Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-2.6
[linux-2.6] / drivers / md / raid1.c
1 /*
2  * raid1.c : Multiple Devices driver for Linux
3  *
4  * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
5  *
6  * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
7  *
8  * RAID-1 management functions.
9  *
10  * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
11  *
12  * Fixes to reconstruction by Jakob Ã˜stergaard" <jakob@ostenfeld.dk>
13  * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
14  *
15  * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
16  * bitmapped intelligence in resync:
17  *
18  *      - bitmap marked during normal i/o
19  *      - bitmap used to skip nondirty blocks during sync
20  *
21  * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
22  * - persistent bitmap code
23  *
24  * This program is free software; you can redistribute it and/or modify
25  * it under the terms of the GNU General Public License as published by
26  * the Free Software Foundation; either version 2, or (at your option)
27  * any later version.
28  *
29  * You should have received a copy of the GNU General Public License
30  * (for example /usr/src/linux/COPYING); if not, write to the Free
31  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
32  */
33
34 #include "dm-bio-list.h"
35 #include <linux/delay.h>
36 #include <linux/raid/raid1.h>
37 #include <linux/raid/bitmap.h>
38
39 #define DEBUG 0
40 #if DEBUG
41 #define PRINTK(x...) printk(x)
42 #else
43 #define PRINTK(x...)
44 #endif
45
46 /*
47  * Number of guaranteed r1bios in case of extreme VM load:
48  */
49 #define NR_RAID1_BIOS 256
50
51
52 static void unplug_slaves(mddev_t *mddev);
53
54 static void allow_barrier(conf_t *conf);
55 static void lower_barrier(conf_t *conf);
56
57 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
58 {
59         struct pool_info *pi = data;
60         r1bio_t *r1_bio;
61         int size = offsetof(r1bio_t, bios[pi->raid_disks]);
62
63         /* allocate a r1bio with room for raid_disks entries in the bios array */
64         r1_bio = kzalloc(size, gfp_flags);
65         if (!r1_bio)
66                 unplug_slaves(pi->mddev);
67
68         return r1_bio;
69 }
70
71 static void r1bio_pool_free(void *r1_bio, void *data)
72 {
73         kfree(r1_bio);
74 }
75
76 #define RESYNC_BLOCK_SIZE (64*1024)
77 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
78 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
79 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
80 #define RESYNC_WINDOW (2048*1024)
81
82 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
83 {
84         struct pool_info *pi = data;
85         struct page *page;
86         r1bio_t *r1_bio;
87         struct bio *bio;
88         int i, j;
89
90         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
91         if (!r1_bio) {
92                 unplug_slaves(pi->mddev);
93                 return NULL;
94         }
95
96         /*
97          * Allocate bios : 1 for reading, n-1 for writing
98          */
99         for (j = pi->raid_disks ; j-- ; ) {
100                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
101                 if (!bio)
102                         goto out_free_bio;
103                 r1_bio->bios[j] = bio;
104         }
105         /*
106          * Allocate RESYNC_PAGES data pages and attach them to
107          * the first bio.
108          * If this is a user-requested check/repair, allocate
109          * RESYNC_PAGES for each bio.
110          */
111         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
112                 j = pi->raid_disks;
113         else
114                 j = 1;
115         while(j--) {
116                 bio = r1_bio->bios[j];
117                 for (i = 0; i < RESYNC_PAGES; i++) {
118                         page = alloc_page(gfp_flags);
119                         if (unlikely(!page))
120                                 goto out_free_pages;
121
122                         bio->bi_io_vec[i].bv_page = page;
123                 }
124         }
125         /* If not user-requests, copy the page pointers to all bios */
126         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
127                 for (i=0; i<RESYNC_PAGES ; i++)
128                         for (j=1; j<pi->raid_disks; j++)
129                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
130                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
131         }
132
133         r1_bio->master_bio = NULL;
134
135         return r1_bio;
136
137 out_free_pages:
138         for (i=0; i < RESYNC_PAGES ; i++)
139                 for (j=0 ; j < pi->raid_disks; j++)
140                         safe_put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
141         j = -1;
142 out_free_bio:
143         while ( ++j < pi->raid_disks )
144                 bio_put(r1_bio->bios[j]);
145         r1bio_pool_free(r1_bio, data);
146         return NULL;
147 }
148
149 static void r1buf_pool_free(void *__r1_bio, void *data)
150 {
151         struct pool_info *pi = data;
152         int i,j;
153         r1bio_t *r1bio = __r1_bio;
154
155         for (i = 0; i < RESYNC_PAGES; i++)
156                 for (j = pi->raid_disks; j-- ;) {
157                         if (j == 0 ||
158                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
159                             r1bio->bios[0]->bi_io_vec[i].bv_page)
160                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
161                 }
162         for (i=0 ; i < pi->raid_disks; i++)
163                 bio_put(r1bio->bios[i]);
164
165         r1bio_pool_free(r1bio, data);
166 }
167
168 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
169 {
170         int i;
171
172         for (i = 0; i < conf->raid_disks; i++) {
173                 struct bio **bio = r1_bio->bios + i;
174                 if (*bio && *bio != IO_BLOCKED)
175                         bio_put(*bio);
176                 *bio = NULL;
177         }
178 }
179
180 static void free_r1bio(r1bio_t *r1_bio)
181 {
182         conf_t *conf = mddev_to_conf(r1_bio->mddev);
183
184         /*
185          * Wake up any possible resync thread that waits for the device
186          * to go idle.
187          */
188         allow_barrier(conf);
189
190         put_all_bios(conf, r1_bio);
191         mempool_free(r1_bio, conf->r1bio_pool);
192 }
193
194 static void put_buf(r1bio_t *r1_bio)
195 {
196         conf_t *conf = mddev_to_conf(r1_bio->mddev);
197         int i;
198
199         for (i=0; i<conf->raid_disks; i++) {
200                 struct bio *bio = r1_bio->bios[i];
201                 if (bio->bi_end_io)
202                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
203         }
204
205         mempool_free(r1_bio, conf->r1buf_pool);
206
207         lower_barrier(conf);
208 }
209
210 static void reschedule_retry(r1bio_t *r1_bio)
211 {
212         unsigned long flags;
213         mddev_t *mddev = r1_bio->mddev;
214         conf_t *conf = mddev_to_conf(mddev);
215
216         spin_lock_irqsave(&conf->device_lock, flags);
217         list_add(&r1_bio->retry_list, &conf->retry_list);
218         conf->nr_queued ++;
219         spin_unlock_irqrestore(&conf->device_lock, flags);
220
221         wake_up(&conf->wait_barrier);
222         md_wakeup_thread(mddev->thread);
223 }
224
225 /*
226  * raid_end_bio_io() is called when we have finished servicing a mirrored
227  * operation and are ready to return a success/failure code to the buffer
228  * cache layer.
229  */
230 static void raid_end_bio_io(r1bio_t *r1_bio)
231 {
232         struct bio *bio = r1_bio->master_bio;
233
234         /* if nobody has done the final endio yet, do it now */
235         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
236                 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
237                         (bio_data_dir(bio) == WRITE) ? "write" : "read",
238                         (unsigned long long) bio->bi_sector,
239                         (unsigned long long) bio->bi_sector +
240                                 (bio->bi_size >> 9) - 1);
241
242                 bio_endio(bio,
243                         test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
244         }
245         free_r1bio(r1_bio);
246 }
247
248 /*
249  * Update disk head position estimator based on IRQ completion info.
250  */
251 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
252 {
253         conf_t *conf = mddev_to_conf(r1_bio->mddev);
254
255         conf->mirrors[disk].head_position =
256                 r1_bio->sector + (r1_bio->sectors);
257 }
258
259 static void raid1_end_read_request(struct bio *bio, int error)
260 {
261         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
262         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
263         int mirror;
264         conf_t *conf = mddev_to_conf(r1_bio->mddev);
265
266         mirror = r1_bio->read_disk;
267         /*
268          * this branch is our 'one mirror IO has finished' event handler:
269          */
270         update_head_pos(mirror, r1_bio);
271
272         if (uptodate)
273                 set_bit(R1BIO_Uptodate, &r1_bio->state);
274         else {
275                 /* If all other devices have failed, we want to return
276                  * the error upwards rather than fail the last device.
277                  * Here we redefine "uptodate" to mean "Don't want to retry"
278                  */
279                 unsigned long flags;
280                 spin_lock_irqsave(&conf->device_lock, flags);
281                 if (r1_bio->mddev->degraded == conf->raid_disks ||
282                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
283                      !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
284                         uptodate = 1;
285                 spin_unlock_irqrestore(&conf->device_lock, flags);
286         }
287
288         if (uptodate)
289                 raid_end_bio_io(r1_bio);
290         else {
291                 /*
292                  * oops, read error:
293                  */
294                 char b[BDEVNAME_SIZE];
295                 if (printk_ratelimit())
296                         printk(KERN_ERR "raid1: %s: rescheduling sector %llu\n",
297                                bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
298                 reschedule_retry(r1_bio);
299         }
300
301         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
302 }
303
304 static void raid1_end_write_request(struct bio *bio, int error)
305 {
306         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
307         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
308         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
309         conf_t *conf = mddev_to_conf(r1_bio->mddev);
310         struct bio *to_put = NULL;
311
312
313         for (mirror = 0; mirror < conf->raid_disks; mirror++)
314                 if (r1_bio->bios[mirror] == bio)
315                         break;
316
317         if (error == -EOPNOTSUPP && test_bit(R1BIO_Barrier, &r1_bio->state)) {
318                 set_bit(BarriersNotsupp, &conf->mirrors[mirror].rdev->flags);
319                 set_bit(R1BIO_BarrierRetry, &r1_bio->state);
320                 r1_bio->mddev->barriers_work = 0;
321                 /* Don't rdev_dec_pending in this branch - keep it for the retry */
322         } else {
323                 /*
324                  * this branch is our 'one mirror IO has finished' event handler:
325                  */
326                 r1_bio->bios[mirror] = NULL;
327                 to_put = bio;
328                 if (!uptodate) {
329                         md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
330                         /* an I/O failed, we can't clear the bitmap */
331                         set_bit(R1BIO_Degraded, &r1_bio->state);
332                 } else
333                         /*
334                          * Set R1BIO_Uptodate in our master bio, so that
335                          * we will return a good error code for to the higher
336                          * levels even if IO on some other mirrored buffer fails.
