]> git.openfabrics.org - ~shefty/rdma-dev.git/blob - drivers/md/raid1.c
Merge branch 'for-3.2/core' of git://git.kernel.dk/linux-block
[~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/seq_file.h>
38 #include <linux/ratelimit.h>
39 #include "md.h"
40 #include "raid1.h"
41 #include "bitmap.h"
42
43 /*
44  * Number of guaranteed r1bios in case of extreme VM load:
45  */
46 #define NR_RAID1_BIOS 256
47
48 /* When there are this many requests queue to be written by
49  * the raid1 thread, we become 'congested' to provide back-pressure
50  * for writeback.
51  */
52 static int max_queued_requests = 1024;
53
54 static void allow_barrier(struct r1conf *conf);
55 static void lower_barrier(struct r1conf *conf);
56
57 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
58 {
59         struct pool_info *pi = data;
60         int size = offsetof(struct r1bio, bios[pi->raid_disks]);
61
62         /* allocate a r1bio with room for raid_disks entries in the bios array */
63         return kzalloc(size, gfp_flags);
64 }
65
66 static void r1bio_pool_free(void *r1_bio, void *data)
67 {
68         kfree(r1_bio);
69 }
70
71 #define RESYNC_BLOCK_SIZE (64*1024)
72 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
73 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
74 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
75 #define RESYNC_WINDOW (2048*1024)
76
77 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
78 {
79         struct pool_info *pi = data;
80         struct page *page;
81         struct r1bio *r1_bio;
82         struct bio *bio;
83         int i, j;
84
85         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
86         if (!r1_bio)
87                 return NULL;
88
89         /*
90          * Allocate bios : 1 for reading, n-1 for writing
91          */
92         for (j = pi->raid_disks ; j-- ; ) {
93                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
94                 if (!bio)
95                         goto out_free_bio;
96                 r1_bio->bios[j] = bio;
97         }
98         /*
99          * Allocate RESYNC_PAGES data pages and attach them to
100          * the first bio.
101          * If this is a user-requested check/repair, allocate
102          * RESYNC_PAGES for each bio.
103          */
104         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
105                 j = pi->raid_disks;
106         else
107                 j = 1;
108         while(j--) {
109                 bio = r1_bio->bios[j];
110                 for (i = 0; i < RESYNC_PAGES; i++) {
111                         page = alloc_page(gfp_flags);
112                         if (unlikely(!page))
113                                 goto out_free_pages;
114
115                         bio->bi_io_vec[i].bv_page = page;
116                         bio->bi_vcnt = i+1;
117                 }
118         }
119         /* If not user-requests, copy the page pointers to all bios */
120         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
121                 for (i=0; i<RESYNC_PAGES ; i++)
122                         for (j=1; j<pi->raid_disks; j++)
123                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
124                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
125         }
126
127         r1_bio->master_bio = NULL;
128
129         return r1_bio;
130
131 out_free_pages:
132         for (j=0 ; j < pi->raid_disks; j++)
133                 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
134                         put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
135         j = -1;
136 out_free_bio:
137         while ( ++j < pi->raid_disks )
138                 bio_put(r1_bio->bios[j]);
139         r1bio_pool_free(r1_bio, data);
140         return NULL;
141 }
142
143 static void r1buf_pool_free(void *__r1_bio, void *data)
144 {
145         struct pool_info *pi = data;
146         int i,j;
147         struct r1bio *r1bio = __r1_bio;
148
149         for (i = 0; i < RESYNC_PAGES; i++)
150                 for (j = pi->raid_disks; j-- ;) {
151                         if (j == 0 ||
152                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
153                             r1bio->bios[0]->bi_io_vec[i].bv_page)
154                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
155                 }
156         for (i=0 ; i < pi->raid_disks; i++)
157                 bio_put(r1bio->bios[i]);
158
159         r1bio_pool_free(r1bio, data);
160 }
161
162 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
163 {
164         int i;
165
166         for (i = 0; i < conf->raid_disks; i++) {
167                 struct bio **bio = r1_bio->bios + i;
168                 if (!BIO_SPECIAL(*bio))
169                         bio_put(*bio);
170                 *bio = NULL;
171         }
172 }
173
174 static void free_r1bio(struct r1bio *r1_bio)
175 {
176         struct r1conf *conf = r1_bio->mddev->private;
177
178         put_all_bios(conf, r1_bio);
179         mempool_free(r1_bio, conf->r1bio_pool);
180 }
181
182 static void put_buf(struct r1bio *r1_bio)
183 {
184         struct r1conf *conf = r1_bio->mddev->private;
185         int i;
186
187         for (i=0; i<conf->raid_disks; i++) {
188                 struct bio *bio = r1_bio->bios[i];
189                 if (bio->bi_end_io)
190                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
191         }
192
193         mempool_free(r1_bio, conf->r1buf_pool);
194
195         lower_barrier(conf);
196 }
197
198 static void reschedule_retry(struct r1bio *r1_bio)
199 {
200         unsigned long flags;
201         struct mddev *mddev = r1_bio->mddev;
202         struct r1conf *conf = mddev->private;
203
204         spin_lock_irqsave(&conf->device_lock, flags);
205         list_add(&r1_bio->retry_list, &conf->retry_list);
206         conf->nr_queued ++;
207         spin_unlock_irqrestore(&conf->device_lock, flags);
208
209         wake_up(&conf->wait_barrier);
210         md_wakeup_thread(mddev->thread);
211 }
212
213 /*
214  * raid_end_bio_io() is called when we have finished servicing a mirrored
215  * operation and are ready to return a success/failure code to the buffer
216  * cache layer.
217  */
218 static void call_bio_endio(struct r1bio *r1_bio)
219 {
220         struct bio *bio = r1_bio->master_bio;
221         int done;
222         struct r1conf *conf = r1_bio->mddev->private;
223
224         if (bio->bi_phys_segments) {
225                 unsigned long flags;
226                 spin_lock_irqsave(&conf->device_lock, flags);
227                 bio->bi_phys_segments--;
228                 done = (bio->bi_phys_segments == 0);
229                 spin_unlock_irqrestore(&conf->device_lock, flags);
230         } else
231                 done = 1;
232
233         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
234                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
235         if (done) {
236                 bio_endio(bio, 0);
237                 /*
238                  * Wake up any possible resync thread that waits for the device
239                  * to go idle.
240                  */
241                 allow_barrier(conf);
242         }
243 }
244
245 static void raid_end_bio_io(struct r1bio *r1_bio)
246 {
247         struct bio *bio = r1_bio->master_bio;
248
249         /* if nobody has done the final endio yet, do it now */
250         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
251                 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
252                          (bio_data_dir(bio) == WRITE) ? "write" : "read",
253                          (unsigned long long) bio->bi_sector,
254                          (unsigned long long) bio->bi_sector +
255                          (bio->bi_size >> 9) - 1);
256
257                 call_bio_endio(r1_bio);
258         }
259         free_r1bio(r1_bio);
260 }
261
262 /*
263  * Update disk head position estimator based on IRQ completion info.
264  */
265 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
266 {
267         struct r1conf *conf = r1_bio->mddev->private;
268
269         conf->mirrors[disk].head_position =
270                 r1_bio->sector + (r1_bio->sectors);
271 }
272
273 /*
274  * Find the disk number which triggered given bio
275  */
276 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
277 {
278         int mirror;
279         int raid_disks = r1_bio->mddev->raid_disks;
280
281         for (mirror = 0; mirror < raid_disks; mirror++)
282                 if (r1_bio->bios[mirror] == bio)
283                         break;
284
285         BUG_ON(mirror == raid_disks);
286         update_head_pos(mirror, r1_bio);
287
288         return mirror;
289 }
290
291 static void raid1_end_read_request(struct bio *bio, int error)
292 {
293         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
294         struct r1bio *r1_bio = bio->bi_private;
295         int mirror;
296         struct r1conf *conf = r1_bio->mddev->private;
297
298         mirror = r1_bio->read_disk;
299         /*
300          * this branch is our 'one mirror IO has finished' event handler:
301          */
302         update_head_pos(mirror, r1_bio);
303
304         if (uptodate)
305                 set_bit(R1BIO_Uptodate, &r1_bio->state);
306         else {
307                 /* If all other devices have failed, we want to return
308                  * the error upwards rather than fail the last device.
309                  * Here we redefine "uptodate" to mean "Don't want to retry"
310                  */
311                 unsigned long flags;
312                 spin_lock_irqsave(&conf->device_lock, flags);
313                 if (r1_bio->mddev->degraded == conf->raid_disks ||
314                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
315                      !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
316                         uptodate = 1;
317                 spin_unlock_irqrestore(&conf->device_lock, flags);
318         }
319
320         if (uptodate)
321                 raid_end_bio_io(r1_bio);
322         else {
323                 /*
324                  * oops, read error:
325                  */
326                 char b[BDEVNAME_SIZE];
327                 printk_ratelimited(
328                         KERN_ERR "md/raid1:%s: %s: "
329                         "rescheduling sector %llu\n",
330                         mdname(conf->mddev),
331                         bdevname(conf->mirrors[mirror].rdev->bdev,
332                                  b),
333                         (unsigned long long)r1_bio->sector);
334                 set_bit(R1BIO_ReadError, &r1_bio->state);
335                 reschedule_retry(r1_bio);
336         }
337
338         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
339 }
340
341 static void close_write(struct r1bio *r1_bio)
342 {
343         /* it really is the end of this request */
344         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
345                 /* free extra copy of the data pages */
346                 int i = r1_bio->behind_page_count;
347                 while (i--)
348                         safe_put_page(r1_bio->behind_bvecs[i].bv_page);
349                 kfree(r1_bio->behind_bvecs);
350                 r1_bio->behind_bvecs = NULL;
351         }
352         /* clear the bitmap if all writes complete successfully */
353         bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
354                         r1_bio->sectors,
355                         !test_bit(R1BIO_Degraded, &r1_bio->state),
356                         test_bit(R1BIO_BehindIO, &r1_bio->state));
357         md_write_end(r1_bio->mddev);
358 }
359
360 static void r1_bio_write_done(struct r1bio *r1_bio)
361 {
362         if (!atomic_dec_and_test(&r1_bio->remaining))
363                 return;
364
365         if (test_bit(R1BIO_WriteError, &r1_bio->state))
366                 reschedule_retry(r1_bio);
367         else {
368                 close_write(r1_bio);
369                 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
370                         reschedule_retry(r1_bio);
371                 else
372                         raid_end_bio_io(r1_bio);
373         }
374 }
375
376 static void raid1_end_write_request(struct bio *bio, int error)
377 {
378         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
379         struct r1bio *r1_bio = bio->bi_private;
380         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
381         struct r1conf *conf = r1_bio->mddev->private;
382         struct bio *to_put = NULL;
383
384         mirror = find_bio_disk(r1_bio, bio);
385
386         /*
387          * 'one mirror IO has finished' event handler:
388          */
389         if (!uptodate) {
390                 set_bit(WriteErrorSeen,
391                         &conf->mirrors[mirror].rdev->flags);
392                 set_bit(R1BIO_WriteError, &r1_bio->state);
393         } else {
394                 /*
395                  * Set R1BIO_Uptodate in our master bio, so that we
396                  * will return a good error code for to the higher
397                  * levels even if IO on some other mirrored buffer
398                  * fails.
399                  *
400                  * The 'master' represents the composite IO operation
401                  * to user-side. So if something waits for IO, then it
402                  * will wait for the 'master' bio.
