e3cfca7cde16e32c92949d8f2e3cc69d26217cae
[~shefty/rdma-dev.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                 set_bit(R1BIO_WriteError, &r1_bio->state);
395         } else {
396                 /*
397                  * Set R1BIO_Uptodate in our master bio, so that we
398                  * will return a good error code for to the higher
399                  * levels even if IO on some other mirrored buffer
400                  * fails.
401                  *
402                  * The 'master' represents the composite IO operation
403                  * to user-side. So if something waits for IO, then it
404                  * will wait for the 'master' bio.
405                  */
406                 sector_t first_bad;
407                 int bad_sectors;
408
409                 r1_bio->bios[mirror] = NULL;
410                 to_put = bio;
411                 set_bit(R1BIO_Uptodate, &r1_bio->state);
412
413                 /* Maybe we can clear some bad blocks. */
414                 if (is_badblock(conf->mirrors[mirror].rdev,
415                                 r1_bio->sector, r1_bio->sectors,
416                                 &first_bad, &bad_sectors)) {
417                         r1_bio->bios[mirror] = IO_MADE_GOOD;
418                         set_bit(R1BIO_MadeGood, &r1_bio->state);
419                 }
420         }
421
422         if (behind) {
423                 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
424                         atomic_dec(&r1_bio->behind_remaining);
425
426                 /*
427                  * In behind mode, we ACK the master bio once the I/O
428                  * has safely reached all non-writemostly
429                  * disks. Setting the Returned bit ensures that this
430                  * gets done only once -- we don't ever want to return
431                  * -EIO here, instead we'll wait
432                  */
433                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
434                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
435                         /* Maybe we can return now */
436                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
437                                 struct bio *mbio = r1_bio->master_bio;
438                                 pr_debug("raid1: behind end write sectors"
439                                          " %llu-%llu\n",
440                                          (unsigned long long) mbio->bi_sector,
441                                          (unsigned long long) mbio->bi_sector +
442                                          (mbio->bi_size >> 9) - 1);
443                                 call_bio_endio(r1_bio);
444                         }
445                 }
446         }
447         if (r1_bio->bios[mirror] == NULL)
448                 rdev_dec_pending(conf->mirrors[mirror].rdev,
449                                  conf->mddev);
450
451         /*
452          * Let's see if all mirrored write operations have finished
453          * already.
454          */
455         r1_bio_write_done(r1_bio);
456
457         if (to_put)
458                 bio_put(to_put);
459 }
460
461
462 /*
463  * This routine returns the disk from which the requested read should
464  * be done. There is a per-array 'next expected sequential IO' sector
465  * number - if this matches on the next IO then we use the last disk.
466  * There is also a per-disk 'last know head position' sector that is
467  * maintained from IRQ contexts, both the normal and the resync IO
468  * completion handlers update this position correctly. If there is no
469  * perfect sequential match then we pick the disk whose head is closest.
470  *
471  * If there are 2 mirrors in the same 2 devices, performance degrades
472  * because position is mirror, not device based.
473  *
474  * The rdev for the device selected will have nr_pending incremented.
475  */
476 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
477 {
478         const sector_t this_sector = r1_bio->sector;
479         int sectors;
480         int best_good_sectors;
481         int start_disk;
482         int best_disk;
483         int i;
484         sector_t best_dist;
485         struct md_rdev *rdev;
486         int choose_first;
487
488         rcu_read_lock();
489         /*
490          * Check if we can balance. We can balance on the whole
491          * device if no resync is going on, or below the resync window.
492          * We take the first readable disk when above the resync window.
493          */
494  retry:
495         sectors = r1_bio->sectors;
496         best_disk = -1;
497         best_dist = MaxSector;
498         best_good_sectors = 0;
499
500         if (conf->mddev->recovery_cp < MaxSector &&
501             (this_sector + sectors >= conf->next_resync)) {
502                 choose_first = 1;
503                 start_disk = 0;
504         } else {
505                 choose_first = 0;
506                 start_disk = conf->last_used;
507         }
508
509         for (i = 0 ; i < conf->raid_disks * 2 ; i++) {
510                 sector_t dist;
511                 sector_t first_bad;
512                 int bad_sectors;
513
514                 int disk = start_disk + i;
515                 if (disk >= conf->raid_disks)
516                         disk -= conf->raid_disks;
517
518                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
519                 if (r1_bio->bios[disk] == IO_BLOCKED
520                     || rdev == NULL
521                     || test_bit(Faulty, &rdev->flags))
522                         continue;
523                 if (!test_bit(In_sync, &rdev->flags) &&
524                     rdev->recovery_offset < this_sector + sectors)
525                         continue;
526                 if (test_bit(WriteMostly, &rdev->flags)) {
527                         /* Don't balance among write-mostly, just
528                          * use the first as a last resort */
529                         if (best_disk < 0)
530                                 best_disk = disk;
531                         continue;
532                 }
533                 /* This is a reasonable device to use.  It might
534                  * even be best.
535                  */
536                 if (is_badblock(rdev, this_sector, sectors,
537                                 &first_bad, &bad_sectors)) {
538                         if (best_dist < MaxSector)
539                                 /* already have a better device */
540                                 continue;
541                         if (first_bad <= this_sector) {
542                                 /* cannot read here. If this is the 'primary'
543                                  * device, then we must not read beyond
544                                  * bad_sectors from another device..
545                                  */
546                                 bad_sectors -= (this_sector - first_bad);
547                                 if (choose_first && sectors > bad_sectors)
548                                         sectors = bad_sectors;
549                                 if (best_good_sectors > sectors)
550                                         best_good_sectors = sectors;
551
552                         } else {
553                                 sector_t good_sectors = first_bad - this_sector;
554                                 if (good_sectors > best_good_sectors) {
555                                         best_good_sectors = good_sectors;
556                                         best_disk = disk;
557                                 }
558                                 if (choose_first)
559                                         break;
560                         }
561                         continue;
562                 } else
563                         best_good_sectors = sectors;
564
565                 dist = abs(this_sector - conf->mirrors[disk].head_position);
566                 if (choose_first
567                     /* Don't change to another disk for sequential reads */
568                     || conf->next_seq_sect == this_sector
569                     || dist == 0
570                     /* If device is idle, use it */
571                     || atomic_read(&rdev->nr_pending) == 0) {
572                         best_disk = disk;
573                         break;
574                 }
575                 if (dist < best_dist) {
576                         best_dist = dist;
577                         best_disk = disk;
578                 }
579         }
580
581         if (best_disk >= 0) {
582                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
583                 if (!rdev)
584                         goto retry;
585                 atomic_inc(&rdev->nr_pending);
586                 if (test_bit(Faulty, &rdev->flags)) {
587                         /* cannot risk returning a device that failed
588                          * before we inc'ed nr_pending
589                          */
590                         rdev_dec_pending(rdev, conf->mddev);
591                         goto retry;
592                 }
593                 sectors = best_good_sectors;
594                 conf->next_seq_sect = this_sector + sectors;
595                 conf->last_used = best_disk;
596         }
597         rcu_read_unlock();
598         *max_sectors = sectors;
599
600         return best_disk;
601 }
602
603 int md_raid1_congested(struct mddev *mddev, int bits)
604 {
605         struct r1conf *conf = mddev->private;
606         int i, ret = 0;
607
608         if ((bits & (1 << BDI_async_congested)) &&
609             conf->pending_count >= max_queued_requests)
610                 return 1;
611
612         rcu_read_lock();
613         for (i = 0; i < conf->raid_disks; i++) {
614                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
615                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
616                         struct request_queue *q = bdev_get_queue(rdev->bdev);
617
618                         BUG_ON(!q);
619
620                         /* Note the '|| 1' - when read_balance prefers
621                          * non-congested targets, it can be removed
622                          */
623                         if ((bits & (1<<BDI_async_congested)) || 1)
624                                 ret |= bdi_congested(&q->backing_dev_info, bits);
625                         else
626                                 ret &= bdi_congested(&q->backing_dev_info, bits);
627                 }
628         }
629         rcu_read_unlock();
630         return ret;
631 }
632 EXPORT_SYMBOL_GPL(md_raid1_congested);
633
634 static int raid1_congested(void *data, int bits)
635 {
636         struct mddev *mddev = data;
637
638         return mddev_congested(mddev, bits) ||
639                 md_raid1_congested(mddev, bits);
640 }
641
642 static void flush_pending_writes(struct r1conf *conf)
643 {
644         /* Any writes that have been queued but are awaiting
645          * bitmap updates get flushed here.
646          */
647         spin_lock_irq(&conf->device_lock);
648
649         if (conf->pending_bio_list.head) {
650                 struct bio *bio;
651                 bio = bio_list_get(&conf->pending_bio_list);
652                 conf->pending_count = 0;
653                 spin_unlock_irq(&conf->device_lock);
654                 /* flush any pending bitmap writes to
655                  * disk before proceeding w/ I/O */
656                 bitmap_unplug(conf->mddev->bitmap);
657                 wake_up(&conf->wait_barrier);
658
659                 while (bio) { /* submit pending writes */
660                         struct bio *next = bio->bi_next;
661                         bio->bi_next = NULL;
662                         generic_make_request(bio);
663                         bio = next;
664                 }
665         } else
666                 spin_unlock_irq(&conf->device_lock);
667 }
668
669 /* Barriers....
670  * Sometimes we need to suspend IO while we do something else,
671  * either some resync/recovery, or reconfigure the array.
672  * To do this we raise a 'barrier'.
673  * The 'barrier' is a counter that can be raised multiple times
674  * to count how many activities are happening which preclude
675  * normal IO.
676  * We can only raise the barrier if there is no pending IO.
677  * i.e. if nr_pending == 0.
678  * We choose only to raise the barrier if no-one is waiting for the
679  * barrier to go down.  This means that as soon as an IO request
680  * is ready, no other operations which require a barrier will start
681  * until the IO request has had a chance.
682  *
683  * So: regular IO calls 'wait_barrier'.  When that returns there
684  *    is no backgroup IO happening,  It must arrange to call
685  *    allow_barrier when it has finished its IO.
686  * backgroup IO calls must call raise_barrier.  Once that returns
687  *    there is no normal IO happeing.  It must arrange to call
688  *    lower_barrier when the particular background IO completes.
