Merge branch 'for-3.6/drivers' of git://git.kernel.dk/linux-block
[linux-drm-fsl-dcu.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/module.h>
25 #include <linux/seq_file.h>
26 #include <linux/ratelimit.h>
27 #include <linux/kthread.h>
28 #include "md.h"
29 #include "raid10.h"
30 #include "raid0.h"
31 #include "bitmap.h"
32
33 /*
34  * RAID10 provides a combination of RAID0 and RAID1 functionality.
35  * The layout of data is defined by
36  *    chunk_size
37  *    raid_disks
38  *    near_copies (stored in low byte of layout)
39  *    far_copies (stored in second byte of layout)
40  *    far_offset (stored in bit 16 of layout )
41  *
42  * The data to be stored is divided into chunks using chunksize.
43  * Each device is divided into far_copies sections.
44  * In each section, chunks are laid out in a style similar to raid0, but
45  * near_copies copies of each chunk is stored (each on a different drive).
46  * The starting device for each section is offset near_copies from the starting
47  * device of the previous section.
48  * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
49  * drive.
50  * near_copies and far_copies must be at least one, and their product is at most
51  * raid_disks.
52  *
53  * If far_offset is true, then the far_copies are handled a bit differently.
54  * The copies are still in different stripes, but instead of be very far apart
55  * on disk, there are adjacent stripes.
56  */
57
58 /*
59  * Number of guaranteed r10bios in case of extreme VM load:
60  */
61 #define NR_RAID10_BIOS 256
62
63 /* when we get a read error on a read-only array, we redirect to another
64  * device without failing the first device, or trying to over-write to
65  * correct the read error.  To keep track of bad blocks on a per-bio
66  * level, we store IO_BLOCKED in the appropriate 'bios' pointer
67  */
68 #define IO_BLOCKED ((struct bio *)1)
69 /* When we successfully write to a known bad-block, we need to remove the
70  * bad-block marking which must be done from process context.  So we record
71  * the success by setting devs[n].bio to IO_MADE_GOOD
72  */
73 #define IO_MADE_GOOD ((struct bio *)2)
74
75 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
76
77 /* When there are this many requests queued to be written by
78  * the raid10 thread, we become 'congested' to provide back-pressure
79  * for writeback.
80  */
81 static int max_queued_requests = 1024;
82
83 static void allow_barrier(struct r10conf *conf);
84 static void lower_barrier(struct r10conf *conf);
85 static int enough(struct r10conf *conf, int ignore);
86 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
87                                 int *skipped);
88 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
89 static void end_reshape_write(struct bio *bio, int error);
90 static void end_reshape(struct r10conf *conf);
91
92 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
93 {
94         struct r10conf *conf = data;
95         int size = offsetof(struct r10bio, devs[conf->copies]);
96
97         /* allocate a r10bio with room for raid_disks entries in the
98          * bios array */
99         return kzalloc(size, gfp_flags);
100 }
101
102 static void r10bio_pool_free(void *r10_bio, void *data)
103 {
104         kfree(r10_bio);
105 }
106
107 /* Maximum size of each resync request */
108 #define RESYNC_BLOCK_SIZE (64*1024)
109 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
110 /* amount of memory to reserve for resync requests */
111 #define RESYNC_WINDOW (1024*1024)
112 /* maximum number of concurrent requests, memory permitting */
113 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
114
115 /*
116  * When performing a resync, we need to read and compare, so
117  * we need as many pages are there are copies.
118  * When performing a recovery, we need 2 bios, one for read,
119  * one for write (we recover only one drive per r10buf)
120  *
121  */
122 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
123 {
124         struct r10conf *conf = data;
125         struct page *page;
126         struct r10bio *r10_bio;
127         struct bio *bio;
128         int i, j;
129         int nalloc;
130
131         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
132         if (!r10_bio)
133                 return NULL;
134
135         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
136             test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
137                 nalloc = conf->copies; /* resync */
138         else
139                 nalloc = 2; /* recovery */
140
141         /*
142          * Allocate bios.
143          */
144         for (j = nalloc ; j-- ; ) {
145                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
146                 if (!bio)
147                         goto out_free_bio;
148                 r10_bio->devs[j].bio = bio;
149                 if (!conf->have_replacement)
150                         continue;
151                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
152                 if (!bio)
153                         goto out_free_bio;
154                 r10_bio->devs[j].repl_bio = bio;
155         }
156         /*
157          * Allocate RESYNC_PAGES data pages and attach them
158          * where needed.
159          */
160         for (j = 0 ; j < nalloc; j++) {
161                 struct bio *rbio = r10_bio->devs[j].repl_bio;
162                 bio = r10_bio->devs[j].bio;
163                 for (i = 0; i < RESYNC_PAGES; i++) {
164                         if (j > 0 && !test_bit(MD_RECOVERY_SYNC,
165                                                &conf->mddev->recovery)) {
166                                 /* we can share bv_page's during recovery
167                                  * and reshape */
168                                 struct bio *rbio = r10_bio->devs[0].bio;
169                                 page = rbio->bi_io_vec[i].bv_page;
170                                 get_page(page);
171                         } else
172                                 page = alloc_page(gfp_flags);
173                         if (unlikely(!page))
174                                 goto out_free_pages;
175
176                         bio->bi_io_vec[i].bv_page = page;
177                         if (rbio)
178                                 rbio->bi_io_vec[i].bv_page = page;
179                 }
180         }
181
182         return r10_bio;
183
184 out_free_pages:
185         for ( ; i > 0 ; i--)
186                 safe_put_page(bio->bi_io_vec[i-1].bv_page);
187         while (j--)
188                 for (i = 0; i < RESYNC_PAGES ; i++)
189                         safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
190         j = 0;
191 out_free_bio:
192         for ( ; j < nalloc; j++) {
193                 if (r10_bio->devs[j].bio)
194                         bio_put(r10_bio->devs[j].bio);
195                 if (r10_bio->devs[j].repl_bio)
196                         bio_put(r10_bio->devs[j].repl_bio);
197         }
198         r10bio_pool_free(r10_bio, conf);
199         return NULL;
200 }
201
202 static void r10buf_pool_free(void *__r10_bio, void *data)
203 {
204         int i;
205         struct r10conf *conf = data;
206         struct r10bio *r10bio = __r10_bio;
207         int j;
208
209         for (j=0; j < conf->copies; j++) {
210                 struct bio *bio = r10bio->devs[j].bio;
211                 if (bio) {
212                         for (i = 0; i < RESYNC_PAGES; i++) {
213                                 safe_put_page(bio->bi_io_vec[i].bv_page);
214                                 bio->bi_io_vec[i].bv_page = NULL;
215                         }
216                         bio_put(bio);
217                 }
218                 bio = r10bio->devs[j].repl_bio;
219                 if (bio)
220                         bio_put(bio);
221         }
222         r10bio_pool_free(r10bio, conf);
223 }
224
225 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
226 {
227         int i;
228
229         for (i = 0; i < conf->copies; i++) {
230                 struct bio **bio = & r10_bio->devs[i].bio;
231                 if (!BIO_SPECIAL(*bio))
232                         bio_put(*bio);
233                 *bio = NULL;
234                 bio = &r10_bio->devs[i].repl_bio;
235                 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
236                         bio_put(*bio);
237                 *bio = NULL;
238         }
239 }
240
241 static void free_r10bio(struct r10bio *r10_bio)
242 {
243         struct r10conf *conf = r10_bio->mddev->private;
244
245         put_all_bios(conf, r10_bio);
246         mempool_free(r10_bio, conf->r10bio_pool);
247 }
248
249 static void put_buf(struct r10bio *r10_bio)
250 {
251         struct r10conf *conf = r10_bio->mddev->private;
252
253         mempool_free(r10_bio, conf->r10buf_pool);
254
255         lower_barrier(conf);
256 }
257
258 static void reschedule_retry(struct r10bio *r10_bio)
259 {
260         unsigned long flags;
261         struct mddev *mddev = r10_bio->mddev;
262         struct r10conf *conf = mddev->private;
263
264         spin_lock_irqsave(&conf->device_lock, flags);
265         list_add(&r10_bio->retry_list, &conf->retry_list);
266         conf->nr_queued ++;
267         spin_unlock_irqrestore(&conf->device_lock, flags);
268
269         /* wake up frozen array... */
270         wake_up(&conf->wait_barrier);
271
272         md_wakeup_thread(mddev->thread);
273 }
274
275 /*
276  * raid_end_bio_io() is called when we have finished servicing a mirrored
277  * operation and are ready to return a success/failure code to the buffer
278  * cache layer.
279  */
280 static void raid_end_bio_io(struct r10bio *r10_bio)
281 {
282         struct bio *bio = r10_bio->master_bio;
283         int done;
284         struct r10conf *conf = r10_bio->mddev->private;
285
286         if (bio->bi_phys_segments) {
287                 unsigned long flags;
288                 spin_lock_irqsave(&conf->device_lock, flags);
289                 bio->bi_phys_segments--;
290                 done = (bio->bi_phys_segments == 0);
291                 spin_unlock_irqrestore(&conf->device_lock, flags);
292         } else
293                 done = 1;
294         if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
295                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
296         if (done) {
297                 bio_endio(bio, 0);
298                 /*
299                  * Wake up any possible resync thread that waits for the device
300                  * to go idle.
301                  */
302                 allow_barrier(conf);
303         }
304         free_r10bio(r10_bio);
305 }
306
307 /*
308  * Update disk head position estimator based on IRQ completion info.
309  */
310 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
311 {
312         struct r10conf *conf = r10_bio->mddev->private;
313
314         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
315                 r10_bio->devs[slot].addr + (r10_bio->sectors);
316 }
317
318 /*
319  * Find the disk number which triggered given bio
320  */
321 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
322                          struct bio *bio, int *slotp, int *replp)
323 {
324         int slot;
325         int repl = 0;
326
327         for (slot = 0; slot < conf->copies; slot++) {
328                 if (r10_bio->devs[slot].bio == bio)
329                         break;
330                 if (r10_bio->devs[slot].repl_bio == bio) {
331                         repl = 1;
332                         break;
333                 }
334         }
335
336         BUG_ON(slot == conf->copies);
337         update_head_pos(slot, r10_bio);
338
339         if (slotp)
340                 *slotp = slot;
341         if (replp)
342                 *replp = repl;
343         return r10_bio->devs[slot].devnum;
344 }
345
346 static void raid10_end_read_request(struct bio *bio, int error)
347 {
348         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
349         struct r10bio *r10_bio = bio->bi_private;
350         int slot, dev;
351         struct md_rdev *rdev;
352         struct r10conf *conf = r10_bio->mddev->private;
353
354
355         slot = r10_bio->read_slot;
356         dev = r10_bio->devs[slot].devnum;
357         rdev = r10_bio->devs[slot].rdev;
358         /*
359          * this branch is our 'one mirror IO has finished' event handler:
360          */
361         update_head_pos(slot, r10_bio);
362
363         if (uptodate) {
364                 /*
365                  * Set R10BIO_Uptodate in our master bio, so that
366                  * we will return a good error code to the higher
367                  * levels even if IO on some other mirrored buffer fails.
368                  *
369                  * The 'master' represents the composite IO operation to
370                  * user-side. So if something waits for IO, then it will
371                  * wait for the 'master' bio.
372                  */
373                 set_bit(R10BIO_Uptodate, &r10_bio->state);
374         } else {
375                 /* If all other devices that store this block have
376                  * failed, we want to return the error upwards rather
377                  * than fail the last device.  Here we redefine
378                  * "uptodate" to mean "Don't want to retry"
379                  */
380                 unsigned long flags;
381                 spin_lock_irqsave(&conf->device_lock, flags);
382                 if (!enough(conf, rdev->raid_disk))
383                         uptodate = 1;
384                 spin_unlock_irqrestore(&conf->device_lock, flags);
385         }
386         if (uptodate) {
387                 raid_end_bio_io(r10_bio);
388                 rdev_dec_pending(rdev, conf->mddev);
389         } else {
390                 /*
391                  * oops, read error - keep the refcount on the rdev
392                  */
393                 char b[BDEVNAME_SIZE];
394                 printk_ratelimited(KERN_ERR
395                                    "md/raid10:%s: %s: rescheduling sector %llu\n",
396                                    mdname(conf->mddev),
397                                    bdevname(rdev->bdev, b),
398                                    (unsigned long long)r10_bio->sector);
399                 set_bit(R10BIO_ReadError, &r10_bio->state);
400                 reschedule_retry(r10_bio);
401         }
402 }
403
404 static void close_write(struct r10bio *r10_bio)
405 {
406         /* clear the bitmap if all writes complete successfully */
407         bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
408                         r10_bio->sectors,
409                         !test_bit(R10BIO_Degraded, &r10_bio->state),
410                         0);
411         md_write_end(r10_bio->mddev);
412 }
413
414 static void one_write_done(struct r10bio *r10_bio)
415 {
416         if (atomic_dec_and_test(&r10_bio->remaining)) {
417                 if (test_bit(R10BIO_WriteError, &r10_bio->state))
418                         reschedule_retry(r10_bio);
419                 else {
420                         close_write(r10_bio);
421                         if (test_bit(R10BIO_MadeGood, &r10_bio->state))
422                                 reschedule_retry(r10_bio);
423                         else
424                                 raid_end_bio_io(r10_bio);
425                 }
426         }
427 }
428
429 static void raid10_end_write_request(struct bio *bio, int error)
430 {
431         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
432         struct r10bio *r10_bio = bio->bi_private;
433         int dev;
434         int dec_rdev = 1;
435         struct r10conf *conf = r10_bio->mddev->private;
436         int slot, repl;
437         struct md_rdev *rdev = NULL;
438
439         dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
440
441         if (repl)
442                 rdev = conf->mirrors[dev].replacement;
443         if (!rdev) {
444                 smp_rmb();
445                 repl = 0;
446                 rdev = conf->mirrors[dev].rdev;
447         }
448         /*
449          * this branch is our 'one mirror IO has finished' event handler:
450          */
451         if (!uptodate) {
452                 if (repl)
453                         /* Never record new bad blocks to replacement,
454                          * just fail it.