403                  */
404                 sector_t first_bad;
405                 int bad_sectors;
406
407                 r1_bio->bios[mirror] = NULL;
408                 to_put = bio;
409                 set_bit(R1BIO_Uptodate, &r1_bio->state);
410
411                 /* Maybe we can clear some bad blocks. */
412                 if (is_badblock(conf->mirrors[mirror].rdev,
413                                 r1_bio->sector, r1_bio->sectors,
414                                 &first_bad, &bad_sectors)) {
415                         r1_bio->bios[mirror] = IO_MADE_GOOD;
416                         set_bit(R1BIO_MadeGood, &r1_bio->state);
417                 }
418         }
419
420         if (behind) {
421                 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
422                         atomic_dec(&r1_bio->behind_remaining);
423
424                 /*
425                  * In behind mode, we ACK the master bio once the I/O
426                  * has safely reached all non-writemostly
427                  * disks. Setting the Returned bit ensures that this
428                  * gets done only once -- we don't ever want to return
429                  * -EIO here, instead we'll wait
430                  */
431                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
432                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
433                         /* Maybe we can return now */
434                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
435                                 struct bio *mbio = r1_bio->master_bio;
436                                 pr_debug("raid1: behind end write sectors"
437                                          " %llu-%llu\n",
438                                          (unsigned long long) mbio->bi_sector,
439                                          (unsigned long long) mbio->bi_sector +
440                                          (mbio->bi_size >> 9) - 1);
441                                 call_bio_endio(r1_bio);
442                         }
443                 }
444         }
445         if (r1_bio->bios[mirror] == NULL)
446                 rdev_dec_pending(conf->mirrors[mirror].rdev,
447                                  conf->mddev);
448
449         /*
450          * Let's see if all mirrored write operations have finished
451          * already.
452          */
453         r1_bio_write_done(r1_bio);
454
455         if (to_put)
456                 bio_put(to_put);
457 }
458
459
460 /*
461  * This routine returns the disk from which the requested read should
462  * be done. There is a per-array 'next expected sequential IO' sector
463  * number - if this matches on the next IO then we use the last disk.
464  * There is also a per-disk 'last know head position' sector that is
465  * maintained from IRQ contexts, both the normal and the resync IO
466  * completion handlers update this position correctly. If there is no
467  * perfect sequential match then we pick the disk whose head is closest.
468  *
469  * If there are 2 mirrors in the same 2 devices, performance degrades
470  * because position is mirror, not device based.
471  *
472  * The rdev for the device selected will have nr_pending incremented.
473  */
474 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
475 {
476         const sector_t this_sector = r1_bio->sector;
477         int sectors;
478         int best_good_sectors;
479         int start_disk;
480         int best_disk;
481         int i;
482         sector_t best_dist;
483         struct md_rdev *rdev;
484         int choose_first;
485
486         rcu_read_lock();
487         /*
488          * Check if we can balance. We can balance on the whole
489          * device if no resync is going on, or below the resync window.
490          * We take the first readable disk when above the resync window.
491          */
492  retry:
493         sectors = r1_bio->sectors;
494         best_disk = -1;
495         best_dist = MaxSector;
496         best_good_sectors = 0;
497
498         if (conf->mddev->recovery_cp < MaxSector &&
499             (this_sector + sectors >= conf->next_resync)) {
500                 choose_first = 1;
501                 start_disk = 0;
502         } else {
503                 choose_first = 0;
504                 start_disk = conf->last_used;
505         }
506
507         for (i = 0 ; i < conf->raid_disks ; i++) {
508                 sector_t dist;
509                 sector_t first_bad;
510                 int bad_sectors;
511
512                 int disk = start_disk + i;
513                 if (disk >= conf->raid_disks)
514                         disk -= conf->raid_disks;
515
516                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
517                 if (r1_bio->bios[disk] == IO_BLOCKED
518                     || rdev == NULL
519                     || test_bit(Faulty, &rdev->flags))
520                         continue;
521                 if (!test_bit(In_sync, &rdev->flags) &&
522                     rdev->recovery_offset < this_sector + sectors)
523                         continue;
524                 if (test_bit(WriteMostly, &rdev->flags)) {
525                         /* Don't balance among write-mostly, just
526                          * use the first as a last resort */
527                         if (best_disk < 0)
528                                 best_disk = disk;
529                         continue;
530                 }
531                 /* This is a reasonable device to use.  It might
532                  * even be best.
533                  */
534                 if (is_badblock(rdev, this_sector, sectors,
535                                 &first_bad, &bad_sectors)) {
536                         if (best_dist < MaxSector)
537                                 /* already have a better device */
538                                 continue;
539                         if (first_bad <= this_sector) {
540                                 /* cannot read here. If this is the 'primary'
541                                  * device, then we must not read beyond
542                                  * bad_sectors from another device..
543                                  */
544                                 bad_sectors -= (this_sector - first_bad);
545                                 if (choose_first && sectors > bad_sectors)
546                                         sectors = bad_sectors;
547                                 if (best_good_sectors > sectors)
548                                         best_good_sectors = sectors;
549
550                         } else {
551                                 sector_t good_sectors = first_bad - this_sector;
552                                 if (good_sectors > best_good_sectors) {
553                                         best_good_sectors = good_sectors;
554                                         best_disk = disk;
555                                 }
556                                 if (choose_first)
557                                         break;
558                         }
559                         continue;
560                 } else
561                         best_good_sectors = sectors;
562
563                 dist = abs(this_sector - conf->mirrors[disk].head_position);
564                 if (choose_first
565                     /* Don't change to another disk for sequential reads */
566                     || conf->next_seq_sect == this_sector
567                     || dist == 0
568                     /* If device is idle, use it */
569                     || atomic_read(&rdev->nr_pending) == 0) {
570                         best_disk = disk;
571                         break;
572                 }
573                 if (dist < best_dist) {
574                         best_dist = dist;
575                         best_disk = disk;
576                 }
577         }
578
579         if (best_disk >= 0) {
580                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
581                 if (!rdev)
582                         goto retry;
583                 atomic_inc(&rdev->nr_pending);
584                 if (test_bit(Faulty, &rdev->flags)) {
585                         /* cannot risk returning a device that failed
586                          * before we inc'ed nr_pending
587                          */
588                         rdev_dec_pending(rdev, conf->mddev);
589                         goto retry;
590                 }
591                 sectors = best_good_sectors;
592                 conf->next_seq_sect = this_sector + sectors;
593                 conf->last_used = best_disk;
594         }
595         rcu_read_unlock();
596         *max_sectors = sectors;
597
598         return best_disk;
599 }
600
601 int md_raid1_congested(struct mddev *mddev, int bits)
602 {
603         struct r1conf *conf = mddev->private;
604         int i, ret = 0;
605
606         if ((bits & (1 << BDI_async_congested)) &&
607             conf->pending_count >= max_queued_requests)
608                 return 1;
609
610         rcu_read_lock();
611         for (i = 0; i < mddev->raid_disks; i++) {
612                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
613                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
614                         struct request_queue *q = bdev_get_queue(rdev->bdev);
615
616                         BUG_ON(!q);
617
618                         /* Note the '|| 1' - when read_balance prefers
619                          * non-congested targets, it can be removed
620                          */
621                         if ((bits & (1<<BDI_async_congested)) || 1)
622                                 ret |= bdi_congested(&q->backing_dev_info, bits);
623                         else
624                                 ret &= bdi_congested(&q->backing_dev_info, bits);
625                 }
626         }
627         rcu_read_unlock();
628         return ret;
629 }
630 EXPORT_SYMBOL_GPL(md_raid1_congested);
631
632 static int raid1_congested(void *data, int bits)
633 {
634         struct mddev *mddev = data;
635
636         return mddev_congested(mddev, bits) ||
637                 md_raid1_congested(mddev, bits);
638 }
639
640 static void flush_pending_writes(struct r1conf *conf)
641 {
642         /* Any writes that have been queued but are awaiting
643          * bitmap updates get flushed here.
644          */
645         spin_lock_irq(&conf->device_lock);
646
647         if (conf->pending_bio_list.head) {
648                 struct bio *bio;
649                 bio = bio_list_get(&conf->pending_bio_list);
650                 conf->pending_count = 0;
651                 spin_unlock_irq(&conf->device_lock);
652                 /* flush any pending bitmap writes to
653                  * disk before proceeding w/ I/O */
654                 bitmap_unplug(conf->mddev->bitmap);
655                 wake_up(&conf->wait_barrier);
656
657                 while (bio) { /* submit pending writes */
658                         struct bio *next = bio->bi_next;
659                         bio->bi_next = NULL;
660                         generic_make_request(bio);
661                         bio = next;
662                 }
663         } else
664                 spin_unlock_irq(&conf->device_lock);
665 }
666
667 /* Barriers....
668  * Sometimes we need to suspend IO while we do something else,
669  * either some resync/recovery, or reconfigure the array.
670  * To do this we raise a 'barrier'.
671  * The 'barrier' is a counter that can be raised multiple times
672  * to count how many activities are happening which preclude
673  * normal IO.
674  * We can only raise the barrier if there is no pending IO.
675  * i.e. if nr_pending == 0.
676  * We choose only to raise the barrier if no-one is waiting for the
677  * barrier to go down.  This means that as soon as an IO request
678  * is ready, no other operations which require a barrier will start
679  * until the IO request has had a chance.
680  *
681  * So: regular IO calls 'wait_barrier'.  When that returns there
682  *    is no backgroup IO happening,  It must arrange to call
683  *    allow_barrier when it has finished its IO.
684  * backgroup IO calls must call raise_barrier.  Once that returns
685  *    there is no normal IO happeing.  It must arrange to call
686  *    lower_barrier when the particular background IO completes.
687  */
688 #define RESYNC_DEPTH 32
689
690 static void raise_barrier(struct r1conf *conf)
691 {
692         spin_lock_irq(&conf->resync_lock);
693
694         /* Wait until no block IO is waiting */
695         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
696                             conf->resync_lock, );
697
698         /* block any new IO from starting */
699         conf->barrier++;
700
701         /* Now wait for all pending IO to complete */
702         wait_event_lock_irq(conf->wait_barrier,
703                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
704                             conf->resync_lock, );
705
706         spin_unlock_irq(&conf->resync_lock);
707 }
708
709 static void lower_barrier(struct r1conf *conf)
710 {
711         unsigned long flags;
712         BUG_ON(conf->barrier <= 0);
713         spin_lock_irqsave(&conf->resync_lock, flags);
714         conf->barrier--;
715         spin_unlock_irqrestore(&conf->resync_lock, flags);
716         wake_up(&conf->wait_barrier);
717 }
718
719 static void wait_barrier(struct r1conf *conf)
720 {
721         spin_lock_irq(&conf->resync_lock);
722         if (conf->barrier) {
723                 conf->nr_waiting++;
724                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
725                                     conf->resync_lock,
726                                     );
727                 conf->nr_waiting--;
728         }
729         conf->nr_pending++;
730         spin_unlock_irq(&conf->resync_lock);
731 }
732
733 static void allow_barrier(struct r1conf *conf)
734 {
735         unsigned long flags;
736         spin_lock_irqsave(&conf->resync_lock, flags);
737         conf->nr_pending--;
738         spin_unlock_irqrestore(&conf->resync_lock, flags);
739         wake_up(&conf->wait_barrier);
740 }
741
742 static void freeze_array(struct r1conf *conf)
743 {
744         /* stop syncio and normal IO and wait for everything to
745          * go quite.
746          * We increment barrier and nr_waiting, and then
747          * wait until nr_pending match nr_queued+1
748          * This is called in the context of one normal IO request
749          * that has failed. Thus any sync request that might be pending
750          * will be blocked by nr_pending, and we need to wait for
751          * pending IO requests to complete or be queued for re-try.
752          * Thus the number queued (nr_queued) plus this request (1)
753          * must match the number of pending IOs (nr_pending) before
754          * we continue.
