Merge branch 'pending' of master.kernel.org:/pub/scm/linux/kernel/git/vxy/lksctp-dev
[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,
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 void raid1_end_read_request(struct bio *bio, 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         mirror = r1_bio->read_disk;
266         /*
267          * this branch is our 'one mirror IO has finished' event handler:
268          */
269         update_head_pos(mirror, r1_bio);
270
271         if (uptodate)
272                 set_bit(R1BIO_Uptodate, &r1_bio->state);
273         else {
274                 /* If all other devices have failed, we want to return
275                  * the error upwards rather than fail the last device.
276                  * Here we redefine "uptodate" to mean "Don't want to retry"
277                  */
278                 unsigned long flags;
279                 spin_lock_irqsave(&conf->device_lock, flags);
280                 if (r1_bio->mddev->degraded == conf->raid_disks ||
281                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
282                      !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
283                         uptodate = 1;
284                 spin_unlock_irqrestore(&conf->device_lock, flags);
285         }
286
287         if (uptodate)
288                 raid_end_bio_io(r1_bio);
289         else {
290                 /*
291                  * oops, read error:
292                  */
293                 char b[BDEVNAME_SIZE];
294                 if (printk_ratelimit())
295                         printk(KERN_ERR "raid1: %s: rescheduling sector %llu\n",
296                                bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
297                 reschedule_retry(r1_bio);
298         }
299
300         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
301 }
302
303 static void raid1_end_write_request(struct bio *bio, 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
312         for (mirror = 0; mirror < conf->raid_disks; mirror++)
313                 if (r1_bio->bios[mirror] == bio)
314                         break;
315
316         if (error == -EOPNOTSUPP && test_bit(R1BIO_Barrier, &r1_bio->state)) {
317                 set_bit(BarriersNotsupp, &conf->mirrors[mirror].rdev->flags);
318                 set_bit(R1BIO_BarrierRetry, &r1_bio->state);
319                 r1_bio->mddev->barriers_work = 0;
320                 /* Don't rdev_dec_pending in this branch - keep it for the retry */
321         } else {
322                 /*
323                  * this branch is our 'one mirror IO has finished' event handler:
324                  */
325                 r1_bio->bios[mirror] = NULL;
326                 to_put = bio;
327                 if (!uptodate) {
328                         md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
329                         /* an I/O failed, we can't clear the bitmap */
330                         set_bit(R1BIO_Degraded, &r1_bio->state);
331                 } else
332                         /*
333                          * Set R1BIO_Uptodate in our master bio, so that
334                          * we will return a good error code for to the higher
335                          * levels even if IO on some other mirrored buffer fails.
336                          *
337                          * The 'master' represents the composite IO operation to
338                          * user-side. So if something waits for IO, then it will
339                          * wait for the 'master' bio.
340                          */
341                         set_bit(R1BIO_Uptodate, &r1_bio->state);
342
343                 update_head_pos(mirror, r1_bio);
344
345                 if (behind) {
346                         if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
347                                 atomic_dec(&r1_bio->behind_remaining);
348
349                         /* In behind mode, we ACK the master bio once the I/O has safely
350                          * reached all non-writemostly disks. Setting the Returned bit
351                          * ensures that this gets done only once -- we don't ever want to
352                          * return -EIO here, instead we'll wait */
353
354                         if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
355                             test_bit(R1BIO_Uptodate, &r1_bio->state)) {
356                                 /* Maybe we can return now */
357                                 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
358                                         struct bio *mbio = r1_bio->master_bio;
359                                         PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
360                                                (unsigned long long) mbio->bi_sector,
361                                                (unsigned long long) mbio->bi_sector +
362                                                (mbio->bi_size >> 9) - 1);
363                                         bio_endio(mbio, 0);
364                                 }
365                         }
366                 }
367                 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
368         }
369         /*
370          *
371          * Let's see if all mirrored write operations have finished
372          * already.
373          */
374         if (atomic_dec_and_test(&r1_bio->remaining)) {
375                 if (test_bit(R1BIO_BarrierRetry, &r1_bio->state))
376                         reschedule_retry(r1_bio);
377                 else {
378                         /* it really is the end of this request */
379                         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
380                                 /* free extra copy of the data pages */
381                                 int i = bio->bi_vcnt;
382                                 while (i--)
383                                         safe_put_page(bio->bi_io_vec[i].bv_page);
384                         }
385                         /* clear the bitmap if all writes complete successfully */
386                         bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
387                                         r1_bio->sectors,
388                                         !test_bit(R1BIO_Degraded, &r1_bio->state),
389                                         behind);
390                         md_write_end(r1_bio->mddev);
391                         raid_end_bio_io(r1_bio);
392                 }
393         }
394
395         if (to_put)
396                 bio_put(to_put);
397 }
398
399
400 /*
401  * This routine returns the disk from which the requested read should
402  * be done. There is a per-array 'next expected sequential IO' sector
403  * number - if this matches on the next IO then we use the last disk.
404  * There is also a per-disk 'last know head position' sector that is
405  * maintained from IRQ contexts, both the normal and the resync IO
406  * completion handlers update this position correctly. If there is no
407  * perfect sequential match then we pick the disk whose head is closest.
408  *
409  * If there are 2 mirrors in the same 2 devices, performance degrades
410  * because position is mirror, not device based.
411  *
412  * The rdev for the device selected will have nr_pending incremented.
413  */
414 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
415 {
416         const unsigned long this_sector = r1_bio->sector;
417         int new_disk = conf->last_used, disk = new_disk;
418         int wonly_disk = -1;
419         const int sectors = r1_bio->sectors;
420         sector_t new_distance, current_distance;
421         mdk_rdev_t *rdev;
422
423         rcu_read_lock();
424         /*
425          * Check if we can balance. We can balance on the whole
426          * device if no resync is going on, or below the resync window.
427          * We take the first readable disk when above the resync window.
428          */
429  retry:
430         if (conf->mddev->recovery_cp < MaxSector &&
431             (this_sector + sectors >= conf->next_resync)) {
432                 /* Choose the first operation device, for consistancy */
433                 new_disk = 0;
434
435                 for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
436                      r1_bio->bios[new_disk] == IO_BLOCKED ||
437                      !rdev || !test_bit(In_sync, &rdev->flags)
438                              || test_bit(WriteMostly, &rdev->flags);
439                      rdev = rcu_dereference(conf->mirrors[++new_disk].rdev)) {
440
441                         if (rdev && test_bit(In_sync, &rdev->flags) &&
442                                 r1_bio->bios[new_disk] != IO_BLOCKED)
443                                 wonly_disk = new_disk;
444
445                         if (new_disk == conf->raid_disks - 1) {
446                                 new_disk = wonly_disk;
447                                 break;
448                         }
449                 }
450                 goto rb_out;
451         }
452
453
454         /* make sure the disk is operational */
455         for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
456              r1_bio->bios[new_disk] == IO_BLOCKED ||
457              !rdev || !test_bit(In_sync, &rdev->flags) ||
458                      test_bit(WriteMostly, &rdev->flags);
459              rdev = rcu_dereference(conf->mirrors[new_disk].rdev)) {
460
461                 if (rdev && test_bit(In_sync, &rdev->flags) &&
462                     r1_bio->bios[new_disk] != IO_BLOCKED)
463                         wonly_disk = new_disk;
464
465                 if (new_disk <= 0)
466                         new_disk = conf->raid_disks;
467                 new_disk--;
468                 if (new_disk == disk) {
469                         new_disk = wonly_disk;
470                         break;
471                 }
472         }
473
474         if (new_disk < 0)
475                 goto rb_out;
476
477         disk = new_disk;
478         /* now disk == new_disk == starting point for search */
479
480         /*
481          * Don't change to another disk for sequential reads:
482          */
483         if (conf->next_seq_sect == this_sector)
484                 goto rb_out;
485         if (this_sector == conf->mirrors[new_disk].head_position)
486                 goto rb_out;
487
488         current_distance = abs(this_sector - conf->mirrors[disk].head_position);
489
