]> git.openfabrics.org - ~shefty/rdma-dev.git/blob - drivers/md/raid10.c
Merge git://git.kernel.org/pub/scm/linux/kernel/git/pkl/squashfs-next
[~shefty/rdma-dev.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 further copyright information.
9  *
10  *
11  * This program is free software; you can redistribute it and/or modify
12  * it under the terms of the GNU General Public License as published by
13  * the Free Software Foundation; either version 2, or (at your option)
14  * any later version.
15  *
16  * You should have received a copy of the GNU General Public License
17  * (for example /usr/src/linux/COPYING); if not, write to the Free
18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19  */
20
21 #include <linux/slab.h>
22 #include <linux/delay.h>
23 #include <linux/blkdev.h>
24 #include <linux/seq_file.h>
25 #include <linux/ratelimit.h>
26 #include "md.h"
27 #include "raid10.h"
28 #include "raid0.h"
29 #include "bitmap.h"
30
31 /*
32  * RAID10 provides a combination of RAID0 and RAID1 functionality.
33  * The layout of data is defined by
34  *    chunk_size
35  *    raid_disks
36  *    near_copies (stored in low byte of layout)
37  *    far_copies (stored in second byte of layout)
38  *    far_offset (stored in bit 16 of layout )
39  *
40  * The data to be stored is divided into chunks using chunksize.
41  * Each device is divided into far_copies sections.
42  * In each section, chunks are laid out in a style similar to raid0, but
43  * near_copies copies of each chunk is stored (each on a different drive).
44  * The starting device for each section is offset near_copies from the starting
45  * device of the previous section.
46  * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
47  * drive.
48  * near_copies and far_copies must be at least one, and their product is at most
49  * raid_disks.
50  *
51  * If far_offset is true, then the far_copies are handled a bit differently.
52  * The copies are still in different stripes, but instead of be very far apart
53  * on disk, there are adjacent stripes.
54  */
55
56 /*
57  * Number of guaranteed r10bios in case of extreme VM load:
58  */
59 #define NR_RAID10_BIOS 256
60
61 /* When there are this many requests queue to be written by
62  * the raid10 thread, we become 'congested' to provide back-pressure
63  * for writeback.
64  */
65 static int max_queued_requests = 1024;
66
67 static void allow_barrier(struct r10conf *conf);
68 static void lower_barrier(struct r10conf *conf);
69
70 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
71 {
72         struct r10conf *conf = data;
73         int size = offsetof(struct r10bio, devs[conf->copies]);
74
75         /* allocate a r10bio with room for raid_disks entries in the bios array */
76         return kzalloc(size, gfp_flags);
77 }
78
79 static void r10bio_pool_free(void *r10_bio, void *data)
80 {
81         kfree(r10_bio);
82 }
83
84 /* Maximum size of each resync request */
85 #define RESYNC_BLOCK_SIZE (64*1024)
86 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
87 /* amount of memory to reserve for resync requests */
88 #define RESYNC_WINDOW (1024*1024)
89 /* maximum number of concurrent requests, memory permitting */
90 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
91
92 /*
93  * When performing a resync, we need to read and compare, so
94  * we need as many pages are there are copies.
95  * When performing a recovery, we need 2 bios, one for read,
96  * one for write (we recover only one drive per r10buf)
97  *
98  */
99 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
100 {
101         struct r10conf *conf = data;
102         struct page *page;
103         struct r10bio *r10_bio;
104         struct bio *bio;
105         int i, j;
106         int nalloc;
107
108         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
109         if (!r10_bio)
110                 return NULL;
111
112         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
113                 nalloc = conf->copies; /* resync */
114         else
115                 nalloc = 2; /* recovery */
116
117         /*
118          * Allocate bios.
119          */
120         for (j = nalloc ; j-- ; ) {
121                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
122                 if (!bio)
123                         goto out_free_bio;
124                 r10_bio->devs[j].bio = bio;
125         }
126         /*
127          * Allocate RESYNC_PAGES data pages and attach them
128          * where needed.
129          */
130         for (j = 0 ; j < nalloc; j++) {
131                 bio = r10_bio->devs[j].bio;
132                 for (i = 0; i < RESYNC_PAGES; i++) {
133                         if (j == 1 && !test_bit(MD_RECOVERY_SYNC,
134                                                 &conf->mddev->recovery)) {
135                                 /* we can share bv_page's during recovery */
136                                 struct bio *rbio = r10_bio->devs[0].bio;
137                                 page = rbio->bi_io_vec[i].bv_page;
138                                 get_page(page);
139                         } else
140                                 page = alloc_page(gfp_flags);
141                         if (unlikely(!page))
142                                 goto out_free_pages;
143
144                         bio->bi_io_vec[i].bv_page = page;
145                 }
146         }
147
148         return r10_bio;
149
150 out_free_pages:
151         for ( ; i > 0 ; i--)
152                 safe_put_page(bio->bi_io_vec[i-1].bv_page);
153         while (j--)
154                 for (i = 0; i < RESYNC_PAGES ; i++)
155                         safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
156         j = -1;
157 out_free_bio:
158         while ( ++j < nalloc )
159                 bio_put(r10_bio->devs[j].bio);
160         r10bio_pool_free(r10_bio, conf);
161         return NULL;
162 }
163
164 static void r10buf_pool_free(void *__r10_bio, void *data)
165 {
166         int i;
167         struct r10conf *conf = data;
168         struct r10bio *r10bio = __r10_bio;
169         int j;
170
171         for (j=0; j < conf->copies; j++) {
172                 struct bio *bio = r10bio->devs[j].bio;
173                 if (bio) {
174                         for (i = 0; i < RESYNC_PAGES; i++) {
175                                 safe_put_page(bio->bi_io_vec[i].bv_page);
176                                 bio->bi_io_vec[i].bv_page = NULL;
177                         }
178                         bio_put(bio);
179                 }
180         }
181         r10bio_pool_free(r10bio, conf);
182 }
183
184 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
185 {
186         int i;
187
188         for (i = 0; i < conf->copies; i++) {
189                 struct bio **bio = & r10_bio->devs[i].bio;
190                 if (!BIO_SPECIAL(*bio))
191                         bio_put(*bio);
192                 *bio = NULL;
193         }
194 }
195
196 static void free_r10bio(struct r10bio *r10_bio)
197 {
198         struct r10conf *conf = r10_bio->mddev->private;
199
200         put_all_bios(conf, r10_bio);
201         mempool_free(r10_bio, conf->r10bio_pool);
202 }
203
204 static void put_buf(struct r10bio *r10_bio)
205 {
206         struct r10conf *conf = r10_bio->mddev->private;
207
208         mempool_free(r10_bio, conf->r10buf_pool);
209
210         lower_barrier(conf);
211 }
212
213 static void reschedule_retry(struct r10bio *r10_bio)
214 {
215         unsigned long flags;
216         struct mddev *mddev = r10_bio->mddev;
217         struct r10conf *conf = mddev->private;
218
219         spin_lock_irqsave(&conf->device_lock, flags);
220         list_add(&r10_bio->retry_list, &conf->retry_list);
221         conf->nr_queued ++;
222         spin_unlock_irqrestore(&conf->device_lock, flags);
223
224         /* wake up frozen array... */
225         wake_up(&conf->wait_barrier);
226
227         md_wakeup_thread(mddev->thread);
228 }
229
230 /*
231  * raid_end_bio_io() is called when we have finished servicing a mirrored
232  * operation and are ready to return a success/failure code to the buffer
233  * cache layer.
234  */
235 static void raid_end_bio_io(struct r10bio *r10_bio)
236 {
237         struct bio *bio = r10_bio->master_bio;
238         int done;
239         struct r10conf *conf = r10_bio->mddev->private;
240
241         if (bio->bi_phys_segments) {
242                 unsigned long flags;
243                 spin_lock_irqsave(&conf->device_lock, flags);
244                 bio->bi_phys_segments--;
245                 done = (bio->bi_phys_segments == 0);
246                 spin_unlock_irqrestore(&conf->device_lock, flags);
247         } else
248                 done = 1;
249         if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
250                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
251         if (done) {
252                 bio_endio(bio, 0);
253                 /*
254                  * Wake up any possible resync thread that waits for the device
255                  * to go idle.
256                  */
257                 allow_barrier(conf);
258         }
259         free_r10bio(r10_bio);
260 }
261
262 /*
263  * Update disk head position estimator based on IRQ completion info.
264  */
265 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
266 {
267         struct r10conf *conf = r10_bio->mddev->private;
268
269         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
270                 r10_bio->devs[slot].addr + (r10_bio->sectors);
271 }
272
273 /*
274  * Find the disk number which triggered given bio
275  */
276 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
277                          struct bio *bio, int *slotp)
278 {
279         int slot;
280
281         for (slot = 0; slot < conf->copies; slot++)
282                 if (r10_bio->devs[slot].bio == bio)
283                         break;
284
285         BUG_ON(slot == conf->copies);
286         update_head_pos(slot, r10_bio);
287
288         if (slotp)
289                 *slotp = slot;
290         return r10_bio->devs[slot].devnum;
291 }
292
293 static void raid10_end_read_request(struct bio *bio, int error)
294 {
295         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
296         struct r10bio *r10_bio = bio->bi_private;
297         int slot, dev;
298         struct r10conf *conf = r10_bio->mddev->private;
299
300
301         slot = r10_bio->read_slot;
302         dev = r10_bio->devs[slot].devnum;
303         /*
304          * this branch is our 'one mirror IO has finished' event handler:
305          */
306         update_head_pos(slot, r10_bio);
307
308         if (uptodate) {
309                 /*
310                  * Set R10BIO_Uptodate in our master bio, so that
311                  * we will return a good error code to the higher
312                  * levels even if IO on some other mirrored buffer fails.
313                  *
314                  * The 'master' represents the composite IO operation to
315                  * user-side. So if something waits for IO, then it will
316                  * wait for the 'master' bio.
317                  */
318                 set_bit(R10BIO_Uptodate, &r10_bio->state);
319                 raid_end_bio_io(r10_bio);
320                 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
321         } else {
322                 /*
323                  * oops, read error - keep the refcount on the rdev
324                  */
325                 char b[BDEVNAME_SIZE];
326                 printk_ratelimited(KERN_ERR
327                                    "md/raid10:%s: %s: rescheduling sector %llu\n",
328                                    mdname(conf->mddev),
329                                    bdevname(conf->mirrors[dev].rdev->bdev, b),
330                                    (unsigned long long)r10_bio->sector);
331                 set_bit(R10BIO_ReadError, &r10_bio->state);
332                 reschedule_retry(r10_bio);
333         }
334 }
335
336 static void close_write(struct r10bio *r10_bio)
337 {
338         /* clear the bitmap if all writes complete successfully */
339         bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
340                         r10_bio->sectors,
341                         !test_bit(R10BIO_Degraded, &r10_bio->state),
342                         0);
343         md_write_end(r10_bio->mddev);
344 }
345
346 static void one_write_done(struct r10bio *r10_bio)
347 {
348         if (atomic_dec_and_test(&r10_bio->remaining)) {
349                 if (test_bit(R10BIO_WriteError, &r10_bio->state))
350                         reschedule_retry(r10_bio);
351                 else {
352                         close_write(r10_bio);
353                         if (test_bit(R10BIO_MadeGood, &r10_bio->state))
354                                 reschedule_retry(r10_bio);
355                         else
356                                 raid_end_bio_io(r10_bio);
357                 }
358         }
359 }
360
361 static void raid10_end_write_request(struct bio *bio, int error)
362 {
363         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
364         struct r10bio *r10_bio = bio->bi_private;
365         int dev;
366         int dec_rdev = 1;
367         struct r10conf *conf = r10_bio->mddev->private;
368         int slot;
369
370         dev = find_bio_disk(conf, r10_bio, bio, &slot);
371
372         /*
373          * this branch is our 'one mirror IO has finished' event handler:
374          */
375         if (!uptodate) {
376                 set_bit(WriteErrorSeen, &conf->mirrors[dev].rdev->flags);
377                 set_bit(R10BIO_WriteError, &r10_bio->state);
378                 dec_rdev = 0;
379         } else {
380                 /*
381                  * Set R10BIO_Uptodate in our master bio, so that
382                  * we will return a good error code for to the higher
383                  * levels even if IO on some other mirrored buffer fails.
384                  *
385                  * The 'master' represents the composite IO operation to
386                  * user-side. So if something waits for IO, then it will
387                  * wait for the 'master' bio.
388                  */
389                 sector_t first_bad;
390                 int bad_sectors;
391
392                 set_bit(R10BIO_Uptodate, &r10_bio->state);
393
394                 /* Maybe we can clear some bad blocks. */
395                 if (is_badblock(conf->mirrors[dev].rdev,
396                                 r10_bio->devs[slot].addr,
397                                 r10_bio->sectors,
398                                 &first_bad, &bad_sectors)) {
399                         bio_put(bio);
400                         r10_bio->devs[slot].bio = IO_MADE_GOOD;
401                         dec_rdev = 0;
402                         set_bit(R10BIO_MadeGood, &r10_bio->state);
403                 }
404         }
405
406         /*
407          *
408          * Let's see if all mirrored write operations have finished
409          * already.
410          */
411         one_write_done(r10_bio);
412         if (dec_rdev)
413                 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
414 }
415
416
417 /*
418  * RAID10 layout manager
419  * As well as the chunksize and raid_disks count, there are two
420  * parameters: near_copies and far_copies.
