Merge branch 'from-linus' into upstream
[linux-2.6] / drivers / md / raid1.c
1 /*
2  * raid1.c : Multiple Devices driver for Linux
3  *
4  * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
5  *
6  * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
7  *
8  * RAID-1 management functions.
9  *
10  * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
11  *
12  * Fixes to reconstruction by Jakob Ã˜stergaard" <jakob@ostenfeld.dk>
13  * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
14  *
15  * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
16  * bitmapped intelligence in resync:
17  *
18  *      - bitmap marked during normal i/o
19  *      - bitmap used to skip nondirty blocks during sync
20  *
21  * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
22  * - persistent bitmap code
23  *
24  * This program is free software; you can redistribute it and/or modify
25  * it under the terms of the GNU General Public License as published by
26  * the Free Software Foundation; either version 2, or (at your option)
27  * any later version.
28  *
29  * You should have received a copy of the GNU General Public License
30  * (for example /usr/src/linux/COPYING); if not, write to the Free
31  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
32  */
33
34 #include "dm-bio-list.h"
35 #include <linux/raid/raid1.h>
36 #include <linux/raid/bitmap.h>
37
38 #define DEBUG 0
39 #if DEBUG
40 #define PRINTK(x...) printk(x)
41 #else
42 #define PRINTK(x...)
43 #endif
44
45 /*
46  * Number of guaranteed r1bios in case of extreme VM load:
47  */
48 #define NR_RAID1_BIOS 256
49
50
51 static void unplug_slaves(mddev_t *mddev);
52
53 static void allow_barrier(conf_t *conf);
54 static void lower_barrier(conf_t *conf);
55
56 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
57 {
58         struct pool_info *pi = data;
59         r1bio_t *r1_bio;
60         int size = offsetof(r1bio_t, bios[pi->raid_disks]);
61
62         /* allocate a r1bio with room for raid_disks entries in the bios array */
63         r1_bio = kzalloc(size, gfp_flags);
64         if (!r1_bio)
65                 unplug_slaves(pi->mddev);
66
67         return r1_bio;
68 }
69
70 static void r1bio_pool_free(void *r1_bio, void *data)
71 {
72         kfree(r1_bio);
73 }
74
75 #define RESYNC_BLOCK_SIZE (64*1024)
76 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
77 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
78 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
79 #define RESYNC_WINDOW (2048*1024)
80
81 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
82 {
83         struct pool_info *pi = data;
84         struct page *page;
85         r1bio_t *r1_bio;
86         struct bio *bio;
87         int i, j;
88
89         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
90         if (!r1_bio) {
91                 unplug_slaves(pi->mddev);
92                 return NULL;
93         }
94
95         /*
96          * Allocate bios : 1 for reading, n-1 for writing
97          */
98         for (j = pi->raid_disks ; j-- ; ) {
99                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
100                 if (!bio)
101                         goto out_free_bio;
102                 r1_bio->bios[j] = bio;
103         }
104         /*
105          * Allocate RESYNC_PAGES data pages and attach them to
106          * the first bio.
107          * If this is a user-requested check/repair, allocate
108          * RESYNC_PAGES for each bio.
109          */
110         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
111                 j = pi->raid_disks;
112         else
113                 j = 1;
114         while(j--) {
115                 bio = r1_bio->bios[j];
116                 for (i = 0; i < RESYNC_PAGES; i++) {
117                         page = alloc_page(gfp_flags);
118                         if (unlikely(!page))
119                                 goto out_free_pages;
120
121                         bio->bi_io_vec[i].bv_page = page;
122                 }
123         }
124         /* If not user-requests, copy the page pointers to all bios */
125         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
126                 for (i=0; i<RESYNC_PAGES ; i++)
127                         for (j=1; j<pi->raid_disks; j++)
128                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
129                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
130         }
131
132         r1_bio->master_bio = NULL;
133
134         return r1_bio;
135
136 out_free_pages:
137         for (i=0; i < RESYNC_PAGES ; i++)
138                 for (j=0 ; j < pi->raid_disks; j++)
139                         safe_put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
140         j = -1;
141 out_free_bio:
142         while ( ++j < pi->raid_disks )
143                 bio_put(r1_bio->bios[j]);
144         r1bio_pool_free(r1_bio, data);
145         return NULL;
146 }
147
148 static void r1buf_pool_free(void *__r1_bio, void *data)
149 {
150         struct pool_info *pi = data;
151         int i,j;
152         r1bio_t *r1bio = __r1_bio;
153
154         for (i = 0; i < RESYNC_PAGES; i++)
155                 for (j = pi->raid_disks; j-- ;) {
156                         if (j == 0 ||
157                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
158                             r1bio->bios[0]->bi_io_vec[i].bv_page)
159                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
160                 }
161         for (i=0 ; i < pi->raid_disks; i++)
162                 bio_put(r1bio->bios[i]);
163
164         r1bio_pool_free(r1bio, data);
165 }
166
167 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
168 {
169         int i;
170
171         for (i = 0; i < conf->raid_disks; i++) {
172                 struct bio **bio = r1_bio->bios + i;
173                 if (*bio && *bio != IO_BLOCKED)
174                         bio_put(*bio);
175                 *bio = NULL;
176         }
177 }
178
179 static void free_r1bio(r1bio_t *r1_bio)
180 {
181         conf_t *conf = mddev_to_conf(r1_bio->mddev);
182
183         /*
184          * Wake up any possible resync thread that waits for the device
185          * to go idle.
186          */
187         allow_barrier(conf);
188
189         put_all_bios(conf, r1_bio);
190         mempool_free(r1_bio, conf->r1bio_pool);
191 }
192
193 static void put_buf(r1bio_t *r1_bio)
194 {
195         conf_t *conf = mddev_to_conf(r1_bio->mddev);
196         int i;
197
198         for (i=0; i<conf->raid_disks; i++) {
199                 struct bio *bio = r1_bio->bios[i];
200                 if (bio->bi_end_io)
201                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
202         }
203
204         mempool_free(r1_bio, conf->r1buf_pool);
205
206         lower_barrier(conf);
207 }
208
209 static void reschedule_retry(r1bio_t *r1_bio)
210 {
211         unsigned long flags;
212         mddev_t *mddev = r1_bio->mddev;
213         conf_t *conf = mddev_to_conf(mddev);
214
215         spin_lock_irqsave(&conf->device_lock, flags);
216         list_add(&r1_bio->retry_list, &conf->retry_list);
217         conf->nr_queued ++;
218         spin_unlock_irqrestore(&conf->device_lock, flags);
219
220         wake_up(&conf->wait_barrier);
221         md_wakeup_thread(mddev->thread);
222 }
223
224 /*
225  * raid_end_bio_io() is called when we have finished servicing a mirrored
226  * operation and are ready to return a success/failure code to the buffer
227  * cache layer.
228  */
229 static void raid_end_bio_io(r1bio_t *r1_bio)
230 {
231         struct bio *bio = r1_bio->master_bio;
232
233         /* if nobody has done the final endio yet, do it now */
234         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
235                 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
236                         (bio_data_dir(bio) == WRITE) ? "write" : "read",
237                         (unsigned long long) bio->bi_sector,
238                         (unsigned long long) bio->bi_sector +
239                                 (bio->bi_size >> 9) - 1);
240
241                 bio_endio(bio, bio->bi_size,
242                         test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
243         }
244         free_r1bio(r1_bio);
245 }
246
247 /*
248  * Update disk head position estimator based on IRQ completion info.
249  */
250 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
251 {
252         conf_t *conf = mddev_to_conf(r1_bio->mddev);
253
254         conf->mirrors[disk].head_position =
255                 r1_bio->sector + (r1_bio->sectors);
256 }
257
258 static int raid1_end_read_request(struct bio *bio, unsigned int bytes_done, int error)
259 {
260         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
261         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
262         int mirror;
263         conf_t *conf = mddev_to_conf(r1_bio->mddev);
264
265         if (bio->bi_size)
266                 return 1;
267         
268         mirror = r1_bio->read_disk;
269         /*
270          * this branch is our 'one mirror IO has finished' event handler:
271          */
272         update_head_pos(mirror, r1_bio);
273
274         if (uptodate || conf->working_disks <= 1) {
275                 /*
276                  * Set R1BIO_Uptodate in our master bio, so that
277                  * we will return a good error code for to the higher
278                  * levels even if IO on some other mirrored buffer fails.
279                  *
280                  * The 'master' represents the composite IO operation to
281                  * user-side. So if something waits for IO, then it will
282                  * wait for the 'master' bio.
