Merge branch 'upstream-fixes'
[linux-2.6] / 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 Ã˜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 "dm-bio-list.h"
35 #include <linux/raid/raid1.h>
36 #include <linux/raid/bitmap.h>
37
38 #define DEBUG 0
39 #if DEBUG
40 #define PRINTK(x...) printk(x)
41 #else
42 #define PRINTK(x...)
43 #endif
44
45 /*
46  * Number of guaranteed r1bios in case of extreme VM load:
47  */
48 #define NR_RAID1_BIOS 256
49
50
51 static void unplug_slaves(mddev_t *mddev);
52
53 static void allow_barrier(conf_t *conf);
54 static void lower_barrier(conf_t *conf);
55
56 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
57 {
58         struct pool_info *pi = data;
59         r1bio_t *r1_bio;
60         int size = offsetof(r1bio_t, bios[pi->raid_disks]);
61
62         /* allocate a r1bio with room for raid_disks entries in the bios array */
63         r1_bio = kzalloc(size, gfp_flags);
64         if (!r1_bio)
65                 unplug_slaves(pi->mddev);
66
67         return r1_bio;
68 }
69
70 static void r1bio_pool_free(void *r1_bio, void *data)
71 {
72         kfree(r1_bio);
73 }
74
75 #define RESYNC_BLOCK_SIZE (64*1024)
76 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
77 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
78 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
79 #define RESYNC_WINDOW (2048*1024)
80
81 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
82 {
83         struct pool_info *pi = data;
84         struct page *page;
85         r1bio_t *r1_bio;
86         struct bio *bio;
87         int i, j;
88
89         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
90         if (!r1_bio) {
91                 unplug_slaves(pi->mddev);
92                 return NULL;
93         }
94
95         /*
96          * Allocate bios : 1 for reading, n-1 for writing
97          */
98         for (j = pi->raid_disks ; j-- ; ) {
99                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
100                 if (!bio)
101                         goto out_free_bio;
102                 r1_bio->bios[j] = bio;
103         }
104         /*
105          * Allocate RESYNC_PAGES data pages and attach them to
106          * the first bio.
107          * If this is a user-requested check/repair, allocate
108          * RESYNC_PAGES for each bio.
109          */
110         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
111                 j = pi->raid_disks;
112         else
113                 j = 1;
114         while(j--) {
115                 bio = r1_bio->bios[j];
116                 for (i = 0; i < RESYNC_PAGES; i++) {
117                         page = alloc_page(gfp_flags);
118                         if (unlikely(!page))
119                                 goto out_free_pages;
120
121                         bio->bi_io_vec[i].bv_page = page;
122                 }
123         }
124         /* If not user-requests, copy the page pointers to all bios */
125         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
126                 for (i=0; i<RESYNC_PAGES ; i++)
127                         for (j=1; j<pi->raid_disks; j++)
128                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
129                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
130         }
131
132         r1_bio->master_bio = NULL;
133
134         return r1_bio;
135
136 out_free_pages:
137         for (i=0; i < RESYNC_PAGES ; i++)
138                 for (j=0 ; j < pi->raid_disks; j++)
139                         safe_put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
140         j = -1;
141 out_free_bio:
142         while ( ++j < pi->raid_disks )
143                 bio_put(r1_bio->bios[j]);
144         r1bio_pool_free(r1_bio, data);
145         return NULL;
146 }
147
148 static void r1buf_pool_free(void *__r1_bio, void *data)
149 {
150         struct pool_info *pi = data;
151         int i,j;
152         r1bio_t *r1bio = __r1_bio;
153
154         for (i = 0; i < RESYNC_PAGES; i++)
155                 for (j = pi->raid_disks; j-- ;) {
156                         if (j == 0 ||
157                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
158                             r1bio->bios[0]->bi_io_vec[i].bv_page)
159                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
160                 }
161         for (i=0 ; i < pi->raid_disks; i++)
162                 bio_put(r1bio->bios[i]);
163
164         r1bio_pool_free(r1bio, data);
165 }
166
167 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
168 {
169         int i;
170
171         for (i = 0; i < conf->raid_disks; i++) {
172                 struct bio **bio = r1_bio->bios + i;
173                 if (*bio && *bio != IO_BLOCKED)
174                         bio_put(*bio);
175                 *bio = NULL;
176         }
177 }
178
179 static void free_r1bio(r1bio_t *r1_bio)
180 {
181         conf_t *conf = mddev_to_conf(r1_bio->mddev);
182
183         /*
184          * Wake up any possible resync thread that waits for the device
185          * to go idle.
186          */
187         allow_barrier(conf);
188
189         put_all_bios(conf, r1_bio);
190         mempool_free(r1_bio, conf->r1bio_pool);
191 }
192
193 static void put_buf(r1bio_t *r1_bio)
194 {
195         conf_t *conf = mddev_to_conf(r1_bio->mddev);
196         int i;
197
198         for (i=0; i<conf->raid_disks; i++) {
199                 struct bio *bio = r1_bio->bios[i];
200                 if (bio->bi_end_io)
201                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
202         }
203
204         mempool_free(r1_bio, conf->r1buf_pool);
205
206         lower_barrier(conf);
207 }
208
209 static void reschedule_retry(r1bio_t *r1_bio)
210 {
211         unsigned long flags;
212         mddev_t *mddev = r1_bio->mddev;
213         conf_t *conf = mddev_to_conf(mddev);
214
215         spin_lock_irqsave(&conf->device_lock, flags);
216         list_add(&r1_bio->retry_list, &conf->retry_list);
217         conf->nr_queued ++;
218         spin_unlock_irqrestore(&conf->device_lock, flags);
219
220         wake_up(&conf->wait_barrier);
221         md_wakeup_thread(mddev->thread);
222 }
223
224 /*
225  * raid_end_bio_io() is called when we have finished servicing a mirrored
226  * operation and are ready to return a success/failure code to the buffer
227  * cache layer.
228  */
229 static void raid_end_bio_io(r1bio_t *r1_bio)
230 {
231         struct bio *bio = r1_bio->master_bio;
232
233         /* if nobody has done the final endio yet, do it now */
234         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
235                 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
236                         (bio_data_dir(bio) == WRITE) ? "write" : "read",
237                         (unsigned long long) bio->bi_sector,
238                         (unsigned long long) bio->bi_sector +
239                                 (bio->bi_size >> 9) - 1);
240
241                 bio_endio(bio, bio->bi_size,
242                         test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
243         }
244         free_r1bio(r1_bio);
245 }
246
247 /*
248  * Update disk head position estimator based on IRQ completion info.
249  */
250 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
251 {
252         conf_t *conf = mddev_to_conf(r1_bio->mddev);
253
254         conf->mirrors[disk].head_position =
255                 r1_bio->sector + (r1_bio->sectors);
256 }
257
258 static int raid1_end_read_request(struct bio *bio, unsigned int bytes_done, int error)
259 {
260         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
261         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
262         int mirror;
263         conf_t *conf = mddev_to_conf(r1_bio->mddev);
264
265         if (bio->bi_size)
266                 return 1;
267         
268         mirror = r1_bio->read_disk;
269         /*
270          * this branch is our 'one mirror IO has finished' event handler:
271          */
272         update_head_pos(mirror, r1_bio);
273
274         if (uptodate || conf->working_disks <= 1) {
275                 /*
276                  * Set R1BIO_Uptodate in our master bio, so that
277                  * we will return a good error code for to the higher
278                  * levels even if IO on some other mirrored buffer fails.
279                  *
280                  * The 'master' represents the composite IO operation to
281                  * user-side. So if something waits for IO, then it will
282                  * wait for the 'master' bio.
283                  */
284                 if (uptodate)
285                         set_bit(R1BIO_Uptodate, &r1_bio->state);
286
287                 raid_end_bio_io(r1_bio);
288         } else {
289                 /*
290                  * oops, read error:
291                  */
292                 char b[BDEVNAME_SIZE];
293                 if (printk_ratelimit())
294                         printk(KERN_ERR "raid1: %s: rescheduling sector %llu\n",
295                                bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
296                 reschedule_retry(r1_bio);
297         }
298
299         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
300         return 0;
301 }
302
303 static int raid1_end_write_request(struct bio *bio, unsigned int bytes_done, int error)
304 {
305         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
306         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
307         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
308         conf_t *conf = mddev_to_conf(r1_bio->mddev);
309         struct bio *to_put = NULL;
310
311         if (bio->bi_size)
312                 return 1;
313
314         for (mirror = 0; mirror < conf->raid_disks; mirror++)
315                 if (r1_bio->bios[mirror] == bio)
316                         break;
317
318         if (error == -ENOTSUPP && test_bit(R1BIO_Barrier, &r1_bio->state)) {
319                 set_bit(BarriersNotsupp, &conf->mirrors[mirror].rdev->flags);
320                 set_bit(R1BIO_BarrierRetry, &r1_bio->state);
321                 r1_bio->mddev->barriers_work = 0;
322         } else {
323                 /*
324                  * this branch is our 'one mirror IO has finished' event handler:
325                  */
326                 r1_bio->bios[mirror] = NULL;
327                 to_put = bio;
328                 if (!uptodate) {
329                         md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
330                         /* an I/O failed, we can't clear the bitmap */
331                         set_bit(R1BIO_Degraded, &r1_bio->state);
332                 } else
333                         /*
334                          * Set R1BIO_Uptodate in our master bio, so that
335                          * we will return a good error code for to the higher
336                          * levels even if IO on some other mirrored buffer fails.
