Merge master.kernel.org:/pub/scm/linux/kernel/git/davej/agpgart
[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 <linux/raid/raid10.h>
22
23 /*
24  * RAID10 provides a combination of RAID0 and RAID1 functionality.
25  * The layout of data is defined by
26  *    chunk_size
27  *    raid_disks
28  *    near_copies (stored in low byte of layout)
29  *    far_copies (stored in second byte of layout)
30  *
31  * The data to be stored is divided into chunks using chunksize.
32  * Each device is divided into far_copies sections.
33  * In each section, chunks are laid out in a style similar to raid0, but
34  * near_copies copies of each chunk is stored (each on a different drive).
35  * The starting device for each section is offset near_copies from the starting
36  * device of the previous section.
37  * Thus there are (near_copies*far_copies) of each chunk, and each is on a different
38  * drive.
39  * near_copies and far_copies must be at least one, and their product is at most
40  * raid_disks.
41  */
42
43 /*
44  * Number of guaranteed r10bios in case of extreme VM load:
45  */
46 #define NR_RAID10_BIOS 256
47
48 static void unplug_slaves(mddev_t *mddev);
49
50 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
51 {
52         conf_t *conf = data;
53         r10bio_t *r10_bio;
54         int size = offsetof(struct r10bio_s, devs[conf->copies]);
55
56         /* allocate a r10bio with room for raid_disks entries in the bios array */
57         r10_bio = kmalloc(size, gfp_flags);
58         if (r10_bio)
59                 memset(r10_bio, 0, size);
60         else
61                 unplug_slaves(conf->mddev);
62
63         return r10_bio;
64 }
65
66 static void r10bio_pool_free(void *r10_bio, void *data)
67 {
68         kfree(r10_bio);
69 }
70
71 #define RESYNC_BLOCK_SIZE (64*1024)
72 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
73 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
74 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
75 #define RESYNC_WINDOW (2048*1024)
76
77 /*
78  * When performing a resync, we need to read and compare, so
79  * we need as many pages are there are copies.
80  * When performing a recovery, we need 2 bios, one for read,
81  * one for write (we recover only one drive per r10buf)
82  *
83  */
84 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
85 {
86         conf_t *conf = data;
87         struct page *page;
88         r10bio_t *r10_bio;
89         struct bio *bio;
90         int i, j;
91         int nalloc;
92
93         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
94         if (!r10_bio) {
95                 unplug_slaves(conf->mddev);
96                 return NULL;
97         }
98
99         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
100                 nalloc = conf->copies; /* resync */
101         else
102                 nalloc = 2; /* recovery */
103
104         /*
105          * Allocate bios.
106          */
107         for (j = nalloc ; j-- ; ) {
108                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
109                 if (!bio)
110                         goto out_free_bio;
111                 r10_bio->devs[j].bio = bio;
112         }
113         /*
114          * Allocate RESYNC_PAGES data pages and attach them
115          * where needed.
116          */
117         for (j = 0 ; j < nalloc; j++) {
118                 bio = r10_bio->devs[j].bio;
119                 for (i = 0; i < RESYNC_PAGES; i++) {
120                         page = alloc_page(gfp_flags);
121                         if (unlikely(!page))
122                                 goto out_free_pages;
123
124                         bio->bi_io_vec[i].bv_page = page;
125                 }
126         }
127
128         return r10_bio;
129
130 out_free_pages:
131         for ( ; i > 0 ; i--)
132                 __free_page(bio->bi_io_vec[i-1].bv_page);
133         while (j--)
134                 for (i = 0; i < RESYNC_PAGES ; i++)
135                         __free_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
136         j = -1;
137 out_free_bio:
138         while ( ++j < nalloc )
139                 bio_put(r10_bio->devs[j].bio);
140         r10bio_pool_free(r10_bio, conf);
141         return NULL;
142 }
143
144 static void r10buf_pool_free(void *__r10_bio, void *data)
145 {
146         int i;
147         conf_t *conf = data;
148         r10bio_t *r10bio = __r10_bio;
149         int j;
150
151         for (j=0; j < conf->copies; j++) {
152                 struct bio *bio = r10bio->devs[j].bio;
153                 if (bio) {
154                         for (i = 0; i < RESYNC_PAGES; i++) {
155                                 __free_page(bio->bi_io_vec[i].bv_page);
156                                 bio->bi_io_vec[i].bv_page = NULL;
157                         }
158                         bio_put(bio);
159                 }
160         }
161         r10bio_pool_free(r10bio, conf);
162 }
163
164 static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
165 {
166         int i;
167
168         for (i = 0; i < conf->copies; i++) {
169                 struct bio **bio = & r10_bio->devs[i].bio;
170                 if (*bio)
171                         bio_put(*bio);
172                 *bio = NULL;
173         }
174 }
175
176 static inline void free_r10bio(r10bio_t *r10_bio)
177 {
178         unsigned long flags;
179
180         conf_t *conf = mddev_to_conf(r10_bio->mddev);
181
182         /*
183          * Wake up any possible resync thread that waits for the device
184          * to go idle.
185          */
186         spin_lock_irqsave(&conf->resync_lock, flags);
187         if (!--conf->nr_pending) {
188                 wake_up(&conf->wait_idle);
189                 wake_up(&conf->wait_resume);
190         }
191         spin_unlock_irqrestore(&conf->resync_lock, flags);
192
193         put_all_bios(conf, r10_bio);
194         mempool_free(r10_bio, conf->r10bio_pool);
195 }
196
197 static inline void put_buf(r10bio_t *r10_bio)
198 {
199         conf_t *conf = mddev_to_conf(r10_bio->mddev);
200         unsigned long flags;
201
202         mempool_free(r10_bio, conf->r10buf_pool);
203
204         spin_lock_irqsave(&conf->resync_lock, flags);
205         if (!conf->barrier)
206                 BUG();
207         --conf->barrier;
208         wake_up(&conf->wait_resume);
209         wake_up(&conf->wait_idle);
210
211         if (!--conf->nr_pending) {
212                 wake_up(&conf->wait_idle);
213                 wake_up(&conf->wait_resume);
214         }
215         spin_unlock_irqrestore(&conf->resync_lock, flags);
216 }
217
218 static void reschedule_retry(r10bio_t *r10_bio)
219 {
220         unsigned long flags;
221         mddev_t *mddev = r10_bio->mddev;
222         conf_t *conf = mddev_to_conf(mddev);
223
224         spin_lock_irqsave(&conf->device_lock, flags);
225         list_add(&r10_bio->retry_list, &conf->retry_list);
226         spin_unlock_irqrestore(&conf->device_lock, flags);
227
228         md_wakeup_thread(mddev->thread);
229 }
230
231 /*
232  * raid_end_bio_io() is called when we have finished servicing a mirrored
233  * operation and are ready to return a success/failure code to the buffer
234  * cache layer.
235  */
236 static void raid_end_bio_io(r10bio_t *r10_bio)
237 {
238         struct bio *bio = r10_bio->master_bio;
239
240         bio_endio(bio, bio->bi_size,
241                 test_bit(R10BIO_Uptodate, &r10_bio->state) ? 0 : -EIO);
242         free_r10bio(r10_bio);
243 }
244
245 /*
246  * Update disk head position estimator based on IRQ completion info.
247  */
248 static inline void update_head_pos(int slot, r10bio_t *r10_bio)
249 {
250         conf_t *conf = mddev_to_conf(r10_bio->mddev);
251
252         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
253                 r10_bio->devs[slot].addr + (r10_bio->sectors);
254 }
255
256 static int raid10_end_read_request(struct bio *bio, unsigned int bytes_done, int error)
257 {
258         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
259         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
260         int slot, dev;
261         conf_t *conf = mddev_to_conf(r10_bio->mddev);
262
263         if (bio->bi_size)
264                 return 1;
265
266         slot = r10_bio->read_slot;
267         dev = r10_bio->devs[slot].devnum;
268         /*
269          * this branch is our 'one mirror IO has finished' event handler:
270          */
271         if (!uptodate)
272                 md_error(r10_bio->mddev, conf->mirrors[dev].rdev);
273         else
274                 /*
275                  * Set R10BIO_Uptodate in our master bio, so that
276                  * we will return a good error code to the higher
277                  * levels even if IO on some other mirrored buffer fails.