283                  */
284                 if (uptodate)
285                         set_bit(R1BIO_Uptodate, &r1_bio->state);
286
287                 raid_end_bio_io(r1_bio);
288         } else {
289                 /*
290                  * oops, read error:
291                  */
292                 char b[BDEVNAME_SIZE];
293                 if (printk_ratelimit())
294                         printk(KERN_ERR "raid1: %s: rescheduling sector %llu\n",
295                                bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
296                 reschedule_retry(r1_bio);
297         }
298
299         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
300         return 0;
301 }
302
303 static int raid1_end_write_request(struct bio *bio, unsigned int bytes_done, int error)
304 {
305         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
306         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
307         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
308         conf_t *conf = mddev_to_conf(r1_bio->mddev);
309         struct bio *to_put = NULL;
310
311         if (bio->bi_size)
312                 return 1;
313
314         for (mirror = 0; mirror < conf->raid_disks; mirror++)
315                 if (r1_bio->bios[mirror] == bio)
316                         break;
317
318         if (error == -EOPNOTSUPP && test_bit(R1BIO_Barrier, &r1_bio->state)) {
319                 set_bit(BarriersNotsupp, &conf->mirrors[mirror].rdev->flags);
320                 set_bit(R1BIO_BarrierRetry, &r1_bio->state);
321                 r1_bio->mddev->barriers_work = 0;
322                 /* Don't rdev_dec_pending in this branch - keep it for the retry */
323         } else {
324                 /*
325                  * this branch is our 'one mirror IO has finished' event handler:
326                  */
327                 r1_bio->bios[mirror] = NULL;
328                 to_put = bio;
329                 if (!uptodate) {
330                         md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
331                         /* an I/O failed, we can't clear the bitmap */
332                         set_bit(R1BIO_Degraded, &r1_bio->state);
333                 } else
334                         /*
335                          * Set R1BIO_Uptodate in our master bio, so that
336                          * we will return a good error code for to the higher
337                          * levels even if IO on some other mirrored buffer fails.
338                          *
339                          * The 'master' represents the composite IO operation to
340                          * user-side. So if something waits for IO, then it will
341                          * wait for the 'master' bio.
342                          */
343                         set_bit(R1BIO_Uptodate, &r1_bio->state);
344
345                 update_head_pos(mirror, r1_bio);
346
347                 if (behind) {
348                         if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
349                                 atomic_dec(&r1_bio->behind_remaining);
350
351                         /* In behind mode, we ACK the master bio once the I/O has safely
352                          * reached all non-writemostly disks. Setting the Returned bit
353                          * ensures that this gets done only once -- we don't ever want to
354                          * return -EIO here, instead we'll wait */
355
356                         if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
357                             test_bit(R1BIO_Uptodate, &r1_bio->state)) {
358                                 /* Maybe we can return now */
359                                 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
360                                         struct bio *mbio = r1_bio->master_bio;
361                                         PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
362                                                (unsigned long long) mbio->bi_sector,
363                                                (unsigned long long) mbio->bi_sector +
364                                                (mbio->bi_size >> 9) - 1);
365                                         bio_endio(mbio, mbio->bi_size, 0);
366                                 }
367                         }
368                 }
369                 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
370         }
371         /*
372          *
373          * Let's see if all mirrored write operations have finished
374          * already.
375          */
376         if (atomic_dec_and_test(&r1_bio->remaining)) {
377                 if (test_bit(R1BIO_BarrierRetry, &r1_bio->state))
378                         reschedule_retry(r1_bio);
379                 else {
380                         /* it really is the end of this request */
381                         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
382                                 /* free extra copy of the data pages */
383                                 int i = bio->bi_vcnt;
384                                 while (i--)
385                                         safe_put_page(bio->bi_io_vec[i].bv_page);
386                         }
387                         /* clear the bitmap if all writes complete successfully */
388                         bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
389                                         r1_bio->sectors,
390                                         !test_bit(R1BIO_Degraded, &r1_bio->state),
391                                         behind);
392                         md_write_end(r1_bio->mddev);
393                         raid_end_bio_io(r1_bio);
394                 }
395         }
396
397         if (to_put)
398                 bio_put(to_put);
399
400         return 0;
401 }
402
403
404 /*
405  * This routine returns the disk from which the requested read should
406  * be done. There is a per-array 'next expected sequential IO' sector
407  * number - if this matches on the next IO then we use the last disk.
408  * There is also a per-disk 'last know head position' sector that is
409  * maintained from IRQ contexts, both the normal and the resync IO
410  * completion handlers update this position correctly. If there is no
411  * perfect sequential match then we pick the disk whose head is closest.
412  *
413  * If there are 2 mirrors in the same 2 devices, performance degrades
414  * because position is mirror, not device based.
415  *
416  * The rdev for the device selected will have nr_pending incremented.
417  */
418 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
419 {
420         const unsigned long this_sector = r1_bio->sector;
421         int new_disk = conf->last_used, disk = new_disk;
422         int wonly_disk = -1;
423         const int sectors = r1_bio->sectors;
424         sector_t new_distance, current_distance;
425         mdk_rdev_t *rdev;
426
427         rcu_read_lock();
428         /*
429          * Check if we can balance. We can balance on the whole
430          * device if no resync is going on, or below the resync window.
431          * We take the first readable disk when above the resync window.
432          */
433  retry:
434         if (conf->mddev->recovery_cp < MaxSector &&
435             (this_sector + sectors >= conf->next_resync)) {
436                 /* Choose the first operation device, for consistancy */
437                 new_disk = 0;
438
439                 for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
440                      r1_bio->bios[new_disk] == IO_BLOCKED ||
441                      !rdev || !test_bit(In_sync, &rdev->flags)
442                              || test_bit(WriteMostly, &rdev->flags);
443                      rdev = rcu_dereference(conf->mirrors[++new_disk].rdev)) {
444
445                         if (rdev && test_bit(In_sync, &rdev->flags) &&
446                                 r1_bio->bios[new_disk] != IO_BLOCKED)
447                                 wonly_disk = new_disk;
448
449                         if (new_disk == conf->raid_disks - 1) {
450                                 new_disk = wonly_disk;
451                                 break;
452                         }
453                 }
454                 goto rb_out;
455         }
456
457
458         /* make sure the disk is operational */
459         for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
460              r1_bio->bios[new_disk] == IO_BLOCKED ||
461              !rdev || !test_bit(In_sync, &rdev->flags) ||
462                      test_bit(WriteMostly, &rdev->flags);
463              rdev = rcu_dereference(conf->mirrors[new_disk].rdev)) {
464
465                 if (rdev && test_bit(In_sync, &rdev->flags) &&
466                     r1_bio->bios[new_disk] != IO_BLOCKED)
467                         wonly_disk = new_disk;
468
469                 if (new_disk <= 0)
470                         new_disk = conf->raid_disks;
471                 new_disk--;
472                 if (new_disk == disk) {
473                         new_disk = wonly_disk;
474                         break;
475                 }
476         }
477
478         if (new_disk < 0)
479                 goto rb_out;
480
481         disk = new_disk;
482         /* now disk == new_disk == starting point for search */
483
484         /*
485          * Don't change to another disk for sequential reads:
486          */
487         if (conf->next_seq_sect == this_sector)
488                 goto rb_out;
489         if (this_sector == conf->mirrors[new_disk].head_position)
490                 goto rb_out;
491
492         current_distance = abs(this_sector - conf->mirrors[disk].head_position);
493
494         /* Find the disk whose head is closest */
495
496         do {
497                 if (disk <= 0)
498                         disk = conf->raid_disks;
499                 disk--;
500
501                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
502
503                 if (!rdev || r1_bio->bios[disk] == IO_BLOCKED ||
504                     !test_bit(In_sync, &rdev->flags) ||
505                     test_bit(WriteMostly, &rdev->flags))
506                         continue;
507
508                 if (!atomic_read(&rdev->nr_pending)) {
509                         new_disk = disk;
510                         break;
511                 }
512                 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
513                 if (new_distance < current_distance) {
514                         current_distance = new_distance;
515                         new_disk = disk;
516                 }
517         } while (disk != conf->last_used);
518
519  rb_out:
520
521
522         if (new_disk >= 0) {
523                 rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
524                 if (!rdev)
525                         goto retry;
526                 atomic_inc(&rdev->nr_pending);
527                 if (!test_bit(In_sync, &rdev->flags)) {
528                         /* cannot risk returning a device that failed
529                          * before we inc'ed nr_pending
530                          */
531                         rdev_dec_pending(rdev, conf->mddev);
532                         goto retry;
533                 }
534                 conf->next_seq_sect = this_sector + sectors;
535                 conf->last_used = new_disk;
536         }
537         rcu_read_unlock();
538
539         return new_disk;
540 }
541
542 static void unplug_slaves(mddev_t *mddev)
543 {
544         conf_t *conf = mddev_to_conf(mddev);
545         int i;
546
547         rcu_read_lock();
548         for (i=0; i<mddev->raid_disks; i++) {
549                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
550                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
551                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
552
553                         atomic_inc(&rdev->nr_pending);
554                         rcu_read_unlock();
555
556                         if (r_queue->unplug_fn)
557                                 r_queue->unplug_fn(r_queue);
558
559                         rdev_dec_pending(rdev, mddev);
560                         rcu_read_lock();
561                 }
562         }
563         rcu_read_unlock();
564 }
565
566 static void raid1_unplug(request_queue_t *q)
567 {
568         mddev_t *mddev = q->queuedata;
569
570         unplug_slaves(mddev);
571         md_wakeup_thread(mddev->thread);
572 }
573
574 static int raid1_issue_flush(request_queue_t *q, struct gendisk *disk,
575                              sector_t *error_sector)
576 {
577         mddev_t *mddev = q->queuedata;
578         conf_t *conf = mddev_to_conf(mddev);
579         int i, ret = 0;
580
581         rcu_read_lock();
582         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
583                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
584                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
585                         struct block_device *bdev = rdev->bdev;
586                         request_queue_t *r_queue = bdev_get_queue(bdev);
587
588                         if (!r_queue->issue_flush_fn)
589                                 ret = -EOPNOTSUPP;
590                         else {
591                                 atomic_inc(&rdev->nr_pending);
592                                 rcu_read_unlock();
593                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
594                                                               error_sector);
595                                 rdev_dec_pending(rdev, mddev);
596                                 rcu_read_lock();
597                         }
598                 }
599         }
600         rcu_read_unlock();
601         return ret;
602 }
603
604 /* Barriers....
