md: the md RAID10 resync thread could cause a md RAID10 array deadlock
[linux-2.6.git] / drivers / md / raid10.c
1 /*
2  * raid10.c : Multiple Devices driver for Linux
3  *
4  * Copyright (C) 2000-2004 Neil Brown
5  *
6  * RAID-10 support for md.
7  *
8  * Base on code in raid1.c.  See raid1.c for futher copyright information.
9  *
10  *
11  * This program is free software; you can redistribute it and/or modify
12  * it under the terms of the GNU General Public License as published by
13  * the Free Software Foundation; either version 2, or (at your option)
14  * any later version.
15  *
16  * You should have received a copy of the GNU General Public License
17  * (for example /usr/src/linux/COPYING); if not, write to the Free
18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19  */
20
21 #include "dm-bio-list.h"
22 #include <linux/raid/raid10.h>
23 #include <linux/raid/bitmap.h>
24
25 /*
26  * RAID10 provides a combination of RAID0 and RAID1 functionality.
27  * The layout of data is defined by
28  *    chunk_size
29  *    raid_disks
30  *    near_copies (stored in low byte of layout)
31  *    far_copies (stored in second byte of layout)
32  *    far_offset (stored in bit 16 of layout )
33  *
34  * The data to be stored is divided into chunks using chunksize.
35  * Each device is divided into far_copies sections.
36  * In each section, chunks are laid out in a style similar to raid0, but
37  * near_copies copies of each chunk is stored (each on a different drive).
38  * The starting device for each section is offset near_copies from the starting
39  * device of the previous section.
40  * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
41  * drive.
42  * near_copies and far_copies must be at least one, and their product is at most
43  * raid_disks.
44  *
45  * If far_offset is true, then the far_copies are handled a bit differently.
46  * The copies are still in different stripes, but instead of be very far apart
47  * on disk, there are adjacent stripes.
48  */
49
50 /*
51  * Number of guaranteed r10bios in case of extreme VM load:
52  */
53 #define NR_RAID10_BIOS 256
54
55 static void unplug_slaves(mddev_t *mddev);
56
57 static void allow_barrier(conf_t *conf);
58 static void lower_barrier(conf_t *conf);
59
60 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
61 {
62         conf_t *conf = data;
63         r10bio_t *r10_bio;
64         int size = offsetof(struct r10bio_s, devs[conf->copies]);
65
66         /* allocate a r10bio with room for raid_disks entries in the bios array */
67         r10_bio = kzalloc(size, gfp_flags);
68         if (!r10_bio)
69                 unplug_slaves(conf->mddev);
70
71         return r10_bio;
72 }
73
74 static void r10bio_pool_free(void *r10_bio, void *data)
75 {
76         kfree(r10_bio);
77 }
78
79 #define RESYNC_BLOCK_SIZE (64*1024)
80 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
81 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
82 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
83 #define RESYNC_WINDOW (2048*1024)
84
85 /*
86  * When performing a resync, we need to read and compare, so
87  * we need as many pages are there are copies.
88  * When performing a recovery, we need 2 bios, one for read,
89  * one for write (we recover only one drive per r10buf)
90  *
91  */
92 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
93 {
94         conf_t *conf = data;
95         struct page *page;
96         r10bio_t *r10_bio;
97         struct bio *bio;
98         int i, j;
99         int nalloc;
100
101         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
102         if (!r10_bio) {
103                 unplug_slaves(conf->mddev);
104                 return NULL;
105         }
106
107         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
108                 nalloc = conf->copies; /* resync */
109         else
110                 nalloc = 2; /* recovery */
111
112         /*
113          * Allocate bios.
114          */
115         for (j = nalloc ; j-- ; ) {
116                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
117                 if (!bio)
118                         goto out_free_bio;
119                 r10_bio->devs[j].bio = bio;
120         }
121         /*
122          * Allocate RESYNC_PAGES data pages and attach them
123          * where needed.
124          */
125         for (j = 0 ; j < nalloc; j++) {
126                 bio = r10_bio->devs[j].bio;
127                 for (i = 0; i < RESYNC_PAGES; i++) {
128                         page = alloc_page(gfp_flags);
129                         if (unlikely(!page))
130                                 goto out_free_pages;
131
132                         bio->bi_io_vec[i].bv_page = page;
133                 }
134         }
135
136         return r10_bio;
137
138 out_free_pages:
139         for ( ; i > 0 ; i--)
140                 safe_put_page(bio->bi_io_vec[i-1].bv_page);
141         while (j--)
142                 for (i = 0; i < RESYNC_PAGES ; i++)
143                         safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
144         j = -1;
145 out_free_bio:
146         while ( ++j < nalloc )
147                 bio_put(r10_bio->devs[j].bio);
148         r10bio_pool_free(r10_bio, conf);
149         return NULL;
150 }
151
152 static void r10buf_pool_free(void *__r10_bio, void *data)
153 {
154         int i;
155         conf_t *conf = data;
156         r10bio_t *r10bio = __r10_bio;
157         int j;
158
159         for (j=0; j < conf->copies; j++) {
160                 struct bio *bio = r10bio->devs[j].bio;
161                 if (bio) {
162                         for (i = 0; i < RESYNC_PAGES; i++) {
163                                 safe_put_page(bio->bi_io_vec[i].bv_page);
164                                 bio->bi_io_vec[i].bv_page = NULL;
165                         }
166                         bio_put(bio);
167                 }
168         }
169         r10bio_pool_free(r10bio, conf);
170 }
171
172 static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
173 {
174         int i;
175
176         for (i = 0; i < conf->copies; i++) {
177                 struct bio **bio = & r10_bio->devs[i].bio;
178                 if (*bio && *bio != IO_BLOCKED)
179                         bio_put(*bio);
180                 *bio = NULL;
181         }
182 }
183
184 static void free_r10bio(r10bio_t *r10_bio)
185 {
186         conf_t *conf = mddev_to_conf(r10_bio->mddev);
187
188         /*
189          * Wake up any possible resync thread that waits for the device
190          * to go idle.
191          */
192         allow_barrier(conf);
193
194         put_all_bios(conf, r10_bio);
195         mempool_free(r10_bio, conf->r10bio_pool);
196 }
197
198 static void put_buf(r10bio_t *r10_bio)
199 {
200         conf_t *conf = mddev_to_conf(r10_bio->mddev);
201
202         mempool_free(r10_bio, conf->r10buf_pool);
203
204         lower_barrier(conf);
205 }
206
207 static void reschedule_retry(r10bio_t *r10_bio)
208 {
209         unsigned long flags;
210         mddev_t *mddev = r10_bio->mddev;
211         conf_t *conf = mddev_to_conf(mddev);
212
213         spin_lock_irqsave(&conf->device_lock, flags);
214         list_add(&r10_bio->retry_list, &conf->retry_list);
215         conf->nr_queued ++;
216         spin_unlock_irqrestore(&conf->device_lock, flags);
217
218         md_wakeup_thread(mddev->thread);
219 }
220
221 /*
222  * raid_end_bio_io() is called when we have finished servicing a mirrored
223  * operation and are ready to return a success/failure code to the buffer
224  * cache layer.
