Merge branch 'for-linus' of git://neil.brown.name/md
[pandora-kernel.git] / drivers / md / raid5.c
1 /*
2  * raid5.c : Multiple Devices driver for Linux
3  *         Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4  *         Copyright (C) 1999, 2000 Ingo Molnar
5  *         Copyright (C) 2002, 2003 H. Peter Anvin
6  *
7  * RAID-4/5/6 management functions.
8  * Thanks to Penguin Computing for making the RAID-6 development possible
9  * by donating a test server!
10  *
11  * This program is free software; you can redistribute it and/or modify
12  * it under the terms of the GNU General Public License as published by
13  * the Free Software Foundation; either version 2, or (at your option)
14  * any later version.
15  *
16  * You should have received a copy of the GNU General Public License
17  * (for example /usr/src/linux/COPYING); if not, write to the Free
18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19  */
20
21 /*
22  * BITMAP UNPLUGGING:
23  *
24  * The sequencing for updating the bitmap reliably is a little
25  * subtle (and I got it wrong the first time) so it deserves some
26  * explanation.
27  *
28  * We group bitmap updates into batches.  Each batch has a number.
29  * We may write out several batches at once, but that isn't very important.
30  * conf->seq_write is the number of the last batch successfully written.
31  * conf->seq_flush is the number of the last batch that was closed to
32  *    new additions.
33  * When we discover that we will need to write to any block in a stripe
34  * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35  * the number of the batch it will be in. This is seq_flush+1.
36  * When we are ready to do a write, if that batch hasn't been written yet,
37  *   we plug the array and queue the stripe for later.
38  * When an unplug happens, we increment bm_flush, thus closing the current
39  *   batch.
40  * When we notice that bm_flush > bm_write, we write out all pending updates
41  * to the bitmap, and advance bm_write to where bm_flush was.
42  * This may occasionally write a bit out twice, but is sure never to
43  * miss any bits.
44  */
45
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/module.h>
51 #include <linux/async.h>
52 #include <linux/seq_file.h>
53 #include <linux/cpu.h>
54 #include <linux/slab.h>
55 #include <linux/ratelimit.h>
56 #include "md.h"
57 #include "raid5.h"
58 #include "raid0.h"
59 #include "bitmap.h"
60
61 /*
62  * Stripe cache
63  */
64
65 #define NR_STRIPES              256
66 #define STRIPE_SIZE             PAGE_SIZE
67 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
68 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
69 #define IO_THRESHOLD            1
70 #define BYPASS_THRESHOLD        1
71 #define NR_HASH                 (PAGE_SIZE / sizeof(struct hlist_head))
72 #define HASH_MASK               (NR_HASH - 1)
73
74 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
75 {
76         int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
77         return &conf->stripe_hashtbl[hash];
78 }
79
80 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
81  * order without overlap.  There may be several bio's per stripe+device, and
82  * a bio could span several devices.
83  * When walking this list for a particular stripe+device, we must never proceed
84  * beyond a bio that extends past this device, as the next bio might no longer
85  * be valid.
86  * This function is used to determine the 'next' bio in the list, given the sector
87  * of the current stripe+device
88  */
89 static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
90 {
91         int sectors = bio->bi_size >> 9;
92         if (bio->bi_sector + sectors < sector + STRIPE_SECTORS)
93                 return bio->bi_next;
94         else
95                 return NULL;
96 }
97
98 /*
99  * We maintain a biased count of active stripes in the bottom 16 bits of
100  * bi_phys_segments, and a count of processed stripes in the upper 16 bits
101  */
102 static inline int raid5_bi_phys_segments(struct bio *bio)
103 {
104         return bio->bi_phys_segments & 0xffff;
105 }
106
107 static inline int raid5_bi_hw_segments(struct bio *bio)
108 {
109         return (bio->bi_phys_segments >> 16) & 0xffff;
110 }
111
112 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
113 {
114         --bio->bi_phys_segments;
115         return raid5_bi_phys_segments(bio);
116 }
117
118 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
119 {
120         unsigned short val = raid5_bi_hw_segments(bio);
121
122         --val;
123         bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
124         return val;
125 }
126
127 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
128 {
129         bio->bi_phys_segments = raid5_bi_phys_segments(bio) | (cnt << 16);
130 }
131
132 /* Find first data disk in a raid6 stripe */
133 static inline int raid6_d0(struct stripe_head *sh)
134 {
135         if (sh->ddf_layout)
136                 /* ddf always start from first device */
137                 return 0;
138         /* md starts just after Q block */
139         if (sh->qd_idx == sh->disks - 1)
140                 return 0;
141         else
142                 return sh->qd_idx + 1;
143 }
144 static inline int raid6_next_disk(int disk, int raid_disks)
145 {
146         disk++;
147         return (disk < raid_disks) ? disk : 0;
148 }
149
150 /* When walking through the disks in a raid5, starting at raid6_d0,
151  * We need to map each disk to a 'slot', where the data disks are slot
152  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
153  * is raid_disks-1.  This help does that mapping.
154  */
155 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
156                              int *count, int syndrome_disks)
157 {
158         int slot = *count;
159
160         if (sh->ddf_layout)
161                 (*count)++;
162         if (idx == sh->pd_idx)
163                 return syndrome_disks;
164         if (idx == sh->qd_idx)
165                 return syndrome_disks + 1;
166         if (!sh->ddf_layout)
167                 (*count)++;
168         return slot;
169 }
170
171 static void return_io(struct bio *return_bi)
172 {
173         struct bio *bi = return_bi;
174         while (bi) {
175
176                 return_bi = bi->bi_next;
177                 bi->bi_next = NULL;
178                 bi->bi_size = 0;
179                 bio_endio(bi, 0);
180                 bi = return_bi;
181         }
182 }
183
184 static void print_raid5_conf (struct r5conf *conf);
185
186 static int stripe_operations_active(struct stripe_head *sh)
187 {
188         return sh->check_state || sh->reconstruct_state ||
189                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
190                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
191 }
192
193 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh)
194 {
195         if (atomic_dec_and_test(&sh->count)) {
196                 BUG_ON(!list_empty(&sh->lru));
197                 BUG_ON(atomic_read(&conf->active_stripes)==0);
198                 if (test_bit(STRIPE_HANDLE, &sh->state)) {
199                         if (test_bit(STRIPE_DELAYED, &sh->state))
200                                 list_add_tail(&sh->lru, &conf->delayed_list);
201                         else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
202                                    sh->bm_seq - conf->seq_write > 0)
203                                 list_add_tail(&sh->lru, &conf->bitmap_list);
204                         else {
205                                 clear_bit(STRIPE_BIT_DELAY, &sh->state);
206                                 list_add_tail(&sh->lru, &conf->handle_list);
207                         }
208                         md_wakeup_thread(conf->mddev->thread);
209                 } else {
210                         BUG_ON(stripe_operations_active(sh));
211                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
212                                 atomic_dec(&conf->preread_active_stripes);
213                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
214                                         md_wakeup_thread(conf->mddev->thread);
215                         }
216                         atomic_dec(&conf->active_stripes);
217                         if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
218                                 list_add_tail(&sh->lru, &conf->inactive_list);
219                                 wake_up(&conf->wait_for_stripe);
220                                 if (conf->retry_read_aligned)
221                                         md_wakeup_thread(conf->mddev->thread);
222                         }
223                 }
224         }
225 }
226
227 static void release_stripe(struct stripe_head *sh)
228 {
229         struct r5conf *conf = sh->raid_conf;
230         unsigned long flags;
231
232         spin_lock_irqsave(&conf->device_lock, flags);
233         __release_stripe(conf, sh);
234         spin_unlock_irqrestore(&conf->device_lock, flags);
235 }
236
237 static inline void remove_hash(struct stripe_head *sh)
238 {
239         pr_debug("remove_hash(), stripe %llu\n",
240                 (unsigned long long)sh->sector);
241
242         hlist_del_init(&sh->hash);
243 }
244
245 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
246 {
247         struct hlist_head *hp = stripe_hash(conf, sh->sector);
248
249         pr_debug("insert_hash(), stripe %llu\n",
250                 (unsigned long long)sh->sector);
251
252         hlist_add_head(&sh->hash, hp);
253 }
254
255
256 /* find an idle stripe, make sure it is unhashed, and return it. */
257 static struct stripe_head *get_free_stripe(struct r5conf *conf)
258 {
259         struct stripe_head *sh = NULL;
260         struct list_head *first;
261
262         if (list_empty(&conf->inactive_list))
263                 goto out;
264         first = conf->inactive_list.next;
265         sh = list_entry(first, struct stripe_head, lru);
266         list_del_init(first);
267         remove_hash(sh);
268         atomic_inc(&conf->active_stripes);
269 out:
270         return sh;
271 }
272
273 static void shrink_buffers(struct stripe_head *sh)
274 {
275         struct page *p;
276         int i;
277         int num = sh->raid_conf->pool_size;
278
279         for (i = 0; i < num ; i++) {
280                 p = sh->dev[i].page;
281                 if (!p)
282                         continue;
283                 sh->dev[i].page = NULL;
284                 put_page(p);
285         }
286 }
287
288 static int grow_buffers(struct stripe_head *sh)
289 {
290         int i;
291         int num = sh->raid_conf->pool_size;
292
293         for (i = 0; i < num; i++) {
294                 struct page *page;
295
296                 if (!(page = alloc_page(GFP_KERNEL))) {
297                         return 1;
298                 }
299                 sh->dev[i].page = page;
300         }
301         return 0;
302 }
303
304 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
305 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
306                             struct stripe_head *sh);
307
308 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
309 {
310         struct r5conf *conf = sh->raid_conf;
311         int i;
312
313         BUG_ON(atomic_read(&sh->count) != 0);
314         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
315         BUG_ON(stripe_operations_active(sh));
316
317         pr_debug("init_stripe called, stripe %llu\n",
318                 (unsigned long long)sh->sector);
319
320         remove_hash(sh);
321
322         sh->generation = conf->generation - previous;
323         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
324         sh->sector = sector;
325         stripe_set_idx(sector, conf, previous, sh);
326         sh->state = 0;
327
328
329         for (i = sh->disks; i--; ) {
330                 struct r5dev *dev = &sh->dev[i];
331
332                 if (dev->toread || dev->read || dev->towrite || dev->written ||
333                     test_bit(R5_LOCKED, &dev->flags)) {
334                         printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
335                                (unsigned long long)sh->sector, i, dev->toread,
336                                dev->read, dev->towrite, dev->written,
337                                test_bit(R5_LOCKED, &dev->flags));
338                         WARN_ON(1);
339                 }
340                 dev->flags = 0;
341                 raid5_build_block(sh, i, previous);
342         }
343         insert_hash(conf, sh);
344 }
345
346 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
347                                          short generation)
348 {
349         struct stripe_head *sh;
350         struct hlist_node *hn;
351
352         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
353         hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
354                 if (sh->sector == sector && sh->generation == generation)
355                         return sh;
356         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
357         return NULL;
358 }
359
360 /*
361  * Need to check if array has failed when deciding whether to:
362  *  - start an array
363  *  - remove non-faulty devices
364  *  - add a spare
365  *  - allow a reshape
366  * This determination is simple when no reshape is happening.
367  * However if there is a reshape, we need to carefully check
368  * both the before and after sections.
369  * This is because some failed devices may only affect one
370  * of the two sections, and some non-in_sync devices may
371  * be insync in the section most affected by failed devices.
372  */
373 static int has_failed(struct r5conf *conf)
374 {
375         int degraded;
376         int i;
377         if (conf->mddev->reshape_position == MaxSector)
378                 return conf->mddev->degraded > conf->max_degraded;
379
380         rcu_read_lock();
381         degraded = 0;
382         for (i = 0; i < conf->previous_raid_disks; i++) {
383                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
384                 if (!rdev || test_bit(Faulty, &rdev->flags))
385                         degraded++;
386                 else if (test_bit(In_sync, &rdev->flags))
387                         ;
388                 else
389                         /* not in-sync or faulty.
390                          * If the reshape increases the number of devices,
391                          * this is being recovered by the reshape, so
392                          * this 'previous' section is not in_sync.
393                          * If the number of devices is being reduced however,
394                          * the device can only be part of the array if
395                          * we are reverting a reshape, so this section will
396                          * be in-sync.
397                          */
398                         if (conf->raid_disks >= conf->previous_raid_disks)
399                                 degraded++;
400         }
401         rcu_read_unlock();
402         if (degraded > conf->max_degraded)
403                 return 1;
404         rcu_read_lock();
405         degraded = 0;
406         for (i = 0; i < conf->raid_disks; i++) {
407                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
408                 if (!rdev || test_bit(Faulty, &rdev->flags))
409                         degraded++;
410                 else if (test_bit(In_sync, &rdev->flags))
411                         ;
412                 else
413                         /* not in-sync or faulty.
414                          * If reshape increases the number of devices, this
415                          * section has already been recovered, else it
416                          * almost certainly hasn't.
417                          */
418                         if (conf->raid_disks <= conf->previous_raid_disks)
419                                 degraded++;
420         }
421         rcu_read_unlock();
422         if (degraded > conf->max_degraded)
423                 return 1;
424         return 0;
425 }
426
427 static struct stripe_head *
428 get_active_stripe(struct r5conf *conf, sector_t sector,
429                   int previous, int noblock, int noquiesce)
430 {
431         struct stripe_head *sh;
432
433         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
434
435         spin_lock_irq(&conf->device_lock);
436
437         do {
438                 wait_event_lock_irq(conf->wait_for_stripe,
439                                     conf->quiesce == 0 || noquiesce,
440                                     conf->device_lock, /* nothing */);
441                 sh = __find_stripe(conf, sector, conf->generation - previous);
442                 if (!sh) {
443                         if (!conf->inactive_blocked)
444                                 sh = get_free_stripe(conf);
445                         if (noblock && sh == NULL)
446                                 break;
447                         if (!sh) {
448                                 conf->inactive_blocked = 1;
449                                 wait_event_lock_irq(conf->wait_for_stripe,
450                                                     !list_empty(&conf->inactive_list) &&
451                                                     (atomic_read(&conf->active_stripes)
452                                                      < (conf->max_nr_stripes *3/4)
453                                                      || !conf->inactive_blocked),
454                                                     conf->device_lock,
455                                                     );
456                                 conf->inactive_blocked = 0;
457                         } else
458                                 init_stripe(sh, sector, previous);
459                 } else {
460                         if (atomic_read(&sh->count)) {
461                                 BUG_ON(!list_empty(&sh->lru)
462                                     && !test_bit(STRIPE_EXPANDING, &sh->state));
463                         } else {
464                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
465                                         atomic_inc(&conf->active_stripes);
466                                 if (list_empty(&sh->lru) &&
467                                     !test_bit(STRIPE_EXPANDING, &sh->state))
468                                         BUG();
469                                 list_del_init(&sh->lru);
470                         }
471                 }
472         } while (sh == NULL);
473
474         if (sh)
475                 atomic_inc(&sh->count);
476
477         spin_unlock_irq(&conf->device_lock);
478         return sh;
479 }
480
481 static void
482 raid5_end_read_request(struct bio *bi, int error);
483 static void
484 raid5_end_write_request(struct bio *bi, int error);
485
486 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
487 {
488         struct r5conf *conf = sh->raid_conf;
489         int i, disks = sh->disks;
490
491         might_sleep();
492
493         for (i = disks; i--; ) {
494                 int rw;
495                 struct bio *bi;
496                 struct md_rdev *rdev;
497                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
498                         if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
499                                 rw = WRITE_FUA;
500                         else
501                                 rw = WRITE;
502                 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
503                         rw = READ;
504                 else
505                         continue;
506
507                 bi = &sh->dev[i].req;
508
509                 bi->bi_rw = rw;
510                 if (rw & WRITE)
511                         bi->bi_end_io = raid5_end_write_request;
512                 else
513                         bi->bi_end_io = raid5_end_read_request;
514
515                 rcu_read_lock();
516                 rdev = rcu_dereference(conf->disks[i].rdev);
517                 if (rdev && test_bit(Faulty, &rdev->flags))
518                         rdev = NULL;
519                 if (rdev)
520                         atomic_inc(&rdev->nr_pending);
521                 rcu_read_unlock();
522
523                 /* We have already checked bad blocks for reads.  Now
524                  * need to check for writes.
525                  */
526                 while ((rw & WRITE) && rdev &&
527                        test_bit(WriteErrorSeen, &rdev->flags)) {
528                         sector_t first_bad;
529                         int bad_sectors;
530                         int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
531                                               &first_bad, &bad_sectors);
532                         if (!bad)
533                                 break;
534
535                         if (bad < 0) {
536                                 set_bit(BlockedBadBlocks, &rdev->flags);
537                                 if (!conf->mddev->external &&
538                                     conf->mddev->flags) {
539                                         /* It is very unlikely, but we might
540                                          * still need to write out the
541                                          * bad block log - better give it
542                                          * a chance*/
543                                         md_check_recovery(conf->mddev);
544                                 }
545                                 md_wait_for_blocked_rdev(rdev, conf->mddev);
546                         } else {
547                                 /* Acknowledged bad block - skip the write */
548                                 rdev_dec_pending(rdev, conf->mddev);
549                                 rdev = NULL;
550                         }
551                 }
552
553                 if (rdev) {
554                         if (s->syncing || s->expanding || s->expanded)
555                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
556
557                         set_bit(STRIPE_IO_STARTED, &sh->state);
558
559                         bi->bi_bdev = rdev->bdev;
560                         pr_debug("%s: for %llu schedule op %ld on disc %d\n",
561                                 __func__, (unsigned long long)sh->sector,
562                                 bi->bi_rw, i);
563                         atomic_inc(&sh->count);
564                         bi->bi_sector = sh->sector + rdev->data_offset;
565                         bi->bi_flags = 1 << BIO_UPTODATE;
566                         bi->bi_vcnt = 1;
567                         bi->bi_max_vecs = 1;
568                         bi->bi_idx = 0;
569                         bi->bi_io_vec = &sh->dev[i].vec;
570                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
571                         bi->bi_io_vec[0].bv_offset = 0;
572                         bi->bi_size = STRIPE_SIZE;
573                         bi->bi_next = NULL;
574                         generic_make_request(bi);
575                 } else {
576                         if (rw & WRITE)
577                                 set_bit(STRIPE_DEGRADED, &sh->state);
578                         pr_debug("skip op %ld on disc %d for sector %llu\n",
579                                 bi->bi_rw, i, (unsigned long long)sh->sector);
580                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
581                         set_bit(STRIPE_HANDLE, &sh->state);
582                 }
583         }
584 }
585
586 static struct dma_async_tx_descriptor *
587 async_copy_data(int frombio, struct bio *bio, struct page *page,
588         sector_t sector, struct dma_async_tx_descriptor *tx)
589 {
590         struct bio_vec *bvl;
591         struct page *bio_page;
592         int i;
593         int page_offset;
594         struct async_submit_ctl submit;
595         enum async_tx_flags flags = 0;
596
597         if (bio->bi_sector >= sector)
598                 page_offset = (signed)(bio->bi_sector - sector) * 512;
599         else
600                 page_offset = (signed)(sector - bio->bi_sector) * -512;
601
602         if (frombio)
603                 flags |= ASYNC_TX_FENCE;
604         init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
605
606         bio_for_each_segment(bvl, bio, i) {
607                 int len = bvl->bv_len;
608                 int clen;
609                 int b_offset = 0;
610
611                 if (page_offset < 0) {
612                         b_offset = -page_offset;
613                         page_offset += b_offset;
614                         len -= b_offset;
615                 }
616
617                 if (len > 0 && page_offset + len > STRIPE_SIZE)
618                         clen = STRIPE_SIZE - page_offset;
619                 else
620                         clen = len;
621
622                 if (clen > 0) {
623                         b_offset += bvl->bv_offset;
624                         bio_page = bvl->bv_page;
625                         if (frombio)
626                                 tx = async_memcpy(page, bio_page, page_offset,
627                                                   b_offset, clen, &submit);
628                         else
629                                 tx = async_memcpy(bio_page, page, b_offset,
630                                                   page_offset, clen, &submit);
631                 }
632                 /* chain the operations */
633                 submit.depend_tx = tx;
634
635                 if (clen < len) /* hit end of page */
636                         break;
637                 page_offset +=  len;
638         }
639
640         return tx;
641 }
642
643 static void ops_complete_biofill(void *stripe_head_ref)
644 {
645         struct stripe_head *sh = stripe_head_ref;
646         struct bio *return_bi = NULL;
647         struct r5conf *conf = sh->raid_conf;
648         int i;
649
650         pr_debug("%s: stripe %llu\n", __func__,
651                 (unsigned long long)sh->sector);
652
653         /* clear completed biofills */
654         spin_lock_irq(&conf->device_lock);
655         for (i = sh->disks; i--; ) {
656                 struct r5dev *dev = &sh->dev[i];
657
658                 /* acknowledge completion of a biofill operation */
659                 /* and check if we need to reply to a read request,
660                  * new R5_Wantfill requests are held off until
661                  * !STRIPE_BIOFILL_RUN
662                  */
663                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
664                         struct bio *rbi, *rbi2;
665
666                         BUG_ON(!dev->read);
667                         rbi = dev->read;
668                         dev->read = NULL;
669                         while (rbi && rbi->bi_sector <
670                                 dev->sector + STRIPE_SECTORS) {
671                                 rbi2 = r5_next_bio(rbi, dev->sector);
672                                 if (!raid5_dec_bi_phys_segments(rbi)) {
673                                         rbi->bi_next = return_bi;
674                                         return_bi = rbi;
675                                 }
676                                 rbi = rbi2;
677                         }
678                 }
679         }
680         spin_unlock_irq(&conf->device_lock);
681         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
682
683         return_io(return_bi);
684
685         set_bit(STRIPE_HANDLE, &sh->state);
686         release_stripe(sh);
687 }
688
689 static void ops_run_biofill(struct stripe_head *sh)
690 {
691         struct dma_async_tx_descriptor *tx = NULL;
692         struct r5conf *conf = sh->raid_conf;
693         struct async_submit_ctl submit;
694         int i;
695
696         pr_debug("%s: stripe %llu\n", __func__,
697                 (unsigned long long)sh->sector);
698
699         for (i = sh->disks; i--; ) {
700                 struct r5dev *dev = &sh->dev[i];
701                 if (test_bit(R5_Wantfill, &dev->flags)) {
702                         struct bio *rbi;
703                         spin_lock_irq(&conf->device_lock);
704                         dev->read = rbi = dev->toread;
705                         dev->toread = NULL;
706                         spin_unlock_irq(&conf->device_lock);
707                         while (rbi && rbi->bi_sector <
708                                 dev->sector + STRIPE_SECTORS) {
709                                 tx = async_copy_data(0, rbi, dev->page,
710                                         dev->sector, tx);
711                                 rbi = r5_next_bio(rbi, dev->sector);
712                         }
713                 }
714         }
715
716         atomic_inc(&sh->count);
717         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
718         async_trigger_callback(&submit);
719 }
720
721 static void mark_target_uptodate(struct stripe_head *sh, int target)
722 {
723         struct r5dev *tgt;
724
725         if (target < 0)
726                 return;
727
728         tgt = &sh->dev[target];
729         set_bit(R5_UPTODATE, &tgt->flags);
730         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
731         clear_bit(R5_Wantcompute, &tgt->flags);
732 }
733
734 static void ops_complete_compute(void *stripe_head_ref)
735 {
736         struct stripe_head *sh = stripe_head_ref;
737
738         pr_debug("%s: stripe %llu\n", __func__,
739                 (unsigned long long)sh->sector);
740
741         /* mark the computed target(s) as uptodate */
742         mark_target_uptodate(sh, sh->ops.target);
743         mark_target_uptodate(sh, sh->ops.target2);
744
745         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
746         if (sh->check_state == check_state_compute_run)
747                 sh->check_state = check_state_compute_result;
748         set_bit(STRIPE_HANDLE, &sh->state);
749         release_stripe(sh);
750 }
751
752 /* return a pointer to the address conversion region of the scribble buffer */
753 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
754                                  struct raid5_percpu *percpu)
755 {
756         return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
757 }
758
759 static struct dma_async_tx_descriptor *
760 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
761 {
762         int disks = sh->disks;
763         struct page **xor_srcs = percpu->scribble;
764         int target = sh->ops.target;
765         struct r5dev *tgt = &sh->dev[target];
766         struct page *xor_dest = tgt->page;
767         int count = 0;
768         struct dma_async_tx_descriptor *tx;
769         struct async_submit_ctl submit;
770         int i;
771
772         pr_debug("%s: stripe %llu block: %d\n",
773                 __func__, (unsigned long long)sh->sector, target);
774         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
775
776         for (i = disks; i--; )
777                 if (i != target)
778                         xor_srcs[count++] = sh->dev[i].page;
779
780         atomic_inc(&sh->count);
781
782         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
783                           ops_complete_compute, sh, to_addr_conv(sh, percpu));
784         if (unlikely(count == 1))
785                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
786         else
787                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
788
789         return tx;
790 }
791
792 /* set_syndrome_sources - populate source buffers for gen_syndrome
793  * @srcs - (struct page *) array of size sh->disks
794  * @sh - stripe_head to parse
795  *
796  * Populates srcs in proper layout order for the stripe and returns the
797  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
798  * destination buffer is recorded in srcs[count] and the Q destination
799  * is recorded in srcs[count+1]].