421  * near_copies * far_copies must be <= raid_disks.
422  * Normally one of these will be 1.
423  * If both are 1, we get raid0.
424  * If near_copies == raid_disks, we get raid1.
425  *
426  * Chunks are laid out in raid0 style with near_copies copies of the
427  * first chunk, followed by near_copies copies of the next chunk and
428  * so on.
429  * If far_copies > 1, then after 1/far_copies of the array has been assigned
430  * as described above, we start again with a device offset of near_copies.
431  * So we effectively have another copy of the whole array further down all
432  * the drives, but with blocks on different drives.
433  * With this layout, and block is never stored twice on the one device.
434  *
435  * raid10_find_phys finds the sector offset of a given virtual sector
436  * on each device that it is on.
437  *
438  * raid10_find_virt does the reverse mapping, from a device and a
439  * sector offset to a virtual address
440  */
441
442 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
443 {
444         int n,f;
445         sector_t sector;
446         sector_t chunk;
447         sector_t stripe;
448         int dev;
449
450         int slot = 0;
451
452         /* now calculate first sector/dev */
453         chunk = r10bio->sector >> conf->chunk_shift;
454         sector = r10bio->sector & conf->chunk_mask;
455
456         chunk *= conf->near_copies;
457         stripe = chunk;
458         dev = sector_div(stripe, conf->raid_disks);
459         if (conf->far_offset)
460                 stripe *= conf->far_copies;
461
462         sector += stripe << conf->chunk_shift;
463
464         /* and calculate all the others */
465         for (n=0; n < conf->near_copies; n++) {
466                 int d = dev;
467                 sector_t s = sector;
468                 r10bio->devs[slot].addr = sector;
469                 r10bio->devs[slot].devnum = d;
470                 slot++;
471
472                 for (f = 1; f < conf->far_copies; f++) {
473                         d += conf->near_copies;
474                         if (d >= conf->raid_disks)
475                                 d -= conf->raid_disks;
476                         s += conf->stride;
477                         r10bio->devs[slot].devnum = d;
478                         r10bio->devs[slot].addr = s;
479                         slot++;
480                 }
481                 dev++;
482                 if (dev >= conf->raid_disks) {
483                         dev = 0;
484                         sector += (conf->chunk_mask + 1);
485                 }
486         }
487         BUG_ON(slot != conf->copies);
488 }
489
490 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
491 {
492         sector_t offset, chunk, vchunk;
493
494         offset = sector & conf->chunk_mask;
495         if (conf->far_offset) {
496                 int fc;
497                 chunk = sector >> conf->chunk_shift;
498                 fc = sector_div(chunk, conf->far_copies);
499                 dev -= fc * conf->near_copies;
500                 if (dev < 0)
501                         dev += conf->raid_disks;
502         } else {
503                 while (sector >= conf->stride) {
504                         sector -= conf->stride;
505                         if (dev < conf->near_copies)
506                                 dev += conf->raid_disks - conf->near_copies;
507                         else
508                                 dev -= conf->near_copies;
509                 }
510                 chunk = sector >> conf->chunk_shift;
511         }
512         vchunk = chunk * conf->raid_disks + dev;
513         sector_div(vchunk, conf->near_copies);
514         return (vchunk << conf->chunk_shift) + offset;
515 }
516
517 /**
518  *      raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
519  *      @q: request queue
520  *      @bvm: properties of new bio
521  *      @biovec: the request that could be merged to it.
522  *
523  *      Return amount of bytes we can accept at this offset
524  *      If near_copies == raid_disk, there are no striping issues,
525  *      but in that case, the function isn't called at all.
526  */
527 static int raid10_mergeable_bvec(struct request_queue *q,
528                                  struct bvec_merge_data *bvm,
529                                  struct bio_vec *biovec)
530 {
531         struct mddev *mddev = q->queuedata;
532         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
533         int max;
534         unsigned int chunk_sectors = mddev->chunk_sectors;
535         unsigned int bio_sectors = bvm->bi_size >> 9;
536
537         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
538         if (max < 0) max = 0; /* bio_add cannot handle a negative return */
539         if (max <= biovec->bv_len && bio_sectors == 0)
540                 return biovec->bv_len;
541         else
542                 return max;
543 }
544
545 /*
546  * This routine returns the disk from which the requested read should
547  * be done. There is a per-array 'next expected sequential IO' sector
548  * number - if this matches on the next IO then we use the last disk.
549  * There is also a per-disk 'last know head position' sector that is
550  * maintained from IRQ contexts, both the normal and the resync IO
551  * completion handlers update this position correctly. If there is no
552  * perfect sequential match then we pick the disk whose head is closest.
553  *
554  * If there are 2 mirrors in the same 2 devices, performance degrades
555  * because position is mirror, not device based.
556  *
557  * The rdev for the device selected will have nr_pending incremented.
558  */
559
560 /*
561  * FIXME: possibly should rethink readbalancing and do it differently
562  * depending on near_copies / far_copies geometry.
563  */
564 static int read_balance(struct r10conf *conf, struct r10bio *r10_bio, int *max_sectors)
565 {
566         const sector_t this_sector = r10_bio->sector;
567         int disk, slot;
568         int sectors = r10_bio->sectors;
569         int best_good_sectors;
570         sector_t new_distance, best_dist;
571         struct md_rdev *rdev;
572         int do_balance;
573         int best_slot;
574
575         raid10_find_phys(conf, r10_bio);
576         rcu_read_lock();
577 retry:
578         sectors = r10_bio->sectors;
579         best_slot = -1;
580         best_dist = MaxSector;
581         best_good_sectors = 0;
582         do_balance = 1;
583         /*
584          * Check if we can balance. We can balance on the whole
585          * device if no resync is going on (recovery is ok), or below
586          * the resync window. We take the first readable disk when
587          * above the resync window.
588          */
589         if (conf->mddev->recovery_cp < MaxSector
590             && (this_sector + sectors >= conf->next_resync))
591                 do_balance = 0;
592
593         for (slot = 0; slot < conf->copies ; slot++) {
594                 sector_t first_bad;
595                 int bad_sectors;
596                 sector_t dev_sector;
597
598                 if (r10_bio->devs[slot].bio == IO_BLOCKED)
599                         continue;
600                 disk = r10_bio->devs[slot].devnum;
601                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
602                 if (rdev == NULL)
603                         continue;
604                 if (!test_bit(In_sync, &rdev->flags))
605                         continue;
606
607                 dev_sector = r10_bio->devs[slot].addr;
608                 if (is_badblock(rdev, dev_sector, sectors,
609                                 &first_bad, &bad_sectors)) {
610                         if (best_dist < MaxSector)
611                                 /* Already have a better slot */
612                                 continue;
613                         if (first_bad <= dev_sector) {
614                                 /* Cannot read here.  If this is the
615                                  * 'primary' device, then we must not read
616                                  * beyond 'bad_sectors' from another device.
617                                  */
618                                 bad_sectors -= (dev_sector - first_bad);
619                                 if (!do_balance && sectors > bad_sectors)
620                                         sectors = bad_sectors;
621                                 if (best_good_sectors > sectors)
622                                         best_good_sectors = sectors;
623                         } else {
624                                 sector_t good_sectors =
625                                         first_bad - dev_sector;
626                                 if (good_sectors > best_good_sectors) {
627                                         best_good_sectors = good_sectors;
628                                         best_slot = slot;
629                                 }
630                                 if (!do_balance)
631                                         /* Must read from here */
632                                         break;
633                         }
634                         continue;
635                 } else
636                         best_good_sectors = sectors;
637
638                 if (!do_balance)
639                         break;
640
641                 /* This optimisation is debatable, and completely destroys
642                  * sequential read speed for 'far copies' arrays.  So only
643                  * keep it for 'near' arrays, and review those later.
644                  */
645                 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending))
646                         break;
647
648                 /* for far > 1 always use the lowest address */
649                 if (conf->far_copies > 1)
650                         new_distance = r10_bio->devs[slot].addr;
651                 else
652                         new_distance = abs(r10_bio->devs[slot].addr -
653                                            conf->mirrors[disk].head_position);
654                 if (new_distance < best_dist) {
655                         best_dist = new_distance;
656                         best_slot = slot;
657                 }
658         }
659         if (slot == conf->copies)
660                 slot = best_slot;
661
662         if (slot >= 0) {
663                 disk = r10_bio->devs[slot].devnum;
664                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
665                 if (!rdev)
666                         goto retry;
667                 atomic_inc(&rdev->nr_pending);
668                 if (test_bit(Faulty, &rdev->flags)) {
669                         /* Cannot risk returning a device that failed
670                          * before we inc'ed nr_pending
671                          */
672                         rdev_dec_pending(rdev, conf->mddev);
673                         goto retry;
674                 }
675                 r10_bio->read_slot = slot;
676         } else
677                 disk = -1;
678         rcu_read_unlock();
679         *max_sectors = best_good_sectors;
680
681         return disk;
682 }
683
684 static int raid10_congested(void *data, int bits)
685 {
686         struct mddev *mddev = data;
687         struct r10conf *conf = mddev->private;
688         int i, ret = 0;
689
690         if ((bits & (1 << BDI_async_congested)) &&
691             conf->pending_count >= max_queued_requests)
692                 return 1;
693
694         if (mddev_congested(mddev, bits))
695                 return 1;
696         rcu_read_lock();
697         for (i = 0; i < conf->raid_disks && ret == 0; i++) {
698                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
699                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
700                         struct request_queue *q = bdev_get_queue(rdev->bdev);
701
702                         ret |= bdi_congested(&q->backing_dev_info, bits);
703                 }
704         }
705         rcu_read_unlock();
706         return ret;
707 }
708
709 static void flush_pending_writes(struct r10conf *conf)
710 {
711         /* Any writes that have been queued but are awaiting
712          * bitmap updates get flushed here.
713          */
714         spin_lock_irq(&conf->device_lock);
715
716         if (conf->pending_bio_list.head) {
717                 struct bio *bio;
718                 bio = bio_list_get(&conf->pending_bio_list);
719                 conf->pending_count = 0;
720                 spin_unlock_irq(&conf->device_lock);
721                 /* flush any pending bitmap writes to disk
722                  * before proceeding w/ I/O */
723                 bitmap_unplug(conf->mddev->bitmap);
724                 wake_up(&conf->wait_barrier);
725
726                 while (bio) { /* submit pending writes */
727                         struct bio *next = bio->bi_next;
728                         bio->bi_next = NULL;
729                         generic_make_request(bio);
730                         bio = next;
731                 }
732         } else
733                 spin_unlock_irq(&conf->device_lock);
734 }
735
736 /* Barriers....
737  * Sometimes we need to suspend IO while we do something else,
738  * either some resync/recovery, or reconfigure the array.
739  * To do this we raise a 'barrier'.
740  * The 'barrier' is a counter that can be raised multiple times
741  * to count how many activities are happening which preclude
742  * normal IO.
743  * We can only raise the barrier if there is no pending IO.
744  * i.e. if nr_pending == 0.
745  * We choose only to raise the barrier if no-one is waiting for the
746  * barrier to go down.  This means that as soon as an IO request
747  * is ready, no other operations which require a barrier will start
748  * until the IO request has had a chance.
749  *
750  * So: regular IO calls 'wait_barrier'.  When that returns there
751  *    is no backgroup IO happening,  It must arrange to call
752  *    allow_barrier when it has finished its IO.
753  * backgroup IO calls must call raise_barrier.  Once that returns
754  *    there is no normal IO happeing.  It must arrange to call
755  *    lower_barrier when the particular background IO completes.
756  */
757
758 static void raise_barrier(struct r10conf *conf, int force)
759 {
760         BUG_ON(force && !conf->barrier);
761         spin_lock_irq(&conf->resync_lock);
762
763         /* Wait until no block IO is waiting (unless 'force') */
764         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
765                             conf->resync_lock, );
766
767         /* block any new IO from starting */
768         conf->barrier++;
769
770         /* Now wait for all pending IO to complete */
771         wait_event_lock_irq(conf->wait_barrier,
772                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
773                             conf->resync_lock, );
774
775         spin_unlock_irq(&conf->resync_lock);
776 }
777
778 static void lower_barrier(struct r10conf *conf)
779 {
780         unsigned long flags;
781         spin_lock_irqsave(&conf->resync_lock, flags);
782         conf->barrier--;
783         spin_unlock_irqrestore(&conf->resync_lock, flags);
784         wake_up(&conf->wait_barrier);
785 }
786
787 static void wait_barrier(struct r10conf *conf)
788 {
789         spin_lock_irq(&conf->resync_lock);
790         if (conf->barrier) {
791                 conf->nr_waiting++;
792                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
793                                     conf->resync_lock,
794                                     );
795                 conf->nr_waiting--;
796         }
797         conf->nr_pending++;
798         spin_unlock_irq(&conf->resync_lock);
799 }
800
801 static void allow_barrier(struct r10conf *conf)
802 {
803         unsigned long flags;
804         spin_lock_irqsave(&conf->resync_lock, flags);
805         conf->nr_pending--;
806         spin_unlock_irqrestore(&conf->resync_lock, flags);
807         wake_up(&conf->wait_barrier);
808 }
809
810 static void freeze_array(struct r10conf *conf)
811 {
812         /* stop syncio and normal IO and wait for everything to
813          * go quiet.
