8726fd7ebce5d7bebfb48012db4785fe4b612c56
[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 <linux/delay.h>
35 #include <linux/blkdev.h>
36 #include <linux/seq_file.h>
37 #include "md.h"
38 #include "raid1.h"
39 #include "bitmap.h"
40
41 #define DEBUG 0
42 #if DEBUG
43 #define PRINTK(x...) printk(x)
44 #else
45 #define PRINTK(x...)
46 #endif
47
48 /*
49  * Number of guaranteed r1bios in case of extreme VM load:
50  */
51 #define NR_RAID1_BIOS 256
52
53
54 static void unplug_slaves(mddev_t *mddev);
55
56 static void allow_barrier(conf_t *conf);
57 static void lower_barrier(conf_t *conf);
58
59 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
60 {
61         struct pool_info *pi = data;
62         r1bio_t *r1_bio;
63         int size = offsetof(r1bio_t, bios[pi->raid_disks]);
64
65         /* allocate a r1bio with room for raid_disks entries in the bios array */
66         r1_bio = kzalloc(size, gfp_flags);
67         if (!r1_bio)
68                 unplug_slaves(pi->mddev);
69
70         return r1_bio;
71 }
72
73 static void r1bio_pool_free(void *r1_bio, void *data)
74 {
75         kfree(r1_bio);
76 }
77
78 #define RESYNC_BLOCK_SIZE (64*1024)
79 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
80 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
81 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
82 #define RESYNC_WINDOW (2048*1024)
83
84 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
85 {
86         struct pool_info *pi = data;
87         struct page *page;
88         r1bio_t *r1_bio;
89         struct bio *bio;
90         int i, j;
91
92         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
93         if (!r1_bio) {
94                 unplug_slaves(pi->mddev);
95                 return NULL;
96         }
97
98         /*
99          * Allocate bios : 1 for reading, n-1 for writing
100          */
101         for (j = pi->raid_disks ; j-- ; ) {
102                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
103                 if (!bio)
104                         goto out_free_bio;
105                 r1_bio->bios[j] = bio;
106         }
107         /*
108          * Allocate RESYNC_PAGES data pages and attach them to
109          * the first bio.
110          * If this is a user-requested check/repair, allocate
111          * RESYNC_PAGES for each bio.
112          */
113         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
114                 j = pi->raid_disks;
115         else
116                 j = 1;
117         while(j--) {
118                 bio = r1_bio->bios[j];
119                 for (i = 0; i < RESYNC_PAGES; i++) {
120                         page = alloc_page(gfp_flags);
121                         if (unlikely(!page))
122                                 goto out_free_pages;
123
124                         bio->bi_io_vec[i].bv_page = page;
125                         bio->bi_vcnt = i+1;
126                 }
127         }
128         /* If not user-requests, copy the page pointers to all bios */
129         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
130                 for (i=0; i<RESYNC_PAGES ; i++)
131                         for (j=1; j<pi->raid_disks; j++)
132                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
133                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
134         }
135
136         r1_bio->master_bio = NULL;
137
138         return r1_bio;
139
140 out_free_pages:
141         for (j=0 ; j < pi->raid_disks; j++)
142                 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
143                         put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
144         j = -1;
145 out_free_bio:
146         while ( ++j < pi->raid_disks )
147                 bio_put(r1_bio->bios[j]);
148         r1bio_pool_free(r1_bio, data);
149         return NULL;
150 }
151
152 static void r1buf_pool_free(void *__r1_bio, void *data)
153 {
154         struct pool_info *pi = data;
155         int i,j;
156         r1bio_t *r1bio = __r1_bio;
157
158         for (i = 0; i < RESYNC_PAGES; i++)
159                 for (j = pi->raid_disks; j-- ;) {
160                         if (j == 0 ||
161                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
162                             r1bio->bios[0]->bi_io_vec[i].bv_page)
163                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
164                 }
165         for (i=0 ; i < pi->raid_disks; i++)
166                 bio_put(r1bio->bios[i]);
167
168         r1bio_pool_free(r1bio, data);
169 }
170
171 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
172 {
173         int i;
174
175         for (i = 0; i < conf->raid_disks; i++) {
176                 struct bio **bio = r1_bio->bios + i;
177                 if (*bio && *bio != IO_BLOCKED)
178                         bio_put(*bio);
179                 *bio = NULL;
180         }
181 }
182
183 static void free_r1bio(r1bio_t *r1_bio)
184 {
185         conf_t *conf = r1_bio->mddev->private;
186
187         /*
188          * Wake up any possible resync thread that waits for the device
189          * to go idle.
190          */
191         allow_barrier(conf);
192
193         put_all_bios(conf, r1_bio);
194         mempool_free(r1_bio, conf->r1bio_pool);
195 }
196
197 static void put_buf(r1bio_t *r1_bio)
198 {
199         conf_t *conf = r1_bio->mddev->private;
200         int i;
201
202         for (i=0; i<conf->raid_disks; i++) {
203                 struct bio *bio = r1_bio->bios[i];
204                 if (bio->bi_end_io)
205                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
206         }
207
208         mempool_free(r1_bio, conf->r1buf_pool);
209
210         lower_barrier(conf);
211 }
212
213 static void reschedule_retry(r1bio_t *r1_bio)
214 {
215         unsigned long flags;
216         mddev_t *mddev = r1_bio->mddev;
217         conf_t *conf = mddev->private;
218
219         spin_lock_irqsave(&conf->device_lock, flags);
220         list_add(&r1_bio->retry_list, &conf->retry_list);
221         conf->nr_queued ++;
222         spin_unlock_irqrestore(&conf->device_lock, flags);
223
224         wake_up(&conf->wait_barrier);
225         md_wakeup_thread(mddev->thread);
226 }
227
228 /*
229  * raid_end_bio_io() is called when we have finished servicing a mirrored
230  * operation and are ready to return a success/failure code to the buffer
231  * cache layer.
232  */
233 static void raid_end_bio_io(r1bio_t *r1_bio)
234 {
235         struct bio *bio = r1_bio->master_bio;
236
237         /* if nobody has done the final endio yet, do it now */
238         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
239                 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
240                         (bio_data_dir(bio) == WRITE) ? "write" : "read",
241                         (unsigned long long) bio->bi_sector,
242                         (unsigned long long) bio->bi_sector +
243                                 (bio->bi_size >> 9) - 1);
244
245                 bio_endio(bio,
246                         test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
247         }
248         free_r1bio(r1_bio);
249 }
250
251 /*
252  * Update disk head position estimator based on IRQ completion info.
253  */
254 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
255 {
256         conf_t *conf = r1_bio->mddev->private;
257
258         conf->mirrors[disk].head_position =
259                 r1_bio->sector + (r1_bio->sectors);
260 }
261
262 static void raid1_end_read_request(struct bio *bio, int error)
263 {
264         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
265         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
266         int mirror;
267         conf_t *conf = r1_bio->mddev->private;
268
269         mirror = r1_bio->read_disk;
270         /*
271          * this branch is our 'one mirror IO has finished' event handler:
272          */
273         update_head_pos(mirror, r1_bio);
274
275         if (uptodate)
276                 set_bit(R1BIO_Uptodate, &r1_bio->state);
277         else {
278                 /* If all other devices have failed, we want to return
279                  * the error upwards rather than fail the last device.
280                  * Here we redefine "uptodate" to mean "Don't want to retry"
281                  */
282                 unsigned long flags;
283                 spin_lock_irqsave(&conf->device_lock, flags);
284                 if (r1_bio->mddev->degraded == conf->raid_disks ||
285                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
286                      !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
287                         uptodate = 1;
288                 spin_unlock_irqrestore(&conf->device_lock, flags);
289         }
290
291         if (uptodate)
292                 raid_end_bio_io(r1_bio);
293         else {
294                 /*
295                  * oops, read error:
296                  */
297                 char b[BDEVNAME_SIZE];
298                 if (printk_ratelimit())
299                         printk(KERN_ERR "raid1: %s: rescheduling sector %llu\n",
300                                bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
301                 reschedule_retry(r1_bio);
302         }
303
304         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
305 }
306
307 static void raid1_end_write_request(struct bio *bio, int error)
308 {
309         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
310         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
311         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
312         conf_t *conf = r1_bio->mddev->private;
313         struct bio *to_put = NULL;
314
315
316         for (mirror = 0; mirror < conf->raid_disks; mirror++)
317                 if (r1_bio->bios[mirror] == bio)
318                         break;
319
320         if (error == -EOPNOTSUPP && test_bit(R1BIO_Barrier, &r1_bio->state)) {
321                 set_bit(BarriersNotsupp, &conf->mirrors[mirror].rdev->flags);
322                 set_bit(R1BIO_BarrierRetry, &r1_bio->state);
323                 r1_bio->mddev->barriers_work = 0;
324                 /* Don't rdev_dec_pending in this branch - keep it for the retry */
325         } else {
326                 /*
327                  * this branch is our 'one mirror IO has finished' event handler:
328                  */
329                 r1_bio->bios[mirror] = NULL;
330                 to_put = bio;
331                 if (!uptodate) {
332                         md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
333                         /* an I/O failed, we can't clear the bitmap */
334                         set_bit(R1BIO_Degraded, &r1_bio->state);
335                 } else
336                         /*
337                          * Set R1BIO_Uptodate in our master bio, so that
338                          * we will return a good error code for to the higher
339                          * levels even if IO on some other mirrored buffer fails.
340                          *
341                          * The 'master' represents the composite IO operation to
342                          * user-side. So if something waits for IO, then it will
343                          * wait for the 'master' bio.
