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/seq_file.h>
25#include <linux/ratelimit.h>
26#include "md.h"
27#include "raid10.h"
28#include "raid0.h"
29#include "bitmap.h"
30
31/*
32 * RAID10 provides a combination of RAID0 and RAID1 functionality.
33 * The layout of data is defined by
34 * chunk_size
35 * raid_disks
36 * near_copies (stored in low byte of layout)
37 * far_copies (stored in second byte of layout)
38 * far_offset (stored in bit 16 of layout )
39 *
40 * The data to be stored is divided into chunks using chunksize.
41 * Each device is divided into far_copies sections.
42 * In each section, chunks are laid out in a style similar to raid0, but
43 * near_copies copies of each chunk is stored (each on a different drive).
44 * The starting device for each section is offset near_copies from the starting
45 * device of the previous section.
46 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
47 * drive.
48 * near_copies and far_copies must be at least one, and their product is at most
49 * raid_disks.
50 *
51 * If far_offset is true, then the far_copies are handled a bit differently.
52 * The copies are still in different stripes, but instead of be very far apart
53 * on disk, there are adjacent stripes.
54 */
55
56/*
57 * Number of guaranteed r10bios in case of extreme VM load:
58 */
59#define NR_RAID10_BIOS 256
60
61static void allow_barrier(conf_t *conf);
62static void lower_barrier(conf_t *conf);
63
64static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
65{
66 conf_t *conf = data;
67 int size = offsetof(struct r10bio_s, devs[conf->copies]);
68
69 /* allocate a r10bio with room for raid_disks entries in the bios array */
70 return kzalloc(size, gfp_flags);
71}
72
73static void r10bio_pool_free(void *r10_bio, void *data)
74{
75 kfree(r10_bio);
76}
77
78/* Maximum size of each resync request */
79#define RESYNC_BLOCK_SIZE (64*1024)
80#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
81/* amount of memory to reserve for resync requests */
82#define RESYNC_WINDOW (1024*1024)
83/* maximum number of concurrent requests, memory permitting */
84#define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
85
86/*
87 * When performing a resync, we need to read and compare, so
88 * we need as many pages are there are copies.
89 * When performing a recovery, we need 2 bios, one for read,
90 * one for write (we recover only one drive per r10buf)
91 *
92 */
93static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
94{
95 conf_t *conf = data;
96 struct page *page;
97 r10bio_t *r10_bio;
98 struct bio *bio;
99 int i, j;
100 int nalloc;
101
102 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
103 if (!r10_bio)
104 return NULL;
105
106 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
107 nalloc = conf->copies; /* resync */
108 else
109 nalloc = 2; /* recovery */
110
111 /*
112 * Allocate bios.
113 */
114 for (j = nalloc ; j-- ; ) {
115 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
116 if (!bio)
117 goto out_free_bio;
118 r10_bio->devs[j].bio = bio;
119 }
120 /*
121 * Allocate RESYNC_PAGES data pages and attach them
122 * where needed.
123 */
124 for (j = 0 ; j < nalloc; j++) {
125 bio = r10_bio->devs[j].bio;
126 for (i = 0; i < RESYNC_PAGES; i++) {
127 if (j == 1 && !test_bit(MD_RECOVERY_SYNC,
128 &conf->mddev->recovery)) {
129 /* we can share bv_page's during recovery */
130 struct bio *rbio = r10_bio->devs[0].bio;
131 page = rbio->bi_io_vec[i].bv_page;
132 get_page(page);
133 } else
134 page = alloc_page(gfp_flags);
135 if (unlikely(!page))
136 goto out_free_pages;
137
138 bio->bi_io_vec[i].bv_page = page;
139 }
140 }
141
142 return r10_bio;
143
144out_free_pages:
145 for ( ; i > 0 ; i--)
146 safe_put_page(bio->bi_io_vec[i-1].bv_page);
147 while (j--)
148 for (i = 0; i < RESYNC_PAGES ; i++)
149 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
150 j = -1;
151out_free_bio:
152 while ( ++j < nalloc )
153 bio_put(r10_bio->devs[j].bio);
154 r10bio_pool_free(r10_bio, conf);
155 return NULL;
156}
157
158static void r10buf_pool_free(void *__r10_bio, void *data)
159{
160 int i;
161 conf_t *conf = data;
162 r10bio_t *r10bio = __r10_bio;
163 int j;
164
165 for (j=0; j < conf->copies; j++) {
166 struct bio *bio = r10bio->devs[j].bio;
167 if (bio) {
168 for (i = 0; i < RESYNC_PAGES; i++) {
169 safe_put_page(bio->bi_io_vec[i].bv_page);
170 bio->bi_io_vec[i].bv_page = NULL;
171 }
172 bio_put(bio);
173 }
174 }
175 r10bio_pool_free(r10bio, conf);
176}
177
178static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
179{
180 int i;
181
182 for (i = 0; i < conf->copies; i++) {
183 struct bio **bio = & r10_bio->devs[i].bio;
184 if (!BIO_SPECIAL(*bio))
185 bio_put(*bio);
186 *bio = NULL;
187 }
188}
189
190static void free_r10bio(r10bio_t *r10_bio)
191{
192 conf_t *conf = r10_bio->mddev->private;
193
194 put_all_bios(conf, r10_bio);
195 mempool_free(r10_bio, conf->r10bio_pool);
196}
197
198static void put_buf(r10bio_t *r10_bio)
199{
200 conf_t *conf = r10_bio->mddev->private;
201
202 mempool_free(r10_bio, conf->r10buf_pool);
203
204 lower_barrier(conf);
205}
206
207static void reschedule_retry(r10bio_t *r10_bio)
208{
209 unsigned long flags;
210 mddev_t *mddev = r10_bio->mddev;
211 conf_t *conf = mddev->private;
212
213 spin_lock_irqsave(&conf->device_lock, flags);
214 list_add(&r10_bio->retry_list, &conf->retry_list);
215 conf->nr_queued ++;
216 spin_unlock_irqrestore(&conf->device_lock, flags);
217
218 /* wake up frozen array... */
219 wake_up(&conf->wait_barrier);
220
221 md_wakeup_thread(mddev->thread);
222}
223
224/*
225 * raid_end_bio_io() is called when we have finished servicing a mirrored
226 * operation and are ready to return a success/failure code to the buffer
227 * cache layer.
228 */
229static void raid_end_bio_io(r10bio_t *r10_bio)
230{
231 struct bio *bio = r10_bio->master_bio;
232 int done;
233 conf_t *conf = r10_bio->mddev->private;
234
235 if (bio->bi_phys_segments) {
236 unsigned long flags;
237 spin_lock_irqsave(&conf->device_lock, flags);
238 bio->bi_phys_segments--;
239 done = (bio->bi_phys_segments == 0);
240 spin_unlock_irqrestore(&conf->device_lock, flags);
241 } else
242 done = 1;
243 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
244 clear_bit(BIO_UPTODATE, &bio->bi_flags);
245 if (done) {
246 bio_endio(bio, 0);
247 /*
248 * Wake up any possible resync thread that waits for the device
249 * to go idle.
250 */
251 allow_barrier(conf);
252 }
253 free_r10bio(r10_bio);
254}
255
256/*
257 * Update disk head position estimator based on IRQ completion info.
258 */
259static inline void update_head_pos(int slot, r10bio_t *r10_bio)
260{
261 conf_t *conf = r10_bio->mddev->private;
262
263 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
264 r10_bio->devs[slot].addr + (r10_bio->sectors);
265}
266
267/*
268 * Find the disk number which triggered given bio
269 */
270static int find_bio_disk(conf_t *conf, r10bio_t *r10_bio,
271 struct bio *bio, int *slotp)
272{
273 int slot;
274
275 for (slot = 0; slot < conf->copies; slot++)
276 if (r10_bio->devs[slot].bio == bio)
277 break;
278
279 BUG_ON(slot == conf->copies);
280 update_head_pos(slot, r10_bio);
281
282 if (slotp)
283 *slotp = slot;
284 return r10_bio->devs[slot].devnum;
285}
286
287static void raid10_end_read_request(struct bio *bio, int error)
288{
289 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
290 r10bio_t *r10_bio = bio->bi_private;
291 int slot, dev;
292 conf_t *conf = r10_bio->mddev->private;
293
294
295 slot = r10_bio->read_slot;
296 dev = r10_bio->devs[slot].devnum;
297 /*
298 * this branch is our 'one mirror IO has finished' event handler:
299 */
300 update_head_pos(slot, r10_bio);
301
302 if (uptodate) {
303 /*
304 * Set R10BIO_Uptodate in our master bio, so that
305 * we will return a good error code to the higher
306 * levels even if IO on some other mirrored buffer fails.
307 *
308 * The 'master' represents the composite IO operation to
309 * user-side. So if something waits for IO, then it will
310 * wait for the 'master' bio.
311 */
312 set_bit(R10BIO_Uptodate, &r10_bio->state);
313 raid_end_bio_io(r10_bio);
314 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
315 } else {
316 /*
317 * oops, read error - keep the refcount on the rdev
318 */
319 char b[BDEVNAME_SIZE];
320 printk_ratelimited(KERN_ERR
321 "md/raid10:%s: %s: rescheduling sector %llu\n",
322 mdname(conf->mddev),
323 bdevname(conf->mirrors[dev].rdev->bdev, b),
324 (unsigned long long)r10_bio->sector);
325 set_bit(R10BIO_ReadError, &r10_bio->state);
326 reschedule_retry(r10_bio);
327 }
328}
329
330static void close_write(r10bio_t *r10_bio)
331{
332 /* clear the bitmap if all writes complete successfully */
333 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
334 r10_bio->sectors,
335 !test_bit(R10BIO_Degraded, &r10_bio->state),
336 0);
337 md_write_end(r10_bio->mddev);
338}
339
340static void one_write_done(r10bio_t *r10_bio)
341{
342 if (atomic_dec_and_test(&r10_bio->remaining)) {
343 if (test_bit(R10BIO_WriteError, &r10_bio->state))
344 reschedule_retry(r10_bio);
345 else {
346 close_write(r10_bio);
347 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
348 reschedule_retry(r10_bio);
349 else
350 raid_end_bio_io(r10_bio);
351 }
352 }
353}
354
355static void raid10_end_write_request(struct bio *bio, int error)
356{
357 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
358 r10bio_t *r10_bio = bio->bi_private;
359 int dev;
360 int dec_rdev = 1;
361 conf_t *conf = r10_bio->mddev->private;
362 int slot;
363
364 dev = find_bio_disk(conf, r10_bio, bio, &slot);
365
366 /*
367 * this branch is our 'one mirror IO has finished' event handler:
368 */
369 if (!uptodate) {
370 set_bit(WriteErrorSeen, &conf->mirrors[dev].rdev->flags);
371 set_bit(R10BIO_WriteError, &r10_bio->state);
372 dec_rdev = 0;
373 } else {
374 /*
375 * Set R10BIO_Uptodate in our master bio, so that
376 * we will return a good error code for to the higher
377 * levels even if IO on some other mirrored buffer fails.
378 *
379 * The 'master' represents the composite IO operation to
380 * user-side. So if something waits for IO, then it will
381 * wait for the 'master' bio.
382 */
383 sector_t first_bad;
384 int bad_sectors;
385
386 set_bit(R10BIO_Uptodate, &r10_bio->state);
387
388 /* Maybe we can clear some bad blocks. */
389 if (is_badblock(conf->mirrors[dev].rdev,
390 r10_bio->devs[slot].addr,
391 r10_bio->sectors,
392 &first_bad, &bad_sectors)) {
393 bio_put(bio);
394 r10_bio->devs[slot].bio = IO_MADE_GOOD;
395 dec_rdev = 0;
396 set_bit(R10BIO_MadeGood, &r10_bio->state);
397 }
398 }
399
400 /*
401 *
402 * Let's see if all mirrored write operations have finished
403 * already.
404 */
405 one_write_done(r10_bio);
406 if (dec_rdev)
407 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
408}
409
410
411/*
412 * RAID10 layout manager
413 * As well as the chunksize and raid_disks count, there are two
414 * parameters: near_copies and far_copies.
415 * near_copies * far_copies must be <= raid_disks.
416 * Normally one of these will be 1.
417 * If both are 1, we get raid0.
418 * If near_copies == raid_disks, we get raid1.
419 *
420 * Chunks are laid out in raid0 style with near_copies copies of the
421 * first chunk, followed by near_copies copies of the next chunk and
422 * so on.
423 * If far_copies > 1, then after 1/far_copies of the array has been assigned
424 * as described above, we start again with a device offset of near_copies.
425 * So we effectively have another copy of the whole array further down all
426 * the drives, but with blocks on different drives.
427 * With this layout, and block is never stored twice on the one device.
428 *
429 * raid10_find_phys finds the sector offset of a given virtual sector
430 * on each device that it is on.
431 *
432 * raid10_find_virt does the reverse mapping, from a device and a
433 * sector offset to a virtual address
434 */
435
436static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
437{
438 int n,f;
439 sector_t sector;
440 sector_t chunk;
441 sector_t stripe;
442 int dev;
443
444 int slot = 0;
445
446 /* now calculate first sector/dev */
447 chunk = r10bio->sector >> conf->chunk_shift;
448 sector = r10bio->sector & conf->chunk_mask;
449
450 chunk *= conf->near_copies;
451 stripe = chunk;
452 dev = sector_div(stripe, conf->raid_disks);
453 if (conf->far_offset)
454 stripe *= conf->far_copies;
455
456 sector += stripe << conf->chunk_shift;
457
458 /* and calculate all the others */
459 for (n=0; n < conf->near_copies; n++) {
460 int d = dev;
461 sector_t s = sector;
462 r10bio->devs[slot].addr = sector;
463 r10bio->devs[slot].devnum = d;
464 slot++;
465
466 for (f = 1; f < conf->far_copies; f++) {
467 d += conf->near_copies;
468 if (d >= conf->raid_disks)
469 d -= conf->raid_disks;
470 s += conf->stride;
471 r10bio->devs[slot].devnum = d;
472 r10bio->devs[slot].addr = s;
473 slot++;
474 }
475 dev++;
476 if (dev >= conf->raid_disks) {
477 dev = 0;
478 sector += (conf->chunk_mask + 1);
479 }
480 }
481 BUG_ON(slot != conf->copies);
482}
483
484static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
485{
486 sector_t offset, chunk, vchunk;
487
488 offset = sector & conf->chunk_mask;
489 if (conf->far_offset) {
490 int fc;
491 chunk = sector >> conf->chunk_shift;
492 fc = sector_div(chunk, conf->far_copies);
493 dev -= fc * conf->near_copies;
494 if (dev < 0)
495 dev += conf->raid_disks;
496 } else {
497 while (sector >= conf->stride) {
498 sector -= conf->stride;
499 if (dev < conf->near_copies)
500 dev += conf->raid_disks - conf->near_copies;
501 else
502 dev -= conf->near_copies;
503 }
504 chunk = sector >> conf->chunk_shift;
505 }
506 vchunk = chunk * conf->raid_disks + dev;
507 sector_div(vchunk, conf->near_copies);
508 return (vchunk << conf->chunk_shift) + offset;
509}
510
511/**
512 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
513 * @q: request queue
514 * @bvm: properties of new bio
515 * @biovec: the request that could be merged to it.
516 *
517 * Return amount of bytes we can accept at this offset
518 * If near_copies == raid_disk, there are no striping issues,
519 * but in that case, the function isn't called at all.
520 */
521static int raid10_mergeable_bvec(struct request_queue *q,
522 struct bvec_merge_data *bvm,
523 struct bio_vec *biovec)
524{
525 mddev_t *mddev = q->queuedata;
526 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
527 int max;
528 unsigned int chunk_sectors = mddev->chunk_sectors;
529 unsigned int bio_sectors = bvm->bi_size >> 9;
530
531 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
532 if (max < 0) max = 0; /* bio_add cannot handle a negative return */
533 if (max <= biovec->bv_len && bio_sectors == 0)
534 return biovec->bv_len;
535 else
536 return max;
537}
538
539/*
540 * This routine returns the disk from which the requested read should
541 * be done. There is a per-array 'next expected sequential IO' sector
542 * number - if this matches on the next IO then we use the last disk.
543 * There is also a per-disk 'last know head position' sector that is
544 * maintained from IRQ contexts, both the normal and the resync IO
545 * completion handlers update this position correctly. If there is no
546 * perfect sequential match then we pick the disk whose head is closest.
547 *
548 * If there are 2 mirrors in the same 2 devices, performance degrades
549 * because position is mirror, not device based.
550 *
551 * The rdev for the device selected will have nr_pending incremented.
552 */
553
554/*
555 * FIXME: possibly should rethink readbalancing and do it differently
556 * depending on near_copies / far_copies geometry.
557 */
558static int read_balance(conf_t *conf, r10bio_t *r10_bio, int *max_sectors)
559{
560 const sector_t this_sector = r10_bio->sector;
561 int disk, slot;
562 int sectors = r10_bio->sectors;
563 int best_good_sectors;
564 sector_t new_distance, best_dist;
565 mdk_rdev_t *rdev;
566 int do_balance;
567 int best_slot;
568
569 raid10_find_phys(conf, r10_bio);
570 rcu_read_lock();
571retry:
572 sectors = r10_bio->sectors;
573 best_slot = -1;
574 best_dist = MaxSector;
575 best_good_sectors = 0;
576 do_balance = 1;
577 /*
578 * Check if we can balance. We can balance on the whole
579 * device if no resync is going on (recovery is ok), or below
580 * the resync window. We take the first readable disk when
581 * above the resync window.
582 */
583 if (conf->mddev->recovery_cp < MaxSector
584 && (this_sector + sectors >= conf->next_resync))
585 do_balance = 0;
586
587 for (slot = 0; slot < conf->copies ; slot++) {
588 sector_t first_bad;
589 int bad_sectors;
590 sector_t dev_sector;
591
592 if (r10_bio->devs[slot].bio == IO_BLOCKED)
593 continue;
594 disk = r10_bio->devs[slot].devnum;
595 rdev = rcu_dereference(conf->mirrors[disk].rdev);
596 if (rdev == NULL)
597 continue;
598 if (!test_bit(In_sync, &rdev->flags))
599 continue;
600
601 dev_sector = r10_bio->devs[slot].addr;
602 if (is_badblock(rdev, dev_sector, sectors,
603 &first_bad, &bad_sectors)) {
604 if (best_dist < MaxSector)
605 /* Already have a better slot */
606 continue;
607 if (first_bad <= dev_sector) {
608 /* Cannot read here. If this is the
609 * 'primary' device, then we must not read
610 * beyond 'bad_sectors' from another device.
611 */
612 bad_sectors -= (dev_sector - first_bad);
613 if (!do_balance && sectors > bad_sectors)
614 sectors = bad_sectors;
615 if (best_good_sectors > sectors)
616 best_good_sectors = sectors;
617 } else {
618 sector_t good_sectors =
619 first_bad - dev_sector;
620 if (good_sectors > best_good_sectors) {
621 best_good_sectors = good_sectors;
622 best_slot = slot;
623 }
624 if (!do_balance)
625 /* Must read from here */
626 break;
627 }
628 continue;
629 } else
630 best_good_sectors = sectors;
631
632 if (!do_balance)
633 break;
634
635 /* This optimisation is debatable, and completely destroys
636 * sequential read speed for 'far copies' arrays. So only
637 * keep it for 'near' arrays, and review those later.
638 */
639 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending))
640 break;
641
642 /* for far > 1 always use the lowest address */
643 if (conf->far_copies > 1)
644 new_distance = r10_bio->devs[slot].addr;
645 else
646 new_distance = abs(r10_bio->devs[slot].addr -
647 conf->mirrors[disk].head_position);
648 if (new_distance < best_dist) {
649 best_dist = new_distance;
650 best_slot = slot;
651 }
652 }
653 if (slot == conf->copies)
654 slot = best_slot;
655
656 if (slot >= 0) {
657 disk = r10_bio->devs[slot].devnum;
658 rdev = rcu_dereference(conf->mirrors[disk].rdev);
659 if (!rdev)
660 goto retry;
661 atomic_inc(&rdev->nr_pending);
662 if (test_bit(Faulty, &rdev->flags)) {
663 /* Cannot risk returning a device that failed
664 * before we inc'ed nr_pending
665 */
666 rdev_dec_pending(rdev, conf->mddev);
667 goto retry;
668 }
669 r10_bio->read_slot = slot;
670 } else
671 disk = -1;
672 rcu_read_unlock();
673 *max_sectors = best_good_sectors;
674
675 return disk;
676}
677
678static int raid10_congested(void *data, int bits)
679{
680 mddev_t *mddev = data;
681 conf_t *conf = mddev->private;
682 int i, ret = 0;
683
684 if (mddev_congested(mddev, bits))
685 return 1;
686 rcu_read_lock();
687 for (i = 0; i < conf->raid_disks && ret == 0; i++) {
688 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
689 if (rdev && !test_bit(Faulty, &rdev->flags)) {
690 struct request_queue *q = bdev_get_queue(rdev->bdev);
691
692 ret |= bdi_congested(&q->backing_dev_info, bits);
693 }
694 }
695 rcu_read_unlock();
696 return ret;
697}
698
699static void flush_pending_writes(conf_t *conf)
700{
701 /* Any writes that have been queued but are awaiting
702 * bitmap updates get flushed here.
703 */
704 spin_lock_irq(&conf->device_lock);
705
706 if (conf->pending_bio_list.head) {
707 struct bio *bio;
708 bio = bio_list_get(&conf->pending_bio_list);
709 spin_unlock_irq(&conf->device_lock);
710 /* flush any pending bitmap writes to disk
711 * before proceeding w/ I/O */
712 bitmap_unplug(conf->mddev->bitmap);
713
714 while (bio) { /* submit pending writes */
715 struct bio *next = bio->bi_next;
716 bio->bi_next = NULL;
717 generic_make_request(bio);
718 bio = next;
719 }
720 } else
721 spin_unlock_irq(&conf->device_lock);
722}
723
724/* Barriers....
725 * Sometimes we need to suspend IO while we do something else,
726 * either some resync/recovery, or reconfigure the array.
727 * To do this we raise a 'barrier'.
728 * The 'barrier' is a counter that can be raised multiple times
729 * to count how many activities are happening which preclude
730 * normal IO.
731 * We can only raise the barrier if there is no pending IO.
732 * i.e. if nr_pending == 0.
733 * We choose only to raise the barrier if no-one is waiting for the
734 * barrier to go down. This means that as soon as an IO request
735 * is ready, no other operations which require a barrier will start
736 * until the IO request has had a chance.
737 *
738 * So: regular IO calls 'wait_barrier'. When that returns there
739 * is no backgroup IO happening, It must arrange to call
740 * allow_barrier when it has finished its IO.
741 * backgroup IO calls must call raise_barrier. Once that returns
742 * there is no normal IO happeing. It must arrange to call
743 * lower_barrier when the particular background IO completes.
744 */
745
746static void raise_barrier(conf_t *conf, int force)
747{
748 BUG_ON(force && !conf->barrier);
749 spin_lock_irq(&conf->resync_lock);
750
751 /* Wait until no block IO is waiting (unless 'force') */
752 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
753 conf->resync_lock, );
754
755 /* block any new IO from starting */
756 conf->barrier++;
757
758 /* Now wait for all pending IO to complete */
759 wait_event_lock_irq(conf->wait_barrier,
760 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
761 conf->resync_lock, );
762
763 spin_unlock_irq(&conf->resync_lock);
764}
765
766static void lower_barrier(conf_t *conf)
767{
768 unsigned long flags;
769 spin_lock_irqsave(&conf->resync_lock, flags);
770 conf->barrier--;
771 spin_unlock_irqrestore(&conf->resync_lock, flags);
772 wake_up(&conf->wait_barrier);
773}
774
775static void wait_barrier(conf_t *conf)
776{
777 spin_lock_irq(&conf->resync_lock);
778 if (conf->barrier) {
779 conf->nr_waiting++;
780 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
781 conf->resync_lock,
782 );
783 conf->nr_waiting--;
784 }
785 conf->nr_pending++;
786 spin_unlock_irq(&conf->resync_lock);
787}
788
789static void allow_barrier(conf_t *conf)
790{
791 unsigned long flags;
792 spin_lock_irqsave(&conf->resync_lock, flags);
793 conf->nr_pending--;
794 spin_unlock_irqrestore(&conf->resync_lock, flags);
795 wake_up(&conf->wait_barrier);
796}
797
798static void freeze_array(conf_t *conf)
799{
800 /* stop syncio and normal IO and wait for everything to
801 * go quiet.
802 * We increment barrier and nr_waiting, and then
803 * wait until nr_pending match nr_queued+1
804 * This is called in the context of one normal IO request
805 * that has failed. Thus any sync request that might be pending
806 * will be blocked by nr_pending, and we need to wait for
807 * pending IO requests to complete or be queued for re-try.
808 * Thus the number queued (nr_queued) plus this request (1)
809 * must match the number of pending IOs (nr_pending) before
810 * we continue.
811 */
812 spin_lock_irq(&conf->resync_lock);
813 conf->barrier++;
814 conf->nr_waiting++;
815 wait_event_lock_irq(conf->wait_barrier,
816 conf->nr_pending == conf->nr_queued+1,
817 conf->resync_lock,
818 flush_pending_writes(conf));
819
820 spin_unlock_irq(&conf->resync_lock);
821}
822
823static void unfreeze_array(conf_t *conf)
824{
825 /* reverse the effect of the freeze */
826 spin_lock_irq(&conf->resync_lock);
827 conf->barrier--;
828 conf->nr_waiting--;
829 wake_up(&conf->wait_barrier);
830 spin_unlock_irq(&conf->resync_lock);
831}
832
833static int make_request(mddev_t *mddev, struct bio * bio)
834{
835 conf_t *conf = mddev->private;
836 mirror_info_t *mirror;
837 r10bio_t *r10_bio;
838 struct bio *read_bio;
839 int i;
840 int chunk_sects = conf->chunk_mask + 1;
841 const int rw = bio_data_dir(bio);
842 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
843 const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
844 unsigned long flags;
845 mdk_rdev_t *blocked_rdev;
846 int plugged;
847 int sectors_handled;
848 int max_sectors;
849
850 if (unlikely(bio->bi_rw & REQ_FLUSH)) {
851 md_flush_request(mddev, bio);
852 return 0;
853 }
854
855 /* If this request crosses a chunk boundary, we need to
856 * split it. This will only happen for 1 PAGE (or less) requests.
857 */
858 if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
859 > chunk_sects &&
860 conf->near_copies < conf->raid_disks)) {
861 struct bio_pair *bp;
862 /* Sanity check -- queue functions should prevent this happening */
863 if (bio->bi_vcnt != 1 ||
864 bio->bi_idx != 0)
865 goto bad_map;
866 /* This is a one page bio that upper layers
867 * refuse to split for us, so we need to split it.
868 */
869 bp = bio_split(bio,
870 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
871
872 /* Each of these 'make_request' calls will call 'wait_barrier'.
873 * If the first succeeds but the second blocks due to the resync
874 * thread raising the barrier, we will deadlock because the
875 * IO to the underlying device will be queued in generic_make_request
876 * and will never complete, so will never reduce nr_pending.
877 * So increment nr_waiting here so no new raise_barriers will
878 * succeed, and so the second wait_barrier cannot block.
879 */
880 spin_lock_irq(&conf->resync_lock);
881 conf->nr_waiting++;
882 spin_unlock_irq(&conf->resync_lock);
883
884 if (make_request(mddev, &bp->bio1))
885 generic_make_request(&bp->bio1);
886 if (make_request(mddev, &bp->bio2))
887 generic_make_request(&bp->bio2);
888
889 spin_lock_irq(&conf->resync_lock);
890 conf->nr_waiting--;
891 wake_up(&conf->wait_barrier);
892 spin_unlock_irq(&conf->resync_lock);
893
894 bio_pair_release(bp);
895 return 0;
896 bad_map:
897 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
898 " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
899 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
900
901 bio_io_error(bio);
902 return 0;
903 }
904
905 md_write_start(mddev, bio);
906
907 /*
908 * Register the new request and wait if the reconstruction
909 * thread has put up a bar for new requests.
910 * Continue immediately if no resync is active currently.
911 */
912 wait_barrier(conf);
913
914 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
915
916 r10_bio->master_bio = bio;
917 r10_bio->sectors = bio->bi_size >> 9;
918
919 r10_bio->mddev = mddev;
920 r10_bio->sector = bio->bi_sector;
921 r10_bio->state = 0;
922
923 /* We might need to issue multiple reads to different
924 * devices if there are bad blocks around, so we keep
925 * track of the number of reads in bio->bi_phys_segments.
926 * If this is 0, there is only one r10_bio and no locking
927 * will be needed when the request completes. If it is
928 * non-zero, then it is the number of not-completed requests.
929 */
930 bio->bi_phys_segments = 0;
931 clear_bit(BIO_SEG_VALID, &bio->bi_flags);
932
933 if (rw == READ) {
934 /*
935 * read balancing logic:
936 */
937 int disk;
938 int slot;
939
940read_again:
941 disk = read_balance(conf, r10_bio, &max_sectors);
942 slot = r10_bio->read_slot;
943 if (disk < 0) {
944 raid_end_bio_io(r10_bio);
945 return 0;
946 }
947 mirror = conf->mirrors + disk;
948
949 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
950 md_trim_bio(read_bio, r10_bio->sector - bio->bi_sector,
951 max_sectors);
952
953 r10_bio->devs[slot].bio = read_bio;
954
955 read_bio->bi_sector = r10_bio->devs[slot].addr +
956 mirror->rdev->data_offset;
957 read_bio->bi_bdev = mirror->rdev->bdev;
958 read_bio->bi_end_io = raid10_end_read_request;
959 read_bio->bi_rw = READ | do_sync;
960 read_bio->bi_private = r10_bio;
961
962 if (max_sectors < r10_bio->sectors) {
963 /* Could not read all from this device, so we will
964 * need another r10_bio.
965 */
966 sectors_handled = (r10_bio->sectors + max_sectors
967 - bio->bi_sector);
968 r10_bio->sectors = max_sectors;
969 spin_lock_irq(&conf->device_lock);
970 if (bio->bi_phys_segments == 0)
971 bio->bi_phys_segments = 2;
972 else
973 bio->bi_phys_segments++;
974 spin_unlock(&conf->device_lock);
975 /* Cannot call generic_make_request directly
976 * as that will be queued in __generic_make_request
977 * and subsequent mempool_alloc might block
978 * waiting for it. so hand bio over to raid10d.
979 */
980 reschedule_retry(r10_bio);
981
982 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
983
984 r10_bio->master_bio = bio;
985 r10_bio->sectors = ((bio->bi_size >> 9)
986 - sectors_handled);
987 r10_bio->state = 0;
988 r10_bio->mddev = mddev;
989 r10_bio->sector = bio->bi_sector + sectors_handled;
990 goto read_again;
991 } else
992 generic_make_request(read_bio);
993 return 0;
994 }
995
996 /*
997 * WRITE:
998 */
999 /* first select target devices under rcu_lock and
1000 * inc refcount on their rdev. Record them by setting
1001 * bios[x] to bio
1002 * If there are known/acknowledged bad blocks on any device
1003 * on which we have seen a write error, we want to avoid
1004 * writing to those blocks. This potentially requires several
1005 * writes to write around the bad blocks. Each set of writes
1006 * gets its own r10_bio with a set of bios attached. The number
1007 * of r10_bios is recored in bio->bi_phys_segments just as with
1008 * the read case.
1009 */
1010 plugged = mddev_check_plugged(mddev);
1011
1012 raid10_find_phys(conf, r10_bio);
1013retry_write:
1014 blocked_rdev = NULL;
1015 rcu_read_lock();
1016 max_sectors = r10_bio->sectors;
1017
1018 for (i = 0; i < conf->copies; i++) {
1019 int d = r10_bio->devs[i].devnum;
1020 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
1021 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1022 atomic_inc(&rdev->nr_pending);
1023 blocked_rdev = rdev;
1024 break;
1025 }
1026 r10_bio->devs[i].bio = NULL;
1027 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1028 set_bit(R10BIO_Degraded, &r10_bio->state);
1029 continue;
1030 }
1031 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1032 sector_t first_bad;
1033 sector_t dev_sector = r10_bio->devs[i].addr;
1034 int bad_sectors;
1035 int is_bad;
1036
1037 is_bad = is_badblock(rdev, dev_sector,
1038 max_sectors,
1039 &first_bad, &bad_sectors);
1040 if (is_bad < 0) {
1041 /* Mustn't write here until the bad block
1042 * is acknowledged
1043 */
1044 atomic_inc(&rdev->nr_pending);
1045 set_bit(BlockedBadBlocks, &rdev->flags);
1046 blocked_rdev = rdev;
1047 break;
1048 }
1049 if (is_bad && first_bad <= dev_sector) {
1050 /* Cannot write here at all */
1051 bad_sectors -= (dev_sector - first_bad);
1052 if (bad_sectors < max_sectors)
1053 /* Mustn't write more than bad_sectors
1054 * to other devices yet
1055 */
1056 max_sectors = bad_sectors;
1057 /* We don't set R10BIO_Degraded as that
1058 * only applies if the disk is missing,
1059 * so it might be re-added, and we want to
1060 * know to recover this chunk.
