Pull sn2-mmio-writes into release branch
[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 NR_HASH                 (PAGE_SIZE / sizeof(struct hlist_head))
39 #define HASH_MASK               (NR_HASH - 1)
40
41 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
42
43 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
44  * order without overlap.  There may be several bio's per stripe+device, and
45  * a bio could span several devices.
46  * When walking this list for a particular stripe+device, we must never proceed
47  * beyond a bio that extends past this device, as the next bio might no longer
48  * be valid.
49  * This macro is used to determine the 'next' bio in the list, given the sector
50  * of the current stripe+device
51  */
52 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
53 /*
54  * The following can be used to debug the driver
55  */
56 #define RAID5_DEBUG     0
57 #define RAID5_PARANOIA  1
58 #if RAID5_PARANOIA && defined(CONFIG_SMP)
59 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
60 #else
61 # define CHECK_DEVLOCK()
62 #endif
63
64 #define PRINTK(x...) ((void)(RAID5_DEBUG && printk(x)))
65 #if RAID5_DEBUG
66 #define inline
67 #define __inline__
68 #endif
69
70 static void print_raid5_conf (raid5_conf_t *conf);
71
72 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
73 {
74         if (atomic_dec_and_test(&sh->count)) {
75                 if (!list_empty(&sh->lru))
76                         BUG();
77                 if (atomic_read(&conf->active_stripes)==0)
78                         BUG();
79                 if (test_bit(STRIPE_HANDLE, &sh->state)) {
80                         if (test_bit(STRIPE_DELAYED, &sh->state))
81                                 list_add_tail(&sh->lru, &conf->delayed_list);
82                         else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
83                                  conf->seq_write == sh->bm_seq)
84                                 list_add_tail(&sh->lru, &conf->bitmap_list);
85                         else {
86                                 clear_bit(STRIPE_BIT_DELAY, &sh->state);
87                                 list_add_tail(&sh->lru, &conf->handle_list);
88                         }
89                         md_wakeup_thread(conf->mddev->thread);
90                 } else {
91                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
92                                 atomic_dec(&conf->preread_active_stripes);
93                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
94                                         md_wakeup_thread(conf->mddev->thread);
95                         }
96                         list_add_tail(&sh->lru, &conf->inactive_list);
97                         atomic_dec(&conf->active_stripes);
98                         if (!conf->inactive_blocked ||
99                             atomic_read(&conf->active_stripes) < (conf->max_nr_stripes*3/4))
100                                 wake_up(&conf->wait_for_stripe);
101                 }
102         }
103 }
104 static void release_stripe(struct stripe_head *sh)
105 {
106         raid5_conf_t *conf = sh->raid_conf;
107         unsigned long flags;
108         
109         spin_lock_irqsave(&conf->device_lock, flags);
110         __release_stripe(conf, sh);
111         spin_unlock_irqrestore(&conf->device_lock, flags);
112 }
113
114 static inline void remove_hash(struct stripe_head *sh)
115 {
116         PRINTK("remove_hash(), stripe %llu\n", (unsigned long long)sh->sector);
117
118         hlist_del_init(&sh->hash);
119 }
120
121 static void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
122 {
123         struct hlist_head *hp = stripe_hash(conf, sh->sector);
124
125         PRINTK("insert_hash(), stripe %llu\n", (unsigned long long)sh->sector);
126
127         CHECK_DEVLOCK();
128         hlist_add_head(&sh->hash, hp);
129 }
130
131
132 /* find an idle stripe, make sure it is unhashed, and return it. */
133 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
134 {
135         struct stripe_head *sh = NULL;
136         struct list_head *first;
137
138         CHECK_DEVLOCK();
139         if (list_empty(&conf->inactive_list))
140                 goto out;
141         first = conf->inactive_list.next;
142         sh = list_entry(first, struct stripe_head, lru);
143         list_del_init(first);
144         remove_hash(sh);
145         atomic_inc(&conf->active_stripes);
146 out:
147         return sh;
148 }
149
150 static void shrink_buffers(struct stripe_head *sh, int num)
151 {
152         struct page *p;
153         int i;
154
155         for (i=0; i<num ; i++) {
156                 p = sh->dev[i].page;
157                 if (!p)
158                         continue;
159                 sh->dev[i].page = NULL;
160                 put_page(p);
161         }
162 }
163
164 static int grow_buffers(struct stripe_head *sh, int num)
165 {
166         int i;
167
168         for (i=0; i<num; i++) {
169                 struct page *page;
170
171                 if (!(page = alloc_page(GFP_KERNEL))) {
172                         return 1;
173                 }
174                 sh->dev[i].page = page;
175         }
176         return 0;
177 }
178
179 static void raid5_build_block (struct stripe_head *sh, int i);
180
181 static void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx)
182 {
183         raid5_conf_t *conf = sh->raid_conf;
184         int disks = conf->raid_disks, i;
185
186         if (atomic_read(&sh->count) != 0)
187                 BUG();
188         if (test_bit(STRIPE_HANDLE, &sh->state))
189                 BUG();
190         
191         CHECK_DEVLOCK();
192         PRINTK("init_stripe called, stripe %llu\n", 
193                 (unsigned long long)sh->sector);
194
195         remove_hash(sh);
196         
197         sh->sector = sector;
198         sh->pd_idx = pd_idx;
199         sh->state = 0;
200
201         for (i=disks; i--; ) {
202                 struct r5dev *dev = &sh->dev[i];
203
204                 if (dev->toread || dev->towrite || dev->written ||
205                     test_bit(R5_LOCKED, &dev->flags)) {
206                         printk("sector=%llx i=%d %p %p %p %d\n",
207                                (unsigned long long)sh->sector, i, dev->toread,
208                                dev->towrite, dev->written,
209                                test_bit(R5_LOCKED, &dev->flags));
210                         BUG();
211                 }
212                 dev->flags = 0;
213                 raid5_build_block(sh, i);
214         }
215         insert_hash(conf, sh);
216 }
217
218 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector)
219 {
220         struct stripe_head *sh;
221         struct hlist_node *hn;
222
223         CHECK_DEVLOCK();
224         PRINTK("__find_stripe, sector %llu\n", (unsigned long long)sector);
225         hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
226                 if (sh->sector == sector)
227                         return sh;
228         PRINTK("__stripe %llu not in cache\n", (unsigned long long)sector);
229         return NULL;
230 }
231
232 static void unplug_slaves(mddev_t *mddev);
233 static void raid5_unplug_device(request_queue_t *q);
234
235 static struct stripe_head *get_active_stripe(raid5_conf_t *conf, sector_t sector,
236                                              int pd_idx, int noblock) 
237 {
238         struct stripe_head *sh;
239
240         PRINTK("get_stripe, sector %llu\n", (unsigned long long)sector);
241
242         spin_lock_irq(&conf->device_lock);
243
244         do {
245                 wait_event_lock_irq(conf->wait_for_stripe,
246                                     conf->quiesce == 0,
247                                     conf->device_lock, /* nothing */);
248                 sh = __find_stripe(conf, sector);
249                 if (!sh) {
250                         if (!conf->inactive_blocked)
251                                 sh = get_free_stripe(conf);
252                         if (noblock && sh == NULL)
253                                 break;
254                         if (!sh) {
255                                 conf->inactive_blocked = 1;
256                                 wait_event_lock_irq(conf->wait_for_stripe,
257                                                     !list_empty(&conf->inactive_list) &&
258                                                     (atomic_read(&conf->active_stripes)
259                                                      < (conf->max_nr_stripes *3/4)
260                                                      || !conf->inactive_blocked),
261                                                     conf->device_lock,
262                                                     unplug_slaves(conf->mddev);
263                                         );
264                                 conf->inactive_blocked = 0;
265                         } else
266                                 init_stripe(sh, sector, pd_idx);
267                 } else {
268                         if (atomic_read(&sh->count)) {
269                                 if (!list_empty(&sh->lru))
270                                         BUG();
271                         } else {
272                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
273                                         atomic_inc(&conf->active_stripes);
274                                 if (list_empty(&sh->lru))
275                                         BUG();
276                                 list_del_init(&sh->lru);
277                         }
278                 }
279         } while (sh == NULL);
280
281         if (sh)
282                 atomic_inc(&sh->count);
283
284         spin_unlock_irq(&conf->device_lock);
285         return sh;
286 }
287
288 static int grow_one_stripe(raid5_conf_t *conf)
289 {
290         struct stripe_head *sh;
291         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
292         if (!sh)
293                 return 0;
294         memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
295         sh->raid_conf = conf;
296         spin_lock_init(&sh->lock);
297
298         if (grow_buffers(sh, conf->raid_disks)) {
299                 shrink_buffers(sh, conf->raid_disks);
300                 kmem_cache_free(conf->slab_cache, sh);
301                 return 0;
302         }
303         /* we just created an active stripe so... */
304         atomic_set(&sh->count, 1);
305         atomic_inc(&conf->active_stripes);
306         INIT_LIST_HEAD(&sh->lru);
307         release_stripe(sh);
308         return 1;
309 }
310
311 static int grow_stripes(raid5_conf_t *conf, int num)
312 {
313         kmem_cache_t *sc;
314         int devs = conf->raid_disks;
315
316         sprintf(conf->cache_name, "raid5/%s", mdname(conf->mddev));
317
318         sc = kmem_cache_create(conf->cache_name, 
319                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
320                                0, 0, NULL, NULL);
321         if (!sc)
322                 return 1;
323         conf->slab_cache = sc;
324         while (num--) {
325                 if (!grow_one_stripe(conf))
326                         return 1;
327         }
328         return 0;
329 }
330
331 static int drop_one_stripe(raid5_conf_t *conf)
332 {
333         struct stripe_head *sh;
334
335         spin_lock_irq(&conf->device_lock);
336         sh = get_free_stripe(conf);
337         spin_unlock_irq(&conf->device_lock);
338         if (!sh)
339                 return 0;
340         if (atomic_read(&sh->count))
341                 BUG();
342         shrink_buffers(sh, conf->raid_disks);
343         kmem_cache_free(conf->slab_cache, sh);
344         atomic_dec(&conf->active_stripes);
345         return 1;
346 }
347
348 static void shrink_stripes(raid5_conf_t *conf)
349 {
350         while (drop_one_stripe(conf))
351                 ;
352
353         if (conf->slab_cache)
354                 kmem_cache_destroy(conf->slab_cache);
355         conf->slab_cache = NULL;
356 }
357
358 static int raid5_end_read_request(struct bio * bi, unsigned int bytes_done,