689  */
690 #define RESYNC_DEPTH 32
691
692 static void raise_barrier(struct r1conf *conf)
693 {
694         spin_lock_irq(&conf->resync_lock);
695
696         /* Wait until no block IO is waiting */
697         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
698                             conf->resync_lock, );
699
700         /* block any new IO from starting */
701         conf->barrier++;
702
703         /* Now wait for all pending IO to complete */
704         wait_event_lock_irq(conf->wait_barrier,
705                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
706                             conf->resync_lock, );
707
708         spin_unlock_irq(&conf->resync_lock);
709 }
710
711 static void lower_barrier(struct r1conf *conf)
712 {
713         unsigned long flags;
714         BUG_ON(conf->barrier <= 0);
715         spin_lock_irqsave(&conf->resync_lock, flags);
716         conf->barrier--;
717         spin_unlock_irqrestore(&conf->resync_lock, flags);
718         wake_up(&conf->wait_barrier);
719 }
720
721 static void wait_barrier(struct r1conf *conf)
722 {
723         spin_lock_irq(&conf->resync_lock);
724         if (conf->barrier) {
725                 conf->nr_waiting++;
726                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
727                                     conf->resync_lock,
728                                     );
729                 conf->nr_waiting--;
730         }
731         conf->nr_pending++;
732         spin_unlock_irq(&conf->resync_lock);
733 }
734
735 static void allow_barrier(struct r1conf *conf)
736 {
737         unsigned long flags;
738         spin_lock_irqsave(&conf->resync_lock, flags);
739         conf->nr_pending--;
740         spin_unlock_irqrestore(&conf->resync_lock, flags);
741         wake_up(&conf->wait_barrier);
742 }
743
744 static void freeze_array(struct r1conf *conf)
745 {
746         /* stop syncio and normal IO and wait for everything to
747          * go quite.
748          * We increment barrier and nr_waiting, and then
749          * wait until nr_pending match nr_queued+1
750          * This is called in the context of one normal IO request
751          * that has failed. Thus any sync request that might be pending
752          * will be blocked by nr_pending, and we need to wait for
753          * pending IO requests to complete or be queued for re-try.
754          * Thus the number queued (nr_queued) plus this request (1)
755          * must match the number of pending IOs (nr_pending) before
756          * we continue.
757          */
758         spin_lock_irq(&conf->resync_lock);
759         conf->barrier++;
760         conf->nr_waiting++;
761         wait_event_lock_irq(conf->wait_barrier,
762                             conf->nr_pending == conf->nr_queued+1,
763                             conf->resync_lock,
764                             flush_pending_writes(conf));
765         spin_unlock_irq(&conf->resync_lock);
766 }
767 static void unfreeze_array(struct r1conf *conf)
768 {
769         /* reverse the effect of the freeze */
770         spin_lock_irq(&conf->resync_lock);
771         conf->barrier--;
772         conf->nr_waiting--;
773         wake_up(&conf->wait_barrier);
774         spin_unlock_irq(&conf->resync_lock);
775 }
776
777
778 /* duplicate the data pages for behind I/O 
779  */
780 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
781 {
782         int i;
783         struct bio_vec *bvec;
784         struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
785                                         GFP_NOIO);
786         if (unlikely(!bvecs))
787                 return;
788
789         bio_for_each_segment(bvec, bio, i) {
790                 bvecs[i] = *bvec;
791                 bvecs[i].bv_page = alloc_page(GFP_NOIO);
792                 if (unlikely(!bvecs[i].bv_page))
793                         goto do_sync_io;
794                 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
795                        kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
796                 kunmap(bvecs[i].bv_page);
797                 kunmap(bvec->bv_page);
798         }
799         r1_bio->behind_bvecs = bvecs;
800         r1_bio->behind_page_count = bio->bi_vcnt;
801         set_bit(R1BIO_BehindIO, &r1_bio->state);
802         return;
803
804 do_sync_io:
805         for (i = 0; i < bio->bi_vcnt; i++)
806                 if (bvecs[i].bv_page)
807                         put_page(bvecs[i].bv_page);
808         kfree(bvecs);
809         pr_debug("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
810 }
811
812 static void make_request(struct mddev *mddev, struct bio * bio)
813 {
814         struct r1conf *conf = mddev->private;
815         struct mirror_info *mirror;
816         struct r1bio *r1_bio;
817         struct bio *read_bio;
818         int i, disks;
819         struct bitmap *bitmap;
820         unsigned long flags;
821         const int rw = bio_data_dir(bio);
822         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
823         const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
824         struct md_rdev *blocked_rdev;
825         int plugged;
826         int first_clone;
827         int sectors_handled;
828         int max_sectors;
829
830         /*
831          * Register the new request and wait if the reconstruction
832          * thread has put up a bar for new requests.
833          * Continue immediately if no resync is active currently.
834          */
835
836         md_write_start(mddev, bio); /* wait on superblock update early */
837
838         if (bio_data_dir(bio) == WRITE &&
839             bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
840             bio->bi_sector < mddev->suspend_hi) {
841                 /* As the suspend_* range is controlled by
842                  * userspace, we want an interruptible
843                  * wait.
844                  */
845                 DEFINE_WAIT(w);
846                 for (;;) {
847                         flush_signals(current);
848                         prepare_to_wait(&conf->wait_barrier,
849                                         &w, TASK_INTERRUPTIBLE);
850                         if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
851                             bio->bi_sector >= mddev->suspend_hi)
852                                 break;
853                         schedule();
854                 }
855                 finish_wait(&conf->wait_barrier, &w);
856         }
857
858         wait_barrier(conf);
859
860         bitmap = mddev->bitmap;
861
862         /*
863          * make_request() can abort the operation when READA is being
864          * used and no empty request is available.
865          *
866          */
867         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
868
869         r1_bio->master_bio = bio;
870         r1_bio->sectors = bio->bi_size >> 9;
871         r1_bio->state = 0;
872         r1_bio->mddev = mddev;
873         r1_bio->sector = bio->bi_sector;
874
875         /* We might need to issue multiple reads to different
876          * devices if there are bad blocks around, so we keep
877          * track of the number of reads in bio->bi_phys_segments.
878          * If this is 0, there is only one r1_bio and no locking
879          * will be needed when requests complete.  If it is
880          * non-zero, then it is the number of not-completed requests.
881          */
882         bio->bi_phys_segments = 0;
883         clear_bit(BIO_SEG_VALID, &bio->bi_flags);
884
885         if (rw == READ) {
886                 /*
887                  * read balancing logic:
888                  */
889                 int rdisk;
890
891 read_again:
892                 rdisk = read_balance(conf, r1_bio, &max_sectors);
893
894                 if (rdisk < 0) {
895                         /* couldn't find anywhere to read from */
896                         raid_end_bio_io(r1_bio);
897                         return;
898                 }
899                 mirror = conf->mirrors + rdisk;
900
901                 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
902                     bitmap) {
903                         /* Reading from a write-mostly device must
904                          * take care not to over-take any writes
905                          * that are 'behind'
906                          */
907                         wait_event(bitmap->behind_wait,
908                                    atomic_read(&bitmap->behind_writes) == 0);
909                 }
910                 r1_bio->read_disk = rdisk;
911
912                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
913                 md_trim_bio(read_bio, r1_bio->sector - bio->bi_sector,
914                             max_sectors);
915
916                 r1_bio->bios[rdisk] = read_bio;
917
918                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
919                 read_bio->bi_bdev = mirror->rdev->bdev;
920                 read_bio->bi_end_io = raid1_end_read_request;
921                 read_bio->bi_rw = READ | do_sync;
922                 read_bio->bi_private = r1_bio;
923
924                 if (max_sectors < r1_bio->sectors) {
925                         /* could not read all from this device, so we will
926                          * need another r1_bio.
927                          */
928
929                         sectors_handled = (r1_bio->sector + max_sectors
930                                            - bio->bi_sector);
931                         r1_bio->sectors = max_sectors;
932                         spin_lock_irq(&conf->device_lock);
933                         if (bio->bi_phys_segments == 0)
934                                 bio->bi_phys_segments = 2;
935                         else
936                                 bio->bi_phys_segments++;
937                         spin_unlock_irq(&conf->device_lock);
938                         /* Cannot call generic_make_request directly
939                          * as that will be queued in __make_request
940                          * and subsequent mempool_alloc might block waiting
941                          * for it.  So hand bio over to raid1d.
942                          */
943                         reschedule_retry(r1_bio);
944
945                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
946
947                         r1_bio->master_bio = bio;
948                         r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
949                         r1_bio->state = 0;
950                         r1_bio->mddev = mddev;
951                         r1_bio->sector = bio->bi_sector + sectors_handled;
952                         goto read_again;
953                 } else
954                         generic_make_request(read_bio);
955                 return;
956         }
957
958         /*
959          * WRITE:
960          */
961         if (conf->pending_count >= max_queued_requests) {
962                 md_wakeup_thread(mddev->thread);
963                 wait_event(conf->wait_barrier,
964                            conf->pending_count < max_queued_requests);
965         }
966         /* first select target devices under rcu_lock and
967          * inc refcount on their rdev.  Record them by setting
968          * bios[x] to bio
969          * If there are known/acknowledged bad blocks on any device on
970          * which we have seen a write error, we want to avoid writing those
971          * blocks.
972          * This potentially requires several writes to write around
973          * the bad blocks.  Each set of writes gets it's own r1bio
974          * with a set of bios attached.
975          */
976         plugged = mddev_check_plugged(mddev);
977
978         disks = conf->raid_disks * 2;
979  retry_write:
980         blocked_rdev = NULL;
981         rcu_read_lock();
982         max_sectors = r1_bio->sectors;
983         for (i = 0;  i < disks; i++) {
984                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
985                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
986                         atomic_inc(&rdev->nr_pending);
987                         blocked_rdev = rdev;
988                         break;
989                 }
990                 r1_bio->bios[i] = NULL;
991                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
992                         if (i < conf->raid_disks)
993                                 set_bit(R1BIO_Degraded, &r1_bio->state);
994                         continue;
995                 }
996
997                 atomic_inc(&rdev->nr_pending);
998                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
999                         sector_t first_bad;
1000                         int bad_sectors;
1001                         int is_bad;
1002
1003                         is_bad = is_badblock(rdev, r1_bio->sector,
1004                                              max_sectors,
1005                                              &first_bad, &bad_sectors);
1006                         if (is_bad < 0) {
1007                                 /* mustn't write here until the bad block is
1008                                  * acknowledged*/
1009                                 set_bit(BlockedBadBlocks, &rdev->flags);
1010                                 blocked_rdev = rdev;
1011                                 break;
1012                         }
1013                         if (is_bad && first_bad <= r1_bio->sector) {
1014                                 /* Cannot write here at all */
1015                                 bad_sectors -= (r1_bio->sector - first_bad);
1016                                 if (bad_sectors < max_sectors)
1017                                         /* mustn't write more than bad_sectors
1018                                          * to other devices yet
1019                                          */
1020                                         max_sectors = bad_sectors;
1021                                 rdev_dec_pending(rdev, mddev);
1022                                 /* We don't set R1BIO_Degraded as that
1023                                  * only applies if the disk is
1024                                  * missing, so it might be re-added,
1025                                  * and we want to know to recover this
1026                                  * chunk.