225  */
226 static void raid_end_bio_io(r10bio_t *r10_bio)
227 {
228         struct bio *bio = r10_bio->master_bio;
229
230         bio_endio(bio,
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
794         if (unlikely(bio_barrier(bio))) {
795                 bio_endio(bio, -EOPNOTSUPP);
796                 return 0;
797         }
798
799         /* If this request crosses a chunk boundary, we need to
800          * split it.  This will only happen for 1 PAGE (or less) requests.
801          */
802         if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
803                       > chunk_sects &&
804                     conf->near_copies < conf->raid_disks)) {
805                 struct bio_pair *bp;
806                 /* Sanity check -- queue functions should prevent this happening */
807                 if (bio->bi_vcnt != 1 ||
808                     bio->bi_idx != 0)
809                         goto bad_map;
810                 /* This is a one page bio that upper layers
811                  * refuse to split for us, so we need to split it.
812                  */
813                 bp = bio_split(bio, bio_split_pool,
814                                chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
815                 if (make_request(q, &bp->bio1))
816                         generic_make_request(&bp->bio1);
817                 if (make_request(q, &bp->bio2))
818                         generic_make_request(&bp->bio2);
819
820                 bio_pair_release(bp);
821                 return 0;
822         bad_map:
823                 printk("raid10_make_request bug: can't convert block across chunks"
824                        " or bigger than %dk %llu %d\n", chunk_sects/2,
825                        (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
826
827                 bio_io_error(bio);
828                 return 0;
829         }
830
831         md_write_start(mddev, bio);
832
833         /*
834          * Register the new request and wait if the reconstruction
835          * thread has put up a bar for new requests.
836          * Continue immediately if no resync is active currently.
837          */
838         wait_barrier(conf);
839
840         disk_stat_inc(mddev->gendisk, ios[rw]);
841         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
842
843         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
844
845         r10_bio->master_bio = bio;
846         r10_bio->sectors = bio->bi_size >> 9;
847
848         r10_bio->mddev = mddev;
849         r10_bio->sector = bio->bi_sector;
850         r10_bio->state = 0;
851
852         if (rw == READ) {
853                 /*
854                  * read balancing logic:
855                  */
856                 int disk = read_balance(conf, r10_bio);
857                 int slot = r10_bio->read_slot;
858                 if (disk < 0) {
859                         raid_end_bio_io(r10_bio);
860                         return 0;
861                 }
862                 mirror = conf->mirrors + disk;
863
864                 read_bio = bio_clone(bio, GFP_NOIO);
865
866                 r10_bio->devs[slot].bio = read_bio;
867
868                 read_bio->bi_sector = r10_bio->devs[slot].addr +
869                         mirror->rdev->data_offset;
870                 read_bio->bi_bdev = mirror->rdev->bdev;
871                 read_bio->bi_end_io = raid10_end_read_request;
872                 read_bio->bi_rw = READ | do_sync;
873                 read_bio->bi_private = r10_bio;
874
875                 generic_make_request(read_bio);
876                 return 0;
877         }
878
879         /*
880          * WRITE:
881          */
882         /* first select target devices under spinlock and
883          * inc refcount on their rdev.  Record them by setting
884          * bios[x] to bio
885          */
886         raid10_find_phys(conf, r10_bio);
887         rcu_read_lock();
888         for (i = 0;  i < conf->copies; i++) {
889                 int d = r10_bio->devs[i].devnum;
890                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
891                 if (rdev &&
892                     !test_bit(Faulty, &rdev->flags)) {
893                         atomic_inc(&rdev->nr_pending);
894                         r10_bio->devs[i].bio = bio;
895                 } else {
896                         r10_bio->devs[i].bio = NULL;
897                         set_bit(R10BIO_Degraded, &r10_bio->state);
898                 }
899         }
900         rcu_read_unlock();
901
902         atomic_set(&r10_bio->remaining, 0);
903
904         bio_list_init(&bl);
905         for (i = 0; i < conf->copies; i++) {
906                 struct bio *mbio;
907                 int d = r10_bio->devs[i].devnum;
908                 if (!r10_bio->devs[i].bio)
909                         continue;
910
911                 mbio = bio_clone(bio, GFP_NOIO);
912                 r10_bio->devs[i].bio = mbio;
913
914                 mbio->bi_sector = r10_bio->devs[i].addr+
915                         conf->mirrors[d].rdev->data_offset;
916                 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
917                 mbio->bi_end_io = raid10_end_write_request;
918                 mbio->bi_rw = WRITE | do_sync;
919                 mbio->bi_private = r10_bio;
920
921                 atomic_inc(&r10_bio->remaining);
922                 bio_list_add(&bl, mbio);
923         }
924
925         if (unlikely(!atomic_read(&r10_bio->remaining))) {
926                 /* the array is dead */
927                 md_write_end(mddev);
928                 raid_end_bio_io(r10_bio);
929                 return 0;
930         }
931
932         bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
933         spin_lock_irqsave(&conf->device_lock, flags);
934         bio_list_merge(&conf->pending_bio_list, &bl);
935         blk_plug_device(mddev->queue);
936         spin_unlock_irqrestore(&conf->device_lock, flags);
937
938         /* In case raid10d snuck in to freeze_array */
939         wake_up(&conf->wait_barrier);
940
941         if (do_sync)
942                 md_wakeup_thread(mddev->thread);
943
944         return 0;
945 }
946
947 static void status(struct seq_file *seq, mddev_t *mddev)
948 {
949         conf_t *conf = mddev_to_conf(mddev);
950         int i;
951
952         if (conf->near_copies < conf->raid_disks)
953                 seq_printf(seq, " %dK chunks", mddev->chunk_size/1024);
954         if (conf->near_copies > 1)
955                 seq_printf(seq, " %d near-copies", conf->near_copies);
956         if (conf->far_copies > 1) {
957                 if (conf->far_offset)
958                         seq_printf(seq, " %d offset-copies", conf->far_copies);
959                 else
960                         seq_printf(seq, " %d far-copies", conf->far_copies);
961         }
962         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
963                                         conf->raid_disks - mddev->degraded);
964         for (i = 0; i < conf->raid_disks; i++)
965                 seq_printf(seq, "%s",
966                               conf->mirrors[i].rdev &&
967                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
968         seq_printf(seq, "]");
969 }
970
971 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
972 {
973         char b[BDEVNAME_SIZE];
974         conf_t *conf = mddev_to_conf(mddev);
975
976         /*
977          * If it is not operational, then we have already marked it as dead
978          * else if it is the last working disks, ignore the error, let the
979          * next level up know.
980          * else mark the drive as failed
981          */
982         if (test_bit(In_sync, &rdev->flags)
983             && conf->raid_disks-mddev->degraded == 1)
984                 /*
985                  * Don't fail the drive, just return an IO error.
986                  * The test should really be more sophisticated than
987                  * "working_disks == 1", but it isn't critical, and
988                  * can wait until we do more sophisticated "is the drive
989                  * really dead" tests...
990                  */
991                 return;
992         if (test_and_clear_bit(In_sync, &rdev->flags)) {
993                 unsigned long flags;
994                 spin_lock_irqsave(&conf->device_lock, flags);
995                 mddev->degraded++;
996                 spin_unlock_irqrestore(&conf->device_lock, flags);
997                 /*
998                  * if recovery is running, make sure it aborts.
