32323f0afd8954714401cb8ed8c5b455b32a9743
[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/seq_file.h>
38 #include <linux/ratelimit.h>
39 #include "md.h"
40 #include "raid1.h"
41 #include "bitmap.h"
42
43 #define DEBUG 0
44 #define PRINTK(x...) do { if (DEBUG) printk(x); } while (0)
45
46 /*
47  * Number of guaranteed r1bios in case of extreme VM load:
48  */
49 #define NR_RAID1_BIOS 256
50
51
52 static void allow_barrier(conf_t *conf);
53 static void lower_barrier(conf_t *conf);
54
55 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
56 {
57         struct pool_info *pi = data;
58         int size = offsetof(r1bio_t, bios[pi->raid_disks]);
59
60         /* allocate a r1bio with room for raid_disks entries in the bios array */
61         return kzalloc(size, gfp_flags);
62 }
63
64 static void r1bio_pool_free(void *r1_bio, void *data)
65 {
66         kfree(r1_bio);
67 }
68
69 #define RESYNC_BLOCK_SIZE (64*1024)
70 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
71 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
72 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
73 #define RESYNC_WINDOW (2048*1024)
74
75 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
76 {
77         struct pool_info *pi = data;
78         struct page *page;
79         r1bio_t *r1_bio;
80         struct bio *bio;
81         int i, j;
82
83         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
84         if (!r1_bio)
85                 return NULL;
86
87         /*
88          * Allocate bios : 1 for reading, n-1 for writing
89          */
90         for (j = pi->raid_disks ; j-- ; ) {
91                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
92                 if (!bio)
93                         goto out_free_bio;
94                 r1_bio->bios[j] = bio;
95         }
96         /*
97          * Allocate RESYNC_PAGES data pages and attach them to
98          * the first bio.
99          * If this is a user-requested check/repair, allocate
100          * RESYNC_PAGES for each bio.
101          */
102         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
103                 j = pi->raid_disks;
104         else
105                 j = 1;
106         while(j--) {
107                 bio = r1_bio->bios[j];
108                 for (i = 0; i < RESYNC_PAGES; i++) {
109                         page = alloc_page(gfp_flags);
110                         if (unlikely(!page))
111                                 goto out_free_pages;
112
113                         bio->bi_io_vec[i].bv_page = page;
114                         bio->bi_vcnt = i+1;
115                 }
116         }
117         /* If not user-requests, copy the page pointers to all bios */
118         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
119                 for (i=0; i<RESYNC_PAGES ; i++)
120                         for (j=1; j<pi->raid_disks; j++)
121                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
122                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
123         }
124
125         r1_bio->master_bio = NULL;
126
127         return r1_bio;
128
129 out_free_pages:
130         for (j=0 ; j < pi->raid_disks; j++)
131                 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
132                         put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
133         j = -1;
134 out_free_bio:
135         while ( ++j < pi->raid_disks )
136                 bio_put(r1_bio->bios[j]);
137         r1bio_pool_free(r1_bio, data);
138         return NULL;
139 }
140
141 static void r1buf_pool_free(void *__r1_bio, void *data)
142 {
143         struct pool_info *pi = data;
144         int i,j;
145         r1bio_t *r1bio = __r1_bio;
146
147         for (i = 0; i < RESYNC_PAGES; i++)
148                 for (j = pi->raid_disks; j-- ;) {
149                         if (j == 0 ||
150                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
151                             r1bio->bios[0]->bi_io_vec[i].bv_page)
152                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
153                 }
154         for (i=0 ; i < pi->raid_disks; i++)
155                 bio_put(r1bio->bios[i]);
156
157         r1bio_pool_free(r1bio, data);
158 }
159
160 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
161 {
162         int i;
163
164         for (i = 0; i < conf->raid_disks; i++) {
165                 struct bio **bio = r1_bio->bios + i;
166                 if (!BIO_SPECIAL(*bio))
167                         bio_put(*bio);
168                 *bio = NULL;
169         }
170 }
171
172 static void free_r1bio(r1bio_t *r1_bio)
173 {
174         conf_t *conf = r1_bio->mddev->private;
175
176         put_all_bios(conf, r1_bio);
177         mempool_free(r1_bio, conf->r1bio_pool);
178 }
179
180 static void put_buf(r1bio_t *r1_bio)
181 {
182         conf_t *conf = r1_bio->mddev->private;
183         int i;
184
185         for (i=0; i<conf->raid_disks; i++) {
186                 struct bio *bio = r1_bio->bios[i];
187                 if (bio->bi_end_io)
188                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
189         }
190
191         mempool_free(r1_bio, conf->r1buf_pool);
192
193         lower_barrier(conf);
194 }
195
196 static void reschedule_retry(r1bio_t *r1_bio)
197 {
198         unsigned long flags;
199         mddev_t *mddev = r1_bio->mddev;
200         conf_t *conf = mddev->private;
201
202         spin_lock_irqsave(&conf->device_lock, flags);
203         list_add(&r1_bio->retry_list, &conf->retry_list);
204         conf->nr_queued ++;
205         spin_unlock_irqrestore(&conf->device_lock, flags);
206
207         wake_up(&conf->wait_barrier);
208         md_wakeup_thread(mddev->thread);
209 }
210
211 /*
212  * raid_end_bio_io() is called when we have finished servicing a mirrored
213  * operation and are ready to return a success/failure code to the buffer
214  * cache layer.
215  */
216 static void call_bio_endio(r1bio_t *r1_bio)
217 {
218         struct bio *bio = r1_bio->master_bio;
219         int done;
220         conf_t *conf = r1_bio->mddev->private;
221
222         if (bio->bi_phys_segments) {
223                 unsigned long flags;
224                 spin_lock_irqsave(&conf->device_lock, flags);
225                 bio->bi_phys_segments--;
226                 done = (bio->bi_phys_segments == 0);
227                 spin_unlock_irqrestore(&conf->device_lock, flags);
228         } else
229                 done = 1;
230
231         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
232                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
233         if (done) {
234                 bio_endio(bio, 0);
235                 /*
236                  * Wake up any possible resync thread that waits for the device
237                  * to go idle.
238                  */
239                 allow_barrier(conf);
240         }
241 }
242
243 static void raid_end_bio_io(r1bio_t *r1_bio)
244 {
245         struct bio *bio = r1_bio->master_bio;
246
247         /* if nobody has done the final endio yet, do it now */
248         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
249                 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
250                         (bio_data_dir(bio) == WRITE) ? "write" : "read",
251                         (unsigned long long) bio->bi_sector,
252                         (unsigned long long) bio->bi_sector +
253                                 (bio->bi_size >> 9) - 1);
254
255                 call_bio_endio(r1_bio);
256         }
257         free_r1bio(r1_bio);
258 }
259
260 /*
261  * Update disk head position estimator based on IRQ completion info.
262  */
263 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
264 {
265         conf_t *conf = r1_bio->mddev->private;
266
267         conf->mirrors[disk].head_position =
268                 r1_bio->sector + (r1_bio->sectors);
269 }
270
271 static void raid1_end_read_request(struct bio *bio, int error)
272 {
273         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
274         r1bio_t *r1_bio = bio->bi_private;
275         int mirror;
276         conf_t *conf = r1_bio->mddev->private;
277
278         mirror = r1_bio->read_disk;
279         /*
280          * this branch is our 'one mirror IO has finished' event handler:
281          */
282         update_head_pos(mirror, r1_bio);
283
284         if (uptodate)
285                 set_bit(R1BIO_Uptodate, &r1_bio->state);
286         else {
287                 /* If all other devices have failed, we want to return
288                  * the error upwards rather than fail the last device.
289                  * Here we redefine "uptodate" to mean "Don't want to retry"
290                  */
291                 unsigned long flags;
292                 spin_lock_irqsave(&conf->device_lock, flags);
293                 if (r1_bio->mddev->degraded == conf->raid_disks ||
294                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
295                      !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
296                         uptodate = 1;
297                 spin_unlock_irqrestore(&conf->device_lock, flags);
298         }
299
300         if (uptodate)
301                 raid_end_bio_io(r1_bio);
302         else {
303                 /*
304                  * oops, read error:
305                  */
306                 char b[BDEVNAME_SIZE];
307                 printk_ratelimited(
308                         KERN_ERR "md/raid1:%s: %s: "
309                         "rescheduling sector %llu\n",
310                         mdname(conf->mddev),
311                         bdevname(conf->mirrors[mirror].rdev->bdev,
312                                  b),
313                         (unsigned long long)r1_bio->sector);
314                 set_bit(R1BIO_ReadError, &r1_bio->state);
315                 reschedule_retry(r1_bio);
316         }
317
318         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
319 }
320
321 static void close_write(r1bio_t *r1_bio)
322 {
323         /* it really is the end of this request */
324         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
325                 /* free extra copy of the data pages */
326                 int i = r1_bio->behind_page_count;
327                 while (i--)
328                         safe_put_page(r1_bio->behind_bvecs[i].bv_page);
329                 kfree(r1_bio->behind_bvecs);
330                 r1_bio->behind_bvecs = NULL;
331         }
332         /* clear the bitmap if all writes complete successfully */
333         bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
334                         r1_bio->sectors,
335                         !test_bit(R1BIO_Degraded, &r1_bio->state),
336                         test_bit(R1BIO_BehindIO, &r1_bio->state));
337         md_write_end(r1_bio->mddev);
338 }
339
340 static void r1_bio_write_done(r1bio_t *r1_bio)
341 {
342         if (!atomic_dec_and_test(&r1_bio->remaining))
343                 return;
344
345         if (test_bit(R1BIO_WriteError, &r1_bio->state))
346                 reschedule_retry(r1_bio);
347         else {
348                 close_write(r1_bio);
349                 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
350                         reschedule_retry(r1_bio);
351                 else
352                         raid_end_bio_io(r1_bio);
353         }
354 }
355
356 static void raid1_end_write_request(struct bio *bio, int error)
357 {
358         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
359         r1bio_t *r1_bio = bio->bi_private;
360         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
361         conf_t *conf = r1_bio->mddev->private;
362         struct bio *to_put = NULL;
363
364
365         for (mirror = 0; mirror < conf->raid_disks; mirror++)
366                 if (r1_bio->bios[mirror] == bio)
367                         break;
368
369         /*
370          * 'one mirror IO has finished' event handler:
371          */
372         if (!uptodate) {
373                 set_bit(WriteErrorSeen,
374                         &conf->mirrors[mirror].rdev->flags);
375                 set_bit(R1BIO_WriteError, &r1_bio->state);
376         } else {
377                 /*
378                  * Set R1BIO_Uptodate in our master bio, so that we
379                  * will return a good error code for to the higher
380                  * levels even if IO on some other mirrored buffer
381                  * fails.
382                  *
383                  * The 'master' represents the composite IO operation
384                  * to user-side. So if something waits for IO, then it
385                  * will wait for the 'master' bio.
