Merge branch 'master' of git://git.kernel.org/pub/scm/linux/kernel/git/linville/wirel...
[linux-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                         if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
511                                 rw = WRITE_FUA;
512                         else
513                                 rw = WRITE;
514                 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
515                         rw = READ;
516                 else
517                         continue;
518
519                 bi = &sh->dev[i].req;
520
521                 bi->bi_rw = rw;
522                 if (rw == WRITE)
523                         bi->bi_end_io = raid5_end_write_request;
524                 else
525                         bi->bi_end_io = raid5_end_read_request;
526
527                 rcu_read_lock();
528                 rdev = rcu_dereference(conf->disks[i].rdev);
529                 if (rdev && test_bit(Faulty, &rdev->flags))
530                         rdev = NULL;
531                 if (rdev)
532                         atomic_inc(&rdev->nr_pending);
533                 rcu_read_unlock();
534
535                 if (rdev) {
536                         if (s->syncing || s->expanding || s->expanded)
537                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
538
539                         set_bit(STRIPE_IO_STARTED, &sh->state);
540
541                         bi->bi_bdev = rdev->bdev;
542                         pr_debug("%s: for %llu schedule op %ld on disc %d\n",
543                                 __func__, (unsigned long long)sh->sector,
544                                 bi->bi_rw, i);
545                         atomic_inc(&sh->count);
546                         bi->bi_sector = sh->sector + rdev->data_offset;
547                         bi->bi_flags = 1 << BIO_UPTODATE;
548                         bi->bi_vcnt = 1;
549                         bi->bi_max_vecs = 1;
550                         bi->bi_idx = 0;
551                         bi->bi_io_vec = &sh->dev[i].vec;
552                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
553                         bi->bi_io_vec[0].bv_offset = 0;
554                         bi->bi_size = STRIPE_SIZE;
555                         bi->bi_next = NULL;
556                         if (rw == WRITE &&
557                             test_bit(R5_ReWrite, &sh->dev[i].flags))
558                                 atomic_add(STRIPE_SECTORS,
559                                         &rdev->corrected_errors);
560                         generic_make_request(bi);
561                 } else {
562                         if (rw == WRITE)
563                                 set_bit(STRIPE_DEGRADED, &sh->state);
564                         pr_debug("skip op %ld on disc %d for sector %llu\n",
565                                 bi->bi_rw, i, (unsigned long long)sh->sector);
566                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
567                         set_bit(STRIPE_HANDLE, &sh->state);
568                 }
569         }
570 }
571
572 static struct dma_async_tx_descriptor *
573 async_copy_data(int frombio, struct bio *bio, struct page *page,
574         sector_t sector, struct dma_async_tx_descriptor *tx)
575 {
576         struct bio_vec *bvl;
577         struct page *bio_page;
578         int i;
579         int page_offset;
580         struct async_submit_ctl submit;
581         enum async_tx_flags flags = 0;
582
583         if (bio->bi_sector >= sector)
584                 page_offset = (signed)(bio->bi_sector - sector) * 512;
585         else
586                 page_offset = (signed)(sector - bio->bi_sector) * -512;
587
588         if (frombio)
589                 flags |= ASYNC_TX_FENCE;
590         init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
591
592         bio_for_each_segment(bvl, bio, i) {
593                 int len = bio_iovec_idx(bio, i)->bv_len;
594                 int clen;
595                 int b_offset = 0;
596
597                 if (page_offset < 0) {
598                         b_offset = -page_offset;
599                         page_offset += b_offset;
600                         len -= b_offset;
601                 }
602
603                 if (len > 0 && page_offset + len > STRIPE_SIZE)
604                         clen = STRIPE_SIZE - page_offset;
605                 else
606                         clen = len;
607
608                 if (clen > 0) {
609                         b_offset += bio_iovec_idx(bio, i)->bv_offset;
610                         bio_page = bio_iovec_idx(bio, i)->bv_page;
611                         if (frombio)
612                                 tx = async_memcpy(page, bio_page, page_offset,
613                                                   b_offset, clen, &submit);
614                         else
615                                 tx = async_memcpy(bio_page, page, b_offset,
616                                                   page_offset, clen, &submit);
617                 }
618                 /* chain the operations */
619                 submit.depend_tx = tx;
620
621                 if (clen < len) /* hit end of page */
622                         break;
623                 page_offset +=  len;
624         }
625
626         return tx;
627 }
628
629 static void ops_complete_biofill(void *stripe_head_ref)
630 {
631         struct stripe_head *sh = stripe_head_ref;
632         struct bio *return_bi = NULL;
633         raid5_conf_t *conf = sh->raid_conf;
634         int i;
635
636         pr_debug("%s: stripe %llu\n", __func__,
637                 (unsigned long long)sh->sector);
638
639         /* clear completed biofills */
640         spin_lock_irq(&conf->device_lock);
641         for (i = sh->disks; i--; ) {
642                 struct r5dev *dev = &sh->dev[i];
643
644                 /* acknowledge completion of a biofill operation */
645                 /* and check if we need to reply to a read request,
646                  * new R5_Wantfill requests are held off until
647                  * !STRIPE_BIOFILL_RUN
648                  */
649                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
650                         struct bio *rbi, *rbi2;
651
652                         BUG_ON(!dev->read);
653                         rbi = dev->read;
654                         dev->read = NULL;
655                         while (rbi && rbi->bi_sector <
656                                 dev->sector + STRIPE_SECTORS) {
657                                 rbi2 = r5_next_bio(rbi, dev->sector);
658                                 if (!raid5_dec_bi_phys_segments(rbi)) {
659                                         rbi->bi_next = return_bi;
660                                         return_bi = rbi;
661                                 }
662                                 rbi = rbi2;
663                         }
664                 }
665         }
666         spin_unlock_irq(&conf->device_lock);
667         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
668
669         return_io(return_bi);
670
671         set_bit(STRIPE_HANDLE, &sh->state);
672         release_stripe(sh);
673 }
674
675 static void ops_run_biofill(struct stripe_head *sh)
676 {
677         struct dma_async_tx_descriptor *tx = NULL;
678         raid5_conf_t *conf = sh->raid_conf;
679         struct async_submit_ctl submit;
680         int i;
681
682         pr_debug("%s: stripe %llu\n", __func__,
683                 (unsigned long long)sh->sector);
684
685         for (i = sh->disks; i--; ) {
686                 struct r5dev *dev = &sh->dev[i];
687                 if (test_bit(R5_Wantfill, &dev->flags)) {
688                         struct bio *rbi;
689                         spin_lock_irq(&conf->device_lock);
690                         dev->read = rbi = dev->toread;
691                         dev->toread = NULL;
692                         spin_unlock_irq(&conf->device_lock);
693                         while (rbi && rbi->bi_sector <
694                                 dev->sector + STRIPE_SECTORS) {
695                                 tx = async_copy_data(0, rbi, dev->page,
696                                         dev->sector, tx);
697                                 rbi = r5_next_bio(rbi, dev->sector);
698                         }
699                 }
700         }
701
702         atomic_inc(&sh->count);
703         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
704         async_trigger_callback(&submit);
705 }
706
707 static void mark_target_uptodate(struct stripe_head *sh, int target)
708 {
709         struct r5dev *tgt;
710
711         if (target < 0)
712                 return;
713
714         tgt = &sh->dev[target];
715         set_bit(R5_UPTODATE, &tgt->flags);
716         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
717         clear_bit(R5_Wantcompute, &tgt->flags);
718 }
719
720 static void ops_complete_compute(void *stripe_head_ref)
721 {
722         struct stripe_head *sh = stripe_head_ref;
723
724         pr_debug("%s: stripe %llu\n", __func__,
725                 (unsigned long long)sh->sector);
726
727         /* mark the computed target(s) as uptodate */
728         mark_target_uptodate(sh, sh->ops.target);
729         mark_target_uptodate(sh, sh->ops.target2);
730
731         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
732         if (sh->check_state == check_state_compute_run)
733                 sh->check_state = check_state_compute_result;
734         set_bit(STRIPE_HANDLE, &sh->state);
735         release_stripe(sh);
736 }
737
738 /* return a pointer to the address conversion region of the scribble buffer */
739 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
740                                  struct raid5_percpu *percpu)
741 {
742         return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
743 }
744
745 static struct dma_async_tx_descriptor *
746 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
747 {
748         int disks = sh->disks;
749         struct page **xor_srcs = percpu->scribble;
750         int target = sh->ops.target;
751         struct r5dev *tgt = &sh->dev[target];
752         struct page *xor_dest = tgt->page;
753         int count = 0;
754         struct dma_async_tx_descriptor *tx;
755         struct async_submit_ctl submit;
756         int i;
757
758         pr_debug("%s: stripe %llu block: %d\n",
759                 __func__, (unsigned long long)sh->sector, target);
760         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
761
762         for (i = disks; i--; )
763                 if (i != target)
764                         xor_srcs[count++] = sh->dev[i].page;
765
766         atomic_inc(&sh->count);
767
768         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
769                           ops_complete_compute, sh, to_addr_conv(sh, percpu));
770         if (unlikely(count == 1))
771                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
772         else
773                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
774
775         return tx;
776 }
777
778 /* set_syndrome_sources - populate source buffers for gen_syndrome
779  * @srcs - (struct page *) array of size sh->disks
780  * @sh - stripe_head to parse
781  *
782  * Populates srcs in proper layout order for the stripe and returns the
783  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
784  * destination buffer is recorded in srcs[count] and the Q destination
785  * is recorded in srcs[count+1]].
786  */
787 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
788 {
789         int disks = sh->disks;
790         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
791         int d0_idx = raid6_d0(sh);
792         int count;
793         int i;
794
795         for (i = 0; i < disks; i++)
796                 srcs[i] = NULL;
797
798         count = 0;
799         i = d0_idx;
800         do {
801                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
802
803                 srcs[slot] = sh->dev[i].page;
804                 i = raid6_next_disk(i, disks);
805         } while (i != d0_idx);
806
807         return syndrome_disks;
808 }
809
810 static struct dma_async_tx_descriptor *
811 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
812 {
813         int disks = sh->disks;
814         struct page **blocks = percpu->scribble;
815         int target;
816         int qd_idx = sh->qd_idx;
817         struct dma_async_tx_descriptor *tx;
818         struct async_submit_ctl submit;
819         struct r5dev *tgt;
820         struct page *dest;
821         int i;
822         int count;
823
824         if (sh->ops.target < 0)
825                 target = sh->ops.target2;
826         else if (sh->ops.target2 < 0)
827                 target = sh->ops.target;
828         else
829                 /* we should only have one valid target */
830                 BUG();
831         BUG_ON(target < 0);
832         pr_debug("%s: stripe %llu block: %d\n",
833                 __func__, (unsigned long long)sh->sector, target);
834
835         tgt = &sh->dev[target];
836         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
837         dest = tgt->page;
838
839         atomic_inc(&sh->count);
840
841         if (target == qd_idx) {
842                 count = set_syndrome_sources(blocks, sh);
843                 blocks[count] = NULL; /* regenerating p is not necessary */
844                 BUG_ON(blocks[count+1] != dest); /* q should already be set */
845                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
846                                   ops_complete_compute, sh,
847                                   to_addr_conv(sh, percpu));
848                 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
849         } else {
850                 /* Compute any data- or p-drive using XOR */
851                 count = 0;
852                 for (i = disks; i-- ; ) {
853                         if (i == target || i == qd_idx)
854                                 continue;
855                         blocks[count++] = sh->dev[i].page;
856                 }
857
858                 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
859                                   NULL, ops_complete_compute, sh,
860                                   to_addr_conv(sh, percpu));
861                 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
862         }
863
864         return tx;
865 }
866
867 static struct dma_async_tx_descriptor *
868 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
869 {
870         int i, count, disks = sh->disks;
871         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
872         int d0_idx = raid6_d0(sh);
873         int faila = -1, failb = -1;
874         int target = sh->ops.target;
875         int target2 = sh->ops.target2;
876         struct r5dev *tgt = &sh->dev[target];
877         struct r5dev *tgt2 = &sh->dev[target2];
878         struct dma_async_tx_descriptor *tx;
879         struct page **blocks = percpu->scribble;
880         struct async_submit_ctl submit;
881
882         pr_debug("%s: stripe %llu block1: %d block2: %d\n",
883                  __func__, (unsigned long long)sh->sector, target, target2);
884         BUG_ON(target < 0 || target2 < 0);
885         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
886         BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
887
888         /* we need to open-code set_syndrome_sources to handle the
889          * slot number conversion for 'faila' and 'failb'
890          */
891         for (i = 0; i < disks ; i++)
892                 blocks[i] = NULL;
893         count = 0;
894         i = d0_idx;
895         do {
896                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
897
898                 blocks[slot] = sh->dev[i].page;
899
900                 if (i == target)
901                         faila = slot;
902                 if (i == target2)
903                         failb = slot;
904                 i = raid6_next_disk(i, disks);
905         } while (i != d0_idx);
906
907         BUG_ON(faila == failb);
908         if (failb < faila)
909                 swap(faila, failb);
910         pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
911                  __func__, (unsigned long long)sh->sector, faila, failb);
912
913         atomic_inc(&sh->count);
914
915         if (failb == syndrome_disks+1) {
916                 /* Q disk is one of the missing disks */
917                 if (faila == syndrome_disks) {
918                         /* Missing P+Q, just recompute */
919                         init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
920                                           ops_complete_compute, sh,
921                                           to_addr_conv(sh, percpu));
922                         return async_gen_syndrome(blocks, 0, syndrome_disks+2,
923                                                   STRIPE_SIZE, &submit);
924                 } else {
925                         struct page *dest;
926                         int data_target;
927                         int qd_idx = sh->qd_idx;
928
929                         /* Missing D+Q: recompute D from P, then recompute Q */
930                         if (target == qd_idx)
931                                 data_target = target2;
932                         else
933                                 data_target = target;
934
935                         count = 0;
936                         for (i = disks; i-- ; ) {
937                                 if (i == data_target || i == qd_idx)
938                                         continue;
939                                 blocks[count++] = sh->dev[i].page;
940                         }
941                         dest = sh->dev[data_target].page;
942                         init_async_submit(&submit,
943                                           ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
944                                           NULL, NULL, NULL,
945                                           to_addr_conv(sh, percpu));
946                         tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
947                                        &submit);
948
949                         count = set_syndrome_sources(blocks, sh);
950                         init_async_submit(&submit, ASYNC_TX_FENCE, tx,
951                                           ops_complete_compute, sh,
952                                           to_addr_conv(sh, percpu));
953                         return async_gen_syndrome(blocks, 0, count+2,
954                                                   STRIPE_SIZE, &submit);
955                 }
956         } else {
957                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
958                                   ops_complete_compute, sh,
959                                   to_addr_conv(sh, percpu));
960                 if (failb == syndrome_disks) {
961                         /* We're missing D+P. */
962                         return async_raid6_datap_recov(syndrome_disks+2,
963                                                        STRIPE_SIZE, faila,
964                                                        blocks, &submit);
965                 } else {
966                         /* We're missing D+D. */
967                         return async_raid6_2data_recov(syndrome_disks+2,
968                                                        STRIPE_SIZE, faila, failb,
969                                                        blocks, &submit);
970                 }
971         }
972 }
973
974
975 static void ops_complete_prexor(void *stripe_head_ref)
976 {
977         struct stripe_head *sh = stripe_head_ref;
978
979         pr_debug("%s: stripe %llu\n", __func__,
980                 (unsigned long long)sh->sector);
981 }
982
983 static struct dma_async_tx_descriptor *
984 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
985                struct dma_async_tx_descriptor *tx)
986 {
987         int disks = sh->disks;
988         struct page **xor_srcs = percpu->scribble;
989         int count = 0, pd_idx = sh->pd_idx, i;
990         struct async_submit_ctl submit;
991
992         /* existing parity data subtracted */
993         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
994
995         pr_debug("%s: stripe %llu\n", __func__,
996                 (unsigned long long)sh->sector);
997
998         for (i = disks; i--; ) {
999                 struct r5dev *dev = &sh->dev[i];
1000                 /* Only process blocks that are known to be uptodate */
1001                 if (test_bit(R5_Wantdrain, &dev->flags))
1002                         xor_srcs[count++] = dev->page;
1003         }
1004
1005         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1006                           ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1007         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1008
1009         return tx;
1010 }
1011
1012 static struct dma_async_tx_descriptor *
1013 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1014 {
1015         int disks = sh->disks;
1016         int i;
1017
1018         pr_debug("%s: stripe %llu\n", __func__,
1019                 (unsigned long long)sh->sector);
1020
1021         for (i = disks; i--; ) {
1022                 struct r5dev *dev = &sh->dev[i];
1023                 struct bio *chosen;
1024
1025                 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1026                         struct bio *wbi;
1027
1028                         spin_lock(&sh->lock);
1029                         chosen = dev->towrite;
1030                         dev->towrite = NULL;
1031                         BUG_ON(dev->written);
1032                         wbi = dev->written = chosen;
1033                         spin_unlock(&sh->lock);
1034
1035                         while (wbi && wbi->bi_sector <
1036                                 dev->sector + STRIPE_SECTORS) {
1037                                 if (wbi->bi_rw & REQ_FUA)
1038                                         set_bit(R5_WantFUA, &dev->flags);
1039                                 tx = async_copy_data(1, wbi, dev->page,
1040                                         dev->sector, tx);
1041                                 wbi = r5_next_bio(wbi, dev->sector);
1042                         }
1043                 }
1044         }
1045
1046         return tx;
1047 }
1048
1049 static void ops_complete_reconstruct(void *stripe_head_ref)
1050 {
1051         struct stripe_head *sh = stripe_head_ref;
1052         int disks = sh->disks;
1053         int pd_idx = sh->pd_idx;
1054         int qd_idx = sh->qd_idx;
1055         int i;
1056         bool fua = false;
1057
1058         pr_debug("%s: stripe %llu\n", __func__,
1059                 (unsigned long long)sh->sector);
1060
1061         for (i = disks; i--; )
1062                 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1063
1064         for (i = disks; i--; ) {
1065                 struct r5dev *dev = &sh->dev[i];
1066
1067                 if (dev->written || i == pd_idx || i == qd_idx) {
1068                         set_bit(R5_UPTODATE, &dev->flags);
1069                         if (fua)
1070                                 set_bit(R5_WantFUA, &dev->flags);
1071                 }
1072         }
1073
1074         if (sh->reconstruct_state == reconstruct_state_drain_run)
1075                 sh->reconstruct_state = reconstruct_state_drain_result;
1076         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1077                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1078         else {
1079                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1080                 sh->reconstruct_state = reconstruct_state_result;
1081         }
1082
1083         set_bit(STRIPE_HANDLE, &sh->state);
1084         release_stripe(sh);
1085 }
1086
1087 static void
1088 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1089                      struct dma_async_tx_descriptor *tx)
1090 {
1091         int disks = sh->disks;
1092         struct page **xor_srcs = percpu->scribble;
1093         struct async_submit_ctl submit;
1094         int count = 0, pd_idx = sh->pd_idx, i;
1095         struct page *xor_dest;
1096         int prexor = 0;
1097         unsigned long flags;
1098
1099         pr_debug("%s: stripe %llu\n", __func__,
1100                 (unsigned long long)sh->sector);
1101
1102         /* check if prexor is active which means only process blocks
1103          * that are part of a read-modify-write (written)
1104          */
1105         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1106                 prexor = 1;
1107                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1108                 for (i = disks; i--; ) {
1109                         struct r5dev *dev = &sh->dev[i];
1110                         if (dev->written)
1111                                 xor_srcs[count++] = dev->page;
1112                 }
1113         } else {
1114                 xor_dest = sh->dev[pd_idx].page;
1115                 for (i = disks; i--; ) {
1116                         struct r5dev *dev = &sh->dev[i];
1117                         if (i != pd_idx)
1118                                 xor_srcs[count++] = dev->page;
1119                 }
1120         }
1121
1122         /* 1/ if we prexor'd then the dest is reused as a source
1123          * 2/ if we did not prexor then we are redoing the parity
1124          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1125          * for the synchronous xor case
1126          */
1127         flags = ASYNC_TX_ACK |
1128                 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1129
1130         atomic_inc(&sh->count);
1131
1132         init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1133                           to_addr_conv(sh, percpu));
