md/raid5: allow new reshape modes to be restarted in the middle.
[efikamx:linux-kernel.git] / drivers / md / raid5.c
1 /*
2  * raid5.c : Multiple Devices driver for Linux
3  *         Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4  *         Copyright (C) 1999, 2000 Ingo Molnar
5  *         Copyright (C) 2002, 2003 H. Peter Anvin
6  *
7  * RAID-4/5/6 management functions.
8  * Thanks to Penguin Computing for making the RAID-6 development possible
9  * by donating a test server!
10  *
11  * This program is free software; you can redistribute it and/or modify
12  * it under the terms of the GNU General Public License as published by
13  * the Free Software Foundation; either version 2, or (at your option)
14  * any later version.
15  *
16  * You should have received a copy of the GNU General Public License
17  * (for example /usr/src/linux/COPYING); if not, write to the Free
18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19  */
20
21 /*
22  * BITMAP UNPLUGGING:
23  *
24  * The sequencing for updating the bitmap reliably is a little
25  * subtle (and I got it wrong the first time) so it deserves some
26  * explanation.
27  *
28  * We group bitmap updates into batches.  Each batch has a number.
29  * We may write out several batches at once, but that isn't very important.
30  * conf->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/seq_file.h>
51 #include "md.h"
52 #include "raid5.h"
53 #include "bitmap.h"
54
55 /*
56  * Stripe cache
57  */
58
59 #define NR_STRIPES              256
60 #define STRIPE_SIZE             PAGE_SIZE
61 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
62 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
63 #define IO_THRESHOLD            1
64 #define BYPASS_THRESHOLD        1
65 #define NR_HASH                 (PAGE_SIZE / sizeof(struct hlist_head))
66 #define HASH_MASK               (NR_HASH - 1)
67
68 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
69
70 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
71  * order without overlap.  There may be several bio's per stripe+device, and
72  * a bio could span several devices.
73  * When walking this list for a particular stripe+device, we must never proceed
74  * beyond a bio that extends past this device, as the next bio might no longer
75  * be valid.
76  * This macro is used to determine the 'next' bio in the list, given the sector
77  * of the current stripe+device
78  */
79 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
80 /*
81  * The following can be used to debug the driver
82  */
83 #define RAID5_PARANOIA  1
84 #if RAID5_PARANOIA && defined(CONFIG_SMP)
85 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
86 #else
87 # define CHECK_DEVLOCK()
88 #endif
89
90 #ifdef DEBUG
91 #define inline
92 #define __inline__
93 #endif
94
95 #define printk_rl(args...) ((void) (printk_ratelimit() && printk(args)))
96
97 /*
98  * We maintain a biased count of active stripes in the bottom 16 bits of
99  * bi_phys_segments, and a count of processed stripes in the upper 16 bits
100  */
101 static inline int raid5_bi_phys_segments(struct bio *bio)
102 {
103         return bio->bi_phys_segments & 0xffff;
104 }
105
106 static inline int raid5_bi_hw_segments(struct bio *bio)
107 {
108         return (bio->bi_phys_segments >> 16) & 0xffff;
109 }
110
111 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
112 {
113         --bio->bi_phys_segments;
114         return raid5_bi_phys_segments(bio);
115 }
116
117 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
118 {
119         unsigned short val = raid5_bi_hw_segments(bio);
120
121         --val;
122         bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
123         return val;
124 }
125
126 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
127 {
128         bio->bi_phys_segments = raid5_bi_phys_segments(bio) || (cnt << 16);
129 }
130
131 /* Find first data disk in a raid6 stripe */
132 static inline int raid6_d0(struct stripe_head *sh)
133 {
134         if (sh->ddf_layout)
135                 /* ddf always start from first device */
136                 return 0;
137         /* md starts just after Q block */
138         if (sh->qd_idx == sh->disks - 1)
139                 return 0;
140         else
141                 return sh->qd_idx + 1;
142 }
143 static inline int raid6_next_disk(int disk, int raid_disks)
144 {
145         disk++;
146         return (disk < raid_disks) ? disk : 0;
147 }
148
149 /* When walking through the disks in a raid5, starting at raid6_d0,
150  * We need to map each disk to a 'slot', where the data disks are slot
151  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
152  * is raid_disks-1.  This help does that mapping.
153  */
154 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
155                              int *count, int syndrome_disks)
156 {
157         int slot;
158
159         if (idx == sh->pd_idx)
160                 return syndrome_disks;
161         if (idx == sh->qd_idx)
162                 return syndrome_disks + 1;
163         slot = (*count)++;
164         return slot;
165 }
166
167 static void return_io(struct bio *return_bi)
168 {
169         struct bio *bi = return_bi;
170         while (bi) {
171
172                 return_bi = bi->bi_next;
173                 bi->bi_next = NULL;
174                 bi->bi_size = 0;
175                 bio_endio(bi, 0);
176                 bi = return_bi;
177         }
178 }
179
180 static void print_raid5_conf (raid5_conf_t *conf);
181
182 static int stripe_operations_active(struct stripe_head *sh)
183 {
184         return sh->check_state || sh->reconstruct_state ||
185                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
186                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
187 }
188
189 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
190 {
191         if (atomic_dec_and_test(&sh->count)) {
192                 BUG_ON(!list_empty(&sh->lru));
193                 BUG_ON(atomic_read(&conf->active_stripes)==0);
194                 if (test_bit(STRIPE_HANDLE, &sh->state)) {
195                         if (test_bit(STRIPE_DELAYED, &sh->state)) {
196                                 list_add_tail(&sh->lru, &conf->delayed_list);
197                                 blk_plug_device(conf->mddev->queue);
198                         } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
199                                    sh->bm_seq - conf->seq_write > 0) {
200                                 list_add_tail(&sh->lru, &conf->bitmap_list);
201                                 blk_plug_device(conf->mddev->queue);
202                         } else {
203                                 clear_bit(STRIPE_BIT_DELAY, &sh->state);
204                                 list_add_tail(&sh->lru, &conf->handle_list);
205                         }
206                         md_wakeup_thread(conf->mddev->thread);
207                 } else {
208                         BUG_ON(stripe_operations_active(sh));
209                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
210                                 atomic_dec(&conf->preread_active_stripes);
211                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
212                                         md_wakeup_thread(conf->mddev->thread);
213                         }
214                         atomic_dec(&conf->active_stripes);
215                         if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
216                                 list_add_tail(&sh->lru, &conf->inactive_list);
217                                 wake_up(&conf->wait_for_stripe);
218                                 if (conf->retry_read_aligned)
219                                         md_wakeup_thread(conf->mddev->thread);
220                         }
221                 }
222         }
223 }
224
225 static void release_stripe(struct stripe_head *sh)
226 {
227         raid5_conf_t *conf = sh->raid_conf;
228         unsigned long flags;
229
230         spin_lock_irqsave(&conf->device_lock, flags);
231         __release_stripe(conf, sh);
232         spin_unlock_irqrestore(&conf->device_lock, flags);
233 }
234
235 static inline void remove_hash(struct stripe_head *sh)
236 {
237         pr_debug("remove_hash(), stripe %llu\n",
238                 (unsigned long long)sh->sector);
239
240         hlist_del_init(&sh->hash);
241 }
242
243 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
244 {
245         struct hlist_head *hp = stripe_hash(conf, sh->sector);
246
247         pr_debug("insert_hash(), stripe %llu\n",
248                 (unsigned long long)sh->sector);
249
250         CHECK_DEVLOCK();
251         hlist_add_head(&sh->hash, hp);
252 }
253
254
255 /* find an idle stripe, make sure it is unhashed, and return it. */
256 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
257 {
258         struct stripe_head *sh = NULL;
259         struct list_head *first;
260
261         CHECK_DEVLOCK();
262         if (list_empty(&conf->inactive_list))
263                 goto out;
264         first = conf->inactive_list.next;