490         /* Find the disk whose head is closest */
491
492         do {
493                 if (disk <= 0)
494                         disk = conf->raid_disks;
495                 disk--;
496
497                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
498
499                 if (!rdev || r1_bio->bios[disk] == IO_BLOCKED ||
500                     !test_bit(In_sync, &rdev->flags) ||
501                     test_bit(WriteMostly, &rdev->flags))
502                         continue;
503
504                 if (!atomic_read(&rdev->nr_pending)) {
505                         new_disk = disk;
506                         break;
507                 }
508                 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
509                 if (new_distance < current_distance) {
510                         current_distance = new_distance;
511                         new_disk = disk;
512                 }
513         } while (disk != conf->last_used);
514
515  rb_out:
516
517
518         if (new_disk >= 0) {
519                 rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
520                 if (!rdev)
521                         goto retry;
522                 atomic_inc(&rdev->nr_pending);
523                 if (!test_bit(In_sync, &rdev->flags)) {
524                         /* cannot risk returning a device that failed
525                          * before we inc'ed nr_pending
526                          */
527                         rdev_dec_pending(rdev, conf->mddev);
528                         goto retry;
529                 }
530                 conf->next_seq_sect = this_sector + sectors;
531                 conf->last_used = new_disk;
532         }
533         rcu_read_unlock();
534
535         return new_disk;
536 }
537
538 static void unplug_slaves(mddev_t *mddev)
539 {
540         conf_t *conf = mddev_to_conf(mddev);
541         int i;
542
543         rcu_read_lock();
544         for (i=0; i<mddev->raid_disks; i++) {
545                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
546                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
547                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
548
549                         atomic_inc(&rdev->nr_pending);
550                         rcu_read_unlock();
551
552                         blk_unplug(r_queue);
553
554                         rdev_dec_pending(rdev, mddev);
555                         rcu_read_lock();
556                 }
557         }
558         rcu_read_unlock();
559 }
560
561 static void raid1_unplug(struct request_queue *q)
562 {
563         mddev_t *mddev = q->queuedata;
564
565         unplug_slaves(mddev);
566         md_wakeup_thread(mddev->thread);
567 }
568
569 static int raid1_congested(void *data, int bits)
570 {
571         mddev_t *mddev = data;
572         conf_t *conf = mddev_to_conf(mddev);
573         int i, ret = 0;
574
575         rcu_read_lock();
576         for (i = 0; i < mddev->raid_disks; i++) {
577                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
578                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
579                         struct request_queue *q = bdev_get_queue(rdev->bdev);
580
581                         /* Note the '|| 1' - when read_balance prefers
582                          * non-congested targets, it can be removed
583                          */
584                         if ((bits & (1<<BDI_write_congested)) || 1)
585                                 ret |= bdi_congested(&q->backing_dev_info, bits);
586                         else
587                                 ret &= bdi_congested(&q->backing_dev_info, bits);
588                 }
589         }
590         rcu_read_unlock();
591         return ret;
592 }
593
594
595 /* Barriers....
596  * Sometimes we need to suspend IO while we do something else,
597  * either some resync/recovery, or reconfigure the array.
598  * To do this we raise a 'barrier'.
599  * The 'barrier' is a counter that can be raised multiple times
600  * to count how many activities are happening which preclude
601  * normal IO.
602  * We can only raise the barrier if there is no pending IO.
603  * i.e. if nr_pending == 0.
604  * We choose only to raise the barrier if no-one is waiting for the
605  * barrier to go down.  This means that as soon as an IO request
606  * is ready, no other operations which require a barrier will start
607  * until the IO request has had a chance.
608  *
609  * So: regular IO calls 'wait_barrier'.  When that returns there
610  *    is no backgroup IO happening,  It must arrange to call
611  *    allow_barrier when it has finished its IO.
612  * backgroup IO calls must call raise_barrier.  Once that returns
613  *    there is no normal IO happeing.  It must arrange to call
614  *    lower_barrier when the particular background IO completes.
615  */
616 #define RESYNC_DEPTH 32
617
618 static void raise_barrier(conf_t *conf)
619 {
620         spin_lock_irq(&conf->resync_lock);
621
622         /* Wait until no block IO is waiting */
623         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
624                             conf->resync_lock,
625                             raid1_unplug(conf->mddev->queue));
626
627         /* block any new IO from starting */
628         conf->barrier++;
629
630         /* No wait for all pending IO to complete */
631         wait_event_lock_irq(conf->wait_barrier,
632                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
633                             conf->resync_lock,
634                             raid1_unplug(conf->mddev->queue));
635
636         spin_unlock_irq(&conf->resync_lock);
637 }
638
639 static void lower_barrier(conf_t *conf)
640 {
641         unsigned long flags;
642         spin_lock_irqsave(&conf->resync_lock, flags);
643         conf->barrier--;
644         spin_unlock_irqrestore(&conf->resync_lock, flags);
645         wake_up(&conf->wait_barrier);
646 }
647
648 static void wait_barrier(conf_t *conf)
649 {
650         spin_lock_irq(&conf->resync_lock);
651         if (conf->barrier) {
652                 conf->nr_waiting++;
653                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
654                                     conf->resync_lock,
655                                     raid1_unplug(conf->mddev->queue));
656                 conf->nr_waiting--;
657         }
658         conf->nr_pending++;
659         spin_unlock_irq(&conf->resync_lock);
660 }
661
662 static void allow_barrier(conf_t *conf)
663 {
664         unsigned long flags;
665         spin_lock_irqsave(&conf->resync_lock, flags);
666         conf->nr_pending--;
667         spin_unlock_irqrestore(&conf->resync_lock, flags);
668         wake_up(&conf->wait_barrier);
669 }
670
671 static void freeze_array(conf_t *conf)
672 {
673         /* stop syncio and normal IO and wait for everything to
674          * go quite.
675          * We increment barrier and nr_waiting, and then
676          * wait until barrier+nr_pending match nr_queued+2
677          */
678         spin_lock_irq(&conf->resync_lock);
679         conf->barrier++;
680         conf->nr_waiting++;
681         wait_event_lock_irq(conf->wait_barrier,
682                             conf->barrier+conf->nr_pending == conf->nr_queued+2,
683                             conf->resync_lock,
684                             raid1_unplug(conf->mddev->queue));
685         spin_unlock_irq(&conf->resync_lock);
686 }
687 static void unfreeze_array(conf_t *conf)
688 {
689         /* reverse the effect of the freeze */
690         spin_lock_irq(&conf->resync_lock);
691         conf->barrier--;
692         conf->nr_waiting--;
693         wake_up(&conf->wait_barrier);
694         spin_unlock_irq(&conf->resync_lock);
695 }
696
697
698 /* duplicate the data pages for behind I/O */
699 static struct page **alloc_behind_pages(struct bio *bio)
700 {
701         int i;
702         struct bio_vec *bvec;
703         struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page *),
704                                         GFP_NOIO);
705         if (unlikely(!pages))
706                 goto do_sync_io;
707
708         bio_for_each_segment(bvec, bio, i) {
709                 pages[i] = alloc_page(GFP_NOIO);
710                 if (unlikely(!pages[i]))
711                         goto do_sync_io;
712                 memcpy(kmap(pages[i]) + bvec->bv_offset,
713                         kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
714                 kunmap(pages[i]);
715                 kunmap(bvec->bv_page);
716         }
717
718         return pages;
719
720 do_sync_io:
721         if (pages)
722                 for (i = 0; i < bio->bi_vcnt && pages[i]; i++)
723                         put_page(pages[i]);
724         kfree(pages);
725         PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
726         return NULL;
727 }
728
729 static int make_request(struct request_queue *q, struct bio * bio)
730 {
731         mddev_t *mddev = q->queuedata;
732         conf_t *conf = mddev_to_conf(mddev);
733         mirror_info_t *mirror;
734         r1bio_t *r1_bio;
735         struct bio *read_bio;
736         int i, targets = 0, disks;
737         mdk_rdev_t *rdev;
738         struct bitmap *bitmap = mddev->bitmap;
739         unsigned long flags;
740         struct bio_list bl;
741         struct page **behind_pages = NULL;
742         const int rw = bio_data_dir(bio);
743         const int do_sync = bio_sync(bio);
744         int do_barriers;
745
746         /*
747          * Register the new request and wait if the reconstruction
748          * thread has put up a bar for new requests.
749          * Continue immediately if no resync is active currently.
750          * We test barriers_work *after* md_write_start as md_write_start
751          * may cause the first superblock write, and that will check out
752          * if barriers work.
753          */
754
755         md_write_start(mddev, bio); /* wait on superblock update early */
756
757         if (unlikely(!mddev->barriers_work && bio_barrier(bio))) {
758                 if (rw == WRITE)
759                         md_write_end(mddev);
760                 bio_endio(bio, -EOPNOTSUPP);
761                 return 0;
762         }
763
764         wait_barrier(conf);
765
766         disk_stat_inc(mddev->gendisk, ios[rw]);
767         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
768
769         /*
770          * make_request() can abort the operation when READA is being
771          * used and no empty request is available.