1027                                  * In this case the device is here,
1028                                  * and the fact that this chunk is not
1029                                  * in-sync is recorded in the bad
1030                                  * block log
1031                                  */
1032                                 continue;
1033                         }
1034                         if (is_bad) {
1035                                 int good_sectors = first_bad - r1_bio->sector;
1036                                 if (good_sectors < max_sectors)
1037                                         max_sectors = good_sectors;
1038                         }
1039                 }
1040                 r1_bio->bios[i] = bio;
1041         }
1042         rcu_read_unlock();
1043
1044         if (unlikely(blocked_rdev)) {
1045                 /* Wait for this device to become unblocked */
1046                 int j;
1047
1048                 for (j = 0; j < i; j++)
1049                         if (r1_bio->bios[j])
1050                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1051                 r1_bio->state = 0;
1052                 allow_barrier(conf);
1053                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1054                 wait_barrier(conf);
1055                 goto retry_write;
1056         }
1057
1058         if (max_sectors < r1_bio->sectors) {
1059                 /* We are splitting this write into multiple parts, so
1060                  * we need to prepare for allocating another r1_bio.
1061                  */
1062                 r1_bio->sectors = max_sectors;
1063                 spin_lock_irq(&conf->device_lock);
1064                 if (bio->bi_phys_segments == 0)
1065                         bio->bi_phys_segments = 2;
1066                 else
1067                         bio->bi_phys_segments++;
1068                 spin_unlock_irq(&conf->device_lock);
1069         }
1070         sectors_handled = r1_bio->sector + max_sectors - bio->bi_sector;
1071
1072         atomic_set(&r1_bio->remaining, 1);
1073         atomic_set(&r1_bio->behind_remaining, 0);
1074
1075         first_clone = 1;
1076         for (i = 0; i < disks; i++) {
1077                 struct bio *mbio;
1078                 if (!r1_bio->bios[i])
1079                         continue;
1080
1081                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1082                 md_trim_bio(mbio, r1_bio->sector - bio->bi_sector, max_sectors);
1083
1084                 if (first_clone) {
1085                         /* do behind I/O ?
1086                          * Not if there are too many, or cannot
1087                          * allocate memory, or a reader on WriteMostly
1088                          * is waiting for behind writes to flush */
1089                         if (bitmap &&
1090                             (atomic_read(&bitmap->behind_writes)
1091                              < mddev->bitmap_info.max_write_behind) &&
1092                             !waitqueue_active(&bitmap->behind_wait))
1093                                 alloc_behind_pages(mbio, r1_bio);
1094
1095                         bitmap_startwrite(bitmap, r1_bio->sector,
1096                                           r1_bio->sectors,
1097                                           test_bit(R1BIO_BehindIO,
1098                                                    &r1_bio->state));
1099                         first_clone = 0;
1100                 }
1101                 if (r1_bio->behind_bvecs) {
1102                         struct bio_vec *bvec;
1103                         int j;
1104
1105                         /* Yes, I really want the '__' version so that
1106                          * we clear any unused pointer in the io_vec, rather
1107                          * than leave them unchanged.  This is important
1108                          * because when we come to free the pages, we won't
1109                          * know the original bi_idx, so we just free
1110                          * them all
1111                          */
1112                         __bio_for_each_segment(bvec, mbio, j, 0)
1113                                 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1114                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1115                                 atomic_inc(&r1_bio->behind_remaining);
1116                 }
1117
1118                 r1_bio->bios[i] = mbio;
1119
1120                 mbio->bi_sector = (r1_bio->sector +
1121                                    conf->mirrors[i].rdev->data_offset);
1122                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1123                 mbio->bi_end_io = raid1_end_write_request;
1124                 mbio->bi_rw = WRITE | do_flush_fua | do_sync;
1125                 mbio->bi_private = r1_bio;
1126
1127                 atomic_inc(&r1_bio->remaining);
1128                 spin_lock_irqsave(&conf->device_lock, flags);
1129                 bio_list_add(&conf->pending_bio_list, mbio);
1130                 conf->pending_count++;
1131                 spin_unlock_irqrestore(&conf->device_lock, flags);
1132         }
1133         /* Mustn't call r1_bio_write_done before this next test,
1134          * as it could result in the bio being freed.
1135          */
1136         if (sectors_handled < (bio->bi_size >> 9)) {
1137                 r1_bio_write_done(r1_bio);
1138                 /* We need another r1_bio.  It has already been counted
1139                  * in bio->bi_phys_segments
1140                  */
1141                 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1142                 r1_bio->master_bio = bio;
1143                 r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1144                 r1_bio->state = 0;
1145                 r1_bio->mddev = mddev;
1146                 r1_bio->sector = bio->bi_sector + sectors_handled;
1147                 goto retry_write;
1148         }
1149
1150         r1_bio_write_done(r1_bio);
1151
1152         /* In case raid1d snuck in to freeze_array */
1153         wake_up(&conf->wait_barrier);
1154
1155         if (do_sync || !bitmap || !plugged)
1156                 md_wakeup_thread(mddev->thread);
1157 }
1158
1159 static void status(struct seq_file *seq, struct mddev *mddev)
1160 {
1161         struct r1conf *conf = mddev->private;
1162         int i;
1163
1164         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1165                    conf->raid_disks - mddev->degraded);
1166         rcu_read_lock();
1167         for (i = 0; i < conf->raid_disks; i++) {
1168                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1169                 seq_printf(seq, "%s",
1170                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1171         }
1172         rcu_read_unlock();
1173         seq_printf(seq, "]");
1174 }
1175
1176
1177 static void error(struct mddev *mddev, struct md_rdev *rdev)
1178 {
1179         char b[BDEVNAME_SIZE];
1180         struct r1conf *conf = mddev->private;
1181
1182         /*
1183          * If it is not operational, then we have already marked it as dead
1184          * else if it is the last working disks, ignore the error, let the
1185          * next level up know.
1186          * else mark the drive as failed
1187          */
1188         if (test_bit(In_sync, &rdev->flags)
1189             && (conf->raid_disks - mddev->degraded) == 1) {
1190                 /*
1191                  * Don't fail the drive, act as though we were just a
1192                  * normal single drive.
1193                  * However don't try a recovery from this drive as
1194                  * it is very likely to fail.
1195                  */
1196                 conf->recovery_disabled = mddev->recovery_disabled;
1197                 return;
1198         }
1199         set_bit(Blocked, &rdev->flags);
1200         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1201                 unsigned long flags;
1202                 spin_lock_irqsave(&conf->device_lock, flags);
1203                 mddev->degraded++;
1204                 set_bit(Faulty, &rdev->flags);
1205                 spin_unlock_irqrestore(&conf->device_lock, flags);
1206                 /*
1207                  * if recovery is running, make sure it aborts.
1208                  */
1209                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1210         } else
1211                 set_bit(Faulty, &rdev->flags);
1212         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1213         printk(KERN_ALERT
1214                "md/raid1:%s: Disk failure on %s, disabling device.\n"
1215                "md/raid1:%s: Operation continuing on %d devices.\n",
1216                mdname(mddev), bdevname(rdev->bdev, b),
1217                mdname(mddev), conf->raid_disks - mddev->degraded);
1218 }
1219
1220 static void print_conf(struct r1conf *conf)
1221 {
1222         int i;
1223
1224         printk(KERN_DEBUG "RAID1 conf printout:\n");
1225         if (!conf) {
1226                 printk(KERN_DEBUG "(!conf)\n");
1227                 return;
1228         }
1229         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1230                 conf->raid_disks);
1231
1232         rcu_read_lock();
1233         for (i = 0; i < conf->raid_disks; i++) {
1234                 char b[BDEVNAME_SIZE];
1235                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1236                 if (rdev)
1237                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1238                                i, !test_bit(In_sync, &rdev->flags),
1239                                !test_bit(Faulty, &rdev->flags),
1240                                bdevname(rdev->bdev,b));
1241         }
1242         rcu_read_unlock();
1243 }
1244
1245 static void close_sync(struct r1conf *conf)
1246 {
1247         wait_barrier(conf);
1248         allow_barrier(conf);
1249
1250         mempool_destroy(conf->r1buf_pool);
1251         conf->r1buf_pool = NULL;
1252 }
1253
1254 static int raid1_spare_active(struct mddev *mddev)
1255 {
1256         int i;
1257         struct r1conf *conf = mddev->private;
1258         int count = 0;
1259         unsigned long flags;
1260
1261         /*
1262          * Find all failed disks within the RAID1 configuration 
1263          * and mark them readable.
1264          * Called under mddev lock, so rcu protection not needed.
1265          */
1266         for (i = 0; i < conf->raid_disks; i++) {
1267                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1268                 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1269                 if (repl
1270                     && repl->recovery_offset == MaxSector
1271                     && !test_bit(Faulty, &repl->flags)
1272                     && !test_and_set_bit(In_sync, &repl->flags)) {
1273                         /* replacement has just become active */
1274                         if (!rdev ||
1275                             !test_and_clear_bit(In_sync, &rdev->flags))
1276                                 count++;
1277                         if (rdev) {
1278                                 /* Replaced device not technically
1279                                  * faulty, but we need to be sure
1280                                  * it gets removed and never re-added
1281                                  */
1282                                 set_bit(Faulty, &rdev->flags);
1283                                 sysfs_notify_dirent_safe(
1284                                         rdev->sysfs_state);
1285                         }
1286                 }
1287                 if (rdev
1288                     && !test_bit(Faulty, &rdev->flags)
1289                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1290                         count++;
1291                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1292                 }
1293         }
1294         spin_lock_irqsave(&conf->device_lock, flags);
1295         mddev->degraded -= count;
1296         spin_unlock_irqrestore(&conf->device_lock, flags);
1297
1298         print_conf(conf);
1299         return count;
1300 }
1301
1302
1303 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1304 {
1305         struct r1conf *conf = mddev->private;
1306         int err = -EEXIST;
1307         int mirror = 0;
1308         struct mirror_info *p;
1309         int first = 0;
1310         int last = conf->raid_disks - 1;
1311
1312         if (mddev->recovery_disabled == conf->recovery_disabled)
1313                 return -EBUSY;
1314
1315         if (rdev->raid_disk >= 0)
1316                 first = last = rdev->raid_disk;
1317
1318         for (mirror = first; mirror <= last; mirror++) {
1319                 p = conf->mirrors+mirror;
1320                 if (!p->rdev) {
1321
1322                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1323                                           rdev->data_offset << 9);
1324                         /* as we don't honour merge_bvec_fn, we must
1325                          * never risk violating it, so limit
1326                          * ->max_segments to one lying with a single
1327                          * page, as a one page request is never in
1328                          * violation.
