Merge branch 'for-2.6.39/stack-plug' into for-2.6.39/core
[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 Ă˜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 "md.h"
39 #include "raid1.h"
40 #include "bitmap.h"
41
42 #define DEBUG 0
43 #if DEBUG
44 #define PRINTK(x...) printk(x)
45 #else
46 #define PRINTK(x...)
47 #endif
48
49 /*
50  * Number of guaranteed r1bios in case of extreme VM load:
51  */
52 #define NR_RAID1_BIOS 256
53
54
55 static void allow_barrier(conf_t *conf);
56 static void lower_barrier(conf_t *conf);
57
58 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
59 {
60         struct pool_info *pi = data;
61         int size = offsetof(r1bio_t, bios[pi->raid_disks]);
62
63         /* allocate a r1bio with room for raid_disks entries in the bios array */
64         return kzalloc(size, gfp_flags);
65 }
66
67 static void r1bio_pool_free(void *r1_bio, void *data)
68 {
69         kfree(r1_bio);
70 }
71
72 #define RESYNC_BLOCK_SIZE (64*1024)
73 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
74 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
75 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
76 #define RESYNC_WINDOW (2048*1024)
77
78 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
79 {
80         struct pool_info *pi = data;
81         struct page *page;
82         r1bio_t *r1_bio;
83         struct bio *bio;
84         int i, j;
85
86         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
87         if (!r1_bio)
88                 return NULL;
89
90         /*
91          * Allocate bios : 1 for reading, n-1 for writing
92          */
93         for (j = pi->raid_disks ; j-- ; ) {
94                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
95                 if (!bio)
96                         goto out_free_bio;
97                 r1_bio->bios[j] = bio;
98         }
99         /*
100          * Allocate RESYNC_PAGES data pages and attach them to
101          * the first bio.
102          * If this is a user-requested check/repair, allocate
103          * RESYNC_PAGES for each bio.
104          */
105         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
106                 j = pi->raid_disks;
107         else
108                 j = 1;
109         while(j--) {
110                 bio = r1_bio->bios[j];
111                 for (i = 0; i < RESYNC_PAGES; i++) {
112                         page = alloc_page(gfp_flags);
113                         if (unlikely(!page))
114                                 goto out_free_pages;
115
116                         bio->bi_io_vec[i].bv_page = page;
117                         bio->bi_vcnt = i+1;
118                 }
119         }
120         /* If not user-requests, copy the page pointers to all bios */
121         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
122                 for (i=0; i<RESYNC_PAGES ; i++)
123                         for (j=1; j<pi->raid_disks; j++)
124                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
125                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
126         }
127
128         r1_bio->master_bio = NULL;
129
130         return r1_bio;
131
132 out_free_pages:
133         for (j=0 ; j < pi->raid_disks; j++)
134                 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
135                         put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
136         j = -1;
137 out_free_bio:
138         while ( ++j < pi->raid_disks )
139                 bio_put(r1_bio->bios[j]);
140         r1bio_pool_free(r1_bio, data);
141         return NULL;
142 }
143
144 static void r1buf_pool_free(void *__r1_bio, void *data)
145 {
146         struct pool_info *pi = data;
147         int i,j;
148         r1bio_t *r1bio = __r1_bio;
149
150         for (i = 0; i < RESYNC_PAGES; i++)
151                 for (j = pi->raid_disks; j-- ;) {
152                         if (j == 0 ||
153                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
154                             r1bio->bios[0]->bi_io_vec[i].bv_page)
155                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
156                 }
157         for (i=0 ; i < pi->raid_disks; i++)
158                 bio_put(r1bio->bios[i]);
159
160         r1bio_pool_free(r1bio, data);
161 }
162
163 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
164 {
165         int i;
166
167         for (i = 0; i < conf->raid_disks; i++) {
168                 struct bio **bio = r1_bio->bios + i;
169                 if (*bio && *bio != IO_BLOCKED)
170                         bio_put(*bio);
171                 *bio = NULL;
172         }
173 }
174
175 static void free_r1bio(r1bio_t *r1_bio)
176 {
177         conf_t *conf = r1_bio->mddev->private;
178
179         /*
180          * Wake up any possible resync thread that waits for the device
181          * to go idle.
182          */
183         allow_barrier(conf);
184
185         put_all_bios(conf, r1_bio);
186         mempool_free(r1_bio, conf->r1bio_pool);
187 }
188
189 static void put_buf(r1bio_t *r1_bio)
190 {
191         conf_t *conf = r1_bio->mddev->private;
192         int i;
193
194         for (i=0; i<conf->raid_disks; i++) {
195                 struct bio *bio = r1_bio->bios[i];
196                 if (bio->bi_end_io)
197                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
198         }
199
200         mempool_free(r1_bio, conf->r1buf_pool);
201
202         lower_barrier(conf);
203 }
204
205 static void reschedule_retry(r1bio_t *r1_bio)
206 {
207         unsigned long flags;
208         mddev_t *mddev = r1_bio->mddev;
209         conf_t *conf = mddev->private;
210
211         spin_lock_irqsave(&conf->device_lock, flags);
212         list_add(&r1_bio->retry_list, &conf->retry_list);
213         conf->nr_queued ++;
214         spin_unlock_irqrestore(&conf->device_lock, flags);
215
216         wake_up(&conf->wait_barrier);
217         md_wakeup_thread(mddev->thread);
218 }
219
220 /*
221  * raid_end_bio_io() is called when we have finished servicing a mirrored
222  * operation and are ready to return a success/failure code to the buffer
223  * cache layer.
224  */
225 static void raid_end_bio_io(r1bio_t *r1_bio)
226 {
227         struct bio *bio = r1_bio->master_bio;
228
229         /* if nobody has done the final endio yet, do it now */
230         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
231                 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
232                         (bio_data_dir(bio) == WRITE) ? "write" : "read",
233                         (unsigned long long) bio->bi_sector,
234                         (unsigned long long) bio->bi_sector +
235                                 (bio->bi_size >> 9) - 1);
236
237                 bio_endio(bio,
238                         test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
239         }
240         free_r1bio(r1_bio);
241 }
242
243 /*
244  * Update disk head position estimator based on IRQ completion info.
245  */
246 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
247 {
248         conf_t *conf = r1_bio->mddev->private;
249
250         conf->mirrors[disk].head_position =
251                 r1_bio->sector + (r1_bio->sectors);
252 }
253
254 static void raid1_end_read_request(struct bio *bio, int error)
255 {
256         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
257         r1bio_t *r1_bio = bio->bi_private;
258         int mirror;
259         conf_t *conf = r1_bio->mddev->private;
260
261         mirror = r1_bio->read_disk;
262         /*
263          * this branch is our 'one mirror IO has finished' event handler:
264          */
265         update_head_pos(mirror, r1_bio);
266
267         if (uptodate)
268                 set_bit(R1BIO_Uptodate, &r1_bio->state);
269         else {
270                 /* If all other devices have failed, we want to return
271                  * the error upwards rather than fail the last device.
272                  * Here we redefine "uptodate" to mean "Don't want to retry"
273                  */
274                 unsigned long flags;
275                 spin_lock_irqsave(&conf->device_lock, flags);
276                 if (r1_bio->mddev->degraded == conf->raid_disks ||
277                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
278                      !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
279                         uptodate = 1;
280                 spin_unlock_irqrestore(&conf->device_lock, flags);
281         }
282
283         if (uptodate)
284                 raid_end_bio_io(r1_bio);
285         else {
286                 /*
287                  * oops, read error:
288                  */
289                 char b[BDEVNAME_SIZE];
290                 if (printk_ratelimit())
291                         printk(KERN_ERR "md/raid1:%s: %s: rescheduling sector %llu\n",
292                                mdname(conf->mddev),
293                                bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
294                 reschedule_retry(r1_bio);
295         }
296
297         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
298 }
299
300 static void r1_bio_write_done(r1bio_t *r1_bio, int vcnt, struct bio_vec *bv,
301                               int behind)
302 {
303         if (atomic_dec_and_test(&r1_bio->remaining))
304         {
305                 /* it really is the end of this request */
306                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
307                         /* free extra copy of the data pages */
308                         int i = vcnt;
309                         while (i--)
310                                 safe_put_page(bv[i].bv_page);
311                 }
312                 /* clear the bitmap if all writes complete successfully */
313                 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
314                                 r1_bio->sectors,
315                                 !test_bit(R1BIO_Degraded, &r1_bio->state),
316                                 behind);
317                 md_write_end(r1_bio->mddev);
318                 raid_end_bio_io(r1_bio);
319         }
320 }
321
322 static void raid1_end_write_request(struct bio *bio, int error)
323 {
324         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
325         r1bio_t *r1_bio = bio->bi_private;
326         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
327         conf_t *conf = r1_bio->mddev->private;
328         struct bio *to_put = NULL;
329
330
331         for (mirror = 0; mirror < conf->raid_disks; mirror++)
332                 if (r1_bio->bios[mirror] == bio)
333                         break;
334
335         /*
336          * 'one mirror IO has finished' event handler:
337          */
338         r1_bio->bios[mirror] = NULL;
339         to_put = bio;
340         if (!uptodate) {
341                 md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
342                 /* an I/O failed, we can't clear the bitmap */
343                 set_bit(R1BIO_Degraded, &r1_bio->state);
344         } else
345                 /*
346                  * Set R1BIO_Uptodate in our master bio, so that we
347                  * will return a good error code for to the higher
348                  * levels even if IO on some other mirrored buffer
349                  * fails.
