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 * 2]);
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 int raid10_congested(struct mddev *mddev, int bits)
852{
853 struct r10conf *conf = mddev->private;
854 int i, ret = 0;
855
856 if ((bits & (1 << WB_async_congested)) &&
857 conf->pending_count >= max_queued_requests)
858 return 1;
859
860 rcu_read_lock();
861 for (i = 0;
862 (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
863 && ret == 0;
864 i++) {
865 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
866 if (rdev && !test_bit(Faulty, &rdev->flags)) {
867 struct request_queue *q = bdev_get_queue(rdev->bdev);
868
869 ret |= bdi_congested(q->backing_dev_info, bits);
870 }
871 }
872 rcu_read_unlock();
873 return ret;
874}
875
876static void flush_pending_writes(struct r10conf *conf)
877{
878 /* Any writes that have been queued but are awaiting
879 * bitmap updates get flushed here.
880 */
881 spin_lock_irq(&conf->device_lock);
882
883 if (conf->pending_bio_list.head) {
884 struct blk_plug plug;
885 struct bio *bio;
886
887 bio = bio_list_get(&conf->pending_bio_list);
888 conf->pending_count = 0;
889 spin_unlock_irq(&conf->device_lock);
890
891 /*
892 * As this is called in a wait_event() loop (see freeze_array),
893 * current->state might be TASK_UNINTERRUPTIBLE which will
894 * cause a warning when we prepare to wait again. As it is
895 * rare that this path is taken, it is perfectly safe to force
896 * us to go around the wait_event() loop again, so the warning
897 * is a false-positive. Silence the warning by resetting
898 * thread state
899 */
900 __set_current_state(TASK_RUNNING);
901
902 blk_start_plug(&plug);
903 /* flush any pending bitmap writes to disk
904 * before proceeding w/ I/O */
905 md_bitmap_unplug(conf->mddev->bitmap);
906 wake_up(&conf->wait_barrier);
907
908 while (bio) { /* submit pending writes */
909 struct bio *next = bio->bi_next;
910 struct md_rdev *rdev = (void*)bio->bi_disk;
911 bio->bi_next = NULL;
912 bio_set_dev(bio, rdev->bdev);
913 if (test_bit(Faulty, &rdev->flags)) {
914 bio_io_error(bio);
915 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
916 !blk_queue_discard(bio->bi_disk->queue)))
917 /* Just ignore it */
918 bio_endio(bio);
919 else
920 generic_make_request(bio);
921 bio = next;
922 }
923 blk_finish_plug(&plug);
924 } else
925 spin_unlock_irq(&conf->device_lock);
926}
927
928/* Barriers....
929 * Sometimes we need to suspend IO while we do something else,
930 * either some resync/recovery, or reconfigure the array.
931 * To do this we raise a 'barrier'.
932 * The 'barrier' is a counter that can be raised multiple times
933 * to count how many activities are happening which preclude
934 * normal IO.
935 * We can only raise the barrier if there is no pending IO.
936 * i.e. if nr_pending == 0.
937 * We choose only to raise the barrier if no-one is waiting for the
938 * barrier to go down. This means that as soon as an IO request
939 * is ready, no other operations which require a barrier will start
940 * until the IO request has had a chance.
941 *
942 * So: regular IO calls 'wait_barrier'. When that returns there
943 * is no backgroup IO happening, It must arrange to call
944 * allow_barrier when it has finished its IO.
945 * backgroup IO calls must call raise_barrier. Once that returns
946 * there is no normal IO happeing. It must arrange to call
947 * lower_barrier when the particular background IO completes.
948 */
949
950static void raise_barrier(struct r10conf *conf, int force)
951{
952 BUG_ON(force && !conf->barrier);
953 spin_lock_irq(&conf->resync_lock);
954
955 /* Wait until no block IO is waiting (unless 'force') */
956 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
957 conf->resync_lock);
958
959 /* block any new IO from starting */
960 conf->barrier++;
961
962 /* Now wait for all pending IO to complete */
963 wait_event_lock_irq(conf->wait_barrier,
964 !atomic_read(&conf->nr_pending) && conf->barrier < RESYNC_DEPTH,
965 conf->resync_lock);
966
967 spin_unlock_irq(&conf->resync_lock);
968}
969
970static void lower_barrier(struct r10conf *conf)
971{
972 unsigned long flags;
973 spin_lock_irqsave(&conf->resync_lock, flags);
974 conf->barrier--;
975 spin_unlock_irqrestore(&conf->resync_lock, flags);
976 wake_up(&conf->wait_barrier);
977}
978
979static void wait_barrier(struct r10conf *conf)
980{
981 spin_lock_irq(&conf->resync_lock);
982 if (conf->barrier) {
983 conf->nr_waiting++;
984 /* Wait for the barrier to drop.
985 * However if there are already pending
986 * requests (preventing the barrier from
987 * rising completely), and the
988 * pre-process bio queue isn't empty,
989 * then don't wait, as we need to empty
990 * that queue to get the nr_pending
991 * count down.
992 */
993 raid10_log(conf->mddev, "wait barrier");
994 wait_event_lock_irq(conf->wait_barrier,
995 !conf->barrier ||
996 (atomic_read(&conf->nr_pending) &&
997 current->bio_list &&
998 (!bio_list_empty(¤t->bio_list[0]) ||
999 !bio_list_empty(¤t->bio_list[1]))),
1000 conf->resync_lock);
1001 conf->nr_waiting--;
1002 if (!conf->nr_waiting)
1003 wake_up(&conf->wait_barrier);
1004 }
1005 atomic_inc(&conf->nr_pending);
1006 spin_unlock_irq(&conf->resync_lock);
1007}
1008
1009static void allow_barrier(struct r10conf *conf)
1010{
1011 if ((atomic_dec_and_test(&conf->nr_pending)) ||
1012 (conf->array_freeze_pending))
1013 wake_up(&conf->wait_barrier);
1014}
1015
1016static void freeze_array(struct r10conf *conf, int extra)
1017{
1018 /* stop syncio and normal IO and wait for everything to
1019 * go quiet.
1020 * We increment barrier and nr_waiting, and then
1021 * wait until nr_pending match nr_queued+extra
1022 * This is called in the context of one normal IO request
1023 * that has failed. Thus any sync request that might be pending
1024 * will be blocked by nr_pending, and we need to wait for
1025 * pending IO requests to complete or be queued for re-try.
1026 * Thus the number queued (nr_queued) plus this request (extra)
1027 * must match the number of pending IOs (nr_pending) before
1028 * we continue.
1029 */
1030 spin_lock_irq(&conf->resync_lock);
1031 conf->array_freeze_pending++;
1032 conf->barrier++;
1033 conf->nr_waiting++;
1034 wait_event_lock_irq_cmd(conf->wait_barrier,
1035 atomic_read(&conf->nr_pending) == conf->nr_queued+extra,
1036 conf->resync_lock,
1037 flush_pending_writes(conf));
1038
1039 conf->array_freeze_pending--;
1040 spin_unlock_irq(&conf->resync_lock);
1041}
1042
1043static void unfreeze_array(struct r10conf *conf)
1044{
1045 /* reverse the effect of the freeze */
1046 spin_lock_irq(&conf->resync_lock);
1047 conf->barrier--;
1048 conf->nr_waiting--;
1049 wake_up(&conf->wait_barrier);
1050 spin_unlock_irq(&conf->resync_lock);
1051}
1052
1053static sector_t choose_data_offset(struct r10bio *r10_bio,
1054 struct md_rdev *rdev)
1055{
1056 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1057 test_bit(R10BIO_Previous, &r10_bio->state))
1058 return rdev->data_offset;
1059 else
1060 return rdev->new_data_offset;
1061}
1062
1063struct raid10_plug_cb {
1064 struct blk_plug_cb cb;
1065 struct bio_list pending;
1066 int pending_cnt;
1067};
1068
1069static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1070{
1071 struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1072 cb);
1073 struct mddev *mddev = plug->cb.data;
1074 struct r10conf *conf = mddev->private;
1075 struct bio *bio;
1076
1077 if (from_schedule || current->bio_list) {
1078 spin_lock_irq(&conf->device_lock);
1079 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1080 conf->pending_count += plug->pending_cnt;
1081 spin_unlock_irq(&conf->device_lock);
1082 wake_up(&conf->wait_barrier);
1083 md_wakeup_thread(mddev->thread);
1084 kfree(plug);
1085 return;
1086 }
1087
1088 /* we aren't scheduling, so we can do the write-out directly. */
1089 bio = bio_list_get(&plug->pending);
1090 md_bitmap_unplug(mddev->bitmap);
1091 wake_up(&conf->wait_barrier);
1092
1093 while (bio) { /* submit pending writes */
1094 struct bio *next = bio->bi_next;
1095 struct md_rdev *rdev = (void*)bio->bi_disk;
1096 bio->bi_next = NULL;
1097 bio_set_dev(bio, rdev->bdev);
1098 if (test_bit(Faulty, &rdev->flags)) {
1099 bio_io_error(bio);
1100 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
1101 !blk_queue_discard(bio->bi_disk->queue)))
1102 /* Just ignore it */
1103 bio_endio(bio);
1104 else
1105 generic_make_request(bio);
1106 bio = next;
1107 }
1108 kfree(plug);
1109}
1110
1111/*
1112 * 1. Register the new request and wait if the reconstruction thread has put
1113 * up a bar for new requests. Continue immediately if no resync is active
1114 * currently.
1115 * 2. If IO spans the reshape position. Need to wait for reshape to pass.
1116 */
1117static void regular_request_wait(struct mddev *mddev, struct r10conf *conf,
1118 struct bio *bio, sector_t sectors)
1119{
1120 wait_barrier(conf);
1121 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1122 bio->bi_iter.bi_sector < conf->reshape_progress &&
1123 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1124 raid10_log(conf->mddev, "wait reshape");
1125 allow_barrier(conf);
1126 wait_event(conf->wait_barrier,
1127 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1128 conf->reshape_progress >= bio->bi_iter.bi_sector +
1129 sectors);
1130 wait_barrier(conf);
1131 }
1132}
1133
1134static void raid10_read_request(struct mddev *mddev, struct bio *bio,
1135 struct r10bio *r10_bio)
1136{
1137 struct r10conf *conf = mddev->private;
1138 struct bio *read_bio;
1139 const int op = bio_op(bio);
1140 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1141 int max_sectors;
1142 struct md_rdev *rdev;
1143 char b[BDEVNAME_SIZE];
1144 int slot = r10_bio->read_slot;
1145 struct md_rdev *err_rdev = NULL;
1146 gfp_t gfp = GFP_NOIO;
1147
1148 if (r10_bio->devs[slot].rdev) {
1149 /*
1150 * This is an error retry, but we cannot
1151 * safely dereference the rdev in the r10_bio,
1152 * we must use the one in conf.
1153 * If it has already been disconnected (unlikely)
1154 * we lose the device name in error messages.
1155 */
1156 int disk;
1157 /*
1158 * As we are blocking raid10, it is a little safer to
1159 * use __GFP_HIGH.
1160 */
1161 gfp = GFP_NOIO | __GFP_HIGH;
1162
1163 rcu_read_lock();
1164 disk = r10_bio->devs[slot].devnum;
1165 err_rdev = rcu_dereference(conf->mirrors[disk].rdev);
1166 if (err_rdev)
1167 bdevname(err_rdev->bdev, b);
1168 else {
1169 strcpy(b, "???");
1170 /* This never gets dereferenced */
1171 err_rdev = r10_bio->devs[slot].rdev;
1172 }
1173 rcu_read_unlock();
1174 }
1175
1176 regular_request_wait(mddev, conf, bio, r10_bio->sectors);
1177 rdev = read_balance(conf, r10_bio, &max_sectors);
1178 if (!rdev) {
1179 if (err_rdev) {
1180 pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
1181 mdname(mddev), b,
1182 (unsigned long long)r10_bio->sector);
1183 }
1184 raid_end_bio_io(r10_bio);
1185 return;
1186 }
1187 if (err_rdev)
1188 pr_err_ratelimited("md/raid10:%s: %s: redirecting sector %llu to another mirror\n",
1189 mdname(mddev),
1190 bdevname(rdev->bdev, b),
1191 (unsigned long long)r10_bio->sector);
1192 if (max_sectors < bio_sectors(bio)) {
1193 struct bio *split = bio_split(bio, max_sectors,
1194 gfp, &conf->bio_split);
1195 bio_chain(split, bio);
1196 allow_barrier(conf);
1197 generic_make_request(bio);
1198 wait_barrier(conf);
1199 bio = split;
1200 r10_bio->master_bio = bio;
1201 r10_bio->sectors = max_sectors;
1202 }
1203 slot = r10_bio->read_slot;
1204
1205 read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1206
1207 r10_bio->devs[slot].bio = read_bio;
1208 r10_bio->devs[slot].rdev = rdev;
1209
1210 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1211 choose_data_offset(r10_bio, rdev);
1212 bio_set_dev(read_bio, rdev->bdev);
1213 read_bio->bi_end_io = raid10_end_read_request;
1214 bio_set_op_attrs(read_bio, op, do_sync);
1215 if (test_bit(FailFast, &rdev->flags) &&
1216 test_bit(R10BIO_FailFast, &r10_bio->state))
1217 read_bio->bi_opf |= MD_FAILFAST;
1218 read_bio->bi_private = r10_bio;
1219
1220 if (mddev->gendisk)
1221 trace_block_bio_remap(read_bio->bi_disk->queue,
1222 read_bio, disk_devt(mddev->gendisk),
1223 r10_bio->sector);
1224 generic_make_request(read_bio);
1225 return;
1226}
1227
1228static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio,
1229 struct bio *bio, bool replacement,
1230 int n_copy)
1231{
1232 const int op = bio_op(bio);
1233 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1234 const unsigned long do_fua = (bio->bi_opf & REQ_FUA);
1235 unsigned long flags;
1236 struct blk_plug_cb *cb;
1237 struct raid10_plug_cb *plug = NULL;
1238 struct r10conf *conf = mddev->private;
1239 struct md_rdev *rdev;
1240 int devnum = r10_bio->devs[n_copy].devnum;
1241 struct bio *mbio;
1242
1243 if (replacement) {
1244 rdev = conf->mirrors[devnum].replacement;
1245 if (rdev == NULL) {
1246 /* Replacement just got moved to main 'rdev' */
1247 smp_mb();
1248 rdev = conf->mirrors[devnum].rdev;
1249 }
1250 } else
1251 rdev = conf->mirrors[devnum].rdev;
1252
1253 mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1254 if (replacement)
1255 r10_bio->devs[n_copy].repl_bio = mbio;
1256 else
1257 r10_bio->devs[n_copy].bio = mbio;
1258
1259 mbio->bi_iter.bi_sector = (r10_bio->devs[n_copy].addr +
1260 choose_data_offset(r10_bio, rdev));
1261 bio_set_dev(mbio, rdev->bdev);
1262 mbio->bi_end_io = raid10_end_write_request;
1263 bio_set_op_attrs(mbio, op, do_sync | do_fua);
1264 if (!replacement && test_bit(FailFast,
1265 &conf->mirrors[devnum].rdev->flags)
1266 && enough(conf, devnum))
1267 mbio->bi_opf |= MD_FAILFAST;
1268 mbio->bi_private = r10_bio;
1269
1270 if (conf->mddev->gendisk)
1271 trace_block_bio_remap(mbio->bi_disk->queue,
1272 mbio, disk_devt(conf->mddev->gendisk),
1273 r10_bio->sector);
1274 /* flush_pending_writes() needs access to the rdev so...*/
1275 mbio->bi_disk = (void *)rdev;
1276
1277 atomic_inc(&r10_bio->remaining);
1278
1279 cb = blk_check_plugged(raid10_unplug, mddev, sizeof(*plug));
1280 if (cb)
1281 plug = container_of(cb, struct raid10_plug_cb, cb);
1282 else
1283 plug = NULL;
1284 if (plug) {
1285 bio_list_add(&plug->pending, mbio);
1286 plug->pending_cnt++;
1287 } else {
1288 spin_lock_irqsave(&conf->device_lock, flags);
1289 bio_list_add(&conf->pending_bio_list, mbio);
1290 conf->pending_count++;
1291 spin_unlock_irqrestore(&conf->device_lock, flags);
1292 md_wakeup_thread(mddev->thread);
1293 }
1294}
1295
1296static void raid10_write_request(struct mddev *mddev, struct bio *bio,
1297 struct r10bio *r10_bio)
1298{
1299 struct r10conf *conf = mddev->private;
1300 int i;
1301 struct md_rdev *blocked_rdev;
1302 sector_t sectors;
1303 int max_sectors;
1304
1305 if ((mddev_is_clustered(mddev) &&
1306 md_cluster_ops->area_resyncing(mddev, WRITE,
1307 bio->bi_iter.bi_sector,
1308 bio_end_sector(bio)))) {
1309 DEFINE_WAIT(w);
1310 for (;;) {
1311 prepare_to_wait(&conf->wait_barrier,
1312 &w, TASK_IDLE);
1313 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1314 bio->bi_iter.bi_sector, bio_end_sector(bio)))
1315 break;
1316 schedule();
1317 }
1318 finish_wait(&conf->wait_barrier, &w);
1319 }
1320
1321 sectors = r10_bio->sectors;
1322 regular_request_wait(mddev, conf, bio, sectors);
1323 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1324 (mddev->reshape_backwards
1325 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1326 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1327 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1328 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1329 /* Need to update reshape_position in metadata */
1330 mddev->reshape_position = conf->reshape_progress;
1331 set_mask_bits(&mddev->sb_flags, 0,
1332 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1333 md_wakeup_thread(mddev->thread);
1334 raid10_log(conf->mddev, "wait reshape metadata");
1335 wait_event(mddev->sb_wait,
1336 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
1337
1338 conf->reshape_safe = mddev->reshape_position;
1339 }
1340
1341 if (conf->pending_count >= max_queued_requests) {
1342 md_wakeup_thread(mddev->thread);
1343 raid10_log(mddev, "wait queued");
1344 wait_event(conf->wait_barrier,
1345 conf->pending_count < max_queued_requests);
1346 }
1347 /* first select target devices under rcu_lock and
1348 * inc refcount on their rdev. Record them by setting
1349 * bios[x] to bio
1350 * If there are known/acknowledged bad blocks on any device
1351 * on which we have seen a write error, we want to avoid
1352 * writing to those blocks. This potentially requires several
1353 * writes to write around the bad blocks. Each set of writes
1354 * gets its own r10_bio with a set of bios attached.
