[PATCH] md: fix potential memalloc deadlock in md
[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         md_allow_write(conf->mddev);
409
410         /* Step 1 */
411         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
412                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
413                                0, 0, NULL, NULL);
414         if (!sc)
415                 return -ENOMEM;
416
417         for (i = conf->max_nr_stripes; i; i--) {
418                 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
419                 if (!nsh)
420                         break;
421
422                 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
423
424                 nsh->raid_conf = conf;
425                 spin_lock_init(&nsh->lock);
426
427                 list_add(&nsh->lru, &newstripes);
428         }
429         if (i) {
430                 /* didn't get enough, give up */
431                 while (!list_empty(&newstripes)) {
432                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
433                         list_del(&nsh->lru);
434                         kmem_cache_free(sc, nsh);
435                 }
436                 kmem_cache_destroy(sc);
437                 return -ENOMEM;
438         }
439         /* Step 2 - Must use GFP_NOIO now.
440          * OK, we have enough stripes, start collecting inactive
441          * stripes and copying them over
442          */
443         list_for_each_entry(nsh, &newstripes, lru) {
444                 spin_lock_irq(&conf->device_lock);
445                 wait_event_lock_irq(conf->wait_for_stripe,
446                                     !list_empty(&conf->inactive_list),
447                                     conf->device_lock,
448                                     unplug_slaves(conf->mddev)
449                         );
450                 osh = get_free_stripe(conf);
451                 spin_unlock_irq(&conf->device_lock);
452                 atomic_set(&nsh->count, 1);
453                 for(i=0; i<conf->pool_size; i++)
454                         nsh->dev[i].page = osh->dev[i].page;
455                 for( ; i<newsize; i++)
456                         nsh->dev[i].page = NULL;
457                 kmem_cache_free(conf->slab_cache, osh);
458         }
459         kmem_cache_destroy(conf->slab_cache);
460
461         /* Step 3.
462          * At this point, we are holding all the stripes so the array
463          * is completely stalled, so now is a good time to resize
464          * conf->disks.
465          */
466         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
467         if (ndisks) {
468                 for (i=0; i<conf->raid_disks; i++)
469                         ndisks[i] = conf->disks[i];
470                 kfree(conf->disks);
471                 conf->disks = ndisks;
472         } else
473                 err = -ENOMEM;
474
475         /* Step 4, return new stripes to service */
476         while(!list_empty(&newstripes)) {
477                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
478                 list_del_init(&nsh->lru);
479                 for (i=conf->raid_disks; i < newsize; i++)
480                         if (nsh->dev[i].page == NULL) {
481                                 struct page *p = alloc_page(GFP_NOIO);
482                                 nsh->dev[i].page = p;
483                                 if (!p)
484                                         err = -ENOMEM;
485                         }
486                 release_stripe(nsh);
487         }
488         /* critical section pass, GFP_NOIO no longer needed */
489
490         conf->slab_cache = sc;
491         conf->active_name = 1-conf->active_name;
492         conf->pool_size = newsize;
493         return err;
494 }
495 #endif
496
497 static int drop_one_stripe(raid5_conf_t *conf)
498 {
499         struct stripe_head *sh;
500
501         spin_lock_irq(&conf->device_lock);
502         sh = get_free_stripe(conf);
503         spin_unlock_irq(&conf->device_lock);
504         if (!sh)
505                 return 0;
506         BUG_ON(atomic_read(&sh->count));
507         shrink_buffers(sh, conf->pool_size);
508         kmem_cache_free(conf->slab_cache, sh);
509         atomic_dec(&conf->active_stripes);
510         return 1;
511 }
512
513 static void shrink_stripes(raid5_conf_t *conf)
514 {
515         while (drop_one_stripe(conf))
516                 ;
517
518         if (conf->slab_cache)
519                 kmem_cache_destroy(conf->slab_cache);
520         conf->slab_cache = NULL;
521 }
522
523 static int raid5_end_read_request(struct bio * bi, unsigned int bytes_done,
524                                    int error)
525 {
526         struct stripe_head *sh = bi->bi_private;
527         raid5_conf_t *conf = sh->raid_conf;
528         int disks = sh->disks, i;
529         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
530         char b[BDEVNAME_SIZE];
531         mdk_rdev_t *rdev;
532
533         if (bi->bi_size)
534                 return 1;
535
536         for (i=0 ; i<disks; i++)
537                 if (bi == &sh->dev[i].req)
538                         break;
539
540         PRINTK("end_read_request %llu/%d, count: %d, uptodate %d.\n", 
541                 (unsigned long long)sh->sector, i, atomic_read(&sh->count), 
542                 uptodate);
543         if (i == disks) {
544                 BUG();
545                 return 0;
546         }
547
548         if (uptodate) {
549                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
550                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
551                         rdev = conf->disks[i].rdev;
552                         printk(KERN_INFO "raid5:%s: read error corrected (%lu sectors at %llu on %s)\n",
553                                mdname(conf->mddev), STRIPE_SECTORS,
554                                (unsigned long long)sh->sector + rdev->data_offset,
555                                bdevname(rdev->bdev, b));
556                         clear_bit(R5_ReadError, &sh->dev[i].flags);
557                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
558                 }
559                 if (atomic_read(&conf->disks[i].rdev->read_errors))
560                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
561         } else {
562                 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
563                 int retry = 0;
564                 rdev = conf->disks[i].rdev;
565
566                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
567                 atomic_inc(&rdev->read_errors);
568                 if (conf->mddev->degraded)
569                         printk(KERN_WARNING "raid5:%s: read error not correctable (sector %llu on %s).\n",
570                                mdname(conf->mddev),
571                                (unsigned long long)sh->sector + rdev->data_offset,
572                                bdn);
573                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
574                         /* Oh, no!!! */
575                         printk(KERN_WARNING "raid5:%s: read error NOT corrected!! (sector %llu on %s).\n",
576                                mdname(conf->mddev),
577                                (unsigned long long)sh->sector + rdev->data_offset,
578                                bdn);
579                 else if (atomic_read(&rdev->read_errors)
580                          > conf->max_nr_stripes)
581                         printk(KERN_WARNING
582                                "raid5:%s: Too many read errors, failing device %s.\n",
583                                mdname(conf->mddev), bdn);
584                 else
585                         retry = 1;
586                 if (retry)
587                         set_bit(R5_ReadError, &sh->dev[i].flags);
588                 else {
589                         clear_bit(R5_ReadError, &sh->dev[i].flags);
590                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
591                         md_error(conf->mddev, rdev);
592                 }
593         }
594         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
595         clear_bit(R5_LOCKED, &sh->dev[i].flags);
596         set_bit(STRIPE_HANDLE, &sh->state);
597         release_stripe(sh);
598         return 0;
599 }
600
601 static int raid5_end_write_request (struct bio *bi, unsigned int bytes_done,
602                                     int error)
603 {
604         struct stripe_head *sh = bi->bi_private;
605         raid5_conf_t *conf = sh->raid_conf;
606         int disks = sh->disks, i;
607         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
608
609         if (bi->bi_size)
610                 return 1;
611
612         for (i=0 ; i<disks; i++)
613                 if (bi == &sh->dev[i].req)
614                         break;
615
616         PRINTK("end_write_request %llu/%d, count %d, uptodate: %d.\n", 
617                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
618                 uptodate);
619         if (i == disks) {
620                 BUG();
621                 return 0;
622         }
623
624         if (!uptodate)
625                 md_error(conf->mddev, conf->disks[i].rdev);
626
627         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
628         
629         clear_bit(R5_LOCKED, &sh->dev[i].flags);
630         set_bit(STRIPE_HANDLE, &sh->state);
631         release_stripe(sh);
632         return 0;
633 }
634
635
636 static sector_t compute_blocknr(struct stripe_head *sh, int i);
637         
638 static void raid5_build_block (struct stripe_head *sh, int i)
639 {
640         struct r5dev *dev = &sh->dev[i];
641
642         bio_init(&dev->req);
643         dev->req.bi_io_vec = &dev->vec;
644         dev->req.bi_vcnt++;
645         dev->req.bi_max_vecs++;
646         dev->vec.bv_page = dev->page;
647         dev->vec.bv_len = STRIPE_SIZE;
648         dev->vec.bv_offset = 0;
649
650         dev->req.bi_sector = sh->sector;
651         dev->req.bi_private = sh;
652
653         dev->flags = 0;
654         dev->sector = compute_blocknr(sh, i);
655 }
656
657 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
658 {
659         char b[BDEVNAME_SIZE];
660         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
661         PRINTK("raid5: error called\n");
662
663         if (!test_bit(Faulty, &rdev->flags)) {
664                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
665                 if (test_and_clear_bit(In_sync, &rdev->flags)) {
666                         unsigned long flags;
667                         spin_lock_irqsave(&conf->device_lock, flags);
668                         mddev->degraded++;
669                         spin_unlock_irqrestore(&conf->device_lock, flags);
670                         /*
671                          * if recovery was running, make sure it aborts.
672                          */
673                         set_bit(MD_RECOVERY_ERR, &mddev->recovery);
674                 }
675                 set_bit(Faulty, &rdev->flags);
676                 printk (KERN_ALERT
677                         "raid5: Disk failure on %s, disabling device."
678                         " Operation continuing on %d devices\n",
679                         bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
680         }
681 }
682
683 /*
684  * Input: a 'big' sector number,
685  * Output: index of the data and parity disk, and the sector # in them.
686  */
687 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
688                         unsigned int data_disks, unsigned int * dd_idx,
689                         unsigned int * pd_idx, raid5_conf_t *conf)
690 {
691         long stripe;
692         unsigned long chunk_number;
693         unsigned int chunk_offset;
694         sector_t new_sector;
695         int sectors_per_chunk = conf->chunk_size >> 9;
696
697         /* First compute the information on this sector */
698
699         /*
700          * Compute the chunk number and the sector offset inside the chunk
701          */
702         chunk_offset = sector_div(r_sector, sectors_per_chunk);
703         chunk_number = r_sector;
704         BUG_ON(r_sector != chunk_number);
705
706         /*
707          * Compute the stripe number
708          */
709         stripe = chunk_number / data_disks;
710
711         /*
712          * Compute the data disk and parity disk indexes inside the stripe
713          */
714         *dd_idx = chunk_number % data_disks;
715
716         /*
717          * Select the parity disk based on the user selected algorithm.
718          */
719         switch(conf->level) {
720         case 4:
721                 *pd_idx = data_disks;
722                 break;
723         case 5:
724                 switch (conf->algorithm) {
725                 case ALGORITHM_LEFT_ASYMMETRIC:
726                         *pd_idx = data_disks - stripe % raid_disks;
727                         if (*dd_idx >= *pd_idx)
728                                 (*dd_idx)++;
729                         break;
730                 case ALGORITHM_RIGHT_ASYMMETRIC:
731                         *pd_idx = stripe % raid_disks;
732                         if (*dd_idx >= *pd_idx)
733                                 (*dd_idx)++;
734                         break;
735                 case ALGORITHM_LEFT_SYMMETRIC:
736                         *pd_idx = data_disks - stripe % raid_disks;
737                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
738                         break;
739                 case ALGORITHM_RIGHT_SYMMETRIC:
740                         *pd_idx = stripe % raid_disks;
741                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
742                         break;
743                 default:
744                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
745                                 conf->algorithm);
746                 }
747                 break;
748         case 6:
749
750                 /**** FIX THIS ****/
751                 switch (conf->algorithm) {
752                 case ALGORITHM_LEFT_ASYMMETRIC:
753                         *pd_idx = raid_disks - 1 - (stripe % raid_disks);
754                         if (*pd_idx == raid_disks-1)
755                                 (*dd_idx)++;    /* Q D D D P */
756                         else if (*dd_idx >= *pd_idx)
757                                 (*dd_idx) += 2; /* D D P Q D */
758                         break;
759                 case ALGORITHM_RIGHT_ASYMMETRIC:
760                         *pd_idx = stripe % raid_disks;
761                         if (*pd_idx == raid_disks-1)
762                                 (*dd_idx)++;    /* Q D D D P */
763                         else if (*dd_idx >= *pd_idx)
764                                 (*dd_idx) += 2; /* D D P Q D */
765                         break;
766                 case ALGORITHM_LEFT_SYMMETRIC:
767                         *pd_idx = raid_disks - 1 - (stripe % raid_disks);
768                         *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
769                         break;
770                 case ALGORITHM_RIGHT_SYMMETRIC:
771                         *pd_idx = stripe % raid_disks;
772                         *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
773                         break;
774                 default:
775                         printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
776                                 conf->algorithm);
777                 }
778                 break;
779         }
780
781         /*
782          * Finally, compute the new sector number
783          */
784         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
785         return new_sector;
786 }
787
788
789 static sector_t compute_blocknr(struct stripe_head *sh, int i)
790 {
791         raid5_conf_t *conf = sh->raid_conf;
792         int raid_disks = sh->disks;
793         int data_disks = raid_disks - conf->max_degraded;
794         sector_t new_sector = sh->sector, check;
795         int sectors_per_chunk = conf->chunk_size >> 9;
796         sector_t stripe;
797         int chunk_offset;
798         int chunk_number, dummy1, dummy2, dd_idx = i;
799         sector_t r_sector;
800
801
802         chunk_offset = sector_div(new_sector, sectors_per_chunk);
803         stripe = new_sector;
804         BUG_ON(new_sector != stripe);
805
806         if (i == sh->pd_idx)
807                 return 0;
808         switch(conf->level) {
809         case 4: break;
810         case 5:
811                 switch (conf->algorithm) {
812                 case ALGORITHM_LEFT_ASYMMETRIC:
813                 case ALGORITHM_RIGHT_ASYMMETRIC:
814                         if (i > sh->pd_idx)
815                                 i--;
816                         break;
817                 case ALGORITHM_LEFT_SYMMETRIC:
818                 case ALGORITHM_RIGHT_SYMMETRIC:
819                         if (i < sh->pd_idx)
820                                 i += raid_disks;
821                         i -= (sh->pd_idx + 1);
822                         break;
823                 default:
824                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
825                                conf->algorithm);
826                 }
827                 break;
828         case 6:
829                 if (i == raid6_next_disk(sh->pd_idx, raid_disks))
830                         return 0; /* It is the Q disk */
831                 switch (conf->algorithm) {
832                 case ALGORITHM_LEFT_ASYMMETRIC:
833                 case ALGORITHM_RIGHT_ASYMMETRIC:
834                         if (sh->pd_idx == raid_disks-1)
835                                 i--;    /* Q D D D P */
836                         else if (i > sh->pd_idx)
837                                 i -= 2; /* D D P Q D */
838                         break;
839                 case ALGORITHM_LEFT_SYMMETRIC:
840                 case ALGORITHM_RIGHT_SYMMETRIC:
841                         if (sh->pd_idx == raid_disks-1)
842                                 i--; /* Q D D D P */
843                         else {
844                                 /* D D P Q D */
845                                 if (i < sh->pd_idx)
846                                         i += raid_disks;
847                                 i -= (sh->pd_idx + 2);
848                         }
849                         break;
850                 default:
851                         printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
852                                 conf->algorithm);
853                 }
854                 break;
855         }
856
857         chunk_number = stripe * data_disks + i;
858         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
859
860         check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
861         if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
862                 printk(KERN_ERR "compute_blocknr: map not correct\n");
863                 return 0;
864         }
865         return r_sector;
866 }
867
868
869
870 /*
871  * Copy data between a page in the stripe cache, and one or more bion
872  * The page could align with the middle of the bio, or there could be
873  * several bion, each with several bio_vecs, which cover part of the page
874  * Multiple bion are linked together on bi_next.  There may be extras
875  * at the end of this list.  We ignore them.