265         sh = list_entry(first, struct stripe_head, lru);
266         list_del_init(first);
267         remove_hash(sh);
268         atomic_inc(&conf->active_stripes);
269 out:
270         return sh;
271 }
272
273 static void shrink_buffers(struct stripe_head *sh, int num)
274 {
275         struct page *p;
276         int i;
277
278         for (i=0; i<num ; i++) {
279                 p = sh->dev[i].page;
280                 if (!p)
281                         continue;
282                 sh->dev[i].page = NULL;
283                 put_page(p);
284         }
285 }
286
287 static int grow_buffers(struct stripe_head *sh, int num)
288 {
289         int i;
290
291         for (i=0; i<num; i++) {
292                 struct page *page;
293
294                 if (!(page = alloc_page(GFP_KERNEL))) {
295                         return 1;
296                 }
297                 sh->dev[i].page = page;
298         }
299         return 0;
300 }
301
302 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
303 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
304                             struct stripe_head *sh);
305
306 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
307 {
308         raid5_conf_t *conf = sh->raid_conf;
309         int i;
310
311         BUG_ON(atomic_read(&sh->count) != 0);
312         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
313         BUG_ON(stripe_operations_active(sh));
314
315         CHECK_DEVLOCK();
316         pr_debug("init_stripe called, stripe %llu\n",
317                 (unsigned long long)sh->sector);
318
319         remove_hash(sh);
320
321         sh->generation = conf->generation - previous;
322         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
323         sh->sector = sector;
324         stripe_set_idx(sector, conf, previous, sh);
325         sh->state = 0;
326
327
328         for (i = sh->disks; i--; ) {
329                 struct r5dev *dev = &sh->dev[i];
330
331                 if (dev->toread || dev->read || dev->towrite || dev->written ||
332                     test_bit(R5_LOCKED, &dev->flags)) {
333                         printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
334                                (unsigned long long)sh->sector, i, dev->toread,
335                                dev->read, dev->towrite, dev->written,
336                                test_bit(R5_LOCKED, &dev->flags));
337                         BUG();
338                 }
339                 dev->flags = 0;
340                 raid5_build_block(sh, i, previous);
341         }
342         insert_hash(conf, sh);
343 }
344
345 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector,
346                                          short generation)
347 {
348         struct stripe_head *sh;
349         struct hlist_node *hn;
350
351         CHECK_DEVLOCK();
352         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
353         hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
354                 if (sh->sector == sector && sh->generation == generation)
355                         return sh;
356         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
357         return NULL;
358 }
359
360 static void unplug_slaves(mddev_t *mddev);
361 static void raid5_unplug_device(struct request_queue *q);
362
363 static struct stripe_head *
364 get_active_stripe(raid5_conf_t *conf, sector_t sector,
365                   int previous, int noblock, int noquiesce)
366 {
367         struct stripe_head *sh;
368
369         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
370
371         spin_lock_irq(&conf->device_lock);
372
373         do {
374                 wait_event_lock_irq(conf->wait_for_stripe,
375                                     conf->quiesce == 0 || noquiesce,
376                                     conf->device_lock, /* nothing */);
377                 sh = __find_stripe(conf, sector, conf->generation - previous);
378                 if (!sh) {
379                         if (!conf->inactive_blocked)
380                                 sh = get_free_stripe(conf);
381                         if (noblock && sh == NULL)
382                                 break;
383                         if (!sh) {
384                                 conf->inactive_blocked = 1;
385                                 wait_event_lock_irq(conf->wait_for_stripe,
386                                                     !list_empty(&conf->inactive_list) &&
387                                                     (atomic_read(&conf->active_stripes)
388                                                      < (conf->max_nr_stripes *3/4)
389                                                      || !conf->inactive_blocked),
390                                                     conf->device_lock,
391                                                     raid5_unplug_device(conf->mddev->queue)
392                                         );
393                                 conf->inactive_blocked = 0;
394                         } else
395                                 init_stripe(sh, sector, previous);
396                 } else {
397                         if (atomic_read(&sh->count)) {
398                                 BUG_ON(!list_empty(&sh->lru)
399                                     && !test_bit(STRIPE_EXPANDING, &sh->state));
400                         } else {
401                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
402                                         atomic_inc(&conf->active_stripes);
403                                 if (list_empty(&sh->lru) &&
404                                     !test_bit(STRIPE_EXPANDING, &sh->state))
405                                         BUG();
406                                 list_del_init(&sh->lru);
407                         }
408                 }
409         } while (sh == NULL);
410
411         if (sh)
412                 atomic_inc(&sh->count);
413
414         spin_unlock_irq(&conf->device_lock);
415         return sh;
416 }
417
418 static void
419 raid5_end_read_request(struct bio *bi, int error);
420 static void
421 raid5_end_write_request(struct bio *bi, int error);
422
423 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
424 {
425         raid5_conf_t *conf = sh->raid_conf;
426         int i, disks = sh->disks;
427
428         might_sleep();
429
430         for (i = disks; i--; ) {
431                 int rw;
432                 struct bio *bi;
433                 mdk_rdev_t *rdev;
434                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
435                         rw = WRITE;
436                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
437                         rw = READ;
438                 else
439                         continue;
440
441                 bi = &sh->dev[i].req;
442
443                 bi->bi_rw = rw;
444                 if (rw == WRITE)
445                         bi->bi_end_io = raid5_end_write_request;
446                 else
447                         bi->bi_end_io = raid5_end_read_request;
448
449                 rcu_read_lock();
450                 rdev = rcu_dereference(conf->disks[i].rdev);
451                 if (rdev && test_bit(Faulty, &rdev->flags))
452                         rdev = NULL;
453                 if (rdev)
454                         atomic_inc(&rdev->nr_pending);
455                 rcu_read_unlock();
456
457                 if (rdev) {
458                         if (s->syncing || s->expanding || s->expanded)
459                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
460
461                         set_bit(STRIPE_IO_STARTED, &sh->state);
462
463                         bi->bi_bdev = rdev->bdev;
464                         pr_debug("%s: for %llu schedule op %ld on disc %d\n",
465                                 __func__, (unsigned long long)sh->sector,
466                                 bi->bi_rw, i);
467                         atomic_inc(&sh->count);
468                         bi->bi_sector = sh->sector + rdev->data_offset;
469                         bi->bi_flags = 1 << BIO_UPTODATE;
470                         bi->bi_vcnt = 1;
471                         bi->bi_max_vecs = 1;
472                         bi->bi_idx = 0;
473                         bi->bi_io_vec = &sh->dev[i].vec;
474                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
475                         bi->bi_io_vec[0].bv_offset = 0;
476                         bi->bi_size = STRIPE_SIZE;
477                         bi->bi_next = NULL;
478                         if (rw == WRITE &&
479                             test_bit(R5_ReWrite, &sh->dev[i].flags))
480                                 atomic_add(STRIPE_SECTORS,
481                                         &rdev->corrected_errors);
482                         generic_make_request(bi);
483                 } else {
484                         if (rw == WRITE)
485                                 set_bit(STRIPE_DEGRADED, &sh->state);
486                         pr_debug("skip op %ld on disc %d for sector %llu\n",