1355 */
1356
1357 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1358 raid10_find_phys(conf, r10_bio);
1359retry_write:
1360 blocked_rdev = NULL;
1361 rcu_read_lock();
1362 max_sectors = r10_bio->sectors;
1363
1364 for (i = 0; i < conf->copies; i++) {
1365 int d = r10_bio->devs[i].devnum;
1366 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1367 struct md_rdev *rrdev = rcu_dereference(
1368 conf->mirrors[d].replacement);
1369 if (rdev == rrdev)
1370 rrdev = NULL;
1371 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1372 atomic_inc(&rdev->nr_pending);
1373 blocked_rdev = rdev;
1374 break;
1375 }
1376 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1377 atomic_inc(&rrdev->nr_pending);
1378 blocked_rdev = rrdev;
1379 break;
1380 }
1381 if (rdev && (test_bit(Faulty, &rdev->flags)))
1382 rdev = NULL;
1383 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1384 rrdev = NULL;
1385
1386 r10_bio->devs[i].bio = NULL;
1387 r10_bio->devs[i].repl_bio = NULL;
1388
1389 if (!rdev && !rrdev) {
1390 set_bit(R10BIO_Degraded, &r10_bio->state);
1391 continue;
1392 }
1393 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1394 sector_t first_bad;
1395 sector_t dev_sector = r10_bio->devs[i].addr;
1396 int bad_sectors;
1397 int is_bad;
1398
1399 is_bad = is_badblock(rdev, dev_sector, max_sectors,
1400 &first_bad, &bad_sectors);
1401 if (is_bad < 0) {
1402 /* Mustn't write here until the bad block
1403 * is acknowledged
1404 */
1405 atomic_inc(&rdev->nr_pending);
1406 set_bit(BlockedBadBlocks, &rdev->flags);
1407 blocked_rdev = rdev;
1408 break;
1409 }
1410 if (is_bad && first_bad <= dev_sector) {
1411 /* Cannot write here at all */
1412 bad_sectors -= (dev_sector - first_bad);
1413 if (bad_sectors < max_sectors)
1414 /* Mustn't write more than bad_sectors
1415 * to other devices yet
1416 */
1417 max_sectors = bad_sectors;
1418 /* We don't set R10BIO_Degraded as that
1419 * only applies if the disk is missing,
1420 * so it might be re-added, and we want to
1421 * know to recover this chunk.
1422 * In this case the device is here, and the
1423 * fact that this chunk is not in-sync is
1424 * recorded in the bad block log.
1425 */
1426 continue;
1427 }
1428 if (is_bad) {
1429 int good_sectors = first_bad - dev_sector;
1430 if (good_sectors < max_sectors)
1431 max_sectors = good_sectors;
1432 }
1433 }
1434 if (rdev) {
1435 r10_bio->devs[i].bio = bio;
1436 atomic_inc(&rdev->nr_pending);
1437 }
1438 if (rrdev) {
1439 r10_bio->devs[i].repl_bio = bio;
1440 atomic_inc(&rrdev->nr_pending);
1441 }
1442 }
1443 rcu_read_unlock();
1444
1445 if (unlikely(blocked_rdev)) {
1446 /* Have to wait for this device to get unblocked, then retry */
1447 int j;
1448 int d;
1449
1450 for (j = 0; j < i; j++) {
1451 if (r10_bio->devs[j].bio) {
1452 d = r10_bio->devs[j].devnum;
1453 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1454 }
1455 if (r10_bio->devs[j].repl_bio) {
1456 struct md_rdev *rdev;
1457 d = r10_bio->devs[j].devnum;
1458 rdev = conf->mirrors[d].replacement;
1459 if (!rdev) {
1460 /* Race with remove_disk */
1461 smp_mb();
1462 rdev = conf->mirrors[d].rdev;
1463 }
1464 rdev_dec_pending(rdev, mddev);
1465 }
1466 }
1467 allow_barrier(conf);
1468 raid10_log(conf->mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1469 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1470 wait_barrier(conf);
1471 goto retry_write;
1472 }
1473
1474 if (max_sectors < r10_bio->sectors)
1475 r10_bio->sectors = max_sectors;
1476
1477 if (r10_bio->sectors < bio_sectors(bio)) {
1478 struct bio *split = bio_split(bio, r10_bio->sectors,
1479 GFP_NOIO, &conf->bio_split);
1480 bio_chain(split, bio);
1481 allow_barrier(conf);
1482 generic_make_request(bio);
1483 wait_barrier(conf);
1484 bio = split;
1485 r10_bio->master_bio = bio;
1486 }
1487
1488 atomic_set(&r10_bio->remaining, 1);
1489 md_bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1490
1491 for (i = 0; i < conf->copies; i++) {
1492 if (r10_bio->devs[i].bio)
1493 raid10_write_one_disk(mddev, r10_bio, bio, false, i);
1494 if (r10_bio->devs[i].repl_bio)
1495 raid10_write_one_disk(mddev, r10_bio, bio, true, i);
1496 }
1497 one_write_done(r10_bio);
1498}
1499
1500static void __make_request(struct mddev *mddev, struct bio *bio, int sectors)
1501{
1502 struct r10conf *conf = mddev->private;
1503 struct r10bio *r10_bio;
1504
1505 r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1506
1507 r10_bio->master_bio = bio;
1508 r10_bio->sectors = sectors;
1509
1510 r10_bio->mddev = mddev;
1511 r10_bio->sector = bio->bi_iter.bi_sector;
1512 r10_bio->state = 0;
1513 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) * conf->copies);
1514
1515 if (bio_data_dir(bio) == READ)
1516 raid10_read_request(mddev, bio, r10_bio);
1517 else
1518 raid10_write_request(mddev, bio, r10_bio);
1519}
1520
1521static bool raid10_make_request(struct mddev *mddev, struct bio *bio)
1522{
1523 struct r10conf *conf = mddev->private;
1524 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1525 int chunk_sects = chunk_mask + 1;
1526 int sectors = bio_sectors(bio);
1527
1528 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1529 md_flush_request(mddev, bio);
1530 return true;
1531 }
1532
1533 if (!md_write_start(mddev, bio))
1534 return false;
1535
1536 /*
1537 * If this request crosses a chunk boundary, we need to split
1538 * it.
1539 */
1540 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1541 sectors > chunk_sects
1542 && (conf->geo.near_copies < conf->geo.raid_disks
1543 || conf->prev.near_copies <
1544 conf->prev.raid_disks)))
1545 sectors = chunk_sects -
1546 (bio->bi_iter.bi_sector &
1547 (chunk_sects - 1));
1548 __make_request(mddev, bio, sectors);
1549
1550 /* In case raid10d snuck in to freeze_array */
1551 wake_up(&conf->wait_barrier);
1552 return true;
1553}
1554
1555static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1556{
1557 struct r10conf *conf = mddev->private;
1558 int i;
1559
1560 if (conf->geo.near_copies < conf->geo.raid_disks)
1561 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1562 if (conf->geo.near_copies > 1)
1563 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1564 if (conf->geo.far_copies > 1) {
1565 if (conf->geo.far_offset)
1566 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1567 else
1568 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1569 if (conf->geo.far_set_size != conf->geo.raid_disks)
1570 seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1571 }
1572 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1573 conf->geo.raid_disks - mddev->degraded);
1574 rcu_read_lock();
1575 for (i = 0; i < conf->geo.raid_disks; i++) {
1576 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1577 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1578 }
1579 rcu_read_unlock();
1580 seq_printf(seq, "]");
1581}
1582
1583/* check if there are enough drives for
1584 * every block to appear on atleast one.
1585 * Don't consider the device numbered 'ignore'
1586 * as we might be about to remove it.
1587 */
1588static int _enough(struct r10conf *conf, int previous, int ignore)
1589{
1590 int first = 0;
1591 int has_enough = 0;
1592 int disks, ncopies;
1593 if (previous) {
1594 disks = conf->prev.raid_disks;
1595 ncopies = conf->prev.near_copies;
1596 } else {
1597 disks = conf->geo.raid_disks;
1598 ncopies = conf->geo.near_copies;
1599 }
1600
1601 rcu_read_lock();
1602 do {
1603 int n = conf->copies;
1604 int cnt = 0;
1605 int this = first;
1606 while (n--) {
1607 struct md_rdev *rdev;
1608 if (this != ignore &&
1609 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1610 test_bit(In_sync, &rdev->flags))
1611 cnt++;
1612 this = (this+1) % disks;
1613 }
1614 if (cnt == 0)
1615 goto out;
1616 first = (first + ncopies) % disks;
1617 } while (first != 0);
1618 has_enough = 1;
1619out:
1620 rcu_read_unlock();
1621 return has_enough;
1622}
1623
1624static int enough(struct r10conf *conf, int ignore)
1625{
1626 /* when calling 'enough', both 'prev' and 'geo' must
1627 * be stable.
1628 * This is ensured if ->reconfig_mutex or ->device_lock
1629 * is held.
1630 */
1631 return _enough(conf, 0, ignore) &&
1632 _enough(conf, 1, ignore);
1633}
1634
1635static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1636{
1637 char b[BDEVNAME_SIZE];
1638 struct r10conf *conf = mddev->private;
1639 unsigned long flags;
1640
1641 /*
1642 * If it is not operational, then we have already marked it as dead
1643 * else if it is the last working disks with "fail_last_dev == false",
1644 * ignore the error, let the next level up know.
1645 * else mark the drive as failed
1646 */
1647 spin_lock_irqsave(&conf->device_lock, flags);
1648 if (test_bit(In_sync, &rdev->flags) && !mddev->fail_last_dev
1649 && !enough(conf, rdev->raid_disk)) {
1650 /*
1651 * Don't fail the drive, just return an IO error.
1652 */
1653 spin_unlock_irqrestore(&conf->device_lock, flags);
1654 return;
1655 }
1656 if (test_and_clear_bit(In_sync, &rdev->flags))
1657 mddev->degraded++;
1658 /*
1659 * If recovery is running, make sure it aborts.
1660 */
1661 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1662 set_bit(Blocked, &rdev->flags);
1663 set_bit(Faulty, &rdev->flags);
1664 set_mask_bits(&mddev->sb_flags, 0,
1665 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1666 spin_unlock_irqrestore(&conf->device_lock, flags);
1667 pr_crit("md/raid10:%s: Disk failure on %s, disabling device.\n"
1668 "md/raid10:%s: Operation continuing on %d devices.\n",
1669 mdname(mddev), bdevname(rdev->bdev, b),
1670 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1671}
1672
1673static void print_conf(struct r10conf *conf)
1674{
1675 int i;
1676 struct md_rdev *rdev;
1677
1678 pr_debug("RAID10 conf printout:\n");
1679 if (!conf) {
1680 pr_debug("(!conf)\n");
1681 return;
1682 }
1683 pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1684 conf->geo.raid_disks);
1685
1686 /* This is only called with ->reconfix_mutex held, so
1687 * rcu protection of rdev is not needed */
1688 for (i = 0; i < conf->geo.raid_disks; i++) {
1689 char b[BDEVNAME_SIZE];
1690 rdev = conf->mirrors[i].rdev;
1691 if (rdev)
1692 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1693 i, !test_bit(In_sync, &rdev->flags),
1694 !test_bit(Faulty, &rdev->flags),
1695 bdevname(rdev->bdev,b));
1696 }
1697}
1698
1699static void close_sync(struct r10conf *conf)
1700{
1701 wait_barrier(conf);
1702 allow_barrier(conf);
1703
1704 mempool_exit(&conf->r10buf_pool);
1705}
1706
1707static int raid10_spare_active(struct mddev *mddev)
1708{
1709 int i;
1710 struct r10conf *conf = mddev->private;
1711 struct raid10_info *tmp;
1712 int count = 0;
1713 unsigned long flags;
1714
1715 /*
1716 * Find all non-in_sync disks within the RAID10 configuration
1717 * and mark them in_sync
1718 */
1719 for (i = 0; i < conf->geo.raid_disks; i++) {
1720 tmp = conf->mirrors + i;
1721 if (tmp->replacement
1722 && tmp->replacement->recovery_offset == MaxSector
1723 && !test_bit(Faulty, &tmp->replacement->flags)
1724 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1725 /* Replacement has just become active */
1726 if (!tmp->rdev
1727 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1728 count++;
1729 if (tmp->rdev) {
1730 /* Replaced device not technically faulty,
1731 * but we need to be sure it gets removed
1732 * and never re-added.
1733 */
1734 set_bit(Faulty, &tmp->rdev->flags);
1735 sysfs_notify_dirent_safe(
1736 tmp->rdev->sysfs_state);
1737 }
1738 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1739 } else if (tmp->rdev
1740 && tmp->rdev->recovery_offset == MaxSector
1741 && !test_bit(Faulty, &tmp->rdev->flags)
1742 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1743 count++;
1744 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1745 }
1746 }
1747 spin_lock_irqsave(&conf->device_lock, flags);
1748 mddev->degraded -= count;
1749 spin_unlock_irqrestore(&conf->device_lock, flags);
1750
1751 print_conf(conf);
1752 return count;
1753}
1754
1755static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1756{
1757 struct r10conf *conf = mddev->private;
1758 int err = -EEXIST;
1759 int mirror;
1760 int first = 0;
1761 int last = conf->geo.raid_disks - 1;
1762
1763 if (mddev->recovery_cp < MaxSector)
1764 /* only hot-add to in-sync arrays, as recovery is
1765 * very different from resync
1766 */
1767 return -EBUSY;
1768 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1769 return -EINVAL;
1770
1771 if (md_integrity_add_rdev(rdev, mddev))
1772 return -ENXIO;
1773
1774 if (rdev->raid_disk >= 0)
1775 first = last = rdev->raid_disk;
1776
1777 if (rdev->saved_raid_disk >= first &&
1778 rdev->saved_raid_disk < conf->geo.raid_disks &&
1779 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1780 mirror = rdev->saved_raid_disk;
1781 else
1782 mirror = first;
1783 for ( ; mirror <= last ; mirror++) {
1784 struct raid10_info *p = &conf->mirrors[mirror];
1785 if (p->recovery_disabled == mddev->recovery_disabled)
1786 continue;
1787 if (p->rdev) {
1788 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1789 p->replacement != NULL)
1790 continue;
1791 clear_bit(In_sync, &rdev->flags);
1792 set_bit(Replacement, &rdev->flags);
1793 rdev->raid_disk = mirror;
1794 err = 0;
1795 if (mddev->gendisk)
1796 disk_stack_limits(mddev->gendisk, rdev->bdev,
1797 rdev->data_offset << 9);
1798 conf->fullsync = 1;
1799 rcu_assign_pointer(p->replacement, rdev);
1800 break;
1801 }
1802
1803 if (mddev->gendisk)
1804 disk_stack_limits(mddev->gendisk, rdev->bdev,
1805 rdev->data_offset << 9);
1806
1807 p->head_position = 0;
1808 p->recovery_disabled = mddev->recovery_disabled - 1;
1809 rdev->raid_disk = mirror;
1810 err = 0;
1811 if (rdev->saved_raid_disk != mirror)
1812 conf->fullsync = 1;
1813 rcu_assign_pointer(p->rdev, rdev);
1814 break;
1815 }
1816 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1817 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1818
1819 print_conf(conf);
1820 return err;
1821}
1822
1823static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1824{
1825 struct r10conf *conf = mddev->private;
1826 int err = 0;
1827 int number = rdev->raid_disk;
1828 struct md_rdev **rdevp;
1829 struct raid10_info *p = conf->mirrors + number;
1830
1831 print_conf(conf);
1832 if (rdev == p->rdev)
1833 rdevp = &p->rdev;
1834 else if (rdev == p->replacement)
1835 rdevp = &p->replacement;
1836 else
1837 return 0;
1838
1839 if (test_bit(In_sync, &rdev->flags) ||
1840 atomic_read(&rdev->nr_pending)) {
1841 err = -EBUSY;
1842 goto abort;
1843 }
1844 /* Only remove non-faulty devices if recovery
1845 * is not possible.
1846 */
1847 if (!test_bit(Faulty, &rdev->flags) &&
1848 mddev->recovery_disabled != p->recovery_disabled &&
1849 (!p->replacement || p->replacement == rdev) &&
1850 number < conf->geo.raid_disks &&
1851 enough(conf, -1)) {
1852 err = -EBUSY;
1853 goto abort;
1854 }
1855 *rdevp = NULL;
1856 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1857 synchronize_rcu();
1858 if (atomic_read(&rdev->nr_pending)) {
1859 /* lost the race, try later */
1860 err = -EBUSY;
1861 *rdevp = rdev;
1862 goto abort;
1863 }
1864 }
1865 if (p->replacement) {
1866 /* We must have just cleared 'rdev' */
1867 p->rdev = p->replacement;
1868 clear_bit(Replacement, &p->replacement->flags);
1869 smp_mb(); /* Make sure other CPUs may see both as identical
1870 * but will never see neither -- if they are careful.
