[PATCH] USB: libusual: fix warning on 64bit boxes
[linux-2.6] / drivers / md / raid6main.c
1 /*
2  * raid6main.c : Multiple Devices driver for Linux
3  *         Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4  *         Copyright (C) 1999, 2000 Ingo Molnar
5  *         Copyright (C) 2002, 2003 H. Peter Anvin
6  *
7  * RAID-6 management functions.  This code is derived from raid5.c.
8  * Last merge from raid5.c bkcvs version 1.79 (kernel 2.6.1).
9  *
10  * Thanks to Penguin Computing for making the RAID-6 development possible
11  * by donating a test server!
12  *
13  * This program is free software; you can redistribute it and/or modify
14  * it under the terms of the GNU General Public License as published by
15  * the Free Software Foundation; either version 2, or (at your option)
16  * any later version.
17  *
18  * You should have received a copy of the GNU General Public License
19  * (for example /usr/src/linux/COPYING); if not, write to the Free
20  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
21  */
22
23
24 #include <linux/config.h>
25 #include <linux/module.h>
26 #include <linux/slab.h>
27 #include <linux/highmem.h>
28 #include <linux/bitops.h>
29 #include <asm/atomic.h>
30 #include "raid6.h"
31
32 #include <linux/raid/bitmap.h>
33
34 /*
35  * Stripe cache
36  */
37
38 #define NR_STRIPES              256
39 #define STRIPE_SIZE             PAGE_SIZE
40 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
41 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
42 #define IO_THRESHOLD            1
43 #define NR_HASH                 (PAGE_SIZE / sizeof(struct hlist_head))
44 #define HASH_MASK               (NR_HASH - 1)
45
46 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
47
48 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
49  * order without overlap.  There may be several bio's per stripe+device, and
50  * a bio could span several devices.
51  * When walking this list for a particular stripe+device, we must never proceed
52  * beyond a bio that extends past this device, as the next bio might no longer
53  * be valid.
54  * This macro is used to determine the 'next' bio in the list, given the sector
55  * of the current stripe+device
56  */
57 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
58 /*
59  * The following can be used to debug the driver
60  */
61 #define RAID6_DEBUG     0       /* Extremely verbose printk */
62 #define RAID6_PARANOIA  1       /* Check spinlocks */
63 #define RAID6_DUMPSTATE 0       /* Include stripe cache state in /proc/mdstat */
64 #if RAID6_PARANOIA && defined(CONFIG_SMP)
65 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
66 #else
67 # define CHECK_DEVLOCK()
68 #endif
69
70 #define PRINTK(x...) ((void)(RAID6_DEBUG && printk(KERN_DEBUG x)))
71 #if RAID6_DEBUG
72 #undef inline
73 #undef __inline__
74 #define inline
75 #define __inline__
76 #endif
77
78 #if !RAID6_USE_EMPTY_ZERO_PAGE
79 /* In .bss so it's zeroed */
80 const char raid6_empty_zero_page[PAGE_SIZE] __attribute__((aligned(256)));
81 #endif
82
83 static inline int raid6_next_disk(int disk, int raid_disks)
84 {
85         disk++;
86         return (disk < raid_disks) ? disk : 0;
87 }
88
89 static void print_raid6_conf (raid6_conf_t *conf);
90
91 static void __release_stripe(raid6_conf_t *conf, struct stripe_head *sh)
92 {
93         if (atomic_dec_and_test(&sh->count)) {
94                 if (!list_empty(&sh->lru))
95                         BUG();
96                 if (atomic_read(&conf->active_stripes)==0)
97                         BUG();
98                 if (test_bit(STRIPE_HANDLE, &sh->state)) {
99                         if (test_bit(STRIPE_DELAYED, &sh->state))
100                                 list_add_tail(&sh->lru, &conf->delayed_list);
101                         else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
102                                  conf->seq_write == sh->bm_seq)
103                                 list_add_tail(&sh->lru, &conf->bitmap_list);
104                         else {
105                                 clear_bit(STRIPE_BIT_DELAY, &sh->state);
106                                 list_add_tail(&sh->lru, &conf->handle_list);
107                         }
108                         md_wakeup_thread(conf->mddev->thread);
109                 } else {
110                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
111                                 atomic_dec(&conf->preread_active_stripes);
112                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
113                                         md_wakeup_thread(conf->mddev->thread);
114                         }
115                         list_add_tail(&sh->lru, &conf->inactive_list);
116                         atomic_dec(&conf->active_stripes);
117                         if (!conf->inactive_blocked ||
118                             atomic_read(&conf->active_stripes) < (NR_STRIPES*3/4))
119                                 wake_up(&conf->wait_for_stripe);
120                 }
121         }
122 }
123 static void release_stripe(struct stripe_head *sh)
124 {
125         raid6_conf_t *conf = sh->raid_conf;
126         unsigned long flags;
127
128         spin_lock_irqsave(&conf->device_lock, flags);
129         __release_stripe(conf, sh);
130         spin_unlock_irqrestore(&conf->device_lock, flags);
131 }
132
133 static inline void remove_hash(struct stripe_head *sh)
134 {
135         PRINTK("remove_hash(), stripe %llu\n", (unsigned long long)sh->sector);
136
137         hlist_del_init(&sh->hash);
138 }
139
140 static inline void insert_hash(raid6_conf_t *conf, struct stripe_head *sh)
141 {
142         struct hlist_head *hp = stripe_hash(conf, sh->sector);
143
144         PRINTK("insert_hash(), stripe %llu\n", (unsigned long long)sh->sector);
145
146         CHECK_DEVLOCK();
147         hlist_add_head(&sh->hash, hp);
148 }
149
150
151 /* find an idle stripe, make sure it is unhashed, and return it. */
152 static struct stripe_head *get_free_stripe(raid6_conf_t *conf)
153 {
154         struct stripe_head *sh = NULL;
155         struct list_head *first;
156
157         CHECK_DEVLOCK();
158         if (list_empty(&conf->inactive_list))
159                 goto out;
160         first = conf->inactive_list.next;
161         sh = list_entry(first, struct stripe_head, lru);
162         list_del_init(first);
163         remove_hash(sh);
164         atomic_inc(&conf->active_stripes);
165 out:
166         return sh;
167 }
168
169 static void shrink_buffers(struct stripe_head *sh, int num)
170 {
171         struct page *p;
172         int i;
173
174         for (i=0; i<num ; i++) {
175                 p = sh->dev[i].page;
176                 if (!p)
177                         continue;
178                 sh->dev[i].page = NULL;
179                 put_page(p);
180         }
181 }
182
183 static int grow_buffers(struct stripe_head *sh, int num)
184 {
185         int i;
186
187         for (i=0; i<num; i++) {
188                 struct page *page;
189
190                 if (!(page = alloc_page(GFP_KERNEL))) {
191                         return 1;
192                 }
193                 sh->dev[i].page = page;
194         }
195         return 0;
196 }
197
198 static void raid6_build_block (struct stripe_head *sh, int i);
199
200 static void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx)
201 {
202         raid6_conf_t *conf = sh->raid_conf;
203         int disks = conf->raid_disks, i;
204
205         if (atomic_read(&sh->count) != 0)
206                 BUG();
207         if (test_bit(STRIPE_HANDLE, &sh->state))
208                 BUG();
209
210         CHECK_DEVLOCK();
211         PRINTK("init_stripe called, stripe %llu\n",
212                 (unsigned long long)sh->sector);
213
214         remove_hash(sh);
215
216         sh->sector = sector;
217         sh->pd_idx = pd_idx;
218         sh->state = 0;
219
220         for (i=disks; i--; ) {
221                 struct r5dev *dev = &sh->dev[i];
222
223                 if (dev->toread || dev->towrite || dev->written ||
224                     test_bit(R5_LOCKED, &dev->flags)) {
225                         PRINTK("sector=%llx i=%d %p %p %p %d\n",
226                                (unsigned long long)sh->sector, i, dev->toread,
227                                dev->towrite, dev->written,
228                                test_bit(R5_LOCKED, &dev->flags));
229                         BUG();
230                 }
231                 dev->flags = 0;
232                 raid6_build_block(sh, i);
233         }
234         insert_hash(conf, sh);
235 }
236
237 static struct stripe_head *__find_stripe(raid6_conf_t *conf, sector_t sector)
238 {
239         struct stripe_head *sh;
240         struct hlist_node *hn;
241
242         CHECK_DEVLOCK();
243         PRINTK("__find_stripe, sector %llu\n", (unsigned long long)sector);
244         hlist_for_each_entry (sh, hn,  stripe_hash(conf, sector), hash)
245                 if (sh->sector == sector)
246                         return sh;
247         PRINTK("__stripe %llu not in cache\n", (unsigned long long)sector);
248         return NULL;
249 }
250
251 static void unplug_slaves(mddev_t *mddev);
252
253 static struct stripe_head *get_active_stripe(raid6_conf_t *conf, sector_t sector,
254                                              int pd_idx, int noblock)
255 {
256         struct stripe_head *sh;
257
258         PRINTK("get_stripe, sector %llu\n", (unsigned long long)sector);
259
260         spin_lock_irq(&conf->device_lock);
261
262         do {
263                 wait_event_lock_irq(conf->wait_for_stripe,
264                                     conf->quiesce == 0,
265                                     conf->device_lock, /* nothing */);
266                 sh = __find_stripe(conf, sector);
267                 if (!sh) {
268                         if (!conf->inactive_blocked)
269                                 sh = get_free_stripe(conf);
270                         if (noblock && sh == NULL)
271                                 break;
272                         if (!sh) {
273                                 conf->inactive_blocked = 1;
274                                 wait_event_lock_irq(conf->wait_for_stripe,
275                                                     !list_empty(&conf->inactive_list) &&
276                                                     (atomic_read(&conf->active_stripes) < (NR_STRIPES *3/4)
277                                                      || !conf->inactive_blocked),
278                                                     conf->device_lock,
279                                                     unplug_slaves(conf->mddev);
280                                         );
281                                 conf->inactive_blocked = 0;
282                         } else
283                                 init_stripe(sh, sector, pd_idx);
284                 } else {
285                         if (atomic_read(&sh->count)) {
286                                 if (!list_empty(&sh->lru))
287                                         BUG();
288                         } else {
289                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
290                                         atomic_inc(&conf->active_stripes);
291                                 if (list_empty(&sh->lru))
292                                         BUG();
293                                 list_del_init(&sh->lru);
294                         }
