md/raid10: unify handling of write completion.
[linux-2.6.git] / drivers / md / raid10.c
1 /*
2  * raid10.c : Multiple Devices driver for Linux
3  *
4  * Copyright (C) 2000-2004 Neil Brown
5  *
6  * RAID-10 support for md.
7  *
8  * Base on code in raid1.c.  See raid1.c for further copyright information.
9  *
10  *
11  * This program is free software; you can redistribute it and/or modify
12  * it under the terms of the GNU General Public License as published by
13  * the Free Software Foundation; either version 2, or (at your option)
14  * any later version.
15  *
16  * You should have received a copy of the GNU General Public License
17  * (for example /usr/src/linux/COPYING); if not, write to the Free
18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19  */
20
21 #include <linux/slab.h>
22 #include <linux/delay.h>
23 #include <linux/blkdev.h>
24 #include <linux/seq_file.h>
25 #include <linux/ratelimit.h>
26 #include "md.h"
27 #include "raid10.h"
28 #include "raid0.h"
29 #include "bitmap.h"
30
31 /*
32  * RAID10 provides a combination of RAID0 and RAID1 functionality.
33  * The layout of data is defined by
34  *    chunk_size
35  *    raid_disks
36  *    near_copies (stored in low byte of layout)
37  *    far_copies (stored in second byte of layout)
38  *    far_offset (stored in bit 16 of layout )
39  *
40  * The data to be stored is divided into chunks using chunksize.
41  * Each device is divided into far_copies sections.
42  * In each section, chunks are laid out in a style similar to raid0, but
43  * near_copies copies of each chunk is stored (each on a different drive).
44  * The starting device for each section is offset near_copies from the starting
45  * device of the previous section.
46  * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
47  * drive.
48  * near_copies and far_copies must be at least one, and their product is at most
49  * raid_disks.
50  *
51  * If far_offset is true, then the far_copies are handled a bit differently.
52  * The copies are still in different stripes, but instead of be very far apart
53  * on disk, there are adjacent stripes.
54  */
55
56 /*
57  * Number of guaranteed r10bios in case of extreme VM load:
58  */
59 #define NR_RAID10_BIOS 256
60
61 static void allow_barrier(conf_t *conf);
62 static void lower_barrier(conf_t *conf);
63
64 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
65 {
66         conf_t *conf = data;
67         int size = offsetof(struct r10bio_s, devs[conf->copies]);
68
69         /* allocate a r10bio with room for raid_disks entries in the bios array */
70         return kzalloc(size, gfp_flags);
71 }
72
73 static void r10bio_pool_free(void *r10_bio, void *data)
74 {
75         kfree(r10_bio);
76 }
77
78 /* Maximum size of each resync request */
79 #define RESYNC_BLOCK_SIZE (64*1024)
80 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
81 /* amount of memory to reserve for resync requests */
82 #define RESYNC_WINDOW (1024*1024)
83 /* maximum number of concurrent requests, memory permitting */
84 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
85
86 /*
87  * When performing a resync, we need to read and compare, so
88  * we need as many pages are there are copies.
89  * When performing a recovery, we need 2 bios, one for read,
90  * one for write (we recover only one drive per r10buf)
91  *
92  */
93 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
94 {
95         conf_t *conf = data;
96         struct page *page;
97         r10bio_t *r10_bio;
98         struct bio *bio;
99         int i, j;
100         int nalloc;
101
102         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
103         if (!r10_bio)
104                 return NULL;
105
106         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
107                 nalloc = conf->copies; /* resync */
108         else
109                 nalloc = 2; /* recovery */
110
111         /*
112          * Allocate bios.
113          */
114         for (j = nalloc ; j-- ; ) {
115                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
116                 if (!bio)
117                         goto out_free_bio;
118                 r10_bio->devs[j].bio = bio;
119         }
120         /*
121          * Allocate RESYNC_PAGES data pages and attach them
122          * where needed.
123          */
124         for (j = 0 ; j < nalloc; j++) {
125                 bio = r10_bio->devs[j].bio;
126                 for (i = 0; i < RESYNC_PAGES; i++) {
127                         if (j == 1 && !test_bit(MD_RECOVERY_SYNC,
128                                                 &conf->mddev->recovery)) {
129                                 /* we can share bv_page's during recovery */
130                                 struct bio *rbio = r10_bio->devs[0].bio;
131                                 page = rbio->bi_io_vec[i].bv_page;
132                                 get_page(page);
133                         } else
134                                 page = alloc_page(gfp_flags);
135                         if (unlikely(!page))
136                                 goto out_free_pages;
137
138                         bio->bi_io_vec[i].bv_page = page;
139                 }
140         }
141
142         return r10_bio;
143
144 out_free_pages:
145         for ( ; i > 0 ; i--)
146                 safe_put_page(bio->bi_io_vec[i-1].bv_page);
147         while (j--)
148                 for (i = 0; i < RESYNC_PAGES ; i++)
149                         safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
150         j = -1;
151 out_free_bio:
152         while ( ++j < nalloc )
153                 bio_put(r10_bio->devs[j].bio);
154         r10bio_pool_free(r10_bio, conf);
155         return NULL;
156 }
157
158 static void r10buf_pool_free(void *__r10_bio, void *data)
159 {
160         int i;
161         conf_t *conf = data;
162         r10bio_t *r10bio = __r10_bio;
163         int j;
164
165         for (j=0; j < conf->copies; j++) {
166                 struct bio *bio = r10bio->devs[j].bio;
167                 if (bio) {
168                         for (i = 0; i < RESYNC_PAGES; i++) {
169                                 safe_put_page(bio->bi_io_vec[i].bv_page);
170                                 bio->bi_io_vec[i].bv_page = NULL;
171                         }
172                         bio_put(bio);
173                 }
174         }
175         r10bio_pool_free(r10bio, conf);
176 }
177
178 static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
179 {
180         int i;
181
182         for (i = 0; i < conf->copies; i++) {
183                 struct bio **bio = & r10_bio->devs[i].bio;
184                 if (!BIO_SPECIAL(*bio))
185                         bio_put(*bio);
186                 *bio = NULL;
187         }
188 }
189
190 static void free_r10bio(r10bio_t *r10_bio)
191 {
192         conf_t *conf = r10_bio->mddev->private;
193
194         put_all_bios(conf, r10_bio);
195         mempool_free(r10_bio, conf->r10bio_pool);
196 }
197
198 static void put_buf(r10bio_t *r10_bio)
199 {
200         conf_t *conf = r10_bio->mddev->private;
201
202         mempool_free(r10_bio, conf->r10buf_pool);
203
204         lower_barrier(conf);
205 }
206
207 static void reschedule_retry(r10bio_t *r10_bio)
208 {
209         unsigned long flags;
210         mddev_t *mddev = r10_bio->mddev;
211         conf_t *conf = mddev->private;
212
213         spin_lock_irqsave(&conf->device_lock, flags);
214         list_add(&r10_bio->retry_list, &conf->retry_list);
215         conf->nr_queued ++;
216         spin_unlock_irqrestore(&conf->device_lock, flags);
217
218         /* wake up frozen array... */
219         wake_up(&conf->wait_barrier);
220
221         md_wakeup_thread(mddev->thread);
222 }
223
224 /*
225  * raid_end_bio_io() is called when we have finished servicing a mirrored
226  * operation and are ready to return a success/failure code to the buffer
227  * cache layer.
228  */
229 static void raid_end_bio_io(r10bio_t *r10_bio)
230 {
231         struct bio *bio = r10_bio->master_bio;
232         int done;
233         conf_t *conf = r10_bio->mddev->private;
234
235         if (bio->bi_phys_segments) {
236                 unsigned long flags;
237                 spin_lock_irqsave(&conf->device_lock, flags);
238                 bio->bi_phys_segments--;
239                 done = (bio->bi_phys_segments == 0);
240                 spin_unlock_irqrestore(&conf->device_lock, flags);
241         } else
242                 done = 1;
243         if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
244                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
245         if (done) {
246                 bio_endio(bio, 0);
247                 /*
248                  * Wake up any possible resync thread that waits for the device
249                  * to go idle.
250                  */
251                 allow_barrier(conf);
252         }
253         free_r10bio(r10_bio);
254 }
255
256 /*
257  * Update disk head position estimator based on IRQ completion info.
258  */
259 static inline void update_head_pos(int slot, r10bio_t *r10_bio)
260 {
261         conf_t *conf = r10_bio->mddev->private;
262
263         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
264                 r10_bio->devs[slot].addr + (r10_bio->sectors);
265 }
266
267 /*
268  * Find the disk number which triggered given bio
269  */
270 static int find_bio_disk(conf_t *conf, r10bio_t *r10_bio,
271                          struct bio *bio, int *slotp)
272 {
273         int slot;
274
275         for (slot = 0; slot < conf->copies; slot++)
276                 if (r10_bio->devs[slot].bio == bio)
277                         break;
278
279         BUG_ON(slot == conf->copies);
280         update_head_pos(slot, r10_bio);
281
282         if (slotp)
283                 *slotp = slot;
284         return r10_bio->devs[slot].devnum;
285 }
286
287 static void raid10_end_read_request(struct bio *bio, int error)
288 {
289         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
290         r10bio_t *r10_bio = bio->bi_private;
291         int slot, dev;
292         conf_t *conf = r10_bio->mddev->private;
293
294
295         slot = r10_bio->read_slot;
296         dev = r10_bio->devs[slot].devnum;
297         /*
298          * this branch is our 'one mirror IO has finished' event handler:
299          */
300         update_head_pos(slot, r10_bio);
301
302         if (uptodate) {
303                 /*
304                  * Set R10BIO_Uptodate in our master bio, so that
305                  * we will return a good error code to the higher
306                  * levels even if IO on some other mirrored buffer fails.
307                  *
308                  * The 'master' represents the composite IO operation to
309                  * user-side. So if something waits for IO, then it will
310                  * wait for the 'master' bio.
311                  */
312                 set_bit(R10BIO_Uptodate, &r10_bio->state);
313                 raid_end_bio_io(r10_bio);
314                 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
315         } else {
316                 /*
317                  * oops, read error - keep the refcount on the rdev
318                  */
319                 char b[BDEVNAME_SIZE];
320                 printk_ratelimited(KERN_ERR
321                                    "md/raid10:%s: %s: rescheduling sector %llu\n",
322                                    mdname(conf->mddev),
323                                    bdevname(conf->mirrors[dev].rdev->bdev, b),
324                                    (unsigned long long)r10_bio->sector);
325                 set_bit(R10BIO_ReadError, &r10_bio->state);
326                 reschedule_retry(r10_bio);
327         }
328 }
329
330 static void close_write(r10bio_t *r10_bio)
331 {
332         /* clear the bitmap if all writes complete successfully */
333         bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
334                         r10_bio->sectors,
335                         !test_bit(R10BIO_Degraded, &r10_bio->state),
336                         0);
337         md_write_end(r10_bio->mddev);
338 }
339
340 static void one_write_done(r10bio_t *r10_bio)
341 {
342         if (atomic_dec_and_test(&r10_bio->remaining)) {
343                 if (test_bit(R10BIO_WriteError, &r10_bio->state))
344                         reschedule_retry(r10_bio);
345                 else {
346                         close_write(r10_bio);
347                         if (test_bit(R10BIO_MadeGood, &r10_bio->state))
348                                 reschedule_retry(r10_bio);
349                         else
350                                 raid_end_bio_io(r10_bio);
351                 }
352         }
353 }
354
355 static void raid10_end_write_request(struct bio *bio, int error)
356 {
357         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
358         r10bio_t *r10_bio = bio->bi_private;
359         int dev;
360         int dec_rdev = 1;
361         conf_t *conf = r10_bio->mddev->private;
362         int slot;
363
364         dev = find_bio_disk(conf, r10_bio, bio, &slot);
365
366         /*
367          * this branch is our 'one mirror IO has finished' event handler:
368          */
369         if (!uptodate) {
370                 set_bit(WriteErrorSeen, &conf->mirrors[dev].rdev->flags);
371                 set_bit(R10BIO_WriteError, &r10_bio->state);
372                 dec_rdev = 0;
373         } else {
374                 /*
375                  * Set R10BIO_Uptodate in our master bio, so that
376                  * we will return a good error code for to the higher
377                  * levels even if IO on some other mirrored buffer fails.
378                  *
379                  * The 'master' represents the composite IO operation to
380                  * user-side. So if something waits for IO, then it will
381                  * wait for the 'master' bio.
382                  */
383                 sector_t first_bad;
384                 int bad_sectors;
385
386                 set_bit(R10BIO_Uptodate, &r10_bio->state);
387
388                 /* Maybe we can clear some bad blocks. */
389                 if (is_badblock(conf->mirrors[dev].rdev,
390                                 r10_bio->devs[slot].addr,
391                                 r10_bio->sectors,
392                                 &first_bad, &bad_sectors)) {
393                         bio_put(bio);
394                         r10_bio->devs[slot].bio = IO_MADE_GOOD;
395                         dec_rdev = 0;
396                         set_bit(R10BIO_MadeGood, &r10_bio->state);
397                 }
398         }
399
400         /*
401          *
402          * Let's see if all mirrored write operations have finished
403          * already.
404          */
405         one_write_done(r10_bio);
406         if (dec_rdev)
407                 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
408 }
409
410
411 /*
412  * RAID10 layout manager
413  * As well as the chunksize and raid_disks count, there are two
414  * parameters: near_copies and far_copies.