876  */
877 static void copy_data(int frombio, struct bio *bio,
878                      struct page *page,
879                      sector_t sector)
880 {
881         char *pa = page_address(page);
882         struct bio_vec *bvl;
883         int i;
884         int page_offset;
885
886         if (bio->bi_sector >= sector)
887                 page_offset = (signed)(bio->bi_sector - sector) * 512;
888         else
889                 page_offset = (signed)(sector - bio->bi_sector) * -512;
890         bio_for_each_segment(bvl, bio, i) {
891                 int len = bio_iovec_idx(bio,i)->bv_len;
892                 int clen;
893                 int b_offset = 0;
894
895                 if (page_offset < 0) {
896                         b_offset = -page_offset;
897                         page_offset += b_offset;
898                         len -= b_offset;
899                 }
900
901                 if (len > 0 && page_offset + len > STRIPE_SIZE)
902                         clen = STRIPE_SIZE - page_offset;
903                 else clen = len;
904
905                 if (clen > 0) {
906                         char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
907                         if (frombio)
908                                 memcpy(pa+page_offset, ba+b_offset, clen);
909                         else
910                                 memcpy(ba+b_offset, pa+page_offset, clen);
911                         __bio_kunmap_atomic(ba, KM_USER0);
912                 }
913                 if (clen < len) /* hit end of page */
914                         break;
915                 page_offset +=  len;
916         }
917 }
918
919 #define check_xor()     do {                                            \
920                            if (count == MAX_XOR_BLOCKS) {               \
921                                 xor_block(count, STRIPE_SIZE, ptr);     \
922                                 count = 1;                              \
923                            }                                            \
924                         } while(0)
925
926
927 static void compute_block(struct stripe_head *sh, int dd_idx)
928 {
929         int i, count, disks = sh->disks;
930         void *ptr[MAX_XOR_BLOCKS], *p;
931
932         PRINTK("compute_block, stripe %llu, idx %d\n", 
933                 (unsigned long long)sh->sector, dd_idx);
934
935         ptr[0] = page_address(sh->dev[dd_idx].page);
936         memset(ptr[0], 0, STRIPE_SIZE);
937         count = 1;
938         for (i = disks ; i--; ) {
939                 if (i == dd_idx)
940                         continue;
941                 p = page_address(sh->dev[i].page);
942                 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
943                         ptr[count++] = p;
944                 else
945                         printk(KERN_ERR "compute_block() %d, stripe %llu, %d"
946                                 " not present\n", dd_idx,
947                                 (unsigned long long)sh->sector, i);
948
949                 check_xor();
950         }
951         if (count != 1)
952                 xor_block(count, STRIPE_SIZE, ptr);
953         set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
954 }
955
956 static void compute_parity5(struct stripe_head *sh, int method)
957 {
958         raid5_conf_t *conf = sh->raid_conf;
959         int i, pd_idx = sh->pd_idx, disks = sh->disks, count;
960         void *ptr[MAX_XOR_BLOCKS];
961         struct bio *chosen;
962
963         PRINTK("compute_parity5, stripe %llu, method %d\n",
964                 (unsigned long long)sh->sector, method);
965
966         count = 1;
967         ptr[0] = page_address(sh->dev[pd_idx].page);
968         switch(method) {
969         case READ_MODIFY_WRITE:
970                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags));
971                 for (i=disks ; i-- ;) {
972                         if (i==pd_idx)
973                                 continue;
974                         if (sh->dev[i].towrite &&
975                             test_bit(R5_UPTODATE, &sh->dev[i].flags)) {
976                                 ptr[count++] = page_address(sh->dev[i].page);
977                                 chosen = sh->dev[i].towrite;
978                                 sh->dev[i].towrite = NULL;
979
980                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
981                                         wake_up(&conf->wait_for_overlap);
982
983                                 BUG_ON(sh->dev[i].written);
984                                 sh->dev[i].written = chosen;
985                                 check_xor();
986                         }
987                 }
988                 break;
989         case RECONSTRUCT_WRITE:
990                 memset(ptr[0], 0, STRIPE_SIZE);
991                 for (i= disks; i-- ;)
992                         if (i!=pd_idx && sh->dev[i].towrite) {
993                                 chosen = sh->dev[i].towrite;
994                                 sh->dev[i].towrite = NULL;
995
996                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
997                                         wake_up(&conf->wait_for_overlap);
998
999                                 BUG_ON(sh->dev[i].written);
1000                                 sh->dev[i].written = chosen;
1001                         }
1002                 break;
1003         case CHECK_PARITY:
1004                 break;
1005         }
1006         if (count>1) {
1007                 xor_block(count, STRIPE_SIZE, ptr);
1008                 count = 1;
1009         }
1010         
1011         for (i = disks; i--;)
1012                 if (sh->dev[i].written) {
1013                         sector_t sector = sh->dev[i].sector;
1014                         struct bio *wbi = sh->dev[i].written;
1015                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1016                                 copy_data(1, wbi, sh->dev[i].page, sector);
1017                                 wbi = r5_next_bio(wbi, sector);
1018                         }
1019
1020                         set_bit(R5_LOCKED, &sh->dev[i].flags);
1021                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
1022                 }
1023
1024         switch(method) {
1025         case RECONSTRUCT_WRITE:
1026         case CHECK_PARITY:
1027                 for (i=disks; i--;)
1028                         if (i != pd_idx) {
1029                                 ptr[count++] = page_address(sh->dev[i].page);
1030                                 check_xor();
1031                         }
1032                 break;
1033         case READ_MODIFY_WRITE:
1034                 for (i = disks; i--;)
1035                         if (sh->dev[i].written) {
1036                                 ptr[count++] = page_address(sh->dev[i].page);
1037                                 check_xor();
1038                         }
1039         }
1040         if (count != 1)
1041                 xor_block(count, STRIPE_SIZE, ptr);
1042         
1043         if (method != CHECK_PARITY) {
1044                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1045                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
1046         } else
1047                 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1048 }
1049
1050 static void compute_parity6(struct stripe_head *sh, int method)
1051 {
1052         raid6_conf_t *conf = sh->raid_conf;
1053         int i, pd_idx = sh->pd_idx, qd_idx, d0_idx, disks = conf->raid_disks, count;
1054         struct bio *chosen;
1055         /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1056         void *ptrs[disks];
1057
1058         qd_idx = raid6_next_disk(pd_idx, disks);
1059         d0_idx = raid6_next_disk(qd_idx, disks);
1060
1061         PRINTK("compute_parity, stripe %llu, method %d\n",
1062                 (unsigned long long)sh->sector, method);
1063
1064         switch(method) {
1065         case READ_MODIFY_WRITE:
1066                 BUG();          /* READ_MODIFY_WRITE N/A for RAID-6 */
1067         case RECONSTRUCT_WRITE:
1068                 for (i= disks; i-- ;)
1069                         if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1070                                 chosen = sh->dev[i].towrite;
1071                                 sh->dev[i].towrite = NULL;
1072
1073                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1074                                         wake_up(&conf->wait_for_overlap);
1075
1076                                 BUG_ON(sh->dev[i].written);
1077                                 sh->dev[i].written = chosen;
1078                         }
1079                 break;
1080         case CHECK_PARITY:
1081                 BUG();          /* Not implemented yet */
1082         }
1083
1084         for (i = disks; i--;)
1085                 if (sh->dev[i].written) {
1086                         sector_t sector = sh->dev[i].sector;
1087                         struct bio *wbi = sh->dev[i].written;
1088                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1089                                 copy_data(1, wbi, sh->dev[i].page, sector);
1090                                 wbi = r5_next_bio(wbi, sector);
1091                         }
1092
1093                         set_bit(R5_LOCKED, &sh->dev[i].flags);
1094                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
1095                 }
1096
1097 //      switch(method) {
1098 //      case RECONSTRUCT_WRITE:
1099 //      case CHECK_PARITY:
1100 //      case UPDATE_PARITY:
1101                 /* Note that unlike RAID-5, the ordering of the disks matters greatly. */
1102                 /* FIX: Is this ordering of drives even remotely optimal? */
1103                 count = 0;
1104                 i = d0_idx;
1105                 do {
1106                         ptrs[count++] = page_address(sh->dev[i].page);
1107                         if (count <= disks-2 && !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1108                                 printk("block %d/%d not uptodate on parity calc\n", i,count);
1109                         i = raid6_next_disk(i, disks);
1110                 } while ( i != d0_idx );
1111 //              break;
1112 //      }
1113
1114         raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs);
1115
1116         switch(method) {
1117         case RECONSTRUCT_WRITE:
1118                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1119                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1120                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
1121                 set_bit(R5_LOCKED,   &sh->dev[qd_idx].flags);
1122                 break;
1123         case UPDATE_PARITY:
1124                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1125                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1126                 break;
1127         }
1128 }
1129
1130
1131 /* Compute one missing block */
1132 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1133 {
1134         raid6_conf_t *conf = sh->raid_conf;
1135         int i, count, disks = conf->raid_disks;
1136         void *ptr[MAX_XOR_BLOCKS], *p;
1137         int pd_idx = sh->pd_idx;
1138         int qd_idx = raid6_next_disk(pd_idx, disks);
1139
1140         PRINTK("compute_block_1, stripe %llu, idx %d\n",
1141                 (unsigned long long)sh->sector, dd_idx);
1142
1143         if ( dd_idx == qd_idx ) {
1144                 /* We're actually computing the Q drive */
1145                 compute_parity6(sh, UPDATE_PARITY);
1146         } else {
1147                 ptr[0] = page_address(sh->dev[dd_idx].page);
1148                 if (!nozero) memset(ptr[0], 0, STRIPE_SIZE);
1149                 count = 1;
1150                 for (i = disks ; i--; ) {
1151                         if (i == dd_idx || i == qd_idx)
1152                                 continue;
1153                         p = page_address(sh->dev[i].page);
1154                         if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1155                                 ptr[count++] = p;
1156                         else
1157                                 printk("compute_block() %d, stripe %llu, %d"
1158                                        " not present\n", dd_idx,
1159                                        (unsigned long long)sh->sector, i);
1160
1161                         check_xor();
1162                 }
1163                 if (count != 1)
1164                         xor_block(count, STRIPE_SIZE, ptr);
1165                 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1166                 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1167         }
1168 }
1169
1170 /* Compute two missing blocks */
1171 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1172 {
1173         raid6_conf_t *conf = sh->raid_conf;
1174         int i, count, disks = conf->raid_disks;
1175         int pd_idx = sh->pd_idx;
1176         int qd_idx = raid6_next_disk(pd_idx, disks);
1177         int d0_idx = raid6_next_disk(qd_idx, disks);
1178         int faila, failb;
1179
1180         /* faila and failb are disk numbers relative to d0_idx */
1181         /* pd_idx become disks-2 and qd_idx become disks-1 */
1182         faila = (dd_idx1 < d0_idx) ? dd_idx1+(disks-d0_idx) : dd_idx1-d0_idx;
1183         failb = (dd_idx2 < d0_idx) ? dd_idx2+(disks-d0_idx) : dd_idx2-d0_idx;
1184
1185         BUG_ON(faila == failb);
1186         if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1187
1188         PRINTK("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1189                (unsigned long long)sh->sector, dd_idx1, dd_idx2, faila, failb);
1190
1191         if ( failb == disks-1 ) {
1192                 /* Q disk is one of the missing disks */
1193                 if ( faila == disks-2 ) {
1194                         /* Missing P+Q, just recompute */
1195                         compute_parity6(sh, UPDATE_PARITY);
1196                         return;
1197                 } else {
1198                         /* We're missing D+Q; recompute D from P */
1199                         compute_block_1(sh, (dd_idx1 == qd_idx) ? dd_idx2 : dd_idx1, 0);
1200                         compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1201                         return;
1202                 }
1203         }
1204
1205         /* We're missing D+P or D+D; build pointer table */
1206         {
1207                 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1208                 void *ptrs[disks];
1209
1210                 count = 0;
1211                 i = d0_idx;
1212                 do {
1213                         ptrs[count++] = page_address(sh->dev[i].page);
1214                         i = raid6_next_disk(i, disks);
1215                         if (i != dd_idx1 && i != dd_idx2 &&
1216                             !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1217                                 printk("compute_2 with missing block %d/%d\n", count, i);
1218                 } while ( i != d0_idx );
1219
1220                 if ( failb == disks-2 ) {
1221                         /* We're missing D+P. */
1222                         raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs);
1223                 } else {
1224                         /* We're missing D+D. */
1225                         raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs);
1226                 }
1227
1228                 /* Both the above update both missing blocks */
1229                 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1230                 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1231         }
1232 }
1233
1234
1235
1236 /*
1237  * Each stripe/dev can have one or more bion attached.
1238  * toread/towrite point to the first in a chain.
1239  * The bi_next chain must be in order.
1240  */
1241 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1242 {
1243         struct bio **bip;
1244         raid5_conf_t *conf = sh->raid_conf;
1245         int firstwrite=0;
1246
1247         PRINTK("adding bh b#%llu to stripe s#%llu\n",
1248                 (unsigned long long)bi->bi_sector,
1249                 (unsigned long long)sh->sector);
1250
1251
1252         spin_lock(&sh->lock);
1253         spin_lock_irq(&conf->device_lock);
1254         if (forwrite) {
1255                 bip = &sh->dev[dd_idx].towrite;
1256                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1257                         firstwrite = 1;
1258         } else
1259                 bip = &sh->dev[dd_idx].toread;
1260         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1261                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1262                         goto overlap;
1263                 bip = & (*bip)->bi_next;
1264         }
1265         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1266                 goto overlap;
1267
1268         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1269         if (*bip)
1270                 bi->bi_next = *bip;
1271         *bip = bi;
1272         bi->bi_phys_segments ++;
1273         spin_unlock_irq(&conf->device_lock);
1274         spin_unlock(&sh->lock);
1275
1276         PRINTK("added bi b#%llu to stripe s#%llu, disk %d.\n",
1277                 (unsigned long long)bi->bi_sector,
1278                 (unsigned long long)sh->sector, dd_idx);
1279
1280         if (conf->mddev->bitmap && firstwrite) {
1281                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1282                                   STRIPE_SECTORS, 0);
1283                 sh->bm_seq = conf->seq_flush+1;
1284                 set_bit(STRIPE_BIT_DELAY, &sh->state);
1285         }
1286
1287         if (forwrite) {
1288                 /* check if page is covered */
1289                 sector_t sector = sh->dev[dd_idx].sector;
1290                 for (bi=sh->dev[dd_idx].towrite;
1291                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1292                              bi && bi->bi_sector <= sector;
1293                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1294                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1295                                 sector = bi->bi_sector + (bi->bi_size>>9);
1296                 }
1297                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1298                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1299         }
1300         return 1;
1301
1302  overlap:
1303         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1304         spin_unlock_irq(&conf->device_lock);
1305         spin_unlock(&sh->lock);
1306         return 0;
1307 }
1308
1309 static void end_reshape(raid5_conf_t *conf);
1310
1311 static int page_is_zero(struct page *p)
1312 {
1313         char *a = page_address(p);
1314         return ((*(u32*)a) == 0 &&
1315                 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1316 }
1317
1318 static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks)
1319 {
1320         int sectors_per_chunk = conf->chunk_size >> 9;
1321         int pd_idx, dd_idx;
1322         int chunk_offset = sector_div(stripe, sectors_per_chunk);
1323
1324         raid5_compute_sector(stripe * (disks - conf->max_degraded)
1325                              *sectors_per_chunk + chunk_offset,
1326                              disks, disks - conf->max_degraded,
1327                              &dd_idx, &pd_idx, conf);
1328         return pd_idx;
1329 }
1330
1331
1332 /*
1333  * handle_stripe - do things to a stripe.
1334  *
1335  * We lock the stripe and then examine the state of various bits
1336  * to see what needs to be done.
1337  * Possible results:
1338  *    return some read request which now have data
1339  *    return some write requests which are safely on disc
1340  *    schedule a read on some buffers
1341  *    schedule a write of some buffers
1342  *    return confirmation of parity correctness
1343  *
1344  * Parity calculations are done inside the stripe lock
1345  * buffers are taken off read_list or write_list, and bh_cache buffers
1346  * get BH_Lock set before the stripe lock is released.