455                          */
456                         md_error(rdev->mddev, rdev);
457                 else {
458                         set_bit(WriteErrorSeen, &rdev->flags);
459                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
460                                 set_bit(MD_RECOVERY_NEEDED,
461                                         &rdev->mddev->recovery);
462                         set_bit(R10BIO_WriteError, &r10_bio->state);
463                         dec_rdev = 0;
464                 }
465         } else {
466                 /*
467                  * Set R10BIO_Uptodate in our master bio, so that
468                  * we will return a good error code for to the higher
469                  * levels even if IO on some other mirrored buffer fails.
470                  *
471                  * The 'master' represents the composite IO operation to
472                  * user-side. So if something waits for IO, then it will
473                  * wait for the 'master' bio.
474                  */
475                 sector_t first_bad;
476                 int bad_sectors;
477
478                 set_bit(R10BIO_Uptodate, &r10_bio->state);
479
480                 /* Maybe we can clear some bad blocks. */
481                 if (is_badblock(rdev,
482                                 r10_bio->devs[slot].addr,
483                                 r10_bio->sectors,
484                                 &first_bad, &bad_sectors)) {
485                         bio_put(bio);
486                         if (repl)
487                                 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
488                         else
489                                 r10_bio->devs[slot].bio = IO_MADE_GOOD;
490                         dec_rdev = 0;
491                         set_bit(R10BIO_MadeGood, &r10_bio->state);
492                 }
493         }
494
495         /*
496          *
497          * Let's see if all mirrored write operations have finished
498          * already.
499          */
500         one_write_done(r10_bio);
501         if (dec_rdev)
502                 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
503 }
504
505 /*
506  * RAID10 layout manager
507  * As well as the chunksize and raid_disks count, there are two
508  * parameters: near_copies and far_copies.
509  * near_copies * far_copies must be <= raid_disks.
510  * Normally one of these will be 1.
511  * If both are 1, we get raid0.
512  * If near_copies == raid_disks, we get raid1.
513  *
514  * Chunks are laid out in raid0 style with near_copies copies of the
515  * first chunk, followed by near_copies copies of the next chunk and
516  * so on.
517  * If far_copies > 1, then after 1/far_copies of the array has been assigned
518  * as described above, we start again with a device offset of near_copies.
519  * So we effectively have another copy of the whole array further down all
520  * the drives, but with blocks on different drives.
521  * With this layout, and block is never stored twice on the one device.
522  *
523  * raid10_find_phys finds the sector offset of a given virtual sector
524  * on each device that it is on.
525  *
526  * raid10_find_virt does the reverse mapping, from a device and a
527  * sector offset to a virtual address
528  */
529
530 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
531 {
532         int n,f;
533         sector_t sector;
534         sector_t chunk;
535         sector_t stripe;
536         int dev;
537         int slot = 0;
538
539         /* now calculate first sector/dev */
540         chunk = r10bio->sector >> geo->chunk_shift;
541         sector = r10bio->sector & geo->chunk_mask;
542
543         chunk *= geo->near_copies;
544         stripe = chunk;
545         dev = sector_div(stripe, geo->raid_disks);
546         if (geo->far_offset)
547                 stripe *= geo->far_copies;
548
549         sector += stripe << geo->chunk_shift;
550
551         /* and calculate all the others */
552         for (n = 0; n < geo->near_copies; n++) {
553                 int d = dev;
554                 sector_t s = sector;
555                 r10bio->devs[slot].addr = sector;
556                 r10bio->devs[slot].devnum = d;
557                 slot++;
558
559                 for (f = 1; f < geo->far_copies; f++) {
560                         d += geo->near_copies;
561                         if (d >= geo->raid_disks)
562                                 d -= geo->raid_disks;
563                         s += geo->stride;
564                         r10bio->devs[slot].devnum = d;
565                         r10bio->devs[slot].addr = s;
566                         slot++;
567                 }
568                 dev++;
569                 if (dev >= geo->raid_disks) {
570                         dev = 0;
571                         sector += (geo->chunk_mask + 1);
572                 }
573         }
574 }
575
576 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
577 {
578         struct geom *geo = &conf->geo;
579
580         if (conf->reshape_progress != MaxSector &&
581             ((r10bio->sector >= conf->reshape_progress) !=
582              conf->mddev->reshape_backwards)) {
583                 set_bit(R10BIO_Previous, &r10bio->state);
584                 geo = &conf->prev;
585         } else
586                 clear_bit(R10BIO_Previous, &r10bio->state);
587
588         __raid10_find_phys(geo, r10bio);
589 }
590
591 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
592 {
593         sector_t offset, chunk, vchunk;
594         /* Never use conf->prev as this is only called during resync
595          * or recovery, so reshape isn't happening
596          */
597         struct geom *geo = &conf->geo;
598
599         offset = sector & geo->chunk_mask;
600         if (geo->far_offset) {
601                 int fc;
602                 chunk = sector >> geo->chunk_shift;
603                 fc = sector_div(chunk, geo->far_copies);
604                 dev -= fc * geo->near_copies;
605                 if (dev < 0)
606                         dev += geo->raid_disks;
607         } else {
608                 while (sector >= geo->stride) {
609                         sector -= geo->stride;
610                         if (dev < geo->near_copies)
611                                 dev += geo->raid_disks - geo->near_copies;
612                         else
613                                 dev -= geo->near_copies;
614                 }
615                 chunk = sector >> geo->chunk_shift;
616         }
617         vchunk = chunk * geo->raid_disks + dev;
618         sector_div(vchunk, geo->near_copies);
619         return (vchunk << geo->chunk_shift) + offset;
620 }
621
622 /**
623  *      raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
624  *      @q: request queue
625  *      @bvm: properties of new bio
626  *      @biovec: the request that could be merged to it.
627  *
628  *      Return amount of bytes we can accept at this offset
629  *      This requires checking for end-of-chunk if near_copies != raid_disks,
630  *      and for subordinate merge_bvec_fns if merge_check_needed.
631  */
632 static int raid10_mergeable_bvec(struct request_queue *q,
633                                  struct bvec_merge_data *bvm,
634                                  struct bio_vec *biovec)
635 {
636         struct mddev *mddev = q->queuedata;
637         struct r10conf *conf = mddev->private;
638         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
639         int max;
640         unsigned int chunk_sectors;
641         unsigned int bio_sectors = bvm->bi_size >> 9;
642         struct geom *geo = &conf->geo;
643
644         chunk_sectors = (conf->geo.chunk_mask & conf->prev.chunk_mask) + 1;
645         if (conf->reshape_progress != MaxSector &&
646             ((sector >= conf->reshape_progress) !=
647              conf->mddev->reshape_backwards))
648                 geo = &conf->prev;
649
650         if (geo->near_copies < geo->raid_disks) {
651                 max = (chunk_sectors - ((sector & (chunk_sectors - 1))
652                                         + bio_sectors)) << 9;
653                 if (max < 0)
654                         /* bio_add cannot handle a negative return */
655                         max = 0;
656                 if (max <= biovec->bv_len && bio_sectors == 0)
657                         return biovec->bv_len;
658         } else
659                 max = biovec->bv_len;
660
661         if (mddev->merge_check_needed) {
662                 struct r10bio r10_bio;
663                 int s;
664                 if (conf->reshape_progress != MaxSector) {
665                         /* Cannot give any guidance during reshape */
666                         if (max <= biovec->bv_len && bio_sectors == 0)
667                                 return biovec->bv_len;
668                         return 0;
669                 }
670                 r10_bio.sector = sector;
671                 raid10_find_phys(conf, &r10_bio);
672                 rcu_read_lock();
673                 for (s = 0; s < conf->copies; s++) {
674                         int disk = r10_bio.devs[s].devnum;
675                         struct md_rdev *rdev = rcu_dereference(
676                                 conf->mirrors[disk].rdev);
677                         if (rdev && !test_bit(Faulty, &rdev->flags)) {
678                                 struct request_queue *q =
679                                         bdev_get_queue(rdev->bdev);
680                                 if (q->merge_bvec_fn) {
681                                         bvm->bi_sector = r10_bio.devs[s].addr
682                                                 + rdev->data_offset;
683                                         bvm->bi_bdev = rdev->bdev;
684                                         max = min(max, q->merge_bvec_fn(
685                                                           q, bvm, biovec));
686                                 }
687                         }
688                         rdev = rcu_dereference(conf->mirrors[disk].replacement);
689                         if (rdev && !test_bit(Faulty, &rdev->flags)) {
690                                 struct request_queue *q =
691                                         bdev_get_queue(rdev->bdev);
692                                 if (q->merge_bvec_fn) {
693                                         bvm->bi_sector = r10_bio.devs[s].addr
694                                                 + rdev->data_offset;
695                                         bvm->bi_bdev = rdev->bdev;
696                                         max = min(max, q->merge_bvec_fn(
697                                                           q, bvm, biovec));
698                                 }
699                         }
700                 }
701                 rcu_read_unlock();
702         }
703         return max;
704 }
705
706 /*
707  * This routine returns the disk from which the requested read should
708  * be done. There is a per-array 'next expected sequential IO' sector
709  * number - if this matches on the next IO then we use the last disk.
710  * There is also a per-disk 'last know head position' sector that is
711  * maintained from IRQ contexts, both the normal and the resync IO
712  * completion handlers update this position correctly. If there is no
713  * perfect sequential match then we pick the disk whose head is closest.
714  *
715  * If there are 2 mirrors in the same 2 devices, performance degrades
716  * because position is mirror, not device based.
717  *
718  * The rdev for the device selected will have nr_pending incremented.
719  */
720
721 /*
722  * FIXME: possibly should rethink readbalancing and do it differently
723  * depending on near_copies / far_copies geometry.
724  */
725 static struct md_rdev *read_balance(struct r10conf *conf,
726                                     struct r10bio *r10_bio,
727                                     int *max_sectors)
728 {
729         const sector_t this_sector = r10_bio->sector;
730         int disk, slot;
731         int sectors = r10_bio->sectors;
732         int best_good_sectors;
733         sector_t new_distance, best_dist;
734         struct md_rdev *best_rdev, *rdev = NULL;
735         int do_balance;
736         int best_slot;
737         struct geom *geo = &conf->geo;
738
739         raid10_find_phys(conf, r10_bio);
740         rcu_read_lock();
741 retry:
742         sectors = r10_bio->sectors;
743         best_slot = -1;
744         best_rdev = NULL;
745         best_dist = MaxSector;
746         best_good_sectors = 0;
747         do_balance = 1;
748         /*
749          * Check if we can balance. We can balance on the whole
750          * device if no resync is going on (recovery is ok), or below
751          * the resync window. We take the first readable disk when
752          * above the resync window.
753          */
754         if (conf->mddev->recovery_cp < MaxSector
755             && (this_sector + sectors >= conf->next_resync))
756                 do_balance = 0;
757
758         for (slot = 0; slot < conf->copies ; slot++) {
759                 sector_t first_bad;
760                 int bad_sectors;
761                 sector_t dev_sector;
762
763                 if (r10_bio->devs[slot].bio == IO_BLOCKED)
764                         continue;
765                 disk = r10_bio->devs[slot].devnum;
766                 rdev = rcu_dereference(conf->mirrors[disk].replacement);
767                 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
768                     test_bit(Unmerged, &rdev->flags) ||
769                     r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
770                         rdev = rcu_dereference(conf->mirrors[disk].rdev);
771                 if (rdev == NULL ||
772                     test_bit(Faulty, &rdev->flags) ||
773                     test_bit(Unmerged, &rdev->flags))
774                         continue;
775                 if (!test_bit(In_sync, &rdev->flags) &&
776                     r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
777                         continue;
778
779                 dev_sector = r10_bio->devs[slot].addr;
780                 if (is_badblock(rdev, dev_sector, sectors,
781                                 &first_bad, &bad_sectors)) {
782                         if (best_dist < MaxSector)
783                                 /* Already have a better slot */
784                                 continue;
785                         if (first_bad <= dev_sector) {
786                                 /* Cannot read here.  If this is the
787                                  * 'primary' device, then we must not read
788                                  * beyond 'bad_sectors' from another device.