337                          *
338                          * The 'master' represents the composite IO operation to
339                          * user-side. So if something waits for IO, then it will
340                          * wait for the 'master' bio.
341                          */
342                         set_bit(R1BIO_Uptodate, &r1_bio->state);
343
344                 update_head_pos(mirror, r1_bio);
345
346                 if (behind) {
347                         if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
348                                 atomic_dec(&r1_bio->behind_remaining);
349
350                         /* In behind mode, we ACK the master bio once the I/O has safely
351                          * reached all non-writemostly disks. Setting the Returned bit
352                          * ensures that this gets done only once -- we don't ever want to
353                          * return -EIO here, instead we'll wait */
354
355                         if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
356                             test_bit(R1BIO_Uptodate, &r1_bio->state)) {
357                                 /* Maybe we can return now */
358                                 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
359                                         struct bio *mbio = r1_bio->master_bio;
360                                         PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
361                                                (unsigned long long) mbio->bi_sector,
362                                                (unsigned long long) mbio->bi_sector +
363                                                (mbio->bi_size >> 9) - 1);
364                                         bio_endio(mbio, 0);
365                                 }
366                         }
367                 }
368                 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
369         }
370         /*
371          *
372          * Let's see if all mirrored write operations have finished
373          * already.
374          */
375         if (atomic_dec_and_test(&r1_bio->remaining)) {
376                 if (test_bit(R1BIO_BarrierRetry, &r1_bio->state))
377                         reschedule_retry(r1_bio);
378                 else {
379                         /* it really is the end of this request */
380                         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
381                                 /* free extra copy of the data pages */
382                                 int i = bio->bi_vcnt;
383                                 while (i--)
384                                         safe_put_page(bio->bi_io_vec[i].bv_page);
385                         }
386                         /* clear the bitmap if all writes complete successfully */
387                         bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
388                                         r1_bio->sectors,
389                                         !test_bit(R1BIO_Degraded, &r1_bio->state),
390                                         behind);
391                         md_write_end(r1_bio->mddev);
392                         raid_end_bio_io(r1_bio);
393                 }
394         }
395
396         if (to_put)
397                 bio_put(to_put);
398 }
399
400
401 /*
402  * This routine returns the disk from which the requested read should
403  * be done. There is a per-array 'next expected sequential IO' sector
404  * number - if this matches on the next IO then we use the last disk.
405  * There is also a per-disk 'last know head position' sector that is
406  * maintained from IRQ contexts, both the normal and the resync IO
407  * completion handlers update this position correctly. If there is no
408  * perfect sequential match then we pick the disk whose head is closest.
409  *
410  * If there are 2 mirrors in the same 2 devices, performance degrades
411  * because position is mirror, not device based.
412  *
413  * The rdev for the device selected will have nr_pending incremented.
414  */
415 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
416 {
417         const unsigned long this_sector = r1_bio->sector;
418         int new_disk = conf->last_used, disk = new_disk;
419         int wonly_disk = -1;
420         const int sectors = r1_bio->sectors;
421         sector_t new_distance, current_distance;
422         mdk_rdev_t *rdev;
423
424         rcu_read_lock();
425         /*
426          * Check if we can balance. We can balance on the whole
427          * device if no resync is going on, or below the resync window.
428          * We take the first readable disk when above the resync window.
429          */
430  retry:
431         if (conf->mddev->recovery_cp < MaxSector &&
432             (this_sector + sectors >= conf->next_resync)) {
433                 /* Choose the first operation device, for consistancy */
434                 new_disk = 0;
435
436                 for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
437                      r1_bio->bios[new_disk] == IO_BLOCKED ||
438                      !rdev || !test_bit(In_sync, &rdev->flags)
439                              || test_bit(WriteMostly, &rdev->flags);
440                      rdev = rcu_dereference(conf->mirrors[++new_disk].rdev)) {
441
442                         if (rdev && test_bit(In_sync, &rdev->flags) &&
443                                 r1_bio->bios[new_disk] != IO_BLOCKED)
444                                 wonly_disk = new_disk;
445
446                         if (new_disk == conf->raid_disks - 1) {
447                                 new_disk = wonly_disk;
448                                 break;
449                         }
450                 }
451                 goto rb_out;
452         }
453
454
455         /* make sure the disk is operational */
456         for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
457              r1_bio->bios[new_disk] == IO_BLOCKED ||
458              !rdev || !test_bit(In_sync, &rdev->flags) ||
459                      test_bit(WriteMostly, &rdev->flags);
460              rdev = rcu_dereference(conf->mirrors[new_disk].rdev)) {
461
462                 if (rdev && test_bit(In_sync, &rdev->flags) &&
463                     r1_bio->bios[new_disk] != IO_BLOCKED)
464                         wonly_disk = new_disk;
465
466                 if (new_disk <= 0)
467                         new_disk = conf->raid_disks;
468                 new_disk--;
469                 if (new_disk == disk) {
470                         new_disk = wonly_disk;
471                         break;
472                 }
473         }
474
475         if (new_disk < 0)
476                 goto rb_out;
477
478         disk = new_disk;
479         /* now disk == new_disk == starting point for search */
480
481         /*
482          * Don't change to another disk for sequential reads:
483          */
484         if (conf->next_seq_sect == this_sector)
485                 goto rb_out;
486         if (this_sector == conf->mirrors[new_disk].head_position)
487                 goto rb_out;
488
489         current_distance = abs(this_sector - conf->mirrors[disk].head_position);
490
491         /* Find the disk whose head is closest */
492
493         do {
494                 if (disk <= 0)
495                         disk = conf->raid_disks;
496                 disk--;
497
498                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
499
500                 if (!rdev || r1_bio->bios[disk] == IO_BLOCKED ||
501                     !test_bit(In_sync, &rdev->flags) ||
502                     test_bit(WriteMostly, &rdev->flags))
503                         continue;
504
505                 if (!atomic_read(&rdev->nr_pending)) {
506                         new_disk = disk;
507                         break;
508                 }
509                 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
510                 if (new_distance < current_distance) {
511                         current_distance = new_distance;
512                         new_disk = disk;
513                 }
514         } while (disk != conf->last_used);
515
516  rb_out:
517
518
519         if (new_disk >= 0) {
520                 rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
521                 if (!rdev)
522                         goto retry;
523                 atomic_inc(&rdev->nr_pending);
524                 if (!test_bit(In_sync, &rdev->flags)) {
525                         /* cannot risk returning a device that failed
526                          * before we inc'ed nr_pending
527                          */
528                         rdev_dec_pending(rdev, conf->mddev);
529                         goto retry;
530                 }
531                 conf->next_seq_sect = this_sector + sectors;
532                 conf->last_used = new_disk;
533         }
534         rcu_read_unlock();
535
536         return new_disk;
537 }
538
539 static void unplug_slaves(mddev_t *mddev)
540 {
541         conf_t *conf = mddev_to_conf(mddev);
542         int i;
543
544         rcu_read_lock();
545         for (i=0; i<mddev->raid_disks; i++) {
546                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
547                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
548                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
549
550                         atomic_inc(&rdev->nr_pending);
551                         rcu_read_unlock();
552
553                         blk_unplug(r_queue);
554
555                         rdev_dec_pending(rdev, mddev);
556                         rcu_read_lock();
557                 }
558         }
559         rcu_read_unlock();
560 }
561
562 static void raid1_unplug(struct request_queue *q)
563 {
564         mddev_t *mddev = q->queuedata;
565
566         unplug_slaves(mddev);
567         md_wakeup_thread(mddev->thread);
568 }
569
570 static int raid1_congested(void *data, int bits)
571 {
572         mddev_t *mddev = data;
573         conf_t *conf = mddev_to_conf(mddev);
574         int i, ret = 0;
575
576         rcu_read_lock();
577         for (i = 0; i < mddev->raid_disks; i++) {
578                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
579                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
580                         struct request_queue *q = bdev_get_queue(rdev->bdev);
581
582                         /* Note the '|| 1' - when read_balance prefers
583                          * non-congested targets, it can be removed
584                          */
585                         if ((bits & (1<<BDI_write_congested)) || 1)
586                                 ret |= bdi_congested(&q->backing_dev_info, bits);
587                         else
588                                 ret &= bdi_congested(&q->backing_dev_info, bits);
589                 }
590         }
591         rcu_read_unlock();
592         return ret;
593 }
594
595
596 static int flush_pending_writes(conf_t *conf)
597 {
598         /* Any writes that have been queued but are awaiting
599          * bitmap updates get flushed here.
600          * We return 1 if any requests were actually submitted.
601          */
602         int rv = 0;
603
604         spin_lock_irq(&conf->device_lock);
605
606         if (conf->pending_bio_list.head) {
607                 struct bio *bio;
608                 bio = bio_list_get(&conf->pending_bio_list);
609                 blk_remove_plug(conf->mddev->queue);
610                 spin_unlock_irq(&conf->device_lock);
611                 /* flush any pending bitmap writes to
612                  * disk before proceeding w/ I/O */
613                 bitmap_unplug(conf->mddev->bitmap);
614
615                 while (bio) { /* submit pending writes */
616                         struct bio *next = bio->bi_next;
617                         bio->bi_next = NULL;
618                         generic_make_request(bio);
619                         bio = next;
620                 }
621                 rv = 1;
622         } else
623                 spin_unlock_irq(&conf->device_lock);
624         return rv;
625 }
626
627 /* Barriers....