605  * Sometimes we need to suspend IO while we do something else,
606  * either some resync/recovery, or reconfigure the array.
607  * To do this we raise a 'barrier'.
608  * The 'barrier' is a counter that can be raised multiple times
609  * to count how many activities are happening which preclude
610  * normal IO.
611  * We can only raise the barrier if there is no pending IO.
612  * i.e. if nr_pending == 0.
613  * We choose only to raise the barrier if no-one is waiting for the
614  * barrier to go down.  This means that as soon as an IO request
615  * is ready, no other operations which require a barrier will start
616  * until the IO request has had a chance.
617  *
618  * So: regular IO calls 'wait_barrier'.  When that returns there
619  *    is no backgroup IO happening,  It must arrange to call
620  *    allow_barrier when it has finished its IO.
621  * backgroup IO calls must call raise_barrier.  Once that returns
622  *    there is no normal IO happeing.  It must arrange to call
623  *    lower_barrier when the particular background IO completes.
624  */
625 #define RESYNC_DEPTH 32
626
627 static void raise_barrier(conf_t *conf)
628 {
629         spin_lock_irq(&conf->resync_lock);
630
631         /* Wait until no block IO is waiting */
632         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
633                             conf->resync_lock,
634                             raid1_unplug(conf->mddev->queue));
635
636         /* block any new IO from starting */
637         conf->barrier++;
638
639         /* No wait for all pending IO to complete */
640         wait_event_lock_irq(conf->wait_barrier,
641                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
642                             conf->resync_lock,
643                             raid1_unplug(conf->mddev->queue));
644
645         spin_unlock_irq(&conf->resync_lock);
646 }
647
648 static void lower_barrier(conf_t *conf)
649 {
650         unsigned long flags;
651         spin_lock_irqsave(&conf->resync_lock, flags);
652         conf->barrier--;
653         spin_unlock_irqrestore(&conf->resync_lock, flags);
654         wake_up(&conf->wait_barrier);
655 }
656
657 static void wait_barrier(conf_t *conf)
658 {
659         spin_lock_irq(&conf->resync_lock);
660         if (conf->barrier) {
661                 conf->nr_waiting++;
662                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
663                                     conf->resync_lock,
664                                     raid1_unplug(conf->mddev->queue));
665                 conf->nr_waiting--;
666         }
667         conf->nr_pending++;
668         spin_unlock_irq(&conf->resync_lock);
669 }
670
671 static void allow_barrier(conf_t *conf)
672 {
673         unsigned long flags;
674         spin_lock_irqsave(&conf->resync_lock, flags);
675         conf->nr_pending--;
676         spin_unlock_irqrestore(&conf->resync_lock, flags);
677         wake_up(&conf->wait_barrier);
678 }
679
680 static void freeze_array(conf_t *conf)
681 {
682         /* stop syncio and normal IO and wait for everything to
683          * go quite.
684          * We increment barrier and nr_waiting, and then
685          * wait until barrier+nr_pending match nr_queued+2
686          */
687         spin_lock_irq(&conf->resync_lock);
688         conf->barrier++;
689         conf->nr_waiting++;
690         wait_event_lock_irq(conf->wait_barrier,
691                             conf->barrier+conf->nr_pending == conf->nr_queued+2,
692                             conf->resync_lock,
693                             raid1_unplug(conf->mddev->queue));
694         spin_unlock_irq(&conf->resync_lock);
695 }
696 static void unfreeze_array(conf_t *conf)
697 {
698         /* reverse the effect of the freeze */
699         spin_lock_irq(&conf->resync_lock);
700         conf->barrier--;
701         conf->nr_waiting--;
702         wake_up(&conf->wait_barrier);
703         spin_unlock_irq(&conf->resync_lock);
704 }
705
706
707 /* duplicate the data pages for behind I/O */
708 static struct page **alloc_behind_pages(struct bio *bio)
709 {
710         int i;
711         struct bio_vec *bvec;
712         struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page *),
713                                         GFP_NOIO);
714         if (unlikely(!pages))
715                 goto do_sync_io;
716
717         bio_for_each_segment(bvec, bio, i) {
718                 pages[i] = alloc_page(GFP_NOIO);
719                 if (unlikely(!pages[i]))
720                         goto do_sync_io;
721                 memcpy(kmap(pages[i]) + bvec->bv_offset,
722                         kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
723                 kunmap(pages[i]);
724                 kunmap(bvec->bv_page);
725         }
726
727         return pages;
728
729 do_sync_io:
730         if (pages)
731                 for (i = 0; i < bio->bi_vcnt && pages[i]; i++)
732                         put_page(pages[i]);
733         kfree(pages);
734         PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
735         return NULL;
736 }
737
738 static int make_request(request_queue_t *q, struct bio * bio)
739 {
740         mddev_t *mddev = q->queuedata;
741         conf_t *conf = mddev_to_conf(mddev);
742         mirror_info_t *mirror;
743         r1bio_t *r1_bio;
744         struct bio *read_bio;
745         int i, targets = 0, disks;
746         mdk_rdev_t *rdev;
747         struct bitmap *bitmap = mddev->bitmap;
748         unsigned long flags;
749         struct bio_list bl;
750         struct page **behind_pages = NULL;
751         const int rw = bio_data_dir(bio);
752         int do_barriers;
753
754         /*
755          * Register the new request and wait if the reconstruction
756          * thread has put up a bar for new requests.
757          * Continue immediately if no resync is active currently.
758          * We test barriers_work *after* md_write_start as md_write_start
759          * may cause the first superblock write, and that will check out
760          * if barriers work.
761          */
762
763         md_write_start(mddev, bio); /* wait on superblock update early */
764
765         if (unlikely(!mddev->barriers_work && bio_barrier(bio))) {
766                 if (rw == WRITE)
767                         md_write_end(mddev);
768                 bio_endio(bio, bio->bi_size, -EOPNOTSUPP);
769                 return 0;
770         }
771
772         wait_barrier(conf);
773
774         disk_stat_inc(mddev->gendisk, ios[rw]);
775         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
776
777         /*
778          * make_request() can abort the operation when READA is being
779          * used and no empty request is available.
