pasemi_mac: flags as passed to spin_*_irqsave() should be unsigned long
[linux-2.6] / 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(struct request_queue *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                         struct request_queue *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(struct request_queue *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(struct request_queue *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                         struct request_queue *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 static int raid10_congested(void *data, int bits)
652 {
653         mddev_t *mddev = data;
654         conf_t *conf = mddev_to_conf(mddev);
655         int i, ret = 0;
656
657         rcu_read_lock();
658         for (i = 0; i < mddev->raid_disks && ret == 0; i++) {
659                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
660                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
661                         struct request_queue *q = bdev_get_queue(rdev->bdev);
662
663                         ret |= bdi_congested(&q->backing_dev_info, bits);
664                 }
665         }
666         rcu_read_unlock();
667         return ret;
668 }
669
670
671 /* Barriers....
672  * Sometimes we need to suspend IO while we do something else,
673  * either some resync/recovery, or reconfigure the array.
674  * To do this we raise a 'barrier'.
675  * The 'barrier' is a counter that can be raised multiple times
676  * to count how many activities are happening which preclude
677  * normal IO.
678  * We can only raise the barrier if there is no pending IO.
679  * i.e. if nr_pending == 0.
680  * We choose only to raise the barrier if no-one is waiting for the
681  * barrier to go down.  This means that as soon as an IO request
682  * is ready, no other operations which require a barrier will start
683  * until the IO request has had a chance.
684  *
685  * So: regular IO calls 'wait_barrier'.  When that returns there
686  *    is no backgroup IO happening,  It must arrange to call
687  *    allow_barrier when it has finished its IO.
688  * backgroup IO calls must call raise_barrier.  Once that returns
689  *    there is no normal IO happeing.  It must arrange to call
690  *    lower_barrier when the particular background IO completes.
691  */
692 #define RESYNC_DEPTH 32
693
694 static void raise_barrier(conf_t *conf, int force)
695 {
696         BUG_ON(force && !conf->barrier);
697         spin_lock_irq(&conf->resync_lock);
698
699         /* Wait until no block IO is waiting (unless 'force') */
700         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
701                             conf->resync_lock,
702                             raid10_unplug(conf->mddev->queue));
703
704         /* block any new IO from starting */
705         conf->barrier++;
706
707         /* No wait for all pending IO to complete */
708         wait_event_lock_irq(conf->wait_barrier,
709                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
710                             conf->resync_lock,
711                             raid10_unplug(conf->mddev->queue));
712
713         spin_unlock_irq(&conf->resync_lock);
714 }
715
716 static void lower_barrier(conf_t *conf)
717 {
718         unsigned long flags;
719         spin_lock_irqsave(&conf->resync_lock, flags);
720         conf->barrier--;
721         spin_unlock_irqrestore(&conf->resync_lock, flags);
722         wake_up(&conf->wait_barrier);
723 }
724
725 static void wait_barrier(conf_t *conf)
726 {
727         spin_lock_irq(&conf->resync_lock);
728         if (conf->barrier) {
729                 conf->nr_waiting++;
730                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
731                                     conf->resync_lock,
732                                     raid10_unplug(conf->mddev->queue));
733                 conf->nr_waiting--;
734         }
735         conf->nr_pending++;
736         spin_unlock_irq(&conf->resync_lock);
737 }
738
739 static void allow_barrier(conf_t *conf)
740 {
741         unsigned long flags;
742         spin_lock_irqsave(&conf->resync_lock, flags);
743         conf->nr_pending--;
744         spin_unlock_irqrestore(&conf->resync_lock, flags);
745         wake_up(&conf->wait_barrier);
746 }
747
748 static void freeze_array(conf_t *conf)
749 {
750         /* stop syncio and normal IO and wait for everything to
751          * go quiet.
752          * We increment barrier and nr_waiting, and then
753          * wait until barrier+nr_pending match nr_queued+2
754          */
755         spin_lock_irq(&conf->resync_lock);
756         conf->barrier++;
757         conf->nr_waiting++;
758         wait_event_lock_irq(conf->wait_barrier,
759                             conf->barrier+conf->nr_pending == conf->nr_queued+2,
760                             conf->resync_lock,
761                             raid10_unplug(conf->mddev->queue));
762         spin_unlock_irq(&conf->resync_lock);
763 }
764
765 static void unfreeze_array(conf_t *conf)
766 {
767         /* reverse the effect of the freeze */
768         spin_lock_irq(&conf->resync_lock);
769         conf->barrier--;
770         conf->nr_waiting--;
771         wake_up(&conf->wait_barrier);
772         spin_unlock_irq(&conf->resync_lock);
773 }
774
775 static int make_request(struct request_queue *q, struct bio * bio)
776 {
777         mddev_t *mddev = q->queuedata;
778         conf_t *conf = mddev_to_conf(mddev);
779         mirror_info_t *mirror;
780         r10bio_t *r10_bio;
781         struct bio *read_bio;
782         int i;
783         int chunk_sects = conf->chunk_mask + 1;
784         const int rw = bio_data_dir(bio);
785         const int do_sync = bio_sync(bio);
786         struct bio_list bl;
787         unsigned long flags;
788
789         if (unlikely(bio_barrier(bio))) {
790                 bio_endio(bio, bio->bi_size, -EOPNOTSUPP);
791                 return 0;
792         }
793
794         /* If this request crosses a chunk boundary, we need to
795          * split it.  This will only happen for 1 PAGE (or less) requests.
796          */
797         if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
798                       > chunk_sects &&
799                     conf->near_copies < conf->raid_disks)) {
800                 struct bio_pair *bp;
801                 /* Sanity check -- queue functions should prevent this happening */
802                 if (bio->bi_vcnt != 1 ||
803                     bio->bi_idx != 0)
804                         goto bad_map;
805                 /* This is a one page bio that upper layers
806                  * refuse to split for us, so we need to split it.
807                  */
808                 bp = bio_split(bio, bio_split_pool,
809                                chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
810                 if (make_request(q, &bp->bio1))
811                         generic_make_request(&bp->bio1);
812                 if (make_request(q, &bp->bio2))
813                         generic_make_request(&bp->bio2);
814
815                 bio_pair_release(bp);
816                 return 0;
817         bad_map:
818                 printk("raid10_make_request bug: can't convert block across chunks"
819                        " or bigger than %dk %llu %d\n", chunk_sects/2,
820                        (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
821
822                 bio_io_error(bio, bio->bi_size);
823                 return 0;
824         }
825
826         md_write_start(mddev, bio);
827
828         /*
829          * Register the new request and wait if the reconstruction
830          * thread has put up a bar for new requests.
831          * Continue immediately if no resync is active currently.