278                  *
279                  * The 'master' represents the composite IO operation to
280                  * user-side. So if something waits for IO, then it will
281                  * wait for the 'master' bio.
282                  */
283                 set_bit(R10BIO_Uptodate, &r10_bio->state);
284
285         update_head_pos(slot, r10_bio);
286
287         /*
288          * we have only one bio on the read side
289          */
290         if (uptodate)
291                 raid_end_bio_io(r10_bio);
292         else {
293                 /*
294                  * oops, read error:
295                  */
296                 char b[BDEVNAME_SIZE];
297                 if (printk_ratelimit())
298                         printk(KERN_ERR "raid10: %s: rescheduling sector %llu\n",
299                                bdevname(conf->mirrors[dev].rdev->bdev,b), (unsigned long long)r10_bio->sector);
300                 reschedule_retry(r10_bio);
301         }
302
303         rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
304         return 0;
305 }
306
307 static int raid10_end_write_request(struct bio *bio, unsigned int bytes_done, int error)
308 {
309         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
310         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
311         int slot, dev;
312         conf_t *conf = mddev_to_conf(r10_bio->mddev);
313
314         if (bio->bi_size)
315                 return 1;
316
317         for (slot = 0; slot < conf->copies; slot++)
318                 if (r10_bio->devs[slot].bio == bio)
319                         break;
320         dev = r10_bio->devs[slot].devnum;
321
322         /*
323          * this branch is our 'one mirror IO has finished' event handler:
324          */
325         if (!uptodate)
326                 md_error(r10_bio->mddev, conf->mirrors[dev].rdev);
327         else
328                 /*
329                  * Set R10BIO_Uptodate in our master bio, so that
330                  * we will return a good error code for to the higher
331                  * levels even if IO on some other mirrored buffer fails.
332                  *
333                  * The 'master' represents the composite IO operation to
334                  * user-side. So if something waits for IO, then it will
335                  * wait for the 'master' bio.
336                  */
337                 set_bit(R10BIO_Uptodate, &r10_bio->state);
338
339         update_head_pos(slot, r10_bio);
340
341         /*
342          *
343          * Let's see if all mirrored write operations have finished
344          * already.
345          */
346         if (atomic_dec_and_test(&r10_bio->remaining)) {
347                 md_write_end(r10_bio->mddev);
348                 raid_end_bio_io(r10_bio);
349         }
350
351         rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
352         return 0;
353 }
354
355
356 /*
357  * RAID10 layout manager
358  * Aswell as the chunksize and raid_disks count, there are two
359  * parameters: near_copies and far_copies.
360  * near_copies * far_copies must be <= raid_disks.
361  * Normally one of these will be 1.
362  * If both are 1, we get raid0.
363  * If near_copies == raid_disks, we get raid1.
364  *
365  * Chunks are layed out in raid0 style with near_copies copies of the
366  * first chunk, followed by near_copies copies of the next chunk and
367  * so on.
368  * If far_copies > 1, then after 1/far_copies of the array has been assigned
369  * as described above, we start again with a device offset of near_copies.
370  * So we effectively have another copy of the whole array further down all
371  * the drives, but with blocks on different drives.
372  * With this layout, and block is never stored twice on the one device.
373  *
374  * raid10_find_phys finds the sector offset of a given virtual sector
375  * on each device that it is on. If a block isn't on a device,
376  * that entry in the array is set to MaxSector.
377  *
378  * raid10_find_virt does the reverse mapping, from a device and a
379  * sector offset to a virtual address
380  */
381
382 static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
383 {
384         int n,f;
385         sector_t sector;
386         sector_t chunk;
387         sector_t stripe;
388         int dev;
389
390         int slot = 0;
391
392         /* now calculate first sector/dev */
393         chunk = r10bio->sector >> conf->chunk_shift;
394         sector = r10bio->sector & conf->chunk_mask;
395
396         chunk *= conf->near_copies;
397         stripe = chunk;
398         dev = sector_div(stripe, conf->raid_disks);
399
400         sector += stripe << conf->chunk_shift;
401
402         /* and calculate all the others */
403         for (n=0; n < conf->near_copies; n++) {
404                 int d = dev;
405                 sector_t s = sector;
406                 r10bio->devs[slot].addr = sector;
407                 r10bio->devs[slot].devnum = d;
408                 slot++;
409
410                 for (f = 1; f < conf->far_copies; f++) {
411                         d += conf->near_copies;
412                         if (d >= conf->raid_disks)
413                                 d -= conf->raid_disks;
414                         s += conf->stride;
415                         r10bio->devs[slot].devnum = d;
416                         r10bio->devs[slot].addr = s;
417                         slot++;
418                 }
419                 dev++;
420                 if (dev >= conf->raid_disks) {
421                         dev = 0;
422                         sector += (conf->chunk_mask + 1);
423                 }
424         }
425         BUG_ON(slot != conf->copies);
426 }
427
428 static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
429 {
430         sector_t offset, chunk, vchunk;
431
432         while (sector > conf->stride) {
433                 sector -= conf->stride;
434                 if (dev < conf->near_copies)
435                         dev += conf->raid_disks - conf->near_copies;
436                 else
437                         dev -= conf->near_copies;
438         }
439
440         offset = sector & conf->chunk_mask;
441         chunk = sector >> conf->chunk_shift;
442         vchunk = chunk * conf->raid_disks + dev;
443         sector_div(vchunk, conf->near_copies);
444         return (vchunk << conf->chunk_shift) + offset;
445 }
446
447 /**
448  *      raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
449  *      @q: request queue
450  *      @bio: the buffer head that's been built up so far
451  *      @biovec: the request that could be merged to it.
452  *
453  *      Return amount of bytes we can accept at this offset
454  *      If near_copies == raid_disk, there are no striping issues,
455  *      but in that case, the function isn't called at all.
456  */
457 static int raid10_mergeable_bvec(request_queue_t *q, struct bio *bio,
458                                 struct bio_vec *bio_vec)
459 {
460         mddev_t *mddev = q->queuedata;
461         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
462         int max;
463         unsigned int chunk_sectors = mddev->chunk_size >> 9;
464         unsigned int bio_sectors = bio->bi_size >> 9;
465
466         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
467         if (max < 0) max = 0; /* bio_add cannot handle a negative return */
468         if (max <= bio_vec->bv_len && bio_sectors == 0)
469                 return bio_vec->bv_len;
470         else
471                 return max;
472 }
473
474 /*
475  * This routine returns the disk from which the requested read should
476  * be done. There is a per-array 'next expected sequential IO' sector
477  * number - if this matches on the next IO then we use the last disk.
478  * There is also a per-disk 'last know head position' sector that is
479  * maintained from IRQ contexts, both the normal and the resync IO
480  * completion handlers update this position correctly. If there is no
481  * perfect sequential match then we pick the disk whose head is closest.
482  *
483  * If there are 2 mirrors in the same 2 devices, performance degrades
484  * because position is mirror, not device based.
485  *
486  * The rdev for the device selected will have nr_pending incremented.
487  */
488
489 /*
490  * FIXME: possibly should rethink readbalancing and do it differently
491  * depending on near_copies / far_copies geometry.