386                  */
387                 sector_t first_bad;
388                 int bad_sectors;
389
390                 r1_bio->bios[mirror] = NULL;
391                 to_put = bio;
392                 set_bit(R1BIO_Uptodate, &r1_bio->state);
393
394                 /* Maybe we can clear some bad blocks. */
395                 if (is_badblock(conf->mirrors[mirror].rdev,
396                                 r1_bio->sector, r1_bio->sectors,
397                                 &first_bad, &bad_sectors)) {
398                         r1_bio->bios[mirror] = IO_MADE_GOOD;
399                         set_bit(R1BIO_MadeGood, &r1_bio->state);
400                 }
401         }
402
403         update_head_pos(mirror, r1_bio);
404
405         if (behind) {
406                 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
407                         atomic_dec(&r1_bio->behind_remaining);
408
409                 /*
410                  * In behind mode, we ACK the master bio once the I/O
411                  * has safely reached all non-writemostly
412                  * disks. Setting the Returned bit ensures that this
413                  * gets done only once -- we don't ever want to return
414                  * -EIO here, instead we'll wait
415                  */
416                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
417                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
418                         /* Maybe we can return now */
419                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
420                                 struct bio *mbio = r1_bio->master_bio;
421                                 PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
422                                        (unsigned long long) mbio->bi_sector,
423                                        (unsigned long long) mbio->bi_sector +
424                                        (mbio->bi_size >> 9) - 1);
425                                 call_bio_endio(r1_bio);
426                         }
427                 }
428         }
429         if (r1_bio->bios[mirror] == NULL)
430                 rdev_dec_pending(conf->mirrors[mirror].rdev,
431                                  conf->mddev);
432
433         /*
434          * Let's see if all mirrored write operations have finished
435          * already.
436          */
437         r1_bio_write_done(r1_bio);
438
439         if (to_put)
440                 bio_put(to_put);
441 }
442
443
444 /*
445  * This routine returns the disk from which the requested read should
446  * be done. There is a per-array 'next expected sequential IO' sector
447  * number - if this matches on the next IO then we use the last disk.
448  * There is also a per-disk 'last know head position' sector that is
449  * maintained from IRQ contexts, both the normal and the resync IO
450  * completion handlers update this position correctly. If there is no
451  * perfect sequential match then we pick the disk whose head is closest.
452  *
453  * If there are 2 mirrors in the same 2 devices, performance degrades
454  * because position is mirror, not device based.
455  *
456  * The rdev for the device selected will have nr_pending incremented.
457  */
458 static int read_balance(conf_t *conf, r1bio_t *r1_bio, int *max_sectors)
459 {
460         const sector_t this_sector = r1_bio->sector;
461         int sectors;
462         int best_good_sectors;
463         int start_disk;
464         int best_disk;
465         int i;
466         sector_t best_dist;
467         mdk_rdev_t *rdev;
468         int choose_first;
469
470         rcu_read_lock();
471         /*
472          * Check if we can balance. We can balance on the whole
473          * device if no resync is going on, or below the resync window.
474          * We take the first readable disk when above the resync window.
475          */
476  retry:
477         sectors = r1_bio->sectors;
478         best_disk = -1;
479         best_dist = MaxSector;
480         best_good_sectors = 0;
481
482         if (conf->mddev->recovery_cp < MaxSector &&
483             (this_sector + sectors >= conf->next_resync)) {
484                 choose_first = 1;
485                 start_disk = 0;
486         } else {
487                 choose_first = 0;
488                 start_disk = conf->last_used;
489         }
490
491         for (i = 0 ; i < conf->raid_disks ; i++) {
492                 sector_t dist;
493                 sector_t first_bad;
494                 int bad_sectors;
495
496                 int disk = start_disk + i;
497                 if (disk >= conf->raid_disks)
498                         disk -= conf->raid_disks;
499
500                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
501                 if (r1_bio->bios[disk] == IO_BLOCKED
502                     || rdev == NULL
503                     || test_bit(Faulty, &rdev->flags))
504                         continue;
505                 if (!test_bit(In_sync, &rdev->flags) &&
506                     rdev->recovery_offset < this_sector + sectors)
507                         continue;
508                 if (test_bit(WriteMostly, &rdev->flags)) {
509                         /* Don't balance among write-mostly, just
510                          * use the first as a last resort */
511                         if (best_disk < 0)
512                                 best_disk = disk;
513                         continue;
514                 }
515                 /* This is a reasonable device to use.  It might
516                  * even be best.
517                  */
518                 if (is_badblock(rdev, this_sector, sectors,
519                                 &first_bad, &bad_sectors)) {
520                         if (best_dist < MaxSector)
521                                 /* already have a better device */
522                                 continue;
523                         if (first_bad <= this_sector) {
524                                 /* cannot read here. If this is the 'primary'
525                                  * device, then we must not read beyond
526                                  * bad_sectors from another device..
527                                  */
528                                 bad_sectors -= (this_sector - first_bad);
529                                 if (choose_first && sectors > bad_sectors)
530                                         sectors = bad_sectors;
531                                 if (best_good_sectors > sectors)
532                                         best_good_sectors = sectors;
533
534                         } else {
535                                 sector_t good_sectors = first_bad - this_sector;
536                                 if (good_sectors > best_good_sectors) {
537                                         best_good_sectors = good_sectors;
538                                         best_disk = disk;
539                                 }
540                                 if (choose_first)
541                                         break;
542                         }
543                         continue;
544                 } else
545                         best_good_sectors = sectors;
546
547                 dist = abs(this_sector - conf->mirrors[disk].head_position);
548                 if (choose_first
549                     /* Don't change to another disk for sequential reads */
550                     || conf->next_seq_sect == this_sector
551                     || dist == 0
552                     /* If device is idle, use it */
553                     || atomic_read(&rdev->nr_pending) == 0) {
554                         best_disk = disk;
555                         break;
556                 }
557                 if (dist < best_dist) {
558                         best_dist = dist;
559                         best_disk = disk;
560                 }
561         }
562
563         if (best_disk >= 0) {
564                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
565                 if (!rdev)
566                         goto retry;
567                 atomic_inc(&rdev->nr_pending);
568                 if (test_bit(Faulty, &rdev->flags)) {
569                         /* cannot risk returning a device that failed
570                          * before we inc'ed nr_pending
571                          */
572                         rdev_dec_pending(rdev, conf->mddev);
573                         goto retry;
574                 }
575                 sectors = best_good_sectors;
576                 conf->next_seq_sect = this_sector + sectors;
577                 conf->last_used = best_disk;
578         }
579         rcu_read_unlock();
580         *max_sectors = sectors;
581
582         return best_disk;
583 }
584
585 int md_raid1_congested(mddev_t *mddev, int bits)
586 {
587         conf_t *conf = mddev->private;
588         int i, ret = 0;
589
590         rcu_read_lock();
591         for (i = 0; i < mddev->raid_disks; i++) {
592                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
593                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
594                         struct request_queue *q = bdev_get_queue(rdev->bdev);
595
596                         BUG_ON(!q);
597
598                         /* Note the '|| 1' - when read_balance prefers
599                          * non-congested targets, it can be removed
600                          */
601                         if ((bits & (1<<BDI_async_congested)) || 1)
602                                 ret |= bdi_congested(&q->backing_dev_info, bits);
603                         else
604                                 ret &= bdi_congested(&q->backing_dev_info, bits);
605                 }
606         }
607         rcu_read_unlock();
608         return ret;
609 }
610 EXPORT_SYMBOL_GPL(md_raid1_congested);
611
612 static int raid1_congested(void *data, int bits)
613 {
614         mddev_t *mddev = data;
615
616         return mddev_congested(mddev, bits) ||
617                 md_raid1_congested(mddev, bits);
618 }
619
620 static void flush_pending_writes(conf_t *conf)
621 {
622         /* Any writes that have been queued but are awaiting
623          * bitmap updates get flushed here.
624          */
625         spin_lock_irq(&conf->device_lock);
626
627         if (conf->pending_bio_list.head) {
628                 struct bio *bio;
629                 bio = bio_list_get(&conf->pending_bio_list);
630                 spin_unlock_irq(&conf->device_lock);
631                 /* flush any pending bitmap writes to
632                  * disk before proceeding w/ I/O */
633                 bitmap_unplug(conf->mddev->bitmap);
634
635                 while (bio) { /* submit pending writes */
636                         struct bio *next = bio->bi_next;
637                         bio->bi_next = NULL;
638                         generic_make_request(bio);
639                         bio = next;
640                 }
641         } else
642                 spin_unlock_irq(&conf->device_lock);
643 }
644
645 /* Barriers....
646  * Sometimes we need to suspend IO while we do something else,
647  * either some resync/recovery, or reconfigure the array.
648  * To do this we raise a 'barrier'.
649  * The 'barrier' is a counter that can be raised multiple times
650  * to count how many activities are happening which preclude
651  * normal IO.
652  * We can only raise the barrier if there is no pending IO.
653  * i.e. if nr_pending == 0.
654  * We choose only to raise the barrier if no-one is waiting for the
655  * barrier to go down.  This means that as soon as an IO request
656  * is ready, no other operations which require a barrier will start
657  * until the IO request has had a chance.
658  *
659  * So: regular IO calls 'wait_barrier'.  When that returns there
660  *    is no backgroup IO happening,  It must arrange to call
661  *    allow_barrier when it has finished its IO.
662  * backgroup IO calls must call raise_barrier.  Once that returns
663  *    there is no normal IO happeing.  It must arrange to call
664  *    lower_barrier when the particular background IO completes.
665  */
666 #define RESYNC_DEPTH 32
667
668 static void raise_barrier(conf_t *conf)
669 {
670         spin_lock_irq(&conf->resync_lock);
671
672         /* Wait until no block IO is waiting */
673         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
674                             conf->resync_lock, );
675
676         /* block any new IO from starting */
677         conf->barrier++;
678
679         /* Now wait for all pending IO to complete */
680         wait_event_lock_irq(conf->wait_barrier,
681                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
682                             conf->resync_lock, );
683
684         spin_unlock_irq(&conf->resync_lock);
685 }
686
687 static void lower_barrier(conf_t *conf)
688 {
689         unsigned long flags;
690         BUG_ON(conf->barrier <= 0);
691         spin_lock_irqsave(&conf->resync_lock, flags);
692         conf->barrier--;
693         spin_unlock_irqrestore(&conf->resync_lock, flags);
694         wake_up(&conf->wait_barrier);
695 }
696
697 static void wait_barrier(conf_t *conf)
698 {
699         spin_lock_irq(&conf->resync_lock);
700         if (conf->barrier) {
701                 conf->nr_waiting++;
702                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
703                                     conf->resync_lock,
704                                     );
705                 conf->nr_waiting--;
706         }
707         conf->nr_pending++;
708         spin_unlock_irq(&conf->resync_lock);
709 }
710
711 static void allow_barrier(conf_t *conf)
712 {
713         unsigned long flags;
714         spin_lock_irqsave(&conf->resync_lock, flags);
715         conf->nr_pending--;
716         spin_unlock_irqrestore(&conf->resync_lock, flags);
717         wake_up(&conf->wait_barrier);
718 }
719
720 static void freeze_array(conf_t *conf)
721 {
722         /* stop syncio and normal IO and wait for everything to
723          * go quite.
724          * We increment barrier and nr_waiting, and then
725          * wait until nr_pending match nr_queued+1
726          * This is called in the context of one normal IO request
727          * that has failed. Thus any sync request that might be pending
728          * will be blocked by nr_pending, and we need to wait for
729          * pending IO requests to complete or be queued for re-try.
730          * Thus the number queued (nr_queued) plus this request (1)
731          * must match the number of pending IOs (nr_pending) before
732          * we continue.
