36a8fc059ac36efc068603c51348132324d5d399
[linux-drm-fsl-dcu.git] / drivers / md / raid1.c
1 /*
2  * raid1.c : Multiple Devices driver for Linux
3  *
4  * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
5  *
6  * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
7  *
8  * RAID-1 management functions.
9  *
10  * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
11  *
12  * Fixes to reconstruction by Jakob Ć˜stergaard" <jakob@ostenfeld.dk>
13  * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
14  *
15  * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
16  * bitmapped intelligence in resync:
17  *
18  *      - bitmap marked during normal i/o
19  *      - bitmap used to skip nondirty blocks during sync
20  *
21  * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
22  * - persistent bitmap code
23  *
24  * This program is free software; you can redistribute it and/or modify
25  * it under the terms of the GNU General Public License as published by
26  * the Free Software Foundation; either version 2, or (at your option)
27  * any later version.
28  *
29  * You should have received a copy of the GNU General Public License
30  * (for example /usr/src/linux/COPYING); if not, write to the Free
31  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
32  */
33
34 #include <linux/slab.h>
35 #include <linux/delay.h>
36 #include <linux/blkdev.h>
37 #include <linux/module.h>
38 #include <linux/seq_file.h>
39 #include <linux/ratelimit.h>
40 #include "md.h"
41 #include "raid1.h"
42 #include "bitmap.h"
43
44 /*
45  * Number of guaranteed r1bios in case of extreme VM load:
46  */
47 #define NR_RAID1_BIOS 256
48
49 /* When there are this many requests queue to be written by
50  * the raid1 thread, we become 'congested' to provide back-pressure
51  * for writeback.
52  */
53 static int max_queued_requests = 1024;
54
55 static void allow_barrier(struct r1conf *conf);
56 static void lower_barrier(struct r1conf *conf);
57
58 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
59 {
60         struct pool_info *pi = data;
61         int size = offsetof(struct r1bio, bios[pi->raid_disks]);
62
63         /* allocate a r1bio with room for raid_disks entries in the bios array */
64         return kzalloc(size, gfp_flags);
65 }
66
67 static void r1bio_pool_free(void *r1_bio, void *data)
68 {
69         kfree(r1_bio);
70 }
71
72 #define RESYNC_BLOCK_SIZE (64*1024)
73 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
74 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
75 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
76 #define RESYNC_WINDOW (2048*1024)
77
78 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
79 {
80         struct pool_info *pi = data;
81         struct page *page;
82         struct r1bio *r1_bio;
83         struct bio *bio;
84         int i, j;
85
86         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
87         if (!r1_bio)
88                 return NULL;
89
90         /*
91          * Allocate bios : 1 for reading, n-1 for writing
92          */
93         for (j = pi->raid_disks ; j-- ; ) {
94                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
95                 if (!bio)
96                         goto out_free_bio;
97                 r1_bio->bios[j] = bio;
98         }
99         /*
100          * Allocate RESYNC_PAGES data pages and attach them to
101          * the first bio.
102          * If this is a user-requested check/repair, allocate
103          * RESYNC_PAGES for each bio.
104          */
105         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
106                 j = pi->raid_disks;
107         else
108                 j = 1;
109         while(j--) {
110                 bio = r1_bio->bios[j];
111                 for (i = 0; i < RESYNC_PAGES; i++) {
112                         page = alloc_page(gfp_flags);
113                         if (unlikely(!page))
114                                 goto out_free_pages;
115
116                         bio->bi_io_vec[i].bv_page = page;
117                         bio->bi_vcnt = i+1;
118                 }
119         }
120         /* If not user-requests, copy the page pointers to all bios */
121         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
122                 for (i=0; i<RESYNC_PAGES ; i++)
123                         for (j=1; j<pi->raid_disks; j++)
124                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
125                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
126         }
127
128         r1_bio->master_bio = NULL;
129
130         return r1_bio;
131
132 out_free_pages:
133         for (j=0 ; j < pi->raid_disks; j++)
134                 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
135                         put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
136         j = -1;
137 out_free_bio:
138         while (++j < pi->raid_disks)
139                 bio_put(r1_bio->bios[j]);
140         r1bio_pool_free(r1_bio, data);
141         return NULL;
142 }
143
144 static void r1buf_pool_free(void *__r1_bio, void *data)
145 {
146         struct pool_info *pi = data;
147         int i,j;
148         struct r1bio *r1bio = __r1_bio;
149
150         for (i = 0; i < RESYNC_PAGES; i++)
151                 for (j = pi->raid_disks; j-- ;) {
152                         if (j == 0 ||
153                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
154                             r1bio->bios[0]->bi_io_vec[i].bv_page)
155                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
156                 }
157         for (i=0 ; i < pi->raid_disks; i++)
158                 bio_put(r1bio->bios[i]);
159
160         r1bio_pool_free(r1bio, data);
161 }
162
163 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
164 {
165         int i;
166
167         for (i = 0; i < conf->raid_disks * 2; i++) {
168                 struct bio **bio = r1_bio->bios + i;
169                 if (!BIO_SPECIAL(*bio))
170                         bio_put(*bio);
171                 *bio = NULL;
172         }
173 }
174
175 static void free_r1bio(struct r1bio *r1_bio)
176 {
177         struct r1conf *conf = r1_bio->mddev->private;
178
179         put_all_bios(conf, r1_bio);
180         mempool_free(r1_bio, conf->r1bio_pool);
181 }
182
183 static void put_buf(struct r1bio *r1_bio)
184 {
185         struct r1conf *conf = r1_bio->mddev->private;
186         int i;
187
188         for (i = 0; i < conf->raid_disks * 2; i++) {
189                 struct bio *bio = r1_bio->bios[i];
190                 if (bio->bi_end_io)
191                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
192         }
193
194         mempool_free(r1_bio, conf->r1buf_pool);
195
196         lower_barrier(conf);
197 }
198
199 static void reschedule_retry(struct r1bio *r1_bio)
200 {
201         unsigned long flags;
202         struct mddev *mddev = r1_bio->mddev;
203         struct r1conf *conf = mddev->private;
204
205         spin_lock_irqsave(&conf->device_lock, flags);
206         list_add(&r1_bio->retry_list, &conf->retry_list);
207         conf->nr_queued ++;
208         spin_unlock_irqrestore(&conf->device_lock, flags);
209
210         wake_up(&conf->wait_barrier);
211         md_wakeup_thread(mddev->thread);
212 }
213
214 /*
215  * raid_end_bio_io() is called when we have finished servicing a mirrored
216  * operation and are ready to return a success/failure code to the buffer
217  * cache layer.
218  */
219 static void call_bio_endio(struct r1bio *r1_bio)
220 {
221         struct bio *bio = r1_bio->master_bio;
222         int done;
223         struct r1conf *conf = r1_bio->mddev->private;
224
225         if (bio->bi_phys_segments) {
226                 unsigned long flags;
227                 spin_lock_irqsave(&conf->device_lock, flags);
228                 bio->bi_phys_segments--;
229                 done = (bio->bi_phys_segments == 0);
230                 spin_unlock_irqrestore(&conf->device_lock, flags);
231         } else
232                 done = 1;
233
234         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
235                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
236         if (done) {
237                 bio_endio(bio, 0);
238                 /*
239                  * Wake up any possible resync thread that waits for the device
240                  * to go idle.
241                  */
242                 allow_barrier(conf);
243         }
244 }
245
246 static void raid_end_bio_io(struct r1bio *r1_bio)
247 {
248         struct bio *bio = r1_bio->master_bio;
249
250         /* if nobody has done the final endio yet, do it now */
251         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
252                 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
253                          (bio_data_dir(bio) == WRITE) ? "write" : "read",
254                          (unsigned long long) bio->bi_sector,
255                          (unsigned long long) bio->bi_sector +
256                          (bio->bi_size >> 9) - 1);
257
258                 call_bio_endio(r1_bio);
259         }
260         free_r1bio(r1_bio);
261 }
262
263 /*
264  * Update disk head position estimator based on IRQ completion info.
265  */
266 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
267 {
268         struct r1conf *conf = r1_bio->mddev->private;
269
270         conf->mirrors[disk].head_position =
271                 r1_bio->sector + (r1_bio->sectors);
272 }
273
274 /*
275  * Find the disk number which triggered given bio
276  */
277 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
278 {
279         int mirror;
280         struct r1conf *conf = r1_bio->mddev->private;
281         int raid_disks = conf->raid_disks;
282
283         for (mirror = 0; mirror < raid_disks * 2; mirror++)
284                 if (r1_bio->bios[mirror] == bio)
285                         break;
286
287         BUG_ON(mirror == raid_disks * 2);
288         update_head_pos(mirror, r1_bio);
289
290         return mirror;
291 }
292
293 static void raid1_end_read_request(struct bio *bio, int error)
294 {
295         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
296         struct r1bio *r1_bio = bio->bi_private;
297         int mirror;
298         struct r1conf *conf = r1_bio->mddev->private;
299
300         mirror = r1_bio->read_disk;
301         /*
302          * this branch is our 'one mirror IO has finished' event handler:
303          */
304         update_head_pos(mirror, r1_bio);
305
306         if (uptodate)
307                 set_bit(R1BIO_Uptodate, &r1_bio->state);
308         else {
309                 /* If all other devices have failed, we want to return
310                  * the error upwards rather than fail the last device.
311                  * Here we redefine "uptodate" to mean "Don't want to retry"
312                  */
313                 unsigned long flags;
314                 spin_lock_irqsave(&conf->device_lock, flags);
315                 if (r1_bio->mddev->degraded == conf->raid_disks ||
316                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
317                      !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
318                         uptodate = 1;
319                 spin_unlock_irqrestore(&conf->device_lock, flags);
320         }
321
322         if (uptodate)
323                 raid_end_bio_io(r1_bio);
324         else {
325                 /*
326                  * oops, read error:
327                  */
328                 char b[BDEVNAME_SIZE];
329                 printk_ratelimited(
330                         KERN_ERR "md/raid1:%s: %s: "
331                         "rescheduling sector %llu\n",
332                         mdname(conf->mddev),
333                         bdevname(conf->mirrors[mirror].rdev->bdev,
334                                  b),
335                         (unsigned long long)r1_bio->sector);
336                 set_bit(R1BIO_ReadError, &r1_bio->state);
337                 reschedule_retry(r1_bio);
338         }
339
340         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
341 }
342
343 static void close_write(struct r1bio *r1_bio)
344 {
345         /* it really is the end of this request */
346         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
347                 /* free extra copy of the data pages */
348                 int i = r1_bio->behind_page_count;
349                 while (i--)
350                         safe_put_page(r1_bio->behind_bvecs[i].bv_page);
351                 kfree(r1_bio->behind_bvecs);
352                 r1_bio->behind_bvecs = NULL;
353         }
354         /* clear the bitmap if all writes complete successfully */
355         bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
356                         r1_bio->sectors,
357                         !test_bit(R1BIO_Degraded, &r1_bio->state),
358                         test_bit(R1BIO_BehindIO, &r1_bio->state));
359         md_write_end(r1_bio->mddev);
360 }
361
362 static void r1_bio_write_done(struct r1bio *r1_bio)
363 {
364         if (!atomic_dec_and_test(&r1_bio->remaining))
365                 return;
366
367         if (test_bit(R1BIO_WriteError, &r1_bio->state))
368                 reschedule_retry(r1_bio);
369         else {
370                 close_write(r1_bio);
371                 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
372                         reschedule_retry(r1_bio);
373                 else
374                         raid_end_bio_io(r1_bio);
375         }
376 }
377
378 static void raid1_end_write_request(struct bio *bio, int error)
379 {
380         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
381         struct r1bio *r1_bio = bio->bi_private;
382         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
383         struct r1conf *conf = r1_bio->mddev->private;
384         struct bio *to_put = NULL;
385
386         mirror = find_bio_disk(r1_bio, bio);
387
388         /*
389          * 'one mirror IO has finished' event handler:
390          */
391         if (!uptodate) {
392                 set_bit(WriteErrorSeen,
393                         &conf->mirrors[mirror].rdev->flags);
394                 if (!test_and_set_bit(WantReplacement,
395                                       &conf->mirrors[mirror].rdev->flags))
396                         set_bit(MD_RECOVERY_NEEDED, &
397                                 conf->mddev->recovery);
398
399                 set_bit(R1BIO_WriteError, &r1_bio->state);
400         } else {
401                 /*
402                  * Set R1BIO_Uptodate in our master bio, so that we
403                  * will return a good error code for to the higher
404                  * levels even if IO on some other mirrored buffer
405                  * fails.
406                  *
407                  * The 'master' represents the composite IO operation
408                  * to user-side. So if something waits for IO, then it
409                  * will wait for the 'master' bio.
