Merge master.kernel.org:/pub/scm/linux/kernel/git/davej/cpufreq
[linux-2.6] / drivers / md / raid1.c
1 /*
2  * raid1.c : Multiple Devices driver for Linux
3  *
4  * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
5  *
6  * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
7  *
8  * RAID-1 management functions.
9  *
10  * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
11  *
12  * Fixes to reconstruction by Jakob Ã˜stergaard" <jakob@ostenfeld.dk>
13  * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
14  *
15  * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
16  * bitmapped intelligence in resync:
17  *
18  *      - bitmap marked during normal i/o
19  *      - bitmap used to skip nondirty blocks during sync
20  *
21  * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
22  * - persistent bitmap code
23  *
24  * This program is free software; you can redistribute it and/or modify
25  * it under the terms of the GNU General Public License as published by
26  * the Free Software Foundation; either version 2, or (at your option)
27  * any later version.
28  *
29  * You should have received a copy of the GNU General Public License
30  * (for example /usr/src/linux/COPYING); if not, write to the Free
31  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
32  */
33
34 #include "dm-bio-list.h"
35 #include <linux/raid/raid1.h>
36 #include <linux/raid/bitmap.h>
37
38 #define DEBUG 0
39 #if DEBUG
40 #define PRINTK(x...) printk(x)
41 #else
42 #define PRINTK(x...)
43 #endif
44
45 /*
46  * Number of guaranteed r1bios in case of extreme VM load:
47  */
48 #define NR_RAID1_BIOS 256
49
50
51 static void unplug_slaves(mddev_t *mddev);
52
53 static void allow_barrier(conf_t *conf);
54 static void lower_barrier(conf_t *conf);
55
56 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
57 {
58         struct pool_info *pi = data;
59         r1bio_t *r1_bio;
60         int size = offsetof(r1bio_t, bios[pi->raid_disks]);
61
62         /* allocate a r1bio with room for raid_disks entries in the bios array */
63         r1_bio = kzalloc(size, gfp_flags);
64         if (!r1_bio)
65                 unplug_slaves(pi->mddev);
66
67         return r1_bio;
68 }
69
70 static void r1bio_pool_free(void *r1_bio, void *data)
71 {
72         kfree(r1_bio);
73 }
74
75 #define RESYNC_BLOCK_SIZE (64*1024)
76 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
77 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
78 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
79 #define RESYNC_WINDOW (2048*1024)
80
81 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
82 {
83         struct pool_info *pi = data;
84         struct page *page;
85         r1bio_t *r1_bio;
86         struct bio *bio;
87         int i, j;
88
89         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
90         if (!r1_bio) {
91                 unplug_slaves(pi->mddev);
92                 return NULL;
93         }
94
95         /*
96          * Allocate bios : 1 for reading, n-1 for writing
97          */
98         for (j = pi->raid_disks ; j-- ; ) {
99                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
100                 if (!bio)
101                         goto out_free_bio;
102                 r1_bio->bios[j] = bio;
103         }
104         /*
105          * Allocate RESYNC_PAGES data pages and attach them to
106          * the first bio.
107          * If this is a user-requested check/repair, allocate
108          * RESYNC_PAGES for each bio.
109          */
110         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
111                 j = pi->raid_disks;
112         else
113                 j = 1;
114         while(j--) {
115                 bio = r1_bio->bios[j];
116                 for (i = 0; i < RESYNC_PAGES; i++) {
117                         page = alloc_page(gfp_flags);
118                         if (unlikely(!page))
119                                 goto out_free_pages;
120
121                         bio->bi_io_vec[i].bv_page = page;
122                 }
123         }
124         /* If not user-requests, copy the page pointers to all bios */
125         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
126                 for (i=0; i<RESYNC_PAGES ; i++)
127                         for (j=1; j<pi->raid_disks; j++)
128                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
129                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
130         }
131
132         r1_bio->master_bio = NULL;
133
134         return r1_bio;
135
136 out_free_pages:
137         for (i=0; i < RESYNC_PAGES ; i++)
138                 for (j=0 ; j < pi->raid_disks; j++)
139                         safe_put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
140         j = -1;
141 out_free_bio:
142         while ( ++j < pi->raid_disks )
143                 bio_put(r1_bio->bios[j]);
144         r1bio_pool_free(r1_bio, data);
145         return NULL;
146 }
147
148 static void r1buf_pool_free(void *__r1_bio, void *data)
149 {
150         struct pool_info *pi = data;
151         int i,j;
152         r1bio_t *r1bio = __r1_bio;
153
154         for (i = 0; i < RESYNC_PAGES; i++)
155                 for (j = pi->raid_disks; j-- ;) {
156                         if (j == 0 ||
157                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
158                             r1bio->bios[0]->bi_io_vec[i].bv_page)
159                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
160                 }
161         for (i=0 ; i < pi->raid_disks; i++)
162                 bio_put(r1bio->bios[i]);
163
164         r1bio_pool_free(r1bio, data);
165 }
166
167 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
168 {
169         int i;
170
171         for (i = 0; i < conf->raid_disks; i++) {
172                 struct bio **bio = r1_bio->bios + i;
173                 if (*bio && *bio != IO_BLOCKED)
174                         bio_put(*bio);
175                 *bio = NULL;
176         }
177 }
178
179 static void free_r1bio(r1bio_t *r1_bio)
180 {
181         conf_t *conf = mddev_to_conf(r1_bio->mddev);
182
183         /*
184          * Wake up any possible resync thread that waits for the device
185          * to go idle.
186          */
187         allow_barrier(conf);
188
189         put_all_bios(conf, r1_bio);
190         mempool_free(r1_bio, conf->r1bio_pool);
191 }
192
193 static void put_buf(r1bio_t *r1_bio)
194 {
195         conf_t *conf = mddev_to_conf(r1_bio->mddev);
196         int i;
197
198         for (i=0; i<conf->raid_disks; i++) {
199                 struct bio *bio = r1_bio->bios[i];
200                 if (bio->bi_end_io)
201                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
202         }
203
204         mempool_free(r1_bio, conf->r1buf_pool);
205
206         lower_barrier(conf);
207 }
208
209 static void reschedule_retry(r1bio_t *r1_bio)
210 {
211         unsigned long flags;
212         mddev_t *mddev = r1_bio->mddev;
213         conf_t *conf = mddev_to_conf(mddev);
214
215         spin_lock_irqsave(&conf->device_lock, flags);
216         list_add(&r1_bio->retry_list, &conf->retry_list);
217         conf->nr_queued ++;
218         spin_unlock_irqrestore(&conf->device_lock, flags);
219
220         wake_up(&conf->wait_barrier);
221         md_wakeup_thread(mddev->thread);
222 }
223
224 /*
225  * raid_end_bio_io() is called when we have finished servicing a mirrored
226  * operation and are ready to return a success/failure code to the buffer
227  * cache layer.
228  */
229 static void raid_end_bio_io(r1bio_t *r1_bio)
230 {
231         struct bio *bio = r1_bio->master_bio;
232
233         /* if nobody has done the final endio yet, do it now */
234         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
235                 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
236                         (bio_data_dir(bio) == WRITE) ? "write" : "read",
237                         (unsigned long long) bio->bi_sector,
238                         (unsigned long long) bio->bi_sector +
239                                 (bio->bi_size >> 9) - 1);
240
241                 bio_endio(bio, bio->bi_size,
242                         test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
243         }
244         free_r1bio(r1_bio);
245 }
246
247 /*
248  * Update disk head position estimator based on IRQ completion info.
249  */
250 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
251 {
252         conf_t *conf = mddev_to_conf(r1_bio->mddev);
253
254         conf->mirrors[disk].head_position =
255                 r1_bio->sector + (r1_bio->sectors);
256 }
257
258 static int raid1_end_read_request(struct bio *bio, unsigned int bytes_done, int error)
259 {
260         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
261         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
262         int mirror;
263         conf_t *conf = mddev_to_conf(r1_bio->mddev);
264
265         if (bio->bi_size)
266                 return 1;
267         
268         mirror = r1_bio->read_disk;
269         /*
270          * this branch is our 'one mirror IO has finished' event handler:
271          */
272         update_head_pos(mirror, r1_bio);
273
274         if (uptodate || (conf->raid_disks - conf->mddev->degraded) <= 1) {
275                 /*
276                  * Set R1BIO_Uptodate in our master bio, so that
277                  * we will return a good error code for to the higher
278                  * levels even if IO on some other mirrored buffer fails.
279                  *
280                  * The 'master' represents the composite IO operation to
281                  * user-side. So if something waits for IO, then it will
282                  * wait for the 'master' bio.
