Merge branch 'for-3.6/drivers' of git://git.kernel.dk/linux-block
[linux-3.10.git] / drivers / md / raid1.c
1 /*
2  * raid1.c : Multiple Devices driver for Linux
3  *
4  * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
5  *
6  * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
7  *
8  * RAID-1 management functions.
9  *
10  * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
11  *
12  * Fixes to reconstruction by Jakob Ć˜stergaard" <jakob@ostenfeld.dk>
13  * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
14  *
15  * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
16  * bitmapped intelligence in resync:
17  *
18  *      - bitmap marked during normal i/o
19  *      - bitmap used to skip nondirty blocks during sync
20  *
21  * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
22  * - persistent bitmap code
23  *
24  * This program is free software; you can redistribute it and/or modify
25  * it under the terms of the GNU General Public License as published by
26  * the Free Software Foundation; either version 2, or (at your option)
27  * any later version.
28  *
29  * You should have received a copy of the GNU General Public License
30  * (for example /usr/src/linux/COPYING); if not, write to the Free
31  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
32  */
33
34 #include <linux/slab.h>
35 #include <linux/delay.h>
36 #include <linux/blkdev.h>
37 #include <linux/module.h>
38 #include <linux/seq_file.h>
39 #include <linux/ratelimit.h>
40 #include "md.h"
41 #include "raid1.h"
42 #include "bitmap.h"
43
44 /*
45  * Number of guaranteed r1bios in case of extreme VM load:
46  */
47 #define NR_RAID1_BIOS 256
48
49 /* when we get a read error on a read-only array, we redirect to another
50  * device without failing the first device, or trying to over-write to
51  * correct the read error.  To keep track of bad blocks on a per-bio
52  * level, we store IO_BLOCKED in the appropriate 'bios' pointer
53  */
54 #define IO_BLOCKED ((struct bio *)1)
55 /* When we successfully write to a known bad-block, we need to remove the
56  * bad-block marking which must be done from process context.  So we record
57  * the success by setting devs[n].bio to IO_MADE_GOOD
58  */
59 #define IO_MADE_GOOD ((struct bio *)2)
60
61 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
62
63 /* When there are this many requests queue to be written by
64  * the raid1 thread, we become 'congested' to provide back-pressure
65  * for writeback.
66  */
67 static int max_queued_requests = 1024;
68
69 static void allow_barrier(struct r1conf *conf);
70 static void lower_barrier(struct r1conf *conf);
71
72 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
73 {
74         struct pool_info *pi = data;
75         int size = offsetof(struct r1bio, bios[pi->raid_disks]);
76
77         /* allocate a r1bio with room for raid_disks entries in the bios array */
78         return kzalloc(size, gfp_flags);
79 }
80
81 static void r1bio_pool_free(void *r1_bio, void *data)
82 {
83         kfree(r1_bio);
84 }
85
86 #define RESYNC_BLOCK_SIZE (64*1024)
87 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
88 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
89 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
90 #define RESYNC_WINDOW (2048*1024)
91
92 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
93 {
94         struct pool_info *pi = data;
95         struct page *page;
96         struct r1bio *r1_bio;
97         struct bio *bio;
98         int i, j;
99
100         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
101         if (!r1_bio)
102                 return NULL;
103
104         /*
105          * Allocate bios : 1 for reading, n-1 for writing
106          */
107         for (j = pi->raid_disks ; j-- ; ) {
108                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
109                 if (!bio)
110                         goto out_free_bio;
111                 r1_bio->bios[j] = bio;
112         }
113         /*
114          * Allocate RESYNC_PAGES data pages and attach them to
115          * the first bio.
116          * If this is a user-requested check/repair, allocate
117          * RESYNC_PAGES for each bio.
118          */
119         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
120                 j = pi->raid_disks;
121         else
122                 j = 1;
123         while(j--) {
124                 bio = r1_bio->bios[j];
125                 for (i = 0; i < RESYNC_PAGES; i++) {
126                         page = alloc_page(gfp_flags);
127                         if (unlikely(!page))
128                                 goto out_free_pages;
129
130                         bio->bi_io_vec[i].bv_page = page;
131                         bio->bi_vcnt = i+1;
132                 }
133         }
134         /* If not user-requests, copy the page pointers to all bios */
135         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
136                 for (i=0; i<RESYNC_PAGES ; i++)
137                         for (j=1; j<pi->raid_disks; j++)
138                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
139                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
140         }
141
142         r1_bio->master_bio = NULL;
143
144         return r1_bio;
145
146 out_free_pages:
147         for (j=0 ; j < pi->raid_disks; j++)
148                 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
149                         put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
150         j = -1;
151 out_free_bio:
152         while (++j < pi->raid_disks)
153                 bio_put(r1_bio->bios[j]);
154         r1bio_pool_free(r1_bio, data);
155         return NULL;
156 }
157
158 static void r1buf_pool_free(void *__r1_bio, void *data)
159 {
160         struct pool_info *pi = data;
161         int i,j;
162         struct r1bio *r1bio = __r1_bio;
163
164         for (i = 0; i < RESYNC_PAGES; i++)
165                 for (j = pi->raid_disks; j-- ;) {
166                         if (j == 0 ||
167                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
168                             r1bio->bios[0]->bi_io_vec[i].bv_page)
169                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
170                 }
171         for (i=0 ; i < pi->raid_disks; i++)
172                 bio_put(r1bio->bios[i]);
173
174         r1bio_pool_free(r1bio, data);
175 }
176
177 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
178 {
179         int i;
180
181         for (i = 0; i < conf->raid_disks * 2; i++) {
182                 struct bio **bio = r1_bio->bios + i;
183                 if (!BIO_SPECIAL(*bio))
184                         bio_put(*bio);
185                 *bio = NULL;
186         }
187 }
188
189 static void free_r1bio(struct r1bio *r1_bio)
190 {
191         struct r1conf *conf = r1_bio->mddev->private;
192
193         put_all_bios(conf, r1_bio);
194         mempool_free(r1_bio, conf->r1bio_pool);
195 }
196
197 static void put_buf(struct r1bio *r1_bio)
198 {
199         struct r1conf *conf = r1_bio->mddev->private;
200         int i;
201
202         for (i = 0; i < conf->raid_disks * 2; i++) {
203                 struct bio *bio = r1_bio->bios[i];
204                 if (bio->bi_end_io)
205                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
206         }
207
208         mempool_free(r1_bio, conf->r1buf_pool);
209
210         lower_barrier(conf);
211 }
212
213 static void reschedule_retry(struct r1bio *r1_bio)
214 {
215         unsigned long flags;
216         struct mddev *mddev = r1_bio->mddev;
217         struct r1conf *conf = mddev->private;
218
219         spin_lock_irqsave(&conf->device_lock, flags);
220         list_add(&r1_bio->retry_list, &conf->retry_list);
221         conf->nr_queued ++;
222         spin_unlock_irqrestore(&conf->device_lock, flags);
223
224         wake_up(&conf->wait_barrier);
225         md_wakeup_thread(mddev->thread);
226 }
227
228 /*
229  * raid_end_bio_io() is called when we have finished servicing a mirrored
230  * operation and are ready to return a success/failure code to the buffer
231  * cache layer.
232  */
233 static void call_bio_endio(struct r1bio *r1_bio)
234 {
235         struct bio *bio = r1_bio->master_bio;
236         int done;
237         struct r1conf *conf = r1_bio->mddev->private;
238
239         if (bio->bi_phys_segments) {
240                 unsigned long flags;
241                 spin_lock_irqsave(&conf->device_lock, flags);
242                 bio->bi_phys_segments--;
243                 done = (bio->bi_phys_segments == 0);
244                 spin_unlock_irqrestore(&conf->device_lock, flags);
245         } else
246                 done = 1;
247
248         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
249                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
250         if (done) {
251                 bio_endio(bio, 0);
252                 /*
253                  * Wake up any possible resync thread that waits for the device
254                  * to go idle.
255                  */
256                 allow_barrier(conf);
257         }
258 }
259
260 static void raid_end_bio_io(struct r1bio *r1_bio)
261 {
262         struct bio *bio = r1_bio->master_bio;
263
264         /* if nobody has done the final endio yet, do it now */
265         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
266                 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
267                          (bio_data_dir(bio) == WRITE) ? "write" : "read",
268                          (unsigned long long) bio->bi_sector,
269                          (unsigned long long) bio->bi_sector +
270                          (bio->bi_size >> 9) - 1);
271
272                 call_bio_endio(r1_bio);
273         }
274         free_r1bio(r1_bio);
275 }
276
277 /*
278  * Update disk head position estimator based on IRQ completion info.
279  */
280 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
281 {
282         struct r1conf *conf = r1_bio->mddev->private;
283
284         conf->mirrors[disk].head_position =
285                 r1_bio->sector + (r1_bio->sectors);
286 }
287
288 /*
289  * Find the disk number which triggered given bio
290  */
291 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
292 {
293         int mirror;
294         struct r1conf *conf = r1_bio->mddev->private;
295         int raid_disks = conf->raid_disks;
296
297         for (mirror = 0; mirror < raid_disks * 2; mirror++)
298                 if (r1_bio->bios[mirror] == bio)
299                         break;
300
301         BUG_ON(mirror == raid_disks * 2);
302         update_head_pos(mirror, r1_bio);
303
304         return mirror;
305 }
306
307 static void raid1_end_read_request(struct bio *bio, int error)
308 {
309         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
310         struct r1bio *r1_bio = bio->bi_private;
311         int mirror;
312         struct r1conf *conf = r1_bio->mddev->private;
313
314         mirror = r1_bio->read_disk;
315         /*
316          * this branch is our 'one mirror IO has finished' event handler:
317          */
318         update_head_pos(mirror, r1_bio);
319
320         if (uptodate)
321                 set_bit(R1BIO_Uptodate, &r1_bio->state);
322         else {
323                 /* If all other devices have failed, we want to return
324                  * the error upwards rather than fail the last device.
325                  * Here we redefine "uptodate" to mean "Don't want to retry"
326                  */
327                 unsigned long flags;
328                 spin_lock_irqsave(&conf->device_lock, flags);
329                 if (r1_bio->mddev->degraded == conf->raid_disks ||
330                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
331                      !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
332                         uptodate = 1;
333                 spin_unlock_irqrestore(&conf->device_lock, flags);
334         }
335
336         if (uptodate)
337                 raid_end_bio_io(r1_bio);
338         else {
339                 /*
340                  * oops, read error:
341                  */
342                 char b[BDEVNAME_SIZE];
343                 printk_ratelimited(
344                         KERN_ERR "md/raid1:%s: %s: "
345                         "rescheduling sector %llu\n",
346                         mdname(conf->mddev),
347                         bdevname(conf->mirrors[mirror].rdev->bdev,
348                                  b),
349                         (unsigned long long)r1_bio->sector);
350                 set_bit(R1BIO_ReadError, &r1_bio->state);
351                 reschedule_retry(r1_bio);
352         }
353
354         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
355 }
356
357 static void close_write(struct r1bio *r1_bio)
358 {
359         /* it really is the end of this request */
360         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
361                 /* free extra copy of the data pages */
362                 int i = r1_bio->behind_page_count;
363                 while (i--)
364                         safe_put_page(r1_bio->behind_bvecs[i].bv_page);
365                 kfree(r1_bio->behind_bvecs);
366                 r1_bio->behind_bvecs = NULL;
367         }
368         /* clear the bitmap if all writes complete successfully */
369         bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
370                         r1_bio->sectors,
371                         !test_bit(R1BIO_Degraded, &r1_bio->state),
372                         test_bit(R1BIO_BehindIO, &r1_bio->state));
373         md_write_end(r1_bio->mddev);
374 }
375
376 static void r1_bio_write_done(struct r1bio *r1_bio)
377 {
378         if (!atomic_dec_and_test(&r1_bio->remaining))
379                 return;
380
381         if (test_bit(R1BIO_WriteError, &r1_bio->state))
382                 reschedule_retry(r1_bio);
383         else {
384                 close_write(r1_bio);
385                 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
386                         reschedule_retry(r1_bio);
387                 else
388                         raid_end_bio_io(r1_bio);
389         }
390 }
391
392 static void raid1_end_write_request(struct bio *bio, int error)
393 {
394         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
395         struct r1bio *r1_bio = bio->bi_private;
396         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
397         struct r1conf *conf = r1_bio->mddev->private;
398         struct bio *to_put = NULL;
399
400         mirror = find_bio_disk(r1_bio, bio);
401
402         /*
403          * 'one mirror IO has finished' event handler:
404          */
405         if (!uptodate) {
406                 set_bit(WriteErrorSeen,
407                         &conf->mirrors[mirror].rdev->flags);
408                 if (!test_and_set_bit(WantReplacement,
409                                       &conf->mirrors[mirror].rdev->flags))
410                         set_bit(MD_RECOVERY_NEEDED, &
411                                 conf->mddev->recovery);
412
413                 set_bit(R1BIO_WriteError, &r1_bio->state);
414         } else {
415                 /*
416                  * Set R1BIO_Uptodate in our master bio, so that we
417                  * will return a good error code for to the higher
418                  * levels even if IO on some other mirrored buffer
419                  * fails.