789                                  */
790                                 bad_sectors -= (dev_sector - first_bad);
791                                 if (!do_balance && sectors > bad_sectors)
792                                         sectors = bad_sectors;
793                                 if (best_good_sectors > sectors)
794                                         best_good_sectors = sectors;
795                         } else {
796                                 sector_t good_sectors =
797                                         first_bad - dev_sector;
798                                 if (good_sectors > best_good_sectors) {
799                                         best_good_sectors = good_sectors;
800                                         best_slot = slot;
801                                         best_rdev = rdev;
802                                 }
803                                 if (!do_balance)
804                                         /* Must read from here */
805                                         break;
806                         }
807                         continue;
808                 } else
809                         best_good_sectors = sectors;
810
811                 if (!do_balance)
812                         break;
813
814                 /* This optimisation is debatable, and completely destroys
815                  * sequential read speed for 'far copies' arrays.  So only
816                  * keep it for 'near' arrays, and review those later.
817                  */
818                 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
819                         break;
820
821                 /* for far > 1 always use the lowest address */
822                 if (geo->far_copies > 1)
823                         new_distance = r10_bio->devs[slot].addr;
824                 else
825                         new_distance = abs(r10_bio->devs[slot].addr -
826                                            conf->mirrors[disk].head_position);
827                 if (new_distance < best_dist) {
828                         best_dist = new_distance;
829                         best_slot = slot;
830                         best_rdev = rdev;
831                 }
832         }
833         if (slot >= conf->copies) {
834                 slot = best_slot;
835                 rdev = best_rdev;
836         }
837
838         if (slot >= 0) {
839                 atomic_inc(&rdev->nr_pending);
840                 if (test_bit(Faulty, &rdev->flags)) {
841                         /* Cannot risk returning a device that failed
842                          * before we inc'ed nr_pending
843                          */
844                         rdev_dec_pending(rdev, conf->mddev);
845                         goto retry;
846                 }
847                 r10_bio->read_slot = slot;
848         } else
849                 rdev = NULL;
850         rcu_read_unlock();
851         *max_sectors = best_good_sectors;
852
853         return rdev;
854 }
855
856 int md_raid10_congested(struct mddev *mddev, int bits)
857 {
858         struct r10conf *conf = mddev->private;
859         int i, ret = 0;
860
861         if ((bits & (1 << BDI_async_congested)) &&
862             conf->pending_count >= max_queued_requests)
863                 return 1;
864
865         rcu_read_lock();
866         for (i = 0;
867              (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
868                      && ret == 0;
869              i++) {
870                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
871                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
872                         struct request_queue *q = bdev_get_queue(rdev->bdev);
873
874                         ret |= bdi_congested(&q->backing_dev_info, bits);
875                 }
876         }
877         rcu_read_unlock();
878         return ret;
879 }
880 EXPORT_SYMBOL_GPL(md_raid10_congested);
881
882 static int raid10_congested(void *data, int bits)
883 {
884         struct mddev *mddev = data;
885
886         return mddev_congested(mddev, bits) ||
887                 md_raid10_congested(mddev, bits);
888 }
889
890 static void flush_pending_writes(struct r10conf *conf)
891 {
892         /* Any writes that have been queued but are awaiting
893          * bitmap updates get flushed here.
894          */
895         spin_lock_irq(&conf->device_lock);
896
897         if (conf->pending_bio_list.head) {
898                 struct bio *bio;
899                 bio = bio_list_get(&conf->pending_bio_list);
900                 conf->pending_count = 0;
901                 spin_unlock_irq(&conf->device_lock);
902                 /* flush any pending bitmap writes to disk
903                  * before proceeding w/ I/O */
904                 bitmap_unplug(conf->mddev->bitmap);
905                 wake_up(&conf->wait_barrier);
906
907                 while (bio) { /* submit pending writes */
908                         struct bio *next = bio->bi_next;
909                         bio->bi_next = NULL;
910                         generic_make_request(bio);
911                         bio = next;
912                 }
913         } else
914                 spin_unlock_irq(&conf->device_lock);
915 }
916
917 /* Barriers....
918  * Sometimes we need to suspend IO while we do something else,
919  * either some resync/recovery, or reconfigure the array.
920  * To do this we raise a 'barrier'.
921  * The 'barrier' is a counter that can be raised multiple times
922  * to count how many activities are happening which preclude
923  * normal IO.
924  * We can only raise the barrier if there is no pending IO.
925  * i.e. if nr_pending == 0.
926  * We choose only to raise the barrier if no-one is waiting for the
927  * barrier to go down.  This means that as soon as an IO request
928  * is ready, no other operations which require a barrier will start
929  * until the IO request has had a chance.
930  *
931  * So: regular IO calls 'wait_barrier'.  When that returns there
932  *    is no backgroup IO happening,  It must arrange to call
933  *    allow_barrier when it has finished its IO.
934  * backgroup IO calls must call raise_barrier.  Once that returns
935  *    there is no normal IO happeing.  It must arrange to call
936  *    lower_barrier when the particular background IO completes.
937  */
938
939 static void raise_barrier(struct r10conf *conf, int force)
940 {
941         BUG_ON(force && !conf->barrier);
942         spin_lock_irq(&conf->resync_lock);
943
944         /* Wait until no block IO is waiting (unless 'force') */
945         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
946                             conf->resync_lock, );
947
948         /* block any new IO from starting */
949         conf->barrier++;
950
951         /* Now wait for all pending IO to complete */
952         wait_event_lock_irq(conf->wait_barrier,
953                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
954                             conf->resync_lock, );
955
956         spin_unlock_irq(&conf->resync_lock);
957 }
958
959 static void lower_barrier(struct r10conf *conf)
960 {
961         unsigned long flags;
962         spin_lock_irqsave(&conf->resync_lock, flags);
963         conf->barrier--;
964         spin_unlock_irqrestore(&conf->resync_lock, flags);
965         wake_up(&conf->wait_barrier);
966 }
967
968 static void wait_barrier(struct r10conf *conf)
969 {
970         spin_lock_irq(&conf->resync_lock);
971         if (conf->barrier) {
972                 conf->nr_waiting++;
973                 /* Wait for the barrier to drop.
974                  * However if there are already pending
975                  * requests (preventing the barrier from
976                  * rising completely), and the
977                  * pre-process bio queue isn't empty,
978                  * then don't wait, as we need to empty
979                  * that queue to get the nr_pending
980                  * count down.
981                  */
982                 wait_event_lock_irq(conf->wait_barrier,
983                                     !conf->barrier ||
984                                     (conf->nr_pending &&
985                                      current->bio_list &&
986                                      !bio_list_empty(current->bio_list)),
987                                     conf->resync_lock,
988                         );
989                 conf->nr_waiting--;
990         }
991         conf->nr_pending++;
992         spin_unlock_irq(&conf->resync_lock);
993 }
994
995 static void allow_barrier(struct r10conf *conf)
996 {
997         unsigned long flags;
998         spin_lock_irqsave(&conf->resync_lock, flags);
999         conf->nr_pending--;
1000         spin_unlock_irqrestore(&conf->resync_lock, flags);
1001         wake_up(&conf->wait_barrier);
1002 }
1003
1004 static void freeze_array(struct r10conf *conf)
1005 {
1006         /* stop syncio and normal IO and wait for everything to
1007          * go quiet.
1008          * We increment barrier and nr_waiting, and then
1009          * wait until nr_pending match nr_queued+1
1010          * This is called in the context of one normal IO request
1011          * that has failed. Thus any sync request that might be pending
1012          * will be blocked by nr_pending, and we need to wait for
1013          * pending IO requests to complete or be queued for re-try.
1014          * Thus the number queued (nr_queued) plus this request (1)
1015          * must match the number of pending IOs (nr_pending) before
1016          * we continue.
1017          */
1018         spin_lock_irq(&conf->resync_lock);
1019         conf->barrier++;
1020         conf->nr_waiting++;
1021         wait_event_lock_irq(conf->wait_barrier,
1022                             conf->nr_pending == conf->nr_queued+1,
1023                             conf->resync_lock,
1024                             flush_pending_writes(conf));
1025
1026         spin_unlock_irq(&conf->resync_lock);
1027 }
1028
1029 static void unfreeze_array(struct r10conf *conf)
1030 {
1031         /* reverse the effect of the freeze */
1032         spin_lock_irq(&conf->resync_lock);
1033         conf->barrier--;
1034         conf->nr_waiting--;
1035         wake_up(&conf->wait_barrier);
1036         spin_unlock_irq(&conf->resync_lock);
1037 }
1038
1039 static sector_t choose_data_offset(struct r10bio *r10_bio,
1040                                    struct md_rdev *rdev)
1041 {
1042         if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1043             test_bit(R10BIO_Previous, &r10_bio->state))
1044                 return rdev->data_offset;
1045         else
1046                 return rdev->new_data_offset;
1047 }
1048
1049 static void make_request(struct mddev *mddev, struct bio * bio)
1050 {
1051         struct r10conf *conf = mddev->private;
1052         struct r10bio *r10_bio;
1053         struct bio *read_bio;
1054         int i;
1055         sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1056         int chunk_sects = chunk_mask + 1;
1057         const int rw = bio_data_dir(bio);
1058         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1059         const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
1060         unsigned long flags;
1061         struct md_rdev *blocked_rdev;
1062         int sectors_handled;
1063         int max_sectors;
1064         int sectors;
1065
1066         if (unlikely(bio->bi_rw & REQ_FLUSH)) {
1067                 md_flush_request(mddev, bio);
1068                 return;
1069         }
1070
1071         /* If this request crosses a chunk boundary, we need to
1072          * split it.  This will only happen for 1 PAGE (or less) requests.
1073          */
1074         if (unlikely((bio->bi_sector & chunk_mask) + (bio->bi_size >> 9)
1075                      > chunk_sects
1076                      && (conf->geo.near_copies < conf->geo.raid_disks
1077                          || conf->prev.near_copies < conf->prev.raid_disks))) {
1078                 struct bio_pair *bp;
1079                 /* Sanity check -- queue functions should prevent this happening */
1080                 if (bio->bi_vcnt != 1 ||
1081                     bio->bi_idx != 0)
1082                         goto bad_map;
1083                 /* This is a one page bio that upper layers
1084                  * refuse to split for us, so we need to split it.
1085                  */
1086                 bp = bio_split(bio,
1087                                chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
1088
1089                 /* Each of these 'make_request' calls will call 'wait_barrier'.
1090                  * If the first succeeds but the second blocks due to the resync
1091                  * thread raising the barrier, we will deadlock because the
1092                  * IO to the underlying device will be queued in generic_make_request
1093                  * and will never complete, so will never reduce nr_pending.
1094                  * So increment nr_waiting here so no new raise_barriers will
1095                  * succeed, and so the second wait_barrier cannot block.
1096                  */
1097                 spin_lock_irq(&conf->resync_lock);
1098                 conf->nr_waiting++;
1099                 spin_unlock_irq(&conf->resync_lock);
1100
1101                 make_request(mddev, &bp->bio1);
1102                 make_request(mddev, &bp->bio2);
1103
1104                 spin_lock_irq(&conf->resync_lock);
1105                 conf->nr_waiting--;
1106                 wake_up(&conf->wait_barrier);
1107                 spin_unlock_irq(&conf->resync_lock);
1108
1109                 bio_pair_release(bp);
1110                 return;
1111         bad_map:
1112                 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
1113                        " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
1114                        (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
1115
1116                 bio_io_error(bio);
1117                 return;
1118         }
1119
1120         md_write_start(mddev, bio);
1121
1122         /*
1123          * Register the new request and wait if the reconstruction
1124          * thread has put up a bar for new requests.
1125          * Continue immediately if no resync is active currently.
1126          */
1127         wait_barrier(conf);
1128
1129         sectors = bio->bi_size >> 9;
1130         while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1131             bio->bi_sector < conf->reshape_progress &&
1132             bio->bi_sector + sectors > conf->reshape_progress) {
1133                 /* IO spans the reshape position.  Need to wait for
1134                  * reshape to pass
1135                  */
1136                 allow_barrier(conf);
1137                 wait_event(conf->wait_barrier,
1138                            conf->reshape_progress <= bio->bi_sector ||
1139                            conf->reshape_progress >= bio->bi_sector + sectors);
1140                 wait_barrier(conf);
1141         }
1142         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1143             bio_data_dir(bio) == WRITE &&
1144             (mddev->reshape_backwards
1145              ? (bio->bi_sector < conf->reshape_safe &&
1146                 bio->bi_sector + sectors > conf->reshape_progress)
1147              : (bio->bi_sector + sectors > conf->reshape_safe &&
1148                 bio->bi_sector < conf->reshape_progress))) {
1149                 /* Need to update reshape_position in metadata */
1150                 mddev->reshape_position = conf->reshape_progress;
1151                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1152                 set_bit(MD_CHANGE_PENDING, &mddev->flags);
1153                 md_wakeup_thread(mddev->thread);
1154                 wait_event(mddev->sb_wait,
1155                            !test_bit(MD_CHANGE_PENDING, &mddev->flags));
1156
1157                 conf->reshape_safe = mddev->reshape_position;
1158         }
1159
1160         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1161
1162         r10_bio->master_bio = bio;
1163         r10_bio->sectors = sectors;
1164
1165         r10_bio->mddev = mddev;
1166         r10_bio->sector = bio->bi_sector;
1167         r10_bio->state = 0;
1168
1169         /* We might need to issue multiple reads to different
1170          * devices if there are bad blocks around, so we keep
1171          * track of the number of reads in bio->bi_phys_segments.