350                  *
351                  * The 'master' represents the composite IO operation
352                  * to user-side. So if something waits for IO, then it
353                  * will wait for the 'master' bio.
354                  */
355                 set_bit(R1BIO_Uptodate, &r1_bio->state);
356
357         update_head_pos(mirror, r1_bio);
358
359         if (behind) {
360                 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
361                         atomic_dec(&r1_bio->behind_remaining);
362
363                 /*
364                  * In behind mode, we ACK the master bio once the I/O
365                  * has safely reached all non-writemostly
366                  * disks. Setting the Returned bit ensures that this
367                  * gets done only once -- we don't ever want to return
368                  * -EIO here, instead we'll wait
369                  */
370                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
371                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
372                         /* Maybe we can return now */
373                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
374                                 struct bio *mbio = r1_bio->master_bio;
375                                 PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
376                                        (unsigned long long) mbio->bi_sector,
377                                        (unsigned long long) mbio->bi_sector +
378                                        (mbio->bi_size >> 9) - 1);
379                                 bio_endio(mbio, 0);
380                         }
381                 }
382         }
383         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
384
385         /*
386          * Let's see if all mirrored write operations have finished
387          * already.
388          */
389         r1_bio_write_done(r1_bio, bio->bi_vcnt, bio->bi_io_vec, behind);
390
391         if (to_put)
392                 bio_put(to_put);
393 }
394
395
396 /*
397  * This routine returns the disk from which the requested read should
398  * be done. There is a per-array 'next expected sequential IO' sector
399  * number - if this matches on the next IO then we use the last disk.
400  * There is also a per-disk 'last know head position' sector that is
401  * maintained from IRQ contexts, both the normal and the resync IO
402  * completion handlers update this position correctly. If there is no
403  * perfect sequential match then we pick the disk whose head is closest.
404  *
405  * If there are 2 mirrors in the same 2 devices, performance degrades
406  * because position is mirror, not device based.
407  *
408  * The rdev for the device selected will have nr_pending incremented.
409  */
410 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
411 {
412         const sector_t this_sector = r1_bio->sector;
413         const int sectors = r1_bio->sectors;
414         int new_disk = -1;
415         int start_disk;
416         int i;
417         sector_t new_distance, current_distance;
418         mdk_rdev_t *rdev;
419         int choose_first;
420
421         rcu_read_lock();
422         /*
423          * Check if we can balance. We can balance on the whole
424          * device if no resync is going on, or below the resync window.
425          * We take the first readable disk when above the resync window.
426          */
427  retry:
428         if (conf->mddev->recovery_cp < MaxSector &&
429             (this_sector + sectors >= conf->next_resync)) {
430                 choose_first = 1;
431                 start_disk = 0;
432         } else {
433                 choose_first = 0;
434                 start_disk = conf->last_used;
435         }
436
437         /* make sure the disk is operational */
438         for (i = 0 ; i < conf->raid_disks ; i++) {
439                 int disk = start_disk + i;
440                 if (disk >= conf->raid_disks)
441                         disk -= conf->raid_disks;
442
443                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
444                 if (r1_bio->bios[disk] == IO_BLOCKED
445                     || rdev == NULL
446                     || !test_bit(In_sync, &rdev->flags))
447                         continue;
448
449                 new_disk = disk;
450                 if (!test_bit(WriteMostly, &rdev->flags))
451                         break;
452         }
453
454         if (new_disk < 0 || choose_first)
455                 goto rb_out;
456
457         /*
458          * Don't change to another disk for sequential reads:
459          */
460         if (conf->next_seq_sect == this_sector)
461                 goto rb_out;
462         if (this_sector == conf->mirrors[new_disk].head_position)
463                 goto rb_out;
464
465         current_distance = abs(this_sector 
466                                - conf->mirrors[new_disk].head_position);
467
468         /* look for a better disk - i.e. head is closer */
469         start_disk = new_disk;
470         for (i = 1; i < conf->raid_disks; i++) {
471                 int disk = start_disk + 1;
472                 if (disk >= conf->raid_disks)
473                         disk -= conf->raid_disks;
474
475                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
476                 if (r1_bio->bios[disk] == IO_BLOCKED
477                     || rdev == NULL
478                     || !test_bit(In_sync, &rdev->flags)
479                     || test_bit(WriteMostly, &rdev->flags))
480                         continue;
481
482                 if (!atomic_read(&rdev->nr_pending)) {
483                         new_disk = disk;
484                         break;
485                 }
486                 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
487                 if (new_distance < current_distance) {
488                         current_distance = new_distance;
489                         new_disk = disk;
490                 }
491         }
492
493  rb_out:
494         if (new_disk >= 0) {
495                 rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
496                 if (!rdev)
497                         goto retry;
498                 atomic_inc(&rdev->nr_pending);
499                 if (!test_bit(In_sync, &rdev->flags)) {
500                         /* cannot risk returning a device that failed
501                          * before we inc'ed nr_pending
502                          */
503                         rdev_dec_pending(rdev, conf->mddev);
504                         goto retry;
505                 }
506                 conf->next_seq_sect = this_sector + sectors;
507                 conf->last_used = new_disk;
508         }
509         rcu_read_unlock();
510
511         return new_disk;
512 }
513
514 static int raid1_congested(void *data, int bits)
515 {
516         mddev_t *mddev = data;
517         conf_t *conf = mddev->private;
518         int i, ret = 0;
519
520         if (mddev_congested(mddev, bits))
521                 return 1;
522
523         rcu_read_lock();
524         for (i = 0; i < mddev->raid_disks; i++) {
525                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
526                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
527                         struct request_queue *q = bdev_get_queue(rdev->bdev);
528
529                         /* Note the '|| 1' - when read_balance prefers
530                          * non-congested targets, it can be removed
531                          */
532                         if ((bits & (1<<BDI_async_congested)) || 1)
533                                 ret |= bdi_congested(&q->backing_dev_info, bits);
534                         else
535                                 ret &= bdi_congested(&q->backing_dev_info, bits);
536                 }
537         }
538         rcu_read_unlock();
539         return ret;
540 }
541
542
543 static void flush_pending_writes(conf_t *conf)
544 {
545         /* Any writes that have been queued but are awaiting
546          * bitmap updates get flushed here.
547          */
548         spin_lock_irq(&conf->device_lock);
549
550         if (conf->pending_bio_list.head) {
551                 struct bio *bio;
552                 bio = bio_list_get(&conf->pending_bio_list);
553                 spin_unlock_irq(&conf->device_lock);
554                 /* flush any pending bitmap writes to
555                  * disk before proceeding w/ I/O */
556                 bitmap_unplug(conf->mddev->bitmap);
557
558                 while (bio) { /* submit pending writes */
559                         struct bio *next = bio->bi_next;
560                         bio->bi_next = NULL;
561                         generic_make_request(bio);
562                         bio = next;
563                 }
564         } else
565                 spin_unlock_irq(&conf->device_lock);
566 }
567
568 static void md_kick_device(mddev_t *mddev)
569 {
570         blk_flush_plug(current);
571         md_wakeup_thread(mddev->thread);
572 }
573
574 /* Barriers....
575  * Sometimes we need to suspend IO while we do something else,
576  * either some resync/recovery, or reconfigure the array.
577  * To do this we raise a 'barrier'.
578  * The 'barrier' is a counter that can be raised multiple times
579  * to count how many activities are happening which preclude
580  * normal IO.
581  * We can only raise the barrier if there is no pending IO.
582  * i.e. if nr_pending == 0.
583  * We choose only to raise the barrier if no-one is waiting for the
584  * barrier to go down.  This means that as soon as an IO request
585  * is ready, no other operations which require a barrier will start
586  * until the IO request has had a chance.
587  *
588  * So: regular IO calls 'wait_barrier'.  When that returns there
589  *    is no backgroup IO happening,  It must arrange to call
590  *    allow_barrier when it has finished its IO.
591  * backgroup IO calls must call raise_barrier.  Once that returns
592  *    there is no normal IO happeing.  It must arrange to call
593  *    lower_barrier when the particular background IO completes.
594  */
595 #define RESYNC_DEPTH 32
596
597 static void raise_barrier(conf_t *conf)
598 {
599         spin_lock_irq(&conf->resync_lock);
600
601         /* Wait until no block IO is waiting */
602         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
603                             conf->resync_lock, md_kick_device(conf->mddev));
604
605         /* block any new IO from starting */
606         conf->barrier++;
607
608         /* Now wait for all pending IO to complete */
609         wait_event_lock_irq(conf->wait_barrier,
610                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
611                             conf->resync_lock, md_kick_device(conf->mddev));
612
613         spin_unlock_irq(&conf->resync_lock);
614 }
615
616 static void lower_barrier(conf_t *conf)
617 {
618         unsigned long flags;
619         BUG_ON(conf->barrier <= 0);
620         spin_lock_irqsave(&conf->resync_lock, flags);
621         conf->barrier--;
622         spin_unlock_irqrestore(&conf->resync_lock, flags);
623         wake_up(&conf->wait_barrier);
624 }
625
626 static void wait_barrier(conf_t *conf)
627 {
628         spin_lock_irq(&conf->resync_lock);
629         if (conf->barrier) {
630                 conf->nr_waiting++;
631                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
632                                     conf->resync_lock,
633                                     md_kick_device(conf->mddev));
634                 conf->nr_waiting--;
635         }
636         conf->nr_pending++;
637         spin_unlock_irq(&conf->resync_lock);
638 }
639
640 static void allow_barrier(conf_t *conf)
641 {
642         unsigned long flags;
643         spin_lock_irqsave(&conf->resync_lock, flags);
644         conf->nr_pending--;
645         spin_unlock_irqrestore(&conf->resync_lock, flags);
646         wake_up(&conf->wait_barrier);
647 }
648
649 static void freeze_array(conf_t *conf)
650 {
651         /* stop syncio and normal IO and wait for everything to
652          * go quite.
