blob: e9dc2876a6269be27b0f726b83e468545be4be88 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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 futher 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/raid/raid10.h>
22
23/*
24 * RAID10 provides a combination of RAID0 and RAID1 functionality.
25 * The layout of data is defined by
26 * chunk_size
27 * raid_disks
28 * near_copies (stored in low byte of layout)
29 * far_copies (stored in second byte of layout)
30 *
31 * The data to be stored is divided into chunks using chunksize.
32 * Each device is divided into far_copies sections.
33 * In each section, chunks are laid out in a style similar to raid0, but
34 * near_copies copies of each chunk is stored (each on a different drive).
35 * The starting device for each section is offset near_copies from the starting
36 * device of the previous section.
37 * Thus there are (near_copies*far_copies) of each chunk, and each is on a different
38 * drive.
39 * near_copies and far_copies must be at least one, and their product is at most
40 * raid_disks.
41 */
42
43/*
44 * Number of guaranteed r10bios in case of extreme VM load:
45 */
46#define NR_RAID10_BIOS 256
47
48static void unplug_slaves(mddev_t *mddev);
49
50static void * r10bio_pool_alloc(unsigned int __nocast gfp_flags, void *data)
51{
52 conf_t *conf = data;
53 r10bio_t *r10_bio;
54 int size = offsetof(struct r10bio_s, devs[conf->copies]);
55
56 /* allocate a r10bio with room for raid_disks entries in the bios array */
57 r10_bio = kmalloc(size, gfp_flags);
58 if (r10_bio)
59 memset(r10_bio, 0, size);
60 else
61 unplug_slaves(conf->mddev);
62
63 return r10_bio;
64}
65
66static void r10bio_pool_free(void *r10_bio, void *data)
67{
68 kfree(r10_bio);
69}
70
71#define RESYNC_BLOCK_SIZE (64*1024)
72//#define RESYNC_BLOCK_SIZE PAGE_SIZE
73#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
74#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
75#define RESYNC_WINDOW (2048*1024)
76
77/*
78 * When performing a resync, we need to read and compare, so
79 * we need as many pages are there are copies.
80 * When performing a recovery, we need 2 bios, one for read,
81 * one for write (we recover only one drive per r10buf)
82 *
83 */
84static void * r10buf_pool_alloc(unsigned int __nocast gfp_flags, void *data)
85{
86 conf_t *conf = data;
87 struct page *page;
88 r10bio_t *r10_bio;
89 struct bio *bio;
90 int i, j;
91 int nalloc;
92
93 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
94 if (!r10_bio) {
95 unplug_slaves(conf->mddev);
96 return NULL;
97 }
98
99 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
100 nalloc = conf->copies; /* resync */
101 else
102 nalloc = 2; /* recovery */
103
104 /*
105 * Allocate bios.
106 */
107 for (j = nalloc ; j-- ; ) {
108 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
109 if (!bio)
110 goto out_free_bio;
111 r10_bio->devs[j].bio = bio;
112 }
113 /*
114 * Allocate RESYNC_PAGES data pages and attach them
115 * where needed.
116 */
117 for (j = 0 ; j < nalloc; j++) {
118 bio = r10_bio->devs[j].bio;
119 for (i = 0; i < RESYNC_PAGES; i++) {
120 page = alloc_page(gfp_flags);
121 if (unlikely(!page))
122 goto out_free_pages;
123
124 bio->bi_io_vec[i].bv_page = page;
125 }
126 }
127
128 return r10_bio;
129
130out_free_pages:
131 for ( ; i > 0 ; i--)
132 __free_page(bio->bi_io_vec[i-1].bv_page);
133 while (j--)
134 for (i = 0; i < RESYNC_PAGES ; i++)
135 __free_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
136 j = -1;
137out_free_bio:
138 while ( ++j < nalloc )
139 bio_put(r10_bio->devs[j].bio);
140 r10bio_pool_free(r10_bio, conf);
141 return NULL;
142}
143
144static void r10buf_pool_free(void *__r10_bio, void *data)
145{
146 int i;
147 conf_t *conf = data;
148 r10bio_t *r10bio = __r10_bio;
149 int j;
150
151 for (j=0; j < conf->copies; j++) {
152 struct bio *bio = r10bio->devs[j].bio;
153 if (bio) {
154 for (i = 0; i < RESYNC_PAGES; i++) {
155 __free_page(bio->bi_io_vec[i].bv_page);
156 bio->bi_io_vec[i].bv_page = NULL;
157 }
158 bio_put(bio);
159 }
160 }
161 r10bio_pool_free(r10bio, conf);
162}
163
164static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
165{
166 int i;
167
168 for (i = 0; i < conf->copies; i++) {
169 struct bio **bio = & r10_bio->devs[i].bio;
170 if (*bio)
171 bio_put(*bio);
172 *bio = NULL;
173 }
174}
175
176static inline void free_r10bio(r10bio_t *r10_bio)
177{
178 unsigned long flags;
179
180 conf_t *conf = mddev_to_conf(r10_bio->mddev);
181
182 /*
183 * Wake up any possible resync thread that waits for the device
184 * to go idle.
185 */
186 spin_lock_irqsave(&conf->resync_lock, flags);
187 if (!--conf->nr_pending) {
188 wake_up(&conf->wait_idle);
189 wake_up(&conf->wait_resume);
190 }
191 spin_unlock_irqrestore(&conf->resync_lock, flags);
192
193 put_all_bios(conf, r10_bio);
194 mempool_free(r10_bio, conf->r10bio_pool);
195}
196
197static inline void put_buf(r10bio_t *r10_bio)
198{
199 conf_t *conf = mddev_to_conf(r10_bio->mddev);
200 unsigned long flags;
201
202 mempool_free(r10_bio, conf->r10buf_pool);
203
204 spin_lock_irqsave(&conf->resync_lock, flags);
205 if (!conf->barrier)
206 BUG();
207 --conf->barrier;
208 wake_up(&conf->wait_resume);
209 wake_up(&conf->wait_idle);
210
211 if (!--conf->nr_pending) {
212 wake_up(&conf->wait_idle);
213 wake_up(&conf->wait_resume);
214 }
215 spin_unlock_irqrestore(&conf->resync_lock, flags);
216}
217
218static void reschedule_retry(r10bio_t *r10_bio)
219{
220 unsigned long flags;
221 mddev_t *mddev = r10_bio->mddev;
222 conf_t *conf = mddev_to_conf(mddev);
223
224 spin_lock_irqsave(&conf->device_lock, flags);
225 list_add(&r10_bio->retry_list, &conf->retry_list);
226 spin_unlock_irqrestore(&conf->device_lock, flags);
227
228 md_wakeup_thread(mddev->thread);
229}
230
231/*
232 * raid_end_bio_io() is called when we have finished servicing a mirrored
233 * operation and are ready to return a success/failure code to the buffer
234 * cache layer.
235 */
236static void raid_end_bio_io(r10bio_t *r10_bio)
237{
238 struct bio *bio = r10_bio->master_bio;
239
240 bio_endio(bio, bio->bi_size,
241 test_bit(R10BIO_Uptodate, &r10_bio->state) ? 0 : -EIO);
242 free_r10bio(r10_bio);
243}
244
245/*
246 * Update disk head position estimator based on IRQ completion info.
