7bf5ce281c4d49fee2499b22112a4f8a361349c4
[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 futher copyright information.
9  *
10  *
11  * This program is free software; you can redistribute it and/or modify
12  * it under the terms of the GNU General Public License as published by
13  * the Free Software Foundation; either version 2, or (at your option)
14  * any later version.
15  *
16  * You should have received a copy of the GNU General Public License
17  * (for example /usr/src/linux/COPYING); if not, write to the Free
18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19  */
20
21 #include "dm-bio-list.h"
22 #include <linux/raid/raid10.h>
23 #include <linux/raid/bitmap.h>
24
25 /*
26  * RAID10 provides a combination of RAID0 and RAID1 functionality.
27  * The layout of data is defined by
28  *    chunk_size
29  *    raid_disks
30  *    near_copies (stored in low byte of layout)
31  *    far_copies (stored in second byte of layout)
32  *    far_offset (stored in bit 16 of layout )
33  *
34  * The data to be stored is divided into chunks using chunksize.
35  * Each device is divided into far_copies sections.
36  * In each section, chunks are laid out in a style similar to raid0, but
37  * near_copies copies of each chunk is stored (each on a different drive).
38  * The starting device for each section is offset near_copies from the starting
39  * device of the previous section.
40  * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
41  * drive.
42  * near_copies and far_copies must be at least one, and their product is at most
43  * raid_disks.
44  *
45  * If far_offset is true, then the far_copies are handled a bit differently.
46  * The copies are still in different stripes, but instead of be very far apart
47  * on disk, there are adjacent stripes.
48  */
49
50 /*
51  * Number of guaranteed r10bios in case of extreme VM load:
52  */
53 #define NR_RAID10_BIOS 256
54
55 static void unplug_slaves(mddev_t *mddev);
56
57 static void allow_barrier(conf_t *conf);
58 static void lower_barrier(conf_t *conf);
59
60 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
61 {
62         conf_t *conf = data;
63         r10bio_t *r10_bio;
64         int size = offsetof(struct r10bio_s, devs[conf->copies]);
65
66         /* allocate a r10bio with room for raid_disks entries in the bios array */
67         r10_bio = kzalloc(size, gfp_flags);
68         if (!r10_bio)
69                 unplug_slaves(conf->mddev);
70
71         return r10_bio;
72 }
73
74 static void r10bio_pool_free(void *r10_bio, void *data)
75 {
76         kfree(r10_bio);
77 }
78
79 #define RESYNC_BLOCK_SIZE (64*1024)
80 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
81 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
82 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
83 #define RESYNC_WINDOW (2048*1024)
84
85 /*
86  * When performing a resync, we need to read and compare, so
87  * we need as many pages are there are copies.
88  * When performing a recovery, we need 2 bios, one for read,
89  * one for write (we recover only one drive per r10buf)
90  *
91  */
92 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
93 {
94         conf_t *conf = data;
95         struct page *page;
96         r10bio_t *r10_bio;
97         struct bio *bio;
98         int i, j;
99         int nalloc;
100
101         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
102         if (!r10_bio) {
103                 unplug_slaves(conf->mddev);
104                 return NULL;
105         }
106
107         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
108                 nalloc = conf->copies; /* resync */
109         else
110                 nalloc = 2; /* recovery */
111
112         /*
113          * Allocate bios.
114          */
115         for (j = nalloc ; j-- ; ) {
116                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
117                 if (!bio)
118                         goto out_free_bio;
119                 r10_bio->devs[j].bio = bio;
120         }
121         /*
122          * Allocate RESYNC_PAGES data pages and attach them
123          * where needed.
124          */
125         for (j = 0 ; j < nalloc; j++) {
126                 bio = r10_bio->devs[j].bio;
127                 for (i = 0; i < RESYNC_PAGES; i++) {
128                         page = alloc_page(gfp_flags);
129                         if (unlikely(!page))
130                                 goto out_free_pages;
131
132                         bio->bi_io_vec[i].bv_page = page;
133                 }
134         }
135
136         return r10_bio;
137
138 out_free_pages:
139         for ( ; i > 0 ; i--)
140                 safe_put_page(bio->bi_io_vec[i-1].bv_page);
141         while (j--)
142                 for (i = 0; i < RESYNC_PAGES ; i++)
143                         safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
144         j = -1;
145 out_free_bio:
146         while ( ++j < nalloc )
147                 bio_put(r10_bio->devs[j].bio);
148         r10bio_pool_free(r10_bio, conf);
149         return NULL;
150 }
151
152 static void r10buf_pool_free(void *__r10_bio, void *data)
153 {
154         int i;
155         conf_t *conf = data;
156         r10bio_t *r10bio = __r10_bio;
157         int j;
158
159         for (j=0; j < conf->copies; j++) {
160                 struct bio *bio = r10bio->devs[j].bio;
161                 if (bio) {
162                         for (i = 0; i < RESYNC_PAGES; i++) {
163                                 safe_put_page(bio->bi_io_vec[i].bv_page);
164                                 bio->bi_io_vec[i].bv_page = NULL;
165                         }
166                         bio_put(bio);
167                 }
168         }
169         r10bio_pool_free(r10bio, conf);
170 }
171
172 static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
173 {
174         int i;
175
176         for (i = 0; i < conf->copies; i++) {
177                 struct bio **bio = & r10_bio->devs[i].bio;
178                 if (*bio && *bio != IO_BLOCKED)
179                         bio_put(*bio);
180                 *bio = NULL;
181         }
182 }
183
184 static void free_r10bio(r10bio_t *r10_bio)
185 {
186         conf_t *conf = mddev_to_conf(r10_bio->mddev);
187
188         /*
189          * Wake up any possible resync thread that waits for the device
190          * to go idle.
191          */
192         allow_barrier(conf);
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 = mddev_to_conf(r10_bio->mddev);
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_to_conf(mddev);
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         md_wakeup_thread(mddev->thread);
219 }
220
221 /*
222  * raid_end_bio_io() is called when we have finished servicing a mirrored
223  * operation and are ready to return a success/failure code to the buffer
224  * cache layer.
225  */
226 static void raid_end_bio_io(r10bio_t *r10_bio)
227 {
228         struct bio *bio = r10_bio->master_bio;
229
230         bio_endio(bio, bio->bi_size,
231                 test_bit(R10BIO_Uptodate, &r10_bio->state) ? 0 : -EIO);
232         free_r10bio(r10_bio);
233 }
234
235 /*
236  * Update disk head position estimator based on IRQ completion info.
237  */
238 static inline void update_head_pos(int slot, r10bio_t *r10_bio)
239 {
240         conf_t *conf = mddev_to_conf(r10_bio->mddev);
241
242         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
243                 r10_bio->devs[slot].addr + (r10_bio->sectors);
244 }
245
246 static int raid10_end_read_request(struct bio *bio, unsigned int bytes_done, int error)
247 {
248         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
249         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
250         int slot, dev;
251         conf_t *conf = mddev_to_conf(r10_bio->mddev);
252
253         if (bio->bi_size)
254                 return 1;
255
256         slot = r10_bio->read_slot;
257         dev = r10_bio->devs[slot].devnum;
258         /*
259          * this branch is our 'one mirror IO has finished' event handler:
260          */
261         update_head_pos(slot, r10_bio);
262
263         if (uptodate) {
264                 /*
265                  * Set R10BIO_Uptodate in our master bio, so that
266                  * we will return a good error code to the higher
267                  * levels even if IO on some other mirrored buffer fails.
268                  *
269                  * The 'master' represents the composite IO operation to
270                  * user-side. So if something waits for IO, then it will
271                  * wait for the 'master' bio.
272                  */
273                 set_bit(R10BIO_Uptodate, &r10_bio->state);
274                 raid_end_bio_io(r10_bio);
275         } else {
276                 /*
277                  * oops, read error:
278                  */
279                 char b[BDEVNAME_SIZE];
280                 if (printk_ratelimit())
281                         printk(KERN_ERR "raid10: %s: rescheduling sector %llu\n",
282                                bdevname(conf->mirrors[dev].rdev->bdev,b), (unsigned long long)r10_bio->sector);
283                 reschedule_retry(r10_bio);
284         }
285
286         rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
287         return 0;
288 }
289
290 static int raid10_end_write_request(struct bio *bio, unsigned int bytes_done, int error)
291 {
292         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
293         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
294         int slot, dev;
295         conf_t *conf = mddev_to_conf(r10_bio->mddev);
296
297         if (bio->bi_size)
298                 return 1;
299
300         for (slot = 0; slot < conf->copies; slot++)
301                 if (r10_bio->devs[slot].bio == bio)
302                         break;
303         dev = r10_bio->devs[slot].devnum;
304
305         /*
306          * this branch is our 'one mirror IO has finished' event handler:
307          */
308         if (!uptodate) {
309                 md_error(r10_bio->mddev, conf->mirrors[dev].rdev);
310                 /* an I/O failed, we can't clear the bitmap */
311                 set_bit(R10BIO_Degraded, &r10_bio->state);
312         } else
313                 /*
314                  * Set R10BIO_Uptodate in our master bio, so that
315                  * we will return a good error code for to the higher
316                  * levels even if IO on some other mirrored buffer fails.
317                  *
318                  * The 'master' represents the composite IO operation to
319                  * user-side. So if something waits for IO, then it will
320                  * wait for the 'master' bio.