420                  *
421                  * The 'master' represents the composite IO operation
422                  * to user-side. So if something waits for IO, then it
423                  * will wait for the 'master' bio.
424                  */
425                 sector_t first_bad;
426                 int bad_sectors;
427
428                 r1_bio->bios[mirror] = NULL;
429                 to_put = bio;
430                 set_bit(R1BIO_Uptodate, &r1_bio->state);
431
432                 /* Maybe we can clear some bad blocks. */
433                 if (is_badblock(conf->mirrors[mirror].rdev,
434                                 r1_bio->sector, r1_bio->sectors,
435                                 &first_bad, &bad_sectors)) {
436                         r1_bio->bios[mirror] = IO_MADE_GOOD;
437                         set_bit(R1BIO_MadeGood, &r1_bio->state);
438                 }
439         }
440
441         if (behind) {
442                 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
443                         atomic_dec(&r1_bio->behind_remaining);
444
445                 /*
446                  * In behind mode, we ACK the master bio once the I/O
447                  * has safely reached all non-writemostly
448                  * disks. Setting the Returned bit ensures that this
449                  * gets done only once -- we don't ever want to return
450                  * -EIO here, instead we'll wait
451                  */
452                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
453                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
454                         /* Maybe we can return now */
455                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
456                                 struct bio *mbio = r1_bio->master_bio;
457                                 pr_debug("raid1: behind end write sectors"
458                                          " %llu-%llu\n",
459                                          (unsigned long long) mbio->bi_sector,
460                                          (unsigned long long) mbio->bi_sector +
461                                          (mbio->bi_size >> 9) - 1);
462                                 call_bio_endio(r1_bio);
463                         }
464                 }
465         }
466         if (r1_bio->bios[mirror] == NULL)
467                 rdev_dec_pending(conf->mirrors[mirror].rdev,
468                                  conf->mddev);
469
470         /*
471          * Let's see if all mirrored write operations have finished
472          * already.
473          */
474         r1_bio_write_done(r1_bio);
475
476         if (to_put)
477                 bio_put(to_put);
478 }
479
480
481 /*
482  * This routine returns the disk from which the requested read should
483  * be done. There is a per-array 'next expected sequential IO' sector
484  * number - if this matches on the next IO then we use the last disk.
485  * There is also a per-disk 'last know head position' sector that is
486  * maintained from IRQ contexts, both the normal and the resync IO
487  * completion handlers update this position correctly. If there is no
488  * perfect sequential match then we pick the disk whose head is closest.
489  *
490  * If there are 2 mirrors in the same 2 devices, performance degrades
491  * because position is mirror, not device based.
492  *
493  * The rdev for the device selected will have nr_pending incremented.
494  */
495 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
496 {
497         const sector_t this_sector = r1_bio->sector;
498         int sectors;
499         int best_good_sectors;
500         int best_disk, best_dist_disk, best_pending_disk;
501         int has_nonrot_disk;
502         int disk;
503         sector_t best_dist;
504         unsigned int min_pending;
505         struct md_rdev *rdev;
506         int choose_first;
507         int choose_next_idle;
508
509         rcu_read_lock();
510         /*
511          * Check if we can balance. We can balance on the whole
512          * device if no resync is going on, or below the resync window.
513          * We take the first readable disk when above the resync window.
514          */
515  retry:
516         sectors = r1_bio->sectors;
517         best_disk = -1;
518         best_dist_disk = -1;
519         best_dist = MaxSector;
520         best_pending_disk = -1;
521         min_pending = UINT_MAX;
522         best_good_sectors = 0;
523         has_nonrot_disk = 0;
524         choose_next_idle = 0;
525
526         if (conf->mddev->recovery_cp < MaxSector &&
527             (this_sector + sectors >= conf->next_resync))
528                 choose_first = 1;
529         else
530                 choose_first = 0;
531
532         for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
533                 sector_t dist;
534                 sector_t first_bad;
535                 int bad_sectors;
536                 unsigned int pending;
537                 bool nonrot;
538
539                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
540                 if (r1_bio->bios[disk] == IO_BLOCKED
541                     || rdev == NULL
542                     || test_bit(Unmerged, &rdev->flags)
543                     || test_bit(Faulty, &rdev->flags))
544                         continue;
545                 if (!test_bit(In_sync, &rdev->flags) &&
546                     rdev->recovery_offset < this_sector + sectors)
547                         continue;
548                 if (test_bit(WriteMostly, &rdev->flags)) {
549                         /* Don't balance among write-mostly, just
550                          * use the first as a last resort */
551                         if (best_disk < 0) {
552                                 if (is_badblock(rdev, this_sector, sectors,
553                                                 &first_bad, &bad_sectors)) {
554                                         if (first_bad < this_sector)
555                                                 /* Cannot use this */
556                                                 continue;
557                                         best_good_sectors = first_bad - this_sector;
558                                 } else
559                                         best_good_sectors = sectors;
560                                 best_disk = disk;
561                         }
562                         continue;
563                 }
564                 /* This is a reasonable device to use.  It might
565                  * even be best.
566                  */
567                 if (is_badblock(rdev, this_sector, sectors,
568                                 &first_bad, &bad_sectors)) {
569                         if (best_dist < MaxSector)
570                                 /* already have a better device */
571                                 continue;
572                         if (first_bad <= this_sector) {
573                                 /* cannot read here. If this is the 'primary'
574                                  * device, then we must not read beyond
575                                  * bad_sectors from another device..
576                                  */
577                                 bad_sectors -= (this_sector - first_bad);
578                                 if (choose_first && sectors > bad_sectors)
579                                         sectors = bad_sectors;
580                                 if (best_good_sectors > sectors)
581                                         best_good_sectors = sectors;
582
583                         } else {
584                                 sector_t good_sectors = first_bad - this_sector;
585                                 if (good_sectors > best_good_sectors) {
586                                         best_good_sectors = good_sectors;
587                                         best_disk = disk;
588                                 }
589                                 if (choose_first)
590                                         break;
591                         }
592                         continue;
593                 } else
594                         best_good_sectors = sectors;
595
596                 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
597                 has_nonrot_disk |= nonrot;
598                 pending = atomic_read(&rdev->nr_pending);
599                 dist = abs(this_sector - conf->mirrors[disk].head_position);
600                 if (choose_first) {
601                         best_disk = disk;
602                         break;
603                 }
604                 /* Don't change to another disk for sequential reads */
605                 if (conf->mirrors[disk].next_seq_sect == this_sector
606                     || dist == 0) {
607                         int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
608                         struct raid1_info *mirror = &conf->mirrors[disk];
609
610                         best_disk = disk;
611                         /*
612                          * If buffered sequential IO size exceeds optimal
613                          * iosize, check if there is idle disk. If yes, choose
614                          * the idle disk. read_balance could already choose an
615                          * idle disk before noticing it's a sequential IO in
616                          * this disk. This doesn't matter because this disk
617                          * will idle, next time it will be utilized after the
618                          * first disk has IO size exceeds optimal iosize. In
619                          * this way, iosize of the first disk will be optimal
620                          * iosize at least. iosize of the second disk might be
621                          * small, but not a big deal since when the second disk
622                          * starts IO, the first disk is likely still busy.
623                          */
624                         if (nonrot && opt_iosize > 0 &&
625                             mirror->seq_start != MaxSector &&
626                             mirror->next_seq_sect > opt_iosize &&
627                             mirror->next_seq_sect - opt_iosize >=
628                             mirror->seq_start) {
629                                 choose_next_idle = 1;
630                                 continue;
631                         }
632                         break;
633                 }
634                 /* If device is idle, use it */
635                 if (pending == 0) {
636                         best_disk = disk;
637                         break;
638                 }
639
640                 if (choose_next_idle)
641                         continue;
642
643                 if (min_pending > pending) {
644                         min_pending = pending;
645                         best_pending_disk = disk;
646                 }
647
648                 if (dist < best_dist) {
649                         best_dist = dist;
650                         best_dist_disk = disk;
651                 }
652         }
653
654         /*
655          * If all disks are rotational, choose the closest disk. If any disk is
656          * non-rotational, choose the disk with less pending request even the
657          * disk is rotational, which might/might not be optimal for raids with
658          * mixed ratation/non-rotational disks depending on workload.
659          */
660         if (best_disk == -1) {
661                 if (has_nonrot_disk)
662                         best_disk = best_pending_disk;
663                 else
664                         best_disk = best_dist_disk;
665         }
666
667         if (best_disk >= 0) {
668                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
669                 if (!rdev)
670                         goto retry;
671                 atomic_inc(&rdev->nr_pending);
672                 if (test_bit(Faulty, &rdev->flags)) {
673                         /* cannot risk returning a device that failed
674                          * before we inc'ed nr_pending
675                          */
676                         rdev_dec_pending(rdev, conf->mddev);
677                         goto retry;
678                 }
679                 sectors = best_good_sectors;
680
681                 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
682                         conf->mirrors[best_disk].seq_start = this_sector;
683
684                 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
685         }
686         rcu_read_unlock();
687         *max_sectors = sectors;
688
689         return best_disk;
690 }
691
692 static int raid1_mergeable_bvec(struct request_queue *q,
693                                 struct bvec_merge_data *bvm,
694                                 struct bio_vec *biovec)
695 {
696         struct mddev *mddev = q->queuedata;
697         struct r1conf *conf = mddev->private;
698         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
699         int max = biovec->bv_len;
700
701         if (mddev->merge_check_needed) {
702                 int disk;
703                 rcu_read_lock();
704                 for (disk = 0; disk < conf->raid_disks * 2; disk++) {
705                         struct md_rdev *rdev = rcu_dereference(
706                                 conf->mirrors[disk].rdev);
707                         if (rdev && !test_bit(Faulty, &rdev->flags)) {
708                                 struct request_queue *q =
709                                         bdev_get_queue(rdev->bdev);
710                                 if (q->merge_bvec_fn) {
711                                         bvm->bi_sector = sector +
712                                                 rdev->data_offset;
713                                         bvm->bi_bdev = rdev->bdev;
714                                         max = min(max, q->merge_bvec_fn(
715                                                           q, bvm, biovec));
716                                 }
717                         }
718                 }
719                 rcu_read_unlock();
720         }
721         return max;
722
723 }
724
725 int md_raid1_congested(struct mddev *mddev, int bits)
726 {
727         struct r1conf *conf = mddev->private;
728         int i, ret = 0;
729
730         if ((bits & (1 << BDI_async_congested)) &&
731             conf->pending_count >= max_queued_requests)
732                 return 1;
733
734         rcu_read_lock();
735         for (i = 0; i < conf->raid_disks * 2; i++) {
736                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
737                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
738                         struct request_queue *q = bdev_get_queue(rdev->bdev);
739
740                         BUG_ON(!q);
741
742                         /* Note the '|| 1' - when read_balance prefers
743                          * non-congested targets, it can be removed
744                          */
745                         if ((bits & (1<<BDI_async_congested)) || 1)
746                                 ret |= bdi_congested(&q->backing_dev_info, bits);
747                         else
748                                 ret &= bdi_congested(&q->backing_dev_info, bits);
749                 }
750         }
751         rcu_read_unlock();
752         return ret;
753 }
754 EXPORT_SYMBOL_GPL(md_raid1_congested);
755
756 static int raid1_congested(void *data, int bits)
757 {
758         struct mddev *mddev = data;
759
760         return mddev_congested(mddev, bits) ||
761                 md_raid1_congested(mddev, bits);
762 }
763
764 static void flush_pending_writes(struct r1conf *conf)
765 {
766         /* Any writes that have been queued but are awaiting
767          * bitmap updates get flushed here.
768          */
769         spin_lock_irq(&conf->device_lock);
770
771         if (conf->pending_bio_list.head) {
772                 struct bio *bio;
773                 bio = bio_list_get(&conf->pending_bio_list);
774                 conf->pending_count = 0;
775                 spin_unlock_irq(&conf->device_lock);
776                 /* flush any pending bitmap writes to
777                  * disk before proceeding w/ I/O */
778                 bitmap_unplug(conf->mddev->bitmap);
779                 wake_up(&conf->wait_barrier);
780
781                 while (bio) { /* submit pending writes */
782                         struct bio *next = bio->bi_next;
783                         bio->bi_next = NULL;
784                         generic_make_request(bio);
785                         bio = next;
786                 }
787         } else
788                 spin_unlock_irq(&conf->device_lock);
789 }
790
791 /* Barriers....
792  * Sometimes we need to suspend IO while we do something else,
793  * either some resync/recovery, or reconfigure the array.
794  * To do this we raise a 'barrier'.
795  * The 'barrier' is a counter that can be raised multiple times
796  * to count how many activities are happening which preclude
797  * normal IO.
798  * We can only raise the barrier if there is no pending IO.
799  * i.e. if nr_pending == 0.
800  * We choose only to raise the barrier if no-one is waiting for the
801  * barrier to go down.  This means that as soon as an IO request
802  * is ready, no other operations which require a barrier will start
803  * until the IO request has had a chance.
