ab74d40cac98c4c35d7785d1c69eb8ca8c96de30
[linux-2.6.git] / drivers / md / raid1.c
1 /*
2  * raid1.c : Multiple Devices driver for Linux
3  *
4  * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
5  *
6  * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
7  *
8  * RAID-1 management functions.
9  *
10  * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
11  *
12  * Fixes to reconstruction by Jakob Ć˜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         const int do_sync = bio_sync(bio);
779         int do_barriers;
780
781         /*
782          * Register the new request and wait if the reconstruction
783          * thread has put up a bar for new requests.
784          * Continue immediately if no resync is active currently.
785          * We test barriers_work *after* md_write_start as md_write_start
786          * may cause the first superblock write, and that will check out
787          * if barriers work.
788          */
789
790         md_write_start(mddev, bio); /* wait on superblock update early */
791
792         if (unlikely(!mddev->barriers_work && bio_barrier(bio))) {
793                 if (rw == WRITE)
794                         md_write_end(mddev);
795                 bio_endio(bio, bio->bi_size, -EOPNOTSUPP);
796                 return 0;
797         }
798
799         wait_barrier(conf);
800
801         disk_stat_inc(mddev->gendisk, ios[rw]);
802         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
803
804         /*
805          * make_request() can abort the operation when READA is being
806          * used and no empty request is available.
807          *
808          */
809         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
810
811         r1_bio->master_bio = bio;
812         r1_bio->sectors = bio->bi_size >> 9;
813         r1_bio->state = 0;
814         r1_bio->mddev = mddev;
815         r1_bio->sector = bio->bi_sector;
816
817         if (rw == READ) {
818                 /*
819                  * read balancing logic:
820                  */
821                 int rdisk = read_balance(conf, r1_bio);
822
823                 if (rdisk < 0) {
824                         /* couldn't find anywhere to read from */
825                         raid_end_bio_io(r1_bio);
826                         return 0;
827                 }
828                 mirror = conf->mirrors + rdisk;
829
830                 r1_bio->read_disk = rdisk;
831
832                 read_bio = bio_clone(bio, GFP_NOIO);
833
834                 r1_bio->bios[rdisk] = read_bio;
835
836                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
837                 read_bio->bi_bdev = mirror->rdev->bdev;
838                 read_bio->bi_end_io = raid1_end_read_request;
839                 read_bio->bi_rw = READ | do_sync;
840                 read_bio->bi_private = r1_bio;
841
842                 generic_make_request(read_bio);
843                 return 0;
844         }
845
846         /*
847          * WRITE:
848          */
849         /* first select target devices under spinlock and
850          * inc refcount on their rdev.  Record them by setting
851          * bios[x] to bio
852          */
853         disks = conf->raid_disks;
854 #if 0
855         { static int first=1;
856         if (first) printk("First Write sector %llu disks %d\n",
857                           (unsigned long long)r1_bio->sector, disks);
858         first = 0;
859         }
860 #endif
861         rcu_read_lock();
862         for (i = 0;  i < disks; i++) {
863                 if ((rdev=rcu_dereference(conf->mirrors[i].rdev)) != NULL &&
864                     !test_bit(Faulty, &rdev->flags)) {
865                         atomic_inc(&rdev->nr_pending);
866                         if (test_bit(Faulty, &rdev->flags)) {
867                                 rdev_dec_pending(rdev, mddev);
868                                 r1_bio->bios[i] = NULL;
869                         } else
870                                 r1_bio->bios[i] = bio;
871                         targets++;
872                 } else
873                         r1_bio->bios[i] = NULL;
874         }
875         rcu_read_unlock();
876
877         BUG_ON(targets == 0); /* we never fail the last device */
878
879         if (targets < conf->raid_disks) {
880                 /* array is degraded, we will not clear the bitmap
881                  * on I/O completion (see raid1_end_write_request) */
882                 set_bit(R1BIO_Degraded, &r1_bio->state);
883         }
884
885         /* do behind I/O ? */
886         if (bitmap &&
887             atomic_read(&bitmap->behind_writes) < bitmap->max_write_behind &&
888             (behind_pages = alloc_behind_pages(bio)) != NULL)
889                 set_bit(R1BIO_BehindIO, &r1_bio->state);
890
891         atomic_set(&r1_bio->remaining, 0);
892         atomic_set(&r1_bio->behind_remaining, 0);
893
894         do_barriers = bio_barrier(bio);
895         if (do_barriers)
896                 set_bit(R1BIO_Barrier, &r1_bio->state);
897
898         bio_list_init(&bl);
899         for (i = 0; i < disks; i++) {
900                 struct bio *mbio;
901                 if (!r1_bio->bios[i])
902                         continue;
903
904                 mbio = bio_clone(bio, GFP_NOIO);
905                 r1_bio->bios[i] = mbio;
906
907                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
908                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
909                 mbio->bi_end_io = raid1_end_write_request;
910                 mbio->bi_rw = WRITE | do_barriers | do_sync;
911                 mbio->bi_private = r1_bio;
912
913                 if (behind_pages) {
914                         struct bio_vec *bvec;
915                         int j;
916
917                         /* Yes, I really want the '__' version so that
918                          * we clear any unused pointer in the io_vec, rather
919                          * than leave them unchanged.  This is important
920                          * because when we come to free the pages, we won't
921                          * know the originial bi_idx, so we just free
922                          * them all
923                          */
924                         __bio_for_each_segment(bvec, mbio, j, 0)
925                                 bvec->bv_page = behind_pages[j];
926                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
927                                 atomic_inc(&r1_bio->behind_remaining);
928                 }
929
930                 atomic_inc(&r1_bio->remaining);
931
932                 bio_list_add(&bl, mbio);
933         }
934         kfree(behind_pages); /* the behind pages are attached to the bios now */
935
936         bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
937                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
938         spin_lock_irqsave(&conf->device_lock, flags);
939         bio_list_merge(&conf->pending_bio_list, &bl);
940         bio_list_init(&bl);
941
942         blk_plug_device(mddev->queue);
943         spin_unlock_irqrestore(&conf->device_lock, flags);
944
945         if (do_sync)
946                 md_wakeup_thread(mddev->thread);
947 #if 0
948         while ((bio = bio_list_pop(&bl)) != NULL)
949                 generic_make_request(bio);
950 #endif
951
952         return 0;
953 }
954
955 static void status(struct seq_file *seq, mddev_t *mddev)
956 {
957         conf_t *conf = mddev_to_conf(mddev);
958         int i;
959
960         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
961                    conf->raid_disks - mddev->degraded);
962         rcu_read_lock();
963         for (i = 0; i < conf->raid_disks; i++) {
964                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
965                 seq_printf(seq, "%s",
966                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
967         }
968         rcu_read_unlock();
969         seq_printf(seq, "]");
970 }
971
972
973 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
974 {
975         char b[BDEVNAME_SIZE];
976         conf_t *conf = mddev_to_conf(mddev);
977
978         /*
979          * If it is not operational, then we have already marked it as dead
980          * else if it is the last working disks, ignore the error, let the
981          * next level up know.
