Merge branch 'master'
[linux-2.6] / drivers / md / raid5.c
1 /*
2  * raid5.c : Multiple Devices driver for Linux
3  *         Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4  *         Copyright (C) 1999, 2000 Ingo Molnar
5  *
6  * RAID-5 management functions.
7  *
8  * This program is free software; you can redistribute it and/or modify
9  * it under the terms of the GNU General Public License as published by
10  * the Free Software Foundation; either version 2, or (at your option)
11  * any later version.
12  *
13  * You should have received a copy of the GNU General Public License
14  * (for example /usr/src/linux/COPYING); if not, write to the Free
15  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
16  */
17
18
19 #include <linux/config.h>
20 #include <linux/module.h>
21 #include <linux/slab.h>
22 #include <linux/raid/raid5.h>
23 #include <linux/highmem.h>
24 #include <linux/bitops.h>
25 #include <asm/atomic.h>
26
27 #include <linux/raid/bitmap.h>
28
29 /*
30  * Stripe cache
31  */
32
33 #define NR_STRIPES              256
34 #define STRIPE_SIZE             PAGE_SIZE
35 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
36 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
37 #define IO_THRESHOLD            1
38 #define HASH_PAGES              1
39 #define HASH_PAGES_ORDER        0
40 #define NR_HASH                 (HASH_PAGES * PAGE_SIZE / sizeof(struct stripe_head *))
41 #define HASH_MASK               (NR_HASH - 1)
42
43 #define stripe_hash(conf, sect) ((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK])
44
45 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
46  * order without overlap.  There may be several bio's per stripe+device, and
47  * a bio could span several devices.
48  * When walking this list for a particular stripe+device, we must never proceed
49  * beyond a bio that extends past this device, as the next bio might no longer
50  * be valid.
51  * This macro is used to determine the 'next' bio in the list, given the sector
52  * of the current stripe+device
53  */
54 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
55 /*
56  * The following can be used to debug the driver
57  */
58 #define RAID5_DEBUG     0
59 #define RAID5_PARANOIA  1
60 #if RAID5_PARANOIA && defined(CONFIG_SMP)
61 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
62 #else
63 # define CHECK_DEVLOCK()
64 #endif
65
66 #define PRINTK(x...) ((void)(RAID5_DEBUG && printk(x)))
67 #if RAID5_DEBUG
68 #define inline
69 #define __inline__
70 #endif
71
72 static void print_raid5_conf (raid5_conf_t *conf);
73
74 static inline void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
75 {
76         if (atomic_dec_and_test(&sh->count)) {
77                 if (!list_empty(&sh->lru))
78                         BUG();
79                 if (atomic_read(&conf->active_stripes)==0)
80                         BUG();
81                 if (test_bit(STRIPE_HANDLE, &sh->state)) {
82                         if (test_bit(STRIPE_DELAYED, &sh->state))
83                                 list_add_tail(&sh->lru, &conf->delayed_list);
84                         else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
85                                  conf->seq_write == sh->bm_seq)
86                                 list_add_tail(&sh->lru, &conf->bitmap_list);
87                         else {
88                                 clear_bit(STRIPE_BIT_DELAY, &sh->state);
89                                 list_add_tail(&sh->lru, &conf->handle_list);
90                         }
91                         md_wakeup_thread(conf->mddev->thread);
92                 } else {
93                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
94                                 atomic_dec(&conf->preread_active_stripes);
95                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
96                                         md_wakeup_thread(conf->mddev->thread);
97                         }
98                         list_add_tail(&sh->lru, &conf->inactive_list);
99                         atomic_dec(&conf->active_stripes);
100                         if (!conf->inactive_blocked ||
101                             atomic_read(&conf->active_stripes) < (conf->max_nr_stripes*3/4))
102                                 wake_up(&conf->wait_for_stripe);
103                 }
104         }
105 }
106 static void release_stripe(struct stripe_head *sh)
107 {
108         raid5_conf_t *conf = sh->raid_conf;
109         unsigned long flags;
110         
111         spin_lock_irqsave(&conf->device_lock, flags);
112         __release_stripe(conf, sh);
113         spin_unlock_irqrestore(&conf->device_lock, flags);
114 }
115
116 static void remove_hash(struct stripe_head *sh)
117 {
118         PRINTK("remove_hash(), stripe %llu\n", (unsigned long long)sh->sector);
119
120         if (sh->hash_pprev) {
121                 if (sh->hash_next)
122                         sh->hash_next->hash_pprev = sh->hash_pprev;
123                 *sh->hash_pprev = sh->hash_next;
124                 sh->hash_pprev = NULL;
125         }
126 }
127
128 static __inline__ void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
129 {
130         struct stripe_head **shp = &stripe_hash(conf, sh->sector);
131
132         PRINTK("insert_hash(), stripe %llu\n", (unsigned long long)sh->sector);
133
134         CHECK_DEVLOCK();
135         if ((sh->hash_next = *shp) != NULL)
136                 (*shp)->hash_pprev = &sh->hash_next;
137         *shp = sh;
138         sh->hash_pprev = shp;
139 }
140
141
142 /* find an idle stripe, make sure it is unhashed, and return it. */
143 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
144 {
145         struct stripe_head *sh = NULL;
146         struct list_head *first;
147
148         CHECK_DEVLOCK();
149         if (list_empty(&conf->inactive_list))
150                 goto out;
151         first = conf->inactive_list.next;
152         sh = list_entry(first, struct stripe_head, lru);
153         list_del_init(first);
154         remove_hash(sh);
155         atomic_inc(&conf->active_stripes);
156 out:
157         return sh;
158 }
159
160 static void shrink_buffers(struct stripe_head *sh, int num)
161 {
162         struct page *p;
163         int i;
164
165         for (i=0; i<num ; i++) {
166                 p = sh->dev[i].page;
167                 if (!p)
168                         continue;
169                 sh->dev[i].page = NULL;
170                 page_cache_release(p);
171         }
172 }
173
174 static int grow_buffers(struct stripe_head *sh, int num)
175 {
176         int i;
177
178         for (i=0; i<num; i++) {
179                 struct page *page;
180
181                 if (!(page = alloc_page(GFP_KERNEL))) {
182                         return 1;
183                 }
184                 sh->dev[i].page = page;
185         }
186         return 0;
187 }
188
189 static void raid5_build_block (struct stripe_head *sh, int i);
190
191 static inline void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx)
192 {
193         raid5_conf_t *conf = sh->raid_conf;
194         int disks = conf->raid_disks, i;
195
196         if (atomic_read(&sh->count) != 0)
197                 BUG();
198         if (test_bit(STRIPE_HANDLE, &sh->state))
199                 BUG();
200         
201         CHECK_DEVLOCK();
202         PRINTK("init_stripe called, stripe %llu\n", 
203                 (unsigned long long)sh->sector);
204
205         remove_hash(sh);
206         
207         sh->sector = sector;
208         sh->pd_idx = pd_idx;
209         sh->state = 0;
210
211         for (i=disks; i--; ) {
212                 struct r5dev *dev = &sh->dev[i];
213
214                 if (dev->toread || dev->towrite || dev->written ||
215                     test_bit(R5_LOCKED, &dev->flags)) {
216                         printk("sector=%llx i=%d %p %p %p %d\n",
217                                (unsigned long long)sh->sector, i, dev->toread,
218                                dev->towrite, dev->written,
219                                test_bit(R5_LOCKED, &dev->flags));
220                         BUG();
221                 }
222                 dev->flags = 0;
223                 raid5_build_block(sh, i);
224         }
225         insert_hash(conf, sh);
226 }
227
228 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector)
229 {
230         struct stripe_head *sh;
231
232         CHECK_DEVLOCK();
233         PRINTK("__find_stripe, sector %llu\n", (unsigned long long)sector);
234         for (sh = stripe_hash(conf, sector); sh; sh = sh->hash_next)
235                 if (sh->sector == sector)
236                         return sh;
237         PRINTK("__stripe %llu not in cache\n", (unsigned long long)sector);
238         return NULL;
239 }
240
241 static void unplug_slaves(mddev_t *mddev);
242 static void raid5_unplug_device(request_queue_t *q);
243
244 static struct stripe_head *get_active_stripe(raid5_conf_t *conf, sector_t sector,
245                                              int pd_idx, int noblock) 
246 {
247         struct stripe_head *sh;
248
249         PRINTK("get_stripe, sector %llu\n", (unsigned long long)sector);
250
251         spin_lock_irq(&conf->device_lock);
252
253         do {
254                 wait_event_lock_irq(conf->wait_for_stripe,
255                                     conf->quiesce == 0,
256                                     conf->device_lock, /* nothing */);
257                 sh = __find_stripe(conf, sector);
258                 if (!sh) {
259                         if (!conf->inactive_blocked)
260                                 sh = get_free_stripe(conf);
261                         if (noblock && sh == NULL)
262                                 break;
263                         if (!sh) {
264                                 conf->inactive_blocked = 1;
265                                 wait_event_lock_irq(conf->wait_for_stripe,
266                                                     !list_empty(&conf->inactive_list) &&
267                                                     (atomic_read(&conf->active_stripes)
268                                                      < (conf->max_nr_stripes *3/4)
269                                                      || !conf->inactive_blocked),
270                                                     conf->device_lock,
271                                                     unplug_slaves(conf->mddev);
272                                         );
273                                 conf->inactive_blocked = 0;
274                         } else
275                                 init_stripe(sh, sector, pd_idx);
276                 } else {
277                         if (atomic_read(&sh->count)) {
278                                 if (!list_empty(&sh->lru))
279                                         BUG();
280                         } else {
281                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
282                                         atomic_inc(&conf->active_stripes);
283                                 if (list_empty(&sh->lru))
284                                         BUG();
285                                 list_del_init(&sh->lru);
