Merge branch 'for-linus' of git://neil.brown.name/md
[sfrench/cifs-2.6.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->bm_write is the number of the last batch successfully written.
31  * conf->bm_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 bm_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/async.h>
51 #include <linux/seq_file.h>
52 #include <linux/cpu.h>
53 #include <linux/slab.h>
54 #include "md.h"
55 #include "raid5.h"
56 #include "raid0.h"
57 #include "bitmap.h"
58
59 /*
60  * Stripe cache
61  */
62
63 #define NR_STRIPES              256
64 #define STRIPE_SIZE             PAGE_SIZE
65 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
66 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
67 #define IO_THRESHOLD            1
68 #define BYPASS_THRESHOLD        1
69 #define NR_HASH                 (PAGE_SIZE / sizeof(struct hlist_head))
70 #define HASH_MASK               (NR_HASH - 1)
71
72 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
73
74 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
75  * order without overlap.  There may be several bio's per stripe+device, and
76  * a bio could span several devices.
77  * When walking this list for a particular stripe+device, we must never proceed
78  * beyond a bio that extends past this device, as the next bio might no longer
79  * be valid.
80  * This macro is used to determine the 'next' bio in the list, given the sector
81  * of the current stripe+device
82  */
83 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
84 /*
85  * The following can be used to debug the driver
86  */
87 #define RAID5_PARANOIA  1
88 #if RAID5_PARANOIA && defined(CONFIG_SMP)
89 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
90 #else
91 # define CHECK_DEVLOCK()
92 #endif
93
94 #ifdef DEBUG
95 #define inline
96 #define __inline__
97 #endif
98
99 #define printk_rl(args...) ((void) (printk_ratelimit() && printk(args)))
100
101 /*
102  * We maintain a biased count of active stripes in the bottom 16 bits of
103  * bi_phys_segments, and a count of processed stripes in the upper 16 bits
104  */
105 static inline int raid5_bi_phys_segments(struct bio *bio)
106 {
107         return bio->bi_phys_segments & 0xffff;
108 }
109
110 static inline int raid5_bi_hw_segments(struct bio *bio)
111 {
112         return (bio->bi_phys_segments >> 16) & 0xffff;
113 }
114
115 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
116 {
117         --bio->bi_phys_segments;
118         return raid5_bi_phys_segments(bio);
119 }
120
121 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
122 {
123         unsigned short val = raid5_bi_hw_segments(bio);
124
125         --val;
126         bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
127         return val;
128 }
129
130 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
131 {
132         bio->bi_phys_segments = raid5_bi_phys_segments(bio) || (cnt << 16);
133 }
134
135 /* Find first data disk in a raid6 stripe */
136 static inline int raid6_d0(struct stripe_head *sh)
137 {
138         if (sh->ddf_layout)
139                 /* ddf always start from first device */
140                 return 0;
141         /* md starts just after Q block */
142         if (sh->qd_idx == sh->disks - 1)
143                 return 0;
144         else
145                 return sh->qd_idx + 1;
146 }
147 static inline int raid6_next_disk(int disk, int raid_disks)
148 {
149         disk++;
150         return (disk < raid_disks) ? disk : 0;
151 }
152
153 /* When walking through the disks in a raid5, starting at raid6_d0,
154  * We need to map each disk to a 'slot', where the data disks are slot
155  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
156  * is raid_disks-1.  This help does that mapping.
157  */
158 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
159                              int *count, int syndrome_disks)
160 {
161         int slot = *count;
162
163         if (sh->ddf_layout)
164                 (*count)++;
165         if (idx == sh->pd_idx)
166                 return syndrome_disks;
167         if (idx == sh->qd_idx)
168                 return syndrome_disks + 1;
169         if (!sh->ddf_layout)
170                 (*count)++;
171         return slot;
172 }
173
174 static void return_io(struct bio *return_bi)
175 {
176         struct bio *bi = return_bi;
177         while (bi) {
178
179                 return_bi = bi->bi_next;
180                 bi->bi_next = NULL;
181                 bi->bi_size = 0;
182                 bio_endio(bi, 0);
183                 bi = return_bi;
184         }
185 }
186
187 static void print_raid5_conf (raid5_conf_t *conf);
188
189 static int stripe_operations_active(struct stripe_head *sh)
190 {
191         return sh->check_state || sh->reconstruct_state ||
192                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
193                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
194 }
195
196 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
197 {
198         if (atomic_dec_and_test(&sh->count)) {
199                 BUG_ON(!list_empty(&sh->lru));
200                 BUG_ON(atomic_read(&conf->active_stripes)==0);
201                 if (test_bit(STRIPE_HANDLE, &sh->state)) {
202                         if (test_bit(STRIPE_DELAYED, &sh->state)) {
203                                 list_add_tail(&sh->lru, &conf->delayed_list);
204                                 plugger_set_plug(&conf->plug);
205                         } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
206                                    sh->bm_seq - conf->seq_write > 0) {
207                                 list_add_tail(&sh->lru, &conf->bitmap_list);
208                                 plugger_set_plug(&conf->plug);
209                         } else {
210                                 clear_bit(STRIPE_BIT_DELAY, &sh->state);
211                                 list_add_tail(&sh->lru, &conf->handle_list);
212                         }
213                         md_wakeup_thread(conf->mddev->thread);
214                 } else {
215                         BUG_ON(stripe_operations_active(sh));
216                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
217                                 atomic_dec(&conf->preread_active_stripes);
218                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
219                                         md_wakeup_thread(conf->mddev->thread);
220                         }
221                         atomic_dec(&conf->active_stripes);
222                         if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
223                                 list_add_tail(&sh->lru, &conf->inactive_list);
224                                 wake_up(&conf->wait_for_stripe);
225                                 if (conf->retry_read_aligned)
226                                         md_wakeup_thread(conf->mddev->thread);
227                         }
228                 }
229         }
230 }
231
232 static void release_stripe(struct stripe_head *sh)
233 {
234         raid5_conf_t *conf = sh->raid_conf;
235         unsigned long flags;
236
237         spin_lock_irqsave(&conf->device_lock, flags);
238         __release_stripe(conf, sh);
239         spin_unlock_irqrestore(&conf->device_lock, flags);
240 }
241
242 static inline void remove_hash(struct stripe_head *sh)
243 {
244         pr_debug("remove_hash(), stripe %llu\n",
245                 (unsigned long long)sh->sector);
246
247         hlist_del_init(&sh->hash);
248 }
249
250 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
251 {
252         struct hlist_head *hp = stripe_hash(conf, sh->sector);
253
254         pr_debug("insert_hash(), stripe %llu\n",
255                 (unsigned long long)sh->sector);
256
257         CHECK_DEVLOCK();
258         hlist_add_head(&sh->hash, hp);
259 }
260
261
262 /* find an idle stripe, make sure it is unhashed, and return it. */
263 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
264 {
265         struct stripe_head *sh = NULL;
266         struct list_head *first;
267
268         CHECK_DEVLOCK();
269         if (list_empty(&conf->inactive_list))
270                 goto out;
271         first = conf->inactive_list.next;
272         sh = list_entry(first, struct stripe_head, lru);
273         list_del_init(first);
274         remove_hash(sh);
275         atomic_inc(&conf->active_stripes);
276 out:
277         return sh;
278 }
279
280 static void shrink_buffers(struct stripe_head *sh)
281 {
282         struct page *p;
283         int i;
284         int num = sh->raid_conf->pool_size;
285
286         for (i = 0; i < num ; i++) {
287                 p = sh->dev[i].page;
288                 if (!p)
289                         continue;
290                 sh->dev[i].page = NULL;
291                 put_page(p);
292         }
293 }
294
295 static int grow_buffers(struct stripe_head *sh)
296 {
297         int i;
298         int num = sh->raid_conf->pool_size;
299
300         for (i = 0; i < num; i++) {
301                 struct page *page;
302
303                 if (!(page = alloc_page(GFP_KERNEL))) {
304                         return 1;
305                 }
306                 sh->dev[i].page = page;
307         }
308         return 0;
309 }
310
311 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
312 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
313                             struct stripe_head *sh);
314
315 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
316 {
317         raid5_conf_t *conf = sh->raid_conf;
318         int i;
319
320         BUG_ON(atomic_read(&sh->count) != 0);
321         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
322         BUG_ON(stripe_operations_active(sh));
323
324         CHECK_DEVLOCK();
325         pr_debug("init_stripe called, stripe %llu\n",
326                 (unsigned long long)sh->sector);
327
328         remove_hash(sh);
329
330         sh->generation = conf->generation - previous;
331         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
332         sh->sector = sector;
333         stripe_set_idx(sector, conf, previous, sh);
334         sh->state = 0;
335
336
337         for (i = sh->disks; i--; ) {
338                 struct r5dev *dev = &sh->dev[i];
339
340                 if (dev->toread || dev->read || dev->towrite || dev->written ||
341                     test_bit(R5_LOCKED, &dev->flags)) {
342                         printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
343                                (unsigned long long)sh->sector, i, dev->toread,
344                                dev->read, dev->towrite, dev->written,
345                                test_bit(R5_LOCKED, &dev->flags));
346                         BUG();
347                 }
348                 dev->flags = 0;
349                 raid5_build_block(sh, i, previous);
350         }
351         insert_hash(conf, sh);
352 }
353
354 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector,
355                                          short generation)
356 {
357         struct stripe_head *sh;
358         struct hlist_node *hn;
359
360         CHECK_DEVLOCK();
361         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
362         hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
363                 if (sh->sector == sector && sh->generation == generation)
364                         return sh;
365         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
366         return NULL;
367 }
368
369 /*
370  * Need to check if array has failed when deciding whether to:
371  *  - start an array
372  *  - remove non-faulty devices
373  *  - add a spare
374  *  - allow a reshape
375  * This determination is simple when no reshape is happening.
376  * However if there is a reshape, we need to carefully check
377  * both the before and after sections.
378  * This is because some failed devices may only affect one
379  * of the two sections, and some non-in_sync devices may
380  * be insync in the section most affected by failed devices.
381  */
382 static int has_failed(raid5_conf_t *conf)
383 {
384         int degraded;
385         int i;
386         if (conf->mddev->reshape_position == MaxSector)
387                 return conf->mddev->degraded > conf->max_degraded;
388
389         rcu_read_lock();
390         degraded = 0;
391         for (i = 0; i < conf->previous_raid_disks; i++) {
392                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
393                 if (!rdev || test_bit(Faulty, &rdev->flags))
394                         degraded++;
395                 else if (test_bit(In_sync, &rdev->flags))
396                         ;
397                 else
398                         /* not in-sync or faulty.
399                          * If the reshape increases the number of devices,
400                          * this is being recovered by the reshape, so
401                          * this 'previous' section is not in_sync.
402                          * If the number of devices is being reduced however,
403                          * the device can only be part of the array if
404                          * we are reverting a reshape, so this section will
405                          * be in-sync.
406                          */
407                         if (conf->raid_disks >= conf->previous_raid_disks)
408                                 degraded++;
409         }
410         rcu_read_unlock();
411         if (degraded > conf->max_degraded)
412                 return 1;
413         rcu_read_lock();
414         degraded = 0;
415         for (i = 0; i < conf->raid_disks; i++) {
416                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
417                 if (!rdev || test_bit(Faulty, &rdev->flags))
418                         degraded++;
419                 else if (test_bit(In_sync, &rdev->flags))
420                         ;
421                 else
422                         /* not in-sync or faulty.
423                          * If reshape increases the number of devices, this
424                          * section has already been recovered, else it
425                          * almost certainly hasn't.