487                                 bi->bi_rw, i, (unsigned long long)sh->sector);
488                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
489                         set_bit(STRIPE_HANDLE, &sh->state);
490                 }
491         }
492 }
493
494 static struct dma_async_tx_descriptor *
495 async_copy_data(int frombio, struct bio *bio, struct page *page,
496         sector_t sector, struct dma_async_tx_descriptor *tx)
497 {
498         struct bio_vec *bvl;
499         struct page *bio_page;
500         int i;
501         int page_offset;
502
503         if (bio->bi_sector >= sector)
504                 page_offset = (signed)(bio->bi_sector - sector) * 512;
505         else
506                 page_offset = (signed)(sector - bio->bi_sector) * -512;
507         bio_for_each_segment(bvl, bio, i) {
508                 int len = bio_iovec_idx(bio, i)->bv_len;
509                 int clen;
510                 int b_offset = 0;
511
512                 if (page_offset < 0) {
513                         b_offset = -page_offset;
514                         page_offset += b_offset;
515                         len -= b_offset;
516                 }
517
518                 if (len > 0 && page_offset + len > STRIPE_SIZE)
519                         clen = STRIPE_SIZE - page_offset;
520                 else
521                         clen = len;
522
523                 if (clen > 0) {
524                         b_offset += bio_iovec_idx(bio, i)->bv_offset;
525                         bio_page = bio_iovec_idx(bio, i)->bv_page;
526                         if (frombio)
527                                 tx = async_memcpy(page, bio_page, page_offset,
528                                         b_offset, clen,
529                                         ASYNC_TX_DEP_ACK,
530                                         tx, NULL, NULL);
531                         else
532                                 tx = async_memcpy(bio_page, page, b_offset,
533                                         page_offset, clen,
534                                         ASYNC_TX_DEP_ACK,
535                                         tx, NULL, NULL);
536                 }
537                 if (clen < len) /* hit end of page */
538                         break;
539                 page_offset +=  len;
540         }
541
542         return tx;
543 }
544
545 static void ops_complete_biofill(void *stripe_head_ref)
546 {
547         struct stripe_head *sh = stripe_head_ref;
548         struct bio *return_bi = NULL;
549         raid5_conf_t *conf = sh->raid_conf;
550         int i;
551
552         pr_debug("%s: stripe %llu\n", __func__,
553                 (unsigned long long)sh->sector);
554
555         /* clear completed biofills */
556         spin_lock_irq(&conf->device_lock);
557         for (i = sh->disks; i--; ) {
558                 struct r5dev *dev = &sh->dev[i];
559
560                 /* acknowledge completion of a biofill operation */
561                 /* and check if we need to reply to a read request,
562                  * new R5_Wantfill requests are held off until
563                  * !STRIPE_BIOFILL_RUN
564                  */
565                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
566                         struct bio *rbi, *rbi2;
567
568                         BUG_ON(!dev->read);
569                         rbi = dev->read;
570                         dev->read = NULL;
571                         while (rbi && rbi->bi_sector <
572                                 dev->sector + STRIPE_SECTORS) {
573                                 rbi2 = r5_next_bio(rbi, dev->sector);
574                                 if (!raid5_dec_bi_phys_segments(rbi)) {
575                                         rbi->bi_next = return_bi;
576                                         return_bi = rbi;
577                                 }
578                                 rbi = rbi2;
579                         }
580                 }
581         }
582         spin_unlock_irq(&conf->device_lock);
583         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
584
585         return_io(return_bi);
586
587         set_bit(STRIPE_HANDLE, &sh->state);
588         release_stripe(sh);
589 }
590
591 static void ops_run_biofill(struct stripe_head *sh)
592 {
593         struct dma_async_tx_descriptor *tx = NULL;
594         raid5_conf_t *conf = sh->raid_conf;
595         int i;
596
597         pr_debug("%s: stripe %llu\n", __func__,
598                 (unsigned long long)sh->sector);
599
600         for (i = sh->disks; i--; ) {
601                 struct r5dev *dev = &sh->dev[i];
602                 if (test_bit(R5_Wantfill, &dev->flags)) {
603                         struct bio *rbi;
604                         spin_lock_irq(&conf->device_lock);
605                         dev->read = rbi = dev->toread;
606                         dev->toread = NULL;
607                         spin_unlock_irq(&conf->device_lock);
608                         while (rbi && rbi->bi_sector <
609                                 dev->sector + STRIPE_SECTORS) {
610                                 tx = async_copy_data(0, rbi, dev->page,
611                                         dev->sector, tx);
612                                 rbi = r5_next_bio(rbi, dev->sector);
613                         }
614                 }
615         }
616
617         atomic_inc(&sh->count);
618         async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
619                 ops_complete_biofill, sh);
620 }
621
622 static void ops_complete_compute5(void *stripe_head_ref)
623 {
624         struct stripe_head *sh = stripe_head_ref;
625         int target = sh->ops.target;
626         struct r5dev *tgt = &sh->dev[target];
627
628         pr_debug("%s: stripe %llu\n", __func__,
629                 (unsigned long long)sh->sector);
630
631         set_bit(R5_UPTODATE, &tgt->flags);
632         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
633         clear_bit(R5_Wantcompute, &tgt->flags);
634         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
635         if (sh->check_state == check_state_compute_run)
636                 sh->check_state = check_state_compute_result;
637         set_bit(STRIPE_HANDLE, &sh->state);
638         release_stripe(sh);
639 }
640
641 static struct dma_async_tx_descriptor *ops_run_compute5(struct stripe_head *sh)
642 {
643         /* kernel stack size limits the total number of disks */
644         int disks = sh->disks;
645         struct page *xor_srcs[disks];
646         int target = sh->ops.target;
647         struct r5dev *tgt = &sh->dev[target];
648         struct page *xor_dest = tgt->page;
649         int count = 0;
650         struct dma_async_tx_descriptor *tx;
651         int i;
652
653         pr_debug("%s: stripe %llu block: %d\n",
654                 __func__, (unsigned long long)sh->sector, target);
655         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
656
657         for (i = disks; i--; )
658                 if (i != target)
659                         xor_srcs[count++] = sh->dev[i].page;
660
661         atomic_inc(&sh->count);
662
663         if (unlikely(count == 1))
664                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
665                         0, NULL, ops_complete_compute5, sh);
666         else
667                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
668                         ASYNC_TX_XOR_ZERO_DST, NULL,
669                         ops_complete_compute5, sh);
670
671         return tx;
672 }
673
674 static void ops_complete_prexor(void *stripe_head_ref)
675 {
676         struct stripe_head *sh = stripe_head_ref;
677
678         pr_debug("%s: stripe %llu\n", __func__,
679                 (unsigned long long)sh->sector);
680 }
681
682 static struct dma_async_tx_descriptor *
683 ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
684 {
685         /* kernel stack size limits the total number of disks */
686         int disks = sh->disks;
687         struct page *xor_srcs[disks];
688         int count = 0, pd_idx = sh->pd_idx, i;
689
690         /* existing parity data subtracted */
691         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
692
693         pr_debug("%s: stripe %llu\n", __func__,
694                 (unsigned long long)sh->sector);
695
696         for (i = disks; i--; ) {
697                 struct r5dev *dev = &sh->dev[i];
698                 /* Only process blocks that are known to be uptodate */
699                 if (test_bit(R5_Wantdrain, &dev->flags))
700                         xor_srcs[count++] = dev->page;
701         }
702
703         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
704                 ASYNC_TX_DEP_ACK | ASYNC_TX_XOR_DROP_DST, tx,
705                 ops_complete_prexor, sh);
706
707         return tx;
708 }
709
710 static struct dma_async_tx_descriptor *
711 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
712 {
713         int disks = sh->disks;
714         int i;
715
716         pr_debug("%s: stripe %llu\n", __func__,