800  */
801 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
802 {
803         int disks = sh->disks;
804         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
805         int d0_idx = raid6_d0(sh);
806         int count;
807         int i;
808
809         for (i = 0; i < disks; i++)
810                 srcs[i] = NULL;
811
812         count = 0;
813         i = d0_idx;
814         do {
815                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
816
817                 srcs[slot] = sh->dev[i].page;
818                 i = raid6_next_disk(i, disks);
819         } while (i != d0_idx);
820
821         return syndrome_disks;
822 }
823
824 static struct dma_async_tx_descriptor *
825 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
826 {
827         int disks = sh->disks;
828         struct page **blocks = percpu->scribble;
829         int target;
830         int qd_idx = sh->qd_idx;
831         struct dma_async_tx_descriptor *tx;
832         struct async_submit_ctl submit;
833         struct r5dev *tgt;
834         struct page *dest;
835         int i;
836         int count;
837
838         if (sh->ops.target < 0)
839                 target = sh->ops.target2;
840         else if (sh->ops.target2 < 0)
841                 target = sh->ops.target;
842         else
843                 /* we should only have one valid target */
844                 BUG();
845         BUG_ON(target < 0);
846         pr_debug("%s: stripe %llu block: %d\n",
847                 __func__, (unsigned long long)sh->sector, target);
848
849         tgt = &sh->dev[target];
850         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
851         dest = tgt->page;
852
853         atomic_inc(&sh->count);
854
855         if (target == qd_idx) {
856                 count = set_syndrome_sources(blocks, sh);
857                 blocks[count] = NULL; /* regenerating p is not necessary */
858                 BUG_ON(blocks[count+1] != dest); /* q should already be set */
859                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
860                                   ops_complete_compute, sh,
861                                   to_addr_conv(sh, percpu));
862                 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
863         } else {
864                 /* Compute any data- or p-drive using XOR */
865                 count = 0;
866                 for (i = disks; i-- ; ) {
867                         if (i == target || i == qd_idx)
868                                 continue;
869                         blocks[count++] = sh->dev[i].page;
870                 }
871
872                 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
873                                   NULL, ops_complete_compute, sh,
874                                   to_addr_conv(sh, percpu));
875                 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
876         }
877
878         return tx;
879 }
880
881 static struct dma_async_tx_descriptor *
882 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
883 {
884         int i, count, disks = sh->disks;
885         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
886         int d0_idx = raid6_d0(sh);
887         int faila = -1, failb = -1;
888         int target = sh->ops.target;
889         int target2 = sh->ops.target2;
890         struct r5dev *tgt = &sh->dev[target];
891         struct r5dev *tgt2 = &sh->dev[target2];
892         struct dma_async_tx_descriptor *tx;
893         struct page **blocks = percpu->scribble;
894         struct async_submit_ctl submit;
895
896         pr_debug("%s: stripe %llu block1: %d block2: %d\n",
897                  __func__, (unsigned long long)sh->sector, target, target2);
898         BUG_ON(target < 0 || target2 < 0);
899         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
900         BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
901
902         /* we need to open-code set_syndrome_sources to handle the
903          * slot number conversion for 'faila' and 'failb'
904          */
905         for (i = 0; i < disks ; i++)
906                 blocks[i] = NULL;
907         count = 0;
908         i = d0_idx;
909         do {
910                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
911
912                 blocks[slot] = sh->dev[i].page;
913
914                 if (i == target)
915                         faila = slot;
916                 if (i == target2)
917                         failb = slot;
918                 i = raid6_next_disk(i, disks);
919         } while (i != d0_idx);
920
921         BUG_ON(faila == failb);
922         if (failb < faila)
923                 swap(faila, failb);
924         pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
925                  __func__, (unsigned long long)sh->sector, faila, failb);
926
927         atomic_inc(&sh->count);
928
929         if (failb == syndrome_disks+1) {
930                 /* Q disk is one of the missing disks */
931                 if (faila == syndrome_disks) {
932                         /* Missing P+Q, just recompute */
933                         init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
934                                           ops_complete_compute, sh,
935                                           to_addr_conv(sh, percpu));
936                         return async_gen_syndrome(blocks, 0, syndrome_disks+2,
937                                                   STRIPE_SIZE, &submit);
938                 } else {
939                         struct page *dest;
940                         int data_target;
941                         int qd_idx = sh->qd_idx;
942
943                         /* Missing D+Q: recompute D from P, then recompute Q */
944                         if (target == qd_idx)
945                                 data_target = target2;
946                         else
947                                 data_target = target;
948
949                         count = 0;
950                         for (i = disks; i-- ; ) {
951                                 if (i == data_target || i == qd_idx)
952                                         continue;
953                                 blocks[count++] = sh->dev[i].page;
954                         }
955                         dest = sh->dev[data_target].page;
956                         init_async_submit(&submit,
957                                           ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
958                                           NULL, NULL, NULL,
959                                           to_addr_conv(sh, percpu));
960                         tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
961                                        &submit);
962
963                         count = set_syndrome_sources(blocks, sh);
964                         init_async_submit(&submit, ASYNC_TX_FENCE, tx,
965                                           ops_complete_compute, sh,
966                                           to_addr_conv(sh, percpu));
967                         return async_gen_syndrome(blocks, 0, count+2,
968                                                   STRIPE_SIZE, &submit);
969                 }
970         } else {
971                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
972                                   ops_complete_compute, sh,
973                                   to_addr_conv(sh, percpu));
974                 if (failb == syndrome_disks) {
975                         /* We're missing D+P. */
976                         return async_raid6_datap_recov(syndrome_disks+2,
977                                                        STRIPE_SIZE, faila,
978                                                        blocks, &submit);
979                 } else {
980                         /* We're missing D+D. */
981                         return async_raid6_2data_recov(syndrome_disks+2,
982                                                        STRIPE_SIZE, faila, failb,
983                                                        blocks, &submit);
984                 }
985         }
986 }
987
988
989 static void ops_complete_prexor(void *stripe_head_ref)
990 {
991         struct stripe_head *sh = stripe_head_ref;
992
993         pr_debug("%s: stripe %llu\n", __func__,
994                 (unsigned long long)sh->sector);
995 }
996
997 static struct dma_async_tx_descriptor *
998 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
999                struct dma_async_tx_descriptor *tx)
1000 {
1001         int disks = sh->disks;
1002         struct page **xor_srcs = percpu->scribble;
1003         int count = 0, pd_idx = sh->pd_idx, i;
1004         struct async_submit_ctl submit;
1005
1006         /* existing parity data subtracted */
1007         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1008
1009         pr_debug("%s: stripe %llu\n", __func__,
1010                 (unsigned long long)sh->sector);
1011
1012         for (i = disks; i--; ) {
1013                 struct r5dev *dev = &sh->dev[i];
1014                 /* Only process blocks that are known to be uptodate */
1015                 if (test_bit(R5_Wantdrain, &dev->flags))
1016                         xor_srcs[count++] = dev->page;
1017         }
1018
1019         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1020                           ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1021         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1022
1023         return tx;
1024 }
1025
1026 static struct dma_async_tx_descriptor *
1027 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1028 {
1029         int disks = sh->disks;
1030         int i;
1031
1032         pr_debug("%s: stripe %llu\n", __func__,
1033                 (unsigned long long)sh->sector);
1034
1035         for (i = disks; i--; ) {
1036                 struct r5dev *dev = &sh->dev[i];
1037                 struct bio *chosen;
1038
1039                 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1040                         struct bio *wbi;
1041
1042                         spin_lock_irq(&sh->raid_conf->device_lock);
1043                         chosen = dev->towrite;
1044                         dev->towrite = NULL;
1045                         BUG_ON(dev->written);
1046                         wbi = dev->written = chosen;
1047                         spin_unlock_irq(&sh->raid_conf->device_lock);
1048
1049                         while (wbi && wbi->bi_sector <
1050                                 dev->sector + STRIPE_SECTORS) {
1051                                 if (wbi->bi_rw & REQ_FUA)
1052                                         set_bit(R5_WantFUA, &dev->flags);
1053                                 tx = async_copy_data(1, wbi, dev->page,
1054                                         dev->sector, tx);
1055                                 wbi = r5_next_bio(wbi, dev->sector);
1056                         }
1057                 }
1058         }
1059
1060         return tx;
1061 }
1062
1063 static void ops_complete_reconstruct(void *stripe_head_ref)
1064 {
1065         struct stripe_head *sh = stripe_head_ref;
1066         int disks = sh->disks;
1067         int pd_idx = sh->pd_idx;
1068         int qd_idx = sh->qd_idx;
1069         int i;
1070         bool fua = false;
1071
1072         pr_debug("%s: stripe %llu\n", __func__,
1073                 (unsigned long long)sh->sector);
1074
1075         for (i = disks; i--; )
1076                 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1077
1078         for (i = disks; i--; ) {
1079                 struct r5dev *dev = &sh->dev[i];
1080
1081                 if (dev->written || i == pd_idx || i == qd_idx) {
1082                         set_bit(R5_UPTODATE, &dev->flags);
1083                         if (fua)
1084                                 set_bit(R5_WantFUA, &dev->flags);
1085                 }
1086         }
1087
1088         if (sh->reconstruct_state == reconstruct_state_drain_run)
1089                 sh->reconstruct_state = reconstruct_state_drain_result;
1090         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1091                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1092         else {
1093                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1094                 sh->reconstruct_state = reconstruct_state_result;
1095         }
1096
1097         set_bit(STRIPE_HANDLE, &sh->state);
1098         release_stripe(sh);
1099 }
1100
1101 static void
1102 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1103                      struct dma_async_tx_descriptor *tx)
1104 {
1105         int disks = sh->disks;
1106         struct page **xor_srcs = percpu->scribble;
1107         struct async_submit_ctl submit;
1108         int count = 0, pd_idx = sh->pd_idx, i;
1109         struct page *xor_dest;
1110         int prexor = 0;
1111         unsigned long flags;
1112
1113         pr_debug("%s: stripe %llu\n", __func__,
1114                 (unsigned long long)sh->sector);
1115
1116         /* check if prexor is active which means only process blocks
1117          * that are part of a read-modify-write (written)
1118          */
1119         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1120                 prexor = 1;
1121                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1122                 for (i = disks; i--; ) {
1123                         struct r5dev *dev = &sh->dev[i];
1124                         if (dev->written)
1125                                 xor_srcs[count++] = dev->page;
1126                 }
1127         } else {
1128                 xor_dest = sh->dev[pd_idx].page;
1129                 for (i = disks; i--; ) {
1130                         struct r5dev *dev = &sh->dev[i];
1131                         if (i != pd_idx)
1132                                 xor_srcs[count++] = dev->page;
1133                 }
1134         }
1135
1136         /* 1/ if we prexor'd then the dest is reused as a source
1137          * 2/ if we did not prexor then we are redoing the parity
1138          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1139          * for the synchronous xor case
1140          */
1141         flags = ASYNC_TX_ACK |
1142                 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1143
1144         atomic_inc(&sh->count);
1145
1146         init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1147                           to_addr_conv(sh, percpu));
1148         if (unlikely(count == 1))
1149                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1150         else
1151                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1152 }
1153
1154 static void
1155 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1156                      struct dma_async_tx_descriptor *tx)
1157 {
1158         struct async_submit_ctl submit;
1159         struct page **blocks = percpu->scribble;
1160         int count;
1161
1162         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1163
1164         count = set_syndrome_sources(blocks, sh);
1165
1166         atomic_inc(&sh->count);
1167
1168         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1169                           sh, to_addr_conv(sh, percpu));
1170         async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1171 }
1172
1173 static void ops_complete_check(void *stripe_head_ref)
1174 {
1175         struct stripe_head *sh = stripe_head_ref;
1176
1177         pr_debug("%s: stripe %llu\n", __func__,
1178                 (unsigned long long)sh->sector);
1179
1180         sh->check_state = check_state_check_result;
1181         set_bit(STRIPE_HANDLE, &sh->state);
1182         release_stripe(sh);
1183 }
1184
1185 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1186 {
1187         int disks = sh->disks;
1188         int pd_idx = sh->pd_idx;
1189         int qd_idx = sh->qd_idx;
1190         struct page *xor_dest;
1191         struct page **xor_srcs = percpu->scribble;
1192         struct dma_async_tx_descriptor *tx;
1193         struct async_submit_ctl submit;
1194         int count;
1195         int i;
1196
1197         pr_debug("%s: stripe %llu\n", __func__,
1198                 (unsigned long long)sh->sector);
1199
1200         count = 0;
1201         xor_dest = sh->dev[pd_idx].page;
1202         xor_srcs[count++] = xor_dest;
1203         for (i = disks; i--; ) {
1204                 if (i == pd_idx || i == qd_idx)
1205                         continue;
1206                 xor_srcs[count++] = sh->dev[i].page;
1207         }
1208
1209         init_async_submit(&submit, 0, NULL, NULL, NULL,
1210                           to_addr_conv(sh, percpu));
1211         tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1212                            &sh->ops.zero_sum_result, &submit);
1213
1214         atomic_inc(&sh->count);
1215         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1216         tx = async_trigger_callback(&submit);
1217 }
1218
1219 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1220 {
1221         struct page **srcs = percpu->scribble;
1222         struct async_submit_ctl submit;
1223         int count;
1224
1225         pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1226                 (unsigned long long)sh->sector, checkp);
1227
1228         count = set_syndrome_sources(srcs, sh);
1229         if (!checkp)
1230                 srcs[count] = NULL;
1231
1232         atomic_inc(&sh->count);
1233         init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1234                           sh, to_addr_conv(sh, percpu));
1235         async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1236                            &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1237 }
1238
1239 static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1240 {
1241         int overlap_clear = 0, i, disks = sh->disks;
1242         struct dma_async_tx_descriptor *tx = NULL;
1243         struct r5conf *conf = sh->raid_conf;
1244         int level = conf->level;
1245         struct raid5_percpu *percpu;
1246         unsigned long cpu;
1247
1248         cpu = get_cpu();
1249         percpu = per_cpu_ptr(conf->percpu, cpu);
1250         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1251                 ops_run_biofill(sh);
1252                 overlap_clear++;
1253         }
1254
1255         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1256                 if (level < 6)
1257                         tx = ops_run_compute5(sh, percpu);
1258                 else {
1259                         if (sh->ops.target2 < 0 || sh->ops.target < 0)
1260                                 tx = ops_run_compute6_1(sh, percpu);
1261                         else
1262                                 tx = ops_run_compute6_2(sh, percpu);
1263                 }
1264                 /* terminate the chain if reconstruct is not set to be run */
1265                 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1266                         async_tx_ack(tx);
1267         }
1268
1269         if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1270                 tx = ops_run_prexor(sh, percpu, tx);
1271
1272         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1273                 tx = ops_run_biodrain(sh, tx);
1274                 overlap_clear++;
1275         }
1276
1277         if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1278                 if (level < 6)
1279                         ops_run_reconstruct5(sh, percpu, tx);
1280                 else
1281                         ops_run_reconstruct6(sh, percpu, tx);
1282         }
1283
1284         if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1285                 if (sh->check_state == check_state_run)
1286                         ops_run_check_p(sh, percpu);
1287                 else if (sh->check_state == check_state_run_q)
1288                         ops_run_check_pq(sh, percpu, 0);
1289                 else if (sh->check_state == check_state_run_pq)
1290                         ops_run_check_pq(sh, percpu, 1);
1291                 else
1292                         BUG();
1293         }
1294
1295         if (overlap_clear)
1296                 for (i = disks; i--; ) {
1297                         struct r5dev *dev = &sh->dev[i];
1298                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
1299                                 wake_up(&sh->raid_conf->wait_for_overlap);
1300                 }
1301         put_cpu();
1302 }
1303
1304 #ifdef CONFIG_MULTICORE_RAID456
1305 static void async_run_ops(void *param, async_cookie_t cookie)
1306 {
1307         struct stripe_head *sh = param;
1308         unsigned long ops_request = sh->ops.request;
1309
1310         clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1311         wake_up(&sh->ops.wait_for_ops);
1312
1313         __raid_run_ops(sh, ops_request);
1314         release_stripe(sh);
1315 }
1316
1317 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1318 {
1319         /* since handle_stripe can be called outside of raid5d context
1320          * we need to ensure sh->ops.request is de-staged before another
1321          * request arrives
1322          */
1323         wait_event(sh->ops.wait_for_ops,
1324                    !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1325         sh->ops.request = ops_request;
1326
1327         atomic_inc(&sh->count);
1328         async_schedule(async_run_ops, sh);
1329 }
1330 #else
1331 #define raid_run_ops __raid_run_ops
1332 #endif
1333
1334 static int grow_one_stripe(struct r5conf *conf)
1335 {
1336         struct stripe_head *sh;
1337         sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1338         if (!sh)
1339                 return 0;
1340
1341         sh->raid_conf = conf;
1342         #ifdef CONFIG_MULTICORE_RAID456
1343         init_waitqueue_head(&sh->ops.wait_for_ops);
1344         #endif
1345
1346         if (grow_buffers(sh)) {
1347                 shrink_buffers(sh);
1348                 kmem_cache_free(conf->slab_cache, sh);
1349                 return 0;
1350         }
1351         /* we just created an active stripe so... */
1352         atomic_set(&sh->count, 1);
1353         atomic_inc(&conf->active_stripes);
1354         INIT_LIST_HEAD(&sh->lru);
1355         release_stripe(sh);
1356         return 1;
1357 }
1358
1359 static int grow_stripes(struct r5conf *conf, int num)
1360 {
1361         struct kmem_cache *sc;
1362         int devs = max(conf->raid_disks, conf->previous_raid_disks);
1363
1364         if (conf->mddev->gendisk)
1365                 sprintf(conf->cache_name[0],
1366                         "raid%d-%s", conf->level, mdname(conf->mddev));
1367         else
1368                 sprintf(conf->cache_name[0],
1369                         "raid%d-%p", conf->level, conf->mddev);
1370         sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1371
1372         conf->active_name = 0;
1373         sc = kmem_cache_create(conf->cache_name[conf->active_name],
1374                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1375                                0, 0, NULL);
1376         if (!sc)
1377                 return 1;
1378         conf->slab_cache = sc;
1379         conf->pool_size = devs;
1380         while (num--)
1381                 if (!grow_one_stripe(conf))
1382                         return 1;
1383         return 0;
1384 }
1385
1386 /**
1387  * scribble_len - return the required size of the scribble region
1388  * @num - total number of disks in the array
1389  *
1390  * The size must be enough to contain:
1391  * 1/ a struct page pointer for each device in the array +2
1392  * 2/ room to convert each entry in (1) to its corresponding dma
1393  *    (dma_map_page()) or page (page_address()) address.
1394  *
1395  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1396  * calculate over all devices (not just the data blocks), using zeros in place
1397  * of the P and Q blocks.
1398  */
1399 static size_t scribble_len(int num)
1400 {
1401         size_t len;
1402
1403         len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1404
1405         return len;
1406 }
1407
1408 static int resize_stripes(struct r5conf *conf, int newsize)
1409 {
1410         /* Make all the stripes able to hold 'newsize' devices.
1411          * New slots in each stripe get 'page' set to a new page.
1412          *
1413          * This happens in stages:
1414          * 1/ create a new kmem_cache and allocate the required number of
1415          *    stripe_heads.
1416          * 2/ gather all the old stripe_heads and tranfer the pages across
1417          *    to the new stripe_heads.  This will have the side effect of
1418          *    freezing the array as once all stripe_heads have been collected,
1419          *    no IO will be possible.  Old stripe heads are freed once their
1420          *    pages have been transferred over, and the old kmem_cache is
1421          *    freed when all stripes are done.
1422          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
1423          *    we simple return a failre status - no need to clean anything up.
1424          * 4/ allocate new pages for the new slots in the new stripe_heads.
1425          *    If this fails, we don't bother trying the shrink the
1426          *    stripe_heads down again, we just leave them as they are.
1427          *    As each stripe_head is processed the new one is released into
1428          *    active service.
1429          *
1430          * Once step2 is started, we cannot afford to wait for a write,
1431          * so we use GFP_NOIO allocations.
1432          */
1433         struct stripe_head *osh, *nsh;
1434         LIST_HEAD(newstripes);
1435         struct disk_info *ndisks;
1436         unsigned long cpu;
1437         int err;
1438         struct kmem_cache *sc;
1439         int i;
1440
1441         if (newsize <= conf->pool_size)
1442                 return 0; /* never bother to shrink */
1443
1444         err = md_allow_write(conf->mddev);
1445         if (err)
1446                 return err;
1447
1448         /* Step 1 */
1449         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1450                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1451                                0, 0, NULL);
1452         if (!sc)
1453                 return -ENOMEM;
1454
1455         for (i = conf->max_nr_stripes; i; i--) {
1456                 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1457                 if (!nsh)
1458                         break;
1459
1460                 nsh->raid_conf = conf;
1461                 #ifdef CONFIG_MULTICORE_RAID456
1462                 init_waitqueue_head(&nsh->ops.wait_for_ops);
1463                 #endif
1464
1465                 list_add(&nsh->lru, &newstripes);
1466         }
1467         if (i) {
1468                 /* didn't get enough, give up */
1469                 while (!list_empty(&newstripes)) {
1470                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
1471                         list_del(&nsh->lru);
1472                         kmem_cache_free(sc, nsh);
1473                 }
1474                 kmem_cache_destroy(sc);
1475                 return -ENOMEM;
1476         }
1477         /* Step 2 - Must use GFP_NOIO now.
1478          * OK, we have enough stripes, start collecting inactive
1479          * stripes and copying them over
1480          */
1481         list_for_each_entry(nsh, &newstripes, lru) {
1482                 spin_lock_irq(&conf->device_lock);
1483                 wait_event_lock_irq(conf->wait_for_stripe,
1484                                     !list_empty(&conf->inactive_list),
1485                                     conf->device_lock,
1486                                     );
1487                 osh = get_free_stripe(conf);
1488                 spin_unlock_irq(&conf->device_lock);
1489                 atomic_set(&nsh->count, 1);
1490                 for(i=0; i<conf->pool_size; i++)
1491                         nsh->dev[i].page = osh->dev[i].page;
1492                 for( ; i<newsize; i++)
1493                         nsh->dev[i].page = NULL;
1494                 kmem_cache_free(conf->slab_cache, osh);
1495         }
1496         kmem_cache_destroy(conf->slab_cache);
1497
1498         /* Step 3.