1329                          */
1330                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1331                                 blk_queue_max_segments(mddev->queue, 1);
1332                                 blk_queue_segment_boundary(mddev->queue,
1333                                                            PAGE_CACHE_SIZE - 1);
1334                         }
1335
1336                         p->head_position = 0;
1337                         rdev->raid_disk = mirror;
1338                         err = 0;
1339                         /* As all devices are equivalent, we don't need a full recovery
1340                          * if this was recently any drive of the array
1341                          */
1342                         if (rdev->saved_raid_disk < 0)
1343                                 conf->fullsync = 1;
1344                         rcu_assign_pointer(p->rdev, rdev);
1345                         break;
1346                 }
1347                 if (test_bit(WantReplacement, &p->rdev->flags) &&
1348                     p[conf->raid_disks].rdev == NULL) {
1349                         /* Add this device as a replacement */
1350                         clear_bit(In_sync, &rdev->flags);
1351                         set_bit(Replacement, &rdev->flags);
1352                         rdev->raid_disk = mirror;
1353                         err = 0;
1354                         conf->fullsync = 1;
1355                         rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1356                         break;
1357                 }
1358         }
1359         md_integrity_add_rdev(rdev, mddev);
1360         print_conf(conf);
1361         return err;
1362 }
1363
1364 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1365 {
1366         struct r1conf *conf = mddev->private;
1367         int err = 0;
1368         int number = rdev->raid_disk;
1369         struct mirror_info *p = conf->mirrors+ number;
1370
1371         if (rdev != p->rdev)
1372                 p = conf->mirrors + conf->raid_disks + number;
1373
1374         print_conf(conf);
1375         if (rdev == p->rdev) {
1376                 if (test_bit(In_sync, &rdev->flags) ||
1377                     atomic_read(&rdev->nr_pending)) {
1378                         err = -EBUSY;
1379                         goto abort;
1380                 }
1381                 /* Only remove non-faulty devices if recovery
1382                  * is not possible.
1383                  */
1384                 if (!test_bit(Faulty, &rdev->flags) &&
1385                     mddev->recovery_disabled != conf->recovery_disabled &&
1386                     mddev->degraded < conf->raid_disks) {
1387                         err = -EBUSY;
1388                         goto abort;
1389                 }
1390                 p->rdev = NULL;
1391                 synchronize_rcu();
1392                 if (atomic_read(&rdev->nr_pending)) {
1393                         /* lost the race, try later */
1394                         err = -EBUSY;
1395                         p->rdev = rdev;
1396                         goto abort;
1397                 } else if (conf->mirrors[conf->raid_disks + number].rdev) {
1398                         /* We just removed a device that is being replaced.
1399                          * Move down the replacement.  We drain all IO before
1400                          * doing this to avoid confusion.
1401                          */
1402                         struct md_rdev *repl =
1403                                 conf->mirrors[conf->raid_disks + number].rdev;
1404                         raise_barrier(conf);
1405                         clear_bit(Replacement, &repl->flags);
1406                         p->rdev = repl;
1407                         conf->mirrors[conf->raid_disks + number].rdev = NULL;
1408                         lower_barrier(conf);
1409                         clear_bit(WantReplacement, &rdev->flags);
1410                 } else
1411                         clear_bit(WantReplacement, &rdev->flags);
1412                 err = md_integrity_register(mddev);
1413         }
1414 abort:
1415
1416         print_conf(conf);
1417         return err;
1418 }
1419
1420
1421 static void end_sync_read(struct bio *bio, int error)
1422 {
1423         struct r1bio *r1_bio = bio->bi_private;
1424
1425         update_head_pos(r1_bio->read_disk, r1_bio);
1426
1427         /*
1428          * we have read a block, now it needs to be re-written,
1429          * or re-read if the read failed.
1430          * We don't do much here, just schedule handling by raid1d
1431          */
1432         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1433                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1434
1435         if (atomic_dec_and_test(&r1_bio->remaining))
1436                 reschedule_retry(r1_bio);
1437 }
1438
1439 static void end_sync_write(struct bio *bio, int error)
1440 {
1441         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1442         struct r1bio *r1_bio = bio->bi_private;
1443         struct mddev *mddev = r1_bio->mddev;
1444         struct r1conf *conf = mddev->private;
1445         int mirror=0;
1446         sector_t first_bad;
1447         int bad_sectors;
1448
1449         mirror = find_bio_disk(r1_bio, bio);
1450
1451         if (!uptodate) {
1452                 sector_t sync_blocks = 0;
1453                 sector_t s = r1_bio->sector;
1454                 long sectors_to_go = r1_bio->sectors;
1455                 /* make sure these bits doesn't get cleared. */
1456                 do {
1457                         bitmap_end_sync(mddev->bitmap, s,
1458                                         &sync_blocks, 1);
1459                         s += sync_blocks;
1460                         sectors_to_go -= sync_blocks;
1461                 } while (sectors_to_go > 0);
1462                 set_bit(WriteErrorSeen,
1463                         &conf->mirrors[mirror].rdev->flags);
1464                 set_bit(R1BIO_WriteError, &r1_bio->state);
1465         } else if (is_badblock(conf->mirrors[mirror].rdev,
1466                                r1_bio->sector,
1467                                r1_bio->sectors,
1468                                &first_bad, &bad_sectors) &&
1469                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1470                                 r1_bio->sector,
1471                                 r1_bio->sectors,
1472                                 &first_bad, &bad_sectors)
1473                 )
1474                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1475
1476         if (atomic_dec_and_test(&r1_bio->remaining)) {
1477                 int s = r1_bio->sectors;
1478                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1479                     test_bit(R1BIO_WriteError, &r1_bio->state))
1480                         reschedule_retry(r1_bio);
1481                 else {
1482                         put_buf(r1_bio);
1483                         md_done_sync(mddev, s, uptodate);
1484                 }
1485         }
1486 }
1487
1488 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1489                             int sectors, struct page *page, int rw)
1490 {
1491         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1492                 /* success */
1493                 return 1;
1494         if (rw == WRITE)
1495                 set_bit(WriteErrorSeen, &rdev->flags);
1496         /* need to record an error - either for the block or the device */
1497         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1498                 md_error(rdev->mddev, rdev);
1499         return 0;
1500 }
1501
1502 static int fix_sync_read_error(struct r1bio *r1_bio)
1503 {
1504         /* Try some synchronous reads of other devices to get
1505          * good data, much like with normal read errors.  Only
1506          * read into the pages we already have so we don't
1507          * need to re-issue the read request.
1508          * We don't need to freeze the array, because being in an
1509          * active sync request, there is no normal IO, and
1510          * no overlapping syncs.
1511          * We don't need to check is_badblock() again as we
1512          * made sure that anything with a bad block in range
1513          * will have bi_end_io clear.
1514          */
1515         struct mddev *mddev = r1_bio->mddev;
1516         struct r1conf *conf = mddev->private;
1517         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1518         sector_t sect = r1_bio->sector;
1519         int sectors = r1_bio->sectors;
1520         int idx = 0;
1521
1522         while(sectors) {
1523                 int s = sectors;
1524                 int d = r1_bio->read_disk;
1525                 int success = 0;
1526                 struct md_rdev *rdev;
1527                 int start;
1528
1529                 if (s > (PAGE_SIZE>>9))
1530                         s = PAGE_SIZE >> 9;
1531                 do {
1532                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1533                                 /* No rcu protection needed here devices
1534                                  * can only be removed when no resync is
1535                                  * active, and resync is currently active
1536                                  */
1537                                 rdev = conf->mirrors[d].rdev;
1538                                 if (sync_page_io(rdev, sect, s<<9,
1539                                                  bio->bi_io_vec[idx].bv_page,
1540                                                  READ, false)) {
1541                                         success = 1;
1542                                         break;
1543                                 }
1544                         }
1545                         d++;
1546                         if (d == conf->raid_disks * 2)
1547                                 d = 0;
1548                 } while (!success && d != r1_bio->read_disk);
1549
1550                 if (!success) {
1551                         char b[BDEVNAME_SIZE];
1552                         int abort = 0;
1553                         /* Cannot read from anywhere, this block is lost.
1554                          * Record a bad block on each device.  If that doesn't
1555                          * work just disable and interrupt the recovery.
1556                          * Don't fail devices as that won't really help.
1557                          */
1558                         printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1559                                " for block %llu\n",
1560                                mdname(mddev),
1561                                bdevname(bio->bi_bdev, b),
1562                                (unsigned long long)r1_bio->sector);
1563                         for (d = 0; d < conf->raid_disks * 2; d++) {
1564                                 rdev = conf->mirrors[d].rdev;
1565                                 if (!rdev || test_bit(Faulty, &rdev->flags))
1566                                         continue;
1567                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
1568                                         abort = 1;
1569                         }
1570                         if (abort) {
1571                                 conf->recovery_disabled =
1572                                         mddev->recovery_disabled;
1573                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1574                                 md_done_sync(mddev, r1_bio->sectors, 0);
1575                                 put_buf(r1_bio);
1576                                 return 0;
1577                         }
1578                         /* Try next page */
1579                         sectors -= s;
1580                         sect += s;
1581                         idx++;
1582                         continue;
1583                 }
1584
1585                 start = d;
1586                 /* write it back and re-read */
1587                 while (d != r1_bio->read_disk) {
1588                         if (d == 0)
1589                                 d = conf->raid_disks * 2;
1590                         d--;
1591                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1592                                 continue;
1593                         rdev = conf->mirrors[d].rdev;
1594                         if (r1_sync_page_io(rdev, sect, s,
1595                                             bio->bi_io_vec[idx].bv_page,
1596                                             WRITE) == 0) {
1597                                 r1_bio->bios[d]->bi_end_io = NULL;
1598                                 rdev_dec_pending(rdev, mddev);
1599                         }
1600                 }
1601                 d = start;
1602                 while (d != r1_bio->read_disk) {
1603                         if (d == 0)
1604                                 d = conf->raid_disks * 2;
1605                         d--;
1606                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1607                                 continue;
1608                         rdev = conf->mirrors[d].rdev;
1609                         if (r1_sync_page_io(rdev, sect, s,
1610                                             bio->bi_io_vec[idx].bv_page,
1611                                             READ) != 0)
1612                                 atomic_add(s, &rdev->corrected_errors);
1613                 }
1614                 sectors -= s;
1615                 sect += s;
1616                 idx ++;
1617         }
1618         set_bit(R1BIO_Uptodate, &r1_bio->state);
1619         set_bit(BIO_UPTODATE, &bio->bi_flags);
1620         return 1;
1621 }
1622
1623 static int process_checks(struct r1bio *r1_bio)
1624 {
1625         /* We have read all readable devices.  If we haven't
1626          * got the block, then there is no hope left.
1627          * If we have, then we want to do a comparison
1628          * and skip the write if everything is the same.
