]> git.openfabrics.org - ~shefty/rdma-dev.git/blob - drivers/md/raid10.c
Merge branch 'for-3.2/core' of git://git.kernel.dk/linux-block
[~shefty/rdma-dev.git] / drivers / md / raid10.c
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
61 /* When there are this many requests queue to be written by
62  * the raid10 thread, we become 'congested' to provide back-pressure
63  * for writeback.
64  */
65 static int max_queued_requests = 1024;
66
67 static void allow_barrier(struct r10conf *conf);
68 static void lower_barrier(struct r10conf *conf);
69
70 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
71 {
72         struct r10conf *conf = data;
73         int size = offsetof(struct r10bio, devs[conf->copies]);
74
75         /* allocate a r10bio with room for raid_disks entries in the bios array */
76         return kzalloc(size, gfp_flags);
77 }
78
79 static void r10bio_pool_free(void *r10_bio, void *data)
80 {
81         kfree(r10_bio);
82 }
83
84 /* Maximum size of each resync request */
85 #define RESYNC_BLOCK_SIZE (64*1024)
86 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
87 /* amount of memory to reserve for resync requests */
88 #define RESYNC_WINDOW (1024*1024)
89 /* maximum number of concurrent requests, memory permitting */
90 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
91
92 /*
93  * When performing a resync, we need to read and compare, so
94  * we need as many pages are there are copies.
95  * When performing a recovery, we need 2 bios, one for read,
96  * one for write (we recover only one drive per r10buf)
97  *
98  */
99 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
100 {
101         struct r10conf *conf = data;
102         struct page *page;
103         struct r10bio *r10_bio;
104         struct bio *bio;
105         int i, j;
106         int nalloc;
107
108         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
109         if (!r10_bio)
110                 return NULL;
111
112         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
113                 nalloc = conf->copies; /* resync */
114         else
115                 nalloc = 2; /* recovery */
116
117         /*
118          * Allocate bios.
119          */
120         for (j = nalloc ; j-- ; ) {
121                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
122                 if (!bio)
123                         goto out_free_bio;
124                 r10_bio->devs[j].bio = bio;
125         }
126         /*
127          * Allocate RESYNC_PAGES data pages and attach them
128          * where needed.
129          */
130         for (j = 0 ; j < nalloc; j++) {
131                 bio = r10_bio->devs[j].bio;
132                 for (i = 0; i < RESYNC_PAGES; i++) {
133                         if (j == 1 && !test_bit(MD_RECOVERY_SYNC,
134                                                 &conf->mddev->recovery)) {
135                                 /* we can share bv_page's during recovery */
136                                 struct bio *rbio = r10_bio->devs[0].bio;
137                                 page = rbio->bi_io_vec[i].bv_page;
138                                 get_page(page);
139                         } else
140                                 page = alloc_page(gfp_flags);
141                         if (unlikely(!page))
142                                 goto out_free_pages;
143
144                         bio->bi_io_vec[i].bv_page = page;
145                 }
146         }
147
148         return r10_bio;
149
150 out_free_pages:
151         for ( ; i > 0 ; i--)
152                 safe_put_page(bio->bi_io_vec[i-1].bv_page);
153         while (j--)
154                 for (i = 0; i < RESYNC_PAGES ; i++)
155                         safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
156         j = -1;
157 out_free_bio:
158         while ( ++j < nalloc )
159                 bio_put(r10_bio->devs[j].bio);
160         r10bio_pool_free(r10_bio, conf);
161         return NULL;
162 }
163
164 static void r10buf_pool_free(void *__r10_bio, void *data)
165 {
166         int i;
167         struct r10conf *conf = data;
168         struct r10bio *r10bio = __r10_bio;
169         int j;
170
171         for (j=0; j < conf->copies; j++) {
172                 struct bio *bio = r10bio->devs[j].bio;
173                 if (bio) {
174                         for (i = 0; i < RESYNC_PAGES; i++) {
175                                 safe_put_page(bio->bi_io_vec[i].bv_page);
176                                 bio->bi_io_vec[i].bv_page = NULL;
177                         }
178                         bio_put(bio);
179                 }
180         }
181         r10bio_pool_free(r10bio, conf);
182 }
183
184 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
185 {
186         int i;
187
188         for (i = 0; i < conf->copies; i++) {
189                 struct bio **bio = & r10_bio->devs[i].bio;
190                 if (!BIO_SPECIAL(*bio))
191                         bio_put(*bio);
192                 *bio = NULL;
193         }
194 }
195
196 static void free_r10bio(struct r10bio *r10_bio)
197 {
198         struct r10conf *conf = r10_bio->mddev->private;
199
200         put_all_bios(conf, r10_bio);
201         mempool_free(r10_bio, conf->r10bio_pool);
202 }
203
204 static void put_buf(struct r10bio *r10_bio)
205 {
206         struct r10conf *conf = r10_bio->mddev->private;
207
208         mempool_free(r10_bio, conf->r10buf_pool);
209
210         lower_barrier(conf);
211 }
212
213 static void reschedule_retry(struct r10bio *r10_bio)
214 {
215         unsigned long flags;
216         struct mddev *mddev = r10_bio->mddev;
217         struct r10conf *conf = mddev->private;
218
219         spin_lock_irqsave(&conf->device_lock, flags);
220         list_add(&r10_bio->retry_list, &conf->retry_list);
221         conf->nr_queued ++;
222         spin_unlock_irqrestore(&conf->device_lock, flags);
223
224         /* wake up frozen array... */
225         wake_up(&conf->wait_barrier);
226
227         md_wakeup_thread(mddev->thread);
228 }
229
230 /*
231  * raid_end_bio_io() is called when we have finished servicing a mirrored
232  * operation and are ready to return a success/failure code to the buffer
233  * cache layer.
234  */
235 static void raid_end_bio_io(struct r10bio *r10_bio)
236 {
237         struct bio *bio = r10_bio->master_bio;
238         int done;
239         struct r10conf *conf = r10_bio->mddev->private;
240
241         if (bio->bi_phys_segments) {
242                 unsigned long flags;
243                 spin_lock_irqsave(&conf->device_lock, flags);
244                 bio->bi_phys_segments--;
245                 done = (bio->bi_phys_segments == 0);
246                 spin_unlock_irqrestore(&conf->device_lock, flags);
247         } else
248                 done = 1;
249         if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
250                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
251         if (done) {
252                 bio_endio(bio, 0);
253                 /*
254                  * Wake up any possible resync thread that waits for the device
255                  * to go idle.
256                  */
257                 allow_barrier(conf);
258         }
259         free_r10bio(r10_bio);
260 }
261
262 /*
263  * Update disk head position estimator based on IRQ completion info.
264  */
265 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
266 {
267         struct r10conf *conf = r10_bio->mddev->private;
268
269         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
270                 r10_bio->devs[slot].addr + (r10_bio->sectors);
271 }
272
273 /*
274  * Find the disk number which triggered given bio
275  */
276 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
277                          struct bio *bio, int *slotp)
278 {
279         int slot;
280
281         for (slot = 0; slot < conf->copies; slot++)
282                 if (r10_bio->devs[slot].bio == bio)
283                         break;
284
285         BUG_ON(slot == conf->copies);
286         update_head_pos(slot, r10_bio);
287
288         if (slotp)
289                 *slotp = slot;
290         return r10_bio->devs[slot].devnum;
291 }
292
293 static void raid10_end_read_request(struct bio *bio, int error)
294 {
295         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
296         struct r10bio *r10_bio = bio->bi_private;
297         int slot, dev;
298         struct r10conf *conf = r10_bio->mddev->private;
299
300
301         slot = r10_bio->read_slot;
302         dev = r10_bio->devs[slot].devnum;
303         /*
304          * this branch is our 'one mirror IO has finished' event handler:
305          */
306         update_head_pos(slot, r10_bio);
307
308         if (uptodate) {
309                 /*
310                  * Set R10BIO_Uptodate in our master bio, so that
311                  * we will return a good error code to the higher
312                  * levels even if IO on some other mirrored buffer fails.
313                  *
314                  * The 'master' represents the composite IO operation to
315                  * user-side. So if something waits for IO, then it will
316                  * wait for the 'master' bio.
317                  */
318                 set_bit(R10BIO_Uptodate, &r10_bio->state);
319                 raid_end_bio_io(r10_bio);
320                 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
321         } else {
322                 /*
323                  * oops, read error - keep the refcount on the rdev
324                  */
325                 char b[BDEVNAME_SIZE];
326                 printk_ratelimited(KERN_ERR
327                                    "md/raid10:%s: %s: rescheduling sector %llu\n",
328                                    mdname(conf->mddev),
329                                    bdevname(conf->mirrors[dev].rdev->bdev, b),
330                                    (unsigned long long)r10_bio->sector);
331                 set_bit(R10BIO_ReadError, &r10_bio->state);
332                 reschedule_retry(r10_bio);
333         }
334 }
335
336 static void close_write(struct r10bio *r10_bio)
337 {
338         /* clear the bitmap if all writes complete successfully */
339         bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
340                         r10_bio->sectors,
341                         !test_bit(R10BIO_Degraded, &r10_bio->state),
342                         0);
343         md_write_end(r10_bio->mddev);
344 }
345
346 static void one_write_done(struct r10bio *r10_bio)
347 {
348         if (atomic_dec_and_test(&r10_bio->remaining)) {
349                 if (test_bit(R10BIO_WriteError, &r10_bio->state))
350                         reschedule_retry(r10_bio);
351                 else {
352                         close_write(r10_bio);
353                         if (test_bit(R10BIO_MadeGood, &r10_bio->state))
354                                 reschedule_retry(r10_bio);
355                         else
356                                 raid_end_bio_io(r10_bio);
357                 }
358         }
359 }
360
361 static void raid10_end_write_request(struct bio *bio, int error)
362 {
363         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
364         struct r10bio *r10_bio = bio->bi_private;
365         int dev;
366         int dec_rdev = 1;
367         struct r10conf *conf = r10_bio->mddev->private;
368         int slot;
369
370         dev = find_bio_disk(conf, r10_bio, bio, &slot);
371
372         /*
373          * this branch is our 'one mirror IO has finished' event handler:
374          */
375         if (!uptodate) {
376                 set_bit(WriteErrorSeen, &conf->mirrors[dev].rdev->flags);
377                 set_bit(R10BIO_WriteError, &r10_bio->state);
378                 dec_rdev = 0;
379         } else {
380                 /*
381                  * Set R10BIO_Uptodate in our master bio, so that
382                  * we will return a good error code for to the higher
383                  * levels even if IO on some other mirrored buffer fails.
384                  *
385                  * The 'master' represents the composite IO operation to
386                  * user-side. So if something waits for IO, then it will
387                  * wait for the 'master' bio.
388                  */
389                 sector_t first_bad;
390                 int bad_sectors;
391
392                 set_bit(R10BIO_Uptodate, &r10_bio->state);
393
394                 /* Maybe we can clear some bad blocks. */
395                 if (is_badblock(conf->mirrors[dev].rdev,
396                                 r10_bio->devs[slot].addr,
397                                 r10_bio->sectors,
398                                 &first_bad, &bad_sectors)) {
399                         bio_put(bio);
400                         r10_bio->devs[slot].bio = IO_MADE_GOOD;
401                         dec_rdev = 0;
402                         set_bit(R10BIO_MadeGood, &r10_bio->state);
403                 }
404         }
405
406         /*
407          *
408          * Let's see if all mirrored write operations have finished
409          * already.
410          */
411         one_write_done(r10_bio);
412         if (dec_rdev)
413                 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
414 }
415
416
417 /*
418  * RAID10 layout manager
419  * As well as the chunksize and raid_disks count, there are two
420  * parameters: near_copies and far_copies.
421  * near_copies * far_copies must be <= raid_disks.
422  * Normally one of these will be 1.
423  * If both are 1, we get raid0.
424  * If near_copies == raid_disks, we get raid1.
425  *
426  * Chunks are laid out in raid0 style with near_copies copies of the
427  * first chunk, followed by near_copies copies of the next chunk and
428  * so on.
429  * If far_copies > 1, then after 1/far_copies of the array has been assigned
430  * as described above, we start again with a device offset of near_copies.
431  * So we effectively have another copy of the whole array further down all
432  * the drives, but with blocks on different drives.
433  * With this layout, and block is never stored twice on the one device.
434  *
435  * raid10_find_phys finds the sector offset of a given virtual sector
436  * on each device that it is on.