832          */
833         wait_barrier(conf);
834
835         disk_stat_inc(mddev->gendisk, ios[rw]);
836         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
837
838         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
839
840         r10_bio->master_bio = bio;
841         r10_bio->sectors = bio->bi_size >> 9;
842
843         r10_bio->mddev = mddev;
844         r10_bio->sector = bio->bi_sector;
845         r10_bio->state = 0;
846
847         if (rw == READ) {
848                 /*
849                  * read balancing logic:
850                  */
851                 int disk = read_balance(conf, r10_bio);
852                 int slot = r10_bio->read_slot;
853                 if (disk < 0) {
854                         raid_end_bio_io(r10_bio);
855                         return 0;
856                 }
857                 mirror = conf->mirrors + disk;
858
859                 read_bio = bio_clone(bio, GFP_NOIO);
860
861                 r10_bio->devs[slot].bio = read_bio;
862
863                 read_bio->bi_sector = r10_bio->devs[slot].addr +
864                         mirror->rdev->data_offset;
865                 read_bio->bi_bdev = mirror->rdev->bdev;
866                 read_bio->bi_end_io = raid10_end_read_request;
867                 read_bio->bi_rw = READ | do_sync;
868                 read_bio->bi_private = r10_bio;
869
870                 generic_make_request(read_bio);
871                 return 0;
872         }
873
874         /*
875          * WRITE:
876          */
877         /* first select target devices under spinlock and
878          * inc refcount on their rdev.  Record them by setting
879          * bios[x] to bio
880          */
881         raid10_find_phys(conf, r10_bio);
882         rcu_read_lock();
883         for (i = 0;  i < conf->copies; i++) {
884                 int d = r10_bio->devs[i].devnum;
885                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
886                 if (rdev &&
887                     !test_bit(Faulty, &rdev->flags)) {
888                         atomic_inc(&rdev->nr_pending);
889                         r10_bio->devs[i].bio = bio;
890                 } else {
891                         r10_bio->devs[i].bio = NULL;
892                         set_bit(R10BIO_Degraded, &r10_bio->state);
893                 }
894         }
895         rcu_read_unlock();
896
897         atomic_set(&r10_bio->remaining, 0);
898
899         bio_list_init(&bl);
900         for (i = 0; i < conf->copies; i++) {
901                 struct bio *mbio;
902                 int d = r10_bio->devs[i].devnum;
903                 if (!r10_bio->devs[i].bio)
904                         continue;
905
906                 mbio = bio_clone(bio, GFP_NOIO);
907                 r10_bio->devs[i].bio = mbio;
908
909                 mbio->bi_sector = r10_bio->devs[i].addr+
910                         conf->mirrors[d].rdev->data_offset;
911                 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
912                 mbio->bi_end_io = raid10_end_write_request;
913                 mbio->bi_rw = WRITE | do_sync;
914                 mbio->bi_private = r10_bio;
915
916                 atomic_inc(&r10_bio->remaining);
917                 bio_list_add(&bl, mbio);
918         }
919
920         if (unlikely(!atomic_read(&r10_bio->remaining))) {
921                 /* the array is dead */
922                 md_write_end(mddev);
923                 raid_end_bio_io(r10_bio);
924                 return 0;
925         }
926
927         bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
928         spin_lock_irqsave(&conf->device_lock, flags);
929         bio_list_merge(&conf->pending_bio_list, &bl);
930         blk_plug_device(mddev->queue);
931         spin_unlock_irqrestore(&conf->device_lock, flags);
932
933         if (do_sync)
934                 md_wakeup_thread(mddev->thread);
935
936         return 0;
937 }
938
939 static void status(struct seq_file *seq, mddev_t *mddev)
940 {
941         conf_t *conf = mddev_to_conf(mddev);
942         int i;
943
944         if (conf->near_copies < conf->raid_disks)
945                 seq_printf(seq, " %dK chunks", mddev->chunk_size/1024);
946         if (conf->near_copies > 1)
947                 seq_printf(seq, " %d near-copies", conf->near_copies);
948         if (conf->far_copies > 1) {
949                 if (conf->far_offset)
950                         seq_printf(seq, " %d offset-copies", conf->far_copies);
951                 else
952                         seq_printf(seq, " %d far-copies", conf->far_copies);
953         }
954         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
955                                         conf->raid_disks - mddev->degraded);
956         for (i = 0; i < conf->raid_disks; i++)
957                 seq_printf(seq, "%s",
958                               conf->mirrors[i].rdev &&
959                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
960         seq_printf(seq, "]");
961 }
962
963 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
964 {
965         char b[BDEVNAME_SIZE];
966         conf_t *conf = mddev_to_conf(mddev);
967
968         /*
969          * If it is not operational, then we have already marked it as dead
970          * else if it is the last working disks, ignore the error, let the
971          * next level up know.
972          * else mark the drive as failed
973          */
974         if (test_bit(In_sync, &rdev->flags)
975             && conf->raid_disks-mddev->degraded == 1)
976                 /*
977                  * Don't fail the drive, just return an IO error.
978                  * The test should really be more sophisticated than
979                  * "working_disks == 1", but it isn't critical, and
980                  * can wait until we do more sophisticated "is the drive
981                  * really dead" tests...
982                  */
983                 return;
984         if (test_and_clear_bit(In_sync, &rdev->flags)) {
985                 unsigned long flags;
986                 spin_lock_irqsave(&conf->device_lock, flags);
987                 mddev->degraded++;
988                 spin_unlock_irqrestore(&conf->device_lock, flags);
989                 /*
990                  * if recovery is running, make sure it aborts.
