f03dd70d12a50502c7a9f1c1d7e2a031163b5c96
[linux-2.6.git] / drivers / md / raid10.c
1 /*
2  * raid10.c : Multiple Devices driver for Linux
3  *
4  * Copyright (C) 2000-2004 Neil Brown
5  *
6  * RAID-10 support for md.
7  *
8  * Base on code in raid1.c.  See raid1.c for futher copyright information.
9  *
10  *
11  * This program is free software; you can redistribute it and/or modify
12  * it under the terms of the GNU General Public License as published by
13  * the Free Software Foundation; either version 2, or (at your option)
14  * any later version.
15  *
16  * You should have received a copy of the GNU General Public License
17  * (for example /usr/src/linux/COPYING); if not, write to the Free
18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19  */
20
21 #include "dm-bio-list.h"
22 #include <linux/delay.h>
23 #include "raid10.h"
24 #include "bitmap.h"
25
26 /*
27  * RAID10 provides a combination of RAID0 and RAID1 functionality.
28  * The layout of data is defined by
29  *    chunk_size
30  *    raid_disks
31  *    near_copies (stored in low byte of layout)
32  *    far_copies (stored in second byte of layout)
33  *    far_offset (stored in bit 16 of layout )
34  *
35  * The data to be stored is divided into chunks using chunksize.
36  * Each device is divided into far_copies sections.
37  * In each section, chunks are laid out in a style similar to raid0, but
38  * near_copies copies of each chunk is stored (each on a different drive).
39  * The starting device for each section is offset near_copies from the starting
40  * device of the previous section.
41  * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
42  * drive.
43  * near_copies and far_copies must be at least one, and their product is at most
44  * raid_disks.
45  *
46  * If far_offset is true, then the far_copies are handled a bit differently.
47  * The copies are still in different stripes, but instead of be very far apart
48  * on disk, there are adjacent stripes.
49  */
50
51 /*
52  * Number of guaranteed r10bios in case of extreme VM load:
53  */
54 #define NR_RAID10_BIOS 256
55
56 static void unplug_slaves(mddev_t *mddev);
57
58 static void allow_barrier(conf_t *conf);
59 static void lower_barrier(conf_t *conf);
60
61 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
62 {
63         conf_t *conf = data;
64         r10bio_t *r10_bio;
65         int size = offsetof(struct r10bio_s, devs[conf->copies]);
66
67         /* allocate a r10bio with room for raid_disks entries in the bios array */
68         r10_bio = kzalloc(size, gfp_flags);
69         if (!r10_bio)
70                 unplug_slaves(conf->mddev);
71
72         return r10_bio;
73 }
74
75 static void r10bio_pool_free(void *r10_bio, void *data)
76 {
77         kfree(r10_bio);
78 }
79
80 /* Maximum size of each resync request */
81 #define RESYNC_BLOCK_SIZE (64*1024)
82 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
83 /* amount of memory to reserve for resync requests */
84 #define RESYNC_WINDOW (1024*1024)
85 /* maximum number of concurrent requests, memory permitting */
86 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
87
88 /*
89  * When performing a resync, we need to read and compare, so
90  * we need as many pages are there are copies.
91  * When performing a recovery, we need 2 bios, one for read,
92  * one for write (we recover only one drive per r10buf)
93  *
94  */
95 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
96 {
97         conf_t *conf = data;
98         struct page *page;
99         r10bio_t *r10_bio;
100         struct bio *bio;
101         int i, j;
102         int nalloc;
103
104         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
105         if (!r10_bio) {
106                 unplug_slaves(conf->mddev);
107                 return NULL;
108         }
109
110         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
111                 nalloc = conf->copies; /* resync */
112         else
113                 nalloc = 2; /* recovery */
114
115         /*
116          * Allocate bios.
117          */
118         for (j = nalloc ; j-- ; ) {
119                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
120                 if (!bio)
121                         goto out_free_bio;
122                 r10_bio->devs[j].bio = bio;
123         }
124         /*
125          * Allocate RESYNC_PAGES data pages and attach them
126          * where needed.
127          */
128         for (j = 0 ; j < nalloc; j++) {
129                 bio = r10_bio->devs[j].bio;
130                 for (i = 0; i < RESYNC_PAGES; i++) {
131                         page = alloc_page(gfp_flags);
132                         if (unlikely(!page))
133                                 goto out_free_pages;
134
135                         bio->bi_io_vec[i].bv_page = page;
136                 }
137         }
138
139         return r10_bio;
140
141 out_free_pages:
142         for ( ; i > 0 ; i--)
143                 safe_put_page(bio->bi_io_vec[i-1].bv_page);
144         while (j--)
145                 for (i = 0; i < RESYNC_PAGES ; i++)
146                         safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
147         j = -1;
148 out_free_bio:
149         while ( ++j < nalloc )
150                 bio_put(r10_bio->devs[j].bio);
151         r10bio_pool_free(r10_bio, conf);
152         return NULL;
153 }
154
155 static void r10buf_pool_free(void *__r10_bio, void *data)
156 {
157         int i;
158         conf_t *conf = data;
159         r10bio_t *r10bio = __r10_bio;
160         int j;
161
162         for (j=0; j < conf->copies; j++) {
163                 struct bio *bio = r10bio->devs[j].bio;
164                 if (bio) {
165                         for (i = 0; i < RESYNC_PAGES; i++) {
166                                 safe_put_page(bio->bi_io_vec[i].bv_page);
167                                 bio->bi_io_vec[i].bv_page = NULL;
168                         }
169                         bio_put(bio);
170                 }
171         }
172         r10bio_pool_free(r10bio, conf);
173 }
174
175 static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
176 {
177         int i;
178
179         for (i = 0; i < conf->copies; i++) {
180                 struct bio **bio = & r10_bio->devs[i].bio;
181                 if (*bio && *bio != IO_BLOCKED)
182                         bio_put(*bio);
183                 *bio = NULL;
184         }
185 }
186
187 static void free_r10bio(r10bio_t *r10_bio)
188 {
189         conf_t *conf = mddev_to_conf(r10_bio->mddev);
190
191         /*
192          * Wake up any possible resync thread that waits for the device
193          * to go idle.
194          */
195         allow_barrier(conf);
196
197         put_all_bios(conf, r10_bio);
198         mempool_free(r10_bio, conf->r10bio_pool);
199 }
200
201 static void put_buf(r10bio_t *r10_bio)
202 {
203         conf_t *conf = mddev_to_conf(r10_bio->mddev);
204
205         mempool_free(r10_bio, conf->r10buf_pool);
206
207         lower_barrier(conf);
208 }
209
210 static void reschedule_retry(r10bio_t *r10_bio)
211 {
212         unsigned long flags;
213         mddev_t *mddev = r10_bio->mddev;
214         conf_t *conf = mddev_to_conf(mddev);
215
216         spin_lock_irqsave(&conf->device_lock, flags);
217         list_add(&r10_bio->retry_list, &conf->retry_list);
218         conf->nr_queued ++;
219         spin_unlock_irqrestore(&conf->device_lock, flags);
220
221         /* wake up frozen array... */
222         wake_up(&conf->wait_barrier);
223
224         md_wakeup_thread(mddev->thread);
225 }
226
227 /*
228  * raid_end_bio_io() is called when we have finished servicing a mirrored
229  * operation and are ready to return a success/failure code to the buffer
230  * cache layer.
231  */
232 static void raid_end_bio_io(r10bio_t *r10_bio)
233 {
234         struct bio *bio = r10_bio->master_bio;
235
236         bio_endio(bio,
237                 test_bit(R10BIO_Uptodate, &r10_bio->state) ? 0 : -EIO);
238         free_r10bio(r10_bio);
239 }
240
241 /*
242  * Update disk head position estimator based on IRQ completion info.
243  */
244 static inline void update_head_pos(int slot, r10bio_t *r10_bio)
245 {
246         conf_t *conf = mddev_to_conf(r10_bio->mddev);
247
248         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
249                 r10_bio->devs[slot].addr + (r10_bio->sectors);
250 }
251
252 static void raid10_end_read_request(struct bio *bio, int error)
253 {
254         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
255         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
256         int slot, dev;
257         conf_t *conf = mddev_to_conf(r10_bio->mddev);
258
259
260         slot = r10_bio->read_slot;
261         dev = r10_bio->devs[slot].devnum;
262         /*
263          * this branch is our 'one mirror IO has finished' event handler:
264          */
265         update_head_pos(slot, r10_bio);
266
267         if (uptodate) {
268                 /*
269                  * Set R10BIO_Uptodate in our master bio, so that
270                  * we will return a good error code to the higher
271                  * levels even if IO on some other mirrored buffer fails.
272                  *
273                  * The 'master' represents the composite IO operation to
274                  * user-side. So if something waits for IO, then it will
275                  * wait for the 'master' bio.