286                         }
287                 }
288         } while (sh == NULL);
289
290         if (sh)
291                 atomic_inc(&sh->count);
292
293         spin_unlock_irq(&conf->device_lock);
294         return sh;
295 }
296
297 static int grow_one_stripe(raid5_conf_t *conf)
298 {
299         struct stripe_head *sh;
300         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
301         if (!sh)
302                 return 0;
303         memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
304         sh->raid_conf = conf;
305         spin_lock_init(&sh->lock);
306
307         if (grow_buffers(sh, conf->raid_disks)) {
308                 shrink_buffers(sh, conf->raid_disks);
309                 kmem_cache_free(conf->slab_cache, sh);
310                 return 0;
311         }
312         /* we just created an active stripe so... */
313         atomic_set(&sh->count, 1);
314         atomic_inc(&conf->active_stripes);
315         INIT_LIST_HEAD(&sh->lru);
316         release_stripe(sh);
317         return 1;
318 }
319
320 static int grow_stripes(raid5_conf_t *conf, int num)
321 {
322         kmem_cache_t *sc;
323         int devs = conf->raid_disks;
324
325         sprintf(conf->cache_name, "raid5/%s", mdname(conf->mddev));
326
327         sc = kmem_cache_create(conf->cache_name, 
328                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
329                                0, 0, NULL, NULL);
330         if (!sc)
331                 return 1;
332         conf->slab_cache = sc;
333         while (num--) {
334                 if (!grow_one_stripe(conf))
335                         return 1;
336         }
337         return 0;
338 }
339
340 static int drop_one_stripe(raid5_conf_t *conf)
341 {
342         struct stripe_head *sh;
343
344         spin_lock_irq(&conf->device_lock);
345         sh = get_free_stripe(conf);
346         spin_unlock_irq(&conf->device_lock);
347         if (!sh)
348                 return 0;
349         if (atomic_read(&sh->count))
350                 BUG();
351         shrink_buffers(sh, conf->raid_disks);
352         kmem_cache_free(conf->slab_cache, sh);
353         atomic_dec(&conf->active_stripes);
354         return 1;
355 }
356
357 static void shrink_stripes(raid5_conf_t *conf)
358 {
359         while (drop_one_stripe(conf))
360                 ;
361
362         kmem_cache_destroy(conf->slab_cache);
363         conf->slab_cache = NULL;
364 }
365
366 static int raid5_end_read_request(struct bio * bi, unsigned int bytes_done,
367                                    int error)
368 {
369         struct stripe_head *sh = bi->bi_private;
370         raid5_conf_t *conf = sh->raid_conf;
371         int disks = conf->raid_disks, i;
372         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
373
374         if (bi->bi_size)
375                 return 1;
376
377         for (i=0 ; i<disks; i++)
378                 if (bi == &sh->dev[i].req)
379                         break;
380
381         PRINTK("end_read_request %llu/%d, count: %d, uptodate %d.\n", 
382                 (unsigned long long)sh->sector, i, atomic_read(&sh->count), 
383                 uptodate);
384         if (i == disks) {
385                 BUG();
386                 return 0;
387         }
388
389         if (uptodate) {
390 #if 0
391                 struct bio *bio;
392                 unsigned long flags;
393                 spin_lock_irqsave(&conf->device_lock, flags);
394                 /* we can return a buffer if we bypassed the cache or
395                  * if the top buffer is not in highmem.  If there are
396                  * multiple buffers, leave the extra work to
397                  * handle_stripe
398                  */
399                 buffer = sh->bh_read[i];
400                 if (buffer &&
401                     (!PageHighMem(buffer->b_page)
402                      || buffer->b_page == bh->b_page )
403                         ) {
404                         sh->bh_read[i] = buffer->b_reqnext;
405                         buffer->b_reqnext = NULL;
406                 } else
407                         buffer = NULL;
408                 spin_unlock_irqrestore(&conf->device_lock, flags);
409                 if (sh->bh_page[i]==bh->b_page)
410                         set_buffer_uptodate(bh);
411                 if (buffer) {
412                         if (buffer->b_page != bh->b_page)
413                                 memcpy(buffer->b_data, bh->b_data, bh->b_size);
414                         buffer->b_end_io(buffer, 1);
415                 }
416 #else
417                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
418 #endif
419                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
420                         printk("R5: read error corrected!!\n");
421                         clear_bit(R5_ReadError, &sh->dev[i].flags);
422                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
423                 }
424                 if (atomic_read(&conf->disks[i].rdev->read_errors))
425                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
426         } else {
427                 int retry = 0;
428                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
429                 atomic_inc(&conf->disks[i].rdev->read_errors);
430                 if (conf->mddev->degraded)
431                         printk("R5: read error not correctable.\n");
432                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
433                         /* Oh, no!!! */
434                         printk("R5: read error NOT corrected!!\n");
435                 else if (atomic_read(&conf->disks[i].rdev->read_errors)
436                          > conf->max_nr_stripes)
437                         printk("raid5: Too many read errors, failing device.\n");
438                 else
439                         retry = 1;
440                 if (retry)
441                         set_bit(R5_ReadError, &sh->dev[i].flags);
442                 else {
443                         clear_bit(R5_ReadError, &sh->dev[i].flags);
444                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
445                         md_error(conf->mddev, conf->disks[i].rdev);
446                 }
447         }
448         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
449 #if 0
450         /* must restore b_page before unlocking buffer... */
451         if (sh->bh_page[i] != bh->b_page) {
452                 bh->b_page = sh->bh_page[i];
453                 bh->b_data = page_address(bh->b_page);
454                 clear_buffer_uptodate(bh);
455         }
456 #endif
457         clear_bit(R5_LOCKED, &sh->dev[i].flags);
458         set_bit(STRIPE_HANDLE, &sh->state);
459         release_stripe(sh);
460         return 0;
461 }
462
463 static int raid5_end_write_request (struct bio *bi, unsigned int bytes_done,
464                                     int error)
465 {
466         struct stripe_head *sh = bi->bi_private;
467         raid5_conf_t *conf = sh->raid_conf;
468         int disks = conf->raid_disks, i;
469         unsigned long flags;
470         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
471
472         if (bi->bi_size)
473                 return 1;
474
475         for (i=0 ; i<disks; i++)
476                 if (bi == &sh->dev[i].req)
477                         break;
478
479         PRINTK("end_write_request %llu/%d, count %d, uptodate: %d.\n", 
480                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
481                 uptodate);
482         if (i == disks) {
483                 BUG();
484                 return 0;
485         }
486
487         spin_lock_irqsave(&conf->device_lock, flags);
488         if (!uptodate)
489                 md_error(conf->mddev, conf->disks[i].rdev);
490
491         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
492         
493         clear_bit(R5_LOCKED, &sh->dev[i].flags);
494         set_bit(STRIPE_HANDLE, &sh->state);
495         __release_stripe(conf, sh);
496         spin_unlock_irqrestore(&conf->device_lock, flags);
497         return 0;
498 }
499
500
501 static sector_t compute_blocknr(struct stripe_head *sh, int i);
502         
503 static void raid5_build_block (struct stripe_head *sh, int i)
504 {
505         struct r5dev *dev = &sh->dev[i];
506
507         bio_init(&dev->req);
508         dev->req.bi_io_vec = &dev->vec;
509         dev->req.bi_vcnt++;
510         dev->req.bi_max_vecs++;
511         dev->vec.bv_page = dev->page;
512         dev->vec.bv_len = STRIPE_SIZE;
513         dev->vec.bv_offset = 0;
514
515         dev->req.bi_sector = sh->sector;
516         dev->req.bi_private = sh;
517
518         dev->flags = 0;
519         if (i != sh->pd_idx)
520                 dev->sector = compute_blocknr(sh, i);
521 }
522
523 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
524 {
525         char b[BDEVNAME_SIZE];
526         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
527         PRINTK("raid5: error called\n");
528
529         if (!test_bit(Faulty, &rdev->flags)) {
530                 mddev->sb_dirty = 1;
531                 if (test_bit(In_sync, &rdev->flags)) {
532                         conf->working_disks--;
533                         mddev->degraded++;
534                         conf->failed_disks++;
535                         clear_bit(In_sync, &rdev->flags);
536                         /*
537                          * if recovery was running, make sure it aborts.
538                          */
539                         set_bit(MD_RECOVERY_ERR, &mddev->recovery);
540                 }
541                 set_bit(Faulty, &rdev->flags);
542                 printk (KERN_ALERT
543                         "raid5: Disk failure on %s, disabling device."
544                         " Operation continuing on %d devices\n",
545                         bdevname(rdev->bdev,b), conf->working_disks);
546         }
547 }       
548
549 /*
550  * Input: a 'big' sector number,
551  * Output: index of the data and parity disk, and the sector # in them.
552  */
553 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
554                         unsigned int data_disks, unsigned int * dd_idx,
555                         unsigned int * pd_idx, raid5_conf_t *conf)
556 {
557         long stripe;
558         unsigned long chunk_number;
559         unsigned int chunk_offset;
560         sector_t new_sector;
561         int sectors_per_chunk = conf->chunk_size >> 9;
562
563         /* First compute the information on this sector */
564
565         /*
566          * Compute the chunk number and the sector offset inside the chunk
567          */
568         chunk_offset = sector_div(r_sector, sectors_per_chunk);
569         chunk_number = r_sector;
570         BUG_ON(r_sector != chunk_number);
571
572         /*
573          * Compute the stripe number
574          */
575         stripe = chunk_number / data_disks;
576
577         /*
578          * Compute the data disk and parity disk indexes inside the stripe
579          */
580         *dd_idx = chunk_number % data_disks;
581
582         /*
583          * Select the parity disk based on the user selected algorithm.