780          *
781          */
782         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
783
784         r1_bio->master_bio = bio;
785         r1_bio->sectors = bio->bi_size >> 9;
786         r1_bio->state = 0;
787         r1_bio->mddev = mddev;
788         r1_bio->sector = bio->bi_sector;
789
790         if (rw == READ) {
791                 /*
792                  * read balancing logic:
793                  */
794                 int rdisk = read_balance(conf, r1_bio);
795
796                 if (rdisk < 0) {
797                         /* couldn't find anywhere to read from */
798                         raid_end_bio_io(r1_bio);
799                         return 0;
800                 }
801                 mirror = conf->mirrors + rdisk;
802
803                 r1_bio->read_disk = rdisk;
804
805                 read_bio = bio_clone(bio, GFP_NOIO);
806
807                 r1_bio->bios[rdisk] = read_bio;
808
809                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
810                 read_bio->bi_bdev = mirror->rdev->bdev;
811                 read_bio->bi_end_io = raid1_end_read_request;
812                 read_bio->bi_rw = READ;
813                 read_bio->bi_private = r1_bio;
814
815                 generic_make_request(read_bio);
816                 return 0;
817         }
818
819         /*
820          * WRITE:
821          */
822         /* first select target devices under spinlock and
823          * inc refcount on their rdev.  Record them by setting
824          * bios[x] to bio
825          */
826         disks = conf->raid_disks;
827 #if 0
828         { static int first=1;
829         if (first) printk("First Write sector %llu disks %d\n",
830                           (unsigned long long)r1_bio->sector, disks);
831         first = 0;
832         }
833 #endif
834         rcu_read_lock();
835         for (i = 0;  i < disks; i++) {
836                 if ((rdev=rcu_dereference(conf->mirrors[i].rdev)) != NULL &&
837                     !test_bit(Faulty, &rdev->flags)) {
838                         atomic_inc(&rdev->nr_pending);
839                         if (test_bit(Faulty, &rdev->flags)) {
840                                 rdev_dec_pending(rdev, mddev);
841                                 r1_bio->bios[i] = NULL;
842                         } else
843                                 r1_bio->bios[i] = bio;
844                         targets++;
845                 } else
846                         r1_bio->bios[i] = NULL;
847         }
848         rcu_read_unlock();
849
850         BUG_ON(targets == 0); /* we never fail the last device */
851
852         if (targets < conf->raid_disks) {
853                 /* array is degraded, we will not clear the bitmap
854                  * on I/O completion (see raid1_end_write_request) */
855                 set_bit(R1BIO_Degraded, &r1_bio->state);
856         }
857
858         /* do behind I/O ? */
859         if (bitmap &&
860             atomic_read(&bitmap->behind_writes) < bitmap->max_write_behind &&
861             (behind_pages = alloc_behind_pages(bio)) != NULL)
862                 set_bit(R1BIO_BehindIO, &r1_bio->state);
863
864         atomic_set(&r1_bio->remaining, 0);
865         atomic_set(&r1_bio->behind_remaining, 0);
866
867         do_barriers = bio_barrier(bio);
868         if (do_barriers)
869                 set_bit(R1BIO_Barrier, &r1_bio->state);
870
871         bio_list_init(&bl);
872         for (i = 0; i < disks; i++) {
873                 struct bio *mbio;
874                 if (!r1_bio->bios[i])
875                         continue;
876
877                 mbio = bio_clone(bio, GFP_NOIO);
878                 r1_bio->bios[i] = mbio;
879
880                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
881                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
882                 mbio->bi_end_io = raid1_end_write_request;
883                 mbio->bi_rw = WRITE | do_barriers;
884                 mbio->bi_private = r1_bio;
885
886                 if (behind_pages) {
887                         struct bio_vec *bvec;
888                         int j;
889
890                         /* Yes, I really want the '__' version so that
891                          * we clear any unused pointer in the io_vec, rather
892                          * than leave them unchanged.  This is important
893                          * because when we come to free the pages, we won't
894                          * know the originial bi_idx, so we just free
895                          * them all
896                          */
897                         __bio_for_each_segment(bvec, mbio, j, 0)
898                                 bvec->bv_page = behind_pages[j];
899                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
900                                 atomic_inc(&r1_bio->behind_remaining);
901                 }
902
903                 atomic_inc(&r1_bio->remaining);
904
905                 bio_list_add(&bl, mbio);
906         }
907         kfree(behind_pages); /* the behind pages are attached to the bios now */
908
909         bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
910                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
911         spin_lock_irqsave(&conf->device_lock, flags);
912         bio_list_merge(&conf->pending_bio_list, &bl);
913         bio_list_init(&bl);
914
915         blk_plug_device(mddev->queue);
916         spin_unlock_irqrestore(&conf->device_lock, flags);
917
918 #if 0
919         while ((bio = bio_list_pop(&bl)) != NULL)
920                 generic_make_request(bio);
921 #endif
922
923         return 0;
924 }
925
926 static void status(struct seq_file *seq, mddev_t *mddev)
927 {
928         conf_t *conf = mddev_to_conf(mddev);
929         int i;
930
931         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
932                                                 conf->working_disks);
933         for (i = 0; i < conf->raid_disks; i++)
934                 seq_printf(seq, "%s",
935                               conf->mirrors[i].rdev &&
936                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
937         seq_printf(seq, "]");
938 }
939
940
941 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
942 {
943         char b[BDEVNAME_SIZE];
944         conf_t *conf = mddev_to_conf(mddev);
945
946         /*
947          * If it is not operational, then we have already marked it as dead
948          * else if it is the last working disks, ignore the error, let the
949          * next level up know.
950          * else mark the drive as failed
951          */
952         if (test_bit(In_sync, &rdev->flags)
953             && conf->working_disks == 1)
954                 /*
955                  * Don't fail the drive, act as though we were just a
956                  * normal single drive
957                  */
958                 return;
959         if (test_bit(In_sync, &rdev->flags)) {
960                 mddev->degraded++;
961                 conf->working_disks--;
962                 /*
963                  * if recovery is running, make sure it aborts.
964                  */
965                 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
966         }
967         clear_bit(In_sync, &rdev->flags);
968         set_bit(Faulty, &rdev->flags);
969         mddev->sb_dirty = 1;
970         printk(KERN_ALERT "raid1: Disk failure on %s, disabling device. \n"
971                 "       Operation continuing on %d devices\n",
972                 bdevname(rdev->bdev,b), conf->working_disks);
973 }
974
975 static void print_conf(conf_t *conf)
976 {
977         int i;
978         mirror_info_t *tmp;
979
980         printk("RAID1 conf printout:\n");
981         if (!conf) {
982                 printk("(!conf)\n");
983                 return;
984         }
985         printk(" --- wd:%d rd:%d\n", conf->working_disks,
986                 conf->raid_disks);
987
988         for (i = 0; i < conf->raid_disks; i++) {
989                 char b[BDEVNAME_SIZE];
990                 tmp = conf->mirrors + i;
991                 if (tmp->rdev)
992                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
993                                 i, !test_bit(In_sync, &tmp->rdev->flags), !test_bit(Faulty, &tmp->rdev->flags),
994                                 bdevname(tmp->rdev->bdev,b));
995         }
996 }
997
998 static void close_sync(conf_t *conf)
999 {
1000         wait_barrier(conf);
1001         allow_barrier(conf);
1002
1003         mempool_destroy(conf->r1buf_pool);
1004         conf->r1buf_pool = NULL;
1005 }
1006
1007 static int raid1_spare_active(mddev_t *mddev)
1008 {
1009         int i;
1010         conf_t *conf = mddev->private;
1011         mirror_info_t *tmp;
1012
1013         /*
1014          * Find all failed disks within the RAID1 configuration 
1015          * and mark them readable
1016          */
1017         for (i = 0; i < conf->raid_disks; i++) {
1018                 tmp = conf->mirrors + i;
1019                 if (tmp->rdev 
1020                     && !test_bit(Faulty, &tmp->rdev->flags)
1021                     && !test_bit(In_sync, &tmp->rdev->flags)) {
1022                         conf->working_disks++;
1023                         mddev->degraded--;
1024                         set_bit(In_sync, &tmp->rdev->flags);
1025                 }
1026         }
1027
1028         print_conf(conf);
1029         return 0;
1030 }
1031
1032
1033 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1034 {
1035         conf_t *conf = mddev->private;
1036         int found = 0;
1037         int mirror = 0;
1038         mirror_info_t *p;
1039
1040         for (mirror=0; mirror < mddev->raid_disks; mirror++)
1041                 if ( !(p=conf->mirrors+mirror)->rdev) {
1042
1043                         blk_queue_stack_limits(mddev->queue,
1044                                                rdev->bdev->bd_disk->queue);
1045                         /* as we don't honour merge_bvec_fn, we must never risk
1046                          * violating it, so limit ->max_sector to one PAGE, as
1047                          * a one page request is never in violation.
1048                          */
1049                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1050                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
1051                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1052
1053                         p->head_position = 0;
1054                         rdev->raid_disk = mirror;
1055                         found = 1;
1056                         /* As all devices are equivalent, we don't need a full recovery
1057                          * if this was recently any drive of the array
1058                          */
1059                         if (rdev->saved_raid_disk < 0)
1060                                 conf->fullsync = 1;
1061                         rcu_assign_pointer(p->rdev, rdev);
1062                         break;
1063                 }
1064
1065         print_conf(conf);
1066         return found;
1067 }
1068
1069 static int raid1_remove_disk(mddev_t *mddev, int number)
1070 {
1071         conf_t *conf = mddev->private;
1072         int err = 0;
1073         mdk_rdev_t *rdev;
1074         mirror_info_t *p = conf->mirrors+ number;
1075
1076         print_conf(conf);
1077         rdev = p->rdev;
1078         if (rdev) {
1079                 if (test_bit(In_sync, &rdev->flags) ||
1080                     atomic_read(&rdev->nr_pending)) {
1081                         err = -EBUSY;
1082                         goto abort;
1083                 }
1084                 p->rdev = NULL;
1085                 synchronize_rcu();
1086                 if (atomic_read(&rdev->nr_pending)) {
1087                         /* lost the race, try later */
1088                         err = -EBUSY;
1089                         p->rdev = rdev;
1090                 }
1091         }
1092 abort:
1093
1094         print_conf(conf);
1095         return err;
1096 }
1097
1098
1099 static int end_sync_read(struct bio *bio, unsigned int bytes_done, int error)
1100 {
1101         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1102         int i;
1103
1104         if (bio->bi_size)
1105                 return 1;
1106
1107         for (i=r1_bio->mddev->raid_disks; i--; )
1108                 if (r1_bio->bios[i] == bio)
1109                         break;
1110         BUG_ON(i < 0);
1111         update_head_pos(i, r1_bio);
1112         /*
1113          * we have read a block, now it needs to be re-written,
1114          * or re-read if the read failed.