1172          * If this is 0, there is only one r10_bio and no locking
1173          * will be needed when the request completes.  If it is
1174          * non-zero, then it is the number of not-completed requests.
1175          */
1176         bio->bi_phys_segments = 0;
1177         clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1178
1179         if (rw == READ) {
1180                 /*
1181                  * read balancing logic:
1182                  */
1183                 struct md_rdev *rdev;
1184                 int slot;
1185
1186 read_again:
1187                 rdev = read_balance(conf, r10_bio, &max_sectors);
1188                 if (!rdev) {
1189                         raid_end_bio_io(r10_bio);
1190                         return;
1191                 }
1192                 slot = r10_bio->read_slot;
1193
1194                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1195                 md_trim_bio(read_bio, r10_bio->sector - bio->bi_sector,
1196                             max_sectors);
1197
1198                 r10_bio->devs[slot].bio = read_bio;
1199                 r10_bio->devs[slot].rdev = rdev;
1200
1201                 read_bio->bi_sector = r10_bio->devs[slot].addr +
1202                         choose_data_offset(r10_bio, rdev);
1203                 read_bio->bi_bdev = rdev->bdev;
1204                 read_bio->bi_end_io = raid10_end_read_request;
1205                 read_bio->bi_rw = READ | do_sync;
1206                 read_bio->bi_private = r10_bio;
1207
1208                 if (max_sectors < r10_bio->sectors) {
1209                         /* Could not read all from this device, so we will
1210                          * need another r10_bio.
1211                          */
1212                         sectors_handled = (r10_bio->sectors + max_sectors
1213                                            - bio->bi_sector);
1214                         r10_bio->sectors = max_sectors;
1215                         spin_lock_irq(&conf->device_lock);
1216                         if (bio->bi_phys_segments == 0)
1217                                 bio->bi_phys_segments = 2;
1218                         else
1219                                 bio->bi_phys_segments++;
1220                         spin_unlock(&conf->device_lock);
1221                         /* Cannot call generic_make_request directly
1222                          * as that will be queued in __generic_make_request
1223                          * and subsequent mempool_alloc might block
1224                          * waiting for it.  so hand bio over to raid10d.
1225                          */
1226                         reschedule_retry(r10_bio);
1227
1228                         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1229
1230                         r10_bio->master_bio = bio;
1231                         r10_bio->sectors = ((bio->bi_size >> 9)
1232                                             - sectors_handled);
1233                         r10_bio->state = 0;
1234                         r10_bio->mddev = mddev;
1235                         r10_bio->sector = bio->bi_sector + sectors_handled;
1236                         goto read_again;
1237                 } else
1238                         generic_make_request(read_bio);
1239                 return;
1240         }
1241
1242         /*
1243          * WRITE:
1244          */
1245         if (conf->pending_count >= max_queued_requests) {
1246                 md_wakeup_thread(mddev->thread);
1247                 wait_event(conf->wait_barrier,
1248                            conf->pending_count < max_queued_requests);
1249         }
1250         /* first select target devices under rcu_lock and
1251          * inc refcount on their rdev.  Record them by setting
1252          * bios[x] to bio
1253          * If there are known/acknowledged bad blocks on any device
1254          * on which we have seen a write error, we want to avoid
1255          * writing to those blocks.  This potentially requires several
1256          * writes to write around the bad blocks.  Each set of writes
1257          * gets its own r10_bio with a set of bios attached.  The number
1258          * of r10_bios is recored in bio->bi_phys_segments just as with
1259          * the read case.
1260          */
1261
1262         r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1263         raid10_find_phys(conf, r10_bio);
1264 retry_write:
1265         blocked_rdev = NULL;
1266         rcu_read_lock();
1267         max_sectors = r10_bio->sectors;
1268
1269         for (i = 0;  i < conf->copies; i++) {
1270                 int d = r10_bio->devs[i].devnum;
1271                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1272                 struct md_rdev *rrdev = rcu_dereference(
1273                         conf->mirrors[d].replacement);
1274                 if (rdev == rrdev)
1275                         rrdev = NULL;
1276                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1277                         atomic_inc(&rdev->nr_pending);
1278                         blocked_rdev = rdev;
1279                         break;
1280                 }
1281                 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1282                         atomic_inc(&rrdev->nr_pending);
1283                         blocked_rdev = rrdev;
1284                         break;
1285                 }
1286                 if (rrdev && (test_bit(Faulty, &rrdev->flags)
1287                               || test_bit(Unmerged, &rrdev->flags)))
1288                         rrdev = NULL;
1289
1290                 r10_bio->devs[i].bio = NULL;
1291                 r10_bio->devs[i].repl_bio = NULL;
1292                 if (!rdev || test_bit(Faulty, &rdev->flags) ||
1293                     test_bit(Unmerged, &rdev->flags)) {
1294                         set_bit(R10BIO_Degraded, &r10_bio->state);
1295                         continue;
1296                 }
1297                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1298                         sector_t first_bad;
1299                         sector_t dev_sector = r10_bio->devs[i].addr;
1300                         int bad_sectors;
1301                         int is_bad;
1302
1303                         is_bad = is_badblock(rdev, dev_sector,
1304                                              max_sectors,
1305                                              &first_bad, &bad_sectors);
1306                         if (is_bad < 0) {
1307                                 /* Mustn't write here until the bad block
1308                                  * is acknowledged
1309                                  */
1310                                 atomic_inc(&rdev->nr_pending);
1311                                 set_bit(BlockedBadBlocks, &rdev->flags);
1312                                 blocked_rdev = rdev;
1313                                 break;
1314                         }
1315                         if (is_bad && first_bad <= dev_sector) {
1316                                 /* Cannot write here at all */
1317                                 bad_sectors -= (dev_sector - first_bad);
1318                                 if (bad_sectors < max_sectors)
1319                                         /* Mustn't write more than bad_sectors
1320                                          * to other devices yet
1321                                          */
1322                                         max_sectors = bad_sectors;
1323                                 /* We don't set R10BIO_Degraded as that
1324                                  * only applies if the disk is missing,
1325                                  * so it might be re-added, and we want to
1326                                  * know to recover this chunk.
1327                                  * In this case the device is here, and the
1328                                  * fact that this chunk is not in-sync is
1329                                  * recorded in the bad block log.
1330                                  */
1331                                 continue;
1332                         }
1333                         if (is_bad) {
1334                                 int good_sectors = first_bad - dev_sector;
1335                                 if (good_sectors < max_sectors)
1336                                         max_sectors = good_sectors;
1337                         }
1338                 }
1339                 r10_bio->devs[i].bio = bio;
1340                 atomic_inc(&rdev->nr_pending);
1341                 if (rrdev) {
1342                         r10_bio->devs[i].repl_bio = bio;
1343                         atomic_inc(&rrdev->nr_pending);
1344                 }
1345         }
1346         rcu_read_unlock();
1347
1348         if (unlikely(blocked_rdev)) {
1349                 /* Have to wait for this device to get unblocked, then retry */
1350                 int j;
1351                 int d;
1352
1353                 for (j = 0; j < i; j++) {
1354                         if (r10_bio->devs[j].bio) {
1355                                 d = r10_bio->devs[j].devnum;
1356                                 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1357                         }
1358                         if (r10_bio->devs[j].repl_bio) {
1359                                 struct md_rdev *rdev;
1360                                 d = r10_bio->devs[j].devnum;
1361                                 rdev = conf->mirrors[d].replacement;
1362                                 if (!rdev) {
1363                                         /* Race with remove_disk */
1364                                         smp_mb();
1365                                         rdev = conf->mirrors[d].rdev;
1366                                 }
1367                                 rdev_dec_pending(rdev, mddev);
1368                         }
1369                 }
1370                 allow_barrier(conf);
1371                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1372                 wait_barrier(conf);
1373                 goto retry_write;
1374         }
1375
1376         if (max_sectors < r10_bio->sectors) {
1377                 /* We are splitting this into multiple parts, so
1378                  * we need to prepare for allocating another r10_bio.
1379                  */
1380                 r10_bio->sectors = max_sectors;
1381                 spin_lock_irq(&conf->device_lock);
1382                 if (bio->bi_phys_segments == 0)
1383                         bio->bi_phys_segments = 2;
1384                 else
1385                         bio->bi_phys_segments++;
1386                 spin_unlock_irq(&conf->device_lock);
1387         }
1388         sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector;
1389
1390         atomic_set(&r10_bio->remaining, 1);
1391         bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1392
1393         for (i = 0; i < conf->copies; i++) {
1394                 struct bio *mbio;
1395                 int d = r10_bio->devs[i].devnum;
1396                 if (!r10_bio->devs[i].bio)
1397                         continue;
1398
1399                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1400                 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1401                             max_sectors);
1402                 r10_bio->devs[i].bio = mbio;
1403
1404                 mbio->bi_sector = (r10_bio->devs[i].addr+
1405                                    choose_data_offset(r10_bio,
1406                                                       conf->mirrors[d].rdev));
1407                 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1408                 mbio->bi_end_io = raid10_end_write_request;
1409                 mbio->bi_rw = WRITE | do_sync | do_fua;
1410                 mbio->bi_private = r10_bio;
1411
1412                 atomic_inc(&r10_bio->remaining);
1413                 spin_lock_irqsave(&conf->device_lock, flags);
1414                 bio_list_add(&conf->pending_bio_list, mbio);
1415                 conf->pending_count++;
1416                 spin_unlock_irqrestore(&conf->device_lock, flags);
1417                 if (!mddev_check_plugged(mddev))
1418                         md_wakeup_thread(mddev->thread);
1419
1420                 if (!r10_bio->devs[i].repl_bio)
1421                         continue;
1422
1423                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1424                 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1425                             max_sectors);
1426                 r10_bio->devs[i].repl_bio = mbio;
1427
1428                 /* We are actively writing to the original device
1429                  * so it cannot disappear, so the replacement cannot
1430                  * become NULL here
1431                  */
1432                 mbio->bi_sector = (r10_bio->devs[i].addr +
1433                                    choose_data_offset(
1434                                            r10_bio,
1435                                            conf->mirrors[d].replacement));
1436                 mbio->bi_bdev = conf->mirrors[d].replacement->bdev;
1437                 mbio->bi_end_io = raid10_end_write_request;
1438                 mbio->bi_rw = WRITE | do_sync | do_fua;
1439                 mbio->bi_private = r10_bio;
1440
1441                 atomic_inc(&r10_bio->remaining);
1442                 spin_lock_irqsave(&conf->device_lock, flags);
1443                 bio_list_add(&conf->pending_bio_list, mbio);
1444                 conf->pending_count++;
1445                 spin_unlock_irqrestore(&conf->device_lock, flags);
1446                 if (!mddev_check_plugged(mddev))
1447                         md_wakeup_thread(mddev->thread);
1448         }
1449
1450         /* Don't remove the bias on 'remaining' (one_write_done) until
1451          * after checking if we need to go around again.
1452          */
1453
1454         if (sectors_handled < (bio->bi_size >> 9)) {
1455                 one_write_done(r10_bio);
1456                 /* We need another r10_bio.  It has already been counted
1457                  * in bio->bi_phys_segments.
1458                  */
1459                 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1460
1461                 r10_bio->master_bio = bio;
1462                 r10_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1463
1464                 r10_bio->mddev = mddev;
1465                 r10_bio->sector = bio->bi_sector + sectors_handled;
1466                 r10_bio->state = 0;
1467                 goto retry_write;
1468         }
1469         one_write_done(r10_bio);
1470
1471         /* In case raid10d snuck in to freeze_array */
1472         wake_up(&conf->wait_barrier);
1473 }
1474
1475 static void status(struct seq_file *seq, struct mddev *mddev)
1476 {
1477         struct r10conf *conf = mddev->private;
1478         int i;
1479
1480         if (conf->geo.near_copies < conf->geo.raid_disks)
1481                 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1482         if (conf->geo.near_copies > 1)
1483                 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1484         if (conf->geo.far_copies > 1) {
1485                 if (conf->geo.far_offset)
1486                         seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1487                 else
1488                         seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1489         }
1490         seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1491                                         conf->geo.raid_disks - mddev->degraded);
1492         for (i = 0; i < conf->geo.raid_disks; i++)
1493                 seq_printf(seq, "%s",
1494                               conf->mirrors[i].rdev &&
1495                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1496         seq_printf(seq, "]");
1497 }
1498
1499 /* check if there are enough drives for
1500  * every block to appear on atleast one.
1501  * Don't consider the device numbered 'ignore'
1502  * as we might be about to remove it.