276                  */
277                 set_bit(R10BIO_Uptodate, &r10_bio->state);
278                 raid_end_bio_io(r10_bio);
279         } else {
280                 /*
281                  * oops, read error:
282                  */
283                 char b[BDEVNAME_SIZE];
284                 if (printk_ratelimit())
285                         printk(KERN_ERR "raid10: %s: rescheduling sector %llu\n",
286                                bdevname(conf->mirrors[dev].rdev->bdev,b), (unsigned long long)r10_bio->sector);
287                 reschedule_retry(r10_bio);
288         }
289
290         rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
291 }
292
293 static void raid10_end_write_request(struct bio *bio, int error)
294 {
295         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
296         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
297         int slot, dev;
298         conf_t *conf = mddev_to_conf(r10_bio->mddev);
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 }
343
344
345 /*
346  * RAID10 layout manager
347  * Aswell as the chunksize and raid_disks count, there are two
348  * parameters: near_copies and far_copies.
349  * near_copies * far_copies must be <= raid_disks.
350  * Normally one of these will be 1.
351  * If both are 1, we get raid0.
352  * If near_copies == raid_disks, we get raid1.
353  *
354  * Chunks are layed out in raid0 style with near_copies copies of the
355  * first chunk, followed by near_copies copies of the next chunk and
356  * so on.
357  * If far_copies > 1, then after 1/far_copies of the array has been assigned
358  * as described above, we start again with a device offset of near_copies.
359  * So we effectively have another copy of the whole array further down all
360  * the drives, but with blocks on different drives.
361  * With this layout, and block is never stored twice on the one device.
362  *
363  * raid10_find_phys finds the sector offset of a given virtual sector
364  * on each device that it is on.
365  *
366  * raid10_find_virt does the reverse mapping, from a device and a
367  * sector offset to a virtual address
368  */
369
370 static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
371 {
372         int n,f;
373         sector_t sector;
374         sector_t chunk;
375         sector_t stripe;
376         int dev;
377
378         int slot = 0;
379
380         /* now calculate first sector/dev */
381         chunk = r10bio->sector >> conf->chunk_shift;
382         sector = r10bio->sector & conf->chunk_mask;
383
384         chunk *= conf->near_copies;
385         stripe = chunk;
386         dev = sector_div(stripe, conf->raid_disks);
387         if (conf->far_offset)
388                 stripe *= conf->far_copies;
389
390         sector += stripe << conf->chunk_shift;
391
392         /* and calculate all the others */
393         for (n=0; n < conf->near_copies; n++) {
394                 int d = dev;
395                 sector_t s = sector;
396                 r10bio->devs[slot].addr = sector;
397                 r10bio->devs[slot].devnum = d;
398                 slot++;
399
400                 for (f = 1; f < conf->far_copies; f++) {
401                         d += conf->near_copies;
402                         if (d >= conf->raid_disks)
403                                 d -= conf->raid_disks;
404                         s += conf->stride;
405                         r10bio->devs[slot].devnum = d;
406                         r10bio->devs[slot].addr = s;
407                         slot++;
408                 }
409                 dev++;
410                 if (dev >= conf->raid_disks) {
411                         dev = 0;
412                         sector += (conf->chunk_mask + 1);
413                 }
414         }
415         BUG_ON(slot != conf->copies);
416 }
417
418 static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
419 {
420         sector_t offset, chunk, vchunk;
421
422         offset = sector & conf->chunk_mask;
423         if (conf->far_offset) {
424                 int fc;
425                 chunk = sector >> conf->chunk_shift;
426                 fc = sector_div(chunk, conf->far_copies);
427                 dev -= fc * conf->near_copies;
428                 if (dev < 0)
429                         dev += conf->raid_disks;
430         } else {
431                 while (sector >= conf->stride) {
432                         sector -= conf->stride;
433                         if (dev < conf->near_copies)
434                                 dev += conf->raid_disks - conf->near_copies;
435                         else
436                                 dev -= conf->near_copies;
437                 }
438                 chunk = sector >> conf->chunk_shift;
439         }
440         vchunk = chunk * conf->raid_disks + dev;
441         sector_div(vchunk, conf->near_copies);
442         return (vchunk << conf->chunk_shift) + offset;
443 }
444
445 /**
446  *      raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
447  *      @q: request queue
448  *      @bvm: properties of new bio
449  *      @biovec: the request that could be merged to it.
450  *
451  *      Return amount of bytes we can accept at this offset
452  *      If near_copies == raid_disk, there are no striping issues,
453  *      but in that case, the function isn't called at all.
454  */
455 static int raid10_mergeable_bvec(struct request_queue *q,
456                                  struct bvec_merge_data *bvm,
457                                  struct bio_vec *biovec)
458 {
459         mddev_t *mddev = q->queuedata;
460         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
461         int max;
462         unsigned int chunk_sectors = mddev->chunk_size >> 9;
463         unsigned int bio_sectors = bvm->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 <= biovec->bv_len && bio_sectors == 0)
468                 return biovec->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          * or - for far > 1 - find the closest to partition beginning */
549
550         for (nslot = slot; nslot < conf->copies; nslot++) {
551                 int ndisk = r10_bio->devs[nslot].devnum;
552
553
554                 if ((rdev=rcu_dereference(conf->mirrors[ndisk].rdev)) == NULL ||
555                     r10_bio->devs[nslot].bio == IO_BLOCKED ||
556                     !test_bit(In_sync, &rdev->flags))
557                         continue;
558
559                 /* This optimisation is debatable, and completely destroys
560                  * sequential read speed for 'far copies' arrays.  So only
561                  * keep it for 'near' arrays, and review those later.
562                  */
563                 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending)) {
564                         disk = ndisk;
565                         slot = nslot;
566                         break;
567                 }
568
569                 /* for far > 1 always use the lowest address */
570                 if (conf->far_copies > 1)
571                         new_distance = r10_bio->devs[nslot].addr;
572                 else
573                         new_distance = abs(r10_bio->devs[nslot].addr -
574                                            conf->mirrors[ndisk].head_position);
575                 if (new_distance < current_distance) {
576                         current_distance = new_distance;
577                         disk = ndisk;
578                         slot = nslot;
579                 }
580         }
581
582 rb_out:
583         r10_bio->read_slot = slot;
584 /*      conf->next_seq_sect = this_sector + sectors;*/
585
586         if (disk >= 0 && (rdev=rcu_dereference(conf->mirrors[disk].rdev))!= NULL)
587                 atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
588         else
589                 disk = -1;
590         rcu_read_unlock();
591
592         return disk;
593 }
594
595 static void unplug_slaves(mddev_t *mddev)
596 {
597         conf_t *conf = mddev_to_conf(mddev);
598         int i;
599
600         rcu_read_lock();
601         for (i=0; i<mddev->raid_disks; i++) {
602                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
603                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
604                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
605
606                         atomic_inc(&rdev->nr_pending);
607                         rcu_read_unlock();
608
609                         blk_unplug(r_queue);
610
611                         rdev_dec_pending(rdev, mddev);
612                         rcu_read_lock();
613                 }
614         }
615         rcu_read_unlock();
616 }
617
618 static void raid10_unplug(struct request_queue *q)
619 {
620         mddev_t *mddev = q->queuedata;
621
622         unplug_slaves(q->queuedata);
623         md_wakeup_thread(mddev->thread);
624 }
625
626 static int raid10_congested(void *data, int bits)
627 {
628         mddev_t *mddev = data;
629         conf_t *conf = mddev_to_conf(mddev);
630         int i, ret = 0;
631
632         rcu_read_lock();
633         for (i = 0; i < mddev->raid_disks && ret == 0; i++) {
634                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
635                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
636                         struct request_queue *q = bdev_get_queue(rdev->bdev);
637
638                         ret |= bdi_congested(&q->backing_dev_info, bits);
639                 }
640         }
641         rcu_read_unlock();
642         return ret;
643 }
644
645 static int flush_pending_writes(conf_t *conf)
646 {
647         /* Any writes that have been queued but are awaiting
648          * bitmap updates get flushed here.
649          * We return 1 if any requests were actually submitted.
650          */
651         int rv = 0;
652
653         spin_lock_irq(&conf->device_lock);
654
655         if (conf->pending_bio_list.head) {
656                 struct bio *bio;
657                 bio = bio_list_get(&conf->pending_bio_list);
658                 blk_remove_plug(conf->mddev->queue);
659                 spin_unlock_irq(&conf->device_lock);
660                 /* flush any pending bitmap writes to disk
661                  * before proceeding w/ I/O */
662                 bitmap_unplug(conf->mddev->bitmap);
663
664                 while (bio) { /* submit pending writes */
665                         struct bio *next = bio->bi_next;
666                         bio->bi_next = NULL;
667                         generic_make_request(bio);
668                         bio = next;
669                 }
670                 rv = 1;
671         } else
672                 spin_unlock_irq(&conf->device_lock);
673         return rv;
674 }
675 /* Barriers....
676  * Sometimes we need to suspend IO while we do something else,
677  * either some resync/recovery, or reconfigure the array.
678  * To do this we raise a 'barrier'.
679  * The 'barrier' is a counter that can be raised multiple times
680  * to count how many activities are happening which preclude
681  * normal IO.
682  * We can only raise the barrier if there is no pending IO.
683  * i.e. if nr_pending == 0.
684  * We choose only to raise the barrier if no-one is waiting for the
685  * barrier to go down.  This means that as soon as an IO request
686  * is ready, no other operations which require a barrier will start
687  * until the IO request has had a chance.
688  *
689  * So: regular IO calls 'wait_barrier'.  When that returns there
690  *    is no backgroup IO happening,  It must arrange to call
691  *    allow_barrier when it has finished its IO.
