[PATCH] md: Final stages of raid5 expand code
[linux-2.6.git] / drivers / md / raid5.c
1 /*
2  * raid5.c : Multiple Devices driver for Linux
3  *         Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4  *         Copyright (C) 1999, 2000 Ingo Molnar
5  *
6  * RAID-5 management functions.
7  *
8  * This program is free software; you can redistribute it and/or modify
9  * it under the terms of the GNU General Public License as published by
10  * the Free Software Foundation; either version 2, or (at your option)
11  * any later version.
12  *
13  * You should have received a copy of the GNU General Public License
14  * (for example /usr/src/linux/COPYING); if not, write to the Free
15  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
16  */
17
18
19 #include <linux/config.h>
20 #include <linux/module.h>
21 #include <linux/slab.h>
22 #include <linux/raid/raid5.h>
23 #include <linux/highmem.h>
24 #include <linux/bitops.h>
25 #include <asm/atomic.h>
26
27 #include <linux/raid/bitmap.h>
28
29 /*
30  * Stripe cache
31  */
32
33 #define NR_STRIPES              256
34 #define STRIPE_SIZE             PAGE_SIZE
35 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
36 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
37 #define IO_THRESHOLD            1
38 #define NR_HASH                 (PAGE_SIZE / sizeof(struct hlist_head))
39 #define HASH_MASK               (NR_HASH - 1)
40
41 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
42
43 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
44  * order without overlap.  There may be several bio's per stripe+device, and
45  * a bio could span several devices.
46  * When walking this list for a particular stripe+device, we must never proceed
47  * beyond a bio that extends past this device, as the next bio might no longer
48  * be valid.
49  * This macro is used to determine the 'next' bio in the list, given the sector
50  * of the current stripe+device
51  */
52 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
53 /*
54  * The following can be used to debug the driver
55  */
56 #define RAID5_DEBUG     0
57 #define RAID5_PARANOIA  1
58 #if RAID5_PARANOIA && defined(CONFIG_SMP)
59 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
60 #else
61 # define CHECK_DEVLOCK()
62 #endif
63
64 #define PRINTK(x...) ((void)(RAID5_DEBUG && printk(x)))
65 #if RAID5_DEBUG
66 #define inline
67 #define __inline__
68 #endif
69
70 static void print_raid5_conf (raid5_conf_t *conf);
71
72 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
73 {
74         if (atomic_dec_and_test(&sh->count)) {
75                 if (!list_empty(&sh->lru))
76                         BUG();
77                 if (atomic_read(&conf->active_stripes)==0)
78                         BUG();
79                 if (test_bit(STRIPE_HANDLE, &sh->state)) {
80                         if (test_bit(STRIPE_DELAYED, &sh->state))
81                                 list_add_tail(&sh->lru, &conf->delayed_list);
82                         else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
83                                  conf->seq_write == sh->bm_seq)
84                                 list_add_tail(&sh->lru, &conf->bitmap_list);
85                         else {
86                                 clear_bit(STRIPE_BIT_DELAY, &sh->state);
87                                 list_add_tail(&sh->lru, &conf->handle_list);
88                         }
89                         md_wakeup_thread(conf->mddev->thread);
90                 } else {
91                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
92                                 atomic_dec(&conf->preread_active_stripes);
93                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
94                                         md_wakeup_thread(conf->mddev->thread);
95                         }
96                         atomic_dec(&conf->active_stripes);
97                         if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
98                                 list_add_tail(&sh->lru, &conf->inactive_list);
99                                 wake_up(&conf->wait_for_stripe);
100                         }
101                 }
102         }
103 }
104 static void release_stripe(struct stripe_head *sh)
105 {
106         raid5_conf_t *conf = sh->raid_conf;
107         unsigned long flags;
108         
109         spin_lock_irqsave(&conf->device_lock, flags);
110         __release_stripe(conf, sh);
111         spin_unlock_irqrestore(&conf->device_lock, flags);
112 }
113
114 static inline void remove_hash(struct stripe_head *sh)
115 {
116         PRINTK("remove_hash(), stripe %llu\n", (unsigned long long)sh->sector);
117
118         hlist_del_init(&sh->hash);
119 }
120
121 static void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
122 {
123         struct hlist_head *hp = stripe_hash(conf, sh->sector);
124
125         PRINTK("insert_hash(), stripe %llu\n", (unsigned long long)sh->sector);
126
127         CHECK_DEVLOCK();
128         hlist_add_head(&sh->hash, hp);
129 }
130
131
132 /* find an idle stripe, make sure it is unhashed, and return it. */
133 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
134 {
135         struct stripe_head *sh = NULL;
136         struct list_head *first;
137
138         CHECK_DEVLOCK();
139         if (list_empty(&conf->inactive_list))
140                 goto out;
141         first = conf->inactive_list.next;
142         sh = list_entry(first, struct stripe_head, lru);
143         list_del_init(first);
144         remove_hash(sh);
145         atomic_inc(&conf->active_stripes);
146 out:
147         return sh;
148 }
149
150 static void shrink_buffers(struct stripe_head *sh, int num)
151 {
152         struct page *p;
153         int i;
154
155         for (i=0; i<num ; i++) {
156                 p = sh->dev[i].page;
157                 if (!p)
158                         continue;
159                 sh->dev[i].page = NULL;
160                 put_page(p);
161         }
162 }
163
164 static int grow_buffers(struct stripe_head *sh, int num)
165 {
166         int i;
167
168         for (i=0; i<num; i++) {
169                 struct page *page;
170
171                 if (!(page = alloc_page(GFP_KERNEL))) {
172                         return 1;
173                 }
174                 sh->dev[i].page = page;
175         }
176         return 0;
177 }
178
179 static void raid5_build_block (struct stripe_head *sh, int i);
180
181 static void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx, int disks)
182 {
183         raid5_conf_t *conf = sh->raid_conf;
184         int i;
185
186         if (atomic_read(&sh->count) != 0)
187                 BUG();
188         if (test_bit(STRIPE_HANDLE, &sh->state))
189                 BUG();
190         
191         CHECK_DEVLOCK();
192         PRINTK("init_stripe called, stripe %llu\n", 
193                 (unsigned long long)sh->sector);
194
195         remove_hash(sh);
196         
197         sh->sector = sector;
198         sh->pd_idx = pd_idx;
199         sh->state = 0;
200
201         sh->disks = disks;
202
203         for (i = sh->disks; i--; ) {
204                 struct r5dev *dev = &sh->dev[i];
205
206                 if (dev->toread || dev->towrite || dev->written ||
207                     test_bit(R5_LOCKED, &dev->flags)) {
208                         printk("sector=%llx i=%d %p %p %p %d\n",
209                                (unsigned long long)sh->sector, i, dev->toread,
210                                dev->towrite, dev->written,
211                                test_bit(R5_LOCKED, &dev->flags));
212                         BUG();
213                 }
214                 dev->flags = 0;
215                 raid5_build_block(sh, i);
216         }
217         insert_hash(conf, sh);
218 }
219
220 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector, int disks)
221 {
222         struct stripe_head *sh;
223         struct hlist_node *hn;
224
225         CHECK_DEVLOCK();
226         PRINTK("__find_stripe, sector %llu\n", (unsigned long long)sector);
227         hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
228                 if (sh->sector == sector && sh->disks == disks)
229                         return sh;
230         PRINTK("__stripe %llu not in cache\n", (unsigned long long)sector);
231         return NULL;
232 }
233
234 static void unplug_slaves(mddev_t *mddev);
235 static void raid5_unplug_device(request_queue_t *q);
236
237 static struct stripe_head *get_active_stripe(raid5_conf_t *conf, sector_t sector, int disks,
238                                              int pd_idx, int noblock)
239 {
240         struct stripe_head *sh;
241
242         PRINTK("get_stripe, sector %llu\n", (unsigned long long)sector);
243
244         spin_lock_irq(&conf->device_lock);
245
246         do {
247                 wait_event_lock_irq(conf->wait_for_stripe,
248                                     conf->quiesce == 0,
249                                     conf->device_lock, /* nothing */);
250                 sh = __find_stripe(conf, sector, disks);
251                 if (!sh) {
252                         if (!conf->inactive_blocked)
253                                 sh = get_free_stripe(conf);
254                         if (noblock && sh == NULL)
255                                 break;
256                         if (!sh) {
257                                 conf->inactive_blocked = 1;
258                                 wait_event_lock_irq(conf->wait_for_stripe,
259                                                     !list_empty(&conf->inactive_list) &&
260                                                     (atomic_read(&conf->active_stripes)
261                                                      < (conf->max_nr_stripes *3/4)
262                                                      || !conf->inactive_blocked),
263                                                     conf->device_lock,
264                                                     unplug_slaves(conf->mddev);
265                                         );
266                                 conf->inactive_blocked = 0;
267                         } else
268                                 init_stripe(sh, sector, pd_idx, disks);
269                 } else {
270                         if (atomic_read(&sh->count)) {
271                                 if (!list_empty(&sh->lru))
272                                         BUG();
273                         } else {
274                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
275                                         atomic_inc(&conf->active_stripes);
276                                 if (!list_empty(&sh->lru))
277                                         list_del_init(&sh->lru);
278                         }
279                 }
280         } while (sh == NULL);
281
282         if (sh)
283                 atomic_inc(&sh->count);
284
285         spin_unlock_irq(&conf->device_lock);
286         return sh;
287 }
288
289 static int grow_one_stripe(raid5_conf_t *conf)
290 {
291         struct stripe_head *sh;
292         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
293         if (!sh)
294                 return 0;
295         memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
296         sh->raid_conf = conf;
297         spin_lock_init(&sh->lock);
298
299         if (grow_buffers(sh, conf->raid_disks)) {
300                 shrink_buffers(sh, conf->raid_disks);
301                 kmem_cache_free(conf->slab_cache, sh);
302                 return 0;
303         }
304         sh->disks = conf->raid_disks;
305         /* we just created an active stripe so... */
306         atomic_set(&sh->count, 1);
307         atomic_inc(&conf->active_stripes);
308         INIT_LIST_HEAD(&sh->lru);
309         release_stripe(sh);
310         return 1;
311 }
312
313 static int grow_stripes(raid5_conf_t *conf, int num)
314 {
315         kmem_cache_t *sc;
316         int devs = conf->raid_disks;
317
318         sprintf(conf->cache_name[0], "raid5/%s", mdname(conf->mddev));
319         sprintf(conf->cache_name[1], "raid5/%s-alt", mdname(conf->mddev));
320         conf->active_name = 0;
321         sc = kmem_cache_create(conf->cache_name[conf->active_name],
322                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
323                                0, 0, NULL, NULL);
324         if (!sc)
325                 return 1;
326         conf->slab_cache = sc;
327         conf->pool_size = devs;
328         while (num--) {
329                 if (!grow_one_stripe(conf))
330                         return 1;
331         }
332         return 0;
333 }
334
335 #ifdef CONFIG_MD_RAID5_RESHAPE
336 static int resize_stripes(raid5_conf_t *conf, int newsize)
337 {
338         /* Make all the stripes able to hold 'newsize' devices.
339          * New slots in each stripe get 'page' set to a new page.
340          *
341          * This happens in stages:
342          * 1/ create a new kmem_cache and allocate the required number of
343          *    stripe_heads.
344          * 2/ gather all the old stripe_heads and tranfer the pages across
345          *    to the new stripe_heads.  This will have the side effect of
346          *    freezing the array as once all stripe_heads have been collected,
347          *    no IO will be possible.  Old stripe heads are freed once their
348          *    pages have been transferred over, and the old kmem_cache is
349          *    freed when all stripes are done.
350          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
351          *    we simple return a failre status - no need to clean anything up.
352          * 4/ allocate new pages for the new slots in the new stripe_heads.
353          *    If this fails, we don't bother trying the shrink the
354          *    stripe_heads down again, we just leave them as they are.
355          *    As each stripe_head is processed the new one is released into
356          *    active service.