733          */
734         spin_lock_irq(&conf->resync_lock);
735         conf->barrier++;
736         conf->nr_waiting++;
737         wait_event_lock_irq(conf->wait_barrier,
738                             conf->nr_pending == conf->nr_queued+1,
739                             conf->resync_lock,
740                             flush_pending_writes(conf));
741         spin_unlock_irq(&conf->resync_lock);
742 }
743 static void unfreeze_array(conf_t *conf)
744 {
745         /* reverse the effect of the freeze */
746         spin_lock_irq(&conf->resync_lock);
747         conf->barrier--;
748         conf->nr_waiting--;
749         wake_up(&conf->wait_barrier);
750         spin_unlock_irq(&conf->resync_lock);
751 }
752
753
754 /* duplicate the data pages for behind I/O 
755  */
756 static void alloc_behind_pages(struct bio *bio, r1bio_t *r1_bio)
757 {
758         int i;
759         struct bio_vec *bvec;
760         struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
761                                         GFP_NOIO);
762         if (unlikely(!bvecs))
763                 return;
764
765         bio_for_each_segment(bvec, bio, i) {
766                 bvecs[i] = *bvec;
767                 bvecs[i].bv_page = alloc_page(GFP_NOIO);
768                 if (unlikely(!bvecs[i].bv_page))
769                         goto do_sync_io;
770                 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
771                        kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
772                 kunmap(bvecs[i].bv_page);
773                 kunmap(bvec->bv_page);
774         }
775         r1_bio->behind_bvecs = bvecs;
776         r1_bio->behind_page_count = bio->bi_vcnt;
777         set_bit(R1BIO_BehindIO, &r1_bio->state);
778         return;
779
780 do_sync_io:
781         for (i = 0; i < bio->bi_vcnt; i++)
782                 if (bvecs[i].bv_page)
783                         put_page(bvecs[i].bv_page);
784         kfree(bvecs);
785         PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
786 }
787
788 static int make_request(mddev_t *mddev, struct bio * bio)
789 {
790         conf_t *conf = mddev->private;
791         mirror_info_t *mirror;
792         r1bio_t *r1_bio;
793         struct bio *read_bio;
794         int i, disks;
795         struct bitmap *bitmap;
796         unsigned long flags;
797         const int rw = bio_data_dir(bio);
798         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
799         const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
800         mdk_rdev_t *blocked_rdev;
801         int plugged;
802         int first_clone;
803         int sectors_handled;
804         int max_sectors;
805
806         /*
807          * Register the new request and wait if the reconstruction
808          * thread has put up a bar for new requests.
809          * Continue immediately if no resync is active currently.
810          */
811
812         md_write_start(mddev, bio); /* wait on superblock update early */
813
814         if (bio_data_dir(bio) == WRITE &&
815             bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
816             bio->bi_sector < mddev->suspend_hi) {
817                 /* As the suspend_* range is controlled by
818                  * userspace, we want an interruptible
819                  * wait.
820                  */
821                 DEFINE_WAIT(w);
822                 for (;;) {
823                         flush_signals(current);
824                         prepare_to_wait(&conf->wait_barrier,
825                                         &w, TASK_INTERRUPTIBLE);
826                         if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
827                             bio->bi_sector >= mddev->suspend_hi)
828                                 break;
829                         schedule();
830                 }
831                 finish_wait(&conf->wait_barrier, &w);
832         }
833
834         wait_barrier(conf);
835
836         bitmap = mddev->bitmap;
837
838         /*
839          * make_request() can abort the operation when READA is being
840          * used and no empty request is available.
841          *
842          */
843         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
844
845         r1_bio->master_bio = bio;
846         r1_bio->sectors = bio->bi_size >> 9;
847         r1_bio->state = 0;
848         r1_bio->mddev = mddev;
849         r1_bio->sector = bio->bi_sector;
850
851         /* We might need to issue multiple reads to different
852          * devices if there are bad blocks around, so we keep
853          * track of the number of reads in bio->bi_phys_segments.
854          * If this is 0, there is only one r1_bio and no locking
855          * will be needed when requests complete.  If it is
856          * non-zero, then it is the number of not-completed requests.
857          */
858         bio->bi_phys_segments = 0;
859         clear_bit(BIO_SEG_VALID, &bio->bi_flags);
860
861         if (rw == READ) {
862                 /*
863                  * read balancing logic:
864                  */
865                 int rdisk;
866
867 read_again:
868                 rdisk = read_balance(conf, r1_bio, &max_sectors);
869
870                 if (rdisk < 0) {
871                         /* couldn't find anywhere to read from */
872                         raid_end_bio_io(r1_bio);
873                         return 0;
874                 }
875                 mirror = conf->mirrors + rdisk;
876
877                 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
878                     bitmap) {
879                         /* Reading from a write-mostly device must
880                          * take care not to over-take any writes
881                          * that are 'behind'
882                          */
883                         wait_event(bitmap->behind_wait,
884                                    atomic_read(&bitmap->behind_writes) == 0);
885                 }
886                 r1_bio->read_disk = rdisk;
887
888                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
889                 md_trim_bio(read_bio, r1_bio->sector - bio->bi_sector,
890                             max_sectors);
891
892                 r1_bio->bios[rdisk] = read_bio;
893
894                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
895                 read_bio->bi_bdev = mirror->rdev->bdev;
896                 read_bio->bi_end_io = raid1_end_read_request;
897                 read_bio->bi_rw = READ | do_sync;
898                 read_bio->bi_private = r1_bio;
899
900                 if (max_sectors < r1_bio->sectors) {
901                         /* could not read all from this device, so we will
902                          * need another r1_bio.
903                          */
904
905                         sectors_handled = (r1_bio->sector + max_sectors
906                                            - bio->bi_sector);
907                         r1_bio->sectors = max_sectors;
908                         spin_lock_irq(&conf->device_lock);
909                         if (bio->bi_phys_segments == 0)
910                                 bio->bi_phys_segments = 2;
911                         else
912                                 bio->bi_phys_segments++;
913                         spin_unlock_irq(&conf->device_lock);
914                         /* Cannot call generic_make_request directly
915                          * as that will be queued in __make_request
916                          * and subsequent mempool_alloc might block waiting
917                          * for it.  So hand bio over to raid1d.
918                          */
919                         reschedule_retry(r1_bio);
920
921                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
922
923                         r1_bio->master_bio = bio;
924                         r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
925                         r1_bio->state = 0;
926                         r1_bio->mddev = mddev;
927                         r1_bio->sector = bio->bi_sector + sectors_handled;
928                         goto read_again;
929                 } else
930                         generic_make_request(read_bio);
931                 return 0;
932         }
933
934         /*
935          * WRITE:
936          */
937         /* first select target devices under rcu_lock and
938          * inc refcount on their rdev.  Record them by setting
939          * bios[x] to bio
940          * If there are known/acknowledged bad blocks on any device on
941          * which we have seen a write error, we want to avoid writing those
942          * blocks.
943          * This potentially requires several writes to write around
944          * the bad blocks.  Each set of writes gets it's own r1bio
945          * with a set of bios attached.
946          */
947         plugged = mddev_check_plugged(mddev);
948
949         disks = conf->raid_disks;
950  retry_write:
951         blocked_rdev = NULL;
952         rcu_read_lock();
953         max_sectors = r1_bio->sectors;
954         for (i = 0;  i < disks; i++) {
955                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
956                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
957                         atomic_inc(&rdev->nr_pending);
958                         blocked_rdev = rdev;
959                         break;
960                 }
961                 r1_bio->bios[i] = NULL;
962                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
963                         set_bit(R1BIO_Degraded, &r1_bio->state);
964                         continue;
965                 }
966
967                 atomic_inc(&rdev->nr_pending);
968                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
969                         sector_t first_bad;
970                         int bad_sectors;
971                         int is_bad;
972
973                         is_bad = is_badblock(rdev, r1_bio->sector,
974                                              max_sectors,
975                                              &first_bad, &bad_sectors);
976                         if (is_bad < 0) {
977                                 /* mustn't write here until the bad block is
978                                  * acknowledged*/
979                                 set_bit(BlockedBadBlocks, &rdev->flags);
980                                 blocked_rdev = rdev;
981                                 break;
982                         }
983                         if (is_bad && first_bad <= r1_bio->sector) {
984                                 /* Cannot write here at all */
985                                 bad_sectors -= (r1_bio->sector - first_bad);
986                                 if (bad_sectors < max_sectors)
987                                         /* mustn't write more than bad_sectors
988                                          * to other devices yet
989                                          */
990                                         max_sectors = bad_sectors;
991                                 rdev_dec_pending(rdev, mddev);
992                                 /* We don't set R1BIO_Degraded as that
993                                  * only applies if the disk is
994                                  * missing, so it might be re-added,
995                                  * and we want to know to recover this
996                                  * chunk.
997                                  * In this case the device is here,
998                                  * and the fact that this chunk is not
999                                  * in-sync is recorded in the bad
1000                                  * block log
1001                                  */
1002                                 continue;
1003                         }
1004                         if (is_bad) {
1005                                 int good_sectors = first_bad - r1_bio->sector;
1006                                 if (good_sectors < max_sectors)
1007                                         max_sectors = good_sectors;
1008                         }
1009                 }
1010                 r1_bio->bios[i] = bio;
1011         }
1012         rcu_read_unlock();
1013
1014         if (unlikely(blocked_rdev)) {
1015                 /* Wait for this device to become unblocked */
1016                 int j;
1017
1018                 for (j = 0; j < i; j++)
1019                         if (r1_bio->bios[j])
1020                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1021                 r1_bio->state = 0;
1022                 allow_barrier(conf);
1023                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1024                 wait_barrier(conf);
1025                 goto retry_write;
1026         }
1027
1028         if (max_sectors < r1_bio->sectors) {
1029                 /* We are splitting this write into multiple parts, so
1030                  * we need to prepare for allocating another r1_bio.
1031                  */
1032                 r1_bio->sectors = max_sectors;
1033                 spin_lock_irq(&conf->device_lock);
1034                 if (bio->bi_phys_segments == 0)
1035                         bio->bi_phys_segments = 2;
1036                 else
1037                         bio->bi_phys_segments++;
1038                 spin_unlock_irq(&conf->device_lock);
1039         }
1040         sectors_handled = r1_bio->sector + max_sectors - bio->bi_sector;
1041
1042         atomic_set(&r1_bio->remaining, 1);
1043         atomic_set(&r1_bio->behind_remaining, 0);
1044
1045         first_clone = 1;
1046         for (i = 0; i < disks; i++) {
1047                 struct bio *mbio;
1048                 if (!r1_bio->bios[i])
1049                         continue;
1050
1051                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1052                 md_trim_bio(mbio, r1_bio->sector - bio->bi_sector, max_sectors);
1053
1054                 if (first_clone) {
1055                         /* do behind I/O ?