410                  */
411                 sector_t first_bad;
412                 int bad_sectors;
413
414                 r1_bio->bios[mirror] = NULL;
415                 to_put = bio;
416                 set_bit(R1BIO_Uptodate, &r1_bio->state);
417
418                 /* Maybe we can clear some bad blocks. */
419                 if (is_badblock(conf->mirrors[mirror].rdev,
420                                 r1_bio->sector, r1_bio->sectors,
421                                 &first_bad, &bad_sectors)) {
422                         r1_bio->bios[mirror] = IO_MADE_GOOD;
423                         set_bit(R1BIO_MadeGood, &r1_bio->state);
424                 }
425         }
426
427         if (behind) {
428                 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
429                         atomic_dec(&r1_bio->behind_remaining);
430
431                 /*
432                  * In behind mode, we ACK the master bio once the I/O
433                  * has safely reached all non-writemostly
434                  * disks. Setting the Returned bit ensures that this
435                  * gets done only once -- we don't ever want to return
436                  * -EIO here, instead we'll wait
437                  */
438                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
439                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
440                         /* Maybe we can return now */
441                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
442                                 struct bio *mbio = r1_bio->master_bio;
443                                 pr_debug("raid1: behind end write sectors"
444                                          " %llu-%llu\n",
445                                          (unsigned long long) mbio->bi_sector,
446                                          (unsigned long long) mbio->bi_sector +
447                                          (mbio->bi_size >> 9) - 1);
448                                 call_bio_endio(r1_bio);
449                         }
450                 }
451         }
452         if (r1_bio->bios[mirror] == NULL)
453                 rdev_dec_pending(conf->mirrors[mirror].rdev,
454                                  conf->mddev);
455
456         /*
457          * Let's see if all mirrored write operations have finished
458          * already.
459          */
460         r1_bio_write_done(r1_bio);
461
462         if (to_put)
463                 bio_put(to_put);
464 }
465
466
467 /*
468  * This routine returns the disk from which the requested read should
469  * be done. There is a per-array 'next expected sequential IO' sector
470  * number - if this matches on the next IO then we use the last disk.
471  * There is also a per-disk 'last know head position' sector that is
472  * maintained from IRQ contexts, both the normal and the resync IO
473  * completion handlers update this position correctly. If there is no
474  * perfect sequential match then we pick the disk whose head is closest.
475  *
476  * If there are 2 mirrors in the same 2 devices, performance degrades
477  * because position is mirror, not device based.
478  *
479  * The rdev for the device selected will have nr_pending incremented.
480  */
481 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
482 {
483         const sector_t this_sector = r1_bio->sector;
484         int sectors;
485         int best_good_sectors;
486         int start_disk;
487         int best_disk;
488         int i;
489         sector_t best_dist;
490         struct md_rdev *rdev;
491         int choose_first;
492
493         rcu_read_lock();
494         /*
495          * Check if we can balance. We can balance on the whole
496          * device if no resync is going on, or below the resync window.
497          * We take the first readable disk when above the resync window.
498          */
499  retry:
500         sectors = r1_bio->sectors;
501         best_disk = -1;
502         best_dist = MaxSector;
503         best_good_sectors = 0;
504
505         if (conf->mddev->recovery_cp < MaxSector &&
506             (this_sector + sectors >= conf->next_resync)) {
507                 choose_first = 1;
508                 start_disk = 0;
509         } else {
510                 choose_first = 0;
511                 start_disk = conf->last_used;
512         }
513
514         for (i = 0 ; i < conf->raid_disks * 2 ; i++) {
515                 sector_t dist;
516                 sector_t first_bad;
517                 int bad_sectors;
518
519                 int disk = start_disk + i;
520                 if (disk >= conf->raid_disks * 2)
521                         disk -= conf->raid_disks * 2;
522
523                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
524                 if (r1_bio->bios[disk] == IO_BLOCKED
525                     || rdev == NULL
526                     || test_bit(Unmerged, &rdev->flags)
527                     || test_bit(Faulty, &rdev->flags))
528                         continue;
529                 if (!test_bit(In_sync, &rdev->flags) &&
530                     rdev->recovery_offset < this_sector + sectors)
531                         continue;
532                 if (test_bit(WriteMostly, &rdev->flags)) {
533                         /* Don't balance among write-mostly, just
534                          * use the first as a last resort */
535                         if (best_disk < 0) {
536                                 if (is_badblock(rdev, this_sector, sectors,
537                                                 &first_bad, &bad_sectors)) {
538                                         if (first_bad < this_sector)
539                                                 /* Cannot use this */
540                                                 continue;
541                                         best_good_sectors = first_bad - this_sector;
542                                 } else
543                                         best_good_sectors = sectors;
544                                 best_disk = disk;
545                         }
546                         continue;
547                 }
548                 /* This is a reasonable device to use.  It might
549                  * even be best.
550                  */
551                 if (is_badblock(rdev, this_sector, sectors,
552                                 &first_bad, &bad_sectors)) {
553                         if (best_dist < MaxSector)
554                                 /* already have a better device */
555                                 continue;
556                         if (first_bad <= this_sector) {
557                                 /* cannot read here. If this is the 'primary'
558                                  * device, then we must not read beyond
559                                  * bad_sectors from another device..
560                                  */
561                                 bad_sectors -= (this_sector - first_bad);
562                                 if (choose_first && sectors > bad_sectors)
563                                         sectors = bad_sectors;
564                                 if (best_good_sectors > sectors)
565                                         best_good_sectors = sectors;
566
567                         } else {
568                                 sector_t good_sectors = first_bad - this_sector;
569                                 if (good_sectors > best_good_sectors) {
570                                         best_good_sectors = good_sectors;
571                                         best_disk = disk;
572                                 }
573                                 if (choose_first)
574                                         break;
575                         }
576                         continue;
577                 } else
578                         best_good_sectors = sectors;
579
580                 dist = abs(this_sector - conf->mirrors[disk].head_position);
581                 if (choose_first
582                     /* Don't change to another disk for sequential reads */
583                     || conf->next_seq_sect == this_sector
584                     || dist == 0
585                     /* If device is idle, use it */
586                     || atomic_read(&rdev->nr_pending) == 0) {
587                         best_disk = disk;
588                         break;
589                 }
590                 if (dist < best_dist) {
591                         best_dist = dist;
592                         best_disk = disk;
593                 }
594         }
595
596         if (best_disk >= 0) {
597                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
598                 if (!rdev)
599                         goto retry;
600                 atomic_inc(&rdev->nr_pending);
601                 if (test_bit(Faulty, &rdev->flags)) {
602                         /* cannot risk returning a device that failed
603                          * before we inc'ed nr_pending
604                          */
605                         rdev_dec_pending(rdev, conf->mddev);
606                         goto retry;
607                 }
608                 sectors = best_good_sectors;
609                 conf->next_seq_sect = this_sector + sectors;
610                 conf->last_used = best_disk;
611         }
612         rcu_read_unlock();
613         *max_sectors = sectors;
614
615         return best_disk;
616 }
617
618 static int raid1_mergeable_bvec(struct request_queue *q,
619                                 struct bvec_merge_data *bvm,
620                                 struct bio_vec *biovec)
621 {
622         struct mddev *mddev = q->queuedata;
623         struct r1conf *conf = mddev->private;
624         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
625         int max = biovec->bv_len;
626
627         if (mddev->merge_check_needed) {
628                 int disk;
629                 rcu_read_lock();
630                 for (disk = 0; disk < conf->raid_disks * 2; disk++) {
631                         struct md_rdev *rdev = rcu_dereference(
632                                 conf->mirrors[disk].rdev);
633                         if (rdev && !test_bit(Faulty, &rdev->flags)) {
634                                 struct request_queue *q =
635                                         bdev_get_queue(rdev->bdev);
636                                 if (q->merge_bvec_fn) {
637                                         bvm->bi_sector = sector +
638                                                 rdev->data_offset;
639                                         bvm->bi_bdev = rdev->bdev;
640                                         max = min(max, q->merge_bvec_fn(
641                                                           q, bvm, biovec));
642                                 }
643                         }
644                 }
645                 rcu_read_unlock();
646         }
647         return max;
648
649 }
650
651 int md_raid1_congested(struct mddev *mddev, int bits)
652 {
653         struct r1conf *conf = mddev->private;
654         int i, ret = 0;
655
656         if ((bits & (1 << BDI_async_congested)) &&
657             conf->pending_count >= max_queued_requests)
658                 return 1;
659
660         rcu_read_lock();
661         for (i = 0; i < conf->raid_disks * 2; i++) {
662                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
663                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
664                         struct request_queue *q = bdev_get_queue(rdev->bdev);
665
666                         BUG_ON(!q);
667
668                         /* Note the '|| 1' - when read_balance prefers
669                          * non-congested targets, it can be removed
670                          */
671                         if ((bits & (1<<BDI_async_congested)) || 1)
672                                 ret |= bdi_congested(&q->backing_dev_info, bits);
673                         else
674                                 ret &= bdi_congested(&q->backing_dev_info, bits);
675                 }
676         }
677         rcu_read_unlock();
678         return ret;
679 }
680 EXPORT_SYMBOL_GPL(md_raid1_congested);
681
682 static int raid1_congested(void *data, int bits)
683 {
684         struct mddev *mddev = data;
685
686         return mddev_congested(mddev, bits) ||
687                 md_raid1_congested(mddev, bits);
688 }
689
690 static void flush_pending_writes(struct r1conf *conf)
691 {
692         /* Any writes that have been queued but are awaiting
693          * bitmap updates get flushed here.
694          */
695         spin_lock_irq(&conf->device_lock);
696
697         if (conf->pending_bio_list.head) {
698                 struct bio *bio;
699                 bio = bio_list_get(&conf->pending_bio_list);
700                 conf->pending_count = 0;
701                 spin_unlock_irq(&conf->device_lock);
702                 /* flush any pending bitmap writes to
703                  * disk before proceeding w/ I/O */
704                 bitmap_unplug(conf->mddev->bitmap);
705                 wake_up(&conf->wait_barrier);
706
707                 while (bio) { /* submit pending writes */
708                         struct bio *next = bio->bi_next;
709                         bio->bi_next = NULL;
710                         generic_make_request(bio);
711                         bio = next;
712                 }
713         } else
714                 spin_unlock_irq(&conf->device_lock);
715 }
716
717 /* Barriers....
718  * Sometimes we need to suspend IO while we do something else,
719  * either some resync/recovery, or reconfigure the array.
720  * To do this we raise a 'barrier'.
721  * The 'barrier' is a counter that can be raised multiple times
722  * to count how many activities are happening which preclude
723  * normal IO.
724  * We can only raise the barrier if there is no pending IO.
725  * i.e. if nr_pending == 0.
726  * We choose only to raise the barrier if no-one is waiting for the
727  * barrier to go down.  This means that as soon as an IO request
728  * is ready, no other operations which require a barrier will start
729  * until the IO request has had a chance.
730  *
731  * So: regular IO calls 'wait_barrier'.  When that returns there
732  *    is no backgroup IO happening,  It must arrange to call
733  *    allow_barrier when it has finished its IO.
734  * backgroup IO calls must call raise_barrier.  Once that returns
735  *    there is no normal IO happeing.  It must arrange to call
736  *    lower_barrier when the particular background IO completes.
737  */
738 #define RESYNC_DEPTH 32
739
740 static void raise_barrier(struct r1conf *conf)
741 {
742         spin_lock_irq(&conf->resync_lock);
743
744         /* Wait until no block IO is waiting */
745         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
746                             conf->resync_lock, );
747
748         /* block any new IO from starting */
749         conf->barrier++;
750
751         /* Now wait for all pending IO to complete */
752         wait_event_lock_irq(conf->wait_barrier,
753                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
754                             conf->resync_lock, );
755
756         spin_unlock_irq(&conf->resync_lock);
757 }
758
759 static void lower_barrier(struct r1conf *conf)
760 {
761         unsigned long flags;
762         BUG_ON(conf->barrier <= 0);
763         spin_lock_irqsave(&conf->resync_lock, flags);
764         conf->barrier--;
765         spin_unlock_irqrestore(&conf->resync_lock, flags);
766         wake_up(&conf->wait_barrier);
767 }
768
769 static void wait_barrier(struct r1conf *conf)
770 {
771         spin_lock_irq(&conf->resync_lock);
772         if (conf->barrier) {
773                 conf->nr_waiting++;
774                 /* Wait for the barrier to drop.
775                  * However if there are already pending
776                  * requests (preventing the barrier from
777                  * rising completely), and the
778                  * pre-process bio queue isn't empty,
779                  * then don't wait, as we need to empty
780                  * that queue to get the nr_pending
781                  * count down.