1061 * In this case the device is here, and the
1062 * fact that this chunk is not in-sync is
1063 * recorded in the bad block log.
1064 */
1065 continue;
1066 }
1067 if (is_bad) {
1068 int good_sectors = first_bad - dev_sector;
1069 if (good_sectors < max_sectors)
1070 max_sectors = good_sectors;
1071 }
1072 }
1073 r10_bio->devs[i].bio = bio;
1074 atomic_inc(&rdev->nr_pending);
1075 }
1076 rcu_read_unlock();
1077
1078 if (unlikely(blocked_rdev)) {
1079 /* Have to wait for this device to get unblocked, then retry */
1080 int j;
1081 int d;
1082
1083 for (j = 0; j < i; j++)
1084 if (r10_bio->devs[j].bio) {
1085 d = r10_bio->devs[j].devnum;
1086 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1087 }
1088 allow_barrier(conf);
1089 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1090 wait_barrier(conf);
1091 goto retry_write;
1092 }
1093
1094 if (max_sectors < r10_bio->sectors) {
1095 /* We are splitting this into multiple parts, so
1096 * we need to prepare for allocating another r10_bio.
1097 */
1098 r10_bio->sectors = max_sectors;
1099 spin_lock_irq(&conf->device_lock);
1100 if (bio->bi_phys_segments == 0)
1101 bio->bi_phys_segments = 2;
1102 else
1103 bio->bi_phys_segments++;
1104 spin_unlock_irq(&conf->device_lock);
1105 }
1106 sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector;
1107
1108 atomic_set(&r10_bio->remaining, 1);
1109 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1110
1111 for (i = 0; i < conf->copies; i++) {
1112 struct bio *mbio;
1113 int d = r10_bio->devs[i].devnum;
1114 if (!r10_bio->devs[i].bio)
1115 continue;
1116
1117 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1118 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1119 max_sectors);
1120 r10_bio->devs[i].bio = mbio;
1121
1122 mbio->bi_sector = (r10_bio->devs[i].addr+
1123 conf->mirrors[d].rdev->data_offset);
1124 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1125 mbio->bi_end_io = raid10_end_write_request;
1126 mbio->bi_rw = WRITE | do_sync | do_fua;
1127 mbio->bi_private = r10_bio;
1128
1129 atomic_inc(&r10_bio->remaining);
1130 spin_lock_irqsave(&conf->device_lock, flags);
1131 bio_list_add(&conf->pending_bio_list, mbio);
1132 spin_unlock_irqrestore(&conf->device_lock, flags);
1133 }
1134
1135 /* Don't remove the bias on 'remaining' (one_write_done) until
1136 * after checking if we need to go around again.
1137 */
1138
1139 if (sectors_handled < (bio->bi_size >> 9)) {
1140 one_write_done(r10_bio);
1141 /* We need another r10_bio. It has already been counted
1142 * in bio->bi_phys_segments.
1143 */
1144 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1145
1146 r10_bio->master_bio = bio;
1147 r10_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1148
1149 r10_bio->mddev = mddev;
1150 r10_bio->sector = bio->bi_sector + sectors_handled;
1151 r10_bio->state = 0;
1152 goto retry_write;
1153 }
1154 one_write_done(r10_bio);
1155
1156 /* In case raid10d snuck in to freeze_array */
1157 wake_up(&conf->wait_barrier);
1158
1159 if (do_sync || !mddev->bitmap || !plugged)
1160 md_wakeup_thread(mddev->thread);
1161 return 0;
1162}
1163
1164static void status(struct seq_file *seq, mddev_t *mddev)
1165{
1166 conf_t *conf = mddev->private;
1167 int i;
1168
1169 if (conf->near_copies < conf->raid_disks)
1170 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1171 if (conf->near_copies > 1)
1172 seq_printf(seq, " %d near-copies", conf->near_copies);
1173 if (conf->far_copies > 1) {
1174 if (conf->far_offset)
1175 seq_printf(seq, " %d offset-copies", conf->far_copies);
1176 else
1177 seq_printf(seq, " %d far-copies", conf->far_copies);
1178 }
1179 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1180 conf->raid_disks - mddev->degraded);
1181 for (i = 0; i < conf->raid_disks; i++)
1182 seq_printf(seq, "%s",
1183 conf->mirrors[i].rdev &&
1184 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1185 seq_printf(seq, "]");
1186}
1187
1188/* check if there are enough drives for
1189 * every block to appear on atleast one.
1190 * Don't consider the device numbered 'ignore'
1191 * as we might be about to remove it.
1192 */
1193static int enough(conf_t *conf, int ignore)
1194{
1195 int first = 0;
1196
1197 do {
1198 int n = conf->copies;
1199 int cnt = 0;
1200 while (n--) {
1201 if (conf->mirrors[first].rdev &&
1202 first != ignore)
1203 cnt++;
1204 first = (first+1) % conf->raid_disks;
1205 }
1206 if (cnt == 0)
1207 return 0;
1208 } while (first != 0);
1209 return 1;
1210}
1211
1212static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1213{
1214 char b[BDEVNAME_SIZE];
1215 conf_t *conf = mddev->private;
1216
1217 /*
1218 * If it is not operational, then we have already marked it as dead
1219 * else if it is the last working disks, ignore the error, let the
1220 * next level up know.
1221 * else mark the drive as failed
1222 */
1223 if (test_bit(In_sync, &rdev->flags)
1224 && !enough(conf, rdev->raid_disk))
1225 /*
1226 * Don't fail the drive, just return an IO error.
1227 */
1228 return;
1229 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1230 unsigned long flags;
1231 spin_lock_irqsave(&conf->device_lock, flags);
1232 mddev->degraded++;
1233 spin_unlock_irqrestore(&conf->device_lock, flags);
1234 /*
1235 * if recovery is running, make sure it aborts.
1236 */
1237 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1238 }
1239 set_bit(Blocked, &rdev->flags);
1240 set_bit(Faulty, &rdev->flags);
1241 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1242 printk(KERN_ALERT
1243 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1244 "md/raid10:%s: Operation continuing on %d devices.\n",
1245 mdname(mddev), bdevname(rdev->bdev, b),
1246 mdname(mddev), conf->raid_disks - mddev->degraded);
1247}
1248
1249static void print_conf(conf_t *conf)
1250{
1251 int i;
1252 mirror_info_t *tmp;
1253
1254 printk(KERN_DEBUG "RAID10 conf printout:\n");
1255 if (!conf) {
1256 printk(KERN_DEBUG "(!conf)\n");
1257 return;
1258 }
1259 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1260 conf->raid_disks);
1261
1262 for (i = 0; i < conf->raid_disks; i++) {
1263 char b[BDEVNAME_SIZE];
1264 tmp = conf->mirrors + i;
1265 if (tmp->rdev)
1266 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1267 i, !test_bit(In_sync, &tmp->rdev->flags),
1268 !test_bit(Faulty, &tmp->rdev->flags),
1269 bdevname(tmp->rdev->bdev,b));
1270 }
1271}
1272
1273static void close_sync(conf_t *conf)
1274{
1275 wait_barrier(conf);
1276 allow_barrier(conf);
1277
1278 mempool_destroy(conf->r10buf_pool);
1279 conf->r10buf_pool = NULL;
1280}
1281
1282static int raid10_spare_active(mddev_t *mddev)
1283{
1284 int i;
1285 conf_t *conf = mddev->private;
1286 mirror_info_t *tmp;
1287 int count = 0;
1288 unsigned long flags;
1289
1290 /*
1291 * Find all non-in_sync disks within the RAID10 configuration
1292 * and mark them in_sync
1293 */
1294 for (i = 0; i < conf->raid_disks; i++) {
1295 tmp = conf->mirrors + i;
1296 if (tmp->rdev
1297 && !test_bit(Faulty, &tmp->rdev->flags)
1298 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1299 count++;
1300 sysfs_notify_dirent(tmp->rdev->sysfs_state);
1301 }
1302 }
1303 spin_lock_irqsave(&conf->device_lock, flags);
1304 mddev->degraded -= count;
1305 spin_unlock_irqrestore(&conf->device_lock, flags);
1306
1307 print_conf(conf);
1308 return count;
1309}
1310
1311
1312static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1313{
1314 conf_t *conf = mddev->private;
1315 int err = -EEXIST;
1316 int mirror;
1317 int first = 0;
1318 int last = conf->raid_disks - 1;
1319
1320 if (mddev->recovery_cp < MaxSector)
1321 /* only hot-add to in-sync arrays, as recovery is
1322 * very different from resync
1323 */
1324 return -EBUSY;
1325 if (!enough(conf, -1))
1326 return -EINVAL;
1327
1328 if (rdev->raid_disk >= 0)
1329 first = last = rdev->raid_disk;
1330
1331 if (rdev->saved_raid_disk >= first &&
1332 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1333 mirror = rdev->saved_raid_disk;
1334 else
1335 mirror = first;
1336 for ( ; mirror <= last ; mirror++) {
1337 mirror_info_t *p = &conf->mirrors[mirror];
1338 if (p->recovery_disabled == mddev->recovery_disabled)
1339 continue;
1340 if (!p->rdev)
1341 continue;
1342
1343 disk_stack_limits(mddev->gendisk, rdev->bdev,
1344 rdev->data_offset << 9);
1345 /* as we don't honour merge_bvec_fn, we must
1346 * never risk violating it, so limit
1347 * ->max_segments to one lying with a single
1348 * page, as a one page request is never in
1349 * violation.
1350 */
1351 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1352 blk_queue_max_segments(mddev->queue, 1);
1353 blk_queue_segment_boundary(mddev->queue,
1354 PAGE_CACHE_SIZE - 1);
1355 }
1356
1357 p->head_position = 0;
1358 rdev->raid_disk = mirror;
1359 err = 0;
1360 if (rdev->saved_raid_disk != mirror)
1361 conf->fullsync = 1;
1362 rcu_assign_pointer(p->rdev, rdev);
1363 break;
1364 }
1365
1366 md_integrity_add_rdev(rdev, mddev);
1367 print_conf(conf);
1368 return err;
1369}
1370
1371static int raid10_remove_disk(mddev_t *mddev, int number)
1372{
1373 conf_t *conf = mddev->private;
1374 int err = 0;
1375 mdk_rdev_t *rdev;
1376 mirror_info_t *p = conf->mirrors+ number;
1377
1378 print_conf(conf);
1379 rdev = p->rdev;
1380 if (rdev) {
1381 if (test_bit(In_sync, &rdev->flags) ||
1382 atomic_read(&rdev->nr_pending)) {
1383 err = -EBUSY;
1384 goto abort;
1385 }
1386 /* Only remove faulty devices in recovery
1387 * is not possible.
1388 */
1389 if (!test_bit(Faulty, &rdev->flags) &&
1390 mddev->recovery_disabled != p->recovery_disabled &&
1391 enough(conf, -1)) {
1392 err = -EBUSY;
1393 goto abort;
1394 }
1395 p->rdev = NULL;
1396 synchronize_rcu();
1397 if (atomic_read(&rdev->nr_pending)) {
1398 /* lost the race, try later */
1399 err = -EBUSY;
1400 p->rdev = rdev;
1401 goto abort;
1402 }
1403 err = md_integrity_register(mddev);
1404 }
1405abort:
1406
1407 print_conf(conf);
1408 return err;
1409}
1410
1411
1412static void end_sync_read(struct bio *bio, int error)
1413{
1414 r10bio_t *r10_bio = bio->bi_private;
1415 conf_t *conf = r10_bio->mddev->private;
1416 int d;
1417
1418 d = find_bio_disk(conf, r10_bio, bio, NULL);
1419
1420 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1421 set_bit(R10BIO_Uptodate, &r10_bio->state);
1422 else
1423 /* The write handler will notice the lack of
1424 * R10BIO_Uptodate and record any errors etc
1425 */
1426 atomic_add(r10_bio->sectors,
1427 &conf->mirrors[d].rdev->corrected_errors);
1428
1429 /* for reconstruct, we always reschedule after a read.
1430 * for resync, only after all reads
1431 */
1432 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1433 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1434 atomic_dec_and_test(&r10_bio->remaining)) {
1435 /* we have read all the blocks,
1436 * do the comparison in process context in raid10d
1437 */
1438 reschedule_retry(r10_bio);
1439 }
1440}
1441
1442static void end_sync_request(r10bio_t *r10_bio)
1443{
1444 mddev_t *mddev = r10_bio->mddev;
1445
1446 while (atomic_dec_and_test(&r10_bio->remaining)) {
1447 if (r10_bio->master_bio == NULL) {
1448 /* the primary of several recovery bios */
1449 sector_t s = r10_bio->sectors;
1450 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1451 test_bit(R10BIO_WriteError, &r10_bio->state))
1452 reschedule_retry(r10_bio);
1453 else
1454 put_buf(r10_bio);
1455 md_done_sync(mddev, s, 1);
1456 break;
1457 } else {
1458 r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1459 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1460 test_bit(R10BIO_WriteError, &r10_bio->state))
1461 reschedule_retry(r10_bio);
1462 else
1463 put_buf(r10_bio);
1464 r10_bio = r10_bio2;
1465 }
1466 }
1467}
1468
1469static void end_sync_write(struct bio *bio, int error)
1470{
1471 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1472 r10bio_t *r10_bio = bio->bi_private;
1473 mddev_t *mddev = r10_bio->mddev;
1474 conf_t *conf = mddev->private;
1475 int d;
1476 sector_t first_bad;
1477 int bad_sectors;
1478 int slot;
1479
1480 d = find_bio_disk(conf, r10_bio, bio, &slot);
1481
1482 if (!uptodate) {
1483 set_bit(WriteErrorSeen, &conf->mirrors[d].rdev->flags);
1484 set_bit(R10BIO_WriteError, &r10_bio->state);
1485 } else if (is_badblock(conf->mirrors[d].rdev,
1486 r10_bio->devs[slot].addr,
1487 r10_bio->sectors,
1488 &first_bad, &bad_sectors))
1489 set_bit(R10BIO_MadeGood, &r10_bio->state);
1490
1491 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1492
1493 end_sync_request(r10_bio);
1494}
1495
1496/*
1497 * Note: sync and recover and handled very differently for raid10
1498 * This code is for resync.
1499 * For resync, we read through virtual addresses and read all blocks.
1500 * If there is any error, we schedule a write. The lowest numbered
1501 * drive is authoritative.
1502 * However requests come for physical address, so we need to map.
1503 * For every physical address there are raid_disks/copies virtual addresses,
1504 * which is always are least one, but is not necessarly an integer.
1505 * This means that a physical address can span multiple chunks, so we may
1506 * have to submit multiple io requests for a single sync request.
1507 */
1508/*
1509 * We check if all blocks are in-sync and only write to blocks that
1510 * aren't in sync
1511 */
1512static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1513{
1514 conf_t *conf = mddev->private;
1515 int i, first;
1516 struct bio *tbio, *fbio;
1517
1518 atomic_set(&r10_bio->remaining, 1);
1519
1520 /* find the first device with a block */
1521 for (i=0; i<conf->copies; i++)
1522 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1523 break;
1524
1525 if (i == conf->copies)
1526 goto done;
1527
1528 first = i;
1529 fbio = r10_bio->devs[i].bio;
1530
1531 /* now find blocks with errors */
1532 for (i=0 ; i < conf->copies ; i++) {
1533 int j, d;
1534 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1535
1536 tbio = r10_bio->devs[i].bio;
1537
1538 if (tbio->bi_end_io != end_sync_read)
1539 continue;
1540 if (i == first)
1541 continue;
1542 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1543 /* We know that the bi_io_vec layout is the same for
1544 * both 'first' and 'i', so we just compare them.
1545 * All vec entries are PAGE_SIZE;
1546 */
1547 for (j = 0; j < vcnt; j++)
1548 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1549 page_address(tbio->bi_io_vec[j].bv_page),
1550 PAGE_SIZE))
1551 break;
1552 if (j == vcnt)
1553 continue;
1554 mddev->resync_mismatches += r10_bio->sectors;
1555 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1556 /* Don't fix anything. */
1557 continue;
1558 }
1559 /* Ok, we need to write this bio, either to correct an
1560 * inconsistency or to correct an unreadable block.
1561 * First we need to fixup bv_offset, bv_len and
1562 * bi_vecs, as the read request might have corrupted these
1563 */
1564 tbio->bi_vcnt = vcnt;
1565 tbio->bi_size = r10_bio->sectors << 9;
1566 tbio->bi_idx = 0;
1567 tbio->bi_phys_segments = 0;
1568 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1569 tbio->bi_flags |= 1 << BIO_UPTODATE;
1570 tbio->bi_next = NULL;
1571 tbio->bi_rw = WRITE;
1572 tbio->bi_private = r10_bio;
1573 tbio->bi_sector = r10_bio->devs[i].addr;
1574
1575 for (j=0; j < vcnt ; j++) {
1576 tbio->bi_io_vec[j].bv_offset = 0;
1577 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1578
1579 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1580 page_address(fbio->bi_io_vec[j].bv_page),
1581 PAGE_SIZE);
1582 }
1583 tbio->bi_end_io = end_sync_write;
1584
1585 d = r10_bio->devs[i].devnum;
1586 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1587 atomic_inc(&r10_bio->remaining);
1588 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1589
1590 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1591 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1592 generic_make_request(tbio);
1593 }
1594
1595done:
1596 if (atomic_dec_and_test(&r10_bio->remaining)) {
1597 md_done_sync(mddev, r10_bio->sectors, 1);
1598 put_buf(r10_bio);
1599 }
1600}
1601
1602/*
1603 * Now for the recovery code.
1604 * Recovery happens across physical sectors.
1605 * We recover all non-is_sync drives by finding the virtual address of
1606 * each, and then choose a working drive that also has that virt address.
1607 * There is a separate r10_bio for each non-in_sync drive.
1608 * Only the first two slots are in use. The first for reading,
1609 * The second for writing.
1610 *
1611 */
1612static void fix_recovery_read_error(r10bio_t *r10_bio)
1613{
1614 /* We got a read error during recovery.
1615 * We repeat the read in smaller page-sized sections.
1616 * If a read succeeds, write it to the new device or record
1617 * a bad block if we cannot.
1618 * If a read fails, record a bad block on both old and
1619 * new devices.
1620 */
1621 mddev_t *mddev = r10_bio->mddev;
1622 conf_t *conf = mddev->private;
1623 struct bio *bio = r10_bio->devs[0].bio;
1624 sector_t sect = 0;
1625 int sectors = r10_bio->sectors;
1626 int idx = 0;
1627 int dr = r10_bio->devs[0].devnum;
1628 int dw = r10_bio->devs[1].devnum;
1629
1630 while (sectors) {
1631 int s = sectors;
1632 mdk_rdev_t *rdev;
1633 sector_t addr;
1634 int ok;
1635
1636 if (s > (PAGE_SIZE>>9))
1637 s = PAGE_SIZE >> 9;
1638
1639 rdev = conf->mirrors[dr].rdev;
1640 addr = r10_bio->devs[0].addr + sect,
1641 ok = sync_page_io(rdev,
1642 addr,
1643 s << 9,
1644 bio->bi_io_vec[idx].bv_page,
1645 READ, false);
1646 if (ok) {
1647 rdev = conf->mirrors[dw].rdev;
1648 addr = r10_bio->devs[1].addr + sect;
1649 ok = sync_page_io(rdev,
1650 addr,
1651 s << 9,
1652 bio->bi_io_vec[idx].bv_page,
1653 WRITE, false);
1654 if (!ok)
1655 set_bit(WriteErrorSeen, &rdev->flags);
1656 }
1657 if (!ok) {
1658 /* We don't worry if we cannot set a bad block -
1659 * it really is bad so there is no loss in not
1660 * recording it yet
1661 */
1662 rdev_set_badblocks(rdev, addr, s, 0);
1663
1664 if (rdev != conf->mirrors[dw].rdev) {
1665 /* need bad block on destination too */
1666 mdk_rdev_t *rdev2 = conf->mirrors[dw].rdev;
1667 addr = r10_bio->devs[1].addr + sect;
1668 ok = rdev_set_badblocks(rdev2, addr, s, 0);
1669 if (!ok) {
1670 /* just abort the recovery */
1671 printk(KERN_NOTICE
1672 "md/raid10:%s: recovery aborted"
1673 " due to read error\n",
1674 mdname(mddev));
1675
1676 conf->mirrors[dw].recovery_disabled
1677 = mddev->recovery_disabled;
1678 set_bit(MD_RECOVERY_INTR,
1679 &mddev->recovery);
1680 break;
1681 }
1682 }
1683 }
1684
1685 sectors -= s;
1686 sect += s;
1687 idx++;
1688 }
1689}
1690
1691static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1692{
1693 conf_t *conf = mddev->private;
1694 int d;
1695 struct bio *wbio;
1696
1697 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
1698 fix_recovery_read_error(r10_bio);
1699 end_sync_request(r10_bio);
1700 return;
1701 }
1702
1703 /*
1704 * share the pages with the first bio
1705 * and submit the write request
1706 */
1707 wbio = r10_bio->devs[1].bio;
1708 d = r10_bio->devs[1].devnum;
1709
1710 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1711 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1712 generic_make_request(wbio);
1713}
1714
1715
1716/*
1717 * Used by fix_read_error() to decay the per rdev read_errors.
1718 * We halve the read error count for every hour that has elapsed
1719 * since the last recorded read error.
1720 *
1721 */
1722static void check_decay_read_errors(mddev_t *mddev, mdk_rdev_t *rdev)
1723{
1724 struct timespec cur_time_mon;
1725 unsigned long hours_since_last;
1726 unsigned int read_errors = atomic_read(&rdev->read_errors);
1727
1728 ktime_get_ts(&cur_time_mon);
1729
1730 if (rdev->last_read_error.tv_sec == 0 &&
1731 rdev->last_read_error.tv_nsec == 0) {
1732 /* first time we've seen a read error */
1733 rdev->last_read_error = cur_time_mon;
1734 return;
1735 }
1736
1737 hours_since_last = (cur_time_mon.tv_sec -
1738 rdev->last_read_error.tv_sec) / 3600;
1739
1740 rdev->last_read_error = cur_time_mon;
1741
1742 /*
1743 * if hours_since_last is > the number of bits in read_errors
1744 * just set read errors to 0. We do this to avoid
1745 * overflowing the shift of read_errors by hours_since_last.
1746 */
1747 if (hours_since_last >= 8 * sizeof(read_errors))
1748 atomic_set(&rdev->read_errors, 0);
1749 else
1750 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
1751}
1752
1753static int r10_sync_page_io(mdk_rdev_t *rdev, sector_t sector,
1754 int sectors, struct page *page, int rw)
1755{
1756 sector_t first_bad;
1757 int bad_sectors;
1758
1759 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
1760 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
1761 return -1;
1762 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1763 /* success */
1764 return 1;
1765 if (rw == WRITE)
1766 set_bit(WriteErrorSeen, &rdev->flags);
1767 /* need to record an error - either for the block or the device */
1768 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1769 md_error(rdev->mddev, rdev);
1770 return 0;
1771}
1772
1773/*
1774 * This is a kernel thread which:
1775 *
1776 * 1. Retries failed read operations on working mirrors.
1777 * 2. Updates the raid superblock when problems encounter.
1778 * 3. Performs writes following reads for array synchronising.
1779 */
1780
1781static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1782{
1783 int sect = 0; /* Offset from r10_bio->sector */
1784 int sectors = r10_bio->sectors;
1785 mdk_rdev_t*rdev;
1786 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
1787 int d = r10_bio->devs[r10_bio->read_slot].devnum;
1788
1789 /* still own a reference to this rdev, so it cannot
1790 * have been cleared recently.
1791 */
1792 rdev = conf->mirrors[d].rdev;
1793
1794 if (test_bit(Faulty, &rdev->flags))
1795 /* drive has already been failed, just ignore any
1796 more fix_read_error() attempts */
1797 return;
1798
1799 check_decay_read_errors(mddev, rdev);
1800 atomic_inc(&rdev->read_errors);
1801 if (atomic_read(&rdev->read_errors) > max_read_errors) {
1802 char b[BDEVNAME_SIZE];
1803 bdevname(rdev->bdev, b);
1804
1805 printk(KERN_NOTICE
1806 "md/raid10:%s: %s: Raid device exceeded "
1807 "read_error threshold [cur %d:max %d]\n",
1808 mdname(mddev), b,
1809 atomic_read(&rdev->read_errors), max_read_errors);
1810 printk(KERN_NOTICE
1811 "md/raid10:%s: %s: Failing raid device\n",
1812 mdname(mddev), b);
1813 md_error(mddev, conf->mirrors[d].rdev);
1814 return;
1815 }
1816
1817 while(sectors) {
1818 int s = sectors;
1819 int sl = r10_bio->read_slot;
1820 int success = 0;
1821 int start;
1822
1823 if (s > (PAGE_SIZE>>9))
1824 s = PAGE_SIZE >> 9;
1825
1826 rcu_read_lock();
1827 do {
1828 sector_t first_bad;
1829 int bad_sectors;
1830
1831 d = r10_bio->devs[sl].devnum;
1832 rdev = rcu_dereference(conf->mirrors[d].rdev);
1833 if (rdev &&
1834 test_bit(In_sync, &rdev->flags) &&
1835 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
1836 &first_bad, &bad_sectors) == 0) {
1837 atomic_inc(&rdev->nr_pending);
1838 rcu_read_unlock();
1839 success = sync_page_io(rdev,
1840 r10_bio->devs[sl].addr +
1841 sect,
1842 s<<9,
1843 conf->tmppage, READ, false);
1844 rdev_dec_pending(rdev, mddev);
1845 rcu_read_lock();
1846 if (success)
1847 break;
1848 }
1849 sl++;
1850 if (sl == conf->copies)
1851 sl = 0;
1852 } while (!success && sl != r10_bio->read_slot);
1853 rcu_read_unlock();
1854
1855 if (!success) {
1856 /* Cannot read from anywhere, just mark the block
1857 * as bad on the first device to discourage future
1858 * reads.
1859 */
1860 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1861 rdev = conf->mirrors[dn].rdev;
1862
1863 if (!rdev_set_badblocks(
1864 rdev,
1865 r10_bio->devs[r10_bio->read_slot].addr
1866 + sect,
1867 s, 0))
1868 md_error(mddev, rdev);
1869 break;
1870 }
1871
1872 start = sl;
1873 /* write it back and re-read */
1874 rcu_read_lock();
1875 while (sl != r10_bio->read_slot) {
1876 char b[BDEVNAME_SIZE];
1877
1878 if (sl==0)
1879 sl = conf->copies;
1880 sl--;
1881 d = r10_bio->devs[sl].devnum;
1882 rdev = rcu_dereference(conf->mirrors[d].rdev);
1883 if (!rdev ||
1884 !test_bit(In_sync, &rdev->flags))
1885 continue;
1886
1887 atomic_inc(&rdev->nr_pending);
1888 rcu_read_unlock();
1889 if (r10_sync_page_io(rdev,
1890 r10_bio->devs[sl].addr +
1891 sect,
1892 s<<9, conf->tmppage, WRITE)
1893 == 0) {
1894 /* Well, this device is dead */
1895 printk(KERN_NOTICE
1896 "md/raid10:%s: read correction "
1897 "write failed"
1898 " (%d sectors at %llu on %s)\n",
1899 mdname(mddev), s,
1900 (unsigned long long)(
1901 sect + rdev->data_offset),
1902 bdevname(rdev->bdev, b));
1903 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
1904 "drive\n",
1905 mdname(mddev),
1906 bdevname(rdev->bdev, b));
1907 }
1908 rdev_dec_pending(rdev, mddev);
1909 rcu_read_lock();
1910 }
1911 sl = start;
1912 while (sl != r10_bio->read_slot) {
1913 char b[BDEVNAME_SIZE];
1914
1915 if (sl==0)
1916 sl = conf->copies;
1917 sl--;
1918 d = r10_bio->devs[sl].devnum;
1919 rdev = rcu_dereference(conf->mirrors[d].rdev);
1920 if (!rdev ||
1921 !test_bit(In_sync, &rdev->flags))
1922 continue;
1923
1924 atomic_inc(&rdev->nr_pending);
1925 rcu_read_unlock();
1926 switch (r10_sync_page_io(rdev,
1927 r10_bio->devs[sl].addr +
1928 sect,
1929 s<<9, conf->tmppage,
1930 READ)) {
1931 case 0:
1932 /* Well, this device is dead */
1933 printk(KERN_NOTICE
1934 "md/raid10:%s: unable to read back "
1935 "corrected sectors"
1936 " (%d sectors at %llu on %s)\n",
1937 mdname(mddev), s,
1938 (unsigned long long)(
1939 sect + rdev->data_offset),
1940 bdevname(rdev->bdev, b));
1941 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
1942 "drive\n",
1943 mdname(mddev),
1944 bdevname(rdev->bdev, b));
1945 break;
1946 case 1:
1947 printk(KERN_INFO
1948 "md/raid10:%s: read error corrected"
1949 " (%d sectors at %llu on %s)\n",
1950 mdname(mddev), s,
1951 (unsigned long long)(
1952 sect + rdev->data_offset),
1953 bdevname(rdev->bdev, b));
1954 atomic_add(s, &rdev->corrected_errors);
1955 }
1956
1957 rdev_dec_pending(rdev, mddev);
1958 rcu_read_lock();
1959 }
1960 rcu_read_unlock();
1961
1962 sectors -= s;
1963 sect += s;
1964 }
1965}
1966
1967static void bi_complete(struct bio *bio, int error)
1968{
1969 complete((struct completion *)bio->bi_private);
1970}
1971
1972static int submit_bio_wait(int rw, struct bio *bio)
1973{
1974 struct completion event;
1975 rw |= REQ_SYNC;
1976
1977 init_completion(&event);
1978 bio->bi_private = &event;
1979 bio->bi_end_io = bi_complete;
1980 submit_bio(rw, bio);
1981 wait_for_completion(&event);
1982
1983 return test_bit(BIO_UPTODATE, &bio->bi_flags);
1984}
1985
1986static int narrow_write_error(r10bio_t *r10_bio, int i)
1987{
1988 struct bio *bio = r10_bio->master_bio;
1989 mddev_t *mddev = r10_bio->mddev;
1990 conf_t *conf = mddev->private;
1991 mdk_rdev_t *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
1992 /* bio has the data to be written to slot 'i' where
1993 * we just recently had a write error.
1994 * We repeatedly clone the bio and trim down to one block,
1995 * then try the write. Where the write fails we record
1996 * a bad block.
1997 * It is conceivable that the bio doesn't exactly align with
1998 * blocks. We must handle this.
1999 *
2000 * We currently own a reference to the rdev.
2001 */
2002
2003 int block_sectors;
2004 sector_t sector;
2005 int sectors;
2006 int sect_to_write = r10_bio->sectors;
2007 int ok = 1;
2008
2009 if (rdev->badblocks.shift < 0)
2010 return 0;
2011
2012 block_sectors = 1 << rdev->badblocks.shift;
2013 sector = r10_bio->sector;
2014 sectors = ((r10_bio->sector + block_sectors)
2015 & ~(sector_t)(block_sectors - 1))
2016 - sector;
2017
2018 while (sect_to_write) {
2019 struct bio *wbio;
2020 if (sectors > sect_to_write)
2021 sectors = sect_to_write;
2022 /* Write at 'sector' for 'sectors' */
2023 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2024 md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2025 wbio->bi_sector = (r10_bio->devs[i].addr+
2026 rdev->data_offset+
2027 (sector - r10_bio->sector));
2028 wbio->bi_bdev = rdev->bdev;
2029 if (submit_bio_wait(WRITE, wbio) == 0)
2030 /* Failure! */
2031 ok = rdev_set_badblocks(rdev, sector,
2032 sectors, 0)
2033 && ok;
2034
2035 bio_put(wbio);
2036 sect_to_write -= sectors;
2037 sector += sectors;
2038 sectors = block_sectors;
2039 }
2040 return ok;
2041}
2042
2043static void handle_read_error(mddev_t *mddev, r10bio_t *r10_bio)
2044{
2045 int slot = r10_bio->read_slot;
2046 int mirror = r10_bio->devs[slot].devnum;
2047 struct bio *bio;
2048 conf_t *conf = mddev->private;
2049 mdk_rdev_t *rdev;
2050 char b[BDEVNAME_SIZE];
2051 unsigned long do_sync;
2052 int max_sectors;
2053
2054 /* we got a read error. Maybe the drive is bad. Maybe just
2055 * the block and we can fix it.
2056 * We freeze all other IO, and try reading the block from
2057 * other devices. When we find one, we re-write
2058 * and check it that fixes the read error.
2059 * This is all done synchronously while the array is
2060 * frozen.