344                          */
345                         set_bit(R1BIO_Uptodate, &r1_bio->state);
346
347                 update_head_pos(mirror, r1_bio);
348
349                 if (behind) {
350                         if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
351                                 atomic_dec(&r1_bio->behind_remaining);
352
353                         /* In behind mode, we ACK the master bio once the I/O has safely
354                          * reached all non-writemostly disks. Setting the Returned bit
355                          * ensures that this gets done only once -- we don't ever want to
356                          * return -EIO here, instead we'll wait */
357
358                         if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
359                             test_bit(R1BIO_Uptodate, &r1_bio->state)) {
360                                 /* Maybe we can return now */
361                                 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
362                                         struct bio *mbio = r1_bio->master_bio;
363                                         PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
364                                                (unsigned long long) mbio->bi_sector,
365                                                (unsigned long long) mbio->bi_sector +
366                                                (mbio->bi_size >> 9) - 1);
367                                         bio_endio(mbio, 0);
368                                 }
369                         }
370                 }
371                 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
372         }
373         /*
374          *
375          * Let's see if all mirrored write operations have finished
376          * already.
377          */
378         if (atomic_dec_and_test(&r1_bio->remaining)) {
379                 if (test_bit(R1BIO_BarrierRetry, &r1_bio->state))
380                         reschedule_retry(r1_bio);
381                 else {
382                         /* it really is the end of this request */
383                         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
384                                 /* free extra copy of the data pages */
385                                 int i = bio->bi_vcnt;
386                                 while (i--)
387                                         safe_put_page(bio->bi_io_vec[i].bv_page);
388                         }
389                         /* clear the bitmap if all writes complete successfully */
390                         bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
391                                         r1_bio->sectors,
392                                         !test_bit(R1BIO_Degraded, &r1_bio->state),
393                                         behind);
394                         md_write_end(r1_bio->mddev);
395                         raid_end_bio_io(r1_bio);
396                 }
397         }
398
399         if (to_put)
400                 bio_put(to_put);
401 }
402
403
404 /*
405  * This routine returns the disk from which the requested read should
406  * be done. There is a per-array 'next expected sequential IO' sector
407  * number - if this matches on the next IO then we use the last disk.
408  * There is also a per-disk 'last know head position' sector that is
409  * maintained from IRQ contexts, both the normal and the resync IO
410  * completion handlers update this position correctly. If there is no
411  * perfect sequential match then we pick the disk whose head is closest.
412  *
413  * If there are 2 mirrors in the same 2 devices, performance degrades
414  * because position is mirror, not device based.
415  *
416  * The rdev for the device selected will have nr_pending incremented.
417  */
418 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
419 {
420         const unsigned long this_sector = r1_bio->sector;
421         int new_disk = conf->last_used, disk = new_disk;
422         int wonly_disk = -1;
423         const int sectors = r1_bio->sectors;
424         sector_t new_distance, current_distance;
425         mdk_rdev_t *rdev;
426
427         rcu_read_lock();
428         /*
429          * Check if we can balance. We can balance on the whole
430          * device if no resync is going on, or below the resync window.
431          * We take the first readable disk when above the resync window.
432          */
433  retry:
434         if (conf->mddev->recovery_cp < MaxSector &&
435             (this_sector + sectors >= conf->next_resync)) {
436                 /* Choose the first operation device, for consistancy */
437                 new_disk = 0;
438
439                 for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
440                      r1_bio->bios[new_disk] == IO_BLOCKED ||
441                      !rdev || !test_bit(In_sync, &rdev->flags)
442                              || test_bit(WriteMostly, &rdev->flags);
443                      rdev = rcu_dereference(conf->mirrors[++new_disk].rdev)) {
444
445                         if (rdev && test_bit(In_sync, &rdev->flags) &&
446                                 r1_bio->bios[new_disk] != IO_BLOCKED)
447                                 wonly_disk = new_disk;
448
449                         if (new_disk == conf->raid_disks - 1) {
450                                 new_disk = wonly_disk;
451                                 break;
452                         }
453                 }
454                 goto rb_out;
455         }
456
457
458         /* make sure the disk is operational */
459         for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
460              r1_bio->bios[new_disk] == IO_BLOCKED ||
461              !rdev || !test_bit(In_sync, &rdev->flags) ||
462                      test_bit(WriteMostly, &rdev->flags);
463              rdev = rcu_dereference(conf->mirrors[new_disk].rdev)) {
464
465                 if (rdev && test_bit(In_sync, &rdev->flags) &&
466                     r1_bio->bios[new_disk] != IO_BLOCKED)
467                         wonly_disk = new_disk;
468
469                 if (new_disk <= 0)
470                         new_disk = conf->raid_disks;
471                 new_disk--;
472                 if (new_disk == disk) {
473                         new_disk = wonly_disk;
474                         break;
475                 }
476         }
477
478         if (new_disk < 0)
479                 goto rb_out;
480
481         disk = new_disk;
482         /* now disk == new_disk == starting point for search */
483
484         /*
485          * Don't change to another disk for sequential reads:
486          */
487         if (conf->next_seq_sect == this_sector)
488                 goto rb_out;
489         if (this_sector == conf->mirrors[new_disk].head_position)
490                 goto rb_out;
491
492         current_distance = abs(this_sector - conf->mirrors[disk].head_position);
493
494         /* Find the disk whose head is closest */
495
496         do {
497                 if (disk <= 0)
498                         disk = conf->raid_disks;
499                 disk--;
500
501                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
502
503                 if (!rdev || r1_bio->bios[disk] == IO_BLOCKED ||
504                     !test_bit(In_sync, &rdev->flags) ||
505                     test_bit(WriteMostly, &rdev->flags))
506                         continue;
507
508                 if (!atomic_read(&rdev->nr_pending)) {
509                         new_disk = disk;
510                         break;
511                 }
512                 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
513                 if (new_distance < current_distance) {
514                         current_distance = new_distance;
515                         new_disk = disk;
516                 }
517         } while (disk != conf->last_used);
518
519  rb_out:
520
521
522         if (new_disk >= 0) {
523                 rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
524                 if (!rdev)
525                         goto retry;
526                 atomic_inc(&rdev->nr_pending);
527                 if (!test_bit(In_sync, &rdev->flags)) {
528                         /* cannot risk returning a device that failed
529                          * before we inc'ed nr_pending
530                          */
531                         rdev_dec_pending(rdev, conf->mddev);
532                         goto retry;
533                 }
534                 conf->next_seq_sect = this_sector + sectors;
535                 conf->last_used = new_disk;
536         }
537         rcu_read_unlock();
538
539         return new_disk;
540 }
541
542 static void unplug_slaves(mddev_t *mddev)
543 {
544         conf_t *conf = mddev->private;
545         int i;
546
547         rcu_read_lock();
548         for (i=0; i<mddev->raid_disks; i++) {
549                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
550                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
551                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
552
553                         atomic_inc(&rdev->nr_pending);
554                         rcu_read_unlock();
555
556                         blk_unplug(r_queue);
557
558                         rdev_dec_pending(rdev, mddev);
559                         rcu_read_lock();
560                 }
561         }
562         rcu_read_unlock();
563 }
564
565 static void raid1_unplug(struct request_queue *q)
566 {
567         mddev_t *mddev = q->queuedata;
568
569         unplug_slaves(mddev);
570         md_wakeup_thread(mddev->thread);
571 }
572
573 static int raid1_congested(void *data, int bits)
574 {
575         mddev_t *mddev = data;
576         conf_t *conf = mddev->private;
577         int i, ret = 0;
578
579         rcu_read_lock();
580         for (i = 0; i < mddev->raid_disks; i++) {
581                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
582                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
583                         struct request_queue *q = bdev_get_queue(rdev->bdev);
584
585                         /* Note the '|| 1' - when read_balance prefers
586                          * non-congested targets, it can be removed
587                          */
588                         if ((bits & (1<<BDI_async_congested)) || 1)
589                                 ret |= bdi_congested(&q->backing_dev_info, bits);
590                         else
591                                 ret &= bdi_congested(&q->backing_dev_info, bits);
592                 }
593         }
594         rcu_read_unlock();
595         return ret;
596 }
597
598
599 static int flush_pending_writes(conf_t *conf)
600 {
601         /* Any writes that have been queued but are awaiting
602          * bitmap updates get flushed here.
603          * We return 1 if any requests were actually submitted.
604          */
605         int rv = 0;
606
607         spin_lock_irq(&conf->device_lock);
608
609         if (conf->pending_bio_list.head) {
610                 struct bio *bio;
611                 bio = bio_list_get(&conf->pending_bio_list);
612                 blk_remove_plug(conf->mddev->queue);
613                 spin_unlock_irq(&conf->device_lock);
614                 /* flush any pending bitmap writes to
615                  * disk before proceeding w/ I/O */
616                 bitmap_unplug(conf->mddev->bitmap);
617
618                 while (bio) { /* submit pending writes */
619                         struct bio *next = bio->bi_next;
620                         bio->bi_next = NULL;
621                         generic_make_request(bio);
622                         bio = next;
623                 }
624                 rv = 1;
625         } else
626                 spin_unlock_irq(&conf->device_lock);
627         return rv;
628 }
629
630 /* Barriers....
631  * Sometimes we need to suspend IO while we do something else,
632  * either some resync/recovery, or reconfigure the array.
633  * To do this we raise a 'barrier'.
634  * The 'barrier' is a counter that can be raised multiple times
635  * to count how many activities are happening which preclude
636  * normal IO.
637  * We can only raise the barrier if there is no pending IO.
638  * i.e. if nr_pending == 0.