357          *
358          * Once step2 is started, we cannot afford to wait for a write,
359          * so we use GFP_NOIO allocations.
360          */
361         struct stripe_head *osh, *nsh;
362         LIST_HEAD(newstripes);
363         struct disk_info *ndisks;
364         int err = 0;
365         kmem_cache_t *sc;
366         int i;
367
368         if (newsize <= conf->pool_size)
369                 return 0; /* never bother to shrink */
370
371         /* Step 1 */
372         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
373                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
374                                0, 0, NULL, NULL);
375         if (!sc)
376                 return -ENOMEM;
377
378         for (i = conf->max_nr_stripes; i; i--) {
379                 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
380                 if (!nsh)
381                         break;
382
383                 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
384
385                 nsh->raid_conf = conf;
386                 spin_lock_init(&nsh->lock);
387
388                 list_add(&nsh->lru, &newstripes);
389         }
390         if (i) {
391                 /* didn't get enough, give up */
392                 while (!list_empty(&newstripes)) {
393                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
394                         list_del(&nsh->lru);
395                         kmem_cache_free(sc, nsh);
396                 }
397                 kmem_cache_destroy(sc);
398                 return -ENOMEM;
399         }
400         /* Step 2 - Must use GFP_NOIO now.
401          * OK, we have enough stripes, start collecting inactive
402          * stripes and copying them over
403          */
404         list_for_each_entry(nsh, &newstripes, lru) {
405                 spin_lock_irq(&conf->device_lock);
406                 wait_event_lock_irq(conf->wait_for_stripe,
407                                     !list_empty(&conf->inactive_list),
408                                     conf->device_lock,
409                                     unplug_slaves(conf->mddev);
410                         );
411                 osh = get_free_stripe(conf);
412                 spin_unlock_irq(&conf->device_lock);
413                 atomic_set(&nsh->count, 1);
414                 for(i=0; i<conf->pool_size; i++)
415                         nsh->dev[i].page = osh->dev[i].page;
416                 for( ; i<newsize; i++)
417                         nsh->dev[i].page = NULL;
418                 kmem_cache_free(conf->slab_cache, osh);
419         }
420         kmem_cache_destroy(conf->slab_cache);
421
422         /* Step 3.
423          * At this point, we are holding all the stripes so the array
424          * is completely stalled, so now is a good time to resize
425          * conf->disks.
426          */
427         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
428         if (ndisks) {
429                 for (i=0; i<conf->raid_disks; i++)
430                         ndisks[i] = conf->disks[i];
431                 kfree(conf->disks);
432                 conf->disks = ndisks;
433         } else
434                 err = -ENOMEM;
435
436         /* Step 4, return new stripes to service */
437         while(!list_empty(&newstripes)) {
438                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
439                 list_del_init(&nsh->lru);
440                 for (i=conf->raid_disks; i < newsize; i++)
441                         if (nsh->dev[i].page == NULL) {
442                                 struct page *p = alloc_page(GFP_NOIO);
443                                 nsh->dev[i].page = p;
444                                 if (!p)
445                                         err = -ENOMEM;
446                         }
447                 release_stripe(nsh);
448         }
449         /* critical section pass, GFP_NOIO no longer needed */
450
451         conf->slab_cache = sc;
452         conf->active_name = 1-conf->active_name;
453         conf->pool_size = newsize;
454         return err;
455 }
456 #endif
457
458 static int drop_one_stripe(raid5_conf_t *conf)
459 {
460         struct stripe_head *sh;
461
462         spin_lock_irq(&conf->device_lock);
463         sh = get_free_stripe(conf);
464         spin_unlock_irq(&conf->device_lock);
465         if (!sh)
466                 return 0;
467         if (atomic_read(&sh->count))
468                 BUG();
469         shrink_buffers(sh, conf->pool_size);
470         kmem_cache_free(conf->slab_cache, sh);
471         atomic_dec(&conf->active_stripes);
472         return 1;
473 }
474
475 static void shrink_stripes(raid5_conf_t *conf)
476 {
477         while (drop_one_stripe(conf))
478                 ;
479
480         if (conf->slab_cache)
481                 kmem_cache_destroy(conf->slab_cache);
482         conf->slab_cache = NULL;
483 }
484
485 static int raid5_end_read_request(struct bio * bi, unsigned int bytes_done,
486                                    int error)
487 {
488         struct stripe_head *sh = bi->bi_private;
489         raid5_conf_t *conf = sh->raid_conf;
490         int disks = sh->disks, i;
491         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
492
493         if (bi->bi_size)
494                 return 1;
495
496         for (i=0 ; i<disks; i++)
497                 if (bi == &sh->dev[i].req)
498                         break;
499
500         PRINTK("end_read_request %llu/%d, count: %d, uptodate %d.\n", 
501                 (unsigned long long)sh->sector, i, atomic_read(&sh->count), 
502                 uptodate);
503         if (i == disks) {
504                 BUG();
505                 return 0;
506         }
507
508         if (uptodate) {
509 #if 0
510                 struct bio *bio;
511                 unsigned long flags;
512                 spin_lock_irqsave(&conf->device_lock, flags);
513                 /* we can return a buffer if we bypassed the cache or
514                  * if the top buffer is not in highmem.  If there are
515                  * multiple buffers, leave the extra work to
516                  * handle_stripe
517                  */
518                 buffer = sh->bh_read[i];
519                 if (buffer &&
520                     (!PageHighMem(buffer->b_page)
521                      || buffer->b_page == bh->b_page )
522                         ) {
523                         sh->bh_read[i] = buffer->b_reqnext;
524                         buffer->b_reqnext = NULL;
525                 } else
526                         buffer = NULL;
527                 spin_unlock_irqrestore(&conf->device_lock, flags);
528                 if (sh->bh_page[i]==bh->b_page)
529                         set_buffer_uptodate(bh);
530                 if (buffer) {
531                         if (buffer->b_page != bh->b_page)
532                                 memcpy(buffer->b_data, bh->b_data, bh->b_size);
533                         buffer->b_end_io(buffer, 1);
534                 }
535 #else
536                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
537 #endif
538                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
539                         printk(KERN_INFO "raid5: read error corrected!!\n");
540                         clear_bit(R5_ReadError, &sh->dev[i].flags);
541                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
542                 }
543                 if (atomic_read(&conf->disks[i].rdev->read_errors))
544                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
545         } else {
546                 int retry = 0;
547                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
548                 atomic_inc(&conf->disks[i].rdev->read_errors);
549                 if (conf->mddev->degraded)
550                         printk(KERN_WARNING "raid5: read error not correctable.\n");
551                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
552                         /* Oh, no!!! */
553                         printk(KERN_WARNING "raid5: read error NOT corrected!!\n");
554                 else if (atomic_read(&conf->disks[i].rdev->read_errors)
555                          > conf->max_nr_stripes)
556                         printk(KERN_WARNING
557                                "raid5: Too many read errors, failing device.\n");
558                 else
559                         retry = 1;
560                 if (retry)
561                         set_bit(R5_ReadError, &sh->dev[i].flags);
562                 else {
563                         clear_bit(R5_ReadError, &sh->dev[i].flags);
564                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
565                         md_error(conf->mddev, conf->disks[i].rdev);
566                 }
567         }
568         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
569 #if 0
570         /* must restore b_page before unlocking buffer... */
571         if (sh->bh_page[i] != bh->b_page) {
572                 bh->b_page = sh->bh_page[i];
573                 bh->b_data = page_address(bh->b_page);
574                 clear_buffer_uptodate(bh);
575         }
576 #endif
577         clear_bit(R5_LOCKED, &sh->dev[i].flags);
578         set_bit(STRIPE_HANDLE, &sh->state);
579         release_stripe(sh);
580         return 0;
581 }
582
583 static int raid5_end_write_request (struct bio *bi, unsigned int bytes_done,
584                                     int error)
585 {
586         struct stripe_head *sh = bi->bi_private;
587         raid5_conf_t *conf = sh->raid_conf;
588         int disks = sh->disks, i;
589         unsigned long flags;
590         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
591
592         if (bi->bi_size)
593                 return 1;
594
595         for (i=0 ; i<disks; i++)
596                 if (bi == &sh->dev[i].req)
597                         break;
598
599         PRINTK("end_write_request %llu/%d, count %d, uptodate: %d.\n", 
600                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
601                 uptodate);
602         if (i == disks) {
603                 BUG();
604                 return 0;
605         }
606
607         spin_lock_irqsave(&conf->device_lock, flags);
608         if (!uptodate)
609                 md_error(conf->mddev, conf->disks[i].rdev);
610
611         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
612         
613         clear_bit(R5_LOCKED, &sh->dev[i].flags);
614         set_bit(STRIPE_HANDLE, &sh->state);
615         __release_stripe(conf, sh);
616         spin_unlock_irqrestore(&conf->device_lock, flags);
617         return 0;
618 }
619
620
621 static sector_t compute_blocknr(struct stripe_head *sh, int i);
622         
623 static void raid5_build_block (struct stripe_head *sh, int i)
624 {
625         struct r5dev *dev = &sh->dev[i];
626
627         bio_init(&dev->req);
628         dev->req.bi_io_vec = &dev->vec;
629         dev->req.bi_vcnt++;
630         dev->req.bi_max_vecs++;
631         dev->vec.bv_page = dev->page;
632         dev->vec.bv_len = STRIPE_SIZE;
633         dev->vec.bv_offset = 0;
634
635         dev->req.bi_sector = sh->sector;
636         dev->req.bi_private = sh;
637
638         dev->flags = 0;
639         if (i != sh->pd_idx)
640                 dev->sector = compute_blocknr(sh, i);
641 }
642
643 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
644 {
645         char b[BDEVNAME_SIZE];
646         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
647         PRINTK("raid5: error called\n");
648
649         if (!test_bit(Faulty, &rdev->flags)) {
650                 mddev->sb_dirty = 1;
651                 if (test_bit(In_sync, &rdev->flags)) {
652                         conf->working_disks--;
653                         mddev->degraded++;
654                         conf->failed_disks++;
655                         clear_bit(In_sync, &rdev->flags);
656                         /*
657                          * if recovery was running, make sure it aborts.
658                          */
659                         set_bit(MD_RECOVERY_ERR, &mddev->recovery);
660                 }
661                 set_bit(Faulty, &rdev->flags);
662                 printk (KERN_ALERT
663                         "raid5: Disk failure on %s, disabling device."
664                         " Operation continuing on %d devices\n",
665                         bdevname(rdev->bdev,b), conf->working_disks);
666         }
667 }       
668
669 /*
670  * Input: a 'big' sector number,
671  * Output: index of the data and parity disk, and the sector # in them.
672  */
673 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
674                         unsigned int data_disks, unsigned int * dd_idx,
675                         unsigned int * pd_idx, raid5_conf_t *conf)
676 {
677         long stripe;
678         unsigned long chunk_number;
679         unsigned int chunk_offset;
680         sector_t new_sector;
681         int sectors_per_chunk = conf->chunk_size >> 9;
682
683         /* First compute the information on this sector */
684
685         /*
686          * Compute the chunk number and the sector offset inside the chunk
687          */
688         chunk_offset = sector_div(r_sector, sectors_per_chunk);
689         chunk_number = r_sector;
690         BUG_ON(r_sector != chunk_number);
691
692         /*
693          * Compute the stripe number
694          */
695         stripe = chunk_number / data_disks;
696
697         /*
698          * Compute the data disk and parity disk indexes inside the stripe
699          */
700         *dd_idx = chunk_number % data_disks;
701
702         /*
703          * Select the parity disk based on the user selected algorithm.