321                  */
322                 set_bit(R10BIO_Uptodate, &r10_bio->state);
323
324         update_head_pos(slot, r10_bio);
325
326         /*
327          *
328          * Let's see if all mirrored write operations have finished
329          * already.
330          */
331         if (atomic_dec_and_test(&r10_bio->remaining)) {
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                 raid_end_bio_io(r10_bio);
339         }
340
341         rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
342         return 0;
343 }
344
345
346 /*
347  * RAID10 layout manager
348  * Aswell as the chunksize and raid_disks count, there are two
349  * parameters: near_copies and far_copies.
350  * near_copies * far_copies must be <= raid_disks.
351  * Normally one of these will be 1.
352  * If both are 1, we get raid0.
353  * If near_copies == raid_disks, we get raid1.
354  *
355  * Chunks are layed out in raid0 style with near_copies copies of the
356  * first chunk, followed by near_copies copies of the next chunk and
357  * so on.
358  * If far_copies > 1, then after 1/far_copies of the array has been assigned
359  * as described above, we start again with a device offset of near_copies.
360  * So we effectively have another copy of the whole array further down all
361  * the drives, but with blocks on different drives.
362  * With this layout, and block is never stored twice on the one device.
363  *
364  * raid10_find_phys finds the sector offset of a given virtual sector
365  * on each device that it is on.
366  *
367  * raid10_find_virt does the reverse mapping, from a device and a
368  * sector offset to a virtual address
369  */
370
371 static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
372 {
373         int n,f;
374         sector_t sector;
375         sector_t chunk;
376         sector_t stripe;
377         int dev;
378
379         int slot = 0;
380
381         /* now calculate first sector/dev */
382         chunk = r10bio->sector >> conf->chunk_shift;
383         sector = r10bio->sector & conf->chunk_mask;
384
385         chunk *= conf->near_copies;
386         stripe = chunk;
387         dev = sector_div(stripe, conf->raid_disks);
388         if (conf->far_offset)
389                 stripe *= conf->far_copies;
390
391         sector += stripe << conf->chunk_shift;
392
393         /* and calculate all the others */
394         for (n=0; n < conf->near_copies; n++) {
395                 int d = dev;
396                 sector_t s = sector;
397                 r10bio->devs[slot].addr = sector;
398                 r10bio->devs[slot].devnum = d;
399                 slot++;
400
401                 for (f = 1; f < conf->far_copies; f++) {
402                         d += conf->near_copies;
403                         if (d >= conf->raid_disks)
404                                 d -= conf->raid_disks;
405                         s += conf->stride;
406                         r10bio->devs[slot].devnum = d;
407                         r10bio->devs[slot].addr = s;
408                         slot++;
409                 }
410                 dev++;
411                 if (dev >= conf->raid_disks) {
412                         dev = 0;
413                         sector += (conf->chunk_mask + 1);
414                 }
415         }
416         BUG_ON(slot != conf->copies);
417 }
418
419 static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
420 {
421         sector_t offset, chunk, vchunk;
422
423         offset = sector & conf->chunk_mask;
424         if (conf->far_offset) {
425                 int fc;
426                 chunk = sector >> conf->chunk_shift;
427                 fc = sector_div(chunk, conf->far_copies);
428                 dev -= fc * conf->near_copies;
429                 if (dev < 0)
430                         dev += conf->raid_disks;
431         } else {
432                 while (sector > conf->stride) {
433                         sector -= conf->stride;
434                         if (dev < conf->near_copies)
435                                 dev += conf->raid_disks - conf->near_copies;
436                         else
437                                 dev -= conf->near_copies;
438                 }
439                 chunk = sector >> conf->chunk_shift;
440         }
441         vchunk = chunk * conf->raid_disks + dev;
442         sector_div(vchunk, conf->near_copies);
443         return (vchunk << conf->chunk_shift) + offset;
444 }
445
446 /**
447  *      raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
448  *      @q: request queue
449  *      @bio: the buffer head that's been built up so far
450  *      @biovec: the request that could be merged to it.
451  *
452  *      Return amount of bytes we can accept at this offset
453  *      If near_copies == raid_disk, there are no striping issues,
454  *      but in that case, the function isn't called at all.
455  */
456 static int raid10_mergeable_bvec(request_queue_t *q, struct bio *bio,
457                                 struct bio_vec *bio_vec)
458 {
459         mddev_t *mddev = q->queuedata;
460         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
461         int max;
462         unsigned int chunk_sectors = mddev->chunk_size >> 9;
463         unsigned int bio_sectors = bio->bi_size >> 9;
464
465         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
466         if (max < 0) max = 0; /* bio_add cannot handle a negative return */
467         if (max <= bio_vec->bv_len && bio_sectors == 0)
468                 return bio_vec->bv_len;
469         else
470                 return max;
471 }
472
473 /*
474  * This routine returns the disk from which the requested read should
475  * be done. There is a per-array 'next expected sequential IO' sector
476  * number - if this matches on the next IO then we use the last disk.
477  * There is also a per-disk 'last know head position' sector that is
478  * maintained from IRQ contexts, both the normal and the resync IO
479  * completion handlers update this position correctly. If there is no
480  * perfect sequential match then we pick the disk whose head is closest.
481  *
482  * If there are 2 mirrors in the same 2 devices, performance degrades
483  * because position is mirror, not device based.
484  *
485  * The rdev for the device selected will have nr_pending incremented.
486  */
487
488 /*
489  * FIXME: possibly should rethink readbalancing and do it differently
490  * depending on near_copies / far_copies geometry.
491  */
492 static int read_balance(conf_t *conf, r10bio_t *r10_bio)
493 {
494         const unsigned long this_sector = r10_bio->sector;
495         int disk, slot, nslot;
496         const int sectors = r10_bio->sectors;
497         sector_t new_distance, current_distance;
498         mdk_rdev_t *rdev;
499
500         raid10_find_phys(conf, r10_bio);
501         rcu_read_lock();
502         /*
503          * Check if we can balance. We can balance on the whole
504          * device if no resync is going on (recovery is ok), or below
505          * the resync window. We take the first readable disk when
506          * above the resync window.
507          */
508         if (conf->mddev->recovery_cp < MaxSector
509             && (this_sector + sectors >= conf->next_resync)) {
510                 /* make sure that disk is operational */
511                 slot = 0;
512                 disk = r10_bio->devs[slot].devnum;
513
514                 while ((rdev = rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
515                        r10_bio->devs[slot].bio == IO_BLOCKED ||
516                        !test_bit(In_sync, &rdev->flags)) {
517                         slot++;
518                         if (slot == conf->copies) {
519                                 slot = 0;
520                                 disk = -1;
521                                 break;
522                         }
523                         disk = r10_bio->devs[slot].devnum;
524                 }
525                 goto rb_out;
526         }
527
528
529         /* make sure the disk is operational */
530         slot = 0;
531         disk = r10_bio->devs[slot].devnum;
532         while ((rdev=rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
533                r10_bio->devs[slot].bio == IO_BLOCKED ||
534                !test_bit(In_sync, &rdev->flags)) {
535                 slot ++;
536                 if (slot == conf->copies) {
537                         disk = -1;
538                         goto rb_out;
539                 }
540                 disk = r10_bio->devs[slot].devnum;
541         }
542
543
544         current_distance = abs(r10_bio->devs[slot].addr -
545                                conf->mirrors[disk].head_position);
546
547         /* Find the disk whose head is closest */
548
549         for (nslot = slot; nslot < conf->copies; nslot++) {
550                 int ndisk = r10_bio->devs[nslot].devnum;
551
552
553                 if ((rdev=rcu_dereference(conf->mirrors[ndisk].rdev)) == NULL ||
554                     r10_bio->devs[nslot].bio == IO_BLOCKED ||
555                     !test_bit(In_sync, &rdev->flags))
556                         continue;
557
558                 /* This optimisation is debatable, and completely destroys
559                  * sequential read speed for 'far copies' arrays.  So only
560                  * keep it for 'near' arrays, and review those later.