692  * backgroup IO calls must call raise_barrier.  Once that returns
693  *    there is no normal IO happeing.  It must arrange to call
694  *    lower_barrier when the particular background IO completes.
695  */
696
697 static void raise_barrier(conf_t *conf, int force)
698 {
699         BUG_ON(force && !conf->barrier);
700         spin_lock_irq(&conf->resync_lock);
701
702         /* Wait until no block IO is waiting (unless 'force') */
703         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
704                             conf->resync_lock,
705                             raid10_unplug(conf->mddev->queue));
706
707         /* block any new IO from starting */
708         conf->barrier++;
709
710         /* No wait for all pending IO to complete */
711         wait_event_lock_irq(conf->wait_barrier,
712                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
713                             conf->resync_lock,
714                             raid10_unplug(conf->mddev->queue));
715
716         spin_unlock_irq(&conf->resync_lock);
717 }
718
719 static void lower_barrier(conf_t *conf)
720 {
721         unsigned long flags;
722         spin_lock_irqsave(&conf->resync_lock, flags);
723         conf->barrier--;
724         spin_unlock_irqrestore(&conf->resync_lock, flags);
725         wake_up(&conf->wait_barrier);
726 }
727
728 static void wait_barrier(conf_t *conf)
729 {
730         spin_lock_irq(&conf->resync_lock);
731         if (conf->barrier) {
732                 conf->nr_waiting++;
733                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
734                                     conf->resync_lock,
735                                     raid10_unplug(conf->mddev->queue));
736                 conf->nr_waiting--;
737         }
738         conf->nr_pending++;
739         spin_unlock_irq(&conf->resync_lock);
740 }
741
742 static void allow_barrier(conf_t *conf)
743 {
744         unsigned long flags;
745         spin_lock_irqsave(&conf->resync_lock, flags);
746         conf->nr_pending--;
747         spin_unlock_irqrestore(&conf->resync_lock, flags);
748         wake_up(&conf->wait_barrier);
749 }
750
751 static void freeze_array(conf_t *conf)
752 {
753         /* stop syncio and normal IO and wait for everything to
754          * go quiet.
755          * We increment barrier and nr_waiting, and then
756          * wait until nr_pending match nr_queued+1
757          * This is called in the context of one normal IO request
758          * that has failed. Thus any sync request that might be pending
759          * will be blocked by nr_pending, and we need to wait for
760          * pending IO requests to complete or be queued for re-try.
761          * Thus the number queued (nr_queued) plus this request (1)
762          * must match the number of pending IOs (nr_pending) before
763          * we continue.
764          */
765         spin_lock_irq(&conf->resync_lock);
766         conf->barrier++;
767         conf->nr_waiting++;
768         wait_event_lock_irq(conf->wait_barrier,
769                             conf->nr_pending == conf->nr_queued+1,
770                             conf->resync_lock,
771                             ({ flush_pending_writes(conf);
772                                raid10_unplug(conf->mddev->queue); }));
773         spin_unlock_irq(&conf->resync_lock);
774 }
775
776 static void unfreeze_array(conf_t *conf)
777 {
778         /* reverse the effect of the freeze */
779         spin_lock_irq(&conf->resync_lock);
780         conf->barrier--;
781         conf->nr_waiting--;
782         wake_up(&conf->wait_barrier);
783         spin_unlock_irq(&conf->resync_lock);
784 }
785
786 static int make_request(struct request_queue *q, struct bio * bio)
787 {
788         mddev_t *mddev = q->queuedata;
789         conf_t *conf = mddev_to_conf(mddev);
790         mirror_info_t *mirror;
791         r10bio_t *r10_bio;
792         struct bio *read_bio;
793         int cpu;
794         int i;
795         int chunk_sects = conf->chunk_mask + 1;
796         const int rw = bio_data_dir(bio);
797         const int do_sync = bio_sync(bio);
798         struct bio_list bl;
799         unsigned long flags;
800         mdk_rdev_t *blocked_rdev;
801
802         if (unlikely(bio_barrier(bio))) {
803                 bio_endio(bio, -EOPNOTSUPP);
804                 return 0;
805         }
806
807         /* If this request crosses a chunk boundary, we need to
808          * split it.  This will only happen for 1 PAGE (or less) requests.
809          */
810         if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
811                       > chunk_sects &&
812                     conf->near_copies < conf->raid_disks)) {
813                 struct bio_pair *bp;
814                 /* Sanity check -- queue functions should prevent this happening */
815                 if (bio->bi_vcnt != 1 ||
816                     bio->bi_idx != 0)
817                         goto bad_map;
818                 /* This is a one page bio that upper layers
819                  * refuse to split for us, so we need to split it.
820                  */
821                 bp = bio_split(bio,
822                                chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
823                 if (make_request(q, &bp->bio1))
824                         generic_make_request(&bp->bio1);
825                 if (make_request(q, &bp->bio2))
826                         generic_make_request(&bp->bio2);
827
828                 bio_pair_release(bp);
829                 return 0;
830         bad_map:
831                 printk("raid10_make_request bug: can't convert block across chunks"
832                        " or bigger than %dk %llu %d\n", chunk_sects/2,
833                        (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
834
835                 bio_io_error(bio);
836                 return 0;
837         }
838
839         md_write_start(mddev, bio);
840
841         /*
842          * Register the new request and wait if the reconstruction
843          * thread has put up a bar for new requests.
844          * Continue immediately if no resync is active currently.
845          */
846         wait_barrier(conf);
847
848         cpu = part_stat_lock();
849         part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
850         part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
851                       bio_sectors(bio));
852         part_stat_unlock();
853
854         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
855
856         r10_bio->master_bio = bio;
857         r10_bio->sectors = bio->bi_size >> 9;
858
859         r10_bio->mddev = mddev;
860         r10_bio->sector = bio->bi_sector;
861         r10_bio->state = 0;
862
863         if (rw == READ) {
864                 /*
865                  * read balancing logic:
866                  */
867                 int disk = read_balance(conf, r10_bio);
868                 int slot = r10_bio->read_slot;
869                 if (disk < 0) {
870                         raid_end_bio_io(r10_bio);
871                         return 0;
872                 }
873                 mirror = conf->mirrors + disk;
874
875                 read_bio = bio_clone(bio, GFP_NOIO);
876
877                 r10_bio->devs[slot].bio = read_bio;
878
879                 read_bio->bi_sector = r10_bio->devs[slot].addr +
880                         mirror->rdev->data_offset;
881                 read_bio->bi_bdev = mirror->rdev->bdev;
882                 read_bio->bi_end_io = raid10_end_read_request;
883                 read_bio->bi_rw = READ | do_sync;
884                 read_bio->bi_private = r10_bio;
885
886                 generic_make_request(read_bio);
887                 return 0;
888         }
889
890         /*
891          * WRITE:
892          */
893         /* first select target devices under rcu_lock and
894          * inc refcount on their rdev.  Record them by setting
895          * bios[x] to bio
896          */
897         raid10_find_phys(conf, r10_bio);
898  retry_write:
899         blocked_rdev = NULL;
900         rcu_read_lock();
901         for (i = 0;  i < conf->copies; i++) {
902                 int d = r10_bio->devs[i].devnum;
903                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
904                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
905                         atomic_inc(&rdev->nr_pending);
906                         blocked_rdev = rdev;
907                         break;
908                 }
909                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
910                         atomic_inc(&rdev->nr_pending);
911                         r10_bio->devs[i].bio = bio;
912                 } else {
913                         r10_bio->devs[i].bio = NULL;
914                         set_bit(R10BIO_Degraded, &r10_bio->state);
915                 }
916         }
917         rcu_read_unlock();
918
919         if (unlikely(blocked_rdev)) {
920                 /* Have to wait for this device to get unblocked, then retry */
921                 int j;
922                 int d;
923
924                 for (j = 0; j < i; j++)
925                         if (r10_bio->devs[j].bio) {
926                                 d = r10_bio->devs[j].devnum;
927                                 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
928                         }
929                 allow_barrier(conf);
930                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
931                 wait_barrier(conf);
932                 goto retry_write;
933         }
934
935         atomic_set(&r10_bio->remaining, 0);
936
937         bio_list_init(&bl);
938         for (i = 0; i < conf->copies; i++) {
939                 struct bio *mbio;
940                 int d = r10_bio->devs[i].devnum;
941                 if (!r10_bio->devs[i].bio)
942                         continue;
943
944                 mbio = bio_clone(bio, GFP_NOIO);
945                 r10_bio->devs[i].bio = mbio;
946
947                 mbio->bi_sector = r10_bio->devs[i].addr+
948                         conf->mirrors[d].rdev->data_offset;
949                 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
950                 mbio->bi_end_io = raid10_end_write_request;
951                 mbio->bi_rw = WRITE | do_sync;
952                 mbio->bi_private = r10_bio;
953
954                 atomic_inc(&r10_bio->remaining);
955                 bio_list_add(&bl, mbio);
956         }
957
958         if (unlikely(!atomic_read(&r10_bio->remaining))) {
959                 /* the array is dead */
960                 md_write_end(mddev);
961                 raid_end_bio_io(r10_bio);
962                 return 0;
963         }
964
965         bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
966         spin_lock_irqsave(&conf->device_lock, flags);
967         bio_list_merge(&conf->pending_bio_list, &bl);
968         blk_plug_device(mddev->queue);
969         spin_unlock_irqrestore(&conf->device_lock, flags);
970
971         /* In case raid10d snuck in to freeze_array */
972         wake_up(&conf->wait_barrier);
973
974         if (do_sync)
975                 md_wakeup_thread(mddev->thread);
976
977         return 0;
978 }
979
980 static void status(struct seq_file *seq, mddev_t *mddev)
981 {
982         conf_t *conf = mddev_to_conf(mddev);
983         int i;
984
985         if (conf->near_copies < conf->raid_disks)
986                 seq_printf(seq, " %dK chunks", mddev->chunk_size/1024);
987         if (conf->near_copies > 1)
988                 seq_printf(seq, " %d near-copies", conf->near_copies);
989         if (conf->far_copies > 1) {
990                 if (conf->far_offset)
991                         seq_printf(seq, " %d offset-copies", conf->far_copies);
992                 else
993                         seq_printf(seq, " %d far-copies", conf->far_copies);
994         }
995         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
996                                         conf->raid_disks - mddev->degraded);
997         for (i = 0; i < conf->raid_disks; i++)
998                 seq_printf(seq, "%s",
999                               conf->mirrors[i].rdev &&
1000                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1001         seq_printf(seq, "]");
1002 }
1003
1004 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1005 {
1006         char b[BDEVNAME_SIZE];
1007         conf_t *conf = mddev_to_conf(mddev);
1008
1009         /*
1010          * If it is not operational, then we have already marked it as dead
1011          * else if it is the last working disks, ignore the error, let the
1012          * next level up know.