359                                    int error)
360 {
361         struct stripe_head *sh = bi->bi_private;
362         raid5_conf_t *conf = sh->raid_conf;
363         int disks = conf->raid_disks, i;
364         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
365
366         if (bi->bi_size)
367                 return 1;
368
369         for (i=0 ; i<disks; i++)
370                 if (bi == &sh->dev[i].req)
371                         break;
372
373         PRINTK("end_read_request %llu/%d, count: %d, uptodate %d.\n", 
374                 (unsigned long long)sh->sector, i, atomic_read(&sh->count), 
375                 uptodate);
376         if (i == disks) {
377                 BUG();
378                 return 0;
379         }
380
381         if (uptodate) {
382 #if 0
383                 struct bio *bio;
384                 unsigned long flags;
385                 spin_lock_irqsave(&conf->device_lock, flags);
386                 /* we can return a buffer if we bypassed the cache or
387                  * if the top buffer is not in highmem.  If there are
388                  * multiple buffers, leave the extra work to
389                  * handle_stripe
390                  */
391                 buffer = sh->bh_read[i];
392                 if (buffer &&
393                     (!PageHighMem(buffer->b_page)
394                      || buffer->b_page == bh->b_page )
395                         ) {
396                         sh->bh_read[i] = buffer->b_reqnext;
397                         buffer->b_reqnext = NULL;
398                 } else
399                         buffer = NULL;
400                 spin_unlock_irqrestore(&conf->device_lock, flags);
401                 if (sh->bh_page[i]==bh->b_page)
402                         set_buffer_uptodate(bh);
403                 if (buffer) {
404                         if (buffer->b_page != bh->b_page)
405                                 memcpy(buffer->b_data, bh->b_data, bh->b_size);
406                         buffer->b_end_io(buffer, 1);
407                 }
408 #else
409                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
410 #endif
411                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
412                         printk(KERN_INFO "raid5: read error corrected!!\n");
413                         clear_bit(R5_ReadError, &sh->dev[i].flags);
414                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
415                 }
416                 if (atomic_read(&conf->disks[i].rdev->read_errors))
417                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
418         } else {
419                 int retry = 0;
420                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
421                 atomic_inc(&conf->disks[i].rdev->read_errors);
422                 if (conf->mddev->degraded)
423                         printk(KERN_WARNING "raid5: read error not correctable.\n");
424                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
425                         /* Oh, no!!! */
426                         printk(KERN_WARNING "raid5: read error NOT corrected!!\n");
427                 else if (atomic_read(&conf->disks[i].rdev->read_errors)
428                          > conf->max_nr_stripes)
429                         printk(KERN_WARNING
430                                "raid5: Too many read errors, failing device.\n");
431                 else
432                         retry = 1;
433                 if (retry)
434                         set_bit(R5_ReadError, &sh->dev[i].flags);
435                 else {
436                         clear_bit(R5_ReadError, &sh->dev[i].flags);
437                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
438                         md_error(conf->mddev, conf->disks[i].rdev);
439                 }
440         }
441         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
442 #if 0
443         /* must restore b_page before unlocking buffer... */
444         if (sh->bh_page[i] != bh->b_page) {
445                 bh->b_page = sh->bh_page[i];
446                 bh->b_data = page_address(bh->b_page);
447                 clear_buffer_uptodate(bh);
448         }
449 #endif
450         clear_bit(R5_LOCKED, &sh->dev[i].flags);
451         set_bit(STRIPE_HANDLE, &sh->state);
452         release_stripe(sh);
453         return 0;
454 }
455
456 static int raid5_end_write_request (struct bio *bi, unsigned int bytes_done,
457                                     int error)
458 {
459         struct stripe_head *sh = bi->bi_private;
460         raid5_conf_t *conf = sh->raid_conf;
461         int disks = conf->raid_disks, i;
462         unsigned long flags;
463         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
464
465         if (bi->bi_size)
466                 return 1;
467
468         for (i=0 ; i<disks; i++)
469                 if (bi == &sh->dev[i].req)
470                         break;
471
472         PRINTK("end_write_request %llu/%d, count %d, uptodate: %d.\n", 
473                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
474                 uptodate);
475         if (i == disks) {
476                 BUG();
477                 return 0;
478         }
479
480         spin_lock_irqsave(&conf->device_lock, flags);
481         if (!uptodate)
482                 md_error(conf->mddev, conf->disks[i].rdev);
483
484         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
485         
486         clear_bit(R5_LOCKED, &sh->dev[i].flags);
487         set_bit(STRIPE_HANDLE, &sh->state);
488         __release_stripe(conf, sh);
489         spin_unlock_irqrestore(&conf->device_lock, flags);
490         return 0;
491 }
492
493
494 static sector_t compute_blocknr(struct stripe_head *sh, int i);
495         
496 static void raid5_build_block (struct stripe_head *sh, int i)
497 {
498         struct r5dev *dev = &sh->dev[i];
499
500         bio_init(&dev->req);
501         dev->req.bi_io_vec = &dev->vec;
502         dev->req.bi_vcnt++;
503         dev->req.bi_max_vecs++;
504         dev->vec.bv_page = dev->page;
505         dev->vec.bv_len = STRIPE_SIZE;
506         dev->vec.bv_offset = 0;
507
508         dev->req.bi_sector = sh->sector;
509         dev->req.bi_private = sh;
510
511         dev->flags = 0;
512         if (i != sh->pd_idx)
513                 dev->sector = compute_blocknr(sh, i);
514 }
515
516 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
517 {
518         char b[BDEVNAME_SIZE];
519         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
520         PRINTK("raid5: error called\n");
521
522         if (!test_bit(Faulty, &rdev->flags)) {
523                 mddev->sb_dirty = 1;
524                 if (test_bit(In_sync, &rdev->flags)) {
525                         conf->working_disks--;
526                         mddev->degraded++;
527                         conf->failed_disks++;
528                         clear_bit(In_sync, &rdev->flags);
529                         /*
530                          * if recovery was running, make sure it aborts.
531                          */
532                         set_bit(MD_RECOVERY_ERR, &mddev->recovery);
533                 }
534                 set_bit(Faulty, &rdev->flags);
535                 printk (KERN_ALERT
536                         "raid5: Disk failure on %s, disabling device."
537                         " Operation continuing on %d devices\n",
538                         bdevname(rdev->bdev,b), conf->working_disks);
539         }
540 }       
541
542 /*
543  * Input: a 'big' sector number,
544  * Output: index of the data and parity disk, and the sector # in them.
545  */
546 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
547                         unsigned int data_disks, unsigned int * dd_idx,
548                         unsigned int * pd_idx, raid5_conf_t *conf)
549 {
550         long stripe;
551         unsigned long chunk_number;
552         unsigned int chunk_offset;
553         sector_t new_sector;
554         int sectors_per_chunk = conf->chunk_size >> 9;
555
556         /* First compute the information on this sector */
557
558         /*
559          * Compute the chunk number and the sector offset inside the chunk
560          */
561         chunk_offset = sector_div(r_sector, sectors_per_chunk);
562         chunk_number = r_sector;
563         BUG_ON(r_sector != chunk_number);
564
565         /*
566          * Compute the stripe number
567          */
568         stripe = chunk_number / data_disks;
569
570         /*
571          * Compute the data disk and parity disk indexes inside the stripe
572          */
573         *dd_idx = chunk_number % data_disks;
574
575         /*
576          * Select the parity disk based on the user selected algorithm.
577          */
578         if (conf->level == 4)
579                 *pd_idx = data_disks;
580         else switch (conf->algorithm) {
581                 case ALGORITHM_LEFT_ASYMMETRIC:
582                         *pd_idx = data_disks - stripe % raid_disks;
583                         if (*dd_idx >= *pd_idx)
584                                 (*dd_idx)++;
585                         break;
586                 case ALGORITHM_RIGHT_ASYMMETRIC:
587                         *pd_idx = stripe % raid_disks;
588                         if (*dd_idx >= *pd_idx)
589                                 (*dd_idx)++;
590                         break;
591                 case ALGORITHM_LEFT_SYMMETRIC:
592                         *pd_idx = data_disks - stripe % raid_disks;
593                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
594                         break;
595                 case ALGORITHM_RIGHT_SYMMETRIC:
596                         *pd_idx = stripe % raid_disks;
597                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
598                         break;
599                 default:
600                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
601                                 conf->algorithm);
602         }
603
604         /*
605          * Finally, compute the new sector number
606          */
607         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
608         return new_sector;
609 }
610
611
612 static sector_t compute_blocknr(struct stripe_head *sh, int i)
613 {
614         raid5_conf_t *conf = sh->raid_conf;
615         int raid_disks = conf->raid_disks, data_disks = raid_disks - 1;
616         sector_t new_sector = sh->sector, check;
617         int sectors_per_chunk = conf->chunk_size >> 9;
618         sector_t stripe;
619         int chunk_offset;
620         int chunk_number, dummy1, dummy2, dd_idx = i;
621         sector_t r_sector;
622
623         chunk_offset = sector_div(new_sector, sectors_per_chunk);
624         stripe = new_sector;
625         BUG_ON(new_sector != stripe);
626
627         
628         switch (conf->algorithm) {
629                 case ALGORITHM_LEFT_ASYMMETRIC:
630                 case ALGORITHM_RIGHT_ASYMMETRIC:
631                         if (i > sh->pd_idx)
632                                 i--;
633                         break;
634                 case ALGORITHM_LEFT_SYMMETRIC:
635                 case ALGORITHM_RIGHT_SYMMETRIC:
636                         if (i < sh->pd_idx)
637                                 i += raid_disks;
638                         i -= (sh->pd_idx + 1);
639                         break;
640                 default:
641                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
642                                 conf->algorithm);
643         }
644
645         chunk_number = stripe * data_disks + i;
646         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
647
648         check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
649         if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
650                 printk(KERN_ERR "compute_blocknr: map not correct\n");
651                 return 0;
652         }
653         return r_sector;
654 }
655
656
657
658 /*
659  * Copy data between a page in the stripe cache, and a bio.