999                  */
1000                 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
1001         }
1002         set_bit(Faulty, &rdev->flags);
1003         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1004         printk(KERN_ALERT "raid10: Disk failure on %s, disabling device. \n"
1005                 "       Operation continuing on %d devices\n",
1006                 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1007 }
1008
1009 static void print_conf(conf_t *conf)
1010 {
1011         int i;
1012         mirror_info_t *tmp;
1013
1014         printk("RAID10 conf printout:\n");
1015         if (!conf) {
1016                 printk("(!conf)\n");
1017                 return;
1018         }
1019         printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1020                 conf->raid_disks);
1021
1022         for (i = 0; i < conf->raid_disks; i++) {
1023                 char b[BDEVNAME_SIZE];
1024                 tmp = conf->mirrors + i;
1025                 if (tmp->rdev)
1026                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
1027                                 i, !test_bit(In_sync, &tmp->rdev->flags),
1028                                 !test_bit(Faulty, &tmp->rdev->flags),
1029                                 bdevname(tmp->rdev->bdev,b));
1030         }
1031 }
1032
1033 static void close_sync(conf_t *conf)
1034 {
1035         wait_barrier(conf);
1036         allow_barrier(conf);
1037
1038         mempool_destroy(conf->r10buf_pool);
1039         conf->r10buf_pool = NULL;
1040 }
1041
1042 /* check if there are enough drives for
1043  * every block to appear on atleast one
1044  */
1045 static int enough(conf_t *conf)
1046 {
1047         int first = 0;
1048
1049         do {
1050                 int n = conf->copies;
1051                 int cnt = 0;
1052                 while (n--) {
1053                         if (conf->mirrors[first].rdev)
1054                                 cnt++;
1055                         first = (first+1) % conf->raid_disks;
1056                 }
1057                 if (cnt == 0)
1058                         return 0;
1059         } while (first != 0);
1060         return 1;
1061 }
1062
1063 static int raid10_spare_active(mddev_t *mddev)
1064 {
1065         int i;
1066         conf_t *conf = mddev->private;
1067         mirror_info_t *tmp;
1068
1069         /*
1070          * Find all non-in_sync disks within the RAID10 configuration
1071          * and mark them in_sync
1072          */
1073         for (i = 0; i < conf->raid_disks; i++) {
1074                 tmp = conf->mirrors + i;
1075                 if (tmp->rdev
1076                     && !test_bit(Faulty, &tmp->rdev->flags)
1077                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1078                         unsigned long flags;
1079                         spin_lock_irqsave(&conf->device_lock, flags);
1080                         mddev->degraded--;
1081                         spin_unlock_irqrestore(&conf->device_lock, flags);
1082                 }
1083         }
1084
1085         print_conf(conf);
1086         return 0;
1087 }
1088
1089
1090 static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1091 {
1092         conf_t *conf = mddev->private;
1093         int found = 0;
1094         int mirror;
1095         mirror_info_t *p;
1096
1097         if (mddev->recovery_cp < MaxSector)
1098                 /* only hot-add to in-sync arrays, as recovery is
1099                  * very different from resync
1100                  */
1101                 return 0;
1102         if (!enough(conf))
1103                 return 0;
1104
1105         if (rdev->saved_raid_disk >= 0 &&
1106             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1107                 mirror = rdev->saved_raid_disk;
1108         else
1109                 mirror = 0;
1110         for ( ; mirror < mddev->raid_disks; mirror++)
1111                 if ( !(p=conf->mirrors+mirror)->rdev) {
1112
1113                         blk_queue_stack_limits(mddev->queue,
1114                                                rdev->bdev->bd_disk->queue);
1115                         /* as we don't honour merge_bvec_fn, we must never risk
1116                          * violating it, so limit ->max_sector to one PAGE, as
1117                          * a one page request is never in violation.
1118                          */
1119                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1120                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
1121                                 mddev->queue->max_sectors = (PAGE_SIZE>>9);
1122
1123                         p->head_position = 0;
1124                         rdev->raid_disk = mirror;
1125                         found = 1;
1126                         if (rdev->saved_raid_disk != mirror)
1127                                 conf->fullsync = 1;
1128                         rcu_assign_pointer(p->rdev, rdev);
1129                         break;
1130                 }
1131
1132         print_conf(conf);
1133         return found;
1134 }
1135
1136 static int raid10_remove_disk(mddev_t *mddev, int number)
1137 {
1138         conf_t *conf = mddev->private;
1139         int err = 0;
1140         mdk_rdev_t *rdev;
1141         mirror_info_t *p = conf->mirrors+ number;
1142
1143         print_conf(conf);
1144         rdev = p->rdev;
1145         if (rdev) {
1146                 if (test_bit(In_sync, &rdev->flags) ||
1147                     atomic_read(&rdev->nr_pending)) {
1148                         err = -EBUSY;
1149                         goto abort;
1150                 }
1151                 p->rdev = NULL;
1152                 synchronize_rcu();
1153                 if (atomic_read(&rdev->nr_pending)) {
1154                         /* lost the race, try later */
1155                         err = -EBUSY;
1156                         p->rdev = rdev;
1157                 }
1158         }
1159 abort:
1160
1161         print_conf(conf);
1162         return err;
1163 }
1164
1165
1166 static void end_sync_read(struct bio *bio, int error)
1167 {
1168         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1169         conf_t *conf = mddev_to_conf(r10_bio->mddev);
1170         int i,d;
1171
1172         for (i=0; i<conf->copies; i++)
1173                 if (r10_bio->devs[i].bio == bio)
1174                         break;
1175         BUG_ON(i == conf->copies);
1176         update_head_pos(i, r10_bio);
1177         d = r10_bio->devs[i].devnum;
1178
1179         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1180                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1181         else {
1182                 atomic_add(r10_bio->sectors,
1183                            &conf->mirrors[d].rdev->corrected_errors);
1184                 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1185                         md_error(r10_bio->mddev,
1186                                  conf->mirrors[d].rdev);
1187         }
1188
1189         /* for reconstruct, we always reschedule after a read.
1190          * for resync, only after all reads
1191          */
1192         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1193             atomic_dec_and_test(&r10_bio->remaining)) {
1194                 /* we have read all the blocks,
1195                  * do the comparison in process context in raid10d
1196                  */
1197                 reschedule_retry(r10_bio);
1198         }
1199         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1200 }
1201
1202 static void end_sync_write(struct bio *bio, int error)
1203 {
1204         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1205         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1206         mddev_t *mddev = r10_bio->mddev;
1207         conf_t *conf = mddev_to_conf(mddev);
1208         int i,d;
1209
1210         for (i = 0; i < conf->copies; i++)
1211                 if (r10_bio->devs[i].bio == bio)
1212                         break;
1213         d = r10_bio->devs[i].devnum;
1214
1215         if (!uptodate)
1216                 md_error(mddev, conf->mirrors[d].rdev);
1217         update_head_pos(i, r10_bio);
1218
1219         while (atomic_dec_and_test(&r10_bio->remaining)) {
1220                 if (r10_bio->master_bio == NULL) {
1221                         /* the primary of several recovery bios */
1222                         md_done_sync(mddev, r10_bio->sectors, 1);
1223                         put_buf(r10_bio);
1224                         break;
1225                 } else {
1226                         r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1227                         put_buf(r10_bio);
1228                         r10_bio = r10_bio2;
1229                 }
1230         }
1231         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1232 }
1233
1234 /*
1235  * Note: sync and recover and handled very differently for raid10
1236  * This code is for resync.