1871 */
1872 p->replacement = NULL;
1873 }
1874
1875 clear_bit(WantReplacement, &rdev->flags);
1876 err = md_integrity_register(mddev);
1877
1878abort:
1879
1880 print_conf(conf);
1881 return err;
1882}
1883
1884static void __end_sync_read(struct r10bio *r10_bio, struct bio *bio, int d)
1885{
1886 struct r10conf *conf = r10_bio->mddev->private;
1887
1888 if (!bio->bi_status)
1889 set_bit(R10BIO_Uptodate, &r10_bio->state);
1890 else
1891 /* The write handler will notice the lack of
1892 * R10BIO_Uptodate and record any errors etc
1893 */
1894 atomic_add(r10_bio->sectors,
1895 &conf->mirrors[d].rdev->corrected_errors);
1896
1897 /* for reconstruct, we always reschedule after a read.
1898 * for resync, only after all reads
1899 */
1900 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1901 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1902 atomic_dec_and_test(&r10_bio->remaining)) {
1903 /* we have read all the blocks,
1904 * do the comparison in process context in raid10d
1905 */
1906 reschedule_retry(r10_bio);
1907 }
1908}
1909
1910static void end_sync_read(struct bio *bio)
1911{
1912 struct r10bio *r10_bio = get_resync_r10bio(bio);
1913 struct r10conf *conf = r10_bio->mddev->private;
1914 int d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1915
1916 __end_sync_read(r10_bio, bio, d);
1917}
1918
1919static void end_reshape_read(struct bio *bio)
1920{
1921 /* reshape read bio isn't allocated from r10buf_pool */
1922 struct r10bio *r10_bio = bio->bi_private;
1923
1924 __end_sync_read(r10_bio, bio, r10_bio->read_slot);
1925}
1926
1927static void end_sync_request(struct r10bio *r10_bio)
1928{
1929 struct mddev *mddev = r10_bio->mddev;
1930
1931 while (atomic_dec_and_test(&r10_bio->remaining)) {
1932 if (r10_bio->master_bio == NULL) {
1933 /* the primary of several recovery bios */
1934 sector_t s = r10_bio->sectors;
1935 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1936 test_bit(R10BIO_WriteError, &r10_bio->state))
1937 reschedule_retry(r10_bio);
1938 else
1939 put_buf(r10_bio);
1940 md_done_sync(mddev, s, 1);
1941 break;
1942 } else {
1943 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1944 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1945 test_bit(R10BIO_WriteError, &r10_bio->state))
1946 reschedule_retry(r10_bio);
1947 else
1948 put_buf(r10_bio);
1949 r10_bio = r10_bio2;
1950 }
1951 }
1952}
1953
1954static void end_sync_write(struct bio *bio)
1955{
1956 struct r10bio *r10_bio = get_resync_r10bio(bio);
1957 struct mddev *mddev = r10_bio->mddev;
1958 struct r10conf *conf = mddev->private;
1959 int d;
1960 sector_t first_bad;
1961 int bad_sectors;
1962 int slot;
1963 int repl;
1964 struct md_rdev *rdev = NULL;
1965
1966 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1967 if (repl)
1968 rdev = conf->mirrors[d].replacement;
1969 else
1970 rdev = conf->mirrors[d].rdev;
1971
1972 if (bio->bi_status) {
1973 if (repl)
1974 md_error(mddev, rdev);
1975 else {
1976 set_bit(WriteErrorSeen, &rdev->flags);
1977 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1978 set_bit(MD_RECOVERY_NEEDED,
1979 &rdev->mddev->recovery);
1980 set_bit(R10BIO_WriteError, &r10_bio->state);
1981 }
1982 } else if (is_badblock(rdev,
1983 r10_bio->devs[slot].addr,
1984 r10_bio->sectors,
1985 &first_bad, &bad_sectors))
1986 set_bit(R10BIO_MadeGood, &r10_bio->state);
1987
1988 rdev_dec_pending(rdev, mddev);
1989
1990 end_sync_request(r10_bio);
1991}
1992
1993/*
1994 * Note: sync and recover and handled very differently for raid10
1995 * This code is for resync.
1996 * For resync, we read through virtual addresses and read all blocks.
1997 * If there is any error, we schedule a write. The lowest numbered
1998 * drive is authoritative.
1999 * However requests come for physical address, so we need to map.
2000 * For every physical address there are raid_disks/copies virtual addresses,
2001 * which is always are least one, but is not necessarly an integer.
2002 * This means that a physical address can span multiple chunks, so we may
2003 * have to submit multiple io requests for a single sync request.
2004 */
2005/*
2006 * We check if all blocks are in-sync and only write to blocks that
2007 * aren't in sync
2008 */
2009static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2010{
2011 struct r10conf *conf = mddev->private;
2012 int i, first;
2013 struct bio *tbio, *fbio;
2014 int vcnt;
2015 struct page **tpages, **fpages;
2016
2017 atomic_set(&r10_bio->remaining, 1);
2018
2019 /* find the first device with a block */
2020 for (i=0; i<conf->copies; i++)
2021 if (!r10_bio->devs[i].bio->bi_status)
2022 break;
2023
2024 if (i == conf->copies)
2025 goto done;
2026
2027 first = i;
2028 fbio = r10_bio->devs[i].bio;
2029 fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2030 fbio->bi_iter.bi_idx = 0;
2031 fpages = get_resync_pages(fbio)->pages;
2032
2033 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2034 /* now find blocks with errors */
2035 for (i=0 ; i < conf->copies ; i++) {
2036 int j, d;
2037 struct md_rdev *rdev;
2038 struct resync_pages *rp;
2039
2040 tbio = r10_bio->devs[i].bio;
2041
2042 if (tbio->bi_end_io != end_sync_read)
2043 continue;
2044 if (i == first)
2045 continue;
2046
2047 tpages = get_resync_pages(tbio)->pages;
2048 d = r10_bio->devs[i].devnum;
2049 rdev = conf->mirrors[d].rdev;
2050 if (!r10_bio->devs[i].bio->bi_status) {
2051 /* We know that the bi_io_vec layout is the same for
2052 * both 'first' and 'i', so we just compare them.
2053 * All vec entries are PAGE_SIZE;
2054 */
2055 int sectors = r10_bio->sectors;
2056 for (j = 0; j < vcnt; j++) {
2057 int len = PAGE_SIZE;
2058 if (sectors < (len / 512))
2059 len = sectors * 512;
2060 if (memcmp(page_address(fpages[j]),
2061 page_address(tpages[j]),
2062 len))
2063 break;
2064 sectors -= len/512;
2065 }
2066 if (j == vcnt)
2067 continue;
2068 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2069 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2070 /* Don't fix anything. */
2071 continue;
2072 } else if (test_bit(FailFast, &rdev->flags)) {
2073 /* Just give up on this device */
2074 md_error(rdev->mddev, rdev);
2075 continue;
2076 }
2077 /* Ok, we need to write this bio, either to correct an
2078 * inconsistency or to correct an unreadable block.
2079 * First we need to fixup bv_offset, bv_len and
2080 * bi_vecs, as the read request might have corrupted these
2081 */
2082 rp = get_resync_pages(tbio);
2083 bio_reset(tbio);
2084
2085 md_bio_reset_resync_pages(tbio, rp, fbio->bi_iter.bi_size);
2086
2087 rp->raid_bio = r10_bio;
2088 tbio->bi_private = rp;
2089 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2090 tbio->bi_end_io = end_sync_write;
2091 bio_set_op_attrs(tbio, REQ_OP_WRITE, 0);
2092
2093 bio_copy_data(tbio, fbio);
2094
2095 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2096 atomic_inc(&r10_bio->remaining);
2097 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2098
2099 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2100 tbio->bi_opf |= MD_FAILFAST;
2101 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2102 bio_set_dev(tbio, conf->mirrors[d].rdev->bdev);
2103 generic_make_request(tbio);
2104 }
2105
2106 /* Now write out to any replacement devices
2107 * that are active
2108 */
2109 for (i = 0; i < conf->copies; i++) {
2110 int d;
2111
2112 tbio = r10_bio->devs[i].repl_bio;
2113 if (!tbio || !tbio->bi_end_io)
2114 continue;
2115 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2116 && r10_bio->devs[i].bio != fbio)
2117 bio_copy_data(tbio, fbio);
2118 d = r10_bio->devs[i].devnum;
2119 atomic_inc(&r10_bio->remaining);
2120 md_sync_acct(conf->mirrors[d].replacement->bdev,
2121 bio_sectors(tbio));
2122 generic_make_request(tbio);
2123 }
2124
2125done:
2126 if (atomic_dec_and_test(&r10_bio->remaining)) {
2127 md_done_sync(mddev, r10_bio->sectors, 1);
2128 put_buf(r10_bio);
2129 }
2130}
2131
2132/*
2133 * Now for the recovery code.
2134 * Recovery happens across physical sectors.
2135 * We recover all non-is_sync drives by finding the virtual address of
2136 * each, and then choose a working drive that also has that virt address.
2137 * There is a separate r10_bio for each non-in_sync drive.
2138 * Only the first two slots are in use. The first for reading,
2139 * The second for writing.
2140 *
2141 */
2142static void fix_recovery_read_error(struct r10bio *r10_bio)
2143{
2144 /* We got a read error during recovery.
2145 * We repeat the read in smaller page-sized sections.
2146 * If a read succeeds, write it to the new device or record
2147 * a bad block if we cannot.
2148 * If a read fails, record a bad block on both old and
2149 * new devices.
2150 */
2151 struct mddev *mddev = r10_bio->mddev;
2152 struct r10conf *conf = mddev->private;
2153 struct bio *bio = r10_bio->devs[0].bio;
2154 sector_t sect = 0;
2155 int sectors = r10_bio->sectors;
2156 int idx = 0;
2157 int dr = r10_bio->devs[0].devnum;
2158 int dw = r10_bio->devs[1].devnum;
2159 struct page **pages = get_resync_pages(bio)->pages;
2160
2161 while (sectors) {
2162 int s = sectors;
2163 struct md_rdev *rdev;
2164 sector_t addr;
2165 int ok;
2166
2167 if (s > (PAGE_SIZE>>9))
2168 s = PAGE_SIZE >> 9;
2169
2170 rdev = conf->mirrors[dr].rdev;
2171 addr = r10_bio->devs[0].addr + sect,
2172 ok = sync_page_io(rdev,
2173 addr,
2174 s << 9,
2175 pages[idx],
2176 REQ_OP_READ, 0, false);
2177 if (ok) {
2178 rdev = conf->mirrors[dw].rdev;
2179 addr = r10_bio->devs[1].addr + sect;
2180 ok = sync_page_io(rdev,
2181 addr,
2182 s << 9,
2183 pages[idx],
2184 REQ_OP_WRITE, 0, false);
2185 if (!ok) {
2186 set_bit(WriteErrorSeen, &rdev->flags);
2187 if (!test_and_set_bit(WantReplacement,
2188 &rdev->flags))
2189 set_bit(MD_RECOVERY_NEEDED,
2190 &rdev->mddev->recovery);
2191 }
2192 }
2193 if (!ok) {
2194 /* We don't worry if we cannot set a bad block -
2195 * it really is bad so there is no loss in not
2196 * recording it yet
2197 */
2198 rdev_set_badblocks(rdev, addr, s, 0);
2199
2200 if (rdev != conf->mirrors[dw].rdev) {
2201 /* need bad block on destination too */
2202 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2203 addr = r10_bio->devs[1].addr + sect;
2204 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2205 if (!ok) {
2206 /* just abort the recovery */
2207 pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2208 mdname(mddev));
2209
2210 conf->mirrors[dw].recovery_disabled
2211 = mddev->recovery_disabled;
2212 set_bit(MD_RECOVERY_INTR,
2213 &mddev->recovery);
2214 break;
2215 }
2216 }
2217 }
2218
2219 sectors -= s;
2220 sect += s;
2221 idx++;
2222 }
2223}
2224
2225static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2226{
2227 struct r10conf *conf = mddev->private;
2228 int d;
2229 struct bio *wbio, *wbio2;
2230
2231 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2232 fix_recovery_read_error(r10_bio);
2233 end_sync_request(r10_bio);
2234 return;
2235 }
2236
2237 /*
2238 * share the pages with the first bio
2239 * and submit the write request
2240 */
2241 d = r10_bio->devs[1].devnum;
2242 wbio = r10_bio->devs[1].bio;
2243 wbio2 = r10_bio->devs[1].repl_bio;
2244 /* Need to test wbio2->bi_end_io before we call
2245 * generic_make_request as if the former is NULL,
2246 * the latter is free to free wbio2.
2247 */
2248 if (wbio2 && !wbio2->bi_end_io)
2249 wbio2 = NULL;
2250 if (wbio->bi_end_io) {
2251 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2252 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2253 generic_make_request(wbio);
2254 }
2255 if (wbio2) {
2256 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2257 md_sync_acct(conf->mirrors[d].replacement->bdev,
2258 bio_sectors(wbio2));
2259 generic_make_request(wbio2);
2260 }
2261}
2262
2263/*
2264 * Used by fix_read_error() to decay the per rdev read_errors.
2265 * We halve the read error count for every hour that has elapsed
2266 * since the last recorded read error.
2267 *
2268 */
2269static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2270{
2271 long cur_time_mon;
2272 unsigned long hours_since_last;
2273 unsigned int read_errors = atomic_read(&rdev->read_errors);
2274
2275 cur_time_mon = ktime_get_seconds();
2276
2277 if (rdev->last_read_error == 0) {
2278 /* first time we've seen a read error */
2279 rdev->last_read_error = cur_time_mon;
2280 return;
2281 }
2282
2283 hours_since_last = (long)(cur_time_mon -
2284 rdev->last_read_error) / 3600;
2285
2286 rdev->last_read_error = cur_time_mon;
2287
2288 /*
2289 * if hours_since_last is > the number of bits in read_errors
2290 * just set read errors to 0. We do this to avoid
2291 * overflowing the shift of read_errors by hours_since_last.
2292 */
2293 if (hours_since_last >= 8 * sizeof(read_errors))
2294 atomic_set(&rdev->read_errors, 0);
2295 else
2296 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2297}
2298
2299static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2300 int sectors, struct page *page, int rw)
2301{
2302 sector_t first_bad;
2303 int bad_sectors;
2304
2305 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2306 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2307 return -1;
2308 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
2309 /* success */
2310 return 1;
2311 if (rw == WRITE) {
2312 set_bit(WriteErrorSeen, &rdev->flags);
2313 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2314 set_bit(MD_RECOVERY_NEEDED,
2315 &rdev->mddev->recovery);
2316 }
2317 /* need to record an error - either for the block or the device */
2318 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2319 md_error(rdev->mddev, rdev);
2320 return 0;
2321}
2322
2323/*
2324 * This is a kernel thread which:
2325 *
2326 * 1. Retries failed read operations on working mirrors.
2327 * 2. Updates the raid superblock when problems encounter.
2328 * 3. Performs writes following reads for array synchronising.
2329 */
2330
2331static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2332{
2333 int sect = 0; /* Offset from r10_bio->sector */
2334 int sectors = r10_bio->sectors;
2335 struct md_rdev *rdev;
2336 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2337 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2338
2339 /* still own a reference to this rdev, so it cannot
2340 * have been cleared recently.
2341 */
2342 rdev = conf->mirrors[d].rdev;
2343
2344 if (test_bit(Faulty, &rdev->flags))
2345 /* drive has already been failed, just ignore any
2346 more fix_read_error() attempts */
2347 return;
2348
2349 check_decay_read_errors(mddev, rdev);
2350 atomic_inc(&rdev->read_errors);
2351 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2352 char b[BDEVNAME_SIZE];
2353 bdevname(rdev->bdev, b);
2354
2355 pr_notice("md/raid10:%s: %s: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2356 mdname(mddev), b,
2357 atomic_read(&rdev->read_errors), max_read_errors);
2358 pr_notice("md/raid10:%s: %s: Failing raid device\n",
2359 mdname(mddev), b);
2360 md_error(mddev, rdev);
2361 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2362 return;
2363 }
2364
2365 while(sectors) {
2366 int s = sectors;
2367 int sl = r10_bio->read_slot;
2368 int success = 0;
2369 int start;
2370
2371 if (s > (PAGE_SIZE>>9))
2372 s = PAGE_SIZE >> 9;
2373
2374 rcu_read_lock();
2375 do {
2376 sector_t first_bad;
2377 int bad_sectors;
2378
2379 d = r10_bio->devs[sl].devnum;
2380 rdev = rcu_dereference(conf->mirrors[d].rdev);
2381 if (rdev &&
2382 test_bit(In_sync, &rdev->flags) &&
2383 !test_bit(Faulty, &rdev->flags) &&
2384 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2385 &first_bad, &bad_sectors) == 0) {
2386 atomic_inc(&rdev->nr_pending);
2387 rcu_read_unlock();
2388 success = sync_page_io(rdev,
2389 r10_bio->devs[sl].addr +
2390 sect,
2391 s<<9,
2392 conf->tmppage,
2393 REQ_OP_READ, 0, false);
2394 rdev_dec_pending(rdev, mddev);
2395 rcu_read_lock();
2396 if (success)
2397 break;
2398 }
2399 sl++;
2400 if (sl == conf->copies)
2401 sl = 0;
2402 } while (!success && sl != r10_bio->read_slot);
2403 rcu_read_unlock();
2404
2405 if (!success) {
2406 /* Cannot read from anywhere, just mark the block
2407 * as bad on the first device to discourage future
2408 * reads.