437  *
438  * raid10_find_virt does the reverse mapping, from a device and a
439  * sector offset to a virtual address
440  */
441
442 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
443 {
444         int n,f;
445         sector_t sector;
446         sector_t chunk;
447         sector_t stripe;
448         int dev;
449
450         int slot = 0;
451
452         /* now calculate first sector/dev */
453         chunk = r10bio->sector >> conf->chunk_shift;
454         sector = r10bio->sector & conf->chunk_mask;
455
456         chunk *= conf->near_copies;
457         stripe = chunk;
458         dev = sector_div(stripe, conf->raid_disks);
459         if (conf->far_offset)
460                 stripe *= conf->far_copies;
461
462         sector += stripe << conf->chunk_shift;
463
464         /* and calculate all the others */
465         for (n=0; n < conf->near_copies; n++) {
466                 int d = dev;
467                 sector_t s = sector;
468                 r10bio->devs[slot].addr = sector;
469                 r10bio->devs[slot].devnum = d;
470                 slot++;
471
472                 for (f = 1; f < conf->far_copies; f++) {
473                         d += conf->near_copies;
474                         if (d >= conf->raid_disks)
475                                 d -= conf->raid_disks;
476                         s += conf->stride;
477                         r10bio->devs[slot].devnum = d;
478                         r10bio->devs[slot].addr = s;
479                         slot++;
480                 }
481                 dev++;
482                 if (dev >= conf->raid_disks) {
483                         dev = 0;
484                         sector += (conf->chunk_mask + 1);
485                 }
486         }
487         BUG_ON(slot != conf->copies);
488 }
489
490 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
491 {
492         sector_t offset, chunk, vchunk;
493
494         offset = sector & conf->chunk_mask;
495         if (conf->far_offset) {
496                 int fc;
497                 chunk = sector >> conf->chunk_shift;
498                 fc = sector_div(chunk, conf->far_copies);
499                 dev -= fc * conf->near_copies;
500                 if (dev < 0)
501                         dev += conf->raid_disks;
502         } else {
503                 while (sector >= conf->stride) {
504                         sector -= conf->stride;
505                         if (dev < conf->near_copies)
506                                 dev += conf->raid_disks - conf->near_copies;
507                         else
508                                 dev -= conf->near_copies;
509                 }
510                 chunk = sector >> conf->chunk_shift;
511         }
512         vchunk = chunk * conf->raid_disks + dev;
513         sector_div(vchunk, conf->near_copies);
514         return (vchunk << conf->chunk_shift) + offset;
515 }
516
517 /**
518  *      raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
519  *      @q: request queue
520  *      @bvm: properties of new bio
521  *      @biovec: the request that could be merged to it.
522  *
523  *      Return amount of bytes we can accept at this offset
524  *      If near_copies == raid_disk, there are no striping issues,
525  *      but in that case, the function isn't called at all.
526  */
527 static int raid10_mergeable_bvec(struct request_queue *q,
528                                  struct bvec_merge_data *bvm,
529                                  struct bio_vec *biovec)
530 {
531         struct mddev *mddev = q->queuedata;
532         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
533         int max;
534         unsigned int chunk_sectors = mddev->chunk_sectors;
535         unsigned int bio_sectors = bvm->bi_size >> 9;
536
537         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
538         if (max < 0) max = 0; /* bio_add cannot handle a negative return */
539         if (max <= biovec->bv_len && bio_sectors == 0)
540                 return biovec->bv_len;
541         else
542                 return max;
543 }
544
545 /*
546  * This routine returns the disk from which the requested read should
547  * be done. There is a per-array 'next expected sequential IO' sector
548  * number - if this matches on the next IO then we use the last disk.
549  * There is also a per-disk 'last know head position' sector that is
550  * maintained from IRQ contexts, both the normal and the resync IO
551  * completion handlers update this position correctly. If there is no
552  * perfect sequential match then we pick the disk whose head is closest.
553  *
554  * If there are 2 mirrors in the same 2 devices, performance degrades
555  * because position is mirror, not device based.
556  *
557  * The rdev for the device selected will have nr_pending incremented.
558  */
559
560 /*
561  * FIXME: possibly should rethink readbalancing and do it differently
562  * depending on near_copies / far_copies geometry.
563  */
564 static int read_balance(struct r10conf *conf, struct r10bio *r10_bio, int *max_sectors)
565 {
566         const sector_t this_sector = r10_bio->sector;
567         int disk, slot;
568         int sectors = r10_bio->sectors;
569         int best_good_sectors;
570         sector_t new_distance, best_dist;
571         struct md_rdev *rdev;
572         int do_balance;
573         int best_slot;
574
575         raid10_find_phys(conf, r10_bio);
576         rcu_read_lock();
577 retry:
578         sectors = r10_bio->sectors;
579         best_slot = -1;
580         best_dist = MaxSector;
581         best_good_sectors = 0;
582         do_balance = 1;
583         /*
584          * Check if we can balance. We can balance on the whole
585          * device if no resync is going on (recovery is ok), or below
586          * the resync window. We take the first readable disk when
587          * above the resync window.
588          */
589         if (conf->mddev->recovery_cp < MaxSector
590             && (this_sector + sectors >= conf->next_resync))
591                 do_balance = 0;
592
593         for (slot = 0; slot < conf->copies ; slot++) {
594                 sector_t first_bad;
595                 int bad_sectors;
596                 sector_t dev_sector;
597
598                 if (r10_bio->devs[slot].bio == IO_BLOCKED)
599                         continue;
600                 disk = r10_bio->devs[slot].devnum;
601                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
602                 if (rdev == NULL)
603                         continue;
604                 if (!test_bit(In_sync, &rdev->flags))
605                         continue;
606
607                 dev_sector = r10_bio->devs[slot].addr;
608                 if (is_badblock(rdev, dev_sector, sectors,
609                                 &first_bad, &bad_sectors)) {
610                         if (best_dist < MaxSector)
611                                 /* Already have a better slot */
612                                 continue;
613                         if (first_bad <= dev_sector) {
614                                 /* Cannot read here.  If this is the
615                                  * 'primary' device, then we must not read
616                                  * beyond 'bad_sectors' from another device.
617                                  */
618                                 bad_sectors -= (dev_sector - first_bad);
619                                 if (!do_balance && sectors > bad_sectors)
620                                         sectors = bad_sectors;
621                                 if (best_good_sectors > sectors)
622                                         best_good_sectors = sectors;
623                         } else {
624                                 sector_t good_sectors =
625                                         first_bad - dev_sector;
626                                 if (good_sectors > best_good_sectors) {
627                                         best_good_sectors = good_sectors;
628                                         best_slot = slot;
629                                 }
630                                 if (!do_balance)
631                                         /* Must read from here */
632                                         break;
633                         }
634                         continue;
635                 } else
636                         best_good_sectors = sectors;
637
638                 if (!do_balance)
639                         break;
640
641                 /* This optimisation is debatable, and completely destroys
642                  * sequential read speed for 'far copies' arrays.  So only
643                  * keep it for 'near' arrays, and review those later.
644                  */
645                 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending))
646                         break;
647
648                 /* for far > 1 always use the lowest address */
649                 if (conf->far_copies > 1)
650                         new_distance = r10_bio->devs[slot].addr;
651                 else
652                         new_distance = abs(r10_bio->devs[slot].addr -
653                                            conf->mirrors[disk].head_position);
654                 if (new_distance < best_dist) {
655                         best_dist = new_distance;
656                         best_slot = slot;
657                 }
658         }
659         if (slot == conf->copies)
660                 slot = best_slot;
661
662         if (slot >= 0) {
663                 disk = r10_bio->devs[slot].devnum;
664                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
665                 if (!rdev)
666                         goto retry;
667                 atomic_inc(&rdev->nr_pending);
668                 if (test_bit(Faulty, &rdev->flags)) {
669                         /* Cannot risk returning a device that failed
670                          * before we inc'ed nr_pending
671                          */
672                         rdev_dec_pending(rdev, conf->mddev);
673                         goto retry;
674                 }
675                 r10_bio->read_slot = slot;
676         } else
677                 disk = -1;
678         rcu_read_unlock();
679         *max_sectors = best_good_sectors;
680
681         return disk;
682 }
683
684 static int raid10_congested(void *data, int bits)
685 {
686         struct mddev *mddev = data;
687         struct r10conf *conf = mddev->private;
688         int i, ret = 0;
689
690         if ((bits & (1 << BDI_async_congested)) &&
691             conf->pending_count >= max_queued_requests)
692                 return 1;
693
694         if (mddev_congested(mddev, bits))
695                 return 1;
696         rcu_read_lock();
697         for (i = 0; i < conf->raid_disks && ret == 0; i++) {
698                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
699                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
700                         struct request_queue *q = bdev_get_queue(rdev->bdev);
701
702                         ret |= bdi_congested(&q->backing_dev_info, bits);
703                 }
704         }
705         rcu_read_unlock();
706         return ret;
707 }
708
709 static void flush_pending_writes(struct r10conf *conf)
710 {
711         /* Any writes that have been queued but are awaiting
712          * bitmap updates get flushed here.
713          */
714         spin_lock_irq(&conf->device_lock);
715
716         if (conf->pending_bio_list.head) {
717                 struct bio *bio;
718                 bio = bio_list_get(&conf->pending_bio_list);
719                 conf->pending_count = 0;
720                 spin_unlock_irq(&conf->device_lock);
721                 /* flush any pending bitmap writes to disk
722                  * before proceeding w/ I/O */
723                 bitmap_unplug(conf->mddev->bitmap);
724                 wake_up(&conf->wait_barrier);
725
726                 while (bio) { /* submit pending writes */
727                         struct bio *next = bio->bi_next;
728                         bio->bi_next = NULL;
729                         generic_make_request(bio);
730                         bio = next;
731                 }
732         } else
733                 spin_unlock_irq(&conf->device_lock);
734 }
735
736 /* Barriers....
737  * Sometimes we need to suspend IO while we do something else,
738  * either some resync/recovery, or reconfigure the array.
739  * To do this we raise a 'barrier'.
740  * The 'barrier' is a counter that can be raised multiple times
741  * to count how many activities are happening which preclude
742  * normal IO.
743  * We can only raise the barrier if there is no pending IO.
744  * i.e. if nr_pending == 0.
745  * We choose only to raise the barrier if no-one is waiting for the
746  * barrier to go down.  This means that as soon as an IO request
747  * is ready, no other operations which require a barrier will start
748  * until the IO request has had a chance.
749  *
750  * So: regular IO calls 'wait_barrier'.  When that returns there
751  *    is no backgroup IO happening,  It must arrange to call
752  *    allow_barrier when it has finished its IO.
753  * backgroup IO calls must call raise_barrier.  Once that returns
754  *    there is no normal IO happeing.  It must arrange to call
755  *    lower_barrier when the particular background IO completes.
756  */
757
758 static void raise_barrier(struct r10conf *conf, int force)
759 {
760         BUG_ON(force && !conf->barrier);
761         spin_lock_irq(&conf->resync_lock);
762
763         /* Wait until no block IO is waiting (unless 'force') */
764         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
765                             conf->resync_lock, );
766
767         /* block any new IO from starting */
768         conf->barrier++;
769
770         /* Now wait for all pending IO to complete */
771         wait_event_lock_irq(conf->wait_barrier,
772                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
773                             conf->resync_lock, );
774
775         spin_unlock_irq(&conf->resync_lock);
776 }
777
778 static void lower_barrier(struct r10conf *conf)
779 {
780         unsigned long flags;
781         spin_lock_irqsave(&conf->resync_lock, flags);
782         conf->barrier--;
783         spin_unlock_irqrestore(&conf->resync_lock, flags);
784         wake_up(&conf->wait_barrier);
785 }
786
787 static void wait_barrier(struct r10conf *conf)
788 {
789         spin_lock_irq(&conf->resync_lock);
790         if (conf->barrier) {
791                 conf->nr_waiting++;
792                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
793                                     conf->resync_lock,
794                                     );
795                 conf->nr_waiting--;
796         }
797         conf->nr_pending++;
798         spin_unlock_irq(&conf->resync_lock);
799 }
800
801 static void allow_barrier(struct r10conf *conf)
802 {
803         unsigned long flags;
804         spin_lock_irqsave(&conf->resync_lock, flags);
805         conf->nr_pending--;
806         spin_unlock_irqrestore(&conf->resync_lock, flags);
807         wake_up(&conf->wait_barrier);
808 }
809
810 static void freeze_array(struct r10conf *conf)
811 {
812         /* stop syncio and normal IO and wait for everything to
813          * go quiet.
814          * We increment barrier and nr_waiting, and then
815          * wait until nr_pending match nr_queued+1
816          * This is called in the context of one normal IO request
817          * that has failed. Thus any sync request that might be pending
818          * will be blocked by nr_pending, and we need to wait for
819          * pending IO requests to complete or be queued for re-try.
820          * Thus the number queued (nr_queued) plus this request (1)
821          * must match the number of pending IOs (nr_pending) before
822          * we continue.
823          */
824         spin_lock_irq(&conf->resync_lock);
825         conf->barrier++;
826         conf->nr_waiting++;
827         wait_event_lock_irq(conf->wait_barrier,
828                             conf->nr_pending == conf->nr_queued+1,
829                             conf->resync_lock,
830                             flush_pending_writes(conf));
831
832         spin_unlock_irq(&conf->resync_lock);
833 }
834
835 static void unfreeze_array(struct r10conf *conf)
836 {
837         /* reverse the effect of the freeze */
838         spin_lock_irq(&conf->resync_lock);
839         conf->barrier--;
840         conf->nr_waiting--;
841         wake_up(&conf->wait_barrier);
842         spin_unlock_irq(&conf->resync_lock);
843 }
844
845 static void make_request(struct mddev *mddev, struct bio * bio)
846 {
847         struct r10conf *conf = mddev->private;
848         struct mirror_info *mirror;
849         struct r10bio *r10_bio;
850         struct bio *read_bio;
851         int i;
852         int chunk_sects = conf->chunk_mask + 1;
853         const int rw = bio_data_dir(bio);
854         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
855         const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
856         unsigned long flags;
857         struct md_rdev *blocked_rdev;
858         int plugged;
859         int sectors_handled;
860         int max_sectors;
861
862         if (unlikely(bio->bi_rw & REQ_FLUSH)) {
863                 md_flush_request(mddev, bio);
864                 return;
865         }
866
867         /* If this request crosses a chunk boundary, we need to
868          * split it.  This will only happen for 1 PAGE (or less) requests.