1115          * We don't do much here, just schedule handling by raid1d
1116          */
1117         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1118                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1119
1120         if (atomic_dec_and_test(&r1_bio->remaining))
1121                 reschedule_retry(r1_bio);
1122         return 0;
1123 }
1124
1125 static int end_sync_write(struct bio *bio, unsigned int bytes_done, int error)
1126 {
1127         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1128         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1129         mddev_t *mddev = r1_bio->mddev;
1130         conf_t *conf = mddev_to_conf(mddev);
1131         int i;
1132         int mirror=0;
1133
1134         if (bio->bi_size)
1135                 return 1;
1136
1137         for (i = 0; i < conf->raid_disks; i++)
1138                 if (r1_bio->bios[i] == bio) {
1139                         mirror = i;
1140                         break;
1141                 }
1142         if (!uptodate) {
1143                 int sync_blocks = 0;
1144                 sector_t s = r1_bio->sector;
1145                 long sectors_to_go = r1_bio->sectors;
1146                 /* make sure these bits doesn't get cleared. */
1147                 do {
1148                         bitmap_end_sync(mddev->bitmap, s,
1149                                         &sync_blocks, 1);
1150                         s += sync_blocks;
1151                         sectors_to_go -= sync_blocks;
1152                 } while (sectors_to_go > 0);
1153                 md_error(mddev, conf->mirrors[mirror].rdev);
1154         }
1155
1156         update_head_pos(mirror, r1_bio);
1157
1158         if (atomic_dec_and_test(&r1_bio->remaining)) {
1159                 md_done_sync(mddev, r1_bio->sectors, uptodate);
1160                 put_buf(r1_bio);
1161         }
1162         return 0;
1163 }
1164
1165 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1166 {
1167         conf_t *conf = mddev_to_conf(mddev);
1168         int i;
1169         int disks = conf->raid_disks;
1170         struct bio *bio, *wbio;
1171
1172         bio = r1_bio->bios[r1_bio->read_disk];
1173
1174
1175         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1176                 /* We have read all readable devices.  If we haven't
1177                  * got the block, then there is no hope left.
1178                  * If we have, then we want to do a comparison
1179                  * and skip the write if everything is the same.
1180                  * If any blocks failed to read, then we need to
1181                  * attempt an over-write
1182                  */
1183                 int primary;
1184                 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1185                         for (i=0; i<mddev->raid_disks; i++)
1186                                 if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1187                                         md_error(mddev, conf->mirrors[i].rdev);
1188
1189                         md_done_sync(mddev, r1_bio->sectors, 1);
1190                         put_buf(r1_bio);
1191                         return;
1192                 }
1193                 for (primary=0; primary<mddev->raid_disks; primary++)
1194                         if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1195                             test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1196                                 r1_bio->bios[primary]->bi_end_io = NULL;
1197                                 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1198                                 break;
1199                         }
1200                 r1_bio->read_disk = primary;
1201                 for (i=0; i<mddev->raid_disks; i++)
1202                         if (r1_bio->bios[i]->bi_end_io == end_sync_read &&
1203                             test_bit(BIO_UPTODATE, &r1_bio->bios[i]->bi_flags)) {
1204                                 int j;
1205                                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1206                                 struct bio *pbio = r1_bio->bios[primary];
1207                                 struct bio *sbio = r1_bio->bios[i];
1208                                 for (j = vcnt; j-- ; )
1209                                         if (memcmp(page_address(pbio->bi_io_vec[j].bv_page),
1210                                                    page_address(sbio->bi_io_vec[j].bv_page),
1211                                                    PAGE_SIZE))
1212                                                 break;
1213                                 if (j >= 0)
1214                                         mddev->resync_mismatches += r1_bio->sectors;
1215                                 if (j < 0 || test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
1216                                         sbio->bi_end_io = NULL;
1217                                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1218                                 } else {
1219                                         /* fixup the bio for reuse */
1220                                         sbio->bi_vcnt = vcnt;
1221                                         sbio->bi_size = r1_bio->sectors << 9;
1222                                         sbio->bi_idx = 0;
1223                                         sbio->bi_phys_segments = 0;
1224                                         sbio->bi_hw_segments = 0;
1225                                         sbio->bi_hw_front_size = 0;
1226                                         sbio->bi_hw_back_size = 0;
1227                                         sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1228                                         sbio->bi_flags |= 1 << BIO_UPTODATE;
1229                                         sbio->bi_next = NULL;
1230                                         sbio->bi_sector = r1_bio->sector +
1231                                                 conf->mirrors[i].rdev->data_offset;
1232                                         sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1233                                 }
1234                         }
1235         }
1236         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1237                 /* ouch - failed to read all of that.
1238                  * Try some synchronous reads of other devices to get
1239                  * good data, much like with normal read errors.  Only
1240                  * read into the pages we already have so they we don't
1241                  * need to re-issue the read request.
1242                  * We don't need to freeze the array, because being in an
1243                  * active sync request, there is no normal IO, and
1244                  * no overlapping syncs.
1245                  */
1246                 sector_t sect = r1_bio->sector;
1247                 int sectors = r1_bio->sectors;
1248                 int idx = 0;
1249
1250                 while(sectors) {
1251                         int s = sectors;
1252                         int d = r1_bio->read_disk;
1253                         int success = 0;
1254                         mdk_rdev_t *rdev;
1255
1256                         if (s > (PAGE_SIZE>>9))
1257                                 s = PAGE_SIZE >> 9;
1258                         do {
1259                                 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1260                                         rdev = conf->mirrors[d].rdev;
1261                                         if (sync_page_io(rdev->bdev,
1262                                                          sect + rdev->data_offset,
1263                                                          s<<9,
1264                                                          bio->bi_io_vec[idx].bv_page,
1265                                                          READ)) {
1266                                                 success = 1;
1267                                                 break;
1268                                         }
1269                                 }
1270                                 d++;
1271                                 if (d == conf->raid_disks)
1272                                         d = 0;
1273                         } while (!success && d != r1_bio->read_disk);
1274
1275                         if (success) {
1276                                 int start = d;
1277                                 /* write it back and re-read */
1278                                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1279                                 while (d != r1_bio->read_disk) {
1280                                         if (d == 0)
1281                                                 d = conf->raid_disks;
1282                                         d--;
1283                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1284                                                 continue;
1285                                         rdev = conf->mirrors[d].rdev;
1286                                         atomic_add(s, &rdev->corrected_errors);
1287                                         if (sync_page_io(rdev->bdev,
1288                                                          sect + rdev->data_offset,
1289                                                          s<<9,
1290                                                          bio->bi_io_vec[idx].bv_page,
1291                                                          WRITE) == 0)
1292                                                 md_error(mddev, rdev);
1293                                 }
1294                                 d = start;
1295                                 while (d != r1_bio->read_disk) {
1296                                         if (d == 0)
1297                                                 d = conf->raid_disks;
1298                                         d--;
1299                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1300                                                 continue;
1301                                         rdev = conf->mirrors[d].rdev;
1302                                         if (sync_page_io(rdev->bdev,
1303                                                          sect + rdev->data_offset,
1304                                                          s<<9,
1305                                                          bio->bi_io_vec[idx].bv_page,
1306                                                          READ) == 0)
1307                                                 md_error(mddev, rdev);
1308                                 }
1309                         } else {
1310                                 char b[BDEVNAME_SIZE];
1311                                 /* Cannot read from anywhere, array is toast */
1312                                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1313                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
1314                                        " for block %llu\n",
1315                                        bdevname(bio->bi_bdev,b),
1316                                        (unsigned long long)r1_bio->sector);
1317                                 md_done_sync(mddev, r1_bio->sectors, 0);
1318                                 put_buf(r1_bio);
1319                                 return;
1320                         }
1321                         sectors -= s;
1322                         sect += s;
1323                         idx ++;
1324                 }
1325         }
1326
1327         /*
1328          * schedule writes
1329          */
1330         atomic_set(&r1_bio->remaining, 1);
1331         for (i = 0; i < disks ; i++) {
1332                 wbio = r1_bio->bios[i];
1333                 if (wbio->bi_end_io == NULL ||
1334                     (wbio->bi_end_io == end_sync_read &&
1335                      (i == r1_bio->read_disk ||
1336                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1337                         continue;
1338
1339                 wbio->bi_rw = WRITE;
1340                 wbio->bi_end_io = end_sync_write;
1341                 atomic_inc(&r1_bio->remaining);
1342                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1343
1344                 generic_make_request(wbio);
1345         }
1346
1347         if (atomic_dec_and_test(&r1_bio->remaining)) {
1348                 /* if we're here, all write(s) have completed, so clean up */
1349                 md_done_sync(mddev, r1_bio->sectors, 1);
1350                 put_buf(r1_bio);
1351         }
1352 }
1353
1354 /*
1355  * This is a kernel thread which:
1356  *
1357  *      1.      Retries failed read operations on working mirrors.
1358  *      2.      Updates the raid superblock when problems encounter.
1359  *      3.      Performs writes following reads for array syncronising.