782                  */
783                 wait_event_lock_irq(conf->wait_barrier,
784                                     !conf->barrier ||
785                                     (conf->nr_pending &&
786                                      current->bio_list &&
787                                      !bio_list_empty(current->bio_list)),
788                                     conf->resync_lock,
789                         );
790                 conf->nr_waiting--;
791         }
792         conf->nr_pending++;
793         spin_unlock_irq(&conf->resync_lock);
794 }
795
796 static void allow_barrier(struct r1conf *conf)
797 {
798         unsigned long flags;
799         spin_lock_irqsave(&conf->resync_lock, flags);
800         conf->nr_pending--;
801         spin_unlock_irqrestore(&conf->resync_lock, flags);
802         wake_up(&conf->wait_barrier);
803 }
804
805 static void freeze_array(struct r1conf *conf)
806 {
807         /* stop syncio and normal IO and wait for everything to
808          * go quite.
809          * We increment barrier and nr_waiting, and then
810          * wait until nr_pending match nr_queued+1
811          * This is called in the context of one normal IO request
812          * that has failed. Thus any sync request that might be pending
813          * will be blocked by nr_pending, and we need to wait for
814          * pending IO requests to complete or be queued for re-try.
815          * Thus the number queued (nr_queued) plus this request (1)
816          * must match the number of pending IOs (nr_pending) before
817          * we continue.
818          */
819         spin_lock_irq(&conf->resync_lock);
820         conf->barrier++;
821         conf->nr_waiting++;
822         wait_event_lock_irq(conf->wait_barrier,
823                             conf->nr_pending == conf->nr_queued+1,
824                             conf->resync_lock,
825                             flush_pending_writes(conf));
826         spin_unlock_irq(&conf->resync_lock);
827 }
828 static void unfreeze_array(struct r1conf *conf)
829 {
830         /* reverse the effect of the freeze */
831         spin_lock_irq(&conf->resync_lock);
832         conf->barrier--;
833         conf->nr_waiting--;
834         wake_up(&conf->wait_barrier);
835         spin_unlock_irq(&conf->resync_lock);
836 }
837
838
839 /* duplicate the data pages for behind I/O 
840  */
841 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
842 {
843         int i;
844         struct bio_vec *bvec;
845         struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
846                                         GFP_NOIO);
847         if (unlikely(!bvecs))
848                 return;
849
850         bio_for_each_segment(bvec, bio, i) {
851                 bvecs[i] = *bvec;
852                 bvecs[i].bv_page = alloc_page(GFP_NOIO);
853                 if (unlikely(!bvecs[i].bv_page))
854                         goto do_sync_io;
855                 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
856                        kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
857                 kunmap(bvecs[i].bv_page);
858                 kunmap(bvec->bv_page);
859         }
860         r1_bio->behind_bvecs = bvecs;
861         r1_bio->behind_page_count = bio->bi_vcnt;
862         set_bit(R1BIO_BehindIO, &r1_bio->state);
863         return;
864
865 do_sync_io:
866         for (i = 0; i < bio->bi_vcnt; i++)
867                 if (bvecs[i].bv_page)
868                         put_page(bvecs[i].bv_page);
869         kfree(bvecs);
870         pr_debug("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
871 }
872
873 static void make_request(struct mddev *mddev, struct bio * bio)
874 {
875         struct r1conf *conf = mddev->private;
876         struct mirror_info *mirror;
877         struct r1bio *r1_bio;
878         struct bio *read_bio;
879         int i, disks;
880         struct bitmap *bitmap;
881         unsigned long flags;
882         const int rw = bio_data_dir(bio);
883         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
884         const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
885         struct md_rdev *blocked_rdev;
886         int first_clone;
887         int sectors_handled;
888         int max_sectors;
889
890         /*
891          * Register the new request and wait if the reconstruction
892          * thread has put up a bar for new requests.
893          * Continue immediately if no resync is active currently.
894          */
895
896         md_write_start(mddev, bio); /* wait on superblock update early */
897
898         if (bio_data_dir(bio) == WRITE &&
899             bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
900             bio->bi_sector < mddev->suspend_hi) {
901                 /* As the suspend_* range is controlled by
902                  * userspace, we want an interruptible
903                  * wait.
904                  */
905                 DEFINE_WAIT(w);
906                 for (;;) {
907                         flush_signals(current);
908                         prepare_to_wait(&conf->wait_barrier,
909                                         &w, TASK_INTERRUPTIBLE);
910                         if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
911                             bio->bi_sector >= mddev->suspend_hi)
912                                 break;
913                         schedule();
914                 }
915                 finish_wait(&conf->wait_barrier, &w);
916         }
917
918         wait_barrier(conf);
919
920         bitmap = mddev->bitmap;
921
922         /*
923          * make_request() can abort the operation when READA is being
924          * used and no empty request is available.
925          *
926          */
927         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
928
929         r1_bio->master_bio = bio;
930         r1_bio->sectors = bio->bi_size >> 9;
931         r1_bio->state = 0;
932         r1_bio->mddev = mddev;
933         r1_bio->sector = bio->bi_sector;
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 r1_bio and no locking
939          * will be needed when requests complete.  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 rdisk;
950
951 read_again:
952                 rdisk = read_balance(conf, r1_bio, &max_sectors);
953
954                 if (rdisk < 0) {
955                         /* couldn't find anywhere to read from */
956                         raid_end_bio_io(r1_bio);
957                         return;
958                 }
959                 mirror = conf->mirrors + rdisk;
960
961                 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
962                     bitmap) {
963                         /* Reading from a write-mostly device must
964                          * take care not to over-take any writes
965                          * that are 'behind'
966                          */
967                         wait_event(bitmap->behind_wait,
968                                    atomic_read(&bitmap->behind_writes) == 0);
969                 }
970                 r1_bio->read_disk = rdisk;
971
972                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
973                 md_trim_bio(read_bio, r1_bio->sector - bio->bi_sector,
974                             max_sectors);
975
976                 r1_bio->bios[rdisk] = read_bio;
977
978                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
979                 read_bio->bi_bdev = mirror->rdev->bdev;
980                 read_bio->bi_end_io = raid1_end_read_request;
981                 read_bio->bi_rw = READ | do_sync;
982                 read_bio->bi_private = r1_bio;
983
984                 if (max_sectors < r1_bio->sectors) {
985                         /* could not read all from this device, so we will
986                          * need another r1_bio.
987                          */
988
989                         sectors_handled = (r1_bio->sector + max_sectors
990                                            - bio->bi_sector);
991                         r1_bio->sectors = max_sectors;
992                         spin_lock_irq(&conf->device_lock);
993                         if (bio->bi_phys_segments == 0)
994                                 bio->bi_phys_segments = 2;
995                         else
996                                 bio->bi_phys_segments++;
997                         spin_unlock_irq(&conf->device_lock);
998                         /* Cannot call generic_make_request directly
999                          * as that will be queued in __make_request
1000                          * and subsequent mempool_alloc might block waiting
1001                          * for it.  So hand bio over to raid1d.
1002                          */
1003                         reschedule_retry(r1_bio);
1004
1005                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1006
1007                         r1_bio->master_bio = bio;
1008                         r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1009                         r1_bio->state = 0;
1010                         r1_bio->mddev = mddev;
1011                         r1_bio->sector = bio->bi_sector + sectors_handled;
1012                         goto read_again;
1013                 } else
1014                         generic_make_request(read_bio);
1015                 return;
1016         }
1017
1018         /*
1019          * WRITE:
1020          */
1021         if (conf->pending_count >= max_queued_requests) {
1022                 md_wakeup_thread(mddev->thread);
1023                 wait_event(conf->wait_barrier,
1024                            conf->pending_count < max_queued_requests);
1025         }
1026         /* first select target devices under rcu_lock and
1027          * inc refcount on their rdev.  Record them by setting
1028          * bios[x] to bio
1029          * If there are known/acknowledged bad blocks on any device on
1030          * which we have seen a write error, we want to avoid writing those
1031          * blocks.
1032          * This potentially requires several writes to write around
1033          * the bad blocks.  Each set of writes gets it's own r1bio
1034          * with a set of bios attached.
1035          */
1036
1037         disks = conf->raid_disks * 2;
1038  retry_write:
1039         blocked_rdev = NULL;
1040         rcu_read_lock();
1041         max_sectors = r1_bio->sectors;
1042         for (i = 0;  i < disks; i++) {
1043                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1044                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1045                         atomic_inc(&rdev->nr_pending);
1046                         blocked_rdev = rdev;
1047                         break;
1048                 }
1049                 r1_bio->bios[i] = NULL;
1050                 if (!rdev || test_bit(Faulty, &rdev->flags)
1051                     || test_bit(Unmerged, &rdev->flags)) {
1052                         if (i < conf->raid_disks)
1053                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1054                         continue;
1055                 }
1056
1057                 atomic_inc(&rdev->nr_pending);
1058                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1059                         sector_t first_bad;
1060                         int bad_sectors;
1061                         int is_bad;
1062
1063                         is_bad = is_badblock(rdev, r1_bio->sector,
1064                                              max_sectors,
1065                                              &first_bad, &bad_sectors);
1066                         if (is_bad < 0) {
1067                                 /* mustn't write here until the bad block is
1068                                  * acknowledged*/
1069                                 set_bit(BlockedBadBlocks, &rdev->flags);
1070                                 blocked_rdev = rdev;
1071                                 break;
1072                         }
1073                         if (is_bad && first_bad <= r1_bio->sector) {
1074                                 /* Cannot write here at all */
1075                                 bad_sectors -= (r1_bio->sector - first_bad);
1076                                 if (bad_sectors < max_sectors)
1077                                         /* mustn't write more than bad_sectors
1078                                          * to other devices yet
1079                                          */
1080                                         max_sectors = bad_sectors;
1081                                 rdev_dec_pending(rdev, mddev);
1082                                 /* We don't set R1BIO_Degraded as that
1083                                  * only applies if the disk is
1084                                  * missing, so it might be re-added,
1085                                  * and we want to know to recover this
1086                                  * chunk.
1087                                  * In this case the device is here,
1088                                  * and the fact that this chunk is not
1089                                  * in-sync is recorded in the bad
1090                                  * block log
1091                                  */
1092                                 continue;
1093                         }
1094                         if (is_bad) {
1095                                 int good_sectors = first_bad - r1_bio->sector;
1096                                 if (good_sectors < max_sectors)
1097                                         max_sectors = good_sectors;
1098                         }
1099                 }
1100                 r1_bio->bios[i] = bio;
1101         }
1102         rcu_read_unlock();
1103
1104         if (unlikely(blocked_rdev)) {
1105                 /* Wait for this device to become unblocked */
1106                 int j;
1107
1108                 for (j = 0; j < i; j++)
1109                         if (r1_bio->bios[j])
1110                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1111                 r1_bio->state = 0;
1112                 allow_barrier(conf);
1113                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1114                 wait_barrier(conf);
1115                 goto retry_write;
1116         }
1117
1118         if (max_sectors < r1_bio->sectors) {
1119                 /* We are splitting this write into multiple parts, so
1120                  * we need to prepare for allocating another r1_bio.
1121                  */
1122                 r1_bio->sectors = max_sectors;
1123                 spin_lock_irq(&conf->device_lock);
1124                 if (bio->bi_phys_segments == 0)
1125                         bio->bi_phys_segments = 2;
1126                 else
1127                         bio->bi_phys_segments++;
1128                 spin_unlock_irq(&conf->device_lock);
1129         }
1130         sectors_handled = r1_bio->sector + max_sectors - bio->bi_sector;
1131
1132         atomic_set(&r1_bio->remaining, 1);
1133         atomic_set(&r1_bio->behind_remaining, 0);
1134
1135         first_clone = 1;
1136         for (i = 0; i < disks; i++) {
1137                 struct bio *mbio;
1138                 if (!r1_bio->bios[i])
1139                         continue;
1140
1141                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1142                 md_trim_bio(mbio, r1_bio->sector - bio->bi_sector, max_sectors);
1143
1144                 if (first_clone) {
1145                         /* do behind I/O ?