337                          *
338                          * The 'master' represents the composite IO operation to
339                          * user-side. So if something waits for IO, then it will
340                          * wait for the 'master' bio.
341                          */
342                         set_bit(R1BIO_Uptodate, &r1_bio->state);
343
344                 update_head_pos(mirror, r1_bio);
345
346                 if (behind) {
347                         if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
348                                 atomic_dec(&r1_bio->behind_remaining);
349
350                         /* In behind mode, we ACK the master bio once the I/O has safely
351                          * reached all non-writemostly disks. Setting the Returned bit
352                          * ensures that this gets done only once -- we don't ever want to
353                          * return -EIO here, instead we'll wait */
354
355                         if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
356                             test_bit(R1BIO_Uptodate, &r1_bio->state)) {
357                                 /* Maybe we can return now */
358                                 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
359                                         struct bio *mbio = r1_bio->master_bio;
360                                         PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
361                                                (unsigned long long) mbio->bi_sector,
362                                                (unsigned long long) mbio->bi_sector +
363                                                (mbio->bi_size >> 9) - 1);
364                                         bio_endio(mbio, mbio->bi_size, 0);
365                                 }
366                         }
367                 }
368         }
369         /*
370          *
371          * Let's see if all mirrored write operations have finished
372          * already.
373          */
374         if (atomic_dec_and_test(&r1_bio->remaining)) {
375                 if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
376                         reschedule_retry(r1_bio);
377                         /* Don't dec_pending yet, we want to hold
378                          * the reference over the retry
379                          */
380                         goto out;
381                 }
382                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
383                         /* free extra copy of the data pages */
384                         int i = bio->bi_vcnt;
385                         while (i--)
386                                 safe_put_page(bio->bi_io_vec[i].bv_page);
387                 }
388                 /* clear the bitmap if all writes complete successfully */
389                 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
390                                 r1_bio->sectors,
391                                 !test_bit(R1BIO_Degraded, &r1_bio->state),
392                                 behind);
393                 md_write_end(r1_bio->mddev);
394                 raid_end_bio_io(r1_bio);
395         }
396
397         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
398  out:
399         if (to_put)
400                 bio_put(to_put);
401
402         return 0;
403 }
404
405
406 /*
407  * This routine returns the disk from which the requested read should
408  * be done. There is a per-array 'next expected sequential IO' sector
409  * number - if this matches on the next IO then we use the last disk.
410  * There is also a per-disk 'last know head position' sector that is
411  * maintained from IRQ contexts, both the normal and the resync IO
412  * completion handlers update this position correctly. If there is no
413  * perfect sequential match then we pick the disk whose head is closest.
414  *
415  * If there are 2 mirrors in the same 2 devices, performance degrades
416  * because position is mirror, not device based.
417  *
418  * The rdev for the device selected will have nr_pending incremented.
419  */
420 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
421 {
422         const unsigned long this_sector = r1_bio->sector;
423         int new_disk = conf->last_used, disk = new_disk;
424         int wonly_disk = -1;
425         const int sectors = r1_bio->sectors;
426         sector_t new_distance, current_distance;
427         mdk_rdev_t *rdev;
428
429         rcu_read_lock();
430         /*
431          * Check if we can balance. We can balance on the whole
432          * device if no resync is going on, or below the resync window.
433          * We take the first readable disk when above the resync window.
434          */
435  retry:
436         if (conf->mddev->recovery_cp < MaxSector &&
437             (this_sector + sectors >= conf->next_resync)) {
438                 /* Choose the first operation device, for consistancy */
439                 new_disk = 0;
440
441                 for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
442                      r1_bio->bios[new_disk] == IO_BLOCKED ||
443                      !rdev || !test_bit(In_sync, &rdev->flags)
444                              || test_bit(WriteMostly, &rdev->flags);
445                      rdev = rcu_dereference(conf->mirrors[++new_disk].rdev)) {
446
447                         if (rdev && test_bit(In_sync, &rdev->flags) &&
448                                 r1_bio->bios[new_disk] != IO_BLOCKED)
449                                 wonly_disk = new_disk;
450
451                         if (new_disk == conf->raid_disks - 1) {
452                                 new_disk = wonly_disk;
453                                 break;
454                         }
455                 }
456                 goto rb_out;
457         }
458
459
460         /* make sure the disk is operational */
461         for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
462              r1_bio->bios[new_disk] == IO_BLOCKED ||
463              !rdev || !test_bit(In_sync, &rdev->flags) ||
464                      test_bit(WriteMostly, &rdev->flags);
465              rdev = rcu_dereference(conf->mirrors[new_disk].rdev)) {
466
467                 if (rdev && test_bit(In_sync, &rdev->flags) &&
468                     r1_bio->bios[new_disk] != IO_BLOCKED)
469                         wonly_disk = new_disk;
470
471                 if (new_disk <= 0)
472                         new_disk = conf->raid_disks;
473                 new_disk--;
474                 if (new_disk == disk) {
475                         new_disk = wonly_disk;
476                         break;
477                 }
478         }
479
480         if (new_disk < 0)
481                 goto rb_out;
482
483         disk = new_disk;
484         /* now disk == new_disk == starting point for search */
485
486         /*
487          * Don't change to another disk for sequential reads:
488          */
489         if (conf->next_seq_sect == this_sector)
490                 goto rb_out;
491         if (this_sector == conf->mirrors[new_disk].head_position)
492                 goto rb_out;
493
494         current_distance = abs(this_sector - conf->mirrors[disk].head_position);
495
496         /* Find the disk whose head is closest */
497
498         do {
499                 if (disk <= 0)
500                         disk = conf->raid_disks;
501                 disk--;
502
503                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
504
505                 if (!rdev || r1_bio->bios[disk] == IO_BLOCKED ||
506                     !test_bit(In_sync, &rdev->flags) ||
507                     test_bit(WriteMostly, &rdev->flags))
508                         continue;
509
510                 if (!atomic_read(&rdev->nr_pending)) {
511                         new_disk = disk;
512                         break;
513                 }
514                 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
515                 if (new_distance < current_distance) {
516                         current_distance = new_distance;
517                         new_disk = disk;
518                 }
519         } while (disk != conf->last_used);
520
521  rb_out:
522
523
524         if (new_disk >= 0) {
525                 rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
526                 if (!rdev)
527                         goto retry;
528                 atomic_inc(&rdev->nr_pending);
529                 if (!test_bit(In_sync, &rdev->flags)) {
530                         /* cannot risk returning a device that failed
531                          * before we inc'ed nr_pending
532                          */
533                         rdev_dec_pending(rdev, conf->mddev);
534                         goto retry;
535                 }
536                 conf->next_seq_sect = this_sector + sectors;
537                 conf->last_used = new_disk;
538         }
539         rcu_read_unlock();
540
541         return new_disk;
542 }
543
544 static void unplug_slaves(mddev_t *mddev)
545 {
546         conf_t *conf = mddev_to_conf(mddev);
547         int i;
548
549         rcu_read_lock();
550         for (i=0; i<mddev->raid_disks; i++) {
551                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
552                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
553                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
554
555                         atomic_inc(&rdev->nr_pending);
556                         rcu_read_unlock();
557
558                         if (r_queue->unplug_fn)
559                                 r_queue->unplug_fn(r_queue);
560
561                         rdev_dec_pending(rdev, mddev);
562                         rcu_read_lock();
563                 }
564         }
565         rcu_read_unlock();
566 }
567
568 static void raid1_unplug(request_queue_t *q)
569 {
570         mddev_t *mddev = q->queuedata;
571
572         unplug_slaves(mddev);
573         md_wakeup_thread(mddev->thread);
574 }
575
576 static int raid1_issue_flush(request_queue_t *q, struct gendisk *disk,
577                              sector_t *error_sector)
578 {
579         mddev_t *mddev = q->queuedata;
580         conf_t *conf = mddev_to_conf(mddev);
581         int i, ret = 0;
582
583         rcu_read_lock();
584         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
585                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
586                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
587                         struct block_device *bdev = rdev->bdev;
588                         request_queue_t *r_queue = bdev_get_queue(bdev);
589
590                         if (!r_queue->issue_flush_fn)
591                                 ret = -EOPNOTSUPP;
592                         else {
593                                 atomic_inc(&rdev->nr_pending);
594                                 rcu_read_unlock();
595                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
596                                                               error_sector);
597                                 rdev_dec_pending(rdev, mddev);
598                                 rcu_read_lock();
599                         }
600                 }
601         }
602         rcu_read_unlock();
603         return ret;
604 }
605
606 /* Barriers....