628  * Sometimes we need to suspend IO while we do something else,
629  * either some resync/recovery, or reconfigure the array.
630  * To do this we raise a 'barrier'.
631  * The 'barrier' is a counter that can be raised multiple times
632  * to count how many activities are happening which preclude
633  * normal IO.
634  * We can only raise the barrier if there is no pending IO.
635  * i.e. if nr_pending == 0.
636  * We choose only to raise the barrier if no-one is waiting for the
637  * barrier to go down.  This means that as soon as an IO request
638  * is ready, no other operations which require a barrier will start
639  * until the IO request has had a chance.
640  *
641  * So: regular IO calls 'wait_barrier'.  When that returns there
642  *    is no backgroup IO happening,  It must arrange to call
643  *    allow_barrier when it has finished its IO.
644  * backgroup IO calls must call raise_barrier.  Once that returns
645  *    there is no normal IO happeing.  It must arrange to call
646  *    lower_barrier when the particular background IO completes.
647  */
648 #define RESYNC_DEPTH 32
649
650 static void raise_barrier(conf_t *conf)
651 {
652         spin_lock_irq(&conf->resync_lock);
653
654         /* Wait until no block IO is waiting */
655         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
656                             conf->resync_lock,
657                             raid1_unplug(conf->mddev->queue));
658
659         /* block any new IO from starting */
660         conf->barrier++;
661
662         /* No wait for all pending IO to complete */
663         wait_event_lock_irq(conf->wait_barrier,
664                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
665                             conf->resync_lock,
666                             raid1_unplug(conf->mddev->queue));
667
668         spin_unlock_irq(&conf->resync_lock);
669 }
670
671 static void lower_barrier(conf_t *conf)
672 {
673         unsigned long flags;
674         spin_lock_irqsave(&conf->resync_lock, flags);
675         conf->barrier--;
676         spin_unlock_irqrestore(&conf->resync_lock, flags);
677         wake_up(&conf->wait_barrier);
678 }
679
680 static void wait_barrier(conf_t *conf)
681 {
682         spin_lock_irq(&conf->resync_lock);
683         if (conf->barrier) {
684                 conf->nr_waiting++;
685                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
686                                     conf->resync_lock,
687                                     raid1_unplug(conf->mddev->queue));
688                 conf->nr_waiting--;
689         }
690         conf->nr_pending++;
691         spin_unlock_irq(&conf->resync_lock);
692 }
693
694 static void allow_barrier(conf_t *conf)
695 {
696         unsigned long flags;
697         spin_lock_irqsave(&conf->resync_lock, flags);
698         conf->nr_pending--;
699         spin_unlock_irqrestore(&conf->resync_lock, flags);
700         wake_up(&conf->wait_barrier);
701 }
702
703 static void freeze_array(conf_t *conf)
704 {
705         /* stop syncio and normal IO and wait for everything to
706          * go quite.
707          * We increment barrier and nr_waiting, and then
708          * wait until nr_pending match nr_queued+1
709          * This is called in the context of one normal IO request
710          * that has failed. Thus any sync request that might be pending
711          * will be blocked by nr_pending, and we need to wait for
712          * pending IO requests to complete or be queued for re-try.
713          * Thus the number queued (nr_queued) plus this request (1)
714          * must match the number of pending IOs (nr_pending) before
715          * we continue.
716          */
717         spin_lock_irq(&conf->resync_lock);
718         conf->barrier++;
719         conf->nr_waiting++;
720         wait_event_lock_irq(conf->wait_barrier,
721                             conf->nr_pending == conf->nr_queued+1,
722                             conf->resync_lock,
723                             ({ flush_pending_writes(conf);
724                                raid1_unplug(conf->mddev->queue); }));
725         spin_unlock_irq(&conf->resync_lock);
726 }
727 static void unfreeze_array(conf_t *conf)
728 {
729         /* reverse the effect of the freeze */
730         spin_lock_irq(&conf->resync_lock);
731         conf->barrier--;
732         conf->nr_waiting--;
733         wake_up(&conf->wait_barrier);
734         spin_unlock_irq(&conf->resync_lock);
735 }
736
737
738 /* duplicate the data pages for behind I/O */
739 static struct page **alloc_behind_pages(struct bio *bio)
740 {
741         int i;
742         struct bio_vec *bvec;
743         struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page *),
744                                         GFP_NOIO);
745         if (unlikely(!pages))
746                 goto do_sync_io;
747
748         bio_for_each_segment(bvec, bio, i) {
749                 pages[i] = alloc_page(GFP_NOIO);
750                 if (unlikely(!pages[i]))
751                         goto do_sync_io;
752                 memcpy(kmap(pages[i]) + bvec->bv_offset,
753                         kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
754                 kunmap(pages[i]);
755                 kunmap(bvec->bv_page);
756         }
757
758         return pages;
759
760 do_sync_io:
761         if (pages)
762                 for (i = 0; i < bio->bi_vcnt && pages[i]; i++)
763                         put_page(pages[i]);
764         kfree(pages);
765         PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
766         return NULL;
767 }
768
769 static int make_request(struct request_queue *q, struct bio * bio)
770 {
771         mddev_t *mddev = q->queuedata;
772         conf_t *conf = mddev_to_conf(mddev);
773         mirror_info_t *mirror;
774         r1bio_t *r1_bio;
775         struct bio *read_bio;
776         int i, targets = 0, disks;
777         struct bitmap *bitmap;
778         unsigned long flags;
779         struct bio_list bl;
780         struct page **behind_pages = NULL;
781         const int rw = bio_data_dir(bio);
782         const int do_sync = bio_sync(bio);
783         int cpu, do_barriers;
784         mdk_rdev_t *blocked_rdev;
785
786         /*
787          * Register the new request and wait if the reconstruction
788          * thread has put up a bar for new requests.
789          * Continue immediately if no resync is active currently.
790          * We test barriers_work *after* md_write_start as md_write_start
791          * may cause the first superblock write, and that will check out
792          * if barriers work.
793          */
794
795         md_write_start(mddev, bio); /* wait on superblock update early */
796
797         if (unlikely(!mddev->barriers_work && bio_barrier(bio))) {
798                 if (rw == WRITE)
799                         md_write_end(mddev);
800                 bio_endio(bio, -EOPNOTSUPP);
801                 return 0;
802         }
803
804         wait_barrier(conf);
805
806         bitmap = mddev->bitmap;
807
808         cpu = part_stat_lock();
809         part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
810         part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
811                       bio_sectors(bio));
812         part_stat_unlock();
813
814         /*
815          * make_request() can abort the operation when READA is being
816          * used and no empty request is available.