1146                          * Not if there are too many, or cannot
1147                          * allocate memory, or a reader on WriteMostly
1148                          * is waiting for behind writes to flush */
1149                         if (bitmap &&
1150                             (atomic_read(&bitmap->behind_writes)
1151                              < mddev->bitmap_info.max_write_behind) &&
1152                             !waitqueue_active(&bitmap->behind_wait))
1153                                 alloc_behind_pages(mbio, r1_bio);
1154
1155                         bitmap_startwrite(bitmap, r1_bio->sector,
1156                                           r1_bio->sectors,
1157                                           test_bit(R1BIO_BehindIO,
1158                                                    &r1_bio->state));
1159                         first_clone = 0;
1160                 }
1161                 if (r1_bio->behind_bvecs) {
1162                         struct bio_vec *bvec;
1163                         int j;
1164
1165                         /* Yes, I really want the '__' version so that
1166                          * we clear any unused pointer in the io_vec, rather
1167                          * than leave them unchanged.  This is important
1168                          * because when we come to free the pages, we won't
1169                          * know the original bi_idx, so we just free
1170                          * them all
1171                          */
1172                         __bio_for_each_segment(bvec, mbio, j, 0)
1173                                 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1174                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1175                                 atomic_inc(&r1_bio->behind_remaining);
1176                 }
1177
1178                 r1_bio->bios[i] = mbio;
1179
1180                 mbio->bi_sector = (r1_bio->sector +
1181                                    conf->mirrors[i].rdev->data_offset);
1182                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1183                 mbio->bi_end_io = raid1_end_write_request;
1184                 mbio->bi_rw = WRITE | do_flush_fua | do_sync;
1185                 mbio->bi_private = r1_bio;
1186
1187                 atomic_inc(&r1_bio->remaining);
1188                 spin_lock_irqsave(&conf->device_lock, flags);
1189                 bio_list_add(&conf->pending_bio_list, mbio);
1190                 conf->pending_count++;
1191                 spin_unlock_irqrestore(&conf->device_lock, flags);
1192                 if (!mddev_check_plugged(mddev))
1193                         md_wakeup_thread(mddev->thread);
1194         }
1195         /* Mustn't call r1_bio_write_done before this next test,
1196          * as it could result in the bio being freed.
1197          */
1198         if (sectors_handled < (bio->bi_size >> 9)) {
1199                 r1_bio_write_done(r1_bio);
1200                 /* We need another r1_bio.  It has already been counted
1201                  * in bio->bi_phys_segments
1202                  */
1203                 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1204                 r1_bio->master_bio = bio;
1205                 r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1206                 r1_bio->state = 0;
1207                 r1_bio->mddev = mddev;
1208                 r1_bio->sector = bio->bi_sector + sectors_handled;
1209                 goto retry_write;
1210         }
1211
1212         r1_bio_write_done(r1_bio);
1213
1214         /* In case raid1d snuck in to freeze_array */
1215         wake_up(&conf->wait_barrier);
1216 }
1217
1218 static void status(struct seq_file *seq, struct mddev *mddev)
1219 {
1220         struct r1conf *conf = mddev->private;
1221         int i;
1222
1223         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1224                    conf->raid_disks - mddev->degraded);
1225         rcu_read_lock();
1226         for (i = 0; i < conf->raid_disks; i++) {
1227                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1228                 seq_printf(seq, "%s",
1229                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1230         }
1231         rcu_read_unlock();
1232         seq_printf(seq, "]");
1233 }
1234
1235
1236 static void error(struct mddev *mddev, struct md_rdev *rdev)
1237 {
1238         char b[BDEVNAME_SIZE];
1239         struct r1conf *conf = mddev->private;
1240
1241         /*
1242          * If it is not operational, then we have already marked it as dead
1243          * else if it is the last working disks, ignore the error, let the
1244          * next level up know.
1245          * else mark the drive as failed
1246          */
1247         if (test_bit(In_sync, &rdev->flags)
1248             && (conf->raid_disks - mddev->degraded) == 1) {
1249                 /*
1250                  * Don't fail the drive, act as though we were just a
1251                  * normal single drive.
1252                  * However don't try a recovery from this drive as
1253                  * it is very likely to fail.
1254                  */
1255                 conf->recovery_disabled = mddev->recovery_disabled;
1256                 return;
1257         }
1258         set_bit(Blocked, &rdev->flags);
1259         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1260                 unsigned long flags;
1261                 spin_lock_irqsave(&conf->device_lock, flags);
1262                 mddev->degraded++;
1263                 set_bit(Faulty, &rdev->flags);
1264                 spin_unlock_irqrestore(&conf->device_lock, flags);
1265                 /*
1266                  * if recovery is running, make sure it aborts.
1267                  */
1268                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1269         } else
1270                 set_bit(Faulty, &rdev->flags);
1271         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1272         printk(KERN_ALERT
1273                "md/raid1:%s: Disk failure on %s, disabling device.\n"
1274                "md/raid1:%s: Operation continuing on %d devices.\n",
1275                mdname(mddev), bdevname(rdev->bdev, b),
1276                mdname(mddev), conf->raid_disks - mddev->degraded);
1277 }
1278
1279 static void print_conf(struct r1conf *conf)
1280 {
1281         int i;
1282
1283         printk(KERN_DEBUG "RAID1 conf printout:\n");
1284         if (!conf) {
1285                 printk(KERN_DEBUG "(!conf)\n");
1286                 return;
1287         }
1288         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1289                 conf->raid_disks);
1290
1291         rcu_read_lock();
1292         for (i = 0; i < conf->raid_disks; i++) {
1293                 char b[BDEVNAME_SIZE];
1294                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1295                 if (rdev)
1296                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1297                                i, !test_bit(In_sync, &rdev->flags),
1298                                !test_bit(Faulty, &rdev->flags),
1299                                bdevname(rdev->bdev,b));
1300         }
1301         rcu_read_unlock();
1302 }
1303
1304 static void close_sync(struct r1conf *conf)
1305 {
1306         wait_barrier(conf);
1307         allow_barrier(conf);
1308
1309         mempool_destroy(conf->r1buf_pool);
1310         conf->r1buf_pool = NULL;
1311 }
1312
1313 static int raid1_spare_active(struct mddev *mddev)
1314 {
1315         int i;
1316         struct r1conf *conf = mddev->private;
1317         int count = 0;
1318         unsigned long flags;
1319
1320         /*
1321          * Find all failed disks within the RAID1 configuration 
1322          * and mark them readable.
1323          * Called under mddev lock, so rcu protection not needed.
1324          */
1325         for (i = 0; i < conf->raid_disks; i++) {
1326                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1327                 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1328                 if (repl
1329                     && repl->recovery_offset == MaxSector
1330                     && !test_bit(Faulty, &repl->flags)
1331                     && !test_and_set_bit(In_sync, &repl->flags)) {
1332                         /* replacement has just become active */
1333                         if (!rdev ||
1334                             !test_and_clear_bit(In_sync, &rdev->flags))
1335                                 count++;
1336                         if (rdev) {
1337                                 /* Replaced device not technically
1338                                  * faulty, but we need to be sure
1339                                  * it gets removed and never re-added
1340                                  */
1341                                 set_bit(Faulty, &rdev->flags);
1342                                 sysfs_notify_dirent_safe(
1343                                         rdev->sysfs_state);
1344                         }
1345                 }
1346                 if (rdev
1347                     && !test_bit(Faulty, &rdev->flags)
1348                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1349                         count++;
1350                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1351                 }
1352         }
1353         spin_lock_irqsave(&conf->device_lock, flags);
1354         mddev->degraded -= count;
1355         spin_unlock_irqrestore(&conf->device_lock, flags);
1356
1357         print_conf(conf);
1358         return count;
1359 }
1360
1361
1362 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1363 {
1364         struct r1conf *conf = mddev->private;
1365         int err = -EEXIST;
1366         int mirror = 0;
1367         struct mirror_info *p;
1368         int first = 0;
1369         int last = conf->raid_disks - 1;
1370         struct request_queue *q = bdev_get_queue(rdev->bdev);
1371
1372         if (mddev->recovery_disabled == conf->recovery_disabled)
1373                 return -EBUSY;
1374
1375         if (rdev->raid_disk >= 0)
1376                 first = last = rdev->raid_disk;
1377
1378         if (q->merge_bvec_fn) {
1379                 set_bit(Unmerged, &rdev->flags);
1380                 mddev->merge_check_needed = 1;
1381         }
1382
1383         for (mirror = first; mirror <= last; mirror++) {
1384                 p = conf->mirrors+mirror;
1385                 if (!p->rdev) {
1386
1387                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1388                                           rdev->data_offset << 9);
1389
1390                         p->head_position = 0;
1391                         rdev->raid_disk = mirror;
1392                         err = 0;
1393                         /* As all devices are equivalent, we don't need a full recovery
1394                          * if this was recently any drive of the array
1395                          */
1396                         if (rdev->saved_raid_disk < 0)
1397                                 conf->fullsync = 1;
1398                         rcu_assign_pointer(p->rdev, rdev);
1399                         break;
1400                 }
1401                 if (test_bit(WantReplacement, &p->rdev->flags) &&
1402                     p[conf->raid_disks].rdev == NULL) {
1403                         /* Add this device as a replacement */
1404                         clear_bit(In_sync, &rdev->flags);
1405                         set_bit(Replacement, &rdev->flags);
1406                         rdev->raid_disk = mirror;
1407                         err = 0;
1408                         conf->fullsync = 1;
1409                         rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1410                         break;
1411                 }
1412         }
1413         if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1414                 /* Some requests might not have seen this new
1415                  * merge_bvec_fn.  We must wait for them to complete
1416                  * before merging the device fully.
1417                  * First we make sure any code which has tested
1418                  * our function has submitted the request, then
1419                  * we wait for all outstanding requests to complete.
1420                  */
1421                 synchronize_sched();
1422                 raise_barrier(conf);
1423                 lower_barrier(conf);
1424                 clear_bit(Unmerged, &rdev->flags);
1425         }
1426         md_integrity_add_rdev(rdev, mddev);
1427         print_conf(conf);
1428         return err;
1429 }
1430
1431 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1432 {
1433         struct r1conf *conf = mddev->private;
1434         int err = 0;
1435         int number = rdev->raid_disk;
1436         struct mirror_info *p = conf->mirrors+ number;
1437
1438         if (rdev != p->rdev)
1439                 p = conf->mirrors + conf->raid_disks + number;
1440
1441         print_conf(conf);
1442         if (rdev == p->rdev) {
1443                 if (test_bit(In_sync, &rdev->flags) ||
1444                     atomic_read(&rdev->nr_pending)) {
1445                         err = -EBUSY;
1446                         goto abort;
1447                 }
1448                 /* Only remove non-faulty devices if recovery
1449                  * is not possible.
1450                  */
1451                 if (!test_bit(Faulty, &rdev->flags) &&
1452                     mddev->recovery_disabled != conf->recovery_disabled &&
1453                     mddev->degraded < conf->raid_disks) {
1454                         err = -EBUSY;
1455                         goto abort;
1456                 }
1457                 p->rdev = NULL;
1458                 synchronize_rcu();
1459                 if (atomic_read(&rdev->nr_pending)) {
1460                         /* lost the race, try later */
1461                         err = -EBUSY;
1462                         p->rdev = rdev;
1463                         goto abort;
1464                 } else if (conf->mirrors[conf->raid_disks + number].rdev) {
1465                         /* We just removed a device that is being replaced.
1466                          * Move down the replacement.  We drain all IO before
1467                          * doing this to avoid confusion.
1468                          */
1469                         struct md_rdev *repl =
1470                                 conf->mirrors[conf->raid_disks + number].rdev;
1471                         raise_barrier(conf);
1472                         clear_bit(Replacement, &repl->flags);
1473                         p->rdev = repl;
1474                         conf->mirrors[conf->raid_disks + number].rdev = NULL;
1475                         lower_barrier(conf);
1476                         clear_bit(WantReplacement, &rdev->flags);
1477                 } else
1478                         clear_bit(WantReplacement, &rdev->flags);
1479                 err = md_integrity_register(mddev);
1480         }
1481 abort:
1482
1483         print_conf(conf);
1484         return err;
1485 }
1486
1487
1488 static void end_sync_read(struct bio *bio, int error)
1489 {
1490         struct r1bio *r1_bio = bio->bi_private;
1491
1492         update_head_pos(r1_bio->read_disk, r1_bio);
1493
1494         /*
1495          * we have read a block, now it needs to be re-written,
1496          * or re-read if the read failed.