772          *
773          */
774         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
775
776         r1_bio->master_bio = bio;
777         r1_bio->sectors = bio->bi_size >> 9;
778         r1_bio->state = 0;
779         r1_bio->mddev = mddev;
780         r1_bio->sector = bio->bi_sector;
781
782         if (rw == READ) {
783                 /*
784                  * read balancing logic:
785                  */
786                 int rdisk = read_balance(conf, r1_bio);
787
788                 if (rdisk < 0) {
789                         /* couldn't find anywhere to read from */
790                         raid_end_bio_io(r1_bio);
791                         return 0;
792                 }
793                 mirror = conf->mirrors + rdisk;
794
795                 r1_bio->read_disk = rdisk;
796
797                 read_bio = bio_clone(bio, GFP_NOIO);
798
799                 r1_bio->bios[rdisk] = read_bio;
800
801                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
802                 read_bio->bi_bdev = mirror->rdev->bdev;
803                 read_bio->bi_end_io = raid1_end_read_request;
804                 read_bio->bi_rw = READ | do_sync;
805                 read_bio->bi_private = r1_bio;
806
807                 generic_make_request(read_bio);
808                 return 0;
809         }
810
811         /*
812          * WRITE:
813          */
814         /* first select target devices under spinlock and
815          * inc refcount on their rdev.  Record them by setting
816          * bios[x] to bio
817          */
818         disks = conf->raid_disks;
819 #if 0
820         { static int first=1;
821         if (first) printk("First Write sector %llu disks %d\n",
822                           (unsigned long long)r1_bio->sector, disks);
823         first = 0;
824         }
825 #endif
826         rcu_read_lock();
827         for (i = 0;  i < disks; i++) {
828                 if ((rdev=rcu_dereference(conf->mirrors[i].rdev)) != NULL &&
829                     !test_bit(Faulty, &rdev->flags)) {
830                         atomic_inc(&rdev->nr_pending);
831                         if (test_bit(Faulty, &rdev->flags)) {
832                                 rdev_dec_pending(rdev, mddev);
833                                 r1_bio->bios[i] = NULL;
834                         } else
835                                 r1_bio->bios[i] = bio;
836                         targets++;
837                 } else
838                         r1_bio->bios[i] = NULL;
839         }
840         rcu_read_unlock();
841
842         BUG_ON(targets == 0); /* we never fail the last device */
843
844         if (targets < conf->raid_disks) {
845                 /* array is degraded, we will not clear the bitmap
846                  * on I/O completion (see raid1_end_write_request) */
847                 set_bit(R1BIO_Degraded, &r1_bio->state);
848         }
849
850         /* do behind I/O ? */
851         if (bitmap &&
852             atomic_read(&bitmap->behind_writes) < bitmap->max_write_behind &&
853             (behind_pages = alloc_behind_pages(bio)) != NULL)
854                 set_bit(R1BIO_BehindIO, &r1_bio->state);
855
856         atomic_set(&r1_bio->remaining, 0);
857         atomic_set(&r1_bio->behind_remaining, 0);
858
859         do_barriers = bio_barrier(bio);
860         if (do_barriers)
861                 set_bit(R1BIO_Barrier, &r1_bio->state);
862
863         bio_list_init(&bl);
864         for (i = 0; i < disks; i++) {
865                 struct bio *mbio;
866                 if (!r1_bio->bios[i])
867                         continue;
868
869                 mbio = bio_clone(bio, GFP_NOIO);
870                 r1_bio->bios[i] = mbio;
871
872                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
873                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
874                 mbio->bi_end_io = raid1_end_write_request;
875                 mbio->bi_rw = WRITE | do_barriers | do_sync;
876                 mbio->bi_private = r1_bio;
877
878                 if (behind_pages) {
879                         struct bio_vec *bvec;
880                         int j;
881
882                         /* Yes, I really want the '__' version so that
883                          * we clear any unused pointer in the io_vec, rather
884                          * than leave them unchanged.  This is important
885                          * because when we come to free the pages, we won't
886                          * know the originial bi_idx, so we just free
887                          * them all
888                          */
889                         __bio_for_each_segment(bvec, mbio, j, 0)
890                                 bvec->bv_page = behind_pages[j];
891                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
892                                 atomic_inc(&r1_bio->behind_remaining);
893                 }
894
895                 atomic_inc(&r1_bio->remaining);
896
897                 bio_list_add(&bl, mbio);
898         }
899         kfree(behind_pages); /* the behind pages are attached to the bios now */
900
901         bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
902                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
903         spin_lock_irqsave(&conf->device_lock, flags);
904         bio_list_merge(&conf->pending_bio_list, &bl);
905         bio_list_init(&bl);
906
907         blk_plug_device(mddev->queue);
908         spin_unlock_irqrestore(&conf->device_lock, flags);
909
910         if (do_sync)
911                 md_wakeup_thread(mddev->thread);
912 #if 0
913         while ((bio = bio_list_pop(&bl)) != NULL)
914                 generic_make_request(bio);
915 #endif
916
917         return 0;
918 }
919
920 static void status(struct seq_file *seq, mddev_t *mddev)
921 {
922         conf_t *conf = mddev_to_conf(mddev);
923         int i;
924
925         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
926                    conf->raid_disks - mddev->degraded);
927         rcu_read_lock();
928         for (i = 0; i < conf->raid_disks; i++) {
929                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
930                 seq_printf(seq, "%s",
931                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
932         }
933         rcu_read_unlock();
934         seq_printf(seq, "]");
935 }
936
937
938 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
939 {
940         char b[BDEVNAME_SIZE];
941         conf_t *conf = mddev_to_conf(mddev);
942
943         /*
944          * If it is not operational, then we have already marked it as dead
945          * else if it is the last working disks, ignore the error, let the
946          * next level up know.
947          * else mark the drive as failed
948          */
949         if (test_bit(In_sync, &rdev->flags)
950             && (conf->raid_disks - mddev->degraded) == 1)
951                 /*
952                  * Don't fail the drive, act as though we were just a
953                  * normal single drive
954                  */
955                 return;
956         if (test_and_clear_bit(In_sync, &rdev->flags)) {
957                 unsigned long flags;
958                 spin_lock_irqsave(&conf->device_lock, flags);
959                 mddev->degraded++;
960                 set_bit(Faulty, &rdev->flags);
961                 spin_unlock_irqrestore(&conf->device_lock, flags);
962                 /*
963                  * if recovery is running, make sure it aborts.
964                  */
965                 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
966         } else
967                 set_bit(Faulty, &rdev->flags);
968         set_bit(MD_CHANGE_DEVS, &mddev->flags);
969         printk(KERN_ALERT "raid1: Disk failure on %s, disabling device. \n"
970                 "       Operation continuing on %d devices\n",
971                 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
972 }
973
974 static void print_conf(conf_t *conf)
975 {
976         int i;
977
978         printk("RAID1 conf printout:\n");
979         if (!conf) {
980                 printk("(!conf)\n");
981                 return;
982         }
983         printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
984                 conf->raid_disks);
985
986         rcu_read_lock();
987         for (i = 0; i < conf->raid_disks; i++) {
988                 char b[BDEVNAME_SIZE];
989                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
990                 if (rdev)
991                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
992                                i, !test_bit(In_sync, &rdev->flags),
993                                !test_bit(Faulty, &rdev->flags),
994                                bdevname(rdev->bdev,b));
995         }
996         rcu_read_unlock();
997 }
998
999 static void close_sync(conf_t *conf)
1000 {
1001         wait_barrier(conf);
1002         allow_barrier(conf);
1003
1004         mempool_destroy(conf->r1buf_pool);
1005         conf->r1buf_pool = NULL;
1006 }
1007
1008 static int raid1_spare_active(mddev_t *mddev)
1009 {
1010         int i;
1011         conf_t *conf = mddev->private;
1012
1013         /*
1014          * Find all failed disks within the RAID1 configuration 
1015          * and mark them readable.
1016          * Called under mddev lock, so rcu protection not needed.