2061 */
2062 if (mddev->ro == 0) {
2063 freeze_array(conf);
2064 fix_read_error(conf, mddev, r10_bio);
2065 unfreeze_array(conf);
2066 }
2067 rdev_dec_pending(conf->mirrors[mirror].rdev, mddev);
2068
2069 bio = r10_bio->devs[slot].bio;
2070 bdevname(bio->bi_bdev, b);
2071 r10_bio->devs[slot].bio =
2072 mddev->ro ? IO_BLOCKED : NULL;
2073read_more:
2074 mirror = read_balance(conf, r10_bio, &max_sectors);
2075 if (mirror == -1) {
2076 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2077 " read error for block %llu\n",
2078 mdname(mddev), b,
2079 (unsigned long long)r10_bio->sector);
2080 raid_end_bio_io(r10_bio);
2081 bio_put(bio);
2082 return;
2083 }
2084
2085 do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2086 if (bio)
2087 bio_put(bio);
2088 slot = r10_bio->read_slot;
2089 rdev = conf->mirrors[mirror].rdev;
2090 printk_ratelimited(
2091 KERN_ERR
2092 "md/raid10:%s: %s: redirecting"
2093 "sector %llu to another mirror\n",
2094 mdname(mddev),
2095 bdevname(rdev->bdev, b),
2096 (unsigned long long)r10_bio->sector);
2097 bio = bio_clone_mddev(r10_bio->master_bio,
2098 GFP_NOIO, mddev);
2099 md_trim_bio(bio,
2100 r10_bio->sector - bio->bi_sector,
2101 max_sectors);
2102 r10_bio->devs[slot].bio = bio;
2103 bio->bi_sector = r10_bio->devs[slot].addr
2104 + rdev->data_offset;
2105 bio->bi_bdev = rdev->bdev;
2106 bio->bi_rw = READ | do_sync;
2107 bio->bi_private = r10_bio;
2108 bio->bi_end_io = raid10_end_read_request;
2109 if (max_sectors < r10_bio->sectors) {
2110 /* Drat - have to split this up more */
2111 struct bio *mbio = r10_bio->master_bio;
2112 int sectors_handled =
2113 r10_bio->sector + max_sectors
2114 - mbio->bi_sector;
2115 r10_bio->sectors = max_sectors;
2116 spin_lock_irq(&conf->device_lock);
2117 if (mbio->bi_phys_segments == 0)
2118 mbio->bi_phys_segments = 2;
2119 else
2120 mbio->bi_phys_segments++;
2121 spin_unlock_irq(&conf->device_lock);
2122 generic_make_request(bio);
2123 bio = NULL;
2124
2125 r10_bio = mempool_alloc(conf->r10bio_pool,
2126 GFP_NOIO);
2127 r10_bio->master_bio = mbio;
2128 r10_bio->sectors = (mbio->bi_size >> 9)
2129 - sectors_handled;
2130 r10_bio->state = 0;
2131 set_bit(R10BIO_ReadError,
2132 &r10_bio->state);
2133 r10_bio->mddev = mddev;
2134 r10_bio->sector = mbio->bi_sector
2135 + sectors_handled;
2136
2137 goto read_more;
2138 } else
2139 generic_make_request(bio);
2140}
2141
2142static void handle_write_completed(conf_t *conf, r10bio_t *r10_bio)
2143{
2144 /* Some sort of write request has finished and it
2145 * succeeded in writing where we thought there was a
2146 * bad block. So forget the bad block.
2147 * Or possibly if failed and we need to record
2148 * a bad block.
2149 */
2150 int m;
2151 mdk_rdev_t *rdev;
2152
2153 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2154 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2155 for (m = 0; m < conf->copies; m++) {
2156 int dev = r10_bio->devs[m].devnum;
2157 rdev = conf->mirrors[dev].rdev;
2158 if (r10_bio->devs[m].bio == NULL)
2159 continue;
2160 if (test_bit(BIO_UPTODATE,
2161 &r10_bio->devs[m].bio->bi_flags)) {
2162 rdev_clear_badblocks(
2163 rdev,
2164 r10_bio->devs[m].addr,
2165 r10_bio->sectors);
2166 } else {
2167 if (!rdev_set_badblocks(
2168 rdev,
2169 r10_bio->devs[m].addr,
2170 r10_bio->sectors, 0))
2171 md_error(conf->mddev, rdev);
2172 }
2173 }
2174 put_buf(r10_bio);
2175 } else {
2176 for (m = 0; m < conf->copies; m++) {
2177 int dev = r10_bio->devs[m].devnum;
2178 struct bio *bio = r10_bio->devs[m].bio;
2179 rdev = conf->mirrors[dev].rdev;
2180 if (bio == IO_MADE_GOOD) {
2181 rdev_clear_badblocks(
2182 rdev,
2183 r10_bio->devs[m].addr,
2184 r10_bio->sectors);
2185 rdev_dec_pending(rdev, conf->mddev);
2186 } else if (bio != NULL &&
2187 !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2188 if (!narrow_write_error(r10_bio, m)) {
2189 md_error(conf->mddev, rdev);
2190 set_bit(R10BIO_Degraded,
2191 &r10_bio->state);
2192 }
2193 rdev_dec_pending(rdev, conf->mddev);
2194 }
2195 }
2196 if (test_bit(R10BIO_WriteError,
2197 &r10_bio->state))
2198 close_write(r10_bio);
2199 raid_end_bio_io(r10_bio);
2200 }
2201}
2202
2203static void raid10d(mddev_t *mddev)
2204{
2205 r10bio_t *r10_bio;
2206 unsigned long flags;
2207 conf_t *conf = mddev->private;
2208 struct list_head *head = &conf->retry_list;
2209 struct blk_plug plug;
2210
2211 md_check_recovery(mddev);
2212
2213 blk_start_plug(&plug);
2214 for (;;) {
2215
2216 flush_pending_writes(conf);
2217
2218 spin_lock_irqsave(&conf->device_lock, flags);
2219 if (list_empty(head)) {
2220 spin_unlock_irqrestore(&conf->device_lock, flags);
2221 break;
2222 }
2223 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
2224 list_del(head->prev);
2225 conf->nr_queued--;
2226 spin_unlock_irqrestore(&conf->device_lock, flags);
2227
2228 mddev = r10_bio->mddev;
2229 conf = mddev->private;
2230 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2231 test_bit(R10BIO_WriteError, &r10_bio->state))
2232 handle_write_completed(conf, r10_bio);
2233 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2234 sync_request_write(mddev, r10_bio);
2235 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2236 recovery_request_write(mddev, r10_bio);
2237 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2238 handle_read_error(mddev, r10_bio);
2239 else {
2240 /* just a partial read to be scheduled from a
2241 * separate context
2242 */
2243 int slot = r10_bio->read_slot;
2244 generic_make_request(r10_bio->devs[slot].bio);
2245 }
2246
2247 cond_resched();
2248 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2249 md_check_recovery(mddev);
2250 }
2251 blk_finish_plug(&plug);
2252}
2253
2254
2255static int init_resync(conf_t *conf)
2256{
2257 int buffs;
2258
2259 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2260 BUG_ON(conf->r10buf_pool);
2261 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2262 if (!conf->r10buf_pool)
2263 return -ENOMEM;
2264 conf->next_resync = 0;
2265 return 0;
2266}
2267
2268/*
2269 * perform a "sync" on one "block"
2270 *
2271 * We need to make sure that no normal I/O request - particularly write
2272 * requests - conflict with active sync requests.
2273 *
2274 * This is achieved by tracking pending requests and a 'barrier' concept
2275 * that can be installed to exclude normal IO requests.
2276 *
2277 * Resync and recovery are handled very differently.
2278 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2279 *
2280 * For resync, we iterate over virtual addresses, read all copies,
2281 * and update if there are differences. If only one copy is live,
2282 * skip it.
2283 * For recovery, we iterate over physical addresses, read a good
2284 * value for each non-in_sync drive, and over-write.
2285 *
2286 * So, for recovery we may have several outstanding complex requests for a
2287 * given address, one for each out-of-sync device. We model this by allocating
2288 * a number of r10_bio structures, one for each out-of-sync device.
2289 * As we setup these structures, we collect all bio's together into a list
2290 * which we then process collectively to add pages, and then process again
2291 * to pass to generic_make_request.
2292 *
2293 * The r10_bio structures are linked using a borrowed master_bio pointer.
2294 * This link is counted in ->remaining. When the r10_bio that points to NULL
2295 * has its remaining count decremented to 0, the whole complex operation
2296 * is complete.
2297 *
2298 */
2299
2300static sector_t sync_request(mddev_t *mddev, sector_t sector_nr,
2301 int *skipped, int go_faster)
2302{
2303 conf_t *conf = mddev->private;
2304 r10bio_t *r10_bio;
2305 struct bio *biolist = NULL, *bio;
2306 sector_t max_sector, nr_sectors;
2307 int i;
2308 int max_sync;
2309 sector_t sync_blocks;
2310 sector_t sectors_skipped = 0;
2311 int chunks_skipped = 0;
2312
2313 if (!conf->r10buf_pool)
2314 if (init_resync(conf))
2315 return 0;
2316
2317 skipped:
2318 max_sector = mddev->dev_sectors;
2319 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2320 max_sector = mddev->resync_max_sectors;
2321 if (sector_nr >= max_sector) {
2322 /* If we aborted, we need to abort the
2323 * sync on the 'current' bitmap chucks (there can
2324 * be several when recovering multiple devices).
2325 * as we may have started syncing it but not finished.
2326 * We can find the current address in
2327 * mddev->curr_resync, but for recovery,
2328 * we need to convert that to several
2329 * virtual addresses.
2330 */
2331 if (mddev->curr_resync < max_sector) { /* aborted */
2332 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2333 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2334 &sync_blocks, 1);
2335 else for (i=0; i<conf->raid_disks; i++) {
2336 sector_t sect =
2337 raid10_find_virt(conf, mddev->curr_resync, i);
2338 bitmap_end_sync(mddev->bitmap, sect,
2339 &sync_blocks, 1);
2340 }
2341 } else /* completed sync */
2342 conf->fullsync = 0;
2343
2344 bitmap_close_sync(mddev->bitmap);
2345 close_sync(conf);
2346 *skipped = 1;
2347 return sectors_skipped;
2348 }
2349 if (chunks_skipped >= conf->raid_disks) {
2350 /* if there has been nothing to do on any drive,
2351 * then there is nothing to do at all..
2352 */
2353 *skipped = 1;
2354 return (max_sector - sector_nr) + sectors_skipped;
2355 }
2356
2357 if (max_sector > mddev->resync_max)
2358 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2359
2360 /* make sure whole request will fit in a chunk - if chunks
2361 * are meaningful
2362 */
2363 if (conf->near_copies < conf->raid_disks &&
2364 max_sector > (sector_nr | conf->chunk_mask))
2365 max_sector = (sector_nr | conf->chunk_mask) + 1;
2366 /*
2367 * If there is non-resync activity waiting for us then
2368 * put in a delay to throttle resync.
2369 */
2370 if (!go_faster && conf->nr_waiting)
2371 msleep_interruptible(1000);
2372
2373 /* Again, very different code for resync and recovery.
2374 * Both must result in an r10bio with a list of bios that
2375 * have bi_end_io, bi_sector, bi_bdev set,
2376 * and bi_private set to the r10bio.
2377 * For recovery, we may actually create several r10bios
2378 * with 2 bios in each, that correspond to the bios in the main one.
2379 * In this case, the subordinate r10bios link back through a
2380 * borrowed master_bio pointer, and the counter in the master
2381 * includes a ref from each subordinate.
2382 */
2383 /* First, we decide what to do and set ->bi_end_io
2384 * To end_sync_read if we want to read, and
2385 * end_sync_write if we will want to write.
2386 */
2387
2388 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2389 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2390 /* recovery... the complicated one */
2391 int j;
2392 r10_bio = NULL;
2393
2394 for (i=0 ; i<conf->raid_disks; i++) {
2395 int still_degraded;
2396 r10bio_t *rb2;
2397 sector_t sect;
2398 int must_sync;
2399 int any_working;
2400
2401 if (conf->mirrors[i].rdev == NULL ||
2402 test_bit(In_sync, &conf->mirrors[i].rdev->flags))
2403 continue;
2404
2405 still_degraded = 0;
2406 /* want to reconstruct this device */
2407 rb2 = r10_bio;
2408 sect = raid10_find_virt(conf, sector_nr, i);
2409 /* Unless we are doing a full sync, we only need
2410 * to recover the block if it is set in the bitmap
2411 */
2412 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2413 &sync_blocks, 1);
2414 if (sync_blocks < max_sync)
2415 max_sync = sync_blocks;
2416 if (!must_sync &&
2417 !conf->fullsync) {
2418 /* yep, skip the sync_blocks here, but don't assume
2419 * that there will never be anything to do here
2420 */
2421 chunks_skipped = -1;
2422 continue;
2423 }
2424
2425 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2426 raise_barrier(conf, rb2 != NULL);
2427 atomic_set(&r10_bio->remaining, 0);
2428
2429 r10_bio->master_bio = (struct bio*)rb2;
2430 if (rb2)
2431 atomic_inc(&rb2->remaining);
2432 r10_bio->mddev = mddev;
2433 set_bit(R10BIO_IsRecover, &r10_bio->state);
2434 r10_bio->sector = sect;
2435
2436 raid10_find_phys(conf, r10_bio);
2437
2438 /* Need to check if the array will still be
2439 * degraded
2440 */
2441 for (j=0; j<conf->raid_disks; j++)
2442 if (conf->mirrors[j].rdev == NULL ||
2443 test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
2444 still_degraded = 1;
2445 break;
2446 }
2447
2448 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2449 &sync_blocks, still_degraded);
2450
2451 any_working = 0;
2452 for (j=0; j<conf->copies;j++) {
2453 int k;
2454 int d = r10_bio->devs[j].devnum;
2455 sector_t from_addr, to_addr;
2456 mdk_rdev_t *rdev;
2457 sector_t sector, first_bad;
2458 int bad_sectors;
2459 if (!conf->mirrors[d].rdev ||
2460 !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
2461 continue;
2462 /* This is where we read from */
2463 any_working = 1;
2464 rdev = conf->mirrors[d].rdev;
2465 sector = r10_bio->devs[j].addr;
2466
2467 if (is_badblock(rdev, sector, max_sync,
2468 &first_bad, &bad_sectors)) {
2469 if (first_bad > sector)
2470 max_sync = first_bad - sector;
2471 else {
2472 bad_sectors -= (sector
2473 - first_bad);
2474 if (max_sync > bad_sectors)
2475 max_sync = bad_sectors;
2476 continue;
2477 }
2478 }
2479 bio = r10_bio->devs[0].bio;
2480 bio->bi_next = biolist;
2481 biolist = bio;
2482 bio->bi_private = r10_bio;
2483 bio->bi_end_io = end_sync_read;
2484 bio->bi_rw = READ;
2485 from_addr = r10_bio->devs[j].addr;
2486 bio->bi_sector = from_addr +
2487 conf->mirrors[d].rdev->data_offset;
2488 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2489 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2490 atomic_inc(&r10_bio->remaining);
2491 /* and we write to 'i' */
2492
2493 for (k=0; k<conf->copies; k++)
2494 if (r10_bio->devs[k].devnum == i)
2495 break;
2496 BUG_ON(k == conf->copies);
2497 bio = r10_bio->devs[1].bio;
2498 bio->bi_next = biolist;
2499 biolist = bio;
2500 bio->bi_private = r10_bio;
2501 bio->bi_end_io = end_sync_write;
2502 bio->bi_rw = WRITE;
2503 to_addr = r10_bio->devs[k].addr;
2504 bio->bi_sector = to_addr +
2505 conf->mirrors[i].rdev->data_offset;
2506 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
2507
2508 r10_bio->devs[0].devnum = d;
2509 r10_bio->devs[0].addr = from_addr;
2510 r10_bio->devs[1].devnum = i;
2511 r10_bio->devs[1].addr = to_addr;
2512
2513 break;
2514 }
2515 if (j == conf->copies) {
2516 /* Cannot recover, so abort the recovery or
2517 * record a bad block */
2518 put_buf(r10_bio);
2519 if (rb2)
2520 atomic_dec(&rb2->remaining);
2521 r10_bio = rb2;
2522 if (any_working) {
2523 /* problem is that there are bad blocks
2524 * on other device(s)
2525 */
2526 int k;
2527 for (k = 0; k < conf->copies; k++)
2528 if (r10_bio->devs[k].devnum == i)
2529 break;
2530 if (!rdev_set_badblocks(
2531 conf->mirrors[i].rdev,
2532 r10_bio->devs[k].addr,
2533 max_sync, 0))
2534 any_working = 0;
2535 }
2536 if (!any_working) {
2537 if (!test_and_set_bit(MD_RECOVERY_INTR,
2538 &mddev->recovery))
2539 printk(KERN_INFO "md/raid10:%s: insufficient "
2540 "working devices for recovery.\n",
2541 mdname(mddev));
2542 conf->mirrors[i].recovery_disabled
2543 = mddev->recovery_disabled;
2544 }
2545 break;
2546 }
2547 }
2548 if (biolist == NULL) {
2549 while (r10_bio) {
2550 r10bio_t *rb2 = r10_bio;
2551 r10_bio = (r10bio_t*) rb2->master_bio;
2552 rb2->master_bio = NULL;
2553 put_buf(rb2);
2554 }
2555 goto giveup;
2556 }
2557 } else {
2558 /* resync. Schedule a read for every block at this virt offset */
2559 int count = 0;
2560
2561 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2562
2563 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2564 &sync_blocks, mddev->degraded) &&
2565 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
2566 &mddev->recovery)) {
2567 /* We can skip this block */
2568 *skipped = 1;
2569 return sync_blocks + sectors_skipped;
2570 }
2571 if (sync_blocks < max_sync)
2572 max_sync = sync_blocks;
2573 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2574
2575 r10_bio->mddev = mddev;
2576 atomic_set(&r10_bio->remaining, 0);
2577 raise_barrier(conf, 0);
2578 conf->next_resync = sector_nr;
2579
2580 r10_bio->master_bio = NULL;
2581 r10_bio->sector = sector_nr;
2582 set_bit(R10BIO_IsSync, &r10_bio->state);
2583 raid10_find_phys(conf, r10_bio);
2584 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
2585
2586 for (i=0; i<conf->copies; i++) {
2587 int d = r10_bio->devs[i].devnum;
2588 sector_t first_bad, sector;
2589 int bad_sectors;
2590
2591 bio = r10_bio->devs[i].bio;
2592 bio->bi_end_io = NULL;
2593 clear_bit(BIO_UPTODATE, &bio->bi_flags);
2594 if (conf->mirrors[d].rdev == NULL ||
2595 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
2596 continue;
2597 sector = r10_bio->devs[i].addr;
2598 if (is_badblock(conf->mirrors[d].rdev,
2599 sector, max_sync,
2600 &first_bad, &bad_sectors)) {
2601 if (first_bad > sector)
2602 max_sync = first_bad - sector;
2603 else {
2604 bad_sectors -= (sector - first_bad);
2605 if (max_sync > bad_sectors)
2606 max_sync = max_sync;
2607 continue;
2608 }
2609 }
2610 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2611 atomic_inc(&r10_bio->remaining);
2612 bio->bi_next = biolist;
2613 biolist = bio;
2614 bio->bi_private = r10_bio;
2615 bio->bi_end_io = end_sync_read;
2616 bio->bi_rw = READ;
2617 bio->bi_sector = sector +
2618 conf->mirrors[d].rdev->data_offset;
2619 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2620 count++;
2621 }
2622
2623 if (count < 2) {
2624 for (i=0; i<conf->copies; i++) {
2625 int d = r10_bio->devs[i].devnum;
2626 if (r10_bio->devs[i].bio->bi_end_io)
2627 rdev_dec_pending(conf->mirrors[d].rdev,
2628 mddev);
2629 }
2630 put_buf(r10_bio);
2631 biolist = NULL;
2632 goto giveup;
2633 }
2634 }
2635
2636 for (bio = biolist; bio ; bio=bio->bi_next) {
2637
2638 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
2639 if (bio->bi_end_io)
2640 bio->bi_flags |= 1 << BIO_UPTODATE;
2641 bio->bi_vcnt = 0;
2642 bio->bi_idx = 0;
2643 bio->bi_phys_segments = 0;
2644 bio->bi_size = 0;
2645 }
2646
2647 nr_sectors = 0;
2648 if (sector_nr + max_sync < max_sector)
2649 max_sector = sector_nr + max_sync;
2650 do {
2651 struct page *page;
2652 int len = PAGE_SIZE;
2653 if (sector_nr + (len>>9) > max_sector)
2654 len = (max_sector - sector_nr) << 9;
2655 if (len == 0)
2656 break;
2657 for (bio= biolist ; bio ; bio=bio->bi_next) {
2658 struct bio *bio2;
2659 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2660 if (bio_add_page(bio, page, len, 0))
2661 continue;
2662
2663 /* stop here */
2664 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2665 for (bio2 = biolist;
2666 bio2 && bio2 != bio;
2667 bio2 = bio2->bi_next) {
2668 /* remove last page from this bio */
2669 bio2->bi_vcnt--;
2670 bio2->bi_size -= len;
2671 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
2672 }
2673 goto bio_full;
2674 }
2675 nr_sectors += len>>9;
2676 sector_nr += len>>9;
2677 } while (biolist->bi_vcnt < RESYNC_PAGES);
2678 bio_full:
2679 r10_bio->sectors = nr_sectors;
2680
2681 while (biolist) {
2682 bio = biolist;
2683 biolist = biolist->bi_next;
2684
2685 bio->bi_next = NULL;
2686 r10_bio = bio->bi_private;
2687 r10_bio->sectors = nr_sectors;
2688
2689 if (bio->bi_end_io == end_sync_read) {
2690 md_sync_acct(bio->bi_bdev, nr_sectors);
2691 generic_make_request(bio);
2692 }
2693 }
2694
2695 if (sectors_skipped)
2696 /* pretend they weren't skipped, it makes
2697 * no important difference in this case
2698 */
2699 md_done_sync(mddev, sectors_skipped, 1);
2700
2701 return sectors_skipped + nr_sectors;
2702 giveup:
2703 /* There is nowhere to write, so all non-sync
2704 * drives must be failed or in resync, all drives
2705 * have a bad block, so try the next chunk...
2706 */
2707 if (sector_nr + max_sync < max_sector)
2708 max_sector = sector_nr + max_sync;
2709
2710 sectors_skipped += (max_sector - sector_nr);
2711 chunks_skipped ++;
2712 sector_nr = max_sector;
2713 goto skipped;
2714}
2715
2716static sector_t
2717raid10_size(mddev_t *mddev, sector_t sectors, int raid_disks)
2718{
2719 sector_t size;
2720 conf_t *conf = mddev->private;
2721
2722 if (!raid_disks)
2723 raid_disks = conf->raid_disks;
2724 if (!sectors)
2725 sectors = conf->dev_sectors;
2726
2727 size = sectors >> conf->chunk_shift;
2728 sector_div(size, conf->far_copies);
2729 size = size * raid_disks;
2730 sector_div(size, conf->near_copies);
2731
2732 return size << conf->chunk_shift;
2733}
2734
2735
2736static conf_t *setup_conf(mddev_t *mddev)
2737{
2738 conf_t *conf = NULL;
2739 int nc, fc, fo;
2740 sector_t stride, size;
2741 int err = -EINVAL;
2742
2743 if (mddev->new_chunk_sectors < (PAGE_SIZE >> 9) ||
2744 !is_power_of_2(mddev->new_chunk_sectors)) {
2745 printk(KERN_ERR "md/raid10:%s: chunk size must be "
2746 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
2747 mdname(mddev), PAGE_SIZE);
2748 goto out;
2749 }
2750
2751 nc = mddev->new_layout & 255;
2752 fc = (mddev->new_layout >> 8) & 255;
2753 fo = mddev->new_layout & (1<<16);
2754
2755 if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
2756 (mddev->new_layout >> 17)) {
2757 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
2758 mdname(mddev), mddev->new_layout);
2759 goto out;
2760 }
2761
2762 err = -ENOMEM;
2763 conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2764 if (!conf)
2765 goto out;
2766
2767 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2768 GFP_KERNEL);
2769 if (!conf->mirrors)
2770 goto out;
2771
2772 conf->tmppage = alloc_page(GFP_KERNEL);
2773 if (!conf->tmppage)
2774 goto out;
2775
2776
2777 conf->raid_disks = mddev->raid_disks;
2778 conf->near_copies = nc;
2779 conf->far_copies = fc;
2780 conf->copies = nc*fc;
2781 conf->far_offset = fo;
2782 conf->chunk_mask = mddev->new_chunk_sectors - 1;
2783 conf->chunk_shift = ffz(~mddev->new_chunk_sectors);
2784
2785 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2786 r10bio_pool_free, conf);
2787 if (!conf->r10bio_pool)
2788 goto out;
2789
2790 size = mddev->dev_sectors >> conf->chunk_shift;
2791 sector_div(size, fc);
2792 size = size * conf->raid_disks;
2793 sector_div(size, nc);
2794 /* 'size' is now the number of chunks in the array */
2795 /* calculate "used chunks per device" in 'stride' */
2796 stride = size * conf->copies;
2797
2798 /* We need to round up when dividing by raid_disks to
2799 * get the stride size.
2800 */
2801 stride += conf->raid_disks - 1;
2802 sector_div(stride, conf->raid_disks);
2803
2804 conf->dev_sectors = stride << conf->chunk_shift;
2805
2806 if (fo)
2807 stride = 1;
2808 else
2809 sector_div(stride, fc);
2810 conf->stride = stride << conf->chunk_shift;
2811
2812
2813 spin_lock_init(&conf->device_lock);
2814 INIT_LIST_HEAD(&conf->retry_list);
2815
2816 spin_lock_init(&conf->resync_lock);
2817 init_waitqueue_head(&conf->wait_barrier);
2818
2819 conf->thread = md_register_thread(raid10d, mddev, NULL);
2820 if (!conf->thread)
2821 goto out;
2822
2823 conf->mddev = mddev;
2824 return conf;
2825
2826 out:
2827 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
2828 mdname(mddev));
2829 if (conf) {
2830 if (conf->r10bio_pool)
2831 mempool_destroy(conf->r10bio_pool);
2832 kfree(conf->mirrors);
2833 safe_put_page(conf->tmppage);
2834 kfree(conf);
2835 }
2836 return ERR_PTR(err);
2837}
2838
2839static int run(mddev_t *mddev)
2840{
2841 conf_t *conf;
2842 int i, disk_idx, chunk_size;
2843 mirror_info_t *disk;
2844 mdk_rdev_t *rdev;
2845 sector_t size;
2846
2847 /*
2848 * copy the already verified devices into our private RAID10
2849 * bookkeeping area. [whatever we allocate in run(),
2850 * should be freed in stop()]
2851 */
2852
2853 if (mddev->private == NULL) {
2854 conf = setup_conf(mddev);
2855 if (IS_ERR(conf))
2856 return PTR_ERR(conf);
2857 mddev->private = conf;
2858 }
2859 conf = mddev->private;
2860 if (!conf)
2861 goto out;
2862
2863 mddev->thread = conf->thread;
2864 conf->thread = NULL;
2865
2866 chunk_size = mddev->chunk_sectors << 9;
2867 blk_queue_io_min(mddev->queue, chunk_size);
2868 if (conf->raid_disks % conf->near_copies)
2869 blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks);
2870 else
2871 blk_queue_io_opt(mddev->queue, chunk_size *
2872 (conf->raid_disks / conf->near_copies));
2873
2874 list_for_each_entry(rdev, &mddev->disks, same_set) {
2875
2876 disk_idx = rdev->raid_disk;
2877 if (disk_idx >= conf->raid_disks
2878 || disk_idx < 0)
2879 continue;
2880 disk = conf->mirrors + disk_idx;
2881
2882 disk->rdev = rdev;
2883 disk_stack_limits(mddev->gendisk, rdev->bdev,
2884 rdev->data_offset << 9);
2885 /* as we don't honour merge_bvec_fn, we must never risk
2886 * violating it, so limit max_segments to 1 lying
2887 * within a single page.
2888 */
2889 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2890 blk_queue_max_segments(mddev->queue, 1);
2891 blk_queue_segment_boundary(mddev->queue,
2892 PAGE_CACHE_SIZE - 1);
2893 }
2894
2895 disk->head_position = 0;
2896 }
2897 /* need to check that every block has at least one working mirror */
2898 if (!enough(conf, -1)) {
2899 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
2900 mdname(mddev));
2901 goto out_free_conf;
2902 }
2903
2904 mddev->degraded = 0;
2905 for (i = 0; i < conf->raid_disks; i++) {
2906
2907 disk = conf->mirrors + i;
2908
2909 if (!disk->rdev ||
2910 !test_bit(In_sync, &disk->rdev->flags)) {
2911 disk->head_position = 0;
2912 mddev->degraded++;
2913 if (disk->rdev)
2914 conf->fullsync = 1;
2915 }
2916 }
2917
2918 if (mddev->recovery_cp != MaxSector)
2919 printk(KERN_NOTICE "md/raid10:%s: not clean"
2920 " -- starting background reconstruction\n",
2921 mdname(mddev));
2922 printk(KERN_INFO
2923 "md/raid10:%s: active with %d out of %d devices\n",
2924 mdname(mddev), conf->raid_disks - mddev->degraded,
2925 conf->raid_disks);
2926 /*
2927 * Ok, everything is just fine now
2928 */
2929 mddev->dev_sectors = conf->dev_sectors;
2930 size = raid10_size(mddev, 0, 0);
2931 md_set_array_sectors(mddev, size);
2932 mddev->resync_max_sectors = size;
2933
2934 mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2935 mddev->queue->backing_dev_info.congested_data = mddev;
2936
2937 /* Calculate max read-ahead size.
2938 * We need to readahead at least twice a whole stripe....
2939 * maybe...
2940 */
2941 {
2942 int stripe = conf->raid_disks *
2943 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
2944 stripe /= conf->near_copies;
2945 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2946 mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2947 }
2948
2949 if (conf->near_copies < conf->raid_disks)
2950 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2951
2952 if (md_integrity_register(mddev))
2953 goto out_free_conf;
2954
2955 return 0;
2956
2957out_free_conf:
2958 md_unregister_thread(&mddev->thread);
2959 if (conf->r10bio_pool)
2960 mempool_destroy(conf->r10bio_pool);
2961 safe_put_page(conf->tmppage);
2962 kfree(conf->mirrors);
2963 kfree(conf);
2964 mddev->private = NULL;
2965out:
2966 return -EIO;
2967}
2968
2969static int stop(mddev_t *mddev)
2970{
2971 conf_t *conf = mddev->private;
2972
2973 raise_barrier(conf, 0);
2974 lower_barrier(conf);
2975
2976 md_unregister_thread(&mddev->thread);
2977 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2978 if (conf->r10bio_pool)
2979 mempool_destroy(conf->r10bio_pool);
2980 kfree(conf->mirrors);
2981 kfree(conf);
2982 mddev->private = NULL;
2983 return 0;
2984}
2985
2986static void raid10_quiesce(mddev_t *mddev, int state)
2987{
2988 conf_t *conf = mddev->private;
2989
2990 switch(state) {
2991 case 1:
2992 raise_barrier(conf, 0);
2993 break;
2994 case 0:
2995 lower_barrier(conf);
2996 break;
2997 }
2998}
2999
3000static void *raid10_takeover_raid0(mddev_t *mddev)
3001{
3002 mdk_rdev_t *rdev;
3003 conf_t *conf;
3004
3005 if (mddev->degraded > 0) {
3006 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3007 mdname(mddev));
3008 return ERR_PTR(-EINVAL);
3009 }
3010
3011 /* Set new parameters */
3012 mddev->new_level = 10;
3013 /* new layout: far_copies = 1, near_copies = 2 */
3014 mddev->new_layout = (1<<8) + 2;
3015 mddev->new_chunk_sectors = mddev->chunk_sectors;
3016 mddev->delta_disks = mddev->raid_disks;
3017 mddev->raid_disks *= 2;
3018 /* make sure it will be not marked as dirty */
3019 mddev->recovery_cp = MaxSector;
3020
3021 conf = setup_conf(mddev);
3022 if (!IS_ERR(conf)) {
3023 list_for_each_entry(rdev, &mddev->disks, same_set)
3024 if (rdev->raid_disk >= 0)
3025 rdev->new_raid_disk = rdev->raid_disk * 2;
3026 conf->barrier = 1;
3027 }
3028
3029 return conf;
3030}
3031
3032static void *raid10_takeover(mddev_t *mddev)
3033{
3034 struct raid0_private_data *raid0_priv;
3035
3036 /* raid10 can take over:
3037 * raid0 - providing it has only two drives
3038 */
3039 if (mddev->level == 0) {
3040 /* for raid0 takeover only one zone is supported */
3041 raid0_priv = mddev->private;
3042 if (raid0_priv->nr_strip_zones > 1) {
3043 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3044 " with more than one zone.\n",
3045 mdname(mddev));
3046 return ERR_PTR(-EINVAL);
3047 }
3048 return raid10_takeover_raid0(mddev);
3049 }
3050 return ERR_PTR(-EINVAL);
3051}
3052
3053static struct mdk_personality raid10_personality =
3054{
3055 .name = "raid10",
3056 .level = 10,
3057 .owner = THIS_MODULE,
3058 .make_request = make_request,
3059 .run = run,
3060 .stop = stop,
3061 .status = status,
3062 .error_handler = error,
3063 .hot_add_disk = raid10_add_disk,
3064 .hot_remove_disk= raid10_remove_disk,
3065 .spare_active = raid10_spare_active,
3066 .sync_request = sync_request,
3067 .quiesce = raid10_quiesce,
3068 .size = raid10_size,
3069 .takeover = raid10_takeover,
3070};
3071
3072static int __init raid_init(void)
3073{
3074 return register_md_personality(&raid10_personality);
3075}
3076
3077static void raid_exit(void)
3078{
3079 unregister_md_personality(&raid10_personality);
3080}
3081
3082module_init(raid_init);
3083module_exit(raid_exit);
3084MODULE_LICENSE("GPL");
3085MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
3086MODULE_ALIAS("md-personality-9"); /* RAID10 */
3087MODULE_ALIAS("md-raid10");
3088MODULE_ALIAS("md-level-10");
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * raid10.c : Multiple Devices driver for Linux
4 *
5 * Copyright (C) 2000-2004 Neil Brown
6 *
7 * RAID-10 support for md.