991                  */
992                 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
993         }
994         set_bit(Faulty, &rdev->flags);
995         set_bit(MD_CHANGE_DEVS, &mddev->flags);
996         printk(KERN_ALERT "raid10: Disk failure on %s, disabling device. \n"
997                 "       Operation continuing on %d devices\n",
998                 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
999 }
1000
1001 static void print_conf(conf_t *conf)
1002 {
1003         int i;
1004         mirror_info_t *tmp;
1005
1006         printk("RAID10 conf printout:\n");
1007         if (!conf) {
1008                 printk("(!conf)\n");
1009                 return;
1010         }
1011         printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1012                 conf->raid_disks);
1013
1014         for (i = 0; i < conf->raid_disks; i++) {
1015                 char b[BDEVNAME_SIZE];
1016                 tmp = conf->mirrors + i;
1017                 if (tmp->rdev)
1018                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
1019                                 i, !test_bit(In_sync, &tmp->rdev->flags),
1020                                 !test_bit(Faulty, &tmp->rdev->flags),
1021                                 bdevname(tmp->rdev->bdev,b));
1022         }
1023 }
1024
1025 static void close_sync(conf_t *conf)
1026 {
1027         wait_barrier(conf);
1028         allow_barrier(conf);
1029
1030         mempool_destroy(conf->r10buf_pool);
1031         conf->r10buf_pool = NULL;
1032 }
1033
1034 /* check if there are enough drives for
1035  * every block to appear on atleast one
1036  */
1037 static int enough(conf_t *conf)
1038 {
1039         int first = 0;
1040
1041         do {
1042                 int n = conf->copies;
1043                 int cnt = 0;
1044                 while (n--) {
1045                         if (conf->mirrors[first].rdev)
1046                                 cnt++;
1047                         first = (first+1) % conf->raid_disks;
1048                 }
1049                 if (cnt == 0)
1050                         return 0;
1051         } while (first != 0);
1052         return 1;
1053 }
1054
1055 static int raid10_spare_active(mddev_t *mddev)
1056 {
1057         int i;
1058         conf_t *conf = mddev->private;
1059         mirror_info_t *tmp;
1060
1061         /*
1062          * Find all non-in_sync disks within the RAID10 configuration
1063          * and mark them in_sync
1064          */
1065         for (i = 0; i < conf->raid_disks; i++) {
1066                 tmp = conf->mirrors + i;
1067                 if (tmp->rdev
1068                     && !test_bit(Faulty, &tmp->rdev->flags)
1069                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1070                         unsigned long flags;
1071                         spin_lock_irqsave(&conf->device_lock, flags);
1072                         mddev->degraded--;
1073                         spin_unlock_irqrestore(&conf->device_lock, flags);
1074                 }
1075         }
1076
1077         print_conf(conf);
1078         return 0;
1079 }
1080
1081
1082 static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1083 {
1084         conf_t *conf = mddev->private;
1085         int found = 0;
1086         int mirror;
1087         mirror_info_t *p;
1088
1089         if (mddev->recovery_cp < MaxSector)
1090                 /* only hot-add to in-sync arrays, as recovery is
1091                  * very different from resync
1092                  */
1093                 return 0;
1094         if (!enough(conf))
1095                 return 0;
1096
1097         if (rdev->saved_raid_disk >= 0 &&
1098             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1099                 mirror = rdev->saved_raid_disk;
1100         else
1101                 mirror = 0;
1102         for ( ; mirror < mddev->raid_disks; mirror++)
1103                 if ( !(p=conf->mirrors+mirror)->rdev) {
1104
1105                         blk_queue_stack_limits(mddev->queue,
1106                                                rdev->bdev->bd_disk->queue);
1107                         /* as we don't honour merge_bvec_fn, we must never risk
1108                          * violating it, so limit ->max_sector to one PAGE, as
1109                          * a one page request is never in violation.
1110                          */
1111                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1112                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
1113                                 mddev->queue->max_sectors = (PAGE_SIZE>>9);
1114
1115                         p->head_position = 0;
1116                         rdev->raid_disk = mirror;
1117                         found = 1;
1118                         if (rdev->saved_raid_disk != mirror)
1119                                 conf->fullsync = 1;
1120                         rcu_assign_pointer(p->rdev, rdev);
1121                         break;
1122                 }
1123
1124         print_conf(conf);
1125         return found;
1126 }
1127
1128 static int raid10_remove_disk(mddev_t *mddev, int number)
1129 {
1130         conf_t *conf = mddev->private;
1131         int err = 0;
1132         mdk_rdev_t *rdev;
1133         mirror_info_t *p = conf->mirrors+ number;
1134
1135         print_conf(conf);
1136         rdev = p->rdev;
1137         if (rdev) {
1138                 if (test_bit(In_sync, &rdev->flags) ||
1139                     atomic_read(&rdev->nr_pending)) {
1140                         err = -EBUSY;
1141                         goto abort;
1142                 }
1143                 p->rdev = NULL;
1144                 synchronize_rcu();
1145                 if (atomic_read(&rdev->nr_pending)) {
1146                         /* lost the race, try later */
1147                         err = -EBUSY;
1148                         p->rdev = rdev;
1149                 }
1150         }
1151 abort:
1152
1153         print_conf(conf);
1154         return err;
1155 }
1156
1157
1158 static int end_sync_read(struct bio *bio, unsigned int bytes_done, int error)
1159 {
1160         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1161         conf_t *conf = mddev_to_conf(r10_bio->mddev);
1162         int i,d;
1163
1164         if (bio->bi_size)
1165                 return 1;
1166
1167         for (i=0; i<conf->copies; i++)
1168                 if (r10_bio->devs[i].bio == bio)
1169                         break;
1170         BUG_ON(i == conf->copies);
1171         update_head_pos(i, r10_bio);
1172         d = r10_bio->devs[i].devnum;
1173
1174         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1175                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1176         else {
1177                 atomic_add(r10_bio->sectors,
1178                            &conf->mirrors[d].rdev->corrected_errors);
1179                 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1180                         md_error(r10_bio->mddev,
1181                                  conf->mirrors[d].rdev);
1182         }
1183
1184         /* for reconstruct, we always reschedule after a read.
1185          * for resync, only after all reads
1186          */
1187         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1188             atomic_dec_and_test(&r10_bio->remaining)) {
1189                 /* we have read all the blocks,
1190                  * do the comparison in process context in raid10d
1191                  */
1192                 reschedule_retry(r10_bio);
1193         }
1194         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1195         return 0;
1196 }
1197
1198 static int end_sync_write(struct bio *bio, unsigned int bytes_done, int error)
1199 {
1200         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1201         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1202         mddev_t *mddev = r10_bio->mddev;
1203         conf_t *conf = mddev_to_conf(mddev);
1204         int i,d;
1205
1206         if (bio->bi_size)
1207                 return 1;
1208
1209         for (i = 0; i < conf->copies; i++)
1210                 if (r10_bio->devs[i].bio == bio)
1211                         break;
1212         d = r10_bio->devs[i].devnum;
1213
1214         if (!uptodate)
1215                 md_error(mddev, conf->mirrors[d].rdev);
1216         update_head_pos(i, r10_bio);
1217
1218         while (atomic_dec_and_test(&r10_bio->remaining)) {
1219                 if (r10_bio->master_bio == NULL) {
1220                         /* the primary of several recovery bios */
1221                         md_done_sync(mddev, r10_bio->sectors, 1);
1222                         put_buf(r10_bio);
1223                         break;
1224                 } else {
1225                         r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1226                         put_buf(r10_bio);
1227                         r10_bio = r10_bio2;
1228                 }
1229         }
1230         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1231         return 0;
1232 }
1233
1234 /*
1235  * Note: sync and recover and handled very differently for raid10
1236  * This code is for resync.
1237  * For resync, we read through virtual addresses and read all blocks.
1238  * If there is any error, we schedule a write.  The lowest numbered
1239  * drive is authoritative.
1240  * However requests come for physical address, so we need to map.
1241  * For every physical address there are raid_disks/copies virtual addresses,
1242  * which is always are least one, but is not necessarly an integer.
1243  * This means that a physical address can span multiple chunks, so we may
1244  * have to submit multiple io requests for a single sync request.
1245  */
1246 /*
1247  * We check if all blocks are in-sync and only write to blocks that
1248  * aren't in sync
1249  */
1250 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1251 {
1252         conf_t *conf = mddev_to_conf(mddev);
1253         int i, first;
1254         struct bio *tbio, *fbio;
1255
1256         atomic_set(&r10_bio->remaining, 1);
1257
1258         /* find the first device with a block */
1259         for (i=0; i<conf->copies; i++)
1260                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1261                         break;
1262
1263         if (i == conf->copies)
1264                 goto done;
1265
1266         first = i;
1267         fbio = r10_bio->devs[i].bio;
1268
1269         /* now find blocks with errors */
1270         for (i=0 ; i < conf->copies ; i++) {
1271                 int  j, d;
1272                 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1273
1274                 tbio = r10_bio->devs[i].bio;
1275
1276                 if (tbio->bi_end_io != end_sync_read)
1277                         continue;
1278                 if (i == first)
1279                         continue;
1280                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1281                         /* We know that the bi_io_vec layout is the same for
1282                          * both 'first' and 'i', so we just compare them.