492  */
493 static int read_balance(conf_t *conf, r10bio_t *r10_bio)
494 {
495         const unsigned long this_sector = r10_bio->sector;
496         int disk, slot, nslot;
497         const int sectors = r10_bio->sectors;
498         sector_t new_distance, current_distance;
499         mdk_rdev_t *rdev;
500
501         raid10_find_phys(conf, r10_bio);
502         rcu_read_lock();
503         /*
504          * Check if we can balance. We can balance on the whole
505          * device if no resync is going on, or below the resync window.
506          * We take the first readable disk when above the resync window.
507          */
508         if (conf->mddev->recovery_cp < MaxSector
509             && (this_sector + sectors >= conf->next_resync)) {
510                 /* make sure that disk is operational */
511                 slot = 0;
512                 disk = r10_bio->devs[slot].devnum;
513
514                 while ((rdev = rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
515                        !test_bit(In_sync, &rdev->flags)) {
516                         slot++;
517                         if (slot == conf->copies) {
518                                 slot = 0;
519                                 disk = -1;
520                                 break;
521                         }
522                         disk = r10_bio->devs[slot].devnum;
523                 }
524                 goto rb_out;
525         }
526
527
528         /* make sure the disk is operational */
529         slot = 0;
530         disk = r10_bio->devs[slot].devnum;
531         while ((rdev=rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
532                !test_bit(In_sync, &rdev->flags)) {
533                 slot ++;
534                 if (slot == conf->copies) {
535                         disk = -1;
536                         goto rb_out;
537                 }
538                 disk = r10_bio->devs[slot].devnum;
539         }
540
541
542         current_distance = abs(r10_bio->devs[slot].addr -
543                                conf->mirrors[disk].head_position);
544
545         /* Find the disk whose head is closest */
546
547         for (nslot = slot; nslot < conf->copies; nslot++) {
548                 int ndisk = r10_bio->devs[nslot].devnum;
549
550
551                 if ((rdev=rcu_dereference(conf->mirrors[ndisk].rdev)) == NULL ||
552                     !test_bit(In_sync, &rdev->flags))
553                         continue;
554
555                 if (!atomic_read(&rdev->nr_pending)) {
556                         disk = ndisk;
557                         slot = nslot;
558                         break;
559                 }
560                 new_distance = abs(r10_bio->devs[nslot].addr -
561                                    conf->mirrors[ndisk].head_position);
562                 if (new_distance < current_distance) {
563                         current_distance = new_distance;
564                         disk = ndisk;
565                         slot = nslot;
566                 }
567         }
568
569 rb_out:
570         r10_bio->read_slot = slot;
571 /*      conf->next_seq_sect = this_sector + sectors;*/
572
573         if (disk >= 0 && (rdev=rcu_dereference(conf->mirrors[disk].rdev))!= NULL)
574                 atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
575         rcu_read_unlock();
576
577         return disk;
578 }
579
580 static void unplug_slaves(mddev_t *mddev)
581 {
582         conf_t *conf = mddev_to_conf(mddev);
583         int i;
584
585         rcu_read_lock();
586         for (i=0; i<mddev->raid_disks; i++) {
587                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
588                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
589                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
590
591                         atomic_inc(&rdev->nr_pending);
592                         rcu_read_unlock();
593
594                         if (r_queue->unplug_fn)
595                                 r_queue->unplug_fn(r_queue);
596
597                         rdev_dec_pending(rdev, mddev);
598                         rcu_read_lock();
599                 }
600         }
601         rcu_read_unlock();
602 }
603
604 static void raid10_unplug(request_queue_t *q)
605 {
606         unplug_slaves(q->queuedata);
607 }
608
609 static int raid10_issue_flush(request_queue_t *q, struct gendisk *disk,
610                              sector_t *error_sector)
611 {
612         mddev_t *mddev = q->queuedata;
613         conf_t *conf = mddev_to_conf(mddev);
614         int i, ret = 0;
615
616         rcu_read_lock();
617         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
618                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
619                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
620                         struct block_device *bdev = rdev->bdev;
621                         request_queue_t *r_queue = bdev_get_queue(bdev);
622
623                         if (!r_queue->issue_flush_fn)
624                                 ret = -EOPNOTSUPP;
625                         else {
626                                 atomic_inc(&rdev->nr_pending);
627                                 rcu_read_unlock();
628                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
629                                                               error_sector);
630                                 rdev_dec_pending(rdev, mddev);
631                                 rcu_read_lock();
632                         }
633                 }
634         }
635         rcu_read_unlock();
636         return ret;
637 }
638
639 /*
640  * Throttle resync depth, so that we can both get proper overlapping of
641  * requests, but are still able to handle normal requests quickly.
642  */
643 #define RESYNC_DEPTH 32
644
645 static void device_barrier(conf_t *conf, sector_t sect)
646 {
647         spin_lock_irq(&conf->resync_lock);
648         wait_event_lock_irq(conf->wait_idle, !waitqueue_active(&conf->wait_resume),
649                             conf->resync_lock, unplug_slaves(conf->mddev));
650
651         if (!conf->barrier++) {
652                 wait_event_lock_irq(conf->wait_idle, !conf->nr_pending,
653                                     conf->resync_lock, unplug_slaves(conf->mddev));
654                 if (conf->nr_pending)
655                         BUG();
656         }
657         wait_event_lock_irq(conf->wait_resume, conf->barrier < RESYNC_DEPTH,
658                             conf->resync_lock, unplug_slaves(conf->mddev));
659         conf->next_resync = sect;
660         spin_unlock_irq(&conf->resync_lock);
661 }
662
663 static int make_request(request_queue_t *q, struct bio * bio)
664 {
665         mddev_t *mddev = q->queuedata;
666         conf_t *conf = mddev_to_conf(mddev);
667         mirror_info_t *mirror;
668         r10bio_t *r10_bio;
669         struct bio *read_bio;
670         int i;
671         int chunk_sects = conf->chunk_mask + 1;
672         const int rw = bio_data_dir(bio);
673
674         if (unlikely(bio_barrier(bio))) {
675                 bio_endio(bio, bio->bi_size, -EOPNOTSUPP);
676                 return 0;
677         }
678
679         /* If this request crosses a chunk boundary, we need to
680          * split it.  This will only happen for 1 PAGE (or less) requests.
681          */
682         if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
683                       > chunk_sects &&
684                     conf->near_copies < conf->raid_disks)) {
685                 struct bio_pair *bp;
686                 /* Sanity check -- queue functions should prevent this happening */
687                 if (bio->bi_vcnt != 1 ||
688                     bio->bi_idx != 0)
689                         goto bad_map;
690                 /* This is a one page bio that upper layers
691                  * refuse to split for us, so we need to split it.
692                  */
693                 bp = bio_split(bio, bio_split_pool,
694                                chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
695                 if (make_request(q, &bp->bio1))
696                         generic_make_request(&bp->bio1);
697                 if (make_request(q, &bp->bio2))
698                         generic_make_request(&bp->bio2);
699
700                 bio_pair_release(bp);
701                 return 0;
702         bad_map:
703                 printk("raid10_make_request bug: can't convert block across chunks"
704                        " or bigger than %dk %llu %d\n", chunk_sects/2,
705                        (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
706
707                 bio_io_error(bio, bio->bi_size);
708                 return 0;
709         }
710
711         md_write_start(mddev, bio);
712
713         /*
714          * Register the new request and wait if the reconstruction
715          * thread has put up a bar for new requests.
716          * Continue immediately if no resync is active currently.