869          */
870         if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
871                       > chunk_sects &&
872                     conf->near_copies < conf->raid_disks)) {
873                 struct bio_pair *bp;
874                 /* Sanity check -- queue functions should prevent this happening */
875                 if (bio->bi_vcnt != 1 ||
876                     bio->bi_idx != 0)
877                         goto bad_map;
878                 /* This is a one page bio that upper layers
879                  * refuse to split for us, so we need to split it.
880                  */
881                 bp = bio_split(bio,
882                                chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
883
884                 /* Each of these 'make_request' calls will call 'wait_barrier'.
885                  * If the first succeeds but the second blocks due to the resync
886                  * thread raising the barrier, we will deadlock because the
887                  * IO to the underlying device will be queued in generic_make_request
888                  * and will never complete, so will never reduce nr_pending.
889                  * So increment nr_waiting here so no new raise_barriers will
890                  * succeed, and so the second wait_barrier cannot block.
891                  */
892                 spin_lock_irq(&conf->resync_lock);
893                 conf->nr_waiting++;
894                 spin_unlock_irq(&conf->resync_lock);
895
896                 make_request(mddev, &bp->bio1);
897                 make_request(mddev, &bp->bio2);
898
899                 spin_lock_irq(&conf->resync_lock);
900                 conf->nr_waiting--;
901                 wake_up(&conf->wait_barrier);
902                 spin_unlock_irq(&conf->resync_lock);
903
904                 bio_pair_release(bp);
905                 return;
906         bad_map:
907                 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
908                        " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
909                        (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
910
911                 bio_io_error(bio);
912                 return;
913         }
914
915         md_write_start(mddev, bio);
916
917         /*
918          * Register the new request and wait if the reconstruction
919          * thread has put up a bar for new requests.
920          * Continue immediately if no resync is active currently.
921          */
922         wait_barrier(conf);
923
924         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
925
926         r10_bio->master_bio = bio;
927         r10_bio->sectors = bio->bi_size >> 9;
928
929         r10_bio->mddev = mddev;
930         r10_bio->sector = bio->bi_sector;
931         r10_bio->state = 0;
932
933         /* We might need to issue multiple reads to different
934          * devices if there are bad blocks around, so we keep
935          * track of the number of reads in bio->bi_phys_segments.
936          * If this is 0, there is only one r10_bio and no locking
937          * will be needed when the request completes.  If it is
938          * non-zero, then it is the number of not-completed requests.
939          */
940         bio->bi_phys_segments = 0;
941         clear_bit(BIO_SEG_VALID, &bio->bi_flags);
942
943         if (rw == READ) {
944                 /*
945                  * read balancing logic:
946                  */
947                 int disk;
948                 int slot;
949
950 read_again:
951                 disk = read_balance(conf, r10_bio, &max_sectors);
952                 slot = r10_bio->read_slot;
953                 if (disk < 0) {
954                         raid_end_bio_io(r10_bio);
955                         return;
956                 }
957                 mirror = conf->mirrors + disk;
958
959                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
960                 md_trim_bio(read_bio, r10_bio->sector - bio->bi_sector,
961                             max_sectors);
962
963                 r10_bio->devs[slot].bio = read_bio;
964
965                 read_bio->bi_sector = r10_bio->devs[slot].addr +
966                         mirror->rdev->data_offset;
967                 read_bio->bi_bdev = mirror->rdev->bdev;
968                 read_bio->bi_end_io = raid10_end_read_request;
969                 read_bio->bi_rw = READ | do_sync;
970                 read_bio->bi_private = r10_bio;
971
972                 if (max_sectors < r10_bio->sectors) {
973                         /* Could not read all from this device, so we will
974                          * need another r10_bio.
975                          */
976                         sectors_handled = (r10_bio->sectors + max_sectors
977                                            - bio->bi_sector);
978                         r10_bio->sectors = max_sectors;
979                         spin_lock_irq(&conf->device_lock);
980                         if (bio->bi_phys_segments == 0)
981                                 bio->bi_phys_segments = 2;
982                         else
983                                 bio->bi_phys_segments++;
984                         spin_unlock(&conf->device_lock);
985                         /* Cannot call generic_make_request directly
986                          * as that will be queued in __generic_make_request
987                          * and subsequent mempool_alloc might block
988                          * waiting for it.  so hand bio over to raid10d.
989                          */
990                         reschedule_retry(r10_bio);
991
992                         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
993
994                         r10_bio->master_bio = bio;
995                         r10_bio->sectors = ((bio->bi_size >> 9)
996                                             - sectors_handled);
997                         r10_bio->state = 0;
998                         r10_bio->mddev = mddev;
999                         r10_bio->sector = bio->bi_sector + sectors_handled;
1000                         goto read_again;
1001                 } else
1002                         generic_make_request(read_bio);
1003                 return;
1004         }
1005
1006         /*
1007          * WRITE:
1008          */
1009         if (conf->pending_count >= max_queued_requests) {
1010                 md_wakeup_thread(mddev->thread);
1011                 wait_event(conf->wait_barrier,
1012                            conf->pending_count < max_queued_requests);
1013         }
1014         /* first select target devices under rcu_lock and
1015          * inc refcount on their rdev.  Record them by setting
1016          * bios[x] to bio
1017          * If there are known/acknowledged bad blocks on any device
1018          * on which we have seen a write error, we want to avoid
1019          * writing to those blocks.  This potentially requires several
1020          * writes to write around the bad blocks.  Each set of writes
1021          * gets its own r10_bio with a set of bios attached.  The number
1022          * of r10_bios is recored in bio->bi_phys_segments just as with
1023          * the read case.
1024          */
1025         plugged = mddev_check_plugged(mddev);
1026
1027         raid10_find_phys(conf, r10_bio);
1028 retry_write:
1029         blocked_rdev = NULL;
1030         rcu_read_lock();
1031         max_sectors = r10_bio->sectors;
1032
1033         for (i = 0;  i < conf->copies; i++) {
1034                 int d = r10_bio->devs[i].devnum;
1035                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1036                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1037                         atomic_inc(&rdev->nr_pending);
1038                         blocked_rdev = rdev;
1039                         break;
1040                 }
1041                 r10_bio->devs[i].bio = NULL;
1042                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1043                         set_bit(R10BIO_Degraded, &r10_bio->state);
1044                         continue;
1045                 }
1046                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1047                         sector_t first_bad;
1048                         sector_t dev_sector = r10_bio->devs[i].addr;
1049                         int bad_sectors;
1050                         int is_bad;
1051
1052                         is_bad = is_badblock(rdev, dev_sector,
1053                                              max_sectors,
1054                                              &first_bad, &bad_sectors);
1055                         if (is_bad < 0) {
1056                                 /* Mustn't write here until the bad block
1057                                  * is acknowledged
1058                                  */
1059                                 atomic_inc(&rdev->nr_pending);
1060                                 set_bit(BlockedBadBlocks, &rdev->flags);
1061                                 blocked_rdev = rdev;
1062                                 break;
1063                         }
1064                         if (is_bad && first_bad <= dev_sector) {
1065                                 /* Cannot write here at all */
1066                                 bad_sectors -= (dev_sector - first_bad);
1067                                 if (bad_sectors < max_sectors)
1068                                         /* Mustn't write more than bad_sectors
1069                                          * to other devices yet
1070                                          */
1071                                         max_sectors = bad_sectors;
1072                                 /* We don't set R10BIO_Degraded as that
1073                                  * only applies if the disk is missing,
1074                                  * so it might be re-added, and we want to
1075                                  * know to recover this chunk.
1076                                  * In this case the device is here, and the
1077                                  * fact that this chunk is not in-sync is
1078                                  * recorded in the bad block log.
1079                                  */
1080                                 continue;
1081                         }
1082                         if (is_bad) {
1083                                 int good_sectors = first_bad - dev_sector;
1084                                 if (good_sectors < max_sectors)
1085                                         max_sectors = good_sectors;
1086                         }
1087                 }
1088                 r10_bio->devs[i].bio = bio;
1089                 atomic_inc(&rdev->nr_pending);
1090         }
1091         rcu_read_unlock();
1092
1093         if (unlikely(blocked_rdev)) {
1094                 /* Have to wait for this device to get unblocked, then retry */
1095                 int j;
1096                 int d;
1097
1098                 for (j = 0; j < i; j++)
1099                         if (r10_bio->devs[j].bio) {
1100                                 d = r10_bio->devs[j].devnum;
1101                                 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1102                         }
1103                 allow_barrier(conf);
1104                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1105                 wait_barrier(conf);
1106                 goto retry_write;
1107         }
1108
1109         if (max_sectors < r10_bio->sectors) {
1110                 /* We are splitting this into multiple parts, so
1111                  * we need to prepare for allocating another r10_bio.
1112                  */
1113                 r10_bio->sectors = max_sectors;
1114                 spin_lock_irq(&conf->device_lock);
1115                 if (bio->bi_phys_segments == 0)
1116                         bio->bi_phys_segments = 2;
1117                 else
1118                         bio->bi_phys_segments++;
1119                 spin_unlock_irq(&conf->device_lock);
1120         }
1121         sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector;
1122
1123         atomic_set(&r10_bio->remaining, 1);
1124         bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1125
1126         for (i = 0; i < conf->copies; i++) {
1127                 struct bio *mbio;
1128                 int d = r10_bio->devs[i].devnum;
1129                 if (!r10_bio->devs[i].bio)
1130                         continue;
1131
1132                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1133                 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1134                             max_sectors);
1135                 r10_bio->devs[i].bio = mbio;
1136
1137                 mbio->bi_sector = (r10_bio->devs[i].addr+
1138                                    conf->mirrors[d].rdev->data_offset);
1139                 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1140                 mbio->bi_end_io = raid10_end_write_request;
1141                 mbio->bi_rw = WRITE | do_sync | do_fua;
1142                 mbio->bi_private = r10_bio;
1143
1144                 atomic_inc(&r10_bio->remaining);
1145                 spin_lock_irqsave(&conf->device_lock, flags);
1146                 bio_list_add(&conf->pending_bio_list, mbio);
1147                 conf->pending_count++;
1148                 spin_unlock_irqrestore(&conf->device_lock, flags);
1149         }
1150
1151         /* Don't remove the bias on 'remaining' (one_write_done) until
1152          * after checking if we need to go around again.
1153          */
1154
1155         if (sectors_handled < (bio->bi_size >> 9)) {
1156                 one_write_done(r10_bio);
1157                 /* We need another r10_bio.  It has already been counted
1158                  * in bio->bi_phys_segments.
1159                  */
1160                 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1161
1162                 r10_bio->master_bio = bio;
1163                 r10_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1164
1165                 r10_bio->mddev = mddev;
1166                 r10_bio->sector = bio->bi_sector + sectors_handled;
1167                 r10_bio->state = 0;
1168                 goto retry_write;
1169         }
1170         one_write_done(r10_bio);
1171
1172         /* In case raid10d snuck in to freeze_array */
1173         wake_up(&conf->wait_barrier);
1174
1175         if (do_sync || !mddev->bitmap || !plugged)
1176                 md_wakeup_thread(mddev->thread);
1177 }
1178
1179 static void status(struct seq_file *seq, struct mddev *mddev)
1180 {
1181         struct r10conf *conf = mddev->private;
1182         int i;
1183
1184         if (conf->near_copies < conf->raid_disks)
1185                 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1186         if (conf->near_copies > 1)
1187                 seq_printf(seq, " %d near-copies", conf->near_copies);
1188         if (conf->far_copies > 1) {
1189                 if (conf->far_offset)
1190                         seq_printf(seq, " %d offset-copies", conf->far_copies);
1191                 else
1192                         seq_printf(seq, " %d far-copies", conf->far_copies);
1193         }
1194         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1195                                         conf->raid_disks - mddev->degraded);
1196         for (i = 0; i < conf->raid_disks; i++)
1197                 seq_printf(seq, "%s",
1198                               conf->mirrors[i].rdev &&
1199                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1200         seq_printf(seq, "]");
1201 }
1202
1203 /* check if there are enough drives for
1204  * every block to appear on atleast one.
1205  * Don't consider the device numbered 'ignore'
1206  * as we might be about to remove it.
