delta_disks;
4799 d++) {
4800 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4801 if (rdev)
4802 clear_bit(In_sync, &rdev->flags);
4803 rdev = rcu_dereference(conf->mirrors[d].replacement);
4804 if (rdev)
4805 clear_bit(In_sync, &rdev->flags);
4806 }
4807 rcu_read_unlock();
4808 }
4809 mddev->layout = mddev->new_layout;
4810 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4811 mddev->reshape_position = MaxSector;
4812 mddev->delta_disks = 0;
4813 mddev->reshape_backwards = 0;
4814}
4815
4816static struct md_personality raid10_personality =
4817{
4818 .name = "raid10",
4819 .level = 10,
4820 .owner = THIS_MODULE,
4821 .make_request = raid10_make_request,
4822 .run = raid10_run,
4823 .free = raid10_free,
4824 .status = raid10_status,
4825 .error_handler = raid10_error,
4826 .hot_add_disk = raid10_add_disk,
4827 .hot_remove_disk= raid10_remove_disk,
4828 .spare_active = raid10_spare_active,
4829 .sync_request = raid10_sync_request,
4830 .quiesce = raid10_quiesce,
4831 .size = raid10_size,
4832 .resize = raid10_resize,
4833 .takeover = raid10_takeover,
4834 .check_reshape = raid10_check_reshape,
4835 .start_reshape = raid10_start_reshape,
4836 .finish_reshape = raid10_finish_reshape,
4837 .congested = raid10_congested,
4838};
4839
4840static int __init raid_init(void)
4841{
4842 return register_md_personality(&raid10_personality);
4843}
4844
4845static void raid_exit(void)
4846{
4847 unregister_md_personality(&raid10_personality);
4848}
4849
4850module_init(raid_init);
4851module_exit(raid_exit);
4852MODULE_LICENSE("GPL");
4853MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4854MODULE_ALIAS("md-personality-9"); /* RAID10 */
4855MODULE_ALIAS("md-raid10");
4856MODULE_ALIAS("md-level-10");
4857
4858module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);