426                          */
427                         if (conf->raid_disks <= conf->previous_raid_disks)
428                                 degraded++;
429         }
430         rcu_read_unlock();
431         if (degraded > conf->max_degraded)
432                 return 1;
433         return 0;
434 }
435
436 static void unplug_slaves(mddev_t *mddev);
437
438 static struct stripe_head *
439 get_active_stripe(raid5_conf_t *conf, sector_t sector,
440                   int previous, int noblock, int noquiesce)
441 {
442         struct stripe_head *sh;
443
444         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
445
446         spin_lock_irq(&conf->device_lock);
447
448         do {
449                 wait_event_lock_irq(conf->wait_for_stripe,
450                                     conf->quiesce == 0 || noquiesce,
451                                     conf->device_lock, /* nothing */);
452                 sh = __find_stripe(conf, sector, conf->generation - previous);
453                 if (!sh) {
454                         if (!conf->inactive_blocked)
455                                 sh = get_free_stripe(conf);
456                         if (noblock && sh == NULL)
457                                 break;
458                         if (!sh) {
459                                 conf->inactive_blocked = 1;
460                                 wait_event_lock_irq(conf->wait_for_stripe,
461                                                     !list_empty(&conf->inactive_list) &&
462                                                     (atomic_read(&conf->active_stripes)
463                                                      < (conf->max_nr_stripes *3/4)
464                                                      || !conf->inactive_blocked),
465                                                     conf->device_lock,
466                                                     md_raid5_unplug_device(conf)
467                                         );
468                                 conf->inactive_blocked = 0;
469                         } else
470                                 init_stripe(sh, sector, previous);
471                 } else {
472                         if (atomic_read(&sh->count)) {
473                                 BUG_ON(!list_empty(&sh->lru)
474                                     && !test_bit(STRIPE_EXPANDING, &sh->state));
475                         } else {
476                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
477                                         atomic_inc(&conf->active_stripes);
478                                 if (list_empty(&sh->lru) &&
479                                     !test_bit(STRIPE_EXPANDING, &sh->state))
480                                         BUG();
481                                 list_del_init(&sh->lru);
482                         }
483                 }
484         } while (sh == NULL);
485
486         if (sh)
487                 atomic_inc(&sh->count);
488
489         spin_unlock_irq(&conf->device_lock);
490         return sh;
491 }
492
493 static void
494 raid5_end_read_request(struct bio *bi, int error);
495 static void
496 raid5_end_write_request(struct bio *bi, int error);
497
498 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
499 {
500         raid5_conf_t *conf = sh->raid_conf;
501         int i, disks = sh->disks;
502
503         might_sleep();
504
505         for (i = disks; i--; ) {
506                 int rw;
507                 struct bio *bi;
508                 mdk_rdev_t *rdev;
509                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
510                         rw = WRITE;
511                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
512                         rw = READ;
513                 else
514                         continue;
515
516                 bi = &sh->dev[i].req;
517
518                 bi->bi_rw = rw;
519                 if (rw == WRITE)
520                         bi->bi_end_io = raid5_end_write_request;
521                 else
522                         bi->bi_end_io = raid5_end_read_request;
523
524                 rcu_read_lock();
525                 rdev = rcu_dereference(conf->disks[i].rdev);
526                 if (rdev && test_bit(Faulty, &rdev->flags))
527                         rdev = NULL;
528                 if (rdev)
529                         atomic_inc(&rdev->nr_pending);
530                 rcu_read_unlock();
531
532                 if (rdev) {
533                         if (s->syncing || s->expanding || s->expanded)
534                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
535
536                         set_bit(STRIPE_IO_STARTED, &sh->state);
537
538                         bi->bi_bdev = rdev->bdev;
539                         pr_debug("%s: for %llu schedule op %ld on disc %d\n",
540                                 __func__, (unsigned long long)sh->sector,
541                                 bi->bi_rw, i);
542                         atomic_inc(&sh->count);
543                         bi->bi_sector = sh->sector + rdev->data_offset;
544                         bi->bi_flags = 1 << BIO_UPTODATE;
545                         bi->bi_vcnt = 1;
546                         bi->bi_max_vecs = 1;
547                         bi->bi_idx = 0;
548                         bi->bi_io_vec = &sh->dev[i].vec;
549                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
550                         bi->bi_io_vec[0].bv_offset = 0;
551                         bi->bi_size = STRIPE_SIZE;
552                         bi->bi_next = NULL;
553                         if (rw == WRITE &&
554                             test_bit(R5_ReWrite, &sh->dev[i].flags))
555                                 atomic_add(STRIPE_SECTORS,
556                                         &rdev->corrected_errors);
557                         generic_make_request(bi);
558                 } else {
559                         if (rw == WRITE)
560                                 set_bit(STRIPE_DEGRADED, &sh->state);
561                         pr_debug("skip op %ld on disc %d for sector %llu\n",
562                                 bi->bi_rw, i, (unsigned long long)sh->sector);
563                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
564                         set_bit(STRIPE_HANDLE, &sh->state);
565                 }
566         }
567 }
568
569 static struct dma_async_tx_descriptor *
570 async_copy_data(int frombio, struct bio *bio, struct page *page,
571         sector_t sector, struct dma_async_tx_descriptor *tx)
572 {
573         struct bio_vec *bvl;
574         struct page *bio_page;
575         int i;
576         int page_offset;
577         struct async_submit_ctl submit;
578         enum async_tx_flags flags = 0;
579
580         if (bio->bi_sector >= sector)
581                 page_offset = (signed)(bio->bi_sector - sector) * 512;
582         else
583                 page_offset = (signed)(sector - bio->bi_sector) * -512;
584
585         if (frombio)
586                 flags |= ASYNC_TX_FENCE;
587         init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
588
589         bio_for_each_segment(bvl, bio, i) {
590                 int len = bio_iovec_idx(bio, i)->bv_len;
591                 int clen;
592                 int b_offset = 0;
593
594                 if (page_offset < 0) {
595                         b_offset = -page_offset;
596                         page_offset += b_offset;
597                         len -= b_offset;
598                 }
599
600                 if (len > 0 && page_offset + len > STRIPE_SIZE)
601                         clen = STRIPE_SIZE - page_offset;
602                 else
603                         clen = len;
604
605                 if (clen > 0) {
606                         b_offset += bio_iovec_idx(bio, i)->bv_offset;
607                         bio_page = bio_iovec_idx(bio, i)->bv_page;
608                         if (frombio)
609                                 tx = async_memcpy(page, bio_page, page_offset,
610                                                   b_offset, clen, &submit);
611                         else
612                                 tx = async_memcpy(bio_page, page, b_offset,
613                                                   page_offset, clen, &submit);
614                 }
615                 /* chain the operations */
616                 submit.depend_tx = tx;
617
618                 if (clen < len) /* hit end of page */
619                         break;
620                 page_offset +=  len;
621         }
622
623         return tx;
624 }
625
626 static void ops_complete_biofill(void *stripe_head_ref)
627 {
628         struct stripe_head *sh = stripe_head_ref;
629         struct bio *return_bi = NULL;
630         raid5_conf_t *conf = sh->raid_conf;
631         int i;
632
633         pr_debug("%s: stripe %llu\n", __func__,
634                 (unsigned long long)sh->sector);
635
636         /* clear completed biofills */
637         spin_lock_irq(&conf->device_lock);
638         for (i = sh->disks; i--; ) {
639                 struct r5dev *dev = &sh->dev[i];
640
641                 /* acknowledge completion of a biofill operation */
642                 /* and check if we need to reply to a read request,
643                  * new R5_Wantfill requests are held off until
644                  * !STRIPE_BIOFILL_RUN
645                  */
646                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
647                         struct bio *rbi, *rbi2;
648
649                         BUG_ON(!dev->read);
650                         rbi = dev->read;
651                         dev->read = NULL;
652                         while (rbi && rbi->bi_sector <
653                                 dev->sector + STRIPE_SECTORS) {
654                                 rbi2 = r5_next_bio(rbi, dev->sector);
655                                 if (!raid5_dec_bi_phys_segments(rbi)) {
656                                         rbi->bi_next = return_bi;
657                                         return_bi = rbi;
658                                 }
659                                 rbi = rbi2;
660                         }
661                 }
662         }
663         spin_unlock_irq(&conf->device_lock);
664         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
665
666         return_io(return_bi);
667
668         set_bit(STRIPE_HANDLE, &sh->state);
669         release_stripe(sh);
670 }
671
672 static void ops_run_biofill(struct stripe_head *sh)
673 {
674         struct dma_async_tx_descriptor *tx = NULL;
675         raid5_conf_t *conf = sh->raid_conf;
676         struct async_submit_ctl submit;
677         int i;
678
679         pr_debug("%s: stripe %llu\n", __func__,
680                 (unsigned long long)sh->sector);
681
682         for (i = sh->disks; i--; ) {
683                 struct r5dev *dev = &sh->dev[i];
684                 if (test_bit(R5_Wantfill, &dev->flags)) {
685                         struct bio *rbi;
686                         spin_lock_irq(&conf->device_lock);
687                         dev->read = rbi = dev->toread;
688                         dev->toread = NULL;
689                         spin_unlock_irq(&conf->device_lock);
690                         while (rbi && rbi->bi_sector <
691                                 dev->sector + STRIPE_SECTORS) {
692                                 tx = async_copy_data(0, rbi, dev->page,
693                                         dev->sector, tx);
694                                 rbi = r5_next_bio(rbi, dev->sector);
695                         }
696                 }
697         }
698
699         atomic_inc(&sh->count);
700         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
701         async_trigger_callback(&submit);
702 }
703
704 static void mark_target_uptodate(struct stripe_head *sh, int target)
705 {
706         struct r5dev *tgt;
707
708         if (target < 0)
709                 return;
710
711         tgt = &sh->dev[target];
712         set_bit(R5_UPTODATE, &tgt->flags);
713         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
714         clear_bit(R5_Wantcompute, &tgt->flags);
715 }
716
717 static void ops_complete_compute(void *stripe_head_ref)
718 {
719         struct stripe_head *sh = stripe_head_ref;
720
721         pr_debug("%s: stripe %llu\n", __func__,
722                 (unsigned long long)sh->sector);
723
724         /* mark the computed target(s) as uptodate */
725         mark_target_uptodate(sh, sh->ops.target);
726         mark_target_uptodate(sh, sh->ops.target2);
727
728         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
729         if (sh->check_state == check_state_compute_run)
730                 sh->check_state = check_state_compute_result;
731         set_bit(STRIPE_HANDLE, &sh->state);
732         release_stripe(sh);
733 }
734
735 /* return a pointer to the address conversion region of the scribble buffer */
736 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
737                                  struct raid5_percpu *percpu)
738 {
739         return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
740 }
741
742 static struct dma_async_tx_descriptor *
743 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
744 {
745         int disks = sh->disks;
746         struct page **xor_srcs = percpu->scribble;
747         int target = sh->ops.target;
748         struct r5dev *tgt = &sh->dev[target];
749         struct page *xor_dest = tgt->page;
750         int count = 0;
751         struct dma_async_tx_descriptor *tx;
752         struct async_submit_ctl submit;
753         int i;
754
755         pr_debug("%s: stripe %llu block: %d\n",
756                 __func__, (unsigned long long)sh->sector, target);
757         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
758
759         for (i = disks; i--; )
760                 if (i != target)
761                         xor_srcs[count++] = sh->dev[i].page;
762
763         atomic_inc(&sh->count);
764
765         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
766                           ops_complete_compute, sh, to_addr_conv(sh, percpu));
767         if (unlikely(count == 1))
768                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
769         else
770                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
771
772         return tx;
773 }
774
775 /* set_syndrome_sources - populate source buffers for gen_syndrome
776  * @srcs - (struct page *) array of size sh->disks
777  * @sh - stripe_head to parse
778  *
779  * Populates srcs in proper layout order for the stripe and returns the
780  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
781  * destination buffer is recorded in srcs[count] and the Q destination
782  * is recorded in srcs[count+1]].
