be008f034ada5e222c18043cf0b766426c86f03c
[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  *         Copyright (C) 2002, 2003 H. Peter Anvin
6  *
7  * RAID-4/5/6 management functions.
8  * Thanks to Penguin Computing for making the RAID-6 development possible
9  * by donating a test server!
10  *
11  * This program is free software; you can redistribute it and/or modify
12  * it under the terms of the GNU General Public License as published by
13  * the Free Software Foundation; either version 2, or (at your option)
14  * any later version.
15  *
16  * You should have received a copy of the GNU General Public License
17  * (for example /usr/src/linux/COPYING); if not, write to the Free
18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19  */
20
21 /*
22  * BITMAP UNPLUGGING:
23  *
24  * The sequencing for updating the bitmap reliably is a little
25  * subtle (and I got it wrong the first time) so it deserves some
26  * explanation.
27  *
28  * We group bitmap updates into batches.  Each batch has a number.
29  * We may write out several batches at once, but that isn't very important.
30  * conf->bm_write is the number of the last batch successfully written.
31  * conf->bm_flush is the number of the last batch that was closed to
32  *    new additions.
33  * When we discover that we will need to write to any block in a stripe
34  * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35  * the number of the batch it will be in. This is bm_flush+1.
36  * When we are ready to do a write, if that batch hasn't been written yet,
37  *   we plug the array and queue the stripe for later.
38  * When an unplug happens, we increment bm_flush, thus closing the current
39  *   batch.
40  * When we notice that bm_flush > bm_write, we write out all pending updates
41  * to the bitmap, and advance bm_write to where bm_flush was.
42  * This may occasionally write a bit out twice, but is sure never to
43  * miss any bits.
44  */
45
46 #include <linux/module.h>
47 #include <linux/slab.h>
48 #include <linux/highmem.h>
49 #include <linux/bitops.h>
50 #include <linux/kthread.h>
51 #include <asm/atomic.h>
52 #include "raid6.h"
53
54 #include <linux/raid/bitmap.h>
55
56 /*
57  * Stripe cache
58  */
59
60 #define NR_STRIPES              256
61 #define STRIPE_SIZE             PAGE_SIZE
62 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
63 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
64 #define IO_THRESHOLD            1
65 #define NR_HASH                 (PAGE_SIZE / sizeof(struct hlist_head))
66 #define HASH_MASK               (NR_HASH - 1)
67
68 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
69
70 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
71  * order without overlap.  There may be several bio's per stripe+device, and
72  * a bio could span several devices.
73  * When walking this list for a particular stripe+device, we must never proceed
74  * beyond a bio that extends past this device, as the next bio might no longer
75  * be valid.
76  * This macro is used to determine the 'next' bio in the list, given the sector
77  * of the current stripe+device
78  */
79 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
80 /*
81  * The following can be used to debug the driver
82  */
83 #define RAID5_DEBUG     0
84 #define RAID5_PARANOIA  1
85 #if RAID5_PARANOIA && defined(CONFIG_SMP)
86 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
87 #else
88 # define CHECK_DEVLOCK()
89 #endif
90
91 #define PRINTK(x...) ((void)(RAID5_DEBUG && printk(x)))
92 #if RAID5_DEBUG
93 #define inline
94 #define __inline__
95 #endif
96
97 #if !RAID6_USE_EMPTY_ZERO_PAGE
98 /* In .bss so it's zeroed */
99 const char raid6_empty_zero_page[PAGE_SIZE] __attribute__((aligned(256)));
100 #endif
101
102 static inline int raid6_next_disk(int disk, int raid_disks)
103 {
104         disk++;
105         return (disk < raid_disks) ? disk : 0;
106 }
107 static void print_raid5_conf (raid5_conf_t *conf);
108
109 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
110 {
111         if (atomic_dec_and_test(&sh->count)) {
112                 BUG_ON(!list_empty(&sh->lru));
113                 BUG_ON(atomic_read(&conf->active_stripes)==0);
114                 if (test_bit(STRIPE_HANDLE, &sh->state)) {
115                         if (test_bit(STRIPE_DELAYED, &sh->state)) {
116                                 list_add_tail(&sh->lru, &conf->delayed_list);
117                                 blk_plug_device(conf->mddev->queue);
118                         } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
119                                    sh->bm_seq - conf->seq_write > 0) {
120                                 list_add_tail(&sh->lru, &conf->bitmap_list);
121                                 blk_plug_device(conf->mddev->queue);
122                         } else {
123                                 clear_bit(STRIPE_BIT_DELAY, &sh->state);
124                                 list_add_tail(&sh->lru, &conf->handle_list);
125                         }
126                         md_wakeup_thread(conf->mddev->thread);
127                 } else {
128                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
129                                 atomic_dec(&conf->preread_active_stripes);
130                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
131                                         md_wakeup_thread(conf->mddev->thread);
132                         }
133                         atomic_dec(&conf->active_stripes);
134                         if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
135                                 list_add_tail(&sh->lru, &conf->inactive_list);
136                                 wake_up(&conf->wait_for_stripe);
137                                 if (conf->retry_read_aligned)
138                                         md_wakeup_thread(conf->mddev->thread);
139                         }
140                 }
141         }
142 }
143 static void release_stripe(struct stripe_head *sh)
144 {
145         raid5_conf_t *conf = sh->raid_conf;
146         unsigned long flags;
147
148         spin_lock_irqsave(&conf->device_lock, flags);
149         __release_stripe(conf, sh);
150         spin_unlock_irqrestore(&conf->device_lock, flags);
151 }
152
153 static inline void remove_hash(struct stripe_head *sh)
154 {
155         PRINTK("remove_hash(), stripe %llu\n", (unsigned long long)sh->sector);
156
157         hlist_del_init(&sh->hash);
158 }
159
160 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
161 {
162         struct hlist_head *hp = stripe_hash(conf, sh->sector);
163
164         PRINTK("insert_hash(), stripe %llu\n", (unsigned long long)sh->sector);
165
166         CHECK_DEVLOCK();
167         hlist_add_head(&sh->hash, hp);
168 }
169
170
171 /* find an idle stripe, make sure it is unhashed, and return it. */
172 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
173 {
174         struct stripe_head *sh = NULL;
175         struct list_head *first;
176
177         CHECK_DEVLOCK();
178         if (list_empty(&conf->inactive_list))
179                 goto out;
180         first = conf->inactive_list.next;
181         sh = list_entry(first, struct stripe_head, lru);
182         list_del_init(first);
183         remove_hash(sh);
184         atomic_inc(&conf->active_stripes);
185 out:
186         return sh;
187 }
188
189 static void shrink_buffers(struct stripe_head *sh, int num)
190 {
191         struct page *p;
192         int i;
193
194         for (i=0; i<num ; i++) {
195                 p = sh->dev[i].page;
196                 if (!p)
197                         continue;
198                 sh->dev[i].page = NULL;
199                 put_page(p);
200         }
201 }
202
203 static int grow_buffers(struct stripe_head *sh, int num)
204 {
205         int i;
206
207         for (i=0; i<num; i++) {
208                 struct page *page;
209
210                 if (!(page = alloc_page(GFP_KERNEL))) {
211                         return 1;
212                 }
213                 sh->dev[i].page = page;
214         }
215         return 0;
216 }
217
218 static void raid5_build_block (struct stripe_head *sh, int i);
219
220 static void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx, int disks)
221 {
222         raid5_conf_t *conf = sh->raid_conf;
223         int i;
224
225         BUG_ON(atomic_read(&sh->count) != 0);
226         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
227         
228         CHECK_DEVLOCK();
229         PRINTK("init_stripe called, stripe %llu\n", 
230                 (unsigned long long)sh->sector);
231
232         remove_hash(sh);
233
234         sh->sector = sector;
235         sh->pd_idx = pd_idx;
236         sh->state = 0;
237
238         sh->disks = disks;
239
240         for (i = sh->disks; i--; ) {
241                 struct r5dev *dev = &sh->dev[i];
242
243                 if (dev->toread || dev->towrite || dev->written ||
244                     test_bit(R5_LOCKED, &dev->flags)) {
245                         printk("sector=%llx i=%d %p %p %p %d\n",
246                                (unsigned long long)sh->sector, i, dev->toread,
247                                dev->towrite, dev->written,
248                                test_bit(R5_LOCKED, &dev->flags));
249                         BUG();
250                 }
251                 dev->flags = 0;
252                 raid5_build_block(sh, i);
253         }
254         insert_hash(conf, sh);
255 }
256
257 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector, int disks)
258 {
259         struct stripe_head *sh;
260         struct hlist_node *hn;
261
262         CHECK_DEVLOCK();
263         PRINTK("__find_stripe, sector %llu\n", (unsigned long long)sector);
264         hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
265                 if (sh->sector == sector && sh->disks == disks)
266                         return sh;
267         PRINTK("__stripe %llu not in cache\n", (unsigned long long)sector);
268         return NULL;
269 }
270
271 static void unplug_slaves(mddev_t *mddev);
272 static void raid5_unplug_device(request_queue_t *q);
273
274 static struct stripe_head *get_active_stripe(raid5_conf_t *conf, sector_t sector, int disks,
275                                              int pd_idx, int noblock)
276 {
277         struct stripe_head *sh;
278
279         PRINTK("get_stripe, sector %llu\n", (unsigned long long)sector);
280
281         spin_lock_irq(&conf->device_lock);
282
283         do {
284                 wait_event_lock_irq(conf->wait_for_stripe,
285                                     conf->quiesce == 0,
286                                     conf->device_lock, /* nothing */);
287                 sh = __find_stripe(conf, sector, disks);
288                 if (!sh) {
289                         if (!conf->inactive_blocked)
290                                 sh = get_free_stripe(conf);
291                         if (noblock && sh == NULL)
292                                 break;
293                         if (!sh) {
294                                 conf->inactive_blocked = 1;
295                                 wait_event_lock_irq(conf->wait_for_stripe,
296                                                     !list_empty(&conf->inactive_list) &&
297                                                     (atomic_read(&conf->active_stripes)
298                                                      < (conf->max_nr_stripes *3/4)
299                                                      || !conf->inactive_blocked),
300                                                     conf->device_lock,
301                                                     raid5_unplug_device(conf->mddev->queue)
302                                         );
303                                 conf->inactive_blocked = 0;
304                         } else
305                                 init_stripe(sh, sector, pd_idx, disks);
306                 } else {
307                         if (atomic_read(&sh->count)) {
308                           BUG_ON(!list_empty(&sh->lru));
309                         } else {
310                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
311                                         atomic_inc(&conf->active_stripes);
312                                 if (list_empty(&sh->lru) &&
313                                     !test_bit(STRIPE_EXPANDING, &sh->state))
314                                         BUG();
315                                 list_del_init(&sh->lru);
316                         }
317                 }
318         } while (sh == NULL);
319
320         if (sh)
321                 atomic_inc(&sh->count);
322
323         spin_unlock_irq(&conf->device_lock);
324         return sh;
325 }
326
327 static int grow_one_stripe(raid5_conf_t *conf)
328 {
329         struct stripe_head *sh;
330         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
331         if (!sh)
332                 return 0;
333         memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
334         sh->raid_conf = conf;
335         spin_lock_init(&sh->lock);
336
337         if (grow_buffers(sh, conf->raid_disks)) {
338                 shrink_buffers(sh, conf->raid_disks);
339                 kmem_cache_free(conf->slab_cache, sh);
340                 return 0;
341         }
342         sh->disks = conf->raid_disks;
343         /* we just created an active stripe so... */
344         atomic_set(&sh->count, 1);
345         atomic_inc(&conf->active_stripes);
346         INIT_LIST_HEAD(&sh->lru);
347         release_stripe(sh);
348         return 1;
349 }
350
351 static int grow_stripes(raid5_conf_t *conf, int num)
352 {
353         struct kmem_cache *sc;
354         int devs = conf->raid_disks;
355
356         sprintf(conf->cache_name[0], "raid5/%s", mdname(conf->mddev));
357         sprintf(conf->cache_name[1], "raid5/%s-alt", mdname(conf->mddev));
358         conf->active_name = 0;
359         sc = kmem_cache_create(conf->cache_name[conf->active_name],
360                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
361                                0, 0, NULL, NULL);
362         if (!sc)
363                 return 1;
364         conf->slab_cache = sc;
365         conf->pool_size = devs;
366         while (num--)
367                 if (!grow_one_stripe(conf))
368                         return 1;
369         return 0;
370 }
371
372 #ifdef CONFIG_MD_RAID5_RESHAPE
373 static int resize_stripes(raid5_conf_t *conf, int newsize)
374 {
375         /* Make all the stripes able to hold 'newsize' devices.
376          * New slots in each stripe get 'page' set to a new page.
377          *
378          * This happens in stages:
379          * 1/ create a new kmem_cache and allocate the required number of
380          *    stripe_heads.
381          * 2/ gather all the old stripe_heads and tranfer the pages across
382          *    to the new stripe_heads.  This will have the side effect of
383          *    freezing the array as once all stripe_heads have been collected,
384          *    no IO will be possible.  Old stripe heads are freed once their
385          *    pages have been transferred over, and the old kmem_cache is
386          *    freed when all stripes are done.
387          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
388          *    we simple return a failre status - no need to clean anything up.
389          * 4/ allocate new pages for the new slots in the new stripe_heads.
390          *    If this fails, we don't bother trying the shrink the
391          *    stripe_heads down again, we just leave them as they are.
392          *    As each stripe_head is processed the new one is released into
393          *    active service.
394          *
395          * Once step2 is started, we cannot afford to wait for a write,
396          * so we use GFP_NOIO allocations.
397          */
398         struct stripe_head *osh, *nsh;
399         LIST_HEAD(newstripes);
400         struct disk_info *ndisks;
401         int err = 0;
402         struct kmem_cache *sc;
403         int i;
404
405         if (newsize <= conf->pool_size)
406                 return 0; /* never bother to shrink */
407
408         /* Step 1 */
409         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
410                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
411                                0, 0, NULL, NULL);
412         if (!sc)
413                 return -ENOMEM;
414
415         for (i = conf->max_nr_stripes; i; i--) {
416                 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
417                 if (!nsh)
418                         break;
419
420                 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
421
422                 nsh->raid_conf = conf;
423                 spin_lock_init(&nsh->lock);
424
425                 list_add(&nsh->lru, &newstripes);
426         }
427         if (i) {
428                 /* didn't get enough, give up */
429                 while (!list_empty(&newstripes)) {
430                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
431                         list_del(&nsh->lru);
432                         kmem_cache_free(sc, nsh);
433                 }
434                 kmem_cache_destroy(sc);
435                 return -ENOMEM;
436         }
437         /* Step 2 - Must use GFP_NOIO now.
438          * OK, we have enough stripes, start collecting inactive
439          * stripes and copying them over
440          */
441         list_for_each_entry(nsh, &newstripes, lru) {
442                 spin_lock_irq(&conf->device_lock);
443                 wait_event_lock_irq(conf->wait_for_stripe,
444                                     !list_empty(&conf->inactive_list),
445                                     conf->device_lock,
446                                     unplug_slaves(conf->mddev)
447                         );
448                 osh = get_free_stripe(conf);
449                 spin_unlock_irq(&conf->device_lock);
450                 atomic_set(&nsh->count, 1);
451                 for(i=0; i<conf->pool_size; i++)
452                         nsh->dev[i].page = osh->dev[i].page;
453                 for( ; i<newsize; i++)
454                         nsh->dev[i].page = NULL;
455                 kmem_cache_free(conf->slab_cache, osh);
456         }
457         kmem_cache_destroy(conf->slab_cache);
458
459         /* Step 3.
460          * At this point, we are holding all the stripes so the array
461          * is completely stalled, so now is a good time to resize
462          * conf->disks.