717          */
718         spin_lock_irq(&conf->resync_lock);
719         wait_event_lock_irq(conf->wait_resume, !conf->barrier, conf->resync_lock, );
720         conf->nr_pending++;
721         spin_unlock_irq(&conf->resync_lock);
722
723         disk_stat_inc(mddev->gendisk, ios[rw]);
724         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
725
726         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
727
728         r10_bio->master_bio = bio;
729         r10_bio->sectors = bio->bi_size >> 9;
730
731         r10_bio->mddev = mddev;
732         r10_bio->sector = bio->bi_sector;
733
734         if (rw == READ) {
735                 /*
736                  * read balancing logic:
737                  */
738                 int disk = read_balance(conf, r10_bio);
739                 int slot = r10_bio->read_slot;
740                 if (disk < 0) {
741                         raid_end_bio_io(r10_bio);
742                         return 0;
743                 }
744                 mirror = conf->mirrors + disk;
745
746                 read_bio = bio_clone(bio, GFP_NOIO);
747
748                 r10_bio->devs[slot].bio = read_bio;
749
750                 read_bio->bi_sector = r10_bio->devs[slot].addr +
751                         mirror->rdev->data_offset;
752                 read_bio->bi_bdev = mirror->rdev->bdev;
753                 read_bio->bi_end_io = raid10_end_read_request;
754                 read_bio->bi_rw = READ;
755                 read_bio->bi_private = r10_bio;
756
757                 generic_make_request(read_bio);
758                 return 0;
759         }
760
761         /*
762          * WRITE:
763          */
764         /* first select target devices under spinlock and
765          * inc refcount on their rdev.  Record them by setting
766          * bios[x] to bio
767          */
768         raid10_find_phys(conf, r10_bio);
769         rcu_read_lock();
770         for (i = 0;  i < conf->copies; i++) {
771                 int d = r10_bio->devs[i].devnum;
772                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
773                 if (rdev &&
774                     !test_bit(Faulty, &rdev->flags)) {
775                         atomic_inc(&rdev->nr_pending);
776                         r10_bio->devs[i].bio = bio;
777                 } else
778                         r10_bio->devs[i].bio = NULL;
779         }
780         rcu_read_unlock();
781
782         atomic_set(&r10_bio->remaining, 1);
783
784         for (i = 0; i < conf->copies; i++) {
785                 struct bio *mbio;
786                 int d = r10_bio->devs[i].devnum;
787                 if (!r10_bio->devs[i].bio)
788                         continue;
789
790                 mbio = bio_clone(bio, GFP_NOIO);
791                 r10_bio->devs[i].bio = mbio;
792
793                 mbio->bi_sector = r10_bio->devs[i].addr+
794                         conf->mirrors[d].rdev->data_offset;
795                 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
796                 mbio->bi_end_io = raid10_end_write_request;
797                 mbio->bi_rw = WRITE;
798                 mbio->bi_private = r10_bio;
799
800                 atomic_inc(&r10_bio->remaining);
801                 generic_make_request(mbio);
802         }
803
804         if (atomic_dec_and_test(&r10_bio->remaining)) {
805                 md_write_end(mddev);
806                 raid_end_bio_io(r10_bio);
807         }
808
809         return 0;
810 }
811
812 static void status(struct seq_file *seq, mddev_t *mddev)
813 {
814         conf_t *conf = mddev_to_conf(mddev);
815         int i;
816
817         if (conf->near_copies < conf->raid_disks)
818                 seq_printf(seq, " %dK chunks", mddev->chunk_size/1024);
819         if (conf->near_copies > 1)
820                 seq_printf(seq, " %d near-copies", conf->near_copies);
821         if (conf->far_copies > 1)
822                 seq_printf(seq, " %d far-copies", conf->far_copies);
823
824         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
825                                                 conf->working_disks);
826         for (i = 0; i < conf->raid_disks; i++)
827                 seq_printf(seq, "%s",
828                               conf->mirrors[i].rdev &&
829                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
830         seq_printf(seq, "]");
831 }
832
833 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
834 {
835         char b[BDEVNAME_SIZE];
836         conf_t *conf = mddev_to_conf(mddev);
837
838         /*
839          * If it is not operational, then we have already marked it as dead
840          * else if it is the last working disks, ignore the error, let the
841          * next level up know.
842          * else mark the drive as failed
843          */
844         if (test_bit(In_sync, &rdev->flags)
845             && conf->working_disks == 1)
846                 /*
847                  * Don't fail the drive, just return an IO error.
848                  * The test should really be more sophisticated than
849                  * "working_disks == 1", but it isn't critical, and
850                  * can wait until we do more sophisticated "is the drive
851                  * really dead" tests...
852                  */
853                 return;
854         if (test_bit(In_sync, &rdev->flags)) {
855                 mddev->degraded++;
856                 conf->working_disks--;
857                 /*
858                  * if recovery is running, make sure it aborts.
859                  */
860                 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
861         }
862         clear_bit(In_sync, &rdev->flags);
863         set_bit(Faulty, &rdev->flags);
864         mddev->sb_dirty = 1;
865         printk(KERN_ALERT "raid10: Disk failure on %s, disabling device. \n"
866                 "       Operation continuing on %d devices\n",
867                 bdevname(rdev->bdev,b), conf->working_disks);
868 }
869
870 static void print_conf(conf_t *conf)
871 {
872         int i;
873         mirror_info_t *tmp;
874
875         printk("RAID10 conf printout:\n");
876         if (!conf) {
877                 printk("(!conf)\n");
878                 return;
879         }
880         printk(" --- wd:%d rd:%d\n", conf->working_disks,
881                 conf->raid_disks);
882
883         for (i = 0; i < conf->raid_disks; i++) {
884                 char b[BDEVNAME_SIZE];
885                 tmp = conf->mirrors + i;
886                 if (tmp->rdev)
887                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
888                                 i, !test_bit(In_sync, &tmp->rdev->flags),
889                                 !test_bit(Faulty, &tmp->rdev->flags),
890                                 bdevname(tmp->rdev->bdev,b));
891         }
892 }
893
894 static void close_sync(conf_t *conf)
895 {
896         spin_lock_irq(&conf->resync_lock);
897         wait_event_lock_irq(conf->wait_resume, !conf->barrier,
898                             conf->resync_lock,  unplug_slaves(conf->mddev));
899         spin_unlock_irq(&conf->resync_lock);
900
901         if (conf->barrier) BUG();
902         if (waitqueue_active(&conf->wait_idle)) BUG();
903
904         mempool_destroy(conf->r10buf_pool);
905         conf->r10buf_pool = NULL;
906 }
907
908 /* check if there are enough drives for
909  * every block to appear on atleast one
910  */
911 static int enough(conf_t *conf)
912 {
913         int first = 0;
914
915         do {
916                 int n = conf->copies;
917                 int cnt = 0;
918                 while (n--) {
919                         if (conf->mirrors[first].rdev)
920                                 cnt++;
921                         first = (first+1) % conf->raid_disks;
922                 }
923                 if (cnt == 0)
924                         return 0;
925         } while (first != 0);
926         return 1;
927 }
928
929 static int raid10_spare_active(mddev_t *mddev)
930 {
931         int i;
932         conf_t *conf = mddev->private;
933         mirror_info_t *tmp;
934
935         /*
936          * Find all non-in_sync disks within the RAID10 configuration
937          * and mark them in_sync
938          */
939         for (i = 0; i < conf->raid_disks; i++) {
940                 tmp = conf->mirrors + i;
941                 if (tmp->rdev
942                     && !test_bit(Faulty, &tmp->rdev->flags)
943                     && !test_bit(In_sync, &tmp->rdev->flags)) {
944                         conf->working_disks++;
945                         mddev->degraded--;
946                         set_bit(In_sync, &tmp->rdev->flags);
947                 }
948         }
949
950         print_conf(conf);
951         return 0;
952 }
953
954
955 static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
956 {
957         conf_t *conf = mddev->private;
958         int found = 0;
959         int mirror;
960         mirror_info_t *p;
961
962         if (mddev->recovery_cp < MaxSector)
963                 /* only hot-add to in-sync arrays, as recovery is
964                  * very different from resync
965                  */
966                 return 0;
967         if (!enough(conf))
968                 return 0;
969
970         for (mirror=0; mirror < mddev->raid_disks; mirror++)
971                 if ( !(p=conf->mirrors+mirror)->rdev) {
972
973                         blk_queue_stack_limits(mddev->queue,
974                                                rdev->bdev->bd_disk->queue);
975                         /* as we don't honour merge_bvec_fn, we must never risk
976                          * violating it, so limit ->max_sector to one PAGE, as
977                          * a one page request is never in violation.