1629          * If any blocks failed to read, then we need to
1630          * attempt an over-write
1631          */
1632         struct mddev *mddev = r1_bio->mddev;
1633         struct r1conf *conf = mddev->private;
1634         int primary;
1635         int i;
1636
1637         for (primary = 0; primary < conf->raid_disks * 2; primary++)
1638                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1639                     test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1640                         r1_bio->bios[primary]->bi_end_io = NULL;
1641                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1642                         break;
1643                 }
1644         r1_bio->read_disk = primary;
1645         for (i = 0; i < conf->raid_disks * 2; i++) {
1646                 int j;
1647                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1648                 struct bio *pbio = r1_bio->bios[primary];
1649                 struct bio *sbio = r1_bio->bios[i];
1650                 int size;
1651
1652                 if (r1_bio->bios[i]->bi_end_io != end_sync_read)
1653                         continue;
1654
1655                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1656                         for (j = vcnt; j-- ; ) {
1657                                 struct page *p, *s;
1658                                 p = pbio->bi_io_vec[j].bv_page;
1659                                 s = sbio->bi_io_vec[j].bv_page;
1660                                 if (memcmp(page_address(p),
1661                                            page_address(s),
1662                                            PAGE_SIZE))
1663                                         break;
1664                         }
1665                 } else
1666                         j = 0;
1667                 if (j >= 0)
1668                         mddev->resync_mismatches += r1_bio->sectors;
1669                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1670                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1671                         /* No need to write to this device. */
1672                         sbio->bi_end_io = NULL;
1673                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1674                         continue;
1675                 }
1676                 /* fixup the bio for reuse */
1677                 sbio->bi_vcnt = vcnt;
1678                 sbio->bi_size = r1_bio->sectors << 9;
1679                 sbio->bi_idx = 0;
1680                 sbio->bi_phys_segments = 0;
1681                 sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1682                 sbio->bi_flags |= 1 << BIO_UPTODATE;
1683                 sbio->bi_next = NULL;
1684                 sbio->bi_sector = r1_bio->sector +
1685                         conf->mirrors[i].rdev->data_offset;
1686                 sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1687                 size = sbio->bi_size;
1688                 for (j = 0; j < vcnt ; j++) {
1689                         struct bio_vec *bi;
1690                         bi = &sbio->bi_io_vec[j];
1691                         bi->bv_offset = 0;
1692                         if (size > PAGE_SIZE)
1693                                 bi->bv_len = PAGE_SIZE;
1694                         else
1695                                 bi->bv_len = size;
1696                         size -= PAGE_SIZE;
1697                         memcpy(page_address(bi->bv_page),
1698                                page_address(pbio->bi_io_vec[j].bv_page),
1699                                PAGE_SIZE);
1700                 }
1701         }
1702         return 0;
1703 }
1704
1705 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
1706 {
1707         struct r1conf *conf = mddev->private;
1708         int i;
1709         int disks = conf->raid_disks * 2;
1710         struct bio *bio, *wbio;
1711
1712         bio = r1_bio->bios[r1_bio->read_disk];
1713
1714         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1715                 /* ouch - failed to read all of that. */
1716                 if (!fix_sync_read_error(r1_bio))
1717                         return;
1718
1719         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1720                 if (process_checks(r1_bio) < 0)
1721                         return;
1722         /*
1723          * schedule writes
1724          */
1725         atomic_set(&r1_bio->remaining, 1);
1726         for (i = 0; i < disks ; i++) {
1727                 wbio = r1_bio->bios[i];
1728                 if (wbio->bi_end_io == NULL ||
1729                     (wbio->bi_end_io == end_sync_read &&
1730                      (i == r1_bio->read_disk ||
1731                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1732                         continue;
1733
1734                 wbio->bi_rw = WRITE;
1735                 wbio->bi_end_io = end_sync_write;
1736                 atomic_inc(&r1_bio->remaining);
1737                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1738
1739                 generic_make_request(wbio);
1740         }
1741
1742         if (atomic_dec_and_test(&r1_bio->remaining)) {
1743                 /* if we're here, all write(s) have completed, so clean up */
1744                 md_done_sync(mddev, r1_bio->sectors, 1);
1745                 put_buf(r1_bio);
1746         }
1747 }
1748
1749 /*
1750  * This is a kernel thread which:
1751  *
1752  *      1.      Retries failed read operations on working mirrors.
1753  *      2.      Updates the raid superblock when problems encounter.
1754  *      3.      Performs writes following reads for array synchronising.
1755  */
1756
1757 static void fix_read_error(struct r1conf *conf, int read_disk,
1758                            sector_t sect, int sectors)
1759 {
1760         struct mddev *mddev = conf->mddev;
1761         while(sectors) {
1762                 int s = sectors;
1763                 int d = read_disk;
1764                 int success = 0;
1765                 int start;
1766                 struct md_rdev *rdev;
1767
1768                 if (s > (PAGE_SIZE>>9))
1769                         s = PAGE_SIZE >> 9;
1770
1771                 do {
1772                         /* Note: no rcu protection needed here
1773                          * as this is synchronous in the raid1d thread
1774                          * which is the thread that might remove
1775                          * a device.  If raid1d ever becomes multi-threaded....
1776                          */
1777                         sector_t first_bad;
1778                         int bad_sectors;
1779
1780                         rdev = conf->mirrors[d].rdev;
1781                         if (rdev &&
1782                             test_bit(In_sync, &rdev->flags) &&
1783                             is_badblock(rdev, sect, s,
1784                                         &first_bad, &bad_sectors) == 0 &&
1785                             sync_page_io(rdev, sect, s<<9,
1786                                          conf->tmppage, READ, false))
1787                                 success = 1;
1788                         else {
1789                                 d++;
1790                                 if (d == conf->raid_disks * 2)
1791                                         d = 0;
1792                         }
1793                 } while (!success && d != read_disk);
1794
1795                 if (!success) {
1796                         /* Cannot read from anywhere - mark it bad */
1797                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
1798                         if (!rdev_set_badblocks(rdev, sect, s, 0))
1799                                 md_error(mddev, rdev);
1800                         break;
1801                 }
1802                 /* write it back and re-read */
1803                 start = d;
1804                 while (d != read_disk) {
1805                         if (d==0)
1806                                 d = conf->raid_disks * 2;
1807                         d--;
1808                         rdev = conf->mirrors[d].rdev;
1809                         if (rdev &&
1810                             test_bit(In_sync, &rdev->flags))
1811                                 r1_sync_page_io(rdev, sect, s,
1812                                                 conf->tmppage, WRITE);
1813                 }
1814                 d = start;
1815                 while (d != read_disk) {
1816                         char b[BDEVNAME_SIZE];
1817                         if (d==0)
1818                                 d = conf->raid_disks * 2;
1819                         d--;
1820                         rdev = conf->mirrors[d].rdev;
1821                         if (rdev &&
1822                             test_bit(In_sync, &rdev->flags)) {
1823                                 if (r1_sync_page_io(rdev, sect, s,
1824                                                     conf->tmppage, READ)) {
1825                                         atomic_add(s, &rdev->corrected_errors);
1826                                         printk(KERN_INFO
1827                                                "md/raid1:%s: read error corrected "
1828                                                "(%d sectors at %llu on %s)\n",
1829                                                mdname(mddev), s,
1830                                                (unsigned long long)(sect +
1831                                                    rdev->data_offset),
1832                                                bdevname(rdev->bdev, b));
1833                                 }
1834                         }
1835                 }
1836                 sectors -= s;
1837                 sect += s;
1838         }
1839 }
1840
1841 static void bi_complete(struct bio *bio, int error)
1842 {
1843         complete((struct completion *)bio->bi_private);
1844 }
1845
1846 static int submit_bio_wait(int rw, struct bio *bio)
1847 {
1848         struct completion event;
1849         rw |= REQ_SYNC;
1850
1851         init_completion(&event);
1852         bio->bi_private = &event;
1853         bio->bi_end_io = bi_complete;
1854         submit_bio(rw, bio);
1855         wait_for_completion(&event);
1856
1857         return test_bit(BIO_UPTODATE, &bio->bi_flags);
1858 }
1859
1860 static int narrow_write_error(struct r1bio *r1_bio, int i)
1861 {
1862         struct mddev *mddev = r1_bio->mddev;
1863         struct r1conf *conf = mddev->private;
1864         struct md_rdev *rdev = conf->mirrors[i].rdev;
1865         int vcnt, idx;
1866         struct bio_vec *vec;
1867
1868         /* bio has the data to be written to device 'i' where
1869          * we just recently had a write error.
1870          * We repeatedly clone the bio and trim down to one block,
1871          * then try the write.  Where the write fails we record
1872          * a bad block.
1873          * It is conceivable that the bio doesn't exactly align with
1874          * blocks.  We must handle this somehow.
1875          *
1876          * We currently own a reference on the rdev.
1877          */
1878
1879         int block_sectors;
1880         sector_t sector;
1881         int sectors;
1882         int sect_to_write = r1_bio->sectors;
1883         int ok = 1;
1884
1885         if (rdev->badblocks.shift < 0)
1886                 return 0;
1887
1888         block_sectors = 1 << rdev->badblocks.shift;
1889         sector = r1_bio->sector;
1890         sectors = ((sector + block_sectors)
1891                    & ~(sector_t)(block_sectors - 1))
1892                 - sector;
1893
1894         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
1895                 vcnt = r1_bio->behind_page_count;
1896                 vec = r1_bio->behind_bvecs;
1897                 idx = 0;
1898                 while (vec[idx].bv_page == NULL)
1899                         idx++;
1900         } else {
1901                 vcnt = r1_bio->master_bio->bi_vcnt;
1902                 vec = r1_bio->master_bio->bi_io_vec;
1903                 idx = r1_bio->master_bio->bi_idx;
1904         }
1905         while (sect_to_write) {
1906                 struct bio *wbio;
1907                 if (sectors > sect_to_write)
1908                         sectors = sect_to_write;
1909                 /* Write at 'sector' for 'sectors'*/
1910
1911                 wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
1912                 memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
1913                 wbio->bi_sector = r1_bio->sector;
1914                 wbio->bi_rw = WRITE;
1915                 wbio->bi_vcnt = vcnt;
1916                 wbio->bi_size = r1_bio->sectors << 9;
1917                 wbio->bi_idx = idx;
1918
1919                 md_trim_bio(wbio, sector - r1_bio->sector, sectors);
1920                 wbio->bi_sector += rdev->data_offset;
1921                 wbio->bi_bdev = rdev->bdev;
1922                 if (submit_bio_wait(WRITE, wbio) == 0)
1923                         /* failure! */
1924                         ok = rdev_set_badblocks(rdev, sector,
1925                                                 sectors, 0)
1926                                 && ok;
1927
1928                 bio_put(wbio);
1929                 sect_to_write -= sectors;
1930                 sector += sectors;
1931                 sectors = block_sectors;
1932         }
1933         return ok;
1934 }
1935
1936 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
1937 {
1938         int m;
1939         int s = r1_bio->sectors;
1940         for (m = 0; m < conf->raid_disks * 2 ; m++) {
1941                 struct md_rdev *rdev = conf->mirrors[m].rdev;
1942                 struct bio *bio = r1_bio->bios[m];
1943                 if (bio->bi_end_io == NULL)
1944                         continue;
1945                 if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
1946                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
1947                         rdev_clear_badblocks(rdev, r1_bio->sector, s);
1948                 }
1949                 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
1950                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
1951                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
1952                                 md_error(conf->mddev, rdev);
1953                 }
1954         }
1955         put_buf(r1_bio);
1956         md_done_sync(conf->mddev, s, 1);
1957 }
1958
1959 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
1960 {
1961         int m;
1962         for (m = 0; m < conf->raid_disks * 2 ; m++)
1963                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
1964                         struct md_rdev *rdev = conf->mirrors[m].rdev;
1965                         rdev_clear_badblocks(rdev,
1966                                              r1_bio->sector,
1967                                              r1_bio->sectors);
1968                         rdev_dec_pending(rdev, conf->mddev);
1969                 } else if (r1_bio->bios[m] != NULL) {
1970                         /* This drive got a write error.  We need to
1971                          * narrow down and record precise write
1972                          * errors.