755          */
756         spin_lock_irq(&conf->resync_lock);
757         conf->barrier++;
758         conf->nr_waiting++;
759         wait_event_lock_irq(conf->wait_barrier,
760                             conf->nr_pending == conf->nr_queued+1,
761                             conf->resync_lock,
762                             flush_pending_writes(conf));
763         spin_unlock_irq(&conf->resync_lock);
764 }
765 static void unfreeze_array(struct r1conf *conf)
766 {
767         /* reverse the effect of the freeze */
768         spin_lock_irq(&conf->resync_lock);
769         conf->barrier--;
770         conf->nr_waiting--;
771         wake_up(&conf->wait_barrier);
772         spin_unlock_irq(&conf->resync_lock);
773 }
774
775
776 /* duplicate the data pages for behind I/O 
777  */
778 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
779 {
780         int i;
781         struct bio_vec *bvec;
782         struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
783                                         GFP_NOIO);
784         if (unlikely(!bvecs))
785                 return;
786
787         bio_for_each_segment(bvec, bio, i) {
788                 bvecs[i] = *bvec;
789                 bvecs[i].bv_page = alloc_page(GFP_NOIO);
790                 if (unlikely(!bvecs[i].bv_page))
791                         goto do_sync_io;
792                 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
793                        kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
794                 kunmap(bvecs[i].bv_page);
795                 kunmap(bvec->bv_page);
796         }
797         r1_bio->behind_bvecs = bvecs;
798         r1_bio->behind_page_count = bio->bi_vcnt;
799         set_bit(R1BIO_BehindIO, &r1_bio->state);
800         return;
801
802 do_sync_io:
803         for (i = 0; i < bio->bi_vcnt; i++)
804                 if (bvecs[i].bv_page)
805                         put_page(bvecs[i].bv_page);
806         kfree(bvecs);
807         pr_debug("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
808 }
809
810 static void make_request(struct mddev *mddev, struct bio * bio)
811 {
812         struct r1conf *conf = mddev->private;
813         struct mirror_info *mirror;
814         struct r1bio *r1_bio;
815         struct bio *read_bio;
816         int i, disks;
817         struct bitmap *bitmap;
818         unsigned long flags;
819         const int rw = bio_data_dir(bio);
820         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
821         const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
822         struct md_rdev *blocked_rdev;
823         int plugged;
824         int first_clone;
825         int sectors_handled;
826         int max_sectors;
827
828         /*
829          * Register the new request and wait if the reconstruction
830          * thread has put up a bar for new requests.
831          * Continue immediately if no resync is active currently.
832          */
833
834         md_write_start(mddev, bio); /* wait on superblock update early */
835
836         if (bio_data_dir(bio) == WRITE &&
837             bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
838             bio->bi_sector < mddev->suspend_hi) {
839                 /* As the suspend_* range is controlled by
840                  * userspace, we want an interruptible
841                  * wait.
842                  */
843                 DEFINE_WAIT(w);
844                 for (;;) {
845                         flush_signals(current);
846                         prepare_to_wait(&conf->wait_barrier,
847                                         &w, TASK_INTERRUPTIBLE);
848                         if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
849                             bio->bi_sector >= mddev->suspend_hi)
850                                 break;
851                         schedule();
852                 }
853                 finish_wait(&conf->wait_barrier, &w);
854         }
855
856         wait_barrier(conf);
857
858         bitmap = mddev->bitmap;
859
860         /*
861          * make_request() can abort the operation when READA is being
862          * used and no empty request is available.
863          *
864          */
865         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
866
867         r1_bio->master_bio = bio;
868         r1_bio->sectors = bio->bi_size >> 9;
869         r1_bio->state = 0;
870         r1_bio->mddev = mddev;
871         r1_bio->sector = bio->bi_sector;
872
873         /* We might need to issue multiple reads to different
874          * devices if there are bad blocks around, so we keep
875          * track of the number of reads in bio->bi_phys_segments.
876          * If this is 0, there is only one r1_bio and no locking
877          * will be needed when requests complete.  If it is
878          * non-zero, then it is the number of not-completed requests.
879          */
880         bio->bi_phys_segments = 0;
881         clear_bit(BIO_SEG_VALID, &bio->bi_flags);
882
883         if (rw == READ) {
884                 /*
885                  * read balancing logic:
886                  */
887                 int rdisk;
888
889 read_again:
890                 rdisk = read_balance(conf, r1_bio, &max_sectors);
891
892                 if (rdisk < 0) {
893                         /* couldn't find anywhere to read from */
894                         raid_end_bio_io(r1_bio);
895                         return;
896                 }
897                 mirror = conf->mirrors + rdisk;
898
899                 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
900                     bitmap) {
901                         /* Reading from a write-mostly device must
902                          * take care not to over-take any writes
903                          * that are 'behind'
904                          */
905                         wait_event(bitmap->behind_wait,
906                                    atomic_read(&bitmap->behind_writes) == 0);
907                 }
908                 r1_bio->read_disk = rdisk;
909
910                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
911                 md_trim_bio(read_bio, r1_bio->sector - bio->bi_sector,
912                             max_sectors);
913
914                 r1_bio->bios[rdisk] = read_bio;
915
916                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
917                 read_bio->bi_bdev = mirror->rdev->bdev;
918                 read_bio->bi_end_io = raid1_end_read_request;
919                 read_bio->bi_rw = READ | do_sync;
920                 read_bio->bi_private = r1_bio;
921
922                 if (max_sectors < r1_bio->sectors) {
923                         /* could not read all from this device, so we will
924                          * need another r1_bio.
925                          */
926
927                         sectors_handled = (r1_bio->sector + max_sectors
928                                            - bio->bi_sector);
929                         r1_bio->sectors = max_sectors;
930                         spin_lock_irq(&conf->device_lock);
931                         if (bio->bi_phys_segments == 0)
932                                 bio->bi_phys_segments = 2;
933                         else
934                                 bio->bi_phys_segments++;
935                         spin_unlock_irq(&conf->device_lock);
936                         /* Cannot call generic_make_request directly
937                          * as that will be queued in __make_request
938                          * and subsequent mempool_alloc might block waiting
939                          * for it.  So hand bio over to raid1d.
940                          */
941                         reschedule_retry(r1_bio);
942
943                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
944
945                         r1_bio->master_bio = bio;
946                         r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
947                         r1_bio->state = 0;
948                         r1_bio->mddev = mddev;
949                         r1_bio->sector = bio->bi_sector + sectors_handled;
950                         goto read_again;
951                 } else
952                         generic_make_request(read_bio);
953                 return;
954         }
955
956         /*
957          * WRITE:
958          */
959         if (conf->pending_count >= max_queued_requests) {
960                 md_wakeup_thread(mddev->thread);
961                 wait_event(conf->wait_barrier,
962                            conf->pending_count < max_queued_requests);
963         }
964         /* first select target devices under rcu_lock and
965          * inc refcount on their rdev.  Record them by setting
966          * bios[x] to bio
967          * If there are known/acknowledged bad blocks on any device on
968          * which we have seen a write error, we want to avoid writing those
969          * blocks.
970          * This potentially requires several writes to write around
971          * the bad blocks.  Each set of writes gets it's own r1bio
972          * with a set of bios attached.
973          */
974         plugged = mddev_check_plugged(mddev);
975
976         disks = conf->raid_disks;
977  retry_write:
978         blocked_rdev = NULL;
979         rcu_read_lock();
980         max_sectors = r1_bio->sectors;
981         for (i = 0;  i < disks; i++) {
982                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
983                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
984                         atomic_inc(&rdev->nr_pending);
985                         blocked_rdev = rdev;
986                         break;
987                 }
988                 r1_bio->bios[i] = NULL;
989                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
990                         set_bit(R1BIO_Degraded, &r1_bio->state);
991                         continue;
992                 }
993
994                 atomic_inc(&rdev->nr_pending);
995                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
996                         sector_t first_bad;
997                         int bad_sectors;
998                         int is_bad;
999
1000                         is_bad = is_badblock(rdev, r1_bio->sector,
1001                                              max_sectors,
1002                                              &first_bad, &bad_sectors);
1003                         if (is_bad < 0) {
1004                                 /* mustn't write here until the bad block is
1005                                  * acknowledged*/
1006                                 set_bit(BlockedBadBlocks, &rdev->flags);
1007                                 blocked_rdev = rdev;
1008                                 break;
1009                         }
1010                         if (is_bad && first_bad <= r1_bio->sector) {
1011                                 /* Cannot write here at all */
1012                                 bad_sectors -= (r1_bio->sector - first_bad);
1013                                 if (bad_sectors < max_sectors)
1014                                         /* mustn't write more than bad_sectors
1015                                          * to other devices yet
1016                                          */
1017                                         max_sectors = bad_sectors;
1018                                 rdev_dec_pending(rdev, mddev);
1019                                 /* We don't set R1BIO_Degraded as that
1020                                  * only applies if the disk is
1021                                  * missing, so it might be re-added,
1022                                  * and we want to know to recover this
1023                                  * chunk.
1024                                  * In this case the device is here,
1025                                  * and the fact that this chunk is not
1026                                  * in-sync is recorded in the bad
1027                                  * block log
1028                                  */
1029                                 continue;
1030                         }
1031                         if (is_bad) {
1032                                 int good_sectors = first_bad - r1_bio->sector;
1033                                 if (good_sectors < max_sectors)
1034                                         max_sectors = good_sectors;
1035                         }
1036                 }
1037                 r1_bio->bios[i] = bio;
1038         }
1039         rcu_read_unlock();
1040
1041         if (unlikely(blocked_rdev)) {
1042                 /* Wait for this device to become unblocked */
1043                 int j;
1044
1045                 for (j = 0; j < i; j++)
1046                         if (r1_bio->bios[j])
1047                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1048                 r1_bio->state = 0;
1049                 allow_barrier(conf);
1050                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1051                 wait_barrier(conf);
1052                 goto retry_write;
1053         }
1054
1055         if (max_sectors < r1_bio->sectors) {
1056                 /* We are splitting this write into multiple parts, so
1057                  * we need to prepare for allocating another r1_bio.
1058                  */
1059                 r1_bio->sectors = max_sectors;
1060                 spin_lock_irq(&conf->device_lock);
1061                 if (bio->bi_phys_segments == 0)
1062                         bio->bi_phys_segments = 2;
1063                 else
1064                         bio->bi_phys_segments++;
1065                 spin_unlock_irq(&conf->device_lock);
1066         }
1067         sectors_handled = r1_bio->sector + max_sectors - bio->bi_sector;
1068
1069         atomic_set(&r1_bio->remaining, 1);
1070         atomic_set(&r1_bio->behind_remaining, 0);
1071
1072         first_clone = 1;
1073         for (i = 0; i < disks; i++) {
1074                 struct bio *mbio;
1075                 if (!r1_bio->bios[i])
1076                         continue;
1077
1078                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1079                 md_trim_bio(mbio, r1_bio->sector - bio->bi_sector, max_sectors);
1080
1081                 if (first_clone) {
1082                         /* do behind I/O ?
