Merge branch 'machtypes' into pxa-palm
[linux-2.6] / drivers / md / raid10.c
1 /*
2  * raid10.c : Multiple Devices driver for Linux
3  *
4  * Copyright (C) 2000-2004 Neil Brown
5  *
6  * RAID-10 support for md.
7  *
8  * Base on code in raid1.c.  See raid1.c for futher copyright information.
9  *
10  *
11  * This program is free software; you can redistribute it and/or modify
12  * it under the terms of the GNU General Public License as published by
13  * the Free Software Foundation; either version 2, or (at your option)
14  * any later version.
15  *
16  * You should have received a copy of the GNU General Public License
17  * (for example /usr/src/linux/COPYING); if not, write to the Free
18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19  */
20
21 #include "dm-bio-list.h"
22 #include <linux/raid/raid10.h>
23 #include <linux/raid/bitmap.h>
24
25 /*
26  * RAID10 provides a combination of RAID0 and RAID1 functionality.
27  * The layout of data is defined by
28  *    chunk_size
29  *    raid_disks
30  *    near_copies (stored in low byte of layout)
31  *    far_copies (stored in second byte of layout)
32  *    far_offset (stored in bit 16 of layout )
33  *
34  * The data to be stored is divided into chunks using chunksize.
35  * Each device is divided into far_copies sections.
36  * In each section, chunks are laid out in a style similar to raid0, but
37  * near_copies copies of each chunk is stored (each on a different drive).
38  * The starting device for each section is offset near_copies from the starting
39  * device of the previous section.
40  * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
41  * drive.
42  * near_copies and far_copies must be at least one, and their product is at most
43  * raid_disks.
44  *
45  * If far_offset is true, then the far_copies are handled a bit differently.
46  * The copies are still in different stripes, but instead of be very far apart
47  * on disk, there are adjacent stripes.
48  */
49
50 /*
51  * Number of guaranteed r10bios in case of extreme VM load:
52  */
53 #define NR_RAID10_BIOS 256
54
55 static void unplug_slaves(mddev_t *mddev);
56
57 static void allow_barrier(conf_t *conf);
58 static void lower_barrier(conf_t *conf);
59
60 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
61 {
62         conf_t *conf = data;
63         r10bio_t *r10_bio;
64         int size = offsetof(struct r10bio_s, devs[conf->copies]);
65
66         /* allocate a r10bio with room for raid_disks entries in the bios array */
67         r10_bio = kzalloc(size, gfp_flags);
68         if (!r10_bio)
69                 unplug_slaves(conf->mddev);
70
71         return r10_bio;
72 }
73
74 static void r10bio_pool_free(void *r10_bio, void *data)
75 {
76         kfree(r10_bio);
77 }
78
79 #define RESYNC_BLOCK_SIZE (64*1024)
80 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
81 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
82 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
83 #define RESYNC_WINDOW (2048*1024)
84
85 /*
86  * When performing a resync, we need to read and compare, so
87  * we need as many pages are there are copies.
88  * When performing a recovery, we need 2 bios, one for read,
89  * one for write (we recover only one drive per r10buf)
90  *
91  */
92 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
93 {
94         conf_t *conf = data;
95         struct page *page;
96         r10bio_t *r10_bio;
97         struct bio *bio;
98         int i, j;
99         int nalloc;
100
101         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
102         if (!r10_bio) {
103                 unplug_slaves(conf->mddev);
104                 return NULL;
105         }
106
107         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
108                 nalloc = conf->copies; /* resync */
109         else
110                 nalloc = 2; /* recovery */
111
112         /*
113          * Allocate bios.
114          */
115         for (j = nalloc ; j-- ; ) {
116                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
117                 if (!bio)
118                         goto out_free_bio;
119                 r10_bio->devs[j].bio = bio;
120         }
121         /*
122          * Allocate RESYNC_PAGES data pages and attach them
123          * where needed.
124          */
125         for (j = 0 ; j < nalloc; j++) {
126                 bio = r10_bio->devs[j].bio;
127                 for (i = 0; i < RESYNC_PAGES; i++) {
128                         page = alloc_page(gfp_flags);
129                         if (unlikely(!page))
130                                 goto out_free_pages;
131
132                         bio->bi_io_vec[i].bv_page = page;
133                 }
134         }
135
136         return r10_bio;
137
138 out_free_pages:
139         for ( ; i > 0 ; i--)
140                 safe_put_page(bio->bi_io_vec[i-1].bv_page);
141         while (j--)
142                 for (i = 0; i < RESYNC_PAGES ; i++)
143                         safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
144         j = -1;
145 out_free_bio:
146         while ( ++j < nalloc )
147                 bio_put(r10_bio->devs[j].bio);
148         r10bio_pool_free(r10_bio, conf);
149         return NULL;
150 }
151
152 static void r10buf_pool_free(void *__r10_bio, void *data)
153 {
154         int i;
155         conf_t *conf = data;
156         r10bio_t *r10bio = __r10_bio;
157         int j;
158
159         for (j=0; j < conf->copies; j++) {
160                 struct bio *bio = r10bio->devs[j].bio;
161                 if (bio) {
162                         for (i = 0; i < RESYNC_PAGES; i++) {
163                                 safe_put_page(bio->bi_io_vec[i].bv_page);
164                                 bio->bi_io_vec[i].bv_page = NULL;
165                         }
166                         bio_put(bio);
167                 }
168         }
169         r10bio_pool_free(r10bio, conf);
170 }
171
172 static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
173 {
174         int i;
175
176         for (i = 0; i < conf->copies; i++) {
177                 struct bio **bio = & r10_bio->devs[i].bio;
178                 if (*bio && *bio != IO_BLOCKED)
179                         bio_put(*bio);
180                 *bio = NULL;
181         }
182 }
183
184 static void free_r10bio(r10bio_t *r10_bio)
185 {
186         conf_t *conf = mddev_to_conf(r10_bio->mddev);
187
188         /*
189          * Wake up any possible resync thread that waits for the device
190          * to go idle.
191          */
192         allow_barrier(conf);
193
194         put_all_bios(conf, r10_bio);
195         mempool_free(r10_bio, conf->r10bio_pool);
196 }
197
198 static void put_buf(r10bio_t *r10_bio)
199 {
200         conf_t *conf = mddev_to_conf(r10_bio->mddev);
201
202         mempool_free(r10_bio, conf->r10buf_pool);
203
204         lower_barrier(conf);
205 }
206
207 static void reschedule_retry(r10bio_t *r10_bio)
208 {
209         unsigned long flags;
210         mddev_t *mddev = r10_bio->mddev;
211         conf_t *conf = mddev_to_conf(mddev);
212
213         spin_lock_irqsave(&conf->device_lock, flags);
214         list_add(&r10_bio->retry_list, &conf->retry_list);
215         conf->nr_queued ++;
216         spin_unlock_irqrestore(&conf->device_lock, flags);
217
218         md_wakeup_thread(mddev->thread);
219 }
220
221 /*
222  * raid_end_bio_io() is called when we have finished servicing a mirrored
223  * operation and are ready to return a success/failure code to the buffer
224  * cache layer.
225  */
226 static void raid_end_bio_io(r10bio_t *r10_bio)
227 {
228         struct bio *bio = r10_bio->master_bio;
229
230         bio_endio(bio,
231                 test_bit(R10BIO_Uptodate, &r10_bio->state) ? 0 : -EIO);
232         free_r10bio(r10_bio);
233 }
234
235 /*
236  * Update disk head position estimator based on IRQ completion info.
237  */
238 static inline void update_head_pos(int slot, r10bio_t *r10_bio)
239 {
240         conf_t *conf = mddev_to_conf(r10_bio->mddev);
241
242         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
243                 r10_bio->devs[slot].addr + (r10_bio->sectors);
244 }
245
246 static void raid10_end_read_request(struct bio *bio, int error)
247 {
248         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
249         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
250         int slot, dev;
251         conf_t *conf = mddev_to_conf(r10_bio->mddev);
252
253
254         slot = r10_bio->read_slot;
255         dev = r10_bio->devs[slot].devnum;
256         /*
257          * this branch is our 'one mirror IO has finished' event handler:
258          */
259         update_head_pos(slot, r10_bio);
260
261         if (uptodate) {
262                 /*
263                  * Set R10BIO_Uptodate in our master bio, so that
264                  * we will return a good error code to the higher
265                  * levels even if IO on some other mirrored buffer fails.
266                  *
267                  * The 'master' represents the composite IO operation to
268                  * user-side. So if something waits for IO, then it will
269                  * wait for the 'master' bio.
270                  */
271                 set_bit(R10BIO_Uptodate, &r10_bio->state);
272                 raid_end_bio_io(r10_bio);
273         } else {
274                 /*
275                  * oops, read error:
276                  */
277                 char b[BDEVNAME_SIZE];
278                 if (printk_ratelimit())
279                         printk(KERN_ERR "raid10: %s: rescheduling sector %llu\n",
280                                bdevname(conf->mirrors[dev].rdev->bdev,b), (unsigned long long)r10_bio->sector);
281                 reschedule_retry(r10_bio);
282         }
283
284         rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
285 }
286
287 static void raid10_end_write_request(struct bio *bio, int error)
288 {
289         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
290         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
291         int slot, dev;
292         conf_t *conf = mddev_to_conf(r10_bio->mddev);
293
294         for (slot = 0; slot < conf->copies; slot++)
295                 if (r10_bio->devs[slot].bio == bio)
296                         break;
297         dev = r10_bio->devs[slot].devnum;
298
299         /*
300          * this branch is our 'one mirror IO has finished' event handler:
301          */
302         if (!uptodate) {
303                 md_error(r10_bio->mddev, conf->mirrors[dev].rdev);
304                 /* an I/O failed, we can't clear the bitmap */
305                 set_bit(R10BIO_Degraded, &r10_bio->state);
306         } else
307                 /*
308                  * Set R10BIO_Uptodate in our master bio, so that
309                  * we will return a good error code for to the higher
310                  * levels even if IO on some other mirrored buffer fails.