1347  *
1348  */
1349  
1350 static void handle_stripe5(struct stripe_head *sh)
1351 {
1352         raid5_conf_t *conf = sh->raid_conf;
1353         int disks = sh->disks;
1354         struct bio *return_bi= NULL;
1355         struct bio *bi;
1356         int i;
1357         int syncing, expanding, expanded;
1358         int locked=0, uptodate=0, to_read=0, to_write=0, failed=0, written=0;
1359         int non_overwrite = 0;
1360         int failed_num=0;
1361         struct r5dev *dev;
1362
1363         PRINTK("handling stripe %llu, cnt=%d, pd_idx=%d\n",
1364                 (unsigned long long)sh->sector, atomic_read(&sh->count),
1365                 sh->pd_idx);
1366
1367         spin_lock(&sh->lock);
1368         clear_bit(STRIPE_HANDLE, &sh->state);
1369         clear_bit(STRIPE_DELAYED, &sh->state);
1370
1371         syncing = test_bit(STRIPE_SYNCING, &sh->state);
1372         expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
1373         expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
1374         /* Now to look around and see what can be done */
1375
1376         rcu_read_lock();
1377         for (i=disks; i--; ) {
1378                 mdk_rdev_t *rdev;
1379                 dev = &sh->dev[i];
1380                 clear_bit(R5_Insync, &dev->flags);
1381
1382                 PRINTK("check %d: state 0x%lx read %p write %p written %p\n",
1383                         i, dev->flags, dev->toread, dev->towrite, dev->written);
1384                 /* maybe we can reply to a read */
1385                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
1386                         struct bio *rbi, *rbi2;
1387                         PRINTK("Return read for disc %d\n", i);
1388                         spin_lock_irq(&conf->device_lock);
1389                         rbi = dev->toread;
1390                         dev->toread = NULL;
1391                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
1392                                 wake_up(&conf->wait_for_overlap);
1393                         spin_unlock_irq(&conf->device_lock);
1394                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1395                                 copy_data(0, rbi, dev->page, dev->sector);
1396                                 rbi2 = r5_next_bio(rbi, dev->sector);
1397                                 spin_lock_irq(&conf->device_lock);
1398                                 if (--rbi->bi_phys_segments == 0) {
1399                                         rbi->bi_next = return_bi;
1400                                         return_bi = rbi;
1401                                 }
1402                                 spin_unlock_irq(&conf->device_lock);
1403                                 rbi = rbi2;
1404                         }
1405                 }
1406
1407                 /* now count some things */
1408                 if (test_bit(R5_LOCKED, &dev->flags)) locked++;
1409                 if (test_bit(R5_UPTODATE, &dev->flags)) uptodate++;
1410
1411                 
1412                 if (dev->toread) to_read++;
1413                 if (dev->towrite) {
1414                         to_write++;
1415                         if (!test_bit(R5_OVERWRITE, &dev->flags))
1416                                 non_overwrite++;
1417                 }
1418                 if (dev->written) written++;
1419                 rdev = rcu_dereference(conf->disks[i].rdev);
1420                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
1421                         /* The ReadError flag will just be confusing now */
1422                         clear_bit(R5_ReadError, &dev->flags);
1423                         clear_bit(R5_ReWrite, &dev->flags);
1424                 }
1425                 if (!rdev || !test_bit(In_sync, &rdev->flags)
1426                     || test_bit(R5_ReadError, &dev->flags)) {
1427                         failed++;
1428                         failed_num = i;
1429                 } else
1430                         set_bit(R5_Insync, &dev->flags);
1431         }
1432         rcu_read_unlock();
1433         PRINTK("locked=%d uptodate=%d to_read=%d"
1434                 " to_write=%d failed=%d failed_num=%d\n",
1435                 locked, uptodate, to_read, to_write, failed, failed_num);
1436         /* check if the array has lost two devices and, if so, some requests might
1437          * need to be failed
1438          */
1439         if (failed > 1 && to_read+to_write+written) {
1440                 for (i=disks; i--; ) {
1441                         int bitmap_end = 0;
1442
1443                         if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1444                                 mdk_rdev_t *rdev;
1445                                 rcu_read_lock();
1446                                 rdev = rcu_dereference(conf->disks[i].rdev);
1447                                 if (rdev && test_bit(In_sync, &rdev->flags))
1448                                         /* multiple read failures in one stripe */
1449                                         md_error(conf->mddev, rdev);
1450                                 rcu_read_unlock();
1451                         }
1452
1453                         spin_lock_irq(&conf->device_lock);
1454                         /* fail all writes first */
1455                         bi = sh->dev[i].towrite;
1456                         sh->dev[i].towrite = NULL;
1457                         if (bi) { to_write--; bitmap_end = 1; }
1458
1459                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1460                                 wake_up(&conf->wait_for_overlap);
1461
1462                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1463                                 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1464                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1465                                 if (--bi->bi_phys_segments == 0) {
1466                                         md_write_end(conf->mddev);
1467                                         bi->bi_next = return_bi;
1468                                         return_bi = bi;
1469                                 }
1470                                 bi = nextbi;
1471                         }
1472                         /* and fail all 'written' */
1473                         bi = sh->dev[i].written;
1474                         sh->dev[i].written = NULL;
1475                         if (bi) bitmap_end = 1;
1476                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS) {
1477                                 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1478                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1479                                 if (--bi->bi_phys_segments == 0) {
1480                                         md_write_end(conf->mddev);
1481                                         bi->bi_next = return_bi;
1482                                         return_bi = bi;
1483                                 }
1484                                 bi = bi2;
1485                         }
1486
1487                         /* fail any reads if this device is non-operational */
1488                         if (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1489                             test_bit(R5_ReadError, &sh->dev[i].flags)) {
1490                                 bi = sh->dev[i].toread;
1491                                 sh->dev[i].toread = NULL;
1492                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1493                                         wake_up(&conf->wait_for_overlap);
1494                                 if (bi) to_read--;
1495                                 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1496                                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1497                                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1498                                         if (--bi->bi_phys_segments == 0) {
1499                                                 bi->bi_next = return_bi;
1500                                                 return_bi = bi;
1501                                         }
1502                                         bi = nextbi;
1503                                 }
1504                         }
1505                         spin_unlock_irq(&conf->device_lock);
1506                         if (bitmap_end)
1507                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1508                                                 STRIPE_SECTORS, 0, 0);
1509                 }
1510         }
1511         if (failed > 1 && syncing) {
1512                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
1513                 clear_bit(STRIPE_SYNCING, &sh->state);
1514                 syncing = 0;
1515         }
1516
1517         /* might be able to return some write requests if the parity block
1518          * is safe, or on a failed drive
1519          */
1520         dev = &sh->dev[sh->pd_idx];
1521         if ( written &&
1522              ( (test_bit(R5_Insync, &dev->flags) && !test_bit(R5_LOCKED, &dev->flags) &&
1523                 test_bit(R5_UPTODATE, &dev->flags))
1524                || (failed == 1 && failed_num == sh->pd_idx))
1525             ) {
1526             /* any written block on an uptodate or failed drive can be returned.
1527              * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but 
1528              * never LOCKED, so we don't need to test 'failed' directly.
1529              */
1530             for (i=disks; i--; )
1531                 if (sh->dev[i].written) {
1532                     dev = &sh->dev[i];
1533                     if (!test_bit(R5_LOCKED, &dev->flags) &&
1534                          test_bit(R5_UPTODATE, &dev->flags) ) {
1535                         /* We can return any write requests */
1536                             struct bio *wbi, *wbi2;
1537                             int bitmap_end = 0;
1538                             PRINTK("Return write for disc %d\n", i);
1539                             spin_lock_irq(&conf->device_lock);
1540                             wbi = dev->written;
1541                             dev->written = NULL;
1542                             while (wbi && wbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1543                                     wbi2 = r5_next_bio(wbi, dev->sector);
1544                                     if (--wbi->bi_phys_segments == 0) {
1545                                             md_write_end(conf->mddev);
1546                                             wbi->bi_next = return_bi;
1547                                             return_bi = wbi;
1548                                     }
1549                                     wbi = wbi2;
1550                             }
1551                             if (dev->towrite == NULL)
1552                                     bitmap_end = 1;
1553                             spin_unlock_irq(&conf->device_lock);
1554                             if (bitmap_end)
1555                                     bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1556                                                     STRIPE_SECTORS,
1557                                                     !test_bit(STRIPE_DEGRADED, &sh->state), 0);
1558                     }
1559                 }
1560         }
1561
1562         /* Now we might consider reading some blocks, either to check/generate
1563          * parity, or to satisfy requests
1564          * or to load a block that is being partially written.
1565          */
1566         if (to_read || non_overwrite || (syncing && (uptodate < disks)) || expanding) {
1567                 for (i=disks; i--;) {
1568                         dev = &sh->dev[i];
1569                         if (!test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1570                             (dev->toread ||
1571                              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1572                              syncing ||
1573                              expanding ||
1574                              (failed && (sh->dev[failed_num].toread ||
1575                                          (sh->dev[failed_num].towrite && !test_bit(R5_OVERWRITE, &sh->dev[failed_num].flags))))
1576                                     )
1577                                 ) {
1578                                 /* we would like to get this block, possibly
1579                                  * by computing it, but we might not be able to
1580                                  */
1581                                 if (uptodate == disks-1) {
1582                                         PRINTK("Computing block %d\n", i);
1583                                         compute_block(sh, i);
1584                                         uptodate++;
1585                                 } else if (test_bit(R5_Insync, &dev->flags)) {
1586                                         set_bit(R5_LOCKED, &dev->flags);
1587                                         set_bit(R5_Wantread, &dev->flags);
1588                                         locked++;
1589                                         PRINTK("Reading block %d (sync=%d)\n", 
1590                                                 i, syncing);
1591                                 }
1592                         }
1593                 }
1594                 set_bit(STRIPE_HANDLE, &sh->state);
1595         }
1596
1597         /* now to consider writing and what else, if anything should be read */
1598         if (to_write) {
1599                 int rmw=0, rcw=0;
1600                 for (i=disks ; i--;) {
1601                         /* would I have to read this buffer for read_modify_write */
1602                         dev = &sh->dev[i];
1603                         if ((dev->towrite || i == sh->pd_idx) &&
1604                             (!test_bit(R5_LOCKED, &dev->flags) 
1605                                     ) &&
1606                             !test_bit(R5_UPTODATE, &dev->flags)) {
1607                                 if (test_bit(R5_Insync, &dev->flags)
1608 /*                                  && !(!mddev->insync && i == sh->pd_idx) */
1609                                         )
1610                                         rmw++;
1611                                 else rmw += 2*disks;  /* cannot read it */
1612                         }
1613                         /* Would I have to read this buffer for reconstruct_write */
1614                         if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
1615                             (!test_bit(R5_LOCKED, &dev->flags) 
1616                                     ) &&
1617                             !test_bit(R5_UPTODATE, &dev->flags)) {
1618                                 if (test_bit(R5_Insync, &dev->flags)) rcw++;
1619                                 else rcw += 2*disks;
1620                         }
1621                 }
1622                 PRINTK("for sector %llu, rmw=%d rcw=%d\n", 
1623                         (unsigned long long)sh->sector, rmw, rcw);
1624                 set_bit(STRIPE_HANDLE, &sh->state);
1625                 if (rmw < rcw && rmw > 0)
1626                         /* prefer read-modify-write, but need to get some data */
1627                         for (i=disks; i--;) {
1628                                 dev = &sh->dev[i];
1629                                 if ((dev->towrite || i == sh->pd_idx) &&
1630                                     !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1631                                     test_bit(R5_Insync, &dev->flags)) {
1632                                         if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1633                                         {
1634                                                 PRINTK("Read_old block %d for r-m-w\n", i);
1635                                                 set_bit(R5_LOCKED, &dev->flags);
1636                                                 set_bit(R5_Wantread, &dev->flags);
1637                                                 locked++;
1638                                         } else {
1639                                                 set_bit(STRIPE_DELAYED, &sh->state);
1640                                                 set_bit(STRIPE_HANDLE, &sh->state);
1641                                         }
1642                                 }
1643                         }
1644                 if (rcw <= rmw && rcw > 0)
1645                         /* want reconstruct write, but need to get some data */
1646                         for (i=disks; i--;) {
1647                                 dev = &sh->dev[i];
1648                                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
1649                                     !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1650                                     test_bit(R5_Insync, &dev->flags)) {
1651                                         if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1652                                         {
1653                                                 PRINTK("Read_old block %d for Reconstruct\n", i);
1654                                                 set_bit(R5_LOCKED, &dev->flags);
1655                                                 set_bit(R5_Wantread, &dev->flags);
1656                                                 locked++;
1657                                         } else {
1658                                                 set_bit(STRIPE_DELAYED, &sh->state);
1659                                                 set_bit(STRIPE_HANDLE, &sh->state);
1660                                         }
1661                                 }
1662                         }
1663                 /* now if nothing is locked, and if we have enough data, we can start a write request */
1664                 if (locked == 0 && (rcw == 0 ||rmw == 0) &&
1665                     !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
1666                         PRINTK("Computing parity...\n");
1667                         compute_parity5(sh, rcw==0 ? RECONSTRUCT_WRITE : READ_MODIFY_WRITE);
1668                         /* now every locked buffer is ready to be written */
1669                         for (i=disks; i--;)
1670                                 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
1671                                         PRINTK("Writing block %d\n", i);
1672                                         locked++;
1673                                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
1674                                         if (!test_bit(R5_Insync, &sh->dev[i].flags)
1675                                             || (i==sh->pd_idx && failed == 0))
1676                                                 set_bit(STRIPE_INSYNC, &sh->state);
1677                                 }
1678                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
1679                                 atomic_dec(&conf->preread_active_stripes);
1680                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
1681                                         md_wakeup_thread(conf->mddev->thread);
1682                         }
1683                 }
1684         }
1685
1686         /* maybe we need to check and possibly fix the parity for this stripe
1687          * Any reads will already have been scheduled, so we just see if enough data
1688          * is available
1689          */
1690         if (syncing && locked == 0 &&
1691             !test_bit(STRIPE_INSYNC, &sh->state)) {
1692                 set_bit(STRIPE_HANDLE, &sh->state);
1693                 if (failed == 0) {
1694                         BUG_ON(uptodate != disks);
1695                         compute_parity5(sh, CHECK_PARITY);
1696                         uptodate--;
1697                         if (page_is_zero(sh->dev[sh->pd_idx].page)) {
1698                                 /* parity is correct (on disc, not in buffer any more) */
1699                                 set_bit(STRIPE_INSYNC, &sh->state);
1700                         } else {
1701                                 conf->mddev->resync_mismatches += STRIPE_SECTORS;
1702                                 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
1703                                         /* don't try to repair!! */
1704                                         set_bit(STRIPE_INSYNC, &sh->state);
1705                                 else {
1706                                         compute_block(sh, sh->pd_idx);
1707                                         uptodate++;
1708                                 }
1709                         }
1710                 }
1711                 if (!test_bit(STRIPE_INSYNC, &sh->state)) {
1712                         /* either failed parity check, or recovery is happening */
1713                         if (failed==0)
1714                                 failed_num = sh->pd_idx;
1715                         dev = &sh->dev[failed_num];
1716                         BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
1717                         BUG_ON(uptodate != disks);
1718
1719                         set_bit(R5_LOCKED, &dev->flags);
1720                         set_bit(R5_Wantwrite, &dev->flags);
1721                         clear_bit(STRIPE_DEGRADED, &sh->state);
1722                         locked++;
1723                         set_bit(STRIPE_INSYNC, &sh->state);
1724                 }
1725         }
1726         if (syncing && locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
1727                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
1728                 clear_bit(STRIPE_SYNCING, &sh->state);
1729         }
1730
1731         /* If the failed drive is just a ReadError, then we might need to progress
1732          * the repair/check process
1733          */
1734         if (failed == 1 && ! conf->mddev->ro &&
1735             test_bit(R5_ReadError, &sh->dev[failed_num].flags)
1736             && !test_bit(R5_LOCKED, &sh->dev[failed_num].flags)
1737             && test_bit(R5_UPTODATE, &sh->dev[failed_num].flags)
1738                 ) {
1739                 dev = &sh->dev[failed_num];
1740                 if (!test_bit(R5_ReWrite, &dev->flags)) {
1741                         set_bit(R5_Wantwrite, &dev->flags);
1742                         set_bit(R5_ReWrite, &dev->flags);
1743                         set_bit(R5_LOCKED, &dev->flags);
1744                         locked++;
1745                 } else {
1746                         /* let's read it back */
1747                         set_bit(R5_Wantread, &dev->flags);
1748                         set_bit(R5_LOCKED, &dev->flags);
1749                         locked++;
1750                 }
1751         }
1752
1753         if (expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
1754                 /* Need to write out all blocks after computing parity */
1755                 sh->disks = conf->raid_disks;
1756                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf, conf->raid_disks);
1757                 compute_parity5(sh, RECONSTRUCT_WRITE);
1758                 for (i= conf->raid_disks; i--;) {
1759                         set_bit(R5_LOCKED, &sh->dev[i].flags);
1760                         locked++;
1761                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
1762                 }
1763                 clear_bit(STRIPE_EXPANDING, &sh->state);
1764         } else if (expanded) {
1765                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
1766                 atomic_dec(&conf->reshape_stripes);
1767                 wake_up(&conf->wait_for_overlap);
1768                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
1769         }
1770
1771         if (expanding && locked == 0) {
1772                 /* We have read all the blocks in this stripe and now we need to
1773                  * copy some of them into a target stripe for expand.