2409 */
2410 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2411 rdev = conf->mirrors[dn].rdev;
2412
2413 if (!rdev_set_badblocks(
2414 rdev,
2415 r10_bio->devs[r10_bio->read_slot].addr
2416 + sect,
2417 s, 0)) {
2418 md_error(mddev, rdev);
2419 r10_bio->devs[r10_bio->read_slot].bio
2420 = IO_BLOCKED;
2421 }
2422 break;
2423 }
2424
2425 start = sl;
2426 /* write it back and re-read */
2427 rcu_read_lock();
2428 while (sl != r10_bio->read_slot) {
2429 char b[BDEVNAME_SIZE];
2430
2431 if (sl==0)
2432 sl = conf->copies;
2433 sl--;
2434 d = r10_bio->devs[sl].devnum;
2435 rdev = rcu_dereference(conf->mirrors[d].rdev);
2436 if (!rdev ||
2437 test_bit(Faulty, &rdev->flags) ||
2438 !test_bit(In_sync, &rdev->flags))
2439 continue;
2440
2441 atomic_inc(&rdev->nr_pending);
2442 rcu_read_unlock();
2443 if (r10_sync_page_io(rdev,
2444 r10_bio->devs[sl].addr +
2445 sect,
2446 s, conf->tmppage, WRITE)
2447 == 0) {
2448 /* Well, this device is dead */
2449 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %s)\n",
2450 mdname(mddev), s,
2451 (unsigned long long)(
2452 sect +
2453 choose_data_offset(r10_bio,
2454 rdev)),
2455 bdevname(rdev->bdev, b));
2456 pr_notice("md/raid10:%s: %s: failing drive\n",
2457 mdname(mddev),
2458 bdevname(rdev->bdev, b));
2459 }
2460 rdev_dec_pending(rdev, mddev);
2461 rcu_read_lock();
2462 }
2463 sl = start;
2464 while (sl != r10_bio->read_slot) {
2465 char b[BDEVNAME_SIZE];
2466
2467 if (sl==0)
2468 sl = conf->copies;
2469 sl--;
2470 d = r10_bio->devs[sl].devnum;
2471 rdev = rcu_dereference(conf->mirrors[d].rdev);
2472 if (!rdev ||
2473 test_bit(Faulty, &rdev->flags) ||
2474 !test_bit(In_sync, &rdev->flags))
2475 continue;
2476
2477 atomic_inc(&rdev->nr_pending);
2478 rcu_read_unlock();
2479 switch (r10_sync_page_io(rdev,
2480 r10_bio->devs[sl].addr +
2481 sect,
2482 s, conf->tmppage,
2483 READ)) {
2484 case 0:
2485 /* Well, this device is dead */
2486 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %s)\n",
2487 mdname(mddev), s,
2488 (unsigned long long)(
2489 sect +
2490 choose_data_offset(r10_bio, rdev)),
2491 bdevname(rdev->bdev, b));
2492 pr_notice("md/raid10:%s: %s: failing drive\n",
2493 mdname(mddev),
2494 bdevname(rdev->bdev, b));
2495 break;
2496 case 1:
2497 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %s)\n",
2498 mdname(mddev), s,
2499 (unsigned long long)(
2500 sect +
2501 choose_data_offset(r10_bio, rdev)),
2502 bdevname(rdev->bdev, b));
2503 atomic_add(s, &rdev->corrected_errors);
2504 }
2505
2506 rdev_dec_pending(rdev, mddev);
2507 rcu_read_lock();
2508 }
2509 rcu_read_unlock();
2510
2511 sectors -= s;
2512 sect += s;
2513 }
2514}
2515
2516static int narrow_write_error(struct r10bio *r10_bio, int i)
2517{
2518 struct bio *bio = r10_bio->master_bio;
2519 struct mddev *mddev = r10_bio->mddev;
2520 struct r10conf *conf = mddev->private;
2521 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2522 /* bio has the data to be written to slot 'i' where
2523 * we just recently had a write error.
2524 * We repeatedly clone the bio and trim down to one block,
2525 * then try the write. Where the write fails we record
2526 * a bad block.
2527 * It is conceivable that the bio doesn't exactly align with
2528 * blocks. We must handle this.
2529 *
2530 * We currently own a reference to the rdev.
2531 */
2532
2533 int block_sectors;
2534 sector_t sector;
2535 int sectors;
2536 int sect_to_write = r10_bio->sectors;
2537 int ok = 1;
2538
2539 if (rdev->badblocks.shift < 0)
2540 return 0;
2541
2542 block_sectors = roundup(1 << rdev->badblocks.shift,
2543 bdev_logical_block_size(rdev->bdev) >> 9);
2544 sector = r10_bio->sector;
2545 sectors = ((r10_bio->sector + block_sectors)
2546 & ~(sector_t)(block_sectors - 1))
2547 - sector;
2548
2549 while (sect_to_write) {
2550 struct bio *wbio;
2551 sector_t wsector;
2552 if (sectors > sect_to_write)
2553 sectors = sect_to_write;
2554 /* Write at 'sector' for 'sectors' */
2555 wbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
2556 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2557 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2558 wbio->bi_iter.bi_sector = wsector +
2559 choose_data_offset(r10_bio, rdev);
2560 bio_set_dev(wbio, rdev->bdev);
2561 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2562
2563 if (submit_bio_wait(wbio) < 0)
2564 /* Failure! */
2565 ok = rdev_set_badblocks(rdev, wsector,
2566 sectors, 0)
2567 && ok;
2568
2569 bio_put(wbio);
2570 sect_to_write -= sectors;
2571 sector += sectors;
2572 sectors = block_sectors;
2573 }
2574 return ok;
2575}
2576
2577static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2578{
2579 int slot = r10_bio->read_slot;
2580 struct bio *bio;
2581 struct r10conf *conf = mddev->private;
2582 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2583
2584 /* we got a read error. Maybe the drive is bad. Maybe just
2585 * the block and we can fix it.
2586 * We freeze all other IO, and try reading the block from
2587 * other devices. When we find one, we re-write
2588 * and check it that fixes the read error.
2589 * This is all done synchronously while the array is
2590 * frozen.
2591 */
2592 bio = r10_bio->devs[slot].bio;
2593 bio_put(bio);
2594 r10_bio->devs[slot].bio = NULL;
2595
2596 if (mddev->ro)
2597 r10_bio->devs[slot].bio = IO_BLOCKED;
2598 else if (!test_bit(FailFast, &rdev->flags)) {
2599 freeze_array(conf, 1);
2600 fix_read_error(conf, mddev, r10_bio);
2601 unfreeze_array(conf);
2602 } else
2603 md_error(mddev, rdev);
2604
2605 rdev_dec_pending(rdev, mddev);
2606 allow_barrier(conf);
2607 r10_bio->state = 0;
2608 raid10_read_request(mddev, r10_bio->master_bio, r10_bio);
2609}
2610
2611static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2612{
2613 /* Some sort of write request has finished and it
2614 * succeeded in writing where we thought there was a
2615 * bad block. So forget the bad block.
2616 * Or possibly if failed and we need to record
2617 * a bad block.
2618 */
2619 int m;
2620 struct md_rdev *rdev;
2621
2622 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2623 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2624 for (m = 0; m < conf->copies; m++) {
2625 int dev = r10_bio->devs[m].devnum;
2626 rdev = conf->mirrors[dev].rdev;
2627 if (r10_bio->devs[m].bio == NULL ||
2628 r10_bio->devs[m].bio->bi_end_io == NULL)
2629 continue;
2630 if (!r10_bio->devs[m].bio->bi_status) {
2631 rdev_clear_badblocks(
2632 rdev,
2633 r10_bio->devs[m].addr,
2634 r10_bio->sectors, 0);
2635 } else {
2636 if (!rdev_set_badblocks(
2637 rdev,
2638 r10_bio->devs[m].addr,
2639 r10_bio->sectors, 0))
2640 md_error(conf->mddev, rdev);
2641 }
2642 rdev = conf->mirrors[dev].replacement;
2643 if (r10_bio->devs[m].repl_bio == NULL ||
2644 r10_bio->devs[m].repl_bio->bi_end_io == NULL)
2645 continue;
2646
2647 if (!r10_bio->devs[m].repl_bio->bi_status) {
2648 rdev_clear_badblocks(
2649 rdev,
2650 r10_bio->devs[m].addr,
2651 r10_bio->sectors, 0);
2652 } else {
2653 if (!rdev_set_badblocks(
2654 rdev,
2655 r10_bio->devs[m].addr,
2656 r10_bio->sectors, 0))
2657 md_error(conf->mddev, rdev);
2658 }
2659 }
2660 put_buf(r10_bio);
2661 } else {
2662 bool fail = false;
2663 for (m = 0; m < conf->copies; m++) {
2664 int dev = r10_bio->devs[m].devnum;
2665 struct bio *bio = r10_bio->devs[m].bio;
2666 rdev = conf->mirrors[dev].rdev;
2667 if (bio == IO_MADE_GOOD) {
2668 rdev_clear_badblocks(
2669 rdev,
2670 r10_bio->devs[m].addr,
2671 r10_bio->sectors, 0);
2672 rdev_dec_pending(rdev, conf->mddev);
2673 } else if (bio != NULL && bio->bi_status) {
2674 fail = true;
2675 if (!narrow_write_error(r10_bio, m)) {
2676 md_error(conf->mddev, rdev);
2677 set_bit(R10BIO_Degraded,
2678 &r10_bio->state);
2679 }
2680 rdev_dec_pending(rdev, conf->mddev);
2681 }
2682 bio = r10_bio->devs[m].repl_bio;
2683 rdev = conf->mirrors[dev].replacement;
2684 if (rdev && bio == IO_MADE_GOOD) {
2685 rdev_clear_badblocks(
2686 rdev,
2687 r10_bio->devs[m].addr,
2688 r10_bio->sectors, 0);
2689 rdev_dec_pending(rdev, conf->mddev);
2690 }
2691 }
2692 if (fail) {
2693 spin_lock_irq(&conf->device_lock);
2694 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2695 conf->nr_queued++;
2696 spin_unlock_irq(&conf->device_lock);
2697 /*
2698 * In case freeze_array() is waiting for condition
2699 * nr_pending == nr_queued + extra to be true.
2700 */
2701 wake_up(&conf->wait_barrier);
2702 md_wakeup_thread(conf->mddev->thread);
2703 } else {
2704 if (test_bit(R10BIO_WriteError,
2705 &r10_bio->state))
2706 close_write(r10_bio);
2707 raid_end_bio_io(r10_bio);
2708 }
2709 }
2710}
2711
2712static void raid10d(struct md_thread *thread)
2713{
2714 struct mddev *mddev = thread->mddev;
2715 struct r10bio *r10_bio;
2716 unsigned long flags;
2717 struct r10conf *conf = mddev->private;
2718 struct list_head *head = &conf->retry_list;
2719 struct blk_plug plug;
2720
2721 md_check_recovery(mddev);
2722
2723 if (!list_empty_careful(&conf->bio_end_io_list) &&
2724 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2725 LIST_HEAD(tmp);
2726 spin_lock_irqsave(&conf->device_lock, flags);
2727 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2728 while (!list_empty(&conf->bio_end_io_list)) {
2729 list_move(conf->bio_end_io_list.prev, &tmp);
2730 conf->nr_queued--;
2731 }
2732 }
2733 spin_unlock_irqrestore(&conf->device_lock, flags);
2734 while (!list_empty(&tmp)) {
2735 r10_bio = list_first_entry(&tmp, struct r10bio,
2736 retry_list);
2737 list_del(&r10_bio->retry_list);
2738 if (mddev->degraded)
2739 set_bit(R10BIO_Degraded, &r10_bio->state);
2740
2741 if (test_bit(R10BIO_WriteError,
2742 &r10_bio->state))
2743 close_write(r10_bio);
2744 raid_end_bio_io(r10_bio);
2745 }
2746 }
2747
2748 blk_start_plug(&plug);
2749 for (;;) {
2750
2751 flush_pending_writes(conf);
2752
2753 spin_lock_irqsave(&conf->device_lock, flags);
2754 if (list_empty(head)) {
2755 spin_unlock_irqrestore(&conf->device_lock, flags);
2756 break;
2757 }
2758 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2759 list_del(head->prev);
2760 conf->nr_queued--;
2761 spin_unlock_irqrestore(&conf->device_lock, flags);
2762
2763 mddev = r10_bio->mddev;
2764 conf = mddev->private;
2765 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2766 test_bit(R10BIO_WriteError, &r10_bio->state))
2767 handle_write_completed(conf, r10_bio);
2768 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2769 reshape_request_write(mddev, r10_bio);
2770 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2771 sync_request_write(mddev, r10_bio);
2772 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2773 recovery_request_write(mddev, r10_bio);
2774 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2775 handle_read_error(mddev, r10_bio);
2776 else
2777 WARN_ON_ONCE(1);
2778
2779 cond_resched();
2780 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2781 md_check_recovery(mddev);
2782 }
2783 blk_finish_plug(&plug);
2784}
2785
2786static int init_resync(struct r10conf *conf)
2787{
2788 int ret, buffs, i;
2789
2790 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2791 BUG_ON(mempool_initialized(&conf->r10buf_pool));
2792 conf->have_replacement = 0;
2793 for (i = 0; i < conf->geo.raid_disks; i++)
2794 if (conf->mirrors[i].replacement)
2795 conf->have_replacement = 1;
2796 ret = mempool_init(&conf->r10buf_pool, buffs,
2797 r10buf_pool_alloc, r10buf_pool_free, conf);
2798 if (ret)
2799 return ret;
2800 conf->next_resync = 0;
2801 return 0;
2802}
2803
2804static struct r10bio *raid10_alloc_init_r10buf(struct r10conf *conf)
2805{
2806 struct r10bio *r10bio = mempool_alloc(&conf->r10buf_pool, GFP_NOIO);
2807 struct rsync_pages *rp;
2808 struct bio *bio;
2809 int nalloc;
2810 int i;
2811
2812 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
2813 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
2814 nalloc = conf->copies; /* resync */
2815 else
2816 nalloc = 2; /* recovery */
2817
2818 for (i = 0; i < nalloc; i++) {
2819 bio = r10bio->devs[i].bio;
2820 rp = bio->bi_private;
2821 bio_reset(bio);
2822 bio->bi_private = rp;
2823 bio = r10bio->devs[i].repl_bio;
2824 if (bio) {
2825 rp = bio->bi_private;
2826 bio_reset(bio);
2827 bio->bi_private = rp;
2828 }
2829 }
2830 return r10bio;
2831}
2832
2833/*
2834 * Set cluster_sync_high since we need other nodes to add the
2835 * range [cluster_sync_low, cluster_sync_high] to suspend list.
2836 */
2837static void raid10_set_cluster_sync_high(struct r10conf *conf)
2838{
2839 sector_t window_size;
2840 int extra_chunk, chunks;
2841
2842 /*
2843 * First, here we define "stripe" as a unit which across
2844 * all member devices one time, so we get chunks by use
2845 * raid_disks / near_copies. Otherwise, if near_copies is
2846 * close to raid_disks, then resync window could increases
2847 * linearly with the increase of raid_disks, which means
2848 * we will suspend a really large IO window while it is not
2849 * necessary. If raid_disks is not divisible by near_copies,
2850 * an extra chunk is needed to ensure the whole "stripe" is
2851 * covered.
2852 */
2853
2854 chunks = conf->geo.raid_disks / conf->geo.near_copies;
2855 if (conf->geo.raid_disks % conf->geo.near_copies == 0)
2856 extra_chunk = 0;
2857 else
2858 extra_chunk = 1;
2859 window_size = (chunks + extra_chunk) * conf->mddev->chunk_sectors;
2860
2861 /*
2862 * At least use a 32M window to align with raid1's resync window
2863 */
2864 window_size = (CLUSTER_RESYNC_WINDOW_SECTORS > window_size) ?
2865 CLUSTER_RESYNC_WINDOW_SECTORS : window_size;
2866
2867 conf->cluster_sync_high = conf->cluster_sync_low + window_size;
2868}
2869
2870/*
2871 * perform a "sync" on one "block"
2872 *
2873 * We need to make sure that no normal I/O request - particularly write
2874 * requests - conflict with active sync requests.
2875 *
2876 * This is achieved by tracking pending requests and a 'barrier' concept
2877 * that can be installed to exclude normal IO requests.
2878 *
2879 * Resync and recovery are handled very differently.
2880 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2881 *
2882 * For resync, we iterate over virtual addresses, read all copies,
2883 * and update if there are differences. If only one copy is live,
2884 * skip it.
2885 * For recovery, we iterate over physical addresses, read a good
2886 * value for each non-in_sync drive, and over-write.
2887 *
2888 * So, for recovery we may have several outstanding complex requests for a
2889 * given address, one for each out-of-sync device. We model this by allocating
2890 * a number of r10_bio structures, one for each out-of-sync device.
2891 * As we setup these structures, we collect all bio's together into a list
2892 * which we then process collectively to add pages, and then process again
2893 * to pass to generic_make_request.
2894 *
2895 * The r10_bio structures are linked using a borrowed master_bio pointer.
2896 * This link is counted in ->remaining. When the r10_bio that points to NULL
2897 * has its remaining count decremented to 0, the whole complex operation
2898 * is complete.