584          */
585         if (conf->level == 4)
586                 *pd_idx = data_disks;
587         else switch (conf->algorithm) {
588                 case ALGORITHM_LEFT_ASYMMETRIC:
589                         *pd_idx = data_disks - stripe % raid_disks;
590                         if (*dd_idx >= *pd_idx)
591                                 (*dd_idx)++;
592                         break;
593                 case ALGORITHM_RIGHT_ASYMMETRIC:
594                         *pd_idx = stripe % raid_disks;
595                         if (*dd_idx >= *pd_idx)
596                                 (*dd_idx)++;
597                         break;
598                 case ALGORITHM_LEFT_SYMMETRIC:
599                         *pd_idx = data_disks - stripe % raid_disks;
600                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
601                         break;
602                 case ALGORITHM_RIGHT_SYMMETRIC:
603                         *pd_idx = stripe % raid_disks;
604                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
605                         break;
606                 default:
607                         printk("raid5: unsupported algorithm %d\n",
608                                 conf->algorithm);
609         }
610
611         /*
612          * Finally, compute the new sector number
613          */
614         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
615         return new_sector;
616 }
617
618
619 static sector_t compute_blocknr(struct stripe_head *sh, int i)
620 {
621         raid5_conf_t *conf = sh->raid_conf;
622         int raid_disks = conf->raid_disks, data_disks = raid_disks - 1;
623         sector_t new_sector = sh->sector, check;
624         int sectors_per_chunk = conf->chunk_size >> 9;
625         sector_t stripe;
626         int chunk_offset;
627         int chunk_number, dummy1, dummy2, dd_idx = i;
628         sector_t r_sector;
629
630         chunk_offset = sector_div(new_sector, sectors_per_chunk);
631         stripe = new_sector;
632         BUG_ON(new_sector != stripe);
633
634         
635         switch (conf->algorithm) {
636                 case ALGORITHM_LEFT_ASYMMETRIC:
637                 case ALGORITHM_RIGHT_ASYMMETRIC:
638                         if (i > sh->pd_idx)
639                                 i--;
640                         break;
641                 case ALGORITHM_LEFT_SYMMETRIC:
642                 case ALGORITHM_RIGHT_SYMMETRIC:
643                         if (i < sh->pd_idx)
644                                 i += raid_disks;
645                         i -= (sh->pd_idx + 1);
646                         break;
647                 default:
648                         printk("raid5: unsupported algorithm %d\n",
649                                 conf->algorithm);
650         }
651
652         chunk_number = stripe * data_disks + i;
653         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
654
655         check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
656         if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
657                 printk("compute_blocknr: map not correct\n");
658                 return 0;
659         }
660         return r_sector;
661 }
662
663
664
665 /*
666  * Copy data between a page in the stripe cache, and a bio.
667  * There are no alignment or size guarantees between the page or the
668  * bio except that there is some overlap.
669  * All iovecs in the bio must be considered.
670  */
671 static void copy_data(int frombio, struct bio *bio,
672                      struct page *page,
673                      sector_t sector)
674 {
675         char *pa = page_address(page);
676         struct bio_vec *bvl;
677         int i;
678         int page_offset;
679
680         if (bio->bi_sector >= sector)
681                 page_offset = (signed)(bio->bi_sector - sector) * 512;
682         else
683                 page_offset = (signed)(sector - bio->bi_sector) * -512;
684         bio_for_each_segment(bvl, bio, i) {
685                 int len = bio_iovec_idx(bio,i)->bv_len;
686                 int clen;
687                 int b_offset = 0;
688
689                 if (page_offset < 0) {
690                         b_offset = -page_offset;
691                         page_offset += b_offset;
692                         len -= b_offset;
693                 }
694
695                 if (len > 0 && page_offset + len > STRIPE_SIZE)
696                         clen = STRIPE_SIZE - page_offset;
697                 else clen = len;
698                         
699                 if (clen > 0) {
700                         char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
701                         if (frombio)
702                                 memcpy(pa+page_offset, ba+b_offset, clen);
703                         else
704                                 memcpy(ba+b_offset, pa+page_offset, clen);
705                         __bio_kunmap_atomic(ba, KM_USER0);
706                 }
707                 if (clen < len) /* hit end of page */
708                         break;
709                 page_offset +=  len;
710         }
711 }
712
713 #define check_xor()     do {                                            \
714                            if (count == MAX_XOR_BLOCKS) {               \
715                                 xor_block(count, STRIPE_SIZE, ptr);     \
716                                 count = 1;                              \
717                            }                                            \
718                         } while(0)
719
720
721 static void compute_block(struct stripe_head *sh, int dd_idx)
722 {
723         raid5_conf_t *conf = sh->raid_conf;
724         int i, count, disks = conf->raid_disks;
725         void *ptr[MAX_XOR_BLOCKS], *p;
726
727         PRINTK("compute_block, stripe %llu, idx %d\n", 
728                 (unsigned long long)sh->sector, dd_idx);
729
730         ptr[0] = page_address(sh->dev[dd_idx].page);
731         memset(ptr[0], 0, STRIPE_SIZE);
732         count = 1;
733         for (i = disks ; i--; ) {
734                 if (i == dd_idx)
735                         continue;
736                 p = page_address(sh->dev[i].page);
737                 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
738                         ptr[count++] = p;
739                 else
740                         printk("compute_block() %d, stripe %llu, %d"
741                                 " not present\n", dd_idx,
742                                 (unsigned long long)sh->sector, i);
743
744                 check_xor();
745         }
746         if (count != 1)
747                 xor_block(count, STRIPE_SIZE, ptr);
748         set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
749 }
750
751 static void compute_parity(struct stripe_head *sh, int method)
752 {
753         raid5_conf_t *conf = sh->raid_conf;
754         int i, pd_idx = sh->pd_idx, disks = conf->raid_disks, count;
755         void *ptr[MAX_XOR_BLOCKS];
756         struct bio *chosen;
757
758         PRINTK("compute_parity, stripe %llu, method %d\n",
759                 (unsigned long long)sh->sector, method);
760
761         count = 1;
762         ptr[0] = page_address(sh->dev[pd_idx].page);
763         switch(method) {
764         case READ_MODIFY_WRITE:
765                 if (!test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags))
766                         BUG();
767                 for (i=disks ; i-- ;) {
768                         if (i==pd_idx)
769                                 continue;
770                         if (sh->dev[i].towrite &&
771                             test_bit(R5_UPTODATE, &sh->dev[i].flags)) {
772                                 ptr[count++] = page_address(sh->dev[i].page);
773                                 chosen = sh->dev[i].towrite;
774                                 sh->dev[i].towrite = NULL;
775
776                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
777                                         wake_up(&conf->wait_for_overlap);
778
779                                 if (sh->dev[i].written) BUG();
780                                 sh->dev[i].written = chosen;
781                                 check_xor();
782                         }
783                 }
784                 break;
785         case RECONSTRUCT_WRITE:
786                 memset(ptr[0], 0, STRIPE_SIZE);
787                 for (i= disks; i-- ;)
788                         if (i!=pd_idx && sh->dev[i].towrite) {
789                                 chosen = sh->dev[i].towrite;
790                                 sh->dev[i].towrite = NULL;
791
792                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
793                                         wake_up(&conf->wait_for_overlap);
794
795                                 if (sh->dev[i].written) BUG();
796                                 sh->dev[i].written = chosen;
797                         }
798                 break;
799         case CHECK_PARITY:
800                 break;
801         }
802         if (count>1) {
803                 xor_block(count, STRIPE_SIZE, ptr);
804                 count = 1;
805         }
806         
807         for (i = disks; i--;)
808                 if (sh->dev[i].written) {
809                         sector_t sector = sh->dev[i].sector;
810                         struct bio *wbi = sh->dev[i].written;
811                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
812                                 copy_data(1, wbi, sh->dev[i].page, sector);
813                                 wbi = r5_next_bio(wbi, sector);
814                         }
815
816                         set_bit(R5_LOCKED, &sh->dev[i].flags);
817                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
818                 }
819
820         switch(method) {
821         case RECONSTRUCT_WRITE:
822         case CHECK_PARITY:
823                 for (i=disks; i--;)
824                         if (i != pd_idx) {
825                                 ptr[count++] = page_address(sh->dev[i].page);
826                                 check_xor();
827                         }
828                 break;
829         case READ_MODIFY_WRITE:
830                 for (i = disks; i--;)
831                         if (sh->dev[i].written) {
832                                 ptr[count++] = page_address(sh->dev[i].page);
833                                 check_xor();
834                         }
835         }
836         if (count != 1)
837                 xor_block(count, STRIPE_SIZE, ptr);
838         
839         if (method != CHECK_PARITY) {
840                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
841                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
842         } else
843                 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
844 }
845
846 /*
847  * Each stripe/dev can have one or more bion attached.
848  * toread/towrite point to the first in a chain. 
849  * The bi_next chain must be in order.
850  */
851 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
852 {
853         struct bio **bip;
854         raid5_conf_t *conf = sh->raid_conf;
855         int firstwrite=0;
856
857         PRINTK("adding bh b#%llu to stripe s#%llu\n",
858                 (unsigned long long)bi->bi_sector,
859                 (unsigned long long)sh->sector);
860
861
862         spin_lock(&sh->lock);
863         spin_lock_irq(&conf->device_lock);
864         if (forwrite) {
865                 bip = &sh->dev[dd_idx].towrite;
866                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
867                         firstwrite = 1;
868         } else
869                 bip = &sh->dev[dd_idx].toread;
870         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
871                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
872                         goto overlap;
873                 bip = & (*bip)->bi_next;
874         }
875         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
876                 goto overlap;
877
878         if (*bip && bi->bi_next && (*bip) != bi->bi_next)
879                 BUG();
880         if (*bip)
881                 bi->bi_next = *bip;
882         *bip = bi;
883         bi->bi_phys_segments ++;
884         spin_unlock_irq(&conf->device_lock);
885         spin_unlock(&sh->lock);
886
887         PRINTK("added bi b#%llu to stripe s#%llu, disk %d.\n",
888                 (unsigned long long)bi->bi_sector,
889                 (unsigned long long)sh->sector, dd_idx);
890
891         if (conf->mddev->bitmap && firstwrite) {
892                 sh->bm_seq = conf->seq_write;
893                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
894                                   STRIPE_SECTORS, 0);
895                 set_bit(STRIPE_BIT_DELAY, &sh->state);
896         }
897
898         if (forwrite) {
899                 /* check if page is covered */
900                 sector_t sector = sh->dev[dd_idx].sector;
901                 for (bi=sh->dev[dd_idx].towrite;
902                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
903                              bi && bi->bi_sector <= sector;
904                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
905                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
906                                 sector = bi->bi_sector + (bi->bi_size>>9);
907                 }
908                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
909                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
910         }
911         return 1;
912
913  overlap:
914         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
915         spin_unlock_irq(&conf->device_lock);
916         spin_unlock(&sh->lock);
917         return 0;
918 }
919
920
921 /*
922  * handle_stripe - do things to a stripe.