982          * else mark the drive as failed
983          */
984         if (test_bit(In_sync, &rdev->flags)
985             && (conf->raid_disks - mddev->degraded) == 1)
986                 /*
987                  * Don't fail the drive, act as though we were just a
988                  * normal single drive
989                  */
990                 return;
991         if (test_and_clear_bit(In_sync, &rdev->flags)) {
992                 unsigned long flags;
993                 spin_lock_irqsave(&conf->device_lock, flags);
994                 mddev->degraded++;
995                 spin_unlock_irqrestore(&conf->device_lock, flags);
996                 /*
997                  * if recovery is running, make sure it aborts.
998                  */
999                 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
1000         }
1001         set_bit(Faulty, &rdev->flags);
1002         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1003         printk(KERN_ALERT "raid1: Disk failure on %s, disabling device. \n"
1004                 "       Operation continuing on %d devices\n",
1005                 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1006 }
1007
1008 static void print_conf(conf_t *conf)
1009 {
1010         int i;
1011
1012         printk("RAID1 conf printout:\n");
1013         if (!conf) {
1014                 printk("(!conf)\n");
1015                 return;
1016         }
1017         printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1018                 conf->raid_disks);
1019
1020         rcu_read_lock();
1021         for (i = 0; i < conf->raid_disks; i++) {
1022                 char b[BDEVNAME_SIZE];
1023                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1024                 if (rdev)
1025                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
1026                                i, !test_bit(In_sync, &rdev->flags),
1027                                !test_bit(Faulty, &rdev->flags),
1028                                bdevname(rdev->bdev,b));
1029         }
1030         rcu_read_unlock();
1031 }
1032
1033 static void close_sync(conf_t *conf)
1034 {
1035         wait_barrier(conf);
1036         allow_barrier(conf);
1037
1038         mempool_destroy(conf->r1buf_pool);
1039         conf->r1buf_pool = NULL;
1040 }
1041
1042 static int raid1_spare_active(mddev_t *mddev)
1043 {
1044         int i;
1045         conf_t *conf = mddev->private;
1046
1047         /*
1048          * Find all failed disks within the RAID1 configuration 
1049          * and mark them readable.
1050          * Called under mddev lock, so rcu protection not needed.
1051          */
1052         for (i = 0; i < conf->raid_disks; i++) {
1053                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1054                 if (rdev
1055                     && !test_bit(Faulty, &rdev->flags)
1056                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1057                         unsigned long flags;
1058                         spin_lock_irqsave(&conf->device_lock, flags);
1059                         mddev->degraded--;
1060                         spin_unlock_irqrestore(&conf->device_lock, flags);
1061                 }
1062         }
1063
1064         print_conf(conf);
1065         return 0;
1066 }
1067
1068
1069 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1070 {
1071         conf_t *conf = mddev->private;
1072         int found = 0;
1073         int mirror = 0;
1074         mirror_info_t *p;
1075
1076         for (mirror=0; mirror < mddev->raid_disks; mirror++)
1077                 if ( !(p=conf->mirrors+mirror)->rdev) {
1078
1079                         blk_queue_stack_limits(mddev->queue,
1080                                                rdev->bdev->bd_disk->queue);
1081                         /* as we don't honour merge_bvec_fn, we must never risk
1082                          * violating it, so limit ->max_sector to one PAGE, as
1083                          * a one page request is never in violation.
1084                          */
1085                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1086                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
1087                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1088
1089                         p->head_position = 0;
1090                         rdev->raid_disk = mirror;
1091                         found = 1;
1092                         /* As all devices are equivalent, we don't need a full recovery
1093                          * if this was recently any drive of the array
1094                          */
1095                         if (rdev->saved_raid_disk < 0)
1096                                 conf->fullsync = 1;
1097                         rcu_assign_pointer(p->rdev, rdev);
1098                         break;
1099                 }
1100
1101         print_conf(conf);
1102         return found;
1103 }
1104
1105 static int raid1_remove_disk(mddev_t *mddev, int number)
1106 {
1107         conf_t *conf = mddev->private;
1108         int err = 0;
1109         mdk_rdev_t *rdev;
1110         mirror_info_t *p = conf->mirrors+ number;
1111
1112         print_conf(conf);
1113         rdev = p->rdev;
1114         if (rdev) {
1115                 if (test_bit(In_sync, &rdev->flags) ||
1116                     atomic_read(&rdev->nr_pending)) {
1117                         err = -EBUSY;
1118                         goto abort;
1119                 }
1120                 p->rdev = NULL;
1121                 synchronize_rcu();
1122                 if (atomic_read(&rdev->nr_pending)) {
1123                         /* lost the race, try later */
1124                         err = -EBUSY;
1125                         p->rdev = rdev;
1126                 }
1127         }
1128 abort:
1129
1130         print_conf(conf);
1131         return err;
1132 }
1133
1134
1135 static int end_sync_read(struct bio *bio, unsigned int bytes_done, int error)
1136 {
1137         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1138         int i;
1139
1140         if (bio->bi_size)
1141                 return 1;
1142
1143         for (i=r1_bio->mddev->raid_disks; i--; )
1144                 if (r1_bio->bios[i] == bio)
1145                         break;
1146         BUG_ON(i < 0);
1147         update_head_pos(i, r1_bio);
1148         /*
1149          * we have read a block, now it needs to be re-written,
1150          * or re-read if the read failed.