804  *
805  * So: regular IO calls 'wait_barrier'.  When that returns there
806  *    is no backgroup IO happening,  It must arrange to call
807  *    allow_barrier when it has finished its IO.
808  * backgroup IO calls must call raise_barrier.  Once that returns
809  *    there is no normal IO happeing.  It must arrange to call
810  *    lower_barrier when the particular background IO completes.
811  */
812 #define RESYNC_DEPTH 32
813
814 static void raise_barrier(struct r1conf *conf)
815 {
816         spin_lock_irq(&conf->resync_lock);
817
818         /* Wait until no block IO is waiting */
819         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
820                             conf->resync_lock, );
821
822         /* block any new IO from starting */
823         conf->barrier++;
824
825         /* Now wait for all pending IO to complete */
826         wait_event_lock_irq(conf->wait_barrier,
827                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
828                             conf->resync_lock, );
829
830         spin_unlock_irq(&conf->resync_lock);
831 }
832
833 static void lower_barrier(struct r1conf *conf)
834 {
835         unsigned long flags;
836         BUG_ON(conf->barrier <= 0);
837         spin_lock_irqsave(&conf->resync_lock, flags);
838         conf->barrier--;
839         spin_unlock_irqrestore(&conf->resync_lock, flags);
840         wake_up(&conf->wait_barrier);
841 }
842
843 static void wait_barrier(struct r1conf *conf)
844 {
845         spin_lock_irq(&conf->resync_lock);
846         if (conf->barrier) {
847                 conf->nr_waiting++;
848                 /* Wait for the barrier to drop.
849                  * However if there are already pending
850                  * requests (preventing the barrier from
851                  * rising completely), and the
852                  * pre-process bio queue isn't empty,
853                  * then don't wait, as we need to empty
854                  * that queue to get the nr_pending
855                  * count down.
856                  */
857                 wait_event_lock_irq(conf->wait_barrier,
858                                     !conf->barrier ||
859                                     (conf->nr_pending &&
860                                      current->bio_list &&
861                                      !bio_list_empty(current->bio_list)),
862                                     conf->resync_lock,
863                         );
864                 conf->nr_waiting--;
865         }
866         conf->nr_pending++;
867         spin_unlock_irq(&conf->resync_lock);
868 }
869
870 static void allow_barrier(struct r1conf *conf)
871 {
872         unsigned long flags;
873         spin_lock_irqsave(&conf->resync_lock, flags);
874         conf->nr_pending--;
875         spin_unlock_irqrestore(&conf->resync_lock, flags);
876         wake_up(&conf->wait_barrier);
877 }
878
879 static void freeze_array(struct r1conf *conf)
880 {
881         /* stop syncio and normal IO and wait for everything to
882          * go quite.
883          * We increment barrier and nr_waiting, and then
884          * wait until nr_pending match nr_queued+1
885          * This is called in the context of one normal IO request
886          * that has failed. Thus any sync request that might be pending
887          * will be blocked by nr_pending, and we need to wait for
888          * pending IO requests to complete or be queued for re-try.
889          * Thus the number queued (nr_queued) plus this request (1)
890          * must match the number of pending IOs (nr_pending) before
891          * we continue.
892          */
893         spin_lock_irq(&conf->resync_lock);
894         conf->barrier++;
895         conf->nr_waiting++;
896         wait_event_lock_irq(conf->wait_barrier,
897                             conf->nr_pending == conf->nr_queued+1,
898                             conf->resync_lock,
899                             flush_pending_writes(conf));
900         spin_unlock_irq(&conf->resync_lock);
901 }
902 static void unfreeze_array(struct r1conf *conf)
903 {
904         /* reverse the effect of the freeze */
905         spin_lock_irq(&conf->resync_lock);
906         conf->barrier--;
907         conf->nr_waiting--;
908         wake_up(&conf->wait_barrier);
909         spin_unlock_irq(&conf->resync_lock);
910 }
911
912
913 /* duplicate the data pages for behind I/O 
914  */
915 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
916 {
917         int i;
918         struct bio_vec *bvec;
919         struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
920                                         GFP_NOIO);
921         if (unlikely(!bvecs))
922                 return;
923
924         bio_for_each_segment(bvec, bio, i) {
925                 bvecs[i] = *bvec;
926                 bvecs[i].bv_page = alloc_page(GFP_NOIO);
927                 if (unlikely(!bvecs[i].bv_page))
928                         goto do_sync_io;
929                 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
930                        kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
931                 kunmap(bvecs[i].bv_page);
932                 kunmap(bvec->bv_page);
933         }
934         r1_bio->behind_bvecs = bvecs;
935         r1_bio->behind_page_count = bio->bi_vcnt;
936         set_bit(R1BIO_BehindIO, &r1_bio->state);
937         return;
938
939 do_sync_io:
940         for (i = 0; i < bio->bi_vcnt; i++)
941                 if (bvecs[i].bv_page)
942                         put_page(bvecs[i].bv_page);
943         kfree(bvecs);
944         pr_debug("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
945 }
946
947 static void make_request(struct mddev *mddev, struct bio * bio)
948 {
949         struct r1conf *conf = mddev->private;
950         struct raid1_info *mirror;
951         struct r1bio *r1_bio;
952         struct bio *read_bio;
953         int i, disks;
954         struct bitmap *bitmap;
955         unsigned long flags;
956         const int rw = bio_data_dir(bio);
957         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
958         const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
959         struct md_rdev *blocked_rdev;
960         int first_clone;
961         int sectors_handled;
962         int max_sectors;
963
964         /*
965          * Register the new request and wait if the reconstruction
966          * thread has put up a bar for new requests.
967          * Continue immediately if no resync is active currently.
968          */
969
970         md_write_start(mddev, bio); /* wait on superblock update early */
971
972         if (bio_data_dir(bio) == WRITE &&
973             bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
974             bio->bi_sector < mddev->suspend_hi) {
975                 /* As the suspend_* range is controlled by
976                  * userspace, we want an interruptible
977                  * wait.
978                  */
979                 DEFINE_WAIT(w);
980                 for (;;) {
981                         flush_signals(current);
982                         prepare_to_wait(&conf->wait_barrier,
983                                         &w, TASK_INTERRUPTIBLE);
984                         if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
985                             bio->bi_sector >= mddev->suspend_hi)
986                                 break;
987                         schedule();
988                 }
989                 finish_wait(&conf->wait_barrier, &w);
990         }
991
992         wait_barrier(conf);
993
994         bitmap = mddev->bitmap;
995
996         /*
997          * make_request() can abort the operation when READA is being
998          * used and no empty request is available.
999          *
1000          */
1001         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1002
1003         r1_bio->master_bio = bio;
1004         r1_bio->sectors = bio->bi_size >> 9;
1005         r1_bio->state = 0;
1006         r1_bio->mddev = mddev;
1007         r1_bio->sector = bio->bi_sector;
1008
1009         /* We might need to issue multiple reads to different
1010          * devices if there are bad blocks around, so we keep
1011          * track of the number of reads in bio->bi_phys_segments.
1012          * If this is 0, there is only one r1_bio and no locking
1013          * will be needed when requests complete.  If it is
1014          * non-zero, then it is the number of not-completed requests.
1015          */
1016         bio->bi_phys_segments = 0;
1017         clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1018
1019         if (rw == READ) {
1020                 /*
1021                  * read balancing logic:
1022                  */
1023                 int rdisk;
1024
1025 read_again:
1026                 rdisk = read_balance(conf, r1_bio, &max_sectors);
1027
1028                 if (rdisk < 0) {
1029                         /* couldn't find anywhere to read from */
1030                         raid_end_bio_io(r1_bio);
1031                         return;
1032                 }
1033                 mirror = conf->mirrors + rdisk;
1034
1035                 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1036                     bitmap) {
1037                         /* Reading from a write-mostly device must
1038                          * take care not to over-take any writes
1039                          * that are 'behind'
1040                          */
1041                         wait_event(bitmap->behind_wait,
1042                                    atomic_read(&bitmap->behind_writes) == 0);
1043                 }
1044                 r1_bio->read_disk = rdisk;
1045
1046                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1047                 md_trim_bio(read_bio, r1_bio->sector - bio->bi_sector,
1048                             max_sectors);
1049
1050                 r1_bio->bios[rdisk] = read_bio;
1051
1052                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
1053                 read_bio->bi_bdev = mirror->rdev->bdev;
1054                 read_bio->bi_end_io = raid1_end_read_request;
1055                 read_bio->bi_rw = READ | do_sync;
1056                 read_bio->bi_private = r1_bio;
1057
1058                 if (max_sectors < r1_bio->sectors) {
1059                         /* could not read all from this device, so we will
1060                          * need another r1_bio.
1061                          */
1062
1063                         sectors_handled = (r1_bio->sector + max_sectors
1064                                            - bio->bi_sector);
1065                         r1_bio->sectors = max_sectors;
1066                         spin_lock_irq(&conf->device_lock);
1067                         if (bio->bi_phys_segments == 0)
1068                                 bio->bi_phys_segments = 2;
1069                         else
1070                                 bio->bi_phys_segments++;
1071                         spin_unlock_irq(&conf->device_lock);
1072                         /* Cannot call generic_make_request directly
1073                          * as that will be queued in __make_request
1074                          * and subsequent mempool_alloc might block waiting
1075                          * for it.  So hand bio over to raid1d.
1076                          */
1077                         reschedule_retry(r1_bio);
1078
1079                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1080
1081                         r1_bio->master_bio = bio;
1082                         r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1083                         r1_bio->state = 0;
1084                         r1_bio->mddev = mddev;
1085                         r1_bio->sector = bio->bi_sector + sectors_handled;
1086                         goto read_again;
1087                 } else
1088                         generic_make_request(read_bio);
1089                 return;
1090         }
1091
1092         /*
1093          * WRITE:
1094          */
1095         if (conf->pending_count >= max_queued_requests) {
1096                 md_wakeup_thread(mddev->thread);
1097                 wait_event(conf->wait_barrier,
1098                            conf->pending_count < max_queued_requests);
1099         }
1100         /* first select target devices under rcu_lock and
1101          * inc refcount on their rdev.  Record them by setting
1102          * bios[x] to bio
1103          * If there are known/acknowledged bad blocks on any device on
1104          * which we have seen a write error, we want to avoid writing those
1105          * blocks.
1106          * This potentially requires several writes to write around
1107          * the bad blocks.  Each set of writes gets it's own r1bio
1108          * with a set of bios attached.
1109          */
1110
1111         disks = conf->raid_disks * 2;
1112  retry_write:
1113         blocked_rdev = NULL;
1114         rcu_read_lock();
1115         max_sectors = r1_bio->sectors;
1116         for (i = 0;  i < disks; i++) {
1117                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1118                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1119                         atomic_inc(&rdev->nr_pending);
1120                         blocked_rdev = rdev;
1121                         break;
1122                 }
1123                 r1_bio->bios[i] = NULL;
1124                 if (!rdev || test_bit(Faulty, &rdev->flags)
1125                     || test_bit(Unmerged, &rdev->flags)) {
1126                         if (i < conf->raid_disks)
1127                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1128                         continue;
1129                 }
1130
1131                 atomic_inc(&rdev->nr_pending);
1132                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1133                         sector_t first_bad;
1134                         int bad_sectors;
1135                         int is_bad;
1136
1137                         is_bad = is_badblock(rdev, r1_bio->sector,
1138                                              max_sectors,
1139                                              &first_bad, &bad_sectors);
1140                         if (is_bad < 0) {
1141                                 /* mustn't write here until the bad block is
1142                                  * acknowledged*/
1143                                 set_bit(BlockedBadBlocks, &rdev->flags);
1144                                 blocked_rdev = rdev;
1145                                 break;
1146                         }
1147                         if (is_bad && first_bad <= r1_bio->sector) {
1148                                 /* Cannot write here at all */
1149                                 bad_sectors -= (r1_bio->sector - first_bad);
1150                                 if (bad_sectors < max_sectors)
1151                                         /* mustn't write more than bad_sectors
1152                                          * to other devices yet
1153                                          */
1154                                         max_sectors = bad_sectors;
1155                                 rdev_dec_pending(rdev, mddev);
1156                                 /* We don't set R1BIO_Degraded as that
1157                                  * only applies if the disk is
1158                                  * missing, so it might be re-added,
1159                                  * and we want to know to recover this
1160                                  * chunk.
1161                                  * In this case the device is here,
1162                                  * and the fact that this chunk is not
1163                                  * in-sync is recorded in the bad
1164                                  * block log
1165                                  */
1166                                 continue;
1167                         }
1168                         if (is_bad) {
1169                                 int good_sectors = first_bad - r1_bio->sector;
1170                                 if (good_sectors < max_sectors)
1171                                         max_sectors = good_sectors;
1172                         }
1173                 }
1174                 r1_bio->bios[i] = bio;
1175         }
1176         rcu_read_unlock();
1177
1178         if (unlikely(blocked_rdev)) {
1179                 /* Wait for this device to become unblocked */
1180                 int j;
1181
1182                 for (j = 0; j < i; j++)
1183                         if (r1_bio->bios[j])
1184                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1185                 r1_bio->state = 0;
1186                 allow_barrier(conf);
1187                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1188                 wait_barrier(conf);
1189                 goto retry_write;
1190         }
1191
1192         if (max_sectors < r1_bio->sectors) {
1193                 /* We are splitting this write into multiple parts, so
1194                  * we need to prepare for allocating another r1_bio.