1360  */
1361
1362 static void raid1d(mddev_t *mddev)
1363 {
1364         r1bio_t *r1_bio;
1365         struct bio *bio;
1366         unsigned long flags;
1367         conf_t *conf = mddev_to_conf(mddev);
1368         struct list_head *head = &conf->retry_list;
1369         int unplug=0;
1370         mdk_rdev_t *rdev;
1371
1372         md_check_recovery(mddev);
1373         
1374         for (;;) {
1375                 char b[BDEVNAME_SIZE];
1376                 spin_lock_irqsave(&conf->device_lock, flags);
1377
1378                 if (conf->pending_bio_list.head) {
1379                         bio = bio_list_get(&conf->pending_bio_list);
1380                         blk_remove_plug(mddev->queue);
1381                         spin_unlock_irqrestore(&conf->device_lock, flags);
1382                         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
1383                         if (bitmap_unplug(mddev->bitmap) != 0)
1384                                 printk("%s: bitmap file write failed!\n", mdname(mddev));
1385
1386                         while (bio) { /* submit pending writes */
1387                                 struct bio *next = bio->bi_next;
1388                                 bio->bi_next = NULL;
1389                                 generic_make_request(bio);
1390                                 bio = next;
1391                         }
1392                         unplug = 1;
1393
1394                         continue;
1395                 }
1396
1397                 if (list_empty(head))
1398                         break;
1399                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1400                 list_del(head->prev);
1401                 conf->nr_queued--;
1402                 spin_unlock_irqrestore(&conf->device_lock, flags);
1403
1404                 mddev = r1_bio->mddev;
1405                 conf = mddev_to_conf(mddev);
1406                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1407                         sync_request_write(mddev, r1_bio);
1408                         unplug = 1;
1409                 } else if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
1410                         /* some requests in the r1bio were BIO_RW_BARRIER
1411                          * requests which failed with -EOPNOTSUPP.  Hohumm..
1412                          * Better resubmit without the barrier.
1413                          * We know which devices to resubmit for, because
1414                          * all others have had their bios[] entry cleared.
1415                          * We already have a nr_pending reference on these rdevs.
1416                          */
1417                         int i;
1418                         clear_bit(R1BIO_BarrierRetry, &r1_bio->state);
1419                         clear_bit(R1BIO_Barrier, &r1_bio->state);
1420                         for (i=0; i < conf->raid_disks; i++)
1421                                 if (r1_bio->bios[i])
1422                                         atomic_inc(&r1_bio->remaining);
1423                         for (i=0; i < conf->raid_disks; i++)
1424                                 if (r1_bio->bios[i]) {
1425                                         struct bio_vec *bvec;
1426                                         int j;
1427
1428                                         bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1429                                         /* copy pages from the failed bio, as
1430                                          * this might be a write-behind device */
1431                                         __bio_for_each_segment(bvec, bio, j, 0)
1432                                                 bvec->bv_page = bio_iovec_idx(r1_bio->bios[i], j)->bv_page;
1433                                         bio_put(r1_bio->bios[i]);
1434                                         bio->bi_sector = r1_bio->sector +
1435                                                 conf->mirrors[i].rdev->data_offset;
1436                                         bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1437                                         bio->bi_end_io = raid1_end_write_request;
1438                                         bio->bi_rw = WRITE;
1439                                         bio->bi_private = r1_bio;
1440                                         r1_bio->bios[i] = bio;
1441                                         generic_make_request(bio);
1442                                 }
1443                 } else {
1444                         int disk;
1445
1446                         /* we got a read error. Maybe the drive is bad.  Maybe just
1447                          * the block and we can fix it.
1448                          * We freeze all other IO, and try reading the block from
1449                          * other devices.  When we find one, we re-write
1450                          * and check it that fixes the read error.
1451                          * This is all done synchronously while the array is
1452                          * frozen
1453                          */
1454                         sector_t sect = r1_bio->sector;
1455                         int sectors = r1_bio->sectors;
1456                         freeze_array(conf);
1457                         if (mddev->ro == 0) while(sectors) {
1458                                 int s = sectors;
1459                                 int d = r1_bio->read_disk;
1460                                 int success = 0;
1461
1462                                 if (s > (PAGE_SIZE>>9))
1463                                         s = PAGE_SIZE >> 9;
1464
1465                                 do {
1466                                         rdev = conf->mirrors[d].rdev;
1467                                         if (rdev &&
1468                                             test_bit(In_sync, &rdev->flags) &&
1469                                             sync_page_io(rdev->bdev,
1470                                                          sect + rdev->data_offset,
1471                                                          s<<9,
1472                                                          conf->tmppage, READ))
1473                                                 success = 1;
1474                                         else {
1475                                                 d++;
1476                                                 if (d == conf->raid_disks)
1477                                                         d = 0;
1478                                         }
1479                                 } while (!success && d != r1_bio->read_disk);
1480
1481                                 if (success) {
1482                                         /* write it back and re-read */
1483                                         int start = d;
1484                                         while (d != r1_bio->read_disk) {
1485                                                 if (d==0)
1486                                                         d = conf->raid_disks;
1487                                                 d--;
1488                                                 rdev = conf->mirrors[d].rdev;
1489                                                 atomic_add(s, &rdev->corrected_errors);
1490                                                 if (rdev &&
1491                                                     test_bit(In_sync, &rdev->flags)) {
1492                                                         if (sync_page_io(rdev->bdev,
1493                                                                          sect + rdev->data_offset,
1494                                                                          s<<9, conf->tmppage, WRITE) == 0)
1495                                                                 /* Well, this device is dead */
1496                                                                 md_error(mddev, rdev);
1497                                                 }
1498                                         }
1499                                         d = start;
1500                                         while (d != r1_bio->read_disk) {
1501                                                 if (d==0)
1502                                                         d = conf->raid_disks;
1503                                                 d--;
1504                                                 rdev = conf->mirrors[d].rdev;
1505                                                 if (rdev &&
1506                                                     test_bit(In_sync, &rdev->flags)) {
1507                                                         if (sync_page_io(rdev->bdev,
1508                                                                          sect + rdev->data_offset,
1509                                                                          s<<9, conf->tmppage, READ) == 0)
1510                                                                 /* Well, this device is dead */
1511                                                                 md_error(mddev, rdev);
1512                                                         else
1513                                                                 printk(KERN_INFO "raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
1514                                                                        mdname(mddev), s, (unsigned long long)(sect + rdev->data_offset), bdevname(rdev->bdev, b));
1515                                                 }
1516                                         }
1517                                 } else {
1518                                         /* Cannot read from anywhere -- bye bye array */
1519                                         md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1520                                         break;
1521                                 }
1522                                 sectors -= s;
1523                                 sect += s;
1524                         }
1525
1526                         unfreeze_array(conf);
1527
1528                         bio = r1_bio->bios[r1_bio->read_disk];
1529                         if ((disk=read_balance(conf, r1_bio)) == -1) {
1530                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
1531                                        " read error for block %llu\n",
1532                                        bdevname(bio->bi_bdev,b),
1533                                        (unsigned long long)r1_bio->sector);
1534                                 raid_end_bio_io(r1_bio);
1535                         } else {
1536                                 r1_bio->bios[r1_bio->read_disk] =
1537                                         mddev->ro ? IO_BLOCKED : NULL;
1538                                 r1_bio->read_disk = disk;
1539                                 bio_put(bio);
1540                                 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1541                                 r1_bio->bios[r1_bio->read_disk] = bio;
1542                                 rdev = conf->mirrors[disk].rdev;
1543                                 if (printk_ratelimit())
1544                                         printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
1545                                                " another mirror\n",
1546                                                bdevname(rdev->bdev,b),
1547                                                (unsigned long long)r1_bio->sector);
1548                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1549                                 bio->bi_bdev = rdev->bdev;
1550                                 bio->bi_end_io = raid1_end_read_request;
1551                                 bio->bi_rw = READ;
1552                                 bio->bi_private = r1_bio;
1553                                 unplug = 1;
1554                                 generic_make_request(bio);
1555                         }
1556                 }
1557         }
1558         spin_unlock_irqrestore(&conf->device_lock, flags);
1559         if (unplug)
1560                 unplug_slaves(mddev);
1561 }
1562
1563
1564 static int init_resync(conf_t *conf)
1565 {
1566         int buffs;
1567
1568         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1569         BUG_ON(conf->r1buf_pool);
1570         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1571                                           conf->poolinfo);
1572         if (!conf->r1buf_pool)
1573                 return -ENOMEM;
1574         conf->next_resync = 0;
1575         return 0;
1576 }
1577
1578 /*
1579  * perform a "sync" on one "block"
1580  *
1581  * We need to make sure that no normal I/O request - particularly write
1582  * requests - conflict with active sync requests.
1583  *
1584  * This is achieved by tracking pending requests and a 'barrier' concept
1585  * that can be installed to exclude normal IO requests.
1586  */
1587
1588 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1589 {
1590         conf_t *conf = mddev_to_conf(mddev);
1591         r1bio_t *r1_bio;
1592         struct bio *bio;
1593         sector_t max_sector, nr_sectors;
1594         int disk = -1;
1595         int i;
1596         int wonly = -1;
1597         int write_targets = 0, read_targets = 0;
1598         int sync_blocks;
1599         int still_degraded = 0;
1600
1601         if (!conf->r1buf_pool)
1602         {
1603 /*
1604                 printk("sync start - bitmap %p\n", mddev->bitmap);
1605 */
1606                 if (init_resync(conf))
1607                         return 0;
1608         }
1609
1610         max_sector = mddev->size << 1;
1611         if (sector_nr >= max_sector) {
1612                 /* If we aborted, we need to abort the
1613                  * sync on the 'current' bitmap chunk (there will
1614                  * only be one in raid1 resync.