817          *
818          */
819         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
820
821         r1_bio->master_bio = bio;
822         r1_bio->sectors = bio->bi_size >> 9;
823         r1_bio->state = 0;
824         r1_bio->mddev = mddev;
825         r1_bio->sector = bio->bi_sector;
826
827         if (rw == READ) {
828                 /*
829                  * read balancing logic:
830                  */
831                 int rdisk = read_balance(conf, r1_bio);
832
833                 if (rdisk < 0) {
834                         /* couldn't find anywhere to read from */
835                         raid_end_bio_io(r1_bio);
836                         return 0;
837                 }
838                 mirror = conf->mirrors + rdisk;
839
840                 r1_bio->read_disk = rdisk;
841
842                 read_bio = bio_clone(bio, GFP_NOIO);
843
844                 r1_bio->bios[rdisk] = read_bio;
845
846                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
847                 read_bio->bi_bdev = mirror->rdev->bdev;
848                 read_bio->bi_end_io = raid1_end_read_request;
849                 read_bio->bi_rw = READ | do_sync;
850                 read_bio->bi_private = r1_bio;
851
852                 generic_make_request(read_bio);
853                 return 0;
854         }
855
856         /*
857          * WRITE:
858          */
859         /* first select target devices under spinlock and
860          * inc refcount on their rdev.  Record them by setting
861          * bios[x] to bio
862          */
863         disks = conf->raid_disks;
864 #if 0
865         { static int first=1;
866         if (first) printk("First Write sector %llu disks %d\n",
867                           (unsigned long long)r1_bio->sector, disks);
868         first = 0;
869         }
870 #endif
871  retry_write:
872         blocked_rdev = NULL;
873         rcu_read_lock();
874         for (i = 0;  i < disks; i++) {
875                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
876                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
877                         atomic_inc(&rdev->nr_pending);
878                         blocked_rdev = rdev;
879                         break;
880                 }
881                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
882                         atomic_inc(&rdev->nr_pending);
883                         if (test_bit(Faulty, &rdev->flags)) {
884                                 rdev_dec_pending(rdev, mddev);
885                                 r1_bio->bios[i] = NULL;
886                         } else
887                                 r1_bio->bios[i] = bio;
888                         targets++;
889                 } else
890                         r1_bio->bios[i] = NULL;
891         }
892         rcu_read_unlock();
893
894         if (unlikely(blocked_rdev)) {
895                 /* Wait for this device to become unblocked */
896                 int j;
897
898                 for (j = 0; j < i; j++)
899                         if (r1_bio->bios[j])
900                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
901
902                 allow_barrier(conf);
903                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
904                 wait_barrier(conf);
905                 goto retry_write;
906         }
907
908         BUG_ON(targets == 0); /* we never fail the last device */
909
910         if (targets < conf->raid_disks) {
911                 /* array is degraded, we will not clear the bitmap
912                  * on I/O completion (see raid1_end_write_request) */
913                 set_bit(R1BIO_Degraded, &r1_bio->state);
914         }
915
916         /* do behind I/O ? */
917         if (bitmap &&
918             atomic_read(&bitmap->behind_writes) < bitmap->max_write_behind &&
919             (behind_pages = alloc_behind_pages(bio)) != NULL)
920                 set_bit(R1BIO_BehindIO, &r1_bio->state);
921
922         atomic_set(&r1_bio->remaining, 0);
923         atomic_set(&r1_bio->behind_remaining, 0);
924
925         do_barriers = bio_barrier(bio);
926         if (do_barriers)
927                 set_bit(R1BIO_Barrier, &r1_bio->state);
928
929         bio_list_init(&bl);
930         for (i = 0; i < disks; i++) {
931                 struct bio *mbio;
932                 if (!r1_bio->bios[i])
933                         continue;
934
935                 mbio = bio_clone(bio, GFP_NOIO);
936                 r1_bio->bios[i] = mbio;
937
938                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
939                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
940                 mbio->bi_end_io = raid1_end_write_request;
941                 mbio->bi_rw = WRITE | do_barriers | do_sync;
942                 mbio->bi_private = r1_bio;
943
944                 if (behind_pages) {
945                         struct bio_vec *bvec;
946                         int j;
947
948                         /* Yes, I really want the '__' version so that
949                          * we clear any unused pointer in the io_vec, rather
950                          * than leave them unchanged.  This is important
951                          * because when we come to free the pages, we won't
952                          * know the originial bi_idx, so we just free
953                          * them all
954                          */
955                         __bio_for_each_segment(bvec, mbio, j, 0)
956                                 bvec->bv_page = behind_pages[j];
957                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
958                                 atomic_inc(&r1_bio->behind_remaining);
959                 }
960
961                 atomic_inc(&r1_bio->remaining);
962
963                 bio_list_add(&bl, mbio);
964         }
965         kfree(behind_pages); /* the behind pages are attached to the bios now */
966
967         bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
968                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
969         spin_lock_irqsave(&conf->device_lock, flags);
970         bio_list_merge(&conf->pending_bio_list, &bl);
971         bio_list_init(&bl);
972
973         blk_plug_device(mddev->queue);
974         spin_unlock_irqrestore(&conf->device_lock, flags);
975
976         /* In case raid1d snuck into freeze_array */
977         wake_up(&conf->wait_barrier);
978
979         if (do_sync)
980                 md_wakeup_thread(mddev->thread);
981 #if 0
982         while ((bio = bio_list_pop(&bl)) != NULL)
983                 generic_make_request(bio);
984 #endif
985
986         return 0;
987 }
988
989 static void status(struct seq_file *seq, mddev_t *mddev)
990 {
991         conf_t *conf = mddev_to_conf(mddev);
992         int i;
993
994         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
995                    conf->raid_disks - mddev->degraded);
996         rcu_read_lock();
997         for (i = 0; i < conf->raid_disks; i++) {
998                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
999                 seq_printf(seq, "%s",
1000                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1001         }
1002         rcu_read_unlock();
1003         seq_printf(seq, "]");
1004 }
1005
1006
1007 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1008 {
1009         char b[BDEVNAME_SIZE];
1010         conf_t *conf = mddev_to_conf(mddev);
1011
1012         /*
1013          * If it is not operational, then we have already marked it as dead
1014          * else if it is the last working disks, ignore the error, let the
1015          * next level up know.
1016          * else mark the drive as failed
1017          */
1018         if (test_bit(In_sync, &rdev->flags)
1019             && (conf->raid_disks - mddev->degraded) == 1) {
1020                 /*
1021                  * Don't fail the drive, act as though we were just a
1022                  * normal single drive.
1023                  * However don't try a recovery from this drive as
1024                  * it is very likely to fail.
1025                  */
1026                 mddev->recovery_disabled = 1;
1027                 return;
1028         }
1029         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1030                 unsigned long flags;
1031                 spin_lock_irqsave(&conf->device_lock, flags);
1032                 mddev->degraded++;
1033                 set_bit(Faulty, &rdev->flags);
1034                 spin_unlock_irqrestore(&conf->device_lock, flags);
1035                 /*
1036                  * if recovery is running, make sure it aborts.
1037                  */
1038                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1039         } else
1040                 set_bit(Faulty, &rdev->flags);
1041         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1042         printk(KERN_ALERT "raid1: Disk failure on %s, disabling device.\n"
1043                 "raid1: Operation continuing on %d devices.\n",
1044                 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1045 }
1046
1047 static void print_conf(conf_t *conf)
1048 {
1049         int i;
1050
1051         printk("RAID1 conf printout:\n");
1052         if (!conf) {
1053                 printk("(!conf)\n");
1054                 return;
1055         }
1056         printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1057                 conf->raid_disks);
1058
1059         rcu_read_lock();
1060         for (i = 0; i < conf->raid_disks; i++) {
1061                 char b[BDEVNAME_SIZE];
1062                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1063                 if (rdev)
1064                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
1065                                i, !test_bit(In_sync, &rdev->flags),
1066                                !test_bit(Faulty, &rdev->flags),
1067                                bdevname(rdev->bdev,b));
1068         }
1069         rcu_read_unlock();
1070 }
1071
1072 static void close_sync(conf_t *conf)
1073 {
1074         wait_barrier(conf);
1075         allow_barrier(conf);
1076
1077         mempool_destroy(conf->r1buf_pool);
1078         conf->r1buf_pool = NULL;
1079 }
1080
1081 static int raid1_spare_active(mddev_t *mddev)
1082 {
1083         int i;
1084         conf_t *conf = mddev->private;
1085
1086         /*
1087          * Find all failed disks within the RAID1 configuration 
1088          * and mark them readable.
1089          * Called under mddev lock, so rcu protection not needed.
1090          */
1091         for (i = 0; i < conf->raid_disks; i++) {
1092                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1093                 if (rdev
1094                     && !test_bit(Faulty, &rdev->flags)
1095                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1096                         unsigned long flags;
1097                         spin_lock_irqsave(&conf->device_lock, flags);
1098                         mddev->degraded--;
1099                         spin_unlock_irqrestore(&conf->device_lock, flags);
1100                 }
1101         }
1102
1103         print_conf(conf);
1104         return 0;
1105 }
1106
1107
1108 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1109 {
1110         conf_t *conf = mddev->private;
1111         int err = -EEXIST;
1112         int mirror = 0;
1113         mirror_info_t *p;
1114         int first = 0;
1115         int last = mddev->raid_disks - 1;
1116
1117         if (rdev->raid_disk >= 0)
1118                 first = last = rdev->raid_disk;
1119
1120         for (mirror = first; mirror <= last; mirror++)
1121                 if ( !(p=conf->mirrors+mirror)->rdev) {
1122
1123                         blk_queue_stack_limits(mddev->queue,
1124                                                rdev->bdev->bd_disk->queue);
1125                         /* as we don't honour merge_bvec_fn, we must never risk
1126                          * violating it, so limit ->max_sector to one PAGE, as
1127                          * a one page request is never in violation.
1128                          */
1129                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1130                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
1131                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1132
1133                         p->head_position = 0;
1134                         rdev->raid_disk = mirror;
1135                         err = 0;
1136                         /* As all devices are equivalent, we don't need a full recovery
1137                          * if this was recently any drive of the array
1138                          */
1139                         if (rdev->saved_raid_disk < 0)
1140                                 conf->fullsync = 1;
1141                         rcu_assign_pointer(p->rdev, rdev);
1142                         break;
1143                 }
1144
1145         print_conf(conf);
1146         return err;
1147 }
1148
1149 static int raid1_remove_disk(mddev_t *mddev, int number)
1150 {
1151         conf_t *conf = mddev->private;
1152         int err = 0;
1153         mdk_rdev_t *rdev;
1154         mirror_info_t *p = conf->mirrors+ number;
1155
1156         print_conf(conf);
1157         rdev = p->rdev;
1158         if (rdev) {
1159                 if (test_bit(In_sync, &rdev->flags) ||
1160                     atomic_read(&rdev->nr_pending)) {
1161                         err = -EBUSY;
1162                         goto abort;
1163                 }
1164                 /* Only remove non-faulty devices is recovery
1165                  * is not possible.
1166                  */
1167                 if (!test_bit(Faulty, &rdev->flags) &&
1168                     mddev->degraded < conf->raid_disks) {
1169                         err = -EBUSY;
1170                         goto abort;
1171                 }
1172                 p->rdev = NULL;
1173                 synchronize_rcu();
1174                 if (atomic_read(&rdev->nr_pending)) {
1175                         /* lost the race, try later */
1176                         err = -EBUSY;
1177                         p->rdev = rdev;
1178                 }
1179         }
1180 abort:
1181
1182         print_conf(conf);
1183         return err;
1184 }
1185
1186
1187 static void end_sync_read(struct bio *bio, int error)
1188 {
1189         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1190         int i;
1191
1192         for (i=r1_bio->mddev->raid_disks; i--; )
1193                 if (r1_bio->bios[i] == bio)
1194                         break;
1195         BUG_ON(i < 0);
1196         update_head_pos(i, r1_bio);
1197         /*
1198          * we have read a block, now it needs to be re-written,
1199          * or re-read if the read failed.