247 */
248static inline void update_head_pos(int slot, r10bio_t *r10_bio)
249{
250 conf_t *conf = mddev_to_conf(r10_bio->mddev);
251
252 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
253 r10_bio->devs[slot].addr + (r10_bio->sectors);
254}
255
256static int raid10_end_read_request(struct bio *bio, unsigned int bytes_done, int error)
257{
258 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
259 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
260 int slot, dev;
261 conf_t *conf = mddev_to_conf(r10_bio->mddev);
262
263 if (bio->bi_size)
264 return 1;
265
266 slot = r10_bio->read_slot;
267 dev = r10_bio->devs[slot].devnum;
268 /*
269 * this branch is our 'one mirror IO has finished' event handler:
270 */
271 if (!uptodate)
272 md_error(r10_bio->mddev, conf->mirrors[dev].rdev);
273 else
274 /*
275 * Set R10BIO_Uptodate in our master bio, so that
276 * we will return a good error code to the higher
277 * levels even if IO on some other mirrored buffer fails.
278 *
279 * The 'master' represents the composite IO operation to
280 * user-side. So if something waits for IO, then it will
281 * wait for the 'master' bio.
282 */
283 set_bit(R10BIO_Uptodate, &r10_bio->state);
284
285 update_head_pos(slot, r10_bio);
286
287 /*
288 * we have only one bio on the read side
289 */
290 if (uptodate)
291 raid_end_bio_io(r10_bio);
292 else {
293 /*
294 * oops, read error:
295 */
296 char b[BDEVNAME_SIZE];
297 if (printk_ratelimit())
298 printk(KERN_ERR "raid10: %s: rescheduling sector %llu\n",
299 bdevname(conf->mirrors[dev].rdev->bdev,b), (unsigned long long)r10_bio->sector);
300 reschedule_retry(r10_bio);
301 }
302
303 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
304 return 0;
305}
306
307static int raid10_end_write_request(struct bio *bio, unsigned int bytes_done, int error)
308{
309 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
310 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
311 int slot, dev;
312 conf_t *conf = mddev_to_conf(r10_bio->mddev);
313
314 if (bio->bi_size)
315 return 1;
316
317 for (slot = 0; slot < conf->copies; slot++)
318 if (r10_bio->devs[slot].bio == bio)
319 break;
320 dev = r10_bio->devs[slot].devnum;
321
322 /*
323 * this branch is our 'one mirror IO has finished' event handler:
324 */
325 if (!uptodate)
326 md_error(r10_bio->mddev, conf->mirrors[dev].rdev);
327 else
328 /*
329 * Set R10BIO_Uptodate in our master bio, so that
330 * we will return a good error code for to the higher
331 * levels even if IO on some other mirrored buffer fails.
332 *
333 * The 'master' represents the composite IO operation to
334 * user-side. So if something waits for IO, then it will
335 * wait for the 'master' bio.
336 */
337 set_bit(R10BIO_Uptodate, &r10_bio->state);
338
339 update_head_pos(slot, r10_bio);
340
341 /*
342 *
343 * Let's see if all mirrored write operations have finished
344 * already.
345 */
346 if (atomic_dec_and_test(&r10_bio->remaining)) {
347 md_write_end(r10_bio->mddev);
348 raid_end_bio_io(r10_bio);
349 }
350
351 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
352 return 0;
353}
354
355
356/*
357 * RAID10 layout manager
358 * Aswell as the chunksize and raid_disks count, there are two
359 * parameters: near_copies and far_copies.
360 * near_copies * far_copies must be <= raid_disks.
361 * Normally one of these will be 1.
362 * If both are 1, we get raid0.
363 * If near_copies == raid_disks, we get raid1.
364 *
365 * Chunks are layed out in raid0 style with near_copies copies of the
366 * first chunk, followed by near_copies copies of the next chunk and
367 * so on.
368 * If far_copies > 1, then after 1/far_copies of the array has been assigned
369 * as described above, we start again with a device offset of near_copies.
370 * So we effectively have another copy of the whole array further down all
371 * the drives, but with blocks on different drives.
372 * With this layout, and block is never stored twice on the one device.
373 *
374 * raid10_find_phys finds the sector offset of a given virtual sector
375 * on each device that it is on. If a block isn't on a device,
376 * that entry in the array is set to MaxSector.
377 *
378 * raid10_find_virt does the reverse mapping, from a device and a
379 * sector offset to a virtual address
380 */
381
382static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
383{
384 int n,f;
385 sector_t sector;
386 sector_t chunk;
387 sector_t stripe;
388 int dev;
389
390 int slot = 0;
391
392 /* now calculate first sector/dev */
393 chunk = r10bio->sector >> conf->chunk_shift;
394 sector = r10bio->sector & conf->chunk_mask;
395
396 chunk *= conf->near_copies;
397 stripe = chunk;
398 dev = sector_div(stripe, conf->raid_disks);
399
400 sector += stripe << conf->chunk_shift;
401
402 /* and calculate all the others */
403 for (n=0; n < conf->near_copies; n++) {
404 int d = dev;
405 sector_t s = sector;
406 r10bio->devs[slot].addr = sector;
407 r10bio->devs[slot].devnum = d;
408 slot++;
409
410 for (f = 1; f < conf->far_copies; f++) {
411 d += conf->near_copies;
412 if (d >= conf->raid_disks)
413 d -= conf->raid_disks;
414 s += conf->stride;
415 r10bio->devs[slot].devnum = d;
416 r10bio->devs[slot].addr = s;
417 slot++;
418 }
419 dev++;
420 if (dev >= conf->raid_disks) {
421 dev = 0;
422 sector += (conf->chunk_mask + 1);
423 }
424 }
425 BUG_ON(slot != conf->copies);
426}
427
428static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
429{
430 sector_t offset, chunk, vchunk;
431
432 while (sector > conf->stride) {
433 sector -= conf->stride;
434 if (dev < conf->near_copies)
435 dev += conf->raid_disks - conf->near_copies;
436 else
437 dev -= conf->near_copies;
438 }
439
440 offset = sector & conf->chunk_mask;
441 chunk = sector >> conf->chunk_shift;
442 vchunk = chunk * conf->raid_disks + dev;
443 sector_div(vchunk, conf->near_copies);
444 return (vchunk << conf->chunk_shift) + offset;
445}
446
447/**
448 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
449 * @q: request queue
450 * @bio: the buffer head that's been built up so far
451 * @biovec: the request that could be merged to it.
452 *
453 * Return amount of bytes we can accept at this offset
454 * If near_copies == raid_disk, there are no striping issues,
455 * but in that case, the function isn't called at all.
456 */
457static int raid10_mergeable_bvec(request_queue_t *q, struct bio *bio,
458 struct bio_vec *bio_vec)
459{
460 mddev_t *mddev = q->queuedata;
461 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
462 int max;
463 unsigned int chunk_sectors = mddev->chunk_size >> 9;
464 unsigned int bio_sectors = bio->bi_size >> 9;
465
466 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
467 if (max < 0) max = 0; /* bio_add cannot handle a negative return */
468 if (max <= bio_vec->bv_len && bio_sectors == 0)
469 return bio_vec->bv_len;
470 else
471 return max;
472}
473
474/*
475 * This routine returns the disk from which the requested read should
476 * be done. There is a per-array 'next expected sequential IO' sector
477 * number - if this matches on the next IO then we use the last disk.
478 * There is also a per-disk 'last know head position' sector that is
479 * maintained from IRQ contexts, both the normal and the resync IO
480 * completion handlers update this position correctly. If there is no
481 * perfect sequential match then we pick the disk whose head is closest.