415  * near_copies * far_copies must be <= raid_disks.
416  * Normally one of these will be 1.
417  * If both are 1, we get raid0.
418  * If near_copies == raid_disks, we get raid1.
419  *
420  * Chunks are laid out in raid0 style with near_copies copies of the
421  * first chunk, followed by near_copies copies of the next chunk and
422  * so on.
423  * If far_copies > 1, then after 1/far_copies of the array has been assigned
424  * as described above, we start again with a device offset of near_copies.
425  * So we effectively have another copy of the whole array further down all
426  * the drives, but with blocks on different drives.
427  * With this layout, and block is never stored twice on the one device.
428  *
429  * raid10_find_phys finds the sector offset of a given virtual sector
430  * on each device that it is on.
431  *
432  * raid10_find_virt does the reverse mapping, from a device and a
433  * sector offset to a virtual address
434  */
435
436 static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
437 {
438         int n,f;
439         sector_t sector;
440         sector_t chunk;
441         sector_t stripe;
442         int dev;
443
444         int slot = 0;
445
446         /* now calculate first sector/dev */
447         chunk = r10bio->sector >> conf->chunk_shift;
448         sector = r10bio->sector & conf->chunk_mask;
449
450         chunk *= conf->near_copies;
451         stripe = chunk;
452         dev = sector_div(stripe, conf->raid_disks);
453         if (conf->far_offset)
454                 stripe *= conf->far_copies;
455
456         sector += stripe << conf->chunk_shift;
457
458         /* and calculate all the others */
459         for (n=0; n < conf->near_copies; n++) {
460                 int d = dev;
461                 sector_t s = sector;
462                 r10bio->devs[slot].addr = sector;
463                 r10bio->devs[slot].devnum = d;
464                 slot++;
465
466                 for (f = 1; f < conf->far_copies; f++) {
467                         d += conf->near_copies;
468                         if (d >= conf->raid_disks)
469                                 d -= conf->raid_disks;
470                         s += conf->stride;
471                         r10bio->devs[slot].devnum = d;
472                         r10bio->devs[slot].addr = s;
473                         slot++;
474                 }
475                 dev++;
476                 if (dev >= conf->raid_disks) {
477                         dev = 0;
478                         sector += (conf->chunk_mask + 1);
479                 }
480         }
481         BUG_ON(slot != conf->copies);
482 }
483
484 static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
485 {
486         sector_t offset, chunk, vchunk;
487
488         offset = sector & conf->chunk_mask;
489         if (conf->far_offset) {
490                 int fc;
491                 chunk = sector >> conf->chunk_shift;
492                 fc = sector_div(chunk, conf->far_copies);
493                 dev -= fc * conf->near_copies;
494                 if (dev < 0)
495                         dev += conf->raid_disks;
496         } else {
497                 while (sector >= conf->stride) {
498                         sector -= conf->stride;
499                         if (dev < conf->near_copies)
500                                 dev += conf->raid_disks - conf->near_copies;
501                         else
502                                 dev -= conf->near_copies;
503                 }
504                 chunk = sector >> conf->chunk_shift;
505         }
506         vchunk = chunk * conf->raid_disks + dev;
507         sector_div(vchunk, conf->near_copies);
508         return (vchunk << conf->chunk_shift) + offset;
509 }
510
511 /**
512  *      raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
513  *      @q: request queue
514  *      @bvm: properties of new bio
515  *      @biovec: the request that could be merged to it.
516  *
517  *      Return amount of bytes we can accept at this offset
518  *      If near_copies == raid_disk, there are no striping issues,
519  *      but in that case, the function isn't called at all.
520  */
521 static int raid10_mergeable_bvec(struct request_queue *q,
522                                  struct bvec_merge_data *bvm,
523                                  struct bio_vec *biovec)
524 {
525         mddev_t *mddev = q->queuedata;
526         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
527         int max;
528         unsigned int chunk_sectors = mddev->chunk_sectors;
529         unsigned int bio_sectors = bvm->bi_size >> 9;
530
531         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
532         if (max < 0) max = 0; /* bio_add cannot handle a negative return */
533         if (max <= biovec->bv_len && bio_sectors == 0)
534                 return biovec->bv_len;
535         else
536                 return max;
537 }
538
539 /*
540  * This routine returns the disk from which the requested read should
541  * be done. There is a per-array 'next expected sequential IO' sector
542  * number - if this matches on the next IO then we use the last disk.
543  * There is also a per-disk 'last know head position' sector that is
544  * maintained from IRQ contexts, both the normal and the resync IO
545  * completion handlers update this position correctly. If there is no
546  * perfect sequential match then we pick the disk whose head is closest.
547  *
548  * If there are 2 mirrors in the same 2 devices, performance degrades
549  * because position is mirror, not device based.
550  *
551  * The rdev for the device selected will have nr_pending incremented.
552  */
553
554 /*
555  * FIXME: possibly should rethink readbalancing and do it differently
556  * depending on near_copies / far_copies geometry.
557  */
558 static int read_balance(conf_t *conf, r10bio_t *r10_bio, int *max_sectors)
559 {
560         const sector_t this_sector = r10_bio->sector;
561         int disk, slot;
562         int sectors = r10_bio->sectors;
563         int best_good_sectors;
564         sector_t new_distance, best_dist;
565         mdk_rdev_t *rdev;
566         int do_balance;
567         int best_slot;
568
569         raid10_find_phys(conf, r10_bio);
570         rcu_read_lock();
571 retry:
572         sectors = r10_bio->sectors;
573         best_slot = -1;
574         best_dist = MaxSector;
575         best_good_sectors = 0;
576         do_balance = 1;
577         /*
578          * Check if we can balance. We can balance on the whole
579          * device if no resync is going on (recovery is ok), or below
580          * the resync window. We take the first readable disk when
581          * above the resync window.
582          */
583         if (conf->mddev->recovery_cp < MaxSector
584             && (this_sector + sectors >= conf->next_resync))
585                 do_balance = 0;
586
587         for (slot = 0; slot < conf->copies ; slot++) {
588                 sector_t first_bad;
589                 int bad_sectors;
590                 sector_t dev_sector;
591
592                 if (r10_bio->devs[slot].bio == IO_BLOCKED)
593                         continue;
594                 disk = r10_bio->devs[slot].devnum;
595                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
596                 if (rdev == NULL)
597                         continue;
598                 if (!test_bit(In_sync, &rdev->flags))
599                         continue;
600
601                 dev_sector = r10_bio->devs[slot].addr;
602                 if (is_badblock(rdev, dev_sector, sectors,
603                                 &first_bad, &bad_sectors)) {
604                         if (best_dist < MaxSector)
605                                 /* Already have a better slot */
606                                 continue;
607                         if (first_bad <= dev_sector) {
608                                 /* Cannot read here.  If this is the
609                                  * 'primary' device, then we must not read
610                                  * beyond 'bad_sectors' from another device.
611                                  */
612                                 bad_sectors -= (dev_sector - first_bad);
613                                 if (!do_balance && sectors > bad_sectors)
614                                         sectors = bad_sectors;
615                                 if (best_good_sectors > sectors)
616                                         best_good_sectors = sectors;
617                         } else {
618                                 sector_t good_sectors =
619                                         first_bad - dev_sector;
620                                 if (good_sectors > best_good_sectors) {
621                                         best_good_sectors = good_sectors;
622                                         best_slot = slot;
623                                 }
624                                 if (!do_balance)
625                                         /* Must read from here */
626                                         break;
627                         }
628                         continue;
629                 } else
630                         best_good_sectors = sectors;
631
632                 if (!do_balance)
633                         break;
634
635                 /* This optimisation is debatable, and completely destroys
636                  * sequential read speed for 'far copies' arrays.  So only
637                  * keep it for 'near' arrays, and review those later.
638                  */
639                 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending))
640                         break;
641
642                 /* for far > 1 always use the lowest address */
643                 if (conf->far_copies > 1)
644                         new_distance = r10_bio->devs[slot].addr;
645                 else
646                         new_distance = abs(r10_bio->devs[slot].addr -
647                                            conf->mirrors[disk].head_position);
648                 if (new_distance < best_dist) {
649                         best_dist = new_distance;
650                         best_slot = slot;
651                 }
652         }
653         if (slot == conf->copies)
654                 slot = best_slot;
655
656         if (slot >= 0) {
657                 disk = r10_bio->devs[slot].devnum;
658                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
659                 if (!rdev)
660                         goto retry;
661                 atomic_inc(&rdev->nr_pending);
662                 if (test_bit(Faulty, &rdev->flags)) {
663                         /* Cannot risk returning a device that failed
664                          * before we inc'ed nr_pending
665                          */
666                         rdev_dec_pending(rdev, conf->mddev);
667                         goto retry;
668                 }
669                 r10_bio->read_slot = slot;
670         } else
671                 disk = -1;
672         rcu_read_unlock();
673         *max_sectors = best_good_sectors;
674
675         return disk;
676 }
677
678 static int raid10_congested(void *data, int bits)
679 {
680         mddev_t *mddev = data;
681         conf_t *conf = mddev->private;
682         int i, ret = 0;
683
684         if (mddev_congested(mddev, bits))
685                 return 1;
686         rcu_read_lock();
687         for (i = 0; i < conf->raid_disks && ret == 0; i++) {
688                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
689                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
690                         struct request_queue *q = bdev_get_queue(rdev->bdev);
691
692                         ret |= bdi_congested(&q->backing_dev_info, bits);
693                 }
694         }
695         rcu_read_unlock();
696         return ret;
697 }
698
699 static void flush_pending_writes(conf_t *conf)
700 {
701         /* Any writes that have been queued but are awaiting
702          * bitmap updates get flushed here.
703          */
704         spin_lock_irq(&conf->device_lock);
705
706         if (conf->pending_bio_list.head) {
707                 struct bio *bio;
708                 bio = bio_list_get(&conf->pending_bio_list);
709                 spin_unlock_irq(&conf->device_lock);
710                 /* flush any pending bitmap writes to disk
711                  * before proceeding w/ I/O */
712                 bitmap_unplug(conf->mddev->bitmap);
713
714                 while (bio) { /* submit pending writes */
715                         struct bio *next = bio->bi_next;
716                         bio->bi_next = NULL;
717                         generic_make_request(bio);
718                         bio = next;
719                 }
720         } else
721                 spin_unlock_irq(&conf->device_lock);
722 }
723
724 /* Barriers....
725  * Sometimes we need to suspend IO while we do something else,
726  * either some resync/recovery, or reconfigure the array.
727  * To do this we raise a 'barrier'.
728  * The 'barrier' is a counter that can be raised multiple times
729  * to count how many activities are happening which preclude
730  * normal IO.
731  * We can only raise the barrier if there is no pending IO.
732  * i.e. if nr_pending == 0.
733  * We choose only to raise the barrier if no-one is waiting for the
734  * barrier to go down.  This means that as soon as an IO request
735  * is ready, no other operations which require a barrier will start
736  * until the IO request has had a chance.
737  *
738  * So: regular IO calls 'wait_barrier'.  When that returns there
739  *    is no backgroup IO happening,  It must arrange to call
740  *    allow_barrier when it has finished its IO.
741  * backgroup IO calls must call raise_barrier.  Once that returns
742  *    there is no normal IO happeing.  It must arrange to call
743  *    lower_barrier when the particular background IO completes.
744  */
745
746 static void raise_barrier(conf_t *conf, int force)
747 {
748         BUG_ON(force && !conf->barrier);
749         spin_lock_irq(&conf->resync_lock);
750
751         /* Wait until no block IO is waiting (unless 'force') */
752         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
753                             conf->resync_lock, );
754
755         /* block any new IO from starting */
756         conf->barrier++;
757
758         /* Now wait for all pending IO to complete */
759         wait_event_lock_irq(conf->wait_barrier,
760                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
761                             conf->resync_lock, );
762
763         spin_unlock_irq(&conf->resync_lock);
764 }
765
766 static void lower_barrier(conf_t *conf)
767 {
768         unsigned long flags;
769         spin_lock_irqsave(&conf->resync_lock, flags);
770         conf->barrier--;
771         spin_unlock_irqrestore(&conf->resync_lock, flags);
772         wake_up(&conf->wait_barrier);
773 }
774
775 static void wait_barrier(conf_t *conf)
776 {
777         spin_lock_irq(&conf->resync_lock);
778         if (conf->barrier) {
779                 conf->nr_waiting++;
780                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
781                                     conf->resync_lock,
782                                     );
783                 conf->nr_waiting--;
784         }
785         conf->nr_pending++;
786         spin_unlock_irq(&conf->resync_lock);
787 }
788
789 static void allow_barrier(conf_t *conf)
790 {
791         unsigned long flags;
792         spin_lock_irqsave(&conf->resync_lock, flags);
793         conf->nr_pending--;
794         spin_unlock_irqrestore(&conf->resync_lock, flags);
795         wake_up(&conf->wait_barrier);
796 }
797
798 static void freeze_array(conf_t *conf)
799 {
800         /* stop syncio and normal IO and wait for everything to
801          * go quiet.
802          * We increment barrier and nr_waiting, and then
803          * wait until nr_pending match nr_queued+1
804          * This is called in the context of one normal IO request
805          * that has failed. Thus any sync request that might be pending
806          * will be blocked by nr_pending, and we need to wait for
807          * pending IO requests to complete or be queued for re-try.