283                  */
284                 if (uptodate)
285                         set_bit(R1BIO_Uptodate, &r1_bio->state);
286
287                 raid_end_bio_io(r1_bio);
288         } else {
289                 /*
290                  * oops, read error:
291                  */
292                 char b[BDEVNAME_SIZE];
293                 if (printk_ratelimit())
294                         printk(KERN_ERR "raid1: %s: rescheduling sector %llu\n",
295                                bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
296                 reschedule_retry(r1_bio);
297         }
298
299         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
300         return 0;
301 }
302
303 static int raid1_end_write_request(struct bio *bio, unsigned int bytes_done, int error)
304 {
305         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
306         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
307         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
308         conf_t *conf = mddev_to_conf(r1_bio->mddev);
309         struct bio *to_put = NULL;
310
311         if (bio->bi_size)
312                 return 1;
313
314         for (mirror = 0; mirror < conf->raid_disks; mirror++)
315                 if (r1_bio->bios[mirror] == bio)
316                         break;
317
318         if (error == -EOPNOTSUPP && test_bit(R1BIO_Barrier, &r1_bio->state)) {
319                 set_bit(BarriersNotsupp, &conf->mirrors[mirror].rdev->flags);
320                 set_bit(R1BIO_BarrierRetry, &r1_bio->state);
321                 r1_bio->mddev->barriers_work = 0;
322                 /* Don't rdev_dec_pending in this branch - keep it for the retry */
323         } else {
324                 /*
325                  * this branch is our 'one mirror IO has finished' event handler:
326                  */
327                 r1_bio->bios[mirror] = NULL;
328                 to_put = bio;
329                 if (!uptodate) {
330                         md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
331                         /* an I/O failed, we can't clear the bitmap */
332                         set_bit(R1BIO_Degraded, &r1_bio->state);
333                 } else
334                         /*
335                          * Set R1BIO_Uptodate in our master bio, so that
336                          * we will return a good error code for to the higher
337                          * levels even if IO on some other mirrored buffer fails.
338                          *
339                          * The 'master' represents the composite IO operation to
340                          * user-side. So if something waits for IO, then it will
341                          * wait for the 'master' bio.
342                          */
343                         set_bit(R1BIO_Uptodate, &r1_bio->state);
344
345                 update_head_pos(mirror, r1_bio);
346
347                 if (behind) {
348                         if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
349                                 atomic_dec(&r1_bio->behind_remaining);
350
351                         /* In behind mode, we ACK the master bio once the I/O has safely
352                          * reached all non-writemostly disks. Setting the Returned bit
353                          * ensures that this gets done only once -- we don't ever want to
354                          * return -EIO here, instead we'll wait */
355
356                         if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
357                             test_bit(R1BIO_Uptodate, &r1_bio->state)) {
358                                 /* Maybe we can return now */
359                                 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
360                                         struct bio *mbio = r1_bio->master_bio;
361                                         PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
362                                                (unsigned long long) mbio->bi_sector,
363                                                (unsigned long long) mbio->bi_sector +
364                                                (mbio->bi_size >> 9) - 1);
365                                         bio_endio(mbio, mbio->bi_size, 0);
366                                 }
367                         }
368                 }
369                 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
370         }
371         /*
372          *
373          * Let's see if all mirrored write operations have finished
374          * already.
375          */
376         if (atomic_dec_and_test(&r1_bio->remaining)) {
377                 if (test_bit(R1BIO_BarrierRetry, &r1_bio->state))
378                         reschedule_retry(r1_bio);
379                 else {
380                         /* it really is the end of this request */
381                         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
382                                 /* free extra copy of the data pages */
383                                 int i = bio->bi_vcnt;
384                                 while (i--)
385                                         safe_put_page(bio->bi_io_vec[i].bv_page);
386                         }
387                         /* clear the bitmap if all writes complete successfully */
388                         bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
389                                         r1_bio->sectors,
390                                         !test_bit(R1BIO_Degraded, &r1_bio->state),
391                                         behind);
392                         md_write_end(r1_bio->mddev);
393                         raid_end_bio_io(r1_bio);
394                 }
395         }
396
397         if (to_put)
398                 bio_put(to_put);
399
400         return 0;
401 }
402
403
404 /*
405  * This routine returns the disk from which the requested read should
406  * be done. There is a per-array 'next expected sequential IO' sector
407  * number - if this matches on the next IO then we use the last disk.
408  * There is also a per-disk 'last know head position' sector that is
409  * maintained from IRQ contexts, both the normal and the resync IO
410  * completion handlers update this position correctly. If there is no
411  * perfect sequential match then we pick the disk whose head is closest.
412  *
413  * If there are 2 mirrors in the same 2 devices, performance degrades
414  * because position is mirror, not device based.
415  *
416  * The rdev for the device selected will have nr_pending incremented.
417  */
418 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
419 {
420         const unsigned long this_sector = r1_bio->sector;
421         int new_disk = conf->last_used, disk = new_disk;
422         int wonly_disk = -1;
423         const int sectors = r1_bio->sectors;
424         sector_t new_distance, current_distance;
425         mdk_rdev_t *rdev;
426
427         rcu_read_lock();
428         /*
429          * Check if we can balance. We can balance on the whole
430          * device if no resync is going on, or below the resync window.
431          * We take the first readable disk when above the resync window.
432          */
433  retry:
434         if (conf->mddev->recovery_cp < MaxSector &&
435             (this_sector + sectors >= conf->next_resync)) {
436                 /* Choose the first operation device, for consistancy */
437                 new_disk = 0;
438
439                 for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
440                      r1_bio->bios[new_disk] == IO_BLOCKED ||
441                      !rdev || !test_bit(In_sync, &rdev->flags)
442                              || test_bit(WriteMostly, &rdev->flags);
443                      rdev = rcu_dereference(conf->mirrors[++new_disk].rdev)) {
444
445                         if (rdev && test_bit(In_sync, &rdev->flags) &&
446                                 r1_bio->bios[new_disk] != IO_BLOCKED)
447                                 wonly_disk = new_disk;
448
449                         if (new_disk == conf->raid_disks - 1) {
450                                 new_disk = wonly_disk;
451                                 break;
452                         }
453                 }
454                 goto rb_out;
455         }
456
457
458         /* make sure the disk is operational */
459         for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
460              r1_bio->bios[new_disk] == IO_BLOCKED ||
461              !rdev || !test_bit(In_sync, &rdev->flags) ||
462                      test_bit(WriteMostly, &rdev->flags);
463              rdev = rcu_dereference(conf->mirrors[new_disk].rdev)) {
464
465                 if (rdev && test_bit(In_sync, &rdev->flags) &&
466                     r1_bio->bios[new_disk] != IO_BLOCKED)
467                         wonly_disk = new_disk;
468
469                 if (new_disk <= 0)
470                         new_disk = conf->raid_disks;
471                 new_disk--;
472                 if (new_disk == disk) {
473                         new_disk = wonly_disk;
474                         break;
475                 }
476         }
477
478         if (new_disk < 0)
479                 goto rb_out;
480
481         disk = new_disk;
482         /* now disk == new_disk == starting point for search */
483
484         /*
485          * Don't change to another disk for sequential reads:
486          */
487         if (conf->next_seq_sect == this_sector)
488                 goto rb_out;
489         if (this_sector == conf->mirrors[new_disk].head_position)
490                 goto rb_out;
491
492         current_distance = abs(this_sector - conf->mirrors[disk].head_position);
493
494         /* Find the disk whose head is closest */
495
496         do {
497                 if (disk <= 0)
498                         disk = conf->raid_disks;
499                 disk--;
500
501                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
502
503                 if (!rdev || r1_bio->bios[disk] == IO_BLOCKED ||
504                     !test_bit(In_sync, &rdev->flags) ||
505                     test_bit(WriteMostly, &rdev->flags))
506                         continue;
507
508                 if (!atomic_read(&rdev->nr_pending)) {
509                         new_disk = disk;
510                         break;
511                 }
512                 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
513                 if (new_distance < current_distance) {
514                         current_distance = new_distance;
515                         new_disk = disk;
516                 }
517         } while (disk != conf->last_used);
518
519  rb_out:
520
521
522         if (new_disk >= 0) {
523                 rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
524                 if (!rdev)
525                         goto retry;
526                 atomic_inc(&rdev->nr_pending);
527                 if (!test_bit(In_sync, &rdev->flags)) {
528                         /* cannot risk returning a device that failed
529                          * before we inc'ed nr_pending
530                          */
531                         rdev_dec_pending(rdev, conf->mddev);
532                         goto retry;
533                 }
534                 conf->next_seq_sect = this_sector + sectors;
535                 conf->last_used = new_disk;
536         }
537         rcu_read_unlock();
538
539         return new_disk;
540 }
541
542 static void unplug_slaves(mddev_t *mddev)
543 {
544         conf_t *conf = mddev_to_conf(mddev);
545         int i;
546
547         rcu_read_lock();
548         for (i=0; i<mddev->raid_disks; i++) {
549                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
550                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
551                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
552
553                         atomic_inc(&rdev->nr_pending);
554                         rcu_read_unlock();
555
556                         if (r_queue->unplug_fn)
557                                 r_queue->unplug_fn(r_queue);
558
559                         rdev_dec_pending(rdev, mddev);
560                         rcu_read_lock();
561                 }
562         }
563         rcu_read_unlock();
564 }
565
566 static void raid1_unplug(request_queue_t *q)
567 {
568         mddev_t *mddev = q->queuedata;
569
570         unplug_slaves(mddev);
571         md_wakeup_thread(mddev->thread);
572 }
573
574 static int raid1_issue_flush(request_queue_t *q, struct gendisk *disk,
575                              sector_t *error_sector)
576 {
577         mddev_t *mddev = q->queuedata;
578         conf_t *conf = mddev_to_conf(mddev);
579         int i, ret = 0;
580
581         rcu_read_lock();
582         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
583                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
584                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
585                         struct block_device *bdev = rdev->bdev;
586                         request_queue_t *r_queue = bdev_get_queue(bdev);
587
588                         if (!r_queue->issue_flush_fn)
589                                 ret = -EOPNOTSUPP;
590                         else {
591                                 atomic_inc(&rdev->nr_pending);
592                                 rcu_read_unlock();
593                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
594                                                               error_sector);
595                                 rdev_dec_pending(rdev, mddev);
596                                 rcu_read_lock();
597                         }
598                 }
599         }
600         rcu_read_unlock();
601         return ret;
602 }
603
604 static int raid1_congested(void *data, int bits)
605 {
606         mddev_t *mddev = data;
607         conf_t *conf = mddev_to_conf(mddev);
608         int i, ret = 0;
609
610         rcu_read_lock();
611         for (i = 0; i < mddev->raid_disks; i++) {
612                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
613                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
614                         request_queue_t *q = bdev_get_queue(rdev->bdev);
615
616                         /* Note the '|| 1' - when read_balance prefers
617                          * non-congested targets, it can be removed
618                          */
619                         if ((bits & (1<<BDI_write_congested)) || 1)
620                                 ret |= bdi_congested(&q->backing_dev_info, bits);
621                         else
622                                 ret &= bdi_congested(&q->backing_dev_info, bits);
623                 }
624         }
625         rcu_read_unlock();
626         return ret;
627 }
628
629
630 /* Barriers....