607  * Sometimes we need to suspend IO while we do something else,
608  * either some resync/recovery, or reconfigure the array.
609  * To do this we raise a 'barrier'.
610  * The 'barrier' is a counter that can be raised multiple times
611  * to count how many activities are happening which preclude
612  * normal IO.
613  * We can only raise the barrier if there is no pending IO.
614  * i.e. if nr_pending == 0.
615  * We choose only to raise the barrier if no-one is waiting for the
616  * barrier to go down.  This means that as soon as an IO request
617  * is ready, no other operations which require a barrier will start
618  * until the IO request has had a chance.
619  *
620  * So: regular IO calls 'wait_barrier'.  When that returns there
621  *    is no backgroup IO happening,  It must arrange to call
622  *    allow_barrier when it has finished its IO.
623  * backgroup IO calls must call raise_barrier.  Once that returns
624  *    there is no normal IO happeing.  It must arrange to call
625  *    lower_barrier when the particular background IO completes.
626  */
627 #define RESYNC_DEPTH 32
628
629 static void raise_barrier(conf_t *conf)
630 {
631         spin_lock_irq(&conf->resync_lock);
632
633         /* Wait until no block IO is waiting */
634         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
635                             conf->resync_lock,
636                             raid1_unplug(conf->mddev->queue));
637
638         /* block any new IO from starting */
639         conf->barrier++;
640
641         /* No wait for all pending IO to complete */
642         wait_event_lock_irq(conf->wait_barrier,
643                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
644                             conf->resync_lock,
645                             raid1_unplug(conf->mddev->queue));
646
647         spin_unlock_irq(&conf->resync_lock);
648 }
649
650 static void lower_barrier(conf_t *conf)
651 {
652         unsigned long flags;
653         spin_lock_irqsave(&conf->resync_lock, flags);
654         conf->barrier--;
655         spin_unlock_irqrestore(&conf->resync_lock, flags);
656         wake_up(&conf->wait_barrier);
657 }
658
659 static void wait_barrier(conf_t *conf)
660 {
661         spin_lock_irq(&conf->resync_lock);
662         if (conf->barrier) {
663                 conf->nr_waiting++;
664                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
665                                     conf->resync_lock,
666                                     raid1_unplug(conf->mddev->queue));
667                 conf->nr_waiting--;
668         }
669         conf->nr_pending++;
670         spin_unlock_irq(&conf->resync_lock);
671 }
672
673 static void allow_barrier(conf_t *conf)
674 {
675         unsigned long flags;
676         spin_lock_irqsave(&conf->resync_lock, flags);
677         conf->nr_pending--;
678         spin_unlock_irqrestore(&conf->resync_lock, flags);
679         wake_up(&conf->wait_barrier);
680 }
681
682 static void freeze_array(conf_t *conf)
683 {
684         /* stop syncio and normal IO and wait for everything to
685          * go quite.
686          * We increment barrier and nr_waiting, and then
687          * wait until barrier+nr_pending match nr_queued+2
688          */
689         spin_lock_irq(&conf->resync_lock);
690         conf->barrier++;
691         conf->nr_waiting++;
692         wait_event_lock_irq(conf->wait_barrier,
693                             conf->barrier+conf->nr_pending == conf->nr_queued+2,
694                             conf->resync_lock,
695                             raid1_unplug(conf->mddev->queue));
696         spin_unlock_irq(&conf->resync_lock);
697 }
698 static void unfreeze_array(conf_t *conf)
699 {
700         /* reverse the effect of the freeze */
701         spin_lock_irq(&conf->resync_lock);
702         conf->barrier--;
703         conf->nr_waiting--;
704         wake_up(&conf->wait_barrier);
705         spin_unlock_irq(&conf->resync_lock);
706 }
707
708
709 /* duplicate the data pages for behind I/O */
710 static struct page **alloc_behind_pages(struct bio *bio)
711 {
712         int i;
713         struct bio_vec *bvec;
714         struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page *),
715                                         GFP_NOIO);
716         if (unlikely(!pages))
717                 goto do_sync_io;
718
719         bio_for_each_segment(bvec, bio, i) {
720                 pages[i] = alloc_page(GFP_NOIO);
721                 if (unlikely(!pages[i]))
722                         goto do_sync_io;
723                 memcpy(kmap(pages[i]) + bvec->bv_offset,
724                         kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
725                 kunmap(pages[i]);
726                 kunmap(bvec->bv_page);
727         }
728
729         return pages;
730
731 do_sync_io:
732         if (pages)
733                 for (i = 0; i < bio->bi_vcnt && pages[i]; i++)
734                         put_page(pages[i]);
735         kfree(pages);
736         PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
737         return NULL;
738 }
739
740 static int make_request(request_queue_t *q, struct bio * bio)
741 {
742         mddev_t *mddev = q->queuedata;
743         conf_t *conf = mddev_to_conf(mddev);
744         mirror_info_t *mirror;
745         r1bio_t *r1_bio;
746         struct bio *read_bio;
747         int i, targets = 0, disks;
748         mdk_rdev_t *rdev;
749         struct bitmap *bitmap = mddev->bitmap;
750         unsigned long flags;
751         struct bio_list bl;
752         struct page **behind_pages = NULL;
753         const int rw = bio_data_dir(bio);
754         int do_barriers;
755
756         if (unlikely(!mddev->barriers_work && bio_barrier(bio))) {
757                 bio_endio(bio, bio->bi_size, -EOPNOTSUPP);
758                 return 0;
759         }
760
761         /*
762          * Register the new request and wait if the reconstruction
763          * thread has put up a bar for new requests.
764          * Continue immediately if no resync is active currently.
765          */
766         md_write_start(mddev, bio); /* wait on superblock update early */
767
768         wait_barrier(conf);
769
770         disk_stat_inc(mddev->gendisk, ios[rw]);
771         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
772
773         /*
774          * make_request() can abort the operation when READA is being
775          * used and no empty request is available.