1083                          * Not if there are too many, or cannot
1084                          * allocate memory, or a reader on WriteMostly
1085                          * is waiting for behind writes to flush */
1086                         if (bitmap &&
1087                             (atomic_read(&bitmap->behind_writes)
1088                              < mddev->bitmap_info.max_write_behind) &&
1089                             !waitqueue_active(&bitmap->behind_wait))
1090                                 alloc_behind_pages(mbio, r1_bio);
1091
1092                         bitmap_startwrite(bitmap, r1_bio->sector,
1093                                           r1_bio->sectors,
1094                                           test_bit(R1BIO_BehindIO,
1095                                                    &r1_bio->state));
1096                         first_clone = 0;
1097                 }
1098                 if (r1_bio->behind_bvecs) {
1099                         struct bio_vec *bvec;
1100                         int j;
1101
1102                         /* Yes, I really want the '__' version so that
1103                          * we clear any unused pointer in the io_vec, rather
1104                          * than leave them unchanged.  This is important
1105                          * because when we come to free the pages, we won't
1106                          * know the original bi_idx, so we just free
1107                          * them all
1108                          */
1109                         __bio_for_each_segment(bvec, mbio, j, 0)
1110                                 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1111                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1112                                 atomic_inc(&r1_bio->behind_remaining);
1113                 }
1114
1115                 r1_bio->bios[i] = mbio;
1116
1117                 mbio->bi_sector = (r1_bio->sector +
1118                                    conf->mirrors[i].rdev->data_offset);
1119                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1120                 mbio->bi_end_io = raid1_end_write_request;
1121                 mbio->bi_rw = WRITE | do_flush_fua | do_sync;
1122                 mbio->bi_private = r1_bio;
1123
1124                 atomic_inc(&r1_bio->remaining);
1125                 spin_lock_irqsave(&conf->device_lock, flags);
1126                 bio_list_add(&conf->pending_bio_list, mbio);
1127                 conf->pending_count++;
1128                 spin_unlock_irqrestore(&conf->device_lock, flags);
1129         }
1130         /* Mustn't call r1_bio_write_done before this next test,
1131          * as it could result in the bio being freed.
1132          */
1133         if (sectors_handled < (bio->bi_size >> 9)) {
1134                 r1_bio_write_done(r1_bio);
1135                 /* We need another r1_bio.  It has already been counted
1136                  * in bio->bi_phys_segments
1137                  */
1138                 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1139                 r1_bio->master_bio = bio;
1140                 r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1141                 r1_bio->state = 0;
1142                 r1_bio->mddev = mddev;
1143                 r1_bio->sector = bio->bi_sector + sectors_handled;
1144                 goto retry_write;
1145         }
1146
1147         r1_bio_write_done(r1_bio);
1148
1149         /* In case raid1d snuck in to freeze_array */
1150         wake_up(&conf->wait_barrier);
1151
1152         if (do_sync || !bitmap || !plugged)
1153                 md_wakeup_thread(mddev->thread);
1154 }
1155
1156 static void status(struct seq_file *seq, struct mddev *mddev)
1157 {
1158         struct r1conf *conf = mddev->private;
1159         int i;
1160
1161         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1162                    conf->raid_disks - mddev->degraded);
1163         rcu_read_lock();
1164         for (i = 0; i < conf->raid_disks; i++) {
1165                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1166                 seq_printf(seq, "%s",
1167                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1168         }
1169         rcu_read_unlock();
1170         seq_printf(seq, "]");
1171 }
1172
1173
1174 static void error(struct mddev *mddev, struct md_rdev *rdev)
1175 {
1176         char b[BDEVNAME_SIZE];
1177         struct r1conf *conf = mddev->private;
1178
1179         /*
1180          * If it is not operational, then we have already marked it as dead
1181          * else if it is the last working disks, ignore the error, let the
1182          * next level up know.
1183          * else mark the drive as failed
1184          */
1185         if (test_bit(In_sync, &rdev->flags)
1186             && (conf->raid_disks - mddev->degraded) == 1) {
1187                 /*
1188                  * Don't fail the drive, act as though we were just a
1189                  * normal single drive.
1190                  * However don't try a recovery from this drive as
1191                  * it is very likely to fail.
1192                  */
1193                 conf->recovery_disabled = mddev->recovery_disabled;
1194                 return;
1195         }
1196         set_bit(Blocked, &rdev->flags);
1197         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1198                 unsigned long flags;
1199                 spin_lock_irqsave(&conf->device_lock, flags);
1200                 mddev->degraded++;
1201                 set_bit(Faulty, &rdev->flags);
1202                 spin_unlock_irqrestore(&conf->device_lock, flags);
1203                 /*
1204                  * if recovery is running, make sure it aborts.
1205                  */
1206                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1207         } else
1208                 set_bit(Faulty, &rdev->flags);
1209         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1210         printk(KERN_ALERT
1211                "md/raid1:%s: Disk failure on %s, disabling device.\n"
1212                "md/raid1:%s: Operation continuing on %d devices.\n",
1213                mdname(mddev), bdevname(rdev->bdev, b),
1214                mdname(mddev), conf->raid_disks - mddev->degraded);
1215 }
1216
1217 static void print_conf(struct r1conf *conf)
1218 {
1219         int i;
1220
1221         printk(KERN_DEBUG "RAID1 conf printout:\n");
1222         if (!conf) {
1223                 printk(KERN_DEBUG "(!conf)\n");
1224                 return;
1225         }
1226         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1227                 conf->raid_disks);
1228
1229         rcu_read_lock();
1230         for (i = 0; i < conf->raid_disks; i++) {
1231                 char b[BDEVNAME_SIZE];
1232                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1233                 if (rdev)
1234                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1235                                i, !test_bit(In_sync, &rdev->flags),
1236                                !test_bit(Faulty, &rdev->flags),
1237                                bdevname(rdev->bdev,b));
1238         }
1239         rcu_read_unlock();
1240 }
1241
1242 static void close_sync(struct r1conf *conf)
1243 {
1244         wait_barrier(conf);
1245         allow_barrier(conf);
1246
1247         mempool_destroy(conf->r1buf_pool);
1248         conf->r1buf_pool = NULL;
1249 }
1250
1251 static int raid1_spare_active(struct mddev *mddev)
1252 {
1253         int i;
1254         struct r1conf *conf = mddev->private;
1255         int count = 0;
1256         unsigned long flags;
1257
1258         /*
1259          * Find all failed disks within the RAID1 configuration 
1260          * and mark them readable.
1261          * Called under mddev lock, so rcu protection not needed.
1262          */
1263         for (i = 0; i < conf->raid_disks; i++) {
1264                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1265                 if (rdev
1266                     && !test_bit(Faulty, &rdev->flags)
1267                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1268                         count++;
1269                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1270                 }
1271         }
1272         spin_lock_irqsave(&conf->device_lock, flags);
1273         mddev->degraded -= count;
1274         spin_unlock_irqrestore(&conf->device_lock, flags);
1275
1276         print_conf(conf);
1277         return count;
1278 }
1279
1280
1281 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1282 {
1283         struct r1conf *conf = mddev->private;
1284         int err = -EEXIST;
1285         int mirror = 0;
1286         struct mirror_info *p;
1287         int first = 0;
1288         int last = mddev->raid_disks - 1;
1289
1290         if (mddev->recovery_disabled == conf->recovery_disabled)
1291                 return -EBUSY;
1292
1293         if (rdev->raid_disk >= 0)
1294                 first = last = rdev->raid_disk;
1295
1296         for (mirror = first; mirror <= last; mirror++)
1297                 if ( !(p=conf->mirrors+mirror)->rdev) {
1298
1299                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1300                                           rdev->data_offset << 9);
1301                         /* as we don't honour merge_bvec_fn, we must
1302                          * never risk violating it, so limit
1303                          * ->max_segments to one lying with a single
1304                          * page, as a one page request is never in
1305                          * violation.
1306                          */
1307                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1308                                 blk_queue_max_segments(mddev->queue, 1);
1309                                 blk_queue_segment_boundary(mddev->queue,
1310                                                            PAGE_CACHE_SIZE - 1);
1311                         }
1312
1313                         p->head_position = 0;
1314                         rdev->raid_disk = mirror;
1315                         err = 0;
1316                         /* As all devices are equivalent, we don't need a full recovery
1317                          * if this was recently any drive of the array
1318                          */
1319                         if (rdev->saved_raid_disk < 0)
1320                                 conf->fullsync = 1;
1321                         rcu_assign_pointer(p->rdev, rdev);
1322                         break;
1323                 }
1324         md_integrity_add_rdev(rdev, mddev);
1325         print_conf(conf);
1326         return err;
1327 }
1328
1329 static int raid1_remove_disk(struct mddev *mddev, int number)
1330 {
1331         struct r1conf *conf = mddev->private;
1332         int err = 0;
1333         struct md_rdev *rdev;
1334         struct mirror_info *p = conf->mirrors+ number;
1335
1336         print_conf(conf);
1337         rdev = p->rdev;
1338         if (rdev) {
1339                 if (test_bit(In_sync, &rdev->flags) ||
1340                     atomic_read(&rdev->nr_pending)) {
1341                         err = -EBUSY;
1342                         goto abort;
1343                 }
1344                 /* Only remove non-faulty devices if recovery
1345                  * is not possible.
1346                  */
1347                 if (!test_bit(Faulty, &rdev->flags) &&
1348                     mddev->recovery_disabled != conf->recovery_disabled &&
1349                     mddev->degraded < conf->raid_disks) {
1350                         err = -EBUSY;
1351                         goto abort;
1352                 }
1353                 p->rdev = NULL;
1354                 synchronize_rcu();
1355                 if (atomic_read(&rdev->nr_pending)) {
1356                         /* lost the race, try later */
1357                         err = -EBUSY;
1358                         p->rdev = rdev;
1359                         goto abort;
1360                 }
1361                 err = md_integrity_register(mddev);
1362         }
1363 abort:
1364
1365         print_conf(conf);
1366         return err;
1367 }
1368
1369
1370 static void end_sync_read(struct bio *bio, int error)
1371 {
1372         struct r1bio *r1_bio = bio->bi_private;
1373
1374         update_head_pos(r1_bio->read_disk, r1_bio);
1375
1376         /*
1377          * we have read a block, now it needs to be re-written,
1378          * or re-read if the read failed.
1379          * We don't do much here, just schedule handling by raid1d
1380          */
1381         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1382                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1383
1384         if (atomic_dec_and_test(&r1_bio->remaining))
1385                 reschedule_retry(r1_bio);
1386 }
1387
1388 static void end_sync_write(struct bio *bio, int error)
1389 {
1390         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1391         struct r1bio *r1_bio = bio->bi_private;
1392         struct mddev *mddev = r1_bio->mddev;
1393         struct r1conf *conf = mddev->private;
1394         int mirror=0;
1395         sector_t first_bad;
1396         int bad_sectors;
1397
1398         mirror = find_bio_disk(r1_bio, bio);
1399
1400         if (!uptodate) {
1401                 sector_t sync_blocks = 0;
1402                 sector_t s = r1_bio->sector;
1403                 long sectors_to_go = r1_bio->sectors;
1404                 /* make sure these bits doesn't get cleared. */
1405                 do {
1406                         bitmap_end_sync(mddev->bitmap, s,
1407                                         &sync_blocks, 1);
1408                         s += sync_blocks;
1409                         sectors_to_go -= sync_blocks;
1410                 } while (sectors_to_go > 0);
1411                 set_bit(WriteErrorSeen,
1412                         &conf->mirrors[mirror].rdev->flags);
1413                 set_bit(R1BIO_WriteError, &r1_bio->state);
1414         } else if (is_badblock(conf->mirrors[mirror].rdev,
1415                                r1_bio->sector,
1416                                r1_bio->sectors,
1417                                &first_bad, &bad_sectors) &&
1418                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1419                                 r1_bio->sector,
1420                                 r1_bio->sectors,
1421                                 &first_bad, &bad_sectors)
1422                 )
1423                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1424
1425         if (atomic_dec_and_test(&r1_bio->remaining)) {
1426                 int s = r1_bio->sectors;
1427                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1428                     test_bit(R1BIO_WriteError, &r1_bio->state))
1429                         reschedule_retry(r1_bio);
1430                 else {
1431                         put_buf(r1_bio);
1432                         md_done_sync(mddev, s, uptodate);
1433                 }
1434         }
1435 }
1436
1437 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1438                             int sectors, struct page *page, int rw)
1439 {
1440         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1441                 /* success */
1442                 return 1;
1443         if (rw == WRITE)
1444                 set_bit(WriteErrorSeen, &rdev->flags);
1445         /* need to record an error - either for the block or the device */
1446         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1447                 md_error(rdev->mddev, rdev);
1448         return 0;
1449 }
1450
1451 static int fix_sync_read_error(struct r1bio *r1_bio)
1452 {
1453         /* Try some synchronous reads of other devices to get
1454          * good data, much like with normal read errors.  Only
1455          * read into the pages we already have so we don't
1456          * need to re-issue the read request.