631  * Sometimes we need to suspend IO while we do something else,
632  * either some resync/recovery, or reconfigure the array.
633  * To do this we raise a 'barrier'.
634  * The 'barrier' is a counter that can be raised multiple times
635  * to count how many activities are happening which preclude
636  * normal IO.
637  * We can only raise the barrier if there is no pending IO.
638  * i.e. if nr_pending == 0.
639  * We choose only to raise the barrier if no-one is waiting for the
640  * barrier to go down.  This means that as soon as an IO request
641  * is ready, no other operations which require a barrier will start
642  * until the IO request has had a chance.
643  *
644  * So: regular IO calls 'wait_barrier'.  When that returns there
645  *    is no backgroup IO happening,  It must arrange to call
646  *    allow_barrier when it has finished its IO.
647  * backgroup IO calls must call raise_barrier.  Once that returns
648  *    there is no normal IO happeing.  It must arrange to call
649  *    lower_barrier when the particular background IO completes.
650  */
651 #define RESYNC_DEPTH 32
652
653 static void raise_barrier(conf_t *conf)
654 {
655         spin_lock_irq(&conf->resync_lock);
656
657         /* Wait until no block IO is waiting */
658         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
659                             conf->resync_lock,
660                             raid1_unplug(conf->mddev->queue));
661
662         /* block any new IO from starting */
663         conf->barrier++;
664
665         /* No wait for all pending IO to complete */
666         wait_event_lock_irq(conf->wait_barrier,
667                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
668                             conf->resync_lock,
669                             raid1_unplug(conf->mddev->queue));
670
671         spin_unlock_irq(&conf->resync_lock);
672 }
673
674 static void lower_barrier(conf_t *conf)
675 {
676         unsigned long flags;
677         spin_lock_irqsave(&conf->resync_lock, flags);
678         conf->barrier--;
679         spin_unlock_irqrestore(&conf->resync_lock, flags);
680         wake_up(&conf->wait_barrier);
681 }
682
683 static void wait_barrier(conf_t *conf)
684 {
685         spin_lock_irq(&conf->resync_lock);
686         if (conf->barrier) {
687                 conf->nr_waiting++;
688                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
689                                     conf->resync_lock,
690                                     raid1_unplug(conf->mddev->queue));
691                 conf->nr_waiting--;
692         }
693         conf->nr_pending++;
694         spin_unlock_irq(&conf->resync_lock);
695 }
696
697 static void allow_barrier(conf_t *conf)
698 {
699         unsigned long flags;
700         spin_lock_irqsave(&conf->resync_lock, flags);
701         conf->nr_pending--;
702         spin_unlock_irqrestore(&conf->resync_lock, flags);
703         wake_up(&conf->wait_barrier);
704 }
705
706 static void freeze_array(conf_t *conf)
707 {
708         /* stop syncio and normal IO and wait for everything to
709          * go quite.
710          * We increment barrier and nr_waiting, and then
711          * wait until barrier+nr_pending match nr_queued+2
712          */
713         spin_lock_irq(&conf->resync_lock);
714         conf->barrier++;
715         conf->nr_waiting++;
716         wait_event_lock_irq(conf->wait_barrier,
717                             conf->barrier+conf->nr_pending == conf->nr_queued+2,
718                             conf->resync_lock,
719                             raid1_unplug(conf->mddev->queue));
720         spin_unlock_irq(&conf->resync_lock);
721 }
722 static void unfreeze_array(conf_t *conf)
723 {
724         /* reverse the effect of the freeze */
725         spin_lock_irq(&conf->resync_lock);
726         conf->barrier--;
727         conf->nr_waiting--;
728         wake_up(&conf->wait_barrier);
729         spin_unlock_irq(&conf->resync_lock);
730 }
731
732
733 /* duplicate the data pages for behind I/O */
734 static struct page **alloc_behind_pages(struct bio *bio)
735 {
736         int i;
737         struct bio_vec *bvec;
738         struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page *),
739                                         GFP_NOIO);
740         if (unlikely(!pages))
741                 goto do_sync_io;
742
743         bio_for_each_segment(bvec, bio, i) {
744                 pages[i] = alloc_page(GFP_NOIO);
745                 if (unlikely(!pages[i]))
746                         goto do_sync_io;
747                 memcpy(kmap(pages[i]) + bvec->bv_offset,
748                         kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
749                 kunmap(pages[i]);
750                 kunmap(bvec->bv_page);
751         }
752
753         return pages;
754
755 do_sync_io:
756         if (pages)
757                 for (i = 0; i < bio->bi_vcnt && pages[i]; i++)
758                         put_page(pages[i]);
759         kfree(pages);
760         PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
761         return NULL;
762 }
763
764 static int make_request(request_queue_t *q, struct bio * bio)
765 {
766         mddev_t *mddev = q->queuedata;
767         conf_t *conf = mddev_to_conf(mddev);
768         mirror_info_t *mirror;
769         r1bio_t *r1_bio;
770         struct bio *read_bio;
771         int i, targets = 0, disks;
772         mdk_rdev_t *rdev;
773         struct bitmap *bitmap = mddev->bitmap;
774         unsigned long flags;
775         struct bio_list bl;
776         struct page **behind_pages = NULL;
777         const int rw = bio_data_dir(bio);
778         int do_barriers;
779
780         /*
781          * Register the new request and wait if the reconstruction
782          * thread has put up a bar for new requests.
783          * Continue immediately if no resync is active currently.
784          * We test barriers_work *after* md_write_start as md_write_start
785          * may cause the first superblock write, and that will check out
786          * if barriers work.
787          */
788
789         md_write_start(mddev, bio); /* wait on superblock update early */
790
791         if (unlikely(!mddev->barriers_work && bio_barrier(bio))) {
792                 if (rw == WRITE)
793                         md_write_end(mddev);
794                 bio_endio(bio, bio->bi_size, -EOPNOTSUPP);
795                 return 0;
796         }
797
798         wait_barrier(conf);
799
800         disk_stat_inc(mddev->gendisk, ios[rw]);
801         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
802
803         /*
804          * make_request() can abort the operation when READA is being
805          * used and no empty request is available.