295                 }
296         } while (sh == NULL);
297
298         if (sh)
299                 atomic_inc(&sh->count);
300
301         spin_unlock_irq(&conf->device_lock);
302         return sh;
303 }
304
305 static int grow_stripes(raid6_conf_t *conf, int num)
306 {
307         struct stripe_head *sh;
308         kmem_cache_t *sc;
309         int devs = conf->raid_disks;
310
311         sprintf(conf->cache_name, "raid6/%s", mdname(conf->mddev));
312
313         sc = kmem_cache_create(conf->cache_name,
314                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
315                                0, 0, NULL, NULL);
316         if (!sc)
317                 return 1;
318         conf->slab_cache = sc;
319         while (num--) {
320                 sh = kmem_cache_alloc(sc, GFP_KERNEL);
321                 if (!sh)
322                         return 1;
323                 memset(sh, 0, sizeof(*sh) + (devs-1)*sizeof(struct r5dev));
324                 sh->raid_conf = conf;
325                 spin_lock_init(&sh->lock);
326
327                 if (grow_buffers(sh, conf->raid_disks)) {
328                         shrink_buffers(sh, conf->raid_disks);
329                         kmem_cache_free(sc, sh);
330                         return 1;
331                 }
332                 /* we just created an active stripe so... */
333                 atomic_set(&sh->count, 1);
334                 atomic_inc(&conf->active_stripes);
335                 INIT_LIST_HEAD(&sh->lru);
336                 release_stripe(sh);
337         }
338         return 0;
339 }
340
341 static void shrink_stripes(raid6_conf_t *conf)
342 {
343         struct stripe_head *sh;
344
345         while (1) {
346                 spin_lock_irq(&conf->device_lock);
347                 sh = get_free_stripe(conf);
348                 spin_unlock_irq(&conf->device_lock);
349                 if (!sh)
350                         break;
351                 if (atomic_read(&sh->count))
352                         BUG();
353                 shrink_buffers(sh, conf->raid_disks);
354                 kmem_cache_free(conf->slab_cache, sh);
355                 atomic_dec(&conf->active_stripes);
356         }
357         kmem_cache_destroy(conf->slab_cache);
358         conf->slab_cache = NULL;
359 }
360
361 static int raid6_end_read_request(struct bio * bi, unsigned int bytes_done,
362                                   int error)
363 {
364         struct stripe_head *sh = bi->bi_private;
365         raid6_conf_t *conf = sh->raid_conf;
366         int disks = conf->raid_disks, i;
367         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
368
369         if (bi->bi_size)
370                 return 1;
371
372         for (i=0 ; i<disks; i++)
373                 if (bi == &sh->dev[i].req)
374                         break;
375
376         PRINTK("end_read_request %llu/%d, count: %d, uptodate %d.\n",
377                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
378                 uptodate);
379         if (i == disks) {
380                 BUG();
381                 return 0;
382         }
383
384         if (uptodate) {
385 #if 0
386                 struct bio *bio;
387                 unsigned long flags;
388                 spin_lock_irqsave(&conf->device_lock, flags);
389                 /* we can return a buffer if we bypassed the cache or
390                  * if the top buffer is not in highmem.  If there are
391                  * multiple buffers, leave the extra work to
392                  * handle_stripe
393                  */
394                 buffer = sh->bh_read[i];
395                 if (buffer &&
396                     (!PageHighMem(buffer->b_page)
397                      || buffer->b_page == bh->b_page )
398                         ) {
399                         sh->bh_read[i] = buffer->b_reqnext;
400                         buffer->b_reqnext = NULL;
401                 } else
402                         buffer = NULL;
403                 spin_unlock_irqrestore(&conf->device_lock, flags);
404                 if (sh->bh_page[i]==bh->b_page)
405                         set_buffer_uptodate(bh);
406                 if (buffer) {
407                         if (buffer->b_page != bh->b_page)
408                                 memcpy(buffer->b_data, bh->b_data, bh->b_size);
409                         buffer->b_end_io(buffer, 1);
410                 }
411 #else
412                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
413 #endif
414                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
415                         printk(KERN_INFO "raid6: read error corrected!!\n");
416                         clear_bit(R5_ReadError, &sh->dev[i].flags);
417                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
418                 }
419                 if (atomic_read(&conf->disks[i].rdev->read_errors))
420                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
421         } else {
422                 int retry = 0;
423                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
424                 atomic_inc(&conf->disks[i].rdev->read_errors);
425                 if (conf->mddev->degraded)
426                         printk(KERN_WARNING "raid6: read error not correctable.\n");
427                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
428                         /* Oh, no!!! */
429                         printk(KERN_WARNING "raid6: read error NOT corrected!!\n");
430                 else if (atomic_read(&conf->disks[i].rdev->read_errors)
431                          > conf->max_nr_stripes)
432                         printk(KERN_WARNING
433                                "raid6: Too many read errors, failing device.\n");
434                 else
435                         retry = 1;
436                 if (retry)
437                         set_bit(R5_ReadError, &sh->dev[i].flags);
438                 else {
439                         clear_bit(R5_ReadError, &sh->dev[i].flags);
440                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
441                         md_error(conf->mddev, conf->disks[i].rdev);
442                 }
443         }
444         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
445 #if 0
446         /* must restore b_page before unlocking buffer... */
447         if (sh->bh_page[i] != bh->b_page) {
448                 bh->b_page = sh->bh_page[i];
449                 bh->b_data = page_address(bh->b_page);
450                 clear_buffer_uptodate(bh);
451         }
452 #endif
453         clear_bit(R5_LOCKED, &sh->dev[i].flags);
454         set_bit(STRIPE_HANDLE, &sh->state);
455         release_stripe(sh);
456         return 0;
457 }
458
459 static int raid6_end_write_request (struct bio *bi, unsigned int bytes_done,
460                                     int error)
461 {
462         struct stripe_head *sh = bi->bi_private;
463         raid6_conf_t *conf = sh->raid_conf;
464         int disks = conf->raid_disks, i;
465         unsigned long flags;
466         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
467
468         if (bi->bi_size)
469                 return 1;
470
471         for (i=0 ; i<disks; i++)
472                 if (bi == &sh->dev[i].req)
473                         break;
474
475         PRINTK("end_write_request %llu/%d, count %d, uptodate: %d.\n",
476                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
477                 uptodate);
478         if (i == disks) {
479                 BUG();
480                 return 0;
481         }
482
483         spin_lock_irqsave(&conf->device_lock, flags);
484         if (!uptodate)
485                 md_error(conf->mddev, conf->disks[i].rdev);
486
487         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
488
489         clear_bit(R5_LOCKED, &sh->dev[i].flags);
490         set_bit(STRIPE_HANDLE, &sh->state);
491         __release_stripe(conf, sh);
492         spin_unlock_irqrestore(&conf->device_lock, flags);
493         return 0;
494 }
495
496
497 static sector_t compute_blocknr(struct stripe_head *sh, int i);
498
499 static void raid6_build_block (struct stripe_head *sh, int i)
500 {
501         struct r5dev *dev = &sh->dev[i];
502         int pd_idx = sh->pd_idx;
503         int qd_idx = raid6_next_disk(pd_idx, sh->raid_conf->raid_disks);
504
505         bio_init(&dev->req);
506         dev->req.bi_io_vec = &dev->vec;
507         dev->req.bi_vcnt++;
508         dev->req.bi_max_vecs++;
509         dev->vec.bv_page = dev->page;
510         dev->vec.bv_len = STRIPE_SIZE;
511         dev->vec.bv_offset = 0;
512
513         dev->req.bi_sector = sh->sector;
514         dev->req.bi_private = sh;
515
516         dev->flags = 0;
517         if (i != pd_idx && i != qd_idx)
518                 dev->sector = compute_blocknr(sh, i);
519 }
520
521 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
522 {
523         char b[BDEVNAME_SIZE];
524         raid6_conf_t *conf = (raid6_conf_t *) mddev->private;
525         PRINTK("raid6: error called\n");
526
527         if (!test_bit(Faulty, &rdev->flags)) {
528                 mddev->sb_dirty = 1;
529                 if (test_bit(In_sync, &rdev->flags)) {
530                         conf->working_disks--;
531                         mddev->degraded++;
532                         conf->failed_disks++;
533                         clear_bit(In_sync, &rdev->flags);
534                         /*
535                          * if recovery was running, make sure it aborts.
536                          */
537                         set_bit(MD_RECOVERY_ERR, &mddev->recovery);
538                 }
539                 set_bit(Faulty, &rdev->flags);
540                 printk (KERN_ALERT
541                         "raid6: Disk failure on %s, disabling device."
542                         " Operation continuing on %d devices\n",
543                         bdevname(rdev->bdev,b), conf->working_disks);
544         }
545 }
546
547 /*
548  * Input: a 'big' sector number,
549  * Output: index of the data and parity disk, and the sector # in them.
550  */
551 static sector_t raid6_compute_sector(sector_t r_sector, unsigned int raid_disks,
552                         unsigned int data_disks, unsigned int * dd_idx,
553                         unsigned int * pd_idx, raid6_conf_t *conf)
554 {
555         long stripe;
556         unsigned long chunk_number;
557         unsigned int chunk_offset;
558         sector_t new_sector;
559         int sectors_per_chunk = conf->chunk_size >> 9;
560
561         /* First compute the information on this sector */
562
563         /*
564          * Compute the chunk number and the sector offset inside the chunk
565          */
566         chunk_offset = sector_div(r_sector, sectors_per_chunk);
567         chunk_number = r_sector;
568         if ( r_sector != chunk_number ) {
569                 printk(KERN_CRIT "raid6: ERROR: r_sector = %llu, chunk_number = %lu\n",
570                        (unsigned long long)r_sector, (unsigned long)chunk_number);
571                 BUG();
572         }
573
574         /*
575          * Compute the stripe number
576          */
577         stripe = chunk_number / data_disks;
578
579         /*
580          * Compute the data disk and parity disk indexes inside the stripe
581          */
582         *dd_idx = chunk_number % data_disks;
583
584         /*
585          * Select the parity disk based on the user selected algorithm.