783  */
784 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
785 {
786         int disks = sh->disks;
787         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
788         int d0_idx = raid6_d0(sh);
789         int count;
790         int i;
791
792         for (i = 0; i < disks; i++)
793                 srcs[i] = NULL;
794
795         count = 0;
796         i = d0_idx;
797         do {
798                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
799
800                 srcs[slot] = sh->dev[i].page;
801                 i = raid6_next_disk(i, disks);
802         } while (i != d0_idx);
803
804         return syndrome_disks;
805 }
806
807 static struct dma_async_tx_descriptor *
808 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
809 {
810         int disks = sh->disks;
811         struct page **blocks = percpu->scribble;
812         int target;
813         int qd_idx = sh->qd_idx;
814         struct dma_async_tx_descriptor *tx;
815         struct async_submit_ctl submit;
816         struct r5dev *tgt;
817         struct page *dest;
818         int i;
819         int count;
820
821         if (sh->ops.target < 0)
822                 target = sh->ops.target2;
823         else if (sh->ops.target2 < 0)
824                 target = sh->ops.target;
825         else
826                 /* we should only have one valid target */
827                 BUG();
828         BUG_ON(target < 0);
829         pr_debug("%s: stripe %llu block: %d\n",
830                 __func__, (unsigned long long)sh->sector, target);
831
832         tgt = &sh->dev[target];
833         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
834         dest = tgt->page;
835
836         atomic_inc(&sh->count);
837
838         if (target == qd_idx) {
839                 count = set_syndrome_sources(blocks, sh);
840                 blocks[count] = NULL; /* regenerating p is not necessary */
841                 BUG_ON(blocks[count+1] != dest); /* q should already be set */
842                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
843                                   ops_complete_compute, sh,
844                                   to_addr_conv(sh, percpu));
845                 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
846         } else {
847                 /* Compute any data- or p-drive using XOR */
848                 count = 0;
849                 for (i = disks; i-- ; ) {
850                         if (i == target || i == qd_idx)
851                                 continue;
852                         blocks[count++] = sh->dev[i].page;
853                 }
854
855                 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
856                                   NULL, ops_complete_compute, sh,
857                                   to_addr_conv(sh, percpu));
858                 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
859         }
860
861         return tx;
862 }
863
864 static struct dma_async_tx_descriptor *
865 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
866 {
867         int i, count, disks = sh->disks;
868         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
869         int d0_idx = raid6_d0(sh);
870         int faila = -1, failb = -1;
871         int target = sh->ops.target;
872         int target2 = sh->ops.target2;
873         struct r5dev *tgt = &sh->dev[target];
874         struct r5dev *tgt2 = &sh->dev[target2];
875         struct dma_async_tx_descriptor *tx;
876         struct page **blocks = percpu->scribble;
877         struct async_submit_ctl submit;
878
879         pr_debug("%s: stripe %llu block1: %d block2: %d\n",
880                  __func__, (unsigned long long)sh->sector, target, target2);
881         BUG_ON(target < 0 || target2 < 0);
882         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
883         BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
884
885         /* we need to open-code set_syndrome_sources to handle the
886          * slot number conversion for 'faila' and 'failb'
887          */
888         for (i = 0; i < disks ; i++)
889                 blocks[i] = NULL;
890         count = 0;
891         i = d0_idx;
892         do {
893                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
894
895                 blocks[slot] = sh->dev[i].page;
896
897                 if (i == target)
898                         faila = slot;
899                 if (i == target2)
900                         failb = slot;
901                 i = raid6_next_disk(i, disks);
902         } while (i != d0_idx);
903
904         BUG_ON(faila == failb);
905         if (failb < faila)
906                 swap(faila, failb);
907         pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
908                  __func__, (unsigned long long)sh->sector, faila, failb);
909
910         atomic_inc(&sh->count);
911
912         if (failb == syndrome_disks+1) {
913                 /* Q disk is one of the missing disks */
914                 if (faila == syndrome_disks) {
915                         /* Missing P+Q, just recompute */
916                         init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
917                                           ops_complete_compute, sh,
918                                           to_addr_conv(sh, percpu));
919                         return async_gen_syndrome(blocks, 0, syndrome_disks+2,
920                                                   STRIPE_SIZE, &submit);
921                 } else {
922                         struct page *dest;
923                         int data_target;
924                         int qd_idx = sh->qd_idx;
925
926                         /* Missing D+Q: recompute D from P, then recompute Q */
927                         if (target == qd_idx)
928                                 data_target = target2;
929                         else
930                                 data_target = target;
931
932                         count = 0;
933                         for (i = disks; i-- ; ) {
934                                 if (i == data_target || i == qd_idx)
935                                         continue;
936                                 blocks[count++] = sh->dev[i].page;
937                         }
938                         dest = sh->dev[data_target].page;
939                         init_async_submit(&submit,
940                                           ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
941                                           NULL, NULL, NULL,
942                                           to_addr_conv(sh, percpu));
943                         tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
944                                        &submit);
945
946                         count = set_syndrome_sources(blocks, sh);
947                         init_async_submit(&submit, ASYNC_TX_FENCE, tx,
948                                           ops_complete_compute, sh,
949                                           to_addr_conv(sh, percpu));
950                         return async_gen_syndrome(blocks, 0, count+2,
951                                                   STRIPE_SIZE, &submit);
952                 }
953         } else {
954                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
955                                   ops_complete_compute, sh,
956                                   to_addr_conv(sh, percpu));
957                 if (failb == syndrome_disks) {
958                         /* We're missing D+P. */
959                         return async_raid6_datap_recov(syndrome_disks+2,
960                                                        STRIPE_SIZE, faila,
961                                                        blocks, &submit);
962                 } else {
963                         /* We're missing D+D. */
964                         return async_raid6_2data_recov(syndrome_disks+2,
965                                                        STRIPE_SIZE, faila, failb,
966                                                        blocks, &submit);
967                 }
968         }
969 }
970
971
972 static void ops_complete_prexor(void *stripe_head_ref)
973 {
974         struct stripe_head *sh = stripe_head_ref;
975
976         pr_debug("%s: stripe %llu\n", __func__,
977                 (unsigned long long)sh->sector);
978 }
979
980 static struct dma_async_tx_descriptor *
981 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
982                struct dma_async_tx_descriptor *tx)
983 {
984         int disks = sh->disks;
985         struct page **xor_srcs = percpu->scribble;
986         int count = 0, pd_idx = sh->pd_idx, i;
987         struct async_submit_ctl submit;
988
989         /* existing parity data subtracted */
990         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
991
992         pr_debug("%s: stripe %llu\n", __func__,
993                 (unsigned long long)sh->sector);
994
995         for (i = disks; i--; ) {
996                 struct r5dev *dev = &sh->dev[i];
997                 /* Only process blocks that are known to be uptodate */
998                 if (test_bit(R5_Wantdrain, &dev->flags))
999                         xor_srcs[count++] = dev->page;
1000         }
1001
1002         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1003                           ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1004         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1005
1006         return tx;
1007 }
1008
1009 static struct dma_async_tx_descriptor *
1010 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1011 {
1012         int disks = sh->disks;
1013         int i;
1014
1015         pr_debug("%s: stripe %llu\n", __func__,
1016                 (unsigned long long)sh->sector);
1017
1018         for (i = disks; i--; ) {
1019                 struct r5dev *dev = &sh->dev[i];
1020                 struct bio *chosen;
1021
1022                 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1023                         struct bio *wbi;
1024
1025                         spin_lock(&sh->lock);
1026                         chosen = dev->towrite;
1027                         dev->towrite = NULL;
1028                         BUG_ON(dev->written);
1029                         wbi = dev->written = chosen;
1030                         spin_unlock(&sh->lock);
1031
1032                         while (wbi && wbi->bi_sector <
1033                                 dev->sector + STRIPE_SECTORS) {
1034                                 tx = async_copy_data(1, wbi, dev->page,
1035                                         dev->sector, tx);
1036                                 wbi = r5_next_bio(wbi, dev->sector);
1037                         }
1038                 }
1039         }
1040
1041         return tx;
1042 }
1043
1044 static void ops_complete_reconstruct(void *stripe_head_ref)
1045 {
1046         struct stripe_head *sh = stripe_head_ref;
1047         int disks = sh->disks;
1048         int pd_idx = sh->pd_idx;
1049         int qd_idx = sh->qd_idx;
1050         int i;
1051
1052         pr_debug("%s: stripe %llu\n", __func__,
1053                 (unsigned long long)sh->sector);
1054
1055         for (i = disks; i--; ) {
1056                 struct r5dev *dev = &sh->dev[i];
1057
1058                 if (dev->written || i == pd_idx || i == qd_idx)
1059                         set_bit(R5_UPTODATE, &dev->flags);
1060         }
1061
1062         if (sh->reconstruct_state == reconstruct_state_drain_run)
1063                 sh->reconstruct_state = reconstruct_state_drain_result;
1064         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1065                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1066         else {
1067                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1068                 sh->reconstruct_state = reconstruct_state_result;
1069         }
1070
1071         set_bit(STRIPE_HANDLE, &sh->state);
1072         release_stripe(sh);
1073 }
1074
1075 static void
1076 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1077                      struct dma_async_tx_descriptor *tx)
1078 {
1079         int disks = sh->disks;
1080         struct page **xor_srcs = percpu->scribble;
1081         struct async_submit_ctl submit;
1082         int count = 0, pd_idx = sh->pd_idx, i;
1083         struct page *xor_dest;
1084         int prexor = 0;
1085         unsigned long flags;
1086
1087         pr_debug("%s: stripe %llu\n", __func__,
1088                 (unsigned long long)sh->sector);
1089
1090         /* check if prexor is active which means only process blocks
1091          * that are part of a read-modify-write (written)
1092          */
1093         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1094                 prexor = 1;
1095                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1096                 for (i = disks; i--; ) {
1097                         struct r5dev *dev = &sh->dev[i];
1098                         if (dev->written)
1099                                 xor_srcs[count++] = dev->page;
1100                 }
1101         } else {
1102                 xor_dest = sh->dev[pd_idx].page;
1103                 for (i = disks; i--; ) {
1104                         struct r5dev *dev = &sh->dev[i];
1105                         if (i != pd_idx)
1106                                 xor_srcs[count++] = dev->page;
1107                 }
1108         }
1109
1110         /* 1/ if we prexor'd then the dest is reused as a source
1111          * 2/ if we did not prexor then we are redoing the parity
1112          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1113          * for the synchronous xor case
1114          */
1115         flags = ASYNC_TX_ACK |
1116                 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1117
1118         atomic_inc(&sh->count);
1119
1120         init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1121                           to_addr_conv(sh, percpu));
1122         if (unlikely(count == 1))
1123                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1124         else
1125                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1126 }
1127
1128 static void
1129 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1130                      struct dma_async_tx_descriptor *tx)
1131 {
1132         struct async_submit_ctl submit;
1133         struct page **blocks = percpu->scribble;
1134         int count;
1135
1136         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1137
1138         count = set_syndrome_sources(blocks, sh);
1139
1140         atomic_inc(&sh->count);
1141
1142         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1143                           sh, to_addr_conv(sh, percpu));
1144         async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1145 }
1146
1147 static void ops_complete_check(void *stripe_head_ref)
1148 {
1149         struct stripe_head *sh = stripe_head_ref;
1150
1151         pr_debug("%s: stripe %llu\n", __func__,
1152                 (unsigned long long)sh->sector);
1153
1154         sh->check_state = check_state_check_result;
1155         set_bit(STRIPE_HANDLE, &sh->state);
1156         release_stripe(sh);
1157 }
1158
1159 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1160 {
1161         int disks = sh->disks;
1162         int pd_idx = sh->pd_idx;
1163         int qd_idx = sh->qd_idx;
1164         struct page *xor_dest;
1165         struct page **xor_srcs = percpu->scribble;
1166         struct dma_async_tx_descriptor *tx;
1167         struct async_submit_ctl submit;
1168         int count;
1169         int i;
1170
1171         pr_debug("%s: stripe %llu\n", __func__,
1172                 (unsigned long long)sh->sector);
1173
1174         count = 0;
1175         xor_dest = sh->dev[pd_idx].page;
1176         xor_srcs[count++] = xor_dest;
1177         for (i = disks; i--; ) {
1178                 if (i == pd_idx || i == qd_idx)
1179                         continue;
1180                 xor_srcs[count++] = sh->dev[i].page;
1181         }
1182
1183         init_async_submit(&submit, 0, NULL, NULL, NULL,
1184                           to_addr_conv(sh, percpu));
1185         tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1186                            &sh->ops.zero_sum_result, &submit);
1187
1188         atomic_inc(&sh->count);
1189         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1190         tx = async_trigger_callback(&submit);
1191 }
1192
1193 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1194 {
1195         struct page **srcs = percpu->scribble;
1196         struct async_submit_ctl submit;
1197         int count;
1198
1199         pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1200                 (unsigned long long)sh->sector, checkp);
1201
1202         count = set_syndrome_sources(srcs, sh);
1203         if (!checkp)
1204                 srcs[count] = NULL;
1205
1206         atomic_inc(&sh->count);
1207         init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1208                           sh, to_addr_conv(sh, percpu));
1209         async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1210                            &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1211 }
1212
1213 static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1214 {
1215         int overlap_clear = 0, i, disks = sh->disks;
1216         struct dma_async_tx_descriptor *tx = NULL;
1217         raid5_conf_t *conf = sh->raid_conf;
1218         int level = conf->level;
1219         struct raid5_percpu *percpu;
1220         unsigned long cpu;
1221
1222         cpu = get_cpu();
1223         percpu = per_cpu_ptr(conf->percpu, cpu);
1224         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1225                 ops_run_biofill(sh);
1226                 overlap_clear++;
1227         }
1228
1229         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1230                 if (level < 6)
1231                         tx = ops_run_compute5(sh, percpu);
1232                 else {
1233                         if (sh->ops.target2 < 0 || sh->ops.target < 0)
1234                                 tx = ops_run_compute6_1(sh, percpu);
1235                         else
1236                                 tx = ops_run_compute6_2(sh, percpu);
1237                 }
1238                 /* terminate the chain if reconstruct is not set to be run */
1239                 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1240                         async_tx_ack(tx);
1241         }
1242
1243         if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1244                 tx = ops_run_prexor(sh, percpu, tx);
1245
1246         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1247                 tx = ops_run_biodrain(sh, tx);