806          *
807          */
808         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
809
810         r1_bio->master_bio = bio;
811         r1_bio->sectors = bio->bi_size >> 9;
812         r1_bio->state = 0;
813         r1_bio->mddev = mddev;
814         r1_bio->sector = bio->bi_sector;
815
816         if (rw == READ) {
817                 /*
818                  * read balancing logic:
819                  */
820                 int rdisk = read_balance(conf, r1_bio);
821
822                 if (rdisk < 0) {
823                         /* couldn't find anywhere to read from */
824                         raid_end_bio_io(r1_bio);
825                         return 0;
826                 }
827                 mirror = conf->mirrors + rdisk;
828
829                 r1_bio->read_disk = rdisk;
830
831                 read_bio = bio_clone(bio, GFP_NOIO);
832
833                 r1_bio->bios[rdisk] = read_bio;
834
835                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
836                 read_bio->bi_bdev = mirror->rdev->bdev;
837                 read_bio->bi_end_io = raid1_end_read_request;
838                 read_bio->bi_rw = READ;
839                 read_bio->bi_private = r1_bio;
840
841                 generic_make_request(read_bio);
842                 return 0;
843         }
844
845         /*
846          * WRITE:
847          */
848         /* first select target devices under spinlock and
849          * inc refcount on their rdev.  Record them by setting
850          * bios[x] to bio
851          */
852         disks = conf->raid_disks;
853 #if 0
854         { static int first=1;
855         if (first) printk("First Write sector %llu disks %d\n",
856                           (unsigned long long)r1_bio->sector, disks);
857         first = 0;
858         }
859 #endif
860         rcu_read_lock();
861         for (i = 0;  i < disks; i++) {
862                 if ((rdev=rcu_dereference(conf->mirrors[i].rdev)) != NULL &&
863                     !test_bit(Faulty, &rdev->flags)) {
864                         atomic_inc(&rdev->nr_pending);
865                         if (test_bit(Faulty, &rdev->flags)) {
866                                 rdev_dec_pending(rdev, mddev);
867                                 r1_bio->bios[i] = NULL;
868                         } else
869                                 r1_bio->bios[i] = bio;
870                         targets++;
871                 } else
872                         r1_bio->bios[i] = NULL;
873         }
874         rcu_read_unlock();
875
876         BUG_ON(targets == 0); /* we never fail the last device */
877
878         if (targets < conf->raid_disks) {
879                 /* array is degraded, we will not clear the bitmap
880                  * on I/O completion (see raid1_end_write_request) */
881                 set_bit(R1BIO_Degraded, &r1_bio->state);
882         }
883
884         /* do behind I/O ? */
885         if (bitmap &&
886             atomic_read(&bitmap->behind_writes) < bitmap->max_write_behind &&
887             (behind_pages = alloc_behind_pages(bio)) != NULL)
888                 set_bit(R1BIO_BehindIO, &r1_bio->state);
889
890         atomic_set(&r1_bio->remaining, 0);
891         atomic_set(&r1_bio->behind_remaining, 0);
892
893         do_barriers = bio_barrier(bio);
894         if (do_barriers)
895                 set_bit(R1BIO_Barrier, &r1_bio->state);
896
897         bio_list_init(&bl);
898         for (i = 0; i < disks; i++) {
899                 struct bio *mbio;
900                 if (!r1_bio->bios[i])
901                         continue;
902
903                 mbio = bio_clone(bio, GFP_NOIO);
904                 r1_bio->bios[i] = mbio;
905
906                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
907                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
908                 mbio->bi_end_io = raid1_end_write_request;
909                 mbio->bi_rw = WRITE | do_barriers;
910                 mbio->bi_private = r1_bio;
911
912                 if (behind_pages) {
913                         struct bio_vec *bvec;
914                         int j;
915
916                         /* Yes, I really want the '__' version so that
917                          * we clear any unused pointer in the io_vec, rather
918                          * than leave them unchanged.  This is important
919                          * because when we come to free the pages, we won't
920                          * know the originial bi_idx, so we just free
921                          * them all
922                          */
923                         __bio_for_each_segment(bvec, mbio, j, 0)
924                                 bvec->bv_page = behind_pages[j];
925                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
926                                 atomic_inc(&r1_bio->behind_remaining);
927                 }
928
929                 atomic_inc(&r1_bio->remaining);
930
931                 bio_list_add(&bl, mbio);
932         }
933         kfree(behind_pages); /* the behind pages are attached to the bios now */
934
935         bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
936                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
937         spin_lock_irqsave(&conf->device_lock, flags);
938         bio_list_merge(&conf->pending_bio_list, &bl);
939         bio_list_init(&bl);
940
941         blk_plug_device(mddev->queue);
942         spin_unlock_irqrestore(&conf->device_lock, flags);
943
944 #if 0
945         while ((bio = bio_list_pop(&bl)) != NULL)
946                 generic_make_request(bio);
947 #endif
948
949         return 0;
950 }
951
952 static void status(struct seq_file *seq, mddev_t *mddev)
953 {
954         conf_t *conf = mddev_to_conf(mddev);
955         int i;
956
957         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
958                    conf->raid_disks - mddev->degraded);
959         rcu_read_lock();
960         for (i = 0; i < conf->raid_disks; i++) {
961                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
962                 seq_printf(seq, "%s",
963                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
964         }
965         rcu_read_unlock();
966         seq_printf(seq, "]");
967 }
968
969
970 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
971 {
972         char b[BDEVNAME_SIZE];
973         conf_t *conf = mddev_to_conf(mddev);
974
975         /*
976          * If it is not operational, then we have already marked it as dead
977          * else if it is the last working disks, ignore the error, let the
978          * next level up know.
979          * else mark the drive as failed
980          */
981         if (test_bit(In_sync, &rdev->flags)
982             && (conf->raid_disks - mddev->degraded) == 1)
983                 /*
984                  * Don't fail the drive, act as though we were just a
985                  * normal single drive
986                  */
987                 return;
988         if (test_and_clear_bit(In_sync, &rdev->flags)) {
989                 unsigned long flags;
990                 spin_lock_irqsave(&conf->device_lock, flags);
991                 mddev->degraded++;
992                 spin_unlock_irqrestore(&conf->device_lock, flags);
993                 /*
994                  * if recovery is running, make sure it aborts.
995                  */
996                 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
997         }
998         set_bit(Faulty, &rdev->flags);
999         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1000         printk(KERN_ALERT "raid1: Disk failure on %s, disabling device. \n"
1001                 "       Operation continuing on %d devices\n",
1002                 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1003 }
1004
1005 static void print_conf(conf_t *conf)
1006 {
1007         int i;
1008
1009         printk("RAID1 conf printout:\n");
1010         if (!conf) {
1011                 printk("(!conf)\n");
1012                 return;
1013         }
1014         printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1015                 conf->raid_disks);
1016
1017         rcu_read_lock();
1018         for (i = 0; i < conf->raid_disks; i++) {
1019                 char b[BDEVNAME_SIZE];
1020                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1021                 if (rdev)
1022                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
1023                                i, !test_bit(In_sync, &rdev->flags),
1024                                !test_bit(Faulty, &rdev->flags),
1025                                bdevname(rdev->bdev,b));
1026         }
1027         rcu_read_unlock();
1028 }
1029
1030 static void close_sync(conf_t *conf)
1031 {
1032         wait_barrier(conf);
1033         allow_barrier(conf);
1034
1035         mempool_destroy(conf->r1buf_pool);
1036         conf->r1buf_pool = NULL;
1037 }
1038
1039 static int raid1_spare_active(mddev_t *mddev)
1040 {
1041         int i;
1042         conf_t *conf = mddev->private;
1043
1044         /*
1045          * Find all failed disks within the RAID1 configuration 
1046          * and mark them readable.
1047          * Called under mddev lock, so rcu protection not needed.
1048          */
1049         for (i = 0; i < conf->raid_disks; i++) {
1050                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1051                 if (rdev
1052                     && !test_bit(Faulty, &rdev->flags)
1053                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1054                         unsigned long flags;
1055                         spin_lock_irqsave(&conf->device_lock, flags);
1056                         mddev->degraded--;
1057                         spin_unlock_irqrestore(&conf->device_lock, flags);
1058                 }
1059         }
1060
1061         print_conf(conf);
1062         return 0;
1063 }
1064
1065
1066 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1067 {
1068         conf_t *conf = mddev->private;
1069         int found = 0;
1070         int mirror = 0;
1071         mirror_info_t *p;
1072
1073         for (mirror=0; mirror < mddev->raid_disks; mirror++)
1074                 if ( !(p=conf->mirrors+mirror)->rdev) {
1075
1076                         blk_queue_stack_limits(mddev->queue,
1077                                                rdev->bdev->bd_disk->queue);
1078                         /* as we don't honour merge_bvec_fn, we must never risk
1079                          * violating it, so limit ->max_sector to one PAGE, as
1080                          * a one page request is never in violation.
1081                          */
1082                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1083                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
1084                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1085
1086                         p->head_position = 0;
1087                         rdev->raid_disk = mirror;
1088                         found = 1;
1089                         /* As all devices are equivalent, we don't need a full recovery
1090                          * if this was recently any drive of the array
1091                          */
1092                         if (rdev->saved_raid_disk < 0)
1093                                 conf->fullsync = 1;
1094                         rcu_assign_pointer(p->rdev, rdev);
1095                         break;
1096                 }
1097
1098         print_conf(conf);
1099         return found;
1100 }
1101
1102 static int raid1_remove_disk(mddev_t *mddev, int number)
1103 {
1104         conf_t *conf = mddev->private;
1105         int err = 0;
1106         mdk_rdev_t *rdev;
1107         mirror_info_t *p = conf->mirrors+ number;
1108
1109         print_conf(conf);
1110         rdev = p->rdev;
1111         if (rdev) {
1112                 if (test_bit(In_sync, &rdev->flags) ||
1113                     atomic_read(&rdev->nr_pending)) {
1114                         err = -EBUSY;
1115                         goto abort;
1116                 }
1117                 p->rdev = NULL;
1118                 synchronize_rcu();
1119                 if (atomic_read(&rdev->nr_pending)) {
1120                         /* lost the race, try later */
1121                         err = -EBUSY;
1122                         p->rdev = rdev;
1123                 }
1124         }
1125 abort:
1126
1127         print_conf(conf);
1128         return err;
1129 }
1130
1131
1132 static int end_sync_read(struct bio *bio, unsigned int bytes_done, int error)
1133 {
1134         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1135         int i;
1136
1137         if (bio->bi_size)
1138                 return 1;
1139
1140         for (i=r1_bio->mddev->raid_disks; i--; )
1141                 if (r1_bio->bios[i] == bio)
1142                         break;
1143         BUG_ON(i < 0);
1144         update_head_pos(i, r1_bio);
1145         /*
1146          * we have read a block, now it needs to be re-written,
1147          * or re-read if the read failed.