1774                  */
1775                 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
1776                 for (i=0; i< sh->disks; i++)
1777                         if (i != sh->pd_idx) {
1778                                 int dd_idx, pd_idx, j;
1779                                 struct stripe_head *sh2;
1780
1781                                 sector_t bn = compute_blocknr(sh, i);
1782                                 sector_t s = raid5_compute_sector(bn, conf->raid_disks,
1783                                                                   conf->raid_disks-1,
1784                                                                   &dd_idx, &pd_idx, conf);
1785                                 sh2 = get_active_stripe(conf, s, conf->raid_disks, pd_idx, 1);
1786                                 if (sh2 == NULL)
1787                                         /* so far only the early blocks of this stripe
1788                                          * have been requested.  When later blocks
1789                                          * get requested, we will try again
1790                                          */
1791                                         continue;
1792                                 if(!test_bit(STRIPE_EXPANDING, &sh2->state) ||
1793                                    test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
1794                                         /* must have already done this block */
1795                                         release_stripe(sh2);
1796                                         continue;
1797                                 }
1798                                 memcpy(page_address(sh2->dev[dd_idx].page),
1799                                        page_address(sh->dev[i].page),
1800                                        STRIPE_SIZE);
1801                                 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
1802                                 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
1803                                 for (j=0; j<conf->raid_disks; j++)
1804                                         if (j != sh2->pd_idx &&
1805                                             !test_bit(R5_Expanded, &sh2->dev[j].flags))
1806                                                 break;
1807                                 if (j == conf->raid_disks) {
1808                                         set_bit(STRIPE_EXPAND_READY, &sh2->state);
1809                                         set_bit(STRIPE_HANDLE, &sh2->state);
1810                                 }
1811                                 release_stripe(sh2);
1812                         }
1813         }
1814
1815         spin_unlock(&sh->lock);
1816
1817         while ((bi=return_bi)) {
1818                 int bytes = bi->bi_size;
1819
1820                 return_bi = bi->bi_next;
1821                 bi->bi_next = NULL;
1822                 bi->bi_size = 0;
1823                 bi->bi_end_io(bi, bytes,
1824                               test_bit(BIO_UPTODATE, &bi->bi_flags)
1825                                 ? 0 : -EIO);
1826         }
1827         for (i=disks; i-- ;) {
1828                 int rw;
1829                 struct bio *bi;
1830                 mdk_rdev_t *rdev;
1831                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
1832                         rw = WRITE;
1833                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1834                         rw = READ;
1835                 else
1836                         continue;
1837  
1838                 bi = &sh->dev[i].req;
1839  
1840                 bi->bi_rw = rw;
1841                 if (rw == WRITE)
1842                         bi->bi_end_io = raid5_end_write_request;
1843                 else
1844                         bi->bi_end_io = raid5_end_read_request;
1845  
1846                 rcu_read_lock();
1847                 rdev = rcu_dereference(conf->disks[i].rdev);
1848                 if (rdev && test_bit(Faulty, &rdev->flags))
1849                         rdev = NULL;
1850                 if (rdev)
1851                         atomic_inc(&rdev->nr_pending);
1852                 rcu_read_unlock();
1853  
1854                 if (rdev) {
1855                         if (syncing || expanding || expanded)
1856                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1857
1858                         bi->bi_bdev = rdev->bdev;
1859                         PRINTK("for %llu schedule op %ld on disc %d\n",
1860                                 (unsigned long long)sh->sector, bi->bi_rw, i);
1861                         atomic_inc(&sh->count);
1862                         bi->bi_sector = sh->sector + rdev->data_offset;
1863                         bi->bi_flags = 1 << BIO_UPTODATE;
1864                         bi->bi_vcnt = 1;        
1865                         bi->bi_max_vecs = 1;
1866                         bi->bi_idx = 0;
1867                         bi->bi_io_vec = &sh->dev[i].vec;
1868                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1869                         bi->bi_io_vec[0].bv_offset = 0;
1870                         bi->bi_size = STRIPE_SIZE;
1871                         bi->bi_next = NULL;
1872                         if (rw == WRITE &&
1873                             test_bit(R5_ReWrite, &sh->dev[i].flags))
1874                                 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1875                         generic_make_request(bi);
1876                 } else {
1877                         if (rw == WRITE)
1878                                 set_bit(STRIPE_DEGRADED, &sh->state);
1879                         PRINTK("skip op %ld on disc %d for sector %llu\n",
1880                                 bi->bi_rw, i, (unsigned long long)sh->sector);
1881                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1882                         set_bit(STRIPE_HANDLE, &sh->state);
1883                 }
1884         }
1885 }
1886
1887 static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
1888 {
1889         raid6_conf_t *conf = sh->raid_conf;
1890         int disks = conf->raid_disks;
1891         struct bio *return_bi= NULL;
1892         struct bio *bi;
1893         int i;
1894         int syncing;
1895         int locked=0, uptodate=0, to_read=0, to_write=0, failed=0, written=0;
1896         int non_overwrite = 0;
1897         int failed_num[2] = {0, 0};
1898         struct r5dev *dev, *pdev, *qdev;
1899         int pd_idx = sh->pd_idx;
1900         int qd_idx = raid6_next_disk(pd_idx, disks);
1901         int p_failed, q_failed;
1902
1903         PRINTK("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d, qd_idx=%d\n",
1904                (unsigned long long)sh->sector, sh->state, atomic_read(&sh->count),
1905                pd_idx, qd_idx);
1906
1907         spin_lock(&sh->lock);
1908         clear_bit(STRIPE_HANDLE, &sh->state);
1909         clear_bit(STRIPE_DELAYED, &sh->state);
1910
1911         syncing = test_bit(STRIPE_SYNCING, &sh->state);
1912         /* Now to look around and see what can be done */
1913
1914         rcu_read_lock();
1915         for (i=disks; i--; ) {
1916                 mdk_rdev_t *rdev;
1917                 dev = &sh->dev[i];
1918                 clear_bit(R5_Insync, &dev->flags);
1919
1920                 PRINTK("check %d: state 0x%lx read %p write %p written %p\n",
1921                         i, dev->flags, dev->toread, dev->towrite, dev->written);
1922                 /* maybe we can reply to a read */
1923                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
1924                         struct bio *rbi, *rbi2;
1925                         PRINTK("Return read for disc %d\n", i);
1926                         spin_lock_irq(&conf->device_lock);
1927                         rbi = dev->toread;
1928                         dev->toread = NULL;
1929                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
1930                                 wake_up(&conf->wait_for_overlap);
1931                         spin_unlock_irq(&conf->device_lock);
1932                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1933                                 copy_data(0, rbi, dev->page, dev->sector);
1934                                 rbi2 = r5_next_bio(rbi, dev->sector);
1935                                 spin_lock_irq(&conf->device_lock);
1936                                 if (--rbi->bi_phys_segments == 0) {
1937                                         rbi->bi_next = return_bi;
1938                                         return_bi = rbi;
1939                                 }
1940                                 spin_unlock_irq(&conf->device_lock);
1941                                 rbi = rbi2;
1942                         }
1943                 }
1944
1945                 /* now count some things */
1946                 if (test_bit(R5_LOCKED, &dev->flags)) locked++;
1947                 if (test_bit(R5_UPTODATE, &dev->flags)) uptodate++;
1948
1949
1950                 if (dev->toread) to_read++;
1951                 if (dev->towrite) {
1952                         to_write++;
1953                         if (!test_bit(R5_OVERWRITE, &dev->flags))
1954                                 non_overwrite++;
1955                 }
1956                 if (dev->written) written++;
1957                 rdev = rcu_dereference(conf->disks[i].rdev);
1958                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
1959                         /* The ReadError flag will just be confusing now */
1960                         clear_bit(R5_ReadError, &dev->flags);
1961                         clear_bit(R5_ReWrite, &dev->flags);
1962                 }
1963                 if (!rdev || !test_bit(In_sync, &rdev->flags)
1964                     || test_bit(R5_ReadError, &dev->flags)) {
1965                         if ( failed < 2 )
1966                                 failed_num[failed] = i;
1967                         failed++;
1968                 } else
1969                         set_bit(R5_Insync, &dev->flags);
1970         }
1971         rcu_read_unlock();
1972         PRINTK("locked=%d uptodate=%d to_read=%d"
1973                " to_write=%d failed=%d failed_num=%d,%d\n",
1974                locked, uptodate, to_read, to_write, failed,
1975                failed_num[0], failed_num[1]);
1976         /* check if the array has lost >2 devices and, if so, some requests might
1977          * need to be failed
1978          */
1979         if (failed > 2 && to_read+to_write+written) {
1980                 for (i=disks; i--; ) {
1981                         int bitmap_end = 0;
1982
1983                         if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1984                                 mdk_rdev_t *rdev;
1985                                 rcu_read_lock();
1986                                 rdev = rcu_dereference(conf->disks[i].rdev);
1987                                 if (rdev && test_bit(In_sync, &rdev->flags))
1988                                         /* multiple read failures in one stripe */
1989                                         md_error(conf->mddev, rdev);
1990                                 rcu_read_unlock();
1991                         }
1992
1993                         spin_lock_irq(&conf->device_lock);
1994                         /* fail all writes first */
1995                         bi = sh->dev[i].towrite;
1996                         sh->dev[i].towrite = NULL;
1997                         if (bi) { to_write--; bitmap_end = 1; }
1998
1999                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2000                                 wake_up(&conf->wait_for_overlap);
2001
2002                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
2003                                 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2004                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2005                                 if (--bi->bi_phys_segments == 0) {
2006                                         md_write_end(conf->mddev);
2007                                         bi->bi_next = return_bi;
2008                                         return_bi = bi;
2009                                 }
2010                                 bi = nextbi;
2011                         }
2012                         /* and fail all 'written' */
2013                         bi = sh->dev[i].written;
2014                         sh->dev[i].written = NULL;
2015                         if (bi) bitmap_end = 1;
2016                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS) {
2017                                 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2018                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2019                                 if (--bi->bi_phys_segments == 0) {
2020                                         md_write_end(conf->mddev);
2021                                         bi->bi_next = return_bi;
2022                                         return_bi = bi;
2023                                 }
2024                                 bi = bi2;
2025                         }
2026
2027                         /* fail any reads if this device is non-operational */
2028                         if (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2029                             test_bit(R5_ReadError, &sh->dev[i].flags)) {
2030                                 bi = sh->dev[i].toread;
2031                                 sh->dev[i].toread = NULL;
2032                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2033                                         wake_up(&conf->wait_for_overlap);
2034                                 if (bi) to_read--;
2035                                 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
2036                                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2037                                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2038                                         if (--bi->bi_phys_segments == 0) {
2039                                                 bi->bi_next = return_bi;
2040                                                 return_bi = bi;
2041                                         }
2042                                         bi = nextbi;
2043                                 }
2044                         }
2045                         spin_unlock_irq(&conf->device_lock);
2046                         if (bitmap_end)
2047                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2048                                                 STRIPE_SECTORS, 0, 0);
2049                 }
2050         }
2051         if (failed > 2 && syncing) {
2052                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2053                 clear_bit(STRIPE_SYNCING, &sh->state);
2054                 syncing = 0;
2055         }
2056
2057         /*
2058          * might be able to return some write requests if the parity blocks
2059          * are safe, or on a failed drive
2060          */
2061         pdev = &sh->dev[pd_idx];
2062         p_failed = (failed >= 1 && failed_num[0] == pd_idx)
2063                 || (failed >= 2 && failed_num[1] == pd_idx);
2064         qdev = &sh->dev[qd_idx];
2065         q_failed = (failed >= 1 && failed_num[0] == qd_idx)
2066                 || (failed >= 2 && failed_num[1] == qd_idx);
2067
2068         if ( written &&
2069              ( p_failed || ((test_bit(R5_Insync, &pdev->flags)
2070                              && !test_bit(R5_LOCKED, &pdev->flags)
2071                              && test_bit(R5_UPTODATE, &pdev->flags))) ) &&
2072              ( q_failed || ((test_bit(R5_Insync, &qdev->flags)
2073                              && !test_bit(R5_LOCKED, &qdev->flags)
2074                              && test_bit(R5_UPTODATE, &qdev->flags))) ) ) {
2075                 /* any written block on an uptodate or failed drive can be
2076                  * returned.  Note that if we 'wrote' to a failed drive,
2077                  * it will be UPTODATE, but never LOCKED, so we don't need
2078                  * to test 'failed' directly.
2079                  */
2080                 for (i=disks; i--; )
2081                         if (sh->dev[i].written) {
2082                                 dev = &sh->dev[i];
2083                                 if (!test_bit(R5_LOCKED, &dev->flags) &&
2084                                     test_bit(R5_UPTODATE, &dev->flags) ) {
2085                                         /* We can return any write requests */
2086                                         int bitmap_end = 0;
2087                                         struct bio *wbi, *wbi2;
2088                                         PRINTK("Return write for stripe %llu disc %d\n",
2089                                                (unsigned long long)sh->sector, i);
2090                                         spin_lock_irq(&conf->device_lock);
2091                                         wbi = dev->written;
2092                                         dev->written = NULL;
2093                                         while (wbi && wbi->bi_sector < dev->sector + STRIPE_SECTORS) {
2094                                                 wbi2 = r5_next_bio(wbi, dev->sector);
2095                                                 if (--wbi->bi_phys_segments == 0) {
2096                                                         md_write_end(conf->mddev);
2097                                                         wbi->bi_next = return_bi;
2098                                                         return_bi = wbi;
2099                                                 }
2100                                                 wbi = wbi2;
2101                                         }
2102                                         if (dev->towrite == NULL)
2103                                                 bitmap_end = 1;
2104                                         spin_unlock_irq(&conf->device_lock);
2105                                         if (bitmap_end)
2106                                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2107                                                                 STRIPE_SECTORS,
2108                                                                 !test_bit(STRIPE_DEGRADED, &sh->state), 0);
2109                                 }
2110                         }
2111         }
2112
2113         /* Now we might consider reading some blocks, either to check/generate
2114          * parity, or to satisfy requests
2115          * or to load a block that is being partially written.