1134         if (unlikely(count == 1))
1135                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1136         else
1137                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1138 }
1139
1140 static void
1141 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1142                      struct dma_async_tx_descriptor *tx)
1143 {
1144         struct async_submit_ctl submit;
1145         struct page **blocks = percpu->scribble;
1146         int count;
1147
1148         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1149
1150         count = set_syndrome_sources(blocks, sh);
1151
1152         atomic_inc(&sh->count);
1153
1154         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1155                           sh, to_addr_conv(sh, percpu));
1156         async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1157 }
1158
1159 static void ops_complete_check(void *stripe_head_ref)
1160 {
1161         struct stripe_head *sh = stripe_head_ref;
1162
1163         pr_debug("%s: stripe %llu\n", __func__,
1164                 (unsigned long long)sh->sector);
1165
1166         sh->check_state = check_state_check_result;
1167         set_bit(STRIPE_HANDLE, &sh->state);
1168         release_stripe(sh);
1169 }
1170
1171 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1172 {
1173         int disks = sh->disks;
1174         int pd_idx = sh->pd_idx;
1175         int qd_idx = sh->qd_idx;
1176         struct page *xor_dest;
1177         struct page **xor_srcs = percpu->scribble;
1178         struct dma_async_tx_descriptor *tx;
1179         struct async_submit_ctl submit;
1180         int count;
1181         int i;
1182
1183         pr_debug("%s: stripe %llu\n", __func__,
1184                 (unsigned long long)sh->sector);
1185
1186         count = 0;
1187         xor_dest = sh->dev[pd_idx].page;
1188         xor_srcs[count++] = xor_dest;
1189         for (i = disks; i--; ) {
1190                 if (i == pd_idx || i == qd_idx)
1191                         continue;
1192                 xor_srcs[count++] = sh->dev[i].page;
1193         }
1194
1195         init_async_submit(&submit, 0, NULL, NULL, NULL,
1196                           to_addr_conv(sh, percpu));
1197         tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1198                            &sh->ops.zero_sum_result, &submit);
1199
1200         atomic_inc(&sh->count);
1201         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1202         tx = async_trigger_callback(&submit);
1203 }
1204
1205 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1206 {
1207         struct page **srcs = percpu->scribble;
1208         struct async_submit_ctl submit;
1209         int count;
1210
1211         pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1212                 (unsigned long long)sh->sector, checkp);
1213
1214         count = set_syndrome_sources(srcs, sh);
1215         if (!checkp)
1216                 srcs[count] = NULL;
1217
1218         atomic_inc(&sh->count);
1219         init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1220                           sh, to_addr_conv(sh, percpu));
1221         async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1222                            &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1223 }
1224
1225 static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1226 {
1227         int overlap_clear = 0, i, disks = sh->disks;
1228         struct dma_async_tx_descriptor *tx = NULL;
1229         raid5_conf_t *conf = sh->raid_conf;
1230         int level = conf->level;
1231         struct raid5_percpu *percpu;
1232         unsigned long cpu;
1233
1234         cpu = get_cpu();
1235         percpu = per_cpu_ptr(conf->percpu, cpu);
1236         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1237                 ops_run_biofill(sh);
1238                 overlap_clear++;
1239         }
1240
1241         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1242                 if (level < 6)
1243                         tx = ops_run_compute5(sh, percpu);
1244                 else {
1245                         if (sh->ops.target2 < 0 || sh->ops.target < 0)
1246                                 tx = ops_run_compute6_1(sh, percpu);
1247                         else
1248                                 tx = ops_run_compute6_2(sh, percpu);
1249                 }
1250                 /* terminate the chain if reconstruct is not set to be run */
1251                 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1252                         async_tx_ack(tx);
1253         }
1254
1255         if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1256                 tx = ops_run_prexor(sh, percpu, tx);
1257
1258         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1259                 tx = ops_run_biodrain(sh, tx);
1260                 overlap_clear++;
1261         }
1262
1263         if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1264                 if (level < 6)
1265                         ops_run_reconstruct5(sh, percpu, tx);
1266                 else
1267                         ops_run_reconstruct6(sh, percpu, tx);
1268         }
1269
1270         if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1271                 if (sh->check_state == check_state_run)
1272                         ops_run_check_p(sh, percpu);
1273                 else if (sh->check_state == check_state_run_q)
1274                         ops_run_check_pq(sh, percpu, 0);
1275                 else if (sh->check_state == check_state_run_pq)
1276                         ops_run_check_pq(sh, percpu, 1);
1277                 else
1278                         BUG();
1279         }
1280
1281         if (overlap_clear)
1282                 for (i = disks; i--; ) {
1283                         struct r5dev *dev = &sh->dev[i];
1284                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
1285                                 wake_up(&sh->raid_conf->wait_for_overlap);
1286                 }
1287         put_cpu();
1288 }
1289
1290 #ifdef CONFIG_MULTICORE_RAID456
1291 static void async_run_ops(void *param, async_cookie_t cookie)
1292 {
1293         struct stripe_head *sh = param;
1294         unsigned long ops_request = sh->ops.request;
1295
1296         clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1297         wake_up(&sh->ops.wait_for_ops);
1298
1299         __raid_run_ops(sh, ops_request);
1300         release_stripe(sh);
1301 }
1302
1303 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1304 {
1305         /* since handle_stripe can be called outside of raid5d context
1306          * we need to ensure sh->ops.request is de-staged before another
1307          * request arrives
1308          */
1309         wait_event(sh->ops.wait_for_ops,
1310                    !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1311         sh->ops.request = ops_request;
1312
1313         atomic_inc(&sh->count);
1314         async_schedule(async_run_ops, sh);
1315 }
1316 #else
1317 #define raid_run_ops __raid_run_ops
1318 #endif
1319
1320 static int grow_one_stripe(raid5_conf_t *conf)
1321 {
1322         struct stripe_head *sh;
1323         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
1324         if (!sh)
1325                 return 0;
1326         memset(sh, 0, sizeof(*sh) + (conf->pool_size-1)*sizeof(struct r5dev));
1327         sh->raid_conf = conf;
1328         spin_lock_init(&sh->lock);
1329         #ifdef CONFIG_MULTICORE_RAID456
1330         init_waitqueue_head(&sh->ops.wait_for_ops);
1331         #endif
1332
1333         if (grow_buffers(sh)) {
1334                 shrink_buffers(sh);
1335                 kmem_cache_free(conf->slab_cache, sh);
1336                 return 0;
1337         }
1338         /* we just created an active stripe so... */
1339         atomic_set(&sh->count, 1);
1340         atomic_inc(&conf->active_stripes);
1341         INIT_LIST_HEAD(&sh->lru);
1342         release_stripe(sh);
1343         return 1;
1344 }
1345
1346 static int grow_stripes(raid5_conf_t *conf, int num)
1347 {
1348         struct kmem_cache *sc;
1349         int devs = max(conf->raid_disks, conf->previous_raid_disks);
1350
1351         if (conf->mddev->gendisk)
1352                 sprintf(conf->cache_name[0],
1353                         "raid%d-%s", conf->level, mdname(conf->mddev));
1354         else
1355                 sprintf(conf->cache_name[0],
1356                         "raid%d-%p", conf->level, conf->mddev);
1357         sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1358
1359         conf->active_name = 0;
1360         sc = kmem_cache_create(conf->cache_name[conf->active_name],
1361                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1362                                0, 0, NULL);
1363         if (!sc)
1364                 return 1;
1365         conf->slab_cache = sc;
1366         conf->pool_size = devs;
1367         while (num--)
1368                 if (!grow_one_stripe(conf))
1369                         return 1;
1370         return 0;
1371 }
1372
1373 /**
1374  * scribble_len - return the required size of the scribble region
1375  * @num - total number of disks in the array
1376  *
1377  * The size must be enough to contain:
1378  * 1/ a struct page pointer for each device in the array +2
1379  * 2/ room to convert each entry in (1) to its corresponding dma
1380  *    (dma_map_page()) or page (page_address()) address.
1381  *
1382  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1383  * calculate over all devices (not just the data blocks), using zeros in place
1384  * of the P and Q blocks.
1385  */
1386 static size_t scribble_len(int num)
1387 {
1388         size_t len;
1389
1390         len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1391
1392         return len;
1393 }
1394
1395 static int resize_stripes(raid5_conf_t *conf, int newsize)
1396 {
1397         /* Make all the stripes able to hold 'newsize' devices.
1398          * New slots in each stripe get 'page' set to a new page.
1399          *
1400          * This happens in stages:
1401          * 1/ create a new kmem_cache and allocate the required number of
1402          *    stripe_heads.
1403          * 2/ gather all the old stripe_heads and tranfer the pages across
1404          *    to the new stripe_heads.  This will have the side effect of
1405          *    freezing the array as once all stripe_heads have been collected,
1406          *    no IO will be possible.  Old stripe heads are freed once their
1407          *    pages have been transferred over, and the old kmem_cache is
1408          *    freed when all stripes are done.
1409          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
1410          *    we simple return a failre status - no need to clean anything up.
1411          * 4/ allocate new pages for the new slots in the new stripe_heads.
1412          *    If this fails, we don't bother trying the shrink the
1413          *    stripe_heads down again, we just leave them as they are.
1414          *    As each stripe_head is processed the new one is released into
1415          *    active service.
1416          *
1417          * Once step2 is started, we cannot afford to wait for a write,
1418          * so we use GFP_NOIO allocations.
1419          */
1420         struct stripe_head *osh, *nsh;
1421         LIST_HEAD(newstripes);
1422         struct disk_info *ndisks;
1423         unsigned long cpu;
1424         int err;
1425         struct kmem_cache *sc;
1426         int i;
1427
1428         if (newsize <= conf->pool_size)
1429                 return 0; /* never bother to shrink */
1430
1431         err = md_allow_write(conf->mddev);
1432         if (err)
1433                 return err;
1434
1435         /* Step 1 */
1436         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1437                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1438                                0, 0, NULL);
1439         if (!sc)
1440                 return -ENOMEM;
1441
1442         for (i = conf->max_nr_stripes; i; i--) {
1443                 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1444                 if (!nsh)
1445                         break;
1446
1447                 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1448
1449                 nsh->raid_conf = conf;
1450                 spin_lock_init(&nsh->lock);
1451                 #ifdef CONFIG_MULTICORE_RAID456
1452                 init_waitqueue_head(&nsh->ops.wait_for_ops);
1453                 #endif
1454
1455                 list_add(&nsh->lru, &newstripes);
1456         }
1457         if (i) {
1458                 /* didn't get enough, give up */
1459                 while (!list_empty(&newstripes)) {
1460                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
1461                         list_del(&nsh->lru);
1462                         kmem_cache_free(sc, nsh);
1463                 }
1464                 kmem_cache_destroy(sc);
1465                 return -ENOMEM;
1466         }
1467         /* Step 2 - Must use GFP_NOIO now.
1468          * OK, we have enough stripes, start collecting inactive
1469          * stripes and copying them over
1470          */
1471         list_for_each_entry(nsh, &newstripes, lru) {
1472                 spin_lock_irq(&conf->device_lock);
1473                 wait_event_lock_irq(conf->wait_for_stripe,
1474                                     !list_empty(&conf->inactive_list),
1475                                     conf->device_lock,
1476                                     unplug_slaves(conf->mddev)
1477                         );
1478                 osh = get_free_stripe(conf);
1479                 spin_unlock_irq(&conf->device_lock);
1480                 atomic_set(&nsh->count, 1);
1481                 for(i=0; i<conf->pool_size; i++)
1482                         nsh->dev[i].page = osh->dev[i].page;
1483                 for( ; i<newsize; i++)
1484                         nsh->dev[i].page = NULL;
1485                 kmem_cache_free(conf->slab_cache, osh);
1486         }
1487         kmem_cache_destroy(conf->slab_cache);
1488
1489         /* Step 3.
1490          * At this point, we are holding all the stripes so the array
1491          * is completely stalled, so now is a good time to resize
1492          * conf->disks and the scribble region
1493          */
1494         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1495         if (ndisks) {
1496                 for (i=0; i<conf->raid_disks; i++)
1497                         ndisks[i] = conf->disks[i];
1498                 kfree(conf->disks);
1499                 conf->disks = ndisks;
1500         } else
1501                 err = -ENOMEM;
1502
1503         get_online_cpus();
1504         conf->scribble_len = scribble_len(newsize);
1505         for_each_present_cpu(cpu) {
1506                 struct raid5_percpu *percpu;
1507                 void *scribble;
1508
1509                 percpu = per_cpu_ptr(conf->percpu, cpu);
1510                 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1511
1512                 if (scribble) {
1513                         kfree(percpu->scribble);
1514                         percpu->scribble = scribble;
1515                 } else {
1516                         err = -ENOMEM;
1517                         break;
1518                 }
1519         }
1520         put_online_cpus();
1521
1522         /* Step 4, return new stripes to service */
1523         while(!list_empty(&newstripes)) {
1524                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1525                 list_del_init(&nsh->lru);
1526
1527                 for (i=conf->raid_disks; i < newsize; i++)
1528                         if (nsh->dev[i].page == NULL) {
1529                                 struct page *p = alloc_page(GFP_NOIO);
1530                                 nsh->dev[i].page = p;
1531                                 if (!p)
1532                                         err = -ENOMEM;
1533                         }
1534                 release_stripe(nsh);
1535         }
1536         /* critical section pass, GFP_NOIO no longer needed */
1537
1538         conf->slab_cache = sc;
1539         conf->active_name = 1-conf->active_name;
1540         conf->pool_size = newsize;
1541         return err;
1542 }
1543
1544 static int drop_one_stripe(raid5_conf_t *conf)
1545 {
1546         struct stripe_head *sh;
1547
1548         spin_lock_irq(&conf->device_lock);
1549         sh = get_free_stripe(conf);
1550         spin_unlock_irq(&conf->device_lock);
1551         if (!sh)
1552                 return 0;
1553         BUG_ON(atomic_read(&sh->count));
1554         shrink_buffers(sh);
1555         kmem_cache_free(conf->slab_cache, sh);
1556         atomic_dec(&conf->active_stripes);
1557         return 1;
1558 }
1559
1560 static void shrink_stripes(raid5_conf_t *conf)
1561 {
1562         while (drop_one_stripe(conf))
1563                 ;
1564
1565         if (conf->slab_cache)
1566                 kmem_cache_destroy(conf->slab_cache);
1567         conf->slab_cache = NULL;
1568 }
1569
1570 static void raid5_end_read_request(struct bio * bi, int error)
1571 {
1572         struct stripe_head *sh = bi->bi_private;
1573         raid5_conf_t *conf = sh->raid_conf;
1574         int disks = sh->disks, i;
1575         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1576         char b[BDEVNAME_SIZE];
1577         mdk_rdev_t *rdev;
1578
1579
1580         for (i=0 ; i<disks; i++)
1581                 if (bi == &sh->dev[i].req)
1582                         break;
1583
1584         pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1585                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1586                 uptodate);
1587         if (i == disks) {
1588                 BUG();
1589                 return;
1590         }
1591
1592         if (uptodate) {
1593                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1594                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1595                         rdev = conf->disks[i].rdev;
1596                         printk_rl(KERN_INFO "md/raid:%s: read error corrected"
1597                                   " (%lu sectors at %llu on %s)\n",
1598                                   mdname(conf->mddev), STRIPE_SECTORS,
1599                                   (unsigned long long)(sh->sector
1600                                                        + rdev->data_offset),
1601                                   bdevname(rdev->bdev, b));
1602                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1603                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1604                 }
1605                 if (atomic_read(&conf->disks[i].rdev->read_errors))
1606                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
1607         } else {
1608                 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1609                 int retry = 0;
1610                 rdev = conf->disks[i].rdev;
1611
1612                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1613                 atomic_inc(&rdev->read_errors);
1614                 if (conf->mddev->degraded >= conf->max_degraded)
1615                         printk_rl(KERN_WARNING
1616                                   "md/raid:%s: read error not correctable "
1617                                   "(sector %llu on %s).\n",
1618                                   mdname(conf->mddev),
1619                                   (unsigned long long)(sh->sector
1620                                                        + rdev->data_offset),
1621                                   bdn);
1622                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1623                         /* Oh, no!!! */
1624                         printk_rl(KERN_WARNING
1625                                   "md/raid:%s: read error NOT corrected!! "
1626                                   "(sector %llu on %s).\n",
1627                                   mdname(conf->mddev),
1628                                   (unsigned long long)(sh->sector
1629                                                        + rdev->data_offset),
1630                                   bdn);
1631                 else if (atomic_read(&rdev->read_errors)
1632                          > conf->max_nr_stripes)
1633                         printk(KERN_WARNING
1634                                "md/raid:%s: Too many read errors, failing device %s.\n",
1635                                mdname(conf->mddev), bdn);
1636                 else
1637                         retry = 1;
1638                 if (retry)
1639                         set_bit(R5_ReadError, &sh->dev[i].flags);
1640                 else {
1641                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1642                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1643                         md_error(conf->mddev, rdev);
1644                 }
1645         }
1646         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1647         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1648         set_bit(STRIPE_HANDLE, &sh->state);
1649         release_stripe(sh);
1650 }
1651
1652 static void raid5_end_write_request(struct bio *bi, int error)
1653 {
1654         struct stripe_head *sh = bi->bi_private;
1655         raid5_conf_t *conf = sh->raid_conf;
1656         int disks = sh->disks, i;
1657         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1658
1659         for (i=0 ; i<disks; i++)
1660                 if (bi == &sh->dev[i].req)
1661                         break;
1662
1663         pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1664                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1665                 uptodate);
1666         if (i == disks) {
1667                 BUG();
1668                 return;
1669         }
1670
1671         if (!uptodate)
1672                 md_error(conf->mddev, conf->disks[i].rdev);
1673
1674         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1675         
1676         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1677         set_bit(STRIPE_HANDLE, &sh->state);
1678         release_stripe(sh);
1679 }
1680
1681
1682 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1683         
1684 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1685 {
1686         struct r5dev *dev = &sh->dev[i];
1687
1688         bio_init(&dev->req);
1689         dev->req.bi_io_vec = &dev->vec;
1690         dev->req.bi_vcnt++;
1691         dev->req.bi_max_vecs++;
1692         dev->vec.bv_page = dev->page;
1693         dev->vec.bv_len = STRIPE_SIZE;
1694         dev->vec.bv_offset = 0;
1695
1696         dev->req.bi_sector = sh->sector;
1697         dev->req.bi_private = sh;
1698
1699         dev->flags = 0;
1700         dev->sector = compute_blocknr(sh, i, previous);
1701 }
1702
1703 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1704 {
1705         char b[BDEVNAME_SIZE];
1706         raid5_conf_t *conf = mddev->private;
1707         pr_debug("raid456: error called\n");
1708
1709         if (!test_bit(Faulty, &rdev->flags)) {
1710                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1711                 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1712                         unsigned long flags;
1713                         spin_lock_irqsave(&conf->device_lock, flags);
1714                         mddev->degraded++;
1715                         spin_unlock_irqrestore(&conf->device_lock, flags);
1716                         /*
1717                          * if recovery was running, make sure it aborts.
1718                          */
1719                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1720                 }
1721                 set_bit(Faulty, &rdev->flags);
1722                 printk(KERN_ALERT
1723                        "md/raid:%s: Disk failure on %s, disabling device.\n"
1724                        "md/raid:%s: Operation continuing on %d devices.\n",
1725                        mdname(mddev),
1726                        bdevname(rdev->bdev, b),
1727                        mdname(mddev),
1728                        conf->raid_disks - mddev->degraded);
1729         }
1730 }
1731
1732 /*
1733  * Input: a 'big' sector number,
1734  * Output: index of the data and parity disk, and the sector # in them.
1735  */
1736 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1737                                      int previous, int *dd_idx,
1738                                      struct stripe_head *sh)
1739 {
1740         sector_t stripe, stripe2;
1741         sector_t chunk_number;
1742         unsigned int chunk_offset;
1743         int pd_idx, qd_idx;
1744         int ddf_layout = 0;
1745         sector_t new_sector;
1746         int algorithm = previous ? conf->prev_algo
1747                                  : conf->algorithm;
1748         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1749                                          : conf->chunk_sectors;
1750         int raid_disks = previous ? conf->previous_raid_disks
1751                                   : conf->raid_disks;
1752         int data_disks = raid_disks - conf->max_degraded;
1753
1754         /* First compute the information on this sector */
1755
1756         /*
1757          * Compute the chunk number and the sector offset inside the chunk
1758          */
1759         chunk_offset = sector_div(r_sector, sectors_per_chunk);
1760         chunk_number = r_sector;
1761
1762         /*
1763          * Compute the stripe number
1764          */
1765         stripe = chunk_number;
1766         *dd_idx = sector_div(stripe, data_disks);
1767         stripe2 = stripe;
1768         /*
1769          * Select the parity disk based on the user selected algorithm.