1457          * We don't need to freeze the array, because being in an
1458          * active sync request, there is no normal IO, and
1459          * no overlapping syncs.
1460          * We don't need to check is_badblock() again as we
1461          * made sure that anything with a bad block in range
1462          * will have bi_end_io clear.
1463          */
1464         struct mddev *mddev = r1_bio->mddev;
1465         struct r1conf *conf = mddev->private;
1466         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1467         sector_t sect = r1_bio->sector;
1468         int sectors = r1_bio->sectors;
1469         int idx = 0;
1470
1471         while(sectors) {
1472                 int s = sectors;
1473                 int d = r1_bio->read_disk;
1474                 int success = 0;
1475                 struct md_rdev *rdev;
1476                 int start;
1477
1478                 if (s > (PAGE_SIZE>>9))
1479                         s = PAGE_SIZE >> 9;
1480                 do {
1481                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1482                                 /* No rcu protection needed here devices
1483                                  * can only be removed when no resync is
1484                                  * active, and resync is currently active
1485                                  */
1486                                 rdev = conf->mirrors[d].rdev;
1487                                 if (sync_page_io(rdev, sect, s<<9,
1488                                                  bio->bi_io_vec[idx].bv_page,
1489                                                  READ, false)) {
1490                                         success = 1;
1491                                         break;
1492                                 }
1493                         }
1494                         d++;
1495                         if (d == conf->raid_disks)
1496                                 d = 0;
1497                 } while (!success && d != r1_bio->read_disk);
1498
1499                 if (!success) {
1500                         char b[BDEVNAME_SIZE];
1501                         int abort = 0;
1502                         /* Cannot read from anywhere, this block is lost.
1503                          * Record a bad block on each device.  If that doesn't
1504                          * work just disable and interrupt the recovery.
1505                          * Don't fail devices as that won't really help.
1506                          */
1507                         printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1508                                " for block %llu\n",
1509                                mdname(mddev),
1510                                bdevname(bio->bi_bdev, b),
1511                                (unsigned long long)r1_bio->sector);
1512                         for (d = 0; d < conf->raid_disks; d++) {
1513                                 rdev = conf->mirrors[d].rdev;
1514                                 if (!rdev || test_bit(Faulty, &rdev->flags))
1515                                         continue;
1516                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
1517                                         abort = 1;
1518                         }
1519                         if (abort) {
1520                                 conf->recovery_disabled =
1521                                         mddev->recovery_disabled;
1522                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1523                                 md_done_sync(mddev, r1_bio->sectors, 0);
1524                                 put_buf(r1_bio);
1525                                 return 0;
1526                         }
1527                         /* Try next page */
1528                         sectors -= s;
1529                         sect += s;
1530                         idx++;
1531                         continue;
1532                 }
1533
1534                 start = d;
1535                 /* write it back and re-read */
1536                 while (d != r1_bio->read_disk) {
1537                         if (d == 0)
1538                                 d = conf->raid_disks;
1539                         d--;
1540                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1541                                 continue;
1542                         rdev = conf->mirrors[d].rdev;
1543                         if (r1_sync_page_io(rdev, sect, s,
1544                                             bio->bi_io_vec[idx].bv_page,
1545                                             WRITE) == 0) {
1546                                 r1_bio->bios[d]->bi_end_io = NULL;
1547                                 rdev_dec_pending(rdev, mddev);
1548                         }
1549                 }
1550                 d = start;
1551                 while (d != r1_bio->read_disk) {
1552                         if (d == 0)
1553                                 d = conf->raid_disks;
1554                         d--;
1555                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1556                                 continue;
1557                         rdev = conf->mirrors[d].rdev;
1558                         if (r1_sync_page_io(rdev, sect, s,
1559                                             bio->bi_io_vec[idx].bv_page,
1560                                             READ) != 0)
1561                                 atomic_add(s, &rdev->corrected_errors);
1562                 }
1563                 sectors -= s;
1564                 sect += s;
1565                 idx ++;
1566         }
1567         set_bit(R1BIO_Uptodate, &r1_bio->state);
1568         set_bit(BIO_UPTODATE, &bio->bi_flags);
1569         return 1;
1570 }
1571
1572 static int process_checks(struct r1bio *r1_bio)
1573 {
1574         /* We have read all readable devices.  If we haven't
1575          * got the block, then there is no hope left.
1576          * If we have, then we want to do a comparison
1577          * and skip the write if everything is the same.
1578          * If any blocks failed to read, then we need to
1579          * attempt an over-write
1580          */
1581         struct mddev *mddev = r1_bio->mddev;
1582         struct r1conf *conf = mddev->private;
1583         int primary;
1584         int i;
1585
1586         for (primary = 0; primary < conf->raid_disks; primary++)
1587                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1588                     test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1589                         r1_bio->bios[primary]->bi_end_io = NULL;
1590                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1591                         break;
1592                 }
1593         r1_bio->read_disk = primary;
1594         for (i = 0; i < conf->raid_disks; i++) {
1595                 int j;
1596                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1597                 struct bio *pbio = r1_bio->bios[primary];
1598                 struct bio *sbio = r1_bio->bios[i];
1599                 int size;
1600
1601                 if (r1_bio->bios[i]->bi_end_io != end_sync_read)
1602                         continue;
1603
1604                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1605                         for (j = vcnt; j-- ; ) {
1606                                 struct page *p, *s;
1607                                 p = pbio->bi_io_vec[j].bv_page;
1608                                 s = sbio->bi_io_vec[j].bv_page;
1609                                 if (memcmp(page_address(p),
1610                                            page_address(s),
1611                                            PAGE_SIZE))
1612                                         break;
1613                         }
1614                 } else
1615                         j = 0;
1616                 if (j >= 0)
1617                         mddev->resync_mismatches += r1_bio->sectors;
1618                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1619                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1620                         /* No need to write to this device. */
1621                         sbio->bi_end_io = NULL;
1622                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1623                         continue;
1624                 }
1625                 /* fixup the bio for reuse */
1626                 sbio->bi_vcnt = vcnt;
1627                 sbio->bi_size = r1_bio->sectors << 9;
1628                 sbio->bi_idx = 0;
1629                 sbio->bi_phys_segments = 0;
1630                 sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1631                 sbio->bi_flags |= 1 << BIO_UPTODATE;
1632                 sbio->bi_next = NULL;
1633                 sbio->bi_sector = r1_bio->sector +
1634                         conf->mirrors[i].rdev->data_offset;
1635                 sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1636                 size = sbio->bi_size;
1637                 for (j = 0; j < vcnt ; j++) {
1638                         struct bio_vec *bi;
1639                         bi = &sbio->bi_io_vec[j];
1640                         bi->bv_offset = 0;
1641                         if (size > PAGE_SIZE)
1642                                 bi->bv_len = PAGE_SIZE;
1643                         else
1644                                 bi->bv_len = size;
1645                         size -= PAGE_SIZE;
1646                         memcpy(page_address(bi->bv_page),
1647                                page_address(pbio->bi_io_vec[j].bv_page),
1648                                PAGE_SIZE);
1649                 }
1650         }
1651         return 0;
1652 }
1653
1654 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
1655 {
1656         struct r1conf *conf = mddev->private;
1657         int i;
1658         int disks = conf->raid_disks;
1659         struct bio *bio, *wbio;
1660
1661         bio = r1_bio->bios[r1_bio->read_disk];
1662
1663         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1664                 /* ouch - failed to read all of that. */
1665                 if (!fix_sync_read_error(r1_bio))
1666                         return;
1667
1668         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1669                 if (process_checks(r1_bio) < 0)
1670                         return;
1671         /*
1672          * schedule writes
1673          */
1674         atomic_set(&r1_bio->remaining, 1);
1675         for (i = 0; i < disks ; i++) {
1676                 wbio = r1_bio->bios[i];
1677                 if (wbio->bi_end_io == NULL ||
1678                     (wbio->bi_end_io == end_sync_read &&
1679                      (i == r1_bio->read_disk ||
1680                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1681                         continue;
1682
1683                 wbio->bi_rw = WRITE;
1684                 wbio->bi_end_io = end_sync_write;
1685                 atomic_inc(&r1_bio->remaining);
1686                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1687
1688                 generic_make_request(wbio);
1689         }
1690
1691         if (atomic_dec_and_test(&r1_bio->remaining)) {
1692                 /* if we're here, all write(s) have completed, so clean up */
1693                 md_done_sync(mddev, r1_bio->sectors, 1);
1694                 put_buf(r1_bio);
1695         }
1696 }
1697
1698 /*
1699  * This is a kernel thread which:
1700  *
1701  *      1.      Retries failed read operations on working mirrors.
1702  *      2.      Updates the raid superblock when problems encounter.
1703  *      3.      Performs writes following reads for array synchronising.
1704  */
1705
1706 static void fix_read_error(struct r1conf *conf, int read_disk,
1707                            sector_t sect, int sectors)
1708 {
1709         struct mddev *mddev = conf->mddev;
1710         while(sectors) {
1711                 int s = sectors;
1712                 int d = read_disk;
1713                 int success = 0;
1714                 int start;
1715                 struct md_rdev *rdev;
1716
1717                 if (s > (PAGE_SIZE>>9))
1718                         s = PAGE_SIZE >> 9;
1719
1720                 do {
1721                         /* Note: no rcu protection needed here
1722                          * as this is synchronous in the raid1d thread
1723                          * which is the thread that might remove
1724                          * a device.  If raid1d ever becomes multi-threaded....
1725                          */
1726                         sector_t first_bad;
1727                         int bad_sectors;
1728
1729                         rdev = conf->mirrors[d].rdev;
1730                         if (rdev &&
1731                             test_bit(In_sync, &rdev->flags) &&
1732                             is_badblock(rdev, sect, s,
1733                                         &first_bad, &bad_sectors) == 0 &&
1734                             sync_page_io(rdev, sect, s<<9,
1735                                          conf->tmppage, READ, false))
1736                                 success = 1;
1737                         else {
1738                                 d++;
1739                                 if (d == conf->raid_disks)
1740                                         d = 0;
1741                         }
1742                 } while (!success && d != read_disk);
1743
1744                 if (!success) {
1745                         /* Cannot read from anywhere - mark it bad */
1746                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
1747                         if (!rdev_set_badblocks(rdev, sect, s, 0))
1748                                 md_error(mddev, rdev);
1749                         break;
1750                 }
1751                 /* write it back and re-read */
1752                 start = d;
1753                 while (d != read_disk) {
1754                         if (d==0)
1755                                 d = conf->raid_disks;
1756                         d--;
1757                         rdev = conf->mirrors[d].rdev;
1758                         if (rdev &&
1759                             test_bit(In_sync, &rdev->flags))
1760                                 r1_sync_page_io(rdev, sect, s,
1761                                                 conf->tmppage, WRITE);
1762                 }
1763                 d = start;
1764                 while (d != read_disk) {
1765                         char b[BDEVNAME_SIZE];
1766                         if (d==0)
1767                                 d = conf->raid_disks;
1768                         d--;
1769                         rdev = conf->mirrors[d].rdev;
1770                         if (rdev &&
1771                             test_bit(In_sync, &rdev->flags)) {
1772                                 if (r1_sync_page_io(rdev, sect, s,
1773                                                     conf->tmppage, READ)) {
1774                                         atomic_add(s, &rdev->corrected_errors);
1775                                         printk(KERN_INFO
1776                                                "md/raid1:%s: read error corrected "
1777                                                "(%d sectors at %llu on %s)\n",
1778                                                mdname(mddev), s,
1779                                                (unsigned long long)(sect +
1780                                                    rdev->data_offset),
1781                                                bdevname(rdev->bdev, b));
1782                                 }
1783                         }
1784                 }
1785                 sectors -= s;
1786                 sect += s;
1787         }
1788 }
1789
1790 static void bi_complete(struct bio *bio, int error)
1791 {
1792         complete((struct completion *)bio->bi_private);
1793 }
1794
1795 static int submit_bio_wait(int rw, struct bio *bio)
1796 {
1797         struct completion event;
1798         rw |= REQ_SYNC;
1799
1800         init_completion(&event);
1801         bio->bi_private = &event;
1802         bio->bi_end_io = bi_complete;
1803         submit_bio(rw, bio);
1804         wait_for_completion(&event);
1805
1806         return test_bit(BIO_UPTODATE, &bio->bi_flags);
1807 }
1808
1809 static int narrow_write_error(struct r1bio *r1_bio, int i)
1810 {
1811         struct mddev *mddev = r1_bio->mddev;
1812         struct r1conf *conf = mddev->private;
1813         struct md_rdev *rdev = conf->mirrors[i].rdev;
1814         int vcnt, idx;
1815         struct bio_vec *vec;
1816
1817         /* bio has the data to be written to device 'i' where
1818          * we just recently had a write error.