482 *
483 * If there are 2 mirrors in the same 2 devices, performance degrades
484 * because position is mirror, not device based.
485 *
486 * The rdev for the device selected will have nr_pending incremented.
487 */
488
489/*
490 * FIXME: possibly should rethink readbalancing and do it differently
491 * depending on near_copies / far_copies geometry.
492 */
493static int read_balance(conf_t *conf, r10bio_t *r10_bio)
494{
495 const unsigned long this_sector = r10_bio->sector;
496 int disk, slot, nslot;
497 const int sectors = r10_bio->sectors;
498 sector_t new_distance, current_distance;
499
500 raid10_find_phys(conf, r10_bio);
501 rcu_read_lock();
502 /*
503 * Check if we can balance. We can balance on the whole
504 * device if no resync is going on, or below the resync window.
505 * We take the first readable disk when above the resync window.
506 */
507 if (conf->mddev->recovery_cp < MaxSector
508 && (this_sector + sectors >= conf->next_resync)) {
509 /* make sure that disk is operational */
510 slot = 0;
511 disk = r10_bio->devs[slot].devnum;
512
513 while (!conf->mirrors[disk].rdev ||
514 !conf->mirrors[disk].rdev->in_sync) {
515 slot++;
516 if (slot == conf->copies) {
517 slot = 0;
518 disk = -1;
519 break;
520 }
521 disk = r10_bio->devs[slot].devnum;
522 }
523 goto rb_out;
524 }
525
526
527 /* make sure the disk is operational */
528 slot = 0;
529 disk = r10_bio->devs[slot].devnum;
530 while (!conf->mirrors[disk].rdev ||
531 !conf->mirrors[disk].rdev->in_sync) {
532 slot ++;
533 if (slot == conf->copies) {
534 disk = -1;
535 goto rb_out;
536 }
537 disk = r10_bio->devs[slot].devnum;
538 }
539
540
541 current_distance = abs(this_sector - conf->mirrors[disk].head_position);
542
543 /* Find the disk whose head is closest */
544
545 for (nslot = slot; nslot < conf->copies; nslot++) {
546 int ndisk = r10_bio->devs[nslot].devnum;
547
548
549 if (!conf->mirrors[ndisk].rdev ||
550 !conf->mirrors[ndisk].rdev->in_sync)
551 continue;
552
553 if (!atomic_read(&conf->mirrors[ndisk].rdev->nr_pending)) {
554 disk = ndisk;
555 slot = nslot;
556 break;
557 }
558 new_distance = abs(r10_bio->devs[nslot].addr -
559 conf->mirrors[ndisk].head_position);
560 if (new_distance < current_distance) {
561 current_distance = new_distance;
562 disk = ndisk;
563 slot = nslot;
564 }
565 }
566
567rb_out:
568 r10_bio->read_slot = slot;
569/* conf->next_seq_sect = this_sector + sectors;*/
570
571 if (disk >= 0 && conf->mirrors[disk].rdev)
572 atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
573 rcu_read_unlock();
574
575 return disk;
576}
577
578static void unplug_slaves(mddev_t *mddev)
579{
580 conf_t *conf = mddev_to_conf(mddev);
581 int i;
582
583 rcu_read_lock();
584 for (i=0; i<mddev->raid_disks; i++) {
585 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
586 if (rdev && !rdev->faulty && atomic_read(&rdev->nr_pending)) {
587 request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
588
589 atomic_inc(&rdev->nr_pending);
590 rcu_read_unlock();
591
592 if (r_queue->unplug_fn)
593 r_queue->unplug_fn(r_queue);
594
595 rdev_dec_pending(rdev, mddev);
596 rcu_read_lock();
597 }
598 }
599 rcu_read_unlock();
600}
601
602static void raid10_unplug(request_queue_t *q)
603{
604 unplug_slaves(q->queuedata);
605}
606
607static int raid10_issue_flush(request_queue_t *q, struct gendisk *disk,
608 sector_t *error_sector)
609{
610 mddev_t *mddev = q->queuedata;
611 conf_t *conf = mddev_to_conf(mddev);
612 int i, ret = 0;
613
614 rcu_read_lock();
615 for (i=0; i<mddev->raid_disks && ret == 0; i++) {
616 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
617 if (rdev && !rdev->faulty) {
618 struct block_device *bdev = rdev->bdev;
619 request_queue_t *r_queue = bdev_get_queue(bdev);
620
621 if (!r_queue->issue_flush_fn)
622 ret = -EOPNOTSUPP;
623 else {
624 atomic_inc(&rdev->nr_pending);
625 rcu_read_unlock();
626 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
627 error_sector);
628 rdev_dec_pending(rdev, mddev);
629 rcu_read_lock();
630 }
631 }
632 }
633 rcu_read_unlock();
634 return ret;
635}
636
637/*
638 * Throttle resync depth, so that we can both get proper overlapping of
639 * requests, but are still able to handle normal requests quickly.
640 */
641#define RESYNC_DEPTH 32
642
643static void device_barrier(conf_t *conf, sector_t sect)
644{
645 spin_lock_irq(&conf->resync_lock);
646 wait_event_lock_irq(conf->wait_idle, !waitqueue_active(&conf->wait_resume),
647 conf->resync_lock, unplug_slaves(conf->mddev));
648
649 if (!conf->barrier++) {
650 wait_event_lock_irq(conf->wait_idle, !conf->nr_pending,
651 conf->resync_lock, unplug_slaves(conf->mddev));
652 if (conf->nr_pending)
653 BUG();
654 }
655 wait_event_lock_irq(conf->wait_resume, conf->barrier < RESYNC_DEPTH,
656 conf->resync_lock, unplug_slaves(conf->mddev));
657 conf->next_resync = sect;
658 spin_unlock_irq(&conf->resync_lock);
659}
660
661static int make_request(request_queue_t *q, struct bio * bio)
662{
663 mddev_t *mddev = q->queuedata;
664 conf_t *conf = mddev_to_conf(mddev);
665 mirror_info_t *mirror;
666 r10bio_t *r10_bio;
667 struct bio *read_bio;
668 int i;
669 int chunk_sects = conf->chunk_mask + 1;
670
671 /* If this request crosses a chunk boundary, we need to
672 * split it. This will only happen for 1 PAGE (or less) requests.
673 */
674 if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
675 > chunk_sects &&
676 conf->near_copies < conf->raid_disks)) {
677 struct bio_pair *bp;
678 /* Sanity check -- queue functions should prevent this happening */
679 if (bio->bi_vcnt != 1 ||
680 bio->bi_idx != 0)
681 goto bad_map;
682 /* This is a one page bio that upper layers
683 * refuse to split for us, so we need to split it.
684 */
685 bp = bio_split(bio, bio_split_pool,
686 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
687 if (make_request(q, &bp->bio1))
688 generic_make_request(&bp->bio1);
689 if (make_request(q, &bp->bio2))
690 generic_make_request(&bp->bio2);
691
692 bio_pair_release(bp);
693 return 0;
694 bad_map:
695 printk("raid10_make_request bug: can't convert block across chunks"
696 " or bigger than %dk %llu %d\n", chunk_sects/2,
697 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
698
699 bio_io_error(bio, bio->bi_size);
700 return 0;
701 }
702
703 /*
704 * Register the new request and wait if the reconstruction
705 * thread has put up a bar for new requests.
706 * Continue immediately if no resync is active currently.
