Pull esi-support into release branch
[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         rcu_read_lock();
934         for (i = 0; i < conf->raid_disks; i++) {
935                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
936                 seq_printf(seq, "%s",
937                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
938         }
939         rcu_read_unlock();
940         seq_printf(seq, "]");
941 }
942
943
944 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
945 {
946         char b[BDEVNAME_SIZE];
947         conf_t *conf = mddev_to_conf(mddev);
948
949         /*
950          * If it is not operational, then we have already marked it as dead
951          * else if it is the last working disks, ignore the error, let the
952          * next level up know.
953          * else mark the drive as failed
954          */
955         if (test_bit(In_sync, &rdev->flags)
956             && conf->working_disks == 1)
957                 /*
958                  * Don't fail the drive, act as though we were just a
959                  * normal single drive
960                  */
961                 return;
962         if (test_bit(In_sync, &rdev->flags)) {
963                 mddev->degraded++;
964                 conf->working_disks--;
965                 /*
966                  * if recovery is running, make sure it aborts.
967                  */
968                 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
969         }
970         clear_bit(In_sync, &rdev->flags);
971         set_bit(Faulty, &rdev->flags);
972         mddev->sb_dirty = 1;
973         printk(KERN_ALERT "raid1: Disk failure on %s, disabling device. \n"
974                 "       Operation continuing on %d devices\n",
975                 bdevname(rdev->bdev,b), conf->working_disks);
976 }
977
978 static void print_conf(conf_t *conf)
979 {
980         int i;
981
982         printk("RAID1 conf printout:\n");
983         if (!conf) {
984                 printk("(!conf)\n");
985                 return;
986         }
987         printk(" --- wd:%d rd:%d\n", conf->working_disks,
988                 conf->raid_disks);
989
990         rcu_read_lock();
991         for (i = 0; i < conf->raid_disks; i++) {
992                 char b[BDEVNAME_SIZE];
993                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
994                 if (rdev)
995                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
996                                i, !test_bit(In_sync, &rdev->flags),
997                                !test_bit(Faulty, &rdev->flags),
998                                bdevname(rdev->bdev,b));
999         }
1000         rcu_read_unlock();
1001 }
1002
1003 static void close_sync(conf_t *conf)
1004 {
1005         wait_barrier(conf);
1006         allow_barrier(conf);
1007
1008         mempool_destroy(conf->r1buf_pool);
1009         conf->r1buf_pool = NULL;
1010 }
1011
1012 static int raid1_spare_active(mddev_t *mddev)
1013 {
1014         int i;
1015         conf_t *conf = mddev->private;
1016
1017         /*
1018          * Find all failed disks within the RAID1 configuration 
1019          * and mark them readable.
1020          * Called under mddev lock, so rcu protection not needed.
1021          */
1022         for (i = 0; i < conf->raid_disks; i++) {
1023                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1024                 if (rdev
1025                     && !test_bit(Faulty, &rdev->flags)
1026                     && !test_bit(In_sync, &rdev->flags)) {
1027                         conf->working_disks++;
1028                         mddev->degraded--;
1029                         set_bit(In_sync, &rdev->flags);
1030                 }
1031         }
1032
1033         print_conf(conf);
1034         return 0;
1035 }
1036
1037
1038 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1039 {
1040         conf_t *conf = mddev->private;
1041         int found = 0;
1042         int mirror = 0;
1043         mirror_info_t *p;
1044
1045         for (mirror=0; mirror < mddev->raid_disks; mirror++)
1046                 if ( !(p=conf->mirrors+mirror)->rdev) {
1047
1048                         blk_queue_stack_limits(mddev->queue,
1049                                                rdev->bdev->bd_disk->queue);
1050                         /* as we don't honour merge_bvec_fn, we must never risk
1051                          * violating it, so limit ->max_sector to one PAGE, as
1052                          * a one page request is never in violation.
1053                          */
1054                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1055                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
1056                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1057
1058                         p->head_position = 0;
1059                         rdev->raid_disk = mirror;
1060                         found = 1;
1061                         /* As all devices are equivalent, we don't need a full recovery
1062                          * if this was recently any drive of the array
1063                          */
1064                         if (rdev->saved_raid_disk < 0)
1065                                 conf->fullsync = 1;
1066                         rcu_assign_pointer(p->rdev, rdev);
1067                         break;
1068                 }
1069
1070         print_conf(conf);
1071         return found;
1072 }
1073
1074 static int raid1_remove_disk(mddev_t *mddev, int number)
1075 {
1076         conf_t *conf = mddev->private;
1077         int err = 0;
1078         mdk_rdev_t *rdev;
1079         mirror_info_t *p = conf->mirrors+ number;
1080
1081         print_conf(conf);
1082         rdev = p->rdev;
1083         if (rdev) {
1084                 if (test_bit(In_sync, &rdev->flags) ||
1085                     atomic_read(&rdev->nr_pending)) {
1086                         err = -EBUSY;
1087                         goto abort;
1088                 }
1089                 p->rdev = NULL;
1090                 synchronize_rcu();
1091                 if (atomic_read(&rdev->nr_pending)) {
1092                         /* lost the race, try later */
1093                         err = -EBUSY;
1094                         p->rdev = rdev;
1095                 }
1096         }
1097 abort:
1098
1099         print_conf(conf);
1100         return err;
1101 }
1102
1103
1104 static int end_sync_read(struct bio *bio, unsigned int bytes_done, int error)
1105 {
1106         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1107         int i;
1108
1109         if (bio->bi_size)
1110                 return 1;
1111
1112         for (i=r1_bio->mddev->raid_disks; i--; )
1113                 if (r1_bio->bios[i] == bio)
1114                         break;
1115         BUG_ON(i < 0);
1116         update_head_pos(i, r1_bio);
1117         /*
1118          * we have read a block, now it needs to be re-written,
1119          * or re-read if the read failed.