1248                 overlap_clear++;
1249         }
1250
1251         if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1252                 if (level < 6)
1253                         ops_run_reconstruct5(sh, percpu, tx);
1254                 else
1255                         ops_run_reconstruct6(sh, percpu, tx);
1256         }
1257
1258         if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1259                 if (sh->check_state == check_state_run)
1260                         ops_run_check_p(sh, percpu);
1261                 else if (sh->check_state == check_state_run_q)
1262                         ops_run_check_pq(sh, percpu, 0);
1263                 else if (sh->check_state == check_state_run_pq)
1264                         ops_run_check_pq(sh, percpu, 1);
1265                 else
1266                         BUG();
1267         }
1268
1269         if (overlap_clear)
1270                 for (i = disks; i--; ) {
1271                         struct r5dev *dev = &sh->dev[i];
1272                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
1273                                 wake_up(&sh->raid_conf->wait_for_overlap);
1274                 }
1275         put_cpu();
1276 }
1277
1278 #ifdef CONFIG_MULTICORE_RAID456
1279 static void async_run_ops(void *param, async_cookie_t cookie)
1280 {
1281         struct stripe_head *sh = param;
1282         unsigned long ops_request = sh->ops.request;
1283
1284         clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1285         wake_up(&sh->ops.wait_for_ops);
1286
1287         __raid_run_ops(sh, ops_request);
1288         release_stripe(sh);
1289 }
1290
1291 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1292 {
1293         /* since handle_stripe can be called outside of raid5d context
1294          * we need to ensure sh->ops.request is de-staged before another
1295          * request arrives
1296          */
1297         wait_event(sh->ops.wait_for_ops,
1298                    !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1299         sh->ops.request = ops_request;
1300
1301         atomic_inc(&sh->count);
1302         async_schedule(async_run_ops, sh);
1303 }
1304 #else
1305 #define raid_run_ops __raid_run_ops
1306 #endif
1307
1308 static int grow_one_stripe(raid5_conf_t *conf)
1309 {
1310         struct stripe_head *sh;
1311         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
1312         if (!sh)
1313                 return 0;
1314         memset(sh, 0, sizeof(*sh) + (conf->pool_size-1)*sizeof(struct r5dev));
1315         sh->raid_conf = conf;
1316         spin_lock_init(&sh->lock);
1317         #ifdef CONFIG_MULTICORE_RAID456
1318         init_waitqueue_head(&sh->ops.wait_for_ops);
1319         #endif
1320
1321         if (grow_buffers(sh)) {
1322                 shrink_buffers(sh);
1323                 kmem_cache_free(conf->slab_cache, sh);
1324                 return 0;
1325         }
1326         /* we just created an active stripe so... */
1327         atomic_set(&sh->count, 1);
1328         atomic_inc(&conf->active_stripes);
1329         INIT_LIST_HEAD(&sh->lru);
1330         release_stripe(sh);
1331         return 1;
1332 }
1333
1334 static int grow_stripes(raid5_conf_t *conf, int num)
1335 {
1336         struct kmem_cache *sc;
1337         int devs = max(conf->raid_disks, conf->previous_raid_disks);
1338
1339         if (conf->mddev->gendisk)
1340                 sprintf(conf->cache_name[0],
1341                         "raid%d-%s", conf->level, mdname(conf->mddev));
1342         else
1343                 sprintf(conf->cache_name[0],
1344                         "raid%d-%p", conf->level, conf->mddev);
1345         sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1346
1347         conf->active_name = 0;
1348         sc = kmem_cache_create(conf->cache_name[conf->active_name],
1349                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1350                                0, 0, NULL);
1351         if (!sc)
1352                 return 1;
1353         conf->slab_cache = sc;
1354         conf->pool_size = devs;
1355         while (num--)
1356                 if (!grow_one_stripe(conf))
1357                         return 1;
1358         return 0;
1359 }
1360
1361 /**
1362  * scribble_len - return the required size of the scribble region
1363  * @num - total number of disks in the array
1364  *
1365  * The size must be enough to contain:
1366  * 1/ a struct page pointer for each device in the array +2
1367  * 2/ room to convert each entry in (1) to its corresponding dma
1368  *    (dma_map_page()) or page (page_address()) address.
1369  *
1370  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1371  * calculate over all devices (not just the data blocks), using zeros in place
1372  * of the P and Q blocks.
1373  */
1374 static size_t scribble_len(int num)
1375 {
1376         size_t len;
1377
1378         len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1379
1380         return len;
1381 }
1382
1383 static int resize_stripes(raid5_conf_t *conf, int newsize)
1384 {
1385         /* Make all the stripes able to hold 'newsize' devices.
1386          * New slots in each stripe get 'page' set to a new page.
1387          *
1388          * This happens in stages:
1389          * 1/ create a new kmem_cache and allocate the required number of
1390          *    stripe_heads.
1391          * 2/ gather all the old stripe_heads and tranfer the pages across
1392          *    to the new stripe_heads.  This will have the side effect of
1393          *    freezing the array as once all stripe_heads have been collected,
1394          *    no IO will be possible.  Old stripe heads are freed once their
1395          *    pages have been transferred over, and the old kmem_cache is
1396          *    freed when all stripes are done.
1397          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
1398          *    we simple return a failre status - no need to clean anything up.
1399          * 4/ allocate new pages for the new slots in the new stripe_heads.
1400          *    If this fails, we don't bother trying the shrink the
1401          *    stripe_heads down again, we just leave them as they are.
1402          *    As each stripe_head is processed the new one is released into
1403          *    active service.
1404          *
1405          * Once step2 is started, we cannot afford to wait for a write,
1406          * so we use GFP_NOIO allocations.
1407          */
1408         struct stripe_head *osh, *nsh;
1409         LIST_HEAD(newstripes);
1410         struct disk_info *ndisks;
1411         unsigned long cpu;
1412         int err;
1413         struct kmem_cache *sc;
1414         int i;
1415
1416         if (newsize <= conf->pool_size)
1417                 return 0; /* never bother to shrink */
1418
1419         err = md_allow_write(conf->mddev);
1420         if (err)
1421                 return err;
1422
1423         /* Step 1 */
1424         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1425                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1426                                0, 0, NULL);
1427         if (!sc)
1428                 return -ENOMEM;
1429
1430         for (i = conf->max_nr_stripes; i; i--) {
1431                 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1432                 if (!nsh)
1433                         break;
1434
1435                 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1436
1437                 nsh->raid_conf = conf;
1438                 spin_lock_init(&nsh->lock);
1439                 #ifdef CONFIG_MULTICORE_RAID456
1440                 init_waitqueue_head(&nsh->ops.wait_for_ops);
1441                 #endif
1442
1443                 list_add(&nsh->lru, &newstripes);
1444         }
1445         if (i) {
1446                 /* didn't get enough, give up */
1447                 while (!list_empty(&newstripes)) {
1448                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
1449                         list_del(&nsh->lru);
1450                         kmem_cache_free(sc, nsh);
1451                 }
1452                 kmem_cache_destroy(sc);
1453                 return -ENOMEM;
1454         }
1455         /* Step 2 - Must use GFP_NOIO now.
1456          * OK, we have enough stripes, start collecting inactive
1457          * stripes and copying them over
1458          */
1459         list_for_each_entry(nsh, &newstripes, lru) {
1460                 spin_lock_irq(&conf->device_lock);
1461                 wait_event_lock_irq(conf->wait_for_stripe,
1462                                     !list_empty(&conf->inactive_list),
1463                                     conf->device_lock,
1464                                     unplug_slaves(conf->mddev)
1465                         );
1466                 osh = get_free_stripe(conf);
1467                 spin_unlock_irq(&conf->device_lock);
1468                 atomic_set(&nsh->count, 1);
1469                 for(i=0; i<conf->pool_size; i++)
1470                         nsh->dev[i].page = osh->dev[i].page;
1471                 for( ; i<newsize; i++)
1472                         nsh->dev[i].page = NULL;
1473                 kmem_cache_free(conf->slab_cache, osh);
1474         }
1475         kmem_cache_destroy(conf->slab_cache);
1476
1477         /* Step 3.
1478          * At this point, we are holding all the stripes so the array
1479          * is completely stalled, so now is a good time to resize
1480          * conf->disks and the scribble region
1481          */
1482         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1483         if (ndisks) {
1484                 for (i=0; i<conf->raid_disks; i++)
1485                         ndisks[i] = conf->disks[i];
1486                 kfree(conf->disks);
1487                 conf->disks = ndisks;
1488         } else
1489                 err = -ENOMEM;
1490
1491         get_online_cpus();
1492         conf->scribble_len = scribble_len(newsize);
1493         for_each_present_cpu(cpu) {
1494                 struct raid5_percpu *percpu;
1495                 void *scribble;
1496
1497                 percpu = per_cpu_ptr(conf->percpu, cpu);
1498                 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1499
1500                 if (scribble) {
1501                         kfree(percpu->scribble);
1502                         percpu->scribble = scribble;
1503                 } else {
1504                         err = -ENOMEM;
1505                         break;
1506                 }
1507         }
1508         put_online_cpus();
1509
1510         /* Step 4, return new stripes to service */
1511         while(!list_empty(&newstripes)) {
1512                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1513                 list_del_init(&nsh->lru);
1514
1515                 for (i=conf->raid_disks; i < newsize; i++)
1516                         if (nsh->dev[i].page == NULL) {
1517                                 struct page *p = alloc_page(GFP_NOIO);
1518                                 nsh->dev[i].page = p;
1519                                 if (!p)
1520                                         err = -ENOMEM;
1521                         }
1522                 release_stripe(nsh);
1523         }
1524         /* critical section pass, GFP_NOIO no longer needed */
1525
1526         conf->slab_cache = sc;
1527         conf->active_name = 1-conf->active_name;
1528         conf->pool_size = newsize;
1529         return err;
1530 }
1531
1532 static int drop_one_stripe(raid5_conf_t *conf)
1533 {
1534         struct stripe_head *sh;
1535
1536         spin_lock_irq(&conf->device_lock);
1537         sh = get_free_stripe(conf);
1538         spin_unlock_irq(&conf->device_lock);
1539         if (!sh)
1540                 return 0;
1541         BUG_ON(atomic_read(&sh->count));
1542         shrink_buffers(sh);
1543         kmem_cache_free(conf->slab_cache, sh);
1544         atomic_dec(&conf->active_stripes);
1545         return 1;
1546 }
1547
1548 static void shrink_stripes(raid5_conf_t *conf)
1549 {
1550         while (drop_one_stripe(conf))
1551                 ;
1552
1553         if (conf->slab_cache)
1554                 kmem_cache_destroy(conf->slab_cache);
1555         conf->slab_cache = NULL;
1556 }
1557
1558 static void raid5_end_read_request(struct bio * bi, int error)
1559 {
1560         struct stripe_head *sh = bi->bi_private;
1561         raid5_conf_t *conf = sh->raid_conf;
1562         int disks = sh->disks, i;
1563         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1564         char b[BDEVNAME_SIZE];
1565         mdk_rdev_t *rdev;
1566
1567
1568         for (i=0 ; i<disks; i++)
1569                 if (bi == &sh->dev[i].req)
1570                         break;
1571
1572         pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1573                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1574                 uptodate);
1575         if (i == disks) {
1576                 BUG();
1577                 return;
1578         }
1579
1580         if (uptodate) {
1581                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1582                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1583                         rdev = conf->disks[i].rdev;
1584                         printk_rl(KERN_INFO "md/raid:%s: read error corrected"
1585                                   " (%lu sectors at %llu on %s)\n",
1586                                   mdname(conf->mddev), STRIPE_SECTORS,
1587                                   (unsigned long long)(sh->sector
1588                                                        + rdev->data_offset),
1589                                   bdevname(rdev->bdev, b));
1590                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1591                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1592                 }
1593                 if (atomic_read(&conf->disks[i].rdev->read_errors))
1594                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
1595         } else {
1596                 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1597                 int retry = 0;
1598                 rdev = conf->disks[i].rdev;
1599
1600                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1601                 atomic_inc(&rdev->read_errors);
1602                 if (conf->mddev->degraded >= conf->max_degraded)
1603                         printk_rl(KERN_WARNING
1604                                   "md/raid:%s: read error not correctable "
1605                                   "(sector %llu on %s).\n",
1606                                   mdname(conf->mddev),
1607                                   (unsigned long long)(sh->sector
1608                                                        + rdev->data_offset),
1609                                   bdn);
1610                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1611                         /* Oh, no!!! */
1612                         printk_rl(KERN_WARNING
1613                                   "md/raid:%s: read error NOT corrected!! "
1614                                   "(sector %llu on %s).\n",
1615                                   mdname(conf->mddev),
1616                                   (unsigned long long)(sh->sector
1617                                                        + rdev->data_offset),
1618                                   bdn);
1619                 else if (atomic_read(&rdev->read_errors)
1620                          > conf->max_nr_stripes)
1621                         printk(KERN_WARNING
1622                                "md/raid:%s: Too many read errors, failing device %s.\n",
1623                                mdname(conf->mddev), bdn);
1624                 else
1625                         retry = 1;
1626                 if (retry)
1627                         set_bit(R5_ReadError, &sh->dev[i].flags);
1628                 else {
1629                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1630                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1631                         md_error(conf->mddev, rdev);
1632                 }
1633         }
1634         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1635         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1636         set_bit(STRIPE_HANDLE, &sh->state);
1637         release_stripe(sh);
1638 }
1639
1640 static void raid5_end_write_request(struct bio *bi, int error)
1641 {
1642         struct stripe_head *sh = bi->bi_private;
1643         raid5_conf_t *conf = sh->raid_conf;
1644         int disks = sh->disks, i;
1645         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1646
1647         for (i=0 ; i<disks; i++)
1648                 if (bi == &sh->dev[i].req)
1649                         break;
1650
1651         pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1652                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1653                 uptodate);
1654         if (i == disks) {
1655                 BUG();
1656                 return;
1657         }
1658
1659         if (!uptodate)
1660                 md_error(conf->mddev, conf->disks[i].rdev);
1661
1662         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1663         
1664         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1665         set_bit(STRIPE_HANDLE, &sh->state);
1666         release_stripe(sh);
1667 }
1668
1669
1670 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1671         
1672 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1673 {
1674         struct r5dev *dev = &sh->dev[i];
1675
1676         bio_init(&dev->req);
1677         dev->req.bi_io_vec = &dev->vec;
1678         dev->req.bi_vcnt++;
1679         dev->req.bi_max_vecs++;
1680         dev->vec.bv_page = dev->page;
1681         dev->vec.bv_len = STRIPE_SIZE;
1682         dev->vec.bv_offset = 0;
1683
1684         dev->req.bi_sector = sh->sector;
1685         dev->req.bi_private = sh;
1686
1687         dev->flags = 0;
1688         dev->sector = compute_blocknr(sh, i, previous);
1689 }
1690
1691 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1692 {
1693         char b[BDEVNAME_SIZE];
1694         raid5_conf_t *conf = mddev->private;
1695         pr_debug("raid456: error called\n");
1696
1697         if (!test_bit(Faulty, &rdev->flags)) {
1698                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1699                 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1700                         unsigned long flags;
1701                         spin_lock_irqsave(&conf->device_lock, flags);
1702                         mddev->degraded++;
1703                         spin_unlock_irqrestore(&conf->device_lock, flags);
1704                         /*
1705                          * if recovery was running, make sure it aborts.