660  * There are no alignment or size guarantees between the page or the
661  * bio except that there is some overlap.
662  * All iovecs in the bio must be considered.
663  */
664 static void copy_data(int frombio, struct bio *bio,
665                      struct page *page,
666                      sector_t sector)
667 {
668         char *pa = page_address(page);
669         struct bio_vec *bvl;
670         int i;
671         int page_offset;
672
673         if (bio->bi_sector >= sector)
674                 page_offset = (signed)(bio->bi_sector - sector) * 512;
675         else
676                 page_offset = (signed)(sector - bio->bi_sector) * -512;
677         bio_for_each_segment(bvl, bio, i) {
678                 int len = bio_iovec_idx(bio,i)->bv_len;
679                 int clen;
680                 int b_offset = 0;
681
682                 if (page_offset < 0) {
683                         b_offset = -page_offset;
684                         page_offset += b_offset;
685                         len -= b_offset;
686                 }
687
688                 if (len > 0 && page_offset + len > STRIPE_SIZE)
689                         clen = STRIPE_SIZE - page_offset;
690                 else clen = len;
691                         
692                 if (clen > 0) {
693                         char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
694                         if (frombio)
695                                 memcpy(pa+page_offset, ba+b_offset, clen);
696                         else
697                                 memcpy(ba+b_offset, pa+page_offset, clen);
698                         __bio_kunmap_atomic(ba, KM_USER0);
699                 }
700                 if (clen < len) /* hit end of page */
701                         break;
702                 page_offset +=  len;
703         }
704 }
705
706 #define check_xor()     do {                                            \
707                            if (count == MAX_XOR_BLOCKS) {               \
708                                 xor_block(count, STRIPE_SIZE, ptr);     \
709                                 count = 1;                              \
710                            }                                            \
711                         } while(0)
712
713
714 static void compute_block(struct stripe_head *sh, int dd_idx)
715 {
716         raid5_conf_t *conf = sh->raid_conf;
717         int i, count, disks = conf->raid_disks;
718         void *ptr[MAX_XOR_BLOCKS], *p;
719
720         PRINTK("compute_block, stripe %llu, idx %d\n", 
721                 (unsigned long long)sh->sector, dd_idx);
722
723         ptr[0] = page_address(sh->dev[dd_idx].page);
724         memset(ptr[0], 0, STRIPE_SIZE);
725         count = 1;
726         for (i = disks ; i--; ) {
727                 if (i == dd_idx)
728                         continue;
729                 p = page_address(sh->dev[i].page);
730                 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
731                         ptr[count++] = p;
732                 else
733                         printk(KERN_ERR "compute_block() %d, stripe %llu, %d"
734                                 " not present\n", dd_idx,
735                                 (unsigned long long)sh->sector, i);
736
737                 check_xor();
738         }
739         if (count != 1)
740                 xor_block(count, STRIPE_SIZE, ptr);
741         set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
742 }
743
744 static void compute_parity(struct stripe_head *sh, int method)
745 {
746         raid5_conf_t *conf = sh->raid_conf;
747         int i, pd_idx = sh->pd_idx, disks = conf->raid_disks, count;
748         void *ptr[MAX_XOR_BLOCKS];
749         struct bio *chosen;
750
751         PRINTK("compute_parity, stripe %llu, method %d\n",
752                 (unsigned long long)sh->sector, method);
753
754         count = 1;
755         ptr[0] = page_address(sh->dev[pd_idx].page);
756         switch(method) {
757         case READ_MODIFY_WRITE:
758                 if (!test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags))
759                         BUG();
760                 for (i=disks ; i-- ;) {
761                         if (i==pd_idx)
762                                 continue;
763                         if (sh->dev[i].towrite &&
764                             test_bit(R5_UPTODATE, &sh->dev[i].flags)) {
765                                 ptr[count++] = page_address(sh->dev[i].page);
766                                 chosen = sh->dev[i].towrite;
767                                 sh->dev[i].towrite = NULL;
768
769                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
770                                         wake_up(&conf->wait_for_overlap);
771
772                                 if (sh->dev[i].written) BUG();
773                                 sh->dev[i].written = chosen;
774                                 check_xor();
775                         }
776                 }
777                 break;
778         case RECONSTRUCT_WRITE:
779                 memset(ptr[0], 0, STRIPE_SIZE);
780                 for (i= disks; i-- ;)
781                         if (i!=pd_idx && sh->dev[i].towrite) {
782                                 chosen = sh->dev[i].towrite;
783                                 sh->dev[i].towrite = NULL;
784
785                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
786                                         wake_up(&conf->wait_for_overlap);
787
788                                 if (sh->dev[i].written) BUG();
789                                 sh->dev[i].written = chosen;
790                         }
791                 break;
792         case CHECK_PARITY:
793                 break;
794         }
795         if (count>1) {
796                 xor_block(count, STRIPE_SIZE, ptr);
797                 count = 1;
798         }
799         
800         for (i = disks; i--;)
801                 if (sh->dev[i].written) {
802                         sector_t sector = sh->dev[i].sector;
803                         struct bio *wbi = sh->dev[i].written;
804                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
805                                 copy_data(1, wbi, sh->dev[i].page, sector);
806                                 wbi = r5_next_bio(wbi, sector);
807                         }
808
809                         set_bit(R5_LOCKED, &sh->dev[i].flags);
810                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
811                 }
812
813         switch(method) {
814         case RECONSTRUCT_WRITE:
815         case CHECK_PARITY:
816                 for (i=disks; i--;)
817                         if (i != pd_idx) {
818                                 ptr[count++] = page_address(sh->dev[i].page);
819                                 check_xor();
820                         }
821                 break;
822         case READ_MODIFY_WRITE:
823                 for (i = disks; i--;)
824                         if (sh->dev[i].written) {
825                                 ptr[count++] = page_address(sh->dev[i].page);
826                                 check_xor();
827                         }
828         }
829         if (count != 1)
830                 xor_block(count, STRIPE_SIZE, ptr);
831         
832         if (method != CHECK_PARITY) {
833                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
834                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
835         } else
836                 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
837 }
838
839 /*
840  * Each stripe/dev can have one or more bion attached.
841  * toread/towrite point to the first in a chain. 
842  * The bi_next chain must be in order.
843  */
844 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
845 {
846         struct bio **bip;
847         raid5_conf_t *conf = sh->raid_conf;
848         int firstwrite=0;
849
850         PRINTK("adding bh b#%llu to stripe s#%llu\n",
851                 (unsigned long long)bi->bi_sector,
852                 (unsigned long long)sh->sector);
853
854
855         spin_lock(&sh->lock);
856         spin_lock_irq(&conf->device_lock);
857         if (forwrite) {
858                 bip = &sh->dev[dd_idx].towrite;
859                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
860                         firstwrite = 1;
861         } else
862                 bip = &sh->dev[dd_idx].toread;
863         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
864                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
865                         goto overlap;
866                 bip = & (*bip)->bi_next;
867         }
868         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
869                 goto overlap;
870
871         if (*bip && bi->bi_next && (*bip) != bi->bi_next)
872                 BUG();
873         if (*bip)
874                 bi->bi_next = *bip;
875         *bip = bi;
876         bi->bi_phys_segments ++;
877         spin_unlock_irq(&conf->device_lock);
878         spin_unlock(&sh->lock);
879
880         PRINTK("added bi b#%llu to stripe s#%llu, disk %d.\n",
881                 (unsigned long long)bi->bi_sector,
882                 (unsigned long long)sh->sector, dd_idx);
883
884         if (conf->mddev->bitmap && firstwrite) {
885                 sh->bm_seq = conf->seq_write;
886                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
887                                   STRIPE_SECTORS, 0);
888                 set_bit(STRIPE_BIT_DELAY, &sh->state);
889         }
890
891         if (forwrite) {
892                 /* check if page is covered */
893                 sector_t sector = sh->dev[dd_idx].sector;
894                 for (bi=sh->dev[dd_idx].towrite;
895                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
896                              bi && bi->bi_sector <= sector;
897                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
898                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
899                                 sector = bi->bi_sector + (bi->bi_size>>9);
900                 }
901                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
902                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
903         }
904         return 1;
905
906  overlap:
907         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
908         spin_unlock_irq(&conf->device_lock);
909         spin_unlock(&sh->lock);
910         return 0;
911 }
912
913
914 /*
915  * handle_stripe - do things to a stripe.
916  *
917  * We lock the stripe and then examine the state of various bits
918  * to see what needs to be done.
919  * Possible results:
920  *    return some read request which now have data
921  *    return some write requests which are safely on disc
922  *    schedule a read on some buffers
923  *    schedule a write of some buffers
924  *    return confirmation of parity correctness
925  *
926  * Parity calculations are done inside the stripe lock
927  * buffers are taken off read_list or write_list, and bh_cache buffers
928  * get BH_Lock set before the stripe lock is released.