639  * We choose only to raise the barrier if no-one is waiting for the
640  * barrier to go down.  This means that as soon as an IO request
641  * is ready, no other operations which require a barrier will start
642  * until the IO request has had a chance.
643  *
644  * So: regular IO calls 'wait_barrier'.  When that returns there
645  *    is no backgroup IO happening,  It must arrange to call
646  *    allow_barrier when it has finished its IO.
647  * backgroup IO calls must call raise_barrier.  Once that returns
648  *    there is no normal IO happeing.  It must arrange to call
649  *    lower_barrier when the particular background IO completes.
650  */
651 #define RESYNC_DEPTH 32
652
653 static void raise_barrier(conf_t *conf)
654 {
655         spin_lock_irq(&conf->resync_lock);
656
657         /* Wait until no block IO is waiting */
658         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
659                             conf->resync_lock,
660                             raid1_unplug(conf->mddev->queue));
661
662         /* block any new IO from starting */
663         conf->barrier++;
664
665         /* No wait for all pending IO to complete */
666         wait_event_lock_irq(conf->wait_barrier,
667                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
668                             conf->resync_lock,
669                             raid1_unplug(conf->mddev->queue));
670
671         spin_unlock_irq(&conf->resync_lock);
672 }
673
674 static void lower_barrier(conf_t *conf)
675 {
676         unsigned long flags;
677         spin_lock_irqsave(&conf->resync_lock, flags);
678         conf->barrier--;
679         spin_unlock_irqrestore(&conf->resync_lock, flags);
680         wake_up(&conf->wait_barrier);
681 }
682
683 static void wait_barrier(conf_t *conf)
684 {
685         spin_lock_irq(&conf->resync_lock);
686         if (conf->barrier) {
687                 conf->nr_waiting++;
688                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
689                                     conf->resync_lock,
690                                     raid1_unplug(conf->mddev->queue));
691                 conf->nr_waiting--;
692         }
693         conf->nr_pending++;
694         spin_unlock_irq(&conf->resync_lock);
695 }
696
697 static void allow_barrier(conf_t *conf)
698 {
699         unsigned long flags;
700         spin_lock_irqsave(&conf->resync_lock, flags);
701         conf->nr_pending--;
702         spin_unlock_irqrestore(&conf->resync_lock, flags);
703         wake_up(&conf->wait_barrier);
704 }
705
706 static void freeze_array(conf_t *conf)
707 {
708         /* stop syncio and normal IO and wait for everything to
709          * go quite.
710          * We increment barrier and nr_waiting, and then
711          * wait until nr_pending match nr_queued+1
712          * This is called in the context of one normal IO request
713          * that has failed. Thus any sync request that might be pending
714          * will be blocked by nr_pending, and we need to wait for
715          * pending IO requests to complete or be queued for re-try.
716          * Thus the number queued (nr_queued) plus this request (1)
717          * must match the number of pending IOs (nr_pending) before
718          * we continue.
719          */
720         spin_lock_irq(&conf->resync_lock);
721         conf->barrier++;
722         conf->nr_waiting++;
723         wait_event_lock_irq(conf->wait_barrier,
724                             conf->nr_pending == conf->nr_queued+1,
725                             conf->resync_lock,
726                             ({ flush_pending_writes(conf);
727                                raid1_unplug(conf->mddev->queue); }));
728         spin_unlock_irq(&conf->resync_lock);
729 }
730 static void unfreeze_array(conf_t *conf)
731 {
732         /* reverse the effect of the freeze */
733         spin_lock_irq(&conf->resync_lock);
734         conf->barrier--;
735         conf->nr_waiting--;
736         wake_up(&conf->wait_barrier);
737         spin_unlock_irq(&conf->resync_lock);
738 }
739
740
741 /* duplicate the data pages for behind I/O */
742 static struct page **alloc_behind_pages(struct bio *bio)
743 {
744         int i;
745         struct bio_vec *bvec;
746         struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page *),
747                                         GFP_NOIO);
748         if (unlikely(!pages))
749                 goto do_sync_io;
750
751         bio_for_each_segment(bvec, bio, i) {
752                 pages[i] = alloc_page(GFP_NOIO);
753                 if (unlikely(!pages[i]))
754                         goto do_sync_io;
755                 memcpy(kmap(pages[i]) + bvec->bv_offset,
756                         kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
757                 kunmap(pages[i]);
758                 kunmap(bvec->bv_page);
759         }
760
761         return pages;
762
763 do_sync_io:
764         if (pages)
765                 for (i = 0; i < bio->bi_vcnt && pages[i]; i++)
766                         put_page(pages[i]);
767         kfree(pages);
768         PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
769         return NULL;
770 }
771
772 static int make_request(struct request_queue *q, struct bio * bio)
773 {
774         mddev_t *mddev = q->queuedata;
775         conf_t *conf = mddev->private;
776         mirror_info_t *mirror;
777         r1bio_t *r1_bio;
778         struct bio *read_bio;
779         int i, targets = 0, disks;
780         struct bitmap *bitmap;
781         unsigned long flags;
782         struct bio_list bl;
783         struct page **behind_pages = NULL;
784         const int rw = bio_data_dir(bio);
785         const int do_sync = bio_sync(bio);
786         int cpu, do_barriers;
787         mdk_rdev_t *blocked_rdev;
788
789         /*
790          * Register the new request and wait if the reconstruction
791          * thread has put up a bar for new requests.
792          * Continue immediately if no resync is active currently.
793          * We test barriers_work *after* md_write_start as md_write_start
794          * may cause the first superblock write, and that will check out
795          * if barriers work.
796          */
797
798         md_write_start(mddev, bio); /* wait on superblock update early */
799
800         if (unlikely(!mddev->barriers_work && bio_barrier(bio))) {
801                 if (rw == WRITE)
802                         md_write_end(mddev);
803                 bio_endio(bio, -EOPNOTSUPP);
804                 return 0;
805         }
806
807         wait_barrier(conf);
808
809         bitmap = mddev->bitmap;
810
811         cpu = part_stat_lock();
812         part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
813         part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
814                       bio_sectors(bio));
815         part_stat_unlock();
816
817         /*
818          * make_request() can abort the operation when READA is being
819          * used and no empty request is available.
820          *
821          */
822         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
823
824         r1_bio->master_bio = bio;
825         r1_bio->sectors = bio->bi_size >> 9;
826         r1_bio->state = 0;
827         r1_bio->mddev = mddev;
828         r1_bio->sector = bio->bi_sector;
829
830         if (rw == READ) {
831                 /*
832                  * read balancing logic:
833                  */
834                 int rdisk = read_balance(conf, r1_bio);
835
836                 if (rdisk < 0) {
837                         /* couldn't find anywhere to read from */
838                         raid_end_bio_io(r1_bio);
839                         return 0;
840                 }
841                 mirror = conf->mirrors + rdisk;
842
843                 r1_bio->read_disk = rdisk;
844
845                 read_bio = bio_clone(bio, GFP_NOIO);
846
847                 r1_bio->bios[rdisk] = read_bio;
848
849                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
850                 read_bio->bi_bdev = mirror->rdev->bdev;
851                 read_bio->bi_end_io = raid1_end_read_request;
852                 read_bio->bi_rw = READ | do_sync;
853                 read_bio->bi_private = r1_bio;
854
855                 generic_make_request(read_bio);
856                 return 0;
857         }
858
859         /*
860          * WRITE:
861          */
862         /* first select target devices under spinlock and
863          * inc refcount on their rdev.  Record them by setting
864          * bios[x] to bio
865          */
866         disks = conf->raid_disks;
867 #if 0
868         { static int first=1;
869         if (first) printk("First Write sector %llu disks %d\n",
870                           (unsigned long long)r1_bio->sector, disks);
871         first = 0;
872         }
873 #endif
874  retry_write:
875         blocked_rdev = NULL;
876         rcu_read_lock();
877         for (i = 0;  i < disks; i++) {
878                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
879                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
880                         atomic_inc(&rdev->nr_pending);
881                         blocked_rdev = rdev;
882                         break;
883                 }
884                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
885                         atomic_inc(&rdev->nr_pending);
886                         if (test_bit(Faulty, &rdev->flags)) {
887                                 rdev_dec_pending(rdev, mddev);
888                                 r1_bio->bios[i] = NULL;
889                         } else
890                                 r1_bio->bios[i] = bio;
891                         targets++;
892                 } else
893                         r1_bio->bios[i] = NULL;
894         }
895         rcu_read_unlock();
896
897         if (unlikely(blocked_rdev)) {
898                 /* Wait for this device to become unblocked */
899                 int j;
900
901                 for (j = 0; j < i; j++)
902                         if (r1_bio->bios[j])
903                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
904
905                 allow_barrier(conf);
906                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
907                 wait_barrier(conf);
908                 goto retry_write;
909         }
910
911         BUG_ON(targets == 0); /* we never fail the last device */
912
913         if (targets < conf->raid_disks) {
914                 /* array is degraded, we will not clear the bitmap
915                  * on I/O completion (see raid1_end_write_request) */
916                 set_bit(R1BIO_Degraded, &r1_bio->state);
917         }
918
919         /* do behind I/O ? */
920         if (bitmap &&
921             atomic_read(&bitmap->behind_writes) < bitmap->max_write_behind &&
922             (behind_pages = alloc_behind_pages(bio)) != NULL)
923                 set_bit(R1BIO_BehindIO, &r1_bio->state);
924
925         atomic_set(&r1_bio->remaining, 0);
926         atomic_set(&r1_bio->behind_remaining, 0);
927
928         do_barriers = bio_barrier(bio);
929         if (do_barriers)
930                 set_bit(R1BIO_Barrier, &r1_bio->state);
931
932         bio_list_init(&bl);
933         for (i = 0; i < disks; i++) {
934                 struct bio *mbio;
935                 if (!r1_bio->bios[i])
936                         continue;
937
938                 mbio = bio_clone(bio, GFP_NOIO);
939                 r1_bio->bios[i] = mbio;
940
941                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
942                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
943                 mbio->bi_end_io = raid1_end_write_request;
944                 mbio->bi_rw = WRITE | do_barriers | do_sync;
945                 mbio->bi_private = r1_bio;
946
947                 if (behind_pages) {
948                         struct bio_vec *bvec;
949                         int j;
950
951                         /* Yes, I really want the '__' version so that
952                          * we clear any unused pointer in the io_vec, rather
953                          * than leave them unchanged.  This is important
954                          * because when we come to free the pages, we won't
955                          * know the originial bi_idx, so we just free
956                          * them all
957                          */
958                         __bio_for_each_segment(bvec, mbio, j, 0)
959                                 bvec->bv_page = behind_pages[j];
960                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
961                                 atomic_inc(&r1_bio->behind_remaining);
962                 }
963
964                 atomic_inc(&r1_bio->remaining);
965
966                 bio_list_add(&bl, mbio);
967         }
968         kfree(behind_pages); /* the behind pages are attached to the bios now */
969
970         bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
971                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
972         spin_lock_irqsave(&conf->device_lock, flags);
973         bio_list_merge(&conf->pending_bio_list, &bl);
974         bio_list_init(&bl);
975
976         blk_plug_device(mddev->queue);
977         spin_unlock_irqrestore(&conf->device_lock, flags);
978
979         /* In case raid1d snuck into freeze_array */
980         wake_up(&conf->wait_barrier);
981
982         if (do_sync)
983                 md_wakeup_thread(mddev->thread);
984 #if 0
985         while ((bio = bio_list_pop(&bl)) != NULL)
986                 generic_make_request(bio);
987 #endif
988
989         return 0;
990 }
991
992 static void status(struct seq_file *seq, mddev_t *mddev)
993 {
994         conf_t *conf = mddev->private;
995         int i;
996
997         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
998                    conf->raid_disks - mddev->degraded);
999         rcu_read_lock();
1000         for (i = 0; i < conf->raid_disks; i++) {
1001                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1002                 seq_printf(seq, "%s",
1003                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1004         }
1005         rcu_read_unlock();
1006         seq_printf(seq, "]");
1007 }
1008
1009
1010 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1011 {
1012         char b[BDEVNAME_SIZE];
1013         conf_t *conf = mddev->private;
1014
1015         /*
1016          * If it is not operational, then we have already marked it as dead
1017          * else if it is the last working disks, ignore the error, let the
1018          * next level up know.