1207  */
1208 static int enough(struct r10conf *conf, int ignore)
1209 {
1210         int first = 0;
1211
1212         do {
1213                 int n = conf->copies;
1214                 int cnt = 0;
1215                 while (n--) {
1216                         if (conf->mirrors[first].rdev &&
1217                             first != ignore)
1218                                 cnt++;
1219                         first = (first+1) % conf->raid_disks;
1220                 }
1221                 if (cnt == 0)
1222                         return 0;
1223         } while (first != 0);
1224         return 1;
1225 }
1226
1227 static void error(struct mddev *mddev, struct md_rdev *rdev)
1228 {
1229         char b[BDEVNAME_SIZE];
1230         struct r10conf *conf = mddev->private;
1231
1232         /*
1233          * If it is not operational, then we have already marked it as dead
1234          * else if it is the last working disks, ignore the error, let the
1235          * next level up know.
1236          * else mark the drive as failed
1237          */
1238         if (test_bit(In_sync, &rdev->flags)
1239             && !enough(conf, rdev->raid_disk))
1240                 /*
1241                  * Don't fail the drive, just return an IO error.
1242                  */
1243                 return;
1244         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1245                 unsigned long flags;
1246                 spin_lock_irqsave(&conf->device_lock, flags);
1247                 mddev->degraded++;
1248                 spin_unlock_irqrestore(&conf->device_lock, flags);
1249                 /*
1250                  * if recovery is running, make sure it aborts.
1251                  */
1252                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1253         }
1254         set_bit(Blocked, &rdev->flags);
1255         set_bit(Faulty, &rdev->flags);
1256         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1257         printk(KERN_ALERT
1258                "md/raid10:%s: Disk failure on %s, disabling device.\n"
1259                "md/raid10:%s: Operation continuing on %d devices.\n",
1260                mdname(mddev), bdevname(rdev->bdev, b),
1261                mdname(mddev), conf->raid_disks - mddev->degraded);
1262 }
1263
1264 static void print_conf(struct r10conf *conf)
1265 {
1266         int i;
1267         struct mirror_info *tmp;
1268
1269         printk(KERN_DEBUG "RAID10 conf printout:\n");
1270         if (!conf) {
1271                 printk(KERN_DEBUG "(!conf)\n");
1272                 return;
1273         }
1274         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1275                 conf->raid_disks);
1276
1277         for (i = 0; i < conf->raid_disks; i++) {
1278                 char b[BDEVNAME_SIZE];
1279                 tmp = conf->mirrors + i;
1280                 if (tmp->rdev)
1281                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1282                                 i, !test_bit(In_sync, &tmp->rdev->flags),
1283                                 !test_bit(Faulty, &tmp->rdev->flags),
1284                                 bdevname(tmp->rdev->bdev,b));
1285         }
1286 }
1287
1288 static void close_sync(struct r10conf *conf)
1289 {
1290         wait_barrier(conf);
1291         allow_barrier(conf);
1292
1293         mempool_destroy(conf->r10buf_pool);
1294         conf->r10buf_pool = NULL;
1295 }
1296
1297 static int raid10_spare_active(struct mddev *mddev)
1298 {
1299         int i;
1300         struct r10conf *conf = mddev->private;
1301         struct mirror_info *tmp;
1302         int count = 0;
1303         unsigned long flags;
1304
1305         /*
1306          * Find all non-in_sync disks within the RAID10 configuration
1307          * and mark them in_sync
1308          */
1309         for (i = 0; i < conf->raid_disks; i++) {
1310                 tmp = conf->mirrors + i;
1311                 if (tmp->rdev
1312                     && !test_bit(Faulty, &tmp->rdev->flags)
1313                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1314                         count++;
1315                         sysfs_notify_dirent(tmp->rdev->sysfs_state);
1316                 }
1317         }
1318         spin_lock_irqsave(&conf->device_lock, flags);
1319         mddev->degraded -= count;
1320         spin_unlock_irqrestore(&conf->device_lock, flags);
1321
1322         print_conf(conf);
1323         return count;
1324 }
1325
1326
1327 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1328 {
1329         struct r10conf *conf = mddev->private;
1330         int err = -EEXIST;
1331         int mirror;
1332         int first = 0;
1333         int last = conf->raid_disks - 1;
1334
1335         if (mddev->recovery_cp < MaxSector)
1336                 /* only hot-add to in-sync arrays, as recovery is
1337                  * very different from resync
1338                  */
1339                 return -EBUSY;
1340         if (!enough(conf, -1))
1341                 return -EINVAL;
1342
1343         if (rdev->raid_disk >= 0)
1344                 first = last = rdev->raid_disk;
1345
1346         if (rdev->saved_raid_disk >= first &&
1347             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1348                 mirror = rdev->saved_raid_disk;
1349         else
1350                 mirror = first;
1351         for ( ; mirror <= last ; mirror++) {
1352                 struct mirror_info *p = &conf->mirrors[mirror];
1353                 if (p->recovery_disabled == mddev->recovery_disabled)
1354                         continue;
1355                 if (p->rdev)
1356                         continue;
1357
1358                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1359                                   rdev->data_offset << 9);
1360                 /* as we don't honour merge_bvec_fn, we must
1361                  * never risk violating it, so limit
1362                  * ->max_segments to one lying with a single
1363                  * page, as a one page request is never in
1364                  * violation.
1365                  */
1366                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1367                         blk_queue_max_segments(mddev->queue, 1);
1368                         blk_queue_segment_boundary(mddev->queue,
1369                                                    PAGE_CACHE_SIZE - 1);
1370                 }
1371
1372                 p->head_position = 0;
1373                 p->recovery_disabled = mddev->recovery_disabled - 1;
1374                 rdev->raid_disk = mirror;
1375                 err = 0;
1376                 if (rdev->saved_raid_disk != mirror)
1377                         conf->fullsync = 1;
1378                 rcu_assign_pointer(p->rdev, rdev);
1379                 break;
1380         }
1381
1382         md_integrity_add_rdev(rdev, mddev);
1383         print_conf(conf);
1384         return err;
1385 }
1386
1387 static int raid10_remove_disk(struct mddev *mddev, int number)
1388 {
1389         struct r10conf *conf = mddev->private;
1390         int err = 0;
1391         struct md_rdev *rdev;
1392         struct mirror_info *p = conf->mirrors+ number;
1393
1394         print_conf(conf);
1395         rdev = p->rdev;
1396         if (rdev) {
1397                 if (test_bit(In_sync, &rdev->flags) ||
1398                     atomic_read(&rdev->nr_pending)) {
1399                         err = -EBUSY;
1400                         goto abort;
1401                 }
1402                 /* Only remove faulty devices in recovery
1403                  * is not possible.
1404                  */
1405                 if (!test_bit(Faulty, &rdev->flags) &&
1406                     mddev->recovery_disabled != p->recovery_disabled &&
1407                     enough(conf, -1)) {
1408                         err = -EBUSY;
1409                         goto abort;
1410                 }
1411                 p->rdev = NULL;
1412                 synchronize_rcu();
1413                 if (atomic_read(&rdev->nr_pending)) {
1414                         /* lost the race, try later */
1415                         err = -EBUSY;
1416                         p->rdev = rdev;
1417                         goto abort;
1418                 }
1419                 err = md_integrity_register(mddev);
1420         }
1421 abort:
1422
1423         print_conf(conf);
1424         return err;
1425 }
1426
1427
1428 static void end_sync_read(struct bio *bio, int error)
1429 {
1430         struct r10bio *r10_bio = bio->bi_private;
1431         struct r10conf *conf = r10_bio->mddev->private;
1432         int d;
1433
1434         d = find_bio_disk(conf, r10_bio, bio, NULL);
1435
1436         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1437                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1438         else
1439                 /* The write handler will notice the lack of
1440                  * R10BIO_Uptodate and record any errors etc
1441                  */
1442                 atomic_add(r10_bio->sectors,
1443                            &conf->mirrors[d].rdev->corrected_errors);
1444
1445         /* for reconstruct, we always reschedule after a read.
1446          * for resync, only after all reads
1447          */
1448         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1449         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1450             atomic_dec_and_test(&r10_bio->remaining)) {
1451                 /* we have read all the blocks,
1452                  * do the comparison in process context in raid10d
1453                  */
1454                 reschedule_retry(r10_bio);
1455         }
1456 }
1457
1458 static void end_sync_request(struct r10bio *r10_bio)
1459 {
1460         struct mddev *mddev = r10_bio->mddev;
1461
1462         while (atomic_dec_and_test(&r10_bio->remaining)) {
1463                 if (r10_bio->master_bio == NULL) {
1464                         /* the primary of several recovery bios */
1465                         sector_t s = r10_bio->sectors;
1466                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1467                             test_bit(R10BIO_WriteError, &r10_bio->state))
1468                                 reschedule_retry(r10_bio);
1469                         else
1470                                 put_buf(r10_bio);
1471                         md_done_sync(mddev, s, 1);
1472                         break;
1473                 } else {
1474                         struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1475                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1476                             test_bit(R10BIO_WriteError, &r10_bio->state))
1477                                 reschedule_retry(r10_bio);
1478                         else
1479                                 put_buf(r10_bio);
1480                         r10_bio = r10_bio2;
1481                 }
1482         }
1483 }
1484
1485 static void end_sync_write(struct bio *bio, int error)
1486 {
1487         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1488         struct r10bio *r10_bio = bio->bi_private;
1489         struct mddev *mddev = r10_bio->mddev;
1490         struct r10conf *conf = mddev->private;
1491         int d;
1492         sector_t first_bad;
1493         int bad_sectors;
1494         int slot;
1495
1496         d = find_bio_disk(conf, r10_bio, bio, &slot);
1497
1498         if (!uptodate) {
1499                 set_bit(WriteErrorSeen, &conf->mirrors[d].rdev->flags);
1500                 set_bit(R10BIO_WriteError, &r10_bio->state);
1501         } else if (is_badblock(conf->mirrors[d].rdev,
1502                              r10_bio->devs[slot].addr,
1503                              r10_bio->sectors,
1504                              &first_bad, &bad_sectors))
1505                 set_bit(R10BIO_MadeGood, &r10_bio->state);
1506
1507         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1508
1509         end_sync_request(r10_bio);
1510 }
1511
1512 /*
1513  * Note: sync and recover and handled very differently for raid10
1514  * This code is for resync.
1515  * For resync, we read through virtual addresses and read all blocks.
1516  * If there is any error, we schedule a write.  The lowest numbered
1517  * drive is authoritative.
1518  * However requests come for physical address, so we need to map.
1519  * For every physical address there are raid_disks/copies virtual addresses,
1520  * which is always are least one, but is not necessarly an integer.
1521  * This means that a physical address can span multiple chunks, so we may
1522  * have to submit multiple io requests for a single sync request.
1523  */
1524 /*
1525  * We check if all blocks are in-sync and only write to blocks that
1526  * aren't in sync
1527  */
1528 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1529 {
1530         struct r10conf *conf = mddev->private;
1531         int i, first;
1532         struct bio *tbio, *fbio;
1533
1534         atomic_set(&r10_bio->remaining, 1);
1535
1536         /* find the first device with a block */
1537         for (i=0; i<conf->copies; i++)
1538                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1539                         break;
1540
1541         if (i == conf->copies)
1542                 goto done;
1543
1544         first = i;
1545         fbio = r10_bio->devs[i].bio;
1546
1547         /* now find blocks with errors */
1548         for (i=0 ; i < conf->copies ; i++) {
1549                 int  j, d;
1550                 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1551
1552                 tbio = r10_bio->devs[i].bio;
1553
1554                 if (tbio->bi_end_io != end_sync_read)
1555                         continue;
1556                 if (i == first)
1557                         continue;
1558                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1559                         /* We know that the bi_io_vec layout is the same for
1560                          * both 'first' and 'i', so we just compare them.
1561                          * All vec entries are PAGE_SIZE;
1562                          */
1563                         for (j = 0; j < vcnt; j++)
1564                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1565                                            page_address(tbio->bi_io_vec[j].bv_page),
1566                                            PAGE_SIZE))
1567                                         break;
1568                         if (j == vcnt)
1569                                 continue;
1570                         mddev->resync_mismatches += r10_bio->sectors;
1571                         if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1572                                 /* Don't fix anything. */
1573                                 continue;
1574                 }
1575                 /* Ok, we need to write this bio, either to correct an
1576                  * inconsistency or to correct an unreadable block.
1577                  * First we need to fixup bv_offset, bv_len and
1578                  * bi_vecs, as the read request might have corrupted these
1579                  */
1580                 tbio->bi_vcnt = vcnt;
1581                 tbio->bi_size = r10_bio->sectors << 9;
1582                 tbio->bi_idx = 0;
1583                 tbio->bi_phys_segments = 0;
1584                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1585                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1586                 tbio->bi_next = NULL;
1587                 tbio->bi_rw = WRITE;
1588                 tbio->bi_private = r10_bio;
1589                 tbio->bi_sector = r10_bio->devs[i].addr;
1590
1591                 for (j=0; j < vcnt ; j++) {
1592                         tbio->bi_io_vec[j].bv_offset = 0;
1593                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1594
1595                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1596                                page_address(fbio->bi_io_vec[j].bv_page),
1597                                PAGE_SIZE);
1598                 }
1599                 tbio->bi_end_io = end_sync_write;
1600
1601                 d = r10_bio->devs[i].devnum;
1602                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1603                 atomic_inc(&r10_bio->remaining);
1604                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1605
1606                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1607                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1608                 generic_make_request(tbio);
1609         }
1610
1611 done:
1612         if (atomic_dec_and_test(&r10_bio->remaining)) {
1613                 md_done_sync(mddev, r10_bio->sectors, 1);
1614                 put_buf(r10_bio);
1615         }
1616 }
1617
1618 /*
1619  * Now for the recovery code.