808          * Thus the number queued (nr_queued) plus this request (1)
809          * must match the number of pending IOs (nr_pending) before
810          * we continue.
811          */
812         spin_lock_irq(&conf->resync_lock);
813         conf->barrier++;
814         conf->nr_waiting++;
815         wait_event_lock_irq(conf->wait_barrier,
816                             conf->nr_pending == conf->nr_queued+1,
817                             conf->resync_lock,
818                             flush_pending_writes(conf));
819
820         spin_unlock_irq(&conf->resync_lock);
821 }
822
823 static void unfreeze_array(conf_t *conf)
824 {
825         /* reverse the effect of the freeze */
826         spin_lock_irq(&conf->resync_lock);
827         conf->barrier--;
828         conf->nr_waiting--;
829         wake_up(&conf->wait_barrier);
830         spin_unlock_irq(&conf->resync_lock);
831 }
832
833 static int make_request(mddev_t *mddev, struct bio * bio)
834 {
835         conf_t *conf = mddev->private;
836         mirror_info_t *mirror;
837         r10bio_t *r10_bio;
838         struct bio *read_bio;
839         int i;
840         int chunk_sects = conf->chunk_mask + 1;
841         const int rw = bio_data_dir(bio);
842         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
843         const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
844         unsigned long flags;
845         mdk_rdev_t *blocked_rdev;
846         int plugged;
847         int sectors_handled;
848         int max_sectors;
849
850         if (unlikely(bio->bi_rw & REQ_FLUSH)) {
851                 md_flush_request(mddev, bio);
852                 return 0;
853         }
854
855         /* If this request crosses a chunk boundary, we need to
856          * split it.  This will only happen for 1 PAGE (or less) requests.
857          */
858         if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
859                       > chunk_sects &&
860                     conf->near_copies < conf->raid_disks)) {
861                 struct bio_pair *bp;
862                 /* Sanity check -- queue functions should prevent this happening */
863                 if (bio->bi_vcnt != 1 ||
864                     bio->bi_idx != 0)
865                         goto bad_map;
866                 /* This is a one page bio that upper layers
867                  * refuse to split for us, so we need to split it.
868                  */
869                 bp = bio_split(bio,
870                                chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
871
872                 /* Each of these 'make_request' calls will call 'wait_barrier'.
873                  * If the first succeeds but the second blocks due to the resync
874                  * thread raising the barrier, we will deadlock because the
875                  * IO to the underlying device will be queued in generic_make_request
876                  * and will never complete, so will never reduce nr_pending.
877                  * So increment nr_waiting here so no new raise_barriers will
878                  * succeed, and so the second wait_barrier cannot block.
879                  */
880                 spin_lock_irq(&conf->resync_lock);
881                 conf->nr_waiting++;
882                 spin_unlock_irq(&conf->resync_lock);
883
884                 if (make_request(mddev, &bp->bio1))
885                         generic_make_request(&bp->bio1);
886                 if (make_request(mddev, &bp->bio2))
887                         generic_make_request(&bp->bio2);
888
889                 spin_lock_irq(&conf->resync_lock);
890                 conf->nr_waiting--;
891                 wake_up(&conf->wait_barrier);
892                 spin_unlock_irq(&conf->resync_lock);
893
894                 bio_pair_release(bp);
895                 return 0;
896         bad_map:
897                 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
898                        " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
899                        (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
900
901                 bio_io_error(bio);
902                 return 0;
903         }
904
905         md_write_start(mddev, bio);
906
907         /*
908          * Register the new request and wait if the reconstruction
909          * thread has put up a bar for new requests.
910          * Continue immediately if no resync is active currently.
911          */
912         wait_barrier(conf);
913
914         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
915
916         r10_bio->master_bio = bio;
917         r10_bio->sectors = bio->bi_size >> 9;
918
919         r10_bio->mddev = mddev;
920         r10_bio->sector = bio->bi_sector;
921         r10_bio->state = 0;
922
923         /* We might need to issue multiple reads to different
924          * devices if there are bad blocks around, so we keep
925          * track of the number of reads in bio->bi_phys_segments.
926          * If this is 0, there is only one r10_bio and no locking
927          * will be needed when the request completes.  If it is
928          * non-zero, then it is the number of not-completed requests.
929          */
930         bio->bi_phys_segments = 0;
931         clear_bit(BIO_SEG_VALID, &bio->bi_flags);
932
933         if (rw == READ) {
934                 /*
935                  * read balancing logic:
936                  */
937                 int disk;
938                 int slot;
939
940 read_again:
941                 disk = read_balance(conf, r10_bio, &max_sectors);
942                 slot = r10_bio->read_slot;
943                 if (disk < 0) {
944                         raid_end_bio_io(r10_bio);
945                         return 0;
946                 }
947                 mirror = conf->mirrors + disk;
948
949                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
950                 md_trim_bio(read_bio, r10_bio->sector - bio->bi_sector,
951                             max_sectors);
952
953                 r10_bio->devs[slot].bio = read_bio;
954
955                 read_bio->bi_sector = r10_bio->devs[slot].addr +
956                         mirror->rdev->data_offset;
957                 read_bio->bi_bdev = mirror->rdev->bdev;
958                 read_bio->bi_end_io = raid10_end_read_request;
959                 read_bio->bi_rw = READ | do_sync;
960                 read_bio->bi_private = r10_bio;
961
962                 if (max_sectors < r10_bio->sectors) {
963                         /* Could not read all from this device, so we will
964                          * need another r10_bio.
965                          */
966                         sectors_handled = (r10_bio->sectors + max_sectors
967                                            - bio->bi_sector);
968                         r10_bio->sectors = max_sectors;
969                         spin_lock_irq(&conf->device_lock);
970                         if (bio->bi_phys_segments == 0)
971                                 bio->bi_phys_segments = 2;
972                         else
973                                 bio->bi_phys_segments++;
974                         spin_unlock(&conf->device_lock);
975                         /* Cannot call generic_make_request directly
976                          * as that will be queued in __generic_make_request
977                          * and subsequent mempool_alloc might block
978                          * waiting for it.  so hand bio over to raid10d.
979                          */
980                         reschedule_retry(r10_bio);
981
982                         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
983
984                         r10_bio->master_bio = bio;
985                         r10_bio->sectors = ((bio->bi_size >> 9)
986                                             - sectors_handled);
987                         r10_bio->state = 0;
988                         r10_bio->mddev = mddev;
989                         r10_bio->sector = bio->bi_sector + sectors_handled;
990                         goto read_again;
991                 } else
992                         generic_make_request(read_bio);
993                 return 0;
994         }
995
996         /*
997          * WRITE:
998          */
999         /* first select target devices under rcu_lock and
1000          * inc refcount on their rdev.  Record them by setting
1001          * bios[x] to bio
1002          * If there are known/acknowledged bad blocks on any device
1003          * on which we have seen a write error, we want to avoid
1004          * writing to those blocks.  This potentially requires several
1005          * writes to write around the bad blocks.  Each set of writes
1006          * gets its own r10_bio with a set of bios attached.  The number
1007          * of r10_bios is recored in bio->bi_phys_segments just as with
1008          * the read case.
1009          */
1010         plugged = mddev_check_plugged(mddev);
1011
1012         raid10_find_phys(conf, r10_bio);
1013 retry_write:
1014         blocked_rdev = NULL;
1015         rcu_read_lock();
1016         max_sectors = r10_bio->sectors;
1017
1018         for (i = 0;  i < conf->copies; i++) {
1019                 int d = r10_bio->devs[i].devnum;
1020                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
1021                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1022                         atomic_inc(&rdev->nr_pending);
1023                         blocked_rdev = rdev;
1024                         break;
1025                 }
1026                 r10_bio->devs[i].bio = NULL;
1027                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1028                         set_bit(R10BIO_Degraded, &r10_bio->state);
1029                         continue;
1030                 }
1031                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1032                         sector_t first_bad;
1033                         sector_t dev_sector = r10_bio->devs[i].addr;
1034                         int bad_sectors;
1035                         int is_bad;
1036
1037                         is_bad = is_badblock(rdev, dev_sector,
1038                                              max_sectors,
1039                                              &first_bad, &bad_sectors);
1040                         if (is_bad < 0) {
1041                                 /* Mustn't write here until the bad block
1042                                  * is acknowledged
1043                                  */
1044                                 atomic_inc(&rdev->nr_pending);
1045                                 set_bit(BlockedBadBlocks, &rdev->flags);
1046                                 blocked_rdev = rdev;
1047                                 break;
1048                         }
1049                         if (is_bad && first_bad <= dev_sector) {
1050                                 /* Cannot write here at all */
1051                                 bad_sectors -= (dev_sector - first_bad);
1052                                 if (bad_sectors < max_sectors)
1053                                         /* Mustn't write more than bad_sectors
1054                                          * to other devices yet
1055                                          */
1056                                         max_sectors = bad_sectors;
1057                                 /* We don't set R10BIO_Degraded as that
1058                                  * only applies if the disk is missing,
1059                                  * so it might be re-added, and we want to
1060                                  * know to recover this chunk.
1061                                  * In this case the device is here, and the
1062                                  * fact that this chunk is not in-sync is
1063                                  * recorded in the bad block log.
1064                                  */
1065                                 continue;
1066                         }
1067                         if (is_bad) {
1068                                 int good_sectors = first_bad - dev_sector;
1069                                 if (good_sectors < max_sectors)
1070                                         max_sectors = good_sectors;
1071                         }
1072                 }
1073                 r10_bio->devs[i].bio = bio;
1074                 atomic_inc(&rdev->nr_pending);
1075         }
1076         rcu_read_unlock();
1077
1078         if (unlikely(blocked_rdev)) {
1079                 /* Have to wait for this device to get unblocked, then retry */
1080                 int j;
1081                 int d;
1082
1083                 for (j = 0; j < i; j++)
1084                         if (r10_bio->devs[j].bio) {
1085                                 d = r10_bio->devs[j].devnum;
1086                                 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1087                         }
1088                 allow_barrier(conf);
1089                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1090                 wait_barrier(conf);
1091                 goto retry_write;
1092         }
1093
1094         if (max_sectors < r10_bio->sectors) {
1095                 /* We are splitting this into multiple parts, so
1096                  * we need to prepare for allocating another r10_bio.
1097                  */
1098                 r10_bio->sectors = max_sectors;
1099                 spin_lock_irq(&conf->device_lock);
1100                 if (bio->bi_phys_segments == 0)
1101                         bio->bi_phys_segments = 2;
1102                 else
1103                         bio->bi_phys_segments++;
1104                 spin_unlock_irq(&conf->device_lock);
1105         }
1106         sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector;
1107
1108         atomic_set(&r10_bio->remaining, 1);
1109         bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1110
1111         for (i = 0; i < conf->copies; i++) {
1112                 struct bio *mbio;
1113                 int d = r10_bio->devs[i].devnum;
1114                 if (!r10_bio->devs[i].bio)
1115                         continue;
1116
1117                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1118                 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1119                             max_sectors);
1120                 r10_bio->devs[i].bio = mbio;
1121
1122                 mbio->bi_sector = (r10_bio->devs[i].addr+
1123                                    conf->mirrors[d].rdev->data_offset);
1124                 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1125                 mbio->bi_end_io = raid10_end_write_request;
1126                 mbio->bi_rw = WRITE | do_sync | do_fua;
1127                 mbio->bi_private = r10_bio;
1128
1129                 atomic_inc(&r10_bio->remaining);
1130                 spin_lock_irqsave(&conf->device_lock, flags);
1131                 bio_list_add(&conf->pending_bio_list, mbio);
1132                 spin_unlock_irqrestore(&conf->device_lock, flags);
1133         }
1134
1135         /* Remove the bias on 'remaining' */
1136         one_write_done(r10_bio);
1137
1138         /* In case raid10d snuck in to freeze_array */
1139         wake_up(&conf->wait_barrier);
1140
1141         if (sectors_handled < (bio->bi_size >> 9)) {
1142                 /* We need another r10_bio.  It has already been counted
1143                  * in bio->bi_phys_segments.
1144                  */
1145                 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1146
1147                 r10_bio->master_bio = bio;
1148                 r10_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1149
1150                 r10_bio->mddev = mddev;
1151                 r10_bio->sector = bio->bi_sector + sectors_handled;
1152                 r10_bio->state = 0;
1153                 goto retry_write;
1154         }
1155
1156         if (do_sync || !mddev->bitmap || !plugged)
1157                 md_wakeup_thread(mddev->thread);
1158         return 0;
1159 }
1160
1161 static void status(struct seq_file *seq, mddev_t *mddev)
1162 {
1163         conf_t *conf = mddev->private;
1164         int i;
1165
1166         if (conf->near_copies < conf->raid_disks)
1167                 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1168         if (conf->near_copies > 1)
1169                 seq_printf(seq, " %d near-copies", conf->near_copies);
1170         if (conf->far_copies > 1) {
1171                 if (conf->far_offset)
1172                         seq_printf(seq, " %d offset-copies", conf->far_copies);
1173                 else
1174                         seq_printf(seq, " %d far-copies", conf->far_copies);
1175         }
1176         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1177                                         conf->raid_disks - mddev->degraded);
1178         for (i = 0; i < conf->raid_disks; i++)
1179                 seq_printf(seq, "%s",
1180                               conf->mirrors[i].rdev &&
1181                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1182         seq_printf(seq, "]");
1183 }
1184
1185 /* check if there are enough drives for
1186  * every block to appear on atleast one.