586          */
587
588         /**** FIX THIS ****/
589         switch (conf->algorithm) {
590         case ALGORITHM_LEFT_ASYMMETRIC:
591                 *pd_idx = raid_disks - 1 - (stripe % raid_disks);
592                 if (*pd_idx == raid_disks-1)
593                         (*dd_idx)++;    /* Q D D D P */
594                 else if (*dd_idx >= *pd_idx)
595                         (*dd_idx) += 2; /* D D P Q D */
596                 break;
597         case ALGORITHM_RIGHT_ASYMMETRIC:
598                 *pd_idx = stripe % raid_disks;
599                 if (*pd_idx == raid_disks-1)
600                         (*dd_idx)++;    /* Q D D D P */
601                 else if (*dd_idx >= *pd_idx)
602                         (*dd_idx) += 2; /* D D P Q D */
603                 break;
604         case ALGORITHM_LEFT_SYMMETRIC:
605                 *pd_idx = raid_disks - 1 - (stripe % raid_disks);
606                 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
607                 break;
608         case ALGORITHM_RIGHT_SYMMETRIC:
609                 *pd_idx = stripe % raid_disks;
610                 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
611                 break;
612         default:
613                 printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
614                         conf->algorithm);
615         }
616
617         PRINTK("raid6: chunk_number = %lu, pd_idx = %u, dd_idx = %u\n",
618                chunk_number, *pd_idx, *dd_idx);
619
620         /*
621          * Finally, compute the new sector number
622          */
623         new_sector = (sector_t) stripe * sectors_per_chunk + chunk_offset;
624         return new_sector;
625 }
626
627
628 static sector_t compute_blocknr(struct stripe_head *sh, int i)
629 {
630         raid6_conf_t *conf = sh->raid_conf;
631         int raid_disks = conf->raid_disks, data_disks = raid_disks - 2;
632         sector_t new_sector = sh->sector, check;
633         int sectors_per_chunk = conf->chunk_size >> 9;
634         sector_t stripe;
635         int chunk_offset;
636         int chunk_number, dummy1, dummy2, dd_idx = i;
637         sector_t r_sector;
638         int i0 = i;
639
640         chunk_offset = sector_div(new_sector, sectors_per_chunk);
641         stripe = new_sector;
642         if ( new_sector != stripe ) {
643                 printk(KERN_CRIT "raid6: ERROR: new_sector = %llu, stripe = %lu\n",
644                        (unsigned long long)new_sector, (unsigned long)stripe);
645                 BUG();
646         }
647
648         switch (conf->algorithm) {
649                 case ALGORITHM_LEFT_ASYMMETRIC:
650                 case ALGORITHM_RIGHT_ASYMMETRIC:
651                         if (sh->pd_idx == raid_disks-1)
652                                 i--;    /* Q D D D P */
653                         else if (i > sh->pd_idx)
654                                 i -= 2; /* D D P Q D */
655                         break;
656                 case ALGORITHM_LEFT_SYMMETRIC:
657                 case ALGORITHM_RIGHT_SYMMETRIC:
658                         if (sh->pd_idx == raid_disks-1)
659                                 i--; /* Q D D D P */
660                         else {
661                                 /* D D P Q D */
662                                 if (i < sh->pd_idx)
663                                         i += raid_disks;
664                                 i -= (sh->pd_idx + 2);
665                         }
666                         break;
667                 default:
668                         printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
669                                 conf->algorithm);
670         }
671
672         PRINTK("raid6: compute_blocknr: pd_idx = %u, i0 = %u, i = %u\n", sh->pd_idx, i0, i);
673
674         chunk_number = stripe * data_disks + i;
675         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
676
677         check = raid6_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
678         if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
679                 printk(KERN_CRIT "raid6: compute_blocknr: map not correct\n");
680                 return 0;
681         }
682         return r_sector;
683 }
684
685
686
687 /*
688  * Copy data between a page in the stripe cache, and one or more bion
689  * The page could align with the middle of the bio, or there could be
690  * several bion, each with several bio_vecs, which cover part of the page
691  * Multiple bion are linked together on bi_next.  There may be extras
692  * at the end of this list.  We ignore them.
693  */
694 static void copy_data(int frombio, struct bio *bio,
695                      struct page *page,
696                      sector_t sector)
697 {
698         char *pa = page_address(page);
699         struct bio_vec *bvl;
700         int i;
701         int page_offset;
702
703         if (bio->bi_sector >= sector)
704                 page_offset = (signed)(bio->bi_sector - sector) * 512;
705         else
706                 page_offset = (signed)(sector - bio->bi_sector) * -512;
707         bio_for_each_segment(bvl, bio, i) {
708                 int len = bio_iovec_idx(bio,i)->bv_len;
709                 int clen;
710                 int b_offset = 0;
711
712                 if (page_offset < 0) {
713                         b_offset = -page_offset;
714                         page_offset += b_offset;
715                         len -= b_offset;
716                 }
717
718                 if (len > 0 && page_offset + len > STRIPE_SIZE)
719                         clen = STRIPE_SIZE - page_offset;
720                 else clen = len;
721
722                 if (clen > 0) {
723                         char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
724                         if (frombio)
725                                 memcpy(pa+page_offset, ba+b_offset, clen);
726                         else
727                                 memcpy(ba+b_offset, pa+page_offset, clen);
728                         __bio_kunmap_atomic(ba, KM_USER0);
729                 }
730                 if (clen < len) /* hit end of page */
731                         break;
732                 page_offset +=  len;
733         }
734 }
735
736 #define check_xor()     do {                                            \
737                            if (count == MAX_XOR_BLOCKS) {               \
738                                 xor_block(count, STRIPE_SIZE, ptr);     \
739                                 count = 1;                              \
740                            }                                            \
741                         } while(0)
742
743 /* Compute P and Q syndromes */
744 static void compute_parity(struct stripe_head *sh, int method)
745 {
746         raid6_conf_t *conf = sh->raid_conf;
747         int i, pd_idx = sh->pd_idx, qd_idx, d0_idx, disks = conf->raid_disks, count;
748         struct bio *chosen;
749         /**** FIX THIS: This could be very bad if disks is close to 256 ****/
750         void *ptrs[disks];
751
752         qd_idx = raid6_next_disk(pd_idx, disks);
753         d0_idx = raid6_next_disk(qd_idx, disks);
754
755         PRINTK("compute_parity, stripe %llu, method %d\n",
756                 (unsigned long long)sh->sector, method);
757
758         switch(method) {
759         case READ_MODIFY_WRITE:
760                 BUG();          /* READ_MODIFY_WRITE N/A for RAID-6 */
761         case RECONSTRUCT_WRITE:
762                 for (i= disks; i-- ;)
763                         if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
764                                 chosen = sh->dev[i].towrite;
765                                 sh->dev[i].towrite = NULL;
766
767                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
768                                         wake_up(&conf->wait_for_overlap);
769
770                                 if (sh->dev[i].written) BUG();
771                                 sh->dev[i].written = chosen;
772                         }
773                 break;
774         case CHECK_PARITY:
775                 BUG();          /* Not implemented yet */
776         }
777
778         for (i = disks; i--;)
779                 if (sh->dev[i].written) {
780                         sector_t sector = sh->dev[i].sector;
781                         struct bio *wbi = sh->dev[i].written;
782                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
783                                 copy_data(1, wbi, sh->dev[i].page, sector);
784                                 wbi = r5_next_bio(wbi, sector);
785                         }
786
787                         set_bit(R5_LOCKED, &sh->dev[i].flags);
788                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
789                 }
790
791 //      switch(method) {
792 //      case RECONSTRUCT_WRITE:
793 //      case CHECK_PARITY:
794 //      case UPDATE_PARITY:
795                 /* Note that unlike RAID-5, the ordering of the disks matters greatly. */
796                 /* FIX: Is this ordering of drives even remotely optimal? */
797                 count = 0;
798                 i = d0_idx;
799                 do {
800                         ptrs[count++] = page_address(sh->dev[i].page);
801                         if (count <= disks-2 && !test_bit(R5_UPTODATE, &sh->dev[i].flags))
802                                 printk("block %d/%d not uptodate on parity calc\n", i,count);
803                         i = raid6_next_disk(i, disks);
804                 } while ( i != d0_idx );
805 //              break;
806 //      }
807
808         raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs);
809
810         switch(method) {
811         case RECONSTRUCT_WRITE:
812                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
813                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
814                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
815                 set_bit(R5_LOCKED,   &sh->dev[qd_idx].flags);
816                 break;
817         case UPDATE_PARITY:
818                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
819                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
820                 break;
821         }
822 }
823
824 /* Compute one missing block */
825 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
826 {
827         raid6_conf_t *conf = sh->raid_conf;
828         int i, count, disks = conf->raid_disks;
829         void *ptr[MAX_XOR_BLOCKS], *p;
830         int pd_idx = sh->pd_idx;
831         int qd_idx = raid6_next_disk(pd_idx, disks);
832
833         PRINTK("compute_block_1, stripe %llu, idx %d\n",
834                 (unsigned long long)sh->sector, dd_idx);
835
836         if ( dd_idx == qd_idx ) {
837                 /* We're actually computing the Q drive */
838                 compute_parity(sh, UPDATE_PARITY);
839         } else {
840                 ptr[0] = page_address(sh->dev[dd_idx].page);
841                 if (!nozero) memset(ptr[0], 0, STRIPE_SIZE);
842                 count = 1;
843                 for (i = disks ; i--; ) {
844                         if (i == dd_idx || i == qd_idx)
845                                 continue;
846                         p = page_address(sh->dev[i].page);
847                         if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
848                                 ptr[count++] = p;
849                         else
850                                 printk("compute_block() %d, stripe %llu, %d"
851                                        " not present\n", dd_idx,
852                                        (unsigned long long)sh->sector, i);
853
854                         check_xor();
855                 }
856                 if (count != 1)
857                         xor_block(count, STRIPE_SIZE, ptr);
858                 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
859                 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
860         }
861 }
862
863 /* Compute two missing blocks */
864 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
865 {
866         raid6_conf_t *conf = sh->raid_conf;
867         int i, count, disks = conf->raid_disks;
868         int pd_idx = sh->pd_idx;
869         int qd_idx = raid6_next_disk(pd_idx, disks);
870         int d0_idx = raid6_next_disk(qd_idx, disks);
871         int faila, failb;
872
873         /* faila and failb are disk numbers relative to d0_idx */
874         /* pd_idx become disks-2 and qd_idx become disks-1 */
875         faila = (dd_idx1 < d0_idx) ? dd_idx1+(disks-d0_idx) : dd_idx1-d0_idx;
876         failb = (dd_idx2 < d0_idx) ? dd_idx2+(disks-d0_idx) : dd_idx2-d0_idx;
877
878         BUG_ON(faila == failb);
879         if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
880
881         PRINTK("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
882                (unsigned long long)sh->sector, dd_idx1, dd_idx2, faila, failb);
883
884         if ( failb == disks-1 ) {
885                 /* Q disk is one of the missing disks */
886                 if ( faila == disks-2 ) {
887                         /* Missing P+Q, just recompute */
888                         compute_parity(sh, UPDATE_PARITY);
889                         return;
890                 } else {
891                         /* We're missing D+Q; recompute D from P */
892                         compute_block_1(sh, (dd_idx1 == qd_idx) ? dd_idx2 : dd_idx1, 0);
893                         compute_parity(sh, UPDATE_PARITY); /* Is this necessary? */
894                         return;
895                 }
896         }
897
898         /* We're missing D+P or D+D; build pointer table */
899         {
900                 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
901                 void *ptrs[disks];
902
903                 count = 0;
904                 i = d0_idx;
905                 do {
906                         ptrs[count++] = page_address(sh->dev[i].page);
907                         i = raid6_next_disk(i, disks);
908                         if (i != dd_idx1 && i != dd_idx2 &&
909                             !test_bit(R5_UPTODATE, &sh->dev[i].flags))
910                                 printk("compute_2 with missing block %d/%d\n", count, i);
911                 } while ( i != d0_idx );
912
913                 if ( failb == disks-2 ) {
914                         /* We're missing D+P. */
915                         raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs);
916                 } else {
917                         /* We're missing D+D. */
918                         raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs);
919                 }
920
921                 /* Both the above update both missing blocks */
922                 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
923                 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
924         }
925 }
926
927
928 /*
929  * Each stripe/dev can have one or more bion attached.