1283                          * All vec entries are PAGE_SIZE;
1284                          */
1285                         for (j = 0; j < vcnt; j++)
1286                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1287                                            page_address(tbio->bi_io_vec[j].bv_page),
1288                                            PAGE_SIZE))
1289                                         break;
1290                         if (j == vcnt)
1291                                 continue;
1292                         mddev->resync_mismatches += r10_bio->sectors;
1293                 }
1294                 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1295                         /* Don't fix anything. */
1296                         continue;
1297                 /* Ok, we need to write this bio
1298                  * First we need to fixup bv_offset, bv_len and
1299                  * bi_vecs, as the read request might have corrupted these
1300                  */
1301                 tbio->bi_vcnt = vcnt;
1302                 tbio->bi_size = r10_bio->sectors << 9;
1303                 tbio->bi_idx = 0;
1304                 tbio->bi_phys_segments = 0;
1305                 tbio->bi_hw_segments = 0;
1306                 tbio->bi_hw_front_size = 0;
1307                 tbio->bi_hw_back_size = 0;
1308                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1309                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1310                 tbio->bi_next = NULL;
1311                 tbio->bi_rw = WRITE;
1312                 tbio->bi_private = r10_bio;
1313                 tbio->bi_sector = r10_bio->devs[i].addr;
1314
1315                 for (j=0; j < vcnt ; j++) {
1316                         tbio->bi_io_vec[j].bv_offset = 0;
1317                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1318
1319                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1320                                page_address(fbio->bi_io_vec[j].bv_page),
1321                                PAGE_SIZE);
1322                 }
1323                 tbio->bi_end_io = end_sync_write;
1324
1325                 d = r10_bio->devs[i].devnum;
1326                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1327                 atomic_inc(&r10_bio->remaining);
1328                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1329
1330                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1331                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1332                 generic_make_request(tbio);
1333         }
1334
1335 done:
1336         if (atomic_dec_and_test(&r10_bio->remaining)) {
1337                 md_done_sync(mddev, r10_bio->sectors, 1);
1338                 put_buf(r10_bio);
1339         }
1340 }
1341
1342 /*
1343  * Now for the recovery code.
1344  * Recovery happens across physical sectors.
1345  * We recover all non-is_sync drives by finding the virtual address of
1346  * each, and then choose a working drive that also has that virt address.
1347  * There is a separate r10_bio for each non-in_sync drive.
1348  * Only the first two slots are in use. The first for reading,
1349  * The second for writing.
1350  *
1351  */
1352
1353 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1354 {
1355         conf_t *conf = mddev_to_conf(mddev);
1356         int i, d;
1357         struct bio *bio, *wbio;
1358
1359
1360         /* move the pages across to the second bio
1361          * and submit the write request
1362          */
1363         bio = r10_bio->devs[0].bio;
1364         wbio = r10_bio->devs[1].bio;
1365         for (i=0; i < wbio->bi_vcnt; i++) {
1366                 struct page *p = bio->bi_io_vec[i].bv_page;
1367                 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1368                 wbio->bi_io_vec[i].bv_page = p;
1369         }
1370         d = r10_bio->devs[1].devnum;
1371
1372         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1373         md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1374         if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1375                 generic_make_request(wbio);
1376         else
1377                 bio_endio(wbio, wbio->bi_size, -EIO);
1378 }
1379
1380
1381 /*
1382  * This is a kernel thread which:
1383  *
1384  *      1.      Retries failed read operations on working mirrors.
1385  *      2.      Updates the raid superblock when problems encounter.
1386  *      3.      Performs writes following reads for array synchronising.
1387  */
1388
1389 static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1390 {
1391         int sect = 0; /* Offset from r10_bio->sector */
1392         int sectors = r10_bio->sectors;
1393         mdk_rdev_t*rdev;
1394         while(sectors) {
1395                 int s = sectors;
1396                 int sl = r10_bio->read_slot;
1397                 int success = 0;
1398                 int start;
1399
1400                 if (s > (PAGE_SIZE>>9))
1401                         s = PAGE_SIZE >> 9;
1402
1403                 rcu_read_lock();
1404                 do {
1405                         int d = r10_bio->devs[sl].devnum;
1406                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1407                         if (rdev &&
1408                             test_bit(In_sync, &rdev->flags)) {
1409                                 atomic_inc(&rdev->nr_pending);
1410                                 rcu_read_unlock();
1411                                 success = sync_page_io(rdev->bdev,
1412                                                        r10_bio->devs[sl].addr +
1413                                                        sect + rdev->data_offset,
1414                                                        s<<9,
1415                                                        conf->tmppage, READ);
1416                                 rdev_dec_pending(rdev, mddev);
1417                                 rcu_read_lock();
1418                                 if (success)
1419                                         break;
1420                         }
1421                         sl++;
1422                         if (sl == conf->copies)
1423                                 sl = 0;
1424                 } while (!success && sl != r10_bio->read_slot);
1425                 rcu_read_unlock();
1426
1427                 if (!success) {
1428                         /* Cannot read from anywhere -- bye bye array */
1429                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1430                         md_error(mddev, conf->mirrors[dn].rdev);
1431                         break;
1432                 }
1433
1434                 start = sl;
1435                 /* write it back and re-read */
1436                 rcu_read_lock();
1437                 while (sl != r10_bio->read_slot) {
1438                         int d;
1439                         if (sl==0)
1440                                 sl = conf->copies;
1441                         sl--;
1442                         d = r10_bio->devs[sl].devnum;
1443                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1444                         if (rdev &&
1445                             test_bit(In_sync, &rdev->flags)) {
1446                                 atomic_inc(&rdev->nr_pending);
1447                                 rcu_read_unlock();
1448                                 atomic_add(s, &rdev->corrected_errors);
1449                                 if (sync_page_io(rdev->bdev,
1450                                                  r10_bio->devs[sl].addr +
1451                                                  sect + rdev->data_offset,
1452                                                  s<<9, conf->tmppage, WRITE)
1453                                     == 0)
1454                                         /* Well, this device is dead */
1455                                         md_error(mddev, rdev);
1456                                 rdev_dec_pending(rdev, mddev);
1457                                 rcu_read_lock();
1458                         }
1459                 }
1460                 sl = start;
1461                 while (sl != r10_bio->read_slot) {
1462                         int d;
1463                         if (sl==0)
1464                                 sl = conf->copies;
1465                         sl--;
1466                         d = r10_bio->devs[sl].devnum;
1467                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1468                         if (rdev &&
1469                             test_bit(In_sync, &rdev->flags)) {
1470                                 char b[BDEVNAME_SIZE];