311                  *
312                  * The 'master' represents the composite IO operation to
313                  * user-side. So if something waits for IO, then it will
314                  * wait for the 'master' bio.
315                  */
316                 set_bit(R10BIO_Uptodate, &r10_bio->state);
317
318         update_head_pos(slot, r10_bio);
319
320         /*
321          *
322          * Let's see if all mirrored write operations have finished
323          * already.
324          */
325         if (atomic_dec_and_test(&r10_bio->remaining)) {
326                 /* clear the bitmap if all writes complete successfully */
327                 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
328                                 r10_bio->sectors,
329                                 !test_bit(R10BIO_Degraded, &r10_bio->state),
330                                 0);
331                 md_write_end(r10_bio->mddev);
332                 raid_end_bio_io(r10_bio);
333         }
334
335         rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
336 }
337
338
339 /*
340  * RAID10 layout manager
341  * Aswell as the chunksize and raid_disks count, there are two
342  * parameters: near_copies and far_copies.
343  * near_copies * far_copies must be <= raid_disks.
344  * Normally one of these will be 1.
345  * If both are 1, we get raid0.
346  * If near_copies == raid_disks, we get raid1.
347  *
348  * Chunks are layed out in raid0 style with near_copies copies of the
349  * first chunk, followed by near_copies copies of the next chunk and
350  * so on.
351  * If far_copies > 1, then after 1/far_copies of the array has been assigned
352  * as described above, we start again with a device offset of near_copies.
353  * So we effectively have another copy of the whole array further down all
354  * the drives, but with blocks on different drives.
355  * With this layout, and block is never stored twice on the one device.
356  *
357  * raid10_find_phys finds the sector offset of a given virtual sector
358  * on each device that it is on.
359  *
360  * raid10_find_virt does the reverse mapping, from a device and a
361  * sector offset to a virtual address
362  */
363
364 static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
365 {
366         int n,f;
367         sector_t sector;
368         sector_t chunk;
369         sector_t stripe;
370         int dev;
371
372         int slot = 0;
373
374         /* now calculate first sector/dev */
375         chunk = r10bio->sector >> conf->chunk_shift;
376         sector = r10bio->sector & conf->chunk_mask;
377
378         chunk *= conf->near_copies;
379         stripe = chunk;
380         dev = sector_div(stripe, conf->raid_disks);
381         if (conf->far_offset)
382                 stripe *= conf->far_copies;
383
384         sector += stripe << conf->chunk_shift;
385
386         /* and calculate all the others */
387         for (n=0; n < conf->near_copies; n++) {
388                 int d = dev;
389                 sector_t s = sector;
390                 r10bio->devs[slot].addr = sector;
391                 r10bio->devs[slot].devnum = d;
392                 slot++;
393
394                 for (f = 1; f < conf->far_copies; f++) {
395                         d += conf->near_copies;
396                         if (d >= conf->raid_disks)
397                                 d -= conf->raid_disks;
398                         s += conf->stride;
399                         r10bio->devs[slot].devnum = d;
400                         r10bio->devs[slot].addr = s;
401                         slot++;
402                 }
403                 dev++;
404                 if (dev >= conf->raid_disks) {
405                         dev = 0;
406                         sector += (conf->chunk_mask + 1);
407                 }
408         }
409         BUG_ON(slot != conf->copies);
410 }
411
412 static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
413 {
414         sector_t offset, chunk, vchunk;
415
416         offset = sector & conf->chunk_mask;
417         if (conf->far_offset) {
418                 int fc;
419                 chunk = sector >> conf->chunk_shift;
420                 fc = sector_div(chunk, conf->far_copies);
421                 dev -= fc * conf->near_copies;
422                 if (dev < 0)
423                         dev += conf->raid_disks;
424         } else {
425                 while (sector >= conf->stride) {
426                         sector -= conf->stride;
427                         if (dev < conf->near_copies)
428                                 dev += conf->raid_disks - conf->near_copies;
429                         else
430                                 dev -= conf->near_copies;
431                 }
432                 chunk = sector >> conf->chunk_shift;
433         }
434         vchunk = chunk * conf->raid_disks + dev;
435         sector_div(vchunk, conf->near_copies);
436         return (vchunk << conf->chunk_shift) + offset;
437 }
438
439 /**
440  *      raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
441  *      @q: request queue
442  *      @bio: the buffer head that's been built up so far
443  *      @biovec: the request that could be merged to it.
444  *
445  *      Return amount of bytes we can accept at this offset
446  *      If near_copies == raid_disk, there are no striping issues,
447  *      but in that case, the function isn't called at all.
448  */
449 static int raid10_mergeable_bvec(struct request_queue *q, struct bio *bio,
450                                 struct bio_vec *bio_vec)
451 {
452         mddev_t *mddev = q->queuedata;
453         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
454         int max;
455         unsigned int chunk_sectors = mddev->chunk_size >> 9;
456         unsigned int bio_sectors = bio->bi_size >> 9;
457
458         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
459         if (max < 0) max = 0; /* bio_add cannot handle a negative return */
460         if (max <= bio_vec->bv_len && bio_sectors == 0)
461                 return bio_vec->bv_len;
462         else
463                 return max;
464 }
465
466 /*
467  * This routine returns the disk from which the requested read should
468  * be done. There is a per-array 'next expected sequential IO' sector
469  * number - if this matches on the next IO then we use the last disk.
470  * There is also a per-disk 'last know head position' sector that is
471  * maintained from IRQ contexts, both the normal and the resync IO
472  * completion handlers update this position correctly. If there is no
473  * perfect sequential match then we pick the disk whose head is closest.
474  *
475  * If there are 2 mirrors in the same 2 devices, performance degrades
476  * because position is mirror, not device based.
477  *
478  * The rdev for the device selected will have nr_pending incremented.
479  */
480
481 /*
482  * FIXME: possibly should rethink readbalancing and do it differently
483  * depending on near_copies / far_copies geometry.
484  */
485 static int read_balance(conf_t *conf, r10bio_t *r10_bio)
486 {
487         const unsigned long this_sector = r10_bio->sector;
488         int disk, slot, nslot;
489         const int sectors = r10_bio->sectors;
490         sector_t new_distance, current_distance;
491         mdk_rdev_t *rdev;
492
493         raid10_find_phys(conf, r10_bio);
494         rcu_read_lock();
495         /*
496          * Check if we can balance. We can balance on the whole
497          * device if no resync is going on (recovery is ok), or below
498          * the resync window. We take the first readable disk when
499          * above the resync window.
500          */
501         if (conf->mddev->recovery_cp < MaxSector
502             && (this_sector + sectors >= conf->next_resync)) {
503                 /* make sure that disk is operational */
504                 slot = 0;
505                 disk = r10_bio->devs[slot].devnum;
506
507                 while ((rdev = rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
508                        r10_bio->devs[slot].bio == IO_BLOCKED ||
509                        !test_bit(In_sync, &rdev->flags)) {
510                         slot++;
511                         if (slot == conf->copies) {
512                                 slot = 0;
513                                 disk = -1;
514                                 break;
515                         }
516                         disk = r10_bio->devs[slot].devnum;
517                 }
518                 goto rb_out;
519         }
520
521
522         /* make sure the disk is operational */
523         slot = 0;
524         disk = r10_bio->devs[slot].devnum;
525         while ((rdev=rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
526                r10_bio->devs[slot].bio == IO_BLOCKED ||
527                !test_bit(In_sync, &rdev->flags)) {
528                 slot ++;
529                 if (slot == conf->copies) {
530                         disk = -1;
531                         goto rb_out;
532                 }
533                 disk = r10_bio->devs[slot].devnum;
534         }
535
536
537         current_distance = abs(r10_bio->devs[slot].addr -
538                                conf->mirrors[disk].head_position);
539
540         /* Find the disk whose head is closest,
541          * or - for far > 1 - find the closest to partition beginning */
542
543         for (nslot = slot; nslot < conf->copies; nslot++) {
544                 int ndisk = r10_bio->devs[nslot].devnum;
545
546
547                 if ((rdev=rcu_dereference(conf->mirrors[ndisk].rdev)) == NULL ||
548                     r10_bio->devs[nslot].bio == IO_BLOCKED ||
549                     !test_bit(In_sync, &rdev->flags))
550                         continue;
551
552                 /* This optimisation is debatable, and completely destroys
553                  * sequential read speed for 'far copies' arrays.  So only
554                  * keep it for 'near' arrays, and review those later.
