block: remove support for bio remapping from ->make_request
[linux-3.10.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/slab.h>
35 #include <linux/delay.h>
36 #include <linux/blkdev.h>
37 #include <linux/seq_file.h>
38 #include <linux/ratelimit.h>
39 #include "md.h"
40 #include "raid1.h"
41 #include "bitmap.h"
42
43 #define DEBUG 0
44 #define PRINTK(x...) do { if (DEBUG) printk(x); } while (0)
45
46 /*
47  * Number of guaranteed r1bios in case of extreme VM load:
48  */
49 #define NR_RAID1_BIOS 256
50
51
52 static void allow_barrier(conf_t *conf);
53 static void lower_barrier(conf_t *conf);
54
55 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
56 {
57         struct pool_info *pi = data;
58         int size = offsetof(r1bio_t, bios[pi->raid_disks]);
59
60         /* allocate a r1bio with room for raid_disks entries in the bios array */
61         return kzalloc(size, gfp_flags);
62 }
63
64 static void r1bio_pool_free(void *r1_bio, void *data)
65 {
66         kfree(r1_bio);
67 }
68
69 #define RESYNC_BLOCK_SIZE (64*1024)
70 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
71 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
72 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
73 #define RESYNC_WINDOW (2048*1024)
74
75 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
76 {
77         struct pool_info *pi = data;
78         struct page *page;
79         r1bio_t *r1_bio;
80         struct bio *bio;
81         int i, j;
82
83         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
84         if (!r1_bio)
85                 return NULL;
86
87         /*
88          * Allocate bios : 1 for reading, n-1 for writing
89          */
90         for (j = pi->raid_disks ; j-- ; ) {
91                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
92                 if (!bio)
93                         goto out_free_bio;
94                 r1_bio->bios[j] = bio;
95         }
96         /*
97          * Allocate RESYNC_PAGES data pages and attach them to
98          * the first bio.
99          * If this is a user-requested check/repair, allocate
100          * RESYNC_PAGES for each bio.
101          */
102         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
103                 j = pi->raid_disks;
104         else
105                 j = 1;
106         while(j--) {
107                 bio = r1_bio->bios[j];
108                 for (i = 0; i < RESYNC_PAGES; i++) {
109                         page = alloc_page(gfp_flags);
110                         if (unlikely(!page))
111                                 goto out_free_pages;
112
113                         bio->bi_io_vec[i].bv_page = page;
114                         bio->bi_vcnt = i+1;
115                 }
116         }
117         /* If not user-requests, copy the page pointers to all bios */
118         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
119                 for (i=0; i<RESYNC_PAGES ; i++)
120                         for (j=1; j<pi->raid_disks; j++)
121                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
122                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
123         }
124
125         r1_bio->master_bio = NULL;
126
127         return r1_bio;
128
129 out_free_pages:
130         for (j=0 ; j < pi->raid_disks; j++)
131                 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
132                         put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
133         j = -1;
134 out_free_bio:
135         while ( ++j < pi->raid_disks )
136                 bio_put(r1_bio->bios[j]);
137         r1bio_pool_free(r1_bio, data);
138         return NULL;
139 }
140
141 static void r1buf_pool_free(void *__r1_bio, void *data)
142 {
143         struct pool_info *pi = data;
144         int i,j;
145         r1bio_t *r1bio = __r1_bio;
146
147         for (i = 0; i < RESYNC_PAGES; i++)
148                 for (j = pi->raid_disks; j-- ;) {
149                         if (j == 0 ||
150                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
151                             r1bio->bios[0]->bi_io_vec[i].bv_page)
152                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
153                 }
154         for (i=0 ; i < pi->raid_disks; i++)
155                 bio_put(r1bio->bios[i]);
156
157         r1bio_pool_free(r1bio, data);
158 }
159
160 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
161 {
162         int i;
163
164         for (i = 0; i < conf->raid_disks; i++) {
165                 struct bio **bio = r1_bio->bios + i;
166                 if (!BIO_SPECIAL(*bio))
167                         bio_put(*bio);
168                 *bio = NULL;
169         }
170 }
171
172 static void free_r1bio(r1bio_t *r1_bio)
173 {
174         conf_t *conf = r1_bio->mddev->private;
175
176         put_all_bios(conf, r1_bio);
177         mempool_free(r1_bio, conf->r1bio_pool);
178 }
179
180 static void put_buf(r1bio_t *r1_bio)
181 {
182         conf_t *conf = r1_bio->mddev->private;
183         int i;
184
185         for (i=0; i<conf->raid_disks; i++) {
186                 struct bio *bio = r1_bio->bios[i];
187                 if (bio->bi_end_io)
188                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
189         }
190
191         mempool_free(r1_bio, conf->r1buf_pool);
192
193         lower_barrier(conf);
194 }
195
196 static void reschedule_retry(r1bio_t *r1_bio)
197 {
198         unsigned long flags;
199         mddev_t *mddev = r1_bio->mddev;
200         conf_t *conf = mddev->private;
201
202         spin_lock_irqsave(&conf->device_lock, flags);
203         list_add(&r1_bio->retry_list, &conf->retry_list);
204         conf->nr_queued ++;
205         spin_unlock_irqrestore(&conf->device_lock, flags);
206
207         wake_up(&conf->wait_barrier);
208         md_wakeup_thread(mddev->thread);
209 }
210
211 /*
212  * raid_end_bio_io() is called when we have finished servicing a mirrored
213  * operation and are ready to return a success/failure code to the buffer
214  * cache layer.
215  */
216 static void call_bio_endio(r1bio_t *r1_bio)
217 {
218         struct bio *bio = r1_bio->master_bio;
219         int done;
220         conf_t *conf = r1_bio->mddev->private;
221
222         if (bio->bi_phys_segments) {
223                 unsigned long flags;
224                 spin_lock_irqsave(&conf->device_lock, flags);
225                 bio->bi_phys_segments--;
226                 done = (bio->bi_phys_segments == 0);
227                 spin_unlock_irqrestore(&conf->device_lock, flags);
228         } else
229                 done = 1;
230
231         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
232                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
233         if (done) {
234                 bio_endio(bio, 0);
235                 /*
236                  * Wake up any possible resync thread that waits for the device
237                  * to go idle.
238                  */
239                 allow_barrier(conf);
240         }
241 }
242
243 static void raid_end_bio_io(r1bio_t *r1_bio)
244 {
245         struct bio *bio = r1_bio->master_bio;
246
247         /* if nobody has done the final endio yet, do it now */
248         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
249                 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
250                         (bio_data_dir(bio) == WRITE) ? "write" : "read",
251                         (unsigned long long) bio->bi_sector,
252                         (unsigned long long) bio->bi_sector +
253                                 (bio->bi_size >> 9) - 1);
254
255                 call_bio_endio(r1_bio);
256         }
257         free_r1bio(r1_bio);
258 }
259
260 /*
261  * Update disk head position estimator based on IRQ completion info.
262  */
263 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
264 {
265         conf_t *conf = r1_bio->mddev->private;
266
267         conf->mirrors[disk].head_position =
268                 r1_bio->sector + (r1_bio->sectors);
269 }
270
271 static void raid1_end_read_request(struct bio *bio, int error)
272 {
273         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
274         r1bio_t *r1_bio = bio->bi_private;
275         int mirror;
276         conf_t *conf = r1_bio->mddev->private;
277
278         mirror = r1_bio->read_disk;
279         /*
280          * this branch is our 'one mirror IO has finished' event handler:
281          */
282         update_head_pos(mirror, r1_bio);
283
284         if (uptodate)
285                 set_bit(R1BIO_Uptodate, &r1_bio->state);
286         else {
287                 /* If all other devices have failed, we want to return
288                  * the error upwards rather than fail the last device.
289                  * Here we redefine "uptodate" to mean "Don't want to retry"
290                  */
291                 unsigned long flags;
292                 spin_lock_irqsave(&conf->device_lock, flags);
293                 if (r1_bio->mddev->degraded == conf->raid_disks ||
294                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
295                      !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
296                         uptodate = 1;
297                 spin_unlock_irqrestore(&conf->device_lock, flags);
298         }
299
300         if (uptodate)
301                 raid_end_bio_io(r1_bio);
302         else {
303                 /*
304                  * oops, read error:
305                  */
306                 char b[BDEVNAME_SIZE];
307                 printk_ratelimited(
308                         KERN_ERR "md/raid1:%s: %s: "
309                         "rescheduling sector %llu\n",
310                         mdname(conf->mddev),
311                         bdevname(conf->mirrors[mirror].rdev->bdev,
312                                  b),
313                         (unsigned long long)r1_bio->sector);
314                 set_bit(R1BIO_ReadError, &r1_bio->state);
315                 reschedule_retry(r1_bio);
316         }
317
318         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
319 }
320
321 static void close_write(r1bio_t *r1_bio)
322 {
323         /* it really is the end of this request */
324         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
325                 /* free extra copy of the data pages */
326                 int i = r1_bio->behind_page_count;
327                 while (i--)
328                         safe_put_page(r1_bio->behind_bvecs[i].bv_page);
329                 kfree(r1_bio->behind_bvecs);
330                 r1_bio->behind_bvecs = NULL;
331         }
332         /* clear the bitmap if all writes complete successfully */
333         bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
334                         r1_bio->sectors,
335                         !test_bit(R1BIO_Degraded, &r1_bio->state),
336                         test_bit(R1BIO_BehindIO, &r1_bio->state));
337         md_write_end(r1_bio->mddev);
338 }
339
340 static void r1_bio_write_done(r1bio_t *r1_bio)
341 {
342         if (!atomic_dec_and_test(&r1_bio->remaining))
343                 return;
344
345         if (test_bit(R1BIO_WriteError, &r1_bio->state))
346                 reschedule_retry(r1_bio);
347         else {
348                 close_write(r1_bio);
349                 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
350                         reschedule_retry(r1_bio);
351                 else
352                         raid_end_bio_io(r1_bio);
353         }
354 }
355
356 static void raid1_end_write_request(struct bio *bio, int error)
357 {
358         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
359         r1bio_t *r1_bio = bio->bi_private;
360         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
361         conf_t *conf = r1_bio->mddev->private;
362         struct bio *to_put = NULL;
363
364
365         for (mirror = 0; mirror < conf->raid_disks; mirror++)
366                 if (r1_bio->bios[mirror] == bio)
367                         break;
368
369         /*
370          * 'one mirror IO has finished' event handler:
371          */
372         if (!uptodate) {
373                 set_bit(WriteErrorSeen,
374                         &conf->mirrors[mirror].rdev->flags);
375                 set_bit(R1BIO_WriteError, &r1_bio->state);
376         } else {
377                 /*
378                  * Set R1BIO_Uptodate in our master bio, so that we
379                  * will return a good error code for to the higher
380                  * levels even if IO on some other mirrored buffer
381                  * fails.
382                  *
383                  * The 'master' represents the composite IO operation
384                  * to user-side. So if something waits for IO, then it
385                  * will wait for the 'master' bio.