1013          * else mark the drive as failed
1014          */
1015         if (test_bit(In_sync, &rdev->flags)
1016             && conf->raid_disks-mddev->degraded == 1)
1017                 /*
1018                  * Don't fail the drive, just return an IO error.
1019                  * The test should really be more sophisticated than
1020                  * "working_disks == 1", but it isn't critical, and
1021                  * can wait until we do more sophisticated "is the drive
1022                  * really dead" tests...
1023                  */
1024                 return;
1025         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1026                 unsigned long flags;
1027                 spin_lock_irqsave(&conf->device_lock, flags);
1028                 mddev->degraded++;
1029                 spin_unlock_irqrestore(&conf->device_lock, flags);
1030                 /*
1031                  * if recovery is running, make sure it aborts.
1032                  */
1033                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1034         }
1035         set_bit(Faulty, &rdev->flags);
1036         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1037         printk(KERN_ALERT "raid10: Disk failure on %s, disabling device.\n"
1038                 "raid10: Operation continuing on %d devices.\n",
1039                 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1040 }
1041
1042 static void print_conf(conf_t *conf)
1043 {
1044         int i;
1045         mirror_info_t *tmp;
1046
1047         printk("RAID10 conf printout:\n");
1048         if (!conf) {
1049                 printk("(!conf)\n");
1050                 return;
1051         }
1052         printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1053                 conf->raid_disks);
1054
1055         for (i = 0; i < conf->raid_disks; i++) {
1056                 char b[BDEVNAME_SIZE];
1057                 tmp = conf->mirrors + i;
1058                 if (tmp->rdev)
1059                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
1060                                 i, !test_bit(In_sync, &tmp->rdev->flags),
1061                                 !test_bit(Faulty, &tmp->rdev->flags),
1062                                 bdevname(tmp->rdev->bdev,b));
1063         }
1064 }
1065
1066 static void close_sync(conf_t *conf)
1067 {
1068         wait_barrier(conf);
1069         allow_barrier(conf);
1070
1071         mempool_destroy(conf->r10buf_pool);
1072         conf->r10buf_pool = NULL;
1073 }
1074
1075 /* check if there are enough drives for
1076  * every block to appear on atleast one
1077  */
1078 static int enough(conf_t *conf)
1079 {
1080         int first = 0;
1081
1082         do {
1083                 int n = conf->copies;
1084                 int cnt = 0;
1085                 while (n--) {
1086                         if (conf->mirrors[first].rdev)
1087                                 cnt++;
1088                         first = (first+1) % conf->raid_disks;
1089                 }
1090                 if (cnt == 0)
1091                         return 0;
1092         } while (first != 0);
1093         return 1;
1094 }
1095
1096 static int raid10_spare_active(mddev_t *mddev)
1097 {
1098         int i;
1099         conf_t *conf = mddev->private;
1100         mirror_info_t *tmp;
1101
1102         /*
1103          * Find all non-in_sync disks within the RAID10 configuration
1104          * and mark them in_sync
1105          */
1106         for (i = 0; i < conf->raid_disks; i++) {
1107                 tmp = conf->mirrors + i;
1108                 if (tmp->rdev
1109                     && !test_bit(Faulty, &tmp->rdev->flags)
1110                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1111                         unsigned long flags;
1112                         spin_lock_irqsave(&conf->device_lock, flags);
1113                         mddev->degraded--;
1114                         spin_unlock_irqrestore(&conf->device_lock, flags);
1115                 }
1116         }
1117
1118         print_conf(conf);
1119         return 0;
1120 }
1121
1122
1123 static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1124 {
1125         conf_t *conf = mddev->private;
1126         int err = -EEXIST;
1127         int mirror;
1128         mirror_info_t *p;
1129         int first = 0;
1130         int last = mddev->raid_disks - 1;
1131
1132         if (mddev->recovery_cp < MaxSector)
1133                 /* only hot-add to in-sync arrays, as recovery is
1134                  * very different from resync
1135                  */
1136                 return -EBUSY;
1137         if (!enough(conf))
1138                 return -EINVAL;
1139
1140         if (rdev->raid_disk >= 0)
1141                 first = last = rdev->raid_disk;
1142
1143         if (rdev->saved_raid_disk >= 0 &&
1144             rdev->saved_raid_disk >= first &&
1145             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1146                 mirror = rdev->saved_raid_disk;
1147         else
1148                 mirror = first;
1149         for ( ; mirror <= last ; mirror++)
1150                 if ( !(p=conf->mirrors+mirror)->rdev) {
1151
1152                         blk_queue_stack_limits(mddev->queue,
1153                                                rdev->bdev->bd_disk->queue);
1154                         /* as we don't honour merge_bvec_fn, we must never risk
1155                          * violating it, so limit ->max_sector to one PAGE, as
1156                          * a one page request is never in violation.
1157                          */
1158                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1159                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
1160                                 mddev->queue->max_sectors = (PAGE_SIZE>>9);
1161
1162                         p->head_position = 0;
1163                         rdev->raid_disk = mirror;
1164                         err = 0;
1165                         if (rdev->saved_raid_disk != mirror)
1166                                 conf->fullsync = 1;
1167                         rcu_assign_pointer(p->rdev, rdev);
1168                         break;
1169                 }
1170
1171         print_conf(conf);
1172         return err;
1173 }
1174
1175 static int raid10_remove_disk(mddev_t *mddev, int number)
1176 {
1177         conf_t *conf = mddev->private;
1178         int err = 0;
1179         mdk_rdev_t *rdev;
1180         mirror_info_t *p = conf->mirrors+ number;
1181
1182         print_conf(conf);
1183         rdev = p->rdev;
1184         if (rdev) {
1185                 if (test_bit(In_sync, &rdev->flags) ||
1186                     atomic_read(&rdev->nr_pending)) {
1187                         err = -EBUSY;
1188                         goto abort;
1189                 }
1190                 /* Only remove faulty devices in recovery
1191                  * is not possible.
1192                  */
1193                 if (!test_bit(Faulty, &rdev->flags) &&
1194                     enough(conf)) {
1195                         err = -EBUSY;
1196                         goto abort;
1197                 }
1198                 p->rdev = NULL;
1199                 synchronize_rcu();
1200                 if (atomic_read(&rdev->nr_pending)) {
1201                         /* lost the race, try later */
1202                         err = -EBUSY;
1203                         p->rdev = rdev;
1204                 }
1205         }
1206 abort:
1207
1208         print_conf(conf);
1209         return err;
1210 }
1211
1212
1213 static void end_sync_read(struct bio *bio, int error)
1214 {
1215         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1216         conf_t *conf = mddev_to_conf(r10_bio->mddev);
1217         int i,d;
1218
1219         for (i=0; i<conf->copies; i++)
1220                 if (r10_bio->devs[i].bio == bio)
1221                         break;
1222         BUG_ON(i == conf->copies);
1223         update_head_pos(i, r10_bio);
1224         d = r10_bio->devs[i].devnum;
1225
1226         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1227                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1228         else {
1229                 atomic_add(r10_bio->sectors,
1230                            &conf->mirrors[d].rdev->corrected_errors);
1231                 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1232                         md_error(r10_bio->mddev,
1233                                  conf->mirrors[d].rdev);
1234         }
1235
1236         /* for reconstruct, we always reschedule after a read.