1187  * Don't consider the device numbered 'ignore'
1188  * as we might be about to remove it.
1189  */
1190 static int enough(conf_t *conf, int ignore)
1191 {
1192         int first = 0;
1193
1194         do {
1195                 int n = conf->copies;
1196                 int cnt = 0;
1197                 while (n--) {
1198                         if (conf->mirrors[first].rdev &&
1199                             first != ignore)
1200                                 cnt++;
1201                         first = (first+1) % conf->raid_disks;
1202                 }
1203                 if (cnt == 0)
1204                         return 0;
1205         } while (first != 0);
1206         return 1;
1207 }
1208
1209 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1210 {
1211         char b[BDEVNAME_SIZE];
1212         conf_t *conf = mddev->private;
1213
1214         /*
1215          * If it is not operational, then we have already marked it as dead
1216          * else if it is the last working disks, ignore the error, let the
1217          * next level up know.
1218          * else mark the drive as failed
1219          */
1220         if (test_bit(In_sync, &rdev->flags)
1221             && !enough(conf, rdev->raid_disk))
1222                 /*
1223                  * Don't fail the drive, just return an IO error.
1224                  */
1225                 return;
1226         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1227                 unsigned long flags;
1228                 spin_lock_irqsave(&conf->device_lock, flags);
1229                 mddev->degraded++;
1230                 spin_unlock_irqrestore(&conf->device_lock, flags);
1231                 /*
1232                  * if recovery is running, make sure it aborts.
1233                  */
1234                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1235         }
1236         set_bit(Blocked, &rdev->flags);
1237         set_bit(Faulty, &rdev->flags);
1238         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1239         printk(KERN_ALERT
1240                "md/raid10:%s: Disk failure on %s, disabling device.\n"
1241                "md/raid10:%s: Operation continuing on %d devices.\n",
1242                mdname(mddev), bdevname(rdev->bdev, b),
1243                mdname(mddev), conf->raid_disks - mddev->degraded);
1244 }
1245
1246 static void print_conf(conf_t *conf)
1247 {
1248         int i;
1249         mirror_info_t *tmp;
1250
1251         printk(KERN_DEBUG "RAID10 conf printout:\n");
1252         if (!conf) {
1253                 printk(KERN_DEBUG "(!conf)\n");
1254                 return;
1255         }
1256         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1257                 conf->raid_disks);
1258
1259         for (i = 0; i < conf->raid_disks; i++) {
1260                 char b[BDEVNAME_SIZE];
1261                 tmp = conf->mirrors + i;
1262                 if (tmp->rdev)
1263                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1264                                 i, !test_bit(In_sync, &tmp->rdev->flags),
1265                                 !test_bit(Faulty, &tmp->rdev->flags),
1266                                 bdevname(tmp->rdev->bdev,b));
1267         }
1268 }
1269
1270 static void close_sync(conf_t *conf)
1271 {
1272         wait_barrier(conf);
1273         allow_barrier(conf);
1274
1275         mempool_destroy(conf->r10buf_pool);
1276         conf->r10buf_pool = NULL;
1277 }
1278
1279 static int raid10_spare_active(mddev_t *mddev)
1280 {
1281         int i;
1282         conf_t *conf = mddev->private;
1283         mirror_info_t *tmp;
1284         int count = 0;
1285         unsigned long flags;
1286
1287         /*
1288          * Find all non-in_sync disks within the RAID10 configuration
1289          * and mark them in_sync
1290          */
1291         for (i = 0; i < conf->raid_disks; i++) {
1292                 tmp = conf->mirrors + i;
1293                 if (tmp->rdev
1294                     && !test_bit(Faulty, &tmp->rdev->flags)
1295                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1296                         count++;
1297                         sysfs_notify_dirent(tmp->rdev->sysfs_state);
1298                 }
1299         }
1300         spin_lock_irqsave(&conf->device_lock, flags);
1301         mddev->degraded -= count;
1302         spin_unlock_irqrestore(&conf->device_lock, flags);
1303
1304         print_conf(conf);
1305         return count;
1306 }
1307
1308
1309 static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1310 {
1311         conf_t *conf = mddev->private;
1312         int err = -EEXIST;
1313         int mirror;
1314         int first = 0;
1315         int last = conf->raid_disks - 1;
1316
1317         if (mddev->recovery_cp < MaxSector)
1318                 /* only hot-add to in-sync arrays, as recovery is
1319                  * very different from resync
1320                  */
1321                 return -EBUSY;
1322         if (!enough(conf, -1))
1323                 return -EINVAL;
1324
1325         if (rdev->raid_disk >= 0)
1326                 first = last = rdev->raid_disk;
1327
1328         if (rdev->saved_raid_disk >= first &&
1329             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1330                 mirror = rdev->saved_raid_disk;
1331         else
1332                 mirror = first;
1333         for ( ; mirror <= last ; mirror++) {
1334                 mirror_info_t *p = &conf->mirrors[mirror];
1335                 if (p->recovery_disabled == mddev->recovery_disabled)
1336                         continue;
1337                 if (!p->rdev)
1338                         continue;
1339
1340                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1341                                   rdev->data_offset << 9);
1342                 /* as we don't honour merge_bvec_fn, we must
1343                  * never risk violating it, so limit
1344                  * ->max_segments to one lying with a single
1345                  * page, as a one page request is never in
1346                  * violation.
1347                  */
1348                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1349                         blk_queue_max_segments(mddev->queue, 1);
1350                         blk_queue_segment_boundary(mddev->queue,
1351                                                    PAGE_CACHE_SIZE - 1);
1352                 }
1353
1354                 p->head_position = 0;
1355                 rdev->raid_disk = mirror;
1356                 err = 0;
1357                 if (rdev->saved_raid_disk != mirror)
1358                         conf->fullsync = 1;
1359                 rcu_assign_pointer(p->rdev, rdev);
1360                 break;
1361         }
1362
1363         md_integrity_add_rdev(rdev, mddev);
1364         print_conf(conf);
1365         return err;
1366 }
1367
1368 static int raid10_remove_disk(mddev_t *mddev, int number)
1369 {
1370         conf_t *conf = mddev->private;
1371         int err = 0;
1372         mdk_rdev_t *rdev;
1373         mirror_info_t *p = conf->mirrors+ number;
1374
1375         print_conf(conf);
1376         rdev = p->rdev;
1377         if (rdev) {
1378                 if (test_bit(In_sync, &rdev->flags) ||
1379                     atomic_read(&rdev->nr_pending)) {
1380                         err = -EBUSY;
1381                         goto abort;
1382                 }
1383                 /* Only remove faulty devices in recovery
1384                  * is not possible.
1385                  */
1386                 if (!test_bit(Faulty, &rdev->flags) &&
1387                     mddev->recovery_disabled != p->recovery_disabled &&
1388                     enough(conf, -1)) {
1389                         err = -EBUSY;
1390                         goto abort;
1391                 }
1392                 p->rdev = NULL;
1393                 synchronize_rcu();
1394                 if (atomic_read(&rdev->nr_pending)) {
1395                         /* lost the race, try later */
1396                         err = -EBUSY;
1397                         p->rdev = rdev;
1398                         goto abort;
1399                 }
1400                 err = md_integrity_register(mddev);
1401         }
1402 abort:
1403
1404         print_conf(conf);
1405         return err;
1406 }
1407
1408
1409 static void end_sync_read(struct bio *bio, int error)
1410 {
1411         r10bio_t *r10_bio = bio->bi_private;
1412         conf_t *conf = r10_bio->mddev->private;
1413         int d;
1414
1415         d = find_bio_disk(conf, r10_bio, bio, NULL);
1416
1417         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1418                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1419         else
1420                 /* The write handler will notice the lack of
1421                  * R10BIO_Uptodate and record any errors etc
1422                  */
1423                 atomic_add(r10_bio->sectors,
1424                            &conf->mirrors[d].rdev->corrected_errors);
1425
1426         /* for reconstruct, we always reschedule after a read.
1427          * for resync, only after all reads
1428          */
1429         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1430         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1431             atomic_dec_and_test(&r10_bio->remaining)) {
1432                 /* we have read all the blocks,
1433                  * do the comparison in process context in raid10d
1434                  */
1435                 reschedule_retry(r10_bio);
1436         }
1437 }
1438
1439 static void end_sync_request(r10bio_t *r10_bio)
1440 {
1441         mddev_t *mddev = r10_bio->mddev;
1442
1443         while (atomic_dec_and_test(&r10_bio->remaining)) {
1444                 if (r10_bio->master_bio == NULL) {
1445                         /* the primary of several recovery bios */
1446                         sector_t s = r10_bio->sectors;
1447                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1448                             test_bit(R10BIO_WriteError, &r10_bio->state))
1449                                 reschedule_retry(r10_bio);
1450                         else
1451                                 put_buf(r10_bio);
1452                         md_done_sync(mddev, s, 1);
1453                         break;
1454                 } else {
1455                         r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1456                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1457                             test_bit(R10BIO_WriteError, &r10_bio->state))
1458                                 reschedule_retry(r10_bio);
1459                         else
1460                                 put_buf(r10_bio);
1461                         r10_bio = r10_bio2;
1462                 }
1463         }
1464 }
1465
1466 static void end_sync_write(struct bio *bio, int error)
1467 {
1468         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1469         r10bio_t *r10_bio = bio->bi_private;
1470         mddev_t *mddev = r10_bio->mddev;
1471         conf_t *conf = mddev->private;
1472         int d;
1473         sector_t first_bad;
1474         int bad_sectors;
1475         int slot;
1476
1477         d = find_bio_disk(conf, r10_bio, bio, &slot);
1478
1479         if (!uptodate) {
1480                 set_bit(WriteErrorSeen, &conf->mirrors[d].rdev->flags);
1481                 set_bit(R10BIO_WriteError, &r10_bio->state);
1482         } else if (is_badblock(conf->mirrors[d].rdev,
1483                              r10_bio->devs[slot].addr,
1484                              r10_bio->sectors,
1485                              &first_bad, &bad_sectors))
1486                 set_bit(R10BIO_MadeGood, &r10_bio->state);
1487
1488         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1489
1490         end_sync_request(r10_bio);
1491 }
1492
1493 /*
1494  * Note: sync and recover and handled very differently for raid10
1495  * This code is for resync.
1496  * For resync, we read through virtual addresses and read all blocks.
1497  * If there is any error, we schedule a write.  The lowest numbered
1498  * drive is authoritative.
1499  * However requests come for physical address, so we need to map.
1500  * For every physical address there are raid_disks/copies virtual addresses,
1501  * which is always are least one, but is not necessarly an integer.
1502  * This means that a physical address can span multiple chunks, so we may
1503  * have to submit multiple io requests for a single sync request.
1504  */
1505 /*
1506  * We check if all blocks are in-sync and only write to blocks that
1507  * aren't in sync
1508  */
1509 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1510 {
1511         conf_t *conf = mddev->private;
1512         int i, first;
1513         struct bio *tbio, *fbio;
1514
1515         atomic_set(&r10_bio->remaining, 1);
1516
1517         /* find the first device with a block */
1518         for (i=0; i<conf->copies; i++)
1519                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1520                         break;
1521
1522         if (i == conf->copies)
1523                 goto done;
1524
1525         first = i;
1526         fbio = r10_bio->devs[i].bio;
1527
1528         /* now find blocks with errors */
1529         for (i=0 ; i < conf->copies ; i++) {
1530                 int  j, d;
1531                 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1532
1533                 tbio = r10_bio->devs[i].bio;
1534
1535                 if (tbio->bi_end_io != end_sync_read)
1536                         continue;
1537                 if (i == first)
1538                         continue;
1539                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1540                         /* We know that the bi_io_vec layout is the same for
1541                          * both 'first' and 'i', so we just compare them.
1542                          * All vec entries are PAGE_SIZE;
1543                          */
1544                         for (j = 0; j < vcnt; j++)
1545                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1546                                            page_address(tbio->bi_io_vec[j].bv_page),
1547                                            PAGE_SIZE))
1548                                         break;
1549                         if (j == vcnt)
1550                                 continue;
1551                         mddev->resync_mismatches += r10_bio->sectors;
1552                         if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1553                                 /* Don't fix anything. */
1554                                 continue;
1555                 }
1556                 /* Ok, we need to write this bio, either to correct an
1557                  * inconsistency or to correct an unreadable block.
1558                  * First we need to fixup bv_offset, bv_len and
1559                  * bi_vecs, as the read request might have corrupted these
1560                  */
1561                 tbio->bi_vcnt = vcnt;
1562                 tbio->bi_size = r10_bio->sectors << 9;
1563                 tbio->bi_idx = 0;
1564                 tbio->bi_phys_segments = 0;
1565                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1566                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1567                 tbio->bi_next = NULL;
1568                 tbio->bi_rw = WRITE;
1569                 tbio->bi_private = r10_bio;
1570                 tbio->bi_sector = r10_bio->devs[i].addr;
1571
1572                 for (j=0; j < vcnt ; j++) {
1573                         tbio->bi_io_vec[j].bv_offset = 0;
1574                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1575
1576                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1577                                page_address(fbio->bi_io_vec[j].bv_page),
1578                                PAGE_SIZE);
1579                 }
1580                 tbio->bi_end_io = end_sync_write;
1581
1582                 d = r10_bio->devs[i].devnum;
1583                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1584                 atomic_inc(&r10_bio->remaining);
1585                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1586
1587                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1588                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1589                 generic_make_request(tbio);
1590         }
1591
1592 done:
1593         if (atomic_dec_and_test(&r10_bio->remaining)) {
1594                 md_done_sync(mddev, r10_bio->sectors, 1);
1595                 put_buf(r10_bio);
1596         }
1597 }
1598
1599 /*
1600  * Now for the recovery code.