1973                          */
1974                         if (!narrow_write_error(r1_bio, m)) {
1975                                 md_error(conf->mddev,
1976                                          conf->mirrors[m].rdev);
1977                                 /* an I/O failed, we can't clear the bitmap */
1978                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1979                         }
1980                         rdev_dec_pending(conf->mirrors[m].rdev,
1981                                          conf->mddev);
1982                 }
1983         if (test_bit(R1BIO_WriteError, &r1_bio->state))
1984                 close_write(r1_bio);
1985         raid_end_bio_io(r1_bio);
1986 }
1987
1988 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
1989 {
1990         int disk;
1991         int max_sectors;
1992         struct mddev *mddev = conf->mddev;
1993         struct bio *bio;
1994         char b[BDEVNAME_SIZE];
1995         struct md_rdev *rdev;
1996
1997         clear_bit(R1BIO_ReadError, &r1_bio->state);
1998         /* we got a read error. Maybe the drive is bad.  Maybe just
1999          * the block and we can fix it.
2000          * We freeze all other IO, and try reading the block from
2001          * other devices.  When we find one, we re-write
2002          * and check it that fixes the read error.
2003          * This is all done synchronously while the array is
2004          * frozen
2005          */
2006         if (mddev->ro == 0) {
2007                 freeze_array(conf);
2008                 fix_read_error(conf, r1_bio->read_disk,
2009                                r1_bio->sector, r1_bio->sectors);
2010                 unfreeze_array(conf);
2011         } else
2012                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2013
2014         bio = r1_bio->bios[r1_bio->read_disk];
2015         bdevname(bio->bi_bdev, b);
2016 read_more:
2017         disk = read_balance(conf, r1_bio, &max_sectors);
2018         if (disk == -1) {
2019                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2020                        " read error for block %llu\n",
2021                        mdname(mddev), b, (unsigned long long)r1_bio->sector);
2022                 raid_end_bio_io(r1_bio);
2023         } else {
2024                 const unsigned long do_sync
2025                         = r1_bio->master_bio->bi_rw & REQ_SYNC;
2026                 if (bio) {
2027                         r1_bio->bios[r1_bio->read_disk] =
2028                                 mddev->ro ? IO_BLOCKED : NULL;
2029                         bio_put(bio);
2030                 }
2031                 r1_bio->read_disk = disk;
2032                 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2033                 md_trim_bio(bio, r1_bio->sector - bio->bi_sector, max_sectors);
2034                 r1_bio->bios[r1_bio->read_disk] = bio;
2035                 rdev = conf->mirrors[disk].rdev;
2036                 printk_ratelimited(KERN_ERR
2037                                    "md/raid1:%s: redirecting sector %llu"
2038                                    " to other mirror: %s\n",
2039                                    mdname(mddev),
2040                                    (unsigned long long)r1_bio->sector,
2041                                    bdevname(rdev->bdev, b));
2042                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
2043                 bio->bi_bdev = rdev->bdev;
2044                 bio->bi_end_io = raid1_end_read_request;
2045                 bio->bi_rw = READ | do_sync;
2046                 bio->bi_private = r1_bio;
2047                 if (max_sectors < r1_bio->sectors) {
2048                         /* Drat - have to split this up more */
2049                         struct bio *mbio = r1_bio->master_bio;
2050                         int sectors_handled = (r1_bio->sector + max_sectors
2051                                                - mbio->bi_sector);
2052                         r1_bio->sectors = max_sectors;
2053                         spin_lock_irq(&conf->device_lock);
2054                         if (mbio->bi_phys_segments == 0)
2055                                 mbio->bi_phys_segments = 2;
2056                         else
2057                                 mbio->bi_phys_segments++;
2058                         spin_unlock_irq(&conf->device_lock);
2059                         generic_make_request(bio);
2060                         bio = NULL;
2061
2062                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2063
2064                         r1_bio->master_bio = mbio;
2065                         r1_bio->sectors = (mbio->bi_size >> 9)
2066                                           - sectors_handled;
2067                         r1_bio->state = 0;
2068                         set_bit(R1BIO_ReadError, &r1_bio->state);
2069                         r1_bio->mddev = mddev;
2070                         r1_bio->sector = mbio->bi_sector + sectors_handled;
2071
2072                         goto read_more;
2073                 } else
2074                         generic_make_request(bio);
2075         }
2076 }
2077
2078 static void raid1d(struct mddev *mddev)
2079 {
2080         struct r1bio *r1_bio;
2081         unsigned long flags;
2082         struct r1conf *conf = mddev->private;
2083         struct list_head *head = &conf->retry_list;
2084         struct blk_plug plug;
2085
2086         md_check_recovery(mddev);
2087
2088         blk_start_plug(&plug);
2089         for (;;) {
2090
2091                 if (atomic_read(&mddev->plug_cnt) == 0)
2092                         flush_pending_writes(conf);
2093
2094                 spin_lock_irqsave(&conf->device_lock, flags);
2095                 if (list_empty(head)) {
2096                         spin_unlock_irqrestore(&conf->device_lock, flags);
2097                         break;
2098                 }
2099                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2100                 list_del(head->prev);
2101                 conf->nr_queued--;
2102                 spin_unlock_irqrestore(&conf->device_lock, flags);
2103
2104                 mddev = r1_bio->mddev;
2105                 conf = mddev->private;
2106                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2107                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2108                             test_bit(R1BIO_WriteError, &r1_bio->state))
2109                                 handle_sync_write_finished(conf, r1_bio);
2110                         else
2111                                 sync_request_write(mddev, r1_bio);
2112                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2113                            test_bit(R1BIO_WriteError, &r1_bio->state))
2114                         handle_write_finished(conf, r1_bio);
2115                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2116                         handle_read_error(conf, r1_bio);
2117                 else
2118                         /* just a partial read to be scheduled from separate
2119                          * context
2120                          */
2121                         generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2122
2123                 cond_resched();
2124                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2125                         md_check_recovery(mddev);
2126         }
2127         blk_finish_plug(&plug);
2128 }
2129
2130
2131 static int init_resync(struct r1conf *conf)
2132 {
2133         int buffs;
2134
2135         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2136         BUG_ON(conf->r1buf_pool);
2137         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2138                                           conf->poolinfo);
2139         if (!conf->r1buf_pool)
2140                 return -ENOMEM;
2141         conf->next_resync = 0;
2142         return 0;
2143 }
2144
2145 /*
2146  * perform a "sync" on one "block"
2147  *
2148  * We need to make sure that no normal I/O request - particularly write
2149  * requests - conflict with active sync requests.
2150  *
2151  * This is achieved by tracking pending requests and a 'barrier' concept
2152  * that can be installed to exclude normal IO requests.
2153  */
2154
2155 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2156 {
2157         struct r1conf *conf = mddev->private;
2158         struct r1bio *r1_bio;
2159         struct bio *bio;
2160         sector_t max_sector, nr_sectors;
2161         int disk = -1;
2162         int i;
2163         int wonly = -1;
2164         int write_targets = 0, read_targets = 0;
2165         sector_t sync_blocks;
2166         int still_degraded = 0;
2167         int good_sectors = RESYNC_SECTORS;
2168         int min_bad = 0; /* number of sectors that are bad in all devices */
2169
2170         if (!conf->r1buf_pool)
2171                 if (init_resync(conf))
2172                         return 0;
2173
2174         max_sector = mddev->dev_sectors;
2175         if (sector_nr >= max_sector) {
2176                 /* If we aborted, we need to abort the
2177                  * sync on the 'current' bitmap chunk (there will
2178                  * only be one in raid1 resync.
2179                  * We can find the current addess in mddev->curr_resync
2180                  */
2181                 if (mddev->curr_resync < max_sector) /* aborted */
2182                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2183                                                 &sync_blocks, 1);
2184                 else /* completed sync */
2185                         conf->fullsync = 0;
2186
2187                 bitmap_close_sync(mddev->bitmap);
2188                 close_sync(conf);
2189                 return 0;
2190         }
2191
2192         if (mddev->bitmap == NULL &&
2193             mddev->recovery_cp == MaxSector &&
2194             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2195             conf->fullsync == 0) {
2196                 *skipped = 1;
2197                 return max_sector - sector_nr;
2198         }
2199         /* before building a request, check if we can skip these blocks..
2200          * This call the bitmap_start_sync doesn't actually record anything
2201          */
2202         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2203             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2204                 /* We can skip this block, and probably several more */
2205                 *skipped = 1;
2206                 return sync_blocks;
2207         }
2208         /*
2209          * If there is non-resync activity waiting for a turn,
2210          * and resync is going fast enough,
2211          * then let it though before starting on this new sync request.
2212          */
2213         if (!go_faster && conf->nr_waiting)
2214                 msleep_interruptible(1000);
2215
2216         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2217         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2218         raise_barrier(conf);
2219
2220         conf->next_resync = sector_nr;
2221
2222         rcu_read_lock();
2223         /*
2224          * If we get a correctably read error during resync or recovery,
2225          * we might want to read from a different device.  So we
2226          * flag all drives that could conceivably be read from for READ,
2227          * and any others (which will be non-In_sync devices) for WRITE.
2228          * If a read fails, we try reading from something else for which READ
2229          * is OK.