929  *
930  */
931  
932 static void handle_stripe(struct stripe_head *sh)
933 {
934         raid5_conf_t *conf = sh->raid_conf;
935         int disks = conf->raid_disks;
936         struct bio *return_bi= NULL;
937         struct bio *bi;
938         int i;
939         int syncing;
940         int locked=0, uptodate=0, to_read=0, to_write=0, failed=0, written=0;
941         int non_overwrite = 0;
942         int failed_num=0;
943         struct r5dev *dev;
944
945         PRINTK("handling stripe %llu, cnt=%d, pd_idx=%d\n",
946                 (unsigned long long)sh->sector, atomic_read(&sh->count),
947                 sh->pd_idx);
948
949         spin_lock(&sh->lock);
950         clear_bit(STRIPE_HANDLE, &sh->state);
951         clear_bit(STRIPE_DELAYED, &sh->state);
952
953         syncing = test_bit(STRIPE_SYNCING, &sh->state);
954         /* Now to look around and see what can be done */
955
956         rcu_read_lock();
957         for (i=disks; i--; ) {
958                 mdk_rdev_t *rdev;
959                 dev = &sh->dev[i];
960                 clear_bit(R5_Insync, &dev->flags);
961
962                 PRINTK("check %d: state 0x%lx read %p write %p written %p\n",
963                         i, dev->flags, dev->toread, dev->towrite, dev->written);
964                 /* maybe we can reply to a read */
965                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
966                         struct bio *rbi, *rbi2;
967                         PRINTK("Return read for disc %d\n", i);
968                         spin_lock_irq(&conf->device_lock);
969                         rbi = dev->toread;
970                         dev->toread = NULL;
971                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
972                                 wake_up(&conf->wait_for_overlap);
973                         spin_unlock_irq(&conf->device_lock);
974                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
975                                 copy_data(0, rbi, dev->page, dev->sector);
976                                 rbi2 = r5_next_bio(rbi, dev->sector);
977                                 spin_lock_irq(&conf->device_lock);
978                                 if (--rbi->bi_phys_segments == 0) {
979                                         rbi->bi_next = return_bi;
980                                         return_bi = rbi;
981                                 }
982                                 spin_unlock_irq(&conf->device_lock);
983                                 rbi = rbi2;
984                         }
985                 }
986
987                 /* now count some things */
988                 if (test_bit(R5_LOCKED, &dev->flags)) locked++;
989                 if (test_bit(R5_UPTODATE, &dev->flags)) uptodate++;
990
991                 
992                 if (dev->toread) to_read++;
993                 if (dev->towrite) {
994                         to_write++;
995                         if (!test_bit(R5_OVERWRITE, &dev->flags))
996                                 non_overwrite++;
997                 }
998                 if (dev->written) written++;
999                 rdev = rcu_dereference(conf->disks[i].rdev);
1000                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
1001                         /* The ReadError flag will just be confusing now */
1002                         clear_bit(R5_ReadError, &dev->flags);
1003                         clear_bit(R5_ReWrite, &dev->flags);
1004                 }
1005                 if (!rdev || !test_bit(In_sync, &rdev->flags)
1006                     || test_bit(R5_ReadError, &dev->flags)) {
1007                         failed++;
1008                         failed_num = i;
1009                 } else
1010                         set_bit(R5_Insync, &dev->flags);
1011         }
1012         rcu_read_unlock();
1013         PRINTK("locked=%d uptodate=%d to_read=%d"
1014                 " to_write=%d failed=%d failed_num=%d\n",
1015                 locked, uptodate, to_read, to_write, failed, failed_num);
1016         /* check if the array has lost two devices and, if so, some requests might
1017          * need to be failed
1018          */
1019         if (failed > 1 && to_read+to_write+written) {
1020                 for (i=disks; i--; ) {
1021                         int bitmap_end = 0;
1022
1023                         if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1024                                 mdk_rdev_t *rdev;
1025                                 rcu_read_lock();
1026                                 rdev = rcu_dereference(conf->disks[i].rdev);
1027                                 if (rdev && test_bit(In_sync, &rdev->flags))
1028                                         /* multiple read failures in one stripe */
1029                                         md_error(conf->mddev, rdev);
1030                                 rcu_read_unlock();
1031                         }
1032
1033                         spin_lock_irq(&conf->device_lock);
1034                         /* fail all writes first */
1035                         bi = sh->dev[i].towrite;
1036                         sh->dev[i].towrite = NULL;
1037                         if (bi) { to_write--; bitmap_end = 1; }
1038
1039                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1040                                 wake_up(&conf->wait_for_overlap);
1041
1042                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1043                                 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1044                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1045                                 if (--bi->bi_phys_segments == 0) {
1046                                         md_write_end(conf->mddev);
1047                                         bi->bi_next = return_bi;
1048                                         return_bi = bi;
1049                                 }
1050                                 bi = nextbi;
1051                         }
1052                         /* and fail all 'written' */
1053                         bi = sh->dev[i].written;
1054                         sh->dev[i].written = NULL;
1055                         if (bi) bitmap_end = 1;
1056                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS) {
1057                                 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1058                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1059                                 if (--bi->bi_phys_segments == 0) {
1060                                         md_write_end(conf->mddev);
1061                                         bi->bi_next = return_bi;
1062                                         return_bi = bi;
1063                                 }
1064                                 bi = bi2;
1065                         }
1066
1067                         /* fail any reads if this device is non-operational */
1068                         if (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1069                             test_bit(R5_ReadError, &sh->dev[i].flags)) {
1070                                 bi = sh->dev[i].toread;
1071                                 sh->dev[i].toread = NULL;
1072                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1073                                         wake_up(&conf->wait_for_overlap);
1074                                 if (bi) to_read--;
1075                                 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1076                                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1077                                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1078                                         if (--bi->bi_phys_segments == 0) {
1079                                                 bi->bi_next = return_bi;
1080                                                 return_bi = bi;
1081                                         }
1082                                         bi = nextbi;
1083                                 }
1084                         }
1085                         spin_unlock_irq(&conf->device_lock);
1086                         if (bitmap_end)
1087                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1088                                                 STRIPE_SECTORS, 0, 0);
1089                 }
1090         }
1091         if (failed > 1 && syncing) {
1092                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
1093                 clear_bit(STRIPE_SYNCING, &sh->state);
1094                 syncing = 0;
1095         }
1096
1097         /* might be able to return some write requests if the parity block
1098          * is safe, or on a failed drive
1099          */
1100         dev = &sh->dev[sh->pd_idx];
1101         if ( written &&
1102              ( (test_bit(R5_Insync, &dev->flags) && !test_bit(R5_LOCKED, &dev->flags) &&
1103                 test_bit(R5_UPTODATE, &dev->flags))
1104                || (failed == 1 && failed_num == sh->pd_idx))
1105             ) {
1106             /* any written block on an uptodate or failed drive can be returned.
1107              * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but 
1108              * never LOCKED, so we don't need to test 'failed' directly.
1109              */
1110             for (i=disks; i--; )
1111                 if (sh->dev[i].written) {
1112                     dev = &sh->dev[i];
1113                     if (!test_bit(R5_LOCKED, &dev->flags) &&
1114                          test_bit(R5_UPTODATE, &dev->flags) ) {
1115                         /* We can return any write requests */
1116                             struct bio *wbi, *wbi2;
1117                             int bitmap_end = 0;
1118                             PRINTK("Return write for disc %d\n", i);
1119                             spin_lock_irq(&conf->device_lock);
1120                             wbi = dev->written;
1121                             dev->written = NULL;
1122                             while (wbi && wbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1123                                     wbi2 = r5_next_bio(wbi, dev->sector);
1124                                     if (--wbi->bi_phys_segments == 0) {
1125                                             md_write_end(conf->mddev);
1126                                             wbi->bi_next = return_bi;
1127                                             return_bi = wbi;
1128                                     }
1129                                     wbi = wbi2;
1130                             }
1131                             if (dev->towrite == NULL)
1132                                     bitmap_end = 1;
1133                             spin_unlock_irq(&conf->device_lock);
1134                             if (bitmap_end)
1135                                     bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1136                                                     STRIPE_SECTORS,
1137                                                     !test_bit(STRIPE_DEGRADED, &sh->state), 0);
1138                     }
1139                 }
1140         }
1141
1142         /* Now we might consider reading some blocks, either to check/generate
1143          * parity, or to satisfy requests
1144          * or to load a block that is being partially written.