1770          */
1771         pd_idx = qd_idx = ~0;
1772         switch(conf->level) {
1773         case 4:
1774                 pd_idx = data_disks;
1775                 break;
1776         case 5:
1777                 switch (algorithm) {
1778                 case ALGORITHM_LEFT_ASYMMETRIC:
1779                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
1780                         if (*dd_idx >= pd_idx)
1781                                 (*dd_idx)++;
1782                         break;
1783                 case ALGORITHM_RIGHT_ASYMMETRIC:
1784                         pd_idx = sector_div(stripe2, raid_disks);
1785                         if (*dd_idx >= pd_idx)
1786                                 (*dd_idx)++;
1787                         break;
1788                 case ALGORITHM_LEFT_SYMMETRIC:
1789                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
1790                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1791                         break;
1792                 case ALGORITHM_RIGHT_SYMMETRIC:
1793                         pd_idx = sector_div(stripe2, raid_disks);
1794                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1795                         break;
1796                 case ALGORITHM_PARITY_0:
1797                         pd_idx = 0;
1798                         (*dd_idx)++;
1799                         break;
1800                 case ALGORITHM_PARITY_N:
1801                         pd_idx = data_disks;
1802                         break;
1803                 default:
1804                         BUG();
1805                 }
1806                 break;
1807         case 6:
1808
1809                 switch (algorithm) {
1810                 case ALGORITHM_LEFT_ASYMMETRIC:
1811                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1812                         qd_idx = pd_idx + 1;
1813                         if (pd_idx == raid_disks-1) {
1814                                 (*dd_idx)++;    /* Q D D D P */
1815                                 qd_idx = 0;
1816                         } else if (*dd_idx >= pd_idx)
1817                                 (*dd_idx) += 2; /* D D P Q D */
1818                         break;
1819                 case ALGORITHM_RIGHT_ASYMMETRIC:
1820                         pd_idx = sector_div(stripe2, raid_disks);
1821                         qd_idx = pd_idx + 1;
1822                         if (pd_idx == raid_disks-1) {
1823                                 (*dd_idx)++;    /* Q D D D P */
1824                                 qd_idx = 0;
1825                         } else if (*dd_idx >= pd_idx)
1826                                 (*dd_idx) += 2; /* D D P Q D */
1827                         break;
1828                 case ALGORITHM_LEFT_SYMMETRIC:
1829                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1830                         qd_idx = (pd_idx + 1) % raid_disks;
1831                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1832                         break;
1833                 case ALGORITHM_RIGHT_SYMMETRIC:
1834                         pd_idx = sector_div(stripe2, raid_disks);
1835                         qd_idx = (pd_idx + 1) % raid_disks;
1836                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1837                         break;
1838
1839                 case ALGORITHM_PARITY_0:
1840                         pd_idx = 0;
1841                         qd_idx = 1;
1842                         (*dd_idx) += 2;
1843                         break;
1844                 case ALGORITHM_PARITY_N:
1845                         pd_idx = data_disks;
1846                         qd_idx = data_disks + 1;
1847                         break;
1848
1849                 case ALGORITHM_ROTATING_ZERO_RESTART:
1850                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
1851                          * of blocks for computing Q is different.
1852                          */
1853                         pd_idx = sector_div(stripe2, raid_disks);
1854                         qd_idx = pd_idx + 1;
1855                         if (pd_idx == raid_disks-1) {
1856                                 (*dd_idx)++;    /* Q D D D P */
1857                                 qd_idx = 0;
1858                         } else if (*dd_idx >= pd_idx)
1859                                 (*dd_idx) += 2; /* D D P Q D */
1860                         ddf_layout = 1;
1861                         break;
1862
1863                 case ALGORITHM_ROTATING_N_RESTART:
1864                         /* Same a left_asymmetric, by first stripe is
1865                          * D D D P Q  rather than
1866                          * Q D D D P
1867                          */
1868                         stripe2 += 1;
1869                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1870                         qd_idx = pd_idx + 1;
1871                         if (pd_idx == raid_disks-1) {
1872                                 (*dd_idx)++;    /* Q D D D P */
1873                                 qd_idx = 0;
1874                         } else if (*dd_idx >= pd_idx)
1875                                 (*dd_idx) += 2; /* D D P Q D */
1876                         ddf_layout = 1;
1877                         break;
1878
1879                 case ALGORITHM_ROTATING_N_CONTINUE:
1880                         /* Same as left_symmetric but Q is before P */
1881                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1882                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1883                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1884                         ddf_layout = 1;
1885                         break;
1886
1887                 case ALGORITHM_LEFT_ASYMMETRIC_6:
1888                         /* RAID5 left_asymmetric, with Q on last device */
1889                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1890                         if (*dd_idx >= pd_idx)
1891                                 (*dd_idx)++;
1892                         qd_idx = raid_disks - 1;
1893                         break;
1894
1895                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1896                         pd_idx = sector_div(stripe2, raid_disks-1);
1897                         if (*dd_idx >= pd_idx)
1898                                 (*dd_idx)++;
1899                         qd_idx = raid_disks - 1;
1900                         break;
1901
1902                 case ALGORITHM_LEFT_SYMMETRIC_6:
1903                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1904                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1905                         qd_idx = raid_disks - 1;
1906                         break;
1907
1908                 case ALGORITHM_RIGHT_SYMMETRIC_6:
1909                         pd_idx = sector_div(stripe2, raid_disks-1);
1910                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1911                         qd_idx = raid_disks - 1;
1912                         break;
1913
1914                 case ALGORITHM_PARITY_0_6:
1915                         pd_idx = 0;
1916                         (*dd_idx)++;
1917                         qd_idx = raid_disks - 1;
1918                         break;
1919
1920                 default:
1921                         BUG();
1922                 }
1923                 break;
1924         }
1925
1926         if (sh) {
1927                 sh->pd_idx = pd_idx;
1928                 sh->qd_idx = qd_idx;
1929                 sh->ddf_layout = ddf_layout;
1930         }
1931         /*
1932          * Finally, compute the new sector number
1933          */
1934         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1935         return new_sector;
1936 }
1937
1938
1939 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1940 {
1941         raid5_conf_t *conf = sh->raid_conf;
1942         int raid_disks = sh->disks;
1943         int data_disks = raid_disks - conf->max_degraded;
1944         sector_t new_sector = sh->sector, check;
1945         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1946                                          : conf->chunk_sectors;
1947         int algorithm = previous ? conf->prev_algo
1948                                  : conf->algorithm;
1949         sector_t stripe;
1950         int chunk_offset;
1951         sector_t chunk_number;
1952         int dummy1, dd_idx = i;
1953         sector_t r_sector;
1954         struct stripe_head sh2;
1955
1956
1957         chunk_offset = sector_div(new_sector, sectors_per_chunk);
1958         stripe = new_sector;
1959
1960         if (i == sh->pd_idx)
1961                 return 0;
1962         switch(conf->level) {
1963         case 4: break;
1964         case 5:
1965                 switch (algorithm) {
1966                 case ALGORITHM_LEFT_ASYMMETRIC:
1967                 case ALGORITHM_RIGHT_ASYMMETRIC:
1968                         if (i > sh->pd_idx)
1969                                 i--;
1970                         break;
1971                 case ALGORITHM_LEFT_SYMMETRIC:
1972                 case ALGORITHM_RIGHT_SYMMETRIC:
1973                         if (i < sh->pd_idx)
1974                                 i += raid_disks;
1975                         i -= (sh->pd_idx + 1);
1976                         break;
1977                 case ALGORITHM_PARITY_0:
1978                         i -= 1;
1979                         break;
1980                 case ALGORITHM_PARITY_N:
1981                         break;
1982                 default:
1983                         BUG();
1984                 }
1985                 break;
1986         case 6:
1987                 if (i == sh->qd_idx)
1988                         return 0; /* It is the Q disk */
1989                 switch (algorithm) {
1990                 case ALGORITHM_LEFT_ASYMMETRIC:
1991                 case ALGORITHM_RIGHT_ASYMMETRIC:
1992                 case ALGORITHM_ROTATING_ZERO_RESTART:
1993                 case ALGORITHM_ROTATING_N_RESTART:
1994                         if (sh->pd_idx == raid_disks-1)
1995                                 i--;    /* Q D D D P */
1996                         else if (i > sh->pd_idx)
1997                                 i -= 2; /* D D P Q D */
1998                         break;
1999                 case ALGORITHM_LEFT_SYMMETRIC:
2000                 case ALGORITHM_RIGHT_SYMMETRIC:
2001                         if (sh->pd_idx == raid_disks-1)
2002                                 i--; /* Q D D D P */
2003                         else {
2004                                 /* D D P Q D */
2005                                 if (i < sh->pd_idx)
2006                                         i += raid_disks;
2007                                 i -= (sh->pd_idx + 2);
2008                         }
2009                         break;
2010                 case ALGORITHM_PARITY_0:
2011                         i -= 2;
2012                         break;
2013                 case ALGORITHM_PARITY_N:
2014                         break;
2015                 case ALGORITHM_ROTATING_N_CONTINUE:
2016                         /* Like left_symmetric, but P is before Q */
2017                         if (sh->pd_idx == 0)
2018                                 i--;    /* P D D D Q */
2019                         else {
2020                                 /* D D Q P D */
2021                                 if (i < sh->pd_idx)
2022                                         i += raid_disks;
2023                                 i -= (sh->pd_idx + 1);
2024                         }
2025                         break;
2026                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2027                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2028                         if (i > sh->pd_idx)
2029                                 i--;
2030                         break;
2031                 case ALGORITHM_LEFT_SYMMETRIC_6:
2032                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2033                         if (i < sh->pd_idx)
2034                                 i += data_disks + 1;
2035                         i -= (sh->pd_idx + 1);
2036                         break;
2037                 case ALGORITHM_PARITY_0_6:
2038                         i -= 1;
2039                         break;
2040                 default:
2041                         BUG();
2042                 }
2043                 break;
2044         }
2045
2046         chunk_number = stripe * data_disks + i;
2047         r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2048
2049         check = raid5_compute_sector(conf, r_sector,
2050                                      previous, &dummy1, &sh2);
2051         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2052                 || sh2.qd_idx != sh->qd_idx) {
2053                 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2054                        mdname(conf->mddev));
2055                 return 0;
2056         }
2057         return r_sector;
2058 }
2059
2060
2061 static void
2062 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2063                          int rcw, int expand)
2064 {
2065         int i, pd_idx = sh->pd_idx, disks = sh->disks;
2066         raid5_conf_t *conf = sh->raid_conf;
2067         int level = conf->level;
2068
2069         if (rcw) {
2070                 /* if we are not expanding this is a proper write request, and
2071                  * there will be bios with new data to be drained into the
2072                  * stripe cache
2073                  */
2074                 if (!expand) {
2075                         sh->reconstruct_state = reconstruct_state_drain_run;
2076                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2077                 } else
2078                         sh->reconstruct_state = reconstruct_state_run;
2079
2080                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2081
2082                 for (i = disks; i--; ) {
2083                         struct r5dev *dev = &sh->dev[i];
2084
2085                         if (dev->towrite) {
2086                                 set_bit(R5_LOCKED, &dev->flags);
2087                                 set_bit(R5_Wantdrain, &dev->flags);
2088                                 if (!expand)
2089                                         clear_bit(R5_UPTODATE, &dev->flags);
2090                                 s->locked++;
2091                         }
2092                 }
2093                 if (s->locked + conf->max_degraded == disks)
2094                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2095                                 atomic_inc(&conf->pending_full_writes);
2096         } else {
2097                 BUG_ON(level == 6);
2098                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2099                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2100
2101                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2102                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2103                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2104                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2105
2106                 for (i = disks; i--; ) {
2107                         struct r5dev *dev = &sh->dev[i];
2108                         if (i == pd_idx)
2109                                 continue;
2110
2111                         if (dev->towrite &&
2112                             (test_bit(R5_UPTODATE, &dev->flags) ||
2113                              test_bit(R5_Wantcompute, &dev->flags))) {
2114                                 set_bit(R5_Wantdrain, &dev->flags);
2115                                 set_bit(R5_LOCKED, &dev->flags);
2116                                 clear_bit(R5_UPTODATE, &dev->flags);
2117                                 s->locked++;
2118                         }
2119                 }
2120         }
2121
2122         /* keep the parity disk(s) locked while asynchronous operations
2123          * are in flight
2124          */
2125         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2126         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2127         s->locked++;
2128
2129         if (level == 6) {
2130                 int qd_idx = sh->qd_idx;
2131                 struct r5dev *dev = &sh->dev[qd_idx];
2132
2133                 set_bit(R5_LOCKED, &dev->flags);
2134                 clear_bit(R5_UPTODATE, &dev->flags);
2135                 s->locked++;
2136         }
2137
2138         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2139                 __func__, (unsigned long long)sh->sector,
2140                 s->locked, s->ops_request);
2141 }
2142
2143 /*
2144  * Each stripe/dev can have one or more bion attached.
2145  * toread/towrite point to the first in a chain.
2146  * The bi_next chain must be in order.
2147  */
2148 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2149 {
2150         struct bio **bip;
2151         raid5_conf_t *conf = sh->raid_conf;
2152         int firstwrite=0;
2153
2154         pr_debug("adding bh b#%llu to stripe s#%llu\n",
2155                 (unsigned long long)bi->bi_sector,
2156                 (unsigned long long)sh->sector);
2157
2158
2159         spin_lock(&sh->lock);
2160         spin_lock_irq(&conf->device_lock);
2161         if (forwrite) {
2162                 bip = &sh->dev[dd_idx].towrite;
2163                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2164                         firstwrite = 1;
2165         } else
2166                 bip = &sh->dev[dd_idx].toread;
2167         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2168                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2169                         goto overlap;
2170                 bip = & (*bip)->bi_next;
2171         }
2172         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2173                 goto overlap;
2174
2175         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2176         if (*bip)
2177                 bi->bi_next = *bip;
2178         *bip = bi;
2179         bi->bi_phys_segments++;
2180         spin_unlock_irq(&conf->device_lock);
2181         spin_unlock(&sh->lock);
2182
2183         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2184                 (unsigned long long)bi->bi_sector,
2185                 (unsigned long long)sh->sector, dd_idx);
2186
2187         if (conf->mddev->bitmap && firstwrite) {
2188                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2189                                   STRIPE_SECTORS, 0);
2190                 sh->bm_seq = conf->seq_flush+1;
2191                 set_bit(STRIPE_BIT_DELAY, &sh->state);
2192         }
2193
2194         if (forwrite) {
2195                 /* check if page is covered */
2196                 sector_t sector = sh->dev[dd_idx].sector;
2197                 for (bi=sh->dev[dd_idx].towrite;
2198                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2199                              bi && bi->bi_sector <= sector;
2200                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2201                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2202                                 sector = bi->bi_sector + (bi->bi_size>>9);
2203                 }
2204                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2205                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2206         }
2207         return 1;
2208
2209  overlap:
2210         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2211         spin_unlock_irq(&conf->device_lock);
2212         spin_unlock(&sh->lock);
2213         return 0;
2214 }
2215
2216 static void end_reshape(raid5_conf_t *conf);
2217
2218 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
2219                             struct stripe_head *sh)
2220 {
2221         int sectors_per_chunk =
2222                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2223         int dd_idx;
2224         int chunk_offset = sector_div(stripe, sectors_per_chunk);
2225         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2226
2227         raid5_compute_sector(conf,
2228                              stripe * (disks - conf->max_degraded)
2229                              *sectors_per_chunk + chunk_offset,
2230                              previous,
2231                              &dd_idx, sh);
2232 }
2233
2234 static void
2235 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2236                                 struct stripe_head_state *s, int disks,
2237                                 struct bio **return_bi)
2238 {
2239         int i;
2240         for (i = disks; i--; ) {
2241                 struct bio *bi;
2242                 int bitmap_end = 0;
2243
2244                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2245                         mdk_rdev_t *rdev;
2246                         rcu_read_lock();
2247                         rdev = rcu_dereference(conf->disks[i].rdev);
2248                         if (rdev && test_bit(In_sync, &rdev->flags))
2249                                 /* multiple read failures in one stripe */
2250                                 md_error(conf->mddev, rdev);
2251                         rcu_read_unlock();
2252                 }
2253                 spin_lock_irq(&conf->device_lock);
2254                 /* fail all writes first */
2255                 bi = sh->dev[i].towrite;
2256                 sh->dev[i].towrite = NULL;
2257                 if (bi) {
2258                         s->to_write--;
2259                         bitmap_end = 1;
2260                 }
2261
2262                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2263                         wake_up(&conf->wait_for_overlap);
2264
2265                 while (bi && bi->bi_sector <
2266                         sh->dev[i].sector + STRIPE_SECTORS) {
2267                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2268                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2269                         if (!raid5_dec_bi_phys_segments(bi)) {
2270                                 md_write_end(conf->mddev);
2271                                 bi->bi_next = *return_bi;
2272                                 *return_bi = bi;
2273                         }
2274                         bi = nextbi;
2275                 }
2276                 /* and fail all 'written' */
2277                 bi = sh->dev[i].written;
2278                 sh->dev[i].written = NULL;
2279                 if (bi) bitmap_end = 1;
2280                 while (bi && bi->bi_sector <
2281                        sh->dev[i].sector + STRIPE_SECTORS) {
2282                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2283                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2284                         if (!raid5_dec_bi_phys_segments(bi)) {
2285                                 md_write_end(conf->mddev);
2286                                 bi->bi_next = *return_bi;
2287                                 *return_bi = bi;
2288                         }
2289                         bi = bi2;
2290                 }
2291
2292                 /* fail any reads if this device is non-operational and
2293                  * the data has not reached the cache yet.
2294                  */
2295                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2296                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2297                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
2298                         bi = sh->dev[i].toread;
2299                         sh->dev[i].toread = NULL;
2300                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2301                                 wake_up(&conf->wait_for_overlap);
2302                         if (bi) s->to_read--;
2303                         while (bi && bi->bi_sector <
2304                                sh->dev[i].sector + STRIPE_SECTORS) {
2305                                 struct bio *nextbi =
2306                                         r5_next_bio(bi, sh->dev[i].sector);
2307                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2308                                 if (!raid5_dec_bi_phys_segments(bi)) {
2309                                         bi->bi_next = *return_bi;
2310                                         *return_bi = bi;
2311                                 }
2312                                 bi = nextbi;
2313                         }
2314                 }
2315                 spin_unlock_irq(&conf->device_lock);
2316                 if (bitmap_end)
2317                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2318                                         STRIPE_SECTORS, 0, 0);
2319         }
2320
2321         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2322                 if (atomic_dec_and_test(&conf->pending_full_writes))
2323                         md_wakeup_thread(conf->mddev->thread);
2324 }
2325
2326 /* fetch_block5 - checks the given member device to see if its data needs
2327  * to be read or computed to satisfy a request.
2328  *
2329  * Returns 1 when no more member devices need to be checked, otherwise returns
2330  * 0 to tell the loop in handle_stripe_fill5 to continue
2331  */
2332 static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s,
2333                         int disk_idx, int disks)
2334 {
2335         struct r5dev *dev = &sh->dev[disk_idx];
2336         struct r5dev *failed_dev = &sh->dev[s->failed_num];
2337
2338         /* is the data in this block needed, and can we get it? */
2339         if (!test_bit(R5_LOCKED, &dev->flags) &&
2340             !test_bit(R5_UPTODATE, &dev->flags) &&
2341             (dev->toread ||
2342              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2343              s->syncing || s->expanding ||
2344              (s->failed &&
2345               (failed_dev->toread ||
2346                (failed_dev->towrite &&
2347                 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) {
2348                 /* We would like to get this block, possibly by computing it,
2349                  * otherwise read it if the backing disk is insync
2350                  */
2351                 if ((s->uptodate == disks - 1) &&
2352                     (s->failed && disk_idx == s->failed_num)) {
2353                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2354                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2355                         set_bit(R5_Wantcompute, &dev->flags);
2356                         sh->ops.target = disk_idx;
2357                         sh->ops.target2 = -1;
2358                         s->req_compute = 1;
2359                         /* Careful: from this point on 'uptodate' is in the eye
2360                          * of raid_run_ops which services 'compute' operations
2361                          * before writes. R5_Wantcompute flags a block that will
2362                          * be R5_UPTODATE by the time it is needed for a
2363                          * subsequent operation.
2364                          */
2365                         s->uptodate++;
2366                         return 1; /* uptodate + compute == disks */
2367                 } else if (test_bit(R5_Insync, &dev->flags)) {
2368                         set_bit(R5_LOCKED, &dev->flags);
2369                         set_bit(R5_Wantread, &dev->flags);
2370                         s->locked++;
2371                         pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2372                                 s->syncing);
2373                 }
2374         }
2375
2376         return 0;
2377 }
2378
2379 /**
2380  * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2381  */
2382 static void handle_stripe_fill5(struct stripe_head *sh,
2383                         struct stripe_head_state *s, int disks)
2384 {
2385         int i;
2386
2387         /* look for blocks to read/compute, skip this if a compute
2388          * is already in flight, or if the stripe contents are in the
2389          * midst of changing due to a write
2390          */
2391         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2392             !sh->reconstruct_state)
2393                 for (i = disks; i--; )
2394                         if (fetch_block5(sh, s, i, disks))
2395                                 break;
2396         set_bit(STRIPE_HANDLE, &sh->state);
2397 }
2398
2399 /* fetch_block6 - checks the given member device to see if its data needs
2400  * to be read or computed to satisfy a request.
