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