1151          * We don't do much here, just schedule handling by raid1d
1152          */
1153         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1154                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1155
1156         if (atomic_dec_and_test(&r1_bio->remaining))
1157                 reschedule_retry(r1_bio);
1158         return 0;
1159 }
1160
1161 static int end_sync_write(struct bio *bio, unsigned int bytes_done, int error)
1162 {
1163         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1164         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1165         mddev_t *mddev = r1_bio->mddev;
1166         conf_t *conf = mddev_to_conf(mddev);
1167         int i;
1168         int mirror=0;
1169
1170         if (bio->bi_size)
1171                 return 1;
1172
1173         for (i = 0; i < conf->raid_disks; i++)
1174                 if (r1_bio->bios[i] == bio) {
1175                         mirror = i;
1176                         break;
1177                 }
1178         if (!uptodate) {
1179                 int sync_blocks = 0;
1180                 sector_t s = r1_bio->sector;
1181                 long sectors_to_go = r1_bio->sectors;
1182                 /* make sure these bits doesn't get cleared. */
1183                 do {
1184                         bitmap_end_sync(mddev->bitmap, s,
1185                                         &sync_blocks, 1);
1186                         s += sync_blocks;
1187                         sectors_to_go -= sync_blocks;
1188                 } while (sectors_to_go > 0);
1189                 md_error(mddev, conf->mirrors[mirror].rdev);
1190         }
1191
1192         update_head_pos(mirror, r1_bio);
1193
1194         if (atomic_dec_and_test(&r1_bio->remaining)) {
1195                 md_done_sync(mddev, r1_bio->sectors, uptodate);
1196                 put_buf(r1_bio);
1197         }
1198         return 0;
1199 }
1200
1201 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1202 {
1203         conf_t *conf = mddev_to_conf(mddev);
1204         int i;
1205         int disks = conf->raid_disks;
1206         struct bio *bio, *wbio;
1207
1208         bio = r1_bio->bios[r1_bio->read_disk];
1209
1210
1211         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1212                 /* We have read all readable devices.  If we haven't
1213                  * got the block, then there is no hope left.
1214                  * If we have, then we want to do a comparison
1215                  * and skip the write if everything is the same.
1216                  * If any blocks failed to read, then we need to
1217                  * attempt an over-write
1218                  */
1219                 int primary;
1220                 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1221                         for (i=0; i<mddev->raid_disks; i++)
1222                                 if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1223                                         md_error(mddev, conf->mirrors[i].rdev);
1224
1225                         md_done_sync(mddev, r1_bio->sectors, 1);
1226                         put_buf(r1_bio);
1227                         return;
1228                 }
1229                 for (primary=0; primary<mddev->raid_disks; primary++)
1230                         if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1231                             test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1232                                 r1_bio->bios[primary]->bi_end_io = NULL;
1233                                 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1234                                 break;
1235                         }
1236                 r1_bio->read_disk = primary;
1237                 for (i=0; i<mddev->raid_disks; i++)
1238                         if (r1_bio->bios[i]->bi_end_io == end_sync_read &&
1239                             test_bit(BIO_UPTODATE, &r1_bio->bios[i]->bi_flags)) {
1240                                 int j;
1241                                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1242                                 struct bio *pbio = r1_bio->bios[primary];
1243                                 struct bio *sbio = r1_bio->bios[i];
1244                                 for (j = vcnt; j-- ; )
1245                                         if (memcmp(page_address(pbio->bi_io_vec[j].bv_page),
1246                                                    page_address(sbio->bi_io_vec[j].bv_page),
1247                                                    PAGE_SIZE))
1248                                                 break;
1249                                 if (j >= 0)
1250                                         mddev->resync_mismatches += r1_bio->sectors;
1251                                 if (j < 0 || test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
1252                                         sbio->bi_end_io = NULL;
1253                                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1254                                 } else {
1255                                         /* fixup the bio for reuse */
1256                                         sbio->bi_vcnt = vcnt;
1257                                         sbio->bi_size = r1_bio->sectors << 9;
1258                                         sbio->bi_idx = 0;
1259                                         sbio->bi_phys_segments = 0;
1260                                         sbio->bi_hw_segments = 0;
1261                                         sbio->bi_hw_front_size = 0;
1262                                         sbio->bi_hw_back_size = 0;
1263                                         sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1264                                         sbio->bi_flags |= 1 << BIO_UPTODATE;
1265                                         sbio->bi_next = NULL;
1266                                         sbio->bi_sector = r1_bio->sector +
1267                                                 conf->mirrors[i].rdev->data_offset;
1268                                         sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1269                                         for (j = 0; j < vcnt ; j++)
1270                                                 memcpy(page_address(sbio->bi_io_vec[j].bv_page),
1271                                                        page_address(pbio->bi_io_vec[j].bv_page),
1272                                                        PAGE_SIZE);
1273
1274                                 }
1275                         }
1276         }
1277         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1278                 /* ouch - failed to read all of that.
1279                  * Try some synchronous reads of other devices to get
1280                  * good data, much like with normal read errors.  Only
1281                  * read into the pages we already have so we don't
1282                  * need to re-issue the read request.
1283                  * We don't need to freeze the array, because being in an
1284                  * active sync request, there is no normal IO, and
1285                  * no overlapping syncs.