1200          * We don't do much here, just schedule handling by raid1d
1201          */
1202         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1203                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1204
1205         if (atomic_dec_and_test(&r1_bio->remaining))
1206                 reschedule_retry(r1_bio);
1207 }
1208
1209 static void end_sync_write(struct bio *bio, int error)
1210 {
1211         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1212         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1213         mddev_t *mddev = r1_bio->mddev;
1214         conf_t *conf = mddev_to_conf(mddev);
1215         int i;
1216         int mirror=0;
1217
1218         for (i = 0; i < conf->raid_disks; i++)
1219                 if (r1_bio->bios[i] == bio) {
1220                         mirror = i;
1221                         break;
1222                 }
1223         if (!uptodate) {
1224                 int sync_blocks = 0;
1225                 sector_t s = r1_bio->sector;
1226                 long sectors_to_go = r1_bio->sectors;
1227                 /* make sure these bits doesn't get cleared. */
1228                 do {
1229                         bitmap_end_sync(mddev->bitmap, s,
1230                                         &sync_blocks, 1);
1231                         s += sync_blocks;
1232                         sectors_to_go -= sync_blocks;
1233                 } while (sectors_to_go > 0);
1234                 md_error(mddev, conf->mirrors[mirror].rdev);
1235         }
1236
1237         update_head_pos(mirror, r1_bio);
1238
1239         if (atomic_dec_and_test(&r1_bio->remaining)) {
1240                 md_done_sync(mddev, r1_bio->sectors, uptodate);
1241                 put_buf(r1_bio);
1242         }
1243 }
1244
1245 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1246 {
1247         conf_t *conf = mddev_to_conf(mddev);
1248         int i;
1249         int disks = conf->raid_disks;
1250         struct bio *bio, *wbio;
1251
1252         bio = r1_bio->bios[r1_bio->read_disk];
1253
1254
1255         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1256                 /* We have read all readable devices.  If we haven't
1257                  * got the block, then there is no hope left.
1258                  * If we have, then we want to do a comparison
1259                  * and skip the write if everything is the same.
1260                  * If any blocks failed to read, then we need to
1261                  * attempt an over-write
1262                  */
1263                 int primary;
1264                 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1265                         for (i=0; i<mddev->raid_disks; i++)
1266                                 if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1267                                         md_error(mddev, conf->mirrors[i].rdev);
1268
1269                         md_done_sync(mddev, r1_bio->sectors, 1);
1270                         put_buf(r1_bio);
1271                         return;
1272                 }
1273                 for (primary=0; primary<mddev->raid_disks; primary++)
1274                         if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1275                             test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1276                                 r1_bio->bios[primary]->bi_end_io = NULL;
1277                                 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1278                                 break;
1279                         }
1280                 r1_bio->read_disk = primary;
1281                 for (i=0; i<mddev->raid_disks; i++)
1282                         if (r1_bio->bios[i]->bi_end_io == end_sync_read) {
1283                                 int j;
1284                                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1285                                 struct bio *pbio = r1_bio->bios[primary];
1286                                 struct bio *sbio = r1_bio->bios[i];
1287
1288                                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1289                                         for (j = vcnt; j-- ; ) {
1290                                                 struct page *p, *s;
1291                                                 p = pbio->bi_io_vec[j].bv_page;
1292                                                 s = sbio->bi_io_vec[j].bv_page;
1293                                                 if (memcmp(page_address(p),
1294                                                            page_address(s),
1295                                                            PAGE_SIZE))
1296                                                         break;
1297                                         }
1298                                 } else
1299                                         j = 0;
1300                                 if (j >= 0)
1301                                         mddev->resync_mismatches += r1_bio->sectors;
1302                                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1303                                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1304                                         sbio->bi_end_io = NULL;
1305                                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1306                                 } else {
1307                                         /* fixup the bio for reuse */
1308                                         int size;
1309                                         sbio->bi_vcnt = vcnt;
1310                                         sbio->bi_size = r1_bio->sectors << 9;
1311                                         sbio->bi_idx = 0;
1312                                         sbio->bi_phys_segments = 0;
1313                                         sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1314                                         sbio->bi_flags |= 1 << BIO_UPTODATE;
1315                                         sbio->bi_next = NULL;
1316                                         sbio->bi_sector = r1_bio->sector +
1317                                                 conf->mirrors[i].rdev->data_offset;
1318                                         sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1319                                         size = sbio->bi_size;
1320                                         for (j = 0; j < vcnt ; j++) {
1321                                                 struct bio_vec *bi;
1322                                                 bi = &sbio->bi_io_vec[j];
1323                                                 bi->bv_offset = 0;
1324                                                 if (size > PAGE_SIZE)
1325                                                         bi->bv_len = PAGE_SIZE;
1326                                                 else
1327                                                         bi->bv_len = size;
1328                                                 size -= PAGE_SIZE;
1329                                                 memcpy(page_address(bi->bv_page),
1330                                                        page_address(pbio->bi_io_vec[j].bv_page),
1331                                                        PAGE_SIZE);
1332                                         }
1333
1334                                 }
1335                         }
1336         }
1337         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1338                 /* ouch - failed to read all of that.
1339                  * Try some synchronous reads of other devices to get
1340                  * good data, much like with normal read errors.  Only
1341                  * read into the pages we already have so we don't
1342                  * need to re-issue the read request.
1343                  * We don't need to freeze the array, because being in an
1344                  * active sync request, there is no normal IO, and
1345                  * no overlapping syncs.
1346                  */
1347                 sector_t sect = r1_bio->sector;
1348                 int sectors = r1_bio->sectors;
1349                 int idx = 0;
1350
1351                 while(sectors) {
1352                         int s = sectors;
1353                         int d = r1_bio->read_disk;
1354                         int success = 0;
1355                         mdk_rdev_t *rdev;
1356
1357                         if (s > (PAGE_SIZE>>9))
1358                                 s = PAGE_SIZE >> 9;
1359                         do {
1360                                 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1361                                         /* No rcu protection needed here devices
1362                                          * can only be removed when no resync is
1363                                          * active, and resync is currently active
1364                                          */
1365                                         rdev = conf->mirrors[d].rdev;
1366                                         if (sync_page_io(rdev->bdev,
1367                                                          sect + rdev->data_offset,
1368                                                          s<<9,
1369                                                          bio->bi_io_vec[idx].bv_page,
1370                                                          READ)) {
1371                                                 success = 1;
1372                                                 break;
1373                                         }
1374                                 }
1375                                 d++;
1376                                 if (d == conf->raid_disks)
1377                                         d = 0;
1378                         } while (!success && d != r1_bio->read_disk);
1379
1380                         if (success) {
1381                                 int start = d;
1382                                 /* write it back and re-read */
1383                                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1384                                 while (d != r1_bio->read_disk) {
1385                                         if (d == 0)
1386                                                 d = conf->raid_disks;
1387                                         d--;
1388                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1389                                                 continue;
1390                                         rdev = conf->mirrors[d].rdev;
1391                                         atomic_add(s, &rdev->corrected_errors);
1392                                         if (sync_page_io(rdev->bdev,
1393                                                          sect + rdev->data_offset,
1394                                                          s<<9,
1395                                                          bio->bi_io_vec[idx].bv_page,
1396                                                          WRITE) == 0)
1397                                                 md_error(mddev, rdev);
1398                                 }
1399                                 d = start;
1400                                 while (d != r1_bio->read_disk) {
1401                                         if (d == 0)
1402                                                 d = conf->raid_disks;
1403                                         d--;
1404                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1405                                                 continue;
1406                                         rdev = conf->mirrors[d].rdev;
1407                                         if (sync_page_io(rdev->bdev,
1408                                                          sect + rdev->data_offset,
1409                                                          s<<9,
1410                                                          bio->bi_io_vec[idx].bv_page,
1411                                                          READ) == 0)
1412                                                 md_error(mddev, rdev);
1413                                 }
1414                         } else {
1415                                 char b[BDEVNAME_SIZE];
1416                                 /* Cannot read from anywhere, array is toast */
1417                                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1418                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
1419                                        " for block %llu\n",
1420                                        bdevname(bio->bi_bdev,b),
1421                                        (unsigned long long)r1_bio->sector);
1422                                 md_done_sync(mddev, r1_bio->sectors, 0);
1423                                 put_buf(r1_bio);
1424                                 return;
1425                         }
1426                         sectors -= s;
1427                         sect += s;
1428                         idx ++;
1429                 }
1430         }
1431
1432         /*
1433          * schedule writes
1434          */
1435         atomic_set(&r1_bio->remaining, 1);
1436         for (i = 0; i < disks ; i++) {
1437                 wbio = r1_bio->bios[i];
1438                 if (wbio->bi_end_io == NULL ||
1439                     (wbio->bi_end_io == end_sync_read &&
1440                      (i == r1_bio->read_disk ||
1441                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1442                         continue;
1443
1444                 wbio->bi_rw = WRITE;
1445                 wbio->bi_end_io = end_sync_write;
1446                 atomic_inc(&r1_bio->remaining);
1447                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1448
1449                 generic_make_request(wbio);
1450         }
1451
1452         if (atomic_dec_and_test(&r1_bio->remaining)) {
1453                 /* if we're here, all write(s) have completed, so clean up */
1454                 md_done_sync(mddev, r1_bio->sectors, 1);
1455                 put_buf(r1_bio);
1456         }
1457 }
1458
1459 /*
1460  * This is a kernel thread which:
1461  *
1462  *      1.      Retries failed read operations on working mirrors.
1463  *      2.      Updates the raid superblock when problems encounter.
1464  *      3.      Performs writes following reads for array syncronising.