386                  */
387                 sector_t first_bad;
388                 int bad_sectors;
389
390                 r1_bio->bios[mirror] = NULL;
391                 to_put = bio;
392                 set_bit(R1BIO_Uptodate, &r1_bio->state);
393
394                 /* Maybe we can clear some bad blocks. */
395                 if (is_badblock(conf->mirrors[mirror].rdev,
396                                 r1_bio->sector, r1_bio->sectors,
397                                 &first_bad, &bad_sectors)) {
398                         r1_bio->bios[mirror] = IO_MADE_GOOD;
399                         set_bit(R1BIO_MadeGood, &r1_bio->state);
400                 }
401         }
402
403         update_head_pos(mirror, r1_bio);
404
405         if (behind) {
406                 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
407                         atomic_dec(&r1_bio->behind_remaining);
408
409                 /*
410                  * In behind mode, we ACK the master bio once the I/O
411                  * has safely reached all non-writemostly
412                  * disks. Setting the Returned bit ensures that this
413                  * gets done only once -- we don't ever want to return
414                  * -EIO here, instead we'll wait
415                  */
416                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
417                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
418                         /* Maybe we can return now */
419                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
420                                 struct bio *mbio = r1_bio->master_bio;
421                                 PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
422                                        (unsigned long long) mbio->bi_sector,
423                                        (unsigned long long) mbio->bi_sector +
424                                        (mbio->bi_size >> 9) - 1);
425                                 call_bio_endio(r1_bio);
426                         }
427                 }
428         }
429         if (r1_bio->bios[mirror] == NULL)
430                 rdev_dec_pending(conf->mirrors[mirror].rdev,
431                                  conf->mddev);
432
433         /*
434          * Let's see if all mirrored write operations have finished
435          * already.
436          */
437         r1_bio_write_done(r1_bio);
438
439         if (to_put)
440                 bio_put(to_put);
441 }
442
443
444 /*
445  * This routine returns the disk from which the requested read should
446  * be done. There is a per-array 'next expected sequential IO' sector
447  * number - if this matches on the next IO then we use the last disk.
448  * There is also a per-disk 'last know head position' sector that is
449  * maintained from IRQ contexts, both the normal and the resync IO
450  * completion handlers update this position correctly. If there is no
451  * perfect sequential match then we pick the disk whose head is closest.
452  *
453  * If there are 2 mirrors in the same 2 devices, performance degrades
454  * because position is mirror, not device based.
455  *
456  * The rdev for the device selected will have nr_pending incremented.
457  */
458 static int read_balance(conf_t *conf, r1bio_t *r1_bio, int *max_sectors)
459 {
460         const sector_t this_sector = r1_bio->sector;
461         int sectors;
462         int best_good_sectors;
463         int start_disk;
464         int best_disk;
465         int i;
466         sector_t best_dist;
467         mdk_rdev_t *rdev;
468         int choose_first;
469
470         rcu_read_lock();
471         /*
472          * Check if we can balance. We can balance on the whole
473          * device if no resync is going on, or below the resync window.
474          * We take the first readable disk when above the resync window.
475          */
476  retry:
477         sectors = r1_bio->sectors;
478         best_disk = -1;
479         best_dist = MaxSector;
480         best_good_sectors = 0;
481
482         if (conf->mddev->recovery_cp < MaxSector &&
483             (this_sector + sectors >= conf->next_resync)) {
484                 choose_first = 1;
485                 start_disk = 0;
486         } else {
487                 choose_first = 0;
488                 start_disk = conf->last_used;
489         }
490
491         for (i = 0 ; i < conf->raid_disks ; i++) {
492                 sector_t dist;
493                 sector_t first_bad;
494                 int bad_sectors;
495
496                 int disk = start_disk + i;
497                 if (disk >= conf->raid_disks)
498                         disk -= conf->raid_disks;
499
500                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
501                 if (r1_bio->bios[disk] == IO_BLOCKED
502                     || rdev == NULL
503                     || test_bit(Faulty, &rdev->flags))
504                         continue;
505                 if (!test_bit(In_sync, &rdev->flags) &&
506                     rdev->recovery_offset < this_sector + sectors)
507                         continue;
508                 if (test_bit(WriteMostly, &rdev->flags)) {
509                         /* Don't balance among write-mostly, just
510                          * use the first as a last resort */
511                         if (best_disk < 0)
512                                 best_disk = disk;
513                         continue;
514                 }
515                 /* This is a reasonable device to use.  It might
516                  * even be best.
517                  */
518                 if (is_badblock(rdev, this_sector, sectors,
519                                 &first_bad, &bad_sectors)) {
520                         if (best_dist < MaxSector)
521                                 /* already have a better device */
522                                 continue;
523                         if (first_bad <= this_sector) {
524                                 /* cannot read here. If this is the 'primary'
525                                  * device, then we must not read beyond
526                                  * bad_sectors from another device..
527                                  */
528                                 bad_sectors -= (this_sector - first_bad);
529                                 if (choose_first && sectors > bad_sectors)
530                                         sectors = bad_sectors;
531                                 if (best_good_sectors > sectors)
532                                         best_good_sectors = sectors;
533
534                         } else {
535                                 sector_t good_sectors = first_bad - this_sector;
536                                 if (good_sectors > best_good_sectors) {
537                                         best_good_sectors = good_sectors;
538                                         best_disk = disk;
539                                 }
540                                 if (choose_first)
541                                         break;
542                         }
543                         continue;
544                 } else
545                         best_good_sectors = sectors;
546
547                 dist = abs(this_sector - conf->mirrors[disk].head_position);
548                 if (choose_first
549                     /* Don't change to another disk for sequential reads */
550                     || conf->next_seq_sect == this_sector
551                     || dist == 0
552                     /* If device is idle, use it */
553                     || atomic_read(&rdev->nr_pending) == 0) {
554                         best_disk = disk;
555                         break;
556                 }
557                 if (dist < best_dist) {
558                         best_dist = dist;
559                         best_disk = disk;
560                 }
561         }
562
563         if (best_disk >= 0) {
564                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
565                 if (!rdev)
566                         goto retry;
567                 atomic_inc(&rdev->nr_pending);
568                 if (test_bit(Faulty, &rdev->flags)) {
569                         /* cannot risk returning a device that failed
570                          * before we inc'ed nr_pending
571                          */
572                         rdev_dec_pending(rdev, conf->mddev);
573                         goto retry;
574                 }
575                 sectors = best_good_sectors;
576                 conf->next_seq_sect = this_sector + sectors;
577                 conf->last_used = best_disk;
578         }
579         rcu_read_unlock();
580         *max_sectors = sectors;
581
582         return best_disk;
583 }
584
585 int md_raid1_congested(mddev_t *mddev, int bits)
586 {
587         conf_t *conf = mddev->private;
588         int i, ret = 0;
589
590         rcu_read_lock();
591         for (i = 0; i < mddev->raid_disks; i++) {
592                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
593                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
594                         struct request_queue *q = bdev_get_queue(rdev->bdev);
595
596                         BUG_ON(!q);
597
598                         /* Note the '|| 1' - when read_balance prefers
599                          * non-congested targets, it can be removed
600                          */
601                         if ((bits & (1<<BDI_async_congested)) || 1)
602                                 ret |= bdi_congested(&q->backing_dev_info, bits);
603                         else
604                                 ret &= bdi_congested(&q->backing_dev_info, bits);
605                 }
606         }
607         rcu_read_unlock();
608         return ret;
609 }
610 EXPORT_SYMBOL_GPL(md_raid1_congested);
611
612 static int raid1_congested(void *data, int bits)
613 {
614         mddev_t *mddev = data;
615
616         return mddev_congested(mddev, bits) ||
617                 md_raid1_congested(mddev, bits);
618 }
619
620 static void flush_pending_writes(conf_t *conf)
621 {
622         /* Any writes that have been queued but are awaiting
623          * bitmap updates get flushed here.
624          */
625         spin_lock_irq(&conf->device_lock);
626
627         if (conf->pending_bio_list.head) {
628                 struct bio *bio;
629                 bio = bio_list_get(&conf->pending_bio_list);
630                 spin_unlock_irq(&conf->device_lock);
631                 /* flush any pending bitmap writes to
632                  * disk before proceeding w/ I/O */
633                 bitmap_unplug(conf->mddev->bitmap);
634
635                 while (bio) { /* submit pending writes */
636                         struct bio *next = bio->bi_next;
637                         bio->bi_next = NULL;
638                         generic_make_request(bio);
639                         bio = next;
640                 }
641         } else
642                 spin_unlock_irq(&conf->device_lock);
643 }
644
645 /* Barriers....
646  * Sometimes we need to suspend IO while we do something else,
647  * either some resync/recovery, or reconfigure the array.
648  * To do this we raise a 'barrier'.
649  * The 'barrier' is a counter that can be raised multiple times
650  * to count how many activities are happening which preclude
651  * normal IO.
652  * We can only raise the barrier if there is no pending IO.
653  * i.e. if nr_pending == 0.
654  * We choose only to raise the barrier if no-one is waiting for the
655  * barrier to go down.  This means that as soon as an IO request
656  * is ready, no other operations which require a barrier will start
657  * until the IO request has had a chance.
658  *
659  * So: regular IO calls 'wait_barrier'.  When that returns there
660  *    is no backgroup IO happening,  It must arrange to call
661  *    allow_barrier when it has finished its IO.
662  * backgroup IO calls must call raise_barrier.  Once that returns
663  *    there is no normal IO happeing.  It must arrange to call
664  *    lower_barrier when the particular background IO completes.
665  */
666 #define RESYNC_DEPTH 32
667
668 static void raise_barrier(conf_t *conf)
669 {
670         spin_lock_irq(&conf->resync_lock);
671
672         /* Wait until no block IO is waiting */
673         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
674                             conf->resync_lock, );
675
676         /* block any new IO from starting */
677         conf->barrier++;
678
679         /* Now wait for all pending IO to complete */
680         wait_event_lock_irq(conf->wait_barrier,
681                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
682                             conf->resync_lock, );
683
684         spin_unlock_irq(&conf->resync_lock);
685 }
686
687 static void lower_barrier(conf_t *conf)
688 {
689         unsigned long flags;
690         BUG_ON(conf->barrier <= 0);
691         spin_lock_irqsave(&conf->resync_lock, flags);
692         conf->barrier--;
693         spin_unlock_irqrestore(&conf->resync_lock, flags);
694         wake_up(&conf->wait_barrier);
695 }
696
697 static void wait_barrier(conf_t *conf)
698 {
699         spin_lock_irq(&conf->resync_lock);
700         if (conf->barrier) {
701                 conf->nr_waiting++;
702                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
703                                     conf->resync_lock,
704                                     );
705                 conf->nr_waiting--;
706         }
707         conf->nr_pending++;
708         spin_unlock_irq(&conf->resync_lock);
709 }
710
711 static void allow_barrier(conf_t *conf)
712 {
713         unsigned long flags;
714         spin_lock_irqsave(&conf->resync_lock, flags);
715         conf->nr_pending--;
716         spin_unlock_irqrestore(&conf->resync_lock, flags);
717         wake_up(&conf->wait_barrier);
718 }
719
720 static void freeze_array(conf_t *conf)
721 {
722         /* stop syncio and normal IO and wait for everything to
723          * go quite.
724          * We increment barrier and nr_waiting, and then
725          * wait until nr_pending match nr_queued+1
726          * This is called in the context of one normal IO request
727          * that has failed. Thus any sync request that might be pending
728          * will be blocked by nr_pending, and we need to wait for
729          * pending IO requests to complete or be queued for re-try.
730          * Thus the number queued (nr_queued) plus this request (1)
731          * must match the number of pending IOs (nr_pending) before
732          * we continue.