8 *
9 * Base on code in raid1.c. See raid1.c for further copyright information.
10 */
11
12#include <linux/slab.h>
13#include <linux/delay.h>
14#include <linux/blkdev.h>
15#include <linux/module.h>
16#include <linux/seq_file.h>
17#include <linux/ratelimit.h>
18#include <linux/kthread.h>
19#include <linux/raid/md_p.h>
20#include <trace/events/block.h>
21#include "md.h"
22#include "raid10.h"
23#include "raid0.h"
24#include "md-bitmap.h"
25
26/*
27 * RAID10 provides a combination of RAID0 and RAID1 functionality.
28 * The layout of data is defined by
29 * chunk_size
30 * raid_disks
31 * near_copies (stored in low byte of layout)
32 * far_copies (stored in second byte of layout)
33 * far_offset (stored in bit 16 of layout )
34 * use_far_sets (stored in bit 17 of layout )
35 * use_far_sets_bugfixed (stored in bit 18 of layout )
36 *
37 * The data to be stored is divided into chunks using chunksize. Each device
38 * is divided into far_copies sections. In each section, chunks are laid out
39 * in a style similar to raid0, but near_copies copies of each chunk is stored
40 * (each on a different drive). The starting device for each section is offset
41 * near_copies from the starting device of the previous section. Thus there
42 * are (near_copies * far_copies) of each chunk, and each is on a different
43 * drive. near_copies and far_copies must be at least one, and their product
44 * is at most raid_disks.
45 *
46 * If far_offset is true, then the far_copies are handled a bit differently.
47 * The copies are still in different stripes, but instead of being very far
48 * apart on disk, there are adjacent stripes.
49 *
50 * The far and offset algorithms are handled slightly differently if
51 * 'use_far_sets' is true. In this case, the array's devices are grouped into
52 * sets that are (near_copies * far_copies) in size. The far copied stripes
53 * are still shifted by 'near_copies' devices, but this shifting stays confined
54 * to the set rather than the entire array. This is done to improve the number
55 * of device combinations that can fail without causing the array to fail.
56 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
57 * on a device):
58 * A B C D A B C D E
59 * ... ...
60 * D A B C E A B C D
61 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
62 * [A B] [C D] [A B] [C D E]
63 * |...| |...| |...| | ... |
64 * [B A] [D C] [B A] [E C D]
65 */
66
67static void allow_barrier(struct r10conf *conf);
68static void lower_barrier(struct r10conf *conf);
69static int _enough(struct r10conf *conf, int previous, int ignore);
70static int enough(struct r10conf *conf, int ignore);
71static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
72 int *skipped);
73static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
74static void end_reshape_write(struct bio *bio);
75static void end_reshape(struct r10conf *conf);
76
77#define raid10_log(md, fmt, args...) \
78 do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid10 " fmt, ##args); } while (0)
79
80#include "raid1-10.c"
81
82/*
83 * for resync bio, r10bio pointer can be retrieved from the per-bio
84 * 'struct resync_pages'.
85 */
86static inline struct r10bio *get_resync_r10bio(struct bio *bio)
87{
88 return get_resync_pages(bio)->raid_bio;
89}
90
91static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
92{
93 struct r10conf *conf = data;
94 int size = offsetof(struct r10bio, devs[conf->copies]);
95
96 /* allocate a r10bio with room for raid_disks entries in the
97 * bios array */
98 return kzalloc(size, gfp_flags);
99}
100
101#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
102/* amount of memory to reserve for resync requests */
103#define RESYNC_WINDOW (1024*1024)
104/* maximum number of concurrent requests, memory permitting */
105#define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
106#define CLUSTER_RESYNC_WINDOW (32 * RESYNC_WINDOW)
107#define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
108
109/*
110 * When performing a resync, we need to read and compare, so
111 * we need as many pages are there are copies.
112 * When performing a recovery, we need 2 bios, one for read,
113 * one for write (we recover only one drive per r10buf)
114 *
115 */
116static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
117{
118 struct r10conf *conf = data;
119 struct r10bio *r10_bio;
120 struct bio *bio;
121 int j;
122 int nalloc, nalloc_rp;
123 struct resync_pages *rps;
124
125 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
126 if (!r10_bio)
127 return NULL;
128
129 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
130 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
131 nalloc = conf->copies; /* resync */
132 else
133 nalloc = 2; /* recovery */
134
135 /* allocate once for all bios */
136 if (!conf->have_replacement)
137 nalloc_rp = nalloc;
138 else
139 nalloc_rp = nalloc * 2;
140 rps = kmalloc_array(nalloc_rp, sizeof(struct resync_pages), gfp_flags);
141 if (!rps)
142 goto out_free_r10bio;
143
144 /*
145 * Allocate bios.
146 */
147 for (j = nalloc ; j-- ; ) {
148 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
149 if (!bio)
150 goto out_free_bio;
151 r10_bio->devs[j].bio = bio;
152 if (!conf->have_replacement)
153 continue;
154 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
155 if (!bio)
156 goto out_free_bio;
157 r10_bio->devs[j].repl_bio = bio;
158 }
159 /*
160 * Allocate RESYNC_PAGES data pages and attach them
161 * where needed.
162 */
163 for (j = 0; j < nalloc; j++) {
164 struct bio *rbio = r10_bio->devs[j].repl_bio;
165 struct resync_pages *rp, *rp_repl;
166
167 rp = &rps[j];
168 if (rbio)
169 rp_repl = &rps[nalloc + j];
170
171 bio = r10_bio->devs[j].bio;
172
173 if (!j || test_bit(MD_RECOVERY_SYNC,
174 &conf->mddev->recovery)) {
175 if (resync_alloc_pages(rp, gfp_flags))
176 goto out_free_pages;
177 } else {
178 memcpy(rp, &rps[0], sizeof(*rp));
179 resync_get_all_pages(rp);
180 }
181
182 rp->raid_bio = r10_bio;
183 bio->bi_private = rp;
184 if (rbio) {
185 memcpy(rp_repl, rp, sizeof(*rp));
186 rbio->bi_private = rp_repl;
187 }
188 }
189
190 return r10_bio;
191
192out_free_pages:
193 while (--j >= 0)
194 resync_free_pages(&rps[j]);
195
196 j = 0;
197out_free_bio:
198 for ( ; j < nalloc; j++) {
199 if (r10_bio->devs[j].bio)
200 bio_put(r10_bio->devs[j].bio);
201 if (r10_bio->devs[j].repl_bio)
202 bio_put(r10_bio->devs[j].repl_bio);
203 }
204 kfree(rps);
205out_free_r10bio:
206 rbio_pool_free(r10_bio, conf);
207 return NULL;
208}
209
210static void r10buf_pool_free(void *__r10_bio, void *data)
211{
212 struct r10conf *conf = data;
213 struct r10bio *r10bio = __r10_bio;
214 int j;
215 struct resync_pages *rp = NULL;
216
217 for (j = conf->copies; j--; ) {
218 struct bio *bio = r10bio->devs[j].bio;
219
220 if (bio) {
221 rp = get_resync_pages(bio);
222 resync_free_pages(rp);
223 bio_put(bio);
224 }
225
226 bio = r10bio->devs[j].repl_bio;
227 if (bio)
228 bio_put(bio);
229 }
230
231 /* resync pages array stored in the 1st bio's .bi_private */
232 kfree(rp);
233
234 rbio_pool_free(r10bio, conf);
235}
236
237static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
238{
239 int i;
240
241 for (i = 0; i < conf->copies; i++) {
242 struct bio **bio = & r10_bio->devs[i].bio;
243 if (!BIO_SPECIAL(*bio))
244 bio_put(*bio);
245 *bio = NULL;
246 bio = &r10_bio->devs[i].repl_bio;
247 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
248 bio_put(*bio);
249 *bio = NULL;
250 }
251}
252
253static void free_r10bio(struct r10bio *r10_bio)
254{
255 struct r10conf *conf = r10_bio->mddev->private;
256
257 put_all_bios(conf, r10_bio);
258 mempool_free(r10_bio, &conf->r10bio_pool);
259}
260
261static void put_buf(struct r10bio *r10_bio)
262{
263 struct r10conf *conf = r10_bio->mddev->private;
264
265 mempool_free(r10_bio, &conf->r10buf_pool);
266
267 lower_barrier(conf);
268}
269
270static void reschedule_retry(struct r10bio *r10_bio)
271{
272 unsigned long flags;
273 struct mddev *mddev = r10_bio->mddev;
274 struct r10conf *conf = mddev->private;
275
276 spin_lock_irqsave(&conf->device_lock, flags);
277 list_add(&r10_bio->retry_list, &conf->retry_list);
278 conf->nr_queued ++;
279 spin_unlock_irqrestore(&conf->device_lock, flags);
280
281 /* wake up frozen array... */
282 wake_up(&conf->wait_barrier);
283
284 md_wakeup_thread(mddev->thread);
285}
286
287/*
288 * raid_end_bio_io() is called when we have finished servicing a mirrored
289 * operation and are ready to return a success/failure code to the buffer
290 * cache layer.
291 */
292static void raid_end_bio_io(struct r10bio *r10_bio)
293{
294 struct bio *bio = r10_bio->master_bio;
295 struct r10conf *conf = r10_bio->mddev->private;
296
297 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
298 bio->bi_status = BLK_STS_IOERR;
299
300 bio_endio(bio);
301 /*
302 * Wake up any possible resync thread that waits for the device
303 * to go idle.
304 */
305 allow_barrier(conf);
306
307 free_r10bio(r10_bio);
308}
309
310/*
311 * Update disk head position estimator based on IRQ completion info.
312 */
313static inline void update_head_pos(int slot, struct r10bio *r10_bio)
314{
315 struct r10conf *conf = r10_bio->mddev->private;
316
317 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
318 r10_bio->devs[slot].addr + (r10_bio->sectors);
319}
320
321/*
322 * Find the disk number which triggered given bio
323 */
324static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
325 struct bio *bio, int *slotp, int *replp)
326{
327 int slot;
328 int repl = 0;
329
330 for (slot = 0; slot < conf->copies; slot++) {
331 if (r10_bio->devs[slot].bio == bio)
332 break;
333 if (r10_bio->devs[slot].repl_bio == bio) {
334 repl = 1;
335 break;
336 }
337 }
338
339 BUG_ON(slot == conf->copies);
340 update_head_pos(slot, r10_bio);
341
342 if (slotp)
343 *slotp = slot;
344 if (replp)
345 *replp = repl;
346 return r10_bio->devs[slot].devnum;
347}
348
349static void raid10_end_read_request(struct bio *bio)
350{
351 int uptodate = !bio->bi_status;
352 struct r10bio *r10_bio = bio->bi_private;
353 int slot;
354 struct md_rdev *rdev;
355 struct r10conf *conf = r10_bio->mddev->private;
356
357 slot = r10_bio->read_slot;
358 rdev = r10_bio->devs[slot].rdev;
359 /*
360 * this branch is our 'one mirror IO has finished' event handler:
361 */
362 update_head_pos(slot, r10_bio);
363
364 if (uptodate) {
365 /*
366 * Set R10BIO_Uptodate in our master bio, so that
367 * we will return a good error code to the higher
368 * levels even if IO on some other mirrored buffer fails.
369 *
370 * The 'master' represents the composite IO operation to
371 * user-side. So if something waits for IO, then it will
372 * wait for the 'master' bio.
373 */
374 set_bit(R10BIO_Uptodate, &r10_bio->state);
375 } else {
376 /* If all other devices that store this block have
377 * failed, we want to return the error upwards rather
378 * than fail the last device. Here we redefine
379 * "uptodate" to mean "Don't want to retry"
380 */
381 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
382 rdev->raid_disk))
383 uptodate = 1;
384 }
385 if (uptodate) {
386 raid_end_bio_io(r10_bio);
387 rdev_dec_pending(rdev, conf->mddev);
388 } else {
389 /*
390 * oops, read error - keep the refcount on the rdev
391 */
392 char b[BDEVNAME_SIZE];
393 pr_err_ratelimited("md/raid10:%s: %s: rescheduling sector %llu\n",
394 mdname(conf->mddev),
395 bdevname(rdev->bdev, b),
396 (unsigned long long)r10_bio->sector);
397 set_bit(R10BIO_ReadError, &r10_bio->state);
398 reschedule_retry(r10_bio);
399 }
400}
401
402static void close_write(struct r10bio *r10_bio)
403{
404 /* clear the bitmap if all writes complete successfully */
405 md_bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
406 r10_bio->sectors,
407 !test_bit(R10BIO_Degraded, &r10_bio->state),
408 0);
409 md_write_end(r10_bio->mddev);
410}
411
412static void one_write_done(struct r10bio *r10_bio)
413{
414 if (atomic_dec_and_test(&r10_bio->remaining)) {
415 if (test_bit(R10BIO_WriteError, &r10_bio->state))
416 reschedule_retry(r10_bio);
417 else {
418 close_write(r10_bio);
419 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
420 reschedule_retry(r10_bio);
421 else
422 raid_end_bio_io(r10_bio);
423 }
424 }
425}
426
427static void raid10_end_write_request(struct bio *bio)
428{
429 struct r10bio *r10_bio = bio->bi_private;
430 int dev;
431 int dec_rdev = 1;
432 struct r10conf *conf = r10_bio->mddev->private;
433 int slot, repl;
434 struct md_rdev *rdev = NULL;
435 struct bio *to_put = NULL;
436 bool discard_error;
437
438 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
439
440 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
441
442 if (repl)
443 rdev = conf->mirrors[dev].replacement;
444 if (!rdev) {
445 smp_rmb();
446 repl = 0;
447 rdev = conf->mirrors[dev].rdev;
448 }
449 /*
450 * this branch is our 'one mirror IO has finished' event handler:
451 */
452 if (bio->bi_status && !discard_error) {
453 if (repl)
454 /* Never record new bad blocks to replacement,
455 * just fail it.
456 */
457 md_error(rdev->mddev, rdev);
458 else {
459 set_bit(WriteErrorSeen, &rdev->flags);
460 if (!test_and_set_bit(WantReplacement, &rdev->flags))
461 set_bit(MD_RECOVERY_NEEDED,
462 &rdev->mddev->recovery);
463
464 dec_rdev = 0;
465 if (test_bit(FailFast, &rdev->flags) &&
466 (bio->bi_opf & MD_FAILFAST)) {
467 md_error(rdev->mddev, rdev);
468 }
469
470 /*
471 * When the device is faulty, it is not necessary to
472 * handle write error.
473 * For failfast, this is the only remaining device,
474 * We need to retry the write without FailFast.
475 */
476 if (!test_bit(Faulty, &rdev->flags))
477 set_bit(R10BIO_WriteError, &r10_bio->state);
478 else {
479 r10_bio->devs[slot].bio = NULL;
480 to_put = bio;
481 dec_rdev = 1;
482 }
483 }
484 } else {
485 /*
486 * Set R10BIO_Uptodate in our master bio, so that
487 * we will return a good error code for to the higher
488 * levels even if IO on some other mirrored buffer fails.
489 *
490 * The 'master' represents the composite IO operation to
491 * user-side. So if something waits for IO, then it will
492 * wait for the 'master' bio.
493 */
494 sector_t first_bad;
495 int bad_sectors;
496
497 /*
498 * Do not set R10BIO_Uptodate if the current device is
499 * rebuilding or Faulty. This is because we cannot use
500 * such device for properly reading the data back (we could
501 * potentially use it, if the current write would have felt
502 * before rdev->recovery_offset, but for simplicity we don't
503 * check this here.
504 */
505 if (test_bit(In_sync, &rdev->flags) &&
506 !test_bit(Faulty, &rdev->flags))
507 set_bit(R10BIO_Uptodate, &r10_bio->state);
508
509 /* Maybe we can clear some bad blocks. */
510 if (is_badblock(rdev,
511 r10_bio->devs[slot].addr,
512 r10_bio->sectors,
513 &first_bad, &bad_sectors) && !discard_error) {
514 bio_put(bio);
515 if (repl)
516 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
517 else
518 r10_bio->devs[slot].bio = IO_MADE_GOOD;
519 dec_rdev = 0;
520 set_bit(R10BIO_MadeGood, &r10_bio->state);
521 }
522 }
523
524 /*
525 *
526 * Let's see if all mirrored write operations have finished
527 * already.
528 */
529 one_write_done(r10_bio);
530 if (dec_rdev)
531 rdev_dec_pending(rdev, conf->mddev);
532 if (to_put)
533 bio_put(to_put);
534}
535
536/*
537 * RAID10 layout manager
538 * As well as the chunksize and raid_disks count, there are two
539 * parameters: near_copies and far_copies.
540 * near_copies * far_copies must be <= raid_disks.
541 * Normally one of these will be 1.
542 * If both are 1, we get raid0.
543 * If near_copies == raid_disks, we get raid1.
544 *
545 * Chunks are laid out in raid0 style with near_copies copies of the
546 * first chunk, followed by near_copies copies of the next chunk and
547 * so on.
548 * If far_copies > 1, then after 1/far_copies of the array has been assigned
549 * as described above, we start again with a device offset of near_copies.
550 * So we effectively have another copy of the whole array further down all
551 * the drives, but with blocks on different drives.
552 * With this layout, and block is never stored twice on the one device.
553 *
554 * raid10_find_phys finds the sector offset of a given virtual sector
555 * on each device that it is on.
556 *
557 * raid10_find_virt does the reverse mapping, from a device and a
558 * sector offset to a virtual address
559 */
560
561static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
562{
563 int n,f;
564 sector_t sector;
565 sector_t chunk;
566 sector_t stripe;
567 int dev;
568 int slot = 0;
569 int last_far_set_start, last_far_set_size;
570
571 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
572 last_far_set_start *= geo->far_set_size;
573
574 last_far_set_size = geo->far_set_size;
575 last_far_set_size += (geo->raid_disks % geo->far_set_size);
576
577 /* now calculate first sector/dev */
578 chunk = r10bio->sector >> geo->chunk_shift;
579 sector = r10bio->sector & geo->chunk_mask;
580
581 chunk *= geo->near_copies;
582 stripe = chunk;
583 dev = sector_div(stripe, geo->raid_disks);
584 if (geo->far_offset)
585 stripe *= geo->far_copies;
586
587 sector += stripe << geo->chunk_shift;
588
589 /* and calculate all the others */
590 for (n = 0; n < geo->near_copies; n++) {
591 int d = dev;
592 int set;
593 sector_t s = sector;
594 r10bio->devs[slot].devnum = d;
595 r10bio->devs[slot].addr = s;
596 slot++;
597
598 for (f = 1; f < geo->far_copies; f++) {
599 set = d / geo->far_set_size;
600 d += geo->near_copies;
601
602 if ((geo->raid_disks % geo->far_set_size) &&
603 (d > last_far_set_start)) {
604 d -= last_far_set_start;
605 d %= last_far_set_size;
606 d += last_far_set_start;
607 } else {
608 d %= geo->far_set_size;
609 d += geo->far_set_size * set;
610 }
611 s += geo->stride;
612 r10bio->devs[slot].devnum = d;
613 r10bio->devs[slot].addr = s;
614 slot++;
615 }
616 dev++;
617 if (dev >= geo->raid_disks) {
618 dev = 0;
619 sector += (geo->chunk_mask + 1);
620 }
621 }
622}
623
624static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
625{
626 struct geom *geo = &conf->geo;
627
628 if (conf->reshape_progress != MaxSector &&
629 ((r10bio->sector >= conf->reshape_progress) !=
630 conf->mddev->reshape_backwards)) {
631 set_bit(R10BIO_Previous, &r10bio->state);
632 geo = &conf->prev;
633 } else
634 clear_bit(R10BIO_Previous, &r10bio->state);
635
636 __raid10_find_phys(geo, r10bio);
637}
638
639static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
640{
641 sector_t offset, chunk, vchunk;
642 /* Never use conf->prev as this is only called during resync
643 * or recovery, so reshape isn't happening
644 */
645 struct geom *geo = &conf->geo;
646 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
647 int far_set_size = geo->far_set_size;
648 int last_far_set_start;
649
650 if (geo->raid_disks % geo->far_set_size) {
651 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
652 last_far_set_start *= geo->far_set_size;
653
654 if (dev >= last_far_set_start) {
655 far_set_size = geo->far_set_size;
656 far_set_size += (geo->raid_disks % geo->far_set_size);
657 far_set_start = last_far_set_start;
658 }
659 }
660
661 offset = sector & geo->chunk_mask;
662 if (geo->far_offset) {
663 int fc;
664 chunk = sector >> geo->chunk_shift;
665 fc = sector_div(chunk, geo->far_copies);
666 dev -= fc * geo->near_copies;
667 if (dev < far_set_start)
668 dev += far_set_size;
669 } else {
670 while (sector >= geo->stride) {
671 sector -= geo->stride;
672 if (dev < (geo->near_copies + far_set_start))
673 dev += far_set_size - geo->near_copies;
674 else
675 dev -= geo->near_copies;
676 }
677 chunk = sector >> geo->chunk_shift;
678 }
679 vchunk = chunk * geo->raid_disks + dev;
680 sector_div(vchunk, geo->near_copies);
681 return (vchunk << geo->chunk_shift) + offset;
682}
683
684/*
685 * This routine returns the disk from which the requested read should
686 * be done. There is a per-array 'next expected sequential IO' sector
687 * number - if this matches on the next IO then we use the last disk.
688 * There is also a per-disk 'last know head position' sector that is
689 * maintained from IRQ contexts, both the normal and the resync IO
690 * completion handlers update this position correctly. If there is no
691 * perfect sequential match then we pick the disk whose head is closest.
692 *
693 * If there are 2 mirrors in the same 2 devices, performance degrades
694 * because position is mirror, not device based.
695 *
696 * The rdev for the device selected will have nr_pending incremented.
697 */
698
699/*
700 * FIXME: possibly should rethink readbalancing and do it differently
701 * depending on near_copies / far_copies geometry.
702 */
703static struct md_rdev *read_balance(struct r10conf *conf,
704 struct r10bio *r10_bio,
705 int *max_sectors)
706{
707 const sector_t this_sector = r10_bio->sector;
708 int disk, slot;
709 int sectors = r10_bio->sectors;
710 int best_good_sectors;
711 sector_t new_distance, best_dist;
712 struct md_rdev *best_dist_rdev, *best_pending_rdev, *rdev = NULL;
713 int do_balance;
714 int best_dist_slot, best_pending_slot;
715 bool has_nonrot_disk = false;
716 unsigned int min_pending;
717 struct geom *geo = &conf->geo;
718
719 raid10_find_phys(conf, r10_bio);
720 rcu_read_lock();
721 best_dist_slot = -1;
722 min_pending = UINT_MAX;
723 best_dist_rdev = NULL;
724 best_pending_rdev = NULL;
725 best_dist = MaxSector;
726 best_good_sectors = 0;
727 do_balance = 1;
728 clear_bit(R10BIO_FailFast, &r10_bio->state);
729 /*
730 * Check if we can balance. We can balance on the whole
731 * device if no resync is going on (recovery is ok), or below
732 * the resync window. We take the first readable disk when
733 * above the resync window.
734 */
735 if ((conf->mddev->recovery_cp < MaxSector
736 && (this_sector + sectors >= conf->next_resync)) ||
737 (mddev_is_clustered(conf->mddev) &&
738 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
739 this_sector + sectors)))
740 do_balance = 0;
741
742 for (slot = 0; slot < conf->copies ; slot++) {
743 sector_t first_bad;
744 int bad_sectors;
745 sector_t dev_sector;
746 unsigned int pending;
747 bool nonrot;
748
749 if (r10_bio->devs[slot].bio == IO_BLOCKED)
750 continue;
751 disk = r10_bio->devs[slot].devnum;
752 rdev = rcu_dereference(conf->mirrors[disk].replacement);
753 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
754 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
755 rdev = rcu_dereference(conf->mirrors[disk].rdev);
756 if (rdev == NULL ||
757 test_bit(Faulty, &rdev->flags))
758 continue;
759 if (!test_bit(In_sync, &rdev->flags) &&
760 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
761 continue;
762
763 dev_sector = r10_bio->devs[slot].addr;
764 if (is_badblock(rdev, dev_sector, sectors,
765 &first_bad, &bad_sectors)) {
766 if (best_dist < MaxSector)
767 /* Already have a better slot */
768 continue;
769 if (first_bad <= dev_sector) {
770 /* Cannot read here. If this is the
771 * 'primary' device, then we must not read
772 * beyond 'bad_sectors' from another device.
773 */
774 bad_sectors -= (dev_sector - first_bad);
775 if (!do_balance && sectors > bad_sectors)
776 sectors = bad_sectors;
777 if (best_good_sectors > sectors)
778 best_good_sectors = sectors;
779 } else {
780 sector_t good_sectors =
781 first_bad - dev_sector;
782 if (good_sectors > best_good_sectors) {
783 best_good_sectors = good_sectors;
784 best_dist_slot = slot;
785 best_dist_rdev = rdev;
786 }
787 if (!do_balance)
788 /* Must read from here */
789 break;
790 }
791 continue;
792 } else
793 best_good_sectors = sectors;
794
795 if (!do_balance)
796 break;
797
798 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
799 has_nonrot_disk |= nonrot;
800 pending = atomic_read(&rdev->nr_pending);
801 if (min_pending > pending && nonrot) {
802 min_pending = pending;
803 best_pending_slot = slot;
804 best_pending_rdev = rdev;
805 }
806
807 if (best_dist_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 && !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
824 if (new_distance < best_dist) {
825 best_dist = new_distance;
826 best_dist_slot = slot;
827 best_dist_rdev = rdev;
828 }
829 }
830 if (slot >= conf->copies) {
831 if (has_nonrot_disk) {
832 slot = best_pending_slot;
833 rdev = best_pending_rdev;
834 } else {
835 slot = best_dist_slot;
836 rdev = best_dist_rdev;
837 }
838 }
839
840 if (slot >= 0) {
841 atomic_inc(&rdev->nr_pending);
842 r10_bio->read_slot = slot;
843 } else
844 rdev = NULL;
845 rcu_read_unlock();
846 *max_sectors = best_good_sectors;
847
848 return rdev;
849}
850
851static void flush_pending_writes(struct r10conf *conf)
852{
853 /* Any writes that have been queued but are awaiting
854 * bitmap updates get flushed here.
855 */
856 spin_lock_irq(&conf->device_lock);
857
858 if (conf->pending_bio_list.head) {
859 struct blk_plug plug;
860 struct bio *bio;
861
862 bio = bio_list_get(&conf->pending_bio_list);
863 conf->pending_count = 0;
864 spin_unlock_irq(&conf->device_lock);
865
866 /*
867 * As this is called in a wait_event() loop (see freeze_array),
868 * current->state might be TASK_UNINTERRUPTIBLE which will
869 * cause a warning when we prepare to wait again. As it is
870 * rare that this path is taken, it is perfectly safe to force
871 * us to go around the wait_event() loop again, so the warning
872 * is a false-positive. Silence the warning by resetting
873 * thread state
874 */
875 __set_current_state(TASK_RUNNING);
876
877 blk_start_plug(&plug);
878 /* flush any pending bitmap writes to disk
879 * before proceeding w/ I/O */
880 md_bitmap_unplug(conf->mddev->bitmap);
881 wake_up(&conf->wait_barrier);
882
883 while (bio) { /* submit pending writes */
884 struct bio *next = bio->bi_next;
885 struct md_rdev *rdev = (void*)bio->bi_disk;
886 bio->bi_next = NULL;
887 bio_set_dev(bio, rdev->bdev);
888 if (test_bit(Faulty, &rdev->flags)) {
889 bio_io_error(bio);
890 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
891 !blk_queue_discard(bio->bi_disk->queue)))
892 /* Just ignore it */
893 bio_endio(bio);
894 else
895 submit_bio_noacct(bio);
896 bio = next;
897 }
898 blk_finish_plug(&plug);
899 } else
900 spin_unlock_irq(&conf->device_lock);
901}
902
903/* Barriers....
904 * Sometimes we need to suspend IO while we do something else,
905 * either some resync/recovery, or reconfigure the array.
906 * To do this we raise a 'barrier'.
907 * The 'barrier' is a counter that can be raised multiple times
908 * to count how many activities are happening which preclude
909 * normal IO.
910 * We can only raise the barrier if there is no pending IO.
911 * i.e. if nr_pending == 0.
912 * We choose only to raise the barrier if no-one is waiting for the
913 * barrier to go down. This means that as soon as an IO request
914 * is ready, no other operations which require a barrier will start
915 * until the IO request has had a chance.
916 *
917 * So: regular IO calls 'wait_barrier'. When that returns there
918 * is no backgroup IO happening, It must arrange to call
919 * allow_barrier when it has finished its IO.
920 * backgroup IO calls must call raise_barrier. Once that returns
921 * there is no normal IO happeing. It must arrange to call
922 * lower_barrier when the particular background IO completes.
923 */
924
925static void raise_barrier(struct r10conf *conf, int force)
926{
927 BUG_ON(force && !conf->barrier);
928 spin_lock_irq(&conf->resync_lock);
929
930 /* Wait until no block IO is waiting (unless 'force') */
931 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
932 conf->resync_lock);
933
934 /* block any new IO from starting */
935 conf->barrier++;
936
937 /* Now wait for all pending IO to complete */
938 wait_event_lock_irq(conf->wait_barrier,
939 !atomic_read(&conf->nr_pending) && conf->barrier < RESYNC_DEPTH,
940 conf->resync_lock);
941
942 spin_unlock_irq(&conf->resync_lock);
943}
944
945static void lower_barrier(struct r10conf *conf)
946{
947 unsigned long flags;
948 spin_lock_irqsave(&conf->resync_lock, flags);
949 conf->barrier--;
950 spin_unlock_irqrestore(&conf->resync_lock, flags);
951 wake_up(&conf->wait_barrier);
952}
953
954static void wait_barrier(struct r10conf *conf)
955{
956 spin_lock_irq(&conf->resync_lock);
957 if (conf->barrier) {
958 struct bio_list *bio_list = current->bio_list;
959 conf->nr_waiting++;
960 /* Wait for the barrier to drop.
961 * However if there are already pending
962 * requests (preventing the barrier from
963 * rising completely), and the
964 * pre-process bio queue isn't empty,
965 * then don't wait, as we need to empty
966 * that queue to get the nr_pending
967 * count down.
968 */
969 raid10_log(conf->mddev, "wait barrier");
970 wait_event_lock_irq(conf->wait_barrier,
971 !conf->barrier ||
972 (atomic_read(&conf->nr_pending) &&
973 bio_list &&
974 (!bio_list_empty(&bio_list[0]) ||
975 !bio_list_empty(&bio_list[1]))) ||
976 /* move on if recovery thread is
977 * blocked by us
978 */
979 (conf->mddev->thread->tsk == current &&
980 test_bit(MD_RECOVERY_RUNNING,
981 &conf->mddev->recovery) &&
982 conf->nr_queued > 0),
983 conf->resync_lock);
984 conf->nr_waiting--;
985 if (!conf->nr_waiting)
986 wake_up(&conf->wait_barrier);
987 }
988 atomic_inc(&conf->nr_pending);
989 spin_unlock_irq(&conf->resync_lock);
990}
991
992static void allow_barrier(struct r10conf *conf)
993{
994 if ((atomic_dec_and_test(&conf->nr_pending)) ||
995 (conf->array_freeze_pending))
996 wake_up(&conf->wait_barrier);
997}
998
999static void freeze_array(struct r10conf *conf, int extra)
1000{
1001 /* stop syncio and normal IO and wait for everything to
1002 * go quiet.
1003 * We increment barrier and nr_waiting, and then
1004 * wait until nr_pending match nr_queued+extra
1005 * This is called in the context of one normal IO request
1006 * that has failed. Thus any sync request that might be pending
1007 * will be blocked by nr_pending, and we need to wait for
1008 * pending IO requests to complete or be queued for re-try.
1009 * Thus the number queued (nr_queued) plus this request (extra)
1010 * must match the number of pending IOs (nr_pending) before
1011 * we continue.