653          * We increment barrier and nr_waiting, and then
654          * wait until nr_pending match nr_queued+1
655          * This is called in the context of one normal IO request
656          * that has failed. Thus any sync request that might be pending
657          * will be blocked by nr_pending, and we need to wait for
658          * pending IO requests to complete or be queued for re-try.
659          * Thus the number queued (nr_queued) plus this request (1)
660          * must match the number of pending IOs (nr_pending) before
661          * we continue.
662          */
663         spin_lock_irq(&conf->resync_lock);
664         conf->barrier++;
665         conf->nr_waiting++;
666         wait_event_lock_irq(conf->wait_barrier,
667                             conf->nr_pending == conf->nr_queued+1,
668                             conf->resync_lock,
669                             ({ flush_pending_writes(conf);
670                                md_kick_device(conf->mddev); }));
671         spin_unlock_irq(&conf->resync_lock);
672 }
673 static void unfreeze_array(conf_t *conf)
674 {
675         /* reverse the effect of the freeze */
676         spin_lock_irq(&conf->resync_lock);
677         conf->barrier--;
678         conf->nr_waiting--;
679         wake_up(&conf->wait_barrier);
680         spin_unlock_irq(&conf->resync_lock);
681 }
682
683
684 /* duplicate the data pages for behind I/O 
685  * We return a list of bio_vec rather than just page pointers
686  * as it makes freeing easier
687  */
688 static struct bio_vec *alloc_behind_pages(struct bio *bio)
689 {
690         int i;
691         struct bio_vec *bvec;
692         struct bio_vec *pages = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
693                                         GFP_NOIO);
694         if (unlikely(!pages))
695                 goto do_sync_io;
696
697         bio_for_each_segment(bvec, bio, i) {
698                 pages[i].bv_page = alloc_page(GFP_NOIO);
699                 if (unlikely(!pages[i].bv_page))
700                         goto do_sync_io;
701                 memcpy(kmap(pages[i].bv_page) + bvec->bv_offset,
702                         kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
703                 kunmap(pages[i].bv_page);
704                 kunmap(bvec->bv_page);
705         }
706
707         return pages;
708
709 do_sync_io:
710         if (pages)
711                 for (i = 0; i < bio->bi_vcnt && pages[i].bv_page; i++)
712                         put_page(pages[i].bv_page);
713         kfree(pages);
714         PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
715         return NULL;
716 }
717
718 static int make_request(mddev_t *mddev, struct bio * bio)
719 {
720         conf_t *conf = mddev->private;
721         mirror_info_t *mirror;
722         r1bio_t *r1_bio;
723         struct bio *read_bio;
724         int i, targets = 0, disks;
725         struct bitmap *bitmap;
726         unsigned long flags;
727         struct bio_vec *behind_pages = NULL;
728         const int rw = bio_data_dir(bio);
729         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
730         const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
731         mdk_rdev_t *blocked_rdev;
732
733         /*
734          * Register the new request and wait if the reconstruction
735          * thread has put up a bar for new requests.
736          * Continue immediately if no resync is active currently.
737          */
738
739         md_write_start(mddev, bio); /* wait on superblock update early */
740
741         if (bio_data_dir(bio) == WRITE &&
742             bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
743             bio->bi_sector < mddev->suspend_hi) {
744                 /* As the suspend_* range is controlled by
745                  * userspace, we want an interruptible
746                  * wait.
747                  */
748                 DEFINE_WAIT(w);
749                 for (;;) {
750                         flush_signals(current);
751                         prepare_to_wait(&conf->wait_barrier,
752                                         &w, TASK_INTERRUPTIBLE);
753                         if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
754                             bio->bi_sector >= mddev->suspend_hi)
755                                 break;
756                         schedule();
757                 }
758                 finish_wait(&conf->wait_barrier, &w);
759         }
760
761         wait_barrier(conf);
762
763         bitmap = mddev->bitmap;
764
765         /*
766          * make_request() can abort the operation when READA is being
767          * used and no empty request is available.
768          *
769          */
770         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
771
772         r1_bio->master_bio = bio;
773         r1_bio->sectors = bio->bi_size >> 9;
774         r1_bio->state = 0;
775         r1_bio->mddev = mddev;
776         r1_bio->sector = bio->bi_sector;
777
778         if (rw == READ) {
779                 /*
780                  * read balancing logic:
781                  */
782                 int rdisk = read_balance(conf, r1_bio);
783
784                 if (rdisk < 0) {
785                         /* couldn't find anywhere to read from */
786                         raid_end_bio_io(r1_bio);
787                         return 0;
788                 }
789                 mirror = conf->mirrors + rdisk;
790
791                 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
792                     bitmap) {
793                         /* Reading from a write-mostly device must
794                          * take care not to over-take any writes
795                          * that are 'behind'
796                          */
797                         wait_event(bitmap->behind_wait,
798                                    atomic_read(&bitmap->behind_writes) == 0);
799                 }
800                 r1_bio->read_disk = rdisk;
801
802                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
803
804                 r1_bio->bios[rdisk] = read_bio;
805
806                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
807                 read_bio->bi_bdev = mirror->rdev->bdev;
808                 read_bio->bi_end_io = raid1_end_read_request;
809                 read_bio->bi_rw = READ | do_sync;
810                 read_bio->bi_private = r1_bio;
811
812                 generic_make_request(read_bio);
813                 return 0;
814         }
815
816         /*
817          * WRITE:
818          */
819         /* first select target devices under spinlock and
820          * inc refcount on their rdev.  Record them by setting
821          * bios[x] to bio
822          */
823         disks = conf->raid_disks;
824  retry_write:
825         blocked_rdev = NULL;
826         rcu_read_lock();
827         for (i = 0;  i < disks; i++) {
828                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
829                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
830                         atomic_inc(&rdev->nr_pending);
831                         blocked_rdev = rdev;
832                         break;
833                 }
834                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
835                         atomic_inc(&rdev->nr_pending);
836                         if (test_bit(Faulty, &rdev->flags)) {
837                                 rdev_dec_pending(rdev, mddev);
838                                 r1_bio->bios[i] = NULL;
839                         } else {
840                                 r1_bio->bios[i] = bio;
841                                 targets++;
842                         }
843                 } else
844                         r1_bio->bios[i] = NULL;
845         }
846         rcu_read_unlock();
847
848         if (unlikely(blocked_rdev)) {
849                 /* Wait for this device to become unblocked */
850                 int j;
851
852                 for (j = 0; j < i; j++)
853                         if (r1_bio->bios[j])
854                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
855
856                 allow_barrier(conf);
857                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
858                 wait_barrier(conf);
859                 goto retry_write;
860         }
861
862         BUG_ON(targets == 0); /* we never fail the last device */
863
864         if (targets < conf->raid_disks) {
865                 /* array is degraded, we will not clear the bitmap
866                  * on I/O completion (see raid1_end_write_request) */
867                 set_bit(R1BIO_Degraded, &r1_bio->state);
868         }
869
870         /* do behind I/O ?
871          * Not if there are too many, or cannot allocate memory,
872          * or a reader on WriteMostly is waiting for behind writes 
873          * to flush */
874         if (bitmap &&
875             (atomic_read(&bitmap->behind_writes)
876              < mddev->bitmap_info.max_write_behind) &&
877             !waitqueue_active(&bitmap->behind_wait) &&
878             (behind_pages = alloc_behind_pages(bio)) != NULL)
879                 set_bit(R1BIO_BehindIO, &r1_bio->state);
880
881         atomic_set(&r1_bio->remaining, 1);
882         atomic_set(&r1_bio->behind_remaining, 0);
883
884         bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
885                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
886         for (i = 0; i < disks; i++) {
887                 struct bio *mbio;
888                 if (!r1_bio->bios[i])
889                         continue;
890
891                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
892                 r1_bio->bios[i] = mbio;
893
894                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
895                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
896                 mbio->bi_end_io = raid1_end_write_request;
897                 mbio->bi_rw = WRITE | do_flush_fua | do_sync;
898                 mbio->bi_private = r1_bio;
899
900                 if (behind_pages) {
901                         struct bio_vec *bvec;
902                         int j;
903
904                         /* Yes, I really want the '__' version so that
905                          * we clear any unused pointer in the io_vec, rather
906                          * than leave them unchanged.  This is important
907                          * because when we come to free the pages, we won't
908                          * know the original bi_idx, so we just free
909                          * them all
910                          */
911                         __bio_for_each_segment(bvec, mbio, j, 0)
912                                 bvec->bv_page = behind_pages[j].bv_page;
913                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
914                                 atomic_inc(&r1_bio->behind_remaining);
915                 }
916
917                 atomic_inc(&r1_bio->remaining);
918                 spin_lock_irqsave(&conf->device_lock, flags);
919                 bio_list_add(&conf->pending_bio_list, mbio);
920                 spin_unlock_irqrestore(&conf->device_lock, flags);
921         }
922         r1_bio_write_done(r1_bio, bio->bi_vcnt, behind_pages, behind_pages != NULL);
923         kfree(behind_pages); /* the behind pages are attached to the bios now */
924
925         /* In case raid1d snuck in to freeze_array */
926         wake_up(&conf->wait_barrier);
927
928         if (do_sync || !bitmap)
929                 md_wakeup_thread(mddev->thread);
930
931         return 0;
932 }
933
934 static void status(struct seq_file *seq, mddev_t *mddev)
935 {
936         conf_t *conf = mddev->private;
937         int i;
938
939         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
940                    conf->raid_disks - mddev->degraded);
941         rcu_read_lock();
942         for (i = 0; i < conf->raid_disks; i++) {
943                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
944                 seq_printf(seq, "%s",
945                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
946         }
947         rcu_read_unlock();
948         seq_printf(seq, "]");
949 }
950
951
952 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
953 {
954         char b[BDEVNAME_SIZE];
955         conf_t *conf = mddev->private;
956
957         /*
958          * If it is not operational, then we have already marked it as dead
959          * else if it is the last working disks, ignore the error, let the
960          * next level up know.