923  *
924  * We lock the stripe and then examine the state of various bits
925  * to see what needs to be done.
926  * Possible results:
927  *    return some read request which now have data
928  *    return some write requests which are safely on disc
929  *    schedule a read on some buffers
930  *    schedule a write of some buffers
931  *    return confirmation of parity correctness
932  *
933  * Parity calculations are done inside the stripe lock
934  * buffers are taken off read_list or write_list, and bh_cache buffers
935  * get BH_Lock set before the stripe lock is released.
936  *
937  */
938  
939 static void handle_stripe(struct stripe_head *sh)
940 {
941         raid5_conf_t *conf = sh->raid_conf;
942         int disks = conf->raid_disks;
943         struct bio *return_bi= NULL;
944         struct bio *bi;
945         int i;
946         int syncing;
947         int locked=0, uptodate=0, to_read=0, to_write=0, failed=0, written=0;
948         int non_overwrite = 0;
949         int failed_num=0;
950         struct r5dev *dev;
951
952         PRINTK("handling stripe %llu, cnt=%d, pd_idx=%d\n",
953                 (unsigned long long)sh->sector, atomic_read(&sh->count),
954                 sh->pd_idx);
955
956         spin_lock(&sh->lock);
957         clear_bit(STRIPE_HANDLE, &sh->state);
958         clear_bit(STRIPE_DELAYED, &sh->state);
959
960         syncing = test_bit(STRIPE_SYNCING, &sh->state);
961         /* Now to look around and see what can be done */
962
963         for (i=disks; i--; ) {
964                 mdk_rdev_t *rdev;
965                 dev = &sh->dev[i];
966                 clear_bit(R5_Insync, &dev->flags);
967                 clear_bit(R5_Syncio, &dev->flags);
968
969                 PRINTK("check %d: state 0x%lx read %p write %p written %p\n",
970                         i, dev->flags, dev->toread, dev->towrite, dev->written);
971                 /* maybe we can reply to a read */
972                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
973                         struct bio *rbi, *rbi2;
974                         PRINTK("Return read for disc %d\n", i);
975                         spin_lock_irq(&conf->device_lock);
976                         rbi = dev->toread;
977                         dev->toread = NULL;
978                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
979                                 wake_up(&conf->wait_for_overlap);
980                         spin_unlock_irq(&conf->device_lock);
981                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
982                                 copy_data(0, rbi, dev->page, dev->sector);
983                                 rbi2 = r5_next_bio(rbi, dev->sector);
984                                 spin_lock_irq(&conf->device_lock);
985                                 if (--rbi->bi_phys_segments == 0) {
986                                         rbi->bi_next = return_bi;
987                                         return_bi = rbi;
988                                 }
989                                 spin_unlock_irq(&conf->device_lock);
990                                 rbi = rbi2;
991                         }
992                 }
993
994                 /* now count some things */
995                 if (test_bit(R5_LOCKED, &dev->flags)) locked++;
996                 if (test_bit(R5_UPTODATE, &dev->flags)) uptodate++;
997
998                 
999                 if (dev->toread) to_read++;
1000                 if (dev->towrite) {
1001                         to_write++;
1002                         if (!test_bit(R5_OVERWRITE, &dev->flags))
1003                                 non_overwrite++;
1004                 }
1005                 if (dev->written) written++;
1006                 rdev = conf->disks[i].rdev; /* FIXME, should I be looking rdev */
1007                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
1008                         /* The ReadError flag wil just be confusing now */
1009                         clear_bit(R5_ReadError, &dev->flags);
1010                         clear_bit(R5_ReWrite, &dev->flags);
1011                 }
1012                 if (!rdev || !test_bit(In_sync, &rdev->flags)
1013                     || test_bit(R5_ReadError, &dev->flags)) {
1014                         failed++;
1015                         failed_num = i;
1016                 } else
1017                         set_bit(R5_Insync, &dev->flags);
1018         }
1019         PRINTK("locked=%d uptodate=%d to_read=%d"
1020                 " to_write=%d failed=%d failed_num=%d\n",
1021                 locked, uptodate, to_read, to_write, failed, failed_num);
1022         /* check if the array has lost two devices and, if so, some requests might
1023          * need to be failed
1024          */
1025         if (failed > 1 && to_read+to_write+written) {
1026                 for (i=disks; i--; ) {
1027                         int bitmap_end = 0;
1028
1029                         if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1030                                 mdk_rdev_t *rdev = conf->disks[i].rdev;
1031                                 if (rdev && test_bit(In_sync, &rdev->flags))
1032                                         /* multiple read failures in one stripe */
1033                                         md_error(conf->mddev, rdev);
1034                         }
1035
1036                         spin_lock_irq(&conf->device_lock);
1037                         /* fail all writes first */
1038                         bi = sh->dev[i].towrite;
1039                         sh->dev[i].towrite = NULL;
1040                         if (bi) { to_write--; bitmap_end = 1; }
1041
1042                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1043                                 wake_up(&conf->wait_for_overlap);
1044
1045                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1046                                 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1047                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1048                                 if (--bi->bi_phys_segments == 0) {
1049                                         md_write_end(conf->mddev);
1050                                         bi->bi_next = return_bi;
1051                                         return_bi = bi;
1052                                 }
1053                                 bi = nextbi;
1054                         }
1055                         /* and fail all 'written' */
1056                         bi = sh->dev[i].written;
1057                         sh->dev[i].written = NULL;
1058                         if (bi) bitmap_end = 1;
1059                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS) {
1060                                 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1061                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1062                                 if (--bi->bi_phys_segments == 0) {
1063                                         md_write_end(conf->mddev);
1064                                         bi->bi_next = return_bi;
1065                                         return_bi = bi;
1066                                 }
1067                                 bi = bi2;
1068                         }
1069
1070                         /* fail any reads if this device is non-operational */
1071                         if (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1072                             test_bit(R5_ReadError, &sh->dev[i].flags)) {
1073                                 bi = sh->dev[i].toread;
1074                                 sh->dev[i].toread = NULL;
1075                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1076                                         wake_up(&conf->wait_for_overlap);
1077                                 if (bi) to_read--;
1078                                 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1079                                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1080                                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1081                                         if (--bi->bi_phys_segments == 0) {
1082                                                 bi->bi_next = return_bi;
1083                                                 return_bi = bi;
1084                                         }
1085                                         bi = nextbi;
1086                                 }
1087                         }
1088                         spin_unlock_irq(&conf->device_lock);
1089                         if (bitmap_end)
1090                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1091                                                 STRIPE_SECTORS, 0, 0);
1092                 }
1093         }
1094         if (failed > 1 && syncing) {
1095                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
1096                 clear_bit(STRIPE_SYNCING, &sh->state);
1097                 syncing = 0;
1098         }
1099
1100         /* might be able to return some write requests if the parity block
1101          * is safe, or on a failed drive
1102          */
1103         dev = &sh->dev[sh->pd_idx];
1104         if ( written &&
1105              ( (test_bit(R5_Insync, &dev->flags) && !test_bit(R5_LOCKED, &dev->flags) &&
1106                 test_bit(R5_UPTODATE, &dev->flags))
1107                || (failed == 1 && failed_num == sh->pd_idx))
1108             ) {
1109             /* any written block on an uptodate or failed drive can be returned.
1110              * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but 
1111              * never LOCKED, so we don't need to test 'failed' directly.
1112              */
1113             for (i=disks; i--; )
1114                 if (sh->dev[i].written) {
1115                     dev = &sh->dev[i];
1116                     if (!test_bit(R5_LOCKED, &dev->flags) &&
1117                          test_bit(R5_UPTODATE, &dev->flags) ) {
1118                         /* We can return any write requests */
1119                             struct bio *wbi, *wbi2;
1120                             int bitmap_end = 0;
1121                             PRINTK("Return write for disc %d\n", i);
1122                             spin_lock_irq(&conf->device_lock);
1123                             wbi = dev->written;
1124                             dev->written = NULL;
1125                             while (wbi && wbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1126                                     wbi2 = r5_next_bio(wbi, dev->sector);
1127                                     if (--wbi->bi_phys_segments == 0) {
1128                                             md_write_end(conf->mddev);
1129                                             wbi->bi_next = return_bi;
1130                                             return_bi = wbi;
1131                                     }
1132                                     wbi = wbi2;
1133                             }
1134                             if (dev->towrite == NULL)
1135                                     bitmap_end = 1;
1136                             spin_unlock_irq(&conf->device_lock);
1137                             if (bitmap_end)
1138                                     bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1139                                                     STRIPE_SECTORS,
1140                                                     !test_bit(STRIPE_DEGRADED, &sh->state), 0);
1141                     }
1142                 }
1143         }
1144
1145         /* Now we might consider reading some blocks, either to check/generate
1146          * parity, or to satisfy requests
1147          * or to load a block that is being partially written.