1012 */
1013 spin_lock_irq(&conf->resync_lock);
1014 conf->array_freeze_pending++;
1015 conf->barrier++;
1016 conf->nr_waiting++;
1017 wait_event_lock_irq_cmd(conf->wait_barrier,
1018 atomic_read(&conf->nr_pending) == conf->nr_queued+extra,
1019 conf->resync_lock,
1020 flush_pending_writes(conf));
1021
1022 conf->array_freeze_pending--;
1023 spin_unlock_irq(&conf->resync_lock);
1024}
1025
1026static void unfreeze_array(struct r10conf *conf)
1027{
1028 /* reverse the effect of the freeze */
1029 spin_lock_irq(&conf->resync_lock);
1030 conf->barrier--;
1031 conf->nr_waiting--;
1032 wake_up(&conf->wait_barrier);
1033 spin_unlock_irq(&conf->resync_lock);
1034}
1035
1036static sector_t choose_data_offset(struct r10bio *r10_bio,
1037 struct md_rdev *rdev)
1038{
1039 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1040 test_bit(R10BIO_Previous, &r10_bio->state))
1041 return rdev->data_offset;
1042 else
1043 return rdev->new_data_offset;
1044}
1045
1046struct raid10_plug_cb {
1047 struct blk_plug_cb cb;
1048 struct bio_list pending;
1049 int pending_cnt;
1050};
1051
1052static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1053{
1054 struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1055 cb);
1056 struct mddev *mddev = plug->cb.data;
1057 struct r10conf *conf = mddev->private;
1058 struct bio *bio;
1059
1060 if (from_schedule || current->bio_list) {
1061 spin_lock_irq(&conf->device_lock);
1062 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1063 conf->pending_count += plug->pending_cnt;
1064 spin_unlock_irq(&conf->device_lock);
1065 wake_up(&conf->wait_barrier);
1066 md_wakeup_thread(mddev->thread);
1067 kfree(plug);
1068 return;
1069 }
1070
1071 /* we aren't scheduling, so we can do the write-out directly. */
1072 bio = bio_list_get(&plug->pending);
1073 md_bitmap_unplug(mddev->bitmap);
1074 wake_up(&conf->wait_barrier);
1075
1076 while (bio) { /* submit pending writes */
1077 struct bio *next = bio->bi_next;
1078 struct md_rdev *rdev = (void*)bio->bi_disk;
1079 bio->bi_next = NULL;
1080 bio_set_dev(bio, rdev->bdev);
1081 if (test_bit(Faulty, &rdev->flags)) {
1082 bio_io_error(bio);
1083 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
1084 !blk_queue_discard(bio->bi_disk->queue)))
1085 /* Just ignore it */
1086 bio_endio(bio);
1087 else
1088 submit_bio_noacct(bio);
1089 bio = next;
1090 }
1091 kfree(plug);
1092}
1093
1094/*
1095 * 1. Register the new request and wait if the reconstruction thread has put
1096 * up a bar for new requests. Continue immediately if no resync is active
1097 * currently.
1098 * 2. If IO spans the reshape position. Need to wait for reshape to pass.
1099 */
1100static void regular_request_wait(struct mddev *mddev, struct r10conf *conf,
1101 struct bio *bio, sector_t sectors)
1102{
1103 wait_barrier(conf);
1104 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1105 bio->bi_iter.bi_sector < conf->reshape_progress &&
1106 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1107 raid10_log(conf->mddev, "wait reshape");
1108 allow_barrier(conf);
1109 wait_event(conf->wait_barrier,
1110 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1111 conf->reshape_progress >= bio->bi_iter.bi_sector +
1112 sectors);
1113 wait_barrier(conf);
1114 }
1115}
1116
1117static void raid10_read_request(struct mddev *mddev, struct bio *bio,
1118 struct r10bio *r10_bio)
1119{
1120 struct r10conf *conf = mddev->private;
1121 struct bio *read_bio;
1122 const int op = bio_op(bio);
1123 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1124 int max_sectors;
1125 struct md_rdev *rdev;
1126 char b[BDEVNAME_SIZE];
1127 int slot = r10_bio->read_slot;
1128 struct md_rdev *err_rdev = NULL;
1129 gfp_t gfp = GFP_NOIO;
1130
1131 if (r10_bio->devs[slot].rdev) {
1132 /*
1133 * This is an error retry, but we cannot
1134 * safely dereference the rdev in the r10_bio,
1135 * we must use the one in conf.
1136 * If it has already been disconnected (unlikely)
1137 * we lose the device name in error messages.
1138 */
1139 int disk;
1140 /*
1141 * As we are blocking raid10, it is a little safer to
1142 * use __GFP_HIGH.
1143 */
1144 gfp = GFP_NOIO | __GFP_HIGH;
1145
1146 rcu_read_lock();
1147 disk = r10_bio->devs[slot].devnum;
1148 err_rdev = rcu_dereference(conf->mirrors[disk].rdev);
1149 if (err_rdev)
1150 bdevname(err_rdev->bdev, b);
1151 else {
1152 strcpy(b, "???");
1153 /* This never gets dereferenced */
1154 err_rdev = r10_bio->devs[slot].rdev;
1155 }
1156 rcu_read_unlock();
1157 }
1158
1159 regular_request_wait(mddev, conf, bio, r10_bio->sectors);
1160 rdev = read_balance(conf, r10_bio, &max_sectors);
1161 if (!rdev) {
1162 if (err_rdev) {
1163 pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
1164 mdname(mddev), b,
1165 (unsigned long long)r10_bio->sector);
1166 }
1167 raid_end_bio_io(r10_bio);
1168 return;
1169 }
1170 if (err_rdev)
1171 pr_err_ratelimited("md/raid10:%s: %s: redirecting sector %llu to another mirror\n",
1172 mdname(mddev),
1173 bdevname(rdev->bdev, b),
1174 (unsigned long long)r10_bio->sector);
1175 if (max_sectors < bio_sectors(bio)) {
1176 struct bio *split = bio_split(bio, max_sectors,
1177 gfp, &conf->bio_split);
1178 bio_chain(split, bio);
1179 allow_barrier(conf);
1180 submit_bio_noacct(bio);
1181 wait_barrier(conf);
1182 bio = split;
1183 r10_bio->master_bio = bio;
1184 r10_bio->sectors = max_sectors;
1185 }
1186 slot = r10_bio->read_slot;
1187
1188 read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1189
1190 r10_bio->devs[slot].bio = read_bio;
1191 r10_bio->devs[slot].rdev = rdev;
1192
1193 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1194 choose_data_offset(r10_bio, rdev);
1195 bio_set_dev(read_bio, rdev->bdev);
1196 read_bio->bi_end_io = raid10_end_read_request;
1197 bio_set_op_attrs(read_bio, op, do_sync);
1198 if (test_bit(FailFast, &rdev->flags) &&
1199 test_bit(R10BIO_FailFast, &r10_bio->state))
1200 read_bio->bi_opf |= MD_FAILFAST;
1201 read_bio->bi_private = r10_bio;
1202
1203 if (mddev->gendisk)
1204 trace_block_bio_remap(read_bio->bi_disk->queue,
1205 read_bio, disk_devt(mddev->gendisk),
1206 r10_bio->sector);
1207 submit_bio_noacct(read_bio);
1208 return;
1209}
1210
1211static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio,
1212 struct bio *bio, bool replacement,
1213 int n_copy)
1214{
1215 const int op = bio_op(bio);
1216 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1217 const unsigned long do_fua = (bio->bi_opf & REQ_FUA);
1218 unsigned long flags;
1219 struct blk_plug_cb *cb;
1220 struct raid10_plug_cb *plug = NULL;
1221 struct r10conf *conf = mddev->private;
1222 struct md_rdev *rdev;
1223 int devnum = r10_bio->devs[n_copy].devnum;
1224 struct bio *mbio;
1225
1226 if (replacement) {
1227 rdev = conf->mirrors[devnum].replacement;
1228 if (rdev == NULL) {
1229 /* Replacement just got moved to main 'rdev' */
1230 smp_mb();
1231 rdev = conf->mirrors[devnum].rdev;
1232 }
1233 } else
1234 rdev = conf->mirrors[devnum].rdev;
1235
1236 mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1237 if (replacement)
1238 r10_bio->devs[n_copy].repl_bio = mbio;
1239 else
1240 r10_bio->devs[n_copy].bio = mbio;
1241
1242 mbio->bi_iter.bi_sector = (r10_bio->devs[n_copy].addr +
1243 choose_data_offset(r10_bio, rdev));
1244 bio_set_dev(mbio, rdev->bdev);
1245 mbio->bi_end_io = raid10_end_write_request;
1246 bio_set_op_attrs(mbio, op, do_sync | do_fua);
1247 if (!replacement && test_bit(FailFast,
1248 &conf->mirrors[devnum].rdev->flags)
1249 && enough(conf, devnum))
1250 mbio->bi_opf |= MD_FAILFAST;
1251 mbio->bi_private = r10_bio;
1252
1253 if (conf->mddev->gendisk)
1254 trace_block_bio_remap(mbio->bi_disk->queue,
1255 mbio, disk_devt(conf->mddev->gendisk),
1256 r10_bio->sector);
1257 /* flush_pending_writes() needs access to the rdev so...*/
1258 mbio->bi_disk = (void *)rdev;
1259
1260 atomic_inc(&r10_bio->remaining);
1261
1262 cb = blk_check_plugged(raid10_unplug, mddev, sizeof(*plug));
1263 if (cb)
1264 plug = container_of(cb, struct raid10_plug_cb, cb);
1265 else
1266 plug = NULL;
1267 if (plug) {
1268 bio_list_add(&plug->pending, mbio);
1269 plug->pending_cnt++;
1270 } else {
1271 spin_lock_irqsave(&conf->device_lock, flags);
1272 bio_list_add(&conf->pending_bio_list, mbio);
1273 conf->pending_count++;
1274 spin_unlock_irqrestore(&conf->device_lock, flags);
1275 md_wakeup_thread(mddev->thread);
1276 }
1277}
1278
1279static void raid10_write_request(struct mddev *mddev, struct bio *bio,
1280 struct r10bio *r10_bio)
1281{
1282 struct r10conf *conf = mddev->private;
1283 int i;
1284 struct md_rdev *blocked_rdev;
1285 sector_t sectors;
1286 int max_sectors;
1287
1288 if ((mddev_is_clustered(mddev) &&
1289 md_cluster_ops->area_resyncing(mddev, WRITE,
1290 bio->bi_iter.bi_sector,
1291 bio_end_sector(bio)))) {
1292 DEFINE_WAIT(w);
1293 for (;;) {
1294 prepare_to_wait(&conf->wait_barrier,
1295 &w, TASK_IDLE);
1296 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1297 bio->bi_iter.bi_sector, bio_end_sector(bio)))
1298 break;
1299 schedule();
1300 }
1301 finish_wait(&conf->wait_barrier, &w);
1302 }
1303
1304 sectors = r10_bio->sectors;
1305 regular_request_wait(mddev, conf, bio, sectors);
1306 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1307 (mddev->reshape_backwards
1308 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1309 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1310 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1311 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1312 /* Need to update reshape_position in metadata */
1313 mddev->reshape_position = conf->reshape_progress;
1314 set_mask_bits(&mddev->sb_flags, 0,
1315 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1316 md_wakeup_thread(mddev->thread);
1317 raid10_log(conf->mddev, "wait reshape metadata");
1318 wait_event(mddev->sb_wait,
1319 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
1320
1321 conf->reshape_safe = mddev->reshape_position;
1322 }
1323
1324 if (conf->pending_count >= max_queued_requests) {
1325 md_wakeup_thread(mddev->thread);
1326 raid10_log(mddev, "wait queued");
1327 wait_event(conf->wait_barrier,
1328 conf->pending_count < max_queued_requests);
1329 }
1330 /* first select target devices under rcu_lock and
1331 * inc refcount on their rdev. Record them by setting
1332 * bios[x] to bio
1333 * If there are known/acknowledged bad blocks on any device
1334 * on which we have seen a write error, we want to avoid
1335 * writing to those blocks. This potentially requires several
1336 * writes to write around the bad blocks. Each set of writes
1337 * gets its own r10_bio with a set of bios attached.
1338 */
1339
1340 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1341 raid10_find_phys(conf, r10_bio);
1342retry_write:
1343 blocked_rdev = NULL;
1344 rcu_read_lock();
1345 max_sectors = r10_bio->sectors;
1346
1347 for (i = 0; i < conf->copies; i++) {
1348 int d = r10_bio->devs[i].devnum;
1349 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1350 struct md_rdev *rrdev = rcu_dereference(
1351 conf->mirrors[d].replacement);
1352 if (rdev == rrdev)
1353 rrdev = NULL;
1354 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1355 atomic_inc(&rdev->nr_pending);
1356 blocked_rdev = rdev;
1357 break;
1358 }
1359 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1360 atomic_inc(&rrdev->nr_pending);
1361 blocked_rdev = rrdev;
1362 break;
1363 }
1364 if (rdev && (test_bit(Faulty, &rdev->flags)))
1365 rdev = NULL;
1366 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1367 rrdev = NULL;
1368
1369 r10_bio->devs[i].bio = NULL;
1370 r10_bio->devs[i].repl_bio = NULL;
1371
1372 if (!rdev && !rrdev) {
1373 set_bit(R10BIO_Degraded, &r10_bio->state);
1374 continue;
1375 }
1376 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1377 sector_t first_bad;
1378 sector_t dev_sector = r10_bio->devs[i].addr;
1379 int bad_sectors;
1380 int is_bad;
1381
1382 is_bad = is_badblock(rdev, dev_sector, max_sectors,
1383 &first_bad, &bad_sectors);
1384 if (is_bad < 0) {
1385 /* Mustn't write here until the bad block
1386 * is acknowledged
1387 */
1388 atomic_inc(&rdev->nr_pending);
1389 set_bit(BlockedBadBlocks, &rdev->flags);
1390 blocked_rdev = rdev;
1391 break;
1392 }
1393 if (is_bad && first_bad <= dev_sector) {
1394 /* Cannot write here at all */
1395 bad_sectors -= (dev_sector - first_bad);
1396 if (bad_sectors < max_sectors)
1397 /* Mustn't write more than bad_sectors
1398 * to other devices yet
1399 */
1400 max_sectors = bad_sectors;
1401 /* We don't set R10BIO_Degraded as that
1402 * only applies if the disk is missing,
1403 * so it might be re-added, and we want to
1404 * know to recover this chunk.
1405 * In this case the device is here, and the
1406 * fact that this chunk is not in-sync is
1407 * recorded in the bad block log.
1408 */
1409 continue;
1410 }
1411 if (is_bad) {
1412 int good_sectors = first_bad - dev_sector;
1413 if (good_sectors < max_sectors)
1414 max_sectors = good_sectors;
1415 }
1416 }
1417 if (rdev) {
1418 r10_bio->devs[i].bio = bio;
1419 atomic_inc(&rdev->nr_pending);
1420 }
1421 if (rrdev) {
1422 r10_bio->devs[i].repl_bio = bio;
1423 atomic_inc(&rrdev->nr_pending);
1424 }
1425 }
1426 rcu_read_unlock();
1427
1428 if (unlikely(blocked_rdev)) {
1429 /* Have to wait for this device to get unblocked, then retry */
1430 int j;
1431 int d;
1432
1433 for (j = 0; j < i; j++) {
1434 if (r10_bio->devs[j].bio) {
1435 d = r10_bio->devs[j].devnum;
1436 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1437 }
1438 if (r10_bio->devs[j].repl_bio) {
1439 struct md_rdev *rdev;
1440 d = r10_bio->devs[j].devnum;
1441 rdev = conf->mirrors[d].replacement;
1442 if (!rdev) {
1443 /* Race with remove_disk */
1444 smp_mb();
1445 rdev = conf->mirrors[d].rdev;
1446 }
1447 rdev_dec_pending(rdev, mddev);
1448 }
1449 }
1450 allow_barrier(conf);
1451 raid10_log(conf->mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1452 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1453 wait_barrier(conf);
1454 goto retry_write;
1455 }
1456
1457 if (max_sectors < r10_bio->sectors)
1458 r10_bio->sectors = max_sectors;
1459
1460 if (r10_bio->sectors < bio_sectors(bio)) {
1461 struct bio *split = bio_split(bio, r10_bio->sectors,
1462 GFP_NOIO, &conf->bio_split);
1463 bio_chain(split, bio);
1464 allow_barrier(conf);
1465 submit_bio_noacct(bio);
1466 wait_barrier(conf);
1467 bio = split;
1468 r10_bio->master_bio = bio;
1469 }
1470
1471 atomic_set(&r10_bio->remaining, 1);
1472 md_bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1473
1474 for (i = 0; i < conf->copies; i++) {
1475 if (r10_bio->devs[i].bio)
1476 raid10_write_one_disk(mddev, r10_bio, bio, false, i);
1477 if (r10_bio->devs[i].repl_bio)
1478 raid10_write_one_disk(mddev, r10_bio, bio, true, i);
1479 }
1480 one_write_done(r10_bio);
1481}
1482
1483static void __make_request(struct mddev *mddev, struct bio *bio, int sectors)
1484{
1485 struct r10conf *conf = mddev->private;
1486 struct r10bio *r10_bio;
1487
1488 r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1489
1490 r10_bio->master_bio = bio;
1491 r10_bio->sectors = sectors;
1492
1493 r10_bio->mddev = mddev;
1494 r10_bio->sector = bio->bi_iter.bi_sector;
1495 r10_bio->state = 0;
1496 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) * conf->copies);
1497
1498 if (bio_data_dir(bio) == READ)
1499 raid10_read_request(mddev, bio, r10_bio);
1500 else
1501 raid10_write_request(mddev, bio, r10_bio);
1502}
1503
1504static bool raid10_make_request(struct mddev *mddev, struct bio *bio)
1505{
1506 struct r10conf *conf = mddev->private;
1507 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1508 int chunk_sects = chunk_mask + 1;
1509 int sectors = bio_sectors(bio);
1510
1511 if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1512 && md_flush_request(mddev, bio))
1513 return true;
1514
1515 if (!md_write_start(mddev, bio))
1516 return false;
1517
1518 /*
1519 * If this request crosses a chunk boundary, we need to split
1520 * it.
1521 */
1522 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1523 sectors > chunk_sects
1524 && (conf->geo.near_copies < conf->geo.raid_disks
1525 || conf->prev.near_copies <
1526 conf->prev.raid_disks)))
1527 sectors = chunk_sects -
1528 (bio->bi_iter.bi_sector &
1529 (chunk_sects - 1));
1530 __make_request(mddev, bio, sectors);
1531
1532 /* In case raid10d snuck in to freeze_array */
1533 wake_up(&conf->wait_barrier);
1534 return true;
1535}
1536
1537static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1538{
1539 struct r10conf *conf = mddev->private;
1540 int i;
1541
1542 if (conf->geo.near_copies < conf->geo.raid_disks)
1543 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1544 if (conf->geo.near_copies > 1)
1545 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1546 if (conf->geo.far_copies > 1) {
1547 if (conf->geo.far_offset)
1548 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1549 else
1550 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1551 if (conf->geo.far_set_size != conf->geo.raid_disks)
1552 seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1553 }
1554 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1555 conf->geo.raid_disks - mddev->degraded);
1556 rcu_read_lock();
1557 for (i = 0; i < conf->geo.raid_disks; i++) {
1558 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1559 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1560 }
1561 rcu_read_unlock();
1562 seq_printf(seq, "]");
1563}
1564
1565/* check if there are enough drives for
1566 * every block to appear on atleast one.
1567 * Don't consider the device numbered 'ignore'
1568 * as we might be about to remove it.
1569 */
1570static int _enough(struct r10conf *conf, int previous, int ignore)
1571{
1572 int first = 0;
1573 int has_enough = 0;
1574 int disks, ncopies;
1575 if (previous) {
1576 disks = conf->prev.raid_disks;
1577 ncopies = conf->prev.near_copies;
1578 } else {
1579 disks = conf->geo.raid_disks;
1580 ncopies = conf->geo.near_copies;
1581 }
1582
1583 rcu_read_lock();
1584 do {
1585 int n = conf->copies;
1586 int cnt = 0;
1587 int this = first;
1588 while (n--) {
1589 struct md_rdev *rdev;
1590 if (this != ignore &&
1591 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1592 test_bit(In_sync, &rdev->flags))
1593 cnt++;
1594 this = (this+1) % disks;
1595 }
1596 if (cnt == 0)
1597 goto out;
1598 first = (first + ncopies) % disks;
1599 } while (first != 0);
1600 has_enough = 1;
1601out:
1602 rcu_read_unlock();
1603 return has_enough;
1604}
1605
1606static int enough(struct r10conf *conf, int ignore)
1607{
1608 /* when calling 'enough', both 'prev' and 'geo' must
1609 * be stable.
1610 * This is ensured if ->reconfig_mutex or ->device_lock
1611 * is held.
1612 */
1613 return _enough(conf, 0, ignore) &&
1614 _enough(conf, 1, ignore);
1615}
1616
1617static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1618{
1619 char b[BDEVNAME_SIZE];
1620 struct r10conf *conf = mddev->private;
1621 unsigned long flags;
1622
1623 /*
1624 * If it is not operational, then we have already marked it as dead
1625 * else if it is the last working disks with "fail_last_dev == false",
1626 * ignore the error, let the next level up know.
1627 * else mark the drive as failed
1628 */
1629 spin_lock_irqsave(&conf->device_lock, flags);
1630 if (test_bit(In_sync, &rdev->flags) && !mddev->fail_last_dev
1631 && !enough(conf, rdev->raid_disk)) {
1632 /*
1633 * Don't fail the drive, just return an IO error.
1634 */
1635 spin_unlock_irqrestore(&conf->device_lock, flags);
1636 return;
1637 }
1638 if (test_and_clear_bit(In_sync, &rdev->flags))
1639 mddev->degraded++;
1640 /*
1641 * If recovery is running, make sure it aborts.
1642 */
1643 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1644 set_bit(Blocked, &rdev->flags);
1645 set_bit(Faulty, &rdev->flags);
1646 set_mask_bits(&mddev->sb_flags, 0,
1647 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1648 spin_unlock_irqrestore(&conf->device_lock, flags);
1649 pr_crit("md/raid10:%s: Disk failure on %s, disabling device.\n"
1650 "md/raid10:%s: Operation continuing on %d devices.\n",
1651 mdname(mddev), bdevname(rdev->bdev, b),
1652 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1653}
1654
1655static void print_conf(struct r10conf *conf)
1656{
1657 int i;
1658 struct md_rdev *rdev;
1659
1660 pr_debug("RAID10 conf printout:\n");
1661 if (!conf) {
1662 pr_debug("(!conf)\n");
1663 return;
1664 }
1665 pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1666 conf->geo.raid_disks);
1667
1668 /* This is only called with ->reconfix_mutex held, so
1669 * rcu protection of rdev is not needed */
1670 for (i = 0; i < conf->geo.raid_disks; i++) {
1671 char b[BDEVNAME_SIZE];
1672 rdev = conf->mirrors[i].rdev;
1673 if (rdev)
1674 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1675 i, !test_bit(In_sync, &rdev->flags),
1676 !test_bit(Faulty, &rdev->flags),
1677 bdevname(rdev->bdev,b));
1678 }
1679}
1680
1681static void close_sync(struct r10conf *conf)
1682{
1683 wait_barrier(conf);
1684 allow_barrier(conf);
1685
1686 mempool_exit(&conf->r10buf_pool);
1687}
1688
1689static int raid10_spare_active(struct mddev *mddev)
1690{
1691 int i;
1692 struct r10conf *conf = mddev->private;
1693 struct raid10_info *tmp;
1694 int count = 0;
1695 unsigned long flags;
1696
1697 /*
1698 * Find all non-in_sync disks within the RAID10 configuration
1699 * and mark them in_sync
1700 */
1701 for (i = 0; i < conf->geo.raid_disks; i++) {
1702 tmp = conf->mirrors + i;
1703 if (tmp->replacement
1704 && tmp->replacement->recovery_offset == MaxSector
1705 && !test_bit(Faulty, &tmp->replacement->flags)
1706 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1707 /* Replacement has just become active */
1708 if (!tmp->rdev
1709 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1710 count++;
1711 if (tmp->rdev) {
1712 /* Replaced device not technically faulty,
1713 * but we need to be sure it gets removed
1714 * and never re-added.
1715 */
1716 set_bit(Faulty, &tmp->rdev->flags);
1717 sysfs_notify_dirent_safe(
1718 tmp->rdev->sysfs_state);
1719 }
1720 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1721 } else if (tmp->rdev
1722 && tmp->rdev->recovery_offset == MaxSector
1723 && !test_bit(Faulty, &tmp->rdev->flags)
1724 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1725 count++;
1726 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1727 }
1728 }
1729 spin_lock_irqsave(&conf->device_lock, flags);
1730 mddev->degraded -= count;
1731 spin_unlock_irqrestore(&conf->device_lock, flags);
1732
1733 print_conf(conf);
1734 return count;
1735}
1736
1737static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1738{
1739 struct r10conf *conf = mddev->private;
1740 int err = -EEXIST;
1741 int mirror;
1742 int first = 0;
1743 int last = conf->geo.raid_disks - 1;
1744
1745 if (mddev->recovery_cp < MaxSector)
1746 /* only hot-add to in-sync arrays, as recovery is
1747 * very different from resync
1748 */
1749 return -EBUSY;
1750 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1751 return -EINVAL;
1752
1753 if (md_integrity_add_rdev(rdev, mddev))
1754 return -ENXIO;
1755
1756 if (rdev->raid_disk >= 0)
1757 first = last = rdev->raid_disk;
1758
1759 if (rdev->saved_raid_disk >= first &&
1760 rdev->saved_raid_disk < conf->geo.raid_disks &&
1761 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1762 mirror = rdev->saved_raid_disk;
1763 else
1764 mirror = first;
1765 for ( ; mirror <= last ; mirror++) {
1766 struct raid10_info *p = &conf->mirrors[mirror];
1767 if (p->recovery_disabled == mddev->recovery_disabled)
1768 continue;
1769 if (p->rdev) {
1770 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1771 p->replacement != NULL)
1772 continue;
1773 clear_bit(In_sync, &rdev->flags);
1774 set_bit(Replacement, &rdev->flags);
1775 rdev->raid_disk = mirror;
1776 err = 0;
1777 if (mddev->gendisk)
1778 disk_stack_limits(mddev->gendisk, rdev->bdev,
1779 rdev->data_offset << 9);
1780 conf->fullsync = 1;
1781 rcu_assign_pointer(p->replacement, rdev);
1782 break;
1783 }
1784
1785 if (mddev->gendisk)
1786 disk_stack_limits(mddev->gendisk, rdev->bdev,
1787 rdev->data_offset << 9);
1788
1789 p->head_position = 0;
1790 p->recovery_disabled = mddev->recovery_disabled - 1;
1791 rdev->raid_disk = mirror;
1792 err = 0;
1793 if (rdev->saved_raid_disk != mirror)
1794 conf->fullsync = 1;
1795 rcu_assign_pointer(p->rdev, rdev);
1796 break;
1797 }
1798 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1799 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1800
1801 print_conf(conf);
1802 return err;
1803}
1804
1805static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1806{
1807 struct r10conf *conf = mddev->private;
1808 int err = 0;
1809 int number = rdev->raid_disk;
1810 struct md_rdev **rdevp;
1811 struct raid10_info *p = conf->mirrors + number;
1812
1813 print_conf(conf);
1814 if (rdev == p->rdev)
1815 rdevp = &p->rdev;
1816 else if (rdev == p->replacement)
1817 rdevp = &p->replacement;
1818 else
1819 return 0;
1820
1821 if (test_bit(In_sync, &rdev->flags) ||
1822 atomic_read(&rdev->nr_pending)) {
1823 err = -EBUSY;
1824 goto abort;
1825 }
1826 /* Only remove non-faulty devices if recovery
1827 * is not possible.
1828 */
1829 if (!test_bit(Faulty, &rdev->flags) &&
1830 mddev->recovery_disabled != p->recovery_disabled &&
1831 (!p->replacement || p->replacement == rdev) &&
1832 number < conf->geo.raid_disks &&
1833 enough(conf, -1)) {
1834 err = -EBUSY;
1835 goto abort;
1836 }
1837 *rdevp = NULL;
1838 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1839 synchronize_rcu();
1840 if (atomic_read(&rdev->nr_pending)) {
1841 /* lost the race, try later */
1842 err = -EBUSY;
1843 *rdevp = rdev;
1844 goto abort;
1845 }
1846 }
1847 if (p->replacement) {
1848 /* We must have just cleared 'rdev' */
1849 p->rdev = p->replacement;
1850 clear_bit(Replacement, &p->replacement->flags);
1851 smp_mb(); /* Make sure other CPUs may see both as identical
1852 * but will never see neither -- if they are careful.
1853 */
1854 p->replacement = NULL;
1855 }
1856
1857 clear_bit(WantReplacement, &rdev->flags);
1858 err = md_integrity_register(mddev);
1859
1860abort:
1861
1862 print_conf(conf);
1863 return err;
1864}
1865
1866static void __end_sync_read(struct r10bio *r10_bio, struct bio *bio, int d)
1867{
1868 struct r10conf *conf = r10_bio->mddev->private;
1869
1870 if (!bio->bi_status)
1871 set_bit(R10BIO_Uptodate, &r10_bio->state);
1872 else
1873 /* The write handler will notice the lack of
1874 * R10BIO_Uptodate and record any errors etc
1875 */
1876 atomic_add(r10_bio->sectors,
1877 &conf->mirrors[d].rdev->corrected_errors);
1878
1879 /* for reconstruct, we always reschedule after a read.
1880 * for resync, only after all reads
1881 */
1882 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1883 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1884 atomic_dec_and_test(&r10_bio->remaining)) {
1885 /* we have read all the blocks,
1886 * do the comparison in process context in raid10d
1887 */
1888 reschedule_retry(r10_bio);
1889 }
1890}
1891
1892static void end_sync_read(struct bio *bio)
1893{
1894 struct r10bio *r10_bio = get_resync_r10bio(bio);
1895 struct r10conf *conf = r10_bio->mddev->private;
1896 int d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1897
1898 __end_sync_read(r10_bio, bio, d);
1899}
1900
1901static void end_reshape_read(struct bio *bio)
1902{
1903 /* reshape read bio isn't allocated from r10buf_pool */
1904 struct r10bio *r10_bio = bio->bi_private;
1905
1906 __end_sync_read(r10_bio, bio, r10_bio->read_slot);
1907}
1908
1909static void end_sync_request(struct r10bio *r10_bio)
1910{
1911 struct mddev *mddev = r10_bio->mddev;
1912
1913 while (atomic_dec_and_test(&r10_bio->remaining)) {
1914 if (r10_bio->master_bio == NULL) {
1915 /* the primary of several recovery bios */
1916 sector_t s = r10_bio->sectors;
1917 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1918 test_bit(R10BIO_WriteError, &r10_bio->state))
1919 reschedule_retry(r10_bio);
1920 else
1921 put_buf(r10_bio);
1922 md_done_sync(mddev, s, 1);
1923 break;
1924 } else {
1925 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1926 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1927 test_bit(R10BIO_WriteError, &r10_bio->state))
1928 reschedule_retry(r10_bio);
1929 else
1930 put_buf(r10_bio);
1931 r10_bio = r10_bio2;
1932 }
1933 }
1934}
1935
1936static void end_sync_write(struct bio *bio)
1937{
1938 struct r10bio *r10_bio = get_resync_r10bio(bio);
1939 struct mddev *mddev = r10_bio->mddev;
1940 struct r10conf *conf = mddev->private;
1941 int d;
1942 sector_t first_bad;
1943 int bad_sectors;
1944 int slot;
1945 int repl;
1946 struct md_rdev *rdev = NULL;
1947
1948 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1949 if (repl)
1950 rdev = conf->mirrors[d].replacement;
1951 else
1952 rdev = conf->mirrors[d].rdev;
1953
1954 if (bio->bi_status) {
1955 if (repl)
1956 md_error(mddev, rdev);
1957 else {
1958 set_bit(WriteErrorSeen, &rdev->flags);
1959 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1960 set_bit(MD_RECOVERY_NEEDED,
1961 &rdev->mddev->recovery);
1962 set_bit(R10BIO_WriteError, &r10_bio->state);
1963 }
1964 } else if (is_badblock(rdev,
1965 r10_bio->devs[slot].addr,
1966 r10_bio->sectors,
1967 &first_bad, &bad_sectors))
1968 set_bit(R10BIO_MadeGood, &r10_bio->state);
1969
1970 rdev_dec_pending(rdev, mddev);
1971
1972 end_sync_request(r10_bio);
1973}
1974
1975/*
1976 * Note: sync and recover and handled very differently for raid10
1977 * This code is for resync.
1978 * For resync, we read through virtual addresses and read all blocks.
1979 * If there is any error, we schedule a write. The lowest numbered
1980 * drive is authoritative.
1981 * However requests come for physical address, so we need to map.
1982 * For every physical address there are raid_disks/copies virtual addresses,
1983 * which is always are least one, but is not necessarly an integer.
1984 * This means that a physical address can span multiple chunks, so we may
1985 * have to submit multiple io requests for a single sync request.