717                 (unsigned long long)sh->sector);
718
719         for (i = disks; i--; ) {
720                 struct r5dev *dev = &sh->dev[i];
721                 struct bio *chosen;
722
723                 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
724                         struct bio *wbi;
725
726                         spin_lock(&sh->lock);
727                         chosen = dev->towrite;
728                         dev->towrite = NULL;
729                         BUG_ON(dev->written);
730                         wbi = dev->written = chosen;
731                         spin_unlock(&sh->lock);
732
733                         while (wbi && wbi->bi_sector <
734                                 dev->sector + STRIPE_SECTORS) {
735                                 tx = async_copy_data(1, wbi, dev->page,
736                                         dev->sector, tx);
737                                 wbi = r5_next_bio(wbi, dev->sector);
738                         }
739                 }
740         }
741
742         return tx;
743 }
744
745 static void ops_complete_postxor(void *stripe_head_ref)
746 {
747         struct stripe_head *sh = stripe_head_ref;
748         int disks = sh->disks, i, pd_idx = sh->pd_idx;
749
750         pr_debug("%s: stripe %llu\n", __func__,
751                 (unsigned long long)sh->sector);
752
753         for (i = disks; i--; ) {
754                 struct r5dev *dev = &sh->dev[i];
755                 if (dev->written || i == pd_idx)
756                         set_bit(R5_UPTODATE, &dev->flags);
757         }
758
759         if (sh->reconstruct_state == reconstruct_state_drain_run)
760                 sh->reconstruct_state = reconstruct_state_drain_result;
761         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
762                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
763         else {
764                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
765                 sh->reconstruct_state = reconstruct_state_result;
766         }
767
768         set_bit(STRIPE_HANDLE, &sh->state);
769         release_stripe(sh);
770 }
771
772 static void
773 ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
774 {
775         /* kernel stack size limits the total number of disks */
776         int disks = sh->disks;
777         struct page *xor_srcs[disks];
778
779         int count = 0, pd_idx = sh->pd_idx, i;
780         struct page *xor_dest;
781         int prexor = 0;
782         unsigned long flags;
783
784         pr_debug("%s: stripe %llu\n", __func__,
785                 (unsigned long long)sh->sector);
786
787         /* check if prexor is active which means only process blocks
788          * that are part of a read-modify-write (written)
789          */
790         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
791                 prexor = 1;
792                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
793                 for (i = disks; i--; ) {
794                         struct r5dev *dev = &sh->dev[i];
795                         if (dev->written)
796                                 xor_srcs[count++] = dev->page;
797                 }
798         } else {
799                 xor_dest = sh->dev[pd_idx].page;
800                 for (i = disks; i--; ) {
801                         struct r5dev *dev = &sh->dev[i];
802                         if (i != pd_idx)
803                                 xor_srcs[count++] = dev->page;
804                 }
805         }
806
807         /* 1/ if we prexor'd then the dest is reused as a source
808          * 2/ if we did not prexor then we are redoing the parity
809          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
810          * for the synchronous xor case
811          */
812         flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK |
813                 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
814
815         atomic_inc(&sh->count);
816
817         if (unlikely(count == 1)) {
818                 flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST);
819                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
820                         flags, tx, ops_complete_postxor, sh);
821         } else
822                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
823                         flags, tx, ops_complete_postxor, sh);
824 }
825
826 static void ops_complete_check(void *stripe_head_ref)
827 {
828         struct stripe_head *sh = stripe_head_ref;
829
830         pr_debug("%s: stripe %llu\n", __func__,
831                 (unsigned long long)sh->sector);
832
833         sh->check_state = check_state_check_result;
834         set_bit(STRIPE_HANDLE, &sh->state);
835         release_stripe(sh);
836 }
837
838 static void ops_run_check(struct stripe_head *sh)
839 {
840         /* kernel stack size limits the total number of disks */
841         int disks = sh->disks;
842         struct page *xor_srcs[disks];
843         struct dma_async_tx_descriptor *tx;
844
845         int count = 0, pd_idx = sh->pd_idx, i;
846         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
847
848         pr_debug("%s: stripe %llu\n", __func__,
849                 (unsigned long long)sh->sector);
850
851         for (i = disks; i--; ) {
852                 struct r5dev *dev = &sh->dev[i];
853                 if (i != pd_idx)
854                         xor_srcs[count++] = dev->page;
855         }
856
857         tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
858                 &sh->ops.zero_sum_result, 0, NULL, NULL, NULL);
859
860         atomic_inc(&sh->count);
861         tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
862                 ops_complete_check, sh);
863 }
864
865 static void raid5_run_ops(struct stripe_head *sh, unsigned long ops_request)
866 {
867         int overlap_clear = 0, i, disks = sh->disks;
868         struct dma_async_tx_descriptor *tx = NULL;
869
870         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
871                 ops_run_biofill(sh);
872                 overlap_clear++;
873         }
874
875         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
876                 tx = ops_run_compute5(sh);
877                 /* terminate the chain if postxor is not set to be run */
878                 if (tx && !test_bit(STRIPE_OP_POSTXOR, &ops_request))
879                         async_tx_ack(tx);
880         }
881
882         if (test_bit(STRIPE_OP_PREXOR, &ops_request))
883                 tx = ops_run_prexor(sh, tx);
884
885         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
886                 tx = ops_run_biodrain(sh, tx);
887                 overlap_clear++;
888         }
889
890         if (test_bit(STRIPE_OP_POSTXOR, &ops_request))
891                 ops_run_postxor(sh, tx);
892
893         if (test_bit(STRIPE_OP_CHECK, &ops_request))
894                 ops_run_check(sh);
895
896         if (overlap_clear)
897                 for (i = disks; i--; ) {
898                         struct r5dev *dev = &sh->dev[i];
899                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
900                                 wake_up(&sh->raid_conf->wait_for_overlap);
901                 }
902 }
903
904 static int grow_one_stripe(raid5_conf_t *conf)
905 {
906         struct stripe_head *sh;
907         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
908         if (!sh)
909                 return 0;
910         memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
911         sh->raid_conf = conf;
912         spin_lock_init(&sh->lock);
913
914         if (grow_buffers(sh, conf->raid_disks)) {
915                 shrink_buffers(sh, conf->raid_disks);
916                 kmem_cache_free(conf->slab_cache, sh);
917                 return 0;
918         }
919         sh->disks = conf->raid_disks;
920         /* we just created an active stripe so... */
921         atomic_set(&sh->count, 1);
922         atomic_inc(&conf->active_stripes);
923         INIT_LIST_HEAD(&sh->lru);
924         release_stripe(sh);
925         return 1;
926 }
927
928 static int grow_stripes(raid5_conf_t *conf, int num)
929 {
930         struct kmem_cache *sc;
931         int devs = conf->raid_disks;
932
933         sprintf(conf->cache_name[0],
934                 "raid%d-%s", conf->level, mdname(conf->mddev));
935         sprintf(conf->cache_name[1],
936                 "raid%d-%s-alt", conf->level, mdname(conf->mddev));
937         conf->active_name = 0;
938         sc = kmem_cache_create(conf->cache_name[conf->active_name],
939                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
940                                0, 0, NULL);
941         if (!sc)
942                 return 1;
943         conf->slab_cache = sc;
944         conf->pool_size = devs;
945         while (num--)
946                 if (!grow_one_stripe(conf))
947                         return 1;
948         return 0;
949 }
950
951 static int resize_stripes(raid5_conf_t *conf, int newsize)
952 {
953         /* Make all the stripes able to hold 'newsize' devices.