561                  */
562                 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending)) {
563                         disk = ndisk;
564                         slot = nslot;
565                         break;
566                 }
567                 new_distance = abs(r10_bio->devs[nslot].addr -
568                                    conf->mirrors[ndisk].head_position);
569                 if (new_distance < current_distance) {
570                         current_distance = new_distance;
571                         disk = ndisk;
572                         slot = nslot;
573                 }
574         }
575
576 rb_out:
577         r10_bio->read_slot = slot;
578 /*      conf->next_seq_sect = this_sector + sectors;*/
579
580         if (disk >= 0 && (rdev=rcu_dereference(conf->mirrors[disk].rdev))!= NULL)
581                 atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
582         else
583                 disk = -1;
584         rcu_read_unlock();
585
586         return disk;
587 }
588
589 static void unplug_slaves(mddev_t *mddev)
590 {
591         conf_t *conf = mddev_to_conf(mddev);
592         int i;
593
594         rcu_read_lock();
595         for (i=0; i<mddev->raid_disks; i++) {
596                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
597                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
598                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
599
600                         atomic_inc(&rdev->nr_pending);
601                         rcu_read_unlock();
602
603                         if (r_queue->unplug_fn)
604                                 r_queue->unplug_fn(r_queue);
605
606                         rdev_dec_pending(rdev, mddev);
607                         rcu_read_lock();
608                 }
609         }
610         rcu_read_unlock();
611 }
612
613 static void raid10_unplug(request_queue_t *q)
614 {
615         mddev_t *mddev = q->queuedata;
616
617         unplug_slaves(q->queuedata);
618         md_wakeup_thread(mddev->thread);
619 }
620
621 static int raid10_issue_flush(request_queue_t *q, struct gendisk *disk,
622                              sector_t *error_sector)
623 {
624         mddev_t *mddev = q->queuedata;
625         conf_t *conf = mddev_to_conf(mddev);
626         int i, ret = 0;
627
628         rcu_read_lock();
629         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
630                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
631                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
632                         struct block_device *bdev = rdev->bdev;
633                         request_queue_t *r_queue = bdev_get_queue(bdev);
634
635                         if (!r_queue->issue_flush_fn)
636                                 ret = -EOPNOTSUPP;
637                         else {
638                                 atomic_inc(&rdev->nr_pending);
639                                 rcu_read_unlock();
640                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
641                                                               error_sector);
642                                 rdev_dec_pending(rdev, mddev);
643                                 rcu_read_lock();
644                         }
645                 }
646         }
647         rcu_read_unlock();
648         return ret;
649 }
650
651 /* Barriers....
652  * Sometimes we need to suspend IO while we do something else,
653  * either some resync/recovery, or reconfigure the array.
654  * To do this we raise a 'barrier'.
655  * The 'barrier' is a counter that can be raised multiple times
656  * to count how many activities are happening which preclude
657  * normal IO.
658  * We can only raise the barrier if there is no pending IO.
659  * i.e. if nr_pending == 0.
660  * We choose only to raise the barrier if no-one is waiting for the
661  * barrier to go down.  This means that as soon as an IO request
662  * is ready, no other operations which require a barrier will start
663  * until the IO request has had a chance.
664  *
665  * So: regular IO calls 'wait_barrier'.  When that returns there
666  *    is no backgroup IO happening,  It must arrange to call
667  *    allow_barrier when it has finished its IO.
668  * backgroup IO calls must call raise_barrier.  Once that returns
669  *    there is no normal IO happeing.  It must arrange to call
670  *    lower_barrier when the particular background IO completes.
671  */
672 #define RESYNC_DEPTH 32
673
674 static void raise_barrier(conf_t *conf, int force)
675 {
676         BUG_ON(force && !conf->barrier);
677         spin_lock_irq(&conf->resync_lock);
678
679         /* Wait until no block IO is waiting (unless 'force') */
680         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
681                             conf->resync_lock,
682                             raid10_unplug(conf->mddev->queue));
683
684         /* block any new IO from starting */
685         conf->barrier++;
686
687         /* No wait for all pending IO to complete */
688         wait_event_lock_irq(conf->wait_barrier,
689                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
690                             conf->resync_lock,
691                             raid10_unplug(conf->mddev->queue));
692
693         spin_unlock_irq(&conf->resync_lock);
694 }
695
696 static void lower_barrier(conf_t *conf)
697 {
698         unsigned long flags;
699         spin_lock_irqsave(&conf->resync_lock, flags);
700         conf->barrier--;
701         spin_unlock_irqrestore(&conf->resync_lock, flags);
702         wake_up(&conf->wait_barrier);
703 }
704
705 static void wait_barrier(conf_t *conf)
706 {
707         spin_lock_irq(&conf->resync_lock);
708         if (conf->barrier) {
709                 conf->nr_waiting++;
710                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
711                                     conf->resync_lock,
712                                     raid10_unplug(conf->mddev->queue));
713                 conf->nr_waiting--;
714         }
715         conf->nr_pending++;
716         spin_unlock_irq(&conf->resync_lock);
717 }
718
719 static void allow_barrier(conf_t *conf)
720 {
721         unsigned long flags;
722         spin_lock_irqsave(&conf->resync_lock, flags);
723         conf->nr_pending--;
724         spin_unlock_irqrestore(&conf->resync_lock, flags);
725         wake_up(&conf->wait_barrier);
726 }
727
728 static void freeze_array(conf_t *conf)
729 {
730         /* stop syncio and normal IO and wait for everything to
731          * go quiet.
732          * We increment barrier and nr_waiting, and then
733          * wait until barrier+nr_pending match nr_queued+2
734          */
735         spin_lock_irq(&conf->resync_lock);
736         conf->barrier++;
737         conf->nr_waiting++;
738         wait_event_lock_irq(conf->wait_barrier,
739                             conf->barrier+conf->nr_pending == conf->nr_queued+2,
740                             conf->resync_lock,
741                             raid10_unplug(conf->mddev->queue));
742         spin_unlock_irq(&conf->resync_lock);
743 }
744
745 static void unfreeze_array(conf_t *conf)
746 {
747         /* reverse the effect of the freeze */
748         spin_lock_irq(&conf->resync_lock);
749         conf->barrier--;
750         conf->nr_waiting--;
751         wake_up(&conf->wait_barrier);
752         spin_unlock_irq(&conf->resync_lock);
753 }
754
755 static int make_request(request_queue_t *q, struct bio * bio)
756 {
757         mddev_t *mddev = q->queuedata;
758         conf_t *conf = mddev_to_conf(mddev);
759         mirror_info_t *mirror;
760         r10bio_t *r10_bio;
761         struct bio *read_bio;
762         int i;
763         int chunk_sects = conf->chunk_mask + 1;
764         const int rw = bio_data_dir(bio);
765         struct bio_list bl;
766         unsigned long flags;
767
768         if (unlikely(bio_barrier(bio))) {
769                 bio_endio(bio, bio->bi_size, -EOPNOTSUPP);
770                 return 0;
771         }
772
773         /* If this request crosses a chunk boundary, we need to
774          * split it.  This will only happen for 1 PAGE (or less) requests.
775          */
776         if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
777                       > chunk_sects &&
778                     conf->near_copies < conf->raid_disks)) {
779                 struct bio_pair *bp;
780                 /* Sanity check -- queue functions should prevent this happening */
781                 if (bio->bi_vcnt != 1 ||
782                     bio->bi_idx != 0)
783                         goto bad_map;
784                 /* This is a one page bio that upper layers
785                  * refuse to split for us, so we need to split it.
786                  */
787                 bp = bio_split(bio, bio_split_pool,
788                                chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
789                 if (make_request(q, &bp->bio1))
790                         generic_make_request(&bp->bio1);
791                 if (make_request(q, &bp->bio2))
792                         generic_make_request(&bp->bio2);
793
794                 bio_pair_release(bp);
795                 return 0;
796         bad_map:
797                 printk("raid10_make_request bug: can't convert block across chunks"
798                        " or bigger than %dk %llu %d\n", chunk_sects/2,
799                        (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
800
801                 bio_io_error(bio, bio->bi_size);
802                 return 0;
803         }
804
805         md_write_start(mddev, bio);
806
807         /*
808          * Register the new request and wait if the reconstruction
809          * thread has put up a bar for new requests.
810          * Continue immediately if no resync is active currently.
811          */
812         wait_barrier(conf);
813
814         disk_stat_inc(mddev->gendisk, ios[rw]);
815         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
816
817         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
818
819         r10_bio->master_bio = bio;
820         r10_bio->sectors = bio->bi_size >> 9;
821
822         r10_bio->mddev = mddev;
823         r10_bio->sector = bio->bi_sector;
824         r10_bio->state = 0;
825
826         if (rw == READ) {
827                 /*
828                  * read balancing logic:
829                  */
830                 int disk = read_balance(conf, r10_bio);
831                 int slot = r10_bio->read_slot;
832                 if (disk < 0) {
833                         raid_end_bio_io(r10_bio);
834                         return 0;
835                 }
836                 mirror = conf->mirrors + disk;
837
838                 read_bio = bio_clone(bio, GFP_NOIO);
839
840                 r10_bio->devs[slot].bio = read_bio;
841
842                 read_bio->bi_sector = r10_bio->devs[slot].addr +
843                         mirror->rdev->data_offset;
844                 read_bio->bi_bdev = mirror->rdev->bdev;
845                 read_bio->bi_end_io = raid10_end_read_request;
846                 read_bio->bi_rw = READ;
847                 read_bio->bi_private = r10_bio;
848
849                 generic_make_request(read_bio);
850                 return 0;
851         }
852
853         /*
854          * WRITE:
855          */
856         /* first select target devices under spinlock and
857          * inc refcount on their rdev.  Record them by setting
858          * bios[x] to bio
859          */
860         raid10_find_phys(conf, r10_bio);
861         rcu_read_lock();
862         for (i = 0;  i < conf->copies; i++) {
863                 int d = r10_bio->devs[i].devnum;
864                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
865                 if (rdev &&
866                     !test_bit(Faulty, &rdev->flags)) {
867                         atomic_inc(&rdev->nr_pending);
868                         r10_bio->devs[i].bio = bio;
869                 } else {
870                         r10_bio->devs[i].bio = NULL;
871                         set_bit(R10BIO_Degraded, &r10_bio->state);
872                 }
873         }
874         rcu_read_unlock();
875
876         atomic_set(&r10_bio->remaining, 0);
877
878         bio_list_init(&bl);
879         for (i = 0; i < conf->copies; i++) {
880                 struct bio *mbio;
881                 int d = r10_bio->devs[i].devnum;
882                 if (!r10_bio->devs[i].bio)
883                         continue;
884
885                 mbio = bio_clone(bio, GFP_NOIO);
886                 r10_bio->devs[i].bio = mbio;
887
888                 mbio->bi_sector = r10_bio->devs[i].addr+
889                         conf->mirrors[d].rdev->data_offset;
890                 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
891                 mbio->bi_end_io = raid10_end_write_request;
892                 mbio->bi_rw = WRITE;
893                 mbio->bi_private = r10_bio;
894
895                 atomic_inc(&r10_bio->remaining);
896                 bio_list_add(&bl, mbio);
897         }
898
899         bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
900         spin_lock_irqsave(&conf->device_lock, flags);
901         bio_list_merge(&conf->pending_bio_list, &bl);
902         blk_plug_device(mddev->queue);
903         spin_unlock_irqrestore(&conf->device_lock, flags);
904
905         return 0;
906 }
907
908 static void status(struct seq_file *seq, mddev_t *mddev)
909 {
910         conf_t *conf = mddev_to_conf(mddev);
911         int i;
912
913         if (conf->near_copies < conf->raid_disks)
914                 seq_printf(seq, " %dK chunks", mddev->chunk_size/1024);
915         if (conf->near_copies > 1)
916                 seq_printf(seq, " %d near-copies", conf->near_copies);
917         if (conf->far_copies > 1) {
918                 if (conf->far_offset)
919                         seq_printf(seq, " %d offset-copies", conf->far_copies);
920                 else
921                         seq_printf(seq, " %d far-copies", conf->far_copies);
922         }
923         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
924                                                 conf->working_disks);
925         for (i = 0; i < conf->raid_disks; i++)
926                 seq_printf(seq, "%s",
927                               conf->mirrors[i].rdev &&
928                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
929         seq_printf(seq, "]");
930 }
931
932 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
933 {
934         char b[BDEVNAME_SIZE];
935         conf_t *conf = mddev_to_conf(mddev);
936
937         /*
938          * If it is not operational, then we have already marked it as dead
939          * else if it is the last working disks, ignore the error, let the
940          * next level up know.