1120          * We don't do much here, just schedule handling by raid1d
1121          */
1122         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1123                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1124
1125         if (atomic_dec_and_test(&r1_bio->remaining))
1126                 reschedule_retry(r1_bio);
1127         return 0;
1128 }
1129
1130 static int end_sync_write(struct bio *bio, unsigned int bytes_done, int error)
1131 {
1132         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1133         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1134         mddev_t *mddev = r1_bio->mddev;
1135         conf_t *conf = mddev_to_conf(mddev);
1136         int i;
1137         int mirror=0;
1138
1139         if (bio->bi_size)
1140                 return 1;
1141
1142         for (i = 0; i < conf->raid_disks; i++)
1143                 if (r1_bio->bios[i] == bio) {
1144                         mirror = i;
1145                         break;
1146                 }
1147         if (!uptodate) {
1148                 int sync_blocks = 0;
1149                 sector_t s = r1_bio->sector;
1150                 long sectors_to_go = r1_bio->sectors;
1151                 /* make sure these bits doesn't get cleared. */
1152                 do {
1153                         bitmap_end_sync(mddev->bitmap, s,
1154                                         &sync_blocks, 1);
1155                         s += sync_blocks;
1156                         sectors_to_go -= sync_blocks;
1157                 } while (sectors_to_go > 0);
1158                 md_error(mddev, conf->mirrors[mirror].rdev);
1159         }
1160
1161         update_head_pos(mirror, r1_bio);
1162
1163         if (atomic_dec_and_test(&r1_bio->remaining)) {
1164                 md_done_sync(mddev, r1_bio->sectors, uptodate);
1165                 put_buf(r1_bio);
1166         }
1167         return 0;
1168 }
1169
1170 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1171 {
1172         conf_t *conf = mddev_to_conf(mddev);
1173         int i;
1174         int disks = conf->raid_disks;
1175         struct bio *bio, *wbio;
1176
1177         bio = r1_bio->bios[r1_bio->read_disk];
1178
1179
1180         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1181                 /* We have read all readable devices.  If we haven't
1182                  * got the block, then there is no hope left.
1183                  * If we have, then we want to do a comparison
1184                  * and skip the write if everything is the same.
1185                  * If any blocks failed to read, then we need to
1186                  * attempt an over-write
1187                  */
1188                 int primary;
1189                 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1190                         for (i=0; i<mddev->raid_disks; i++)
1191                                 if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1192                                         md_error(mddev, conf->mirrors[i].rdev);
1193
1194                         md_done_sync(mddev, r1_bio->sectors, 1);
1195                         put_buf(r1_bio);
1196                         return;
1197                 }
1198                 for (primary=0; primary<mddev->raid_disks; primary++)
1199                         if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1200                             test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1201                                 r1_bio->bios[primary]->bi_end_io = NULL;
1202                                 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1203                                 break;
1204                         }
1205                 r1_bio->read_disk = primary;
1206                 for (i=0; i<mddev->raid_disks; i++)
1207                         if (r1_bio->bios[i]->bi_end_io == end_sync_read &&
1208                             test_bit(BIO_UPTODATE, &r1_bio->bios[i]->bi_flags)) {
1209                                 int j;
1210                                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1211                                 struct bio *pbio = r1_bio->bios[primary];
1212                                 struct bio *sbio = r1_bio->bios[i];
1213                                 for (j = vcnt; j-- ; )
1214                                         if (memcmp(page_address(pbio->bi_io_vec[j].bv_page),
1215                                                    page_address(sbio->bi_io_vec[j].bv_page),
1216                                                    PAGE_SIZE))
1217                                                 break;
1218                                 if (j >= 0)
1219                                         mddev->resync_mismatches += r1_bio->sectors;
1220                                 if (j < 0 || test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
1221                                         sbio->bi_end_io = NULL;
1222                                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1223                                 } else {
1224                                         /* fixup the bio for reuse */
1225                                         sbio->bi_vcnt = vcnt;
1226                                         sbio->bi_size = r1_bio->sectors << 9;
1227                                         sbio->bi_idx = 0;
1228                                         sbio->bi_phys_segments = 0;
1229                                         sbio->bi_hw_segments = 0;
1230                                         sbio->bi_hw_front_size = 0;
1231                                         sbio->bi_hw_back_size = 0;
1232                                         sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1233                                         sbio->bi_flags |= 1 << BIO_UPTODATE;
1234                                         sbio->bi_next = NULL;
1235                                         sbio->bi_sector = r1_bio->sector +
1236                                                 conf->mirrors[i].rdev->data_offset;
1237                                         sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1238                                 }
1239                         }
1240         }
1241         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1242                 /* ouch - failed to read all of that.
1243                  * Try some synchronous reads of other devices to get
1244                  * good data, much like with normal read errors.  Only
1245                  * read into the pages we already have so we don't
1246                  * need to re-issue the read request.
1247                  * We don't need to freeze the array, because being in an
1248                  * active sync request, there is no normal IO, and
1249                  * no overlapping syncs.