2899 *
2900 */
2901
2902static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2903 int *skipped)
2904{
2905 struct r10conf *conf = mddev->private;
2906 struct r10bio *r10_bio;
2907 struct bio *biolist = NULL, *bio;
2908 sector_t max_sector, nr_sectors;
2909 int i;
2910 int max_sync;
2911 sector_t sync_blocks;
2912 sector_t sectors_skipped = 0;
2913 int chunks_skipped = 0;
2914 sector_t chunk_mask = conf->geo.chunk_mask;
2915 int page_idx = 0;
2916
2917 if (!mempool_initialized(&conf->r10buf_pool))
2918 if (init_resync(conf))
2919 return 0;
2920
2921 /*
2922 * Allow skipping a full rebuild for incremental assembly
2923 * of a clean array, like RAID1 does.
2924 */
2925 if (mddev->bitmap == NULL &&
2926 mddev->recovery_cp == MaxSector &&
2927 mddev->reshape_position == MaxSector &&
2928 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2929 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2930 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2931 conf->fullsync == 0) {
2932 *skipped = 1;
2933 return mddev->dev_sectors - sector_nr;
2934 }
2935
2936 skipped:
2937 max_sector = mddev->dev_sectors;
2938 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2939 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2940 max_sector = mddev->resync_max_sectors;
2941 if (sector_nr >= max_sector) {
2942 conf->cluster_sync_low = 0;
2943 conf->cluster_sync_high = 0;
2944
2945 /* If we aborted, we need to abort the
2946 * sync on the 'current' bitmap chucks (there can
2947 * be several when recovering multiple devices).
2948 * as we may have started syncing it but not finished.
2949 * We can find the current address in
2950 * mddev->curr_resync, but for recovery,
2951 * we need to convert that to several
2952 * virtual addresses.
2953 */
2954 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2955 end_reshape(conf);
2956 close_sync(conf);
2957 return 0;
2958 }
2959
2960 if (mddev->curr_resync < max_sector) { /* aborted */
2961 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2962 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2963 &sync_blocks, 1);
2964 else for (i = 0; i < conf->geo.raid_disks; i++) {
2965 sector_t sect =
2966 raid10_find_virt(conf, mddev->curr_resync, i);
2967 md_bitmap_end_sync(mddev->bitmap, sect,
2968 &sync_blocks, 1);
2969 }
2970 } else {
2971 /* completed sync */
2972 if ((!mddev->bitmap || conf->fullsync)
2973 && conf->have_replacement
2974 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2975 /* Completed a full sync so the replacements
2976 * are now fully recovered.
2977 */
2978 rcu_read_lock();
2979 for (i = 0; i < conf->geo.raid_disks; i++) {
2980 struct md_rdev *rdev =
2981 rcu_dereference(conf->mirrors[i].replacement);
2982 if (rdev)
2983 rdev->recovery_offset = MaxSector;
2984 }
2985 rcu_read_unlock();
2986 }
2987 conf->fullsync = 0;
2988 }
2989 md_bitmap_close_sync(mddev->bitmap);
2990 close_sync(conf);
2991 *skipped = 1;
2992 return sectors_skipped;
2993 }
2994
2995 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2996 return reshape_request(mddev, sector_nr, skipped);
2997
2998 if (chunks_skipped >= conf->geo.raid_disks) {
2999 /* if there has been nothing to do on any drive,
3000 * then there is nothing to do at all..
3001 */
3002 *skipped = 1;
3003 return (max_sector - sector_nr) + sectors_skipped;
3004 }
3005
3006 if (max_sector > mddev->resync_max)
3007 max_sector = mddev->resync_max; /* Don't do IO beyond here */
3008
3009 /* make sure whole request will fit in a chunk - if chunks
3010 * are meaningful
3011 */
3012 if (conf->geo.near_copies < conf->geo.raid_disks &&
3013 max_sector > (sector_nr | chunk_mask))
3014 max_sector = (sector_nr | chunk_mask) + 1;
3015
3016 /*
3017 * If there is non-resync activity waiting for a turn, then let it
3018 * though before starting on this new sync request.
3019 */
3020 if (conf->nr_waiting)
3021 schedule_timeout_uninterruptible(1);
3022
3023 /* Again, very different code for resync and recovery.
3024 * Both must result in an r10bio with a list of bios that
3025 * have bi_end_io, bi_sector, bi_disk set,
3026 * and bi_private set to the r10bio.
3027 * For recovery, we may actually create several r10bios
3028 * with 2 bios in each, that correspond to the bios in the main one.
3029 * In this case, the subordinate r10bios link back through a
3030 * borrowed master_bio pointer, and the counter in the master
3031 * includes a ref from each subordinate.
3032 */
3033 /* First, we decide what to do and set ->bi_end_io
3034 * To end_sync_read if we want to read, and
3035 * end_sync_write if we will want to write.
3036 */
3037
3038 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3039 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3040 /* recovery... the complicated one */
3041 int j;
3042 r10_bio = NULL;
3043
3044 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3045 int still_degraded;
3046 struct r10bio *rb2;
3047 sector_t sect;
3048 int must_sync;
3049 int any_working;
3050 int need_recover = 0;
3051 int need_replace = 0;
3052 struct raid10_info *mirror = &conf->mirrors[i];
3053 struct md_rdev *mrdev, *mreplace;
3054
3055 rcu_read_lock();
3056 mrdev = rcu_dereference(mirror->rdev);
3057 mreplace = rcu_dereference(mirror->replacement);
3058
3059 if (mrdev != NULL &&
3060 !test_bit(Faulty, &mrdev->flags) &&
3061 !test_bit(In_sync, &mrdev->flags))
3062 need_recover = 1;
3063 if (mreplace != NULL &&
3064 !test_bit(Faulty, &mreplace->flags))
3065 need_replace = 1;
3066
3067 if (!need_recover && !need_replace) {
3068 rcu_read_unlock();
3069 continue;
3070 }
3071
3072 still_degraded = 0;
3073 /* want to reconstruct this device */
3074 rb2 = r10_bio;
3075 sect = raid10_find_virt(conf, sector_nr, i);
3076 if (sect >= mddev->resync_max_sectors) {
3077 /* last stripe is not complete - don't
3078 * try to recover this sector.
3079 */
3080 rcu_read_unlock();
3081 continue;
3082 }
3083 if (mreplace && test_bit(Faulty, &mreplace->flags))
3084 mreplace = NULL;
3085 /* Unless we are doing a full sync, or a replacement
3086 * we only need to recover the block if it is set in
3087 * the bitmap
3088 */
3089 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3090 &sync_blocks, 1);
3091 if (sync_blocks < max_sync)
3092 max_sync = sync_blocks;
3093 if (!must_sync &&
3094 mreplace == NULL &&
3095 !conf->fullsync) {
3096 /* yep, skip the sync_blocks here, but don't assume
3097 * that there will never be anything to do here
3098 */
3099 chunks_skipped = -1;
3100 rcu_read_unlock();
3101 continue;
3102 }
3103 atomic_inc(&mrdev->nr_pending);
3104 if (mreplace)
3105 atomic_inc(&mreplace->nr_pending);
3106 rcu_read_unlock();
3107
3108 r10_bio = raid10_alloc_init_r10buf(conf);
3109 r10_bio->state = 0;
3110 raise_barrier(conf, rb2 != NULL);
3111 atomic_set(&r10_bio->remaining, 0);
3112
3113 r10_bio->master_bio = (struct bio*)rb2;
3114 if (rb2)
3115 atomic_inc(&rb2->remaining);
3116 r10_bio->mddev = mddev;
3117 set_bit(R10BIO_IsRecover, &r10_bio->state);
3118 r10_bio->sector = sect;
3119
3120 raid10_find_phys(conf, r10_bio);
3121
3122 /* Need to check if the array will still be
3123 * degraded
3124 */
3125 rcu_read_lock();
3126 for (j = 0; j < conf->geo.raid_disks; j++) {
3127 struct md_rdev *rdev = rcu_dereference(
3128 conf->mirrors[j].rdev);
3129 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3130 still_degraded = 1;
3131 break;
3132 }
3133 }
3134
3135 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3136 &sync_blocks, still_degraded);
3137
3138 any_working = 0;
3139 for (j=0; j<conf->copies;j++) {
3140 int k;
3141 int d = r10_bio->devs[j].devnum;
3142 sector_t from_addr, to_addr;
3143 struct md_rdev *rdev =
3144 rcu_dereference(conf->mirrors[d].rdev);
3145 sector_t sector, first_bad;
3146 int bad_sectors;
3147 if (!rdev ||
3148 !test_bit(In_sync, &rdev->flags))
3149 continue;
3150 /* This is where we read from */
3151 any_working = 1;
3152 sector = r10_bio->devs[j].addr;
3153
3154 if (is_badblock(rdev, sector, max_sync,
3155 &first_bad, &bad_sectors)) {
3156 if (first_bad > sector)
3157 max_sync = first_bad - sector;
3158 else {
3159 bad_sectors -= (sector
3160 - first_bad);
3161 if (max_sync > bad_sectors)
3162 max_sync = bad_sectors;
3163 continue;
3164 }
3165 }
3166 bio = r10_bio->devs[0].bio;
3167 bio->bi_next = biolist;
3168 biolist = bio;
3169 bio->bi_end_io = end_sync_read;
3170 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3171 if (test_bit(FailFast, &rdev->flags))
3172 bio->bi_opf |= MD_FAILFAST;
3173 from_addr = r10_bio->devs[j].addr;
3174 bio->bi_iter.bi_sector = from_addr +
3175 rdev->data_offset;
3176 bio_set_dev(bio, rdev->bdev);
3177 atomic_inc(&rdev->nr_pending);
3178 /* and we write to 'i' (if not in_sync) */
3179
3180 for (k=0; k<conf->copies; k++)
3181 if (r10_bio->devs[k].devnum == i)
3182 break;
3183 BUG_ON(k == conf->copies);
3184 to_addr = r10_bio->devs[k].addr;
3185 r10_bio->devs[0].devnum = d;
3186 r10_bio->devs[0].addr = from_addr;
3187 r10_bio->devs[1].devnum = i;
3188 r10_bio->devs[1].addr = to_addr;
3189
3190 if (need_recover) {
3191 bio = r10_bio->devs[1].bio;
3192 bio->bi_next = biolist;
3193 biolist = bio;
3194 bio->bi_end_io = end_sync_write;
3195 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3196 bio->bi_iter.bi_sector = to_addr
3197 + mrdev->data_offset;
3198 bio_set_dev(bio, mrdev->bdev);
3199 atomic_inc(&r10_bio->remaining);
3200 } else
3201 r10_bio->devs[1].bio->bi_end_io = NULL;
3202
3203 /* and maybe write to replacement */
3204 bio = r10_bio->devs[1].repl_bio;
3205 if (bio)
3206 bio->bi_end_io = NULL;
3207 /* Note: if need_replace, then bio
3208 * cannot be NULL as r10buf_pool_alloc will
3209 * have allocated it.
3210 */
3211 if (!need_replace)
3212 break;
3213 bio->bi_next = biolist;
3214 biolist = bio;
3215 bio->bi_end_io = end_sync_write;
3216 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3217 bio->bi_iter.bi_sector = to_addr +
3218 mreplace->data_offset;
3219 bio_set_dev(bio, mreplace->bdev);
3220 atomic_inc(&r10_bio->remaining);
3221 break;
3222 }
3223 rcu_read_unlock();
3224 if (j == conf->copies) {
3225 /* Cannot recover, so abort the recovery or
3226 * record a bad block */
3227 if (any_working) {
3228 /* problem is that there are bad blocks
3229 * on other device(s)
3230 */
3231 int k;
3232 for (k = 0; k < conf->copies; k++)
3233 if (r10_bio->devs[k].devnum == i)
3234 break;
3235 if (!test_bit(In_sync,
3236 &mrdev->flags)
3237 && !rdev_set_badblocks(
3238 mrdev,
3239 r10_bio->devs[k].addr,
3240 max_sync, 0))
3241 any_working = 0;
3242 if (mreplace &&
3243 !rdev_set_badblocks(
3244 mreplace,
3245 r10_bio->devs[k].addr,
3246 max_sync, 0))
3247 any_working = 0;
3248 }
3249 if (!any_working) {
3250 if (!test_and_set_bit(MD_RECOVERY_INTR,
3251 &mddev->recovery))
3252 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3253 mdname(mddev));
3254 mirror->recovery_disabled
3255 = mddev->recovery_disabled;
3256 }
3257 put_buf(r10_bio);
3258 if (rb2)
3259 atomic_dec(&rb2->remaining);
3260 r10_bio = rb2;
3261 rdev_dec_pending(mrdev, mddev);
3262 if (mreplace)
3263 rdev_dec_pending(mreplace, mddev);
3264 break;
3265 }
3266 rdev_dec_pending(mrdev, mddev);
3267 if (mreplace)
3268 rdev_dec_pending(mreplace, mddev);
3269 if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3270 /* Only want this if there is elsewhere to
3271 * read from. 'j' is currently the first
3272 * readable copy.
3273 */
3274 int targets = 1;
3275 for (; j < conf->copies; j++) {
3276 int d = r10_bio->devs[j].devnum;
3277 if (conf->mirrors[d].rdev &&
3278 test_bit(In_sync,
3279 &conf->mirrors[d].rdev->flags))
3280 targets++;
3281 }
3282 if (targets == 1)
3283 r10_bio->devs[0].bio->bi_opf
3284 &= ~MD_FAILFAST;
3285 }
3286 }
3287 if (biolist == NULL) {
3288 while (r10_bio) {
3289 struct r10bio *rb2 = r10_bio;
3290 r10_bio = (struct r10bio*) rb2->master_bio;
3291 rb2->master_bio = NULL;
3292 put_buf(rb2);
3293 }
3294 goto giveup;
3295 }
3296 } else {
3297 /* resync. Schedule a read for every block at this virt offset */
3298 int count = 0;
3299
3300 /*
3301 * Since curr_resync_completed could probably not update in
3302 * time, and we will set cluster_sync_low based on it.
3303 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3304 * safety reason, which ensures curr_resync_completed is
3305 * updated in bitmap_cond_end_sync.
3306 */
3307 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
3308 mddev_is_clustered(mddev) &&
3309 (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
3310
3311 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
3312 &sync_blocks, mddev->degraded) &&
3313 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3314 &mddev->recovery)) {
3315 /* We can skip this block */
3316 *skipped = 1;
3317 return sync_blocks + sectors_skipped;
3318 }
3319 if (sync_blocks < max_sync)
3320 max_sync = sync_blocks;
3321 r10_bio = raid10_alloc_init_r10buf(conf);
3322 r10_bio->state = 0;
3323
3324 r10_bio->mddev = mddev;
3325 atomic_set(&r10_bio->remaining, 0);
3326 raise_barrier(conf, 0);
3327 conf->next_resync = sector_nr;
3328
3329 r10_bio->master_bio = NULL;
3330 r10_bio->sector = sector_nr;
3331 set_bit(R10BIO_IsSync, &r10_bio->state);
3332 raid10_find_phys(conf, r10_bio);
3333 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3334
3335 for (i = 0; i < conf->copies; i++) {
3336 int d = r10_bio->devs[i].devnum;
3337 sector_t first_bad, sector;
3338 int bad_sectors;
3339 struct md_rdev *rdev;
3340
3341 if (r10_bio->devs[i].repl_bio)
3342 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3343
3344 bio = r10_bio->devs[i].bio;
3345 bio->bi_status = BLK_STS_IOERR;
3346 rcu_read_lock();
3347 rdev = rcu_dereference(conf->mirrors[d].rdev);
3348 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3349 rcu_read_unlock();
3350 continue;
3351 }
3352 sector = r10_bio->devs[i].addr;
3353 if (is_badblock(rdev, sector, max_sync,
3354 &first_bad, &bad_sectors)) {
3355 if (first_bad > sector)
3356 max_sync = first_bad - sector;
3357 else {
3358 bad_sectors -= (sector - first_bad);
3359 if (max_sync > bad_sectors)
3360 max_sync = bad_sectors;
3361 rcu_read_unlock();
3362 continue;
3363 }
3364 }
3365 atomic_inc(&rdev->nr_pending);
3366 atomic_inc(&r10_bio->remaining);
3367 bio->bi_next = biolist;
3368 biolist = bio;
3369 bio->bi_end_io = end_sync_read;
3370 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3371 if (test_bit(FailFast, &rdev->flags))
3372 bio->bi_opf |= MD_FAILFAST;
3373 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3374 bio_set_dev(bio, rdev->bdev);
3375 count++;
3376
3377 rdev = rcu_dereference(conf->mirrors[d].replacement);
3378 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3379 rcu_read_unlock();
3380 continue;
3381 }
3382 atomic_inc(&rdev->nr_pending);
3383
3384 /* Need to set up for writing to the replacement */
3385 bio = r10_bio->devs[i].repl_bio;
3386 bio->bi_status = BLK_STS_IOERR;
3387
3388 sector = r10_bio->devs[i].addr;
3389 bio->bi_next = biolist;
3390 biolist = bio;
3391 bio->bi_end_io = end_sync_write;
3392 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3393 if (test_bit(FailFast, &rdev->flags))
3394 bio->bi_opf |= MD_FAILFAST;
3395 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3396 bio_set_dev(bio, rdev->bdev);
3397 count++;
3398 rcu_read_unlock();
3399 }
3400
3401 if (count < 2) {
3402 for (i=0; i<conf->copies; i++) {
3403 int d = r10_bio->devs[i].devnum;
3404 if (r10_bio->devs[i].bio->bi_end_io)
3405 rdev_dec_pending(conf->mirrors[d].rdev,
3406 mddev);
3407 if (r10_bio->devs[i].repl_bio &&
3408 r10_bio->devs[i].repl_bio->bi_end_io)
3409 rdev_dec_pending(
3410 conf->mirrors[d].replacement,
3411 mddev);
3412 }
3413 put_buf(r10_bio);
3414 biolist = NULL;
3415 goto giveup;
3416 }
3417 }
3418
3419 nr_sectors = 0;
3420 if (sector_nr + max_sync < max_sector)
3421 max_sector = sector_nr + max_sync;
3422 do {
3423 struct page *page;
3424 int len = PAGE_SIZE;
3425 if (sector_nr + (len>>9) > max_sector)
3426 len = (max_sector - sector_nr) << 9;
3427 if (len == 0)
3428 break;
3429 for (bio= biolist ; bio ; bio=bio->bi_next) {
3430 struct resync_pages *rp = get_resync_pages(bio);
3431 page = resync_fetch_page(rp, page_idx);
3432 /*
3433 * won't fail because the vec table is big enough
3434 * to hold all these pages
3435 */
3436 bio_add_page(bio, page, len, 0);
3437 }
3438 nr_sectors += len>>9;
3439 sector_nr += len>>9;
3440 } while (++page_idx < RESYNC_PAGES);
3441 r10_bio->sectors = nr_sectors;
3442
3443 if (mddev_is_clustered(mddev) &&
3444 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3445 /* It is resync not recovery */
3446 if (conf->cluster_sync_high < sector_nr + nr_sectors) {
3447 conf->cluster_sync_low = mddev->curr_resync_completed;
3448 raid10_set_cluster_sync_high(conf);
3449 /* Send resync message */
3450 md_cluster_ops->resync_info_update(mddev,
3451 conf->cluster_sync_low,
3452 conf->cluster_sync_high);
3453 }
3454 } else if (mddev_is_clustered(mddev)) {
3455 /* This is recovery not resync */
3456 sector_t sect_va1, sect_va2;
3457 bool broadcast_msg = false;
3458
3459 for (i = 0; i < conf->geo.raid_disks; i++) {
3460 /*
3461 * sector_nr is a device address for recovery, so we
3462 * need translate it to array address before compare
3463 * with cluster_sync_high.