1286                  */
1287                 sector_t sect = r1_bio->sector;
1288                 int sectors = r1_bio->sectors;
1289                 int idx = 0;
1290
1291                 while(sectors) {
1292                         int s = sectors;
1293                         int d = r1_bio->read_disk;
1294                         int success = 0;
1295                         mdk_rdev_t *rdev;
1296
1297                         if (s > (PAGE_SIZE>>9))
1298                                 s = PAGE_SIZE >> 9;
1299                         do {
1300                                 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1301                                         /* No rcu protection needed here devices
1302                                          * can only be removed when no resync is
1303                                          * active, and resync is currently active
1304                                          */
1305                                         rdev = conf->mirrors[d].rdev;
1306                                         if (sync_page_io(rdev->bdev,
1307                                                          sect + rdev->data_offset,
1308                                                          s<<9,
1309                                                          bio->bi_io_vec[idx].bv_page,
1310                                                          READ)) {
1311                                                 success = 1;
1312                                                 break;
1313                                         }
1314                                 }
1315                                 d++;
1316                                 if (d == conf->raid_disks)
1317                                         d = 0;
1318                         } while (!success && d != r1_bio->read_disk);
1319
1320                         if (success) {
1321                                 int start = d;
1322                                 /* write it back and re-read */
1323                                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1324                                 while (d != r1_bio->read_disk) {
1325                                         if (d == 0)
1326                                                 d = conf->raid_disks;
1327                                         d--;
1328                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1329                                                 continue;
1330                                         rdev = conf->mirrors[d].rdev;
1331                                         atomic_add(s, &rdev->corrected_errors);
1332                                         if (sync_page_io(rdev->bdev,
1333                                                          sect + rdev->data_offset,
1334                                                          s<<9,
1335                                                          bio->bi_io_vec[idx].bv_page,
1336                                                          WRITE) == 0)
1337                                                 md_error(mddev, rdev);
1338                                 }
1339                                 d = start;
1340                                 while (d != r1_bio->read_disk) {
1341                                         if (d == 0)
1342                                                 d = conf->raid_disks;
1343                                         d--;
1344                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1345                                                 continue;
1346                                         rdev = conf->mirrors[d].rdev;
1347                                         if (sync_page_io(rdev->bdev,
1348                                                          sect + rdev->data_offset,
1349                                                          s<<9,
1350                                                          bio->bi_io_vec[idx].bv_page,
1351                                                          READ) == 0)
1352                                                 md_error(mddev, rdev);
1353                                 }
1354                         } else {
1355                                 char b[BDEVNAME_SIZE];
1356                                 /* Cannot read from anywhere, array is toast */
1357                                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1358                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
1359                                        " for block %llu\n",
1360                                        bdevname(bio->bi_bdev,b),
1361                                        (unsigned long long)r1_bio->sector);
1362                                 md_done_sync(mddev, r1_bio->sectors, 0);
1363                                 put_buf(r1_bio);
1364                                 return;
1365                         }
1366                         sectors -= s;
1367                         sect += s;
1368                         idx ++;
1369                 }
1370         }
1371
1372         /*
1373          * schedule writes
1374          */
1375         atomic_set(&r1_bio->remaining, 1);
1376         for (i = 0; i < disks ; i++) {
1377                 wbio = r1_bio->bios[i];
1378                 if (wbio->bi_end_io == NULL ||
1379                     (wbio->bi_end_io == end_sync_read &&
1380                      (i == r1_bio->read_disk ||
1381                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1382                         continue;
1383
1384                 wbio->bi_rw = WRITE;
1385                 wbio->bi_end_io = end_sync_write;
1386                 atomic_inc(&r1_bio->remaining);
1387                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1388
1389                 generic_make_request(wbio);
1390         }
1391
1392         if (atomic_dec_and_test(&r1_bio->remaining)) {
1393                 /* if we're here, all write(s) have completed, so clean up */
1394                 md_done_sync(mddev, r1_bio->sectors, 1);
1395                 put_buf(r1_bio);
1396         }
1397 }
1398
1399 /*
1400  * This is a kernel thread which:
1401  *
1402  *      1.      Retries failed read operations on working mirrors.
1403  *      2.      Updates the raid superblock when problems encounter.
1404  *      3.      Performs writes following reads for array syncronising.
1405  */
1406
1407 static void fix_read_error(conf_t *conf, int read_disk,
1408                            sector_t sect, int sectors)
1409 {
1410         mddev_t *mddev = conf->mddev;
1411         while(sectors) {
1412                 int s = sectors;
1413                 int d = read_disk;
1414                 int success = 0;
1415                 int start;
1416                 mdk_rdev_t *rdev;
1417
1418                 if (s > (PAGE_SIZE>>9))
1419                         s = PAGE_SIZE >> 9;
1420
1421                 do {
1422                         /* Note: no rcu protection needed here
1423                          * as this is synchronous in the raid1d thread
1424                          * which is the thread that might remove
1425                          * a device.  If raid1d ever becomes multi-threaded....
1426                          */
1427                         rdev = conf->mirrors[d].rdev;
1428                         if (rdev &&
1429                             test_bit(In_sync, &rdev->flags) &&
1430                             sync_page_io(rdev->bdev,
1431                                          sect + rdev->data_offset,
1432                                          s<<9,
1433                                          conf->tmppage, READ))
1434                                 success = 1;
1435                         else {
1436                                 d++;
1437                                 if (d == conf->raid_disks)
1438                                         d = 0;
1439                         }
1440                 } while (!success && d != read_disk);
1441
1442                 if (!success) {
1443                         /* Cannot read from anywhere -- bye bye array */
1444                         md_error(mddev, conf->mirrors[read_disk].rdev);
1445                         break;
1446                 }
1447                 /* write it back and re-read */
1448                 start = d;
1449                 while (d != read_disk) {
1450                         if (d==0)
1451                                 d = conf->raid_disks;
1452                         d--;
1453                         rdev = conf->mirrors[d].rdev;
1454                         if (rdev &&
1455                             test_bit(In_sync, &rdev->flags)) {
1456                                 if (sync_page_io(rdev->bdev,
1457                                                  sect + rdev->data_offset,
1458                                                  s<<9, conf->tmppage, WRITE)
1459                                     == 0)
1460                                         /* Well, this device is dead */
1461                                         md_error(mddev, rdev);
1462                         }
1463                 }
1464                 d = start;
1465                 while (d != read_disk) {
1466                         char b[BDEVNAME_SIZE];
1467                         if (d==0)
1468                                 d = conf->raid_disks;
1469                         d--;
1470                         rdev = conf->mirrors[d].rdev;
1471                         if (rdev &&
1472                             test_bit(In_sync, &rdev->flags)) {
1473                                 if (sync_page_io(rdev->bdev,
1474                                                  sect + rdev->data_offset,
1475                                                  s<<9, conf->tmppage, READ)
1476                                     == 0)
1477                                         /* Well, this device is dead */
1478                                         md_error(mddev, rdev);
1479                                 else {
1480                                         atomic_add(s, &rdev->corrected_errors);
1481                                         printk(KERN_INFO
1482                                                "raid1:%s: read error corrected "
1483                                                "(%d sectors at %llu on %s)\n",
1484                                                mdname(mddev), s,
1485                                                (unsigned long long)(sect +
1486                                                    rdev->data_offset),
1487                                                bdevname(rdev->bdev, b));
1488                                 }
1489                         }
1490                 }
1491                 sectors -= s;
1492                 sect += s;
1493         }
1494 }
1495
1496 static void raid1d(mddev_t *mddev)
1497 {
1498         r1bio_t *r1_bio;
1499         struct bio *bio;
1500         unsigned long flags;
1501         conf_t *conf = mddev_to_conf(mddev);
1502         struct list_head *head = &conf->retry_list;
1503         int unplug=0;
1504         mdk_rdev_t *rdev;
1505
1506         md_check_recovery(mddev);
1507         
1508         for (;;) {
1509                 char b[BDEVNAME_SIZE];
1510                 spin_lock_irqsave(&conf->device_lock, flags);
1511
1512                 if (conf->pending_bio_list.head) {
1513                         bio = bio_list_get(&conf->pending_bio_list);
1514                         blk_remove_plug(mddev->queue);
1515                         spin_unlock_irqrestore(&conf->device_lock, flags);
1516                         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
1517                         if (bitmap_unplug(mddev->bitmap) != 0)
1518                                 printk("%s: bitmap file write failed!\n", mdname(mddev));
1519
1520                         while (bio) { /* submit pending writes */
1521                                 struct bio *next = bio->bi_next;
1522                                 bio->bi_next = NULL;
1523                                 generic_make_request(bio);
1524                                 bio = next;
1525                         }
1526                         unplug = 1;
1527
1528                         continue;
1529                 }
1530
1531                 if (list_empty(head))
1532                         break;
1533                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1534                 list_del(head->prev);
1535                 conf->nr_queued--;
1536                 spin_unlock_irqrestore(&conf->device_lock, flags);
1537
1538                 mddev = r1_bio->mddev;
1539                 conf = mddev_to_conf(mddev);
1540                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1541                         sync_request_write(mddev, r1_bio);
1542                         unplug = 1;
1543                 } else if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
1544                         /* some requests in the r1bio were BIO_RW_BARRIER
1545                          * requests which failed with -EOPNOTSUPP.  Hohumm..