961          * else mark the drive as failed
962          */
963         if (test_bit(In_sync, &rdev->flags)
964             && (conf->raid_disks - mddev->degraded) == 1) {
965                 /*
966                  * Don't fail the drive, act as though we were just a
967                  * normal single drive.
968                  * However don't try a recovery from this drive as
969                  * it is very likely to fail.
970                  */
971                 mddev->recovery_disabled = 1;
972                 return;
973         }
974         if (test_and_clear_bit(In_sync, &rdev->flags)) {
975                 unsigned long flags;
976                 spin_lock_irqsave(&conf->device_lock, flags);
977                 mddev->degraded++;
978                 set_bit(Faulty, &rdev->flags);
979                 spin_unlock_irqrestore(&conf->device_lock, flags);
980                 /*
981                  * if recovery is running, make sure it aborts.
982                  */
983                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
984         } else
985                 set_bit(Faulty, &rdev->flags);
986         set_bit(MD_CHANGE_DEVS, &mddev->flags);
987         printk(KERN_ALERT
988                "md/raid1:%s: Disk failure on %s, disabling device.\n"
989                "md/raid1:%s: Operation continuing on %d devices.\n",
990                mdname(mddev), bdevname(rdev->bdev, b),
991                mdname(mddev), conf->raid_disks - mddev->degraded);
992 }
993
994 static void print_conf(conf_t *conf)
995 {
996         int i;
997
998         printk(KERN_DEBUG "RAID1 conf printout:\n");
999         if (!conf) {
1000                 printk(KERN_DEBUG "(!conf)\n");
1001                 return;
1002         }
1003         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1004                 conf->raid_disks);
1005
1006         rcu_read_lock();
1007         for (i = 0; i < conf->raid_disks; i++) {
1008                 char b[BDEVNAME_SIZE];
1009                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1010                 if (rdev)
1011                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1012                                i, !test_bit(In_sync, &rdev->flags),
1013                                !test_bit(Faulty, &rdev->flags),
1014                                bdevname(rdev->bdev,b));
1015         }
1016         rcu_read_unlock();
1017 }
1018
1019 static void close_sync(conf_t *conf)
1020 {
1021         wait_barrier(conf);
1022         allow_barrier(conf);
1023
1024         mempool_destroy(conf->r1buf_pool);
1025         conf->r1buf_pool = NULL;
1026 }
1027
1028 static int raid1_spare_active(mddev_t *mddev)
1029 {
1030         int i;
1031         conf_t *conf = mddev->private;
1032         int count = 0;
1033         unsigned long flags;
1034
1035         /*
1036          * Find all failed disks within the RAID1 configuration 
1037          * and mark them readable.
1038          * Called under mddev lock, so rcu protection not needed.
1039          */
1040         for (i = 0; i < conf->raid_disks; i++) {
1041                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1042                 if (rdev
1043                     && !test_bit(Faulty, &rdev->flags)
1044                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1045                         count++;
1046                         sysfs_notify_dirent(rdev->sysfs_state);
1047                 }
1048         }
1049         spin_lock_irqsave(&conf->device_lock, flags);
1050         mddev->degraded -= count;
1051         spin_unlock_irqrestore(&conf->device_lock, flags);
1052
1053         print_conf(conf);
1054         return count;
1055 }
1056
1057
1058 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1059 {
1060         conf_t *conf = mddev->private;
1061         int err = -EEXIST;
1062         int mirror = 0;
1063         mirror_info_t *p;
1064         int first = 0;
1065         int last = mddev->raid_disks - 1;
1066
1067         if (rdev->raid_disk >= 0)
1068                 first = last = rdev->raid_disk;
1069
1070         for (mirror = first; mirror <= last; mirror++)
1071                 if ( !(p=conf->mirrors+mirror)->rdev) {
1072
1073                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1074                                           rdev->data_offset << 9);
1075                         /* as we don't honour merge_bvec_fn, we must
1076                          * never risk violating it, so limit
1077                          * ->max_segments to one lying with a single
1078                          * page, as a one page request is never in
1079                          * violation.
1080                          */
1081                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1082                                 blk_queue_max_segments(mddev->queue, 1);
1083                                 blk_queue_segment_boundary(mddev->queue,
1084                                                            PAGE_CACHE_SIZE - 1);
1085                         }
1086
1087                         p->head_position = 0;
1088                         rdev->raid_disk = mirror;
1089                         err = 0;
1090                         /* As all devices are equivalent, we don't need a full recovery
1091                          * if this was recently any drive of the array
1092                          */
1093                         if (rdev->saved_raid_disk < 0)
1094                                 conf->fullsync = 1;
1095                         rcu_assign_pointer(p->rdev, rdev);
1096                         break;
1097                 }
1098         md_integrity_add_rdev(rdev, mddev);
1099         print_conf(conf);
1100         return err;
1101 }
1102
1103 static int raid1_remove_disk(mddev_t *mddev, int number)
1104 {
1105         conf_t *conf = mddev->private;
1106         int err = 0;
1107         mdk_rdev_t *rdev;
1108         mirror_info_t *p = conf->mirrors+ number;
1109
1110         print_conf(conf);
1111         rdev = p->rdev;
1112         if (rdev) {
1113                 if (test_bit(In_sync, &rdev->flags) ||
1114                     atomic_read(&rdev->nr_pending)) {
1115                         err = -EBUSY;
1116                         goto abort;
1117                 }
1118                 /* Only remove non-faulty devices if recovery
1119                  * is not possible.
1120                  */
1121                 if (!test_bit(Faulty, &rdev->flags) &&
1122                     !mddev->recovery_disabled &&
1123                     mddev->degraded < conf->raid_disks) {
1124                         err = -EBUSY;
1125                         goto abort;
1126                 }
1127                 p->rdev = NULL;
1128                 synchronize_rcu();
1129                 if (atomic_read(&rdev->nr_pending)) {
1130                         /* lost the race, try later */
1131                         err = -EBUSY;
1132                         p->rdev = rdev;
1133                         goto abort;
1134                 }
1135                 md_integrity_register(mddev);
1136         }
1137 abort:
1138
1139         print_conf(conf);
1140         return err;
1141 }
1142
1143
1144 static void end_sync_read(struct bio *bio, int error)
1145 {
1146         r1bio_t *r1_bio = bio->bi_private;
1147         int i;
1148
1149         for (i=r1_bio->mddev->raid_disks; i--; )
1150                 if (r1_bio->bios[i] == bio)
1151                         break;
1152         BUG_ON(i < 0);
1153         update_head_pos(i, r1_bio);
1154         /*
1155          * we have read a block, now it needs to be re-written,
1156          * or re-read if the read failed.
1157          * We don't do much here, just schedule handling by raid1d
1158          */
1159         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1160                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1161
1162         if (atomic_dec_and_test(&r1_bio->remaining))
1163                 reschedule_retry(r1_bio);
1164 }
1165
1166 static void end_sync_write(struct bio *bio, int error)
1167 {
1168         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1169         r1bio_t *r1_bio = bio->bi_private;
1170         mddev_t *mddev = r1_bio->mddev;
1171         conf_t *conf = mddev->private;
1172         int i;
1173         int mirror=0;
1174
1175         for (i = 0; i < conf->raid_disks; i++)
1176                 if (r1_bio->bios[i] == bio) {
1177                         mirror = i;
1178                         break;
1179                 }
1180         if (!uptodate) {
1181                 sector_t sync_blocks = 0;
1182                 sector_t s = r1_bio->sector;
1183                 long sectors_to_go = r1_bio->sectors;
1184                 /* make sure these bits doesn't get cleared. */
1185                 do {
1186                         bitmap_end_sync(mddev->bitmap, s,
1187                                         &sync_blocks, 1);
1188                         s += sync_blocks;
1189                         sectors_to_go -= sync_blocks;
1190                 } while (sectors_to_go > 0);
1191                 md_error(mddev, conf->mirrors[mirror].rdev);
1192         }
1193
1194         update_head_pos(mirror, r1_bio);
1195
1196         if (atomic_dec_and_test(&r1_bio->remaining)) {
1197                 sector_t s = r1_bio->sectors;
1198                 put_buf(r1_bio);
1199                 md_done_sync(mddev, s, uptodate);
1200         }
1201 }
1202
1203 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1204 {
1205         conf_t *conf = mddev->private;
1206         int i;
1207         int disks = conf->raid_disks;
1208         struct bio *bio, *wbio;
1209
1210         bio = r1_bio->bios[r1_bio->read_disk];
1211
1212
1213         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1214                 /* We have read all readable devices.  If we haven't
1215                  * got the block, then there is no hope left.