1237  * For resync, we read through virtual addresses and read all blocks.
1238  * If there is any error, we schedule a write.  The lowest numbered
1239  * drive is authoritative.
1240  * However requests come for physical address, so we need to map.
1241  * For every physical address there are raid_disks/copies virtual addresses,
1242  * which is always are least one, but is not necessarly an integer.
1243  * This means that a physical address can span multiple chunks, so we may
1244  * have to submit multiple io requests for a single sync request.
1245  */
1246 /*
1247  * We check if all blocks are in-sync and only write to blocks that
1248  * aren't in sync
1249  */
1250 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1251 {
1252         conf_t *conf = mddev_to_conf(mddev);
1253         int i, first;
1254         struct bio *tbio, *fbio;
1255
1256         atomic_set(&r10_bio->remaining, 1);
1257
1258         /* find the first device with a block */
1259         for (i=0; i<conf->copies; i++)
1260                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1261                         break;
1262
1263         if (i == conf->copies)
1264                 goto done;
1265
1266         first = i;
1267         fbio = r10_bio->devs[i].bio;
1268
1269         /* now find blocks with errors */
1270         for (i=0 ; i < conf->copies ; i++) {
1271                 int  j, d;
1272                 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1273
1274                 tbio = r10_bio->devs[i].bio;
1275
1276                 if (tbio->bi_end_io != end_sync_read)
1277                         continue;
1278                 if (i == first)
1279                         continue;
1280                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1281                         /* We know that the bi_io_vec layout is the same for
1282                          * both 'first' and 'i', so we just compare them.
1283                          * All vec entries are PAGE_SIZE;
1284                          */
1285                         for (j = 0; j < vcnt; j++)
1286                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1287                                            page_address(tbio->bi_io_vec[j].bv_page),
1288                                            PAGE_SIZE))
1289                                         break;
1290                         if (j == vcnt)
1291                                 continue;
1292                         mddev->resync_mismatches += r10_bio->sectors;
1293                 }
1294                 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1295                         /* Don't fix anything. */
1296                         continue;
1297                 /* Ok, we need to write this bio
1298                  * First we need to fixup bv_offset, bv_len and
1299                  * bi_vecs, as the read request might have corrupted these
1300                  */
1301                 tbio->bi_vcnt = vcnt;
1302                 tbio->bi_size = r10_bio->sectors << 9;
1303                 tbio->bi_idx = 0;
1304                 tbio->bi_phys_segments = 0;
1305                 tbio->bi_hw_segments = 0;
1306                 tbio->bi_hw_front_size = 0;
1307                 tbio->bi_hw_back_size = 0;
1308                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1309                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1310                 tbio->bi_next = NULL;
1311                 tbio->bi_rw = WRITE;
1312                 tbio->bi_private = r10_bio;
1313                 tbio->bi_sector = r10_bio->devs[i].addr;
1314
1315                 for (j=0; j < vcnt ; j++) {
1316                         tbio->bi_io_vec[j].bv_offset = 0;
1317                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1318
1319                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1320                                page_address(fbio->bi_io_vec[j].bv_page),
1321                                PAGE_SIZE);
1322                 }
1323                 tbio->bi_end_io = end_sync_write;
1324
1325                 d = r10_bio->devs[i].devnum;
1326                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1327                 atomic_inc(&r10_bio->remaining);
1328                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1329
1330                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1331                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1332                 generic_make_request(tbio);
1333         }
1334
1335 done:
1336         if (atomic_dec_and_test(&r10_bio->remaining)) {
1337                 md_done_sync(mddev, r10_bio->sectors, 1);
1338                 put_buf(r10_bio);
1339         }
1340 }
1341
1342 /*
1343  * Now for the recovery code.
1344  * Recovery happens across physical sectors.
1345  * We recover all non-is_sync drives by finding the virtual address of
1346  * each, and then choose a working drive that also has that virt address.
1347  * There is a separate r10_bio for each non-in_sync drive.
1348  * Only the first two slots are in use. The first for reading,
1349  * The second for writing.
1350  *
1351  */
1352
1353 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1354 {
1355         conf_t *conf = mddev_to_conf(mddev);
1356         int i, d;
1357         struct bio *bio, *wbio;
1358
1359
1360         /* move the pages across to the second bio
1361          * and submit the write request
1362          */
1363         bio = r10_bio->devs[0].bio;
1364         wbio = r10_bio->devs[1].bio;
1365         for (i=0; i < wbio->bi_vcnt; i++) {
1366                 struct page *p = bio->bi_io_vec[i].bv_page;
1367                 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1368                 wbio->bi_io_vec[i].bv_page = p;
1369         }
1370         d = r10_bio->devs[1].devnum;
1371
1372         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1373         md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1374         if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1375                 generic_make_request(wbio);
1376         else
1377                 bio_endio(wbio, -EIO);
1378 }
1379
1380
1381 /*
1382  * This is a kernel thread which:
1383  *
1384  *      1.      Retries failed read operations on working mirrors.
1385  *      2.      Updates the raid superblock when problems encounter.
1386  *      3.      Performs writes following reads for array synchronising.