1195                  */
1196                 r1_bio->sectors = max_sectors;
1197                 spin_lock_irq(&conf->device_lock);
1198                 if (bio->bi_phys_segments == 0)
1199                         bio->bi_phys_segments = 2;
1200                 else
1201                         bio->bi_phys_segments++;
1202                 spin_unlock_irq(&conf->device_lock);
1203         }
1204         sectors_handled = r1_bio->sector + max_sectors - bio->bi_sector;
1205
1206         atomic_set(&r1_bio->remaining, 1);
1207         atomic_set(&r1_bio->behind_remaining, 0);
1208
1209         first_clone = 1;
1210         for (i = 0; i < disks; i++) {
1211                 struct bio *mbio;
1212                 if (!r1_bio->bios[i])
1213                         continue;
1214
1215                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1216                 md_trim_bio(mbio, r1_bio->sector - bio->bi_sector, max_sectors);
1217
1218                 if (first_clone) {
1219                         /* do behind I/O ?
1220                          * Not if there are too many, or cannot
1221                          * allocate memory, or a reader on WriteMostly
1222                          * is waiting for behind writes to flush */
1223                         if (bitmap &&
1224                             (atomic_read(&bitmap->behind_writes)
1225                              < mddev->bitmap_info.max_write_behind) &&
1226                             !waitqueue_active(&bitmap->behind_wait))
1227                                 alloc_behind_pages(mbio, r1_bio);
1228
1229                         bitmap_startwrite(bitmap, r1_bio->sector,
1230                                           r1_bio->sectors,
1231                                           test_bit(R1BIO_BehindIO,
1232                                                    &r1_bio->state));
1233                         first_clone = 0;
1234                 }
1235                 if (r1_bio->behind_bvecs) {
1236                         struct bio_vec *bvec;
1237                         int j;
1238
1239                         /* Yes, I really want the '__' version so that
1240                          * we clear any unused pointer in the io_vec, rather
1241                          * than leave them unchanged.  This is important
1242                          * because when we come to free the pages, we won't
1243                          * know the original bi_idx, so we just free
1244                          * them all
1245                          */
1246                         __bio_for_each_segment(bvec, mbio, j, 0)
1247                                 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1248                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1249                                 atomic_inc(&r1_bio->behind_remaining);
1250                 }
1251
1252                 r1_bio->bios[i] = mbio;
1253
1254                 mbio->bi_sector = (r1_bio->sector +
1255                                    conf->mirrors[i].rdev->data_offset);
1256                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1257                 mbio->bi_end_io = raid1_end_write_request;
1258                 mbio->bi_rw = WRITE | do_flush_fua | do_sync;
1259                 mbio->bi_private = r1_bio;
1260
1261                 atomic_inc(&r1_bio->remaining);
1262                 spin_lock_irqsave(&conf->device_lock, flags);
1263                 bio_list_add(&conf->pending_bio_list, mbio);
1264                 conf->pending_count++;
1265                 spin_unlock_irqrestore(&conf->device_lock, flags);
1266                 if (!mddev_check_plugged(mddev))
1267                         md_wakeup_thread(mddev->thread);
1268         }
1269         /* Mustn't call r1_bio_write_done before this next test,
1270          * as it could result in the bio being freed.
1271          */
1272         if (sectors_handled < (bio->bi_size >> 9)) {
1273                 r1_bio_write_done(r1_bio);
1274                 /* We need another r1_bio.  It has already been counted
1275                  * in bio->bi_phys_segments
1276                  */
1277                 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1278                 r1_bio->master_bio = bio;
1279                 r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1280                 r1_bio->state = 0;
1281                 r1_bio->mddev = mddev;
1282                 r1_bio->sector = bio->bi_sector + sectors_handled;
1283                 goto retry_write;
1284         }
1285
1286         r1_bio_write_done(r1_bio);
1287
1288         /* In case raid1d snuck in to freeze_array */
1289         wake_up(&conf->wait_barrier);
1290 }
1291
1292 static void status(struct seq_file *seq, struct mddev *mddev)
1293 {
1294         struct r1conf *conf = mddev->private;
1295         int i;
1296
1297         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1298                    conf->raid_disks - mddev->degraded);
1299         rcu_read_lock();
1300         for (i = 0; i < conf->raid_disks; i++) {
1301                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1302                 seq_printf(seq, "%s",
1303                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1304         }
1305         rcu_read_unlock();
1306         seq_printf(seq, "]");
1307 }
1308
1309
1310 static void error(struct mddev *mddev, struct md_rdev *rdev)
1311 {
1312         char b[BDEVNAME_SIZE];
1313         struct r1conf *conf = mddev->private;
1314
1315         /*
1316          * If it is not operational, then we have already marked it as dead
1317          * else if it is the last working disks, ignore the error, let the
1318          * next level up know.
1319          * else mark the drive as failed
1320          */
1321         if (test_bit(In_sync, &rdev->flags)
1322             && (conf->raid_disks - mddev->degraded) == 1) {
1323                 /*
1324                  * Don't fail the drive, act as though we were just a
1325                  * normal single drive.
1326                  * However don't try a recovery from this drive as
1327                  * it is very likely to fail.
1328                  */
1329                 conf->recovery_disabled = mddev->recovery_disabled;
1330                 return;
1331         }
1332         set_bit(Blocked, &rdev->flags);
1333         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1334                 unsigned long flags;
1335                 spin_lock_irqsave(&conf->device_lock, flags);
1336                 mddev->degraded++;
1337                 set_bit(Faulty, &rdev->flags);
1338                 spin_unlock_irqrestore(&conf->device_lock, flags);
1339                 /*
1340                  * if recovery is running, make sure it aborts.
1341                  */
1342                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1343         } else
1344                 set_bit(Faulty, &rdev->flags);
1345         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1346         printk(KERN_ALERT
1347                "md/raid1:%s: Disk failure on %s, disabling device.\n"
1348                "md/raid1:%s: Operation continuing on %d devices.\n",
1349                mdname(mddev), bdevname(rdev->bdev, b),
1350                mdname(mddev), conf->raid_disks - mddev->degraded);
1351 }
1352
1353 static void print_conf(struct r1conf *conf)
1354 {
1355         int i;
1356
1357         printk(KERN_DEBUG "RAID1 conf printout:\n");
1358         if (!conf) {
1359                 printk(KERN_DEBUG "(!conf)\n");
1360                 return;
1361         }
1362         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1363                 conf->raid_disks);
1364
1365         rcu_read_lock();
1366         for (i = 0; i < conf->raid_disks; i++) {
1367                 char b[BDEVNAME_SIZE];
1368                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1369                 if (rdev)
1370                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1371                                i, !test_bit(In_sync, &rdev->flags),
1372                                !test_bit(Faulty, &rdev->flags),
1373                                bdevname(rdev->bdev,b));
1374         }
1375         rcu_read_unlock();
1376 }
1377
1378 static void close_sync(struct r1conf *conf)
1379 {
1380         wait_barrier(conf);
1381         allow_barrier(conf);
1382
1383         mempool_destroy(conf->r1buf_pool);
1384         conf->r1buf_pool = NULL;
1385 }
1386
1387 static int raid1_spare_active(struct mddev *mddev)
1388 {
1389         int i;
1390         struct r1conf *conf = mddev->private;
1391         int count = 0;
1392         unsigned long flags;
1393
1394         /*
1395          * Find all failed disks within the RAID1 configuration 
1396          * and mark them readable.
1397          * Called under mddev lock, so rcu protection not needed.
1398          */
1399         for (i = 0; i < conf->raid_disks; i++) {
1400                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1401                 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1402                 if (repl
1403                     && repl->recovery_offset == MaxSector
1404                     && !test_bit(Faulty, &repl->flags)
1405                     && !test_and_set_bit(In_sync, &repl->flags)) {
1406                         /* replacement has just become active */
1407                         if (!rdev ||
1408                             !test_and_clear_bit(In_sync, &rdev->flags))
1409                                 count++;
1410                         if (rdev) {
1411                                 /* Replaced device not technically
1412                                  * faulty, but we need to be sure
1413                                  * it gets removed and never re-added
1414                                  */
1415                                 set_bit(Faulty, &rdev->flags);
1416                                 sysfs_notify_dirent_safe(
1417                                         rdev->sysfs_state);
1418                         }
1419                 }
1420                 if (rdev
1421                     && !test_bit(Faulty, &rdev->flags)
1422                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1423                         count++;
1424                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1425                 }
1426         }
1427         spin_lock_irqsave(&conf->device_lock, flags);
1428         mddev->degraded -= count;
1429         spin_unlock_irqrestore(&conf->device_lock, flags);
1430
1431         print_conf(conf);
1432         return count;
1433 }
1434
1435
1436 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1437 {
1438         struct r1conf *conf = mddev->private;
1439         int err = -EEXIST;
1440         int mirror = 0;
1441         struct raid1_info *p;
1442         int first = 0;
1443         int last = conf->raid_disks - 1;
1444         struct request_queue *q = bdev_get_queue(rdev->bdev);
1445
1446         if (mddev->recovery_disabled == conf->recovery_disabled)
1447                 return -EBUSY;
1448
1449         if (rdev->raid_disk >= 0)
1450                 first = last = rdev->raid_disk;
1451
1452         if (q->merge_bvec_fn) {
1453                 set_bit(Unmerged, &rdev->flags);
1454                 mddev->merge_check_needed = 1;
1455         }
1456
1457         for (mirror = first; mirror <= last; mirror++) {
1458                 p = conf->mirrors+mirror;
1459                 if (!p->rdev) {
1460
1461                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1462                                           rdev->data_offset << 9);
1463
1464                         p->head_position = 0;
1465                         rdev->raid_disk = mirror;
1466                         err = 0;
1467                         /* As all devices are equivalent, we don't need a full recovery
1468                          * if this was recently any drive of the array
1469                          */
1470                         if (rdev->saved_raid_disk < 0)
1471                                 conf->fullsync = 1;
1472                         rcu_assign_pointer(p->rdev, rdev);
1473                         break;
1474                 }
1475                 if (test_bit(WantReplacement, &p->rdev->flags) &&
1476                     p[conf->raid_disks].rdev == NULL) {
1477                         /* Add this device as a replacement */
1478                         clear_bit(In_sync, &rdev->flags);
1479                         set_bit(Replacement, &rdev->flags);
1480                         rdev->raid_disk = mirror;
1481                         err = 0;
1482                         conf->fullsync = 1;
1483                         rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1484                         break;
1485                 }
1486         }
1487         if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1488                 /* Some requests might not have seen this new
1489                  * merge_bvec_fn.  We must wait for them to complete
1490                  * before merging the device fully.
1491                  * First we make sure any code which has tested
1492                  * our function has submitted the request, then
1493                  * we wait for all outstanding requests to complete.
1494                  */
1495                 synchronize_sched();
1496                 raise_barrier(conf);
1497                 lower_barrier(conf);
1498                 clear_bit(Unmerged, &rdev->flags);
1499         }
1500         md_integrity_add_rdev(rdev, mddev);
1501         print_conf(conf);
1502         return err;
1503 }
1504
1505 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1506 {
1507         struct r1conf *conf = mddev->private;
1508         int err = 0;
1509         int number = rdev->raid_disk;
1510         struct raid1_info *p = conf->mirrors + number;
1511
1512         if (rdev != p->rdev)
1513                 p = conf->mirrors + conf->raid_disks + number;
1514
1515         print_conf(conf);
1516         if (rdev == p->rdev) {
1517                 if (test_bit(In_sync, &rdev->flags) ||
1518                     atomic_read(&rdev->nr_pending)) {
1519                         err = -EBUSY;
1520                         goto abort;
1521                 }
1522                 /* Only remove non-faulty devices if recovery
1523                  * is not possible.
1524                  */
1525                 if (!test_bit(Faulty, &rdev->flags) &&
1526                     mddev->recovery_disabled != conf->recovery_disabled &&
1527                     mddev->degraded < conf->raid_disks) {
1528                         err = -EBUSY;
1529                         goto abort;
1530                 }
1531                 p->rdev = NULL;
1532                 synchronize_rcu();
1533                 if (atomic_read(&rdev->nr_pending)) {
1534                         /* lost the race, try later */
1535                         err = -EBUSY;
1536                         p->rdev = rdev;
1537                         goto abort;
1538                 } else if (conf->mirrors[conf->raid_disks + number].rdev) {
1539                         /* We just removed a device that is being replaced.
1540                          * Move down the replacement.  We drain all IO before
1541                          * doing this to avoid confusion.