814          * We increment barrier and nr_waiting, and then
815          * wait until nr_pending match nr_queued+1
816          * This is called in the context of one normal IO request
817          * that has failed. Thus any sync request that might be pending
818          * will be blocked by nr_pending, and we need to wait for
819          * pending IO requests to complete or be queued for re-try.
820          * Thus the number queued (nr_queued) plus this request (1)
821          * must match the number of pending IOs (nr_pending) before
822          * we continue.
823          */
824         spin_lock_irq(&conf->resync_lock);
825         conf->barrier++;
826         conf->nr_waiting++;
827         wait_event_lock_irq(conf->wait_barrier,
828                             conf->nr_pending == conf->nr_queued+1,
829                             conf->resync_lock,
830                             flush_pending_writes(conf));
831
832         spin_unlock_irq(&conf->resync_lock);
833 }
834
835 static void unfreeze_array(struct r10conf *conf)
836 {
837         /* reverse the effect of the freeze */
838         spin_lock_irq(&conf->resync_lock);
839         conf->barrier--;
840         conf->nr_waiting--;
841         wake_up(&conf->wait_barrier);
842         spin_unlock_irq(&conf->resync_lock);
843 }
844
845 static int make_request(struct mddev *mddev, struct bio * bio)
846 {
847         struct r10conf *conf = mddev->private;
848         struct mirror_info *mirror;
849         struct r10bio *r10_bio;
850         struct bio *read_bio;
851         int i;
852         int chunk_sects = conf->chunk_mask + 1;
853         const int rw = bio_data_dir(bio);
854         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
855         const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
856         unsigned long flags;
857         struct md_rdev *blocked_rdev;
858         int plugged;
859         int sectors_handled;
860         int max_sectors;
861
862         if (unlikely(bio->bi_rw & REQ_FLUSH)) {
863                 md_flush_request(mddev, bio);
864                 return 0;
865         }
866
867         /* If this request crosses a chunk boundary, we need to
868          * split it.  This will only happen for 1 PAGE (or less) requests.
869          */
870         if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
871                       > chunk_sects &&
872                     conf->near_copies < conf->raid_disks)) {
873                 struct bio_pair *bp;
874                 /* Sanity check -- queue functions should prevent this happening */
875                 if (bio->bi_vcnt != 1 ||
876                     bio->bi_idx != 0)
877                         goto bad_map;
878                 /* This is a one page bio that upper layers
879                  * refuse to split for us, so we need to split it.
880                  */
881                 bp = bio_split(bio,
882                                chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
883
884                 /* Each of these 'make_request' calls will call 'wait_barrier'.
885                  * If the first succeeds but the second blocks due to the resync
886                  * thread raising the barrier, we will deadlock because the
887                  * IO to the underlying device will be queued in generic_make_request
888                  * and will never complete, so will never reduce nr_pending.
889                  * So increment nr_waiting here so no new raise_barriers will
890                  * succeed, and so the second wait_barrier cannot block.
891                  */
892                 spin_lock_irq(&conf->resync_lock);
893                 conf->nr_waiting++;
894                 spin_unlock_irq(&conf->resync_lock);
895
896                 if (make_request(mddev, &bp->bio1))
897                         generic_make_request(&bp->bio1);
898                 if (make_request(mddev, &bp->bio2))
899                         generic_make_request(&bp->bio2);
900
901                 spin_lock_irq(&conf->resync_lock);
902                 conf->nr_waiting--;
903                 wake_up(&conf->wait_barrier);
904                 spin_unlock_irq(&conf->resync_lock);
905
906                 bio_pair_release(bp);
907                 return 0;
908         bad_map:
909                 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
910                        " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
911                        (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
912
913                 bio_io_error(bio);
914                 return 0;
915         }
916
917         md_write_start(mddev, bio);
918
919         /*
920          * Register the new request and wait if the reconstruction
921          * thread has put up a bar for new requests.
922          * Continue immediately if no resync is active currently.
923          */
924         wait_barrier(conf);
925
926         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
927
928         r10_bio->master_bio = bio;
929         r10_bio->sectors = bio->bi_size >> 9;
930
931         r10_bio->mddev = mddev;
932         r10_bio->sector = bio->bi_sector;
933         r10_bio->state = 0;
934
935         /* We might need to issue multiple reads to different
936          * devices if there are bad blocks around, so we keep
937          * track of the number of reads in bio->bi_phys_segments.
938          * If this is 0, there is only one r10_bio and no locking
939          * will be needed when the request completes.  If it is
940          * non-zero, then it is the number of not-completed requests.
941          */
942         bio->bi_phys_segments = 0;
943         clear_bit(BIO_SEG_VALID, &bio->bi_flags);
944
945         if (rw == READ) {
946                 /*
947                  * read balancing logic:
948                  */
949                 int disk;
950                 int slot;
951
952 read_again:
953                 disk = read_balance(conf, r10_bio, &max_sectors);
954                 slot = r10_bio->read_slot;
955                 if (disk < 0) {
956                         raid_end_bio_io(r10_bio);
957                         return 0;
958                 }
959                 mirror = conf->mirrors + disk;
960
961                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
962                 md_trim_bio(read_bio, r10_bio->sector - bio->bi_sector,
963                             max_sectors);
964
965                 r10_bio->devs[slot].bio = read_bio;
966
967                 read_bio->bi_sector = r10_bio->devs[slot].addr +
968                         mirror->rdev->data_offset;
969                 read_bio->bi_bdev = mirror->rdev->bdev;
970                 read_bio->bi_end_io = raid10_end_read_request;
971                 read_bio->bi_rw = READ | do_sync;
972                 read_bio->bi_private = r10_bio;
973
974                 if (max_sectors < r10_bio->sectors) {
975                         /* Could not read all from this device, so we will
976                          * need another r10_bio.
977                          */
978                         sectors_handled = (r10_bio->sectors + max_sectors
979                                            - bio->bi_sector);
980                         r10_bio->sectors = max_sectors;
981                         spin_lock_irq(&conf->device_lock);
982                         if (bio->bi_phys_segments == 0)
983                                 bio->bi_phys_segments = 2;
984                         else
985                                 bio->bi_phys_segments++;
986                         spin_unlock(&conf->device_lock);
987                         /* Cannot call generic_make_request directly
988                          * as that will be queued in __generic_make_request
989                          * and subsequent mempool_alloc might block
990                          * waiting for it.  so hand bio over to raid10d.
991                          */
992                         reschedule_retry(r10_bio);
993
994                         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
995
996                         r10_bio->master_bio = bio;
997                         r10_bio->sectors = ((bio->bi_size >> 9)
998                                             - sectors_handled);
999                         r10_bio->state = 0;
1000                         r10_bio->mddev = mddev;
1001                         r10_bio->sector = bio->bi_sector + sectors_handled;
1002                         goto read_again;
1003                 } else
1004                         generic_make_request(read_bio);
1005                 return 0;
1006         }
1007
1008         /*
1009          * WRITE:
1010          */
1011         if (conf->pending_count >= max_queued_requests) {
1012                 md_wakeup_thread(mddev->thread);
1013                 wait_event(conf->wait_barrier,
1014                            conf->pending_count < max_queued_requests);
1015         }
1016         /* first select target devices under rcu_lock and
1017          * inc refcount on their rdev.  Record them by setting
1018          * bios[x] to bio
1019          * If there are known/acknowledged bad blocks on any device
1020          * on which we have seen a write error, we want to avoid
1021          * writing to those blocks.  This potentially requires several
1022          * writes to write around the bad blocks.  Each set of writes
1023          * gets its own r10_bio with a set of bios attached.  The number
1024          * of r10_bios is recored in bio->bi_phys_segments just as with
1025          * the read case.
1026          */
1027         plugged = mddev_check_plugged(mddev);
1028
1029         raid10_find_phys(conf, r10_bio);
1030 retry_write:
1031         blocked_rdev = NULL;
1032         rcu_read_lock();
1033         max_sectors = r10_bio->sectors;
1034
1035         for (i = 0;  i < conf->copies; i++) {
1036                 int d = r10_bio->devs[i].devnum;
1037                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1038                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1039                         atomic_inc(&rdev->nr_pending);
1040                         blocked_rdev = rdev;
1041                         break;
1042                 }
1043                 r10_bio->devs[i].bio = NULL;
1044                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1045                         set_bit(R10BIO_Degraded, &r10_bio->state);
1046                         continue;
1047                 }
1048                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1049                         sector_t first_bad;
1050                         sector_t dev_sector = r10_bio->devs[i].addr;
1051                         int bad_sectors;
1052                         int is_bad;
1053
1054                         is_bad = is_badblock(rdev, dev_sector,
1055                                              max_sectors,
1056                                              &first_bad, &bad_sectors);
1057                         if (is_bad < 0) {
1058                                 /* Mustn't write here until the bad block
1059                                  * is acknowledged
1060                                  */
1061                                 atomic_inc(&rdev->nr_pending);
1062                                 set_bit(BlockedBadBlocks, &rdev->flags);
1063                                 blocked_rdev = rdev;
1064                                 break;
1065                         }
1066                         if (is_bad && first_bad <= dev_sector) {
1067                                 /* Cannot write here at all */
1068                                 bad_sectors -= (dev_sector - first_bad);
1069                                 if (bad_sectors < max_sectors)
1070                                         /* Mustn't write more than bad_sectors
1071                                          * to other devices yet
1072                                          */
1073                                         max_sectors = bad_sectors;
1074                                 /* We don't set R10BIO_Degraded as that
1075                                  * only applies if the disk is missing,
1076                                  * so it might be re-added, and we want to
1077                                  * know to recover this chunk.
1078                                  * In this case the device is here, and the
1079                                  * fact that this chunk is not in-sync is
1080                                  * recorded in the bad block log.
1081                                  */
1082                                 continue;
1083                         }
1084                         if (is_bad) {
1085                                 int good_sectors = first_bad - dev_sector;
1086                                 if (good_sectors < max_sectors)
1087                                         max_sectors = good_sectors;
1088                         }
1089                 }
1090                 r10_bio->devs[i].bio = bio;
1091                 atomic_inc(&rdev->nr_pending);
1092         }
1093         rcu_read_unlock();
1094
1095         if (unlikely(blocked_rdev)) {
1096                 /* Have to wait for this device to get unblocked, then retry */
1097                 int j;
1098                 int d;
1099
1100                 for (j = 0; j < i; j++)
1101                         if (r10_bio->devs[j].bio) {
1102                                 d = r10_bio->devs[j].devnum;
1103                                 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1104                         }
1105                 allow_barrier(conf);
1106                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1107                 wait_barrier(conf);
1108                 goto retry_write;
1109         }
1110
1111         if (max_sectors < r10_bio->sectors) {
1112                 /* We are splitting this into multiple parts, so
1113                  * we need to prepare for allocating another r10_bio.
1114                  */
1115                 r10_bio->sectors = max_sectors;
1116                 spin_lock_irq(&conf->device_lock);
1117                 if (bio->bi_phys_segments == 0)
1118                         bio->bi_phys_segments = 2;
1119                 else
1120                         bio->bi_phys_segments++;
1121                 spin_unlock_irq(&conf->device_lock);
1122         }
1123         sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector;
1124
1125         atomic_set(&r10_bio->remaining, 1);
1126         bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1127
1128         for (i = 0; i < conf->copies; i++) {
1129                 struct bio *mbio;
1130                 int d = r10_bio->devs[i].devnum;
1131                 if (!r10_bio->devs[i].bio)
1132                         continue;
1133
1134                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1135                 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1136                             max_sectors);
1137                 r10_bio->devs[i].bio = mbio;
1138
1139                 mbio->bi_sector = (r10_bio->devs[i].addr+
1140                                    conf->mirrors[d].rdev->data_offset);
1141                 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1142                 mbio->bi_end_io = raid10_end_write_request;
1143                 mbio->bi_rw = WRITE | do_sync | do_fua;
1144                 mbio->bi_private = r10_bio;
1145
1146                 atomic_inc(&r10_bio->remaining);
1147                 spin_lock_irqsave(&conf->device_lock, flags);
1148                 bio_list_add(&conf->pending_bio_list, mbio);
1149                 conf->pending_count++;
1150                 spin_unlock_irqrestore(&conf->device_lock, flags);
1151         }
1152
1153         /* Don't remove the bias on 'remaining' (one_write_done) until
1154          * after checking if we need to go around again.
1155          */
1156
1157         if (sectors_handled < (bio->bi_size >> 9)) {
1158                 one_write_done(r10_bio);
1159                 /* We need another r10_bio.  It has already been counted
1160                  * in bio->bi_phys_segments.
1161                  */
1162                 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1163
1164                 r10_bio->master_bio = bio;
1165                 r10_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1166
1167                 r10_bio->mddev = mddev;
1168                 r10_bio->sector = bio->bi_sector + sectors_handled;
1169                 r10_bio->state = 0;
1170                 goto retry_write;
1171         }
1172         one_write_done(r10_bio);
1173
1174         /* In case raid10d snuck in to freeze_array */
1175         wake_up(&conf->wait_barrier);
1176
1177         if (do_sync || !mddev->bitmap || !plugged)
1178                 md_wakeup_thread(mddev->thread);
1179         return 0;
1180 }
1181
1182 static void status(struct seq_file *seq, struct mddev *mddev)
1183 {
1184         struct r10conf *conf = mddev->private;
1185         int i;
1186
1187         if (conf->near_copies < conf->raid_disks)
1188                 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1189         if (conf->near_copies > 1)
1190                 seq_printf(seq, " %d near-copies", conf->near_copies);
1191         if (conf->far_copies > 1) {
1192                 if (conf->far_offset)
1193                         seq_printf(seq, " %d offset-copies", conf->far_copies);
1194                 else
1195                         seq_printf(seq, " %d far-copies", conf->far_copies);
1196         }
1197         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1198                                         conf->raid_disks - mddev->degraded);
1199         for (i = 0; i < conf->raid_disks; i++)
1200                 seq_printf(seq, "%s",
1201                               conf->mirrors[i].rdev &&
1202                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1203         seq_printf(seq, "]");
1204 }
1205
1206 /* check if there are enough drives for
1207  * every block to appear on atleast one.