1237          * for resync, only after all reads
1238          */
1239         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1240         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1241             atomic_dec_and_test(&r10_bio->remaining)) {
1242                 /* we have read all the blocks,
1243                  * do the comparison in process context in raid10d
1244                  */
1245                 reschedule_retry(r10_bio);
1246         }
1247 }
1248
1249 static void end_sync_write(struct bio *bio, int error)
1250 {
1251         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1252         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1253         mddev_t *mddev = r10_bio->mddev;
1254         conf_t *conf = mddev_to_conf(mddev);
1255         int i,d;
1256
1257         for (i = 0; i < conf->copies; i++)
1258                 if (r10_bio->devs[i].bio == bio)
1259                         break;
1260         d = r10_bio->devs[i].devnum;
1261
1262         if (!uptodate)
1263                 md_error(mddev, conf->mirrors[d].rdev);
1264
1265         update_head_pos(i, r10_bio);
1266
1267         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1268         while (atomic_dec_and_test(&r10_bio->remaining)) {
1269                 if (r10_bio->master_bio == NULL) {
1270                         /* the primary of several recovery bios */
1271                         sector_t s = r10_bio->sectors;
1272                         put_buf(r10_bio);
1273                         md_done_sync(mddev, s, 1);
1274                         break;
1275                 } else {
1276                         r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1277                         put_buf(r10_bio);
1278                         r10_bio = r10_bio2;
1279                 }
1280         }
1281 }
1282
1283 /*
1284  * Note: sync and recover and handled very differently for raid10
1285  * This code is for resync.
1286  * For resync, we read through virtual addresses and read all blocks.
1287  * If there is any error, we schedule a write.  The lowest numbered
1288  * drive is authoritative.
1289  * However requests come for physical address, so we need to map.
1290  * For every physical address there are raid_disks/copies virtual addresses,
1291  * which is always are least one, but is not necessarly an integer.
1292  * This means that a physical address can span multiple chunks, so we may
1293  * have to submit multiple io requests for a single sync request.
1294  */
1295 /*
1296  * We check if all blocks are in-sync and only write to blocks that
1297  * aren't in sync
1298  */
1299 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1300 {
1301         conf_t *conf = mddev_to_conf(mddev);
1302         int i, first;
1303         struct bio *tbio, *fbio;
1304
1305         atomic_set(&r10_bio->remaining, 1);
1306
1307         /* find the first device with a block */
1308         for (i=0; i<conf->copies; i++)
1309                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1310                         break;
1311
1312         if (i == conf->copies)
1313                 goto done;
1314
1315         first = i;
1316         fbio = r10_bio->devs[i].bio;
1317
1318         /* now find blocks with errors */
1319         for (i=0 ; i < conf->copies ; i++) {
1320                 int  j, d;
1321                 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1322
1323                 tbio = r10_bio->devs[i].bio;
1324
1325                 if (tbio->bi_end_io != end_sync_read)
1326                         continue;
1327                 if (i == first)
1328                         continue;
1329                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1330                         /* We know that the bi_io_vec layout is the same for
1331                          * both 'first' and 'i', so we just compare them.
1332                          * All vec entries are PAGE_SIZE;
1333                          */
1334                         for (j = 0; j < vcnt; j++)
1335                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1336                                            page_address(tbio->bi_io_vec[j].bv_page),
1337                                            PAGE_SIZE))
1338                                         break;
1339                         if (j == vcnt)
1340                                 continue;
1341                         mddev->resync_mismatches += r10_bio->sectors;
1342                 }
1343                 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1344                         /* Don't fix anything. */
1345                         continue;
1346                 /* Ok, we need to write this bio
1347                  * First we need to fixup bv_offset, bv_len and
1348                  * bi_vecs, as the read request might have corrupted these
1349                  */
1350                 tbio->bi_vcnt = vcnt;
1351                 tbio->bi_size = r10_bio->sectors << 9;
1352                 tbio->bi_idx = 0;
1353                 tbio->bi_phys_segments = 0;
1354                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1355                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1356                 tbio->bi_next = NULL;
1357                 tbio->bi_rw = WRITE;
1358                 tbio->bi_private = r10_bio;
1359                 tbio->bi_sector = r10_bio->devs[i].addr;
1360
1361                 for (j=0; j < vcnt ; j++) {
1362                         tbio->bi_io_vec[j].bv_offset = 0;
1363                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1364
1365                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1366                                page_address(fbio->bi_io_vec[j].bv_page),
1367                                PAGE_SIZE);
1368                 }
1369                 tbio->bi_end_io = end_sync_write;
1370
1371                 d = r10_bio->devs[i].devnum;
1372                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1373                 atomic_inc(&r10_bio->remaining);
1374                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1375
1376                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1377                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1378                 generic_make_request(tbio);
1379         }
1380
1381 done:
1382         if (atomic_dec_and_test(&r10_bio->remaining)) {
1383                 md_done_sync(mddev, r10_bio->sectors, 1);
1384                 put_buf(r10_bio);
1385         }
1386 }
1387
1388 /*
1389  * Now for the recovery code.
1390  * Recovery happens across physical sectors.
1391  * We recover all non-is_sync drives by finding the virtual address of
1392  * each, and then choose a working drive that also has that virt address.
1393  * There is a separate r10_bio for each non-in_sync drive.
1394  * Only the first two slots are in use. The first for reading,
1395  * The second for writing.
1396  *
1397  */
1398
1399 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1400 {
1401         conf_t *conf = mddev_to_conf(mddev);
1402         int i, d;
1403         struct bio *bio, *wbio;
1404
1405
1406         /* move the pages across to the second bio
1407          * and submit the write request
1408          */
1409         bio = r10_bio->devs[0].bio;
1410         wbio = r10_bio->devs[1].bio;
1411         for (i=0; i < wbio->bi_vcnt; i++) {
1412                 struct page *p = bio->bi_io_vec[i].bv_page;
1413                 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1414                 wbio->bi_io_vec[i].bv_page = p;
1415         }
1416         d = r10_bio->devs[1].devnum;
1417
1418         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1419         md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1420         if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1421                 generic_make_request(wbio);
1422         else
1423                 bio_endio(wbio, -EIO);
1424 }
1425
1426
1427 /*
1428  * This is a kernel thread which:
1429  *
1430  *      1.      Retries failed read operations on working mirrors.
1431  *      2.      Updates the raid superblock when problems encounter.
1432  *      3.      Performs writes following reads for array synchronising.
1433  */
1434
1435 static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1436 {
1437         int sect = 0; /* Offset from r10_bio->sector */
1438         int sectors = r10_bio->sectors;
1439         mdk_rdev_t*rdev;
1440         while(sectors) {
1441                 int s = sectors;
1442                 int sl = r10_bio->read_slot;
1443                 int success = 0;
1444                 int start;
1445
1446                 if (s > (PAGE_SIZE>>9))
1447                         s = PAGE_SIZE >> 9;
1448
1449                 rcu_read_lock();
1450                 do {
1451                         int d = r10_bio->devs[sl].devnum;
1452                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1453                         if (rdev &&
1454                             test_bit(In_sync, &rdev->flags)) {
1455                                 atomic_inc(&rdev->nr_pending);
1456                                 rcu_read_unlock();
1457                                 success = sync_page_io(rdev->bdev,
1458                                                        r10_bio->devs[sl].addr +
1459                                                        sect + rdev->data_offset,
1460                                                        s<<9,
1461                                                        conf->tmppage, READ);
1462                                 rdev_dec_pending(rdev, mddev);
1463                                 rcu_read_lock();
1464                                 if (success)
1465                                         break;
1466                         }
1467                         sl++;
1468                         if (sl == conf->copies)
1469                                 sl = 0;
1470                 } while (!success && sl != r10_bio->read_slot);
1471                 rcu_read_unlock();
1472
1473                 if (!success) {
1474                         /* Cannot read from anywhere -- bye bye array */
1475                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1476                         md_error(mddev, conf->mirrors[dn].rdev);
1477                         break;
1478                 }
1479
1480                 start = sl;
1481                 /* write it back and re-read */
1482                 rcu_read_lock();
1483                 while (sl != r10_bio->read_slot) {
1484                         int d;
1485                         if (sl==0)
1486                                 sl = conf->copies;
1487                         sl--;
1488                         d = r10_bio->devs[sl].devnum;
1489                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1490                         if (rdev &&
1491                             test_bit(In_sync, &rdev->flags)) {
1492                                 atomic_inc(&rdev->nr_pending);
1493                                 rcu_read_unlock();
1494                                 atomic_add(s, &rdev->corrected_errors);
1495                                 if (sync_page_io(rdev->bdev,
1496                                                  r10_bio->devs[sl].addr +
1497                                                  sect + rdev->data_offset,
1498                                                  s<<9, conf->tmppage, WRITE)
1499                                     == 0)
1500                                         /* Well, this device is dead */
1501                                         md_error(mddev, rdev);
1502                                 rdev_dec_pending(rdev, mddev);
1503                                 rcu_read_lock();
1504                         }
1505                 }
1506                 sl = start;
1507                 while (sl != r10_bio->read_slot) {
1508                         int d;
1509                         if (sl==0)
1510                                 sl = conf->copies;
1511                         sl--;
1512                         d = r10_bio->devs[sl].devnum;
1513                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1514                         if (rdev &&
1515                             test_bit(In_sync, &rdev->flags)) {
1516                                 char b[BDEVNAME_SIZE];
1517                                 atomic_inc(&rdev->nr_pending);
1518                                 rcu_read_unlock();
1519                                 if (sync_page_io(rdev->bdev,
1520                                                  r10_bio->devs[sl].addr +
1521                                                  sect + rdev->data_offset,
1522                                                  s<<9, conf->tmppage, READ) == 0)
1523                                         /* Well, this device is dead */
1524                                         md_error(mddev, rdev);
1525                                 else
1526                                         printk(KERN_INFO
1527                                                "raid10:%s: read error corrected"
1528                                                " (%d sectors at %llu on %s)\n",
1529                                                mdname(mddev), s,
1530                                                (unsigned long long)(sect+
1531                                                     rdev->data_offset),
1532                                                bdevname(rdev->bdev, b));
1533
1534                                 rdev_dec_pending(rdev, mddev);
1535                                 rcu_read_lock();
1536                         }
1537                 }
1538                 rcu_read_unlock();
1539
1540                 sectors -= s;
1541                 sect += s;
1542         }
1543 }
1544
1545 static void raid10d(mddev_t *mddev)
1546 {
1547         r10bio_t *r10_bio;
1548         struct bio *bio;
1549         unsigned long flags;
1550         conf_t *conf = mddev_to_conf(mddev);
1551         struct list_head *head = &conf->retry_list;
1552         int unplug=0;
1553         mdk_rdev_t *rdev;
1554
1555         md_check_recovery(mddev);
1556
1557         for (;;) {
1558                 char b[BDEVNAME_SIZE];
1559
1560                 unplug += flush_pending_writes(conf);
1561
1562                 spin_lock_irqsave(&conf->device_lock, flags);
1563                 if (list_empty(head)) {
1564                         spin_unlock_irqrestore(&conf->device_lock, flags);
1565                         break;
1566                 }
1567                 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1568                 list_del(head->prev);
1569                 conf->nr_queued--;
1570                 spin_unlock_irqrestore(&conf->device_lock, flags);
1571
1572                 mddev = r10_bio->mddev;
1573                 conf = mddev_to_conf(mddev);
1574                 if (test_bit(R10BIO_IsSync, &r10_bio->state)) {
1575                         sync_request_write(mddev, r10_bio);
1576                         unplug = 1;
1577                 } else  if (test_bit(R10BIO_IsRecover, &r10_bio->state)) {
1578                         recovery_request_write(mddev, r10_bio);
1579                         unplug = 1;
1580                 } else {
1581                         int mirror;
1582                         /* we got a read error. Maybe the drive is bad.  Maybe just
1583                          * the block and we can fix it.