941          * else mark the drive as failed
942          */
943         if (test_bit(In_sync, &rdev->flags)
944             && conf->working_disks == 1)
945                 /*
946                  * Don't fail the drive, just return an IO error.
947                  * The test should really be more sophisticated than
948                  * "working_disks == 1", but it isn't critical, and
949                  * can wait until we do more sophisticated "is the drive
950                  * really dead" tests...
951                  */
952                 return;
953         if (test_bit(In_sync, &rdev->flags)) {
954                 mddev->degraded++;
955                 conf->working_disks--;
956                 /*
957                  * if recovery is running, make sure it aborts.
958                  */
959                 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
960         }
961         clear_bit(In_sync, &rdev->flags);
962         set_bit(Faulty, &rdev->flags);
963         mddev->sb_dirty = 1;
964         printk(KERN_ALERT "raid10: Disk failure on %s, disabling device. \n"
965                 "       Operation continuing on %d devices\n",
966                 bdevname(rdev->bdev,b), conf->working_disks);
967 }
968
969 static void print_conf(conf_t *conf)
970 {
971         int i;
972         mirror_info_t *tmp;
973
974         printk("RAID10 conf printout:\n");
975         if (!conf) {
976                 printk("(!conf)\n");
977                 return;
978         }
979         printk(" --- wd:%d rd:%d\n", conf->working_disks,
980                 conf->raid_disks);
981
982         for (i = 0; i < conf->raid_disks; i++) {
983                 char b[BDEVNAME_SIZE];
984                 tmp = conf->mirrors + i;
985                 if (tmp->rdev)
986                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
987                                 i, !test_bit(In_sync, &tmp->rdev->flags),
988                                 !test_bit(Faulty, &tmp->rdev->flags),
989                                 bdevname(tmp->rdev->bdev,b));
990         }
991 }
992
993 static void close_sync(conf_t *conf)
994 {
995         wait_barrier(conf);
996         allow_barrier(conf);
997
998         mempool_destroy(conf->r10buf_pool);
999         conf->r10buf_pool = NULL;
1000 }
1001
1002 /* check if there are enough drives for
1003  * every block to appear on atleast one
1004  */
1005 static int enough(conf_t *conf)
1006 {
1007         int first = 0;
1008
1009         do {
1010                 int n = conf->copies;
1011                 int cnt = 0;
1012                 while (n--) {
1013                         if (conf->mirrors[first].rdev)
1014                                 cnt++;
1015                         first = (first+1) % conf->raid_disks;
1016                 }
1017                 if (cnt == 0)
1018                         return 0;
1019         } while (first != 0);
1020         return 1;
1021 }
1022
1023 static int raid10_spare_active(mddev_t *mddev)
1024 {
1025         int i;
1026         conf_t *conf = mddev->private;
1027         mirror_info_t *tmp;
1028
1029         /*
1030          * Find all non-in_sync disks within the RAID10 configuration
1031          * and mark them in_sync
1032          */
1033         for (i = 0; i < conf->raid_disks; i++) {
1034                 tmp = conf->mirrors + i;
1035                 if (tmp->rdev
1036                     && !test_bit(Faulty, &tmp->rdev->flags)
1037                     && !test_bit(In_sync, &tmp->rdev->flags)) {
1038                         conf->working_disks++;
1039                         mddev->degraded--;
1040                         set_bit(In_sync, &tmp->rdev->flags);
1041                 }
1042         }
1043
1044         print_conf(conf);
1045         return 0;
1046 }
1047
1048
1049 static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1050 {
1051         conf_t *conf = mddev->private;
1052         int found = 0;
1053         int mirror;
1054         mirror_info_t *p;
1055
1056         if (mddev->recovery_cp < MaxSector)
1057                 /* only hot-add to in-sync arrays, as recovery is
1058                  * very different from resync
1059                  */
1060                 return 0;
1061         if (!enough(conf))
1062                 return 0;
1063
1064         if (rdev->saved_raid_disk >= 0 &&
1065             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1066                 mirror = rdev->saved_raid_disk;
1067         else
1068                 mirror = 0;
1069         for ( ; mirror < mddev->raid_disks; mirror++)
1070                 if ( !(p=conf->mirrors+mirror)->rdev) {
1071
1072                         blk_queue_stack_limits(mddev->queue,
1073                                                rdev->bdev->bd_disk->queue);
1074                         /* as we don't honour merge_bvec_fn, we must never risk
1075                          * violating it, so limit ->max_sector to one PAGE, as
1076                          * a one page request is never in violation.
1077                          */
1078                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1079                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
1080                                 mddev->queue->max_sectors = (PAGE_SIZE>>9);
1081
1082                         p->head_position = 0;
1083                         rdev->raid_disk = mirror;
1084                         found = 1;
1085                         if (rdev->saved_raid_disk != mirror)
1086                                 conf->fullsync = 1;
1087                         rcu_assign_pointer(p->rdev, rdev);
1088                         break;
1089                 }
1090
1091         print_conf(conf);
1092         return found;
1093 }
1094
1095 static int raid10_remove_disk(mddev_t *mddev, int number)
1096 {
1097         conf_t *conf = mddev->private;
1098         int err = 0;
1099         mdk_rdev_t *rdev;
1100         mirror_info_t *p = conf->mirrors+ number;
1101
1102         print_conf(conf);
1103         rdev = p->rdev;
1104         if (rdev) {
1105                 if (test_bit(In_sync, &rdev->flags) ||
1106                     atomic_read(&rdev->nr_pending)) {
1107                         err = -EBUSY;
1108                         goto abort;
1109                 }
1110                 p->rdev = NULL;
1111                 synchronize_rcu();
1112                 if (atomic_read(&rdev->nr_pending)) {
1113                         /* lost the race, try later */
1114                         err = -EBUSY;
1115                         p->rdev = rdev;
1116                 }
1117         }
1118 abort:
1119
1120         print_conf(conf);
1121         return err;
1122 }
1123
1124
1125 static int end_sync_read(struct bio *bio, unsigned int bytes_done, int error)
1126 {
1127         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1128         conf_t *conf = mddev_to_conf(r10_bio->mddev);
1129         int i,d;
1130
1131         if (bio->bi_size)
1132                 return 1;
1133
1134         for (i=0; i<conf->copies; i++)
1135                 if (r10_bio->devs[i].bio == bio)
1136                         break;
1137         BUG_ON(i == conf->copies);
1138         update_head_pos(i, r10_bio);
1139         d = r10_bio->devs[i].devnum;
1140
1141         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1142                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1143         else {
1144                 atomic_add(r10_bio->sectors,
1145                            &conf->mirrors[d].rdev->corrected_errors);
1146                 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1147                         md_error(r10_bio->mddev,
1148                                  conf->mirrors[d].rdev);
1149         }
1150
1151         /* for reconstruct, we always reschedule after a read.