1819          * We repeatedly clone the bio and trim down to one block,
1820          * then try the write.  Where the write fails we record
1821          * a bad block.
1822          * It is conceivable that the bio doesn't exactly align with
1823          * blocks.  We must handle this somehow.
1824          *
1825          * We currently own a reference on the rdev.
1826          */
1827
1828         int block_sectors;
1829         sector_t sector;
1830         int sectors;
1831         int sect_to_write = r1_bio->sectors;
1832         int ok = 1;
1833
1834         if (rdev->badblocks.shift < 0)
1835                 return 0;
1836
1837         block_sectors = 1 << rdev->badblocks.shift;
1838         sector = r1_bio->sector;
1839         sectors = ((sector + block_sectors)
1840                    & ~(sector_t)(block_sectors - 1))
1841                 - sector;
1842
1843         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
1844                 vcnt = r1_bio->behind_page_count;
1845                 vec = r1_bio->behind_bvecs;
1846                 idx = 0;
1847                 while (vec[idx].bv_page == NULL)
1848                         idx++;
1849         } else {
1850                 vcnt = r1_bio->master_bio->bi_vcnt;
1851                 vec = r1_bio->master_bio->bi_io_vec;
1852                 idx = r1_bio->master_bio->bi_idx;
1853         }
1854         while (sect_to_write) {
1855                 struct bio *wbio;
1856                 if (sectors > sect_to_write)
1857                         sectors = sect_to_write;
1858                 /* Write at 'sector' for 'sectors'*/
1859
1860                 wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
1861                 memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
1862                 wbio->bi_sector = r1_bio->sector;
1863                 wbio->bi_rw = WRITE;
1864                 wbio->bi_vcnt = vcnt;
1865                 wbio->bi_size = r1_bio->sectors << 9;
1866                 wbio->bi_idx = idx;
1867
1868                 md_trim_bio(wbio, sector - r1_bio->sector, sectors);
1869                 wbio->bi_sector += rdev->data_offset;
1870                 wbio->bi_bdev = rdev->bdev;
1871                 if (submit_bio_wait(WRITE, wbio) == 0)
1872                         /* failure! */
1873                         ok = rdev_set_badblocks(rdev, sector,
1874                                                 sectors, 0)
1875                                 && ok;
1876
1877                 bio_put(wbio);
1878                 sect_to_write -= sectors;
1879                 sector += sectors;
1880                 sectors = block_sectors;
1881         }
1882         return ok;
1883 }
1884
1885 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
1886 {
1887         int m;
1888         int s = r1_bio->sectors;
1889         for (m = 0; m < conf->raid_disks ; m++) {
1890                 struct md_rdev *rdev = conf->mirrors[m].rdev;
1891                 struct bio *bio = r1_bio->bios[m];
1892                 if (bio->bi_end_io == NULL)
1893                         continue;
1894                 if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
1895                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
1896                         rdev_clear_badblocks(rdev, r1_bio->sector, s);
1897                 }
1898                 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
1899                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
1900                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
1901                                 md_error(conf->mddev, rdev);
1902                 }
1903         }
1904         put_buf(r1_bio);
1905         md_done_sync(conf->mddev, s, 1);
1906 }
1907
1908 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
1909 {
1910         int m;
1911         for (m = 0; m < conf->raid_disks ; m++)
1912                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
1913                         struct md_rdev *rdev = conf->mirrors[m].rdev;
1914                         rdev_clear_badblocks(rdev,
1915                                              r1_bio->sector,
1916                                              r1_bio->sectors);
1917                         rdev_dec_pending(rdev, conf->mddev);
1918                 } else if (r1_bio->bios[m] != NULL) {
1919                         /* This drive got a write error.  We need to
1920                          * narrow down and record precise write
1921                          * errors.
1922                          */
1923                         if (!narrow_write_error(r1_bio, m)) {
1924                                 md_error(conf->mddev,
1925                                          conf->mirrors[m].rdev);
1926                                 /* an I/O failed, we can't clear the bitmap */
1927                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1928                         }
1929                         rdev_dec_pending(conf->mirrors[m].rdev,
1930                                          conf->mddev);
1931                 }
1932         if (test_bit(R1BIO_WriteError, &r1_bio->state))
1933                 close_write(r1_bio);
1934         raid_end_bio_io(r1_bio);
1935 }
1936
1937 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
1938 {
1939         int disk;
1940         int max_sectors;
1941         struct mddev *mddev = conf->mddev;
1942         struct bio *bio;
1943         char b[BDEVNAME_SIZE];
1944         struct md_rdev *rdev;
1945
1946         clear_bit(R1BIO_ReadError, &r1_bio->state);
1947         /* we got a read error. Maybe the drive is bad.  Maybe just
1948          * the block and we can fix it.
1949          * We freeze all other IO, and try reading the block from
1950          * other devices.  When we find one, we re-write
1951          * and check it that fixes the read error.
1952          * This is all done synchronously while the array is
1953          * frozen
1954          */
1955         if (mddev->ro == 0) {
1956                 freeze_array(conf);
1957                 fix_read_error(conf, r1_bio->read_disk,
1958                                r1_bio->sector, r1_bio->sectors);
1959                 unfreeze_array(conf);
1960         } else
1961                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1962
1963         bio = r1_bio->bios[r1_bio->read_disk];
1964         bdevname(bio->bi_bdev, b);
1965 read_more:
1966         disk = read_balance(conf, r1_bio, &max_sectors);
1967         if (disk == -1) {
1968                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
1969                        " read error for block %llu\n",
1970                        mdname(mddev), b, (unsigned long long)r1_bio->sector);
1971                 raid_end_bio_io(r1_bio);
1972         } else {
1973                 const unsigned long do_sync
1974                         = r1_bio->master_bio->bi_rw & REQ_SYNC;
1975                 if (bio) {
1976                         r1_bio->bios[r1_bio->read_disk] =
1977                                 mddev->ro ? IO_BLOCKED : NULL;
1978                         bio_put(bio);
1979                 }
1980                 r1_bio->read_disk = disk;
1981                 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
1982                 md_trim_bio(bio, r1_bio->sector - bio->bi_sector, max_sectors);
1983                 r1_bio->bios[r1_bio->read_disk] = bio;
1984                 rdev = conf->mirrors[disk].rdev;
1985                 printk_ratelimited(KERN_ERR
1986                                    "md/raid1:%s: redirecting sector %llu"
1987                                    " to other mirror: %s\n",
1988                                    mdname(mddev),
1989                                    (unsigned long long)r1_bio->sector,
1990                                    bdevname(rdev->bdev, b));
1991                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1992                 bio->bi_bdev = rdev->bdev;
1993                 bio->bi_end_io = raid1_end_read_request;
1994                 bio->bi_rw = READ | do_sync;
1995                 bio->bi_private = r1_bio;
1996                 if (max_sectors < r1_bio->sectors) {
1997                         /* Drat - have to split this up more */
1998                         struct bio *mbio = r1_bio->master_bio;
1999                         int sectors_handled = (r1_bio->sector + max_sectors
2000                                                - mbio->bi_sector);
2001                         r1_bio->sectors = max_sectors;
2002                         spin_lock_irq(&conf->device_lock);
2003                         if (mbio->bi_phys_segments == 0)
2004                                 mbio->bi_phys_segments = 2;
2005                         else
2006                                 mbio->bi_phys_segments++;
2007                         spin_unlock_irq(&conf->device_lock);
2008                         generic_make_request(bio);
2009                         bio = NULL;
2010
2011                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2012
2013                         r1_bio->master_bio = mbio;
2014                         r1_bio->sectors = (mbio->bi_size >> 9)
2015                                           - sectors_handled;
2016                         r1_bio->state = 0;
2017                         set_bit(R1BIO_ReadError, &r1_bio->state);
2018                         r1_bio->mddev = mddev;
2019                         r1_bio->sector = mbio->bi_sector + sectors_handled;
2020
2021                         goto read_more;
2022                 } else
2023                         generic_make_request(bio);
2024         }
2025 }
2026
2027 static void raid1d(struct mddev *mddev)
2028 {
2029         struct r1bio *r1_bio;
2030         unsigned long flags;
2031         struct r1conf *conf = mddev->private;
2032         struct list_head *head = &conf->retry_list;
2033         struct blk_plug plug;
2034
2035         md_check_recovery(mddev);
2036
2037         blk_start_plug(&plug);
2038         for (;;) {
2039
2040                 if (atomic_read(&mddev->plug_cnt) == 0)
2041                         flush_pending_writes(conf);
2042
2043                 spin_lock_irqsave(&conf->device_lock, flags);
2044                 if (list_empty(head)) {
2045                         spin_unlock_irqrestore(&conf->device_lock, flags);
2046                         break;
2047                 }
2048                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2049                 list_del(head->prev);
2050                 conf->nr_queued--;
2051                 spin_unlock_irqrestore(&conf->device_lock, flags);
2052
2053                 mddev = r1_bio->mddev;
2054                 conf = mddev->private;
2055                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2056                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2057                             test_bit(R1BIO_WriteError, &r1_bio->state))
2058                                 handle_sync_write_finished(conf, r1_bio);
2059                         else
2060                                 sync_request_write(mddev, r1_bio);
2061                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2062                            test_bit(R1BIO_WriteError, &r1_bio->state))
2063                         handle_write_finished(conf, r1_bio);
2064                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2065                         handle_read_error(conf, r1_bio);
2066                 else
2067                         /* just a partial read to be scheduled from separate
2068                          * context
2069                          */
2070                         generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2071
2072                 cond_resched();
2073                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2074                         md_check_recovery(mddev);
2075         }
2076         blk_finish_plug(&plug);
2077 }
2078
2079
2080 static int init_resync(struct r1conf *conf)
2081 {
2082         int buffs;
2083
2084         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2085         BUG_ON(conf->r1buf_pool);
2086         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2087                                           conf->poolinfo);
2088         if (!conf->r1buf_pool)
2089                 return -ENOMEM;
2090         conf->next_resync = 0;
2091         return 0;
2092 }
2093
2094 /*
2095  * perform a "sync" on one "block"
2096  *
2097  * We need to make sure that no normal I/O request - particularly write
2098  * requests - conflict with active sync requests.
2099  *
2100  * This is achieved by tracking pending requests and a 'barrier' concept
2101  * that can be installed to exclude normal IO requests.