1056                          * Not if there are too many, or cannot
1057                          * allocate memory, or a reader on WriteMostly
1058                          * is waiting for behind writes to flush */
1059                         if (bitmap &&
1060                             (atomic_read(&bitmap->behind_writes)
1061                              < mddev->bitmap_info.max_write_behind) &&
1062                             !waitqueue_active(&bitmap->behind_wait))
1063                                 alloc_behind_pages(mbio, r1_bio);
1064
1065                         bitmap_startwrite(bitmap, r1_bio->sector,
1066                                           r1_bio->sectors,
1067                                           test_bit(R1BIO_BehindIO,
1068                                                    &r1_bio->state));
1069                         first_clone = 0;
1070                 }
1071                 if (r1_bio->behind_bvecs) {
1072                         struct bio_vec *bvec;
1073                         int j;
1074
1075                         /* Yes, I really want the '__' version so that
1076                          * we clear any unused pointer in the io_vec, rather
1077                          * than leave them unchanged.  This is important
1078                          * because when we come to free the pages, we won't
1079                          * know the original bi_idx, so we just free
1080                          * them all
1081                          */
1082                         __bio_for_each_segment(bvec, mbio, j, 0)
1083                                 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1084                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1085                                 atomic_inc(&r1_bio->behind_remaining);
1086                 }
1087
1088                 r1_bio->bios[i] = mbio;
1089
1090                 mbio->bi_sector = (r1_bio->sector +
1091                                    conf->mirrors[i].rdev->data_offset);
1092                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1093                 mbio->bi_end_io = raid1_end_write_request;
1094                 mbio->bi_rw = WRITE | do_flush_fua | do_sync;
1095                 mbio->bi_private = r1_bio;
1096
1097                 atomic_inc(&r1_bio->remaining);
1098                 spin_lock_irqsave(&conf->device_lock, flags);
1099                 bio_list_add(&conf->pending_bio_list, mbio);
1100                 spin_unlock_irqrestore(&conf->device_lock, flags);
1101         }
1102         r1_bio_write_done(r1_bio);
1103
1104         /* In case raid1d snuck in to freeze_array */
1105         wake_up(&conf->wait_barrier);
1106
1107         if (sectors_handled < (bio->bi_size >> 9)) {
1108                 /* We need another r1_bio.  It has already been counted
1109                  * in bio->bi_phys_segments
1110                  */
1111                 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1112                 r1_bio->master_bio = bio;
1113                 r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1114                 r1_bio->state = 0;
1115                 r1_bio->mddev = mddev;
1116                 r1_bio->sector = bio->bi_sector + sectors_handled;
1117                 goto retry_write;
1118         }
1119
1120         if (do_sync || !bitmap || !plugged)
1121                 md_wakeup_thread(mddev->thread);
1122
1123         return 0;
1124 }
1125
1126 static void status(struct seq_file *seq, mddev_t *mddev)
1127 {
1128         conf_t *conf = mddev->private;
1129         int i;
1130
1131         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1132                    conf->raid_disks - mddev->degraded);
1133         rcu_read_lock();
1134         for (i = 0; i < conf->raid_disks; i++) {
1135                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1136                 seq_printf(seq, "%s",
1137                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1138         }
1139         rcu_read_unlock();
1140         seq_printf(seq, "]");
1141 }
1142
1143
1144 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1145 {
1146         char b[BDEVNAME_SIZE];
1147         conf_t *conf = mddev->private;
1148
1149         /*
1150          * If it is not operational, then we have already marked it as dead
1151          * else if it is the last working disks, ignore the error, let the
1152          * next level up know.
1153          * else mark the drive as failed
1154          */
1155         if (test_bit(In_sync, &rdev->flags)
1156             && (conf->raid_disks - mddev->degraded) == 1) {
1157                 /*
1158                  * Don't fail the drive, act as though we were just a
1159                  * normal single drive.
1160                  * However don't try a recovery from this drive as
1161                  * it is very likely to fail.
1162                  */
1163                 conf->recovery_disabled = mddev->recovery_disabled;
1164                 return;
1165         }
1166         set_bit(Blocked, &rdev->flags);
1167         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1168                 unsigned long flags;
1169                 spin_lock_irqsave(&conf->device_lock, flags);
1170                 mddev->degraded++;
1171                 set_bit(Faulty, &rdev->flags);
1172                 spin_unlock_irqrestore(&conf->device_lock, flags);
1173                 /*
1174                  * if recovery is running, make sure it aborts.
1175                  */
1176                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1177         } else
1178                 set_bit(Faulty, &rdev->flags);
1179         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1180         printk(KERN_ALERT
1181                "md/raid1:%s: Disk failure on %s, disabling device.\n"
1182                "md/raid1:%s: Operation continuing on %d devices.\n",
1183                mdname(mddev), bdevname(rdev->bdev, b),
1184                mdname(mddev), conf->raid_disks - mddev->degraded);
1185 }
1186
1187 static void print_conf(conf_t *conf)
1188 {
1189         int i;
1190
1191         printk(KERN_DEBUG "RAID1 conf printout:\n");
1192         if (!conf) {
1193                 printk(KERN_DEBUG "(!conf)\n");
1194                 return;
1195         }
1196         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1197                 conf->raid_disks);
1198
1199         rcu_read_lock();
1200         for (i = 0; i < conf->raid_disks; i++) {
1201                 char b[BDEVNAME_SIZE];
1202                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1203                 if (rdev)
1204                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1205                                i, !test_bit(In_sync, &rdev->flags),
1206                                !test_bit(Faulty, &rdev->flags),
1207                                bdevname(rdev->bdev,b));
1208         }
1209         rcu_read_unlock();
1210 }
1211
1212 static void close_sync(conf_t *conf)
1213 {
1214         wait_barrier(conf);
1215         allow_barrier(conf);
1216
1217         mempool_destroy(conf->r1buf_pool);
1218         conf->r1buf_pool = NULL;
1219 }
1220
1221 static int raid1_spare_active(mddev_t *mddev)
1222 {
1223         int i;
1224         conf_t *conf = mddev->private;
1225         int count = 0;
1226         unsigned long flags;
1227
1228         /*
1229          * Find all failed disks within the RAID1 configuration 
1230          * and mark them readable.
1231          * Called under mddev lock, so rcu protection not needed.
1232          */
1233         for (i = 0; i < conf->raid_disks; i++) {
1234                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1235                 if (rdev
1236                     && !test_bit(Faulty, &rdev->flags)
1237                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1238                         count++;
1239                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1240                 }
1241         }
1242         spin_lock_irqsave(&conf->device_lock, flags);
1243         mddev->degraded -= count;
1244         spin_unlock_irqrestore(&conf->device_lock, flags);
1245
1246         print_conf(conf);
1247         return count;
1248 }
1249
1250
1251 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1252 {
1253         conf_t *conf = mddev->private;
1254         int err = -EEXIST;
1255         int mirror = 0;
1256         mirror_info_t *p;
1257         int first = 0;
1258         int last = mddev->raid_disks - 1;
1259
1260         if (mddev->recovery_disabled == conf->recovery_disabled)
1261                 return -EBUSY;
1262
1263         if (rdev->raid_disk >= 0)
1264                 first = last = rdev->raid_disk;
1265
1266         for (mirror = first; mirror <= last; mirror++)
1267                 if ( !(p=conf->mirrors+mirror)->rdev) {
1268
1269                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1270                                           rdev->data_offset << 9);
1271                         /* as we don't honour merge_bvec_fn, we must
1272                          * never risk violating it, so limit
1273                          * ->max_segments to one lying with a single
1274                          * page, as a one page request is never in
1275                          * violation.
1276                          */
1277                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1278                                 blk_queue_max_segments(mddev->queue, 1);
1279                                 blk_queue_segment_boundary(mddev->queue,
1280                                                            PAGE_CACHE_SIZE - 1);
1281                         }
1282
1283                         p->head_position = 0;
1284                         rdev->raid_disk = mirror;
1285                         err = 0;
1286                         /* As all devices are equivalent, we don't need a full recovery
1287                          * if this was recently any drive of the array
1288                          */
1289                         if (rdev->saved_raid_disk < 0)
1290                                 conf->fullsync = 1;
1291                         rcu_assign_pointer(p->rdev, rdev);
1292                         break;
1293                 }
1294         md_integrity_add_rdev(rdev, mddev);
1295         print_conf(conf);
1296         return err;
1297 }
1298
1299 static int raid1_remove_disk(mddev_t *mddev, int number)
1300 {
1301         conf_t *conf = mddev->private;
1302         int err = 0;
1303         mdk_rdev_t *rdev;
1304         mirror_info_t *p = conf->mirrors+ number;
1305
1306         print_conf(conf);
1307         rdev = p->rdev;
1308         if (rdev) {
1309                 if (test_bit(In_sync, &rdev->flags) ||
1310                     atomic_read(&rdev->nr_pending)) {
1311                         err = -EBUSY;
1312                         goto abort;
1313                 }
1314                 /* Only remove non-faulty devices if recovery
1315                  * is not possible.
1316                  */
1317                 if (!test_bit(Faulty, &rdev->flags) &&
1318                     mddev->recovery_disabled != conf->recovery_disabled &&
1319                     mddev->degraded < conf->raid_disks) {
1320                         err = -EBUSY;
1321                         goto abort;
1322                 }
1323                 p->rdev = NULL;
1324                 synchronize_rcu();
1325                 if (atomic_read(&rdev->nr_pending)) {
1326                         /* lost the race, try later */
1327                         err = -EBUSY;
1328                         p->rdev = rdev;
1329                         goto abort;
1330                 }
1331                 err = md_integrity_register(mddev);
1332         }
1333 abort:
1334
1335         print_conf(conf);
1336         return err;
1337 }
1338
1339
1340 static void end_sync_read(struct bio *bio, int error)
1341 {
1342         r1bio_t *r1_bio = bio->bi_private;
1343         int i;
1344
1345         for (i=r1_bio->mddev->raid_disks; i--; )
1346                 if (r1_bio->bios[i] == bio)
1347                         break;
1348         BUG_ON(i < 0);
1349         update_head_pos(i, r1_bio);
1350         /*
1351          * we have read a block, now it needs to be re-written,
1352          * or re-read if the read failed.
1353          * We don't do much here, just schedule handling by raid1d
1354          */
1355         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1356                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1357
1358         if (atomic_dec_and_test(&r1_bio->remaining))
1359                 reschedule_retry(r1_bio);
1360 }
1361
1362 static void end_sync_write(struct bio *bio, int error)
1363 {
1364         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1365         r1bio_t *r1_bio = bio->bi_private;
1366         mddev_t *mddev = r1_bio->mddev;
1367         conf_t *conf = mddev->private;
1368         int i;
1369         int mirror=0;
1370         sector_t first_bad;
1371         int bad_sectors;
1372
1373         for (i = 0; i < conf->raid_disks; i++)
1374                 if (r1_bio->bios[i] == bio) {
1375                         mirror = i;
1376                         break;
1377                 }
1378         if (!uptodate) {
1379                 sector_t sync_blocks = 0;
1380                 sector_t s = r1_bio->sector;
1381                 long sectors_to_go = r1_bio->sectors;
1382                 /* make sure these bits doesn't get cleared. */
1383                 do {
1384                         bitmap_end_sync(mddev->bitmap, s,
1385                                         &sync_blocks, 1);
1386                         s += sync_blocks;
1387                         sectors_to_go -= sync_blocks;
1388                 } while (sectors_to_go > 0);
1389                 set_bit(WriteErrorSeen,
1390                         &conf->mirrors[mirror].rdev->flags);
1391                 set_bit(R1BIO_WriteError, &r1_bio->state);
1392         } else if (is_badblock(conf->mirrors[mirror].rdev,
1393                                r1_bio->sector,
1394                                r1_bio->sectors,
1395                                &first_bad, &bad_sectors) &&
1396                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1397                                 r1_bio->sector,
1398                                 r1_bio->sectors,
1399                                 &first_bad, &bad_sectors)
1400                 )
1401                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1402
1403         update_head_pos(mirror, r1_bio);
1404
1405         if (atomic_dec_and_test(&r1_bio->remaining)) {
1406                 int s = r1_bio->sectors;
1407                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1408                     test_bit(R1BIO_WriteError, &r1_bio->state))
1409                         reschedule_retry(r1_bio);
1410                 else {
1411                         put_buf(r1_bio);
1412                         md_done_sync(mddev, s, uptodate);
1413                 }
1414         }
1415 }
1416
1417 static int r1_sync_page_io(mdk_rdev_t *rdev, sector_t sector,
1418                             int sectors, struct page *page, int rw)
1419 {
1420         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1421                 /* success */
1422                 return 1;
1423         if (rw == WRITE)
1424                 set_bit(WriteErrorSeen, &rdev->flags);
1425         /* need to record an error - either for the block or the device */
1426         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1427                 md_error(rdev->mddev, rdev);
1428         return 0;
1429 }
1430
1431 static int fix_sync_read_error(r1bio_t *r1_bio)
1432 {
1433         /* Try some synchronous reads of other devices to get
1434          * good data, much like with normal read errors.  Only
1435          * read into the pages we already have so we don't
1436          * need to re-issue the read request.
1437          * We don't need to freeze the array, because being in an
1438          * active sync request, there is no normal IO, and
1439          * no overlapping syncs.