1615                  * We can find the current addess in mddev->curr_resync
1616                  */
1617                 if (mddev->curr_resync < max_sector) /* aborted */
1618                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1619                                                 &sync_blocks, 1);
1620                 else /* completed sync */
1621                         conf->fullsync = 0;
1622
1623                 bitmap_close_sync(mddev->bitmap);
1624                 close_sync(conf);
1625                 return 0;
1626         }
1627
1628         if (mddev->bitmap == NULL &&
1629             mddev->recovery_cp == MaxSector &&
1630             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1631             conf->fullsync == 0) {
1632                 *skipped = 1;
1633                 return max_sector - sector_nr;
1634         }
1635         /* before building a request, check if we can skip these blocks..
1636          * This call the bitmap_start_sync doesn't actually record anything
1637          */
1638         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1639             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1640                 /* We can skip this block, and probably several more */
1641                 *skipped = 1;
1642                 return sync_blocks;
1643         }
1644         /*
1645          * If there is non-resync activity waiting for a turn,
1646          * and resync is going fast enough,
1647          * then let it though before starting on this new sync request.
1648          */
1649         if (!go_faster && conf->nr_waiting)
1650                 msleep_interruptible(1000);
1651
1652         raise_barrier(conf);
1653
1654         conf->next_resync = sector_nr;
1655
1656         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1657         rcu_read_lock();
1658         /*
1659          * If we get a correctably read error during resync or recovery,
1660          * we might want to read from a different device.  So we
1661          * flag all drives that could conceivably be read from for READ,
1662          * and any others (which will be non-In_sync devices) for WRITE.
1663          * If a read fails, we try reading from something else for which READ
1664          * is OK.
1665          */
1666
1667         r1_bio->mddev = mddev;
1668         r1_bio->sector = sector_nr;
1669         r1_bio->state = 0;
1670         set_bit(R1BIO_IsSync, &r1_bio->state);
1671
1672         for (i=0; i < conf->raid_disks; i++) {
1673                 mdk_rdev_t *rdev;
1674                 bio = r1_bio->bios[i];
1675
1676                 /* take from bio_init */
1677                 bio->bi_next = NULL;
1678                 bio->bi_flags |= 1 << BIO_UPTODATE;
1679                 bio->bi_rw = 0;
1680                 bio->bi_vcnt = 0;
1681                 bio->bi_idx = 0;
1682                 bio->bi_phys_segments = 0;
1683                 bio->bi_hw_segments = 0;
1684                 bio->bi_size = 0;
1685                 bio->bi_end_io = NULL;
1686                 bio->bi_private = NULL;
1687
1688                 rdev = rcu_dereference(conf->mirrors[i].rdev);
1689                 if (rdev == NULL ||
1690                            test_bit(Faulty, &rdev->flags)) {
1691                         still_degraded = 1;
1692                         continue;
1693                 } else if (!test_bit(In_sync, &rdev->flags)) {
1694                         bio->bi_rw = WRITE;
1695                         bio->bi_end_io = end_sync_write;
1696                         write_targets ++;
1697                 } else {
1698                         /* may need to read from here */
1699                         bio->bi_rw = READ;
1700                         bio->bi_end_io = end_sync_read;
1701                         if (test_bit(WriteMostly, &rdev->flags)) {
1702                                 if (wonly < 0)
1703                                         wonly = i;
1704                         } else {
1705                                 if (disk < 0)
1706                                         disk = i;
1707                         }
1708                         read_targets++;
1709                 }
1710                 atomic_inc(&rdev->nr_pending);
1711                 bio->bi_sector = sector_nr + rdev->data_offset;
1712                 bio->bi_bdev = rdev->bdev;
1713                 bio->bi_private = r1_bio;
1714         }
1715         rcu_read_unlock();
1716         if (disk < 0)
1717                 disk = wonly;
1718         r1_bio->read_disk = disk;
1719
1720         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1721                 /* extra read targets are also write targets */
1722                 write_targets += read_targets-1;
1723
1724         if (write_targets == 0 || read_targets == 0) {
1725                 /* There is nowhere to write, so all non-sync
1726                  * drives must be failed - so we are finished
1727                  */
1728                 sector_t rv = max_sector - sector_nr;
1729                 *skipped = 1;
1730                 put_buf(r1_bio);
1731                 return rv;
1732         }
1733
1734         nr_sectors = 0;
1735         sync_blocks = 0;
1736         do {
1737                 struct page *page;
1738                 int len = PAGE_SIZE;
1739                 if (sector_nr + (len>>9) > max_sector)
1740                         len = (max_sector - sector_nr) << 9;
1741                 if (len == 0)
1742                         break;
1743                 if (sync_blocks == 0) {
1744                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1745                                                &sync_blocks, still_degraded) &&
1746                             !conf->fullsync &&
1747                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1748                                 break;
1749                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1750                         if (len > (sync_blocks<<9))
1751                                 len = sync_blocks<<9;
1752                 }
1753
1754                 for (i=0 ; i < conf->raid_disks; i++) {
1755                         bio = r1_bio->bios[i];
1756                         if (bio->bi_end_io) {
1757                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1758                                 if (bio_add_page(bio, page, len, 0) == 0) {
1759                                         /* stop here */
1760                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1761                                         while (i > 0) {
1762                                                 i--;
1763                                                 bio = r1_bio->bios[i];
1764                                                 if (bio->bi_end_io==NULL)
1765                                                         continue;
1766                                                 /* remove last page from this bio */
1767                                                 bio->bi_vcnt--;
1768                                                 bio->bi_size -= len;
1769                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1770                                         }
1771                                         goto bio_full;
1772                                 }
1773                         }
1774                 }
1775                 nr_sectors += len>>9;
1776                 sector_nr += len>>9;
1777                 sync_blocks -= (len>>9);
1778         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1779  bio_full:
1780         r1_bio->sectors = nr_sectors;
1781
1782         /* For a user-requested sync, we read all readable devices and do a
1783          * compare
1784          */
1785         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1786                 atomic_set(&r1_bio->remaining, read_targets);
1787                 for (i=0; i<conf->raid_disks; i++) {
1788                         bio = r1_bio->bios[i];
1789                         if (bio->bi_end_io == end_sync_read) {
1790                                 md_sync_acct(conf->mirrors[i].rdev->bdev, nr_sectors);
1791                                 generic_make_request(bio);
1792                         }
1793                 }
1794         } else {
1795                 atomic_set(&r1_bio->remaining, 1);
1796                 bio = r1_bio->bios[r1_bio->read_disk];
1797                 md_sync_acct(conf->mirrors[r1_bio->read_disk].rdev->bdev,
1798                              nr_sectors);
1799                 generic_make_request(bio);
1800
1801         }
1802
1803         return nr_sectors;
1804 }
1805
1806 static int run(mddev_t *mddev)
1807 {
1808         conf_t *conf;
1809         int i, j, disk_idx;
1810         mirror_info_t *disk;
1811         mdk_rdev_t *rdev;
1812         struct list_head *tmp;
1813
1814         if (mddev->level != 1) {
1815                 printk("raid1: %s: raid level not set to mirroring (%d)\n",
1816                        mdname(mddev), mddev->level);
1817                 goto out;
1818         }
1819         if (mddev->reshape_position != MaxSector) {
1820                 printk("raid1: %s: reshape_position set but not supported\n",
1821                        mdname(mddev));
1822                 goto out;
1823         }
1824         /*
1825          * copy the already verified devices into our private RAID1
1826          * bookkeeping area. [whatever we allocate in run(),
1827          * should be freed in stop()]
1828          */
1829         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1830         mddev->private = conf;
1831         if (!conf)
1832                 goto out_no_mem;
1833
1834         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1835                                  GFP_KERNEL);
1836         if (!conf->mirrors)
1837                 goto out_no_mem;
1838
1839         conf->tmppage = alloc_page(GFP_KERNEL);
1840         if (!conf->tmppage)
1841                 goto out_no_mem;
1842
1843         conf->poolinfo = kmalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1844         if (!conf->poolinfo)
1845                 goto out_no_mem;
1846         conf->poolinfo->mddev = mddev;
1847         conf->poolinfo->raid_disks = mddev->raid_disks;
1848         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1849                                           r1bio_pool_free,
1850                                           conf->poolinfo);
1851         if (!conf->r1bio_pool)
1852                 goto out_no_mem;
1853
1854         ITERATE_RDEV(mddev, rdev, tmp) {
1855                 disk_idx = rdev->raid_disk;
1856                 if (disk_idx >= mddev->raid_disks
1857                     || disk_idx < 0)
1858                         continue;
1859                 disk = conf->mirrors + disk_idx;
1860
1861                 disk->rdev = rdev;
1862
1863                 blk_queue_stack_limits(mddev->queue,
1864                                        rdev->bdev->bd_disk->queue);
1865                 /* as we don't honour merge_bvec_fn, we must never risk
1866                  * violating it, so limit ->max_sector to one PAGE, as
1867                  * a one page request is never in violation.