1499          * At this point, we are holding all the stripes so the array
1500          * is completely stalled, so now is a good time to resize
1501          * conf->disks and the scribble region
1502          */
1503         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1504         if (ndisks) {
1505                 for (i=0; i<conf->raid_disks; i++)
1506                         ndisks[i] = conf->disks[i];
1507                 kfree(conf->disks);
1508                 conf->disks = ndisks;
1509         } else
1510                 err = -ENOMEM;
1511
1512         get_online_cpus();
1513         conf->scribble_len = scribble_len(newsize);
1514         for_each_present_cpu(cpu) {
1515                 struct raid5_percpu *percpu;
1516                 void *scribble;
1517
1518                 percpu = per_cpu_ptr(conf->percpu, cpu);
1519                 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1520
1521                 if (scribble) {
1522                         kfree(percpu->scribble);
1523                         percpu->scribble = scribble;
1524                 } else {
1525                         err = -ENOMEM;
1526                         break;
1527                 }
1528         }
1529         put_online_cpus();
1530
1531         /* Step 4, return new stripes to service */
1532         while(!list_empty(&newstripes)) {
1533                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1534                 list_del_init(&nsh->lru);
1535
1536                 for (i=conf->raid_disks; i < newsize; i++)
1537                         if (nsh->dev[i].page == NULL) {
1538                                 struct page *p = alloc_page(GFP_NOIO);
1539                                 nsh->dev[i].page = p;
1540                                 if (!p)
1541                                         err = -ENOMEM;
1542                         }
1543                 release_stripe(nsh);
1544         }
1545         /* critical section pass, GFP_NOIO no longer needed */
1546
1547         conf->slab_cache = sc;
1548         conf->active_name = 1-conf->active_name;
1549         conf->pool_size = newsize;
1550         return err;
1551 }
1552
1553 static int drop_one_stripe(struct r5conf *conf)
1554 {
1555         struct stripe_head *sh;
1556
1557         spin_lock_irq(&conf->device_lock);
1558         sh = get_free_stripe(conf);
1559         spin_unlock_irq(&conf->device_lock);
1560         if (!sh)
1561                 return 0;
1562         BUG_ON(atomic_read(&sh->count));
1563         shrink_buffers(sh);
1564         kmem_cache_free(conf->slab_cache, sh);
1565         atomic_dec(&conf->active_stripes);
1566         return 1;
1567 }
1568
1569 static void shrink_stripes(struct r5conf *conf)
1570 {
1571         while (drop_one_stripe(conf))
1572                 ;
1573
1574         if (conf->slab_cache)
1575                 kmem_cache_destroy(conf->slab_cache);
1576         conf->slab_cache = NULL;
1577 }
1578
1579 static void raid5_end_read_request(struct bio * bi, int error)
1580 {
1581         struct stripe_head *sh = bi->bi_private;
1582         struct r5conf *conf = sh->raid_conf;
1583         int disks = sh->disks, i;
1584         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1585         char b[BDEVNAME_SIZE];
1586         struct md_rdev *rdev;
1587
1588
1589         for (i=0 ; i<disks; i++)
1590                 if (bi == &sh->dev[i].req)
1591                         break;
1592
1593         pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1594                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1595                 uptodate);
1596         if (i == disks) {
1597                 BUG();
1598                 return;
1599         }
1600
1601         if (uptodate) {
1602                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1603                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1604                         rdev = conf->disks[i].rdev;
1605                         printk_ratelimited(
1606                                 KERN_INFO
1607                                 "md/raid:%s: read error corrected"
1608                                 " (%lu sectors at %llu on %s)\n",
1609                                 mdname(conf->mddev), STRIPE_SECTORS,
1610                                 (unsigned long long)(sh->sector
1611                                                      + rdev->data_offset),
1612                                 bdevname(rdev->bdev, b));
1613                         atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1614                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1615                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1616                 }
1617                 if (atomic_read(&conf->disks[i].rdev->read_errors))
1618                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
1619         } else {
1620                 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1621                 int retry = 0;
1622                 rdev = conf->disks[i].rdev;
1623
1624                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1625                 atomic_inc(&rdev->read_errors);
1626                 if (conf->mddev->degraded >= conf->max_degraded)
1627                         printk_ratelimited(
1628                                 KERN_WARNING
1629                                 "md/raid:%s: read error not correctable "
1630                                 "(sector %llu on %s).\n",
1631                                 mdname(conf->mddev),
1632                                 (unsigned long long)(sh->sector
1633                                                      + rdev->data_offset),
1634                                 bdn);
1635                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1636                         /* Oh, no!!! */
1637                         printk_ratelimited(
1638                                 KERN_WARNING
1639                                 "md/raid:%s: read error NOT corrected!! "
1640                                 "(sector %llu on %s).\n",
1641                                 mdname(conf->mddev),
1642                                 (unsigned long long)(sh->sector
1643                                                      + rdev->data_offset),
1644                                 bdn);
1645                 else if (atomic_read(&rdev->read_errors)
1646                          > conf->max_nr_stripes)
1647                         printk(KERN_WARNING
1648                                "md/raid:%s: Too many read errors, failing device %s.\n",
1649                                mdname(conf->mddev), bdn);
1650                 else
1651                         retry = 1;
1652                 if (retry)
1653                         set_bit(R5_ReadError, &sh->dev[i].flags);
1654                 else {
1655                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1656                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1657                         md_error(conf->mddev, rdev);
1658                 }
1659         }
1660         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1661         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1662         set_bit(STRIPE_HANDLE, &sh->state);
1663         release_stripe(sh);
1664 }
1665
1666 static void raid5_end_write_request(struct bio *bi, int error)
1667 {
1668         struct stripe_head *sh = bi->bi_private;
1669         struct r5conf *conf = sh->raid_conf;
1670         int disks = sh->disks, i;
1671         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1672         sector_t first_bad;
1673         int bad_sectors;
1674
1675         for (i=0 ; i<disks; i++)
1676                 if (bi == &sh->dev[i].req)
1677                         break;
1678
1679         pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1680                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1681                 uptodate);
1682         if (i == disks) {
1683                 BUG();
1684                 return;
1685         }
1686
1687         if (!uptodate) {
1688                 set_bit(WriteErrorSeen, &conf->disks[i].rdev->flags);
1689                 set_bit(R5_WriteError, &sh->dev[i].flags);
1690         } else if (is_badblock(conf->disks[i].rdev, sh->sector, STRIPE_SECTORS,
1691                                &first_bad, &bad_sectors))
1692                 set_bit(R5_MadeGood, &sh->dev[i].flags);
1693
1694         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1695         
1696         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1697         set_bit(STRIPE_HANDLE, &sh->state);
1698         release_stripe(sh);
1699 }
1700
1701
1702 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1703         
1704 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1705 {
1706         struct r5dev *dev = &sh->dev[i];
1707
1708         bio_init(&dev->req);
1709         dev->req.bi_io_vec = &dev->vec;
1710         dev->req.bi_vcnt++;
1711         dev->req.bi_max_vecs++;
1712         dev->vec.bv_page = dev->page;
1713         dev->vec.bv_len = STRIPE_SIZE;
1714         dev->vec.bv_offset = 0;
1715
1716         dev->req.bi_sector = sh->sector;
1717         dev->req.bi_private = sh;
1718
1719         dev->flags = 0;
1720         dev->sector = compute_blocknr(sh, i, previous);
1721 }
1722
1723 static void error(struct mddev *mddev, struct md_rdev *rdev)
1724 {
1725         char b[BDEVNAME_SIZE];
1726         struct r5conf *conf = mddev->private;
1727         pr_debug("raid456: error called\n");
1728
1729         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1730                 unsigned long flags;
1731                 spin_lock_irqsave(&conf->device_lock, flags);
1732                 mddev->degraded++;
1733                 spin_unlock_irqrestore(&conf->device_lock, flags);
1734                 /*
1735                  * if recovery was running, make sure it aborts.
1736                  */
1737                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1738         }
1739         set_bit(Blocked, &rdev->flags);
1740         set_bit(Faulty, &rdev->flags);
1741         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1742         printk(KERN_ALERT
1743                "md/raid:%s: Disk failure on %s, disabling device.\n"
1744                "md/raid:%s: Operation continuing on %d devices.\n",
1745                mdname(mddev),
1746                bdevname(rdev->bdev, b),
1747                mdname(mddev),
1748                conf->raid_disks - mddev->degraded);
1749 }
1750
1751 /*
1752  * Input: a 'big' sector number,
1753  * Output: index of the data and parity disk, and the sector # in them.
1754  */
1755 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
1756                                      int previous, int *dd_idx,
1757                                      struct stripe_head *sh)
1758 {
1759         sector_t stripe, stripe2;
1760         sector_t chunk_number;
1761         unsigned int chunk_offset;
1762         int pd_idx, qd_idx;
1763         int ddf_layout = 0;
1764         sector_t new_sector;
1765         int algorithm = previous ? conf->prev_algo
1766                                  : conf->algorithm;
1767         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1768                                          : conf->chunk_sectors;
1769         int raid_disks = previous ? conf->previous_raid_disks
1770                                   : conf->raid_disks;
1771         int data_disks = raid_disks - conf->max_degraded;
1772
1773         /* First compute the information on this sector */
1774
1775         /*
1776          * Compute the chunk number and the sector offset inside the chunk
1777          */
1778         chunk_offset = sector_div(r_sector, sectors_per_chunk);
1779         chunk_number = r_sector;
1780
1781         /*
1782          * Compute the stripe number
1783          */
1784         stripe = chunk_number;
1785         *dd_idx = sector_div(stripe, data_disks);
1786         stripe2 = stripe;
1787         /*
1788          * Select the parity disk based on the user selected algorithm.
1789          */
1790         pd_idx = qd_idx = -1;
1791         switch(conf->level) {
1792         case 4:
1793                 pd_idx = data_disks;
1794                 break;
1795         case 5:
1796                 switch (algorithm) {
1797                 case ALGORITHM_LEFT_ASYMMETRIC:
1798                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
1799                         if (*dd_idx >= pd_idx)
1800                                 (*dd_idx)++;
1801                         break;
1802                 case ALGORITHM_RIGHT_ASYMMETRIC:
1803                         pd_idx = sector_div(stripe2, raid_disks);
1804                         if (*dd_idx >= pd_idx)
1805                                 (*dd_idx)++;
1806                         break;
1807                 case ALGORITHM_LEFT_SYMMETRIC:
1808                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
1809                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1810                         break;
1811                 case ALGORITHM_RIGHT_SYMMETRIC:
1812                         pd_idx = sector_div(stripe2, raid_disks);
1813                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1814                         break;
1815                 case ALGORITHM_PARITY_0:
1816                         pd_idx = 0;
1817                         (*dd_idx)++;
1818                         break;
1819                 case ALGORITHM_PARITY_N:
1820                         pd_idx = data_disks;
1821                         break;
1822                 default:
1823                         BUG();
1824                 }
1825                 break;
1826         case 6:
1827
1828                 switch (algorithm) {
1829                 case ALGORITHM_LEFT_ASYMMETRIC:
1830                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1831                         qd_idx = pd_idx + 1;
1832                         if (pd_idx == raid_disks-1) {
1833                                 (*dd_idx)++;    /* Q D D D P */
1834                                 qd_idx = 0;
1835                         } else if (*dd_idx >= pd_idx)
1836                                 (*dd_idx) += 2; /* D D P Q D */
1837                         break;
1838                 case ALGORITHM_RIGHT_ASYMMETRIC:
1839                         pd_idx = sector_div(stripe2, raid_disks);
1840                         qd_idx = pd_idx + 1;
1841                         if (pd_idx == raid_disks-1) {
1842                                 (*dd_idx)++;    /* Q D D D P */
1843                                 qd_idx = 0;
1844                         } else if (*dd_idx >= pd_idx)
1845                                 (*dd_idx) += 2; /* D D P Q D */
1846                         break;
1847                 case ALGORITHM_LEFT_SYMMETRIC:
1848                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1849                         qd_idx = (pd_idx + 1) % raid_disks;
1850                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1851                         break;
1852                 case ALGORITHM_RIGHT_SYMMETRIC:
1853                         pd_idx = sector_div(stripe2, raid_disks);
1854                         qd_idx = (pd_idx + 1) % raid_disks;
1855                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1856                         break;
1857
1858                 case ALGORITHM_PARITY_0:
1859                         pd_idx = 0;
1860                         qd_idx = 1;
1861                         (*dd_idx) += 2;
1862                         break;
1863                 case ALGORITHM_PARITY_N:
1864                         pd_idx = data_disks;
1865                         qd_idx = data_disks + 1;
1866                         break;
1867
1868                 case ALGORITHM_ROTATING_ZERO_RESTART:
1869                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
1870                          * of blocks for computing Q is different.
1871                          */
1872                         pd_idx = sector_div(stripe2, raid_disks);
1873                         qd_idx = pd_idx + 1;
1874                         if (pd_idx == raid_disks-1) {
1875                                 (*dd_idx)++;    /* Q D D D P */
1876                                 qd_idx = 0;
1877                         } else if (*dd_idx >= pd_idx)
1878                                 (*dd_idx) += 2; /* D D P Q D */
1879                         ddf_layout = 1;
1880                         break;
1881
1882                 case ALGORITHM_ROTATING_N_RESTART:
1883                         /* Same a left_asymmetric, by first stripe is
1884                          * D D D P Q  rather than
1885                          * Q D D D P
1886                          */
1887                         stripe2 += 1;
1888                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1889                         qd_idx = pd_idx + 1;
1890                         if (pd_idx == raid_disks-1) {
1891                                 (*dd_idx)++;    /* Q D D D P */
1892                                 qd_idx = 0;
1893                         } else if (*dd_idx >= pd_idx)
1894                                 (*dd_idx) += 2; /* D D P Q D */
1895                         ddf_layout = 1;
1896                         break;
1897
1898                 case ALGORITHM_ROTATING_N_CONTINUE:
1899                         /* Same as left_symmetric but Q is before P */
1900                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1901                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1902                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1903                         ddf_layout = 1;
1904                         break;
1905
1906                 case ALGORITHM_LEFT_ASYMMETRIC_6:
1907                         /* RAID5 left_asymmetric, with Q on last device */
1908                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1909                         if (*dd_idx >= pd_idx)
1910                                 (*dd_idx)++;
1911                         qd_idx = raid_disks - 1;
1912                         break;
1913
1914                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1915                         pd_idx = sector_div(stripe2, raid_disks-1);
1916                         if (*dd_idx >= pd_idx)
1917                                 (*dd_idx)++;
1918                         qd_idx = raid_disks - 1;
1919                         break;
1920
1921                 case ALGORITHM_LEFT_SYMMETRIC_6:
1922                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1923                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1924                         qd_idx = raid_disks - 1;
1925                         break;
1926
1927                 case ALGORITHM_RIGHT_SYMMETRIC_6:
1928                         pd_idx = sector_div(stripe2, raid_disks-1);
1929                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1930                         qd_idx = raid_disks - 1;
1931                         break;
1932
1933                 case ALGORITHM_PARITY_0_6:
1934                         pd_idx = 0;
1935                         (*dd_idx)++;
1936                         qd_idx = raid_disks - 1;
1937                         break;
1938
1939                 default:
1940                         BUG();
1941                 }
1942                 break;
1943         }
1944
1945         if (sh) {
1946                 sh->pd_idx = pd_idx;
1947                 sh->qd_idx = qd_idx;
1948                 sh->ddf_layout = ddf_layout;
1949         }
1950         /*
1951          * Finally, compute the new sector number
1952          */
1953         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1954         return new_sector;
1955 }
1956
1957
1958 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1959 {
1960         struct r5conf *conf = sh->raid_conf;
1961         int raid_disks = sh->disks;
1962         int data_disks = raid_disks - conf->max_degraded;
1963         sector_t new_sector = sh->sector, check;
1964         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1965                                          : conf->chunk_sectors;
1966         int algorithm = previous ? conf->prev_algo
1967                                  : conf->algorithm;
1968         sector_t stripe;
1969         int chunk_offset;
1970         sector_t chunk_number;
1971         int dummy1, dd_idx = i;
1972         sector_t r_sector;
1973         struct stripe_head sh2;
1974
1975
1976         chunk_offset = sector_div(new_sector, sectors_per_chunk);
1977         stripe = new_sector;
1978
1979         if (i == sh->pd_idx)
1980                 return 0;
1981         switch(conf->level) {
1982         case 4: break;
1983         case 5:
1984                 switch (algorithm) {
1985                 case ALGORITHM_LEFT_ASYMMETRIC:
1986                 case ALGORITHM_RIGHT_ASYMMETRIC:
1987                         if (i > sh->pd_idx)
1988                                 i--;
1989                         break;
1990                 case ALGORITHM_LEFT_SYMMETRIC:
1991                 case ALGORITHM_RIGHT_SYMMETRIC:
1992                         if (i < sh->pd_idx)
1993                                 i += raid_disks;
1994                         i -= (sh->pd_idx + 1);
1995                         break;
1996                 case ALGORITHM_PARITY_0:
1997                         i -= 1;
1998                         break;
1999                 case ALGORITHM_PARITY_N:
2000                         break;
2001                 default:
2002                         BUG();
2003                 }
2004                 break;
2005         case 6:
2006                 if (i == sh->qd_idx)
2007                         return 0; /* It is the Q disk */
2008                 switch (algorithm) {
2009                 case ALGORITHM_LEFT_ASYMMETRIC:
2010                 case ALGORITHM_RIGHT_ASYMMETRIC:
2011                 case ALGORITHM_ROTATING_ZERO_RESTART:
2012                 case ALGORITHM_ROTATING_N_RESTART:
2013                         if (sh->pd_idx == raid_disks-1)
2014                                 i--;    /* Q D D D P */
2015                         else if (i > sh->pd_idx)
2016                                 i -= 2; /* D D P Q D */
2017                         break;
2018                 case ALGORITHM_LEFT_SYMMETRIC:
2019                 case ALGORITHM_RIGHT_SYMMETRIC:
2020                         if (sh->pd_idx == raid_disks-1)
2021                                 i--; /* Q D D D P */
2022                         else {
2023                                 /* D D P Q D */
2024                                 if (i < sh->pd_idx)
2025                                         i += raid_disks;
2026                                 i -= (sh->pd_idx + 2);
2027                         }
2028                         break;
2029                 case ALGORITHM_PARITY_0:
2030                         i -= 2;
2031                         break;
2032                 case ALGORITHM_PARITY_N:
2033                         break;
2034                 case ALGORITHM_ROTATING_N_CONTINUE:
2035                         /* Like left_symmetric, but P is before Q */
2036                         if (sh->pd_idx == 0)
2037                                 i--;    /* P D D D Q */
2038                         else {
2039                                 /* D D Q P D */
2040                                 if (i < sh->pd_idx)
2041                                         i += raid_disks;
2042                                 i -= (sh->pd_idx + 1);
2043                         }
2044                         break;
2045                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2046                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2047                         if (i > sh->pd_idx)
2048                                 i--;
2049                         break;
2050                 case ALGORITHM_LEFT_SYMMETRIC_6:
2051                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2052                         if (i < sh->pd_idx)
2053                                 i += data_disks + 1;
2054                         i -= (sh->pd_idx + 1);
2055                         break;
2056                 case ALGORITHM_PARITY_0_6:
2057                         i -= 1;
2058                         break;
2059                 default:
2060                         BUG();
2061                 }
2062                 break;
2063         }
2064
2065         chunk_number = stripe * data_disks + i;
2066         r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2067
2068         check = raid5_compute_sector(conf, r_sector,
2069                                      previous, &dummy1, &sh2);
2070         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2071                 || sh2.qd_idx != sh->qd_idx) {
2072                 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2073                        mdname(conf->mddev));
2074                 return 0;
2075         }
2076         return r_sector;
2077 }
2078
2079
2080 static void
2081 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2082                          int rcw, int expand)
2083 {
2084         int i, pd_idx = sh->pd_idx, disks = sh->disks;
2085         struct r5conf *conf = sh->raid_conf;
2086         int level = conf->level;
2087
2088         if (rcw) {
2089                 /* if we are not expanding this is a proper write request, and
2090                  * there will be bios with new data to be drained into the
2091                  * stripe cache
2092                  */
2093                 if (!expand) {
2094                         sh->reconstruct_state = reconstruct_state_drain_run;
2095                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2096                 } else
2097                         sh->reconstruct_state = reconstruct_state_run;
2098
2099                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2100
2101                 for (i = disks; i--; ) {
2102                         struct r5dev *dev = &sh->dev[i];
2103
2104                         if (dev->towrite) {
2105                                 set_bit(R5_LOCKED, &dev->flags);
2106                                 set_bit(R5_Wantdrain, &dev->flags);
2107                                 if (!expand)
2108                                         clear_bit(R5_UPTODATE, &dev->flags);
2109                                 s->locked++;
2110                         }
2111                 }
2112                 if (s->locked + conf->max_degraded == disks)
2113                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2114                                 atomic_inc(&conf->pending_full_writes);
2115         } else {
2116                 BUG_ON(level == 6);
2117                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2118                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2119
2120                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2121                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2122                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2123                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2124
2125                 for (i = disks; i--; ) {
2126                         struct r5dev *dev = &sh->dev[i];
2127                         if (i == pd_idx)
2128                                 continue;
2129
2130                         if (dev->towrite &&
2131                             (test_bit(R5_UPTODATE, &dev->flags) ||
2132                              test_bit(R5_Wantcompute, &dev->flags))) {
2133                                 set_bit(R5_Wantdrain, &dev->flags);
2134                                 set_bit(R5_LOCKED, &dev->flags);
2135                                 clear_bit(R5_UPTODATE, &dev->flags);
2136                                 s->locked++;
2137                         }
2138                 }
2139         }
2140
2141         /* keep the parity disk(s) locked while asynchronous operations
2142          * are in flight
2143          */
2144         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2145         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2146         s->locked++;
2147
2148         if (level == 6) {
2149                 int qd_idx = sh->qd_idx;
2150                 struct r5dev *dev = &sh->dev[qd_idx];
2151
2152                 set_bit(R5_LOCKED, &dev->flags);
2153                 clear_bit(R5_UPTODATE, &dev->flags);
2154                 s->locked++;
2155         }
2156
2157         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2158                 __func__, (unsigned long long)sh->sector,
2159                 s->locked, s->ops_request);
2160 }
2161
2162 /*
2163  * Each stripe/dev can have one or more bion attached.
2164  * toread/towrite point to the first in a chain.
2165  * The bi_next chain must be in order.
2166  */
2167 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2168 {
2169         struct bio **bip;
2170         struct r5conf *conf = sh->raid_conf;
2171         int firstwrite=0;
2172
2173         pr_debug("adding bi b#%llu to stripe s#%llu\n",
2174                 (unsigned long long)bi->bi_sector,
2175                 (unsigned long long)sh->sector);
2176
2177
2178         spin_lock_irq(&conf->device_lock);
2179         if (forwrite) {
2180                 bip = &sh->dev[dd_idx].towrite;
2181                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2182                         firstwrite = 1;
2183         } else
2184                 bip = &sh->dev[dd_idx].toread;
2185         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2186                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2187                         goto overlap;
2188                 bip = & (*bip)->bi_next;
2189         }
2190         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2191                 goto overlap;
2192
2193         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2194         if (*bip)
2195                 bi->bi_next = *bip;
2196         *bip = bi;
2197         bi->bi_phys_segments++;
2198
2199         if (forwrite) {
2200                 /* check if page is covered */
2201                 sector_t sector = sh->dev[dd_idx].sector;
2202                 for (bi=sh->dev[dd_idx].towrite;
2203                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2204                              bi && bi->bi_sector <= sector;
2205                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2206                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2207                                 sector = bi->bi_sector + (bi->bi_size>>9);
2208                 }
2209                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2210                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2211         }
2212         spin_unlock_irq(&conf->device_lock);
2213
2214         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2215                 (unsigned long long)(*bip)->bi_sector,
2216                 (unsigned long long)sh->sector, dd_idx);
2217
2218         if (conf->mddev->bitmap && firstwrite) {
2219                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2220                                   STRIPE_SECTORS, 0);
2221                 sh->bm_seq = conf->seq_flush+1;
2222                 set_bit(STRIPE_BIT_DELAY, &sh->state);
2223         }
2224         return 1;
2225
2226  overlap:
2227         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2228         spin_unlock_irq(&conf->device_lock);
2229         return 0;
2230 }
2231
2232 static void end_reshape(struct r5conf *conf);
2233
2234 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2235                             struct stripe_head *sh)
2236 {
2237         int sectors_per_chunk =
2238                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2239         int dd_idx;
2240         int chunk_offset = sector_div(stripe, sectors_per_chunk);
2241         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2242
2243         raid5_compute_sector(conf,
2244                              stripe * (disks - conf->max_degraded)
2245                              *sectors_per_chunk + chunk_offset,
2246                              previous,
2247                              &dd_idx, sh);
2248 }
2249
2250 static void
2251 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2252                                 struct stripe_head_state *s, int disks,
2253                                 struct bio **return_bi)
2254 {
2255         int i;
2256         for (i = disks; i--; ) {
2257                 struct bio *bi;
2258                 int bitmap_end = 0;
2259
2260                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2261                         struct md_rdev *rdev;
2262                         rcu_read_lock();
2263                         rdev = rcu_dereference(conf->disks[i].rdev);
2264                         if (rdev && test_bit(In_sync, &rdev->flags))
2265                                 atomic_inc(&rdev->nr_pending);
2266                         else
2267                                 rdev = NULL;
2268                         rcu_read_unlock();
2269                         if (rdev) {
2270                                 if (!rdev_set_badblocks(
2271                                             rdev,
2272                                             sh->sector,
2273                                             STRIPE_SECTORS, 0))
2274                                         md_error(conf->mddev, rdev);
2275                                 rdev_dec_pending(rdev, conf->mddev);
2276                         }
2277                 }
2278                 spin_lock_irq(&conf->device_lock);
2279                 /* fail all writes first */
2280                 bi = sh->dev[i].towrite;
2281                 sh->dev[i].towrite = NULL;
2282                 if (bi) {
2283                         s->to_write--;
2284                         bitmap_end = 1;
2285                 }
2286
2287                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2288                         wake_up(&conf->wait_for_overlap);
2289
2290                 while (bi && bi->bi_sector <
2291                         sh->dev[i].sector + STRIPE_SECTORS) {
2292                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2293                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2294                         if (!raid5_dec_bi_phys_segments(bi)) {
2295                                 md_write_end(conf->mddev);
2296                                 bi->bi_next = *return_bi;
2297                                 *return_bi = bi;
2298                         }
2299                         bi = nextbi;
2300                 }
2301                 /* and fail all 'written' */
2302                 bi = sh->dev[i].written;
2303                 sh->dev[i].written = NULL;
2304                 if (bi) bitmap_end = 1;
2305                 while (bi && bi->bi_sector <
2306                        sh->dev[i].sector + STRIPE_SECTORS) {
2307                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2308                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2309                         if (!raid5_dec_bi_phys_segments(bi)) {
2310                                 md_write_end(conf->mddev);
2311                                 bi->bi_next = *return_bi;
2312                                 *return_bi = bi;
2313                         }
2314                         bi = bi2;
2315                 }
2316
2317                 /* fail any reads if this device is non-operational and
2318                  * the data has not reached the cache yet.