1471                                 atomic_inc(&rdev->nr_pending);
1472                                 rcu_read_unlock();
1473                                 if (sync_page_io(rdev->bdev,
1474                                                  r10_bio->devs[sl].addr +
1475                                                  sect + rdev->data_offset,
1476                                                  s<<9, conf->tmppage, READ) == 0)
1477                                         /* Well, this device is dead */
1478                                         md_error(mddev, rdev);
1479                                 else
1480                                         printk(KERN_INFO
1481                                                "raid10:%s: read error corrected"
1482                                                " (%d sectors at %llu on %s)\n",
1483                                                mdname(mddev), s,
1484                                                (unsigned long long)(sect+
1485                                                     rdev->data_offset),
1486                                                bdevname(rdev->bdev, b));
1487
1488                                 rdev_dec_pending(rdev, mddev);
1489                                 rcu_read_lock();
1490                         }
1491                 }
1492                 rcu_read_unlock();
1493
1494                 sectors -= s;
1495                 sect += s;
1496         }
1497 }
1498
1499 static void raid10d(mddev_t *mddev)
1500 {
1501         r10bio_t *r10_bio;
1502         struct bio *bio;
1503         unsigned long flags;
1504         conf_t *conf = mddev_to_conf(mddev);
1505         struct list_head *head = &conf->retry_list;
1506         int unplug=0;
1507         mdk_rdev_t *rdev;
1508
1509         md_check_recovery(mddev);
1510
1511         for (;;) {
1512                 char b[BDEVNAME_SIZE];
1513                 spin_lock_irqsave(&conf->device_lock, flags);
1514
1515                 if (conf->pending_bio_list.head) {
1516                         bio = bio_list_get(&conf->pending_bio_list);
1517                         blk_remove_plug(mddev->queue);
1518                         spin_unlock_irqrestore(&conf->device_lock, flags);
1519                         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
1520                         bitmap_unplug(mddev->bitmap);
1521
1522                         while (bio) { /* submit pending writes */
1523                                 struct bio *next = bio->bi_next;
1524                                 bio->bi_next = NULL;
1525                                 generic_make_request(bio);
1526                                 bio = next;
1527                         }
1528                         unplug = 1;
1529
1530                         continue;
1531                 }
1532
1533                 if (list_empty(head))
1534                         break;
1535                 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1536                 list_del(head->prev);
1537                 conf->nr_queued--;
1538                 spin_unlock_irqrestore(&conf->device_lock, flags);
1539
1540                 mddev = r10_bio->mddev;
1541                 conf = mddev_to_conf(mddev);
1542                 if (test_bit(R10BIO_IsSync, &r10_bio->state)) {
1543                         sync_request_write(mddev, r10_bio);
1544                         unplug = 1;
1545                 } else  if (test_bit(R10BIO_IsRecover, &r10_bio->state)) {
1546                         recovery_request_write(mddev, r10_bio);
1547                         unplug = 1;
1548                 } else {
1549                         int mirror;
1550                         /* we got a read error. Maybe the drive is bad.  Maybe just
1551                          * the block and we can fix it.
1552                          * We freeze all other IO, and try reading the block from
1553                          * other devices.  When we find one, we re-write
1554                          * and check it that fixes the read error.
1555                          * This is all done synchronously while the array is
1556                          * frozen.
1557                          */
1558                         if (mddev->ro == 0) {
1559                                 freeze_array(conf);
1560                                 fix_read_error(conf, mddev, r10_bio);
1561                                 unfreeze_array(conf);
1562                         }
1563
1564                         bio = r10_bio->devs[r10_bio->read_slot].bio;
1565                         r10_bio->devs[r10_bio->read_slot].bio =
1566                                 mddev->ro ? IO_BLOCKED : NULL;
1567                         mirror = read_balance(conf, r10_bio);
1568                         if (mirror == -1) {
1569                                 printk(KERN_ALERT "raid10: %s: unrecoverable I/O"
1570                                        " read error for block %llu\n",
1571                                        bdevname(bio->bi_bdev,b),
1572                                        (unsigned long long)r10_bio->sector);
1573                                 raid_end_bio_io(r10_bio);
1574                                 bio_put(bio);
1575                         } else {
1576                                 const int do_sync = bio_sync(r10_bio->master_bio);
1577                                 bio_put(bio);
1578                                 rdev = conf->mirrors[mirror].rdev;
1579                                 if (printk_ratelimit())
1580                                         printk(KERN_ERR "raid10: %s: redirecting sector %llu to"
1581                                                " another mirror\n",
1582                                                bdevname(rdev->bdev,b),
1583                                                (unsigned long long)r10_bio->sector);
1584                                 bio = bio_clone(r10_bio->master_bio, GFP_NOIO);
1585                                 r10_bio->devs[r10_bio->read_slot].bio = bio;
1586                                 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1587                                         + rdev->data_offset;
1588                                 bio->bi_bdev = rdev->bdev;
1589                                 bio->bi_rw = READ | do_sync;
1590                                 bio->bi_private = r10_bio;
1591                                 bio->bi_end_io = raid10_end_read_request;
1592                                 unplug = 1;
1593                                 generic_make_request(bio);
1594                         }
1595                 }
1596         }
1597         spin_unlock_irqrestore(&conf->device_lock, flags);
1598         if (unplug)
1599                 unplug_slaves(mddev);
1600 }
1601
1602
1603 static int init_resync(conf_t *conf)
1604 {
1605         int buffs;
1606
1607         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1608         BUG_ON(conf->r10buf_pool);
1609         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1610         if (!conf->r10buf_pool)
1611                 return -ENOMEM;
1612         conf->next_resync = 0;
1613         return 0;
1614 }
1615
1616 /*
1617  * perform a "sync" on one "block"
1618  *
1619  * We need to make sure that no normal I/O request - particularly write
1620  * requests - conflict with active sync requests.
1621  *
1622  * This is achieved by tracking pending requests and a 'barrier' concept
1623  * that can be installed to exclude normal IO requests.
1624  *
1625  * Resync and recovery are handled very differently.
1626  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1627  *
1628  * For resync, we iterate over virtual addresses, read all copies,
1629  * and update if there are differences.  If only one copy is live,
1630  * skip it.
1631  * For recovery, we iterate over physical addresses, read a good
1632  * value for each non-in_sync drive, and over-write.
1633  *
1634  * So, for recovery we may have several outstanding complex requests for a
1635  * given address, one for each out-of-sync device.  We model this by allocating
1636  * a number of r10_bio structures, one for each out-of-sync device.
1637  * As we setup these structures, we collect all bio's together into a list
1638  * which we then process collectively to add pages, and then process again
1639  * to pass to generic_make_request.
1640  *
1641  * The r10_bio structures are linked using a borrowed master_bio pointer.
1642  * This link is counted in ->remaining.  When the r10_bio that points to NULL
1643  * has its remaining count decremented to 0, the whole complex operation
1644  * is complete.