1497          * We don't do much here, just schedule handling by raid1d
1498          */
1499         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1500                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1501
1502         if (atomic_dec_and_test(&r1_bio->remaining))
1503                 reschedule_retry(r1_bio);
1504 }
1505
1506 static void end_sync_write(struct bio *bio, int error)
1507 {
1508         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1509         struct r1bio *r1_bio = bio->bi_private;
1510         struct mddev *mddev = r1_bio->mddev;
1511         struct r1conf *conf = mddev->private;
1512         int mirror=0;
1513         sector_t first_bad;
1514         int bad_sectors;
1515
1516         mirror = find_bio_disk(r1_bio, bio);
1517
1518         if (!uptodate) {
1519                 sector_t sync_blocks = 0;
1520                 sector_t s = r1_bio->sector;
1521                 long sectors_to_go = r1_bio->sectors;
1522                 /* make sure these bits doesn't get cleared. */
1523                 do {
1524                         bitmap_end_sync(mddev->bitmap, s,
1525                                         &sync_blocks, 1);
1526                         s += sync_blocks;
1527                         sectors_to_go -= sync_blocks;
1528                 } while (sectors_to_go > 0);
1529                 set_bit(WriteErrorSeen,
1530                         &conf->mirrors[mirror].rdev->flags);
1531                 if (!test_and_set_bit(WantReplacement,
1532                                       &conf->mirrors[mirror].rdev->flags))
1533                         set_bit(MD_RECOVERY_NEEDED, &
1534                                 mddev->recovery);
1535                 set_bit(R1BIO_WriteError, &r1_bio->state);
1536         } else if (is_badblock(conf->mirrors[mirror].rdev,
1537                                r1_bio->sector,
1538                                r1_bio->sectors,
1539                                &first_bad, &bad_sectors) &&
1540                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1541                                 r1_bio->sector,
1542                                 r1_bio->sectors,
1543                                 &first_bad, &bad_sectors)
1544                 )
1545                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1546
1547         if (atomic_dec_and_test(&r1_bio->remaining)) {
1548                 int s = r1_bio->sectors;
1549                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1550                     test_bit(R1BIO_WriteError, &r1_bio->state))
1551                         reschedule_retry(r1_bio);
1552                 else {
1553                         put_buf(r1_bio);
1554                         md_done_sync(mddev, s, uptodate);
1555                 }
1556         }
1557 }
1558
1559 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1560                             int sectors, struct page *page, int rw)
1561 {
1562         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1563                 /* success */
1564                 return 1;
1565         if (rw == WRITE) {
1566                 set_bit(WriteErrorSeen, &rdev->flags);
1567                 if (!test_and_set_bit(WantReplacement,
1568                                       &rdev->flags))
1569                         set_bit(MD_RECOVERY_NEEDED, &
1570                                 rdev->mddev->recovery);
1571         }
1572         /* need to record an error - either for the block or the device */
1573         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1574                 md_error(rdev->mddev, rdev);
1575         return 0;
1576 }
1577
1578 static int fix_sync_read_error(struct r1bio *r1_bio)
1579 {
1580         /* Try some synchronous reads of other devices to get
1581          * good data, much like with normal read errors.  Only
1582          * read into the pages we already have so we don't
1583          * need to re-issue the read request.
1584          * We don't need to freeze the array, because being in an
1585          * active sync request, there is no normal IO, and
1586          * no overlapping syncs.
1587          * We don't need to check is_badblock() again as we
1588          * made sure that anything with a bad block in range
1589          * will have bi_end_io clear.
1590          */
1591         struct mddev *mddev = r1_bio->mddev;
1592         struct r1conf *conf = mddev->private;
1593         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1594         sector_t sect = r1_bio->sector;
1595         int sectors = r1_bio->sectors;
1596         int idx = 0;
1597
1598         while(sectors) {
1599                 int s = sectors;
1600                 int d = r1_bio->read_disk;
1601                 int success = 0;
1602                 struct md_rdev *rdev;
1603                 int start;
1604
1605                 if (s > (PAGE_SIZE>>9))
1606                         s = PAGE_SIZE >> 9;
1607                 do {
1608                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1609                                 /* No rcu protection needed here devices
1610                                  * can only be removed when no resync is
1611                                  * active, and resync is currently active
1612                                  */
1613                                 rdev = conf->mirrors[d].rdev;
1614                                 if (sync_page_io(rdev, sect, s<<9,
1615                                                  bio->bi_io_vec[idx].bv_page,
1616                                                  READ, false)) {
1617                                         success = 1;
1618                                         break;
1619                                 }
1620                         }
1621                         d++;
1622                         if (d == conf->raid_disks * 2)
1623                                 d = 0;
1624                 } while (!success && d != r1_bio->read_disk);
1625
1626                 if (!success) {
1627                         char b[BDEVNAME_SIZE];
1628                         int abort = 0;
1629                         /* Cannot read from anywhere, this block is lost.
1630                          * Record a bad block on each device.  If that doesn't
1631                          * work just disable and interrupt the recovery.
1632                          * Don't fail devices as that won't really help.
1633                          */
1634                         printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1635                                " for block %llu\n",
1636                                mdname(mddev),
1637                                bdevname(bio->bi_bdev, b),
1638                                (unsigned long long)r1_bio->sector);
1639                         for (d = 0; d < conf->raid_disks * 2; d++) {
1640                                 rdev = conf->mirrors[d].rdev;
1641                                 if (!rdev || test_bit(Faulty, &rdev->flags))
1642                                         continue;
1643                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
1644                                         abort = 1;
1645                         }
1646                         if (abort) {
1647                                 conf->recovery_disabled =
1648                                         mddev->recovery_disabled;
1649                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1650                                 md_done_sync(mddev, r1_bio->sectors, 0);
1651                                 put_buf(r1_bio);
1652                                 return 0;
1653                         }
1654                         /* Try next page */
1655                         sectors -= s;
1656                         sect += s;
1657                         idx++;
1658                         continue;
1659                 }
1660
1661                 start = d;
1662                 /* write it back and re-read */
1663                 while (d != r1_bio->read_disk) {
1664                         if (d == 0)
1665                                 d = conf->raid_disks * 2;
1666                         d--;
1667                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1668                                 continue;
1669                         rdev = conf->mirrors[d].rdev;
1670                         if (r1_sync_page_io(rdev, sect, s,
1671                                             bio->bi_io_vec[idx].bv_page,
1672                                             WRITE) == 0) {
1673                                 r1_bio->bios[d]->bi_end_io = NULL;
1674                                 rdev_dec_pending(rdev, mddev);
1675                         }
1676                 }
1677                 d = start;
1678                 while (d != r1_bio->read_disk) {
1679                         if (d == 0)
1680                                 d = conf->raid_disks * 2;
1681                         d--;
1682                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1683                                 continue;
1684                         rdev = conf->mirrors[d].rdev;
1685                         if (r1_sync_page_io(rdev, sect, s,
1686                                             bio->bi_io_vec[idx].bv_page,
1687                                             READ) != 0)
1688                                 atomic_add(s, &rdev->corrected_errors);
1689                 }
1690                 sectors -= s;
1691                 sect += s;
1692                 idx ++;
1693         }
1694         set_bit(R1BIO_Uptodate, &r1_bio->state);
1695         set_bit(BIO_UPTODATE, &bio->bi_flags);
1696         return 1;
1697 }
1698
1699 static int process_checks(struct r1bio *r1_bio)
1700 {
1701         /* We have read all readable devices.  If we haven't
1702          * got the block, then there is no hope left.
1703          * If we have, then we want to do a comparison
1704          * and skip the write if everything is the same.
1705          * If any blocks failed to read, then we need to
1706          * attempt an over-write
1707          */
1708         struct mddev *mddev = r1_bio->mddev;
1709         struct r1conf *conf = mddev->private;
1710         int primary;
1711         int i;
1712         int vcnt;
1713
1714         for (primary = 0; primary < conf->raid_disks * 2; primary++)
1715                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1716                     test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1717                         r1_bio->bios[primary]->bi_end_io = NULL;
1718                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1719                         break;
1720                 }
1721         r1_bio->read_disk = primary;
1722         vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
1723         for (i = 0; i < conf->raid_disks * 2; i++) {
1724                 int j;
1725                 struct bio *pbio = r1_bio->bios[primary];
1726                 struct bio *sbio = r1_bio->bios[i];
1727                 int size;
1728
1729                 if (r1_bio->bios[i]->bi_end_io != end_sync_read)
1730                         continue;
1731
1732                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1733                         for (j = vcnt; j-- ; ) {
1734                                 struct page *p, *s;
1735                                 p = pbio->bi_io_vec[j].bv_page;
1736                                 s = sbio->bi_io_vec[j].bv_page;
1737                                 if (memcmp(page_address(p),
1738                                            page_address(s),
1739                                            sbio->bi_io_vec[j].bv_len))
1740                                         break;
1741                         }
1742                 } else
1743                         j = 0;
1744                 if (j >= 0)
1745                         mddev->resync_mismatches += r1_bio->sectors;
1746                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1747                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1748                         /* No need to write to this device. */
1749                         sbio->bi_end_io = NULL;
1750                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1751                         continue;
1752                 }
1753                 /* fixup the bio for reuse */
1754                 sbio->bi_vcnt = vcnt;
1755                 sbio->bi_size = r1_bio->sectors << 9;
1756                 sbio->bi_idx = 0;
1757                 sbio->bi_phys_segments = 0;
1758                 sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1759                 sbio->bi_flags |= 1 << BIO_UPTODATE;
1760                 sbio->bi_next = NULL;
1761                 sbio->bi_sector = r1_bio->sector +
1762                         conf->mirrors[i].rdev->data_offset;
1763                 sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1764                 size = sbio->bi_size;
1765                 for (j = 0; j < vcnt ; j++) {
1766                         struct bio_vec *bi;
1767                         bi = &sbio->bi_io_vec[j];
1768                         bi->bv_offset = 0;
1769                         if (size > PAGE_SIZE)
1770                                 bi->bv_len = PAGE_SIZE;
1771                         else
1772                                 bi->bv_len = size;
1773                         size -= PAGE_SIZE;
1774                         memcpy(page_address(bi->bv_page),
1775                                page_address(pbio->bi_io_vec[j].bv_page),
1776                                PAGE_SIZE);
1777                 }
1778         }
1779         return 0;
1780 }
1781
1782 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
1783 {
1784         struct r1conf *conf = mddev->private;
1785         int i;
1786         int disks = conf->raid_disks * 2;
1787         struct bio *bio, *wbio;
1788
1789         bio = r1_bio->bios[r1_bio->read_disk];
1790
1791         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1792                 /* ouch - failed to read all of that. */
1793                 if (!fix_sync_read_error(r1_bio))
1794                         return;
1795
1796         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1797                 if (process_checks(r1_bio) < 0)
1798                         return;
1799         /*
1800          * schedule writes
1801          */
1802         atomic_set(&r1_bio->remaining, 1);
1803         for (i = 0; i < disks ; i++) {
1804                 wbio = r1_bio->bios[i];
1805                 if (wbio->bi_end_io == NULL ||
1806                     (wbio->bi_end_io == end_sync_read &&
1807                      (i == r1_bio->read_disk ||
1808                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1809                         continue;
1810
1811                 wbio->bi_rw = WRITE;
1812                 wbio->bi_end_io = end_sync_write;
1813                 atomic_inc(&r1_bio->remaining);
1814                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1815
1816                 generic_make_request(wbio);
1817         }
1818
1819         if (atomic_dec_and_test(&r1_bio->remaining)) {
1820                 /* if we're here, all write(s) have completed, so clean up */
1821                 int s = r1_bio->sectors;
1822                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1823                     test_bit(R1BIO_WriteError, &r1_bio->state))
1824                         reschedule_retry(r1_bio);
1825                 else {
1826                         put_buf(r1_bio);
1827                         md_done_sync(mddev, s, 1);
1828                 }
1829         }
1830 }
1831
1832 /*
1833  * This is a kernel thread which:
1834  *
1835  *      1.      Retries failed read operations on working mirrors.
1836  *      2.      Updates the raid superblock when problems encounter.
1837  *      3.      Performs writes following reads for array synchronising.
1838  */
1839
1840 static void fix_read_error(struct r1conf *conf, int read_disk,
1841                            sector_t sect, int sectors)
1842 {
1843         struct mddev *mddev = conf->mddev;
1844         while(sectors) {
1845                 int s = sectors;
1846                 int d = read_disk;
1847                 int success = 0;
1848                 int start;
1849                 struct md_rdev *rdev;
1850
1851                 if (s > (PAGE_SIZE>>9))
1852                         s = PAGE_SIZE >> 9;
1853
1854                 do {
1855                         /* Note: no rcu protection needed here
1856                          * as this is synchronous in the raid1d thread
1857                          * which is the thread that might remove
1858                          * a device.  If raid1d ever becomes multi-threaded....