1145          */
1146         if (to_read || non_overwrite || (syncing && (uptodate < disks))) {
1147                 for (i=disks; i--;) {
1148                         dev = &sh->dev[i];
1149                         if (!test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1150                             (dev->toread ||
1151                              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1152                              syncing ||
1153                              (failed && (sh->dev[failed_num].toread ||
1154                                          (sh->dev[failed_num].towrite && !test_bit(R5_OVERWRITE, &sh->dev[failed_num].flags))))
1155                                     )
1156                                 ) {
1157                                 /* we would like to get this block, possibly
1158                                  * by computing it, but we might not be able to
1159                                  */
1160                                 if (uptodate == disks-1) {
1161                                         PRINTK("Computing block %d\n", i);
1162                                         compute_block(sh, i);
1163                                         uptodate++;
1164                                 } else if (test_bit(R5_Insync, &dev->flags)) {
1165                                         set_bit(R5_LOCKED, &dev->flags);
1166                                         set_bit(R5_Wantread, &dev->flags);
1167 #if 0
1168                                         /* if I am just reading this block and we don't have
1169                                            a failed drive, or any pending writes then sidestep the cache */
1170                                         if (sh->bh_read[i] && !sh->bh_read[i]->b_reqnext &&
1171                                             ! syncing && !failed && !to_write) {
1172                                                 sh->bh_cache[i]->b_page =  sh->bh_read[i]->b_page;
1173                                                 sh->bh_cache[i]->b_data =  sh->bh_read[i]->b_data;
1174                                         }
1175 #endif
1176                                         locked++;
1177                                         PRINTK("Reading block %d (sync=%d)\n", 
1178                                                 i, syncing);
1179                                 }
1180                         }
1181                 }
1182                 set_bit(STRIPE_HANDLE, &sh->state);
1183         }
1184
1185         /* now to consider writing and what else, if anything should be read */
1186         if (to_write) {
1187                 int rmw=0, rcw=0;
1188                 for (i=disks ; i--;) {
1189                         /* would I have to read this buffer for read_modify_write */
1190                         dev = &sh->dev[i];
1191                         if ((dev->towrite || i == sh->pd_idx) &&
1192                             (!test_bit(R5_LOCKED, &dev->flags) 
1193 #if 0
1194 || sh->bh_page[i]!=bh->b_page
1195 #endif
1196                                     ) &&
1197                             !test_bit(R5_UPTODATE, &dev->flags)) {
1198                                 if (test_bit(R5_Insync, &dev->flags)
1199 /*                                  && !(!mddev->insync && i == sh->pd_idx) */
1200                                         )
1201                                         rmw++;
1202                                 else rmw += 2*disks;  /* cannot read it */
1203                         }
1204                         /* Would I have to read this buffer for reconstruct_write */
1205                         if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
1206                             (!test_bit(R5_LOCKED, &dev->flags) 
1207 #if 0
1208 || sh->bh_page[i] != bh->b_page
1209 #endif
1210                                     ) &&
1211                             !test_bit(R5_UPTODATE, &dev->flags)) {
1212                                 if (test_bit(R5_Insync, &dev->flags)) rcw++;
1213                                 else rcw += 2*disks;
1214                         }
1215                 }
1216                 PRINTK("for sector %llu, rmw=%d rcw=%d\n", 
1217                         (unsigned long long)sh->sector, rmw, rcw);
1218                 set_bit(STRIPE_HANDLE, &sh->state);
1219                 if (rmw < rcw && rmw > 0)
1220                         /* prefer read-modify-write, but need to get some data */
1221                         for (i=disks; i--;) {
1222                                 dev = &sh->dev[i];
1223                                 if ((dev->towrite || i == sh->pd_idx) &&
1224                                     !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1225                                     test_bit(R5_Insync, &dev->flags)) {
1226                                         if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1227                                         {
1228                                                 PRINTK("Read_old block %d for r-m-w\n", i);
1229                                                 set_bit(R5_LOCKED, &dev->flags);
1230                                                 set_bit(R5_Wantread, &dev->flags);
1231                                                 locked++;
1232                                         } else {
1233                                                 set_bit(STRIPE_DELAYED, &sh->state);
1234                                                 set_bit(STRIPE_HANDLE, &sh->state);
1235                                         }
1236                                 }
1237                         }
1238                 if (rcw <= rmw && rcw > 0)
1239                         /* want reconstruct write, but need to get some data */
1240                         for (i=disks; i--;) {
1241                                 dev = &sh->dev[i];
1242                                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
1243                                     !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1244                                     test_bit(R5_Insync, &dev->flags)) {
1245                                         if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1246                                         {
1247                                                 PRINTK("Read_old block %d for Reconstruct\n", i);
1248                                                 set_bit(R5_LOCKED, &dev->flags);
1249                                                 set_bit(R5_Wantread, &dev->flags);
1250                                                 locked++;
1251                                         } else {
1252                                                 set_bit(STRIPE_DELAYED, &sh->state);
1253                                                 set_bit(STRIPE_HANDLE, &sh->state);
1254                                         }
1255                                 }
1256                         }
1257                 /* now if nothing is locked, and if we have enough data, we can start a write request */
1258                 if (locked == 0 && (rcw == 0 ||rmw == 0) &&
1259                     !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
1260                         PRINTK("Computing parity...\n");
1261                         compute_parity(sh, rcw==0 ? RECONSTRUCT_WRITE : READ_MODIFY_WRITE);
1262                         /* now every locked buffer is ready to be written */
1263                         for (i=disks; i--;)
1264                                 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
1265                                         PRINTK("Writing block %d\n", i);
1266                                         locked++;
1267                                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
1268                                         if (!test_bit(R5_Insync, &sh->dev[i].flags)
1269                                             || (i==sh->pd_idx && failed == 0))
1270                                                 set_bit(STRIPE_INSYNC, &sh->state);
1271                                 }
1272                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
1273                                 atomic_dec(&conf->preread_active_stripes);
1274                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
1275                                         md_wakeup_thread(conf->mddev->thread);
1276                         }
1277                 }
1278         }
1279
1280         /* maybe we need to check and possibly fix the parity for this stripe
1281          * Any reads will already have been scheduled, so we just see if enough data
1282          * is available
1283          */
1284         if (syncing && locked == 0 &&
1285             !test_bit(STRIPE_INSYNC, &sh->state)) {
1286                 set_bit(STRIPE_HANDLE, &sh->state);
1287                 if (failed == 0) {
1288                         char *pagea;
1289                         if (uptodate != disks)
1290                                 BUG();
1291                         compute_parity(sh, CHECK_PARITY);
1292                         uptodate--;
1293                         pagea = page_address(sh->dev[sh->pd_idx].page);
1294                         if ((*(u32*)pagea) == 0 &&
1295                             !memcmp(pagea, pagea+4, STRIPE_SIZE-4)) {
1296                                 /* parity is correct (on disc, not in buffer any more) */
1297                                 set_bit(STRIPE_INSYNC, &sh->state);
1298                         } else {
1299                                 conf->mddev->resync_mismatches += STRIPE_SECTORS;
1300                                 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
1301                                         /* don't try to repair!! */
1302                                         set_bit(STRIPE_INSYNC, &sh->state);
1303                                 else {
1304                                         compute_block(sh, sh->pd_idx);
1305                                         uptodate++;
1306                                 }
1307                         }
1308                 }
1309                 if (!test_bit(STRIPE_INSYNC, &sh->state)) {
1310                         /* either failed parity check, or recovery is happening */
1311                         if (failed==0)
1312                                 failed_num = sh->pd_idx;
1313                         dev = &sh->dev[failed_num];
1314                         BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
1315                         BUG_ON(uptodate != disks);
1316
1317                         set_bit(R5_LOCKED, &dev->flags);
1318                         set_bit(R5_Wantwrite, &dev->flags);
1319                         clear_bit(STRIPE_DEGRADED, &sh->state);
1320                         locked++;
1321                         set_bit(STRIPE_INSYNC, &sh->state);
1322                 }
1323         }
1324         if (syncing && locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
1325                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
1326                 clear_bit(STRIPE_SYNCING, &sh->state);
1327         }
1328
1329         /* If the failed drive is just a ReadError, then we might need to progress
1330          * the repair/check process
1331          */
1332         if (failed == 1 && ! conf->mddev->ro &&
1333             test_bit(R5_ReadError, &sh->dev[failed_num].flags)
1334             && !test_bit(R5_LOCKED, &sh->dev[failed_num].flags)
1335             && test_bit(R5_UPTODATE, &sh->dev[failed_num].flags)
1336                 ) {
1337                 dev = &sh->dev[failed_num];
1338                 if (!test_bit(R5_ReWrite, &dev->flags)) {
1339                         set_bit(R5_Wantwrite, &dev->flags);
1340                         set_bit(R5_ReWrite, &dev->flags);
1341                         set_bit(R5_LOCKED, &dev->flags);
1342                 } else {
1343                         /* let's read it back */
1344                         set_bit(R5_Wantread, &dev->flags);
1345                         set_bit(R5_LOCKED, &dev->flags);
1346                 }
1347         }
1348
1349         spin_unlock(&sh->lock);
1350
1351         while ((bi=return_bi)) {
1352                 int bytes = bi->bi_size;
1353
1354                 return_bi = bi->bi_next;
1355                 bi->bi_next = NULL;
1356                 bi->bi_size = 0;
1357                 bi->bi_end_io(bi, bytes, 0);
1358         }
1359         for (i=disks; i-- ;) {
1360                 int rw;
1361                 struct bio *bi;
1362                 mdk_rdev_t *rdev;
1363                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
1364                         rw = 1;
1365                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1366                         rw = 0;
1367                 else
1368                         continue;
1369  
1370                 bi = &sh->dev[i].req;
1371  
1372                 bi->bi_rw = rw;
1373                 if (rw)
1374                         bi->bi_end_io = raid5_end_write_request;
1375                 else
1376                         bi->bi_end_io = raid5_end_read_request;
1377  