733          */
734         spin_lock_irq(&conf->resync_lock);
735         conf->barrier++;
736         conf->nr_waiting++;
737         wait_event_lock_irq(conf->wait_barrier,
738                             conf->nr_pending == conf->nr_queued+1,
739                             conf->resync_lock,
740                             flush_pending_writes(conf));
741         spin_unlock_irq(&conf->resync_lock);
742 }
743 static void unfreeze_array(conf_t *conf)
744 {
745         /* reverse the effect of the freeze */
746         spin_lock_irq(&conf->resync_lock);
747         conf->barrier--;
748         conf->nr_waiting--;
749         wake_up(&conf->wait_barrier);
750         spin_unlock_irq(&conf->resync_lock);
751 }
752
753
754 /* duplicate the data pages for behind I/O 
755  */
756 static void alloc_behind_pages(struct bio *bio, r1bio_t *r1_bio)
757 {
758         int i;
759         struct bio_vec *bvec;
760         struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
761                                         GFP_NOIO);
762         if (unlikely(!bvecs))
763                 return;
764
765         bio_for_each_segment(bvec, bio, i) {
766                 bvecs[i] = *bvec;
767                 bvecs[i].bv_page = alloc_page(GFP_NOIO);
768                 if (unlikely(!bvecs[i].bv_page))
769                         goto do_sync_io;
770                 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
771                        kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
772                 kunmap(bvecs[i].bv_page);
773                 kunmap(bvec->bv_page);
774         }
775         r1_bio->behind_bvecs = bvecs;
776         r1_bio->behind_page_count = bio->bi_vcnt;
777         set_bit(R1BIO_BehindIO, &r1_bio->state);
778         return;
779
780 do_sync_io:
781         for (i = 0; i < bio->bi_vcnt; i++)
782                 if (bvecs[i].bv_page)
783                         put_page(bvecs[i].bv_page);
784         kfree(bvecs);
785         PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
786 }
787
788 static void make_request(mddev_t *mddev, struct bio * bio)
789 {
790         conf_t *conf = mddev->private;
791         mirror_info_t *mirror;
792         r1bio_t *r1_bio;
793         struct bio *read_bio;
794         int i, disks;
795         struct bitmap *bitmap;
796         unsigned long flags;
797         const int rw = bio_data_dir(bio);
798         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
799         const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
800         mdk_rdev_t *blocked_rdev;
801         int plugged;
802         int first_clone;
803         int sectors_handled;
804         int max_sectors;
805
806         /*
807          * Register the new request and wait if the reconstruction
808          * thread has put up a bar for new requests.
809          * Continue immediately if no resync is active currently.
810          */
811
812         md_write_start(mddev, bio); /* wait on superblock update early */
813
814         if (bio_data_dir(bio) == WRITE &&
815             bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
816             bio->bi_sector < mddev->suspend_hi) {
817                 /* As the suspend_* range is controlled by
818                  * userspace, we want an interruptible
819                  * wait.
820                  */
821                 DEFINE_WAIT(w);
822                 for (;;) {
823                         flush_signals(current);
824                         prepare_to_wait(&conf->wait_barrier,
825                                         &w, TASK_INTERRUPTIBLE);
826                         if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
827                             bio->bi_sector >= mddev->suspend_hi)
828                                 break;
829                         schedule();
830                 }
831                 finish_wait(&conf->wait_barrier, &w);
832         }
833
834         wait_barrier(conf);
835
836         bitmap = mddev->bitmap;
837
838         /*
839          * make_request() can abort the operation when READA is being
840          * used and no empty request is available.
841          *
842          */
843         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
844
845         r1_bio->master_bio = bio;
846         r1_bio->sectors = bio->bi_size >> 9;
847         r1_bio->state = 0;
848         r1_bio->mddev = mddev;
849         r1_bio->sector = bio->bi_sector;
850
851         /* We might need to issue multiple reads to different
852          * devices if there are bad blocks around, so we keep
853          * track of the number of reads in bio->bi_phys_segments.
854          * If this is 0, there is only one r1_bio and no locking
855          * will be needed when requests complete.  If it is
856          * non-zero, then it is the number of not-completed requests.
857          */
858         bio->bi_phys_segments = 0;
859         clear_bit(BIO_SEG_VALID, &bio->bi_flags);
860
861         if (rw == READ) {
862                 /*
863                  * read balancing logic:
864                  */
865                 int rdisk;
866
867 read_again:
868                 rdisk = read_balance(conf, r1_bio, &max_sectors);
869
870                 if (rdisk < 0) {
871                         /* couldn't find anywhere to read from */
872                         raid_end_bio_io(r1_bio);
873                         return;
874                 }
875                 mirror = conf->mirrors + rdisk;
876
877                 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
878                     bitmap) {
879                         /* Reading from a write-mostly device must
880                          * take care not to over-take any writes
881                          * that are 'behind'
882                          */
883                         wait_event(bitmap->behind_wait,
884                                    atomic_read(&bitmap->behind_writes) == 0);
885                 }
886                 r1_bio->read_disk = rdisk;
887
888                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
889                 md_trim_bio(read_bio, r1_bio->sector - bio->bi_sector,
890                             max_sectors);
891
892                 r1_bio->bios[rdisk] = read_bio;
893
894                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
895                 read_bio->bi_bdev = mirror->rdev->bdev;
896                 read_bio->bi_end_io = raid1_end_read_request;
897                 read_bio->bi_rw = READ | do_sync;
898                 read_bio->bi_private = r1_bio;
899
900                 if (max_sectors < r1_bio->sectors) {
901                         /* could not read all from this device, so we will
902                          * need another r1_bio.
903                          */
904
905                         sectors_handled = (r1_bio->sector + max_sectors
906                                            - bio->bi_sector);
907                         r1_bio->sectors = max_sectors;
908                         spin_lock_irq(&conf->device_lock);
909                         if (bio->bi_phys_segments == 0)
910                                 bio->bi_phys_segments = 2;
911                         else
912                                 bio->bi_phys_segments++;
913                         spin_unlock_irq(&conf->device_lock);
914                         /* Cannot call generic_make_request directly
915                          * as that will be queued in __make_request
916                          * and subsequent mempool_alloc might block waiting
917                          * for it.  So hand bio over to raid1d.
918                          */
919                         reschedule_retry(r1_bio);
920
921                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
922
923                         r1_bio->master_bio = bio;
924                         r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
925                         r1_bio->state = 0;
926                         r1_bio->mddev = mddev;
927                         r1_bio->sector = bio->bi_sector + sectors_handled;
928                         goto read_again;
929                 } else
930                         generic_make_request(read_bio);
931                 return;
932         }
933
934         /*
935          * WRITE:
936          */
937         /* first select target devices under rcu_lock and
938          * inc refcount on their rdev.  Record them by setting
939          * bios[x] to bio
940          * If there are known/acknowledged bad blocks on any device on
941          * which we have seen a write error, we want to avoid writing those
942          * blocks.
943          * This potentially requires several writes to write around
944          * the bad blocks.  Each set of writes gets it's own r1bio
945          * with a set of bios attached.
946          */
947         plugged = mddev_check_plugged(mddev);
948
949         disks = conf->raid_disks;
950  retry_write:
951         blocked_rdev = NULL;
952         rcu_read_lock();
953         max_sectors = r1_bio->sectors;
954         for (i = 0;  i < disks; i++) {
955                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
956                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
957                         atomic_inc(&rdev->nr_pending);
958                         blocked_rdev = rdev;
959                         break;
960                 }
961                 r1_bio->bios[i] = NULL;
962                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
963                         set_bit(R1BIO_Degraded, &r1_bio->state);
964                         continue;
965                 }
966
967                 atomic_inc(&rdev->nr_pending);
968                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
969                         sector_t first_bad;
970                         int bad_sectors;
971                         int is_bad;
972
973                         is_bad = is_badblock(rdev, r1_bio->sector,
974                                              max_sectors,
975                                              &first_bad, &bad_sectors);
976                         if (is_bad < 0) {
977                                 /* mustn't write here until the bad block is
978                                  * acknowledged*/
979                                 set_bit(BlockedBadBlocks, &rdev->flags);
980                                 blocked_rdev = rdev;
981                                 break;
982                         }
983                         if (is_bad && first_bad <= r1_bio->sector) {
984                                 /* Cannot write here at all */
985                                 bad_sectors -= (r1_bio->sector - first_bad);
986                                 if (bad_sectors < max_sectors)
987                                         /* mustn't write more than bad_sectors
988                                          * to other devices yet
989                                          */
990                                         max_sectors = bad_sectors;
991                                 rdev_dec_pending(rdev, mddev);
992                                 /* We don't set R1BIO_Degraded as that
993                                  * only applies if the disk is
994                                  * missing, so it might be re-added,
995                                  * and we want to know to recover this
996                                  * chunk.
997                                  * In this case the device is here,
998                                  * and the fact that this chunk is not
999                                  * in-sync is recorded in the bad
1000                                  * block log
1001                                  */
1002                                 continue;
1003                         }
1004                         if (is_bad) {
1005                                 int good_sectors = first_bad - r1_bio->sector;
1006                                 if (good_sectors < max_sectors)
1007                                         max_sectors = good_sectors;
1008                         }
1009                 }
1010                 r1_bio->bios[i] = bio;
1011         }
1012         rcu_read_unlock();
1013
1014         if (unlikely(blocked_rdev)) {
1015                 /* Wait for this device to become unblocked */
1016                 int j;
1017
1018                 for (j = 0; j < i; j++)
1019                         if (r1_bio->bios[j])
1020                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1021                 r1_bio->state = 0;
1022                 allow_barrier(conf);
1023                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1024                 wait_barrier(conf);
1025                 goto retry_write;
1026         }
1027
1028         if (max_sectors < r1_bio->sectors) {
1029                 /* We are splitting this write into multiple parts, so
1030                  * we need to prepare for allocating another r1_bio.
1031                  */
1032                 r1_bio->sectors = max_sectors;
1033                 spin_lock_irq(&conf->device_lock);
1034                 if (bio->bi_phys_segments == 0)
1035                         bio->bi_phys_segments = 2;
1036                 else
1037                         bio->bi_phys_segments++;
1038                 spin_unlock_irq(&conf->device_lock);
1039         }
1040         sectors_handled = r1_bio->sector + max_sectors - bio->bi_sector;
1041
1042         atomic_set(&r1_bio->remaining, 1);
1043         atomic_set(&r1_bio->behind_remaining, 0);
1044
1045         first_clone = 1;
1046         for (i = 0; i < disks; i++) {
1047                 struct bio *mbio;
1048                 if (!r1_bio->bios[i])
1049                         continue;
1050
1051                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1052                 md_trim_bio(mbio, r1_bio->sector - bio->bi_sector, max_sectors);
1053
1054                 if (first_clone) {
1055                         /* do behind I/O ?