1440          * We don't need to check is_badblock() again as we
1441          * made sure that anything with a bad block in range
1442          * will have bi_end_io clear.
1443          */
1444         mddev_t *mddev = r1_bio->mddev;
1445         conf_t *conf = mddev->private;
1446         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1447         sector_t sect = r1_bio->sector;
1448         int sectors = r1_bio->sectors;
1449         int idx = 0;
1450
1451         while(sectors) {
1452                 int s = sectors;
1453                 int d = r1_bio->read_disk;
1454                 int success = 0;
1455                 mdk_rdev_t *rdev;
1456                 int start;
1457
1458                 if (s > (PAGE_SIZE>>9))
1459                         s = PAGE_SIZE >> 9;
1460                 do {
1461                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1462                                 /* No rcu protection needed here devices
1463                                  * can only be removed when no resync is
1464                                  * active, and resync is currently active
1465                                  */
1466                                 rdev = conf->mirrors[d].rdev;
1467                                 if (sync_page_io(rdev, sect, s<<9,
1468                                                  bio->bi_io_vec[idx].bv_page,
1469                                                  READ, false)) {
1470                                         success = 1;
1471                                         break;
1472                                 }
1473                         }
1474                         d++;
1475                         if (d == conf->raid_disks)
1476                                 d = 0;
1477                 } while (!success && d != r1_bio->read_disk);
1478
1479                 if (!success) {
1480                         char b[BDEVNAME_SIZE];
1481                         int abort = 0;
1482                         /* Cannot read from anywhere, this block is lost.
1483                          * Record a bad block on each device.  If that doesn't
1484                          * work just disable and interrupt the recovery.
1485                          * Don't fail devices as that won't really help.
1486                          */
1487                         printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1488                                " for block %llu\n",
1489                                mdname(mddev),
1490                                bdevname(bio->bi_bdev, b),
1491                                (unsigned long long)r1_bio->sector);
1492                         for (d = 0; d < conf->raid_disks; d++) {
1493                                 rdev = conf->mirrors[d].rdev;
1494                                 if (!rdev || test_bit(Faulty, &rdev->flags))
1495                                         continue;
1496                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
1497                                         abort = 1;
1498                         }
1499                         if (abort) {
1500                                 mddev->recovery_disabled = 1;
1501                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1502                                 md_done_sync(mddev, r1_bio->sectors, 0);
1503                                 put_buf(r1_bio);
1504                                 return 0;
1505                         }
1506                         /* Try next page */
1507                         sectors -= s;
1508                         sect += s;
1509                         idx++;
1510                         continue;
1511                 }
1512
1513                 start = d;
1514                 /* write it back and re-read */
1515                 while (d != r1_bio->read_disk) {
1516                         if (d == 0)
1517                                 d = conf->raid_disks;
1518                         d--;
1519                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1520                                 continue;
1521                         rdev = conf->mirrors[d].rdev;
1522                         if (r1_sync_page_io(rdev, sect, s,
1523                                             bio->bi_io_vec[idx].bv_page,
1524                                             WRITE) == 0) {
1525                                 r1_bio->bios[d]->bi_end_io = NULL;
1526                                 rdev_dec_pending(rdev, mddev);
1527                         }
1528                 }
1529                 d = start;
1530                 while (d != r1_bio->read_disk) {
1531                         if (d == 0)
1532                                 d = conf->raid_disks;
1533                         d--;
1534                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1535                                 continue;
1536                         rdev = conf->mirrors[d].rdev;
1537                         if (r1_sync_page_io(rdev, sect, s,
1538                                             bio->bi_io_vec[idx].bv_page,
1539                                             READ) != 0)
1540                                 atomic_add(s, &rdev->corrected_errors);
1541                 }
1542                 sectors -= s;
1543                 sect += s;
1544                 idx ++;
1545         }
1546         set_bit(R1BIO_Uptodate, &r1_bio->state);
1547         set_bit(BIO_UPTODATE, &bio->bi_flags);
1548         return 1;
1549 }
1550
1551 static int process_checks(r1bio_t *r1_bio)
1552 {
1553         /* We have read all readable devices.  If we haven't
1554          * got the block, then there is no hope left.
1555          * If we have, then we want to do a comparison
1556          * and skip the write if everything is the same.
1557          * If any blocks failed to read, then we need to
1558          * attempt an over-write
1559          */
1560         mddev_t *mddev = r1_bio->mddev;
1561         conf_t *conf = mddev->private;
1562         int primary;
1563         int i;
1564
1565         for (primary = 0; primary < conf->raid_disks; primary++)
1566                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1567                     test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1568                         r1_bio->bios[primary]->bi_end_io = NULL;
1569                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1570                         break;
1571                 }
1572         r1_bio->read_disk = primary;
1573         for (i = 0; i < conf->raid_disks; i++) {
1574                 int j;
1575                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1576                 struct bio *pbio = r1_bio->bios[primary];
1577                 struct bio *sbio = r1_bio->bios[i];
1578                 int size;
1579
1580                 if (r1_bio->bios[i]->bi_end_io != end_sync_read)
1581                         continue;
1582
1583                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1584                         for (j = vcnt; j-- ; ) {
1585                                 struct page *p, *s;
1586                                 p = pbio->bi_io_vec[j].bv_page;
1587                                 s = sbio->bi_io_vec[j].bv_page;
1588                                 if (memcmp(page_address(p),
1589                                            page_address(s),
1590                                            PAGE_SIZE))
1591                                         break;
1592                         }
1593                 } else
1594                         j = 0;
1595                 if (j >= 0)
1596                         mddev->resync_mismatches += r1_bio->sectors;
1597                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1598                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1599                         /* No need to write to this device. */
1600                         sbio->bi_end_io = NULL;
1601                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1602                         continue;
1603                 }
1604                 /* fixup the bio for reuse */
1605                 sbio->bi_vcnt = vcnt;
1606                 sbio->bi_size = r1_bio->sectors << 9;
1607                 sbio->bi_idx = 0;
1608                 sbio->bi_phys_segments = 0;
1609                 sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1610                 sbio->bi_flags |= 1 << BIO_UPTODATE;
1611                 sbio->bi_next = NULL;
1612                 sbio->bi_sector = r1_bio->sector +
1613                         conf->mirrors[i].rdev->data_offset;
1614                 sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1615                 size = sbio->bi_size;
1616                 for (j = 0; j < vcnt ; j++) {
1617                         struct bio_vec *bi;
1618                         bi = &sbio->bi_io_vec[j];
1619                         bi->bv_offset = 0;
1620                         if (size > PAGE_SIZE)
1621                                 bi->bv_len = PAGE_SIZE;
1622                         else
1623                                 bi->bv_len = size;
1624                         size -= PAGE_SIZE;
1625                         memcpy(page_address(bi->bv_page),
1626                                page_address(pbio->bi_io_vec[j].bv_page),
1627                                PAGE_SIZE);
1628                 }
1629         }
1630         return 0;
1631 }
1632
1633 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1634 {
1635         conf_t *conf = mddev->private;
1636         int i;
1637         int disks = conf->raid_disks;
1638         struct bio *bio, *wbio;
1639
1640         bio = r1_bio->bios[r1_bio->read_disk];
1641
1642         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1643                 /* ouch - failed to read all of that. */
1644                 if (!fix_sync_read_error(r1_bio))
1645                         return;
1646
1647         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1648                 if (process_checks(r1_bio) < 0)
1649                         return;
1650         /*
1651          * schedule writes
1652          */
1653         atomic_set(&r1_bio->remaining, 1);
1654         for (i = 0; i < disks ; i++) {
1655                 wbio = r1_bio->bios[i];
1656                 if (wbio->bi_end_io == NULL ||
1657                     (wbio->bi_end_io == end_sync_read &&
1658                      (i == r1_bio->read_disk ||
1659                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1660                         continue;
1661
1662                 wbio->bi_rw = WRITE;
1663                 wbio->bi_end_io = end_sync_write;
1664                 atomic_inc(&r1_bio->remaining);
1665                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1666
1667                 generic_make_request(wbio);
1668         }
1669
1670         if (atomic_dec_and_test(&r1_bio->remaining)) {
1671                 /* if we're here, all write(s) have completed, so clean up */
1672                 md_done_sync(mddev, r1_bio->sectors, 1);
1673                 put_buf(r1_bio);
1674         }
1675 }
1676
1677 /*
1678  * This is a kernel thread which:
1679  *
1680  *      1.      Retries failed read operations on working mirrors.
1681  *      2.      Updates the raid superblock when problems encounter.
1682  *      3.      Performs writes following reads for array synchronising.
1683  */
1684
1685 static void fix_read_error(conf_t *conf, int read_disk,
1686                            sector_t sect, int sectors)
1687 {
1688         mddev_t *mddev = conf->mddev;
1689         while(sectors) {
1690                 int s = sectors;
1691                 int d = read_disk;
1692                 int success = 0;
1693                 int start;
1694                 mdk_rdev_t *rdev;
1695
1696                 if (s > (PAGE_SIZE>>9))
1697                         s = PAGE_SIZE >> 9;
1698
1699                 do {
1700                         /* Note: no rcu protection needed here
1701                          * as this is synchronous in the raid1d thread
1702                          * which is the thread that might remove
1703                          * a device.  If raid1d ever becomes multi-threaded....
1704                          */
1705                         sector_t first_bad;
1706                         int bad_sectors;
1707
1708                         rdev = conf->mirrors[d].rdev;
1709                         if (rdev &&
1710                             test_bit(In_sync, &rdev->flags) &&
1711                             is_badblock(rdev, sect, s,
1712                                         &first_bad, &bad_sectors) == 0 &&
1713                             sync_page_io(rdev, sect, s<<9,
1714                                          conf->tmppage, READ, false))
1715                                 success = 1;
1716                         else {
1717                                 d++;
1718                                 if (d == conf->raid_disks)
1719                                         d = 0;
1720                         }
1721                 } while (!success && d != read_disk);
1722
1723                 if (!success) {
1724                         /* Cannot read from anywhere - mark it bad */
1725                         mdk_rdev_t *rdev = conf->mirrors[read_disk].rdev;
1726                         if (!rdev_set_badblocks(rdev, sect, s, 0))
1727                                 md_error(mddev, rdev);
1728                         break;
1729                 }
1730                 /* write it back and re-read */
1731                 start = d;
1732                 while (d != read_disk) {
1733                         if (d==0)
1734                                 d = conf->raid_disks;
1735                         d--;
1736                         rdev = conf->mirrors[d].rdev;
1737                         if (rdev &&
1738                             test_bit(In_sync, &rdev->flags))
1739                                 r1_sync_page_io(rdev, sect, s,
1740                                                 conf->tmppage, WRITE);
1741                 }
1742                 d = start;
1743                 while (d != read_disk) {
1744                         char b[BDEVNAME_SIZE];
1745                         if (d==0)
1746                                 d = conf->raid_disks;
1747                         d--;
1748                         rdev = conf->mirrors[d].rdev;
1749                         if (rdev &&
1750                             test_bit(In_sync, &rdev->flags)) {
1751                                 if (r1_sync_page_io(rdev, sect, s,
1752                                                     conf->tmppage, READ)) {
1753                                         atomic_add(s, &rdev->corrected_errors);
1754                                         printk(KERN_INFO
1755                                                "md/raid1:%s: read error corrected "
1756                                                "(%d sectors at %llu on %s)\n",
1757                                                mdname(mddev), s,
1758                                                (unsigned long long)(sect +
1759                                                    rdev->data_offset),
1760                                                bdevname(rdev->bdev, b));
1761                                 }
1762                         }
1763                 }
1764                 sectors -= s;
1765                 sect += s;
1766         }
1767 }
1768
1769 static void bi_complete(struct bio *bio, int error)
1770 {
1771         complete((struct completion *)bio->bi_private);
1772 }
1773
1774 static int submit_bio_wait(int rw, struct bio *bio)
1775 {
1776         struct completion event;
1777         rw |= REQ_SYNC;
1778
1779         init_completion(&event);
1780         bio->bi_private = &event;
1781         bio->bi_end_io = bi_complete;
1782         submit_bio(rw, bio);
1783         wait_for_completion(&event);
1784
1785         return test_bit(BIO_UPTODATE, &bio->bi_flags);
1786 }
1787
1788 static int narrow_write_error(r1bio_t *r1_bio, int i)
1789 {
1790         mddev_t *mddev = r1_bio->mddev;
1791         conf_t *conf = mddev->private;
1792         mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1793         int vcnt, idx;
1794         struct bio_vec *vec;
1795
1796         /* bio has the data to be written to device 'i' where
1797          * we just recently had a write error.