1017          */
1018         for (i = 0; i < conf->raid_disks; i++) {
1019                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1020                 if (rdev
1021                     && !test_bit(Faulty, &rdev->flags)
1022                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1023                         unsigned long flags;
1024                         spin_lock_irqsave(&conf->device_lock, flags);
1025                         mddev->degraded--;
1026                         spin_unlock_irqrestore(&conf->device_lock, flags);
1027                 }
1028         }
1029
1030         print_conf(conf);
1031         return 0;
1032 }
1033
1034
1035 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1036 {
1037         conf_t *conf = mddev->private;
1038         int found = 0;
1039         int mirror = 0;
1040         mirror_info_t *p;
1041
1042         for (mirror=0; mirror < mddev->raid_disks; mirror++)
1043                 if ( !(p=conf->mirrors+mirror)->rdev) {
1044
1045                         blk_queue_stack_limits(mddev->queue,
1046                                                rdev->bdev->bd_disk->queue);
1047                         /* as we don't honour merge_bvec_fn, we must never risk
1048                          * violating it, so limit ->max_sector to one PAGE, as
1049                          * a one page request is never in violation.
1050                          */
1051                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1052                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
1053                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1054
1055                         p->head_position = 0;
1056                         rdev->raid_disk = mirror;
1057                         found = 1;
1058                         /* As all devices are equivalent, we don't need a full recovery
1059                          * if this was recently any drive of the array
1060                          */
1061                         if (rdev->saved_raid_disk < 0)
1062                                 conf->fullsync = 1;
1063                         rcu_assign_pointer(p->rdev, rdev);
1064                         break;
1065                 }
1066
1067         print_conf(conf);
1068         return found;
1069 }
1070
1071 static int raid1_remove_disk(mddev_t *mddev, int number)
1072 {
1073         conf_t *conf = mddev->private;
1074         int err = 0;
1075         mdk_rdev_t *rdev;
1076         mirror_info_t *p = conf->mirrors+ number;
1077
1078         print_conf(conf);
1079         rdev = p->rdev;
1080         if (rdev) {
1081                 if (test_bit(In_sync, &rdev->flags) ||
1082                     atomic_read(&rdev->nr_pending)) {
1083                         err = -EBUSY;
1084                         goto abort;
1085                 }
1086                 p->rdev = NULL;
1087                 synchronize_rcu();
1088                 if (atomic_read(&rdev->nr_pending)) {
1089                         /* lost the race, try later */
1090                         err = -EBUSY;
1091                         p->rdev = rdev;
1092                 }
1093         }
1094 abort:
1095
1096         print_conf(conf);
1097         return err;
1098 }
1099
1100
1101 static void end_sync_read(struct bio *bio, int error)
1102 {
1103         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1104         int i;
1105
1106         for (i=r1_bio->mddev->raid_disks; i--; )
1107                 if (r1_bio->bios[i] == bio)
1108                         break;
1109         BUG_ON(i < 0);
1110         update_head_pos(i, r1_bio);
1111         /*
1112          * we have read a block, now it needs to be re-written,
1113          * or re-read if the read failed.
1114          * We don't do much here, just schedule handling by raid1d
1115          */
1116         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1117                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1118
1119         if (atomic_dec_and_test(&r1_bio->remaining))
1120                 reschedule_retry(r1_bio);
1121 }
1122
1123 static void end_sync_write(struct bio *bio, int error)
1124 {
1125         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1126         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1127         mddev_t *mddev = r1_bio->mddev;
1128         conf_t *conf = mddev_to_conf(mddev);
1129         int i;
1130         int mirror=0;
1131
1132         for (i = 0; i < conf->raid_disks; i++)
1133                 if (r1_bio->bios[i] == bio) {
1134                         mirror = i;
1135                         break;
1136                 }
1137         if (!uptodate) {
1138                 int sync_blocks = 0;
1139                 sector_t s = r1_bio->sector;
1140                 long sectors_to_go = r1_bio->sectors;
1141                 /* make sure these bits doesn't get cleared. */
1142                 do {
1143                         bitmap_end_sync(mddev->bitmap, s,
1144                                         &sync_blocks, 1);
1145                         s += sync_blocks;
1146                         sectors_to_go -= sync_blocks;
1147                 } while (sectors_to_go > 0);
1148                 md_error(mddev, conf->mirrors[mirror].rdev);
1149         }
1150
1151         update_head_pos(mirror, r1_bio);
1152
1153         if (atomic_dec_and_test(&r1_bio->remaining)) {
1154                 md_done_sync(mddev, r1_bio->sectors, uptodate);
1155                 put_buf(r1_bio);
1156         }
1157 }
1158
1159 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1160 {
1161         conf_t *conf = mddev_to_conf(mddev);
1162         int i;
1163         int disks = conf->raid_disks;
1164         struct bio *bio, *wbio;
1165
1166         bio = r1_bio->bios[r1_bio->read_disk];
1167
1168
1169         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1170                 /* We have read all readable devices.  If we haven't
1171                  * got the block, then there is no hope left.
1172                  * If we have, then we want to do a comparison
1173                  * and skip the write if everything is the same.
1174                  * If any blocks failed to read, then we need to
1175                  * attempt an over-write
1176                  */
1177                 int primary;
1178                 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1179                         for (i=0; i<mddev->raid_disks; i++)
1180                                 if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1181                                         md_error(mddev, conf->mirrors[i].rdev);
1182
1183                         md_done_sync(mddev, r1_bio->sectors, 1);
1184                         put_buf(r1_bio);
1185                         return;
1186                 }
1187                 for (primary=0; primary<mddev->raid_disks; primary++)
1188                         if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1189                             test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1190                                 r1_bio->bios[primary]->bi_end_io = NULL;
1191                                 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1192                                 break;
1193                         }
1194                 r1_bio->read_disk = primary;
1195                 for (i=0; i<mddev->raid_disks; i++)
1196                         if (r1_bio->bios[i]->bi_end_io == end_sync_read) {
1197                                 int j;
1198                                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1199                                 struct bio *pbio = r1_bio->bios[primary];
1200                                 struct bio *sbio = r1_bio->bios[i];
1201
1202                                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1203                                         for (j = vcnt; j-- ; ) {
1204                                                 struct page *p, *s;
1205                                                 p = pbio->bi_io_vec[j].bv_page;
1206                                                 s = sbio->bi_io_vec[j].bv_page;
1207                                                 if (memcmp(page_address(p),
1208                                                            page_address(s),
1209                                                            PAGE_SIZE))
1210                                                         break;
1211                                         }
1212                                 } else
1213                                         j = 0;
1214                                 if (j >= 0)
1215                                         mddev->resync_mismatches += r1_bio->sectors;
1216                                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1217                                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1218                                         sbio->bi_end_io = NULL;
1219                                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1220                                 } else {
1221                                         /* fixup the bio for reuse */
1222                                         sbio->bi_vcnt = vcnt;
1223                                         sbio->bi_size = r1_bio->sectors << 9;
1224                                         sbio->bi_idx = 0;
1225                                         sbio->bi_phys_segments = 0;
1226                                         sbio->bi_hw_segments = 0;
1227                                         sbio->bi_hw_front_size = 0;
1228                                         sbio->bi_hw_back_size = 0;
1229                                         sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1230                                         sbio->bi_flags |= 1 << BIO_UPTODATE;
1231                                         sbio->bi_next = NULL;
1232                                         sbio->bi_sector = r1_bio->sector +
1233                                                 conf->mirrors[i].rdev->data_offset;
1234                                         sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1235                                         for (j = 0; j < vcnt ; j++)
1236                                                 memcpy(page_address(sbio->bi_io_vec[j].bv_page),
1237                                                        page_address(pbio->bi_io_vec[j].bv_page),
1238                                                        PAGE_SIZE);
1239
1240                                 }
1241                         }
1242         }
1243         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1244                 /* ouch - failed to read all of that.
1245                  * Try some synchronous reads of other devices to get
1246                  * good data, much like with normal read errors.  Only
1247                  * read into the pages we already have so we don't
1248                  * need to re-issue the read request.
1249                  * We don't need to freeze the array, because being in an
1250                  * active sync request, there is no normal IO, and
1251                  * no overlapping syncs.