463          */
464         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
465         if (ndisks) {
466                 for (i=0; i<conf->raid_disks; i++)
467                         ndisks[i] = conf->disks[i];
468                 kfree(conf->disks);
469                 conf->disks = ndisks;
470         } else
471                 err = -ENOMEM;
472
473         /* Step 4, return new stripes to service */
474         while(!list_empty(&newstripes)) {
475                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
476                 list_del_init(&nsh->lru);
477                 for (i=conf->raid_disks; i < newsize; i++)
478                         if (nsh->dev[i].page == NULL) {
479                                 struct page *p = alloc_page(GFP_NOIO);
480                                 nsh->dev[i].page = p;
481                                 if (!p)
482                                         err = -ENOMEM;
483                         }
484                 release_stripe(nsh);
485         }
486         /* critical section pass, GFP_NOIO no longer needed */
487
488         conf->slab_cache = sc;
489         conf->active_name = 1-conf->active_name;
490         conf->pool_size = newsize;
491         return err;
492 }
493 #endif
494
495 static int drop_one_stripe(raid5_conf_t *conf)
496 {
497         struct stripe_head *sh;
498
499         spin_lock_irq(&conf->device_lock);
500         sh = get_free_stripe(conf);
501         spin_unlock_irq(&conf->device_lock);
502         if (!sh)
503                 return 0;
504         BUG_ON(atomic_read(&sh->count));
505         shrink_buffers(sh, conf->pool_size);
506         kmem_cache_free(conf->slab_cache, sh);
507         atomic_dec(&conf->active_stripes);
508         return 1;
509 }
510
511 static void shrink_stripes(raid5_conf_t *conf)
512 {
513         while (drop_one_stripe(conf))
514                 ;
515
516         if (conf->slab_cache)
517                 kmem_cache_destroy(conf->slab_cache);
518         conf->slab_cache = NULL;
519 }
520
521 static int raid5_end_read_request(struct bio * bi, unsigned int bytes_done,
522                                    int error)
523 {
524         struct stripe_head *sh = bi->bi_private;
525         raid5_conf_t *conf = sh->raid_conf;
526         int disks = sh->disks, i;
527         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
528         char b[BDEVNAME_SIZE];
529         mdk_rdev_t *rdev;
530
531         if (bi->bi_size)
532                 return 1;
533
534         for (i=0 ; i<disks; i++)
535                 if (bi == &sh->dev[i].req)
536                         break;
537
538         PRINTK("end_read_request %llu/%d, count: %d, uptodate %d.\n", 
539                 (unsigned long long)sh->sector, i, atomic_read(&sh->count), 
540                 uptodate);
541         if (i == disks) {
542                 BUG();
543                 return 0;
544         }
545
546         if (uptodate) {
547                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
548                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
549                         rdev = conf->disks[i].rdev;
550                         printk(KERN_INFO "raid5:%s: read error corrected (%lu sectors at %llu on %s)\n",
551                                mdname(conf->mddev), STRIPE_SECTORS,
552                                (unsigned long long)sh->sector + rdev->data_offset,
553                                bdevname(rdev->bdev, b));
554                         clear_bit(R5_ReadError, &sh->dev[i].flags);
555                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
556                 }
557                 if (atomic_read(&conf->disks[i].rdev->read_errors))
558                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
559         } else {
560                 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
561                 int retry = 0;
562                 rdev = conf->disks[i].rdev;
563
564                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
565                 atomic_inc(&rdev->read_errors);
566                 if (conf->mddev->degraded)
567                         printk(KERN_WARNING "raid5:%s: read error not correctable (sector %llu on %s).\n",
568                                mdname(conf->mddev),
569                                (unsigned long long)sh->sector + rdev->data_offset,
570                                bdn);
571                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
572                         /* Oh, no!!! */
573                         printk(KERN_WARNING "raid5:%s: read error NOT corrected!! (sector %llu on %s).\n",
574                                mdname(conf->mddev),
575                                (unsigned long long)sh->sector + rdev->data_offset,
576                                bdn);
577                 else if (atomic_read(&rdev->read_errors)
578                          > conf->max_nr_stripes)
579                         printk(KERN_WARNING
580                                "raid5:%s: Too many read errors, failing device %s.\n",
581                                mdname(conf->mddev), bdn);
582                 else
583                         retry = 1;
584                 if (retry)
585                         set_bit(R5_ReadError, &sh->dev[i].flags);
586                 else {
587                         clear_bit(R5_ReadError, &sh->dev[i].flags);
588                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
589                         md_error(conf->mddev, rdev);
590                 }
591         }
592         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
593         clear_bit(R5_LOCKED, &sh->dev[i].flags);
594         set_bit(STRIPE_HANDLE, &sh->state);
595         release_stripe(sh);
596         return 0;
597 }
598
599 static int raid5_end_write_request (struct bio *bi, unsigned int bytes_done,
600                                     int error)
601 {
602         struct stripe_head *sh = bi->bi_private;
603         raid5_conf_t *conf = sh->raid_conf;
604         int disks = sh->disks, i;
605         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
606
607         if (bi->bi_size)
608                 return 1;
609
610         for (i=0 ; i<disks; i++)
611                 if (bi == &sh->dev[i].req)
612                         break;
613
614         PRINTK("end_write_request %llu/%d, count %d, uptodate: %d.\n", 
615                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
616                 uptodate);
617         if (i == disks) {
618                 BUG();
619                 return 0;
620         }
621
622         if (!uptodate)
623                 md_error(conf->mddev, conf->disks[i].rdev);
624
625         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
626         
627         clear_bit(R5_LOCKED, &sh->dev[i].flags);
628         set_bit(STRIPE_HANDLE, &sh->state);
629         release_stripe(sh);
630         return 0;
631 }
632
633
634 static sector_t compute_blocknr(struct stripe_head *sh, int i);
635         
636 static void raid5_build_block (struct stripe_head *sh, int i)
637 {
638         struct r5dev *dev = &sh->dev[i];
639
640         bio_init(&dev->req);
641         dev->req.bi_io_vec = &dev->vec;
642         dev->req.bi_vcnt++;
643         dev->req.bi_max_vecs++;
644         dev->vec.bv_page = dev->page;
645         dev->vec.bv_len = STRIPE_SIZE;
646         dev->vec.bv_offset = 0;
647
648         dev->req.bi_sector = sh->sector;
649         dev->req.bi_private = sh;
650
651         dev->flags = 0;
652         dev->sector = compute_blocknr(sh, i);
653 }
654
655 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
656 {
657         char b[BDEVNAME_SIZE];
658         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
659         PRINTK("raid5: error called\n");
660
661         if (!test_bit(Faulty, &rdev->flags)) {
662                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
663                 if (test_and_clear_bit(In_sync, &rdev->flags)) {
664                         unsigned long flags;
665                         spin_lock_irqsave(&conf->device_lock, flags);
666                         mddev->degraded++;
667                         spin_unlock_irqrestore(&conf->device_lock, flags);
668                         /*
669                          * if recovery was running, make sure it aborts.
670                          */
671                         set_bit(MD_RECOVERY_ERR, &mddev->recovery);
672                 }
673                 set_bit(Faulty, &rdev->flags);
674                 printk (KERN_ALERT
675                         "raid5: Disk failure on %s, disabling device."
676                         " Operation continuing on %d devices\n",
677                         bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
678         }
679 }
680
681 /*
682  * Input: a 'big' sector number,
683  * Output: index of the data and parity disk, and the sector # in them.
684  */
685 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
686                         unsigned int data_disks, unsigned int * dd_idx,
687                         unsigned int * pd_idx, raid5_conf_t *conf)
688 {
689         long stripe;
690         unsigned long chunk_number;
691         unsigned int chunk_offset;
692         sector_t new_sector;
693         int sectors_per_chunk = conf->chunk_size >> 9;
694
695         /* First compute the information on this sector */
696
697         /*
698          * Compute the chunk number and the sector offset inside the chunk
699          */
700         chunk_offset = sector_div(r_sector, sectors_per_chunk);
701         chunk_number = r_sector;
702         BUG_ON(r_sector != chunk_number);
703
704         /*
705          * Compute the stripe number
706          */
707         stripe = chunk_number / data_disks;
708
709         /*
710          * Compute the data disk and parity disk indexes inside the stripe
711          */
712         *dd_idx = chunk_number % data_disks;
713
714         /*
715          * Select the parity disk based on the user selected algorithm.
716          */
717         switch(conf->level) {
718         case 4:
719                 *pd_idx = data_disks;
720                 break;
721         case 5:
722                 switch (conf->algorithm) {
723                 case ALGORITHM_LEFT_ASYMMETRIC:
724                         *pd_idx = data_disks - stripe % raid_disks;
725                         if (*dd_idx >= *pd_idx)
726                                 (*dd_idx)++;
727                         break;
728                 case ALGORITHM_RIGHT_ASYMMETRIC:
729                         *pd_idx = stripe % raid_disks;
730                         if (*dd_idx >= *pd_idx)
731                                 (*dd_idx)++;
732                         break;
733                 case ALGORITHM_LEFT_SYMMETRIC:
734                         *pd_idx = data_disks - stripe % raid_disks;
735                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
736                         break;
737                 case ALGORITHM_RIGHT_SYMMETRIC:
738                         *pd_idx = stripe % raid_disks;
739                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
740                         break;
741                 default:
742                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
743                                 conf->algorithm);
744                 }
745                 break;
746         case 6:
747
748                 /**** FIX THIS ****/
749                 switch (conf->algorithm) {
750                 case ALGORITHM_LEFT_ASYMMETRIC:
751                         *pd_idx = raid_disks - 1 - (stripe % raid_disks);
752                         if (*pd_idx == raid_disks-1)
753                                 (*dd_idx)++;    /* Q D D D P */
754                         else if (*dd_idx >= *pd_idx)
755                                 (*dd_idx) += 2; /* D D P Q D */
756                         break;
757                 case ALGORITHM_RIGHT_ASYMMETRIC:
758                         *pd_idx = stripe % raid_disks;
759                         if (*pd_idx == raid_disks-1)
760                                 (*dd_idx)++;    /* Q D D D P */
761                         else if (*dd_idx >= *pd_idx)
762                                 (*dd_idx) += 2; /* D D P Q D */
763                         break;
764                 case ALGORITHM_LEFT_SYMMETRIC:
765                         *pd_idx = raid_disks - 1 - (stripe % raid_disks);
766                         *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
767                         break;
768                 case ALGORITHM_RIGHT_SYMMETRIC:
769                         *pd_idx = stripe % raid_disks;
770                         *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
771                         break;
772                 default:
773                         printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
774                                 conf->algorithm);
775                 }
776                 break;
777         }
778
779         /*
780          * Finally, compute the new sector number
781          */
782         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
783         return new_sector;
784 }
785
786
787 static sector_t compute_blocknr(struct stripe_head *sh, int i)
788 {
789         raid5_conf_t *conf = sh->raid_conf;
790         int raid_disks = sh->disks;
791         int data_disks = raid_disks - conf->max_degraded;
792         sector_t new_sector = sh->sector, check;
793         int sectors_per_chunk = conf->chunk_size >> 9;
794         sector_t stripe;
795         int chunk_offset;
796         int chunk_number, dummy1, dummy2, dd_idx = i;
797         sector_t r_sector;
798
799
800         chunk_offset = sector_div(new_sector, sectors_per_chunk);
801         stripe = new_sector;
802         BUG_ON(new_sector != stripe);
803
804         if (i == sh->pd_idx)
805                 return 0;
806         switch(conf->level) {
807         case 4: break;
808         case 5:
809                 switch (conf->algorithm) {
810                 case ALGORITHM_LEFT_ASYMMETRIC:
811                 case ALGORITHM_RIGHT_ASYMMETRIC:
812                         if (i > sh->pd_idx)
813                                 i--;
814                         break;
815                 case ALGORITHM_LEFT_SYMMETRIC:
816                 case ALGORITHM_RIGHT_SYMMETRIC:
817                         if (i < sh->pd_idx)
818                                 i += raid_disks;
819                         i -= (sh->pd_idx + 1);
820                         break;
821                 default:
822                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
823                                conf->algorithm);
824                 }
825                 break;
826         case 6:
827                 if (i == raid6_next_disk(sh->pd_idx, raid_disks))
828                         return 0; /* It is the Q disk */
829                 switch (conf->algorithm) {
830                 case ALGORITHM_LEFT_ASYMMETRIC:
831                 case ALGORITHM_RIGHT_ASYMMETRIC:
832                         if (sh->pd_idx == raid_disks-1)
833                                 i--;    /* Q D D D P */
834                         else if (i > sh->pd_idx)
835                                 i -= 2; /* D D P Q D */
836                         break;
837                 case ALGORITHM_LEFT_SYMMETRIC:
838                 case ALGORITHM_RIGHT_SYMMETRIC:
839                         if (sh->pd_idx == raid_disks-1)
840                                 i--; /* Q D D D P */
841                         else {
842                                 /* D D P Q D */
843                                 if (i < sh->pd_idx)
844                                         i += raid_disks;
845                                 i -= (sh->pd_idx + 2);
846                         }
847                         break;
848                 default:
849                         printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
850                                 conf->algorithm);
851                 }
852                 break;
853         }
854
855         chunk_number = stripe * data_disks + i;
856         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
857
858         check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
859         if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
860                 printk(KERN_ERR "compute_blocknr: map not correct\n");
861                 return 0;
862         }
863         return r_sector;
864 }
865
866
867
868 /*
869  * Copy data between a page in the stripe cache, and one or more bion
870  * The page could align with the middle of the bio, or there could be
871  * several bion, each with several bio_vecs, which cover part of the page
872  * Multiple bion are linked together on bi_next.  There may be extras
873  * at the end of this list.  We ignore them.
874  */
875 static void copy_data(int frombio, struct bio *bio,
876                      struct page *page,
877                      sector_t sector)
878 {
879         char *pa = page_address(page);
880         struct bio_vec *bvl;
881         int i;
882         int page_offset;
883
884         if (bio->bi_sector >= sector)
885                 page_offset = (signed)(bio->bi_sector - sector) * 512;
886         else
887                 page_offset = (signed)(sector - bio->bi_sector) * -512;
888         bio_for_each_segment(bvl, bio, i) {
889                 int len = bio_iovec_idx(bio,i)->bv_len;
890                 int clen;
891                 int b_offset = 0;
892
893                 if (page_offset < 0) {
894                         b_offset = -page_offset;
895                         page_offset += b_offset;
896                         len -= b_offset;
897                 }
898
899                 if (len > 0 && page_offset + len > STRIPE_SIZE)
900                         clen = STRIPE_SIZE - page_offset;
901                 else clen = len;
902
903                 if (clen > 0) {
904                         char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
905                         if (frombio)
906                                 memcpy(pa+page_offset, ba+b_offset, clen);
907                         else
908                                 memcpy(ba+b_offset, pa+page_offset, clen);
909                         __bio_kunmap_atomic(ba, KM_USER0);
910                 }
911                 if (clen < len) /* hit end of page */
912                         break;
913                 page_offset +=  len;
914         }
915 }
916
917 #define check_xor()     do {                                            \
918                            if (count == MAX_XOR_BLOCKS) {               \
919                                 xor_block(count, STRIPE_SIZE, ptr);     \
920                                 count = 1;                              \
921                            }                                            \
922                         } while(0)
923
924
925 static void compute_block(struct stripe_head *sh, int dd_idx)
926 {
927         int i, count, disks = sh->disks;
928         void *ptr[MAX_XOR_BLOCKS], *p;
929
930         PRINTK("compute_block, stripe %llu, idx %d\n", 
931                 (unsigned long long)sh->sector, dd_idx);
932
933         ptr[0] = page_address(sh->dev[dd_idx].page);
934         memset(ptr[0], 0, STRIPE_SIZE);
935         count = 1;
936         for (i = disks ; i--; ) {
937                 if (i == dd_idx)
938                         continue;
939                 p = page_address(sh->dev[i].page);
940                 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
941                         ptr[count++] = p;
942                 else
943                         printk(KERN_ERR "compute_block() %d, stripe %llu, %d"
944                                 " not present\n", dd_idx,
945                                 (unsigned long long)sh->sector, i);
946
947                 check_xor();
948         }
949         if (count != 1)
950                 xor_block(count, STRIPE_SIZE, ptr);
951         set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
952 }
953
954 static void compute_parity5(struct stripe_head *sh, int method)
955 {
956         raid5_conf_t *conf = sh->raid_conf;
957         int i, pd_idx = sh->pd_idx, disks = sh->disks, count;
958         void *ptr[MAX_XOR_BLOCKS];
959         struct bio *chosen;
960
961         PRINTK("compute_parity5, stripe %llu, method %d\n",
962                 (unsigned long long)sh->sector, method);
963
964         count = 1;
965         ptr[0] = page_address(sh->dev[pd_idx].page);
966         switch(method) {
967         case READ_MODIFY_WRITE:
968                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags));
969                 for (i=disks ; i-- ;) {
970                         if (i==pd_idx)
971                                 continue;
972                         if (sh->dev[i].towrite &&
973                             test_bit(R5_UPTODATE, &sh->dev[i].flags)) {
974                                 ptr[count++] = page_address(sh->dev[i].page);
975                                 chosen = sh->dev[i].towrite;
976                                 sh->dev[i].towrite = NULL;
977
978                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
979                                         wake_up(&conf->wait_for_overlap);
980
981                                 BUG_ON(sh->dev[i].written);
982                                 sh->dev[i].written = chosen;
983                                 check_xor();
984                         }
985                 }
986                 break;
987         case RECONSTRUCT_WRITE:
988                 memset(ptr[0], 0, STRIPE_SIZE);
989                 for (i= disks; i-- ;)
990                         if (i!=pd_idx && sh->dev[i].towrite) {
991                                 chosen = sh->dev[i].towrite;
992                                 sh->dev[i].towrite = NULL;
993
994                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
995                                         wake_up(&conf->wait_for_overlap);
996
997                                 BUG_ON(sh->dev[i].written);
998                                 sh->dev[i].written = chosen;
999                         }
1000                 break;
1001         case CHECK_PARITY:
1002                 break;
1003         }
1004         if (count>1) {
1005                 xor_block(count, STRIPE_SIZE, ptr);
1006                 count = 1;
1007         }
1008         
1009         for (i = disks; i--;)
1010                 if (sh->dev[i].written) {
1011                         sector_t sector = sh->dev[i].sector;
1012                         struct bio *wbi = sh->dev[i].written;
1013                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1014                                 copy_data(1, wbi, sh->dev[i].page, sector);
1015                                 wbi = r5_next_bio(wbi, sector);
1016                         }
1017
1018                         set_bit(R5_LOCKED, &sh->dev[i].flags);
1019                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
1020                 }
1021
1022         switch(method) {
1023         case RECONSTRUCT_WRITE:
1024         case CHECK_PARITY:
1025                 for (i=disks; i--;)
1026                         if (i != pd_idx) {
1027                                 ptr[count++] = page_address(sh->dev[i].page);
1028                                 check_xor();
1029                         }
1030                 break;
1031         case READ_MODIFY_WRITE:
1032                 for (i = disks; i--;)
1033                         if (sh->dev[i].written) {
1034                                 ptr[count++] = page_address(sh->dev[i].page);
1035                                 check_xor();
1036                         }
1037         }
1038         if (count != 1)
1039                 xor_block(count, STRIPE_SIZE, ptr);
1040         
1041         if (method != CHECK_PARITY) {
1042                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1043                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
1044         } else
1045                 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1046 }
1047
1048 static void compute_parity6(struct stripe_head *sh, int method)
1049 {
1050         raid6_conf_t *conf = sh->raid_conf;