2116          */
2117         if (to_read || non_overwrite || (to_write && failed) || (syncing && (uptodate < disks))) {
2118                 for (i=disks; i--;) {
2119                         dev = &sh->dev[i];
2120                         if (!test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
2121                             (dev->toread ||
2122                              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2123                              syncing ||
2124                              (failed >= 1 && (sh->dev[failed_num[0]].toread || to_write)) ||
2125                              (failed >= 2 && (sh->dev[failed_num[1]].toread || to_write))
2126                                     )
2127                                 ) {
2128                                 /* we would like to get this block, possibly
2129                                  * by computing it, but we might not be able to
2130                                  */
2131                                 if (uptodate == disks-1) {
2132                                         PRINTK("Computing stripe %llu block %d\n",
2133                                                (unsigned long long)sh->sector, i);
2134                                         compute_block_1(sh, i, 0);
2135                                         uptodate++;
2136                                 } else if ( uptodate == disks-2 && failed >= 2 ) {
2137                                         /* Computing 2-failure is *very* expensive; only do it if failed >= 2 */
2138                                         int other;
2139                                         for (other=disks; other--;) {
2140                                                 if ( other == i )
2141                                                         continue;
2142                                                 if ( !test_bit(R5_UPTODATE, &sh->dev[other].flags) )
2143                                                         break;
2144                                         }
2145                                         BUG_ON(other < 0);
2146                                         PRINTK("Computing stripe %llu blocks %d,%d\n",
2147                                                (unsigned long long)sh->sector, i, other);
2148                                         compute_block_2(sh, i, other);
2149                                         uptodate += 2;
2150                                 } else if (test_bit(R5_Insync, &dev->flags)) {
2151                                         set_bit(R5_LOCKED, &dev->flags);
2152                                         set_bit(R5_Wantread, &dev->flags);
2153                                         locked++;
2154                                         PRINTK("Reading block %d (sync=%d)\n",
2155                                                 i, syncing);
2156                                 }
2157                         }
2158                 }
2159                 set_bit(STRIPE_HANDLE, &sh->state);
2160         }
2161
2162         /* now to consider writing and what else, if anything should be read */
2163         if (to_write) {
2164                 int rcw=0, must_compute=0;
2165                 for (i=disks ; i--;) {
2166                         dev = &sh->dev[i];
2167                         /* Would I have to read this buffer for reconstruct_write */
2168                         if (!test_bit(R5_OVERWRITE, &dev->flags)
2169                             && i != pd_idx && i != qd_idx
2170                             && (!test_bit(R5_LOCKED, &dev->flags)
2171                                     ) &&
2172                             !test_bit(R5_UPTODATE, &dev->flags)) {
2173                                 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2174                                 else {
2175                                         PRINTK("raid6: must_compute: disk %d flags=%#lx\n", i, dev->flags);
2176                                         must_compute++;
2177                                 }
2178                         }
2179                 }
2180                 PRINTK("for sector %llu, rcw=%d, must_compute=%d\n",
2181                        (unsigned long long)sh->sector, rcw, must_compute);
2182                 set_bit(STRIPE_HANDLE, &sh->state);
2183
2184                 if (rcw > 0)
2185                         /* want reconstruct write, but need to get some data */
2186                         for (i=disks; i--;) {
2187                                 dev = &sh->dev[i];
2188                                 if (!test_bit(R5_OVERWRITE, &dev->flags)
2189                                     && !(failed == 0 && (i == pd_idx || i == qd_idx))
2190                                     && !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
2191                                     test_bit(R5_Insync, &dev->flags)) {
2192                                         if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
2193                                         {
2194                                                 PRINTK("Read_old stripe %llu block %d for Reconstruct\n",
2195                                                        (unsigned long long)sh->sector, i);
2196                                                 set_bit(R5_LOCKED, &dev->flags);
2197                                                 set_bit(R5_Wantread, &dev->flags);
2198                                                 locked++;
2199                                         } else {
2200                                                 PRINTK("Request delayed stripe %llu block %d for Reconstruct\n",
2201                                                        (unsigned long long)sh->sector, i);
2202                                                 set_bit(STRIPE_DELAYED, &sh->state);
2203                                                 set_bit(STRIPE_HANDLE, &sh->state);
2204                                         }
2205                                 }
2206                         }
2207                 /* now if nothing is locked, and if we have enough data, we can start a write request */
2208                 if (locked == 0 && rcw == 0 &&
2209                     !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2210                         if ( must_compute > 0 ) {
2211                                 /* We have failed blocks and need to compute them */
2212                                 switch ( failed ) {
2213                                 case 0: BUG();
2214                                 case 1: compute_block_1(sh, failed_num[0], 0); break;
2215                                 case 2: compute_block_2(sh, failed_num[0], failed_num[1]); break;
2216                                 default: BUG(); /* This request should have been failed? */
2217                                 }
2218                         }
2219
2220                         PRINTK("Computing parity for stripe %llu\n", (unsigned long long)sh->sector);
2221                         compute_parity6(sh, RECONSTRUCT_WRITE);
2222                         /* now every locked buffer is ready to be written */
2223                         for (i=disks; i--;)
2224                                 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2225                                         PRINTK("Writing stripe %llu block %d\n",
2226                                                (unsigned long long)sh->sector, i);
2227                                         locked++;
2228                                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
2229                                 }
2230                         /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2231                         set_bit(STRIPE_INSYNC, &sh->state);
2232
2233                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2234                                 atomic_dec(&conf->preread_active_stripes);
2235                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
2236                                         md_wakeup_thread(conf->mddev->thread);
2237                         }
2238                 }
2239         }
2240
2241         /* maybe we need to check and possibly fix the parity for this stripe
2242          * Any reads will already have been scheduled, so we just see if enough data
2243          * is available
2244          */
2245         if (syncing && locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state)) {
2246                 int update_p = 0, update_q = 0;
2247                 struct r5dev *dev;
2248
2249                 set_bit(STRIPE_HANDLE, &sh->state);
2250
2251                 BUG_ON(failed>2);
2252                 BUG_ON(uptodate < disks);
2253                 /* Want to check and possibly repair P and Q.
2254                  * However there could be one 'failed' device, in which
2255                  * case we can only check one of them, possibly using the
2256                  * other to generate missing data
2257                  */
2258
2259                 /* If !tmp_page, we cannot do the calculations,
2260                  * but as we have set STRIPE_HANDLE, we will soon be called
2261                  * by stripe_handle with a tmp_page - just wait until then.
2262                  */
2263                 if (tmp_page) {
2264                         if (failed == q_failed) {
2265                                 /* The only possible failed device holds 'Q', so it makes
2266                                  * sense to check P (If anything else were failed, we would
2267                                  * have used P to recreate it).
2268                                  */
2269                                 compute_block_1(sh, pd_idx, 1);
2270                                 if (!page_is_zero(sh->dev[pd_idx].page)) {
2271                                         compute_block_1(sh,pd_idx,0);
2272                                         update_p = 1;
2273                                 }
2274                         }
2275                         if (!q_failed && failed < 2) {
2276                                 /* q is not failed, and we didn't use it to generate
2277                                  * anything, so it makes sense to check it
2278                                  */
2279                                 memcpy(page_address(tmp_page),
2280                                        page_address(sh->dev[qd_idx].page),
2281                                        STRIPE_SIZE);
2282                                 compute_parity6(sh, UPDATE_PARITY);
2283                                 if (memcmp(page_address(tmp_page),
2284                                            page_address(sh->dev[qd_idx].page),
2285                                            STRIPE_SIZE)!= 0) {
2286                                         clear_bit(STRIPE_INSYNC, &sh->state);
2287                                         update_q = 1;
2288                                 }
2289                         }
2290                         if (update_p || update_q) {
2291                                 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2292                                 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2293                                         /* don't try to repair!! */
2294                                         update_p = update_q = 0;
2295                         }
2296
2297                         /* now write out any block on a failed drive,
2298                          * or P or Q if they need it
2299                          */
2300
2301                         if (failed == 2) {
2302                                 dev = &sh->dev[failed_num[1]];
2303                                 locked++;
2304                                 set_bit(R5_LOCKED, &dev->flags);
2305                                 set_bit(R5_Wantwrite, &dev->flags);
2306                         }
2307                         if (failed >= 1) {
2308                                 dev = &sh->dev[failed_num[0]];
2309                                 locked++;
2310                                 set_bit(R5_LOCKED, &dev->flags);
2311                                 set_bit(R5_Wantwrite, &dev->flags);
2312                         }
2313
2314                         if (update_p) {
2315                                 dev = &sh->dev[pd_idx];
2316                                 locked ++;
2317                                 set_bit(R5_LOCKED, &dev->flags);
2318                                 set_bit(R5_Wantwrite, &dev->flags);
2319                         }
2320                         if (update_q) {
2321                                 dev = &sh->dev[qd_idx];
2322                                 locked++;
2323                                 set_bit(R5_LOCKED, &dev->flags);
2324                                 set_bit(R5_Wantwrite, &dev->flags);
2325                         }
2326                         clear_bit(STRIPE_DEGRADED, &sh->state);
2327
2328                         set_bit(STRIPE_INSYNC, &sh->state);
2329                 }
2330         }
2331
2332         if (syncing && locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2333                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2334                 clear_bit(STRIPE_SYNCING, &sh->state);
2335         }
2336
2337         /* If the failed drives are just a ReadError, then we might need
2338          * to progress the repair/check process
2339          */
2340         if (failed <= 2 && ! conf->mddev->ro)
2341                 for (i=0; i<failed;i++) {
2342                         dev = &sh->dev[failed_num[i]];
2343                         if (test_bit(R5_ReadError, &dev->flags)
2344                             && !test_bit(R5_LOCKED, &dev->flags)
2345                             && test_bit(R5_UPTODATE, &dev->flags)
2346                                 ) {
2347                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
2348                                         set_bit(R5_Wantwrite, &dev->flags);
2349                                         set_bit(R5_ReWrite, &dev->flags);
2350                                         set_bit(R5_LOCKED, &dev->flags);
2351                                 } else {
2352                                         /* let's read it back */
2353                                         set_bit(R5_Wantread, &dev->flags);
2354                                         set_bit(R5_LOCKED, &dev->flags);
2355                                 }
2356                         }
2357                 }
2358         spin_unlock(&sh->lock);
2359
2360         while ((bi=return_bi)) {
2361                 int bytes = bi->bi_size;
2362
2363                 return_bi = bi->bi_next;
2364                 bi->bi_next = NULL;
2365                 bi->bi_size = 0;
2366                 bi->bi_end_io(bi, bytes,
2367                               test_bit(BIO_UPTODATE, &bi->bi_flags)
2368                                 ? 0 : -EIO);
2369         }
2370         for (i=disks; i-- ;) {
2371                 int rw;
2372                 struct bio *bi;
2373                 mdk_rdev_t *rdev;
2374                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
2375                         rw = WRITE;
2376                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
2377                         rw = READ;
2378                 else
2379                         continue;
2380
2381                 bi = &sh->dev[i].req;
2382
2383                 bi->bi_rw = rw;
2384                 if (rw == WRITE)
2385                         bi->bi_end_io = raid5_end_write_request;
2386                 else
2387                         bi->bi_end_io = raid5_end_read_request;
2388
2389                 rcu_read_lock();
2390                 rdev = rcu_dereference(conf->disks[i].rdev);
2391                 if (rdev && test_bit(Faulty, &rdev->flags))
2392                         rdev = NULL;
2393                 if (rdev)
2394                         atomic_inc(&rdev->nr_pending);
2395                 rcu_read_unlock();
2396
2397                 if (rdev) {
2398                         if (syncing)
2399                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
2400
2401                         bi->bi_bdev = rdev->bdev;
2402                         PRINTK("for %llu schedule op %ld on disc %d\n",
2403                                 (unsigned long long)sh->sector, bi->bi_rw, i);
2404                         atomic_inc(&sh->count);
2405                         bi->bi_sector = sh->sector + rdev->data_offset;
2406                         bi->bi_flags = 1 << BIO_UPTODATE;
2407                         bi->bi_vcnt = 1;
2408                         bi->bi_max_vecs = 1;
2409                         bi->bi_idx = 0;
2410                         bi->bi_io_vec = &sh->dev[i].vec;
2411                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
2412                         bi->bi_io_vec[0].bv_offset = 0;
2413                         bi->bi_size = STRIPE_SIZE;
2414                         bi->bi_next = NULL;
2415                         if (rw == WRITE &&
2416                             test_bit(R5_ReWrite, &sh->dev[i].flags))
2417                                 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2418                         generic_make_request(bi);
2419                 } else {
2420                         if (rw == WRITE)
2421                                 set_bit(STRIPE_DEGRADED, &sh->state);
2422                         PRINTK("skip op %ld on disc %d for sector %llu\n",
2423                                 bi->bi_rw, i, (unsigned long long)sh->sector);
2424                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
2425                         set_bit(STRIPE_HANDLE, &sh->state);
2426                 }
2427         }
2428 }
2429
2430 static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
2431 {
2432         if (sh->raid_conf->level == 6)
2433                 handle_stripe6(sh, tmp_page);
2434         else
2435                 handle_stripe5(sh);
2436 }
2437
2438
2439
2440 static void raid5_activate_delayed(raid5_conf_t *conf)
2441 {
2442         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
2443                 while (!list_empty(&conf->delayed_list)) {
2444                         struct list_head *l = conf->delayed_list.next;
2445                         struct stripe_head *sh;
2446                         sh = list_entry(l, struct stripe_head, lru);
2447                         list_del_init(l);
2448                         clear_bit(STRIPE_DELAYED, &sh->state);
2449                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
2450                                 atomic_inc(&conf->preread_active_stripes);
2451                         list_add_tail(&sh->lru, &conf->handle_list);
2452                 }
2453         }
2454 }
2455
2456 static void activate_bit_delay(raid5_conf_t *conf)
2457 {
2458         /* device_lock is held */
2459         struct list_head head;
2460         list_add(&head, &conf->bitmap_list);
2461         list_del_init(&conf->bitmap_list);
2462         while (!list_empty(&head)) {
2463                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
2464                 list_del_init(&sh->lru);
2465                 atomic_inc(&sh->count);
2466                 __release_stripe(conf, sh);
2467         }
2468 }
2469
2470 static void unplug_slaves(mddev_t *mddev)
2471 {
2472         raid5_conf_t *conf = mddev_to_conf(mddev);
2473         int i;
2474
2475         rcu_read_lock();
2476         for (i=0; i<mddev->raid_disks; i++) {
2477                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
2478                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
2479                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
2480
2481                         atomic_inc(&rdev->nr_pending);
2482                         rcu_read_unlock();
2483
2484                         if (r_queue->unplug_fn)
2485                                 r_queue->unplug_fn(r_queue);
2486
2487                         rdev_dec_pending(rdev, mddev);
2488                         rcu_read_lock();
2489                 }
2490         }
2491         rcu_read_unlock();
2492 }
2493
2494 static void raid5_unplug_device(request_queue_t *q)
2495 {
2496         mddev_t *mddev = q->queuedata;
2497         raid5_conf_t *conf = mddev_to_conf(mddev);
2498         unsigned long flags;
2499
2500         spin_lock_irqsave(&conf->device_lock, flags);
2501
2502         if (blk_remove_plug(q)) {
2503                 conf->seq_flush++;
2504                 raid5_activate_delayed(conf);
2505         }
2506         md_wakeup_thread(mddev->thread);
2507
2508         spin_unlock_irqrestore(&conf->device_lock, flags);
2509
2510         unplug_slaves(mddev);
2511 }
2512
2513 static int raid5_issue_flush(request_queue_t *q, struct gendisk *disk,
2514                              sector_t *error_sector)
2515 {
2516         mddev_t *mddev = q->queuedata;
2517         raid5_conf_t *conf = mddev_to_conf(mddev);
2518         int i, ret = 0;
2519
2520         rcu_read_lock();
2521         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
2522                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
2523                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
2524                         struct block_device *bdev = rdev->bdev;
2525                         request_queue_t *r_queue = bdev_get_queue(bdev);
2526
2527                         if (!r_queue->issue_flush_fn)
2528                                 ret = -EOPNOTSUPP;
2529                         else {
2530                                 atomic_inc(&rdev->nr_pending);
2531                                 rcu_read_unlock();
2532                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
2533                                                               error_sector);
2534                                 rdev_dec_pending(rdev, mddev);
2535                                 rcu_read_lock();
2536                         }
2537                 }
2538         }
2539         rcu_read_unlock();
2540         return ret;
2541 }
2542
2543 static int raid5_congested(void *data, int bits)
2544 {
2545         mddev_t *mddev = data;
2546         raid5_conf_t *conf = mddev_to_conf(mddev);
2547
2548         /* No difference between reads and writes.  Just check
2549          * how busy the stripe_cache is
2550          */
2551         if (conf->inactive_blocked)
2552                 return 1;
2553         if (conf->quiesce)
2554                 return 1;
2555         if (list_empty_careful(&conf->inactive_list))
2556                 return 1;
2557
2558         return 0;
2559 }
2560
2561 /* We want read requests to align with chunks where possible,
2562  * but write requests don't need to.
2563  */
2564 static int raid5_mergeable_bvec(request_queue_t *q, struct bio *bio, struct bio_vec *biovec)
2565 {
2566         mddev_t *mddev = q->queuedata;
2567         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
2568         int max;
2569         unsigned int chunk_sectors = mddev->chunk_size >> 9;
2570         unsigned int bio_sectors = bio->bi_size >> 9;
2571
2572         if (bio_data_dir(bio) == WRITE)
2573                 return biovec->bv_len; /* always allow writes to be mergeable */
2574
2575         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
2576         if (max < 0) max = 0;
2577         if (max <= biovec->bv_len && bio_sectors == 0)
2578                 return biovec->bv_len;
2579         else
2580                 return max;
2581 }
2582
2583
2584 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
2585 {
2586         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
2587         unsigned int chunk_sectors = mddev->chunk_size >> 9;
2588         unsigned int bio_sectors = bio->bi_size >> 9;
2589
2590         return  chunk_sectors >=
2591                 ((sector & (chunk_sectors - 1)) + bio_sectors);
2592 }
2593
2594 /*
2595  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
2596  *  later sampled by raid5d.
2597  */
2598 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
2599 {
2600         unsigned long flags;
2601
2602         spin_lock_irqsave(&conf->device_lock, flags);
2603
2604         bi->bi_next = conf->retry_read_aligned_list;
2605         conf->retry_read_aligned_list = bi;
2606
2607         spin_unlock_irqrestore(&conf->device_lock, flags);
2608         md_wakeup_thread(conf->mddev->thread);
2609 }
2610
2611
2612 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
2613 {
2614         struct bio *bi;
2615
2616         bi = conf->retry_read_aligned;
2617         if (bi) {
2618                 conf->retry_read_aligned = NULL;
2619                 return bi;
2620         }
2621         bi = conf->retry_read_aligned_list;
2622         if(bi) {
2623                 conf->retry_read_aligned = bi->bi_next;
2624                 bi->bi_next = NULL;
2625                 bi->bi_phys_segments = 1; /* biased count of active stripes */
2626                 bi->bi_hw_segments = 0; /* count of processed stripes */
2627         }
2628
2629         return bi;
2630 }
2631
2632
2633 /*
2634  *  The "raid5_align_endio" should check if the read succeeded and if it
2635  *  did, call bio_endio on the original bio (having bio_put the new bio
2636  *  first).