1546                          * Better resubmit without the barrier.
1547                          * We know which devices to resubmit for, because
1548                          * all others have had their bios[] entry cleared.
1549                          * We already have a nr_pending reference on these rdevs.
1550                          */
1551                         int i;
1552                         const int do_sync = bio_sync(r1_bio->master_bio);
1553                         clear_bit(R1BIO_BarrierRetry, &r1_bio->state);
1554                         clear_bit(R1BIO_Barrier, &r1_bio->state);
1555                         for (i=0; i < conf->raid_disks; i++)
1556                                 if (r1_bio->bios[i])
1557                                         atomic_inc(&r1_bio->remaining);
1558                         for (i=0; i < conf->raid_disks; i++)
1559                                 if (r1_bio->bios[i]) {
1560                                         struct bio_vec *bvec;
1561                                         int j;
1562
1563                                         bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1564                                         /* copy pages from the failed bio, as
1565                                          * this might be a write-behind device */
1566                                         __bio_for_each_segment(bvec, bio, j, 0)
1567                                                 bvec->bv_page = bio_iovec_idx(r1_bio->bios[i], j)->bv_page;
1568                                         bio_put(r1_bio->bios[i]);
1569                                         bio->bi_sector = r1_bio->sector +
1570                                                 conf->mirrors[i].rdev->data_offset;
1571                                         bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1572                                         bio->bi_end_io = raid1_end_write_request;
1573                                         bio->bi_rw = WRITE | do_sync;
1574                                         bio->bi_private = r1_bio;
1575                                         r1_bio->bios[i] = bio;
1576                                         generic_make_request(bio);
1577                                 }
1578                 } else {
1579                         int disk;
1580
1581                         /* we got a read error. Maybe the drive is bad.  Maybe just
1582                          * the block and we can fix it.
1583                          * We freeze all other IO, and try reading the block from
1584                          * other devices.  When we find one, we re-write
1585                          * and check it that fixes the read error.
1586                          * This is all done synchronously while the array is
1587                          * frozen
1588                          */
1589                         if (mddev->ro == 0) {
1590                                 freeze_array(conf);
1591                                 fix_read_error(conf, r1_bio->read_disk,
1592                                                r1_bio->sector,
1593                                                r1_bio->sectors);
1594                                 unfreeze_array(conf);
1595                         }
1596
1597                         bio = r1_bio->bios[r1_bio->read_disk];
1598                         if ((disk=read_balance(conf, r1_bio)) == -1) {
1599                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
1600                                        " read error for block %llu\n",
1601                                        bdevname(bio->bi_bdev,b),
1602                                        (unsigned long long)r1_bio->sector);
1603                                 raid_end_bio_io(r1_bio);
1604                         } else {
1605                                 const int do_sync = bio_sync(r1_bio->master_bio);
1606                                 r1_bio->bios[r1_bio->read_disk] =
1607                                         mddev->ro ? IO_BLOCKED : NULL;
1608                                 r1_bio->read_disk = disk;
1609                                 bio_put(bio);
1610                                 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1611                                 r1_bio->bios[r1_bio->read_disk] = bio;
1612                                 rdev = conf->mirrors[disk].rdev;
1613                                 if (printk_ratelimit())
1614                                         printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
1615                                                " another mirror\n",
1616                                                bdevname(rdev->bdev,b),
1617                                                (unsigned long long)r1_bio->sector);
1618                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1619                                 bio->bi_bdev = rdev->bdev;
1620                                 bio->bi_end_io = raid1_end_read_request;
1621                                 bio->bi_rw = READ | do_sync;
1622                                 bio->bi_private = r1_bio;
1623                                 unplug = 1;
1624                                 generic_make_request(bio);
1625                         }
1626                 }
1627         }
1628         spin_unlock_irqrestore(&conf->device_lock, flags);
1629         if (unplug)
1630                 unplug_slaves(mddev);
1631 }
1632
1633
1634 static int init_resync(conf_t *conf)
1635 {
1636         int buffs;
1637
1638         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1639         BUG_ON(conf->r1buf_pool);
1640         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1641                                           conf->poolinfo);
1642         if (!conf->r1buf_pool)
1643                 return -ENOMEM;
1644         conf->next_resync = 0;
1645         return 0;
1646 }
1647
1648 /*
1649  * perform a "sync" on one "block"
1650  *
1651  * We need to make sure that no normal I/O request - particularly write
1652  * requests - conflict with active sync requests.
1653  *
1654  * This is achieved by tracking pending requests and a 'barrier' concept
1655  * that can be installed to exclude normal IO requests.
1656  */
1657
1658 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1659 {
1660         conf_t *conf = mddev_to_conf(mddev);
1661         r1bio_t *r1_bio;
1662         struct bio *bio;
1663         sector_t max_sector, nr_sectors;
1664         int disk = -1;
1665         int i;
1666         int wonly = -1;
1667         int write_targets = 0, read_targets = 0;
1668         int sync_blocks;
1669         int still_degraded = 0;
1670
1671         if (!conf->r1buf_pool)
1672         {
1673 /*
1674                 printk("sync start - bitmap %p\n", mddev->bitmap);
1675 */
1676                 if (init_resync(conf))
1677                         return 0;
1678         }
1679
1680         max_sector = mddev->size << 1;
1681         if (sector_nr >= max_sector) {
1682                 /* If we aborted, we need to abort the
1683                  * sync on the 'current' bitmap chunk (there will
1684                  * only be one in raid1 resync.