1148          * We don't do much here, just schedule handling by raid1d
1149          */
1150         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1151                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1152
1153         if (atomic_dec_and_test(&r1_bio->remaining))
1154                 reschedule_retry(r1_bio);
1155         return 0;
1156 }
1157
1158 static int end_sync_write(struct bio *bio, unsigned int bytes_done, int error)
1159 {
1160         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1161         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1162         mddev_t *mddev = r1_bio->mddev;
1163         conf_t *conf = mddev_to_conf(mddev);
1164         int i;
1165         int mirror=0;
1166
1167         if (bio->bi_size)
1168                 return 1;
1169
1170         for (i = 0; i < conf->raid_disks; i++)
1171                 if (r1_bio->bios[i] == bio) {
1172                         mirror = i;
1173                         break;
1174                 }
1175         if (!uptodate) {
1176                 int sync_blocks = 0;
1177                 sector_t s = r1_bio->sector;
1178                 long sectors_to_go = r1_bio->sectors;
1179                 /* make sure these bits doesn't get cleared. */
1180                 do {
1181                         bitmap_end_sync(mddev->bitmap, s,
1182                                         &sync_blocks, 1);
1183                         s += sync_blocks;
1184                         sectors_to_go -= sync_blocks;
1185                 } while (sectors_to_go > 0);
1186                 md_error(mddev, conf->mirrors[mirror].rdev);
1187         }
1188
1189         update_head_pos(mirror, r1_bio);
1190
1191         if (atomic_dec_and_test(&r1_bio->remaining)) {
1192                 md_done_sync(mddev, r1_bio->sectors, uptodate);
1193                 put_buf(r1_bio);
1194         }
1195         return 0;
1196 }
1197
1198 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1199 {
1200         conf_t *conf = mddev_to_conf(mddev);
1201         int i;
1202         int disks = conf->raid_disks;
1203         struct bio *bio, *wbio;
1204
1205         bio = r1_bio->bios[r1_bio->read_disk];
1206
1207
1208         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1209                 /* We have read all readable devices.  If we haven't
1210                  * got the block, then there is no hope left.
1211                  * If we have, then we want to do a comparison
1212                  * and skip the write if everything is the same.
1213                  * If any blocks failed to read, then we need to
1214                  * attempt an over-write
1215                  */
1216                 int primary;
1217                 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1218                         for (i=0; i<mddev->raid_disks; i++)
1219                                 if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1220                                         md_error(mddev, conf->mirrors[i].rdev);
1221
1222                         md_done_sync(mddev, r1_bio->sectors, 1);
1223                         put_buf(r1_bio);
1224                         return;
1225                 }
1226                 for (primary=0; primary<mddev->raid_disks; primary++)
1227                         if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1228                             test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1229                                 r1_bio->bios[primary]->bi_end_io = NULL;
1230                                 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1231                                 break;
1232                         }
1233                 r1_bio->read_disk = primary;
1234                 for (i=0; i<mddev->raid_disks; i++)
1235                         if (r1_bio->bios[i]->bi_end_io == end_sync_read &&
1236                             test_bit(BIO_UPTODATE, &r1_bio->bios[i]->bi_flags)) {
1237                                 int j;
1238                                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1239                                 struct bio *pbio = r1_bio->bios[primary];
1240                                 struct bio *sbio = r1_bio->bios[i];
1241                                 for (j = vcnt; j-- ; )
1242                                         if (memcmp(page_address(pbio->bi_io_vec[j].bv_page),
1243                                                    page_address(sbio->bi_io_vec[j].bv_page),
1244                                                    PAGE_SIZE))
1245                                                 break;
1246                                 if (j >= 0)
1247                                         mddev->resync_mismatches += r1_bio->sectors;
1248                                 if (j < 0 || test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
1249                                         sbio->bi_end_io = NULL;
1250                                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1251                                 } else {
1252                                         /* fixup the bio for reuse */
1253                                         sbio->bi_vcnt = vcnt;
1254                                         sbio->bi_size = r1_bio->sectors << 9;
1255                                         sbio->bi_idx = 0;
1256                                         sbio->bi_phys_segments = 0;
1257                                         sbio->bi_hw_segments = 0;
1258                                         sbio->bi_hw_front_size = 0;
1259                                         sbio->bi_hw_back_size = 0;
1260                                         sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1261                                         sbio->bi_flags |= 1 << BIO_UPTODATE;
1262                                         sbio->bi_next = NULL;
1263                                         sbio->bi_sector = r1_bio->sector +
1264                                                 conf->mirrors[i].rdev->data_offset;
1265                                         sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1266                                 }
1267                         }
1268         }
1269         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1270                 /* ouch - failed to read all of that.
1271                  * Try some synchronous reads of other devices to get
1272                  * good data, much like with normal read errors.  Only
1273                  * read into the pages we already have so we don't
1274                  * need to re-issue the read request.
1275                  * We don't need to freeze the array, because being in an
1276                  * active sync request, there is no normal IO, and
1277                  * no overlapping syncs.
1278                  */
1279                 sector_t sect = r1_bio->sector;
1280                 int sectors = r1_bio->sectors;
1281                 int idx = 0;
1282
1283                 while(sectors) {
1284                         int s = sectors;
1285                         int d = r1_bio->read_disk;
1286                         int success = 0;
1287                         mdk_rdev_t *rdev;
1288
1289                         if (s > (PAGE_SIZE>>9))
1290                                 s = PAGE_SIZE >> 9;
1291                         do {
1292                                 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1293                                         /* No rcu protection needed here devices
1294                                          * can only be removed when no resync is
1295                                          * active, and resync is currently active
1296                                          */
1297                                         rdev = conf->mirrors[d].rdev;
1298                                         if (sync_page_io(rdev->bdev,
1299                                                          sect + rdev->data_offset,
1300                                                          s<<9,
1301                                                          bio->bi_io_vec[idx].bv_page,
1302                                                          READ)) {
1303                                                 success = 1;
1304                                                 break;
1305                                         }
1306                                 }
1307                                 d++;
1308                                 if (d == conf->raid_disks)
1309                                         d = 0;
1310                         } while (!success && d != r1_bio->read_disk);
1311
1312                         if (success) {
1313                                 int start = d;
1314                                 /* write it back and re-read */
1315                                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1316                                 while (d != r1_bio->read_disk) {
1317                                         if (d == 0)
1318                                                 d = conf->raid_disks;
1319                                         d--;
1320                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1321                                                 continue;
1322                                         rdev = conf->mirrors[d].rdev;
1323                                         atomic_add(s, &rdev->corrected_errors);
1324                                         if (sync_page_io(rdev->bdev,
1325                                                          sect + rdev->data_offset,
1326                                                          s<<9,
1327                                                          bio->bi_io_vec[idx].bv_page,
1328                                                          WRITE) == 0)
1329                                                 md_error(mddev, rdev);
1330                                 }
1331                                 d = start;
1332                                 while (d != r1_bio->read_disk) {
1333                                         if (d == 0)
1334                                                 d = conf->raid_disks;
1335                                         d--;
1336                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1337                                                 continue;
1338                                         rdev = conf->mirrors[d].rdev;
1339                                         if (sync_page_io(rdev->bdev,
1340                                                          sect + rdev->data_offset,
1341                                                          s<<9,
1342                                                          bio->bi_io_vec[idx].bv_page,
1343                                                          READ) == 0)
1344                                                 md_error(mddev, rdev);
1345                                 }
1346                         } else {
1347                                 char b[BDEVNAME_SIZE];
1348                                 /* Cannot read from anywhere, array is toast */
1349                                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1350                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
1351                                        " for block %llu\n",
1352                                        bdevname(bio->bi_bdev,b),
1353                                        (unsigned long long)r1_bio->sector);
1354                                 md_done_sync(mddev, r1_bio->sectors, 0);
1355                                 put_buf(r1_bio);
1356                                 return;
1357                         }
1358                         sectors -= s;
1359                         sect += s;
1360                         idx ++;
1361                 }
1362         }
1363
1364         /*
1365          * schedule writes
1366          */
1367         atomic_set(&r1_bio->remaining, 1);
1368         for (i = 0; i < disks ; i++) {
1369                 wbio = r1_bio->bios[i];
1370                 if (wbio->bi_end_io == NULL ||
1371                     (wbio->bi_end_io == end_sync_read &&
1372                      (i == r1_bio->read_disk ||
1373                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1374                         continue;
1375
1376                 wbio->bi_rw = WRITE;
1377                 wbio->bi_end_io = end_sync_write;
1378                 atomic_inc(&r1_bio->remaining);
1379                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1380
1381                 generic_make_request(wbio);
1382         }
1383
1384         if (atomic_dec_and_test(&r1_bio->remaining)) {
1385                 /* if we're here, all write(s) have completed, so clean up */
1386                 md_done_sync(mddev, r1_bio->sectors, 1);
1387                 put_buf(r1_bio);
1388         }
1389 }
1390
1391 /*
1392  * This is a kernel thread which:
1393  *
1394  *      1.      Retries failed read operations on working mirrors.
1395  *      2.      Updates the raid superblock when problems encounter.
1396  *      3.      Performs writes following reads for array syncronising.
1397  */
1398
1399 static void fix_read_error(conf_t *conf, int read_disk,
1400                            sector_t sect, int sectors)
1401 {
1402         mddev_t *mddev = conf->mddev;
1403         while(sectors) {
1404                 int s = sectors;
1405                 int d = read_disk;
1406                 int success = 0;
1407                 int start;
1408                 mdk_rdev_t *rdev;
1409
1410                 if (s > (PAGE_SIZE>>9))
1411                         s = PAGE_SIZE >> 9;
1412
1413                 do {
1414                         /* Note: no rcu protection needed here
1415                          * as this is synchronous in the raid1d thread
1416                          * which is the thread that might remove
1417                          * a device.  If raid1d ever becomes multi-threaded....