3464 */
3465 sect_va1 = raid10_find_virt(conf, sector_nr, i);
3466
3467 if (conf->cluster_sync_high < sect_va1 + nr_sectors) {
3468 broadcast_msg = true;
3469 /*
3470 * curr_resync_completed is similar as
3471 * sector_nr, so make the translation too.
3472 */
3473 sect_va2 = raid10_find_virt(conf,
3474 mddev->curr_resync_completed, i);
3475
3476 if (conf->cluster_sync_low == 0 ||
3477 conf->cluster_sync_low > sect_va2)
3478 conf->cluster_sync_low = sect_va2;
3479 }
3480 }
3481 if (broadcast_msg) {
3482 raid10_set_cluster_sync_high(conf);
3483 md_cluster_ops->resync_info_update(mddev,
3484 conf->cluster_sync_low,
3485 conf->cluster_sync_high);
3486 }
3487 }
3488
3489 while (biolist) {
3490 bio = biolist;
3491 biolist = biolist->bi_next;
3492
3493 bio->bi_next = NULL;
3494 r10_bio = get_resync_r10bio(bio);
3495 r10_bio->sectors = nr_sectors;
3496
3497 if (bio->bi_end_io == end_sync_read) {
3498 md_sync_acct_bio(bio, nr_sectors);
3499 bio->bi_status = 0;
3500 generic_make_request(bio);
3501 }
3502 }
3503
3504 if (sectors_skipped)
3505 /* pretend they weren't skipped, it makes
3506 * no important difference in this case
3507 */
3508 md_done_sync(mddev, sectors_skipped, 1);
3509
3510 return sectors_skipped + nr_sectors;
3511 giveup:
3512 /* There is nowhere to write, so all non-sync
3513 * drives must be failed or in resync, all drives
3514 * have a bad block, so try the next chunk...
3515 */
3516 if (sector_nr + max_sync < max_sector)
3517 max_sector = sector_nr + max_sync;
3518
3519 sectors_skipped += (max_sector - sector_nr);
3520 chunks_skipped ++;
3521 sector_nr = max_sector;
3522 goto skipped;
3523}
3524
3525static sector_t
3526raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3527{
3528 sector_t size;
3529 struct r10conf *conf = mddev->private;
3530
3531 if (!raid_disks)
3532 raid_disks = min(conf->geo.raid_disks,
3533 conf->prev.raid_disks);
3534 if (!sectors)
3535 sectors = conf->dev_sectors;
3536
3537 size = sectors >> conf->geo.chunk_shift;
3538 sector_div(size, conf->geo.far_copies);
3539 size = size * raid_disks;
3540 sector_div(size, conf->geo.near_copies);
3541
3542 return size << conf->geo.chunk_shift;
3543}
3544
3545static void calc_sectors(struct r10conf *conf, sector_t size)
3546{
3547 /* Calculate the number of sectors-per-device that will
3548 * actually be used, and set conf->dev_sectors and
3549 * conf->stride
3550 */
3551
3552 size = size >> conf->geo.chunk_shift;
3553 sector_div(size, conf->geo.far_copies);
3554 size = size * conf->geo.raid_disks;
3555 sector_div(size, conf->geo.near_copies);
3556 /* 'size' is now the number of chunks in the array */
3557 /* calculate "used chunks per device" */
3558 size = size * conf->copies;
3559
3560 /* We need to round up when dividing by raid_disks to
3561 * get the stride size.
3562 */
3563 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3564
3565 conf->dev_sectors = size << conf->geo.chunk_shift;
3566
3567 if (conf->geo.far_offset)
3568 conf->geo.stride = 1 << conf->geo.chunk_shift;
3569 else {
3570 sector_div(size, conf->geo.far_copies);
3571 conf->geo.stride = size << conf->geo.chunk_shift;
3572 }
3573}
3574
3575enum geo_type {geo_new, geo_old, geo_start};
3576static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3577{
3578 int nc, fc, fo;
3579 int layout, chunk, disks;
3580 switch (new) {
3581 case geo_old:
3582 layout = mddev->layout;
3583 chunk = mddev->chunk_sectors;
3584 disks = mddev->raid_disks - mddev->delta_disks;
3585 break;
3586 case geo_new:
3587 layout = mddev->new_layout;
3588 chunk = mddev->new_chunk_sectors;
3589 disks = mddev->raid_disks;
3590 break;
3591 default: /* avoid 'may be unused' warnings */
3592 case geo_start: /* new when starting reshape - raid_disks not
3593 * updated yet. */
3594 layout = mddev->new_layout;
3595 chunk = mddev->new_chunk_sectors;
3596 disks = mddev->raid_disks + mddev->delta_disks;
3597 break;
3598 }
3599 if (layout >> 19)
3600 return -1;
3601 if (chunk < (PAGE_SIZE >> 9) ||
3602 !is_power_of_2(chunk))
3603 return -2;
3604 nc = layout & 255;
3605 fc = (layout >> 8) & 255;
3606 fo = layout & (1<<16);
3607 geo->raid_disks = disks;
3608 geo->near_copies = nc;
3609 geo->far_copies = fc;
3610 geo->far_offset = fo;
3611 switch (layout >> 17) {
3612 case 0: /* original layout. simple but not always optimal */
3613 geo->far_set_size = disks;
3614 break;
3615 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3616 * actually using this, but leave code here just in case.*/
3617 geo->far_set_size = disks/fc;
3618 WARN(geo->far_set_size < fc,
3619 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3620 break;
3621 case 2: /* "improved" layout fixed to match documentation */
3622 geo->far_set_size = fc * nc;
3623 break;
3624 default: /* Not a valid layout */
3625 return -1;
3626 }
3627 geo->chunk_mask = chunk - 1;
3628 geo->chunk_shift = ffz(~chunk);
3629 return nc*fc;
3630}
3631
3632static struct r10conf *setup_conf(struct mddev *mddev)
3633{
3634 struct r10conf *conf = NULL;
3635 int err = -EINVAL;
3636 struct geom geo;
3637 int copies;
3638
3639 copies = setup_geo(&geo, mddev, geo_new);
3640
3641 if (copies == -2) {
3642 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3643 mdname(mddev), PAGE_SIZE);
3644 goto out;
3645 }
3646
3647 if (copies < 2 || copies > mddev->raid_disks) {
3648 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3649 mdname(mddev), mddev->new_layout);
3650 goto out;
3651 }
3652
3653 err = -ENOMEM;
3654 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3655 if (!conf)
3656 goto out;
3657
3658 /* FIXME calc properly */
3659 conf->mirrors = kcalloc(mddev->raid_disks + max(0, -mddev->delta_disks),
3660 sizeof(struct raid10_info),
3661 GFP_KERNEL);
3662 if (!conf->mirrors)
3663 goto out;
3664
3665 conf->tmppage = alloc_page(GFP_KERNEL);
3666 if (!conf->tmppage)
3667 goto out;
3668
3669 conf->geo = geo;
3670 conf->copies = copies;
3671 err = mempool_init(&conf->r10bio_pool, NR_RAID_BIOS, r10bio_pool_alloc,
3672 rbio_pool_free, conf);
3673 if (err)
3674 goto out;
3675
3676 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3677 if (err)
3678 goto out;
3679
3680 calc_sectors(conf, mddev->dev_sectors);
3681 if (mddev->reshape_position == MaxSector) {
3682 conf->prev = conf->geo;
3683 conf->reshape_progress = MaxSector;
3684 } else {
3685 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3686 err = -EINVAL;
3687 goto out;
3688 }
3689 conf->reshape_progress = mddev->reshape_position;
3690 if (conf->prev.far_offset)
3691 conf->prev.stride = 1 << conf->prev.chunk_shift;
3692 else
3693 /* far_copies must be 1 */
3694 conf->prev.stride = conf->dev_sectors;
3695 }
3696 conf->reshape_safe = conf->reshape_progress;
3697 spin_lock_init(&conf->device_lock);
3698 INIT_LIST_HEAD(&conf->retry_list);
3699 INIT_LIST_HEAD(&conf->bio_end_io_list);
3700
3701 spin_lock_init(&conf->resync_lock);
3702 init_waitqueue_head(&conf->wait_barrier);
3703 atomic_set(&conf->nr_pending, 0);
3704
3705 err = -ENOMEM;
3706 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3707 if (!conf->thread)
3708 goto out;
3709
3710 conf->mddev = mddev;
3711 return conf;
3712
3713 out:
3714 if (conf) {
3715 mempool_exit(&conf->r10bio_pool);
3716 kfree(conf->mirrors);
3717 safe_put_page(conf->tmppage);
3718 bioset_exit(&conf->bio_split);
3719 kfree(conf);
3720 }
3721 return ERR_PTR(err);
3722}
3723
3724static int raid10_run(struct mddev *mddev)
3725{
3726 struct r10conf *conf;
3727 int i, disk_idx, chunk_size;
3728 struct raid10_info *disk;
3729 struct md_rdev *rdev;
3730 sector_t size;
3731 sector_t min_offset_diff = 0;
3732 int first = 1;
3733 bool discard_supported = false;
3734
3735 if (mddev_init_writes_pending(mddev) < 0)
3736 return -ENOMEM;
3737
3738 if (mddev->private == NULL) {
3739 conf = setup_conf(mddev);
3740 if (IS_ERR(conf))
3741 return PTR_ERR(conf);
3742 mddev->private = conf;
3743 }
3744 conf = mddev->private;
3745 if (!conf)
3746 goto out;
3747
3748 if (mddev_is_clustered(conf->mddev)) {
3749 int fc, fo;
3750
3751 fc = (mddev->layout >> 8) & 255;
3752 fo = mddev->layout & (1<<16);
3753 if (fc > 1 || fo > 0) {
3754 pr_err("only near layout is supported by clustered"
3755 " raid10\n");
3756 goto out_free_conf;
3757 }
3758 }
3759
3760 mddev->thread = conf->thread;
3761 conf->thread = NULL;
3762
3763 chunk_size = mddev->chunk_sectors << 9;
3764 if (mddev->queue) {
3765 blk_queue_max_discard_sectors(mddev->queue,
3766 mddev->chunk_sectors);
3767 blk_queue_max_write_same_sectors(mddev->queue, 0);
3768 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3769 blk_queue_io_min(mddev->queue, chunk_size);
3770 if (conf->geo.raid_disks % conf->geo.near_copies)
3771 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3772 else
3773 blk_queue_io_opt(mddev->queue, chunk_size *
3774 (conf->geo.raid_disks / conf->geo.near_copies));
3775 }
3776
3777 rdev_for_each(rdev, mddev) {
3778 long long diff;
3779
3780 disk_idx = rdev->raid_disk;
3781 if (disk_idx < 0)
3782 continue;
3783 if (disk_idx >= conf->geo.raid_disks &&
3784 disk_idx >= conf->prev.raid_disks)
3785 continue;
3786 disk = conf->mirrors + disk_idx;
3787
3788 if (test_bit(Replacement, &rdev->flags)) {
3789 if (disk->replacement)
3790 goto out_free_conf;
3791 disk->replacement = rdev;
3792 } else {
3793 if (disk->rdev)
3794 goto out_free_conf;
3795 disk->rdev = rdev;
3796 }
3797 diff = (rdev->new_data_offset - rdev->data_offset);
3798 if (!mddev->reshape_backwards)
3799 diff = -diff;
3800 if (diff < 0)
3801 diff = 0;
3802 if (first || diff < min_offset_diff)
3803 min_offset_diff = diff;
3804
3805 if (mddev->gendisk)
3806 disk_stack_limits(mddev->gendisk, rdev->bdev,
3807 rdev->data_offset << 9);
3808
3809 disk->head_position = 0;
3810
3811 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3812 discard_supported = true;
3813 first = 0;
3814 }
3815
3816 if (mddev->queue) {
3817 if (discard_supported)
3818 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3819 mddev->queue);
3820 else
3821 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3822 mddev->queue);
3823 }
3824 /* need to check that every block has at least one working mirror */
3825 if (!enough(conf, -1)) {
3826 pr_err("md/raid10:%s: not enough operational mirrors.\n",
3827 mdname(mddev));
3828 goto out_free_conf;
3829 }
3830
3831 if (conf->reshape_progress != MaxSector) {
3832 /* must ensure that shape change is supported */
3833 if (conf->geo.far_copies != 1 &&
3834 conf->geo.far_offset == 0)
3835 goto out_free_conf;
3836 if (conf->prev.far_copies != 1 &&
3837 conf->prev.far_offset == 0)
3838 goto out_free_conf;
3839 }
3840
3841 mddev->degraded = 0;
3842 for (i = 0;
3843 i < conf->geo.raid_disks
3844 || i < conf->prev.raid_disks;
3845 i++) {
3846
3847 disk = conf->mirrors + i;
3848
3849 if (!disk->rdev && disk->replacement) {
3850 /* The replacement is all we have - use it */
3851 disk->rdev = disk->replacement;
3852 disk->replacement = NULL;
3853 clear_bit(Replacement, &disk->rdev->flags);
3854 }
3855
3856 if (!disk->rdev ||
3857 !test_bit(In_sync, &disk->rdev->flags)) {
3858 disk->head_position = 0;
3859 mddev->degraded++;
3860 if (disk->rdev &&
3861 disk->rdev->saved_raid_disk < 0)
3862 conf->fullsync = 1;
3863 }
3864
3865 if (disk->replacement &&
3866 !test_bit(In_sync, &disk->replacement->flags) &&
3867 disk->replacement->saved_raid_disk < 0) {
3868 conf->fullsync = 1;
3869 }
3870
3871 disk->recovery_disabled = mddev->recovery_disabled - 1;
3872 }
3873
3874 if (mddev->recovery_cp != MaxSector)
3875 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
3876 mdname(mddev));
3877 pr_info("md/raid10:%s: active with %d out of %d devices\n",
3878 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3879 conf->geo.raid_disks);
3880 /*
3881 * Ok, everything is just fine now
3882 */
3883 mddev->dev_sectors = conf->dev_sectors;
3884 size = raid10_size(mddev, 0, 0);
3885 md_set_array_sectors(mddev, size);
3886 mddev->resync_max_sectors = size;
3887 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3888
3889 if (mddev->queue) {
3890 int stripe = conf->geo.raid_disks *
3891 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3892
3893 /* Calculate max read-ahead size.
3894 * We need to readahead at least twice a whole stripe....
3895 * maybe...