1051         int i, pd_idx = sh->pd_idx, qd_idx, d0_idx, disks = conf->raid_disks, count;
1052         struct bio *chosen;
1053         /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1054         void *ptrs[disks];
1055
1056         qd_idx = raid6_next_disk(pd_idx, disks);
1057         d0_idx = raid6_next_disk(qd_idx, disks);
1058
1059         PRINTK("compute_parity, stripe %llu, method %d\n",
1060                 (unsigned long long)sh->sector, method);
1061
1062         switch(method) {
1063         case READ_MODIFY_WRITE:
1064                 BUG();          /* READ_MODIFY_WRITE N/A for RAID-6 */
1065         case RECONSTRUCT_WRITE:
1066                 for (i= disks; i-- ;)
1067                         if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1068                                 chosen = sh->dev[i].towrite;
1069                                 sh->dev[i].towrite = NULL;
1070
1071                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1072                                         wake_up(&conf->wait_for_overlap);
1073
1074                                 BUG_ON(sh->dev[i].written);
1075                                 sh->dev[i].written = chosen;
1076                         }
1077                 break;
1078         case CHECK_PARITY:
1079                 BUG();          /* Not implemented yet */
1080         }
1081
1082         for (i = disks; i--;)
1083                 if (sh->dev[i].written) {
1084                         sector_t sector = sh->dev[i].sector;
1085                         struct bio *wbi = sh->dev[i].written;
1086                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1087                                 copy_data(1, wbi, sh->dev[i].page, sector);
1088                                 wbi = r5_next_bio(wbi, sector);
1089                         }
1090
1091                         set_bit(R5_LOCKED, &sh->dev[i].flags);
1092                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
1093                 }
1094
1095 //      switch(method) {
1096 //      case RECONSTRUCT_WRITE:
1097 //      case CHECK_PARITY:
1098 //      case UPDATE_PARITY:
1099                 /* Note that unlike RAID-5, the ordering of the disks matters greatly. */
1100                 /* FIX: Is this ordering of drives even remotely optimal? */
1101                 count = 0;
1102                 i = d0_idx;
1103                 do {
1104                         ptrs[count++] = page_address(sh->dev[i].page);
1105                         if (count <= disks-2 && !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1106                                 printk("block %d/%d not uptodate on parity calc\n", i,count);
1107                         i = raid6_next_disk(i, disks);
1108                 } while ( i != d0_idx );
1109 //              break;
1110 //      }
1111
1112         raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs);
1113
1114         switch(method) {
1115         case RECONSTRUCT_WRITE:
1116                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1117                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1118                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
1119                 set_bit(R5_LOCKED,   &sh->dev[qd_idx].flags);
1120                 break;
1121         case UPDATE_PARITY:
1122                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1123                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1124                 break;
1125         }
1126 }
1127
1128
1129 /* Compute one missing block */
1130 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1131 {
1132         raid6_conf_t *conf = sh->raid_conf;
1133         int i, count, disks = conf->raid_disks;
1134         void *ptr[MAX_XOR_BLOCKS], *p;
1135         int pd_idx = sh->pd_idx;
1136         int qd_idx = raid6_next_disk(pd_idx, disks);
1137
1138         PRINTK("compute_block_1, stripe %llu, idx %d\n",
1139                 (unsigned long long)sh->sector, dd_idx);
1140
1141         if ( dd_idx == qd_idx ) {
1142                 /* We're actually computing the Q drive */
1143                 compute_parity6(sh, UPDATE_PARITY);
1144         } else {
1145                 ptr[0] = page_address(sh->dev[dd_idx].page);
1146                 if (!nozero) memset(ptr[0], 0, STRIPE_SIZE);
1147                 count = 1;
1148                 for (i = disks ; i--; ) {
1149                         if (i == dd_idx || i == qd_idx)
1150                                 continue;
1151                         p = page_address(sh->dev[i].page);
1152                         if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1153                                 ptr[count++] = p;
1154                         else
1155                                 printk("compute_block() %d, stripe %llu, %d"
1156                                        " not present\n", dd_idx,
1157                                        (unsigned long long)sh->sector, i);
1158
1159                         check_xor();
1160                 }
1161                 if (count != 1)
1162                         xor_block(count, STRIPE_SIZE, ptr);
1163                 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1164                 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1165         }
1166 }
1167
1168 /* Compute two missing blocks */
1169 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1170 {
1171         raid6_conf_t *conf = sh->raid_conf;
1172         int i, count, disks = conf->raid_disks;
1173         int pd_idx = sh->pd_idx;
1174         int qd_idx = raid6_next_disk(pd_idx, disks);
1175         int d0_idx = raid6_next_disk(qd_idx, disks);
1176         int faila, failb;
1177
1178         /* faila and failb are disk numbers relative to d0_idx */
1179         /* pd_idx become disks-2 and qd_idx become disks-1 */
1180         faila = (dd_idx1 < d0_idx) ? dd_idx1+(disks-d0_idx) : dd_idx1-d0_idx;
1181         failb = (dd_idx2 < d0_idx) ? dd_idx2+(disks-d0_idx) : dd_idx2-d0_idx;
1182
1183         BUG_ON(faila == failb);
1184         if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1185
1186         PRINTK("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1187                (unsigned long long)sh->sector, dd_idx1, dd_idx2, faila, failb);
1188
1189         if ( failb == disks-1 ) {
1190                 /* Q disk is one of the missing disks */
1191                 if ( faila == disks-2 ) {
1192                         /* Missing P+Q, just recompute */
1193                         compute_parity6(sh, UPDATE_PARITY);
1194                         return;
1195                 } else {
1196                         /* We're missing D+Q; recompute D from P */
1197                         compute_block_1(sh, (dd_idx1 == qd_idx) ? dd_idx2 : dd_idx1, 0);
1198                         compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1199                         return;
1200                 }
1201         }
1202
1203         /* We're missing D+P or D+D; build pointer table */
1204         {
1205                 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1206                 void *ptrs[disks];
1207
1208                 count = 0;
1209                 i = d0_idx;
1210                 do {
1211                         ptrs[count++] = page_address(sh->dev[i].page);
1212                         i = raid6_next_disk(i, disks);
1213                         if (i != dd_idx1 && i != dd_idx2 &&
1214                             !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1215                                 printk("compute_2 with missing block %d/%d\n", count, i);
1216                 } while ( i != d0_idx );
1217
1218                 if ( failb == disks-2 ) {
1219                         /* We're missing D+P. */
1220                         raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs);
1221                 } else {
1222                         /* We're missing D+D. */
1223                         raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs);
1224                 }
1225
1226                 /* Both the above update both missing blocks */
1227                 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1228                 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1229         }
1230 }
1231
1232
1233
1234 /*
1235  * Each stripe/dev can have one or more bion attached.
1236  * toread/towrite point to the first in a chain.
1237  * The bi_next chain must be in order.
1238  */
1239 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1240 {
1241         struct bio **bip;
1242         raid5_conf_t *conf = sh->raid_conf;
1243         int firstwrite=0;
1244
1245         PRINTK("adding bh b#%llu to stripe s#%llu\n",
1246                 (unsigned long long)bi->bi_sector,
1247                 (unsigned long long)sh->sector);
1248
1249
1250         spin_lock(&sh->lock);
1251         spin_lock_irq(&conf->device_lock);
1252         if (forwrite) {
1253                 bip = &sh->dev[dd_idx].towrite;
1254                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1255                         firstwrite = 1;
1256         } else
1257                 bip = &sh->dev[dd_idx].toread;
1258         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1259                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1260                         goto overlap;
1261                 bip = & (*bip)->bi_next;
1262         }
1263         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1264                 goto overlap;
1265
1266         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1267         if (*bip)
1268                 bi->bi_next = *bip;
1269         *bip = bi;
1270         bi->bi_phys_segments ++;
1271         spin_unlock_irq(&conf->device_lock);
1272         spin_unlock(&sh->lock);
1273
1274         PRINTK("added bi b#%llu to stripe s#%llu, disk %d.\n",
1275                 (unsigned long long)bi->bi_sector,
1276                 (unsigned long long)sh->sector, dd_idx);
1277
1278         if (conf->mddev->bitmap && firstwrite) {
1279                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1280                                   STRIPE_SECTORS, 0);
1281                 sh->bm_seq = conf->seq_flush+1;
1282                 set_bit(STRIPE_BIT_DELAY, &sh->state);
1283         }
1284
1285         if (forwrite) {
1286                 /* check if page is covered */
1287                 sector_t sector = sh->dev[dd_idx].sector;
1288                 for (bi=sh->dev[dd_idx].towrite;
1289                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1290                              bi && bi->bi_sector <= sector;
1291                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1292                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1293                                 sector = bi->bi_sector + (bi->bi_size>>9);
1294                 }
1295                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1296                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1297         }
1298         return 1;
1299
1300  overlap:
1301         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1302         spin_unlock_irq(&conf->device_lock);
1303         spin_unlock(&sh->lock);
1304         return 0;
1305 }
1306
1307 static void end_reshape(raid5_conf_t *conf);
1308
1309 static int page_is_zero(struct page *p)
1310 {
1311         char *a = page_address(p);
1312         return ((*(u32*)a) == 0 &&
1313                 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1314 }
1315
1316 static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks)
1317 {
1318         int sectors_per_chunk = conf->chunk_size >> 9;
1319         int pd_idx, dd_idx;
1320         int chunk_offset = sector_div(stripe, sectors_per_chunk);
1321
1322         raid5_compute_sector(stripe * (disks - conf->max_degraded)
1323                              *sectors_per_chunk + chunk_offset,
1324                              disks, disks - conf->max_degraded,
1325                              &dd_idx, &pd_idx, conf);
1326         return pd_idx;
1327 }
1328
1329
1330 /*
1331  * handle_stripe - do things to a stripe.
1332  *
1333  * We lock the stripe and then examine the state of various bits
1334  * to see what needs to be done.
1335  * Possible results:
1336  *    return some read request which now have data
1337  *    return some write requests which are safely on disc
1338  *    schedule a read on some buffers
1339  *    schedule a write of some buffers
1340  *    return confirmation of parity correctness
1341  *
1342  * Parity calculations are done inside the stripe lock
1343  * buffers are taken off read_list or write_list, and bh_cache buffers
1344  * get BH_Lock set before the stripe lock is released.
1345  *
1346  */
1347  
1348 static void handle_stripe5(struct stripe_head *sh)
1349 {
1350         raid5_conf_t *conf = sh->raid_conf;
1351         int disks = sh->disks;
1352         struct bio *return_bi= NULL;
1353         struct bio *bi;
1354         int i;
1355         int syncing, expanding, expanded;
1356         int locked=0, uptodate=0, to_read=0, to_write=0, failed=0, written=0;
1357         int non_overwrite = 0;
1358         int failed_num=0;
1359         struct r5dev *dev;
1360
1361         PRINTK("handling stripe %llu, cnt=%d, pd_idx=%d\n",
1362                 (unsigned long long)sh->sector, atomic_read(&sh->count),
1363                 sh->pd_idx);
1364
1365         spin_lock(&sh->lock);
1366         clear_bit(STRIPE_HANDLE, &sh->state);
1367         clear_bit(STRIPE_DELAYED, &sh->state);
1368
1369         syncing = test_bit(STRIPE_SYNCING, &sh->state);
1370         expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
1371         expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
1372         /* Now to look around and see what can be done */
1373
1374         rcu_read_lock();
1375         for (i=disks; i--; ) {
1376                 mdk_rdev_t *rdev;
1377                 dev = &sh->dev[i];
1378                 clear_bit(R5_Insync, &dev->flags);
1379
1380                 PRINTK("check %d: state 0x%lx read %p write %p written %p\n",
1381                         i, dev->flags, dev->toread, dev->towrite, dev->written);
1382                 /* maybe we can reply to a read */
1383                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
1384                         struct bio *rbi, *rbi2;
1385                         PRINTK("Return read for disc %d\n", i);
1386                         spin_lock_irq(&conf->device_lock);
1387                         rbi = dev->toread;
1388                         dev->toread = NULL;
1389                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
1390                                 wake_up(&conf->wait_for_overlap);
1391                         spin_unlock_irq(&conf->device_lock);
1392                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1393                                 copy_data(0, rbi, dev->page, dev->sector);
1394                                 rbi2 = r5_next_bio(rbi, dev->sector);
1395                                 spin_lock_irq(&conf->device_lock);
1396                                 if (--rbi->bi_phys_segments == 0) {
1397                                         rbi->bi_next = return_bi;
1398                                         return_bi = rbi;
1399                                 }
1400                                 spin_unlock_irq(&conf->device_lock);
1401                                 rbi = rbi2;
1402                         }
1403                 }
1404
1405                 /* now count some things */
1406                 if (test_bit(R5_LOCKED, &dev->flags)) locked++;
1407                 if (test_bit(R5_UPTODATE, &dev->flags)) uptodate++;
1408
1409                 
1410                 if (dev->toread) to_read++;
1411                 if (dev->towrite) {
1412                         to_write++;
1413                         if (!test_bit(R5_OVERWRITE, &dev->flags))
1414                                 non_overwrite++;
1415                 }
1416                 if (dev->written) written++;
1417                 rdev = rcu_dereference(conf->disks[i].rdev);
1418                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
1419                         /* The ReadError flag will just be confusing now */
1420                         clear_bit(R5_ReadError, &dev->flags);
1421                         clear_bit(R5_ReWrite, &dev->flags);
1422                 }
1423                 if (!rdev || !test_bit(In_sync, &rdev->flags)
1424                     || test_bit(R5_ReadError, &dev->flags)) {
1425                         failed++;
1426                         failed_num = i;
1427                 } else
1428                         set_bit(R5_Insync, &dev->flags);
1429         }
1430         rcu_read_unlock();
1431         PRINTK("locked=%d uptodate=%d to_read=%d"
1432                 " to_write=%d failed=%d failed_num=%d\n",
1433                 locked, uptodate, to_read, to_write, failed, failed_num);
1434         /* check if the array has lost two devices and, if so, some requests might
1435          * need to be failed
1436          */
1437         if (failed > 1 && to_read+to_write+written) {
1438                 for (i=disks; i--; ) {
1439                         int bitmap_end = 0;
1440
1441                         if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1442                                 mdk_rdev_t *rdev;
1443                                 rcu_read_lock();
1444                                 rdev = rcu_dereference(conf->disks[i].rdev);
1445                                 if (rdev && test_bit(In_sync, &rdev->flags))
1446                                         /* multiple read failures in one stripe */
1447                                         md_error(conf->mddev, rdev);
1448                                 rcu_read_unlock();
1449                         }
1450
1451                         spin_lock_irq(&conf->device_lock);
1452                         /* fail all writes first */
1453                         bi = sh->dev[i].towrite;
1454                         sh->dev[i].towrite = NULL;
1455                         if (bi) { to_write--; bitmap_end = 1; }
1456
1457                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1458                                 wake_up(&conf->wait_for_overlap);
1459
1460                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1461                                 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1462                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1463                                 if (--bi->bi_phys_segments == 0) {
1464                                         md_write_end(conf->mddev);
1465                                         bi->bi_next = return_bi;
1466                                         return_bi = bi;
1467                                 }
1468                                 bi = nextbi;
1469                         }
1470                         /* and fail all 'written' */
1471                         bi = sh->dev[i].written;
1472                         sh->dev[i].written = NULL;
1473                         if (bi) bitmap_end = 1;
1474                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS) {
1475                                 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1476                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1477                                 if (--bi->bi_phys_segments == 0) {
1478                                         md_write_end(conf->mddev);
1479                                         bi->bi_next = return_bi;
1480                                         return_bi = bi;
1481                                 }
1482                                 bi = bi2;
1483                         }
1484
1485                         /* fail any reads if this device is non-operational */
1486                         if (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1487                             test_bit(R5_ReadError, &sh->dev[i].flags)) {
1488                                 bi = sh->dev[i].toread;
1489                                 sh->dev[i].toread = NULL;
1490                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1491                                         wake_up(&conf->wait_for_overlap);
1492                                 if (bi) to_read--;
1493                                 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1494                                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1495                                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1496                                         if (--bi->bi_phys_segments == 0) {
1497                                                 bi->bi_next = return_bi;
1498                                                 return_bi = bi;
1499                                         }
1500                                         bi = nextbi;
1501                                 }
1502                         }
1503                         spin_unlock_irq(&conf->device_lock);
1504                         if (bitmap_end)
1505                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1506                                                 STRIPE_SECTORS, 0, 0);
1507                 }
1508         }
1509         if (failed > 1 && syncing) {
1510                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
1511                 clear_bit(STRIPE_SYNCING, &sh->state);
1512                 syncing = 0;
1513         }
1514
1515         /* might be able to return some write requests if the parity block
1516          * is safe, or on a failed drive
1517          */
1518         dev = &sh->dev[sh->pd_idx];
1519         if ( written &&
1520              ( (test_bit(R5_Insync, &dev->flags) && !test_bit(R5_LOCKED, &dev->flags) &&
1521                 test_bit(R5_UPTODATE, &dev->flags))
1522                || (failed == 1 && failed_num == sh->pd_idx))
1523             ) {
1524             /* any written block on an uptodate or failed drive can be returned.
1525              * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but 
1526              * never LOCKED, so we don't need to test 'failed' directly.
1527              */
1528             for (i=disks; i--; )
1529                 if (sh->dev[i].written) {
1530                     dev = &sh->dev[i];
1531                     if (!test_bit(R5_LOCKED, &dev->flags) &&
1532                          test_bit(R5_UPTODATE, &dev->flags) ) {
1533                         /* We can return any write requests */
1534                             struct bio *wbi, *wbi2;
1535                             int bitmap_end = 0;
1536                             PRINTK("Return write for disc %d\n", i);
1537                             spin_lock_irq(&conf->device_lock);
1538                             wbi = dev->written;
1539                             dev->written = NULL;
1540                             while (wbi && wbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1541                                     wbi2 = r5_next_bio(wbi, dev->sector);
1542                                     if (--wbi->bi_phys_segments == 0) {
1543                                             md_write_end(conf->mddev);
1544                                             wbi->bi_next = return_bi;
1545                                             return_bi = wbi;
1546                                     }
1547                                     wbi = wbi2;
1548                             }
1549                             if (dev->towrite == NULL)
1550                                     bitmap_end = 1;
1551                             spin_unlock_irq(&conf->device_lock);
1552                             if (bitmap_end)
1553                                     bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1554                                                     STRIPE_SECTORS,
1555                                                     !test_bit(STRIPE_DEGRADED, &sh->state), 0);
1556                     }
1557                 }
1558         }
1559
1560         /* Now we might consider reading some blocks, either to check/generate
1561          * parity, or to satisfy requests
1562          * or to load a block that is being partially written.