1216                  * If we have, then we want to do a comparison
1217                  * and skip the write if everything is the same.
1218                  * If any blocks failed to read, then we need to
1219                  * attempt an over-write
1220                  */
1221                 int primary;
1222                 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1223                         for (i=0; i<mddev->raid_disks; i++)
1224                                 if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1225                                         md_error(mddev, conf->mirrors[i].rdev);
1226
1227                         md_done_sync(mddev, r1_bio->sectors, 1);
1228                         put_buf(r1_bio);
1229                         return;
1230                 }
1231                 for (primary=0; primary<mddev->raid_disks; primary++)
1232                         if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1233                             test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1234                                 r1_bio->bios[primary]->bi_end_io = NULL;
1235                                 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1236                                 break;
1237                         }
1238                 r1_bio->read_disk = primary;
1239                 for (i=0; i<mddev->raid_disks; i++)
1240                         if (r1_bio->bios[i]->bi_end_io == end_sync_read) {
1241                                 int j;
1242                                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1243                                 struct bio *pbio = r1_bio->bios[primary];
1244                                 struct bio *sbio = r1_bio->bios[i];
1245
1246                                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1247                                         for (j = vcnt; j-- ; ) {
1248                                                 struct page *p, *s;
1249                                                 p = pbio->bi_io_vec[j].bv_page;
1250                                                 s = sbio->bi_io_vec[j].bv_page;
1251                                                 if (memcmp(page_address(p),
1252                                                            page_address(s),
1253                                                            PAGE_SIZE))
1254                                                         break;
1255                                         }
1256                                 } else
1257                                         j = 0;
1258                                 if (j >= 0)
1259                                         mddev->resync_mismatches += r1_bio->sectors;
1260                                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1261                                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1262                                         sbio->bi_end_io = NULL;
1263                                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1264                                 } else {
1265                                         /* fixup the bio for reuse */
1266                                         int size;
1267                                         sbio->bi_vcnt = vcnt;
1268                                         sbio->bi_size = r1_bio->sectors << 9;
1269                                         sbio->bi_idx = 0;
1270                                         sbio->bi_phys_segments = 0;
1271                                         sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1272                                         sbio->bi_flags |= 1 << BIO_UPTODATE;
1273                                         sbio->bi_next = NULL;
1274                                         sbio->bi_sector = r1_bio->sector +
1275                                                 conf->mirrors[i].rdev->data_offset;
1276                                         sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1277                                         size = sbio->bi_size;
1278                                         for (j = 0; j < vcnt ; j++) {
1279                                                 struct bio_vec *bi;
1280                                                 bi = &sbio->bi_io_vec[j];
1281                                                 bi->bv_offset = 0;
1282                                                 if (size > PAGE_SIZE)
1283                                                         bi->bv_len = PAGE_SIZE;
1284                                                 else
1285                                                         bi->bv_len = size;
1286                                                 size -= PAGE_SIZE;
1287                                                 memcpy(page_address(bi->bv_page),
1288                                                        page_address(pbio->bi_io_vec[j].bv_page),
1289                                                        PAGE_SIZE);
1290                                         }
1291
1292                                 }
1293                         }
1294         }
1295         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1296                 /* ouch - failed to read all of that.
1297                  * Try some synchronous reads of other devices to get
1298                  * good data, much like with normal read errors.  Only
1299                  * read into the pages we already have so we don't
1300                  * need to re-issue the read request.
1301                  * We don't need to freeze the array, because being in an
1302                  * active sync request, there is no normal IO, and
1303                  * no overlapping syncs.
1304                  */
1305                 sector_t sect = r1_bio->sector;
1306                 int sectors = r1_bio->sectors;
1307                 int idx = 0;
1308
1309                 while(sectors) {
1310                         int s = sectors;
1311                         int d = r1_bio->read_disk;
1312                         int success = 0;
1313                         mdk_rdev_t *rdev;
1314
1315                         if (s > (PAGE_SIZE>>9))
1316                                 s = PAGE_SIZE >> 9;
1317                         do {
1318                                 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1319                                         /* No rcu protection needed here devices
1320                                          * can only be removed when no resync is
1321                                          * active, and resync is currently active
1322                                          */
1323                                         rdev = conf->mirrors[d].rdev;
1324                                         if (sync_page_io(rdev,
1325                                                          sect,
1326                                                          s<<9,
1327                                                          bio->bi_io_vec[idx].bv_page,
1328                                                          READ, false)) {
1329                                                 success = 1;
1330                                                 break;
1331                                         }
1332                                 }
1333                                 d++;
1334                                 if (d == conf->raid_disks)
1335                                         d = 0;
1336                         } while (!success && d != r1_bio->read_disk);
1337
1338                         if (success) {
1339                                 int start = d;
1340                                 /* write it back and re-read */
1341                                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1342                                 while (d != r1_bio->read_disk) {
1343                                         if (d == 0)
1344                                                 d = conf->raid_disks;
1345                                         d--;
1346                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1347                                                 continue;
1348                                         rdev = conf->mirrors[d].rdev;
1349                                         atomic_add(s, &rdev->corrected_errors);
1350                                         if (sync_page_io(rdev,
1351                                                          sect,
1352                                                          s<<9,
1353                                                          bio->bi_io_vec[idx].bv_page,
1354                                                          WRITE, false) == 0)
1355                                                 md_error(mddev, rdev);
1356                                 }
1357                                 d = start;
1358                                 while (d != r1_bio->read_disk) {
1359                                         if (d == 0)
1360                                                 d = conf->raid_disks;
1361                                         d--;
1362                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1363                                                 continue;
1364                                         rdev = conf->mirrors[d].rdev;
1365                                         if (sync_page_io(rdev,
1366                                                          sect,
1367                                                          s<<9,
1368                                                          bio->bi_io_vec[idx].bv_page,
1369                                                          READ, false) == 0)
1370                                                 md_error(mddev, rdev);
1371                                 }
1372                         } else {
1373                                 char b[BDEVNAME_SIZE];
1374                                 /* Cannot read from anywhere, array is toast */
1375                                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1376                                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1377                                        " for block %llu\n",
1378                                        mdname(mddev),
1379                                        bdevname(bio->bi_bdev, b),
1380                                        (unsigned long long)r1_bio->sector);
1381                                 md_done_sync(mddev, r1_bio->sectors, 0);
1382                                 put_buf(r1_bio);
1383                                 return;
1384                         }
1385                         sectors -= s;
1386                         sect += s;
1387                         idx ++;
1388                 }
1389         }
1390
1391         /*
1392          * schedule writes
1393          */
1394         atomic_set(&r1_bio->remaining, 1);
1395         for (i = 0; i < disks ; i++) {
1396                 wbio = r1_bio->bios[i];
1397                 if (wbio->bi_end_io == NULL ||
1398                     (wbio->bi_end_io == end_sync_read &&
1399                      (i == r1_bio->read_disk ||
1400                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1401                         continue;
1402
1403                 wbio->bi_rw = WRITE;
1404                 wbio->bi_end_io = end_sync_write;
1405                 atomic_inc(&r1_bio->remaining);
1406                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1407
1408                 generic_make_request(wbio);
1409         }
1410
1411         if (atomic_dec_and_test(&r1_bio->remaining)) {
1412                 /* if we're here, all write(s) have completed, so clean up */
1413                 md_done_sync(mddev, r1_bio->sectors, 1);
1414                 put_buf(r1_bio);
1415         }
1416 }
1417
1418 /*
1419  * This is a kernel thread which:
1420  *
1421  *      1.      Retries failed read operations on working mirrors.
1422  *      2.      Updates the raid superblock when problems encounter.
1423  *      3.      Performs writes following reads for array syncronising.
1424  */
1425
1426 static void fix_read_error(conf_t *conf, int read_disk,
1427                            sector_t sect, int sectors)
1428 {
1429         mddev_t *mddev = conf->mddev;
1430         while(sectors) {
1431                 int s = sectors;
1432                 int d = read_disk;
1433                 int success = 0;
1434                 int start;
1435                 mdk_rdev_t *rdev;
1436
1437                 if (s > (PAGE_SIZE>>9))
1438                         s = PAGE_SIZE >> 9;
1439
1440                 do {
1441                         /* Note: no rcu protection needed here
1442                          * as this is synchronous in the raid1d thread
1443                          * which is the thread that might remove
1444                          * a device.  If raid1d ever becomes multi-threaded....