1706                          */
1707                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1708                 }
1709                 set_bit(Faulty, &rdev->flags);
1710                 printk(KERN_ALERT
1711                        "md/raid:%s: Disk failure on %s, disabling device.\n"
1712                        KERN_ALERT
1713                        "md/raid:%s: Operation continuing on %d devices.\n",
1714                        mdname(mddev),
1715                        bdevname(rdev->bdev, b),
1716                        mdname(mddev),
1717                        conf->raid_disks - mddev->degraded);
1718         }
1719 }
1720
1721 /*
1722  * Input: a 'big' sector number,
1723  * Output: index of the data and parity disk, and the sector # in them.
1724  */
1725 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1726                                      int previous, int *dd_idx,
1727                                      struct stripe_head *sh)
1728 {
1729         sector_t stripe, stripe2;
1730         sector_t chunk_number;
1731         unsigned int chunk_offset;
1732         int pd_idx, qd_idx;
1733         int ddf_layout = 0;
1734         sector_t new_sector;
1735         int algorithm = previous ? conf->prev_algo
1736                                  : conf->algorithm;
1737         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1738                                          : conf->chunk_sectors;
1739         int raid_disks = previous ? conf->previous_raid_disks
1740                                   : conf->raid_disks;
1741         int data_disks = raid_disks - conf->max_degraded;
1742
1743         /* First compute the information on this sector */
1744
1745         /*
1746          * Compute the chunk number and the sector offset inside the chunk
1747          */
1748         chunk_offset = sector_div(r_sector, sectors_per_chunk);
1749         chunk_number = r_sector;
1750
1751         /*
1752          * Compute the stripe number
1753          */
1754         stripe = chunk_number;
1755         *dd_idx = sector_div(stripe, data_disks);
1756         stripe2 = stripe;
1757         /*
1758          * Select the parity disk based on the user selected algorithm.
1759          */
1760         pd_idx = qd_idx = ~0;
1761         switch(conf->level) {
1762         case 4:
1763                 pd_idx = data_disks;
1764                 break;
1765         case 5:
1766                 switch (algorithm) {
1767                 case ALGORITHM_LEFT_ASYMMETRIC:
1768                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
1769                         if (*dd_idx >= pd_idx)
1770                                 (*dd_idx)++;
1771                         break;
1772                 case ALGORITHM_RIGHT_ASYMMETRIC:
1773                         pd_idx = sector_div(stripe2, raid_disks);
1774                         if (*dd_idx >= pd_idx)
1775                                 (*dd_idx)++;
1776                         break;
1777                 case ALGORITHM_LEFT_SYMMETRIC:
1778                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
1779                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1780                         break;
1781                 case ALGORITHM_RIGHT_SYMMETRIC:
1782                         pd_idx = sector_div(stripe2, raid_disks);
1783                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1784                         break;
1785                 case ALGORITHM_PARITY_0:
1786                         pd_idx = 0;
1787                         (*dd_idx)++;
1788                         break;
1789                 case ALGORITHM_PARITY_N:
1790                         pd_idx = data_disks;
1791                         break;
1792                 default:
1793                         BUG();
1794                 }
1795                 break;
1796         case 6:
1797
1798                 switch (algorithm) {
1799                 case ALGORITHM_LEFT_ASYMMETRIC:
1800                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1801                         qd_idx = pd_idx + 1;
1802                         if (pd_idx == raid_disks-1) {
1803                                 (*dd_idx)++;    /* Q D D D P */
1804                                 qd_idx = 0;
1805                         } else if (*dd_idx >= pd_idx)
1806                                 (*dd_idx) += 2; /* D D P Q D */
1807                         break;
1808                 case ALGORITHM_RIGHT_ASYMMETRIC:
1809                         pd_idx = sector_div(stripe2, raid_disks);
1810                         qd_idx = pd_idx + 1;
1811                         if (pd_idx == raid_disks-1) {
1812                                 (*dd_idx)++;    /* Q D D D P */
1813                                 qd_idx = 0;
1814                         } else if (*dd_idx >= pd_idx)
1815                                 (*dd_idx) += 2; /* D D P Q D */
1816                         break;
1817                 case ALGORITHM_LEFT_SYMMETRIC:
1818                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1819                         qd_idx = (pd_idx + 1) % raid_disks;
1820                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1821                         break;
1822                 case ALGORITHM_RIGHT_SYMMETRIC:
1823                         pd_idx = sector_div(stripe2, raid_disks);
1824                         qd_idx = (pd_idx + 1) % raid_disks;
1825                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1826                         break;
1827
1828                 case ALGORITHM_PARITY_0:
1829                         pd_idx = 0;
1830                         qd_idx = 1;
1831                         (*dd_idx) += 2;
1832                         break;
1833                 case ALGORITHM_PARITY_N:
1834                         pd_idx = data_disks;
1835                         qd_idx = data_disks + 1;
1836                         break;
1837
1838                 case ALGORITHM_ROTATING_ZERO_RESTART:
1839                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
1840                          * of blocks for computing Q is different.
1841                          */
1842                         pd_idx = sector_div(stripe2, raid_disks);
1843                         qd_idx = pd_idx + 1;
1844                         if (pd_idx == raid_disks-1) {
1845                                 (*dd_idx)++;    /* Q D D D P */
1846                                 qd_idx = 0;
1847                         } else if (*dd_idx >= pd_idx)
1848                                 (*dd_idx) += 2; /* D D P Q D */
1849                         ddf_layout = 1;
1850                         break;
1851
1852                 case ALGORITHM_ROTATING_N_RESTART:
1853                         /* Same a left_asymmetric, by first stripe is
1854                          * D D D P Q  rather than
1855                          * Q D D D P
1856                          */
1857                         stripe2 += 1;
1858                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1859                         qd_idx = pd_idx + 1;
1860                         if (pd_idx == raid_disks-1) {
1861                                 (*dd_idx)++;    /* Q D D D P */
1862                                 qd_idx = 0;
1863                         } else if (*dd_idx >= pd_idx)
1864                                 (*dd_idx) += 2; /* D D P Q D */
1865                         ddf_layout = 1;
1866                         break;
1867
1868                 case ALGORITHM_ROTATING_N_CONTINUE:
1869                         /* Same as left_symmetric but Q is before P */
1870                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1871                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1872                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1873                         ddf_layout = 1;
1874                         break;
1875
1876                 case ALGORITHM_LEFT_ASYMMETRIC_6:
1877                         /* RAID5 left_asymmetric, with Q on last device */
1878                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1879                         if (*dd_idx >= pd_idx)
1880                                 (*dd_idx)++;
1881                         qd_idx = raid_disks - 1;
1882                         break;
1883
1884                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1885                         pd_idx = sector_div(stripe2, raid_disks-1);
1886                         if (*dd_idx >= pd_idx)
1887                                 (*dd_idx)++;
1888                         qd_idx = raid_disks - 1;
1889                         break;
1890
1891                 case ALGORITHM_LEFT_SYMMETRIC_6:
1892                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1893                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1894                         qd_idx = raid_disks - 1;
1895                         break;
1896
1897                 case ALGORITHM_RIGHT_SYMMETRIC_6:
1898                         pd_idx = sector_div(stripe2, raid_disks-1);
1899                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1900                         qd_idx = raid_disks - 1;
1901                         break;
1902
1903                 case ALGORITHM_PARITY_0_6:
1904                         pd_idx = 0;
1905                         (*dd_idx)++;
1906                         qd_idx = raid_disks - 1;
1907                         break;
1908
1909                 default:
1910                         BUG();
1911                 }
1912                 break;
1913         }
1914
1915         if (sh) {
1916                 sh->pd_idx = pd_idx;
1917                 sh->qd_idx = qd_idx;
1918                 sh->ddf_layout = ddf_layout;
1919         }
1920         /*
1921          * Finally, compute the new sector number
1922          */
1923         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1924         return new_sector;
1925 }
1926
1927
1928 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1929 {
1930         raid5_conf_t *conf = sh->raid_conf;
1931         int raid_disks = sh->disks;
1932         int data_disks = raid_disks - conf->max_degraded;
1933         sector_t new_sector = sh->sector, check;
1934         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1935                                          : conf->chunk_sectors;
1936         int algorithm = previous ? conf->prev_algo
1937                                  : conf->algorithm;
1938         sector_t stripe;
1939         int chunk_offset;
1940         sector_t chunk_number;
1941         int dummy1, dd_idx = i;
1942         sector_t r_sector;
1943         struct stripe_head sh2;
1944
1945
1946         chunk_offset = sector_div(new_sector, sectors_per_chunk);
1947         stripe = new_sector;
1948
1949         if (i == sh->pd_idx)
1950                 return 0;
1951         switch(conf->level) {
1952         case 4: break;
1953         case 5:
1954                 switch (algorithm) {
1955                 case ALGORITHM_LEFT_ASYMMETRIC:
1956                 case ALGORITHM_RIGHT_ASYMMETRIC:
1957                         if (i > sh->pd_idx)
1958                                 i--;
1959                         break;
1960                 case ALGORITHM_LEFT_SYMMETRIC:
1961                 case ALGORITHM_RIGHT_SYMMETRIC:
1962                         if (i < sh->pd_idx)
1963                                 i += raid_disks;
1964                         i -= (sh->pd_idx + 1);
1965                         break;
1966                 case ALGORITHM_PARITY_0:
1967                         i -= 1;
1968                         break;
1969                 case ALGORITHM_PARITY_N:
1970                         break;
1971                 default:
1972                         BUG();
1973                 }
1974                 break;
1975         case 6:
1976                 if (i == sh->qd_idx)
1977                         return 0; /* It is the Q disk */
1978                 switch (algorithm) {
1979                 case ALGORITHM_LEFT_ASYMMETRIC:
1980                 case ALGORITHM_RIGHT_ASYMMETRIC:
1981                 case ALGORITHM_ROTATING_ZERO_RESTART:
1982                 case ALGORITHM_ROTATING_N_RESTART:
1983                         if (sh->pd_idx == raid_disks-1)
1984                                 i--;    /* Q D D D P */
1985                         else if (i > sh->pd_idx)
1986                                 i -= 2; /* D D P Q D */
1987                         break;
1988                 case ALGORITHM_LEFT_SYMMETRIC:
1989                 case ALGORITHM_RIGHT_SYMMETRIC:
1990                         if (sh->pd_idx == raid_disks-1)
1991                                 i--; /* Q D D D P */
1992                         else {
1993                                 /* D D P Q D */
1994                                 if (i < sh->pd_idx)
1995                                         i += raid_disks;
1996                                 i -= (sh->pd_idx + 2);
1997                         }
1998                         break;
1999                 case ALGORITHM_PARITY_0:
2000                         i -= 2;
2001                         break;
2002                 case ALGORITHM_PARITY_N:
2003                         break;
2004                 case ALGORITHM_ROTATING_N_CONTINUE:
2005                         /* Like left_symmetric, but P is before Q */
2006                         if (sh->pd_idx == 0)
2007                                 i--;    /* P D D D Q */
2008                         else {
2009                                 /* D D Q P D */
2010                                 if (i < sh->pd_idx)
2011                                         i += raid_disks;