2401  *
2402  * Returns 1 when no more member devices need to be checked, otherwise returns
2403  * 0 to tell the loop in handle_stripe_fill6 to continue
2404  */
2405 static int fetch_block6(struct stripe_head *sh, struct stripe_head_state *s,
2406                          struct r6_state *r6s, int disk_idx, int disks)
2407 {
2408         struct r5dev *dev = &sh->dev[disk_idx];
2409         struct r5dev *fdev[2] = { &sh->dev[r6s->failed_num[0]],
2410                                   &sh->dev[r6s->failed_num[1]] };
2411
2412         if (!test_bit(R5_LOCKED, &dev->flags) &&
2413             !test_bit(R5_UPTODATE, &dev->flags) &&
2414             (dev->toread ||
2415              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2416              s->syncing || s->expanding ||
2417              (s->failed >= 1 &&
2418               (fdev[0]->toread || s->to_write)) ||
2419              (s->failed >= 2 &&
2420               (fdev[1]->toread || s->to_write)))) {
2421                 /* we would like to get this block, possibly by computing it,
2422                  * otherwise read it if the backing disk is insync
2423                  */
2424                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2425                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2426                 if ((s->uptodate == disks - 1) &&
2427                     (s->failed && (disk_idx == r6s->failed_num[0] ||
2428                                    disk_idx == r6s->failed_num[1]))) {
2429                         /* have disk failed, and we're requested to fetch it;
2430                          * do compute it
2431                          */
2432                         pr_debug("Computing stripe %llu block %d\n",
2433                                (unsigned long long)sh->sector, disk_idx);
2434                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2435                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2436                         set_bit(R5_Wantcompute, &dev->flags);
2437                         sh->ops.target = disk_idx;
2438                         sh->ops.target2 = -1; /* no 2nd target */
2439                         s->req_compute = 1;
2440                         s->uptodate++;
2441                         return 1;
2442                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2443                         /* Computing 2-failure is *very* expensive; only
2444                          * do it if failed >= 2
2445                          */
2446                         int other;
2447                         for (other = disks; other--; ) {
2448                                 if (other == disk_idx)
2449                                         continue;
2450                                 if (!test_bit(R5_UPTODATE,
2451                                       &sh->dev[other].flags))
2452                                         break;
2453                         }
2454                         BUG_ON(other < 0);
2455                         pr_debug("Computing stripe %llu blocks %d,%d\n",
2456                                (unsigned long long)sh->sector,
2457                                disk_idx, other);
2458                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2459                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2460                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2461                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
2462                         sh->ops.target = disk_idx;
2463                         sh->ops.target2 = other;
2464                         s->uptodate += 2;
2465                         s->req_compute = 1;
2466                         return 1;
2467                 } else if (test_bit(R5_Insync, &dev->flags)) {
2468                         set_bit(R5_LOCKED, &dev->flags);
2469                         set_bit(R5_Wantread, &dev->flags);
2470                         s->locked++;
2471                         pr_debug("Reading block %d (sync=%d)\n",
2472                                 disk_idx, s->syncing);
2473                 }
2474         }
2475
2476         return 0;
2477 }
2478
2479 /**
2480  * handle_stripe_fill6 - read or compute data to satisfy pending requests.
2481  */
2482 static void handle_stripe_fill6(struct stripe_head *sh,
2483                         struct stripe_head_state *s, struct r6_state *r6s,
2484                         int disks)
2485 {
2486         int i;
2487
2488         /* look for blocks to read/compute, skip this if a compute
2489          * is already in flight, or if the stripe contents are in the
2490          * midst of changing due to a write
2491          */
2492         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2493             !sh->reconstruct_state)
2494                 for (i = disks; i--; )
2495                         if (fetch_block6(sh, s, r6s, i, disks))
2496                                 break;
2497         set_bit(STRIPE_HANDLE, &sh->state);
2498 }
2499
2500
2501 /* handle_stripe_clean_event
2502  * any written block on an uptodate or failed drive can be returned.
2503  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2504  * never LOCKED, so we don't need to test 'failed' directly.
2505  */
2506 static void handle_stripe_clean_event(raid5_conf_t *conf,
2507         struct stripe_head *sh, int disks, struct bio **return_bi)
2508 {
2509         int i;
2510         struct r5dev *dev;
2511
2512         for (i = disks; i--; )
2513                 if (sh->dev[i].written) {
2514                         dev = &sh->dev[i];
2515                         if (!test_bit(R5_LOCKED, &dev->flags) &&
2516                                 test_bit(R5_UPTODATE, &dev->flags)) {
2517                                 /* We can return any write requests */
2518                                 struct bio *wbi, *wbi2;
2519                                 int bitmap_end = 0;
2520                                 pr_debug("Return write for disc %d\n", i);
2521                                 spin_lock_irq(&conf->device_lock);
2522                                 wbi = dev->written;
2523                                 dev->written = NULL;
2524                                 while (wbi && wbi->bi_sector <
2525                                         dev->sector + STRIPE_SECTORS) {
2526                                         wbi2 = r5_next_bio(wbi, dev->sector);
2527                                         if (!raid5_dec_bi_phys_segments(wbi)) {
2528                                                 md_write_end(conf->mddev);
2529                                                 wbi->bi_next = *return_bi;
2530                                                 *return_bi = wbi;
2531                                         }
2532                                         wbi = wbi2;
2533                                 }
2534                                 if (dev->towrite == NULL)
2535                                         bitmap_end = 1;
2536                                 spin_unlock_irq(&conf->device_lock);
2537                                 if (bitmap_end)
2538                                         bitmap_endwrite(conf->mddev->bitmap,
2539                                                         sh->sector,
2540                                                         STRIPE_SECTORS,
2541                                          !test_bit(STRIPE_DEGRADED, &sh->state),
2542                                                         0);
2543                         }
2544                 }
2545
2546         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2547                 if (atomic_dec_and_test(&conf->pending_full_writes))
2548                         md_wakeup_thread(conf->mddev->thread);
2549 }
2550
2551 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2552                 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2553 {
2554         int rmw = 0, rcw = 0, i;
2555         for (i = disks; i--; ) {
2556                 /* would I have to read this buffer for read_modify_write */
2557                 struct r5dev *dev = &sh->dev[i];
2558                 if ((dev->towrite || i == sh->pd_idx) &&
2559                     !test_bit(R5_LOCKED, &dev->flags) &&
2560                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2561                       test_bit(R5_Wantcompute, &dev->flags))) {
2562                         if (test_bit(R5_Insync, &dev->flags))
2563                                 rmw++;
2564                         else
2565                                 rmw += 2*disks;  /* cannot read it */
2566                 }
2567                 /* Would I have to read this buffer for reconstruct_write */
2568                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2569                     !test_bit(R5_LOCKED, &dev->flags) &&
2570                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2571                     test_bit(R5_Wantcompute, &dev->flags))) {
2572                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2573                         else
2574                                 rcw += 2*disks;
2575                 }
2576         }
2577         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2578                 (unsigned long long)sh->sector, rmw, rcw);
2579         set_bit(STRIPE_HANDLE, &sh->state);
2580         if (rmw < rcw && rmw > 0)
2581                 /* prefer read-modify-write, but need to get some data */
2582                 for (i = disks; i--; ) {
2583                         struct r5dev *dev = &sh->dev[i];
2584                         if ((dev->towrite || i == sh->pd_idx) &&
2585                             !test_bit(R5_LOCKED, &dev->flags) &&
2586                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2587                             test_bit(R5_Wantcompute, &dev->flags)) &&
2588                             test_bit(R5_Insync, &dev->flags)) {
2589                                 if (
2590                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2591                                         pr_debug("Read_old block "
2592                                                 "%d for r-m-w\n", i);
2593                                         set_bit(R5_LOCKED, &dev->flags);
2594                                         set_bit(R5_Wantread, &dev->flags);
2595                                         s->locked++;
2596                                 } else {
2597                                         set_bit(STRIPE_DELAYED, &sh->state);
2598                                         set_bit(STRIPE_HANDLE, &sh->state);
2599                                 }
2600                         }
2601                 }
2602         if (rcw <= rmw && rcw > 0)
2603                 /* want reconstruct write, but need to get some data */
2604                 for (i = disks; i--; ) {
2605                         struct r5dev *dev = &sh->dev[i];
2606                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2607                             i != sh->pd_idx &&
2608                             !test_bit(R5_LOCKED, &dev->flags) &&
2609                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2610                             test_bit(R5_Wantcompute, &dev->flags)) &&
2611                             test_bit(R5_Insync, &dev->flags)) {
2612                                 if (
2613                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2614                                         pr_debug("Read_old block "
2615                                                 "%d for Reconstruct\n", i);
2616                                         set_bit(R5_LOCKED, &dev->flags);
2617                                         set_bit(R5_Wantread, &dev->flags);
2618                                         s->locked++;
2619                                 } else {
2620                                         set_bit(STRIPE_DELAYED, &sh->state);
2621                                         set_bit(STRIPE_HANDLE, &sh->state);
2622                                 }
2623                         }
2624                 }
2625         /* now if nothing is locked, and if we have enough data,
2626          * we can start a write request
2627          */
2628         /* since handle_stripe can be called at any time we need to handle the
2629          * case where a compute block operation has been submitted and then a
2630          * subsequent call wants to start a write request.  raid_run_ops only
2631          * handles the case where compute block and reconstruct are requested
2632          * simultaneously.  If this is not the case then new writes need to be
2633          * held off until the compute completes.
2634          */
2635         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2636             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2637             !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2638                 schedule_reconstruction(sh, s, rcw == 0, 0);
2639 }
2640
2641 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2642                 struct stripe_head *sh, struct stripe_head_state *s,
2643                 struct r6_state *r6s, int disks)
2644 {
2645         int rcw = 0, pd_idx = sh->pd_idx, i;
2646         int qd_idx = sh->qd_idx;
2647
2648         set_bit(STRIPE_HANDLE, &sh->state);
2649         for (i = disks; i--; ) {
2650                 struct r5dev *dev = &sh->dev[i];
2651                 /* check if we haven't enough data */
2652                 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2653                     i != pd_idx && i != qd_idx &&
2654                     !test_bit(R5_LOCKED, &dev->flags) &&
2655                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2656                       test_bit(R5_Wantcompute, &dev->flags))) {
2657                         rcw++;
2658                         if (!test_bit(R5_Insync, &dev->flags))
2659                                 continue; /* it's a failed drive */
2660
2661                         if (
2662                           test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2663                                 pr_debug("Read_old stripe %llu "
2664                                         "block %d for Reconstruct\n",
2665                                      (unsigned long long)sh->sector, i);
2666                                 set_bit(R5_LOCKED, &dev->flags);
2667                                 set_bit(R5_Wantread, &dev->flags);
2668                                 s->locked++;
2669                         } else {
2670                                 pr_debug("Request delayed stripe %llu "
2671                                         "block %d for Reconstruct\n",
2672                                      (unsigned long long)sh->sector, i);
2673                                 set_bit(STRIPE_DELAYED, &sh->state);
2674                                 set_bit(STRIPE_HANDLE, &sh->state);
2675                         }
2676                 }
2677         }
2678         /* now if nothing is locked, and if we have enough data, we can start a
2679          * write request
2680          */
2681         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2682             s->locked == 0 && rcw == 0 &&
2683             !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2684                 schedule_reconstruction(sh, s, 1, 0);
2685         }
2686 }
2687
2688 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2689                                 struct stripe_head_state *s, int disks)
2690 {
2691         struct r5dev *dev = NULL;
2692
2693         set_bit(STRIPE_HANDLE, &sh->state);
2694
2695         switch (sh->check_state) {
2696         case check_state_idle:
2697                 /* start a new check operation if there are no failures */
2698                 if (s->failed == 0) {
2699                         BUG_ON(s->uptodate != disks);
2700                         sh->check_state = check_state_run;
2701                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
2702                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2703                         s->uptodate--;
2704                         break;
2705                 }
2706                 dev = &sh->dev[s->failed_num];
2707                 /* fall through */
2708         case check_state_compute_result:
2709                 sh->check_state = check_state_idle;
2710                 if (!dev)
2711                         dev = &sh->dev[sh->pd_idx];
2712
2713                 /* check that a write has not made the stripe insync */
2714                 if (test_bit(STRIPE_INSYNC, &sh->state))
2715                         break;
2716
2717                 /* either failed parity check, or recovery is happening */
2718                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2719                 BUG_ON(s->uptodate != disks);
2720
2721                 set_bit(R5_LOCKED, &dev->flags);
2722                 s->locked++;
2723                 set_bit(R5_Wantwrite, &dev->flags);
2724
2725                 clear_bit(STRIPE_DEGRADED, &sh->state);
2726                 set_bit(STRIPE_INSYNC, &sh->state);
2727                 break;
2728         case check_state_run:
2729                 break; /* we will be called again upon completion */
2730         case check_state_check_result:
2731                 sh->check_state = check_state_idle;
2732
2733                 /* if a failure occurred during the check operation, leave
2734                  * STRIPE_INSYNC not set and let the stripe be handled again
2735                  */
2736                 if (s->failed)
2737                         break;
2738
2739                 /* handle a successful check operation, if parity is correct
2740                  * we are done.  Otherwise update the mismatch count and repair
2741                  * parity if !MD_RECOVERY_CHECK
2742                  */
2743                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2744                         /* parity is correct (on disc,
2745                          * not in buffer any more)
2746                          */
2747                         set_bit(STRIPE_INSYNC, &sh->state);
2748                 else {
2749                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2750                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2751                                 /* don't try to repair!! */
2752                                 set_bit(STRIPE_INSYNC, &sh->state);
2753                         else {
2754                                 sh->check_state = check_state_compute_run;
2755                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2756                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2757                                 set_bit(R5_Wantcompute,
2758                                         &sh->dev[sh->pd_idx].flags);
2759                                 sh->ops.target = sh->pd_idx;
2760                                 sh->ops.target2 = -1;
2761                                 s->uptodate++;
2762                         }
2763                 }
2764                 break;
2765         case check_state_compute_run:
2766                 break;
2767         default:
2768                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2769                        __func__, sh->check_state,
2770                        (unsigned long long) sh->sector);
2771                 BUG();
2772         }
2773 }
2774
2775
2776 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2777                                   struct stripe_head_state *s,
2778                                   struct r6_state *r6s, int disks)
2779 {
2780         int pd_idx = sh->pd_idx;
2781         int qd_idx = sh->qd_idx;
2782         struct r5dev *dev;
2783
2784         set_bit(STRIPE_HANDLE, &sh->state);
2785
2786         BUG_ON(s->failed > 2);
2787
2788         /* Want to check and possibly repair P and Q.
2789          * However there could be one 'failed' device, in which
2790          * case we can only check one of them, possibly using the
2791          * other to generate missing data
2792          */
2793
2794         switch (sh->check_state) {
2795         case check_state_idle:
2796                 /* start a new check operation if there are < 2 failures */
2797                 if (s->failed == r6s->q_failed) {
2798                         /* The only possible failed device holds Q, so it
2799                          * makes sense to check P (If anything else were failed,
2800                          * we would have used P to recreate it).
2801                          */
2802                         sh->check_state = check_state_run;
2803                 }
2804                 if (!r6s->q_failed && s->failed < 2) {
2805                         /* Q is not failed, and we didn't use it to generate
2806                          * anything, so it makes sense to check it
2807                          */
2808                         if (sh->check_state == check_state_run)
2809                                 sh->check_state = check_state_run_pq;
2810                         else
2811                                 sh->check_state = check_state_run_q;
2812                 }
2813
2814                 /* discard potentially stale zero_sum_result */
2815                 sh->ops.zero_sum_result = 0;
2816
2817                 if (sh->check_state == check_state_run) {
2818                         /* async_xor_zero_sum destroys the contents of P */
2819                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2820                         s->uptodate--;
2821                 }
2822                 if (sh->check_state >= check_state_run &&
2823                     sh->check_state <= check_state_run_pq) {
2824                         /* async_syndrome_zero_sum preserves P and Q, so
2825                          * no need to mark them !uptodate here
2826                          */
2827                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
2828                         break;
2829                 }
2830
2831                 /* we have 2-disk failure */
2832                 BUG_ON(s->failed != 2);
2833                 /* fall through */
2834         case check_state_compute_result:
2835                 sh->check_state = check_state_idle;
2836
2837                 /* check that a write has not made the stripe insync */
2838                 if (test_bit(STRIPE_INSYNC, &sh->state))
2839                         break;
2840
2841                 /* now write out any block on a failed drive,
2842                  * or P or Q if they were recomputed
2843                  */
2844                 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2845                 if (s->failed == 2) {
2846                         dev = &sh->dev[r6s->failed_num[1]];
2847                         s->locked++;
2848                         set_bit(R5_LOCKED, &dev->flags);
2849                         set_bit(R5_Wantwrite, &dev->flags);
2850                 }
2851                 if (s->failed >= 1) {
2852                         dev = &sh->dev[r6s->failed_num[0]];
2853                         s->locked++;
2854                         set_bit(R5_LOCKED, &dev->flags);
2855                         set_bit(R5_Wantwrite, &dev->flags);
2856                 }
2857                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2858                         dev = &sh->dev[pd_idx];
2859                         s->locked++;
2860                         set_bit(R5_LOCKED, &dev->flags);
2861                         set_bit(R5_Wantwrite, &dev->flags);
2862                 }
2863                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2864                         dev = &sh->dev[qd_idx];
2865                         s->locked++;
2866                         set_bit(R5_LOCKED, &dev->flags);
2867                         set_bit(R5_Wantwrite, &dev->flags);
2868                 }
2869                 clear_bit(STRIPE_DEGRADED, &sh->state);
2870
2871                 set_bit(STRIPE_INSYNC, &sh->state);
2872                 break;
2873         case check_state_run:
2874         case check_state_run_q:
2875         case check_state_run_pq:
2876                 break; /* we will be called again upon completion */
2877         case check_state_check_result:
2878                 sh->check_state = check_state_idle;
2879
2880                 /* handle a successful check operation, if parity is correct
2881                  * we are done.  Otherwise update the mismatch count and repair
2882                  * parity if !MD_RECOVERY_CHECK
2883                  */
2884                 if (sh->ops.zero_sum_result == 0) {
2885                         /* both parities are correct */
2886                         if (!s->failed)
2887                                 set_bit(STRIPE_INSYNC, &sh->state);
2888                         else {
2889                                 /* in contrast to the raid5 case we can validate
2890                                  * parity, but still have a failure to write
2891                                  * back
2892                                  */
2893                                 sh->check_state = check_state_compute_result;
2894                                 /* Returning at this point means that we may go
2895                                  * off and bring p and/or q uptodate again so
2896                                  * we make sure to check zero_sum_result again
2897                                  * to verify if p or q need writeback
2898                                  */
2899                         }
2900                 } else {
2901                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2902                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2903                                 /* don't try to repair!! */
2904                                 set_bit(STRIPE_INSYNC, &sh->state);
2905                         else {
2906                                 int *target = &sh->ops.target;
2907
2908                                 sh->ops.target = -1;
2909                                 sh->ops.target2 = -1;
2910                                 sh->check_state = check_state_compute_run;
2911                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2912                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2913                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2914                                         set_bit(R5_Wantcompute,
2915                                                 &sh->dev[pd_idx].flags);
2916                                         *target = pd_idx;
2917                                         target = &sh->ops.target2;
2918                                         s->uptodate++;
2919                                 }
2920                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2921                                         set_bit(R5_Wantcompute,
2922                                                 &sh->dev[qd_idx].flags);
2923                                         *target = qd_idx;
2924                                         s->uptodate++;
2925                                 }
2926                         }
2927                 }
2928                 break;
2929         case check_state_compute_run:
2930                 break;
2931         default:
2932                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2933                        __func__, sh->check_state,
2934                        (unsigned long long) sh->sector);
2935                 BUG();
2936         }
2937 }
2938
2939 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2940                                 struct r6_state *r6s)
2941 {
2942         int i;
2943
2944         /* We have read all the blocks in this stripe and now we need to
2945          * copy some of them into a target stripe for expand.