1542                          */
1543                         struct md_rdev *repl =
1544                                 conf->mirrors[conf->raid_disks + number].rdev;
1545                         raise_barrier(conf);
1546                         clear_bit(Replacement, &repl->flags);
1547                         p->rdev = repl;
1548                         conf->mirrors[conf->raid_disks + number].rdev = NULL;
1549                         lower_barrier(conf);
1550                         clear_bit(WantReplacement, &rdev->flags);
1551                 } else
1552                         clear_bit(WantReplacement, &rdev->flags);
1553                 err = md_integrity_register(mddev);
1554         }
1555 abort:
1556
1557         print_conf(conf);
1558         return err;
1559 }
1560
1561
1562 static void end_sync_read(struct bio *bio, int error)
1563 {
1564         struct r1bio *r1_bio = bio->bi_private;
1565
1566         update_head_pos(r1_bio->read_disk, r1_bio);
1567
1568         /*
1569          * we have read a block, now it needs to be re-written,
1570          * or re-read if the read failed.
1571          * We don't do much here, just schedule handling by raid1d
1572          */
1573         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1574                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1575
1576         if (atomic_dec_and_test(&r1_bio->remaining))
1577                 reschedule_retry(r1_bio);
1578 }
1579
1580 static void end_sync_write(struct bio *bio, int error)
1581 {
1582         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1583         struct r1bio *r1_bio = bio->bi_private;
1584         struct mddev *mddev = r1_bio->mddev;
1585         struct r1conf *conf = mddev->private;
1586         int mirror=0;
1587         sector_t first_bad;
1588         int bad_sectors;
1589
1590         mirror = find_bio_disk(r1_bio, bio);
1591
1592         if (!uptodate) {
1593                 sector_t sync_blocks = 0;
1594                 sector_t s = r1_bio->sector;
1595                 long sectors_to_go = r1_bio->sectors;
1596                 /* make sure these bits doesn't get cleared. */
1597                 do {
1598                         bitmap_end_sync(mddev->bitmap, s,
1599                                         &sync_blocks, 1);
1600                         s += sync_blocks;
1601                         sectors_to_go -= sync_blocks;
1602                 } while (sectors_to_go > 0);
1603                 set_bit(WriteErrorSeen,
1604                         &conf->mirrors[mirror].rdev->flags);
1605                 if (!test_and_set_bit(WantReplacement,
1606                                       &conf->mirrors[mirror].rdev->flags))
1607                         set_bit(MD_RECOVERY_NEEDED, &
1608                                 mddev->recovery);
1609                 set_bit(R1BIO_WriteError, &r1_bio->state);
1610         } else if (is_badblock(conf->mirrors[mirror].rdev,
1611                                r1_bio->sector,
1612                                r1_bio->sectors,
1613                                &first_bad, &bad_sectors) &&
1614                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1615                                 r1_bio->sector,
1616                                 r1_bio->sectors,
1617                                 &first_bad, &bad_sectors)
1618                 )
1619                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1620
1621         if (atomic_dec_and_test(&r1_bio->remaining)) {
1622                 int s = r1_bio->sectors;
1623                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1624                     test_bit(R1BIO_WriteError, &r1_bio->state))
1625                         reschedule_retry(r1_bio);
1626                 else {
1627                         put_buf(r1_bio);
1628                         md_done_sync(mddev, s, uptodate);
1629                 }
1630         }
1631 }
1632
1633 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1634                             int sectors, struct page *page, int rw)
1635 {
1636         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1637                 /* success */
1638                 return 1;
1639         if (rw == WRITE) {
1640                 set_bit(WriteErrorSeen, &rdev->flags);
1641                 if (!test_and_set_bit(WantReplacement,
1642                                       &rdev->flags))
1643                         set_bit(MD_RECOVERY_NEEDED, &
1644                                 rdev->mddev->recovery);
1645         }
1646         /* need to record an error - either for the block or the device */
1647         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1648                 md_error(rdev->mddev, rdev);
1649         return 0;
1650 }
1651
1652 static int fix_sync_read_error(struct r1bio *r1_bio)
1653 {
1654         /* Try some synchronous reads of other devices to get
1655          * good data, much like with normal read errors.  Only
1656          * read into the pages we already have so we don't
1657          * need to re-issue the read request.
1658          * We don't need to freeze the array, because being in an
1659          * active sync request, there is no normal IO, and
1660          * no overlapping syncs.
1661          * We don't need to check is_badblock() again as we
1662          * made sure that anything with a bad block in range
1663          * will have bi_end_io clear.
1664          */
1665         struct mddev *mddev = r1_bio->mddev;
1666         struct r1conf *conf = mddev->private;
1667         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1668         sector_t sect = r1_bio->sector;
1669         int sectors = r1_bio->sectors;
1670         int idx = 0;
1671
1672         while(sectors) {
1673                 int s = sectors;
1674                 int d = r1_bio->read_disk;
1675                 int success = 0;
1676                 struct md_rdev *rdev;
1677                 int start;
1678
1679                 if (s > (PAGE_SIZE>>9))
1680                         s = PAGE_SIZE >> 9;
1681                 do {
1682                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1683                                 /* No rcu protection needed here devices
1684                                  * can only be removed when no resync is
1685                                  * active, and resync is currently active
1686                                  */
1687                                 rdev = conf->mirrors[d].rdev;
1688                                 if (sync_page_io(rdev, sect, s<<9,
1689                                                  bio->bi_io_vec[idx].bv_page,
1690                                                  READ, false)) {
1691                                         success = 1;
1692                                         break;
1693                                 }
1694                         }
1695                         d++;
1696                         if (d == conf->raid_disks * 2)
1697                                 d = 0;
1698                 } while (!success && d != r1_bio->read_disk);
1699
1700                 if (!success) {
1701                         char b[BDEVNAME_SIZE];
1702                         int abort = 0;
1703                         /* Cannot read from anywhere, this block is lost.
1704                          * Record a bad block on each device.  If that doesn't
1705                          * work just disable and interrupt the recovery.
1706                          * Don't fail devices as that won't really help.
1707                          */
1708                         printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1709                                " for block %llu\n",
1710                                mdname(mddev),
1711                                bdevname(bio->bi_bdev, b),
1712                                (unsigned long long)r1_bio->sector);
1713                         for (d = 0; d < conf->raid_disks * 2; d++) {
1714                                 rdev = conf->mirrors[d].rdev;
1715                                 if (!rdev || test_bit(Faulty, &rdev->flags))
1716                                         continue;
1717                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
1718                                         abort = 1;
1719                         }
1720                         if (abort) {
1721                                 conf->recovery_disabled =
1722                                         mddev->recovery_disabled;
1723                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1724                                 md_done_sync(mddev, r1_bio->sectors, 0);
1725                                 put_buf(r1_bio);
1726                                 return 0;
1727                         }
1728                         /* Try next page */
1729                         sectors -= s;
1730                         sect += s;
1731                         idx++;
1732                         continue;
1733                 }
1734
1735                 start = d;
1736                 /* write it back and re-read */
1737                 while (d != r1_bio->read_disk) {
1738                         if (d == 0)
1739                                 d = conf->raid_disks * 2;
1740                         d--;
1741                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1742                                 continue;
1743                         rdev = conf->mirrors[d].rdev;
1744                         if (r1_sync_page_io(rdev, sect, s,
1745                                             bio->bi_io_vec[idx].bv_page,
1746                                             WRITE) == 0) {
1747                                 r1_bio->bios[d]->bi_end_io = NULL;
1748                                 rdev_dec_pending(rdev, mddev);
1749                         }
1750                 }
1751                 d = start;
1752                 while (d != r1_bio->read_disk) {
1753                         if (d == 0)
1754                                 d = conf->raid_disks * 2;
1755                         d--;
1756                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1757                                 continue;
1758                         rdev = conf->mirrors[d].rdev;
1759                         if (r1_sync_page_io(rdev, sect, s,
1760                                             bio->bi_io_vec[idx].bv_page,
1761                                             READ) != 0)
1762                                 atomic_add(s, &rdev->corrected_errors);
1763                 }
1764                 sectors -= s;
1765                 sect += s;
1766                 idx ++;
1767         }
1768         set_bit(R1BIO_Uptodate, &r1_bio->state);
1769         set_bit(BIO_UPTODATE, &bio->bi_flags);
1770         return 1;
1771 }
1772
1773 static int process_checks(struct r1bio *r1_bio)
1774 {
1775         /* We have read all readable devices.  If we haven't
1776          * got the block, then there is no hope left.
1777          * If we have, then we want to do a comparison
1778          * and skip the write if everything is the same.
1779          * If any blocks failed to read, then we need to
1780          * attempt an over-write
1781          */
1782         struct mddev *mddev = r1_bio->mddev;
1783         struct r1conf *conf = mddev->private;
1784         int primary;
1785         int i;
1786         int vcnt;
1787
1788         for (primary = 0; primary < conf->raid_disks * 2; primary++)
1789                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1790                     test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1791                         r1_bio->bios[primary]->bi_end_io = NULL;
1792                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1793                         break;
1794                 }
1795         r1_bio->read_disk = primary;
1796         vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
1797         for (i = 0; i < conf->raid_disks * 2; i++) {
1798                 int j;
1799                 struct bio *pbio = r1_bio->bios[primary];
1800                 struct bio *sbio = r1_bio->bios[i];
1801                 int size;
1802
1803                 if (r1_bio->bios[i]->bi_end_io != end_sync_read)
1804                         continue;
1805
1806                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1807                         for (j = vcnt; j-- ; ) {
1808                                 struct page *p, *s;
1809                                 p = pbio->bi_io_vec[j].bv_page;
1810                                 s = sbio->bi_io_vec[j].bv_page;
1811                                 if (memcmp(page_address(p),
1812                                            page_address(s),
1813                                            sbio->bi_io_vec[j].bv_len))
1814                                         break;
1815                         }
1816                 } else
1817                         j = 0;
1818                 if (j >= 0)
1819                         mddev->resync_mismatches += r1_bio->sectors;
1820                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1821                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1822                         /* No need to write to this device. */
1823                         sbio->bi_end_io = NULL;
1824                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1825                         continue;
1826                 }
1827                 /* fixup the bio for reuse */
1828                 sbio->bi_vcnt = vcnt;
1829                 sbio->bi_size = r1_bio->sectors << 9;
1830                 sbio->bi_idx = 0;
1831                 sbio->bi_phys_segments = 0;
1832                 sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1833                 sbio->bi_flags |= 1 << BIO_UPTODATE;
1834                 sbio->bi_next = NULL;
1835                 sbio->bi_sector = r1_bio->sector +
1836                         conf->mirrors[i].rdev->data_offset;
1837                 sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1838                 size = sbio->bi_size;
1839                 for (j = 0; j < vcnt ; j++) {
1840                         struct bio_vec *bi;
1841                         bi = &sbio->bi_io_vec[j];
1842                         bi->bv_offset = 0;
1843                         if (size > PAGE_SIZE)
1844                                 bi->bv_len = PAGE_SIZE;
1845                         else
1846                                 bi->bv_len = size;
1847                         size -= PAGE_SIZE;
1848                         memcpy(page_address(bi->bv_page),
1849                                page_address(pbio->bi_io_vec[j].bv_page),
1850                                PAGE_SIZE);
1851                 }
1852         }
1853         return 0;
1854 }
1855
1856 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
1857 {
1858         struct r1conf *conf = mddev->private;
1859         int i;
1860         int disks = conf->raid_disks * 2;
1861         struct bio *bio, *wbio;
1862
1863         bio = r1_bio->bios[r1_bio->read_disk];
1864
1865         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1866                 /* ouch - failed to read all of that. */
1867                 if (!fix_sync_read_error(r1_bio))
1868                         return;
1869
1870         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1871                 if (process_checks(r1_bio) < 0)
1872                         return;
1873         /*
1874          * schedule writes
1875          */
1876         atomic_set(&r1_bio->remaining, 1);
1877         for (i = 0; i < disks ; i++) {
1878                 wbio = r1_bio->bios[i];
1879                 if (wbio->bi_end_io == NULL ||
1880                     (wbio->bi_end_io == end_sync_read &&
1881                      (i == r1_bio->read_disk ||
1882                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1883                         continue;
1884
1885                 wbio->bi_rw = WRITE;
1886                 wbio->bi_end_io = end_sync_write;
1887                 atomic_inc(&r1_bio->remaining);
1888                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1889
1890                 generic_make_request(wbio);
1891         }
1892
1893         if (atomic_dec_and_test(&r1_bio->remaining)) {
1894                 /* if we're here, all write(s) have completed, so clean up */
1895                 int s = r1_bio->sectors;
1896                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1897                     test_bit(R1BIO_WriteError, &r1_bio->state))
1898                         reschedule_retry(r1_bio);
1899                 else {
1900                         put_buf(r1_bio);
1901                         md_done_sync(mddev, s, 1);
1902                 }
1903         }
1904 }
1905
1906 /*
1907  * This is a kernel thread which:
1908  *
1909  *      1.      Retries failed read operations on working mirrors.
1910  *      2.      Updates the raid superblock when problems encounter.
1911  *      3.      Performs writes following reads for array synchronising.