930  * toread/towrite point to the first in a chain.
931  * The bi_next chain must be in order.
932  */
933 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
934 {
935         struct bio **bip;
936         raid6_conf_t *conf = sh->raid_conf;
937         int firstwrite=0;
938
939         PRINTK("adding bh b#%llu to stripe s#%llu\n",
940                 (unsigned long long)bi->bi_sector,
941                 (unsigned long long)sh->sector);
942
943
944         spin_lock(&sh->lock);
945         spin_lock_irq(&conf->device_lock);
946         if (forwrite) {
947                 bip = &sh->dev[dd_idx].towrite;
948                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
949                         firstwrite = 1;
950         } else
951                 bip = &sh->dev[dd_idx].toread;
952         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
953                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
954                         goto overlap;
955                 bip = &(*bip)->bi_next;
956         }
957         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
958                 goto overlap;
959
960         if (*bip && bi->bi_next && (*bip) != bi->bi_next)
961                 BUG();
962         if (*bip)
963                 bi->bi_next = *bip;
964         *bip = bi;
965         bi->bi_phys_segments ++;
966         spin_unlock_irq(&conf->device_lock);
967         spin_unlock(&sh->lock);
968
969         PRINTK("added bi b#%llu to stripe s#%llu, disk %d.\n",
970                 (unsigned long long)bi->bi_sector,
971                 (unsigned long long)sh->sector, dd_idx);
972
973         if (conf->mddev->bitmap && firstwrite) {
974                 sh->bm_seq = conf->seq_write;
975                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
976                                   STRIPE_SECTORS, 0);
977                 set_bit(STRIPE_BIT_DELAY, &sh->state);
978         }
979
980         if (forwrite) {
981                 /* check if page is covered */
982                 sector_t sector = sh->dev[dd_idx].sector;
983                 for (bi=sh->dev[dd_idx].towrite;
984                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
985                              bi && bi->bi_sector <= sector;
986                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
987                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
988                                 sector = bi->bi_sector + (bi->bi_size>>9);
989                 }
990                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
991                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
992         }
993         return 1;
994
995  overlap:
996         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
997         spin_unlock_irq(&conf->device_lock);
998         spin_unlock(&sh->lock);
999         return 0;
1000 }
1001
1002
1003 static int page_is_zero(struct page *p)
1004 {
1005         char *a = page_address(p);
1006         return ((*(u32*)a) == 0 &&
1007                 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1008 }
1009 /*
1010  * handle_stripe - do things to a stripe.
1011  *
1012  * We lock the stripe and then examine the state of various bits
1013  * to see what needs to be done.
1014  * Possible results:
1015  *    return some read request which now have data
1016  *    return some write requests which are safely on disc
1017  *    schedule a read on some buffers
1018  *    schedule a write of some buffers
1019  *    return confirmation of parity correctness
1020  *
1021  * Parity calculations are done inside the stripe lock
1022  * buffers are taken off read_list or write_list, and bh_cache buffers
1023  * get BH_Lock set before the stripe lock is released.
1024  *
1025  */
1026
1027 static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
1028 {
1029         raid6_conf_t *conf = sh->raid_conf;
1030         int disks = conf->raid_disks;
1031         struct bio *return_bi= NULL;
1032         struct bio *bi;
1033         int i;
1034         int syncing;
1035         int locked=0, uptodate=0, to_read=0, to_write=0, failed=0, written=0;
1036         int non_overwrite = 0;
1037         int failed_num[2] = {0, 0};
1038         struct r5dev *dev, *pdev, *qdev;
1039         int pd_idx = sh->pd_idx;
1040         int qd_idx = raid6_next_disk(pd_idx, disks);
1041         int p_failed, q_failed;
1042
1043         PRINTK("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d, qd_idx=%d\n",
1044                (unsigned long long)sh->sector, sh->state, atomic_read(&sh->count),
1045                pd_idx, qd_idx);
1046
1047         spin_lock(&sh->lock);
1048         clear_bit(STRIPE_HANDLE, &sh->state);
1049         clear_bit(STRIPE_DELAYED, &sh->state);
1050
1051         syncing = test_bit(STRIPE_SYNCING, &sh->state);
1052         /* Now to look around and see what can be done */
1053
1054         rcu_read_lock();
1055         for (i=disks; i--; ) {
1056                 mdk_rdev_t *rdev;
1057                 dev = &sh->dev[i];
1058                 clear_bit(R5_Insync, &dev->flags);
1059
1060                 PRINTK("check %d: state 0x%lx read %p write %p written %p\n",
1061                         i, dev->flags, dev->toread, dev->towrite, dev->written);
1062                 /* maybe we can reply to a read */
1063                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
1064                         struct bio *rbi, *rbi2;
1065                         PRINTK("Return read for disc %d\n", i);
1066                         spin_lock_irq(&conf->device_lock);
1067                         rbi = dev->toread;
1068                         dev->toread = NULL;
1069                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
1070                                 wake_up(&conf->wait_for_overlap);
1071                         spin_unlock_irq(&conf->device_lock);
1072                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1073                                 copy_data(0, rbi, dev->page, dev->sector);
1074                                 rbi2 = r5_next_bio(rbi, dev->sector);
1075                                 spin_lock_irq(&conf->device_lock);
1076                                 if (--rbi->bi_phys_segments == 0) {
1077                                         rbi->bi_next = return_bi;
1078                                         return_bi = rbi;
1079                                 }
1080                                 spin_unlock_irq(&conf->device_lock);
1081                                 rbi = rbi2;
1082                         }
1083                 }
1084
1085                 /* now count some things */
1086                 if (test_bit(R5_LOCKED, &dev->flags)) locked++;
1087                 if (test_bit(R5_UPTODATE, &dev->flags)) uptodate++;
1088
1089
1090                 if (dev->toread) to_read++;
1091                 if (dev->towrite) {
1092                         to_write++;
1093                         if (!test_bit(R5_OVERWRITE, &dev->flags))
1094                                 non_overwrite++;
1095                 }
1096                 if (dev->written) written++;
1097                 rdev = rcu_dereference(conf->disks[i].rdev);
1098                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
1099                         /* The ReadError flag will just be confusing now */
1100                         clear_bit(R5_ReadError, &dev->flags);
1101                         clear_bit(R5_ReWrite, &dev->flags);
1102                 }
1103                 if (!rdev || !test_bit(In_sync, &rdev->flags)
1104                     || test_bit(R5_ReadError, &dev->flags)) {
1105                         if ( failed < 2 )
1106                                 failed_num[failed] = i;
1107                         failed++;
1108                 } else
1109                         set_bit(R5_Insync, &dev->flags);
1110         }
1111         rcu_read_unlock();
1112         PRINTK("locked=%d uptodate=%d to_read=%d"
1113                " to_write=%d failed=%d failed_num=%d,%d\n",
1114                locked, uptodate, to_read, to_write, failed,
1115                failed_num[0], failed_num[1]);
1116         /* check if the array has lost >2 devices and, if so, some requests might
1117          * need to be failed
1118          */
1119         if (failed > 2 && to_read+to_write+written) {
1120                 for (i=disks; i--; ) {
1121                         int bitmap_end = 0;
1122
1123                         if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1124                                 mdk_rdev_t *rdev;
1125                                 rcu_read_lock();
1126                                 rdev = rcu_dereference(conf->disks[i].rdev);
1127                                 if (rdev && test_bit(In_sync, &rdev->flags))
1128                                         /* multiple read failures in one stripe */
1129                                         md_error(conf->mddev, rdev);
1130                                 rcu_read_unlock();
1131                         }
1132
1133                         spin_lock_irq(&conf->device_lock);
1134                         /* fail all writes first */
1135                         bi = sh->dev[i].towrite;
1136                         sh->dev[i].towrite = NULL;
1137                         if (bi) { to_write--; bitmap_end = 1; }
1138
1139                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1140                                 wake_up(&conf->wait_for_overlap);
1141
1142                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1143                                 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1144                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1145                                 if (--bi->bi_phys_segments == 0) {
1146                                         md_write_end(conf->mddev);
1147                                         bi->bi_next = return_bi;
1148                                         return_bi = bi;
1149                                 }
1150                                 bi = nextbi;
1151                         }
1152                         /* and fail all 'written' */
1153                         bi = sh->dev[i].written;
1154                         sh->dev[i].written = NULL;
1155                         if (bi) bitmap_end = 1;
1156                         while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS) {
1157                                 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1158                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1159                                 if (--bi->bi_phys_segments == 0) {
1160                                         md_write_end(conf->mddev);
1161                                         bi->bi_next = return_bi;
1162                                         return_bi = bi;
1163                                 }
1164                                 bi = bi2;
1165                         }
1166
1167                         /* fail any reads if this device is non-operational */
1168                         if (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1169                             test_bit(R5_ReadError, &sh->dev[i].flags)) {
1170                                 bi = sh->dev[i].toread;
1171                                 sh->dev[i].toread = NULL;
1172                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1173                                         wake_up(&conf->wait_for_overlap);
1174                                 if (bi) to_read--;
1175                                 while (bi && bi->bi_sector < sh->dev[i].sector + STRIPE_SECTORS){
1176                                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1177                                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1178                                         if (--bi->bi_phys_segments == 0) {
1179                                                 bi->bi_next = return_bi;
1180                                                 return_bi = bi;
1181                                         }
1182                                         bi = nextbi;
1183                                 }
1184                         }
1185                         spin_unlock_irq(&conf->device_lock);
1186                         if (bitmap_end)
1187                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1188                                                 STRIPE_SECTORS, 0, 0);
1189                 }
1190         }
1191         if (failed > 2 && syncing) {
1192                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
1193                 clear_bit(STRIPE_SYNCING, &sh->state);
1194                 syncing = 0;
1195         }
1196
1197         /*
1198          * might be able to return some write requests if the parity blocks
1199          * are safe, or on a failed drive
1200          */
1201         pdev = &sh->dev[pd_idx];
1202         p_failed = (failed >= 1 && failed_num[0] == pd_idx)
1203                 || (failed >= 2 && failed_num[1] == pd_idx);
1204         qdev = &sh->dev[qd_idx];
1205         q_failed = (failed >= 1 && failed_num[0] == qd_idx)
1206                 || (failed >= 2 && failed_num[1] == qd_idx);
1207
1208         if ( written &&
1209              ( p_failed || ((test_bit(R5_Insync, &pdev->flags)
1210                              && !test_bit(R5_LOCKED, &pdev->flags)
1211                              && test_bit(R5_UPTODATE, &pdev->flags))) ) &&
1212              ( q_failed || ((test_bit(R5_Insync, &qdev->flags)
1213                              && !test_bit(R5_LOCKED, &qdev->flags)
1214                              && test_bit(R5_UPTODATE, &qdev->flags))) ) ) {
1215                 /* any written block on an uptodate or failed drive can be
1216                  * returned.  Note that if we 'wrote' to a failed drive,
1217                  * it will be UPTODATE, but never LOCKED, so we don't need
1218                  * to test 'failed' directly.