2230          */
2231
2232         r1_bio->mddev = mddev;
2233         r1_bio->sector = sector_nr;
2234         r1_bio->state = 0;
2235         set_bit(R1BIO_IsSync, &r1_bio->state);
2236
2237         for (i = 0; i < conf->raid_disks * 2; i++) {
2238                 struct md_rdev *rdev;
2239                 bio = r1_bio->bios[i];
2240
2241                 /* take from bio_init */
2242                 bio->bi_next = NULL;
2243                 bio->bi_flags &= ~(BIO_POOL_MASK-1);
2244                 bio->bi_flags |= 1 << BIO_UPTODATE;
2245                 bio->bi_rw = READ;
2246                 bio->bi_vcnt = 0;
2247                 bio->bi_idx = 0;
2248                 bio->bi_phys_segments = 0;
2249                 bio->bi_size = 0;
2250                 bio->bi_end_io = NULL;
2251                 bio->bi_private = NULL;
2252
2253                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2254                 if (rdev == NULL ||
2255                     test_bit(Faulty, &rdev->flags)) {
2256                         if (i < conf->raid_disks)
2257                                 still_degraded = 1;
2258                 } else if (!test_bit(In_sync, &rdev->flags)) {
2259                         bio->bi_rw = WRITE;
2260                         bio->bi_end_io = end_sync_write;
2261                         write_targets ++;
2262                 } else {
2263                         /* may need to read from here */
2264                         sector_t first_bad = MaxSector;
2265                         int bad_sectors;
2266
2267                         if (is_badblock(rdev, sector_nr, good_sectors,
2268                                         &first_bad, &bad_sectors)) {
2269                                 if (first_bad > sector_nr)
2270                                         good_sectors = first_bad - sector_nr;
2271                                 else {
2272                                         bad_sectors -= (sector_nr - first_bad);
2273                                         if (min_bad == 0 ||
2274                                             min_bad > bad_sectors)
2275                                                 min_bad = bad_sectors;
2276                                 }
2277                         }
2278                         if (sector_nr < first_bad) {
2279                                 if (test_bit(WriteMostly, &rdev->flags)) {
2280                                         if (wonly < 0)
2281                                                 wonly = i;
2282                                 } else {
2283                                         if (disk < 0)
2284                                                 disk = i;
2285                                 }
2286                                 bio->bi_rw = READ;
2287                                 bio->bi_end_io = end_sync_read;
2288                                 read_targets++;
2289                         }
2290                 }
2291                 if (bio->bi_end_io) {
2292                         atomic_inc(&rdev->nr_pending);
2293                         bio->bi_sector = sector_nr + rdev->data_offset;
2294                         bio->bi_bdev = rdev->bdev;
2295                         bio->bi_private = r1_bio;
2296                 }
2297         }
2298         rcu_read_unlock();
2299         if (disk < 0)
2300                 disk = wonly;
2301         r1_bio->read_disk = disk;
2302
2303         if (read_targets == 0 && min_bad > 0) {
2304                 /* These sectors are bad on all InSync devices, so we
2305                  * need to mark them bad on all write targets
2306                  */
2307                 int ok = 1;
2308                 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2309                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2310                                 struct md_rdev *rdev =
2311                                         rcu_dereference(conf->mirrors[i].rdev);
2312                                 ok = rdev_set_badblocks(rdev, sector_nr,
2313                                                         min_bad, 0
2314                                         ) && ok;
2315                         }
2316                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2317                 *skipped = 1;
2318                 put_buf(r1_bio);
2319
2320                 if (!ok) {
2321                         /* Cannot record the badblocks, so need to
2322                          * abort the resync.
2323                          * If there are multiple read targets, could just
2324                          * fail the really bad ones ???
2325                          */
2326                         conf->recovery_disabled = mddev->recovery_disabled;
2327                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2328                         return 0;
2329                 } else
2330                         return min_bad;
2331
2332         }
2333         if (min_bad > 0 && min_bad < good_sectors) {
2334                 /* only resync enough to reach the next bad->good
2335                  * transition */
2336                 good_sectors = min_bad;
2337         }
2338
2339         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2340                 /* extra read targets are also write targets */
2341                 write_targets += read_targets-1;
2342
2343         if (write_targets == 0 || read_targets == 0) {
2344                 /* There is nowhere to write, so all non-sync
2345                  * drives must be failed - so we are finished
2346                  */
2347                 sector_t rv = max_sector - sector_nr;
2348                 *skipped = 1;
2349                 put_buf(r1_bio);
2350                 return rv;
2351         }
2352
2353         if (max_sector > mddev->resync_max)
2354                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2355         if (max_sector > sector_nr + good_sectors)
2356                 max_sector = sector_nr + good_sectors;
2357         nr_sectors = 0;
2358         sync_blocks = 0;
2359         do {
2360                 struct page *page;
2361                 int len = PAGE_SIZE;
2362                 if (sector_nr + (len>>9) > max_sector)
2363                         len = (max_sector - sector_nr) << 9;
2364                 if (len == 0)
2365                         break;
2366                 if (sync_blocks == 0) {
2367                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2368                                                &sync_blocks, still_degraded) &&
2369                             !conf->fullsync &&
2370                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2371                                 break;
2372                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2373                         if ((len >> 9) > sync_blocks)
2374                                 len = sync_blocks<<9;
2375                 }
2376
2377                 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2378                         bio = r1_bio->bios[i];
2379                         if (bio->bi_end_io) {
2380                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2381                                 if (bio_add_page(bio, page, len, 0) == 0) {
2382                                         /* stop here */
2383                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2384                                         while (i > 0) {
2385                                                 i--;
2386                                                 bio = r1_bio->bios[i];
2387                                                 if (bio->bi_end_io==NULL)
2388                                                         continue;
2389                                                 /* remove last page from this bio */
2390                                                 bio->bi_vcnt--;
2391                                                 bio->bi_size -= len;
2392                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2393                                         }
2394                                         goto bio_full;
2395                                 }
2396                         }
2397                 }
2398                 nr_sectors += len>>9;
2399                 sector_nr += len>>9;
2400                 sync_blocks -= (len>>9);
2401         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2402  bio_full:
2403         r1_bio->sectors = nr_sectors;
2404
2405         /* For a user-requested sync, we read all readable devices and do a
2406          * compare
2407          */
2408         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2409                 atomic_set(&r1_bio->remaining, read_targets);
2410                 for (i = 0; i < conf->raid_disks * 2; i++) {
2411                         bio = r1_bio->bios[i];
2412                         if (bio->bi_end_io == end_sync_read) {
2413                                 md_sync_acct(bio->bi_bdev, nr_sectors);
2414                                 generic_make_request(bio);
2415                         }
2416                 }
2417         } else {
2418                 atomic_set(&r1_bio->remaining, 1);
2419                 bio = r1_bio->bios[r1_bio->read_disk];
2420                 md_sync_acct(bio->bi_bdev, nr_sectors);
2421                 generic_make_request(bio);
2422
2423         }
2424         return nr_sectors;
2425 }
2426
2427 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2428 {
2429         if (sectors)
2430                 return sectors;
2431
2432         return mddev->dev_sectors;
2433 }
2434
2435 static struct r1conf *setup_conf(struct mddev *mddev)
2436 {
2437         struct r1conf *conf;
2438         int i;
2439         struct mirror_info *disk;
2440         struct md_rdev *rdev;
2441         int err = -ENOMEM;
2442
2443         conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2444         if (!conf)
2445                 goto abort;
2446
2447         conf->mirrors = kzalloc(sizeof(struct mirror_info)
2448                                 * mddev->raid_disks * 2,
2449                                  GFP_KERNEL);
2450         if (!conf->mirrors)
2451                 goto abort;
2452
2453         conf->tmppage = alloc_page(GFP_KERNEL);
2454         if (!conf->tmppage)
2455                 goto abort;
2456
2457         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2458         if (!conf->poolinfo)
2459                 goto abort;
2460         conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2461         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2462                                           r1bio_pool_free,
2463                                           conf->poolinfo);
2464         if (!conf->r1bio_pool)
2465                 goto abort;
2466
2467         conf->poolinfo->mddev = mddev;
2468
2469         err = -EINVAL;
2470         spin_lock_init(&conf->device_lock);
2471         list_for_each_entry(rdev, &mddev->disks, same_set) {
2472                 int disk_idx = rdev->raid_disk;
2473                 if (disk_idx >= mddev->raid_disks
2474                     || disk_idx < 0)
2475                         continue;
2476                 if (test_bit(Replacement, &rdev->flags))
2477                         disk = conf->mirrors + conf->raid_disks + disk_idx;
2478                 else
2479                         disk = conf->mirrors + disk_idx;
2480
2481                 if (disk->rdev)
2482                         goto abort;
2483                 disk->rdev = rdev;
2484
2485                 disk->head_position = 0;
2486         }
2487         conf->raid_disks = mddev->raid_disks;
2488         conf->mddev = mddev;
2489         INIT_LIST_HEAD(&conf->retry_list);
2490
2491         spin_lock_init(&conf->resync_lock);
2492         init_waitqueue_head(&conf->wait_barrier);
2493
2494         bio_list_init(&conf->pending_bio_list);
2495         conf->pending_count = 0;
2496         conf->recovery_disabled = mddev->recovery_disabled - 1;
2497
2498         err = -EIO;
2499         conf->last_used = -1;
2500         for (i = 0; i < conf->raid_disks * 2; i++) {
2501
2502                 disk = conf->mirrors + i;
2503
2504                 if (i < conf->raid_disks &&
2505                     disk[conf->raid_disks].rdev) {
2506                         /* This slot has a replacement. */
2507                         if (!disk->rdev) {
2508                                 /* No original, just make the replacement
2509                                  * a recovering spare
2510                                  */
2511                                 disk->rdev =
2512                                         disk[conf->raid_disks].rdev;
2513                                 disk[conf->raid_disks].rdev = NULL;
2514                         } else if (!test_bit(In_sync, &disk->rdev->flags))
2515                                 /* Original is not in_sync - bad */
2516                                 goto abort;
2517                 }
2518
2519                 if (!disk->rdev ||
2520                     !test_bit(In_sync, &disk->rdev->flags)) {
2521                         disk->head_position = 0;
2522                         if (disk->rdev)
2523                                 conf->fullsync = 1;
2524                 } else if (conf->last_used < 0)
2525                         /*
2526                          * The first working device is used as a
2527                          * starting point to read balancing.