704          */
705         if (conf->level == 4)
706                 *pd_idx = data_disks;
707         else switch (conf->algorithm) {
708                 case ALGORITHM_LEFT_ASYMMETRIC:
709                         *pd_idx = data_disks - stripe % raid_disks;
710                         if (*dd_idx >= *pd_idx)
711                                 (*dd_idx)++;
712                         break;
713                 case ALGORITHM_RIGHT_ASYMMETRIC:
714                         *pd_idx = stripe % raid_disks;
715                         if (*dd_idx >= *pd_idx)
716                                 (*dd_idx)++;
717                         break;
718                 case ALGORITHM_LEFT_SYMMETRIC:
719                         *pd_idx = data_disks - stripe % raid_disks;
720                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
721                         break;
722                 case ALGORITHM_RIGHT_SYMMETRIC:
723                         *pd_idx = stripe % raid_disks;
724                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
725                         break;
726                 default:
727                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
728                                 conf->algorithm);
729         }
730
731         /*
732          * Finally, compute the new sector number
733          */
734         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
735         return new_sector;
736 }
737
738
739 static sector_t compute_blocknr(struct stripe_head *sh, int i)
740 {
741         raid5_conf_t *conf = sh->raid_conf;
742         int raid_disks = sh->disks, data_disks = raid_disks - 1;
743         sector_t new_sector = sh->sector, check;
744         int sectors_per_chunk = conf->chunk_size >> 9;
745         sector_t stripe;
746         int chunk_offset;
747         int chunk_number, dummy1, dummy2, dd_idx = i;
748         sector_t r_sector;
749
750         chunk_offset = sector_div(new_sector, sectors_per_chunk);
751         stripe = new_sector;
752         BUG_ON(new_sector != stripe);
753
754         
755         switch (conf->algorithm) {
756                 case ALGORITHM_LEFT_ASYMMETRIC:
757                 case ALGORITHM_RIGHT_ASYMMETRIC:
758                         if (i > sh->pd_idx)
759                                 i--;
760                         break;
761                 case ALGORITHM_LEFT_SYMMETRIC:
762                 case ALGORITHM_RIGHT_SYMMETRIC:
763                         if (i < sh->pd_idx)
764                                 i += raid_disks;
765                         i -= (sh->pd_idx + 1);
766                         break;
767                 default:
768                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
769                                 conf->algorithm);
770         }
771
772         chunk_number = stripe * data_disks + i;
773         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
774
775         check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
776         if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
777                 printk(KERN_ERR "compute_blocknr: map not correct\n");
778                 return 0;
779         }
780         return r_sector;
781 }
782
783
784
785 /*
786  * Copy data between a page in the stripe cache, and a bio.
787  * There are no alignment or size guarantees between the page or the
788  * bio except that there is some overlap.
789  * All iovecs in the bio must be considered.
790  */
791 static void copy_data(int frombio, struct bio *bio,
792                      struct page *page,
793                      sector_t sector)
794 {
795         char *pa = page_address(page);
796         struct bio_vec *bvl;
797         int i;
798         int page_offset;
799
800         if (bio->bi_sector >= sector)
801                 page_offset = (signed)(bio->bi_sector - sector) * 512;
802         else
803                 page_offset = (signed)(sector - bio->bi_sector) * -512;
804         bio_for_each_segment(bvl, bio, i) {
805                 int len = bio_iovec_idx(bio,i)->bv_len;
806                 int clen;
807                 int b_offset = 0;
808
809                 if (page_offset < 0) {
810                         b_offset = -page_offset;
811                         page_offset += b_offset;
812                         len -= b_offset;
813                 }
814
815                 if (len > 0 && page_offset + len > STRIPE_SIZE)
816                         clen = STRIPE_SIZE - page_offset;
817                 else clen = len;
818                         
819                 if (clen > 0) {
820                         char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
821                         if (frombio)
822                                 memcpy(pa+page_offset, ba+b_offset, clen);
823                         else
824                                 memcpy(ba+b_offset, pa+page_offset, clen);
825                         __bio_kunmap_atomic(ba, KM_USER0);
826                 }
827                 if (clen < len) /* hit end of page */
828                         break;
829                 page_offset +=  len;
830         }
831 }
832
833 #define check_xor()     do {                                            \
834                            if (count == MAX_XOR_BLOCKS) {               \
835                                 xor_block(count, STRIPE_SIZE, ptr);     \
836                                 count = 1;                              \
837                            }                                            \
838                         } while(0)
839
840
841 static void compute_block(struct stripe_head *sh, int dd_idx)
842 {
843         int i, count, disks = sh->disks;
844         void *ptr[MAX_XOR_BLOCKS], *p;
845
846         PRINTK("compute_block, stripe %llu, idx %d\n", 
847                 (unsigned long long)sh->sector, dd_idx);
848
849         ptr[0] = page_address(sh->dev[dd_idx].page);
850         memset(ptr[0], 0, STRIPE_SIZE);
851         count = 1;
852         for (i = disks ; i--; ) {
853                 if (i == dd_idx)
854                         continue;
855                 p = page_address(sh->dev[i].page);
856                 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
857                         ptr[count++] = p;
858                 else
859                         printk(KERN_ERR "compute_block() %d, stripe %llu, %d"
860                                 " not present\n", dd_idx,
861                                 (unsigned long long)sh->sector, i);
862
863                 check_xor();
864         }
865         if (count != 1)
866                 xor_block(count, STRIPE_SIZE, ptr);
867         set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
868 }
869
870 static void compute_parity(struct stripe_head *sh, int method)
871 {
872         raid5_conf_t *conf = sh->raid_conf;
873         int i, pd_idx = sh->pd_idx, disks = sh->disks, count;
874         void *ptr[MAX_XOR_BLOCKS];
875         struct bio *chosen;
876
877         PRINTK("compute_parity, stripe %llu, method %d\n",
878                 (unsigned long long)sh->sector, method);
879
880         count = 1;
881         ptr[0] = page_address(sh->dev[pd_idx].page);
882         switch(method) {
883         case READ_MODIFY_WRITE:
884                 if (!test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags))
885                         BUG();
886                 for (i=disks ; i-- ;) {
887                         if (i==pd_idx)
888                                 continue;
889                         if (sh->dev[i].towrite &&
890                             test_bit(R5_UPTODATE, &sh->dev[i].flags)) {
891                                 ptr[count++] = page_address(sh->dev[i].page);
892                                 chosen = sh->dev[i].towrite;
893                                 sh->dev[i].towrite = NULL;
894
895                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
896                                         wake_up(&conf->wait_for_overlap);
897
898                                 if (sh->dev[i].written) BUG();
899                                 sh->dev[i].written = chosen;
900                                 check_xor();
901                         }
902                 }
903                 break;
904         case RECONSTRUCT_WRITE:
905                 memset(ptr[0], 0, STRIPE_SIZE);
906                 for (i= disks; i-- ;)
907                         if (i!=pd_idx && sh->dev[i].towrite) {
908                                 chosen = sh->dev[i].towrite;
909                                 sh->dev[i].towrite = NULL;
910
911                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
912                                         wake_up(&conf->wait_for_overlap);
913
914                                 if (sh->dev[i].written) BUG();
915                                 sh->dev[i].written = chosen;
916                         }
917                 break;
918         case CHECK_PARITY:
919                 break;
920         }
921         if (count>1) {
922                 xor_block(count, STRIPE_SIZE, ptr);
923                 count = 1;
924         }
925         
926         for (i = disks; i--;)
927                 if (sh->dev[i].written) {
928                         sector_t sector = sh->dev[i].sector;
929                         struct bio *wbi = sh->dev[i].written;
930                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
931                                 copy_data(1, wbi, sh->dev[i].page, sector);
932                                 wbi = r5_next_bio(wbi, sector);
933                         }
934
935                         set_bit(R5_LOCKED, &sh->dev[i].flags);
936                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
937                 }
938
939         switch(method) {
940         case RECONSTRUCT_WRITE:
941         case CHECK_PARITY:
942                 for (i=disks; i--;)
943                         if (i != pd_idx) {
944                                 ptr[count++] = page_address(sh->dev[i].page);
945                                 check_xor();
946                         }
947                 break;
948         case READ_MODIFY_WRITE:
949                 for (i = disks; i--;)
950                         if (sh->dev[i].written) {
951                                 ptr[count++] = page_address(sh->dev[i].page);
952                                 check_xor();
953                         }
954         }
955         if (count != 1)
956                 xor_block(count, STRIPE_SIZE, ptr);
957         
958         if (method != CHECK_PARITY) {
959                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
960                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
961         } else
962                 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
963 }
964
965 /*
966  * Each stripe/dev can have one or more bion attached.
967  * toread/towrite point to the first in a chain. 
968  * The bi_next chain must be in order.
969  */
970 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
971 {
972         struct bio **bip;
973         raid5_conf_t *conf = sh->raid_conf;
974         int firstwrite=0;
975
976         PRINTK("adding bh b#%llu to stripe s#%llu\n",
977                 (unsigned long long)bi->bi_sector,
978                 (unsigned long long)sh->sector);
979
980
981         spin_lock(&sh->lock);
982         spin_lock_irq(&conf->device_lock);
983         if (forwrite) {
984                 bip = &sh->dev[dd_idx].towrite;
985                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
986                         firstwrite = 1;
987         } else
988                 bip = &sh->dev[dd_idx].toread;
989         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
990                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
991                         goto overlap;
992                 bip = & (*bip)->bi_next;
993         }
994         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
995                 goto overlap;
996
997         if (*bip && bi->bi_next && (*bip) != bi->bi_next)
998                 BUG();
999         if (*bip)
1000                 bi->bi_next = *bip;
1001         *bip = bi;
1002         bi->bi_phys_segments ++;
1003         spin_unlock_irq(&conf->device_lock);
1004         spin_unlock(&sh->lock);
1005
1006         PRINTK("added bi b#%llu to stripe s#%llu, disk %d.\n",
1007                 (unsigned long long)bi->bi_sector,
1008                 (unsigned long long)sh->sector, dd_idx);
1009
1010         if (conf->mddev->bitmap && firstwrite) {
1011                 sh->bm_seq = conf->seq_write;
1012                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1013                                   STRIPE_SECTORS, 0);
1014                 set_bit(STRIPE_BIT_DELAY, &sh->state);
1015         }
1016
1017         if (forwrite) {
1018                 /* check if page is covered */
1019                 sector_t sector = sh->dev[dd_idx].sector;
1020                 for (bi=sh->dev[dd_idx].towrite;
1021                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1022                              bi && bi->bi_sector <= sector;
1023                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1024                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1025                                 sector = bi->bi_sector + (bi->bi_size>>9);
1026                 }
1027                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1028                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1029         }
1030         return 1;
1031
1032  overlap:
1033         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1034         spin_unlock_irq(&conf->device_lock);
1035         spin_unlock(&sh->lock);
1036         return 0;
1037 }
1038
1039 static void end_reshape(raid5_conf_t *conf);
1040
1041 static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks)
1042 {
1043         int sectors_per_chunk = conf->chunk_size >> 9;
1044         sector_t x = stripe;
1045         int pd_idx, dd_idx;
1046         int chunk_offset = sector_div(x, sectors_per_chunk);
1047         stripe = x;
1048         raid5_compute_sector(stripe*(disks-1)*sectors_per_chunk
1049                              + chunk_offset, disks, disks-1, &dd_idx, &pd_idx, conf);
1050         return pd_idx;
1051 }
1052
1053
1054 /*
1055  * handle_stripe - do things to a stripe.
1056  *
1057  * We lock the stripe and then examine the state of various bits
1058  * to see what needs to be done.
1059  * Possible results:
1060  *    return some read request which now have data
1061  *    return some write requests which are safely on disc
1062  *    schedule a read on some buffers
1063  *    schedule a write of some buffers
1064  *    return confirmation of parity correctness
1065  *
1066  * Parity calculations are done inside the stripe lock
1067  * buffers are taken off read_list or write_list, and bh_cache buffers
1068  * get BH_Lock set before the stripe lock is released.