978                          */
979                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
980                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
981                                 mddev->queue->max_sectors = (PAGE_SIZE>>9);
982
983                         p->head_position = 0;
984                         rdev->raid_disk = mirror;
985                         found = 1;
986                         rcu_assign_pointer(p->rdev, rdev);
987                         break;
988                 }
989
990         print_conf(conf);
991         return found;
992 }
993
994 static int raid10_remove_disk(mddev_t *mddev, int number)
995 {
996         conf_t *conf = mddev->private;
997         int err = 0;
998         mdk_rdev_t *rdev;
999         mirror_info_t *p = conf->mirrors+ number;
1000
1001         print_conf(conf);
1002         rdev = p->rdev;
1003         if (rdev) {
1004                 if (test_bit(In_sync, &rdev->flags) ||
1005                     atomic_read(&rdev->nr_pending)) {
1006                         err = -EBUSY;
1007                         goto abort;
1008                 }
1009                 p->rdev = NULL;
1010                 synchronize_rcu();
1011                 if (atomic_read(&rdev->nr_pending)) {
1012                         /* lost the race, try later */
1013                         err = -EBUSY;
1014                         p->rdev = rdev;
1015                 }
1016         }
1017 abort:
1018
1019         print_conf(conf);
1020         return err;
1021 }
1022
1023
1024 static int end_sync_read(struct bio *bio, unsigned int bytes_done, int error)
1025 {
1026         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1027         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1028         conf_t *conf = mddev_to_conf(r10_bio->mddev);
1029         int i,d;
1030
1031         if (bio->bi_size)
1032                 return 1;
1033
1034         for (i=0; i<conf->copies; i++)
1035                 if (r10_bio->devs[i].bio == bio)
1036                         break;
1037         if (i == conf->copies)
1038                 BUG();
1039         update_head_pos(i, r10_bio);
1040         d = r10_bio->devs[i].devnum;
1041         if (!uptodate)
1042                 md_error(r10_bio->mddev,
1043                          conf->mirrors[d].rdev);
1044
1045         /* for reconstruct, we always reschedule after a read.
1046          * for resync, only after all reads
1047          */
1048         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1049             atomic_dec_and_test(&r10_bio->remaining)) {
1050                 /* we have read all the blocks,
1051                  * do the comparison in process context in raid10d
1052                  */
1053                 reschedule_retry(r10_bio);
1054         }
1055         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1056         return 0;
1057 }
1058
1059 static int end_sync_write(struct bio *bio, unsigned int bytes_done, int error)
1060 {
1061         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1062         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1063         mddev_t *mddev = r10_bio->mddev;
1064         conf_t *conf = mddev_to_conf(mddev);
1065         int i,d;
1066
1067         if (bio->bi_size)
1068                 return 1;
1069
1070         for (i = 0; i < conf->copies; i++)
1071                 if (r10_bio->devs[i].bio == bio)
1072                         break;
1073         d = r10_bio->devs[i].devnum;
1074
1075         if (!uptodate)
1076                 md_error(mddev, conf->mirrors[d].rdev);
1077         update_head_pos(i, r10_bio);
1078
1079         while (atomic_dec_and_test(&r10_bio->remaining)) {
1080                 if (r10_bio->master_bio == NULL) {
1081                         /* the primary of several recovery bios */
1082                         md_done_sync(mddev, r10_bio->sectors, 1);
1083                         put_buf(r10_bio);
1084                         break;
1085                 } else {
1086                         r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1087                         put_buf(r10_bio);
1088                         r10_bio = r10_bio2;
1089                 }
1090         }
1091         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1092         return 0;
1093 }
1094
1095 /*
1096  * Note: sync and recover and handled very differently for raid10
1097  * This code is for resync.
1098  * For resync, we read through virtual addresses and read all blocks.
1099  * If there is any error, we schedule a write.  The lowest numbered
1100  * drive is authoritative.
1101  * However requests come for physical address, so we need to map.
1102  * For every physical address there are raid_disks/copies virtual addresses,
1103  * which is always are least one, but is not necessarly an integer.
1104  * This means that a physical address can span multiple chunks, so we may
1105  * have to submit multiple io requests for a single sync request.
1106  */
1107 /*
1108  * We check if all blocks are in-sync and only write to blocks that
1109  * aren't in sync
1110  */
1111 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1112 {
1113         conf_t *conf = mddev_to_conf(mddev);
1114         int i, first;
1115         struct bio *tbio, *fbio;
1116
1117         atomic_set(&r10_bio->remaining, 1);
1118
1119         /* find the first device with a block */
1120         for (i=0; i<conf->copies; i++)
1121                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1122                         break;
1123
1124         if (i == conf->copies)
1125                 goto done;
1126
1127         first = i;
1128         fbio = r10_bio->devs[i].bio;
1129
1130         /* now find blocks with errors */
1131         for (i=first+1 ; i < conf->copies ; i++) {
1132                 int vcnt, j, d;
1133
1134                 if (!test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1135                         continue;
1136                 /* We know that the bi_io_vec layout is the same for
1137                  * both 'first' and 'i', so we just compare them.
1138                  * All vec entries are PAGE_SIZE;
1139                  */
1140                 tbio = r10_bio->devs[i].bio;
1141                 vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1142                 for (j = 0; j < vcnt; j++)
1143                         if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1144                                    page_address(tbio->bi_io_vec[j].bv_page),
1145                                    PAGE_SIZE))
1146                                 break;
1147                 if (j == vcnt)
1148                         continue;
1149                 /* Ok, we need to write this bio
1150                  * First we need to fixup bv_offset, bv_len and
1151                  * bi_vecs, as the read request might have corrupted these
1152                  */
1153                 tbio->bi_vcnt = vcnt;
1154                 tbio->bi_size = r10_bio->sectors << 9;
1155                 tbio->bi_idx = 0;
1156                 tbio->bi_phys_segments = 0;
1157                 tbio->bi_hw_segments = 0;
1158                 tbio->bi_hw_front_size = 0;
1159                 tbio->bi_hw_back_size = 0;
1160                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1161                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1162                 tbio->bi_next = NULL;
1163                 tbio->bi_rw = WRITE;
1164                 tbio->bi_private = r10_bio;
1165                 tbio->bi_sector = r10_bio->devs[i].addr;
1166
1167                 for (j=0; j < vcnt ; j++) {
1168                         tbio->bi_io_vec[j].bv_offset = 0;
1169                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1170
1171                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1172                                page_address(fbio->bi_io_vec[j].bv_page),
1173                                PAGE_SIZE);
1174                 }
1175                 tbio->bi_end_io = end_sync_write;
1176
1177                 d = r10_bio->devs[i].devnum;
1178                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1179                 atomic_inc(&r10_bio->remaining);
1180                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1181
1182                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1183                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1184                 generic_make_request(tbio);
1185         }
1186
1187 done:
1188         if (atomic_dec_and_test(&r10_bio->remaining)) {
1189                 md_done_sync(mddev, r10_bio->sectors, 1);
1190                 put_buf(r10_bio);
1191         }
1192 }
1193
1194 /*
1195  * Now for the recovery code.
1196  * Recovery happens across physical sectors.
1197  * We recover all non-is_sync drives by finding the virtual address of
1198  * each, and then choose a working drive that also has that virt address.
1199  * There is a separate r10_bio for each non-in_sync drive.