1912  */
1913
1914 static void fix_read_error(struct r1conf *conf, int read_disk,
1915                            sector_t sect, int sectors)
1916 {
1917         struct mddev *mddev = conf->mddev;
1918         while(sectors) {
1919                 int s = sectors;
1920                 int d = read_disk;
1921                 int success = 0;
1922                 int start;
1923                 struct md_rdev *rdev;
1924
1925                 if (s > (PAGE_SIZE>>9))
1926                         s = PAGE_SIZE >> 9;
1927
1928                 do {
1929                         /* Note: no rcu protection needed here
1930                          * as this is synchronous in the raid1d thread
1931                          * which is the thread that might remove
1932                          * a device.  If raid1d ever becomes multi-threaded....
1933                          */
1934                         sector_t first_bad;
1935                         int bad_sectors;
1936
1937                         rdev = conf->mirrors[d].rdev;
1938                         if (rdev &&
1939                             (test_bit(In_sync, &rdev->flags) ||
1940                              (!test_bit(Faulty, &rdev->flags) &&
1941                               rdev->recovery_offset >= sect + s)) &&
1942                             is_badblock(rdev, sect, s,
1943                                         &first_bad, &bad_sectors) == 0 &&
1944                             sync_page_io(rdev, sect, s<<9,
1945                                          conf->tmppage, READ, false))
1946                                 success = 1;
1947                         else {
1948                                 d++;
1949                                 if (d == conf->raid_disks * 2)
1950                                         d = 0;
1951                         }
1952                 } while (!success && d != read_disk);
1953
1954                 if (!success) {
1955                         /* Cannot read from anywhere - mark it bad */
1956                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
1957                         if (!rdev_set_badblocks(rdev, sect, s, 0))
1958                                 md_error(mddev, rdev);
1959                         break;
1960                 }
1961                 /* write it back and re-read */
1962                 start = d;
1963                 while (d != read_disk) {
1964                         if (d==0)
1965                                 d = conf->raid_disks * 2;
1966                         d--;
1967                         rdev = conf->mirrors[d].rdev;
1968                         if (rdev &&
1969                             test_bit(In_sync, &rdev->flags))
1970                                 r1_sync_page_io(rdev, sect, s,
1971                                                 conf->tmppage, WRITE);
1972                 }
1973                 d = start;
1974                 while (d != read_disk) {
1975                         char b[BDEVNAME_SIZE];
1976                         if (d==0)
1977                                 d = conf->raid_disks * 2;
1978                         d--;
1979                         rdev = conf->mirrors[d].rdev;
1980                         if (rdev &&
1981                             test_bit(In_sync, &rdev->flags)) {
1982                                 if (r1_sync_page_io(rdev, sect, s,
1983                                                     conf->tmppage, READ)) {
1984                                         atomic_add(s, &rdev->corrected_errors);
1985                                         printk(KERN_INFO
1986                                                "md/raid1:%s: read error corrected "
1987                                                "(%d sectors at %llu on %s)\n",
1988                                                mdname(mddev), s,
1989                                                (unsigned long long)(sect +
1990                                                    rdev->data_offset),
1991                                                bdevname(rdev->bdev, b));
1992                                 }
1993                         }
1994                 }
1995                 sectors -= s;
1996                 sect += s;
1997         }
1998 }
1999
2000 static void bi_complete(struct bio *bio, int error)
2001 {
2002         complete((struct completion *)bio->bi_private);
2003 }
2004
2005 static int submit_bio_wait(int rw, struct bio *bio)
2006 {
2007         struct completion event;
2008         rw |= REQ_SYNC;
2009
2010         init_completion(&event);
2011         bio->bi_private = &event;
2012         bio->bi_end_io = bi_complete;
2013         submit_bio(rw, bio);
2014         wait_for_completion(&event);
2015
2016         return test_bit(BIO_UPTODATE, &bio->bi_flags);
2017 }
2018
2019 static int narrow_write_error(struct r1bio *r1_bio, int i)
2020 {
2021         struct mddev *mddev = r1_bio->mddev;
2022         struct r1conf *conf = mddev->private;
2023         struct md_rdev *rdev = conf->mirrors[i].rdev;
2024         int vcnt, idx;
2025         struct bio_vec *vec;
2026
2027         /* bio has the data to be written to device 'i' where
2028          * we just recently had a write error.
2029          * We repeatedly clone the bio and trim down to one block,
2030          * then try the write.  Where the write fails we record
2031          * a bad block.
2032          * It is conceivable that the bio doesn't exactly align with
2033          * blocks.  We must handle this somehow.
2034          *
2035          * We currently own a reference on the rdev.
2036          */
2037
2038         int block_sectors;
2039         sector_t sector;
2040         int sectors;
2041         int sect_to_write = r1_bio->sectors;
2042         int ok = 1;
2043
2044         if (rdev->badblocks.shift < 0)
2045                 return 0;
2046
2047         block_sectors = 1 << rdev->badblocks.shift;
2048         sector = r1_bio->sector;
2049         sectors = ((sector + block_sectors)
2050                    & ~(sector_t)(block_sectors - 1))
2051                 - sector;
2052
2053         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2054                 vcnt = r1_bio->behind_page_count;
2055                 vec = r1_bio->behind_bvecs;
2056                 idx = 0;
2057                 while (vec[idx].bv_page == NULL)
2058                         idx++;
2059         } else {
2060                 vcnt = r1_bio->master_bio->bi_vcnt;
2061                 vec = r1_bio->master_bio->bi_io_vec;
2062                 idx = r1_bio->master_bio->bi_idx;
2063         }
2064         while (sect_to_write) {
2065                 struct bio *wbio;
2066                 if (sectors > sect_to_write)
2067                         sectors = sect_to_write;
2068                 /* Write at 'sector' for 'sectors'*/
2069
2070                 wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
2071                 memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
2072                 wbio->bi_sector = r1_bio->sector;
2073                 wbio->bi_rw = WRITE;
2074                 wbio->bi_vcnt = vcnt;
2075                 wbio->bi_size = r1_bio->sectors << 9;
2076                 wbio->bi_idx = idx;
2077
2078                 md_trim_bio(wbio, sector - r1_bio->sector, sectors);
2079                 wbio->bi_sector += rdev->data_offset;
2080                 wbio->bi_bdev = rdev->bdev;
2081                 if (submit_bio_wait(WRITE, wbio) == 0)
2082                         /* failure! */
2083                         ok = rdev_set_badblocks(rdev, sector,
2084                                                 sectors, 0)
2085                                 && ok;
2086
2087                 bio_put(wbio);
2088                 sect_to_write -= sectors;
2089                 sector += sectors;
2090                 sectors = block_sectors;
2091         }
2092         return ok;
2093 }
2094
2095 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2096 {
2097         int m;
2098         int s = r1_bio->sectors;
2099         for (m = 0; m < conf->raid_disks * 2 ; m++) {
2100                 struct md_rdev *rdev = conf->mirrors[m].rdev;
2101                 struct bio *bio = r1_bio->bios[m];
2102                 if (bio->bi_end_io == NULL)
2103                         continue;
2104                 if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2105                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2106                         rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2107                 }
2108                 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2109                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
2110                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2111                                 md_error(conf->mddev, rdev);
2112                 }
2113         }
2114         put_buf(r1_bio);
2115         md_done_sync(conf->mddev, s, 1);
2116 }
2117
2118 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2119 {
2120         int m;
2121         for (m = 0; m < conf->raid_disks * 2 ; m++)
2122                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2123                         struct md_rdev *rdev = conf->mirrors[m].rdev;
2124                         rdev_clear_badblocks(rdev,
2125                                              r1_bio->sector,
2126                                              r1_bio->sectors, 0);
2127                         rdev_dec_pending(rdev, conf->mddev);
2128                 } else if (r1_bio->bios[m] != NULL) {
2129                         /* This drive got a write error.  We need to
2130                          * narrow down and record precise write
2131                          * errors.
2132                          */
2133                         if (!narrow_write_error(r1_bio, m)) {
2134                                 md_error(conf->mddev,
2135                                          conf->mirrors[m].rdev);
2136                                 /* an I/O failed, we can't clear the bitmap */
2137                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2138                         }
2139                         rdev_dec_pending(conf->mirrors[m].rdev,
2140                                          conf->mddev);
2141                 }
2142         if (test_bit(R1BIO_WriteError, &r1_bio->state))
2143                 close_write(r1_bio);
2144         raid_end_bio_io(r1_bio);
2145 }
2146
2147 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2148 {
2149         int disk;
2150         int max_sectors;
2151         struct mddev *mddev = conf->mddev;
2152         struct bio *bio;
2153         char b[BDEVNAME_SIZE];
2154         struct md_rdev *rdev;
2155
2156         clear_bit(R1BIO_ReadError, &r1_bio->state);
2157         /* we got a read error. Maybe the drive is bad.  Maybe just
2158          * the block and we can fix it.
2159          * We freeze all other IO, and try reading the block from
2160          * other devices.  When we find one, we re-write
2161          * and check it that fixes the read error.
2162          * This is all done synchronously while the array is
2163          * frozen
2164          */
2165         if (mddev->ro == 0) {
2166                 freeze_array(conf);
2167                 fix_read_error(conf, r1_bio->read_disk,
2168                                r1_bio->sector, r1_bio->sectors);
2169                 unfreeze_array(conf);
2170         } else
2171                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2172
2173         bio = r1_bio->bios[r1_bio->read_disk];
2174         bdevname(bio->bi_bdev, b);
2175 read_more:
2176         disk = read_balance(conf, r1_bio, &max_sectors);
2177         if (disk == -1) {
2178                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2179                        " read error for block %llu\n",
2180                        mdname(mddev), b, (unsigned long long)r1_bio->sector);
2181                 raid_end_bio_io(r1_bio);
2182         } else {
2183                 const unsigned long do_sync
2184                         = r1_bio->master_bio->bi_rw & REQ_SYNC;
2185                 if (bio) {
2186                         r1_bio->bios[r1_bio->read_disk] =
2187                                 mddev->ro ? IO_BLOCKED : NULL;
2188                         bio_put(bio);
2189                 }
2190                 r1_bio->read_disk = disk;
2191                 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2192                 md_trim_bio(bio, r1_bio->sector - bio->bi_sector, max_sectors);
2193                 r1_bio->bios[r1_bio->read_disk] = bio;
2194                 rdev = conf->mirrors[disk].rdev;
2195                 printk_ratelimited(KERN_ERR
2196                                    "md/raid1:%s: redirecting sector %llu"
2197                                    " to other mirror: %s\n",
2198                                    mdname(mddev),
2199                                    (unsigned long long)r1_bio->sector,
2200                                    bdevname(rdev->bdev, b));
2201                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
2202                 bio->bi_bdev = rdev->bdev;
2203                 bio->bi_end_io = raid1_end_read_request;
2204                 bio->bi_rw = READ | do_sync;
2205                 bio->bi_private = r1_bio;
2206                 if (max_sectors < r1_bio->sectors) {
2207                         /* Drat - have to split this up more */
2208                         struct bio *mbio = r1_bio->master_bio;
2209                         int sectors_handled = (r1_bio->sector + max_sectors
2210                                                - mbio->bi_sector);
2211                         r1_bio->sectors = max_sectors;
2212                         spin_lock_irq(&conf->device_lock);
2213                         if (mbio->bi_phys_segments == 0)
2214                                 mbio->bi_phys_segments = 2;
2215                         else
2216                                 mbio->bi_phys_segments++;
2217                         spin_unlock_irq(&conf->device_lock);
2218                         generic_make_request(bio);
2219                         bio = NULL;
2220
2221                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2222
2223                         r1_bio->master_bio = mbio;
2224                         r1_bio->sectors = (mbio->bi_size >> 9)
2225                                           - sectors_handled;
2226                         r1_bio->state = 0;
2227                         set_bit(R1BIO_ReadError, &r1_bio->state);
2228                         r1_bio->mddev = mddev;
2229                         r1_bio->sector = mbio->bi_sector + sectors_handled;
2230
2231                         goto read_more;
2232                 } else
2233                         generic_make_request(bio);
2234         }
2235 }
2236
2237 static void raid1d(struct mddev *mddev)
2238 {
2239         struct r1bio *r1_bio;
2240         unsigned long flags;
2241         struct r1conf *conf = mddev->private;
2242         struct list_head *head = &conf->retry_list;
2243         struct blk_plug plug;
2244
2245         md_check_recovery(mddev);
2246
2247         blk_start_plug(&plug);
2248         for (;;) {
2249
2250                 flush_pending_writes(conf);
2251
2252                 spin_lock_irqsave(&conf->device_lock, flags);
2253                 if (list_empty(head)) {
2254                         spin_unlock_irqrestore(&conf->device_lock, flags);
2255                         break;
2256                 }
2257                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2258                 list_del(head->prev);
2259                 conf->nr_queued--;
2260                 spin_unlock_irqrestore(&conf->device_lock, flags);
2261
2262                 mddev = r1_bio->mddev;
2263                 conf = mddev->private;
2264                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2265                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2266                             test_bit(R1BIO_WriteError, &r1_bio->state))
2267                                 handle_sync_write_finished(conf, r1_bio);
2268                         else
2269                                 sync_request_write(mddev, r1_bio);
2270                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2271                            test_bit(R1BIO_WriteError, &r1_bio->state))
2272                         handle_write_finished(conf, r1_bio);
2273                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2274                         handle_read_error(conf, r1_bio);
2275                 else
2276                         /* just a partial read to be scheduled from separate
2277                          * context
2278                          */
2279                         generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2280
2281                 cond_resched();
2282                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2283                         md_check_recovery(mddev);
2284         }
2285         blk_finish_plug(&plug);
2286 }
2287
2288
2289 static int init_resync(struct r1conf *conf)
2290 {
2291         int buffs;
2292
2293         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2294         BUG_ON(conf->r1buf_pool);
2295         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2296                                           conf->poolinfo);
2297         if (!conf->r1buf_pool)
2298                 return -ENOMEM;
2299         conf->next_resync = 0;
2300         return 0;
2301 }
2302
2303 /*
2304  * perform a "sync" on one "block"
2305  *
2306  * We need to make sure that no normal I/O request - particularly write
2307  * requests - conflict with active sync requests.