2012                                 i -= (sh->pd_idx + 1);
2013                         }
2014                         break;
2015                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2016                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2017                         if (i > sh->pd_idx)
2018                                 i--;
2019                         break;
2020                 case ALGORITHM_LEFT_SYMMETRIC_6:
2021                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2022                         if (i < sh->pd_idx)
2023                                 i += data_disks + 1;
2024                         i -= (sh->pd_idx + 1);
2025                         break;
2026                 case ALGORITHM_PARITY_0_6:
2027                         i -= 1;
2028                         break;
2029                 default:
2030                         BUG();
2031                 }
2032                 break;
2033         }
2034
2035         chunk_number = stripe * data_disks + i;
2036         r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2037
2038         check = raid5_compute_sector(conf, r_sector,
2039                                      previous, &dummy1, &sh2);
2040         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2041                 || sh2.qd_idx != sh->qd_idx) {
2042                 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2043                        mdname(conf->mddev));
2044                 return 0;
2045         }
2046         return r_sector;
2047 }
2048
2049
2050 static void
2051 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2052                          int rcw, int expand)
2053 {
2054         int i, pd_idx = sh->pd_idx, disks = sh->disks;
2055         raid5_conf_t *conf = sh->raid_conf;
2056         int level = conf->level;
2057
2058         if (rcw) {
2059                 /* if we are not expanding this is a proper write request, and
2060                  * there will be bios with new data to be drained into the
2061                  * stripe cache
2062                  */
2063                 if (!expand) {
2064                         sh->reconstruct_state = reconstruct_state_drain_run;
2065                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2066                 } else
2067                         sh->reconstruct_state = reconstruct_state_run;
2068
2069                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2070
2071                 for (i = disks; i--; ) {
2072                         struct r5dev *dev = &sh->dev[i];
2073
2074                         if (dev->towrite) {
2075                                 set_bit(R5_LOCKED, &dev->flags);
2076                                 set_bit(R5_Wantdrain, &dev->flags);
2077                                 if (!expand)
2078                                         clear_bit(R5_UPTODATE, &dev->flags);
2079                                 s->locked++;
2080                         }
2081                 }
2082                 if (s->locked + conf->max_degraded == disks)
2083                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2084                                 atomic_inc(&conf->pending_full_writes);
2085         } else {
2086                 BUG_ON(level == 6);
2087                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2088                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2089
2090                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2091                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2092                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2093                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2094
2095                 for (i = disks; i--; ) {
2096                         struct r5dev *dev = &sh->dev[i];
2097                         if (i == pd_idx)
2098                                 continue;
2099
2100                         if (dev->towrite &&
2101                             (test_bit(R5_UPTODATE, &dev->flags) ||
2102                              test_bit(R5_Wantcompute, &dev->flags))) {
2103                                 set_bit(R5_Wantdrain, &dev->flags);
2104                                 set_bit(R5_LOCKED, &dev->flags);
2105                                 clear_bit(R5_UPTODATE, &dev->flags);
2106                                 s->locked++;
2107                         }
2108                 }
2109         }
2110
2111         /* keep the parity disk(s) locked while asynchronous operations
2112          * are in flight
2113          */
2114         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2115         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2116         s->locked++;
2117
2118         if (level == 6) {
2119                 int qd_idx = sh->qd_idx;
2120                 struct r5dev *dev = &sh->dev[qd_idx];
2121
2122                 set_bit(R5_LOCKED, &dev->flags);
2123                 clear_bit(R5_UPTODATE, &dev->flags);
2124                 s->locked++;
2125         }
2126
2127         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2128                 __func__, (unsigned long long)sh->sector,
2129                 s->locked, s->ops_request);
2130 }
2131
2132 /*
2133  * Each stripe/dev can have one or more bion attached.
2134  * toread/towrite point to the first in a chain.
2135  * The bi_next chain must be in order.
2136  */
2137 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2138 {
2139         struct bio **bip;
2140         raid5_conf_t *conf = sh->raid_conf;
2141         int firstwrite=0;
2142
2143         pr_debug("adding bh b#%llu to stripe s#%llu\n",
2144                 (unsigned long long)bi->bi_sector,
2145                 (unsigned long long)sh->sector);
2146
2147
2148         spin_lock(&sh->lock);
2149         spin_lock_irq(&conf->device_lock);
2150         if (forwrite) {
2151                 bip = &sh->dev[dd_idx].towrite;
2152                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2153                         firstwrite = 1;
2154         } else
2155                 bip = &sh->dev[dd_idx].toread;
2156         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2157                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2158                         goto overlap;
2159                 bip = & (*bip)->bi_next;
2160         }
2161         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2162                 goto overlap;
2163
2164         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2165         if (*bip)
2166                 bi->bi_next = *bip;
2167         *bip = bi;
2168         bi->bi_phys_segments++;
2169         spin_unlock_irq(&conf->device_lock);
2170         spin_unlock(&sh->lock);
2171
2172         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2173                 (unsigned long long)bi->bi_sector,
2174                 (unsigned long long)sh->sector, dd_idx);
2175
2176         if (conf->mddev->bitmap && firstwrite) {
2177                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2178                                   STRIPE_SECTORS, 0);
2179                 sh->bm_seq = conf->seq_flush+1;
2180                 set_bit(STRIPE_BIT_DELAY, &sh->state);
2181         }
2182
2183         if (forwrite) {
2184                 /* check if page is covered */
2185                 sector_t sector = sh->dev[dd_idx].sector;
2186                 for (bi=sh->dev[dd_idx].towrite;
2187                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2188                              bi && bi->bi_sector <= sector;
2189                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2190                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2191                                 sector = bi->bi_sector + (bi->bi_size>>9);
2192                 }
2193                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2194                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2195         }
2196         return 1;
2197
2198  overlap:
2199         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2200         spin_unlock_irq(&conf->device_lock);
2201         spin_unlock(&sh->lock);
2202         return 0;
2203 }
2204
2205 static void end_reshape(raid5_conf_t *conf);
2206
2207 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
2208                             struct stripe_head *sh)
2209 {
2210         int sectors_per_chunk =
2211                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2212         int dd_idx;
2213         int chunk_offset = sector_div(stripe, sectors_per_chunk);
2214         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2215
2216         raid5_compute_sector(conf,
2217                              stripe * (disks - conf->max_degraded)
2218                              *sectors_per_chunk + chunk_offset,
2219                              previous,
2220                              &dd_idx, sh);
2221 }
2222
2223 static void
2224 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2225                                 struct stripe_head_state *s, int disks,
2226                                 struct bio **return_bi)
2227 {
2228         int i;
2229         for (i = disks; i--; ) {
2230                 struct bio *bi;
2231                 int bitmap_end = 0;
2232
2233                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2234                         mdk_rdev_t *rdev;
2235                         rcu_read_lock();
2236                         rdev = rcu_dereference(conf->disks[i].rdev);
2237                         if (rdev && test_bit(In_sync, &rdev->flags))
2238                                 /* multiple read failures in one stripe */
2239                                 md_error(conf->mddev, rdev);
2240                         rcu_read_unlock();
2241                 }
2242                 spin_lock_irq(&conf->device_lock);
2243                 /* fail all writes first */
2244                 bi = sh->dev[i].towrite;
2245                 sh->dev[i].towrite = NULL;
2246                 if (bi) {
2247                         s->to_write--;
2248                         bitmap_end = 1;
2249                 }
2250
2251                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2252                         wake_up(&conf->wait_for_overlap);
2253
2254                 while (bi && bi->bi_sector <
2255                         sh->dev[i].sector + STRIPE_SECTORS) {
2256                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2257                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2258                         if (!raid5_dec_bi_phys_segments(bi)) {
2259                                 md_write_end(conf->mddev);
2260                                 bi->bi_next = *return_bi;
2261                                 *return_bi = bi;
2262                         }
2263                         bi = nextbi;
2264                 }
2265                 /* and fail all 'written' */
2266                 bi = sh->dev[i].written;
2267                 sh->dev[i].written = NULL;
2268                 if (bi) bitmap_end = 1;
2269                 while (bi && bi->bi_sector <
2270                        sh->dev[i].sector + STRIPE_SECTORS) {
2271                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2272                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2273                         if (!raid5_dec_bi_phys_segments(bi)) {
2274                                 md_write_end(conf->mddev);
2275                                 bi->bi_next = *return_bi;
2276                                 *return_bi = bi;
2277                         }
2278                         bi = bi2;
2279                 }
2280
2281                 /* fail any reads if this device is non-operational and
2282                  * the data has not reached the cache yet.
2283                  */
2284                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2285                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2286                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
2287                         bi = sh->dev[i].toread;
2288                         sh->dev[i].toread = NULL;
2289                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2290                                 wake_up(&conf->wait_for_overlap);
2291                         if (bi) s->to_read--;
2292                         while (bi && bi->bi_sector <
2293                                sh->dev[i].sector + STRIPE_SECTORS) {
2294                                 struct bio *nextbi =
2295                                         r5_next_bio(bi, sh->dev[i].sector);
2296                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2297                                 if (!raid5_dec_bi_phys_segments(bi)) {
2298                                         bi->bi_next = *return_bi;
2299                                         *return_bi = bi;
2300                                 }
2301                                 bi = nextbi;
2302                         }
2303                 }
2304                 spin_unlock_irq(&conf->device_lock);
2305                 if (bitmap_end)
2306                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2307                                         STRIPE_SECTORS, 0, 0);
2308         }
2309
2310         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2311                 if (atomic_dec_and_test(&conf->pending_full_writes))
2312                         md_wakeup_thread(conf->mddev->thread);
2313 }
2314
2315 /* fetch_block5 - checks the given member device to see if its data needs
2316  * to be read or computed to satisfy a request.