1152          * for resync, only after all reads
1153          */
1154         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1155             atomic_dec_and_test(&r10_bio->remaining)) {
1156                 /* we have read all the blocks,
1157                  * do the comparison in process context in raid10d
1158                  */
1159                 reschedule_retry(r10_bio);
1160         }
1161         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1162         return 0;
1163 }
1164
1165 static int end_sync_write(struct bio *bio, unsigned int bytes_done, int error)
1166 {
1167         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1168         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1169         mddev_t *mddev = r10_bio->mddev;
1170         conf_t *conf = mddev_to_conf(mddev);
1171         int i,d;
1172
1173         if (bio->bi_size)
1174                 return 1;
1175
1176         for (i = 0; i < conf->copies; i++)
1177                 if (r10_bio->devs[i].bio == bio)
1178                         break;
1179         d = r10_bio->devs[i].devnum;
1180
1181         if (!uptodate)
1182                 md_error(mddev, conf->mirrors[d].rdev);
1183         update_head_pos(i, r10_bio);
1184
1185         while (atomic_dec_and_test(&r10_bio->remaining)) {
1186                 if (r10_bio->master_bio == NULL) {
1187                         /* the primary of several recovery bios */
1188                         md_done_sync(mddev, r10_bio->sectors, 1);
1189                         put_buf(r10_bio);
1190                         break;
1191                 } else {
1192                         r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1193                         put_buf(r10_bio);
1194                         r10_bio = r10_bio2;
1195                 }
1196         }
1197         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1198         return 0;
1199 }
1200
1201 /*
1202  * Note: sync and recover and handled very differently for raid10
1203  * This code is for resync.
1204  * For resync, we read through virtual addresses and read all blocks.
1205  * If there is any error, we schedule a write.  The lowest numbered
1206  * drive is authoritative.
1207  * However requests come for physical address, so we need to map.
1208  * For every physical address there are raid_disks/copies virtual addresses,
1209  * which is always are least one, but is not necessarly an integer.
1210  * This means that a physical address can span multiple chunks, so we may
1211  * have to submit multiple io requests for a single sync request.
1212  */
1213 /*
1214  * We check if all blocks are in-sync and only write to blocks that
1215  * aren't in sync
1216  */
1217 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1218 {
1219         conf_t *conf = mddev_to_conf(mddev);
1220         int i, first;
1221         struct bio *tbio, *fbio;
1222
1223         atomic_set(&r10_bio->remaining, 1);
1224
1225         /* find the first device with a block */
1226         for (i=0; i<conf->copies; i++)
1227                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1228                         break;
1229
1230         if (i == conf->copies)
1231                 goto done;
1232
1233         first = i;
1234         fbio = r10_bio->devs[i].bio;
1235
1236         /* now find blocks with errors */
1237         for (i=0 ; i < conf->copies ; i++) {
1238                 int  j, d;
1239                 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1240
1241                 tbio = r10_bio->devs[i].bio;
1242
1243                 if (tbio->bi_end_io != end_sync_read)
1244                         continue;
1245                 if (i == first)
1246                         continue;
1247                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1248                         /* We know that the bi_io_vec layout is the same for
1249                          * both 'first' and 'i', so we just compare them.
1250                          * All vec entries are PAGE_SIZE;
1251                          */
1252                         for (j = 0; j < vcnt; j++)
1253                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1254                                            page_address(tbio->bi_io_vec[j].bv_page),
1255                                            PAGE_SIZE))
1256                                         break;
1257                         if (j == vcnt)
1258                                 continue;
1259                         mddev->resync_mismatches += r10_bio->sectors;
1260                 }
1261                 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1262                         /* Don't fix anything. */
1263                         continue;
1264                 /* Ok, we need to write this bio
1265                  * First we need to fixup bv_offset, bv_len and
1266                  * bi_vecs, as the read request might have corrupted these
1267                  */
1268                 tbio->bi_vcnt = vcnt;
1269                 tbio->bi_size = r10_bio->sectors << 9;
1270                 tbio->bi_idx = 0;
1271                 tbio->bi_phys_segments = 0;
1272                 tbio->bi_hw_segments = 0;
1273                 tbio->bi_hw_front_size = 0;
1274                 tbio->bi_hw_back_size = 0;
1275                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1276                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1277                 tbio->bi_next = NULL;
1278                 tbio->bi_rw = WRITE;
1279                 tbio->bi_private = r10_bio;
1280                 tbio->bi_sector = r10_bio->devs[i].addr;
1281
1282                 for (j=0; j < vcnt ; j++) {
1283                         tbio->bi_io_vec[j].bv_offset = 0;
1284                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1285
1286                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1287                                page_address(fbio->bi_io_vec[j].bv_page),
1288                                PAGE_SIZE);
1289                 }
1290                 tbio->bi_end_io = end_sync_write;
1291
1292                 d = r10_bio->devs[i].devnum;
1293                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1294                 atomic_inc(&r10_bio->remaining);
1295                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1296
1297                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1298                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1299                 generic_make_request(tbio);
1300         }
1301
1302 done:
1303         if (atomic_dec_and_test(&r10_bio->remaining)) {
1304                 md_done_sync(mddev, r10_bio->sectors, 1);
1305                 put_buf(r10_bio);
1306         }
1307 }
1308
1309 /*
1310  * Now for the recovery code.
1311  * Recovery happens across physical sectors.
1312  * We recover all non-is_sync drives by finding the virtual address of
1313  * each, and then choose a working drive that also has that virt address.
1314  * There is a separate r10_bio for each non-in_sync drive.
1315  * Only the first two slots are in use. The first for reading,
1316  * The second for writing.
1317  *
1318  */
1319
1320 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1321 {
1322         conf_t *conf = mddev_to_conf(mddev);
1323         int i, d;
1324         struct bio *bio, *wbio;
1325
1326
1327         /* move the pages across to the second bio
1328          * and submit the write request
1329          */
1330         bio = r10_bio->devs[0].bio;
1331         wbio = r10_bio->devs[1].bio;
1332         for (i=0; i < wbio->bi_vcnt; i++) {
1333                 struct page *p = bio->bi_io_vec[i].bv_page;
1334                 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1335                 wbio->bi_io_vec[i].bv_page = p;
1336         }
1337         d = r10_bio->devs[1].devnum;
1338
1339         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1340         md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1341         if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1342                 generic_make_request(wbio);
1343         else
1344                 bio_endio(wbio, wbio->bi_size, -EIO);
1345 }
1346
1347
1348 /*
1349  * This is a kernel thread which:
1350  *
1351  *      1.      Retries failed read operations on working mirrors.
1352  *      2.      Updates the raid superblock when problems encounter.
1353  *      3.      Performs writes following reads for array synchronising.
1354  */
1355
1356 static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1357 {
1358         int sect = 0; /* Offset from r10_bio->sector */
1359         int sectors = r10_bio->sectors;
1360         mdk_rdev_t*rdev;
1361         while(sectors) {
1362                 int s = sectors;
1363                 int sl = r10_bio->read_slot;
1364                 int success = 0;
1365                 int start;
1366
1367                 if (s > (PAGE_SIZE>>9))
1368                         s = PAGE_SIZE >> 9;
1369
1370                 rcu_read_lock();
1371                 do {
1372                         int d = r10_bio->devs[sl].devnum;
1373                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1374                         if (rdev &&
1375                             test_bit(In_sync, &rdev->flags)) {
1376                                 atomic_inc(&rdev->nr_pending);
1377                                 rcu_read_unlock();
1378                                 success = sync_page_io(rdev->bdev,
1379                                                        r10_bio->devs[sl].addr +
1380                                                        sect + rdev->data_offset,
1381                                                        s<<9,
1382                                                        conf->tmppage, READ);
1383                                 rdev_dec_pending(rdev, mddev);
1384                                 rcu_read_lock();
1385                                 if (success)
1386                                         break;
1387                         }
1388                         sl++;
1389                         if (sl == conf->copies)
1390                                 sl = 0;
1391                 } while (!success && sl != r10_bio->read_slot);
1392                 rcu_read_unlock();
1393
1394                 if (!success) {
1395                         /* Cannot read from anywhere -- bye bye array */
1396                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1397                         md_error(mddev, conf->mirrors[dn].rdev);
1398                         break;
1399                 }
1400
1401                 start = sl;
1402                 /* write it back and re-read */
1403                 rcu_read_lock();
1404                 while (sl != r10_bio->read_slot) {
1405                         int d;
1406                         if (sl==0)
1407                                 sl = conf->copies;
1408                         sl--;
1409                         d = r10_bio->devs[sl].devnum;
1410                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1411                         if (rdev &&
1412                             test_bit(In_sync, &rdev->flags)) {
1413                                 atomic_inc(&rdev->nr_pending);
1414                                 rcu_read_unlock();
1415                                 atomic_add(s, &rdev->corrected_errors);
1416                                 if (sync_page_io(rdev->bdev,
1417                                                  r10_bio->devs[sl].addr +
1418                                                  sect + rdev->data_offset,
1419                                                  s<<9, conf->tmppage, WRITE)
1420                                     == 0)
1421                                         /* Well, this device is dead */
1422                                         md_error(mddev, rdev);
1423                                 rdev_dec_pending(rdev, mddev);
1424                                 rcu_read_lock();
1425                         }
1426                 }
1427                 sl = start;