707 */
708 spin_lock_irq(&conf->resync_lock);
709 wait_event_lock_irq(conf->wait_resume, !conf->barrier, conf->resync_lock, );
710 conf->nr_pending++;
711 spin_unlock_irq(&conf->resync_lock);
712
713 if (bio_data_dir(bio)==WRITE) {
714 disk_stat_inc(mddev->gendisk, writes);
715 disk_stat_add(mddev->gendisk, write_sectors, bio_sectors(bio));
716 } else {
717 disk_stat_inc(mddev->gendisk, reads);
718 disk_stat_add(mddev->gendisk, read_sectors, bio_sectors(bio));
719 }
720
721 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
722
723 r10_bio->master_bio = bio;
724 r10_bio->sectors = bio->bi_size >> 9;
725
726 r10_bio->mddev = mddev;
727 r10_bio->sector = bio->bi_sector;
728
729 if (bio_data_dir(bio) == READ) {
730 /*
731 * read balancing logic:
732 */
733 int disk = read_balance(conf, r10_bio);
734 int slot = r10_bio->read_slot;
735 if (disk < 0) {
736 raid_end_bio_io(r10_bio);
737 return 0;
738 }
739 mirror = conf->mirrors + disk;
740
741 read_bio = bio_clone(bio, GFP_NOIO);
742
743 r10_bio->devs[slot].bio = read_bio;
744
745 read_bio->bi_sector = r10_bio->devs[slot].addr +
746 mirror->rdev->data_offset;
747 read_bio->bi_bdev = mirror->rdev->bdev;
748 read_bio->bi_end_io = raid10_end_read_request;
749 read_bio->bi_rw = READ;
750 read_bio->bi_private = r10_bio;
751
752 generic_make_request(read_bio);
753 return 0;
754 }
755
756 /*
757 * WRITE:
758 */
759 /* first select target devices under spinlock and
760 * inc refcount on their rdev. Record them by setting
761 * bios[x] to bio
762 */
763 raid10_find_phys(conf, r10_bio);
764 rcu_read_lock();
765 for (i = 0; i < conf->copies; i++) {
766 int d = r10_bio->devs[i].devnum;
767 if (conf->mirrors[d].rdev &&
768 !conf->mirrors[d].rdev->faulty) {
769 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
770 r10_bio->devs[i].bio = bio;
771 } else
772 r10_bio->devs[i].bio = NULL;
773 }
774 rcu_read_unlock();
775
776 atomic_set(&r10_bio->remaining, 1);
777 md_write_start(mddev);
778 for (i = 0; i < conf->copies; i++) {
779 struct bio *mbio;
780 int d = r10_bio->devs[i].devnum;
781 if (!r10_bio->devs[i].bio)
782 continue;
783
784 mbio = bio_clone(bio, GFP_NOIO);
785 r10_bio->devs[i].bio = mbio;
786
787 mbio->bi_sector = r10_bio->devs[i].addr+
788 conf->mirrors[d].rdev->data_offset;
789 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
790 mbio->bi_end_io = raid10_end_write_request;
791 mbio->bi_rw = WRITE;
792 mbio->bi_private = r10_bio;
793
794 atomic_inc(&r10_bio->remaining);
795 generic_make_request(mbio);
796 }
797
798 if (atomic_dec_and_test(&r10_bio->remaining)) {
799 md_write_end(mddev);
800 raid_end_bio_io(r10_bio);
801 }
802
803 return 0;
804}
805
806static void status(struct seq_file *seq, mddev_t *mddev)
807{
808 conf_t *conf = mddev_to_conf(mddev);
809 int i;
810
811 if (conf->near_copies < conf->raid_disks)
812 seq_printf(seq, " %dK chunks", mddev->chunk_size/1024);
813 if (conf->near_copies > 1)
814 seq_printf(seq, " %d near-copies", conf->near_copies);
815 if (conf->far_copies > 1)
816 seq_printf(seq, " %d far-copies", conf->far_copies);
817
818 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
819 conf->working_disks);
820 for (i = 0; i < conf->raid_disks; i++)
821 seq_printf(seq, "%s",
822 conf->mirrors[i].rdev &&
823 conf->mirrors[i].rdev->in_sync ? "U" : "_");
824 seq_printf(seq, "]");
825}
826
827static void error(mddev_t *mddev, mdk_rdev_t *rdev)
828{
829 char b[BDEVNAME_SIZE];
830 conf_t *conf = mddev_to_conf(mddev);
831
832 /*
833 * If it is not operational, then we have already marked it as dead
834 * else if it is the last working disks, ignore the error, let the
835 * next level up know.
836 * else mark the drive as failed
837 */
838 if (rdev->in_sync
839 && conf->working_disks == 1)
840 /*
841 * Don't fail the drive, just return an IO error.
842 * The test should really be more sophisticated than
843 * "working_disks == 1", but it isn't critical, and
844 * can wait until we do more sophisticated "is the drive
845 * really dead" tests...
846 */
847 return;
848 if (rdev->in_sync) {
849 mddev->degraded++;
850 conf->working_disks--;
851 /*
852 * if recovery is running, make sure it aborts.
853 */
854 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
855 }
856 rdev->in_sync = 0;
857 rdev->faulty = 1;
858 mddev->sb_dirty = 1;
859 printk(KERN_ALERT "raid10: Disk failure on %s, disabling device. \n"
860 " Operation continuing on %d devices\n",
861 bdevname(rdev->bdev,b), conf->working_disks);
862}
863
864static void print_conf(conf_t *conf)
865{
866 int i;
867 mirror_info_t *tmp;
868
869 printk("RAID10 conf printout:\n");
870 if (!conf) {
871 printk("(!conf)\n");
872 return;
873 }
874 printk(" --- wd:%d rd:%d\n", conf->working_disks,
875 conf->raid_disks);
876
877 for (i = 0; i < conf->raid_disks; i++) {
878 char b[BDEVNAME_SIZE];
879 tmp = conf->mirrors + i;
880 if (tmp->rdev)
881 printk(" disk %d, wo:%d, o:%d, dev:%s\n",
882 i, !tmp->rdev->in_sync, !tmp->rdev->faulty,
883 bdevname(tmp->rdev->bdev,b));
884 }
885}
886
887static void close_sync(conf_t *conf)
888{
889 spin_lock_irq(&conf->resync_lock);
890 wait_event_lock_irq(conf->wait_resume, !conf->barrier,
891 conf->resync_lock, unplug_slaves(conf->mddev));
892 spin_unlock_irq(&conf->resync_lock);
893
894 if (conf->barrier) BUG();
895 if (waitqueue_active(&conf->wait_idle)) BUG();
896
897 mempool_destroy(conf->r10buf_pool);
898 conf->r10buf_pool = NULL;
899}
900
901static int raid10_spare_active(mddev_t *mddev)
902{
903 int i;
904 conf_t *conf = mddev->private;
905 mirror_info_t *tmp;
906
907 /*
908 * Find all non-in_sync disks within the RAID10 configuration
909 * and mark them in_sync
910 */
911 for (i = 0; i < conf->raid_disks; i++) {
912 tmp = conf->mirrors + i;
913 if (tmp->rdev
914 && !tmp->rdev->faulty
915 && !tmp->rdev->in_sync) {
916 conf->working_disks++;
917 mddev->degraded--;
918 tmp->rdev->in_sync = 1;
919 }
920 }
921
922 print_conf(conf);
923 return 0;
924}
925
926
927static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
928{
929 conf_t *conf = mddev->private;
930 int found = 0;
931 int mirror;
932 mirror_info_t *p;
933
934 if (mddev->recovery_cp < MaxSector)
935 /* only hot-add to in-sync arrays, as recovery is
936 * very different from resync
937 */
938 return 0;
939
940 for (mirror=0; mirror < mddev->raid_disks; mirror++)
941 if ( !(p=conf->mirrors+mirror)->rdev) {
942
943 blk_queue_stack_limits(mddev->queue,
944 rdev->bdev->bd_disk->queue);
945 /* as we don't honour merge_bvec_fn, we must never risk
946 * violating it, so limit ->max_sector to one PAGE, as
947 * a one page request is never in violation.