1056                          * Not if there are too many, or cannot
1057                          * allocate memory, or a reader on WriteMostly
1058                          * is waiting for behind writes to flush */
1059                         if (bitmap &&
1060                             (atomic_read(&bitmap->behind_writes)
1061                              < mddev->bitmap_info.max_write_behind) &&
1062                             !waitqueue_active(&bitmap->behind_wait))
1063                                 alloc_behind_pages(mbio, r1_bio);
1064
1065                         bitmap_startwrite(bitmap, r1_bio->sector,
1066                                           r1_bio->sectors,
1067                                           test_bit(R1BIO_BehindIO,
1068                                                    &r1_bio->state));
1069                         first_clone = 0;
1070                 }
1071                 if (r1_bio->behind_bvecs) {
1072                         struct bio_vec *bvec;
1073                         int j;
1074
1075                         /* Yes, I really want the '__' version so that
1076                          * we clear any unused pointer in the io_vec, rather
1077                          * than leave them unchanged.  This is important
1078                          * because when we come to free the pages, we won't
1079                          * know the original bi_idx, so we just free
1080                          * them all
1081                          */
1082                         __bio_for_each_segment(bvec, mbio, j, 0)
1083                                 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1084                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1085                                 atomic_inc(&r1_bio->behind_remaining);
1086                 }
1087
1088                 r1_bio->bios[i] = mbio;
1089
1090                 mbio->bi_sector = (r1_bio->sector +
1091                                    conf->mirrors[i].rdev->data_offset);
1092                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1093                 mbio->bi_end_io = raid1_end_write_request;
1094                 mbio->bi_rw = WRITE | do_flush_fua | do_sync;
1095                 mbio->bi_private = r1_bio;
1096
1097                 atomic_inc(&r1_bio->remaining);
1098                 spin_lock_irqsave(&conf->device_lock, flags);
1099                 bio_list_add(&conf->pending_bio_list, mbio);
1100                 spin_unlock_irqrestore(&conf->device_lock, flags);
1101         }
1102         r1_bio_write_done(r1_bio);
1103
1104         /* In case raid1d snuck in to freeze_array */
1105         wake_up(&conf->wait_barrier);
1106
1107         if (sectors_handled < (bio->bi_size >> 9)) {
1108                 /* We need another r1_bio.  It has already been counted
1109                  * in bio->bi_phys_segments
1110                  */
1111                 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1112                 r1_bio->master_bio = bio;
1113                 r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1114                 r1_bio->state = 0;
1115                 r1_bio->mddev = mddev;
1116                 r1_bio->sector = bio->bi_sector + sectors_handled;
1117                 goto retry_write;
1118         }
1119
1120         if (do_sync || !bitmap || !plugged)
1121                 md_wakeup_thread(mddev->thread);
1122 }
1123
1124 static void status(struct seq_file *seq, mddev_t *mddev)
1125 {
1126         conf_t *conf = mddev->private;
1127         int i;
1128
1129         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1130                    conf->raid_disks - mddev->degraded);
1131         rcu_read_lock();
1132         for (i = 0; i < conf->raid_disks; i++) {
1133                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1134                 seq_printf(seq, "%s",
1135                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1136         }
1137         rcu_read_unlock();
1138         seq_printf(seq, "]");
1139 }
1140
1141
1142 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1143 {
1144         char b[BDEVNAME_SIZE];
1145         conf_t *conf = mddev->private;
1146
1147         /*
1148          * If it is not operational, then we have already marked it as dead
1149          * else if it is the last working disks, ignore the error, let the
1150          * next level up know.
1151          * else mark the drive as failed
1152          */
1153         if (test_bit(In_sync, &rdev->flags)
1154             && (conf->raid_disks - mddev->degraded) == 1) {
1155                 /*
1156                  * Don't fail the drive, act as though we were just a
1157                  * normal single drive.
1158                  * However don't try a recovery from this drive as
1159                  * it is very likely to fail.
1160                  */
1161                 conf->recovery_disabled = mddev->recovery_disabled;
1162                 return;
1163         }
1164         set_bit(Blocked, &rdev->flags);
1165         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1166                 unsigned long flags;
1167                 spin_lock_irqsave(&conf->device_lock, flags);
1168                 mddev->degraded++;
1169                 set_bit(Faulty, &rdev->flags);
1170                 spin_unlock_irqrestore(&conf->device_lock, flags);
1171                 /*
1172                  * if recovery is running, make sure it aborts.
1173                  */
1174                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1175         } else
1176                 set_bit(Faulty, &rdev->flags);
1177         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1178         printk(KERN_ALERT
1179                "md/raid1:%s: Disk failure on %s, disabling device.\n"
1180                "md/raid1:%s: Operation continuing on %d devices.\n",
1181                mdname(mddev), bdevname(rdev->bdev, b),
1182                mdname(mddev), conf->raid_disks - mddev->degraded);
1183 }
1184
1185 static void print_conf(conf_t *conf)
1186 {
1187         int i;
1188
1189         printk(KERN_DEBUG "RAID1 conf printout:\n");
1190         if (!conf) {
1191                 printk(KERN_DEBUG "(!conf)\n");
1192                 return;
1193         }
1194         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1195                 conf->raid_disks);
1196
1197         rcu_read_lock();
1198         for (i = 0; i < conf->raid_disks; i++) {
1199                 char b[BDEVNAME_SIZE];
1200                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1201                 if (rdev)
1202                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1203                                i, !test_bit(In_sync, &rdev->flags),
1204                                !test_bit(Faulty, &rdev->flags),
1205                                bdevname(rdev->bdev,b));
1206         }
1207         rcu_read_unlock();
1208 }
1209
1210 static void close_sync(conf_t *conf)
1211 {
1212         wait_barrier(conf);
1213         allow_barrier(conf);
1214
1215         mempool_destroy(conf->r1buf_pool);
1216         conf->r1buf_pool = NULL;
1217 }
1218
1219 static int raid1_spare_active(mddev_t *mddev)
1220 {
1221         int i;
1222         conf_t *conf = mddev->private;
1223         int count = 0;
1224         unsigned long flags;
1225
1226         /*
1227          * Find all failed disks within the RAID1 configuration 
1228          * and mark them readable.
1229          * Called under mddev lock, so rcu protection not needed.
1230          */
1231         for (i = 0; i < conf->raid_disks; i++) {
1232                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1233                 if (rdev
1234                     && !test_bit(Faulty, &rdev->flags)
1235                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1236                         count++;
1237                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1238                 }
1239         }
1240         spin_lock_irqsave(&conf->device_lock, flags);
1241         mddev->degraded -= count;
1242         spin_unlock_irqrestore(&conf->device_lock, flags);
1243
1244         print_conf(conf);
1245         return count;
1246 }
1247
1248
1249 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1250 {
1251         conf_t *conf = mddev->private;
1252         int err = -EEXIST;
1253         int mirror = 0;
1254         mirror_info_t *p;
1255         int first = 0;
1256         int last = mddev->raid_disks - 1;
1257
1258         if (mddev->recovery_disabled == conf->recovery_disabled)
1259                 return -EBUSY;
1260
1261         if (rdev->raid_disk >= 0)
1262                 first = last = rdev->raid_disk;
1263
1264         for (mirror = first; mirror <= last; mirror++)
1265                 if ( !(p=conf->mirrors+mirror)->rdev) {
1266
1267                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1268                                           rdev->data_offset << 9);
1269                         /* as we don't honour merge_bvec_fn, we must
1270                          * never risk violating it, so limit
1271                          * ->max_segments to one lying with a single
1272                          * page, as a one page request is never in
1273                          * violation.
1274                          */
1275                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1276                                 blk_queue_max_segments(mddev->queue, 1);
1277                                 blk_queue_segment_boundary(mddev->queue,
1278                                                            PAGE_CACHE_SIZE - 1);
1279                         }
1280
1281                         p->head_position = 0;
1282                         rdev->raid_disk = mirror;
1283                         err = 0;
1284                         /* As all devices are equivalent, we don't need a full recovery
1285                          * if this was recently any drive of the array
1286                          */
1287                         if (rdev->saved_raid_disk < 0)
1288                                 conf->fullsync = 1;
1289                         rcu_assign_pointer(p->rdev, rdev);
1290                         break;
1291                 }
1292         md_integrity_add_rdev(rdev, mddev);
1293         print_conf(conf);
1294         return err;
1295 }
1296
1297 static int raid1_remove_disk(mddev_t *mddev, int number)
1298 {
1299         conf_t *conf = mddev->private;
1300         int err = 0;
1301         mdk_rdev_t *rdev;
1302         mirror_info_t *p = conf->mirrors+ number;
1303
1304         print_conf(conf);
1305         rdev = p->rdev;
1306         if (rdev) {
1307                 if (test_bit(In_sync, &rdev->flags) ||
1308                     atomic_read(&rdev->nr_pending)) {
1309                         err = -EBUSY;
1310                         goto abort;
1311                 }
1312                 /* Only remove non-faulty devices if recovery
1313                  * is not possible.
1314                  */
1315                 if (!test_bit(Faulty, &rdev->flags) &&
1316                     mddev->recovery_disabled != conf->recovery_disabled &&
1317                     mddev->degraded < conf->raid_disks) {
1318                         err = -EBUSY;
1319                         goto abort;
1320                 }
1321                 p->rdev = NULL;
1322                 synchronize_rcu();
1323                 if (atomic_read(&rdev->nr_pending)) {
1324                         /* lost the race, try later */
1325                         err = -EBUSY;
1326                         p->rdev = rdev;
1327                         goto abort;
1328                 }
1329                 err = md_integrity_register(mddev);
1330         }
1331 abort:
1332
1333         print_conf(conf);
1334         return err;
1335 }
1336
1337
1338 static void end_sync_read(struct bio *bio, int error)
1339 {
1340         r1bio_t *r1_bio = bio->bi_private;
1341         int i;
1342
1343         for (i=r1_bio->mddev->raid_disks; i--; )
1344                 if (r1_bio->bios[i] == bio)
1345                         break;
1346         BUG_ON(i < 0);
1347         update_head_pos(i, r1_bio);
1348         /*
1349          * we have read a block, now it needs to be re-written,
1350          * or re-read if the read failed.
1351          * We don't do much here, just schedule handling by raid1d
1352          */
1353         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1354                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1355
1356         if (atomic_dec_and_test(&r1_bio->remaining))
1357                 reschedule_retry(r1_bio);
1358 }
1359
1360 static void end_sync_write(struct bio *bio, int error)
1361 {
1362         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1363         r1bio_t *r1_bio = bio->bi_private;
1364         mddev_t *mddev = r1_bio->mddev;
1365         conf_t *conf = mddev->private;
1366         int i;
1367         int mirror=0;
1368         sector_t first_bad;
1369         int bad_sectors;
1370
1371         for (i = 0; i < conf->raid_disks; i++)
1372                 if (r1_bio->bios[i] == bio) {
1373                         mirror = i;
1374                         break;
1375                 }
1376         if (!uptodate) {
1377                 sector_t sync_blocks = 0;
1378                 sector_t s = r1_bio->sector;
1379                 long sectors_to_go = r1_bio->sectors;
1380                 /* make sure these bits doesn't get cleared. */
1381                 do {
1382                         bitmap_end_sync(mddev->bitmap, s,
1383                                         &sync_blocks, 1);
1384                         s += sync_blocks;
1385                         sectors_to_go -= sync_blocks;
1386                 } while (sectors_to_go > 0);
1387                 set_bit(WriteErrorSeen,
1388                         &conf->mirrors[mirror].rdev->flags);
1389                 set_bit(R1BIO_WriteError, &r1_bio->state);
1390         } else if (is_badblock(conf->mirrors[mirror].rdev,
1391                                r1_bio->sector,
1392                                r1_bio->sectors,
1393                                &first_bad, &bad_sectors) &&
1394                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1395                                 r1_bio->sector,
1396                                 r1_bio->sectors,
1397                                 &first_bad, &bad_sectors)
1398                 )
1399                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1400
1401         update_head_pos(mirror, r1_bio);
1402
1403         if (atomic_dec_and_test(&r1_bio->remaining)) {
1404                 int s = r1_bio->sectors;
1405                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1406                     test_bit(R1BIO_WriteError, &r1_bio->state))
1407                         reschedule_retry(r1_bio);
1408                 else {
1409                         put_buf(r1_bio);
1410                         md_done_sync(mddev, s, uptodate);
1411                 }
1412         }
1413 }
1414
1415 static int r1_sync_page_io(mdk_rdev_t *rdev, sector_t sector,
1416                             int sectors, struct page *page, int rw)
1417 {
1418         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1419                 /* success */
1420                 return 1;
1421         if (rw == WRITE)
1422                 set_bit(WriteErrorSeen, &rdev->flags);
1423         /* need to record an error - either for the block or the device */
1424         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1425                 md_error(rdev->mddev, rdev);
1426         return 0;
1427 }
1428
1429 static int fix_sync_read_error(r1bio_t *r1_bio)
1430 {
1431         /* Try some synchronous reads of other devices to get
1432          * good data, much like with normal read errors.  Only
1433          * read into the pages we already have so we don't
1434          * need to re-issue the read request.
1435          * We don't need to freeze the array, because being in an
1436          * active sync request, there is no normal IO, and
1437          * no overlapping syncs.