1645  *
1646  */
1647
1648 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1649 {
1650         conf_t *conf = mddev_to_conf(mddev);
1651         r10bio_t *r10_bio;
1652         struct bio *biolist = NULL, *bio;
1653         sector_t max_sector, nr_sectors;
1654         int disk;
1655         int i;
1656         int max_sync;
1657         int sync_blocks;
1658
1659         sector_t sectors_skipped = 0;
1660         int chunks_skipped = 0;
1661
1662         if (!conf->r10buf_pool)
1663                 if (init_resync(conf))
1664                         return 0;
1665
1666  skipped:
1667         max_sector = mddev->size << 1;
1668         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1669                 max_sector = mddev->resync_max_sectors;
1670         if (sector_nr >= max_sector) {
1671                 /* If we aborted, we need to abort the
1672                  * sync on the 'current' bitmap chucks (there can
1673                  * be several when recovering multiple devices).
1674                  * as we may have started syncing it but not finished.
1675                  * We can find the current address in
1676                  * mddev->curr_resync, but for recovery,
1677                  * we need to convert that to several
1678                  * virtual addresses.
1679                  */
1680                 if (mddev->curr_resync < max_sector) { /* aborted */
1681                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1682                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1683                                                 &sync_blocks, 1);
1684                         else for (i=0; i<conf->raid_disks; i++) {
1685                                 sector_t sect =
1686                                         raid10_find_virt(conf, mddev->curr_resync, i);
1687                                 bitmap_end_sync(mddev->bitmap, sect,
1688                                                 &sync_blocks, 1);
1689                         }
1690                 } else /* completed sync */
1691                         conf->fullsync = 0;
1692
1693                 bitmap_close_sync(mddev->bitmap);
1694                 close_sync(conf);
1695                 *skipped = 1;
1696                 return sectors_skipped;
1697         }
1698         if (chunks_skipped >= conf->raid_disks) {
1699                 /* if there has been nothing to do on any drive,
1700                  * then there is nothing to do at all..
1701                  */
1702                 *skipped = 1;
1703                 return (max_sector - sector_nr) + sectors_skipped;
1704         }
1705
1706         /* make sure whole request will fit in a chunk - if chunks
1707          * are meaningful
1708          */
1709         if (conf->near_copies < conf->raid_disks &&
1710             max_sector > (sector_nr | conf->chunk_mask))
1711                 max_sector = (sector_nr | conf->chunk_mask) + 1;
1712         /*
1713          * If there is non-resync activity waiting for us then
1714          * put in a delay to throttle resync.
1715          */
1716         if (!go_faster && conf->nr_waiting)
1717                 msleep_interruptible(1000);
1718
1719         /* Again, very different code for resync and recovery.
1720          * Both must result in an r10bio with a list of bios that
1721          * have bi_end_io, bi_sector, bi_bdev set,
1722          * and bi_private set to the r10bio.
1723          * For recovery, we may actually create several r10bios
1724          * with 2 bios in each, that correspond to the bios in the main one.
1725          * In this case, the subordinate r10bios link back through a
1726          * borrowed master_bio pointer, and the counter in the master
1727          * includes a ref from each subordinate.
1728          */
1729         /* First, we decide what to do and set ->bi_end_io
1730          * To end_sync_read if we want to read, and
1731          * end_sync_write if we will want to write.
1732          */
1733
1734         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
1735         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1736                 /* recovery... the complicated one */
1737                 int i, j, k;
1738                 r10_bio = NULL;
1739
1740                 for (i=0 ; i<conf->raid_disks; i++)
1741                         if (conf->mirrors[i].rdev &&
1742                             !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) {
1743                                 int still_degraded = 0;
1744                                 /* want to reconstruct this device */
1745                                 r10bio_t *rb2 = r10_bio;
1746                                 sector_t sect = raid10_find_virt(conf, sector_nr, i);
1747                                 int must_sync;
1748                                 /* Unless we are doing a full sync, we only need
1749                                  * to recover the block if it is set in the bitmap
1750                                  */
1751                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1752                                                               &sync_blocks, 1);
1753                                 if (sync_blocks < max_sync)
1754                                         max_sync = sync_blocks;
1755                                 if (!must_sync &&
1756                                     !conf->fullsync) {
1757                                         /* yep, skip the sync_blocks here, but don't assume
1758                                          * that there will never be anything to do here
1759                                          */
1760                                         chunks_skipped = -1;
1761                                         continue;
1762                                 }
1763
1764                                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1765                                 raise_barrier(conf, rb2 != NULL);
1766                                 atomic_set(&r10_bio->remaining, 0);
1767
1768                                 r10_bio->master_bio = (struct bio*)rb2;
1769                                 if (rb2)
1770                                         atomic_inc(&rb2->remaining);
1771                                 r10_bio->mddev = mddev;
1772                                 set_bit(R10BIO_IsRecover, &r10_bio->state);
1773                                 r10_bio->sector = sect;
1774
1775                                 raid10_find_phys(conf, r10_bio);
1776                                 /* Need to check if this section will still be
1777                                  * degraded
1778                                  */
1779                                 for (j=0; j<conf->copies;j++) {
1780                                         int d = r10_bio->devs[j].devnum;
1781                                         if (conf->mirrors[d].rdev == NULL ||
1782                                             test_bit(Faulty, &conf->mirrors[d].rdev->flags)) {
1783                                                 still_degraded = 1;
1784                                                 break;
1785                                         }
1786                                 }
1787                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1788                                                               &sync_blocks, still_degraded);
1789
1790                                 for (j=0; j<conf->copies;j++) {
1791                                         int d = r10_bio->devs[j].devnum;
1792                                         if (conf->mirrors[d].rdev &&
1793                                             test_bit(In_sync, &conf->mirrors[d].rdev->flags)) {
1794                                                 /* This is where we read from */
1795                                                 bio = r10_bio->devs[0].bio;
1796                                                 bio->bi_next = biolist;
1797                                                 biolist = bio;
1798                                                 bio->bi_private = r10_bio;
1799                                                 bio->bi_end_io = end_sync_read;
1800                                                 bio->bi_rw = READ;
1801                                                 bio->bi_sector = r10_bio->devs[j].addr +
1802                                                         conf->mirrors[d].rdev->data_offset;
1803                                                 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1804                                                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1805                                                 atomic_inc(&r10_bio->remaining);
1806                                                 /* and we write to 'i' */
1807
1808                                                 for (k=0; k<conf->copies; k++)
1809                                                         if (r10_bio->devs[k].devnum == i)
1810                                                                 break;
1811                                                 BUG_ON(k == conf->copies);
1812                                                 bio = r10_bio->devs[1].bio;
1813                                                 bio->bi_next = biolist;
1814                                                 biolist = bio;
1815                                                 bio->bi_private = r10_bio;
1816                                                 bio->bi_end_io = end_sync_write;
1817                                                 bio->bi_rw = WRITE;
1818                                                 bio->bi_sector = r10_bio->devs[k].addr +
1819                                                         conf->mirrors[i].rdev->data_offset;
1820                                                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1821
1822                                                 r10_bio->devs[0].devnum = d;
1823                                                 r10_bio->devs[1].devnum = i;
1824