1859                          */
1860                         sector_t first_bad;
1861                         int bad_sectors;
1862
1863                         rdev = conf->mirrors[d].rdev;
1864                         if (rdev &&
1865                             (test_bit(In_sync, &rdev->flags) ||
1866                              (!test_bit(Faulty, &rdev->flags) &&
1867                               rdev->recovery_offset >= sect + s)) &&
1868                             is_badblock(rdev, sect, s,
1869                                         &first_bad, &bad_sectors) == 0 &&
1870                             sync_page_io(rdev, sect, s<<9,
1871                                          conf->tmppage, READ, false))
1872                                 success = 1;
1873                         else {
1874                                 d++;
1875                                 if (d == conf->raid_disks * 2)
1876                                         d = 0;
1877                         }
1878                 } while (!success && d != read_disk);
1879
1880                 if (!success) {
1881                         /* Cannot read from anywhere - mark it bad */
1882                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
1883                         if (!rdev_set_badblocks(rdev, sect, s, 0))
1884                                 md_error(mddev, rdev);
1885                         break;
1886                 }
1887                 /* write it back and re-read */
1888                 start = d;
1889                 while (d != read_disk) {
1890                         if (d==0)
1891                                 d = conf->raid_disks * 2;
1892                         d--;
1893                         rdev = conf->mirrors[d].rdev;
1894                         if (rdev &&
1895                             test_bit(In_sync, &rdev->flags))
1896                                 r1_sync_page_io(rdev, sect, s,
1897                                                 conf->tmppage, WRITE);
1898                 }
1899                 d = start;
1900                 while (d != read_disk) {
1901                         char b[BDEVNAME_SIZE];
1902                         if (d==0)
1903                                 d = conf->raid_disks * 2;
1904                         d--;
1905                         rdev = conf->mirrors[d].rdev;
1906                         if (rdev &&
1907                             test_bit(In_sync, &rdev->flags)) {
1908                                 if (r1_sync_page_io(rdev, sect, s,
1909                                                     conf->tmppage, READ)) {
1910                                         atomic_add(s, &rdev->corrected_errors);
1911                                         printk(KERN_INFO
1912                                                "md/raid1:%s: read error corrected "
1913                                                "(%d sectors at %llu on %s)\n",
1914                                                mdname(mddev), s,
1915                                                (unsigned long long)(sect +
1916                                                    rdev->data_offset),
1917                                                bdevname(rdev->bdev, b));
1918                                 }
1919                         }
1920                 }
1921                 sectors -= s;
1922                 sect += s;
1923         }
1924 }
1925
1926 static void bi_complete(struct bio *bio, int error)
1927 {
1928         complete((struct completion *)bio->bi_private);
1929 }
1930
1931 static int submit_bio_wait(int rw, struct bio *bio)
1932 {
1933         struct completion event;
1934         rw |= REQ_SYNC;
1935
1936         init_completion(&event);
1937         bio->bi_private = &event;
1938         bio->bi_end_io = bi_complete;
1939         submit_bio(rw, bio);
1940         wait_for_completion(&event);
1941
1942         return test_bit(BIO_UPTODATE, &bio->bi_flags);
1943 }
1944
1945 static int narrow_write_error(struct r1bio *r1_bio, int i)
1946 {
1947         struct mddev *mddev = r1_bio->mddev;
1948         struct r1conf *conf = mddev->private;
1949         struct md_rdev *rdev = conf->mirrors[i].rdev;
1950         int vcnt, idx;
1951         struct bio_vec *vec;
1952
1953         /* bio has the data to be written to device 'i' where
1954          * we just recently had a write error.
1955          * We repeatedly clone the bio and trim down to one block,
1956          * then try the write.  Where the write fails we record
1957          * a bad block.
1958          * It is conceivable that the bio doesn't exactly align with
1959          * blocks.  We must handle this somehow.
1960          *
1961          * We currently own a reference on the rdev.
1962          */
1963
1964         int block_sectors;
1965         sector_t sector;
1966         int sectors;
1967         int sect_to_write = r1_bio->sectors;
1968         int ok = 1;
1969
1970         if (rdev->badblocks.shift < 0)
1971                 return 0;
1972
1973         block_sectors = 1 << rdev->badblocks.shift;
1974         sector = r1_bio->sector;
1975         sectors = ((sector + block_sectors)
1976                    & ~(sector_t)(block_sectors - 1))
1977                 - sector;
1978
1979         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
1980                 vcnt = r1_bio->behind_page_count;
1981                 vec = r1_bio->behind_bvecs;
1982                 idx = 0;
1983                 while (vec[idx].bv_page == NULL)
1984                         idx++;
1985         } else {
1986                 vcnt = r1_bio->master_bio->bi_vcnt;
1987                 vec = r1_bio->master_bio->bi_io_vec;
1988                 idx = r1_bio->master_bio->bi_idx;
1989         }
1990         while (sect_to_write) {
1991                 struct bio *wbio;
1992                 if (sectors > sect_to_write)
1993                         sectors = sect_to_write;
1994                 /* Write at 'sector' for 'sectors'*/
1995
1996                 wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
1997                 memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
1998                 wbio->bi_sector = r1_bio->sector;
1999                 wbio->bi_rw = WRITE;
2000                 wbio->bi_vcnt = vcnt;
2001                 wbio->bi_size = r1_bio->sectors << 9;
2002                 wbio->bi_idx = idx;
2003
2004                 md_trim_bio(wbio, sector - r1_bio->sector, sectors);
2005                 wbio->bi_sector += rdev->data_offset;
2006                 wbio->bi_bdev = rdev->bdev;
2007                 if (submit_bio_wait(WRITE, wbio) == 0)
2008                         /* failure! */
2009                         ok = rdev_set_badblocks(rdev, sector,
2010                                                 sectors, 0)
2011                                 && ok;
2012
2013                 bio_put(wbio);
2014                 sect_to_write -= sectors;
2015                 sector += sectors;
2016                 sectors = block_sectors;
2017         }
2018         return ok;
2019 }
2020
2021 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2022 {
2023         int m;
2024         int s = r1_bio->sectors;
2025         for (m = 0; m < conf->raid_disks * 2 ; m++) {
2026                 struct md_rdev *rdev = conf->mirrors[m].rdev;
2027                 struct bio *bio = r1_bio->bios[m];
2028                 if (bio->bi_end_io == NULL)
2029                         continue;
2030                 if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2031                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2032                         rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2033                 }
2034                 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2035                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
2036                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2037                                 md_error(conf->mddev, rdev);
2038                 }
2039         }
2040         put_buf(r1_bio);
2041         md_done_sync(conf->mddev, s, 1);
2042 }
2043
2044 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2045 {
2046         int m;
2047         for (m = 0; m < conf->raid_disks * 2 ; m++)
2048                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2049                         struct md_rdev *rdev = conf->mirrors[m].rdev;
2050                         rdev_clear_badblocks(rdev,
2051                                              r1_bio->sector,
2052                                              r1_bio->sectors, 0);
2053                         rdev_dec_pending(rdev, conf->mddev);
2054                 } else if (r1_bio->bios[m] != NULL) {
2055                         /* This drive got a write error.  We need to
2056                          * narrow down and record precise write
2057                          * errors.
2058                          */
2059                         if (!narrow_write_error(r1_bio, m)) {
2060                                 md_error(conf->mddev,
2061                                          conf->mirrors[m].rdev);
2062                                 /* an I/O failed, we can't clear the bitmap */
2063                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2064                         }
2065                         rdev_dec_pending(conf->mirrors[m].rdev,
2066                                          conf->mddev);
2067                 }
2068         if (test_bit(R1BIO_WriteError, &r1_bio->state))
2069                 close_write(r1_bio);
2070         raid_end_bio_io(r1_bio);
2071 }
2072
2073 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2074 {
2075         int disk;
2076         int max_sectors;
2077         struct mddev *mddev = conf->mddev;
2078         struct bio *bio;
2079         char b[BDEVNAME_SIZE];
2080         struct md_rdev *rdev;
2081
2082         clear_bit(R1BIO_ReadError, &r1_bio->state);
2083         /* we got a read error. Maybe the drive is bad.  Maybe just
2084          * the block and we can fix it.
2085          * We freeze all other IO, and try reading the block from
2086          * other devices.  When we find one, we re-write
2087          * and check it that fixes the read error.
2088          * This is all done synchronously while the array is
2089          * frozen
2090          */
2091         if (mddev->ro == 0) {
2092                 freeze_array(conf);
2093                 fix_read_error(conf, r1_bio->read_disk,
2094                                r1_bio->sector, r1_bio->sectors);
2095                 unfreeze_array(conf);
2096         } else
2097                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2098
2099         bio = r1_bio->bios[r1_bio->read_disk];
2100         bdevname(bio->bi_bdev, b);
2101 read_more:
2102         disk = read_balance(conf, r1_bio, &max_sectors);
2103         if (disk == -1) {
2104                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2105                        " read error for block %llu\n",
2106                        mdname(mddev), b, (unsigned long long)r1_bio->sector);
2107                 raid_end_bio_io(r1_bio);
2108         } else {
2109                 const unsigned long do_sync
2110                         = r1_bio->master_bio->bi_rw & REQ_SYNC;
2111                 if (bio) {
2112                         r1_bio->bios[r1_bio->read_disk] =
2113                                 mddev->ro ? IO_BLOCKED : NULL;
2114                         bio_put(bio);
2115                 }
2116                 r1_bio->read_disk = disk;
2117                 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2118                 md_trim_bio(bio, r1_bio->sector - bio->bi_sector, max_sectors);
2119                 r1_bio->bios[r1_bio->read_disk] = bio;
2120                 rdev = conf->mirrors[disk].rdev;
2121                 printk_ratelimited(KERN_ERR
2122                                    "md/raid1:%s: redirecting sector %llu"
2123                                    " to other mirror: %s\n",
2124                                    mdname(mddev),
2125                                    (unsigned long long)r1_bio->sector,
2126                                    bdevname(rdev->bdev, b));
2127                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
2128                 bio->bi_bdev = rdev->bdev;
2129                 bio->bi_end_io = raid1_end_read_request;
2130                 bio->bi_rw = READ | do_sync;
2131                 bio->bi_private = r1_bio;
2132                 if (max_sectors < r1_bio->sectors) {
2133                         /* Drat - have to split this up more */
2134                         struct bio *mbio = r1_bio->master_bio;
2135                         int sectors_handled = (r1_bio->sector + max_sectors
2136                                                - mbio->bi_sector);
2137                         r1_bio->sectors = max_sectors;
2138                         spin_lock_irq(&conf->device_lock);
2139                         if (mbio->bi_phys_segments == 0)
2140                                 mbio->bi_phys_segments = 2;
2141                         else
2142                                 mbio->bi_phys_segments++;
2143                         spin_unlock_irq(&conf->device_lock);
2144                         generic_make_request(bio);
2145                         bio = NULL;
2146
2147                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2148
2149                         r1_bio->master_bio = mbio;
2150                         r1_bio->sectors = (mbio->bi_size >> 9)
2151                                           - sectors_handled;
2152                         r1_bio->state = 0;
2153                         set_bit(R1BIO_ReadError, &r1_bio->state);
2154                         r1_bio->mddev = mddev;
2155                         r1_bio->sector = mbio->bi_sector + sectors_handled;
2156
2157                         goto read_more;
2158                 } else
2159                         generic_make_request(bio);
2160         }
2161 }
2162
2163 static void raid1d(struct mddev *mddev)
2164 {
2165         struct r1bio *r1_bio;
2166         unsigned long flags;
2167         struct r1conf *conf = mddev->private;
2168         struct list_head *head = &conf->retry_list;
2169         struct blk_plug plug;
2170
2171         md_check_recovery(mddev);
2172
2173         blk_start_plug(&plug);
2174         for (;;) {
2175
2176                 flush_pending_writes(conf);
2177
2178                 spin_lock_irqsave(&conf->device_lock, flags);
2179                 if (list_empty(head)) {
2180                         spin_unlock_irqrestore(&conf->device_lock, flags);
2181                         break;
2182                 }
2183                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2184                 list_del(head->prev);
2185                 conf->nr_queued--;
2186                 spin_unlock_irqrestore(&conf->device_lock, flags);
2187
2188                 mddev = r1_bio->mddev;
2189                 conf = mddev->private;
2190                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2191                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2192                             test_bit(R1BIO_WriteError, &r1_bio->state))
2193                                 handle_sync_write_finished(conf, r1_bio);
2194                         else
2195                                 sync_request_write(mddev, r1_bio);
2196                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2197                            test_bit(R1BIO_WriteError, &r1_bio->state))
2198                         handle_write_finished(conf, r1_bio);
2199                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2200                         handle_read_error(conf, r1_bio);
2201                 else
2202                         /* just a partial read to be scheduled from separate
2203                          * context
2204                          */
2205                         generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2206
2207                 cond_resched();
2208                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2209                         md_check_recovery(mddev);
2210         }
2211         blk_finish_plug(&plug);
2212 }
2213
2214
2215 static int init_resync(struct r1conf *conf)
2216 {
2217         int buffs;
2218
2219         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2220         BUG_ON(conf->r1buf_pool);
2221         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2222                                           conf->poolinfo);
2223         if (!conf->r1buf_pool)
2224                 return -ENOMEM;
2225         conf->next_resync = 0;
2226         return 0;
2227 }
2228
2229 /*
2230  * perform a "sync" on one "block"
2231  *
2232  * We need to make sure that no normal I/O request - particularly write
2233  * requests - conflict with active sync requests.