1465  */
1466
1467 static void fix_read_error(conf_t *conf, int read_disk,
1468                            sector_t sect, int sectors)
1469 {
1470         mddev_t *mddev = conf->mddev;
1471         while(sectors) {
1472                 int s = sectors;
1473                 int d = read_disk;
1474                 int success = 0;
1475                 int start;
1476                 mdk_rdev_t *rdev;
1477
1478                 if (s > (PAGE_SIZE>>9))
1479                         s = PAGE_SIZE >> 9;
1480
1481                 do {
1482                         /* Note: no rcu protection needed here
1483                          * as this is synchronous in the raid1d thread
1484                          * which is the thread that might remove
1485                          * a device.  If raid1d ever becomes multi-threaded....
1486                          */
1487                         rdev = conf->mirrors[d].rdev;
1488                         if (rdev &&
1489                             test_bit(In_sync, &rdev->flags) &&
1490                             sync_page_io(rdev->bdev,
1491                                          sect + rdev->data_offset,
1492                                          s<<9,
1493                                          conf->tmppage, READ))
1494                                 success = 1;
1495                         else {
1496                                 d++;
1497                                 if (d == conf->raid_disks)
1498                                         d = 0;
1499                         }
1500                 } while (!success && d != read_disk);
1501
1502                 if (!success) {
1503                         /* Cannot read from anywhere -- bye bye array */
1504                         md_error(mddev, conf->mirrors[read_disk].rdev);
1505                         break;
1506                 }
1507                 /* write it back and re-read */
1508                 start = d;
1509                 while (d != read_disk) {
1510                         if (d==0)
1511                                 d = conf->raid_disks;
1512                         d--;
1513                         rdev = conf->mirrors[d].rdev;
1514                         if (rdev &&
1515                             test_bit(In_sync, &rdev->flags)) {
1516                                 if (sync_page_io(rdev->bdev,
1517                                                  sect + rdev->data_offset,
1518                                                  s<<9, conf->tmppage, WRITE)
1519                                     == 0)
1520                                         /* Well, this device is dead */
1521                                         md_error(mddev, rdev);
1522                         }
1523                 }
1524                 d = start;
1525                 while (d != read_disk) {
1526                         char b[BDEVNAME_SIZE];
1527                         if (d==0)
1528                                 d = conf->raid_disks;
1529                         d--;
1530                         rdev = conf->mirrors[d].rdev;
1531                         if (rdev &&
1532                             test_bit(In_sync, &rdev->flags)) {
1533                                 if (sync_page_io(rdev->bdev,
1534                                                  sect + rdev->data_offset,
1535                                                  s<<9, conf->tmppage, READ)
1536                                     == 0)
1537                                         /* Well, this device is dead */
1538                                         md_error(mddev, rdev);
1539                                 else {
1540                                         atomic_add(s, &rdev->corrected_errors);
1541                                         printk(KERN_INFO
1542                                                "raid1:%s: read error corrected "
1543                                                "(%d sectors at %llu on %s)\n",
1544                                                mdname(mddev), s,
1545                                                (unsigned long long)(sect +
1546                                                    rdev->data_offset),
1547                                                bdevname(rdev->bdev, b));
1548                                 }
1549                         }
1550                 }
1551                 sectors -= s;
1552                 sect += s;
1553         }
1554 }
1555
1556 static void raid1d(mddev_t *mddev)
1557 {
1558         r1bio_t *r1_bio;
1559         struct bio *bio;
1560         unsigned long flags;
1561         conf_t *conf = mddev_to_conf(mddev);
1562         struct list_head *head = &conf->retry_list;
1563         int unplug=0;
1564         mdk_rdev_t *rdev;
1565
1566         md_check_recovery(mddev);
1567         
1568         for (;;) {
1569                 char b[BDEVNAME_SIZE];
1570
1571                 unplug += flush_pending_writes(conf);
1572
1573                 spin_lock_irqsave(&conf->device_lock, flags);
1574                 if (list_empty(head)) {
1575                         spin_unlock_irqrestore(&conf->device_lock, flags);
1576                         break;
1577                 }
1578                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1579                 list_del(head->prev);
1580                 conf->nr_queued--;
1581                 spin_unlock_irqrestore(&conf->device_lock, flags);
1582
1583                 mddev = r1_bio->mddev;
1584                 conf = mddev_to_conf(mddev);
1585                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1586                         sync_request_write(mddev, r1_bio);
1587                         unplug = 1;
1588                 } else if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
1589                         /* some requests in the r1bio were BIO_RW_BARRIER
1590                          * requests which failed with -EOPNOTSUPP.  Hohumm..
1591                          * Better resubmit without the barrier.
1592                          * We know which devices to resubmit for, because
1593                          * all others have had their bios[] entry cleared.
1594                          * We already have a nr_pending reference on these rdevs.
1595                          */
1596                         int i;
1597                         const int do_sync = bio_sync(r1_bio->master_bio);
1598                         clear_bit(R1BIO_BarrierRetry, &r1_bio->state);
1599                         clear_bit(R1BIO_Barrier, &r1_bio->state);
1600                         for (i=0; i < conf->raid_disks; i++)
1601                                 if (r1_bio->bios[i])
1602                                         atomic_inc(&r1_bio->remaining);
1603                         for (i=0; i < conf->raid_disks; i++)
1604                                 if (r1_bio->bios[i]) {
1605                                         struct bio_vec *bvec;
1606                                         int j;
1607
1608                                         bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1609                                         /* copy pages from the failed bio, as
1610                                          * this might be a write-behind device */
1611                                         __bio_for_each_segment(bvec, bio, j, 0)
1612                                                 bvec->bv_page = bio_iovec_idx(r1_bio->bios[i], j)->bv_page;
1613                                         bio_put(r1_bio->bios[i]);
1614                                         bio->bi_sector = r1_bio->sector +
1615                                                 conf->mirrors[i].rdev->data_offset;
1616                                         bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1617                                         bio->bi_end_io = raid1_end_write_request;
1618                                         bio->bi_rw = WRITE | do_sync;
1619                                         bio->bi_private = r1_bio;
1620                                         r1_bio->bios[i] = bio;
1621                                         generic_make_request(bio);
1622                                 }
1623                 } else {
1624                         int disk;
1625
1626                         /* we got a read error. Maybe the drive is bad.  Maybe just
1627                          * the block and we can fix it.
1628                          * We freeze all other IO, and try reading the block from
1629                          * other devices.  When we find one, we re-write
1630                          * and check it that fixes the read error.
1631                          * This is all done synchronously while the array is
1632                          * frozen
1633                          */
1634                         if (mddev->ro == 0) {
1635                                 freeze_array(conf);
1636                                 fix_read_error(conf, r1_bio->read_disk,
1637                                                r1_bio->sector,
1638                                                r1_bio->sectors);
1639                                 unfreeze_array(conf);
1640                         }
1641
1642                         bio = r1_bio->bios[r1_bio->read_disk];
1643                         if ((disk=read_balance(conf, r1_bio)) == -1 ||
1644                             disk == r1_bio->read_disk) {
1645                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
1646                                        " read error for block %llu\n",
1647                                        bdevname(bio->bi_bdev,b),
1648                                        (unsigned long long)r1_bio->sector);
1649                                 raid_end_bio_io(r1_bio);
1650                         } else {
1651                                 const int do_sync = bio_sync(r1_bio->master_bio);
1652                                 r1_bio->bios[r1_bio->read_disk] =
1653                                         mddev->ro ? IO_BLOCKED : NULL;
1654                                 r1_bio->read_disk = disk;
1655                                 bio_put(bio);
1656                                 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1657                                 r1_bio->bios[r1_bio->read_disk] = bio;
1658                                 rdev = conf->mirrors[disk].rdev;
1659                                 if (printk_ratelimit())
1660                                         printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
1661                                                " another mirror\n",
1662                                                bdevname(rdev->bdev,b),
1663                                                (unsigned long long)r1_bio->sector);
1664                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1665                                 bio->bi_bdev = rdev->bdev;
1666                                 bio->bi_end_io = raid1_end_read_request;
1667                                 bio->bi_rw = READ | do_sync;
1668                                 bio->bi_private = r1_bio;
1669                                 unplug = 1;
1670                                 generic_make_request(bio);
1671                         }
1672                 }
1673         }
1674         if (unplug)
1675                 unplug_slaves(mddev);
1676 }
1677
1678
1679 static int init_resync(conf_t *conf)
1680 {
1681         int buffs;
1682
1683         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1684         BUG_ON(conf->r1buf_pool);
1685         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1686                                           conf->poolinfo);
1687         if (!conf->r1buf_pool)
1688                 return -ENOMEM;
1689         conf->next_resync = 0;
1690         return 0;
1691 }
1692
1693 /*
1694  * perform a "sync" on one "block"
1695  *
1696  * We need to make sure that no normal I/O request - particularly write
1697  * requests - conflict with active sync requests.
1698  *
1699  * This is achieved by tracking pending requests and a 'barrier' concept
1700  * that can be installed to exclude normal IO requests.
1701  */
1702
1703 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1704 {
1705         conf_t *conf = mddev_to_conf(mddev);
1706         r1bio_t *r1_bio;
1707         struct bio *bio;
1708         sector_t max_sector, nr_sectors;
1709         int disk = -1;
1710         int i;
1711         int wonly = -1;
1712         int write_targets = 0, read_targets = 0;
1713         int sync_blocks;
1714         int still_degraded = 0;
1715
1716         if (!conf->r1buf_pool)
1717         {
1718 /*
1719                 printk("sync start - bitmap %p\n", mddev->bitmap);
1720 */
1721                 if (init_resync(conf))
1722                         return 0;
1723         }
1724
1725         max_sector = mddev->size << 1;
1726         if (sector_nr >= max_sector) {
1727                 /* If we aborted, we need to abort the
1728                  * sync on the 'current' bitmap chunk (there will
1729                  * only be one in raid1 resync.