1250                  */
1251                 sector_t sect = r1_bio->sector;
1252                 int sectors = r1_bio->sectors;
1253                 int idx = 0;
1254
1255                 while(sectors) {
1256                         int s = sectors;
1257                         int d = r1_bio->read_disk;
1258                         int success = 0;
1259                         mdk_rdev_t *rdev;
1260
1261                         if (s > (PAGE_SIZE>>9))
1262                                 s = PAGE_SIZE >> 9;
1263                         do {
1264                                 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1265                                         /* No rcu protection needed here devices
1266                                          * can only be removed when no resync is
1267                                          * active, and resync is currently active
1268                                          */
1269                                         rdev = conf->mirrors[d].rdev;
1270                                         if (sync_page_io(rdev->bdev,
1271                                                          sect + rdev->data_offset,
1272                                                          s<<9,
1273                                                          bio->bi_io_vec[idx].bv_page,
1274                                                          READ)) {
1275                                                 success = 1;
1276                                                 break;
1277                                         }
1278                                 }
1279                                 d++;
1280                                 if (d == conf->raid_disks)
1281                                         d = 0;
1282                         } while (!success && d != r1_bio->read_disk);
1283
1284                         if (success) {
1285                                 int start = d;
1286                                 /* write it back and re-read */
1287                                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1288                                 while (d != r1_bio->read_disk) {
1289                                         if (d == 0)
1290                                                 d = conf->raid_disks;
1291                                         d--;
1292                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1293                                                 continue;
1294                                         rdev = conf->mirrors[d].rdev;
1295                                         atomic_add(s, &rdev->corrected_errors);
1296                                         if (sync_page_io(rdev->bdev,
1297                                                          sect + rdev->data_offset,
1298                                                          s<<9,
1299                                                          bio->bi_io_vec[idx].bv_page,
1300                                                          WRITE) == 0)
1301                                                 md_error(mddev, rdev);
1302                                 }
1303                                 d = start;
1304                                 while (d != r1_bio->read_disk) {
1305                                         if (d == 0)
1306                                                 d = conf->raid_disks;
1307                                         d--;
1308                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1309                                                 continue;
1310                                         rdev = conf->mirrors[d].rdev;
1311                                         if (sync_page_io(rdev->bdev,
1312                                                          sect + rdev->data_offset,
1313                                                          s<<9,
1314                                                          bio->bi_io_vec[idx].bv_page,
1315                                                          READ) == 0)
1316                                                 md_error(mddev, rdev);
1317                                 }
1318                         } else {
1319                                 char b[BDEVNAME_SIZE];
1320                                 /* Cannot read from anywhere, array is toast */
1321                                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1322                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
1323                                        " for block %llu\n",
1324                                        bdevname(bio->bi_bdev,b),
1325                                        (unsigned long long)r1_bio->sector);
1326                                 md_done_sync(mddev, r1_bio->sectors, 0);
1327                                 put_buf(r1_bio);
1328                                 return;
1329                         }
1330                         sectors -= s;
1331                         sect += s;
1332                         idx ++;
1333                 }
1334         }
1335
1336         /*
1337          * schedule writes
1338          */
1339         atomic_set(&r1_bio->remaining, 1);
1340         for (i = 0; i < disks ; i++) {
1341                 wbio = r1_bio->bios[i];
1342                 if (wbio->bi_end_io == NULL ||
1343                     (wbio->bi_end_io == end_sync_read &&
1344                      (i == r1_bio->read_disk ||
1345                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1346                         continue;
1347
1348                 wbio->bi_rw = WRITE;
1349                 wbio->bi_end_io = end_sync_write;
1350                 atomic_inc(&r1_bio->remaining);
1351                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1352
1353                 generic_make_request(wbio);
1354         }
1355
1356         if (atomic_dec_and_test(&r1_bio->remaining)) {
1357                 /* if we're here, all write(s) have completed, so clean up */
1358                 md_done_sync(mddev, r1_bio->sectors, 1);
1359                 put_buf(r1_bio);
1360         }
1361 }
1362
1363 /*
1364  * This is a kernel thread which:
1365  *
1366  *      1.      Retries failed read operations on working mirrors.
1367  *      2.      Updates the raid superblock when problems encounter.
1368  *      3.      Performs writes following reads for array syncronising.
1369  */
1370
1371 static void raid1d(mddev_t *mddev)
1372 {
1373         r1bio_t *r1_bio;
1374         struct bio *bio;
1375         unsigned long flags;
1376         conf_t *conf = mddev_to_conf(mddev);
1377         struct list_head *head = &conf->retry_list;
1378         int unplug=0;
1379         mdk_rdev_t *rdev;
1380
1381         md_check_recovery(mddev);
1382         
1383         for (;;) {
1384                 char b[BDEVNAME_SIZE];
1385                 spin_lock_irqsave(&conf->device_lock, flags);
1386
1387                 if (conf->pending_bio_list.head) {
1388                         bio = bio_list_get(&conf->pending_bio_list);
1389                         blk_remove_plug(mddev->queue);
1390                         spin_unlock_irqrestore(&conf->device_lock, flags);
1391                         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
1392                         if (bitmap_unplug(mddev->bitmap) != 0)
1393                                 printk("%s: bitmap file write failed!\n", mdname(mddev));
1394
1395                         while (bio) { /* submit pending writes */
1396                                 struct bio *next = bio->bi_next;
1397                                 bio->bi_next = NULL;
1398                                 generic_make_request(bio);
1399                                 bio = next;
1400                         }
1401                         unplug = 1;
1402
1403                         continue;
1404                 }
1405
1406                 if (list_empty(head))
1407                         break;
1408                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1409                 list_del(head->prev);
1410                 conf->nr_queued--;
1411                 spin_unlock_irqrestore(&conf->device_lock, flags);
1412
1413                 mddev = r1_bio->mddev;
1414                 conf = mddev_to_conf(mddev);
1415                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1416                         sync_request_write(mddev, r1_bio);
1417                         unplug = 1;
1418                 } else if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
1419                         /* some requests in the r1bio were BIO_RW_BARRIER
1420                          * requests which failed with -EOPNOTSUPP.  Hohumm..