1620  * Recovery happens across physical sectors.
1621  * We recover all non-is_sync drives by finding the virtual address of
1622  * each, and then choose a working drive that also has that virt address.
1623  * There is a separate r10_bio for each non-in_sync drive.
1624  * Only the first two slots are in use. The first for reading,
1625  * The second for writing.
1626  *
1627  */
1628 static void fix_recovery_read_error(struct r10bio *r10_bio)
1629 {
1630         /* We got a read error during recovery.
1631          * We repeat the read in smaller page-sized sections.
1632          * If a read succeeds, write it to the new device or record
1633          * a bad block if we cannot.
1634          * If a read fails, record a bad block on both old and
1635          * new devices.
1636          */
1637         struct mddev *mddev = r10_bio->mddev;
1638         struct r10conf *conf = mddev->private;
1639         struct bio *bio = r10_bio->devs[0].bio;
1640         sector_t sect = 0;
1641         int sectors = r10_bio->sectors;
1642         int idx = 0;
1643         int dr = r10_bio->devs[0].devnum;
1644         int dw = r10_bio->devs[1].devnum;
1645
1646         while (sectors) {
1647                 int s = sectors;
1648                 struct md_rdev *rdev;
1649                 sector_t addr;
1650                 int ok;
1651
1652                 if (s > (PAGE_SIZE>>9))
1653                         s = PAGE_SIZE >> 9;
1654
1655                 rdev = conf->mirrors[dr].rdev;
1656                 addr = r10_bio->devs[0].addr + sect,
1657                 ok = sync_page_io(rdev,
1658                                   addr,
1659                                   s << 9,
1660                                   bio->bi_io_vec[idx].bv_page,
1661                                   READ, false);
1662                 if (ok) {
1663                         rdev = conf->mirrors[dw].rdev;
1664                         addr = r10_bio->devs[1].addr + sect;
1665                         ok = sync_page_io(rdev,
1666                                           addr,
1667                                           s << 9,
1668                                           bio->bi_io_vec[idx].bv_page,
1669                                           WRITE, false);
1670                         if (!ok)
1671                                 set_bit(WriteErrorSeen, &rdev->flags);
1672                 }
1673                 if (!ok) {
1674                         /* We don't worry if we cannot set a bad block -
1675                          * it really is bad so there is no loss in not
1676                          * recording it yet
1677                          */
1678                         rdev_set_badblocks(rdev, addr, s, 0);
1679
1680                         if (rdev != conf->mirrors[dw].rdev) {
1681                                 /* need bad block on destination too */
1682                                 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
1683                                 addr = r10_bio->devs[1].addr + sect;
1684                                 ok = rdev_set_badblocks(rdev2, addr, s, 0);
1685                                 if (!ok) {
1686                                         /* just abort the recovery */
1687                                         printk(KERN_NOTICE
1688                                                "md/raid10:%s: recovery aborted"
1689                                                " due to read error\n",
1690                                                mdname(mddev));
1691
1692                                         conf->mirrors[dw].recovery_disabled
1693                                                 = mddev->recovery_disabled;
1694                                         set_bit(MD_RECOVERY_INTR,
1695                                                 &mddev->recovery);
1696                                         break;
1697                                 }
1698                         }
1699                 }
1700
1701                 sectors -= s;
1702                 sect += s;
1703                 idx++;
1704         }
1705 }
1706
1707 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1708 {
1709         struct r10conf *conf = mddev->private;
1710         int d;
1711         struct bio *wbio;
1712
1713         if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
1714                 fix_recovery_read_error(r10_bio);
1715                 end_sync_request(r10_bio);
1716                 return;
1717         }
1718
1719         /*
1720          * share the pages with the first bio
1721          * and submit the write request
1722          */
1723         wbio = r10_bio->devs[1].bio;
1724         d = r10_bio->devs[1].devnum;
1725
1726         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1727         md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1728         generic_make_request(wbio);
1729 }
1730
1731
1732 /*
1733  * Used by fix_read_error() to decay the per rdev read_errors.
1734  * We halve the read error count for every hour that has elapsed
1735  * since the last recorded read error.
1736  *
1737  */
1738 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
1739 {
1740         struct timespec cur_time_mon;
1741         unsigned long hours_since_last;
1742         unsigned int read_errors = atomic_read(&rdev->read_errors);
1743
1744         ktime_get_ts(&cur_time_mon);
1745
1746         if (rdev->last_read_error.tv_sec == 0 &&
1747             rdev->last_read_error.tv_nsec == 0) {
1748                 /* first time we've seen a read error */
1749                 rdev->last_read_error = cur_time_mon;
1750                 return;
1751         }
1752
1753         hours_since_last = (cur_time_mon.tv_sec -
1754                             rdev->last_read_error.tv_sec) / 3600;
1755
1756         rdev->last_read_error = cur_time_mon;
1757
1758         /*
1759          * if hours_since_last is > the number of bits in read_errors
1760          * just set read errors to 0. We do this to avoid
1761          * overflowing the shift of read_errors by hours_since_last.
1762          */
1763         if (hours_since_last >= 8 * sizeof(read_errors))
1764                 atomic_set(&rdev->read_errors, 0);
1765         else
1766                 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
1767 }
1768
1769 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
1770                             int sectors, struct page *page, int rw)
1771 {
1772         sector_t first_bad;
1773         int bad_sectors;
1774
1775         if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
1776             && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
1777                 return -1;
1778         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1779                 /* success */
1780                 return 1;
1781         if (rw == WRITE)
1782                 set_bit(WriteErrorSeen, &rdev->flags);
1783         /* need to record an error - either for the block or the device */
1784         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1785                 md_error(rdev->mddev, rdev);
1786         return 0;
1787 }
1788
1789 /*
1790  * This is a kernel thread which:
1791  *
1792  *      1.      Retries failed read operations on working mirrors.
1793  *      2.      Updates the raid superblock when problems encounter.
1794  *      3.      Performs writes following reads for array synchronising.
1795  */
1796
1797 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
1798 {
1799         int sect = 0; /* Offset from r10_bio->sector */
1800         int sectors = r10_bio->sectors;
1801         struct md_rdev*rdev;
1802         int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
1803         int d = r10_bio->devs[r10_bio->read_slot].devnum;
1804
1805         /* still own a reference to this rdev, so it cannot
1806          * have been cleared recently.
1807          */
1808         rdev = conf->mirrors[d].rdev;
1809
1810         if (test_bit(Faulty, &rdev->flags))
1811                 /* drive has already been failed, just ignore any
1812                    more fix_read_error() attempts */
1813                 return;
1814
1815         check_decay_read_errors(mddev, rdev);
1816         atomic_inc(&rdev->read_errors);
1817         if (atomic_read(&rdev->read_errors) > max_read_errors) {
1818                 char b[BDEVNAME_SIZE];
1819                 bdevname(rdev->bdev, b);
1820
1821                 printk(KERN_NOTICE
1822                        "md/raid10:%s: %s: Raid device exceeded "
1823                        "read_error threshold [cur %d:max %d]\n",
1824                        mdname(mddev), b,
1825                        atomic_read(&rdev->read_errors), max_read_errors);
1826                 printk(KERN_NOTICE
1827                        "md/raid10:%s: %s: Failing raid device\n",
1828                        mdname(mddev), b);
1829                 md_error(mddev, conf->mirrors[d].rdev);
1830                 return;
1831         }
1832
1833         while(sectors) {
1834                 int s = sectors;
1835                 int sl = r10_bio->read_slot;
1836                 int success = 0;
1837                 int start;
1838
1839                 if (s > (PAGE_SIZE>>9))
1840                         s = PAGE_SIZE >> 9;
1841
1842                 rcu_read_lock();
1843                 do {
1844                         sector_t first_bad;
1845                         int bad_sectors;
1846
1847                         d = r10_bio->devs[sl].devnum;
1848                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1849                         if (rdev &&
1850                             test_bit(In_sync, &rdev->flags) &&
1851                             is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
1852                                         &first_bad, &bad_sectors) == 0) {
1853                                 atomic_inc(&rdev->nr_pending);
1854                                 rcu_read_unlock();
1855                                 success = sync_page_io(rdev,
1856                                                        r10_bio->devs[sl].addr +
1857                                                        sect,
1858                                                        s<<9,
1859                                                        conf->tmppage, READ, false);
1860                                 rdev_dec_pending(rdev, mddev);
1861                                 rcu_read_lock();
1862                                 if (success)
1863                                         break;
1864                         }
1865                         sl++;
1866                         if (sl == conf->copies)
1867                                 sl = 0;
1868                 } while (!success && sl != r10_bio->read_slot);
1869                 rcu_read_unlock();
1870
1871                 if (!success) {
1872                         /* Cannot read from anywhere, just mark the block
1873                          * as bad on the first device to discourage future
1874                          * reads.
1875                          */
1876                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1877                         rdev = conf->mirrors[dn].rdev;
1878
1879                         if (!rdev_set_badblocks(
1880                                     rdev,
1881                                     r10_bio->devs[r10_bio->read_slot].addr
1882                                     + sect,
1883                                     s, 0))
1884                                 md_error(mddev, rdev);
1885                         break;
1886                 }
1887
1888                 start = sl;
1889                 /* write it back and re-read */
1890                 rcu_read_lock();
1891                 while (sl != r10_bio->read_slot) {
1892                         char b[BDEVNAME_SIZE];
1893
1894                         if (sl==0)
1895                                 sl = conf->copies;
1896                         sl--;
1897                         d = r10_bio->devs[sl].devnum;
1898                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1899                         if (!rdev ||
1900                             !test_bit(In_sync, &rdev->flags))
1901                                 continue;
1902
1903                         atomic_inc(&rdev->nr_pending);
1904                         rcu_read_unlock();
1905                         if (r10_sync_page_io(rdev,
1906                                              r10_bio->devs[sl].addr +
1907                                              sect,
1908                                              s<<9, conf->tmppage, WRITE)
1909                             == 0) {
1910                                 /* Well, this device is dead */
1911                                 printk(KERN_NOTICE
1912                                        "md/raid10:%s: read correction "
1913                                        "write failed"
1914                                        " (%d sectors at %llu on %s)\n",
1915                                        mdname(mddev), s,
1916                                        (unsigned long long)(
1917                                                sect + rdev->data_offset),
1918                                        bdevname(rdev->bdev, b));
1919                                 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
1920                                        "drive\n",
1921                                        mdname(mddev),
1922                                        bdevname(rdev->bdev, b));
1923                         }
1924                         rdev_dec_pending(rdev, mddev);
1925                         rcu_read_lock();
1926                 }
1927                 sl = start;
1928                 while (sl != r10_bio->read_slot) {
1929                         char b[BDEVNAME_SIZE];
1930
1931                         if (sl==0)
1932                                 sl = conf->copies;
1933                         sl--;
1934                         d = r10_bio->devs[sl].devnum;
1935                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1936                         if (!rdev ||
1937                             !test_bit(In_sync, &rdev->flags))
1938                                 continue;
1939
1940                         atomic_inc(&rdev->nr_pending);
1941                         rcu_read_unlock();
1942                         switch (r10_sync_page_io(rdev,
1943                                              r10_bio->devs[sl].addr +
1944                                              sect,
1945                                              s<<9, conf->tmppage,
1946                                                  READ)) {
1947                         case 0:
1948                                 /* Well, this device is dead */
1949                                 printk(KERN_NOTICE
1950                                        "md/raid10:%s: unable to read back "
1951                                        "corrected sectors"
1952                                        " (%d sectors at %llu on %s)\n",
1953                                        mdname(mddev), s,
1954                                        (unsigned long long)(
1955                                                sect + rdev->data_offset),
1956                                        bdevname(rdev->bdev, b));
1957                                 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
1958                                        "drive\n",
1959                                        mdname(mddev),
1960                                        bdevname(rdev->bdev, b));
1961                                 break;
1962                         case 1:
1963                                 printk(KERN_INFO
1964                                        "md/raid10:%s: read error corrected"
1965                                        " (%d sectors at %llu on %s)\n",
1966                                        mdname(mddev), s,
1967                                        (unsigned long long)(
1968                                                sect + rdev->data_offset),
1969                                        bdevname(rdev->bdev, b));
1970                                 atomic_add(s, &rdev->corrected_errors);
1971                         }
1972
1973                         rdev_dec_pending(rdev, mddev);
1974                         rcu_read_lock();
1975                 }
1976                 rcu_read_unlock();
1977
1978                 sectors -= s;
1979                 sect += s;
1980         }
1981 }
1982
1983 static void bi_complete(struct bio *bio, int error)
1984 {
1985         complete((struct completion *)bio->bi_private);
1986 }
1987
1988 static int submit_bio_wait(int rw, struct bio *bio)
1989 {
1990         struct completion event;
1991         rw |= REQ_SYNC;
1992
1993         init_completion(&event);
1994         bio->bi_private = &event;
1995         bio->bi_end_io = bi_complete;
1996         submit_bio(rw, bio);
1997         wait_for_completion(&event);
1998
1999         return test_bit(BIO_UPTODATE, &bio->bi_flags);
2000 }
2001
2002 static int narrow_write_error(struct r10bio *r10_bio, int i)
2003 {
2004         struct bio *bio = r10_bio->master_bio;
2005         struct mddev *mddev = r10_bio->mddev;
2006         struct r10conf *conf = mddev->private;
2007         struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2008         /* bio has the data to be written to slot 'i' where
2009          * we just recently had a write error.