1445                          */
1446                         rdev = conf->mirrors[d].rdev;
1447                         if (rdev &&
1448                             test_bit(In_sync, &rdev->flags) &&
1449                             sync_page_io(rdev, sect, s<<9,
1450                                          conf->tmppage, READ, false))
1451                                 success = 1;
1452                         else {
1453                                 d++;
1454                                 if (d == conf->raid_disks)
1455                                         d = 0;
1456                         }
1457                 } while (!success && d != read_disk);
1458
1459                 if (!success) {
1460                         /* Cannot read from anywhere -- bye bye array */
1461                         md_error(mddev, conf->mirrors[read_disk].rdev);
1462                         break;
1463                 }
1464                 /* write it back and re-read */
1465                 start = d;
1466                 while (d != read_disk) {
1467                         if (d==0)
1468                                 d = conf->raid_disks;
1469                         d--;
1470                         rdev = conf->mirrors[d].rdev;
1471                         if (rdev &&
1472                             test_bit(In_sync, &rdev->flags)) {
1473                                 if (sync_page_io(rdev, sect, s<<9,
1474                                                  conf->tmppage, WRITE, false)
1475                                     == 0)
1476                                         /* Well, this device is dead */
1477                                         md_error(mddev, rdev);
1478                         }
1479                 }
1480                 d = start;
1481                 while (d != read_disk) {
1482                         char b[BDEVNAME_SIZE];
1483                         if (d==0)
1484                                 d = conf->raid_disks;
1485                         d--;
1486                         rdev = conf->mirrors[d].rdev;
1487                         if (rdev &&
1488                             test_bit(In_sync, &rdev->flags)) {
1489                                 if (sync_page_io(rdev, sect, s<<9,
1490                                                  conf->tmppage, READ, false)
1491                                     == 0)
1492                                         /* Well, this device is dead */
1493                                         md_error(mddev, rdev);
1494                                 else {
1495                                         atomic_add(s, &rdev->corrected_errors);
1496                                         printk(KERN_INFO
1497                                                "md/raid1:%s: read error corrected "
1498                                                "(%d sectors at %llu on %s)\n",
1499                                                mdname(mddev), s,
1500                                                (unsigned long long)(sect +
1501                                                    rdev->data_offset),
1502                                                bdevname(rdev->bdev, b));
1503                                 }
1504                         }
1505                 }
1506                 sectors -= s;
1507                 sect += s;
1508         }
1509 }
1510
1511 static void raid1d(mddev_t *mddev)
1512 {
1513         r1bio_t *r1_bio;
1514         struct bio *bio;
1515         unsigned long flags;
1516         conf_t *conf = mddev->private;
1517         struct list_head *head = &conf->retry_list;
1518         mdk_rdev_t *rdev;
1519
1520         md_check_recovery(mddev);
1521         
1522         for (;;) {
1523                 char b[BDEVNAME_SIZE];
1524
1525                 flush_pending_writes(conf);
1526
1527                 spin_lock_irqsave(&conf->device_lock, flags);
1528                 if (list_empty(head)) {
1529                         spin_unlock_irqrestore(&conf->device_lock, flags);
1530                         break;
1531                 }
1532                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1533                 list_del(head->prev);
1534                 conf->nr_queued--;
1535                 spin_unlock_irqrestore(&conf->device_lock, flags);
1536
1537                 mddev = r1_bio->mddev;
1538                 conf = mddev->private;
1539                 if (test_bit(R1BIO_IsSync, &r1_bio->state))
1540                         sync_request_write(mddev, r1_bio);
1541                 else {
1542                         int disk;
1543
1544                         /* we got a read error. Maybe the drive is bad.  Maybe just
1545                          * the block and we can fix it.
1546                          * We freeze all other IO, and try reading the block from
1547                          * other devices.  When we find one, we re-write
1548                          * and check it that fixes the read error.
1549                          * This is all done synchronously while the array is
1550                          * frozen
1551                          */
1552                         if (mddev->ro == 0) {
1553                                 freeze_array(conf);
1554                                 fix_read_error(conf, r1_bio->read_disk,
1555                                                r1_bio->sector,
1556                                                r1_bio->sectors);
1557                                 unfreeze_array(conf);
1558                         } else
1559                                 md_error(mddev,
1560                                          conf->mirrors[r1_bio->read_disk].rdev);
1561
1562                         bio = r1_bio->bios[r1_bio->read_disk];
1563                         if ((disk=read_balance(conf, r1_bio)) == -1) {
1564                                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
1565                                        " read error for block %llu\n",
1566                                        mdname(mddev),
1567                                        bdevname(bio->bi_bdev,b),
1568                                        (unsigned long long)r1_bio->sector);
1569                                 raid_end_bio_io(r1_bio);
1570                         } else {
1571                                 const unsigned long do_sync = r1_bio->master_bio->bi_rw & REQ_SYNC;
1572                                 r1_bio->bios[r1_bio->read_disk] =
1573                                         mddev->ro ? IO_BLOCKED : NULL;
1574                                 r1_bio->read_disk = disk;
1575                                 bio_put(bio);
1576                                 bio = bio_clone_mddev(r1_bio->master_bio,
1577                                                       GFP_NOIO, mddev);
1578                                 r1_bio->bios[r1_bio->read_disk] = bio;
1579                                 rdev = conf->mirrors[disk].rdev;
1580                                 if (printk_ratelimit())
1581                                         printk(KERN_ERR "md/raid1:%s: redirecting sector %llu to"
1582                                                " other mirror: %s\n",
1583                                                mdname(mddev),
1584                                                (unsigned long long)r1_bio->sector,
1585                                                bdevname(rdev->bdev,b));
1586                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1587                                 bio->bi_bdev = rdev->bdev;
1588                                 bio->bi_end_io = raid1_end_read_request;
1589                                 bio->bi_rw = READ | do_sync;
1590                                 bio->bi_private = r1_bio;
1591                                 generic_make_request(bio);
1592                         }
1593                 }
1594                 cond_resched();
1595         }
1596 }
1597
1598
1599 static int init_resync(conf_t *conf)
1600 {
1601         int buffs;
1602
1603         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1604         BUG_ON(conf->r1buf_pool);
1605         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1606                                           conf->poolinfo);
1607         if (!conf->r1buf_pool)
1608                 return -ENOMEM;
1609         conf->next_resync = 0;
1610         return 0;
1611 }
1612
1613 /*
1614  * perform a "sync" on one "block"
1615  *
1616  * We need to make sure that no normal I/O request - particularly write
1617  * requests - conflict with active sync requests.
1618  *
1619  * This is achieved by tracking pending requests and a 'barrier' concept
1620  * that can be installed to exclude normal IO requests.
1621  */
1622
1623 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1624 {
1625         conf_t *conf = mddev->private;
1626         r1bio_t *r1_bio;
1627         struct bio *bio;
1628         sector_t max_sector, nr_sectors;
1629         int disk = -1;
1630         int i;
1631         int wonly = -1;
1632         int write_targets = 0, read_targets = 0;
1633         sector_t sync_blocks;
1634         int still_degraded = 0;
1635
1636         if (!conf->r1buf_pool)
1637                 if (init_resync(conf))
1638                         return 0;
1639
1640         max_sector = mddev->dev_sectors;
1641         if (sector_nr >= max_sector) {
1642                 /* If we aborted, we need to abort the
1643                  * sync on the 'current' bitmap chunk (there will
1644                  * only be one in raid1 resync.
1645                  * We can find the current addess in mddev->curr_resync
1646                  */
1647                 if (mddev->curr_resync < max_sector) /* aborted */
1648                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1649                                                 &sync_blocks, 1);
1650                 else /* completed sync */
1651                         conf->fullsync = 0;
1652
1653                 bitmap_close_sync(mddev->bitmap);
1654                 close_sync(conf);
1655                 return 0;
1656         }
1657
1658         if (mddev->bitmap == NULL &&
1659             mddev->recovery_cp == MaxSector &&
1660             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1661             conf->fullsync == 0) {
1662                 *skipped = 1;
1663                 return max_sector - sector_nr;
1664         }
1665         /* before building a request, check if we can skip these blocks..
1666          * This call the bitmap_start_sync doesn't actually record anything
1667          */
1668         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1669             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1670                 /* We can skip this block, and probably several more */
1671                 *skipped = 1;
1672                 return sync_blocks;
1673         }
1674         /*
1675          * If there is non-resync activity waiting for a turn,
1676          * and resync is going fast enough,
1677          * then let it though before starting on this new sync request.
1678          */
1679         if (!go_faster && conf->nr_waiting)
1680                 msleep_interruptible(1000);
1681
1682         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1683         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1684         raise_barrier(conf);
1685
1686         conf->next_resync = sector_nr;
1687
1688         rcu_read_lock();
1689         /*
1690          * If we get a correctably read error during resync or recovery,
1691          * we might want to read from a different device.  So we
1692          * flag all drives that could conceivably be read from for READ,
1693          * and any others (which will be non-In_sync devices) for WRITE.
1694          * If a read fails, we try reading from something else for which READ
1695          * is OK.