1685                  * We can find the current addess in mddev->curr_resync
1686                  */
1687                 if (mddev->curr_resync < max_sector) /* aborted */
1688                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1689                                                 &sync_blocks, 1);
1690                 else /* completed sync */
1691                         conf->fullsync = 0;
1692
1693                 bitmap_close_sync(mddev->bitmap);
1694                 close_sync(conf);
1695                 return 0;
1696         }
1697
1698         if (mddev->bitmap == NULL &&
1699             mddev->recovery_cp == MaxSector &&
1700             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1701             conf->fullsync == 0) {
1702                 *skipped = 1;
1703                 return max_sector - sector_nr;
1704         }
1705         /* before building a request, check if we can skip these blocks..
1706          * This call the bitmap_start_sync doesn't actually record anything
1707          */
1708         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1709             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1710                 /* We can skip this block, and probably several more */
1711                 *skipped = 1;
1712                 return sync_blocks;
1713         }
1714         /*
1715          * If there is non-resync activity waiting for a turn,
1716          * and resync is going fast enough,
1717          * then let it though before starting on this new sync request.
1718          */
1719         if (!go_faster && conf->nr_waiting)
1720                 msleep_interruptible(1000);
1721
1722         raise_barrier(conf);
1723
1724         conf->next_resync = sector_nr;
1725
1726         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1727         rcu_read_lock();
1728         /*
1729          * If we get a correctably read error during resync or recovery,
1730          * we might want to read from a different device.  So we
1731          * flag all drives that could conceivably be read from for READ,
1732          * and any others (which will be non-In_sync devices) for WRITE.
1733          * If a read fails, we try reading from something else for which READ
1734          * is OK.
1735          */
1736
1737         r1_bio->mddev = mddev;
1738         r1_bio->sector = sector_nr;
1739         r1_bio->state = 0;
1740         set_bit(R1BIO_IsSync, &r1_bio->state);
1741
1742         for (i=0; i < conf->raid_disks; i++) {
1743                 mdk_rdev_t *rdev;
1744                 bio = r1_bio->bios[i];
1745
1746                 /* take from bio_init */
1747                 bio->bi_next = NULL;
1748                 bio->bi_flags |= 1 << BIO_UPTODATE;
1749                 bio->bi_rw = READ;
1750                 bio->bi_vcnt = 0;
1751                 bio->bi_idx = 0;
1752                 bio->bi_phys_segments = 0;
1753                 bio->bi_hw_segments = 0;
1754                 bio->bi_size = 0;
1755                 bio->bi_end_io = NULL;
1756                 bio->bi_private = NULL;
1757
1758                 rdev = rcu_dereference(conf->mirrors[i].rdev);
1759                 if (rdev == NULL ||
1760                            test_bit(Faulty, &rdev->flags)) {
1761                         still_degraded = 1;
1762                         continue;
1763                 } else if (!test_bit(In_sync, &rdev->flags)) {
1764                         bio->bi_rw = WRITE;
1765                         bio->bi_end_io = end_sync_write;
1766                         write_targets ++;
1767                 } else {
1768                         /* may need to read from here */
1769                         bio->bi_rw = READ;
1770                         bio->bi_end_io = end_sync_read;
1771                         if (test_bit(WriteMostly, &rdev->flags)) {
1772                                 if (wonly < 0)
1773                                         wonly = i;
1774                         } else {
1775                                 if (disk < 0)
1776                                         disk = i;
1777                         }
1778                         read_targets++;
1779                 }
1780                 atomic_inc(&rdev->nr_pending);
1781                 bio->bi_sector = sector_nr + rdev->data_offset;
1782                 bio->bi_bdev = rdev->bdev;
1783                 bio->bi_private = r1_bio;
1784         }
1785         rcu_read_unlock();
1786         if (disk < 0)
1787                 disk = wonly;
1788         r1_bio->read_disk = disk;
1789
1790         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1791                 /* extra read targets are also write targets */
1792                 write_targets += read_targets-1;
1793
1794         if (write_targets == 0 || read_targets == 0) {
1795                 /* There is nowhere to write, so all non-sync
1796                  * drives must be failed - so we are finished
1797                  */
1798                 sector_t rv = max_sector - sector_nr;
1799                 *skipped = 1;
1800                 put_buf(r1_bio);
1801                 return rv;
1802         }
1803
1804         nr_sectors = 0;
1805         sync_blocks = 0;
1806         do {
1807                 struct page *page;
1808                 int len = PAGE_SIZE;
1809                 if (sector_nr + (len>>9) > max_sector)
1810                         len = (max_sector - sector_nr) << 9;
1811                 if (len == 0)
1812                         break;
1813                 if (sync_blocks == 0) {
1814                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1815                                                &sync_blocks, still_degraded) &&
1816                             !conf->fullsync &&
1817                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1818                                 break;
1819                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1820                         if (len > (sync_blocks<<9))
1821                                 len = sync_blocks<<9;
1822                 }
1823
1824                 for (i=0 ; i < conf->raid_disks; i++) {
1825                         bio = r1_bio->bios[i];
1826                         if (bio->bi_end_io) {
1827                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1828                                 if (bio_add_page(bio, page, len, 0) == 0) {
1829                                         /* stop here */
1830                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1831                                         while (i > 0) {
1832                                                 i--;
1833                                                 bio = r1_bio->bios[i];
1834                                                 if (bio->bi_end_io==NULL)
1835                                                         continue;
1836                                                 /* remove last page from this bio */
1837                                                 bio->bi_vcnt--;
1838                                                 bio->bi_size -= len;
1839                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1840                                         }
1841                                         goto bio_full;
1842                                 }
1843                         }
1844                 }
1845                 nr_sectors += len>>9;
1846                 sector_nr += len>>9;
1847                 sync_blocks -= (len>>9);
1848         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1849  bio_full:
1850         r1_bio->sectors = nr_sectors;
1851
1852         /* For a user-requested sync, we read all readable devices and do a
1853          * compare
1854          */
1855         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1856                 atomic_set(&r1_bio->remaining, read_targets);
1857                 for (i=0; i<conf->raid_disks; i++) {