1730                  * We can find the current addess in mddev->curr_resync
1731                  */
1732                 if (mddev->curr_resync < max_sector) /* aborted */
1733                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1734                                                 &sync_blocks, 1);
1735                 else /* completed sync */
1736                         conf->fullsync = 0;
1737
1738                 bitmap_close_sync(mddev->bitmap);
1739                 close_sync(conf);
1740                 return 0;
1741         }
1742
1743         if (mddev->bitmap == NULL &&
1744             mddev->recovery_cp == MaxSector &&
1745             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1746             conf->fullsync == 0) {
1747                 *skipped = 1;
1748                 return max_sector - sector_nr;
1749         }
1750         /* before building a request, check if we can skip these blocks..
1751          * This call the bitmap_start_sync doesn't actually record anything
1752          */
1753         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1754             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1755                 /* We can skip this block, and probably several more */
1756                 *skipped = 1;
1757                 return sync_blocks;
1758         }
1759         /*
1760          * If there is non-resync activity waiting for a turn,
1761          * and resync is going fast enough,
1762          * then let it though before starting on this new sync request.
1763          */
1764         if (!go_faster && conf->nr_waiting)
1765                 msleep_interruptible(1000);
1766
1767         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1768         raise_barrier(conf);
1769
1770         conf->next_resync = sector_nr;
1771
1772         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1773         rcu_read_lock();
1774         /*
1775          * If we get a correctably read error during resync or recovery,
1776          * we might want to read from a different device.  So we
1777          * flag all drives that could conceivably be read from for READ,
1778          * and any others (which will be non-In_sync devices) for WRITE.
1779          * If a read fails, we try reading from something else for which READ
1780          * is OK.
1781          */
1782
1783         r1_bio->mddev = mddev;
1784         r1_bio->sector = sector_nr;
1785         r1_bio->state = 0;
1786         set_bit(R1BIO_IsSync, &r1_bio->state);
1787
1788         for (i=0; i < conf->raid_disks; i++) {
1789                 mdk_rdev_t *rdev;
1790                 bio = r1_bio->bios[i];
1791
1792                 /* take from bio_init */
1793                 bio->bi_next = NULL;
1794                 bio->bi_flags |= 1 << BIO_UPTODATE;
1795                 bio->bi_rw = READ;
1796                 bio->bi_vcnt = 0;
1797                 bio->bi_idx = 0;
1798                 bio->bi_phys_segments = 0;
1799                 bio->bi_size = 0;
1800                 bio->bi_end_io = NULL;
1801                 bio->bi_private = NULL;
1802
1803                 rdev = rcu_dereference(conf->mirrors[i].rdev);
1804                 if (rdev == NULL ||
1805                            test_bit(Faulty, &rdev->flags)) {
1806                         still_degraded = 1;
1807                         continue;
1808                 } else if (!test_bit(In_sync, &rdev->flags)) {
1809                         bio->bi_rw = WRITE;
1810                         bio->bi_end_io = end_sync_write;
1811                         write_targets ++;
1812                 } else {
1813                         /* may need to read from here */
1814                         bio->bi_rw = READ;
1815                         bio->bi_end_io = end_sync_read;
1816                         if (test_bit(WriteMostly, &rdev->flags)) {
1817                                 if (wonly < 0)
1818                                         wonly = i;
1819                         } else {
1820                                 if (disk < 0)
1821                                         disk = i;
1822                         }
1823                         read_targets++;
1824                 }
1825                 atomic_inc(&rdev->nr_pending);
1826                 bio->bi_sector = sector_nr + rdev->data_offset;
1827                 bio->bi_bdev = rdev->bdev;
1828                 bio->bi_private = r1_bio;
1829         }
1830         rcu_read_unlock();
1831         if (disk < 0)
1832                 disk = wonly;
1833         r1_bio->read_disk = disk;
1834
1835         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1836                 /* extra read targets are also write targets */
1837                 write_targets += read_targets-1;
1838
1839         if (write_targets == 0 || read_targets == 0) {
1840                 /* There is nowhere to write, so all non-sync
1841                  * drives must be failed - so we are finished
1842                  */
1843                 sector_t rv = max_sector - sector_nr;
1844                 *skipped = 1;
1845                 put_buf(r1_bio);
1846                 return rv;
1847         }
1848
1849         if (max_sector > mddev->resync_max)
1850                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1851         nr_sectors = 0;
1852         sync_blocks = 0;
1853         do {
1854                 struct page *page;
1855                 int len = PAGE_SIZE;
1856                 if (sector_nr + (len>>9) > max_sector)
1857                         len = (max_sector - sector_nr) << 9;
1858                 if (len == 0)
1859                         break;
1860                 if (sync_blocks == 0) {
1861                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1862                                                &sync_blocks, still_degraded) &&
1863                             !conf->fullsync &&
1864                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1865                                 break;
1866                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1867                         if (len > (sync_blocks<<9))
1868                                 len = sync_blocks<<9;
1869                 }
1870
1871                 for (i=0 ; i < conf->raid_disks; i++) {
1872                         bio = r1_bio->bios[i];
1873                         if (bio->bi_end_io) {
1874                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1875                                 if (bio_add_page(bio, page, len, 0) == 0) {
1876                                         /* stop here */
1877                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1878                                         while (i > 0) {
1879                                                 i--;
1880                                                 bio = r1_bio->bios[i];
1881                                                 if (bio->bi_end_io==NULL)
1882                                                         continue;
1883                                                 /* remove last page from this bio */
1884                                                 bio->bi_vcnt--;
1885                                                 bio->bi_size -= len;
1886                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1887                                         }
1888                                         goto bio_full;
1889                                 }
1890                         }
1891                 }
1892                 nr_sectors += len>>9;
1893                 sector_nr += len>>9;
1894                 sync_blocks -= (len>>9);
1895         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1896  bio_full:
1897         r1_bio->sectors = nr_sectors;
1898
1899         /* For a user-requested sync, we read all readable devices and do a
1900          * compare
1901          */
1902         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1903                 atomic_set(&r1_bio->remaining, read_targets);
1904                 for (i=0; i<conf->raid_disks; i++) {
1905                         bio = r1_bio->bios[i];
1906                         if (bio->bi_end_io == end_sync_read) {
1907                                 md_sync_acct(bio->bi_bdev, nr_sectors);
1908                                 generic_make_request(bio);
1909                         }
1910                 }
1911         } else {
1912                 atomic_set(&r1_bio->remaining, 1);
1913                 bio = r1_bio->bios[r1_bio->read_disk];
1914                 md_sync_acct(bio->bi_bdev, nr_sectors);
1915                 generic_make_request(bio);
1916
1917         }
1918         return nr_sectors;
1919 }
1920
1921 static int run(mddev_t *mddev)
1922 {
1923         conf_t *conf;
1924         int i, j, disk_idx;
1925         mirror_info_t *disk;
1926         mdk_rdev_t *rdev;
1927
1928         if (mddev->level != 1) {
1929                 printk("raid1: %s: raid level not set to mirroring (%d)\n",
1930                        mdname(mddev), mddev->level);
1931                 goto out;
1932         }
1933         if (mddev->reshape_position != MaxSector) {
1934                 printk("raid1: %s: reshape_position set but not supported\n",
1935                        mdname(mddev));
1936                 goto out;
1937         }
1938         /*
1939          * copy the already verified devices into our private RAID1
1940          * bookkeeping area. [whatever we allocate in run(),
1941          * should be freed in stop()]
1942          */
1943         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1944         mddev->private = conf;
1945         if (!conf)
1946                 goto out_no_mem;
1947
1948         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1949                                  GFP_KERNEL);
1950         if (!conf->mirrors)
1951                 goto out_no_mem;
1952
1953         conf->tmppage = alloc_page(GFP_KERNEL);
1954         if (!conf->tmppage)
1955                 goto out_no_mem;
1956
1957         conf->poolinfo = kmalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1958         if (!conf->poolinfo)
1959                 goto out_no_mem;
1960         conf->poolinfo->mddev = mddev;
1961         conf->poolinfo->raid_disks = mddev->raid_disks;
1962         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1963                                           r1bio_pool_free,
1964                                           conf->poolinfo);
1965         if (!conf->r1bio_pool)
1966                 goto out_no_mem;
1967
1968         spin_lock_init(&conf->device_lock);
1969         mddev->queue->queue_lock = &conf->device_lock;
1970
1971         list_for_each_entry(rdev, &mddev->disks, same_set) {
1972                 disk_idx = rdev->raid_disk;
1973                 if (disk_idx >= mddev->raid_disks
1974                     || disk_idx < 0)
1975                         continue;
1976                 disk = conf->mirrors + disk_idx;
1977
1978                 disk->rdev = rdev;
1979
1980                 blk_queue_stack_limits(mddev->queue,
1981                                        rdev->bdev->bd_disk->queue);
1982                 /* as we don't honour merge_bvec_fn, we must never risk
1983                  * violating it, so limit ->max_sector to one PAGE, as
1984                  * a one page request is never in violation.
1985                  */
1986                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1987                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
1988                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1989
1990                 disk->head_position = 0;
1991         }
1992         conf->raid_disks = mddev->raid_disks;
1993         conf->mddev = mddev;
1994         INIT_LIST_HEAD(&conf->retry_list);
1995
1996         spin_lock_init(&conf->resync_lock);
1997         init_waitqueue_head(&conf->wait_barrier);
1998
1999         bio_list_init(&conf->pending_bio_list);
2000         bio_list_init(&conf->flushing_bio_list);
2001
2002
2003         mddev->degraded = 0;
2004         for (i = 0; i < conf->raid_disks; i++) {
2005
2006                 disk = conf->mirrors + i;
2007
2008                 if (!disk->rdev ||
2009                     !test_bit(In_sync, &disk->rdev->flags)) {
2010                         disk->head_position = 0;
2011                         mddev->degraded++;
2012                         if (disk->rdev)
2013                                 conf->fullsync = 1;
2014                 }
2015         }
2016         if (mddev->degraded == conf->raid_disks) {
2017                 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
2018                         mdname(mddev));
2019                 goto out_free_conf;
2020         }
2021         if (conf->raid_disks - mddev->degraded == 1)
2022                 mddev->recovery_cp = MaxSector;
2023
2024         /*
2025          * find the first working one and use it as a starting point
2026          * to read balancing.