555                  */
556                 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending)) {
557                         disk = ndisk;
558                         slot = nslot;
559                         break;
560                 }
561
562                 /* for far > 1 always use the lowest address */
563                 if (conf->far_copies > 1)
564                         new_distance = r10_bio->devs[nslot].addr;
565                 else
566                         new_distance = abs(r10_bio->devs[nslot].addr -
567                                            conf->mirrors[ndisk].head_position);
568                 if (new_distance < current_distance) {
569                         current_distance = new_distance;
570                         disk = ndisk;
571                         slot = nslot;
572                 }
573         }
574
575 rb_out:
576         r10_bio->read_slot = slot;
577 /*      conf->next_seq_sect = this_sector + sectors;*/
578
579         if (disk >= 0 && (rdev=rcu_dereference(conf->mirrors[disk].rdev))!= NULL)
580                 atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
581         else
582                 disk = -1;
583         rcu_read_unlock();
584
585         return disk;
586 }
587
588 static void unplug_slaves(mddev_t *mddev)
589 {
590         conf_t *conf = mddev_to_conf(mddev);
591         int i;
592
593         rcu_read_lock();
594         for (i=0; i<mddev->raid_disks; i++) {
595                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
596                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
597                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
598
599                         atomic_inc(&rdev->nr_pending);
600                         rcu_read_unlock();
601
602                         blk_unplug(r_queue);
603
604                         rdev_dec_pending(rdev, mddev);
605                         rcu_read_lock();
606                 }
607         }
608         rcu_read_unlock();
609 }
610
611 static void raid10_unplug(struct request_queue *q)
612 {
613         mddev_t *mddev = q->queuedata;
614
615         unplug_slaves(q->queuedata);
616         md_wakeup_thread(mddev->thread);
617 }
618
619 static int raid10_congested(void *data, int bits)
620 {
621         mddev_t *mddev = data;
622         conf_t *conf = mddev_to_conf(mddev);
623         int i, ret = 0;
624
625         rcu_read_lock();
626         for (i = 0; i < mddev->raid_disks && ret == 0; i++) {
627                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
628                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
629                         struct request_queue *q = bdev_get_queue(rdev->bdev);
630
631                         ret |= bdi_congested(&q->backing_dev_info, bits);
632                 }
633         }
634         rcu_read_unlock();
635         return ret;
636 }
637
638 static int flush_pending_writes(conf_t *conf)
639 {
640         /* Any writes that have been queued but are awaiting
641          * bitmap updates get flushed here.
642          * We return 1 if any requests were actually submitted.
643          */
644         int rv = 0;
645
646         spin_lock_irq(&conf->device_lock);
647
648         if (conf->pending_bio_list.head) {
649                 struct bio *bio;
650                 bio = bio_list_get(&conf->pending_bio_list);
651                 blk_remove_plug(conf->mddev->queue);
652                 spin_unlock_irq(&conf->device_lock);
653                 /* flush any pending bitmap writes to disk
654                  * before proceeding w/ I/O */
655                 bitmap_unplug(conf->mddev->bitmap);
656
657                 while (bio) { /* submit pending writes */
658                         struct bio *next = bio->bi_next;
659                         bio->bi_next = NULL;
660                         generic_make_request(bio);
661                         bio = next;
662                 }
663                 rv = 1;
664         } else
665                 spin_unlock_irq(&conf->device_lock);
666         return rv;
667 }
668 /* Barriers....
669  * Sometimes we need to suspend IO while we do something else,
670  * either some resync/recovery, or reconfigure the array.
671  * To do this we raise a 'barrier'.
672  * The 'barrier' is a counter that can be raised multiple times
673  * to count how many activities are happening which preclude
674  * normal IO.
675  * We can only raise the barrier if there is no pending IO.
676  * i.e. if nr_pending == 0.
677  * We choose only to raise the barrier if no-one is waiting for the
678  * barrier to go down.  This means that as soon as an IO request
679  * is ready, no other operations which require a barrier will start
680  * until the IO request has had a chance.
681  *
682  * So: regular IO calls 'wait_barrier'.  When that returns there
683  *    is no backgroup IO happening,  It must arrange to call
684  *    allow_barrier when it has finished its IO.
685  * backgroup IO calls must call raise_barrier.  Once that returns
686  *    there is no normal IO happeing.  It must arrange to call
687  *    lower_barrier when the particular background IO completes.
688  */
689 #define RESYNC_DEPTH 32
690
691 static void raise_barrier(conf_t *conf, int force)
692 {
693         BUG_ON(force && !conf->barrier);
694         spin_lock_irq(&conf->resync_lock);
695
696         /* Wait until no block IO is waiting (unless 'force') */
697         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
698                             conf->resync_lock,
699                             raid10_unplug(conf->mddev->queue));
700
701         /* block any new IO from starting */
702         conf->barrier++;
703
704         /* No wait for all pending IO to complete */
705         wait_event_lock_irq(conf->wait_barrier,
706                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
707                             conf->resync_lock,
708                             raid10_unplug(conf->mddev->queue));
709
710         spin_unlock_irq(&conf->resync_lock);
711 }
712
713 static void lower_barrier(conf_t *conf)
714 {
715         unsigned long flags;
716         spin_lock_irqsave(&conf->resync_lock, flags);
717         conf->barrier--;
718         spin_unlock_irqrestore(&conf->resync_lock, flags);
719         wake_up(&conf->wait_barrier);
720 }
721
722 static void wait_barrier(conf_t *conf)
723 {
724         spin_lock_irq(&conf->resync_lock);
725         if (conf->barrier) {
726                 conf->nr_waiting++;
727                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
728                                     conf->resync_lock,
729                                     raid10_unplug(conf->mddev->queue));
730                 conf->nr_waiting--;
731         }
732         conf->nr_pending++;
733         spin_unlock_irq(&conf->resync_lock);
734 }
735
736 static void allow_barrier(conf_t *conf)
737 {
738         unsigned long flags;
739         spin_lock_irqsave(&conf->resync_lock, flags);
740         conf->nr_pending--;
741         spin_unlock_irqrestore(&conf->resync_lock, flags);
742         wake_up(&conf->wait_barrier);
743 }
744
745 static void freeze_array(conf_t *conf)
746 {
747         /* stop syncio and normal IO and wait for everything to
748          * go quiet.
749          * We increment barrier and nr_waiting, and then
750          * wait until nr_pending match nr_queued+1
751          * This is called in the context of one normal IO request
752          * that has failed. Thus any sync request that might be pending
753          * will be blocked by nr_pending, and we need to wait for
754          * pending IO requests to complete or be queued for re-try.
755          * Thus the number queued (nr_queued) plus this request (1)
756          * must match the number of pending IOs (nr_pending) before
757          * we continue.
758          */
759         spin_lock_irq(&conf->resync_lock);
760         conf->barrier++;
761         conf->nr_waiting++;
762         wait_event_lock_irq(conf->wait_barrier,
763                             conf->nr_pending == conf->nr_queued+1,
764                             conf->resync_lock,
765                             ({ flush_pending_writes(conf);
766                                raid10_unplug(conf->mddev->queue); }));
767         spin_unlock_irq(&conf->resync_lock);
768 }
769
770 static void unfreeze_array(conf_t *conf)
771 {
772         /* reverse the effect of the freeze */
773         spin_lock_irq(&conf->resync_lock);
774         conf->barrier--;
775         conf->nr_waiting--;
776         wake_up(&conf->wait_barrier);
777         spin_unlock_irq(&conf->resync_lock);
778 }
779
780 static int make_request(struct request_queue *q, struct bio * bio)
781 {
782         mddev_t *mddev = q->queuedata;
783         conf_t *conf = mddev_to_conf(mddev);
784         mirror_info_t *mirror;
785         r10bio_t *r10_bio;
786         struct bio *read_bio;
787         int i;
788         int chunk_sects = conf->chunk_mask + 1;
789         const int rw = bio_data_dir(bio);
790         const int do_sync = bio_sync(bio);
791         struct bio_list bl;
792         unsigned long flags;
793         mdk_rdev_t *blocked_rdev;
794
795         if (unlikely(bio_barrier(bio))) {
796                 bio_endio(bio, -EOPNOTSUPP);
797                 return 0;
798         }
799
800         /* If this request crosses a chunk boundary, we need to
801          * split it.  This will only happen for 1 PAGE (or less) requests.
802          */
803         if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
804                       > chunk_sects &&
805                     conf->near_copies < conf->raid_disks)) {
806                 struct bio_pair *bp;
807                 /* Sanity check -- queue functions should prevent this happening */
808                 if (bio->bi_vcnt != 1 ||
809                     bio->bi_idx != 0)
810                         goto bad_map;
811                 /* This is a one page bio that upper layers
812                  * refuse to split for us, so we need to split it.
813                  */
814                 bp = bio_split(bio, bio_split_pool,
815                                chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
816                 if (make_request(q, &bp->bio1))
817                         generic_make_request(&bp->bio1);
818                 if (make_request(q, &bp->bio2))
819                         generic_make_request(&bp->bio2);
820
821                 bio_pair_release(bp);
822                 return 0;
823         bad_map:
824                 printk("raid10_make_request bug: can't convert block across chunks"
825                        " or bigger than %dk %llu %d\n", chunk_sects/2,
826                        (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
827
828                 bio_io_error(bio);
829                 return 0;
830         }
831
832         md_write_start(mddev, bio);
833
834         /*
835          * Register the new request and wait if the reconstruction
836          * thread has put up a bar for new requests.
837          * Continue immediately if no resync is active currently.