1069  *
1070  */
1071  
1072 static void handle_stripe(struct stripe_head *sh)
1073 {
1074         raid5_conf_t *conf = sh->raid_conf;
1075         int disks = sh->disks;
1076         struct bio *return_bi= NULL;
1077         struct bio *bi;
1078         int i;
1079         int syncing, expanding, expanded;
1080         int locked=0, uptodate=0, to_read=0, to_write=0, failed=0, written=0;
1081         int non_overwrite = 0;
1082         int failed_num=0;
1083         struct r5dev *dev;
1084
1085         PRINTK("handling stripe %llu, cnt=%d, pd_idx=%d\n",
1086                 (unsigned long long)sh->sector, atomic_read(&sh->count),
1087                 sh->pd_idx);
1088
1089         spin_lock(&sh->lock);
1090         clear_bit(STRIPE_HANDLE, &sh->state);
1091         clear_bit(STRIPE_DELAYED, &sh->state);
1092
1093         syncing = test_bit(STRIPE_SYNCING, &sh->state);
1094         expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
1095         expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
1096         /* Now to look around and see what can be done */
1097
1098         rcu_read_lock();
1099         for (i=disks; i--; ) {
1100                 mdk_rdev_t *rdev;
1101                 dev = &sh->dev[i];
1102                 clear_bit(R5_Insync, &dev->flags);
1103
1104                 PRINTK("check %d: state 0x%lx read %p write %p written %p\n",
1105                         i, dev->flags, dev->toread, dev->towrite, dev->written);
1106                 /* maybe we can reply to a read */
1107                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
1108                         struct bio *rbi, *rbi2;
1109                         PRINTK("Return read for disc %d\n", i);
1110                         spin_lock_irq(&conf->device_lock);
1111                         rbi = dev->toread;
1112                         dev->toread = NULL;
1113                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
1114                                 wake_up(&conf->wait_for_overlap);
1115                         spin_unlock_irq(&conf->device_lock);
1116                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1117                                 copy_data(0, rbi, dev->page, dev->sector);
1118                                 rbi2 = r5_next_bio(rbi, dev->sector);
1119                                 spin_lock_irq(&conf->device_lock);
1120                                 if (--rbi->bi_phys_segments == 0) {
1121                                         rbi->bi_next = return_bi;
1122                                         return_bi = rbi;
1123                                 }
1124                                 spin_unlock_irq(&conf->device_lock);
1125                                 rbi = rbi2;
1126                         }
1127                 }
1128
1129                 /* now count some things */
1130                 if (test_bit(R5_LOCKED, &dev->flags)) locked++;
1131                 if (test_bit(R5_UPTODATE, &dev->flags)) uptodate++;
1132
1133                 
1134                 if (dev->toread) to_read++;
1135                 if (dev->towrite) {
1136                         to_write++;
1137                         if (!test_bit(R5_OVERWRITE, &dev->flags))
1138                                 non_overwrite++;
1139                 }
1140                 if (dev->written) written++;
1141                 rdev = rcu_dereference(conf->disks[i].rdev);
1142                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
1143                         /* The ReadError flag will just be confusing now */
1144                         clear_bit(R5_ReadError, &dev->flags);
1145                         clear_bit(R5_ReWrite, &dev->flags);
1146                 }
1147                 if (!rdev || !test_bit(In_sync, &rdev->flags)
1148                     || test_bit(R5_ReadError, &dev->flags)) {
1149                         failed++;
1150                         failed_num = i;
1151                 } else
1152                         set_bit(R5_Insync, &dev->flags);
1153         }
1154         rcu_read_unlock();
1155         PRINTK("locked=%d uptodate=%d to_read=%d"
1156                 " to_write=%d failed=%d failed_num=%d\n",
1157                 locked, uptodate, to_read, to_write, failed, failed_num);
1158         /* check if the array has lost two devices and, if so, some requests might
1159          * need to be failed
1160          */
1161         if (failed > 1 && to_read+to_write+written) {
1162                 for (i=disks; i--; ) {
1163                         int bitmap_end = 0;
1164
1165                         if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1166                                 mdk_rdev_t *rdev;
1167                                 rcu_read_lock();
1168                                 rdev = rcu_dereference(conf->disks[i].rdev);
1169                                 if (rdev && test_bit(In_sync, &rdev->flags))
1170                                         /* multiple read failures in one stripe */
1171                                         md_error(conf->mddev, rdev);
1172                                 rcu_read_unlock();
1173                         }
1174
1175                         spin_lock_irq(&conf->device_lock);
1176                         /* fail all writes first */
1177                         bi = sh->dev[i].towrite;
1178                         sh->dev[i].towrite = NULL;
1179                         if (bi) { to_write--; bitmap_end = 1; }
1180
1181                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1182                                 wake_up(&conf->wait_for_overlap);
1183
1184                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1185                                 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1186                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1187                                 if (--bi->bi_phys_segments == 0) {
1188                                         md_write_end(conf->mddev);
1189                                         bi->bi_next = return_bi;
1190                                         return_bi = bi;
1191                                 }
1192                                 bi = nextbi;
1193                         }
1194                         /* and fail all 'written' */
1195                         bi = sh->dev[i].written;
1196                         sh->dev[i].written = NULL;
1197                         if (bi) bitmap_end = 1;
1198                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS) {
1199                                 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1200                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1201                                 if (--bi->bi_phys_segments == 0) {
1202                                         md_write_end(conf->mddev);
1203                                         bi->bi_next = return_bi;
1204                                         return_bi = bi;
1205                                 }
1206                                 bi = bi2;
1207                         }
1208
1209                         /* fail any reads if this device is non-operational */
1210                         if (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1211                             test_bit(R5_ReadError, &sh->dev[i].flags)) {
1212                                 bi = sh->dev[i].toread;
1213                                 sh->dev[i].toread = NULL;
1214                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1215                                         wake_up(&conf->wait_for_overlap);
1216                                 if (bi) to_read--;
1217                                 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1218                                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1219                                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1220                                         if (--bi->bi_phys_segments == 0) {
1221                                                 bi->bi_next = return_bi;
1222                                                 return_bi = bi;
1223                                         }
1224                                         bi = nextbi;
1225                                 }
1226                         }
1227                         spin_unlock_irq(&conf->device_lock);
1228                         if (bitmap_end)
1229                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1230                                                 STRIPE_SECTORS, 0, 0);
1231                 }
1232         }
1233         if (failed > 1 && syncing) {
1234                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
1235                 clear_bit(STRIPE_SYNCING, &sh->state);
1236                 syncing = 0;
1237         }
1238
1239         /* might be able to return some write requests if the parity block
1240          * is safe, or on a failed drive
1241          */
1242         dev = &sh->dev[sh->pd_idx];
1243         if ( written &&
1244              ( (test_bit(R5_Insync, &dev->flags) && !test_bit(R5_LOCKED, &dev->flags) &&
1245                 test_bit(R5_UPTODATE, &dev->flags))
1246                || (failed == 1 && failed_num == sh->pd_idx))
1247             ) {
1248             /* any written block on an uptodate or failed drive can be returned.
1249              * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but 
1250              * never LOCKED, so we don't need to test 'failed' directly.
1251              */
1252             for (i=disks; i--; )
1253                 if (sh->dev[i].written) {
1254                     dev = &sh->dev[i];
1255                     if (!test_bit(R5_LOCKED, &dev->flags) &&
1256                          test_bit(R5_UPTODATE, &dev->flags) ) {
1257                         /* We can return any write requests */
1258                             struct bio *wbi, *wbi2;
1259                             int bitmap_end = 0;
1260                             PRINTK("Return write for disc %d\n", i);
1261                             spin_lock_irq(&conf->device_lock);
1262                             wbi = dev->written;
1263                             dev->written = NULL;
1264                             while (wbi && wbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1265                                     wbi2 = r5_next_bio(wbi, dev->sector);
1266                                     if (--wbi->bi_phys_segments == 0) {
1267                                             md_write_end(conf->mddev);
1268                                             wbi->bi_next = return_bi;
1269                                             return_bi = wbi;
1270                                     }
1271                                     wbi = wbi2;
1272                             }
1273                             if (dev->towrite == NULL)
1274                                     bitmap_end = 1;
1275                             spin_unlock_irq(&conf->device_lock);
1276                             if (bitmap_end)
1277                                     bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1278                                                     STRIPE_SECTORS,
1279                                                     !test_bit(STRIPE_DEGRADED, &sh->state), 0);
1280                     }
1281                 }
1282         }
1283
1284         /* Now we might consider reading some blocks, either to check/generate
1285          * parity, or to satisfy requests
1286          * or to load a block that is being partially written.
1287          */
1288         if (to_read || non_overwrite || (syncing && (uptodate < disks)) || expanding) {
1289                 for (i=disks; i--;) {
1290                         dev = &sh->dev[i];
1291                         if (!test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1292                             (dev->toread ||
1293                              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1294                              syncing ||
1295                              expanding ||
1296                              (failed && (sh->dev[failed_num].toread ||
1297                                          (sh->dev[failed_num].towrite && !test_bit(R5_OVERWRITE, &sh->dev[failed_num].flags))))
1298                                     )
1299                                 ) {
1300                                 /* we would like to get this block, possibly
1301                                  * by computing it, but we might not be able to
1302                                  */
1303                                 if (uptodate == disks-1) {
1304                                         PRINTK("Computing block %d\n", i);
1305                                         compute_block(sh, i);
1306                                         uptodate++;
1307                                 } else if (test_bit(R5_Insync, &dev->flags)) {
1308                                         set_bit(R5_LOCKED, &dev->flags);
1309                                         set_bit(R5_Wantread, &dev->flags);
1310 #if 0
1311                                         /* if I am just reading this block and we don't have
1312                                            a failed drive, or any pending writes then sidestep the cache */
1313                                         if (sh->bh_read[i] && !sh->bh_read[i]->b_reqnext &&
1314                                             ! syncing && !failed && !to_write) {
1315                                                 sh->bh_cache[i]->b_page =  sh->bh_read[i]->b_page;
1316                                                 sh->bh_cache[i]->b_data =  sh->bh_read[i]->b_data;
1317                                         }
1318 #endif
1319                                         locked++;
1320                                         PRINTK("Reading block %d (sync=%d)\n", 
1321                                                 i, syncing);
1322                                 }
1323                         }
1324                 }
1325                 set_bit(STRIPE_HANDLE, &sh->state);
1326         }
1327
1328         /* now to consider writing and what else, if anything should be read */
1329         if (to_write) {
1330                 int rmw=0, rcw=0;
1331                 for (i=disks ; i--;) {
1332                         /* would I have to read this buffer for read_modify_write */
1333                         dev = &sh->dev[i];
1334                         if ((dev->towrite || i == sh->pd_idx) &&
1335                             (!test_bit(R5_LOCKED, &dev->flags) 
1336 #if 0
1337 || sh->bh_page[i]!=bh->b_page
1338 #endif
1339                                     ) &&
1340                             !test_bit(R5_UPTODATE, &dev->flags)) {
1341                                 if (test_bit(R5_Insync, &dev->flags)
1342 /*                                  && !(!mddev->insync && i == sh->pd_idx) */
1343                                         )
1344                                         rmw++;
1345                                 else rmw += 2*disks;  /* cannot read it */
1346                         }
1347                         /* Would I have to read this buffer for reconstruct_write */
1348                         if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
1349                             (!test_bit(R5_LOCKED, &dev->flags) 
1350 #if 0
1351 || sh->bh_page[i] != bh->b_page
1352 #endif
1353                                     ) &&
1354                             !test_bit(R5_UPTODATE, &dev->flags)) {
1355                                 if (test_bit(R5_Insync, &dev->flags)) rcw++;
1356                                 else rcw += 2*disks;
1357                         }
1358                 }
1359                 PRINTK("for sector %llu, rmw=%d rcw=%d\n", 
1360                         (unsigned long long)sh->sector, rmw, rcw);