1563          */
1564         if (to_read || non_overwrite || (syncing && (uptodate < disks)) || expanding) {
1565                 for (i=disks; i--;) {
1566                         dev = &sh->dev[i];
1567                         if (!test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1568                             (dev->toread ||
1569                              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1570                              syncing ||
1571                              expanding ||
1572                              (failed && (sh->dev[failed_num].toread ||
1573                                          (sh->dev[failed_num].towrite && !test_bit(R5_OVERWRITE, &sh->dev[failed_num].flags))))
1574                                     )
1575                                 ) {
1576                                 /* we would like to get this block, possibly
1577                                  * by computing it, but we might not be able to
1578                                  */
1579                                 if (uptodate == disks-1) {
1580                                         PRINTK("Computing block %d\n", i);
1581                                         compute_block(sh, i);
1582                                         uptodate++;
1583                                 } else if (test_bit(R5_Insync, &dev->flags)) {
1584                                         set_bit(R5_LOCKED, &dev->flags);
1585                                         set_bit(R5_Wantread, &dev->flags);
1586                                         locked++;
1587                                         PRINTK("Reading block %d (sync=%d)\n", 
1588                                                 i, syncing);
1589                                 }
1590                         }
1591                 }
1592                 set_bit(STRIPE_HANDLE, &sh->state);
1593         }
1594
1595         /* now to consider writing and what else, if anything should be read */
1596         if (to_write) {
1597                 int rmw=0, rcw=0;
1598                 for (i=disks ; i--;) {
1599                         /* would I have to read this buffer for read_modify_write */
1600                         dev = &sh->dev[i];
1601                         if ((dev->towrite || i == sh->pd_idx) &&
1602                             (!test_bit(R5_LOCKED, &dev->flags) 
1603                                     ) &&
1604                             !test_bit(R5_UPTODATE, &dev->flags)) {
1605                                 if (test_bit(R5_Insync, &dev->flags)
1606 /*                                  && !(!mddev->insync && i == sh->pd_idx) */
1607                                         )
1608                                         rmw++;
1609                                 else rmw += 2*disks;  /* cannot read it */
1610                         }
1611                         /* Would I have to read this buffer for reconstruct_write */
1612                         if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
1613                             (!test_bit(R5_LOCKED, &dev->flags) 
1614                                     ) &&
1615                             !test_bit(R5_UPTODATE, &dev->flags)) {
1616                                 if (test_bit(R5_Insync, &dev->flags)) rcw++;
1617                                 else rcw += 2*disks;
1618                         }
1619                 }
1620                 PRINTK("for sector %llu, rmw=%d rcw=%d\n", 
1621                         (unsigned long long)sh->sector, rmw, rcw);
1622                 set_bit(STRIPE_HANDLE, &sh->state);
1623                 if (rmw < rcw && rmw > 0)
1624                         /* prefer read-modify-write, but need to get some data */
1625                         for (i=disks; i--;) {
1626                                 dev = &sh->dev[i];
1627                                 if ((dev->towrite || i == sh->pd_idx) &&
1628                                     !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1629                                     test_bit(R5_Insync, &dev->flags)) {
1630                                         if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1631                                         {
1632                                                 PRINTK("Read_old block %d for r-m-w\n", i);
1633                                                 set_bit(R5_LOCKED, &dev->flags);
1634                                                 set_bit(R5_Wantread, &dev->flags);
1635                                                 locked++;
1636                                         } else {
1637                                                 set_bit(STRIPE_DELAYED, &sh->state);
1638                                                 set_bit(STRIPE_HANDLE, &sh->state);
1639                                         }
1640                                 }
1641                         }
1642                 if (rcw <= rmw && rcw > 0)
1643                         /* want reconstruct write, but need to get some data */
1644                         for (i=disks; i--;) {
1645                                 dev = &sh->dev[i];
1646                                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
1647                                     !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1648                                     test_bit(R5_Insync, &dev->flags)) {
1649                                         if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1650                                         {
1651                                                 PRINTK("Read_old block %d for Reconstruct\n", i);
1652                                                 set_bit(R5_LOCKED, &dev->flags);
1653                                                 set_bit(R5_Wantread, &dev->flags);
1654                                                 locked++;
1655                                         } else {
1656                                                 set_bit(STRIPE_DELAYED, &sh->state);
1657                                                 set_bit(STRIPE_HANDLE, &sh->state);
1658                                         }
1659                                 }
1660                         }
1661                 /* now if nothing is locked, and if we have enough data, we can start a write request */
1662                 if (locked == 0 && (rcw == 0 ||rmw == 0) &&
1663                     !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
1664                         PRINTK("Computing parity...\n");
1665                         compute_parity5(sh, rcw==0 ? RECONSTRUCT_WRITE : READ_MODIFY_WRITE);
1666                         /* now every locked buffer is ready to be written */
1667                         for (i=disks; i--;)
1668                                 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
1669                                         PRINTK("Writing block %d\n", i);
1670                                         locked++;
1671                                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
1672                                         if (!test_bit(R5_Insync, &sh->dev[i].flags)
1673                                             || (i==sh->pd_idx && failed == 0))
1674                                                 set_bit(STRIPE_INSYNC, &sh->state);
1675                                 }
1676                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
1677                                 atomic_dec(&conf->preread_active_stripes);
1678                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
1679                                         md_wakeup_thread(conf->mddev->thread);
1680                         }
1681                 }
1682         }
1683
1684         /* maybe we need to check and possibly fix the parity for this stripe
1685          * Any reads will already have been scheduled, so we just see if enough data
1686          * is available
1687          */
1688         if (syncing && locked == 0 &&
1689             !test_bit(STRIPE_INSYNC, &sh->state)) {
1690                 set_bit(STRIPE_HANDLE, &sh->state);
1691                 if (failed == 0) {
1692                         BUG_ON(uptodate != disks);
1693                         compute_parity5(sh, CHECK_PARITY);
1694                         uptodate--;
1695                         if (page_is_zero(sh->dev[sh->pd_idx].page)) {
1696                                 /* parity is correct (on disc, not in buffer any more) */
1697                                 set_bit(STRIPE_INSYNC, &sh->state);
1698                         } else {
1699                                 conf->mddev->resync_mismatches += STRIPE_SECTORS;
1700                                 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
1701                                         /* don't try to repair!! */
1702                                         set_bit(STRIPE_INSYNC, &sh->state);
1703                                 else {
1704                                         compute_block(sh, sh->pd_idx);
1705                                         uptodate++;
1706                                 }
1707                         }
1708                 }
1709                 if (!test_bit(STRIPE_INSYNC, &sh->state)) {
1710                         /* either failed parity check, or recovery is happening */
1711                         if (failed==0)
1712                                 failed_num = sh->pd_idx;
1713                         dev = &sh->dev[failed_num];
1714                         BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
1715                         BUG_ON(uptodate != disks);
1716
1717                         set_bit(R5_LOCKED, &dev->flags);
1718                         set_bit(R5_Wantwrite, &dev->flags);
1719                         clear_bit(STRIPE_DEGRADED, &sh->state);
1720                         locked++;
1721                         set_bit(STRIPE_INSYNC, &sh->state);
1722                 }
1723         }
1724         if (syncing && locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
1725                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
1726                 clear_bit(STRIPE_SYNCING, &sh->state);
1727         }
1728
1729         /* If the failed drive is just a ReadError, then we might need to progress
1730          * the repair/check process
1731          */
1732         if (failed == 1 && ! conf->mddev->ro &&
1733             test_bit(R5_ReadError, &sh->dev[failed_num].flags)
1734             && !test_bit(R5_LOCKED, &sh->dev[failed_num].flags)
1735             && test_bit(R5_UPTODATE, &sh->dev[failed_num].flags)
1736                 ) {
1737                 dev = &sh->dev[failed_num];
1738                 if (!test_bit(R5_ReWrite, &dev->flags)) {
1739                         set_bit(R5_Wantwrite, &dev->flags);
1740                         set_bit(R5_ReWrite, &dev->flags);
1741                         set_bit(R5_LOCKED, &dev->flags);
1742                         locked++;
1743                 } else {
1744                         /* let's read it back */
1745                         set_bit(R5_Wantread, &dev->flags);
1746                         set_bit(R5_LOCKED, &dev->flags);
1747                         locked++;
1748                 }
1749         }
1750
1751         if (expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
1752                 /* Need to write out all blocks after computing parity */
1753                 sh->disks = conf->raid_disks;
1754                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf, conf->raid_disks);
1755                 compute_parity5(sh, RECONSTRUCT_WRITE);
1756                 for (i= conf->raid_disks; i--;) {
1757                         set_bit(R5_LOCKED, &sh->dev[i].flags);
1758                         locked++;
1759                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
1760                 }
1761                 clear_bit(STRIPE_EXPANDING, &sh->state);
1762         } else if (expanded) {
1763                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
1764                 atomic_dec(&conf->reshape_stripes);
1765                 wake_up(&conf->wait_for_overlap);
1766                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
1767         }
1768
1769         if (expanding && locked == 0) {
1770                 /* We have read all the blocks in this stripe and now we need to
1771                  * copy some of them into a target stripe for expand.
1772                  */
1773                 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
1774                 for (i=0; i< sh->disks; i++)
1775                         if (i != sh->pd_idx) {
1776                                 int dd_idx, pd_idx, j;
1777                                 struct stripe_head *sh2;
1778
1779                                 sector_t bn = compute_blocknr(sh, i);
1780                                 sector_t s = raid5_compute_sector(bn, conf->raid_disks,
1781                                                                   conf->raid_disks-1,
1782                                                                   &dd_idx, &pd_idx, conf);
1783                                 sh2 = get_active_stripe(conf, s, conf->raid_disks, pd_idx, 1);
1784                                 if (sh2 == NULL)
1785                                         /* so far only the early blocks of this stripe
1786                                          * have been requested.  When later blocks
1787                                          * get requested, we will try again
1788                                          */
1789                                         continue;
1790                                 if(!test_bit(STRIPE_EXPANDING, &sh2->state) ||
1791                                    test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
1792                                         /* must have already done this block */
1793                                         release_stripe(sh2);
1794                                         continue;
1795                                 }
1796                                 memcpy(page_address(sh2->dev[dd_idx].page),
1797                                        page_address(sh->dev[i].page),
1798                                        STRIPE_SIZE);
1799                                 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
1800                                 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
1801                                 for (j=0; j<conf->raid_disks; j++)
1802                                         if (j != sh2->pd_idx &&
1803                                             !test_bit(R5_Expanded, &sh2->dev[j].flags))
1804                                                 break;
1805                                 if (j == conf->raid_disks) {
1806                                         set_bit(STRIPE_EXPAND_READY, &sh2->state);
1807                                         set_bit(STRIPE_HANDLE, &sh2->state);
1808                                 }
1809                                 release_stripe(sh2);
1810                         }
1811         }
1812
1813         spin_unlock(&sh->lock);
1814
1815         while ((bi=return_bi)) {
1816                 int bytes = bi->bi_size;
1817
1818                 return_bi = bi->bi_next;
1819                 bi->bi_next = NULL;
1820                 bi->bi_size = 0;
1821                 bi->bi_end_io(bi, bytes,
1822                               test_bit(BIO_UPTODATE, &bi->bi_flags)
1823                                 ? 0 : -EIO);
1824         }
1825         for (i=disks; i-- ;) {
1826                 int rw;
1827                 struct bio *bi;
1828                 mdk_rdev_t *rdev;
1829                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
1830                         rw = WRITE;
1831                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1832                         rw = READ;
1833                 else
1834                         continue;
1835  
1836                 bi = &sh->dev[i].req;
1837  
1838                 bi->bi_rw = rw;
1839                 if (rw == WRITE)
1840                         bi->bi_end_io = raid5_end_write_request;
1841                 else
1842                         bi->bi_end_io = raid5_end_read_request;
1843  
1844                 rcu_read_lock();
1845                 rdev = rcu_dereference(conf->disks[i].rdev);
1846                 if (rdev && test_bit(Faulty, &rdev->flags))
1847                         rdev = NULL;
1848                 if (rdev)
1849                         atomic_inc(&rdev->nr_pending);
1850                 rcu_read_unlock();
1851  
1852                 if (rdev) {
1853                         if (syncing || expanding || expanded)
1854                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1855
1856                         bi->bi_bdev = rdev->bdev;
1857                         PRINTK("for %llu schedule op %ld on disc %d\n",
1858                                 (unsigned long long)sh->sector, bi->bi_rw, i);
1859                         atomic_inc(&sh->count);
1860                         bi->bi_sector = sh->sector + rdev->data_offset;
1861                         bi->bi_flags = 1 << BIO_UPTODATE;
1862                         bi->bi_vcnt = 1;        
1863                         bi->bi_max_vecs = 1;
1864                         bi->bi_idx = 0;
1865                         bi->bi_io_vec = &sh->dev[i].vec;
1866                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1867                         bi->bi_io_vec[0].bv_offset = 0;
1868                         bi->bi_size = STRIPE_SIZE;
1869                         bi->bi_next = NULL;
1870                         if (rw == WRITE &&
1871                             test_bit(R5_ReWrite, &sh->dev[i].flags))
1872                                 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1873                         generic_make_request(bi);
1874                 } else {
1875                         if (rw == WRITE)
1876                                 set_bit(STRIPE_DEGRADED, &sh->state);
1877                         PRINTK("skip op %ld on disc %d for sector %llu\n",
1878                                 bi->bi_rw, i, (unsigned long long)sh->sector);
1879                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1880                         set_bit(STRIPE_HANDLE, &sh->state);
1881                 }
1882         }
1883 }
1884
1885 static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
1886 {
1887         raid6_conf_t *conf = sh->raid_conf;
1888         int disks = conf->raid_disks;
1889         struct bio *return_bi= NULL;
1890         struct bio *bi;
1891         int i;
1892         int syncing;
1893         int locked=0, uptodate=0, to_read=0, to_write=0, failed=0, written=0;
1894         int non_overwrite = 0;
1895         int failed_num[2] = {0, 0};
1896         struct r5dev *dev, *pdev, *qdev;
1897         int pd_idx = sh->pd_idx;
1898         int qd_idx = raid6_next_disk(pd_idx, disks);
1899         int p_failed, q_failed;
1900
1901         PRINTK("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d, qd_idx=%d\n",
1902                (unsigned long long)sh->sector, sh->state, atomic_read(&sh->count),
1903                pd_idx, qd_idx);
1904
1905         spin_lock(&sh->lock);
1906         clear_bit(STRIPE_HANDLE, &sh->state);
1907         clear_bit(STRIPE_DELAYED, &sh->state);
1908
1909         syncing = test_bit(STRIPE_SYNCING, &sh->state);
1910         /* Now to look around and see what can be done */
1911
1912         rcu_read_lock();
1913         for (i=disks; i--; ) {
1914                 mdk_rdev_t *rdev;
1915                 dev = &sh->dev[i];
1916                 clear_bit(R5_Insync, &dev->flags);
1917
1918                 PRINTK("check %d: state 0x%lx read %p write %p written %p\n",
1919                         i, dev->flags, dev->toread, dev->towrite, dev->written);
1920                 /* maybe we can reply to a read */
1921                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
1922                         struct bio *rbi, *rbi2;
1923                         PRINTK("Return read for disc %d\n", i);
1924                         spin_lock_irq(&conf->device_lock);
1925                         rbi = dev->toread;
1926                         dev->toread = NULL;
1927                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
1928                                 wake_up(&conf->wait_for_overlap);
1929                         spin_unlock_irq(&conf->device_lock);
1930                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1931                                 copy_data(0, rbi, dev->page, dev->sector);
1932                                 rbi2 = r5_next_bio(rbi, dev->sector);
1933                                 spin_lock_irq(&conf->device_lock);
1934                                 if (--rbi->bi_phys_segments == 0) {
1935                                         rbi->bi_next = return_bi;
1936                                         return_bi = rbi;
1937                                 }
1938                                 spin_unlock_irq(&conf->device_lock);
1939                                 rbi = rbi2;
1940                         }
1941                 }
1942
1943                 /* now count some things */
1944                 if (test_bit(R5_LOCKED, &dev->flags)) locked++;
1945                 if (test_bit(R5_UPTODATE, &dev->flags)) uptodate++;
1946
1947
1948                 if (dev->toread) to_read++;
1949                 if (dev->towrite) {
1950                         to_write++;
1951                         if (!test_bit(R5_OVERWRITE, &dev->flags))
1952                                 non_overwrite++;
1953                 }
1954                 if (dev->written) written++;
1955                 rdev = rcu_dereference(conf->disks[i].rdev);
1956                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
1957                         /* The ReadError flag will just be confusing now */
1958                         clear_bit(R5_ReadError, &dev->flags);
1959                         clear_bit(R5_ReWrite, &dev->flags);
1960                 }
1961                 if (!rdev || !test_bit(In_sync, &rdev->flags)
1962                     || test_bit(R5_ReadError, &dev->flags)) {
1963                         if ( failed < 2 )
1964                                 failed_num[failed] = i;
1965                         failed++;
1966                 } else
1967                         set_bit(R5_Insync, &dev->flags);
1968         }
1969         rcu_read_unlock();
1970         PRINTK("locked=%d uptodate=%d to_read=%d"
1971                " to_write=%d failed=%d failed_num=%d,%d\n",
1972                locked, uptodate, to_read, to_write, failed,
1973                failed_num[0], failed_num[1]);
1974         /* check if the array has lost >2 devices and, if so, some requests might
1975          * need to be failed
1976          */
1977         if (failed > 2 && to_read+to_write+written) {
1978                 for (i=disks; i--; ) {
1979                         int bitmap_end = 0;
1980
1981                         if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1982                                 mdk_rdev_t *rdev;
1983                                 rcu_read_lock();
1984                                 rdev = rcu_dereference(conf->disks[i].rdev);
1985                                 if (rdev && test_bit(In_sync, &rdev->flags))
1986                                         /* multiple read failures in one stripe */
1987                                         md_error(conf->mddev, rdev);
1988                                 rcu_read_unlock();
1989                         }
1990
1991                         spin_lock_irq(&conf->device_lock);
1992                         /* fail all writes first */
1993                         bi = sh->dev[i].towrite;
1994                         sh->dev[i].towrite = NULL;
1995                         if (bi) { to_write--; bitmap_end = 1; }
1996
1997                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1998                                 wake_up(&conf->wait_for_overlap);
1999
2000                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
2001                                 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2002                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2003                                 if (--bi->bi_phys_segments == 0) {
2004                                         md_write_end(conf->mddev);
2005                                         bi->bi_next = return_bi;
2006                                         return_bi = bi;
2007                                 }
2008                                 bi = nextbi;
2009                         }
2010                         /* and fail all 'written' */
2011                         bi = sh->dev[i].written;
2012                         sh->dev[i].written = NULL;
2013                         if (bi) bitmap_end = 1;
2014                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS) {
2015                                 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2016                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2017                                 if (--bi->bi_phys_segments == 0) {
2018                                         md_write_end(conf->mddev);
2019                                         bi->bi_next = return_bi;
2020                                         return_bi = bi;
2021                                 }
2022                                 bi = bi2;
2023                         }
2024
2025                         /* fail any reads if this device is non-operational */
2026                         if (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2027                             test_bit(R5_ReadError, &sh->dev[i].flags)) {
2028                                 bi = sh->dev[i].toread;
2029                                 sh->dev[i].toread = NULL;
2030                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2031                                         wake_up(&conf->wait_for_overlap);
2032                                 if (bi) to_read--;
2033                                 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
2034                                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2035                                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2036                                         if (--bi->bi_phys_segments == 0) {
2037                                                 bi->bi_next = return_bi;
2038                                                 return_bi = bi;
2039                                         }
2040                                         bi = nextbi;
2041                                 }
2042                         }
2043                         spin_unlock_irq(&conf->device_lock);
2044                         if (bitmap_end)
2045                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2046                                                 STRIPE_SECTORS, 0, 0);
2047                 }
2048         }
2049         if (failed > 2 && syncing) {
2050                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2051                 clear_bit(STRIPE_SYNCING, &sh->state);
2052                 syncing = 0;
2053         }
2054
2055         /*
2056          * might be able to return some write requests if the parity blocks
2057          * are safe, or on a failed drive
2058          */
2059         pdev = &sh->dev[pd_idx];
2060         p_failed = (failed >= 1 && failed_num[0] == pd_idx)
2061                 || (failed >= 2 && failed_num[1] == pd_idx);
2062         qdev = &sh->dev[qd_idx];
2063         q_failed = (failed >= 1 && failed_num[0] == qd_idx)
2064                 || (failed >= 2 && failed_num[1] == qd_idx);
2065
2066         if ( written &&
2067              ( p_failed || ((test_bit(R5_Insync, &pdev->flags)
2068                              && !test_bit(R5_LOCKED, &pdev->flags)
2069                              && test_bit(R5_UPTODATE, &pdev->flags))) ) &&
2070              ( q_failed || ((test_bit(R5_Insync, &qdev->flags)
2071                              && !test_bit(R5_LOCKED, &qdev->flags)
2072                              && test_bit(R5_UPTODATE, &qdev->flags))) ) ) {
2073                 /* any written block on an uptodate or failed drive can be
2074                  * returned.  Note that if we 'wrote' to a failed drive,
2075                  * it will be UPTODATE, but never LOCKED, so we don't need
2076                  * to test 'failed' directly.