838          */
839         wait_barrier(conf);
840
841         disk_stat_inc(mddev->gendisk, ios[rw]);
842         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
843
844         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
845
846         r10_bio->master_bio = bio;
847         r10_bio->sectors = bio->bi_size >> 9;
848
849         r10_bio->mddev = mddev;
850         r10_bio->sector = bio->bi_sector;
851         r10_bio->state = 0;
852
853         if (rw == READ) {
854                 /*
855                  * read balancing logic:
856                  */
857                 int disk = read_balance(conf, r10_bio);
858                 int slot = r10_bio->read_slot;
859                 if (disk < 0) {
860                         raid_end_bio_io(r10_bio);
861                         return 0;
862                 }
863                 mirror = conf->mirrors + disk;
864
865                 read_bio = bio_clone(bio, GFP_NOIO);
866
867                 r10_bio->devs[slot].bio = read_bio;
868
869                 read_bio->bi_sector = r10_bio->devs[slot].addr +
870                         mirror->rdev->data_offset;
871                 read_bio->bi_bdev = mirror->rdev->bdev;
872                 read_bio->bi_end_io = raid10_end_read_request;
873                 read_bio->bi_rw = READ | do_sync;
874                 read_bio->bi_private = r10_bio;
875
876                 generic_make_request(read_bio);
877                 return 0;
878         }
879
880         /*
881          * WRITE:
882          */
883         /* first select target devices under rcu_lock and
884          * inc refcount on their rdev.  Record them by setting
885          * bios[x] to bio
886          */
887         raid10_find_phys(conf, r10_bio);
888  retry_write:
889         blocked_rdev = NULL;
890         rcu_read_lock();
891         for (i = 0;  i < conf->copies; i++) {
892                 int d = r10_bio->devs[i].devnum;
893                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
894                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
895                         atomic_inc(&rdev->nr_pending);
896                         blocked_rdev = rdev;
897                         break;
898                 }
899                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
900                         atomic_inc(&rdev->nr_pending);
901                         r10_bio->devs[i].bio = bio;
902                 } else {
903                         r10_bio->devs[i].bio = NULL;
904                         set_bit(R10BIO_Degraded, &r10_bio->state);
905                 }
906         }
907         rcu_read_unlock();
908
909         if (unlikely(blocked_rdev)) {
910                 /* Have to wait for this device to get unblocked, then retry */
911                 int j;
912                 int d;
913
914                 for (j = 0; j < i; j++)
915                         if (r10_bio->devs[j].bio) {
916                                 d = r10_bio->devs[j].devnum;
917                                 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
918                         }
919                 allow_barrier(conf);
920                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
921                 wait_barrier(conf);
922                 goto retry_write;
923         }
924
925         atomic_set(&r10_bio->remaining, 0);
926
927         bio_list_init(&bl);
928         for (i = 0; i < conf->copies; i++) {
929                 struct bio *mbio;
930                 int d = r10_bio->devs[i].devnum;
931                 if (!r10_bio->devs[i].bio)
932                         continue;
933
934                 mbio = bio_clone(bio, GFP_NOIO);
935                 r10_bio->devs[i].bio = mbio;
936
937                 mbio->bi_sector = r10_bio->devs[i].addr+
938                         conf->mirrors[d].rdev->data_offset;
939                 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
940                 mbio->bi_end_io = raid10_end_write_request;
941                 mbio->bi_rw = WRITE | do_sync;
942                 mbio->bi_private = r10_bio;
943
944                 atomic_inc(&r10_bio->remaining);
945                 bio_list_add(&bl, mbio);
946         }
947
948         if (unlikely(!atomic_read(&r10_bio->remaining))) {
949                 /* the array is dead */
950                 md_write_end(mddev);
951                 raid_end_bio_io(r10_bio);
952                 return 0;
953         }
954
955         bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
956         spin_lock_irqsave(&conf->device_lock, flags);
957         bio_list_merge(&conf->pending_bio_list, &bl);
958         blk_plug_device(mddev->queue);
959         spin_unlock_irqrestore(&conf->device_lock, flags);
960
961         /* In case raid10d snuck in to freeze_array */
962         wake_up(&conf->wait_barrier);
963
964         if (do_sync)
965                 md_wakeup_thread(mddev->thread);
966
967         return 0;
968 }
969
970 static void status(struct seq_file *seq, mddev_t *mddev)
971 {
972         conf_t *conf = mddev_to_conf(mddev);
973         int i;
974
975         if (conf->near_copies < conf->raid_disks)
976                 seq_printf(seq, " %dK chunks", mddev->chunk_size/1024);
977         if (conf->near_copies > 1)
978                 seq_printf(seq, " %d near-copies", conf->near_copies);
979         if (conf->far_copies > 1) {
980                 if (conf->far_offset)
981                         seq_printf(seq, " %d offset-copies", conf->far_copies);
982                 else
983                         seq_printf(seq, " %d far-copies", conf->far_copies);
984         }
985         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
986                                         conf->raid_disks - mddev->degraded);
987         for (i = 0; i < conf->raid_disks; i++)
988                 seq_printf(seq, "%s",
989                               conf->mirrors[i].rdev &&
990                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
991         seq_printf(seq, "]");
992 }
993
994 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
995 {
996         char b[BDEVNAME_SIZE];
997         conf_t *conf = mddev_to_conf(mddev);
998
999         /*
1000          * If it is not operational, then we have already marked it as dead
1001          * else if it is the last working disks, ignore the error, let the
1002          * next level up know.
1003          * else mark the drive as failed
1004          */
1005         if (test_bit(In_sync, &rdev->flags)
1006             && conf->raid_disks-mddev->degraded == 1)
1007                 /*
1008                  * Don't fail the drive, just return an IO error.
1009                  * The test should really be more sophisticated than
1010                  * "working_disks == 1", but it isn't critical, and
1011                  * can wait until we do more sophisticated "is the drive
1012                  * really dead" tests...
1013                  */
1014                 return;
1015         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1016                 unsigned long flags;
1017                 spin_lock_irqsave(&conf->device_lock, flags);
1018                 mddev->degraded++;
1019                 spin_unlock_irqrestore(&conf->device_lock, flags);
1020                 /*
1021                  * if recovery is running, make sure it aborts.
1022                  */
1023                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1024         }
1025         set_bit(Faulty, &rdev->flags);
1026         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1027         printk(KERN_ALERT "raid10: Disk failure on %s, disabling device.\n"
1028                 "raid10: Operation continuing on %d devices.\n",
1029                 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1030 }
1031
1032 static void print_conf(conf_t *conf)
1033 {
1034         int i;
1035         mirror_info_t *tmp;
1036
1037         printk("RAID10 conf printout:\n");
1038         if (!conf) {
1039                 printk("(!conf)\n");
1040                 return;
1041         }
1042         printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1043                 conf->raid_disks);
1044
1045         for (i = 0; i < conf->raid_disks; i++) {
1046                 char b[BDEVNAME_SIZE];
1047                 tmp = conf->mirrors + i;
1048                 if (tmp->rdev)
1049                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
1050                                 i, !test_bit(In_sync, &tmp->rdev->flags),
1051                                 !test_bit(Faulty, &tmp->rdev->flags),
1052                                 bdevname(tmp->rdev->bdev,b));
1053         }
1054 }
1055
1056 static void close_sync(conf_t *conf)
1057 {
1058         wait_barrier(conf);
1059         allow_barrier(conf);
1060
1061         mempool_destroy(conf->r10buf_pool);
1062         conf->r10buf_pool = NULL;
1063 }
1064
1065 /* check if there are enough drives for
1066  * every block to appear on atleast one
1067  */
1068 static int enough(conf_t *conf)
1069 {
1070         int first = 0;
1071
1072         do {
1073                 int n = conf->copies;
1074                 int cnt = 0;
1075                 while (n--) {
1076                         if (conf->mirrors[first].rdev)
1077                                 cnt++;
1078                         first = (first+1) % conf->raid_disks;
1079                 }
1080                 if (cnt == 0)
1081                         return 0;
1082         } while (first != 0);
1083         return 1;
1084 }
1085
1086 static int raid10_spare_active(mddev_t *mddev)
1087 {
1088         int i;
1089         conf_t *conf = mddev->private;
1090         mirror_info_t *tmp;
1091
1092         /*
1093          * Find all non-in_sync disks within the RAID10 configuration
1094          * and mark them in_sync
1095          */
1096         for (i = 0; i < conf->raid_disks; i++) {
1097                 tmp = conf->mirrors + i;
1098                 if (tmp->rdev
1099                     && !test_bit(Faulty, &tmp->rdev->flags)
1100                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1101                         unsigned long flags;
1102                         spin_lock_irqsave(&conf->device_lock, flags);
1103                         mddev->degraded--;
1104                         spin_unlock_irqrestore(&conf->device_lock, flags);
1105                 }
1106         }
1107
1108         print_conf(conf);
1109         return 0;
1110 }
1111
1112
1113 static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1114 {
1115         conf_t *conf = mddev->private;
1116         int found = 0;
1117         int mirror;
1118         mirror_info_t *p;
1119
1120         if (mddev->recovery_cp < MaxSector)
1121                 /* only hot-add to in-sync arrays, as recovery is
1122                  * very different from resync
1123                  */
1124                 return 0;
1125         if (!enough(conf))
1126                 return 0;
1127
1128         if (rdev->saved_raid_disk >= 0 &&
1129             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1130                 mirror = rdev->saved_raid_disk;
1131         else
1132                 mirror = 0;
1133         for ( ; mirror < mddev->raid_disks; mirror++)
1134                 if ( !(p=conf->mirrors+mirror)->rdev) {
1135
1136                         blk_queue_stack_limits(mddev->queue,
1137                                                rdev->bdev->bd_disk->queue);
1138                         /* as we don't honour merge_bvec_fn, we must never risk
1139                          * violating it, so limit ->max_sector to one PAGE, as
1140                          * a one page request is never in violation.
1141                          */
1142                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1143                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
1144                                 mddev->queue->max_sectors = (PAGE_SIZE>>9);
1145
1146                         p->head_position = 0;
1147                         rdev->raid_disk = mirror;
1148                         found = 1;
1149                         if (rdev->saved_raid_disk != mirror)
1150                                 conf->fullsync = 1;
1151                         rcu_assign_pointer(p->rdev, rdev);
1152                         break;
1153                 }
1154
1155         print_conf(conf);
1156         return found;
1157 }
1158
1159 static int raid10_remove_disk(mddev_t *mddev, int number)
1160 {
1161         conf_t *conf = mddev->private;
1162         int err = 0;
1163         mdk_rdev_t *rdev;
1164         mirror_info_t *p = conf->mirrors+ number;
1165
1166         print_conf(conf);
1167         rdev = p->rdev;
1168         if (rdev) {
1169                 if (test_bit(In_sync, &rdev->flags) ||
1170                     atomic_read(&rdev->nr_pending)) {
1171                         err = -EBUSY;
1172                         goto abort;
1173                 }
1174                 /* Only remove faulty devices in recovery
1175                  * is not possible.