2319                  */
2320                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2321                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2322                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
2323                         bi = sh->dev[i].toread;
2324                         sh->dev[i].toread = NULL;
2325                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2326                                 wake_up(&conf->wait_for_overlap);
2327                         if (bi) s->to_read--;
2328                         while (bi && bi->bi_sector <
2329                                sh->dev[i].sector + STRIPE_SECTORS) {
2330                                 struct bio *nextbi =
2331                                         r5_next_bio(bi, sh->dev[i].sector);
2332                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2333                                 if (!raid5_dec_bi_phys_segments(bi)) {
2334                                         bi->bi_next = *return_bi;
2335                                         *return_bi = bi;
2336                                 }
2337                                 bi = nextbi;
2338                         }
2339                 }
2340                 spin_unlock_irq(&conf->device_lock);
2341                 if (bitmap_end)
2342                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2343                                         STRIPE_SECTORS, 0, 0);
2344                 /* If we were in the middle of a write the parity block might
2345                  * still be locked - so just clear all R5_LOCKED flags
2346                  */
2347                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2348         }
2349
2350         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2351                 if (atomic_dec_and_test(&conf->pending_full_writes))
2352                         md_wakeup_thread(conf->mddev->thread);
2353 }
2354
2355 static void
2356 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
2357                    struct stripe_head_state *s)
2358 {
2359         int abort = 0;
2360         int i;
2361
2362         md_done_sync(conf->mddev, STRIPE_SECTORS, 0);
2363         clear_bit(STRIPE_SYNCING, &sh->state);
2364         s->syncing = 0;
2365         /* There is nothing more to do for sync/check/repair.
2366          * For recover we need to record a bad block on all
2367          * non-sync devices, or abort the recovery
2368          */
2369         if (!test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery))
2370                 return;
2371         /* During recovery devices cannot be removed, so locking and
2372          * refcounting of rdevs is not needed
2373          */
2374         for (i = 0; i < conf->raid_disks; i++) {
2375                 struct md_rdev *rdev = conf->disks[i].rdev;
2376                 if (!rdev
2377                     || test_bit(Faulty, &rdev->flags)
2378                     || test_bit(In_sync, &rdev->flags))
2379                         continue;
2380                 if (!rdev_set_badblocks(rdev, sh->sector,
2381                                         STRIPE_SECTORS, 0))
2382                         abort = 1;
2383         }
2384         if (abort) {
2385                 conf->recovery_disabled = conf->mddev->recovery_disabled;
2386                 set_bit(MD_RECOVERY_INTR, &conf->mddev->recovery);
2387         }
2388 }
2389
2390 /* fetch_block - checks the given member device to see if its data needs
2391  * to be read or computed to satisfy a request.
2392  *
2393  * Returns 1 when no more member devices need to be checked, otherwise returns
2394  * 0 to tell the loop in handle_stripe_fill to continue
2395  */
2396 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2397                        int disk_idx, int disks)
2398 {
2399         struct r5dev *dev = &sh->dev[disk_idx];
2400         struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2401                                   &sh->dev[s->failed_num[1]] };
2402
2403         /* is the data in this block needed, and can we get it? */
2404         if (!test_bit(R5_LOCKED, &dev->flags) &&
2405             !test_bit(R5_UPTODATE, &dev->flags) &&
2406             (dev->toread ||
2407              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2408              s->syncing || s->expanding ||
2409              (s->failed >= 1 && fdev[0]->toread) ||
2410              (s->failed >= 2 && fdev[1]->toread) ||
2411              (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2412               !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2413              (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
2414                 /* we would like to get this block, possibly by computing it,
2415                  * otherwise read it if the backing disk is insync
2416                  */
2417                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2418                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2419                 if ((s->uptodate == disks - 1) &&
2420                     (s->failed && (disk_idx == s->failed_num[0] ||
2421                                    disk_idx == s->failed_num[1]))) {
2422                         /* have disk failed, and we're requested to fetch it;
2423                          * do compute it
2424                          */
2425                         pr_debug("Computing stripe %llu block %d\n",
2426                                (unsigned long long)sh->sector, disk_idx);
2427                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2428                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2429                         set_bit(R5_Wantcompute, &dev->flags);
2430                         sh->ops.target = disk_idx;
2431                         sh->ops.target2 = -1; /* no 2nd target */
2432                         s->req_compute = 1;
2433                         /* Careful: from this point on 'uptodate' is in the eye
2434                          * of raid_run_ops which services 'compute' operations
2435                          * before writes. R5_Wantcompute flags a block that will
2436                          * be R5_UPTODATE by the time it is needed for a
2437                          * subsequent operation.
2438                          */
2439                         s->uptodate++;
2440                         return 1;
2441                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2442                         /* Computing 2-failure is *very* expensive; only
2443                          * do it if failed >= 2
2444                          */
2445                         int other;
2446                         for (other = disks; other--; ) {
2447                                 if (other == disk_idx)
2448                                         continue;
2449                                 if (!test_bit(R5_UPTODATE,
2450                                       &sh->dev[other].flags))
2451                                         break;
2452                         }
2453                         BUG_ON(other < 0);
2454                         pr_debug("Computing stripe %llu blocks %d,%d\n",
2455                                (unsigned long long)sh->sector,
2456                                disk_idx, other);
2457                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2458                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2459                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2460                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
2461                         sh->ops.target = disk_idx;
2462                         sh->ops.target2 = other;
2463                         s->uptodate += 2;
2464                         s->req_compute = 1;
2465                         return 1;
2466                 } else if (test_bit(R5_Insync, &dev->flags)) {
2467                         set_bit(R5_LOCKED, &dev->flags);
2468                         set_bit(R5_Wantread, &dev->flags);
2469                         s->locked++;
2470                         pr_debug("Reading block %d (sync=%d)\n",
2471                                 disk_idx, s->syncing);
2472                 }
2473         }
2474
2475         return 0;
2476 }
2477
2478 /**
2479  * handle_stripe_fill - read or compute data to satisfy pending requests.
2480  */
2481 static void handle_stripe_fill(struct stripe_head *sh,
2482                                struct stripe_head_state *s,
2483                                int disks)
2484 {
2485         int i;
2486
2487         /* look for blocks to read/compute, skip this if a compute
2488          * is already in flight, or if the stripe contents are in the
2489          * midst of changing due to a write
2490          */
2491         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2492             !sh->reconstruct_state)
2493                 for (i = disks; i--; )
2494                         if (fetch_block(sh, s, i, disks))
2495                                 break;
2496         set_bit(STRIPE_HANDLE, &sh->state);
2497 }
2498
2499
2500 /* handle_stripe_clean_event
2501  * any written block on an uptodate or failed drive can be returned.
2502  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2503  * never LOCKED, so we don't need to test 'failed' directly.
2504  */
2505 static void handle_stripe_clean_event(struct r5conf *conf,
2506         struct stripe_head *sh, int disks, struct bio **return_bi)
2507 {
2508         int i;
2509         struct r5dev *dev;
2510
2511         for (i = disks; i--; )
2512                 if (sh->dev[i].written) {
2513                         dev = &sh->dev[i];
2514                         if (!test_bit(R5_LOCKED, &dev->flags) &&
2515                                 test_bit(R5_UPTODATE, &dev->flags)) {
2516                                 /* We can return any write requests */
2517                                 struct bio *wbi, *wbi2;
2518                                 int bitmap_end = 0;
2519                                 pr_debug("Return write for disc %d\n", i);
2520                                 spin_lock_irq(&conf->device_lock);
2521                                 wbi = dev->written;
2522                                 dev->written = NULL;
2523                                 while (wbi && wbi->bi_sector <
2524                                         dev->sector + STRIPE_SECTORS) {
2525                                         wbi2 = r5_next_bio(wbi, dev->sector);
2526                                         if (!raid5_dec_bi_phys_segments(wbi)) {
2527                                                 md_write_end(conf->mddev);
2528                                                 wbi->bi_next = *return_bi;
2529                                                 *return_bi = wbi;
2530                                         }
2531                                         wbi = wbi2;
2532                                 }
2533                                 if (dev->towrite == NULL)
2534                                         bitmap_end = 1;
2535                                 spin_unlock_irq(&conf->device_lock);
2536                                 if (bitmap_end)
2537                                         bitmap_endwrite(conf->mddev->bitmap,
2538                                                         sh->sector,
2539                                                         STRIPE_SECTORS,
2540                                          !test_bit(STRIPE_DEGRADED, &sh->state),
2541                                                         0);
2542                         }
2543                 }
2544
2545         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2546                 if (atomic_dec_and_test(&conf->pending_full_writes))
2547                         md_wakeup_thread(conf->mddev->thread);
2548 }
2549
2550 static void handle_stripe_dirtying(struct r5conf *conf,
2551                                    struct stripe_head *sh,
2552                                    struct stripe_head_state *s,
2553                                    int disks)
2554 {
2555         int rmw = 0, rcw = 0, i;
2556         if (conf->max_degraded == 2) {
2557                 /* RAID6 requires 'rcw' in current implementation
2558                  * Calculate the real rcw later - for now fake it
2559                  * look like rcw is cheaper
2560                  */
2561                 rcw = 1; rmw = 2;
2562         } else for (i = disks; i--; ) {
2563                 /* would I have to read this buffer for read_modify_write */
2564                 struct r5dev *dev = &sh->dev[i];
2565                 if ((dev->towrite || i == sh->pd_idx) &&
2566                     !test_bit(R5_LOCKED, &dev->flags) &&
2567                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2568                       test_bit(R5_Wantcompute, &dev->flags))) {
2569                         if (test_bit(R5_Insync, &dev->flags))
2570                                 rmw++;
2571                         else
2572                                 rmw += 2*disks;  /* cannot read it */
2573                 }
2574                 /* Would I have to read this buffer for reconstruct_write */
2575                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2576                     !test_bit(R5_LOCKED, &dev->flags) &&
2577                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2578                     test_bit(R5_Wantcompute, &dev->flags))) {
2579                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2580                         else
2581                                 rcw += 2*disks;
2582                 }
2583         }
2584         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2585                 (unsigned long long)sh->sector, rmw, rcw);
2586         set_bit(STRIPE_HANDLE, &sh->state);
2587         if (rmw < rcw && rmw > 0)
2588                 /* prefer read-modify-write, but need to get some data */
2589                 for (i = disks; i--; ) {
2590                         struct r5dev *dev = &sh->dev[i];
2591                         if ((dev->towrite || i == sh->pd_idx) &&
2592                             !test_bit(R5_LOCKED, &dev->flags) &&
2593                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2594                             test_bit(R5_Wantcompute, &dev->flags)) &&
2595                             test_bit(R5_Insync, &dev->flags)) {
2596                                 if (
2597                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2598                                         pr_debug("Read_old block "
2599                                                 "%d for r-m-w\n", i);
2600                                         set_bit(R5_LOCKED, &dev->flags);
2601                                         set_bit(R5_Wantread, &dev->flags);
2602                                         s->locked++;
2603                                 } else {
2604                                         set_bit(STRIPE_DELAYED, &sh->state);
2605                                         set_bit(STRIPE_HANDLE, &sh->state);
2606                                 }
2607                         }
2608                 }
2609         if (rcw <= rmw && rcw > 0) {
2610                 /* want reconstruct write, but need to get some data */
2611                 rcw = 0;
2612                 for (i = disks; i--; ) {
2613                         struct r5dev *dev = &sh->dev[i];
2614                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2615                             i != sh->pd_idx && i != sh->qd_idx &&
2616                             !test_bit(R5_LOCKED, &dev->flags) &&
2617                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2618                               test_bit(R5_Wantcompute, &dev->flags))) {
2619                                 rcw++;
2620                                 if (!test_bit(R5_Insync, &dev->flags))
2621                                         continue; /* it's a failed drive */
2622                                 if (
2623                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2624                                         pr_debug("Read_old block "
2625                                                 "%d for Reconstruct\n", i);
2626                                         set_bit(R5_LOCKED, &dev->flags);
2627                                         set_bit(R5_Wantread, &dev->flags);
2628                                         s->locked++;
2629                                 } else {
2630                                         set_bit(STRIPE_DELAYED, &sh->state);
2631                                         set_bit(STRIPE_HANDLE, &sh->state);
2632                                 }
2633                         }
2634                 }
2635         }
2636         /* now if nothing is locked, and if we have enough data,
2637          * we can start a write request
2638          */
2639         /* since handle_stripe can be called at any time we need to handle the
2640          * case where a compute block operation has been submitted and then a
2641          * subsequent call wants to start a write request.  raid_run_ops only
2642          * handles the case where compute block and reconstruct are requested
2643          * simultaneously.  If this is not the case then new writes need to be
2644          * held off until the compute completes.
2645          */
2646         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2647             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2648             !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2649                 schedule_reconstruction(sh, s, rcw == 0, 0);
2650 }
2651
2652 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
2653                                 struct stripe_head_state *s, int disks)
2654 {
2655         struct r5dev *dev = NULL;
2656
2657         set_bit(STRIPE_HANDLE, &sh->state);
2658
2659         switch (sh->check_state) {
2660         case check_state_idle:
2661                 /* start a new check operation if there are no failures */
2662                 if (s->failed == 0) {
2663                         BUG_ON(s->uptodate != disks);
2664                         sh->check_state = check_state_run;
2665                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
2666                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2667                         s->uptodate--;
2668                         break;
2669                 }
2670                 dev = &sh->dev[s->failed_num[0]];
2671                 /* fall through */
2672         case check_state_compute_result:
2673                 sh->check_state = check_state_idle;
2674                 if (!dev)
2675                         dev = &sh->dev[sh->pd_idx];
2676
2677                 /* check that a write has not made the stripe insync */
2678                 if (test_bit(STRIPE_INSYNC, &sh->state))
2679                         break;
2680
2681                 /* either failed parity check, or recovery is happening */
2682                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2683                 BUG_ON(s->uptodate != disks);
2684
2685                 set_bit(R5_LOCKED, &dev->flags);
2686                 s->locked++;
2687                 set_bit(R5_Wantwrite, &dev->flags);
2688
2689                 clear_bit(STRIPE_DEGRADED, &sh->state);
2690                 set_bit(STRIPE_INSYNC, &sh->state);
2691                 break;
2692         case check_state_run:
2693                 break; /* we will be called again upon completion */
2694         case check_state_check_result:
2695                 sh->check_state = check_state_idle;
2696
2697                 /* if a failure occurred during the check operation, leave
2698                  * STRIPE_INSYNC not set and let the stripe be handled again
2699                  */
2700                 if (s->failed)
2701                         break;
2702
2703                 /* handle a successful check operation, if parity is correct
2704                  * we are done.  Otherwise update the mismatch count and repair
2705                  * parity if !MD_RECOVERY_CHECK
2706                  */
2707                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2708                         /* parity is correct (on disc,
2709                          * not in buffer any more)
2710                          */
2711                         set_bit(STRIPE_INSYNC, &sh->state);
2712                 else {
2713                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2714                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2715                                 /* don't try to repair!! */
2716                                 set_bit(STRIPE_INSYNC, &sh->state);
2717                         else {
2718                                 sh->check_state = check_state_compute_run;
2719                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2720                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2721                                 set_bit(R5_Wantcompute,
2722                                         &sh->dev[sh->pd_idx].flags);
2723                                 sh->ops.target = sh->pd_idx;
2724                                 sh->ops.target2 = -1;
2725                                 s->uptodate++;
2726                         }
2727                 }
2728                 break;
2729         case check_state_compute_run:
2730                 break;
2731         default:
2732                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2733                        __func__, sh->check_state,
2734                        (unsigned long long) sh->sector);
2735                 BUG();
2736         }
2737 }
2738
2739
2740 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
2741                                   struct stripe_head_state *s,
2742                                   int disks)
2743 {
2744         int pd_idx = sh->pd_idx;
2745         int qd_idx = sh->qd_idx;
2746         struct r5dev *dev;
2747
2748         set_bit(STRIPE_HANDLE, &sh->state);
2749
2750         BUG_ON(s->failed > 2);
2751
2752         /* Want to check and possibly repair P and Q.
2753          * However there could be one 'failed' device, in which
2754          * case we can only check one of them, possibly using the
2755          * other to generate missing data
2756          */
2757
2758         switch (sh->check_state) {
2759         case check_state_idle:
2760                 /* start a new check operation if there are < 2 failures */
2761                 if (s->failed == s->q_failed) {
2762                         /* The only possible failed device holds Q, so it
2763                          * makes sense to check P (If anything else were failed,
2764                          * we would have used P to recreate it).
2765                          */
2766                         sh->check_state = check_state_run;
2767                 }
2768                 if (!s->q_failed && s->failed < 2) {
2769                         /* Q is not failed, and we didn't use it to generate
2770                          * anything, so it makes sense to check it
2771                          */
2772                         if (sh->check_state == check_state_run)
2773                                 sh->check_state = check_state_run_pq;
2774                         else
2775                                 sh->check_state = check_state_run_q;
2776                 }
2777
2778                 /* discard potentially stale zero_sum_result */
2779                 sh->ops.zero_sum_result = 0;
2780
2781                 if (sh->check_state == check_state_run) {
2782                         /* async_xor_zero_sum destroys the contents of P */
2783                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2784                         s->uptodate--;
2785                 }
2786                 if (sh->check_state >= check_state_run &&
2787                     sh->check_state <= check_state_run_pq) {
2788                         /* async_syndrome_zero_sum preserves P and Q, so
2789                          * no need to mark them !uptodate here
2790                          */
2791                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
2792                         break;
2793                 }
2794
2795                 /* we have 2-disk failure */
2796                 BUG_ON(s->failed != 2);
2797                 /* fall through */
2798         case check_state_compute_result:
2799                 sh->check_state = check_state_idle;
2800
2801                 /* check that a write has not made the stripe insync */
2802                 if (test_bit(STRIPE_INSYNC, &sh->state))
2803                         break;
2804
2805                 /* now write out any block on a failed drive,
2806                  * or P or Q if they were recomputed
2807                  */
2808                 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2809                 if (s->failed == 2) {
2810                         dev = &sh->dev[s->failed_num[1]];
2811                         s->locked++;
2812                         set_bit(R5_LOCKED, &dev->flags);
2813                         set_bit(R5_Wantwrite, &dev->flags);
2814                 }
2815                 if (s->failed >= 1) {
2816                         dev = &sh->dev[s->failed_num[0]];
2817                         s->locked++;
2818                         set_bit(R5_LOCKED, &dev->flags);
2819                         set_bit(R5_Wantwrite, &dev->flags);
2820                 }
2821                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2822                         dev = &sh->dev[pd_idx];
2823                         s->locked++;
2824                         set_bit(R5_LOCKED, &dev->flags);
2825                         set_bit(R5_Wantwrite, &dev->flags);
2826                 }
2827                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2828                         dev = &sh->dev[qd_idx];
2829                         s->locked++;
2830                         set_bit(R5_LOCKED, &dev->flags);
2831                         set_bit(R5_Wantwrite, &dev->flags);
2832                 }
2833                 clear_bit(STRIPE_DEGRADED, &sh->state);
2834
2835                 set_bit(STRIPE_INSYNC, &sh->state);
2836                 break;
2837         case check_state_run:
2838         case check_state_run_q:
2839         case check_state_run_pq:
2840                 break; /* we will be called again upon completion */
2841         case check_state_check_result:
2842                 sh->check_state = check_state_idle;
2843
2844                 /* handle a successful check operation, if parity is correct
2845                  * we are done.  Otherwise update the mismatch count and repair
2846                  * parity if !MD_RECOVERY_CHECK
2847                  */
2848                 if (sh->ops.zero_sum_result == 0) {
2849                         /* both parities are correct */
2850                         if (!s->failed)
2851                                 set_bit(STRIPE_INSYNC, &sh->state);
2852                         else {
2853                                 /* in contrast to the raid5 case we can validate
2854                                  * parity, but still have a failure to write
2855                                  * back
2856                                  */
2857                                 sh->check_state = check_state_compute_result;
2858                                 /* Returning at this point means that we may go
2859                                  * off and bring p and/or q uptodate again so
2860                                  * we make sure to check zero_sum_result again
2861                                  * to verify if p or q need writeback
2862                                  */
2863                         }
2864                 } else {
2865                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2866                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2867                                 /* don't try to repair!! */
2868                                 set_bit(STRIPE_INSYNC, &sh->state);
2869                         else {
2870                                 int *target = &sh->ops.target;
2871
2872                                 sh->ops.target = -1;
2873                                 sh->ops.target2 = -1;
2874                                 sh->check_state = check_state_compute_run;
2875                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2876                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2877                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2878                                         set_bit(R5_Wantcompute,
2879                                                 &sh->dev[pd_idx].flags);
2880                                         *target = pd_idx;
2881                                         target = &sh->ops.target2;
2882                                         s->uptodate++;
2883                                 }
2884                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2885                                         set_bit(R5_Wantcompute,
2886                                                 &sh->dev[qd_idx].flags);
2887                                         *target = qd_idx;
2888                                         s->uptodate++;
2889                                 }
2890                         }
2891                 }
2892                 break;
2893         case check_state_compute_run:
2894                 break;
2895         default:
2896                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2897                        __func__, sh->check_state,
2898                        (unsigned long long) sh->sector);
2899                 BUG();
2900         }
2901 }
2902
2903 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
2904 {
2905         int i;
2906
2907         /* We have read all the blocks in this stripe and now we need to
2908          * copy some of them into a target stripe for expand.
2909          */
2910         struct dma_async_tx_descriptor *tx = NULL;
2911         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2912         for (i = 0; i < sh->disks; i++)
2913                 if (i != sh->pd_idx && i != sh->qd_idx) {
2914                         int dd_idx, j;
2915                         struct stripe_head *sh2;
2916                         struct async_submit_ctl submit;
2917
2918                         sector_t bn = compute_blocknr(sh, i, 1);
2919                         sector_t s = raid5_compute_sector(conf, bn, 0,
2920                                                           &dd_idx, NULL);
2921                         sh2 = get_active_stripe(conf, s, 0, 1, 1);
2922                         if (sh2 == NULL)
2923                                 /* so far only the early blocks of this stripe
2924                                  * have been requested.  When later blocks
2925                                  * get requested, we will try again
2926                                  */
2927                                 continue;
2928                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2929                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2930                                 /* must have already done this block */
2931                                 release_stripe(sh2);
2932                                 continue;
2933                         }
2934
2935                         /* place all the copies on one channel */
2936                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
2937                         tx = async_memcpy(sh2->dev[dd_idx].page,
2938                                           sh->dev[i].page, 0, 0, STRIPE_SIZE,
2939                                           &submit);
2940
2941                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2942                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2943                         for (j = 0; j < conf->raid_disks; j++)
2944                                 if (j != sh2->pd_idx &&
2945                                     j != sh2->qd_idx &&
2946                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
2947                                         break;
2948                         if (j == conf->raid_disks) {
2949                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2950                                 set_bit(STRIPE_HANDLE, &sh2->state);
2951                         }
2952                         release_stripe(sh2);
2953
2954                 }
2955         /* done submitting copies, wait for them to complete */
2956         if (tx) {
2957                 async_tx_ack(tx);
2958                 dma_wait_for_async_tx(tx);
2959         }
2960 }
2961
2962
2963 /*
2964  * handle_stripe - do things to a stripe.
2965  *
2966  * We lock the stripe and then examine the state of various bits
2967  * to see what needs to be done.
2968  * Possible results:
2969  *    return some read request which now have data
2970  *    return some write requests which are safely on disc
2971  *    schedule a read on some buffers
2972  *    schedule a write of some buffers
2973  *    return confirmation of parity correctness
2974  *
2975  * buffers are taken off read_list or write_list, and bh_cache buffers
2976  * get BH_Lock set before the stripe lock is released.