2102  */
2103
2104 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2105 {
2106         struct r1conf *conf = mddev->private;
2107         struct r1bio *r1_bio;
2108         struct bio *bio;
2109         sector_t max_sector, nr_sectors;
2110         int disk = -1;
2111         int i;
2112         int wonly = -1;
2113         int write_targets = 0, read_targets = 0;
2114         sector_t sync_blocks;
2115         int still_degraded = 0;
2116         int good_sectors = RESYNC_SECTORS;
2117         int min_bad = 0; /* number of sectors that are bad in all devices */
2118
2119         if (!conf->r1buf_pool)
2120                 if (init_resync(conf))
2121                         return 0;
2122
2123         max_sector = mddev->dev_sectors;
2124         if (sector_nr >= max_sector) {
2125                 /* If we aborted, we need to abort the
2126                  * sync on the 'current' bitmap chunk (there will
2127                  * only be one in raid1 resync.
2128                  * We can find the current addess in mddev->curr_resync
2129                  */
2130                 if (mddev->curr_resync < max_sector) /* aborted */
2131                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2132                                                 &sync_blocks, 1);
2133                 else /* completed sync */
2134                         conf->fullsync = 0;
2135
2136                 bitmap_close_sync(mddev->bitmap);
2137                 close_sync(conf);
2138                 return 0;
2139         }
2140
2141         if (mddev->bitmap == NULL &&
2142             mddev->recovery_cp == MaxSector &&
2143             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2144             conf->fullsync == 0) {
2145                 *skipped = 1;
2146                 return max_sector - sector_nr;
2147         }
2148         /* before building a request, check if we can skip these blocks..
2149          * This call the bitmap_start_sync doesn't actually record anything
2150          */
2151         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2152             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2153                 /* We can skip this block, and probably several more */
2154                 *skipped = 1;
2155                 return sync_blocks;
2156         }
2157         /*
2158          * If there is non-resync activity waiting for a turn,
2159          * and resync is going fast enough,
2160          * then let it though before starting on this new sync request.
2161          */
2162         if (!go_faster && conf->nr_waiting)
2163                 msleep_interruptible(1000);
2164
2165         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2166         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2167         raise_barrier(conf);
2168
2169         conf->next_resync = sector_nr;
2170
2171         rcu_read_lock();
2172         /*
2173          * If we get a correctably read error during resync or recovery,
2174          * we might want to read from a different device.  So we
2175          * flag all drives that could conceivably be read from for READ,
2176          * and any others (which will be non-In_sync devices) for WRITE.
2177          * If a read fails, we try reading from something else for which READ
2178          * is OK.
2179          */
2180
2181         r1_bio->mddev = mddev;
2182         r1_bio->sector = sector_nr;
2183         r1_bio->state = 0;
2184         set_bit(R1BIO_IsSync, &r1_bio->state);
2185
2186         for (i=0; i < conf->raid_disks; i++) {
2187                 struct md_rdev *rdev;
2188                 bio = r1_bio->bios[i];
2189
2190                 /* take from bio_init */
2191                 bio->bi_next = NULL;
2192                 bio->bi_flags &= ~(BIO_POOL_MASK-1);
2193                 bio->bi_flags |= 1 << BIO_UPTODATE;
2194                 bio->bi_rw = READ;
2195                 bio->bi_vcnt = 0;
2196                 bio->bi_idx = 0;
2197                 bio->bi_phys_segments = 0;
2198                 bio->bi_size = 0;
2199                 bio->bi_end_io = NULL;
2200                 bio->bi_private = NULL;
2201
2202                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2203                 if (rdev == NULL ||
2204                     test_bit(Faulty, &rdev->flags)) {
2205                         still_degraded = 1;
2206                 } else if (!test_bit(In_sync, &rdev->flags)) {
2207                         bio->bi_rw = WRITE;
2208                         bio->bi_end_io = end_sync_write;
2209                         write_targets ++;
2210                 } else {
2211                         /* may need to read from here */
2212                         sector_t first_bad = MaxSector;
2213                         int bad_sectors;
2214
2215                         if (is_badblock(rdev, sector_nr, good_sectors,
2216                                         &first_bad, &bad_sectors)) {
2217                                 if (first_bad > sector_nr)
2218                                         good_sectors = first_bad - sector_nr;
2219                                 else {
2220                                         bad_sectors -= (sector_nr - first_bad);
2221                                         if (min_bad == 0 ||
2222                                             min_bad > bad_sectors)
2223                                                 min_bad = bad_sectors;
2224                                 }
2225                         }
2226                         if (sector_nr < first_bad) {
2227                                 if (test_bit(WriteMostly, &rdev->flags)) {
2228                                         if (wonly < 0)
2229                                                 wonly = i;
2230                                 } else {
2231                                         if (disk < 0)
2232                                                 disk = i;
2233                                 }
2234                                 bio->bi_rw = READ;
2235                                 bio->bi_end_io = end_sync_read;
2236                                 read_targets++;
2237                         }
2238                 }
2239                 if (bio->bi_end_io) {
2240                         atomic_inc(&rdev->nr_pending);
2241                         bio->bi_sector = sector_nr + rdev->data_offset;
2242                         bio->bi_bdev = rdev->bdev;
2243                         bio->bi_private = r1_bio;
2244                 }
2245         }
2246         rcu_read_unlock();
2247         if (disk < 0)
2248                 disk = wonly;
2249         r1_bio->read_disk = disk;
2250
2251         if (read_targets == 0 && min_bad > 0) {
2252                 /* These sectors are bad on all InSync devices, so we
2253                  * need to mark them bad on all write targets
2254                  */
2255                 int ok = 1;
2256                 for (i = 0 ; i < conf->raid_disks ; i++)
2257                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2258                                 struct md_rdev *rdev =
2259                                         rcu_dereference(conf->mirrors[i].rdev);
2260                                 ok = rdev_set_badblocks(rdev, sector_nr,
2261                                                         min_bad, 0
2262                                         ) && ok;
2263                         }
2264                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2265                 *skipped = 1;
2266                 put_buf(r1_bio);
2267
2268                 if (!ok) {
2269                         /* Cannot record the badblocks, so need to
2270                          * abort the resync.
2271                          * If there are multiple read targets, could just
2272                          * fail the really bad ones ???
2273                          */
2274                         conf->recovery_disabled = mddev->recovery_disabled;
2275                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2276                         return 0;
2277                 } else
2278                         return min_bad;
2279
2280         }
2281         if (min_bad > 0 && min_bad < good_sectors) {
2282                 /* only resync enough to reach the next bad->good
2283                  * transition */
2284                 good_sectors = min_bad;
2285         }
2286
2287         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2288                 /* extra read targets are also write targets */
2289                 write_targets += read_targets-1;
2290
2291         if (write_targets == 0 || read_targets == 0) {
2292                 /* There is nowhere to write, so all non-sync
2293                  * drives must be failed - so we are finished
2294                  */
2295                 sector_t rv = max_sector - sector_nr;
2296                 *skipped = 1;
2297                 put_buf(r1_bio);
2298                 return rv;
2299         }
2300
2301         if (max_sector > mddev->resync_max)
2302                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2303         if (max_sector > sector_nr + good_sectors)
2304                 max_sector = sector_nr + good_sectors;
2305         nr_sectors = 0;
2306         sync_blocks = 0;
2307         do {
2308                 struct page *page;
2309                 int len = PAGE_SIZE;
2310                 if (sector_nr + (len>>9) > max_sector)
2311                         len = (max_sector - sector_nr) << 9;
2312                 if (len == 0)
2313                         break;
2314                 if (sync_blocks == 0) {
2315                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2316                                                &sync_blocks, still_degraded) &&
2317                             !conf->fullsync &&
2318                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2319                                 break;
2320                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2321                         if ((len >> 9) > sync_blocks)
2322                                 len = sync_blocks<<9;
2323                 }
2324
2325                 for (i=0 ; i < conf->raid_disks; i++) {
2326                         bio = r1_bio->bios[i];
2327                         if (bio->bi_end_io) {
2328                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2329                                 if (bio_add_page(bio, page, len, 0) == 0) {
2330                                         /* stop here */
2331                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2332                                         while (i > 0) {
2333                                                 i--;
2334                                                 bio = r1_bio->bios[i];
2335                                                 if (bio->bi_end_io==NULL)
2336                                                         continue;
2337                                                 /* remove last page from this bio */
2338                                                 bio->bi_vcnt--;
2339                                                 bio->bi_size -= len;
2340                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2341                                         }
2342                                         goto bio_full;
2343                                 }
2344                         }
2345                 }
2346                 nr_sectors += len>>9;
2347                 sector_nr += len>>9;
2348                 sync_blocks -= (len>>9);
2349         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2350  bio_full:
2351         r1_bio->sectors = nr_sectors;
2352
2353         /* For a user-requested sync, we read all readable devices and do a
2354          * compare
2355          */
2356         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2357                 atomic_set(&r1_bio->remaining, read_targets);
2358                 for (i=0; i<conf->raid_disks; i++) {
2359                         bio = r1_bio->bios[i];
2360                         if (bio->bi_end_io == end_sync_read) {
2361                                 md_sync_acct(bio->bi_bdev, nr_sectors);
2362                                 generic_make_request(bio);
2363                         }
2364                 }
2365         } else {
2366                 atomic_set(&r1_bio->remaining, 1);
2367                 bio = r1_bio->bios[r1_bio->read_disk];
2368                 md_sync_acct(bio->bi_bdev, nr_sectors);
2369                 generic_make_request(bio);
2370
2371         }
2372         return nr_sectors;
2373 }
2374
2375 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2376 {
2377         if (sectors)
2378                 return sectors;
2379
2380         return mddev->dev_sectors;
2381 }
2382
2383 static struct r1conf *setup_conf(struct mddev *mddev)
2384 {
2385         struct r1conf *conf;
2386         int i;
2387         struct mirror_info *disk;
2388         struct md_rdev *rdev;
2389         int err = -ENOMEM;
2390
2391         conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2392         if (!conf)
2393                 goto abort;
2394
2395         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2396                                  GFP_KERNEL);
2397         if (!conf->mirrors)
2398                 goto abort;
2399
2400         conf->tmppage = alloc_page(GFP_KERNEL);
2401         if (!conf->tmppage)
2402                 goto abort;
2403
2404         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2405         if (!conf->poolinfo)
2406                 goto abort;
2407         conf->poolinfo->raid_disks = mddev->raid_disks;
2408         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2409                                           r1bio_pool_free,
2410                                           conf->poolinfo);
2411         if (!conf->r1bio_pool)
2412                 goto abort;
2413
2414         conf->poolinfo->mddev = mddev;
2415
2416         spin_lock_init(&conf->device_lock);
2417         list_for_each_entry(rdev, &mddev->disks, same_set) {
2418                 int disk_idx = rdev->raid_disk;
2419                 if (disk_idx >= mddev->raid_disks
2420                     || disk_idx < 0)
2421                         continue;
2422                 disk = conf->mirrors + disk_idx;
2423
2424                 disk->rdev = rdev;
2425
2426                 disk->head_position = 0;
2427         }
2428         conf->raid_disks = mddev->raid_disks;
2429         conf->mddev = mddev;
2430         INIT_LIST_HEAD(&conf->retry_list);
2431
2432         spin_lock_init(&conf->resync_lock);
2433         init_waitqueue_head(&conf->wait_barrier);
2434
2435         bio_list_init(&conf->pending_bio_list);
2436         conf->pending_count = 0;
2437         conf->recovery_disabled = mddev->recovery_disabled - 1;
2438
2439         conf->last_used = -1;
2440         for (i = 0; i < conf->raid_disks; i++) {
2441
2442                 disk = conf->mirrors + i;
2443
2444                 if (!disk->rdev ||
2445                     !test_bit(In_sync, &disk->rdev->flags)) {
2446                         disk->head_position = 0;
2447                         if (disk->rdev)
2448                                 conf->fullsync = 1;
2449                 } else if (conf->last_used < 0)
2450                         /*
2451                          * The first working device is used as a
2452                          * starting point to read balancing.