1503  */
1504 static int _enough(struct r10conf *conf, struct geom *geo, int ignore)
1505 {
1506         int first = 0;
1507
1508         do {
1509                 int n = conf->copies;
1510                 int cnt = 0;
1511                 while (n--) {
1512                         if (conf->mirrors[first].rdev &&
1513                             first != ignore)
1514                                 cnt++;
1515                         first = (first+1) % geo->raid_disks;
1516                 }
1517                 if (cnt == 0)
1518                         return 0;
1519         } while (first != 0);
1520         return 1;
1521 }
1522
1523 static int enough(struct r10conf *conf, int ignore)
1524 {
1525         return _enough(conf, &conf->geo, ignore) &&
1526                 _enough(conf, &conf->prev, ignore);
1527 }
1528
1529 static void error(struct mddev *mddev, struct md_rdev *rdev)
1530 {
1531         char b[BDEVNAME_SIZE];
1532         struct r10conf *conf = mddev->private;
1533
1534         /*
1535          * If it is not operational, then we have already marked it as dead
1536          * else if it is the last working disks, ignore the error, let the
1537          * next level up know.
1538          * else mark the drive as failed
1539          */
1540         if (test_bit(In_sync, &rdev->flags)
1541             && !enough(conf, rdev->raid_disk))
1542                 /*
1543                  * Don't fail the drive, just return an IO error.
1544                  */
1545                 return;
1546         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1547                 unsigned long flags;
1548                 spin_lock_irqsave(&conf->device_lock, flags);
1549                 mddev->degraded++;
1550                 spin_unlock_irqrestore(&conf->device_lock, flags);
1551                 /*
1552                  * if recovery is running, make sure it aborts.
1553                  */
1554                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1555         }
1556         set_bit(Blocked, &rdev->flags);
1557         set_bit(Faulty, &rdev->flags);
1558         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1559         printk(KERN_ALERT
1560                "md/raid10:%s: Disk failure on %s, disabling device.\n"
1561                "md/raid10:%s: Operation continuing on %d devices.\n",
1562                mdname(mddev), bdevname(rdev->bdev, b),
1563                mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1564 }
1565
1566 static void print_conf(struct r10conf *conf)
1567 {
1568         int i;
1569         struct raid10_info *tmp;
1570
1571         printk(KERN_DEBUG "RAID10 conf printout:\n");
1572         if (!conf) {
1573                 printk(KERN_DEBUG "(!conf)\n");
1574                 return;
1575         }
1576         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1577                 conf->geo.raid_disks);
1578
1579         for (i = 0; i < conf->geo.raid_disks; i++) {
1580                 char b[BDEVNAME_SIZE];
1581                 tmp = conf->mirrors + i;
1582                 if (tmp->rdev)
1583                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1584                                 i, !test_bit(In_sync, &tmp->rdev->flags),
1585                                 !test_bit(Faulty, &tmp->rdev->flags),
1586                                 bdevname(tmp->rdev->bdev,b));
1587         }
1588 }
1589
1590 static void close_sync(struct r10conf *conf)
1591 {
1592         wait_barrier(conf);
1593         allow_barrier(conf);
1594
1595         mempool_destroy(conf->r10buf_pool);
1596         conf->r10buf_pool = NULL;
1597 }
1598
1599 static int raid10_spare_active(struct mddev *mddev)
1600 {
1601         int i;
1602         struct r10conf *conf = mddev->private;
1603         struct raid10_info *tmp;
1604         int count = 0;
1605         unsigned long flags;
1606
1607         /*
1608          * Find all non-in_sync disks within the RAID10 configuration
1609          * and mark them in_sync
1610          */
1611         for (i = 0; i < conf->geo.raid_disks; i++) {
1612                 tmp = conf->mirrors + i;
1613                 if (tmp->replacement
1614                     && tmp->replacement->recovery_offset == MaxSector
1615                     && !test_bit(Faulty, &tmp->replacement->flags)
1616                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1617                         /* Replacement has just become active */
1618                         if (!tmp->rdev
1619                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1620                                 count++;
1621                         if (tmp->rdev) {
1622                                 /* Replaced device not technically faulty,
1623                                  * but we need to be sure it gets removed
1624                                  * and never re-added.
1625                                  */
1626                                 set_bit(Faulty, &tmp->rdev->flags);
1627                                 sysfs_notify_dirent_safe(
1628                                         tmp->rdev->sysfs_state);
1629                         }
1630                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1631                 } else if (tmp->rdev
1632                            && !test_bit(Faulty, &tmp->rdev->flags)
1633                            && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1634                         count++;
1635                         sysfs_notify_dirent(tmp->rdev->sysfs_state);
1636                 }
1637         }
1638         spin_lock_irqsave(&conf->device_lock, flags);
1639         mddev->degraded -= count;
1640         spin_unlock_irqrestore(&conf->device_lock, flags);
1641
1642         print_conf(conf);
1643         return count;
1644 }
1645
1646
1647 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1648 {
1649         struct r10conf *conf = mddev->private;
1650         int err = -EEXIST;
1651         int mirror;
1652         int first = 0;
1653         int last = conf->geo.raid_disks - 1;
1654         struct request_queue *q = bdev_get_queue(rdev->bdev);
1655
1656         if (mddev->recovery_cp < MaxSector)
1657                 /* only hot-add to in-sync arrays, as recovery is
1658                  * very different from resync
1659                  */
1660                 return -EBUSY;
1661         if (rdev->saved_raid_disk < 0 && !_enough(conf, &conf->prev, -1))
1662                 return -EINVAL;
1663
1664         if (rdev->raid_disk >= 0)
1665                 first = last = rdev->raid_disk;
1666
1667         if (q->merge_bvec_fn) {
1668                 set_bit(Unmerged, &rdev->flags);
1669                 mddev->merge_check_needed = 1;
1670         }
1671
1672         if (rdev->saved_raid_disk >= first &&
1673             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1674                 mirror = rdev->saved_raid_disk;
1675         else
1676                 mirror = first;
1677         for ( ; mirror <= last ; mirror++) {
1678                 struct raid10_info *p = &conf->mirrors[mirror];
1679                 if (p->recovery_disabled == mddev->recovery_disabled)
1680                         continue;
1681                 if (p->rdev) {
1682                         if (!test_bit(WantReplacement, &p->rdev->flags) ||
1683                             p->replacement != NULL)
1684                                 continue;
1685                         clear_bit(In_sync, &rdev->flags);
1686                         set_bit(Replacement, &rdev->flags);
1687                         rdev->raid_disk = mirror;
1688                         err = 0;
1689                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1690                                           rdev->data_offset << 9);
1691                         conf->fullsync = 1;
1692                         rcu_assign_pointer(p->replacement, rdev);
1693                         break;
1694                 }
1695
1696                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1697                                   rdev->data_offset << 9);
1698
1699                 p->head_position = 0;
1700                 p->recovery_disabled = mddev->recovery_disabled - 1;
1701                 rdev->raid_disk = mirror;
1702                 err = 0;
1703                 if (rdev->saved_raid_disk != mirror)
1704                         conf->fullsync = 1;
1705                 rcu_assign_pointer(p->rdev, rdev);
1706                 break;
1707         }
1708         if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1709                 /* Some requests might not have seen this new
1710                  * merge_bvec_fn.  We must wait for them to complete
1711                  * before merging the device fully.
1712                  * First we make sure any code which has tested
1713                  * our function has submitted the request, then
1714                  * we wait for all outstanding requests to complete.
1715                  */
1716                 synchronize_sched();
1717                 raise_barrier(conf, 0);
1718                 lower_barrier(conf);
1719                 clear_bit(Unmerged, &rdev->flags);
1720         }
1721         md_integrity_add_rdev(rdev, mddev);
1722         print_conf(conf);
1723         return err;
1724 }
1725
1726 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1727 {
1728         struct r10conf *conf = mddev->private;
1729         int err = 0;
1730         int number = rdev->raid_disk;
1731         struct md_rdev **rdevp;
1732         struct raid10_info *p = conf->mirrors + number;
1733
1734         print_conf(conf);
1735         if (rdev == p->rdev)
1736                 rdevp = &p->rdev;
1737         else if (rdev == p->replacement)
1738                 rdevp = &p->replacement;
1739         else
1740                 return 0;
1741
1742         if (test_bit(In_sync, &rdev->flags) ||
1743             atomic_read(&rdev->nr_pending)) {
1744                 err = -EBUSY;
1745                 goto abort;
1746         }
1747         /* Only remove faulty devices if recovery
1748          * is not possible.
1749          */
1750         if (!test_bit(Faulty, &rdev->flags) &&
1751             mddev->recovery_disabled != p->recovery_disabled &&
1752             (!p->replacement || p->replacement == rdev) &&
1753             number < conf->geo.raid_disks &&
1754             enough(conf, -1)) {
1755                 err = -EBUSY;
1756                 goto abort;
1757         }
1758         *rdevp = NULL;
1759         synchronize_rcu();
1760         if (atomic_read(&rdev->nr_pending)) {
1761                 /* lost the race, try later */
1762                 err = -EBUSY;
1763                 *rdevp = rdev;
1764                 goto abort;
1765         } else if (p->replacement) {
1766                 /* We must have just cleared 'rdev' */
1767                 p->rdev = p->replacement;
1768                 clear_bit(Replacement, &p->replacement->flags);
1769                 smp_mb(); /* Make sure other CPUs may see both as identical
1770                            * but will never see neither -- if they are careful.
1771                            */
1772                 p->replacement = NULL;
1773                 clear_bit(WantReplacement, &rdev->flags);
1774         } else
1775                 /* We might have just remove the Replacement as faulty
1776                  * Clear the flag just in case
1777                  */
1778                 clear_bit(WantReplacement, &rdev->flags);
1779
1780         err = md_integrity_register(mddev);
1781
1782 abort:
1783
1784         print_conf(conf);
1785         return err;
1786 }
1787
1788
1789 static void end_sync_read(struct bio *bio, int error)
1790 {
1791         struct r10bio *r10_bio = bio->bi_private;
1792         struct r10conf *conf = r10_bio->mddev->private;
1793         int d;
1794
1795         if (bio == r10_bio->master_bio) {
1796                 /* this is a reshape read */
1797                 d = r10_bio->read_slot; /* really the read dev */
1798         } else
1799                 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1800
1801         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1802                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1803         else
1804                 /* The write handler will notice the lack of
1805                  * R10BIO_Uptodate and record any errors etc
1806                  */
1807                 atomic_add(r10_bio->sectors,
1808                            &conf->mirrors[d].rdev->corrected_errors);
1809
1810         /* for reconstruct, we always reschedule after a read.
1811          * for resync, only after all reads
1812          */
1813         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1814         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1815             atomic_dec_and_test(&r10_bio->remaining)) {
1816                 /* we have read all the blocks,
1817                  * do the comparison in process context in raid10d
1818                  */
1819                 reschedule_retry(r10_bio);
1820         }
1821 }
1822
1823 static void end_sync_request(struct r10bio *r10_bio)
1824 {
1825         struct mddev *mddev = r10_bio->mddev;
1826
1827         while (atomic_dec_and_test(&r10_bio->remaining)) {
1828                 if (r10_bio->master_bio == NULL) {
1829                         /* the primary of several recovery bios */
1830                         sector_t s = r10_bio->sectors;
1831                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1832                             test_bit(R10BIO_WriteError, &r10_bio->state))
1833                                 reschedule_retry(r10_bio);
1834                         else
1835                                 put_buf(r10_bio);
1836                         md_done_sync(mddev, s, 1);
1837                         break;
1838                 } else {
1839                         struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1840                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1841                             test_bit(R10BIO_WriteError, &r10_bio->state))
1842                                 reschedule_retry(r10_bio);
1843                         else
1844                                 put_buf(r10_bio);
1845                         r10_bio = r10_bio2;
1846                 }
1847         }
1848 }
1849
1850 static void end_sync_write(struct bio *bio, int error)
1851 {
1852         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1853         struct r10bio *r10_bio = bio->bi_private;
1854         struct mddev *mddev = r10_bio->mddev;
1855         struct r10conf *conf = mddev->private;
1856         int d;
1857         sector_t first_bad;
1858         int bad_sectors;
1859         int slot;
1860         int repl;
1861         struct md_rdev *rdev = NULL;
1862
1863         d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1864         if (repl)
1865                 rdev = conf->mirrors[d].replacement;
1866         else
1867                 rdev = conf->mirrors[d].rdev;
1868
1869         if (!uptodate) {
1870                 if (repl)
1871                         md_error(mddev, rdev);
1872                 else {
1873                         set_bit(WriteErrorSeen, &rdev->flags);
1874                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
1875                                 set_bit(MD_RECOVERY_NEEDED,
1876                                         &rdev->mddev->recovery);
1877                         set_bit(R10BIO_WriteError, &r10_bio->state);
1878                 }
1879         } else if (is_badblock(rdev,
1880                              r10_bio->devs[slot].addr,
1881                              r10_bio->sectors,
1882                              &first_bad, &bad_sectors))
1883                 set_bit(R10BIO_MadeGood, &r10_bio->state);
1884
1885         rdev_dec_pending(rdev, mddev);
1886
1887         end_sync_request(r10_bio);
1888 }
1889
1890 /*
1891  * Note: sync and recover and handled very differently for raid10
1892  * This code is for resync.
1893  * For resync, we read through virtual addresses and read all blocks.
1894  * If there is any error, we schedule a write.  The lowest numbered
1895  * drive is authoritative.
1896  * However requests come for physical address, so we need to map.
1897  * For every physical address there are raid_disks/copies virtual addresses,
1898  * which is always are least one, but is not necessarly an integer.