1438          * We don't need to check is_badblock() again as we
1439          * made sure that anything with a bad block in range
1440          * will have bi_end_io clear.
1441          */
1442         mddev_t *mddev = r1_bio->mddev;
1443         conf_t *conf = mddev->private;
1444         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1445         sector_t sect = r1_bio->sector;
1446         int sectors = r1_bio->sectors;
1447         int idx = 0;
1448
1449         while(sectors) {
1450                 int s = sectors;
1451                 int d = r1_bio->read_disk;
1452                 int success = 0;
1453                 mdk_rdev_t *rdev;
1454                 int start;
1455
1456                 if (s > (PAGE_SIZE>>9))
1457                         s = PAGE_SIZE >> 9;
1458                 do {
1459                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1460                                 /* No rcu protection needed here devices
1461                                  * can only be removed when no resync is
1462                                  * active, and resync is currently active
1463                                  */
1464                                 rdev = conf->mirrors[d].rdev;
1465                                 if (sync_page_io(rdev, sect, s<<9,
1466                                                  bio->bi_io_vec[idx].bv_page,
1467                                                  READ, false)) {
1468                                         success = 1;
1469                                         break;
1470                                 }
1471                         }
1472                         d++;
1473                         if (d == conf->raid_disks)
1474                                 d = 0;
1475                 } while (!success && d != r1_bio->read_disk);
1476
1477                 if (!success) {
1478                         char b[BDEVNAME_SIZE];
1479                         int abort = 0;
1480                         /* Cannot read from anywhere, this block is lost.
1481                          * Record a bad block on each device.  If that doesn't
1482                          * work just disable and interrupt the recovery.
1483                          * Don't fail devices as that won't really help.
1484                          */
1485                         printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1486                                " for block %llu\n",
1487                                mdname(mddev),
1488                                bdevname(bio->bi_bdev, b),
1489                                (unsigned long long)r1_bio->sector);
1490                         for (d = 0; d < conf->raid_disks; d++) {
1491                                 rdev = conf->mirrors[d].rdev;
1492                                 if (!rdev || test_bit(Faulty, &rdev->flags))
1493                                         continue;
1494                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
1495                                         abort = 1;
1496                         }
1497                         if (abort) {
1498                                 mddev->recovery_disabled = 1;
1499                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1500                                 md_done_sync(mddev, r1_bio->sectors, 0);
1501                                 put_buf(r1_bio);
1502                                 return 0;
1503                         }
1504                         /* Try next page */
1505                         sectors -= s;
1506                         sect += s;
1507                         idx++;
1508                         continue;
1509                 }
1510
1511                 start = d;
1512                 /* write it back and re-read */
1513                 while (d != r1_bio->read_disk) {
1514                         if (d == 0)
1515                                 d = conf->raid_disks;
1516                         d--;
1517                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1518                                 continue;
1519                         rdev = conf->mirrors[d].rdev;
1520                         if (r1_sync_page_io(rdev, sect, s,
1521                                             bio->bi_io_vec[idx].bv_page,
1522                                             WRITE) == 0) {
1523                                 r1_bio->bios[d]->bi_end_io = NULL;
1524                                 rdev_dec_pending(rdev, mddev);
1525                         }
1526                 }
1527                 d = start;
1528                 while (d != r1_bio->read_disk) {
1529                         if (d == 0)
1530                                 d = conf->raid_disks;
1531                         d--;
1532                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1533                                 continue;
1534                         rdev = conf->mirrors[d].rdev;
1535                         if (r1_sync_page_io(rdev, sect, s,
1536                                             bio->bi_io_vec[idx].bv_page,
1537                                             READ) != 0)
1538                                 atomic_add(s, &rdev->corrected_errors);
1539                 }
1540                 sectors -= s;
1541                 sect += s;
1542                 idx ++;
1543         }
1544         set_bit(R1BIO_Uptodate, &r1_bio->state);
1545         set_bit(BIO_UPTODATE, &bio->bi_flags);
1546         return 1;
1547 }
1548
1549 static int process_checks(r1bio_t *r1_bio)
1550 {
1551         /* We have read all readable devices.  If we haven't
1552          * got the block, then there is no hope left.
1553          * If we have, then we want to do a comparison
1554          * and skip the write if everything is the same.
1555          * If any blocks failed to read, then we need to
1556          * attempt an over-write
1557          */
1558         mddev_t *mddev = r1_bio->mddev;
1559         conf_t *conf = mddev->private;
1560         int primary;
1561         int i;
1562
1563         for (primary = 0; primary < conf->raid_disks; primary++)
1564                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1565                     test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1566                         r1_bio->bios[primary]->bi_end_io = NULL;
1567                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1568                         break;
1569                 }
1570         r1_bio->read_disk = primary;
1571         for (i = 0; i < conf->raid_disks; i++) {
1572                 int j;
1573                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1574                 struct bio *pbio = r1_bio->bios[primary];
1575                 struct bio *sbio = r1_bio->bios[i];
1576                 int size;
1577
1578                 if (r1_bio->bios[i]->bi_end_io != end_sync_read)
1579                         continue;
1580
1581                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1582                         for (j = vcnt; j-- ; ) {
1583                                 struct page *p, *s;
1584                                 p = pbio->bi_io_vec[j].bv_page;
1585                                 s = sbio->bi_io_vec[j].bv_page;
1586                                 if (memcmp(page_address(p),
1587                                            page_address(s),
1588                                            PAGE_SIZE))
1589                                         break;
1590                         }
1591                 } else
1592                         j = 0;
1593                 if (j >= 0)
1594                         mddev->resync_mismatches += r1_bio->sectors;
1595                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1596                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1597                         /* No need to write to this device. */
1598                         sbio->bi_end_io = NULL;
1599                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1600                         continue;
1601                 }
1602                 /* fixup the bio for reuse */
1603                 sbio->bi_vcnt = vcnt;
1604                 sbio->bi_size = r1_bio->sectors << 9;
1605                 sbio->bi_idx = 0;
1606                 sbio->bi_phys_segments = 0;
1607                 sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1608                 sbio->bi_flags |= 1 << BIO_UPTODATE;
1609                 sbio->bi_next = NULL;
1610                 sbio->bi_sector = r1_bio->sector +
1611                         conf->mirrors[i].rdev->data_offset;
1612                 sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1613                 size = sbio->bi_size;
1614                 for (j = 0; j < vcnt ; j++) {
1615                         struct bio_vec *bi;
1616                         bi = &sbio->bi_io_vec[j];
1617                         bi->bv_offset = 0;
1618                         if (size > PAGE_SIZE)
1619                                 bi->bv_len = PAGE_SIZE;
1620                         else
1621                                 bi->bv_len = size;
1622                         size -= PAGE_SIZE;
1623                         memcpy(page_address(bi->bv_page),
1624                                page_address(pbio->bi_io_vec[j].bv_page),
1625                                PAGE_SIZE);
1626                 }
1627         }
1628         return 0;
1629 }
1630
1631 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1632 {
1633         conf_t *conf = mddev->private;
1634         int i;
1635         int disks = conf->raid_disks;
1636         struct bio *bio, *wbio;
1637
1638         bio = r1_bio->bios[r1_bio->read_disk];
1639
1640         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1641                 /* ouch - failed to read all of that. */
1642                 if (!fix_sync_read_error(r1_bio))
1643                         return;
1644
1645         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1646                 if (process_checks(r1_bio) < 0)
1647                         return;
1648         /*
1649          * schedule writes
1650          */
1651         atomic_set(&r1_bio->remaining, 1);
1652         for (i = 0; i < disks ; i++) {
1653                 wbio = r1_bio->bios[i];
1654                 if (wbio->bi_end_io == NULL ||
1655                     (wbio->bi_end_io == end_sync_read &&
1656                      (i == r1_bio->read_disk ||
1657                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1658                         continue;
1659
1660                 wbio->bi_rw = WRITE;
1661                 wbio->bi_end_io = end_sync_write;
1662                 atomic_inc(&r1_bio->remaining);
1663                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1664
1665                 generic_make_request(wbio);
1666         }
1667
1668         if (atomic_dec_and_test(&r1_bio->remaining)) {
1669                 /* if we're here, all write(s) have completed, so clean up */
1670                 md_done_sync(mddev, r1_bio->sectors, 1);
1671                 put_buf(r1_bio);
1672         }
1673 }
1674
1675 /*
1676  * This is a kernel thread which:
1677  *
1678  *      1.      Retries failed read operations on working mirrors.
1679  *      2.      Updates the raid superblock when problems encounter.
1680  *      3.      Performs writes following reads for array synchronising.
1681  */
1682
1683 static void fix_read_error(conf_t *conf, int read_disk,
1684                            sector_t sect, int sectors)
1685 {
1686         mddev_t *mddev = conf->mddev;
1687         while(sectors) {
1688                 int s = sectors;
1689                 int d = read_disk;
1690                 int success = 0;
1691                 int start;
1692                 mdk_rdev_t *rdev;
1693
1694                 if (s > (PAGE_SIZE>>9))
1695                         s = PAGE_SIZE >> 9;
1696
1697                 do {
1698                         /* Note: no rcu protection needed here
1699                          * as this is synchronous in the raid1d thread
1700                          * which is the thread that might remove
1701                          * a device.  If raid1d ever becomes multi-threaded....
1702                          */
1703                         sector_t first_bad;
1704                         int bad_sectors;
1705
1706                         rdev = conf->mirrors[d].rdev;
1707                         if (rdev &&
1708                             test_bit(In_sync, &rdev->flags) &&
1709                             is_badblock(rdev, sect, s,
1710                                         &first_bad, &bad_sectors) == 0 &&
1711                             sync_page_io(rdev, sect, s<<9,
1712                                          conf->tmppage, READ, false))
1713                                 success = 1;
1714                         else {
1715                                 d++;
1716                                 if (d == conf->raid_disks)
1717                                         d = 0;
1718                         }
1719                 } while (!success && d != read_disk);
1720
1721                 if (!success) {
1722                         /* Cannot read from anywhere - mark it bad */
1723                         mdk_rdev_t *rdev = conf->mirrors[read_disk].rdev;
1724                         if (!rdev_set_badblocks(rdev, sect, s, 0))
1725                                 md_error(mddev, rdev);
1726                         break;
1727                 }
1728                 /* write it back and re-read */
1729                 start = d;
1730                 while (d != read_disk) {
1731                         if (d==0)
1732                                 d = conf->raid_disks;
1733                         d--;
1734                         rdev = conf->mirrors[d].rdev;
1735                         if (rdev &&
1736                             test_bit(In_sync, &rdev->flags))
1737                                 r1_sync_page_io(rdev, sect, s,
1738                                                 conf->tmppage, WRITE);
1739                 }
1740                 d = start;
1741                 while (d != read_disk) {
1742                         char b[BDEVNAME_SIZE];
1743                         if (d==0)
1744                                 d = conf->raid_disks;
1745                         d--;
1746                         rdev = conf->mirrors[d].rdev;
1747                         if (rdev &&
1748                             test_bit(In_sync, &rdev->flags)) {
1749                                 if (r1_sync_page_io(rdev, sect, s,
1750                                                     conf->tmppage, READ)) {
1751                                         atomic_add(s, &rdev->corrected_errors);
1752                                         printk(KERN_INFO
1753                                                "md/raid1:%s: read error corrected "
1754                                                "(%d sectors at %llu on %s)\n",
1755                                                mdname(mddev), s,
1756                                                (unsigned long long)(sect +
1757                                                    rdev->data_offset),
1758                                                bdevname(rdev->bdev, b));
1759                                 }
1760                         }
1761                 }
1762                 sectors -= s;
1763                 sect += s;
1764         }
1765 }
1766
1767 static void bi_complete(struct bio *bio, int error)
1768 {
1769         complete((struct completion *)bio->bi_private);
1770 }
1771
1772 static int submit_bio_wait(int rw, struct bio *bio)
1773 {
1774         struct completion event;
1775         rw |= REQ_SYNC;
1776
1777         init_completion(&event);
1778         bio->bi_private = &event;
1779         bio->bi_end_io = bi_complete;
1780         submit_bio(rw, bio);
1781         wait_for_completion(&event);
1782
1783         return test_bit(BIO_UPTODATE, &bio->bi_flags);
1784 }
1785
1786 static int narrow_write_error(r1bio_t *r1_bio, int i)
1787 {
1788         mddev_t *mddev = r1_bio->mddev;
1789         conf_t *conf = mddev->private;
1790         mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1791         int vcnt, idx;
1792         struct bio_vec *vec;
1793
1794         /* bio has the data to be written to device 'i' where
1795          * we just recently had a write error.