1176                  */
1177                 if (!test_bit(Faulty, &rdev->flags) &&
1178                     enough(conf)) {
1179                         err = -EBUSY;
1180                         goto abort;
1181                 }
1182                 p->rdev = NULL;
1183                 synchronize_rcu();
1184                 if (atomic_read(&rdev->nr_pending)) {
1185                         /* lost the race, try later */
1186                         err = -EBUSY;
1187                         p->rdev = rdev;
1188                 }
1189         }
1190 abort:
1191
1192         print_conf(conf);
1193         return err;
1194 }
1195
1196
1197 static void end_sync_read(struct bio *bio, int error)
1198 {
1199         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1200         conf_t *conf = mddev_to_conf(r10_bio->mddev);
1201         int i,d;
1202
1203         for (i=0; i<conf->copies; i++)
1204                 if (r10_bio->devs[i].bio == bio)
1205                         break;
1206         BUG_ON(i == conf->copies);
1207         update_head_pos(i, r10_bio);
1208         d = r10_bio->devs[i].devnum;
1209
1210         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1211                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1212         else {
1213                 atomic_add(r10_bio->sectors,
1214                            &conf->mirrors[d].rdev->corrected_errors);
1215                 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1216                         md_error(r10_bio->mddev,
1217                                  conf->mirrors[d].rdev);
1218         }
1219
1220         /* for reconstruct, we always reschedule after a read.
1221          * for resync, only after all reads
1222          */
1223         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1224             atomic_dec_and_test(&r10_bio->remaining)) {
1225                 /* we have read all the blocks,
1226                  * do the comparison in process context in raid10d
1227                  */
1228                 reschedule_retry(r10_bio);
1229         }
1230         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1231 }
1232
1233 static void end_sync_write(struct bio *bio, int error)
1234 {
1235         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1236         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1237         mddev_t *mddev = r10_bio->mddev;
1238         conf_t *conf = mddev_to_conf(mddev);
1239         int i,d;
1240
1241         for (i = 0; i < conf->copies; i++)
1242                 if (r10_bio->devs[i].bio == bio)
1243                         break;
1244         d = r10_bio->devs[i].devnum;
1245
1246         if (!uptodate)
1247                 md_error(mddev, conf->mirrors[d].rdev);
1248
1249         update_head_pos(i, r10_bio);
1250
1251         while (atomic_dec_and_test(&r10_bio->remaining)) {
1252                 if (r10_bio->master_bio == NULL) {
1253                         /* the primary of several recovery bios */
1254                         md_done_sync(mddev, r10_bio->sectors, 1);
1255                         put_buf(r10_bio);
1256                         break;
1257                 } else {
1258                         r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1259                         put_buf(r10_bio);
1260                         r10_bio = r10_bio2;
1261                 }
1262         }
1263         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1264 }
1265
1266 /*
1267  * Note: sync and recover and handled very differently for raid10
1268  * This code is for resync.
1269  * For resync, we read through virtual addresses and read all blocks.
1270  * If there is any error, we schedule a write.  The lowest numbered
1271  * drive is authoritative.
1272  * However requests come for physical address, so we need to map.
1273  * For every physical address there are raid_disks/copies virtual addresses,
1274  * which is always are least one, but is not necessarly an integer.
1275  * This means that a physical address can span multiple chunks, so we may
1276  * have to submit multiple io requests for a single sync request.
1277  */
1278 /*
1279  * We check if all blocks are in-sync and only write to blocks that
1280  * aren't in sync
1281  */
1282 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1283 {
1284         conf_t *conf = mddev_to_conf(mddev);
1285         int i, first;
1286         struct bio *tbio, *fbio;
1287
1288         atomic_set(&r10_bio->remaining, 1);
1289
1290         /* find the first device with a block */
1291         for (i=0; i<conf->copies; i++)
1292                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1293                         break;
1294
1295         if (i == conf->copies)
1296                 goto done;
1297
1298         first = i;
1299         fbio = r10_bio->devs[i].bio;
1300
1301         /* now find blocks with errors */
1302         for (i=0 ; i < conf->copies ; i++) {
1303                 int  j, d;
1304                 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1305
1306                 tbio = r10_bio->devs[i].bio;
1307
1308                 if (tbio->bi_end_io != end_sync_read)
1309                         continue;
1310                 if (i == first)
1311                         continue;
1312                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1313                         /* We know that the bi_io_vec layout is the same for
1314                          * both 'first' and 'i', so we just compare them.
1315                          * All vec entries are PAGE_SIZE;
1316                          */
1317                         for (j = 0; j < vcnt; j++)
1318                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1319                                            page_address(tbio->bi_io_vec[j].bv_page),
1320                                            PAGE_SIZE))
1321                                         break;
1322                         if (j == vcnt)
1323                                 continue;
1324                         mddev->resync_mismatches += r10_bio->sectors;
1325                 }
1326                 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1327                         /* Don't fix anything. */
1328                         continue;
1329                 /* Ok, we need to write this bio
1330                  * First we need to fixup bv_offset, bv_len and
1331                  * bi_vecs, as the read request might have corrupted these
1332                  */
1333                 tbio->bi_vcnt = vcnt;
1334                 tbio->bi_size = r10_bio->sectors << 9;
1335                 tbio->bi_idx = 0;
1336                 tbio->bi_phys_segments = 0;
1337                 tbio->bi_hw_segments = 0;
1338                 tbio->bi_hw_front_size = 0;
1339                 tbio->bi_hw_back_size = 0;
1340                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1341                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1342                 tbio->bi_next = NULL;
1343                 tbio->bi_rw = WRITE;
1344                 tbio->bi_private = r10_bio;
1345                 tbio->bi_sector = r10_bio->devs[i].addr;
1346
1347                 for (j=0; j < vcnt ; j++) {
1348                         tbio->bi_io_vec[j].bv_offset = 0;
1349                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1350
1351                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1352                                page_address(fbio->bi_io_vec[j].bv_page),
1353                                PAGE_SIZE);
1354                 }
1355                 tbio->bi_end_io = end_sync_write;
1356
1357                 d = r10_bio->devs[i].devnum;
1358                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1359                 atomic_inc(&r10_bio->remaining);
1360                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1361
1362                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1363                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1364                 generic_make_request(tbio);
1365         }
1366
1367 done:
1368         if (atomic_dec_and_test(&r10_bio->remaining)) {
1369                 md_done_sync(mddev, r10_bio->sectors, 1);
1370                 put_buf(r10_bio);
1371         }
1372 }
1373
1374 /*
1375  * Now for the recovery code.
1376  * Recovery happens across physical sectors.
1377  * We recover all non-is_sync drives by finding the virtual address of
1378  * each, and then choose a working drive that also has that virt address.
1379  * There is a separate r10_bio for each non-in_sync drive.
1380  * Only the first two slots are in use. The first for reading,
1381  * The second for writing.
1382  *
1383  */
1384
1385 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1386 {
1387         conf_t *conf = mddev_to_conf(mddev);
1388         int i, d;
1389         struct bio *bio, *wbio;
1390
1391
1392         /* move the pages across to the second bio
1393          * and submit the write request
1394          */
1395         bio = r10_bio->devs[0].bio;
1396         wbio = r10_bio->devs[1].bio;
1397         for (i=0; i < wbio->bi_vcnt; i++) {
1398                 struct page *p = bio->bi_io_vec[i].bv_page;
1399                 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1400                 wbio->bi_io_vec[i].bv_page = p;
1401         }
1402         d = r10_bio->devs[1].devnum;
1403
1404         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1405         md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1406         if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1407                 generic_make_request(wbio);
1408         else
1409                 bio_endio(wbio, -EIO);
1410 }
1411
1412
1413 /*
1414  * This is a kernel thread which:
1415  *
1416  *      1.      Retries failed read operations on working mirrors.
1417  *      2.      Updates the raid superblock when problems encounter.
1418  *      3.      Performs writes following reads for array synchronising.