954          * New slots in each stripe get 'page' set to a new page.
955          *
956          * This happens in stages:
957          * 1/ create a new kmem_cache and allocate the required number of
958          *    stripe_heads.
959          * 2/ gather all the old stripe_heads and tranfer the pages across
960          *    to the new stripe_heads.  This will have the side effect of
961          *    freezing the array as once all stripe_heads have been collected,
962          *    no IO will be possible.  Old stripe heads are freed once their
963          *    pages have been transferred over, and the old kmem_cache is
964          *    freed when all stripes are done.
965          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
966          *    we simple return a failre status - no need to clean anything up.
967          * 4/ allocate new pages for the new slots in the new stripe_heads.
968          *    If this fails, we don't bother trying the shrink the
969          *    stripe_heads down again, we just leave them as they are.
970          *    As each stripe_head is processed the new one is released into
971          *    active service.
972          *
973          * Once step2 is started, we cannot afford to wait for a write,
974          * so we use GFP_NOIO allocations.
975          */
976         struct stripe_head *osh, *nsh;
977         LIST_HEAD(newstripes);
978         struct disk_info *ndisks;
979         int err;
980         struct kmem_cache *sc;
981         int i;
982
983         if (newsize <= conf->pool_size)
984                 return 0; /* never bother to shrink */
985
986         err = md_allow_write(conf->mddev);
987         if (err)
988                 return err;
989
990         /* Step 1 */
991         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
992                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
993                                0, 0, NULL);
994         if (!sc)
995                 return -ENOMEM;
996
997         for (i = conf->max_nr_stripes; i; i--) {
998                 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
999                 if (!nsh)
1000                         break;
1001
1002                 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1003
1004                 nsh->raid_conf = conf;
1005                 spin_lock_init(&nsh->lock);
1006
1007                 list_add(&nsh->lru, &newstripes);
1008         }
1009         if (i) {
1010                 /* didn't get enough, give up */
1011                 while (!list_empty(&newstripes)) {
1012                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
1013                         list_del(&nsh->lru);
1014                         kmem_cache_free(sc, nsh);
1015                 }
1016                 kmem_cache_destroy(sc);
1017                 return -ENOMEM;
1018         }
1019         /* Step 2 - Must use GFP_NOIO now.
1020          * OK, we have enough stripes, start collecting inactive
1021          * stripes and copying them over
1022          */
1023         list_for_each_entry(nsh, &newstripes, lru) {
1024                 spin_lock_irq(&conf->device_lock);
1025                 wait_event_lock_irq(conf->wait_for_stripe,
1026                                     !list_empty(&conf->inactive_list),
1027                                     conf->device_lock,
1028                                     unplug_slaves(conf->mddev)
1029                         );
1030                 osh = get_free_stripe(conf);
1031                 spin_unlock_irq(&conf->device_lock);
1032                 atomic_set(&nsh->count, 1);
1033                 for(i=0; i<conf->pool_size; i++)
1034                         nsh->dev[i].page = osh->dev[i].page;
1035                 for( ; i<newsize; i++)
1036                         nsh->dev[i].page = NULL;
1037                 kmem_cache_free(conf->slab_cache, osh);
1038         }
1039         kmem_cache_destroy(conf->slab_cache);
1040
1041         /* Step 3.
1042          * At this point, we are holding all the stripes so the array
1043          * is completely stalled, so now is a good time to resize
1044          * conf->disks.
1045          */
1046         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1047         if (ndisks) {
1048                 for (i=0; i<conf->raid_disks; i++)
1049                         ndisks[i] = conf->disks[i];
1050                 kfree(conf->disks);
1051                 conf->disks = ndisks;
1052         } else
1053                 err = -ENOMEM;
1054
1055         /* Step 4, return new stripes to service */
1056         while(!list_empty(&newstripes)) {
1057                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1058                 list_del_init(&nsh->lru);
1059                 for (i=conf->raid_disks; i < newsize; i++)
1060                         if (nsh->dev[i].page == NULL) {
1061                                 struct page *p = alloc_page(GFP_NOIO);
1062                                 nsh->dev[i].page = p;
1063                                 if (!p)
1064                                         err = -ENOMEM;
1065                         }
1066                 release_stripe(nsh);
1067         }
1068         /* critical section pass, GFP_NOIO no longer needed */
1069
1070         conf->slab_cache = sc;
1071         conf->active_name = 1-conf->active_name;
1072         conf->pool_size = newsize;
1073         return err;
1074 }
1075
1076 static int drop_one_stripe(raid5_conf_t *conf)
1077 {
1078         struct stripe_head *sh;
1079
1080         spin_lock_irq(&conf->device_lock);
1081         sh = get_free_stripe(conf);
1082         spin_unlock_irq(&conf->device_lock);
1083         if (!sh)
1084                 return 0;
1085         BUG_ON(atomic_read(&sh->count));
1086         shrink_buffers(sh, conf->pool_size);
1087         kmem_cache_free(conf->slab_cache, sh);
1088         atomic_dec(&conf->active_stripes);
1089         return 1;
1090 }
1091
1092 static void shrink_stripes(raid5_conf_t *conf)
1093 {
1094         while (drop_one_stripe(conf))
1095                 ;
1096
1097         if (conf->slab_cache)
1098                 kmem_cache_destroy(conf->slab_cache);
1099         conf->slab_cache = NULL;
1100 }
1101
1102 static void raid5_end_read_request(struct bio * bi, int error)
1103 {
1104         struct stripe_head *sh = bi->bi_private;
1105         raid5_conf_t *conf = sh->raid_conf;
1106         int disks = sh->disks, i;
1107         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1108         char b[BDEVNAME_SIZE];
1109         mdk_rdev_t *rdev;
1110
1111
1112         for (i=0 ; i<disks; i++)
1113                 if (bi == &sh->dev[i].req)
1114                         break;
1115
1116         pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1117                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1118                 uptodate);
1119         if (i == disks) {
1120                 BUG();
1121                 return;
1122         }
1123
1124         if (uptodate) {
1125                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1126                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1127                         rdev = conf->disks[i].rdev;
1128                         printk_rl(KERN_INFO "raid5:%s: read error corrected"
1129                                   " (%lu sectors at %llu on %s)\n",
1130                                   mdname(conf->mddev), STRIPE_SECTORS,
1131                                   (unsigned long long)(sh->sector