1601  * Recovery happens across physical sectors.
1602  * We recover all non-is_sync drives by finding the virtual address of
1603  * each, and then choose a working drive that also has that virt address.
1604  * There is a separate r10_bio for each non-in_sync drive.
1605  * Only the first two slots are in use. The first for reading,
1606  * The second for writing.
1607  *
1608  */
1609 static void fix_recovery_read_error(r10bio_t *r10_bio)
1610 {
1611         /* We got a read error during recovery.
1612          * We repeat the read in smaller page-sized sections.
1613          * If a read succeeds, write it to the new device or record
1614          * a bad block if we cannot.
1615          * If a read fails, record a bad block on both old and
1616          * new devices.
1617          */
1618         mddev_t *mddev = r10_bio->mddev;
1619         conf_t *conf = mddev->private;
1620         struct bio *bio = r10_bio->devs[0].bio;
1621         sector_t sect = 0;
1622         int sectors = r10_bio->sectors;
1623         int idx = 0;
1624         int dr = r10_bio->devs[0].devnum;
1625         int dw = r10_bio->devs[1].devnum;
1626
1627         while (sectors) {
1628                 int s = sectors;
1629                 mdk_rdev_t *rdev;
1630                 sector_t addr;
1631                 int ok;
1632
1633                 if (s > (PAGE_SIZE>>9))
1634                         s = PAGE_SIZE >> 9;
1635
1636                 rdev = conf->mirrors[dr].rdev;
1637                 addr = r10_bio->devs[0].addr + sect,
1638                 ok = sync_page_io(rdev,
1639                                   addr,
1640                                   s << 9,
1641                                   bio->bi_io_vec[idx].bv_page,
1642                                   READ, false);
1643                 if (ok) {
1644                         rdev = conf->mirrors[dw].rdev;
1645                         addr = r10_bio->devs[1].addr + sect;
1646                         ok = sync_page_io(rdev,
1647                                           addr,
1648                                           s << 9,
1649                                           bio->bi_io_vec[idx].bv_page,
1650                                           WRITE, false);
1651                         if (!ok)
1652                                 set_bit(WriteErrorSeen, &rdev->flags);
1653                 }
1654                 if (!ok) {
1655                         /* We don't worry if we cannot set a bad block -
1656                          * it really is bad so there is no loss in not
1657                          * recording it yet
1658                          */
1659                         rdev_set_badblocks(rdev, addr, s, 0);
1660
1661                         if (rdev != conf->mirrors[dw].rdev) {
1662                                 /* need bad block on destination too */
1663                                 mdk_rdev_t *rdev2 = conf->mirrors[dw].rdev;
1664                                 addr = r10_bio->devs[1].addr + sect;
1665                                 ok = rdev_set_badblocks(rdev2, addr, s, 0);
1666                                 if (!ok) {
1667                                         /* just abort the recovery */
1668                                         printk(KERN_NOTICE
1669                                                "md/raid10:%s: recovery aborted"
1670                                                " due to read error\n",
1671                                                mdname(mddev));
1672
1673                                         conf->mirrors[dw].recovery_disabled
1674                                                 = mddev->recovery_disabled;
1675                                         set_bit(MD_RECOVERY_INTR,
1676                                                 &mddev->recovery);
1677                                         break;
1678                                 }
1679                         }
1680                 }
1681
1682                 sectors -= s;
1683                 sect += s;
1684                 idx++;
1685         }
1686 }
1687
1688 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1689 {
1690         conf_t *conf = mddev->private;
1691         int d;
1692         struct bio *wbio;
1693
1694         if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
1695                 fix_recovery_read_error(r10_bio);
1696                 end_sync_request(r10_bio);
1697                 return;
1698         }
1699
1700         /*
1701          * share the pages with the first bio
1702          * and submit the write request
1703          */
1704         wbio = r10_bio->devs[1].bio;
1705         d = r10_bio->devs[1].devnum;
1706
1707         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1708         md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1709         generic_make_request(wbio);
1710 }
1711
1712
1713 /*
1714  * Used by fix_read_error() to decay the per rdev read_errors.
1715  * We halve the read error count for every hour that has elapsed
1716  * since the last recorded read error.
1717  *
1718  */
1719 static void check_decay_read_errors(mddev_t *mddev, mdk_rdev_t *rdev)
1720 {
1721         struct timespec cur_time_mon;
1722         unsigned long hours_since_last;
1723         unsigned int read_errors = atomic_read(&rdev->read_errors);
1724
1725         ktime_get_ts(&cur_time_mon);
1726
1727         if (rdev->last_read_error.tv_sec == 0 &&
1728             rdev->last_read_error.tv_nsec == 0) {
1729                 /* first time we've seen a read error */
1730                 rdev->last_read_error = cur_time_mon;
1731                 return;
1732         }
1733
1734         hours_since_last = (cur_time_mon.tv_sec -
1735                             rdev->last_read_error.tv_sec) / 3600;
1736
1737         rdev->last_read_error = cur_time_mon;
1738
1739         /*
1740          * if hours_since_last is > the number of bits in read_errors
1741          * just set read errors to 0. We do this to avoid
1742          * overflowing the shift of read_errors by hours_since_last.
1743          */
1744         if (hours_since_last >= 8 * sizeof(read_errors))
1745                 atomic_set(&rdev->read_errors, 0);
1746         else
1747                 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
1748 }
1749
1750 static int r10_sync_page_io(mdk_rdev_t *rdev, sector_t sector,
1751                             int sectors, struct page *page, int rw)
1752 {
1753         sector_t first_bad;
1754         int bad_sectors;
1755
1756         if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
1757             && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
1758                 return -1;
1759         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1760                 /* success */
1761                 return 1;
1762         if (rw == WRITE)
1763                 set_bit(WriteErrorSeen, &rdev->flags);
1764         /* need to record an error - either for the block or the device */
1765         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1766                 md_error(rdev->mddev, rdev);
1767         return 0;
1768 }
1769
1770 /*
1771  * This is a kernel thread which:
1772  *
1773  *      1.      Retries failed read operations on working mirrors.
1774  *      2.      Updates the raid superblock when problems encounter.
1775  *      3.      Performs writes following reads for array synchronising.
1776  */
1777
1778 static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1779 {
1780         int sect = 0; /* Offset from r10_bio->sector */
1781         int sectors = r10_bio->sectors;
1782         mdk_rdev_t*rdev;
1783         int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
1784         int d = r10_bio->devs[r10_bio->read_slot].devnum;
1785
1786         /* still own a reference to this rdev, so it cannot
1787          * have been cleared recently.
1788          */
1789         rdev = conf->mirrors[d].rdev;
1790
1791         if (test_bit(Faulty, &rdev->flags))
1792                 /* drive has already been failed, just ignore any
1793                    more fix_read_error() attempts */
1794                 return;
1795
1796         check_decay_read_errors(mddev, rdev);
1797         atomic_inc(&rdev->read_errors);
1798         if (atomic_read(&rdev->read_errors) > max_read_errors) {
1799                 char b[BDEVNAME_SIZE];
1800                 bdevname(rdev->bdev, b);
1801
1802                 printk(KERN_NOTICE
1803                        "md/raid10:%s: %s: Raid device exceeded "
1804                        "read_error threshold [cur %d:max %d]\n",
1805                        mdname(mddev), b,
1806                        atomic_read(&rdev->read_errors), max_read_errors);
1807                 printk(KERN_NOTICE
1808                        "md/raid10:%s: %s: Failing raid device\n",
1809                        mdname(mddev), b);
1810                 md_error(mddev, conf->mirrors[d].rdev);
1811                 return;
1812         }
1813
1814         while(sectors) {
1815                 int s = sectors;
1816                 int sl = r10_bio->read_slot;
1817                 int success = 0;
1818                 int start;
1819
1820                 if (s > (PAGE_SIZE>>9))
1821                         s = PAGE_SIZE >> 9;
1822
1823                 rcu_read_lock();
1824                 do {
1825                         sector_t first_bad;
1826                         int bad_sectors;
1827
1828                         d = r10_bio->devs[sl].devnum;
1829                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1830                         if (rdev &&
1831                             test_bit(In_sync, &rdev->flags) &&
1832                             is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
1833                                         &first_bad, &bad_sectors) == 0) {
1834                                 atomic_inc(&rdev->nr_pending);
1835                                 rcu_read_unlock();
1836                                 success = sync_page_io(rdev,
1837                                                        r10_bio->devs[sl].addr +
1838                                                        sect,
1839                                                        s<<9,
1840                                                        conf->tmppage, READ, false);
1841                                 rdev_dec_pending(rdev, mddev);
1842                                 rcu_read_lock();
1843                                 if (success)
1844                                         break;
1845                         }
1846                         sl++;
1847                         if (sl == conf->copies)
1848                                 sl = 0;
1849                 } while (!success && sl != r10_bio->read_slot);
1850                 rcu_read_unlock();
1851
1852                 if (!success) {
1853                         /* Cannot read from anywhere, just mark the block
1854                          * as bad on the first device to discourage future
1855                          * reads.
1856                          */
1857                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1858                         rdev = conf->mirrors[dn].rdev;
1859
1860                         if (!rdev_set_badblocks(
1861                                     rdev,
1862                                     r10_bio->devs[r10_bio->read_slot].addr
1863                                     + sect,
1864                                     s, 0))
1865                                 md_error(mddev, rdev);
1866                         break;
1867                 }
1868
1869                 start = sl;
1870                 /* write it back and re-read */
1871                 rcu_read_lock();
1872                 while (sl != r10_bio->read_slot) {
1873                         char b[BDEVNAME_SIZE];
1874
1875                         if (sl==0)
1876                                 sl = conf->copies;
1877                         sl--;
1878                         d = r10_bio->devs[sl].devnum;
1879                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1880                         if (!rdev ||
1881                             !test_bit(In_sync, &rdev->flags))
1882                                 continue;
1883
1884                         atomic_inc(&rdev->nr_pending);
1885                         rcu_read_unlock();
1886                         if (r10_sync_page_io(rdev,
1887                                              r10_bio->devs[sl].addr +
1888                                              sect,
1889                                              s<<9, conf->tmppage, WRITE)
1890                             == 0) {
1891                                 /* Well, this device is dead */
1892                                 printk(KERN_NOTICE
1893                                        "md/raid10:%s: read correction "
1894                                        "write failed"
1895                                        " (%d sectors at %llu on %s)\n",
1896                                        mdname(mddev), s,
1897                                        (unsigned long long)(
1898                                                sect + rdev->data_offset),
1899                                        bdevname(rdev->bdev, b));
1900                                 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
1901                                        "drive\n",
1902                                        mdname(mddev),
1903                                        bdevname(rdev->bdev, b));
1904                         }
1905                         rdev_dec_pending(rdev, mddev);
1906                         rcu_read_lock();
1907                 }
1908                 sl = start;
1909                 while (sl != r10_bio->read_slot) {
1910                         char b[BDEVNAME_SIZE];
1911
1912                         if (sl==0)
1913                                 sl = conf->copies;
1914                         sl--;
1915                         d = r10_bio->devs[sl].devnum;
1916                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1917                         if (!rdev ||
1918                             !test_bit(In_sync, &rdev->flags))
1919                                 continue;
1920
1921                         atomic_inc(&rdev->nr_pending);
1922                         rcu_read_unlock();
1923                         switch (r10_sync_page_io(rdev,
1924                                              r10_bio->devs[sl].addr +
1925                                              sect,
1926                                              s<<9, conf->tmppage,
1927                                                  READ)) {
1928                         case 0:
1929                                 /* Well, this device is dead */
1930                                 printk(KERN_NOTICE
1931                                        "md/raid10:%s: unable to read back "
1932                                        "corrected sectors"
1933                                        " (%d sectors at %llu on %s)\n",
1934                                        mdname(mddev), s,
1935                                        (unsigned long long)(
1936                                                sect + rdev->data_offset),
1937                                        bdevname(rdev->bdev, b));
1938                                 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
1939                                        "drive\n",
1940                                        mdname(mddev),
1941                                        bdevname(rdev->bdev, b));
1942                                 break;
1943                         case 1:
1944                                 printk(KERN_INFO
1945                                        "md/raid10:%s: read error corrected"
1946                                        " (%d sectors at %llu on %s)\n",
1947                                        mdname(mddev), s,
1948                                        (unsigned long long)(
1949                                                sect + rdev->data_offset),
1950                                        bdevname(rdev->bdev, b));
1951                                 atomic_add(s, &rdev->corrected_errors);
1952                         }
1953
1954                         rdev_dec_pending(rdev, mddev);
1955                         rcu_read_lock();
1956                 }
1957                 rcu_read_unlock();
1958
1959                 sectors -= s;
1960                 sect += s;
1961         }
1962 }
1963
1964 static void bi_complete(struct bio *bio, int error)
1965 {
1966         complete((struct completion *)bio->bi_private);
1967 }
1968
1969 static int submit_bio_wait(int rw, struct bio *bio)
1970 {
1971         struct completion event;
1972         rw |= REQ_SYNC;
1973
1974         init_completion(&event);
1975         bio->bi_private = &event;
1976         bio->bi_end_io = bi_complete;
1977         submit_bio(rw, bio);
1978         wait_for_completion(&event);
1979
1980         return test_bit(BIO_UPTODATE, &bio->bi_flags);
1981 }
1982
1983 static int narrow_write_error(r10bio_t *r10_bio, int i)
1984 {
1985         struct bio *bio = r10_bio->master_bio;
1986         mddev_t *mddev = r10_bio->mddev;
1987         conf_t *conf = mddev->private;
1988         mdk_rdev_t *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
1989         /* bio has the data to be written to slot 'i' where
1990          * we just recently had a write error.