1148          */
1149         if (to_read || non_overwrite || (syncing && (uptodate < disks))) {
1150                 for (i=disks; i--;) {
1151                         dev = &sh->dev[i];
1152                         if (!test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1153                             (dev->toread ||
1154                              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1155                              syncing ||
1156                              (failed && (sh->dev[failed_num].toread ||
1157                                          (sh->dev[failed_num].towrite && !test_bit(R5_OVERWRITE, &sh->dev[failed_num].flags))))
1158                                     )
1159                                 ) {
1160                                 /* we would like to get this block, possibly
1161                                  * by computing it, but we might not be able to
1162                                  */
1163                                 if (uptodate == disks-1) {
1164                                         PRINTK("Computing block %d\n", i);
1165                                         compute_block(sh, i);
1166                                         uptodate++;
1167                                 } else if (test_bit(R5_Insync, &dev->flags)) {
1168                                         set_bit(R5_LOCKED, &dev->flags);
1169                                         set_bit(R5_Wantread, &dev->flags);
1170 #if 0
1171                                         /* if I am just reading this block and we don't have
1172                                            a failed drive, or any pending writes then sidestep the cache */
1173                                         if (sh->bh_read[i] && !sh->bh_read[i]->b_reqnext &&
1174                                             ! syncing && !failed && !to_write) {
1175                                                 sh->bh_cache[i]->b_page =  sh->bh_read[i]->b_page;
1176                                                 sh->bh_cache[i]->b_data =  sh->bh_read[i]->b_data;
1177                                         }
1178 #endif
1179                                         locked++;
1180                                         PRINTK("Reading block %d (sync=%d)\n", 
1181                                                 i, syncing);
1182                                         if (syncing)
1183                                                 md_sync_acct(conf->disks[i].rdev->bdev,
1184                                                              STRIPE_SECTORS);
1185                                 }
1186                         }
1187                 }
1188                 set_bit(STRIPE_HANDLE, &sh->state);
1189         }
1190
1191         /* now to consider writing and what else, if anything should be read */
1192         if (to_write) {
1193                 int rmw=0, rcw=0;
1194                 for (i=disks ; i--;) {
1195                         /* would I have to read this buffer for read_modify_write */
1196                         dev = &sh->dev[i];
1197                         if ((dev->towrite || i == sh->pd_idx) &&
1198                             (!test_bit(R5_LOCKED, &dev->flags) 
1199 #if 0
1200 || sh->bh_page[i]!=bh->b_page
1201 #endif
1202                                     ) &&
1203                             !test_bit(R5_UPTODATE, &dev->flags)) {
1204                                 if (test_bit(R5_Insync, &dev->flags)
1205 /*                                  && !(!mddev->insync && i == sh->pd_idx) */
1206                                         )
1207                                         rmw++;
1208                                 else rmw += 2*disks;  /* cannot read it */
1209                         }
1210                         /* Would I have to read this buffer for reconstruct_write */
1211                         if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
1212                             (!test_bit(R5_LOCKED, &dev->flags) 
1213 #if 0
1214 || sh->bh_page[i] != bh->b_page
1215 #endif
1216                                     ) &&
1217                             !test_bit(R5_UPTODATE, &dev->flags)) {
1218                                 if (test_bit(R5_Insync, &dev->flags)) rcw++;
1219                                 else rcw += 2*disks;
1220                         }
1221                 }
1222                 PRINTK("for sector %llu, rmw=%d rcw=%d\n", 
1223                         (unsigned long long)sh->sector, rmw, rcw);
1224                 set_bit(STRIPE_HANDLE, &sh->state);
1225                 if (rmw < rcw && rmw > 0)
1226                         /* prefer read-modify-write, but need to get some data */
1227                         for (i=disks; i--;) {
1228                                 dev = &sh->dev[i];
1229                                 if ((dev->towrite || i == sh->pd_idx) &&
1230                                     !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1231                                     test_bit(R5_Insync, &dev->flags)) {
1232                                         if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1233                                         {
1234                                                 PRINTK("Read_old block %d for r-m-w\n", i);
1235                                                 set_bit(R5_LOCKED, &dev->flags);
1236                                                 set_bit(R5_Wantread, &dev->flags);
1237                                                 locked++;
1238                                         } else {
1239                                                 set_bit(STRIPE_DELAYED, &sh->state);
1240                                                 set_bit(STRIPE_HANDLE, &sh->state);
1241                                         }
1242                                 }
1243                         }
1244                 if (rcw <= rmw && rcw > 0)
1245                         /* want reconstruct write, but need to get some data */
1246                         for (i=disks; i--;) {
1247                                 dev = &sh->dev[i];
1248                                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
1249                                     !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1250                                     test_bit(R5_Insync, &dev->flags)) {
1251                                         if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1252                                         {
1253                                                 PRINTK("Read_old block %d for Reconstruct\n", i);
1254                                                 set_bit(R5_LOCKED, &dev->flags);
1255                                                 set_bit(R5_Wantread, &dev->flags);
1256                                                 locked++;
1257                                         } else {
1258                                                 set_bit(STRIPE_DELAYED, &sh->state);
1259                                                 set_bit(STRIPE_HANDLE, &sh->state);
1260                                         }
1261                                 }
1262                         }
1263                 /* now if nothing is locked, and if we have enough data, we can start a write request */
1264                 if (locked == 0 && (rcw == 0 ||rmw == 0) &&
1265                     !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
1266                         PRINTK("Computing parity...\n");
1267                         compute_parity(sh, rcw==0 ? RECONSTRUCT_WRITE : READ_MODIFY_WRITE);
1268                         /* now every locked buffer is ready to be written */
1269                         for (i=disks; i--;)
1270                                 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
1271                                         PRINTK("Writing block %d\n", i);
1272                                         locked++;
1273                                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
1274                                         if (!test_bit(R5_Insync, &sh->dev[i].flags)
1275                                             || (i==sh->pd_idx && failed == 0))
1276                                                 set_bit(STRIPE_INSYNC, &sh->state);
1277                                 }
1278                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
1279                                 atomic_dec(&conf->preread_active_stripes);
1280                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
1281                                         md_wakeup_thread(conf->mddev->thread);
1282                         }
1283                 }
1284         }
1285
1286         /* maybe we need to check and possibly fix the parity for this stripe
1287          * Any reads will already have been scheduled, so we just see if enough data
1288          * is available
1289          */
1290         if (syncing && locked == 0 &&
1291             !test_bit(STRIPE_INSYNC, &sh->state) && failed <= 1) {
1292                 set_bit(STRIPE_HANDLE, &sh->state);
1293                 if (failed == 0) {
1294                         char *pagea;
1295                         if (uptodate != disks)
1296                                 BUG();
1297                         compute_parity(sh, CHECK_PARITY);
1298                         uptodate--;
1299                         pagea = page_address(sh->dev[sh->pd_idx].page);
1300                         if ((*(u32*)pagea) == 0 &&
1301                             !memcmp(pagea, pagea+4, STRIPE_SIZE-4)) {
1302                                 /* parity is correct (on disc, not in buffer any more) */
1303                                 set_bit(STRIPE_INSYNC, &sh->state);
1304                         } else {
1305                                 conf->mddev->resync_mismatches += STRIPE_SECTORS;
1306                                 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
1307                                         /* don't try to repair!! */
1308                                         set_bit(STRIPE_INSYNC, &sh->state);
1309                         }
1310                 }
1311                 if (!test_bit(STRIPE_INSYNC, &sh->state)) {
1312                         if (failed==0)
1313                                 failed_num = sh->pd_idx;
1314                         /* should be able to compute the missing block and write it to spare */
1315                         if (!test_bit(R5_UPTODATE, &sh->dev[failed_num].flags)) {
1316                                 if (uptodate+1 != disks)
1317                                         BUG();
1318                                 compute_block(sh, failed_num);
1319                                 uptodate++;
1320                         }
1321                         if (uptodate != disks)
1322                                 BUG();
1323                         dev = &sh->dev[failed_num];
1324                         set_bit(R5_LOCKED, &dev->flags);
1325                         set_bit(R5_Wantwrite, &dev->flags);
1326                         clear_bit(STRIPE_DEGRADED, &sh->state);
1327                         locked++;
1328                         set_bit(STRIPE_INSYNC, &sh->state);
1329                         set_bit(R5_Syncio, &dev->flags);
1330                 }
1331         }
1332         if (syncing && locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
1333                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
1334                 clear_bit(STRIPE_SYNCING, &sh->state);
1335         }
1336
1337         /* If the failed drive is just a ReadError, then we might need to progress
1338          * the repair/check process
1339          */
1340         if (failed == 1 && ! conf->mddev->ro &&
1341             test_bit(R5_ReadError, &sh->dev[failed_num].flags)
1342             && !test_bit(R5_LOCKED, &sh->dev[failed_num].flags)
1343             && test_bit(R5_UPTODATE, &sh->dev[failed_num].flags)
1344                 ) {
1345                 dev = &sh->dev[failed_num];
1346                 if (!test_bit(R5_ReWrite, &dev->flags)) {
1347                         set_bit(R5_Wantwrite, &dev->flags);
1348                         set_bit(R5_ReWrite, &dev->flags);
1349                         set_bit(R5_LOCKED, &dev->flags);
1350                 } else {
1351                         /* let's read it back */
1352                         set_bit(R5_Wantread, &dev->flags);
1353                         set_bit(R5_LOCKED, &dev->flags);
1354                 }
1355         }
1356
1357         spin_unlock(&sh->lock);
1358
1359         while ((bi=return_bi)) {
1360                 int bytes = bi->bi_size;
1361
1362                 return_bi = bi->bi_next;
1363                 bi->bi_next = NULL;
1364                 bi->bi_size = 0;
1365                 bi->bi_end_io(bi, bytes, 0);
1366         }
1367         for (i=disks; i-- ;) {
1368                 int rw;
1369                 struct bio *bi;
1370                 mdk_rdev_t *rdev;
1371                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
1372                         rw = 1;
1373                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1374                         rw = 0;