1208  * Don't consider the device numbered 'ignore'
1209  * as we might be about to remove it.
1210  */
1211 static int enough(struct r10conf *conf, int ignore)
1212 {
1213         int first = 0;
1214
1215         do {
1216                 int n = conf->copies;
1217                 int cnt = 0;
1218                 while (n--) {
1219                         if (conf->mirrors[first].rdev &&
1220                             first != ignore)
1221                                 cnt++;
1222                         first = (first+1) % conf->raid_disks;
1223                 }
1224                 if (cnt == 0)
1225                         return 0;
1226         } while (first != 0);
1227         return 1;
1228 }
1229
1230 static void error(struct mddev *mddev, struct md_rdev *rdev)
1231 {
1232         char b[BDEVNAME_SIZE];
1233         struct r10conf *conf = mddev->private;
1234
1235         /*
1236          * If it is not operational, then we have already marked it as dead
1237          * else if it is the last working disks, ignore the error, let the
1238          * next level up know.
1239          * else mark the drive as failed
1240          */
1241         if (test_bit(In_sync, &rdev->flags)
1242             && !enough(conf, rdev->raid_disk))
1243                 /*
1244                  * Don't fail the drive, just return an IO error.
1245                  */
1246                 return;
1247         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1248                 unsigned long flags;
1249                 spin_lock_irqsave(&conf->device_lock, flags);
1250                 mddev->degraded++;
1251                 spin_unlock_irqrestore(&conf->device_lock, flags);
1252                 /*
1253                  * if recovery is running, make sure it aborts.
1254                  */
1255                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1256         }
1257         set_bit(Blocked, &rdev->flags);
1258         set_bit(Faulty, &rdev->flags);
1259         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1260         printk(KERN_ALERT
1261                "md/raid10:%s: Disk failure on %s, disabling device.\n"
1262                "md/raid10:%s: Operation continuing on %d devices.\n",
1263                mdname(mddev), bdevname(rdev->bdev, b),
1264                mdname(mddev), conf->raid_disks - mddev->degraded);
1265 }
1266
1267 static void print_conf(struct r10conf *conf)
1268 {
1269         int i;
1270         struct mirror_info *tmp;
1271
1272         printk(KERN_DEBUG "RAID10 conf printout:\n");
1273         if (!conf) {
1274                 printk(KERN_DEBUG "(!conf)\n");
1275                 return;
1276         }
1277         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1278                 conf->raid_disks);
1279
1280         for (i = 0; i < conf->raid_disks; i++) {
1281                 char b[BDEVNAME_SIZE];
1282                 tmp = conf->mirrors + i;
1283                 if (tmp->rdev)
1284                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1285                                 i, !test_bit(In_sync, &tmp->rdev->flags),
1286                                 !test_bit(Faulty, &tmp->rdev->flags),
1287                                 bdevname(tmp->rdev->bdev,b));
1288         }
1289 }
1290
1291 static void close_sync(struct r10conf *conf)
1292 {
1293         wait_barrier(conf);
1294         allow_barrier(conf);
1295
1296         mempool_destroy(conf->r10buf_pool);
1297         conf->r10buf_pool = NULL;
1298 }
1299
1300 static int raid10_spare_active(struct mddev *mddev)
1301 {
1302         int i;
1303         struct r10conf *conf = mddev->private;
1304         struct mirror_info *tmp;
1305         int count = 0;
1306         unsigned long flags;
1307
1308         /*
1309          * Find all non-in_sync disks within the RAID10 configuration
1310          * and mark them in_sync
1311          */
1312         for (i = 0; i < conf->raid_disks; i++) {
1313                 tmp = conf->mirrors + i;
1314                 if (tmp->rdev
1315                     && !test_bit(Faulty, &tmp->rdev->flags)
1316                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1317                         count++;
1318                         sysfs_notify_dirent(tmp->rdev->sysfs_state);
1319                 }
1320         }
1321         spin_lock_irqsave(&conf->device_lock, flags);
1322         mddev->degraded -= count;
1323         spin_unlock_irqrestore(&conf->device_lock, flags);
1324
1325         print_conf(conf);
1326         return count;
1327 }
1328
1329
1330 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1331 {
1332         struct r10conf *conf = mddev->private;
1333         int err = -EEXIST;
1334         int mirror;
1335         int first = 0;
1336         int last = conf->raid_disks - 1;
1337
1338         if (mddev->recovery_cp < MaxSector)
1339                 /* only hot-add to in-sync arrays, as recovery is
1340                  * very different from resync
1341                  */
1342                 return -EBUSY;
1343         if (!enough(conf, -1))
1344                 return -EINVAL;
1345
1346         if (rdev->raid_disk >= 0)
1347                 first = last = rdev->raid_disk;
1348
1349         if (rdev->saved_raid_disk >= first &&
1350             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1351                 mirror = rdev->saved_raid_disk;
1352         else
1353                 mirror = first;
1354         for ( ; mirror <= last ; mirror++) {
1355                 struct mirror_info *p = &conf->mirrors[mirror];
1356                 if (p->recovery_disabled == mddev->recovery_disabled)
1357                         continue;
1358                 if (p->rdev)
1359                         continue;
1360
1361                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1362                                   rdev->data_offset << 9);
1363                 /* as we don't honour merge_bvec_fn, we must
1364                  * never risk violating it, so limit
1365                  * ->max_segments to one lying with a single
1366                  * page, as a one page request is never in
1367                  * violation.
1368                  */
1369                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1370                         blk_queue_max_segments(mddev->queue, 1);
1371                         blk_queue_segment_boundary(mddev->queue,
1372                                                    PAGE_CACHE_SIZE - 1);
1373                 }
1374
1375                 p->head_position = 0;
1376                 p->recovery_disabled = mddev->recovery_disabled - 1;
1377                 rdev->raid_disk = mirror;
1378                 err = 0;
1379                 if (rdev->saved_raid_disk != mirror)
1380                         conf->fullsync = 1;
1381                 rcu_assign_pointer(p->rdev, rdev);
1382                 break;
1383         }
1384
1385         md_integrity_add_rdev(rdev, mddev);
1386         print_conf(conf);
1387         return err;
1388 }
1389
1390 static int raid10_remove_disk(struct mddev *mddev, int number)
1391 {
1392         struct r10conf *conf = mddev->private;
1393         int err = 0;
1394         struct md_rdev *rdev;
1395         struct mirror_info *p = conf->mirrors+ number;
1396
1397         print_conf(conf);
1398         rdev = p->rdev;
1399         if (rdev) {
1400                 if (test_bit(In_sync, &rdev->flags) ||
1401                     atomic_read(&rdev->nr_pending)) {
1402                         err = -EBUSY;
1403                         goto abort;
1404                 }
1405                 /* Only remove faulty devices in recovery
1406                  * is not possible.
1407                  */
1408                 if (!test_bit(Faulty, &rdev->flags) &&
1409                     mddev->recovery_disabled != p->recovery_disabled &&
1410                     enough(conf, -1)) {
1411                         err = -EBUSY;
1412                         goto abort;
1413                 }
1414                 p->rdev = NULL;
1415                 synchronize_rcu();
1416                 if (atomic_read(&rdev->nr_pending)) {
1417                         /* lost the race, try later */
1418                         err = -EBUSY;
1419                         p->rdev = rdev;
1420                         goto abort;
1421                 }
1422                 err = md_integrity_register(mddev);
1423         }
1424 abort:
1425
1426         print_conf(conf);
1427         return err;
1428 }
1429
1430
1431 static void end_sync_read(struct bio *bio, int error)
1432 {
1433         struct r10bio *r10_bio = bio->bi_private;
1434         struct r10conf *conf = r10_bio->mddev->private;
1435         int d;
1436
1437         d = find_bio_disk(conf, r10_bio, bio, NULL);
1438
1439         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1440                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1441         else
1442                 /* The write handler will notice the lack of
1443                  * R10BIO_Uptodate and record any errors etc
1444                  */
1445                 atomic_add(r10_bio->sectors,
1446                            &conf->mirrors[d].rdev->corrected_errors);
1447
1448         /* for reconstruct, we always reschedule after a read.
1449          * for resync, only after all reads
1450          */
1451         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1452         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1453             atomic_dec_and_test(&r10_bio->remaining)) {
1454                 /* we have read all the blocks,
1455                  * do the comparison in process context in raid10d
1456                  */
1457                 reschedule_retry(r10_bio);
1458         }
1459 }
1460
1461 static void end_sync_request(struct r10bio *r10_bio)
1462 {
1463         struct mddev *mddev = r10_bio->mddev;
1464
1465         while (atomic_dec_and_test(&r10_bio->remaining)) {
1466                 if (r10_bio->master_bio == NULL) {
1467                         /* the primary of several recovery bios */
1468                         sector_t s = r10_bio->sectors;
1469                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1470                             test_bit(R10BIO_WriteError, &r10_bio->state))
1471                                 reschedule_retry(r10_bio);
1472                         else
1473                                 put_buf(r10_bio);
1474                         md_done_sync(mddev, s, 1);
1475                         break;
1476                 } else {
1477                         struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1478                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1479                             test_bit(R10BIO_WriteError, &r10_bio->state))
1480                                 reschedule_retry(r10_bio);
1481                         else
1482                                 put_buf(r10_bio);
1483                         r10_bio = r10_bio2;
1484                 }
1485         }
1486 }
1487
1488 static void end_sync_write(struct bio *bio, int error)
1489 {
1490         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1491         struct r10bio *r10_bio = bio->bi_private;
1492         struct mddev *mddev = r10_bio->mddev;
1493         struct r10conf *conf = mddev->private;
1494         int d;
1495         sector_t first_bad;
1496         int bad_sectors;
1497         int slot;
1498
1499         d = find_bio_disk(conf, r10_bio, bio, &slot);
1500
1501         if (!uptodate) {
1502                 set_bit(WriteErrorSeen, &conf->mirrors[d].rdev->flags);
1503                 set_bit(R10BIO_WriteError, &r10_bio->state);
1504         } else if (is_badblock(conf->mirrors[d].rdev,
1505                              r10_bio->devs[slot].addr,
1506                              r10_bio->sectors,
1507                              &first_bad, &bad_sectors))
1508                 set_bit(R10BIO_MadeGood, &r10_bio->state);
1509
1510         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1511
1512         end_sync_request(r10_bio);
1513 }
1514
1515 /*
1516  * Note: sync and recover and handled very differently for raid10
1517  * This code is for resync.
1518  * For resync, we read through virtual addresses and read all blocks.
1519  * If there is any error, we schedule a write.  The lowest numbered
1520  * drive is authoritative.
1521  * However requests come for physical address, so we need to map.
1522  * For every physical address there are raid_disks/copies virtual addresses,
1523  * which is always are least one, but is not necessarly an integer.
1524  * This means that a physical address can span multiple chunks, so we may
1525  * have to submit multiple io requests for a single sync request.
1526  */
1527 /*
1528  * We check if all blocks are in-sync and only write to blocks that
1529  * aren't in sync
1530  */
1531 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1532 {
1533         struct r10conf *conf = mddev->private;
1534         int i, first;
1535         struct bio *tbio, *fbio;
1536
1537         atomic_set(&r10_bio->remaining, 1);
1538
1539         /* find the first device with a block */
1540         for (i=0; i<conf->copies; i++)
1541                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1542                         break;
1543
1544         if (i == conf->copies)
1545                 goto done;
1546
1547         first = i;
1548         fbio = r10_bio->devs[i].bio;
1549
1550         /* now find blocks with errors */
1551         for (i=0 ; i < conf->copies ; i++) {
1552                 int  j, d;
1553                 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1554
1555                 tbio = r10_bio->devs[i].bio;
1556
1557                 if (tbio->bi_end_io != end_sync_read)
1558                         continue;
1559                 if (i == first)
1560                         continue;
1561                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1562                         /* We know that the bi_io_vec layout is the same for
1563                          * both 'first' and 'i', so we just compare them.
1564                          * All vec entries are PAGE_SIZE;
1565                          */
1566                         for (j = 0; j < vcnt; j++)
1567                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1568                                            page_address(tbio->bi_io_vec[j].bv_page),
1569                                            PAGE_SIZE))
1570                                         break;
1571                         if (j == vcnt)
1572                                 continue;
1573                         mddev->resync_mismatches += r10_bio->sectors;
1574                         if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1575                                 /* Don't fix anything. */
1576                                 continue;
1577                 }
1578                 /* Ok, we need to write this bio, either to correct an
1579                  * inconsistency or to correct an unreadable block.