1991          * We repeatedly clone the bio and trim down to one block,
1992          * then try the write.  Where the write fails we record
1993          * a bad block.
1994          * It is conceivable that the bio doesn't exactly align with
1995          * blocks.  We must handle this.
1996          *
1997          * We currently own a reference to the rdev.
1998          */
1999
2000         int block_sectors;
2001         sector_t sector;
2002         int sectors;
2003         int sect_to_write = r10_bio->sectors;
2004         int ok = 1;
2005
2006         if (rdev->badblocks.shift < 0)
2007                 return 0;
2008
2009         block_sectors = 1 << rdev->badblocks.shift;
2010         sector = r10_bio->sector;
2011         sectors = ((r10_bio->sector + block_sectors)
2012                    & ~(sector_t)(block_sectors - 1))
2013                 - sector;
2014
2015         while (sect_to_write) {
2016                 struct bio *wbio;
2017                 if (sectors > sect_to_write)
2018                         sectors = sect_to_write;
2019                 /* Write at 'sector' for 'sectors' */
2020                 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2021                 md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2022                 wbio->bi_sector = (r10_bio->devs[i].addr+
2023                                    rdev->data_offset+
2024                                    (sector - r10_bio->sector));
2025                 wbio->bi_bdev = rdev->bdev;
2026                 if (submit_bio_wait(WRITE, wbio) == 0)
2027                         /* Failure! */
2028                         ok = rdev_set_badblocks(rdev, sector,
2029                                                 sectors, 0)
2030                                 && ok;
2031
2032                 bio_put(wbio);
2033                 sect_to_write -= sectors;
2034                 sector += sectors;
2035                 sectors = block_sectors;
2036         }
2037         return ok;
2038 }
2039
2040 static void handle_read_error(mddev_t *mddev, r10bio_t *r10_bio)
2041 {
2042         int slot = r10_bio->read_slot;
2043         int mirror = r10_bio->devs[slot].devnum;
2044         struct bio *bio;
2045         conf_t *conf = mddev->private;
2046         mdk_rdev_t *rdev;
2047         char b[BDEVNAME_SIZE];
2048         unsigned long do_sync;
2049         int max_sectors;
2050
2051         /* we got a read error. Maybe the drive is bad.  Maybe just
2052          * the block and we can fix it.
2053          * We freeze all other IO, and try reading the block from
2054          * other devices.  When we find one, we re-write
2055          * and check it that fixes the read error.
2056          * This is all done synchronously while the array is
2057          * frozen.
2058          */
2059         if (mddev->ro == 0) {
2060                 freeze_array(conf);
2061                 fix_read_error(conf, mddev, r10_bio);
2062                 unfreeze_array(conf);
2063         }
2064         rdev_dec_pending(conf->mirrors[mirror].rdev, mddev);
2065
2066         bio = r10_bio->devs[slot].bio;
2067         bdevname(bio->bi_bdev, b);
2068         r10_bio->devs[slot].bio =
2069                 mddev->ro ? IO_BLOCKED : NULL;
2070 read_more:
2071         mirror = read_balance(conf, r10_bio, &max_sectors);
2072         if (mirror == -1) {
2073                 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2074                        " read error for block %llu\n",
2075                        mdname(mddev), b,
2076                        (unsigned long long)r10_bio->sector);
2077                 raid_end_bio_io(r10_bio);
2078                 bio_put(bio);
2079                 return;
2080         }
2081
2082         do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2083         if (bio)
2084                 bio_put(bio);
2085         slot = r10_bio->read_slot;
2086         rdev = conf->mirrors[mirror].rdev;
2087         printk_ratelimited(
2088                 KERN_ERR
2089                 "md/raid10:%s: %s: redirecting"
2090                 "sector %llu to another mirror\n",
2091                 mdname(mddev),
2092                 bdevname(rdev->bdev, b),
2093                 (unsigned long long)r10_bio->sector);
2094         bio = bio_clone_mddev(r10_bio->master_bio,
2095                               GFP_NOIO, mddev);
2096         md_trim_bio(bio,
2097                     r10_bio->sector - bio->bi_sector,
2098                     max_sectors);
2099         r10_bio->devs[slot].bio = bio;
2100         bio->bi_sector = r10_bio->devs[slot].addr
2101                 + rdev->data_offset;
2102         bio->bi_bdev = rdev->bdev;
2103         bio->bi_rw = READ | do_sync;
2104         bio->bi_private = r10_bio;
2105         bio->bi_end_io = raid10_end_read_request;
2106         if (max_sectors < r10_bio->sectors) {
2107                 /* Drat - have to split this up more */
2108                 struct bio *mbio = r10_bio->master_bio;
2109                 int sectors_handled =
2110                         r10_bio->sector + max_sectors
2111                         - mbio->bi_sector;
2112                 r10_bio->sectors = max_sectors;
2113                 spin_lock_irq(&conf->device_lock);
2114                 if (mbio->bi_phys_segments == 0)
2115                         mbio->bi_phys_segments = 2;
2116                 else
2117                         mbio->bi_phys_segments++;
2118                 spin_unlock_irq(&conf->device_lock);
2119                 generic_make_request(bio);
2120                 bio = NULL;
2121
2122                 r10_bio = mempool_alloc(conf->r10bio_pool,
2123                                         GFP_NOIO);
2124                 r10_bio->master_bio = mbio;
2125                 r10_bio->sectors = (mbio->bi_size >> 9)
2126                         - sectors_handled;
2127                 r10_bio->state = 0;
2128                 set_bit(R10BIO_ReadError,
2129                         &r10_bio->state);
2130                 r10_bio->mddev = mddev;
2131                 r10_bio->sector = mbio->bi_sector
2132                         + sectors_handled;
2133
2134                 goto read_more;
2135         } else
2136                 generic_make_request(bio);
2137 }
2138
2139 static void handle_write_completed(conf_t *conf, r10bio_t *r10_bio)
2140 {
2141         /* Some sort of write request has finished and it
2142          * succeeded in writing where we thought there was a
2143          * bad block.  So forget the bad block.
2144          * Or possibly if failed and we need to record
2145          * a bad block.
2146          */
2147         int m;
2148         mdk_rdev_t *rdev;
2149
2150         if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2151             test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2152                 for (m = 0; m < conf->copies; m++) {
2153                         int dev = r10_bio->devs[m].devnum;
2154                         rdev = conf->mirrors[dev].rdev;
2155                         if (r10_bio->devs[m].bio == NULL)
2156                                 continue;
2157                         if (test_bit(BIO_UPTODATE,
2158                                      &r10_bio->devs[m].bio->bi_flags)) {
2159                                 rdev_clear_badblocks(
2160                                         rdev,
2161                                         r10_bio->devs[m].addr,
2162                                         r10_bio->sectors);
2163                         } else {
2164                                 if (!rdev_set_badblocks(
2165                                             rdev,
2166                                             r10_bio->devs[m].addr,
2167                                             r10_bio->sectors, 0))
2168                                         md_error(conf->mddev, rdev);
2169                         }
2170                 }
2171                 put_buf(r10_bio);
2172         } else {
2173                 for (m = 0; m < conf->copies; m++) {
2174                         int dev = r10_bio->devs[m].devnum;
2175                         struct bio *bio = r10_bio->devs[m].bio;
2176                         rdev = conf->mirrors[dev].rdev;
2177                         if (bio == IO_MADE_GOOD) {
2178                                 rdev_clear_badblocks(
2179                                         rdev,
2180                                         r10_bio->devs[m].addr,
2181                                         r10_bio->sectors);
2182                                 rdev_dec_pending(rdev, conf->mddev);
2183                         } else if (bio != NULL &&
2184                                    !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2185                                 if (!narrow_write_error(r10_bio, m)) {
2186                                         md_error(conf->mddev, rdev);
2187                                         set_bit(R10BIO_Degraded,
2188                                                 &r10_bio->state);
2189                                 }
2190                                 rdev_dec_pending(rdev, conf->mddev);
2191                         }
2192                 }
2193                 if (test_bit(R10BIO_WriteError,
2194                              &r10_bio->state))
2195                         close_write(r10_bio);
2196                 raid_end_bio_io(r10_bio);
2197         }
2198 }
2199
2200 static void raid10d(mddev_t *mddev)
2201 {
2202         r10bio_t *r10_bio;
2203         unsigned long flags;
2204         conf_t *conf = mddev->private;
2205         struct list_head *head = &conf->retry_list;
2206         struct blk_plug plug;
2207
2208         md_check_recovery(mddev);
2209
2210         blk_start_plug(&plug);
2211         for (;;) {
2212
2213                 flush_pending_writes(conf);
2214
2215                 spin_lock_irqsave(&conf->device_lock, flags);
2216                 if (list_empty(head)) {
2217                         spin_unlock_irqrestore(&conf->device_lock, flags);
2218                         break;
2219                 }
2220                 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
2221                 list_del(head->prev);
2222                 conf->nr_queued--;
2223                 spin_unlock_irqrestore(&conf->device_lock, flags);
2224
2225                 mddev = r10_bio->mddev;
2226                 conf = mddev->private;
2227                 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2228                     test_bit(R10BIO_WriteError, &r10_bio->state))
2229                         handle_write_completed(conf, r10_bio);
2230                 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2231                         sync_request_write(mddev, r10_bio);
2232                 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2233                         recovery_request_write(mddev, r10_bio);
2234                 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2235                         handle_read_error(mddev, r10_bio);
2236                 else {
2237                         /* just a partial read to be scheduled from a
2238                          * separate context
2239                          */
2240                         int slot = r10_bio->read_slot;
2241                         generic_make_request(r10_bio->devs[slot].bio);
2242                 }
2243
2244                 cond_resched();
2245                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2246                         md_check_recovery(mddev);
2247         }
2248         blk_finish_plug(&plug);
2249 }
2250
2251
2252 static int init_resync(conf_t *conf)
2253 {
2254         int buffs;
2255
2256         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2257         BUG_ON(conf->r10buf_pool);
2258         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2259         if (!conf->r10buf_pool)
2260                 return -ENOMEM;
2261         conf->next_resync = 0;
2262         return 0;
2263 }
2264
2265 /*
2266  * perform a "sync" on one "block"
2267  *
2268  * We need to make sure that no normal I/O request - particularly write
2269  * requests - conflict with active sync requests.
2270  *
2271  * This is achieved by tracking pending requests and a 'barrier' concept
2272  * that can be installed to exclude normal IO requests.
2273  *
2274  * Resync and recovery are handled very differently.
2275  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2276  *
2277  * For resync, we iterate over virtual addresses, read all copies,
2278  * and update if there are differences.  If only one copy is live,
2279  * skip it.
2280  * For recovery, we iterate over physical addresses, read a good
2281  * value for each non-in_sync drive, and over-write.
2282  *
2283  * So, for recovery we may have several outstanding complex requests for a
2284  * given address, one for each out-of-sync device.  We model this by allocating
2285  * a number of r10_bio structures, one for each out-of-sync device.
2286  * As we setup these structures, we collect all bio's together into a list
2287  * which we then process collectively to add pages, and then process again
2288  * to pass to generic_make_request.
2289  *
2290  * The r10_bio structures are linked using a borrowed master_bio pointer.
2291  * This link is counted in ->remaining.  When the r10_bio that points to NULL
2292  * has its remaining count decremented to 0, the whole complex operation
2293  * is complete.
2294  *
2295  */
2296
2297 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr,
2298                              int *skipped, int go_faster)
2299 {
2300         conf_t *conf = mddev->private;
2301         r10bio_t *r10_bio;
2302         struct bio *biolist = NULL, *bio;
2303         sector_t max_sector, nr_sectors;
2304         int i;
2305         int max_sync;
2306         sector_t sync_blocks;
2307         sector_t sectors_skipped = 0;
2308         int chunks_skipped = 0;
2309
2310         if (!conf->r10buf_pool)
2311                 if (init_resync(conf))
2312                         return 0;
2313
2314  skipped:
2315         max_sector = mddev->dev_sectors;
2316         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2317                 max_sector = mddev->resync_max_sectors;
2318         if (sector_nr >= max_sector) {
2319                 /* If we aborted, we need to abort the
2320                  * sync on the 'current' bitmap chucks (there can
2321                  * be several when recovering multiple devices).
2322                  * as we may have started syncing it but not finished.
2323                  * We can find the current address in
2324                  * mddev->curr_resync, but for recovery,
2325                  * we need to convert that to several
2326                  * virtual addresses.