2637  *  If the read failed..
2638  */
2639 static int raid5_align_endio(struct bio *bi, unsigned int bytes, int error)
2640 {
2641         struct bio* raid_bi  = bi->bi_private;
2642         mddev_t *mddev;
2643         raid5_conf_t *conf;
2644         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
2645         mdk_rdev_t *rdev;
2646
2647         if (bi->bi_size)
2648                 return 1;
2649         bio_put(bi);
2650
2651         mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
2652         conf = mddev_to_conf(mddev);
2653         rdev = (void*)raid_bi->bi_next;
2654         raid_bi->bi_next = NULL;
2655
2656         rdev_dec_pending(rdev, conf->mddev);
2657
2658         if (!error && uptodate) {
2659                 bio_endio(raid_bi, bytes, 0);
2660                 if (atomic_dec_and_test(&conf->active_aligned_reads))
2661                         wake_up(&conf->wait_for_stripe);
2662                 return 0;
2663         }
2664
2665
2666         PRINTK("raid5_align_endio : io error...handing IO for a retry\n");
2667
2668         add_bio_to_retry(raid_bi, conf);
2669         return 0;
2670 }
2671
2672 static int chunk_aligned_read(request_queue_t *q, struct bio * raid_bio)
2673 {
2674         mddev_t *mddev = q->queuedata;
2675         raid5_conf_t *conf = mddev_to_conf(mddev);
2676         const unsigned int raid_disks = conf->raid_disks;
2677         const unsigned int data_disks = raid_disks - conf->max_degraded;
2678         unsigned int dd_idx, pd_idx;
2679         struct bio* align_bi;
2680         mdk_rdev_t *rdev;
2681
2682         if (!in_chunk_boundary(mddev, raid_bio)) {
2683                 printk("chunk_aligned_read : non aligned\n");
2684                 return 0;
2685         }
2686         /*
2687          * use bio_clone to make a copy of the bio
2688          */
2689         align_bi = bio_clone(raid_bio, GFP_NOIO);
2690         if (!align_bi)
2691                 return 0;
2692         /*
2693          *   set bi_end_io to a new function, and set bi_private to the
2694          *     original bio.
2695          */
2696         align_bi->bi_end_io  = raid5_align_endio;
2697         align_bi->bi_private = raid_bio;
2698         /*
2699          *      compute position
2700          */
2701         align_bi->bi_sector =  raid5_compute_sector(raid_bio->bi_sector,
2702                                         raid_disks,
2703                                         data_disks,
2704                                         &dd_idx,
2705                                         &pd_idx,
2706                                         conf);
2707
2708         rcu_read_lock();
2709         rdev = rcu_dereference(conf->disks[dd_idx].rdev);
2710         if (rdev && test_bit(In_sync, &rdev->flags)) {
2711                 atomic_inc(&rdev->nr_pending);
2712                 rcu_read_unlock();
2713                 raid_bio->bi_next = (void*)rdev;
2714                 align_bi->bi_bdev =  rdev->bdev;
2715                 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
2716                 align_bi->bi_sector += rdev->data_offset;
2717
2718                 spin_lock_irq(&conf->device_lock);
2719                 wait_event_lock_irq(conf->wait_for_stripe,
2720                                     conf->quiesce == 0,
2721                                     conf->device_lock, /* nothing */);
2722                 atomic_inc(&conf->active_aligned_reads);
2723                 spin_unlock_irq(&conf->device_lock);
2724
2725                 generic_make_request(align_bi);
2726                 return 1;
2727         } else {
2728                 rcu_read_unlock();
2729                 bio_put(align_bi);
2730                 return 0;
2731         }
2732 }
2733
2734
2735 static int make_request(request_queue_t *q, struct bio * bi)
2736 {
2737         mddev_t *mddev = q->queuedata;
2738         raid5_conf_t *conf = mddev_to_conf(mddev);
2739         unsigned int dd_idx, pd_idx;
2740         sector_t new_sector;
2741         sector_t logical_sector, last_sector;
2742         struct stripe_head *sh;
2743         const int rw = bio_data_dir(bi);
2744         int remaining;
2745
2746         if (unlikely(bio_barrier(bi))) {
2747                 bio_endio(bi, bi->bi_size, -EOPNOTSUPP);
2748                 return 0;
2749         }
2750
2751         md_write_start(mddev, bi);
2752
2753         disk_stat_inc(mddev->gendisk, ios[rw]);
2754         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
2755
2756         if (rw == READ &&
2757              mddev->reshape_position == MaxSector &&
2758              chunk_aligned_read(q,bi))
2759                 return 0;
2760
2761         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
2762         last_sector = bi->bi_sector + (bi->bi_size>>9);
2763         bi->bi_next = NULL;
2764         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
2765
2766         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
2767                 DEFINE_WAIT(w);
2768                 int disks, data_disks;
2769
2770         retry:
2771                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
2772                 if (likely(conf->expand_progress == MaxSector))
2773                         disks = conf->raid_disks;
2774                 else {
2775                         /* spinlock is needed as expand_progress may be
2776                          * 64bit on a 32bit platform, and so it might be
2777                          * possible to see a half-updated value
2778                          * Ofcourse expand_progress could change after
2779                          * the lock is dropped, so once we get a reference
2780                          * to the stripe that we think it is, we will have
2781                          * to check again.
2782                          */
2783                         spin_lock_irq(&conf->device_lock);
2784                         disks = conf->raid_disks;
2785                         if (logical_sector >= conf->expand_progress)
2786                                 disks = conf->previous_raid_disks;
2787                         else {
2788                                 if (logical_sector >= conf->expand_lo) {
2789                                         spin_unlock_irq(&conf->device_lock);
2790                                         schedule();
2791                                         goto retry;
2792                                 }
2793                         }
2794                         spin_unlock_irq(&conf->device_lock);
2795                 }
2796                 data_disks = disks - conf->max_degraded;
2797
2798                 new_sector = raid5_compute_sector(logical_sector, disks, data_disks,
2799                                                   &dd_idx, &pd_idx, conf);
2800                 PRINTK("raid5: make_request, sector %llu logical %llu\n",
2801                         (unsigned long long)new_sector, 
2802                         (unsigned long long)logical_sector);
2803
2804                 sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
2805                 if (sh) {
2806                         if (unlikely(conf->expand_progress != MaxSector)) {
2807                                 /* expansion might have moved on while waiting for a
2808                                  * stripe, so we must do the range check again.
2809                                  * Expansion could still move past after this
2810                                  * test, but as we are holding a reference to
2811                                  * 'sh', we know that if that happens,
2812                                  *  STRIPE_EXPANDING will get set and the expansion
2813                                  * won't proceed until we finish with the stripe.
2814                                  */
2815                                 int must_retry = 0;
2816                                 spin_lock_irq(&conf->device_lock);
2817                                 if (logical_sector <  conf->expand_progress &&
2818                                     disks == conf->previous_raid_disks)
2819                                         /* mismatch, need to try again */
2820                                         must_retry = 1;
2821                                 spin_unlock_irq(&conf->device_lock);
2822                                 if (must_retry) {
2823                                         release_stripe(sh);
2824                                         goto retry;
2825                                 }
2826                         }
2827                         /* FIXME what if we get a false positive because these
2828                          * are being updated.
2829                          */
2830                         if (logical_sector >= mddev->suspend_lo &&
2831                             logical_sector < mddev->suspend_hi) {
2832                                 release_stripe(sh);
2833                                 schedule();
2834                                 goto retry;
2835                         }
2836
2837                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
2838                             !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
2839                                 /* Stripe is busy expanding or
2840                                  * add failed due to overlap.  Flush everything
2841                                  * and wait a while
2842                                  */
2843                                 raid5_unplug_device(mddev->queue);
2844                                 release_stripe(sh);
2845                                 schedule();
2846                                 goto retry;
2847                         }
2848                         finish_wait(&conf->wait_for_overlap, &w);
2849                         handle_stripe(sh, NULL);
2850                         release_stripe(sh);
2851                 } else {
2852                         /* cannot get stripe for read-ahead, just give-up */
2853                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2854                         finish_wait(&conf->wait_for_overlap, &w);
2855                         break;
2856                 }
2857                         
2858         }
2859         spin_lock_irq(&conf->device_lock);
2860         remaining = --bi->bi_phys_segments;
2861         spin_unlock_irq(&conf->device_lock);
2862         if (remaining == 0) {
2863                 int bytes = bi->bi_size;
2864
2865                 if ( rw == WRITE )
2866                         md_write_end(mddev);
2867                 bi->bi_size = 0;
2868                 bi->bi_end_io(bi, bytes,
2869                               test_bit(BIO_UPTODATE, &bi->bi_flags)
2870                                 ? 0 : -EIO);
2871         }
2872         return 0;
2873 }
2874
2875 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
2876 {
2877         /* reshaping is quite different to recovery/resync so it is
2878          * handled quite separately ... here.
2879          *
2880          * On each call to sync_request, we gather one chunk worth of
2881          * destination stripes and flag them as expanding.
2882          * Then we find all the source stripes and request reads.
2883          * As the reads complete, handle_stripe will copy the data
2884          * into the destination stripe and release that stripe.
2885          */
2886         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
2887         struct stripe_head *sh;
2888         int pd_idx;
2889         sector_t first_sector, last_sector;
2890         int raid_disks;
2891         int data_disks;
2892         int i;
2893         int dd_idx;
2894         sector_t writepos, safepos, gap;
2895
2896         if (sector_nr == 0 &&
2897             conf->expand_progress != 0) {
2898                 /* restarting in the middle, skip the initial sectors */
2899                 sector_nr = conf->expand_progress;
2900                 sector_div(sector_nr, conf->raid_disks-1);
2901                 *skipped = 1;
2902                 return sector_nr;
2903         }
2904
2905         /* we update the metadata when there is more than 3Meg
2906          * in the block range (that is rather arbitrary, should
2907          * probably be time based) or when the data about to be
2908          * copied would over-write the source of the data at
2909          * the front of the range.
2910          * i.e. one new_stripe forward from expand_progress new_maps
2911          * to after where expand_lo old_maps to
2912          */
2913         writepos = conf->expand_progress +
2914                 conf->chunk_size/512*(conf->raid_disks-1);
2915         sector_div(writepos, conf->raid_disks-1);
2916         safepos = conf->expand_lo;
2917         sector_div(safepos, conf->previous_raid_disks-1);
2918         gap = conf->expand_progress - conf->expand_lo;
2919
2920         if (writepos >= safepos ||
2921             gap > (conf->raid_disks-1)*3000*2 /*3Meg*/) {
2922                 /* Cannot proceed until we've updated the superblock... */
2923                 wait_event(conf->wait_for_overlap,
2924                            atomic_read(&conf->reshape_stripes)==0);
2925                 mddev->reshape_position = conf->expand_progress;
2926                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2927                 md_wakeup_thread(mddev->thread);
2928                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
2929                            kthread_should_stop());
2930                 spin_lock_irq(&conf->device_lock);
2931                 conf->expand_lo = mddev->reshape_position;
2932                 spin_unlock_irq(&conf->device_lock);
2933                 wake_up(&conf->wait_for_overlap);
2934         }
2935
2936         for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
2937                 int j;
2938                 int skipped = 0;
2939                 pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks);
2940                 sh = get_active_stripe(conf, sector_nr+i,
2941                                        conf->raid_disks, pd_idx, 0);
2942                 set_bit(STRIPE_EXPANDING, &sh->state);
2943                 atomic_inc(&conf->reshape_stripes);
2944                 /* If any of this stripe is beyond the end of the old
2945                  * array, then we need to zero those blocks
2946                  */
2947                 for (j=sh->disks; j--;) {
2948                         sector_t s;
2949                         if (j == sh->pd_idx)
2950                                 continue;
2951                         s = compute_blocknr(sh, j);
2952                         if (s < (mddev->array_size<<1)) {
2953                                 skipped = 1;
2954                                 continue;
2955                         }
2956                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
2957                         set_bit(R5_Expanded, &sh->dev[j].flags);
2958                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
2959                 }
2960                 if (!skipped) {
2961                         set_bit(STRIPE_EXPAND_READY, &sh->state);
2962                         set_bit(STRIPE_HANDLE, &sh->state);
2963                 }
2964                 release_stripe(sh);
2965         }
2966         spin_lock_irq(&conf->device_lock);
2967         conf->expand_progress = (sector_nr + i)*(conf->raid_disks-1);
2968         spin_unlock_irq(&conf->device_lock);
2969         /* Ok, those stripe are ready. We can start scheduling
2970          * reads on the source stripes.
2971          * The source stripes are determined by mapping the first and last
2972          * block on the destination stripes.
2973          */
2974         raid_disks = conf->previous_raid_disks;
2975         data_disks = raid_disks - 1;
2976         first_sector =
2977                 raid5_compute_sector(sector_nr*(conf->raid_disks-1),
2978                                      raid_disks, data_disks,
2979                                      &dd_idx, &pd_idx, conf);
2980         last_sector =
2981                 raid5_compute_sector((sector_nr+conf->chunk_size/512)
2982                                      *(conf->raid_disks-1) -1,
2983                                      raid_disks, data_disks,
2984                                      &dd_idx, &pd_idx, conf);
2985         if (last_sector >= (mddev->size<<1))
2986                 last_sector = (mddev->size<<1)-1;
2987         while (first_sector <= last_sector) {
2988                 pd_idx = stripe_to_pdidx(first_sector, conf, conf->previous_raid_disks);
2989                 sh = get_active_stripe(conf, first_sector,
2990                                        conf->previous_raid_disks, pd_idx, 0);
2991                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2992                 set_bit(STRIPE_HANDLE, &sh->state);
2993                 release_stripe(sh);
2994                 first_sector += STRIPE_SECTORS;
2995         }
2996         return conf->chunk_size>>9;
2997 }
2998
2999 /* FIXME go_faster isn't used */
3000 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3001 {
3002         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3003         struct stripe_head *sh;
3004         int pd_idx;
3005         int raid_disks = conf->raid_disks;
3006         sector_t max_sector = mddev->size << 1;
3007         int sync_blocks;
3008         int still_degraded = 0;
3009         int i;
3010
3011         if (sector_nr >= max_sector) {
3012                 /* just being told to finish up .. nothing much to do */
3013                 unplug_slaves(mddev);
3014                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3015                         end_reshape(conf);
3016                         return 0;
3017                 }
3018
3019                 if (mddev->curr_resync < max_sector) /* aborted */
3020                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3021                                         &sync_blocks, 1);
3022                 else /* completed sync */
3023                         conf->fullsync = 0;
3024                 bitmap_close_sync(mddev->bitmap);
3025
3026                 return 0;
3027         }
3028
3029         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3030                 return reshape_request(mddev, sector_nr, skipped);
3031
3032         /* if there is too many failed drives and we are trying
3033          * to resync, then assert that we are finished, because there is
3034          * nothing we can do.
3035          */
3036         if (mddev->degraded >= conf->max_degraded &&
3037             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3038                 sector_t rv = (mddev->size << 1) - sector_nr;
3039                 *skipped = 1;
3040                 return rv;
3041         }
3042         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3043             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3044             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3045                 /* we can skip this block, and probably more */
3046                 sync_blocks /= STRIPE_SECTORS;
3047                 *skipped = 1;
3048                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3049         }
3050
3051         pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks);
3052         sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1);
3053         if (sh == NULL) {
3054                 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0);
3055                 /* make sure we don't swamp the stripe cache if someone else
3056                  * is trying to get access
3057                  */
3058                 schedule_timeout_uninterruptible(1);
3059         }
3060         /* Need to check if array will still be degraded after recovery/resync
3061          * We don't need to check the 'failed' flag as when that gets set,
3062          * recovery aborts.
3063          */
3064         for (i=0; i<mddev->raid_disks; i++)
3065                 if (conf->disks[i].rdev == NULL)
3066                         still_degraded = 1;
3067
3068         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
3069
3070         spin_lock(&sh->lock);
3071         set_bit(STRIPE_SYNCING, &sh->state);
3072         clear_bit(STRIPE_INSYNC, &sh->state);
3073         spin_unlock(&sh->lock);
3074
3075         handle_stripe(sh, NULL);
3076         release_stripe(sh);
3077
3078         return STRIPE_SECTORS;
3079 }
3080
3081 static int  retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
3082 {
3083         /* We may not be able to submit a whole bio at once as there
3084          * may not be enough stripe_heads available.