776          *
777          */
778         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
779
780         r1_bio->master_bio = bio;
781         r1_bio->sectors = bio->bi_size >> 9;
782         r1_bio->state = 0;
783         r1_bio->mddev = mddev;
784         r1_bio->sector = bio->bi_sector;
785
786         if (rw == READ) {
787                 /*
788                  * read balancing logic:
789                  */
790                 int rdisk = read_balance(conf, r1_bio);
791
792                 if (rdisk < 0) {
793                         /* couldn't find anywhere to read from */
794                         raid_end_bio_io(r1_bio);
795                         return 0;
796                 }
797                 mirror = conf->mirrors + rdisk;
798
799                 r1_bio->read_disk = rdisk;
800
801                 read_bio = bio_clone(bio, GFP_NOIO);
802
803                 r1_bio->bios[rdisk] = read_bio;
804
805                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
806                 read_bio->bi_bdev = mirror->rdev->bdev;
807                 read_bio->bi_end_io = raid1_end_read_request;
808                 read_bio->bi_rw = READ;
809                 read_bio->bi_private = r1_bio;
810
811                 generic_make_request(read_bio);
812                 return 0;
813         }
814
815         /*
816          * WRITE:
817          */
818         /* first select target devices under spinlock and
819          * inc refcount on their rdev.  Record them by setting
820          * bios[x] to bio
821          */
822         disks = conf->raid_disks;
823 #if 0
824         { static int first=1;
825         if (first) printk("First Write sector %llu disks %d\n",
826                           (unsigned long long)r1_bio->sector, disks);
827         first = 0;
828         }
829 #endif
830         rcu_read_lock();
831         for (i = 0;  i < disks; i++) {
832                 if ((rdev=rcu_dereference(conf->mirrors[i].rdev)) != NULL &&
833                     !test_bit(Faulty, &rdev->flags)) {
834                         atomic_inc(&rdev->nr_pending);
835                         if (test_bit(Faulty, &rdev->flags)) {
836                                 rdev_dec_pending(rdev, mddev);
837                                 r1_bio->bios[i] = NULL;
838                         } else
839                                 r1_bio->bios[i] = bio;
840                         targets++;
841                 } else
842                         r1_bio->bios[i] = NULL;
843         }
844         rcu_read_unlock();
845
846         BUG_ON(targets == 0); /* we never fail the last device */
847
848         if (targets < conf->raid_disks) {
849                 /* array is degraded, we will not clear the bitmap
850                  * on I/O completion (see raid1_end_write_request) */
851                 set_bit(R1BIO_Degraded, &r1_bio->state);
852         }
853
854         /* do behind I/O ? */
855         if (bitmap &&
856             atomic_read(&bitmap->behind_writes) < bitmap->max_write_behind &&
857             (behind_pages = alloc_behind_pages(bio)) != NULL)
858                 set_bit(R1BIO_BehindIO, &r1_bio->state);
859
860         atomic_set(&r1_bio->remaining, 0);
861         atomic_set(&r1_bio->behind_remaining, 0);
862
863         do_barriers = bio_barrier(bio);
864         if (do_barriers)
865                 set_bit(R1BIO_Barrier, &r1_bio->state);
866
867         bio_list_init(&bl);
868         for (i = 0; i < disks; i++) {
869                 struct bio *mbio;
870                 if (!r1_bio->bios[i])
871                         continue;
872
873                 mbio = bio_clone(bio, GFP_NOIO);
874                 r1_bio->bios[i] = mbio;
875
876                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
877                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
878                 mbio->bi_end_io = raid1_end_write_request;
879                 mbio->bi_rw = WRITE | do_barriers;
880                 mbio->bi_private = r1_bio;
881
882                 if (behind_pages) {
883                         struct bio_vec *bvec;
884                         int j;
885
886                         /* Yes, I really want the '__' version so that
887                          * we clear any unused pointer in the io_vec, rather
888                          * than leave them unchanged.  This is important
889                          * because when we come to free the pages, we won't
890                          * know the originial bi_idx, so we just free
891                          * them all
892                          */
893                         __bio_for_each_segment(bvec, mbio, j, 0)
894                                 bvec->bv_page = behind_pages[j];
895                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
896                                 atomic_inc(&r1_bio->behind_remaining);
897                 }
898
899                 atomic_inc(&r1_bio->remaining);
900
901                 bio_list_add(&bl, mbio);
902         }
903         kfree(behind_pages); /* the behind pages are attached to the bios now */
904
905         bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
906                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
907         spin_lock_irqsave(&conf->device_lock, flags);
908         bio_list_merge(&conf->pending_bio_list, &bl);
909         bio_list_init(&bl);
910
911         blk_plug_device(mddev->queue);
912         spin_unlock_irqrestore(&conf->device_lock, flags);
913
914 #if 0
915         while ((bio = bio_list_pop(&bl)) != NULL)
916                 generic_make_request(bio);
917 #endif
918
919         return 0;
920 }
921
922 static void status(struct seq_file *seq, mddev_t *mddev)
923 {
924         conf_t *conf = mddev_to_conf(mddev);
925         int i;
926
927         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
928                                                 conf->working_disks);
929         for (i = 0; i < conf->raid_disks; i++)
930                 seq_printf(seq, "%s",
931                               conf->mirrors[i].rdev &&
932                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
933         seq_printf(seq, "]");
934 }
935
936
937 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
938 {
939         char b[BDEVNAME_SIZE];
940         conf_t *conf = mddev_to_conf(mddev);
941
942         /*
943          * If it is not operational, then we have already marked it as dead
944          * else if it is the last working disks, ignore the error, let the
945          * next level up know.
946          * else mark the drive as failed
947          */
948         if (test_bit(In_sync, &rdev->flags)
949             && conf->working_disks == 1)
950                 /*
951                  * Don't fail the drive, act as though we were just a
952                  * normal single drive
953                  */
954                 return;
955         if (test_bit(In_sync, &rdev->flags)) {
956                 mddev->degraded++;
957                 conf->working_disks--;
958                 /*
959                  * if recovery is running, make sure it aborts.
960                  */
961                 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
962         }
963         clear_bit(In_sync, &rdev->flags);
964         set_bit(Faulty, &rdev->flags);
965         mddev->sb_dirty = 1;
966         printk(KERN_ALERT "raid1: Disk failure on %s, disabling device. \n"
967                 "       Operation continuing on %d devices\n",
968                 bdevname(rdev->bdev,b), conf->working_disks);
969 }
970
971 static void print_conf(conf_t *conf)
972 {
973         int i;
974         mirror_info_t *tmp;
975
976         printk("RAID1 conf printout:\n");
977         if (!conf) {
978                 printk("(!conf)\n");
979                 return;
980         }
981         printk(" --- wd:%d rd:%d\n", conf->working_disks,
982                 conf->raid_disks);
983
984         for (i = 0; i < conf->raid_disks; i++) {
985                 char b[BDEVNAME_SIZE];
986                 tmp = conf->mirrors + i;
987                 if (tmp->rdev)
988                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
989                                 i, !test_bit(In_sync, &tmp->rdev->flags), !test_bit(Faulty, &tmp->rdev->flags),
990                                 bdevname(tmp->rdev->bdev,b));
991         }
992 }
993
994 static void close_sync(conf_t *conf)
995 {
996         wait_barrier(conf);
997         allow_barrier(conf);
998
999         mempool_destroy(conf->r1buf_pool);
1000         conf->r1buf_pool = NULL;
1001 }
1002
1003 static int raid1_spare_active(mddev_t *mddev)
1004 {
1005         int i;
1006         conf_t *conf = mddev->private;
1007         mirror_info_t *tmp;
1008
1009         /*
1010          * Find all failed disks within the RAID1 configuration 
1011          * and mark them readable
1012          */
1013         for (i = 0; i < conf->raid_disks; i++) {
1014                 tmp = conf->mirrors + i;
1015                 if (tmp->rdev 
1016                     && !test_bit(Faulty, &tmp->rdev->flags)
1017                     && !test_bit(In_sync, &tmp->rdev->flags)) {
1018                         conf->working_disks++;
1019                         mddev->degraded--;
1020                         set_bit(In_sync, &tmp->rdev->flags);
1021                 }
1022         }
1023
1024         print_conf(conf);
1025         return 0;
1026 }
1027
1028
1029 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1030 {
1031         conf_t *conf = mddev->private;
1032         int found = 0;
1033         int mirror = 0;
1034         mirror_info_t *p;
1035
1036         for (mirror=0; mirror < mddev->raid_disks; mirror++)
1037                 if ( !(p=conf->mirrors+mirror)->rdev) {
1038
1039                         blk_queue_stack_limits(mddev->queue,
1040                                                rdev->bdev->bd_disk->queue);
1041                         /* as we don't honour merge_bvec_fn, we must never risk
1042                          * violating it, so limit ->max_sector to one PAGE, as
1043                          * a one page request is never in violation.
1044                          */
1045                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1046                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
1047                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1048
1049                         p->head_position = 0;
1050                         rdev->raid_disk = mirror;
1051                         found = 1;
1052                         /* As all devices are equivalent, we don't need a full recovery
1053                          * if this was recently any drive of the array
1054                          */
1055                         if (rdev->saved_raid_disk < 0)
1056                                 conf->fullsync = 1;
1057                         rcu_assign_pointer(p->rdev, rdev);
1058                         break;
1059                 }
1060
1061         print_conf(conf);
1062         return found;
1063 }
1064
1065 static int raid1_remove_disk(mddev_t *mddev, int number)
1066 {
1067         conf_t *conf = mddev->private;
1068         int err = 0;
1069         mdk_rdev_t *rdev;
1070         mirror_info_t *p = conf->mirrors+ number;
1071
1072         print_conf(conf);
1073         rdev = p->rdev;
1074         if (rdev) {
1075                 if (test_bit(In_sync, &rdev->flags) ||
1076                     atomic_read(&rdev->nr_pending)) {
1077                         err = -EBUSY;
1078                         goto abort;
1079                 }
1080                 p->rdev = NULL;
1081                 synchronize_rcu();
1082                 if (atomic_read(&rdev->nr_pending)) {
1083                         /* lost the race, try later */
1084                         err = -EBUSY;
1085                         p->rdev = rdev;
1086                 }
1087         }
1088 abort:
1089
1090         print_conf(conf);
1091         return err;
1092 }
1093
1094
1095 static int end_sync_read(struct bio *bio, unsigned int bytes_done, int error)
1096 {
1097         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1098         int i;
1099
1100         if (bio->bi_size)
1101                 return 1;
1102
1103         for (i=r1_bio->mddev->raid_disks; i--; )
1104                 if (r1_bio->bios[i] == bio)
1105                         break;
1106         BUG_ON(i < 0);
1107         update_head_pos(i, r1_bio);
1108         /*
1109          * we have read a block, now it needs to be re-written,
1110          * or re-read if the read failed.