1132                                                        + rdev->data_offset),
1133                                   bdevname(rdev->bdev, b));
1134                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1135                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1136                 }
1137                 if (atomic_read(&conf->disks[i].rdev->read_errors))
1138                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
1139         } else {
1140                 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1141                 int retry = 0;
1142                 rdev = conf->disks[i].rdev;
1143
1144                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1145                 atomic_inc(&rdev->read_errors);
1146                 if (conf->mddev->degraded)
1147                         printk_rl(KERN_WARNING
1148                                   "raid5:%s: read error not correctable "
1149                                   "(sector %llu on %s).\n",
1150                                   mdname(conf->mddev),
1151                                   (unsigned long long)(sh->sector
1152                                                        + rdev->data_offset),
1153                                   bdn);
1154                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1155                         /* Oh, no!!! */
1156                         printk_rl(KERN_WARNING
1157                                   "raid5:%s: read error NOT corrected!! "
1158                                   "(sector %llu on %s).\n",
1159                                   mdname(conf->mddev),
1160                                   (unsigned long long)(sh->sector
1161                                                        + rdev->data_offset),
1162                                   bdn);
1163                 else if (atomic_read(&rdev->read_errors)
1164                          > conf->max_nr_stripes)
1165                         printk(KERN_WARNING
1166                                "raid5:%s: Too many read errors, failing device %s.\n",
1167                                mdname(conf->mddev), bdn);
1168                 else
1169                         retry = 1;
1170                 if (retry)
1171                         set_bit(R5_ReadError, &sh->dev[i].flags);
1172                 else {
1173                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1174                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1175                         md_error(conf->mddev, rdev);
1176                 }
1177         }
1178         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1179         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1180         set_bit(STRIPE_HANDLE, &sh->state);
1181         release_stripe(sh);
1182 }
1183
1184 static void raid5_end_write_request(struct bio *bi, int error)
1185 {
1186         struct stripe_head *sh = bi->bi_private;
1187         raid5_conf_t *conf = sh->raid_conf;
1188         int disks = sh->disks, i;
1189         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1190
1191         for (i=0 ; i<disks; i++)
1192                 if (bi == &sh->dev[i].req)
1193                         break;
1194
1195         pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1196                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1197                 uptodate);
1198         if (i == disks) {
1199                 BUG();
1200                 return;
1201         }
1202
1203         if (!uptodate)
1204                 md_error(conf->mddev, conf->disks[i].rdev);
1205
1206         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1207         
1208         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1209         set_bit(STRIPE_HANDLE, &sh->state);
1210         release_stripe(sh);
1211 }
1212
1213
1214 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1215         
1216 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1217 {
1218         struct r5dev *dev = &sh->dev[i];
1219
1220         bio_init(&dev->req);
1221         dev->req.bi_io_vec = &dev->vec;
1222         dev->req.bi_vcnt++;
1223         dev->req.bi_max_vecs++;
1224         dev->vec.bv_page = dev->page;
1225         dev->vec.bv_len = STRIPE_SIZE;
1226         dev->vec.bv_offset = 0;
1227
1228         dev->req.bi_sector = sh->sector;
1229         dev->req.bi_private = sh;
1230
1231         dev->flags = 0;
1232         dev->sector = compute_blocknr(sh, i, previous);
1233 }
1234
1235 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1236 {
1237         char b[BDEVNAME_SIZE];
1238         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1239         pr_debug("raid5: error called\n");
1240
1241         if (!test_bit(Faulty, &rdev->flags)) {
1242                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1243                 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1244                         unsigned long flags;
1245                         spin_lock_irqsave(&conf->device_lock, flags);
1246                         mddev->degraded++;
1247                         spin_unlock_irqrestore(&conf->device_lock, flags);
1248                         /*
1249                          * if recovery was running, make sure it aborts.
1250                          */
1251                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1252                 }
1253                 set_bit(Faulty, &rdev->flags);
1254                 printk(KERN_ALERT
1255                        "raid5: Disk failure on %s, disabling device.\n"
1256                        "raid5: Operation continuing on %d devices.\n",
1257                        bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1258         }
1259 }
1260
1261 /*
1262  * Input: a 'big' sector number,
1263  * Output: index of the data and parity disk, and the sector # in them.
1264  */
1265 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1266                                      int previous, int *dd_idx,
1267                                      struct stripe_head *sh)
1268 {
1269         long stripe;
1270         unsigned long chunk_number;
1271         unsigned int chunk_offset;
1272         int pd_idx, qd_idx;
1273         int ddf_layout = 0;
1274         sector_t new_sector;
1275         int algorithm = previous ? conf->prev_algo
1276                                  : conf->algorithm;
1277         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1278                                          : conf->chunk_sectors;
1279         int raid_disks = previous ? conf->previous_raid_disks
1280                                   : conf->raid_disks;
1281         int data_disks = raid_disks - conf->max_degraded;
1282
1283         /* First compute the information on this sector */
1284
1285         /*
1286          * Compute the chunk number and the sector offset inside the chunk
1287          */
1288         chunk_offset = sector_div(r_sector, sectors_per_chunk);
1289         chunk_number = r_sector;
1290         BUG_ON(r_sector != chunk_number);
1291
1292         /*
1293          * Compute the stripe number
1294          */
1295         stripe = chunk_number / data_disks;
1296
1297         /*
1298          * Compute the data disk and parity disk indexes inside the stripe
1299          */
1300         *dd_idx = chunk_number % data_disks;
1301
1302         /*
1303          * Select the parity disk based on the user selected algorithm.