2946          */
2947         struct dma_async_tx_descriptor *tx = NULL;
2948         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2949         for (i = 0; i < sh->disks; i++)
2950                 if (i != sh->pd_idx && i != sh->qd_idx) {
2951                         int dd_idx, j;
2952                         struct stripe_head *sh2;
2953                         struct async_submit_ctl submit;
2954
2955                         sector_t bn = compute_blocknr(sh, i, 1);
2956                         sector_t s = raid5_compute_sector(conf, bn, 0,
2957                                                           &dd_idx, NULL);
2958                         sh2 = get_active_stripe(conf, s, 0, 1, 1);
2959                         if (sh2 == NULL)
2960                                 /* so far only the early blocks of this stripe
2961                                  * have been requested.  When later blocks
2962                                  * get requested, we will try again
2963                                  */
2964                                 continue;
2965                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2966                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2967                                 /* must have already done this block */
2968                                 release_stripe(sh2);
2969                                 continue;
2970                         }
2971
2972                         /* place all the copies on one channel */
2973                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
2974                         tx = async_memcpy(sh2->dev[dd_idx].page,
2975                                           sh->dev[i].page, 0, 0, STRIPE_SIZE,
2976                                           &submit);
2977
2978                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2979                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2980                         for (j = 0; j < conf->raid_disks; j++)
2981                                 if (j != sh2->pd_idx &&
2982                                     (!r6s || j != sh2->qd_idx) &&
2983                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
2984                                         break;
2985                         if (j == conf->raid_disks) {
2986                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2987                                 set_bit(STRIPE_HANDLE, &sh2->state);
2988                         }
2989                         release_stripe(sh2);
2990
2991                 }
2992         /* done submitting copies, wait for them to complete */
2993         if (tx) {
2994                 async_tx_ack(tx);
2995                 dma_wait_for_async_tx(tx);
2996         }
2997 }
2998
2999
3000 /*
3001  * handle_stripe - do things to a stripe.
3002  *
3003  * We lock the stripe and then examine the state of various bits
3004  * to see what needs to be done.
3005  * Possible results:
3006  *    return some read request which now have data
3007  *    return some write requests which are safely on disc
3008  *    schedule a read on some buffers
3009  *    schedule a write of some buffers
3010  *    return confirmation of parity correctness
3011  *
3012  * buffers are taken off read_list or write_list, and bh_cache buffers
3013  * get BH_Lock set before the stripe lock is released.
3014  *
3015  */
3016
3017 static void handle_stripe5(struct stripe_head *sh)
3018 {
3019         raid5_conf_t *conf = sh->raid_conf;
3020         int disks = sh->disks, i;
3021         struct bio *return_bi = NULL;
3022         struct stripe_head_state s;
3023         struct r5dev *dev;
3024         mdk_rdev_t *blocked_rdev = NULL;
3025         int prexor;
3026         int dec_preread_active = 0;
3027
3028         memset(&s, 0, sizeof(s));
3029         pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
3030                  "reconstruct:%d\n", (unsigned long long)sh->sector, sh->state,
3031                  atomic_read(&sh->count), sh->pd_idx, sh->check_state,
3032                  sh->reconstruct_state);
3033
3034         spin_lock(&sh->lock);
3035         clear_bit(STRIPE_HANDLE, &sh->state);
3036         clear_bit(STRIPE_DELAYED, &sh->state);
3037
3038         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3039         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3040         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3041
3042         /* Now to look around and see what can be done */
3043         rcu_read_lock();
3044         for (i=disks; i--; ) {
3045                 mdk_rdev_t *rdev;
3046
3047                 dev = &sh->dev[i];
3048
3049                 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
3050                         "written %p\n", i, dev->flags, dev->toread, dev->read,
3051                         dev->towrite, dev->written);
3052
3053                 /* maybe we can request a biofill operation
3054                  *
3055                  * new wantfill requests are only permitted while
3056                  * ops_complete_biofill is guaranteed to be inactive
3057                  */
3058                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3059                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3060                         set_bit(R5_Wantfill, &dev->flags);
3061
3062                 /* now count some things */
3063                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3064                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3065                 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
3066
3067                 if (test_bit(R5_Wantfill, &dev->flags))
3068                         s.to_fill++;
3069                 else if (dev->toread)
3070                         s.to_read++;
3071                 if (dev->towrite) {
3072                         s.to_write++;
3073                         if (!test_bit(R5_OVERWRITE, &dev->flags))
3074                                 s.non_overwrite++;
3075                 }
3076                 if (dev->written)
3077                         s.written++;
3078                 rdev = rcu_dereference(conf->disks[i].rdev);
3079                 if (blocked_rdev == NULL &&
3080                     rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3081                         blocked_rdev = rdev;
3082                         atomic_inc(&rdev->nr_pending);
3083                 }
3084                 clear_bit(R5_Insync, &dev->flags);
3085                 if (!rdev)
3086                         /* Not in-sync */;
3087                 else if (test_bit(In_sync, &rdev->flags))
3088                         set_bit(R5_Insync, &dev->flags);
3089                 else {
3090                         /* could be in-sync depending on recovery/reshape status */
3091                         if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3092                                 set_bit(R5_Insync, &dev->flags);
3093                 }
3094                 if (!test_bit(R5_Insync, &dev->flags)) {
3095                         /* The ReadError flag will just be confusing now */
3096                         clear_bit(R5_ReadError, &dev->flags);
3097                         clear_bit(R5_ReWrite, &dev->flags);
3098                 }
3099                 if (test_bit(R5_ReadError, &dev->flags))
3100                         clear_bit(R5_Insync, &dev->flags);
3101                 if (!test_bit(R5_Insync, &dev->flags)) {
3102                         s.failed++;
3103                         s.failed_num = i;
3104                 }
3105         }
3106         rcu_read_unlock();
3107
3108         if (unlikely(blocked_rdev)) {
3109                 if (s.syncing || s.expanding || s.expanded ||
3110                     s.to_write || s.written) {
3111                         set_bit(STRIPE_HANDLE, &sh->state);
3112                         goto unlock;
3113                 }
3114                 /* There is nothing for the blocked_rdev to block */
3115                 rdev_dec_pending(blocked_rdev, conf->mddev);
3116                 blocked_rdev = NULL;
3117         }
3118
3119         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3120                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3121                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3122         }
3123
3124         pr_debug("locked=%d uptodate=%d to_read=%d"
3125                 " to_write=%d failed=%d failed_num=%d\n",
3126                 s.locked, s.uptodate, s.to_read, s.to_write,
3127                 s.failed, s.failed_num);
3128         /* check if the array has lost two devices and, if so, some requests might
3129          * need to be failed
3130          */
3131         if (s.failed > 1 && s.to_read+s.to_write+s.written)
3132                 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3133         if (s.failed > 1 && s.syncing) {
3134                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3135                 clear_bit(STRIPE_SYNCING, &sh->state);
3136                 s.syncing = 0;
3137         }
3138
3139         /* might be able to return some write requests if the parity block
3140          * is safe, or on a failed drive
3141          */
3142         dev = &sh->dev[sh->pd_idx];
3143         if ( s.written &&
3144              ((test_bit(R5_Insync, &dev->flags) &&
3145                !test_bit(R5_LOCKED, &dev->flags) &&
3146                test_bit(R5_UPTODATE, &dev->flags)) ||
3147                (s.failed == 1 && s.failed_num == sh->pd_idx)))
3148                 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3149
3150         /* Now we might consider reading some blocks, either to check/generate
3151          * parity, or to satisfy requests
3152          * or to load a block that is being partially written.
3153          */
3154         if (s.to_read || s.non_overwrite ||
3155             (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3156                 handle_stripe_fill5(sh, &s, disks);
3157
3158         /* Now we check to see if any write operations have recently
3159          * completed
3160          */
3161         prexor = 0;
3162         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3163                 prexor = 1;
3164         if (sh->reconstruct_state == reconstruct_state_drain_result ||
3165             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3166                 sh->reconstruct_state = reconstruct_state_idle;
3167
3168                 /* All the 'written' buffers and the parity block are ready to
3169                  * be written back to disk
3170                  */
3171                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3172                 for (i = disks; i--; ) {
3173                         dev = &sh->dev[i];
3174                         if (test_bit(R5_LOCKED, &dev->flags) &&
3175                                 (i == sh->pd_idx || dev->written)) {
3176                                 pr_debug("Writing block %d\n", i);
3177                                 set_bit(R5_Wantwrite, &dev->flags);
3178                                 if (prexor)
3179                                         continue;
3180                                 if (!test_bit(R5_Insync, &dev->flags) ||
3181                                     (i == sh->pd_idx && s.failed == 0))
3182                                         set_bit(STRIPE_INSYNC, &sh->state);
3183                         }
3184                 }
3185                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3186                         dec_preread_active = 1;
3187         }
3188
3189         /* Now to consider new write requests and what else, if anything
3190          * should be read.  We do not handle new writes when:
3191          * 1/ A 'write' operation (copy+xor) is already in flight.
3192          * 2/ A 'check' operation is in flight, as it may clobber the parity
3193          *    block.
3194          */
3195         if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3196                 handle_stripe_dirtying5(conf, sh, &s, disks);
3197
3198         /* maybe we need to check and possibly fix the parity for this stripe
3199          * Any reads will already have been scheduled, so we just see if enough
3200          * data is available.  The parity check is held off while parity
3201          * dependent operations are in flight.
3202          */
3203         if (sh->check_state ||
3204             (s.syncing && s.locked == 0 &&
3205              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3206              !test_bit(STRIPE_INSYNC, &sh->state)))
3207                 handle_parity_checks5(conf, sh, &s, disks);
3208
3209         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3210                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3211                 clear_bit(STRIPE_SYNCING, &sh->state);
3212         }
3213
3214         /* If the failed drive is just a ReadError, then we might need to progress
3215          * the repair/check process
3216          */
3217         if (s.failed == 1 && !conf->mddev->ro &&
3218             test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
3219             && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
3220             && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
3221                 ) {
3222                 dev = &sh->dev[s.failed_num];
3223                 if (!test_bit(R5_ReWrite, &dev->flags)) {
3224                         set_bit(R5_Wantwrite, &dev->flags);
3225                         set_bit(R5_ReWrite, &dev->flags);
3226                         set_bit(R5_LOCKED, &dev->flags);
3227                         s.locked++;
3228                 } else {
3229                         /* let's read it back */
3230                         set_bit(R5_Wantread, &dev->flags);
3231                         set_bit(R5_LOCKED, &dev->flags);
3232                         s.locked++;
3233                 }
3234         }
3235
3236         /* Finish reconstruct operations initiated by the expansion process */
3237         if (sh->reconstruct_state == reconstruct_state_result) {
3238                 struct stripe_head *sh2
3239                         = get_active_stripe(conf, sh->sector, 1, 1, 1);
3240                 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3241                         /* sh cannot be written until sh2 has been read.
3242                          * so arrange for sh to be delayed a little
3243                          */
3244                         set_bit(STRIPE_DELAYED, &sh->state);
3245                         set_bit(STRIPE_HANDLE, &sh->state);
3246                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3247                                               &sh2->state))
3248                                 atomic_inc(&conf->preread_active_stripes);
3249                         release_stripe(sh2);
3250                         goto unlock;
3251                 }
3252                 if (sh2)
3253                         release_stripe(sh2);
3254
3255                 sh->reconstruct_state = reconstruct_state_idle;
3256                 clear_bit(STRIPE_EXPANDING, &sh->state);
3257                 for (i = conf->raid_disks; i--; ) {
3258                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
3259                         set_bit(R5_LOCKED, &sh->dev[i].flags);
3260                         s.locked++;
3261                 }
3262         }
3263
3264         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3265             !sh->reconstruct_state) {
3266                 /* Need to write out all blocks after computing parity */
3267                 sh->disks = conf->raid_disks;
3268                 stripe_set_idx(sh->sector, conf, 0, sh);
3269                 schedule_reconstruction(sh, &s, 1, 1);
3270         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3271                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3272                 atomic_dec(&conf->reshape_stripes);
3273                 wake_up(&conf->wait_for_overlap);
3274                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3275         }
3276
3277         if (s.expanding && s.locked == 0 &&
3278             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3279                 handle_stripe_expansion(conf, sh, NULL);
3280
3281  unlock:
3282         spin_unlock(&sh->lock);
3283
3284         /* wait for this device to become unblocked */
3285         if (unlikely(blocked_rdev))
3286                 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3287
3288         if (s.ops_request)
3289                 raid_run_ops(sh, s.ops_request);
3290
3291         ops_run_io(sh, &s);
3292
3293         if (dec_preread_active) {
3294                 /* We delay this until after ops_run_io so that if make_request
3295                  * is waiting on a flush, it won't continue until the writes
3296                  * have actually been submitted.
3297                  */
3298                 atomic_dec(&conf->preread_active_stripes);
3299                 if (atomic_read(&conf->preread_active_stripes) <
3300                     IO_THRESHOLD)
3301                         md_wakeup_thread(conf->mddev->thread);
3302         }
3303         return_io(return_bi);
3304 }
3305
3306 static void handle_stripe6(struct stripe_head *sh)
3307 {
3308         raid5_conf_t *conf = sh->raid_conf;
3309         int disks = sh->disks;
3310         struct bio *return_bi = NULL;
3311         int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx;
3312         struct stripe_head_state s;
3313         struct r6_state r6s;
3314         struct r5dev *dev, *pdev, *qdev;
3315         mdk_rdev_t *blocked_rdev = NULL;
3316         int dec_preread_active = 0;
3317
3318         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3319                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3320                (unsigned long long)sh->sector, sh->state,
3321                atomic_read(&sh->count), pd_idx, qd_idx,
3322                sh->check_state, sh->reconstruct_state);
3323         memset(&s, 0, sizeof(s));
3324
3325         spin_lock(&sh->lock);
3326         clear_bit(STRIPE_HANDLE, &sh->state);
3327         clear_bit(STRIPE_DELAYED, &sh->state);
3328
3329         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3330         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3331         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3332         /* Now to look around and see what can be done */
3333
3334         rcu_read_lock();
3335         for (i=disks; i--; ) {
3336                 mdk_rdev_t *rdev;
3337                 dev = &sh->dev[i];
3338
3339                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3340                         i, dev->flags, dev->toread, dev->towrite, dev->written);
3341                 /* maybe we can reply to a read
3342                  *
3343                  * new wantfill requests are only permitted while
3344                  * ops_complete_biofill is guaranteed to be inactive
3345                  */
3346                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3347                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3348                         set_bit(R5_Wantfill, &dev->flags);
3349
3350                 /* now count some things */
3351                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3352                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3353                 if (test_bit(R5_Wantcompute, &dev->flags)) {
3354                         s.compute++;
3355                         BUG_ON(s.compute > 2);
3356                 }
3357
3358                 if (test_bit(R5_Wantfill, &dev->flags)) {
3359                         s.to_fill++;
3360                 } else if (dev->toread)
3361                         s.to_read++;
3362                 if (dev->towrite) {
3363                         s.to_write++;
3364                         if (!test_bit(R5_OVERWRITE, &dev->flags))
3365                                 s.non_overwrite++;
3366                 }
3367                 if (dev->written)
3368                         s.written++;
3369                 rdev = rcu_dereference(conf->disks[i].rdev);
3370                 if (blocked_rdev == NULL &&
3371                     rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3372                         blocked_rdev = rdev;
3373                         atomic_inc(&rdev->nr_pending);
3374                 }
3375                 clear_bit(R5_Insync, &dev->flags);
3376                 if (!rdev)
3377                         /* Not in-sync */;
3378                 else if (test_bit(In_sync, &rdev->flags))
3379                         set_bit(R5_Insync, &dev->flags);
3380                 else {
3381                         /* in sync if before recovery_offset */
3382                         if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3383                                 set_bit(R5_Insync, &dev->flags);
3384                 }
3385                 if (!test_bit(R5_Insync, &dev->flags)) {
3386                         /* The ReadError flag will just be confusing now */
3387                         clear_bit(R5_ReadError, &dev->flags);
3388                         clear_bit(R5_ReWrite, &dev->flags);
3389                 }
3390                 if (test_bit(R5_ReadError, &dev->flags))
3391                         clear_bit(R5_Insync, &dev->flags);
3392                 if (!test_bit(R5_Insync, &dev->flags)) {
3393                         if (s.failed < 2)
3394                                 r6s.failed_num[s.failed] = i;
3395                         s.failed++;
3396                 }
3397         }
3398         rcu_read_unlock();
3399
3400         if (unlikely(blocked_rdev)) {
3401                 if (s.syncing || s.expanding || s.expanded ||
3402                     s.to_write || s.written) {
3403                         set_bit(STRIPE_HANDLE, &sh->state);
3404                         goto unlock;
3405                 }
3406                 /* There is nothing for the blocked_rdev to block */
3407                 rdev_dec_pending(blocked_rdev, conf->mddev);
3408                 blocked_rdev = NULL;
3409         }
3410
3411         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3412                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3413                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3414         }
3415
3416         pr_debug("locked=%d uptodate=%d to_read=%d"
3417                " to_write=%d failed=%d failed_num=%d,%d\n",
3418                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3419                r6s.failed_num[0], r6s.failed_num[1]);
3420         /* check if the array has lost >2 devices and, if so, some requests
3421          * might need to be failed
3422          */
3423         if (s.failed > 2 && s.to_read+s.to_write+s.written)
3424                 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3425         if (s.failed > 2 && s.syncing) {
3426                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3427                 clear_bit(STRIPE_SYNCING, &sh->state);
3428                 s.syncing = 0;
3429         }
3430
3431         /*
3432          * might be able to return some write requests if the parity blocks
3433          * are safe, or on a failed drive
3434          */
3435         pdev = &sh->dev[pd_idx];
3436         r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3437                 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3438         qdev = &sh->dev[qd_idx];
3439         r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == qd_idx)
3440                 || (s.failed >= 2 && r6s.failed_num[1] == qd_idx);
3441
3442         if ( s.written &&
3443              ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3444                              && !test_bit(R5_LOCKED, &pdev->flags)
3445                              && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3446              ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3447                              && !test_bit(R5_LOCKED, &qdev->flags)
3448                              && test_bit(R5_UPTODATE, &qdev->flags)))))
3449                 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3450
3451         /* Now we might consider reading some blocks, either to check/generate
3452          * parity, or to satisfy requests
3453          * or to load a block that is being partially written.
3454          */
3455         if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3456             (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3457                 handle_stripe_fill6(sh, &s, &r6s, disks);
3458
3459         /* Now we check to see if any write operations have recently
3460          * completed
3461          */
3462         if (sh->reconstruct_state == reconstruct_state_drain_result) {
3463
3464                 sh->reconstruct_state = reconstruct_state_idle;
3465                 /* All the 'written' buffers and the parity blocks are ready to
3466                  * be written back to disk
3467                  */
3468                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3469                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags));
3470                 for (i = disks; i--; ) {
3471                         dev = &sh->dev[i];
3472                         if (test_bit(R5_LOCKED, &dev->flags) &&
3473                             (i == sh->pd_idx || i == qd_idx ||
3474                              dev->written)) {
3475                                 pr_debug("Writing block %d\n", i);
3476                                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3477                                 set_bit(R5_Wantwrite, &dev->flags);
3478                                 if (!test_bit(R5_Insync, &dev->flags) ||
3479                                     ((i == sh->pd_idx || i == qd_idx) &&
3480                                       s.failed == 0))
3481                                         set_bit(STRIPE_INSYNC, &sh->state);
3482                         }
3483                 }
3484                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3485                         dec_preread_active = 1;
3486         }
3487
3488         /* Now to consider new write requests and what else, if anything
3489          * should be read.  We do not handle new writes when:
3490          * 1/ A 'write' operation (copy+gen_syndrome) is already in flight.
3491          * 2/ A 'check' operation is in flight, as it may clobber the parity
3492          *    block.
3493          */
3494         if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3495                 handle_stripe_dirtying6(conf, sh, &s, &r6s, disks);
3496
3497         /* maybe we need to check and possibly fix the parity for this stripe
3498          * Any reads will already have been scheduled, so we just see if enough
3499          * data is available.  The parity check is held off while parity
3500          * dependent operations are in flight.
3501          */
3502         if (sh->check_state ||
3503             (s.syncing && s.locked == 0 &&
3504              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3505              !test_bit(STRIPE_INSYNC, &sh->state)))
3506                 handle_parity_checks6(conf, sh, &s, &r6s, disks);
3507
3508         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3509                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3510                 clear_bit(STRIPE_SYNCING, &sh->state);
3511         }
3512
3513         /* If the failed drives are just a ReadError, then we might need
3514          * to progress the repair/check process
3515          */
3516         if (s.failed <= 2 && !conf->mddev->ro)
3517                 for (i = 0; i < s.failed; i++) {
3518                         dev = &sh->dev[r6s.failed_num[i]];
3519                         if (test_bit(R5_ReadError, &dev->flags)
3520                             && !test_bit(R5_LOCKED, &dev->flags)
3521                             && test_bit(R5_UPTODATE, &dev->flags)
3522                                 ) {
3523                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
3524                                         set_bit(R5_Wantwrite, &dev->flags);
3525                                         set_bit(R5_ReWrite, &dev->flags);
3526                                         set_bit(R5_LOCKED, &dev->flags);
3527                                         s.locked++;
3528                                 } else {
3529                                         /* let's read it back */
3530                                         set_bit(R5_Wantread, &dev->flags);
3531                                         set_bit(R5_LOCKED, &dev->flags);
3532                                         s.locked++;
3533                                 }
3534                         }
3535                 }
3536
3537         /* Finish reconstruct operations initiated by the expansion process */
3538         if (sh->reconstruct_state == reconstruct_state_result) {
3539                 sh->reconstruct_state = reconstruct_state_idle;
3540                 clear_bit(STRIPE_EXPANDING, &sh->state);
3541                 for (i = conf->raid_disks; i--; ) {
3542                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
3543                         set_bit(R5_LOCKED, &sh->dev[i].flags);
3544                         s.locked++;
3545                 }
3546         }
3547
3548         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3549             !sh->reconstruct_state) {
3550                 struct stripe_head *sh2
3551                         = get_active_stripe(conf, sh->sector, 1, 1, 1);
3552                 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3553                         /* sh cannot be written until sh2 has been read.