2010          * We repeatedly clone the bio and trim down to one block,
2011          * then try the write.  Where the write fails we record
2012          * a bad block.
2013          * It is conceivable that the bio doesn't exactly align with
2014          * blocks.  We must handle this.
2015          *
2016          * We currently own a reference to the rdev.
2017          */
2018
2019         int block_sectors;
2020         sector_t sector;
2021         int sectors;
2022         int sect_to_write = r10_bio->sectors;
2023         int ok = 1;
2024
2025         if (rdev->badblocks.shift < 0)
2026                 return 0;
2027
2028         block_sectors = 1 << rdev->badblocks.shift;
2029         sector = r10_bio->sector;
2030         sectors = ((r10_bio->sector + block_sectors)
2031                    & ~(sector_t)(block_sectors - 1))
2032                 - sector;
2033
2034         while (sect_to_write) {
2035                 struct bio *wbio;
2036                 if (sectors > sect_to_write)
2037                         sectors = sect_to_write;
2038                 /* Write at 'sector' for 'sectors' */
2039                 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2040                 md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2041                 wbio->bi_sector = (r10_bio->devs[i].addr+
2042                                    rdev->data_offset+
2043                                    (sector - r10_bio->sector));
2044                 wbio->bi_bdev = rdev->bdev;
2045                 if (submit_bio_wait(WRITE, wbio) == 0)
2046                         /* Failure! */
2047                         ok = rdev_set_badblocks(rdev, sector,
2048                                                 sectors, 0)
2049                                 && ok;
2050
2051                 bio_put(wbio);
2052                 sect_to_write -= sectors;
2053                 sector += sectors;
2054                 sectors = block_sectors;
2055         }
2056         return ok;
2057 }
2058
2059 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2060 {
2061         int slot = r10_bio->read_slot;
2062         int mirror = r10_bio->devs[slot].devnum;
2063         struct bio *bio;
2064         struct r10conf *conf = mddev->private;
2065         struct md_rdev *rdev;
2066         char b[BDEVNAME_SIZE];
2067         unsigned long do_sync;
2068         int max_sectors;
2069
2070         /* we got a read error. Maybe the drive is bad.  Maybe just
2071          * the block and we can fix it.
2072          * We freeze all other IO, and try reading the block from
2073          * other devices.  When we find one, we re-write
2074          * and check it that fixes the read error.
2075          * This is all done synchronously while the array is
2076          * frozen.
2077          */
2078         if (mddev->ro == 0) {
2079                 freeze_array(conf);
2080                 fix_read_error(conf, mddev, r10_bio);
2081                 unfreeze_array(conf);
2082         }
2083         rdev_dec_pending(conf->mirrors[mirror].rdev, mddev);
2084
2085         bio = r10_bio->devs[slot].bio;
2086         bdevname(bio->bi_bdev, b);
2087         r10_bio->devs[slot].bio =
2088                 mddev->ro ? IO_BLOCKED : NULL;
2089 read_more:
2090         mirror = read_balance(conf, r10_bio, &max_sectors);
2091         if (mirror == -1) {
2092                 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2093                        " read error for block %llu\n",
2094                        mdname(mddev), b,
2095                        (unsigned long long)r10_bio->sector);
2096                 raid_end_bio_io(r10_bio);
2097                 bio_put(bio);
2098                 return;
2099         }
2100
2101         do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2102         if (bio)
2103                 bio_put(bio);
2104         slot = r10_bio->read_slot;
2105         rdev = conf->mirrors[mirror].rdev;
2106         printk_ratelimited(
2107                 KERN_ERR
2108                 "md/raid10:%s: %s: redirecting"
2109                 "sector %llu to another mirror\n",
2110                 mdname(mddev),
2111                 bdevname(rdev->bdev, b),
2112                 (unsigned long long)r10_bio->sector);
2113         bio = bio_clone_mddev(r10_bio->master_bio,
2114                               GFP_NOIO, mddev);
2115         md_trim_bio(bio,
2116                     r10_bio->sector - bio->bi_sector,
2117                     max_sectors);
2118         r10_bio->devs[slot].bio = bio;
2119         bio->bi_sector = r10_bio->devs[slot].addr
2120                 + rdev->data_offset;
2121         bio->bi_bdev = rdev->bdev;
2122         bio->bi_rw = READ | do_sync;
2123         bio->bi_private = r10_bio;
2124         bio->bi_end_io = raid10_end_read_request;
2125         if (max_sectors < r10_bio->sectors) {
2126                 /* Drat - have to split this up more */
2127                 struct bio *mbio = r10_bio->master_bio;
2128                 int sectors_handled =
2129                         r10_bio->sector + max_sectors
2130                         - mbio->bi_sector;
2131                 r10_bio->sectors = max_sectors;
2132                 spin_lock_irq(&conf->device_lock);
2133                 if (mbio->bi_phys_segments == 0)
2134                         mbio->bi_phys_segments = 2;
2135                 else
2136                         mbio->bi_phys_segments++;
2137                 spin_unlock_irq(&conf->device_lock);
2138                 generic_make_request(bio);
2139                 bio = NULL;
2140
2141                 r10_bio = mempool_alloc(conf->r10bio_pool,
2142                                         GFP_NOIO);
2143                 r10_bio->master_bio = mbio;
2144                 r10_bio->sectors = (mbio->bi_size >> 9)
2145                         - sectors_handled;
2146                 r10_bio->state = 0;
2147                 set_bit(R10BIO_ReadError,
2148                         &r10_bio->state);
2149                 r10_bio->mddev = mddev;
2150                 r10_bio->sector = mbio->bi_sector
2151                         + sectors_handled;
2152
2153                 goto read_more;
2154         } else
2155                 generic_make_request(bio);
2156 }
2157
2158 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2159 {
2160         /* Some sort of write request has finished and it
2161          * succeeded in writing where we thought there was a
2162          * bad block.  So forget the bad block.
2163          * Or possibly if failed and we need to record
2164          * a bad block.
2165          */
2166         int m;
2167         struct md_rdev *rdev;
2168
2169         if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2170             test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2171                 for (m = 0; m < conf->copies; m++) {
2172                         int dev = r10_bio->devs[m].devnum;
2173                         rdev = conf->mirrors[dev].rdev;
2174                         if (r10_bio->devs[m].bio == NULL)
2175                                 continue;
2176                         if (test_bit(BIO_UPTODATE,
2177                                      &r10_bio->devs[m].bio->bi_flags)) {
2178                                 rdev_clear_badblocks(
2179                                         rdev,
2180                                         r10_bio->devs[m].addr,
2181                                         r10_bio->sectors);
2182                         } else {
2183                                 if (!rdev_set_badblocks(
2184                                             rdev,
2185                                             r10_bio->devs[m].addr,
2186                                             r10_bio->sectors, 0))
2187                                         md_error(conf->mddev, rdev);
2188                         }
2189                 }
2190                 put_buf(r10_bio);
2191         } else {
2192                 for (m = 0; m < conf->copies; m++) {
2193                         int dev = r10_bio->devs[m].devnum;
2194                         struct bio *bio = r10_bio->devs[m].bio;
2195                         rdev = conf->mirrors[dev].rdev;
2196                         if (bio == IO_MADE_GOOD) {
2197                                 rdev_clear_badblocks(
2198                                         rdev,
2199                                         r10_bio->devs[m].addr,
2200                                         r10_bio->sectors);
2201                                 rdev_dec_pending(rdev, conf->mddev);
2202                         } else if (bio != NULL &&
2203                                    !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2204                                 if (!narrow_write_error(r10_bio, m)) {
2205                                         md_error(conf->mddev, rdev);
2206                                         set_bit(R10BIO_Degraded,
2207                                                 &r10_bio->state);
2208                                 }
2209                                 rdev_dec_pending(rdev, conf->mddev);
2210                         }
2211                 }
2212                 if (test_bit(R10BIO_WriteError,
2213                              &r10_bio->state))
2214                         close_write(r10_bio);
2215                 raid_end_bio_io(r10_bio);
2216         }
2217 }
2218
2219 static void raid10d(struct mddev *mddev)
2220 {
2221         struct r10bio *r10_bio;
2222         unsigned long flags;
2223         struct r10conf *conf = mddev->private;
2224         struct list_head *head = &conf->retry_list;
2225         struct blk_plug plug;
2226
2227         md_check_recovery(mddev);
2228
2229         blk_start_plug(&plug);
2230         for (;;) {
2231
2232                 flush_pending_writes(conf);
2233
2234                 spin_lock_irqsave(&conf->device_lock, flags);
2235                 if (list_empty(head)) {
2236                         spin_unlock_irqrestore(&conf->device_lock, flags);
2237                         break;
2238                 }
2239                 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2240                 list_del(head->prev);
2241                 conf->nr_queued--;
2242                 spin_unlock_irqrestore(&conf->device_lock, flags);
2243
2244                 mddev = r10_bio->mddev;
2245                 conf = mddev->private;
2246                 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2247                     test_bit(R10BIO_WriteError, &r10_bio->state))
2248                         handle_write_completed(conf, r10_bio);
2249                 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2250                         sync_request_write(mddev, r10_bio);
2251                 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2252                         recovery_request_write(mddev, r10_bio);
2253                 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2254                         handle_read_error(mddev, r10_bio);
2255                 else {
2256                         /* just a partial read to be scheduled from a
2257                          * separate context
2258                          */
2259                         int slot = r10_bio->read_slot;
2260                         generic_make_request(r10_bio->devs[slot].bio);
2261                 }
2262
2263                 cond_resched();
2264                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2265                         md_check_recovery(mddev);
2266         }
2267         blk_finish_plug(&plug);
2268 }
2269
2270
2271 static int init_resync(struct r10conf *conf)
2272 {
2273         int buffs;
2274
2275         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2276         BUG_ON(conf->r10buf_pool);
2277         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2278         if (!conf->r10buf_pool)
2279                 return -ENOMEM;
2280         conf->next_resync = 0;
2281         return 0;
2282 }
2283
2284 /*
2285  * perform a "sync" on one "block"
2286  *
2287  * We need to make sure that no normal I/O request - particularly write
2288  * requests - conflict with active sync requests.
2289  *
2290  * This is achieved by tracking pending requests and a 'barrier' concept
2291  * that can be installed to exclude normal IO requests.
2292  *
2293  * Resync and recovery are handled very differently.
2294  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2295  *
2296  * For resync, we iterate over virtual addresses, read all copies,
2297  * and update if there are differences.  If only one copy is live,
2298  * skip it.
2299  * For recovery, we iterate over physical addresses, read a good
2300  * value for each non-in_sync drive, and over-write.
2301  *
2302  * So, for recovery we may have several outstanding complex requests for a
2303  * given address, one for each out-of-sync device.  We model this by allocating
2304  * a number of r10_bio structures, one for each out-of-sync device.
2305  * As we setup these structures, we collect all bio's together into a list
2306  * which we then process collectively to add pages, and then process again
2307  * to pass to generic_make_request.
2308  *
2309  * The r10_bio structures are linked using a borrowed master_bio pointer.
2310  * This link is counted in ->remaining.  When the r10_bio that points to NULL
2311  * has its remaining count decremented to 0, the whole complex operation
2312  * is complete.
2313  *
2314  */
2315
2316 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
2317                              int *skipped, int go_faster)
2318 {
2319         struct r10conf *conf = mddev->private;
2320         struct r10bio *r10_bio;
2321         struct bio *biolist = NULL, *bio;
2322         sector_t max_sector, nr_sectors;
2323         int i;
2324         int max_sync;
2325         sector_t sync_blocks;
2326         sector_t sectors_skipped = 0;
2327         int chunks_skipped = 0;
2328
2329         if (!conf->r10buf_pool)
2330                 if (init_resync(conf))
2331                         return 0;
2332
2333  skipped:
2334         max_sector = mddev->dev_sectors;
2335         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2336                 max_sector = mddev->resync_max_sectors;
2337         if (sector_nr >= max_sector) {
2338                 /* If we aborted, we need to abort the
2339                  * sync on the 'current' bitmap chucks (there can
2340                  * be several when recovering multiple devices).
2341                  * as we may have started syncing it but not finished.
2342                  * We can find the current address in
2343                  * mddev->curr_resync, but for recovery,
2344                  * we need to convert that to several
2345                  * virtual addresses.