1219                  */
1220                 for (i=disks; i--; )
1221                         if (sh->dev[i].written) {
1222                                 dev = &sh->dev[i];
1223                                 if (!test_bit(R5_LOCKED, &dev->flags) &&
1224                                     test_bit(R5_UPTODATE, &dev->flags) ) {
1225                                         /* We can return any write requests */
1226                                         int bitmap_end = 0;
1227                                         struct bio *wbi, *wbi2;
1228                                         PRINTK("Return write for stripe %llu disc %d\n",
1229                                                (unsigned long long)sh->sector, i);
1230                                         spin_lock_irq(&conf->device_lock);
1231                                         wbi = dev->written;
1232                                         dev->written = NULL;
1233                                         while (wbi && wbi->bi_sector < dev->sector + STRIPE_SECTORS) {
1234                                                 wbi2 = r5_next_bio(wbi, dev->sector);
1235                                                 if (--wbi->bi_phys_segments == 0) {
1236                                                         md_write_end(conf->mddev);
1237                                                         wbi->bi_next = return_bi;
1238                                                         return_bi = wbi;
1239                                                 }
1240                                                 wbi = wbi2;
1241                                         }
1242                                         if (dev->towrite == NULL)
1243                                                 bitmap_end = 1;
1244                                         spin_unlock_irq(&conf->device_lock);
1245                                         if (bitmap_end)
1246                                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1247                                                                 STRIPE_SECTORS,
1248                                                                 !test_bit(STRIPE_DEGRADED, &sh->state), 0);
1249                                 }
1250                         }
1251         }
1252
1253         /* Now we might consider reading some blocks, either to check/generate
1254          * parity, or to satisfy requests
1255          * or to load a block that is being partially written.
1256          */
1257         if (to_read || non_overwrite || (to_write && failed) || (syncing && (uptodate < disks))) {
1258                 for (i=disks; i--;) {
1259                         dev = &sh->dev[i];
1260                         if (!test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1261                             (dev->toread ||
1262                              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1263                              syncing ||
1264                              (failed >= 1 && (sh->dev[failed_num[0]].toread || to_write)) ||
1265                              (failed >= 2 && (sh->dev[failed_num[1]].toread || to_write))
1266                                     )
1267                                 ) {
1268                                 /* we would like to get this block, possibly
1269                                  * by computing it, but we might not be able to
1270                                  */
1271                                 if (uptodate == disks-1) {
1272                                         PRINTK("Computing stripe %llu block %d\n",
1273                                                (unsigned long long)sh->sector, i);
1274                                         compute_block_1(sh, i, 0);
1275                                         uptodate++;
1276                                 } else if ( uptodate == disks-2 && failed >= 2 ) {
1277                                         /* Computing 2-failure is *very* expensive; only do it if failed >= 2 */
1278                                         int other;
1279                                         for (other=disks; other--;) {
1280                                                 if ( other == i )
1281                                                         continue;
1282                                                 if ( !test_bit(R5_UPTODATE, &sh->dev[other].flags) )
1283                                                         break;
1284                                         }
1285                                         BUG_ON(other < 0);
1286                                         PRINTK("Computing stripe %llu blocks %d,%d\n",
1287                                                (unsigned long long)sh->sector, i, other);
1288                                         compute_block_2(sh, i, other);
1289                                         uptodate += 2;
1290                                 } else if (test_bit(R5_Insync, &dev->flags)) {
1291                                         set_bit(R5_LOCKED, &dev->flags);
1292                                         set_bit(R5_Wantread, &dev->flags);
1293 #if 0
1294                                         /* if I am just reading this block and we don't have
1295                                            a failed drive, or any pending writes then sidestep the cache */
1296                                         if (sh->bh_read[i] && !sh->bh_read[i]->b_reqnext &&
1297                                             ! syncing && !failed && !to_write) {
1298                                                 sh->bh_cache[i]->b_page =  sh->bh_read[i]->b_page;
1299                                                 sh->bh_cache[i]->b_data =  sh->bh_read[i]->b_data;
1300                                         }
1301 #endif
1302                                         locked++;
1303                                         PRINTK("Reading block %d (sync=%d)\n",
1304                                                 i, syncing);
1305                                 }
1306                         }
1307                 }
1308                 set_bit(STRIPE_HANDLE, &sh->state);
1309         }
1310
1311         /* now to consider writing and what else, if anything should be read */
1312         if (to_write) {
1313                 int rcw=0, must_compute=0;
1314                 for (i=disks ; i--;) {
1315                         dev = &sh->dev[i];
1316                         /* Would I have to read this buffer for reconstruct_write */
1317                         if (!test_bit(R5_OVERWRITE, &dev->flags)
1318                             && i != pd_idx && i != qd_idx
1319                             && (!test_bit(R5_LOCKED, &dev->flags)
1320 #if 0
1321                                 || sh->bh_page[i] != bh->b_page
1322 #endif
1323                                     ) &&
1324                             !test_bit(R5_UPTODATE, &dev->flags)) {
1325                                 if (test_bit(R5_Insync, &dev->flags)) rcw++;
1326                                 else {
1327                                         PRINTK("raid6: must_compute: disk %d flags=%#lx\n", i, dev->flags);
1328                                         must_compute++;
1329                                 }
1330                         }
1331                 }
1332                 PRINTK("for sector %llu, rcw=%d, must_compute=%d\n",
1333                        (unsigned long long)sh->sector, rcw, must_compute);
1334                 set_bit(STRIPE_HANDLE, &sh->state);
1335
1336                 if (rcw > 0)
1337                         /* want reconstruct write, but need to get some data */
1338                         for (i=disks; i--;) {
1339                                 dev = &sh->dev[i];
1340                                 if (!test_bit(R5_OVERWRITE, &dev->flags)
1341                                     && !(failed == 0 && (i == pd_idx || i == qd_idx))
1342                                     && !test_bit(R5_LOCKED, &dev->flags) && !test_bit(R5_UPTODATE, &dev->flags) &&
1343                                     test_bit(R5_Insync, &dev->flags)) {
1344                                         if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1345                                         {
1346                                                 PRINTK("Read_old stripe %llu block %d for Reconstruct\n",
1347                                                        (unsigned long long)sh->sector, i);
1348                                                 set_bit(R5_LOCKED, &dev->flags);
1349                                                 set_bit(R5_Wantread, &dev->flags);
1350                                                 locked++;
1351                                         } else {
1352                                                 PRINTK("Request delayed stripe %llu block %d for Reconstruct\n",
1353                                                        (unsigned long long)sh->sector, i);
1354                                                 set_bit(STRIPE_DELAYED, &sh->state);
1355                                                 set_bit(STRIPE_HANDLE, &sh->state);
1356                                         }
1357                                 }
1358                         }
1359                 /* now if nothing is locked, and if we have enough data, we can start a write request */
1360                 if (locked == 0 && rcw == 0 &&
1361                     !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
1362                         if ( must_compute > 0 ) {
1363                                 /* We have failed blocks and need to compute them */
1364                                 switch ( failed ) {
1365                                 case 0: BUG();
1366                                 case 1: compute_block_1(sh, failed_num[0], 0); break;
1367                                 case 2: compute_block_2(sh, failed_num[0], failed_num[1]); break;
1368                                 default: BUG(); /* This request should have been failed? */
1369                                 }
1370                         }
1371
1372                         PRINTK("Computing parity for stripe %llu\n", (unsigned long long)sh->sector);
1373                         compute_parity(sh, RECONSTRUCT_WRITE);
1374                         /* now every locked buffer is ready to be written */
1375                         for (i=disks; i--;)
1376                                 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
1377                                         PRINTK("Writing stripe %llu block %d\n",
1378                                                (unsigned long long)sh->sector, i);
1379                                         locked++;
1380                                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
1381                                 }
1382                         /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
1383                         set_bit(STRIPE_INSYNC, &sh->state);
1384
1385                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
1386                                 atomic_dec(&conf->preread_active_stripes);
1387                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
1388                                         md_wakeup_thread(conf->mddev->thread);
1389                         }
1390                 }
1391         }
1392
1393         /* maybe we need to check and possibly fix the parity for this stripe
1394          * Any reads will already have been scheduled, so we just see if enough data
1395          * is available
1396          */
1397         if (syncing && locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state)) {
1398                 int update_p = 0, update_q = 0;
1399                 struct r5dev *dev;
1400
1401                 set_bit(STRIPE_HANDLE, &sh->state);
1402
1403                 BUG_ON(failed>2);
1404                 BUG_ON(uptodate < disks);
1405                 /* Want to check and possibly repair P and Q.
1406                  * However there could be one 'failed' device, in which
1407                  * case we can only check one of them, possibly using the
1408                  * other to generate missing data
1409                  */
1410
1411                 /* If !tmp_page, we cannot do the calculations,
1412                  * but as we have set STRIPE_HANDLE, we will soon be called
1413                  * by stripe_handle with a tmp_page - just wait until then.
1414                  */
1415                 if (tmp_page) {
1416                         if (failed == q_failed) {
1417                                 /* The only possible failed device holds 'Q', so it makes
1418                                  * sense to check P (If anything else were failed, we would
1419                                  * have used P to recreate it).