2453                          */
2454                         conf->last_used = i;
2455         }
2456
2457         err = -EIO;
2458         if (conf->last_used < 0) {
2459                 printk(KERN_ERR "md/raid1:%s: no operational mirrors\n",
2460                        mdname(mddev));
2461                 goto abort;
2462         }
2463         err = -ENOMEM;
2464         conf->thread = md_register_thread(raid1d, mddev, NULL);
2465         if (!conf->thread) {
2466                 printk(KERN_ERR
2467                        "md/raid1:%s: couldn't allocate thread\n",
2468                        mdname(mddev));
2469                 goto abort;
2470         }
2471
2472         return conf;
2473
2474  abort:
2475         if (conf) {
2476                 if (conf->r1bio_pool)
2477                         mempool_destroy(conf->r1bio_pool);
2478                 kfree(conf->mirrors);
2479                 safe_put_page(conf->tmppage);
2480                 kfree(conf->poolinfo);
2481                 kfree(conf);
2482         }
2483         return ERR_PTR(err);
2484 }
2485
2486 static int run(struct mddev *mddev)
2487 {
2488         struct r1conf *conf;
2489         int i;
2490         struct md_rdev *rdev;
2491
2492         if (mddev->level != 1) {
2493                 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2494                        mdname(mddev), mddev->level);
2495                 return -EIO;
2496         }
2497         if (mddev->reshape_position != MaxSector) {
2498                 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2499                        mdname(mddev));
2500                 return -EIO;
2501         }
2502         /*
2503          * copy the already verified devices into our private RAID1
2504          * bookkeeping area. [whatever we allocate in run(),
2505          * should be freed in stop()]
2506          */
2507         if (mddev->private == NULL)
2508                 conf = setup_conf(mddev);
2509         else
2510                 conf = mddev->private;
2511
2512         if (IS_ERR(conf))
2513                 return PTR_ERR(conf);
2514
2515         list_for_each_entry(rdev, &mddev->disks, same_set) {
2516                 if (!mddev->gendisk)
2517                         continue;
2518                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2519                                   rdev->data_offset << 9);
2520                 /* as we don't honour merge_bvec_fn, we must never risk
2521                  * violating it, so limit ->max_segments to 1 lying within
2522                  * a single page, as a one page request is never in violation.
2523                  */
2524                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2525                         blk_queue_max_segments(mddev->queue, 1);
2526                         blk_queue_segment_boundary(mddev->queue,
2527                                                    PAGE_CACHE_SIZE - 1);
2528                 }
2529         }
2530
2531         mddev->degraded = 0;
2532         for (i=0; i < conf->raid_disks; i++)
2533                 if (conf->mirrors[i].rdev == NULL ||
2534                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2535                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2536                         mddev->degraded++;
2537
2538         if (conf->raid_disks - mddev->degraded == 1)
2539                 mddev->recovery_cp = MaxSector;
2540
2541         if (mddev->recovery_cp != MaxSector)
2542                 printk(KERN_NOTICE "md/raid1:%s: not clean"
2543                        " -- starting background reconstruction\n",
2544                        mdname(mddev));
2545         printk(KERN_INFO 
2546                 "md/raid1:%s: active with %d out of %d mirrors\n",
2547                 mdname(mddev), mddev->raid_disks - mddev->degraded, 
2548                 mddev->raid_disks);
2549
2550         /*
2551          * Ok, everything is just fine now
2552          */
2553         mddev->thread = conf->thread;
2554         conf->thread = NULL;
2555         mddev->private = conf;
2556
2557         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2558
2559         if (mddev->queue) {
2560                 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2561                 mddev->queue->backing_dev_info.congested_data = mddev;
2562         }
2563         return md_integrity_register(mddev);
2564 }
2565
2566 static int stop(struct mddev *mddev)
2567 {
2568         struct r1conf *conf = mddev->private;
2569         struct bitmap *bitmap = mddev->bitmap;
2570
2571         /* wait for behind writes to complete */
2572         if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2573                 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2574                        mdname(mddev));
2575                 /* need to kick something here to make sure I/O goes? */
2576                 wait_event(bitmap->behind_wait,
2577                            atomic_read(&bitmap->behind_writes) == 0);
2578         }
2579
2580         raise_barrier(conf);
2581         lower_barrier(conf);
2582
2583         md_unregister_thread(&mddev->thread);
2584         if (conf->r1bio_pool)
2585                 mempool_destroy(conf->r1bio_pool);
2586         kfree(conf->mirrors);
2587         kfree(conf->poolinfo);
2588         kfree(conf);
2589         mddev->private = NULL;
2590         return 0;
2591 }
2592
2593 static int raid1_resize(struct mddev *mddev, sector_t sectors)
2594 {
2595         /* no resync is happening, and there is enough space
2596          * on all devices, so we can resize.
2597          * We need to make sure resync covers any new space.
2598          * If the array is shrinking we should possibly wait until
2599          * any io in the removed space completes, but it hardly seems
2600          * worth it.
2601          */
2602         md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2603         if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2604                 return -EINVAL;
2605         set_capacity(mddev->gendisk, mddev->array_sectors);
2606         revalidate_disk(mddev->gendisk);
2607         if (sectors > mddev->dev_sectors &&
2608             mddev->recovery_cp > mddev->dev_sectors) {
2609                 mddev->recovery_cp = mddev->dev_sectors;
2610                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2611         }
2612         mddev->dev_sectors = sectors;
2613         mddev->resync_max_sectors = sectors;
2614         return 0;
2615 }
2616
2617 static int raid1_reshape(struct mddev *mddev)
2618 {
2619         /* We need to:
2620          * 1/ resize the r1bio_pool
2621          * 2/ resize conf->mirrors
2622          *
2623          * We allocate a new r1bio_pool if we can.
2624          * Then raise a device barrier and wait until all IO stops.
2625          * Then resize conf->mirrors and swap in the new r1bio pool.
2626          *
2627          * At the same time, we "pack" the devices so that all the missing
2628          * devices have the higher raid_disk numbers.
2629          */
2630         mempool_t *newpool, *oldpool;
2631         struct pool_info *newpoolinfo;
2632         struct mirror_info *newmirrors;
2633         struct r1conf *conf = mddev->private;
2634         int cnt, raid_disks;
2635         unsigned long flags;
2636         int d, d2, err;
2637
2638         /* Cannot change chunk_size, layout, or level */
2639         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2640             mddev->layout != mddev->new_layout ||
2641             mddev->level != mddev->new_level) {
2642                 mddev->new_chunk_sectors = mddev->chunk_sectors;
2643                 mddev->new_layout = mddev->layout;
2644                 mddev->new_level = mddev->level;
2645                 return -EINVAL;
2646         }
2647
2648         err = md_allow_write(mddev);
2649         if (err)
2650                 return err;
2651
2652         raid_disks = mddev->raid_disks + mddev->delta_disks;
2653
2654         if (raid_disks < conf->raid_disks) {
2655                 cnt=0;
2656                 for (d= 0; d < conf->raid_disks; d++)
2657                         if (conf->mirrors[d].rdev)
2658                                 cnt++;
2659                 if (cnt > raid_disks)
2660                         return -EBUSY;
2661         }
2662
2663         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2664         if (!newpoolinfo)
2665                 return -ENOMEM;
2666         newpoolinfo->mddev = mddev;
2667         newpoolinfo->raid_disks = raid_disks;
2668
2669         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2670                                  r1bio_pool_free, newpoolinfo);
2671         if (!newpool) {
2672                 kfree(newpoolinfo);
2673                 return -ENOMEM;
2674         }
2675         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2676         if (!newmirrors) {
2677                 kfree(newpoolinfo);
2678                 mempool_destroy(newpool);
2679                 return -ENOMEM;
2680         }
2681
2682         raise_barrier(conf);
2683
2684         /* ok, everything is stopped */
2685         oldpool = conf->r1bio_pool;
2686         conf->r1bio_pool = newpool;
2687
2688         for (d = d2 = 0; d < conf->raid_disks; d++) {
2689                 struct md_rdev *rdev = conf->mirrors[d].rdev;
2690                 if (rdev && rdev->raid_disk != d2) {
2691                         sysfs_unlink_rdev(mddev, rdev);
2692                         rdev->raid_disk = d2;
2693                         sysfs_unlink_rdev(mddev, rdev);
2694                         if (sysfs_link_rdev(mddev, rdev))
2695                                 printk(KERN_WARNING
2696                                        "md/raid1:%s: cannot register rd%d\n",
2697                                        mdname(mddev), rdev->raid_disk);
2698                 }
2699                 if (rdev)
2700                         newmirrors[d2++].rdev = rdev;
2701         }
2702         kfree(conf->mirrors);
2703         conf->mirrors = newmirrors;
2704         kfree(conf->poolinfo);
2705         conf->poolinfo = newpoolinfo;
2706
2707         spin_lock_irqsave(&conf->device_lock, flags);
2708         mddev->degraded += (raid_disks - conf->raid_disks);
2709         spin_unlock_irqrestore(&conf->device_lock, flags);
2710         conf->raid_disks = mddev->raid_disks = raid_disks;
2711         mddev->delta_disks = 0;
2712
2713         conf->last_used = 0; /* just make sure it is in-range */
2714         lower_barrier(conf);
2715
2716         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2717         md_wakeup_thread(mddev->thread);
2718
2719         mempool_destroy(oldpool);
2720         return 0;
2721 }
2722
2723 static void raid1_quiesce(struct mddev *mddev, int state)
2724 {
2725         struct r1conf *conf = mddev->private;
2726
2727         switch(state) {
2728         case 2: /* wake for suspend */
2729                 wake_up(&conf->wait_barrier);
2730                 break;
2731         case 1:
2732                 raise_barrier(conf);
2733                 break;
2734         case 0:
2735                 lower_barrier(conf);
2736                 break;
2737         }
2738 }
2739
2740 static void *raid1_takeover(struct mddev *mddev)
2741 {
2742         /* raid1 can take over:
2743          *  raid5 with 2 devices, any layout or chunk size
2744          */
2745         if (mddev->level == 5 && mddev->raid_disks == 2) {
2746                 struct r1conf *conf;
2747                 mddev->new_level = 1;
2748                 mddev->new_layout = 0;
2749                 mddev->new_chunk_sectors = 0;
2750                 conf = setup_conf(mddev);
2751                 if (!IS_ERR(conf))
2752                         conf->barrier = 1;
2753                 return conf;
2754         }
2755         return ERR_PTR(-EINVAL);
2756 }
2757
2758 static struct md_personality raid1_personality =
2759 {
2760         .name           = "raid1",
2761         .level          = 1,
2762         .owner          = THIS_MODULE,
2763         .make_request   = make_request,
2764         .run            = run,
2765         .stop           = stop,
2766         .status         = status,
2767         .error_handler  = error,
2768         .hot_add_disk   = raid1_add_disk,
2769         .hot_remove_disk= raid1_remove_disk,
2770         .spare_active   = raid1_spare_active,
2771         .sync_request   = sync_request,
2772         .resize         = raid1_resize,
2773         .size           = raid1_size,
2774         .check_reshape  = raid1_reshape,
2775         .quiesce        = raid1_quiesce,
2776         .takeover       = raid1_takeover,
2777 };
2778
2779 static int __init raid_init(void)
2780 {
2781         return register_md_personality(&raid1_personality);
2782 }
2783
2784 static void raid_exit(void)
2785 {
2786         unregister_md_personality(&raid1_personality);
2787 }
2788
2789 module_init(raid_init);
2790 module_exit(raid_exit);
2791 MODULE_LICENSE("GPL");
2792 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2793 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2794 MODULE_ALIAS("md-raid1");
2795 MODULE_ALIAS("md-level-1");
2796
2797 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);