1580                  * First we need to fixup bv_offset, bv_len and
1581                  * bi_vecs, as the read request might have corrupted these
1582                  */
1583                 tbio->bi_vcnt = vcnt;
1584                 tbio->bi_size = r10_bio->sectors << 9;
1585                 tbio->bi_idx = 0;
1586                 tbio->bi_phys_segments = 0;
1587                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1588                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1589                 tbio->bi_next = NULL;
1590                 tbio->bi_rw = WRITE;
1591                 tbio->bi_private = r10_bio;
1592                 tbio->bi_sector = r10_bio->devs[i].addr;
1593
1594                 for (j=0; j < vcnt ; j++) {
1595                         tbio->bi_io_vec[j].bv_offset = 0;
1596                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1597
1598                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1599                                page_address(fbio->bi_io_vec[j].bv_page),
1600                                PAGE_SIZE);
1601                 }
1602                 tbio->bi_end_io = end_sync_write;
1603
1604                 d = r10_bio->devs[i].devnum;
1605                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1606                 atomic_inc(&r10_bio->remaining);
1607                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1608
1609                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1610                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1611                 generic_make_request(tbio);
1612         }
1613
1614 done:
1615         if (atomic_dec_and_test(&r10_bio->remaining)) {
1616                 md_done_sync(mddev, r10_bio->sectors, 1);
1617                 put_buf(r10_bio);
1618         }
1619 }
1620
1621 /*
1622  * Now for the recovery code.
1623  * Recovery happens across physical sectors.
1624  * We recover all non-is_sync drives by finding the virtual address of
1625  * each, and then choose a working drive that also has that virt address.
1626  * There is a separate r10_bio for each non-in_sync drive.
1627  * Only the first two slots are in use. The first for reading,
1628  * The second for writing.
1629  *
1630  */
1631 static void fix_recovery_read_error(struct r10bio *r10_bio)
1632 {
1633         /* We got a read error during recovery.
1634          * We repeat the read in smaller page-sized sections.
1635          * If a read succeeds, write it to the new device or record
1636          * a bad block if we cannot.
1637          * If a read fails, record a bad block on both old and
1638          * new devices.
1639          */
1640         struct mddev *mddev = r10_bio->mddev;
1641         struct r10conf *conf = mddev->private;
1642         struct bio *bio = r10_bio->devs[0].bio;
1643         sector_t sect = 0;
1644         int sectors = r10_bio->sectors;
1645         int idx = 0;
1646         int dr = r10_bio->devs[0].devnum;
1647         int dw = r10_bio->devs[1].devnum;
1648
1649         while (sectors) {
1650                 int s = sectors;
1651                 struct md_rdev *rdev;
1652                 sector_t addr;
1653                 int ok;
1654
1655                 if (s > (PAGE_SIZE>>9))
1656                         s = PAGE_SIZE >> 9;
1657
1658                 rdev = conf->mirrors[dr].rdev;
1659                 addr = r10_bio->devs[0].addr + sect,
1660                 ok = sync_page_io(rdev,
1661                                   addr,
1662                                   s << 9,
1663                                   bio->bi_io_vec[idx].bv_page,
1664                                   READ, false);
1665                 if (ok) {
1666                         rdev = conf->mirrors[dw].rdev;
1667                         addr = r10_bio->devs[1].addr + sect;
1668                         ok = sync_page_io(rdev,
1669                                           addr,
1670                                           s << 9,
1671                                           bio->bi_io_vec[idx].bv_page,
1672                                           WRITE, false);
1673                         if (!ok)
1674                                 set_bit(WriteErrorSeen, &rdev->flags);
1675                 }
1676                 if (!ok) {
1677                         /* We don't worry if we cannot set a bad block -
1678                          * it really is bad so there is no loss in not
1679                          * recording it yet
1680                          */
1681                         rdev_set_badblocks(rdev, addr, s, 0);
1682
1683                         if (rdev != conf->mirrors[dw].rdev) {
1684                                 /* need bad block on destination too */
1685                                 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
1686                                 addr = r10_bio->devs[1].addr + sect;
1687                                 ok = rdev_set_badblocks(rdev2, addr, s, 0);
1688                                 if (!ok) {
1689                                         /* just abort the recovery */
1690                                         printk(KERN_NOTICE
1691                                                "md/raid10:%s: recovery aborted"
1692                                                " due to read error\n",
1693                                                mdname(mddev));
1694
1695                                         conf->mirrors[dw].recovery_disabled
1696                                                 = mddev->recovery_disabled;
1697                                         set_bit(MD_RECOVERY_INTR,
1698                                                 &mddev->recovery);
1699                                         break;
1700                                 }
1701                         }
1702                 }
1703
1704                 sectors -= s;
1705                 sect += s;
1706                 idx++;
1707         }
1708 }
1709
1710 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1711 {
1712         struct r10conf *conf = mddev->private;
1713         int d;
1714         struct bio *wbio;
1715
1716         if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
1717                 fix_recovery_read_error(r10_bio);
1718                 end_sync_request(r10_bio);
1719                 return;
1720         }
1721
1722         /*
1723          * share the pages with the first bio
1724          * and submit the write request
1725          */
1726         wbio = r10_bio->devs[1].bio;
1727         d = r10_bio->devs[1].devnum;
1728
1729         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1730         md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1731         generic_make_request(wbio);
1732 }
1733
1734
1735 /*
1736  * Used by fix_read_error() to decay the per rdev read_errors.
1737  * We halve the read error count for every hour that has elapsed
1738  * since the last recorded read error.
1739  *
1740  */
1741 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
1742 {
1743         struct timespec cur_time_mon;
1744         unsigned long hours_since_last;
1745         unsigned int read_errors = atomic_read(&rdev->read_errors);
1746
1747         ktime_get_ts(&cur_time_mon);
1748
1749         if (rdev->last_read_error.tv_sec == 0 &&
1750             rdev->last_read_error.tv_nsec == 0) {
1751                 /* first time we've seen a read error */
1752                 rdev->last_read_error = cur_time_mon;
1753                 return;
1754         }
1755
1756         hours_since_last = (cur_time_mon.tv_sec -
1757                             rdev->last_read_error.tv_sec) / 3600;
1758
1759         rdev->last_read_error = cur_time_mon;
1760
1761         /*
1762          * if hours_since_last is > the number of bits in read_errors
1763          * just set read errors to 0. We do this to avoid
1764          * overflowing the shift of read_errors by hours_since_last.
1765          */
1766         if (hours_since_last >= 8 * sizeof(read_errors))
1767                 atomic_set(&rdev->read_errors, 0);
1768         else
1769                 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
1770 }
1771
1772 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
1773                             int sectors, struct page *page, int rw)
1774 {
1775         sector_t first_bad;
1776         int bad_sectors;
1777
1778         if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
1779             && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
1780                 return -1;
1781         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1782                 /* success */
1783                 return 1;
1784         if (rw == WRITE)
1785                 set_bit(WriteErrorSeen, &rdev->flags);
1786         /* need to record an error - either for the block or the device */
1787         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1788                 md_error(rdev->mddev, rdev);
1789         return 0;
1790 }
1791
1792 /*
1793  * This is a kernel thread which:
1794  *
1795  *      1.      Retries failed read operations on working mirrors.
1796  *      2.      Updates the raid superblock when problems encounter.
1797  *      3.      Performs writes following reads for array synchronising.
1798  */
1799
1800 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
1801 {
1802         int sect = 0; /* Offset from r10_bio->sector */
1803         int sectors = r10_bio->sectors;
1804         struct md_rdev*rdev;
1805         int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
1806         int d = r10_bio->devs[r10_bio->read_slot].devnum;
1807
1808         /* still own a reference to this rdev, so it cannot
1809          * have been cleared recently.
1810          */
1811         rdev = conf->mirrors[d].rdev;
1812
1813         if (test_bit(Faulty, &rdev->flags))
1814                 /* drive has already been failed, just ignore any
1815                    more fix_read_error() attempts */
1816                 return;
1817
1818         check_decay_read_errors(mddev, rdev);
1819         atomic_inc(&rdev->read_errors);
1820         if (atomic_read(&rdev->read_errors) > max_read_errors) {
1821                 char b[BDEVNAME_SIZE];
1822                 bdevname(rdev->bdev, b);
1823
1824                 printk(KERN_NOTICE
1825                        "md/raid10:%s: %s: Raid device exceeded "
1826                        "read_error threshold [cur %d:max %d]\n",
1827                        mdname(mddev), b,
1828                        atomic_read(&rdev->read_errors), max_read_errors);
1829                 printk(KERN_NOTICE
1830                        "md/raid10:%s: %s: Failing raid device\n",
1831                        mdname(mddev), b);
1832                 md_error(mddev, conf->mirrors[d].rdev);
1833                 return;
1834         }
1835
1836         while(sectors) {
1837                 int s = sectors;
1838                 int sl = r10_bio->read_slot;
1839                 int success = 0;
1840                 int start;
1841
1842                 if (s > (PAGE_SIZE>>9))
1843                         s = PAGE_SIZE >> 9;
1844
1845                 rcu_read_lock();
1846                 do {
1847                         sector_t first_bad;
1848                         int bad_sectors;
1849
1850                         d = r10_bio->devs[sl].devnum;
1851                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1852                         if (rdev &&
1853                             test_bit(In_sync, &rdev->flags) &&
1854                             is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
1855                                         &first_bad, &bad_sectors) == 0) {
1856                                 atomic_inc(&rdev->nr_pending);
1857                                 rcu_read_unlock();
1858                                 success = sync_page_io(rdev,
1859                                                        r10_bio->devs[sl].addr +
1860                                                        sect,
1861                                                        s<<9,
1862                                                        conf->tmppage, READ, false);
1863                                 rdev_dec_pending(rdev, mddev);
1864                                 rcu_read_lock();
1865                                 if (success)
1866                                         break;
1867                         }
1868                         sl++;
1869                         if (sl == conf->copies)
1870                                 sl = 0;
1871                 } while (!success && sl != r10_bio->read_slot);
1872                 rcu_read_unlock();
1873
1874                 if (!success) {
1875                         /* Cannot read from anywhere, just mark the block
1876                          * as bad on the first device to discourage future
1877                          * reads.
1878                          */
1879                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1880                         rdev = conf->mirrors[dn].rdev;
1881
1882                         if (!rdev_set_badblocks(
1883                                     rdev,
1884                                     r10_bio->devs[r10_bio->read_slot].addr
1885                                     + sect,
1886                                     s, 0))
1887                                 md_error(mddev, rdev);
1888                         break;
1889                 }
1890
1891                 start = sl;
1892                 /* write it back and re-read */
1893                 rcu_read_lock();
1894                 while (sl != r10_bio->read_slot) {
1895                         char b[BDEVNAME_SIZE];
1896
1897                         if (sl==0)
1898                                 sl = conf->copies;
1899                         sl--;
1900                         d = r10_bio->devs[sl].devnum;
1901                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1902                         if (!rdev ||
1903                             !test_bit(In_sync, &rdev->flags))
1904                                 continue;
1905
1906                         atomic_inc(&rdev->nr_pending);
1907                         rcu_read_unlock();
1908                         if (r10_sync_page_io(rdev,
1909                                              r10_bio->devs[sl].addr +
1910                                              sect,
1911                                              s<<9, conf->tmppage, WRITE)
1912                             == 0) {
1913                                 /* Well, this device is dead */
1914                                 printk(KERN_NOTICE
1915                                        "md/raid10:%s: read correction "
1916                                        "write failed"
1917                                        " (%d sectors at %llu on %s)\n",
1918                                        mdname(mddev), s,
1919                                        (unsigned long long)(
1920                                                sect + rdev->data_offset),
1921                                        bdevname(rdev->bdev, b));
1922                                 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
1923                                        "drive\n",
1924                                        mdname(mddev),
1925                                        bdevname(rdev->bdev, b));
1926                         }
1927                         rdev_dec_pending(rdev, mddev);
1928                         rcu_read_lock();
1929                 }
1930                 sl = start;
1931                 while (sl != r10_bio->read_slot) {
1932                         char b[BDEVNAME_SIZE];
1933
1934                         if (sl==0)
1935                                 sl = conf->copies;
1936                         sl--;
1937                         d = r10_bio->devs[sl].devnum;
1938                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1939                         if (!rdev ||
1940                             !test_bit(In_sync, &rdev->flags))
1941                                 continue;
1942
1943                         atomic_inc(&rdev->nr_pending);
1944                         rcu_read_unlock();
1945                         switch (r10_sync_page_io(rdev,
1946                                              r10_bio->devs[sl].addr +
1947                                              sect,
1948                                              s<<9, conf->tmppage,
1949                                                  READ)) {
1950                         case 0:
1951                                 /* Well, this device is dead */
1952                                 printk(KERN_NOTICE
1953                                        "md/raid10:%s: unable to read back "
1954                                        "corrected sectors"
1955                                        " (%d sectors at %llu on %s)\n",
1956                                        mdname(mddev), s,
1957                                        (unsigned long long)(
1958                                                sect + rdev->data_offset),
1959                                        bdevname(rdev->bdev, b));
1960                                 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
1961                                        "drive\n",
1962                                        mdname(mddev),
1963                                        bdevname(rdev->bdev, b));
1964                                 break;
1965                         case 1:
1966                                 printk(KERN_INFO
1967                                        "md/raid10:%s: read error corrected"
1968                                        " (%d sectors at %llu on %s)\n",
1969                                        mdname(mddev), s,
1970                                        (unsigned long long)(
1971                                                sect + rdev->data_offset),
1972                                        bdevname(rdev->bdev, b));
1973                                 atomic_add(s, &rdev->corrected_errors);
1974                         }
1975
1976                         rdev_dec_pending(rdev, mddev);
1977                         rcu_read_lock();
1978                 }
1979                 rcu_read_unlock();
1980
1981                 sectors -= s;
1982                 sect += s;
1983         }
1984 }
1985
1986 static void bi_complete(struct bio *bio, int error)
1987 {
1988         complete((struct completion *)bio->bi_private);
1989 }
1990
1991 static int submit_bio_wait(int rw, struct bio *bio)
1992 {
1993         struct completion event;
1994         rw |= REQ_SYNC;
1995
1996         init_completion(&event);
1997         bio->bi_private = &event;
1998         bio->bi_end_io = bi_complete;
1999         submit_bio(rw, bio);
2000         wait_for_completion(&event);
2001
2002         return test_bit(BIO_UPTODATE, &bio->bi_flags);
2003 }
2004
2005 static int narrow_write_error(struct r10bio *r10_bio, int i)
2006 {
2007         struct bio *bio = r10_bio->master_bio;
2008         struct mddev *mddev = r10_bio->mddev;
2009         struct r10conf *conf = mddev->private;
2010         struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2011         /* bio has the data to be written to slot 'i' where
2012          * we just recently had a write error.