2308  *
2309  * This is achieved by tracking pending requests and a 'barrier' concept
2310  * that can be installed to exclude normal IO requests.
2311  */
2312
2313 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2314 {
2315         struct r1conf *conf = mddev->private;
2316         struct r1bio *r1_bio;
2317         struct bio *bio;
2318         sector_t max_sector, nr_sectors;
2319         int disk = -1;
2320         int i;
2321         int wonly = -1;
2322         int write_targets = 0, read_targets = 0;
2323         sector_t sync_blocks;
2324         int still_degraded = 0;
2325         int good_sectors = RESYNC_SECTORS;
2326         int min_bad = 0; /* number of sectors that are bad in all devices */
2327
2328         if (!conf->r1buf_pool)
2329                 if (init_resync(conf))
2330                         return 0;
2331
2332         max_sector = mddev->dev_sectors;
2333         if (sector_nr >= max_sector) {
2334                 /* If we aborted, we need to abort the
2335                  * sync on the 'current' bitmap chunk (there will
2336                  * only be one in raid1 resync.
2337                  * We can find the current addess in mddev->curr_resync
2338                  */
2339                 if (mddev->curr_resync < max_sector) /* aborted */
2340                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2341                                                 &sync_blocks, 1);
2342                 else /* completed sync */
2343                         conf->fullsync = 0;
2344
2345                 bitmap_close_sync(mddev->bitmap);
2346                 close_sync(conf);
2347                 return 0;
2348         }
2349
2350         if (mddev->bitmap == NULL &&
2351             mddev->recovery_cp == MaxSector &&
2352             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2353             conf->fullsync == 0) {
2354                 *skipped = 1;
2355                 return max_sector - sector_nr;
2356         }
2357         /* before building a request, check if we can skip these blocks..
2358          * This call the bitmap_start_sync doesn't actually record anything
2359          */
2360         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2361             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2362                 /* We can skip this block, and probably several more */
2363                 *skipped = 1;
2364                 return sync_blocks;
2365         }
2366         /*
2367          * If there is non-resync activity waiting for a turn,
2368          * and resync is going fast enough,
2369          * then let it though before starting on this new sync request.
2370          */
2371         if (!go_faster && conf->nr_waiting)
2372                 msleep_interruptible(1000);
2373
2374         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2375         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2376         raise_barrier(conf);
2377
2378         conf->next_resync = sector_nr;
2379
2380         rcu_read_lock();
2381         /*
2382          * If we get a correctably read error during resync or recovery,
2383          * we might want to read from a different device.  So we
2384          * flag all drives that could conceivably be read from for READ,
2385          * and any others (which will be non-In_sync devices) for WRITE.
2386          * If a read fails, we try reading from something else for which READ
2387          * is OK.
2388          */
2389
2390         r1_bio->mddev = mddev;
2391         r1_bio->sector = sector_nr;
2392         r1_bio->state = 0;
2393         set_bit(R1BIO_IsSync, &r1_bio->state);
2394
2395         for (i = 0; i < conf->raid_disks * 2; i++) {
2396                 struct md_rdev *rdev;
2397                 bio = r1_bio->bios[i];
2398
2399                 /* take from bio_init */
2400                 bio->bi_next = NULL;
2401                 bio->bi_flags &= ~(BIO_POOL_MASK-1);
2402                 bio->bi_flags |= 1 << BIO_UPTODATE;
2403                 bio->bi_rw = READ;
2404                 bio->bi_vcnt = 0;
2405                 bio->bi_idx = 0;
2406                 bio->bi_phys_segments = 0;
2407                 bio->bi_size = 0;
2408                 bio->bi_end_io = NULL;
2409                 bio->bi_private = NULL;
2410
2411                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2412                 if (rdev == NULL ||
2413                     test_bit(Faulty, &rdev->flags)) {
2414                         if (i < conf->raid_disks)
2415                                 still_degraded = 1;
2416                 } else if (!test_bit(In_sync, &rdev->flags)) {
2417                         bio->bi_rw = WRITE;
2418                         bio->bi_end_io = end_sync_write;
2419                         write_targets ++;
2420                 } else {
2421                         /* may need to read from here */
2422                         sector_t first_bad = MaxSector;
2423                         int bad_sectors;
2424
2425                         if (is_badblock(rdev, sector_nr, good_sectors,
2426                                         &first_bad, &bad_sectors)) {
2427                                 if (first_bad > sector_nr)
2428                                         good_sectors = first_bad - sector_nr;
2429                                 else {
2430                                         bad_sectors -= (sector_nr - first_bad);
2431                                         if (min_bad == 0 ||
2432                                             min_bad > bad_sectors)
2433                                                 min_bad = bad_sectors;
2434                                 }
2435                         }
2436                         if (sector_nr < first_bad) {
2437                                 if (test_bit(WriteMostly, &rdev->flags)) {
2438                                         if (wonly < 0)
2439                                                 wonly = i;
2440                                 } else {
2441                                         if (disk < 0)
2442                                                 disk = i;
2443                                 }
2444                                 bio->bi_rw = READ;
2445                                 bio->bi_end_io = end_sync_read;
2446                                 read_targets++;
2447                         } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2448                                 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2449                                 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2450                                 /*
2451                                  * The device is suitable for reading (InSync),
2452                                  * but has bad block(s) here. Let's try to correct them,
2453                                  * if we are doing resync or repair. Otherwise, leave
2454                                  * this device alone for this sync request.
2455                                  */
2456                                 bio->bi_rw = WRITE;
2457                                 bio->bi_end_io = end_sync_write;
2458                                 write_targets++;
2459                         }
2460                 }
2461                 if (bio->bi_end_io) {
2462                         atomic_inc(&rdev->nr_pending);
2463                         bio->bi_sector = sector_nr + rdev->data_offset;
2464                         bio->bi_bdev = rdev->bdev;
2465                         bio->bi_private = r1_bio;
2466                 }
2467         }
2468         rcu_read_unlock();
2469         if (disk < 0)
2470                 disk = wonly;
2471         r1_bio->read_disk = disk;
2472
2473         if (read_targets == 0 && min_bad > 0) {
2474                 /* These sectors are bad on all InSync devices, so we
2475                  * need to mark them bad on all write targets
2476                  */
2477                 int ok = 1;
2478                 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2479                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2480                                 struct md_rdev *rdev = conf->mirrors[i].rdev;
2481                                 ok = rdev_set_badblocks(rdev, sector_nr,
2482                                                         min_bad, 0
2483                                         ) && ok;
2484                         }
2485                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2486                 *skipped = 1;
2487                 put_buf(r1_bio);
2488
2489                 if (!ok) {
2490                         /* Cannot record the badblocks, so need to
2491                          * abort the resync.
2492                          * If there are multiple read targets, could just
2493                          * fail the really bad ones ???
2494                          */
2495                         conf->recovery_disabled = mddev->recovery_disabled;
2496                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2497                         return 0;
2498                 } else
2499                         return min_bad;
2500
2501         }
2502         if (min_bad > 0 && min_bad < good_sectors) {
2503                 /* only resync enough to reach the next bad->good
2504                  * transition */
2505                 good_sectors = min_bad;
2506         }
2507
2508         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2509                 /* extra read targets are also write targets */
2510                 write_targets += read_targets-1;
2511
2512         if (write_targets == 0 || read_targets == 0) {
2513                 /* There is nowhere to write, so all non-sync
2514                  * drives must be failed - so we are finished
2515                  */
2516                 sector_t rv;
2517                 if (min_bad > 0)
2518                         max_sector = sector_nr + min_bad;
2519                 rv = max_sector - sector_nr;
2520                 *skipped = 1;
2521                 put_buf(r1_bio);
2522                 return rv;
2523         }
2524
2525         if (max_sector > mddev->resync_max)
2526                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2527         if (max_sector > sector_nr + good_sectors)
2528                 max_sector = sector_nr + good_sectors;
2529         nr_sectors = 0;
2530         sync_blocks = 0;
2531         do {
2532                 struct page *page;
2533                 int len = PAGE_SIZE;
2534                 if (sector_nr + (len>>9) > max_sector)
2535                         len = (max_sector - sector_nr) << 9;
2536                 if (len == 0)
2537                         break;
2538                 if (sync_blocks == 0) {
2539                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2540                                                &sync_blocks, still_degraded) &&
2541                             !conf->fullsync &&
2542                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2543                                 break;
2544                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2545                         if ((len >> 9) > sync_blocks)
2546                                 len = sync_blocks<<9;
2547                 }
2548
2549                 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2550                         bio = r1_bio->bios[i];
2551                         if (bio->bi_end_io) {
2552                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2553                                 if (bio_add_page(bio, page, len, 0) == 0) {
2554                                         /* stop here */
2555                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2556                                         while (i > 0) {
2557                                                 i--;
2558                                                 bio = r1_bio->bios[i];
2559                                                 if (bio->bi_end_io==NULL)
2560                                                         continue;
2561                                                 /* remove last page from this bio */
2562                                                 bio->bi_vcnt--;
2563                                                 bio->bi_size -= len;
2564                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2565                                         }
2566                                         goto bio_full;
2567                                 }
2568                         }
2569                 }
2570                 nr_sectors += len>>9;
2571                 sector_nr += len>>9;
2572                 sync_blocks -= (len>>9);
2573         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2574  bio_full:
2575         r1_bio->sectors = nr_sectors;
2576
2577         /* For a user-requested sync, we read all readable devices and do a
2578          * compare
2579          */
2580         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2581                 atomic_set(&r1_bio->remaining, read_targets);
2582                 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2583                         bio = r1_bio->bios[i];
2584                         if (bio->bi_end_io == end_sync_read) {
2585                                 read_targets--;
2586                                 md_sync_acct(bio->bi_bdev, nr_sectors);
2587                                 generic_make_request(bio);
2588                         }
2589                 }
2590         } else {
2591                 atomic_set(&r1_bio->remaining, 1);
2592                 bio = r1_bio->bios[r1_bio->read_disk];
2593                 md_sync_acct(bio->bi_bdev, nr_sectors);
2594                 generic_make_request(bio);
2595
2596         }
2597         return nr_sectors;
2598 }
2599
2600 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2601 {
2602         if (sectors)
2603                 return sectors;
2604
2605         return mddev->dev_sectors;
2606 }
2607
2608 static struct r1conf *setup_conf(struct mddev *mddev)
2609 {
2610         struct r1conf *conf;
2611         int i;
2612         struct raid1_info *disk;
2613         struct md_rdev *rdev;
2614         int err = -ENOMEM;
2615
2616         conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2617         if (!conf)
2618                 goto abort;
2619
2620         conf->mirrors = kzalloc(sizeof(struct raid1_info)
2621                                 * mddev->raid_disks * 2,
2622                                  GFP_KERNEL);
2623         if (!conf->mirrors)
2624                 goto abort;
2625
2626         conf->tmppage = alloc_page(GFP_KERNEL);
2627         if (!conf->tmppage)
2628                 goto abort;
2629
2630         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2631         if (!conf->poolinfo)
2632                 goto abort;
2633         conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2634         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2635                                           r1bio_pool_free,
2636                                           conf->poolinfo);
2637         if (!conf->r1bio_pool)
2638                 goto abort;
2639
2640         conf->poolinfo->mddev = mddev;
2641
2642         err = -EINVAL;
2643         spin_lock_init(&conf->device_lock);