2077                  */
2078                 for (i=disks; i--; )
2079                         if (sh->dev[i].written) {
2080                                 dev = &sh->dev[i];
2081                                 if (!test_bit(R5_LOCKED, &dev->flags) &&
2082                                     test_bit(R5_UPTODATE, &dev->flags) ) {
2083                                         /* We can return any write requests */
2084                                         int bitmap_end = 0;
2085                                         struct bio *wbi, *wbi2;
2086                                         PRINTK("Return write for stripe %llu disc %d\n",
2087                                                (unsigned long long)sh->sector, i);
2088                                         spin_lock_irq(&conf->device_lock);
2089                                         wbi = dev->written;
2090                                         dev->written = NULL;
2091                                         while (wbi && wbi->bi_sector < dev->sector + STRIPE_SECTORS) {
2092                                                 wbi2 = r5_next_bio(wbi, dev->sector);
2093                                                 if (--wbi->bi_phys_segments == 0) {
2094                                                         md_write_end(conf->mddev);
2095                                                         wbi->bi_next = return_bi;
2096                                                         return_bi = wbi;
2097                                                 }
2098                                                 wbi = wbi2;
2099                                         }
2100                                         if (dev->towrite == NULL)
2101                                                 bitmap_end = 1;
2102                                         spin_unlock_irq(&conf->device_lock);
2103                                         if (bitmap_end)
2104                                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2105                                                                 STRIPE_SECTORS,
2106                                                                 !test_bit(STRIPE_DEGRADED, &sh->state), 0);
2107                                 }
2108                         }
2109         }
2110
2111         /* Now we might consider reading some blocks, either to check/generate
2112          * parity, or to satisfy requests
2113          * or to load a block that is being partially written.
2114          */
2115         if (to_read || non_overwrite || (to_write && failed) || (syncing && (uptodate < disks))) {
2116                 for (i=disks; i--;) {
2117                         dev = &sh->dev[i];
2118                         if (!test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
2119                             (dev->toread ||
2120                              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2121                              syncing ||
2122                              (failed >= 1 && (sh->dev[failed_num[0]].toread || to_write)) ||
2123                              (failed >= 2 && (sh->dev[failed_num[1]].toread || to_write))
2124                                     )
2125                                 ) {
2126                                 /* we would like to get this block, possibly
2127                                  * by computing it, but we might not be able to
2128                                  */
2129                                 if (uptodate == disks-1) {
2130                                         PRINTK("Computing stripe %llu block %d\n",
2131                                                (unsigned long long)sh->sector, i);
2132                                         compute_block_1(sh, i, 0);
2133                                         uptodate++;
2134                                 } else if ( uptodate == disks-2 && failed >= 2 ) {
2135                                         /* Computing 2-failure is *very* expensive; only do it if failed >= 2 */
2136                                         int other;
2137                                         for (other=disks; other--;) {
2138                                                 if ( other == i )
2139                                                         continue;
2140                                                 if ( !test_bit(R5_UPTODATE, &sh->dev[other].flags) )
2141                                                         break;
2142                                         }
2143                                         BUG_ON(other < 0);
2144                                         PRINTK("Computing stripe %llu blocks %d,%d\n",
2145                                                (unsigned long long)sh->sector, i, other);
2146                                         compute_block_2(sh, i, other);
2147                                         uptodate += 2;
2148                                 } else if (test_bit(R5_Insync, &dev->flags)) {
2149                                         set_bit(R5_LOCKED, &dev->flags);
2150                                         set_bit(R5_Wantread, &dev->flags);
2151                                         locked++;
2152                                         PRINTK("Reading block %d (sync=%d)\n",
2153                                                 i, syncing);
2154                                 }
2155                         }
2156                 }
2157                 set_bit(STRIPE_HANDLE, &sh->state);
2158         }
2159
2160         /* now to consider writing and what else, if anything should be read */
2161         if (to_write) {
2162                 int rcw=0, must_compute=0;
2163                 for (i=disks ; i--;) {
2164                         dev = &sh->dev[i];
2165                         /* Would I have to read this buffer for reconstruct_write */
2166                         if (!test_bit(R5_OVERWRITE, &dev->flags)
2167                             && i != pd_idx && i != qd_idx
2168                             && (!test_bit(R5_LOCKED, &dev->flags)
2169                                     ) &&
2170                             !test_bit(R5_UPTODATE, &dev->flags)) {
2171                                 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2172                                 else {
2173                                         PRINTK("raid6: must_compute: disk %d flags=%#lx\n", i, dev->flags);
2174                                         must_compute++;
2175                                 }
2176                         }
2177                 }
2178                 PRINTK("for sector %llu, rcw=%d, must_compute=%d\n",
2179                        (unsigned long long)sh->sector, rcw, must_compute);
2180                 set_bit(STRIPE_HANDLE, &sh->state);
2181
2182                 if (rcw > 0)
2183                         /* want reconstruct write, but need to get some data */
2184                         for (i=disks; i--;) {
2185                                 dev = &sh->dev[i];
2186                                 if (!test_bit(R5_OVERWRITE, &dev->flags)
2187                                     && !(failed == 0 && (i == pd_idx || i == qd_idx))
2188                                     && !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
2189                                     test_bit(R5_Insync, &dev->flags)) {
2190                                         if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
2191                                         {
2192                                                 PRINTK("Read_old stripe %llu block %d for Reconstruct\n",
2193                                                        (unsigned long long)sh->sector, i);
2194                                                 set_bit(R5_LOCKED, &dev->flags);
2195                                                 set_bit(R5_Wantread, &dev->flags);
2196                                                 locked++;
2197                                         } else {
2198                                                 PRINTK("Request delayed stripe %llu block %d for Reconstruct\n",
2199                                                        (unsigned long long)sh->sector, i);
2200                                                 set_bit(STRIPE_DELAYED, &sh->state);
2201                                                 set_bit(STRIPE_HANDLE, &sh->state);
2202                                         }
2203                                 }
2204                         }
2205                 /* now if nothing is locked, and if we have enough data, we can start a write request */
2206                 if (locked == 0 && rcw == 0 &&
2207                     !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2208                         if ( must_compute > 0 ) {
2209                                 /* We have failed blocks and need to compute them */
2210                                 switch ( failed ) {
2211                                 case 0: BUG();
2212                                 case 1: compute_block_1(sh, failed_num[0], 0); break;
2213                                 case 2: compute_block_2(sh, failed_num[0], failed_num[1]); break;
2214                                 default: BUG(); /* This request should have been failed? */
2215                                 }
2216                         }
2217
2218                         PRINTK("Computing parity for stripe %llu\n", (unsigned long long)sh->sector);
2219                         compute_parity6(sh, RECONSTRUCT_WRITE);
2220                         /* now every locked buffer is ready to be written */
2221                         for (i=disks; i--;)
2222                                 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2223                                         PRINTK("Writing stripe %llu block %d\n",
2224                                                (unsigned long long)sh->sector, i);
2225                                         locked++;
2226                                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
2227                                 }
2228                         /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2229                         set_bit(STRIPE_INSYNC, &sh->state);
2230
2231                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2232                                 atomic_dec(&conf->preread_active_stripes);
2233                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
2234                                         md_wakeup_thread(conf->mddev->thread);
2235                         }
2236                 }
2237         }
2238
2239         /* maybe we need to check and possibly fix the parity for this stripe
2240          * Any reads will already have been scheduled, so we just see if enough data
2241          * is available
2242          */
2243         if (syncing && locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state)) {
2244                 int update_p = 0, update_q = 0;
2245                 struct r5dev *dev;
2246
2247                 set_bit(STRIPE_HANDLE, &sh->state);
2248
2249                 BUG_ON(failed>2);
2250                 BUG_ON(uptodate < disks);
2251                 /* Want to check and possibly repair P and Q.
2252                  * However there could be one 'failed' device, in which
2253                  * case we can only check one of them, possibly using the
2254                  * other to generate missing data
2255                  */
2256
2257                 /* If !tmp_page, we cannot do the calculations,
2258                  * but as we have set STRIPE_HANDLE, we will soon be called
2259                  * by stripe_handle with a tmp_page - just wait until then.
2260                  */
2261                 if (tmp_page) {
2262                         if (failed == q_failed) {
2263                                 /* The only possible failed device holds 'Q', so it makes
2264                                  * sense to check P (If anything else were failed, we would
2265                                  * have used P to recreate it).
2266                                  */
2267                                 compute_block_1(sh, pd_idx, 1);
2268                                 if (!page_is_zero(sh->dev[pd_idx].page)) {
2269                                         compute_block_1(sh,pd_idx,0);
2270                                         update_p = 1;
2271                                 }
2272                         }
2273                         if (!q_failed && failed < 2) {
2274                                 /* q is not failed, and we didn't use it to generate
2275                                  * anything, so it makes sense to check it
2276                                  */
2277                                 memcpy(page_address(tmp_page),
2278                                        page_address(sh->dev[qd_idx].page),
2279                                        STRIPE_SIZE);
2280                                 compute_parity6(sh, UPDATE_PARITY);
2281                                 if (memcmp(page_address(tmp_page),
2282                                            page_address(sh->dev[qd_idx].page),
2283                                            STRIPE_SIZE)!= 0) {
2284                                         clear_bit(STRIPE_INSYNC, &sh->state);
2285                                         update_q = 1;
2286                                 }
2287                         }
2288                         if (update_p || update_q) {
2289                                 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2290                                 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2291                                         /* don't try to repair!! */
2292                                         update_p = update_q = 0;
2293                         }
2294
2295                         /* now write out any block on a failed drive,
2296                          * or P or Q if they need it
2297                          */
2298
2299                         if (failed == 2) {
2300                                 dev = &sh->dev[failed_num[1]];
2301                                 locked++;
2302                                 set_bit(R5_LOCKED, &dev->flags);
2303                                 set_bit(R5_Wantwrite, &dev->flags);
2304                         }
2305                         if (failed >= 1) {
2306                                 dev = &sh->dev[failed_num[0]];
2307                                 locked++;
2308                                 set_bit(R5_LOCKED, &dev->flags);
2309                                 set_bit(R5_Wantwrite, &dev->flags);
2310                         }
2311
2312                         if (update_p) {
2313                                 dev = &sh->dev[pd_idx];
2314                                 locked ++;
2315                                 set_bit(R5_LOCKED, &dev->flags);
2316                                 set_bit(R5_Wantwrite, &dev->flags);
2317                         }
2318                         if (update_q) {
2319                                 dev = &sh->dev[qd_idx];
2320                                 locked++;
2321                                 set_bit(R5_LOCKED, &dev->flags);
2322                                 set_bit(R5_Wantwrite, &dev->flags);
2323                         }
2324                         clear_bit(STRIPE_DEGRADED, &sh->state);
2325
2326                         set_bit(STRIPE_INSYNC, &sh->state);
2327                 }
2328         }
2329
2330         if (syncing && locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2331                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2332                 clear_bit(STRIPE_SYNCING, &sh->state);
2333         }
2334
2335         /* If the failed drives are just a ReadError, then we might need
2336          * to progress the repair/check process
2337          */
2338         if (failed <= 2 && ! conf->mddev->ro)
2339                 for (i=0; i<failed;i++) {
2340                         dev = &sh->dev[failed_num[i]];
2341                         if (test_bit(R5_ReadError, &dev->flags)
2342                             && !test_bit(R5_LOCKED, &dev->flags)
2343                             && test_bit(R5_UPTODATE, &dev->flags)
2344                                 ) {
2345                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
2346                                         set_bit(R5_Wantwrite, &dev->flags);
2347                                         set_bit(R5_ReWrite, &dev->flags);
2348                                         set_bit(R5_LOCKED, &dev->flags);
2349                                 } else {
2350                                         /* let's read it back */
2351                                         set_bit(R5_Wantread, &dev->flags);
2352                                         set_bit(R5_LOCKED, &dev->flags);
2353                                 }
2354                         }
2355                 }
2356         spin_unlock(&sh->lock);
2357
2358         while ((bi=return_bi)) {
2359                 int bytes = bi->bi_size;
2360
2361                 return_bi = bi->bi_next;
2362                 bi->bi_next = NULL;
2363                 bi->bi_size = 0;
2364                 bi->bi_end_io(bi, bytes,
2365                               test_bit(BIO_UPTODATE, &bi->bi_flags)
2366                                 ? 0 : -EIO);
2367         }
2368         for (i=disks; i-- ;) {
2369                 int rw;
2370                 struct bio *bi;
2371                 mdk_rdev_t *rdev;
2372                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
2373                         rw = WRITE;
2374                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
2375                         rw = READ;
2376                 else
2377                         continue;
2378
2379                 bi = &sh->dev[i].req;
2380
2381                 bi->bi_rw = rw;
2382                 if (rw == WRITE)
2383                         bi->bi_end_io = raid5_end_write_request;
2384                 else
2385                         bi->bi_end_io = raid5_end_read_request;
2386
2387                 rcu_read_lock();
2388                 rdev = rcu_dereference(conf->disks[i].rdev);
2389                 if (rdev && test_bit(Faulty, &rdev->flags))
2390                         rdev = NULL;
2391                 if (rdev)
2392                         atomic_inc(&rdev->nr_pending);
2393                 rcu_read_unlock();
2394
2395                 if (rdev) {
2396                         if (syncing)
2397                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
2398
2399                         bi->bi_bdev = rdev->bdev;
2400                         PRINTK("for %llu schedule op %ld on disc %d\n",
2401                                 (unsigned long long)sh->sector, bi->bi_rw, i);
2402                         atomic_inc(&sh->count);
2403                         bi->bi_sector = sh->sector + rdev->data_offset;
2404                         bi->bi_flags = 1 << BIO_UPTODATE;
2405                         bi->bi_vcnt = 1;
2406                         bi->bi_max_vecs = 1;
2407                         bi->bi_idx = 0;
2408                         bi->bi_io_vec = &sh->dev[i].vec;
2409                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
2410                         bi->bi_io_vec[0].bv_offset = 0;
2411                         bi->bi_size = STRIPE_SIZE;
2412                         bi->bi_next = NULL;
2413                         if (rw == WRITE &&
2414                             test_bit(R5_ReWrite, &sh->dev[i].flags))
2415                                 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2416                         generic_make_request(bi);
2417                 } else {
2418                         if (rw == WRITE)
2419                                 set_bit(STRIPE_DEGRADED, &sh->state);
2420                         PRINTK("skip op %ld on disc %d for sector %llu\n",
2421                                 bi->bi_rw, i, (unsigned long long)sh->sector);
2422                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
2423                         set_bit(STRIPE_HANDLE, &sh->state);
2424                 }
2425         }
2426 }
2427
2428 static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
2429 {
2430         if (sh->raid_conf->level == 6)
2431                 handle_stripe6(sh, tmp_page);
2432         else
2433                 handle_stripe5(sh);
2434 }
2435
2436
2437
2438 static void raid5_activate_delayed(raid5_conf_t *conf)
2439 {
2440         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
2441                 while (!list_empty(&conf->delayed_list)) {
2442                         struct list_head *l = conf->delayed_list.next;
2443                         struct stripe_head *sh;
2444                         sh = list_entry(l, struct stripe_head, lru);
2445                         list_del_init(l);
2446                         clear_bit(STRIPE_DELAYED, &sh->state);
2447                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
2448                                 atomic_inc(&conf->preread_active_stripes);
2449                         list_add_tail(&sh->lru, &conf->handle_list);
2450                 }
2451         }
2452 }
2453
2454 static void activate_bit_delay(raid5_conf_t *conf)
2455 {
2456         /* device_lock is held */
2457         struct list_head head;
2458         list_add(&head, &conf->bitmap_list);
2459         list_del_init(&conf->bitmap_list);
2460         while (!list_empty(&head)) {
2461                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
2462                 list_del_init(&sh->lru);
2463                 atomic_inc(&sh->count);
2464                 __release_stripe(conf, sh);
2465         }
2466 }
2467
2468 static void unplug_slaves(mddev_t *mddev)
2469 {
2470         raid5_conf_t *conf = mddev_to_conf(mddev);
2471         int i;
2472
2473         rcu_read_lock();
2474         for (i=0; i<mddev->raid_disks; i++) {
2475                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
2476                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
2477                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
2478
2479                         atomic_inc(&rdev->nr_pending);
2480                         rcu_read_unlock();
2481
2482                         if (r_queue->unplug_fn)
2483                                 r_queue->unplug_fn(r_queue);
2484
2485                         rdev_dec_pending(rdev, mddev);
2486                         rcu_read_lock();
2487                 }
2488         }
2489         rcu_read_unlock();
2490 }
2491
2492 static void raid5_unplug_device(request_queue_t *q)
2493 {
2494         mddev_t *mddev = q->queuedata;
2495         raid5_conf_t *conf = mddev_to_conf(mddev);
2496         unsigned long flags;
2497
2498         spin_lock_irqsave(&conf->device_lock, flags);
2499
2500         if (blk_remove_plug(q)) {
2501                 conf->seq_flush++;
2502                 raid5_activate_delayed(conf);
2503         }
2504         md_wakeup_thread(mddev->thread);
2505
2506         spin_unlock_irqrestore(&conf->device_lock, flags);
2507
2508         unplug_slaves(mddev);
2509 }
2510
2511 static int raid5_issue_flush(request_queue_t *q, struct gendisk *disk,
2512                              sector_t *error_sector)
2513 {
2514         mddev_t *mddev = q->queuedata;
2515         raid5_conf_t *conf = mddev_to_conf(mddev);
2516         int i, ret = 0;
2517
2518         rcu_read_lock();
2519         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
2520                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
2521                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
2522                         struct block_device *bdev = rdev->bdev;
2523                         request_queue_t *r_queue = bdev_get_queue(bdev);
2524
2525                         if (!r_queue->issue_flush_fn)
2526                                 ret = -EOPNOTSUPP;
2527                         else {
2528                                 atomic_inc(&rdev->nr_pending);
2529                                 rcu_read_unlock();
2530                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
2531                                                               error_sector);
2532                                 rdev_dec_pending(rdev, mddev);
2533                                 rcu_read_lock();
2534                         }
2535                 }
2536         }
2537         rcu_read_unlock();
2538         return ret;
2539 }
2540
2541 static int raid5_congested(void *data, int bits)
2542 {
2543         mddev_t *mddev = data;
2544         raid5_conf_t *conf = mddev_to_conf(mddev);
2545
2546         /* No difference between reads and writes.  Just check
2547          * how busy the stripe_cache is
2548          */
2549         if (conf->inactive_blocked)
2550                 return 1;
2551         if (conf->quiesce)
2552                 return 1;
2553         if (list_empty_careful(&conf->inactive_list))
2554                 return 1;
2555
2556         return 0;
2557 }
2558
2559 /* We want read requests to align with chunks where possible,
2560  * but write requests don't need to.