1986 */
1987/*
1988 * We check if all blocks are in-sync and only write to blocks that
1989 * aren't in sync
1990 */
1991static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1992{
1993 struct r10conf *conf = mddev->private;
1994 int i, first;
1995 struct bio *tbio, *fbio;
1996 int vcnt;
1997 struct page **tpages, **fpages;
1998
1999 atomic_set(&r10_bio->remaining, 1);
2000
2001 /* find the first device with a block */
2002 for (i=0; i<conf->copies; i++)
2003 if (!r10_bio->devs[i].bio->bi_status)
2004 break;
2005
2006 if (i == conf->copies)
2007 goto done;
2008
2009 first = i;
2010 fbio = r10_bio->devs[i].bio;
2011 fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2012 fbio->bi_iter.bi_idx = 0;
2013 fpages = get_resync_pages(fbio)->pages;
2014
2015 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2016 /* now find blocks with errors */
2017 for (i=0 ; i < conf->copies ; i++) {
2018 int j, d;
2019 struct md_rdev *rdev;
2020 struct resync_pages *rp;
2021
2022 tbio = r10_bio->devs[i].bio;
2023
2024 if (tbio->bi_end_io != end_sync_read)
2025 continue;
2026 if (i == first)
2027 continue;
2028
2029 tpages = get_resync_pages(tbio)->pages;
2030 d = r10_bio->devs[i].devnum;
2031 rdev = conf->mirrors[d].rdev;
2032 if (!r10_bio->devs[i].bio->bi_status) {
2033 /* We know that the bi_io_vec layout is the same for
2034 * both 'first' and 'i', so we just compare them.
2035 * All vec entries are PAGE_SIZE;
2036 */
2037 int sectors = r10_bio->sectors;
2038 for (j = 0; j < vcnt; j++) {
2039 int len = PAGE_SIZE;
2040 if (sectors < (len / 512))
2041 len = sectors * 512;
2042 if (memcmp(page_address(fpages[j]),
2043 page_address(tpages[j]),
2044 len))
2045 break;
2046 sectors -= len/512;
2047 }
2048 if (j == vcnt)
2049 continue;
2050 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2051 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2052 /* Don't fix anything. */
2053 continue;
2054 } else if (test_bit(FailFast, &rdev->flags)) {
2055 /* Just give up on this device */
2056 md_error(rdev->mddev, rdev);
2057 continue;
2058 }
2059 /* Ok, we need to write this bio, either to correct an
2060 * inconsistency or to correct an unreadable block.
2061 * First we need to fixup bv_offset, bv_len and
2062 * bi_vecs, as the read request might have corrupted these
2063 */
2064 rp = get_resync_pages(tbio);
2065 bio_reset(tbio);
2066
2067 md_bio_reset_resync_pages(tbio, rp, fbio->bi_iter.bi_size);
2068
2069 rp->raid_bio = r10_bio;
2070 tbio->bi_private = rp;
2071 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2072 tbio->bi_end_io = end_sync_write;
2073 bio_set_op_attrs(tbio, REQ_OP_WRITE, 0);
2074
2075 bio_copy_data(tbio, fbio);
2076
2077 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2078 atomic_inc(&r10_bio->remaining);
2079 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2080
2081 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2082 tbio->bi_opf |= MD_FAILFAST;
2083 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2084 bio_set_dev(tbio, conf->mirrors[d].rdev->bdev);
2085 submit_bio_noacct(tbio);
2086 }
2087
2088 /* Now write out to any replacement devices
2089 * that are active
2090 */
2091 for (i = 0; i < conf->copies; i++) {
2092 int d;
2093
2094 tbio = r10_bio->devs[i].repl_bio;
2095 if (!tbio || !tbio->bi_end_io)
2096 continue;
2097 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2098 && r10_bio->devs[i].bio != fbio)
2099 bio_copy_data(tbio, fbio);
2100 d = r10_bio->devs[i].devnum;
2101 atomic_inc(&r10_bio->remaining);
2102 md_sync_acct(conf->mirrors[d].replacement->bdev,
2103 bio_sectors(tbio));
2104 submit_bio_noacct(tbio);
2105 }
2106
2107done:
2108 if (atomic_dec_and_test(&r10_bio->remaining)) {
2109 md_done_sync(mddev, r10_bio->sectors, 1);
2110 put_buf(r10_bio);
2111 }
2112}
2113
2114/*
2115 * Now for the recovery code.
2116 * Recovery happens across physical sectors.
2117 * We recover all non-is_sync drives by finding the virtual address of
2118 * each, and then choose a working drive that also has that virt address.
2119 * There is a separate r10_bio for each non-in_sync drive.
2120 * Only the first two slots are in use. The first for reading,
2121 * The second for writing.
2122 *
2123 */
2124static void fix_recovery_read_error(struct r10bio *r10_bio)
2125{
2126 /* We got a read error during recovery.
2127 * We repeat the read in smaller page-sized sections.
2128 * If a read succeeds, write it to the new device or record
2129 * a bad block if we cannot.
2130 * If a read fails, record a bad block on both old and
2131 * new devices.
2132 */
2133 struct mddev *mddev = r10_bio->mddev;
2134 struct r10conf *conf = mddev->private;
2135 struct bio *bio = r10_bio->devs[0].bio;
2136 sector_t sect = 0;
2137 int sectors = r10_bio->sectors;
2138 int idx = 0;
2139 int dr = r10_bio->devs[0].devnum;
2140 int dw = r10_bio->devs[1].devnum;
2141 struct page **pages = get_resync_pages(bio)->pages;
2142
2143 while (sectors) {
2144 int s = sectors;
2145 struct md_rdev *rdev;
2146 sector_t addr;
2147 int ok;
2148
2149 if (s > (PAGE_SIZE>>9))
2150 s = PAGE_SIZE >> 9;
2151
2152 rdev = conf->mirrors[dr].rdev;
2153 addr = r10_bio->devs[0].addr + sect,
2154 ok = sync_page_io(rdev,
2155 addr,
2156 s << 9,
2157 pages[idx],
2158 REQ_OP_READ, 0, false);
2159 if (ok) {
2160 rdev = conf->mirrors[dw].rdev;
2161 addr = r10_bio->devs[1].addr + sect;
2162 ok = sync_page_io(rdev,
2163 addr,
2164 s << 9,
2165 pages[idx],
2166 REQ_OP_WRITE, 0, false);
2167 if (!ok) {
2168 set_bit(WriteErrorSeen, &rdev->flags);
2169 if (!test_and_set_bit(WantReplacement,
2170 &rdev->flags))
2171 set_bit(MD_RECOVERY_NEEDED,
2172 &rdev->mddev->recovery);
2173 }
2174 }
2175 if (!ok) {
2176 /* We don't worry if we cannot set a bad block -
2177 * it really is bad so there is no loss in not
2178 * recording it yet
2179 */
2180 rdev_set_badblocks(rdev, addr, s, 0);
2181
2182 if (rdev != conf->mirrors[dw].rdev) {
2183 /* need bad block on destination too */
2184 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2185 addr = r10_bio->devs[1].addr + sect;
2186 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2187 if (!ok) {
2188 /* just abort the recovery */
2189 pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2190 mdname(mddev));
2191
2192 conf->mirrors[dw].recovery_disabled
2193 = mddev->recovery_disabled;
2194 set_bit(MD_RECOVERY_INTR,
2195 &mddev->recovery);
2196 break;
2197 }
2198 }
2199 }
2200
2201 sectors -= s;
2202 sect += s;
2203 idx++;
2204 }
2205}
2206
2207static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2208{
2209 struct r10conf *conf = mddev->private;
2210 int d;
2211 struct bio *wbio, *wbio2;
2212
2213 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2214 fix_recovery_read_error(r10_bio);
2215 end_sync_request(r10_bio);
2216 return;
2217 }
2218
2219 /*
2220 * share the pages with the first bio
2221 * and submit the write request
2222 */
2223 d = r10_bio->devs[1].devnum;
2224 wbio = r10_bio->devs[1].bio;
2225 wbio2 = r10_bio->devs[1].repl_bio;
2226 /* Need to test wbio2->bi_end_io before we call
2227 * submit_bio_noacct as if the former is NULL,
2228 * the latter is free to free wbio2.
2229 */
2230 if (wbio2 && !wbio2->bi_end_io)
2231 wbio2 = NULL;
2232 if (wbio->bi_end_io) {
2233 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2234 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2235 submit_bio_noacct(wbio);
2236 }
2237 if (wbio2) {
2238 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2239 md_sync_acct(conf->mirrors[d].replacement->bdev,
2240 bio_sectors(wbio2));
2241 submit_bio_noacct(wbio2);
2242 }
2243}
2244
2245/*
2246 * Used by fix_read_error() to decay the per rdev read_errors.
2247 * We halve the read error count for every hour that has elapsed
2248 * since the last recorded read error.
2249 *
2250 */
2251static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2252{
2253 long cur_time_mon;
2254 unsigned long hours_since_last;
2255 unsigned int read_errors = atomic_read(&rdev->read_errors);
2256
2257 cur_time_mon = ktime_get_seconds();
2258
2259 if (rdev->last_read_error == 0) {
2260 /* first time we've seen a read error */
2261 rdev->last_read_error = cur_time_mon;
2262 return;
2263 }
2264
2265 hours_since_last = (long)(cur_time_mon -
2266 rdev->last_read_error) / 3600;
2267
2268 rdev->last_read_error = cur_time_mon;
2269
2270 /*
2271 * if hours_since_last is > the number of bits in read_errors
2272 * just set read errors to 0. We do this to avoid
2273 * overflowing the shift of read_errors by hours_since_last.
2274 */
2275 if (hours_since_last >= 8 * sizeof(read_errors))
2276 atomic_set(&rdev->read_errors, 0);
2277 else
2278 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2279}
2280
2281static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2282 int sectors, struct page *page, int rw)
2283{
2284 sector_t first_bad;
2285 int bad_sectors;
2286
2287 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2288 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2289 return -1;
2290 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
2291 /* success */
2292 return 1;
2293 if (rw == WRITE) {
2294 set_bit(WriteErrorSeen, &rdev->flags);
2295 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2296 set_bit(MD_RECOVERY_NEEDED,
2297 &rdev->mddev->recovery);
2298 }
2299 /* need to record an error - either for the block or the device */
2300 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2301 md_error(rdev->mddev, rdev);
2302 return 0;
2303}
2304
2305/*
2306 * This is a kernel thread which:
2307 *
2308 * 1. Retries failed read operations on working mirrors.
2309 * 2. Updates the raid superblock when problems encounter.
2310 * 3. Performs writes following reads for array synchronising.
2311 */
2312
2313static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2314{
2315 int sect = 0; /* Offset from r10_bio->sector */
2316 int sectors = r10_bio->sectors;
2317 struct md_rdev *rdev;
2318 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2319 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2320
2321 /* still own a reference to this rdev, so it cannot
2322 * have been cleared recently.
2323 */
2324 rdev = conf->mirrors[d].rdev;
2325
2326 if (test_bit(Faulty, &rdev->flags))
2327 /* drive has already been failed, just ignore any
2328 more fix_read_error() attempts */
2329 return;
2330
2331 check_decay_read_errors(mddev, rdev);
2332 atomic_inc(&rdev->read_errors);
2333 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2334 char b[BDEVNAME_SIZE];
2335 bdevname(rdev->bdev, b);
2336
2337 pr_notice("md/raid10:%s: %s: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2338 mdname(mddev), b,
2339 atomic_read(&rdev->read_errors), max_read_errors);
2340 pr_notice("md/raid10:%s: %s: Failing raid device\n",
2341 mdname(mddev), b);
2342 md_error(mddev, rdev);
2343 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2344 return;
2345 }
2346
2347 while(sectors) {
2348 int s = sectors;
2349 int sl = r10_bio->read_slot;
2350 int success = 0;
2351 int start;
2352
2353 if (s > (PAGE_SIZE>>9))
2354 s = PAGE_SIZE >> 9;
2355
2356 rcu_read_lock();
2357 do {
2358 sector_t first_bad;
2359 int bad_sectors;
2360
2361 d = r10_bio->devs[sl].devnum;
2362 rdev = rcu_dereference(conf->mirrors[d].rdev);
2363 if (rdev &&
2364 test_bit(In_sync, &rdev->flags) &&
2365 !test_bit(Faulty, &rdev->flags) &&
2366 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2367 &first_bad, &bad_sectors) == 0) {
2368 atomic_inc(&rdev->nr_pending);
2369 rcu_read_unlock();
2370 success = sync_page_io(rdev,
2371 r10_bio->devs[sl].addr +
2372 sect,
2373 s<<9,
2374 conf->tmppage,
2375 REQ_OP_READ, 0, false);
2376 rdev_dec_pending(rdev, mddev);
2377 rcu_read_lock();
2378 if (success)
2379 break;
2380 }
2381 sl++;
2382 if (sl == conf->copies)
2383 sl = 0;
2384 } while (!success && sl != r10_bio->read_slot);
2385 rcu_read_unlock();
2386
2387 if (!success) {
2388 /* Cannot read from anywhere, just mark the block
2389 * as bad on the first device to discourage future
2390 * reads.
2391 */
2392 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2393 rdev = conf->mirrors[dn].rdev;
2394
2395 if (!rdev_set_badblocks(
2396 rdev,
2397 r10_bio->devs[r10_bio->read_slot].addr
2398 + sect,
2399 s, 0)) {
2400 md_error(mddev, rdev);
2401 r10_bio->devs[r10_bio->read_slot].bio
2402 = IO_BLOCKED;
2403 }
2404 break;
2405 }
2406
2407 start = sl;
2408 /* write it back and re-read */
2409 rcu_read_lock();
2410 while (sl != r10_bio->read_slot) {
2411 char b[BDEVNAME_SIZE];
2412
2413 if (sl==0)
2414 sl = conf->copies;
2415 sl--;
2416 d = r10_bio->devs[sl].devnum;
2417 rdev = rcu_dereference(conf->mirrors[d].rdev);
2418 if (!rdev ||
2419 test_bit(Faulty, &rdev->flags) ||
2420 !test_bit(In_sync, &rdev->flags))
2421 continue;
2422
2423 atomic_inc(&rdev->nr_pending);
2424 rcu_read_unlock();
2425 if (r10_sync_page_io(rdev,
2426 r10_bio->devs[sl].addr +
2427 sect,
2428 s, conf->tmppage, WRITE)
2429 == 0) {
2430 /* Well, this device is dead */
2431 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %s)\n",
2432 mdname(mddev), s,
2433 (unsigned long long)(
2434 sect +
2435 choose_data_offset(r10_bio,
2436 rdev)),
2437 bdevname(rdev->bdev, b));
2438 pr_notice("md/raid10:%s: %s: failing drive\n",
2439 mdname(mddev),
2440 bdevname(rdev->bdev, b));
2441 }
2442 rdev_dec_pending(rdev, mddev);
2443 rcu_read_lock();
2444 }
2445 sl = start;
2446 while (sl != r10_bio->read_slot) {
2447 char b[BDEVNAME_SIZE];
2448
2449 if (sl==0)
2450 sl = conf->copies;
2451 sl--;
2452 d = r10_bio->devs[sl].devnum;
2453 rdev = rcu_dereference(conf->mirrors[d].rdev);
2454 if (!rdev ||
2455 test_bit(Faulty, &rdev->flags) ||
2456 !test_bit(In_sync, &rdev->flags))
2457 continue;
2458
2459 atomic_inc(&rdev->nr_pending);
2460 rcu_read_unlock();
2461 switch (r10_sync_page_io(rdev,
2462 r10_bio->devs[sl].addr +
2463 sect,
2464 s, conf->tmppage,
2465 READ)) {
2466 case 0:
2467 /* Well, this device is dead */
2468 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %s)\n",
2469 mdname(mddev), s,
2470 (unsigned long long)(
2471 sect +
2472 choose_data_offset(r10_bio, rdev)),
2473 bdevname(rdev->bdev, b));
2474 pr_notice("md/raid10:%s: %s: failing drive\n",
2475 mdname(mddev),
2476 bdevname(rdev->bdev, b));
2477 break;
2478 case 1:
2479 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %s)\n",
2480 mdname(mddev), s,
2481 (unsigned long long)(
2482 sect +
2483 choose_data_offset(r10_bio, rdev)),
2484 bdevname(rdev->bdev, b));
2485 atomic_add(s, &rdev->corrected_errors);
2486 }
2487
2488 rdev_dec_pending(rdev, mddev);
2489 rcu_read_lock();
2490 }
2491 rcu_read_unlock();
2492
2493 sectors -= s;
2494 sect += s;
2495 }
2496}
2497
2498static int narrow_write_error(struct r10bio *r10_bio, int i)
2499{
2500 struct bio *bio = r10_bio->master_bio;
2501 struct mddev *mddev = r10_bio->mddev;
2502 struct r10conf *conf = mddev->private;
2503 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2504 /* bio has the data to be written to slot 'i' where
2505 * we just recently had a write error.
2506 * We repeatedly clone the bio and trim down to one block,
2507 * then try the write. Where the write fails we record
2508 * a bad block.
2509 * It is conceivable that the bio doesn't exactly align with
2510 * blocks. We must handle this.
2511 *
2512 * We currently own a reference to the rdev.
2513 */
2514
2515 int block_sectors;
2516 sector_t sector;
2517 int sectors;
2518 int sect_to_write = r10_bio->sectors;
2519 int ok = 1;
2520
2521 if (rdev->badblocks.shift < 0)
2522 return 0;
2523
2524 block_sectors = roundup(1 << rdev->badblocks.shift,
2525 bdev_logical_block_size(rdev->bdev) >> 9);
2526 sector = r10_bio->sector;
2527 sectors = ((r10_bio->sector + block_sectors)
2528 & ~(sector_t)(block_sectors - 1))
2529 - sector;
2530
2531 while (sect_to_write) {
2532 struct bio *wbio;
2533 sector_t wsector;
2534 if (sectors > sect_to_write)
2535 sectors = sect_to_write;
2536 /* Write at 'sector' for 'sectors' */
2537 wbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
2538 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2539 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2540 wbio->bi_iter.bi_sector = wsector +
2541 choose_data_offset(r10_bio, rdev);
2542 bio_set_dev(wbio, rdev->bdev);
2543 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2544
2545 if (submit_bio_wait(wbio) < 0)
2546 /* Failure! */
2547 ok = rdev_set_badblocks(rdev, wsector,
2548 sectors, 0)
2549 && ok;
2550
2551 bio_put(wbio);
2552 sect_to_write -= sectors;
2553 sector += sectors;
2554 sectors = block_sectors;
2555 }
2556 return ok;
2557}
2558
2559static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2560{
2561 int slot = r10_bio->read_slot;
2562 struct bio *bio;
2563 struct r10conf *conf = mddev->private;
2564 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2565
2566 /* we got a read error. Maybe the drive is bad. Maybe just
2567 * the block and we can fix it.
2568 * We freeze all other IO, and try reading the block from
2569 * other devices. When we find one, we re-write
2570 * and check it that fixes the read error.
2571 * This is all done synchronously while the array is
2572 * frozen.
2573 */
2574 bio = r10_bio->devs[slot].bio;
2575 bio_put(bio);
2576 r10_bio->devs[slot].bio = NULL;
2577
2578 if (mddev->ro)
2579 r10_bio->devs[slot].bio = IO_BLOCKED;
2580 else if (!test_bit(FailFast, &rdev->flags)) {
2581 freeze_array(conf, 1);
2582 fix_read_error(conf, mddev, r10_bio);
2583 unfreeze_array(conf);
2584 } else
2585 md_error(mddev, rdev);
2586
2587 rdev_dec_pending(rdev, mddev);
2588 allow_barrier(conf);
2589 r10_bio->state = 0;
2590 raid10_read_request(mddev, r10_bio->master_bio, r10_bio);
2591}
2592
2593static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2594{
2595 /* Some sort of write request has finished and it
2596 * succeeded in writing where we thought there was a
2597 * bad block. So forget the bad block.
2598 * Or possibly if failed and we need to record
2599 * a bad block.
2600 */
2601 int m;
2602 struct md_rdev *rdev;
2603
2604 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2605 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2606 for (m = 0; m < conf->copies; m++) {
2607 int dev = r10_bio->devs[m].devnum;
2608 rdev = conf->mirrors[dev].rdev;
2609 if (r10_bio->devs[m].bio == NULL ||
2610 r10_bio->devs[m].bio->bi_end_io == NULL)
2611 continue;
2612 if (!r10_bio->devs[m].bio->bi_status) {
2613 rdev_clear_badblocks(
2614 rdev,
2615 r10_bio->devs[m].addr,
2616 r10_bio->sectors, 0);
2617 } else {
2618 if (!rdev_set_badblocks(
2619 rdev,
2620 r10_bio->devs[m].addr,
2621 r10_bio->sectors, 0))
2622 md_error(conf->mddev, rdev);
2623 }
2624 rdev = conf->mirrors[dev].replacement;
2625 if (r10_bio->devs[m].repl_bio == NULL ||
2626 r10_bio->devs[m].repl_bio->bi_end_io == NULL)
2627 continue;
2628
2629 if (!r10_bio->devs[m].repl_bio->bi_status) {
2630 rdev_clear_badblocks(
2631 rdev,
2632 r10_bio->devs[m].addr,
2633 r10_bio->sectors, 0);
2634 } else {
2635 if (!rdev_set_badblocks(
2636 rdev,
2637 r10_bio->devs[m].addr,
2638 r10_bio->sectors, 0))
2639 md_error(conf->mddev, rdev);
2640 }
2641 }
2642 put_buf(r10_bio);
2643 } else {
2644 bool fail = false;
2645 for (m = 0; m < conf->copies; m++) {
2646 int dev = r10_bio->devs[m].devnum;
2647 struct bio *bio = r10_bio->devs[m].bio;
2648 rdev = conf->mirrors[dev].rdev;
2649 if (bio == IO_MADE_GOOD) {
2650 rdev_clear_badblocks(
2651 rdev,
2652 r10_bio->devs[m].addr,
2653 r10_bio->sectors, 0);
2654 rdev_dec_pending(rdev, conf->mddev);
2655 } else if (bio != NULL && bio->bi_status) {
2656 fail = true;
2657 if (!narrow_write_error(r10_bio, m)) {
2658 md_error(conf->mddev, rdev);
2659 set_bit(R10BIO_Degraded,
2660 &r10_bio->state);
2661 }
2662 rdev_dec_pending(rdev, conf->mddev);
2663 }
2664 bio = r10_bio->devs[m].repl_bio;
2665 rdev = conf->mirrors[dev].replacement;
2666 if (rdev && bio == IO_MADE_GOOD) {
2667 rdev_clear_badblocks(
2668 rdev,
2669 r10_bio->devs[m].addr,
2670 r10_bio->sectors, 0);
2671 rdev_dec_pending(rdev, conf->mddev);
2672 }
2673 }
2674 if (fail) {
2675 spin_lock_irq(&conf->device_lock);
2676 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2677 conf->nr_queued++;
2678 spin_unlock_irq(&conf->device_lock);
2679 /*
2680 * In case freeze_array() is waiting for condition
2681 * nr_pending == nr_queued + extra to be true.
2682 */
2683 wake_up(&conf->wait_barrier);
2684 md_wakeup_thread(conf->mddev->thread);
2685 } else {
2686 if (test_bit(R10BIO_WriteError,
2687 &r10_bio->state))
2688 close_write(r10_bio);
2689 raid_end_bio_io(r10_bio);
2690 }
2691 }
2692}
2693
2694static void raid10d(struct md_thread *thread)
2695{
2696 struct mddev *mddev = thread->mddev;
2697 struct r10bio *r10_bio;
2698 unsigned long flags;
2699 struct r10conf *conf = mddev->private;
2700 struct list_head *head = &conf->retry_list;
2701 struct blk_plug plug;
2702
2703 md_check_recovery(mddev);
2704
2705 if (!list_empty_careful(&conf->bio_end_io_list) &&
2706 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2707 LIST_HEAD(tmp);
2708 spin_lock_irqsave(&conf->device_lock, flags);
2709 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2710 while (!list_empty(&conf->bio_end_io_list)) {
2711 list_move(conf->bio_end_io_list.prev, &tmp);
2712 conf->nr_queued--;
2713 }
2714 }
2715 spin_unlock_irqrestore(&conf->device_lock, flags);
2716 while (!list_empty(&tmp)) {
2717 r10_bio = list_first_entry(&tmp, struct r10bio,
2718 retry_list);
2719 list_del(&r10_bio->retry_list);
2720 if (mddev->degraded)
2721 set_bit(R10BIO_Degraded, &r10_bio->state);
2722
2723 if (test_bit(R10BIO_WriteError,
2724 &r10_bio->state))
2725 close_write(r10_bio);
2726 raid_end_bio_io(r10_bio);
2727 }
2728 }
2729
2730 blk_start_plug(&plug);
2731 for (;;) {
2732
2733 flush_pending_writes(conf);
2734
2735 spin_lock_irqsave(&conf->device_lock, flags);
2736 if (list_empty(head)) {
2737 spin_unlock_irqrestore(&conf->device_lock, flags);
2738 break;
2739 }
2740 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2741 list_del(head->prev);
2742 conf->nr_queued--;
2743 spin_unlock_irqrestore(&conf->device_lock, flags);
2744
2745 mddev = r10_bio->mddev;
2746 conf = mddev->private;
2747 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2748 test_bit(R10BIO_WriteError, &r10_bio->state))
2749 handle_write_completed(conf, r10_bio);
2750 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2751 reshape_request_write(mddev, r10_bio);
2752 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2753 sync_request_write(mddev, r10_bio);
2754 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2755 recovery_request_write(mddev, r10_bio);
2756 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2757 handle_read_error(mddev, r10_bio);
2758 else
2759 WARN_ON_ONCE(1);
2760
2761 cond_resched();
2762 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2763 md_check_recovery(mddev);
2764 }
2765 blk_finish_plug(&plug);
2766}
2767
2768static int init_resync(struct r10conf *conf)
2769{
2770 int ret, buffs, i;
2771
2772 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2773 BUG_ON(mempool_initialized(&conf->r10buf_pool));
2774 conf->have_replacement = 0;
2775 for (i = 0; i < conf->geo.raid_disks; i++)
2776 if (conf->mirrors[i].replacement)
2777 conf->have_replacement = 1;
2778 ret = mempool_init(&conf->r10buf_pool, buffs,
2779 r10buf_pool_alloc, r10buf_pool_free, conf);
2780 if (ret)
2781 return ret;
2782 conf->next_resync = 0;
2783 return 0;
2784}
2785
2786static struct r10bio *raid10_alloc_init_r10buf(struct r10conf *conf)
2787{
2788 struct r10bio *r10bio = mempool_alloc(&conf->r10buf_pool, GFP_NOIO);
2789 struct rsync_pages *rp;
2790 struct bio *bio;
2791 int nalloc;
2792 int i;
2793
2794 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
2795 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
2796 nalloc = conf->copies; /* resync */
2797 else
2798 nalloc = 2; /* recovery */
2799
2800 for (i = 0; i < nalloc; i++) {
2801 bio = r10bio->devs[i].bio;
2802 rp = bio->bi_private;
2803 bio_reset(bio);
2804 bio->bi_private = rp;
2805 bio = r10bio->devs[i].repl_bio;
2806 if (bio) {
2807 rp = bio->bi_private;
2808 bio_reset(bio);
2809 bio->bi_private = rp;
2810 }
2811 }
2812 return r10bio;
2813}
2814
2815/*
2816 * Set cluster_sync_high since we need other nodes to add the
2817 * range [cluster_sync_low, cluster_sync_high] to suspend list.
2818 */
2819static void raid10_set_cluster_sync_high(struct r10conf *conf)
2820{
2821 sector_t window_size;
2822 int extra_chunk, chunks;
2823
2824 /*
2825 * First, here we define "stripe" as a unit which across
2826 * all member devices one time, so we get chunks by use
2827 * raid_disks / near_copies. Otherwise, if near_copies is
2828 * close to raid_disks, then resync window could increases
2829 * linearly with the increase of raid_disks, which means
2830 * we will suspend a really large IO window while it is not
2831 * necessary. If raid_disks is not divisible by near_copies,
2832 * an extra chunk is needed to ensure the whole "stripe" is
2833 * covered.
2834 */
2835
2836 chunks = conf->geo.raid_disks / conf->geo.near_copies;
2837 if (conf->geo.raid_disks % conf->geo.near_copies == 0)
2838 extra_chunk = 0;
2839 else
2840 extra_chunk = 1;
2841 window_size = (chunks + extra_chunk) * conf->mddev->chunk_sectors;
2842
2843 /*
2844 * At least use a 32M window to align with raid1's resync window
2845 */
2846 window_size = (CLUSTER_RESYNC_WINDOW_SECTORS > window_size) ?
2847 CLUSTER_RESYNC_WINDOW_SECTORS : window_size;
2848
2849 conf->cluster_sync_high = conf->cluster_sync_low + window_size;
2850}
2851
2852/*
2853 * perform a "sync" on one "block"
2854 *
2855 * We need to make sure that no normal I/O request - particularly write
2856 * requests - conflict with active sync requests.
2857 *
2858 * This is achieved by tracking pending requests and a 'barrier' concept
2859 * that can be installed to exclude normal IO requests.
2860 *
2861 * Resync and recovery are handled very differently.
2862 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2863 *
2864 * For resync, we iterate over virtual addresses, read all copies,
2865 * and update if there are differences. If only one copy is live,
2866 * skip it.
2867 * For recovery, we iterate over physical addresses, read a good
2868 * value for each non-in_sync drive, and over-write.
2869 *
2870 * So, for recovery we may have several outstanding complex requests for a
2871 * given address, one for each out-of-sync device. We model this by allocating
2872 * a number of r10_bio structures, one for each out-of-sync device.
2873 * As we setup these structures, we collect all bio's together into a list
2874 * which we then process collectively to add pages, and then process again
2875 * to pass to submit_bio_noacct.
2876 *
2877 * The r10_bio structures are linked using a borrowed master_bio pointer.
2878 * This link is counted in ->remaining. When the r10_bio that points to NULL
2879 * has its remaining count decremented to 0, the whole complex operation
2880 * is complete.
2881 *
2882 */
2883
2884static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2885 int *skipped)
2886{
2887 struct r10conf *conf = mddev->private;
2888 struct r10bio *r10_bio;
2889 struct bio *biolist = NULL, *bio;
2890 sector_t max_sector, nr_sectors;
2891 int i;
2892 int max_sync;
2893 sector_t sync_blocks;
2894 sector_t sectors_skipped = 0;
2895 int chunks_skipped = 0;
2896 sector_t chunk_mask = conf->geo.chunk_mask;
2897 int page_idx = 0;
2898
2899 if (!mempool_initialized(&conf->r10buf_pool))
2900 if (init_resync(conf))
2901 return 0;
2902
2903 /*
2904 * Allow skipping a full rebuild for incremental assembly
2905 * of a clean array, like RAID1 does.
2906 */
2907 if (mddev->bitmap == NULL &&
2908 mddev->recovery_cp == MaxSector &&
2909 mddev->reshape_position == MaxSector &&
2910 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2911 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2912 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2913 conf->fullsync == 0) {
2914 *skipped = 1;
2915 return mddev->dev_sectors - sector_nr;
2916 }
2917
2918 skipped:
2919 max_sector = mddev->dev_sectors;
2920 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2921 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2922 max_sector = mddev->resync_max_sectors;
2923 if (sector_nr >= max_sector) {
2924 conf->cluster_sync_low = 0;
2925 conf->cluster_sync_high = 0;
2926
2927 /* If we aborted, we need to abort the
2928 * sync on the 'current' bitmap chucks (there can
2929 * be several when recovering multiple devices).
2930 * as we may have started syncing it but not finished.
2931 * We can find the current address in
2932 * mddev->curr_resync, but for recovery,
2933 * we need to convert that to several
2934 * virtual addresses.
2935 */
2936 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2937 end_reshape(conf);
2938 close_sync(conf);
2939 return 0;
2940 }
2941
2942 if (mddev->curr_resync < max_sector) { /* aborted */
2943 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2944 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2945 &sync_blocks, 1);
2946 else for (i = 0; i < conf->geo.raid_disks; i++) {
2947 sector_t sect =
2948 raid10_find_virt(conf, mddev->curr_resync, i);
2949 md_bitmap_end_sync(mddev->bitmap, sect,
2950 &sync_blocks, 1);
2951 }
2952 } else {
2953 /* completed sync */
2954 if ((!mddev->bitmap || conf->fullsync)
2955 && conf->have_replacement
2956 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2957 /* Completed a full sync so the replacements
2958 * are now fully recovered.
2959 */
2960 rcu_read_lock();
2961 for (i = 0; i < conf->geo.raid_disks; i++) {
2962 struct md_rdev *rdev =
2963 rcu_dereference(conf->mirrors[i].replacement);
2964 if (rdev)
2965 rdev->recovery_offset = MaxSector;
2966 }
2967 rcu_read_unlock();
2968 }
2969 conf->fullsync = 0;
2970 }
2971 md_bitmap_close_sync(mddev->bitmap);
2972 close_sync(conf);
2973 *skipped = 1;
2974 return sectors_skipped;
2975 }
2976
2977 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2978 return reshape_request(mddev, sector_nr, skipped);
2979
2980 if (chunks_skipped >= conf->geo.raid_disks) {
2981 /* if there has been nothing to do on any drive,
2982 * then there is nothing to do at all..