1378                 rcu_read_lock();
1379                 rdev = rcu_dereference(conf->disks[i].rdev);
1380                 if (rdev && test_bit(Faulty, &rdev->flags))
1381                         rdev = NULL;
1382                 if (rdev)
1383                         atomic_inc(&rdev->nr_pending);
1384                 rcu_read_unlock();
1385  
1386                 if (rdev) {
1387                         if (syncing)
1388                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1389
1390                         bi->bi_bdev = rdev->bdev;
1391                         PRINTK("for %llu schedule op %ld on disc %d\n",
1392                                 (unsigned long long)sh->sector, bi->bi_rw, i);
1393                         atomic_inc(&sh->count);
1394                         bi->bi_sector = sh->sector + rdev->data_offset;
1395                         bi->bi_flags = 1 << BIO_UPTODATE;
1396                         bi->bi_vcnt = 1;        
1397                         bi->bi_max_vecs = 1;
1398                         bi->bi_idx = 0;
1399                         bi->bi_io_vec = &sh->dev[i].vec;
1400                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1401                         bi->bi_io_vec[0].bv_offset = 0;
1402                         bi->bi_size = STRIPE_SIZE;
1403                         bi->bi_next = NULL;
1404                         if (rw == WRITE &&
1405                             test_bit(R5_ReWrite, &sh->dev[i].flags))
1406                                 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1407                         generic_make_request(bi);
1408                 } else {
1409                         if (rw == 1)
1410                                 set_bit(STRIPE_DEGRADED, &sh->state);
1411                         PRINTK("skip op %ld on disc %d for sector %llu\n",
1412                                 bi->bi_rw, i, (unsigned long long)sh->sector);
1413                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1414                         set_bit(STRIPE_HANDLE, &sh->state);
1415                 }
1416         }
1417 }
1418
1419 static void raid5_activate_delayed(raid5_conf_t *conf)
1420 {
1421         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
1422                 while (!list_empty(&conf->delayed_list)) {
1423                         struct list_head *l = conf->delayed_list.next;
1424                         struct stripe_head *sh;
1425                         sh = list_entry(l, struct stripe_head, lru);
1426                         list_del_init(l);
1427                         clear_bit(STRIPE_DELAYED, &sh->state);
1428                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1429                                 atomic_inc(&conf->preread_active_stripes);
1430                         list_add_tail(&sh->lru, &conf->handle_list);
1431                 }
1432         }
1433 }
1434
1435 static void activate_bit_delay(raid5_conf_t *conf)
1436 {
1437         /* device_lock is held */
1438         struct list_head head;
1439         list_add(&head, &conf->bitmap_list);
1440         list_del_init(&conf->bitmap_list);
1441         while (!list_empty(&head)) {
1442                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
1443                 list_del_init(&sh->lru);
1444                 atomic_inc(&sh->count);
1445                 __release_stripe(conf, sh);
1446         }
1447 }
1448
1449 static void unplug_slaves(mddev_t *mddev)
1450 {
1451         raid5_conf_t *conf = mddev_to_conf(mddev);
1452         int i;
1453
1454         rcu_read_lock();
1455         for (i=0; i<mddev->raid_disks; i++) {
1456                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
1457                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
1458                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
1459
1460                         atomic_inc(&rdev->nr_pending);
1461                         rcu_read_unlock();
1462
1463                         if (r_queue->unplug_fn)
1464                                 r_queue->unplug_fn(r_queue);
1465
1466                         rdev_dec_pending(rdev, mddev);
1467                         rcu_read_lock();
1468                 }
1469         }
1470         rcu_read_unlock();
1471 }
1472
1473 static void raid5_unplug_device(request_queue_t *q)
1474 {
1475         mddev_t *mddev = q->queuedata;
1476         raid5_conf_t *conf = mddev_to_conf(mddev);
1477         unsigned long flags;
1478
1479         spin_lock_irqsave(&conf->device_lock, flags);
1480
1481         if (blk_remove_plug(q)) {
1482                 conf->seq_flush++;
1483                 raid5_activate_delayed(conf);
1484         }
1485         md_wakeup_thread(mddev->thread);
1486
1487         spin_unlock_irqrestore(&conf->device_lock, flags);
1488
1489         unplug_slaves(mddev);
1490 }
1491
1492 static int raid5_issue_flush(request_queue_t *q, struct gendisk *disk,
1493                              sector_t *error_sector)
1494 {
1495         mddev_t *mddev = q->queuedata;
1496         raid5_conf_t *conf = mddev_to_conf(mddev);
1497         int i, ret = 0;
1498
1499         rcu_read_lock();
1500         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
1501                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
1502                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
1503                         struct block_device *bdev = rdev->bdev;
1504                         request_queue_t *r_queue = bdev_get_queue(bdev);
1505
1506                         if (!r_queue->issue_flush_fn)
1507                                 ret = -EOPNOTSUPP;
1508                         else {
1509                                 atomic_inc(&rdev->nr_pending);
1510                                 rcu_read_unlock();
1511                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
1512                                                               error_sector);
1513                                 rdev_dec_pending(rdev, mddev);
1514                                 rcu_read_lock();
1515                         }
1516                 }
1517         }
1518         rcu_read_unlock();
1519         return ret;
1520 }
1521
1522 static inline void raid5_plug_device(raid5_conf_t *conf)
1523 {
1524         spin_lock_irq(&conf->device_lock);
1525         blk_plug_device(conf->mddev->queue);
1526         spin_unlock_irq(&conf->device_lock);
1527 }
1528
1529 static int make_request (request_queue_t *q, struct bio * bi)
1530 {
1531         mddev_t *mddev = q->queuedata;
1532         raid5_conf_t *conf = mddev_to_conf(mddev);
1533         const unsigned int raid_disks = conf->raid_disks;
1534         const unsigned int data_disks = raid_disks - 1;
1535         unsigned int dd_idx, pd_idx;
1536         sector_t new_sector;
1537         sector_t logical_sector, last_sector;
1538         struct stripe_head *sh;
1539         const int rw = bio_data_dir(bi);
1540
1541         if (unlikely(bio_barrier(bi))) {
1542                 bio_endio(bi, bi->bi_size, -EOPNOTSUPP);
1543                 return 0;
1544         }
1545
1546         md_write_start(mddev, bi);
1547
1548         disk_stat_inc(mddev->gendisk, ios[rw]);
1549         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
1550
1551         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
1552         last_sector = bi->bi_sector + (bi->bi_size>>9);
1553         bi->bi_next = NULL;
1554         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
1555
1556         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
1557                 DEFINE_WAIT(w);
1558                 
1559                 new_sector = raid5_compute_sector(logical_sector,
1560                                                   raid_disks, data_disks, &dd_idx, &pd_idx, conf);
1561
1562                 PRINTK("raid5: make_request, sector %llu logical %llu\n",
1563                         (unsigned long long)new_sector, 
1564                         (unsigned long long)logical_sector);
1565
1566         retry:
1567                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
1568                 sh = get_active_stripe(conf, new_sector, pd_idx, (bi->bi_rw&RWA_MASK));
1569                 if (sh) {
1570                         if (!add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
1571                                 /* Add failed due to overlap.  Flush everything
1572                                  * and wait a while
1573                                  */
1574                                 raid5_unplug_device(mddev->queue);
1575                                 release_stripe(sh);
1576                                 schedule();
1577                                 goto retry;
1578                         }
1579                         finish_wait(&conf->wait_for_overlap, &w);
1580                         raid5_plug_device(conf);
1581                         handle_stripe(sh);
1582                         release_stripe(sh);
1583
1584                 } else {
1585                         /* cannot get stripe for read-ahead, just give-up */
1586                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1587                         finish_wait(&conf->wait_for_overlap, &w);
1588                         break;
1589                 }
1590                         
1591         }
1592         spin_lock_irq(&conf->device_lock);
1593         if (--bi->bi_phys_segments == 0) {
1594                 int bytes = bi->bi_size;
1595
1596                 if ( bio_data_dir(bi) == WRITE )
1597                         md_write_end(mddev);
1598                 bi->bi_size = 0;
1599                 bi->bi_end_io(bi, bytes, 0);
1600         }
1601         spin_unlock_irq(&conf->device_lock);
1602         return 0;
1603 }
1604
1605 /* FIXME go_faster isn't used */
1606 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1607 {
1608         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1609         struct stripe_head *sh;
1610         int sectors_per_chunk = conf->chunk_size >> 9;
1611         sector_t x;
1612         unsigned long stripe;
1613         int chunk_offset;
1614         int dd_idx, pd_idx;
1615         sector_t first_sector;
1616         int raid_disks = conf->raid_disks;
1617         int data_disks = raid_disks-1;
1618         sector_t max_sector = mddev->size << 1;
1619         int sync_blocks;
1620
1621         if (sector_nr >= max_sector) {
1622                 /* just being told to finish up .. nothing much to do */
1623                 unplug_slaves(mddev);
1624
1625                 if (mddev->curr_resync < max_sector) /* aborted */
1626                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1627                                         &sync_blocks, 1);
1628                 else /* compelted sync */
1629                         conf->fullsync = 0;
1630                 bitmap_close_sync(mddev->bitmap);
1631
1632                 return 0;
1633         }
1634         /* if there is 1 or more failed drives and we are trying
1635          * to resync, then assert that we are finished, because there is
1636          * nothing we can do.
1637          */
1638         if (mddev->degraded >= 1 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1639                 sector_t rv = (mddev->size << 1) - sector_nr;
1640                 *skipped = 1;
1641                 return rv;
1642         }
1643         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1644             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1645             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
1646                 /* we can skip this block, and probably more */
1647                 sync_blocks /= STRIPE_SECTORS;
1648                 *skipped = 1;
1649                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
1650         }
1651
1652         x = sector_nr;
1653         chunk_offset = sector_div(x, sectors_per_chunk);
1654         stripe = x;
1655         BUG_ON(x != stripe);
1656
1657         first_sector = raid5_compute_sector((sector_t)stripe*data_disks*sectors_per_chunk
1658                 + chunk_offset, raid_disks, data_disks, &dd_idx, &pd_idx, conf);
1659         sh = get_active_stripe(conf, sector_nr, pd_idx, 1);
1660         if (sh == NULL) {
1661                 sh = get_active_stripe(conf, sector_nr, pd_idx, 0);
1662                 /* make sure we don't swamp the stripe cache if someone else
1663                  * is trying to get access 
1664                  */
1665                 schedule_timeout_uninterruptible(1);
1666         }
1667         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 0);
1668         spin_lock(&sh->lock);   
1669         set_bit(STRIPE_SYNCING, &sh->state);
1670         clear_bit(STRIPE_INSYNC, &sh->state);
1671         spin_unlock(&sh->lock);
1672
1673         handle_stripe(sh);
1674         release_stripe(sh);
1675
1676         return STRIPE_SECTORS;
1677 }
1678
1679 /*
1680  * This is our raid5 kernel thread.
1681  *
1682  * We scan the hash table for stripes which can be handled now.
1683  * During the scan, completed stripes are saved for us by the interrupt
1684  * handler, so that they will not have to wait for our next wakeup.