1421                          * Better resubmit without the barrier.
1422                          * We know which devices to resubmit for, because
1423                          * all others have had their bios[] entry cleared.
1424                          * We already have a nr_pending reference on these rdevs.
1425                          */
1426                         int i;
1427                         clear_bit(R1BIO_BarrierRetry, &r1_bio->state);
1428                         clear_bit(R1BIO_Barrier, &r1_bio->state);
1429                         for (i=0; i < conf->raid_disks; i++)
1430                                 if (r1_bio->bios[i])
1431                                         atomic_inc(&r1_bio->remaining);
1432                         for (i=0; i < conf->raid_disks; i++)
1433                                 if (r1_bio->bios[i]) {
1434                                         struct bio_vec *bvec;
1435                                         int j;
1436
1437                                         bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1438                                         /* copy pages from the failed bio, as
1439                                          * this might be a write-behind device */
1440                                         __bio_for_each_segment(bvec, bio, j, 0)
1441                                                 bvec->bv_page = bio_iovec_idx(r1_bio->bios[i], j)->bv_page;
1442                                         bio_put(r1_bio->bios[i]);
1443                                         bio->bi_sector = r1_bio->sector +
1444                                                 conf->mirrors[i].rdev->data_offset;
1445                                         bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1446                                         bio->bi_end_io = raid1_end_write_request;
1447                                         bio->bi_rw = WRITE;
1448                                         bio->bi_private = r1_bio;
1449                                         r1_bio->bios[i] = bio;
1450                                         generic_make_request(bio);
1451                                 }
1452                 } else {
1453                         int disk;
1454
1455                         /* we got a read error. Maybe the drive is bad.  Maybe just
1456                          * the block and we can fix it.
1457                          * We freeze all other IO, and try reading the block from
1458                          * other devices.  When we find one, we re-write
1459                          * and check it that fixes the read error.
1460                          * This is all done synchronously while the array is
1461                          * frozen
1462                          */
1463                         sector_t sect = r1_bio->sector;
1464                         int sectors = r1_bio->sectors;
1465                         freeze_array(conf);
1466                         if (mddev->ro == 0) while(sectors) {
1467                                 int s = sectors;
1468                                 int d = r1_bio->read_disk;
1469                                 int success = 0;
1470
1471                                 if (s > (PAGE_SIZE>>9))
1472                                         s = PAGE_SIZE >> 9;
1473
1474                                 do {
1475                                         /* Note: no rcu protection needed here
1476                                          * as this is synchronous in the raid1d thread
1477                                          * which is the thread that might remove
1478                                          * a device.  If raid1d ever becomes multi-threaded....
1479                                          */
1480                                         rdev = conf->mirrors[d].rdev;
1481                                         if (rdev &&
1482                                             test_bit(In_sync, &rdev->flags) &&
1483                                             sync_page_io(rdev->bdev,
1484                                                          sect + rdev->data_offset,
1485                                                          s<<9,
1486                                                          conf->tmppage, READ))
1487                                                 success = 1;
1488                                         else {
1489                                                 d++;
1490                                                 if (d == conf->raid_disks)
1491                                                         d = 0;
1492                                         }
1493                                 } while (!success && d != r1_bio->read_disk);
1494
1495                                 if (success) {
1496                                         /* write it back and re-read */
1497                                         int start = d;
1498                                         while (d != r1_bio->read_disk) {
1499                                                 if (d==0)
1500                                                         d = conf->raid_disks;
1501                                                 d--;
1502                                                 rdev = conf->mirrors[d].rdev;
1503                                                 if (rdev &&
1504                                                     test_bit(In_sync, &rdev->flags)) {
1505                                                         if (sync_page_io(rdev->bdev,
1506                                                                          sect + rdev->data_offset,
1507                                                                          s<<9, conf->tmppage, WRITE) == 0)
1508                                                                 /* Well, this device is dead */
1509                                                                 md_error(mddev, rdev);
1510                                                 }
1511                                         }
1512                                         d = start;
1513                                         while (d != r1_bio->read_disk) {
1514                                                 if (d==0)
1515                                                         d = conf->raid_disks;
1516                                                 d--;
1517                                                 rdev = conf->mirrors[d].rdev;
1518                                                 if (rdev &&
1519                                                     test_bit(In_sync, &rdev->flags)) {
1520                                                         if (sync_page_io(rdev->bdev,
1521                                                                          sect + rdev->data_offset,
1522                                                                          s<<9, conf->tmppage, READ) == 0)
1523                                                                 /* Well, this device is dead */
1524                                                                 md_error(mddev, rdev);
1525                                                         else {