2013          * We repeatedly clone the bio and trim down to one block,
2014          * then try the write.  Where the write fails we record
2015          * a bad block.
2016          * It is conceivable that the bio doesn't exactly align with
2017          * blocks.  We must handle this.
2018          *
2019          * We currently own a reference to the rdev.
2020          */
2021
2022         int block_sectors;
2023         sector_t sector;
2024         int sectors;
2025         int sect_to_write = r10_bio->sectors;
2026         int ok = 1;
2027
2028         if (rdev->badblocks.shift < 0)
2029                 return 0;
2030
2031         block_sectors = 1 << rdev->badblocks.shift;
2032         sector = r10_bio->sector;
2033         sectors = ((r10_bio->sector + block_sectors)
2034                    & ~(sector_t)(block_sectors - 1))
2035                 - sector;
2036
2037         while (sect_to_write) {
2038                 struct bio *wbio;
2039                 if (sectors > sect_to_write)
2040                         sectors = sect_to_write;
2041                 /* Write at 'sector' for 'sectors' */
2042                 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2043                 md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2044                 wbio->bi_sector = (r10_bio->devs[i].addr+
2045                                    rdev->data_offset+
2046                                    (sector - r10_bio->sector));
2047                 wbio->bi_bdev = rdev->bdev;
2048                 if (submit_bio_wait(WRITE, wbio) == 0)
2049                         /* Failure! */
2050                         ok = rdev_set_badblocks(rdev, sector,
2051                                                 sectors, 0)
2052                                 && ok;
2053
2054                 bio_put(wbio);
2055                 sect_to_write -= sectors;
2056                 sector += sectors;
2057                 sectors = block_sectors;
2058         }
2059         return ok;
2060 }
2061
2062 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2063 {
2064         int slot = r10_bio->read_slot;
2065         int mirror = r10_bio->devs[slot].devnum;
2066         struct bio *bio;
2067         struct r10conf *conf = mddev->private;
2068         struct md_rdev *rdev;
2069         char b[BDEVNAME_SIZE];
2070         unsigned long do_sync;
2071         int max_sectors;
2072
2073         /* we got a read error. Maybe the drive is bad.  Maybe just
2074          * the block and we can fix it.
2075          * We freeze all other IO, and try reading the block from
2076          * other devices.  When we find one, we re-write
2077          * and check it that fixes the read error.
2078          * This is all done synchronously while the array is
2079          * frozen.
2080          */
2081         if (mddev->ro == 0) {
2082                 freeze_array(conf);
2083                 fix_read_error(conf, mddev, r10_bio);
2084                 unfreeze_array(conf);
2085         }
2086         rdev_dec_pending(conf->mirrors[mirror].rdev, mddev);
2087
2088         bio = r10_bio->devs[slot].bio;
2089         bdevname(bio->bi_bdev, b);
2090         r10_bio->devs[slot].bio =
2091                 mddev->ro ? IO_BLOCKED : NULL;
2092 read_more:
2093         mirror = read_balance(conf, r10_bio, &max_sectors);
2094         if (mirror == -1) {
2095                 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2096                        " read error for block %llu\n",
2097                        mdname(mddev), b,
2098                        (unsigned long long)r10_bio->sector);
2099                 raid_end_bio_io(r10_bio);
2100                 bio_put(bio);
2101                 return;
2102         }
2103
2104         do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2105         if (bio)
2106                 bio_put(bio);
2107         slot = r10_bio->read_slot;
2108         rdev = conf->mirrors[mirror].rdev;
2109         printk_ratelimited(
2110                 KERN_ERR
2111                 "md/raid10:%s: %s: redirecting"
2112                 "sector %llu to another mirror\n",
2113                 mdname(mddev),
2114                 bdevname(rdev->bdev, b),
2115                 (unsigned long long)r10_bio->sector);
2116         bio = bio_clone_mddev(r10_bio->master_bio,
2117                               GFP_NOIO, mddev);
2118         md_trim_bio(bio,
2119                     r10_bio->sector - bio->bi_sector,
2120                     max_sectors);
2121         r10_bio->devs[slot].bio = bio;
2122         bio->bi_sector = r10_bio->devs[slot].addr
2123                 + rdev->data_offset;
2124         bio->bi_bdev = rdev->bdev;
2125         bio->bi_rw = READ | do_sync;
2126         bio->bi_private = r10_bio;
2127         bio->bi_end_io = raid10_end_read_request;
2128         if (max_sectors < r10_bio->sectors) {
2129                 /* Drat - have to split this up more */
2130                 struct bio *mbio = r10_bio->master_bio;
2131                 int sectors_handled =
2132                         r10_bio->sector + max_sectors
2133                         - mbio->bi_sector;
2134                 r10_bio->sectors = max_sectors;
2135                 spin_lock_irq(&conf->device_lock);
2136                 if (mbio->bi_phys_segments == 0)
2137                         mbio->bi_phys_segments = 2;
2138                 else
2139                         mbio->bi_phys_segments++;
2140                 spin_unlock_irq(&conf->device_lock);
2141                 generic_make_request(bio);
2142                 bio = NULL;
2143
2144                 r10_bio = mempool_alloc(conf->r10bio_pool,
2145                                         GFP_NOIO);
2146                 r10_bio->master_bio = mbio;
2147                 r10_bio->sectors = (mbio->bi_size >> 9)
2148                         - sectors_handled;
2149                 r10_bio->state = 0;
2150                 set_bit(R10BIO_ReadError,
2151                         &r10_bio->state);
2152                 r10_bio->mddev = mddev;
2153                 r10_bio->sector = mbio->bi_sector
2154                         + sectors_handled;
2155
2156                 goto read_more;
2157         } else
2158                 generic_make_request(bio);
2159 }
2160
2161 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2162 {
2163         /* Some sort of write request has finished and it
2164          * succeeded in writing where we thought there was a
2165          * bad block.  So forget the bad block.
2166          * Or possibly if failed and we need to record
2167          * a bad block.
2168          */
2169         int m;
2170         struct md_rdev *rdev;
2171
2172         if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2173             test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2174                 for (m = 0; m < conf->copies; m++) {
2175                         int dev = r10_bio->devs[m].devnum;
2176                         rdev = conf->mirrors[dev].rdev;
2177                         if (r10_bio->devs[m].bio == NULL)
2178                                 continue;
2179                         if (test_bit(BIO_UPTODATE,
2180                                      &r10_bio->devs[m].bio->bi_flags)) {
2181                                 rdev_clear_badblocks(
2182                                         rdev,
2183                                         r10_bio->devs[m].addr,
2184                                         r10_bio->sectors);
2185                         } else {
2186                                 if (!rdev_set_badblocks(
2187                                             rdev,
2188                                             r10_bio->devs[m].addr,
2189                                             r10_bio->sectors, 0))
2190                                         md_error(conf->mddev, rdev);
2191                         }
2192                 }
2193                 put_buf(r10_bio);
2194         } else {
2195                 for (m = 0; m < conf->copies; m++) {
2196                         int dev = r10_bio->devs[m].devnum;
2197                         struct bio *bio = r10_bio->devs[m].bio;
2198                         rdev = conf->mirrors[dev].rdev;
2199                         if (bio == IO_MADE_GOOD) {
2200                                 rdev_clear_badblocks(
2201                                         rdev,
2202                                         r10_bio->devs[m].addr,
2203                                         r10_bio->sectors);
2204                                 rdev_dec_pending(rdev, conf->mddev);
2205                         } else if (bio != NULL &&
2206                                    !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2207                                 if (!narrow_write_error(r10_bio, m)) {
2208                                         md_error(conf->mddev, rdev);
2209                                         set_bit(R10BIO_Degraded,
2210                                                 &r10_bio->state);
2211                                 }
2212                                 rdev_dec_pending(rdev, conf->mddev);
2213                         }
2214                 }
2215                 if (test_bit(R10BIO_WriteError,
2216                              &r10_bio->state))
2217                         close_write(r10_bio);
2218                 raid_end_bio_io(r10_bio);
2219         }
2220 }
2221
2222 static void raid10d(struct mddev *mddev)
2223 {
2224         struct r10bio *r10_bio;
2225         unsigned long flags;
2226         struct r10conf *conf = mddev->private;
2227         struct list_head *head = &conf->retry_list;
2228         struct blk_plug plug;
2229
2230         md_check_recovery(mddev);
2231
2232         blk_start_plug(&plug);
2233         for (;;) {
2234
2235                 flush_pending_writes(conf);
2236
2237                 spin_lock_irqsave(&conf->device_lock, flags);
2238                 if (list_empty(head)) {
2239                         spin_unlock_irqrestore(&conf->device_lock, flags);
2240                         break;
2241                 }
2242                 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2243                 list_del(head->prev);
2244                 conf->nr_queued--;
2245                 spin_unlock_irqrestore(&conf->device_lock, flags);
2246
2247                 mddev = r10_bio->mddev;
2248                 conf = mddev->private;
2249                 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2250                     test_bit(R10BIO_WriteError, &r10_bio->state))
2251                         handle_write_completed(conf, r10_bio);
2252                 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2253                         sync_request_write(mddev, r10_bio);
2254                 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2255                         recovery_request_write(mddev, r10_bio);
2256                 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2257                         handle_read_error(mddev, r10_bio);
2258                 else {
2259                         /* just a partial read to be scheduled from a
2260                          * separate context
2261                          */
2262                         int slot = r10_bio->read_slot;
2263                         generic_make_request(r10_bio->devs[slot].bio);
2264                 }
2265
2266                 cond_resched();
2267                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2268                         md_check_recovery(mddev);
2269         }
2270         blk_finish_plug(&plug);
2271 }
2272
2273
2274 static int init_resync(struct r10conf *conf)
2275 {
2276         int buffs;
2277
2278         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2279         BUG_ON(conf->r10buf_pool);
2280         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2281         if (!conf->r10buf_pool)
2282                 return -ENOMEM;
2283         conf->next_resync = 0;
2284         return 0;
2285 }
2286
2287 /*
2288  * perform a "sync" on one "block"
2289  *
2290  * We need to make sure that no normal I/O request - particularly write
2291  * requests - conflict with active sync requests.
2292  *
2293  * This is achieved by tracking pending requests and a 'barrier' concept
2294  * that can be installed to exclude normal IO requests.
2295  *
2296  * Resync and recovery are handled very differently.
2297  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2298  *
2299  * For resync, we iterate over virtual addresses, read all copies,
2300  * and update if there are differences.  If only one copy is live,
2301  * skip it.
2302  * For recovery, we iterate over physical addresses, read a good
2303  * value for each non-in_sync drive, and over-write.
2304  *
2305  * So, for recovery we may have several outstanding complex requests for a
2306  * given address, one for each out-of-sync device.  We model this by allocating
2307  * a number of r10_bio structures, one for each out-of-sync device.
2308  * As we setup these structures, we collect all bio's together into a list
2309  * which we then process collectively to add pages, and then process again
2310  * to pass to generic_make_request.
2311  *
2312  * The r10_bio structures are linked using a borrowed master_bio pointer.
2313  * This link is counted in ->remaining.  When the r10_bio that points to NULL
2314  * has its remaining count decremented to 0, the whole complex operation
2315  * is complete.
2316  *
2317  */
2318
2319 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
2320                              int *skipped, int go_faster)
2321 {
2322         struct r10conf *conf = mddev->private;
2323         struct r10bio *r10_bio;
2324         struct bio *biolist = NULL, *bio;
2325         sector_t max_sector, nr_sectors;
2326         int i;
2327         int max_sync;
2328         sector_t sync_blocks;
2329         sector_t sectors_skipped = 0;
2330         int chunks_skipped = 0;
2331
2332         if (!conf->r10buf_pool)
2333                 if (init_resync(conf))
2334                         return 0;
2335
2336  skipped:
2337         max_sector = mddev->dev_sectors;
2338         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2339                 max_sector = mddev->resync_max_sectors;
2340         if (sector_nr >= max_sector) {
2341                 /* If we aborted, we need to abort the
2342                  * sync on the 'current' bitmap chucks (there can
2343                  * be several when recovering multiple devices).
2344                  * as we may have started syncing it but not finished.