2234  *
2235  * This is achieved by tracking pending requests and a 'barrier' concept
2236  * that can be installed to exclude normal IO requests.
2237  */
2238
2239 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2240 {
2241         struct r1conf *conf = mddev->private;
2242         struct r1bio *r1_bio;
2243         struct bio *bio;
2244         sector_t max_sector, nr_sectors;
2245         int disk = -1;
2246         int i;
2247         int wonly = -1;
2248         int write_targets = 0, read_targets = 0;
2249         sector_t sync_blocks;
2250         int still_degraded = 0;
2251         int good_sectors = RESYNC_SECTORS;
2252         int min_bad = 0; /* number of sectors that are bad in all devices */
2253
2254         if (!conf->r1buf_pool)
2255                 if (init_resync(conf))
2256                         return 0;
2257
2258         max_sector = mddev->dev_sectors;
2259         if (sector_nr >= max_sector) {
2260                 /* If we aborted, we need to abort the
2261                  * sync on the 'current' bitmap chunk (there will
2262                  * only be one in raid1 resync.
2263                  * We can find the current addess in mddev->curr_resync
2264                  */
2265                 if (mddev->curr_resync < max_sector) /* aborted */
2266                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2267                                                 &sync_blocks, 1);
2268                 else /* completed sync */
2269                         conf->fullsync = 0;
2270
2271                 bitmap_close_sync(mddev->bitmap);
2272                 close_sync(conf);
2273                 return 0;
2274         }
2275
2276         if (mddev->bitmap == NULL &&
2277             mddev->recovery_cp == MaxSector &&
2278             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2279             conf->fullsync == 0) {
2280                 *skipped = 1;
2281                 return max_sector - sector_nr;
2282         }
2283         /* before building a request, check if we can skip these blocks..
2284          * This call the bitmap_start_sync doesn't actually record anything
2285          */
2286         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2287             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2288                 /* We can skip this block, and probably several more */
2289                 *skipped = 1;
2290                 return sync_blocks;
2291         }
2292         /*
2293          * If there is non-resync activity waiting for a turn,
2294          * and resync is going fast enough,
2295          * then let it though before starting on this new sync request.
2296          */
2297         if (!go_faster && conf->nr_waiting)
2298                 msleep_interruptible(1000);
2299
2300         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2301         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2302         raise_barrier(conf);
2303
2304         conf->next_resync = sector_nr;
2305
2306         rcu_read_lock();
2307         /*
2308          * If we get a correctably read error during resync or recovery,
2309          * we might want to read from a different device.  So we
2310          * flag all drives that could conceivably be read from for READ,
2311          * and any others (which will be non-In_sync devices) for WRITE.
2312          * If a read fails, we try reading from something else for which READ
2313          * is OK.
2314          */
2315
2316         r1_bio->mddev = mddev;
2317         r1_bio->sector = sector_nr;
2318         r1_bio->state = 0;
2319         set_bit(R1BIO_IsSync, &r1_bio->state);
2320
2321         for (i = 0; i < conf->raid_disks * 2; i++) {
2322                 struct md_rdev *rdev;
2323                 bio = r1_bio->bios[i];
2324
2325                 /* take from bio_init */
2326                 bio->bi_next = NULL;
2327                 bio->bi_flags &= ~(BIO_POOL_MASK-1);
2328                 bio->bi_flags |= 1 << BIO_UPTODATE;
2329                 bio->bi_rw = READ;
2330                 bio->bi_vcnt = 0;
2331                 bio->bi_idx = 0;
2332                 bio->bi_phys_segments = 0;
2333                 bio->bi_size = 0;
2334                 bio->bi_end_io = NULL;
2335                 bio->bi_private = NULL;
2336
2337                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2338                 if (rdev == NULL ||
2339                     test_bit(Faulty, &rdev->flags)) {
2340                         if (i < conf->raid_disks)
2341                                 still_degraded = 1;
2342                 } else if (!test_bit(In_sync, &rdev->flags)) {
2343                         bio->bi_rw = WRITE;
2344                         bio->bi_end_io = end_sync_write;
2345                         write_targets ++;
2346                 } else {
2347                         /* may need to read from here */
2348                         sector_t first_bad = MaxSector;
2349                         int bad_sectors;
2350
2351                         if (is_badblock(rdev, sector_nr, good_sectors,
2352                                         &first_bad, &bad_sectors)) {
2353                                 if (first_bad > sector_nr)
2354                                         good_sectors = first_bad - sector_nr;
2355                                 else {
2356                                         bad_sectors -= (sector_nr - first_bad);
2357                                         if (min_bad == 0 ||
2358                                             min_bad > bad_sectors)
2359                                                 min_bad = bad_sectors;
2360                                 }
2361                         }
2362                         if (sector_nr < first_bad) {
2363                                 if (test_bit(WriteMostly, &rdev->flags)) {
2364                                         if (wonly < 0)
2365                                                 wonly = i;
2366                                 } else {
2367                                         if (disk < 0)
2368                                                 disk = i;
2369                                 }
2370                                 bio->bi_rw = READ;
2371                                 bio->bi_end_io = end_sync_read;
2372                                 read_targets++;
2373                         }
2374                 }
2375                 if (bio->bi_end_io) {
2376                         atomic_inc(&rdev->nr_pending);
2377                         bio->bi_sector = sector_nr + rdev->data_offset;
2378                         bio->bi_bdev = rdev->bdev;
2379                         bio->bi_private = r1_bio;
2380                 }
2381         }
2382         rcu_read_unlock();
2383         if (disk < 0)
2384                 disk = wonly;
2385         r1_bio->read_disk = disk;
2386
2387         if (read_targets == 0 && min_bad > 0) {
2388                 /* These sectors are bad on all InSync devices, so we
2389                  * need to mark them bad on all write targets
2390                  */
2391                 int ok = 1;
2392                 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2393                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2394                                 struct md_rdev *rdev = conf->mirrors[i].rdev;
2395                                 ok = rdev_set_badblocks(rdev, sector_nr,
2396                                                         min_bad, 0
2397                                         ) && ok;
2398                         }
2399                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2400                 *skipped = 1;
2401                 put_buf(r1_bio);
2402
2403                 if (!ok) {
2404                         /* Cannot record the badblocks, so need to
2405                          * abort the resync.
2406                          * If there are multiple read targets, could just
2407                          * fail the really bad ones ???
2408                          */
2409                         conf->recovery_disabled = mddev->recovery_disabled;
2410                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2411                         return 0;
2412                 } else
2413                         return min_bad;
2414
2415         }
2416         if (min_bad > 0 && min_bad < good_sectors) {
2417                 /* only resync enough to reach the next bad->good
2418                  * transition */
2419                 good_sectors = min_bad;
2420         }
2421
2422         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2423                 /* extra read targets are also write targets */
2424                 write_targets += read_targets-1;
2425
2426         if (write_targets == 0 || read_targets == 0) {
2427                 /* There is nowhere to write, so all non-sync
2428                  * drives must be failed - so we are finished
2429                  */
2430                 sector_t rv = max_sector - sector_nr;
2431                 *skipped = 1;
2432                 put_buf(r1_bio);
2433                 return rv;
2434         }
2435
2436         if (max_sector > mddev->resync_max)
2437                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2438         if (max_sector > sector_nr + good_sectors)
2439                 max_sector = sector_nr + good_sectors;
2440         nr_sectors = 0;
2441         sync_blocks = 0;
2442         do {
2443                 struct page *page;
2444                 int len = PAGE_SIZE;
2445                 if (sector_nr + (len>>9) > max_sector)
2446                         len = (max_sector - sector_nr) << 9;
2447                 if (len == 0)
2448                         break;
2449                 if (sync_blocks == 0) {
2450                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2451                                                &sync_blocks, still_degraded) &&
2452                             !conf->fullsync &&
2453                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2454                                 break;
2455                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2456                         if ((len >> 9) > sync_blocks)
2457                                 len = sync_blocks<<9;
2458                 }
2459
2460                 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2461                         bio = r1_bio->bios[i];
2462                         if (bio->bi_end_io) {
2463                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2464                                 if (bio_add_page(bio, page, len, 0) == 0) {
2465                                         /* stop here */
2466                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2467                                         while (i > 0) {
2468                                                 i--;
2469                                                 bio = r1_bio->bios[i];
2470                                                 if (bio->bi_end_io==NULL)
2471                                                         continue;
2472                                                 /* remove last page from this bio */
2473                                                 bio->bi_vcnt--;
2474                                                 bio->bi_size -= len;
2475                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2476                                         }
2477                                         goto bio_full;
2478                                 }
2479                         }
2480                 }
2481                 nr_sectors += len>>9;
2482                 sector_nr += len>>9;
2483                 sync_blocks -= (len>>9);
2484         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2485  bio_full:
2486         r1_bio->sectors = nr_sectors;
2487
2488         /* For a user-requested sync, we read all readable devices and do a
2489          * compare
2490          */
2491         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2492                 atomic_set(&r1_bio->remaining, read_targets);
2493                 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2494                         bio = r1_bio->bios[i];
2495                         if (bio->bi_end_io == end_sync_read) {
2496                                 read_targets--;
2497                                 md_sync_acct(bio->bi_bdev, nr_sectors);
2498                                 generic_make_request(bio);
2499                         }
2500                 }
2501         } else {
2502                 atomic_set(&r1_bio->remaining, 1);
2503                 bio = r1_bio->bios[r1_bio->read_disk];
2504                 md_sync_acct(bio->bi_bdev, nr_sectors);
2505                 generic_make_request(bio);
2506
2507         }
2508         return nr_sectors;
2509 }
2510
2511 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2512 {
2513         if (sectors)
2514                 return sectors;
2515
2516         return mddev->dev_sectors;
2517 }
2518
2519 static struct r1conf *setup_conf(struct mddev *mddev)
2520 {
2521         struct r1conf *conf;
2522         int i;
2523         struct mirror_info *disk;
2524         struct md_rdev *rdev;
2525         int err = -ENOMEM;
2526
2527         conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2528         if (!conf)
2529                 goto abort;
2530
2531         conf->mirrors = kzalloc(sizeof(struct mirror_info)
2532                                 * mddev->raid_disks * 2,
2533                                  GFP_KERNEL);
2534         if (!conf->mirrors)
2535                 goto abort;
2536
2537         conf->tmppage = alloc_page(GFP_KERNEL);
2538         if (!conf->tmppage)
2539                 goto abort;
2540
2541         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2542         if (!conf->poolinfo)
2543                 goto abort;
2544         conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2545         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2546                                           r1bio_pool_free,
2547                                           conf->poolinfo);
2548         if (!conf->r1bio_pool)
2549                 goto abort;
2550
2551         conf->poolinfo->mddev = mddev;
2552
2553         err = -EINVAL;
2554         spin_lock_init(&conf->device_lock);
2555         rdev_for_each(rdev, mddev) {
2556                 struct request_queue *q;
2557                 int disk_idx = rdev->raid_disk;
2558                 if (disk_idx >= mddev->raid_disks
2559                     || disk_idx < 0)
2560                         continue;
2561                 if (test_bit(Replacement, &rdev->flags))
2562                         disk = conf->mirrors + conf->raid_disks + disk_idx;
2563                 else
2564                         disk = conf->mirrors + disk_idx;
2565
2566                 if (disk->rdev)
2567                         goto abort;
2568                 disk->rdev = rdev;
2569                 q = bdev_get_queue(rdev->bdev);
2570                 if (q->merge_bvec_fn)
2571                         mddev->merge_check_needed = 1;
2572
2573                 disk->head_position = 0;
2574         }
2575         conf->raid_disks = mddev->raid_disks;
2576         conf->mddev = mddev;
2577         INIT_LIST_HEAD(&conf->retry_list);
2578
2579         spin_lock_init(&conf->resync_lock);
2580         init_waitqueue_head(&conf->wait_barrier);
2581
2582         bio_list_init(&conf->pending_bio_list);
2583         conf->pending_count = 0;
2584         conf->recovery_disabled = mddev->recovery_disabled - 1;
2585
2586         err = -EIO;
2587         conf->last_used = -1;
2588         for (i = 0; i < conf->raid_disks * 2; i++) {
2589
2590                 disk = conf->mirrors + i;
2591
2592                 if (i < conf->raid_disks &&
2593                     disk[conf->raid_disks].rdev) {
2594                         /* This slot has a replacement. */
2595                         if (!disk->rdev) {
2596                                 /* No original, just make the replacement
2597                                  * a recovering spare
2598                                  */
2599                                 disk->rdev =
2600                                         disk[conf->raid_disks].rdev;
2601                                 disk[conf->raid_disks].rdev = NULL;
2602                         } else if (!test_bit(In_sync, &disk->rdev->flags))
2603                                 /* Original is not in_sync - bad */
2604                                 goto abort;
2605                 }
2606
2607                 if (!disk->rdev ||
2608                     !test_bit(In_sync, &disk->rdev->flags)) {
2609                         disk->head_position = 0;
2610                         if (disk->rdev &&
2611                             (disk->rdev->saved_raid_disk < 0))
2612                                 conf->fullsync = 1;
2613                 } else if (conf->last_used < 0)
2614                         /*
2615                          * The first working device is used as a
2616                          * starting point to read balancing.