1899  * This means that a physical address can span multiple chunks, so we may
1900  * have to submit multiple io requests for a single sync request.
1901  */
1902 /*
1903  * We check if all blocks are in-sync and only write to blocks that
1904  * aren't in sync
1905  */
1906 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1907 {
1908         struct r10conf *conf = mddev->private;
1909         int i, first;
1910         struct bio *tbio, *fbio;
1911         int vcnt;
1912
1913         atomic_set(&r10_bio->remaining, 1);
1914
1915         /* find the first device with a block */
1916         for (i=0; i<conf->copies; i++)
1917                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1918                         break;
1919
1920         if (i == conf->copies)
1921                 goto done;
1922
1923         first = i;
1924         fbio = r10_bio->devs[i].bio;
1925
1926         vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
1927         /* now find blocks with errors */
1928         for (i=0 ; i < conf->copies ; i++) {
1929                 int  j, d;
1930
1931                 tbio = r10_bio->devs[i].bio;
1932
1933                 if (tbio->bi_end_io != end_sync_read)
1934                         continue;
1935                 if (i == first)
1936                         continue;
1937                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1938                         /* We know that the bi_io_vec layout is the same for
1939                          * both 'first' and 'i', so we just compare them.
1940                          * All vec entries are PAGE_SIZE;
1941                          */
1942                         for (j = 0; j < vcnt; j++)
1943                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1944                                            page_address(tbio->bi_io_vec[j].bv_page),
1945                                            fbio->bi_io_vec[j].bv_len))
1946                                         break;
1947                         if (j == vcnt)
1948                                 continue;
1949                         mddev->resync_mismatches += r10_bio->sectors;
1950                         if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1951                                 /* Don't fix anything. */
1952                                 continue;
1953                 }
1954                 /* Ok, we need to write this bio, either to correct an
1955                  * inconsistency or to correct an unreadable block.
1956                  * First we need to fixup bv_offset, bv_len and
1957                  * bi_vecs, as the read request might have corrupted these
1958                  */
1959                 tbio->bi_vcnt = vcnt;
1960                 tbio->bi_size = r10_bio->sectors << 9;
1961                 tbio->bi_idx = 0;
1962                 tbio->bi_phys_segments = 0;
1963                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1964                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1965                 tbio->bi_next = NULL;
1966                 tbio->bi_rw = WRITE;
1967                 tbio->bi_private = r10_bio;
1968                 tbio->bi_sector = r10_bio->devs[i].addr;
1969
1970                 for (j=0; j < vcnt ; j++) {
1971                         tbio->bi_io_vec[j].bv_offset = 0;
1972                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1973
1974                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1975                                page_address(fbio->bi_io_vec[j].bv_page),
1976                                PAGE_SIZE);
1977                 }
1978                 tbio->bi_end_io = end_sync_write;
1979
1980                 d = r10_bio->devs[i].devnum;
1981                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1982                 atomic_inc(&r10_bio->remaining);
1983                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1984
1985                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1986                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1987                 generic_make_request(tbio);
1988         }
1989
1990         /* Now write out to any replacement devices
1991          * that are active
1992          */
1993         for (i = 0; i < conf->copies; i++) {
1994                 int j, d;
1995
1996                 tbio = r10_bio->devs[i].repl_bio;
1997                 if (!tbio || !tbio->bi_end_io)
1998                         continue;
1999                 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2000                     && r10_bio->devs[i].bio != fbio)
2001                         for (j = 0; j < vcnt; j++)
2002                                 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2003                                        page_address(fbio->bi_io_vec[j].bv_page),
2004                                        PAGE_SIZE);
2005                 d = r10_bio->devs[i].devnum;
2006                 atomic_inc(&r10_bio->remaining);
2007                 md_sync_acct(conf->mirrors[d].replacement->bdev,
2008                              tbio->bi_size >> 9);
2009                 generic_make_request(tbio);
2010         }
2011
2012 done:
2013         if (atomic_dec_and_test(&r10_bio->remaining)) {
2014                 md_done_sync(mddev, r10_bio->sectors, 1);
2015                 put_buf(r10_bio);
2016         }
2017 }
2018
2019 /*
2020  * Now for the recovery code.
2021  * Recovery happens across physical sectors.
2022  * We recover all non-is_sync drives by finding the virtual address of
2023  * each, and then choose a working drive that also has that virt address.
2024  * There is a separate r10_bio for each non-in_sync drive.
2025  * Only the first two slots are in use. The first for reading,
2026  * The second for writing.
2027  *
2028  */
2029 static void fix_recovery_read_error(struct r10bio *r10_bio)
2030 {
2031         /* We got a read error during recovery.
2032          * We repeat the read in smaller page-sized sections.
2033          * If a read succeeds, write it to the new device or record
2034          * a bad block if we cannot.
2035          * If a read fails, record a bad block on both old and
2036          * new devices.
2037          */
2038         struct mddev *mddev = r10_bio->mddev;
2039         struct r10conf *conf = mddev->private;
2040         struct bio *bio = r10_bio->devs[0].bio;
2041         sector_t sect = 0;
2042         int sectors = r10_bio->sectors;
2043         int idx = 0;
2044         int dr = r10_bio->devs[0].devnum;
2045         int dw = r10_bio->devs[1].devnum;
2046
2047         while (sectors) {
2048                 int s = sectors;
2049                 struct md_rdev *rdev;
2050                 sector_t addr;
2051                 int ok;
2052
2053                 if (s > (PAGE_SIZE>>9))
2054                         s = PAGE_SIZE >> 9;
2055
2056                 rdev = conf->mirrors[dr].rdev;
2057                 addr = r10_bio->devs[0].addr + sect,
2058                 ok = sync_page_io(rdev,
2059                                   addr,
2060                                   s << 9,
2061                                   bio->bi_io_vec[idx].bv_page,
2062                                   READ, false);
2063                 if (ok) {
2064                         rdev = conf->mirrors[dw].rdev;
2065                         addr = r10_bio->devs[1].addr + sect;
2066                         ok = sync_page_io(rdev,
2067                                           addr,
2068                                           s << 9,
2069                                           bio->bi_io_vec[idx].bv_page,
2070                                           WRITE, false);
2071                         if (!ok) {
2072                                 set_bit(WriteErrorSeen, &rdev->flags);
2073                                 if (!test_and_set_bit(WantReplacement,
2074                                                       &rdev->flags))
2075                                         set_bit(MD_RECOVERY_NEEDED,
2076                                                 &rdev->mddev->recovery);
2077                         }
2078                 }
2079                 if (!ok) {
2080                         /* We don't worry if we cannot set a bad block -
2081                          * it really is bad so there is no loss in not
2082                          * recording it yet
2083                          */
2084                         rdev_set_badblocks(rdev, addr, s, 0);
2085
2086                         if (rdev != conf->mirrors[dw].rdev) {
2087                                 /* need bad block on destination too */
2088                                 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2089                                 addr = r10_bio->devs[1].addr + sect;
2090                                 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2091                                 if (!ok) {
2092                                         /* just abort the recovery */
2093                                         printk(KERN_NOTICE
2094                                                "md/raid10:%s: recovery aborted"
2095                                                " due to read error\n",
2096                                                mdname(mddev));
2097
2098                                         conf->mirrors[dw].recovery_disabled
2099                                                 = mddev->recovery_disabled;
2100                                         set_bit(MD_RECOVERY_INTR,
2101                                                 &mddev->recovery);
2102                                         break;
2103                                 }
2104                         }
2105                 }
2106
2107                 sectors -= s;
2108                 sect += s;
2109                 idx++;
2110         }
2111 }
2112
2113 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2114 {
2115         struct r10conf *conf = mddev->private;
2116         int d;
2117         struct bio *wbio, *wbio2;
2118
2119         if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2120                 fix_recovery_read_error(r10_bio);
2121                 end_sync_request(r10_bio);
2122                 return;
2123         }
2124
2125         /*
2126          * share the pages with the first bio
2127          * and submit the write request
2128          */
2129         d = r10_bio->devs[1].devnum;
2130         wbio = r10_bio->devs[1].bio;
2131         wbio2 = r10_bio->devs[1].repl_bio;
2132         if (wbio->bi_end_io) {
2133                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2134                 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
2135                 generic_make_request(wbio);
2136         }
2137         if (wbio2 && wbio2->bi_end_io) {
2138                 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2139                 md_sync_acct(conf->mirrors[d].replacement->bdev,
2140                              wbio2->bi_size >> 9);
2141                 generic_make_request(wbio2);
2142         }
2143 }
2144
2145
2146 /*
2147  * Used by fix_read_error() to decay the per rdev read_errors.
2148  * We halve the read error count for every hour that has elapsed
2149  * since the last recorded read error.
2150  *
2151  */
2152 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2153 {
2154         struct timespec cur_time_mon;
2155         unsigned long hours_since_last;
2156         unsigned int read_errors = atomic_read(&rdev->read_errors);
2157
2158         ktime_get_ts(&cur_time_mon);
2159
2160         if (rdev->last_read_error.tv_sec == 0 &&
2161             rdev->last_read_error.tv_nsec == 0) {
2162                 /* first time we've seen a read error */
2163                 rdev->last_read_error = cur_time_mon;
2164                 return;
2165         }
2166
2167         hours_since_last = (cur_time_mon.tv_sec -
2168                             rdev->last_read_error.tv_sec) / 3600;
2169
2170         rdev->last_read_error = cur_time_mon;
2171
2172         /*
2173          * if hours_since_last is > the number of bits in read_errors
2174          * just set read errors to 0. We do this to avoid
2175          * overflowing the shift of read_errors by hours_since_last.
2176          */
2177         if (hours_since_last >= 8 * sizeof(read_errors))
2178                 atomic_set(&rdev->read_errors, 0);
2179         else
2180                 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2181 }
2182
2183 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2184                             int sectors, struct page *page, int rw)
2185 {
2186         sector_t first_bad;
2187         int bad_sectors;
2188
2189         if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2190             && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2191                 return -1;
2192         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2193                 /* success */
2194                 return 1;
2195         if (rw == WRITE) {
2196                 set_bit(WriteErrorSeen, &rdev->flags);
2197                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2198                         set_bit(MD_RECOVERY_NEEDED,
2199                                 &rdev->mddev->recovery);
2200         }
2201         /* need to record an error - either for the block or the device */
2202         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2203                 md_error(rdev->mddev, rdev);
2204         return 0;
2205 }
2206
2207 /*
2208  * This is a kernel thread which:
2209  *
2210  *      1.      Retries failed read operations on working mirrors.
2211  *      2.      Updates the raid superblock when problems encounter.
2212  *      3.      Performs writes following reads for array synchronising.
2213  */
2214
2215 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2216 {
2217         int sect = 0; /* Offset from r10_bio->sector */
2218         int sectors = r10_bio->sectors;
2219         struct md_rdev*rdev;
2220         int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2221         int d = r10_bio->devs[r10_bio->read_slot].devnum;
2222
2223         /* still own a reference to this rdev, so it cannot
2224          * have been cleared recently.
2225          */
2226         rdev = conf->mirrors[d].rdev;
2227
2228         if (test_bit(Faulty, &rdev->flags))
2229                 /* drive has already been failed, just ignore any
2230                    more fix_read_error() attempts */
2231                 return;
2232
2233         check_decay_read_errors(mddev, rdev);
2234         atomic_inc(&rdev->read_errors);
2235         if (atomic_read(&rdev->read_errors) > max_read_errors) {
2236                 char b[BDEVNAME_SIZE];
2237                 bdevname(rdev->bdev, b);
2238
2239                 printk(KERN_NOTICE
2240                        "md/raid10:%s: %s: Raid device exceeded "
2241                        "read_error threshold [cur %d:max %d]\n",
2242                        mdname(mddev), b,
2243                        atomic_read(&rdev->read_errors), max_read_errors);
2244                 printk(KERN_NOTICE
2245                        "md/raid10:%s: %s: Failing raid device\n",
2246                        mdname(mddev), b);
2247                 md_error(mddev, conf->mirrors[d].rdev);
2248                 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2249                 return;
2250         }
2251
2252         while(sectors) {
2253                 int s = sectors;
2254                 int sl = r10_bio->read_slot;
2255                 int success = 0;
2256                 int start;
2257
2258                 if (s > (PAGE_SIZE>>9))
2259                         s = PAGE_SIZE >> 9;
2260
2261                 rcu_read_lock();
2262                 do {
2263                         sector_t first_bad;
2264                         int bad_sectors;
2265
2266                         d = r10_bio->devs[sl].devnum;
2267                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2268                         if (rdev &&
2269                             !test_bit(Unmerged, &rdev->flags) &&
2270                             test_bit(In_sync, &rdev->flags) &&
2271                             is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2272                                         &first_bad, &bad_sectors) == 0) {
2273                                 atomic_inc(&rdev->nr_pending);
2274                                 rcu_read_unlock();
2275                                 success = sync_page_io(rdev,
2276                                                        r10_bio->devs[sl].addr +
2277                                                        sect,
2278                                                        s<<9,
2279                                                        conf->tmppage, READ, false);
2280                                 rdev_dec_pending(rdev, mddev);
2281                                 rcu_read_lock();
2282                                 if (success)
2283                                         break;
2284                         }
2285                         sl++;
2286                         if (sl == conf->copies)
2287                                 sl = 0;
2288                 } while (!success && sl != r10_bio->read_slot);
2289                 rcu_read_unlock();
2290
2291                 if (!success) {
2292                         /* Cannot read from anywhere, just mark the block
2293                          * as bad on the first device to discourage future
2294                          * reads.