2561  */
2562 static int raid5_mergeable_bvec(request_queue_t *q, struct bio *bio, struct bio_vec *biovec)
2563 {
2564         mddev_t *mddev = q->queuedata;
2565         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
2566         int max;
2567         unsigned int chunk_sectors = mddev->chunk_size >> 9;
2568         unsigned int bio_sectors = bio->bi_size >> 9;
2569
2570         if (bio_data_dir(bio) == WRITE)
2571                 return biovec->bv_len; /* always allow writes to be mergeable */
2572
2573         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
2574         if (max < 0) max = 0;
2575         if (max <= biovec->bv_len && bio_sectors == 0)
2576                 return biovec->bv_len;
2577         else
2578                 return max;
2579 }
2580
2581
2582 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
2583 {
2584         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
2585         unsigned int chunk_sectors = mddev->chunk_size >> 9;
2586         unsigned int bio_sectors = bio->bi_size >> 9;
2587
2588         return  chunk_sectors >=
2589                 ((sector & (chunk_sectors - 1)) + bio_sectors);
2590 }
2591
2592 /*
2593  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
2594  *  later sampled by raid5d.
2595  */
2596 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
2597 {
2598         unsigned long flags;
2599
2600         spin_lock_irqsave(&conf->device_lock, flags);
2601
2602         bi->bi_next = conf->retry_read_aligned_list;
2603         conf->retry_read_aligned_list = bi;
2604
2605         spin_unlock_irqrestore(&conf->device_lock, flags);
2606         md_wakeup_thread(conf->mddev->thread);
2607 }
2608
2609
2610 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
2611 {
2612         struct bio *bi;
2613
2614         bi = conf->retry_read_aligned;
2615         if (bi) {
2616                 conf->retry_read_aligned = NULL;
2617                 return bi;
2618         }
2619         bi = conf->retry_read_aligned_list;
2620         if(bi) {
2621                 conf->retry_read_aligned = bi->bi_next;
2622                 bi->bi_next = NULL;
2623                 bi->bi_phys_segments = 1; /* biased count of active stripes */
2624                 bi->bi_hw_segments = 0; /* count of processed stripes */
2625         }
2626
2627         return bi;
2628 }
2629
2630
2631 /*
2632  *  The "raid5_align_endio" should check if the read succeeded and if it
2633  *  did, call bio_endio on the original bio (having bio_put the new bio
2634  *  first).
2635  *  If the read failed..
2636  */
2637 static int raid5_align_endio(struct bio *bi, unsigned int bytes, int error)
2638 {
2639         struct bio* raid_bi  = bi->bi_private;
2640         mddev_t *mddev;
2641         raid5_conf_t *conf;
2642         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
2643         mdk_rdev_t *rdev;
2644
2645         if (bi->bi_size)
2646                 return 1;
2647         bio_put(bi);
2648
2649         mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
2650         conf = mddev_to_conf(mddev);
2651         rdev = (void*)raid_bi->bi_next;
2652         raid_bi->bi_next = NULL;
2653
2654         rdev_dec_pending(rdev, conf->mddev);
2655
2656         if (!error && uptodate) {
2657                 bio_endio(raid_bi, bytes, 0);
2658                 if (atomic_dec_and_test(&conf->active_aligned_reads))
2659                         wake_up(&conf->wait_for_stripe);
2660                 return 0;
2661         }
2662
2663
2664         PRINTK("raid5_align_endio : io error...handing IO for a retry\n");
2665
2666         add_bio_to_retry(raid_bi, conf);
2667         return 0;
2668 }
2669
2670 static int chunk_aligned_read(request_queue_t *q, struct bio * raid_bio)
2671 {
2672         mddev_t *mddev = q->queuedata;
2673         raid5_conf_t *conf = mddev_to_conf(mddev);
2674         const unsigned int raid_disks = conf->raid_disks;
2675         const unsigned int data_disks = raid_disks - conf->max_degraded;
2676         unsigned int dd_idx, pd_idx;
2677         struct bio* align_bi;
2678         mdk_rdev_t *rdev;
2679
2680         if (!in_chunk_boundary(mddev, raid_bio)) {
2681                 printk("chunk_aligned_read : non aligned\n");
2682                 return 0;
2683         }
2684         /*
2685          * use bio_clone to make a copy of the bio
2686          */
2687         align_bi = bio_clone(raid_bio, GFP_NOIO);
2688         if (!align_bi)
2689                 return 0;
2690         /*
2691          *   set bi_end_io to a new function, and set bi_private to the
2692          *     original bio.
2693          */
2694         align_bi->bi_end_io  = raid5_align_endio;
2695         align_bi->bi_private = raid_bio;
2696         /*
2697          *      compute position
2698          */
2699         align_bi->bi_sector =  raid5_compute_sector(raid_bio->bi_sector,
2700                                         raid_disks,
2701                                         data_disks,
2702                                         &dd_idx,
2703                                         &pd_idx,
2704                                         conf);
2705
2706         rcu_read_lock();
2707         rdev = rcu_dereference(conf->disks[dd_idx].rdev);
2708         if (rdev && test_bit(In_sync, &rdev->flags)) {
2709                 atomic_inc(&rdev->nr_pending);
2710                 rcu_read_unlock();
2711                 raid_bio->bi_next = (void*)rdev;
2712                 align_bi->bi_bdev =  rdev->bdev;
2713                 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
2714                 align_bi->bi_sector += rdev->data_offset;
2715
2716                 spin_lock_irq(&conf->device_lock);
2717                 wait_event_lock_irq(conf->wait_for_stripe,
2718                                     conf->quiesce == 0,
2719                                     conf->device_lock, /* nothing */);
2720                 atomic_inc(&conf->active_aligned_reads);
2721                 spin_unlock_irq(&conf->device_lock);
2722
2723                 generic_make_request(align_bi);
2724                 return 1;
2725         } else {
2726                 rcu_read_unlock();
2727                 bio_put(align_bi);
2728                 return 0;
2729         }
2730 }
2731
2732
2733 static int make_request(request_queue_t *q, struct bio * bi)
2734 {
2735         mddev_t *mddev = q->queuedata;
2736         raid5_conf_t *conf = mddev_to_conf(mddev);
2737         unsigned int dd_idx, pd_idx;
2738         sector_t new_sector;
2739         sector_t logical_sector, last_sector;
2740         struct stripe_head *sh;
2741         const int rw = bio_data_dir(bi);
2742         int remaining;
2743
2744         if (unlikely(bio_barrier(bi))) {
2745                 bio_endio(bi, bi->bi_size, -EOPNOTSUPP);
2746                 return 0;
2747         }
2748
2749         md_write_start(mddev, bi);
2750
2751         disk_stat_inc(mddev->gendisk, ios[rw]);
2752         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
2753
2754         if (rw == READ &&
2755              mddev->reshape_position == MaxSector &&
2756              chunk_aligned_read(q,bi))
2757                 return 0;
2758
2759         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
2760         last_sector = bi->bi_sector + (bi->bi_size>>9);
2761         bi->bi_next = NULL;
2762         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
2763
2764         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
2765                 DEFINE_WAIT(w);
2766                 int disks, data_disks;
2767
2768         retry:
2769                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
2770                 if (likely(conf->expand_progress == MaxSector))
2771                         disks = conf->raid_disks;
2772                 else {
2773                         /* spinlock is needed as expand_progress may be
2774                          * 64bit on a 32bit platform, and so it might be
2775                          * possible to see a half-updated value
2776                          * Ofcourse expand_progress could change after
2777                          * the lock is dropped, so once we get a reference
2778                          * to the stripe that we think it is, we will have
2779                          * to check again.
2780                          */
2781                         spin_lock_irq(&conf->device_lock);
2782                         disks = conf->raid_disks;
2783                         if (logical_sector >= conf->expand_progress)
2784                                 disks = conf->previous_raid_disks;
2785                         else {
2786                                 if (logical_sector >= conf->expand_lo) {
2787                                         spin_unlock_irq(&conf->device_lock);
2788                                         schedule();
2789                                         goto retry;
2790                                 }
2791                         }
2792                         spin_unlock_irq(&conf->device_lock);
2793                 }
2794                 data_disks = disks - conf->max_degraded;
2795
2796                 new_sector = raid5_compute_sector(logical_sector, disks, data_disks,
2797                                                   &dd_idx, &pd_idx, conf);
2798                 PRINTK("raid5: make_request, sector %llu logical %llu\n",
2799                         (unsigned long long)new_sector, 
2800                         (unsigned long long)logical_sector);
2801
2802                 sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
2803                 if (sh) {
2804                         if (unlikely(conf->expand_progress != MaxSector)) {
2805                                 /* expansion might have moved on while waiting for a
2806                                  * stripe, so we must do the range check again.
2807                                  * Expansion could still move past after this
2808                                  * test, but as we are holding a reference to
2809                                  * 'sh', we know that if that happens,
2810                                  *  STRIPE_EXPANDING will get set and the expansion
2811                                  * won't proceed until we finish with the stripe.
2812                                  */
2813                                 int must_retry = 0;
2814                                 spin_lock_irq(&conf->device_lock);
2815                                 if (logical_sector <  conf->expand_progress &&
2816                                     disks == conf->previous_raid_disks)
2817                                         /* mismatch, need to try again */
2818                                         must_retry = 1;
2819                                 spin_unlock_irq(&conf->device_lock);
2820                                 if (must_retry) {
2821                                         release_stripe(sh);
2822                                         goto retry;
2823                                 }
2824                         }
2825                         /* FIXME what if we get a false positive because these
2826                          * are being updated.
2827                          */
2828                         if (logical_sector >= mddev->suspend_lo &&
2829                             logical_sector < mddev->suspend_hi) {
2830                                 release_stripe(sh);
2831                                 schedule();
2832                                 goto retry;
2833                         }
2834
2835                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
2836                             !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
2837                                 /* Stripe is busy expanding or
2838                                  * add failed due to overlap.  Flush everything
2839                                  * and wait a while
2840                                  */
2841                                 raid5_unplug_device(mddev->queue);
2842                                 release_stripe(sh);
2843                                 schedule();
2844                                 goto retry;
2845                         }
2846                         finish_wait(&conf->wait_for_overlap, &w);
2847                         handle_stripe(sh, NULL);
2848                         release_stripe(sh);
2849                 } else {
2850                         /* cannot get stripe for read-ahead, just give-up */
2851                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2852                         finish_wait(&conf->wait_for_overlap, &w);
2853                         break;
2854                 }
2855                         
2856         }
2857         spin_lock_irq(&conf->device_lock);
2858         remaining = --bi->bi_phys_segments;
2859         spin_unlock_irq(&conf->device_lock);
2860         if (remaining == 0) {
2861                 int bytes = bi->bi_size;
2862
2863                 if ( rw == WRITE )
2864                         md_write_end(mddev);
2865                 bi->bi_size = 0;
2866                 bi->bi_end_io(bi, bytes,
2867                               test_bit(BIO_UPTODATE, &bi->bi_flags)
2868                                 ? 0 : -EIO);
2869         }
2870         return 0;
2871 }
2872
2873 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
2874 {
2875         /* reshaping is quite different to recovery/resync so it is
2876          * handled quite separately ... here.
2877          *
2878          * On each call to sync_request, we gather one chunk worth of
2879          * destination stripes and flag them as expanding.
2880          * Then we find all the source stripes and request reads.
2881          * As the reads complete, handle_stripe will copy the data
2882          * into the destination stripe and release that stripe.
2883          */
2884         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
2885         struct stripe_head *sh;
2886         int pd_idx;
2887         sector_t first_sector, last_sector;
2888         int raid_disks;
2889         int data_disks;
2890         int i;
2891         int dd_idx;
2892         sector_t writepos, safepos, gap;
2893
2894         if (sector_nr == 0 &&
2895             conf->expand_progress != 0) {
2896                 /* restarting in the middle, skip the initial sectors */
2897                 sector_nr = conf->expand_progress;
2898                 sector_div(sector_nr, conf->raid_disks-1);
2899                 *skipped = 1;
2900                 return sector_nr;
2901         }
2902
2903         /* we update the metadata when there is more than 3Meg
2904          * in the block range (that is rather arbitrary, should
2905          * probably be time based) or when the data about to be
2906          * copied would over-write the source of the data at
2907          * the front of the range.
2908          * i.e. one new_stripe forward from expand_progress new_maps
2909          * to after where expand_lo old_maps to
2910          */
2911         writepos = conf->expand_progress +
2912                 conf->chunk_size/512*(conf->raid_disks-1);
2913         sector_div(writepos, conf->raid_disks-1);
2914         safepos = conf->expand_lo;
2915         sector_div(safepos, conf->previous_raid_disks-1);
2916         gap = conf->expand_progress - conf->expand_lo;
2917
2918         if (writepos >= safepos ||
2919             gap > (conf->raid_disks-1)*3000*2 /*3Meg*/) {
2920                 /* Cannot proceed until we've updated the superblock... */
2921                 wait_event(conf->wait_for_overlap,
2922                            atomic_read(&conf->reshape_stripes)==0);
2923                 mddev->reshape_position = conf->expand_progress;
2924                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2925                 md_wakeup_thread(mddev->thread);
2926                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
2927                            kthread_should_stop());
2928                 spin_lock_irq(&conf->device_lock);
2929                 conf->expand_lo = mddev->reshape_position;
2930                 spin_unlock_irq(&conf->device_lock);
2931                 wake_up(&conf->wait_for_overlap);
2932         }
2933
2934         for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
2935                 int j;
2936                 int skipped = 0;
2937                 pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks);
2938                 sh = get_active_stripe(conf, sector_nr+i,
2939                                        conf->raid_disks, pd_idx, 0);
2940                 set_bit(STRIPE_EXPANDING, &sh->state);
2941                 atomic_inc(&conf->reshape_stripes);
2942                 /* If any of this stripe is beyond the end of the old
2943                  * array, then we need to zero those blocks
2944                  */
2945                 for (j=sh->disks; j--;) {
2946                         sector_t s;
2947                         if (j == sh->pd_idx)
2948                                 continue;
2949                         s = compute_blocknr(sh, j);
2950                         if (s < (mddev->array_size<<1)) {
2951                                 skipped = 1;
2952                                 continue;
2953                         }
2954                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
2955                         set_bit(R5_Expanded, &sh->dev[j].flags);
2956                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
2957                 }
2958                 if (!skipped) {
2959                         set_bit(STRIPE_EXPAND_READY, &sh->state);
2960                         set_bit(STRIPE_HANDLE, &sh->state);
2961                 }
2962                 release_stripe(sh);
2963         }
2964         spin_lock_irq(&conf->device_lock);
2965         conf->expand_progress = (sector_nr + i)*(conf->raid_disks-1);
2966         spin_unlock_irq(&conf->device_lock);
2967         /* Ok, those stripe are ready. We can start scheduling
2968          * reads on the source stripes.
2969          * The source stripes are determined by mapping the first and last
2970          * block on the destination stripes.