1825                                                 break;
1826                                         }
1827                                 }
1828                                 if (j == conf->copies) {
1829                                         /* Cannot recover, so abort the recovery */
1830                                         put_buf(r10_bio);
1831                                         r10_bio = rb2;
1832                                         if (!test_and_set_bit(MD_RECOVERY_ERR, &mddev->recovery))
1833                                                 printk(KERN_INFO "raid10: %s: insufficient working devices for recovery.\n",
1834                                                        mdname(mddev));
1835                                         break;
1836                                 }
1837                         }
1838                 if (biolist == NULL) {
1839                         while (r10_bio) {
1840                                 r10bio_t *rb2 = r10_bio;
1841                                 r10_bio = (r10bio_t*) rb2->master_bio;
1842                                 rb2->master_bio = NULL;
1843                                 put_buf(rb2);
1844                         }
1845                         goto giveup;
1846                 }
1847         } else {
1848                 /* resync. Schedule a read for every block at this virt offset */
1849                 int count = 0;
1850
1851                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1852                                        &sync_blocks, mddev->degraded) &&
1853                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1854                         /* We can skip this block */
1855                         *skipped = 1;
1856                         return sync_blocks + sectors_skipped;
1857                 }
1858                 if (sync_blocks < max_sync)
1859                         max_sync = sync_blocks;
1860                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1861
1862                 r10_bio->mddev = mddev;
1863                 atomic_set(&r10_bio->remaining, 0);
1864                 raise_barrier(conf, 0);
1865                 conf->next_resync = sector_nr;
1866
1867                 r10_bio->master_bio = NULL;
1868                 r10_bio->sector = sector_nr;
1869                 set_bit(R10BIO_IsSync, &r10_bio->state);
1870                 raid10_find_phys(conf, r10_bio);
1871                 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1872
1873                 for (i=0; i<conf->copies; i++) {
1874                         int d = r10_bio->devs[i].devnum;
1875                         bio = r10_bio->devs[i].bio;
1876                         bio->bi_end_io = NULL;
1877                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
1878                         if (conf->mirrors[d].rdev == NULL ||
1879                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
1880                                 continue;
1881                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1882                         atomic_inc(&r10_bio->remaining);
1883                         bio->bi_next = biolist;
1884                         biolist = bio;
1885                         bio->bi_private = r10_bio;
1886                         bio->bi_end_io = end_sync_read;
1887                         bio->bi_rw = READ;
1888                         bio->bi_sector = r10_bio->devs[i].addr +
1889                                 conf->mirrors[d].rdev->data_offset;
1890                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1891                         count++;
1892                 }
1893
1894                 if (count < 2) {
1895                         for (i=0; i<conf->copies; i++) {
1896                                 int d = r10_bio->devs[i].devnum;
1897                                 if (r10_bio->devs[i].bio->bi_end_io)
1898                                         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1899                         }
1900                         put_buf(r10_bio);
1901                         biolist = NULL;
1902                         goto giveup;
1903                 }
1904         }
1905
1906         for (bio = biolist; bio ; bio=bio->bi_next) {
1907
1908                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1909                 if (bio->bi_end_io)
1910                         bio->bi_flags |= 1 << BIO_UPTODATE;
1911                 bio->bi_vcnt = 0;
1912                 bio->bi_idx = 0;
1913                 bio->bi_phys_segments = 0;
1914                 bio->bi_hw_segments = 0;
1915                 bio->bi_size = 0;
1916         }
1917
1918         nr_sectors = 0;
1919         if (sector_nr + max_sync < max_sector)
1920                 max_sector = sector_nr + max_sync;
1921         do {
1922                 struct page *page;
1923                 int len = PAGE_SIZE;
1924                 disk = 0;
1925                 if (sector_nr + (len>>9) > max_sector)
1926                         len = (max_sector - sector_nr) << 9;
1927                 if (len == 0)
1928                         break;
1929                 for (bio= biolist ; bio ; bio=bio->bi_next) {
1930                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1931                         if (bio_add_page(bio, page, len, 0) == 0) {
1932                                 /* stop here */
1933                                 struct bio *bio2;
1934                                 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1935                                 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
1936                                         /* remove last page from this bio */
1937                                         bio2->bi_vcnt--;
1938                                         bio2->bi_size -= len;
1939                                         bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
1940                                 }
1941                                 goto bio_full;
1942                         }
1943                         disk = i;
1944                 }
1945                 nr_sectors += len>>9;
1946                 sector_nr += len>>9;
1947         } while (biolist->bi_vcnt < RESYNC_PAGES);
1948  bio_full:
1949         r10_bio->sectors = nr_sectors;
1950
1951         while (biolist) {
1952                 bio = biolist;
1953                 biolist = biolist->bi_next;
1954
1955                 bio->bi_next = NULL;
1956                 r10_bio = bio->bi_private;
1957                 r10_bio->sectors = nr_sectors;
1958
1959                 if (bio->bi_end_io == end_sync_read) {
1960                         md_sync_acct(bio->bi_bdev, nr_sectors);
1961                         generic_make_request(bio);
1962                 }
1963         }
1964
1965         if (sectors_skipped)
1966                 /* pretend they weren't skipped, it makes
1967                  * no important difference in this case
1968                  */
1969                 md_done_sync(mddev, sectors_skipped, 1);
1970
1971         return sectors_skipped + nr_sectors;
1972  giveup:
1973         /* There is nowhere to write, so all non-sync
1974          * drives must be failed, so try the next chunk...
1975          */
1976         {
1977         sector_t sec = max_sector - sector_nr;
1978         sectors_skipped += sec;
1979         chunks_skipped ++;
1980         sector_nr = max_sector;
1981         goto skipped;
1982         }
1983 }
1984
1985 static int run(mddev_t *mddev)
1986 {
1987         conf_t *conf;
1988         int i, disk_idx;
1989         mirror_info_t *disk;
1990         mdk_rdev_t *rdev;
1991         struct list_head *tmp;
1992         int nc, fc, fo;
1993         sector_t stride, size;
1994
1995         if (mddev->chunk_size == 0) {
1996                 printk(KERN_ERR "md/raid10: non-zero chunk size required.\n");
1997                 return -EINVAL;
1998         }
1999
2000         nc = mddev->layout & 255;
2001         fc = (mddev->layout >> 8) & 255;
2002         fo = mddev->layout & (1<<16);
2003         if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
2004             (mddev->layout >> 17)) {
2005                 printk(KERN_ERR "raid10: %s: unsupported raid10 layout: 0x%8x\n",
2006                        mdname(mddev), mddev->layout);
2007                 goto out;
2008         }
2009         /*
2010          * copy the already verified devices into our private RAID10
2011          * bookkeeping area. [whatever we allocate in run(),
2012          * should be freed in stop()]
2013          */
2014         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2015         mddev->private = conf;
2016         if (!conf) {
2017                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2018                         mdname(mddev));
2019                 goto out;
2020         }
2021         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2022                                  GFP_KERNEL);
2023         if (!conf->mirrors) {
2024                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2025                        mdname(mddev));
2026                 goto out_free_conf;
2027         }
2028
2029         conf->tmppage = alloc_page(GFP_KERNEL);
2030         if (!conf->tmppage)
2031                 goto out_free_conf;
2032
2033         conf->mddev = mddev;
2034         conf->raid_disks = mddev->raid_disks;
2035         conf->near_copies = nc;
2036         conf->far_copies = fc;
2037         conf->copies = nc*fc;
2038         conf->far_offset = fo;
2039         conf->chunk_mask = (sector_t)(mddev->chunk_size>>9)-1;
2040         conf->chunk_shift = ffz(~mddev->chunk_size) - 9;
2041         size = mddev->size >> (conf->chunk_shift-1);
2042         sector_div(size, fc);
2043         size = size * conf->raid_disks;
2044         sector_div(size, nc);
2045         /* 'size' is now the number of chunks in the array */
2046         /* calculate "used chunks per device" in 'stride' */
2047         stride = size * conf->copies;
2048
2049         /* We need to round up when dividing by raid_disks to
2050          * get the stride size.