2346                  */
2347                 if (mddev->curr_resync < max_sector) { /* aborted */
2348                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2349                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2350                                                 &sync_blocks, 1);
2351                         else for (i=0; i<conf->raid_disks; i++) {
2352                                 sector_t sect =
2353                                         raid10_find_virt(conf, mddev->curr_resync, i);
2354                                 bitmap_end_sync(mddev->bitmap, sect,
2355                                                 &sync_blocks, 1);
2356                         }
2357                 } else /* completed sync */
2358                         conf->fullsync = 0;
2359
2360                 bitmap_close_sync(mddev->bitmap);
2361                 close_sync(conf);
2362                 *skipped = 1;
2363                 return sectors_skipped;
2364         }
2365         if (chunks_skipped >= conf->raid_disks) {
2366                 /* if there has been nothing to do on any drive,
2367                  * then there is nothing to do at all..
2368                  */
2369                 *skipped = 1;
2370                 return (max_sector - sector_nr) + sectors_skipped;
2371         }
2372
2373         if (max_sector > mddev->resync_max)
2374                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2375
2376         /* make sure whole request will fit in a chunk - if chunks
2377          * are meaningful
2378          */
2379         if (conf->near_copies < conf->raid_disks &&
2380             max_sector > (sector_nr | conf->chunk_mask))
2381                 max_sector = (sector_nr | conf->chunk_mask) + 1;
2382         /*
2383          * If there is non-resync activity waiting for us then
2384          * put in a delay to throttle resync.
2385          */
2386         if (!go_faster && conf->nr_waiting)
2387                 msleep_interruptible(1000);
2388
2389         /* Again, very different code for resync and recovery.
2390          * Both must result in an r10bio with a list of bios that
2391          * have bi_end_io, bi_sector, bi_bdev set,
2392          * and bi_private set to the r10bio.
2393          * For recovery, we may actually create several r10bios
2394          * with 2 bios in each, that correspond to the bios in the main one.
2395          * In this case, the subordinate r10bios link back through a
2396          * borrowed master_bio pointer, and the counter in the master
2397          * includes a ref from each subordinate.
2398          */
2399         /* First, we decide what to do and set ->bi_end_io
2400          * To end_sync_read if we want to read, and
2401          * end_sync_write if we will want to write.
2402          */
2403
2404         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2405         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2406                 /* recovery... the complicated one */
2407                 int j;
2408                 r10_bio = NULL;
2409
2410                 for (i=0 ; i<conf->raid_disks; i++) {
2411                         int still_degraded;
2412                         struct r10bio *rb2;
2413                         sector_t sect;
2414                         int must_sync;
2415                         int any_working;
2416
2417                         if (conf->mirrors[i].rdev == NULL ||
2418                             test_bit(In_sync, &conf->mirrors[i].rdev->flags)) 
2419                                 continue;
2420
2421                         still_degraded = 0;
2422                         /* want to reconstruct this device */
2423                         rb2 = r10_bio;
2424                         sect = raid10_find_virt(conf, sector_nr, i);
2425                         /* Unless we are doing a full sync, we only need
2426                          * to recover the block if it is set in the bitmap
2427                          */
2428                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
2429                                                       &sync_blocks, 1);
2430                         if (sync_blocks < max_sync)
2431                                 max_sync = sync_blocks;
2432                         if (!must_sync &&
2433                             !conf->fullsync) {
2434                                 /* yep, skip the sync_blocks here, but don't assume
2435                                  * that there will never be anything to do here
2436                                  */
2437                                 chunks_skipped = -1;
2438                                 continue;
2439                         }
2440
2441                         r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2442                         raise_barrier(conf, rb2 != NULL);
2443                         atomic_set(&r10_bio->remaining, 0);
2444
2445                         r10_bio->master_bio = (struct bio*)rb2;
2446                         if (rb2)
2447                                 atomic_inc(&rb2->remaining);
2448                         r10_bio->mddev = mddev;
2449                         set_bit(R10BIO_IsRecover, &r10_bio->state);
2450                         r10_bio->sector = sect;
2451
2452                         raid10_find_phys(conf, r10_bio);
2453
2454                         /* Need to check if the array will still be
2455                          * degraded
2456                          */
2457                         for (j=0; j<conf->raid_disks; j++)
2458                                 if (conf->mirrors[j].rdev == NULL ||
2459                                     test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
2460                                         still_degraded = 1;
2461                                         break;
2462                                 }
2463
2464                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
2465                                                       &sync_blocks, still_degraded);
2466
2467                         any_working = 0;
2468                         for (j=0; j<conf->copies;j++) {
2469                                 int k;
2470                                 int d = r10_bio->devs[j].devnum;
2471                                 sector_t from_addr, to_addr;
2472                                 struct md_rdev *rdev;
2473                                 sector_t sector, first_bad;
2474                                 int bad_sectors;
2475                                 if (!conf->mirrors[d].rdev ||
2476                                     !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
2477                                         continue;
2478                                 /* This is where we read from */
2479                                 any_working = 1;
2480                                 rdev = conf->mirrors[d].rdev;
2481                                 sector = r10_bio->devs[j].addr;
2482
2483                                 if (is_badblock(rdev, sector, max_sync,
2484                                                 &first_bad, &bad_sectors)) {
2485                                         if (first_bad > sector)
2486                                                 max_sync = first_bad - sector;
2487                                         else {
2488                                                 bad_sectors -= (sector
2489                                                                 - first_bad);
2490                                                 if (max_sync > bad_sectors)
2491                                                         max_sync = bad_sectors;
2492                                                 continue;
2493                                         }
2494                                 }
2495                                 bio = r10_bio->devs[0].bio;
2496                                 bio->bi_next = biolist;
2497                                 biolist = bio;
2498                                 bio->bi_private = r10_bio;
2499                                 bio->bi_end_io = end_sync_read;
2500                                 bio->bi_rw = READ;
2501                                 from_addr = r10_bio->devs[j].addr;
2502                                 bio->bi_sector = from_addr +
2503                                         conf->mirrors[d].rdev->data_offset;
2504                                 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2505                                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2506                                 atomic_inc(&r10_bio->remaining);
2507                                 /* and we write to 'i' */
2508
2509                                 for (k=0; k<conf->copies; k++)
2510                                         if (r10_bio->devs[k].devnum == i)
2511                                                 break;
2512                                 BUG_ON(k == conf->copies);
2513                                 bio = r10_bio->devs[1].bio;
2514                                 bio->bi_next = biolist;
2515                                 biolist = bio;
2516                                 bio->bi_private = r10_bio;
2517                                 bio->bi_end_io = end_sync_write;
2518                                 bio->bi_rw = WRITE;
2519                                 to_addr = r10_bio->devs[k].addr;
2520                                 bio->bi_sector = to_addr +
2521                                         conf->mirrors[i].rdev->data_offset;
2522                                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
2523
2524                                 r10_bio->devs[0].devnum = d;
2525                                 r10_bio->devs[0].addr = from_addr;
2526                                 r10_bio->devs[1].devnum = i;
2527                                 r10_bio->devs[1].addr = to_addr;
2528
2529                                 break;
2530                         }
2531                         if (j == conf->copies) {
2532                                 /* Cannot recover, so abort the recovery or
2533                                  * record a bad block */
2534                                 put_buf(r10_bio);
2535                                 if (rb2)
2536                                         atomic_dec(&rb2->remaining);
2537                                 r10_bio = rb2;
2538                                 if (any_working) {
2539                                         /* problem is that there are bad blocks
2540                                          * on other device(s)
2541                                          */
2542                                         int k;
2543                                         for (k = 0; k < conf->copies; k++)
2544                                                 if (r10_bio->devs[k].devnum == i)
2545                                                         break;
2546                                         if (!rdev_set_badblocks(
2547                                                     conf->mirrors[i].rdev,
2548                                                     r10_bio->devs[k].addr,
2549                                                     max_sync, 0))
2550                                                 any_working = 0;
2551                                 }
2552                                 if (!any_working)  {
2553                                         if (!test_and_set_bit(MD_RECOVERY_INTR,
2554                                                               &mddev->recovery))
2555                                                 printk(KERN_INFO "md/raid10:%s: insufficient "
2556                                                        "working devices for recovery.\n",
2557                                                        mdname(mddev));
2558                                         conf->mirrors[i].recovery_disabled
2559                                                 = mddev->recovery_disabled;
2560                                 }
2561                                 break;
2562                         }
2563                 }
2564                 if (biolist == NULL) {
2565                         while (r10_bio) {
2566                                 struct r10bio *rb2 = r10_bio;
2567                                 r10_bio = (struct r10bio*) rb2->master_bio;
2568                                 rb2->master_bio = NULL;
2569                                 put_buf(rb2);
2570                         }
2571                         goto giveup;
2572                 }
2573         } else {
2574                 /* resync. Schedule a read for every block at this virt offset */
2575                 int count = 0;
2576
2577                 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2578
2579                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2580                                        &sync_blocks, mddev->degraded) &&
2581                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
2582                                                  &mddev->recovery)) {
2583                         /* We can skip this block */
2584                         *skipped = 1;
2585                         return sync_blocks + sectors_skipped;
2586                 }
2587                 if (sync_blocks < max_sync)
2588                         max_sync = sync_blocks;
2589                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2590
2591                 r10_bio->mddev = mddev;
2592                 atomic_set(&r10_bio->remaining, 0);
2593                 raise_barrier(conf, 0);
2594                 conf->next_resync = sector_nr;
2595
2596                 r10_bio->master_bio = NULL;
2597                 r10_bio->sector = sector_nr;
2598                 set_bit(R10BIO_IsSync, &r10_bio->state);
2599                 raid10_find_phys(conf, r10_bio);
2600                 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
2601
2602                 for (i=0; i<conf->copies; i++) {
2603                         int d = r10_bio->devs[i].devnum;
2604                         sector_t first_bad, sector;
2605                         int bad_sectors;
2606
2607                         bio = r10_bio->devs[i].bio;
2608                         bio->bi_end_io = NULL;
2609                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
2610                         if (conf->mirrors[d].rdev == NULL ||
2611                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
2612                                 continue;
2613                         sector = r10_bio->devs[i].addr;
2614                         if (is_badblock(conf->mirrors[d].rdev,
2615                                         sector, max_sync,
2616                                         &first_bad, &bad_sectors)) {
2617                                 if (first_bad > sector)
2618                                         max_sync = first_bad - sector;
2619                                 else {
2620                                         bad_sectors -= (sector - first_bad);
2621                                         if (max_sync > bad_sectors)
2622                                                 max_sync = max_sync;
2623                                         continue;
2624                                 }
2625                         }
2626                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2627                         atomic_inc(&r10_bio->remaining);
2628                         bio->bi_next = biolist;
2629                         biolist = bio;
2630                         bio->bi_private = r10_bio;
2631                         bio->bi_end_io = end_sync_read;
2632                         bio->bi_rw = READ;
2633                         bio->bi_sector = sector +
2634                                 conf->mirrors[d].rdev->data_offset;
2635                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2636                         count++;
2637                 }
2638
2639                 if (count < 2) {
2640                         for (i=0; i<conf->copies; i++) {
2641                                 int d = r10_bio->devs[i].devnum;
2642                                 if (r10_bio->devs[i].bio->bi_end_io)
2643                                         rdev_dec_pending(conf->mirrors[d].rdev,
2644                                                          mddev);
2645                         }
2646                         put_buf(r10_bio);
2647                         biolist = NULL;
2648                         goto giveup;
2649                 }
2650         }
2651
2652         for (bio = biolist; bio ; bio=bio->bi_next) {
2653
2654                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
2655                 if (bio->bi_end_io)
2656                         bio->bi_flags |= 1 << BIO_UPTODATE;
2657                 bio->bi_vcnt = 0;
2658                 bio->bi_idx = 0;
2659                 bio->bi_phys_segments = 0;
2660                 bio->bi_size = 0;
2661         }
2662
2663         nr_sectors = 0;
2664         if (sector_nr + max_sync < max_sector)
2665                 max_sector = sector_nr + max_sync;
2666         do {
2667                 struct page *page;
2668                 int len = PAGE_SIZE;
2669                 if (sector_nr + (len>>9) > max_sector)
2670                         len = (max_sector - sector_nr) << 9;
2671                 if (len == 0)
2672                         break;
2673                 for (bio= biolist ; bio ; bio=bio->bi_next) {
2674                         struct bio *bio2;
2675                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2676                         if (bio_add_page(bio, page, len, 0))
2677                                 continue;
2678
2679                         /* stop here */
2680                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2681                         for (bio2 = biolist;
2682                              bio2 && bio2 != bio;
2683                              bio2 = bio2->bi_next) {
2684                                 /* remove last page from this bio */
2685                                 bio2->bi_vcnt--;
2686                                 bio2->bi_size -= len;
2687                                 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
2688                         }
2689                         goto bio_full;
2690                 }
2691                 nr_sectors += len>>9;
2692                 sector_nr += len>>9;
2693         } while (biolist->bi_vcnt < RESYNC_PAGES);
2694  bio_full:
2695         r10_bio->sectors = nr_sectors;
2696
2697         while (biolist) {
2698                 bio = biolist;
2699                 biolist = biolist->bi_next;
2700
2701                 bio->bi_next = NULL;
2702                 r10_bio = bio->bi_private;
2703                 r10_bio->sectors = nr_sectors;
2704
2705                 if (bio->bi_end_io == end_sync_read) {
2706                         md_sync_acct(bio->bi_bdev, nr_sectors);
2707                         generic_make_request(bio);
2708                 }
2709         }
2710
2711         if (sectors_skipped)
2712                 /* pretend they weren't skipped, it makes
2713                  * no important difference in this case
2714                  */
2715                 md_done_sync(mddev, sectors_skipped, 1);
2716
2717         return sectors_skipped + nr_sectors;
2718  giveup:
2719         /* There is nowhere to write, so all non-sync
2720          * drives must be failed or in resync, all drives
2721          * have a bad block, so try the next chunk...