1798          * We repeatedly clone the bio and trim down to one block,
1799          * then try the write.  Where the write fails we record
1800          * a bad block.
1801          * It is conceivable that the bio doesn't exactly align with
1802          * blocks.  We must handle this somehow.
1803          *
1804          * We currently own a reference on the rdev.
1805          */
1806
1807         int block_sectors;
1808         sector_t sector;
1809         int sectors;
1810         int sect_to_write = r1_bio->sectors;
1811         int ok = 1;
1812
1813         if (rdev->badblocks.shift < 0)
1814                 return 0;
1815
1816         block_sectors = 1 << rdev->badblocks.shift;
1817         sector = r1_bio->sector;
1818         sectors = ((sector + block_sectors)
1819                    & ~(sector_t)(block_sectors - 1))
1820                 - sector;
1821
1822         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
1823                 vcnt = r1_bio->behind_page_count;
1824                 vec = r1_bio->behind_bvecs;
1825                 idx = 0;
1826                 while (vec[idx].bv_page == NULL)
1827                         idx++;
1828         } else {
1829                 vcnt = r1_bio->master_bio->bi_vcnt;
1830                 vec = r1_bio->master_bio->bi_io_vec;
1831                 idx = r1_bio->master_bio->bi_idx;
1832         }
1833         while (sect_to_write) {
1834                 struct bio *wbio;
1835                 if (sectors > sect_to_write)
1836                         sectors = sect_to_write;
1837                 /* Write at 'sector' for 'sectors'*/
1838
1839                 wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
1840                 memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
1841                 wbio->bi_sector = r1_bio->sector;
1842                 wbio->bi_rw = WRITE;
1843                 wbio->bi_vcnt = vcnt;
1844                 wbio->bi_size = r1_bio->sectors << 9;
1845                 wbio->bi_idx = idx;
1846
1847                 md_trim_bio(wbio, sector - r1_bio->sector, sectors);
1848                 wbio->bi_sector += rdev->data_offset;
1849                 wbio->bi_bdev = rdev->bdev;
1850                 if (submit_bio_wait(WRITE, wbio) == 0)
1851                         /* failure! */
1852                         ok = rdev_set_badblocks(rdev, sector,
1853                                                 sectors, 0)
1854                                 && ok;
1855
1856                 bio_put(wbio);
1857                 sect_to_write -= sectors;
1858                 sector += sectors;
1859                 sectors = block_sectors;
1860         }
1861         return ok;
1862 }
1863
1864 static void handle_sync_write_finished(conf_t *conf, r1bio_t *r1_bio)
1865 {
1866         int m;
1867         int s = r1_bio->sectors;
1868         for (m = 0; m < conf->raid_disks ; m++) {
1869                 mdk_rdev_t *rdev = conf->mirrors[m].rdev;
1870                 struct bio *bio = r1_bio->bios[m];
1871                 if (bio->bi_end_io == NULL)
1872                         continue;
1873                 if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
1874                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
1875                         rdev_clear_badblocks(rdev, r1_bio->sector, s);
1876                 }
1877                 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
1878                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
1879                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
1880                                 md_error(conf->mddev, rdev);
1881                 }
1882         }
1883         put_buf(r1_bio);
1884         md_done_sync(conf->mddev, s, 1);
1885 }
1886
1887 static void handle_write_finished(conf_t *conf, r1bio_t *r1_bio)
1888 {
1889         int m;
1890         for (m = 0; m < conf->raid_disks ; m++)
1891                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
1892                         mdk_rdev_t *rdev = conf->mirrors[m].rdev;
1893                         rdev_clear_badblocks(rdev,
1894                                              r1_bio->sector,
1895                                              r1_bio->sectors);
1896                         rdev_dec_pending(rdev, conf->mddev);
1897                 } else if (r1_bio->bios[m] != NULL) {
1898                         /* This drive got a write error.  We need to
1899                          * narrow down and record precise write
1900                          * errors.
1901                          */
1902                         if (!narrow_write_error(r1_bio, m)) {
1903                                 md_error(conf->mddev,
1904                                          conf->mirrors[m].rdev);
1905                                 /* an I/O failed, we can't clear the bitmap */
1906                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1907                         }
1908                         rdev_dec_pending(conf->mirrors[m].rdev,
1909                                          conf->mddev);
1910                 }
1911         if (test_bit(R1BIO_WriteError, &r1_bio->state))
1912                 close_write(r1_bio);
1913         raid_end_bio_io(r1_bio);
1914 }
1915
1916 static void handle_read_error(conf_t *conf, r1bio_t *r1_bio)
1917 {
1918         int disk;
1919         int max_sectors;
1920         mddev_t *mddev = conf->mddev;
1921         struct bio *bio;
1922         char b[BDEVNAME_SIZE];
1923         mdk_rdev_t *rdev;
1924
1925         clear_bit(R1BIO_ReadError, &r1_bio->state);
1926         /* we got a read error. Maybe the drive is bad.  Maybe just
1927          * the block and we can fix it.
1928          * We freeze all other IO, and try reading the block from
1929          * other devices.  When we find one, we re-write
1930          * and check it that fixes the read error.
1931          * This is all done synchronously while the array is
1932          * frozen
1933          */
1934         if (mddev->ro == 0) {
1935                 freeze_array(conf);
1936                 fix_read_error(conf, r1_bio->read_disk,
1937                                r1_bio->sector, r1_bio->sectors);
1938                 unfreeze_array(conf);
1939         } else
1940                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1941
1942         bio = r1_bio->bios[r1_bio->read_disk];
1943         bdevname(bio->bi_bdev, b);
1944 read_more:
1945         disk = read_balance(conf, r1_bio, &max_sectors);
1946         if (disk == -1) {
1947                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
1948                        " read error for block %llu\n",
1949                        mdname(mddev), b, (unsigned long long)r1_bio->sector);
1950                 raid_end_bio_io(r1_bio);
1951         } else {
1952                 const unsigned long do_sync
1953                         = r1_bio->master_bio->bi_rw & REQ_SYNC;
1954                 if (bio) {
1955                         r1_bio->bios[r1_bio->read_disk] =
1956                                 mddev->ro ? IO_BLOCKED : NULL;
1957                         bio_put(bio);
1958                 }
1959                 r1_bio->read_disk = disk;
1960                 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
1961                 md_trim_bio(bio, r1_bio->sector - bio->bi_sector, max_sectors);
1962                 r1_bio->bios[r1_bio->read_disk] = bio;
1963                 rdev = conf->mirrors[disk].rdev;
1964                 printk_ratelimited(KERN_ERR
1965                                    "md/raid1:%s: redirecting sector %llu"
1966                                    " to other mirror: %s\n",
1967                                    mdname(mddev),
1968                                    (unsigned long long)r1_bio->sector,
1969                                    bdevname(rdev->bdev, b));
1970                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1971                 bio->bi_bdev = rdev->bdev;
1972                 bio->bi_end_io = raid1_end_read_request;
1973                 bio->bi_rw = READ | do_sync;
1974                 bio->bi_private = r1_bio;
1975                 if (max_sectors < r1_bio->sectors) {
1976                         /* Drat - have to split this up more */
1977                         struct bio *mbio = r1_bio->master_bio;
1978                         int sectors_handled = (r1_bio->sector + max_sectors
1979                                                - mbio->bi_sector);
1980                         r1_bio->sectors = max_sectors;
1981                         spin_lock_irq(&conf->device_lock);
1982                         if (mbio->bi_phys_segments == 0)
1983                                 mbio->bi_phys_segments = 2;
1984                         else
1985                                 mbio->bi_phys_segments++;
1986                         spin_unlock_irq(&conf->device_lock);
1987                         generic_make_request(bio);
1988                         bio = NULL;
1989
1990                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1991
1992                         r1_bio->master_bio = mbio;
1993                         r1_bio->sectors = (mbio->bi_size >> 9)
1994                                           - sectors_handled;
1995                         r1_bio->state = 0;
1996                         set_bit(R1BIO_ReadError, &r1_bio->state);
1997                         r1_bio->mddev = mddev;
1998                         r1_bio->sector = mbio->bi_sector + sectors_handled;
1999
2000                         goto read_more;
2001                 } else
2002                         generic_make_request(bio);
2003         }
2004 }
2005
2006 static void raid1d(mddev_t *mddev)
2007 {
2008         r1bio_t *r1_bio;
2009         unsigned long flags;
2010         conf_t *conf = mddev->private;
2011         struct list_head *head = &conf->retry_list;
2012         struct blk_plug plug;
2013
2014         md_check_recovery(mddev);
2015
2016         blk_start_plug(&plug);
2017         for (;;) {
2018
2019                 if (atomic_read(&mddev->plug_cnt) == 0)
2020                         flush_pending_writes(conf);
2021
2022                 spin_lock_irqsave(&conf->device_lock, flags);
2023                 if (list_empty(head)) {
2024                         spin_unlock_irqrestore(&conf->device_lock, flags);
2025                         break;
2026                 }
2027                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
2028                 list_del(head->prev);
2029                 conf->nr_queued--;
2030                 spin_unlock_irqrestore(&conf->device_lock, flags);
2031
2032                 mddev = r1_bio->mddev;
2033                 conf = mddev->private;
2034                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2035                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2036                             test_bit(R1BIO_WriteError, &r1_bio->state))
2037                                 handle_sync_write_finished(conf, r1_bio);
2038                         else
2039                                 sync_request_write(mddev, r1_bio);
2040                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2041                            test_bit(R1BIO_WriteError, &r1_bio->state))
2042                         handle_write_finished(conf, r1_bio);
2043                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2044                         handle_read_error(conf, r1_bio);
2045                 else
2046                         /* just a partial read to be scheduled from separate
2047                          * context
2048                          */
2049                         generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2050
2051                 cond_resched();
2052                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2053                         md_check_recovery(mddev);
2054         }
2055         blk_finish_plug(&plug);
2056 }
2057
2058
2059 static int init_resync(conf_t *conf)
2060 {
2061         int buffs;
2062
2063         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2064         BUG_ON(conf->r1buf_pool);
2065         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2066                                           conf->poolinfo);
2067         if (!conf->r1buf_pool)
2068                 return -ENOMEM;
2069         conf->next_resync = 0;
2070         return 0;
2071 }
2072
2073 /*
2074  * perform a "sync" on one "block"
2075  *
2076  * We need to make sure that no normal I/O request - particularly write
2077  * requests - conflict with active sync requests.
2078  *
2079  * This is achieved by tracking pending requests and a 'barrier' concept
2080  * that can be installed to exclude normal IO requests.