1019          * else mark the drive as failed
1020          */
1021         if (test_bit(In_sync, &rdev->flags)
1022             && (conf->raid_disks - mddev->degraded) == 1) {
1023                 /*
1024                  * Don't fail the drive, act as though we were just a
1025                  * normal single drive.
1026                  * However don't try a recovery from this drive as
1027                  * it is very likely to fail.
1028                  */
1029                 mddev->recovery_disabled = 1;
1030                 return;
1031         }
1032         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1033                 unsigned long flags;
1034                 spin_lock_irqsave(&conf->device_lock, flags);
1035                 mddev->degraded++;
1036                 set_bit(Faulty, &rdev->flags);
1037                 spin_unlock_irqrestore(&conf->device_lock, flags);
1038                 /*
1039                  * if recovery is running, make sure it aborts.
1040                  */
1041                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1042         } else
1043                 set_bit(Faulty, &rdev->flags);
1044         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1045         printk(KERN_ALERT "raid1: Disk failure on %s, disabling device.\n"
1046                 "raid1: Operation continuing on %d devices.\n",
1047                 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1048 }
1049
1050 static void print_conf(conf_t *conf)
1051 {
1052         int i;
1053
1054         printk("RAID1 conf printout:\n");
1055         if (!conf) {
1056                 printk("(!conf)\n");
1057                 return;
1058         }
1059         printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1060                 conf->raid_disks);
1061
1062         rcu_read_lock();
1063         for (i = 0; i < conf->raid_disks; i++) {
1064                 char b[BDEVNAME_SIZE];
1065                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1066                 if (rdev)
1067                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
1068                                i, !test_bit(In_sync, &rdev->flags),
1069                                !test_bit(Faulty, &rdev->flags),
1070                                bdevname(rdev->bdev,b));
1071         }
1072         rcu_read_unlock();
1073 }
1074
1075 static void close_sync(conf_t *conf)
1076 {
1077         wait_barrier(conf);
1078         allow_barrier(conf);
1079
1080         mempool_destroy(conf->r1buf_pool);
1081         conf->r1buf_pool = NULL;
1082 }
1083
1084 static int raid1_spare_active(mddev_t *mddev)
1085 {
1086         int i;
1087         conf_t *conf = mddev->private;
1088
1089         /*
1090          * Find all failed disks within the RAID1 configuration 
1091          * and mark them readable.
1092          * Called under mddev lock, so rcu protection not needed.
1093          */
1094         for (i = 0; i < conf->raid_disks; i++) {
1095                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1096                 if (rdev
1097                     && !test_bit(Faulty, &rdev->flags)
1098                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1099                         unsigned long flags;
1100                         spin_lock_irqsave(&conf->device_lock, flags);
1101                         mddev->degraded--;
1102                         spin_unlock_irqrestore(&conf->device_lock, flags);
1103                 }
1104         }
1105
1106         print_conf(conf);
1107         return 0;
1108 }
1109
1110
1111 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1112 {
1113         conf_t *conf = mddev->private;
1114         int err = -EEXIST;
1115         int mirror = 0;
1116         mirror_info_t *p;
1117         int first = 0;
1118         int last = mddev->raid_disks - 1;
1119
1120         if (rdev->raid_disk >= 0)
1121                 first = last = rdev->raid_disk;
1122
1123         for (mirror = first; mirror <= last; mirror++)
1124                 if ( !(p=conf->mirrors+mirror)->rdev) {
1125
1126                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1127                                           rdev->data_offset << 9);
1128                         /* as we don't honour merge_bvec_fn, we must never risk
1129                          * violating it, so limit ->max_sector to one PAGE, as
1130                          * a one page request is never in violation.
1131                          */
1132                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1133                             queue_max_sectors(mddev->queue) > (PAGE_SIZE>>9))
1134                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1135
1136                         p->head_position = 0;
1137                         rdev->raid_disk = mirror;
1138                         err = 0;
1139                         /* As all devices are equivalent, we don't need a full recovery
1140                          * if this was recently any drive of the array
1141                          */
1142                         if (rdev->saved_raid_disk < 0)
1143                                 conf->fullsync = 1;
1144                         rcu_assign_pointer(p->rdev, rdev);
1145                         break;
1146                 }
1147         md_integrity_add_rdev(rdev, mddev);
1148         print_conf(conf);
1149         return err;
1150 }
1151
1152 static int raid1_remove_disk(mddev_t *mddev, int number)
1153 {
1154         conf_t *conf = mddev->private;
1155         int err = 0;
1156         mdk_rdev_t *rdev;
1157         mirror_info_t *p = conf->mirrors+ number;
1158
1159         print_conf(conf);
1160         rdev = p->rdev;
1161         if (rdev) {
1162                 if (test_bit(In_sync, &rdev->flags) ||
1163                     atomic_read(&rdev->nr_pending)) {
1164                         err = -EBUSY;
1165                         goto abort;
1166                 }
1167                 /* Only remove non-faulty devices is recovery
1168                  * is not possible.
1169                  */
1170                 if (!test_bit(Faulty, &rdev->flags) &&
1171                     mddev->degraded < conf->raid_disks) {
1172                         err = -EBUSY;
1173                         goto abort;
1174                 }
1175                 p->rdev = NULL;
1176                 synchronize_rcu();
1177                 if (atomic_read(&rdev->nr_pending)) {
1178                         /* lost the race, try later */
1179                         err = -EBUSY;
1180                         p->rdev = rdev;
1181                         goto abort;
1182                 }
1183                 md_integrity_register(mddev);
1184         }
1185 abort:
1186
1187         print_conf(conf);
1188         return err;
1189 }
1190
1191
1192 static void end_sync_read(struct bio *bio, int error)
1193 {
1194         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1195         int i;
1196
1197         for (i=r1_bio->mddev->raid_disks; i--; )
1198                 if (r1_bio->bios[i] == bio)
1199                         break;
1200         BUG_ON(i < 0);
1201         update_head_pos(i, r1_bio);
1202         /*
1203          * we have read a block, now it needs to be re-written,
1204          * or re-read if the read failed.
1205          * We don't do much here, just schedule handling by raid1d
1206          */
1207         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1208                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1209
1210         if (atomic_dec_and_test(&r1_bio->remaining))
1211                 reschedule_retry(r1_bio);
1212 }
1213
1214 static void end_sync_write(struct bio *bio, int error)
1215 {
1216         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1217         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1218         mddev_t *mddev = r1_bio->mddev;
1219         conf_t *conf = mddev->private;
1220         int i;
1221         int mirror=0;
1222
1223         for (i = 0; i < conf->raid_disks; i++)
1224                 if (r1_bio->bios[i] == bio) {
1225                         mirror = i;
1226                         break;
1227                 }
1228         if (!uptodate) {
1229                 int sync_blocks = 0;
1230                 sector_t s = r1_bio->sector;
1231                 long sectors_to_go = r1_bio->sectors;
1232                 /* make sure these bits doesn't get cleared. */
1233                 do {
1234                         bitmap_end_sync(mddev->bitmap, s,
1235                                         &sync_blocks, 1);
1236                         s += sync_blocks;
1237                         sectors_to_go -= sync_blocks;
1238                 } while (sectors_to_go > 0);
1239                 md_error(mddev, conf->mirrors[mirror].rdev);
1240         }
1241
1242         update_head_pos(mirror, r1_bio);
1243
1244         if (atomic_dec_and_test(&r1_bio->remaining)) {
1245                 sector_t s = r1_bio->sectors;
1246                 put_buf(r1_bio);
1247                 md_done_sync(mddev, s, uptodate);
1248         }
1249 }
1250
1251 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1252 {
1253         conf_t *conf = mddev->private;
1254         int i;
1255         int disks = conf->raid_disks;
1256         struct bio *bio, *wbio;
1257
1258         bio = r1_bio->bios[r1_bio->read_disk];
1259
1260
1261         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1262                 /* We have read all readable devices.  If we haven't
1263                  * got the block, then there is no hope left.