3554                          * so arrange for sh to be delayed a little
3555                          */
3556                         set_bit(STRIPE_DELAYED, &sh->state);
3557                         set_bit(STRIPE_HANDLE, &sh->state);
3558                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3559                                               &sh2->state))
3560                                 atomic_inc(&conf->preread_active_stripes);
3561                         release_stripe(sh2);
3562                         goto unlock;
3563                 }
3564                 if (sh2)
3565                         release_stripe(sh2);
3566
3567                 /* Need to write out all blocks after computing P&Q */
3568                 sh->disks = conf->raid_disks;
3569                 stripe_set_idx(sh->sector, conf, 0, sh);
3570                 schedule_reconstruction(sh, &s, 1, 1);
3571         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3572                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3573                 atomic_dec(&conf->reshape_stripes);
3574                 wake_up(&conf->wait_for_overlap);
3575                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3576         }
3577
3578         if (s.expanding && s.locked == 0 &&
3579             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3580                 handle_stripe_expansion(conf, sh, &r6s);
3581
3582  unlock:
3583         spin_unlock(&sh->lock);
3584
3585         /* wait for this device to become unblocked */
3586         if (unlikely(blocked_rdev))
3587                 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3588
3589         if (s.ops_request)
3590                 raid_run_ops(sh, s.ops_request);
3591
3592         ops_run_io(sh, &s);
3593
3594
3595         if (dec_preread_active) {
3596                 /* We delay this until after ops_run_io so that if make_request
3597                  * is waiting on a flush, it won't continue until the writes
3598                  * have actually been submitted.
3599                  */
3600                 atomic_dec(&conf->preread_active_stripes);
3601                 if (atomic_read(&conf->preread_active_stripes) <
3602                     IO_THRESHOLD)
3603                         md_wakeup_thread(conf->mddev->thread);
3604         }
3605
3606         return_io(return_bi);
3607 }
3608
3609 static void handle_stripe(struct stripe_head *sh)
3610 {
3611         if (sh->raid_conf->level == 6)
3612                 handle_stripe6(sh);
3613         else
3614                 handle_stripe5(sh);
3615 }
3616
3617 static void raid5_activate_delayed(raid5_conf_t *conf)
3618 {
3619         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3620                 while (!list_empty(&conf->delayed_list)) {
3621                         struct list_head *l = conf->delayed_list.next;
3622                         struct stripe_head *sh;
3623                         sh = list_entry(l, struct stripe_head, lru);
3624                         list_del_init(l);
3625                         clear_bit(STRIPE_DELAYED, &sh->state);
3626                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3627                                 atomic_inc(&conf->preread_active_stripes);
3628                         list_add_tail(&sh->lru, &conf->hold_list);
3629                 }
3630         } else
3631                 plugger_set_plug(&conf->plug);
3632 }
3633
3634 static void activate_bit_delay(raid5_conf_t *conf)
3635 {
3636         /* device_lock is held */
3637         struct list_head head;
3638         list_add(&head, &conf->bitmap_list);
3639         list_del_init(&conf->bitmap_list);
3640         while (!list_empty(&head)) {
3641                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3642                 list_del_init(&sh->lru);
3643                 atomic_inc(&sh->count);
3644                 __release_stripe(conf, sh);
3645         }
3646 }
3647
3648 static void unplug_slaves(mddev_t *mddev)
3649 {
3650         raid5_conf_t *conf = mddev->private;
3651         int i;
3652         int devs = max(conf->raid_disks, conf->previous_raid_disks);
3653
3654         rcu_read_lock();
3655         for (i = 0; i < devs; i++) {
3656                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3657                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3658                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3659
3660                         atomic_inc(&rdev->nr_pending);
3661                         rcu_read_unlock();
3662
3663                         blk_unplug(r_queue);
3664
3665                         rdev_dec_pending(rdev, mddev);
3666                         rcu_read_lock();
3667                 }
3668         }
3669         rcu_read_unlock();
3670 }
3671
3672 void md_raid5_unplug_device(raid5_conf_t *conf)
3673 {
3674         unsigned long flags;
3675
3676         spin_lock_irqsave(&conf->device_lock, flags);
3677
3678         if (plugger_remove_plug(&conf->plug)) {
3679                 conf->seq_flush++;
3680                 raid5_activate_delayed(conf);
3681         }
3682         md_wakeup_thread(conf->mddev->thread);
3683
3684         spin_unlock_irqrestore(&conf->device_lock, flags);
3685
3686         unplug_slaves(conf->mddev);
3687 }
3688 EXPORT_SYMBOL_GPL(md_raid5_unplug_device);
3689
3690 static void raid5_unplug(struct plug_handle *plug)
3691 {
3692         raid5_conf_t *conf = container_of(plug, raid5_conf_t, plug);
3693         md_raid5_unplug_device(conf);
3694 }
3695
3696 static void raid5_unplug_queue(struct request_queue *q)
3697 {
3698         mddev_t *mddev = q->queuedata;
3699         md_raid5_unplug_device(mddev->private);
3700 }
3701
3702 int md_raid5_congested(mddev_t *mddev, int bits)
3703 {
3704         raid5_conf_t *conf = mddev->private;
3705
3706         /* No difference between reads and writes.  Just check
3707          * how busy the stripe_cache is
3708          */
3709
3710         if (conf->inactive_blocked)
3711                 return 1;
3712         if (conf->quiesce)
3713                 return 1;
3714         if (list_empty_careful(&conf->inactive_list))
3715                 return 1;
3716
3717         return 0;
3718 }
3719 EXPORT_SYMBOL_GPL(md_raid5_congested);
3720
3721 static int raid5_congested(void *data, int bits)
3722 {
3723         mddev_t *mddev = data;
3724
3725         return mddev_congested(mddev, bits) ||
3726                 md_raid5_congested(mddev, bits);
3727 }
3728
3729 /* We want read requests to align with chunks where possible,
3730  * but write requests don't need to.
3731  */
3732 static int raid5_mergeable_bvec(struct request_queue *q,
3733                                 struct bvec_merge_data *bvm,
3734                                 struct bio_vec *biovec)
3735 {
3736         mddev_t *mddev = q->queuedata;
3737         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3738         int max;
3739         unsigned int chunk_sectors = mddev->chunk_sectors;
3740         unsigned int bio_sectors = bvm->bi_size >> 9;
3741
3742         if ((bvm->bi_rw & 1) == WRITE)
3743                 return biovec->bv_len; /* always allow writes to be mergeable */
3744
3745         if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3746                 chunk_sectors = mddev->new_chunk_sectors;
3747         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3748         if (max < 0) max = 0;
3749         if (max <= biovec->bv_len && bio_sectors == 0)
3750                 return biovec->bv_len;
3751         else
3752                 return max;
3753 }
3754
3755
3756 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3757 {
3758         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3759         unsigned int chunk_sectors = mddev->chunk_sectors;
3760         unsigned int bio_sectors = bio->bi_size >> 9;
3761
3762         if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3763                 chunk_sectors = mddev->new_chunk_sectors;
3764         return  chunk_sectors >=
3765                 ((sector & (chunk_sectors - 1)) + bio_sectors);
3766 }
3767
3768 /*
3769  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3770  *  later sampled by raid5d.
3771  */
3772 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3773 {
3774         unsigned long flags;
3775
3776         spin_lock_irqsave(&conf->device_lock, flags);
3777
3778         bi->bi_next = conf->retry_read_aligned_list;
3779         conf->retry_read_aligned_list = bi;
3780
3781         spin_unlock_irqrestore(&conf->device_lock, flags);
3782         md_wakeup_thread(conf->mddev->thread);
3783 }
3784
3785
3786 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3787 {
3788         struct bio *bi;
3789
3790         bi = conf->retry_read_aligned;
3791         if (bi) {
3792                 conf->retry_read_aligned = NULL;
3793                 return bi;
3794         }
3795         bi = conf->retry_read_aligned_list;
3796         if(bi) {
3797                 conf->retry_read_aligned_list = bi->bi_next;
3798                 bi->bi_next = NULL;
3799                 /*
3800                  * this sets the active strip count to 1 and the processed
3801                  * strip count to zero (upper 8 bits)
3802                  */
3803                 bi->bi_phys_segments = 1; /* biased count of active stripes */
3804         }
3805
3806         return bi;
3807 }
3808
3809
3810 /*
3811  *  The "raid5_align_endio" should check if the read succeeded and if it
3812  *  did, call bio_endio on the original bio (having bio_put the new bio
3813  *  first).
3814  *  If the read failed..
3815  */
3816 static void raid5_align_endio(struct bio *bi, int error)
3817 {
3818         struct bio* raid_bi  = bi->bi_private;
3819         mddev_t *mddev;
3820         raid5_conf_t *conf;
3821         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3822         mdk_rdev_t *rdev;
3823
3824         bio_put(bi);
3825
3826         rdev = (void*)raid_bi->bi_next;
3827         raid_bi->bi_next = NULL;
3828         mddev = rdev->mddev;
3829         conf = mddev->private;
3830
3831         rdev_dec_pending(rdev, conf->mddev);
3832
3833         if (!error && uptodate) {
3834                 bio_endio(raid_bi, 0);
3835                 if (atomic_dec_and_test(&conf->active_aligned_reads))
3836                         wake_up(&conf->wait_for_stripe);
3837                 return;
3838         }
3839
3840
3841         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3842
3843         add_bio_to_retry(raid_bi, conf);
3844 }
3845
3846 static int bio_fits_rdev(struct bio *bi)
3847 {
3848         struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3849
3850         if ((bi->bi_size>>9) > queue_max_sectors(q))
3851                 return 0;
3852         blk_recount_segments(q, bi);
3853         if (bi->bi_phys_segments > queue_max_segments(q))
3854                 return 0;
3855
3856         if (q->merge_bvec_fn)
3857                 /* it's too hard to apply the merge_bvec_fn at this stage,
3858                  * just just give up
3859                  */
3860                 return 0;
3861
3862         return 1;
3863 }
3864
3865
3866 static int chunk_aligned_read(mddev_t *mddev, struct bio * raid_bio)
3867 {
3868         raid5_conf_t *conf = mddev->private;
3869         int dd_idx;
3870         struct bio* align_bi;
3871         mdk_rdev_t *rdev;
3872
3873         if (!in_chunk_boundary(mddev, raid_bio)) {
3874                 pr_debug("chunk_aligned_read : non aligned\n");
3875                 return 0;
3876         }
3877         /*
3878          * use bio_clone_mddev to make a copy of the bio
3879          */
3880         align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
3881         if (!align_bi)
3882                 return 0;
3883         /*
3884          *   set bi_end_io to a new function, and set bi_private to the
3885          *     original bio.
3886          */
3887         align_bi->bi_end_io  = raid5_align_endio;
3888         align_bi->bi_private = raid_bio;
3889         /*
3890          *      compute position
3891          */
3892         align_bi->bi_sector =  raid5_compute_sector(conf, raid_bio->bi_sector,
3893                                                     0,
3894                                                     &dd_idx, NULL);
3895
3896         rcu_read_lock();
3897         rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3898         if (rdev && test_bit(In_sync, &rdev->flags)) {
3899                 atomic_inc(&rdev->nr_pending);
3900                 rcu_read_unlock();
3901                 raid_bio->bi_next = (void*)rdev;
3902                 align_bi->bi_bdev =  rdev->bdev;
3903                 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3904                 align_bi->bi_sector += rdev->data_offset;
3905
3906                 if (!bio_fits_rdev(align_bi)) {
3907                         /* too big in some way */
3908                         bio_put(align_bi);
3909                         rdev_dec_pending(rdev, mddev);
3910                         return 0;
3911                 }
3912
3913                 spin_lock_irq(&conf->device_lock);
3914                 wait_event_lock_irq(conf->wait_for_stripe,
3915                                     conf->quiesce == 0,
3916                                     conf->device_lock, /* nothing */);
3917                 atomic_inc(&conf->active_aligned_reads);
3918                 spin_unlock_irq(&conf->device_lock);
3919
3920                 generic_make_request(align_bi);
3921                 return 1;
3922         } else {
3923                 rcu_read_unlock();
3924                 bio_put(align_bi);
3925                 return 0;
3926         }
3927 }
3928
3929 /* __get_priority_stripe - get the next stripe to process
3930  *
3931  * Full stripe writes are allowed to pass preread active stripes up until
3932  * the bypass_threshold is exceeded.  In general the bypass_count
3933  * increments when the handle_list is handled before the hold_list; however, it
3934  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3935  * stripe with in flight i/o.  The bypass_count will be reset when the
3936  * head of the hold_list has changed, i.e. the head was promoted to the
3937  * handle_list.
3938  */
3939 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3940 {
3941         struct stripe_head *sh;
3942
3943         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3944                   __func__,
3945                   list_empty(&conf->handle_list) ? "empty" : "busy",
3946                   list_empty(&conf->hold_list) ? "empty" : "busy",
3947                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
3948
3949         if (!list_empty(&conf->handle_list)) {
3950                 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3951
3952                 if (list_empty(&conf->hold_list))
3953                         conf->bypass_count = 0;
3954                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3955                         if (conf->hold_list.next == conf->last_hold)
3956                                 conf->bypass_count++;
3957                         else {
3958                                 conf->last_hold = conf->hold_list.next;
3959                                 conf->bypass_count -= conf->bypass_threshold;
3960                                 if (conf->bypass_count < 0)
3961                                         conf->bypass_count = 0;
3962                         }
3963                 }
3964         } else if (!list_empty(&conf->hold_list) &&
3965                    ((conf->bypass_threshold &&
3966                      conf->bypass_count > conf->bypass_threshold) ||
3967                     atomic_read(&conf->pending_full_writes) == 0)) {
3968                 sh = list_entry(conf->hold_list.next,
3969                                 typeof(*sh), lru);
3970                 conf->bypass_count -= conf->bypass_threshold;
3971                 if (conf->bypass_count < 0)
3972                         conf->bypass_count = 0;
3973         } else
3974                 return NULL;
3975
3976         list_del_init(&sh->lru);
3977         atomic_inc(&sh->count);
3978         BUG_ON(atomic_read(&sh->count) != 1);
3979         return sh;
3980 }
3981
3982 static int make_request(mddev_t *mddev, struct bio * bi)
3983 {
3984         raid5_conf_t *conf = mddev->private;
3985         int dd_idx;
3986         sector_t new_sector;
3987         sector_t logical_sector, last_sector;
3988         struct stripe_head *sh;
3989         const int rw = bio_data_dir(bi);
3990         int remaining;
3991
3992         if (unlikely(bi->bi_rw & REQ_FLUSH)) {
3993                 md_flush_request(mddev, bi);
3994                 return 0;
3995         }
3996
3997         md_write_start(mddev, bi);
3998
3999         if (rw == READ &&
4000              mddev->reshape_position == MaxSector &&
4001              chunk_aligned_read(mddev,bi))
4002                 return 0;
4003
4004         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4005         last_sector = bi->bi_sector + (bi->bi_size>>9);
4006         bi->bi_next = NULL;
4007         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
4008
4009         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
4010                 DEFINE_WAIT(w);
4011                 int disks, data_disks;
4012                 int previous;
4013
4014         retry:
4015                 previous = 0;
4016                 disks = conf->raid_disks;
4017                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
4018                 if (unlikely(conf->reshape_progress != MaxSector)) {
4019                         /* spinlock is needed as reshape_progress may be
4020                          * 64bit on a 32bit platform, and so it might be
4021                          * possible to see a half-updated value
4022                          * Ofcourse reshape_progress could change after
4023                          * the lock is dropped, so once we get a reference
4024                          * to the stripe that we think it is, we will have
4025                          * to check again.
4026                          */
4027                         spin_lock_irq(&conf->device_lock);
4028                         if (mddev->delta_disks < 0
4029                             ? logical_sector < conf->reshape_progress
4030                             : logical_sector >= conf->reshape_progress) {
4031                                 disks = conf->previous_raid_disks;
4032                                 previous = 1;
4033                         } else {
4034                                 if (mddev->delta_disks < 0
4035                                     ? logical_sector < conf->reshape_safe
4036                                     : logical_sector >= conf->reshape_safe) {
4037                                         spin_unlock_irq(&conf->device_lock);
4038                                         schedule();
4039                                         goto retry;
4040                                 }
4041                         }
4042                         spin_unlock_irq(&conf->device_lock);
4043                 }
4044                 data_disks = disks - conf->max_degraded;
4045
4046                 new_sector = raid5_compute_sector(conf, logical_sector,
4047                                                   previous,
4048                                                   &dd_idx, NULL);
4049                 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4050                         (unsigned long long)new_sector, 
4051                         (unsigned long long)logical_sector);
4052
4053                 sh = get_active_stripe(conf, new_sector, previous,
4054                                        (bi->bi_rw&RWA_MASK), 0);
4055                 if (sh) {
4056                         if (unlikely(previous)) {
4057                                 /* expansion might have moved on while waiting for a
4058                                  * stripe, so we must do the range check again.
4059                                  * Expansion could still move past after this
4060                                  * test, but as we are holding a reference to
4061                                  * 'sh', we know that if that happens,
4062                                  *  STRIPE_EXPANDING will get set and the expansion
4063                                  * won't proceed until we finish with the stripe.
4064                                  */
4065                                 int must_retry = 0;
4066                                 spin_lock_irq(&conf->device_lock);
4067                                 if (mddev->delta_disks < 0
4068                                     ? logical_sector >= conf->reshape_progress
4069                                     : logical_sector < conf->reshape_progress)
4070                                         /* mismatch, need to try again */
4071                                         must_retry = 1;
4072                                 spin_unlock_irq(&conf->device_lock);
4073                                 if (must_retry) {
4074                                         release_stripe(sh);
4075                                         schedule();
4076                                         goto retry;
4077                                 }
4078                         }
4079
4080                         if (bio_data_dir(bi) == WRITE &&
4081                             logical_sector >= mddev->suspend_lo &&
4082                             logical_sector < mddev->suspend_hi) {
4083                                 release_stripe(sh);
4084                                 /* As the suspend_* range is controlled by
4085                                  * userspace, we want an interruptible
4086                                  * wait.
4087                                  */
4088                                 flush_signals(current);
4089                                 prepare_to_wait(&conf->wait_for_overlap,
4090                                                 &w, TASK_INTERRUPTIBLE);
4091                                 if (logical_sector >= mddev->suspend_lo &&
4092                                     logical_sector < mddev->suspend_hi)
4093                                         schedule();
4094                                 goto retry;
4095                         }
4096
4097                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
4098                             !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
4099                                 /* Stripe is busy expanding or
4100                                  * add failed due to overlap.  Flush everything
4101                                  * and wait a while
4102                                  */
4103                                 md_raid5_unplug_device(conf);
4104                                 release_stripe(sh);
4105                                 schedule();
4106                                 goto retry;
4107                         }
4108                         finish_wait(&conf->wait_for_overlap, &w);
4109                         set_bit(STRIPE_HANDLE, &sh->state);
4110                         clear_bit(STRIPE_DELAYED, &sh->state);
4111                         if ((bi->bi_rw & REQ_SYNC) &&
4112                             !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4113                                 atomic_inc(&conf->preread_active_stripes);
4114                         release_stripe(sh);
4115                 } else {
4116                         /* cannot get stripe for read-ahead, just give-up */
4117                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
4118                         finish_wait(&conf->wait_for_overlap, &w);
4119                         break;
4120                 }
4121                         
4122         }
4123         spin_lock_irq(&conf->device_lock);
4124         remaining = raid5_dec_bi_phys_segments(bi);
4125         spin_unlock_irq(&conf->device_lock);
4126         if (remaining == 0) {
4127
4128                 if ( rw == WRITE )
4129                         md_write_end(mddev);
4130
4131                 bio_endio(bi, 0);
4132         }
4133
4134         return 0;
4135 }
4136
4137 static sector_t raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks);
4138
4139 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
4140 {
4141         /* reshaping is quite different to recovery/resync so it is
4142          * handled quite separately ... here.
4143          *
4144          * On each call to sync_request, we gather one chunk worth of
4145          * destination stripes and flag them as expanding.
4146          * Then we find all the source stripes and request reads.
4147          * As the reads complete, handle_stripe will copy the data
4148          * into the destination stripe and release that stripe.