1252                  */
1253                 sector_t sect = r1_bio->sector;
1254                 int sectors = r1_bio->sectors;
1255                 int idx = 0;
1256
1257                 while(sectors) {
1258                         int s = sectors;
1259                         int d = r1_bio->read_disk;
1260                         int success = 0;
1261                         mdk_rdev_t *rdev;
1262
1263                         if (s > (PAGE_SIZE>>9))
1264                                 s = PAGE_SIZE >> 9;
1265                         do {
1266                                 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1267                                         /* No rcu protection needed here devices
1268                                          * can only be removed when no resync is
1269                                          * active, and resync is currently active
1270                                          */
1271                                         rdev = conf->mirrors[d].rdev;
1272                                         if (sync_page_io(rdev->bdev,
1273                                                          sect + rdev->data_offset,
1274                                                          s<<9,
1275                                                          bio->bi_io_vec[idx].bv_page,
1276                                                          READ)) {
1277                                                 success = 1;
1278                                                 break;
1279                                         }
1280                                 }
1281                                 d++;
1282                                 if (d == conf->raid_disks)
1283                                         d = 0;
1284                         } while (!success && d != r1_bio->read_disk);
1285
1286                         if (success) {
1287                                 int start = d;
1288                                 /* write it back and re-read */
1289                                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1290                                 while (d != r1_bio->read_disk) {
1291                                         if (d == 0)
1292                                                 d = conf->raid_disks;
1293                                         d--;
1294                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1295                                                 continue;
1296                                         rdev = conf->mirrors[d].rdev;
1297                                         atomic_add(s, &rdev->corrected_errors);
1298                                         if (sync_page_io(rdev->bdev,
1299                                                          sect + rdev->data_offset,
1300                                                          s<<9,
1301                                                          bio->bi_io_vec[idx].bv_page,
1302                                                          WRITE) == 0)
1303                                                 md_error(mddev, rdev);
1304                                 }
1305                                 d = start;
1306                                 while (d != r1_bio->read_disk) {
1307                                         if (d == 0)
1308                                                 d = conf->raid_disks;
1309                                         d--;
1310                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1311                                                 continue;
1312                                         rdev = conf->mirrors[d].rdev;
1313                                         if (sync_page_io(rdev->bdev,
1314                                                          sect + rdev->data_offset,
1315                                                          s<<9,
1316                                                          bio->bi_io_vec[idx].bv_page,
1317                                                          READ) == 0)
1318                                                 md_error(mddev, rdev);
1319                                 }
1320                         } else {
1321                                 char b[BDEVNAME_SIZE];
1322                                 /* Cannot read from anywhere, array is toast */
1323                                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1324                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
1325                                        " for block %llu\n",
1326                                        bdevname(bio->bi_bdev,b),
1327                                        (unsigned long long)r1_bio->sector);
1328                                 md_done_sync(mddev, r1_bio->sectors, 0);
1329                                 put_buf(r1_bio);
1330                                 return;
1331                         }
1332                         sectors -= s;
1333                         sect += s;
1334                         idx ++;
1335                 }
1336         }
1337
1338         /*
1339          * schedule writes
1340          */
1341         atomic_set(&r1_bio->remaining, 1);
1342         for (i = 0; i < disks ; i++) {
1343                 wbio = r1_bio->bios[i];
1344                 if (wbio->bi_end_io == NULL ||
1345                     (wbio->bi_end_io == end_sync_read &&
1346                      (i == r1_bio->read_disk ||
1347                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1348                         continue;
1349
1350                 wbio->bi_rw = WRITE;
1351                 wbio->bi_end_io = end_sync_write;
1352                 atomic_inc(&r1_bio->remaining);
1353                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1354
1355                 generic_make_request(wbio);
1356         }
1357
1358         if (atomic_dec_and_test(&r1_bio->remaining)) {
1359                 /* if we're here, all write(s) have completed, so clean up */
1360                 md_done_sync(mddev, r1_bio->sectors, 1);
1361                 put_buf(r1_bio);
1362         }
1363 }
1364
1365 /*
1366  * This is a kernel thread which:
1367  *
1368  *      1.      Retries failed read operations on working mirrors.
1369  *      2.      Updates the raid superblock when problems encounter.
1370  *      3.      Performs writes following reads for array syncronising.
1371  */
1372
1373 static void fix_read_error(conf_t *conf, int read_disk,
1374                            sector_t sect, int sectors)
1375 {
1376         mddev_t *mddev = conf->mddev;
1377         while(sectors) {
1378                 int s = sectors;
1379                 int d = read_disk;
1380                 int success = 0;
1381                 int start;
1382                 mdk_rdev_t *rdev;
1383
1384                 if (s > (PAGE_SIZE>>9))
1385                         s = PAGE_SIZE >> 9;
1386
1387                 do {
1388                         /* Note: no rcu protection needed here
1389                          * as this is synchronous in the raid1d thread
1390                          * which is the thread that might remove
1391                          * a device.  If raid1d ever becomes multi-threaded....
1392                          */
1393                         rdev = conf->mirrors[d].rdev;
1394                         if (rdev &&
1395                             test_bit(In_sync, &rdev->flags) &&
1396                             sync_page_io(rdev->bdev,
1397                                          sect + rdev->data_offset,
1398                                          s<<9,
1399                                          conf->tmppage, READ))
1400                                 success = 1;
1401                         else {
1402                                 d++;
1403                                 if (d == conf->raid_disks)
1404                                         d = 0;
1405                         }
1406                 } while (!success && d != read_disk);
1407
1408                 if (!success) {
1409                         /* Cannot read from anywhere -- bye bye array */
1410                         md_error(mddev, conf->mirrors[read_disk].rdev);
1411                         break;
1412                 }
1413                 /* write it back and re-read */
1414                 start = d;
1415                 while (d != read_disk) {
1416                         if (d==0)
1417                                 d = conf->raid_disks;
1418                         d--;
1419                         rdev = conf->mirrors[d].rdev;
1420                         if (rdev &&
1421                             test_bit(In_sync, &rdev->flags)) {
1422                                 if (sync_page_io(rdev->bdev,
1423                                                  sect + rdev->data_offset,
1424                                                  s<<9, conf->tmppage, WRITE)
1425                                     == 0)
1426                                         /* Well, this device is dead */
1427                                         md_error(mddev, rdev);
1428                         }
1429                 }
1430                 d = start;
1431                 while (d != read_disk) {
1432                         char b[BDEVNAME_SIZE];
1433                         if (d==0)
1434                                 d = conf->raid_disks;
1435                         d--;
1436                         rdev = conf->mirrors[d].rdev;
1437                         if (rdev &&
1438                             test_bit(In_sync, &rdev->flags)) {
1439                                 if (sync_page_io(rdev->bdev,
1440                                                  sect + rdev->data_offset,
1441                                                  s<<9, conf->tmppage, READ)
1442                                     == 0)
1443                                         /* Well, this device is dead */
1444                                         md_error(mddev, rdev);
1445                                 else {
1446                                         atomic_add(s, &rdev->corrected_errors);
1447                                         printk(KERN_INFO
1448                                                "raid1:%s: read error corrected "
1449                                                "(%d sectors at %llu on %s)\n",
1450                                                mdname(mddev), s,
1451                                                (unsigned long long)(sect +
1452                                                    rdev->data_offset),
1453                                                bdevname(rdev->bdev, b));
1454                                 }
1455                         }
1456                 }
1457                 sectors -= s;
1458                 sect += s;
1459         }
1460 }
1461
1462 static void raid1d(mddev_t *mddev)
1463 {
1464         r1bio_t *r1_bio;
1465         struct bio *bio;
1466         unsigned long flags;
1467         conf_t *conf = mddev_to_conf(mddev);
1468         struct list_head *head = &conf->retry_list;
1469         int unplug=0;
1470         mdk_rdev_t *rdev;
1471
1472         md_check_recovery(mddev);
1473         
1474         for (;;) {
1475                 char b[BDEVNAME_SIZE];
1476                 spin_lock_irqsave(&conf->device_lock, flags);
1477
1478                 if (conf->pending_bio_list.head) {
1479                         bio = bio_list_get(&conf->pending_bio_list);
1480                         blk_remove_plug(mddev->queue);
1481                         spin_unlock_irqrestore(&conf->device_lock, flags);
1482                         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
1483                         bitmap_unplug(mddev->bitmap);
1484
1485                         while (bio) { /* submit pending writes */
1486                                 struct bio *next = bio->bi_next;
1487                                 bio->bi_next = NULL;
1488                                 generic_make_request(bio);
1489                                 bio = next;
1490                         }
1491                         unplug = 1;
1492
1493                         continue;
1494                 }
1495
1496                 if (list_empty(head))
1497                         break;
1498                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1499                 list_del(head->prev);
1500                 conf->nr_queued--;
1501                 spin_unlock_irqrestore(&conf->device_lock, flags);
1502
1503                 mddev = r1_bio->mddev;
1504                 conf = mddev_to_conf(mddev);
1505                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1506                         sync_request_write(mddev, r1_bio);
1507                         unplug = 1;
1508                 } else if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
1509                         /* some requests in the r1bio were BIO_RW_BARRIER
1510                          * requests which failed with -EOPNOTSUPP.  Hohumm..