1420                                  */
1421                                 compute_block_1(sh, pd_idx, 1);
1422                                 if (!page_is_zero(sh->dev[pd_idx].page)) {
1423                                         compute_block_1(sh,pd_idx,0);
1424                                         update_p = 1;
1425                                 }
1426                         }
1427                         if (!q_failed && failed < 2) {
1428                                 /* q is not failed, and we didn't use it to generate
1429                                  * anything, so it makes sense to check it
1430                                  */
1431                                 memcpy(page_address(tmp_page),
1432                                        page_address(sh->dev[qd_idx].page),
1433                                        STRIPE_SIZE);
1434                                 compute_parity(sh, UPDATE_PARITY);
1435                                 if (memcmp(page_address(tmp_page),
1436                                            page_address(sh->dev[qd_idx].page),
1437                                            STRIPE_SIZE)!= 0) {
1438                                         clear_bit(STRIPE_INSYNC, &sh->state);
1439                                         update_q = 1;
1440                                 }
1441                         }
1442                         if (update_p || update_q) {
1443                                 conf->mddev->resync_mismatches += STRIPE_SECTORS;
1444                                 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
1445                                         /* don't try to repair!! */
1446                                         update_p = update_q = 0;
1447                         }
1448
1449                         /* now write out any block on a failed drive,
1450                          * or P or Q if they need it
1451                          */
1452
1453                         if (failed == 2) {
1454                                 dev = &sh->dev[failed_num[1]];
1455                                 locked++;
1456                                 set_bit(R5_LOCKED, &dev->flags);
1457                                 set_bit(R5_Wantwrite, &dev->flags);
1458                         }
1459                         if (failed >= 1) {
1460                                 dev = &sh->dev[failed_num[0]];
1461                                 locked++;
1462                                 set_bit(R5_LOCKED, &dev->flags);
1463                                 set_bit(R5_Wantwrite, &dev->flags);
1464                         }
1465
1466                         if (update_p) {
1467                                 dev = &sh->dev[pd_idx];
1468                                 locked ++;
1469                                 set_bit(R5_LOCKED, &dev->flags);
1470                                 set_bit(R5_Wantwrite, &dev->flags);
1471                         }
1472                         if (update_q) {
1473                                 dev = &sh->dev[qd_idx];
1474                                 locked++;
1475                                 set_bit(R5_LOCKED, &dev->flags);
1476                                 set_bit(R5_Wantwrite, &dev->flags);
1477                         }
1478                         clear_bit(STRIPE_DEGRADED, &sh->state);
1479
1480                         set_bit(STRIPE_INSYNC, &sh->state);
1481                 }
1482         }
1483
1484         if (syncing && locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
1485                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
1486                 clear_bit(STRIPE_SYNCING, &sh->state);
1487         }
1488
1489         /* If the failed drives are just a ReadError, then we might need
1490          * to progress the repair/check process
1491          */
1492         if (failed <= 2 && ! conf->mddev->ro)
1493                 for (i=0; i<failed;i++) {
1494                         dev = &sh->dev[failed_num[i]];
1495                         if (test_bit(R5_ReadError, &dev->flags)
1496                             && !test_bit(R5_LOCKED, &dev->flags)
1497                             && test_bit(R5_UPTODATE, &dev->flags)
1498                                 ) {
1499                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
1500                                         set_bit(R5_Wantwrite, &dev->flags);
1501                                         set_bit(R5_ReWrite, &dev->flags);
1502                                         set_bit(R5_LOCKED, &dev->flags);
1503                                 } else {
1504                                         /* let's read it back */
1505                                         set_bit(R5_Wantread, &dev->flags);
1506                                         set_bit(R5_LOCKED, &dev->flags);
1507                                 }
1508                         }
1509                 }
1510         spin_unlock(&sh->lock);
1511
1512         while ((bi=return_bi)) {
1513                 int bytes = bi->bi_size;
1514
1515                 return_bi = bi->bi_next;
1516                 bi->bi_next = NULL;
1517                 bi->bi_size = 0;
1518                 bi->bi_end_io(bi, bytes, 0);
1519         }
1520         for (i=disks; i-- ;) {
1521                 int rw;
1522                 struct bio *bi;
1523                 mdk_rdev_t *rdev;
1524                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
1525                         rw = 1;
1526                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1527                         rw = 0;
1528                 else
1529                         continue;
1530
1531                 bi = &sh->dev[i].req;
1532
1533                 bi->bi_rw = rw;
1534                 if (rw)
1535                         bi->bi_end_io = raid6_end_write_request;
1536                 else
1537                         bi->bi_end_io = raid6_end_read_request;
1538
1539                 rcu_read_lock();
1540                 rdev = rcu_dereference(conf->disks[i].rdev);
1541                 if (rdev && test_bit(Faulty, &rdev->flags))
1542                         rdev = NULL;
1543                 if (rdev)
1544                         atomic_inc(&rdev->nr_pending);
1545                 rcu_read_unlock();
1546
1547                 if (rdev) {
1548                         if (syncing)
1549                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1550
1551                         bi->bi_bdev = rdev->bdev;
1552                         PRINTK("for %llu schedule op %ld on disc %d\n",
1553                                 (unsigned long long)sh->sector, bi->bi_rw, i);
1554                         atomic_inc(&sh->count);
1555                         bi->bi_sector = sh->sector + rdev->data_offset;
1556                         bi->bi_flags = 1 << BIO_UPTODATE;
1557                         bi->bi_vcnt = 1;
1558                         bi->bi_max_vecs = 1;
1559                         bi->bi_idx = 0;
1560                         bi->bi_io_vec = &sh->dev[i].vec;
1561                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1562                         bi->bi_io_vec[0].bv_offset = 0;
1563                         bi->bi_size = STRIPE_SIZE;
1564                         bi->bi_next = NULL;
1565                         if (rw == WRITE &&
1566                             test_bit(R5_ReWrite, &sh->dev[i].flags))
1567                                 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1568                         generic_make_request(bi);
1569                 } else {
1570                         if (rw == 1)
1571                                 set_bit(STRIPE_DEGRADED, &sh->state);
1572                         PRINTK("skip op %ld on disc %d for sector %llu\n",
1573                                 bi->bi_rw, i, (unsigned long long)sh->sector);
1574                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1575                         set_bit(STRIPE_HANDLE, &sh->state);
1576                 }
1577         }
1578 }
1579
1580 static void raid6_activate_delayed(raid6_conf_t *conf)
1581 {
1582         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
1583                 while (!list_empty(&conf->delayed_list)) {
1584                         struct list_head *l = conf->delayed_list.next;
1585                         struct stripe_head *sh;
1586                         sh = list_entry(l, struct stripe_head, lru);
1587                         list_del_init(l);
1588                         clear_bit(STRIPE_DELAYED, &sh->state);
1589                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1590                                 atomic_inc(&conf->preread_active_stripes);
1591                         list_add_tail(&sh->lru, &conf->handle_list);
1592                 }
1593         }
1594 }
1595
1596 static void activate_bit_delay(raid6_conf_t *conf)
1597 {
1598         /* device_lock is held */
1599         struct list_head head;
1600         list_add(&head, &conf->bitmap_list);
1601         list_del_init(&conf->bitmap_list);
1602         while (!list_empty(&head)) {
1603                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
1604                 list_del_init(&sh->lru);
1605                 atomic_inc(&sh->count);
1606                 __release_stripe(conf, sh);
1607         }
1608 }
1609
1610 static void unplug_slaves(mddev_t *mddev)
1611 {
1612         raid6_conf_t *conf = mddev_to_conf(mddev);
1613         int i;
1614
1615         rcu_read_lock();
1616         for (i=0; i<mddev->raid_disks; i++) {
1617                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
1618                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
1619                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
1620
1621                         atomic_inc(&rdev->nr_pending);
1622                         rcu_read_unlock();
1623
1624                         if (r_queue->unplug_fn)
1625                                 r_queue->unplug_fn(r_queue);
1626
1627                         rdev_dec_pending(rdev, mddev);
1628                         rcu_read_lock();
1629                 }
1630         }
1631         rcu_read_unlock();
1632 }
1633
1634 static void raid6_unplug_device(request_queue_t *q)
1635 {
1636         mddev_t *mddev = q->queuedata;
1637         raid6_conf_t *conf = mddev_to_conf(mddev);
1638         unsigned long flags;
1639
1640         spin_lock_irqsave(&conf->device_lock, flags);
1641
1642         if (blk_remove_plug(q)) {
1643                 conf->seq_flush++;
1644                 raid6_activate_delayed(conf);
1645         }
1646         md_wakeup_thread(mddev->thread);
1647
1648         spin_unlock_irqrestore(&conf->device_lock, flags);
1649
1650         unplug_slaves(mddev);
1651 }
1652
1653 static int raid6_issue_flush(request_queue_t *q, struct gendisk *disk,
1654                              sector_t *error_sector)
1655 {
1656         mddev_t *mddev = q->queuedata;
1657         raid6_conf_t *conf = mddev_to_conf(mddev);
1658         int i, ret = 0;
1659
1660         rcu_read_lock();
1661         for (i=0; i<mddev->raid_disks && ret == 0; i++) {
1662                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
1663                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
1664                         struct block_device *bdev = rdev->bdev;
1665                         request_queue_t *r_queue = bdev_get_queue(bdev);
1666
1667                         if (!r_queue->issue_flush_fn)
1668                                 ret = -EOPNOTSUPP;
1669                         else {
1670                                 atomic_inc(&rdev->nr_pending);
1671                                 rcu_read_unlock();
1672                                 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
1673                                                               error_sector);
1674                                 rdev_dec_pending(rdev, mddev);
1675                                 rcu_read_lock();
1676                         }
1677                 }
1678         }
1679         rcu_read_unlock();
1680         return ret;
1681 }
1682
1683 static inline void raid6_plug_device(raid6_conf_t *conf)
1684 {
1685         spin_lock_irq(&conf->device_lock);
1686         blk_plug_device(conf->mddev->queue);
1687         spin_unlock_irq(&conf->device_lock);
1688 }
1689
1690 static int make_request (request_queue_t *q, struct bio * bi)
1691 {
1692         mddev_t *mddev = q->queuedata;
1693         raid6_conf_t *conf = mddev_to_conf(mddev);
1694         const unsigned int raid_disks = conf->raid_disks;
1695         const unsigned int data_disks = raid_disks - 2;
1696         unsigned int dd_idx, pd_idx;
1697         sector_t new_sector;
1698         sector_t logical_sector, last_sector;
1699         struct stripe_head *sh;
1700         const int rw = bio_data_dir(bi);
1701
1702         if (unlikely(bio_barrier(bi))) {
1703                 bio_endio(bi, bi->bi_size, -EOPNOTSUPP);
1704                 return 0;
1705         }
1706
1707         md_write_start(mddev, bi);
1708
1709         disk_stat_inc(mddev->gendisk, ios[rw]);
1710         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
1711
1712         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
1713         last_sector = bi->bi_sector + (bi->bi_size>>9);
1714
1715         bi->bi_next = NULL;
1716         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
1717
1718         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
1719                 DEFINE_WAIT(w);
1720
1721                 new_sector = raid6_compute_sector(logical_sector,
1722                                                   raid_disks, data_disks, &dd_idx, &pd_idx, conf);
1723
1724                 PRINTK("raid6: make_request, sector %llu logical %llu\n",
1725                        (unsigned long long)new_sector,
1726                        (unsigned long long)logical_sector);
1727
1728         retry:
1729                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
1730                 sh = get_active_stripe(conf, new_sector, pd_idx, (bi->bi_rw&RWA_MASK));
1731                 if (sh) {
1732                         if (!add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
1733                                 /* Add failed due to overlap.  Flush everything
1734                                  * and wait a while
1735                                  */
1736                                 raid6_unplug_device(mddev->queue);
1737                                 release_stripe(sh);
1738                                 schedule();
1739                                 goto retry;
1740                         }
1741                         finish_wait(&conf->wait_for_overlap, &w);
1742                         raid6_plug_device(conf);
1743                         handle_stripe(sh, NULL);
1744                         release_stripe(sh);
1745                 } else {
1746                         /* cannot get stripe for read-ahead, just give-up */
1747                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1748                         finish_wait(&conf->wait_for_overlap, &w);
1749                         break;
1750                 }
1751
1752         }
1753         spin_lock_irq(&conf->device_lock);
1754         if (--bi->bi_phys_segments == 0) {
1755                 int bytes = bi->bi_size;
1756
1757                 if (rw == WRITE )
1758                         md_write_end(mddev);
1759                 bi->bi_size = 0;
1760                 bi->bi_end_io(bi, bytes, 0);
1761         }
1762         spin_unlock_irq(&conf->device_lock);
1763         return 0;
1764 }
1765
1766 /* FIXME go_faster isn't used */
1767 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1768 {
1769         raid6_conf_t *conf = (raid6_conf_t *) mddev->private;
1770         struct stripe_head *sh;
1771         int sectors_per_chunk = conf->chunk_size >> 9;
1772         sector_t x;
1773         unsigned long stripe;
1774         int chunk_offset;
1775         int dd_idx, pd_idx;
1776         sector_t first_sector;
1777         int raid_disks = conf->raid_disks;
1778         int data_disks = raid_disks - 2;
1779         sector_t max_sector = mddev->size << 1;
1780         int sync_blocks;
1781         int still_degraded = 0;
1782         int i;
1783
1784         if (sector_nr >= max_sector) {
1785                 /* just being told to finish up .. nothing much to do */
1786                 unplug_slaves(mddev);
1787
1788                 if (mddev->curr_resync < max_sector) /* aborted */
1789                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1790                                         &sync_blocks, 1);
1791                 else /* completed sync */
1792                         conf->fullsync = 0;
1793                 bitmap_close_sync(mddev->bitmap);
1794
1795                 return 0;
1796         }
1797         /* if there are 2 or more failed drives and we are trying
1798          * to resync, then assert that we are finished, because there is
1799          * nothing we can do.