2977  *
2978  */
2979
2980 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
2981 {
2982         struct r5conf *conf = sh->raid_conf;
2983         int disks = sh->disks;
2984         struct r5dev *dev;
2985         int i;
2986
2987         memset(s, 0, sizeof(*s));
2988
2989         s->syncing = test_bit(STRIPE_SYNCING, &sh->state);
2990         s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2991         s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2992         s->failed_num[0] = -1;
2993         s->failed_num[1] = -1;
2994
2995         /* Now to look around and see what can be done */
2996         rcu_read_lock();
2997         spin_lock_irq(&conf->device_lock);
2998         for (i=disks; i--; ) {
2999                 struct md_rdev *rdev;
3000                 sector_t first_bad;
3001                 int bad_sectors;
3002                 int is_bad = 0;
3003
3004                 dev = &sh->dev[i];
3005
3006                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3007                         i, dev->flags, dev->toread, dev->towrite, dev->written);
3008                 /* maybe we can reply to a read
3009                  *
3010                  * new wantfill requests are only permitted while
3011                  * ops_complete_biofill is guaranteed to be inactive
3012                  */
3013                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3014                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3015                         set_bit(R5_Wantfill, &dev->flags);
3016
3017                 /* now count some things */
3018                 if (test_bit(R5_LOCKED, &dev->flags))
3019                         s->locked++;
3020                 if (test_bit(R5_UPTODATE, &dev->flags))
3021                         s->uptodate++;
3022                 if (test_bit(R5_Wantcompute, &dev->flags)) {
3023                         s->compute++;
3024                         BUG_ON(s->compute > 2);
3025                 }
3026
3027                 if (test_bit(R5_Wantfill, &dev->flags))
3028                         s->to_fill++;
3029                 else if (dev->toread)
3030                         s->to_read++;
3031                 if (dev->towrite) {
3032                         s->to_write++;
3033                         if (!test_bit(R5_OVERWRITE, &dev->flags))
3034                                 s->non_overwrite++;
3035                 }
3036                 if (dev->written)
3037                         s->written++;
3038                 rdev = rcu_dereference(conf->disks[i].rdev);
3039                 if (rdev && test_bit(Faulty, &rdev->flags))
3040                         rdev = NULL;
3041                 if (rdev) {
3042                         is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3043                                              &first_bad, &bad_sectors);
3044                         if (s->blocked_rdev == NULL
3045                             && (test_bit(Blocked, &rdev->flags)
3046                                 || is_bad < 0)) {
3047                                 if (is_bad < 0)
3048                                         set_bit(BlockedBadBlocks,
3049                                                 &rdev->flags);
3050                                 s->blocked_rdev = rdev;
3051                                 atomic_inc(&rdev->nr_pending);
3052                         }
3053                 }
3054                 clear_bit(R5_Insync, &dev->flags);
3055                 if (!rdev)
3056                         /* Not in-sync */;
3057                 else if (is_bad) {
3058                         /* also not in-sync */
3059                         if (!test_bit(WriteErrorSeen, &rdev->flags)) {
3060                                 /* treat as in-sync, but with a read error
3061                                  * which we can now try to correct
3062                                  */
3063                                 set_bit(R5_Insync, &dev->flags);
3064                                 set_bit(R5_ReadError, &dev->flags);
3065                         }
3066                 } else if (test_bit(In_sync, &rdev->flags))
3067                         set_bit(R5_Insync, &dev->flags);
3068                 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3069                         /* in sync if before recovery_offset */
3070                         set_bit(R5_Insync, &dev->flags);
3071                 else if (test_bit(R5_UPTODATE, &dev->flags) &&
3072                          test_bit(R5_Expanded, &dev->flags))
3073                         /* If we've reshaped into here, we assume it is Insync.
3074                          * We will shortly update recovery_offset to make
3075                          * it official.
3076                          */
3077                         set_bit(R5_Insync, &dev->flags);
3078
3079                 if (rdev && test_bit(R5_WriteError, &dev->flags)) {
3080                         clear_bit(R5_Insync, &dev->flags);
3081                         if (!test_bit(Faulty, &rdev->flags)) {
3082                                 s->handle_bad_blocks = 1;
3083                                 atomic_inc(&rdev->nr_pending);
3084                         } else
3085                                 clear_bit(R5_WriteError, &dev->flags);
3086                 }
3087                 if (rdev && test_bit(R5_MadeGood, &dev->flags)) {
3088                         if (!test_bit(Faulty, &rdev->flags)) {
3089                                 s->handle_bad_blocks = 1;
3090                                 atomic_inc(&rdev->nr_pending);
3091                         } else
3092                                 clear_bit(R5_MadeGood, &dev->flags);
3093                 }
3094                 if (!test_bit(R5_Insync, &dev->flags)) {
3095                         /* The ReadError flag will just be confusing now */
3096                         clear_bit(R5_ReadError, &dev->flags);
3097                         clear_bit(R5_ReWrite, &dev->flags);
3098                 }
3099                 if (test_bit(R5_ReadError, &dev->flags))
3100                         clear_bit(R5_Insync, &dev->flags);
3101                 if (!test_bit(R5_Insync, &dev->flags)) {
3102                         if (s->failed < 2)
3103                                 s->failed_num[s->failed] = i;
3104                         s->failed++;
3105                 }
3106         }
3107         spin_unlock_irq(&conf->device_lock);
3108         rcu_read_unlock();
3109 }
3110
3111 static void handle_stripe(struct stripe_head *sh)
3112 {
3113         struct stripe_head_state s;
3114         struct r5conf *conf = sh->raid_conf;
3115         int i;
3116         int prexor;
3117         int disks = sh->disks;
3118         struct r5dev *pdev, *qdev;
3119
3120         clear_bit(STRIPE_HANDLE, &sh->state);
3121         if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
3122                 /* already being handled, ensure it gets handled
3123                  * again when current action finishes */
3124                 set_bit(STRIPE_HANDLE, &sh->state);
3125                 return;
3126         }
3127
3128         if (test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3129                 set_bit(STRIPE_SYNCING, &sh->state);
3130                 clear_bit(STRIPE_INSYNC, &sh->state);
3131         }
3132         clear_bit(STRIPE_DELAYED, &sh->state);
3133
3134         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3135                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3136                (unsigned long long)sh->sector, sh->state,
3137                atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3138                sh->check_state, sh->reconstruct_state);
3139
3140         analyse_stripe(sh, &s);
3141
3142         if (s.handle_bad_blocks) {
3143                 set_bit(STRIPE_HANDLE, &sh->state);
3144                 goto finish;
3145         }
3146
3147         if (unlikely(s.blocked_rdev)) {
3148                 if (s.syncing || s.expanding || s.expanded ||
3149                     s.to_write || s.written) {
3150                         set_bit(STRIPE_HANDLE, &sh->state);
3151                         goto finish;
3152                 }
3153                 /* There is nothing for the blocked_rdev to block */
3154                 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3155                 s.blocked_rdev = NULL;
3156         }
3157
3158         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3159                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3160                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3161         }
3162
3163         pr_debug("locked=%d uptodate=%d to_read=%d"
3164                " to_write=%d failed=%d failed_num=%d,%d\n",
3165                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3166                s.failed_num[0], s.failed_num[1]);
3167         /* check if the array has lost more than max_degraded devices and,
3168          * if so, some requests might need to be failed.
3169          */
3170         if (s.failed > conf->max_degraded) {
3171                 sh->check_state = 0;
3172                 sh->reconstruct_state = 0;
3173                 if (s.to_read+s.to_write+s.written)
3174                         handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3175                 if (s.syncing)
3176                         handle_failed_sync(conf, sh, &s);
3177         }
3178
3179         /*
3180          * might be able to return some write requests if the parity blocks
3181          * are safe, or on a failed drive
3182          */
3183         pdev = &sh->dev[sh->pd_idx];
3184         s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3185                 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3186         qdev = &sh->dev[sh->qd_idx];
3187         s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3188                 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3189                 || conf->level < 6;
3190
3191         if (s.written &&
3192             (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3193                              && !test_bit(R5_LOCKED, &pdev->flags)
3194                              && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3195             (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3196                              && !test_bit(R5_LOCKED, &qdev->flags)
3197                              && test_bit(R5_UPTODATE, &qdev->flags)))))
3198                 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3199
3200         /* Now we might consider reading some blocks, either to check/generate
3201          * parity, or to satisfy requests
3202          * or to load a block that is being partially written.
3203          */
3204         if (s.to_read || s.non_overwrite
3205             || (conf->level == 6 && s.to_write && s.failed)
3206             || (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3207                 handle_stripe_fill(sh, &s, disks);
3208
3209         /* Now we check to see if any write operations have recently
3210          * completed
3211          */
3212         prexor = 0;
3213         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3214                 prexor = 1;
3215         if (sh->reconstruct_state == reconstruct_state_drain_result ||
3216             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3217                 sh->reconstruct_state = reconstruct_state_idle;
3218
3219                 /* All the 'written' buffers and the parity block are ready to
3220                  * be written back to disk
3221                  */
3222                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3223                 BUG_ON(sh->qd_idx >= 0 &&
3224                        !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags));
3225                 for (i = disks; i--; ) {
3226                         struct r5dev *dev = &sh->dev[i];
3227                         if (test_bit(R5_LOCKED, &dev->flags) &&
3228                                 (i == sh->pd_idx || i == sh->qd_idx ||
3229                                  dev->written)) {
3230                                 pr_debug("Writing block %d\n", i);
3231                                 set_bit(R5_Wantwrite, &dev->flags);
3232                                 if (prexor)
3233                                         continue;
3234                                 if (!test_bit(R5_Insync, &dev->flags) ||
3235                                     ((i == sh->pd_idx || i == sh->qd_idx)  &&
3236                                      s.failed == 0))
3237                                         set_bit(STRIPE_INSYNC, &sh->state);
3238                         }
3239                 }
3240                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3241                         s.dec_preread_active = 1;
3242         }
3243
3244         /* Now to consider new write requests and what else, if anything
3245          * should be read.  We do not handle new writes when:
3246          * 1/ A 'write' operation (copy+xor) is already in flight.
3247          * 2/ A 'check' operation is in flight, as it may clobber the parity
3248          *    block.
3249          */
3250         if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3251                 handle_stripe_dirtying(conf, sh, &s, disks);
3252
3253         /* maybe we need to check and possibly fix the parity for this stripe
3254          * Any reads will already have been scheduled, so we just see if enough
3255          * data is available.  The parity check is held off while parity
3256          * dependent operations are in flight.
3257          */
3258         if (sh->check_state ||
3259             (s.syncing && s.locked == 0 &&
3260              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3261              !test_bit(STRIPE_INSYNC, &sh->state))) {
3262                 if (conf->level == 6)
3263                         handle_parity_checks6(conf, sh, &s, disks);
3264                 else
3265                         handle_parity_checks5(conf, sh, &s, disks);
3266         }
3267
3268         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3269                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3270                 clear_bit(STRIPE_SYNCING, &sh->state);
3271         }
3272
3273         /* If the failed drives are just a ReadError, then we might need
3274          * to progress the repair/check process
3275          */
3276         if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3277                 for (i = 0; i < s.failed; i++) {
3278                         struct r5dev *dev = &sh->dev[s.failed_num[i]];
3279                         if (test_bit(R5_ReadError, &dev->flags)
3280                             && !test_bit(R5_LOCKED, &dev->flags)
3281                             && test_bit(R5_UPTODATE, &dev->flags)
3282                                 ) {
3283                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
3284                                         set_bit(R5_Wantwrite, &dev->flags);
3285                                         set_bit(R5_ReWrite, &dev->flags);
3286                                         set_bit(R5_LOCKED, &dev->flags);
3287                                         s.locked++;
3288                                 } else {
3289                                         /* let's read it back */
3290                                         set_bit(R5_Wantread, &dev->flags);
3291                                         set_bit(R5_LOCKED, &dev->flags);
3292                                         s.locked++;
3293                                 }
3294                         }
3295                 }
3296
3297
3298         /* Finish reconstruct operations initiated by the expansion process */
3299         if (sh->reconstruct_state == reconstruct_state_result) {
3300                 struct stripe_head *sh_src
3301                         = get_active_stripe(conf, sh->sector, 1, 1, 1);
3302                 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3303                         /* sh cannot be written until sh_src has been read.
3304                          * so arrange for sh to be delayed a little
3305                          */
3306                         set_bit(STRIPE_DELAYED, &sh->state);
3307                         set_bit(STRIPE_HANDLE, &sh->state);
3308                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3309                                               &sh_src->state))
3310                                 atomic_inc(&conf->preread_active_stripes);
3311                         release_stripe(sh_src);
3312                         goto finish;
3313                 }
3314                 if (sh_src)
3315                         release_stripe(sh_src);
3316
3317                 sh->reconstruct_state = reconstruct_state_idle;
3318                 clear_bit(STRIPE_EXPANDING, &sh->state);
3319                 for (i = conf->raid_disks; i--; ) {
3320                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
3321                         set_bit(R5_LOCKED, &sh->dev[i].flags);
3322                         s.locked++;
3323                 }
3324         }
3325
3326         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3327             !sh->reconstruct_state) {
3328                 /* Need to write out all blocks after computing parity */
3329                 sh->disks = conf->raid_disks;
3330                 stripe_set_idx(sh->sector, conf, 0, sh);
3331                 schedule_reconstruction(sh, &s, 1, 1);
3332         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3333                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3334                 atomic_dec(&conf->reshape_stripes);
3335                 wake_up(&conf->wait_for_overlap);
3336                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3337         }
3338
3339         if (s.expanding && s.locked == 0 &&
3340             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3341                 handle_stripe_expansion(conf, sh);
3342
3343 finish:
3344         /* wait for this device to become unblocked */
3345         if (conf->mddev->external && unlikely(s.blocked_rdev))
3346                 md_wait_for_blocked_rdev(s.blocked_rdev, conf->mddev);
3347
3348         if (s.handle_bad_blocks)
3349                 for (i = disks; i--; ) {
3350                         struct md_rdev *rdev;
3351                         struct r5dev *dev = &sh->dev[i];
3352                         if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
3353                                 /* We own a safe reference to the rdev */
3354                                 rdev = conf->disks[i].rdev;
3355                                 if (!rdev_set_badblocks(rdev, sh->sector,
3356                                                         STRIPE_SECTORS, 0))
3357                                         md_error(conf->mddev, rdev);
3358                                 rdev_dec_pending(rdev, conf->mddev);
3359                         }
3360                         if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
3361                                 rdev = conf->disks[i].rdev;
3362                                 rdev_clear_badblocks(rdev, sh->sector,
3363                                                      STRIPE_SECTORS);
3364                                 rdev_dec_pending(rdev, conf->mddev);
3365                         }
3366                 }
3367
3368         if (s.ops_request)
3369                 raid_run_ops(sh, s.ops_request);
3370
3371         ops_run_io(sh, &s);
3372
3373         if (s.dec_preread_active) {
3374                 /* We delay this until after ops_run_io so that if make_request
3375                  * is waiting on a flush, it won't continue until the writes
3376                  * have actually been submitted.
3377                  */
3378                 atomic_dec(&conf->preread_active_stripes);
3379                 if (atomic_read(&conf->preread_active_stripes) <
3380                     IO_THRESHOLD)
3381                         md_wakeup_thread(conf->mddev->thread);
3382         }
3383
3384         return_io(s.return_bi);
3385
3386         clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
3387 }
3388
3389 static void raid5_activate_delayed(struct r5conf *conf)
3390 {
3391         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3392                 while (!list_empty(&conf->delayed_list)) {
3393                         struct list_head *l = conf->delayed_list.next;
3394                         struct stripe_head *sh;
3395                         sh = list_entry(l, struct stripe_head, lru);
3396                         list_del_init(l);
3397                         clear_bit(STRIPE_DELAYED, &sh->state);
3398                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3399                                 atomic_inc(&conf->preread_active_stripes);
3400                         list_add_tail(&sh->lru, &conf->hold_list);
3401                 }
3402         }
3403 }
3404
3405 static void activate_bit_delay(struct r5conf *conf)
3406 {
3407         /* device_lock is held */
3408         struct list_head head;
3409         list_add(&head, &conf->bitmap_list);
3410         list_del_init(&conf->bitmap_list);
3411         while (!list_empty(&head)) {
3412                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3413                 list_del_init(&sh->lru);
3414                 atomic_inc(&sh->count);
3415                 __release_stripe(conf, sh);
3416         }
3417 }
3418
3419 int md_raid5_congested(struct mddev *mddev, int bits)
3420 {
3421         struct r5conf *conf = mddev->private;
3422
3423         /* No difference between reads and writes.  Just check
3424          * how busy the stripe_cache is
3425          */
3426
3427         if (conf->inactive_blocked)
3428                 return 1;
3429         if (conf->quiesce)
3430                 return 1;
3431         if (list_empty_careful(&conf->inactive_list))
3432                 return 1;
3433
3434         return 0;
3435 }
3436 EXPORT_SYMBOL_GPL(md_raid5_congested);
3437
3438 static int raid5_congested(void *data, int bits)
3439 {
3440         struct mddev *mddev = data;
3441
3442         return mddev_congested(mddev, bits) ||
3443                 md_raid5_congested(mddev, bits);
3444 }
3445
3446 /* We want read requests to align with chunks where possible,
3447  * but write requests don't need to.
3448  */
3449 static int raid5_mergeable_bvec(struct request_queue *q,
3450                                 struct bvec_merge_data *bvm,
3451                                 struct bio_vec *biovec)
3452 {
3453         struct mddev *mddev = q->queuedata;
3454         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3455         int max;
3456         unsigned int chunk_sectors = mddev->chunk_sectors;
3457         unsigned int bio_sectors = bvm->bi_size >> 9;
3458
3459         if ((bvm->bi_rw & 1) == WRITE)
3460                 return biovec->bv_len; /* always allow writes to be mergeable */
3461
3462         if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3463                 chunk_sectors = mddev->new_chunk_sectors;
3464         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3465         if (max < 0) max = 0;
3466         if (max <= biovec->bv_len && bio_sectors == 0)
3467                 return biovec->bv_len;
3468         else
3469                 return max;
3470 }
3471
3472
3473 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
3474 {
3475         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3476         unsigned int chunk_sectors = mddev->chunk_sectors;
3477         unsigned int bio_sectors = bio->bi_size >> 9;
3478
3479         if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3480                 chunk_sectors = mddev->new_chunk_sectors;
3481         return  chunk_sectors >=
3482                 ((sector & (chunk_sectors - 1)) + bio_sectors);
3483 }
3484
3485 /*
3486  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3487  *  later sampled by raid5d.
3488  */
3489 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
3490 {
3491         unsigned long flags;
3492
3493         spin_lock_irqsave(&conf->device_lock, flags);
3494
3495         bi->bi_next = conf->retry_read_aligned_list;
3496         conf->retry_read_aligned_list = bi;
3497
3498         spin_unlock_irqrestore(&conf->device_lock, flags);
3499         md_wakeup_thread(conf->mddev->thread);
3500 }
3501
3502
3503 static struct bio *remove_bio_from_retry(struct r5conf *conf)
3504 {
3505         struct bio *bi;
3506
3507         bi = conf->retry_read_aligned;
3508         if (bi) {
3509                 conf->retry_read_aligned = NULL;
3510                 return bi;
3511         }
3512         bi = conf->retry_read_aligned_list;
3513         if(bi) {
3514                 conf->retry_read_aligned_list = bi->bi_next;
3515                 bi->bi_next = NULL;
3516                 /*
3517                  * this sets the active strip count to 1 and the processed
3518                  * strip count to zero (upper 8 bits)
3519                  */
3520                 bi->bi_phys_segments = 1; /* biased count of active stripes */
3521         }
3522
3523         return bi;
3524 }
3525
3526
3527 /*
3528  *  The "raid5_align_endio" should check if the read succeeded and if it
3529  *  did, call bio_endio on the original bio (having bio_put the new bio
3530  *  first).
3531  *  If the read failed..
3532  */
3533 static void raid5_align_endio(struct bio *bi, int error)
3534 {
3535         struct bio* raid_bi  = bi->bi_private;
3536         struct mddev *mddev;
3537         struct r5conf *conf;
3538         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3539         struct md_rdev *rdev;
3540
3541         bio_put(bi);
3542
3543         rdev = (void*)raid_bi->bi_next;
3544         raid_bi->bi_next = NULL;
3545         mddev = rdev->mddev;
3546         conf = mddev->private;
3547
3548         rdev_dec_pending(rdev, conf->mddev);
3549
3550         if (!error && uptodate) {
3551                 bio_endio(raid_bi, 0);
3552                 if (atomic_dec_and_test(&conf->active_aligned_reads))
3553                         wake_up(&conf->wait_for_stripe);
3554                 return;
3555         }
3556
3557
3558         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3559
3560         add_bio_to_retry(raid_bi, conf);
3561 }
3562
3563 static int bio_fits_rdev(struct bio *bi)
3564 {
3565         struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3566
3567         if ((bi->bi_size>>9) > queue_max_sectors(q))
3568                 return 0;
3569         blk_recount_segments(q, bi);
3570         if (bi->bi_phys_segments > queue_max_segments(q))
3571                 return 0;
3572
3573         if (q->merge_bvec_fn)
3574                 /* it's too hard to apply the merge_bvec_fn at this stage,
3575                  * just just give up
3576                  */
3577                 return 0;
3578
3579         return 1;
3580 }
3581
3582
3583 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
3584 {
3585         struct r5conf *conf = mddev->private;
3586         int dd_idx;
3587         struct bio* align_bi;
3588         struct md_rdev *rdev;
3589
3590         if (!in_chunk_boundary(mddev, raid_bio)) {
3591                 pr_debug("chunk_aligned_read : non aligned\n");
3592                 return 0;
3593         }
3594         /*
3595          * use bio_clone_mddev to make a copy of the bio
3596          */
3597         align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
3598         if (!align_bi)
3599                 return 0;
3600         /*
3601          *   set bi_end_io to a new function, and set bi_private to the
3602          *     original bio.
3603          */
3604         align_bi->bi_end_io  = raid5_align_endio;
3605         align_bi->bi_private = raid_bio;
3606         /*
3607          *      compute position
3608          */
3609         align_bi->bi_sector =  raid5_compute_sector(conf, raid_bio->bi_sector,
3610                                                     0,
3611                                                     &dd_idx, NULL);
3612
3613         rcu_read_lock();
3614         rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3615         if (rdev && test_bit(In_sync, &rdev->flags)) {
3616                 sector_t first_bad;
3617                 int bad_sectors;
3618
3619                 atomic_inc(&rdev->nr_pending);
3620                 rcu_read_unlock();
3621                 raid_bio->bi_next = (void*)rdev;
3622                 align_bi->bi_bdev =  rdev->bdev;
3623                 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3624                 align_bi->bi_sector += rdev->data_offset;
3625
3626                 if (!bio_fits_rdev(align_bi) ||
3627                     is_badblock(rdev, align_bi->bi_sector, align_bi->bi_size>>9,
3628                                 &first_bad, &bad_sectors)) {
3629                         /* too big in some way, or has a known bad block */
3630                         bio_put(align_bi);
3631                         rdev_dec_pending(rdev, mddev);
3632                         return 0;
3633                 }
3634
3635                 spin_lock_irq(&conf->device_lock);
3636                 wait_event_lock_irq(conf->wait_for_stripe,
3637                                     conf->quiesce == 0,
3638                                     conf->device_lock, /* nothing */);
3639                 atomic_inc(&conf->active_aligned_reads);
3640                 spin_unlock_irq(&conf->device_lock);
3641
3642                 generic_make_request(align_bi);
3643                 return 1;
3644         } else {
3645                 rcu_read_unlock();
3646                 bio_put(align_bi);
3647                 return 0;
3648         }
3649 }
3650
3651 /* __get_priority_stripe - get the next stripe to process
3652  *
3653  * Full stripe writes are allowed to pass preread active stripes up until
3654  * the bypass_threshold is exceeded.  In general the bypass_count
3655  * increments when the handle_list is handled before the hold_list; however, it
3656  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3657  * stripe with in flight i/o.  The bypass_count will be reset when the
3658  * head of the hold_list has changed, i.e. the head was promoted to the
3659  * handle_list.