3085          * We cannot pre-allocate enough stripe_heads as we may need
3086          * more than exist in the cache (if we allow ever large chunks).
3087          * So we do one stripe head at a time and record in
3088          * ->bi_hw_segments how many have been done.
3089          *
3090          * We *know* that this entire raid_bio is in one chunk, so
3091          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3092          */
3093         struct stripe_head *sh;
3094         int dd_idx, pd_idx;
3095         sector_t sector, logical_sector, last_sector;
3096         int scnt = 0;
3097         int remaining;
3098         int handled = 0;
3099
3100         logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3101         sector = raid5_compute_sector(  logical_sector,
3102                                         conf->raid_disks,
3103                                         conf->raid_disks - conf->max_degraded,
3104                                         &dd_idx,
3105                                         &pd_idx,
3106                                         conf);
3107         last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
3108
3109         for (; logical_sector < last_sector;
3110              logical_sector += STRIPE_SECTORS, scnt++) {
3111
3112                 if (scnt < raid_bio->bi_hw_segments)
3113                         /* already done this stripe */
3114                         continue;
3115
3116                 sh = get_active_stripe(conf, sector, conf->raid_disks, pd_idx, 1);
3117
3118                 if (!sh) {
3119                         /* failed to get a stripe - must wait */
3120                         raid_bio->bi_hw_segments = scnt;
3121                         conf->retry_read_aligned = raid_bio;
3122                         return handled;
3123                 }
3124
3125                 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
3126                 add_stripe_bio(sh, raid_bio, dd_idx, 0);
3127                 handle_stripe(sh, NULL);
3128                 release_stripe(sh);
3129                 handled++;
3130         }
3131         spin_lock_irq(&conf->device_lock);
3132         remaining = --raid_bio->bi_phys_segments;
3133         spin_unlock_irq(&conf->device_lock);
3134         if (remaining == 0) {
3135                 int bytes = raid_bio->bi_size;
3136
3137                 raid_bio->bi_size = 0;
3138                 raid_bio->bi_end_io(raid_bio, bytes,
3139                               test_bit(BIO_UPTODATE, &raid_bio->bi_flags)
3140                                 ? 0 : -EIO);
3141         }
3142         if (atomic_dec_and_test(&conf->active_aligned_reads))
3143                 wake_up(&conf->wait_for_stripe);
3144         return handled;
3145 }
3146
3147
3148
3149 /*
3150  * This is our raid5 kernel thread.
3151  *
3152  * We scan the hash table for stripes which can be handled now.
3153  * During the scan, completed stripes are saved for us by the interrupt
3154  * handler, so that they will not have to wait for our next wakeup.
3155  */
3156 static void raid5d (mddev_t *mddev)
3157 {
3158         struct stripe_head *sh;
3159         raid5_conf_t *conf = mddev_to_conf(mddev);
3160         int handled;
3161
3162         PRINTK("+++ raid5d active\n");
3163
3164         md_check_recovery(mddev);
3165
3166         handled = 0;
3167         spin_lock_irq(&conf->device_lock);
3168         while (1) {
3169                 struct list_head *first;
3170                 struct bio *bio;
3171
3172                 if (conf->seq_flush != conf->seq_write) {
3173                         int seq = conf->seq_flush;
3174                         spin_unlock_irq(&conf->device_lock);
3175                         bitmap_unplug(mddev->bitmap);
3176                         spin_lock_irq(&conf->device_lock);
3177                         conf->seq_write = seq;
3178                         activate_bit_delay(conf);
3179                 }
3180
3181                 if (list_empty(&conf->handle_list) &&
3182                     atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD &&
3183                     !blk_queue_plugged(mddev->queue) &&
3184                     !list_empty(&conf->delayed_list))
3185                         raid5_activate_delayed(conf);
3186
3187                 while ((bio = remove_bio_from_retry(conf))) {
3188                         int ok;
3189                         spin_unlock_irq(&conf->device_lock);
3190                         ok = retry_aligned_read(conf, bio);
3191                         spin_lock_irq(&conf->device_lock);
3192                         if (!ok)
3193                                 break;
3194                         handled++;
3195                 }
3196
3197                 if (list_empty(&conf->handle_list))
3198                         break;
3199
3200                 first = conf->handle_list.next;
3201                 sh = list_entry(first, struct stripe_head, lru);
3202
3203                 list_del_init(first);
3204                 atomic_inc(&sh->count);
3205                 BUG_ON(atomic_read(&sh->count)!= 1);
3206                 spin_unlock_irq(&conf->device_lock);
3207                 
3208                 handled++;
3209                 handle_stripe(sh, conf->spare_page);
3210                 release_stripe(sh);
3211
3212                 spin_lock_irq(&conf->device_lock);
3213         }
3214         PRINTK("%d stripes handled\n", handled);
3215
3216         spin_unlock_irq(&conf->device_lock);
3217
3218         unplug_slaves(mddev);
3219
3220         PRINTK("--- raid5d inactive\n");
3221 }
3222
3223 static ssize_t
3224 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
3225 {
3226         raid5_conf_t *conf = mddev_to_conf(mddev);
3227         if (conf)
3228                 return sprintf(page, "%d\n", conf->max_nr_stripes);
3229         else
3230                 return 0;
3231 }
3232
3233 static ssize_t
3234 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
3235 {
3236         raid5_conf_t *conf = mddev_to_conf(mddev);
3237         char *end;
3238         int new;
3239         if (len >= PAGE_SIZE)
3240                 return -EINVAL;
3241         if (!conf)
3242                 return -ENODEV;
3243
3244         new = simple_strtoul(page, &end, 10);
3245         if (!*page || (*end && *end != '\n') )
3246                 return -EINVAL;
3247         if (new <= 16 || new > 32768)
3248                 return -EINVAL;
3249         while (new < conf->max_nr_stripes) {
3250                 if (drop_one_stripe(conf))
3251                         conf->max_nr_stripes--;
3252                 else
3253                         break;
3254         }
3255         md_allow_write(mddev);
3256         while (new > conf->max_nr_stripes) {
3257                 if (grow_one_stripe(conf))
3258                         conf->max_nr_stripes++;
3259                 else break;
3260         }
3261         return len;
3262 }
3263
3264 static struct md_sysfs_entry
3265 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
3266                                 raid5_show_stripe_cache_size,
3267                                 raid5_store_stripe_cache_size);
3268
3269 static ssize_t
3270 stripe_cache_active_show(mddev_t *mddev, char *page)
3271 {
3272         raid5_conf_t *conf = mddev_to_conf(mddev);
3273         if (conf)
3274                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
3275         else
3276                 return 0;
3277 }
3278
3279 static struct md_sysfs_entry
3280 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
3281
3282 static struct attribute *raid5_attrs[] =  {
3283         &raid5_stripecache_size.attr,
3284         &raid5_stripecache_active.attr,
3285         NULL,
3286 };
3287 static struct attribute_group raid5_attrs_group = {
3288         .name = NULL,
3289         .attrs = raid5_attrs,
3290 };
3291
3292 static int run(mddev_t *mddev)
3293 {
3294         raid5_conf_t *conf;
3295         int raid_disk, memory;
3296         mdk_rdev_t *rdev;
3297         struct disk_info *disk;
3298         struct list_head *tmp;
3299         int working_disks = 0;
3300
3301         if (mddev->level != 5 && mddev->level != 4 && mddev->level != 6) {
3302                 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
3303                        mdname(mddev), mddev->level);
3304                 return -EIO;
3305         }
3306
3307         if (mddev->reshape_position != MaxSector) {
3308                 /* Check that we can continue the reshape.
3309                  * Currently only disks can change, it must
3310                  * increase, and we must be past the point where
3311                  * a stripe over-writes itself
3312                  */
3313                 sector_t here_new, here_old;
3314                 int old_disks;
3315
3316                 if (mddev->new_level != mddev->level ||
3317                     mddev->new_layout != mddev->layout ||
3318                     mddev->new_chunk != mddev->chunk_size) {
3319                         printk(KERN_ERR "raid5: %s: unsupported reshape required - aborting.\n",
3320                                mdname(mddev));
3321                         return -EINVAL;
3322                 }
3323                 if (mddev->delta_disks <= 0) {
3324                         printk(KERN_ERR "raid5: %s: unsupported reshape (reduce disks) required - aborting.\n",
3325                                mdname(mddev));
3326                         return -EINVAL;
3327                 }
3328                 old_disks = mddev->raid_disks - mddev->delta_disks;
3329                 /* reshape_position must be on a new-stripe boundary, and one
3330                  * further up in new geometry must map after here in old geometry.
3331                  */
3332                 here_new = mddev->reshape_position;
3333                 if (sector_div(here_new, (mddev->chunk_size>>9)*(mddev->raid_disks-1))) {
3334                         printk(KERN_ERR "raid5: reshape_position not on a stripe boundary\n");
3335                         return -EINVAL;
3336                 }
3337                 /* here_new is the stripe we will write to */
3338                 here_old = mddev->reshape_position;
3339                 sector_div(here_old, (mddev->chunk_size>>9)*(old_disks-1));
3340                 /* here_old is the first stripe that we might need to read from */
3341                 if (here_new >= here_old) {
3342                         /* Reading from the same stripe as writing to - bad */
3343                         printk(KERN_ERR "raid5: reshape_position too early for auto-recovery - aborting.\n");
3344                         return -EINVAL;
3345                 }
3346                 printk(KERN_INFO "raid5: reshape will continue\n");
3347                 /* OK, we should be able to continue; */
3348         }
3349
3350
3351         mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
3352         if ((conf = mddev->private) == NULL)
3353                 goto abort;
3354         if (mddev->reshape_position == MaxSector) {
3355                 conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks;
3356         } else {
3357                 conf->raid_disks = mddev->raid_disks;
3358                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
3359         }
3360
3361         conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
3362                               GFP_KERNEL);
3363         if (!conf->disks)
3364                 goto abort;
3365
3366         conf->mddev = mddev;
3367
3368         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
3369                 goto abort;
3370
3371         if (mddev->level == 6) {
3372                 conf->spare_page = alloc_page(GFP_KERNEL);
3373                 if (!conf->spare_page)
3374                         goto abort;
3375         }
3376         spin_lock_init(&conf->device_lock);
3377         init_waitqueue_head(&conf->wait_for_stripe);
3378         init_waitqueue_head(&conf->wait_for_overlap);
3379         INIT_LIST_HEAD(&conf->handle_list);
3380         INIT_LIST_HEAD(&conf->delayed_list);
3381         INIT_LIST_HEAD(&conf->bitmap_list);
3382         INIT_LIST_HEAD(&conf->inactive_list);
3383         atomic_set(&conf->active_stripes, 0);
3384         atomic_set(&conf->preread_active_stripes, 0);
3385         atomic_set(&conf->active_aligned_reads, 0);
3386
3387         PRINTK("raid5: run(%s) called.\n", mdname(mddev));
3388
3389         ITERATE_RDEV(mddev,rdev,tmp) {
3390                 raid_disk = rdev->raid_disk;
3391                 if (raid_disk >= conf->raid_disks
3392                     || raid_disk < 0)
3393                         continue;
3394                 disk = conf->disks + raid_disk;
3395
3396                 disk->rdev = rdev;
3397
3398                 if (test_bit(In_sync, &rdev->flags)) {
3399                         char b[BDEVNAME_SIZE];
3400                         printk(KERN_INFO "raid5: device %s operational as raid"
3401                                 " disk %d\n", bdevname(rdev->bdev,b),
3402                                 raid_disk);
3403                         working_disks++;
3404                 }
3405         }
3406
3407         /*
3408          * 0 for a fully functional array, 1 or 2 for a degraded array.
3409          */
3410         mddev->degraded = conf->raid_disks - working_disks;
3411         conf->mddev = mddev;
3412         conf->chunk_size = mddev->chunk_size;
3413         conf->level = mddev->level;
3414         if (conf->level == 6)
3415                 conf->max_degraded = 2;
3416         else
3417                 conf->max_degraded = 1;
3418         conf->algorithm = mddev->layout;
3419         conf->max_nr_stripes = NR_STRIPES;
3420         conf->expand_progress = mddev->reshape_position;
3421
3422         /* device size must be a multiple of chunk size */
3423         mddev->size &= ~(mddev->chunk_size/1024 -1);
3424         mddev->resync_max_sectors = mddev->size << 1;
3425
3426         if (conf->level == 6 && conf->raid_disks < 4) {
3427                 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
3428                        mdname(mddev), conf->raid_disks);
3429                 goto abort;
3430         }
3431         if (!conf->chunk_size || conf->chunk_size % 4) {
3432                 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
3433                         conf->chunk_size, mdname(mddev));
3434                 goto abort;
3435         }
3436         if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
3437                 printk(KERN_ERR 
3438                         "raid5: unsupported parity algorithm %d for %s\n",
3439                         conf->algorithm, mdname(mddev));
3440                 goto abort;
3441         }
3442         if (mddev->degraded > conf->max_degraded) {
3443                 printk(KERN_ERR "raid5: not enough operational devices for %s"
3444                         " (%d/%d failed)\n",
3445                         mdname(mddev), mddev->degraded, conf->raid_disks);
3446                 goto abort;
3447         }
3448
3449         if (mddev->degraded > 0 &&
3450             mddev->recovery_cp != MaxSector) {
3451                 if (mddev->ok_start_degraded)
3452                         printk(KERN_WARNING
3453                                "raid5: starting dirty degraded array: %s"
3454                                "- data corruption possible.\n",
3455                                mdname(mddev));
3456                 else {
3457                         printk(KERN_ERR
3458                                "raid5: cannot start dirty degraded array for %s\n",
3459                                mdname(mddev));
3460                         goto abort;
3461                 }
3462         }
3463
3464         {
3465                 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
3466                 if (!mddev->thread) {
3467                         printk(KERN_ERR 
3468                                 "raid5: couldn't allocate thread for %s\n",
3469                                 mdname(mddev));
3470                         goto abort;
3471                 }
3472         }
3473         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
3474                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
3475         if (grow_stripes(conf, conf->max_nr_stripes)) {
3476                 printk(KERN_ERR 
3477                         "raid5: couldn't allocate %dkB for buffers\n", memory);
3478                 shrink_stripes(conf);
3479                 md_unregister_thread(mddev->thread);
3480                 goto abort;
3481         } else
3482                 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
3483                         memory, mdname(mddev));
3484
3485         if (mddev->degraded == 0)
3486                 printk("raid5: raid level %d set %s active with %d out of %d"
3487                         " devices, algorithm %d\n", conf->level, mdname(mddev), 
3488                         mddev->raid_disks-mddev->degraded, mddev->raid_disks,
3489                         conf->algorithm);
3490         else
3491                 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
3492                         " out of %d devices, algorithm %d\n", conf->level,
3493                         mdname(mddev), mddev->raid_disks - mddev->degraded,
3494                         mddev->raid_disks, conf->algorithm);
3495
3496         print_raid5_conf(conf);
3497
3498         if (conf->expand_progress != MaxSector) {
3499                 printk("...ok start reshape thread\n");
3500                 conf->expand_lo = conf->expand_progress;
3501                 atomic_set(&conf->reshape_stripes, 0);
3502                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3503                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3504                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3505                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3506                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3507                                                         "%s_reshape");
3508         }
3509
3510         /* read-ahead size must cover two whole stripes, which is
3511          * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
3512          */
3513         {
3514                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
3515                 int stripe = data_disks *
3516                         (mddev->chunk_size / PAGE_SIZE);
3517                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3518                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3519         }
3520
3521         /* Ok, everything is just fine now */
3522         sysfs_create_group(&mddev->kobj, &raid5_attrs_group);
3523
3524         mddev->queue->unplug_fn = raid5_unplug_device;
3525         mddev->queue->issue_flush_fn = raid5_issue_flush;
3526         mddev->queue->backing_dev_info.congested_fn = raid5_congested;
3527         mddev->queue->backing_dev_info.congested_data = mddev;
3528
3529         mddev->array_size =  mddev->size * (conf->previous_raid_disks -
3530                                             conf->max_degraded);
3531
3532         blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
3533
3534         return 0;
3535 abort:
3536         if (conf) {
3537                 print_raid5_conf(conf);
3538                 safe_put_page(conf->spare_page);
3539                 kfree(conf->disks);
3540                 kfree(conf->stripe_hashtbl);
3541                 kfree(conf);
3542         }
3543         mddev->private = NULL;
3544         printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
3545         return -EIO;
3546 }
3547
3548
3549
3550 static int stop(mddev_t *mddev)
3551 {
3552         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3553
3554         md_unregister_thread(mddev->thread);
3555         mddev->thread = NULL;
3556         shrink_stripes(conf);
3557         kfree(conf->stripe_hashtbl);
3558         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
3559         sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
3560         kfree(conf->disks);
3561         kfree(conf);
3562         mddev->private = NULL;
3563         return 0;
3564 }
3565
3566 #if RAID5_DEBUG
3567 static void print_sh (struct seq_file *seq, struct stripe_head *sh)
3568 {
3569         int i;
3570
3571         seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
3572                    (unsigned long long)sh->sector, sh->pd_idx, sh->state);
3573         seq_printf(seq, "sh %llu,  count %d.\n",
3574                    (unsigned long long)sh->sector, atomic_read(&sh->count));
3575         seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
3576         for (i = 0; i < sh->disks; i++) {
3577                 seq_printf(seq, "(cache%d: %p %ld) ",
3578                            i, sh->dev[i].page, sh->dev[i].flags);
3579         }
3580         seq_printf(seq, "\n");
3581 }
3582
3583 static void printall (struct seq_file *seq, raid5_conf_t *conf)
3584 {
3585         struct stripe_head *sh;
3586         struct hlist_node *hn;
3587         int i;
3588
3589         spin_lock_irq(&conf->device_lock);
3590         for (i = 0; i < NR_HASH; i++) {
3591                 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
3592                         if (sh->raid_conf != conf)
3593                                 continue;
3594                         print_sh(seq, sh);
3595                 }
3596         }
3597         spin_unlock_irq(&conf->device_lock);
3598 }
3599 #endif
3600
3601 static void status (struct seq_file *seq, mddev_t *mddev)
3602 {
3603         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3604         int i;
3605
3606         seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
3607         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
3608         for (i = 0; i < conf->raid_disks; i++)
3609                 seq_printf (seq, "%s",
3610                                conf->disks[i].rdev &&
3611                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
3612         seq_printf (seq, "]");
3613 #if RAID5_DEBUG
3614         seq_printf (seq, "\n");
3615         printall(seq, conf);
3616 #endif
3617 }
3618
3619 static void print_raid5_conf (raid5_conf_t *conf)
3620 {
3621         int i;
3622         struct disk_info *tmp;
3623
3624         printk("RAID5 conf printout:\n");
3625         if (!conf) {
3626                 printk("(conf==NULL)\n");
3627                 return;
3628         }
3629         printk(" --- rd:%d wd:%d\n", conf->raid_disks,
3630                  conf->raid_disks - conf->mddev->degraded);
3631
3632         for (i = 0; i < conf->raid_disks; i++) {
3633                 char b[BDEVNAME_SIZE];
3634                 tmp = conf->disks + i;
3635                 if (tmp->rdev)
3636                 printk(" disk %d, o:%d, dev:%s\n",
3637                         i, !test_bit(Faulty, &tmp->rdev->flags),
3638                         bdevname(tmp->rdev->bdev,b));
3639         }
3640 }
3641
3642 static int raid5_spare_active(mddev_t *mddev)
3643 {
3644         int i;
3645         raid5_conf_t *conf = mddev->private;
3646         struct disk_info *tmp;
3647
3648         for (i = 0; i < conf->raid_disks; i++) {
3649                 tmp = conf->disks + i;
3650                 if (tmp->rdev
3651                     && !test_bit(Faulty, &tmp->rdev->flags)
3652                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
3653                         unsigned long flags;
3654                         spin_lock_irqsave(&conf->device_lock, flags);
3655                         mddev->degraded--;
3656                         spin_unlock_irqrestore(&conf->device_lock, flags);
3657                 }
3658         }
3659         print_raid5_conf(conf);
3660         return 0;
3661 }
3662
3663 static int raid5_remove_disk(mddev_t *mddev, int number)
3664 {
3665         raid5_conf_t *conf = mddev->private;
3666         int err = 0;
3667         mdk_rdev_t *rdev;
3668         struct disk_info *p = conf->disks + number;
3669
3670         print_raid5_conf(conf);
3671         rdev = p->rdev;
3672         if (rdev) {
3673                 if (test_bit(In_sync, &rdev->flags) ||
3674                     atomic_read(&rdev->nr_pending)) {
3675                         err = -EBUSY;
3676                         goto abort;
3677                 }
3678                 p->rdev = NULL;
3679                 synchronize_rcu();
3680                 if (atomic_read(&rdev->nr_pending)) {
3681                         /* lost the race, try later */
3682                         err = -EBUSY;
3683                         p->rdev = rdev;
3684                 }
3685         }
3686 abort:
3687
3688         print_raid5_conf(conf);
3689         return err;
3690 }
3691
3692 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
3693 {
3694         raid5_conf_t *conf = mddev->private;
3695         int found = 0;
3696         int disk;
3697         struct disk_info *p;
3698
3699         if (mddev->degraded > conf->max_degraded)
3700                 /* no point adding a device */
3701                 return 0;
3702
3703         /*
3704          * find the disk ... but prefer rdev->saved_raid_disk
3705          * if possible.