2027          */
2028         for (j = 0; j < conf->raid_disks &&
2029                      (!conf->mirrors[j].rdev ||
2030                       !test_bit(In_sync, &conf->mirrors[j].rdev->flags)) ; j++)
2031                 /* nothing */;
2032         conf->last_used = j;
2033
2034
2035         mddev->thread = md_register_thread(raid1d, mddev, "%s_raid1");
2036         if (!mddev->thread) {
2037                 printk(KERN_ERR
2038                        "raid1: couldn't allocate thread for %s\n",
2039                        mdname(mddev));
2040                 goto out_free_conf;
2041         }
2042
2043         printk(KERN_INFO 
2044                 "raid1: raid set %s active with %d out of %d mirrors\n",
2045                 mdname(mddev), mddev->raid_disks - mddev->degraded, 
2046                 mddev->raid_disks);
2047         /*
2048          * Ok, everything is just fine now
2049          */
2050         mddev->array_sectors = mddev->size * 2;
2051
2052         mddev->queue->unplug_fn = raid1_unplug;
2053         mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2054         mddev->queue->backing_dev_info.congested_data = mddev;
2055
2056         return 0;
2057
2058 out_no_mem:
2059         printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
2060                mdname(mddev));
2061
2062 out_free_conf:
2063         if (conf) {
2064                 if (conf->r1bio_pool)
2065                         mempool_destroy(conf->r1bio_pool);
2066                 kfree(conf->mirrors);
2067                 safe_put_page(conf->tmppage);
2068                 kfree(conf->poolinfo);
2069                 kfree(conf);
2070                 mddev->private = NULL;
2071         }
2072 out:
2073         return -EIO;
2074 }
2075
2076 static int stop(mddev_t *mddev)
2077 {
2078         conf_t *conf = mddev_to_conf(mddev);
2079         struct bitmap *bitmap = mddev->bitmap;
2080         int behind_wait = 0;
2081
2082         /* wait for behind writes to complete */
2083         while (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2084                 behind_wait++;
2085                 printk(KERN_INFO "raid1: behind writes in progress on device %s, waiting to stop (%d)\n", mdname(mddev), behind_wait);
2086                 set_current_state(TASK_UNINTERRUPTIBLE);
2087                 schedule_timeout(HZ); /* wait a second */
2088                 /* need to kick something here to make sure I/O goes? */
2089         }
2090
2091         md_unregister_thread(mddev->thread);
2092         mddev->thread = NULL;
2093         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2094         if (conf->r1bio_pool)
2095                 mempool_destroy(conf->r1bio_pool);
2096         kfree(conf->mirrors);
2097         kfree(conf->poolinfo);
2098         kfree(conf);
2099         mddev->private = NULL;
2100         return 0;
2101 }
2102
2103 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2104 {
2105         /* no resync is happening, and there is enough space
2106          * on all devices, so we can resize.
2107          * We need to make sure resync covers any new space.
2108          * If the array is shrinking we should possibly wait until
2109          * any io in the removed space completes, but it hardly seems
2110          * worth it.
2111          */
2112         mddev->array_sectors = sectors;
2113         set_capacity(mddev->gendisk, mddev->array_sectors);
2114         mddev->changed = 1;
2115         if (mddev->array_sectors / 2 > mddev->size &&
2116             mddev->recovery_cp == MaxSector) {
2117                 mddev->recovery_cp = mddev->size << 1;
2118                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2119         }
2120         mddev->size = mddev->array_sectors / 2;
2121         mddev->resync_max_sectors = sectors;
2122         return 0;
2123 }
2124
2125 static int raid1_reshape(mddev_t *mddev)
2126 {
2127         /* We need to:
2128          * 1/ resize the r1bio_pool
2129          * 2/ resize conf->mirrors
2130          *
2131          * We allocate a new r1bio_pool if we can.
2132          * Then raise a device barrier and wait until all IO stops.
2133          * Then resize conf->mirrors and swap in the new r1bio pool.
2134          *
2135          * At the same time, we "pack" the devices so that all the missing
2136          * devices have the higher raid_disk numbers.
2137          */
2138         mempool_t *newpool, *oldpool;
2139         struct pool_info *newpoolinfo;
2140         mirror_info_t *newmirrors;
2141         conf_t *conf = mddev_to_conf(mddev);
2142         int cnt, raid_disks;
2143         unsigned long flags;
2144         int d, d2, err;
2145
2146         /* Cannot change chunk_size, layout, or level */
2147         if (mddev->chunk_size != mddev->new_chunk ||
2148             mddev->layout != mddev->new_layout ||
2149             mddev->level != mddev->new_level) {
2150                 mddev->new_chunk = mddev->chunk_size;
2151                 mddev->new_layout = mddev->layout;
2152                 mddev->new_level = mddev->level;
2153                 return -EINVAL;
2154         }
2155
2156         err = md_allow_write(mddev);
2157         if (err)
2158                 return err;
2159
2160         raid_disks = mddev->raid_disks + mddev->delta_disks;
2161
2162         if (raid_disks < conf->raid_disks) {
2163                 cnt=0;
2164                 for (d= 0; d < conf->raid_disks; d++)
2165                         if (conf->mirrors[d].rdev)
2166                                 cnt++;
2167                 if (cnt > raid_disks)
2168                         return -EBUSY;
2169         }
2170
2171         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2172         if (!newpoolinfo)
2173                 return -ENOMEM;
2174         newpoolinfo->mddev = mddev;
2175         newpoolinfo->raid_disks = raid_disks;
2176
2177         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2178                                  r1bio_pool_free, newpoolinfo);
2179         if (!newpool) {
2180                 kfree(newpoolinfo);
2181                 return -ENOMEM;
2182         }
2183         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2184         if (!newmirrors) {
2185                 kfree(newpoolinfo);
2186                 mempool_destroy(newpool);
2187                 return -ENOMEM;
2188         }
2189
2190         raise_barrier(conf);
2191
2192         /* ok, everything is stopped */
2193         oldpool = conf->r1bio_pool;
2194         conf->r1bio_pool = newpool;
2195
2196         for (d = d2 = 0; d < conf->raid_disks; d++) {
2197                 mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2198                 if (rdev && rdev->raid_disk != d2) {
2199                         char nm[20];
2200                         sprintf(nm, "rd%d", rdev->raid_disk);
2201                         sysfs_remove_link(&mddev->kobj, nm);
2202                         rdev->raid_disk = d2;
2203                         sprintf(nm, "rd%d", rdev->raid_disk);
2204                         sysfs_remove_link(&mddev->kobj, nm);
2205                         if (sysfs_create_link(&mddev->kobj,
2206                                               &rdev->kobj, nm))
2207                                 printk(KERN_WARNING
2208                                        "md/raid1: cannot register "
2209                                        "%s for %s\n",
2210                                        nm, mdname(mddev));
2211                 }
2212                 if (rdev)
2213                         newmirrors[d2++].rdev = rdev;
2214         }
2215         kfree(conf->mirrors);
2216         conf->mirrors = newmirrors;
2217         kfree(conf->poolinfo);
2218         conf->poolinfo = newpoolinfo;
2219
2220         spin_lock_irqsave(&conf->device_lock, flags);
2221         mddev->degraded += (raid_disks - conf->raid_disks);
2222         spin_unlock_irqrestore(&conf->device_lock, flags);
2223         conf->raid_disks = mddev->raid_disks = raid_disks;
2224         mddev->delta_disks = 0;
2225
2226         conf->last_used = 0; /* just make sure it is in-range */
2227         lower_barrier(conf);
2228
2229         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2230         md_wakeup_thread(mddev->thread);
2231
2232         mempool_destroy(oldpool);
2233         return 0;
2234 }
2235
2236 static void raid1_quiesce(mddev_t *mddev, int state)
2237 {
2238         conf_t *conf = mddev_to_conf(mddev);
2239
2240         switch(state) {
2241         case 1:
2242                 raise_barrier(conf);
2243                 break;
2244         case 0:
2245                 lower_barrier(conf);
2246                 break;
2247         }
2248 }
2249
2250
2251 static struct mdk_personality raid1_personality =
2252 {
2253         .name           = "raid1",
2254         .level          = 1,
2255         .owner          = THIS_MODULE,
2256         .make_request   = make_request,
2257         .run            = run,
2258         .stop           = stop,
2259         .status         = status,
2260         .error_handler  = error,
2261         .hot_add_disk   = raid1_add_disk,
2262         .hot_remove_disk= raid1_remove_disk,
2263         .spare_active   = raid1_spare_active,
2264         .sync_request   = sync_request,
2265         .resize         = raid1_resize,
2266         .check_reshape  = raid1_reshape,
2267         .quiesce        = raid1_quiesce,
2268 };
2269
2270 static int __init raid_init(void)
2271 {
2272         return register_md_personality(&raid1_personality);
2273 }
2274
2275 static void raid_exit(void)
2276 {
2277         unregister_md_personality(&raid1_personality);
2278 }
2279
2280 module_init(raid_init);
2281 module_exit(raid_exit);
2282 MODULE_LICENSE("GPL");
2283 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2284 MODULE_ALIAS("md-raid1");
2285 MODULE_ALIAS("md-level-1");