1418                          */
1419                         rdev = conf->mirrors[d].rdev;
1420                         if (rdev &&
1421                             test_bit(In_sync, &rdev->flags) &&
1422                             sync_page_io(rdev->bdev,
1423                                          sect + rdev->data_offset,
1424                                          s<<9,
1425                                          conf->tmppage, READ))
1426                                 success = 1;
1427                         else {
1428                                 d++;
1429                                 if (d == conf->raid_disks)
1430                                         d = 0;
1431                         }
1432                 } while (!success && d != read_disk);
1433
1434                 if (!success) {
1435                         /* Cannot read from anywhere -- bye bye array */
1436                         md_error(mddev, conf->mirrors[read_disk].rdev);
1437                         break;
1438                 }
1439                 /* write it back and re-read */
1440                 start = d;
1441                 while (d != read_disk) {
1442                         if (d==0)
1443                                 d = conf->raid_disks;
1444                         d--;
1445                         rdev = conf->mirrors[d].rdev;
1446                         if (rdev &&
1447                             test_bit(In_sync, &rdev->flags)) {
1448                                 if (sync_page_io(rdev->bdev,
1449                                                  sect + rdev->data_offset,
1450                                                  s<<9, conf->tmppage, WRITE)
1451                                     == 0)
1452                                         /* Well, this device is dead */
1453                                         md_error(mddev, rdev);
1454                         }
1455                 }
1456                 d = start;
1457                 while (d != read_disk) {
1458                         char b[BDEVNAME_SIZE];
1459                         if (d==0)
1460                                 d = conf->raid_disks;
1461                         d--;
1462                         rdev = conf->mirrors[d].rdev;
1463                         if (rdev &&
1464                             test_bit(In_sync, &rdev->flags)) {
1465                                 if (sync_page_io(rdev->bdev,
1466                                                  sect + rdev->data_offset,
1467                                                  s<<9, conf->tmppage, READ)
1468                                     == 0)
1469                                         /* Well, this device is dead */
1470                                         md_error(mddev, rdev);
1471                                 else {
1472                                         atomic_add(s, &rdev->corrected_errors);
1473                                         printk(KERN_INFO
1474                                                "raid1:%s: read error corrected "
1475                                                "(%d sectors at %llu on %s)\n",
1476                                                mdname(mddev), s,
1477                                                (unsigned long long)(sect +
1478                                                    rdev->data_offset),
1479                                                bdevname(rdev->bdev, b));
1480                                 }
1481                         }
1482                 }
1483                 sectors -= s;
1484                 sect += s;
1485         }
1486 }
1487
1488 static void raid1d(mddev_t *mddev)
1489 {
1490         r1bio_t *r1_bio;
1491         struct bio *bio;
1492         unsigned long flags;
1493         conf_t *conf = mddev_to_conf(mddev);
1494         struct list_head *head = &conf->retry_list;
1495         int unplug=0;
1496         mdk_rdev_t *rdev;
1497
1498         md_check_recovery(mddev);
1499         
1500         for (;;) {
1501                 char b[BDEVNAME_SIZE];
1502                 spin_lock_irqsave(&conf->device_lock, flags);
1503
1504                 if (conf->pending_bio_list.head) {
1505                         bio = bio_list_get(&conf->pending_bio_list);
1506                         blk_remove_plug(mddev->queue);
1507                         spin_unlock_irqrestore(&conf->device_lock, flags);
1508                         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
1509                         if (bitmap_unplug(mddev->bitmap) != 0)
1510                                 printk("%s: bitmap file write failed!\n", mdname(mddev));
1511
1512                         while (bio) { /* submit pending writes */
1513                                 struct bio *next = bio->bi_next;
1514                                 bio->bi_next = NULL;
1515                                 generic_make_request(bio);
1516                                 bio = next;
1517                         }
1518                         unplug = 1;
1519
1520                         continue;
1521                 }
1522
1523                 if (list_empty(head))
1524                         break;
1525                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1526                 list_del(head->prev);
1527                 conf->nr_queued--;
1528                 spin_unlock_irqrestore(&conf->device_lock, flags);
1529
1530                 mddev = r1_bio->mddev;
1531                 conf = mddev_to_conf(mddev);
1532                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1533                         sync_request_write(mddev, r1_bio);
1534                         unplug = 1;
1535                 } else if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
1536                         /* some requests in the r1bio were BIO_RW_BARRIER
1537                          * requests which failed with -EOPNOTSUPP.  Hohumm..
1538                          * Better resubmit without the barrier.
1539                          * We know which devices to resubmit for, because
1540                          * all others have had their bios[] entry cleared.
1541                          * We already have a nr_pending reference on these rdevs.
1542                          */
1543                         int i;
1544                         clear_bit(R1BIO_BarrierRetry, &r1_bio->state);
1545                         clear_bit(R1BIO_Barrier, &r1_bio->state);
1546                         for (i=0; i < conf->raid_disks; i++)
1547                                 if (r1_bio->bios[i])
1548                                         atomic_inc(&r1_bio->remaining);
1549                         for (i=0; i < conf->raid_disks; i++)
1550                                 if (r1_bio->bios[i]) {
1551                                         struct bio_vec *bvec;
1552                                         int j;
1553
1554                                         bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1555                                         /* copy pages from the failed bio, as
1556                                          * this might be a write-behind device */
1557                                         __bio_for_each_segment(bvec, bio, j, 0)
1558                                                 bvec->bv_page = bio_iovec_idx(r1_bio->bios[i], j)->bv_page;
1559                                         bio_put(r1_bio->bios[i]);
1560                                         bio->bi_sector = r1_bio->sector +
1561                                                 conf->mirrors[i].rdev->data_offset;
1562                                         bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1563                                         bio->bi_end_io = raid1_end_write_request;
1564                                         bio->bi_rw = WRITE;
1565                                         bio->bi_private = r1_bio;
1566                                         r1_bio->bios[i] = bio;
1567                                         generic_make_request(bio);
1568                                 }
1569                 } else {
1570                         int disk;
1571
1572                         /* we got a read error. Maybe the drive is bad.  Maybe just
1573                          * the block and we can fix it.
1574                          * We freeze all other IO, and try reading the block from
1575                          * other devices.  When we find one, we re-write
1576                          * and check it that fixes the read error.
1577                          * This is all done synchronously while the array is
1578                          * frozen
1579                          */
1580                         if (mddev->ro == 0) {
1581                                 freeze_array(conf);
1582                                 fix_read_error(conf, r1_bio->read_disk,
1583                                                r1_bio->sector,
1584                                                r1_bio->sectors);
1585                                 unfreeze_array(conf);
1586                         }
1587
1588                         bio = r1_bio->bios[r1_bio->read_disk];
1589                         if ((disk=read_balance(conf, r1_bio)) == -1) {
1590                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
1591                                        " read error for block %llu\n",
1592                                        bdevname(bio->bi_bdev,b),
1593                                        (unsigned long long)r1_bio->sector);
1594                                 raid_end_bio_io(r1_bio);
1595                         } else {
1596                                 r1_bio->bios[r1_bio->read_disk] =
1597                                         mddev->ro ? IO_BLOCKED : NULL;
1598                                 r1_bio->read_disk = disk;
1599                                 bio_put(bio);
1600                                 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1601                                 r1_bio->bios[r1_bio->read_disk] = bio;
1602                                 rdev = conf->mirrors[disk].rdev;
1603                                 if (printk_ratelimit())
1604                                         printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
1605                                                " another mirror\n",
1606                                                bdevname(rdev->bdev,b),
1607                                                (unsigned long long)r1_bio->sector);
1608                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1609                                 bio->bi_bdev = rdev->bdev;
1610                                 bio->bi_end_io = raid1_end_read_request;
1611                                 bio->bi_rw = READ;
1612                                 bio->bi_private = r1_bio;
1613                                 unplug = 1;
1614                                 generic_make_request(bio);
1615                         }
1616                 }
1617         }
1618         spin_unlock_irqrestore(&conf->device_lock, flags);
1619         if (unplug)
1620                 unplug_slaves(mddev);
1621 }
1622
1623
1624 static int init_resync(conf_t *conf)
1625 {
1626         int buffs;
1627
1628         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1629         BUG_ON(conf->r1buf_pool);
1630         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1631                                           conf->poolinfo);
1632         if (!conf->r1buf_pool)
1633                 return -ENOMEM;
1634         conf->next_resync = 0;
1635         return 0;
1636 }
1637
1638 /*
1639  * perform a "sync" on one "block"
1640  *
1641  * We need to make sure that no normal I/O request - particularly write
1642  * requests - conflict with active sync requests.
1643  *
1644  * This is achieved by tracking pending requests and a 'barrier' concept
1645  * that can be installed to exclude normal IO requests.
1646  */
1647
1648 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1649 {
1650         conf_t *conf = mddev_to_conf(mddev);
1651         r1bio_t *r1_bio;
1652         struct bio *bio;
1653         sector_t max_sector, nr_sectors;
1654         int disk = -1;
1655         int i;
1656         int wonly = -1;
1657         int write_targets = 0, read_targets = 0;
1658         int sync_blocks;
1659         int still_degraded = 0;
1660
1661         if (!conf->r1buf_pool)
1662         {
1663 /*
1664                 printk("sync start - bitmap %p\n", mddev->bitmap);
1665 */
1666                 if (init_resync(conf))
1667                         return 0;
1668         }
1669
1670         max_sector = mddev->size << 1;
1671         if (sector_nr >= max_sector) {
1672                 /* If we aborted, we need to abort the
1673                  * sync on the 'current' bitmap chunk (there will
1674                  * only be one in raid1 resync.