1526                                                                 atomic_add(s, &rdev->corrected_errors);
1527                                                                 printk(KERN_INFO "raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
1528                                                                        mdname(mddev), s, (unsigned long long)(sect + rdev->data_offset), bdevname(rdev->bdev, b));
1529                                                         }
1530                                                 }
1531                                         }
1532                                 } else {
1533                                         /* Cannot read from anywhere -- bye bye array */
1534                                         md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1535                                         break;
1536                                 }
1537                                 sectors -= s;
1538                                 sect += s;
1539                         }
1540
1541                         unfreeze_array(conf);
1542
1543                         bio = r1_bio->bios[r1_bio->read_disk];
1544                         if ((disk=read_balance(conf, r1_bio)) == -1) {
1545                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
1546                                        " read error for block %llu\n",
1547                                        bdevname(bio->bi_bdev,b),
1548                                        (unsigned long long)r1_bio->sector);
1549                                 raid_end_bio_io(r1_bio);
1550                         } else {
1551                                 r1_bio->bios[r1_bio->read_disk] =
1552                                         mddev->ro ? IO_BLOCKED : NULL;
1553                                 r1_bio->read_disk = disk;
1554                                 bio_put(bio);
1555                                 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1556                                 r1_bio->bios[r1_bio->read_disk] = bio;
1557                                 rdev = conf->mirrors[disk].rdev;
1558                                 if (printk_ratelimit())
1559                                         printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
1560                                                " another mirror\n",
1561                                                bdevname(rdev->bdev,b),
1562                                                (unsigned long long)r1_bio->sector);
1563                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1564                                 bio->bi_bdev = rdev->bdev;
1565                                 bio->bi_end_io = raid1_end_read_request;
1566                                 bio->bi_rw = READ;
1567                                 bio->bi_private = r1_bio;
1568                                 unplug = 1;
1569                                 generic_make_request(bio);
1570                         }
1571                 }
1572         }
1573         spin_unlock_irqrestore(&conf->device_lock, flags);
1574         if (unplug)
1575                 unplug_slaves(mddev);
1576 }
1577
1578
1579 static int init_resync(conf_t *conf)
1580 {
1581         int buffs;
1582
1583         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1584         BUG_ON(conf->r1buf_pool);
1585         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1586                                           conf->poolinfo);
1587         if (!conf->r1buf_pool)
1588                 return -ENOMEM;
1589         conf->next_resync = 0;
1590         return 0;
1591 }
1592
1593 /*
1594  * perform a "sync" on one "block"
1595  *
1596  * We need to make sure that no normal I/O request - particularly write
1597  * requests - conflict with active sync requests.
1598  *
1599  * This is achieved by tracking pending requests and a 'barrier' concept
1600  * that can be installed to exclude normal IO requests.
1601  */
1602
1603 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1604 {
1605         conf_t *conf = mddev_to_conf(mddev);
1606         r1bio_t *r1_bio;
1607         struct bio *bio;
1608         sector_t max_sector, nr_sectors;
1609         int disk = -1;
1610         int i;
1611         int wonly = -1;
1612         int write_targets = 0, read_targets = 0;
1613         int sync_blocks;
1614         int still_degraded = 0;
1615
1616         if (!conf->r1buf_pool)
1617         {
1618 /*
1619                 printk("sync start - bitmap %p\n", mddev->bitmap);
1620 */
1621                 if (init_resync(conf))
1622                         return 0;
1623         }
1624
1625         max_sector = mddev->size << 1;
1626         if (sector_nr >= max_sector) {
1627                 /* If we aborted, we need to abort the
1628                  * sync on the 'current' bitmap chunk (there will
1629                  * only be one in raid1 resync.
1630                  * We can find the current addess in mddev->curr_resync
1631                  */
1632                 if (mddev->curr_resync < max_sector) /* aborted */
1633                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1634                                                 &sync_blocks, 1);
1635                 else /* completed sync */
1636                         conf->fullsync = 0;
1637
1638                 bitmap_close_sync(mddev->bitmap);
1639                 close_sync(conf);
1640                 return 0;
1641         }
1642
1643         if (mddev->bitmap == NULL &&
1644             mddev->recovery_cp == MaxSector &&
1645             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1646             conf->fullsync == 0) {
1647                 *skipped = 1;
1648                 return max_sector - sector_nr;
1649         }
1650         /* before building a request, check if we can skip these blocks..
1651          * This call the bitmap_start_sync doesn't actually record anything
1652          */
1653         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1654             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1655                 /* We can skip this block, and probably several more */
1656                 *skipped = 1;
1657                 return sync_blocks;
1658         }
1659         /*
1660          * If there is non-resync activity waiting for a turn,
1661          * and resync is going fast enough,
1662          * then let it though before starting on this new sync request.
1663          */
1664         if (!go_faster && conf->nr_waiting)
1665                 msleep_interruptible(1000);
1666
1667         raise_barrier(conf);
1668
1669         conf->next_resync = sector_nr;
1670
1671         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1672         rcu_read_lock();
1673         /*
1674          * If we get a correctably read error during resync or recovery,
1675          * we might want to read from a different device.  So we
1676          * flag all drives that could conceivably be read from for READ,
1677          * and any others (which will be non-In_sync devices) for WRITE.
1678          * If a read fails, we try reading from something else for which READ
1679          * is OK.