1111          * We don't do much here, just schedule handling by raid1d
1112          */
1113         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1114                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1115
1116         if (atomic_dec_and_test(&r1_bio->remaining))
1117                 reschedule_retry(r1_bio);
1118         return 0;
1119 }
1120
1121 static int end_sync_write(struct bio *bio, unsigned int bytes_done, int error)
1122 {
1123         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1124         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1125         mddev_t *mddev = r1_bio->mddev;
1126         conf_t *conf = mddev_to_conf(mddev);
1127         int i;
1128         int mirror=0;
1129
1130         if (bio->bi_size)
1131                 return 1;
1132
1133         for (i = 0; i < conf->raid_disks; i++)
1134                 if (r1_bio->bios[i] == bio) {
1135                         mirror = i;
1136                         break;
1137                 }
1138         if (!uptodate)
1139                 md_error(mddev, conf->mirrors[mirror].rdev);
1140
1141         update_head_pos(mirror, r1_bio);
1142
1143         if (atomic_dec_and_test(&r1_bio->remaining)) {
1144                 md_done_sync(mddev, r1_bio->sectors, uptodate);
1145                 put_buf(r1_bio);
1146         }
1147         return 0;
1148 }
1149
1150 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1151 {
1152         conf_t *conf = mddev_to_conf(mddev);
1153         int i;
1154         int disks = conf->raid_disks;
1155         struct bio *bio, *wbio;
1156
1157         bio = r1_bio->bios[r1_bio->read_disk];
1158
1159
1160         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1161                 /* We have read all readable devices.  If we haven't
1162                  * got the block, then there is no hope left.
1163                  * If we have, then we want to do a comparison
1164                  * and skip the write if everything is the same.
1165                  * If any blocks failed to read, then we need to
1166                  * attempt an over-write
1167                  */
1168                 int primary;
1169                 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1170                         for (i=0; i<mddev->raid_disks; i++)
1171                                 if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1172                                         md_error(mddev, conf->mirrors[i].rdev);
1173
1174                         md_done_sync(mddev, r1_bio->sectors, 1);
1175                         put_buf(r1_bio);
1176                         return;
1177                 }
1178                 for (primary=0; primary<mddev->raid_disks; primary++)
1179                         if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1180                             test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1181                                 r1_bio->bios[primary]->bi_end_io = NULL;
1182                                 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1183                                 break;
1184                         }
1185                 r1_bio->read_disk = primary;
1186                 for (i=0; i<mddev->raid_disks; i++)
1187                         if (r1_bio->bios[i]->bi_end_io == end_sync_read &&
1188                             test_bit(BIO_UPTODATE, &r1_bio->bios[i]->bi_flags)) {
1189                                 int j;
1190                                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1191                                 struct bio *pbio = r1_bio->bios[primary];
1192                                 struct bio *sbio = r1_bio->bios[i];
1193                                 for (j = vcnt; j-- ; )
1194                                         if (memcmp(page_address(pbio->bi_io_vec[j].bv_page),
1195                                                    page_address(sbio->bi_io_vec[j].bv_page),
1196                                                    PAGE_SIZE))
1197                                                 break;
1198                                 if (j >= 0)
1199                                         mddev->resync_mismatches += r1_bio->sectors;
1200                                 if (j < 0 || test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
1201                                         sbio->bi_end_io = NULL;
1202                                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1203                                 } else {
1204                                         /* fixup the bio for reuse */
1205                                         sbio->bi_vcnt = vcnt;
1206                                         sbio->bi_size = r1_bio->sectors << 9;
1207                                         sbio->bi_idx = 0;
1208                                         sbio->bi_phys_segments = 0;
1209                                         sbio->bi_hw_segments = 0;
1210                                         sbio->bi_hw_front_size = 0;
1211                                         sbio->bi_hw_back_size = 0;
1212                                         sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1213                                         sbio->bi_flags |= 1 << BIO_UPTODATE;
1214                                         sbio->bi_next = NULL;
1215                                         sbio->bi_sector = r1_bio->sector +
1216                                                 conf->mirrors[i].rdev->data_offset;
1217                                         sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1218                                 }
1219                         }
1220         }
1221         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1222                 /* ouch - failed to read all of that.
1223                  * Try some synchronous reads of other devices to get
1224                  * good data, much like with normal read errors.  Only
1225                  * read into the pages we already have so they we don't
1226                  * need to re-issue the read request.
1227                  * We don't need to freeze the array, because being in an
1228                  * active sync request, there is no normal IO, and
1229                  * no overlapping syncs.
1230                  */
1231                 sector_t sect = r1_bio->sector;
1232                 int sectors = r1_bio->sectors;
1233                 int idx = 0;
1234
1235                 while(sectors) {
1236                         int s = sectors;
1237                         int d = r1_bio->read_disk;
1238                         int success = 0;
1239                         mdk_rdev_t *rdev;
1240
1241                         if (s > (PAGE_SIZE>>9))
1242                                 s = PAGE_SIZE >> 9;
1243                         do {
1244                                 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1245                                         rdev = conf->mirrors[d].rdev;
1246                                         if (sync_page_io(rdev->bdev,
1247                                                          sect + rdev->data_offset,
1248                                                          s<<9,
1249                                                          bio->bi_io_vec[idx].bv_page,
1250                                                          READ)) {
1251                                                 success = 1;
1252                                                 break;
1253                                         }
1254                                 }
1255                                 d++;
1256                                 if (d == conf->raid_disks)
1257                                         d = 0;
1258                         } while (!success && d != r1_bio->read_disk);
1259
1260                         if (success) {
1261                                 int start = d;
1262                                 /* write it back and re-read */
1263                                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1264                                 while (d != r1_bio->read_disk) {
1265                                         if (d == 0)
1266                                                 d = conf->raid_disks;
1267                                         d--;
1268                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1269                                                 continue;
1270                                         rdev = conf->mirrors[d].rdev;
1271                                         atomic_add(s, &rdev->corrected_errors);
1272                                         if (sync_page_io(rdev->bdev,
1273                                                          sect + rdev->data_offset,
1274                                                          s<<9,
1275                                                          bio->bi_io_vec[idx].bv_page,
1276                                                          WRITE) == 0)
1277                                                 md_error(mddev, rdev);
1278                                 }
1279                                 d = start;
1280                                 while (d != r1_bio->read_disk) {
1281                                         if (d == 0)
1282                                                 d = conf->raid_disks;
1283                                         d--;
1284                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1285                                                 continue;
1286                                         rdev = conf->mirrors[d].rdev;
1287                                         if (sync_page_io(rdev->bdev,
1288                                                          sect + rdev->data_offset,
1289                                                          s<<9,
1290                                                          bio->bi_io_vec[idx].bv_page,
1291                                                          READ) == 0)
1292                                                 md_error(mddev, rdev);
1293                                 }
1294                         } else {
1295                                 char b[BDEVNAME_SIZE];
1296                                 /* Cannot read from anywhere, array is toast */
1297                                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1298                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
1299                                        " for block %llu\n",
1300                                        bdevname(bio->bi_bdev,b),
1301                                        (unsigned long long)r1_bio->sector);
1302                                 md_done_sync(mddev, r1_bio->sectors, 0);
1303                                 put_buf(r1_bio);
1304                                 return;
1305                         }
1306                         sectors -= s;
1307                         sect += s;
1308                         idx ++;
1309                 }
1310         }
1311
1312         /*
1313          * schedule writes
1314          */
1315         atomic_set(&r1_bio->remaining, 1);
1316         for (i = 0; i < disks ; i++) {
1317                 wbio = r1_bio->bios[i];
1318                 if (wbio->bi_end_io == NULL ||
1319                     (wbio->bi_end_io == end_sync_read &&
1320                      (i == r1_bio->read_disk ||
1321                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1322                         continue;
1323
1324                 wbio->bi_rw = WRITE;
1325                 wbio->bi_end_io = end_sync_write;
1326                 atomic_inc(&r1_bio->remaining);
1327                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1328
1329                 generic_make_request(wbio);
1330         }
1331
1332         if (atomic_dec_and_test(&r1_bio->remaining)) {
1333                 /* if we're here, all write(s) have completed, so clean up */
1334                 md_done_sync(mddev, r1_bio->sectors, 1);
1335                 put_buf(r1_bio);
1336         }
1337 }
1338
1339 /*
1340  * This is a kernel thread which:
1341  *
1342  *      1.      Retries failed read operations on working mirrors.
1343  *      2.      Updates the raid superblock when problems encounter.
1344  *      3.      Performs writes following reads for array syncronising.