1800          */
1801         if (mddev->degraded >= 2 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1802                 sector_t rv = (mddev->size << 1) - sector_nr;
1803                 *skipped = 1;
1804                 return rv;
1805         }
1806         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1807             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1808             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
1809                 /* we can skip this block, and probably more */
1810                 sync_blocks /= STRIPE_SECTORS;
1811                 *skipped = 1;
1812                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
1813         }
1814
1815         x = sector_nr;
1816         chunk_offset = sector_div(x, sectors_per_chunk);
1817         stripe = x;
1818         BUG_ON(x != stripe);
1819
1820         first_sector = raid6_compute_sector((sector_t)stripe*data_disks*sectors_per_chunk
1821                 + chunk_offset, raid_disks, data_disks, &dd_idx, &pd_idx, conf);
1822         sh = get_active_stripe(conf, sector_nr, pd_idx, 1);
1823         if (sh == NULL) {
1824                 sh = get_active_stripe(conf, sector_nr, pd_idx, 0);
1825                 /* make sure we don't swamp the stripe cache if someone else
1826                  * is trying to get access
1827                  */
1828                 schedule_timeout_uninterruptible(1);
1829         }
1830         /* Need to check if array will still be degraded after recovery/resync
1831          * We don't need to check the 'failed' flag as when that gets set,
1832          * recovery aborts.
1833          */
1834         for (i=0; i<mddev->raid_disks; i++)
1835                 if (conf->disks[i].rdev == NULL)
1836                         still_degraded = 1;
1837
1838         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
1839
1840         spin_lock(&sh->lock);
1841         set_bit(STRIPE_SYNCING, &sh->state);
1842         clear_bit(STRIPE_INSYNC, &sh->state);
1843         spin_unlock(&sh->lock);
1844
1845         handle_stripe(sh, NULL);
1846         release_stripe(sh);
1847
1848         return STRIPE_SECTORS;
1849 }
1850
1851 /*
1852  * This is our raid6 kernel thread.
1853  *
1854  * We scan the hash table for stripes which can be handled now.
1855  * During the scan, completed stripes are saved for us by the interrupt
1856  * handler, so that they will not have to wait for our next wakeup.
1857  */
1858 static void raid6d (mddev_t *mddev)
1859 {
1860         struct stripe_head *sh;
1861         raid6_conf_t *conf = mddev_to_conf(mddev);
1862         int handled;
1863
1864         PRINTK("+++ raid6d active\n");
1865
1866         md_check_recovery(mddev);
1867
1868         handled = 0;
1869         spin_lock_irq(&conf->device_lock);
1870         while (1) {
1871                 struct list_head *first;
1872
1873                 if (conf->seq_flush - conf->seq_write > 0) {
1874                         int seq = conf->seq_flush;
1875                         spin_unlock_irq(&conf->device_lock);
1876                         bitmap_unplug(mddev->bitmap);
1877                         spin_lock_irq(&conf->device_lock);
1878                         conf->seq_write = seq;
1879                         activate_bit_delay(conf);
1880                 }
1881
1882                 if (list_empty(&conf->handle_list) &&
1883                     atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD &&
1884                     !blk_queue_plugged(mddev->queue) &&
1885                     !list_empty(&conf->delayed_list))
1886                         raid6_activate_delayed(conf);
1887
1888                 if (list_empty(&conf->handle_list))
1889                         break;
1890
1891                 first = conf->handle_list.next;
1892                 sh = list_entry(first, struct stripe_head, lru);
1893
1894                 list_del_init(first);
1895                 atomic_inc(&sh->count);
1896                 if (atomic_read(&sh->count)!= 1)
1897                         BUG();
1898                 spin_unlock_irq(&conf->device_lock);
1899
1900                 handled++;
1901                 handle_stripe(sh, conf->spare_page);
1902                 release_stripe(sh);
1903
1904                 spin_lock_irq(&conf->device_lock);
1905         }
1906         PRINTK("%d stripes handled\n", handled);
1907
1908         spin_unlock_irq(&conf->device_lock);
1909
1910         unplug_slaves(mddev);
1911
1912         PRINTK("--- raid6d inactive\n");
1913 }
1914
1915 static int run(mddev_t *mddev)
1916 {
1917         raid6_conf_t *conf;
1918         int raid_disk, memory;
1919         mdk_rdev_t *rdev;
1920         struct disk_info *disk;
1921         struct list_head *tmp;
1922
1923         if (mddev->level != 6) {
1924                 PRINTK("raid6: %s: raid level not set to 6 (%d)\n", mdname(mddev), mddev->level);
1925                 return -EIO;
1926         }
1927
1928         mddev->private = kzalloc(sizeof (raid6_conf_t)
1929                                  + mddev->raid_disks * sizeof(struct disk_info),
1930                                  GFP_KERNEL);
1931         if ((conf = mddev->private) == NULL)
1932                 goto abort;
1933         conf->mddev = mddev;
1934
1935         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
1936                 goto abort;
1937
1938         conf->spare_page = alloc_page(GFP_KERNEL);
1939         if (!conf->spare_page)
1940                 goto abort;
1941
1942         spin_lock_init(&conf->device_lock);
1943         init_waitqueue_head(&conf->wait_for_stripe);
1944         init_waitqueue_head(&conf->wait_for_overlap);
1945         INIT_LIST_HEAD(&conf->handle_list);
1946         INIT_LIST_HEAD(&conf->delayed_list);
1947         INIT_LIST_HEAD(&conf->bitmap_list);
1948         INIT_LIST_HEAD(&conf->inactive_list);
1949         atomic_set(&conf->active_stripes, 0);
1950         atomic_set(&conf->preread_active_stripes, 0);
1951
1952         PRINTK("raid6: run(%s) called.\n", mdname(mddev));
1953
1954         ITERATE_RDEV(mddev,rdev,tmp) {
1955                 raid_disk = rdev->raid_disk;
1956                 if (raid_disk >= mddev->raid_disks
1957                     || raid_disk < 0)
1958                         continue;
1959                 disk = conf->disks + raid_disk;
1960
1961                 disk->rdev = rdev;
1962
1963                 if (test_bit(In_sync, &rdev->flags)) {
1964                         char b[BDEVNAME_SIZE];
1965                         printk(KERN_INFO "raid6: device %s operational as raid"
1966                                " disk %d\n", bdevname(rdev->bdev,b),
1967                                raid_disk);
1968                         conf->working_disks++;
1969                 }
1970         }
1971
1972         conf->raid_disks = mddev->raid_disks;
1973
1974         /*
1975          * 0 for a fully functional array, 1 or 2 for a degraded array.