1361                 set_bit(STRIPE_HANDLE, &sh->state);
1362                 if (rmw < rcw && rmw > 0)
1363                         /* prefer read-modify-write, but need to get some data */
1364                         for (i=disks; i--;) {
1365                                 dev = &sh->dev[i];
1366                                 if ((dev->towrite || i == sh->pd_idx) &&
1367                                     !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1368                                     test_bit(R5_Insync, &dev->flags)) {
1369                                         if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1370                                         {
1371                                                 PRINTK("Read_old block %d for r-m-w\n", i);
1372                                                 set_bit(R5_LOCKED, &dev->flags);
1373                                                 set_bit(R5_Wantread, &dev->flags);
1374                                                 locked++;
1375                                         } else {
1376                                                 set_bit(STRIPE_DELAYED, &sh->state);
1377                                                 set_bit(STRIPE_HANDLE, &sh->state);
1378                                         }
1379                                 }
1380                         }
1381                 if (rcw <= rmw && rcw > 0)
1382                         /* want reconstruct write, but need to get some data */
1383                         for (i=disks; i--;) {
1384                                 dev = &sh->dev[i];
1385                                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
1386                                     !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1387                                     test_bit(R5_Insync, &dev->flags)) {
1388                                         if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1389                                         {
1390                                                 PRINTK("Read_old block %d for Reconstruct\n", i);
1391                                                 set_bit(R5_LOCKED, &dev->flags);
1392                                                 set_bit(R5_Wantread, &dev->flags);
1393                                                 locked++;
1394                                         } else {
1395                                                 set_bit(STRIPE_DELAYED, &sh->state);
1396                                                 set_bit(STRIPE_HANDLE, &sh->state);
1397                                         }
1398                                 }
1399                         }
1400                 /* now if nothing is locked, and if we have enough data, we can start a write request */
1401                 if (locked == 0 && (rcw == 0 ||rmw == 0) &&
1402                     !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
1403                         PRINTK("Computing parity...\n");
1404                         compute_parity(sh, rcw==0 ? RECONSTRUCT_WRITE : READ_MODIFY_WRITE);
1405                         /* now every locked buffer is ready to be written */
1406                         for (i=disks; i--;)
1407                                 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
1408                                         PRINTK("Writing block %d\n", i);
1409                                         locked++;
1410                                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
1411                                         if (!test_bit(R5_Insync, &sh->dev[i].flags)
1412                                             || (i==sh->pd_idx && failed == 0))
1413                                                 set_bit(STRIPE_INSYNC, &sh->state);
1414                                 }
1415                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
1416                                 atomic_dec(&conf->preread_active_stripes);
1417                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
1418                                         md_wakeup_thread(conf->mddev->thread);
1419                         }
1420                 }
1421         }
1422
1423         /* maybe we need to check and possibly fix the parity for this stripe
1424          * Any reads will already have been scheduled, so we just see if enough data
1425          * is available
1426          */
1427         if (syncing && locked == 0 &&
1428             !test_bit(STRIPE_INSYNC, &sh->state)) {
1429                 set_bit(STRIPE_HANDLE, &sh->state);
1430                 if (failed == 0) {
1431                         char *pagea;
1432                         if (uptodate != disks)
1433                                 BUG();
1434                         compute_parity(sh, CHECK_PARITY);
1435                         uptodate--;
1436                         pagea = page_address(sh->dev[sh->pd_idx].page);
1437                         if ((*(u32*)pagea) == 0 &&
1438                             !memcmp(pagea, pagea+4, STRIPE_SIZE-4)) {
1439                                 /* parity is correct (on disc, not in buffer any more) */
1440                                 set_bit(STRIPE_INSYNC, &sh->state);
1441                         } else {
1442                                 conf->mddev->resync_mismatches += STRIPE_SECTORS;
1443                                 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
1444                                         /* don't try to repair!! */
1445                                         set_bit(STRIPE_INSYNC, &sh->state);
1446                                 else {
1447                                         compute_block(sh, sh->pd_idx);
1448                                         uptodate++;
1449                                 }
1450                         }
1451                 }
1452                 if (!test_bit(STRIPE_INSYNC, &sh->state)) {
1453                         /* either failed parity check, or recovery is happening */
1454                         if (failed==0)
1455                                 failed_num = sh->pd_idx;
1456                         dev = &sh->dev[failed_num];
1457                         BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
1458                         BUG_ON(uptodate != disks);
1459
1460                         set_bit(R5_LOCKED, &dev->flags);
1461                         set_bit(R5_Wantwrite, &dev->flags);
1462                         clear_bit(STRIPE_DEGRADED, &sh->state);
1463                         locked++;
1464                         set_bit(STRIPE_INSYNC, &sh->state);
1465                 }
1466         }
1467         if (syncing && locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
1468                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
1469                 clear_bit(STRIPE_SYNCING, &sh->state);
1470         }
1471
1472         /* If the failed drive is just a ReadError, then we might need to progress
1473          * the repair/check process
1474          */
1475         if (failed == 1 && ! conf->mddev->ro &&
1476             test_bit(R5_ReadError, &sh->dev[failed_num].flags)
1477             && !test_bit(R5_LOCKED, &sh->dev[failed_num].flags)
1478             && test_bit(R5_UPTODATE, &sh->dev[failed_num].flags)
1479                 ) {
1480                 dev = &sh->dev[failed_num];
1481                 if (!test_bit(R5_ReWrite, &dev->flags)) {
1482                         set_bit(R5_Wantwrite, &dev->flags);
1483                         set_bit(R5_ReWrite, &dev->flags);
1484                         set_bit(R5_LOCKED, &dev->flags);
1485                         locked++;
1486                 } else {
1487                         /* let's read it back */
1488                         set_bit(R5_Wantread, &dev->flags);
1489                         set_bit(R5_LOCKED, &dev->flags);
1490                         locked++;
1491                 }
1492         }
1493
1494         if (expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
1495                 /* Need to write out all blocks after computing parity */
1496                 sh->disks = conf->raid_disks;
1497                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf, conf->raid_disks);
1498                 compute_parity(sh, RECONSTRUCT_WRITE);
1499                 for (i= conf->raid_disks; i--;) {
1500                         set_bit(R5_LOCKED, &sh->dev[i].flags);
1501                         locked++;
1502                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
1503                 }
1504                 clear_bit(STRIPE_EXPANDING, &sh->state);
1505         } else if (expanded) {
1506                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
1507                 wake_up(&conf->wait_for_overlap);
1508                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
1509         }
1510
1511         if (expanding && locked == 0) {
1512                 /* We have read all the blocks in this stripe and now we need to
1513                  * copy some of them into a target stripe for expand.
1514                  */
1515                 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
1516                 for (i=0; i< sh->disks; i++)
1517                         if (i != sh->pd_idx) {
1518                                 int dd_idx, pd_idx, j;
1519                                 struct stripe_head *sh2;
1520
1521                                 sector_t bn = compute_blocknr(sh, i);
1522                                 sector_t s = raid5_compute_sector(bn, conf->raid_disks,
1523                                                                   conf->raid_disks-1,
1524                                                                   &dd_idx, &pd_idx, conf);
1525                                 sh2 = get_active_stripe(conf, s, conf->raid_disks, pd_idx, 1);
1526                                 if (sh2 == NULL)
1527                                         /* so far only the early blocks of this stripe
1528                                          * have been requested.  When later blocks
1529                                          * get requested, we will try again
1530                                          */
1531                                         continue;
1532                                 if(!test_bit(STRIPE_EXPANDING, &sh2->state) ||
1533                                    test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
1534                                         /* must have already done this block */
1535                                         release_stripe(sh2);
1536                                         continue;
1537                                 }
1538                                 memcpy(page_address(sh2->dev[dd_idx].page),
1539                                        page_address(sh->dev[i].page),
1540                                        STRIPE_SIZE);
1541                                 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
1542                                 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
1543                                 for (j=0; j<conf->raid_disks; j++)
1544                                         if (j != sh2->pd_idx &&
1545                                             !test_bit(R5_Expanded, &sh2->dev[j].flags))
1546                                                 break;
1547                                 if (j == conf->raid_disks) {
1548                                         set_bit(STRIPE_EXPAND_READY, &sh2->state);
1549                                         set_bit(STRIPE_HANDLE, &sh2->state);
1550                                 }
1551                                 release_stripe(sh2);
1552                         }
1553         }
1554
1555         spin_unlock(&sh->lock);
1556
1557         while ((bi=return_bi)) {
1558                 int bytes = bi->bi_size;
1559
1560                 return_bi = bi->bi_next;
1561                 bi->bi_next = NULL;
1562                 bi->bi_size = 0;
1563                 bi->bi_end_io(bi, bytes, 0);
1564         }
1565         for (i=disks; i-- ;) {
1566                 int rw;
1567                 struct bio *bi;
1568                 mdk_rdev_t *rdev;
1569                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
1570                         rw = 1;
1571                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1572                         rw = 0;
1573                 else
1574                         continue;
1575  
1576                 bi = &sh->dev[i].req;
1577  
1578                 bi->bi_rw = rw;
1579                 if (rw)
1580                         bi->bi_end_io = raid5_end_write_request;
1581                 else
1582                         bi->bi_end_io = raid5_end_read_request;
1583  
1584                 rcu_read_lock();
1585                 rdev = rcu_dereference(conf->disks[i].rdev);
1586                 if (rdev && test_bit(Faulty, &rdev->flags))
1587                         rdev = NULL;
1588                 if (rdev)
1589                         atomic_inc(&rdev->nr_pending);
1590                 rcu_read_unlock();
1591  
1592                 if (rdev) {
1593                         if (syncing || expanding || expanded)
1594                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1595
1596                         bi->bi_bdev = rdev->bdev;
1597                         PRINTK("for %llu schedule op %ld on disc %d\n",
1598                                 (unsigned long long)sh->sector, bi->bi_rw, i);
1599                         atomic_inc(&sh->count);
1600                         bi->bi_sector = sh->sector + rdev->data_offset;
1601                         bi->bi_flags = 1 << BIO_UPTODATE;
1602                         bi->bi_vcnt = 1;        
1603                         bi->bi_max_vecs = 1;
1604                         bi->bi_idx = 0;
1605                         bi->bi_io_vec = &sh->dev[i].vec;
1606                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1607                         bi->bi_io_vec[0].bv_offset = 0;
1608                         bi->bi_size = STRIPE_SIZE;
1609                         bi->bi_next = NULL;
1610                         if (rw == WRITE &&
1611                             test_bit(R5_ReWrite, &sh->dev[i].flags))
1612                                 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1613                         generic_make_request(bi);
1614                 } else {
1615                         if (rw == 1)
1616                                 set_bit(STRIPE_DEGRADED, &sh->state);
1617                         PRINTK("skip op %ld on disc %d for sector %llu\n",
1618                                 bi->bi_rw, i, (unsigned long long)sh->sector);
1619                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1620                         set_bit(STRIPE_HANDLE, &sh->state);
1621                 }
1622         }
1623 }
1624
1625 static void raid5_activate_delayed(raid5_conf_t *conf)
1626 {
1627         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
1628                 while (!list_empty(&conf->delayed_list)) {
1629                         struct list_head *l = conf->delayed_list.next;
1630                         struct stripe_head *sh;
1631                         sh = list_entry(l, struct stripe_head, lru);
1632                         list_del_init(l);
1633                         clear_bit(STRIPE_DELAYED, &sh->state);
1634                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1635                                 atomic_inc(&conf->preread_active_stripes);
1636                         list_add_tail(&sh->lru, &conf->handle_list);
1637                 }
1638         }
1639 }
1640
1641 static void activate_bit_delay(raid5_conf_t *conf)
1642 {
1643         /* device_lock is held */
1644         struct list_head head;
1645         list_add(&head, &conf->bitmap_list);