1345  */
1346
1347 static void raid1d(mddev_t *mddev)
1348 {
1349         r1bio_t *r1_bio;
1350         struct bio *bio;
1351         unsigned long flags;
1352         conf_t *conf = mddev_to_conf(mddev);
1353         struct list_head *head = &conf->retry_list;
1354         int unplug=0;
1355         mdk_rdev_t *rdev;
1356
1357         md_check_recovery(mddev);
1358         
1359         for (;;) {
1360                 char b[BDEVNAME_SIZE];
1361                 spin_lock_irqsave(&conf->device_lock, flags);
1362
1363                 if (conf->pending_bio_list.head) {
1364                         bio = bio_list_get(&conf->pending_bio_list);
1365                         blk_remove_plug(mddev->queue);
1366                         spin_unlock_irqrestore(&conf->device_lock, flags);
1367                         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
1368                         if (bitmap_unplug(mddev->bitmap) != 0)
1369                                 printk("%s: bitmap file write failed!\n", mdname(mddev));
1370
1371                         while (bio) { /* submit pending writes */
1372                                 struct bio *next = bio->bi_next;
1373                                 bio->bi_next = NULL;
1374                                 generic_make_request(bio);
1375                                 bio = next;
1376                         }
1377                         unplug = 1;
1378
1379                         continue;
1380                 }
1381
1382                 if (list_empty(head))
1383                         break;
1384                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1385                 list_del(head->prev);
1386                 conf->nr_queued--;
1387                 spin_unlock_irqrestore(&conf->device_lock, flags);
1388
1389                 mddev = r1_bio->mddev;
1390                 conf = mddev_to_conf(mddev);
1391                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1392                         sync_request_write(mddev, r1_bio);
1393                         unplug = 1;
1394                 } else if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
1395                         /* some requests in the r1bio were BIO_RW_BARRIER
1396                          * requests which failed with -ENOTSUPP.  Hohumm..
1397                          * Better resubmit without the barrier.
1398                          * We know which devices to resubmit for, because
1399                          * all others have had their bios[] entry cleared.
1400                          */
1401                         int i;
1402                         clear_bit(R1BIO_BarrierRetry, &r1_bio->state);
1403                         clear_bit(R1BIO_Barrier, &r1_bio->state);
1404                         for (i=0; i < conf->raid_disks; i++)
1405                                 if (r1_bio->bios[i]) {
1406                                         struct bio_vec *bvec;
1407                                         int j;
1408
1409                                         bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1410                                         /* copy pages from the failed bio, as
1411                                          * this might be a write-behind device */
1412                                         __bio_for_each_segment(bvec, bio, j, 0)
1413                                                 bvec->bv_page = bio_iovec_idx(r1_bio->bios[i], j)->bv_page;
1414                                         bio_put(r1_bio->bios[i]);
1415                                         bio->bi_sector = r1_bio->sector +
1416                                                 conf->mirrors[i].rdev->data_offset;
1417                                         bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1418                                         bio->bi_end_io = raid1_end_write_request;
1419                                         bio->bi_rw = WRITE;
1420                                         bio->bi_private = r1_bio;
1421                                         r1_bio->bios[i] = bio;
1422                                         generic_make_request(bio);
1423                                 }
1424                 } else {
1425                         int disk;
1426
1427                         /* we got a read error. Maybe the drive is bad.  Maybe just
1428                          * the block and we can fix it.
1429                          * We freeze all other IO, and try reading the block from
1430                          * other devices.  When we find one, we re-write
1431                          * and check it that fixes the read error.
1432                          * This is all done synchronously while the array is
1433                          * frozen
1434                          */
1435                         sector_t sect = r1_bio->sector;
1436                         int sectors = r1_bio->sectors;
1437                         freeze_array(conf);
1438                         if (mddev->ro == 0) while(sectors) {
1439                                 int s = sectors;
1440                                 int d = r1_bio->read_disk;
1441                                 int success = 0;
1442
1443                                 if (s > (PAGE_SIZE>>9))
1444                                         s = PAGE_SIZE >> 9;
1445
1446                                 do {
1447                                         rdev = conf->mirrors[d].rdev;
1448                                         if (rdev &&
1449                                             test_bit(In_sync, &rdev->flags) &&
1450                                             sync_page_io(rdev->bdev,
1451                                                          sect + rdev->data_offset,
1452                                                          s<<9,
1453                                                          conf->tmppage, READ))
1454                                                 success = 1;
1455                                         else {
1456                                                 d++;
1457                                                 if (d == conf->raid_disks)
1458                                                         d = 0;
1459                                         }
1460                                 } while (!success && d != r1_bio->read_disk);
1461
1462                                 if (success) {
1463                                         /* write it back and re-read */
1464                                         int start = d;
1465                                         while (d != r1_bio->read_disk) {
1466                                                 if (d==0)
1467                                                         d = conf->raid_disks;
1468                                                 d--;
1469                                                 rdev = conf->mirrors[d].rdev;
1470                                                 atomic_add(s, &rdev->corrected_errors);
1471                                                 if (rdev &&
1472                                                     test_bit(In_sync, &rdev->flags)) {
1473                                                         if (sync_page_io(rdev->bdev,
1474                                                                          sect + rdev->data_offset,
1475                                                                          s<<9, conf->tmppage, WRITE) == 0)
1476                                                                 /* Well, this device is dead */
1477                                                                 md_error(mddev, rdev);
1478                                                 }
1479                                         }
1480                                         d = start;
1481                                         while (d != r1_bio->read_disk) {
1482                                                 if (d==0)
1483                                                         d = conf->raid_disks;
1484                                                 d--;
1485                                                 rdev = conf->mirrors[d].rdev;
1486                                                 if (rdev &&
1487                                                     test_bit(In_sync, &rdev->flags)) {
1488                                                         if (sync_page_io(rdev->bdev,
1489                                                                          sect + rdev->data_offset,
1490                                                                          s<<9, conf->tmppage, READ) == 0)
1491                                                                 /* Well, this device is dead */
1492                                                                 md_error(mddev, rdev);
1493                                                 }
1494                                         }
1495                                 } else {
1496                                         /* Cannot read from anywhere -- bye bye array */
1497                                         md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1498                                         break;
1499                                 }
1500                                 sectors -= s;
1501                                 sect += s;
1502                         }
1503
1504                         unfreeze_array(conf);
1505
1506                         bio = r1_bio->bios[r1_bio->read_disk];
1507                         if ((disk=read_balance(conf, r1_bio)) == -1) {
1508                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
1509                                        " read error for block %llu\n",
1510                                        bdevname(bio->bi_bdev,b),
1511                                        (unsigned long long)r1_bio->sector);
1512                                 raid_end_bio_io(r1_bio);
1513                         } else {
1514                                 r1_bio->bios[r1_bio->read_disk] =
1515                                         mddev->ro ? IO_BLOCKED : NULL;
1516                                 r1_bio->read_disk = disk;
1517                                 bio_put(bio);
1518                                 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1519                                 r1_bio->bios[r1_bio->read_disk] = bio;
1520                                 rdev = conf->mirrors[disk].rdev;
1521                                 if (printk_ratelimit())
1522                                         printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
1523                                                " another mirror\n",
1524                                                bdevname(rdev->bdev,b),
1525                                                (unsigned long long)r1_bio->sector);
1526                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1527                                 bio->bi_bdev = rdev->bdev;
1528                                 bio->bi_end_io = raid1_end_read_request;
1529                                 bio->bi_rw = READ;
1530                                 bio->bi_private = r1_bio;
1531                                 unplug = 1;
1532                                 generic_make_request(bio);
1533                         }
1534                 }
1535         }
1536         spin_unlock_irqrestore(&conf->device_lock, flags);
1537         if (unplug)
1538                 unplug_slaves(mddev);
1539 }
1540
1541
1542 static int init_resync(conf_t *conf)
1543 {
1544         int buffs;
1545
1546         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1547         if (conf->r1buf_pool)
1548                 BUG();
1549         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1550                                           conf->poolinfo);
1551         if (!conf->r1buf_pool)
1552                 return -ENOMEM;
1553         conf->next_resync = 0;
1554         return 0;
1555 }
1556
1557 /*
1558  * perform a "sync" on one "block"
1559  *
1560  * We need to make sure that no normal I/O request - particularly write
1561  * requests - conflict with active sync requests.
1562  *
1563  * This is achieved by tracking pending requests and a 'barrier' concept
1564  * that can be installed to exclude normal IO requests.
1565  */
1566
1567 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1568 {
1569         conf_t *conf = mddev_to_conf(mddev);
1570         r1bio_t *r1_bio;
1571         struct bio *bio;
1572         sector_t max_sector, nr_sectors;
1573         int disk = -1;
1574         int i;
1575         int wonly = -1;
1576         int write_targets = 0, read_targets = 0;
1577         int sync_blocks;
1578         int still_degraded = 0;
1579
1580         if (!conf->r1buf_pool)
1581         {
1582 /*
1583                 printk("sync start - bitmap %p\n", mddev->bitmap);
1584 */
1585                 if (init_resync(conf))
1586                         return 0;
1587         }
1588
1589         max_sector = mddev->size << 1;
1590         if (sector_nr >= max_sector) {
1591                 /* If we aborted, we need to abort the
1592                  * sync on the 'current' bitmap chunk (there will
1593                  * only be one in raid1 resync.