3660  */
3661 static struct stripe_head *__get_priority_stripe(struct r5conf *conf)
3662 {
3663         struct stripe_head *sh;
3664
3665         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3666                   __func__,
3667                   list_empty(&conf->handle_list) ? "empty" : "busy",
3668                   list_empty(&conf->hold_list) ? "empty" : "busy",
3669                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
3670
3671         if (!list_empty(&conf->handle_list)) {
3672                 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3673
3674                 if (list_empty(&conf->hold_list))
3675                         conf->bypass_count = 0;
3676                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3677                         if (conf->hold_list.next == conf->last_hold)
3678                                 conf->bypass_count++;
3679                         else {
3680                                 conf->last_hold = conf->hold_list.next;
3681                                 conf->bypass_count -= conf->bypass_threshold;
3682                                 if (conf->bypass_count < 0)
3683                                         conf->bypass_count = 0;
3684                         }
3685                 }
3686         } else if (!list_empty(&conf->hold_list) &&
3687                    ((conf->bypass_threshold &&
3688                      conf->bypass_count > conf->bypass_threshold) ||
3689                     atomic_read(&conf->pending_full_writes) == 0)) {
3690                 sh = list_entry(conf->hold_list.next,
3691                                 typeof(*sh), lru);
3692                 conf->bypass_count -= conf->bypass_threshold;
3693                 if (conf->bypass_count < 0)
3694                         conf->bypass_count = 0;
3695         } else
3696                 return NULL;
3697
3698         list_del_init(&sh->lru);
3699         atomic_inc(&sh->count);
3700         BUG_ON(atomic_read(&sh->count) != 1);
3701         return sh;
3702 }
3703
3704 static void make_request(struct mddev *mddev, struct bio * bi)
3705 {
3706         struct r5conf *conf = mddev->private;
3707         int dd_idx;
3708         sector_t new_sector;
3709         sector_t logical_sector, last_sector;
3710         struct stripe_head *sh;
3711         const int rw = bio_data_dir(bi);
3712         int remaining;
3713         int plugged;
3714
3715         if (unlikely(bi->bi_rw & REQ_FLUSH)) {
3716                 md_flush_request(mddev, bi);
3717                 return;
3718         }
3719
3720         md_write_start(mddev, bi);
3721
3722         if (rw == READ &&
3723              mddev->reshape_position == MaxSector &&
3724              chunk_aligned_read(mddev,bi))
3725                 return;
3726
3727         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3728         last_sector = bi->bi_sector + (bi->bi_size>>9);
3729         bi->bi_next = NULL;
3730         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
3731
3732         plugged = mddev_check_plugged(mddev);
3733         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3734                 DEFINE_WAIT(w);
3735                 int disks, data_disks;
3736                 int previous;
3737
3738         retry:
3739                 previous = 0;
3740                 disks = conf->raid_disks;
3741                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3742                 if (unlikely(conf->reshape_progress != MaxSector)) {
3743                         /* spinlock is needed as reshape_progress may be
3744                          * 64bit on a 32bit platform, and so it might be
3745                          * possible to see a half-updated value
3746                          * Of course reshape_progress could change after
3747                          * the lock is dropped, so once we get a reference
3748                          * to the stripe that we think it is, we will have
3749                          * to check again.
3750                          */
3751                         spin_lock_irq(&conf->device_lock);
3752                         if (mddev->delta_disks < 0
3753                             ? logical_sector < conf->reshape_progress
3754                             : logical_sector >= conf->reshape_progress) {
3755                                 disks = conf->previous_raid_disks;
3756                                 previous = 1;
3757                         } else {
3758                                 if (mddev->delta_disks < 0
3759                                     ? logical_sector < conf->reshape_safe
3760                                     : logical_sector >= conf->reshape_safe) {
3761                                         spin_unlock_irq(&conf->device_lock);
3762                                         schedule();
3763                                         goto retry;
3764                                 }
3765                         }
3766                         spin_unlock_irq(&conf->device_lock);
3767                 }
3768                 data_disks = disks - conf->max_degraded;
3769
3770                 new_sector = raid5_compute_sector(conf, logical_sector,
3771                                                   previous,
3772                                                   &dd_idx, NULL);
3773                 pr_debug("raid456: make_request, sector %llu logical %llu\n",
3774                         (unsigned long long)new_sector, 
3775                         (unsigned long long)logical_sector);
3776
3777                 sh = get_active_stripe(conf, new_sector, previous,
3778                                        (bi->bi_rw&RWA_MASK), 0);
3779                 if (sh) {
3780                         if (unlikely(previous)) {
3781                                 /* expansion might have moved on while waiting for a
3782                                  * stripe, so we must do the range check again.
3783                                  * Expansion could still move past after this
3784                                  * test, but as we are holding a reference to
3785                                  * 'sh', we know that if that happens,
3786                                  *  STRIPE_EXPANDING will get set and the expansion
3787                                  * won't proceed until we finish with the stripe.
3788                                  */
3789                                 int must_retry = 0;
3790                                 spin_lock_irq(&conf->device_lock);
3791                                 if (mddev->delta_disks < 0
3792                                     ? logical_sector >= conf->reshape_progress
3793                                     : logical_sector < conf->reshape_progress)
3794                                         /* mismatch, need to try again */
3795                                         must_retry = 1;
3796                                 spin_unlock_irq(&conf->device_lock);
3797                                 if (must_retry) {
3798                                         release_stripe(sh);
3799                                         schedule();
3800                                         goto retry;
3801                                 }
3802                         }
3803
3804                         if (rw == WRITE &&
3805                             logical_sector >= mddev->suspend_lo &&
3806                             logical_sector < mddev->suspend_hi) {
3807                                 release_stripe(sh);
3808                                 /* As the suspend_* range is controlled by
3809                                  * userspace, we want an interruptible
3810                                  * wait.
3811                                  */
3812                                 flush_signals(current);
3813                                 prepare_to_wait(&conf->wait_for_overlap,
3814                                                 &w, TASK_INTERRUPTIBLE);
3815                                 if (logical_sector >= mddev->suspend_lo &&
3816                                     logical_sector < mddev->suspend_hi)
3817                                         schedule();
3818                                 goto retry;
3819                         }
3820
3821                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3822                             !add_stripe_bio(sh, bi, dd_idx, rw)) {
3823                                 /* Stripe is busy expanding or
3824                                  * add failed due to overlap.  Flush everything
3825                                  * and wait a while
3826                                  */
3827                                 md_wakeup_thread(mddev->thread);
3828                                 release_stripe(sh);
3829                                 schedule();
3830                                 goto retry;
3831                         }
3832                         finish_wait(&conf->wait_for_overlap, &w);
3833                         set_bit(STRIPE_HANDLE, &sh->state);
3834                         clear_bit(STRIPE_DELAYED, &sh->state);
3835                         if ((bi->bi_rw & REQ_SYNC) &&
3836                             !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3837                                 atomic_inc(&conf->preread_active_stripes);
3838                         release_stripe(sh);
3839                 } else {
3840                         /* cannot get stripe for read-ahead, just give-up */
3841                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
3842                         finish_wait(&conf->wait_for_overlap, &w);
3843                         break;
3844                 }
3845                         
3846         }
3847         if (!plugged)
3848                 md_wakeup_thread(mddev->thread);
3849
3850         spin_lock_irq(&conf->device_lock);
3851         remaining = raid5_dec_bi_phys_segments(bi);
3852         spin_unlock_irq(&conf->device_lock);
3853         if (remaining == 0) {
3854
3855                 if ( rw == WRITE )
3856                         md_write_end(mddev);
3857
3858                 bio_endio(bi, 0);
3859         }
3860 }
3861
3862 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
3863
3864 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
3865 {
3866         /* reshaping is quite different to recovery/resync so it is
3867          * handled quite separately ... here.
3868          *
3869          * On each call to sync_request, we gather one chunk worth of
3870          * destination stripes and flag them as expanding.
3871          * Then we find all the source stripes and request reads.
3872          * As the reads complete, handle_stripe will copy the data
3873          * into the destination stripe and release that stripe.
3874          */
3875         struct r5conf *conf = mddev->private;
3876         struct stripe_head *sh;
3877         sector_t first_sector, last_sector;
3878         int raid_disks = conf->previous_raid_disks;
3879         int data_disks = raid_disks - conf->max_degraded;
3880         int new_data_disks = conf->raid_disks - conf->max_degraded;
3881         int i;
3882         int dd_idx;
3883         sector_t writepos, readpos, safepos;
3884         sector_t stripe_addr;
3885         int reshape_sectors;
3886         struct list_head stripes;
3887
3888         if (sector_nr == 0) {
3889                 /* If restarting in the middle, skip the initial sectors */
3890                 if (mddev->delta_disks < 0 &&
3891                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
3892                         sector_nr = raid5_size(mddev, 0, 0)
3893                                 - conf->reshape_progress;
3894                 } else if (mddev->delta_disks >= 0 &&
3895                            conf->reshape_progress > 0)
3896                         sector_nr = conf->reshape_progress;
3897                 sector_div(sector_nr, new_data_disks);
3898                 if (sector_nr) {
3899                         mddev->curr_resync_completed = sector_nr;
3900                         sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3901                         *skipped = 1;
3902                         return sector_nr;
3903                 }
3904         }
3905
3906         /* We need to process a full chunk at a time.
3907          * If old and new chunk sizes differ, we need to process the
3908          * largest of these
3909          */
3910         if (mddev->new_chunk_sectors > mddev->chunk_sectors)
3911                 reshape_sectors = mddev->new_chunk_sectors;
3912         else
3913                 reshape_sectors = mddev->chunk_sectors;
3914
3915         /* we update the metadata when there is more than 3Meg
3916          * in the block range (that is rather arbitrary, should
3917          * probably be time based) or when the data about to be
3918          * copied would over-write the source of the data at
3919          * the front of the range.
3920          * i.e. one new_stripe along from reshape_progress new_maps
3921          * to after where reshape_safe old_maps to
3922          */
3923         writepos = conf->reshape_progress;
3924         sector_div(writepos, new_data_disks);
3925         readpos = conf->reshape_progress;
3926         sector_div(readpos, data_disks);
3927         safepos = conf->reshape_safe;
3928         sector_div(safepos, data_disks);
3929         if (mddev->delta_disks < 0) {
3930                 writepos -= min_t(sector_t, reshape_sectors, writepos);
3931                 readpos += reshape_sectors;
3932                 safepos += reshape_sectors;
3933         } else {
3934                 writepos += reshape_sectors;
3935                 readpos -= min_t(sector_t, reshape_sectors, readpos);
3936                 safepos -= min_t(sector_t, reshape_sectors, safepos);
3937         }
3938
3939         /* 'writepos' is the most advanced device address we might write.
3940          * 'readpos' is the least advanced device address we might read.
3941          * 'safepos' is the least address recorded in the metadata as having
3942          *     been reshaped.
3943          * If 'readpos' is behind 'writepos', then there is no way that we can
3944          * ensure safety in the face of a crash - that must be done by userspace
3945          * making a backup of the data.  So in that case there is no particular
3946          * rush to update metadata.
3947          * Otherwise if 'safepos' is behind 'writepos', then we really need to
3948          * update the metadata to advance 'safepos' to match 'readpos' so that
3949          * we can be safe in the event of a crash.
3950          * So we insist on updating metadata if safepos is behind writepos and
3951          * readpos is beyond writepos.
3952          * In any case, update the metadata every 10 seconds.
3953          * Maybe that number should be configurable, but I'm not sure it is
3954          * worth it.... maybe it could be a multiple of safemode_delay???
3955          */
3956         if ((mddev->delta_disks < 0
3957              ? (safepos > writepos && readpos < writepos)
3958              : (safepos < writepos && readpos > writepos)) ||
3959             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
3960                 /* Cannot proceed until we've updated the superblock... */
3961                 wait_event(conf->wait_for_overlap,
3962                            atomic_read(&conf->reshape_stripes)==0);
3963                 mddev->reshape_position = conf->reshape_progress;
3964                 mddev->curr_resync_completed = sector_nr;
3965                 conf->reshape_checkpoint = jiffies;
3966                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3967                 md_wakeup_thread(mddev->thread);
3968                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3969                            kthread_should_stop());
3970                 spin_lock_irq(&conf->device_lock);
3971                 conf->reshape_safe = mddev->reshape_position;
3972                 spin_unlock_irq(&conf->device_lock);
3973                 wake_up(&conf->wait_for_overlap);
3974                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3975         }
3976
3977         if (mddev->delta_disks < 0) {
3978                 BUG_ON(conf->reshape_progress == 0);
3979                 stripe_addr = writepos;
3980                 BUG_ON((mddev->dev_sectors &
3981                         ~((sector_t)reshape_sectors - 1))
3982                        - reshape_sectors - stripe_addr
3983                        != sector_nr);
3984         } else {
3985                 BUG_ON(writepos != sector_nr + reshape_sectors);
3986                 stripe_addr = sector_nr;
3987         }
3988         INIT_LIST_HEAD(&stripes);
3989         for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
3990                 int j;
3991                 int skipped_disk = 0;
3992                 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
3993                 set_bit(STRIPE_EXPANDING, &sh->state);
3994                 atomic_inc(&conf->reshape_stripes);
3995                 /* If any of this stripe is beyond the end of the old
3996                  * array, then we need to zero those blocks
3997                  */
3998                 for (j=sh->disks; j--;) {
3999                         sector_t s;
4000                         if (j == sh->pd_idx)
4001                                 continue;
4002                         if (conf->level == 6 &&
4003                             j == sh->qd_idx)
4004                                 continue;
4005                         s = compute_blocknr(sh, j, 0);
4006                         if (s < raid5_size(mddev, 0, 0)) {
4007                                 skipped_disk = 1;
4008                                 continue;
4009                         }
4010                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4011                         set_bit(R5_Expanded, &sh->dev[j].flags);
4012                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
4013                 }
4014                 if (!skipped_disk) {
4015                         set_bit(STRIPE_EXPAND_READY, &sh->state);
4016                         set_bit(STRIPE_HANDLE, &sh->state);
4017                 }
4018                 list_add(&sh->lru, &stripes);
4019         }
4020         spin_lock_irq(&conf->device_lock);
4021         if (mddev->delta_disks < 0)
4022                 conf->reshape_progress -= reshape_sectors * new_data_disks;
4023         else
4024                 conf->reshape_progress += reshape_sectors * new_data_disks;
4025         spin_unlock_irq(&conf->device_lock);
4026         /* Ok, those stripe are ready. We can start scheduling
4027          * reads on the source stripes.
4028          * The source stripes are determined by mapping the first and last
4029          * block on the destination stripes.
4030          */
4031         first_sector =
4032                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4033                                      1, &dd_idx, NULL);
4034         last_sector =
4035                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4036                                             * new_data_disks - 1),
4037                                      1, &dd_idx, NULL);
4038         if (last_sector >= mddev->dev_sectors)
4039                 last_sector = mddev->dev_sectors - 1;
4040         while (first_sector <= last_sector) {
4041                 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4042                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4043                 set_bit(STRIPE_HANDLE, &sh->state);
4044                 release_stripe(sh);
4045                 first_sector += STRIPE_SECTORS;
4046         }
4047         /* Now that the sources are clearly marked, we can release
4048          * the destination stripes
4049          */
4050         while (!list_empty(&stripes)) {
4051                 sh = list_entry(stripes.next, struct stripe_head, lru);
4052                 list_del_init(&sh->lru);
4053                 release_stripe(sh);
4054         }
4055         /* If this takes us to the resync_max point where we have to pause,
4056          * then we need to write out the superblock.
4057          */
4058         sector_nr += reshape_sectors;
4059         if ((sector_nr - mddev->curr_resync_completed) * 2
4060             >= mddev->resync_max - mddev->curr_resync_completed) {
4061                 /* Cannot proceed until we've updated the superblock... */
4062                 wait_event(conf->wait_for_overlap,
4063                            atomic_read(&conf->reshape_stripes) == 0);
4064                 mddev->reshape_position = conf->reshape_progress;
4065                 mddev->curr_resync_completed = sector_nr;
4066                 conf->reshape_checkpoint = jiffies;
4067                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4068                 md_wakeup_thread(mddev->thread);
4069                 wait_event(mddev->sb_wait,
4070                            !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4071                            || kthread_should_stop());
4072                 spin_lock_irq(&conf->device_lock);
4073                 conf->reshape_safe = mddev->reshape_position;
4074                 spin_unlock_irq(&conf->device_lock);
4075                 wake_up(&conf->wait_for_overlap);
4076                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4077         }
4078         return reshape_sectors;
4079 }
4080
4081 /* FIXME go_faster isn't used */
4082 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
4083 {
4084         struct r5conf *conf = mddev->private;
4085         struct stripe_head *sh;
4086         sector_t max_sector = mddev->dev_sectors;
4087         sector_t sync_blocks;
4088         int still_degraded = 0;
4089         int i;
4090
4091         if (sector_nr >= max_sector) {
4092                 /* just being told to finish up .. nothing much to do */
4093
4094                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4095                         end_reshape(conf);
4096                         return 0;
4097                 }
4098
4099                 if (mddev->curr_resync < max_sector) /* aborted */
4100                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4101                                         &sync_blocks, 1);
4102                 else /* completed sync */
4103                         conf->fullsync = 0;
4104                 bitmap_close_sync(mddev->bitmap);
4105
4106                 return 0;
4107         }
4108
4109         /* Allow raid5_quiesce to complete */
4110         wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4111
4112         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4113                 return reshape_request(mddev, sector_nr, skipped);
4114
4115         /* No need to check resync_max as we never do more than one
4116          * stripe, and as resync_max will always be on a chunk boundary,
4117          * if the check in md_do_sync didn't fire, there is no chance
4118          * of overstepping resync_max here
4119          */
4120
4121         /* if there is too many failed drives and we are trying
4122          * to resync, then assert that we are finished, because there is
4123          * nothing we can do.
4124          */
4125         if (mddev->degraded >= conf->max_degraded &&
4126             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4127                 sector_t rv = mddev->dev_sectors - sector_nr;
4128                 *skipped = 1;
4129                 return rv;
4130         }
4131         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4132             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4133             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4134                 /* we can skip this block, and probably more */
4135                 sync_blocks /= STRIPE_SECTORS;
4136                 *skipped = 1;
4137                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4138         }
4139
4140
4141         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4142
4143         sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4144         if (sh == NULL) {
4145                 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4146                 /* make sure we don't swamp the stripe cache if someone else
4147                  * is trying to get access
4148                  */
4149                 schedule_timeout_uninterruptible(1);
4150         }
4151         /* Need to check if array will still be degraded after recovery/resync
4152          * We don't need to check the 'failed' flag as when that gets set,
4153          * recovery aborts.
4154          */
4155         for (i = 0; i < conf->raid_disks; i++)
4156                 if (conf->disks[i].rdev == NULL)
4157                         still_degraded = 1;
4158
4159         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4160
4161         set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4162
4163         handle_stripe(sh);
4164         release_stripe(sh);
4165
4166         return STRIPE_SECTORS;
4167 }
4168
4169 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
4170 {
4171         /* We may not be able to submit a whole bio at once as there
4172          * may not be enough stripe_heads available.
4173          * We cannot pre-allocate enough stripe_heads as we may need
4174          * more than exist in the cache (if we allow ever large chunks).
4175          * So we do one stripe head at a time and record in
4176          * ->bi_hw_segments how many have been done.
4177          *
4178          * We *know* that this entire raid_bio is in one chunk, so
4179          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4180          */
4181         struct stripe_head *sh;
4182         int dd_idx;
4183         sector_t sector, logical_sector, last_sector;
4184         int scnt = 0;
4185         int remaining;
4186         int handled = 0;
4187
4188         logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4189         sector = raid5_compute_sector(conf, logical_sector,
4190                                       0, &dd_idx, NULL);
4191         last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4192
4193         for (; logical_sector < last_sector;
4194              logical_sector += STRIPE_SECTORS,
4195                      sector += STRIPE_SECTORS,
4196                      scnt++) {
4197
4198                 if (scnt < raid5_bi_hw_segments(raid_bio))
4199                         /* already done this stripe */
4200                         continue;
4201
4202                 sh = get_active_stripe(conf, sector, 0, 1, 0);
4203
4204                 if (!sh) {
4205                         /* failed to get a stripe - must wait */
4206                         raid5_set_bi_hw_segments(raid_bio, scnt);
4207                         conf->retry_read_aligned = raid_bio;
4208                         return handled;
4209                 }
4210
4211                 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4212                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4213                         release_stripe(sh);
4214                         raid5_set_bi_hw_segments(raid_bio, scnt);
4215                         conf->retry_read_aligned = raid_bio;
4216                         return handled;
4217                 }
4218
4219                 handle_stripe(sh);
4220                 release_stripe(sh);
4221                 handled++;
4222         }
4223         spin_lock_irq(&conf->device_lock);
4224         remaining = raid5_dec_bi_phys_segments(raid_bio);
4225         spin_unlock_irq(&conf->device_lock);
4226         if (remaining == 0)
4227                 bio_endio(raid_bio, 0);
4228         if (atomic_dec_and_test(&conf->active_aligned_reads))
4229                 wake_up(&conf->wait_for_stripe);
4230         return handled;
4231 }
4232
4233
4234 /*
4235  * This is our raid5 kernel thread.
4236  *
4237  * We scan the hash table for stripes which can be handled now.
4238  * During the scan, completed stripes are saved for us by the interrupt
4239  * handler, so that they will not have to wait for our next wakeup.