1511                          * Better resubmit without the barrier.
1512                          * We know which devices to resubmit for, because
1513                          * all others have had their bios[] entry cleared.
1514                          * We already have a nr_pending reference on these rdevs.
1515                          */
1516                         int i;
1517                         const int do_sync = bio_sync(r1_bio->master_bio);
1518                         clear_bit(R1BIO_BarrierRetry, &r1_bio->state);
1519                         clear_bit(R1BIO_Barrier, &r1_bio->state);
1520                         for (i=0; i < conf->raid_disks; i++)
1521                                 if (r1_bio->bios[i])
1522                                         atomic_inc(&r1_bio->remaining);
1523                         for (i=0; i < conf->raid_disks; i++)
1524                                 if (r1_bio->bios[i]) {
1525                                         struct bio_vec *bvec;
1526                                         int j;
1527
1528                                         bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1529                                         /* copy pages from the failed bio, as
1530                                          * this might be a write-behind device */
1531                                         __bio_for_each_segment(bvec, bio, j, 0)
1532                                                 bvec->bv_page = bio_iovec_idx(r1_bio->bios[i], j)->bv_page;
1533                                         bio_put(r1_bio->bios[i]);
1534                                         bio->bi_sector = r1_bio->sector +
1535                                                 conf->mirrors[i].rdev->data_offset;
1536                                         bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1537                                         bio->bi_end_io = raid1_end_write_request;
1538                                         bio->bi_rw = WRITE | do_sync;
1539                                         bio->bi_private = r1_bio;
1540                                         r1_bio->bios[i] = bio;
1541                                         generic_make_request(bio);
1542                                 }
1543                 } else {
1544                         int disk;
1545
1546                         /* we got a read error. Maybe the drive is bad.  Maybe just
1547                          * the block and we can fix it.
1548                          * We freeze all other IO, and try reading the block from
1549                          * other devices.  When we find one, we re-write
1550                          * and check it that fixes the read error.
1551                          * This is all done synchronously while the array is
1552                          * frozen
1553                          */
1554                         if (mddev->ro == 0) {
1555                                 freeze_array(conf);
1556                                 fix_read_error(conf, r1_bio->read_disk,
1557                                                r1_bio->sector,
1558                                                r1_bio->sectors);
1559                                 unfreeze_array(conf);
1560                         }
1561
1562                         bio = r1_bio->bios[r1_bio->read_disk];
1563                         if ((disk=read_balance(conf, r1_bio)) == -1) {
1564                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
1565                                        " read error for block %llu\n",
1566                                        bdevname(bio->bi_bdev,b),
1567                                        (unsigned long long)r1_bio->sector);
1568                                 raid_end_bio_io(r1_bio);
1569                         } else {
1570                                 const int do_sync = bio_sync(r1_bio->master_bio);
1571                                 r1_bio->bios[r1_bio->read_disk] =
1572                                         mddev->ro ? IO_BLOCKED : NULL;
1573                                 r1_bio->read_disk = disk;
1574                                 bio_put(bio);
1575                                 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1576                                 r1_bio->bios[r1_bio->read_disk] = bio;
1577                                 rdev = conf->mirrors[disk].rdev;
1578                                 if (printk_ratelimit())
1579                                         printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
1580                                                " another mirror\n",
1581                                                bdevname(rdev->bdev,b),
1582                                                (unsigned long long)r1_bio->sector);
1583                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1584                                 bio->bi_bdev = rdev->bdev;
1585                                 bio->bi_end_io = raid1_end_read_request;
1586                                 bio->bi_rw = READ | do_sync;
1587                                 bio->bi_private = r1_bio;
1588                                 unplug = 1;
1589                                 generic_make_request(bio);
1590                         }
1591                 }
1592         }
1593         spin_unlock_irqrestore(&conf->device_lock, flags);
1594         if (unplug)
1595                 unplug_slaves(mddev);
1596 }
1597
1598
1599 static int init_resync(conf_t *conf)
1600 {
1601         int buffs;
1602
1603         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1604         BUG_ON(conf->r1buf_pool);
1605         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1606                                           conf->poolinfo);
1607         if (!conf->r1buf_pool)
1608                 return -ENOMEM;
1609         conf->next_resync = 0;
1610         return 0;
1611 }
1612
1613 /*
1614  * perform a "sync" on one "block"
1615  *
1616  * We need to make sure that no normal I/O request - particularly write
1617  * requests - conflict with active sync requests.
1618  *
1619  * This is achieved by tracking pending requests and a 'barrier' concept
1620  * that can be installed to exclude normal IO requests.
1621  */
1622
1623 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1624 {
1625         conf_t *conf = mddev_to_conf(mddev);
1626         r1bio_t *r1_bio;
1627         struct bio *bio;
1628         sector_t max_sector, nr_sectors;
1629         int disk = -1;
1630         int i;
1631         int wonly = -1;
1632         int write_targets = 0, read_targets = 0;
1633         int sync_blocks;
1634         int still_degraded = 0;
1635
1636         if (!conf->r1buf_pool)
1637         {
1638 /*
1639                 printk("sync start - bitmap %p\n", mddev->bitmap);
1640 */
1641                 if (init_resync(conf))
1642                         return 0;
1643         }
1644
1645         max_sector = mddev->size << 1;
1646         if (sector_nr >= max_sector) {
1647                 /* If we aborted, we need to abort the
1648                  * sync on the 'current' bitmap chunk (there will
1649                  * only be one in raid1 resync.
1650                  * We can find the current addess in mddev->curr_resync
1651                  */
1652                 if (mddev->curr_resync < max_sector) /* aborted */
1653                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1654                                                 &sync_blocks, 1);
1655                 else /* completed sync */
1656                         conf->fullsync = 0;
1657
1658                 bitmap_close_sync(mddev->bitmap);
1659                 close_sync(conf);
1660                 return 0;
1661         }
1662
1663         if (mddev->bitmap == NULL &&
1664             mddev->recovery_cp == MaxSector &&
1665             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1666             conf->fullsync == 0) {
1667                 *skipped = 1;
1668                 return max_sector - sector_nr;
1669         }
1670         /* before building a request, check if we can skip these blocks..
1671          * This call the bitmap_start_sync doesn't actually record anything
1672          */
1673         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1674             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1675                 /* We can skip this block, and probably several more */
1676                 *skipped = 1;
1677                 return sync_blocks;
1678         }
1679         /*
1680          * If there is non-resync activity waiting for a turn,
1681          * and resync is going fast enough,
1682          * then let it though before starting on this new sync request.
1683          */
1684         if (!go_faster && conf->nr_waiting)
1685                 msleep_interruptible(1000);
1686
1687         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1688         raise_barrier(conf);
1689
1690         conf->next_resync = sector_nr;
1691
1692         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1693         rcu_read_lock();
1694         /*
1695          * If we get a correctably read error during resync or recovery,
1696          * we might want to read from a different device.  So we
1697          * flag all drives that could conceivably be read from for READ,
1698          * and any others (which will be non-In_sync devices) for WRITE.
1699          * If a read fails, we try reading from something else for which READ
1700          * is OK.