2617                          */
2618                         conf->last_used = i;
2619         }
2620
2621         if (conf->last_used < 0) {
2622                 printk(KERN_ERR "md/raid1:%s: no operational mirrors\n",
2623                        mdname(mddev));
2624                 goto abort;
2625         }
2626         err = -ENOMEM;
2627         conf->thread = md_register_thread(raid1d, mddev, "raid1");
2628         if (!conf->thread) {
2629                 printk(KERN_ERR
2630                        "md/raid1:%s: couldn't allocate thread\n",
2631                        mdname(mddev));
2632                 goto abort;
2633         }
2634
2635         return conf;
2636
2637  abort:
2638         if (conf) {
2639                 if (conf->r1bio_pool)
2640                         mempool_destroy(conf->r1bio_pool);
2641                 kfree(conf->mirrors);
2642                 safe_put_page(conf->tmppage);
2643                 kfree(conf->poolinfo);
2644                 kfree(conf);
2645         }
2646         return ERR_PTR(err);
2647 }
2648
2649 static int stop(struct mddev *mddev);
2650 static int run(struct mddev *mddev)
2651 {
2652         struct r1conf *conf;
2653         int i;
2654         struct md_rdev *rdev;
2655         int ret;
2656
2657         if (mddev->level != 1) {
2658                 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2659                        mdname(mddev), mddev->level);
2660                 return -EIO;
2661         }
2662         if (mddev->reshape_position != MaxSector) {
2663                 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2664                        mdname(mddev));
2665                 return -EIO;
2666         }
2667         /*
2668          * copy the already verified devices into our private RAID1
2669          * bookkeeping area. [whatever we allocate in run(),
2670          * should be freed in stop()]
2671          */
2672         if (mddev->private == NULL)
2673                 conf = setup_conf(mddev);
2674         else
2675                 conf = mddev->private;
2676
2677         if (IS_ERR(conf))
2678                 return PTR_ERR(conf);
2679
2680         rdev_for_each(rdev, mddev) {
2681                 if (!mddev->gendisk)
2682                         continue;
2683                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2684                                   rdev->data_offset << 9);
2685         }
2686
2687         mddev->degraded = 0;
2688         for (i=0; i < conf->raid_disks; i++)
2689                 if (conf->mirrors[i].rdev == NULL ||
2690                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2691                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2692                         mddev->degraded++;
2693
2694         if (conf->raid_disks - mddev->degraded == 1)
2695                 mddev->recovery_cp = MaxSector;
2696
2697         if (mddev->recovery_cp != MaxSector)
2698                 printk(KERN_NOTICE "md/raid1:%s: not clean"
2699                        " -- starting background reconstruction\n",
2700                        mdname(mddev));
2701         printk(KERN_INFO 
2702                 "md/raid1:%s: active with %d out of %d mirrors\n",
2703                 mdname(mddev), mddev->raid_disks - mddev->degraded, 
2704                 mddev->raid_disks);
2705
2706         /*
2707          * Ok, everything is just fine now
2708          */
2709         mddev->thread = conf->thread;
2710         conf->thread = NULL;
2711         mddev->private = conf;
2712
2713         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2714
2715         if (mddev->queue) {
2716                 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2717                 mddev->queue->backing_dev_info.congested_data = mddev;
2718                 blk_queue_merge_bvec(mddev->queue, raid1_mergeable_bvec);
2719         }
2720
2721         ret =  md_integrity_register(mddev);
2722         if (ret)
2723                 stop(mddev);
2724         return ret;
2725 }
2726
2727 static int stop(struct mddev *mddev)
2728 {
2729         struct r1conf *conf = mddev->private;
2730         struct bitmap *bitmap = mddev->bitmap;
2731
2732         /* wait for behind writes to complete */
2733         if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2734                 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2735                        mdname(mddev));
2736                 /* need to kick something here to make sure I/O goes? */
2737                 wait_event(bitmap->behind_wait,
2738                            atomic_read(&bitmap->behind_writes) == 0);
2739         }
2740
2741         raise_barrier(conf);
2742         lower_barrier(conf);
2743
2744         md_unregister_thread(&mddev->thread);
2745         if (conf->r1bio_pool)
2746                 mempool_destroy(conf->r1bio_pool);
2747         kfree(conf->mirrors);
2748         kfree(conf->poolinfo);
2749         kfree(conf);
2750         mddev->private = NULL;
2751         return 0;
2752 }
2753
2754 static int raid1_resize(struct mddev *mddev, sector_t sectors)
2755 {
2756         /* no resync is happening, and there is enough space
2757          * on all devices, so we can resize.
2758          * We need to make sure resync covers any new space.
2759          * If the array is shrinking we should possibly wait until
2760          * any io in the removed space completes, but it hardly seems
2761          * worth it.
2762          */
2763         sector_t newsize = raid1_size(mddev, sectors, 0);
2764         if (mddev->external_size &&
2765             mddev->array_sectors > newsize)
2766                 return -EINVAL;
2767         if (mddev->bitmap) {
2768                 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
2769                 if (ret)
2770                         return ret;
2771         }
2772         md_set_array_sectors(mddev, newsize);
2773         set_capacity(mddev->gendisk, mddev->array_sectors);
2774         revalidate_disk(mddev->gendisk);
2775         if (sectors > mddev->dev_sectors &&
2776             mddev->recovery_cp > mddev->dev_sectors) {
2777                 mddev->recovery_cp = mddev->dev_sectors;
2778                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2779         }
2780         mddev->dev_sectors = sectors;
2781         mddev->resync_max_sectors = sectors;
2782         return 0;
2783 }
2784
2785 static int raid1_reshape(struct mddev *mddev)
2786 {
2787         /* We need to:
2788          * 1/ resize the r1bio_pool
2789          * 2/ resize conf->mirrors
2790          *
2791          * We allocate a new r1bio_pool if we can.
2792          * Then raise a device barrier and wait until all IO stops.
2793          * Then resize conf->mirrors and swap in the new r1bio pool.
2794          *
2795          * At the same time, we "pack" the devices so that all the missing
2796          * devices have the higher raid_disk numbers.
2797          */
2798         mempool_t *newpool, *oldpool;
2799         struct pool_info *newpoolinfo;
2800         struct mirror_info *newmirrors;
2801         struct r1conf *conf = mddev->private;
2802         int cnt, raid_disks;
2803         unsigned long flags;
2804         int d, d2, err;
2805
2806         /* Cannot change chunk_size, layout, or level */
2807         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2808             mddev->layout != mddev->new_layout ||
2809             mddev->level != mddev->new_level) {
2810                 mddev->new_chunk_sectors = mddev->chunk_sectors;
2811                 mddev->new_layout = mddev->layout;
2812                 mddev->new_level = mddev->level;
2813                 return -EINVAL;
2814         }
2815
2816         err = md_allow_write(mddev);
2817         if (err)
2818                 return err;
2819
2820         raid_disks = mddev->raid_disks + mddev->delta_disks;
2821
2822         if (raid_disks < conf->raid_disks) {
2823                 cnt=0;
2824                 for (d= 0; d < conf->raid_disks; d++)
2825                         if (conf->mirrors[d].rdev)
2826                                 cnt++;
2827                 if (cnt > raid_disks)
2828                         return -EBUSY;
2829         }
2830
2831         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2832         if (!newpoolinfo)
2833                 return -ENOMEM;
2834         newpoolinfo->mddev = mddev;
2835         newpoolinfo->raid_disks = raid_disks * 2;
2836
2837         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2838                                  r1bio_pool_free, newpoolinfo);
2839         if (!newpool) {
2840                 kfree(newpoolinfo);
2841                 return -ENOMEM;
2842         }
2843         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks * 2,
2844                              GFP_KERNEL);
2845         if (!newmirrors) {
2846                 kfree(newpoolinfo);
2847                 mempool_destroy(newpool);
2848                 return -ENOMEM;
2849         }
2850
2851         raise_barrier(conf);
2852
2853         /* ok, everything is stopped */
2854         oldpool = conf->r1bio_pool;
2855         conf->r1bio_pool = newpool;
2856
2857         for (d = d2 = 0; d < conf->raid_disks; d++) {
2858                 struct md_rdev *rdev = conf->mirrors[d].rdev;
2859                 if (rdev && rdev->raid_disk != d2) {
2860                         sysfs_unlink_rdev(mddev, rdev);
2861                         rdev->raid_disk = d2;
2862                         sysfs_unlink_rdev(mddev, rdev);
2863                         if (sysfs_link_rdev(mddev, rdev))
2864                                 printk(KERN_WARNING
2865                                        "md/raid1:%s: cannot register rd%d\n",
2866                                        mdname(mddev), rdev->raid_disk);
2867                 }
2868                 if (rdev)
2869                         newmirrors[d2++].rdev = rdev;
2870         }
2871         kfree(conf->mirrors);
2872         conf->mirrors = newmirrors;
2873         kfree(conf->poolinfo);
2874         conf->poolinfo = newpoolinfo;
2875
2876         spin_lock_irqsave(&conf->device_lock, flags);
2877         mddev->degraded += (raid_disks - conf->raid_disks);
2878         spin_unlock_irqrestore(&conf->device_lock, flags);
2879         conf->raid_disks = mddev->raid_disks = raid_disks;
2880         mddev->delta_disks = 0;
2881
2882         conf->last_used = 0; /* just make sure it is in-range */
2883         lower_barrier(conf);
2884
2885         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2886         md_wakeup_thread(mddev->thread);
2887
2888         mempool_destroy(oldpool);
2889         return 0;
2890 }
2891
2892 static void raid1_quiesce(struct mddev *mddev, int state)
2893 {
2894         struct r1conf *conf = mddev->private;
2895
2896         switch(state) {
2897         case 2: /* wake for suspend */
2898                 wake_up(&conf->wait_barrier);
2899                 break;
2900         case 1:
2901                 raise_barrier(conf);
2902                 break;
2903         case 0:
2904                 lower_barrier(conf);
2905                 break;
2906         }
2907 }
2908
2909 static void *raid1_takeover(struct mddev *mddev)
2910 {
2911         /* raid1 can take over:
2912          *  raid5 with 2 devices, any layout or chunk size
2913          */
2914         if (mddev->level == 5 && mddev->raid_disks == 2) {
2915                 struct r1conf *conf;
2916                 mddev->new_level = 1;
2917                 mddev->new_layout = 0;
2918                 mddev->new_chunk_sectors = 0;
2919                 conf = setup_conf(mddev);
2920                 if (!IS_ERR(conf))
2921                         conf->barrier = 1;
2922                 return conf;
2923         }
2924         return ERR_PTR(-EINVAL);
2925 }
2926
2927 static struct md_personality raid1_personality =
2928 {
2929         .name           = "raid1",
2930         .level          = 1,
2931         .owner          = THIS_MODULE,
2932         .make_request   = make_request,
2933         .run            = run,
2934         .stop           = stop,
2935         .status         = status,
2936         .error_handler  = error,
2937         .hot_add_disk   = raid1_add_disk,
2938         .hot_remove_disk= raid1_remove_disk,
2939         .spare_active   = raid1_spare_active,
2940         .sync_request   = sync_request,
2941         .resize         = raid1_resize,
2942         .size           = raid1_size,
2943         .check_reshape  = raid1_reshape,
2944         .quiesce        = raid1_quiesce,
2945         .takeover       = raid1_takeover,
2946 };
2947
2948 static int __init raid_init(void)
2949 {
2950         return register_md_personality(&raid1_personality);
2951 }
2952
2953 static void raid_exit(void)
2954 {
2955         unregister_md_personality(&raid1_personality);
2956 }
2957
2958 module_init(raid_init);
2959 module_exit(raid_exit);
2960 MODULE_LICENSE("GPL");
2961 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2962 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2963 MODULE_ALIAS("md-raid1");
2964 MODULE_ALIAS("md-level-1");
2965
2966 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);