1646         list_del_init(&conf->bitmap_list);
1647         while (!list_empty(&head)) {
1648                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
1649                 list_del_init(&sh->lru);
1650                 atomic_inc(&sh->count);
1651                 __release_stripe(conf, sh);
1652         }
1653 }
1654
1655 static void unplug_slaves(mddev_t *mddev)
1656 {
1657         raid5_conf_t *conf = mddev_to_conf(mddev);
1658         int i;
1659
1660         rcu_read_lock();
1661         for (i=0; i<mddev->raid_disks; i++) {
1662                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
1663                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
1664                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
1665
1666                         atomic_inc(&rdev->nr_pending);
1667                         rcu_read_unlock();
1668
1669                         if (r_queue->unplug_fn)
1670                                 r_queue->unplug_fn(r_queue);
1671
1672                         rdev_dec_pending(rdev, mddev);
1673                         rcu_read_lock();
1674                 }
1675         }
1676         rcu_read_unlock();
1677 }
1678
1679 static void raid5_unplug_device(request_queue_t *q)
1680 {
1681         mddev_t *mddev = q->queuedata;
1682         raid5_conf_t *conf = mddev_to_conf(mddev);
1683         unsigned long flags;
1684
1685         spin_lock_irqsave(&conf->device_lock, flags);
1686
1687         if (blk_remove_plug(q)) {
1688                 conf->seq_flush++;
1689                 raid5_activate_delayed(conf);
1690         }
1691         md_wakeup_thread(mddev->thread);
1692
1693         spin_unlock_irqrestore(&conf->device_lock, flags);
1694
1695         unplug_slaves(mddev);
1696 }
1697
1698 static int raid5_issue_flush(request_queue_t *q, struct gendisk *disk,
1699                              sector_t *error_sector)
1700 {
1701         mddev_t *mddev = q->queuedata;
1702         raid5_conf_t *conf = mddev_to_conf(mddev);
1703         int i, ret = 0;
1704
1705         rcu_read_lock();
1706         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
1707                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
1708                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
1709                         struct block_device *bdev = rdev->bdev;
1710                         request_queue_t *r_queue = bdev_get_queue(bdev);
1711
1712                         if (!r_queue->issue_flush_fn)
1713                                 ret = -EOPNOTSUPP;
1714                         else {
1715                                 atomic_inc(&rdev->nr_pending);
1716                                 rcu_read_unlock();
1717                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
1718                                                               error_sector);
1719                                 rdev_dec_pending(rdev, mddev);
1720                                 rcu_read_lock();
1721                         }
1722                 }
1723         }
1724         rcu_read_unlock();
1725         return ret;
1726 }
1727
1728 static inline void raid5_plug_device(raid5_conf_t *conf)
1729 {
1730         spin_lock_irq(&conf->device_lock);
1731         blk_plug_device(conf->mddev->queue);
1732         spin_unlock_irq(&conf->device_lock);
1733 }
1734
1735 static int make_request(request_queue_t *q, struct bio * bi)
1736 {
1737         mddev_t *mddev = q->queuedata;
1738         raid5_conf_t *conf = mddev_to_conf(mddev);
1739         unsigned int dd_idx, pd_idx;
1740         sector_t new_sector;
1741         sector_t logical_sector, last_sector;
1742         struct stripe_head *sh;
1743         const int rw = bio_data_dir(bi);
1744
1745         if (unlikely(bio_barrier(bi))) {
1746                 bio_endio(bi, bi->bi_size, -EOPNOTSUPP);
1747                 return 0;
1748         }
1749
1750         md_write_start(mddev, bi);
1751
1752         disk_stat_inc(mddev->gendisk, ios[rw]);
1753         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
1754
1755         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
1756         last_sector = bi->bi_sector + (bi->bi_size>>9);
1757         bi->bi_next = NULL;
1758         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
1759
1760         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
1761                 DEFINE_WAIT(w);
1762                 int disks;
1763                 
1764         retry:
1765                 if (likely(conf->expand_progress == MaxSector))
1766                         disks = conf->raid_disks;
1767                 else {
1768                         spin_lock_irq(&conf->device_lock);
1769                         disks = conf->raid_disks;
1770                         if (logical_sector >= conf->expand_progress)
1771                                 disks = conf->previous_raid_disks;
1772                         spin_unlock_irq(&conf->device_lock);
1773                 }
1774                 new_sector = raid5_compute_sector(logical_sector, disks, disks - 1,
1775                                                   &dd_idx, &pd_idx, conf);
1776                 PRINTK("raid5: make_request, sector %llu logical %llu\n",
1777                         (unsigned long long)new_sector, 
1778                         (unsigned long long)logical_sector);
1779
1780                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
1781                 sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
1782                 if (sh) {
1783                         if (unlikely(conf->expand_progress != MaxSector)) {
1784                                 /* expansion might have moved on while waiting for a
1785                                  * stripe, so we much do the range check again.
1786                                  */
1787                                 int must_retry = 0;
1788                                 spin_lock_irq(&conf->device_lock);
1789                                 if (logical_sector <  conf->expand_progress &&
1790                                     disks == conf->previous_raid_disks)
1791                                         /* mismatch, need to try again */
1792                                         must_retry = 1;
1793                                 spin_unlock_irq(&conf->device_lock);
1794                                 if (must_retry) {
1795                                         release_stripe(sh);
1796                                         goto retry;
1797                                 }
1798                         }
1799
1800                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
1801                             !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
1802                                 /* Stripe is busy expanding or
1803                                  * add failed due to overlap.  Flush everything
1804                                  * and wait a while
1805                                  */
1806                                 raid5_unplug_device(mddev->queue);
1807                                 release_stripe(sh);
1808                                 schedule();
1809                                 goto retry;
1810                         }
1811                         finish_wait(&conf->wait_for_overlap, &w);
1812                         raid5_plug_device(conf);
1813                         handle_stripe(sh);
1814                         release_stripe(sh);
1815                 } else {
1816                         /* cannot get stripe for read-ahead, just give-up */
1817                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1818                         finish_wait(&conf->wait_for_overlap, &w);
1819                         break;
1820                 }
1821                         
1822         }
1823         spin_lock_irq(&conf->device_lock);
1824         if (--bi->bi_phys_segments == 0) {
1825                 int bytes = bi->bi_size;
1826
1827                 if ( bio_data_dir(bi) == WRITE )
1828                         md_write_end(mddev);
1829                 bi->bi_size = 0;
1830                 bi->bi_end_io(bi, bytes, 0);
1831         }
1832         spin_unlock_irq(&conf->device_lock);
1833         return 0;
1834 }
1835
1836 /* FIXME go_faster isn't used */
1837 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1838 {
1839         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1840         struct stripe_head *sh;
1841         int pd_idx;
1842         sector_t first_sector, last_sector;
1843         int raid_disks = conf->raid_disks;
1844         int data_disks = raid_disks-1;
1845         sector_t max_sector = mddev->size << 1;
1846         int sync_blocks;
1847
1848         if (sector_nr >= max_sector) {
1849                 /* just being told to finish up .. nothing much to do */
1850                 unplug_slaves(mddev);
1851                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
1852                         end_reshape(conf);
1853                         return 0;
1854                 }
1855
1856                 if (mddev->curr_resync < max_sector) /* aborted */
1857                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1858                                         &sync_blocks, 1);
1859                 else /* compelted sync */
1860                         conf->fullsync = 0;
1861                 bitmap_close_sync(mddev->bitmap);
1862
1863                 return 0;
1864         }
1865
1866         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
1867                 /* reshaping is quite different to recovery/resync so it is
1868                  * handled quite separately ... here.
1869                  *
1870                  * On each call to sync_request, we gather one chunk worth of
1871                  * destination stripes and flag them as expanding.
1872                  * Then we find all the source stripes and request reads.
1873                  * As the reads complete, handle_stripe will copy the data
1874                  * into the destination stripe and release that stripe.
1875                  */
1876                 int i;
1877                 int dd_idx;
1878                 for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
1879                         int j;
1880                         int skipped = 0;
1881                         pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks);
1882                         sh = get_active_stripe(conf, sector_nr+i,
1883                                                conf->raid_disks, pd_idx, 0);
1884                         set_bit(STRIPE_EXPANDING, &sh->state);
1885                         /* If any of this stripe is beyond the end of the old
1886                          * array, then we need to zero those blocks
1887                          */
1888                         for (j=sh->disks; j--;) {
1889                                 sector_t s;
1890                                 if (j == sh->pd_idx)
1891                                         continue;
1892                                 s = compute_blocknr(sh, j);
1893                                 if (s < (mddev->array_size<<1)) {
1894                                         skipped = 1;
1895                                         continue;
1896                                 }
1897                                 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
1898                                 set_bit(R5_Expanded, &sh->dev[j].flags);
1899                                 set_bit(R5_UPTODATE, &sh->dev[j].flags);
1900                         }
1901                         if (!skipped) {
1902                                 set_bit(STRIPE_EXPAND_READY, &sh->state);
1903                                 set_bit(STRIPE_HANDLE, &sh->state);
1904                         }
1905                         release_stripe(sh);
1906                 }
1907                 spin_lock_irq(&conf->device_lock);
1908                 conf->expand_progress = (sector_nr + i)*(conf->raid_disks-1);
1909                 spin_unlock_irq(&conf->device_lock);
1910                 /* Ok, those stripe are ready. We can start scheduling
1911                  * reads on the source stripes.
1912                  * The source stripes are determined by mapping the first and last
1913                  * block on the destination stripes.
1914                  */
1915                 raid_disks = conf->previous_raid_disks;
1916                 data_disks = raid_disks - 1;
1917                 first_sector =
1918                         raid5_compute_sector(sector_nr*(conf->raid_disks-1),
1919                                              raid_disks, data_disks,
1920                                              &dd_idx, &pd_idx, conf);
1921                 last_sector =
1922                         raid5_compute_sector((sector_nr+conf->chunk_size/512)
1923                                                *(conf->raid_disks-1) -1,
1924                                              raid_disks, data_disks,
1925                                              &dd_idx, &pd_idx, conf);
1926                 if (last_sector >= (mddev->size<<1))
1927                         last_sector = (mddev->size<<1)-1;
1928                 while (first_sector <= last_sector) {
1929                         pd_idx = stripe_to_pdidx(first_sector, conf, conf->previous_raid_disks);
1930                         sh = get_active_stripe(conf, first_sector,
1931                                                conf->previous_raid_disks, pd_idx, 0);
1932                         set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
1933                         set_bit(STRIPE_HANDLE, &sh->state);
1934                         release_stripe(sh);
1935                         first_sector += STRIPE_SECTORS;
1936                 }
1937                 return conf->chunk_size>>9;
1938         }
1939         /* if there is 1 or more failed drives and we are trying
1940          * to resync, then assert that we are finished, because there is
1941          * nothing we can do.
1942          */
1943         if (mddev->degraded >= 1 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1944                 sector_t rv = (mddev->size << 1) - sector_nr;
1945                 *skipped = 1;
1946                 return rv;
1947         }
1948         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1949             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1950             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
1951                 /* we can skip this block, and probably more */
1952                 sync_blocks /= STRIPE_SECTORS;
1953                 *skipped = 1;
1954                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
1955         }
1956
1957         pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks);
1958         sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1);
1959         if (sh == NULL) {
1960                 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0);
1961                 /* make sure we don't swamp the stripe cache if someone else
1962                  * is trying to get access 
1963                  */
1964                 schedule_timeout_uninterruptible(1);
1965         }
1966         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 0);
1967         spin_lock(&sh->lock);   
1968         set_bit(STRIPE_SYNCING, &sh->state);
1969         clear_bit(STRIPE_INSYNC, &sh->state);
1970         spin_unlock(&sh->lock);
1971
1972         handle_stripe(sh);
1973         release_stripe(sh);
1974
1975         return STRIPE_SECTORS;
1976 }
1977
1978 /*
1979  * This is our raid5 kernel thread.
1980  *
1981  * We scan the hash table for stripes which can be handled now.
1982  * During the scan, completed stripes are saved for us by the interrupt
1983  * handler, so that they will not have to wait for our next wakeup.