1419  */
1420
1421 static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1422 {
1423         int sect = 0; /* Offset from r10_bio->sector */
1424         int sectors = r10_bio->sectors;
1425         mdk_rdev_t*rdev;
1426         while(sectors) {
1427                 int s = sectors;
1428                 int sl = r10_bio->read_slot;
1429                 int success = 0;
1430                 int start;
1431
1432                 if (s > (PAGE_SIZE>>9))
1433                         s = PAGE_SIZE >> 9;
1434
1435                 rcu_read_lock();
1436                 do {
1437                         int d = r10_bio->devs[sl].devnum;
1438                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1439                         if (rdev &&
1440                             test_bit(In_sync, &rdev->flags)) {
1441                                 atomic_inc(&rdev->nr_pending);
1442                                 rcu_read_unlock();
1443                                 success = sync_page_io(rdev->bdev,
1444                                                        r10_bio->devs[sl].addr +
1445                                                        sect + rdev->data_offset,
1446                                                        s<<9,
1447                                                        conf->tmppage, READ);
1448                                 rdev_dec_pending(rdev, mddev);
1449                                 rcu_read_lock();
1450                                 if (success)
1451                                         break;
1452                         }
1453                         sl++;
1454                         if (sl == conf->copies)
1455                                 sl = 0;
1456                 } while (!success && sl != r10_bio->read_slot);
1457                 rcu_read_unlock();
1458
1459                 if (!success) {
1460                         /* Cannot read from anywhere -- bye bye array */
1461                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1462                         md_error(mddev, conf->mirrors[dn].rdev);
1463                         break;
1464                 }
1465
1466                 start = sl;
1467                 /* write it back and re-read */
1468                 rcu_read_lock();
1469                 while (sl != r10_bio->read_slot) {
1470                         int d;
1471                         if (sl==0)
1472                                 sl = conf->copies;
1473                         sl--;
1474                         d = r10_bio->devs[sl].devnum;
1475                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1476                         if (rdev &&
1477                             test_bit(In_sync, &rdev->flags)) {
1478                                 atomic_inc(&rdev->nr_pending);
1479                                 rcu_read_unlock();
1480                                 atomic_add(s, &rdev->corrected_errors);
1481                                 if (sync_page_io(rdev->bdev,
1482                                                  r10_bio->devs[sl].addr +
1483                                                  sect + rdev->data_offset,
1484                                                  s<<9, conf->tmppage, WRITE)
1485                                     == 0)
1486                                         /* Well, this device is dead */
1487                                         md_error(mddev, rdev);
1488                                 rdev_dec_pending(rdev, mddev);
1489                                 rcu_read_lock();
1490                         }
1491                 }
1492                 sl = start;
1493                 while (sl != r10_bio->read_slot) {
1494                         int d;
1495                         if (sl==0)
1496                                 sl = conf->copies;
1497                         sl--;
1498                         d = r10_bio->devs[sl].devnum;
1499                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1500                         if (rdev &&
1501                             test_bit(In_sync, &rdev->flags)) {
1502                                 char b[BDEVNAME_SIZE];
1503                                 atomic_inc(&rdev->nr_pending);
1504                                 rcu_read_unlock();
1505                                 if (sync_page_io(rdev->bdev,
1506                                                  r10_bio->devs[sl].addr +
1507                                                  sect + rdev->data_offset,
1508                                                  s<<9, conf->tmppage, READ) == 0)
1509                                         /* Well, this device is dead */
1510                                         md_error(mddev, rdev);
1511                                 else
1512                                         printk(KERN_INFO
1513                                                "raid10:%s: read error corrected"
1514                                                " (%d sectors at %llu on %s)\n",
1515                                                mdname(mddev), s,
1516                                                (unsigned long long)(sect+
1517                                                     rdev->data_offset),
1518                                                bdevname(rdev->bdev, b));
1519
1520                                 rdev_dec_pending(rdev, mddev);
1521                                 rcu_read_lock();
1522                         }
1523                 }
1524                 rcu_read_unlock();
1525
1526                 sectors -= s;
1527                 sect += s;
1528         }
1529 }
1530
1531 static void raid10d(mddev_t *mddev)
1532 {
1533         r10bio_t *r10_bio;
1534         struct bio *bio;
1535         unsigned long flags;
1536         conf_t *conf = mddev_to_conf(mddev);
1537         struct list_head *head = &conf->retry_list;
1538         int unplug=0;
1539         mdk_rdev_t *rdev;
1540
1541         md_check_recovery(mddev);
1542
1543         for (;;) {
1544                 char b[BDEVNAME_SIZE];
1545
1546                 unplug += flush_pending_writes(conf);
1547
1548                 spin_lock_irqsave(&conf->device_lock, flags);
1549                 if (list_empty(head)) {
1550                         spin_unlock_irqrestore(&conf->device_lock, flags);
1551                         break;
1552                 }
1553                 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1554                 list_del(head->prev);
1555                 conf->nr_queued--;
1556                 spin_unlock_irqrestore(&conf->device_lock, flags);
1557
1558                 mddev = r10_bio->mddev;
1559                 conf = mddev_to_conf(mddev);
1560                 if (test_bit(R10BIO_IsSync, &r10_bio->state)) {
1561                         sync_request_write(mddev, r10_bio);
1562                         unplug = 1;
1563                 } else  if (test_bit(R10BIO_IsRecover, &r10_bio->state)) {
1564                         recovery_request_write(mddev, r10_bio);
1565                         unplug = 1;
1566                 } else {
1567                         int mirror;
1568                         /* we got a read error. Maybe the drive is bad.  Maybe just
1569                          * the block and we can fix it.
1570                          * We freeze all other IO, and try reading the block from
1571                          * other devices.  When we find one, we re-write
1572                          * and check it that fixes the read error.
1573                          * This is all done synchronously while the array is
1574                          * frozen.
1575                          */
1576                         if (mddev->ro == 0) {
1577                                 freeze_array(conf);
1578                                 fix_read_error(conf, mddev, r10_bio);
1579                                 unfreeze_array(conf);
1580                         }
1581
1582                         bio = r10_bio->devs[r10_bio->read_slot].bio;
1583                         r10_bio->devs[r10_bio->read_slot].bio =
1584                                 mddev->ro ? IO_BLOCKED : NULL;
1585                         mirror = read_balance(conf, r10_bio);
1586                         if (mirror == -1) {
1587                                 printk(KERN_ALERT "raid10: %s: unrecoverable I/O"
1588                                        " read error for block %llu\n",
1589                                        bdevname(bio->bi_bdev,b),
1590                                        (unsigned long long)r10_bio->sector);
1591                                 raid_end_bio_io(r10_bio);
1592                                 bio_put(bio);
1593                         } else {
1594                                 const int do_sync = bio_sync(r10_bio->master_bio);
1595                                 bio_put(bio);
1596                                 rdev = conf->mirrors[mirror].rdev;
1597                                 if (printk_ratelimit())
1598                                         printk(KERN_ERR "raid10: %s: redirecting sector %llu to"
1599                                                " another mirror\n",
1600                                                bdevname(rdev->bdev,b),
1601                                                (unsigned long long)r10_bio->sector);
1602                                 bio = bio_clone(r10_bio->master_bio, GFP_NOIO);
1603                                 r10_bio->devs[r10_bio->read_slot].bio = bio;
1604                                 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1605                                         + rdev->data_offset;
1606                                 bio->bi_bdev = rdev->bdev;
1607                                 bio->bi_rw = READ | do_sync;
1608                                 bio->bi_private = r10_bio;
1609                                 bio->bi_end_io = raid10_end_read_request;
1610                                 unplug = 1;
1611                                 generic_make_request(bio);
1612                         }
1613                 }
1614         }
1615         if (unplug)
1616                 unplug_slaves(mddev);
1617 }
1618
1619
1620 static int init_resync(conf_t *conf)
1621 {
1622         int buffs;
1623
1624         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1625         BUG_ON(conf->r10buf_pool);
1626         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1627         if (!conf->r10buf_pool)
1628                 return -ENOMEM;
1629         conf->next_resync = 0;
1630         return 0;
1631 }
1632
1633 /*
1634  * perform a "sync" on one "block"
1635  *
1636  * We need to make sure that no normal I/O request - particularly write
1637  * requests - conflict with active sync requests.
1638  *
1639  * This is achieved by tracking pending requests and a 'barrier' concept
1640  * that can be installed to exclude normal IO requests.
1641  *
1642  * Resync and recovery are handled very differently.
1643  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1644  *
1645  * For resync, we iterate over virtual addresses, read all copies,
1646  * and update if there are differences.  If only one copy is live,
1647  * skip it.
1648  * For recovery, we iterate over physical addresses, read a good
1649  * value for each non-in_sync drive, and over-write.
1650  *
1651  * So, for recovery we may have several outstanding complex requests for a
1652  * given address, one for each out-of-sync device.  We model this by allocating
1653  * a number of r10_bio structures, one for each out-of-sync device.
1654  * As we setup these structures, we collect all bio's together into a list
1655  * which we then process collectively to add pages, and then process again
1656  * to pass to generic_make_request.
1657  *
1658  * The r10_bio structures are linked using a borrowed master_bio pointer.
1659  * This link is counted in ->remaining.  When the r10_bio that points to NULL
1660  * has its remaining count decremented to 0, the whole complex operation
1661  * is complete.
1662  *
1663  */
1664
1665 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1666 {
1667         conf_t *conf = mddev_to_conf(mddev);
1668         r10bio_t *r10_bio;
1669         struct bio *biolist = NULL, *bio;
1670         sector_t max_sector, nr_sectors;
1671         int disk;
1672         int i;
1673         int max_sync;
1674         int sync_blocks;
1675
1676         sector_t sectors_skipped = 0;
1677         int chunks_skipped = 0;
1678
1679         if (!conf->r10buf_pool)
1680                 if (init_resync(conf))
1681                         return 0;
1682
1683  skipped:
1684         max_sector = mddev->size << 1;
1685         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1686                 max_sector = mddev->resync_max_sectors;
1687         if (sector_nr >= max_sector) {
1688                 /* If we aborted, we need to abort the
1689                  * sync on the 'current' bitmap chucks (there can
1690                  * be several when recovering multiple devices).
1691                  * as we may have started syncing it but not finished.
1692                  * We can find the current address in
1693                  * mddev->curr_resync, but for recovery,
1694                  * we need to convert that to several
1695                  * virtual addresses.
1696                  */
1697                 if (mddev->curr_resync < max_sector) { /* aborted */
1698                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1699                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1700                                                 &sync_blocks, 1);
1701                         else for (i=0; i<conf->raid_disks; i++) {
1702                                 sector_t sect =
1703                                         raid10_find_virt(conf, mddev->curr_resync, i);
1704                                 bitmap_end_sync(mddev->bitmap, sect,
1705                                                 &sync_blocks, 1);
1706                         }
1707                 } else /* completed sync */
1708                         conf->fullsync = 0;
1709
1710                 bitmap_close_sync(mddev->bitmap);
1711                 close_sync(conf);
1712                 *skipped = 1;
1713                 return sectors_skipped;
1714         }
1715         if (chunks_skipped >= conf->raid_disks) {
1716                 /* if there has been nothing to do on any drive,
1717                  * then there is nothing to do at all..