1375                 else
1376                         continue;
1377  
1378                 bi = &sh->dev[i].req;
1379  
1380                 bi->bi_rw = rw;
1381                 if (rw)
1382                         bi->bi_end_io = raid5_end_write_request;
1383                 else
1384                         bi->bi_end_io = raid5_end_read_request;
1385  
1386                 rcu_read_lock();
1387                 rdev = rcu_dereference(conf->disks[i].rdev);
1388                 if (rdev && test_bit(Faulty, &rdev->flags))
1389                         rdev = NULL;
1390                 if (rdev)
1391                         atomic_inc(&rdev->nr_pending);
1392                 rcu_read_unlock();
1393  
1394                 if (rdev) {
1395                         if (test_bit(R5_Syncio, &sh->dev[i].flags))
1396                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1397
1398                         bi->bi_bdev = rdev->bdev;
1399                         PRINTK("for %llu schedule op %ld on disc %d\n",
1400                                 (unsigned long long)sh->sector, bi->bi_rw, i);
1401                         atomic_inc(&sh->count);
1402                         bi->bi_sector = sh->sector + rdev->data_offset;
1403                         bi->bi_flags = 1 << BIO_UPTODATE;
1404                         bi->bi_vcnt = 1;        
1405                         bi->bi_max_vecs = 1;
1406                         bi->bi_idx = 0;
1407                         bi->bi_io_vec = &sh->dev[i].vec;
1408                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1409                         bi->bi_io_vec[0].bv_offset = 0;
1410                         bi->bi_size = STRIPE_SIZE;
1411                         bi->bi_next = NULL;
1412                         generic_make_request(bi);
1413                 } else {
1414                         if (rw == 1)
1415                                 set_bit(STRIPE_DEGRADED, &sh->state);
1416                         PRINTK("skip op %ld on disc %d for sector %llu\n",
1417                                 bi->bi_rw, i, (unsigned long long)sh->sector);
1418                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1419                         set_bit(STRIPE_HANDLE, &sh->state);
1420                 }
1421         }
1422 }
1423
1424 static inline void raid5_activate_delayed(raid5_conf_t *conf)
1425 {
1426         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
1427                 while (!list_empty(&conf->delayed_list)) {
1428                         struct list_head *l = conf->delayed_list.next;
1429                         struct stripe_head *sh;
1430                         sh = list_entry(l, struct stripe_head, lru);
1431                         list_del_init(l);
1432                         clear_bit(STRIPE_DELAYED, &sh->state);
1433                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1434                                 atomic_inc(&conf->preread_active_stripes);
1435                         list_add_tail(&sh->lru, &conf->handle_list);
1436                 }
1437         }
1438 }
1439
1440 static inline void activate_bit_delay(raid5_conf_t *conf)
1441 {
1442         /* device_lock is held */
1443         struct list_head head;
1444         list_add(&head, &conf->bitmap_list);
1445         list_del_init(&conf->bitmap_list);
1446         while (!list_empty(&head)) {
1447                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
1448                 list_del_init(&sh->lru);
1449                 atomic_inc(&sh->count);
1450                 __release_stripe(conf, sh);
1451         }
1452 }
1453
1454 static void unplug_slaves(mddev_t *mddev)
1455 {
1456         raid5_conf_t *conf = mddev_to_conf(mddev);
1457         int i;
1458
1459         rcu_read_lock();
1460         for (i=0; i<mddev->raid_disks; i++) {
1461                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
1462                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
1463                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
1464
1465                         atomic_inc(&rdev->nr_pending);
1466                         rcu_read_unlock();
1467
1468                         if (r_queue->unplug_fn)
1469                                 r_queue->unplug_fn(r_queue);
1470
1471                         rdev_dec_pending(rdev, mddev);
1472                         rcu_read_lock();
1473                 }
1474         }
1475         rcu_read_unlock();
1476 }
1477
1478 static void raid5_unplug_device(request_queue_t *q)
1479 {
1480         mddev_t *mddev = q->queuedata;
1481         raid5_conf_t *conf = mddev_to_conf(mddev);
1482         unsigned long flags;
1483
1484         spin_lock_irqsave(&conf->device_lock, flags);
1485
1486         if (blk_remove_plug(q)) {
1487                 conf->seq_flush++;
1488                 raid5_activate_delayed(conf);
1489         }
1490         md_wakeup_thread(mddev->thread);
1491
1492         spin_unlock_irqrestore(&conf->device_lock, flags);
1493
1494         unplug_slaves(mddev);
1495 }
1496
1497 static int raid5_issue_flush(request_queue_t *q, struct gendisk *disk,
1498                              sector_t *error_sector)
1499 {
1500         mddev_t *mddev = q->queuedata;
1501         raid5_conf_t *conf = mddev_to_conf(mddev);
1502         int i, ret = 0;
1503
1504         rcu_read_lock();
1505         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
1506                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
1507                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
1508                         struct block_device *bdev = rdev->bdev;
1509                         request_queue_t *r_queue = bdev_get_queue(bdev);
1510
1511                         if (!r_queue->issue_flush_fn)
1512                                 ret = -EOPNOTSUPP;
1513                         else {
1514                                 atomic_inc(&rdev->nr_pending);
1515                                 rcu_read_unlock();
1516                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
1517                                                               error_sector);
1518                                 rdev_dec_pending(rdev, mddev);
1519                                 rcu_read_lock();
1520                         }
1521                 }
1522         }
1523         rcu_read_unlock();
1524         return ret;
1525 }
1526
1527 static inline void raid5_plug_device(raid5_conf_t *conf)
1528 {
1529         spin_lock_irq(&conf->device_lock);
1530         blk_plug_device(conf->mddev->queue);
1531         spin_unlock_irq(&conf->device_lock);
1532 }
1533
1534 static int make_request (request_queue_t *q, struct bio * bi)
1535 {
1536         mddev_t *mddev = q->queuedata;
1537         raid5_conf_t *conf = mddev_to_conf(mddev);
1538         const unsigned int raid_disks = conf->raid_disks;
1539         const unsigned int data_disks = raid_disks - 1;
1540         unsigned int dd_idx, pd_idx;
1541         sector_t new_sector;
1542         sector_t logical_sector, last_sector;
1543         struct stripe_head *sh;
1544         const int rw = bio_data_dir(bi);
1545
1546         if (unlikely(bio_barrier(bi))) {
1547                 bio_endio(bi, bi->bi_size, -EOPNOTSUPP);
1548                 return 0;
1549         }
1550
1551         md_write_start(mddev, bi);
1552
1553         disk_stat_inc(mddev->gendisk, ios[rw]);
1554         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
1555
1556         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
1557         last_sector = bi->bi_sector + (bi->bi_size>>9);
1558         bi->bi_next = NULL;
1559         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
1560
1561         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
1562                 DEFINE_WAIT(w);
1563                 
1564                 new_sector = raid5_compute_sector(logical_sector,
1565                                                   raid_disks, data_disks, &dd_idx, &pd_idx, conf);
1566
1567                 PRINTK("raid5: make_request, sector %llu logical %llu\n",
1568                         (unsigned long long)new_sector, 
1569                         (unsigned long long)logical_sector);
1570
1571         retry:
1572                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
1573                 sh = get_active_stripe(conf, new_sector, pd_idx, (bi->bi_rw&RWA_MASK));
1574                 if (sh) {
1575                         if (!add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
1576                                 /* Add failed due to overlap.  Flush everything
1577                                  * and wait a while
1578                                  */
1579                                 raid5_unplug_device(mddev->queue);
1580                                 release_stripe(sh);
1581                                 schedule();
1582                                 goto retry;
1583                         }
1584                         finish_wait(&conf->wait_for_overlap, &w);
1585                         raid5_plug_device(conf);
1586                         handle_stripe(sh);
1587                         release_stripe(sh);
1588
1589                 } else {
1590                         /* cannot get stripe for read-ahead, just give-up */
1591                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1592                         finish_wait(&conf->wait_for_overlap, &w);
1593                         break;
1594                 }
1595                         
1596         }
1597         spin_lock_irq(&conf->device_lock);
1598         if (--bi->bi_phys_segments == 0) {
1599                 int bytes = bi->bi_size;
1600
1601                 if ( bio_data_dir(bi) == WRITE )
1602                         md_write_end(mddev);
1603                 bi->bi_size = 0;
1604                 bi->bi_end_io(bi, bytes, 0);
1605         }
1606         spin_unlock_irq(&conf->device_lock);
1607         return 0;
1608 }
1609
1610 /* FIXME go_faster isn't used */
1611 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1612 {
1613         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1614         struct stripe_head *sh;
1615         int sectors_per_chunk = conf->chunk_size >> 9;
1616         sector_t x;
1617         unsigned long stripe;
1618         int chunk_offset;
1619         int dd_idx, pd_idx;
1620         sector_t first_sector;
1621         int raid_disks = conf->raid_disks;
1622         int data_disks = raid_disks-1;
1623         sector_t max_sector = mddev->size << 1;
1624         int sync_blocks;
1625
1626         if (sector_nr >= max_sector) {
1627                 /* just being told to finish up .. nothing much to do */
1628                 unplug_slaves(mddev);
1629
1630                 if (mddev->curr_resync < max_sector) /* aborted */
1631                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1632                                         &sync_blocks, 1);
1633                 else /* compelted sync */
1634                         conf->fullsync = 0;
1635                 bitmap_close_sync(mddev->bitmap);
1636
1637                 return 0;
1638         }
1639         /* if there is 1 or more failed drives and we are trying
1640          * to resync, then assert that we are finished, because there is
1641          * nothing we can do.
1642          */
1643         if (mddev->degraded >= 1 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1644                 sector_t rv = (mddev->size << 1) - sector_nr;
1645                 *skipped = 1;
1646                 return rv;
1647         }
1648         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1649             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1650             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
1651                 /* we can skip this block, and probably more */
1652                 sync_blocks /= STRIPE_SECTORS;
1653                 *skipped = 1;
1654                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
1655         }
1656
1657         x = sector_nr;
1658         chunk_offset = sector_div(x, sectors_per_chunk);
1659         stripe = x;
1660         BUG_ON(x != stripe);
1661
1662         first_sector = raid5_compute_sector((sector_t)stripe*data_disks*sectors_per_chunk
1663                 + chunk_offset, raid_disks, data_disks, &dd_idx, &pd_idx, conf);
1664         sh = get_active_stripe(conf, sector_nr, pd_idx, 1);
1665         if (sh == NULL) {
1666                 sh = get_active_stripe(conf, sector_nr, pd_idx, 0);
1667                 /* make sure we don't swamp the stripe cache if someone else
1668                  * is trying to get access 
1669                  */
1670                 schedule_timeout_uninterruptible(1);
1671         }
1672         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 0);
1673         spin_lock(&sh->lock);   
1674         set_bit(STRIPE_SYNCING, &sh->state);
1675         clear_bit(STRIPE_INSYNC, &sh->state);
1676         spin_unlock(&sh->lock);
1677
1678         handle_stripe(sh);
1679         release_stripe(sh);
1680
1681         return STRIPE_SECTORS;
1682 }
1683
1684 /*
1685  * This is our raid5 kernel thread.