1584                          * We freeze all other IO, and try reading the block from
1585                          * other devices.  When we find one, we re-write
1586                          * and check it that fixes the read error.
1587                          * This is all done synchronously while the array is
1588                          * frozen.
1589                          */
1590                         if (mddev->ro == 0) {
1591                                 freeze_array(conf);
1592                                 fix_read_error(conf, mddev, r10_bio);
1593                                 unfreeze_array(conf);
1594                         }
1595
1596                         bio = r10_bio->devs[r10_bio->read_slot].bio;
1597                         r10_bio->devs[r10_bio->read_slot].bio =
1598                                 mddev->ro ? IO_BLOCKED : NULL;
1599                         mirror = read_balance(conf, r10_bio);
1600                         if (mirror == -1) {
1601                                 printk(KERN_ALERT "raid10: %s: unrecoverable I/O"
1602                                        " read error for block %llu\n",
1603                                        bdevname(bio->bi_bdev,b),
1604                                        (unsigned long long)r10_bio->sector);
1605                                 raid_end_bio_io(r10_bio);
1606                                 bio_put(bio);
1607                         } else {
1608                                 const int do_sync = bio_sync(r10_bio->master_bio);
1609                                 bio_put(bio);
1610                                 rdev = conf->mirrors[mirror].rdev;
1611                                 if (printk_ratelimit())
1612                                         printk(KERN_ERR "raid10: %s: redirecting sector %llu to"
1613                                                " another mirror\n",
1614                                                bdevname(rdev->bdev,b),
1615                                                (unsigned long long)r10_bio->sector);
1616                                 bio = bio_clone(r10_bio->master_bio, GFP_NOIO);
1617                                 r10_bio->devs[r10_bio->read_slot].bio = bio;
1618                                 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1619                                         + rdev->data_offset;
1620                                 bio->bi_bdev = rdev->bdev;
1621                                 bio->bi_rw = READ | do_sync;
1622                                 bio->bi_private = r10_bio;
1623                                 bio->bi_end_io = raid10_end_read_request;
1624                                 unplug = 1;
1625                                 generic_make_request(bio);
1626                         }
1627                 }
1628         }
1629         if (unplug)
1630                 unplug_slaves(mddev);
1631 }
1632
1633
1634 static int init_resync(conf_t *conf)
1635 {
1636         int buffs;
1637
1638         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1639         BUG_ON(conf->r10buf_pool);
1640         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1641         if (!conf->r10buf_pool)
1642                 return -ENOMEM;
1643         conf->next_resync = 0;
1644         return 0;
1645 }
1646
1647 /*
1648  * perform a "sync" on one "block"
1649  *
1650  * We need to make sure that no normal I/O request - particularly write
1651  * requests - conflict with active sync requests.
1652  *
1653  * This is achieved by tracking pending requests and a 'barrier' concept
1654  * that can be installed to exclude normal IO requests.
1655  *
1656  * Resync and recovery are handled very differently.
1657  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1658  *
1659  * For resync, we iterate over virtual addresses, read all copies,
1660  * and update if there are differences.  If only one copy is live,
1661  * skip it.
1662  * For recovery, we iterate over physical addresses, read a good
1663  * value for each non-in_sync drive, and over-write.
1664  *
1665  * So, for recovery we may have several outstanding complex requests for a
1666  * given address, one for each out-of-sync device.  We model this by allocating
1667  * a number of r10_bio structures, one for each out-of-sync device.
1668  * As we setup these structures, we collect all bio's together into a list
1669  * which we then process collectively to add pages, and then process again
1670  * to pass to generic_make_request.
1671  *
1672  * The r10_bio structures are linked using a borrowed master_bio pointer.
1673  * This link is counted in ->remaining.  When the r10_bio that points to NULL
1674  * has its remaining count decremented to 0, the whole complex operation
1675  * is complete.
1676  *
1677  */
1678
1679 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1680 {
1681         conf_t *conf = mddev_to_conf(mddev);
1682         r10bio_t *r10_bio;
1683         struct bio *biolist = NULL, *bio;
1684         sector_t max_sector, nr_sectors;
1685         int disk;
1686         int i;
1687         int max_sync;
1688         int sync_blocks;
1689
1690         sector_t sectors_skipped = 0;
1691         int chunks_skipped = 0;
1692
1693         if (!conf->r10buf_pool)
1694                 if (init_resync(conf))
1695                         return 0;
1696
1697  skipped:
1698         max_sector = mddev->size << 1;
1699         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1700                 max_sector = mddev->resync_max_sectors;
1701         if (sector_nr >= max_sector) {
1702                 /* If we aborted, we need to abort the
1703                  * sync on the 'current' bitmap chucks (there can
1704                  * be several when recovering multiple devices).
1705                  * as we may have started syncing it but not finished.
1706                  * We can find the current address in
1707                  * mddev->curr_resync, but for recovery,
1708                  * we need to convert that to several
1709                  * virtual addresses.
1710                  */
1711                 if (mddev->curr_resync < max_sector) { /* aborted */
1712                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1713                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1714                                                 &sync_blocks, 1);
1715                         else for (i=0; i<conf->raid_disks; i++) {
1716                                 sector_t sect =
1717                                         raid10_find_virt(conf, mddev->curr_resync, i);
1718                                 bitmap_end_sync(mddev->bitmap, sect,
1719                                                 &sync_blocks, 1);
1720                         }
1721                 } else /* completed sync */
1722                         conf->fullsync = 0;
1723
1724                 bitmap_close_sync(mddev->bitmap);
1725                 close_sync(conf);
1726                 *skipped = 1;
1727                 return sectors_skipped;
1728         }
1729         if (chunks_skipped >= conf->raid_disks) {
1730                 /* if there has been nothing to do on any drive,
1731                  * then there is nothing to do at all..
1732                  */
1733                 *skipped = 1;
1734                 return (max_sector - sector_nr) + sectors_skipped;
1735         }
1736
1737         if (max_sector > mddev->resync_max)
1738                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1739
1740         /* make sure whole request will fit in a chunk - if chunks
1741          * are meaningful
1742          */
1743         if (conf->near_copies < conf->raid_disks &&
1744             max_sector > (sector_nr | conf->chunk_mask))
1745                 max_sector = (sector_nr | conf->chunk_mask) + 1;
1746         /*
1747          * If there is non-resync activity waiting for us then
1748          * put in a delay to throttle resync.
1749          */
1750         if (!go_faster && conf->nr_waiting)
1751                 msleep_interruptible(1000);
1752
1753         /* Again, very different code for resync and recovery.
1754          * Both must result in an r10bio with a list of bios that
1755          * have bi_end_io, bi_sector, bi_bdev set,
1756          * and bi_private set to the r10bio.
1757          * For recovery, we may actually create several r10bios
1758          * with 2 bios in each, that correspond to the bios in the main one.
