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