1686  *
1687  * We scan the hash table for stripes which can be handled now.
1688  * During the scan, completed stripes are saved for us by the interrupt
1689  * handler, so that they will not have to wait for our next wakeup.
1690  */
1691 static void raid5d (mddev_t *mddev)
1692 {
1693         struct stripe_head *sh;
1694         raid5_conf_t *conf = mddev_to_conf(mddev);
1695         int handled;
1696
1697         PRINTK("+++ raid5d active\n");
1698
1699         md_check_recovery(mddev);
1700
1701         handled = 0;
1702         spin_lock_irq(&conf->device_lock);
1703         while (1) {
1704                 struct list_head *first;
1705
1706                 if (conf->seq_flush - conf->seq_write > 0) {
1707                         int seq = conf->seq_flush;
1708                         spin_unlock_irq(&conf->device_lock);
1709                         bitmap_unplug(mddev->bitmap);
1710                         spin_lock_irq(&conf->device_lock);
1711                         conf->seq_write = seq;
1712                         activate_bit_delay(conf);
1713                 }
1714
1715                 if (list_empty(&conf->handle_list) &&
1716                     atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD &&
1717                     !blk_queue_plugged(mddev->queue) &&
1718                     !list_empty(&conf->delayed_list))
1719                         raid5_activate_delayed(conf);
1720
1721                 if (list_empty(&conf->handle_list))
1722                         break;
1723
1724                 first = conf->handle_list.next;
1725                 sh = list_entry(first, struct stripe_head, lru);
1726
1727                 list_del_init(first);
1728                 atomic_inc(&sh->count);
1729                 if (atomic_read(&sh->count)!= 1)
1730                         BUG();
1731                 spin_unlock_irq(&conf->device_lock);
1732                 
1733                 handled++;
1734                 handle_stripe(sh);
1735                 release_stripe(sh);
1736
1737                 spin_lock_irq(&conf->device_lock);
1738         }
1739         PRINTK("%d stripes handled\n", handled);
1740
1741         spin_unlock_irq(&conf->device_lock);
1742
1743         unplug_slaves(mddev);
1744
1745         PRINTK("--- raid5d inactive\n");
1746 }
1747
1748 static ssize_t
1749 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
1750 {
1751         raid5_conf_t *conf = mddev_to_conf(mddev);
1752         if (conf)
1753                 return sprintf(page, "%d\n", conf->max_nr_stripes);
1754         else
1755                 return 0;
1756 }
1757
1758 static ssize_t
1759 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
1760 {
1761         raid5_conf_t *conf = mddev_to_conf(mddev);
1762         char *end;
1763         int new;
1764         if (len >= PAGE_SIZE)
1765                 return -EINVAL;
1766         if (!conf)
1767                 return -ENODEV;
1768
1769         new = simple_strtoul(page, &end, 10);
1770         if (!*page || (*end && *end != '\n') )
1771                 return -EINVAL;
1772         if (new <= 16 || new > 32768)
1773                 return -EINVAL;
1774         while (new < conf->max_nr_stripes) {
1775                 if (drop_one_stripe(conf))
1776                         conf->max_nr_stripes--;
1777                 else
1778                         break;
1779         }
1780         while (new > conf->max_nr_stripes) {
1781                 if (grow_one_stripe(conf))
1782                         conf->max_nr_stripes++;
1783                 else break;
1784         }
1785         return len;
1786 }
1787
1788 static struct md_sysfs_entry
1789 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
1790                                 raid5_show_stripe_cache_size,
1791                                 raid5_store_stripe_cache_size);
1792
1793 static ssize_t
1794 stripe_cache_active_show(mddev_t *mddev, char *page)
1795 {
1796         raid5_conf_t *conf = mddev_to_conf(mddev);
1797         if (conf)
1798                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
1799         else
1800                 return 0;
1801 }
1802
1803 static struct md_sysfs_entry
1804 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
1805
1806 static struct attribute *raid5_attrs[] =  {
1807         &raid5_stripecache_size.attr,
1808         &raid5_stripecache_active.attr,
1809         NULL,
1810 };
1811 static struct attribute_group raid5_attrs_group = {
1812         .name = NULL,
1813         .attrs = raid5_attrs,
1814 };
1815
1816 static int run(mddev_t *mddev)
1817 {
1818         raid5_conf_t *conf;
1819         int raid_disk, memory;
1820         mdk_rdev_t *rdev;
1821         struct disk_info *disk;
1822         struct list_head *tmp;
1823
1824         if (mddev->level != 5 && mddev->level != 4) {
1825                 printk("raid5: %s: raid level not set to 4/5 (%d)\n", mdname(mddev), mddev->level);
1826                 return -EIO;
1827         }
1828
1829         mddev->private = kmalloc (sizeof (raid5_conf_t)
1830                                   + mddev->raid_disks * sizeof(struct disk_info),
1831                                   GFP_KERNEL);
1832         if ((conf = mddev->private) == NULL)
1833                 goto abort;
1834         memset (conf, 0, sizeof (*conf) + mddev->raid_disks * sizeof(struct disk_info) );
1835         conf->mddev = mddev;
1836
1837         if ((conf->stripe_hashtbl = (struct stripe_head **) __get_free_pages(GFP_ATOMIC, HASH_PAGES_ORDER)) == NULL)
1838                 goto abort;
1839         memset(conf->stripe_hashtbl, 0, HASH_PAGES * PAGE_SIZE);
1840
1841         spin_lock_init(&conf->device_lock);
1842         init_waitqueue_head(&conf->wait_for_stripe);
1843         init_waitqueue_head(&conf->wait_for_overlap);
1844         INIT_LIST_HEAD(&conf->handle_list);
1845         INIT_LIST_HEAD(&conf->delayed_list);
1846         INIT_LIST_HEAD(&conf->bitmap_list);
1847         INIT_LIST_HEAD(&conf->inactive_list);
1848         atomic_set(&conf->active_stripes, 0);
1849         atomic_set(&conf->preread_active_stripes, 0);
1850
1851         PRINTK("raid5: run(%s) called.\n", mdname(mddev));
1852
1853         ITERATE_RDEV(mddev,rdev,tmp) {
1854                 raid_disk = rdev->raid_disk;
1855                 if (raid_disk >= mddev->raid_disks
1856                     || raid_disk < 0)
1857                         continue;
1858                 disk = conf->disks + raid_disk;
1859
1860                 disk->rdev = rdev;
1861
1862                 if (test_bit(In_sync, &rdev->flags)) {
1863                         char b[BDEVNAME_SIZE];
1864                         printk(KERN_INFO "raid5: device %s operational as raid"
1865                                 " disk %d\n", bdevname(rdev->bdev,b),
1866                                 raid_disk);
1867                         conf->working_disks++;
1868                 }
1869         }
1870
1871         conf->raid_disks = mddev->raid_disks;
1872         /*
1873          * 0 for a fully functional array, 1 for a degraded array.