1387  */
1388
1389 static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1390 {
1391         int sect = 0; /* Offset from r10_bio->sector */
1392         int sectors = r10_bio->sectors;
1393         mdk_rdev_t*rdev;
1394         while(sectors) {
1395                 int s = sectors;
1396                 int sl = r10_bio->read_slot;
1397                 int success = 0;
1398                 int start;
1399
1400                 if (s > (PAGE_SIZE>>9))
1401                         s = PAGE_SIZE >> 9;
1402
1403                 rcu_read_lock();
1404                 do {
1405                         int d = r10_bio->devs[sl].devnum;
1406                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1407                         if (rdev &&
1408                             test_bit(In_sync, &rdev->flags)) {
1409                                 atomic_inc(&rdev->nr_pending);
1410                                 rcu_read_unlock();
1411                                 success = sync_page_io(rdev->bdev,
1412                                                        r10_bio->devs[sl].addr +
1413                                                        sect + rdev->data_offset,
1414                                                        s<<9,
1415                                                        conf->tmppage, READ);
1416                                 rdev_dec_pending(rdev, mddev);
1417                                 rcu_read_lock();
1418                                 if (success)
1419                                         break;
1420                         }
1421                         sl++;
1422                         if (sl == conf->copies)
1423                                 sl = 0;
1424                 } while (!success && sl != r10_bio->read_slot);
1425                 rcu_read_unlock();
1426
1427                 if (!success) {
1428                         /* Cannot read from anywhere -- bye bye array */
1429                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1430                         md_error(mddev, conf->mirrors[dn].rdev);
1431                         break;
1432                 }
1433
1434                 start = sl;
1435                 /* write it back and re-read */
1436                 rcu_read_lock();
1437                 while (sl != r10_bio->read_slot) {
1438                         int d;
1439                         if (sl==0)
1440                                 sl = conf->copies;
1441                         sl--;
1442                         d = r10_bio->devs[sl].devnum;
1443                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1444                         if (rdev &&
1445                             test_bit(In_sync, &rdev->flags)) {
1446                                 atomic_inc(&rdev->nr_pending);
1447                                 rcu_read_unlock();
1448                                 atomic_add(s, &rdev->corrected_errors);
1449                                 if (sync_page_io(rdev->bdev,
1450                                                  r10_bio->devs[sl].addr +
1451                                                  sect + rdev->data_offset,
1452                                                  s<<9, conf->tmppage, WRITE)
1453                                     == 0)
1454                                         /* Well, this device is dead */
1455                                         md_error(mddev, rdev);
1456                                 rdev_dec_pending(rdev, mddev);
1457                                 rcu_read_lock();
1458                         }
1459                 }
1460                 sl = start;
1461                 while (sl != r10_bio->read_slot) {
1462                         int d;
1463                         if (sl==0)
1464                                 sl = conf->copies;
1465                         sl--;
1466                         d = r10_bio->devs[sl].devnum;
1467                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1468                         if (rdev &&
1469                             test_bit(In_sync, &rdev->flags)) {
1470                                 char b[BDEVNAME_SIZE];
1471                                 atomic_inc(&rdev->nr_pending);
1472                                 rcu_read_unlock();
1473                                 if (sync_page_io(rdev->bdev,
1474                                                  r10_bio->devs[sl].addr +
1475                                                  sect + rdev->data_offset,
1476                                                  s<<9, conf->tmppage, READ) == 0)
1477                                         /* Well, this device is dead */
1478                                         md_error(mddev, rdev);
1479                                 else
1480                                         printk(KERN_INFO
1481                                                "raid10:%s: read error corrected"
1482                                                " (%d sectors at %llu on %s)\n",
1483                                                mdname(mddev), s,
1484                                                (unsigned long long)(sect+
1485                                                     rdev->data_offset),
1486                                                bdevname(rdev->bdev, b));
1487
1488                                 rdev_dec_pending(rdev, mddev);
1489                                 rcu_read_lock();
1490                         }
1491                 }
1492                 rcu_read_unlock();
1493
1494                 sectors -= s;
1495                 sect += s;
1496         }
1497 }
1498
1499 static void raid10d(mddev_t *mddev)
1500 {
1501         r10bio_t *r10_bio;
1502         struct bio *bio;
1503         unsigned long flags;
1504         conf_t *conf = mddev_to_conf(mddev);
1505         struct list_head *head = &conf->retry_list;
1506         int unplug=0;
1507         mdk_rdev_t *rdev;
1508
1509         md_check_recovery(mddev);
1510
1511         for (;;) {
1512                 char b[BDEVNAME_SIZE];
1513
1514                 unplug += flush_pending_writes(conf);
1515
1516                 spin_lock_irqsave(&conf->device_lock, flags);
1517                 if (list_empty(head)) {
1518                         spin_unlock_irqrestore(&conf->device_lock, flags);
1519                         break;
1520                 }
1521                 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1522                 list_del(head->prev);
1523                 conf->nr_queued--;
1524                 spin_unlock_irqrestore(&conf->device_lock, flags);
1525
1526                 mddev = r10_bio->mddev;
1527                 conf = mddev_to_conf(mddev);
1528                 if (test_bit(R10BIO_IsSync, &r10_bio->state)) {
1529                         sync_request_write(mddev, r10_bio);
1530                         unplug = 1;
1531                 } else  if (test_bit(R10BIO_IsRecover, &r10_bio->state)) {
1532                         recovery_request_write(mddev, r10_bio);
1533                         unplug = 1;
1534                 } else {
1535                         int mirror;
1536                         /* we got a read error. Maybe the drive is bad.  Maybe just
1537                          * the block and we can fix it.
1538                          * We freeze all other IO, and try reading the block from
1539                          * other devices.  When we find one, we re-write
1540                          * and check it that fixes the read error.
1541                          * This is all done synchronously while the array is
1542                          * frozen.
1543                          */
1544                         if (mddev->ro == 0) {
1545                                 freeze_array(conf);
1546                                 fix_read_error(conf, mddev, r10_bio);
1547                                 unfreeze_array(conf);
1548                         }
1549
1550                         bio = r10_bio->devs[r10_bio->read_slot].bio;
1551                         r10_bio->devs[r10_bio->read_slot].bio =
1552                                 mddev->ro ? IO_BLOCKED : NULL;
1553                         mirror = read_balance(conf, r10_bio);
1554                         if (mirror == -1) {
1555                                 printk(KERN_ALERT "raid10: %s: unrecoverable I/O"
1556                                        " read error for block %llu\n",
1557                                        bdevname(bio->bi_bdev,b),
1558                                        (unsigned long long)r10_bio->sector);
1559                                 raid_end_bio_io(r10_bio);
1560                                 bio_put(bio);
1561                         } else {
1562                                 const int do_sync = bio_sync(r10_bio->master_bio);
1563                                 bio_put(bio);
1564                                 rdev = conf->mirrors[mirror].rdev;
1565                                 if (printk_ratelimit())
1566                                         printk(KERN_ERR "raid10: %s: redirecting sector %llu to"
1567                                                " another mirror\n",
1568                                                bdevname(rdev->bdev,b),
1569                                                (unsigned long long)r10_bio->sector);
1570                                 bio = bio_clone(r10_bio->master_bio, GFP_NOIO);
1571                                 r10_bio->devs[r10_bio->read_slot].bio = bio;
1572                                 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1573                                         + rdev->data_offset;
1574                                 bio->bi_bdev = rdev->bdev;
1575                                 bio->bi_rw = READ | do_sync;
1576                                 bio->bi_private = r10_bio;
1577                                 bio->bi_end_io = raid10_end_read_request;
1578                                 unplug = 1;
1579                                 generic_make_request(bio);
1580                         }
1581                 }
1582         }
1583         if (unplug)
1584                 unplug_slaves(mddev);
1585 }
1586
1587
1588 static int init_resync(conf_t *conf)
1589 {
1590         int buffs;
1591
1592         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1593         BUG_ON(conf->r10buf_pool);
1594         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1595         if (!conf->r10buf_pool)
1596                 return -ENOMEM;
1597         conf->next_resync = 0;
1598         return 0;
1599 }
1600
1601 /*
1602  * perform a "sync" on one "block"
1603  *
1604  * We need to make sure that no normal I/O request - particularly write
1605  * requests - conflict with active sync requests.
1606  *
1607  * This is achieved by tracking pending requests and a 'barrier' concept
1608  * that can be installed to exclude normal IO requests.
1609  *
1610  * Resync and recovery are handled very differently.
1611  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1612  *
1613  * For resync, we iterate over virtual addresses, read all copies,
1614  * and update if there are differences.  If only one copy is live,
1615  * skip it.
1616  * For recovery, we iterate over physical addresses, read a good
1617  * value for each non-in_sync drive, and over-write.
1618  *
1619  * So, for recovery we may have several outstanding complex requests for a
1620  * given address, one for each out-of-sync device.  We model this by allocating
1621  * a number of r10_bio structures, one for each out-of-sync device.
1622  * As we setup these structures, we collect all bio's together into a list
1623  * which we then process collectively to add pages, and then process again
1624  * to pass to generic_make_request.