1200  * Only the first two slots are in use. The first for reading,
1201  * The second for writing.
1202  *
1203  */
1204
1205 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1206 {
1207         conf_t *conf = mddev_to_conf(mddev);
1208         int i, d;
1209         struct bio *bio, *wbio;
1210
1211
1212         /* move the pages across to the second bio
1213          * and submit the write request
1214          */
1215         bio = r10_bio->devs[0].bio;
1216         wbio = r10_bio->devs[1].bio;
1217         for (i=0; i < wbio->bi_vcnt; i++) {
1218                 struct page *p = bio->bi_io_vec[i].bv_page;
1219                 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1220                 wbio->bi_io_vec[i].bv_page = p;
1221         }
1222         d = r10_bio->devs[1].devnum;
1223
1224         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1225         md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1226         generic_make_request(wbio);
1227 }
1228
1229
1230 /*
1231  * This is a kernel thread which:
1232  *
1233  *      1.      Retries failed read operations on working mirrors.
1234  *      2.      Updates the raid superblock when problems encounter.
1235  *      3.      Performs writes following reads for array syncronising.
1236  */
1237
1238 static void raid10d(mddev_t *mddev)
1239 {
1240         r10bio_t *r10_bio;
1241         struct bio *bio;
1242         unsigned long flags;
1243         conf_t *conf = mddev_to_conf(mddev);
1244         struct list_head *head = &conf->retry_list;
1245         int unplug=0;
1246         mdk_rdev_t *rdev;
1247
1248         md_check_recovery(mddev);
1249
1250         for (;;) {
1251                 char b[BDEVNAME_SIZE];
1252                 spin_lock_irqsave(&conf->device_lock, flags);
1253                 if (list_empty(head))
1254                         break;
1255                 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1256                 list_del(head->prev);
1257                 spin_unlock_irqrestore(&conf->device_lock, flags);
1258
1259                 mddev = r10_bio->mddev;
1260                 conf = mddev_to_conf(mddev);
1261                 if (test_bit(R10BIO_IsSync, &r10_bio->state)) {
1262                         sync_request_write(mddev, r10_bio);
1263                         unplug = 1;
1264                 } else  if (test_bit(R10BIO_IsRecover, &r10_bio->state)) {
1265                         recovery_request_write(mddev, r10_bio);
1266                         unplug = 1;
1267                 } else {
1268                         int mirror;
1269                         bio = r10_bio->devs[r10_bio->read_slot].bio;
1270                         r10_bio->devs[r10_bio->read_slot].bio = NULL;
1271                         bio_put(bio);
1272                         mirror = read_balance(conf, r10_bio);
1273                         if (mirror == -1) {
1274                                 printk(KERN_ALERT "raid10: %s: unrecoverable I/O"
1275                                        " read error for block %llu\n",
1276                                        bdevname(bio->bi_bdev,b),
1277                                        (unsigned long long)r10_bio->sector);
1278                                 raid_end_bio_io(r10_bio);
1279                         } else {
1280                                 rdev = conf->mirrors[mirror].rdev;
1281                                 if (printk_ratelimit())
1282                                         printk(KERN_ERR "raid10: %s: redirecting sector %llu to"
1283                                                " another mirror\n",
1284                                                bdevname(rdev->bdev,b),
1285                                                (unsigned long long)r10_bio->sector);
1286                                 bio = bio_clone(r10_bio->master_bio, GFP_NOIO);
1287                                 r10_bio->devs[r10_bio->read_slot].bio = bio;
1288                                 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1289                                         + rdev->data_offset;
1290                                 bio->bi_bdev = rdev->bdev;
1291                                 bio->bi_rw = READ;
1292                                 bio->bi_private = r10_bio;
1293                                 bio->bi_end_io = raid10_end_read_request;
1294                                 unplug = 1;
1295                                 generic_make_request(bio);
1296                         }
1297                 }
1298         }
1299         spin_unlock_irqrestore(&conf->device_lock, flags);
1300         if (unplug)
1301                 unplug_slaves(mddev);
1302 }
1303
1304
1305 static int init_resync(conf_t *conf)
1306 {
1307         int buffs;
1308
1309         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1310         if (conf->r10buf_pool)
1311                 BUG();
1312         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1313         if (!conf->r10buf_pool)
1314                 return -ENOMEM;
1315         conf->next_resync = 0;
1316         return 0;
1317 }
1318
1319 /*
1320  * perform a "sync" on one "block"
1321  *
1322  * We need to make sure that no normal I/O request - particularly write
1323  * requests - conflict with active sync requests.
1324  *
1325  * This is achieved by tracking pending requests and a 'barrier' concept
1326  * that can be installed to exclude normal IO requests.
1327  *
1328  * Resync and recovery are handled very differently.
1329  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1330  *
1331  * For resync, we iterate over virtual addresses, read all copies,
1332  * and update if there are differences.  If only one copy is live,
1333  * skip it.
1334  * For recovery, we iterate over physical addresses, read a good
1335  * value for each non-in_sync drive, and over-write.
1336  *
1337  * So, for recovery we may have several outstanding complex requests for a
1338  * given address, one for each out-of-sync device.  We model this by allocating
1339  * a number of r10_bio structures, one for each out-of-sync device.
1340  * As we setup these structures, we collect all bio's together into a list
1341  * which we then process collectively to add pages, and then process again
1342  * to pass to generic_make_request.
1343  *
1344  * The r10_bio structures are linked using a borrowed master_bio pointer.
1345  * This link is counted in ->remaining.  When the r10_bio that points to NULL
1346  * has its remaining count decremented to 0, the whole complex operation
1347  * is complete.
1348  *
1349  */
1350
1351 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1352 {
1353         conf_t *conf = mddev_to_conf(mddev);
1354         r10bio_t *r10_bio;
1355         struct bio *biolist = NULL, *bio;
1356         sector_t max_sector, nr_sectors;
1357         int disk;
1358         int i;
1359
1360         sector_t sectors_skipped = 0;
1361         int chunks_skipped = 0;
1362
1363         if (!conf->r10buf_pool)
1364                 if (init_resync(conf))
1365                         return 0;
1366
1367  skipped:
1368         max_sector = mddev->size << 1;
1369         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1370                 max_sector = mddev->resync_max_sectors;
1371         if (sector_nr >= max_sector) {
1372                 close_sync(conf);
1373                 *skipped = 1;
1374                 return sectors_skipped;
1375         }
1376         if (chunks_skipped >= conf->raid_disks) {
1377                 /* if there has been nothing to do on any drive,
1378                  * then there is nothing to do at all..
1379                  */
1380                 *skipped = 1;
1381                 return (max_sector - sector_nr) + sectors_skipped;
1382         }
1383
1384         /* make sure whole request will fit in a chunk - if chunks
1385          * are meaningful
1386          */
1387         if (conf->near_copies < conf->raid_disks &&
1388             max_sector > (sector_nr | conf->chunk_mask))
1389                 max_sector = (sector_nr | conf->chunk_mask) + 1;
1390         /*
1391          * If there is non-resync activity waiting for us then
1392          * put in a delay to throttle resync.
1393          */
1394         if (!go_faster && waitqueue_active(&conf->wait_resume))
1395                 msleep_interruptible(1000);
1396         device_barrier(conf, sector_nr + RESYNC_SECTORS);
1397
1398         /* Again, very different code for resync and recovery.
1399          * Both must result in an r10bio with a list of bios that
1400          * have bi_end_io, bi_sector, bi_bdev set,
1401          * and bi_private set to the r10bio.
1402          * For recovery, we may actually create several r10bios
1403          * with 2 bios in each, that correspond to the bios in the main one.
1404          * In this case, the subordinate r10bios link back through a
1405          * borrowed master_bio pointer, and the counter in the master
1406          * includes a ref from each subordinate.