2971          */
2972         raid_disks = conf->previous_raid_disks;
2973         data_disks = raid_disks - 1;
2974         first_sector =
2975                 raid5_compute_sector(sector_nr*(conf->raid_disks-1),
2976                                      raid_disks, data_disks,
2977                                      &dd_idx, &pd_idx, conf);
2978         last_sector =
2979                 raid5_compute_sector((sector_nr+conf->chunk_size/512)
2980                                      *(conf->raid_disks-1) -1,
2981                                      raid_disks, data_disks,
2982                                      &dd_idx, &pd_idx, conf);
2983         if (last_sector >= (mddev->size<<1))
2984                 last_sector = (mddev->size<<1)-1;
2985         while (first_sector <= last_sector) {
2986                 pd_idx = stripe_to_pdidx(first_sector, conf, conf->previous_raid_disks);
2987                 sh = get_active_stripe(conf, first_sector,
2988                                        conf->previous_raid_disks, pd_idx, 0);
2989                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2990                 set_bit(STRIPE_HANDLE, &sh->state);
2991                 release_stripe(sh);
2992                 first_sector += STRIPE_SECTORS;
2993         }
2994         return conf->chunk_size>>9;
2995 }
2996
2997 /* FIXME go_faster isn't used */
2998 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
2999 {
3000         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3001         struct stripe_head *sh;
3002         int pd_idx;
3003         int raid_disks = conf->raid_disks;
3004         sector_t max_sector = mddev->size << 1;
3005         int sync_blocks;
3006         int still_degraded = 0;
3007         int i;
3008
3009         if (sector_nr >= max_sector) {
3010                 /* just being told to finish up .. nothing much to do */
3011                 unplug_slaves(mddev);
3012                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3013                         end_reshape(conf);
3014                         return 0;
3015                 }
3016
3017                 if (mddev->curr_resync < max_sector) /* aborted */
3018                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3019                                         &sync_blocks, 1);
3020                 else /* completed sync */
3021                         conf->fullsync = 0;
3022                 bitmap_close_sync(mddev->bitmap);
3023
3024                 return 0;
3025         }
3026
3027         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3028                 return reshape_request(mddev, sector_nr, skipped);
3029
3030         /* if there is too many failed drives and we are trying
3031          * to resync, then assert that we are finished, because there is
3032          * nothing we can do.
3033          */
3034         if (mddev->degraded >= conf->max_degraded &&
3035             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3036                 sector_t rv = (mddev->size << 1) - sector_nr;
3037                 *skipped = 1;
3038                 return rv;
3039         }
3040         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3041             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3042             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3043                 /* we can skip this block, and probably more */
3044                 sync_blocks /= STRIPE_SECTORS;
3045                 *skipped = 1;
3046                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3047         }
3048
3049         pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks);
3050         sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1);
3051         if (sh == NULL) {
3052                 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0);
3053                 /* make sure we don't swamp the stripe cache if someone else
3054                  * is trying to get access
3055                  */
3056                 schedule_timeout_uninterruptible(1);
3057         }
3058         /* Need to check if array will still be degraded after recovery/resync
3059          * We don't need to check the 'failed' flag as when that gets set,
3060          * recovery aborts.
3061          */
3062         for (i=0; i<mddev->raid_disks; i++)
3063                 if (conf->disks[i].rdev == NULL)
3064                         still_degraded = 1;
3065
3066         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
3067
3068         spin_lock(&sh->lock);
3069         set_bit(STRIPE_SYNCING, &sh->state);
3070         clear_bit(STRIPE_INSYNC, &sh->state);
3071         spin_unlock(&sh->lock);
3072
3073         handle_stripe(sh, NULL);
3074         release_stripe(sh);
3075
3076         return STRIPE_SECTORS;
3077 }
3078
3079 static int  retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
3080 {
3081         /* We may not be able to submit a whole bio at once as there
3082          * may not be enough stripe_heads available.
3083          * We cannot pre-allocate enough stripe_heads as we may need
3084          * more than exist in the cache (if we allow ever large chunks).
3085          * So we do one stripe head at a time and record in
3086          * ->bi_hw_segments how many have been done.
3087          *
3088          * We *know* that this entire raid_bio is in one chunk, so
3089          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3090          */
3091         struct stripe_head *sh;
3092         int dd_idx, pd_idx;
3093         sector_t sector, logical_sector, last_sector;
3094         int scnt = 0;
3095         int remaining;
3096         int handled = 0;
3097
3098         logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3099         sector = raid5_compute_sector(  logical_sector,
3100                                         conf->raid_disks,
3101                                         conf->raid_disks - conf->max_degraded,
3102                                         &dd_idx,
3103                                         &pd_idx,
3104                                         conf);
3105         last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
3106
3107         for (; logical_sector < last_sector;
3108              logical_sector += STRIPE_SECTORS, scnt++) {
3109
3110                 if (scnt < raid_bio->bi_hw_segments)
3111                         /* already done this stripe */
3112                         continue;
3113
3114                 sh = get_active_stripe(conf, sector, conf->raid_disks, pd_idx, 1);
3115
3116                 if (!sh) {
3117                         /* failed to get a stripe - must wait */
3118                         raid_bio->bi_hw_segments = scnt;
3119                         conf->retry_read_aligned = raid_bio;
3120                         return handled;
3121                 }
3122
3123                 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
3124                 add_stripe_bio(sh, raid_bio, dd_idx, 0);
3125                 handle_stripe(sh, NULL);
3126                 release_stripe(sh);
3127                 handled++;
3128         }
3129         spin_lock_irq(&conf->device_lock);
3130         remaining = --raid_bio->bi_phys_segments;
3131         spin_unlock_irq(&conf->device_lock);
3132         if (remaining == 0) {
3133                 int bytes = raid_bio->bi_size;
3134
3135                 raid_bio->bi_size = 0;
3136                 raid_bio->bi_end_io(raid_bio, bytes,
3137                               test_bit(BIO_UPTODATE, &raid_bio->bi_flags)
3138                                 ? 0 : -EIO);
3139         }
3140         if (atomic_dec_and_test(&conf->active_aligned_reads))
3141                 wake_up(&conf->wait_for_stripe);
3142         return handled;
3143 }
3144
3145
3146
3147 /*
3148  * This is our raid5 kernel thread.
3149  *
3150  * We scan the hash table for stripes which can be handled now.
3151  * During the scan, completed stripes are saved for us by the interrupt
3152  * handler, so that they will not have to wait for our next wakeup.
3153  */
3154 static void raid5d (mddev_t *mddev)
3155 {
3156         struct stripe_head *sh;
3157         raid5_conf_t *conf = mddev_to_conf(mddev);
3158         int handled;
3159
3160         PRINTK("+++ raid5d active\n");
3161
3162         md_check_recovery(mddev);
3163
3164         handled = 0;
3165         spin_lock_irq(&conf->device_lock);
3166         while (1) {
3167                 struct list_head *first;
3168                 struct bio *bio;
3169
3170                 if (conf->seq_flush != conf->seq_write) {
3171                         int seq = conf->seq_flush;
3172                         spin_unlock_irq(&conf->device_lock);
3173                         bitmap_unplug(mddev->bitmap);
3174                         spin_lock_irq(&conf->device_lock);
3175                         conf->seq_write = seq;
3176                         activate_bit_delay(conf);
3177                 }
3178
3179                 if (list_empty(&conf->handle_list) &&
3180                     atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD &&
3181                     !blk_queue_plugged(mddev->queue) &&
3182                     !list_empty(&conf->delayed_list))
3183                         raid5_activate_delayed(conf);
3184
3185                 while ((bio = remove_bio_from_retry(conf))) {
3186                         int ok;
3187                         spin_unlock_irq(&conf->device_lock);
3188                         ok = retry_aligned_read(conf, bio);
3189                         spin_lock_irq(&conf->device_lock);
3190                         if (!ok)
3191                                 break;
3192                         handled++;
3193                 }
3194
3195                 if (list_empty(&conf->handle_list))
3196                         break;
3197
3198                 first = conf->handle_list.next;
3199                 sh = list_entry(first, struct stripe_head, lru);
3200
3201                 list_del_init(first);
3202                 atomic_inc(&sh->count);
3203                 BUG_ON(atomic_read(&sh->count)!= 1);
3204                 spin_unlock_irq(&conf->device_lock);
3205                 
3206                 handled++;
3207                 handle_stripe(sh, conf->spare_page);
3208                 release_stripe(sh);
3209
3210                 spin_lock_irq(&conf->device_lock);
3211         }
3212         PRINTK("%d stripes handled\n", handled);
3213
3214         spin_unlock_irq(&conf->device_lock);
3215
3216         unplug_slaves(mddev);
3217
3218         PRINTK("--- raid5d inactive\n");
3219 }
3220
3221 static ssize_t
3222 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
3223 {
3224         raid5_conf_t *conf = mddev_to_conf(mddev);
3225         if (conf)
3226                 return sprintf(page, "%d\n", conf->max_nr_stripes);
3227         else
3228                 return 0;
3229 }
3230
3231 static ssize_t
3232 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
3233 {
3234         raid5_conf_t *conf = mddev_to_conf(mddev);
3235         char *end;
3236         int new;
3237         if (len >= PAGE_SIZE)
3238                 return -EINVAL;
3239         if (!conf)
3240                 return -ENODEV;
3241
3242         new = simple_strtoul(page, &end, 10);
3243         if (!*page || (*end && *end != '\n') )
3244                 return -EINVAL;
3245         if (new <= 16 || new > 32768)
3246                 return -EINVAL;
3247         while (new < conf->max_nr_stripes) {
3248                 if (drop_one_stripe(conf))
3249                         conf->max_nr_stripes--;
3250                 else
3251                         break;
3252         }
3253         while (new > conf->max_nr_stripes) {
3254                 if (grow_one_stripe(conf))
3255                         conf->max_nr_stripes++;
3256                 else break;
3257         }
3258         return len;
3259 }
3260
3261 static struct md_sysfs_entry
3262 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
3263                                 raid5_show_stripe_cache_size,
3264                                 raid5_store_stripe_cache_size);
3265
3266 static ssize_t
3267 stripe_cache_active_show(mddev_t *mddev, char *page)
3268 {
3269         raid5_conf_t *conf = mddev_to_conf(mddev);
3270         if (conf)
3271                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
3272         else
3273                 return 0;
3274 }
3275
3276 static struct md_sysfs_entry
3277 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
3278
3279 static struct attribute *raid5_attrs[] =  {
3280         &raid5_stripecache_size.attr,
3281         &raid5_stripecache_active.attr,
3282         NULL,
3283 };
3284 static struct attribute_group raid5_attrs_group = {
3285         .name = NULL,
3286         .attrs = raid5_attrs,
3287 };
3288
3289 static int run(mddev_t *mddev)
3290 {
3291         raid5_conf_t *conf;
3292         int raid_disk, memory;
3293         mdk_rdev_t *rdev;
3294         struct disk_info *disk;
3295         struct list_head *tmp;
3296         int working_disks = 0;
3297
3298         if (mddev->level != 5 && mddev->level != 4 && mddev->level != 6) {
3299                 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
3300                        mdname(mddev), mddev->level);
3301                 return -EIO;
3302         }
3303
3304         if (mddev->reshape_position != MaxSector) {
3305                 /* Check that we can continue the reshape.
3306                  * Currently only disks can change, it must
3307                  * increase, and we must be past the point where
3308                  * a stripe over-writes itself
3309                  */
3310                 sector_t here_new, here_old;
3311                 int old_disks;
3312
3313                 if (mddev->new_level != mddev->level ||
3314                     mddev->new_layout != mddev->layout ||
3315                     mddev->new_chunk != mddev->chunk_size) {
3316                         printk(KERN_ERR "raid5: %s: unsupported reshape required - aborting.\n",
3317                                mdname(mddev));
3318                         return -EINVAL;
3319                 }
3320                 if (mddev->delta_disks <= 0) {
3321                         printk(KERN_ERR "raid5: %s: unsupported reshape (reduce disks) required - aborting.\n",
3322                                mdname(mddev));
3323                         return -EINVAL;
3324                 }
3325                 old_disks = mddev->raid_disks - mddev->delta_disks;
3326                 /* reshape_position must be on a new-stripe boundary, and one
3327                  * further up in new geometry must map after here in old geometry.
3328                  */
3329                 here_new = mddev->reshape_position;
3330                 if (sector_div(here_new, (mddev->chunk_size>>9)*(mddev->raid_disks-1))) {
3331                         printk(KERN_ERR "raid5: reshape_position not on a stripe boundary\n");
3332                         return -EINVAL;
3333                 }
3334                 /* here_new is the stripe we will write to */
3335                 here_old = mddev->reshape_position;
3336                 sector_div(here_old, (mddev->chunk_size>>9)*(old_disks-1));
3337                 /* here_old is the first stripe that we might need to read from */
3338                 if (here_new >= here_old) {
3339                         /* Reading from the same stripe as writing to - bad */
3340                         printk(KERN_ERR "raid5: reshape_position too early for auto-recovery - aborting.\n");
3341                         return -EINVAL;
3342                 }
3343                 printk(KERN_INFO "raid5: reshape will continue\n");
3344                 /* OK, we should be able to continue; */
3345         }
3346
3347
3348         mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
3349         if ((conf = mddev->private) == NULL)
3350                 goto abort;
3351         if (mddev->reshape_position == MaxSector) {
3352                 conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks;
3353         } else {
3354                 conf->raid_disks = mddev->raid_disks;
3355                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
3356         }
3357
3358         conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
3359                               GFP_KERNEL);
3360         if (!conf->disks)
3361                 goto abort;
3362
3363         conf->mddev = mddev;
3364
3365         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
3366                 goto abort;
3367
3368         if (mddev->level == 6) {
3369                 conf->spare_page = alloc_page(GFP_KERNEL);
3370                 if (!conf->spare_page)
3371                         goto abort;
3372         }
3373         spin_lock_init(&conf->device_lock);
3374         init_waitqueue_head(&conf->wait_for_stripe);
3375         init_waitqueue_head(&conf->wait_for_overlap);
3376         INIT_LIST_HEAD(&conf->handle_list);
3377         INIT_LIST_HEAD(&conf->delayed_list);
3378         INIT_LIST_HEAD(&conf->bitmap_list);
3379         INIT_LIST_HEAD(&conf->inactive_list);
3380         atomic_set(&conf->active_stripes, 0);
3381         atomic_set(&conf->preread_active_stripes, 0);
3382         atomic_set(&conf->active_aligned_reads, 0);
3383
3384         PRINTK("raid5: run(%s) called.\n", mdname(mddev));
3385
3386         ITERATE_RDEV(mddev,rdev,tmp) {
3387                 raid_disk = rdev->raid_disk;
3388                 if (raid_disk >= conf->raid_disks
3389                     || raid_disk < 0)
3390                         continue;
3391                 disk = conf->disks + raid_disk;
3392
3393                 disk->rdev = rdev;
3394
3395                 if (test_bit(In_sync, &rdev->flags)) {
3396                         char b[BDEVNAME_SIZE];
3397                         printk(KERN_INFO "raid5: device %s operational as raid"
3398                                 " disk %d\n", bdevname(rdev->bdev,b),
3399                                 raid_disk);
3400                         working_disks++;
3401                 }
3402         }
3403
3404         /*
3405          * 0 for a fully functional array, 1 or 2 for a degraded array.