1696          */
1697
1698         r1_bio->mddev = mddev;
1699         r1_bio->sector = sector_nr;
1700         r1_bio->state = 0;
1701         set_bit(R1BIO_IsSync, &r1_bio->state);
1702
1703         for (i=0; i < conf->raid_disks; i++) {
1704                 mdk_rdev_t *rdev;
1705                 bio = r1_bio->bios[i];
1706
1707                 /* take from bio_init */
1708                 bio->bi_next = NULL;
1709                 bio->bi_flags &= ~(BIO_POOL_MASK-1);
1710                 bio->bi_flags |= 1 << BIO_UPTODATE;
1711                 bio->bi_comp_cpu = -1;
1712                 bio->bi_rw = READ;
1713                 bio->bi_vcnt = 0;
1714                 bio->bi_idx = 0;
1715                 bio->bi_phys_segments = 0;
1716                 bio->bi_size = 0;
1717                 bio->bi_end_io = NULL;
1718                 bio->bi_private = NULL;
1719
1720                 rdev = rcu_dereference(conf->mirrors[i].rdev);
1721                 if (rdev == NULL ||
1722                            test_bit(Faulty, &rdev->flags)) {
1723                         still_degraded = 1;
1724                         continue;
1725                 } else if (!test_bit(In_sync, &rdev->flags)) {
1726                         bio->bi_rw = WRITE;
1727                         bio->bi_end_io = end_sync_write;
1728                         write_targets ++;
1729                 } else {
1730                         /* may need to read from here */
1731                         bio->bi_rw = READ;
1732                         bio->bi_end_io = end_sync_read;
1733                         if (test_bit(WriteMostly, &rdev->flags)) {
1734                                 if (wonly < 0)
1735                                         wonly = i;
1736                         } else {
1737                                 if (disk < 0)
1738                                         disk = i;
1739                         }
1740                         read_targets++;
1741                 }
1742                 atomic_inc(&rdev->nr_pending);
1743                 bio->bi_sector = sector_nr + rdev->data_offset;
1744                 bio->bi_bdev = rdev->bdev;
1745                 bio->bi_private = r1_bio;
1746         }
1747         rcu_read_unlock();
1748         if (disk < 0)
1749                 disk = wonly;
1750         r1_bio->read_disk = disk;
1751
1752         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1753                 /* extra read targets are also write targets */
1754                 write_targets += read_targets-1;
1755
1756         if (write_targets == 0 || read_targets == 0) {
1757                 /* There is nowhere to write, so all non-sync
1758                  * drives must be failed - so we are finished
1759                  */
1760                 sector_t rv = max_sector - sector_nr;
1761                 *skipped = 1;
1762                 put_buf(r1_bio);
1763                 return rv;
1764         }
1765
1766         if (max_sector > mddev->resync_max)
1767                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1768         nr_sectors = 0;
1769         sync_blocks = 0;
1770         do {
1771                 struct page *page;
1772                 int len = PAGE_SIZE;
1773                 if (sector_nr + (len>>9) > max_sector)
1774                         len = (max_sector - sector_nr) << 9;
1775                 if (len == 0)
1776                         break;
1777                 if (sync_blocks == 0) {
1778                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1779                                                &sync_blocks, still_degraded) &&
1780                             !conf->fullsync &&
1781                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1782                                 break;
1783                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1784                         if ((len >> 9) > sync_blocks)
1785                                 len = sync_blocks<<9;
1786                 }
1787
1788                 for (i=0 ; i < conf->raid_disks; i++) {
1789                         bio = r1_bio->bios[i];
1790                         if (bio->bi_end_io) {
1791                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1792                                 if (bio_add_page(bio, page, len, 0) == 0) {
1793                                         /* stop here */
1794                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1795                                         while (i > 0) {
1796                                                 i--;
1797                                                 bio = r1_bio->bios[i];
1798                                                 if (bio->bi_end_io==NULL)
1799                                                         continue;
1800                                                 /* remove last page from this bio */
1801                                                 bio->bi_vcnt--;
1802                                                 bio->bi_size -= len;
1803                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1804                                         }
1805                                         goto bio_full;
1806                                 }
1807                         }
1808                 }
1809                 nr_sectors += len>>9;
1810                 sector_nr += len>>9;
1811                 sync_blocks -= (len>>9);
1812         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1813  bio_full:
1814         r1_bio->sectors = nr_sectors;
1815
1816         /* For a user-requested sync, we read all readable devices and do a
1817          * compare
1818          */
1819         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1820                 atomic_set(&r1_bio->remaining, read_targets);
1821                 for (i=0; i<conf->raid_disks; i++) {
1822                         bio = r1_bio->bios[i];
1823                         if (bio->bi_end_io == end_sync_read) {
1824                                 md_sync_acct(bio->bi_bdev, nr_sectors);
1825                                 generic_make_request(bio);
1826                         }
1827                 }
1828         } else {
1829                 atomic_set(&r1_bio->remaining, 1);
1830                 bio = r1_bio->bios[r1_bio->read_disk];
1831                 md_sync_acct(bio->bi_bdev, nr_sectors);
1832                 generic_make_request(bio);
1833
1834         }
1835         return nr_sectors;
1836 }
1837
1838 static sector_t raid1_size(mddev_t *mddev, sector_t sectors, int raid_disks)
1839 {
1840         if (sectors)
1841                 return sectors;
1842
1843         return mddev->dev_sectors;
1844 }
1845
1846 static conf_t *setup_conf(mddev_t *mddev)
1847 {
1848         conf_t *conf;
1849         int i;
1850         mirror_info_t *disk;
1851         mdk_rdev_t *rdev;
1852         int err = -ENOMEM;
1853
1854         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1855         if (!conf)
1856                 goto abort;
1857
1858         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1859                                  GFP_KERNEL);
1860         if (!conf->mirrors)
1861                 goto abort;
1862
1863         conf->tmppage = alloc_page(GFP_KERNEL);
1864         if (!conf->tmppage)
1865                 goto abort;
1866
1867         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1868         if (!conf->poolinfo)
1869                 goto abort;
1870         conf->poolinfo->raid_disks = mddev->raid_disks;
1871         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1872                                           r1bio_pool_free,
1873                                           conf->poolinfo);
1874         if (!conf->r1bio_pool)
1875                 goto abort;
1876
1877         conf->poolinfo->mddev = mddev;
1878
1879         spin_lock_init(&conf->device_lock);
1880         list_for_each_entry(rdev, &mddev->disks, same_set) {
1881                 int disk_idx = rdev->raid_disk;
1882                 if (disk_idx >= mddev->raid_disks
1883                     || disk_idx < 0)
1884                         continue;
1885                 disk = conf->mirrors + disk_idx;
1886
1887                 disk->rdev = rdev;
1888
1889                 disk->head_position = 0;
1890         }
1891         conf->raid_disks = mddev->raid_disks;
1892         conf->mddev = mddev;
1893         INIT_LIST_HEAD(&conf->retry_list);
1894
1895         spin_lock_init(&conf->resync_lock);
1896         init_waitqueue_head(&conf->wait_barrier);
1897
1898         bio_list_init(&conf->pending_bio_list);
1899
1900         conf->last_used = -1;
1901         for (i = 0; i < conf->raid_disks; i++) {
1902
1903                 disk = conf->mirrors + i;
1904
1905                 if (!disk->rdev ||
1906                     !test_bit(In_sync, &disk->rdev->flags)) {
1907                         disk->head_position = 0;
1908                         if (disk->rdev)
1909                                 conf->fullsync = 1;
1910                 } else if (conf->last_used < 0)
1911                         /*
1912                          * The first working device is used as a
1913                          * starting point to read balancing.
1914                          */
1915                         conf->last_used = i;
1916         }
1917
1918         err = -EIO;
1919         if (conf->last_used < 0) {
1920                 printk(KERN_ERR "md/raid1:%s: no operational mirrors\n",
1921                        mdname(mddev));
1922                 goto abort;
1923         }
1924         err = -ENOMEM;
1925         conf->thread = md_register_thread(raid1d, mddev, NULL);
1926         if (!conf->thread) {
1927                 printk(KERN_ERR
1928                        "md/raid1:%s: couldn't allocate thread\n",
1929                        mdname(mddev));
1930                 goto abort;
1931         }
1932
1933         return conf;
1934
1935  abort:
1936         if (conf) {
1937                 if (conf->r1bio_pool)
1938                         mempool_destroy(conf->r1bio_pool);
1939                 kfree(conf->mirrors);
1940                 safe_put_page(conf->tmppage);
1941                 kfree(conf->poolinfo);
1942                 kfree(conf);
1943         }
1944         return ERR_PTR(err);
1945 }
1946
1947 static int run(mddev_t *mddev)
1948 {
1949         conf_t *conf;
1950         int i;
1951         mdk_rdev_t *rdev;
1952
1953         if (mddev->level != 1) {
1954                 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
1955                        mdname(mddev), mddev->level);
1956                 return -EIO;
1957         }
1958         if (mddev->reshape_position != MaxSector) {
1959                 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
1960                        mdname(mddev));
1961                 return -EIO;
1962         }
1963         /*
1964          * copy the already verified devices into our private RAID1
1965          * bookkeeping area. [whatever we allocate in run(),
1966          * should be freed in stop()]
1967          */
1968         if (mddev->private == NULL)
1969                 conf = setup_conf(mddev);
1970         else
1971                 conf = mddev->private;
1972
1973         if (IS_ERR(conf))
1974                 return PTR_ERR(conf);
1975
1976         list_for_each_entry(rdev, &mddev->disks, same_set) {
1977                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1978                                   rdev->data_offset << 9);
1979                 /* as we don't honour merge_bvec_fn, we must never risk
1980                  * violating it, so limit ->max_segments to 1 lying within
1981                  * a single page, as a one page request is never in violation.