1594                  * We can find the current addess in mddev->curr_resync
1595                  */
1596                 if (mddev->curr_resync < max_sector) /* aborted */
1597                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1598                                                 &sync_blocks, 1);
1599                 else /* completed sync */
1600                         conf->fullsync = 0;
1601
1602                 bitmap_close_sync(mddev->bitmap);
1603                 close_sync(conf);
1604                 return 0;
1605         }
1606
1607         /* before building a request, check if we can skip these blocks..
1608          * This call the bitmap_start_sync doesn't actually record anything
1609          */
1610         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1611             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1612                 /* We can skip this block, and probably several more */
1613                 *skipped = 1;
1614                 return sync_blocks;
1615         }
1616         /*
1617          * If there is non-resync activity waiting for a turn,
1618          * and resync is going fast enough,
1619          * then let it though before starting on this new sync request.
1620          */
1621         if (!go_faster && conf->nr_waiting)
1622                 msleep_interruptible(1000);
1623
1624         raise_barrier(conf);
1625
1626         conf->next_resync = sector_nr;
1627
1628         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1629         rcu_read_lock();
1630         /*
1631          * If we get a correctably read error during resync or recovery,
1632          * we might want to read from a different device.  So we
1633          * flag all drives that could conceivably be read from for READ,
1634          * and any others (which will be non-In_sync devices) for WRITE.
1635          * If a read fails, we try reading from something else for which READ
1636          * is OK.
1637          */
1638
1639         r1_bio->mddev = mddev;
1640         r1_bio->sector = sector_nr;
1641         r1_bio->state = 0;
1642         set_bit(R1BIO_IsSync, &r1_bio->state);
1643
1644         for (i=0; i < conf->raid_disks; i++) {
1645                 mdk_rdev_t *rdev;
1646                 bio = r1_bio->bios[i];
1647
1648                 /* take from bio_init */
1649                 bio->bi_next = NULL;
1650                 bio->bi_flags |= 1 << BIO_UPTODATE;
1651                 bio->bi_rw = 0;
1652                 bio->bi_vcnt = 0;
1653                 bio->bi_idx = 0;
1654                 bio->bi_phys_segments = 0;
1655                 bio->bi_hw_segments = 0;
1656                 bio->bi_size = 0;
1657                 bio->bi_end_io = NULL;
1658                 bio->bi_private = NULL;
1659
1660                 rdev = rcu_dereference(conf->mirrors[i].rdev);
1661                 if (rdev == NULL ||
1662                            test_bit(Faulty, &rdev->flags)) {
1663                         still_degraded = 1;
1664                         continue;
1665                 } else if (!test_bit(In_sync, &rdev->flags)) {
1666                         bio->bi_rw = WRITE;
1667                         bio->bi_end_io = end_sync_write;
1668                         write_targets ++;
1669                 } else {
1670                         /* may need to read from here */
1671                         bio->bi_rw = READ;
1672                         bio->bi_end_io = end_sync_read;
1673                         if (test_bit(WriteMostly, &rdev->flags)) {
1674                                 if (wonly < 0)
1675                                         wonly = i;
1676                         } else {
1677                                 if (disk < 0)
1678                                         disk = i;
1679                         }
1680                         read_targets++;
1681                 }
1682                 atomic_inc(&rdev->nr_pending);
1683                 bio->bi_sector = sector_nr + rdev->data_offset;
1684                 bio->bi_bdev = rdev->bdev;
1685                 bio->bi_private = r1_bio;
1686         }
1687         rcu_read_unlock();
1688         if (disk < 0)
1689                 disk = wonly;
1690         r1_bio->read_disk = disk;
1691
1692         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1693                 /* extra read targets are also write targets */
1694                 write_targets += read_targets-1;
1695
1696         if (write_targets == 0 || read_targets == 0) {
1697                 /* There is nowhere to write, so all non-sync
1698                  * drives must be failed - so we are finished
1699                  */
1700                 sector_t rv = max_sector - sector_nr;
1701                 *skipped = 1;
1702                 put_buf(r1_bio);
1703                 return rv;
1704         }
1705
1706         nr_sectors = 0;
1707         sync_blocks = 0;
1708         do {
1709                 struct page *page;
1710                 int len = PAGE_SIZE;
1711                 if (sector_nr + (len>>9) > max_sector)
1712                         len = (max_sector - sector_nr) << 9;
1713                 if (len == 0)
1714                         break;
1715                 if (sync_blocks == 0) {
1716                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1717                                                &sync_blocks, still_degraded) &&
1718                             !conf->fullsync &&
1719                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1720                                 break;
1721                         if (sync_blocks < (PAGE_SIZE>>9))
1722                                 BUG();
1723                         if (len > (sync_blocks<<9))
1724                                 len = sync_blocks<<9;
1725                 }
1726
1727                 for (i=0 ; i < conf->raid_disks; i++) {
1728                         bio = r1_bio->bios[i];
1729                         if (bio->bi_end_io) {
1730                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1731                                 if (bio_add_page(bio, page, len, 0) == 0) {
1732                                         /* stop here */
1733                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1734                                         while (i > 0) {
1735                                                 i--;
1736                                                 bio = r1_bio->bios[i];
1737                                                 if (bio->bi_end_io==NULL)
1738                                                         continue;
1739                                                 /* remove last page from this bio */
1740                                                 bio->bi_vcnt--;
1741                                                 bio->bi_size -= len;
1742                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1743                                         }
1744                                         goto bio_full;
1745                                 }
1746                         }
1747                 }
1748                 nr_sectors += len>>9;
1749                 sector_nr += len>>9;
1750                 sync_blocks -= (len>>9);
1751         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1752  bio_full:
1753         r1_bio->sectors = nr_sectors;
1754
1755         /* For a user-requested sync, we read all readable devices and do a
1756          * compare
1757          */
1758         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1759                 atomic_set(&r1_bio->remaining, read_targets);
1760                 for (i=0; i<conf->raid_disks; i++) {
1761                         bio = r1_bio->bios[i];
1762                         if (bio->bi_end_io == end_sync_read) {
1763                                 md_sync_acct(conf->mirrors[i].rdev->bdev, nr_sectors);
1764                                 generic_make_request(bio);
1765                         }
1766                 }
1767         } else {
1768                 atomic_set(&r1_bio->remaining, 1);
1769                 bio = r1_bio->bios[r1_bio->read_disk];
1770                 md_sync_acct(conf->mirrors[r1_bio->read_disk].rdev->bdev,
1771                              nr_sectors);
1772                 generic_make_request(bio);
1773
1774         }
1775
1776         return nr_sectors;
1777 }
1778
1779 static int run(mddev_t *mddev)
1780 {
1781         conf_t *conf;
1782         int i, j, disk_idx;
1783         mirror_info_t *disk;
1784         mdk_rdev_t *rdev;
1785         struct list_head *tmp;
1786
1787         if (mddev->level != 1) {
1788                 printk("raid1: %s: raid level not set to mirroring (%d)\n",
1789                        mdname(mddev), mddev->level);
1790                 goto out;
1791         }
1792         /*
1793          * copy the already verified devices into our private RAID1
1794          * bookkeeping area. [whatever we allocate in run(),
1795          * should be freed in stop()]
1796          */
1797         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1798         mddev->private = conf;
1799         if (!conf)
1800                 goto out_no_mem;
1801
1802         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1803                                  GFP_KERNEL);
1804         if (!conf->mirrors)
1805                 goto out_no_mem;
1806
1807         conf->tmppage = alloc_page(GFP_KERNEL);
1808         if (!conf->tmppage)
1809                 goto out_no_mem;
1810
1811         conf->poolinfo = kmalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1812         if (!conf->poolinfo)
1813                 goto out_no_mem;
1814         conf->poolinfo->mddev = mddev;
1815         conf->poolinfo->raid_disks = mddev->raid_disks;
1816         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1817                                           r1bio_pool_free,
1818                                           conf->poolinfo);
1819         if (!conf->r1bio_pool)
1820                 goto out_no_mem;
1821
1822         ITERATE_RDEV(mddev, rdev, tmp) {
1823                 disk_idx = rdev->raid_disk;
1824                 if (disk_idx >= mddev->raid_disks
1825                     || disk_idx < 0)
1826                         continue;
1827                 disk = conf->mirrors + disk_idx;
1828
1829                 disk->rdev = rdev;
1830
1831                 blk_queue_stack_limits(mddev->queue,
1832                                        rdev->bdev->bd_disk->queue);
1833                 /* as we don't honour merge_bvec_fn, we must never risk
1834                  * violating it, so limit ->max_sector to one PAGE, as
1835                  * a one page request is never in violation.