1701          */
1702
1703         r1_bio->mddev = mddev;
1704         r1_bio->sector = sector_nr;
1705         r1_bio->state = 0;
1706         set_bit(R1BIO_IsSync, &r1_bio->state);
1707
1708         for (i=0; i < conf->raid_disks; i++) {
1709                 mdk_rdev_t *rdev;
1710                 bio = r1_bio->bios[i];
1711
1712                 /* take from bio_init */
1713                 bio->bi_next = NULL;
1714                 bio->bi_flags |= 1 << BIO_UPTODATE;
1715                 bio->bi_rw = READ;
1716                 bio->bi_vcnt = 0;
1717                 bio->bi_idx = 0;
1718                 bio->bi_phys_segments = 0;
1719                 bio->bi_hw_segments = 0;
1720                 bio->bi_size = 0;
1721                 bio->bi_end_io = NULL;
1722                 bio->bi_private = NULL;
1723
1724                 rdev = rcu_dereference(conf->mirrors[i].rdev);
1725                 if (rdev == NULL ||
1726                            test_bit(Faulty, &rdev->flags)) {
1727                         still_degraded = 1;
1728                         continue;
1729                 } else if (!test_bit(In_sync, &rdev->flags)) {
1730                         bio->bi_rw = WRITE;
1731                         bio->bi_end_io = end_sync_write;
1732                         write_targets ++;
1733                 } else {
1734                         /* may need to read from here */
1735                         bio->bi_rw = READ;
1736                         bio->bi_end_io = end_sync_read;
1737                         if (test_bit(WriteMostly, &rdev->flags)) {
1738                                 if (wonly < 0)
1739                                         wonly = i;
1740                         } else {
1741                                 if (disk < 0)
1742                                         disk = i;
1743                         }
1744                         read_targets++;
1745                 }
1746                 atomic_inc(&rdev->nr_pending);
1747                 bio->bi_sector = sector_nr + rdev->data_offset;
1748                 bio->bi_bdev = rdev->bdev;
1749                 bio->bi_private = r1_bio;
1750         }
1751         rcu_read_unlock();
1752         if (disk < 0)
1753                 disk = wonly;
1754         r1_bio->read_disk = disk;
1755
1756         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1757                 /* extra read targets are also write targets */
1758                 write_targets += read_targets-1;
1759
1760         if (write_targets == 0 || read_targets == 0) {
1761                 /* There is nowhere to write, so all non-sync
1762                  * drives must be failed - so we are finished
1763                  */
1764                 sector_t rv = max_sector - sector_nr;
1765                 *skipped = 1;
1766                 put_buf(r1_bio);
1767                 return rv;
1768         }
1769
1770         if (max_sector > mddev->resync_max)
1771                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1772         nr_sectors = 0;
1773         sync_blocks = 0;
1774         do {
1775                 struct page *page;
1776                 int len = PAGE_SIZE;
1777                 if (sector_nr + (len>>9) > max_sector)
1778                         len = (max_sector - sector_nr) << 9;
1779                 if (len == 0)
1780                         break;
1781                 if (sync_blocks == 0) {
1782                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1783                                                &sync_blocks, still_degraded) &&
1784                             !conf->fullsync &&
1785                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1786                                 break;
1787                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1788                         if (len > (sync_blocks<<9))
1789                                 len = sync_blocks<<9;
1790                 }
1791
1792                 for (i=0 ; i < conf->raid_disks; i++) {
1793                         bio = r1_bio->bios[i];
1794                         if (bio->bi_end_io) {
1795                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1796                                 if (bio_add_page(bio, page, len, 0) == 0) {
1797                                         /* stop here */
1798                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1799                                         while (i > 0) {
1800                                                 i--;
1801                                                 bio = r1_bio->bios[i];
1802                                                 if (bio->bi_end_io==NULL)
1803                                                         continue;
1804                                                 /* remove last page from this bio */
1805                                                 bio->bi_vcnt--;
1806                                                 bio->bi_size -= len;
1807                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1808                                         }
1809                                         goto bio_full;
1810                                 }
1811                         }
1812                 }
1813                 nr_sectors += len>>9;
1814                 sector_nr += len>>9;
1815                 sync_blocks -= (len>>9);
1816         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1817  bio_full:
1818         r1_bio->sectors = nr_sectors;
1819
1820         /* For a user-requested sync, we read all readable devices and do a
1821          * compare
1822          */
1823         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1824                 atomic_set(&r1_bio->remaining, read_targets);
1825                 for (i=0; i<conf->raid_disks; i++) {
1826                         bio = r1_bio->bios[i];
1827                         if (bio->bi_end_io == end_sync_read) {
1828                                 md_sync_acct(bio->bi_bdev, nr_sectors);
1829                                 generic_make_request(bio);
1830                         }
1831                 }
1832         } else {
1833                 atomic_set(&r1_bio->remaining, 1);
1834                 bio = r1_bio->bios[r1_bio->read_disk];
1835                 md_sync_acct(bio->bi_bdev, nr_sectors);
1836                 generic_make_request(bio);
1837
1838         }
1839         return nr_sectors;
1840 }
1841
1842 static int run(mddev_t *mddev)
1843 {
1844         conf_t *conf;
1845         int i, j, disk_idx;
1846         mirror_info_t *disk;
1847         mdk_rdev_t *rdev;
1848         struct list_head *tmp;
1849
1850         if (mddev->level != 1) {
1851                 printk("raid1: %s: raid level not set to mirroring (%d)\n",
1852                        mdname(mddev), mddev->level);
1853                 goto out;
1854         }
1855         if (mddev->reshape_position != MaxSector) {
1856                 printk("raid1: %s: reshape_position set but not supported\n",
1857                        mdname(mddev));
1858                 goto out;
1859         }
1860         /*
1861          * copy the already verified devices into our private RAID1
1862          * bookkeeping area. [whatever we allocate in run(),
1863          * should be freed in stop()]
1864          */
1865         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1866         mddev->private = conf;
1867         if (!conf)
1868                 goto out_no_mem;
1869
1870         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1871                                  GFP_KERNEL);
1872         if (!conf->mirrors)
1873                 goto out_no_mem;
1874
1875         conf->tmppage = alloc_page(GFP_KERNEL);
1876         if (!conf->tmppage)
1877                 goto out_no_mem;
1878
1879         conf->poolinfo = kmalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1880         if (!conf->poolinfo)
1881                 goto out_no_mem;
1882         conf->poolinfo->mddev = mddev;
1883         conf->poolinfo->raid_disks = mddev->raid_disks;
1884         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1885                                           r1bio_pool_free,
1886                                           conf->poolinfo);
1887         if (!conf->r1bio_pool)
1888                 goto out_no_mem;
1889
1890         rdev_for_each(rdev, tmp, mddev) {
1891                 disk_idx = rdev->raid_disk;
1892                 if (disk_idx >= mddev->raid_disks
1893                     || disk_idx < 0)
1894                         continue;
1895                 disk = conf->mirrors + disk_idx;
1896
1897                 disk->rdev = rdev;
1898
1899                 blk_queue_stack_limits(mddev->queue,
1900                                        rdev->bdev->bd_disk->queue);
1901                 /* as we don't honour merge_bvec_fn, we must never risk
1902                  * violating it, so limit ->max_sector to one PAGE, as
1903                  * a one page request is never in violation.
1904                  */
1905                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1906                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
1907                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1908
1909                 disk->head_position = 0;
1910         }
1911         conf->raid_disks = mddev->raid_disks;
1912         conf->mddev = mddev;
1913         spin_lock_init(&conf->device_lock);
1914         INIT_LIST_HEAD(&conf->retry_list);
1915
1916         spin_lock_init(&conf->resync_lock);
1917         init_waitqueue_head(&conf->wait_barrier);
1918
1919         bio_list_init(&conf->pending_bio_list);
1920         bio_list_init(&conf->flushing_bio_list);
1921
1922
1923         mddev->degraded = 0;
1924         for (i = 0; i < conf->raid_disks; i++) {
1925
1926                 disk = conf->mirrors + i;
1927
1928                 if (!disk->rdev ||
1929                     !test_bit(In_sync, &disk->rdev->flags)) {
1930                         disk->head_position = 0;
1931                         mddev->degraded++;
1932                         if (disk->rdev)
1933                                 conf->fullsync = 1;
1934                 }
1935         }
1936         if (mddev->degraded == conf->raid_disks) {
1937                 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
1938                         mdname(mddev));
1939                 goto out_free_conf;
1940         }
1941         if (conf->raid_disks - mddev->degraded == 1)
1942                 mddev->recovery_cp = MaxSector;
1943
1944         /*
1945          * find the first working one and use it as a starting point
1946          * to read balancing.