2295                          */
2296                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2297                         rdev = conf->mirrors[dn].rdev;
2298
2299                         if (!rdev_set_badblocks(
2300                                     rdev,
2301                                     r10_bio->devs[r10_bio->read_slot].addr
2302                                     + sect,
2303                                     s, 0)) {
2304                                 md_error(mddev, rdev);
2305                                 r10_bio->devs[r10_bio->read_slot].bio
2306                                         = IO_BLOCKED;
2307                         }
2308                         break;
2309                 }
2310
2311                 start = sl;
2312                 /* write it back and re-read */
2313                 rcu_read_lock();
2314                 while (sl != r10_bio->read_slot) {
2315                         char b[BDEVNAME_SIZE];
2316
2317                         if (sl==0)
2318                                 sl = conf->copies;
2319                         sl--;
2320                         d = r10_bio->devs[sl].devnum;
2321                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2322                         if (!rdev ||
2323                             test_bit(Unmerged, &rdev->flags) ||
2324                             !test_bit(In_sync, &rdev->flags))
2325                                 continue;
2326
2327                         atomic_inc(&rdev->nr_pending);
2328                         rcu_read_unlock();
2329                         if (r10_sync_page_io(rdev,
2330                                              r10_bio->devs[sl].addr +
2331                                              sect,
2332                                              s, conf->tmppage, WRITE)
2333                             == 0) {
2334                                 /* Well, this device is dead */
2335                                 printk(KERN_NOTICE
2336                                        "md/raid10:%s: read correction "
2337                                        "write failed"
2338                                        " (%d sectors at %llu on %s)\n",
2339                                        mdname(mddev), s,
2340                                        (unsigned long long)(
2341                                                sect +
2342                                                choose_data_offset(r10_bio,
2343                                                                   rdev)),
2344                                        bdevname(rdev->bdev, b));
2345                                 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2346                                        "drive\n",
2347                                        mdname(mddev),
2348                                        bdevname(rdev->bdev, b));
2349                         }
2350                         rdev_dec_pending(rdev, mddev);
2351                         rcu_read_lock();
2352                 }
2353                 sl = start;
2354                 while (sl != r10_bio->read_slot) {
2355                         char b[BDEVNAME_SIZE];
2356
2357                         if (sl==0)
2358                                 sl = conf->copies;
2359                         sl--;
2360                         d = r10_bio->devs[sl].devnum;
2361                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2362                         if (!rdev ||
2363                             !test_bit(In_sync, &rdev->flags))
2364                                 continue;
2365
2366                         atomic_inc(&rdev->nr_pending);
2367                         rcu_read_unlock();
2368                         switch (r10_sync_page_io(rdev,
2369                                              r10_bio->devs[sl].addr +
2370                                              sect,
2371                                              s, conf->tmppage,
2372                                                  READ)) {
2373                         case 0:
2374                                 /* Well, this device is dead */
2375                                 printk(KERN_NOTICE
2376                                        "md/raid10:%s: unable to read back "
2377                                        "corrected sectors"
2378                                        " (%d sectors at %llu on %s)\n",
2379                                        mdname(mddev), s,
2380                                        (unsigned long long)(
2381                                                sect +
2382                                                choose_data_offset(r10_bio, rdev)),
2383                                        bdevname(rdev->bdev, b));
2384                                 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2385                                        "drive\n",
2386                                        mdname(mddev),
2387                                        bdevname(rdev->bdev, b));
2388                                 break;
2389                         case 1:
2390                                 printk(KERN_INFO
2391                                        "md/raid10:%s: read error corrected"
2392                                        " (%d sectors at %llu on %s)\n",
2393                                        mdname(mddev), s,
2394                                        (unsigned long long)(
2395                                                sect +
2396                                                choose_data_offset(r10_bio, rdev)),
2397                                        bdevname(rdev->bdev, b));
2398                                 atomic_add(s, &rdev->corrected_errors);
2399                         }
2400
2401                         rdev_dec_pending(rdev, mddev);
2402                         rcu_read_lock();
2403                 }
2404                 rcu_read_unlock();
2405
2406                 sectors -= s;
2407                 sect += s;
2408         }
2409 }
2410
2411 static void bi_complete(struct bio *bio, int error)
2412 {
2413         complete((struct completion *)bio->bi_private);
2414 }
2415
2416 static int submit_bio_wait(int rw, struct bio *bio)
2417 {
2418         struct completion event;
2419         rw |= REQ_SYNC;
2420
2421         init_completion(&event);
2422         bio->bi_private = &event;
2423         bio->bi_end_io = bi_complete;
2424         submit_bio(rw, bio);
2425         wait_for_completion(&event);
2426
2427         return test_bit(BIO_UPTODATE, &bio->bi_flags);
2428 }
2429
2430 static int narrow_write_error(struct r10bio *r10_bio, int i)
2431 {
2432         struct bio *bio = r10_bio->master_bio;
2433         struct mddev *mddev = r10_bio->mddev;
2434         struct r10conf *conf = mddev->private;
2435         struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2436         /* bio has the data to be written to slot 'i' where
2437          * we just recently had a write error.
2438          * We repeatedly clone the bio and trim down to one block,
2439          * then try the write.  Where the write fails we record
2440          * a bad block.
2441          * It is conceivable that the bio doesn't exactly align with
2442          * blocks.  We must handle this.
2443          *
2444          * We currently own a reference to the rdev.
2445          */
2446
2447         int block_sectors;
2448         sector_t sector;
2449         int sectors;
2450         int sect_to_write = r10_bio->sectors;
2451         int ok = 1;
2452
2453         if (rdev->badblocks.shift < 0)
2454                 return 0;
2455
2456         block_sectors = 1 << rdev->badblocks.shift;
2457         sector = r10_bio->sector;
2458         sectors = ((r10_bio->sector + block_sectors)
2459                    & ~(sector_t)(block_sectors - 1))
2460                 - sector;
2461
2462         while (sect_to_write) {
2463                 struct bio *wbio;
2464                 if (sectors > sect_to_write)
2465                         sectors = sect_to_write;
2466                 /* Write at 'sector' for 'sectors' */
2467                 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2468                 md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2469                 wbio->bi_sector = (r10_bio->devs[i].addr+
2470                                    choose_data_offset(r10_bio, rdev) +
2471                                    (sector - r10_bio->sector));
2472                 wbio->bi_bdev = rdev->bdev;
2473                 if (submit_bio_wait(WRITE, wbio) == 0)
2474                         /* Failure! */
2475                         ok = rdev_set_badblocks(rdev, sector,
2476                                                 sectors, 0)
2477                                 && ok;
2478
2479                 bio_put(wbio);
2480                 sect_to_write -= sectors;
2481                 sector += sectors;
2482                 sectors = block_sectors;
2483         }
2484         return ok;
2485 }
2486
2487 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2488 {
2489         int slot = r10_bio->read_slot;
2490         struct bio *bio;
2491         struct r10conf *conf = mddev->private;
2492         struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2493         char b[BDEVNAME_SIZE];
2494         unsigned long do_sync;
2495         int max_sectors;
2496
2497         /* we got a read error. Maybe the drive is bad.  Maybe just
2498          * the block and we can fix it.
2499          * We freeze all other IO, and try reading the block from
2500          * other devices.  When we find one, we re-write
2501          * and check it that fixes the read error.
2502          * This is all done synchronously while the array is
2503          * frozen.
2504          */
2505         bio = r10_bio->devs[slot].bio;
2506         bdevname(bio->bi_bdev, b);
2507         bio_put(bio);
2508         r10_bio->devs[slot].bio = NULL;
2509
2510         if (mddev->ro == 0) {
2511                 freeze_array(conf);
2512                 fix_read_error(conf, mddev, r10_bio);
2513                 unfreeze_array(conf);
2514         } else
2515                 r10_bio->devs[slot].bio = IO_BLOCKED;
2516
2517         rdev_dec_pending(rdev, mddev);
2518
2519 read_more:
2520         rdev = read_balance(conf, r10_bio, &max_sectors);
2521         if (rdev == NULL) {
2522                 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2523                        " read error for block %llu\n",
2524                        mdname(mddev), b,
2525                        (unsigned long long)r10_bio->sector);
2526                 raid_end_bio_io(r10_bio);
2527                 return;
2528         }
2529
2530         do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2531         slot = r10_bio->read_slot;
2532         printk_ratelimited(
2533                 KERN_ERR
2534                 "md/raid10:%s: %s: redirecting "
2535                 "sector %llu to another mirror\n",
2536                 mdname(mddev),
2537                 bdevname(rdev->bdev, b),
2538                 (unsigned long long)r10_bio->sector);
2539         bio = bio_clone_mddev(r10_bio->master_bio,
2540                               GFP_NOIO, mddev);
2541         md_trim_bio(bio,
2542                     r10_bio->sector - bio->bi_sector,
2543                     max_sectors);
2544         r10_bio->devs[slot].bio = bio;
2545         r10_bio->devs[slot].rdev = rdev;
2546         bio->bi_sector = r10_bio->devs[slot].addr
2547                 + choose_data_offset(r10_bio, rdev);
2548         bio->bi_bdev = rdev->bdev;
2549         bio->bi_rw = READ | do_sync;
2550         bio->bi_private = r10_bio;
2551         bio->bi_end_io = raid10_end_read_request;
2552         if (max_sectors < r10_bio->sectors) {
2553                 /* Drat - have to split this up more */
2554                 struct bio *mbio = r10_bio->master_bio;
2555                 int sectors_handled =
2556                         r10_bio->sector + max_sectors
2557                         - mbio->bi_sector;
2558                 r10_bio->sectors = max_sectors;
2559                 spin_lock_irq(&conf->device_lock);
2560                 if (mbio->bi_phys_segments == 0)
2561                         mbio->bi_phys_segments = 2;
2562                 else
2563                         mbio->bi_phys_segments++;
2564                 spin_unlock_irq(&conf->device_lock);
2565                 generic_make_request(bio);
2566
2567                 r10_bio = mempool_alloc(conf->r10bio_pool,
2568                                         GFP_NOIO);
2569                 r10_bio->master_bio = mbio;
2570                 r10_bio->sectors = (mbio->bi_size >> 9)
2571                         - sectors_handled;
2572                 r10_bio->state = 0;
2573                 set_bit(R10BIO_ReadError,
2574                         &r10_bio->state);
2575                 r10_bio->mddev = mddev;
2576                 r10_bio->sector = mbio->bi_sector
2577                         + sectors_handled;
2578
2579                 goto read_more;
2580         } else
2581                 generic_make_request(bio);
2582 }
2583
2584 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2585 {
2586         /* Some sort of write request has finished and it
2587          * succeeded in writing where we thought there was a
2588          * bad block.  So forget the bad block.
2589          * Or possibly if failed and we need to record
2590          * a bad block.