1675                  * We can find the current addess in mddev->curr_resync
1676                  */
1677                 if (mddev->curr_resync < max_sector) /* aborted */
1678                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1679                                                 &sync_blocks, 1);
1680                 else /* completed sync */
1681                         conf->fullsync = 0;
1682
1683                 bitmap_close_sync(mddev->bitmap);
1684                 close_sync(conf);
1685                 return 0;
1686         }
1687
1688         if (mddev->bitmap == NULL &&
1689             mddev->recovery_cp == MaxSector &&
1690             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1691             conf->fullsync == 0) {
1692                 *skipped = 1;
1693                 return max_sector - sector_nr;
1694         }
1695         /* before building a request, check if we can skip these blocks..
1696          * This call the bitmap_start_sync doesn't actually record anything
1697          */
1698         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1699             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1700                 /* We can skip this block, and probably several more */
1701                 *skipped = 1;
1702                 return sync_blocks;
1703         }
1704         /*
1705          * If there is non-resync activity waiting for a turn,
1706          * and resync is going fast enough,
1707          * then let it though before starting on this new sync request.
1708          */
1709         if (!go_faster && conf->nr_waiting)
1710                 msleep_interruptible(1000);
1711
1712         raise_barrier(conf);
1713
1714         conf->next_resync = sector_nr;
1715
1716         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1717         rcu_read_lock();
1718         /*
1719          * If we get a correctably read error during resync or recovery,
1720          * we might want to read from a different device.  So we
1721          * flag all drives that could conceivably be read from for READ,
1722          * and any others (which will be non-In_sync devices) for WRITE.
1723          * If a read fails, we try reading from something else for which READ
1724          * is OK.
1725          */
1726
1727         r1_bio->mddev = mddev;
1728         r1_bio->sector = sector_nr;
1729         r1_bio->state = 0;
1730         set_bit(R1BIO_IsSync, &r1_bio->state);
1731
1732         for (i=0; i < conf->raid_disks; i++) {
1733                 mdk_rdev_t *rdev;
1734                 bio = r1_bio->bios[i];
1735
1736                 /* take from bio_init */
1737                 bio->bi_next = NULL;
1738                 bio->bi_flags |= 1 << BIO_UPTODATE;
1739                 bio->bi_rw = READ;
1740                 bio->bi_vcnt = 0;
1741                 bio->bi_idx = 0;
1742                 bio->bi_phys_segments = 0;
1743                 bio->bi_hw_segments = 0;
1744                 bio->bi_size = 0;
1745                 bio->bi_end_io = NULL;
1746                 bio->bi_private = NULL;
1747
1748                 rdev = rcu_dereference(conf->mirrors[i].rdev);
1749                 if (rdev == NULL ||
1750                            test_bit(Faulty, &rdev->flags)) {
1751                         still_degraded = 1;
1752                         continue;
1753                 } else if (!test_bit(In_sync, &rdev->flags)) {
1754                         bio->bi_rw = WRITE;
1755                         bio->bi_end_io = end_sync_write;
1756                         write_targets ++;
1757                 } else {
1758                         /* may need to read from here */
1759                         bio->bi_rw = READ;
1760                         bio->bi_end_io = end_sync_read;
1761                         if (test_bit(WriteMostly, &rdev->flags)) {
1762                                 if (wonly < 0)
1763                                         wonly = i;
1764                         } else {
1765                                 if (disk < 0)
1766                                         disk = i;
1767                         }
1768                         read_targets++;
1769                 }
1770                 atomic_inc(&rdev->nr_pending);
1771                 bio->bi_sector = sector_nr + rdev->data_offset;
1772                 bio->bi_bdev = rdev->bdev;
1773                 bio->bi_private = r1_bio;
1774         }
1775         rcu_read_unlock();
1776         if (disk < 0)
1777                 disk = wonly;
1778         r1_bio->read_disk = disk;
1779
1780         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1781                 /* extra read targets are also write targets */
1782                 write_targets += read_targets-1;
1783
1784         if (write_targets == 0 || read_targets == 0) {
1785                 /* There is nowhere to write, so all non-sync
1786                  * drives must be failed - so we are finished
1787                  */
1788                 sector_t rv = max_sector - sector_nr;
1789                 *skipped = 1;
1790                 put_buf(r1_bio);
1791                 return rv;
1792         }
1793
1794         nr_sectors = 0;
1795         sync_blocks = 0;
1796         do {
1797                 struct page *page;
1798                 int len = PAGE_SIZE;
1799                 if (sector_nr + (len>>9) > max_sector)
1800                         len = (max_sector - sector_nr) << 9;
1801                 if (len == 0)
1802                         break;
1803                 if (sync_blocks == 0) {
1804                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1805                                                &sync_blocks, still_degraded) &&
1806                             !conf->fullsync &&
1807                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1808                                 break;
1809                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1810                         if (len > (sync_blocks<<9))
1811                                 len = sync_blocks<<9;
1812                 }
1813
1814                 for (i=0 ; i < conf->raid_disks; i++) {
1815                         bio = r1_bio->bios[i];
1816                         if (bio->bi_end_io) {
1817                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1818                                 if (bio_add_page(bio, page, len, 0) == 0) {
1819                                         /* stop here */
1820                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1821                                         while (i > 0) {
1822                                                 i--;
1823                                                 bio = r1_bio->bios[i];
1824                                                 if (bio->bi_end_io==NULL)
1825                                                         continue;
1826                                                 /* remove last page from this bio */
1827                                                 bio->bi_vcnt--;
1828                                                 bio->bi_size -= len;
1829                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1830                                         }
1831                                         goto bio_full;
1832                                 }
1833                         }
1834                 }
1835                 nr_sectors += len>>9;
1836                 sector_nr += len>>9;
1837                 sync_blocks -= (len>>9);
1838         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1839  bio_full:
1840         r1_bio->sectors = nr_sectors;
1841
1842         /* For a user-requested sync, we read all readable devices and do a
1843          * compare
1844          */
1845         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1846                 atomic_set(&r1_bio->remaining, read_targets);
1847                 for (i=0; i<conf->raid_disks; i++) {
1848                         bio = r1_bio->bios[i];
1849                         if (bio->bi_end_io == end_sync_read) {
1850                                 md_sync_acct(bio->bi_bdev, nr_sectors);
1851                                 generic_make_request(bio);
1852                         }
1853                 }
1854         } else {
1855                 atomic_set(&r1_bio->remaining, 1);
1856                 bio = r1_bio->bios[r1_bio->read_disk];
1857                 md_sync_acct(bio->bi_bdev, nr_sectors);
1858                 generic_make_request(bio);
1859
1860         }
1861         return nr_sectors;
1862 }
1863
1864 static int run(mddev_t *mddev)
1865 {
1866         conf_t *conf;
1867         int i, j, disk_idx;
1868         mirror_info_t *disk;
1869         mdk_rdev_t *rdev;
1870         struct list_head *tmp;
1871
1872         if (mddev->level != 1) {
1873                 printk("raid1: %s: raid level not set to mirroring (%d)\n",
1874                        mdname(mddev), mddev->level);
1875                 goto out;
1876         }
1877         if (mddev->reshape_position != MaxSector) {
1878                 printk("raid1: %s: reshape_position set but not supported\n",
1879                        mdname(mddev));
1880                 goto out;
1881         }
1882         /*
1883          * copy the already verified devices into our private RAID1
1884          * bookkeeping area. [whatever we allocate in run(),
1885          * should be freed in stop()]
1886          */
1887         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1888         mddev->private = conf;
1889         if (!conf)
1890                 goto out_no_mem;
1891
1892         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1893                                  GFP_KERNEL);
1894         if (!conf->mirrors)
1895                 goto out_no_mem;
1896
1897         conf->tmppage = alloc_page(GFP_KERNEL);
1898         if (!conf->tmppage)
1899                 goto out_no_mem;
1900
1901         conf->poolinfo = kmalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1902         if (!conf->poolinfo)
1903                 goto out_no_mem;
1904         conf->poolinfo->mddev = mddev;
1905         conf->poolinfo->raid_disks = mddev->raid_disks;
1906         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1907                                           r1bio_pool_free,
1908                                           conf->poolinfo);
1909         if (!conf->r1bio_pool)
1910                 goto out_no_mem;
1911
1912         ITERATE_RDEV(mddev, rdev, tmp) {
1913                 disk_idx = rdev->raid_disk;
1914                 if (disk_idx >= mddev->raid_disks
1915                     || disk_idx < 0)
1916                         continue;
1917                 disk = conf->mirrors + disk_idx;
1918
1919                 disk->rdev = rdev;
1920
1921                 blk_queue_stack_limits(mddev->queue,
1922                                        rdev->bdev->bd_disk->queue);
1923                 /* as we don't honour merge_bvec_fn, we must never risk
1924                  * violating it, so limit ->max_sector to one PAGE, as
1925                  * a one page request is never in violation.