1759          * In this case, the subordinate r10bios link back through a
1760          * borrowed master_bio pointer, and the counter in the master
1761          * includes a ref from each subordinate.
1762          */
1763         /* First, we decide what to do and set ->bi_end_io
1764          * To end_sync_read if we want to read, and
1765          * end_sync_write if we will want to write.
1766          */
1767
1768         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
1769         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1770                 /* recovery... the complicated one */
1771                 int i, j, k;
1772                 r10_bio = NULL;
1773
1774                 for (i=0 ; i<conf->raid_disks; i++)
1775                         if (conf->mirrors[i].rdev &&
1776                             !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) {
1777                                 int still_degraded = 0;
1778                                 /* want to reconstruct this device */
1779                                 r10bio_t *rb2 = r10_bio;
1780                                 sector_t sect = raid10_find_virt(conf, sector_nr, i);
1781                                 int must_sync;
1782                                 /* Unless we are doing a full sync, we only need
1783                                  * to recover the block if it is set in the bitmap
1784                                  */
1785                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1786                                                               &sync_blocks, 1);
1787                                 if (sync_blocks < max_sync)
1788                                         max_sync = sync_blocks;
1789                                 if (!must_sync &&
1790                                     !conf->fullsync) {
1791                                         /* yep, skip the sync_blocks here, but don't assume
1792                                          * that there will never be anything to do here
1793                                          */
1794                                         chunks_skipped = -1;
1795                                         continue;
1796                                 }
1797
1798                                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1799                                 raise_barrier(conf, rb2 != NULL);
1800                                 atomic_set(&r10_bio->remaining, 0);
1801
1802                                 r10_bio->master_bio = (struct bio*)rb2;
1803                                 if (rb2)
1804                                         atomic_inc(&rb2->remaining);
1805                                 r10_bio->mddev = mddev;
1806                                 set_bit(R10BIO_IsRecover, &r10_bio->state);
1807                                 r10_bio->sector = sect;
1808
1809                                 raid10_find_phys(conf, r10_bio);
1810                                 /* Need to check if this section will still be
1811                                  * degraded
1812                                  */
1813                                 for (j=0; j<conf->copies;j++) {
1814                                         int d = r10_bio->devs[j].devnum;
1815                                         if (conf->mirrors[d].rdev == NULL ||
1816                                             test_bit(Faulty, &conf->mirrors[d].rdev->flags)) {
1817                                                 still_degraded = 1;
1818                                                 break;
1819                                         }
1820                                 }
1821                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1822                                                               &sync_blocks, still_degraded);
1823
1824                                 for (j=0; j<conf->copies;j++) {
1825                                         int d = r10_bio->devs[j].devnum;
1826                                         if (conf->mirrors[d].rdev &&
1827                                             test_bit(In_sync, &conf->mirrors[d].rdev->flags)) {
1828                                                 /* This is where we read from */
1829                                                 bio = r10_bio->devs[0].bio;
1830                                                 bio->bi_next = biolist;
1831                                                 biolist = bio;
1832                                                 bio->bi_private = r10_bio;
1833                                                 bio->bi_end_io = end_sync_read;
1834                                                 bio->bi_rw = READ;
1835                                                 bio->bi_sector = r10_bio->devs[j].addr +
1836                                                         conf->mirrors[d].rdev->data_offset;
1837                                                 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1838                                                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1839                                                 atomic_inc(&r10_bio->remaining);
1840                                                 /* and we write to 'i' */
1841
1842                                                 for (k=0; k<conf->copies; k++)
1843                                                         if (r10_bio->devs[k].devnum == i)
1844                                                                 break;
1845                                                 BUG_ON(k == conf->copies);
1846                                                 bio = r10_bio->devs[1].bio;
1847                                                 bio->bi_next = biolist;
1848                                                 biolist = bio;
1849                                                 bio->bi_private = r10_bio;
1850                                                 bio->bi_end_io = end_sync_write;
1851                                                 bio->bi_rw = WRITE;
1852                                                 bio->bi_sector = r10_bio->devs[k].addr +
1853                                                         conf->mirrors[i].rdev->data_offset;
1854                                                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1855
1856                                                 r10_bio->devs[0].devnum = d;
1857                                                 r10_bio->devs[1].devnum = i;
1858
1859                                                 break;
1860                                         }
1861                                 }
1862                                 if (j == conf->copies) {
1863                                         /* Cannot recover, so abort the recovery */
1864                                         put_buf(r10_bio);
1865                                         if (rb2)
1866                                                 atomic_dec(&rb2->remaining);
1867                                         r10_bio = rb2;
1868                                         if (!test_and_set_bit(MD_RECOVERY_INTR,
1869                                                               &mddev->recovery))
1870                                                 printk(KERN_INFO "raid10: %s: insufficient working devices for recovery.\n",
1871                                                        mdname(mddev));
1872                                         break;
1873                                 }
1874                         }
1875                 if (biolist == NULL) {
1876                         while (r10_bio) {
1877                                 r10bio_t *rb2 = r10_bio;
1878                                 r10_bio = (r10bio_t*) rb2->master_bio;
1879                                 rb2->master_bio = NULL;
1880                                 put_buf(rb2);
1881                         }
1882                         goto giveup;
1883                 }
1884         } else {
1885                 /* resync. Schedule a read for every block at this virt offset */
1886                 int count = 0;
1887
1888                 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1889
1890                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1891                                        &sync_blocks, mddev->degraded) &&
1892                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1893                         /* We can skip this block */
1894                         *skipped = 1;
1895                         return sync_blocks + sectors_skipped;
1896                 }
1897                 if (sync_blocks < max_sync)
1898                         max_sync = sync_blocks;
1899                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1900
1901                 r10_bio->mddev = mddev;
1902                 atomic_set(&r10_bio->remaining, 0);
1903                 raise_barrier(conf, 0);
1904                 conf->next_resync = sector_nr;
1905
1906                 r10_bio->master_bio = NULL;
1907                 r10_bio->sector = sector_nr;
1908                 set_bit(R10BIO_IsSync, &r10_bio->state);
1909                 raid10_find_phys(conf, r10_bio);
1910                 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1911
1912                 for (i=0; i<conf->copies; i++) {
1913                         int d = r10_bio->devs[i].devnum;
1914                         bio = r10_bio->devs[i].bio;
1915                         bio->bi_end_io = NULL;
1916                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
1917                         if (conf->mirrors[d].rdev == NULL ||
1918                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
1919                                 continue;
1920                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1921                         atomic_inc(&r10_bio->remaining);
1922                         bio->bi_next = biolist;
1923                         biolist = bio;
1924                         bio->bi_private = r10_bio;
1925                         bio->bi_end_io = end_sync_read;
1926                         bio->bi_rw = READ;
1927                         bio->bi_sector = r10_bio->devs[i].addr +
1928                                 conf->mirrors[d].rdev->data_offset;
1929                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1930                         count++;
1931                 }
1932
1933                 if (count < 2) {
1934                         for (i=0; i<conf->copies; i++) {
1935                                 int d = r10_bio->devs[i].devnum;
1936                                 if (r10_bio->devs[i].bio->bi_end_io)
1937                                         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1938                         }
1939                         put_buf(r10_bio);
1940                         biolist = NULL;
1941                         goto giveup;
1942                 }
1943         }
1944
1945         for (bio = biolist; bio ; bio=bio->bi_next) {
1946
1947                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1948                 if (bio->bi_end_io)
1949                         bio->bi_flags |= 1 << BIO_UPTODATE;
1950                 bio->bi_vcnt = 0;
1951                 bio->bi_idx = 0;
1952                 bio->bi_phys_segments = 0;
1953                 bio->bi_size = 0;
1954         }
1955
1956         nr_sectors = 0;
1957         if (sector_nr + max_sync < max_sector)
1958                 max_sector = sector_nr + max_sync;
1959         do {
1960                 struct page *page;
1961                 int len = PAGE_SIZE;
1962                 disk = 0;
1963                 if (sector_nr + (len>>9) > max_sector)
1964                         len = (max_sector - sector_nr) << 9;
1965                 if (len == 0)
1966                         break;
1967                 for (bio= biolist ; bio ; bio=bio->bi_next) {
1968                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1969                         if (bio_add_page(bio, page, len, 0) == 0) {
1970                                 /* stop here */
1971                                 struct bio *bio2;
1972                                 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1973                                 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
1974                                         /* remove last page from this bio */
1975                                         bio2->bi_vcnt--;
1976                                         bio2->bi_size -= len;
1977                                         bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
1978                                 }
1979                                 goto bio_full;
1980                         }
1981                         disk = i;
1982                 }
1983                 nr_sectors += len>>9;
1984                 sector_nr += len>>9;
1985         } while (biolist->bi_vcnt < RESYNC_PAGES);
1986  bio_full:
1987         r10_bio->sectors = nr_sectors;
1988
1989         while (biolist) {
1990                 bio = biolist;
1991                 biolist = biolist->bi_next;
1992
1993                 bio->bi_next = NULL;
1994                 r10_bio = bio->bi_private;
1995                 r10_bio->sectors = nr_sectors;
1996
1997                 if (bio->bi_end_io == end_sync_read) {
1998                         md_sync_acct(bio->bi_bdev, nr_sectors);
1999                         generic_make_request(bio);
2000                 }
2001         }
2002
2003         if (sectors_skipped)
2004                 /* pretend they weren't skipped, it makes
2005                  * no important difference in this case
2006                  */
2007                 md_done_sync(mddev, sectors_skipped, 1);
2008
2009         return sectors_skipped + nr_sectors;
2010  giveup:
2011         /* There is nowhere to write, so all non-sync
2012          * drives must be failed, so try the next chunk...