2317  *
2318  * Returns 1 when no more member devices need to be checked, otherwise returns
2319  * 0 to tell the loop in handle_stripe_fill5 to continue
2320  */
2321 static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s,
2322                         int disk_idx, int disks)
2323 {
2324         struct r5dev *dev = &sh->dev[disk_idx];
2325         struct r5dev *failed_dev = &sh->dev[s->failed_num];
2326
2327         /* is the data in this block needed, and can we get it? */
2328         if (!test_bit(R5_LOCKED, &dev->flags) &&
2329             !test_bit(R5_UPTODATE, &dev->flags) &&
2330             (dev->toread ||
2331              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2332              s->syncing || s->expanding ||
2333              (s->failed &&
2334               (failed_dev->toread ||
2335                (failed_dev->towrite &&
2336                 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) {
2337                 /* We would like to get this block, possibly by computing it,
2338                  * otherwise read it if the backing disk is insync
2339                  */
2340                 if ((s->uptodate == disks - 1) &&
2341                     (s->failed && disk_idx == s->failed_num)) {
2342                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2343                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2344                         set_bit(R5_Wantcompute, &dev->flags);
2345                         sh->ops.target = disk_idx;
2346                         sh->ops.target2 = -1;
2347                         s->req_compute = 1;
2348                         /* Careful: from this point on 'uptodate' is in the eye
2349                          * of raid_run_ops which services 'compute' operations
2350                          * before writes. R5_Wantcompute flags a block that will
2351                          * be R5_UPTODATE by the time it is needed for a
2352                          * subsequent operation.
2353                          */
2354                         s->uptodate++;
2355                         return 1; /* uptodate + compute == disks */
2356                 } else if (test_bit(R5_Insync, &dev->flags)) {
2357                         set_bit(R5_LOCKED, &dev->flags);
2358                         set_bit(R5_Wantread, &dev->flags);
2359                         s->locked++;
2360                         pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2361                                 s->syncing);
2362                 }
2363         }
2364
2365         return 0;
2366 }
2367
2368 /**
2369  * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2370  */
2371 static void handle_stripe_fill5(struct stripe_head *sh,
2372                         struct stripe_head_state *s, int disks)
2373 {
2374         int i;
2375
2376         /* look for blocks to read/compute, skip this if a compute
2377          * is already in flight, or if the stripe contents are in the
2378          * midst of changing due to a write
2379          */
2380         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2381             !sh->reconstruct_state)
2382                 for (i = disks; i--; )
2383                         if (fetch_block5(sh, s, i, disks))
2384                                 break;
2385         set_bit(STRIPE_HANDLE, &sh->state);
2386 }
2387
2388 /* fetch_block6 - checks the given member device to see if its data needs
2389  * to be read or computed to satisfy a request.
2390  *
2391  * Returns 1 when no more member devices need to be checked, otherwise returns
2392  * 0 to tell the loop in handle_stripe_fill6 to continue
2393  */
2394 static int fetch_block6(struct stripe_head *sh, struct stripe_head_state *s,
2395                          struct r6_state *r6s, int disk_idx, int disks)
2396 {
2397         struct r5dev *dev = &sh->dev[disk_idx];
2398         struct r5dev *fdev[2] = { &sh->dev[r6s->failed_num[0]],
2399                                   &sh->dev[r6s->failed_num[1]] };
2400
2401         if (!test_bit(R5_LOCKED, &dev->flags) &&
2402             !test_bit(R5_UPTODATE, &dev->flags) &&
2403             (dev->toread ||
2404              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2405              s->syncing || s->expanding ||
2406              (s->failed >= 1 &&
2407               (fdev[0]->toread || s->to_write)) ||
2408              (s->failed >= 2 &&
2409               (fdev[1]->toread || s->to_write)))) {
2410                 /* we would like to get this block, possibly by computing it,
2411                  * otherwise read it if the backing disk is insync
2412                  */
2413                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2414                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2415                 if ((s->uptodate == disks - 1) &&
2416                     (s->failed && (disk_idx == r6s->failed_num[0] ||
2417                                    disk_idx == r6s->failed_num[1]))) {
2418                         /* have disk failed, and we're requested to fetch it;
2419                          * do compute it
2420                          */
2421                         pr_debug("Computing stripe %llu block %d\n",
2422                                (unsigned long long)sh->sector, disk_idx);
2423                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2424                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2425                         set_bit(R5_Wantcompute, &dev->flags);
2426                         sh->ops.target = disk_idx;
2427                         sh->ops.target2 = -1; /* no 2nd target */
2428                         s->req_compute = 1;
2429                         s->uptodate++;
2430                         return 1;
2431                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2432                         /* Computing 2-failure is *very* expensive; only
2433                          * do it if failed >= 2
2434                          */
2435                         int other;
2436                         for (other = disks; other--; ) {
2437                                 if (other == disk_idx)
2438                                         continue;
2439                                 if (!test_bit(R5_UPTODATE,
2440                                       &sh->dev[other].flags))
2441                                         break;
2442                         }
2443                         BUG_ON(other < 0);
2444                         pr_debug("Computing stripe %llu blocks %d,%d\n",
2445                                (unsigned long long)sh->sector,
2446                                disk_idx, other);
2447                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2448                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2449                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2450                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
2451                         sh->ops.target = disk_idx;
2452                         sh->ops.target2 = other;
2453                         s->uptodate += 2;
2454                         s->req_compute = 1;
2455                         return 1;
2456                 } else if (test_bit(R5_Insync, &dev->flags)) {
2457                         set_bit(R5_LOCKED, &dev->flags);
2458                         set_bit(R5_Wantread, &dev->flags);
2459                         s->locked++;
2460                         pr_debug("Reading block %d (sync=%d)\n",
2461                                 disk_idx, s->syncing);
2462                 }
2463         }
2464
2465         return 0;
2466 }
2467
2468 /**
2469  * handle_stripe_fill6 - read or compute data to satisfy pending requests.
2470  */
2471 static void handle_stripe_fill6(struct stripe_head *sh,
2472                         struct stripe_head_state *s, struct r6_state *r6s,
2473                         int disks)
2474 {
2475         int i;
2476
2477         /* look for blocks to read/compute, skip this if a compute
2478          * is already in flight, or if the stripe contents are in the
2479          * midst of changing due to a write
2480          */
2481         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2482             !sh->reconstruct_state)
2483                 for (i = disks; i--; )
2484                         if (fetch_block6(sh, s, r6s, i, disks))
2485                                 break;
2486         set_bit(STRIPE_HANDLE, &sh->state);
2487 }
2488
2489
2490 /* handle_stripe_clean_event
2491  * any written block on an uptodate or failed drive can be returned.
2492  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2493  * never LOCKED, so we don't need to test 'failed' directly.
2494  */
2495 static void handle_stripe_clean_event(raid5_conf_t *conf,
2496         struct stripe_head *sh, int disks, struct bio **return_bi)
2497 {
2498         int i;
2499         struct r5dev *dev;
2500
2501         for (i = disks; i--; )
2502                 if (sh->dev[i].written) {
2503                         dev = &sh->dev[i];
2504                         if (!test_bit(R5_LOCKED, &dev->flags) &&
2505                                 test_bit(R5_UPTODATE, &dev->flags)) {
2506                                 /* We can return any write requests */
2507                                 struct bio *wbi, *wbi2;
2508                                 int bitmap_end = 0;
2509                                 pr_debug("Return write for disc %d\n", i);
2510                                 spin_lock_irq(&conf->device_lock);
2511                                 wbi = dev->written;
2512                                 dev->written = NULL;
2513                                 while (wbi && wbi->bi_sector <
2514                                         dev->sector + STRIPE_SECTORS) {
2515                                         wbi2 = r5_next_bio(wbi, dev->sector);
2516                                         if (!raid5_dec_bi_phys_segments(wbi)) {
2517                                                 md_write_end(conf->mddev);
2518                                                 wbi->bi_next = *return_bi;
2519                                                 *return_bi = wbi;
2520                                         }
2521                                         wbi = wbi2;
2522                                 }
2523                                 if (dev->towrite == NULL)
2524                                         bitmap_end = 1;
2525                                 spin_unlock_irq(&conf->device_lock);
2526                                 if (bitmap_end)
2527                                         bitmap_endwrite(conf->mddev->bitmap,
2528                                                         sh->sector,
2529                                                         STRIPE_SECTORS,
2530                                          !test_bit(STRIPE_DEGRADED, &sh->state),
2531                                                         0);
2532                         }
2533                 }
2534
2535         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2536                 if (atomic_dec_and_test(&conf->pending_full_writes))
2537                         md_wakeup_thread(conf->mddev->thread);
2538 }
2539
2540 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2541                 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2542 {
2543         int rmw = 0, rcw = 0, i;
2544         for (i = disks; i--; ) {
2545                 /* would I have to read this buffer for read_modify_write */
2546                 struct r5dev *dev = &sh->dev[i];
2547                 if ((dev->towrite || i == sh->pd_idx) &&
2548                     !test_bit(R5_LOCKED, &dev->flags) &&
2549                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2550                       test_bit(R5_Wantcompute, &dev->flags))) {
2551                         if (test_bit(R5_Insync, &dev->flags))
2552                                 rmw++;
2553                         else
2554                                 rmw += 2*disks;  /* cannot read it */
2555                 }
2556                 /* Would I have to read this buffer for reconstruct_write */
2557                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2558                     !test_bit(R5_LOCKED, &dev->flags) &&
2559                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2560                     test_bit(R5_Wantcompute, &dev->flags))) {
2561                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2562                         else
2563                                 rcw += 2*disks;
2564                 }
2565         }
2566         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2567                 (unsigned long long)sh->sector, rmw, rcw);
2568         set_bit(STRIPE_HANDLE, &sh->state);
2569         if (rmw < rcw && rmw > 0)
2570                 /* prefer read-modify-write, but need to get some data */
2571                 for (i = disks; i--; ) {
2572                         struct r5dev *dev = &sh->dev[i];
2573                         if ((dev->towrite || i == sh->pd_idx) &&
2574                             !test_bit(R5_LOCKED, &dev->flags) &&
2575                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2576                             test_bit(R5_Wantcompute, &dev->flags)) &&
2577                             test_bit(R5_Insync, &dev->flags)) {
2578                                 if (
2579                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2580                                         pr_debug("Read_old block "
2581                                                 "%d for r-m-w\n", i);
2582                                         set_bit(R5_LOCKED, &dev->flags);
2583                                         set_bit(R5_Wantread, &dev->flags);
2584                                         s->locked++;
2585                                 } else {
2586                                         set_bit(STRIPE_DELAYED, &sh->state);
2587                                         set_bit(STRIPE_HANDLE, &sh->state);
2588                                 }
2589                         }
2590                 }
2591         if (rcw <= rmw && rcw > 0)
2592                 /* want reconstruct write, but need to get some data */
2593                 for (i = disks; i--; ) {
2594                         struct r5dev *dev = &sh->dev[i];
2595                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2596                             i != sh->pd_idx &&
2597                             !test_bit(R5_LOCKED, &dev->flags) &&
2598                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2599                             test_bit(R5_Wantcompute, &dev->flags)) &&
2600                             test_bit(R5_Insync, &dev->flags)) {
2601                                 if (
2602                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2603                                         pr_debug("Read_old block "
2604                                                 "%d for Reconstruct\n", i);
2605                                         set_bit(R5_LOCKED, &dev->flags);
2606                                         set_bit(R5_Wantread, &dev->flags);
2607                                         s->locked++;
2608                                 } else {
2609                                         set_bit(STRIPE_DELAYED, &sh->state);
2610                                         set_bit(STRIPE_HANDLE, &sh->state);
2611                                 }
2612                         }
2613                 }
2614         /* now if nothing is locked, and if we have enough data,
2615          * we can start a write request
2616          */
2617         /* since handle_stripe can be called at any time we need to handle the
2618          * case where a compute block operation has been submitted and then a
2619          * subsequent call wants to start a write request.  raid_run_ops only
2620          * handles the case where compute block and reconstruct are requested
2621          * simultaneously.  If this is not the case then new writes need to be
2622          * held off until the compute completes.