1947          */
1948         for (j = 0; j < conf->raid_disks &&
1949                      (!conf->mirrors[j].rdev ||
1950                       !test_bit(In_sync, &conf->mirrors[j].rdev->flags)) ; j++)
1951                 /* nothing */;
1952         conf->last_used = j;
1953
1954
1955         mddev->thread = md_register_thread(raid1d, mddev, "%s_raid1");
1956         if (!mddev->thread) {
1957                 printk(KERN_ERR
1958                        "raid1: couldn't allocate thread for %s\n",
1959                        mdname(mddev));
1960                 goto out_free_conf;
1961         }
1962
1963         printk(KERN_INFO 
1964                 "raid1: raid set %s active with %d out of %d mirrors\n",
1965                 mdname(mddev), mddev->raid_disks - mddev->degraded, 
1966                 mddev->raid_disks);
1967         /*
1968          * Ok, everything is just fine now
1969          */
1970         mddev->array_size = mddev->size;
1971
1972         mddev->queue->unplug_fn = raid1_unplug;
1973         mddev->queue->backing_dev_info.congested_fn = raid1_congested;
1974         mddev->queue->backing_dev_info.congested_data = mddev;
1975
1976         return 0;
1977
1978 out_no_mem:
1979         printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
1980                mdname(mddev));
1981
1982 out_free_conf:
1983         if (conf) {
1984                 if (conf->r1bio_pool)
1985                         mempool_destroy(conf->r1bio_pool);
1986                 kfree(conf->mirrors);
1987                 safe_put_page(conf->tmppage);
1988                 kfree(conf->poolinfo);
1989                 kfree(conf);
1990                 mddev->private = NULL;
1991         }
1992 out:
1993         return -EIO;
1994 }
1995
1996 static int stop(mddev_t *mddev)
1997 {
1998         conf_t *conf = mddev_to_conf(mddev);
1999         struct bitmap *bitmap = mddev->bitmap;
2000         int behind_wait = 0;
2001
2002         /* wait for behind writes to complete */
2003         while (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2004                 behind_wait++;
2005                 printk(KERN_INFO "raid1: behind writes in progress on device %s, waiting to stop (%d)\n", mdname(mddev), behind_wait);
2006                 set_current_state(TASK_UNINTERRUPTIBLE);
2007                 schedule_timeout(HZ); /* wait a second */
2008                 /* need to kick something here to make sure I/O goes? */
2009         }
2010
2011         md_unregister_thread(mddev->thread);
2012         mddev->thread = NULL;
2013         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2014         if (conf->r1bio_pool)
2015                 mempool_destroy(conf->r1bio_pool);
2016         kfree(conf->mirrors);
2017         kfree(conf->poolinfo);
2018         kfree(conf);
2019         mddev->private = NULL;
2020         return 0;
2021 }
2022
2023 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2024 {
2025         /* no resync is happening, and there is enough space
2026          * on all devices, so we can resize.
2027          * We need to make sure resync covers any new space.
2028          * If the array is shrinking we should possibly wait until
2029          * any io in the removed space completes, but it hardly seems
2030          * worth it.
2031          */
2032         mddev->array_size = sectors>>1;
2033         set_capacity(mddev->gendisk, mddev->array_size << 1);
2034         mddev->changed = 1;
2035         if (mddev->array_size > mddev->size && mddev->recovery_cp == MaxSector) {
2036                 mddev->recovery_cp = mddev->size << 1;
2037                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2038         }
2039         mddev->size = mddev->array_size;
2040         mddev->resync_max_sectors = sectors;
2041         return 0;
2042 }
2043
2044 static int raid1_reshape(mddev_t *mddev)
2045 {
2046         /* We need to:
2047          * 1/ resize the r1bio_pool
2048          * 2/ resize conf->mirrors
2049          *
2050          * We allocate a new r1bio_pool if we can.
2051          * Then raise a device barrier and wait until all IO stops.
2052          * Then resize conf->mirrors and swap in the new r1bio pool.
2053          *
2054          * At the same time, we "pack" the devices so that all the missing
2055          * devices have the higher raid_disk numbers.
2056          */
2057         mempool_t *newpool, *oldpool;
2058         struct pool_info *newpoolinfo;
2059         mirror_info_t *newmirrors;
2060         conf_t *conf = mddev_to_conf(mddev);
2061         int cnt, raid_disks;
2062         unsigned long flags;
2063         int d, d2;
2064
2065         /* Cannot change chunk_size, layout, or level */
2066         if (mddev->chunk_size != mddev->new_chunk ||
2067             mddev->layout != mddev->new_layout ||
2068             mddev->level != mddev->new_level) {
2069                 mddev->new_chunk = mddev->chunk_size;
2070                 mddev->new_layout = mddev->layout;
2071                 mddev->new_level = mddev->level;
2072                 return -EINVAL;
2073         }
2074
2075         md_allow_write(mddev);
2076
2077         raid_disks = mddev->raid_disks + mddev->delta_disks;
2078
2079         if (raid_disks < conf->raid_disks) {
2080                 cnt=0;
2081                 for (d= 0; d < conf->raid_disks; d++)
2082                         if (conf->mirrors[d].rdev)
2083                                 cnt++;
2084                 if (cnt > raid_disks)
2085                         return -EBUSY;
2086         }
2087
2088         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2089         if (!newpoolinfo)
2090                 return -ENOMEM;
2091         newpoolinfo->mddev = mddev;
2092         newpoolinfo->raid_disks = raid_disks;
2093
2094         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2095                                  r1bio_pool_free, newpoolinfo);
2096         if (!newpool) {
2097                 kfree(newpoolinfo);
2098                 return -ENOMEM;
2099         }
2100         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2101         if (!newmirrors) {
2102                 kfree(newpoolinfo);
2103                 mempool_destroy(newpool);
2104                 return -ENOMEM;
2105         }
2106
2107         raise_barrier(conf);
2108
2109         /* ok, everything is stopped */
2110         oldpool = conf->r1bio_pool;
2111         conf->r1bio_pool = newpool;
2112
2113         for (d = d2 = 0; d < conf->raid_disks; d++) {
2114                 mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2115                 if (rdev && rdev->raid_disk != d2) {
2116                         char nm[20];
2117                         sprintf(nm, "rd%d", rdev->raid_disk);
2118                         sysfs_remove_link(&mddev->kobj, nm);
2119                         rdev->raid_disk = d2;
2120                         sprintf(nm, "rd%d", rdev->raid_disk);
2121                         sysfs_remove_link(&mddev->kobj, nm);
2122                         if (sysfs_create_link(&mddev->kobj,
2123                                               &rdev->kobj, nm))
2124                                 printk(KERN_WARNING
2125                                        "md/raid1: cannot register "
2126                                        "%s for %s\n",
2127                                        nm, mdname(mddev));
2128                 }
2129                 if (rdev)
2130                         newmirrors[d2++].rdev = rdev;
2131         }
2132         kfree(conf->mirrors);
2133         conf->mirrors = newmirrors;
2134         kfree(conf->poolinfo);
2135         conf->poolinfo = newpoolinfo;
2136
2137         spin_lock_irqsave(&conf->device_lock, flags);
2138         mddev->degraded += (raid_disks - conf->raid_disks);
2139         spin_unlock_irqrestore(&conf->device_lock, flags);
2140         conf->raid_disks = mddev->raid_disks = raid_disks;
2141         mddev->delta_disks = 0;
2142
2143         conf->last_used = 0; /* just make sure it is in-range */
2144         lower_barrier(conf);
2145
2146         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2147         md_wakeup_thread(mddev->thread);
2148
2149         mempool_destroy(oldpool);
2150         return 0;
2151 }
2152
2153 static void raid1_quiesce(mddev_t *mddev, int state)
2154 {
2155         conf_t *conf = mddev_to_conf(mddev);
2156
2157         switch(state) {
2158         case 1:
2159                 raise_barrier(conf);
2160                 break;
2161         case 0:
2162                 lower_barrier(conf);
2163                 break;
2164         }
2165 }
2166
2167
2168 static struct mdk_personality raid1_personality =
2169 {
2170         .name           = "raid1",
2171         .level          = 1,
2172         .owner          = THIS_MODULE,
2173         .make_request   = make_request,
2174         .run            = run,
2175         .stop           = stop,
2176         .status         = status,
2177         .error_handler  = error,
2178         .hot_add_disk   = raid1_add_disk,
2179         .hot_remove_disk= raid1_remove_disk,
2180         .spare_active   = raid1_spare_active,
2181         .sync_request   = sync_request,
2182         .resize         = raid1_resize,
2183         .check_reshape  = raid1_reshape,
2184         .quiesce        = raid1_quiesce,
2185 };
2186
2187 static int __init raid_init(void)
2188 {
2189         return register_md_personality(&raid1_personality);
2190 }
2191
2192 static void raid_exit(void)
2193 {
2194         unregister_md_personality(&raid1_personality);
2195 }
2196
2197 module_init(raid_init);
2198 module_exit(raid_exit);
2199 MODULE_LICENSE("GPL");
2200 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2201 MODULE_ALIAS("md-raid1");
2202 MODULE_ALIAS("md-level-1");