2722          */
2723         if (sector_nr + max_sync < max_sector)
2724                 max_sector = sector_nr + max_sync;
2725
2726         sectors_skipped += (max_sector - sector_nr);
2727         chunks_skipped ++;
2728         sector_nr = max_sector;
2729         goto skipped;
2730 }
2731
2732 static sector_t
2733 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2734 {
2735         sector_t size;
2736         struct r10conf *conf = mddev->private;
2737
2738         if (!raid_disks)
2739                 raid_disks = conf->raid_disks;
2740         if (!sectors)
2741                 sectors = conf->dev_sectors;
2742
2743         size = sectors >> conf->chunk_shift;
2744         sector_div(size, conf->far_copies);
2745         size = size * raid_disks;
2746         sector_div(size, conf->near_copies);
2747
2748         return size << conf->chunk_shift;
2749 }
2750
2751
2752 static struct r10conf *setup_conf(struct mddev *mddev)
2753 {
2754         struct r10conf *conf = NULL;
2755         int nc, fc, fo;
2756         sector_t stride, size;
2757         int err = -EINVAL;
2758
2759         if (mddev->new_chunk_sectors < (PAGE_SIZE >> 9) ||
2760             !is_power_of_2(mddev->new_chunk_sectors)) {
2761                 printk(KERN_ERR "md/raid10:%s: chunk size must be "
2762                        "at least PAGE_SIZE(%ld) and be a power of 2.\n",
2763                        mdname(mddev), PAGE_SIZE);
2764                 goto out;
2765         }
2766
2767         nc = mddev->new_layout & 255;
2768         fc = (mddev->new_layout >> 8) & 255;
2769         fo = mddev->new_layout & (1<<16);
2770
2771         if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
2772             (mddev->new_layout >> 17)) {
2773                 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
2774                        mdname(mddev), mddev->new_layout);
2775                 goto out;
2776         }
2777
2778         err = -ENOMEM;
2779         conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
2780         if (!conf)
2781                 goto out;
2782
2783         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2784                                 GFP_KERNEL);
2785         if (!conf->mirrors)
2786                 goto out;
2787
2788         conf->tmppage = alloc_page(GFP_KERNEL);
2789         if (!conf->tmppage)
2790                 goto out;
2791
2792
2793         conf->raid_disks = mddev->raid_disks;
2794         conf->near_copies = nc;
2795         conf->far_copies = fc;
2796         conf->copies = nc*fc;
2797         conf->far_offset = fo;
2798         conf->chunk_mask = mddev->new_chunk_sectors - 1;
2799         conf->chunk_shift = ffz(~mddev->new_chunk_sectors);
2800
2801         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2802                                            r10bio_pool_free, conf);
2803         if (!conf->r10bio_pool)
2804                 goto out;
2805
2806         size = mddev->dev_sectors >> conf->chunk_shift;
2807         sector_div(size, fc);
2808         size = size * conf->raid_disks;
2809         sector_div(size, nc);
2810         /* 'size' is now the number of chunks in the array */
2811         /* calculate "used chunks per device" in 'stride' */
2812         stride = size * conf->copies;
2813
2814         /* We need to round up when dividing by raid_disks to
2815          * get the stride size.
2816          */
2817         stride += conf->raid_disks - 1;
2818         sector_div(stride, conf->raid_disks);
2819
2820         conf->dev_sectors = stride << conf->chunk_shift;
2821
2822         if (fo)
2823                 stride = 1;
2824         else
2825                 sector_div(stride, fc);
2826         conf->stride = stride << conf->chunk_shift;
2827
2828
2829         spin_lock_init(&conf->device_lock);
2830         INIT_LIST_HEAD(&conf->retry_list);
2831
2832         spin_lock_init(&conf->resync_lock);
2833         init_waitqueue_head(&conf->wait_barrier);
2834
2835         conf->thread = md_register_thread(raid10d, mddev, NULL);
2836         if (!conf->thread)
2837                 goto out;
2838
2839         conf->mddev = mddev;
2840         return conf;
2841
2842  out:
2843         printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
2844                mdname(mddev));
2845         if (conf) {
2846                 if (conf->r10bio_pool)
2847                         mempool_destroy(conf->r10bio_pool);
2848                 kfree(conf->mirrors);
2849                 safe_put_page(conf->tmppage);
2850                 kfree(conf);
2851         }
2852         return ERR_PTR(err);
2853 }
2854
2855 static int run(struct mddev *mddev)
2856 {
2857         struct r10conf *conf;
2858         int i, disk_idx, chunk_size;
2859         struct mirror_info *disk;
2860         struct md_rdev *rdev;
2861         sector_t size;
2862
2863         /*
2864          * copy the already verified devices into our private RAID10
2865          * bookkeeping area. [whatever we allocate in run(),
2866          * should be freed in stop()]
2867          */
2868
2869         if (mddev->private == NULL) {
2870                 conf = setup_conf(mddev);
2871                 if (IS_ERR(conf))
2872                         return PTR_ERR(conf);
2873                 mddev->private = conf;
2874         }
2875         conf = mddev->private;
2876         if (!conf)
2877                 goto out;
2878
2879         mddev->thread = conf->thread;
2880         conf->thread = NULL;
2881
2882         chunk_size = mddev->chunk_sectors << 9;
2883         blk_queue_io_min(mddev->queue, chunk_size);
2884         if (conf->raid_disks % conf->near_copies)
2885                 blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks);
2886         else
2887                 blk_queue_io_opt(mddev->queue, chunk_size *
2888                                  (conf->raid_disks / conf->near_copies));
2889
2890         list_for_each_entry(rdev, &mddev->disks, same_set) {
2891
2892                 disk_idx = rdev->raid_disk;
2893                 if (disk_idx >= conf->raid_disks
2894                     || disk_idx < 0)
2895                         continue;
2896                 disk = conf->mirrors + disk_idx;
2897
2898                 disk->rdev = rdev;
2899                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2900                                   rdev->data_offset << 9);
2901                 /* as we don't honour merge_bvec_fn, we must never risk
2902                  * violating it, so limit max_segments to 1 lying
2903                  * within a single page.
2904                  */
2905                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2906                         blk_queue_max_segments(mddev->queue, 1);
2907                         blk_queue_segment_boundary(mddev->queue,
2908                                                    PAGE_CACHE_SIZE - 1);
2909                 }
2910
2911                 disk->head_position = 0;
2912         }
2913         /* need to check that every block has at least one working mirror */
2914         if (!enough(conf, -1)) {
2915                 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
2916                        mdname(mddev));
2917                 goto out_free_conf;
2918         }
2919
2920         mddev->degraded = 0;
2921         for (i = 0; i < conf->raid_disks; i++) {
2922
2923                 disk = conf->mirrors + i;
2924
2925                 if (!disk->rdev ||
2926                     !test_bit(In_sync, &disk->rdev->flags)) {
2927                         disk->head_position = 0;
2928                         mddev->degraded++;
2929                         if (disk->rdev)
2930                                 conf->fullsync = 1;
2931                 }
2932                 disk->recovery_disabled = mddev->recovery_disabled - 1;
2933         }
2934
2935         if (mddev->recovery_cp != MaxSector)
2936                 printk(KERN_NOTICE "md/raid10:%s: not clean"
2937                        " -- starting background reconstruction\n",
2938                        mdname(mddev));
2939         printk(KERN_INFO
2940                 "md/raid10:%s: active with %d out of %d devices\n",
2941                 mdname(mddev), conf->raid_disks - mddev->degraded,
2942                 conf->raid_disks);
2943         /*
2944          * Ok, everything is just fine now
2945          */
2946         mddev->dev_sectors = conf->dev_sectors;
2947         size = raid10_size(mddev, 0, 0);
2948         md_set_array_sectors(mddev, size);
2949         mddev->resync_max_sectors = size;
2950
2951         mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2952         mddev->queue->backing_dev_info.congested_data = mddev;
2953
2954         /* Calculate max read-ahead size.
2955          * We need to readahead at least twice a whole stripe....
2956          * maybe...
2957          */
2958         {
2959                 int stripe = conf->raid_disks *
2960                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
2961                 stripe /= conf->near_copies;
2962                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2963                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2964         }
2965
2966         if (conf->near_copies < conf->raid_disks)
2967                 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2968
2969         if (md_integrity_register(mddev))
2970                 goto out_free_conf;
2971
2972         return 0;
2973
2974 out_free_conf:
2975         md_unregister_thread(&mddev->thread);
2976         if (conf->r10bio_pool)
2977                 mempool_destroy(conf->r10bio_pool);
2978         safe_put_page(conf->tmppage);
2979         kfree(conf->mirrors);
2980         kfree(conf);
2981         mddev->private = NULL;
2982 out:
2983         return -EIO;
2984 }
2985
2986 static int stop(struct mddev *mddev)
2987 {
2988         struct r10conf *conf = mddev->private;
2989
2990         raise_barrier(conf, 0);
2991         lower_barrier(conf);
2992
2993         md_unregister_thread(&mddev->thread);
2994         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2995         if (conf->r10bio_pool)
2996                 mempool_destroy(conf->r10bio_pool);
2997         kfree(conf->mirrors);
2998         kfree(conf);
2999         mddev->private = NULL;
3000         return 0;
3001 }
3002
3003 static void raid10_quiesce(struct mddev *mddev, int state)
3004 {
3005         struct r10conf *conf = mddev->private;
3006
3007         switch(state) {
3008         case 1:
3009                 raise_barrier(conf, 0);
3010                 break;
3011         case 0:
3012                 lower_barrier(conf);
3013                 break;
3014         }
3015 }
3016
3017 static void *raid10_takeover_raid0(struct mddev *mddev)
3018 {
3019         struct md_rdev *rdev;
3020         struct r10conf *conf;
3021
3022         if (mddev->degraded > 0) {
3023                 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3024                        mdname(mddev));
3025                 return ERR_PTR(-EINVAL);
3026         }
3027
3028         /* Set new parameters */
3029         mddev->new_level = 10;
3030         /* new layout: far_copies = 1, near_copies = 2 */
3031         mddev->new_layout = (1<<8) + 2;
3032         mddev->new_chunk_sectors = mddev->chunk_sectors;
3033         mddev->delta_disks = mddev->raid_disks;
3034         mddev->raid_disks *= 2;
3035         /* make sure it will be not marked as dirty */
3036         mddev->recovery_cp = MaxSector;
3037
3038         conf = setup_conf(mddev);
3039         if (!IS_ERR(conf)) {
3040                 list_for_each_entry(rdev, &mddev->disks, same_set)
3041                         if (rdev->raid_disk >= 0)
3042                                 rdev->new_raid_disk = rdev->raid_disk * 2;
3043                 conf->barrier = 1;
3044         }
3045
3046         return conf;
3047 }
3048
3049 static void *raid10_takeover(struct mddev *mddev)
3050 {
3051         struct r0conf *raid0_conf;
3052
3053         /* raid10 can take over:
3054          *  raid0 - providing it has only two drives
3055          */
3056         if (mddev->level == 0) {
3057                 /* for raid0 takeover only one zone is supported */
3058                 raid0_conf = mddev->private;
3059                 if (raid0_conf->nr_strip_zones > 1) {
3060                         printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3061                                " with more than one zone.\n",
3062                                mdname(mddev));
3063                         return ERR_PTR(-EINVAL);
3064                 }
3065                 return raid10_takeover_raid0(mddev);
3066         }
3067         return ERR_PTR(-EINVAL);
3068 }
3069
3070 static struct md_personality raid10_personality =
3071 {
3072         .name           = "raid10",
3073         .level          = 10,
3074         .owner          = THIS_MODULE,
3075         .make_request   = make_request,
3076         .run            = run,
3077         .stop           = stop,
3078         .status         = status,
3079         .error_handler  = error,
3080         .hot_add_disk   = raid10_add_disk,
3081         .hot_remove_disk= raid10_remove_disk,
3082         .spare_active   = raid10_spare_active,
3083         .sync_request   = sync_request,
3084         .quiesce        = raid10_quiesce,
3085         .size           = raid10_size,
3086         .takeover       = raid10_takeover,
3087 };
3088
3089 static int __init raid_init(void)
3090 {
3091         return register_md_personality(&raid10_personality);
3092 }
3093
3094 static void raid_exit(void)
3095 {
3096         unregister_md_personality(&raid10_personality);
3097 }
3098
3099 module_init(raid_init);
3100 module_exit(raid_exit);
3101 MODULE_LICENSE("GPL");
3102 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
3103 MODULE_ALIAS("md-personality-9"); /* RAID10 */
3104 MODULE_ALIAS("md-raid10");
3105 MODULE_ALIAS("md-level-10");
3106
3107 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);