2983 */
2984 *skipped = 1;
2985 return (max_sector - sector_nr) + sectors_skipped;
2986 }
2987
2988 if (max_sector > mddev->resync_max)
2989 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2990
2991 /* make sure whole request will fit in a chunk - if chunks
2992 * are meaningful
2993 */
2994 if (conf->geo.near_copies < conf->geo.raid_disks &&
2995 max_sector > (sector_nr | chunk_mask))
2996 max_sector = (sector_nr | chunk_mask) + 1;
2997
2998 /*
2999 * If there is non-resync activity waiting for a turn, then let it
3000 * though before starting on this new sync request.
3001 */
3002 if (conf->nr_waiting)
3003 schedule_timeout_uninterruptible(1);
3004
3005 /* Again, very different code for resync and recovery.
3006 * Both must result in an r10bio with a list of bios that
3007 * have bi_end_io, bi_sector, bi_disk set,
3008 * and bi_private set to the r10bio.
3009 * For recovery, we may actually create several r10bios
3010 * with 2 bios in each, that correspond to the bios in the main one.
3011 * In this case, the subordinate r10bios link back through a
3012 * borrowed master_bio pointer, and the counter in the master
3013 * includes a ref from each subordinate.
3014 */
3015 /* First, we decide what to do and set ->bi_end_io
3016 * To end_sync_read if we want to read, and
3017 * end_sync_write if we will want to write.
3018 */
3019
3020 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3021 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3022 /* recovery... the complicated one */
3023 int j;
3024 r10_bio = NULL;
3025
3026 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3027 int still_degraded;
3028 struct r10bio *rb2;
3029 sector_t sect;
3030 int must_sync;
3031 int any_working;
3032 int need_recover = 0;
3033 int need_replace = 0;
3034 struct raid10_info *mirror = &conf->mirrors[i];
3035 struct md_rdev *mrdev, *mreplace;
3036
3037 rcu_read_lock();
3038 mrdev = rcu_dereference(mirror->rdev);
3039 mreplace = rcu_dereference(mirror->replacement);
3040
3041 if (mrdev != NULL &&
3042 !test_bit(Faulty, &mrdev->flags) &&
3043 !test_bit(In_sync, &mrdev->flags))
3044 need_recover = 1;
3045 if (mreplace != NULL &&
3046 !test_bit(Faulty, &mreplace->flags))
3047 need_replace = 1;
3048
3049 if (!need_recover && !need_replace) {
3050 rcu_read_unlock();
3051 continue;
3052 }
3053
3054 still_degraded = 0;
3055 /* want to reconstruct this device */
3056 rb2 = r10_bio;
3057 sect = raid10_find_virt(conf, sector_nr, i);
3058 if (sect >= mddev->resync_max_sectors) {
3059 /* last stripe is not complete - don't
3060 * try to recover this sector.
3061 */
3062 rcu_read_unlock();
3063 continue;
3064 }
3065 if (mreplace && test_bit(Faulty, &mreplace->flags))
3066 mreplace = NULL;
3067 /* Unless we are doing a full sync, or a replacement
3068 * we only need to recover the block if it is set in
3069 * the bitmap
3070 */
3071 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3072 &sync_blocks, 1);
3073 if (sync_blocks < max_sync)
3074 max_sync = sync_blocks;
3075 if (!must_sync &&
3076 mreplace == NULL &&
3077 !conf->fullsync) {
3078 /* yep, skip the sync_blocks here, but don't assume
3079 * that there will never be anything to do here
3080 */
3081 chunks_skipped = -1;
3082 rcu_read_unlock();
3083 continue;
3084 }
3085 atomic_inc(&mrdev->nr_pending);
3086 if (mreplace)
3087 atomic_inc(&mreplace->nr_pending);
3088 rcu_read_unlock();
3089
3090 r10_bio = raid10_alloc_init_r10buf(conf);
3091 r10_bio->state = 0;
3092 raise_barrier(conf, rb2 != NULL);
3093 atomic_set(&r10_bio->remaining, 0);
3094
3095 r10_bio->master_bio = (struct bio*)rb2;
3096 if (rb2)
3097 atomic_inc(&rb2->remaining);
3098 r10_bio->mddev = mddev;
3099 set_bit(R10BIO_IsRecover, &r10_bio->state);
3100 r10_bio->sector = sect;
3101
3102 raid10_find_phys(conf, r10_bio);
3103
3104 /* Need to check if the array will still be
3105 * degraded
3106 */
3107 rcu_read_lock();
3108 for (j = 0; j < conf->geo.raid_disks; j++) {
3109 struct md_rdev *rdev = rcu_dereference(
3110 conf->mirrors[j].rdev);
3111 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3112 still_degraded = 1;
3113 break;
3114 }
3115 }
3116
3117 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3118 &sync_blocks, still_degraded);
3119
3120 any_working = 0;
3121 for (j=0; j<conf->copies;j++) {
3122 int k;
3123 int d = r10_bio->devs[j].devnum;
3124 sector_t from_addr, to_addr;
3125 struct md_rdev *rdev =
3126 rcu_dereference(conf->mirrors[d].rdev);
3127 sector_t sector, first_bad;
3128 int bad_sectors;
3129 if (!rdev ||
3130 !test_bit(In_sync, &rdev->flags))
3131 continue;
3132 /* This is where we read from */
3133 any_working = 1;
3134 sector = r10_bio->devs[j].addr;
3135
3136 if (is_badblock(rdev, sector, max_sync,
3137 &first_bad, &bad_sectors)) {
3138 if (first_bad > sector)
3139 max_sync = first_bad - sector;
3140 else {
3141 bad_sectors -= (sector
3142 - first_bad);
3143 if (max_sync > bad_sectors)
3144 max_sync = bad_sectors;
3145 continue;
3146 }
3147 }
3148 bio = r10_bio->devs[0].bio;
3149 bio->bi_next = biolist;
3150 biolist = bio;
3151 bio->bi_end_io = end_sync_read;
3152 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3153 if (test_bit(FailFast, &rdev->flags))
3154 bio->bi_opf |= MD_FAILFAST;
3155 from_addr = r10_bio->devs[j].addr;
3156 bio->bi_iter.bi_sector = from_addr +
3157 rdev->data_offset;
3158 bio_set_dev(bio, rdev->bdev);
3159 atomic_inc(&rdev->nr_pending);
3160 /* and we write to 'i' (if not in_sync) */
3161
3162 for (k=0; k<conf->copies; k++)
3163 if (r10_bio->devs[k].devnum == i)
3164 break;
3165 BUG_ON(k == conf->copies);
3166 to_addr = r10_bio->devs[k].addr;
3167 r10_bio->devs[0].devnum = d;
3168 r10_bio->devs[0].addr = from_addr;
3169 r10_bio->devs[1].devnum = i;
3170 r10_bio->devs[1].addr = to_addr;
3171
3172 if (need_recover) {
3173 bio = r10_bio->devs[1].bio;
3174 bio->bi_next = biolist;
3175 biolist = bio;
3176 bio->bi_end_io = end_sync_write;
3177 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3178 bio->bi_iter.bi_sector = to_addr
3179 + mrdev->data_offset;
3180 bio_set_dev(bio, mrdev->bdev);
3181 atomic_inc(&r10_bio->remaining);
3182 } else
3183 r10_bio->devs[1].bio->bi_end_io = NULL;
3184
3185 /* and maybe write to replacement */
3186 bio = r10_bio->devs[1].repl_bio;
3187 if (bio)
3188 bio->bi_end_io = NULL;
3189 /* Note: if need_replace, then bio
3190 * cannot be NULL as r10buf_pool_alloc will
3191 * have allocated it.
3192 */
3193 if (!need_replace)
3194 break;
3195 bio->bi_next = biolist;
3196 biolist = bio;
3197 bio->bi_end_io = end_sync_write;
3198 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3199 bio->bi_iter.bi_sector = to_addr +
3200 mreplace->data_offset;
3201 bio_set_dev(bio, mreplace->bdev);
3202 atomic_inc(&r10_bio->remaining);
3203 break;
3204 }
3205 rcu_read_unlock();
3206 if (j == conf->copies) {
3207 /* Cannot recover, so abort the recovery or
3208 * record a bad block */
3209 if (any_working) {
3210 /* problem is that there are bad blocks
3211 * on other device(s)
3212 */
3213 int k;
3214 for (k = 0; k < conf->copies; k++)
3215 if (r10_bio->devs[k].devnum == i)
3216 break;
3217 if (!test_bit(In_sync,
3218 &mrdev->flags)
3219 && !rdev_set_badblocks(
3220 mrdev,
3221 r10_bio->devs[k].addr,
3222 max_sync, 0))
3223 any_working = 0;
3224 if (mreplace &&
3225 !rdev_set_badblocks(
3226 mreplace,
3227 r10_bio->devs[k].addr,
3228 max_sync, 0))
3229 any_working = 0;
3230 }
3231 if (!any_working) {
3232 if (!test_and_set_bit(MD_RECOVERY_INTR,
3233 &mddev->recovery))
3234 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3235 mdname(mddev));
3236 mirror->recovery_disabled
3237 = mddev->recovery_disabled;
3238 }
3239 put_buf(r10_bio);
3240 if (rb2)
3241 atomic_dec(&rb2->remaining);
3242 r10_bio = rb2;
3243 rdev_dec_pending(mrdev, mddev);
3244 if (mreplace)
3245 rdev_dec_pending(mreplace, mddev);
3246 break;
3247 }
3248 rdev_dec_pending(mrdev, mddev);
3249 if (mreplace)
3250 rdev_dec_pending(mreplace, mddev);
3251 if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3252 /* Only want this if there is elsewhere to
3253 * read from. 'j' is currently the first
3254 * readable copy.
3255 */
3256 int targets = 1;
3257 for (; j < conf->copies; j++) {
3258 int d = r10_bio->devs[j].devnum;
3259 if (conf->mirrors[d].rdev &&
3260 test_bit(In_sync,
3261 &conf->mirrors[d].rdev->flags))
3262 targets++;
3263 }
3264 if (targets == 1)
3265 r10_bio->devs[0].bio->bi_opf
3266 &= ~MD_FAILFAST;
3267 }
3268 }
3269 if (biolist == NULL) {
3270 while (r10_bio) {
3271 struct r10bio *rb2 = r10_bio;
3272 r10_bio = (struct r10bio*) rb2->master_bio;
3273 rb2->master_bio = NULL;
3274 put_buf(rb2);
3275 }
3276 goto giveup;
3277 }
3278 } else {
3279 /* resync. Schedule a read for every block at this virt offset */
3280 int count = 0;
3281
3282 /*
3283 * Since curr_resync_completed could probably not update in
3284 * time, and we will set cluster_sync_low based on it.
3285 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3286 * safety reason, which ensures curr_resync_completed is
3287 * updated in bitmap_cond_end_sync.
3288 */
3289 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
3290 mddev_is_clustered(mddev) &&
3291 (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
3292
3293 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
3294 &sync_blocks, mddev->degraded) &&
3295 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3296 &mddev->recovery)) {
3297 /* We can skip this block */
3298 *skipped = 1;
3299 return sync_blocks + sectors_skipped;
3300 }
3301 if (sync_blocks < max_sync)
3302 max_sync = sync_blocks;
3303 r10_bio = raid10_alloc_init_r10buf(conf);
3304 r10_bio->state = 0;
3305
3306 r10_bio->mddev = mddev;
3307 atomic_set(&r10_bio->remaining, 0);
3308 raise_barrier(conf, 0);
3309 conf->next_resync = sector_nr;
3310
3311 r10_bio->master_bio = NULL;
3312 r10_bio->sector = sector_nr;
3313 set_bit(R10BIO_IsSync, &r10_bio->state);
3314 raid10_find_phys(conf, r10_bio);
3315 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3316
3317 for (i = 0; i < conf->copies; i++) {
3318 int d = r10_bio->devs[i].devnum;
3319 sector_t first_bad, sector;
3320 int bad_sectors;
3321 struct md_rdev *rdev;
3322
3323 if (r10_bio->devs[i].repl_bio)
3324 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3325
3326 bio = r10_bio->devs[i].bio;
3327 bio->bi_status = BLK_STS_IOERR;
3328 rcu_read_lock();
3329 rdev = rcu_dereference(conf->mirrors[d].rdev);
3330 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3331 rcu_read_unlock();
3332 continue;
3333 }
3334 sector = r10_bio->devs[i].addr;
3335 if (is_badblock(rdev, sector, max_sync,
3336 &first_bad, &bad_sectors)) {
3337 if (first_bad > sector)
3338 max_sync = first_bad - sector;
3339 else {
3340 bad_sectors -= (sector - first_bad);
3341 if (max_sync > bad_sectors)
3342 max_sync = bad_sectors;
3343 rcu_read_unlock();
3344 continue;
3345 }
3346 }
3347 atomic_inc(&rdev->nr_pending);
3348 atomic_inc(&r10_bio->remaining);
3349 bio->bi_next = biolist;
3350 biolist = bio;
3351 bio->bi_end_io = end_sync_read;
3352 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3353 if (test_bit(FailFast, &rdev->flags))
3354 bio->bi_opf |= MD_FAILFAST;
3355 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3356 bio_set_dev(bio, rdev->bdev);
3357 count++;
3358
3359 rdev = rcu_dereference(conf->mirrors[d].replacement);
3360 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3361 rcu_read_unlock();
3362 continue;
3363 }
3364 atomic_inc(&rdev->nr_pending);
3365
3366 /* Need to set up for writing to the replacement */
3367 bio = r10_bio->devs[i].repl_bio;
3368 bio->bi_status = BLK_STS_IOERR;
3369
3370 sector = r10_bio->devs[i].addr;
3371 bio->bi_next = biolist;
3372 biolist = bio;
3373 bio->bi_end_io = end_sync_write;
3374 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3375 if (test_bit(FailFast, &rdev->flags))
3376 bio->bi_opf |= MD_FAILFAST;
3377 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3378 bio_set_dev(bio, rdev->bdev);
3379 count++;
3380 rcu_read_unlock();
3381 }
3382
3383 if (count < 2) {
3384 for (i=0; i<conf->copies; i++) {
3385 int d = r10_bio->devs[i].devnum;
3386 if (r10_bio->devs[i].bio->bi_end_io)
3387 rdev_dec_pending(conf->mirrors[d].rdev,
3388 mddev);
3389 if (r10_bio->devs[i].repl_bio &&
3390 r10_bio->devs[i].repl_bio->bi_end_io)
3391 rdev_dec_pending(
3392 conf->mirrors[d].replacement,
3393 mddev);
3394 }
3395 put_buf(r10_bio);
3396 biolist = NULL;
3397 goto giveup;
3398 }
3399 }
3400
3401 nr_sectors = 0;
3402 if (sector_nr + max_sync < max_sector)
3403 max_sector = sector_nr + max_sync;
3404 do {
3405 struct page *page;
3406 int len = PAGE_SIZE;
3407 if (sector_nr + (len>>9) > max_sector)
3408 len = (max_sector - sector_nr) << 9;
3409 if (len == 0)
3410 break;
3411 for (bio= biolist ; bio ; bio=bio->bi_next) {
3412 struct resync_pages *rp = get_resync_pages(bio);
3413 page = resync_fetch_page(rp, page_idx);
3414 /*
3415 * won't fail because the vec table is big enough
3416 * to hold all these pages
3417 */
3418 bio_add_page(bio, page, len, 0);
3419 }
3420 nr_sectors += len>>9;
3421 sector_nr += len>>9;
3422 } while (++page_idx < RESYNC_PAGES);
3423 r10_bio->sectors = nr_sectors;
3424
3425 if (mddev_is_clustered(mddev) &&
3426 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3427 /* It is resync not recovery */
3428 if (conf->cluster_sync_high < sector_nr + nr_sectors) {
3429 conf->cluster_sync_low = mddev->curr_resync_completed;
3430 raid10_set_cluster_sync_high(conf);
3431 /* Send resync message */
3432 md_cluster_ops->resync_info_update(mddev,
3433 conf->cluster_sync_low,
3434 conf->cluster_sync_high);
3435 }
3436 } else if (mddev_is_clustered(mddev)) {
3437 /* This is recovery not resync */
3438 sector_t sect_va1, sect_va2;
3439 bool broadcast_msg = false;
3440
3441 for (i = 0; i < conf->geo.raid_disks; i++) {
3442 /*
3443 * sector_nr is a device address for recovery, so we
3444 * need translate it to array address before compare
3445 * with cluster_sync_high.
3446 */
3447 sect_va1 = raid10_find_virt(conf, sector_nr, i);
3448
3449 if (conf->cluster_sync_high < sect_va1 + nr_sectors) {
3450 broadcast_msg = true;
3451 /*
3452 * curr_resync_completed is similar as
3453 * sector_nr, so make the translation too.
3454 */
3455 sect_va2 = raid10_find_virt(conf,
3456 mddev->curr_resync_completed, i);
3457
3458 if (conf->cluster_sync_low == 0 ||
3459 conf->cluster_sync_low > sect_va2)
3460 conf->cluster_sync_low = sect_va2;
3461 }
3462 }
3463 if (broadcast_msg) {
3464 raid10_set_cluster_sync_high(conf);
3465 md_cluster_ops->resync_info_update(mddev,
3466 conf->cluster_sync_low,
3467 conf->cluster_sync_high);
3468 }
3469 }
3470
3471 while (biolist) {
3472 bio = biolist;
3473 biolist = biolist->bi_next;
3474
3475 bio->bi_next = NULL;
3476 r10_bio = get_resync_r10bio(bio);
3477 r10_bio->sectors = nr_sectors;
3478
3479 if (bio->bi_end_io == end_sync_read) {
3480 md_sync_acct_bio(bio, nr_sectors);
3481 bio->bi_status = 0;
3482 submit_bio_noacct(bio);
3483 }
3484 }
3485
3486 if (sectors_skipped)
3487 /* pretend they weren't skipped, it makes
3488 * no important difference in this case
3489 */
3490 md_done_sync(mddev, sectors_skipped, 1);
3491
3492 return sectors_skipped + nr_sectors;
3493 giveup:
3494 /* There is nowhere to write, so all non-sync
3495 * drives must be failed or in resync, all drives
3496 * have a bad block, so try the next chunk...
3497 */
3498 if (sector_nr + max_sync < max_sector)
3499 max_sector = sector_nr + max_sync;
3500
3501 sectors_skipped += (max_sector - sector_nr);
3502 chunks_skipped ++;
3503 sector_nr = max_sector;
3504 goto skipped;
3505}
3506
3507static sector_t
3508raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3509{
3510 sector_t size;
3511 struct r10conf *conf = mddev->private;
3512
3513 if (!raid_disks)
3514 raid_disks = min(conf->geo.raid_disks,
3515 conf->prev.raid_disks);
3516 if (!sectors)
3517 sectors = conf->dev_sectors;
3518
3519 size = sectors >> conf->geo.chunk_shift;
3520 sector_div(size, conf->geo.far_copies);
3521 size = size * raid_disks;
3522 sector_div(size, conf->geo.near_copies);
3523
3524 return size << conf->geo.chunk_shift;
3525}
3526
3527static void calc_sectors(struct r10conf *conf, sector_t size)
3528{
3529 /* Calculate the number of sectors-per-device that will
3530 * actually be used, and set conf->dev_sectors and
3531 * conf->stride
3532 */
3533
3534 size = size >> conf->geo.chunk_shift;
3535 sector_div(size, conf->geo.far_copies);
3536 size = size * conf->geo.raid_disks;
3537 sector_div(size, conf->geo.near_copies);
3538 /* 'size' is now the number of chunks in the array */
3539 /* calculate "used chunks per device" */
3540 size = size * conf->copies;
3541
3542 /* We need to round up when dividing by raid_disks to
3543 * get the stride size.
3544 */
3545 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3546
3547 conf->dev_sectors = size << conf->geo.chunk_shift;
3548
3549 if (conf->geo.far_offset)
3550 conf->geo.stride = 1 << conf->geo.chunk_shift;
3551 else {
3552 sector_div(size, conf->geo.far_copies);
3553 conf->geo.stride = size << conf->geo.chunk_shift;
3554 }
3555}
3556
3557enum geo_type {geo_new, geo_old, geo_start};
3558static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3559{
3560 int nc, fc, fo;
3561 int layout, chunk, disks;
3562 switch (new) {
3563 case geo_old:
3564 layout = mddev->layout;
3565 chunk = mddev->chunk_sectors;
3566 disks = mddev->raid_disks - mddev->delta_disks;
3567 break;
3568 case geo_new:
3569 layout = mddev->new_layout;
3570 chunk = mddev->new_chunk_sectors;
3571 disks = mddev->raid_disks;
3572 break;
3573 default: /* avoid 'may be unused' warnings */
3574 case geo_start: /* new when starting reshape - raid_disks not
3575 * updated yet. */
3576 layout = mddev->new_layout;
3577 chunk = mddev->new_chunk_sectors;
3578 disks = mddev->raid_disks + mddev->delta_disks;
3579 break;
3580 }
3581 if (layout >> 19)
3582 return -1;
3583 if (chunk < (PAGE_SIZE >> 9) ||
3584 !is_power_of_2(chunk))
3585 return -2;
3586 nc = layout & 255;
3587 fc = (layout >> 8) & 255;
3588 fo = layout & (1<<16);
3589 geo->raid_disks = disks;
3590 geo->near_copies = nc;
3591 geo->far_copies = fc;
3592 geo->far_offset = fo;
3593 switch (layout >> 17) {
3594 case 0: /* original layout. simple but not always optimal */
3595 geo->far_set_size = disks;
3596 break;
3597 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3598 * actually using this, but leave code here just in case.*/
3599 geo->far_set_size = disks/fc;
3600 WARN(geo->far_set_size < fc,
3601 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3602 break;
3603 case 2: /* "improved" layout fixed to match documentation */
3604 geo->far_set_size = fc * nc;
3605 break;
3606 default: /* Not a valid layout */
3607 return -1;
3608 }
3609 geo->chunk_mask = chunk - 1;
3610 geo->chunk_shift = ffz(~chunk);
3611 return nc*fc;
3612}
3613
3614static struct r10conf *setup_conf(struct mddev *mddev)
3615{
3616 struct r10conf *conf = NULL;
3617 int err = -EINVAL;
3618 struct geom geo;
3619 int copies;
3620
3621 copies = setup_geo(&geo, mddev, geo_new);
3622
3623 if (copies == -2) {
3624 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3625 mdname(mddev), PAGE_SIZE);
3626 goto out;
3627 }
3628
3629 if (copies < 2 || copies > mddev->raid_disks) {
3630 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3631 mdname(mddev), mddev->new_layout);
3632 goto out;
3633 }
3634
3635 err = -ENOMEM;
3636 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3637 if (!conf)
3638 goto out;
3639
3640 /* FIXME calc properly */
3641 conf->mirrors = kcalloc(mddev->raid_disks + max(0, -mddev->delta_disks),
3642 sizeof(struct raid10_info),
3643 GFP_KERNEL);
3644 if (!conf->mirrors)
3645 goto out;
3646
3647 conf->tmppage = alloc_page(GFP_KERNEL);
3648 if (!conf->tmppage)
3649 goto out;
3650
3651 conf->geo = geo;
3652 conf->copies = copies;
3653 err = mempool_init(&conf->r10bio_pool, NR_RAID_BIOS, r10bio_pool_alloc,
3654 rbio_pool_free, conf);
3655 if (err)
3656 goto out;
3657
3658 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3659 if (err)
3660 goto out;
3661
3662 calc_sectors(conf, mddev->dev_sectors);
3663 if (mddev->reshape_position == MaxSector) {
3664 conf->prev = conf->geo;
3665 conf->reshape_progress = MaxSector;
3666 } else {
3667 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3668 err = -EINVAL;
3669 goto out;
3670 }
3671 conf->reshape_progress = mddev->reshape_position;
3672 if (conf->prev.far_offset)
3673 conf->prev.stride = 1 << conf->prev.chunk_shift;
3674 else
3675 /* far_copies must be 1 */
3676 conf->prev.stride = conf->dev_sectors;
3677 }
3678 conf->reshape_safe = conf->reshape_progress;
3679 spin_lock_init(&conf->device_lock);
3680 INIT_LIST_HEAD(&conf->retry_list);
3681 INIT_LIST_HEAD(&conf->bio_end_io_list);
3682
3683 spin_lock_init(&conf->resync_lock);
3684 init_waitqueue_head(&conf->wait_barrier);
3685 atomic_set(&conf->nr_pending, 0);
3686
3687 err = -ENOMEM;
3688 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3689 if (!conf->thread)
3690 goto out;
3691
3692 conf->mddev = mddev;
3693 return conf;
3694
3695 out:
3696 if (conf) {
3697 mempool_exit(&conf->r10bio_pool);
3698 kfree(conf->mirrors);
3699 safe_put_page(conf->tmppage);
3700 bioset_exit(&conf->bio_split);
3701 kfree(conf);
3702 }
3703 return ERR_PTR(err);
3704}
3705
3706static int raid10_run(struct mddev *mddev)
3707{
3708 struct r10conf *conf;
3709 int i, disk_idx, chunk_size;
3710 struct raid10_info *disk;
3711 struct md_rdev *rdev;
3712 sector_t size;
3713 sector_t min_offset_diff = 0;
3714 int first = 1;
3715 bool discard_supported = false;
3716
3717 if (mddev_init_writes_pending(mddev) < 0)
3718 return -ENOMEM;
3719
3720 if (mddev->private == NULL) {
3721 conf = setup_conf(mddev);
3722 if (IS_ERR(conf))
3723 return PTR_ERR(conf);
3724 mddev->private = conf;
3725 }
3726 conf = mddev->private;
3727 if (!conf)
3728 goto out;
3729
3730 if (mddev_is_clustered(conf->mddev)) {
3731 int fc, fo;
3732
3733 fc = (mddev->layout >> 8) & 255;
3734 fo = mddev->layout & (1<<16);
3735 if (fc > 1 || fo > 0) {
3736 pr_err("only near layout is supported by clustered"
3737 " raid10\n");
3738 goto out_free_conf;
3739 }
3740 }
3741
3742 mddev->thread = conf->thread;
3743 conf->thread = NULL;
3744
3745 chunk_size = mddev->chunk_sectors << 9;
3746 if (mddev->queue) {
3747 blk_queue_max_discard_sectors(mddev->queue,
3748 mddev->chunk_sectors);
3749 blk_queue_max_write_same_sectors(mddev->queue, 0);
3750 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3751 blk_queue_io_min(mddev->queue, chunk_size);
3752 if (conf->geo.raid_disks % conf->geo.near_copies)
3753 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3754 else
3755 blk_queue_io_opt(mddev->queue, chunk_size *
3756 (conf->geo.raid_disks / conf->geo.near_copies));
3757 }
3758
3759 rdev_for_each(rdev, mddev) {
3760 long long diff;
3761
3762 disk_idx = rdev->raid_disk;
3763 if (disk_idx < 0)
3764 continue;
3765 if (disk_idx >= conf->geo.raid_disks &&
3766 disk_idx >= conf->prev.raid_disks)
3767 continue;
3768 disk = conf->mirrors + disk_idx;
3769
3770 if (test_bit(Replacement, &rdev->flags)) {
3771 if (disk->replacement)
3772 goto out_free_conf;
3773 disk->replacement = rdev;
3774 } else {
3775 if (disk->rdev)
3776 goto out_free_conf;
3777 disk->rdev = rdev;
3778 }
3779 diff = (rdev->new_data_offset - rdev->data_offset);
3780 if (!mddev->reshape_backwards)
3781 diff = -diff;
3782 if (diff < 0)
3783 diff = 0;
3784 if (first || diff < min_offset_diff)
3785 min_offset_diff = diff;
3786
3787 if (mddev->gendisk)
3788 disk_stack_limits(mddev->gendisk, rdev->bdev,
3789 rdev->data_offset << 9);
3790
3791 disk->head_position = 0;
3792
3793 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3794 discard_supported = true;
3795 first = 0;
3796 }
3797
3798 if (mddev->queue) {
3799 if (discard_supported)
3800 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3801 mddev->queue);
3802 else
3803 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3804 mddev->queue);
3805 }
3806 /* need to check that every block has at least one working mirror */
3807 if (!enough(conf, -1)) {
3808 pr_err("md/raid10:%s: not enough operational mirrors.\n",
3809 mdname(mddev));
3810 goto out_free_conf;
3811 }
3812
3813 if (conf->reshape_progress != MaxSector) {
3814 /* must ensure that shape change is supported */
3815 if (conf->geo.far_copies != 1 &&
3816 conf->geo.far_offset == 0)
3817 goto out_free_conf;
3818 if (conf->prev.far_copies != 1 &&
3819 conf->prev.far_offset == 0)
3820 goto out_free_conf;
3821 }
3822
3823 mddev->degraded = 0;
3824 for (i = 0;
3825 i < conf->geo.raid_disks
3826 || i < conf->prev.raid_disks;
3827 i++) {
3828
3829 disk = conf->mirrors + i;
3830
3831 if (!disk->rdev && disk->replacement) {
3832 /* The replacement is all we have - use it */
3833 disk->rdev = disk->replacement;
3834 disk->replacement = NULL;
3835 clear_bit(Replacement, &disk->rdev->flags);
3836 }
3837
3838 if (!disk->rdev ||
3839 !test_bit(In_sync, &disk->rdev->flags)) {
3840 disk->head_position = 0;
3841 mddev->degraded++;
3842 if (disk->rdev &&
3843 disk->rdev->saved_raid_disk < 0)
3844 conf->fullsync = 1;
3845 }
3846
3847 if (disk->replacement &&
3848 !test_bit(In_sync, &disk->replacement->flags) &&
3849 disk->replacement->saved_raid_disk < 0) {
3850 conf->fullsync = 1;
3851 }
3852
3853 disk->recovery_disabled = mddev->recovery_disabled - 1;
3854 }
3855
3856 if (mddev->recovery_cp != MaxSector)
3857 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
3858 mdname(mddev));
3859 pr_info("md/raid10:%s: active with %d out of %d devices\n",
3860 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3861 conf->geo.raid_disks);
3862 /*
3863 * Ok, everything is just fine now
3864 */
3865 mddev->dev_sectors = conf->dev_sectors;
3866 size = raid10_size(mddev, 0, 0);
3867 md_set_array_sectors(mddev, size);
3868 mddev->resync_max_sectors = size;
3869 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3870
3871 if (mddev->queue) {
3872 int stripe = conf->geo.raid_disks *
3873 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3874
3875 /* Calculate max read-ahead size.
3876 * We need to readahead at least twice a whole stripe....
3877 * maybe...
3878 */
3879 stripe /= conf->geo.near_copies;
3880 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
3881 mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
3882 }
3883
3884 if (md_integrity_register(mddev))
3885 goto out_free_conf;
3886
3887 if (conf->reshape_progress != MaxSector) {
3888 unsigned long before_length, after_length;
3889
3890 before_length = ((1 << conf->prev.chunk_shift) *
3891 conf->prev.far_copies);
3892 after_length = ((1 << conf->geo.chunk_shift) *
3893 conf->geo.far_copies);
3894
3895 if (max(before_length, after_length) > min_offset_diff) {
3896 /* This cannot work */
3897 pr_warn("md/raid10: offset difference not enough to continue reshape\n");
3898 goto out_free_conf;
3899 }
3900 conf->offset_diff = min_offset_diff;
3901
3902 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3903 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3904 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3905 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3906 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3907 "reshape");
3908 if (!mddev->sync_thread)
3909 goto out_free_conf;
3910 }
3911
3912 return 0;
3913
3914out_free_conf:
3915 md_unregister_thread(&mddev->thread);
3916 mempool_exit(&conf->r10bio_pool);
3917 safe_put_page(conf->tmppage);
3918 kfree(conf->mirrors);
3919 kfree(conf);
3920 mddev->private = NULL;
3921out:
3922 return -EIO;
3923}
3924
3925static void raid10_free(struct mddev *mddev, void *priv)
3926{
3927 struct r10conf *conf = priv;
3928
3929 mempool_exit(&conf->r10bio_pool);
3930 safe_put_page(conf->tmppage);
3931 kfree(conf->mirrors);
3932 kfree(conf->mirrors_old);
3933 kfree(conf->mirrors_new);
3934 bioset_exit(&conf->bio_split);
3935 kfree(conf);
3936}
3937
3938static void raid10_quiesce(struct mddev *mddev, int quiesce)
3939{
3940 struct r10conf *conf = mddev->private;
3941
3942 if (quiesce)
3943 raise_barrier(conf, 0);
3944 else
3945 lower_barrier(conf);
3946}
3947
3948static int raid10_resize(struct mddev *mddev, sector_t sectors)
3949{
3950 /* Resize of 'far' arrays is not supported.
3951 * For 'near' and 'offset' arrays we can set the
3952 * number of sectors used to be an appropriate multiple
3953 * of the chunk size.
3954 * For 'offset', this is far_copies*chunksize.
3955 * For 'near' the multiplier is the LCM of
3956 * near_copies and raid_disks.
3957 * So if far_copies > 1 && !far_offset, fail.
3958 * Else find LCM(raid_disks, near_copy)*far_copies and
3959 * multiply by chunk_size. Then round to this number.