4149          */
4150         raid5_conf_t *conf = mddev->private;
4151         struct stripe_head *sh;
4152         sector_t first_sector, last_sector;
4153         int raid_disks = conf->previous_raid_disks;
4154         int data_disks = raid_disks - conf->max_degraded;
4155         int new_data_disks = conf->raid_disks - conf->max_degraded;
4156         int i;
4157         int dd_idx;
4158         sector_t writepos, readpos, safepos;
4159         sector_t stripe_addr;
4160         int reshape_sectors;
4161         struct list_head stripes;
4162
4163         if (sector_nr == 0) {
4164                 /* If restarting in the middle, skip the initial sectors */
4165                 if (mddev->delta_disks < 0 &&
4166                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4167                         sector_nr = raid5_size(mddev, 0, 0)
4168                                 - conf->reshape_progress;
4169                 } else if (mddev->delta_disks >= 0 &&
4170                            conf->reshape_progress > 0)
4171                         sector_nr = conf->reshape_progress;
4172                 sector_div(sector_nr, new_data_disks);
4173                 if (sector_nr) {
4174                         mddev->curr_resync_completed = sector_nr;
4175                         sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4176                         *skipped = 1;
4177                         return sector_nr;
4178                 }
4179         }
4180
4181         /* We need to process a full chunk at a time.
4182          * If old and new chunk sizes differ, we need to process the
4183          * largest of these
4184          */
4185         if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4186                 reshape_sectors = mddev->new_chunk_sectors;
4187         else
4188                 reshape_sectors = mddev->chunk_sectors;
4189
4190         /* we update the metadata when there is more than 3Meg
4191          * in the block range (that is rather arbitrary, should
4192          * probably be time based) or when the data about to be
4193          * copied would over-write the source of the data at
4194          * the front of the range.
4195          * i.e. one new_stripe along from reshape_progress new_maps
4196          * to after where reshape_safe old_maps to
4197          */
4198         writepos = conf->reshape_progress;
4199         sector_div(writepos, new_data_disks);
4200         readpos = conf->reshape_progress;
4201         sector_div(readpos, data_disks);
4202         safepos = conf->reshape_safe;
4203         sector_div(safepos, data_disks);
4204         if (mddev->delta_disks < 0) {
4205                 writepos -= min_t(sector_t, reshape_sectors, writepos);
4206                 readpos += reshape_sectors;
4207                 safepos += reshape_sectors;
4208         } else {
4209                 writepos += reshape_sectors;
4210                 readpos -= min_t(sector_t, reshape_sectors, readpos);
4211                 safepos -= min_t(sector_t, reshape_sectors, safepos);
4212         }
4213
4214         /* 'writepos' is the most advanced device address we might write.
4215          * 'readpos' is the least advanced device address we might read.
4216          * 'safepos' is the least address recorded in the metadata as having
4217          *     been reshaped.
4218          * If 'readpos' is behind 'writepos', then there is no way that we can
4219          * ensure safety in the face of a crash - that must be done by userspace
4220          * making a backup of the data.  So in that case there is no particular
4221          * rush to update metadata.
4222          * Otherwise if 'safepos' is behind 'writepos', then we really need to
4223          * update the metadata to advance 'safepos' to match 'readpos' so that
4224          * we can be safe in the event of a crash.
4225          * So we insist on updating metadata if safepos is behind writepos and
4226          * readpos is beyond writepos.
4227          * In any case, update the metadata every 10 seconds.
4228          * Maybe that number should be configurable, but I'm not sure it is
4229          * worth it.... maybe it could be a multiple of safemode_delay???
4230          */
4231         if ((mddev->delta_disks < 0
4232              ? (safepos > writepos && readpos < writepos)
4233              : (safepos < writepos && readpos > writepos)) ||
4234             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4235                 /* Cannot proceed until we've updated the superblock... */
4236                 wait_event(conf->wait_for_overlap,
4237                            atomic_read(&conf->reshape_stripes)==0);
4238                 mddev->reshape_position = conf->reshape_progress;
4239                 mddev->curr_resync_completed = sector_nr;
4240                 conf->reshape_checkpoint = jiffies;
4241                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4242                 md_wakeup_thread(mddev->thread);
4243                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4244                            kthread_should_stop());
4245                 spin_lock_irq(&conf->device_lock);
4246                 conf->reshape_safe = mddev->reshape_position;
4247                 spin_unlock_irq(&conf->device_lock);
4248                 wake_up(&conf->wait_for_overlap);
4249                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4250         }
4251
4252         if (mddev->delta_disks < 0) {
4253                 BUG_ON(conf->reshape_progress == 0);
4254                 stripe_addr = writepos;
4255                 BUG_ON((mddev->dev_sectors &
4256                         ~((sector_t)reshape_sectors - 1))
4257                        - reshape_sectors - stripe_addr
4258                        != sector_nr);
4259         } else {
4260                 BUG_ON(writepos != sector_nr + reshape_sectors);
4261                 stripe_addr = sector_nr;
4262         }
4263         INIT_LIST_HEAD(&stripes);
4264         for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4265                 int j;
4266                 int skipped_disk = 0;
4267                 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4268                 set_bit(STRIPE_EXPANDING, &sh->state);
4269                 atomic_inc(&conf->reshape_stripes);
4270                 /* If any of this stripe is beyond the end of the old
4271                  * array, then we need to zero those blocks
4272                  */
4273                 for (j=sh->disks; j--;) {
4274                         sector_t s;
4275                         if (j == sh->pd_idx)
4276                                 continue;
4277                         if (conf->level == 6 &&
4278                             j == sh->qd_idx)
4279                                 continue;
4280                         s = compute_blocknr(sh, j, 0);
4281                         if (s < raid5_size(mddev, 0, 0)) {
4282                                 skipped_disk = 1;
4283                                 continue;
4284                         }
4285                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4286                         set_bit(R5_Expanded, &sh->dev[j].flags);
4287                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
4288                 }
4289                 if (!skipped_disk) {
4290                         set_bit(STRIPE_EXPAND_READY, &sh->state);
4291                         set_bit(STRIPE_HANDLE, &sh->state);
4292                 }
4293                 list_add(&sh->lru, &stripes);
4294         }
4295         spin_lock_irq(&conf->device_lock);
4296         if (mddev->delta_disks < 0)
4297                 conf->reshape_progress -= reshape_sectors * new_data_disks;
4298         else
4299                 conf->reshape_progress += reshape_sectors * new_data_disks;
4300         spin_unlock_irq(&conf->device_lock);
4301         /* Ok, those stripe are ready. We can start scheduling
4302          * reads on the source stripes.
4303          * The source stripes are determined by mapping the first and last
4304          * block on the destination stripes.
4305          */
4306         first_sector =
4307                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4308                                      1, &dd_idx, NULL);
4309         last_sector =
4310                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4311                                             * new_data_disks - 1),
4312                                      1, &dd_idx, NULL);
4313         if (last_sector >= mddev->dev_sectors)
4314                 last_sector = mddev->dev_sectors - 1;
4315         while (first_sector <= last_sector) {
4316                 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4317                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4318                 set_bit(STRIPE_HANDLE, &sh->state);
4319                 release_stripe(sh);
4320                 first_sector += STRIPE_SECTORS;
4321         }
4322         /* Now that the sources are clearly marked, we can release
4323          * the destination stripes
4324          */
4325         while (!list_empty(&stripes)) {
4326                 sh = list_entry(stripes.next, struct stripe_head, lru);
4327                 list_del_init(&sh->lru);
4328                 release_stripe(sh);
4329         }
4330         /* If this takes us to the resync_max point where we have to pause,
4331          * then we need to write out the superblock.
4332          */
4333         sector_nr += reshape_sectors;
4334         if ((sector_nr - mddev->curr_resync_completed) * 2
4335             >= mddev->resync_max - mddev->curr_resync_completed) {
4336                 /* Cannot proceed until we've updated the superblock... */
4337                 wait_event(conf->wait_for_overlap,
4338                            atomic_read(&conf->reshape_stripes) == 0);
4339                 mddev->reshape_position = conf->reshape_progress;
4340                 mddev->curr_resync_completed = sector_nr;
4341                 conf->reshape_checkpoint = jiffies;
4342                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4343                 md_wakeup_thread(mddev->thread);
4344                 wait_event(mddev->sb_wait,
4345                            !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4346                            || kthread_should_stop());
4347                 spin_lock_irq(&conf->device_lock);
4348                 conf->reshape_safe = mddev->reshape_position;
4349                 spin_unlock_irq(&conf->device_lock);
4350                 wake_up(&conf->wait_for_overlap);
4351                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4352         }
4353         return reshape_sectors;
4354 }
4355
4356 /* FIXME go_faster isn't used */
4357 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
4358 {
4359         raid5_conf_t *conf = mddev->private;
4360         struct stripe_head *sh;
4361         sector_t max_sector = mddev->dev_sectors;
4362         sector_t sync_blocks;
4363         int still_degraded = 0;
4364         int i;
4365
4366         if (sector_nr >= max_sector) {
4367                 /* just being told to finish up .. nothing much to do */
4368                 unplug_slaves(mddev);
4369
4370                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4371                         end_reshape(conf);
4372                         return 0;
4373                 }
4374
4375                 if (mddev->curr_resync < max_sector) /* aborted */
4376                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4377                                         &sync_blocks, 1);
4378                 else /* completed sync */
4379                         conf->fullsync = 0;
4380                 bitmap_close_sync(mddev->bitmap);
4381
4382                 return 0;
4383         }
4384
4385         /* Allow raid5_quiesce to complete */
4386         wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4387
4388         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4389                 return reshape_request(mddev, sector_nr, skipped);
4390
4391         /* No need to check resync_max as we never do more than one
4392          * stripe, and as resync_max will always be on a chunk boundary,
4393          * if the check in md_do_sync didn't fire, there is no chance
4394          * of overstepping resync_max here
4395          */
4396
4397         /* if there is too many failed drives and we are trying
4398          * to resync, then assert that we are finished, because there is
4399          * nothing we can do.
4400          */
4401         if (mddev->degraded >= conf->max_degraded &&
4402             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4403                 sector_t rv = mddev->dev_sectors - sector_nr;
4404                 *skipped = 1;
4405                 return rv;
4406         }
4407         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4408             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4409             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4410                 /* we can skip this block, and probably more */
4411                 sync_blocks /= STRIPE_SECTORS;
4412                 *skipped = 1;
4413                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4414         }
4415
4416
4417         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4418
4419         sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4420         if (sh == NULL) {
4421                 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4422                 /* make sure we don't swamp the stripe cache if someone else
4423                  * is trying to get access
4424                  */
4425                 schedule_timeout_uninterruptible(1);
4426         }
4427         /* Need to check if array will still be degraded after recovery/resync
4428          * We don't need to check the 'failed' flag as when that gets set,
4429          * recovery aborts.
4430          */
4431         for (i = 0; i < conf->raid_disks; i++)
4432                 if (conf->disks[i].rdev == NULL)
4433                         still_degraded = 1;
4434
4435         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4436
4437         spin_lock(&sh->lock);
4438         set_bit(STRIPE_SYNCING, &sh->state);
4439         clear_bit(STRIPE_INSYNC, &sh->state);
4440         spin_unlock(&sh->lock);
4441
4442         handle_stripe(sh);
4443         release_stripe(sh);
4444
4445         return STRIPE_SECTORS;
4446 }
4447
4448 static int  retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
4449 {
4450         /* We may not be able to submit a whole bio at once as there
4451          * may not be enough stripe_heads available.
4452          * We cannot pre-allocate enough stripe_heads as we may need
4453          * more than exist in the cache (if we allow ever large chunks).
4454          * So we do one stripe head at a time and record in
4455          * ->bi_hw_segments how many have been done.
4456          *
4457          * We *know* that this entire raid_bio is in one chunk, so
4458          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4459          */
4460         struct stripe_head *sh;
4461         int dd_idx;
4462         sector_t sector, logical_sector, last_sector;
4463         int scnt = 0;
4464         int remaining;
4465         int handled = 0;
4466
4467         logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4468         sector = raid5_compute_sector(conf, logical_sector,
4469                                       0, &dd_idx, NULL);
4470         last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4471
4472         for (; logical_sector < last_sector;
4473              logical_sector += STRIPE_SECTORS,
4474                      sector += STRIPE_SECTORS,
4475                      scnt++) {
4476
4477                 if (scnt < raid5_bi_hw_segments(raid_bio))
4478                         /* already done this stripe */
4479                         continue;
4480
4481                 sh = get_active_stripe(conf, sector, 0, 1, 0);
4482
4483                 if (!sh) {
4484                         /* failed to get a stripe - must wait */
4485                         raid5_set_bi_hw_segments(raid_bio, scnt);
4486                         conf->retry_read_aligned = raid_bio;
4487                         return handled;
4488                 }
4489
4490                 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4491                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4492                         release_stripe(sh);
4493                         raid5_set_bi_hw_segments(raid_bio, scnt);
4494                         conf->retry_read_aligned = raid_bio;
4495                         return handled;
4496                 }
4497
4498                 handle_stripe(sh);
4499                 release_stripe(sh);
4500                 handled++;
4501         }
4502         spin_lock_irq(&conf->device_lock);
4503         remaining = raid5_dec_bi_phys_segments(raid_bio);
4504         spin_unlock_irq(&conf->device_lock);
4505         if (remaining == 0)
4506                 bio_endio(raid_bio, 0);
4507         if (atomic_dec_and_test(&conf->active_aligned_reads))
4508                 wake_up(&conf->wait_for_stripe);
4509         return handled;
4510 }
4511
4512
4513 /*
4514  * This is our raid5 kernel thread.
4515  *
4516  * We scan the hash table for stripes which can be handled now.
4517  * During the scan, completed stripes are saved for us by the interrupt
4518  * handler, so that they will not have to wait for our next wakeup.
4519  */
4520 static void raid5d(mddev_t *mddev)
4521 {
4522         struct stripe_head *sh;
4523         raid5_conf_t *conf = mddev->private;
4524         int handled;
4525
4526         pr_debug("+++ raid5d active\n");
4527
4528         md_check_recovery(mddev);
4529
4530         handled = 0;
4531         spin_lock_irq(&conf->device_lock);
4532         while (1) {
4533                 struct bio *bio;
4534
4535                 if (conf->seq_flush != conf->seq_write) {
4536                         int seq = conf->seq_flush;
4537                         spin_unlock_irq(&conf->device_lock);
4538                         bitmap_unplug(mddev->bitmap);
4539                         spin_lock_irq(&conf->device_lock);
4540                         conf->seq_write = seq;
4541                         activate_bit_delay(conf);
4542                 }
4543
4544                 while ((bio = remove_bio_from_retry(conf))) {
4545                         int ok;
4546                         spin_unlock_irq(&conf->device_lock);
4547                         ok = retry_aligned_read(conf, bio);
4548                         spin_lock_irq(&conf->device_lock);
4549                         if (!ok)
4550                                 break;
4551                         handled++;
4552                 }
4553
4554                 sh = __get_priority_stripe(conf);
4555
4556                 if (!sh)
4557                         break;
4558                 spin_unlock_irq(&conf->device_lock);
4559                 
4560                 handled++;
4561                 handle_stripe(sh);
4562                 release_stripe(sh);
4563                 cond_resched();
4564
4565                 spin_lock_irq(&conf->device_lock);
4566         }
4567         pr_debug("%d stripes handled\n", handled);
4568
4569         spin_unlock_irq(&conf->device_lock);
4570
4571         async_tx_issue_pending_all();
4572         unplug_slaves(mddev);
4573
4574         pr_debug("--- raid5d inactive\n");
4575 }
4576
4577 static ssize_t
4578 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4579 {
4580         raid5_conf_t *conf = mddev->private;
4581         if (conf)
4582                 return sprintf(page, "%d\n", conf->max_nr_stripes);
4583         else
4584                 return 0;
4585 }
4586
4587 int
4588 raid5_set_cache_size(mddev_t *mddev, int size)
4589 {
4590         raid5_conf_t *conf = mddev->private;
4591         int err;
4592
4593         if (size <= 16 || size > 32768)
4594                 return -EINVAL;
4595         while (size < conf->max_nr_stripes) {
4596                 if (drop_one_stripe(conf))
4597                         conf->max_nr_stripes--;
4598                 else
4599                         break;
4600         }
4601         err = md_allow_write(mddev);
4602         if (err)
4603                 return err;
4604         while (size > conf->max_nr_stripes) {
4605                 if (grow_one_stripe(conf))
4606                         conf->max_nr_stripes++;
4607                 else break;
4608         }
4609         return 0;
4610 }
4611 EXPORT_SYMBOL(raid5_set_cache_size);
4612
4613 static ssize_t
4614 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4615 {
4616         raid5_conf_t *conf = mddev->private;
4617         unsigned long new;
4618         int err;
4619
4620         if (len >= PAGE_SIZE)
4621                 return -EINVAL;
4622         if (!conf)
4623                 return -ENODEV;
4624
4625         if (strict_strtoul(page, 10, &new))
4626                 return -EINVAL;
4627         err = raid5_set_cache_size(mddev, new);
4628         if (err)
4629                 return err;
4630         return len;
4631 }
4632
4633 static struct md_sysfs_entry
4634 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4635                                 raid5_show_stripe_cache_size,
4636                                 raid5_store_stripe_cache_size);
4637
4638 static ssize_t
4639 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4640 {
4641         raid5_conf_t *conf = mddev->private;
4642         if (conf)
4643                 return sprintf(page, "%d\n", conf->bypass_threshold);
4644         else
4645                 return 0;
4646 }
4647
4648 static ssize_t
4649 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4650 {
4651         raid5_conf_t *conf = mddev->private;
4652         unsigned long new;
4653         if (len >= PAGE_SIZE)
4654                 return -EINVAL;
4655         if (!conf)
4656                 return -ENODEV;
4657
4658         if (strict_strtoul(page, 10, &new))
4659                 return -EINVAL;
4660         if (new > conf->max_nr_stripes)
4661                 return -EINVAL;
4662         conf->bypass_threshold = new;
4663         return len;
4664 }
4665
4666 static struct md_sysfs_entry
4667 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4668                                         S_IRUGO | S_IWUSR,
4669                                         raid5_show_preread_threshold,
4670                                         raid5_store_preread_threshold);
4671
4672 static ssize_t
4673 stripe_cache_active_show(mddev_t *mddev, char *page)
4674 {
4675         raid5_conf_t *conf = mddev->private;
4676         if (conf)
4677                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4678         else
4679                 return 0;
4680 }
4681
4682 static struct md_sysfs_entry
4683 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4684
4685 static struct attribute *raid5_attrs[] =  {
4686         &raid5_stripecache_size.attr,
4687         &raid5_stripecache_active.attr,
4688         &raid5_preread_bypass_threshold.attr,
4689         NULL,
4690 };
4691 static struct attribute_group raid5_attrs_group = {
4692         .name = NULL,
4693         .attrs = raid5_attrs,
4694 };
4695
4696 static sector_t
4697 raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks)
4698 {
4699         raid5_conf_t *conf = mddev->private;
4700
4701         if (!sectors)
4702                 sectors = mddev->dev_sectors;
4703         if (!raid_disks)
4704                 /* size is defined by the smallest of previous and new size */
4705                 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4706
4707         sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4708         sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4709         return sectors * (raid_disks - conf->max_degraded);
4710 }
4711
4712 static void raid5_free_percpu(raid5_conf_t *conf)
4713 {
4714         struct raid5_percpu *percpu;
4715         unsigned long cpu;
4716
4717         if (!conf->percpu)
4718                 return;
4719
4720         get_online_cpus();
4721         for_each_possible_cpu(cpu) {
4722                 percpu = per_cpu_ptr(conf->percpu, cpu);
4723                 safe_put_page(percpu->spare_page);
4724                 kfree(percpu->scribble);
4725         }
4726 #ifdef CONFIG_HOTPLUG_CPU
4727         unregister_cpu_notifier(&conf->cpu_notify);
4728 #endif
4729         put_online_cpus();
4730
4731         free_percpu(conf->percpu);
4732 }
4733
4734 static void free_conf(raid5_conf_t *conf)
4735 {
4736         shrink_stripes(conf);
4737         raid5_free_percpu(conf);
4738         kfree(conf->disks);
4739         kfree(conf->stripe_hashtbl);
4740         kfree(conf);
4741 }
4742
4743 #ifdef CONFIG_HOTPLUG_CPU
4744 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4745                               void *hcpu)
4746 {
4747         raid5_conf_t *conf = container_of(nfb, raid5_conf_t, cpu_notify);
4748         long cpu = (long)hcpu;
4749         struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4750
4751         switch (action) {
4752         case CPU_UP_PREPARE:
4753         case CPU_UP_PREPARE_FROZEN:
4754                 if (conf->level == 6 && !percpu->spare_page)