1685  */
1686 static void raid5d (mddev_t *mddev)
1687 {
1688         struct stripe_head *sh;
1689         raid5_conf_t *conf = mddev_to_conf(mddev);
1690         int handled;
1691
1692         PRINTK("+++ raid5d active\n");
1693
1694         md_check_recovery(mddev);
1695
1696         handled = 0;
1697         spin_lock_irq(&conf->device_lock);
1698         while (1) {
1699                 struct list_head *first;
1700
1701                 if (conf->seq_flush - conf->seq_write > 0) {
1702                         int seq = conf->seq_flush;
1703                         spin_unlock_irq(&conf->device_lock);
1704                         bitmap_unplug(mddev->bitmap);
1705                         spin_lock_irq(&conf->device_lock);
1706                         conf->seq_write = seq;
1707                         activate_bit_delay(conf);
1708                 }
1709
1710                 if (list_empty(&conf->handle_list) &&
1711                     atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD &&
1712                     !blk_queue_plugged(mddev->queue) &&
1713                     !list_empty(&conf->delayed_list))
1714                         raid5_activate_delayed(conf);
1715
1716                 if (list_empty(&conf->handle_list))
1717                         break;
1718
1719                 first = conf->handle_list.next;
1720                 sh = list_entry(first, struct stripe_head, lru);
1721
1722                 list_del_init(first);
1723                 atomic_inc(&sh->count);
1724                 if (atomic_read(&sh->count)!= 1)
1725                         BUG();
1726                 spin_unlock_irq(&conf->device_lock);
1727                 
1728                 handled++;
1729                 handle_stripe(sh);
1730                 release_stripe(sh);
1731
1732                 spin_lock_irq(&conf->device_lock);
1733         }
1734         PRINTK("%d stripes handled\n", handled);
1735
1736         spin_unlock_irq(&conf->device_lock);
1737
1738         unplug_slaves(mddev);
1739
1740         PRINTK("--- raid5d inactive\n");
1741 }
1742
1743 static ssize_t
1744 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
1745 {
1746         raid5_conf_t *conf = mddev_to_conf(mddev);
1747         if (conf)
1748                 return sprintf(page, "%d\n", conf->max_nr_stripes);
1749         else
1750                 return 0;
1751 }
1752
1753 static ssize_t
1754 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
1755 {
1756         raid5_conf_t *conf = mddev_to_conf(mddev);
1757         char *end;
1758         int new;
1759         if (len >= PAGE_SIZE)
1760                 return -EINVAL;
1761         if (!conf)
1762                 return -ENODEV;
1763
1764         new = simple_strtoul(page, &end, 10);
1765         if (!*page || (*end && *end != '\n') )
1766                 return -EINVAL;
1767         if (new <= 16 || new > 32768)
1768                 return -EINVAL;
1769         while (new < conf->max_nr_stripes) {
1770                 if (drop_one_stripe(conf))
1771                         conf->max_nr_stripes--;
1772                 else
1773                         break;
1774         }
1775         while (new > conf->max_nr_stripes) {
1776                 if (grow_one_stripe(conf))
1777                         conf->max_nr_stripes++;
1778                 else break;
1779         }
1780         return len;
1781 }
1782
1783 static struct md_sysfs_entry
1784 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
1785                                 raid5_show_stripe_cache_size,
1786                                 raid5_store_stripe_cache_size);
1787
1788 static ssize_t
1789 stripe_cache_active_show(mddev_t *mddev, char *page)
1790 {
1791         raid5_conf_t *conf = mddev_to_conf(mddev);
1792         if (conf)
1793                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
1794         else
1795                 return 0;
1796 }
1797
1798 static struct md_sysfs_entry
1799 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
1800
1801 static struct attribute *raid5_attrs[] =  {
1802         &raid5_stripecache_size.attr,
1803         &raid5_stripecache_active.attr,
1804         NULL,
1805 };
1806 static struct attribute_group raid5_attrs_group = {
1807         .name = NULL,
1808         .attrs = raid5_attrs,
1809 };
1810
1811 static int run(mddev_t *mddev)
1812 {
1813         raid5_conf_t *conf;
1814         int raid_disk, memory;
1815         mdk_rdev_t *rdev;
1816         struct disk_info *disk;
1817         struct list_head *tmp;
1818
1819         if (mddev->level != 5 && mddev->level != 4) {
1820                 printk(KERN_ERR "raid5: %s: raid level not set to 4/5 (%d)\n",
1821                        mdname(mddev), mddev->level);
1822                 return -EIO;
1823         }
1824
1825         mddev->private = kzalloc(sizeof (raid5_conf_t)
1826                                  + mddev->raid_disks * sizeof(struct disk_info),
1827                                  GFP_KERNEL);
1828         if ((conf = mddev->private) == NULL)
1829                 goto abort;
1830
1831         conf->mddev = mddev;
1832
1833         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
1834                 goto abort;
1835
1836         spin_lock_init(&conf->device_lock);
1837         init_waitqueue_head(&conf->wait_for_stripe);
1838         init_waitqueue_head(&conf->wait_for_overlap);
1839         INIT_LIST_HEAD(&conf->handle_list);
1840         INIT_LIST_HEAD(&conf->delayed_list);
1841         INIT_LIST_HEAD(&conf->bitmap_list);
1842         INIT_LIST_HEAD(&conf->inactive_list);
1843         atomic_set(&conf->active_stripes, 0);
1844         atomic_set(&conf->preread_active_stripes, 0);
1845
1846         PRINTK("raid5: run(%s) called.\n", mdname(mddev));
1847
1848         ITERATE_RDEV(mddev,rdev,tmp) {
1849                 raid_disk = rdev->raid_disk;
1850                 if (raid_disk >= mddev->raid_disks
1851                     || raid_disk < 0)
1852                         continue;
1853                 disk = conf->disks + raid_disk;
1854
1855                 disk->rdev = rdev;
1856
1857                 if (test_bit(In_sync, &rdev->flags)) {
1858                         char b[BDEVNAME_SIZE];
1859                         printk(KERN_INFO "raid5: device %s operational as raid"
1860                                 " disk %d\n", bdevname(rdev->bdev,b),
1861                                 raid_disk);
1862                         conf->working_disks++;
1863                 }
1864         }
1865
1866         conf->raid_disks = mddev->raid_disks;
1867         /*
1868          * 0 for a fully functional array, 1 for a degraded array.
1869          */
1870         mddev->degraded = conf->failed_disks = conf->raid_disks - conf->working_disks;
1871         conf->mddev = mddev;
1872         conf->chunk_size = mddev->chunk_size;
1873         conf->level = mddev->level;
1874         conf->algorithm = mddev->layout;
1875         conf->max_nr_stripes = NR_STRIPES;
1876
1877         /* device size must be a multiple of chunk size */
1878         mddev->size &= ~(mddev->chunk_size/1024 -1);
1879         mddev->resync_max_sectors = mddev->size << 1;
1880
1881         if (!conf->chunk_size || conf->chunk_size % 4) {
1882                 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
1883                         conf->chunk_size, mdname(mddev));
1884                 goto abort;
1885         }
1886         if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
1887                 printk(KERN_ERR 
1888                         "raid5: unsupported parity algorithm %d for %s\n",
1889                         conf->algorithm, mdname(mddev));
1890                 goto abort;
1891         }
1892         if (mddev->degraded > 1) {
1893                 printk(KERN_ERR "raid5: not enough operational devices for %s"
1894                         " (%d/%d failed)\n",
1895                         mdname(mddev), conf->failed_disks, conf->raid_disks);
1896                 goto abort;
1897         }
1898
1899         if (mddev->degraded == 1 &&
1900             mddev->recovery_cp != MaxSector) {
1901                 if (mddev->ok_start_degraded)
1902                         printk(KERN_WARNING
1903                                "raid5: starting dirty degraded array: %s"
1904                                "- data corruption possible.\n",
1905                                mdname(mddev));
1906                 else {
1907                         printk(KERN_ERR
1908                                "raid5: cannot start dirty degraded array for %s\n",
1909                                mdname(mddev));
1910                         goto abort;
1911                 }
1912         }
1913
1914         {
1915                 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
1916                 if (!mddev->thread) {
1917                         printk(KERN_ERR 
1918                                 "raid5: couldn't allocate thread for %s\n",
1919                                 mdname(mddev));
1920                         goto abort;
1921                 }
1922         }
1923         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
1924                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
1925         if (grow_stripes(conf, conf->max_nr_stripes)) {
1926                 printk(KERN_ERR 
1927                         "raid5: couldn't allocate %dkB for buffers\n", memory);
1928                 shrink_stripes(conf);
1929                 md_unregister_thread(mddev->thread);
1930                 goto abort;
1931         } else
1932                 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
1933                         memory, mdname(mddev));
1934
1935         if (mddev->degraded == 0)
1936                 printk("raid5: raid level %d set %s active with %d out of %d"
1937                         " devices, algorithm %d\n", conf->level, mdname(mddev), 
1938                         mddev->raid_disks-mddev->degraded, mddev->raid_disks,
1939                         conf->algorithm);
1940         else
1941                 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
1942                         " out of %d devices, algorithm %d\n", conf->level,
1943                         mdname(mddev), mddev->raid_disks - mddev->degraded,
1944                         mddev->raid_disks, conf->algorithm);
1945
1946         print_raid5_conf(conf);
1947
1948         /* read-ahead size must cover two whole stripes, which is
1949          * 2 * (n-1) * chunksize where 'n' is the number of raid devices
1950          */
1951         {
1952                 int stripe = (mddev->raid_disks-1) * mddev->chunk_size
1953                         / PAGE_SIZE;