2081  */
2082
2083 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
2084 {
2085         conf_t *conf = mddev->private;
2086         r1bio_t *r1_bio;
2087         struct bio *bio;
2088         sector_t max_sector, nr_sectors;
2089         int disk = -1;
2090         int i;
2091         int wonly = -1;
2092         int write_targets = 0, read_targets = 0;
2093         sector_t sync_blocks;
2094         int still_degraded = 0;
2095         int good_sectors = RESYNC_SECTORS;
2096         int min_bad = 0; /* number of sectors that are bad in all devices */
2097
2098         if (!conf->r1buf_pool)
2099                 if (init_resync(conf))
2100                         return 0;
2101
2102         max_sector = mddev->dev_sectors;
2103         if (sector_nr >= max_sector) {
2104                 /* If we aborted, we need to abort the
2105                  * sync on the 'current' bitmap chunk (there will
2106                  * only be one in raid1 resync.
2107                  * We can find the current addess in mddev->curr_resync
2108                  */
2109                 if (mddev->curr_resync < max_sector) /* aborted */
2110                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2111                                                 &sync_blocks, 1);
2112                 else /* completed sync */
2113                         conf->fullsync = 0;
2114
2115                 bitmap_close_sync(mddev->bitmap);
2116                 close_sync(conf);
2117                 return 0;
2118         }
2119
2120         if (mddev->bitmap == NULL &&
2121             mddev->recovery_cp == MaxSector &&
2122             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2123             conf->fullsync == 0) {
2124                 *skipped = 1;
2125                 return max_sector - sector_nr;
2126         }
2127         /* before building a request, check if we can skip these blocks..
2128          * This call the bitmap_start_sync doesn't actually record anything
2129          */
2130         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2131             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2132                 /* We can skip this block, and probably several more */
2133                 *skipped = 1;
2134                 return sync_blocks;
2135         }
2136         /*
2137          * If there is non-resync activity waiting for a turn,
2138          * and resync is going fast enough,
2139          * then let it though before starting on this new sync request.
2140          */
2141         if (!go_faster && conf->nr_waiting)
2142                 msleep_interruptible(1000);
2143
2144         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2145         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2146         raise_barrier(conf);
2147
2148         conf->next_resync = sector_nr;
2149
2150         rcu_read_lock();
2151         /*
2152          * If we get a correctably read error during resync or recovery,
2153          * we might want to read from a different device.  So we
2154          * flag all drives that could conceivably be read from for READ,
2155          * and any others (which will be non-In_sync devices) for WRITE.
2156          * If a read fails, we try reading from something else for which READ
2157          * is OK.
2158          */
2159
2160         r1_bio->mddev = mddev;
2161         r1_bio->sector = sector_nr;
2162         r1_bio->state = 0;
2163         set_bit(R1BIO_IsSync, &r1_bio->state);
2164
2165         for (i=0; i < conf->raid_disks; i++) {
2166                 mdk_rdev_t *rdev;
2167                 bio = r1_bio->bios[i];
2168
2169                 /* take from bio_init */
2170                 bio->bi_next = NULL;
2171                 bio->bi_flags &= ~(BIO_POOL_MASK-1);
2172                 bio->bi_flags |= 1 << BIO_UPTODATE;
2173                 bio->bi_comp_cpu = -1;
2174                 bio->bi_rw = READ;
2175                 bio->bi_vcnt = 0;
2176                 bio->bi_idx = 0;
2177                 bio->bi_phys_segments = 0;
2178                 bio->bi_size = 0;
2179                 bio->bi_end_io = NULL;
2180                 bio->bi_private = NULL;
2181
2182                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2183                 if (rdev == NULL ||
2184                     test_bit(Faulty, &rdev->flags)) {
2185                         still_degraded = 1;
2186                 } else if (!test_bit(In_sync, &rdev->flags)) {
2187                         bio->bi_rw = WRITE;
2188                         bio->bi_end_io = end_sync_write;
2189                         write_targets ++;
2190                 } else {
2191                         /* may need to read from here */
2192                         sector_t first_bad = MaxSector;
2193                         int bad_sectors;
2194
2195                         if (is_badblock(rdev, sector_nr, good_sectors,
2196                                         &first_bad, &bad_sectors)) {
2197                                 if (first_bad > sector_nr)
2198                                         good_sectors = first_bad - sector_nr;
2199                                 else {
2200                                         bad_sectors -= (sector_nr - first_bad);
2201                                         if (min_bad == 0 ||
2202                                             min_bad > bad_sectors)
2203                                                 min_bad = bad_sectors;
2204                                 }
2205                         }
2206                         if (sector_nr < first_bad) {
2207                                 if (test_bit(WriteMostly, &rdev->flags)) {
2208                                         if (wonly < 0)
2209                                                 wonly = i;
2210                                 } else {
2211                                         if (disk < 0)
2212                                                 disk = i;
2213                                 }
2214                                 bio->bi_rw = READ;
2215                                 bio->bi_end_io = end_sync_read;
2216                                 read_targets++;
2217                         }
2218                 }
2219                 if (bio->bi_end_io) {
2220                         atomic_inc(&rdev->nr_pending);
2221                         bio->bi_sector = sector_nr + rdev->data_offset;
2222                         bio->bi_bdev = rdev->bdev;
2223                         bio->bi_private = r1_bio;
2224                 }
2225         }
2226         rcu_read_unlock();
2227         if (disk < 0)
2228                 disk = wonly;
2229         r1_bio->read_disk = disk;
2230
2231         if (read_targets == 0 && min_bad > 0) {
2232                 /* These sectors are bad on all InSync devices, so we
2233                  * need to mark them bad on all write targets
2234                  */
2235                 int ok = 1;
2236                 for (i = 0 ; i < conf->raid_disks ; i++)
2237                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2238                                 mdk_rdev_t *rdev =
2239                                         rcu_dereference(conf->mirrors[i].rdev);
2240                                 ok = rdev_set_badblocks(rdev, sector_nr,
2241                                                         min_bad, 0
2242                                         ) && ok;
2243                         }
2244                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2245                 *skipped = 1;
2246                 put_buf(r1_bio);
2247
2248                 if (!ok) {
2249                         /* Cannot record the badblocks, so need to
2250                          * abort the resync.
2251                          * If there are multiple read targets, could just
2252                          * fail the really bad ones ???
2253                          */
2254                         conf->recovery_disabled = mddev->recovery_disabled;
2255                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2256                         return 0;
2257                 } else
2258                         return min_bad;
2259
2260         }
2261         if (min_bad > 0 && min_bad < good_sectors) {
2262                 /* only resync enough to reach the next bad->good
2263                  * transition */
2264                 good_sectors = min_bad;
2265         }
2266
2267         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2268                 /* extra read targets are also write targets */
2269                 write_targets += read_targets-1;
2270
2271         if (write_targets == 0 || read_targets == 0) {
2272                 /* There is nowhere to write, so all non-sync
2273                  * drives must be failed - so we are finished
2274                  */
2275                 sector_t rv = max_sector - sector_nr;
2276                 *skipped = 1;
2277                 put_buf(r1_bio);
2278                 return rv;
2279         }
2280
2281         if (max_sector > mddev->resync_max)
2282                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2283         if (max_sector > sector_nr + good_sectors)
2284                 max_sector = sector_nr + good_sectors;
2285         nr_sectors = 0;
2286         sync_blocks = 0;
2287         do {
2288                 struct page *page;
2289                 int len = PAGE_SIZE;
2290                 if (sector_nr + (len>>9) > max_sector)
2291                         len = (max_sector - sector_nr) << 9;
2292                 if (len == 0)
2293                         break;
2294                 if (sync_blocks == 0) {
2295                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2296                                                &sync_blocks, still_degraded) &&
2297                             !conf->fullsync &&
2298                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2299                                 break;
2300                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2301                         if ((len >> 9) > sync_blocks)
2302                                 len = sync_blocks<<9;
2303                 }
2304
2305                 for (i=0 ; i < conf->raid_disks; i++) {
2306                         bio = r1_bio->bios[i];
2307                         if (bio->bi_end_io) {
2308                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2309                                 if (bio_add_page(bio, page, len, 0) == 0) {
2310                                         /* stop here */
2311                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2312                                         while (i > 0) {
2313                                                 i--;
2314                                                 bio = r1_bio->bios[i];
2315                                                 if (bio->bi_end_io==NULL)
2316                                                         continue;
2317                                                 /* remove last page from this bio */
2318                                                 bio->bi_vcnt--;
2319                                                 bio->bi_size -= len;
2320                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2321                                         }
2322                                         goto bio_full;
2323                                 }
2324                         }
2325                 }
2326                 nr_sectors += len>>9;
2327                 sector_nr += len>>9;
2328                 sync_blocks -= (len>>9);
2329         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2330  bio_full:
2331         r1_bio->sectors = nr_sectors;
2332
2333         /* For a user-requested sync, we read all readable devices and do a
2334          * compare
2335          */
2336         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2337                 atomic_set(&r1_bio->remaining, read_targets);
2338                 for (i=0; i<conf->raid_disks; i++) {
2339                         bio = r1_bio->bios[i];
2340                         if (bio->bi_end_io == end_sync_read) {
2341                                 md_sync_acct(bio->bi_bdev, nr_sectors);
2342                                 generic_make_request(bio);
2343                         }
2344                 }
2345         } else {
2346                 atomic_set(&r1_bio->remaining, 1);
2347                 bio = r1_bio->bios[r1_bio->read_disk];
2348                 md_sync_acct(bio->bi_bdev, nr_sectors);
2349                 generic_make_request(bio);
2350
2351         }
2352         return nr_sectors;
2353 }
2354
2355 static sector_t raid1_size(mddev_t *mddev, sector_t sectors, int raid_disks)
2356 {
2357         if (sectors)
2358                 return sectors;
2359
2360         return mddev->dev_sectors;
2361 }
2362
2363 static conf_t *setup_conf(mddev_t *mddev)
2364 {
2365         conf_t *conf;
2366         int i;
2367         mirror_info_t *disk;
2368         mdk_rdev_t *rdev;
2369         int err = -ENOMEM;
2370
2371         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2372         if (!conf)
2373                 goto abort;
2374
2375         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2376                                  GFP_KERNEL);
2377         if (!conf->mirrors)
2378                 goto abort;
2379
2380         conf->tmppage = alloc_page(GFP_KERNEL);
2381         if (!conf->tmppage)
2382                 goto abort;
2383
2384         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2385         if (!conf->poolinfo)
2386                 goto abort;
2387         conf->poolinfo->raid_disks = mddev->raid_disks;
2388         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2389                                           r1bio_pool_free,
2390                                           conf->poolinfo);
2391         if (!conf->r1bio_pool)
2392                 goto abort;
2393
2394         conf->poolinfo->mddev = mddev;
2395
2396         spin_lock_init(&conf->device_lock);
2397         list_for_each_entry(rdev, &mddev->disks, same_set) {
2398                 int disk_idx = rdev->raid_disk;
2399                 if (disk_idx >= mddev->raid_disks
2400                     || disk_idx < 0)
2401                         continue;
2402                 disk = conf->mirrors + disk_idx;
2403
2404                 disk->rdev = rdev;
2405
2406                 disk->head_position = 0;
2407         }
2408         conf->raid_disks = mddev->raid_disks;
2409         conf->mddev = mddev;
2410         INIT_LIST_HEAD(&conf->retry_list);
2411
2412         spin_lock_init(&conf->resync_lock);
2413         init_waitqueue_head(&conf->wait_barrier);
2414
2415         bio_list_init(&conf->pending_bio_list);
2416
2417         conf->last_used = -1;
2418         for (i = 0; i < conf->raid_disks; i++) {
2419
2420                 disk = conf->mirrors + i;
2421
2422                 if (!disk->rdev ||
2423                     !test_bit(In_sync, &disk->rdev->flags)) {
2424                         disk->head_position = 0;
2425                         if (disk->rdev)
2426                                 conf->fullsync = 1;
2427                 } else if (conf->last_used < 0)
2428                         /*
2429                          * The first working device is used as a
2430                          * starting point to read balancing.