1874          */
1875         mddev->degraded = conf->failed_disks = conf->raid_disks - conf->working_disks;
1876         conf->mddev = mddev;
1877         conf->chunk_size = mddev->chunk_size;
1878         conf->level = mddev->level;
1879         conf->algorithm = mddev->layout;
1880         conf->max_nr_stripes = NR_STRIPES;
1881
1882         /* device size must be a multiple of chunk size */
1883         mddev->size &= ~(mddev->chunk_size/1024 -1);
1884         mddev->resync_max_sectors = mddev->size << 1;
1885
1886         if (!conf->chunk_size || conf->chunk_size % 4) {
1887                 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
1888                         conf->chunk_size, mdname(mddev));
1889                 goto abort;
1890         }
1891         if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
1892                 printk(KERN_ERR 
1893                         "raid5: unsupported parity algorithm %d for %s\n",
1894                         conf->algorithm, mdname(mddev));
1895                 goto abort;
1896         }
1897         if (mddev->degraded > 1) {
1898                 printk(KERN_ERR "raid5: not enough operational devices for %s"
1899                         " (%d/%d failed)\n",
1900                         mdname(mddev), conf->failed_disks, conf->raid_disks);
1901                 goto abort;
1902         }
1903
1904         if (mddev->degraded == 1 &&
1905             mddev->recovery_cp != MaxSector) {
1906                 printk(KERN_ERR 
1907                         "raid5: cannot start dirty degraded array for %s\n",
1908                         mdname(mddev));
1909                 goto abort;
1910         }
1911
1912         {
1913                 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
1914                 if (!mddev->thread) {
1915                         printk(KERN_ERR 
1916                                 "raid5: couldn't allocate thread for %s\n",
1917                                 mdname(mddev));
1918                         goto abort;
1919                 }
1920         }
1921         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
1922                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
1923         if (grow_stripes(conf, conf->max_nr_stripes)) {
1924                 printk(KERN_ERR 
1925                         "raid5: couldn't allocate %dkB for buffers\n", memory);
1926                 shrink_stripes(conf);
1927                 md_unregister_thread(mddev->thread);
1928                 goto abort;
1929         } else
1930                 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
1931                         memory, mdname(mddev));
1932
1933         if (mddev->degraded == 0)
1934                 printk("raid5: raid level %d set %s active with %d out of %d"
1935                         " devices, algorithm %d\n", conf->level, mdname(mddev), 
1936                         mddev->raid_disks-mddev->degraded, mddev->raid_disks,
1937                         conf->algorithm);
1938         else
1939                 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
1940                         " out of %d devices, algorithm %d\n", conf->level,
1941                         mdname(mddev), mddev->raid_disks - mddev->degraded,
1942                         mddev->raid_disks, conf->algorithm);
1943
1944         print_raid5_conf(conf);
1945
1946         /* read-ahead size must cover two whole stripes, which is
1947          * 2 * (n-1) * chunksize where 'n' is the number of raid devices
1948          */
1949         {
1950                 int stripe = (mddev->raid_disks-1) * mddev->chunk_size
1951                         / PAGE_CACHE_SIZE;
1952                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
1953                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
1954         }
1955
1956         /* Ok, everything is just fine now */
1957         sysfs_create_group(&mddev->kobj, &raid5_attrs_group);
1958
1959         if (mddev->bitmap)
1960                 mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
1961
1962         mddev->queue->unplug_fn = raid5_unplug_device;
1963         mddev->queue->issue_flush_fn = raid5_issue_flush;
1964
1965         mddev->array_size =  mddev->size * (mddev->raid_disks - 1);
1966         return 0;
1967 abort:
1968         if (conf) {
1969                 print_raid5_conf(conf);
1970                 if (conf->stripe_hashtbl)
1971                         free_pages((unsigned long) conf->stripe_hashtbl,
1972                                                         HASH_PAGES_ORDER);
1973                 kfree(conf);
1974         }
1975         mddev->private = NULL;
1976         printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
1977         return -EIO;
1978 }
1979
1980
1981
1982 static int stop(mddev_t *mddev)
1983 {
1984         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1985
1986         md_unregister_thread(mddev->thread);
1987         mddev->thread = NULL;
1988         shrink_stripes(conf);
1989         free_pages((unsigned long) conf->stripe_hashtbl, HASH_PAGES_ORDER);
1990         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
1991         sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
1992         kfree(conf);
1993         mddev->private = NULL;
1994         return 0;
1995 }
1996
1997 #if RAID5_DEBUG
1998 static void print_sh (struct stripe_head *sh)
1999 {
2000         int i;
2001
2002         printk("sh %llu, pd_idx %d, state %ld.\n",
2003                 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
2004         printk("sh %llu,  count %d.\n",
2005                 (unsigned long long)sh->sector, atomic_read(&sh->count));
2006         printk("sh %llu, ", (unsigned long long)sh->sector);
2007         for (i = 0; i < sh->raid_conf->raid_disks; i++) {
2008                 printk("(cache%d: %p %ld) ", 
2009                         i, sh->dev[i].page, sh->dev[i].flags);
2010         }
2011         printk("\n");
2012 }
2013
2014 static void printall (raid5_conf_t *conf)
2015 {
2016         struct stripe_head *sh;
2017         int i;
2018
2019         spin_lock_irq(&conf->device_lock);
2020         for (i = 0; i < NR_HASH; i++) {
2021                 sh = conf->stripe_hashtbl[i];
2022                 for (; sh; sh = sh->hash_next) {
2023                         if (sh->raid_conf != conf)
2024                                 continue;
2025                         print_sh(sh);
2026                 }
2027         }
2028         spin_unlock_irq(&conf->device_lock);
2029 }
2030 #endif
2031
2032 static void status (struct seq_file *seq, mddev_t *mddev)
2033 {
2034         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
2035         int i;
2036
2037         seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
2038         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->working_disks);
2039         for (i = 0; i < conf->raid_disks; i++)
2040                 seq_printf (seq, "%s",
2041                                conf->disks[i].rdev &&
2042                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
2043         seq_printf (seq, "]");
2044 #if RAID5_DEBUG
2045 #define D(x) \
2046         seq_printf (seq, "<"#x":%d>", atomic_read(&conf->x))
2047         printall(conf);
2048 #endif
2049 }
2050
2051 static void print_raid5_conf (raid5_conf_t *conf)
2052 {
2053         int i;
2054         struct disk_info *tmp;
2055
2056         printk("RAID5 conf printout:\n");
2057         if (!conf) {
2058                 printk("(conf==NULL)\n");
2059                 return;
2060         }
2061         printk(" --- rd:%d wd:%d fd:%d\n", conf->raid_disks,
2062                  conf->working_disks, conf->failed_disks);
2063
2064         for (i = 0; i < conf->raid_disks; i++) {
2065                 char b[BDEVNAME_SIZE];
2066                 tmp = conf->disks + i;
2067                 if (tmp->rdev)
2068                 printk(" disk %d, o:%d, dev:%s\n",
2069                         i, !test_bit(Faulty, &tmp->rdev->flags),
2070                         bdevname(tmp->rdev->bdev,b));
2071         }
2072 }
2073
2074 static int raid5_spare_active(mddev_t *mddev)
2075 {
2076         int i;
2077         raid5_conf_t *conf = mddev->private;
2078         struct disk_info *tmp;
2079
2080         for (i = 0; i < conf->raid_disks; i++) {
2081                 tmp = conf->disks + i;
2082                 if (tmp->rdev
2083                     && !test_bit(Faulty, &tmp->rdev->flags)
2084                     && !test_bit(In_sync, &tmp->rdev->flags)) {
2085                         mddev->degraded--;
2086                         conf->failed_disks--;
2087                         conf->working_disks++;
2088                         set_bit(In_sync, &tmp->rdev->flags);
2089                 }
2090         }
2091         print_raid5_conf(conf);
2092         return 0;
2093 }
2094
2095 static int raid5_remove_disk(mddev_t *mddev, int number)
2096 {
2097         raid5_conf_t *conf = mddev->private;
2098         int err = 0;
2099         mdk_rdev_t *rdev;
2100         struct disk_info *p = conf->disks + number;
2101
2102         print_raid5_conf(conf);
2103         rdev = p->rdev;
2104         if (rdev) {
2105                 if (test_bit(In_sync, &rdev->flags) ||
2106                     atomic_read(&rdev->nr_pending)) {
2107                         err = -EBUSY;
2108                         goto abort;
2109                 }
2110                 p->rdev = NULL;
2111                 synchronize_rcu();
2112                 if (atomic_read(&rdev->nr_pending)) {
2113                         /* lost the race, try later */
2114                         err = -EBUSY;
2115                         p->rdev = rdev;
2116                 }
2117         }
2118 abort:
2119
2120         print_raid5_conf(conf);
2121         return err;
2122 }
2123
2124 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
2125 {
2126         raid5_conf_t *conf = mddev->private;
2127         int found = 0;
2128         int disk;
2129         struct disk_info *p;
2130
2131         if (mddev->degraded > 1)
2132                 /* no point adding a device */
2133                 return 0;
2134
2135         /*
2136          * find the disk ...
2137          */
2138         for (disk=0; disk < mddev->raid_disks; disk++)
2139                 if ((p=conf->disks + disk)->rdev == NULL) {
2140                         clear_bit(In_sync, &rdev->flags);
2141                         rdev->raid_disk = disk;
2142                         found = 1;
2143                         if (rdev->saved_raid_disk != disk)
2144                                 conf->fullsync = 1;
2145                         rcu_assign_pointer(p->rdev, rdev);
2146                         break;
2147                 }
2148         print_raid5_conf(conf);
2149         return found;
2150 }
2151
2152 static int raid5_resize(mddev_t *mddev, sector_t sectors)
2153 {
2154         /* no resync is happening, and there is enough space
2155          * on all devices, so we can resize.
2156          * We need to make sure resync covers any new space.
2157          * If the array is shrinking we should possibly wait until
2158          * any io in the removed space completes, but it hardly seems
2159          * worth it.
2160          */
2161         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
2162         mddev->array_size = (sectors * (mddev->raid_disks-1))>>1;
2163         set_capacity(mddev->gendisk, mddev->array_size << 1);
2164         mddev->changed = 1;
2165         if (sectors/2  > mddev->size && mddev->recovery_cp == MaxSector) {
2166                 mddev->recovery_cp = mddev->size << 1;
2167                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2168         }
2169         mddev->size = sectors /2;
2170         mddev->resync_max_sectors = sectors;
2171         return 0;
2172 }
2173
2174 static void raid5_quiesce(mddev_t *mddev, int state)
2175 {
2176         raid5_conf_t *conf = mddev_to_conf(mddev);
2177
2178         switch(state) {
2179         case 1: /* stop all writes */
2180                 spin_lock_irq(&conf->device_lock);
2181                 conf->quiesce = 1;
2182                 wait_event_lock_irq(conf->wait_for_stripe,
2183                                     atomic_read(&conf->active_stripes) == 0,
2184                                     conf->device_lock, /* nothing */);
2185                 spin_unlock_irq(&conf->device_lock);
2186                 break;
2187
2188         case 0: /* re-enable writes */
2189                 spin_lock_irq(&conf->device_lock);
2190                 conf->quiesce = 0;
2191                 wake_up(&conf->wait_for_stripe);
2192                 spin_unlock_irq(&conf->device_lock);
2193                 break;
2194         }
2195         if (mddev->thread) {
2196                 if (mddev->bitmap)
2197                         mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
2198                 else
2199                         mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
2200                 md_wakeup_thread(mddev->thread);
2201         }
2202 }
2203 static mdk_personality_t raid5_personality=
2204 {
2205         .name           = "raid5",
2206         .owner          = THIS_MODULE,
2207         .make_request   = make_request,
2208         .run            = run,
2209         .stop           = stop,
2210         .status         = status,
2211         .error_handler  = error,
2212         .hot_add_disk   = raid5_add_disk,
2213         .hot_remove_disk= raid5_remove_disk,
2214         .spare_active   = raid5_spare_active,
2215         .sync_request   = sync_request,
2216         .resize         = raid5_resize,
2217         .quiesce        = raid5_quiesce,
2218 };
2219
2220 static int __init raid5_init (void)
2221 {
2222         return register_md_personality (RAID5, &raid5_personality);
2223 }
2224
2225 static void raid5_exit (void)
2226 {
2227         unregister_md_personality (RAID5);
2228 }
2229
2230 module_init(raid5_init);
2231 module_exit(raid5_exit);
2232 MODULE_LICENSE("GPL");
2233 MODULE_ALIAS("md-personality-4"); /* RAID5 */