1428                 while (sl != r10_bio->read_slot) {
1429                         int d;
1430                         if (sl==0)
1431                                 sl = conf->copies;
1432                         sl--;
1433                         d = r10_bio->devs[sl].devnum;
1434                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1435                         if (rdev &&
1436                             test_bit(In_sync, &rdev->flags)) {
1437                                 char b[BDEVNAME_SIZE];
1438                                 atomic_inc(&rdev->nr_pending);
1439                                 rcu_read_unlock();
1440                                 if (sync_page_io(rdev->bdev,
1441                                                  r10_bio->devs[sl].addr +
1442                                                  sect + rdev->data_offset,
1443                                                  s<<9, conf->tmppage, READ) == 0)
1444                                         /* Well, this device is dead */
1445                                         md_error(mddev, rdev);
1446                                 else
1447                                         printk(KERN_INFO
1448                                                "raid10:%s: read error corrected"
1449                                                " (%d sectors at %llu on %s)\n",
1450                                                mdname(mddev), s,
1451                                                (unsigned long long)sect+
1452                                                     rdev->data_offset,
1453                                                bdevname(rdev->bdev, b));
1454
1455                                 rdev_dec_pending(rdev, mddev);
1456                                 rcu_read_lock();
1457                         }
1458                 }
1459                 rcu_read_unlock();
1460
1461                 sectors -= s;
1462                 sect += s;
1463         }
1464 }
1465
1466 static void raid10d(mddev_t *mddev)
1467 {
1468         r10bio_t *r10_bio;
1469         struct bio *bio;
1470         unsigned long flags;
1471         conf_t *conf = mddev_to_conf(mddev);
1472         struct list_head *head = &conf->retry_list;
1473         int unplug=0;
1474         mdk_rdev_t *rdev;
1475
1476         md_check_recovery(mddev);
1477
1478         for (;;) {
1479                 char b[BDEVNAME_SIZE];
1480                 spin_lock_irqsave(&conf->device_lock, flags);
1481
1482                 if (conf->pending_bio_list.head) {
1483                         bio = bio_list_get(&conf->pending_bio_list);
1484                         blk_remove_plug(mddev->queue);
1485                         spin_unlock_irqrestore(&conf->device_lock, flags);
1486                         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
1487                         if (bitmap_unplug(mddev->bitmap) != 0)
1488                                 printk("%s: bitmap file write failed!\n", mdname(mddev));
1489
1490                         while (bio) { /* submit pending writes */
1491                                 struct bio *next = bio->bi_next;
1492                                 bio->bi_next = NULL;
1493                                 generic_make_request(bio);
1494                                 bio = next;
1495                         }
1496                         unplug = 1;
1497
1498                         continue;
1499                 }
1500
1501                 if (list_empty(head))
1502                         break;
1503                 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1504                 list_del(head->prev);
1505                 conf->nr_queued--;
1506                 spin_unlock_irqrestore(&conf->device_lock, flags);
1507
1508                 mddev = r10_bio->mddev;
1509                 conf = mddev_to_conf(mddev);
1510                 if (test_bit(R10BIO_IsSync, &r10_bio->state)) {
1511                         sync_request_write(mddev, r10_bio);
1512                         unplug = 1;
1513                 } else  if (test_bit(R10BIO_IsRecover, &r10_bio->state)) {
1514                         recovery_request_write(mddev, r10_bio);
1515                         unplug = 1;
1516                 } else {
1517                         int mirror;
1518                         /* we got a read error. Maybe the drive is bad.  Maybe just
1519                          * the block and we can fix it.
1520                          * We freeze all other IO, and try reading the block from
1521                          * other devices.  When we find one, we re-write
1522                          * and check it that fixes the read error.
1523                          * This is all done synchronously while the array is
1524                          * frozen.
1525                          */
1526                         if (mddev->ro == 0) {
1527                                 freeze_array(conf);
1528                                 fix_read_error(conf, mddev, r10_bio);
1529                                 unfreeze_array(conf);
1530                         }
1531
1532                         bio = r10_bio->devs[r10_bio->read_slot].bio;
1533                         r10_bio->devs[r10_bio->read_slot].bio =
1534                                 mddev->ro ? IO_BLOCKED : NULL;
1535                         bio_put(bio);
1536                         mirror = read_balance(conf, r10_bio);
1537                         if (mirror == -1) {
1538                                 printk(KERN_ALERT "raid10: %s: unrecoverable I/O"
1539                                        " read error for block %llu\n",
1540                                        bdevname(bio->bi_bdev,b),
1541                                        (unsigned long long)r10_bio->sector);
1542                                 raid_end_bio_io(r10_bio);
1543                         } else {
1544                                 rdev = conf->mirrors[mirror].rdev;
1545                                 if (printk_ratelimit())
1546                                         printk(KERN_ERR "raid10: %s: redirecting sector %llu to"
1547                                                " another mirror\n",
1548                                                bdevname(rdev->bdev,b),
1549                                                (unsigned long long)r10_bio->sector);
1550                                 bio = bio_clone(r10_bio->master_bio, GFP_NOIO);
1551                                 r10_bio->devs[r10_bio->read_slot].bio = bio;
1552                                 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1553                                         + rdev->data_offset;
1554                                 bio->bi_bdev = rdev->bdev;
1555                                 bio->bi_rw = READ;
1556                                 bio->bi_private = r10_bio;
1557                                 bio->bi_end_io = raid10_end_read_request;
1558                                 unplug = 1;
1559                                 generic_make_request(bio);
1560                         }
1561                 }
1562         }
1563         spin_unlock_irqrestore(&conf->device_lock, flags);
1564         if (unplug)
1565                 unplug_slaves(mddev);
1566 }
1567
1568
1569 static int init_resync(conf_t *conf)
1570 {
1571         int buffs;
1572
1573         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1574         BUG_ON(conf->r10buf_pool);
1575         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1576         if (!conf->r10buf_pool)
1577                 return -ENOMEM;
1578         conf->next_resync = 0;
1579         return 0;
1580 }
1581
1582 /*
1583  * perform a "sync" on one "block"
1584  *
1585  * We need to make sure that no normal I/O request - particularly write
1586  * requests - conflict with active sync requests.
1587  *
1588  * This is achieved by tracking pending requests and a 'barrier' concept
1589  * that can be installed to exclude normal IO requests.
1590  *
1591  * Resync and recovery are handled very differently.
1592  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1593  *
1594  * For resync, we iterate over virtual addresses, read all copies,
1595  * and update if there are differences.  If only one copy is live,
1596  * skip it.
1597  * For recovery, we iterate over physical addresses, read a good
1598  * value for each non-in_sync drive, and over-write.
1599  *
1600  * So, for recovery we may have several outstanding complex requests for a
1601  * given address, one for each out-of-sync device.  We model this by allocating
1602  * a number of r10_bio structures, one for each out-of-sync device.
1603  * As we setup these structures, we collect all bio's together into a list
1604  * which we then process collectively to add pages, and then process again
1605  * to pass to generic_make_request.
1606  *
1607  * The r10_bio structures are linked using a borrowed master_bio pointer.
1608  * This link is counted in ->remaining.  When the r10_bio that points to NULL
1609  * has its remaining count decremented to 0, the whole complex operation
1610  * is complete.
1611  *
1612  */
1613
1614 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1615 {
1616         conf_t *conf = mddev_to_conf(mddev);
1617         r10bio_t *r10_bio;
1618         struct bio *biolist = NULL, *bio;
1619         sector_t max_sector, nr_sectors;
1620         int disk;
1621         int i;
1622         int max_sync;
1623         int sync_blocks;
1624
1625         sector_t sectors_skipped = 0;
1626         int chunks_skipped = 0;
1627
1628         if (!conf->r10buf_pool)
1629                 if (init_resync(conf))
1630                         return 0;
1631
1632  skipped:
1633         max_sector = mddev->size << 1;
1634         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1635                 max_sector = mddev->resync_max_sectors;
1636         if (sector_nr >= max_sector) {
1637                 /* If we aborted, we need to abort the
1638                  * sync on the 'current' bitmap chucks (there can
1639                  * be several when recovering multiple devices).
1640                  * as we may have started syncing it but not finished.
1641                  * We can find the current address in
1642                  * mddev->curr_resync, but for recovery,
1643                  * we need to convert that to several
1644                  * virtual addresses.
1645                  */
1646                 if (mddev->curr_resync < max_sector) { /* aborted */
1647                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1648                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1649                                                 &sync_blocks, 1);
1650                         else for (i=0; i<conf->raid_disks; i++) {
1651                                 sector_t sect =
1652                                         raid10_find_virt(conf, mddev->curr_resync, i);
1653                                 bitmap_end_sync(mddev->bitmap, sect,
1654                                                 &sync_blocks, 1);
1655                         }
1656                 } else /* completed sync */
1657                         conf->fullsync = 0;
1658
1659                 bitmap_close_sync(mddev->bitmap);
1660                 close_sync(conf);
1661                 *skipped = 1;
1662                 return sectors_skipped;
1663         }
1664         if (chunks_skipped >= conf->raid_disks) {
1665                 /* if there has been nothing to do on any drive,
1666                  * then there is nothing to do at all..
1667                  */
1668                 *skipped = 1;
1669                 return (max_sector - sector_nr) + sectors_skipped;
1670         }
1671
1672         /* make sure whole request will fit in a chunk - if chunks
1673          * are meaningful
1674          */
1675         if (conf->near_copies < conf->raid_disks &&
1676             max_sector > (sector_nr | conf->chunk_mask))
1677                 max_sector = (sector_nr | conf->chunk_mask) + 1;
1678         /*
1679          * If there is non-resync activity waiting for us then
1680          * put in a delay to throttle resync.