3706          */
3707         if (rdev->saved_raid_disk >= 0 &&
3708             conf->disks[rdev->saved_raid_disk].rdev == NULL)
3709                 disk = rdev->saved_raid_disk;
3710         else
3711                 disk = 0;
3712         for ( ; disk < conf->raid_disks; disk++)
3713                 if ((p=conf->disks + disk)->rdev == NULL) {
3714                         clear_bit(In_sync, &rdev->flags);
3715                         rdev->raid_disk = disk;
3716                         found = 1;
3717                         if (rdev->saved_raid_disk != disk)
3718                                 conf->fullsync = 1;
3719                         rcu_assign_pointer(p->rdev, rdev);
3720                         break;
3721                 }
3722         print_raid5_conf(conf);
3723         return found;
3724 }
3725
3726 static int raid5_resize(mddev_t *mddev, sector_t sectors)
3727 {
3728         /* no resync is happening, and there is enough space
3729          * on all devices, so we can resize.
3730          * We need to make sure resync covers any new space.
3731          * If the array is shrinking we should possibly wait until
3732          * any io in the removed space completes, but it hardly seems
3733          * worth it.
3734          */
3735         raid5_conf_t *conf = mddev_to_conf(mddev);
3736
3737         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
3738         mddev->array_size = (sectors * (mddev->raid_disks-conf->max_degraded))>>1;
3739         set_capacity(mddev->gendisk, mddev->array_size << 1);
3740         mddev->changed = 1;
3741         if (sectors/2  > mddev->size && mddev->recovery_cp == MaxSector) {
3742                 mddev->recovery_cp = mddev->size << 1;
3743                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3744         }
3745         mddev->size = sectors /2;
3746         mddev->resync_max_sectors = sectors;
3747         return 0;
3748 }
3749
3750 #ifdef CONFIG_MD_RAID5_RESHAPE
3751 static int raid5_check_reshape(mddev_t *mddev)
3752 {
3753         raid5_conf_t *conf = mddev_to_conf(mddev);
3754         int err;
3755
3756         if (mddev->delta_disks < 0 ||
3757             mddev->new_level != mddev->level)
3758                 return -EINVAL; /* Cannot shrink array or change level yet */
3759         if (mddev->delta_disks == 0)
3760                 return 0; /* nothing to do */
3761
3762         /* Can only proceed if there are plenty of stripe_heads.
3763          * We need a minimum of one full stripe,, and for sensible progress
3764          * it is best to have about 4 times that.
3765          * If we require 4 times, then the default 256 4K stripe_heads will
3766          * allow for chunk sizes up to 256K, which is probably OK.
3767          * If the chunk size is greater, user-space should request more
3768          * stripe_heads first.
3769          */
3770         if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
3771             (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
3772                 printk(KERN_WARNING "raid5: reshape: not enough stripes.  Needed %lu\n",
3773                        (mddev->chunk_size / STRIPE_SIZE)*4);
3774                 return -ENOSPC;
3775         }
3776
3777         err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
3778         if (err)
3779                 return err;
3780
3781         /* looks like we might be able to manage this */
3782         return 0;
3783 }
3784
3785 static int raid5_start_reshape(mddev_t *mddev)
3786 {
3787         raid5_conf_t *conf = mddev_to_conf(mddev);
3788         mdk_rdev_t *rdev;
3789         struct list_head *rtmp;
3790         int spares = 0;
3791         int added_devices = 0;
3792         unsigned long flags;
3793
3794         if (mddev->degraded ||
3795             test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
3796                 return -EBUSY;
3797
3798         ITERATE_RDEV(mddev, rdev, rtmp)
3799                 if (rdev->raid_disk < 0 &&
3800                     !test_bit(Faulty, &rdev->flags))
3801                         spares++;
3802
3803         if (spares < mddev->delta_disks-1)
3804                 /* Not enough devices even to make a degraded array
3805                  * of that size
3806                  */
3807                 return -EINVAL;
3808
3809         atomic_set(&conf->reshape_stripes, 0);
3810         spin_lock_irq(&conf->device_lock);
3811         conf->previous_raid_disks = conf->raid_disks;
3812         conf->raid_disks += mddev->delta_disks;
3813         conf->expand_progress = 0;
3814         conf->expand_lo = 0;
3815         spin_unlock_irq(&conf->device_lock);
3816
3817         /* Add some new drives, as many as will fit.
3818          * We know there are enough to make the newly sized array work.
3819          */
3820         ITERATE_RDEV(mddev, rdev, rtmp)
3821                 if (rdev->raid_disk < 0 &&
3822                     !test_bit(Faulty, &rdev->flags)) {
3823                         if (raid5_add_disk(mddev, rdev)) {
3824                                 char nm[20];
3825                                 set_bit(In_sync, &rdev->flags);
3826                                 added_devices++;
3827                                 rdev->recovery_offset = 0;
3828                                 sprintf(nm, "rd%d", rdev->raid_disk);
3829                                 sysfs_create_link(&mddev->kobj, &rdev->kobj, nm);
3830                         } else
3831                                 break;
3832                 }
3833
3834         spin_lock_irqsave(&conf->device_lock, flags);
3835         mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
3836         spin_unlock_irqrestore(&conf->device_lock, flags);
3837         mddev->raid_disks = conf->raid_disks;
3838         mddev->reshape_position = 0;
3839         set_bit(MD_CHANGE_DEVS, &mddev->flags);
3840
3841         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3842         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3843         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3844         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3845         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3846                                                 "%s_reshape");
3847         if (!mddev->sync_thread) {
3848                 mddev->recovery = 0;
3849                 spin_lock_irq(&conf->device_lock);
3850                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
3851                 conf->expand_progress = MaxSector;
3852                 spin_unlock_irq(&conf->device_lock);
3853                 return -EAGAIN;
3854         }
3855         md_wakeup_thread(mddev->sync_thread);
3856         md_new_event(mddev);
3857         return 0;
3858 }
3859 #endif
3860
3861 static void end_reshape(raid5_conf_t *conf)
3862 {
3863         struct block_device *bdev;
3864
3865         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
3866                 conf->mddev->array_size = conf->mddev->size * (conf->raid_disks-1);
3867                 set_capacity(conf->mddev->gendisk, conf->mddev->array_size << 1);
3868                 conf->mddev->changed = 1;
3869
3870                 bdev = bdget_disk(conf->mddev->gendisk, 0);
3871                 if (bdev) {
3872                         mutex_lock(&bdev->bd_inode->i_mutex);
3873                         i_size_write(bdev->bd_inode, (loff_t)conf->mddev->array_size << 10);
3874                         mutex_unlock(&bdev->bd_inode->i_mutex);
3875                         bdput(bdev);
3876                 }
3877                 spin_lock_irq(&conf->device_lock);
3878                 conf->expand_progress = MaxSector;
3879                 spin_unlock_irq(&conf->device_lock);
3880                 conf->mddev->reshape_position = MaxSector;
3881
3882                 /* read-ahead size must cover two whole stripes, which is
3883                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
3884                  */
3885                 {
3886                         int data_disks = conf->previous_raid_disks - conf->max_degraded;
3887                         int stripe = data_disks *
3888                                 (conf->mddev->chunk_size / PAGE_SIZE);
3889                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3890                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3891                 }
3892         }
3893 }
3894
3895 static void raid5_quiesce(mddev_t *mddev, int state)
3896 {
3897         raid5_conf_t *conf = mddev_to_conf(mddev);
3898
3899         switch(state) {
3900         case 2: /* resume for a suspend */
3901                 wake_up(&conf->wait_for_overlap);
3902                 break;
3903
3904         case 1: /* stop all writes */
3905                 spin_lock_irq(&conf->device_lock);
3906                 conf->quiesce = 1;
3907                 wait_event_lock_irq(conf->wait_for_stripe,
3908                                     atomic_read(&conf->active_stripes) == 0 &&
3909                                     atomic_read(&conf->active_aligned_reads) == 0,
3910                                     conf->device_lock, /* nothing */);
3911                 spin_unlock_irq(&conf->device_lock);
3912                 break;
3913
3914         case 0: /* re-enable writes */
3915                 spin_lock_irq(&conf->device_lock);
3916                 conf->quiesce = 0;
3917                 wake_up(&conf->wait_for_stripe);
3918                 wake_up(&conf->wait_for_overlap);
3919                 spin_unlock_irq(&conf->device_lock);
3920                 break;
3921         }
3922 }
3923
3924 static struct mdk_personality raid6_personality =
3925 {
3926         .name           = "raid6",
3927         .level          = 6,
3928         .owner          = THIS_MODULE,
3929         .make_request   = make_request,
3930         .run            = run,
3931         .stop           = stop,
3932         .status         = status,
3933         .error_handler  = error,
3934         .hot_add_disk   = raid5_add_disk,
3935         .hot_remove_disk= raid5_remove_disk,
3936         .spare_active   = raid5_spare_active,
3937         .sync_request   = sync_request,
3938         .resize         = raid5_resize,
3939         .quiesce        = raid5_quiesce,
3940 };
3941 static struct mdk_personality raid5_personality =
3942 {
3943         .name           = "raid5",
3944         .level          = 5,
3945         .owner          = THIS_MODULE,
3946         .make_request   = make_request,
3947         .run            = run,
3948         .stop           = stop,
3949         .status         = status,
3950         .error_handler  = error,
3951         .hot_add_disk   = raid5_add_disk,
3952         .hot_remove_disk= raid5_remove_disk,
3953         .spare_active   = raid5_spare_active,
3954         .sync_request   = sync_request,
3955         .resize         = raid5_resize,
3956 #ifdef CONFIG_MD_RAID5_RESHAPE
3957         .check_reshape  = raid5_check_reshape,
3958         .start_reshape  = raid5_start_reshape,
3959 #endif
3960         .quiesce        = raid5_quiesce,
3961 };
3962
3963 static struct mdk_personality raid4_personality =
3964 {
3965         .name           = "raid4",
3966         .level          = 4,
3967         .owner          = THIS_MODULE,
3968         .make_request   = make_request,
3969         .run            = run,
3970         .stop           = stop,
3971         .status         = status,
3972         .error_handler  = error,
3973         .hot_add_disk   = raid5_add_disk,
3974         .hot_remove_disk= raid5_remove_disk,
3975         .spare_active   = raid5_spare_active,
3976         .sync_request   = sync_request,
3977         .resize         = raid5_resize,
3978         .quiesce        = raid5_quiesce,
3979 };
3980
3981 static int __init raid5_init(void)
3982 {
3983         int e;
3984
3985         e = raid6_select_algo();
3986         if ( e )
3987                 return e;
3988         register_md_personality(&raid6_personality);
3989         register_md_personality(&raid5_personality);
3990         register_md_personality(&raid4_personality);
3991         return 0;
3992 }
3993
3994 static void raid5_exit(void)
3995 {
3996         unregister_md_personality(&raid6_personality);
3997         unregister_md_personality(&raid5_personality);
3998         unregister_md_personality(&raid4_personality);
3999 }
4000
4001 module_init(raid5_init);
4002 module_exit(raid5_exit);
4003 MODULE_LICENSE("GPL");
4004 MODULE_ALIAS("md-personality-4"); /* RAID5 */
4005 MODULE_ALIAS("md-raid5");
4006 MODULE_ALIAS("md-raid4");
4007 MODULE_ALIAS("md-level-5");
4008 MODULE_ALIAS("md-level-4");
4009 MODULE_ALIAS("md-personality-8"); /* RAID6 */
4010 MODULE_ALIAS("md-raid6");
4011 MODULE_ALIAS("md-level-6");
4012
4013 /* This used to be two separate modules, they were: */
4014 MODULE_ALIAS("raid5");
4015 MODULE_ALIAS("raid6");