1407          */
1408         /* First, we decide what to do and set ->bi_end_io
1409          * To end_sync_read if we want to read, and
1410          * end_sync_write if we will want to write.
1411          */
1412
1413         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1414                 /* recovery... the complicated one */
1415                 int i, j, k;
1416                 r10_bio = NULL;
1417
1418                 for (i=0 ; i<conf->raid_disks; i++)
1419                         if (conf->mirrors[i].rdev &&
1420                             !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) {
1421                                 /* want to reconstruct this device */
1422                                 r10bio_t *rb2 = r10_bio;
1423
1424                                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1425                                 spin_lock_irq(&conf->resync_lock);
1426                                 conf->nr_pending++;
1427                                 if (rb2) conf->barrier++;
1428                                 spin_unlock_irq(&conf->resync_lock);
1429                                 atomic_set(&r10_bio->remaining, 0);
1430
1431                                 r10_bio->master_bio = (struct bio*)rb2;
1432                                 if (rb2)
1433                                         atomic_inc(&rb2->remaining);
1434                                 r10_bio->mddev = mddev;
1435                                 set_bit(R10BIO_IsRecover, &r10_bio->state);
1436                                 r10_bio->sector = raid10_find_virt(conf, sector_nr, i);
1437                                 raid10_find_phys(conf, r10_bio);
1438                                 for (j=0; j<conf->copies;j++) {
1439                                         int d = r10_bio->devs[j].devnum;
1440                                         if (conf->mirrors[d].rdev &&
1441                                             test_bit(In_sync, &conf->mirrors[d].rdev->flags)) {
1442                                                 /* This is where we read from */
1443                                                 bio = r10_bio->devs[0].bio;
1444                                                 bio->bi_next = biolist;
1445                                                 biolist = bio;
1446                                                 bio->bi_private = r10_bio;
1447                                                 bio->bi_end_io = end_sync_read;
1448                                                 bio->bi_rw = 0;
1449                                                 bio->bi_sector = r10_bio->devs[j].addr +
1450                                                         conf->mirrors[d].rdev->data_offset;
1451                                                 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1452                                                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1453                                                 atomic_inc(&r10_bio->remaining);
1454                                                 /* and we write to 'i' */
1455
1456                                                 for (k=0; k<conf->copies; k++)
1457                                                         if (r10_bio->devs[k].devnum == i)
1458                                                                 break;
1459                                                 bio = r10_bio->devs[1].bio;
1460                                                 bio->bi_next = biolist;
1461                                                 biolist = bio;
1462                                                 bio->bi_private = r10_bio;
1463                                                 bio->bi_end_io = end_sync_write;
1464                                                 bio->bi_rw = 1;
1465                                                 bio->bi_sector = r10_bio->devs[k].addr +
1466                                                         conf->mirrors[i].rdev->data_offset;
1467                                                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1468
1469                                                 r10_bio->devs[0].devnum = d;
1470                                                 r10_bio->devs[1].devnum = i;
1471
1472                                                 break;
1473                                         }
1474                                 }
1475                                 if (j == conf->copies) {
1476                                         /* Cannot recover, so abort the recovery */
1477                                         put_buf(r10_bio);
1478                                         r10_bio = rb2;
1479                                         if (!test_and_set_bit(MD_RECOVERY_ERR, &mddev->recovery))
1480                                                 printk(KERN_INFO "raid10: %s: insufficient working devices for recovery.\n",
1481                                                        mdname(mddev));
1482                                         break;
1483                                 }
1484                         }
1485                 if (biolist == NULL) {
1486                         while (r10_bio) {
1487                                 r10bio_t *rb2 = r10_bio;
1488                                 r10_bio = (r10bio_t*) rb2->master_bio;
1489                                 rb2->master_bio = NULL;
1490                                 put_buf(rb2);
1491                         }
1492                         goto giveup;
1493                 }
1494         } else {
1495                 /* resync. Schedule a read for every block at this virt offset */
1496                 int count = 0;
1497                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1498
1499                 spin_lock_irq(&conf->resync_lock);
1500                 conf->nr_pending++;
1501                 spin_unlock_irq(&conf->resync_lock);
1502
1503                 r10_bio->mddev = mddev;
1504                 atomic_set(&r10_bio->remaining, 0);
1505
1506                 r10_bio->master_bio = NULL;
1507                 r10_bio->sector = sector_nr;
1508                 set_bit(R10BIO_IsSync, &r10_bio->state);
1509                 raid10_find_phys(conf, r10_bio);
1510                 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1511
1512                 for (i=0; i<conf->copies; i++) {
1513                         int d = r10_bio->devs[i].devnum;
1514                         bio = r10_bio->devs[i].bio;
1515                         bio->bi_end_io = NULL;
1516                         if (conf->mirrors[d].rdev == NULL ||
1517                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
1518                                 continue;
1519                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1520                         atomic_inc(&r10_bio->remaining);
1521                         bio->bi_next = biolist;
1522                         biolist = bio;
1523                         bio->bi_private = r10_bio;
1524                         bio->bi_end_io = end_sync_read;
1525                         bio->bi_rw = 0;
1526                         bio->bi_sector = r10_bio->devs[i].addr +
1527                                 conf->mirrors[d].rdev->data_offset;
1528                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1529                         count++;
1530                 }
1531
1532                 if (count < 2) {
1533                         for (i=0; i<conf->copies; i++) {
1534                                 int d = r10_bio->devs[i].devnum;
1535                                 if (r10_bio->devs[i].bio->bi_end_io)
1536                                         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1537                         }
1538                         put_buf(r10_bio);
1539                         biolist = NULL;
1540                         goto giveup;
1541                 }
1542         }
1543
1544         for (bio = biolist; bio ; bio=bio->bi_next) {
1545
1546                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1547                 if (bio->bi_end_io)
1548                         bio->bi_flags |= 1 << BIO_UPTODATE;
1549                 bio->bi_vcnt = 0;
1550                 bio->bi_idx = 0;
1551                 bio->bi_phys_segments = 0;
1552                 bio->bi_hw_segments = 0;
1553                 bio->bi_size = 0;
1554         }
1555
1556         nr_sectors = 0;
1557         do {
1558                 struct page *page;
1559                 int len = PAGE_SIZE;
1560                 disk = 0;
1561                 if (sector_nr + (len>>9) > max_sector)
1562                         len = (max_sector - sector_nr) << 9;
1563                 if (len == 0)
1564                         break;
1565                 for (bio= biolist ; bio ; bio=bio->bi_next) {
1566                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1567                         if (bio_add_page(bio, page, len, 0) == 0) {
1568                                 /* stop here */
1569                                 struct bio *bio2;
1570                                 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1571                                 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
1572                                         /* remove last page from this bio */
1573                                         bio2->bi_vcnt--;
1574                                         bio2->bi_size -= len;
1575                                         bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
1576                                 }
1577                                 goto bio_full;
1578                         }
1579                         disk = i;
1580                 }
1581                 nr_sectors += len>>9;
1582                 sector_nr += len>>9;
1583         } while (biolist->bi_vcnt < RESYNC_PAGES);
1584  bio_full:
1585         r10_bio->sectors = nr_sectors;
1586
1587         while (biolist) {
1588                 bio = biolist;
1589                 biolist = biolist->bi_next;
1590
1591                 bio->bi_next = NULL;
1592                 r10_bio = bio->bi_private;
1593                 r10_bio->sectors = nr_sectors;
1594
1595                 if (bio->bi_end_io == end_sync_read) {
1596                         md_sync_acct(bio->bi_bdev, nr_sectors);
1597                         generic_make_request(bio);
1598                 }
1599         }
1600
1601         if (sectors_skipped)
1602                 /* pretend they weren't skipped, it makes
1603                  * no important difference in this case
1604                  */
1605                 md_done_sync(mddev, sectors_skipped, 1);
1606
1607         return sectors_skipped + nr_sectors;
1608  giveup:
1609         /* There is nowhere to write, so all non-sync
1610          * drives must be failed, so try the next chunk...