1680          */
1681
1682         r1_bio->mddev = mddev;
1683         r1_bio->sector = sector_nr;
1684         r1_bio->state = 0;
1685         set_bit(R1BIO_IsSync, &r1_bio->state);
1686
1687         for (i=0; i < conf->raid_disks; i++) {
1688                 mdk_rdev_t *rdev;
1689                 bio = r1_bio->bios[i];
1690
1691                 /* take from bio_init */
1692                 bio->bi_next = NULL;
1693                 bio->bi_flags |= 1 << BIO_UPTODATE;
1694                 bio->bi_rw = 0;
1695                 bio->bi_vcnt = 0;
1696                 bio->bi_idx = 0;
1697                 bio->bi_phys_segments = 0;
1698                 bio->bi_hw_segments = 0;
1699                 bio->bi_size = 0;
1700                 bio->bi_end_io = NULL;
1701                 bio->bi_private = NULL;
1702
1703                 rdev = rcu_dereference(conf->mirrors[i].rdev);
1704                 if (rdev == NULL ||
1705                            test_bit(Faulty, &rdev->flags)) {
1706                         still_degraded = 1;
1707                         continue;
1708                 } else if (!test_bit(In_sync, &rdev->flags)) {
1709                         bio->bi_rw = WRITE;
1710                         bio->bi_end_io = end_sync_write;
1711                         write_targets ++;
1712                 } else {
1713                         /* may need to read from here */
1714                         bio->bi_rw = READ;
1715                         bio->bi_end_io = end_sync_read;
1716                         if (test_bit(WriteMostly, &rdev->flags)) {
1717                                 if (wonly < 0)
1718                                         wonly = i;
1719                         } else {
1720                                 if (disk < 0)
1721                                         disk = i;
1722                         }
1723                         read_targets++;
1724                 }
1725                 atomic_inc(&rdev->nr_pending);
1726                 bio->bi_sector = sector_nr + rdev->data_offset;
1727                 bio->bi_bdev = rdev->bdev;
1728                 bio->bi_private = r1_bio;
1729         }
1730         rcu_read_unlock();
1731         if (disk < 0)
1732                 disk = wonly;
1733         r1_bio->read_disk = disk;
1734
1735         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1736                 /* extra read targets are also write targets */
1737                 write_targets += read_targets-1;
1738
1739         if (write_targets == 0 || read_targets == 0) {
1740                 /* There is nowhere to write, so all non-sync
1741                  * drives must be failed - so we are finished
1742                  */
1743                 sector_t rv = max_sector - sector_nr;
1744                 *skipped = 1;
1745                 put_buf(r1_bio);
1746                 return rv;
1747         }
1748
1749         nr_sectors = 0;
1750         sync_blocks = 0;
1751         do {
1752                 struct page *page;
1753                 int len = PAGE_SIZE;
1754                 if (sector_nr + (len>>9) > max_sector)
1755                         len = (max_sector - sector_nr) << 9;
1756                 if (len == 0)
1757                         break;
1758                 if (sync_blocks == 0) {
1759                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1760                                                &sync_blocks, still_degraded) &&
1761                             !conf->fullsync &&
1762                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1763                                 break;
1764                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1765                         if (len > (sync_blocks<<9))
1766                                 len = sync_blocks<<9;
1767                 }
1768
1769                 for (i=0 ; i < conf->raid_disks; i++) {
1770                         bio = r1_bio->bios[i];
1771                         if (bio->bi_end_io) {
1772                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1773                                 if (bio_add_page(bio, page, len, 0) == 0) {
1774                                         /* stop here */
1775                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1776                                         while (i > 0) {
1777                                                 i--;
1778                                                 bio = r1_bio->bios[i];
1779                                                 if (bio->bi_end_io==NULL)
1780                                                         continue;
1781                                                 /* remove last page from this bio */
1782                                                 bio->bi_vcnt--;
1783                                                 bio->bi_size -= len;
1784                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1785                                         }
1786                                         goto bio_full;
1787                                 }
1788                         }
1789                 }
1790                 nr_sectors += len>>9;
1791                 sector_nr += len>>9;
1792                 sync_blocks -= (len>>9);
1793         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1794  bio_full:
1795         r1_bio->sectors = nr_sectors;
1796
1797         /* For a user-requested sync, we read all readable devices and do a
1798          * compare
1799          */
1800         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1801                 atomic_set(&r1_bio->remaining, read_targets);
1802                 for (i=0; i<conf->raid_disks; i++) {
1803                         bio = r1_bio->bios[i];
1804                         if (bio->bi_end_io == end_sync_read) {
1805                                 md_sync_acct(bio->bi_bdev, nr_sectors);
1806                                 generic_make_request(bio);
1807                         }
1808                 }
1809         } else {
1810                 atomic_set(&r1_bio->remaining, 1);
1811                 bio = r1_bio->bios[r1_bio->read_disk];
1812                 md_sync_acct(bio->bi_bdev, nr_sectors);
1813                 generic_make_request(bio);
1814
1815         }
1816         return nr_sectors;
1817 }
1818
1819 static int run(mddev_t *mddev)
1820 {
1821         conf_t *conf;
1822         int i, j, disk_idx;
1823         mirror_info_t *disk;
1824         mdk_rdev_t *rdev;
1825         struct list_head *tmp;
1826
1827         if (mddev->level != 1) {
1828                 printk("raid1: %s: raid level not set to mirroring (%d)\n",
1829                        mdname(mddev), mddev->level);
1830                 goto out;
1831         }
1832         if (mddev->reshape_position != MaxSector) {
1833                 printk("raid1: %s: reshape_position set but not supported\n",
1834                        mdname(mddev));
1835                 goto out;
1836         }
1837         /*
1838          * copy the already verified devices into our private RAID1
1839          * bookkeeping area. [whatever we allocate in run(),
1840          * should be freed in stop()]
1841          */
1842         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1843         mddev->private = conf;
1844         if (!conf)
1845                 goto out_no_mem;
1846
1847         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1848                                  GFP_KERNEL);
1849         if (!conf->mirrors)
1850                 goto out_no_mem;
1851
1852         conf->tmppage = alloc_page(GFP_KERNEL);
1853         if (!conf->tmppage)
1854                 goto out_no_mem;
1855
1856         conf->poolinfo = kmalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1857         if (!conf->poolinfo)
1858                 goto out_no_mem;
1859         conf->poolinfo->mddev = mddev;
1860         conf->poolinfo->raid_disks = mddev->raid_disks;
1861         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1862                                           r1bio_pool_free,
1863                                           conf->poolinfo);
1864         if (!conf->r1bio_pool)
1865                 goto out_no_mem;
1866
1867         ITERATE_RDEV(mddev, rdev, tmp) {
1868                 disk_idx = rdev->raid_disk;
1869                 if (disk_idx >= mddev->raid_disks
1870                     || disk_idx < 0)
1871                         continue;
1872                 disk = conf->mirrors + disk_idx;
1873
1874                 disk->rdev = rdev;
1875
1876                 blk_queue_stack_limits(mddev->queue,
1877                                        rdev->bdev->bd_disk->queue);
1878                 /* as we don't honour merge_bvec_fn, we must never risk
1879                  * violating it, so limit ->max_sector to one PAGE, as
1880                  * a one page request is never in violation.