2345                  * We can find the current address in
2346                  * mddev->curr_resync, but for recovery,
2347                  * we need to convert that to several
2348                  * virtual addresses.
2349                  */
2350                 if (mddev->curr_resync < max_sector) { /* aborted */
2351                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2352                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2353                                                 &sync_blocks, 1);
2354                         else for (i=0; i<conf->raid_disks; i++) {
2355                                 sector_t sect =
2356                                         raid10_find_virt(conf, mddev->curr_resync, i);
2357                                 bitmap_end_sync(mddev->bitmap, sect,
2358                                                 &sync_blocks, 1);
2359                         }
2360                 } else /* completed sync */
2361                         conf->fullsync = 0;
2362
2363                 bitmap_close_sync(mddev->bitmap);
2364                 close_sync(conf);
2365                 *skipped = 1;
2366                 return sectors_skipped;
2367         }
2368         if (chunks_skipped >= conf->raid_disks) {
2369                 /* if there has been nothing to do on any drive,
2370                  * then there is nothing to do at all..
2371                  */
2372                 *skipped = 1;
2373                 return (max_sector - sector_nr) + sectors_skipped;
2374         }
2375
2376         if (max_sector > mddev->resync_max)
2377                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2378
2379         /* make sure whole request will fit in a chunk - if chunks
2380          * are meaningful
2381          */
2382         if (conf->near_copies < conf->raid_disks &&
2383             max_sector > (sector_nr | conf->chunk_mask))
2384                 max_sector = (sector_nr | conf->chunk_mask) + 1;
2385         /*
2386          * If there is non-resync activity waiting for us then
2387          * put in a delay to throttle resync.
2388          */
2389         if (!go_faster && conf->nr_waiting)
2390                 msleep_interruptible(1000);
2391
2392         /* Again, very different code for resync and recovery.
2393          * Both must result in an r10bio with a list of bios that
2394          * have bi_end_io, bi_sector, bi_bdev set,
2395          * and bi_private set to the r10bio.
2396          * For recovery, we may actually create several r10bios
2397          * with 2 bios in each, that correspond to the bios in the main one.
2398          * In this case, the subordinate r10bios link back through a
2399          * borrowed master_bio pointer, and the counter in the master
2400          * includes a ref from each subordinate.
2401          */
2402         /* First, we decide what to do and set ->bi_end_io
2403          * To end_sync_read if we want to read, and
2404          * end_sync_write if we will want to write.
2405          */
2406
2407         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2408         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2409                 /* recovery... the complicated one */
2410                 int j;
2411                 r10_bio = NULL;
2412
2413                 for (i=0 ; i<conf->raid_disks; i++) {
2414                         int still_degraded;
2415                         struct r10bio *rb2;
2416                         sector_t sect;
2417                         int must_sync;
2418                         int any_working;
2419
2420                         if (conf->mirrors[i].rdev == NULL ||
2421                             test_bit(In_sync, &conf->mirrors[i].rdev->flags)) 
2422                                 continue;
2423
2424                         still_degraded = 0;
2425                         /* want to reconstruct this device */
2426                         rb2 = r10_bio;
2427                         sect = raid10_find_virt(conf, sector_nr, i);
2428                         /* Unless we are doing a full sync, we only need
2429                          * to recover the block if it is set in the bitmap
2430                          */
2431                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
2432                                                       &sync_blocks, 1);
2433                         if (sync_blocks < max_sync)
2434                                 max_sync = sync_blocks;
2435                         if (!must_sync &&
2436                             !conf->fullsync) {
2437                                 /* yep, skip the sync_blocks here, but don't assume
2438                                  * that there will never be anything to do here
2439                                  */
2440                                 chunks_skipped = -1;
2441                                 continue;
2442                         }
2443
2444                         r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2445                         raise_barrier(conf, rb2 != NULL);
2446                         atomic_set(&r10_bio->remaining, 0);
2447
2448                         r10_bio->master_bio = (struct bio*)rb2;
2449                         if (rb2)
2450                                 atomic_inc(&rb2->remaining);
2451                         r10_bio->mddev = mddev;
2452                         set_bit(R10BIO_IsRecover, &r10_bio->state);
2453                         r10_bio->sector = sect;
2454
2455                         raid10_find_phys(conf, r10_bio);
2456
2457                         /* Need to check if the array will still be
2458                          * degraded
2459                          */
2460                         for (j=0; j<conf->raid_disks; j++)
2461                                 if (conf->mirrors[j].rdev == NULL ||
2462                                     test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
2463                                         still_degraded = 1;
2464                                         break;
2465                                 }
2466
2467                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
2468                                                       &sync_blocks, still_degraded);
2469
2470                         any_working = 0;
2471                         for (j=0; j<conf->copies;j++) {
2472                                 int k;
2473                                 int d = r10_bio->devs[j].devnum;
2474                                 sector_t from_addr, to_addr;
2475                                 struct md_rdev *rdev;
2476                                 sector_t sector, first_bad;
2477                                 int bad_sectors;
2478                                 if (!conf->mirrors[d].rdev ||
2479                                     !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
2480                                         continue;
2481                                 /* This is where we read from */
2482                                 any_working = 1;
2483                                 rdev = conf->mirrors[d].rdev;
2484                                 sector = r10_bio->devs[j].addr;
2485
2486                                 if (is_badblock(rdev, sector, max_sync,
2487                                                 &first_bad, &bad_sectors)) {
2488                                         if (first_bad > sector)
2489                                                 max_sync = first_bad - sector;
2490                                         else {
2491                                                 bad_sectors -= (sector
2492                                                                 - first_bad);
2493                                                 if (max_sync > bad_sectors)
2494                                                         max_sync = bad_sectors;
2495                                                 continue;
2496                                         }
2497                                 }
2498                                 bio = r10_bio->devs[0].bio;
2499                                 bio->bi_next = biolist;
2500                                 biolist = bio;
2501                                 bio->bi_private = r10_bio;
2502                                 bio->bi_end_io = end_sync_read;
2503                                 bio->bi_rw = READ;
2504                                 from_addr = r10_bio->devs[j].addr;
2505                                 bio->bi_sector = from_addr +
2506                                         conf->mirrors[d].rdev->data_offset;
2507                                 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2508                                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2509                                 atomic_inc(&r10_bio->remaining);
2510                                 /* and we write to 'i' */
2511
2512                                 for (k=0; k<conf->copies; k++)
2513                                         if (r10_bio->devs[k].devnum == i)
2514                                                 break;
2515                                 BUG_ON(k == conf->copies);
2516                                 bio = r10_bio->devs[1].bio;
2517                                 bio->bi_next = biolist;
2518                                 biolist = bio;
2519                                 bio->bi_private = r10_bio;
2520                                 bio->bi_end_io = end_sync_write;
2521                                 bio->bi_rw = WRITE;
2522                                 to_addr = r10_bio->devs[k].addr;
2523                                 bio->bi_sector = to_addr +
2524                                         conf->mirrors[i].rdev->data_offset;
2525                                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
2526
2527                                 r10_bio->devs[0].devnum = d;
2528                                 r10_bio->devs[0].addr = from_addr;
2529                                 r10_bio->devs[1].devnum = i;
2530                                 r10_bio->devs[1].addr = to_addr;
2531
2532                                 break;
2533                         }
2534                         if (j == conf->copies) {
2535                                 /* Cannot recover, so abort the recovery or
2536                                  * record a bad block */
2537                                 put_buf(r10_bio);
2538                                 if (rb2)
2539                                         atomic_dec(&rb2->remaining);
2540                                 r10_bio = rb2;
2541                                 if (any_working) {
2542                                         /* problem is that there are bad blocks
2543                                          * on other device(s)
2544                                          */
2545                                         int k;
2546                                         for (k = 0; k < conf->copies; k++)
2547                                                 if (r10_bio->devs[k].devnum == i)
2548                                                         break;
2549                                         if (!rdev_set_badblocks(
2550                                                     conf->mirrors[i].rdev,
2551                                                     r10_bio->devs[k].addr,
2552                                                     max_sync, 0))
2553                                                 any_working = 0;
2554                                 }
2555                                 if (!any_working)  {
2556                                         if (!test_and_set_bit(MD_RECOVERY_INTR,
2557                                                               &mddev->recovery))
2558                                                 printk(KERN_INFO "md/raid10:%s: insufficient "
2559                                                        "working devices for recovery.\n",
2560                                                        mdname(mddev));
2561                                         conf->mirrors[i].recovery_disabled
2562                                                 = mddev->recovery_disabled;
2563                                 }
2564                                 break;
2565                         }
2566                 }
2567                 if (biolist == NULL) {
2568                         while (r10_bio) {
2569                                 struct r10bio *rb2 = r10_bio;
2570                                 r10_bio = (struct r10bio*) rb2->master_bio;
2571                                 rb2->master_bio = NULL;
2572                                 put_buf(rb2);
2573                         }
2574                         goto giveup;
2575                 }
2576         } else {
2577                 /* resync. Schedule a read for every block at this virt offset */
2578                 int count = 0;
2579
2580                 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2581
2582                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2583                                        &sync_blocks, mddev->degraded) &&
2584                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
2585                                                  &mddev->recovery)) {
2586                         /* We can skip this block */
2587                         *skipped = 1;
2588                         return sync_blocks + sectors_skipped;
2589                 }
2590                 if (sync_blocks < max_sync)
2591                         max_sync = sync_blocks;
2592                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2593
2594                 r10_bio->mddev = mddev;
2595                 atomic_set(&r10_bio->remaining, 0);
2596                 raise_barrier(conf, 0);
2597                 conf->next_resync = sector_nr;
2598
2599                 r10_bio->master_bio = NULL;
2600                 r10_bio->sector = sector_nr;
2601                 set_bit(R10BIO_IsSync, &r10_bio->state);
2602                 raid10_find_phys(conf, r10_bio);
2603                 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
2604
2605                 for (i=0; i<conf->copies; i++) {
2606                         int d = r10_bio->devs[i].devnum;
2607                         sector_t first_bad, sector;
2608                         int bad_sectors;
2609
2610                         bio = r10_bio->devs[i].bio;
2611                         bio->bi_end_io = NULL;
2612                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
2613                         if (conf->mirrors[d].rdev == NULL ||
2614                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
2615                                 continue;
2616                         sector = r10_bio->devs[i].addr;
2617                         if (is_badblock(conf->mirrors[d].rdev,
2618                                         sector, max_sync,
2619                                         &first_bad, &bad_sectors)) {
2620                                 if (first_bad > sector)
2621                                         max_sync = first_bad - sector;
2622                                 else {
2623                                         bad_sectors -= (sector - first_bad);
2624                                         if (max_sync > bad_sectors)
2625                                                 max_sync = max_sync;
2626                                         continue;
2627                                 }
2628                         }
2629                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2630                         atomic_inc(&r10_bio->remaining);
2631                         bio->bi_next = biolist;
2632                         biolist = bio;
2633                         bio->bi_private = r10_bio;
2634                         bio->bi_end_io = end_sync_read;
2635                         bio->bi_rw = READ;
2636                         bio->bi_sector = sector +
2637                                 conf->mirrors[d].rdev->data_offset;
2638                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2639                         count++;
2640                 }
2641
2642                 if (count < 2) {
2643                         for (i=0; i<conf->copies; i++) {
2644                                 int d = r10_bio->devs[i].devnum;
2645                                 if (r10_bio->devs[i].bio->bi_end_io)
2646                                         rdev_dec_pending(conf->mirrors[d].rdev,
2647                                                          mddev);
2648                         }
2649                         put_buf(r10_bio);
2650                         biolist = NULL;
2651                         goto giveup;
2652                 }
2653         }
2654
2655         for (bio = biolist; bio ; bio=bio->bi_next) {
2656
2657                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
2658                 if (bio->bi_end_io)
2659                         bio->bi_flags |= 1 << BIO_UPTODATE;
2660                 bio->bi_vcnt = 0;
2661                 bio->bi_idx = 0;
2662                 bio->bi_phys_segments = 0;
2663                 bio->bi_size = 0;
2664         }
2665
2666         nr_sectors = 0;
2667         if (sector_nr + max_sync < max_sector)
2668                 max_sector = sector_nr + max_sync;
2669         do {
2670                 struct page *page;
2671                 int len = PAGE_SIZE;
2672                 if (sector_nr + (len>>9) > max_sector)
2673                         len = (max_sector - sector_nr) << 9;
2674                 if (len == 0)
2675                         break;
2676                 for (bio= biolist ; bio ; bio=bio->bi_next) {
2677                         struct bio *bio2;
2678                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2679                         if (bio_add_page(bio, page, len, 0))
2680                                 continue;
2681
2682                         /* stop here */
2683                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2684                         for (bio2 = biolist;
2685                              bio2 && bio2 != bio;
2686                              bio2 = bio2->bi_next) {
2687                                 /* remove last page from this bio */
2688                                 bio2->bi_vcnt--;
2689                                 bio2->bi_size -= len;
2690                                 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
2691                         }
2692                         goto bio_full;
2693                 }
2694                 nr_sectors += len>>9;
2695                 sector_nr += len>>9;
2696         } while (biolist->bi_vcnt < RESYNC_PAGES);
2697  bio_full:
2698         r10_bio->sectors = nr_sectors;
2699
2700         while (biolist) {
2701                 bio = biolist;
2702                 biolist = biolist->bi_next;
2703
2704                 bio->bi_next = NULL;
2705                 r10_bio = bio->bi_private;
2706                 r10_bio->sectors = nr_sectors;
2707
2708                 if (bio->bi_end_io == end_sync_read) {
2709                         md_sync_acct(bio->bi_bdev, nr_sectors);
2710                         generic_make_request(bio);
2711                 }
2712         }
2713
2714         if (sectors_skipped)
2715                 /* pretend they weren't skipped, it makes
2716                  * no important difference in this case
2717                  */
2718                 md_done_sync(mddev, sectors_skipped, 1);
2719
2720         return sectors_skipped + nr_sectors;
2721  giveup:
2722         /* There is nowhere to write, so all non-sync
2723          * drives must be failed or in resync, all drives
2724          * have a bad block, so try the next chunk...