948 */
949 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
950 mddev->queue->max_sectors > (PAGE_SIZE>>9))
951 mddev->queue->max_sectors = (PAGE_SIZE>>9);
952
953 p->head_position = 0;
954 rdev->raid_disk = mirror;
955 found = 1;
956 p->rdev = rdev;
957 break;
958 }
959
960 print_conf(conf);
961 return found;
962}
963
964static int raid10_remove_disk(mddev_t *mddev, int number)
965{
966 conf_t *conf = mddev->private;
967 int err = 0;
968 mdk_rdev_t *rdev;
969 mirror_info_t *p = conf->mirrors+ number;
970
971 print_conf(conf);
972 rdev = p->rdev;
973 if (rdev) {
974 if (rdev->in_sync ||
975 atomic_read(&rdev->nr_pending)) {
976 err = -EBUSY;
977 goto abort;
978 }
979 p->rdev = NULL;
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -0700980 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700981 if (atomic_read(&rdev->nr_pending)) {
982 /* lost the race, try later */
983 err = -EBUSY;
984 p->rdev = rdev;
985 }
986 }
987abort:
988
989 print_conf(conf);
990 return err;
991}
992
993
994static int end_sync_read(struct bio *bio, unsigned int bytes_done, int error)
995{
996 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
997 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
998 conf_t *conf = mddev_to_conf(r10_bio->mddev);
999 int i,d;
1000
1001 if (bio->bi_size)
1002 return 1;
1003
1004 for (i=0; i<conf->copies; i++)
1005 if (r10_bio->devs[i].bio == bio)
1006 break;
1007 if (i == conf->copies)
1008 BUG();
1009 update_head_pos(i, r10_bio);
1010 d = r10_bio->devs[i].devnum;
1011 if (!uptodate)
1012 md_error(r10_bio->mddev,
1013 conf->mirrors[d].rdev);
1014
1015 /* for reconstruct, we always reschedule after a read.
1016 * for resync, only after all reads
1017 */
1018 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1019 atomic_dec_and_test(&r10_bio->remaining)) {
1020 /* we have read all the blocks,
1021 * do the comparison in process context in raid10d
1022 */
1023 reschedule_retry(r10_bio);
1024 }
1025 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1026 return 0;
1027}
1028
1029static int end_sync_write(struct bio *bio, unsigned int bytes_done, int error)
1030{
1031 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1032 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1033 mddev_t *mddev = r10_bio->mddev;
1034 conf_t *conf = mddev_to_conf(mddev);
1035 int i,d;
1036
1037 if (bio->bi_size)
1038 return 1;
1039
1040 for (i = 0; i < conf->copies; i++)
1041 if (r10_bio->devs[i].bio == bio)
1042 break;
1043 d = r10_bio->devs[i].devnum;
1044
1045 if (!uptodate)
1046 md_error(mddev, conf->mirrors[d].rdev);
1047 update_head_pos(i, r10_bio);
1048
1049 while (atomic_dec_and_test(&r10_bio->remaining)) {
1050 if (r10_bio->master_bio == NULL) {
1051 /* the primary of several recovery bios */
1052 md_done_sync(mddev, r10_bio->sectors, 1);
1053 put_buf(r10_bio);
1054 break;
1055 } else {
1056 r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1057 put_buf(r10_bio);
1058 r10_bio = r10_bio2;
1059 }
1060 }
1061 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1062 return 0;
1063}
1064
1065/*
1066 * Note: sync and recover and handled very differently for raid10
1067 * This code is for resync.
1068 * For resync, we read through virtual addresses and read all blocks.
1069 * If there is any error, we schedule a write. The lowest numbered
1070 * drive is authoritative.
1071 * However requests come for physical address, so we need to map.
1072 * For every physical address there are raid_disks/copies virtual addresses,
1073 * which is always are least one, but is not necessarly an integer.
1074 * This means that a physical address can span multiple chunks, so we may
1075 * have to submit multiple io requests for a single sync request.
1076 */
1077/*
1078 * We check if all blocks are in-sync and only write to blocks that
1079 * aren't in sync
1080 */
1081static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1082{
1083 conf_t *conf = mddev_to_conf(mddev);
1084 int i, first;
1085 struct bio *tbio, *fbio;
1086
1087 atomic_set(&r10_bio->remaining, 1);
1088
1089 /* find the first device with a block */
1090 for (i=0; i<conf->copies; i++)
1091 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1092 break;
1093
1094 if (i == conf->copies)
1095 goto done;
1096
1097 first = i;
1098 fbio = r10_bio->devs[i].bio;
1099
1100 /* now find blocks with errors */
1101 for (i=first+1 ; i < conf->copies ; i++) {
1102 int vcnt, j, d;
1103
1104 if (!test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1105 continue;
1106 /* We know that the bi_io_vec layout is the same for
1107 * both 'first' and 'i', so we just compare them.
1108 * All vec entries are PAGE_SIZE;
1109 */
1110 tbio = r10_bio->devs[i].bio;
1111 vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1112 for (j = 0; j < vcnt; j++)
1113 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1114 page_address(tbio->bi_io_vec[j].bv_page),
1115 PAGE_SIZE))
1116 break;
1117 if (j == vcnt)
1118 continue;
1119 /* Ok, we need to write this bio
1120 * First we need to fixup bv_offset, bv_len and
1121 * bi_vecs, as the read request might have corrupted these
1122 */
1123 tbio->bi_vcnt = vcnt;
1124 tbio->bi_size = r10_bio->sectors << 9;
1125 tbio->bi_idx = 0;
1126 tbio->bi_phys_segments = 0;
1127 tbio->bi_hw_segments = 0;
1128 tbio->bi_hw_front_size = 0;
1129 tbio->bi_hw_back_size = 0;
1130 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1131 tbio->bi_flags |= 1 << BIO_UPTODATE;
1132 tbio->bi_next = NULL;
1133 tbio->bi_rw = WRITE;
1134 tbio->bi_private = r10_bio;
1135 tbio->bi_sector = r10_bio->devs[i].addr;
1136
1137 for (j=0; j < vcnt ; j++) {
1138 tbio->bi_io_vec[j].bv_offset = 0;
1139 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1140
1141 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1142 page_address(fbio->bi_io_vec[j].bv_page),
1143 PAGE_SIZE);
1144 }
1145 tbio->bi_end_io = end_sync_write;
1146
1147 d = r10_bio->devs[i].devnum;
1148 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1149 atomic_inc(&r10_bio->remaining);
1150 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1151
1152 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1153 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1154 generic_make_request(tbio);
1155 }
1156
1157done:
1158 if (atomic_dec_and_test(&r10_bio->remaining)) {
1159 md_done_sync(mddev, r10_bio->sectors, 1);
1160 put_buf(r10_bio);
1161 }
1162}
1163
1164/*
1165 * Now for the recovery code.