1264                  * If we have, then we want to do a comparison
1265                  * and skip the write if everything is the same.
1266                  * If any blocks failed to read, then we need to
1267                  * attempt an over-write
1268                  */
1269                 int primary;
1270                 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1271                         for (i=0; i<mddev->raid_disks; i++)
1272                                 if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1273                                         md_error(mddev, conf->mirrors[i].rdev);
1274
1275                         md_done_sync(mddev, r1_bio->sectors, 1);
1276                         put_buf(r1_bio);
1277                         return;
1278                 }
1279                 for (primary=0; primary<mddev->raid_disks; primary++)
1280                         if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1281                             test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1282                                 r1_bio->bios[primary]->bi_end_io = NULL;
1283                                 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1284                                 break;
1285                         }
1286                 r1_bio->read_disk = primary;
1287                 for (i=0; i<mddev->raid_disks; i++)
1288                         if (r1_bio->bios[i]->bi_end_io == end_sync_read) {
1289                                 int j;
1290                                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1291                                 struct bio *pbio = r1_bio->bios[primary];
1292                                 struct bio *sbio = r1_bio->bios[i];
1293
1294                                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1295                                         for (j = vcnt; j-- ; ) {
1296                                                 struct page *p, *s;
1297                                                 p = pbio->bi_io_vec[j].bv_page;
1298                                                 s = sbio->bi_io_vec[j].bv_page;
1299                                                 if (memcmp(page_address(p),
1300                                                            page_address(s),
1301                                                            PAGE_SIZE))
1302                                                         break;
1303                                         }
1304                                 } else
1305                                         j = 0;
1306                                 if (j >= 0)
1307                                         mddev->resync_mismatches += r1_bio->sectors;
1308                                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1309                                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1310                                         sbio->bi_end_io = NULL;
1311                                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1312                                 } else {
1313                                         /* fixup the bio for reuse */
1314                                         int size;
1315                                         sbio->bi_vcnt = vcnt;
1316                                         sbio->bi_size = r1_bio->sectors << 9;
1317                                         sbio->bi_idx = 0;
1318                                         sbio->bi_phys_segments = 0;
1319                                         sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1320                                         sbio->bi_flags |= 1 << BIO_UPTODATE;
1321                                         sbio->bi_next = NULL;
1322                                         sbio->bi_sector = r1_bio->sector +
1323                                                 conf->mirrors[i].rdev->data_offset;
1324                                         sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1325                                         size = sbio->bi_size;
1326                                         for (j = 0; j < vcnt ; j++) {
1327                                                 struct bio_vec *bi;
1328                                                 bi = &sbio->bi_io_vec[j];
1329                                                 bi->bv_offset = 0;
1330                                                 if (size > PAGE_SIZE)
1331                                                         bi->bv_len = PAGE_SIZE;
1332                                                 else
1333                                                         bi->bv_len = size;
1334                                                 size -= PAGE_SIZE;
1335                                                 memcpy(page_address(bi->bv_page),
1336                                                        page_address(pbio->bi_io_vec[j].bv_page),
1337                                                        PAGE_SIZE);
1338                                         }
1339
1340                                 }
1341                         }
1342         }
1343         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1344                 /* ouch - failed to read all of that.
1345                  * Try some synchronous reads of other devices to get
1346                  * good data, much like with normal read errors.  Only
1347                  * read into the pages we already have so we don't
1348                  * need to re-issue the read request.
1349                  * We don't need to freeze the array, because being in an
1350                  * active sync request, there is no normal IO, and
1351                  * no overlapping syncs.
1352                  */
1353                 sector_t sect = r1_bio->sector;
1354                 int sectors = r1_bio->sectors;
1355                 int idx = 0;
1356
1357                 while(sectors) {
1358                         int s = sectors;
1359                         int d = r1_bio->read_disk;
1360                         int success = 0;
1361                         mdk_rdev_t *rdev;
1362
1363                         if (s > (PAGE_SIZE>>9))
1364                                 s = PAGE_SIZE >> 9;
1365                         do {
1366                                 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1367                                         /* No rcu protection needed here devices
1368                                          * can only be removed when no resync is
1369                                          * active, and resync is currently active
1370                                          */
1371                                         rdev = conf->mirrors[d].rdev;
1372                                         if (sync_page_io(rdev->bdev,
1373                                                          sect + rdev->data_offset,
1374                                                          s<<9,
1375                                                          bio->bi_io_vec[idx].bv_page,
1376                                                          READ)) {
1377                                                 success = 1;
1378                                                 break;
1379                                         }
1380                                 }
1381                                 d++;
1382                                 if (d == conf->raid_disks)
1383                                         d = 0;
1384                         } while (!success && d != r1_bio->read_disk);
1385
1386                         if (success) {
1387                                 int start = d;
1388                                 /* write it back and re-read */
1389                                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1390                                 while (d != r1_bio->read_disk) {
1391                                         if (d == 0)
1392                                                 d = conf->raid_disks;
1393                                         d--;
1394                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1395                                                 continue;
1396                                         rdev = conf->mirrors[d].rdev;
1397                                         atomic_add(s, &rdev->corrected_errors);
1398                                         if (sync_page_io(rdev->bdev,
1399                                                          sect + rdev->data_offset,
1400                                                          s<<9,
1401                                                          bio->bi_io_vec[idx].bv_page,
1402                                                          WRITE) == 0)
1403                                                 md_error(mddev, rdev);
1404                                 }
1405                                 d = start;
1406                                 while (d != r1_bio->read_disk) {
1407                                         if (d == 0)
1408                                                 d = conf->raid_disks;
1409                                         d--;
1410                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1411                                                 continue;
1412                                         rdev = conf->mirrors[d].rdev;
1413                                         if (sync_page_io(rdev->bdev,
1414                                                          sect + rdev->data_offset,
1415                                                          s<<9,
1416                                                          bio->bi_io_vec[idx].bv_page,
1417                                                          READ) == 0)
1418                                                 md_error(mddev, rdev);
1419                                 }
1420                         } else {
1421                                 char b[BDEVNAME_SIZE];
1422                                 /* Cannot read from anywhere, array is toast */
1423                                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1424                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
1425                                        " for block %llu\n",
1426                                        bdevname(bio->bi_bdev,b),
1427                                        (unsigned long long)r1_bio->sector);
1428                                 md_done_sync(mddev, r1_bio->sectors, 0);
1429                                 put_buf(r1_bio);
1430                                 return;
1431                         }
1432                         sectors -= s;
1433                         sect += s;
1434                         idx ++;
1435                 }
1436         }
1437
1438         /*
1439          * schedule writes
1440          */
1441         atomic_set(&r1_bio->remaining, 1);
1442         for (i = 0; i < disks ; i++) {
1443                 wbio = r1_bio->bios[i];
1444                 if (wbio->bi_end_io == NULL ||
1445                     (wbio->bi_end_io == end_sync_read &&
1446                      (i == r1_bio->read_disk ||
1447                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1448                         continue;
1449
1450                 wbio->bi_rw = WRITE;
1451                 wbio->bi_end_io = end_sync_write;
1452                 atomic_inc(&r1_bio->remaining);
1453                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1454
1455                 generic_make_request(wbio);
1456         }
1457
1458         if (atomic_dec_and_test(&r1_bio->remaining)) {
1459                 /* if we're here, all write(s) have completed, so clean up */
1460                 md_done_sync(mddev, r1_bio->sectors, 1);
1461                 put_buf(r1_bio);
1462         }
1463 }
1464
1465 /*
1466  * This is a kernel thread which:
1467  *
1468  *      1.      Retries failed read operations on working mirrors.
1469  *      2.      Updates the raid superblock when problems encounter.
1470  *      3.      Performs writes following reads for array syncronising.
1471  */
1472
1473 static void fix_read_error(conf_t *conf, int read_disk,
1474                            sector_t sect, int sectors)
1475 {
1476         mddev_t *mddev = conf->mddev;
1477         while(sectors) {
1478                 int s = sectors;
1479                 int d = read_disk;
1480                 int success = 0;
1481                 int start;
1482                 mdk_rdev_t *rdev;
1483
1484                 if (s > (PAGE_SIZE>>9))
1485                         s = PAGE_SIZE >> 9;
1486
1487                 do {
1488                         /* Note: no rcu protection needed here
1489                          * as this is synchronous in the raid1d thread
1490                          * which is the thread that might remove
1491                          * a device.  If raid1d ever becomes multi-threaded....