2327                  */
2328                 if (mddev->curr_resync < max_sector) { /* aborted */
2329                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2330                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2331                                                 &sync_blocks, 1);
2332                         else for (i=0; i<conf->raid_disks; i++) {
2333                                 sector_t sect =
2334                                         raid10_find_virt(conf, mddev->curr_resync, i);
2335                                 bitmap_end_sync(mddev->bitmap, sect,
2336                                                 &sync_blocks, 1);
2337                         }
2338                 } else /* completed sync */
2339                         conf->fullsync = 0;
2340
2341                 bitmap_close_sync(mddev->bitmap);
2342                 close_sync(conf);
2343                 *skipped = 1;
2344                 return sectors_skipped;
2345         }
2346         if (chunks_skipped >= conf->raid_disks) {
2347                 /* if there has been nothing to do on any drive,
2348                  * then there is nothing to do at all..
2349                  */
2350                 *skipped = 1;
2351                 return (max_sector - sector_nr) + sectors_skipped;
2352         }
2353
2354         if (max_sector > mddev->resync_max)
2355                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2356
2357         /* make sure whole request will fit in a chunk - if chunks
2358          * are meaningful
2359          */
2360         if (conf->near_copies < conf->raid_disks &&
2361             max_sector > (sector_nr | conf->chunk_mask))
2362                 max_sector = (sector_nr | conf->chunk_mask) + 1;
2363         /*
2364          * If there is non-resync activity waiting for us then
2365          * put in a delay to throttle resync.
2366          */
2367         if (!go_faster && conf->nr_waiting)
2368                 msleep_interruptible(1000);
2369
2370         /* Again, very different code for resync and recovery.
2371          * Both must result in an r10bio with a list of bios that
2372          * have bi_end_io, bi_sector, bi_bdev set,
2373          * and bi_private set to the r10bio.
2374          * For recovery, we may actually create several r10bios
2375          * with 2 bios in each, that correspond to the bios in the main one.
2376          * In this case, the subordinate r10bios link back through a
2377          * borrowed master_bio pointer, and the counter in the master
2378          * includes a ref from each subordinate.
2379          */
2380         /* First, we decide what to do and set ->bi_end_io
2381          * To end_sync_read if we want to read, and
2382          * end_sync_write if we will want to write.
2383          */
2384
2385         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2386         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2387                 /* recovery... the complicated one */
2388                 int j;
2389                 r10_bio = NULL;
2390
2391                 for (i=0 ; i<conf->raid_disks; i++) {
2392                         int still_degraded;
2393                         r10bio_t *rb2;
2394                         sector_t sect;
2395                         int must_sync;
2396                         int any_working;
2397
2398                         if (conf->mirrors[i].rdev == NULL ||
2399                             test_bit(In_sync, &conf->mirrors[i].rdev->flags)) 
2400                                 continue;
2401
2402                         still_degraded = 0;
2403                         /* want to reconstruct this device */
2404                         rb2 = r10_bio;
2405                         sect = raid10_find_virt(conf, sector_nr, i);
2406                         /* Unless we are doing a full sync, we only need
2407                          * to recover the block if it is set in the bitmap
2408                          */
2409                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
2410                                                       &sync_blocks, 1);
2411                         if (sync_blocks < max_sync)
2412                                 max_sync = sync_blocks;
2413                         if (!must_sync &&
2414                             !conf->fullsync) {
2415                                 /* yep, skip the sync_blocks here, but don't assume
2416                                  * that there will never be anything to do here
2417                                  */
2418                                 chunks_skipped = -1;
2419                                 continue;
2420                         }
2421
2422                         r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2423                         raise_barrier(conf, rb2 != NULL);
2424                         atomic_set(&r10_bio->remaining, 0);
2425
2426                         r10_bio->master_bio = (struct bio*)rb2;
2427                         if (rb2)
2428                                 atomic_inc(&rb2->remaining);
2429                         r10_bio->mddev = mddev;
2430                         set_bit(R10BIO_IsRecover, &r10_bio->state);
2431                         r10_bio->sector = sect;
2432
2433                         raid10_find_phys(conf, r10_bio);
2434
2435                         /* Need to check if the array will still be
2436                          * degraded
2437                          */
2438                         for (j=0; j<conf->raid_disks; j++)
2439                                 if (conf->mirrors[j].rdev == NULL ||
2440                                     test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
2441                                         still_degraded = 1;
2442                                         break;
2443                                 }
2444
2445                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
2446                                                       &sync_blocks, still_degraded);
2447
2448                         any_working = 0;
2449                         for (j=0; j<conf->copies;j++) {
2450                                 int k;
2451                                 int d = r10_bio->devs[j].devnum;
2452                                 sector_t from_addr, to_addr;
2453                                 mdk_rdev_t *rdev;
2454                                 sector_t sector, first_bad;
2455                                 int bad_sectors;
2456                                 if (!conf->mirrors[d].rdev ||
2457                                     !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
2458                                         continue;
2459                                 /* This is where we read from */
2460                                 any_working = 1;
2461                                 rdev = conf->mirrors[d].rdev;
2462                                 sector = r10_bio->devs[j].addr;
2463
2464                                 if (is_badblock(rdev, sector, max_sync,
2465                                                 &first_bad, &bad_sectors)) {
2466                                         if (first_bad > sector)
2467                                                 max_sync = first_bad - sector;
2468                                         else {
2469                                                 bad_sectors -= (sector
2470                                                                 - first_bad);
2471                                                 if (max_sync > bad_sectors)
2472                                                         max_sync = bad_sectors;
2473                                                 continue;
2474                                         }
2475                                 }
2476                                 bio = r10_bio->devs[0].bio;
2477                                 bio->bi_next = biolist;
2478                                 biolist = bio;
2479                                 bio->bi_private = r10_bio;
2480                                 bio->bi_end_io = end_sync_read;
2481                                 bio->bi_rw = READ;
2482                                 from_addr = r10_bio->devs[j].addr;
2483                                 bio->bi_sector = from_addr +
2484                                         conf->mirrors[d].rdev->data_offset;
2485                                 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2486                                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2487                                 atomic_inc(&r10_bio->remaining);
2488                                 /* and we write to 'i' */
2489
2490                                 for (k=0; k<conf->copies; k++)
2491                                         if (r10_bio->devs[k].devnum == i)
2492                                                 break;
2493                                 BUG_ON(k == conf->copies);
2494                                 bio = r10_bio->devs[1].bio;
2495                                 bio->bi_next = biolist;
2496                                 biolist = bio;
2497                                 bio->bi_private = r10_bio;
2498                                 bio->bi_end_io = end_sync_write;
2499                                 bio->bi_rw = WRITE;
2500                                 to_addr = r10_bio->devs[k].addr;
2501                                 bio->bi_sector = to_addr +
2502                                         conf->mirrors[i].rdev->data_offset;
2503                                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
2504
2505                                 r10_bio->devs[0].devnum = d;
2506                                 r10_bio->devs[0].addr = from_addr;
2507                                 r10_bio->devs[1].devnum = i;
2508                                 r10_bio->devs[1].addr = to_addr;
2509
2510                                 break;
2511                         }
2512                         if (j == conf->copies) {
2513                                 /* Cannot recover, so abort the recovery or
2514                                  * record a bad block */
2515                                 put_buf(r10_bio);
2516                                 if (rb2)
2517                                         atomic_dec(&rb2->remaining);
2518                                 r10_bio = rb2;
2519                                 if (any_working) {
2520                                         /* problem is that there are bad blocks
2521                                          * on other device(s)
2522                                          */
2523                                         int k;
2524                                         for (k = 0; k < conf->copies; k++)
2525                                                 if (r10_bio->devs[k].devnum == i)
2526                                                         break;
2527                                         if (!rdev_set_badblocks(
2528                                                     conf->mirrors[i].rdev,
2529                                                     r10_bio->devs[k].addr,
2530                                                     max_sync, 0))
2531                                                 any_working = 0;
2532                                 }
2533                                 if (!any_working)  {
2534                                         if (!test_and_set_bit(MD_RECOVERY_INTR,
2535                                                               &mddev->recovery))
2536                                                 printk(KERN_INFO "md/raid10:%s: insufficient "
2537                                                        "working devices for recovery.\n",
2538                                                        mdname(mddev));
2539                                         conf->mirrors[i].recovery_disabled
2540                                                 = mddev->recovery_disabled;
2541                                 }
2542                                 break;
2543                         }
2544                 }
2545                 if (biolist == NULL) {
2546                         while (r10_bio) {
2547                                 r10bio_t *rb2 = r10_bio;
2548                                 r10_bio = (r10bio_t*) rb2->master_bio;
2549                                 rb2->master_bio = NULL;
2550                                 put_buf(rb2);
2551                         }
2552                         goto giveup;
2553                 }
2554         } else {
2555                 /* resync. Schedule a read for every block at this virt offset */
2556                 int count = 0;
2557
2558                 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2559
2560                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2561                                        &sync_blocks, mddev->degraded) &&
2562                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
2563                                                  &mddev->recovery)) {
2564                         /* We can skip this block */
2565                         *skipped = 1;
2566                         return sync_blocks + sectors_skipped;
2567                 }
2568                 if (sync_blocks < max_sync)
2569                         max_sync = sync_blocks;
2570                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2571
2572                 r10_bio->mddev = mddev;
2573                 atomic_set(&r10_bio->remaining, 0);
2574                 raise_barrier(conf, 0);
2575                 conf->next_resync = sector_nr;
2576
2577                 r10_bio->master_bio = NULL;
2578                 r10_bio->sector = sector_nr;
2579                 set_bit(R10BIO_IsSync, &r10_bio->state);
2580                 raid10_find_phys(conf, r10_bio);
2581                 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
2582
2583                 for (i=0; i<conf->copies; i++) {
2584                         int d = r10_bio->devs[i].devnum;
2585                         sector_t first_bad, sector;
2586                         int bad_sectors;
2587
2588                         bio = r10_bio->devs[i].bio;
2589                         bio->bi_end_io = NULL;
2590                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
2591                         if (conf->mirrors[d].rdev == NULL ||
2592                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
2593                                 continue;
2594                         sector = r10_bio->devs[i].addr;
2595                         if (is_badblock(conf->mirrors[d].rdev,
2596                                         sector, max_sync,
2597                                         &first_bad, &bad_sectors)) {
2598                                 if (first_bad > sector)
2599                                         max_sync = first_bad - sector;
2600                                 else {
2601                                         bad_sectors -= (sector - first_bad);
2602                                         if (max_sync > bad_sectors)
2603                                                 max_sync = max_sync;
2604                                         continue;
2605                                 }
2606                         }
2607                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2608                         atomic_inc(&r10_bio->remaining);
2609                         bio->bi_next = biolist;
2610                         biolist = bio;
2611                         bio->bi_private = r10_bio;
2612                         bio->bi_end_io = end_sync_read;
2613                         bio->bi_rw = READ;
2614                         bio->bi_sector = sector +
2615                                 conf->mirrors[d].rdev->data_offset;
2616                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2617                         count++;
2618                 }
2619
2620                 if (count < 2) {
2621                         for (i=0; i<conf->copies; i++) {
2622                                 int d = r10_bio->devs[i].devnum;
2623                                 if (r10_bio->devs[i].bio->bi_end_io)
2624                                         rdev_dec_pending(conf->mirrors[d].rdev,
2625                                                          mddev);
2626                         }
2627                         put_buf(r10_bio);
2628                         biolist = NULL;
2629                         goto giveup;
2630                 }
2631         }
2632
2633         for (bio = biolist; bio ; bio=bio->bi_next) {
2634
2635                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
2636                 if (bio->bi_end_io)
2637                         bio->bi_flags |= 1 << BIO_UPTODATE;
2638                 bio->bi_vcnt = 0;
2639                 bio->bi_idx = 0;
2640                 bio->bi_phys_segments = 0;
2641                 bio->bi_size = 0;
2642         }
2643
2644         nr_sectors = 0;
2645         if (sector_nr + max_sync < max_sector)
2646                 max_sector = sector_nr + max_sync;
2647         do {
2648                 struct page *page;
2649                 int len = PAGE_SIZE;
2650                 if (sector_nr + (len>>9) > max_sector)
2651                         len = (max_sector - sector_nr) << 9;
2652                 if (len == 0)
2653                         break;
2654                 for (bio= biolist ; bio ; bio=bio->bi_next) {
2655                         struct bio *bio2;
2656                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2657                         if (bio_add_page(bio, page, len, 0))
2658                                 continue;
2659
2660                         /* stop here */
2661                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2662                         for (bio2 = biolist;
2663                              bio2 && bio2 != bio;
2664                              bio2 = bio2->bi_next) {
2665                                 /* remove last page from this bio */
2666                                 bio2->bi_vcnt--;
2667                                 bio2->bi_size -= len;
2668                                 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
2669                         }
2670                         goto bio_full;
2671                 }
2672                 nr_sectors += len>>9;
2673                 sector_nr += len>>9;
2674         } while (biolist->bi_vcnt < RESYNC_PAGES);
2675  bio_full:
2676         r10_bio->sectors = nr_sectors;
2677
2678         while (biolist) {
2679                 bio = biolist;
2680                 biolist = biolist->bi_next;
2681
2682                 bio->bi_next = NULL;
2683                 r10_bio = bio->bi_private;
2684                 r10_bio->sectors = nr_sectors;
2685
2686                 if (bio->bi_end_io == end_sync_read) {
2687                         md_sync_acct(bio->bi_bdev, nr_sectors);
2688                         generic_make_request(bio);
2689                 }
2690         }
2691
2692         if (sectors_skipped)
2693                 /* pretend they weren't skipped, it makes
2694                  * no important difference in this case
2695                  */
2696                 md_done_sync(mddev, sectors_skipped, 1);
2697
2698         return sectors_skipped + nr_sectors;
2699  giveup:
2700         /* There is nowhere to write, so all non-sync
2701          * drives must be failed or in resync, all drives
2702          * have a bad block, so try the next chunk...