2051          */
2052         stride += conf->raid_disks - 1;
2053         sector_div(stride, conf->raid_disks);
2054         mddev->size = stride  << (conf->chunk_shift-1);
2055
2056         if (fo)
2057                 stride = 1;
2058         else
2059                 sector_div(stride, fc);
2060         conf->stride = stride << conf->chunk_shift;
2061
2062         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2063                                                 r10bio_pool_free, conf);
2064         if (!conf->r10bio_pool) {
2065                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2066                         mdname(mddev));
2067                 goto out_free_conf;
2068         }
2069
2070         ITERATE_RDEV(mddev, rdev, tmp) {
2071                 disk_idx = rdev->raid_disk;
2072                 if (disk_idx >= mddev->raid_disks
2073                     || disk_idx < 0)
2074                         continue;
2075                 disk = conf->mirrors + disk_idx;
2076
2077                 disk->rdev = rdev;
2078
2079                 blk_queue_stack_limits(mddev->queue,
2080                                        rdev->bdev->bd_disk->queue);
2081                 /* as we don't honour merge_bvec_fn, we must never risk
2082                  * violating it, so limit ->max_sector to one PAGE, as
2083                  * a one page request is never in violation.
2084                  */
2085                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
2086                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
2087                         mddev->queue->max_sectors = (PAGE_SIZE>>9);
2088
2089                 disk->head_position = 0;
2090         }
2091         spin_lock_init(&conf->device_lock);
2092         INIT_LIST_HEAD(&conf->retry_list);
2093
2094         spin_lock_init(&conf->resync_lock);
2095         init_waitqueue_head(&conf->wait_barrier);
2096
2097         /* need to check that every block has at least one working mirror */
2098         if (!enough(conf)) {
2099                 printk(KERN_ERR "raid10: not enough operational mirrors for %s\n",
2100                        mdname(mddev));
2101                 goto out_free_conf;
2102         }
2103
2104         mddev->degraded = 0;
2105         for (i = 0; i < conf->raid_disks; i++) {
2106
2107                 disk = conf->mirrors + i;
2108
2109                 if (!disk->rdev ||
2110                     !test_bit(In_sync, &disk->rdev->flags)) {
2111                         disk->head_position = 0;
2112                         mddev->degraded++;
2113                 }
2114         }
2115
2116
2117         mddev->thread = md_register_thread(raid10d, mddev, "%s_raid10");
2118         if (!mddev->thread) {
2119                 printk(KERN_ERR
2120                        "raid10: couldn't allocate thread for %s\n",
2121                        mdname(mddev));
2122                 goto out_free_conf;
2123         }
2124
2125         printk(KERN_INFO
2126                 "raid10: raid set %s active with %d out of %d devices\n",
2127                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2128                 mddev->raid_disks);
2129         /*
2130          * Ok, everything is just fine now
2131          */
2132         mddev->array_size = size << (conf->chunk_shift-1);
2133         mddev->resync_max_sectors = size << conf->chunk_shift;
2134
2135         mddev->queue->unplug_fn = raid10_unplug;
2136         mddev->queue->issue_flush_fn = raid10_issue_flush;
2137         mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2138         mddev->queue->backing_dev_info.congested_data = mddev;
2139
2140         /* Calculate max read-ahead size.
2141          * We need to readahead at least twice a whole stripe....
2142          * maybe...
2143          */
2144         {
2145                 int stripe = conf->raid_disks * (mddev->chunk_size / PAGE_SIZE);
2146                 stripe /= conf->near_copies;
2147                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2148                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2149         }
2150
2151         if (conf->near_copies < mddev->raid_disks)
2152                 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2153         return 0;
2154
2155 out_free_conf:
2156         if (conf->r10bio_pool)
2157                 mempool_destroy(conf->r10bio_pool);
2158         safe_put_page(conf->tmppage);
2159         kfree(conf->mirrors);
2160         kfree(conf);
2161         mddev->private = NULL;
2162 out:
2163         return -EIO;
2164 }
2165
2166 static int stop(mddev_t *mddev)
2167 {
2168         conf_t *conf = mddev_to_conf(mddev);
2169
2170         md_unregister_thread(mddev->thread);
2171         mddev->thread = NULL;
2172         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2173         if (conf->r10bio_pool)
2174                 mempool_destroy(conf->r10bio_pool);
2175         kfree(conf->mirrors);
2176         kfree(conf);
2177         mddev->private = NULL;
2178         return 0;
2179 }
2180
2181 static void raid10_quiesce(mddev_t *mddev, int state)
2182 {
2183         conf_t *conf = mddev_to_conf(mddev);
2184
2185         switch(state) {
2186         case 1:
2187                 raise_barrier(conf, 0);
2188                 break;
2189         case 0:
2190                 lower_barrier(conf);
2191                 break;
2192         }
2193         if (mddev->thread) {
2194                 if (mddev->bitmap)
2195                         mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
2196                 else
2197                         mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
2198                 md_wakeup_thread(mddev->thread);
2199         }
2200 }
2201
2202 static struct mdk_personality raid10_personality =
2203 {
2204         .name           = "raid10",
2205         .level          = 10,
2206         .owner          = THIS_MODULE,
2207         .make_request   = make_request,
2208         .run            = run,
2209         .stop           = stop,
2210         .status         = status,
2211         .error_handler  = error,
2212         .hot_add_disk   = raid10_add_disk,
2213         .hot_remove_disk= raid10_remove_disk,
2214         .spare_active   = raid10_spare_active,
2215         .sync_request   = sync_request,
2216         .quiesce        = raid10_quiesce,
2217 };
2218
2219 static int __init raid_init(void)
2220 {
2221         return register_md_personality(&raid10_personality);
2222 }
2223
2224 static void raid_exit(void)
2225 {
2226         unregister_md_personality(&raid10_personality);
2227 }
2228
2229 module_init(raid_init);
2230 module_exit(raid_exit);
2231 MODULE_LICENSE("GPL");
2232 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2233 MODULE_ALIAS("md-raid10");
2234 MODULE_ALIAS("md-level-10");