1858                         bio = r1_bio->bios[i];
1859                         if (bio->bi_end_io == end_sync_read) {
1860                                 md_sync_acct(bio->bi_bdev, nr_sectors);
1861                                 generic_make_request(bio);
1862                         }
1863                 }
1864         } else {
1865                 atomic_set(&r1_bio->remaining, 1);
1866                 bio = r1_bio->bios[r1_bio->read_disk];
1867                 md_sync_acct(bio->bi_bdev, nr_sectors);
1868                 generic_make_request(bio);
1869
1870         }
1871         return nr_sectors;
1872 }
1873
1874 static int run(mddev_t *mddev)
1875 {
1876         conf_t *conf;
1877         int i, j, disk_idx;
1878         mirror_info_t *disk;
1879         mdk_rdev_t *rdev;
1880         struct list_head *tmp;
1881
1882         if (mddev->level != 1) {
1883                 printk("raid1: %s: raid level not set to mirroring (%d)\n",
1884                        mdname(mddev), mddev->level);
1885                 goto out;
1886         }
1887         if (mddev->reshape_position != MaxSector) {
1888                 printk("raid1: %s: reshape_position set but not supported\n",
1889                        mdname(mddev));
1890                 goto out;
1891         }
1892         /*
1893          * copy the already verified devices into our private RAID1
1894          * bookkeeping area. [whatever we allocate in run(),
1895          * should be freed in stop()]
1896          */
1897         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1898         mddev->private = conf;
1899         if (!conf)
1900                 goto out_no_mem;
1901
1902         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1903                                  GFP_KERNEL);
1904         if (!conf->mirrors)
1905                 goto out_no_mem;
1906
1907         conf->tmppage = alloc_page(GFP_KERNEL);
1908         if (!conf->tmppage)
1909                 goto out_no_mem;
1910
1911         conf->poolinfo = kmalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1912         if (!conf->poolinfo)
1913                 goto out_no_mem;
1914         conf->poolinfo->mddev = mddev;
1915         conf->poolinfo->raid_disks = mddev->raid_disks;
1916         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1917                                           r1bio_pool_free,
1918                                           conf->poolinfo);
1919         if (!conf->r1bio_pool)
1920                 goto out_no_mem;
1921
1922         ITERATE_RDEV(mddev, rdev, tmp) {
1923                 disk_idx = rdev->raid_disk;
1924                 if (disk_idx >= mddev->raid_disks
1925                     || disk_idx < 0)
1926                         continue;
1927                 disk = conf->mirrors + disk_idx;
1928
1929                 disk->rdev = rdev;
1930
1931                 blk_queue_stack_limits(mddev->queue,
1932                                        rdev->bdev->bd_disk->queue);
1933                 /* as we don't honour merge_bvec_fn, we must never risk
1934                  * violating it, so limit ->max_sector to one PAGE, as
1935                  * a one page request is never in violation.
1936                  */
1937                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1938                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
1939                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1940
1941                 disk->head_position = 0;
1942         }
1943         conf->raid_disks = mddev->raid_disks;
1944         conf->mddev = mddev;
1945         spin_lock_init(&conf->device_lock);
1946         INIT_LIST_HEAD(&conf->retry_list);
1947
1948         spin_lock_init(&conf->resync_lock);
1949         init_waitqueue_head(&conf->wait_barrier);
1950
1951         bio_list_init(&conf->pending_bio_list);
1952         bio_list_init(&conf->flushing_bio_list);
1953
1954
1955         mddev->degraded = 0;
1956         for (i = 0; i < conf->raid_disks; i++) {
1957
1958                 disk = conf->mirrors + i;
1959
1960                 if (!disk->rdev ||
1961                     !test_bit(In_sync, &disk->rdev->flags)) {
1962                         disk->head_position = 0;
1963                         mddev->degraded++;
1964                         conf->fullsync = 1;
1965                 }
1966         }
1967         if (mddev->degraded == conf->raid_disks) {
1968                 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
1969                         mdname(mddev));
1970                 goto out_free_conf;
1971         }
1972         if (conf->raid_disks - mddev->degraded == 1)
1973                 mddev->recovery_cp = MaxSector;
1974
1975         /*
1976          * find the first working one and use it as a starting point
1977          * to read balancing.
1978          */
1979         for (j = 0; j < conf->raid_disks &&
1980                      (!conf->mirrors[j].rdev ||
1981                       !test_bit(In_sync, &conf->mirrors[j].rdev->flags)) ; j++)
1982                 /* nothing */;
1983         conf->last_used = j;
1984
1985
1986         mddev->thread = md_register_thread(raid1d, mddev, "%s_raid1");
1987         if (!mddev->thread) {
1988                 printk(KERN_ERR
1989                        "raid1: couldn't allocate thread for %s\n",
1990                        mdname(mddev));
1991                 goto out_free_conf;
1992         }
1993
1994         printk(KERN_INFO 
1995                 "raid1: raid set %s active with %d out of %d mirrors\n",
1996                 mdname(mddev), mddev->raid_disks - mddev->degraded, 
1997                 mddev->raid_disks);
1998         /*
1999          * Ok, everything is just fine now
2000          */
2001         mddev->array_size = mddev->size;
2002
2003         mddev->queue->unplug_fn = raid1_unplug;
2004         mddev->queue->issue_flush_fn = raid1_issue_flush;
2005         mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2006         mddev->queue->backing_dev_info.congested_data = mddev;
2007
2008         return 0;
2009
2010 out_no_mem:
2011         printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
2012                mdname(mddev));
2013
2014 out_free_conf:
2015         if (conf) {
2016                 if (conf->r1bio_pool)
2017                         mempool_destroy(conf->r1bio_pool);
2018                 kfree(conf->mirrors);
2019                 safe_put_page(conf->tmppage);
2020                 kfree(conf->poolinfo);
2021                 kfree(conf);
2022                 mddev->private = NULL;
2023         }
2024 out:
2025         return -EIO;
2026 }
2027
2028 static int stop(mddev_t *mddev)
2029 {
2030         conf_t *conf = mddev_to_conf(mddev);
2031         struct bitmap *bitmap = mddev->bitmap;
2032         int behind_wait = 0;
2033
2034         /* wait for behind writes to complete */
2035         while (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2036                 behind_wait++;
2037                 printk(KERN_INFO "raid1: behind writes in progress on device %s, waiting to stop (%d)\n", mdname(mddev), behind_wait);
2038                 set_current_state(TASK_UNINTERRUPTIBLE);
2039                 schedule_timeout(HZ); /* wait a second */
2040                 /* need to kick something here to make sure I/O goes? */
2041         }
2042
2043         md_unregister_thread(mddev->thread);
2044         mddev->thread = NULL;
2045         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2046         if (conf->r1bio_pool)
2047                 mempool_destroy(conf->r1bio_pool);
2048         kfree(conf->mirrors);
2049         kfree(conf->poolinfo);
2050         kfree(conf);
2051         mddev->private = NULL;
2052         return 0;
2053 }
2054
2055 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2056 {
2057         /* no resync is happening, and there is enough space
2058          * on all devices, so we can resize.