1166 * Recovery happens across physical sectors.
1167 * We recover all non-is_sync drives by finding the virtual address of
1168 * each, and then choose a working drive that also has that virt address.
1169 * There is a separate r10_bio for each non-in_sync drive.
1170 * Only the first two slots are in use. The first for reading,
1171 * The second for writing.
1172 *
1173 */
1174
1175static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1176{
1177 conf_t *conf = mddev_to_conf(mddev);
1178 int i, d;
1179 struct bio *bio, *wbio;
1180
1181
1182 /* move the pages across to the second bio
1183 * and submit the write request
1184 */
1185 bio = r10_bio->devs[0].bio;
1186 wbio = r10_bio->devs[1].bio;
1187 for (i=0; i < wbio->bi_vcnt; i++) {
1188 struct page *p = bio->bi_io_vec[i].bv_page;
1189 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1190 wbio->bi_io_vec[i].bv_page = p;
1191 }
1192 d = r10_bio->devs[1].devnum;
1193
1194 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1195 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1196 generic_make_request(wbio);
1197}
1198
1199
1200/*
1201 * This is a kernel thread which:
1202 *
1203 * 1. Retries failed read operations on working mirrors.
1204 * 2. Updates the raid superblock when problems encounter.
1205 * 3. Performs writes following reads for array syncronising.
1206 */
1207
1208static void raid10d(mddev_t *mddev)
1209{
1210 r10bio_t *r10_bio;
1211 struct bio *bio;
1212 unsigned long flags;
1213 conf_t *conf = mddev_to_conf(mddev);
1214 struct list_head *head = &conf->retry_list;
1215 int unplug=0;
1216 mdk_rdev_t *rdev;
1217
1218 md_check_recovery(mddev);
1219 md_handle_safemode(mddev);
1220
1221 for (;;) {
1222 char b[BDEVNAME_SIZE];
1223 spin_lock_irqsave(&conf->device_lock, flags);
1224 if (list_empty(head))
1225 break;
1226 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1227 list_del(head->prev);
1228 spin_unlock_irqrestore(&conf->device_lock, flags);
1229
1230 mddev = r10_bio->mddev;
1231 conf = mddev_to_conf(mddev);
1232 if (test_bit(R10BIO_IsSync, &r10_bio->state)) {
1233 sync_request_write(mddev, r10_bio);
1234 unplug = 1;
1235 } else if (test_bit(R10BIO_IsRecover, &r10_bio->state)) {
1236 recovery_request_write(mddev, r10_bio);
1237 unplug = 1;
1238 } else {
1239 int mirror;
1240 bio = r10_bio->devs[r10_bio->read_slot].bio;
1241 r10_bio->devs[r10_bio->read_slot].bio = NULL;
1242 bio_put(bio);
1243 mirror = read_balance(conf, r10_bio);
1244 if (mirror == -1) {
1245 printk(KERN_ALERT "raid10: %s: unrecoverable I/O"
1246 " read error for block %llu\n",
1247 bdevname(bio->bi_bdev,b),
1248 (unsigned long long)r10_bio->sector);
1249 raid_end_bio_io(r10_bio);
1250 } else {
1251 rdev = conf->mirrors[mirror].rdev;
1252 if (printk_ratelimit())
1253 printk(KERN_ERR "raid10: %s: redirecting sector %llu to"
1254 " another mirror\n",
1255 bdevname(rdev->bdev,b),
1256 (unsigned long long)r10_bio->sector);
1257 bio = bio_clone(r10_bio->master_bio, GFP_NOIO);
1258 r10_bio->devs[r10_bio->read_slot].bio = bio;
1259 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1260 + rdev->data_offset;
1261 bio->bi_bdev = rdev->bdev;
1262 bio->bi_rw = READ;
1263 bio->bi_private = r10_bio;
1264 bio->bi_end_io = raid10_end_read_request;
1265 unplug = 1;
1266 generic_make_request(bio);
1267 }
1268 }
1269 }
1270 spin_unlock_irqrestore(&conf->device_lock, flags);
1271 if (unplug)
1272 unplug_slaves(mddev);
1273}
1274
1275
1276static int init_resync(conf_t *conf)
1277{
1278 int buffs;
1279
1280 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1281 if (conf->r10buf_pool)
1282 BUG();
1283 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1284 if (!conf->r10buf_pool)
1285 return -ENOMEM;
1286 conf->next_resync = 0;
1287 return 0;
1288}
1289
1290/*
1291 * perform a "sync" on one "block"
1292 *
1293 * We need to make sure that no normal I/O request - particularly write
1294 * requests - conflict with active sync requests.
1295 *
1296 * This is achieved by tracking pending requests and a 'barrier' concept
1297 * that can be installed to exclude normal IO requests.
1298 *
1299 * Resync and recovery are handled very differently.
1300 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1301 *
1302 * For resync, we iterate over virtual addresses, read all copies,
1303 * and update if there are differences. If only one copy is live,
1304 * skip it.
1305 * For recovery, we iterate over physical addresses, read a good
1306 * value for each non-in_sync drive, and over-write.
1307 *
1308 * So, for recovery we may have several outstanding complex requests for a
1309 * given address, one for each out-of-sync device. We model this by allocating
1310 * a number of r10_bio structures, one for each out-of-sync device.
1311 * As we setup these structures, we collect all bio's together into a list
1312 * which we then process collectively to add pages, and then process again
1313 * to pass to generic_make_request.
1314 *
1315 * The r10_bio structures are linked using a borrowed master_bio pointer.
1316 * This link is counted in ->remaining. When the r10_bio that points to NULL
1317 * has its remaining count decremented to 0, the whole complex operation
1318 * is complete.
1319 *
1320 */
1321
1322static int sync_request(mddev_t *mddev, sector_t sector_nr, int go_faster)
1323{
1324 conf_t *conf = mddev_to_conf(mddev);
1325 r10bio_t *r10_bio;
1326 struct bio *biolist = NULL, *bio;
1327 sector_t max_sector, nr_sectors;
1328 int disk;
1329 int i;
1330
1331 sector_t sectors_skipped = 0;
1332 int chunks_skipped = 0;
1333
1334 if (!conf->r10buf_pool)
1335 if (init_resync(conf))
1336 return -ENOMEM;
1337
1338 skipped:
1339 max_sector = mddev->size << 1;
1340 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1341 max_sector = mddev->resync_max_sectors;
1342 if (sector_nr >= max_sector) {
1343 close_sync(conf);
1344 return sectors_skipped;
1345 }
1346 if (chunks_skipped >= conf->raid_disks) {
1347 /* if there has been nothing to do on any drive,
1348 * then there is nothing to do at all..
1349 */
1350 sector_t sec = max_sector - sector_nr;
1351 md_done_sync(mddev, sec, 1);
1352 return sec + sectors_skipped;
1353 }
1354
1355 /* make sure whole request will fit in a chunk - if chunks
1356 * are meaningful
1357 */
1358 if (conf->near_copies < conf->raid_disks &&
1359 max_sector > (sector_nr | conf->chunk_mask))
1360 max_sector = (sector_nr | conf->chunk_mask) + 1;
1361 /*
1362 * If there is non-resync activity waiting for us then
1363 * put in a delay to throttle resync.
1364 */
1365 if (!go_faster && waitqueue_active(&conf->wait_resume))
1366 msleep_interruptible(1000);
1367 device_barrier(conf, sector_nr + RESYNC_SECTORS);
1368
1369 /* Again, very different code for resync and recovery.