1796          * We repeatedly clone the bio and trim down to one block,
1797          * then try the write.  Where the write fails we record
1798          * a bad block.
1799          * It is conceivable that the bio doesn't exactly align with
1800          * blocks.  We must handle this somehow.
1801          *
1802          * We currently own a reference on the rdev.
1803          */
1804
1805         int block_sectors;
1806         sector_t sector;
1807         int sectors;
1808         int sect_to_write = r1_bio->sectors;
1809         int ok = 1;
1810
1811         if (rdev->badblocks.shift < 0)
1812                 return 0;
1813
1814         block_sectors = 1 << rdev->badblocks.shift;
1815         sector = r1_bio->sector;
1816         sectors = ((sector + block_sectors)
1817                    & ~(sector_t)(block_sectors - 1))
1818                 - sector;
1819
1820         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
1821                 vcnt = r1_bio->behind_page_count;
1822                 vec = r1_bio->behind_bvecs;
1823                 idx = 0;
1824                 while (vec[idx].bv_page == NULL)
1825                         idx++;
1826         } else {
1827                 vcnt = r1_bio->master_bio->bi_vcnt;
1828                 vec = r1_bio->master_bio->bi_io_vec;
1829                 idx = r1_bio->master_bio->bi_idx;
1830         }
1831         while (sect_to_write) {
1832                 struct bio *wbio;
1833                 if (sectors > sect_to_write)
1834                         sectors = sect_to_write;
1835                 /* Write at 'sector' for 'sectors'*/
1836
1837                 wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
1838                 memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
1839                 wbio->bi_sector = r1_bio->sector;
1840                 wbio->bi_rw = WRITE;
1841                 wbio->bi_vcnt = vcnt;
1842                 wbio->bi_size = r1_bio->sectors << 9;
1843                 wbio->bi_idx = idx;
1844
1845                 md_trim_bio(wbio, sector - r1_bio->sector, sectors);
1846                 wbio->bi_sector += rdev->data_offset;
1847                 wbio->bi_bdev = rdev->bdev;
1848                 if (submit_bio_wait(WRITE, wbio) == 0)
1849                         /* failure! */
1850                         ok = rdev_set_badblocks(rdev, sector,
1851                                                 sectors, 0)
1852                                 && ok;
1853
1854                 bio_put(wbio);
1855                 sect_to_write -= sectors;
1856                 sector += sectors;
1857                 sectors = block_sectors;
1858         }
1859         return ok;
1860 }
1861
1862 static void handle_sync_write_finished(conf_t *conf, r1bio_t *r1_bio)
1863 {
1864         int m;
1865         int s = r1_bio->sectors;
1866         for (m = 0; m < conf->raid_disks ; m++) {
1867                 mdk_rdev_t *rdev = conf->mirrors[m].rdev;
1868                 struct bio *bio = r1_bio->bios[m];
1869                 if (bio->bi_end_io == NULL)
1870                         continue;
1871                 if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
1872                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
1873                         rdev_clear_badblocks(rdev, r1_bio->sector, s);
1874                 }
1875                 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
1876                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
1877                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
1878                                 md_error(conf->mddev, rdev);
1879                 }
1880         }
1881         put_buf(r1_bio);
1882         md_done_sync(conf->mddev, s, 1);
1883 }
1884
1885 static void handle_write_finished(conf_t *conf, r1bio_t *r1_bio)
1886 {
1887         int m;
1888         for (m = 0; m < conf->raid_disks ; m++)
1889                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
1890                         mdk_rdev_t *rdev = conf->mirrors[m].rdev;
1891                         rdev_clear_badblocks(rdev,
1892                                              r1_bio->sector,
1893                                              r1_bio->sectors);
1894                         rdev_dec_pending(rdev, conf->mddev);
1895                 } else if (r1_bio->bios[m] != NULL) {
1896                         /* This drive got a write error.  We need to
1897                          * narrow down and record precise write
1898                          * errors.
1899                          */
1900                         if (!narrow_write_error(r1_bio, m)) {
1901                                 md_error(conf->mddev,
1902                                          conf->mirrors[m].rdev);
1903                                 /* an I/O failed, we can't clear the bitmap */
1904                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1905                         }
1906                         rdev_dec_pending(conf->mirrors[m].rdev,
1907                                          conf->mddev);
1908                 }
1909         if (test_bit(R1BIO_WriteError, &r1_bio->state))
1910                 close_write(r1_bio);
1911         raid_end_bio_io(r1_bio);
1912 }
1913
1914 static void handle_read_error(conf_t *conf, r1bio_t *r1_bio)
1915 {
1916         int disk;
1917         int max_sectors;
1918         mddev_t *mddev = conf->mddev;
1919         struct bio *bio;
1920         char b[BDEVNAME_SIZE];
1921         mdk_rdev_t *rdev;
1922
1923         clear_bit(R1BIO_ReadError, &r1_bio->state);
1924         /* we got a read error. Maybe the drive is bad.  Maybe just
1925          * the block and we can fix it.
1926          * We freeze all other IO, and try reading the block from
1927          * other devices.  When we find one, we re-write
1928          * and check it that fixes the read error.
1929          * This is all done synchronously while the array is
1930          * frozen
1931          */
1932         if (mddev->ro == 0) {
1933                 freeze_array(conf);
1934                 fix_read_error(conf, r1_bio->read_disk,
1935                                r1_bio->sector, r1_bio->sectors);
1936                 unfreeze_array(conf);
1937         } else
1938                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1939
1940         bio = r1_bio->bios[r1_bio->read_disk];
1941         bdevname(bio->bi_bdev, b);
1942 read_more:
1943         disk = read_balance(conf, r1_bio, &max_sectors);
1944         if (disk == -1) {
1945                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
1946                        " read error for block %llu\n",
1947                        mdname(mddev), b, (unsigned long long)r1_bio->sector);
1948                 raid_end_bio_io(r1_bio);
1949         } else {
1950                 const unsigned long do_sync
1951                         = r1_bio->master_bio->bi_rw & REQ_SYNC;
1952                 if (bio) {
1953                         r1_bio->bios[r1_bio->read_disk] =
1954                                 mddev->ro ? IO_BLOCKED : NULL;
1955                         bio_put(bio);
1956                 }
1957                 r1_bio->read_disk = disk;
1958                 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
1959                 md_trim_bio(bio, r1_bio->sector - bio->bi_sector, max_sectors);
1960                 r1_bio->bios[r1_bio->read_disk] = bio;
1961                 rdev = conf->mirrors[disk].rdev;
1962                 printk_ratelimited(KERN_ERR
1963                                    "md/raid1:%s: redirecting sector %llu"
1964                                    " to other mirror: %s\n",
1965                                    mdname(mddev),
1966                                    (unsigned long long)r1_bio->sector,
1967                                    bdevname(rdev->bdev, b));
1968                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1969                 bio->bi_bdev = rdev->bdev;
1970                 bio->bi_end_io = raid1_end_read_request;
1971                 bio->bi_rw = READ | do_sync;
1972                 bio->bi_private = r1_bio;
1973                 if (max_sectors < r1_bio->sectors) {
1974                         /* Drat - have to split this up more */
1975                         struct bio *mbio = r1_bio->master_bio;
1976                         int sectors_handled = (r1_bio->sector + max_sectors
1977                                                - mbio->bi_sector);
1978                         r1_bio->sectors = max_sectors;
1979                         spin_lock_irq(&conf->device_lock);
1980                         if (mbio->bi_phys_segments == 0)
1981                                 mbio->bi_phys_segments = 2;
1982                         else
1983                                 mbio->bi_phys_segments++;
1984                         spin_unlock_irq(&conf->device_lock);
1985                         generic_make_request(bio);
1986                         bio = NULL;
1987
1988                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1989
1990                         r1_bio->master_bio = mbio;
1991                         r1_bio->sectors = (mbio->bi_size >> 9)
1992                                           - sectors_handled;
1993                         r1_bio->state = 0;
1994                         set_bit(R1BIO_ReadError, &r1_bio->state);
1995                         r1_bio->mddev = mddev;
1996                         r1_bio->sector = mbio->bi_sector + sectors_handled;
1997
1998                         goto read_more;
1999                 } else
2000                         generic_make_request(bio);
2001         }
2002 }
2003
2004 static void raid1d(mddev_t *mddev)
2005 {
2006         r1bio_t *r1_bio;
2007         unsigned long flags;
2008         conf_t *conf = mddev->private;
2009         struct list_head *head = &conf->retry_list;
2010         struct blk_plug plug;
2011
2012         md_check_recovery(mddev);
2013
2014         blk_start_plug(&plug);
2015         for (;;) {
2016
2017                 if (atomic_read(&mddev->plug_cnt) == 0)
2018                         flush_pending_writes(conf);
2019
2020                 spin_lock_irqsave(&conf->device_lock, flags);
2021                 if (list_empty(head)) {
2022                         spin_unlock_irqrestore(&conf->device_lock, flags);
2023                         break;
2024                 }
2025                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
2026                 list_del(head->prev);
2027                 conf->nr_queued--;
2028                 spin_unlock_irqrestore(&conf->device_lock, flags);
2029
2030                 mddev = r1_bio->mddev;
2031                 conf = mddev->private;
2032                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2033                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2034                             test_bit(R1BIO_WriteError, &r1_bio->state))
2035                                 handle_sync_write_finished(conf, r1_bio);
2036                         else
2037                                 sync_request_write(mddev, r1_bio);
2038                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2039                            test_bit(R1BIO_WriteError, &r1_bio->state))
2040                         handle_write_finished(conf, r1_bio);
2041                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2042                         handle_read_error(conf, r1_bio);
2043                 else
2044                         /* just a partial read to be scheduled from separate
2045                          * context
2046                          */
2047                         generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2048
2049                 cond_resched();
2050                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2051                         md_check_recovery(mddev);
2052         }
2053         blk_finish_plug(&plug);
2054 }
2055
2056
2057 static int init_resync(conf_t *conf)
2058 {
2059         int buffs;
2060
2061         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2062         BUG_ON(conf->r1buf_pool);
2063         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2064                                           conf->poolinfo);
2065         if (!conf->r1buf_pool)
2066                 return -ENOMEM;
2067         conf->next_resync = 0;
2068         return 0;
2069 }
2070
2071 /*
2072  * perform a "sync" on one "block"
2073  *
2074  * We need to make sure that no normal I/O request - particularly write
2075  * requests - conflict with active sync requests.
2076  *
2077  * This is achieved by tracking pending requests and a 'barrier' concept
2078  * that can be installed to exclude normal IO requests.