1492                          */
1493                         rdev = conf->mirrors[d].rdev;
1494                         if (rdev &&
1495                             test_bit(In_sync, &rdev->flags) &&
1496                             sync_page_io(rdev->bdev,
1497                                          sect + rdev->data_offset,
1498                                          s<<9,
1499                                          conf->tmppage, READ))
1500                                 success = 1;
1501                         else {
1502                                 d++;
1503                                 if (d == conf->raid_disks)
1504                                         d = 0;
1505                         }
1506                 } while (!success && d != read_disk);
1507
1508                 if (!success) {
1509                         /* Cannot read from anywhere -- bye bye array */
1510                         md_error(mddev, conf->mirrors[read_disk].rdev);
1511                         break;
1512                 }
1513                 /* write it back and re-read */
1514                 start = d;
1515                 while (d != read_disk) {
1516                         if (d==0)
1517                                 d = conf->raid_disks;
1518                         d--;
1519                         rdev = conf->mirrors[d].rdev;
1520                         if (rdev &&
1521                             test_bit(In_sync, &rdev->flags)) {
1522                                 if (sync_page_io(rdev->bdev,
1523                                                  sect + rdev->data_offset,
1524                                                  s<<9, conf->tmppage, WRITE)
1525                                     == 0)
1526                                         /* Well, this device is dead */
1527                                         md_error(mddev, rdev);
1528                         }
1529                 }
1530                 d = start;
1531                 while (d != read_disk) {
1532                         char b[BDEVNAME_SIZE];
1533                         if (d==0)
1534                                 d = conf->raid_disks;
1535                         d--;
1536                         rdev = conf->mirrors[d].rdev;
1537                         if (rdev &&
1538                             test_bit(In_sync, &rdev->flags)) {
1539                                 if (sync_page_io(rdev->bdev,
1540                                                  sect + rdev->data_offset,
1541                                                  s<<9, conf->tmppage, READ)
1542                                     == 0)
1543                                         /* Well, this device is dead */
1544                                         md_error(mddev, rdev);
1545                                 else {
1546                                         atomic_add(s, &rdev->corrected_errors);
1547                                         printk(KERN_INFO
1548                                                "raid1:%s: read error corrected "
1549                                                "(%d sectors at %llu on %s)\n",
1550                                                mdname(mddev), s,
1551                                                (unsigned long long)(sect +
1552                                                    rdev->data_offset),
1553                                                bdevname(rdev->bdev, b));
1554                                 }
1555                         }
1556                 }
1557                 sectors -= s;
1558                 sect += s;
1559         }
1560 }
1561
1562 static void raid1d(mddev_t *mddev)
1563 {
1564         r1bio_t *r1_bio;
1565         struct bio *bio;
1566         unsigned long flags;
1567         conf_t *conf = mddev->private;
1568         struct list_head *head = &conf->retry_list;
1569         int unplug=0;
1570         mdk_rdev_t *rdev;
1571
1572         md_check_recovery(mddev);
1573         
1574         for (;;) {
1575                 char b[BDEVNAME_SIZE];
1576
1577                 unplug += flush_pending_writes(conf);
1578
1579                 spin_lock_irqsave(&conf->device_lock, flags);
1580                 if (list_empty(head)) {
1581                         spin_unlock_irqrestore(&conf->device_lock, flags);
1582                         break;
1583                 }
1584                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1585                 list_del(head->prev);
1586                 conf->nr_queued--;
1587                 spin_unlock_irqrestore(&conf->device_lock, flags);
1588
1589                 mddev = r1_bio->mddev;
1590                 conf = mddev->private;
1591                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1592                         sync_request_write(mddev, r1_bio);
1593                         unplug = 1;
1594                 } else if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
1595                         /* some requests in the r1bio were BIO_RW_BARRIER
1596                          * requests which failed with -EOPNOTSUPP.  Hohumm..
1597                          * Better resubmit without the barrier.
1598                          * We know which devices to resubmit for, because
1599                          * all others have had their bios[] entry cleared.
1600                          * We already have a nr_pending reference on these rdevs.
1601                          */
1602                         int i;
1603                         const int do_sync = bio_sync(r1_bio->master_bio);
1604                         clear_bit(R1BIO_BarrierRetry, &r1_bio->state);
1605                         clear_bit(R1BIO_Barrier, &r1_bio->state);
1606                         for (i=0; i < conf->raid_disks; i++)
1607                                 if (r1_bio->bios[i])
1608                                         atomic_inc(&r1_bio->remaining);
1609                         for (i=0; i < conf->raid_disks; i++)
1610                                 if (r1_bio->bios[i]) {
1611                                         struct bio_vec *bvec;
1612                                         int j;
1613
1614                                         bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1615                                         /* copy pages from the failed bio, as
1616                                          * this might be a write-behind device */
1617                                         __bio_for_each_segment(bvec, bio, j, 0)
1618                                                 bvec->bv_page = bio_iovec_idx(r1_bio->bios[i], j)->bv_page;
1619                                         bio_put(r1_bio->bios[i]);
1620                                         bio->bi_sector = r1_bio->sector +
1621                                                 conf->mirrors[i].rdev->data_offset;
1622                                         bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1623                                         bio->bi_end_io = raid1_end_write_request;
1624                                         bio->bi_rw = WRITE | do_sync;
1625                                         bio->bi_private = r1_bio;
1626                                         r1_bio->bios[i] = bio;
1627                                         generic_make_request(bio);
1628                                 }
1629                 } else {
1630                         int disk;
1631
1632                         /* we got a read error. Maybe the drive is bad.  Maybe just
1633                          * the block and we can fix it.
1634                          * We freeze all other IO, and try reading the block from
1635                          * other devices.  When we find one, we re-write
1636                          * and check it that fixes the read error.
1637                          * This is all done synchronously while the array is
1638                          * frozen
1639                          */
1640                         if (mddev->ro == 0) {
1641                                 freeze_array(conf);
1642                                 fix_read_error(conf, r1_bio->read_disk,
1643                                                r1_bio->sector,
1644                                                r1_bio->sectors);
1645                                 unfreeze_array(conf);
1646                         }
1647
1648                         bio = r1_bio->bios[r1_bio->read_disk];
1649                         if ((disk=read_balance(conf, r1_bio)) == -1 ||
1650                             disk == r1_bio->read_disk) {
1651                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
1652                                        " read error for block %llu\n",
1653                                        bdevname(bio->bi_bdev,b),
1654                                        (unsigned long long)r1_bio->sector);
1655                                 raid_end_bio_io(r1_bio);
1656                         } else {
1657                                 const int do_sync = bio_sync(r1_bio->master_bio);
1658                                 r1_bio->bios[r1_bio->read_disk] =
1659                                         mddev->ro ? IO_BLOCKED : NULL;
1660                                 r1_bio->read_disk = disk;
1661                                 bio_put(bio);
1662                                 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1663                                 r1_bio->bios[r1_bio->read_disk] = bio;
1664                                 rdev = conf->mirrors[disk].rdev;
1665                                 if (printk_ratelimit())
1666                                         printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
1667                                                " another mirror\n",
1668                                                bdevname(rdev->bdev,b),
1669                                                (unsigned long long)r1_bio->sector);
1670                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1671                                 bio->bi_bdev = rdev->bdev;
1672                                 bio->bi_end_io = raid1_end_read_request;
1673                                 bio->bi_rw = READ | do_sync;
1674                                 bio->bi_private = r1_bio;
1675                                 unplug = 1;
1676                                 generic_make_request(bio);
1677                         }
1678                 }
1679         }
1680         if (unplug)
1681                 unplug_slaves(mddev);
1682 }
1683
1684
1685 static int init_resync(conf_t *conf)
1686 {
1687         int buffs;
1688
1689         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1690         BUG_ON(conf->r1buf_pool);
1691         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1692                                           conf->poolinfo);
1693         if (!conf->r1buf_pool)
1694                 return -ENOMEM;
1695         conf->next_resync = 0;
1696         return 0;
1697 }
1698
1699 /*
1700  * perform a "sync" on one "block"
1701  *
1702  * We need to make sure that no normal I/O request - particularly write
1703  * requests - conflict with active sync requests.
1704  *
1705  * This is achieved by tracking pending requests and a 'barrier' concept
1706  * that can be installed to exclude normal IO requests.
1707  */
1708
1709 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1710 {
1711         conf_t *conf = mddev->private;
1712         r1bio_t *r1_bio;
1713         struct bio *bio;
1714         sector_t max_sector, nr_sectors;
1715         int disk = -1;
1716         int i;
1717         int wonly = -1;
1718         int write_targets = 0, read_targets = 0;
1719         int sync_blocks;
1720         int still_degraded = 0;
1721
1722         if (!conf->r1buf_pool)
1723         {
1724 /*
1725                 printk("sync start - bitmap %p\n", mddev->bitmap);
1726 */
1727                 if (init_resync(conf))
1728                         return 0;
1729         }
1730
1731         max_sector = mddev->dev_sectors;
1732         if (sector_nr >= max_sector) {
1733                 /* If we aborted, we need to abort the
1734                  * sync on the 'current' bitmap chunk (there will
1735                  * only be one in raid1 resync.
1736                  * We can find the current addess in mddev->curr_resync
1737                  */
1738                 if (mddev->curr_resync < max_sector) /* aborted */
1739                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1740                                                 &sync_blocks, 1);
1741                 else /* completed sync */
1742                         conf->fullsync = 0;
1743
1744                 bitmap_close_sync(mddev->bitmap);
1745                 close_sync(conf);
1746                 return 0;
1747         }
1748
1749         if (mddev->bitmap == NULL &&
1750             mddev->recovery_cp == MaxSector &&
1751             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1752             conf->fullsync == 0) {
1753                 *skipped = 1;
1754                 return max_sector - sector_nr;
1755         }
1756         /* before building a request, check if we can skip these blocks..