4240  */
4241 static void raid5d(struct mddev *mddev)
4242 {
4243         struct stripe_head *sh;
4244         struct r5conf *conf = mddev->private;
4245         int handled;
4246         struct blk_plug plug;
4247
4248         pr_debug("+++ raid5d active\n");
4249
4250         md_check_recovery(mddev);
4251
4252         blk_start_plug(&plug);
4253         handled = 0;
4254         spin_lock_irq(&conf->device_lock);
4255         while (1) {
4256                 struct bio *bio;
4257
4258                 if (atomic_read(&mddev->plug_cnt) == 0 &&
4259                     !list_empty(&conf->bitmap_list)) {
4260                         /* Now is a good time to flush some bitmap updates */
4261                         conf->seq_flush++;
4262                         spin_unlock_irq(&conf->device_lock);
4263                         bitmap_unplug(mddev->bitmap);
4264                         spin_lock_irq(&conf->device_lock);
4265                         conf->seq_write = conf->seq_flush;
4266                         activate_bit_delay(conf);
4267                 }
4268                 if (atomic_read(&mddev->plug_cnt) == 0)
4269                         raid5_activate_delayed(conf);
4270
4271                 while ((bio = remove_bio_from_retry(conf))) {
4272                         int ok;
4273                         spin_unlock_irq(&conf->device_lock);
4274                         ok = retry_aligned_read(conf, bio);
4275                         spin_lock_irq(&conf->device_lock);
4276                         if (!ok)
4277                                 break;
4278                         handled++;
4279                 }
4280
4281                 sh = __get_priority_stripe(conf);
4282
4283                 if (!sh)
4284                         break;
4285                 spin_unlock_irq(&conf->device_lock);
4286                 
4287                 handled++;
4288                 handle_stripe(sh);
4289                 release_stripe(sh);
4290                 cond_resched();
4291
4292                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
4293                         md_check_recovery(mddev);
4294
4295                 spin_lock_irq(&conf->device_lock);
4296         }
4297         pr_debug("%d stripes handled\n", handled);
4298
4299         spin_unlock_irq(&conf->device_lock);
4300
4301         async_tx_issue_pending_all();
4302         blk_finish_plug(&plug);
4303
4304         pr_debug("--- raid5d inactive\n");
4305 }
4306
4307 static ssize_t
4308 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
4309 {
4310         struct r5conf *conf = mddev->private;
4311         if (conf)
4312                 return sprintf(page, "%d\n", conf->max_nr_stripes);
4313         else
4314                 return 0;
4315 }
4316
4317 int
4318 raid5_set_cache_size(struct mddev *mddev, int size)
4319 {
4320         struct r5conf *conf = mddev->private;
4321         int err;
4322
4323         if (size <= 16 || size > 32768)
4324                 return -EINVAL;
4325         while (size < conf->max_nr_stripes) {
4326                 if (drop_one_stripe(conf))
4327                         conf->max_nr_stripes--;
4328                 else
4329                         break;
4330         }
4331         err = md_allow_write(mddev);
4332         if (err)
4333                 return err;
4334         while (size > conf->max_nr_stripes) {
4335                 if (grow_one_stripe(conf))
4336                         conf->max_nr_stripes++;
4337                 else break;
4338         }
4339         return 0;
4340 }
4341 EXPORT_SYMBOL(raid5_set_cache_size);
4342
4343 static ssize_t
4344 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
4345 {
4346         struct r5conf *conf = mddev->private;
4347         unsigned long new;
4348         int err;
4349
4350         if (len >= PAGE_SIZE)
4351                 return -EINVAL;
4352         if (!conf)
4353                 return -ENODEV;
4354
4355         if (strict_strtoul(page, 10, &new))
4356                 return -EINVAL;
4357         err = raid5_set_cache_size(mddev, new);
4358         if (err)
4359                 return err;
4360         return len;
4361 }
4362
4363 static struct md_sysfs_entry
4364 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4365                                 raid5_show_stripe_cache_size,
4366                                 raid5_store_stripe_cache_size);
4367
4368 static ssize_t
4369 raid5_show_preread_threshold(struct mddev *mddev, char *page)
4370 {
4371         struct r5conf *conf = mddev->private;
4372         if (conf)
4373                 return sprintf(page, "%d\n", conf->bypass_threshold);
4374         else
4375                 return 0;
4376 }
4377
4378 static ssize_t
4379 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
4380 {
4381         struct r5conf *conf = mddev->private;
4382         unsigned long new;
4383         if (len >= PAGE_SIZE)
4384                 return -EINVAL;
4385         if (!conf)
4386                 return -ENODEV;
4387
4388         if (strict_strtoul(page, 10, &new))
4389                 return -EINVAL;
4390         if (new > conf->max_nr_stripes)
4391                 return -EINVAL;
4392         conf->bypass_threshold = new;
4393         return len;
4394 }
4395
4396 static struct md_sysfs_entry
4397 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4398                                         S_IRUGO | S_IWUSR,
4399                                         raid5_show_preread_threshold,
4400                                         raid5_store_preread_threshold);
4401
4402 static ssize_t
4403 stripe_cache_active_show(struct mddev *mddev, char *page)
4404 {
4405         struct r5conf *conf = mddev->private;
4406         if (conf)
4407                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4408         else
4409                 return 0;
4410 }
4411
4412 static struct md_sysfs_entry
4413 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4414
4415 static struct attribute *raid5_attrs[] =  {
4416         &raid5_stripecache_size.attr,
4417         &raid5_stripecache_active.attr,
4418         &raid5_preread_bypass_threshold.attr,
4419         NULL,
4420 };
4421 static struct attribute_group raid5_attrs_group = {
4422         .name = NULL,
4423         .attrs = raid5_attrs,
4424 };
4425
4426 static sector_t
4427 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
4428 {
4429         struct r5conf *conf = mddev->private;
4430
4431         if (!sectors)
4432                 sectors = mddev->dev_sectors;
4433         if (!raid_disks)
4434                 /* size is defined by the smallest of previous and new size */
4435                 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4436
4437         sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4438         sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4439         return sectors * (raid_disks - conf->max_degraded);
4440 }
4441
4442 static void raid5_free_percpu(struct r5conf *conf)
4443 {
4444         struct raid5_percpu *percpu;
4445         unsigned long cpu;
4446
4447         if (!conf->percpu)
4448                 return;
4449
4450         get_online_cpus();
4451         for_each_possible_cpu(cpu) {
4452                 percpu = per_cpu_ptr(conf->percpu, cpu);
4453                 safe_put_page(percpu->spare_page);
4454                 kfree(percpu->scribble);
4455         }
4456 #ifdef CONFIG_HOTPLUG_CPU
4457         unregister_cpu_notifier(&conf->cpu_notify);
4458 #endif
4459         put_online_cpus();
4460
4461         free_percpu(conf->percpu);
4462 }
4463
4464 static void free_conf(struct r5conf *conf)
4465 {
4466         shrink_stripes(conf);
4467         raid5_free_percpu(conf);
4468         kfree(conf->disks);
4469         kfree(conf->stripe_hashtbl);
4470         kfree(conf);
4471 }
4472
4473 #ifdef CONFIG_HOTPLUG_CPU
4474 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4475                               void *hcpu)
4476 {
4477         struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
4478         long cpu = (long)hcpu;
4479         struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4480
4481         switch (action) {
4482         case CPU_UP_PREPARE:
4483         case CPU_UP_PREPARE_FROZEN:
4484                 if (conf->level == 6 && !percpu->spare_page)
4485                         percpu->spare_page = alloc_page(GFP_KERNEL);
4486                 if (!percpu->scribble)
4487                         percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4488
4489                 if (!percpu->scribble ||
4490                     (conf->level == 6 && !percpu->spare_page)) {
4491                         safe_put_page(percpu->spare_page);
4492                         kfree(percpu->scribble);
4493                         pr_err("%s: failed memory allocation for cpu%ld\n",
4494                                __func__, cpu);
4495                         return notifier_from_errno(-ENOMEM);
4496                 }
4497                 break;
4498         case CPU_DEAD:
4499         case CPU_DEAD_FROZEN:
4500                 safe_put_page(percpu->spare_page);
4501                 kfree(percpu->scribble);
4502                 percpu->spare_page = NULL;
4503                 percpu->scribble = NULL;
4504                 break;
4505         default:
4506                 break;
4507         }
4508         return NOTIFY_OK;
4509 }
4510 #endif
4511
4512 static int raid5_alloc_percpu(struct r5conf *conf)
4513 {
4514         unsigned long cpu;
4515         struct page *spare_page;
4516         struct raid5_percpu __percpu *allcpus;
4517         void *scribble;
4518         int err;
4519
4520         allcpus = alloc_percpu(struct raid5_percpu);
4521         if (!allcpus)
4522                 return -ENOMEM;
4523         conf->percpu = allcpus;
4524
4525         get_online_cpus();
4526         err = 0;
4527         for_each_present_cpu(cpu) {
4528                 if (conf->level == 6) {
4529                         spare_page = alloc_page(GFP_KERNEL);
4530                         if (!spare_page) {
4531                                 err = -ENOMEM;
4532                                 break;
4533                         }
4534                         per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4535                 }
4536                 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4537                 if (!scribble) {
4538                         err = -ENOMEM;
4539                         break;
4540                 }
4541                 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4542         }
4543 #ifdef CONFIG_HOTPLUG_CPU
4544         conf->cpu_notify.notifier_call = raid456_cpu_notify;
4545         conf->cpu_notify.priority = 0;
4546         if (err == 0)
4547                 err = register_cpu_notifier(&conf->cpu_notify);
4548 #endif
4549         put_online_cpus();
4550
4551         return err;
4552 }
4553
4554 static struct r5conf *setup_conf(struct mddev *mddev)
4555 {
4556         struct r5conf *conf;
4557         int raid_disk, memory, max_disks;
4558         struct md_rdev *rdev;
4559         struct disk_info *disk;
4560
4561         if (mddev->new_level != 5
4562             && mddev->new_level != 4
4563             && mddev->new_level != 6) {
4564                 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4565                        mdname(mddev), mddev->new_level);
4566                 return ERR_PTR(-EIO);
4567         }
4568         if ((mddev->new_level == 5
4569              && !algorithm_valid_raid5(mddev->new_layout)) ||
4570             (mddev->new_level == 6
4571              && !algorithm_valid_raid6(mddev->new_layout))) {
4572                 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
4573                        mdname(mddev), mddev->new_layout);
4574                 return ERR_PTR(-EIO);
4575         }
4576         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4577                 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4578                        mdname(mddev), mddev->raid_disks);
4579                 return ERR_PTR(-EINVAL);
4580         }
4581
4582         if (!mddev->new_chunk_sectors ||
4583             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4584             !is_power_of_2(mddev->new_chunk_sectors)) {
4585                 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
4586                        mdname(mddev), mddev->new_chunk_sectors << 9);
4587                 return ERR_PTR(-EINVAL);
4588         }
4589
4590         conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
4591         if (conf == NULL)
4592                 goto abort;
4593         spin_lock_init(&conf->device_lock);
4594         init_waitqueue_head(&conf->wait_for_stripe);
4595         init_waitqueue_head(&conf->wait_for_overlap);
4596         INIT_LIST_HEAD(&conf->handle_list);
4597         INIT_LIST_HEAD(&conf->hold_list);
4598         INIT_LIST_HEAD(&conf->delayed_list);
4599         INIT_LIST_HEAD(&conf->bitmap_list);
4600         INIT_LIST_HEAD(&conf->inactive_list);
4601         atomic_set(&conf->active_stripes, 0);
4602         atomic_set(&conf->preread_active_stripes, 0);
4603         atomic_set(&conf->active_aligned_reads, 0);
4604         conf->bypass_threshold = BYPASS_THRESHOLD;
4605         conf->recovery_disabled = mddev->recovery_disabled - 1;
4606
4607         conf->raid_disks = mddev->raid_disks;
4608         if (mddev->reshape_position == MaxSector)
4609                 conf->previous_raid_disks = mddev->raid_disks;
4610         else
4611                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4612         max_disks = max(conf->raid_disks, conf->previous_raid_disks);
4613         conf->scribble_len = scribble_len(max_disks);
4614
4615         conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
4616                               GFP_KERNEL);
4617         if (!conf->disks)
4618                 goto abort;
4619
4620         conf->mddev = mddev;
4621
4622         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4623                 goto abort;
4624
4625         conf->level = mddev->new_level;
4626         if (raid5_alloc_percpu(conf) != 0)
4627                 goto abort;
4628
4629         pr_debug("raid456: run(%s) called.\n", mdname(mddev));
4630
4631         list_for_each_entry(rdev, &mddev->disks, same_set) {
4632                 raid_disk = rdev->raid_disk;
4633                 if (raid_disk >= max_disks
4634                     || raid_disk < 0)
4635                         continue;
4636                 disk = conf->disks + raid_disk;
4637
4638                 disk->rdev = rdev;
4639
4640                 if (test_bit(In_sync, &rdev->flags)) {
4641                         char b[BDEVNAME_SIZE];
4642                         printk(KERN_INFO "md/raid:%s: device %s operational as raid"
4643                                " disk %d\n",
4644                                mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
4645                 } else if (rdev->saved_raid_disk != raid_disk)
4646                         /* Cannot rely on bitmap to complete recovery */
4647                         conf->fullsync = 1;
4648         }
4649
4650         conf->chunk_sectors = mddev->new_chunk_sectors;
4651         conf->level = mddev->new_level;
4652         if (conf->level == 6)
4653                 conf->max_degraded = 2;
4654         else
4655                 conf->max_degraded = 1;
4656         conf->algorithm = mddev->new_layout;
4657         conf->max_nr_stripes = NR_STRIPES;
4658         conf->reshape_progress = mddev->reshape_position;
4659         if (conf->reshape_progress != MaxSector) {
4660                 conf->prev_chunk_sectors = mddev->chunk_sectors;
4661                 conf->prev_algo = mddev->layout;
4662         }
4663
4664         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4665                  max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4666         if (grow_stripes(conf, conf->max_nr_stripes)) {
4667                 printk(KERN_ERR
4668                        "md/raid:%s: couldn't allocate %dkB for buffers\n",
4669                        mdname(mddev), memory);
4670                 goto abort;
4671         } else
4672                 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
4673                        mdname(mddev), memory);
4674
4675         conf->thread = md_register_thread(raid5d, mddev, NULL);
4676         if (!conf->thread) {
4677                 printk(KERN_ERR
4678                        "md/raid:%s: couldn't allocate thread.\n",
4679                        mdname(mddev));
4680                 goto abort;
4681         }
4682
4683         return conf;
4684
4685  abort:
4686         if (conf) {
4687                 free_conf(conf);
4688                 return ERR_PTR(-EIO);
4689         } else
4690                 return ERR_PTR(-ENOMEM);
4691 }
4692
4693
4694 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
4695 {
4696         switch (algo) {
4697         case ALGORITHM_PARITY_0:
4698                 if (raid_disk < max_degraded)
4699                         return 1;
4700                 break;
4701         case ALGORITHM_PARITY_N:
4702                 if (raid_disk >= raid_disks - max_degraded)
4703                         return 1;
4704                 break;
4705         case ALGORITHM_PARITY_0_6:
4706                 if (raid_disk == 0 || 
4707                     raid_disk == raid_disks - 1)
4708                         return 1;
4709                 break;
4710         case ALGORITHM_LEFT_ASYMMETRIC_6:
4711         case ALGORITHM_RIGHT_ASYMMETRIC_6:
4712         case ALGORITHM_LEFT_SYMMETRIC_6:
4713         case ALGORITHM_RIGHT_SYMMETRIC_6:
4714                 if (raid_disk == raid_disks - 1)
4715                         return 1;
4716         }
4717         return 0;
4718 }
4719
4720 static int run(struct mddev *mddev)
4721 {
4722         struct r5conf *conf;
4723         int working_disks = 0;
4724         int dirty_parity_disks = 0;
4725         struct md_rdev *rdev;
4726         sector_t reshape_offset = 0;
4727
4728         if (mddev->recovery_cp != MaxSector)
4729                 printk(KERN_NOTICE "md/raid:%s: not clean"
4730                        " -- starting background reconstruction\n",
4731                        mdname(mddev));
4732         if (mddev->reshape_position != MaxSector) {
4733                 /* Check that we can continue the reshape.
4734                  * Currently only disks can change, it must
4735                  * increase, and we must be past the point where
4736                  * a stripe over-writes itself
4737                  */
4738                 sector_t here_new, here_old;
4739                 int old_disks;
4740                 int max_degraded = (mddev->level == 6 ? 2 : 1);
4741
4742                 if (mddev->new_level != mddev->level) {
4743                         printk(KERN_ERR "md/raid:%s: unsupported reshape "
4744                                "required - aborting.\n",
4745                                mdname(mddev));
4746                         return -EINVAL;
4747                 }
4748                 old_disks = mddev->raid_disks - mddev->delta_disks;
4749                 /* reshape_position must be on a new-stripe boundary, and one
4750                  * further up in new geometry must map after here in old
4751                  * geometry.
4752                  */
4753                 here_new = mddev->reshape_position;
4754                 if (sector_div(here_new, mddev->new_chunk_sectors *
4755                                (mddev->raid_disks - max_degraded))) {
4756                         printk(KERN_ERR "md/raid:%s: reshape_position not "
4757                                "on a stripe boundary\n", mdname(mddev));
4758                         return -EINVAL;
4759                 }
4760                 reshape_offset = here_new * mddev->new_chunk_sectors;
4761                 /* here_new is the stripe we will write to */
4762                 here_old = mddev->reshape_position;
4763                 sector_div(here_old, mddev->chunk_sectors *
4764                            (old_disks-max_degraded));
4765                 /* here_old is the first stripe that we might need to read
4766                  * from */
4767                 if (mddev->delta_disks == 0) {
4768                         /* We cannot be sure it is safe to start an in-place
4769                          * reshape.  It is only safe if user-space if monitoring
4770                          * and taking constant backups.
4771                          * mdadm always starts a situation like this in
4772                          * readonly mode so it can take control before
4773                          * allowing any writes.  So just check for that.
4774                          */
4775                         if ((here_new * mddev->new_chunk_sectors != 
4776                              here_old * mddev->chunk_sectors) ||
4777                             mddev->ro == 0) {
4778                                 printk(KERN_ERR "md/raid:%s: in-place reshape must be started"
4779                                        " in read-only mode - aborting\n",
4780                                        mdname(mddev));
4781                                 return -EINVAL;
4782                         }
4783                 } else if (mddev->delta_disks < 0
4784                     ? (here_new * mddev->new_chunk_sectors <=
4785                        here_old * mddev->chunk_sectors)
4786                     : (here_new * mddev->new_chunk_sectors >=
4787                        here_old * mddev->chunk_sectors)) {
4788                         /* Reading from the same stripe as writing to - bad */
4789                         printk(KERN_ERR "md/raid:%s: reshape_position too early for "
4790                                "auto-recovery - aborting.\n",
4791                                mdname(mddev));
4792                         return -EINVAL;
4793                 }
4794                 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
4795                        mdname(mddev));
4796                 /* OK, we should be able to continue; */
4797         } else {
4798                 BUG_ON(mddev->level != mddev->new_level);
4799                 BUG_ON(mddev->layout != mddev->new_layout);
4800                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
4801                 BUG_ON(mddev->delta_disks != 0);
4802         }
4803
4804         if (mddev->private == NULL)
4805                 conf = setup_conf(mddev);
4806         else
4807                 conf = mddev->private;
4808
4809         if (IS_ERR(conf))
4810                 return PTR_ERR(conf);
4811
4812         mddev->thread = conf->thread;
4813         conf->thread = NULL;
4814         mddev->private = conf;
4815
4816         /*
4817          * 0 for a fully functional array, 1 or 2 for a degraded array.
4818          */
4819         list_for_each_entry(rdev, &mddev->disks, same_set) {
4820                 if (rdev->raid_disk < 0)
4821                         continue;
4822                 if (test_bit(In_sync, &rdev->flags)) {
4823                         working_disks++;
4824                         continue;
4825                 }
4826                 /* This disc is not fully in-sync.  However if it
4827                  * just stored parity (beyond the recovery_offset),
4828                  * when we don't need to be concerned about the
4829                  * array being dirty.
4830                  * When reshape goes 'backwards', we never have
4831                  * partially completed devices, so we only need
4832                  * to worry about reshape going forwards.
4833                  */
4834                 /* Hack because v0.91 doesn't store recovery_offset properly. */
4835                 if (mddev->major_version == 0 &&
4836                     mddev->minor_version > 90)
4837                         rdev->recovery_offset = reshape_offset;
4838                         
4839                 if (rdev->recovery_offset < reshape_offset) {
4840                         /* We need to check old and new layout */
4841                         if (!only_parity(rdev->raid_disk,
4842                                          conf->algorithm,
4843                                          conf->raid_disks,
4844                                          conf->max_degraded))
4845                                 continue;
4846                 }
4847                 if (!only_parity(rdev->raid_disk,
4848                                  conf->prev_algo,
4849                                  conf->previous_raid_disks,
4850                                  conf->max_degraded))
4851                         continue;
4852                 dirty_parity_disks++;
4853         }
4854
4855         mddev->degraded = (max(conf->raid_disks, conf->previous_raid_disks)
4856                            - working_disks);
4857
4858         if (has_failed(conf)) {
4859                 printk(KERN_ERR "md/raid:%s: not enough operational devices"
4860                         " (%d/%d failed)\n",
4861                         mdname(mddev), mddev->degraded, conf->raid_disks);
4862                 goto abort;
4863         }
4864
4865         /* device size must be a multiple of chunk size */
4866         mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
4867         mddev->resync_max_sectors = mddev->dev_sectors;
4868
4869         if (mddev->degraded > dirty_parity_disks &&
4870             mddev->recovery_cp != MaxSector) {
4871                 if (mddev->ok_start_degraded)
4872                         printk(KERN_WARNING
4873                                "md/raid:%s: starting dirty degraded array"
4874                                " - data corruption possible.\n",
4875                                mdname(mddev));
4876                 else {
4877                         printk(KERN_ERR
4878                                "md/raid:%s: cannot start dirty degraded array.\n",
4879                                mdname(mddev));
4880                         goto abort;
4881                 }
4882         }
4883
4884         if (mddev->degraded == 0)
4885                 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
4886                        " devices, algorithm %d\n", mdname(mddev), conf->level,
4887                        mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4888                        mddev->new_layout);
4889         else
4890                 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
4891                        " out of %d devices, algorithm %d\n",
4892                        mdname(mddev), conf->level,
4893                        mddev->raid_disks - mddev->degraded,
4894                        mddev->raid_disks, mddev->new_layout);
4895
4896         print_raid5_conf(conf);
4897
4898         if (conf->reshape_progress != MaxSector) {
4899                 conf->reshape_safe = conf->reshape_progress;
4900                 atomic_set(&conf->reshape_stripes, 0);
4901                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4902                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4903                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4904                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4905                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4906                                                         "reshape");
4907         }
4908
4909
4910         /* Ok, everything is just fine now */
4911         if (mddev->to_remove == &raid5_attrs_group)
4912                 mddev->to_remove = NULL;
4913         else if (mddev->kobj.sd &&
4914             sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4915                 printk(KERN_WARNING
4916                        "raid5: failed to create sysfs attributes for %s\n",
4917                        mdname(mddev));
4918         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
4919
4920         if (mddev->queue) {
4921                 int chunk_size;
4922                 /* read-ahead size must cover two whole stripes, which
4923                  * is 2 * (datadisks) * chunksize where 'n' is the
4924                  * number of raid devices
4925                  */
4926                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4927                 int stripe = data_disks *
4928                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
4929                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4930                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4931
4932                 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4933
4934                 mddev->queue->backing_dev_info.congested_data = mddev;
4935                 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4936
4937                 chunk_size = mddev->chunk_sectors << 9;
4938                 blk_queue_io_min(mddev->queue, chunk_size);
4939                 blk_queue_io_opt(mddev->queue, chunk_size *
4940                                  (conf->raid_disks - conf->max_degraded));
4941
4942                 list_for_each_entry(rdev, &mddev->disks, same_set)
4943                         disk_stack_limits(mddev->gendisk, rdev->bdev,
4944                                           rdev->data_offset << 9);
4945         }
4946
4947         return 0;
4948 abort:
4949         md_unregister_thread(&mddev->thread);
4950         print_raid5_conf(conf);
4951         free_conf(conf);
4952         mddev->private = NULL;
4953         printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
4954         return -EIO;
4955 }
4956
4957 static int stop(struct mddev *mddev)
4958 {
4959         struct r5conf *conf = mddev->private;
4960
4961         md_unregister_thread(&mddev->thread);
4962         if (mddev->queue)
4963                 mddev->queue->backing_dev_info.congested_fn = NULL;
4964         free_conf(conf);
4965         mddev->private = NULL;
4966         mddev->to_remove = &raid5_attrs_group;
4967         return 0;
4968 }
4969
4970 static void status(struct seq_file *seq, struct mddev *mddev)
4971 {
4972         struct r5conf *conf = mddev->private;
4973         int i;
4974
4975         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
4976                 mddev->chunk_sectors / 2, mddev->layout);
4977         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4978         for (i = 0; i < conf->raid_disks; i++)
4979                 seq_printf (seq, "%s",
4980                                conf->disks[i].rdev &&
4981                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4982         seq_printf (seq, "]");
4983 }
4984
4985 static void print_raid5_conf (struct r5conf *conf)
4986 {
4987         int i;
4988         struct disk_info *tmp;
4989
4990         printk(KERN_DEBUG "RAID conf printout:\n");
4991         if (!conf) {
4992                 printk("(conf==NULL)\n");
4993                 return;
4994         }
4995         printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
4996                conf->raid_disks,
4997                conf->raid_disks - conf->mddev->degraded);
4998
4999         for (i = 0; i < conf->raid_disks; i++) {
5000                 char b[BDEVNAME_SIZE];
5001                 tmp = conf->disks + i;
5002                 if (tmp->rdev)
5003                         printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5004                                i, !test_bit(Faulty, &tmp->rdev->flags),
5005                                bdevname(tmp->rdev->bdev, b));
5006         }
5007 }
5008
5009 static int raid5_spare_active(struct mddev *mddev)
5010 {
5011         int i;
5012         struct r5conf *conf = mddev->private;
5013         struct disk_info *tmp;
5014         int count = 0;
5015         unsigned long flags;
5016
5017         for (i = 0; i < conf->raid_disks; i++) {
5018                 tmp = conf->disks + i;
5019                 if (tmp->rdev
5020                     && tmp->rdev->recovery_offset == MaxSector
5021                     && !test_bit(Faulty, &tmp->rdev->flags)
5022                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5023                         count++;
5024                         sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
5025                 }
5026         }
5027         spin_lock_irqsave(&conf->device_lock, flags);
5028         mddev->degraded -= count;
5029         spin_unlock_irqrestore(&conf->device_lock, flags);
5030         print_raid5_conf(conf);
5031         return count;
5032 }
5033
5034 static int raid5_remove_disk(struct mddev *mddev, int number)
5035 {
5036         struct r5conf *conf = mddev->private;
5037         int err = 0;
5038         struct md_rdev *rdev;
5039         struct disk_info *p = conf->disks + number;
5040
5041         print_raid5_conf(conf);
5042         rdev = p->rdev;
5043         if (rdev) {
5044                 if (number >= conf->raid_disks &&
5045                     conf->reshape_progress == MaxSector)
5046                         clear_bit(In_sync, &rdev->flags);
5047
5048                 if (test_bit(In_sync, &rdev->flags) ||
5049                     atomic_read(&rdev->nr_pending)) {
5050                         err = -EBUSY;
5051                         goto abort;
5052                 }
5053                 /* Only remove non-faulty devices if recovery
5054                  * isn't possible.