1836                  */
1837                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1838                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
1839                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1840
1841                 disk->head_position = 0;
1842                 if (!test_bit(Faulty, &rdev->flags) && test_bit(In_sync, &rdev->flags))
1843                         conf->working_disks++;
1844         }
1845         conf->raid_disks = mddev->raid_disks;
1846         conf->mddev = mddev;
1847         spin_lock_init(&conf->device_lock);
1848         INIT_LIST_HEAD(&conf->retry_list);
1849         if (conf->working_disks == 1)
1850                 mddev->recovery_cp = MaxSector;
1851
1852         spin_lock_init(&conf->resync_lock);
1853         init_waitqueue_head(&conf->wait_barrier);
1854
1855         bio_list_init(&conf->pending_bio_list);
1856         bio_list_init(&conf->flushing_bio_list);
1857
1858         if (!conf->working_disks) {
1859                 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
1860                         mdname(mddev));
1861                 goto out_free_conf;
1862         }
1863
1864         mddev->degraded = 0;
1865         for (i = 0; i < conf->raid_disks; i++) {
1866
1867                 disk = conf->mirrors + i;
1868
1869                 if (!disk->rdev) {
1870                         disk->head_position = 0;
1871                         mddev->degraded++;
1872                 }
1873         }
1874
1875         /*
1876          * find the first working one and use it as a starting point
1877          * to read balancing.
1878          */
1879         for (j = 0; j < conf->raid_disks &&
1880                      (!conf->mirrors[j].rdev ||
1881                       !test_bit(In_sync, &conf->mirrors[j].rdev->flags)) ; j++)
1882                 /* nothing */;
1883         conf->last_used = j;
1884
1885
1886         mddev->thread = md_register_thread(raid1d, mddev, "%s_raid1");
1887         if (!mddev->thread) {
1888                 printk(KERN_ERR
1889                        "raid1: couldn't allocate thread for %s\n",
1890                        mdname(mddev));
1891                 goto out_free_conf;
1892         }
1893
1894         printk(KERN_INFO 
1895                 "raid1: raid set %s active with %d out of %d mirrors\n",
1896                 mdname(mddev), mddev->raid_disks - mddev->degraded, 
1897                 mddev->raid_disks);
1898         /*
1899          * Ok, everything is just fine now
1900          */
1901         mddev->array_size = mddev->size;
1902
1903         mddev->queue->unplug_fn = raid1_unplug;
1904         mddev->queue->issue_flush_fn = raid1_issue_flush;
1905
1906         return 0;
1907
1908 out_no_mem:
1909         printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
1910                mdname(mddev));
1911
1912 out_free_conf:
1913         if (conf) {
1914                 if (conf->r1bio_pool)
1915                         mempool_destroy(conf->r1bio_pool);
1916                 kfree(conf->mirrors);
1917                 safe_put_page(conf->tmppage);
1918                 kfree(conf->poolinfo);
1919                 kfree(conf);
1920                 mddev->private = NULL;
1921         }
1922 out:
1923         return -EIO;
1924 }
1925
1926 static int stop(mddev_t *mddev)
1927 {
1928         conf_t *conf = mddev_to_conf(mddev);
1929         struct bitmap *bitmap = mddev->bitmap;
1930         int behind_wait = 0;
1931
1932         /* wait for behind writes to complete */
1933         while (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
1934                 behind_wait++;
1935                 printk(KERN_INFO "raid1: behind writes in progress on device %s, waiting to stop (%d)\n", mdname(mddev), behind_wait);
1936                 set_current_state(TASK_UNINTERRUPTIBLE);
1937                 schedule_timeout(HZ); /* wait a second */
1938                 /* need to kick something here to make sure I/O goes? */
1939         }
1940
1941         md_unregister_thread(mddev->thread);
1942         mddev->thread = NULL;
1943         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
1944         if (conf->r1bio_pool)
1945                 mempool_destroy(conf->r1bio_pool);
1946         kfree(conf->mirrors);
1947         kfree(conf->poolinfo);
1948         kfree(conf);
1949         mddev->private = NULL;
1950         return 0;
1951 }
1952
1953 static int raid1_resize(mddev_t *mddev, sector_t sectors)
1954 {
1955         /* no resync is happening, and there is enough space
1956          * on all devices, so we can resize.
1957          * We need to make sure resync covers any new space.
1958          * If the array is shrinking we should possibly wait until
1959          * any io in the removed space completes, but it hardly seems
1960          * worth it.
1961          */
1962         mddev->array_size = sectors>>1;
1963         set_capacity(mddev->gendisk, mddev->array_size << 1);
1964         mddev->changed = 1;
1965         if (mddev->array_size > mddev->size && mddev->recovery_cp == MaxSector) {
1966                 mddev->recovery_cp = mddev->size << 1;
1967                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
1968         }
1969         mddev->size = mddev->array_size;
1970         mddev->resync_max_sectors = sectors;
1971         return 0;
1972 }
1973
1974 static int raid1_reshape(mddev_t *mddev, int raid_disks)
1975 {
1976         /* We need to:
1977          * 1/ resize the r1bio_pool
1978          * 2/ resize conf->mirrors
1979          *
1980          * We allocate a new r1bio_pool if we can.
1981          * Then raise a device barrier and wait until all IO stops.
1982          * Then resize conf->mirrors and swap in the new r1bio pool.
1983          *
1984          * At the same time, we "pack" the devices so that all the missing
1985          * devices have the higher raid_disk numbers.
1986          */
1987         mempool_t *newpool, *oldpool;
1988         struct pool_info *newpoolinfo;
1989         mirror_info_t *newmirrors;
1990         conf_t *conf = mddev_to_conf(mddev);
1991         int cnt;
1992
1993         int d, d2;
1994
1995         if (raid_disks < conf->raid_disks) {
1996                 cnt=0;
1997                 for (d= 0; d < conf->raid_disks; d++)
1998                         if (conf->mirrors[d].rdev)
1999                                 cnt++;
2000                 if (cnt > raid_disks)
2001                         return -EBUSY;
2002         }
2003
2004         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2005         if (!newpoolinfo)
2006                 return -ENOMEM;
2007         newpoolinfo->mddev = mddev;
2008         newpoolinfo->raid_disks = raid_disks;
2009
2010         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2011                                  r1bio_pool_free, newpoolinfo);
2012         if (!newpool) {
2013                 kfree(newpoolinfo);
2014                 return -ENOMEM;
2015         }
2016         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2017         if (!newmirrors) {
2018                 kfree(newpoolinfo);
2019                 mempool_destroy(newpool);
2020                 return -ENOMEM;
2021         }
2022
2023         raise_barrier(conf);
2024
2025         /* ok, everything is stopped */
2026         oldpool = conf->r1bio_pool;
2027         conf->r1bio_pool = newpool;
2028
2029         for (d=d2=0; d < conf->raid_disks; d++)
2030                 if (conf->mirrors[d].rdev) {
2031                         conf->mirrors[d].rdev->raid_disk = d2;
2032                         newmirrors[d2++].rdev = conf->mirrors[d].rdev;
2033                 }
2034         kfree(conf->mirrors);
2035         conf->mirrors = newmirrors;
2036         kfree(conf->poolinfo);
2037         conf->poolinfo = newpoolinfo;
2038
2039         mddev->degraded += (raid_disks - conf->raid_disks);
2040         conf->raid_disks = mddev->raid_disks = raid_disks;
2041
2042         conf->last_used = 0; /* just make sure it is in-range */
2043         lower_barrier(conf);
2044
2045         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2046         md_wakeup_thread(mddev->thread);
2047
2048         mempool_destroy(oldpool);
2049         return 0;
2050 }
2051
2052 static void raid1_quiesce(mddev_t *mddev, int state)
2053 {
2054         conf_t *conf = mddev_to_conf(mddev);
2055
2056         switch(state) {
2057         case 1:
2058                 raise_barrier(conf);
2059                 break;
2060         case 0:
2061                 lower_barrier(conf);
2062                 break;
2063         }
2064 }
2065
2066
2067 static struct mdk_personality raid1_personality =
2068 {
2069         .name           = "raid1",
2070         .level          = 1,
2071         .owner          = THIS_MODULE,
2072         .make_request   = make_request,
2073         .run            = run,
2074         .stop           = stop,
2075         .status         = status,
2076         .error_handler  = error,
2077         .hot_add_disk   = raid1_add_disk,
2078         .hot_remove_disk= raid1_remove_disk,
2079         .spare_active   = raid1_spare_active,
2080         .sync_request   = sync_request,
2081         .resize         = raid1_resize,
2082         .reshape        = raid1_reshape,
2083         .quiesce        = raid1_quiesce,
2084 };
2085
2086 static int __init raid_init(void)
2087 {
2088         return register_md_personality(&raid1_personality);
2089 }
2090
2091 static void raid_exit(void)
2092 {
2093         unregister_md_personality(&raid1_personality);
2094 }
2095
2096 module_init(raid_init);
2097 module_exit(raid_exit);
2098 MODULE_LICENSE("GPL");
2099 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2100 MODULE_ALIAS("md-raid1");
2101 MODULE_ALIAS("md-level-1");