2079  */
2080
2081 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
2082 {
2083         conf_t *conf = mddev->private;
2084         r1bio_t *r1_bio;
2085         struct bio *bio;
2086         sector_t max_sector, nr_sectors;
2087         int disk = -1;
2088         int i;
2089         int wonly = -1;
2090         int write_targets = 0, read_targets = 0;
2091         sector_t sync_blocks;
2092         int still_degraded = 0;
2093         int good_sectors = RESYNC_SECTORS;
2094         int min_bad = 0; /* number of sectors that are bad in all devices */
2095
2096         if (!conf->r1buf_pool)
2097                 if (init_resync(conf))
2098                         return 0;
2099
2100         max_sector = mddev->dev_sectors;
2101         if (sector_nr >= max_sector) {
2102                 /* If we aborted, we need to abort the
2103                  * sync on the 'current' bitmap chunk (there will
2104                  * only be one in raid1 resync.
2105                  * We can find the current addess in mddev->curr_resync
2106                  */
2107                 if (mddev->curr_resync < max_sector) /* aborted */
2108                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2109                                                 &sync_blocks, 1);
2110                 else /* completed sync */
2111                         conf->fullsync = 0;
2112
2113                 bitmap_close_sync(mddev->bitmap);
2114                 close_sync(conf);
2115                 return 0;
2116         }
2117
2118         if (mddev->bitmap == NULL &&
2119             mddev->recovery_cp == MaxSector &&
2120             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2121             conf->fullsync == 0) {
2122                 *skipped = 1;
2123                 return max_sector - sector_nr;
2124         }
2125         /* before building a request, check if we can skip these blocks..
2126          * This call the bitmap_start_sync doesn't actually record anything
2127          */
2128         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2129             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2130                 /* We can skip this block, and probably several more */
2131                 *skipped = 1;
2132                 return sync_blocks;
2133         }
2134         /*
2135          * If there is non-resync activity waiting for a turn,
2136          * and resync is going fast enough,
2137          * then let it though before starting on this new sync request.
2138          */
2139         if (!go_faster && conf->nr_waiting)
2140                 msleep_interruptible(1000);
2141
2142         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2143         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2144         raise_barrier(conf);
2145
2146         conf->next_resync = sector_nr;
2147
2148         rcu_read_lock();
2149         /*
2150          * If we get a correctably read error during resync or recovery,
2151          * we might want to read from a different device.  So we
2152          * flag all drives that could conceivably be read from for READ,
2153          * and any others (which will be non-In_sync devices) for WRITE.
2154          * If a read fails, we try reading from something else for which READ
2155          * is OK.
2156          */
2157
2158         r1_bio->mddev = mddev;
2159         r1_bio->sector = sector_nr;
2160         r1_bio->state = 0;
2161         set_bit(R1BIO_IsSync, &r1_bio->state);
2162
2163         for (i=0; i < conf->raid_disks; i++) {
2164                 mdk_rdev_t *rdev;
2165                 bio = r1_bio->bios[i];
2166
2167                 /* take from bio_init */
2168                 bio->bi_next = NULL;
2169                 bio->bi_flags &= ~(BIO_POOL_MASK-1);
2170                 bio->bi_flags |= 1 << BIO_UPTODATE;
2171                 bio->bi_comp_cpu = -1;
2172                 bio->bi_rw = READ;
2173                 bio->bi_vcnt = 0;
2174                 bio->bi_idx = 0;
2175                 bio->bi_phys_segments = 0;
2176                 bio->bi_size = 0;
2177                 bio->bi_end_io = NULL;
2178                 bio->bi_private = NULL;
2179
2180                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2181                 if (rdev == NULL ||
2182                     test_bit(Faulty, &rdev->flags)) {
2183                         still_degraded = 1;
2184                 } else if (!test_bit(In_sync, &rdev->flags)) {
2185                         bio->bi_rw = WRITE;
2186                         bio->bi_end_io = end_sync_write;
2187                         write_targets ++;
2188                 } else {
2189                         /* may need to read from here */
2190                         sector_t first_bad = MaxSector;
2191                         int bad_sectors;
2192
2193                         if (is_badblock(rdev, sector_nr, good_sectors,
2194                                         &first_bad, &bad_sectors)) {
2195                                 if (first_bad > sector_nr)
2196                                         good_sectors = first_bad - sector_nr;
2197                                 else {
2198                                         bad_sectors -= (sector_nr - first_bad);
2199                                         if (min_bad == 0 ||
2200                                             min_bad > bad_sectors)
2201                                                 min_bad = bad_sectors;
2202                                 }
2203                         }
2204                         if (sector_nr < first_bad) {
2205                                 if (test_bit(WriteMostly, &rdev->flags)) {
2206                                         if (wonly < 0)
2207                                                 wonly = i;
2208                                 } else {
2209                                         if (disk < 0)
2210                                                 disk = i;
2211                                 }
2212                                 bio->bi_rw = READ;
2213                                 bio->bi_end_io = end_sync_read;
2214                                 read_targets++;
2215                         }
2216                 }
2217                 if (bio->bi_end_io) {
2218                         atomic_inc(&rdev->nr_pending);
2219                         bio->bi_sector = sector_nr + rdev->data_offset;
2220                         bio->bi_bdev = rdev->bdev;
2221                         bio->bi_private = r1_bio;
2222                 }
2223         }
2224         rcu_read_unlock();
2225         if (disk < 0)
2226                 disk = wonly;
2227         r1_bio->read_disk = disk;
2228
2229         if (read_targets == 0 && min_bad > 0) {
2230                 /* These sectors are bad on all InSync devices, so we
2231                  * need to mark them bad on all write targets
2232                  */
2233                 int ok = 1;
2234                 for (i = 0 ; i < conf->raid_disks ; i++)
2235                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2236                                 mdk_rdev_t *rdev =
2237                                         rcu_dereference(conf->mirrors[i].rdev);
2238                                 ok = rdev_set_badblocks(rdev, sector_nr,
2239                                                         min_bad, 0
2240                                         ) && ok;
2241                         }
2242                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2243                 *skipped = 1;
2244                 put_buf(r1_bio);
2245
2246                 if (!ok) {
2247                         /* Cannot record the badblocks, so need to
2248                          * abort the resync.
2249                          * If there are multiple read targets, could just
2250                          * fail the really bad ones ???
2251                          */
2252                         conf->recovery_disabled = mddev->recovery_disabled;
2253                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2254                         return 0;
2255                 } else
2256                         return min_bad;
2257
2258         }
2259         if (min_bad > 0 && min_bad < good_sectors) {
2260                 /* only resync enough to reach the next bad->good
2261                  * transition */
2262                 good_sectors = min_bad;
2263         }
2264
2265         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2266                 /* extra read targets are also write targets */
2267                 write_targets += read_targets-1;
2268
2269         if (write_targets == 0 || read_targets == 0) {
2270                 /* There is nowhere to write, so all non-sync
2271                  * drives must be failed - so we are finished
2272                  */
2273                 sector_t rv = max_sector - sector_nr;
2274                 *skipped = 1;
2275                 put_buf(r1_bio);
2276                 return rv;
2277         }
2278
2279         if (max_sector > mddev->resync_max)
2280                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2281         if (max_sector > sector_nr + good_sectors)
2282                 max_sector = sector_nr + good_sectors;
2283         nr_sectors = 0;
2284         sync_blocks = 0;
2285         do {
2286                 struct page *page;
2287                 int len = PAGE_SIZE;
2288                 if (sector_nr + (len>>9) > max_sector)
2289                         len = (max_sector - sector_nr) << 9;
2290                 if (len == 0)
2291                         break;
2292                 if (sync_blocks == 0) {
2293                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2294                                                &sync_blocks, still_degraded) &&
2295                             !conf->fullsync &&
2296                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2297                                 break;
2298                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2299                         if ((len >> 9) > sync_blocks)
2300                                 len = sync_blocks<<9;
2301                 }
2302
2303                 for (i=0 ; i < conf->raid_disks; i++) {
2304                         bio = r1_bio->bios[i];
2305                         if (bio->bi_end_io) {
2306                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2307                                 if (bio_add_page(bio, page, len, 0) == 0) {
2308                                         /* stop here */
2309                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2310                                         while (i > 0) {
2311                                                 i--;
2312                                                 bio = r1_bio->bios[i];
2313                                                 if (bio->bi_end_io==NULL)
2314                                                         continue;
2315                                                 /* remove last page from this bio */
2316                                                 bio->bi_vcnt--;
2317                                                 bio->bi_size -= len;
2318                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2319                                         }
2320                                         goto bio_full;
2321                                 }
2322                         }
2323                 }
2324                 nr_sectors += len>>9;
2325                 sector_nr += len>>9;
2326                 sync_blocks -= (len>>9);
2327         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2328  bio_full:
2329         r1_bio->sectors = nr_sectors;
2330
2331         /* For a user-requested sync, we read all readable devices and do a
2332          * compare
2333          */
2334         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2335                 atomic_set(&r1_bio->remaining, read_targets);
2336                 for (i=0; i<conf->raid_disks; i++) {
2337                         bio = r1_bio->bios[i];
2338                         if (bio->bi_end_io == end_sync_read) {
2339                                 md_sync_acct(bio->bi_bdev, nr_sectors);
2340                                 generic_make_request(bio);
2341                         }
2342                 }
2343         } else {
2344                 atomic_set(&r1_bio->remaining, 1);
2345                 bio = r1_bio->bios[r1_bio->read_disk];
2346                 md_sync_acct(bio->bi_bdev, nr_sectors);
2347                 generic_make_request(bio);
2348
2349         }
2350         return nr_sectors;
2351 }
2352
2353 static sector_t raid1_size(mddev_t *mddev, sector_t sectors, int raid_disks)
2354 {
2355         if (sectors)
2356                 return sectors;
2357
2358         return mddev->dev_sectors;
2359 }
2360
2361 static conf_t *setup_conf(mddev_t *mddev)
2362 {
2363         conf_t *conf;
2364         int i;
2365         mirror_info_t *disk;
2366         mdk_rdev_t *rdev;
2367         int err = -ENOMEM;
2368
2369         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2370         if (!conf)
2371                 goto abort;
2372
2373         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2374                                  GFP_KERNEL);
2375         if (!conf->mirrors)
2376                 goto abort;
2377
2378         conf->tmppage = alloc_page(GFP_KERNEL);
2379         if (!conf->tmppage)
2380                 goto abort;
2381
2382         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2383         if (!conf->poolinfo)
2384                 goto abort;
2385         conf->poolinfo->raid_disks = mddev->raid_disks;
2386         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2387                                           r1bio_pool_free,
2388                                           conf->poolinfo);
2389         if (!conf->r1bio_pool)
2390                 goto abort;
2391
2392         conf->poolinfo->mddev = mddev;
2393
2394         spin_lock_init(&conf->device_lock);
2395         list_for_each_entry(rdev, &mddev->disks, same_set) {
2396                 int disk_idx = rdev->raid_disk;
2397                 if (disk_idx >= mddev->raid_disks
2398                     || disk_idx < 0)
2399                         continue;
2400                 disk = conf->mirrors + disk_idx;
2401
2402                 disk->rdev = rdev;
2403
2404                 disk->head_position = 0;
2405         }
2406         conf->raid_disks = mddev->raid_disks;
2407         conf->mddev = mddev;
2408         INIT_LIST_HEAD(&conf->retry_list);
2409
2410         spin_lock_init(&conf->resync_lock);
2411         init_waitqueue_head(&conf->wait_barrier);
2412
2413         bio_list_init(&conf->pending_bio_list);
2414
2415         conf->last_used = -1;
2416         for (i = 0; i < conf->raid_disks; i++) {
2417
2418                 disk = conf->mirrors + i;
2419
2420                 if (!disk->rdev ||
2421                     !test_bit(In_sync, &disk->rdev->flags)) {
2422                         disk->head_position = 0;
2423                         if (disk->rdev)
2424                                 conf->fullsync = 1;
2425                 } else if (conf->last_used < 0)
2426                         /*
2427                          * The first working device is used as a
2428                          * starting point to read balancing.