2644         rdev_for_each(rdev, mddev) {
2645                 struct request_queue *q;
2646                 int disk_idx = rdev->raid_disk;
2647                 if (disk_idx >= mddev->raid_disks
2648                     || disk_idx < 0)
2649                         continue;
2650                 if (test_bit(Replacement, &rdev->flags))
2651                         disk = conf->mirrors + conf->raid_disks + disk_idx;
2652                 else
2653                         disk = conf->mirrors + disk_idx;
2654
2655                 if (disk->rdev)
2656                         goto abort;
2657                 disk->rdev = rdev;
2658                 q = bdev_get_queue(rdev->bdev);
2659                 if (q->merge_bvec_fn)
2660                         mddev->merge_check_needed = 1;
2661
2662                 disk->head_position = 0;
2663                 disk->seq_start = MaxSector;
2664         }
2665         conf->raid_disks = mddev->raid_disks;
2666         conf->mddev = mddev;
2667         INIT_LIST_HEAD(&conf->retry_list);
2668
2669         spin_lock_init(&conf->resync_lock);
2670         init_waitqueue_head(&conf->wait_barrier);
2671
2672         bio_list_init(&conf->pending_bio_list);
2673         conf->pending_count = 0;
2674         conf->recovery_disabled = mddev->recovery_disabled - 1;
2675
2676         err = -EIO;
2677         for (i = 0; i < conf->raid_disks * 2; i++) {
2678
2679                 disk = conf->mirrors + i;
2680
2681                 if (i < conf->raid_disks &&
2682                     disk[conf->raid_disks].rdev) {
2683                         /* This slot has a replacement. */
2684                         if (!disk->rdev) {
2685                                 /* No original, just make the replacement
2686                                  * a recovering spare
2687                                  */
2688                                 disk->rdev =
2689                                         disk[conf->raid_disks].rdev;
2690                                 disk[conf->raid_disks].rdev = NULL;
2691                         } else if (!test_bit(In_sync, &disk->rdev->flags))
2692                                 /* Original is not in_sync - bad */
2693                                 goto abort;
2694                 }
2695
2696                 if (!disk->rdev ||
2697                     !test_bit(In_sync, &disk->rdev->flags)) {
2698                         disk->head_position = 0;
2699                         if (disk->rdev &&
2700                             (disk->rdev->saved_raid_disk < 0))
2701                                 conf->fullsync = 1;
2702                 }
2703         }
2704
2705         err = -ENOMEM;
2706         conf->thread = md_register_thread(raid1d, mddev, "raid1");
2707         if (!conf->thread) {
2708                 printk(KERN_ERR
2709                        "md/raid1:%s: couldn't allocate thread\n",
2710                        mdname(mddev));
2711                 goto abort;
2712         }
2713
2714         return conf;
2715
2716  abort:
2717         if (conf) {
2718                 if (conf->r1bio_pool)
2719                         mempool_destroy(conf->r1bio_pool);
2720                 kfree(conf->mirrors);
2721                 safe_put_page(conf->tmppage);
2722                 kfree(conf->poolinfo);
2723                 kfree(conf);
2724         }
2725         return ERR_PTR(err);
2726 }
2727
2728 static int stop(struct mddev *mddev);
2729 static int run(struct mddev *mddev)
2730 {
2731         struct r1conf *conf;
2732         int i;
2733         struct md_rdev *rdev;
2734         int ret;
2735
2736         if (mddev->level != 1) {
2737                 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2738                        mdname(mddev), mddev->level);
2739                 return -EIO;
2740         }
2741         if (mddev->reshape_position != MaxSector) {
2742                 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2743                        mdname(mddev));
2744                 return -EIO;
2745         }
2746         /*
2747          * copy the already verified devices into our private RAID1
2748          * bookkeeping area. [whatever we allocate in run(),
2749          * should be freed in stop()]
2750          */
2751         if (mddev->private == NULL)
2752                 conf = setup_conf(mddev);
2753         else
2754                 conf = mddev->private;
2755
2756         if (IS_ERR(conf))
2757                 return PTR_ERR(conf);
2758
2759         rdev_for_each(rdev, mddev) {
2760                 if (!mddev->gendisk)
2761                         continue;
2762                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2763                                   rdev->data_offset << 9);
2764         }
2765
2766         mddev->degraded = 0;
2767         for (i=0; i < conf->raid_disks; i++)
2768                 if (conf->mirrors[i].rdev == NULL ||
2769                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2770                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2771                         mddev->degraded++;
2772
2773         if (conf->raid_disks - mddev->degraded == 1)
2774                 mddev->recovery_cp = MaxSector;
2775
2776         if (mddev->recovery_cp != MaxSector)
2777                 printk(KERN_NOTICE "md/raid1:%s: not clean"
2778                        " -- starting background reconstruction\n",
2779                        mdname(mddev));
2780         printk(KERN_INFO 
2781                 "md/raid1:%s: active with %d out of %d mirrors\n",
2782                 mdname(mddev), mddev->raid_disks - mddev->degraded, 
2783                 mddev->raid_disks);
2784
2785         /*
2786          * Ok, everything is just fine now
2787          */
2788         mddev->thread = conf->thread;
2789         conf->thread = NULL;
2790         mddev->private = conf;
2791
2792         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2793
2794         if (mddev->queue) {
2795                 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2796                 mddev->queue->backing_dev_info.congested_data = mddev;
2797                 blk_queue_merge_bvec(mddev->queue, raid1_mergeable_bvec);
2798         }
2799
2800         ret =  md_integrity_register(mddev);
2801         if (ret)
2802                 stop(mddev);
2803         return ret;
2804 }
2805
2806 static int stop(struct mddev *mddev)
2807 {
2808         struct r1conf *conf = mddev->private;
2809         struct bitmap *bitmap = mddev->bitmap;
2810
2811         /* wait for behind writes to complete */
2812         if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2813                 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2814                        mdname(mddev));
2815                 /* need to kick something here to make sure I/O goes? */
2816                 wait_event(bitmap->behind_wait,
2817                            atomic_read(&bitmap->behind_writes) == 0);
2818         }
2819
2820         raise_barrier(conf);
2821         lower_barrier(conf);
2822
2823         md_unregister_thread(&mddev->thread);
2824         if (conf->r1bio_pool)
2825                 mempool_destroy(conf->r1bio_pool);
2826         kfree(conf->mirrors);
2827         kfree(conf->poolinfo);
2828         kfree(conf);
2829         mddev->private = NULL;
2830         return 0;
2831 }
2832
2833 static int raid1_resize(struct mddev *mddev, sector_t sectors)
2834 {
2835         /* no resync is happening, and there is enough space
2836          * on all devices, so we can resize.
2837          * We need to make sure resync covers any new space.
2838          * If the array is shrinking we should possibly wait until
2839          * any io in the removed space completes, but it hardly seems
2840          * worth it.
2841          */
2842         sector_t newsize = raid1_size(mddev, sectors, 0);
2843         if (mddev->external_size &&
2844             mddev->array_sectors > newsize)
2845                 return -EINVAL;
2846         if (mddev->bitmap) {
2847                 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
2848                 if (ret)
2849                         return ret;
2850         }
2851         md_set_array_sectors(mddev, newsize);
2852         set_capacity(mddev->gendisk, mddev->array_sectors);
2853         revalidate_disk(mddev->gendisk);
2854         if (sectors > mddev->dev_sectors &&
2855             mddev->recovery_cp > mddev->dev_sectors) {
2856                 mddev->recovery_cp = mddev->dev_sectors;
2857                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2858         }
2859         mddev->dev_sectors = sectors;
2860         mddev->resync_max_sectors = sectors;
2861         return 0;
2862 }
2863
2864 static int raid1_reshape(struct mddev *mddev)
2865 {
2866         /* We need to:
2867          * 1/ resize the r1bio_pool
2868          * 2/ resize conf->mirrors
2869          *
2870          * We allocate a new r1bio_pool if we can.
2871          * Then raise a device barrier and wait until all IO stops.
2872          * Then resize conf->mirrors and swap in the new r1bio pool.
2873          *
2874          * At the same time, we "pack" the devices so that all the missing
2875          * devices have the higher raid_disk numbers.
2876          */
2877         mempool_t *newpool, *oldpool;
2878         struct pool_info *newpoolinfo;
2879         struct raid1_info *newmirrors;
2880         struct r1conf *conf = mddev->private;
2881         int cnt, raid_disks;
2882         unsigned long flags;
2883         int d, d2, err;
2884
2885         /* Cannot change chunk_size, layout, or level */
2886         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2887             mddev->layout != mddev->new_layout ||
2888             mddev->level != mddev->new_level) {
2889                 mddev->new_chunk_sectors = mddev->chunk_sectors;
2890                 mddev->new_layout = mddev->layout;
2891                 mddev->new_level = mddev->level;
2892                 return -EINVAL;
2893         }
2894
2895         err = md_allow_write(mddev);
2896         if (err)
2897                 return err;
2898
2899         raid_disks = mddev->raid_disks + mddev->delta_disks;
2900
2901         if (raid_disks < conf->raid_disks) {
2902                 cnt=0;
2903                 for (d= 0; d < conf->raid_disks; d++)
2904                         if (conf->mirrors[d].rdev)
2905                                 cnt++;
2906                 if (cnt > raid_disks)
2907                         return -EBUSY;
2908         }
2909
2910         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2911         if (!newpoolinfo)
2912                 return -ENOMEM;
2913         newpoolinfo->mddev = mddev;
2914         newpoolinfo->raid_disks = raid_disks * 2;
2915
2916         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2917                                  r1bio_pool_free, newpoolinfo);
2918         if (!newpool) {
2919                 kfree(newpoolinfo);
2920                 return -ENOMEM;
2921         }
2922         newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
2923                              GFP_KERNEL);
2924         if (!newmirrors) {
2925                 kfree(newpoolinfo);
2926                 mempool_destroy(newpool);
2927                 return -ENOMEM;
2928         }
2929
2930         raise_barrier(conf);
2931
2932         /* ok, everything is stopped */
2933         oldpool = conf->r1bio_pool;
2934         conf->r1bio_pool = newpool;
2935
2936         for (d = d2 = 0; d < conf->raid_disks; d++) {
2937                 struct md_rdev *rdev = conf->mirrors[d].rdev;
2938                 if (rdev && rdev->raid_disk != d2) {
2939                         sysfs_unlink_rdev(mddev, rdev);
2940                         rdev->raid_disk = d2;
2941                         sysfs_unlink_rdev(mddev, rdev);
2942                         if (sysfs_link_rdev(mddev, rdev))
2943                                 printk(KERN_WARNING
2944                                        "md/raid1:%s: cannot register rd%d\n",
2945                                        mdname(mddev), rdev->raid_disk);
2946                 }
2947                 if (rdev)
2948                         newmirrors[d2++].rdev = rdev;
2949         }
2950         kfree(conf->mirrors);
2951         conf->mirrors = newmirrors;
2952         kfree(conf->poolinfo);
2953         conf->poolinfo = newpoolinfo;
2954
2955         spin_lock_irqsave(&conf->device_lock, flags);
2956         mddev->degraded += (raid_disks - conf->raid_disks);
2957         spin_unlock_irqrestore(&conf->device_lock, flags);
2958         conf->raid_disks = mddev->raid_disks = raid_disks;
2959         mddev->delta_disks = 0;
2960
2961         lower_barrier(conf);
2962
2963         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2964         md_wakeup_thread(mddev->thread);
2965
2966         mempool_destroy(oldpool);
2967         return 0;
2968 }
2969
2970 static void raid1_quiesce(struct mddev *mddev, int state)
2971 {
2972         struct r1conf *conf = mddev->private;
2973
2974         switch(state) {
2975         case 2: /* wake for suspend */
2976                 wake_up(&conf->wait_barrier);
2977                 break;
2978         case 1:
2979                 raise_barrier(conf);
2980                 break;
2981         case 0:
2982                 lower_barrier(conf);
2983                 break;
2984         }
2985 }
2986
2987 static void *raid1_takeover(struct mddev *mddev)
2988 {
2989         /* raid1 can take over:
2990          *  raid5 with 2 devices, any layout or chunk size
2991          */
2992         if (mddev->level == 5 && mddev->raid_disks == 2) {
2993                 struct r1conf *conf;
2994                 mddev->new_level = 1;
2995                 mddev->new_layout = 0;
2996                 mddev->new_chunk_sectors = 0;
2997                 conf = setup_conf(mddev);
2998                 if (!IS_ERR(conf))
2999                         conf->barrier = 1;
3000                 return conf;
3001         }
3002         return ERR_PTR(-EINVAL);
3003 }
3004
3005 static struct md_personality raid1_personality =
3006 {
3007         .name           = "raid1",
3008         .level          = 1,
3009         .owner          = THIS_MODULE,
3010         .make_request   = make_request,
3011         .run            = run,
3012         .stop           = stop,
3013         .status         = status,
3014         .error_handler  = error,
3015         .hot_add_disk   = raid1_add_disk,
3016         .hot_remove_disk= raid1_remove_disk,
3017         .spare_active   = raid1_spare_active,
3018         .sync_request   = sync_request,
3019         .resize         = raid1_resize,
3020         .size           = raid1_size,
3021         .check_reshape  = raid1_reshape,
3022         .quiesce        = raid1_quiesce,
3023         .takeover       = raid1_takeover,
3024 };
3025
3026 static int __init raid_init(void)
3027 {
3028         return register_md_personality(&raid1_personality);
3029 }
3030
3031 static void raid_exit(void)
3032 {
3033         unregister_md_personality(&raid1_personality);
3034 }
3035
3036 module_init(raid_init);
3037 module_exit(raid_exit);
3038 MODULE_LICENSE("GPL");
3039 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3040 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3041 MODULE_ALIAS("md-raid1");
3042 MODULE_ALIAS("md-level-1");
3043
3044 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);