1718                  */
1719                 *skipped = 1;
1720                 return (max_sector - sector_nr) + sectors_skipped;
1721         }
1722
1723         if (max_sector > mddev->resync_max)
1724                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1725
1726         /* make sure whole request will fit in a chunk - if chunks
1727          * are meaningful
1728          */
1729         if (conf->near_copies < conf->raid_disks &&
1730             max_sector > (sector_nr | conf->chunk_mask))
1731                 max_sector = (sector_nr | conf->chunk_mask) + 1;
1732         /*
1733          * If there is non-resync activity waiting for us then
1734          * put in a delay to throttle resync.
1735          */
1736         if (!go_faster && conf->nr_waiting)
1737                 msleep_interruptible(1000);
1738
1739         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1740
1741         /* Again, very different code for resync and recovery.
1742          * Both must result in an r10bio with a list of bios that
1743          * have bi_end_io, bi_sector, bi_bdev set,
1744          * and bi_private set to the r10bio.
1745          * For recovery, we may actually create several r10bios
1746          * with 2 bios in each, that correspond to the bios in the main one.
1747          * In this case, the subordinate r10bios link back through a
1748          * borrowed master_bio pointer, and the counter in the master
1749          * includes a ref from each subordinate.
1750          */
1751         /* First, we decide what to do and set ->bi_end_io
1752          * To end_sync_read if we want to read, and
1753          * end_sync_write if we will want to write.
1754          */
1755
1756         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
1757         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1758                 /* recovery... the complicated one */
1759                 int i, j, k;
1760                 r10_bio = NULL;
1761
1762                 for (i=0 ; i<conf->raid_disks; i++)
1763                         if (conf->mirrors[i].rdev &&
1764                             !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) {
1765                                 int still_degraded = 0;
1766                                 /* want to reconstruct this device */
1767                                 r10bio_t *rb2 = r10_bio;
1768                                 sector_t sect = raid10_find_virt(conf, sector_nr, i);
1769                                 int must_sync;
1770                                 /* Unless we are doing a full sync, we only need
1771                                  * to recover the block if it is set in the bitmap
1772                                  */
1773                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1774                                                               &sync_blocks, 1);
1775                                 if (sync_blocks < max_sync)
1776                                         max_sync = sync_blocks;
1777                                 if (!must_sync &&
1778                                     !conf->fullsync) {
1779                                         /* yep, skip the sync_blocks here, but don't assume
1780                                          * that there will never be anything to do here
1781                                          */
1782                                         chunks_skipped = -1;
1783                                         continue;
1784                                 }
1785
1786                                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1787                                 raise_barrier(conf, rb2 != NULL);
1788                                 atomic_set(&r10_bio->remaining, 0);
1789
1790                                 r10_bio->master_bio = (struct bio*)rb2;
1791                                 if (rb2)
1792                                         atomic_inc(&rb2->remaining);
1793                                 r10_bio->mddev = mddev;
1794                                 set_bit(R10BIO_IsRecover, &r10_bio->state);
1795                                 r10_bio->sector = sect;
1796
1797                                 raid10_find_phys(conf, r10_bio);
1798                                 /* Need to check if this section will still be
1799                                  * degraded
1800                                  */
1801                                 for (j=0; j<conf->copies;j++) {
1802                                         int d = r10_bio->devs[j].devnum;
1803                                         if (conf->mirrors[d].rdev == NULL ||
1804                                             test_bit(Faulty, &conf->mirrors[d].rdev->flags)) {
1805                                                 still_degraded = 1;
1806                                                 break;
1807                                         }
1808                                 }
1809                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1810                                                               &sync_blocks, still_degraded);
1811
1812                                 for (j=0; j<conf->copies;j++) {
1813                                         int d = r10_bio->devs[j].devnum;
1814                                         if (conf->mirrors[d].rdev &&
1815                                             test_bit(In_sync, &conf->mirrors[d].rdev->flags)) {
1816                                                 /* This is where we read from */
1817                                                 bio = r10_bio->devs[0].bio;
1818                                                 bio->bi_next = biolist;
1819                                                 biolist = bio;
1820                                                 bio->bi_private = r10_bio;
1821                                                 bio->bi_end_io = end_sync_read;
1822                                                 bio->bi_rw = READ;
1823                                                 bio->bi_sector = r10_bio->devs[j].addr +
1824                                                         conf->mirrors[d].rdev->data_offset;
1825                                                 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1826                                                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1827                                                 atomic_inc(&r10_bio->remaining);
1828                                                 /* and we write to 'i' */
1829
1830                                                 for (k=0; k<conf->copies; k++)
1831                                                         if (r10_bio->devs[k].devnum == i)
1832                                                                 break;
1833                                                 BUG_ON(k == conf->copies);
1834                                                 bio = r10_bio->devs[1].bio;
1835                                                 bio->bi_next = biolist;
1836                                                 biolist = bio;
1837                                                 bio->bi_private = r10_bio;
1838                                                 bio->bi_end_io = end_sync_write;
1839                                                 bio->bi_rw = WRITE;
1840                                                 bio->bi_sector = r10_bio->devs[k].addr +
1841                                                         conf->mirrors[i].rdev->data_offset;
1842                                                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1843
1844                                                 r10_bio->devs[0].devnum = d;
1845                                                 r10_bio->devs[1].devnum = i;
1846
1847                                                 break;
1848                                         }
1849                                 }
1850                                 if (j == conf->copies) {
1851                                         /* Cannot recover, so abort the recovery */
1852                                         put_buf(r10_bio);
1853                                         if (rb2)
1854                                                 atomic_dec(&rb2->remaining);
1855                                         r10_bio = rb2;
1856                                         if (!test_and_set_bit(MD_RECOVERY_INTR,
1857                                                               &mddev->recovery))
1858                                                 printk(KERN_INFO "raid10: %s: insufficient working devices for recovery.\n",
1859                                                        mdname(mddev));
1860                                         break;
1861                                 }
1862                         }
1863                 if (biolist == NULL) {
1864                         while (r10_bio) {
1865                                 r10bio_t *rb2 = r10_bio;
1866                                 r10_bio = (r10bio_t*) rb2->master_bio;
1867                                 rb2->master_bio = NULL;
1868                                 put_buf(rb2);
1869                         }
1870                         goto giveup;
1871                 }
1872         } else {
1873                 /* resync. Schedule a read for every block at this virt offset */
1874                 int count = 0;
1875
1876                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1877                                        &sync_blocks, mddev->degraded) &&
1878                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1879                         /* We can skip this block */
1880                         *skipped = 1;
1881                         return sync_blocks + sectors_skipped;
1882                 }
1883                 if (sync_blocks < max_sync)
1884                         max_sync = sync_blocks;
1885                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1886
1887                 r10_bio->mddev = mddev;
1888                 atomic_set(&r10_bio->remaining, 0);
1889                 raise_barrier(conf, 0);
1890                 conf->next_resync = sector_nr;
1891
1892                 r10_bio->master_bio = NULL;
1893                 r10_bio->sector = sector_nr;
1894                 set_bit(R10BIO_IsSync, &r10_bio->state);
1895                 raid10_find_phys(conf, r10_bio);
1896                 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1897
1898                 for (i=0; i<conf->copies; i++) {
1899                         int d = r10_bio->devs[i].devnum;
1900                         bio = r10_bio->devs[i].bio;
1901                         bio->bi_end_io = NULL;
1902                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
1903                         if (conf->mirrors[d].rdev == NULL ||
1904                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
1905                                 continue;
1906                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1907                         atomic_inc(&r10_bio->remaining);
1908                         bio->bi_next = biolist;
1909                         biolist = bio;
1910                         bio->bi_private = r10_bio;
1911                         bio->bi_end_io = end_sync_read;
1912                         bio->bi_rw = READ;
1913                         bio->bi_sector = r10_bio->devs[i].addr +
1914                                 conf->mirrors[d].rdev->data_offset;
1915                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1916                         count++;
1917                 }
1918
1919                 if (count < 2) {
1920                         for (i=0; i<conf->copies; i++) {
1921                                 int d = r10_bio->devs[i].devnum;
1922                                 if (r10_bio->devs[i].bio->bi_end_io)
1923                                         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1924                         }
1925                         put_buf(r10_bio);
1926                         biolist = NULL;
1927                         goto giveup;
1928                 }
1929         }
1930
1931         for (bio = biolist; bio ; bio=bio->bi_next) {
1932
1933                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1934                 if (bio->bi_end_io)
1935                         bio->bi_flags |= 1 << BIO_UPTODATE;
1936                 bio->bi_vcnt = 0;
1937                 bio->bi_idx = 0;
1938                 bio->bi_phys_segments = 0;
1939                 bio->bi_hw_segments = 0;
1940                 bio->bi_size = 0;
1941         }
1942
1943         nr_sectors = 0;
1944         if (sector_nr + max_sync < max_sector)
1945                 max_sector = sector_nr + max_sync;
1946         do {
1947                 struct page *page;
1948                 int len = PAGE_SIZE;
1949                 disk = 0;
1950                 if (sector_nr + (len>>9) > max_sector)
1951                         len = (max_sector - sector_nr) << 9;
1952                 if (len == 0)
1953                         break;
1954                 for (bio= biolist ; bio ; bio=bio->bi_next) {
1955                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1956                         if (bio_add_page(bio, page, len, 0) == 0) {
1957                                 /* stop here */
1958                                 struct bio *bio2;
1959                                 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1960                                 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
1961                                         /* remove last page from this bio */
1962                                         bio2->bi_vcnt--;
1963                                         bio2->bi_size -= len;
1964                                         bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
1965                                 }
1966                                 goto bio_full;
1967                         }
1968                         disk = i;
1969                 }
1970                 nr_sectors += len>>9;
1971                 sector_nr += len>>9;
1972         } while (biolist->bi_vcnt < RESYNC_PAGES);
1973  bio_full:
1974         r10_bio->sectors = nr_sectors;
1975
1976         while (biolist) {
1977                 bio = biolist;
1978                 biolist = biolist->bi_next;
1979
1980                 bio->bi_next = NULL;
1981                 r10_bio = bio->bi_private;
1982                 r10_bio->sectors = nr_sectors;
1983
1984                 if (bio->bi_end_io == end_sync_read) {
1985                         md_sync_acct(bio->bi_bdev, nr_sectors);
1986                         generic_make_request(bio);
1987                 }
1988         }
1989
1990         if (sectors_skipped)
1991                 /* pretend they weren't skipped, it makes
1992                  * no important difference in this case
1993                  */
1994                 md_done_sync(mddev, sectors_skipped, 1);
1995
1996         return sectors_skipped + nr_sectors;
1997  giveup:
1998         /* There is nowhere to write, so all non-sync
1999          * drives must be failed, so try the next chunk...