2429                          */
2430                         conf->last_used = i;
2431         }
2432
2433         err = -EIO;
2434         if (conf->last_used < 0) {
2435                 printk(KERN_ERR "md/raid1:%s: no operational mirrors\n",
2436                        mdname(mddev));
2437                 goto abort;
2438         }
2439         err = -ENOMEM;
2440         conf->thread = md_register_thread(raid1d, mddev, NULL);
2441         if (!conf->thread) {
2442                 printk(KERN_ERR
2443                        "md/raid1:%s: couldn't allocate thread\n",
2444                        mdname(mddev));
2445                 goto abort;
2446         }
2447
2448         return conf;
2449
2450  abort:
2451         if (conf) {
2452                 if (conf->r1bio_pool)
2453                         mempool_destroy(conf->r1bio_pool);
2454                 kfree(conf->mirrors);
2455                 safe_put_page(conf->tmppage);
2456                 kfree(conf->poolinfo);
2457                 kfree(conf);
2458         }
2459         return ERR_PTR(err);
2460 }
2461
2462 static int run(mddev_t *mddev)
2463 {
2464         conf_t *conf;
2465         int i;
2466         mdk_rdev_t *rdev;
2467
2468         if (mddev->level != 1) {
2469                 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2470                        mdname(mddev), mddev->level);
2471                 return -EIO;
2472         }
2473         if (mddev->reshape_position != MaxSector) {
2474                 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2475                        mdname(mddev));
2476                 return -EIO;
2477         }
2478         /*
2479          * copy the already verified devices into our private RAID1
2480          * bookkeeping area. [whatever we allocate in run(),
2481          * should be freed in stop()]
2482          */
2483         if (mddev->private == NULL)
2484                 conf = setup_conf(mddev);
2485         else
2486                 conf = mddev->private;
2487
2488         if (IS_ERR(conf))
2489                 return PTR_ERR(conf);
2490
2491         list_for_each_entry(rdev, &mddev->disks, same_set) {
2492                 if (!mddev->gendisk)
2493                         continue;
2494                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2495                                   rdev->data_offset << 9);
2496                 /* as we don't honour merge_bvec_fn, we must never risk
2497                  * violating it, so limit ->max_segments to 1 lying within
2498                  * a single page, as a one page request is never in violation.
2499                  */
2500                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2501                         blk_queue_max_segments(mddev->queue, 1);
2502                         blk_queue_segment_boundary(mddev->queue,
2503                                                    PAGE_CACHE_SIZE - 1);
2504                 }
2505         }
2506
2507         mddev->degraded = 0;
2508         for (i=0; i < conf->raid_disks; i++)
2509                 if (conf->mirrors[i].rdev == NULL ||
2510                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2511                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2512                         mddev->degraded++;
2513
2514         if (conf->raid_disks - mddev->degraded == 1)
2515                 mddev->recovery_cp = MaxSector;
2516
2517         if (mddev->recovery_cp != MaxSector)
2518                 printk(KERN_NOTICE "md/raid1:%s: not clean"
2519                        " -- starting background reconstruction\n",
2520                        mdname(mddev));
2521         printk(KERN_INFO 
2522                 "md/raid1:%s: active with %d out of %d mirrors\n",
2523                 mdname(mddev), mddev->raid_disks - mddev->degraded, 
2524                 mddev->raid_disks);
2525
2526         /*
2527          * Ok, everything is just fine now
2528          */
2529         mddev->thread = conf->thread;
2530         conf->thread = NULL;
2531         mddev->private = conf;
2532
2533         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2534
2535         if (mddev->queue) {
2536                 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2537                 mddev->queue->backing_dev_info.congested_data = mddev;
2538         }
2539         return md_integrity_register(mddev);
2540 }
2541
2542 static int stop(mddev_t *mddev)
2543 {
2544         conf_t *conf = mddev->private;
2545         struct bitmap *bitmap = mddev->bitmap;
2546
2547         /* wait for behind writes to complete */
2548         if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2549                 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2550                        mdname(mddev));
2551                 /* need to kick something here to make sure I/O goes? */
2552                 wait_event(bitmap->behind_wait,
2553                            atomic_read(&bitmap->behind_writes) == 0);
2554         }
2555
2556         raise_barrier(conf);
2557         lower_barrier(conf);
2558
2559         md_unregister_thread(mddev->thread);
2560         mddev->thread = NULL;
2561         if (conf->r1bio_pool)
2562                 mempool_destroy(conf->r1bio_pool);
2563         kfree(conf->mirrors);
2564         kfree(conf->poolinfo);
2565         kfree(conf);
2566         mddev->private = NULL;
2567         return 0;
2568 }
2569
2570 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2571 {
2572         /* no resync is happening, and there is enough space
2573          * on all devices, so we can resize.
2574          * We need to make sure resync covers any new space.
2575          * If the array is shrinking we should possibly wait until
2576          * any io in the removed space completes, but it hardly seems
2577          * worth it.
2578          */
2579         md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2580         if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2581                 return -EINVAL;
2582         set_capacity(mddev->gendisk, mddev->array_sectors);
2583         revalidate_disk(mddev->gendisk);
2584         if (sectors > mddev->dev_sectors &&
2585             mddev->recovery_cp > mddev->dev_sectors) {
2586                 mddev->recovery_cp = mddev->dev_sectors;
2587                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2588         }
2589         mddev->dev_sectors = sectors;
2590         mddev->resync_max_sectors = sectors;
2591         return 0;
2592 }
2593
2594 static int raid1_reshape(mddev_t *mddev)
2595 {
2596         /* We need to:
2597          * 1/ resize the r1bio_pool
2598          * 2/ resize conf->mirrors
2599          *
2600          * We allocate a new r1bio_pool if we can.
2601          * Then raise a device barrier and wait until all IO stops.
2602          * Then resize conf->mirrors and swap in the new r1bio pool.
2603          *
2604          * At the same time, we "pack" the devices so that all the missing
2605          * devices have the higher raid_disk numbers.
2606          */
2607         mempool_t *newpool, *oldpool;
2608         struct pool_info *newpoolinfo;
2609         mirror_info_t *newmirrors;
2610         conf_t *conf = mddev->private;
2611         int cnt, raid_disks;
2612         unsigned long flags;
2613         int d, d2, err;
2614
2615         /* Cannot change chunk_size, layout, or level */
2616         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2617             mddev->layout != mddev->new_layout ||
2618             mddev->level != mddev->new_level) {
2619                 mddev->new_chunk_sectors = mddev->chunk_sectors;
2620                 mddev->new_layout = mddev->layout;
2621                 mddev->new_level = mddev->level;
2622                 return -EINVAL;
2623         }
2624
2625         err = md_allow_write(mddev);
2626         if (err)
2627                 return err;
2628
2629         raid_disks = mddev->raid_disks + mddev->delta_disks;
2630
2631         if (raid_disks < conf->raid_disks) {
2632                 cnt=0;
2633                 for (d= 0; d < conf->raid_disks; d++)
2634                         if (conf->mirrors[d].rdev)
2635                                 cnt++;
2636                 if (cnt > raid_disks)
2637                         return -EBUSY;
2638         }
2639
2640         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2641         if (!newpoolinfo)
2642                 return -ENOMEM;
2643         newpoolinfo->mddev = mddev;
2644         newpoolinfo->raid_disks = raid_disks;
2645
2646         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2647                                  r1bio_pool_free, newpoolinfo);
2648         if (!newpool) {
2649                 kfree(newpoolinfo);
2650                 return -ENOMEM;
2651         }
2652         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2653         if (!newmirrors) {
2654                 kfree(newpoolinfo);
2655                 mempool_destroy(newpool);
2656                 return -ENOMEM;
2657         }
2658
2659         raise_barrier(conf);
2660
2661         /* ok, everything is stopped */
2662         oldpool = conf->r1bio_pool;
2663         conf->r1bio_pool = newpool;
2664
2665         for (d = d2 = 0; d < conf->raid_disks; d++) {
2666                 mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2667                 if (rdev && rdev->raid_disk != d2) {
2668                         sysfs_unlink_rdev(mddev, rdev);
2669                         rdev->raid_disk = d2;
2670                         sysfs_unlink_rdev(mddev, rdev);
2671                         if (sysfs_link_rdev(mddev, rdev))
2672                                 printk(KERN_WARNING
2673                                        "md/raid1:%s: cannot register rd%d\n",
2674                                        mdname(mddev), rdev->raid_disk);
2675                 }
2676                 if (rdev)
2677                         newmirrors[d2++].rdev = rdev;
2678         }
2679         kfree(conf->mirrors);
2680         conf->mirrors = newmirrors;
2681         kfree(conf->poolinfo);
2682         conf->poolinfo = newpoolinfo;
2683
2684         spin_lock_irqsave(&conf->device_lock, flags);
2685         mddev->degraded += (raid_disks - conf->raid_disks);
2686         spin_unlock_irqrestore(&conf->device_lock, flags);
2687         conf->raid_disks = mddev->raid_disks = raid_disks;
2688         mddev->delta_disks = 0;
2689
2690         conf->last_used = 0; /* just make sure it is in-range */
2691         lower_barrier(conf);
2692
2693         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2694         md_wakeup_thread(mddev->thread);
2695
2696         mempool_destroy(oldpool);
2697         return 0;
2698 }
2699
2700 static void raid1_quiesce(mddev_t *mddev, int state)
2701 {
2702         conf_t *conf = mddev->private;
2703
2704         switch(state) {
2705         case 2: /* wake for suspend */
2706                 wake_up(&conf->wait_barrier);
2707                 break;
2708         case 1:
2709                 raise_barrier(conf);
2710                 break;
2711         case 0:
2712                 lower_barrier(conf);
2713                 break;
2714         }
2715 }
2716
2717 static void *raid1_takeover(mddev_t *mddev)
2718 {
2719         /* raid1 can take over:
2720          *  raid5 with 2 devices, any layout or chunk size
2721          */
2722         if (mddev->level == 5 && mddev->raid_disks == 2) {
2723                 conf_t *conf;
2724                 mddev->new_level = 1;
2725                 mddev->new_layout = 0;
2726                 mddev->new_chunk_sectors = 0;
2727                 conf = setup_conf(mddev);
2728                 if (!IS_ERR(conf))
2729                         conf->barrier = 1;
2730                 return conf;
2731         }
2732         return ERR_PTR(-EINVAL);
2733 }
2734
2735 static struct mdk_personality raid1_personality =
2736 {
2737         .name           = "raid1",
2738         .level          = 1,
2739         .owner          = THIS_MODULE,
2740         .make_request   = make_request,
2741         .run            = run,
2742         .stop           = stop,
2743         .status         = status,
2744         .error_handler  = error,
2745         .hot_add_disk   = raid1_add_disk,
2746         .hot_remove_disk= raid1_remove_disk,
2747         .spare_active   = raid1_spare_active,
2748         .sync_request   = sync_request,
2749         .resize         = raid1_resize,
2750         .size           = raid1_size,
2751         .check_reshape  = raid1_reshape,
2752         .quiesce        = raid1_quiesce,
2753         .takeover       = raid1_takeover,
2754 };
2755
2756 static int __init raid_init(void)
2757 {
2758         return register_md_personality(&raid1_personality);
2759 }
2760
2761 static void raid_exit(void)
2762 {
2763         unregister_md_personality(&raid1_personality);
2764 }
2765
2766 module_init(raid_init);
2767 module_exit(raid_exit);
2768 MODULE_LICENSE("GPL");
2769 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2770 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2771 MODULE_ALIAS("md-raid1");
2772 MODULE_ALIAS("md-level-1");