1984  */
1985 static void raid5d (mddev_t *mddev)
1986 {
1987         struct stripe_head *sh;
1988         raid5_conf_t *conf = mddev_to_conf(mddev);
1989         int handled;
1990
1991         PRINTK("+++ raid5d active\n");
1992
1993         md_check_recovery(mddev);
1994
1995         handled = 0;
1996         spin_lock_irq(&conf->device_lock);
1997         while (1) {
1998                 struct list_head *first;
1999
2000                 if (conf->seq_flush - conf->seq_write > 0) {
2001                         int seq = conf->seq_flush;
2002                         spin_unlock_irq(&conf->device_lock);
2003                         bitmap_unplug(mddev->bitmap);
2004                         spin_lock_irq(&conf->device_lock);
2005                         conf->seq_write = seq;
2006                         activate_bit_delay(conf);
2007                 }
2008
2009                 if (list_empty(&conf->handle_list) &&
2010                     atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD &&
2011                     !blk_queue_plugged(mddev->queue) &&
2012                     !list_empty(&conf->delayed_list))
2013                         raid5_activate_delayed(conf);
2014
2015                 if (list_empty(&conf->handle_list))
2016                         break;
2017
2018                 first = conf->handle_list.next;
2019                 sh = list_entry(first, struct stripe_head, lru);
2020
2021                 list_del_init(first);
2022                 atomic_inc(&sh->count);
2023                 if (atomic_read(&sh->count)!= 1)
2024                         BUG();
2025                 spin_unlock_irq(&conf->device_lock);
2026                 
2027                 handled++;
2028                 handle_stripe(sh);
2029                 release_stripe(sh);
2030
2031                 spin_lock_irq(&conf->device_lock);
2032         }
2033         PRINTK("%d stripes handled\n", handled);
2034
2035         spin_unlock_irq(&conf->device_lock);
2036
2037         unplug_slaves(mddev);
2038
2039         PRINTK("--- raid5d inactive\n");
2040 }
2041
2042 static ssize_t
2043 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
2044 {
2045         raid5_conf_t *conf = mddev_to_conf(mddev);
2046         if (conf)
2047                 return sprintf(page, "%d\n", conf->max_nr_stripes);
2048         else
2049                 return 0;
2050 }
2051
2052 static ssize_t
2053 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
2054 {
2055         raid5_conf_t *conf = mddev_to_conf(mddev);
2056         char *end;
2057         int new;
2058         if (len >= PAGE_SIZE)
2059                 return -EINVAL;
2060         if (!conf)
2061                 return -ENODEV;
2062
2063         new = simple_strtoul(page, &end, 10);
2064         if (!*page || (*end && *end != '\n') )
2065                 return -EINVAL;
2066         if (new <= 16 || new > 32768)
2067                 return -EINVAL;
2068         while (new < conf->max_nr_stripes) {
2069                 if (drop_one_stripe(conf))
2070                         conf->max_nr_stripes--;
2071                 else
2072                         break;
2073         }
2074         while (new > conf->max_nr_stripes) {
2075                 if (grow_one_stripe(conf))
2076                         conf->max_nr_stripes++;
2077                 else break;
2078         }
2079         return len;
2080 }
2081
2082 static struct md_sysfs_entry
2083 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
2084                                 raid5_show_stripe_cache_size,
2085                                 raid5_store_stripe_cache_size);
2086
2087 static ssize_t
2088 stripe_cache_active_show(mddev_t *mddev, char *page)
2089 {
2090         raid5_conf_t *conf = mddev_to_conf(mddev);
2091         if (conf)
2092                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
2093         else
2094                 return 0;
2095 }
2096
2097 static struct md_sysfs_entry
2098 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
2099
2100 static struct attribute *raid5_attrs[] =  {
2101         &raid5_stripecache_size.attr,
2102         &raid5_stripecache_active.attr,
2103         NULL,
2104 };
2105 static struct attribute_group raid5_attrs_group = {
2106         .name = NULL,
2107         .attrs = raid5_attrs,
2108 };
2109
2110 static int run(mddev_t *mddev)
2111 {
2112         raid5_conf_t *conf;
2113         int raid_disk, memory;
2114         mdk_rdev_t *rdev;
2115         struct disk_info *disk;
2116         struct list_head *tmp;
2117
2118         if (mddev->level != 5 && mddev->level != 4) {
2119                 printk(KERN_ERR "raid5: %s: raid level not set to 4/5 (%d)\n",
2120                        mdname(mddev), mddev->level);
2121                 return -EIO;
2122         }
2123
2124         mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
2125         if ((conf = mddev->private) == NULL)
2126                 goto abort;
2127         conf->disks = kzalloc(mddev->raid_disks * sizeof(struct disk_info),
2128                               GFP_KERNEL);
2129         if (!conf->disks)
2130                 goto abort;
2131
2132         conf->mddev = mddev;
2133
2134         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
2135                 goto abort;
2136
2137         spin_lock_init(&conf->device_lock);
2138         init_waitqueue_head(&conf->wait_for_stripe);
2139         init_waitqueue_head(&conf->wait_for_overlap);
2140         INIT_LIST_HEAD(&conf->handle_list);
2141         INIT_LIST_HEAD(&conf->delayed_list);
2142         INIT_LIST_HEAD(&conf->bitmap_list);
2143         INIT_LIST_HEAD(&conf->inactive_list);
2144         atomic_set(&conf->active_stripes, 0);
2145         atomic_set(&conf->preread_active_stripes, 0);
2146
2147         PRINTK("raid5: run(%s) called.\n", mdname(mddev));
2148
2149         ITERATE_RDEV(mddev,rdev,tmp) {
2150                 raid_disk = rdev->raid_disk;
2151                 if (raid_disk >= mddev->raid_disks
2152                     || raid_disk < 0)
2153                         continue;
2154                 disk = conf->disks + raid_disk;
2155
2156                 disk->rdev = rdev;
2157
2158                 if (test_bit(In_sync, &rdev->flags)) {
2159                         char b[BDEVNAME_SIZE];
2160                         printk(KERN_INFO "raid5: device %s operational as raid"
2161                                 " disk %d\n", bdevname(rdev->bdev,b),
2162                                 raid_disk);
2163                         conf->working_disks++;
2164                 }
2165         }
2166
2167         conf->raid_disks = mddev->raid_disks;
2168         /*
2169          * 0 for a fully functional array, 1 for a degraded array.
2170          */
2171         mddev->degraded = conf->failed_disks = conf->raid_disks - conf->working_disks;
2172         conf->mddev = mddev;
2173         conf->chunk_size = mddev->chunk_size;
2174         conf->level = mddev->level;
2175         conf->algorithm = mddev->layout;
2176         conf->max_nr_stripes = NR_STRIPES;
2177         conf->expand_progress = MaxSector;
2178
2179         /* device size must be a multiple of chunk size */
2180         mddev->size &= ~(mddev->chunk_size/1024 -1);
2181         mddev->resync_max_sectors = mddev->size << 1;
2182
2183         if (!conf->chunk_size || conf->chunk_size % 4) {
2184                 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
2185                         conf->chunk_size, mdname(mddev));
2186                 goto abort;
2187         }
2188         if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
2189                 printk(KERN_ERR 
2190                         "raid5: unsupported parity algorithm %d for %s\n",
2191                         conf->algorithm, mdname(mddev));
2192                 goto abort;
2193         }
2194         if (mddev->degraded > 1) {
2195                 printk(KERN_ERR "raid5: not enough operational devices for %s"
2196                         " (%d/%d failed)\n",
2197                         mdname(mddev), conf->failed_disks, conf->raid_disks);
2198                 goto abort;
2199         }
2200
2201         if (mddev->degraded == 1 &&
2202             mddev->recovery_cp != MaxSector) {
2203                 if (mddev->ok_start_degraded)
2204                         printk(KERN_WARNING
2205                                "raid5: starting dirty degraded array: %s"
2206                                "- data corruption possible.\n",
2207                                mdname(mddev));
2208                 else {
2209                         printk(KERN_ERR
2210                                "raid5: cannot start dirty degraded array for %s\n",
2211                                mdname(mddev));
2212                         goto abort;
2213                 }
2214         }
2215
2216         {
2217                 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
2218                 if (!mddev->thread) {
2219                         printk(KERN_ERR 
2220                                 "raid5: couldn't allocate thread for %s\n",
2221                                 mdname(mddev));
2222                         goto abort;
2223                 }
2224         }
2225         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
2226                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
2227         if (grow_stripes(conf, conf->max_nr_stripes)) {
2228                 printk(KERN_ERR 
2229                         "raid5: couldn't allocate %dkB for buffers\n", memory);
2230                 shrink_stripes(conf);
2231                 md_unregister_thread(mddev->thread);
2232                 goto abort;
2233         } else
2234                 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
2235                         memory, mdname(mddev));
2236
2237         if (mddev->degraded == 0)
2238                 printk("raid5: raid level %d set %s active with %d out of %d"
2239                         " devices, algorithm %d\n", conf->level, mdname(mddev), 
2240                         mddev->raid_disks-mddev->degraded, mddev->raid_disks,
2241                         conf->algorithm);
2242         else
2243                 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
2244                         " out of %d devices, algorithm %d\n", conf->level,
2245                         mdname(mddev), mddev->raid_disks - mddev->degraded,
2246                         mddev->raid_disks, conf->algorithm);
2247
2248         print_raid5_conf(conf);
2249
2250         /* read-ahead size must cover two whole stripes, which is
2251          * 2 * (n-1) * chunksize where 'n' is the number of raid devices
2252          */
2253         {
2254                 int stripe = (mddev->raid_disks-1) * mddev->chunk_size
2255                         / PAGE_SIZE;
2256                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
2257                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
2258         }
2259
2260         /* Ok, everything is just fine now */
2261         sysfs_create_group(&mddev->kobj, &raid5_attrs_group);
2262
2263         mddev->queue->unplug_fn = raid5_unplug_device;
2264         mddev->queue->issue_flush_fn = raid5_issue_flush;
2265
2266         mddev->array_size =  mddev->size * (mddev->raid_disks - 1);
2267         return 0;
2268 abort:
2269         if (conf) {
2270                 print_raid5_conf(conf);
2271                 kfree(conf->disks);
2272                 kfree(conf->stripe_hashtbl);
2273                 kfree(conf);
2274         }
2275         mddev->private = NULL;
2276         printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
2277         return -EIO;
2278 }
2279
2280
2281
2282 static int stop(mddev_t *mddev)
2283 {
2284         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
2285
2286         md_unregister_thread(mddev->thread);
2287         mddev->thread = NULL;
2288         shrink_stripes(conf);
2289         kfree(conf->stripe_hashtbl);
2290         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2291         sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
2292         kfree(conf->disks);
2293         kfree(conf);
2294         mddev->private = NULL;
2295         return 0;
2296 }
2297
2298 #if RAID5_DEBUG
2299 static void print_sh (struct stripe_head *sh)
2300 {
2301         int i;
2302
2303         printk("sh %llu, pd_idx %d, state %ld.\n",
2304                 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
2305         printk("sh %llu,  count %d.\n",
2306                 (unsigned long long)sh->sector, atomic_read(&sh->count));
2307         printk("sh %llu, ", (unsigned long long)sh->sector);
2308         for (i = 0; i < sh->disks; i++) {
2309                 printk("(cache%d: %p %ld) ", 
2310                         i, sh->dev[i].page, sh->dev[i].flags);
2311         }
2312         printk("\n");
2313 }
2314
2315 static void printall (raid5_conf_t *conf)
2316 {
2317         struct stripe_head *sh;
2318         struct hlist_node *hn;
2319         int i;
2320
2321         spin_lock_irq(&conf->device_lock);
2322         for (i = 0; i < NR_HASH; i++) {
2323                 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
2324                         if (sh->raid_conf != conf)
2325                                 continue;
2326                         print_sh(sh);
2327                 }
2328         }
2329         spin_unlock_irq(&conf->device_lock);
2330 }
2331 #endif
2332
2333 static void status (struct seq_file *seq, mddev_t *mddev)
2334 {
2335         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
2336         int i;
2337
2338         seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
2339         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->working_disks);
2340         for (i = 0; i < conf->raid_disks; i++)
2341                 seq_printf (seq, "%s",
2342                                conf->disks[i].rdev &&
2343                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
2344         seq_printf (seq, "]");
2345 #if RAID5_DEBUG
2346 #define D(x) \
2347         seq_printf (seq, "<"#x":%d>", atomic_read(&conf->x))
2348         printall(conf);
2349 #endif
2350 }
2351
2352 static void print_raid5_conf (raid5_conf_t *conf)
2353 {
2354         int i;
2355         struct disk_info *tmp;
2356
2357         printk("RAID5 conf printout:\n");
2358         if (!conf) {
2359                 printk("(conf==NULL)\n");
2360                 return;
2361         }
2362         printk(" --- rd:%d wd:%d fd:%d\n", conf->raid_disks,
2363                  conf->working_disks, conf->failed_disks);
2364
2365         for (i = 0; i < conf->raid_disks; i++) {
2366                 char b[BDEVNAME_SIZE];
2367                 tmp = conf->disks + i;
2368                 if (tmp->rdev)
2369                 printk(" disk %d, o:%d, dev:%s\n",
2370                         i, !test_bit(Faulty, &tmp->rdev->flags),
2371                         bdevname(tmp->rdev->bdev,b));
2372         }
2373 }
2374
2375 static int raid5_spare_active(mddev_t *mddev)
2376 {
2377         int i;
2378         raid5_conf_t *conf = mddev->private;
2379         struct disk_info *tmp;
2380
2381         for (i = 0; i < conf->raid_disks; i++) {
2382                 tmp = conf->disks + i;
2383                 if (tmp->rdev
2384                     && !test_bit(Faulty, &tmp->rdev->flags)
2385                     && !test_bit(In_sync, &tmp->rdev->flags)) {
2386                         mddev->degraded--;
2387                         conf->failed_disks--;
2388                         conf->working_disks++;
2389                         set_bit(In_sync, &tmp->rdev->flags);
2390                 }
2391         }
2392         print_raid5_conf(conf);
2393         return 0;
2394 }
2395
2396 static int raid5_remove_disk(mddev_t *mddev, int number)
2397 {
2398         raid5_conf_t *conf = mddev->private;
2399         int err = 0;
2400         mdk_rdev_t *rdev;
2401         struct disk_info *p = conf->disks + number;
2402
2403         print_raid5_conf(conf);
2404         rdev = p->rdev;
2405         if (rdev) {
2406                 if (test_bit(In_sync, &rdev->flags) ||
2407                     atomic_read(&rdev->nr_pending)) {
2408                         err = -EBUSY;
2409                         goto abort;
2410                 }
2411                 p->rdev = NULL;
2412                 synchronize_rcu();
2413                 if (atomic_read(&rdev->nr_pending)) {
2414                         /* lost the race, try later */
2415                         err = -EBUSY;
2416                         p->rdev = rdev;
2417                 }
2418         }
2419 abort:
2420
2421         print_raid5_conf(conf);
2422         return err;
2423 }
2424
2425 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
2426 {
2427         raid5_conf_t *conf = mddev->private;
2428         int found = 0;
2429         int disk;
2430         struct disk_info *p;
2431
2432         if (mddev->degraded > 1)
2433                 /* no point adding a device */
2434                 return 0;
2435
2436         /*
2437          * find the disk ...