1881                  */
1882                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1883                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
1884                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1885
1886                 disk->head_position = 0;
1887                 if (!test_bit(Faulty, &rdev->flags) && test_bit(In_sync, &rdev->flags))
1888                         conf->working_disks++;
1889         }
1890         conf->raid_disks = mddev->raid_disks;
1891         conf->mddev = mddev;
1892         spin_lock_init(&conf->device_lock);
1893         INIT_LIST_HEAD(&conf->retry_list);
1894         if (conf->working_disks == 1)
1895                 mddev->recovery_cp = MaxSector;
1896
1897         spin_lock_init(&conf->resync_lock);
1898         init_waitqueue_head(&conf->wait_barrier);
1899
1900         bio_list_init(&conf->pending_bio_list);
1901         bio_list_init(&conf->flushing_bio_list);
1902
1903         if (!conf->working_disks) {
1904                 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
1905                         mdname(mddev));
1906                 goto out_free_conf;
1907         }
1908
1909         mddev->degraded = 0;
1910         for (i = 0; i < conf->raid_disks; i++) {
1911
1912                 disk = conf->mirrors + i;
1913
1914                 if (!disk->rdev ||
1915                     !test_bit(In_sync, &disk->rdev->flags)) {
1916                         disk->head_position = 0;
1917                         mddev->degraded++;
1918                 }
1919         }
1920
1921         /*
1922          * find the first working one and use it as a starting point
1923          * to read balancing.
1924          */
1925         for (j = 0; j < conf->raid_disks &&
1926                      (!conf->mirrors[j].rdev ||
1927                       !test_bit(In_sync, &conf->mirrors[j].rdev->flags)) ; j++)
1928                 /* nothing */;
1929         conf->last_used = j;
1930
1931
1932         mddev->thread = md_register_thread(raid1d, mddev, "%s_raid1");
1933         if (!mddev->thread) {
1934                 printk(KERN_ERR
1935                        "raid1: couldn't allocate thread for %s\n",
1936                        mdname(mddev));
1937                 goto out_free_conf;
1938         }
1939
1940         printk(KERN_INFO 
1941                 "raid1: raid set %s active with %d out of %d mirrors\n",
1942                 mdname(mddev), mddev->raid_disks - mddev->degraded, 
1943                 mddev->raid_disks);
1944         /*
1945          * Ok, everything is just fine now
1946          */
1947         mddev->array_size = mddev->size;
1948
1949         mddev->queue->unplug_fn = raid1_unplug;
1950         mddev->queue->issue_flush_fn = raid1_issue_flush;
1951
1952         return 0;
1953
1954 out_no_mem:
1955         printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
1956                mdname(mddev));
1957
1958 out_free_conf:
1959         if (conf) {
1960                 if (conf->r1bio_pool)
1961                         mempool_destroy(conf->r1bio_pool);
1962                 kfree(conf->mirrors);
1963                 safe_put_page(conf->tmppage);
1964                 kfree(conf->poolinfo);
1965                 kfree(conf);
1966                 mddev->private = NULL;
1967         }
1968 out:
1969         return -EIO;
1970 }
1971
1972 static int stop(mddev_t *mddev)
1973 {
1974         conf_t *conf = mddev_to_conf(mddev);
1975         struct bitmap *bitmap = mddev->bitmap;
1976         int behind_wait = 0;
1977
1978         /* wait for behind writes to complete */
1979         while (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
1980                 behind_wait++;
1981                 printk(KERN_INFO "raid1: behind writes in progress on device %s, waiting to stop (%d)\n", mdname(mddev), behind_wait);
1982                 set_current_state(TASK_UNINTERRUPTIBLE);
1983                 schedule_timeout(HZ); /* wait a second */
1984                 /* need to kick something here to make sure I/O goes? */
1985         }
1986
1987         md_unregister_thread(mddev->thread);
1988         mddev->thread = NULL;
1989         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
1990         if (conf->r1bio_pool)
1991                 mempool_destroy(conf->r1bio_pool);
1992         kfree(conf->mirrors);
1993         kfree(conf->poolinfo);
1994         kfree(conf);
1995         mddev->private = NULL;
1996         return 0;
1997 }
1998
1999 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2000 {
2001         /* no resync is happening, and there is enough space
2002          * on all devices, so we can resize.
2003          * We need to make sure resync covers any new space.