1681          */
1682         if (!go_faster && conf->nr_waiting)
1683                 msleep_interruptible(1000);
1684
1685         /* Again, very different code for resync and recovery.
1686          * Both must result in an r10bio with a list of bios that
1687          * have bi_end_io, bi_sector, bi_bdev set,
1688          * and bi_private set to the r10bio.
1689          * For recovery, we may actually create several r10bios
1690          * with 2 bios in each, that correspond to the bios in the main one.
1691          * In this case, the subordinate r10bios link back through a
1692          * borrowed master_bio pointer, and the counter in the master
1693          * includes a ref from each subordinate.
1694          */
1695         /* First, we decide what to do and set ->bi_end_io
1696          * To end_sync_read if we want to read, and
1697          * end_sync_write if we will want to write.
1698          */
1699
1700         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
1701         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1702                 /* recovery... the complicated one */
1703                 int i, j, k;
1704                 r10_bio = NULL;
1705
1706                 for (i=0 ; i<conf->raid_disks; i++)
1707                         if (conf->mirrors[i].rdev &&
1708                             !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) {
1709                                 int still_degraded = 0;
1710                                 /* want to reconstruct this device */
1711                                 r10bio_t *rb2 = r10_bio;
1712                                 sector_t sect = raid10_find_virt(conf, sector_nr, i);
1713                                 int must_sync;
1714                                 /* Unless we are doing a full sync, we only need
1715                                  * to recover the block if it is set in the bitmap
1716                                  */
1717                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1718                                                               &sync_blocks, 1);
1719                                 if (sync_blocks < max_sync)
1720                                         max_sync = sync_blocks;
1721                                 if (!must_sync &&
1722                                     !conf->fullsync) {
1723                                         /* yep, skip the sync_blocks here, but don't assume
1724                                          * that there will never be anything to do here
1725                                          */
1726                                         chunks_skipped = -1;
1727                                         continue;
1728                                 }
1729
1730                                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1731                                 raise_barrier(conf, rb2 != NULL);
1732                                 atomic_set(&r10_bio->remaining, 0);
1733
1734                                 r10_bio->master_bio = (struct bio*)rb2;
1735                                 if (rb2)
1736                                         atomic_inc(&rb2->remaining);
1737                                 r10_bio->mddev = mddev;
1738                                 set_bit(R10BIO_IsRecover, &r10_bio->state);
1739                                 r10_bio->sector = sect;
1740
1741                                 raid10_find_phys(conf, r10_bio);
1742                                 /* Need to check if this section will still be
1743                                  * degraded
1744                                  */
1745                                 for (j=0; j<conf->copies;j++) {
1746                                         int d = r10_bio->devs[j].devnum;
1747                                         if (conf->mirrors[d].rdev == NULL ||
1748                                             test_bit(Faulty, &conf->mirrors[d].rdev->flags)) {
1749                                                 still_degraded = 1;
1750                                                 break;
1751                                         }
1752                                 }
1753                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1754                                                               &sync_blocks, still_degraded);
1755
1756                                 for (j=0; j<conf->copies;j++) {
1757                                         int d = r10_bio->devs[j].devnum;
1758                                         if (conf->mirrors[d].rdev &&
1759                                             test_bit(In_sync, &conf->mirrors[d].rdev->flags)) {
1760                                                 /* This is where we read from */
1761                                                 bio = r10_bio->devs[0].bio;
1762                                                 bio->bi_next = biolist;
1763                                                 biolist = bio;
1764                                                 bio->bi_private = r10_bio;
1765                                                 bio->bi_end_io = end_sync_read;
1766                                                 bio->bi_rw = 0;
1767                                                 bio->bi_sector = r10_bio->devs[j].addr +
1768                                                         conf->mirrors[d].rdev->data_offset;
1769                                                 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1770                                                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1771                                                 atomic_inc(&r10_bio->remaining);
1772                                                 /* and we write to 'i' */
1773
1774                                                 for (k=0; k<conf->copies; k++)
1775                                                         if (r10_bio->devs[k].devnum == i)
1776                                                                 break;
1777                                                 bio = r10_bio->devs[1].bio;
1778                                                 bio->bi_next = biolist;
1779                                                 biolist = bio;
1780                                                 bio->bi_private = r10_bio;
1781                                                 bio->bi_end_io = end_sync_write;
1782                                                 bio->bi_rw = 1;
1783                                                 bio->bi_sector = r10_bio->devs[k].addr +
1784                                                         conf->mirrors[i].rdev->data_offset;
1785                                                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1786
1787                                                 r10_bio->devs[0].devnum = d;
1788                                                 r10_bio->devs[1].devnum = i;
1789
1790                                                 break;
1791                                         }
1792                                 }
1793                                 if (j == conf->copies) {
1794                                         /* Cannot recover, so abort the recovery */
1795                                         put_buf(r10_bio);
1796                                         r10_bio = rb2;
1797                                         if (!test_and_set_bit(MD_RECOVERY_ERR, &mddev->recovery))
1798                                                 printk(KERN_INFO "raid10: %s: insufficient working devices for recovery.\n",
1799                                                        mdname(mddev));
1800                                         break;
1801                                 }
1802                         }
1803                 if (biolist == NULL) {
1804                         while (r10_bio) {
1805                                 r10bio_t *rb2 = r10_bio;
1806                                 r10_bio = (r10bio_t*) rb2->master_bio;
1807                                 rb2->master_bio = NULL;
1808                                 put_buf(rb2);
1809                         }
1810                         goto giveup;
1811                 }
1812         } else {
1813                 /* resync. Schedule a read for every block at this virt offset */
1814                 int count = 0;
1815
1816                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1817                                        &sync_blocks, mddev->degraded) &&
1818                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1819                         /* We can skip this block */
1820                         *skipped = 1;
1821                         return sync_blocks + sectors_skipped;
1822                 }
1823                 if (sync_blocks < max_sync)
1824                         max_sync = sync_blocks;
1825                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1826
1827                 r10_bio->mddev = mddev;
1828                 atomic_set(&r10_bio->remaining, 0);
1829                 raise_barrier(conf, 0);
1830                 conf->next_resync = sector_nr;
1831
1832                 r10_bio->master_bio = NULL;
1833                 r10_bio->sector = sector_nr;
1834                 set_bit(R10BIO_IsSync, &r10_bio->state);
1835                 raid10_find_phys(conf, r10_bio);
1836                 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1837
1838                 for (i=0; i<conf->copies; i++) {
1839                         int d = r10_bio->devs[i].devnum;
1840                         bio = r10_bio->devs[i].bio;
1841                         bio->bi_end_io = NULL;
1842                         if (conf->mirrors[d].rdev == NULL ||
1843                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
1844                                 continue;
1845                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1846                         atomic_inc(&r10_bio->remaining);
1847                         bio->bi_next = biolist;
1848                         biolist = bio;
1849                         bio->bi_private = r10_bio;
1850                         bio->bi_end_io = end_sync_read;
1851                         bio->bi_rw = 0;
1852                         bio->bi_sector = r10_bio->devs[i].addr +
1853                                 conf->mirrors[d].rdev->data_offset;
1854                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1855                         count++;
1856                 }
1857
1858                 if (count < 2) {
1859                         for (i=0; i<conf->copies; i++) {
1860                                 int d = r10_bio->devs[i].devnum;
1861                                 if (r10_bio->devs[i].bio->bi_end_io)
1862                                         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1863                         }
1864                         put_buf(r10_bio);
1865                         biolist = NULL;
1866                         goto giveup;
1867                 }
1868         }
1869
1870         for (bio = biolist; bio ; bio=bio->bi_next) {
1871
1872                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1873                 if (bio->bi_end_io)
1874                         bio->bi_flags |= 1 << BIO_UPTODATE;
1875                 bio->bi_vcnt = 0;
1876                 bio->bi_idx = 0;
1877                 bio->bi_phys_segments = 0;
1878                 bio->bi_hw_segments = 0;
1879                 bio->bi_size = 0;
1880         }
1881
1882         nr_sectors = 0;
1883         if (sector_nr + max_sync < max_sector)
1884                 max_sector = sector_nr + max_sync;
1885         do {
1886                 struct page *page;
1887                 int len = PAGE_SIZE;
1888                 disk = 0;
1889                 if (sector_nr + (len>>9) > max_sector)
1890                         len = (max_sector - sector_nr) << 9;
1891                 if (len == 0)
1892                         break;
1893                 for (bio= biolist ; bio ; bio=bio->bi_next) {
1894                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1895                         if (bio_add_page(bio, page, len, 0) == 0) {
1896                                 /* stop here */
1897                                 struct bio *bio2;
1898                                 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1899                                 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
1900                                         /* remove last page from this bio */
1901                                         bio2->bi_vcnt--;
1902                                         bio2->bi_size -= len;
1903                                         bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
1904                                 }
1905                                 goto bio_full;
1906                         }
1907                         disk = i;
1908                 }
1909                 nr_sectors += len>>9;
1910                 sector_nr += len>>9;
1911         } while (biolist->bi_vcnt < RESYNC_PAGES);
1912  bio_full:
1913         r10_bio->sectors = nr_sectors;
1914
1915         while (biolist) {
1916                 bio = biolist;
1917                 biolist = biolist->bi_next;
1918
1919                 bio->bi_next = NULL;
1920                 r10_bio = bio->bi_private;
1921                 r10_bio->sectors = nr_sectors;
1922
1923                 if (bio->bi_end_io == end_sync_read) {
1924                         md_sync_acct(bio->bi_bdev, nr_sectors);
1925                         generic_make_request(bio);
1926                 }
1927         }
1928
1929         if (sectors_skipped)
1930                 /* pretend they weren't skipped, it makes
1931                  * no important difference in this case
1932                  */
1933                 md_done_sync(mddev, sectors_skipped, 1);
1934
1935         return sectors_skipped + nr_sectors;
1936  giveup:
1937         /* There is nowhere to write, so all non-sync
1938          * drives must be failed, so try the next chunk...