1370 * Both must result in an r10bio with a list of bios that
1371 * have bi_end_io, bi_sector, bi_bdev set,
1372 * and bi_private set to the r10bio.
1373 * For recovery, we may actually create several r10bios
1374 * with 2 bios in each, that correspond to the bios in the main one.
1375 * In this case, the subordinate r10bios link back through a
1376 * borrowed master_bio pointer, and the counter in the master
1377 * includes a ref from each subordinate.
1378 */
1379 /* First, we decide what to do and set ->bi_end_io
1380 * To end_sync_read if we want to read, and
1381 * end_sync_write if we will want to write.
1382 */
1383
1384 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1385 /* recovery... the complicated one */
1386 int i, j, k;
1387 r10_bio = NULL;
1388
1389 for (i=0 ; i<conf->raid_disks; i++)
1390 if (conf->mirrors[i].rdev &&
1391 !conf->mirrors[i].rdev->in_sync) {
1392 /* want to reconstruct this device */
1393 r10bio_t *rb2 = r10_bio;
1394
1395 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1396 spin_lock_irq(&conf->resync_lock);
1397 conf->nr_pending++;
1398 if (rb2) conf->barrier++;
1399 spin_unlock_irq(&conf->resync_lock);
1400 atomic_set(&r10_bio->remaining, 0);
1401
1402 r10_bio->master_bio = (struct bio*)rb2;
1403 if (rb2)
1404 atomic_inc(&rb2->remaining);
1405 r10_bio->mddev = mddev;
1406 set_bit(R10BIO_IsRecover, &r10_bio->state);
1407 r10_bio->sector = raid10_find_virt(conf, sector_nr, i);
1408 raid10_find_phys(conf, r10_bio);
1409 for (j=0; j<conf->copies;j++) {
1410 int d = r10_bio->devs[j].devnum;
1411 if (conf->mirrors[d].rdev &&
1412 conf->mirrors[d].rdev->in_sync) {
1413 /* This is where we read from */
1414 bio = r10_bio->devs[0].bio;
1415 bio->bi_next = biolist;
1416 biolist = bio;
1417 bio->bi_private = r10_bio;
1418 bio->bi_end_io = end_sync_read;
1419 bio->bi_rw = 0;
1420 bio->bi_sector = r10_bio->devs[j].addr +
1421 conf->mirrors[d].rdev->data_offset;
1422 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1423 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1424 atomic_inc(&r10_bio->remaining);
1425 /* and we write to 'i' */
1426
1427 for (k=0; k<conf->copies; k++)
1428 if (r10_bio->devs[k].devnum == i)
1429 break;
1430 bio = r10_bio->devs[1].bio;
1431 bio->bi_next = biolist;
1432 biolist = bio;
1433 bio->bi_private = r10_bio;
1434 bio->bi_end_io = end_sync_write;
1435 bio->bi_rw = 1;
1436 bio->bi_sector = r10_bio->devs[k].addr +
1437 conf->mirrors[i].rdev->data_offset;
1438 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1439
1440 r10_bio->devs[0].devnum = d;
1441 r10_bio->devs[1].devnum = i;
1442
1443 break;
1444 }
1445 }
1446 if (j == conf->copies) {
1447 BUG();
1448 }
1449 }
1450 if (biolist == NULL) {
1451 while (r10_bio) {
1452 r10bio_t *rb2 = r10_bio;
1453 r10_bio = (r10bio_t*) rb2->master_bio;
1454 rb2->master_bio = NULL;
1455 put_buf(rb2);
1456 }
1457 goto giveup;
1458 }
1459 } else {
1460 /* resync. Schedule a read for every block at this virt offset */
1461 int count = 0;
1462 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1463
1464 spin_lock_irq(&conf->resync_lock);
1465 conf->nr_pending++;
1466 spin_unlock_irq(&conf->resync_lock);
1467
1468 r10_bio->mddev = mddev;
1469 atomic_set(&r10_bio->remaining, 0);
1470
1471 r10_bio->master_bio = NULL;
1472 r10_bio->sector = sector_nr;
1473 set_bit(R10BIO_IsSync, &r10_bio->state);
1474 raid10_find_phys(conf, r10_bio);
1475 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1476
1477 for (i=0; i<conf->copies; i++) {
1478 int d = r10_bio->devs[i].devnum;
1479 bio = r10_bio->devs[i].bio;
1480 bio->bi_end_io = NULL;
1481 if (conf->mirrors[d].rdev == NULL ||
1482 conf->mirrors[d].rdev->faulty)
1483 continue;
1484 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1485 atomic_inc(&r10_bio->remaining);
1486 bio->bi_next = biolist;
1487 biolist = bio;
1488 bio->bi_private = r10_bio;
1489 bio->bi_end_io = end_sync_read;
1490 bio->bi_rw = 0;
1491 bio->bi_sector = r10_bio->devs[i].addr +
1492 conf->mirrors[d].rdev->data_offset;
1493 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1494 count++;
1495 }
1496
1497 if (count < 2) {
1498 for (i=0; i<conf->copies; i++) {
1499 int d = r10_bio->devs[i].devnum;
1500 if (r10_bio->devs[i].bio->bi_end_io)
1501 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1502 }
1503 put_buf(r10_bio);
1504 biolist = NULL;
1505 goto giveup;
1506 }
1507 }
1508
1509 for (bio = biolist; bio ; bio=bio->bi_next) {
1510
1511 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1512 if (bio->bi_end_io)
1513 bio->bi_flags |= 1 << BIO_UPTODATE;
1514 bio->bi_vcnt = 0;
1515 bio->bi_idx = 0;
1516 bio->bi_phys_segments = 0;
1517 bio->bi_hw_segments = 0;
1518 bio->bi_size = 0;
1519 }
1520
1521 nr_sectors = 0;
1522 do {
1523 struct page *page;
1524 int len = PAGE_SIZE;
1525 disk = 0;
1526 if (sector_nr + (len>>9) > max_sector)
1527 len = (max_sector - sector_nr) << 9;
1528 if (len == 0)
1529 break;
1530 for (bio= biolist ; bio ; bio=bio->bi_next) {
1531 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1532 if (bio_add_page(bio, page, len, 0) == 0) {
1533 /* stop here */
1534 struct bio *bio2;
1535 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1536 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
1537 /* remove last page from this bio */
1538 bio2->bi_vcnt--;
1539 bio2->bi_size -= len;
1540 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
1541 }
1542 goto bio_full;
1543 }
1544 disk = i;
1545 }
1546 nr_sectors += len>>9;
1547 sector_nr += len>>9;
1548 } while (biolist->bi_vcnt < RESYNC_PAGES);
1549 bio_full:
1550 r10_bio->sectors = nr_sectors;
1551
1552 while (biolist) {
1553 bio = biolist;
1554 biolist = biolist->bi_next;
1555
1556 bio->bi_next = NULL;
1557 r10_bio = bio->bi_private;
1558 r10_bio->sectors = nr_sectors;
1559
1560 if (bio->bi_end_io == end_sync_read) {
1561 md_sync_acct(bio->bi_bdev, nr_sectors);
1562 generic_make_request(bio);
1563 }
1564 }
1565
1566 return sectors_skipped + nr_sectors;
1567 giveup:
1568 /* There is nowhere to write, so all non-sync
1569 * drives must be failed, so try the next chunk...