2000          */
2001         {
2002         sector_t sec = max_sector - sector_nr;
2003         sectors_skipped += sec;
2004         chunks_skipped ++;
2005         sector_nr = max_sector;
2006         goto skipped;
2007         }
2008 }
2009
2010 static int run(mddev_t *mddev)
2011 {
2012         conf_t *conf;
2013         int i, disk_idx;
2014         mirror_info_t *disk;
2015         mdk_rdev_t *rdev;
2016         struct list_head *tmp;
2017         int nc, fc, fo;
2018         sector_t stride, size;
2019
2020         if (mddev->chunk_size == 0) {
2021                 printk(KERN_ERR "md/raid10: non-zero chunk size required.\n");
2022                 return -EINVAL;
2023         }
2024
2025         nc = mddev->layout & 255;
2026         fc = (mddev->layout >> 8) & 255;
2027         fo = mddev->layout & (1<<16);
2028         if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
2029             (mddev->layout >> 17)) {
2030                 printk(KERN_ERR "raid10: %s: unsupported raid10 layout: 0x%8x\n",
2031                        mdname(mddev), mddev->layout);
2032                 goto out;
2033         }
2034         /*
2035          * copy the already verified devices into our private RAID10
2036          * bookkeeping area. [whatever we allocate in run(),
2037          * should be freed in stop()]
2038          */
2039         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2040         mddev->private = conf;
2041         if (!conf) {
2042                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2043                         mdname(mddev));
2044                 goto out;
2045         }
2046         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2047                                  GFP_KERNEL);
2048         if (!conf->mirrors) {
2049                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2050                        mdname(mddev));
2051                 goto out_free_conf;
2052         }
2053
2054         conf->tmppage = alloc_page(GFP_KERNEL);
2055         if (!conf->tmppage)
2056                 goto out_free_conf;
2057
2058         conf->mddev = mddev;
2059         conf->raid_disks = mddev->raid_disks;
2060         conf->near_copies = nc;
2061         conf->far_copies = fc;
2062         conf->copies = nc*fc;
2063         conf->far_offset = fo;
2064         conf->chunk_mask = (sector_t)(mddev->chunk_size>>9)-1;
2065         conf->chunk_shift = ffz(~mddev->chunk_size) - 9;
2066         size = mddev->size >> (conf->chunk_shift-1);
2067         sector_div(size, fc);
2068         size = size * conf->raid_disks;
2069         sector_div(size, nc);
2070         /* 'size' is now the number of chunks in the array */
2071         /* calculate "used chunks per device" in 'stride' */
2072         stride = size * conf->copies;
2073
2074         /* We need to round up when dividing by raid_disks to
2075          * get the stride size.
2076          */
2077         stride += conf->raid_disks - 1;
2078         sector_div(stride, conf->raid_disks);
2079         mddev->size = stride  << (conf->chunk_shift-1);
2080
2081         if (fo)
2082                 stride = 1;
2083         else
2084                 sector_div(stride, fc);
2085         conf->stride = stride << conf->chunk_shift;
2086
2087         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2088                                                 r10bio_pool_free, conf);
2089         if (!conf->r10bio_pool) {
2090                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2091                         mdname(mddev));
2092                 goto out_free_conf;
2093         }
2094
2095         spin_lock_init(&conf->device_lock);
2096         mddev->queue->queue_lock = &conf->device_lock;
2097
2098         rdev_for_each(rdev, tmp, mddev) {
2099                 disk_idx = rdev->raid_disk;
2100                 if (disk_idx >= mddev->raid_disks
2101                     || disk_idx < 0)
2102                         continue;
2103                 disk = conf->mirrors + disk_idx;
2104
2105                 disk->rdev = rdev;
2106
2107                 blk_queue_stack_limits(mddev->queue,
2108                                        rdev->bdev->bd_disk->queue);
2109                 /* as we don't honour merge_bvec_fn, we must never risk
2110                  * violating it, so limit ->max_sector to one PAGE, as
2111                  * a one page request is never in violation.
2112                  */
2113                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
2114                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
2115                         mddev->queue->max_sectors = (PAGE_SIZE>>9);
2116
2117                 disk->head_position = 0;
2118         }
2119         INIT_LIST_HEAD(&conf->retry_list);
2120
2121         spin_lock_init(&conf->resync_lock);
2122         init_waitqueue_head(&conf->wait_barrier);
2123
2124         /* need to check that every block has at least one working mirror */
2125         if (!enough(conf)) {
2126                 printk(KERN_ERR "raid10: not enough operational mirrors for %s\n",
2127                        mdname(mddev));
2128                 goto out_free_conf;
2129         }
2130
2131         mddev->degraded = 0;
2132         for (i = 0; i < conf->raid_disks; i++) {
2133
2134                 disk = conf->mirrors + i;
2135
2136                 if (!disk->rdev ||
2137                     !test_bit(In_sync, &disk->rdev->flags)) {
2138                         disk->head_position = 0;
2139                         mddev->degraded++;
2140                 }
2141         }
2142
2143
2144         mddev->thread = md_register_thread(raid10d, mddev, "%s_raid10");
2145         if (!mddev->thread) {
2146                 printk(KERN_ERR
2147                        "raid10: couldn't allocate thread for %s\n",
2148                        mdname(mddev));
2149                 goto out_free_conf;
2150         }
2151
2152         printk(KERN_INFO
2153                 "raid10: raid set %s active with %d out of %d devices\n",
2154                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2155                 mddev->raid_disks);
2156         /*
2157          * Ok, everything is just fine now
2158          */
2159         mddev->array_size = size << (conf->chunk_shift-1);
2160         mddev->resync_max_sectors = size << conf->chunk_shift;
2161
2162         mddev->queue->unplug_fn = raid10_unplug;
2163         mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2164         mddev->queue->backing_dev_info.congested_data = mddev;
2165
2166         /* Calculate max read-ahead size.
2167          * We need to readahead at least twice a whole stripe....
2168          * maybe...
2169          */
2170         {
2171                 int stripe = conf->raid_disks * (mddev->chunk_size / PAGE_SIZE);
2172                 stripe /= conf->near_copies;
2173                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2174                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2175         }
2176
2177         if (conf->near_copies < mddev->raid_disks)
2178                 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2179         return 0;
2180
2181 out_free_conf:
2182         if (conf->r10bio_pool)
2183                 mempool_destroy(conf->r10bio_pool);
2184         safe_put_page(conf->tmppage);
2185         kfree(conf->mirrors);
2186         kfree(conf);
2187         mddev->private = NULL;
2188 out:
2189         return -EIO;
2190 }
2191
2192 static int stop(mddev_t *mddev)
2193 {
2194         conf_t *conf = mddev_to_conf(mddev);
2195
2196         md_unregister_thread(mddev->thread);
2197         mddev->thread = NULL;
2198         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2199         if (conf->r10bio_pool)
2200                 mempool_destroy(conf->r10bio_pool);
2201         kfree(conf->mirrors);
2202         kfree(conf);
2203         mddev->private = NULL;
2204         return 0;
2205 }
2206
2207 static void raid10_quiesce(mddev_t *mddev, int state)
2208 {
2209         conf_t *conf = mddev_to_conf(mddev);
2210
2211         switch(state) {
2212         case 1:
2213                 raise_barrier(conf, 0);
2214                 break;
2215         case 0:
2216                 lower_barrier(conf);
2217                 break;
2218         }
2219         if (mddev->thread) {
2220                 if (mddev->bitmap)
2221                         mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
2222                 else
2223                         mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
2224                 md_wakeup_thread(mddev->thread);
2225         }
2226 }
2227
2228 static struct mdk_personality raid10_personality =
2229 {
2230         .name           = "raid10",
2231         .level          = 10,
2232         .owner          = THIS_MODULE,
2233         .make_request   = make_request,
2234         .run            = run,
2235         .stop           = stop,
2236         .status         = status,
2237         .error_handler  = error,
2238         .hot_add_disk   = raid10_add_disk,
2239         .hot_remove_disk= raid10_remove_disk,
2240         .spare_active   = raid10_spare_active,
2241         .sync_request   = sync_request,
2242         .quiesce        = raid10_quiesce,
2243 };
2244
2245 static int __init raid_init(void)
2246 {
2247         return register_md_personality(&raid10_personality);
2248 }
2249
2250 static void raid_exit(void)
2251 {
2252         unregister_md_personality(&raid10_personality);
2253 }
2254
2255 module_init(raid_init);
2256 module_exit(raid_exit);
2257 MODULE_LICENSE("GPL");
2258 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2259 MODULE_ALIAS("md-raid10");
2260 MODULE_ALIAS("md-level-10");