2059          * We need to make sure resync covers any new space.
2060          * If the array is shrinking we should possibly wait until
2061          * any io in the removed space completes, but it hardly seems
2062          * worth it.
2063          */
2064         mddev->array_size = sectors>>1;
2065         set_capacity(mddev->gendisk, mddev->array_size << 1);
2066         mddev->changed = 1;
2067         if (mddev->array_size > mddev->size && mddev->recovery_cp == MaxSector) {
2068                 mddev->recovery_cp = mddev->size << 1;
2069                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2070         }
2071         mddev->size = mddev->array_size;
2072         mddev->resync_max_sectors = sectors;
2073         return 0;
2074 }
2075
2076 static int raid1_reshape(mddev_t *mddev)
2077 {
2078         /* We need to:
2079          * 1/ resize the r1bio_pool
2080          * 2/ resize conf->mirrors
2081          *
2082          * We allocate a new r1bio_pool if we can.
2083          * Then raise a device barrier and wait until all IO stops.
2084          * Then resize conf->mirrors and swap in the new r1bio pool.
2085          *
2086          * At the same time, we "pack" the devices so that all the missing
2087          * devices have the higher raid_disk numbers.
2088          */
2089         mempool_t *newpool, *oldpool;
2090         struct pool_info *newpoolinfo;
2091         mirror_info_t *newmirrors;
2092         conf_t *conf = mddev_to_conf(mddev);
2093         int cnt, raid_disks;
2094         unsigned long flags;
2095         int d, d2;
2096
2097         /* Cannot change chunk_size, layout, or level */
2098         if (mddev->chunk_size != mddev->new_chunk ||
2099             mddev->layout != mddev->new_layout ||
2100             mddev->level != mddev->new_level) {
2101                 mddev->new_chunk = mddev->chunk_size;
2102                 mddev->new_layout = mddev->layout;
2103                 mddev->new_level = mddev->level;
2104                 return -EINVAL;
2105         }
2106
2107         raid_disks = mddev->raid_disks + mddev->delta_disks;
2108
2109         if (raid_disks < conf->raid_disks) {
2110                 cnt=0;
2111                 for (d= 0; d < conf->raid_disks; d++)
2112                         if (conf->mirrors[d].rdev)
2113                                 cnt++;
2114                 if (cnt > raid_disks)
2115                         return -EBUSY;
2116         }
2117
2118         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2119         if (!newpoolinfo)
2120                 return -ENOMEM;
2121         newpoolinfo->mddev = mddev;
2122         newpoolinfo->raid_disks = raid_disks;
2123
2124         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2125                                  r1bio_pool_free, newpoolinfo);
2126         if (!newpool) {
2127                 kfree(newpoolinfo);
2128                 return -ENOMEM;
2129         }
2130         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2131         if (!newmirrors) {
2132                 kfree(newpoolinfo);
2133                 mempool_destroy(newpool);
2134                 return -ENOMEM;
2135         }
2136
2137         raise_barrier(conf);
2138
2139         /* ok, everything is stopped */
2140         oldpool = conf->r1bio_pool;
2141         conf->r1bio_pool = newpool;
2142
2143         for (d=d2=0; d < conf->raid_disks; d++)
2144                 if (conf->mirrors[d].rdev) {
2145                         conf->mirrors[d].rdev->raid_disk = d2;
2146                         newmirrors[d2++].rdev = conf->mirrors[d].rdev;
2147                 }
2148         kfree(conf->mirrors);
2149         conf->mirrors = newmirrors;
2150         kfree(conf->poolinfo);
2151         conf->poolinfo = newpoolinfo;
2152
2153         spin_lock_irqsave(&conf->device_lock, flags);
2154         mddev->degraded += (raid_disks - conf->raid_disks);
2155         spin_unlock_irqrestore(&conf->device_lock, flags);
2156         conf->raid_disks = mddev->raid_disks = raid_disks;
2157         mddev->delta_disks = 0;
2158
2159         conf->last_used = 0; /* just make sure it is in-range */
2160         lower_barrier(conf);
2161
2162         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2163         md_wakeup_thread(mddev->thread);
2164
2165         mempool_destroy(oldpool);
2166         return 0;
2167 }
2168
2169 static void raid1_quiesce(mddev_t *mddev, int state)
2170 {
2171         conf_t *conf = mddev_to_conf(mddev);
2172
2173         switch(state) {
2174         case 1:
2175                 raise_barrier(conf);
2176                 break;
2177         case 0:
2178                 lower_barrier(conf);
2179                 break;
2180         }
2181 }
2182
2183
2184 static struct mdk_personality raid1_personality =
2185 {
2186         .name           = "raid1",
2187         .level          = 1,
2188         .owner          = THIS_MODULE,
2189         .make_request   = make_request,
2190         .run            = run,
2191         .stop           = stop,
2192         .status         = status,
2193         .error_handler  = error,
2194         .hot_add_disk   = raid1_add_disk,
2195         .hot_remove_disk= raid1_remove_disk,
2196         .spare_active   = raid1_spare_active,
2197         .sync_request   = sync_request,
2198         .resize         = raid1_resize,
2199         .check_reshape  = raid1_reshape,
2200         .quiesce        = raid1_quiesce,
2201 };
2202
2203 static int __init raid_init(void)
2204 {
2205         return register_md_personality(&raid1_personality);
2206 }
2207
2208 static void raid_exit(void)
2209 {
2210         unregister_md_personality(&raid1_personality);
2211 }
2212
2213 module_init(raid_init);
2214 module_exit(raid_exit);
2215 MODULE_LICENSE("GPL");
2216 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2217 MODULE_ALIAS("md-raid1");
2218 MODULE_ALIAS("md-level-1");