1570 */
1571 {
1572 int sec = max_sector - sector_nr;
1573 sectors_skipped += sec;
1574 chunks_skipped ++;
1575 sector_nr = max_sector;
1576 md_done_sync(mddev, sec, 1);
1577 goto skipped;
1578 }
1579}
1580
1581static int run(mddev_t *mddev)
1582{
1583 conf_t *conf;
1584 int i, disk_idx;
1585 mirror_info_t *disk;
1586 mdk_rdev_t *rdev;
1587 struct list_head *tmp;
1588 int nc, fc;
1589 sector_t stride, size;
1590
1591 if (mddev->level != 10) {
1592 printk(KERN_ERR "raid10: %s: raid level not set correctly... (%d)\n",
1593 mdname(mddev), mddev->level);
1594 goto out;
1595 }
1596 nc = mddev->layout & 255;
1597 fc = (mddev->layout >> 8) & 255;
1598 if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
1599 (mddev->layout >> 16)) {
1600 printk(KERN_ERR "raid10: %s: unsupported raid10 layout: 0x%8x\n",
1601 mdname(mddev), mddev->layout);
1602 goto out;
1603 }
1604 /*
1605 * copy the already verified devices into our private RAID10
1606 * bookkeeping area. [whatever we allocate in run(),
1607 * should be freed in stop()]
1608 */
1609 conf = kmalloc(sizeof(conf_t), GFP_KERNEL);
1610 mddev->private = conf;
1611 if (!conf) {
1612 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1613 mdname(mddev));
1614 goto out;
1615 }
1616 memset(conf, 0, sizeof(*conf));
1617 conf->mirrors = kmalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1618 GFP_KERNEL);
1619 if (!conf->mirrors) {
1620 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1621 mdname(mddev));
1622 goto out_free_conf;
1623 }
1624 memset(conf->mirrors, 0, sizeof(struct mirror_info)*mddev->raid_disks);
1625
1626 conf->near_copies = nc;
1627 conf->far_copies = fc;
1628 conf->copies = nc*fc;
1629 conf->chunk_mask = (sector_t)(mddev->chunk_size>>9)-1;
1630 conf->chunk_shift = ffz(~mddev->chunk_size) - 9;
1631 stride = mddev->size >> (conf->chunk_shift-1);
1632 sector_div(stride, fc);
1633 conf->stride = stride << conf->chunk_shift;
1634
1635 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
1636 r10bio_pool_free, conf);
1637 if (!conf->r10bio_pool) {
1638 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1639 mdname(mddev));
1640 goto out_free_conf;
1641 }
1642 mddev->queue->unplug_fn = raid10_unplug;
1643
1644 mddev->queue->issue_flush_fn = raid10_issue_flush;
1645
1646 ITERATE_RDEV(mddev, rdev, tmp) {
1647 disk_idx = rdev->raid_disk;
1648 if (disk_idx >= mddev->raid_disks
1649 || disk_idx < 0)
1650 continue;
1651 disk = conf->mirrors + disk_idx;
1652
1653 disk->rdev = rdev;
1654
1655 blk_queue_stack_limits(mddev->queue,
1656 rdev->bdev->bd_disk->queue);
1657 /* as we don't honour merge_bvec_fn, we must never risk
1658 * violating it, so limit ->max_sector to one PAGE, as
1659 * a one page request is never in violation.
1660 */
1661 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1662 mddev->queue->max_sectors > (PAGE_SIZE>>9))
1663 mddev->queue->max_sectors = (PAGE_SIZE>>9);
1664
1665 disk->head_position = 0;
1666 if (!rdev->faulty && rdev->in_sync)
1667 conf->working_disks++;
1668 }
1669 conf->raid_disks = mddev->raid_disks;
1670 conf->mddev = mddev;
1671 spin_lock_init(&conf->device_lock);
1672 INIT_LIST_HEAD(&conf->retry_list);
1673
1674 spin_lock_init(&conf->resync_lock);
1675 init_waitqueue_head(&conf->wait_idle);
1676 init_waitqueue_head(&conf->wait_resume);
1677
1678 if (!conf->working_disks) {
1679 printk(KERN_ERR "raid10: no operational mirrors for %s\n",
1680 mdname(mddev));
1681 goto out_free_conf;
1682 }
1683
1684 mddev->degraded = 0;
1685 for (i = 0; i < conf->raid_disks; i++) {
1686
1687 disk = conf->mirrors + i;
1688
1689 if (!disk->rdev) {
1690 disk->head_position = 0;
1691 mddev->degraded++;
1692 }
1693 }
1694
1695
1696 mddev->thread = md_register_thread(raid10d, mddev, "%s_raid10");
1697 if (!mddev->thread) {
1698 printk(KERN_ERR
1699 "raid10: couldn't allocate thread for %s\n",
1700 mdname(mddev));
1701 goto out_free_conf;
1702 }
1703
1704 printk(KERN_INFO
1705 "raid10: raid set %s active with %d out of %d devices\n",
1706 mdname(mddev), mddev->raid_disks - mddev->degraded,
1707 mddev->raid_disks);
1708 /*
1709 * Ok, everything is just fine now
1710 */
1711 size = conf->stride * conf->raid_disks;
1712 sector_div(size, conf->near_copies);
1713 mddev->array_size = size/2;
1714 mddev->resync_max_sectors = size;
1715
1716 /* Calculate max read-ahead size.
1717 * We need to readahead at least twice a whole stripe....
1718 * maybe...
1719 */
1720 {
1721 int stripe = conf->raid_disks * mddev->chunk_size / PAGE_CACHE_SIZE;
1722 stripe /= conf->near_copies;
1723 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
1724 mddev->queue->backing_dev_info.ra_pages = 2* stripe;
1725 }
1726
1727 if (conf->near_copies < mddev->raid_disks)
1728 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
1729 return 0;
1730
1731out_free_conf:
1732 if (conf->r10bio_pool)
1733 mempool_destroy(conf->r10bio_pool);
1734 if (conf->mirrors)
1735 kfree(conf->mirrors);
1736 kfree(conf);
1737 mddev->private = NULL;
1738out:
1739 return -EIO;
1740}
1741
1742static int stop(mddev_t *mddev)
1743{
1744 conf_t *conf = mddev_to_conf(mddev);
1745
1746 md_unregister_thread(mddev->thread);
1747 mddev->thread = NULL;
1748 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
1749 if (conf->r10bio_pool)
1750 mempool_destroy(conf->r10bio_pool);
1751 if (conf->mirrors)
1752 kfree(conf->mirrors);
1753 kfree(conf);
1754 mddev->private = NULL;
1755 return 0;
1756}
1757
1758
1759static mdk_personality_t raid10_personality =
1760{
1761 .name = "raid10",
1762 .owner = THIS_MODULE,
1763 .make_request = make_request,
1764 .run = run,
1765 .stop = stop,
1766 .status = status,
1767 .error_handler = error,
1768 .hot_add_disk = raid10_add_disk,
1769 .hot_remove_disk= raid10_remove_disk,
1770 .spare_active = raid10_spare_active,
1771 .sync_request = sync_request,
1772};
1773
1774static int __init raid_init(void)
1775{
1776 return register_md_personality(RAID10, &raid10_personality);
1777}
1778
1779static void raid_exit(void)
1780{
1781 unregister_md_personality(RAID10);
1782}
1783
1784module_init(raid_init);
1785module_exit(raid_exit);
1786MODULE_LICENSE("GPL");
1787MODULE_ALIAS("md-personality-9"); /* RAID10 */