3896 */
3897 stripe /= conf->geo.near_copies;
3898 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
3899 mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
3900 }
3901
3902 if (md_integrity_register(mddev))
3903 goto out_free_conf;
3904
3905 if (conf->reshape_progress != MaxSector) {
3906 unsigned long before_length, after_length;
3907
3908 before_length = ((1 << conf->prev.chunk_shift) *
3909 conf->prev.far_copies);
3910 after_length = ((1 << conf->geo.chunk_shift) *
3911 conf->geo.far_copies);
3912
3913 if (max(before_length, after_length) > min_offset_diff) {
3914 /* This cannot work */
3915 pr_warn("md/raid10: offset difference not enough to continue reshape\n");
3916 goto out_free_conf;
3917 }
3918 conf->offset_diff = min_offset_diff;
3919
3920 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3921 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3922 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3923 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3924 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3925 "reshape");
3926 if (!mddev->sync_thread)
3927 goto out_free_conf;
3928 }
3929
3930 return 0;
3931
3932out_free_conf:
3933 md_unregister_thread(&mddev->thread);
3934 mempool_exit(&conf->r10bio_pool);
3935 safe_put_page(conf->tmppage);
3936 kfree(conf->mirrors);
3937 kfree(conf);
3938 mddev->private = NULL;
3939out:
3940 return -EIO;
3941}
3942
3943static void raid10_free(struct mddev *mddev, void *priv)
3944{
3945 struct r10conf *conf = priv;
3946
3947 mempool_exit(&conf->r10bio_pool);
3948 safe_put_page(conf->tmppage);
3949 kfree(conf->mirrors);
3950 kfree(conf->mirrors_old);
3951 kfree(conf->mirrors_new);
3952 bioset_exit(&conf->bio_split);
3953 kfree(conf);
3954}
3955
3956static void raid10_quiesce(struct mddev *mddev, int quiesce)
3957{
3958 struct r10conf *conf = mddev->private;
3959
3960 if (quiesce)
3961 raise_barrier(conf, 0);
3962 else
3963 lower_barrier(conf);
3964}
3965
3966static int raid10_resize(struct mddev *mddev, sector_t sectors)
3967{
3968 /* Resize of 'far' arrays is not supported.
3969 * For 'near' and 'offset' arrays we can set the
3970 * number of sectors used to be an appropriate multiple
3971 * of the chunk size.
3972 * For 'offset', this is far_copies*chunksize.
3973 * For 'near' the multiplier is the LCM of
3974 * near_copies and raid_disks.
3975 * So if far_copies > 1 && !far_offset, fail.
3976 * Else find LCM(raid_disks, near_copy)*far_copies and
3977 * multiply by chunk_size. Then round to this number.
3978 * This is mostly done by raid10_size()
3979 */
3980 struct r10conf *conf = mddev->private;
3981 sector_t oldsize, size;
3982
3983 if (mddev->reshape_position != MaxSector)
3984 return -EBUSY;
3985
3986 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3987 return -EINVAL;
3988
3989 oldsize = raid10_size(mddev, 0, 0);
3990 size = raid10_size(mddev, sectors, 0);
3991 if (mddev->external_size &&
3992 mddev->array_sectors > size)
3993 return -EINVAL;
3994 if (mddev->bitmap) {
3995 int ret = md_bitmap_resize(mddev->bitmap, size, 0, 0);
3996 if (ret)
3997 return ret;
3998 }
3999 md_set_array_sectors(mddev, size);
4000 if (sectors > mddev->dev_sectors &&
4001 mddev->recovery_cp > oldsize) {
4002 mddev->recovery_cp = oldsize;
4003 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4004 }
4005 calc_sectors(conf, sectors);
4006 mddev->dev_sectors = conf->dev_sectors;
4007 mddev->resync_max_sectors = size;
4008 return 0;
4009}
4010
4011static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
4012{
4013 struct md_rdev *rdev;
4014 struct r10conf *conf;
4015
4016 if (mddev->degraded > 0) {
4017 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
4018 mdname(mddev));
4019 return ERR_PTR(-EINVAL);
4020 }
4021 sector_div(size, devs);
4022
4023 /* Set new parameters */
4024 mddev->new_level = 10;
4025 /* new layout: far_copies = 1, near_copies = 2 */
4026 mddev->new_layout = (1<<8) + 2;
4027 mddev->new_chunk_sectors = mddev->chunk_sectors;
4028 mddev->delta_disks = mddev->raid_disks;
4029 mddev->raid_disks *= 2;
4030 /* make sure it will be not marked as dirty */
4031 mddev->recovery_cp = MaxSector;
4032 mddev->dev_sectors = size;
4033
4034 conf = setup_conf(mddev);
4035 if (!IS_ERR(conf)) {
4036 rdev_for_each(rdev, mddev)
4037 if (rdev->raid_disk >= 0) {
4038 rdev->new_raid_disk = rdev->raid_disk * 2;
4039 rdev->sectors = size;
4040 }
4041 conf->barrier = 1;
4042 }
4043
4044 return conf;
4045}
4046
4047static void *raid10_takeover(struct mddev *mddev)
4048{
4049 struct r0conf *raid0_conf;
4050
4051 /* raid10 can take over:
4052 * raid0 - providing it has only two drives
4053 */
4054 if (mddev->level == 0) {
4055 /* for raid0 takeover only one zone is supported */
4056 raid0_conf = mddev->private;
4057 if (raid0_conf->nr_strip_zones > 1) {
4058 pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4059 mdname(mddev));
4060 return ERR_PTR(-EINVAL);
4061 }
4062 return raid10_takeover_raid0(mddev,
4063 raid0_conf->strip_zone->zone_end,
4064 raid0_conf->strip_zone->nb_dev);
4065 }
4066 return ERR_PTR(-EINVAL);
4067}
4068
4069static int raid10_check_reshape(struct mddev *mddev)
4070{
4071 /* Called when there is a request to change
4072 * - layout (to ->new_layout)
4073 * - chunk size (to ->new_chunk_sectors)
4074 * - raid_disks (by delta_disks)
4075 * or when trying to restart a reshape that was ongoing.
4076 *
4077 * We need to validate the request and possibly allocate
4078 * space if that might be an issue later.
4079 *
4080 * Currently we reject any reshape of a 'far' mode array,
4081 * allow chunk size to change if new is generally acceptable,
4082 * allow raid_disks to increase, and allow
4083 * a switch between 'near' mode and 'offset' mode.
4084 */
4085 struct r10conf *conf = mddev->private;
4086 struct geom geo;
4087
4088 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
4089 return -EINVAL;
4090
4091 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
4092 /* mustn't change number of copies */
4093 return -EINVAL;
4094 if (geo.far_copies > 1 && !geo.far_offset)
4095 /* Cannot switch to 'far' mode */
4096 return -EINVAL;
4097
4098 if (mddev->array_sectors & geo.chunk_mask)
4099 /* not factor of array size */
4100 return -EINVAL;
4101
4102 if (!enough(conf, -1))
4103 return -EINVAL;
4104
4105 kfree(conf->mirrors_new);
4106 conf->mirrors_new = NULL;
4107 if (mddev->delta_disks > 0) {
4108 /* allocate new 'mirrors' list */
4109 conf->mirrors_new =
4110 kcalloc(mddev->raid_disks + mddev->delta_disks,
4111 sizeof(struct raid10_info),
4112 GFP_KERNEL);
4113 if (!conf->mirrors_new)
4114 return -ENOMEM;
4115 }
4116 return 0;
4117}
4118
4119/*
4120 * Need to check if array has failed when deciding whether to:
4121 * - start an array
4122 * - remove non-faulty devices
4123 * - add a spare
4124 * - allow a reshape
4125 * This determination is simple when no reshape is happening.
4126 * However if there is a reshape, we need to carefully check
4127 * both the before and after sections.
4128 * This is because some failed devices may only affect one
4129 * of the two sections, and some non-in_sync devices may
4130 * be insync in the section most affected by failed devices.
4131 */
4132static int calc_degraded(struct r10conf *conf)
4133{
4134 int degraded, degraded2;
4135 int i;
4136
4137 rcu_read_lock();
4138 degraded = 0;
4139 /* 'prev' section first */
4140 for (i = 0; i < conf->prev.raid_disks; i++) {
4141 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4142 if (!rdev || test_bit(Faulty, &rdev->flags))
4143 degraded++;
4144 else if (!test_bit(In_sync, &rdev->flags))
4145 /* When we can reduce the number of devices in
4146 * an array, this might not contribute to
4147 * 'degraded'. It does now.
4148 */
4149 degraded++;
4150 }
4151 rcu_read_unlock();
4152 if (conf->geo.raid_disks == conf->prev.raid_disks)
4153 return degraded;
4154 rcu_read_lock();
4155 degraded2 = 0;
4156 for (i = 0; i < conf->geo.raid_disks; i++) {
4157 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4158 if (!rdev || test_bit(Faulty, &rdev->flags))
4159 degraded2++;
4160 else if (!test_bit(In_sync, &rdev->flags)) {
4161 /* If reshape is increasing the number of devices,
4162 * this section has already been recovered, so
4163 * it doesn't contribute to degraded.
4164 * else it does.
4165 */
4166 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4167 degraded2++;
4168 }
4169 }
4170 rcu_read_unlock();
4171 if (degraded2 > degraded)
4172 return degraded2;
4173 return degraded;
4174}
4175
4176static int raid10_start_reshape(struct mddev *mddev)
4177{
4178 /* A 'reshape' has been requested. This commits
4179 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4180 * This also checks if there are enough spares and adds them
4181 * to the array.
4182 * We currently require enough spares to make the final
4183 * array non-degraded. We also require that the difference
4184 * between old and new data_offset - on each device - is
4185 * enough that we never risk over-writing.
4186 */
4187
4188 unsigned long before_length, after_length;
4189 sector_t min_offset_diff = 0;
4190 int first = 1;
4191 struct geom new;
4192 struct r10conf *conf = mddev->private;
4193 struct md_rdev *rdev;
4194 int spares = 0;
4195 int ret;
4196
4197 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4198 return -EBUSY;
4199
4200 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4201 return -EINVAL;
4202
4203 before_length = ((1 << conf->prev.chunk_shift) *
4204 conf->prev.far_copies);
4205 after_length = ((1 << conf->geo.chunk_shift) *
4206 conf->geo.far_copies);
4207
4208 rdev_for_each(rdev, mddev) {
4209 if (!test_bit(In_sync, &rdev->flags)
4210 && !test_bit(Faulty, &rdev->flags))
4211 spares++;
4212 if (rdev->raid_disk >= 0) {
4213 long long diff = (rdev->new_data_offset
4214 - rdev->data_offset);
4215 if (!mddev->reshape_backwards)
4216 diff = -diff;
4217 if (diff < 0)
4218 diff = 0;
4219 if (first || diff < min_offset_diff)
4220 min_offset_diff = diff;
4221 first = 0;
4222 }
4223 }
4224
4225 if (max(before_length, after_length) > min_offset_diff)
4226 return -EINVAL;
4227
4228 if (spares < mddev->delta_disks)
4229 return -EINVAL;
4230
4231 conf->offset_diff = min_offset_diff;
4232 spin_lock_irq(&conf->device_lock);
4233 if (conf->mirrors_new) {
4234 memcpy(conf->mirrors_new, conf->mirrors,
4235 sizeof(struct raid10_info)*conf->prev.raid_disks);
4236 smp_mb();
4237 kfree(conf->mirrors_old);
4238 conf->mirrors_old = conf->mirrors;
4239 conf->mirrors = conf->mirrors_new;
4240 conf->mirrors_new = NULL;
4241 }
4242 setup_geo(&conf->geo, mddev, geo_start);
4243 smp_mb();
4244 if (mddev->reshape_backwards) {
4245 sector_t size = raid10_size(mddev, 0, 0);
4246 if (size < mddev->array_sectors) {
4247 spin_unlock_irq(&conf->device_lock);
4248 pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4249 mdname(mddev));
4250 return -EINVAL;
4251 }
4252 mddev->resync_max_sectors = size;
4253 conf->reshape_progress = size;
4254 } else
4255 conf->reshape_progress = 0;
4256 conf->reshape_safe = conf->reshape_progress;
4257 spin_unlock_irq(&conf->device_lock);
4258
4259 if (mddev->delta_disks && mddev->bitmap) {
4260 struct mdp_superblock_1 *sb = NULL;
4261 sector_t oldsize, newsize;
4262
4263 oldsize = raid10_size(mddev, 0, 0);
4264 newsize = raid10_size(mddev, 0, conf->geo.raid_disks);
4265
4266 if (!mddev_is_clustered(mddev)) {
4267 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4268 if (ret)
4269 goto abort;
4270 else
4271 goto out;
4272 }
4273
4274 rdev_for_each(rdev, mddev) {
4275 if (rdev->raid_disk > -1 &&
4276 !test_bit(Faulty, &rdev->flags))
4277 sb = page_address(rdev->sb_page);
4278 }
4279
4280 /*
4281 * some node is already performing reshape, and no need to
4282 * call md_bitmap_resize again since it should be called when
4283 * receiving BITMAP_RESIZE msg
4284 */
4285 if ((sb && (le32_to_cpu(sb->feature_map) &
4286 MD_FEATURE_RESHAPE_ACTIVE)) || (oldsize == newsize))
4287 goto out;
4288
4289 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4290 if (ret)
4291 goto abort;
4292
4293 ret = md_cluster_ops->resize_bitmaps(mddev, newsize, oldsize);
4294 if (ret) {
4295 md_bitmap_resize(mddev->bitmap, oldsize, 0, 0);
4296 goto abort;
4297 }
4298 }
4299out:
4300 if (mddev->delta_disks > 0) {
4301 rdev_for_each(rdev, mddev)
4302 if (rdev->raid_disk < 0 &&
4303 !test_bit(Faulty, &rdev->flags)) {
4304 if (raid10_add_disk(mddev, rdev) == 0) {
4305 if (rdev->raid_disk >=
4306 conf->prev.raid_disks)
4307 set_bit(In_sync, &rdev->flags);
4308 else
4309 rdev->recovery_offset = 0;
4310
4311 if (sysfs_link_rdev(mddev, rdev))
4312 /* Failure here is OK */;
4313 }
4314 } else if (rdev->raid_disk >= conf->prev.raid_disks
4315 && !test_bit(Faulty, &rdev->flags)) {
4316 /* This is a spare that was manually added */
4317 set_bit(In_sync, &rdev->flags);
4318 }
4319 }
4320 /* When a reshape changes the number of devices,
4321 * ->degraded is measured against the larger of the
4322 * pre and post numbers.
4323 */
4324 spin_lock_irq(&conf->device_lock);
4325 mddev->degraded = calc_degraded(conf);
4326 spin_unlock_irq(&conf->device_lock);
4327 mddev->raid_disks = conf->geo.raid_disks;
4328 mddev->reshape_position = conf->reshape_progress;
4329 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4330
4331 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4332 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4333 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4334 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4335 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4336
4337 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4338 "reshape");
4339 if (!mddev->sync_thread) {
4340 ret = -EAGAIN;
4341 goto abort;
4342 }
4343 conf->reshape_checkpoint = jiffies;
4344 md_wakeup_thread(mddev->sync_thread);
4345 md_new_event(mddev);
4346 return 0;
4347
4348abort:
4349 mddev->recovery = 0;
4350 spin_lock_irq(&conf->device_lock);
4351 conf->geo = conf->prev;
4352 mddev->raid_disks = conf->geo.raid_disks;
4353 rdev_for_each(rdev, mddev)
4354 rdev->new_data_offset = rdev->data_offset;
4355 smp_wmb();
4356 conf->reshape_progress = MaxSector;
4357 conf->reshape_safe = MaxSector;
4358 mddev->reshape_position = MaxSector;
4359 spin_unlock_irq(&conf->device_lock);
4360 return ret;
4361}
4362
4363/* Calculate the last device-address that could contain
4364 * any block from the chunk that includes the array-address 's'
4365 * and report the next address.
4366 * i.e. the address returned will be chunk-aligned and after
4367 * any data that is in the chunk containing 's'.
4368 */
4369static sector_t last_dev_address(sector_t s, struct geom *geo)
4370{
4371 s = (s | geo->chunk_mask) + 1;
4372 s >>= geo->chunk_shift;
4373 s *= geo->near_copies;
4374 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4375 s *= geo->far_copies;
4376 s <<= geo->chunk_shift;
4377 return s;
4378}
4379
4380/* Calculate the first device-address that could contain
4381 * any block from the chunk that includes the array-address 's'.
4382 * This too will be the start of a chunk
4383 */
4384static sector_t first_dev_address(sector_t s, struct geom *geo)
4385{
4386 s >>= geo->chunk_shift;
4387 s *= geo->near_copies;
4388 sector_div(s, geo->raid_disks);
4389 s *= geo->far_copies;
4390 s <<= geo->chunk_shift;
4391 return s;
4392}
4393
4394static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4395 int *skipped)
4396{
4397 /* We simply copy at most one chunk (smallest of old and new)
4398 * at a time, possibly less if that exceeds RESYNC_PAGES,
4399 * or we hit a bad block or something.
4400 * This might mean we pause for normal IO in the middle of
4401 * a chunk, but that is not a problem as mddev->reshape_position
4402 * can record any location.
4403 *
4404 * If we will want to write to a location that isn't
4405 * yet recorded as 'safe' (i.e. in metadata on disk) then
4406 * we need to flush all reshape requests and update the metadata.
4407 *
4408 * When reshaping forwards (e.g. to more devices), we interpret
4409 * 'safe' as the earliest block which might not have been copied
4410 * down yet. We divide this by previous stripe size and multiply
4411 * by previous stripe length to get lowest device offset that we
4412 * cannot write to yet.
4413 * We interpret 'sector_nr' as an address that we want to write to.
4414 * From this we use last_device_address() to find where we might
4415 * write to, and first_device_address on the 'safe' position.
4416 * If this 'next' write position is after the 'safe' position,
4417 * we must update the metadata to increase the 'safe' position.
4418 *
4419 * When reshaping backwards, we round in the opposite direction
4420 * and perform the reverse test: next write position must not be
4421 * less than current safe position.
4422 *
4423 * In all this the minimum difference in data offsets
4424 * (conf->offset_diff - always positive) allows a bit of slack,
4425 * so next can be after 'safe', but not by more than offset_diff
4426 *
4427 * We need to prepare all the bios here before we start any IO
4428 * to ensure the size we choose is acceptable to all devices.