4755                         percpu->spare_page = alloc_page(GFP_KERNEL);
4756                 if (!percpu->scribble)
4757                         percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4758
4759                 if (!percpu->scribble ||
4760                     (conf->level == 6 && !percpu->spare_page)) {
4761                         safe_put_page(percpu->spare_page);
4762                         kfree(percpu->scribble);
4763                         pr_err("%s: failed memory allocation for cpu%ld\n",
4764                                __func__, cpu);
4765                         return notifier_from_errno(-ENOMEM);
4766                 }
4767                 break;
4768         case CPU_DEAD:
4769         case CPU_DEAD_FROZEN:
4770                 safe_put_page(percpu->spare_page);
4771                 kfree(percpu->scribble);
4772                 percpu->spare_page = NULL;
4773                 percpu->scribble = NULL;
4774                 break;
4775         default:
4776                 break;
4777         }
4778         return NOTIFY_OK;
4779 }
4780 #endif
4781
4782 static int raid5_alloc_percpu(raid5_conf_t *conf)
4783 {
4784         unsigned long cpu;
4785         struct page *spare_page;
4786         struct raid5_percpu __percpu *allcpus;
4787         void *scribble;
4788         int err;
4789
4790         allcpus = alloc_percpu(struct raid5_percpu);
4791         if (!allcpus)
4792                 return -ENOMEM;
4793         conf->percpu = allcpus;
4794
4795         get_online_cpus();
4796         err = 0;
4797         for_each_present_cpu(cpu) {
4798                 if (conf->level == 6) {
4799                         spare_page = alloc_page(GFP_KERNEL);
4800                         if (!spare_page) {
4801                                 err = -ENOMEM;
4802                                 break;
4803                         }
4804                         per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4805                 }
4806                 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4807                 if (!scribble) {
4808                         err = -ENOMEM;
4809                         break;
4810                 }
4811                 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4812         }
4813 #ifdef CONFIG_HOTPLUG_CPU
4814         conf->cpu_notify.notifier_call = raid456_cpu_notify;
4815         conf->cpu_notify.priority = 0;
4816         if (err == 0)
4817                 err = register_cpu_notifier(&conf->cpu_notify);
4818 #endif
4819         put_online_cpus();
4820
4821         return err;
4822 }
4823
4824 static raid5_conf_t *setup_conf(mddev_t *mddev)
4825 {
4826         raid5_conf_t *conf;
4827         int raid_disk, memory, max_disks;
4828         mdk_rdev_t *rdev;
4829         struct disk_info *disk;
4830
4831         if (mddev->new_level != 5
4832             && mddev->new_level != 4
4833             && mddev->new_level != 6) {
4834                 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4835                        mdname(mddev), mddev->new_level);
4836                 return ERR_PTR(-EIO);
4837         }
4838         if ((mddev->new_level == 5
4839              && !algorithm_valid_raid5(mddev->new_layout)) ||
4840             (mddev->new_level == 6
4841              && !algorithm_valid_raid6(mddev->new_layout))) {
4842                 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
4843                        mdname(mddev), mddev->new_layout);
4844                 return ERR_PTR(-EIO);
4845         }
4846         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4847                 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4848                        mdname(mddev), mddev->raid_disks);
4849                 return ERR_PTR(-EINVAL);
4850         }
4851
4852         if (!mddev->new_chunk_sectors ||
4853             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4854             !is_power_of_2(mddev->new_chunk_sectors)) {
4855                 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
4856                        mdname(mddev), mddev->new_chunk_sectors << 9);
4857                 return ERR_PTR(-EINVAL);
4858         }
4859
4860         conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL);
4861         if (conf == NULL)
4862                 goto abort;
4863         spin_lock_init(&conf->device_lock);
4864         init_waitqueue_head(&conf->wait_for_stripe);
4865         init_waitqueue_head(&conf->wait_for_overlap);
4866         INIT_LIST_HEAD(&conf->handle_list);
4867         INIT_LIST_HEAD(&conf->hold_list);
4868         INIT_LIST_HEAD(&conf->delayed_list);
4869         INIT_LIST_HEAD(&conf->bitmap_list);
4870         INIT_LIST_HEAD(&conf->inactive_list);
4871         atomic_set(&conf->active_stripes, 0);
4872         atomic_set(&conf->preread_active_stripes, 0);
4873         atomic_set(&conf->active_aligned_reads, 0);
4874         conf->bypass_threshold = BYPASS_THRESHOLD;
4875
4876         conf->raid_disks = mddev->raid_disks;
4877         if (mddev->reshape_position == MaxSector)
4878                 conf->previous_raid_disks = mddev->raid_disks;
4879         else
4880                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4881         max_disks = max(conf->raid_disks, conf->previous_raid_disks);
4882         conf->scribble_len = scribble_len(max_disks);
4883
4884         conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
4885                               GFP_KERNEL);
4886         if (!conf->disks)
4887                 goto abort;
4888
4889         conf->mddev = mddev;
4890
4891         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4892                 goto abort;
4893
4894         conf->level = mddev->new_level;
4895         if (raid5_alloc_percpu(conf) != 0)
4896                 goto abort;
4897
4898         pr_debug("raid456: run(%s) called.\n", mdname(mddev));
4899
4900         list_for_each_entry(rdev, &mddev->disks, same_set) {
4901                 raid_disk = rdev->raid_disk;
4902                 if (raid_disk >= max_disks
4903                     || raid_disk < 0)
4904                         continue;
4905                 disk = conf->disks + raid_disk;
4906
4907                 disk->rdev = rdev;
4908
4909                 if (test_bit(In_sync, &rdev->flags)) {
4910                         char b[BDEVNAME_SIZE];
4911                         printk(KERN_INFO "md/raid:%s: device %s operational as raid"
4912                                " disk %d\n",
4913                                mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
4914                 } else
4915                         /* Cannot rely on bitmap to complete recovery */
4916                         conf->fullsync = 1;
4917         }
4918
4919         conf->chunk_sectors = mddev->new_chunk_sectors;
4920         conf->level = mddev->new_level;
4921         if (conf->level == 6)
4922                 conf->max_degraded = 2;
4923         else
4924                 conf->max_degraded = 1;
4925         conf->algorithm = mddev->new_layout;
4926         conf->max_nr_stripes = NR_STRIPES;
4927         conf->reshape_progress = mddev->reshape_position;
4928         if (conf->reshape_progress != MaxSector) {
4929                 conf->prev_chunk_sectors = mddev->chunk_sectors;
4930                 conf->prev_algo = mddev->layout;
4931         }
4932
4933         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4934                  max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4935         if (grow_stripes(conf, conf->max_nr_stripes)) {
4936                 printk(KERN_ERR
4937                        "md/raid:%s: couldn't allocate %dkB for buffers\n",
4938                        mdname(mddev), memory);
4939                 goto abort;
4940         } else
4941                 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
4942                        mdname(mddev), memory);
4943
4944         conf->thread = md_register_thread(raid5d, mddev, NULL);
4945         if (!conf->thread) {
4946                 printk(KERN_ERR
4947                        "md/raid:%s: couldn't allocate thread.\n",
4948                        mdname(mddev));
4949                 goto abort;
4950         }
4951
4952         return conf;
4953
4954  abort:
4955         if (conf) {
4956                 free_conf(conf);
4957                 return ERR_PTR(-EIO);
4958         } else
4959                 return ERR_PTR(-ENOMEM);
4960 }
4961
4962
4963 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
4964 {
4965         switch (algo) {
4966         case ALGORITHM_PARITY_0:
4967                 if (raid_disk < max_degraded)
4968                         return 1;
4969                 break;
4970         case ALGORITHM_PARITY_N:
4971                 if (raid_disk >= raid_disks - max_degraded)
4972                         return 1;
4973                 break;
4974         case ALGORITHM_PARITY_0_6:
4975                 if (raid_disk == 0 || 
4976                     raid_disk == raid_disks - 1)
4977                         return 1;
4978                 break;
4979         case ALGORITHM_LEFT_ASYMMETRIC_6:
4980         case ALGORITHM_RIGHT_ASYMMETRIC_6:
4981         case ALGORITHM_LEFT_SYMMETRIC_6:
4982         case ALGORITHM_RIGHT_SYMMETRIC_6:
4983                 if (raid_disk == raid_disks - 1)
4984                         return 1;
4985         }
4986         return 0;
4987 }
4988
4989 static int run(mddev_t *mddev)
4990 {
4991         raid5_conf_t *conf;
4992         int working_disks = 0;
4993         int dirty_parity_disks = 0;
4994         mdk_rdev_t *rdev;
4995         sector_t reshape_offset = 0;
4996
4997         if (mddev->recovery_cp != MaxSector)
4998                 printk(KERN_NOTICE "md/raid:%s: not clean"
4999                        " -- starting background reconstruction\n",
5000                        mdname(mddev));
5001         if (mddev->reshape_position != MaxSector) {
5002                 /* Check that we can continue the reshape.
5003                  * Currently only disks can change, it must
5004                  * increase, and we must be past the point where
5005                  * a stripe over-writes itself
5006                  */
5007                 sector_t here_new, here_old;
5008                 int old_disks;
5009                 int max_degraded = (mddev->level == 6 ? 2 : 1);
5010
5011                 if (mddev->new_level != mddev->level) {
5012                         printk(KERN_ERR "md/raid:%s: unsupported reshape "
5013                                "required - aborting.\n",
5014                                mdname(mddev));
5015                         return -EINVAL;
5016                 }
5017                 old_disks = mddev->raid_disks - mddev->delta_disks;
5018                 /* reshape_position must be on a new-stripe boundary, and one
5019                  * further up in new geometry must map after here in old
5020                  * geometry.
5021                  */
5022                 here_new = mddev->reshape_position;
5023                 if (sector_div(here_new, mddev->new_chunk_sectors *
5024                                (mddev->raid_disks - max_degraded))) {
5025                         printk(KERN_ERR "md/raid:%s: reshape_position not "
5026                                "on a stripe boundary\n", mdname(mddev));
5027                         return -EINVAL;
5028                 }
5029                 reshape_offset = here_new * mddev->new_chunk_sectors;
5030                 /* here_new is the stripe we will write to */
5031                 here_old = mddev->reshape_position;
5032                 sector_div(here_old, mddev->chunk_sectors *
5033                            (old_disks-max_degraded));
5034                 /* here_old is the first stripe that we might need to read
5035                  * from */
5036                 if (mddev->delta_disks == 0) {
5037                         /* We cannot be sure it is safe to start an in-place
5038                          * reshape.  It is only safe if user-space if monitoring
5039                          * and taking constant backups.
5040                          * mdadm always starts a situation like this in
5041                          * readonly mode so it can take control before
5042                          * allowing any writes.  So just check for that.
5043                          */
5044                         if ((here_new * mddev->new_chunk_sectors != 
5045                              here_old * mddev->chunk_sectors) ||
5046                             mddev->ro == 0) {
5047                                 printk(KERN_ERR "md/raid:%s: in-place reshape must be started"
5048                                        " in read-only mode - aborting\n",
5049                                        mdname(mddev));
5050                                 return -EINVAL;
5051                         }
5052                 } else if (mddev->delta_disks < 0
5053                     ? (here_new * mddev->new_chunk_sectors <=
5054                        here_old * mddev->chunk_sectors)
5055                     : (here_new * mddev->new_chunk_sectors >=
5056                        here_old * mddev->chunk_sectors)) {
5057                         /* Reading from the same stripe as writing to - bad */
5058                         printk(KERN_ERR "md/raid:%s: reshape_position too early for "
5059                                "auto-recovery - aborting.\n",
5060                                mdname(mddev));
5061                         return -EINVAL;
5062                 }
5063                 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
5064                        mdname(mddev));
5065                 /* OK, we should be able to continue; */
5066         } else {
5067                 BUG_ON(mddev->level != mddev->new_level);
5068                 BUG_ON(mddev->layout != mddev->new_layout);
5069                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
5070                 BUG_ON(mddev->delta_disks != 0);
5071         }
5072
5073         if (mddev->private == NULL)
5074                 conf = setup_conf(mddev);
5075         else
5076                 conf = mddev->private;
5077
5078         if (IS_ERR(conf))
5079                 return PTR_ERR(conf);
5080
5081         mddev->thread = conf->thread;
5082         conf->thread = NULL;
5083         mddev->private = conf;
5084
5085         /*
5086          * 0 for a fully functional array, 1 or 2 for a degraded array.
5087          */
5088         list_for_each_entry(rdev, &mddev->disks, same_set) {
5089                 if (rdev->raid_disk < 0)
5090                         continue;
5091                 if (test_bit(In_sync, &rdev->flags)) {
5092                         working_disks++;
5093                         continue;
5094                 }
5095                 /* This disc is not fully in-sync.  However if it
5096                  * just stored parity (beyond the recovery_offset),
5097                  * when we don't need to be concerned about the
5098                  * array being dirty.
5099                  * When reshape goes 'backwards', we never have
5100                  * partially completed devices, so we only need
5101                  * to worry about reshape going forwards.
5102                  */
5103                 /* Hack because v0.91 doesn't store recovery_offset properly. */
5104                 if (mddev->major_version == 0 &&
5105                     mddev->minor_version > 90)
5106                         rdev->recovery_offset = reshape_offset;
5107                         
5108                 if (rdev->recovery_offset < reshape_offset) {
5109                         /* We need to check old and new layout */
5110                         if (!only_parity(rdev->raid_disk,
5111                                          conf->algorithm,
5112                                          conf->raid_disks,
5113                                          conf->max_degraded))
5114                                 continue;
5115                 }
5116                 if (!only_parity(rdev->raid_disk,
5117                                  conf->prev_algo,
5118                                  conf->previous_raid_disks,
5119                                  conf->max_degraded))
5120                         continue;
5121                 dirty_parity_disks++;
5122         }
5123
5124         mddev->degraded = (max(conf->raid_disks, conf->previous_raid_disks)
5125                            - working_disks);
5126
5127         if (has_failed(conf)) {
5128                 printk(KERN_ERR "md/raid:%s: not enough operational devices"
5129                         " (%d/%d failed)\n",
5130                         mdname(mddev), mddev->degraded, conf->raid_disks);
5131                 goto abort;
5132         }
5133
5134         /* device size must be a multiple of chunk size */
5135         mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
5136         mddev->resync_max_sectors = mddev->dev_sectors;
5137
5138         if (mddev->degraded > dirty_parity_disks &&
5139             mddev->recovery_cp != MaxSector) {
5140                 if (mddev->ok_start_degraded)
5141                         printk(KERN_WARNING
5142                                "md/raid:%s: starting dirty degraded array"
5143                                " - data corruption possible.\n",
5144                                mdname(mddev));
5145                 else {
5146                         printk(KERN_ERR
5147                                "md/raid:%s: cannot start dirty degraded array.\n",
5148                                mdname(mddev));
5149                         goto abort;
5150                 }
5151         }
5152
5153         if (mddev->degraded == 0)
5154                 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
5155                        " devices, algorithm %d\n", mdname(mddev), conf->level,
5156                        mddev->raid_disks-mddev->degraded, mddev->raid_disks,
5157                        mddev->new_layout);
5158         else
5159                 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
5160                        " out of %d devices, algorithm %d\n",
5161                        mdname(mddev), conf->level,
5162                        mddev->raid_disks - mddev->degraded,
5163                        mddev->raid_disks, mddev->new_layout);
5164
5165         print_raid5_conf(conf);
5166
5167         if (conf->reshape_progress != MaxSector) {
5168                 conf->reshape_safe = conf->reshape_progress;
5169                 atomic_set(&conf->reshape_stripes, 0);
5170                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5171                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5172                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5173                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5174                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5175                                                         "reshape");
5176         }
5177
5178
5179         /* Ok, everything is just fine now */
5180         if (mddev->to_remove == &raid5_attrs_group)
5181                 mddev->to_remove = NULL;
5182         else if (mddev->kobj.sd &&
5183             sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5184                 printk(KERN_WARNING
5185                        "raid5: failed to create sysfs attributes for %s\n",
5186                        mdname(mddev));
5187         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5188
5189         plugger_init(&conf->plug, raid5_unplug);
5190         mddev->plug = &conf->plug;
5191         if (mddev->queue) {
5192                 int chunk_size;
5193                 /* read-ahead size must cover two whole stripes, which
5194                  * is 2 * (datadisks) * chunksize where 'n' is the
5195                  * number of raid devices
5196                  */
5197                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
5198                 int stripe = data_disks *
5199                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
5200                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5201                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5202
5203                 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
5204
5205                 mddev->queue->backing_dev_info.congested_data = mddev;
5206                 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
5207                 mddev->queue->queue_lock = &conf->device_lock;
5208                 mddev->queue->unplug_fn = raid5_unplug_queue;
5209
5210                 chunk_size = mddev->chunk_sectors << 9;
5211                 blk_queue_io_min(mddev->queue, chunk_size);
5212                 blk_queue_io_opt(mddev->queue, chunk_size *
5213                                  (conf->raid_disks - conf->max_degraded));
5214
5215                 list_for_each_entry(rdev, &mddev->disks, same_set)
5216                         disk_stack_limits(mddev->gendisk, rdev->bdev,
5217                                           rdev->data_offset << 9);
5218         }
5219
5220         return 0;
5221 abort:
5222         md_unregister_thread(mddev->thread);
5223         mddev->thread = NULL;
5224         if (conf) {
5225                 print_raid5_conf(conf);
5226                 free_conf(conf);
5227         }
5228         mddev->private = NULL;
5229         printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
5230         return -EIO;
5231 }
5232
5233 static int stop(mddev_t *mddev)
5234 {
5235         raid5_conf_t *conf = mddev->private;
5236
5237         md_unregister_thread(mddev->thread);
5238         mddev->thread = NULL;
5239         if (mddev->queue)
5240                 mddev->queue->backing_dev_info.congested_fn = NULL;
5241         plugger_flush(&conf->plug); /* the unplug fn references 'conf'*/
5242         free_conf(conf);
5243         mddev->private = NULL;
5244         mddev->to_remove = &raid5_attrs_group;
5245         return 0;
5246 }
5247
5248 #ifdef DEBUG
5249 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
5250 {
5251         int i;
5252
5253         seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
5254                    (unsigned long long)sh->sector, sh->pd_idx, sh->state);
5255         seq_printf(seq, "sh %llu,  count %d.\n",
5256                    (unsigned long long)sh->sector, atomic_read(&sh->count));
5257         seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
5258         for (i = 0; i < sh->disks; i++) {
5259                 seq_printf(seq, "(cache%d: %p %ld) ",
5260                            i, sh->dev[i].page, sh->dev[i].flags);
5261         }
5262         seq_printf(seq, "\n");
5263 }
5264
5265 static void printall(struct seq_file *seq, raid5_conf_t *conf)
5266 {
5267         struct stripe_head *sh;
5268         struct hlist_node *hn;
5269         int i;
5270
5271         spin_lock_irq(&conf->device_lock);
5272         for (i = 0; i < NR_HASH; i++) {
5273                 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
5274                         if (sh->raid_conf != conf)
5275                                 continue;
5276                         print_sh(seq, sh);
5277                 }
5278         }
5279         spin_unlock_irq(&conf->device_lock);
5280 }
5281 #endif
5282
5283 static void status(struct seq_file *seq, mddev_t *mddev)
5284 {
5285         raid5_conf_t *conf = mddev->private;
5286         int i;
5287
5288         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5289                 mddev->chunk_sectors / 2, mddev->layout);
5290         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5291         for (i = 0; i < conf->raid_disks; i++)
5292                 seq_printf (seq, "%s",
5293                                conf->disks[i].rdev &&
5294                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5295         seq_printf (seq, "]");
5296 #ifdef DEBUG
5297         seq_printf (seq, "\n");
5298         printall(seq, conf);
5299 #endif
5300 }
5301
5302 static void print_raid5_conf (raid5_conf_t *conf)
5303 {
5304         int i;
5305         struct disk_info *tmp;
5306
5307         printk(KERN_DEBUG "RAID conf printout:\n");
5308         if (!conf) {
5309                 printk("(conf==NULL)\n");
5310                 return;
5311         }
5312         printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5313                conf->raid_disks,
5314                conf->raid_disks - conf->mddev->degraded);
5315
5316         for (i = 0; i < conf->raid_disks; i++) {
5317                 char b[BDEVNAME_SIZE];
5318                 tmp = conf->disks + i;
5319                 if (tmp->rdev)
5320                         printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5321                                i, !test_bit(Faulty, &tmp->rdev->flags),
5322                                bdevname(tmp->rdev->bdev, b));
5323         }
5324 }
5325
5326 static int raid5_spare_active(mddev_t *mddev)
5327 {
5328         int i;
5329         raid5_conf_t *conf = mddev->private;
5330         struct disk_info *tmp;
5331         int count = 0;
5332         unsigned long flags;
5333
5334         for (i = 0; i < conf->raid_disks; i++) {
5335                 tmp = conf->disks + i;
5336                 if (tmp->rdev
5337                     && tmp->rdev->recovery_offset == MaxSector
5338                     && !test_bit(Faulty, &tmp->rdev->flags)