5055                  */
5056                 if (!test_bit(Faulty, &rdev->flags) &&
5057                     mddev->recovery_disabled != conf->recovery_disabled &&
5058                     !has_failed(conf) &&
5059                     number < conf->raid_disks) {
5060                         err = -EBUSY;
5061                         goto abort;
5062                 }
5063                 p->rdev = NULL;
5064                 synchronize_rcu();
5065                 if (atomic_read(&rdev->nr_pending)) {
5066                         /* lost the race, try later */
5067                         err = -EBUSY;
5068                         p->rdev = rdev;
5069                 }
5070         }
5071 abort:
5072
5073         print_raid5_conf(conf);
5074         return err;
5075 }
5076
5077 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
5078 {
5079         struct r5conf *conf = mddev->private;
5080         int err = -EEXIST;
5081         int disk;
5082         struct disk_info *p;
5083         int first = 0;
5084         int last = conf->raid_disks - 1;
5085
5086         if (mddev->recovery_disabled == conf->recovery_disabled)
5087                 return -EBUSY;
5088
5089         if (has_failed(conf))
5090                 /* no point adding a device */
5091                 return -EINVAL;
5092
5093         if (rdev->raid_disk >= 0)
5094                 first = last = rdev->raid_disk;
5095
5096         /*
5097          * find the disk ... but prefer rdev->saved_raid_disk
5098          * if possible.
5099          */
5100         if (rdev->saved_raid_disk >= 0 &&
5101             rdev->saved_raid_disk >= first &&
5102             conf->disks[rdev->saved_raid_disk].rdev == NULL)
5103                 disk = rdev->saved_raid_disk;
5104         else
5105                 disk = first;
5106         for ( ; disk <= last ; disk++)
5107                 if ((p=conf->disks + disk)->rdev == NULL) {
5108                         clear_bit(In_sync, &rdev->flags);
5109                         rdev->raid_disk = disk;
5110                         err = 0;
5111                         if (rdev->saved_raid_disk != disk)
5112                                 conf->fullsync = 1;
5113                         rcu_assign_pointer(p->rdev, rdev);
5114                         break;
5115                 }
5116         print_raid5_conf(conf);
5117         return err;
5118 }
5119
5120 static int raid5_resize(struct mddev *mddev, sector_t sectors)
5121 {
5122         /* no resync is happening, and there is enough space
5123          * on all devices, so we can resize.
5124          * We need to make sure resync covers any new space.
5125          * If the array is shrinking we should possibly wait until
5126          * any io in the removed space completes, but it hardly seems
5127          * worth it.
5128          */
5129         sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5130         md_set_array_sectors(mddev, raid5_size(mddev, sectors,
5131                                                mddev->raid_disks));
5132         if (mddev->array_sectors >
5133             raid5_size(mddev, sectors, mddev->raid_disks))
5134                 return -EINVAL;
5135         set_capacity(mddev->gendisk, mddev->array_sectors);
5136         revalidate_disk(mddev->gendisk);
5137         if (sectors > mddev->dev_sectors &&
5138             mddev->recovery_cp > mddev->dev_sectors) {
5139                 mddev->recovery_cp = mddev->dev_sectors;
5140                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5141         }
5142         mddev->dev_sectors = sectors;
5143         mddev->resync_max_sectors = sectors;
5144         return 0;
5145 }
5146
5147 static int check_stripe_cache(struct mddev *mddev)
5148 {
5149         /* Can only proceed if there are plenty of stripe_heads.
5150          * We need a minimum of one full stripe,, and for sensible progress
5151          * it is best to have about 4 times that.
5152          * If we require 4 times, then the default 256 4K stripe_heads will
5153          * allow for chunk sizes up to 256K, which is probably OK.
5154          * If the chunk size is greater, user-space should request more
5155          * stripe_heads first.
5156          */
5157         struct r5conf *conf = mddev->private;
5158         if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5159             > conf->max_nr_stripes ||
5160             ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5161             > conf->max_nr_stripes) {
5162                 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
5163                        mdname(mddev),
5164                        ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5165                         / STRIPE_SIZE)*4);
5166                 return 0;
5167         }
5168         return 1;
5169 }
5170
5171 static int check_reshape(struct mddev *mddev)
5172 {
5173         struct r5conf *conf = mddev->private;
5174
5175         if (mddev->delta_disks == 0 &&
5176             mddev->new_layout == mddev->layout &&
5177             mddev->new_chunk_sectors == mddev->chunk_sectors)
5178                 return 0; /* nothing to do */
5179         if (mddev->bitmap)
5180                 /* Cannot grow a bitmap yet */
5181                 return -EBUSY;
5182         if (has_failed(conf))
5183                 return -EINVAL;
5184         if (mddev->delta_disks < 0) {
5185                 /* We might be able to shrink, but the devices must
5186                  * be made bigger first.
5187                  * For raid6, 4 is the minimum size.
5188                  * Otherwise 2 is the minimum
5189                  */
5190                 int min = 2;
5191                 if (mddev->level == 6)
5192                         min = 4;
5193                 if (mddev->raid_disks + mddev->delta_disks < min)
5194                         return -EINVAL;
5195         }
5196
5197         if (!check_stripe_cache(mddev))
5198                 return -ENOSPC;
5199
5200         return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5201 }
5202
5203 static int raid5_start_reshape(struct mddev *mddev)
5204 {
5205         struct r5conf *conf = mddev->private;
5206         struct md_rdev *rdev;
5207         int spares = 0;
5208         unsigned long flags;
5209
5210         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5211                 return -EBUSY;
5212
5213         if (!check_stripe_cache(mddev))
5214                 return -ENOSPC;
5215
5216         list_for_each_entry(rdev, &mddev->disks, same_set)
5217                 if (!test_bit(In_sync, &rdev->flags)
5218                     && !test_bit(Faulty, &rdev->flags))
5219                         spares++;
5220
5221         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5222                 /* Not enough devices even to make a degraded array
5223                  * of that size
5224                  */
5225                 return -EINVAL;
5226
5227         /* Refuse to reduce size of the array.  Any reductions in
5228          * array size must be through explicit setting of array_size
5229          * attribute.
5230          */
5231         if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5232             < mddev->array_sectors) {
5233                 printk(KERN_ERR "md/raid:%s: array size must be reduced "
5234                        "before number of disks\n", mdname(mddev));
5235                 return -EINVAL;
5236         }
5237
5238         atomic_set(&conf->reshape_stripes, 0);
5239         spin_lock_irq(&conf->device_lock);
5240         conf->previous_raid_disks = conf->raid_disks;
5241         conf->raid_disks += mddev->delta_disks;
5242         conf->prev_chunk_sectors = conf->chunk_sectors;
5243         conf->chunk_sectors = mddev->new_chunk_sectors;
5244         conf->prev_algo = conf->algorithm;
5245         conf->algorithm = mddev->new_layout;
5246         if (mddev->delta_disks < 0)
5247                 conf->reshape_progress = raid5_size(mddev, 0, 0);
5248         else
5249                 conf->reshape_progress = 0;
5250         conf->reshape_safe = conf->reshape_progress;
5251         conf->generation++;
5252         spin_unlock_irq(&conf->device_lock);
5253
5254         /* Add some new drives, as many as will fit.
5255          * We know there are enough to make the newly sized array work.
5256          * Don't add devices if we are reducing the number of
5257          * devices in the array.  This is because it is not possible
5258          * to correctly record the "partially reconstructed" state of
5259          * such devices during the reshape and confusion could result.
5260          */
5261         if (mddev->delta_disks >= 0) {
5262                 int added_devices = 0;
5263                 list_for_each_entry(rdev, &mddev->disks, same_set)
5264                         if (rdev->raid_disk < 0 &&
5265                             !test_bit(Faulty, &rdev->flags)) {
5266                                 if (raid5_add_disk(mddev, rdev) == 0) {
5267                                         if (rdev->raid_disk
5268                                             >= conf->previous_raid_disks) {
5269                                                 set_bit(In_sync, &rdev->flags);
5270                                                 added_devices++;
5271                                         } else
5272                                                 rdev->recovery_offset = 0;
5273
5274                                         if (sysfs_link_rdev(mddev, rdev))
5275                                                 /* Failure here is OK */;
5276                                 }
5277                         } else if (rdev->raid_disk >= conf->previous_raid_disks
5278                                    && !test_bit(Faulty, &rdev->flags)) {
5279                                 /* This is a spare that was manually added */
5280                                 set_bit(In_sync, &rdev->flags);
5281                                 added_devices++;
5282                         }
5283
5284                 /* When a reshape changes the number of devices,
5285                  * ->degraded is measured against the larger of the
5286                  * pre and post number of devices.
5287                  */
5288                 spin_lock_irqsave(&conf->device_lock, flags);
5289                 mddev->degraded += (conf->raid_disks - conf->previous_raid_disks)
5290                         - added_devices;
5291                 spin_unlock_irqrestore(&conf->device_lock, flags);
5292         }
5293         mddev->raid_disks = conf->raid_disks;
5294         mddev->reshape_position = conf->reshape_progress;
5295         set_bit(MD_CHANGE_DEVS, &mddev->flags);
5296
5297         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5298         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5299         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5300         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5301         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5302                                                 "reshape");
5303         if (!mddev->sync_thread) {
5304                 mddev->recovery = 0;
5305                 spin_lock_irq(&conf->device_lock);
5306                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5307                 conf->reshape_progress = MaxSector;
5308                 spin_unlock_irq(&conf->device_lock);
5309                 return -EAGAIN;
5310         }
5311         conf->reshape_checkpoint = jiffies;
5312         md_wakeup_thread(mddev->sync_thread);
5313         md_new_event(mddev);
5314         return 0;
5315 }
5316
5317 /* This is called from the reshape thread and should make any
5318  * changes needed in 'conf'
5319  */
5320 static void end_reshape(struct r5conf *conf)
5321 {
5322
5323         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5324
5325                 spin_lock_irq(&conf->device_lock);
5326                 conf->previous_raid_disks = conf->raid_disks;
5327                 conf->reshape_progress = MaxSector;
5328                 spin_unlock_irq(&conf->device_lock);
5329                 wake_up(&conf->wait_for_overlap);
5330
5331                 /* read-ahead size must cover two whole stripes, which is
5332                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5333                  */
5334                 if (conf->mddev->queue) {
5335                         int data_disks = conf->raid_disks - conf->max_degraded;
5336                         int stripe = data_disks * ((conf->chunk_sectors << 9)
5337                                                    / PAGE_SIZE);
5338                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5339                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5340                 }
5341         }
5342 }
5343
5344 /* This is called from the raid5d thread with mddev_lock held.
5345  * It makes config changes to the device.
5346  */
5347 static void raid5_finish_reshape(struct mddev *mddev)
5348 {
5349         struct r5conf *conf = mddev->private;
5350
5351         if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5352
5353                 if (mddev->delta_disks > 0) {
5354                         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5355                         set_capacity(mddev->gendisk, mddev->array_sectors);
5356                         revalidate_disk(mddev->gendisk);
5357                 } else {
5358                         int d;
5359                         mddev->degraded = conf->raid_disks;
5360                         for (d = 0; d < conf->raid_disks ; d++)
5361                                 if (conf->disks[d].rdev &&
5362                                     test_bit(In_sync,
5363                                              &conf->disks[d].rdev->flags))
5364                                         mddev->degraded--;
5365                         for (d = conf->raid_disks ;
5366                              d < conf->raid_disks - mddev->delta_disks;
5367                              d++) {
5368                                 struct md_rdev *rdev = conf->disks[d].rdev;
5369                                 if (rdev && raid5_remove_disk(mddev, d) == 0) {
5370                                         sysfs_unlink_rdev(mddev, rdev);
5371                                         rdev->raid_disk = -1;
5372                                 }
5373                         }
5374                 }
5375                 mddev->layout = conf->algorithm;
5376                 mddev->chunk_sectors = conf->chunk_sectors;
5377                 mddev->reshape_position = MaxSector;
5378                 mddev->delta_disks = 0;
5379         }
5380 }
5381
5382 static void raid5_quiesce(struct mddev *mddev, int state)
5383 {
5384         struct r5conf *conf = mddev->private;
5385
5386         switch(state) {
5387         case 2: /* resume for a suspend */
5388                 wake_up(&conf->wait_for_overlap);
5389                 break;
5390
5391         case 1: /* stop all writes */
5392                 spin_lock_irq(&conf->device_lock);
5393                 /* '2' tells resync/reshape to pause so that all
5394                  * active stripes can drain
5395                  */
5396                 conf->quiesce = 2;
5397                 wait_event_lock_irq(conf->wait_for_stripe,
5398                                     atomic_read(&conf->active_stripes) == 0 &&
5399                                     atomic_read(&conf->active_aligned_reads) == 0,
5400                                     conf->device_lock, /* nothing */);
5401                 conf->quiesce = 1;
5402                 spin_unlock_irq(&conf->device_lock);
5403                 /* allow reshape to continue */
5404                 wake_up(&conf->wait_for_overlap);
5405                 break;
5406
5407         case 0: /* re-enable writes */
5408                 spin_lock_irq(&conf->device_lock);
5409                 conf->quiesce = 0;
5410                 wake_up(&conf->wait_for_stripe);
5411                 wake_up(&conf->wait_for_overlap);
5412                 spin_unlock_irq(&conf->device_lock);
5413                 break;
5414         }
5415 }
5416
5417
5418 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
5419 {
5420         struct r0conf *raid0_conf = mddev->private;
5421         sector_t sectors;
5422
5423         /* for raid0 takeover only one zone is supported */
5424         if (raid0_conf->nr_strip_zones > 1) {
5425                 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5426                        mdname(mddev));
5427                 return ERR_PTR(-EINVAL);
5428         }
5429
5430         sectors = raid0_conf->strip_zone[0].zone_end;
5431         sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
5432         mddev->dev_sectors = sectors;
5433         mddev->new_level = level;
5434         mddev->new_layout = ALGORITHM_PARITY_N;
5435         mddev->new_chunk_sectors = mddev->chunk_sectors;
5436         mddev->raid_disks += 1;
5437         mddev->delta_disks = 1;
5438         /* make sure it will be not marked as dirty */
5439         mddev->recovery_cp = MaxSector;
5440
5441         return setup_conf(mddev);
5442 }
5443
5444
5445 static void *raid5_takeover_raid1(struct mddev *mddev)
5446 {
5447         int chunksect;
5448
5449         if (mddev->raid_disks != 2 ||
5450             mddev->degraded > 1)
5451                 return ERR_PTR(-EINVAL);
5452
5453         /* Should check if there are write-behind devices? */
5454
5455         chunksect = 64*2; /* 64K by default */
5456
5457         /* The array must be an exact multiple of chunksize */
5458         while (chunksect && (mddev->array_sectors & (chunksect-1)))
5459                 chunksect >>= 1;
5460
5461         if ((chunksect<<9) < STRIPE_SIZE)
5462                 /* array size does not allow a suitable chunk size */
5463                 return ERR_PTR(-EINVAL);
5464
5465         mddev->new_level = 5;
5466         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5467         mddev->new_chunk_sectors = chunksect;
5468
5469         return setup_conf(mddev);
5470 }
5471
5472 static void *raid5_takeover_raid6(struct mddev *mddev)
5473 {
5474         int new_layout;
5475
5476         switch (mddev->layout) {
5477         case ALGORITHM_LEFT_ASYMMETRIC_6:
5478                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5479                 break;
5480         case ALGORITHM_RIGHT_ASYMMETRIC_6:
5481                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5482                 break;
5483         case ALGORITHM_LEFT_SYMMETRIC_6:
5484                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5485                 break;
5486         case ALGORITHM_RIGHT_SYMMETRIC_6:
5487                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5488                 break;
5489         case ALGORITHM_PARITY_0_6:
5490                 new_layout = ALGORITHM_PARITY_0;
5491                 break;
5492         case ALGORITHM_PARITY_N:
5493                 new_layout = ALGORITHM_PARITY_N;
5494                 break;
5495         default:
5496                 return ERR_PTR(-EINVAL);
5497         }
5498         mddev->new_level = 5;
5499         mddev->new_layout = new_layout;
5500         mddev->delta_disks = -1;
5501         mddev->raid_disks -= 1;
5502         return setup_conf(mddev);
5503 }
5504
5505
5506 static int raid5_check_reshape(struct mddev *mddev)
5507 {
5508         /* For a 2-drive array, the layout and chunk size can be changed
5509          * immediately as not restriping is needed.
5510          * For larger arrays we record the new value - after validation
5511          * to be used by a reshape pass.
5512          */
5513         struct r5conf *conf = mddev->private;
5514         int new_chunk = mddev->new_chunk_sectors;
5515
5516         if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
5517                 return -EINVAL;
5518         if (new_chunk > 0) {
5519                 if (!is_power_of_2(new_chunk))
5520                         return -EINVAL;
5521                 if (new_chunk < (PAGE_SIZE>>9))
5522                         return -EINVAL;
5523                 if (mddev->array_sectors & (new_chunk-1))
5524                         /* not factor of array size */
5525                         return -EINVAL;
5526         }
5527
5528         /* They look valid */
5529
5530         if (mddev->raid_disks == 2) {
5531                 /* can make the change immediately */
5532                 if (mddev->new_layout >= 0) {
5533                         conf->algorithm = mddev->new_layout;
5534                         mddev->layout = mddev->new_layout;
5535                 }
5536                 if (new_chunk > 0) {
5537                         conf->chunk_sectors = new_chunk ;
5538                         mddev->chunk_sectors = new_chunk;
5539                 }
5540                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5541                 md_wakeup_thread(mddev->thread);
5542         }
5543         return check_reshape(mddev);
5544 }
5545
5546 static int raid6_check_reshape(struct mddev *mddev)
5547 {
5548         int new_chunk = mddev->new_chunk_sectors;
5549
5550         if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
5551                 return -EINVAL;
5552         if (new_chunk > 0) {
5553                 if (!is_power_of_2(new_chunk))
5554                         return -EINVAL;
5555                 if (new_chunk < (PAGE_SIZE >> 9))
5556                         return -EINVAL;
5557                 if (mddev->array_sectors & (new_chunk-1))
5558                         /* not factor of array size */
5559                         return -EINVAL;
5560         }
5561
5562         /* They look valid */
5563         return check_reshape(mddev);
5564 }
5565
5566 static void *raid5_takeover(struct mddev *mddev)
5567 {
5568         /* raid5 can take over:
5569          *  raid0 - if there is only one strip zone - make it a raid4 layout
5570          *  raid1 - if there are two drives.  We need to know the chunk size
5571          *  raid4 - trivial - just use a raid4 layout.
5572          *  raid6 - Providing it is a *_6 layout
5573          */
5574         if (mddev->level == 0)
5575                 return raid45_takeover_raid0(mddev, 5);
5576         if (mddev->level == 1)
5577                 return raid5_takeover_raid1(mddev);
5578         if (mddev->level == 4) {
5579                 mddev->new_layout = ALGORITHM_PARITY_N;
5580                 mddev->new_level = 5;
5581                 return setup_conf(mddev);
5582         }
5583         if (mddev->level == 6)
5584                 return raid5_takeover_raid6(mddev);
5585
5586         return ERR_PTR(-EINVAL);
5587 }
5588
5589 static void *raid4_takeover(struct mddev *mddev)
5590 {
5591         /* raid4 can take over:
5592          *  raid0 - if there is only one strip zone
5593          *  raid5 - if layout is right
5594          */
5595         if (mddev->level == 0)
5596                 return raid45_takeover_raid0(mddev, 4);
5597         if (mddev->level == 5 &&
5598             mddev->layout == ALGORITHM_PARITY_N) {
5599                 mddev->new_layout = 0;
5600                 mddev->new_level = 4;
5601                 return setup_conf(mddev);
5602         }
5603         return ERR_PTR(-EINVAL);
5604 }
5605
5606 static struct md_personality raid5_personality;
5607
5608 static void *raid6_takeover(struct mddev *mddev)
5609 {
5610         /* Currently can only take over a raid5.  We map the
5611          * personality to an equivalent raid6 personality
5612          * with the Q block at the end.
5613          */
5614         int new_layout;
5615
5616         if (mddev->pers != &raid5_personality)
5617                 return ERR_PTR(-EINVAL);
5618         if (mddev->degraded > 1)
5619                 return ERR_PTR(-EINVAL);
5620         if (mddev->raid_disks > 253)
5621                 return ERR_PTR(-EINVAL);
5622         if (mddev->raid_disks < 3)
5623                 return ERR_PTR(-EINVAL);
5624
5625         switch (mddev->layout) {
5626         case ALGORITHM_LEFT_ASYMMETRIC:
5627                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5628                 break;
5629         case ALGORITHM_RIGHT_ASYMMETRIC:
5630                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5631                 break;
5632         case ALGORITHM_LEFT_SYMMETRIC:
5633                 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5634                 break;
5635         case ALGORITHM_RIGHT_SYMMETRIC:
5636                 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5637                 break;
5638         case ALGORITHM_PARITY_0:
5639                 new_layout = ALGORITHM_PARITY_0_6;
5640                 break;
5641         case ALGORITHM_PARITY_N:
5642                 new_layout = ALGORITHM_PARITY_N;
5643                 break;
5644         default:
5645                 return ERR_PTR(-EINVAL);
5646         }
5647         mddev->new_level = 6;
5648         mddev->new_layout = new_layout;
5649         mddev->delta_disks = 1;
5650         mddev->raid_disks += 1;
5651         return setup_conf(mddev);
5652 }
5653
5654
5655 static struct md_personality raid6_personality =
5656 {
5657         .name           = "raid6",
5658         .level          = 6,
5659         .owner          = THIS_MODULE,
5660         .make_request   = make_request,
5661         .run            = run,
5662         .stop           = stop,
5663         .status         = status,
5664         .error_handler  = error,
5665         .hot_add_disk   = raid5_add_disk,
5666         .hot_remove_disk= raid5_remove_disk,
5667         .spare_active   = raid5_spare_active,
5668         .sync_request   = sync_request,
5669         .resize         = raid5_resize,
5670         .size           = raid5_size,
5671         .check_reshape  = raid6_check_reshape,
5672         .start_reshape  = raid5_start_reshape,
5673         .finish_reshape = raid5_finish_reshape,
5674         .quiesce        = raid5_quiesce,
5675         .takeover       = raid6_takeover,
5676 };
5677 static struct md_personality raid5_personality =
5678 {
5679         .name           = "raid5",
5680         .level          = 5,
5681         .owner          = THIS_MODULE,
5682         .make_request   = make_request,
5683         .run            = run,
5684         .stop           = stop,
5685         .status         = status,
5686         .error_handler  = error,
5687         .hot_add_disk   = raid5_add_disk,
5688         .hot_remove_disk= raid5_remove_disk,
5689         .spare_active   = raid5_spare_active,
5690         .sync_request   = sync_request,
5691         .resize         = raid5_resize,
5692         .size           = raid5_size,
5693         .check_reshape  = raid5_check_reshape,
5694         .start_reshape  = raid5_start_reshape,
5695         .finish_reshape = raid5_finish_reshape,
5696         .quiesce        = raid5_quiesce,
5697         .takeover       = raid5_takeover,
5698 };
5699
5700 static struct md_personality raid4_personality =
5701 {
5702         .name           = "raid4",
5703         .level          = 4,
5704         .owner          = THIS_MODULE,
5705         .make_request   = make_request,
5706         .run            = run,
5707         .stop           = stop,
5708         .status         = status,
5709         .error_handler  = error,
5710         .hot_add_disk   = raid5_add_disk,
5711         .hot_remove_disk= raid5_remove_disk,
5712         .spare_active   = raid5_spare_active,
5713         .sync_request   = sync_request,
5714         .resize         = raid5_resize,
5715         .size           = raid5_size,
5716         .check_reshape  = raid5_check_reshape,
5717         .start_reshape  = raid5_start_reshape,
5718         .finish_reshape = raid5_finish_reshape,
5719         .quiesce        = raid5_quiesce,
5720         .takeover       = raid4_takeover,
5721 };
5722
5723 static int __init raid5_init(void)
5724 {
5725         register_md_personality(&raid6_personality);
5726         register_md_personality(&raid5_personality);
5727         register_md_personality(&raid4_personality);
5728         return 0;
5729 }
5730
5731 static void raid5_exit(void)
5732 {
5733         unregister_md_personality(&raid6_personality);
5734         unregister_md_personality(&raid5_personality);
5735         unregister_md_personality(&raid4_personality);
5736 }
5737
5738 module_init(raid5_init);
5739 module_exit(raid5_exit);
5740 MODULE_LICENSE("GPL");
5741 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
5742 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5743 MODULE_ALIAS("md-raid5");
5744 MODULE_ALIAS("md-raid4");
5745 MODULE_ALIAS("md-level-5");
5746 MODULE_ALIAS("md-level-4");
5747 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5748 MODULE_ALIAS("md-raid6");
5749 MODULE_ALIAS("md-level-6");
5750
5751 /* This used to be two separate modules, they were: */
5752 MODULE_ALIAS("raid5");
5753 MODULE_ALIAS("raid6");