3960 * This is mostly done by raid10_size()
3961 */
3962 struct r10conf *conf = mddev->private;
3963 sector_t oldsize, size;
3964
3965 if (mddev->reshape_position != MaxSector)
3966 return -EBUSY;
3967
3968 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3969 return -EINVAL;
3970
3971 oldsize = raid10_size(mddev, 0, 0);
3972 size = raid10_size(mddev, sectors, 0);
3973 if (mddev->external_size &&
3974 mddev->array_sectors > size)
3975 return -EINVAL;
3976 if (mddev->bitmap) {
3977 int ret = md_bitmap_resize(mddev->bitmap, size, 0, 0);
3978 if (ret)
3979 return ret;
3980 }
3981 md_set_array_sectors(mddev, size);
3982 if (sectors > mddev->dev_sectors &&
3983 mddev->recovery_cp > oldsize) {
3984 mddev->recovery_cp = oldsize;
3985 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3986 }
3987 calc_sectors(conf, sectors);
3988 mddev->dev_sectors = conf->dev_sectors;
3989 mddev->resync_max_sectors = size;
3990 return 0;
3991}
3992
3993static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
3994{
3995 struct md_rdev *rdev;
3996 struct r10conf *conf;
3997
3998 if (mddev->degraded > 0) {
3999 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
4000 mdname(mddev));
4001 return ERR_PTR(-EINVAL);
4002 }
4003 sector_div(size, devs);
4004
4005 /* Set new parameters */
4006 mddev->new_level = 10;
4007 /* new layout: far_copies = 1, near_copies = 2 */
4008 mddev->new_layout = (1<<8) + 2;
4009 mddev->new_chunk_sectors = mddev->chunk_sectors;
4010 mddev->delta_disks = mddev->raid_disks;
4011 mddev->raid_disks *= 2;
4012 /* make sure it will be not marked as dirty */
4013 mddev->recovery_cp = MaxSector;
4014 mddev->dev_sectors = size;
4015
4016 conf = setup_conf(mddev);
4017 if (!IS_ERR(conf)) {
4018 rdev_for_each(rdev, mddev)
4019 if (rdev->raid_disk >= 0) {
4020 rdev->new_raid_disk = rdev->raid_disk * 2;
4021 rdev->sectors = size;
4022 }
4023 conf->barrier = 1;
4024 }
4025
4026 return conf;
4027}
4028
4029static void *raid10_takeover(struct mddev *mddev)
4030{
4031 struct r0conf *raid0_conf;
4032
4033 /* raid10 can take over:
4034 * raid0 - providing it has only two drives
4035 */
4036 if (mddev->level == 0) {
4037 /* for raid0 takeover only one zone is supported */
4038 raid0_conf = mddev->private;
4039 if (raid0_conf->nr_strip_zones > 1) {
4040 pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4041 mdname(mddev));
4042 return ERR_PTR(-EINVAL);
4043 }
4044 return raid10_takeover_raid0(mddev,
4045 raid0_conf->strip_zone->zone_end,
4046 raid0_conf->strip_zone->nb_dev);
4047 }
4048 return ERR_PTR(-EINVAL);
4049}
4050
4051static int raid10_check_reshape(struct mddev *mddev)
4052{
4053 /* Called when there is a request to change
4054 * - layout (to ->new_layout)
4055 * - chunk size (to ->new_chunk_sectors)
4056 * - raid_disks (by delta_disks)
4057 * or when trying to restart a reshape that was ongoing.
4058 *
4059 * We need to validate the request and possibly allocate
4060 * space if that might be an issue later.
4061 *
4062 * Currently we reject any reshape of a 'far' mode array,
4063 * allow chunk size to change if new is generally acceptable,
4064 * allow raid_disks to increase, and allow
4065 * a switch between 'near' mode and 'offset' mode.
4066 */
4067 struct r10conf *conf = mddev->private;
4068 struct geom geo;
4069
4070 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
4071 return -EINVAL;
4072
4073 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
4074 /* mustn't change number of copies */
4075 return -EINVAL;
4076 if (geo.far_copies > 1 && !geo.far_offset)
4077 /* Cannot switch to 'far' mode */
4078 return -EINVAL;
4079
4080 if (mddev->array_sectors & geo.chunk_mask)
4081 /* not factor of array size */
4082 return -EINVAL;
4083
4084 if (!enough(conf, -1))
4085 return -EINVAL;
4086
4087 kfree(conf->mirrors_new);
4088 conf->mirrors_new = NULL;
4089 if (mddev->delta_disks > 0) {
4090 /* allocate new 'mirrors' list */
4091 conf->mirrors_new =
4092 kcalloc(mddev->raid_disks + mddev->delta_disks,
4093 sizeof(struct raid10_info),
4094 GFP_KERNEL);
4095 if (!conf->mirrors_new)
4096 return -ENOMEM;
4097 }
4098 return 0;
4099}
4100
4101/*
4102 * Need to check if array has failed when deciding whether to:
4103 * - start an array
4104 * - remove non-faulty devices
4105 * - add a spare
4106 * - allow a reshape
4107 * This determination is simple when no reshape is happening.
4108 * However if there is a reshape, we need to carefully check
4109 * both the before and after sections.
4110 * This is because some failed devices may only affect one
4111 * of the two sections, and some non-in_sync devices may
4112 * be insync in the section most affected by failed devices.
4113 */
4114static int calc_degraded(struct r10conf *conf)
4115{
4116 int degraded, degraded2;
4117 int i;
4118
4119 rcu_read_lock();
4120 degraded = 0;
4121 /* 'prev' section first */
4122 for (i = 0; i < conf->prev.raid_disks; i++) {
4123 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4124 if (!rdev || test_bit(Faulty, &rdev->flags))
4125 degraded++;
4126 else if (!test_bit(In_sync, &rdev->flags))
4127 /* When we can reduce the number of devices in
4128 * an array, this might not contribute to
4129 * 'degraded'. It does now.
4130 */
4131 degraded++;
4132 }
4133 rcu_read_unlock();
4134 if (conf->geo.raid_disks == conf->prev.raid_disks)
4135 return degraded;
4136 rcu_read_lock();
4137 degraded2 = 0;
4138 for (i = 0; i < conf->geo.raid_disks; i++) {
4139 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4140 if (!rdev || test_bit(Faulty, &rdev->flags))
4141 degraded2++;
4142 else if (!test_bit(In_sync, &rdev->flags)) {
4143 /* If reshape is increasing the number of devices,
4144 * this section has already been recovered, so
4145 * it doesn't contribute to degraded.
4146 * else it does.
4147 */
4148 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4149 degraded2++;
4150 }
4151 }
4152 rcu_read_unlock();
4153 if (degraded2 > degraded)
4154 return degraded2;
4155 return degraded;
4156}
4157
4158static int raid10_start_reshape(struct mddev *mddev)
4159{
4160 /* A 'reshape' has been requested. This commits
4161 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4162 * This also checks if there are enough spares and adds them
4163 * to the array.
4164 * We currently require enough spares to make the final
4165 * array non-degraded. We also require that the difference
4166 * between old and new data_offset - on each device - is
4167 * enough that we never risk over-writing.
4168 */
4169
4170 unsigned long before_length, after_length;
4171 sector_t min_offset_diff = 0;
4172 int first = 1;
4173 struct geom new;
4174 struct r10conf *conf = mddev->private;
4175 struct md_rdev *rdev;
4176 int spares = 0;
4177 int ret;
4178
4179 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4180 return -EBUSY;
4181
4182 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4183 return -EINVAL;
4184
4185 before_length = ((1 << conf->prev.chunk_shift) *
4186 conf->prev.far_copies);
4187 after_length = ((1 << conf->geo.chunk_shift) *
4188 conf->geo.far_copies);
4189
4190 rdev_for_each(rdev, mddev) {
4191 if (!test_bit(In_sync, &rdev->flags)
4192 && !test_bit(Faulty, &rdev->flags))
4193 spares++;
4194 if (rdev->raid_disk >= 0) {
4195 long long diff = (rdev->new_data_offset
4196 - rdev->data_offset);
4197 if (!mddev->reshape_backwards)
4198 diff = -diff;
4199 if (diff < 0)
4200 diff = 0;
4201 if (first || diff < min_offset_diff)
4202 min_offset_diff = diff;
4203 first = 0;
4204 }
4205 }
4206
4207 if (max(before_length, after_length) > min_offset_diff)
4208 return -EINVAL;
4209
4210 if (spares < mddev->delta_disks)
4211 return -EINVAL;
4212
4213 conf->offset_diff = min_offset_diff;
4214 spin_lock_irq(&conf->device_lock);
4215 if (conf->mirrors_new) {
4216 memcpy(conf->mirrors_new, conf->mirrors,
4217 sizeof(struct raid10_info)*conf->prev.raid_disks);
4218 smp_mb();
4219 kfree(conf->mirrors_old);
4220 conf->mirrors_old = conf->mirrors;
4221 conf->mirrors = conf->mirrors_new;
4222 conf->mirrors_new = NULL;
4223 }
4224 setup_geo(&conf->geo, mddev, geo_start);
4225 smp_mb();
4226 if (mddev->reshape_backwards) {
4227 sector_t size = raid10_size(mddev, 0, 0);
4228 if (size < mddev->array_sectors) {
4229 spin_unlock_irq(&conf->device_lock);
4230 pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4231 mdname(mddev));
4232 return -EINVAL;
4233 }
4234 mddev->resync_max_sectors = size;
4235 conf->reshape_progress = size;
4236 } else
4237 conf->reshape_progress = 0;
4238 conf->reshape_safe = conf->reshape_progress;
4239 spin_unlock_irq(&conf->device_lock);
4240
4241 if (mddev->delta_disks && mddev->bitmap) {
4242 struct mdp_superblock_1 *sb = NULL;
4243 sector_t oldsize, newsize;
4244
4245 oldsize = raid10_size(mddev, 0, 0);
4246 newsize = raid10_size(mddev, 0, conf->geo.raid_disks);
4247
4248 if (!mddev_is_clustered(mddev)) {
4249 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4250 if (ret)
4251 goto abort;
4252 else
4253 goto out;
4254 }
4255
4256 rdev_for_each(rdev, mddev) {
4257 if (rdev->raid_disk > -1 &&
4258 !test_bit(Faulty, &rdev->flags))
4259 sb = page_address(rdev->sb_page);
4260 }
4261
4262 /*
4263 * some node is already performing reshape, and no need to
4264 * call md_bitmap_resize again since it should be called when
4265 * receiving BITMAP_RESIZE msg
4266 */
4267 if ((sb && (le32_to_cpu(sb->feature_map) &
4268 MD_FEATURE_RESHAPE_ACTIVE)) || (oldsize == newsize))
4269 goto out;
4270
4271 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4272 if (ret)
4273 goto abort;
4274
4275 ret = md_cluster_ops->resize_bitmaps(mddev, newsize, oldsize);
4276 if (ret) {
4277 md_bitmap_resize(mddev->bitmap, oldsize, 0, 0);
4278 goto abort;
4279 }
4280 }
4281out:
4282 if (mddev->delta_disks > 0) {
4283 rdev_for_each(rdev, mddev)
4284 if (rdev->raid_disk < 0 &&
4285 !test_bit(Faulty, &rdev->flags)) {
4286 if (raid10_add_disk(mddev, rdev) == 0) {
4287 if (rdev->raid_disk >=
4288 conf->prev.raid_disks)
4289 set_bit(In_sync, &rdev->flags);
4290 else
4291 rdev->recovery_offset = 0;
4292
4293 /* Failure here is OK */
4294 sysfs_link_rdev(mddev, rdev);
4295 }
4296 } else if (rdev->raid_disk >= conf->prev.raid_disks
4297 && !test_bit(Faulty, &rdev->flags)) {
4298 /* This is a spare that was manually added */
4299 set_bit(In_sync, &rdev->flags);
4300 }
4301 }
4302 /* When a reshape changes the number of devices,
4303 * ->degraded is measured against the larger of the
4304 * pre and post numbers.
4305 */
4306 spin_lock_irq(&conf->device_lock);
4307 mddev->degraded = calc_degraded(conf);
4308 spin_unlock_irq(&conf->device_lock);
4309 mddev->raid_disks = conf->geo.raid_disks;
4310 mddev->reshape_position = conf->reshape_progress;
4311 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4312
4313 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4314 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4315 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4316 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4317 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4318
4319 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4320 "reshape");
4321 if (!mddev->sync_thread) {
4322 ret = -EAGAIN;
4323 goto abort;
4324 }
4325 conf->reshape_checkpoint = jiffies;
4326 md_wakeup_thread(mddev->sync_thread);
4327 md_new_event(mddev);
4328 return 0;
4329
4330abort:
4331 mddev->recovery = 0;
4332 spin_lock_irq(&conf->device_lock);
4333 conf->geo = conf->prev;
4334 mddev->raid_disks = conf->geo.raid_disks;
4335 rdev_for_each(rdev, mddev)
4336 rdev->new_data_offset = rdev->data_offset;
4337 smp_wmb();
4338 conf->reshape_progress = MaxSector;
4339 conf->reshape_safe = MaxSector;
4340 mddev->reshape_position = MaxSector;
4341 spin_unlock_irq(&conf->device_lock);
4342 return ret;
4343}
4344
4345/* Calculate the last device-address that could contain
4346 * any block from the chunk that includes the array-address 's'
4347 * and report the next address.
4348 * i.e. the address returned will be chunk-aligned and after
4349 * any data that is in the chunk containing 's'.
4350 */
4351static sector_t last_dev_address(sector_t s, struct geom *geo)
4352{
4353 s = (s | geo->chunk_mask) + 1;
4354 s >>= geo->chunk_shift;
4355 s *= geo->near_copies;
4356 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4357 s *= geo->far_copies;
4358 s <<= geo->chunk_shift;
4359 return s;
4360}
4361
4362/* Calculate the first device-address that could contain
4363 * any block from the chunk that includes the array-address 's'.
4364 * This too will be the start of a chunk
4365 */
4366static sector_t first_dev_address(sector_t s, struct geom *geo)
4367{
4368 s >>= geo->chunk_shift;
4369 s *= geo->near_copies;
4370 sector_div(s, geo->raid_disks);
4371 s *= geo->far_copies;
4372 s <<= geo->chunk_shift;
4373 return s;
4374}
4375
4376static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4377 int *skipped)
4378{
4379 /* We simply copy at most one chunk (smallest of old and new)
4380 * at a time, possibly less if that exceeds RESYNC_PAGES,
4381 * or we hit a bad block or something.
4382 * This might mean we pause for normal IO in the middle of
4383 * a chunk, but that is not a problem as mddev->reshape_position
4384 * can record any location.
4385 *
4386 * If we will want to write to a location that isn't
4387 * yet recorded as 'safe' (i.e. in metadata on disk) then
4388 * we need to flush all reshape requests and update the metadata.
4389 *
4390 * When reshaping forwards (e.g. to more devices), we interpret
4391 * 'safe' as the earliest block which might not have been copied
4392 * down yet. We divide this by previous stripe size and multiply
4393 * by previous stripe length to get lowest device offset that we
4394 * cannot write to yet.
4395 * We interpret 'sector_nr' as an address that we want to write to.
4396 * From this we use last_device_address() to find where we might
4397 * write to, and first_device_address on the 'safe' position.
4398 * If this 'next' write position is after the 'safe' position,
4399 * we must update the metadata to increase the 'safe' position.
4400 *
4401 * When reshaping backwards, we round in the opposite direction
4402 * and perform the reverse test: next write position must not be
4403 * less than current safe position.
4404 *
4405 * In all this the minimum difference in data offsets
4406 * (conf->offset_diff - always positive) allows a bit of slack,
4407 * so next can be after 'safe', but not by more than offset_diff
4408 *
4409 * We need to prepare all the bios here before we start any IO
4410 * to ensure the size we choose is acceptable to all devices.
4411 * The means one for each copy for write-out and an extra one for
4412 * read-in.
4413 * We store the read-in bio in ->master_bio and the others in
4414 * ->devs[x].bio and ->devs[x].repl_bio.
4415 */
4416 struct r10conf *conf = mddev->private;
4417 struct r10bio *r10_bio;
4418 sector_t next, safe, last;
4419 int max_sectors;
4420 int nr_sectors;
4421 int s;
4422 struct md_rdev *rdev;
4423 int need_flush = 0;
4424 struct bio *blist;
4425 struct bio *bio, *read_bio;
4426 int sectors_done = 0;
4427 struct page **pages;
4428
4429 if (sector_nr == 0) {
4430 /* If restarting in the middle, skip the initial sectors */
4431 if (mddev->reshape_backwards &&
4432 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4433 sector_nr = (raid10_size(mddev, 0, 0)
4434 - conf->reshape_progress);
4435 } else if (!mddev->reshape_backwards &&
4436 conf->reshape_progress > 0)
4437 sector_nr = conf->reshape_progress;
4438 if (sector_nr) {
4439 mddev->curr_resync_completed = sector_nr;
4440 sysfs_notify_dirent_safe(mddev->sysfs_completed);
4441 *skipped = 1;
4442 return sector_nr;
4443 }
4444 }
4445
4446 /* We don't use sector_nr to track where we are up to
4447 * as that doesn't work well for ->reshape_backwards.
4448 * So just use ->reshape_progress.
4449 */
4450 if (mddev->reshape_backwards) {
4451 /* 'next' is the earliest device address that we might
4452 * write to for this chunk in the new layout
4453 */
4454 next = first_dev_address(conf->reshape_progress - 1,
4455 &conf->geo);
4456
4457 /* 'safe' is the last device address that we might read from
4458 * in the old layout after a restart
4459 */
4460 safe = last_dev_address(conf->reshape_safe - 1,
4461 &conf->prev);
4462
4463 if (next + conf->offset_diff < safe)
4464 need_flush = 1;
4465
4466 last = conf->reshape_progress - 1;
4467 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4468 & conf->prev.chunk_mask);
4469 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4470 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4471 } else {
4472 /* 'next' is after the last device address that we
4473 * might write to for this chunk in the new layout
4474 */
4475 next = last_dev_address(conf->reshape_progress, &conf->geo);
4476
4477 /* 'safe' is the earliest device address that we might
4478 * read from in the old layout after a restart
4479 */
4480 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4481
4482 /* Need to update metadata if 'next' might be beyond 'safe'
4483 * as that would possibly corrupt data
4484 */
4485 if (next > safe + conf->offset_diff)
4486 need_flush = 1;
4487
4488 sector_nr = conf->reshape_progress;
4489 last = sector_nr | (conf->geo.chunk_mask
4490 & conf->prev.chunk_mask);
4491
4492 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4493 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4494 }
4495
4496 if (need_flush ||
4497 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4498 /* Need to update reshape_position in metadata */
4499 wait_barrier(conf);
4500 mddev->reshape_position = conf->reshape_progress;
4501 if (mddev->reshape_backwards)
4502 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4503 - conf->reshape_progress;
4504 else
4505 mddev->curr_resync_completed = conf->reshape_progress;
4506 conf->reshape_checkpoint = jiffies;
4507 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4508 md_wakeup_thread(mddev->thread);
4509 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4510 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4511 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4512 allow_barrier(conf);
4513 return sectors_done;
4514 }
4515 conf->reshape_safe = mddev->reshape_position;
4516 allow_barrier(conf);
4517 }
4518
4519 raise_barrier(conf, 0);
4520read_more:
4521 /* Now schedule reads for blocks from sector_nr to last */
4522 r10_bio = raid10_alloc_init_r10buf(conf);
4523 r10_bio->state = 0;
4524 raise_barrier(conf, 1);
4525 atomic_set(&r10_bio->remaining, 0);
4526 r10_bio->mddev = mddev;
4527 r10_bio->sector = sector_nr;
4528 set_bit(R10BIO_IsReshape, &r10_bio->state);
4529 r10_bio->sectors = last - sector_nr + 1;
4530 rdev = read_balance(conf, r10_bio, &max_sectors);
4531 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4532
4533 if (!rdev) {
4534 /* Cannot read from here, so need to record bad blocks
4535 * on all the target devices.
4536 */
4537 // FIXME
4538 mempool_free(r10_bio, &conf->r10buf_pool);
4539 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4540 return sectors_done;
4541 }
4542
4543 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4544
4545 bio_set_dev(read_bio, rdev->bdev);
4546 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4547 + rdev->data_offset);
4548 read_bio->bi_private = r10_bio;
4549 read_bio->bi_end_io = end_reshape_read;
4550 bio_set_op_attrs(read_bio, REQ_OP_READ, 0);
4551 read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4552 read_bio->bi_status = 0;
4553 read_bio->bi_vcnt = 0;
4554 read_bio->bi_iter.bi_size = 0;
4555 r10_bio->master_bio = read_bio;
4556 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4557
4558 /*
4559 * Broadcast RESYNC message to other nodes, so all nodes would not
4560 * write to the region to avoid conflict.
4561 */
4562 if (mddev_is_clustered(mddev) && conf->cluster_sync_high <= sector_nr) {
4563 struct mdp_superblock_1 *sb = NULL;
4564 int sb_reshape_pos = 0;
4565
4566 conf->cluster_sync_low = sector_nr;
4567 conf->cluster_sync_high = sector_nr + CLUSTER_RESYNC_WINDOW_SECTORS;
4568 sb = page_address(rdev->sb_page);
4569 if (sb) {
4570 sb_reshape_pos = le64_to_cpu(sb->reshape_position);
4571 /*
4572 * Set cluster_sync_low again if next address for array
4573 * reshape is less than cluster_sync_low. Since we can't
4574 * update cluster_sync_low until it has finished reshape.
4575 */
4576 if (sb_reshape_pos < conf->cluster_sync_low)
4577 conf->cluster_sync_low = sb_reshape_pos;
4578 }
4579
4580 md_cluster_ops->resync_info_update(mddev, conf->cluster_sync_low,
4581 conf->cluster_sync_high);
4582 }
4583
4584 /* Now find the locations in the new layout */
4585 __raid10_find_phys(&conf->geo, r10_bio);
4586
4587 blist = read_bio;
4588 read_bio->bi_next = NULL;
4589
4590 rcu_read_lock();
4591 for (s = 0; s < conf->copies*2; s++) {
4592 struct bio *b;
4593 int d = r10_bio->devs[s/2].devnum;
4594 struct md_rdev *rdev2;
4595 if (s&1) {
4596 rdev2 = rcu_dereference(conf->mirrors[d].replacement);
4597 b = r10_bio->devs[s/2].repl_bio;
4598 } else {
4599 rdev2 = rcu_dereference(conf->mirrors[d].rdev);
4600 b = r10_bio->devs[s/2].bio;
4601 }
4602 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4603 continue;
4604
4605 bio_set_dev(b, rdev2->bdev);
4606 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4607 rdev2->new_data_offset;
4608 b->bi_end_io = end_reshape_write;
4609 bio_set_op_attrs(b, REQ_OP_WRITE, 0);
4610 b->bi_next = blist;
4611 blist = b;
4612 }
4613
4614 /* Now add as many pages as possible to all of these bios. */
4615
4616 nr_sectors = 0;
4617 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4618 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4619 struct page *page = pages[s / (PAGE_SIZE >> 9)];
4620 int len = (max_sectors - s) << 9;
4621 if (len > PAGE_SIZE)
4622 len = PAGE_SIZE;
4623 for (bio = blist; bio ; bio = bio->bi_next) {
4624 /*
4625 * won't fail because the vec table is big enough
4626 * to hold all these pages
4627 */
4628 bio_add_page(bio, page, len, 0);
4629 }
4630 sector_nr += len >> 9;
4631 nr_sectors += len >> 9;
4632 }
4633 rcu_read_unlock();
4634 r10_bio->sectors = nr_sectors;
4635
4636 /* Now submit the read */
4637 md_sync_acct_bio(read_bio, r10_bio->sectors);
4638 atomic_inc(&r10_bio->remaining);
4639 read_bio->bi_next = NULL;
4640 submit_bio_noacct(read_bio);
4641 sectors_done += nr_sectors;
4642 if (sector_nr <= last)
4643 goto read_more;
4644
4645 lower_barrier(conf);
4646
4647 /* Now that we have done the whole section we can
4648 * update reshape_progress
4649 */
4650 if (mddev->reshape_backwards)
4651 conf->reshape_progress -= sectors_done;
4652 else
4653 conf->reshape_progress += sectors_done;
4654
4655 return sectors_done;
4656}
4657
4658static void end_reshape_request(struct r10bio *r10_bio);
4659static int handle_reshape_read_error(struct mddev *mddev,
4660 struct r10bio *r10_bio);
4661static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4662{
4663 /* Reshape read completed. Hopefully we have a block
4664 * to write out.
4665 * If we got a read error then we do sync 1-page reads from
4666 * elsewhere until we find the data - or give up.
4667 */
4668 struct r10conf *conf = mddev->private;
4669 int s;
4670
4671 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4672 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4673 /* Reshape has been aborted */
4674 md_done_sync(mddev, r10_bio->sectors, 0);
4675 return;
4676 }
4677
4678 /* We definitely have the data in the pages, schedule the
4679 * writes.
4680 */
4681 atomic_set(&r10_bio->remaining, 1);
4682 for (s = 0; s < conf->copies*2; s++) {
4683 struct bio *b;
4684 int d = r10_bio->devs[s/2].devnum;
4685 struct md_rdev *rdev;
4686 rcu_read_lock();
4687 if (s&1) {
4688 rdev = rcu_dereference(conf->mirrors[d].replacement);
4689 b = r10_bio->devs[s/2].repl_bio;
4690 } else {
4691 rdev = rcu_dereference(conf->mirrors[d].rdev);
4692 b = r10_bio->devs[s/2].bio;
4693 }
4694 if (!rdev || test_bit(Faulty, &rdev->flags)) {
4695 rcu_read_unlock();
4696 continue;
4697 }
4698 atomic_inc(&rdev->nr_pending);
4699 rcu_read_unlock();
4700 md_sync_acct_bio(b, r10_bio->sectors);
4701 atomic_inc(&r10_bio->remaining);
4702 b->bi_next = NULL;
4703 submit_bio_noacct(b);
4704 }
4705 end_reshape_request(r10_bio);
4706}
4707
4708static void end_reshape(struct r10conf *conf)
4709{
4710 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4711 return;
4712
4713 spin_lock_irq(&conf->device_lock);
4714 conf->prev = conf->geo;
4715 md_finish_reshape(conf->mddev);
4716 smp_wmb();
4717 conf->reshape_progress = MaxSector;
4718 conf->reshape_safe = MaxSector;
4719 spin_unlock_irq(&conf->device_lock);
4720
4721 /* read-ahead size must cover two whole stripes, which is
4722 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4723 */
4724 if (conf->mddev->queue) {
4725 int stripe = conf->geo.raid_disks *
4726 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4727 stripe /= conf->geo.near_copies;
4728 if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
4729 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
4730 }
4731 conf->fullsync = 0;
4732}
4733
4734static void raid10_update_reshape_pos(struct mddev *mddev)
4735{
4736 struct r10conf *conf = mddev->private;
4737 sector_t lo, hi;
4738
4739 md_cluster_ops->resync_info_get(mddev, &lo, &hi);
4740 if (((mddev->reshape_position <= hi) && (mddev->reshape_position >= lo))
4741 || mddev->reshape_position == MaxSector)
4742 conf->reshape_progress = mddev->reshape_position;
4743 else
4744 WARN_ON_ONCE(1);
4745}
4746
4747static int handle_reshape_read_error(struct mddev *mddev,
4748 struct r10bio *r10_bio)
4749{
4750 /* Use sync reads to get the blocks from somewhere else */
4751 int sectors = r10_bio->sectors;
4752 struct r10conf *conf = mddev->private;
4753 struct r10bio *r10b;
4754 int slot = 0;
4755 int idx = 0;
4756 struct page **pages;
4757
4758 r10b = kmalloc(struct_size(r10b, devs, conf->copies), GFP_NOIO);
4759 if (!r10b) {
4760 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4761 return -ENOMEM;
4762 }
4763
4764 /* reshape IOs share pages from .devs[0].bio */
4765 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4766
4767 r10b->sector = r10_bio->sector;
4768 __raid10_find_phys(&conf->prev, r10b);
4769
4770 while (sectors) {
4771 int s = sectors;
4772 int success = 0;
4773 int first_slot = slot;
4774
4775 if (s > (PAGE_SIZE >> 9))
4776 s = PAGE_SIZE >> 9;
4777
4778 rcu_read_lock();
4779 while (!success) {
4780 int d = r10b->devs[slot].devnum;
4781 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4782 sector_t addr;
4783 if (rdev == NULL ||
4784 test_bit(Faulty, &rdev->flags) ||
4785 !test_bit(In_sync, &rdev->flags))
4786 goto failed;
4787
4788 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4789 atomic_inc(&rdev->nr_pending);
4790 rcu_read_unlock();
4791 success = sync_page_io(rdev,
4792 addr,
4793 s << 9,
4794 pages[idx],
4795 REQ_OP_READ, 0, false);
4796 rdev_dec_pending(rdev, mddev);
4797 rcu_read_lock();
4798 if (success)
4799 break;
4800 failed:
4801 slot++;
4802 if (slot >= conf->copies)
4803 slot = 0;
4804 if (slot == first_slot)
4805 break;
4806 }
4807 rcu_read_unlock();
4808 if (!success) {
4809 /* couldn't read this block, must give up */
4810 set_bit(MD_RECOVERY_INTR,
4811 &mddev->recovery);
4812 kfree(r10b);
4813 return -EIO;
4814 }
4815 sectors -= s;
4816 idx++;
4817 }
4818 kfree(r10b);
4819 return 0;
4820}
4821
4822static void end_reshape_write(struct bio *bio)
4823{
4824 struct r10bio *r10_bio = get_resync_r10bio(bio);
4825 struct mddev *mddev = r10_bio->mddev;
4826 struct r10conf *conf = mddev->private;
4827 int d;
4828 int slot;
4829 int repl;
4830 struct md_rdev *rdev = NULL;
4831
4832 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4833 if (repl)
4834 rdev = conf->mirrors[d].replacement;
4835 if (!rdev) {
4836 smp_mb();
4837 rdev = conf->mirrors[d].rdev;
4838 }
4839
4840 if (bio->bi_status) {
4841 /* FIXME should record badblock */
4842 md_error(mddev, rdev);
4843 }
4844
4845 rdev_dec_pending(rdev, mddev);
4846 end_reshape_request(r10_bio);
4847}
4848
4849static void end_reshape_request(struct r10bio *r10_bio)
4850{
4851 if (!atomic_dec_and_test(&r10_bio->remaining))
4852 return;
4853 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4854 bio_put(r10_bio->master_bio);
4855 put_buf(r10_bio);
4856}
4857
4858static void raid10_finish_reshape(struct mddev *mddev)
4859{
4860 struct r10conf *conf = mddev->private;
4861
4862 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4863 return;
4864
4865 if (mddev->delta_disks > 0) {
4866 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4867 mddev->recovery_cp = mddev->resync_max_sectors;
4868 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4869 }
4870 mddev->resync_max_sectors = mddev->array_sectors;
4871 } else {
4872 int d;
4873 rcu_read_lock();
4874 for (d = conf->geo.raid_disks ;
4875 d < conf->geo.raid_disks - mddev->delta_disks;
4876 d++) {
4877 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4878 if (rdev)
4879 clear_bit(In_sync, &rdev->flags);
4880 rdev = rcu_dereference(conf->mirrors[d].replacement);
4881 if (rdev)
4882 clear_bit(In_sync, &rdev->flags);
4883 }
4884 rcu_read_unlock();
4885 }
4886 mddev->layout = mddev->new_layout;
4887 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4888 mddev->reshape_position = MaxSector;
4889 mddev->delta_disks = 0;
4890 mddev->reshape_backwards = 0;
4891}
4892
4893static struct md_personality raid10_personality =
4894{
4895 .name = "raid10",
4896 .level = 10,
4897 .owner = THIS_MODULE,
4898 .make_request = raid10_make_request,
4899 .run = raid10_run,
4900 .free = raid10_free,
4901 .status = raid10_status,
4902 .error_handler = raid10_error,
4903 .hot_add_disk = raid10_add_disk,
4904 .hot_remove_disk= raid10_remove_disk,
4905 .spare_active = raid10_spare_active,
4906 .sync_request = raid10_sync_request,
4907 .quiesce = raid10_quiesce,
4908 .size = raid10_size,
4909 .resize = raid10_resize,
4910 .takeover = raid10_takeover,
4911 .check_reshape = raid10_check_reshape,
4912 .start_reshape = raid10_start_reshape,
4913 .finish_reshape = raid10_finish_reshape,
4914 .update_reshape_pos = raid10_update_reshape_pos,
4915};
4916
4917static int __init raid_init(void)
4918{
4919 return register_md_personality(&raid10_personality);
4920}
4921
4922static void raid_exit(void)
4923{
4924 unregister_md_personality(&raid10_personality);
4925}
4926
4927module_init(raid_init);
4928module_exit(raid_exit);
4929MODULE_LICENSE("GPL");
4930MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4931MODULE_ALIAS("md-personality-9"); /* RAID10 */
4932MODULE_ALIAS("md-raid10");
4933MODULE_ALIAS("md-level-10");
4934
4935module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);