2703          */
2704         if (sector_nr + max_sync < max_sector)
2705                 max_sector = sector_nr + max_sync;
2706
2707         sectors_skipped += (max_sector - sector_nr);
2708         chunks_skipped ++;
2709         sector_nr = max_sector;
2710         goto skipped;
2711 }
2712
2713 static sector_t
2714 raid10_size(mddev_t *mddev, sector_t sectors, int raid_disks)
2715 {
2716         sector_t size;
2717         conf_t *conf = mddev->private;
2718
2719         if (!raid_disks)
2720                 raid_disks = conf->raid_disks;
2721         if (!sectors)
2722                 sectors = conf->dev_sectors;
2723
2724         size = sectors >> conf->chunk_shift;
2725         sector_div(size, conf->far_copies);
2726         size = size * raid_disks;
2727         sector_div(size, conf->near_copies);
2728
2729         return size << conf->chunk_shift;
2730 }
2731
2732
2733 static conf_t *setup_conf(mddev_t *mddev)
2734 {
2735         conf_t *conf = NULL;
2736         int nc, fc, fo;
2737         sector_t stride, size;
2738         int err = -EINVAL;
2739
2740         if (mddev->new_chunk_sectors < (PAGE_SIZE >> 9) ||
2741             !is_power_of_2(mddev->new_chunk_sectors)) {
2742                 printk(KERN_ERR "md/raid10:%s: chunk size must be "
2743                        "at least PAGE_SIZE(%ld) and be a power of 2.\n",
2744                        mdname(mddev), PAGE_SIZE);
2745                 goto out;
2746         }
2747
2748         nc = mddev->new_layout & 255;
2749         fc = (mddev->new_layout >> 8) & 255;
2750         fo = mddev->new_layout & (1<<16);
2751
2752         if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
2753             (mddev->new_layout >> 17)) {
2754                 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
2755                        mdname(mddev), mddev->new_layout);
2756                 goto out;
2757         }
2758
2759         err = -ENOMEM;
2760         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2761         if (!conf)
2762                 goto out;
2763
2764         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2765                                 GFP_KERNEL);
2766         if (!conf->mirrors)
2767                 goto out;
2768
2769         conf->tmppage = alloc_page(GFP_KERNEL);
2770         if (!conf->tmppage)
2771                 goto out;
2772
2773
2774         conf->raid_disks = mddev->raid_disks;
2775         conf->near_copies = nc;
2776         conf->far_copies = fc;
2777         conf->copies = nc*fc;
2778         conf->far_offset = fo;
2779         conf->chunk_mask = mddev->new_chunk_sectors - 1;
2780         conf->chunk_shift = ffz(~mddev->new_chunk_sectors);
2781
2782         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2783                                            r10bio_pool_free, conf);
2784         if (!conf->r10bio_pool)
2785                 goto out;
2786
2787         size = mddev->dev_sectors >> conf->chunk_shift;
2788         sector_div(size, fc);
2789         size = size * conf->raid_disks;
2790         sector_div(size, nc);
2791         /* 'size' is now the number of chunks in the array */
2792         /* calculate "used chunks per device" in 'stride' */
2793         stride = size * conf->copies;
2794
2795         /* We need to round up when dividing by raid_disks to
2796          * get the stride size.
2797          */
2798         stride += conf->raid_disks - 1;
2799         sector_div(stride, conf->raid_disks);
2800
2801         conf->dev_sectors = stride << conf->chunk_shift;
2802
2803         if (fo)
2804                 stride = 1;
2805         else
2806                 sector_div(stride, fc);
2807         conf->stride = stride << conf->chunk_shift;
2808
2809
2810         spin_lock_init(&conf->device_lock);
2811         INIT_LIST_HEAD(&conf->retry_list);
2812
2813         spin_lock_init(&conf->resync_lock);
2814         init_waitqueue_head(&conf->wait_barrier);
2815
2816         conf->thread = md_register_thread(raid10d, mddev, NULL);
2817         if (!conf->thread)
2818                 goto out;
2819
2820         conf->mddev = mddev;
2821         return conf;
2822
2823  out:
2824         printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
2825                mdname(mddev));
2826         if (conf) {
2827                 if (conf->r10bio_pool)
2828                         mempool_destroy(conf->r10bio_pool);
2829                 kfree(conf->mirrors);
2830                 safe_put_page(conf->tmppage);
2831                 kfree(conf);
2832         }
2833         return ERR_PTR(err);
2834 }
2835
2836 static int run(mddev_t *mddev)
2837 {
2838         conf_t *conf;
2839         int i, disk_idx, chunk_size;
2840         mirror_info_t *disk;
2841         mdk_rdev_t *rdev;
2842         sector_t size;
2843
2844         /*
2845          * copy the already verified devices into our private RAID10
2846          * bookkeeping area. [whatever we allocate in run(),
2847          * should be freed in stop()]
2848          */
2849
2850         if (mddev->private == NULL) {
2851                 conf = setup_conf(mddev);
2852                 if (IS_ERR(conf))
2853                         return PTR_ERR(conf);
2854                 mddev->private = conf;
2855         }
2856         conf = mddev->private;
2857         if (!conf)
2858                 goto out;
2859
2860         mddev->thread = conf->thread;
2861         conf->thread = NULL;
2862
2863         chunk_size = mddev->chunk_sectors << 9;
2864         blk_queue_io_min(mddev->queue, chunk_size);
2865         if (conf->raid_disks % conf->near_copies)
2866                 blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks);
2867         else
2868                 blk_queue_io_opt(mddev->queue, chunk_size *
2869                                  (conf->raid_disks / conf->near_copies));
2870
2871         list_for_each_entry(rdev, &mddev->disks, same_set) {
2872
2873                 disk_idx = rdev->raid_disk;
2874                 if (disk_idx >= conf->raid_disks
2875                     || disk_idx < 0)
2876                         continue;
2877                 disk = conf->mirrors + disk_idx;
2878
2879                 disk->rdev = rdev;
2880                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2881                                   rdev->data_offset << 9);
2882                 /* as we don't honour merge_bvec_fn, we must never risk
2883                  * violating it, so limit max_segments to 1 lying
2884                  * within a single page.
2885                  */
2886                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2887                         blk_queue_max_segments(mddev->queue, 1);
2888                         blk_queue_segment_boundary(mddev->queue,
2889                                                    PAGE_CACHE_SIZE - 1);
2890                 }
2891
2892                 disk->head_position = 0;
2893         }
2894         /* need to check that every block has at least one working mirror */
2895         if (!enough(conf, -1)) {
2896                 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
2897                        mdname(mddev));
2898                 goto out_free_conf;
2899         }
2900
2901         mddev->degraded = 0;
2902         for (i = 0; i < conf->raid_disks; i++) {
2903
2904                 disk = conf->mirrors + i;
2905
2906                 if (!disk->rdev ||
2907                     !test_bit(In_sync, &disk->rdev->flags)) {
2908                         disk->head_position = 0;
2909                         mddev->degraded++;
2910                         if (disk->rdev)
2911                                 conf->fullsync = 1;
2912                 }
2913         }
2914
2915         if (mddev->recovery_cp != MaxSector)
2916                 printk(KERN_NOTICE "md/raid10:%s: not clean"
2917                        " -- starting background reconstruction\n",
2918                        mdname(mddev));
2919         printk(KERN_INFO
2920                 "md/raid10:%s: active with %d out of %d devices\n",
2921                 mdname(mddev), conf->raid_disks - mddev->degraded,
2922                 conf->raid_disks);
2923         /*
2924          * Ok, everything is just fine now
2925          */
2926         mddev->dev_sectors = conf->dev_sectors;
2927         size = raid10_size(mddev, 0, 0);
2928         md_set_array_sectors(mddev, size);
2929         mddev->resync_max_sectors = size;
2930
2931         mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2932         mddev->queue->backing_dev_info.congested_data = mddev;
2933
2934         /* Calculate max read-ahead size.
2935          * We need to readahead at least twice a whole stripe....
2936          * maybe...
2937          */
2938         {
2939                 int stripe = conf->raid_disks *
2940                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
2941                 stripe /= conf->near_copies;
2942                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2943                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2944         }
2945
2946         if (conf->near_copies < conf->raid_disks)
2947                 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2948
2949         if (md_integrity_register(mddev))
2950                 goto out_free_conf;
2951
2952         return 0;
2953
2954 out_free_conf:
2955         md_unregister_thread(mddev->thread);
2956         if (conf->r10bio_pool)
2957                 mempool_destroy(conf->r10bio_pool);
2958         safe_put_page(conf->tmppage);
2959         kfree(conf->mirrors);
2960         kfree(conf);
2961         mddev->private = NULL;
2962 out:
2963         return -EIO;
2964 }
2965
2966 static int stop(mddev_t *mddev)
2967 {
2968         conf_t *conf = mddev->private;
2969
2970         raise_barrier(conf, 0);
2971         lower_barrier(conf);
2972
2973         md_unregister_thread(mddev->thread);
2974         mddev->thread = NULL;
2975         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2976         if (conf->r10bio_pool)
2977                 mempool_destroy(conf->r10bio_pool);
2978         kfree(conf->mirrors);
2979         kfree(conf);
2980         mddev->private = NULL;
2981         return 0;
2982 }
2983
2984 static void raid10_quiesce(mddev_t *mddev, int state)
2985 {
2986         conf_t *conf = mddev->private;
2987
2988         switch(state) {
2989         case 1:
2990                 raise_barrier(conf, 0);
2991                 break;
2992         case 0:
2993                 lower_barrier(conf);
2994                 break;
2995         }
2996 }
2997
2998 static void *raid10_takeover_raid0(mddev_t *mddev)
2999 {
3000         mdk_rdev_t *rdev;
3001         conf_t *conf;
3002
3003         if (mddev->degraded > 0) {
3004                 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3005                        mdname(mddev));
3006                 return ERR_PTR(-EINVAL);
3007         }
3008
3009         /* Set new parameters */
3010         mddev->new_level = 10;
3011         /* new layout: far_copies = 1, near_copies = 2 */
3012         mddev->new_layout = (1<<8) + 2;
3013         mddev->new_chunk_sectors = mddev->chunk_sectors;
3014         mddev->delta_disks = mddev->raid_disks;
3015         mddev->raid_disks *= 2;
3016         /* make sure it will be not marked as dirty */
3017         mddev->recovery_cp = MaxSector;
3018
3019         conf = setup_conf(mddev);
3020         if (!IS_ERR(conf)) {
3021                 list_for_each_entry(rdev, &mddev->disks, same_set)
3022                         if (rdev->raid_disk >= 0)
3023                                 rdev->new_raid_disk = rdev->raid_disk * 2;
3024                 conf->barrier = 1;
3025         }
3026
3027         return conf;
3028 }
3029
3030 static void *raid10_takeover(mddev_t *mddev)
3031 {
3032         struct raid0_private_data *raid0_priv;
3033
3034         /* raid10 can take over:
3035          *  raid0 - providing it has only two drives
3036          */
3037         if (mddev->level == 0) {
3038                 /* for raid0 takeover only one zone is supported */
3039                 raid0_priv = mddev->private;
3040                 if (raid0_priv->nr_strip_zones > 1) {
3041                         printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3042                                " with more than one zone.\n",
3043                                mdname(mddev));
3044                         return ERR_PTR(-EINVAL);
3045                 }
3046                 return raid10_takeover_raid0(mddev);
3047         }
3048         return ERR_PTR(-EINVAL);
3049 }
3050
3051 static struct mdk_personality raid10_personality =
3052 {
3053         .name           = "raid10",
3054         .level          = 10,
3055         .owner          = THIS_MODULE,
3056         .make_request   = make_request,
3057         .run            = run,
3058         .stop           = stop,
3059         .status         = status,
3060         .error_handler  = error,
3061         .hot_add_disk   = raid10_add_disk,
3062         .hot_remove_disk= raid10_remove_disk,
3063         .spare_active   = raid10_spare_active,
3064         .sync_request   = sync_request,
3065         .quiesce        = raid10_quiesce,
3066         .size           = raid10_size,
3067         .takeover       = raid10_takeover,
3068 };
3069
3070 static int __init raid_init(void)
3071 {
3072         return register_md_personality(&raid10_personality);
3073 }
3074
3075 static void raid_exit(void)
3076 {
3077         unregister_md_personality(&raid10_personality);
3078 }
3079
3080 module_init(raid_init);
3081 module_exit(raid_exit);
3082 MODULE_LICENSE("GPL");
3083 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
3084 MODULE_ALIAS("md-personality-9"); /* RAID10 */
3085 MODULE_ALIAS("md-raid10");
3086 MODULE_ALIAS("md-level-10");