1926                  */
1927                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1928                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
1929                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1930
1931                 disk->head_position = 0;
1932         }
1933         conf->raid_disks = mddev->raid_disks;
1934         conf->mddev = mddev;
1935         spin_lock_init(&conf->device_lock);
1936         INIT_LIST_HEAD(&conf->retry_list);
1937
1938         spin_lock_init(&conf->resync_lock);
1939         init_waitqueue_head(&conf->wait_barrier);
1940
1941         bio_list_init(&conf->pending_bio_list);
1942         bio_list_init(&conf->flushing_bio_list);
1943
1944
1945         mddev->degraded = 0;
1946         for (i = 0; i < conf->raid_disks; i++) {
1947
1948                 disk = conf->mirrors + i;
1949
1950                 if (!disk->rdev ||
1951                     !test_bit(In_sync, &disk->rdev->flags)) {
1952                         disk->head_position = 0;
1953                         mddev->degraded++;
1954                         conf->fullsync = 1;
1955                 }
1956         }
1957         if (mddev->degraded == conf->raid_disks) {
1958                 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
1959                         mdname(mddev));
1960                 goto out_free_conf;
1961         }
1962         if (conf->raid_disks - mddev->degraded == 1)
1963                 mddev->recovery_cp = MaxSector;
1964
1965         /*
1966          * find the first working one and use it as a starting point
1967          * to read balancing.
1968          */
1969         for (j = 0; j < conf->raid_disks &&
1970                      (!conf->mirrors[j].rdev ||
1971                       !test_bit(In_sync, &conf->mirrors[j].rdev->flags)) ; j++)
1972                 /* nothing */;
1973         conf->last_used = j;
1974
1975
1976         mddev->thread = md_register_thread(raid1d, mddev, "%s_raid1");
1977         if (!mddev->thread) {
1978                 printk(KERN_ERR
1979                        "raid1: couldn't allocate thread for %s\n",
1980                        mdname(mddev));
1981                 goto out_free_conf;
1982         }
1983
1984         printk(KERN_INFO 
1985                 "raid1: raid set %s active with %d out of %d mirrors\n",
1986                 mdname(mddev), mddev->raid_disks - mddev->degraded, 
1987                 mddev->raid_disks);
1988         /*
1989          * Ok, everything is just fine now
1990          */
1991         mddev->array_size = mddev->size;
1992
1993         mddev->queue->unplug_fn = raid1_unplug;
1994         mddev->queue->issue_flush_fn = raid1_issue_flush;
1995         mddev->queue->backing_dev_info.congested_fn = raid1_congested;
1996         mddev->queue->backing_dev_info.congested_data = mddev;
1997
1998         return 0;
1999
2000 out_no_mem:
2001         printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
2002                mdname(mddev));
2003
2004 out_free_conf:
2005         if (conf) {
2006                 if (conf->r1bio_pool)
2007                         mempool_destroy(conf->r1bio_pool);
2008                 kfree(conf->mirrors);
2009                 safe_put_page(conf->tmppage);
2010                 kfree(conf->poolinfo);
2011                 kfree(conf);
2012                 mddev->private = NULL;
2013         }
2014 out:
2015         return -EIO;
2016 }
2017
2018 static int stop(mddev_t *mddev)
2019 {
2020         conf_t *conf = mddev_to_conf(mddev);
2021         struct bitmap *bitmap = mddev->bitmap;
2022         int behind_wait = 0;
2023
2024         /* wait for behind writes to complete */
2025         while (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2026                 behind_wait++;
2027                 printk(KERN_INFO "raid1: behind writes in progress on device %s, waiting to stop (%d)\n", mdname(mddev), behind_wait);
2028                 set_current_state(TASK_UNINTERRUPTIBLE);
2029                 schedule_timeout(HZ); /* wait a second */
2030                 /* need to kick something here to make sure I/O goes? */
2031         }
2032
2033         md_unregister_thread(mddev->thread);
2034         mddev->thread = NULL;
2035         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2036         if (conf->r1bio_pool)
2037                 mempool_destroy(conf->r1bio_pool);
2038         kfree(conf->mirrors);
2039         kfree(conf->poolinfo);
2040         kfree(conf);
2041         mddev->private = NULL;
2042         return 0;
2043 }
2044
2045 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2046 {
2047         /* no resync is happening, and there is enough space
2048          * on all devices, so we can resize.
2049          * We need to make sure resync covers any new space.
2050          * If the array is shrinking we should possibly wait until
2051          * any io in the removed space completes, but it hardly seems
2052          * worth it.
2053          */
2054         mddev->array_size = sectors>>1;
2055         set_capacity(mddev->gendisk, mddev->array_size << 1);
2056         mddev->changed = 1;
2057         if (mddev->array_size > mddev->size && mddev->recovery_cp == MaxSector) {
2058                 mddev->recovery_cp = mddev->size << 1;
2059                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2060         }
2061         mddev->size = mddev->array_size;
2062         mddev->resync_max_sectors = sectors;
2063         return 0;
2064 }
2065
2066 static int raid1_reshape(mddev_t *mddev)
2067 {
2068         /* We need to:
2069          * 1/ resize the r1bio_pool
2070          * 2/ resize conf->mirrors
2071          *
2072          * We allocate a new r1bio_pool if we can.
2073          * Then raise a device barrier and wait until all IO stops.
2074          * Then resize conf->mirrors and swap in the new r1bio pool.
2075          *
2076          * At the same time, we "pack" the devices so that all the missing
2077          * devices have the higher raid_disk numbers.
2078          */
2079         mempool_t *newpool, *oldpool;
2080         struct pool_info *newpoolinfo;
2081         mirror_info_t *newmirrors;
2082         conf_t *conf = mddev_to_conf(mddev);
2083         int cnt, raid_disks;
2084         unsigned long flags;
2085         int d, d2;
2086
2087         /* Cannot change chunk_size, layout, or level */
2088         if (mddev->chunk_size != mddev->new_chunk ||
2089             mddev->layout != mddev->new_layout ||
2090             mddev->level != mddev->new_level) {
2091                 mddev->new_chunk = mddev->chunk_size;
2092                 mddev->new_layout = mddev->layout;
2093                 mddev->new_level = mddev->level;
2094                 return -EINVAL;
2095         }
2096
2097         raid_disks = mddev->raid_disks + mddev->delta_disks;
2098
2099         if (raid_disks < conf->raid_disks) {
2100                 cnt=0;
2101                 for (d= 0; d < conf->raid_disks; d++)
2102                         if (conf->mirrors[d].rdev)
2103                                 cnt++;
2104                 if (cnt > raid_disks)
2105                         return -EBUSY;
2106         }
2107
2108         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2109         if (!newpoolinfo)
2110                 return -ENOMEM;
2111         newpoolinfo->mddev = mddev;
2112         newpoolinfo->raid_disks = raid_disks;
2113
2114         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2115                                  r1bio_pool_free, newpoolinfo);
2116         if (!newpool) {
2117                 kfree(newpoolinfo);
2118                 return -ENOMEM;
2119         }
2120         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2121         if (!newmirrors) {
2122                 kfree(newpoolinfo);
2123                 mempool_destroy(newpool);
2124                 return -ENOMEM;
2125         }
2126
2127         raise_barrier(conf);
2128
2129         /* ok, everything is stopped */
2130         oldpool = conf->r1bio_pool;
2131         conf->r1bio_pool = newpool;
2132
2133         for (d=d2=0; d < conf->raid_disks; d++)
2134                 if (conf->mirrors[d].rdev) {
2135                         conf->mirrors[d].rdev->raid_disk = d2;
2136                         newmirrors[d2++].rdev = conf->mirrors[d].rdev;
2137                 }
2138         kfree(conf->mirrors);
2139         conf->mirrors = newmirrors;
2140         kfree(conf->poolinfo);
2141         conf->poolinfo = newpoolinfo;
2142
2143         spin_lock_irqsave(&conf->device_lock, flags);
2144         mddev->degraded += (raid_disks - conf->raid_disks);
2145         spin_unlock_irqrestore(&conf->device_lock, flags);
2146         conf->raid_disks = mddev->raid_disks = raid_disks;
2147         mddev->delta_disks = 0;
2148
2149         conf->last_used = 0; /* just make sure it is in-range */
2150         lower_barrier(conf);
2151
2152         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2153         md_wakeup_thread(mddev->thread);
2154
2155         mempool_destroy(oldpool);
2156         return 0;
2157 }
2158
2159 static void raid1_quiesce(mddev_t *mddev, int state)
2160 {
2161         conf_t *conf = mddev_to_conf(mddev);
2162
2163         switch(state) {
2164         case 1:
2165                 raise_barrier(conf);
2166                 break;
2167         case 0:
2168                 lower_barrier(conf);
2169                 break;
2170         }
2171 }
2172
2173
2174 static struct mdk_personality raid1_personality =
2175 {
2176         .name           = "raid1",
2177         .level          = 1,
2178         .owner          = THIS_MODULE,
2179         .make_request   = make_request,
2180         .run            = run,
2181         .stop           = stop,
2182         .status         = status,
2183         .error_handler  = error,
2184         .hot_add_disk   = raid1_add_disk,
2185         .hot_remove_disk= raid1_remove_disk,
2186         .spare_active   = raid1_spare_active,
2187         .sync_request   = sync_request,
2188         .resize         = raid1_resize,
2189         .check_reshape  = raid1_reshape,
2190         .quiesce        = raid1_quiesce,
2191 };
2192
2193 static int __init raid_init(void)
2194 {
2195         return register_md_personality(&raid1_personality);
2196 }
2197
2198 static void raid_exit(void)
2199 {
2200         unregister_md_personality(&raid1_personality);
2201 }
2202
2203 module_init(raid_init);
2204 module_exit(raid_exit);
2205 MODULE_LICENSE("GPL");
2206 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2207 MODULE_ALIAS("md-raid1");
2208 MODULE_ALIAS("md-level-1");