2591          */
2592         int m;
2593         struct md_rdev *rdev;
2594
2595         if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2596             test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2597                 for (m = 0; m < conf->copies; m++) {
2598                         int dev = r10_bio->devs[m].devnum;
2599                         rdev = conf->mirrors[dev].rdev;
2600                         if (r10_bio->devs[m].bio == NULL)
2601                                 continue;
2602                         if (test_bit(BIO_UPTODATE,
2603                                      &r10_bio->devs[m].bio->bi_flags)) {
2604                                 rdev_clear_badblocks(
2605                                         rdev,
2606                                         r10_bio->devs[m].addr,
2607                                         r10_bio->sectors, 0);
2608                         } else {
2609                                 if (!rdev_set_badblocks(
2610                                             rdev,
2611                                             r10_bio->devs[m].addr,
2612                                             r10_bio->sectors, 0))
2613                                         md_error(conf->mddev, rdev);
2614                         }
2615                         rdev = conf->mirrors[dev].replacement;
2616                         if (r10_bio->devs[m].repl_bio == NULL)
2617                                 continue;
2618                         if (test_bit(BIO_UPTODATE,
2619                                      &r10_bio->devs[m].repl_bio->bi_flags)) {
2620                                 rdev_clear_badblocks(
2621                                         rdev,
2622                                         r10_bio->devs[m].addr,
2623                                         r10_bio->sectors, 0);
2624                         } else {
2625                                 if (!rdev_set_badblocks(
2626                                             rdev,
2627                                             r10_bio->devs[m].addr,
2628                                             r10_bio->sectors, 0))
2629                                         md_error(conf->mddev, rdev);
2630                         }
2631                 }
2632                 put_buf(r10_bio);
2633         } else {
2634                 for (m = 0; m < conf->copies; m++) {
2635                         int dev = r10_bio->devs[m].devnum;
2636                         struct bio *bio = r10_bio->devs[m].bio;
2637                         rdev = conf->mirrors[dev].rdev;
2638                         if (bio == IO_MADE_GOOD) {
2639                                 rdev_clear_badblocks(
2640                                         rdev,
2641                                         r10_bio->devs[m].addr,
2642                                         r10_bio->sectors, 0);
2643                                 rdev_dec_pending(rdev, conf->mddev);
2644                         } else if (bio != NULL &&
2645                                    !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2646                                 if (!narrow_write_error(r10_bio, m)) {
2647                                         md_error(conf->mddev, rdev);
2648                                         set_bit(R10BIO_Degraded,
2649                                                 &r10_bio->state);
2650                                 }
2651                                 rdev_dec_pending(rdev, conf->mddev);
2652                         }
2653                         bio = r10_bio->devs[m].repl_bio;
2654                         rdev = conf->mirrors[dev].replacement;
2655                         if (rdev && bio == IO_MADE_GOOD) {
2656                                 rdev_clear_badblocks(
2657                                         rdev,
2658                                         r10_bio->devs[m].addr,
2659                                         r10_bio->sectors, 0);
2660                                 rdev_dec_pending(rdev, conf->mddev);
2661                         }
2662                 }
2663                 if (test_bit(R10BIO_WriteError,
2664                              &r10_bio->state))
2665                         close_write(r10_bio);
2666                 raid_end_bio_io(r10_bio);
2667         }
2668 }
2669
2670 static void raid10d(struct mddev *mddev)
2671 {
2672         struct r10bio *r10_bio;
2673         unsigned long flags;
2674         struct r10conf *conf = mddev->private;
2675         struct list_head *head = &conf->retry_list;
2676         struct blk_plug plug;
2677
2678         md_check_recovery(mddev);
2679
2680         blk_start_plug(&plug);
2681         for (;;) {
2682
2683                 flush_pending_writes(conf);
2684
2685                 spin_lock_irqsave(&conf->device_lock, flags);
2686                 if (list_empty(head)) {
2687                         spin_unlock_irqrestore(&conf->device_lock, flags);
2688                         break;
2689                 }
2690                 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2691                 list_del(head->prev);
2692                 conf->nr_queued--;
2693                 spin_unlock_irqrestore(&conf->device_lock, flags);
2694
2695                 mddev = r10_bio->mddev;
2696                 conf = mddev->private;
2697                 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2698                     test_bit(R10BIO_WriteError, &r10_bio->state))
2699                         handle_write_completed(conf, r10_bio);
2700                 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2701                         reshape_request_write(mddev, r10_bio);
2702                 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2703                         sync_request_write(mddev, r10_bio);
2704                 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2705                         recovery_request_write(mddev, r10_bio);
2706                 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2707                         handle_read_error(mddev, r10_bio);
2708                 else {
2709                         /* just a partial read to be scheduled from a
2710                          * separate context
2711                          */
2712                         int slot = r10_bio->read_slot;
2713                         generic_make_request(r10_bio->devs[slot].bio);
2714                 }
2715
2716                 cond_resched();
2717                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2718                         md_check_recovery(mddev);
2719         }
2720         blk_finish_plug(&plug);
2721 }
2722
2723
2724 static int init_resync(struct r10conf *conf)
2725 {
2726         int buffs;
2727         int i;
2728
2729         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2730         BUG_ON(conf->r10buf_pool);
2731         conf->have_replacement = 0;
2732         for (i = 0; i < conf->geo.raid_disks; i++)
2733                 if (conf->mirrors[i].replacement)
2734                         conf->have_replacement = 1;
2735         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2736         if (!conf->r10buf_pool)
2737                 return -ENOMEM;
2738         conf->next_resync = 0;
2739         return 0;
2740 }
2741
2742 /*
2743  * perform a "sync" on one "block"
2744  *
2745  * We need to make sure that no normal I/O request - particularly write
2746  * requests - conflict with active sync requests.
2747  *
2748  * This is achieved by tracking pending requests and a 'barrier' concept
2749  * that can be installed to exclude normal IO requests.
2750  *
2751  * Resync and recovery are handled very differently.
2752  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2753  *
2754  * For resync, we iterate over virtual addresses, read all copies,
2755  * and update if there are differences.  If only one copy is live,
2756  * skip it.
2757  * For recovery, we iterate over physical addresses, read a good
2758  * value for each non-in_sync drive, and over-write.
2759  *
2760  * So, for recovery we may have several outstanding complex requests for a
2761  * given address, one for each out-of-sync device.  We model this by allocating
2762  * a number of r10_bio structures, one for each out-of-sync device.
2763  * As we setup these structures, we collect all bio's together into a list
2764  * which we then process collectively to add pages, and then process again
2765  * to pass to generic_make_request.
2766  *
2767  * The r10_bio structures are linked using a borrowed master_bio pointer.
2768  * This link is counted in ->remaining.  When the r10_bio that points to NULL
2769  * has its remaining count decremented to 0, the whole complex operation
2770  * is complete.
2771  *
2772  */
2773
2774 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
2775                              int *skipped, int go_faster)
2776 {
2777         struct r10conf *conf = mddev->private;
2778         struct r10bio *r10_bio;
2779         struct bio *biolist = NULL, *bio;
2780         sector_t max_sector, nr_sectors;
2781         int i;
2782         int max_sync;
2783         sector_t sync_blocks;
2784         sector_t sectors_skipped = 0;
2785         int chunks_skipped = 0;
2786         sector_t chunk_mask = conf->geo.chunk_mask;
2787
2788         if (!conf->r10buf_pool)
2789                 if (init_resync(conf))
2790                         return 0;
2791
2792  skipped:
2793         max_sector = mddev->dev_sectors;
2794         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2795             test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2796                 max_sector = mddev->resync_max_sectors;
2797         if (sector_nr >= max_sector) {
2798                 /* If we aborted, we need to abort the
2799                  * sync on the 'current' bitmap chucks (there can
2800                  * be several when recovering multiple devices).
2801                  * as we may have started syncing it but not finished.
2802                  * We can find the current address in
2803                  * mddev->curr_resync, but for recovery,
2804                  * we need to convert that to several
2805                  * virtual addresses.
2806                  */
2807                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2808                         end_reshape(conf);
2809                         return 0;
2810                 }
2811
2812                 if (mddev->curr_resync < max_sector) { /* aborted */
2813                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2814                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2815                                                 &sync_blocks, 1);
2816                         else for (i = 0; i < conf->geo.raid_disks; i++) {
2817                                 sector_t sect =
2818                                         raid10_find_virt(conf, mddev->curr_resync, i);
2819                                 bitmap_end_sync(mddev->bitmap, sect,
2820                                                 &sync_blocks, 1);
2821                         }
2822                 } else {
2823                         /* completed sync */
2824                         if ((!mddev->bitmap || conf->fullsync)
2825                             && conf->have_replacement
2826                             && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2827                                 /* Completed a full sync so the replacements
2828                                  * are now fully recovered.
2829                                  */
2830                                 for (i = 0; i < conf->geo.raid_disks; i++)
2831                                         if (conf->mirrors[i].replacement)
2832                                                 conf->mirrors[i].replacement
2833                                                         ->recovery_offset
2834                                                         = MaxSector;
2835                         }
2836                         conf->fullsync = 0;
2837                 }
2838                 bitmap_close_sync(mddev->bitmap);
2839                 close_sync(conf);
2840                 *skipped = 1;
2841                 return sectors_skipped;
2842         }
2843
2844         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2845                 return reshape_request(mddev, sector_nr, skipped);
2846
2847         if (chunks_skipped >= conf->geo.raid_disks) {
2848                 /* if there has been nothing to do on any drive,
2849                  * then there is nothing to do at all..
2850                  */
2851                 *skipped = 1;
2852                 return (max_sector - sector_nr) + sectors_skipped;
2853         }
2854
2855         if (max_sector > mddev->resync_max)
2856                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2857
2858         /* make sure whole request will fit in a chunk - if chunks
2859          * are meaningful
2860          */
2861         if (conf->geo.near_copies < conf->geo.raid_disks &&
2862             max_sector > (sector_nr | chunk_mask))
2863                 max_sector = (sector_nr | chunk_mask) + 1;
2864         /*
2865          * If there is non-resync activity waiting for us then
2866          * put in a delay to throttle resync.
2867          */
2868         if (!go_faster && conf->nr_waiting)
2869                 msleep_interruptible(1000);
2870
2871         /* Again, very different code for resync and recovery.
2872          * Both must result in an r10bio with a list of bios that
2873          * have bi_end_io, bi_sector, bi_bdev set,
2874          * and bi_private set to the r10bio.
2875          * For recovery, we may actually create several r10bios
2876          * with 2 bios in each, that correspond to the bios in the main one.
2877          * In this case, the subordinate r10bios link back through a
2878          * borrowed master_bio pointer, and the counter in the master
2879          * includes a ref from each subordinate.
2880          */
2881         /* First, we decide what to do and set ->bi_end_io
2882          * To end_sync_read if we want to read, and
2883          * end_sync_write if we will want to write.
2884          */
2885
2886         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2887         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2888                 /* recovery... the complicated one */
2889                 int j;
2890                 r10_bio = NULL;
2891
2892                 for (i = 0 ; i < conf->geo.raid_disks; i++) {
2893                         int still_degraded;
2894                         struct r10bio *rb2;
2895                         sector_t sect;
2896                         int must_sync;
2897                         int any_working;
2898                         struct raid10_info *mirror = &conf->mirrors[i];
2899
2900                         if ((mirror->rdev == NULL ||
2901                              test_bit(In_sync, &mirror->rdev->flags))
2902                             &&
2903                             (mirror->replacement == NULL ||
2904                              test_bit(Faulty,
2905                                       &mirror->replacement->flags)))
2906                                 continue;
2907
2908                         still_degraded = 0;
2909                         /* want to reconstruct this device */
2910                         rb2 = r10_bio;
2911                         sect = raid10_find_virt(conf, sector_nr, i);
2912                         if (sect >= mddev->resync_max_sectors) {
2913                                 /* last stripe is not complete - don't
2914                                  * try to recover this sector.
2915                                  */
2916                                 continue;
2917                         }
2918                         /* Unless we are doing a full sync, or a replacement
2919                          * we only need to recover the block if it is set in
2920                          * the bitmap
2921                          */
2922                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
2923                                                       &sync_blocks, 1);
2924                         if (sync_blocks < max_sync)
2925                                 max_sync = sync_blocks;
2926                         if (!must_sync &&
2927                             mirror->replacement == NULL &&
2928                             !conf->fullsync) {
2929                                 /* yep, skip the sync_blocks here, but don't assume
2930                                  * that there will never be anything to do here
2931                                  */
2932                                 chunks_skipped = -1;
2933                                 continue;
2934                         }
2935
2936                         r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2937                         raise_barrier(conf, rb2 != NULL);
2938                         atomic_set(&r10_bio->remaining, 0);
2939
2940                         r10_bio->master_bio = (struct bio*)rb2;
2941                         if (rb2)
2942                                 atomic_inc(&rb2->remaining);
2943                         r10_bio->mddev = mddev;
2944                         set_bit(R10BIO_IsRecover, &r10_bio->state);
2945                         r10_bio->sector = sect;
2946
2947                         raid10_find_phys(conf, r10_bio);
2948
2949                         /* Need to check if the array will still be
2950                          * degraded
2951                          */
2952                         for (j = 0; j < conf->geo.raid_disks; j++)
2953                                 if (conf->mirrors[j].rdev == NULL ||
2954                                     test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
2955                                         still_degraded = 1;
2956                                         break;
2957                                 }
2958
2959                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
2960                                                       &sync_blocks, still_degraded);
2961
2962                         any_working = 0;
2963                         for (j=0; j<conf->copies;j++) {
2964                                 int k;
2965                                 int d = r10_bio->devs[j].devnum;
2966                                 sector_t from_addr, to_addr;
2967                                 struct md_rdev *rdev;
2968                                 sector_t sector, first_bad;
2969                                 int bad_sectors;
2970                                 if (!conf->mirrors[d].rdev ||
2971                                     !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
2972                                         continue;
2973                                 /* This is where we read from */
2974                                 any_working = 1;
2975                                 rdev = conf->mirrors[d].rdev;
2976                                 sector = r10_bio->devs[j].addr;
2977
2978                                 if (is_badblock(rdev, sector, max_sync,
2979                                                 &first_bad, &bad_sectors)) {
2980                                         if (first_bad > sector)
2981                                                 max_sync = first_bad - sector;
2982                                         else {
2983                                                 bad_sectors -= (sector
2984                                                                 - first_bad);
2985                                                 if (max_sync > bad_sectors)
2986                                                         max_sync = bad_sectors;
2987                                                 continue;
2988                                         }
2989                                 }
2990                                 bio = r10_bio->devs[0].bio;
2991                                 bio->bi_next = biolist;
2992                                 biolist = bio;
2993                                 bio->bi_private = r10_bio;
2994                                 bio->bi_end_io = end_sync_read;
2995                                 bio->bi_rw = READ;
2996                                 from_addr = r10_bio->devs[j].addr;
2997                                 bio->bi_sector = from_addr + rdev->data_offset;
2998                                 bio->bi_bdev = rdev->bdev;
2999                                 atomic_inc(&rdev->nr_pending);
3000                                 /* and we write to 'i' (if not in_sync) */