1757          * This call the bitmap_start_sync doesn't actually record anything
1758          */
1759         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1760             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1761                 /* We can skip this block, and probably several more */
1762                 *skipped = 1;
1763                 return sync_blocks;
1764         }
1765         /*
1766          * If there is non-resync activity waiting for a turn,
1767          * and resync is going fast enough,
1768          * then let it though before starting on this new sync request.
1769          */
1770         if (!go_faster && conf->nr_waiting)
1771                 msleep_interruptible(1000);
1772
1773         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1774         raise_barrier(conf);
1775
1776         conf->next_resync = sector_nr;
1777
1778         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1779         rcu_read_lock();
1780         /*
1781          * If we get a correctably read error during resync or recovery,
1782          * we might want to read from a different device.  So we
1783          * flag all drives that could conceivably be read from for READ,
1784          * and any others (which will be non-In_sync devices) for WRITE.
1785          * If a read fails, we try reading from something else for which READ
1786          * is OK.
1787          */
1788
1789         r1_bio->mddev = mddev;
1790         r1_bio->sector = sector_nr;
1791         r1_bio->state = 0;
1792         set_bit(R1BIO_IsSync, &r1_bio->state);
1793
1794         for (i=0; i < conf->raid_disks; i++) {
1795                 mdk_rdev_t *rdev;
1796                 bio = r1_bio->bios[i];
1797
1798                 /* take from bio_init */
1799                 bio->bi_next = NULL;
1800                 bio->bi_flags |= 1 << BIO_UPTODATE;
1801                 bio->bi_rw = READ;
1802                 bio->bi_vcnt = 0;
1803                 bio->bi_idx = 0;
1804                 bio->bi_phys_segments = 0;
1805                 bio->bi_size = 0;
1806                 bio->bi_end_io = NULL;
1807                 bio->bi_private = NULL;
1808
1809                 rdev = rcu_dereference(conf->mirrors[i].rdev);
1810                 if (rdev == NULL ||
1811                            test_bit(Faulty, &rdev->flags)) {
1812                         still_degraded = 1;
1813                         continue;
1814                 } else if (!test_bit(In_sync, &rdev->flags)) {
1815                         bio->bi_rw = WRITE;
1816                         bio->bi_end_io = end_sync_write;
1817                         write_targets ++;
1818                 } else {
1819                         /* may need to read from here */
1820                         bio->bi_rw = READ;
1821                         bio->bi_end_io = end_sync_read;
1822                         if (test_bit(WriteMostly, &rdev->flags)) {
1823                                 if (wonly < 0)
1824                                         wonly = i;
1825                         } else {
1826                                 if (disk < 0)
1827                                         disk = i;
1828                         }
1829                         read_targets++;
1830                 }
1831                 atomic_inc(&rdev->nr_pending);
1832                 bio->bi_sector = sector_nr + rdev->data_offset;
1833                 bio->bi_bdev = rdev->bdev;
1834                 bio->bi_private = r1_bio;
1835         }
1836         rcu_read_unlock();
1837         if (disk < 0)
1838                 disk = wonly;
1839         r1_bio->read_disk = disk;
1840
1841         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1842                 /* extra read targets are also write targets */
1843                 write_targets += read_targets-1;
1844
1845         if (write_targets == 0 || read_targets == 0) {
1846                 /* There is nowhere to write, so all non-sync
1847                  * drives must be failed - so we are finished
1848                  */
1849                 sector_t rv = max_sector - sector_nr;
1850                 *skipped = 1;
1851                 put_buf(r1_bio);
1852                 return rv;
1853         }
1854
1855         if (max_sector > mddev->resync_max)
1856                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1857         nr_sectors = 0;
1858         sync_blocks = 0;
1859         do {
1860                 struct page *page;
1861                 int len = PAGE_SIZE;
1862                 if (sector_nr + (len>>9) > max_sector)
1863                         len = (max_sector - sector_nr) << 9;
1864                 if (len == 0)
1865                         break;
1866                 if (sync_blocks == 0) {
1867                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1868                                                &sync_blocks, still_degraded) &&
1869                             !conf->fullsync &&
1870                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1871                                 break;
1872                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1873                         if (len > (sync_blocks<<9))
1874                                 len = sync_blocks<<9;
1875                 }
1876
1877                 for (i=0 ; i < conf->raid_disks; i++) {
1878                         bio = r1_bio->bios[i];
1879                         if (bio->bi_end_io) {
1880                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1881                                 if (bio_add_page(bio, page, len, 0) == 0) {
1882                                         /* stop here */
1883                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1884                                         while (i > 0) {
1885                                                 i--;
1886                                                 bio = r1_bio->bios[i];
1887                                                 if (bio->bi_end_io==NULL)
1888                                                         continue;
1889                                                 /* remove last page from this bio */
1890                                                 bio->bi_vcnt--;
1891                                                 bio->bi_size -= len;
1892                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1893                                         }
1894                                         goto bio_full;
1895                                 }
1896                         }
1897                 }
1898                 nr_sectors += len>>9;
1899                 sector_nr += len>>9;
1900                 sync_blocks -= (len>>9);
1901         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1902  bio_full:
1903         r1_bio->sectors = nr_sectors;
1904
1905         /* For a user-requested sync, we read all readable devices and do a
1906          * compare
1907          */
1908         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1909                 atomic_set(&r1_bio->remaining, read_targets);
1910                 for (i=0; i<conf->raid_disks; i++) {
1911                         bio = r1_bio->bios[i];
1912                         if (bio->bi_end_io == end_sync_read) {
1913                                 md_sync_acct(bio->bi_bdev, nr_sectors);
1914                                 generic_make_request(bio);
1915                         }
1916                 }
1917         } else {
1918                 atomic_set(&r1_bio->remaining, 1);
1919                 bio = r1_bio->bios[r1_bio->read_disk];
1920                 md_sync_acct(bio->bi_bdev, nr_sectors);
1921                 generic_make_request(bio);
1922
1923         }
1924         return nr_sectors;
1925 }
1926
1927 static sector_t raid1_size(mddev_t *mddev, sector_t sectors, int raid_disks)
1928 {
1929         if (sectors)
1930                 return sectors;
1931
1932         return mddev->dev_sectors;
1933 }
1934
1935 static int run(mddev_t *mddev)
1936 {
1937         conf_t *conf;
1938         int i, j, disk_idx;
1939         mirror_info_t *disk;
1940         mdk_rdev_t *rdev;
1941
1942         if (mddev->level != 1) {
1943                 printk("raid1: %s: raid level not set to mirroring (%d)\n",
1944                        mdname(mddev), mddev->level);
1945                 goto out;
1946         }
1947         if (mddev->reshape_position != MaxSector) {
1948                 printk("raid1: %s: reshape_position set but not supported\n",
1949                        mdname(mddev));
1950                 goto out;
1951         }
1952         /*
1953          * copy the already verified devices into our private RAID1
1954          * bookkeeping area. [whatever we allocate in run(),
1955          * should be freed in stop()]
1956          */
1957         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1958         mddev->private = conf;
1959         if (!conf)
1960                 goto out_no_mem;
1961
1962         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1963                                  GFP_KERNEL);
1964         if (!conf->mirrors)
1965                 goto out_no_mem;
1966
1967         conf->tmppage = alloc_page(GFP_KERNEL);
1968         if (!conf->tmppage)
1969                 goto out_no_mem;
1970
1971         conf->poolinfo = kmalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1972         if (!conf->poolinfo)
1973                 goto out_no_mem;
1974         conf->poolinfo->mddev = mddev;
1975         conf->poolinfo->raid_disks = mddev->raid_disks;
1976         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1977                                           r1bio_pool_free,
1978                                           conf->poolinfo);
1979         if (!conf->r1bio_pool)
1980                 goto out_no_mem;
1981
1982         spin_lock_init(&conf->device_lock);
1983         mddev->queue->queue_lock = &conf->device_lock;
1984
1985         list_for_each_entry(rdev, &mddev->disks, same_set) {
1986                 disk_idx = rdev->raid_disk;
1987                 if (disk_idx >= mddev->raid_disks
1988                     || disk_idx < 0)
1989                         continue;
1990                 disk = conf->mirrors + disk_idx;
1991
1992                 disk->rdev = rdev;
1993                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1994                                   rdev->data_offset << 9);
1995                 /* as we don't honour merge_bvec_fn, we must never risk
1996                  * violating it, so limit ->max_sector to one PAGE, as
1997                  * a one page request is never in violation.
1998                  */
1999                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
2000                     queue_max_sectors(mddev->queue) > (PAGE_SIZE>>9))
2001                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
2002
2003                 disk->head_position = 0;
2004         }
2005         conf->raid_disks = mddev->raid_disks;
2006         conf->mddev = mddev;
2007         INIT_LIST_HEAD(&conf->retry_list);
2008
2009         spin_lock_init(&conf->resync_lock);
2010         init_waitqueue_head(&conf->wait_barrier);
2011
2012         bio_list_init(&conf->pending_bio_list);
2013         bio_list_init(&conf->flushing_bio_list);
2014
2015
2016         mddev->degraded = 0;
2017         for (i = 0; i < conf->raid_disks; i++) {
2018
2019                 disk = conf->mirrors + i;
2020
2021                 if (!disk->rdev ||
2022                     !test_bit(In_sync, &disk->rdev->flags)) {
2023                         disk->head_position = 0;
2024                         mddev->degraded++;
2025                         if (disk->rdev)
2026                                 conf->fullsync = 1;
2027                 }
2028         }
2029         if (mddev->degraded == conf->raid_disks) {
2030                 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
2031                         mdname(mddev));
2032                 goto out_free_conf;
2033         }
2034         if (conf->raid_disks - mddev->degraded == 1)
2035                 mddev->recovery_cp = MaxSector;
2036
2037         /*
2038          * find the first working one and use it as a starting point
2039          * to read balancing.