2725          */
2726         if (sector_nr + max_sync < max_sector)
2727                 max_sector = sector_nr + max_sync;
2728
2729         sectors_skipped += (max_sector - sector_nr);
2730         chunks_skipped ++;
2731         sector_nr = max_sector;
2732         goto skipped;
2733 }
2734
2735 static sector_t
2736 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2737 {
2738         sector_t size;
2739         struct r10conf *conf = mddev->private;
2740
2741         if (!raid_disks)
2742                 raid_disks = conf->raid_disks;
2743         if (!sectors)
2744                 sectors = conf->dev_sectors;
2745
2746         size = sectors >> conf->chunk_shift;
2747         sector_div(size, conf->far_copies);
2748         size = size * raid_disks;
2749         sector_div(size, conf->near_copies);
2750
2751         return size << conf->chunk_shift;
2752 }
2753
2754
2755 static struct r10conf *setup_conf(struct mddev *mddev)
2756 {
2757         struct r10conf *conf = NULL;
2758         int nc, fc, fo;
2759         sector_t stride, size;
2760         int err = -EINVAL;
2761
2762         if (mddev->new_chunk_sectors < (PAGE_SIZE >> 9) ||
2763             !is_power_of_2(mddev->new_chunk_sectors)) {
2764                 printk(KERN_ERR "md/raid10:%s: chunk size must be "
2765                        "at least PAGE_SIZE(%ld) and be a power of 2.\n",
2766                        mdname(mddev), PAGE_SIZE);
2767                 goto out;
2768         }
2769
2770         nc = mddev->new_layout & 255;
2771         fc = (mddev->new_layout >> 8) & 255;
2772         fo = mddev->new_layout & (1<<16);
2773
2774         if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
2775             (mddev->new_layout >> 17)) {
2776                 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
2777                        mdname(mddev), mddev->new_layout);
2778                 goto out;
2779         }
2780
2781         err = -ENOMEM;
2782         conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
2783         if (!conf)
2784                 goto out;
2785
2786         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2787                                 GFP_KERNEL);
2788         if (!conf->mirrors)
2789                 goto out;
2790
2791         conf->tmppage = alloc_page(GFP_KERNEL);
2792         if (!conf->tmppage)
2793                 goto out;
2794
2795
2796         conf->raid_disks = mddev->raid_disks;
2797         conf->near_copies = nc;
2798         conf->far_copies = fc;
2799         conf->copies = nc*fc;
2800         conf->far_offset = fo;
2801         conf->chunk_mask = mddev->new_chunk_sectors - 1;
2802         conf->chunk_shift = ffz(~mddev->new_chunk_sectors);
2803
2804         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2805                                            r10bio_pool_free, conf);
2806         if (!conf->r10bio_pool)
2807                 goto out;
2808
2809         size = mddev->dev_sectors >> conf->chunk_shift;
2810         sector_div(size, fc);
2811         size = size * conf->raid_disks;
2812         sector_div(size, nc);
2813         /* 'size' is now the number of chunks in the array */
2814         /* calculate "used chunks per device" in 'stride' */
2815         stride = size * conf->copies;
2816
2817         /* We need to round up when dividing by raid_disks to
2818          * get the stride size.
2819          */
2820         stride += conf->raid_disks - 1;
2821         sector_div(stride, conf->raid_disks);
2822
2823         conf->dev_sectors = stride << conf->chunk_shift;
2824
2825         if (fo)
2826                 stride = 1;
2827         else
2828                 sector_div(stride, fc);
2829         conf->stride = stride << conf->chunk_shift;
2830
2831
2832         spin_lock_init(&conf->device_lock);
2833         INIT_LIST_HEAD(&conf->retry_list);
2834
2835         spin_lock_init(&conf->resync_lock);
2836         init_waitqueue_head(&conf->wait_barrier);
2837
2838         conf->thread = md_register_thread(raid10d, mddev, NULL);
2839         if (!conf->thread)
2840                 goto out;
2841
2842         conf->mddev = mddev;
2843         return conf;
2844
2845  out:
2846         printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
2847                mdname(mddev));
2848         if (conf) {
2849                 if (conf->r10bio_pool)
2850                         mempool_destroy(conf->r10bio_pool);
2851                 kfree(conf->mirrors);
2852                 safe_put_page(conf->tmppage);
2853                 kfree(conf);
2854         }
2855         return ERR_PTR(err);
2856 }
2857
2858 static int run(struct mddev *mddev)
2859 {
2860         struct r10conf *conf;
2861         int i, disk_idx, chunk_size;
2862         struct mirror_info *disk;
2863         struct md_rdev *rdev;
2864         sector_t size;
2865
2866         /*
2867          * copy the already verified devices into our private RAID10
2868          * bookkeeping area. [whatever we allocate in run(),
2869          * should be freed in stop()]
2870          */
2871
2872         if (mddev->private == NULL) {
2873                 conf = setup_conf(mddev);
2874                 if (IS_ERR(conf))
2875                         return PTR_ERR(conf);
2876                 mddev->private = conf;
2877         }
2878         conf = mddev->private;
2879         if (!conf)
2880                 goto out;
2881
2882         mddev->thread = conf->thread;
2883         conf->thread = NULL;
2884
2885         chunk_size = mddev->chunk_sectors << 9;
2886         blk_queue_io_min(mddev->queue, chunk_size);
2887         if (conf->raid_disks % conf->near_copies)
2888                 blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks);
2889         else
2890                 blk_queue_io_opt(mddev->queue, chunk_size *
2891                                  (conf->raid_disks / conf->near_copies));
2892
2893         list_for_each_entry(rdev, &mddev->disks, same_set) {
2894
2895                 disk_idx = rdev->raid_disk;
2896                 if (disk_idx >= conf->raid_disks
2897                     || disk_idx < 0)
2898                         continue;
2899                 disk = conf->mirrors + disk_idx;
2900
2901                 disk->rdev = rdev;
2902                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2903                                   rdev->data_offset << 9);
2904                 /* as we don't honour merge_bvec_fn, we must never risk
2905                  * violating it, so limit max_segments to 1 lying
2906                  * within a single page.
2907                  */
2908                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2909                         blk_queue_max_segments(mddev->queue, 1);
2910                         blk_queue_segment_boundary(mddev->queue,
2911                                                    PAGE_CACHE_SIZE - 1);
2912                 }
2913
2914                 disk->head_position = 0;
2915         }
2916         /* need to check that every block has at least one working mirror */
2917         if (!enough(conf, -1)) {
2918                 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
2919                        mdname(mddev));
2920                 goto out_free_conf;
2921         }
2922
2923         mddev->degraded = 0;
2924         for (i = 0; i < conf->raid_disks; i++) {
2925
2926                 disk = conf->mirrors + i;
2927
2928                 if (!disk->rdev ||
2929                     !test_bit(In_sync, &disk->rdev->flags)) {
2930                         disk->head_position = 0;
2931                         mddev->degraded++;
2932                         if (disk->rdev)
2933                                 conf->fullsync = 1;
2934                 }
2935                 disk->recovery_disabled = mddev->recovery_disabled - 1;
2936         }
2937
2938         if (mddev->recovery_cp != MaxSector)
2939                 printk(KERN_NOTICE "md/raid10:%s: not clean"
2940                        " -- starting background reconstruction\n",
2941                        mdname(mddev));
2942         printk(KERN_INFO
2943                 "md/raid10:%s: active with %d out of %d devices\n",
2944                 mdname(mddev), conf->raid_disks - mddev->degraded,
2945                 conf->raid_disks);
2946         /*
2947          * Ok, everything is just fine now
2948          */
2949         mddev->dev_sectors = conf->dev_sectors;
2950         size = raid10_size(mddev, 0, 0);
2951         md_set_array_sectors(mddev, size);
2952         mddev->resync_max_sectors = size;
2953
2954         mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2955         mddev->queue->backing_dev_info.congested_data = mddev;
2956
2957         /* Calculate max read-ahead size.
2958          * We need to readahead at least twice a whole stripe....
2959          * maybe...
2960          */
2961         {
2962                 int stripe = conf->raid_disks *
2963                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
2964                 stripe /= conf->near_copies;
2965                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2966                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2967         }
2968
2969         if (conf->near_copies < conf->raid_disks)
2970                 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2971
2972         if (md_integrity_register(mddev))
2973                 goto out_free_conf;
2974
2975         return 0;
2976
2977 out_free_conf:
2978         md_unregister_thread(&mddev->thread);
2979         if (conf->r10bio_pool)
2980                 mempool_destroy(conf->r10bio_pool);
2981         safe_put_page(conf->tmppage);
2982         kfree(conf->mirrors);
2983         kfree(conf);
2984         mddev->private = NULL;
2985 out:
2986         return -EIO;
2987 }
2988
2989 static int stop(struct mddev *mddev)
2990 {
2991         struct r10conf *conf = mddev->private;
2992
2993         raise_barrier(conf, 0);
2994         lower_barrier(conf);
2995
2996         md_unregister_thread(&mddev->thread);
2997         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2998         if (conf->r10bio_pool)
2999                 mempool_destroy(conf->r10bio_pool);
3000         kfree(conf->mirrors);
3001         kfree(conf);
3002         mddev->private = NULL;
3003         return 0;
3004 }
3005
3006 static void raid10_quiesce(struct mddev *mddev, int state)
3007 {
3008         struct r10conf *conf = mddev->private;
3009
3010         switch(state) {
3011         case 1:
3012                 raise_barrier(conf, 0);
3013                 break;
3014         case 0:
3015                 lower_barrier(conf);
3016                 break;
3017         }
3018 }
3019
3020 static void *raid10_takeover_raid0(struct mddev *mddev)
3021 {
3022         struct md_rdev *rdev;
3023         struct r10conf *conf;
3024
3025         if (mddev->degraded > 0) {
3026                 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3027                        mdname(mddev));
3028                 return ERR_PTR(-EINVAL);
3029         }
3030
3031         /* Set new parameters */
3032         mddev->new_level = 10;
3033         /* new layout: far_copies = 1, near_copies = 2 */
3034         mddev->new_layout = (1<<8) + 2;
3035         mddev->new_chunk_sectors = mddev->chunk_sectors;
3036         mddev->delta_disks = mddev->raid_disks;
3037         mddev->raid_disks *= 2;
3038         /* make sure it will be not marked as dirty */
3039         mddev->recovery_cp = MaxSector;
3040
3041         conf = setup_conf(mddev);
3042         if (!IS_ERR(conf)) {
3043                 list_for_each_entry(rdev, &mddev->disks, same_set)
3044                         if (rdev->raid_disk >= 0)
3045                                 rdev->new_raid_disk = rdev->raid_disk * 2;
3046                 conf->barrier = 1;
3047         }
3048
3049         return conf;
3050 }
3051
3052 static void *raid10_takeover(struct mddev *mddev)
3053 {
3054         struct r0conf *raid0_conf;
3055
3056         /* raid10 can take over:
3057          *  raid0 - providing it has only two drives
3058          */
3059         if (mddev->level == 0) {
3060                 /* for raid0 takeover only one zone is supported */
3061                 raid0_conf = mddev->private;
3062                 if (raid0_conf->nr_strip_zones > 1) {
3063                         printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3064                                " with more than one zone.\n",
3065                                mdname(mddev));
3066                         return ERR_PTR(-EINVAL);
3067                 }
3068                 return raid10_takeover_raid0(mddev);
3069         }
3070         return ERR_PTR(-EINVAL);
3071 }
3072
3073 static struct md_personality raid10_personality =
3074 {
3075         .name           = "raid10",
3076         .level          = 10,
3077         .owner          = THIS_MODULE,
3078         .make_request   = make_request,
3079         .run            = run,
3080         .stop           = stop,
3081         .status         = status,
3082         .error_handler  = error,
3083         .hot_add_disk   = raid10_add_disk,
3084         .hot_remove_disk= raid10_remove_disk,
3085         .spare_active   = raid10_spare_active,
3086         .sync_request   = sync_request,
3087         .quiesce        = raid10_quiesce,
3088         .size           = raid10_size,
3089         .takeover       = raid10_takeover,
3090 };
3091
3092 static int __init raid_init(void)
3093 {
3094         return register_md_personality(&raid10_personality);
3095 }
3096
3097 static void raid_exit(void)
3098 {
3099         unregister_md_personality(&raid10_personality);
3100 }
3101
3102 module_init(raid_init);
3103 module_exit(raid_exit);
3104 MODULE_LICENSE("GPL");
3105 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
3106 MODULE_ALIAS("md-personality-9"); /* RAID10 */
3107 MODULE_ALIAS("md-raid10");
3108 MODULE_ALIAS("md-level-10");
3109
3110 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);