1625  *
1626  * The r10_bio structures are linked using a borrowed master_bio pointer.
1627  * This link is counted in ->remaining.  When the r10_bio that points to NULL
1628  * has its remaining count decremented to 0, the whole complex operation
1629  * is complete.
1630  *
1631  */
1632
1633 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1634 {
1635         conf_t *conf = mddev_to_conf(mddev);
1636         r10bio_t *r10_bio;
1637         struct bio *biolist = NULL, *bio;
1638         sector_t max_sector, nr_sectors;
1639         int disk;
1640         int i;
1641         int max_sync;
1642         int sync_blocks;
1643
1644         sector_t sectors_skipped = 0;
1645         int chunks_skipped = 0;
1646
1647         if (!conf->r10buf_pool)
1648                 if (init_resync(conf))
1649                         return 0;
1650
1651  skipped:
1652         max_sector = mddev->size << 1;
1653         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1654                 max_sector = mddev->resync_max_sectors;
1655         if (sector_nr >= max_sector) {
1656                 /* If we aborted, we need to abort the
1657                  * sync on the 'current' bitmap chucks (there can
1658                  * be several when recovering multiple devices).
1659                  * as we may have started syncing it but not finished.
1660                  * We can find the current address in
1661                  * mddev->curr_resync, but for recovery,
1662                  * we need to convert that to several
1663                  * virtual addresses.
1664                  */
1665                 if (mddev->curr_resync < max_sector) { /* aborted */
1666                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1667                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1668                                                 &sync_blocks, 1);
1669                         else for (i=0; i<conf->raid_disks; i++) {
1670                                 sector_t sect =
1671                                         raid10_find_virt(conf, mddev->curr_resync, i);
1672                                 bitmap_end_sync(mddev->bitmap, sect,
1673                                                 &sync_blocks, 1);
1674                         }
1675                 } else /* completed sync */
1676                         conf->fullsync = 0;
1677
1678                 bitmap_close_sync(mddev->bitmap);
1679                 close_sync(conf);
1680                 *skipped = 1;
1681                 return sectors_skipped;
1682         }
1683         if (chunks_skipped >= conf->raid_disks) {
1684                 /* if there has been nothing to do on any drive,
1685                  * then there is nothing to do at all..
1686                  */
1687                 *skipped = 1;
1688                 return (max_sector - sector_nr) + sectors_skipped;
1689         }
1690
1691         if (max_sector > mddev->resync_max)
1692                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1693
1694         /* make sure whole request will fit in a chunk - if chunks
1695          * are meaningful
1696          */
1697         if (conf->near_copies < conf->raid_disks &&
1698             max_sector > (sector_nr | conf->chunk_mask))
1699                 max_sector = (sector_nr | conf->chunk_mask) + 1;
1700         /*
1701          * If there is non-resync activity waiting for us then
1702          * put in a delay to throttle resync.
1703          */
1704         if (!go_faster && conf->nr_waiting)
1705                 msleep_interruptible(1000);
1706
1707         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1708
1709         /* Again, very different code for resync and recovery.
1710          * Both must result in an r10bio with a list of bios that
1711          * have bi_end_io, bi_sector, bi_bdev set,
1712          * and bi_private set to the r10bio.
1713          * For recovery, we may actually create several r10bios
1714          * with 2 bios in each, that correspond to the bios in the main one.
1715          * In this case, the subordinate r10bios link back through a
1716          * borrowed master_bio pointer, and the counter in the master
1717          * includes a ref from each subordinate.
1718          */
1719         /* First, we decide what to do and set ->bi_end_io
1720          * To end_sync_read if we want to read, and
1721          * end_sync_write if we will want to write.
1722          */
1723
1724         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
1725         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1726                 /* recovery... the complicated one */
1727                 int i, j, k;
1728                 r10_bio = NULL;
1729
1730                 for (i=0 ; i<conf->raid_disks; i++)
1731                         if (conf->mirrors[i].rdev &&
1732                             !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) {
1733                                 int still_degraded = 0;
1734                                 /* want to reconstruct this device */
1735                                 r10bio_t *rb2 = r10_bio;
1736                                 sector_t sect = raid10_find_virt(conf, sector_nr, i);
1737                                 int must_sync;
1738                                 /* Unless we are doing a full sync, we only need
1739                                  * to recover the block if it is set in the bitmap
1740                                  */
1741                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1742                                                               &sync_blocks, 1);
1743                                 if (sync_blocks < max_sync)
1744                                         max_sync = sync_blocks;
1745                                 if (!must_sync &&
1746                                     !conf->fullsync) {
1747                                         /* yep, skip the sync_blocks here, but don't assume
1748                                          * that there will never be anything to do here
1749                                          */
1750                                         chunks_skipped = -1;
1751                                         continue;
1752                                 }
1753
1754                                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1755                                 raise_barrier(conf, rb2 != NULL);
1756                                 atomic_set(&r10_bio->remaining, 0);
1757
1758                                 r10_bio->master_bio = (struct bio*)rb2;
1759                                 if (rb2)
1760                                         atomic_inc(&rb2->remaining);
1761                                 r10_bio->mddev = mddev;
1762                                 set_bit(R10BIO_IsRecover, &r10_bio->state);
1763                                 r10_bio->sector = sect;
1764
1765                                 raid10_find_phys(conf, r10_bio);
1766                                 /* Need to check if this section will still be
1767                                  * degraded
1768                                  */
1769                                 for (j=0; j<conf->copies;j++) {
1770                                         int d = r10_bio->devs[j].devnum;
1771                                         if (conf->mirrors[d].rdev == NULL ||
1772                                             test_bit(Faulty, &conf->mirrors[d].rdev->flags)) {
1773                                                 still_degraded = 1;
1774                                                 break;
1775                                         }
1776                                 }
1777                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1778                                                               &sync_blocks, still_degraded);
1779
1780                                 for (j=0; j<conf->copies;j++) {
1781                                         int d = r10_bio->devs[j].devnum;
1782                                         if (conf->mirrors[d].rdev &&
1783                                             test_bit(In_sync, &conf->mirrors[d].rdev->flags)) {
1784                                                 /* This is where we read from */
1785                                                 bio = r10_bio->devs[0].bio;
1786                                                 bio->bi_next = biolist;
1787                                                 biolist = bio;
1788                                                 bio->bi_private = r10_bio;
1789                                                 bio->bi_end_io = end_sync_read;
1790                                                 bio->bi_rw = READ;
1791                                                 bio->bi_sector = r10_bio->devs[j].addr +
1792                                                         conf->mirrors[d].rdev->data_offset;
1793                                                 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1794                                                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1795                                                 atomic_inc(&r10_bio->remaining);
1796                                                 /* and we write to 'i' */
1797
1798                                                 for (k=0; k<conf->copies; k++)
1799                                                         if (r10_bio->devs[k].devnum == i)
1800                                                                 break;
1801                                                 BUG_ON(k == conf->copies);
1802                                                 bio = r10_bio->devs[1].bio;
1803                                                 bio->bi_next = biolist;
1804                                                 biolist = bio;
1805                                                 bio->bi_private = r10_bio;
1806                                                 bio->bi_end_io = end_sync_write;
1807                                                 bio->bi_rw = WRITE;
1808                                                 bio->bi_sector = r10_bio->devs[k].addr +
1809                                                         conf->mirrors[i].rdev->data_offset;
1810                                                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1811