1611          */
1612         {
1613         sector_t sec = max_sector - sector_nr;
1614         sectors_skipped += sec;
1615         chunks_skipped ++;
1616         sector_nr = max_sector;
1617         goto skipped;
1618         }
1619 }
1620
1621 static int run(mddev_t *mddev)
1622 {
1623         conf_t *conf;
1624         int i, disk_idx;
1625         mirror_info_t *disk;
1626         mdk_rdev_t *rdev;
1627         struct list_head *tmp;
1628         int nc, fc;
1629         sector_t stride, size;
1630
1631         if (mddev->level != 10) {
1632                 printk(KERN_ERR "raid10: %s: raid level not set correctly... (%d)\n",
1633                        mdname(mddev), mddev->level);
1634                 goto out;
1635         }
1636         nc = mddev->layout & 255;
1637         fc = (mddev->layout >> 8) & 255;
1638         if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
1639             (mddev->layout >> 16)) {
1640                 printk(KERN_ERR "raid10: %s: unsupported raid10 layout: 0x%8x\n",
1641                        mdname(mddev), mddev->layout);
1642                 goto out;
1643         }
1644         /*
1645          * copy the already verified devices into our private RAID10
1646          * bookkeeping area. [whatever we allocate in run(),
1647          * should be freed in stop()]
1648          */
1649         conf = kmalloc(sizeof(conf_t), GFP_KERNEL);
1650         mddev->private = conf;
1651         if (!conf) {
1652                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1653                         mdname(mddev));
1654                 goto out;
1655         }
1656         memset(conf, 0, sizeof(*conf));
1657         conf->mirrors = kmalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1658                                  GFP_KERNEL);
1659         if (!conf->mirrors) {
1660                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1661                        mdname(mddev));
1662                 goto out_free_conf;
1663         }
1664         memset(conf->mirrors, 0, sizeof(struct mirror_info)*mddev->raid_disks);
1665
1666         conf->near_copies = nc;
1667         conf->far_copies = fc;
1668         conf->copies = nc*fc;
1669         conf->chunk_mask = (sector_t)(mddev->chunk_size>>9)-1;
1670         conf->chunk_shift = ffz(~mddev->chunk_size) - 9;
1671         stride = mddev->size >> (conf->chunk_shift-1);
1672         sector_div(stride, fc);
1673         conf->stride = stride << conf->chunk_shift;
1674
1675         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
1676                                                 r10bio_pool_free, conf);
1677         if (!conf->r10bio_pool) {
1678                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1679                         mdname(mddev));
1680                 goto out_free_conf;
1681         }
1682
1683         ITERATE_RDEV(mddev, rdev, tmp) {
1684                 disk_idx = rdev->raid_disk;
1685                 if (disk_idx >= mddev->raid_disks
1686                     || disk_idx < 0)
1687                         continue;
1688                 disk = conf->mirrors + disk_idx;
1689
1690                 disk->rdev = rdev;
1691
1692                 blk_queue_stack_limits(mddev->queue,
1693                                        rdev->bdev->bd_disk->queue);
1694                 /* as we don't honour merge_bvec_fn, we must never risk
1695                  * violating it, so limit ->max_sector to one PAGE, as
1696                  * a one page request is never in violation.
1697                  */
1698                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1699                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
1700                         mddev->queue->max_sectors = (PAGE_SIZE>>9);
1701
1702                 disk->head_position = 0;
1703                 if (!test_bit(Faulty, &rdev->flags) && test_bit(In_sync, &rdev->flags))
1704                         conf->working_disks++;
1705         }
1706         conf->raid_disks = mddev->raid_disks;
1707         conf->mddev = mddev;
1708         spin_lock_init(&conf->device_lock);
1709         INIT_LIST_HEAD(&conf->retry_list);
1710
1711         spin_lock_init(&conf->resync_lock);
1712         init_waitqueue_head(&conf->wait_idle);
1713         init_waitqueue_head(&conf->wait_resume);
1714
1715         /* need to check that every block has at least one working mirror */
1716         if (!enough(conf)) {
1717                 printk(KERN_ERR "raid10: not enough operational mirrors for %s\n",
1718                        mdname(mddev));
1719                 goto out_free_conf;
1720         }
1721
1722         mddev->degraded = 0;
1723         for (i = 0; i < conf->raid_disks; i++) {
1724
1725                 disk = conf->mirrors + i;
1726
1727                 if (!disk->rdev) {
1728                         disk->head_position = 0;
1729                         mddev->degraded++;
1730                 }
1731         }
1732
1733
1734         mddev->thread = md_register_thread(raid10d, mddev, "%s_raid10");
1735         if (!mddev->thread) {
1736                 printk(KERN_ERR
1737                        "raid10: couldn't allocate thread for %s\n",
1738                        mdname(mddev));
1739                 goto out_free_conf;
1740         }
1741
1742         printk(KERN_INFO
1743                 "raid10: raid set %s active with %d out of %d devices\n",
1744                 mdname(mddev), mddev->raid_disks - mddev->degraded,
1745                 mddev->raid_disks);
1746         /*
1747          * Ok, everything is just fine now
1748          */
1749         size = conf->stride * conf->raid_disks;
1750         sector_div(size, conf->near_copies);
1751         mddev->array_size = size/2;
1752         mddev->resync_max_sectors = size;
1753
1754         mddev->queue->unplug_fn = raid10_unplug;
1755         mddev->queue->issue_flush_fn = raid10_issue_flush;
1756
1757         /* Calculate max read-ahead size.
1758          * We need to readahead at least twice a whole stripe....
1759          * maybe...
1760          */
1761         {
1762                 int stripe = conf->raid_disks * mddev->chunk_size / PAGE_CACHE_SIZE;
1763                 stripe /= conf->near_copies;
1764                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
1765                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
1766         }
1767
1768         if (conf->near_copies < mddev->raid_disks)
1769                 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
1770         return 0;
1771
1772 out_free_conf:
1773         if (conf->r10bio_pool)
1774                 mempool_destroy(conf->r10bio_pool);
1775         kfree(conf->mirrors);
1776         kfree(conf);
1777         mddev->private = NULL;
1778 out:
1779         return -EIO;
1780 }
1781
1782 static int stop(mddev_t *mddev)
1783 {
1784         conf_t *conf = mddev_to_conf(mddev);
1785
1786         md_unregister_thread(mddev->thread);
1787         mddev->thread = NULL;
1788         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
1789         if (conf->r10bio_pool)
1790                 mempool_destroy(conf->r10bio_pool);
1791         kfree(conf->mirrors);
1792         kfree(conf);
1793         mddev->private = NULL;
1794         return 0;
1795 }
1796
1797
1798 static mdk_personality_t raid10_personality =
1799 {
1800         .name           = "raid10",
1801         .owner          = THIS_MODULE,
1802         .make_request   = make_request,
1803         .run            = run,
1804         .stop           = stop,
1805         .status         = status,
1806         .error_handler  = error,
1807         .hot_add_disk   = raid10_add_disk,
1808         .hot_remove_disk= raid10_remove_disk,
1809         .spare_active   = raid10_spare_active,
1810         .sync_request   = sync_request,
1811 };
1812
1813 static int __init raid_init(void)
1814 {
1815         return register_md_personality(RAID10, &raid10_personality);
1816 }
1817
1818 static void raid_exit(void)
1819 {
1820         unregister_md_personality(RAID10);
1821 }
1822
1823 module_init(raid_init);
1824 module_exit(raid_exit);
1825 MODULE_LICENSE("GPL");
1826 MODULE_ALIAS("md-personality-9"); /* RAID10 */