1982                  */
1983                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1984                         blk_queue_max_segments(mddev->queue, 1);
1985                         blk_queue_segment_boundary(mddev->queue,
1986                                                    PAGE_CACHE_SIZE - 1);
1987                 }
1988         }
1989
1990         mddev->degraded = 0;
1991         for (i=0; i < conf->raid_disks; i++)
1992                 if (conf->mirrors[i].rdev == NULL ||
1993                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
1994                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
1995                         mddev->degraded++;
1996
1997         if (conf->raid_disks - mddev->degraded == 1)
1998                 mddev->recovery_cp = MaxSector;
1999
2000         if (mddev->recovery_cp != MaxSector)
2001                 printk(KERN_NOTICE "md/raid1:%s: not clean"
2002                        " -- starting background reconstruction\n",
2003                        mdname(mddev));
2004         printk(KERN_INFO 
2005                 "md/raid1:%s: active with %d out of %d mirrors\n",
2006                 mdname(mddev), mddev->raid_disks - mddev->degraded, 
2007                 mddev->raid_disks);
2008
2009         /*
2010          * Ok, everything is just fine now
2011          */
2012         mddev->thread = conf->thread;
2013         conf->thread = NULL;
2014         mddev->private = conf;
2015
2016         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2017
2018         mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2019         mddev->queue->backing_dev_info.congested_data = mddev;
2020         md_integrity_register(mddev);
2021         return 0;
2022 }
2023
2024 static int stop(mddev_t *mddev)
2025 {
2026         conf_t *conf = mddev->private;
2027         struct bitmap *bitmap = mddev->bitmap;
2028
2029         /* wait for behind writes to complete */
2030         if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2031                 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2032                        mdname(mddev));
2033                 /* need to kick something here to make sure I/O goes? */
2034                 wait_event(bitmap->behind_wait,
2035                            atomic_read(&bitmap->behind_writes) == 0);
2036         }
2037
2038         raise_barrier(conf);
2039         lower_barrier(conf);
2040
2041         md_unregister_thread(mddev->thread);
2042         mddev->thread = NULL;
2043         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2044         if (conf->r1bio_pool)
2045                 mempool_destroy(conf->r1bio_pool);
2046         kfree(conf->mirrors);
2047         kfree(conf->poolinfo);
2048         kfree(conf);
2049         mddev->private = NULL;
2050         return 0;
2051 }
2052
2053 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2054 {
2055         /* no resync is happening, and there is enough space
2056          * on all devices, so we can resize.
2057          * We need to make sure resync covers any new space.
2058          * If the array is shrinking we should possibly wait until
2059          * any io in the removed space completes, but it hardly seems
2060          * worth it.
2061          */
2062         md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2063         if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2064                 return -EINVAL;
2065         set_capacity(mddev->gendisk, mddev->array_sectors);
2066         revalidate_disk(mddev->gendisk);
2067         if (sectors > mddev->dev_sectors &&
2068             mddev->recovery_cp == MaxSector) {
2069                 mddev->recovery_cp = mddev->dev_sectors;
2070                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2071         }
2072         mddev->dev_sectors = sectors;
2073         mddev->resync_max_sectors = sectors;
2074         return 0;
2075 }
2076
2077 static int raid1_reshape(mddev_t *mddev)
2078 {
2079         /* We need to:
2080          * 1/ resize the r1bio_pool
2081          * 2/ resize conf->mirrors
2082          *
2083          * We allocate a new r1bio_pool if we can.
2084          * Then raise a device barrier and wait until all IO stops.
2085          * Then resize conf->mirrors and swap in the new r1bio pool.
2086          *
2087          * At the same time, we "pack" the devices so that all the missing
2088          * devices have the higher raid_disk numbers.
2089          */
2090         mempool_t *newpool, *oldpool;
2091         struct pool_info *newpoolinfo;
2092         mirror_info_t *newmirrors;
2093         conf_t *conf = mddev->private;
2094         int cnt, raid_disks;
2095         unsigned long flags;
2096         int d, d2, err;
2097
2098         /* Cannot change chunk_size, layout, or level */
2099         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2100             mddev->layout != mddev->new_layout ||
2101             mddev->level != mddev->new_level) {
2102                 mddev->new_chunk_sectors = mddev->chunk_sectors;
2103                 mddev->new_layout = mddev->layout;
2104                 mddev->new_level = mddev->level;
2105                 return -EINVAL;
2106         }
2107
2108         err = md_allow_write(mddev);
2109         if (err)
2110                 return err;
2111
2112         raid_disks = mddev->raid_disks + mddev->delta_disks;
2113
2114         if (raid_disks < conf->raid_disks) {
2115                 cnt=0;
2116                 for (d= 0; d < conf->raid_disks; d++)
2117                         if (conf->mirrors[d].rdev)
2118                                 cnt++;
2119                 if (cnt > raid_disks)
2120                         return -EBUSY;
2121         }
2122
2123         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2124         if (!newpoolinfo)
2125                 return -ENOMEM;
2126         newpoolinfo->mddev = mddev;
2127         newpoolinfo->raid_disks = raid_disks;
2128
2129         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2130                                  r1bio_pool_free, newpoolinfo);
2131         if (!newpool) {
2132                 kfree(newpoolinfo);
2133                 return -ENOMEM;
2134         }
2135         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2136         if (!newmirrors) {
2137                 kfree(newpoolinfo);
2138                 mempool_destroy(newpool);
2139                 return -ENOMEM;
2140         }
2141
2142         raise_barrier(conf);
2143
2144         /* ok, everything is stopped */
2145         oldpool = conf->r1bio_pool;
2146         conf->r1bio_pool = newpool;
2147
2148         for (d = d2 = 0; d < conf->raid_disks; d++) {
2149                 mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2150                 if (rdev && rdev->raid_disk != d2) {
2151                         char nm[20];
2152                         sprintf(nm, "rd%d", rdev->raid_disk);
2153                         sysfs_remove_link(&mddev->kobj, nm);
2154                         rdev->raid_disk = d2;
2155                         sprintf(nm, "rd%d", rdev->raid_disk);
2156                         sysfs_remove_link(&mddev->kobj, nm);
2157                         if (sysfs_create_link(&mddev->kobj,
2158                                               &rdev->kobj, nm))
2159                                 printk(KERN_WARNING
2160                                        "md/raid1:%s: cannot register "
2161                                        "%s\n",
2162                                        mdname(mddev), nm);
2163                 }
2164                 if (rdev)
2165                         newmirrors[d2++].rdev = rdev;
2166         }
2167         kfree(conf->mirrors);
2168         conf->mirrors = newmirrors;
2169         kfree(conf->poolinfo);
2170         conf->poolinfo = newpoolinfo;
2171
2172         spin_lock_irqsave(&conf->device_lock, flags);
2173         mddev->degraded += (raid_disks - conf->raid_disks);
2174         spin_unlock_irqrestore(&conf->device_lock, flags);
2175         conf->raid_disks = mddev->raid_disks = raid_disks;
2176         mddev->delta_disks = 0;
2177
2178         conf->last_used = 0; /* just make sure it is in-range */
2179         lower_barrier(conf);
2180
2181         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2182         md_wakeup_thread(mddev->thread);
2183
2184         mempool_destroy(oldpool);
2185         return 0;
2186 }
2187
2188 static void raid1_quiesce(mddev_t *mddev, int state)
2189 {
2190         conf_t *conf = mddev->private;
2191
2192         switch(state) {
2193         case 2: /* wake for suspend */
2194                 wake_up(&conf->wait_barrier);
2195                 break;
2196         case 1:
2197                 raise_barrier(conf);
2198                 break;
2199         case 0:
2200                 lower_barrier(conf);
2201                 break;
2202         }
2203 }
2204
2205 static void *raid1_takeover(mddev_t *mddev)
2206 {
2207         /* raid1 can take over:
2208          *  raid5 with 2 devices, any layout or chunk size
2209          */
2210         if (mddev->level == 5 && mddev->raid_disks == 2) {
2211                 conf_t *conf;
2212                 mddev->new_level = 1;
2213                 mddev->new_layout = 0;
2214                 mddev->new_chunk_sectors = 0;
2215                 conf = setup_conf(mddev);
2216                 if (!IS_ERR(conf))
2217                         conf->barrier = 1;
2218                 return conf;
2219         }
2220         return ERR_PTR(-EINVAL);
2221 }
2222
2223 static struct mdk_personality raid1_personality =
2224 {
2225         .name           = "raid1",
2226         .level          = 1,
2227         .owner          = THIS_MODULE,
2228         .make_request   = make_request,
2229         .run            = run,
2230         .stop           = stop,
2231         .status         = status,
2232         .error_handler  = error,
2233         .hot_add_disk   = raid1_add_disk,
2234         .hot_remove_disk= raid1_remove_disk,
2235         .spare_active   = raid1_spare_active,
2236         .sync_request   = sync_request,
2237         .resize         = raid1_resize,
2238         .size           = raid1_size,
2239         .check_reshape  = raid1_reshape,
2240         .quiesce        = raid1_quiesce,
2241         .takeover       = raid1_takeover,
2242 };
2243
2244 static int __init raid_init(void)
2245 {
2246         return register_md_personality(&raid1_personality);
2247 }
2248
2249 static void raid_exit(void)
2250 {
2251         unregister_md_personality(&raid1_personality);
2252 }
2253
2254 module_init(raid_init);
2255 module_exit(raid_exit);
2256 MODULE_LICENSE("GPL");
2257 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2258 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2259 MODULE_ALIAS("md-raid1");
2260 MODULE_ALIAS("md-level-1");