4429 * The means one for each copy for write-out and an extra one for
4430 * read-in.
4431 * We store the read-in bio in ->master_bio and the others in
4432 * ->devs[x].bio and ->devs[x].repl_bio.
4433 */
4434 struct r10conf *conf = mddev->private;
4435 struct r10bio *r10_bio;
4436 sector_t next, safe, last;
4437 int max_sectors;
4438 int nr_sectors;
4439 int s;
4440 struct md_rdev *rdev;
4441 int need_flush = 0;
4442 struct bio *blist;
4443 struct bio *bio, *read_bio;
4444 int sectors_done = 0;
4445 struct page **pages;
4446
4447 if (sector_nr == 0) {
4448 /* If restarting in the middle, skip the initial sectors */
4449 if (mddev->reshape_backwards &&
4450 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4451 sector_nr = (raid10_size(mddev, 0, 0)
4452 - conf->reshape_progress);
4453 } else if (!mddev->reshape_backwards &&
4454 conf->reshape_progress > 0)
4455 sector_nr = conf->reshape_progress;
4456 if (sector_nr) {
4457 mddev->curr_resync_completed = sector_nr;
4458 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4459 *skipped = 1;
4460 return sector_nr;
4461 }
4462 }
4463
4464 /* We don't use sector_nr to track where we are up to
4465 * as that doesn't work well for ->reshape_backwards.
4466 * So just use ->reshape_progress.
4467 */
4468 if (mddev->reshape_backwards) {
4469 /* 'next' is the earliest device address that we might
4470 * write to for this chunk in the new layout
4471 */
4472 next = first_dev_address(conf->reshape_progress - 1,
4473 &conf->geo);
4474
4475 /* 'safe' is the last device address that we might read from
4476 * in the old layout after a restart
4477 */
4478 safe = last_dev_address(conf->reshape_safe - 1,
4479 &conf->prev);
4480
4481 if (next + conf->offset_diff < safe)
4482 need_flush = 1;
4483
4484 last = conf->reshape_progress - 1;
4485 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4486 & conf->prev.chunk_mask);
4487 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4488 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4489 } else {
4490 /* 'next' is after the last device address that we
4491 * might write to for this chunk in the new layout
4492 */
4493 next = last_dev_address(conf->reshape_progress, &conf->geo);
4494
4495 /* 'safe' is the earliest device address that we might
4496 * read from in the old layout after a restart
4497 */
4498 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4499
4500 /* Need to update metadata if 'next' might be beyond 'safe'
4501 * as that would possibly corrupt data
4502 */
4503 if (next > safe + conf->offset_diff)
4504 need_flush = 1;
4505
4506 sector_nr = conf->reshape_progress;
4507 last = sector_nr | (conf->geo.chunk_mask
4508 & conf->prev.chunk_mask);
4509
4510 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4511 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4512 }
4513
4514 if (need_flush ||
4515 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4516 /* Need to update reshape_position in metadata */
4517 wait_barrier(conf);
4518 mddev->reshape_position = conf->reshape_progress;
4519 if (mddev->reshape_backwards)
4520 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4521 - conf->reshape_progress;
4522 else
4523 mddev->curr_resync_completed = conf->reshape_progress;
4524 conf->reshape_checkpoint = jiffies;
4525 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4526 md_wakeup_thread(mddev->thread);
4527 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4528 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4529 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4530 allow_barrier(conf);
4531 return sectors_done;
4532 }
4533 conf->reshape_safe = mddev->reshape_position;
4534 allow_barrier(conf);
4535 }
4536
4537 raise_barrier(conf, 0);
4538read_more:
4539 /* Now schedule reads for blocks from sector_nr to last */
4540 r10_bio = raid10_alloc_init_r10buf(conf);
4541 r10_bio->state = 0;
4542 raise_barrier(conf, 1);
4543 atomic_set(&r10_bio->remaining, 0);
4544 r10_bio->mddev = mddev;
4545 r10_bio->sector = sector_nr;
4546 set_bit(R10BIO_IsReshape, &r10_bio->state);
4547 r10_bio->sectors = last - sector_nr + 1;
4548 rdev = read_balance(conf, r10_bio, &max_sectors);
4549 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4550
4551 if (!rdev) {
4552 /* Cannot read from here, so need to record bad blocks
4553 * on all the target devices.
4554 */
4555 // FIXME
4556 mempool_free(r10_bio, &conf->r10buf_pool);
4557 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4558 return sectors_done;
4559 }
4560
4561 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4562
4563 bio_set_dev(read_bio, rdev->bdev);
4564 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4565 + rdev->data_offset);
4566 read_bio->bi_private = r10_bio;
4567 read_bio->bi_end_io = end_reshape_read;
4568 bio_set_op_attrs(read_bio, REQ_OP_READ, 0);
4569 read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4570 read_bio->bi_status = 0;
4571 read_bio->bi_vcnt = 0;
4572 read_bio->bi_iter.bi_size = 0;
4573 r10_bio->master_bio = read_bio;
4574 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4575
4576 /*
4577 * Broadcast RESYNC message to other nodes, so all nodes would not
4578 * write to the region to avoid conflict.
4579 */
4580 if (mddev_is_clustered(mddev) && conf->cluster_sync_high <= sector_nr) {
4581 struct mdp_superblock_1 *sb = NULL;
4582 int sb_reshape_pos = 0;
4583
4584 conf->cluster_sync_low = sector_nr;
4585 conf->cluster_sync_high = sector_nr + CLUSTER_RESYNC_WINDOW_SECTORS;
4586 sb = page_address(rdev->sb_page);
4587 if (sb) {
4588 sb_reshape_pos = le64_to_cpu(sb->reshape_position);
4589 /*
4590 * Set cluster_sync_low again if next address for array
4591 * reshape is less than cluster_sync_low. Since we can't
4592 * update cluster_sync_low until it has finished reshape.
4593 */
4594 if (sb_reshape_pos < conf->cluster_sync_low)
4595 conf->cluster_sync_low = sb_reshape_pos;
4596 }
4597
4598 md_cluster_ops->resync_info_update(mddev, conf->cluster_sync_low,
4599 conf->cluster_sync_high);
4600 }
4601
4602 /* Now find the locations in the new layout */
4603 __raid10_find_phys(&conf->geo, r10_bio);
4604
4605 blist = read_bio;
4606 read_bio->bi_next = NULL;
4607
4608 rcu_read_lock();
4609 for (s = 0; s < conf->copies*2; s++) {
4610 struct bio *b;
4611 int d = r10_bio->devs[s/2].devnum;
4612 struct md_rdev *rdev2;
4613 if (s&1) {
4614 rdev2 = rcu_dereference(conf->mirrors[d].replacement);
4615 b = r10_bio->devs[s/2].repl_bio;
4616 } else {
4617 rdev2 = rcu_dereference(conf->mirrors[d].rdev);
4618 b = r10_bio->devs[s/2].bio;
4619 }
4620 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4621 continue;
4622
4623 bio_set_dev(b, rdev2->bdev);
4624 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4625 rdev2->new_data_offset;
4626 b->bi_end_io = end_reshape_write;
4627 bio_set_op_attrs(b, REQ_OP_WRITE, 0);
4628 b->bi_next = blist;
4629 blist = b;
4630 }
4631
4632 /* Now add as many pages as possible to all of these bios. */
4633
4634 nr_sectors = 0;
4635 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4636 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4637 struct page *page = pages[s / (PAGE_SIZE >> 9)];
4638 int len = (max_sectors - s) << 9;
4639 if (len > PAGE_SIZE)
4640 len = PAGE_SIZE;
4641 for (bio = blist; bio ; bio = bio->bi_next) {
4642 /*
4643 * won't fail because the vec table is big enough
4644 * to hold all these pages
4645 */
4646 bio_add_page(bio, page, len, 0);
4647 }
4648 sector_nr += len >> 9;
4649 nr_sectors += len >> 9;
4650 }
4651 rcu_read_unlock();
4652 r10_bio->sectors = nr_sectors;
4653
4654 /* Now submit the read */
4655 md_sync_acct_bio(read_bio, r10_bio->sectors);
4656 atomic_inc(&r10_bio->remaining);
4657 read_bio->bi_next = NULL;
4658 generic_make_request(read_bio);
4659 sectors_done += nr_sectors;
4660 if (sector_nr <= last)
4661 goto read_more;
4662
4663 lower_barrier(conf);
4664
4665 /* Now that we have done the whole section we can
4666 * update reshape_progress
4667 */
4668 if (mddev->reshape_backwards)
4669 conf->reshape_progress -= sectors_done;
4670 else
4671 conf->reshape_progress += sectors_done;
4672
4673 return sectors_done;
4674}
4675
4676static void end_reshape_request(struct r10bio *r10_bio);
4677static int handle_reshape_read_error(struct mddev *mddev,
4678 struct r10bio *r10_bio);
4679static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4680{
4681 /* Reshape read completed. Hopefully we have a block
4682 * to write out.
4683 * If we got a read error then we do sync 1-page reads from
4684 * elsewhere until we find the data - or give up.
4685 */
4686 struct r10conf *conf = mddev->private;
4687 int s;
4688
4689 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4690 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4691 /* Reshape has been aborted */
4692 md_done_sync(mddev, r10_bio->sectors, 0);
4693 return;
4694 }
4695
4696 /* We definitely have the data in the pages, schedule the
4697 * writes.
4698 */
4699 atomic_set(&r10_bio->remaining, 1);
4700 for (s = 0; s < conf->copies*2; s++) {
4701 struct bio *b;
4702 int d = r10_bio->devs[s/2].devnum;
4703 struct md_rdev *rdev;
4704 rcu_read_lock();
4705 if (s&1) {
4706 rdev = rcu_dereference(conf->mirrors[d].replacement);
4707 b = r10_bio->devs[s/2].repl_bio;
4708 } else {
4709 rdev = rcu_dereference(conf->mirrors[d].rdev);
4710 b = r10_bio->devs[s/2].bio;
4711 }
4712 if (!rdev || test_bit(Faulty, &rdev->flags)) {
4713 rcu_read_unlock();
4714 continue;
4715 }
4716 atomic_inc(&rdev->nr_pending);
4717 rcu_read_unlock();
4718 md_sync_acct_bio(b, r10_bio->sectors);
4719 atomic_inc(&r10_bio->remaining);
4720 b->bi_next = NULL;
4721 generic_make_request(b);
4722 }
4723 end_reshape_request(r10_bio);
4724}
4725
4726static void end_reshape(struct r10conf *conf)
4727{
4728 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4729 return;
4730
4731 spin_lock_irq(&conf->device_lock);
4732 conf->prev = conf->geo;
4733 md_finish_reshape(conf->mddev);
4734 smp_wmb();
4735 conf->reshape_progress = MaxSector;
4736 conf->reshape_safe = MaxSector;
4737 spin_unlock_irq(&conf->device_lock);
4738
4739 /* read-ahead size must cover two whole stripes, which is
4740 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4741 */
4742 if (conf->mddev->queue) {
4743 int stripe = conf->geo.raid_disks *
4744 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4745 stripe /= conf->geo.near_copies;
4746 if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
4747 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
4748 }
4749 conf->fullsync = 0;
4750}
4751
4752static void raid10_update_reshape_pos(struct mddev *mddev)
4753{
4754 struct r10conf *conf = mddev->private;
4755 sector_t lo, hi;
4756
4757 md_cluster_ops->resync_info_get(mddev, &lo, &hi);
4758 if (((mddev->reshape_position <= hi) && (mddev->reshape_position >= lo))
4759 || mddev->reshape_position == MaxSector)
4760 conf->reshape_progress = mddev->reshape_position;
4761 else
4762 WARN_ON_ONCE(1);
4763}
4764
4765static int handle_reshape_read_error(struct mddev *mddev,
4766 struct r10bio *r10_bio)
4767{
4768 /* Use sync reads to get the blocks from somewhere else */
4769 int sectors = r10_bio->sectors;
4770 struct r10conf *conf = mddev->private;
4771 struct r10bio *r10b;
4772 int slot = 0;
4773 int idx = 0;
4774 struct page **pages;
4775
4776 r10b = kmalloc(struct_size(r10b, devs, conf->copies), GFP_NOIO);
4777 if (!r10b) {
4778 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4779 return -ENOMEM;
4780 }
4781
4782 /* reshape IOs share pages from .devs[0].bio */
4783 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4784
4785 r10b->sector = r10_bio->sector;
4786 __raid10_find_phys(&conf->prev, r10b);
4787
4788 while (sectors) {
4789 int s = sectors;
4790 int success = 0;
4791 int first_slot = slot;
4792
4793 if (s > (PAGE_SIZE >> 9))
4794 s = PAGE_SIZE >> 9;
4795
4796 rcu_read_lock();
4797 while (!success) {
4798 int d = r10b->devs[slot].devnum;
4799 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4800 sector_t addr;
4801 if (rdev == NULL ||
4802 test_bit(Faulty, &rdev->flags) ||
4803 !test_bit(In_sync, &rdev->flags))
4804 goto failed;
4805
4806 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4807 atomic_inc(&rdev->nr_pending);
4808 rcu_read_unlock();
4809 success = sync_page_io(rdev,
4810 addr,
4811 s << 9,
4812 pages[idx],
4813 REQ_OP_READ, 0, false);
4814 rdev_dec_pending(rdev, mddev);
4815 rcu_read_lock();
4816 if (success)
4817 break;
4818 failed:
4819 slot++;
4820 if (slot >= conf->copies)
4821 slot = 0;
4822 if (slot == first_slot)
4823 break;
4824 }
4825 rcu_read_unlock();
4826 if (!success) {
4827 /* couldn't read this block, must give up */
4828 set_bit(MD_RECOVERY_INTR,
4829 &mddev->recovery);
4830 kfree(r10b);
4831 return -EIO;
4832 }
4833 sectors -= s;
4834 idx++;
4835 }
4836 kfree(r10b);
4837 return 0;
4838}
4839
4840static void end_reshape_write(struct bio *bio)
4841{
4842 struct r10bio *r10_bio = get_resync_r10bio(bio);
4843 struct mddev *mddev = r10_bio->mddev;
4844 struct r10conf *conf = mddev->private;
4845 int d;
4846 int slot;
4847 int repl;
4848 struct md_rdev *rdev = NULL;
4849
4850 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4851 if (repl)
4852 rdev = conf->mirrors[d].replacement;
4853 if (!rdev) {
4854 smp_mb();
4855 rdev = conf->mirrors[d].rdev;
4856 }
4857
4858 if (bio->bi_status) {
4859 /* FIXME should record badblock */
4860 md_error(mddev, rdev);
4861 }
4862
4863 rdev_dec_pending(rdev, mddev);
4864 end_reshape_request(r10_bio);
4865}
4866
4867static void end_reshape_request(struct r10bio *r10_bio)
4868{
4869 if (!atomic_dec_and_test(&r10_bio->remaining))
4870 return;
4871 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4872 bio_put(r10_bio->master_bio);
4873 put_buf(r10_bio);
4874}
4875
4876static void raid10_finish_reshape(struct mddev *mddev)
4877{
4878 struct r10conf *conf = mddev->private;
4879
4880 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4881 return;
4882
4883 if (mddev->delta_disks > 0) {
4884 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4885 mddev->recovery_cp = mddev->resync_max_sectors;
4886 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4887 }
4888 mddev->resync_max_sectors = mddev->array_sectors;
4889 } else {
4890 int d;
4891 rcu_read_lock();
4892 for (d = conf->geo.raid_disks ;
4893 d < conf->geo.raid_disks - mddev->delta_disks;
4894 d++) {
4895 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4896 if (rdev)
4897 clear_bit(In_sync, &rdev->flags);
4898 rdev = rcu_dereference(conf->mirrors[d].replacement);
4899 if (rdev)
4900 clear_bit(In_sync, &rdev->flags);
4901 }
4902 rcu_read_unlock();
4903 }
4904 mddev->layout = mddev->new_layout;
4905 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4906 mddev->reshape_position = MaxSector;
4907 mddev->delta_disks = 0;
4908 mddev->reshape_backwards = 0;
4909}
4910
4911static struct md_personality raid10_personality =
4912{
4913 .name = "raid10",
4914 .level = 10,
4915 .owner = THIS_MODULE,
4916 .make_request = raid10_make_request,
4917 .run = raid10_run,
4918 .free = raid10_free,
4919 .status = raid10_status,
4920 .error_handler = raid10_error,
4921 .hot_add_disk = raid10_add_disk,
4922 .hot_remove_disk= raid10_remove_disk,
4923 .spare_active = raid10_spare_active,
4924 .sync_request = raid10_sync_request,
4925 .quiesce = raid10_quiesce,
4926 .size = raid10_size,
4927 .resize = raid10_resize,
4928 .takeover = raid10_takeover,
4929 .check_reshape = raid10_check_reshape,
4930 .start_reshape = raid10_start_reshape,
4931 .finish_reshape = raid10_finish_reshape,
4932 .update_reshape_pos = raid10_update_reshape_pos,
4933 .congested = raid10_congested,
4934};
4935
4936static int __init raid_init(void)
4937{
4938 return register_md_personality(&raid10_personality);
4939}
4940
4941static void raid_exit(void)
4942{
4943 unregister_md_personality(&raid10_personality);
4944}
4945
4946module_init(raid_init);
4947module_exit(raid_exit);
4948MODULE_LICENSE("GPL");
4949MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4950MODULE_ALIAS("md-personality-9"); /* RAID10 */
4951MODULE_ALIAS("md-raid10");
4952MODULE_ALIAS("md-level-10");
4953
4954module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);