2528                          */
2529                         conf->last_used = i;
2530         }
2531
2532         if (conf->last_used < 0) {
2533                 printk(KERN_ERR "md/raid1:%s: no operational mirrors\n",
2534                        mdname(mddev));
2535                 goto abort;
2536         }
2537         err = -ENOMEM;
2538         conf->thread = md_register_thread(raid1d, mddev, NULL);
2539         if (!conf->thread) {
2540                 printk(KERN_ERR
2541                        "md/raid1:%s: couldn't allocate thread\n",
2542                        mdname(mddev));
2543                 goto abort;
2544         }
2545
2546         return conf;
2547
2548  abort:
2549         if (conf) {
2550                 if (conf->r1bio_pool)
2551                         mempool_destroy(conf->r1bio_pool);
2552                 kfree(conf->mirrors);
2553                 safe_put_page(conf->tmppage);
2554                 kfree(conf->poolinfo);
2555                 kfree(conf);
2556         }
2557         return ERR_PTR(err);
2558 }
2559
2560 static int run(struct mddev *mddev)
2561 {
2562         struct r1conf *conf;
2563         int i;
2564         struct md_rdev *rdev;
2565
2566         if (mddev->level != 1) {
2567                 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2568                        mdname(mddev), mddev->level);
2569                 return -EIO;
2570         }
2571         if (mddev->reshape_position != MaxSector) {
2572                 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2573                        mdname(mddev));
2574                 return -EIO;
2575         }
2576         /*
2577          * copy the already verified devices into our private RAID1
2578          * bookkeeping area. [whatever we allocate in run(),
2579          * should be freed in stop()]
2580          */
2581         if (mddev->private == NULL)
2582                 conf = setup_conf(mddev);
2583         else
2584                 conf = mddev->private;
2585
2586         if (IS_ERR(conf))
2587                 return PTR_ERR(conf);
2588
2589         list_for_each_entry(rdev, &mddev->disks, same_set) {
2590                 if (!mddev->gendisk)
2591                         continue;
2592                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2593                                   rdev->data_offset << 9);
2594                 /* as we don't honour merge_bvec_fn, we must never risk
2595                  * violating it, so limit ->max_segments to 1 lying within
2596                  * a single page, as a one page request is never in violation.
2597                  */
2598                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2599                         blk_queue_max_segments(mddev->queue, 1);
2600                         blk_queue_segment_boundary(mddev->queue,
2601                                                    PAGE_CACHE_SIZE - 1);
2602                 }
2603         }
2604
2605         mddev->degraded = 0;
2606         for (i=0; i < conf->raid_disks; i++)
2607                 if (conf->mirrors[i].rdev == NULL ||
2608                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2609                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2610                         mddev->degraded++;
2611
2612         if (conf->raid_disks - mddev->degraded == 1)
2613                 mddev->recovery_cp = MaxSector;
2614
2615         if (mddev->recovery_cp != MaxSector)
2616                 printk(KERN_NOTICE "md/raid1:%s: not clean"
2617                        " -- starting background reconstruction\n",
2618                        mdname(mddev));
2619         printk(KERN_INFO 
2620                 "md/raid1:%s: active with %d out of %d mirrors\n",
2621                 mdname(mddev), mddev->raid_disks - mddev->degraded, 
2622                 mddev->raid_disks);
2623
2624         /*
2625          * Ok, everything is just fine now
2626          */
2627         mddev->thread = conf->thread;
2628         conf->thread = NULL;
2629         mddev->private = conf;
2630
2631         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2632
2633         if (mddev->queue) {
2634                 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2635                 mddev->queue->backing_dev_info.congested_data = mddev;
2636         }
2637         return md_integrity_register(mddev);
2638 }
2639
2640 static int stop(struct mddev *mddev)
2641 {
2642         struct r1conf *conf = mddev->private;
2643         struct bitmap *bitmap = mddev->bitmap;
2644
2645         /* wait for behind writes to complete */
2646         if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2647                 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2648                        mdname(mddev));
2649                 /* need to kick something here to make sure I/O goes? */
2650                 wait_event(bitmap->behind_wait,
2651                            atomic_read(&bitmap->behind_writes) == 0);
2652         }
2653
2654         raise_barrier(conf);
2655         lower_barrier(conf);
2656
2657         md_unregister_thread(&mddev->thread);
2658         if (conf->r1bio_pool)
2659                 mempool_destroy(conf->r1bio_pool);
2660         kfree(conf->mirrors);
2661         kfree(conf->poolinfo);
2662         kfree(conf);
2663         mddev->private = NULL;
2664         return 0;
2665 }
2666
2667 static int raid1_resize(struct mddev *mddev, sector_t sectors)
2668 {
2669         /* no resync is happening, and there is enough space
2670          * on all devices, so we can resize.
2671          * We need to make sure resync covers any new space.
2672          * If the array is shrinking we should possibly wait until
2673          * any io in the removed space completes, but it hardly seems
2674          * worth it.
2675          */
2676         md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2677         if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2678                 return -EINVAL;
2679         set_capacity(mddev->gendisk, mddev->array_sectors);
2680         revalidate_disk(mddev->gendisk);
2681         if (sectors > mddev->dev_sectors &&
2682             mddev->recovery_cp > mddev->dev_sectors) {
2683                 mddev->recovery_cp = mddev->dev_sectors;
2684                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2685         }
2686         mddev->dev_sectors = sectors;
2687         mddev->resync_max_sectors = sectors;
2688         return 0;
2689 }
2690
2691 static int raid1_reshape(struct mddev *mddev)
2692 {
2693         /* We need to:
2694          * 1/ resize the r1bio_pool
2695          * 2/ resize conf->mirrors
2696          *
2697          * We allocate a new r1bio_pool if we can.
2698          * Then raise a device barrier and wait until all IO stops.
2699          * Then resize conf->mirrors and swap in the new r1bio pool.
2700          *
2701          * At the same time, we "pack" the devices so that all the missing
2702          * devices have the higher raid_disk numbers.
2703          */
2704         mempool_t *newpool, *oldpool;
2705         struct pool_info *newpoolinfo;
2706         struct mirror_info *newmirrors;
2707         struct r1conf *conf = mddev->private;
2708         int cnt, raid_disks;
2709         unsigned long flags;
2710         int d, d2, err;
2711
2712         /* Cannot change chunk_size, layout, or level */
2713         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2714             mddev->layout != mddev->new_layout ||
2715             mddev->level != mddev->new_level) {
2716                 mddev->new_chunk_sectors = mddev->chunk_sectors;
2717                 mddev->new_layout = mddev->layout;
2718                 mddev->new_level = mddev->level;
2719                 return -EINVAL;
2720         }
2721
2722         err = md_allow_write(mddev);
2723         if (err)
2724                 return err;
2725
2726         raid_disks = mddev->raid_disks + mddev->delta_disks;
2727
2728         if (raid_disks < conf->raid_disks) {
2729                 cnt=0;
2730                 for (d= 0; d < conf->raid_disks; d++)
2731                         if (conf->mirrors[d].rdev)
2732                                 cnt++;
2733                 if (cnt > raid_disks)
2734                         return -EBUSY;
2735         }
2736
2737         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2738         if (!newpoolinfo)
2739                 return -ENOMEM;
2740         newpoolinfo->mddev = mddev;
2741         newpoolinfo->raid_disks = raid_disks * 2;
2742
2743         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2744                                  r1bio_pool_free, newpoolinfo);
2745         if (!newpool) {
2746                 kfree(newpoolinfo);
2747                 return -ENOMEM;
2748         }
2749         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks * 2,
2750                              GFP_KERNEL);
2751         if (!newmirrors) {
2752                 kfree(newpoolinfo);
2753                 mempool_destroy(newpool);
2754                 return -ENOMEM;
2755         }
2756
2757         raise_barrier(conf);
2758
2759         /* ok, everything is stopped */
2760         oldpool = conf->r1bio_pool;
2761         conf->r1bio_pool = newpool;
2762
2763         for (d = d2 = 0; d < conf->raid_disks; d++) {
2764                 struct md_rdev *rdev = conf->mirrors[d].rdev;
2765                 if (rdev && rdev->raid_disk != d2) {
2766                         sysfs_unlink_rdev(mddev, rdev);
2767                         rdev->raid_disk = d2;
2768                         sysfs_unlink_rdev(mddev, rdev);
2769                         if (sysfs_link_rdev(mddev, rdev))
2770                                 printk(KERN_WARNING
2771                                        "md/raid1:%s: cannot register rd%d\n",
2772                                        mdname(mddev), rdev->raid_disk);
2773                 }
2774                 if (rdev)
2775                         newmirrors[d2++].rdev = rdev;
2776         }
2777         kfree(conf->mirrors);
2778         conf->mirrors = newmirrors;
2779         kfree(conf->poolinfo);
2780         conf->poolinfo = newpoolinfo;
2781
2782         spin_lock_irqsave(&conf->device_lock, flags);
2783         mddev->degraded += (raid_disks - conf->raid_disks);
2784         spin_unlock_irqrestore(&conf->device_lock, flags);
2785         conf->raid_disks = mddev->raid_disks = raid_disks;
2786         mddev->delta_disks = 0;
2787
2788         conf->last_used = 0; /* just make sure it is in-range */
2789         lower_barrier(conf);
2790
2791         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2792         md_wakeup_thread(mddev->thread);
2793
2794         mempool_destroy(oldpool);
2795         return 0;
2796 }
2797
2798 static void raid1_quiesce(struct mddev *mddev, int state)
2799 {
2800         struct r1conf *conf = mddev->private;
2801
2802         switch(state) {
2803         case 2: /* wake for suspend */
2804                 wake_up(&conf->wait_barrier);
2805                 break;
2806         case 1:
2807                 raise_barrier(conf);
2808                 break;
2809         case 0:
2810                 lower_barrier(conf);
2811                 break;
2812         }
2813 }
2814
2815 static void *raid1_takeover(struct mddev *mddev)
2816 {
2817         /* raid1 can take over:
2818          *  raid5 with 2 devices, any layout or chunk size
2819          */
2820         if (mddev->level == 5 && mddev->raid_disks == 2) {
2821                 struct r1conf *conf;
2822                 mddev->new_level = 1;
2823                 mddev->new_layout = 0;
2824                 mddev->new_chunk_sectors = 0;
2825                 conf = setup_conf(mddev);
2826                 if (!IS_ERR(conf))
2827                         conf->barrier = 1;
2828                 return conf;
2829         }
2830         return ERR_PTR(-EINVAL);
2831 }
2832
2833 static struct md_personality raid1_personality =
2834 {
2835         .name           = "raid1",
2836         .level          = 1,
2837         .owner          = THIS_MODULE,
2838         .make_request   = make_request,
2839         .run            = run,
2840         .stop           = stop,
2841         .status         = status,
2842         .error_handler  = error,
2843         .hot_add_disk   = raid1_add_disk,
2844         .hot_remove_disk= raid1_remove_disk,
2845         .spare_active   = raid1_spare_active,
2846         .sync_request   = sync_request,
2847         .resize         = raid1_resize,
2848         .size           = raid1_size,
2849         .check_reshape  = raid1_reshape,
2850         .quiesce        = raid1_quiesce,
2851         .takeover       = raid1_takeover,
2852 };
2853
2854 static int __init raid_init(void)
2855 {
2856         return register_md_personality(&raid1_personality);
2857 }
2858
2859 static void raid_exit(void)
2860 {
2861         unregister_md_personality(&raid1_personality);
2862 }
2863
2864 module_init(raid_init);
2865 module_exit(raid_exit);
2866 MODULE_LICENSE("GPL");
2867 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2868 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2869 MODULE_ALIAS("md-raid1");
2870 MODULE_ALIAS("md-level-1");
2871
2872 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);