1939          */
1940         {
1941         sector_t sec = max_sector - sector_nr;
1942         sectors_skipped += sec;
1943         chunks_skipped ++;
1944         sector_nr = max_sector;
1945         goto skipped;
1946         }
1947 }
1948
1949 static int run(mddev_t *mddev)
1950 {
1951         conf_t *conf;
1952         int i, disk_idx;
1953         mirror_info_t *disk;
1954         mdk_rdev_t *rdev;
1955         struct list_head *tmp;
1956         int nc, fc, fo;
1957         sector_t stride, size;
1958
1959         if (mddev->chunk_size == 0) {
1960                 printk(KERN_ERR "md/raid10: non-zero chunk size required.\n");
1961                 return -EINVAL;
1962         }
1963
1964         nc = mddev->layout & 255;
1965         fc = (mddev->layout >> 8) & 255;
1966         fo = mddev->layout & (1<<16);
1967         if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
1968             (mddev->layout >> 17)) {
1969                 printk(KERN_ERR "raid10: %s: unsupported raid10 layout: 0x%8x\n",
1970                        mdname(mddev), mddev->layout);
1971                 goto out;
1972         }
1973         /*
1974          * copy the already verified devices into our private RAID10
1975          * bookkeeping area. [whatever we allocate in run(),
1976          * should be freed in stop()]
1977          */
1978         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1979         mddev->private = conf;
1980         if (!conf) {
1981                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1982                         mdname(mddev));
1983                 goto out;
1984         }
1985         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1986                                  GFP_KERNEL);
1987         if (!conf->mirrors) {
1988                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1989                        mdname(mddev));
1990                 goto out_free_conf;
1991         }
1992
1993         conf->tmppage = alloc_page(GFP_KERNEL);
1994         if (!conf->tmppage)
1995                 goto out_free_conf;
1996
1997         conf->near_copies = nc;
1998         conf->far_copies = fc;
1999         conf->copies = nc*fc;
2000         conf->far_offset = fo;
2001         conf->chunk_mask = (sector_t)(mddev->chunk_size>>9)-1;
2002         conf->chunk_shift = ffz(~mddev->chunk_size) - 9;
2003         if (fo)
2004                 conf->stride = 1 << conf->chunk_shift;
2005         else {
2006                 stride = mddev->size >> (conf->chunk_shift-1);
2007                 sector_div(stride, fc);
2008                 conf->stride = stride << conf->chunk_shift;
2009         }
2010         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2011                                                 r10bio_pool_free, conf);
2012         if (!conf->r10bio_pool) {
2013                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2014                         mdname(mddev));
2015                 goto out_free_conf;
2016         }
2017
2018         ITERATE_RDEV(mddev, rdev, tmp) {
2019                 disk_idx = rdev->raid_disk;
2020                 if (disk_idx >= mddev->raid_disks
2021                     || disk_idx < 0)
2022                         continue;
2023                 disk = conf->mirrors + disk_idx;
2024
2025                 disk->rdev = rdev;
2026
2027                 blk_queue_stack_limits(mddev->queue,
2028                                        rdev->bdev->bd_disk->queue);
2029                 /* as we don't honour merge_bvec_fn, we must never risk
2030                  * violating it, so limit ->max_sector to one PAGE, as
2031                  * a one page request is never in violation.
2032                  */
2033                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
2034                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
2035                         mddev->queue->max_sectors = (PAGE_SIZE>>9);
2036
2037                 disk->head_position = 0;
2038                 if (!test_bit(Faulty, &rdev->flags) && test_bit(In_sync, &rdev->flags))
2039                         conf->working_disks++;
2040         }
2041         conf->raid_disks = mddev->raid_disks;
2042         conf->mddev = mddev;
2043         spin_lock_init(&conf->device_lock);
2044         INIT_LIST_HEAD(&conf->retry_list);
2045
2046         spin_lock_init(&conf->resync_lock);
2047         init_waitqueue_head(&conf->wait_barrier);
2048
2049         /* need to check that every block has at least one working mirror */
2050         if (!enough(conf)) {
2051                 printk(KERN_ERR "raid10: not enough operational mirrors for %s\n",
2052                        mdname(mddev));
2053                 goto out_free_conf;
2054         }
2055
2056         mddev->degraded = 0;
2057         for (i = 0; i < conf->raid_disks; i++) {
2058
2059                 disk = conf->mirrors + i;
2060
2061                 if (!disk->rdev ||
2062                     !test_bit(In_sync, &rdev->flags)) {
2063                         disk->head_position = 0;
2064                         mddev->degraded++;
2065                 }
2066         }
2067
2068
2069         mddev->thread = md_register_thread(raid10d, mddev, "%s_raid10");
2070         if (!mddev->thread) {
2071                 printk(KERN_ERR
2072                        "raid10: couldn't allocate thread for %s\n",
2073                        mdname(mddev));
2074                 goto out_free_conf;
2075         }
2076
2077         printk(KERN_INFO
2078                 "raid10: raid set %s active with %d out of %d devices\n",
2079                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2080                 mddev->raid_disks);
2081         /*
2082          * Ok, everything is just fine now
2083          */
2084         if (conf->far_offset) {
2085                 size = mddev->size >> (conf->chunk_shift-1);
2086                 size *= conf->raid_disks;
2087                 size <<= conf->chunk_shift;
2088                 sector_div(size, conf->far_copies);
2089         } else
2090                 size = conf->stride * conf->raid_disks;
2091         sector_div(size, conf->near_copies);
2092         mddev->array_size = size/2;
2093         mddev->resync_max_sectors = size;
2094
2095         mddev->queue->unplug_fn = raid10_unplug;
2096         mddev->queue->issue_flush_fn = raid10_issue_flush;
2097
2098         /* Calculate max read-ahead size.
2099          * We need to readahead at least twice a whole stripe....
2100          * maybe...
2101          */
2102         {
2103                 int stripe = conf->raid_disks * (mddev->chunk_size / PAGE_SIZE);
2104                 stripe /= conf->near_copies;
2105                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2106                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2107         }
2108
2109         if (conf->near_copies < mddev->raid_disks)
2110                 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2111         return 0;
2112
2113 out_free_conf:
2114         if (conf->r10bio_pool)
2115                 mempool_destroy(conf->r10bio_pool);
2116         safe_put_page(conf->tmppage);
2117         kfree(conf->mirrors);
2118         kfree(conf);
2119         mddev->private = NULL;
2120 out:
2121         return -EIO;
2122 }
2123
2124 static int stop(mddev_t *mddev)
2125 {
2126         conf_t *conf = mddev_to_conf(mddev);
2127
2128         md_unregister_thread(mddev->thread);
2129         mddev->thread = NULL;
2130         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2131         if (conf->r10bio_pool)
2132                 mempool_destroy(conf->r10bio_pool);
2133         kfree(conf->mirrors);
2134         kfree(conf);
2135         mddev->private = NULL;
2136         return 0;
2137 }
2138
2139 static void raid10_quiesce(mddev_t *mddev, int state)
2140 {
2141         conf_t *conf = mddev_to_conf(mddev);
2142
2143         switch(state) {
2144         case 1:
2145                 raise_barrier(conf, 0);
2146                 break;
2147         case 0:
2148                 lower_barrier(conf);
2149                 break;
2150         }
2151         if (mddev->thread) {
2152                 if (mddev->bitmap)
2153                         mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
2154                 else
2155                         mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
2156                 md_wakeup_thread(mddev->thread);
2157         }
2158 }
2159
2160 static struct mdk_personality raid10_personality =
2161 {
2162         .name           = "raid10",
2163         .level          = 10,
2164         .owner          = THIS_MODULE,
2165         .make_request   = make_request,
2166         .run            = run,
2167         .stop           = stop,
2168         .status         = status,
2169         .error_handler  = error,
2170         .hot_add_disk   = raid10_add_disk,
2171         .hot_remove_disk= raid10_remove_disk,
2172         .spare_active   = raid10_spare_active,
2173         .sync_request   = sync_request,
2174         .quiesce        = raid10_quiesce,
2175 };
2176
2177 static int __init raid_init(void)
2178 {
2179         return register_md_personality(&raid10_personality);
2180 }
2181
2182 static void raid_exit(void)
2183 {
2184         unregister_md_personality(&raid10_personality);
2185 }
2186
2187 module_init(raid_init);
2188 module_exit(raid_exit);
2189 MODULE_LICENSE("GPL");
2190 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2191 MODULE_ALIAS("md-raid10");
2192 MODULE_ALIAS("md-level-10");