2438          */
2439         for (disk=0; disk < mddev->raid_disks; disk++)
2440                 if ((p=conf->disks + disk)->rdev == NULL) {
2441                         clear_bit(In_sync, &rdev->flags);
2442                         rdev->raid_disk = disk;
2443                         found = 1;
2444                         if (rdev->saved_raid_disk != disk)
2445                                 conf->fullsync = 1;
2446                         rcu_assign_pointer(p->rdev, rdev);
2447                         break;
2448                 }
2449         print_raid5_conf(conf);
2450         return found;
2451 }
2452
2453 static int raid5_resize(mddev_t *mddev, sector_t sectors)
2454 {
2455         /* no resync is happening, and there is enough space
2456          * on all devices, so we can resize.
2457          * We need to make sure resync covers any new space.
2458          * If the array is shrinking we should possibly wait until
2459          * any io in the removed space completes, but it hardly seems
2460          * worth it.
2461          */
2462         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
2463         mddev->array_size = (sectors * (mddev->raid_disks-1))>>1;
2464         set_capacity(mddev->gendisk, mddev->array_size << 1);
2465         mddev->changed = 1;
2466         if (sectors/2  > mddev->size && mddev->recovery_cp == MaxSector) {
2467                 mddev->recovery_cp = mddev->size << 1;
2468                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2469         }
2470         mddev->size = sectors /2;
2471         mddev->resync_max_sectors = sectors;
2472         return 0;
2473 }
2474
2475 #ifdef CONFIG_MD_RAID5_RESHAPE
2476 static int raid5_reshape(mddev_t *mddev, int raid_disks)
2477 {
2478         raid5_conf_t *conf = mddev_to_conf(mddev);
2479         int err;
2480         mdk_rdev_t *rdev;
2481         struct list_head *rtmp;
2482         int spares = 0;
2483         int added_devices = 0;
2484
2485         if (mddev->degraded ||
2486             test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
2487                 return -EBUSY;
2488         if (conf->raid_disks > raid_disks)
2489                 return -EINVAL; /* Cannot shrink array yet */
2490         if (conf->raid_disks == raid_disks)
2491                 return 0; /* nothing to do */
2492
2493         /* Can only proceed if there are plenty of stripe_heads.
2494          * We need a minimum of one full stripe,, and for sensible progress
2495          * it is best to have about 4 times that.
2496          * If we require 4 times, then the default 256 4K stripe_heads will
2497          * allow for chunk sizes up to 256K, which is probably OK.
2498          * If the chunk size is greater, user-space should request more
2499          * stripe_heads first.
2500          */
2501         if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
2502                 printk(KERN_WARNING "raid5: reshape: not enough stripes.  Needed %lu\n",
2503                        (mddev->chunk_size / STRIPE_SIZE)*4);
2504                 return -ENOSPC;
2505         }
2506
2507         ITERATE_RDEV(mddev, rdev, rtmp)
2508                 if (rdev->raid_disk < 0 &&
2509                     !test_bit(Faulty, &rdev->flags))
2510                         spares++;
2511         if (conf->raid_disks + spares < raid_disks-1)
2512                 /* Not enough devices even to make a degraded array
2513                  * of that size
2514                  */
2515                 return -EINVAL;
2516
2517         err = resize_stripes(conf, raid_disks);
2518         if (err)
2519                 return err;
2520
2521         spin_lock_irq(&conf->device_lock);
2522         conf->previous_raid_disks = conf->raid_disks;
2523         mddev->raid_disks = conf->raid_disks = raid_disks;
2524         conf->expand_progress = 0;
2525         spin_unlock_irq(&conf->device_lock);
2526
2527         /* Add some new drives, as many as will fit.
2528          * We know there are enough to make the newly sized array work.
2529          */
2530         ITERATE_RDEV(mddev, rdev, rtmp)
2531                 if (rdev->raid_disk < 0 &&
2532                     !test_bit(Faulty, &rdev->flags)) {
2533                         if (raid5_add_disk(mddev, rdev)) {
2534                                 char nm[20];
2535                                 set_bit(In_sync, &rdev->flags);
2536                                 conf->working_disks++;
2537                                 added_devices++;
2538                                 sprintf(nm, "rd%d", rdev->raid_disk);
2539                                 sysfs_create_link(&mddev->kobj, &rdev->kobj, nm);
2540                         } else
2541                                 break;
2542                 }
2543
2544         mddev->degraded = (raid_disks - conf->previous_raid_disks) - added_devices;
2545         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
2546         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
2547         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
2548         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
2549         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
2550                                                 "%s_reshape");
2551         if (!mddev->sync_thread) {
2552                 mddev->recovery = 0;
2553                 spin_lock_irq(&conf->device_lock);
2554                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
2555                 conf->expand_progress = MaxSector;
2556                 spin_unlock_irq(&conf->device_lock);
2557                 return -EAGAIN;
2558         }
2559         md_wakeup_thread(mddev->sync_thread);
2560         md_new_event(mddev);
2561         return 0;
2562 }
2563 #endif
2564
2565 static void end_reshape(raid5_conf_t *conf)
2566 {
2567         struct block_device *bdev;
2568
2569         conf->mddev->array_size = conf->mddev->size * (conf->mddev->raid_disks-1);
2570         set_capacity(conf->mddev->gendisk, conf->mddev->array_size << 1);
2571         conf->mddev->changed = 1;
2572
2573         bdev = bdget_disk(conf->mddev->gendisk, 0);
2574         if (bdev) {
2575                 mutex_lock(&bdev->bd_inode->i_mutex);
2576                 i_size_write(bdev->bd_inode, conf->mddev->array_size << 10);
2577                 mutex_unlock(&bdev->bd_inode->i_mutex);
2578                 bdput(bdev);
2579         }
2580         spin_lock_irq(&conf->device_lock);
2581         conf->expand_progress = MaxSector;
2582         spin_unlock_irq(&conf->device_lock);
2583 }
2584
2585 static void raid5_quiesce(mddev_t *mddev, int state)
2586 {
2587         raid5_conf_t *conf = mddev_to_conf(mddev);
2588
2589         switch(state) {
2590         case 1: /* stop all writes */
2591                 spin_lock_irq(&conf->device_lock);
2592                 conf->quiesce = 1;
2593                 wait_event_lock_irq(conf->wait_for_stripe,
2594                                     atomic_read(&conf->active_stripes) == 0,
2595                                     conf->device_lock, /* nothing */);
2596                 spin_unlock_irq(&conf->device_lock);
2597                 break;
2598
2599         case 0: /* re-enable writes */
2600                 spin_lock_irq(&conf->device_lock);
2601                 conf->quiesce = 0;
2602                 wake_up(&conf->wait_for_stripe);
2603                 spin_unlock_irq(&conf->device_lock);
2604                 break;
2605         }
2606 }
2607
2608 static struct mdk_personality raid5_personality =
2609 {
2610         .name           = "raid5",
2611         .level          = 5,
2612         .owner          = THIS_MODULE,
2613         .make_request   = make_request,
2614         .run            = run,
2615         .stop           = stop,
2616         .status         = status,
2617         .error_handler  = error,
2618         .hot_add_disk   = raid5_add_disk,
2619         .hot_remove_disk= raid5_remove_disk,
2620         .spare_active   = raid5_spare_active,
2621         .sync_request   = sync_request,
2622         .resize         = raid5_resize,
2623 #ifdef CONFIG_MD_RAID5_RESHAPE
2624         .reshape        = raid5_reshape,
2625 #endif
2626         .quiesce        = raid5_quiesce,
2627 };
2628
2629 static struct mdk_personality raid4_personality =
2630 {
2631         .name           = "raid4",
2632         .level          = 4,
2633         .owner          = THIS_MODULE,
2634         .make_request   = make_request,
2635         .run            = run,
2636         .stop           = stop,
2637         .status         = status,
2638         .error_handler  = error,
2639         .hot_add_disk   = raid5_add_disk,
2640         .hot_remove_disk= raid5_remove_disk,
2641         .spare_active   = raid5_spare_active,
2642         .sync_request   = sync_request,
2643         .resize         = raid5_resize,
2644         .quiesce        = raid5_quiesce,
2645 };
2646
2647 static int __init raid5_init(void)
2648 {
2649         register_md_personality(&raid5_personality);
2650         register_md_personality(&raid4_personality);
2651         return 0;
2652 }
2653
2654 static void raid5_exit(void)
2655 {
2656         unregister_md_personality(&raid5_personality);
2657         unregister_md_personality(&raid4_personality);
2658 }
2659
2660 module_init(raid5_init);
2661 module_exit(raid5_exit);
2662 MODULE_LICENSE("GPL");
2663 MODULE_ALIAS("md-personality-4"); /* RAID5 */
2664 MODULE_ALIAS("md-raid5");
2665 MODULE_ALIAS("md-raid4");
2666 MODULE_ALIAS("md-level-5");
2667 MODULE_ALIAS("md-level-4");