2004          * If the array is shrinking we should possibly wait until
2005          * any io in the removed space completes, but it hardly seems
2006          * worth it.
2007          */
2008         mddev->array_size = sectors>>1;
2009         set_capacity(mddev->gendisk, mddev->array_size << 1);
2010         mddev->changed = 1;
2011         if (mddev->array_size > mddev->size && mddev->recovery_cp == MaxSector) {
2012                 mddev->recovery_cp = mddev->size << 1;
2013                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2014         }
2015         mddev->size = mddev->array_size;
2016         mddev->resync_max_sectors = sectors;
2017         return 0;
2018 }
2019
2020 static int raid1_reshape(mddev_t *mddev)
2021 {
2022         /* We need to:
2023          * 1/ resize the r1bio_pool
2024          * 2/ resize conf->mirrors
2025          *
2026          * We allocate a new r1bio_pool if we can.
2027          * Then raise a device barrier and wait until all IO stops.
2028          * Then resize conf->mirrors and swap in the new r1bio pool.
2029          *
2030          * At the same time, we "pack" the devices so that all the missing
2031          * devices have the higher raid_disk numbers.
2032          */
2033         mempool_t *newpool, *oldpool;
2034         struct pool_info *newpoolinfo;
2035         mirror_info_t *newmirrors;
2036         conf_t *conf = mddev_to_conf(mddev);
2037         int cnt, raid_disks;
2038
2039         int d, d2;
2040
2041         /* Cannot change chunk_size, layout, or level */
2042         if (mddev->chunk_size != mddev->new_chunk ||
2043             mddev->layout != mddev->new_layout ||
2044             mddev->level != mddev->new_level) {
2045                 mddev->new_chunk = mddev->chunk_size;
2046                 mddev->new_layout = mddev->layout;
2047                 mddev->new_level = mddev->level;
2048                 return -EINVAL;
2049         }
2050
2051         raid_disks = mddev->raid_disks + mddev->delta_disks;
2052
2053         if (raid_disks < conf->raid_disks) {
2054                 cnt=0;
2055                 for (d= 0; d < conf->raid_disks; d++)
2056                         if (conf->mirrors[d].rdev)
2057                                 cnt++;
2058                 if (cnt > raid_disks)
2059                         return -EBUSY;
2060         }
2061
2062         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2063         if (!newpoolinfo)
2064                 return -ENOMEM;
2065         newpoolinfo->mddev = mddev;
2066         newpoolinfo->raid_disks = raid_disks;
2067
2068         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2069                                  r1bio_pool_free, newpoolinfo);
2070         if (!newpool) {
2071                 kfree(newpoolinfo);
2072                 return -ENOMEM;
2073         }
2074         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2075         if (!newmirrors) {
2076                 kfree(newpoolinfo);
2077                 mempool_destroy(newpool);
2078                 return -ENOMEM;
2079         }
2080
2081         raise_barrier(conf);
2082
2083         /* ok, everything is stopped */
2084         oldpool = conf->r1bio_pool;
2085         conf->r1bio_pool = newpool;
2086
2087         for (d=d2=0; d < conf->raid_disks; d++)
2088                 if (conf->mirrors[d].rdev) {
2089                         conf->mirrors[d].rdev->raid_disk = d2;
2090                         newmirrors[d2++].rdev = conf->mirrors[d].rdev;
2091                 }
2092         kfree(conf->mirrors);
2093         conf->mirrors = newmirrors;
2094         kfree(conf->poolinfo);
2095         conf->poolinfo = newpoolinfo;
2096
2097         mddev->degraded += (raid_disks - conf->raid_disks);
2098         conf->raid_disks = mddev->raid_disks = raid_disks;
2099         mddev->delta_disks = 0;
2100
2101         conf->last_used = 0; /* just make sure it is in-range */
2102         lower_barrier(conf);
2103
2104         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2105         md_wakeup_thread(mddev->thread);
2106
2107         mempool_destroy(oldpool);
2108         return 0;
2109 }
2110
2111 static void raid1_quiesce(mddev_t *mddev, int state)
2112 {
2113         conf_t *conf = mddev_to_conf(mddev);
2114
2115         switch(state) {
2116         case 1:
2117                 raise_barrier(conf);
2118                 break;
2119         case 0:
2120                 lower_barrier(conf);
2121                 break;
2122         }
2123 }
2124
2125
2126 static struct mdk_personality raid1_personality =
2127 {
2128         .name           = "raid1",
2129         .level          = 1,
2130         .owner          = THIS_MODULE,
2131         .make_request   = make_request,
2132         .run            = run,
2133         .stop           = stop,
2134         .status         = status,
2135         .error_handler  = error,
2136         .hot_add_disk   = raid1_add_disk,
2137         .hot_remove_disk= raid1_remove_disk,
2138         .spare_active   = raid1_spare_active,
2139         .sync_request   = sync_request,
2140         .resize         = raid1_resize,
2141         .check_reshape  = raid1_reshape,
2142         .quiesce        = raid1_quiesce,
2143 };
2144
2145 static int __init raid_init(void)
2146 {
2147         return register_md_personality(&raid1_personality);
2148 }
2149
2150 static void raid_exit(void)
2151 {
2152         unregister_md_personality(&raid1_personality);
2153 }
2154
2155 module_init(raid_init);
2156 module_exit(raid_exit);
2157 MODULE_LICENSE("GPL");
2158 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2159 MODULE_ALIAS("md-raid1");
2160 MODULE_ALIAS("md-level-1");