1868                  */
1869                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1870                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
1871                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1872
1873                 disk->head_position = 0;
1874                 if (!test_bit(Faulty, &rdev->flags) && test_bit(In_sync, &rdev->flags))
1875                         conf->working_disks++;
1876         }
1877         conf->raid_disks = mddev->raid_disks;
1878         conf->mddev = mddev;
1879         spin_lock_init(&conf->device_lock);
1880         INIT_LIST_HEAD(&conf->retry_list);
1881         if (conf->working_disks == 1)
1882                 mddev->recovery_cp = MaxSector;
1883
1884         spin_lock_init(&conf->resync_lock);
1885         init_waitqueue_head(&conf->wait_barrier);
1886
1887         bio_list_init(&conf->pending_bio_list);
1888         bio_list_init(&conf->flushing_bio_list);
1889
1890         if (!conf->working_disks) {
1891                 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
1892                         mdname(mddev));
1893                 goto out_free_conf;
1894         }
1895
1896         mddev->degraded = 0;
1897         for (i = 0; i < conf->raid_disks; i++) {
1898
1899                 disk = conf->mirrors + i;
1900
1901                 if (!disk->rdev ||
1902                     !test_bit(In_sync, &disk->rdev->flags)) {
1903                         disk->head_position = 0;
1904                         mddev->degraded++;
1905                 }
1906         }
1907
1908         /*
1909          * find the first working one and use it as a starting point
1910          * to read balancing.
1911          */
1912         for (j = 0; j < conf->raid_disks &&
1913                      (!conf->mirrors[j].rdev ||
1914                       !test_bit(In_sync, &conf->mirrors[j].rdev->flags)) ; j++)
1915                 /* nothing */;
1916         conf->last_used = j;
1917
1918
1919         mddev->thread = md_register_thread(raid1d, mddev, "%s_raid1");
1920         if (!mddev->thread) {
1921                 printk(KERN_ERR
1922                        "raid1: couldn't allocate thread for %s\n",
1923                        mdname(mddev));
1924                 goto out_free_conf;
1925         }
1926
1927         printk(KERN_INFO 
1928                 "raid1: raid set %s active with %d out of %d mirrors\n",
1929                 mdname(mddev), mddev->raid_disks - mddev->degraded, 
1930                 mddev->raid_disks);
1931         /*
1932          * Ok, everything is just fine now
1933          */
1934         mddev->array_size = mddev->size;
1935
1936         mddev->queue->unplug_fn = raid1_unplug;
1937         mddev->queue->issue_flush_fn = raid1_issue_flush;
1938
1939         return 0;
1940
1941 out_no_mem:
1942         printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
1943                mdname(mddev));
1944
1945 out_free_conf:
1946         if (conf) {
1947                 if (conf->r1bio_pool)
1948                         mempool_destroy(conf->r1bio_pool);
1949                 kfree(conf->mirrors);
1950                 safe_put_page(conf->tmppage);
1951                 kfree(conf->poolinfo);
1952                 kfree(conf);
1953                 mddev->private = NULL;
1954         }
1955 out:
1956         return -EIO;
1957 }
1958
1959 static int stop(mddev_t *mddev)
1960 {
1961         conf_t *conf = mddev_to_conf(mddev);
1962         struct bitmap *bitmap = mddev->bitmap;
1963         int behind_wait = 0;
1964
1965         /* wait for behind writes to complete */
1966         while (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
1967                 behind_wait++;
1968                 printk(KERN_INFO "raid1: behind writes in progress on device %s, waiting to stop (%d)\n", mdname(mddev), behind_wait);
1969                 set_current_state(TASK_UNINTERRUPTIBLE);
1970                 schedule_timeout(HZ); /* wait a second */
1971                 /* need to kick something here to make sure I/O goes? */
1972         }
1973
1974         md_unregister_thread(mddev->thread);
1975         mddev->thread = NULL;
1976         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
1977         if (conf->r1bio_pool)
1978                 mempool_destroy(conf->r1bio_pool);
1979         kfree(conf->mirrors);
1980         kfree(conf->poolinfo);
1981         kfree(conf);
1982         mddev->private = NULL;
1983         return 0;
1984 }
1985
1986 static int raid1_resize(mddev_t *mddev, sector_t sectors)
1987 {
1988         /* no resync is happening, and there is enough space
1989          * on all devices, so we can resize.
1990          * We need to make sure resync covers any new space.
1991          * If the array is shrinking we should possibly wait until
1992          * any io in the removed space completes, but it hardly seems
1993          * worth it.
1994          */
1995         mddev->array_size = sectors>>1;
1996         set_capacity(mddev->gendisk, mddev->array_size << 1);
1997         mddev->changed = 1;
1998         if (mddev->array_size > mddev->size && mddev->recovery_cp == MaxSector) {
1999                 mddev->recovery_cp = mddev->size << 1;
2000                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2001         }
2002         mddev->size = mddev->array_size;
2003         mddev->resync_max_sectors = sectors;
2004         return 0;
2005 }
2006
2007 static int raid1_reshape(mddev_t *mddev)
2008 {
2009         /* We need to:
2010          * 1/ resize the r1bio_pool
2011          * 2/ resize conf->mirrors
2012          *
2013          * We allocate a new r1bio_pool if we can.
2014          * Then raise a device barrier and wait until all IO stops.
2015          * Then resize conf->mirrors and swap in the new r1bio pool.
2016          *
2017          * At the same time, we "pack" the devices so that all the missing
2018          * devices have the higher raid_disk numbers.
2019          */
2020         mempool_t *newpool, *oldpool;
2021         struct pool_info *newpoolinfo;
2022         mirror_info_t *newmirrors;
2023         conf_t *conf = mddev_to_conf(mddev);
2024         int cnt, raid_disks;
2025
2026         int d, d2;
2027
2028         /* Cannot change chunk_size, layout, or level */
2029         if (mddev->chunk_size != mddev->new_chunk ||
2030             mddev->layout != mddev->new_layout ||
2031             mddev->level != mddev->new_level) {
2032                 mddev->new_chunk = mddev->chunk_size;
2033                 mddev->new_layout = mddev->layout;
2034                 mddev->new_level = mddev->level;
2035                 return -EINVAL;
2036         }
2037
2038         raid_disks = mddev->raid_disks + mddev->delta_disks;
2039
2040         if (raid_disks < conf->raid_disks) {
2041                 cnt=0;
2042                 for (d= 0; d < conf->raid_disks; d++)
2043                         if (conf->mirrors[d].rdev)
2044                                 cnt++;
2045                 if (cnt > raid_disks)
2046                         return -EBUSY;
2047         }
2048
2049         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2050         if (!newpoolinfo)
2051                 return -ENOMEM;
2052         newpoolinfo->mddev = mddev;
2053         newpoolinfo->raid_disks = raid_disks;
2054
2055         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2056                                  r1bio_pool_free, newpoolinfo);
2057         if (!newpool) {
2058                 kfree(newpoolinfo);
2059                 return -ENOMEM;
2060         }
2061         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2062         if (!newmirrors) {
2063                 kfree(newpoolinfo);
2064                 mempool_destroy(newpool);
2065                 return -ENOMEM;
2066         }
2067
2068         raise_barrier(conf);
2069
2070         /* ok, everything is stopped */
2071         oldpool = conf->r1bio_pool;
2072         conf->r1bio_pool = newpool;
2073
2074         for (d=d2=0; d < conf->raid_disks; d++)
2075                 if (conf->mirrors[d].rdev) {
2076                         conf->mirrors[d].rdev->raid_disk = d2;
2077                         newmirrors[d2++].rdev = conf->mirrors[d].rdev;
2078                 }
2079         kfree(conf->mirrors);
2080         conf->mirrors = newmirrors;
2081         kfree(conf->poolinfo);
2082         conf->poolinfo = newpoolinfo;
2083
2084         mddev->degraded += (raid_disks - conf->raid_disks);
2085         conf->raid_disks = mddev->raid_disks = raid_disks;
2086         mddev->delta_disks = 0;
2087
2088         conf->last_used = 0; /* just make sure it is in-range */
2089         lower_barrier(conf);
2090
2091         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2092         md_wakeup_thread(mddev->thread);
2093
2094         mempool_destroy(oldpool);
2095         return 0;
2096 }
2097
2098 static void raid1_quiesce(mddev_t *mddev, int state)
2099 {
2100         conf_t *conf = mddev_to_conf(mddev);
2101
2102         switch(state) {
2103         case 1:
2104                 raise_barrier(conf);
2105                 break;
2106         case 0:
2107                 lower_barrier(conf);
2108                 break;
2109         }
2110 }
2111
2112
2113 static struct mdk_personality raid1_personality =
2114 {
2115         .name           = "raid1",
2116         .level          = 1,
2117         .owner          = THIS_MODULE,
2118         .make_request   = make_request,
2119         .run            = run,
2120         .stop           = stop,
2121         .status         = status,
2122         .error_handler  = error,
2123         .hot_add_disk   = raid1_add_disk,
2124         .hot_remove_disk= raid1_remove_disk,
2125         .spare_active   = raid1_spare_active,
2126         .sync_request   = sync_request,
2127         .resize         = raid1_resize,
2128         .check_reshape  = raid1_reshape,
2129         .quiesce        = raid1_quiesce,
2130 };
2131
2132 static int __init raid_init(void)
2133 {
2134         return register_md_personality(&raid1_personality);
2135 }
2136
2137 static void raid_exit(void)
2138 {
2139         unregister_md_personality(&raid1_personality);
2140 }
2141
2142 module_init(raid_init);
2143 module_exit(raid_exit);
2144 MODULE_LICENSE("GPL");
2145 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2146 MODULE_ALIAS("md-raid1");
2147 MODULE_ALIAS("md-level-1");