2623          */
2624         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2625             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2626             !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2627                 schedule_reconstruction(sh, s, rcw == 0, 0);
2628 }
2629
2630 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2631                 struct stripe_head *sh, struct stripe_head_state *s,
2632                 struct r6_state *r6s, int disks)
2633 {
2634         int rcw = 0, pd_idx = sh->pd_idx, i;
2635         int qd_idx = sh->qd_idx;
2636
2637         set_bit(STRIPE_HANDLE, &sh->state);
2638         for (i = disks; i--; ) {
2639                 struct r5dev *dev = &sh->dev[i];
2640                 /* check if we haven't enough data */
2641                 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2642                     i != pd_idx && i != qd_idx &&
2643                     !test_bit(R5_LOCKED, &dev->flags) &&
2644                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2645                       test_bit(R5_Wantcompute, &dev->flags))) {
2646                         rcw++;
2647                         if (!test_bit(R5_Insync, &dev->flags))
2648                                 continue; /* it's a failed drive */
2649
2650                         if (
2651                           test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2652                                 pr_debug("Read_old stripe %llu "
2653                                         "block %d for Reconstruct\n",
2654                                      (unsigned long long)sh->sector, i);
2655                                 set_bit(R5_LOCKED, &dev->flags);
2656                                 set_bit(R5_Wantread, &dev->flags);
2657                                 s->locked++;
2658                         } else {
2659                                 pr_debug("Request delayed stripe %llu "
2660                                         "block %d for Reconstruct\n",
2661                                      (unsigned long long)sh->sector, i);
2662                                 set_bit(STRIPE_DELAYED, &sh->state);
2663                                 set_bit(STRIPE_HANDLE, &sh->state);
2664                         }
2665                 }
2666         }
2667         /* now if nothing is locked, and if we have enough data, we can start a
2668          * write request
2669          */
2670         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2671             s->locked == 0 && rcw == 0 &&
2672             !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2673                 schedule_reconstruction(sh, s, 1, 0);
2674         }
2675 }
2676
2677 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2678                                 struct stripe_head_state *s, int disks)
2679 {
2680         struct r5dev *dev = NULL;
2681
2682         set_bit(STRIPE_HANDLE, &sh->state);
2683
2684         switch (sh->check_state) {
2685         case check_state_idle:
2686                 /* start a new check operation if there are no failures */
2687                 if (s->failed == 0) {
2688                         BUG_ON(s->uptodate != disks);
2689                         sh->check_state = check_state_run;
2690                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
2691                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2692                         s->uptodate--;
2693                         break;
2694                 }
2695                 dev = &sh->dev[s->failed_num];
2696                 /* fall through */
2697         case check_state_compute_result:
2698                 sh->check_state = check_state_idle;
2699                 if (!dev)
2700                         dev = &sh->dev[sh->pd_idx];
2701
2702                 /* check that a write has not made the stripe insync */
2703                 if (test_bit(STRIPE_INSYNC, &sh->state))
2704                         break;
2705
2706                 /* either failed parity check, or recovery is happening */
2707                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2708                 BUG_ON(s->uptodate != disks);
2709
2710                 set_bit(R5_LOCKED, &dev->flags);
2711                 s->locked++;
2712                 set_bit(R5_Wantwrite, &dev->flags);
2713
2714                 clear_bit(STRIPE_DEGRADED, &sh->state);
2715                 set_bit(STRIPE_INSYNC, &sh->state);
2716                 break;
2717         case check_state_run:
2718                 break; /* we will be called again upon completion */
2719         case check_state_check_result:
2720                 sh->check_state = check_state_idle;
2721
2722                 /* if a failure occurred during the check operation, leave
2723                  * STRIPE_INSYNC not set and let the stripe be handled again
2724                  */
2725                 if (s->failed)
2726                         break;
2727
2728                 /* handle a successful check operation, if parity is correct
2729                  * we are done.  Otherwise update the mismatch count and repair
2730                  * parity if !MD_RECOVERY_CHECK
2731                  */
2732                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2733                         /* parity is correct (on disc,
2734                          * not in buffer any more)
2735                          */
2736                         set_bit(STRIPE_INSYNC, &sh->state);
2737                 else {
2738                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2739                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2740                                 /* don't try to repair!! */
2741                                 set_bit(STRIPE_INSYNC, &sh->state);
2742                         else {
2743                                 sh->check_state = check_state_compute_run;
2744                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2745                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2746                                 set_bit(R5_Wantcompute,
2747                                         &sh->dev[sh->pd_idx].flags);
2748                                 sh->ops.target = sh->pd_idx;
2749                                 sh->ops.target2 = -1;
2750                                 s->uptodate++;
2751                         }
2752                 }
2753                 break;
2754         case check_state_compute_run:
2755                 break;
2756         default:
2757                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2758                        __func__, sh->check_state,
2759                        (unsigned long long) sh->sector);
2760                 BUG();
2761         }
2762 }
2763
2764
2765 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2766                                   struct stripe_head_state *s,
2767                                   struct r6_state *r6s, int disks)
2768 {
2769         int pd_idx = sh->pd_idx;
2770         int qd_idx = sh->qd_idx;
2771         struct r5dev *dev;
2772
2773         set_bit(STRIPE_HANDLE, &sh->state);
2774
2775         BUG_ON(s->failed > 2);
2776
2777         /* Want to check and possibly repair P and Q.
2778          * However there could be one 'failed' device, in which
2779          * case we can only check one of them, possibly using the
2780          * other to generate missing data
2781          */
2782
2783         switch (sh->check_state) {
2784         case check_state_idle:
2785                 /* start a new check operation if there are < 2 failures */
2786                 if (s->failed == r6s->q_failed) {
2787                         /* The only possible failed device holds Q, so it
2788                          * makes sense to check P (If anything else were failed,
2789                          * we would have used P to recreate it).
2790                          */
2791                         sh->check_state = check_state_run;
2792                 }
2793                 if (!r6s->q_failed && s->failed < 2) {
2794                         /* Q is not failed, and we didn't use it to generate
2795                          * anything, so it makes sense to check it
2796                          */
2797                         if (sh->check_state == check_state_run)
2798                                 sh->check_state = check_state_run_pq;
2799                         else
2800                                 sh->check_state = check_state_run_q;
2801                 }
2802
2803                 /* discard potentially stale zero_sum_result */
2804                 sh->ops.zero_sum_result = 0;
2805
2806                 if (sh->check_state == check_state_run) {
2807                         /* async_xor_zero_sum destroys the contents of P */
2808                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2809                         s->uptodate--;
2810                 }
2811                 if (sh->check_state >= check_state_run &&
2812                     sh->check_state <= check_state_run_pq) {
2813                         /* async_syndrome_zero_sum preserves P and Q, so
2814                          * no need to mark them !uptodate here
2815                          */
2816                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
2817                         break;
2818                 }
2819
2820                 /* we have 2-disk failure */
2821                 BUG_ON(s->failed != 2);
2822                 /* fall through */
2823         case check_state_compute_result:
2824                 sh->check_state = check_state_idle;
2825
2826                 /* check that a write has not made the stripe insync */
2827                 if (test_bit(STRIPE_INSYNC, &sh->state))
2828                         break;
2829
2830                 /* now write out any block on a failed drive,
2831                  * or P or Q if they were recomputed
2832                  */
2833                 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2834                 if (s->failed == 2) {
2835                         dev = &sh->dev[r6s->failed_num[1]];
2836                         s->locked++;
2837                         set_bit(R5_LOCKED, &dev->flags);
2838                         set_bit(R5_Wantwrite, &dev->flags);
2839                 }
2840                 if (s->failed >= 1) {
2841                         dev = &sh->dev[r6s->failed_num[0]];
2842                         s->locked++;
2843                         set_bit(R5_LOCKED, &dev->flags);
2844                         set_bit(R5_Wantwrite, &dev->flags);
2845                 }
2846                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2847                         dev = &sh->dev[pd_idx];
2848                         s->locked++;
2849                         set_bit(R5_LOCKED, &dev->flags);
2850                         set_bit(R5_Wantwrite, &dev->flags);
2851                 }
2852                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2853                         dev = &sh->dev[qd_idx];
2854                         s->locked++;
2855                         set_bit(R5_LOCKED, &dev->flags);
2856                         set_bit(R5_Wantwrite, &dev->flags);
2857                 }
2858                 clear_bit(STRIPE_DEGRADED, &sh->state);
2859
2860                 set_bit(STRIPE_INSYNC, &sh->state);
2861                 break;
2862         case check_state_run:
2863         case check_state_run_q:
2864         case check_state_run_pq:
2865                 break; /* we will be called again upon completion */
2866         case check_state_check_result:
2867                 sh->check_state = check_state_idle;
2868
2869                 /* handle a successful check operation, if parity is correct
2870                  * we are done.  Otherwise update the mismatch count and repair
2871                  * parity if !MD_RECOVERY_CHECK
2872                  */
2873                 if (sh->ops.zero_sum_result == 0) {
2874                         /* both parities are correct */
2875                         if (!s->failed)
2876                                 set_bit(STRIPE_INSYNC, &sh->state);
2877                         else {
2878                                 /* in contrast to the raid5 case we can validate
2879                                  * parity, but still have a failure to write
2880                                  * back
2881                                  */
2882                                 sh->check_state = check_state_compute_result;
2883                                 /* Returning at this point means that we may go
2884                                  * off and bring p and/or q uptodate again so
2885                                  * we make sure to check zero_sum_result again
2886                                  * to verify if p or q need writeback
2887                                  */
2888                         }
2889                 } else {
2890                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2891                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2892                                 /* don't try to repair!! */
2893                                 set_bit(STRIPE_INSYNC, &sh->state);
2894                         else {
2895                                 int *target = &sh->ops.target;
2896
2897                                 sh->ops.target = -1;
2898                                 sh->ops.target2 = -1;
2899                                 sh->check_state = check_state_compute_run;
2900                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2901                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2902                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2903                                         set_bit(R5_Wantcompute,
2904                                                 &sh->dev[pd_idx].flags);
2905                                         *target = pd_idx;
2906                                         target = &sh->ops.target2;
2907                                         s->uptodate++;
2908                                 }
2909                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2910                                         set_bit(R5_Wantcompute,
2911                                                 &sh->dev[qd_idx].flags);
2912                                         *target = qd_idx;
2913                                         s->uptodate++;
2914                                 }
2915                         }
2916                 }
2917                 break;
2918         case check_state_compute_run:
2919                 break;
2920         default:
2921                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2922                        __func__, sh->check_state,
2923                        (unsigned long long) sh->sector);
2924                 BUG();
2925         }
2926 }
2927
2928 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2929                                 struct r6_state *r6s)
2930 {
2931         int i;
2932
2933         /* We have read all the blocks in this stripe and now we need to
2934          * copy some of them into a target stripe for expand.
2935          */
2936         struct dma_async_tx_descriptor *tx = NULL;
2937         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2938         for (i = 0; i < sh->disks; i++)
2939                 if (i != sh->pd_idx && i != sh->qd_idx) {
2940                         int dd_idx, j;
2941                         struct stripe_head *sh2;
2942                         struct async_submit_ctl submit;
2943
2944                         sector_t bn = compute_blocknr(sh, i, 1);
2945                         sector_t s = raid5_compute_sector(conf, bn, 0,
2946                                                           &dd_idx, NULL);
2947                         sh2 = get_active_stripe(conf, s, 0, 1, 1);
2948                         if (sh2 == NULL)
2949                                 /* so far only the early blocks of this stripe
2950                                  * have been requested.  When later blocks
2951                                  * get requested, we will try again
2952                                  */
2953                                 continue;
2954                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2955                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2956                                 /* must have already done this block */
2957                                 release_stripe(sh2);
2958                                 continue;
2959                         }
2960
2961                         /* place all the copies on one channel */
2962                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
2963                         tx = async_memcpy(sh2->dev[dd_idx].page,
2964                                           sh->dev[i].page, 0, 0, STRIPE_SIZE,
2965                                           &submit);
2966
2967                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2968                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2969                         for (j = 0; j < conf->raid_disks; j++)
2970                                 if (j != sh2->pd_idx &&
2971                                     (!r6s || j != sh2->qd_idx) &&
2972                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
2973                                         break;
2974                         if (j == conf->raid_disks) {
2975                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2976                                 set_bit(STRIPE_HANDLE, &sh2->state);
2977                         }
2978                         release_stripe(sh2);
2979
2980                 }
2981         /* done submitting copies, wait for them to complete */
2982         if (tx) {
2983                 async_tx_ack(tx);
2984                 dma_wait_for_async_tx(tx);
2985         }
2986 }
2987
2988
2989 /*
2990  * handle_stripe - do things to a stripe.
2991  *
2992  * We lock the stripe and then examine the state of various bits
2993  * to see what needs to be done.
2994  * Possible results:
2995  *    return some read request which now have data
2996  *    return some write requests which are safely on disc
2997  *    schedule a read on some buffers
2998  *    schedule a write of some buffers
2999  *    return confirmation of parity correctness
3000  *
3001  * buffers are taken off read_list or write_list, and bh_cache buffers
3002  * get BH_Lock set before the stripe lock is released.
3003  *
3004  */
3005
3006 static void handle_stripe5(struct stripe_head *sh)
3007 {
3008         raid5_conf_t *conf = sh->raid_conf;
3009   &nbs