2040          */
2041         for (j = 0; j < conf->raid_disks &&
2042                      (!conf->mirrors[j].rdev ||
2043                       !test_bit(In_sync, &conf->mirrors[j].rdev->flags)) ; j++)
2044                 /* nothing */;
2045         conf->last_used = j;
2046
2047
2048         mddev->thread = md_register_thread(raid1d, mddev, "%s_raid1");
2049         if (!mddev->thread) {
2050                 printk(KERN_ERR
2051                        "raid1: couldn't allocate thread for %s\n",
2052                        mdname(mddev));
2053                 goto out_free_conf;
2054         }
2055
2056         if (mddev->recovery_cp != MaxSector)
2057                 printk(KERN_NOTICE "raid1: %s is not clean"
2058                        " -- starting background reconstruction\n",
2059                        mdname(mddev));
2060         printk(KERN_INFO 
2061                 "raid1: raid set %s active with %d out of %d mirrors\n",
2062                 mdname(mddev), mddev->raid_disks - mddev->degraded, 
2063                 mddev->raid_disks);
2064         /*
2065          * Ok, everything is just fine now
2066          */
2067         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2068
2069         mddev->queue->unplug_fn = raid1_unplug;
2070         mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2071         mddev->queue->backing_dev_info.congested_data = mddev;
2072         md_integrity_register(mddev);
2073         return 0;
2074
2075 out_no_mem:
2076         printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
2077                mdname(mddev));
2078
2079 out_free_conf:
2080         if (conf) {
2081                 if (conf->r1bio_pool)
2082                         mempool_destroy(conf->r1bio_pool);
2083                 kfree(conf->mirrors);
2084                 safe_put_page(conf->tmppage);
2085                 kfree(conf->poolinfo);
2086                 kfree(conf);
2087                 mddev->private = NULL;
2088         }
2089 out:
2090         return -EIO;
2091 }
2092
2093 static int stop(mddev_t *mddev)
2094 {
2095         conf_t *conf = mddev->private;
2096         struct bitmap *bitmap = mddev->bitmap;
2097         int behind_wait = 0;
2098
2099         /* wait for behind writes to complete */
2100         while (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2101                 behind_wait++;
2102                 printk(KERN_INFO "raid1: behind writes in progress on device %s, waiting to stop (%d)\n", mdname(mddev), behind_wait);
2103                 set_current_state(TASK_UNINTERRUPTIBLE);
2104                 schedule_timeout(HZ); /* wait a second */
2105                 /* need to kick something here to make sure I/O goes? */
2106         }
2107
2108         raise_barrier(conf);
2109         lower_barrier(conf);
2110
2111         md_unregister_thread(mddev->thread);
2112         mddev->thread = NULL;
2113         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2114         if (conf->r1bio_pool)
2115                 mempool_destroy(conf->r1bio_pool);
2116         kfree(conf->mirrors);
2117         kfree(conf->poolinfo);
2118         kfree(conf);
2119         mddev->private = NULL;
2120         return 0;
2121 }
2122
2123 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2124 {
2125         /* no resync is happening, and there is enough space
2126          * on all devices, so we can resize.
2127          * We need to make sure resync covers any new space.
2128          * If the array is shrinking we should possibly wait until
2129          * any io in the removed space completes, but it hardly seems
2130          * worth it.
2131          */
2132         md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2133         if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2134                 return -EINVAL;
2135         set_capacity(mddev->gendisk, mddev->array_sectors);
2136         mddev->changed = 1;
2137         revalidate_disk(mddev->gendisk);
2138         if (sectors > mddev->dev_sectors &&
2139             mddev->recovery_cp == MaxSector) {
2140                 mddev->recovery_cp = mddev->dev_sectors;
2141                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2142         }
2143         mddev->dev_sectors = sectors;
2144         mddev->resync_max_sectors = sectors;
2145         return 0;
2146 }
2147
2148 static int raid1_reshape(mddev_t *mddev)
2149 {
2150         /* We need to:
2151          * 1/ resize the r1bio_pool
2152          * 2/ resize conf->mirrors
2153          *
2154          * We allocate a new r1bio_pool if we can.
2155          * Then raise a device barrier and wait until all IO stops.
2156          * Then resize conf->mirrors and swap in the new r1bio pool.
2157          *
2158          * At the same time, we "pack" the devices so that all the missing
2159          * devices have the higher raid_disk numbers.
2160          */
2161         mempool_t *newpool, *oldpool;
2162         struct pool_info *newpoolinfo;
2163         mirror_info_t *newmirrors;
2164         conf_t *conf = mddev->private;
2165         int cnt, raid_disks;
2166         unsigned long flags;
2167         int d, d2, err;
2168
2169         /* Cannot change chunk_size, layout, or level */
2170         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2171             mddev->layout != mddev->new_layout ||
2172             mddev->level != mddev->new_level) {
2173                 mddev->new_chunk_sectors = mddev->chunk_sectors;
2174                 mddev->new_layout = mddev->layout;
2175                 mddev->new_level = mddev->level;
2176                 return -EINVAL;
2177         }
2178
2179         err = md_allow_write(mddev);
2180         if (err)
2181                 return err;
2182
2183         raid_disks = mddev->raid_disks + mddev->delta_disks;
2184
2185         if (raid_disks < conf->raid_disks) {
2186                 cnt=0;
2187                 for (d= 0; d < conf->raid_disks; d++)
2188                         if (conf->mirrors[d].rdev)
2189                                 cnt++;
2190                 if (cnt > raid_disks)
2191                         return -EBUSY;
2192         }
2193
2194         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2195         if (!newpoolinfo)
2196                 return -ENOMEM;
2197         newpoolinfo->mddev = mddev;
2198         newpoolinfo->raid_disks = raid_disks;
2199
2200         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2201                                  r1bio_pool_free, newpoolinfo);
2202         if (!newpool) {
2203                 kfree(newpoolinfo);
2204                 return -ENOMEM;
2205         }
2206         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2207         if (!newmirrors) {
2208                 kfree(newpoolinfo);
2209                 mempool_destroy(newpool);
2210                 return -ENOMEM;
2211         }
2212
2213         raise_barrier(conf);
2214
2215         /* ok, everything is stopped */
2216         oldpool = conf->r1bio_pool;
2217         conf->r1bio_pool = newpool;
2218
2219         for (d = d2 = 0; d < conf->raid_disks; d++) {
2220                 mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2221                 if (rdev && rdev->raid_disk != d2) {
2222                         char nm[20];
2223                         sprintf(nm, "rd%d", rdev->raid_disk);
2224                         sysfs_remove_link(&mddev->kobj, nm);
2225                         rdev->raid_disk = d2;
2226                         sprintf(nm, "rd%d", rdev->raid_disk);
2227                         sysfs_remove_link(&mddev->kobj, nm);
2228                         if (sysfs_create_link(&mddev->kobj,
2229                                               &rdev->kobj, nm))
2230                                 printk(KERN_WARNING
2231                                        "md/raid1: cannot register "
2232                                        "%s for %s\n",
2233                                        nm, mdname(mddev));
2234                 }
2235                 if (rdev)
2236                         newmirrors[d2++].rdev = rdev;
2237         }
2238         kfree(conf->mirrors);
2239         conf->mirrors = newmirrors;
2240         kfree(conf->poolinfo);
2241         conf->poolinfo = newpoolinfo;
2242
2243         spin_lock_irqsave(&conf->device_lock, flags);
2244         mddev->degraded += (raid_disks - conf->raid_disks);
2245         spin_unlock_irqrestore(&conf->device_lock, flags);
2246         conf->raid_disks = mddev->raid_disks = raid_disks;
2247         mddev->delta_disks = 0;
2248
2249         conf->last_used = 0; /* just make sure it is in-range */
2250         lower_barrier(conf);
2251
2252         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2253         md_wakeup_thread(mddev->thread);
2254
2255         mempool_destroy(oldpool);
2256         return 0;
2257 }
2258
2259 static void raid1_quiesce(mddev_t *mddev, int state)
2260 {
2261         conf_t *conf = mddev->private;
2262
2263         switch(state) {
2264         case 1:
2265                 raise_barrier(conf);
2266                 break;
2267         case 0:
2268                 lower_barrier(conf);
2269                 break;
2270         }
2271 }
2272
2273
2274 static struct mdk_personality raid1_personality =
2275 {
2276         .name           = "raid1",
2277         .level          = 1,
2278         .owner          = THIS_MODULE,
2279         .make_request   = make_request,
2280         .run            = run,
2281         .stop           = stop,
2282         .status         = status,
2283         .error_handler  = error,
2284         .hot_add_disk   = raid1_add_disk,
2285         .hot_remove_disk= raid1_remove_disk,
2286         .spare_active   = raid1_spare_active,
2287         .sync_request   = sync_request,
2288         .resize         = raid1_resize,
2289         .size           = raid1_size,
2290         .check_reshape  = raid1_reshape,
2291         .quiesce        = raid1_quiesce,
2292 };
2293
2294 static int __init raid_init(void)
2295 {
2296         return register_md_personality(&raid1_personality);
2297 }
2298
2299 static void raid_exit(void)
2300 {
2301         unregister_md_personality(&raid1_personality);
2302 }
2303
2304 module_init(raid_init);
2305 module_exit(raid_exit);
2306 MODULE_LICENSE("GPL");
2307 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2308 MODULE_ALIAS("md-raid1");
2309 MODULE_ALIAS("md-level-1");