1976          */
1977         mddev->degraded = conf->failed_disks = conf->raid_disks - conf->working_disks;
1978         conf->mddev = mddev;
1979         conf->chunk_size = mddev->chunk_size;
1980         conf->level = mddev->level;
1981         conf->algorithm = mddev->layout;
1982         conf->max_nr_stripes = NR_STRIPES;
1983
1984         /* device size must be a multiple of chunk size */
1985         mddev->size &= ~(mddev->chunk_size/1024 -1);
1986         mddev->resync_max_sectors = mddev->size << 1;
1987
1988         if (conf->raid_disks < 4) {
1989                 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
1990                        mdname(mddev), conf->raid_disks);
1991                 goto abort;
1992         }
1993         if (!conf->chunk_size || conf->chunk_size % 4) {
1994                 printk(KERN_ERR "raid6: invalid chunk size %d for %s\n",
1995                        conf->chunk_size, mdname(mddev));
1996                 goto abort;
1997         }
1998         if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
1999                 printk(KERN_ERR
2000                        "raid6: unsupported parity algorithm %d for %s\n",
2001                        conf->algorithm, mdname(mddev));
2002                 goto abort;
2003         }
2004         if (mddev->degraded > 2) {
2005                 printk(KERN_ERR "raid6: not enough operational devices for %s"
2006                        " (%d/%d failed)\n",
2007                        mdname(mddev), conf->failed_disks, conf->raid_disks);
2008                 goto abort;
2009         }
2010
2011         if (mddev->degraded > 0 &&
2012             mddev->recovery_cp != MaxSector) {
2013                 if (mddev->ok_start_degraded)
2014                         printk(KERN_WARNING "raid6: starting dirty degraded array:%s"
2015                                "- data corruption possible.\n",
2016                                mdname(mddev));
2017                 else {
2018                         printk(KERN_ERR "raid6: cannot start dirty degraded array"
2019                                " for %s\n", mdname(mddev));
2020                         goto abort;
2021                 }
2022         }
2023
2024         {
2025                 mddev->thread = md_register_thread(raid6d, mddev, "%s_raid6");
2026                 if (!mddev->thread) {
2027                         printk(KERN_ERR
2028                                "raid6: couldn't allocate thread for %s\n",
2029                                mdname(mddev));
2030                         goto abort;
2031                 }
2032         }
2033
2034         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
2035                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
2036         if (grow_stripes(conf, conf->max_nr_stripes)) {
2037                 printk(KERN_ERR
2038                        "raid6: couldn't allocate %dkB for buffers\n", memory);
2039                 shrink_stripes(conf);
2040                 md_unregister_thread(mddev->thread);
2041                 goto abort;
2042         } else
2043                 printk(KERN_INFO "raid6: allocated %dkB for %s\n",
2044                        memory, mdname(mddev));
2045
2046         if (mddev->degraded == 0)
2047                 printk(KERN_INFO "raid6: raid level %d set %s active with %d out of %d"
2048                        " devices, algorithm %d\n", conf->level, mdname(mddev),
2049                        mddev->raid_disks-mddev->degraded, mddev->raid_disks,
2050                        conf->algorithm);
2051         else
2052                 printk(KERN_ALERT "raid6: raid level %d set %s active with %d"
2053                        " out of %d devices, algorithm %d\n", conf->level,
2054                        mdname(mddev), mddev->raid_disks - mddev->degraded,
2055                        mddev->raid_disks, conf->algorithm);
2056
2057         print_raid6_conf(conf);
2058
2059         /* read-ahead size must cover two whole stripes, which is
2060          * 2 * (n-2) * chunksize where 'n' is the number of raid devices
2061          */
2062         {
2063                 int stripe = (mddev->raid_disks-2) * mddev->chunk_size
2064                         / PAGE_SIZE;
2065                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
2066                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
2067         }
2068
2069         /* Ok, everything is just fine now */
2070         mddev->array_size =  mddev->size * (mddev->raid_disks - 2);
2071
2072         mddev->queue->unplug_fn = raid6_unplug_device;
2073         mddev->queue->issue_flush_fn = raid6_issue_flush;
2074         return 0;
2075 abort:
2076         if (conf) {
2077                 print_raid6_conf(conf);
2078                 safe_put_page(conf->spare_page);
2079                 kfree(conf->stripe_hashtbl);
2080                 kfree(conf);
2081         }
2082         mddev->private = NULL;
2083         printk(KERN_ALERT "raid6: failed to run raid set %s\n", mdname(mddev));
2084         return -EIO;
2085 }
2086
2087
2088
2089 static int stop (mddev_t *mddev)
2090 {
2091         raid6_conf_t *conf = (raid6_conf_t *) mddev->private;
2092
2093         md_unregister_thread(mddev->thread);
2094         mddev->thread = NULL;
2095         shrink_stripes(conf);
2096         kfree(conf->stripe_hashtbl);
2097         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2098         kfree(conf);
2099         mddev->private = NULL;
2100         return 0;
2101 }
2102
2103 #if RAID6_DUMPSTATE
2104 static void print_sh (struct seq_file *seq, struct stripe_head *sh)
2105 {
2106         int i;
2107
2108         seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
2109                    (unsigned long long)sh->sector, sh->pd_idx, sh->state);
2110         seq_printf(seq, "sh %llu,  count %d.\n",
2111                    (unsigned long long)sh->sector, atomic_read(&sh->count));
2112         seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
2113         for (i = 0; i < sh->raid_conf->raid_disks; i++) {
2114                 seq_printf(seq, "(cache%d: %p %ld) ",
2115                            i, sh->dev[i].page, sh->dev[i].flags);
2116         }
2117         seq_printf(seq, "\n");
2118 }
2119
2120 static void printall (struct seq_file *seq, raid6_conf_t *conf)
2121 {
2122         struct stripe_head *sh;
2123         struct hlist_node *hn;
2124         int i;
2125
2126         spin_lock_irq(&conf->device_lock);
2127         for (i = 0; i < NR_HASH; i++) {
2128                 sh = conf->stripe_hashtbl[i];
2129                 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
2130                         if (sh->raid_conf != conf)
2131                                 continue;
2132                         print_sh(seq, sh);
2133                 }
2134         }
2135         spin_unlock_irq(&conf->device_lock);
2136 }
2137 #endif
2138
2139 static void status (struct seq_file *seq, mddev_t *mddev)
2140 {
2141         raid6_conf_t *conf = (raid6_conf_t *) mddev->private;
2142         int i;
2143
2144         seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
2145         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->working_disks);
2146         for (i = 0; i < conf->raid_disks; i++)
2147                 seq_printf (seq, "%s",
2148                             conf->disks[i].rdev &&
2149                             test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
2150         seq_printf (seq, "]");
2151 #if RAID6_DUMPSTATE
2152         seq_printf (seq, "\n");
2153         printall(seq, conf);
2154 #endif
2155 }
2156
2157 static void print_raid6_conf (raid6_conf_t *conf)
2158 {
2159         int i;
2160         struct disk_info *tmp;
2161
2162         printk("RAID6 conf printout:\n");
2163         if (!conf) {
2164                 printk("(conf==NULL)\n");
2165                 return;
2166         }
2167         printk(" --- rd:%d wd:%d fd:%d\n", conf->raid_disks,
2168                  conf->working_disks, conf->failed_disks);
2169
2170         for (i = 0; i < conf->raid_disks; i++) {
2171                 char b[BDEVNAME_SIZE];
2172                 tmp = conf->disks + i;
2173                 if (tmp->rdev)
2174                 printk(" disk %d, o:%d, dev:%s\n",
2175                         i, !test_bit(Faulty, &tmp->rdev->flags),
2176                         bdevname(tmp->rdev->bdev,b));
2177         }
2178 }
2179
2180 static int raid6_spare_active(mddev_t *mddev)
2181 {
2182         int i;
2183         raid6_conf_t *conf = mddev->private;
2184         struct disk_info *tmp;
2185
2186         for (i = 0; i < conf->raid_disks; i++) {
2187                 tmp = conf->disks + i;
2188                 if (tmp->rdev
2189                     && !test_bit(Faulty, &tmp->rdev->flags)
2190                     && !test_bit(In_sync, &tmp->rdev->flags)) {
2191                         mddev->degraded--;
2192                         conf->failed_disks--;
2193                         conf->working_disks++;
2194                         set_bit(In_sync, &tmp->rdev->flags);
2195                 }
2196         }
2197         print_raid6_conf(conf);
2198         return 0;
2199 }
2200
2201 static int raid6_remove_disk(mddev_t *mddev, int number)
2202 {
2203         raid6_conf_t *conf = mddev->private;
2204         int err = 0;
2205         mdk_rdev_t *rdev;
2206         struct disk_info *p = conf->disks + number;
2207
2208         print_raid6_conf(conf);
2209         rdev = p->rdev;
2210         if (rdev) {
2211                 if (test_bit(In_sync, &rdev->flags) ||
2212                     atomic_read(&rdev->nr_pending)) {
2213                         err = -EBUSY;
2214                         goto abort;
2215                 }
2216                 p->rdev = NULL;
2217                 synchronize_rcu();
2218                 if (atomic_read(&rdev->nr_pending)) {
2219                         /* lost the race, try later */
2220                         err = -EBUSY;
2221                         p->rdev = rdev;
2222                 }
2223         }
2224
2225 abort:
2226
2227         print_raid6_conf(conf);
2228         return err;
2229 }
2230
2231 static int raid6_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
2232 {
2233         raid6_conf_t *conf = mddev->private;
2234         int found = 0;
2235         int disk;
2236         struct disk_info *p;
2237
2238         if (mddev->degraded > 2)
2239                 /* no point adding a device */
2240                 return 0;
2241         /*
2242          * find the disk ... but prefer rdev->saved_raid_disk
2243          * if possible.
2244          */
2245         if (rdev->saved_raid_disk >= 0 &&
2246             conf->disks[rdev->saved_raid_disk].rdev == NULL)
2247                 disk = rdev->saved_raid_disk;
2248         else
2249                 disk = 0;
2250         for ( ; disk < mddev->raid_disks; disk++)
2251                 if ((p=conf->disks + disk)->rdev == NULL) {
2252                         clear_bit(In_sync, &rdev->flags);
2253                         rdev->raid_disk = disk;
2254                         found = 1;
2255                         if (rdev->saved_raid_disk != disk)
2256                                 conf->fullsync = 1;
2257                         rcu_assign_pointer(p->rdev, rdev);
2258                         break;
2259                 }
2260         print_raid6_conf(conf);
2261         return found;
2262 }
2263
2264 static int raid6_resize(mddev_t *mddev, sector_t sectors)
2265 {
2266         /* no resync is happening, and there is enough space
2267          * on all devices, so we can resize.
2268          * We need to make sure resync covers any new space.
2269          * If the array is shrinking we should possibly wait until
2270          * any io in the removed space completes, but it hardly seems
2271          * worth it.
2272          */
2273         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
2274         mddev->array_size = (sectors * (mddev->raid_disks-2))>>1;
2275         set_capacity(mddev->gendisk, mddev->array_size << 1);
2276         mddev->changed = 1;
2277         if (sectors/2  > mddev->size && mddev->recovery_cp == MaxSector) {
2278                 mddev->recovery_cp = mddev->size << 1;
2279                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2280         }
2281         mddev->size = sectors /2;
2282         mddev->resync_max_sectors = sectors;
2283         return 0;
2284 }
2285
2286 static void raid6_quiesce(mddev_t *mddev, int state)
2287 {
2288         raid6_conf_t *conf = mddev_to_conf(mddev);
2289
2290         switch(state) {
2291         case 1: /* stop all writes */
2292                 spin_lock_irq(&conf->device_lock);
2293                 conf->quiesce = 1;
2294                 wait_event_lock_irq(conf->wait_for_stripe,
2295                                     atomic_read(&conf->active_stripes) == 0,
2296                                     conf->device_lock, /* nothing */);
2297                 spin_unlock_irq(&conf->device_lock);
2298                 break;
2299
2300         case 0: /* re-enable writes */
2301                 spin_lock_irq(&conf->device_lock);
2302                 conf->quiesce = 0;
2303                 wake_up(&conf->wait_for_stripe);
2304                 spin_unlock_irq(&conf->device_lock);
2305                 break;
2306         }
2307 }
2308
2309 static struct mdk_personality raid6_personality =
2310 {
2311         .name           = "raid6",
2312         .level          = 6,
2313         .owner          = THIS_MODULE,
2314         .make_request   = make_request,
2315         .run            = run,
2316         .stop           = stop,
2317         .status         = status,
2318         .error_handler  = error,
2319         .hot_add_disk   = raid6_add_disk,
2320         .hot_remove_disk= raid6_remove_disk,
2321         .spare_active   = raid6_spare_active,
2322         .sync_request   = sync_request,
2323         .resize         = raid6_resize,
2324         .quiesce        = raid6_quiesce,
2325 };
2326
2327 static int __init raid6_init(void)
2328 {
2329         int e;
2330
2331         e = raid6_select_algo();
2332         if ( e )
2333                 return e;
2334
2335         return register_md_personality(&raid6_personality);
2336 }
2337
2338 static void raid6_exit (void)
2339 {
2340         unregister_md_personality(&raid6_personality);
2341 }
2342
2343 module_init(raid6_init);
2344 module_exit(raid6_exit);
2345 MODULE_LICENSE("GPL");
2346 MODULE_ALIAS("md-personality-8"); /* RAID6 */
2347 MODULE_ALIAS("md-raid6");
2348 MODULE_ALIAS("md-level-6");