2013          */
2014         if (sector_nr + max_sync < max_sector)
2015                 max_sector = sector_nr + max_sync;
2016
2017         sectors_skipped += (max_sector - sector_nr);
2018         chunks_skipped ++;
2019         sector_nr = max_sector;
2020         goto skipped;
2021 }
2022
2023 static int run(mddev_t *mddev)
2024 {
2025         conf_t *conf;
2026         int i, disk_idx;
2027         mirror_info_t *disk;
2028         mdk_rdev_t *rdev;
2029         int nc, fc, fo;
2030         sector_t stride, size;
2031
2032         if (mddev->chunk_size < PAGE_SIZE) {
2033                 printk(KERN_ERR "md/raid10: chunk size must be "
2034                        "at least PAGE_SIZE(%ld).\n", PAGE_SIZE);
2035                 return -EINVAL;
2036         }
2037
2038         nc = mddev->layout & 255;
2039         fc = (mddev->layout >> 8) & 255;
2040         fo = mddev->layout & (1<<16);
2041         if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
2042             (mddev->layout >> 17)) {
2043                 printk(KERN_ERR "raid10: %s: unsupported raid10 layout: 0x%8x\n",
2044                        mdname(mddev), mddev->layout);
2045                 goto out;
2046         }
2047         /*
2048          * copy the already verified devices into our private RAID10
2049          * bookkeeping area. [whatever we allocate in run(),
2050          * should be freed in stop()]
2051          */
2052         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2053         mddev->private = conf;
2054         if (!conf) {
2055                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2056                         mdname(mddev));
2057                 goto out;
2058         }
2059         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2060                                  GFP_KERNEL);
2061         if (!conf->mirrors) {
2062                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2063                        mdname(mddev));
2064                 goto out_free_conf;
2065         }
2066
2067         conf->tmppage = alloc_page(GFP_KERNEL);
2068         if (!conf->tmppage)
2069                 goto out_free_conf;
2070
2071         conf->mddev = mddev;
2072         conf->raid_disks = mddev->raid_disks;
2073         conf->near_copies = nc;
2074         conf->far_copies = fc;
2075         conf->copies = nc*fc;
2076         conf->far_offset = fo;
2077         conf->chunk_mask = (sector_t)(mddev->chunk_size>>9)-1;
2078         conf->chunk_shift = ffz(~mddev->chunk_size) - 9;
2079         size = mddev->size >> (conf->chunk_shift-1);
2080         sector_div(size, fc);
2081         size = size * conf->raid_disks;
2082         sector_div(size, nc);
2083         /* 'size' is now the number of chunks in the array */
2084         /* calculate "used chunks per device" in 'stride' */
2085         stride = size * conf->copies;
2086
2087         /* We need to round up when dividing by raid_disks to
2088          * get the stride size.
2089          */
2090         stride += conf->raid_disks - 1;
2091         sector_div(stride, conf->raid_disks);
2092         mddev->size = stride  << (conf->chunk_shift-1);
2093
2094         if (fo)
2095                 stride = 1;
2096         else
2097                 sector_div(stride, fc);
2098         conf->stride = stride << conf->chunk_shift;
2099
2100         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2101                                                 r10bio_pool_free, conf);
2102         if (!conf->r10bio_pool) {
2103                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2104                         mdname(mddev));
2105                 goto out_free_conf;
2106         }
2107
2108         spin_lock_init(&conf->device_lock);
2109         mddev->queue->queue_lock = &conf->device_lock;
2110
2111         list_for_each_entry(rdev, &mddev->disks, same_set) {
2112                 disk_idx = rdev->raid_disk;
2113                 if (disk_idx >= mddev->raid_disks
2114                     || disk_idx < 0)
2115                         continue;
2116                 disk = conf->mirrors + disk_idx;
2117
2118                 disk->rdev = rdev;
2119
2120                 blk_queue_stack_limits(mddev->queue,
2121                                        rdev->bdev->bd_disk->queue);
2122                 /* as we don't honour merge_bvec_fn, we must never risk
2123                  * violating it, so limit ->max_sector to one PAGE, as
2124                  * a one page request is never in violation.
2125                  */
2126                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
2127                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
2128                         mddev->queue->max_sectors = (PAGE_SIZE>>9);
2129
2130                 disk->head_position = 0;
2131         }
2132         INIT_LIST_HEAD(&conf->retry_list);
2133
2134         spin_lock_init(&conf->resync_lock);
2135         init_waitqueue_head(&conf->wait_barrier);
2136
2137         /* need to check that every block has at least one working mirror */
2138         if (!enough(conf)) {
2139                 printk(KERN_ERR "raid10: not enough operational mirrors for %s\n",
2140                        mdname(mddev));
2141                 goto out_free_conf;
2142         }
2143
2144         mddev->degraded = 0;
2145         for (i = 0; i < conf->raid_disks; i++) {
2146
2147                 disk = conf->mirrors + i;
2148
2149                 if (!disk->rdev ||
2150                     !test_bit(In_sync, &disk->rdev->flags)) {
2151                         disk->head_position = 0;
2152                         mddev->degraded++;
2153                         if (disk->rdev)
2154                                 conf->fullsync = 1;
2155                 }
2156         }
2157
2158
2159         mddev->thread = md_register_thread(raid10d, mddev, "%s_raid10");
2160         if (!mddev->thread) {
2161                 printk(KERN_ERR
2162                        "raid10: couldn't allocate thread for %s\n",
2163                        mdname(mddev));
2164                 goto out_free_conf;
2165         }
2166
2167         printk(KERN_INFO
2168                 "raid10: raid set %s active with %d out of %d devices\n",
2169                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2170                 mddev->raid_disks);
2171         /*
2172          * Ok, everything is just fine now
2173          */
2174         mddev->array_sectors = size << conf->chunk_shift;
2175         mddev->resync_max_sectors = size << conf->chunk_shift;
2176
2177         mddev->queue->unplug_fn = raid10_unplug;
2178         mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2179         mddev->queue->backing_dev_info.congested_data = mddev;
2180
2181         /* Calculate max read-ahead size.
2182          * We need to readahead at least twice a whole stripe....
2183          * maybe...
2184          */
2185         {
2186                 int stripe = conf->raid_disks * (mddev->chunk_size / PAGE_SIZE);
2187                 stripe /= conf->near_copies;
2188                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2189                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2190         }
2191
2192         if (conf->near_copies < mddev->raid_disks)
2193                 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2194         return 0;
2195
2196 out_free_conf:
2197         if (conf->r10bio_pool)
2198                 mempool_destroy(conf->r10bio_pool);
2199         safe_put_page(conf->tmppage);
2200         kfree(conf->mirrors);
2201         kfree(conf);
2202         mddev->private = NULL;
2203 out:
2204         return -EIO;
2205 }
2206
2207 static int stop(mddev_t *mddev)
2208 {
2209         conf_t *conf = mddev_to_conf(mddev);
2210
2211         md_unregister_thread(mddev->thread);
2212         mddev->thread = NULL;
2213         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2214         if (conf->r10bio_pool)
2215                 mempool_destroy(conf->r10bio_pool);
2216         kfree(conf->mirrors);
2217         kfree(conf);
2218         mddev->private = NULL;
2219         return 0;
2220 }
2221
2222 static void raid10_quiesce(mddev_t *mddev, int state)
2223 {
2224         conf_t *conf = mddev_to_conf(mddev);
2225
2226         switch(state) {
2227         case 1:
2228                 raise_barrier(conf, 0);
2229                 break;
2230         case 0:
2231                 lower_barrier(conf);
2232                 break;
2233         }
2234         if (mddev->thread) {
2235                 if (mddev->bitmap)
2236                         mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
2237                 else
2238                         mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
2239                 md_wakeup_thread(mddev->thread);
2240         }
2241 }
2242
2243 static struct mdk_personality raid10_personality =
2244 {
2245         .name           = "raid10",
2246         .level          = 10,
2247         .owner          = THIS_MODULE,
2248         .make_request   = make_request,
2249         .run            = run,
2250         .stop           = stop,
2251         .status         = status,
2252         .error_handler  = error,
2253         .hot_add_disk   = raid10_add_disk,
2254         .hot_remove_disk= raid10_remove_disk,
2255         .spare_active   = raid10_spare_active,
2256         .sync_request   = sync_request,
2257         .quiesce        = raid10_quiesce,
2258 };
2259
2260 static int __init raid_init(void)
2261 {
2262         return register_md_personality(&raid10_personality);
2263 }
2264
2265 static void raid_exit(void)
2266 {
2267         unregister_md_personality(&raid10_personality);
2268 }
2269
2270 module_init(raid_init);
2271 module_exit(raid_exit);
2272 MODULE_LICENSE("GPL");
2273 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2274 MODULE_ALIAS("md-raid10");
2275 MODULE_ALIAS("md-level-10");