1954                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
1955                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
1956         }
1957
1958         /* Ok, everything is just fine now */
1959         sysfs_create_group(&mddev->kobj, &raid5_attrs_group);
1960
1961         mddev->queue->unplug_fn = raid5_unplug_device;
1962         mddev->queue->issue_flush_fn = raid5_issue_flush;
1963
1964         mddev->array_size =  mddev->size * (mddev->raid_disks - 1);
1965         return 0;
1966 abort:
1967         if (conf) {
1968                 print_raid5_conf(conf);
1969                 kfree(conf->stripe_hashtbl);
1970                 kfree(conf);
1971         }
1972         mddev->private = NULL;
1973         printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
1974         return -EIO;
1975 }
1976
1977
1978
1979 static int stop(mddev_t *mddev)
1980 {
1981         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1982
1983         md_unregister_thread(mddev->thread);
1984         mddev->thread = NULL;
1985         shrink_stripes(conf);
1986         kfree(conf->stripe_hashtbl);
1987         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
1988         sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
1989         kfree(conf);
1990         mddev->private = NULL;
1991         return 0;
1992 }
1993
1994 #if RAID5_DEBUG
1995 static void print_sh (struct stripe_head *sh)
1996 {
1997         int i;
1998
1999         printk("sh %llu, pd_idx %d, state %ld.\n",
2000                 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
2001         printk("sh %llu,  count %d.\n",
2002                 (unsigned long long)sh->sector, atomic_read(&sh->count));
2003         printk("sh %llu, ", (unsigned long long)sh->sector);
2004         for (i = 0; i < sh->raid_conf->raid_disks; i++) {
2005                 printk("(cache%d: %p %ld) ", 
2006                         i, sh->dev[i].page, sh->dev[i].flags);
2007         }
2008         printk("\n");
2009 }
2010
2011 static void printall (raid5_conf_t *conf)
2012 {
2013         struct stripe_head *sh;
2014         struct hlist_node *hn;
2015         int i;
2016
2017         spin_lock_irq(&conf->device_lock);
2018         for (i = 0; i < NR_HASH; i++) {
2019                 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
2020                         if (sh->raid_conf != conf)
2021                                 continue;
2022                         print_sh(sh);
2023                 }
2024         }
2025         spin_unlock_irq(&conf->device_lock);
2026 }
2027 #endif
2028
2029 static void status (struct seq_file *seq, mddev_t *mddev)
2030 {
2031         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
2032         int i;
2033
2034         seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
2035         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->working_disks);
2036         for (i = 0; i < conf->raid_disks; i++)
2037                 seq_printf (seq, "%s",
2038                                conf->disks[i].rdev &&
2039                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
2040         seq_printf (seq, "]");
2041 #if RAID5_DEBUG
2042 #define D(x) \
2043         seq_printf (seq, "<"#x":%d>", atomic_read(&conf->x))
2044         printall(conf);
2045 #endif
2046 }
2047
2048 static void print_raid5_conf (raid5_conf_t *conf)
2049 {
2050         int i;
2051         struct disk_info *tmp;
2052
2053         printk("RAID5 conf printout:\n");
2054         if (!conf) {
2055                 printk("(conf==NULL)\n");
2056                 return;
2057         }
2058         printk(" --- rd:%d wd:%d fd:%d\n", conf->raid_disks,
2059                  conf->working_disks, conf->failed_disks);
2060
2061         for (i = 0; i < conf->raid_disks; i++) {
2062                 char b[BDEVNAME_SIZE];
2063                 tmp = conf->disks + i;
2064                 if (tmp->rdev)
2065                 printk(" disk %d, o:%d, dev:%s\n",
2066                         i, !test_bit(Faulty, &tmp->rdev->flags),
2067                         bdevname(tmp->rdev->bdev,b));
2068         }
2069 }
2070
2071 static int raid5_spare_active(mddev_t *mddev)
2072 {
2073         int i;
2074         raid5_conf_t *conf = mddev->private;
2075         struct disk_info *tmp;
2076
2077         for (i = 0; i < conf->raid_disks; i++) {
2078                 tmp = conf->disks + i;
2079                 if (tmp->rdev
2080                     && !test_bit(Faulty, &tmp->rdev->flags)
2081                     && !test_bit(In_sync, &tmp->rdev->flags)) {
2082                         mddev->degraded--;
2083                         conf->failed_disks--;
2084                         conf->working_disks++;
2085                         set_bit(In_sync, &tmp->rdev->flags);
2086                 }
2087         }
2088         print_raid5_conf(conf);
2089         return 0;
2090 }
2091
2092 static int raid5_remove_disk(mddev_t *mddev, int number)
2093 {
2094         raid5_conf_t *conf = mddev->private;
2095         int err = 0;
2096         mdk_rdev_t *rdev;
2097         struct disk_info *p = conf->disks + number;
2098
2099         print_raid5_conf(conf);
2100         rdev = p->rdev;
2101         if (rdev) {
2102                 if (test_bit(In_sync, &rdev->flags) ||
2103                     atomic_read(&rdev->nr_pending)) {
2104                         err = -EBUSY;
2105                         goto abort;
2106                 }
2107                 p->rdev = NULL;
2108                 synchronize_rcu();
2109                 if (atomic_read(&rdev->nr_pending)) {
2110                         /* lost the race, try later */
2111                         err = -EBUSY;
2112                         p->rdev = rdev;
2113                 }
2114         }
2115 abort:
2116
2117         print_raid5_conf(conf);
2118         return err;
2119 }
2120
2121 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
2122 {
2123         raid5_conf_t *conf = mddev->private;
2124         int found = 0;
2125         int disk;
2126         struct disk_info *p;
2127
2128         if (mddev->degraded > 1)
2129                 /* no point adding a device */
2130                 return 0;
2131
2132         /*
2133          * find the disk ...
2134          */
2135         for (disk=0; disk < mddev->raid_disks; disk++)
2136                 if ((p=conf->disks + disk)->rdev == NULL) {
2137                         clear_bit(In_sync, &rdev->flags);
2138                         rdev->raid_disk = disk;
2139                         found = 1;
2140                         if (rdev->saved_raid_disk != disk)
2141                                 conf->fullsync = 1;
2142                         rcu_assign_pointer(p->rdev, rdev);
2143                         break;
2144                 }
2145         print_raid5_conf(conf);
2146         return found;
2147 }
2148
2149 static int raid5_resize(mddev_t *mddev, sector_t sectors)
2150 {
2151         /* no resync is happening, and there is enough space
2152          * on all devices, so we can resize.
2153          * We need to make sure resync covers any new space.
2154          * If the array is shrinking we should possibly wait until
2155          * any io in the removed space completes, but it hardly seems
2156          * worth it.
2157          */
2158         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
2159         mddev->array_size = (sectors * (mddev->raid_disks-1))>>1;
2160         set_capacity(mddev->gendisk, mddev->array_size << 1);
2161         mddev->changed = 1;
2162         if (sectors/2  > mddev->size && mddev->recovery_cp == MaxSector) {
2163                 mddev->recovery_cp = mddev->size << 1;
2164                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2165         }
2166         mddev->size = sectors /2;
2167         mddev->resync_max_sectors = sectors;
2168         return 0;
2169 }
2170
2171 static void raid5_quiesce(mddev_t *mddev, int state)
2172 {
2173         raid5_conf_t *conf = mddev_to_conf(mddev);
2174
2175         switch(state) {
2176         case 1: /* stop all writes */
2177                 spin_lock_irq(&conf->device_lock);
2178                 conf->quiesce = 1;
2179                 wait_event_lock_irq(conf->wait_for_stripe,
2180                                     atomic_read(&conf->active_stripes) == 0,
2181                                     conf->device_lock, /* nothing */);
2182                 spin_unlock_irq(&conf->device_lock);
2183                 break;
2184
2185         case 0: /* re-enable writes */
2186                 spin_lock_irq(&conf->device_lock);
2187                 conf->quiesce = 0;
2188                 wake_up(&conf->wait_for_stripe);
2189                 spin_unlock_irq(&conf->device_lock);
2190                 break;
2191         }
2192 }
2193
2194 static struct mdk_personality raid5_personality =
2195 {
2196         .name           = "raid5",
2197         .level          = 5,
2198         .owner          = THIS_MODULE,
2199         .make_request   = make_request,
2200         .run            = run,
2201         .stop           = stop,
2202         .status         = status,
2203         .error_handler  = error,
2204         .hot_add_disk   = raid5_add_disk,
2205         .hot_remove_disk= raid5_remove_disk,
2206         .spare_active   = raid5_spare_active,
2207         .sync_request   = sync_request,
2208         .resize         = raid5_resize,
2209         .quiesce        = raid5_quiesce,
2210 };
2211
2212 static struct mdk_personality raid4_personality =
2213 {
2214         .name           = "raid4",
2215         .level          = 4,
2216         .owner          = THIS_MODULE,
2217         .make_request   = make_request,
2218         .run            = run,
2219         .stop           = stop,
2220         .status         = status,
2221         .error_handler  = error,
2222         .hot_add_disk   = raid5_add_disk,
2223         .hot_remove_disk= raid5_remove_disk,
2224         .spare_active   = raid5_spare_active,
2225         .sync_request   = sync_request,
2226         .resize         = raid5_resize,
2227         .quiesce        = raid5_quiesce,
2228 };
2229
2230 static int __init raid5_init(void)
2231 {
2232         register_md_personality(&raid5_personality);
2233         register_md_personality(&raid4_personality);
2234         return 0;
2235 }
2236
2237 static void raid5_exit(void)
2238 {
2239         unregister_md_personality(&raid5_personality);
2240         unregister_md_personality(&raid4_personality);
2241 }
2242
2243 module_init(raid5_init);
2244 module_exit(raid5_exit);
2245 MODULE_LICENSE("GPL");
2246 MODULE_ALIAS("md-personality-4"); /* RAID5 */
2247 MODULE_ALIAS("md-raid5");
2248 MODULE_ALIAS("md-raid4");
2249 MODULE_ALIAS("md-level-5");
2250 MODULE_ALIAS("md-level-4");