3406          */
3407         mddev->degraded = conf->raid_disks - working_disks;
3408         conf->mddev = mddev;
3409         conf->chunk_size = mddev->chunk_size;
3410         conf->level = mddev->level;
3411         if (conf->level == 6)
3412                 conf->max_degraded = 2;
3413         else
3414                 conf->max_degraded = 1;
3415         conf->algorithm = mddev->layout;
3416         conf->max_nr_stripes = NR_STRIPES;
3417         conf->expand_progress = mddev->reshape_position;
3418
3419         /* device size must be a multiple of chunk size */
3420         mddev->size &= ~(mddev->chunk_size/1024 -1);
3421         mddev->resync_max_sectors = mddev->size << 1;
3422
3423         if (conf->level == 6 && conf->raid_disks < 4) {
3424                 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
3425                        mdname(mddev), conf->raid_disks);
3426                 goto abort;
3427         }
3428         if (!conf->chunk_size || conf->chunk_size % 4) {
3429                 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
3430                         conf->chunk_size, mdname(mddev));
3431                 goto abort;
3432         }
3433         if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
3434                 printk(KERN_ERR 
3435                         "raid5: unsupported parity algorithm %d for %s\n",
3436                         conf->algorithm, mdname(mddev));
3437                 goto abort;
3438         }
3439         if (mddev->degraded > conf->max_degraded) {
3440                 printk(KERN_ERR "raid5: not enough operational devices for %s"
3441                         " (%d/%d failed)\n",
3442                         mdname(mddev), mddev->degraded, conf->raid_disks);
3443                 goto abort;
3444         }
3445
3446         if (mddev->degraded > 0 &&
3447             mddev->recovery_cp != MaxSector) {
3448                 if (mddev->ok_start_degraded)
3449                         printk(KERN_WARNING
3450                                "raid5: starting dirty degraded array: %s"
3451                                "- data corruption possible.\n",
3452                                mdname(mddev));
3453                 else {
3454                         printk(KERN_ERR
3455                                "raid5: cannot start dirty degraded array for %s\n",
3456                                mdname(mddev));
3457                         goto abort;
3458                 }
3459         }
3460
3461         {
3462                 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
3463                 if (!mddev->thread) {
3464                         printk(KERN_ERR 
3465                                 "raid5: couldn't allocate thread for %s\n",
3466                                 mdname(mddev));
3467                         goto abort;
3468                 }
3469         }
3470         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
3471                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
3472         if (grow_stripes(conf, conf->max_nr_stripes)) {
3473                 printk(KERN_ERR 
3474                         "raid5: couldn't allocate %dkB for buffers\n", memory);
3475                 shrink_stripes(conf);
3476                 md_unregister_thread(mddev->thread);
3477                 goto abort;
3478         } else
3479                 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
3480                         memory, mdname(mddev));
3481
3482         if (mddev->degraded == 0)
3483                 printk("raid5: raid level %d set %s active with %d out of %d"
3484                         " devices, algorithm %d\n", conf->level, mdname(mddev), 
3485                         mddev->raid_disks-mddev->degraded, mddev->raid_disks,
3486                         conf->algorithm);
3487         else
3488                 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
3489                         " out of %d devices, algorithm %d\n", conf->level,
3490                         mdname(mddev), mddev->raid_disks - mddev->degraded,
3491                         mddev->raid_disks, conf->algorithm);
3492
3493         print_raid5_conf(conf);
3494
3495         if (conf->expand_progress != MaxSector) {
3496                 printk("...ok start reshape thread\n");
3497                 conf->expand_lo = conf->expand_progress;
3498                 atomic_set(&conf->reshape_stripes, 0);
3499                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3500                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3501                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3502                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3503                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3504                                                         "%s_reshape");
3505         }
3506
3507         /* read-ahead size must cover two whole stripes, which is
3508          * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
3509          */
3510         {
3511                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
3512                 int stripe = data_disks *
3513                         (mddev->chunk_size / PAGE_SIZE);
3514                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3515                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3516         }
3517
3518         /* Ok, everything is just fine now */
3519         sysfs_create_group(&mddev->kobj, &raid5_attrs_group);
3520
3521         mddev->queue->unplug_fn = raid5_unplug_device;
3522         mddev->queue->issue_flush_fn = raid5_issue_flush;
3523         mddev->queue->backing_dev_info.congested_fn = raid5_congested;
3524         mddev->queue->backing_dev_info.congested_data = mddev;
3525
3526         mddev->array_size =  mddev->size * (conf->previous_raid_disks -
3527                                             conf->max_degraded);
3528
3529         blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
3530
3531         return 0;
3532 abort:
3533         if (conf) {
3534                 print_raid5_conf(conf);
3535                 safe_put_page(conf->spare_page);
3536                 kfree(conf->disks);
3537                 kfree(conf->stripe_hashtbl);
3538                 kfree(conf);
3539         }
3540         mddev->private = NULL;
3541         printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
3542         return -EIO;
3543 }
3544
3545
3546
3547 static int stop(mddev_t *mddev)
3548 {
3549         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3550
3551         md_unregister_thread(mddev->thread);
3552         mddev->thread = NULL;
3553         shrink_stripes(conf);
3554         kfree(conf->stripe_hashtbl);
3555         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
3556         sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
3557         kfree(conf->disks);
3558         kfree(conf);
3559         mddev->private = NULL;
3560         return 0;
3561 }
3562
3563 #if RAID5_DEBUG
3564 static void print_sh (struct seq_file *seq, struct stripe_head *sh)
3565 {
3566         int i;
3567
3568         seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
3569                    (unsigned long long)sh->sector, sh->pd_idx, sh->state);
3570         seq_printf(seq, "sh %llu,  count %d.\n",
3571                    (unsigned long long)sh->sector, atomic_read(&sh->count));
3572         seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
3573         for (i = 0; i < sh->disks; i++) {
3574                 seq_printf(seq, "(cache%d: %p %ld) ",
3575                            i, sh->dev[i].page, sh->dev[i].flags);
3576         }
3577         seq_printf(seq, "\n");
3578 }
3579
3580 static void printall (struct seq_file *seq, raid5_conf_t *conf)
3581 {
3582         struct stripe_head *sh;
3583         struct hlist_node *hn;
3584         int i;
3585
3586         spin_lock_irq(&conf->device_lock);
3587         for (i = 0; i < NR_HASH; i++) {
3588                 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
3589                         if (sh->raid_conf != conf)
3590                                 continue;
3591                         print_sh(seq, sh);
3592                 }
3593         }
3594         spin_unlock_irq(&conf->device_lock);
3595 }
3596 #endif
3597
3598 static void status (struct seq_file *seq, mddev_t *mddev)
3599 {
3600         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3601         int i;
3602
3603         seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
3604         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
3605         for (i = 0; i < conf->raid_disks; i++)
3606                 seq_printf (seq, "%s",
3607                                conf->disks[i].rdev &&
3608                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
3609         seq_printf (seq, "]");
3610 #if RAID5_DEBUG
3611         seq_printf (seq, "\n");
3612         printall(seq, conf);
3613 #endif
3614 }
3615
3616 static void print_raid5_conf (raid5_conf_t *conf)
3617 {
3618         int i;
3619         struct disk_info *tmp;
3620
3621         printk("RAID5 conf printout:\n");
3622         if (!conf) {
3623                 printk("(conf==NULL)\n");
3624                 return;
3625         }
3626         printk(" --- rd:%d wd:%d\n", conf->raid_disks,
3627                  conf->raid_disks - conf->mddev->degraded);
3628
3629         for (i = 0; i < conf->raid_disks; i++) {
3630                 char b[BDEVNAME_SIZE];
3631                 tmp = conf->disks + i;
3632                 if (tmp->rdev)
3633                 printk(" disk %d, o:%d, dev:%s\n",
3634                         i, !test_bit(Faulty, &tmp->rdev->flags),
3635                         bdevname(tmp->rdev->bdev,b));
3636         }
3637 }
3638
3639 static int raid5_spare_active(mddev_t *mddev)
3640 {
3641         int i;
3642         raid5_conf_t *conf = mddev->private;
3643         struct disk_info *tmp;
3644
3645         for (i = 0; i < conf->raid_disks; i++) {
3646                 tmp = conf->disks + i;
3647                 if (tmp->rdev
3648                     && !test_bit(Faulty, &tmp->rdev->flags)
3649                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
3650                         unsigned long flags;
3651                         spin_lock_irqsave(&conf->device_lock, flags);
3652                         mddev->degraded--;
3653                         spin_unlock_irqrestore(&conf->device_lock, flags);
3654                 }
3655         }
3656         print_raid5_conf(conf);
3657         return 0;
3658 }
3659
3660 static int raid5_remove_disk(mddev_t *mddev, int number)
3661 {
3662         raid5_conf_t *conf = mddev->private;
3663         int err = 0;
3664         mdk_rdev_t *rdev;
3665         struct disk_info *p = conf->disks + number;
3666
3667         print_raid5_conf(conf);
3668         rdev = p->rdev;
3669         if (rdev) {
3670                 if (test_bit(In_sync, &rdev->flags) ||
3671                     atomic_read(&rdev->nr_pending)) {
3672                         err = -EBUSY;
3673                         goto abort;
3674                 }
3675                 p->rdev = NULL;
3676                 synchronize_rcu();
3677                 if (atomic_read(&rdev->nr_pending)) {
3678                         /* lost the race, try later */
3679                         err = -EBUSY;
3680                         p->rdev = rdev;
3681                 }
3682         }
3683 abort:
3684
3685         print_raid5_conf(conf);
3686         return err;
3687 }
3688
3689 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
3690 {
3691         raid5_conf_t *conf = mddev->private;
3692         int found = 0;
3693         int disk;
3694         struct disk_info *p;
3695
3696         if (mddev->degraded > conf->max_degraded)
3697                 /* no point adding a device */
3698                 return 0;
3699
3700         /*
3701          * find the disk ... but prefer rdev->saved_raid_disk
3702          * if possible.
3703          */
3704         if (rdev->saved_raid_disk >= 0 &&
3705             conf->disks[rdev->saved_raid_disk].rdev == NULL)
3706                 disk = rdev->saved_raid_disk;
3707         else
3708                 disk = 0;
3709         for ( ; disk < conf->raid_disks; disk++)
3710                 if ((p=conf->disks + disk)->rdev == NULL) {
3711                         clear_bit(In_sync, &rdev->flags);
3712                         rdev->raid_disk = disk;
3713                         found = 1;
3714                         if (rdev->saved_raid_disk != disk)
3715                                 conf->fullsync = 1;
3716                         rcu_assign_pointer(p->rdev, rdev);
3717                         break;
3718                 }
3719         print_raid5_conf(conf);
3720         return found;
3721 }
3722
3723 static int raid5_resize(mddev_t *mddev, sector_t sectors)
3724 {
3725         /* no resync is happening, and there is enough space
3726          * on all devices, so we can resize.
3727          * We need to make sure resync covers any new space.
3728          * If the array is shrinking we should possibly wait until
3729          * any io in the removed space completes, but it hardly seems
3730          * worth it.
3731          */
3732         raid5_conf_t *conf = mddev_to_conf(mddev);
3733
3734         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
3735         mddev->array_size = (sectors * (mddev->raid_disks-conf->max_degraded))>>1;
3736         set_capacity(mddev->gendisk, mddev->array_size << 1);
3737         mddev->changed = 1;
3738         if (sectors/2  > mddev->size && mddev->recovery_cp == MaxSector) {
3739                 mddev->recovery_cp = mddev->size << 1;
3740                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3741         }
3742         mddev->size = sectors /2;
3743         mddev->resync_max_sectors = sectors;
3744         return 0;
3745 }
3746
3747 #ifdef CONFIG_MD_RAID5_RESHAPE
3748 static int raid5_check_reshape(mddev_t *mddev)
3749 {
3750         raid5_conf_t *conf = mddev_to_conf(mddev);
3751         int err;
3752
3753         if (mddev->delta_disks < 0 ||
3754             mddev->new_level != mddev->level)
3755                 return -EINVAL; /* Cannot shrink array or change level yet */
3756         if (mddev->delta_disks == 0)
3757                 return 0; /* nothing to do */
3758
3759         /* Can only proceed if there are plenty of stripe_heads.
3760          * We need a minimum of one full stripe,, and for sensible progress
3761          * it is best to have about 4 times that.
3762          * If we require 4 times, then the default 256 4K stripe_heads will
3763          * allow for chunk sizes up to 256K, which is probably OK.
3764          * If the chunk size is greater, user-space should request more
3765          * stripe_heads first.
3766          */
3767         if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
3768             (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
3769                 printk(KERN_WARNING "raid5: reshape: not enough stripes.  Needed %lu\n",
3770                        (mddev->chunk_size / STRIPE_SIZE)*4);
3771                 return -ENOSPC;
3772         }
3773
3774         err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
3775         if (err)
3776                 return err;
3777
3778         /* looks like we might be able to manage this */
3779         return 0;
3780 }
3781
3782 static int raid5_start_reshape(mddev_t *mddev)
3783 {
3784         raid5_conf_t *conf = mddev_to_conf(mddev);
3785         mdk_rdev_t *rdev;
3786         struct list_head *rtmp;
3787         int spares = 0;
3788         int added_devices = 0;
3789         unsigned long flags;
3790
3791         if (mddev->degraded ||
3792             test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
3793                 return -EBUSY;
3794
3795         ITERATE_RDEV(mddev, rdev, rtmp)
3796                 if (rdev->raid_disk < 0 &&
3797                     !test_bit(Faulty, &rdev->flags))
3798                         spares++;
3799
3800         if (spares < mddev->delta_disks-1)
3801                 /* Not enough devices even to make a degraded array
3802                  * of that size
3803                  */
3804                 return -EINVAL;
3805
3806         atomic_set(&conf->reshape_stripes, 0);
3807         spin_lock_irq(&conf->device_lock);
3808         conf->previous_raid_disks = conf->raid_disks;
3809         conf->raid_disks += mddev->delta_disks;
3810         conf->expand_progress = 0;
3811         conf->expand_lo = 0;
3812         spin_unlock_irq(&conf->device_lock);
3813
3814         /* Add some new drives, as many as will fit.
3815          * We know there are enough to make the newly sized array work.
3816          */
3817         ITERATE_RDEV(mddev, rdev, rtmp)
3818                 if (rdev->raid_disk < 0 &&
3819                     !test_bit(Faulty, &rdev->flags)) {
3820                         if (raid5_add_disk(mddev, rdev)) {
3821                                 char nm[20];
3822                                 set_bit(In_sync, &rdev->flags);
3823                                 added_devices++;
3824                                 rdev->recovery_offset = 0;
3825                                 sprintf(nm, "rd%d", rdev->raid_disk);
3826                                 sysfs_create_link(&mddev->kobj, &rdev->kobj, nm);
3827                         } else
3828                                 break;
3829                 }
3830
3831         spin_lock_irqsave(&conf->device_lock, flags);
3832         mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
3833         spin_unlock_irqrestore(&conf->device_lock, flags);
3834         mddev->raid_disks = conf->raid_disks;
3835         mddev->reshape_position = 0;
3836         set_bit(MD_CHANGE_DEVS, &mddev->flags);
3837
3838         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3839         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3840         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3841         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3842         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3843                                                 "%s_reshape");
3844         if (!mddev->sync_thread) {
3845                 mddev->recovery = 0;
3846                 spin_lock_irq(&conf->device_lock);
3847                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
3848                 conf->expand_progress = MaxSector;
3849                 spin_unlock_irq(&conf->device_lock);
3850                 return -EAGAIN;
3851         }
3852         md_wakeup_thread(mddev->sync_thread);
3853         md_new_event(mddev);
3854         return 0;
3855 }
3856 #endif
3857
3858 static void end_reshape(raid5_conf_t *conf)
3859 {
3860         struct block_device *bdev;
3861
3862         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
3863                 conf->mddev->array_size = conf->mddev->size * (conf->raid_disks-1);
3864                 set_capacity(conf->mddev->gendisk, conf->mddev->array_size << 1);
3865                 conf->mddev->changed = 1;
3866
3867                 bdev = bdget_disk(conf->mddev->gendisk, 0);
3868                 if (bdev) {
3869                         mutex_lock(&bdev->bd_inode->i_mutex);
3870                         i_size_write(bdev->bd_inode, (loff_t)conf->mddev->array_size << 10);
3871                         mutex_unlock(&bdev->bd_inode->i_mutex);
3872                         bdput(bdev);
3873                 }
3874                 spin_lock_irq(&conf->device_lock);
3875                 conf->expand_progress = MaxSector;
3876                 spin_unlock_irq(&conf->device_lock);
3877                 conf->mddev->reshape_position = MaxSector;
3878
3879                 /* read-ahead size must cover two whole stripes, which is
3880                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
3881                  */
3882                 {
3883                         int data_disks = conf->previous_raid_disks - conf->max_degraded;
3884                         int stripe = data_disks *
3885                                 (conf->mddev->chunk_size / PAGE_SIZE);
3886                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3887                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3888                 }
3889         }
3890 }
3891
3892 static void raid5_quiesce(mddev_t *mddev, int state)
3893 {
3894         raid5_conf_t *conf = mddev_to_conf(mddev);
3895
3896         switch(state) {
3897         case 2: /* resume for a suspend */
3898                 wake_up(&conf->wait_for_overlap);
3899                 break;
3900
3901         case 1: /* stop all writes */
3902                 spin_lock_irq(&conf->device_lock);
3903                 conf->quiesce = 1;
3904                 wait_event_lock_irq(conf->wait_for_stripe,
3905                                     atomic_read(&conf->active_stripes) == 0 &&
3906                                     atomic_read(&conf->active_aligned_reads) == 0,
3907                                     conf->device_lock, /* nothing */);
3908                 spin_unlock_irq(&conf->device_lock);
3909                 break;
3910
3911         case 0: /* re-enable writes */
3912                 spin_lock_irq(&conf->device_lock);
3913                 conf->quiesce = 0;
3914                 wake_up(&conf->wait_for_stripe);
3915                 wake_up(&conf->wait_for_overlap);
3916                 spin_unlock_irq(&conf->device_lock);
3917                 break;
3918         }
3919 }
3920
3921 static struct mdk_personality raid6_personality =
3922 {
3923         .name           = "raid6",
3924         .level          = 6,
3925         .owner          = THIS_MODULE,
3926         .make_request   = make_request,
3927         .run            = run,
3928         .stop           = stop,
3929         .status         = status,
3930         .error_handler  = error,
3931         .hot_add_disk   = raid5_add_disk,
3932         .hot_remove_disk= raid5_remove_disk,
3933         .spare_active   = raid5_spare_active,
3934         .sync_request   = sync_request,
3935         .resize         = raid5_resize,
3936         .quiesce        = raid5_quiesce,
3937 };
3938 static struct mdk_personality raid5_personality =
3939 {
3940         .name           = "raid5",
3941         .level          = 5,
3942         .owner          = THIS_MODULE,
3943         .make_request   = make_request,
3944         .run            = run,
3945         .stop           = stop,
3946         .status         = status,
3947         .error_handler  = error,
3948         .hot_add_disk   = raid5_add_disk,
3949         .hot_remove_disk= raid5_remove_disk,
3950         .spare_active   = raid5_spare_active,
3951         .sync_request   = sync_request,
3952         .resize         = raid5_resize,
3953 #ifdef CONFIG_MD_RAID5_RESHAPE
3954         .check_reshape  = raid5_check_reshape,
3955         .start_reshape  = raid5_start_reshape,
3956 #endif
3957         .quiesce        = raid5_quiesce,
3958 };
3959
3960 static struct mdk_personality raid4_personality =
3961 {
3962         .name           = "raid4",
3963         .level          = 4,
3964         .owner          = THIS_MODULE,
3965         .make_request   = make_request,
3966         .run            = run,
3967         .stop           = stop,
3968         .status         = status,
3969         .error_handler  = error,
3970         .hot_add_disk   = raid5_add_disk,
3971         .hot_remove_disk= raid5_remove_disk,
3972         .spare_active   = raid5_spare_active,
3973         .sync_request   = sync_request,
3974         .resize         = raid5_resize,
3975         .quiesce        = raid5_quiesce,
3976 };
3977
3978 static int __init raid5_init(void)
3979 {
3980         int e;
3981
3982         e = raid6_select_algo();
3983         if ( e )
3984                 return e;
3985         register_md_personality(&raid6_personality);
3986         register_md_personality(&raid5_personality);
3987         register_md_personality(&raid4_personality);
3988         return 0;
3989 }
3990
3991 static void raid5_exit(void)
3992 {
3993         unregister_md_personality(&raid6_personality);
3994         unregister_md_personality(&raid5_personality);
3995         unregister_md_personality(&raid4_personality);
3996 }
3997
3998 module_init(raid5_init);
3999 module_exit(raid5_exit);
4000 MODULE_LICENSE("GPL");
4001 MODULE_ALIAS("md-personality-4"); /* RAID5 */
4002 MODULE_ALIAS("md-raid5");
4003 MODULE_ALIAS("md-raid4");
4004 MODULE_ALIAS("md-level-5");
4005 MODULE_ALIAS("md-level-4");
4006 MODULE_ALIAS("md-personality-8"); /* RAID6 */
4007 MODULE_ALIAS("md-raid6");
4008 MODULE_ALIAS("md-level-6");
4009
4010 /* This used to be two separate modules, they were: */
4011 MODULE_ALIAS("raid5");
4012 MODULE_ALIAS("raid6");