1812                                                 r10_bio->devs[0].devnum = d;
1813                                                 r10_bio->devs[1].devnum = i;
1814
1815                                                 break;
1816                                         }
1817                                 }
1818                                 if (j == conf->copies) {
1819                                         /* Cannot recover, so abort the recovery */
1820                                         put_buf(r10_bio);
1821                                         if (rb2)
1822                                                 atomic_dec(&rb2->remaining);
1823                                         r10_bio = rb2;
1824                                         if (!test_and_set_bit(MD_RECOVERY_ERR, &mddev->recovery))
1825                                                 printk(KERN_INFO "raid10: %s: insufficient working devices for recovery.\n",
1826                                                        mdname(mddev));
1827                                         break;
1828                                 }
1829                         }
1830                 if (biolist == NULL) {
1831                         while (r10_bio) {
1832                                 r10bio_t *rb2 = r10_bio;
1833                                 r10_bio = (r10bio_t*) rb2->master_bio;
1834                                 rb2->master_bio = NULL;
1835                                 put_buf(rb2);
1836                         }
1837                         goto giveup;
1838                 }
1839         } else {
1840                 /* resync. Schedule a read for every block at this virt offset */
1841                 int count = 0;
1842
1843                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1844                                        &sync_blocks, mddev->degraded) &&
1845                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1846                         /* We can skip this block */
1847                         *skipped = 1;
1848                         return sync_blocks + sectors_skipped;
1849                 }
1850                 if (sync_blocks < max_sync)
1851                         max_sync = sync_blocks;
1852                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1853
1854                 r10_bio->mddev = mddev;
1855                 atomic_set(&r10_bio->remaining, 0);
1856                 raise_barrier(conf, 0);
1857                 conf->next_resync = sector_nr;
1858
1859                 r10_bio->master_bio = NULL;
1860                 r10_bio->sector = sector_nr;
1861                 set_bit(R10BIO_IsSync, &r10_bio->state);
1862                 raid10_find_phys(conf, r10_bio);
1863                 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1864
1865                 for (i=0; i<conf->copies; i++) {
1866                         int d = r10_bio->devs[i].devnum;
1867                         bio = r10_bio->devs[i].bio;
1868                         bio->bi_end_io = NULL;
1869                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
1870                         if (conf->mirrors[d].rdev == NULL ||
1871                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
1872                                 continue;
1873                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1874                         atomic_inc(&r10_bio->remaining);
1875                         bio->bi_next = biolist;
1876                         biolist = bio;
1877                         bio->bi_private = r10_bio;
1878                         bio->bi_end_io = end_sync_read;
1879                         bio->bi_rw = READ;
1880                         bio->bi_sector = r10_bio->devs[i].addr +
1881                                 conf->mirrors[d].rdev->data_offset;
1882                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1883                         count++;
1884                 }
1885
1886                 if (count < 2) {
1887                         for (i=0; i<conf->copies; i++) {
1888                                 int d = r10_bio->devs[i].devnum;
1889                                 if (r10_bio->devs[i].bio->bi_end_io)
1890                                         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1891                         }
1892                         put_buf(r10_bio);
1893                         biolist = NULL;
1894                         goto giveup;
1895                 }
1896         }
1897
1898         for (bio = biolist; bio ; bio=bio->bi_next) {
1899
1900                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1901                 if (bio->bi_end_io)
1902                         bio->bi_flags |= 1 << BIO_UPTODATE;
1903                 bio->bi_vcnt = 0;
1904                 bio->bi_idx = 0;
1905                 bio->bi_phys_segments = 0;
1906                 bio->bi_hw_segments = 0;
1907                 bio->bi_size = 0;
1908         }
1909
1910         nr_sectors = 0;
1911         if (sector_nr + max_sync < max_sector)
1912                 max_sector = sector_nr + max_sync;
1913         do {
1914                 struct page *page;
1915                 int len = PAGE_SIZE;
1916                 disk = 0;
1917                 if (sector_nr + (len>>9) > max_sector)
1918                         len = (max_sector - sector_nr) << 9;
1919                 if (len == 0)
1920                         break;
1921                 for (bio= biolist ; bio ; bio=bio->bi_next) {
1922                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1923                         if (bio_add_page(bio, page, len, 0) == 0) {
1924                                 /* stop here */
1925                                 struct bio *bio2;
1926                                 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1927                                 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
1928                                         /* remove last page from this bio */
1929                                         bio2->bi_vcnt--;
1930                                         bio2->bi_size -= len;
1931                                         bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
1932                                 }
1933                                 goto bio_full;
1934                         }
1935                         disk = i;
1936                 }
1937                 nr_sectors += len>>9;
1938                 sector_nr += len>>9;
1939         } while (biolist->bi_vcnt < RESYNC_PAGES);
1940  bio_full:
1941         r10_bio->sectors = nr_sectors;
1942
1943         while (biolist) {
1944                 bio = biolist;
1945                 biolist = biolist->bi_next;
1946
1947                 bio->bi_next = NULL;
1948                 r10_bio = bio->bi_private;
1949                 r10_bio->sectors = nr_sectors;
1950
1951                 if (bio->bi_end_io == end_sync_read) {
1952                         md_sync_acct(bio->bi_bdev, nr_sectors);
1953                         generic_make_request(bio);
1954                 }
1955         }
1956
1957         if (sectors_skipped)
1958                 /* pretend they weren't skipped, it makes
1959                  * no important difference in this case
1960                  */
1961                 md_done_sync(mddev, sectors_skipped, 1);
1962
1963         return sectors_skipped + nr_sectors;
1964  giveup:
1965         /* There is nowhere to write, so all non-sync
1966          * drives must be failed, so try the next chunk...
1967          */
1968         {
1969         sector_t sec = max_sector - sector_nr;
1970         sectors_skipped += sec;
1971         chunks_skipped ++;
1972         sector_nr = max_sector;
1973         goto skipped;
1974         }
1975 }
1976
1977 static int run(mddev_t *mddev)
1978 {
1979         conf_t *conf;
1980         int i, disk_idx;
1981         mirror_info_t *disk;
1982         mdk_rdev_t *rdev;
1983         struct list_head *tmp;
1984         int nc, fc, fo;
1985         sector_t stride, size;
1986
1987         if (mddev->chunk_size == 0) {
1988                 printk(KERN_ERR "md/raid10: non-zero chunk size required.\n");
1989                 return -EINVAL;
1990         }
1991
1992         nc = mddev->layout & 255;
1993         fc = (mddev->layout >> 8) & 255;
1994         fo = mddev->layout & (1<<16);
1995         if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
1996             (mddev->layout >> 17)) {
1997                 printk(KERN_ERR "raid10: %s: unsupported raid10 layout: 0x%8x\n",
1998                        mdname(mddev), mddev->layout);
1999                 goto out;
2000         }
2001         /*
2002          * copy the already verified devices into our private RAID10
2003          * bookkeeping area. [whatever we allocate in run(),
2004          * should be freed in stop()]
2005          */
2006         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2007         mddev->private = conf;
2008         if (!conf) {
2009                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2010                         mdname(mddev));
2011                 goto out;
2012         }
2013         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2014                                  GFP_KERNEL);
2015         if (!conf->mirrors) {
2016                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2017                        mdname(mddev));
2018                 goto out_free_conf;
2019         }
2020
2021         conf->tmppage = alloc_page(GFP_KERNEL);
2022         if (!conf->tmppage)
2023                 goto out_free_conf;
2024
2025         conf->mddev = mddev;
2026         conf->raid_disks = mddev->raid_disks;
2027         conf->near_copies = nc;
2028         conf->far_copies = fc;
2029         conf->copies = nc*fc;
2030         conf->far_offset = fo;
2031         conf->chunk_mask = (sector_t)(mddev->chunk_size>>9)-1;
2032         conf->chunk_shift = ffz(~mddev->chunk_size) - 9;
2033         size = mddev->size >> (conf->chunk_shift-1);
2034         sector_div(size, fc);
2035         size = size * conf->raid_disks;
2036         sector_div(size, nc);
2037         /* 'size' is now the number of chunks in the array */
2038         /* calculate "used chunks per device" in 'stride' */
2039         stride = size * conf->copies;
2040
2041         /* We need to round up when dividing by raid_disks to
2042          * get the stride size.