2431                          */
2432                         conf->last_used = i;
2433         }
2434
2435         err = -EIO;
2436         if (conf->last_used < 0) {
2437                 printk(KERN_ERR "md/raid1:%s: no operational mirrors\n",
2438                        mdname(mddev));
2439                 goto abort;
2440         }
2441         err = -ENOMEM;
2442         conf->thread = md_register_thread(raid1d, mddev, NULL);
2443         if (!conf->thread) {
2444                 printk(KERN_ERR
2445                        "md/raid1:%s: couldn't allocate thread\n",
2446                        mdname(mddev));
2447                 goto abort;
2448         }
2449
2450         return conf;
2451
2452  abort:
2453         if (conf) {
2454                 if (conf->r1bio_pool)
2455                         mempool_destroy(conf->r1bio_pool);
2456                 kfree(conf->mirrors);
2457                 safe_put_page(conf->tmppage);
2458                 kfree(conf->poolinfo);
2459                 kfree(conf);
2460         }
2461         return ERR_PTR(err);
2462 }
2463
2464 static int run(mddev_t *mddev)
2465 {
2466         conf_t *conf;
2467         int i;
2468         mdk_rdev_t *rdev;
2469
2470         if (mddev->level != 1) {
2471                 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2472                        mdname(mddev), mddev->level);
2473                 return -EIO;
2474         }
2475         if (mddev->reshape_position != MaxSector) {
2476                 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2477                        mdname(mddev));
2478                 return -EIO;
2479         }
2480         /*
2481          * copy the already verified devices into our private RAID1
2482          * bookkeeping area. [whatever we allocate in run(),
2483          * should be freed in stop()]
2484          */
2485         if (mddev->private == NULL)
2486                 conf = setup_conf(mddev);
2487         else
2488                 conf = mddev->private;
2489
2490         if (IS_ERR(conf))
2491                 return PTR_ERR(conf);
2492
2493         list_for_each_entry(rdev, &mddev->disks, same_set) {
2494                 if (!mddev->gendisk)
2495                         continue;
2496                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2497                                   rdev->data_offset << 9);
2498                 /* as we don't honour merge_bvec_fn, we must never risk
2499                  * violating it, so limit ->max_segments to 1 lying within
2500                  * a single page, as a one page request is never in violation.
2501                  */
2502                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2503                         blk_queue_max_segments(mddev->queue, 1);
2504                         blk_queue_segment_boundary(mddev->queue,
2505                                                    PAGE_CACHE_SIZE - 1);
2506                 }
2507         }
2508
2509         mddev->degraded = 0;
2510         for (i=0; i < conf->raid_disks; i++)
2511                 if (conf->mirrors[i].rdev == NULL ||
2512                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2513                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2514                         mddev->degraded++;
2515
2516         if (conf->raid_disks - mddev->degraded == 1)
2517                 mddev->recovery_cp = MaxSector;
2518
2519         if (mddev->recovery_cp != MaxSector)
2520                 printk(KERN_NOTICE "md/raid1:%s: not clean"
2521                        " -- starting background reconstruction\n",
2522                        mdname(mddev));
2523         printk(KERN_INFO 
2524                 "md/raid1:%s: active with %d out of %d mirrors\n",
2525                 mdname(mddev), mddev->raid_disks - mddev->degraded, 
2526                 mddev->raid_disks);
2527
2528         /*
2529          * Ok, everything is just fine now
2530          */
2531         mddev->thread = conf->thread;
2532         conf->thread = NULL;
2533         mddev->private = conf;
2534
2535         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2536
2537         if (mddev->queue) {
2538                 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2539                 mddev->queue->backing_dev_info.congested_data = mddev;
2540         }
2541         return md_integrity_register(mddev);
2542 }
2543
2544 static int stop(mddev_t *mddev)
2545 {
2546         conf_t *conf = mddev->private;
2547         struct bitmap *bitmap = mddev->bitmap;
2548
2549         /* wait for behind writes to complete */
2550         if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2551                 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2552                        mdname(mddev));
2553                 /* need to kick something here to make sure I/O goes? */
2554                 wait_event(bitmap->behind_wait,
2555                            atomic_read(&bitmap->behind_writes) == 0);
2556         }
2557
2558         raise_barrier(conf);
2559         lower_barrier(conf);
2560
2561         md_unregister_thread(mddev->thread);
2562         mddev->thread = NULL;
2563         if (conf->r1bio_pool)
2564                 mempool_destroy(conf->r1bio_pool);
2565         kfree(conf->mirrors);
2566         kfree(conf->poolinfo);
2567         kfree(conf);
2568         mddev->private = NULL;
2569         return 0;
2570 }
2571
2572 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2573 {
2574         /* no resync is happening, and there is enough space
2575          * on all devices, so we can resize.
2576          * We need to make sure resync covers any new space.
2577          * If the array is shrinking we should possibly wait until
2578          * any io in the removed space completes, but it hardly seems
2579          * worth it.
2580          */
2581         md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2582         if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2583                 return -EINVAL;
2584         set_capacity(mddev->gendisk, mddev->array_sectors);
2585         revalidate_disk(mddev->gendisk);
2586         if (sectors > mddev->dev_sectors &&
2587             mddev->recovery_cp > mddev->dev_sectors) {
2588                 mddev->recovery_cp = mddev->dev_sectors;
2589                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2590         }
2591         mddev->dev_sectors = sectors;
2592         mddev->resync_max_sectors = sectors;
2593         return 0;
2594 }
2595
2596 static int raid1_reshape(mddev_t *mddev)
2597 {
2598         /* We need to:
2599          * 1/ resize the r1bio_pool
2600          * 2/ resize conf->mirrors
2601          *
2602          * We allocate a new r1bio_pool if we can.
2603          * Then raise a device barrier and wait until all IO stops.
2604          * Then resize conf->mirrors and swap in the new r1bio pool.
2605          *
2606          * At the same time, we "pack" the devices so that all the missing
2607          * devices have the higher raid_disk numbers.
2608          */
2609         mempool_t *newpool, *oldpool;
2610         struct pool_info *newpoolinfo;
2611         mirror_info_t *newmirrors;
2612         conf_t *conf = mddev->private;
2613         int cnt, raid_disks;
2614         unsigned long flags;
2615         int d, d2, err;
2616
2617         /* Cannot change chunk_size, layout, or level */
2618         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2619             mddev->layout != mddev->new_layout ||
2620             mddev->level != mddev->new_level) {
2621                 mddev->new_chunk_sectors = mddev->chunk_sectors;
2622                 mddev->new_layout = mddev->layout;
2623                 mddev->new_level = mddev->level;
2624                 return -EINVAL;
2625         }
2626
2627         err = md_allow_write(mddev);
2628         if (err)
2629                 return err;
2630
2631         raid_disks = mddev->raid_disks + mddev->delta_disks;
2632
2633         if (raid_disks < conf->raid_disks) {
2634                 cnt=0;
2635                 for (d= 0; d < conf->raid_disks; d++)
2636                         if (conf->mirrors[d].rdev)
2637                                 cnt++;
2638                 if (cnt > raid_disks)
2639                         return -EBUSY;
2640         }
2641
2642         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2643         if (!newpoolinfo)
2644                 return -ENOMEM;
2645         newpoolinfo->mddev = mddev;
2646         newpoolinfo->raid_disks = raid_disks;
2647
2648         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2649                                  r1bio_pool_free, newpoolinfo);
2650         if (!newpool) {
2651                 kfree(newpoolinfo);
2652                 return -ENOMEM;
2653         }
2654         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2655         if (!newmirrors) {
2656                 kfree(newpoolinfo);
2657                 mempool_destroy(newpool);
2658                 return -ENOMEM;
2659         }
2660
2661         raise_barrier(conf);
2662
2663         /* ok, everything is stopped */
2664         oldpool = conf->r1bio_pool;
2665         conf->r1bio_pool = newpool;
2666
2667         for (d = d2 = 0; d < conf->raid_disks; d++) {
2668                 mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2669                 if (rdev && rdev->raid_disk != d2) {
2670                         sysfs_unlink_rdev(mddev, rdev);
2671                         rdev->raid_disk = d2;
2672                         sysfs_unlink_rdev(mddev, rdev);
2673                         if (sysfs_link_rdev(mddev, rdev))
2674                                 printk(KERN_WARNING
2675                                        "md/raid1:%s: cannot register rd%d\n",
2676                                        mdname(mddev), rdev->raid_disk);
2677                 }
2678                 if (rdev)
2679                         newmirrors[d2++].rdev = rdev;
2680         }
2681         kfree(conf->mirrors);
2682         conf->mirrors = newmirrors;
2683         kfree(conf->poolinfo);
2684         conf->poolinfo = newpoolinfo;
2685
2686         spin_lock_irqsave(&conf->device_lock, flags);
2687         mddev->degraded += (raid_disks - conf->raid_disks);
2688         spin_unlock_irqrestore(&conf->device_lock, flags);
2689         conf->raid_disks = mddev->raid_disks = raid_disks;
2690         mddev->delta_disks = 0;
2691
2692         conf->last_used = 0; /* just make sure it is in-range */
2693         lower_barrier(conf);
2694
2695         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2696         md_wakeup_thread(mddev->thread);
2697
2698         mempool_destroy(oldpool);
2699         return 0;
2700 }
2701
2702 static void raid1_quiesce(mddev_t *mddev, int state)
2703 {
2704         conf_t *conf = mddev->private;
2705
2706         switch(state) {
2707         case 2: /* wake for suspend */
2708                 wake_up(&conf->wait_barrier);
2709                 break;
2710         case 1:
2711                 raise_barrier(conf);
2712                 break;
2713         case 0:
2714                 lower_barrier(conf);
2715                 break;
2716         }
2717 }
2718
2719 static void *raid1_takeover(mddev_t *mddev)
2720 {
2721         /* raid1 can take over:
2722          *  raid5 with 2 devices, any layout or chunk size
2723          */
2724         if (mddev->level == 5 && mddev->raid_disks == 2) {
2725                 conf_t *conf;
2726                 mddev->new_level = 1;
2727                 mddev->new_layout = 0;
2728                 mddev->new_chunk_sectors = 0;
2729                 conf = setup_conf(mddev);
2730                 if (!IS_ERR(conf))
2731                         conf->barrier = 1;
2732                 return conf;
2733         }
2734         return ERR_PTR(-EINVAL);
2735 }
2736
2737 static struct mdk_personality raid1_personality =
2738 {
2739         .name           = "raid1",
2740         .level          = 1,
2741         .owner          = THIS_MODULE,
2742         .make_request   = make_request,
2743         .run            = run,
2744         .stop           = stop,
2745         .status         = status,
2746         .error_handler  = error,
2747         .hot_add_disk   = raid1_add_disk,
2748         .hot_remove_disk= raid1_remove_disk,
2749         .spare_active   = raid1_spare_active,
2750         .sync_request   = sync_request,
2751         .resize         = raid1_resize,
2752         .size           = raid1_size,
2753         .check_reshape  = raid1_reshape,
2754         .quiesce        = raid1_quiesce,
2755         .takeover       = raid1_takeover,
2756 };
2757
2758 static int __init raid_init(void)
2759 {
2760         return register_md_personality(&raid1_personality);
2761 }
2762
2763 static void raid_exit(void)
2764 {
2765         unregister_md_personality(&raid1_personality);
2766 }
2767
2768 module_init(raid_init);
2769 module_exit(raid_exit);
2770 MODULE_LICENSE("GPL");
2771 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2772 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2773 MODULE_ALIAS("md-raid1");
2774 MODULE_ALIAS("md-level-1");