2043          */
2044         stride += conf->raid_disks - 1;
2045         sector_div(stride, conf->raid_disks);
2046         mddev->size = stride  << (conf->chunk_shift-1);
2047
2048         if (fo)
2049                 stride = 1;
2050         else
2051                 sector_div(stride, fc);
2052         conf->stride = stride << conf->chunk_shift;
2053
2054         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2055                                                 r10bio_pool_free, conf);
2056         if (!conf->r10bio_pool) {
2057                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2058                         mdname(mddev));
2059                 goto out_free_conf;
2060         }
2061
2062         rdev_for_each(rdev, tmp, mddev) {
2063                 disk_idx = rdev->raid_disk;
2064                 if (disk_idx >= mddev->raid_disks
2065                     || disk_idx < 0)
2066                         continue;
2067                 disk = conf->mirrors + disk_idx;
2068
2069                 disk->rdev = rdev;
2070
2071                 blk_queue_stack_limits(mddev->queue,
2072                                        rdev->bdev->bd_disk->queue);
2073                 /* as we don't honour merge_bvec_fn, we must never risk
2074                  * violating it, so limit ->max_sector to one PAGE, as
2075                  * a one page request is never in violation.
2076                  */
2077                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
2078                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
2079                         mddev->queue->max_sectors = (PAGE_SIZE>>9);
2080
2081                 disk->head_position = 0;
2082         }
2083         spin_lock_init(&conf->device_lock);
2084         INIT_LIST_HEAD(&conf->retry_list);
2085
2086         spin_lock_init(&conf->resync_lock);
2087         init_waitqueue_head(&conf->wait_barrier);
2088
2089         /* need to check that every block has at least one working mirror */
2090         if (!enough(conf)) {
2091                 printk(KERN_ERR "raid10: not enough operational mirrors for %s\n",
2092                        mdname(mddev));
2093                 goto out_free_conf;
2094         }
2095
2096         mddev->degraded = 0;
2097         for (i = 0; i < conf->raid_disks; i++) {
2098
2099                 disk = conf->mirrors + i;
2100
2101                 if (!disk->rdev ||
2102                     !test_bit(In_sync, &disk->rdev->flags)) {
2103                         disk->head_position = 0;
2104                         mddev->degraded++;
2105                 }
2106         }
2107
2108
2109         mddev->thread = md_register_thread(raid10d, mddev, "%s_raid10");
2110         if (!mddev->thread) {
2111                 printk(KERN_ERR
2112                        "raid10: couldn't allocate thread for %s\n",
2113                        mdname(mddev));
2114                 goto out_free_conf;
2115         }
2116
2117         printk(KERN_INFO
2118                 "raid10: raid set %s active with %d out of %d devices\n",
2119                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2120                 mddev->raid_disks);
2121         /*
2122          * Ok, everything is just fine now
2123          */
2124         mddev->array_size = size << (conf->chunk_shift-1);
2125         mddev->resync_max_sectors = size << conf->chunk_shift;
2126
2127         mddev->queue->unplug_fn = raid10_unplug;
2128         mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2129         mddev->queue->backing_dev_info.congested_data = mddev;
2130
2131         /* Calculate max read-ahead size.
2132          * We need to readahead at least twice a whole stripe....
2133          * maybe...
2134          */
2135         {
2136                 int stripe = conf->raid_disks * (mddev->chunk_size / PAGE_SIZE);
2137                 stripe /= conf->near_copies;
2138                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2139                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2140         }
2141
2142         if (conf->near_copies < mddev->raid_disks)
2143                 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2144         return 0;
2145
2146 out_free_conf:
2147         if (conf->r10bio_pool)
2148                 mempool_destroy(conf->r10bio_pool);
2149         safe_put_page(conf->tmppage);
2150         kfree(conf->mirrors);
2151         kfree(conf);
2152         mddev->private = NULL;
2153 out:
2154         return -EIO;
2155 }
2156
2157 static int stop(mddev_t *mddev)
2158 {
2159         conf_t *conf = mddev_to_conf(mddev);
2160
2161         md_unregister_thread(mddev->thread);
2162         mddev->thread = NULL;
2163         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2164         if (conf->r10bio_pool)
2165                 mempool_destroy(conf->r10bio_pool);
2166         kfree(conf->mirrors);
2167         kfree(conf);
2168         mddev->private = NULL;
2169         return 0;
2170 }
2171
2172 static void raid10_quiesce(mddev_t *mddev, int state)
2173 {
2174         conf_t *conf = mddev_to_conf(mddev);
2175
2176         switch(state) {
2177         case 1:
2178                 raise_barrier(conf, 0);
2179                 break;
2180         case 0:
2181                 lower_barrier(conf);
2182                 break;
2183         }
2184         if (mddev->thread) {
2185                 if (mddev->bitmap)
2186                         mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
2187                 else
2188                         mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
2189                 md_wakeup_thread(mddev->thread);
2190         }
2191 }
2192
2193 static struct mdk_personality raid10_personality =
2194 {
2195         .name           = "raid10",
2196         .level          = 10,
2197         .owner          = THIS_MODULE,
2198         .make_request   = make_request,
2199         .run            = run,
2200         .stop           = stop,
2201         .status         = status,
2202         .error_handler  = error,
2203         .hot_add_disk   = raid10_add_disk,
2204         .hot_remove_disk= raid10_remove_disk,
2205         .spare_active   = raid10_spare_active,
2206         .sync_request   = sync_request,
2207         .quiesce        = raid10_quiesce,
2208 };
2209
2210 static int __init raid_init(void)
2211 {
2212         return register_md_personality(&raid10_personality);
2213 }
2214
2215 static void raid_exit(void)
2216 {
2217         unregister_md_personality(&raid10_personality);
2218 }
2219
2220 module_init(raid_init);
2221 module_exit(raid_exit);
2222 MODULE_LICENSE("GPL");
2223 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2224 MODULE_ALIAS("md-raid10");
2225 MODULE_ALIAS("md-level-10");