1304          */
1305         pd_idx = qd_idx = ~0;
1306         switch(conf->level) {
1307         case 4:
1308                 pd_idx = data_disks;
1309                 break;
1310         case 5:
1311                 switch (algorithm) {
1312                 case ALGORITHM_LEFT_ASYMMETRIC:
1313                         pd_idx = data_disks - stripe % raid_disks;
1314                         if (*dd_idx >= pd_idx)
1315                                 (*dd_idx)++;
1316                         break;
1317                 case ALGORITHM_RIGHT_ASYMMETRIC:
1318                         pd_idx = stripe % raid_disks;
1319                         if (*dd_idx >= pd_idx)
1320                                 (*dd_idx)++;
1321                         break;
1322                 case ALGORITHM_LEFT_SYMMETRIC:
1323                         pd_idx = data_disks - stripe % raid_disks;
1324                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1325                         break;
1326                 case ALGORITHM_RIGHT_SYMMETRIC:
1327                         pd_idx = stripe % raid_disks;
1328                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1329                         break;
1330                 case ALGORITHM_PARITY_0:
1331                         pd_idx = 0;
1332                         (*dd_idx)++;
1333                         break;
1334                 case ALGORITHM_PARITY_N:
1335                         pd_idx = data_disks;
1336                         break;
1337                 default:
1338                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1339                                 algorithm);
1340                         BUG();
1341                 }
1342                 break;
1343         case 6:
1344
1345                 switch (algorithm) {
1346                 case ALGORITHM_LEFT_ASYMMETRIC:
1347                         pd_idx = raid_disks - 1 - (stripe % raid_disks);
1348                         qd_idx = pd_idx + 1;
1349                         if (pd_idx == raid_disks-1) {
1350                                 (*dd_idx)++;    /* Q D D D P */
1351                                 qd_idx = 0;
1352                         } else if (*dd_idx >= pd_idx)
1353                                 (*dd_idx) += 2; /* D D P Q D */
1354                         break;
1355                 case ALGORITHM_RIGHT_ASYMMETRIC:
1356                         pd_idx = stripe % raid_disks;
1357                         qd_idx = pd_idx + 1;
1358                         if (pd_idx == raid_disks-1) {
1359                                 (*dd_idx)++;    /* Q D D D P */
1360                                 qd_idx = 0;
1361                         } else if (*dd_idx >= pd_idx)
1362                                 (*dd_idx) += 2; /* D D P Q D */
1363                         break;
1364                 case ALGORITHM_LEFT_SYMMETRIC:
1365                         pd_idx = raid_disks - 1 - (stripe % raid_disks);
1366                         qd_idx = (pd_idx + 1) % raid_disks;
1367                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1368                         break;
1369                 case ALGORITHM_RIGHT_SYMMETRIC:
1370                         pd_idx = stripe % raid_disks;
1371                         qd_idx = (pd_idx + 1) % raid_disks;
1372                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1373                         break;
1374
1375                 case ALGORITHM_PARITY_0:
1376                         pd_idx = 0;
1377                         qd_idx = 1;
1378                         (*dd_idx) += 2;
1379                         break;
1380                 case ALGORITHM_PARITY_N:
1381                         pd_idx = data_disks;
1382                         qd_idx = data_disks + 1;
1383                         break;
1384
1385                 case ALGORITHM_ROTATING_ZERO_RESTART:
1386                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
1387                          * of blocks for computing Q is different.
1388                          */
1389                         pd_idx = stripe % raid_disks;
1390                         qd_idx = pd_idx + 1;
1391                         if (pd_idx == raid_disks-1) {
1392                                 (*dd_idx)++;    /* Q D D D P */
1393                                 qd_idx = 0;
1394                         } else if (*dd_idx >= pd_idx)
1395                                 (*dd_idx) += 2; /* D D P Q D */
1396                         ddf_layout = 1;
1397                         break;
1398
1399                 case ALGORITHM_ROTATING_N_RESTART:
1400                         /* Same a left_asymmetric, by first stripe is
1401                          * D D D P Q  rather than
1402                          * Q D D D P
1403                          */
1404                         pd_idx = raid_disks - 1 - ((stripe + 1) % raid_disks);
1405                         qd_idx = pd_idx + 1;
1406                         if (pd_idx == raid_disks-1) {
1407                                 (*dd_idx)++;    /* Q D D D P */
1408                                 qd_idx = 0;
1409                         } else if (*dd_idx >= pd_idx)
1410                                 (*dd_idx) += 2; /* D D P Q D */
1411                         ddf_layout = 1;
1412                         break;
1413
1414                 case ALGORITHM_ROTATING_N_CONTINUE:
1415                         /* Same as left_symmetric but Q is before P */
1416                         pd_idx = raid_disks - 1 - (stripe % raid_disks);
1417                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1418                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1419                         ddf_layout = 1;
1420                         break;
1421
1422                 case ALGORITHM_LEFT_ASYMMETRIC_6:
1423                         /* RAID5 left_asymmetric, with Q on last device */
1424                         pd_idx = data_disks - stripe % (raid_disks-1);
1425                         if (*dd_idx >= pd_idx)
1426                                 (*dd_idx)++;
1427                         qd_idx = raid_disks - 1;
1428                         break;
1429
1430                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1431                         pd_idx = stripe % (raid_disks-1);
1432                         if (*dd_idx >= pd_idx)
1433                                 (*dd_idx)++;
1434                         qd_idx = raid_disks - 1;
1435                         break;
1436
1437                 case ALGORITHM_LEFT_SYMMETRIC_6:
1438                         pd_idx = data_disks - stripe % (raid_disks-1);
1439                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1440                         qd_idx = raid_disks - 1;
1441                         break;
1442
1443                 case ALGORITHM_RIGHT_SYMMETRIC_6:
1444                         pd_idx = stripe % (raid_disks-1);
1445                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1446                         qd_idx = raid_disks - 1;
1447                         break;
1448
1449                 case ALGORITHM_PARITY_0_6:
1450                         pd_idx = 0;
1451                         (*dd_idx)++;
1452                         qd_idx = raid_disks - 1;
1453                         break;
1454
1455
1456                 default:
1457                         printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1458                                algorithm);
1459                         BUG();
1460                 }
1461                 break;
1462         }
1463
1464         if (sh) {
1465                 sh->pd_idx = pd_idx;
1466                 sh->qd_idx = qd_idx;
1467                 sh->ddf_layout = ddf_layout;
1468         }
1469         /*
1470          * Finally, compute the new sector number
1471          */
1472         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1473         return new_sector;
1474 }
1475
1476
1477 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1478 {
1479         raid5_conf_t *conf = sh->raid_conf;
1480         int raid_disks = sh->disks;
1481         int data_disks = raid_disks - conf->max_degraded;
1482         sector_t new_sector = sh->sector, check;
1483         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1484                                          : conf->chunk_sectors;
1485         int algorithm = previous ? conf->prev_algo
1486                                  : conf->algorithm;
1487         sector_t stripe;
1488         int chunk_offset;
1489         int chunk_number, dummy1, dd_idx = i;
1490         sector_t r_sector;
1491         struct stripe_head sh2;
1492
1493
1494         chunk_offset = sector_div(new_sector, sectors_per_chunk);
1495         stripe = new_sector;
1496         BUG_ON(new_sector != stripe);
1497
1498         if (i == sh->pd_idx)
1499                 return 0;
1500         switch(conf->level) {
1501         case 4: break;
1502         case 5:
1503                 switch (algorithm) {
1504                 case ALGORITHM_LEFT_ASYMMETRIC:
1505                 case ALGORITHM_RIGHT_ASYMMETRIC:
1506                         if (i > sh->pd_idx)
1507                                 i--;
1508                         break;
1509                 case ALGORITHM_LEFT_SYMMETRIC:
1510                 case ALGORITHM_RIGHT_SYMMETRIC:
1511                         if (i < sh->pd_idx)
1512                                 i += raid_disks;
1513                         i -= (sh->pd_idx + 1);
1514                         break;
1515                 case ALGORITHM_PARITY_0:
1516                         i -= 1;
1517                         break;
1518                 case ALGORITHM_PARITY_N:
1519                         break;
1520                 default:
1521                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1522                                algorithm);
1523                         BUG();
1524                 }
1525                 break;
1526         case 6:
1527                 if (i == sh->qd_idx)
1528                         return 0; /* It is the Q disk */
1529                 switch (algorithm) {
1530                 case ALGORITHM_LEFT_ASYMMETRIC:
1531                 case ALGORITHM_RIGHT_ASYMMETRIC:
1532                 case ALGORITHM_ROTATING_ZERO_RESTART:
1533                 case ALGORITHM_ROTATING_N_RESTART:
1534                         if (sh->pd_idx == raid_disks-1)
1535                                 i--;    /* Q D D D P */
1536                         else if (i > sh->pd_idx)
1537                                 i -= 2; /* D D P Q D */
1538                         break;
1539                 case ALGORITHM_LEFT_SYMMETRIC:
1540                 case ALGORITHM_RIGHT_SYMMETRIC:
1541                         if (sh->pd_idx == raid_disks-1)
1542                                 i--; /* Q D D D P */
1543                         else {
1544                                 /* D D P Q D */
1545                                 if (i < sh->pd_idx)
1546                                         i += raid_disks;
1547                                 i -= (sh->pd_idx + 2);
1548                         }
1549                         break;
1550                 case ALGORITHM_PARITY_0:
1551                         i -= 2;
1552                         break;
1553                 case ALGORITHM_PARITY_N:
1554                         break;
1555                 case ALGORITHM_ROTATING_N_CONTINUE:
1556                         if (sh->pd_idx == 0)
1557                                 i--;    /* P D D D Q */
1558                         else if (i > sh->pd_idx)
1559                                 i -= 2; /* D D Q P D */
1560                         break;
1561                 case ALGORITHM_LEFT_ASYMMETRIC_6:
1562                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1563                         if (i > sh->pd_idx)
1564                                 i--;
1565                         break;
1566                 case ALGORITHM_LEFT_SYMMETRIC_6:
1567                 case ALGORITHM_RIGHT_SYMMETRIC_6:
1568                         if (i < sh->pd_idx)
1569                                 i += data_disks + 1;
1570                         i -= (sh->pd_idx + 1);
1571                         break;
1572                 case ALGORITHM_PARITY_0_6:
1573                         i -= 1;
1574                         break;
1575                 default:
1576                         printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1577                                algorithm);
1578                         BUG();
1579                 }
1580                 break;
1581         }
1582
1583         chunk_number = stripe * data_disks + i;
1584         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1585
1586         check = raid5_compute_sector(conf, r_sector,
1587                                      previous, &dummy1, &sh2);
1588         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
1589                 || sh2.qd_idx != sh->qd_idx) {
1590                 printk(KERN_ERR "compute_blocknr: map not correct\n");
1591                 return 0;
1592         }
1593         return r_sector;
1594 }
1595
1596
1597
1598 /*
1599  * Copy data between a page in the stripe cache, and one or more bion
1600  * The page could align with the middle of the bio, or there could be
1601  * several bion, each with several bio_vecs, which cover part of the page
1602  * Multiple bion are linked together on bi_next.  There may be extras
1603  * at the end of this list.  We ignore them.
1604  */
1605 static void copy_data(int frombio, struct bio *bio,
1606                      struct page *page,
1607                      sector_t sector)
1608 {
1609         char *pa = page_address(page);
1610         struct bio_vec *bvl;
1611         int i;
1612         int page_offset;
1613
1614         if (bio->bi_sector >= sector)
1615                 page_offset = (signed)(bio->bi_sector - sector) * 512;
1616         else
1617                 page_offset = (signed)(sector - bio->bi_sector) * -512;
1618         bio_for_each_segment(bvl, bio, i) {
1619                 int len = bio_iovec_idx(bio,i)->bv_len;
1620                 int clen;
1621                 int b_offset = 0;
1622
1623                 if (page_offset < 0) {
1624                         b_offset = -page_offset;
1625                         page_offset += b_offset;
1626                         len -= b_offset;
1627                 }
1628
1629                 if (len > 0 && page_offset + len > STRIPE_SIZE)
1630                         clen = STRIPE_SIZE - page_offset;
1631                 else clen = len;
1632
1633                 if (clen > 0) {
1634                         char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1635                         if (frombio)
1636                                 memcpy(pa+page_offset, ba+b_offset, clen);
1637                         else
1638                                 memcpy(ba+b_offset, pa+page_offset, clen);
1639                         __bio_kunmap_atomic(ba, KM_USER0);
1640                 }
1641                 if (clen < len) /* hit end of page */
1642                         break;
1643                 page_offset +=  len;
1644         }
1645 }
1646
1647 #define check_xor()     do {                                              \
1648                                 if (count == MAX_XOR_BLOCKS) {            \
1649                                 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1650                                 count = 0;                                \
1651                            }                                              \
1652                         } while(0)
1653
1654 static void compute_parity6(struct stripe_head *sh, int method)
1655 {
1656         raid5_conf_t *conf = sh->raid_conf;
1657         int i, pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1658         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1659         struct bio *chosen;
1660         /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1661         void *ptrs[syndrome_disks+2];
1662
1663         pd_idx = sh->pd_idx;
1664         qd_idx = sh->qd_idx;
1665         d0_idx = raid6_d0(sh);
1666
1667         pr_debug("compute_parity, stripe %llu, method %d\n",
1668                 (unsigned long long)sh->sector, method);
1669
1670         switch(method) {
1671         case READ_MODIFY_WRITE:
1672                 BUG();          /* READ_MODIFY_WRITE N/A for RAID-6 */
1673         case RECONSTRUCT_WRITE:
1674                 for (i= disks; i-- ;)
1675                         if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1676                                 chosen = sh->dev[i].towrite;
1677                                 sh->dev[i].towrite = NULL;
1678
1679                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1680                                         wake_up(&conf->wait_for_overlap);
1681
1682                                 BUG_ON(sh->dev[i].written);
1683                                 sh->dev[i].written = chosen;
1684                         }
1685                 break;
1686         case CHECK_PARITY:
1687                 BUG();          /* Not implemented yet */
1688         }
1689
1690         for (i = disks; i--;)
1691                 if (sh->dev[i].written) {
1692                         sector_t sector = sh->dev[i].sector;
1693                         struct bio *wbi = sh->dev[i].written;
1694                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1695                                 copy_data(1, wbi, sh->dev[i].page, sector);
1696                                 wbi = r5_next_bio(wbi, sector);
1697                         }
1698
1699                         set_bit(R5_LOCKED, &sh->dev[i].flags);
1700                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
1701                 }
1702
1703         /* Note that unlike RAID-5, the ordering of the disks matters greatly.*/
1704
1705         for (i = 0; i < disks; i++)
1706                 ptrs[i] = (void *)raid6_empty_zero_page;
1707
1708         count = 0;
1709         i = d0_idx;
1710         do {
1711                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1712
1713                 ptrs[slot] = page_address(sh->dev[i].page);
1714                 if (slot < syndrome_disks &&
1715                     !test_bit(R5_UPTODATE, &sh->dev[i].flags)) {
1716                         printk(KERN_ERR "block %d/%d not uptodate "
1717                                "on parity calc\n", i, count);
1718                         BUG();
1719                 }
1720
1721                 i = raid6_next_disk(i, disks);
1722         } while (i != d0_idx);
1723         BUG_ON(count != syndrome_disks);
1724
1725         raid6_call.gen_syndrome(syndrome_disks+2, STRIPE_SIZE, ptrs);
1726
1727         switch(method) {
1728         case RECONSTRUCT_WRITE:
1729                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1730                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1731                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
1732                 set_bit(R5_LOCKED,   &sh->dev[qd_idx].flags);
1733                 break;
1734         case UPDATE_PARITY:
1735                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1736                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1737                 break;
1738         }
1739 }
1740
1741
1742 /* Compute one missing block */
1743 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1744 {
1745         int i, count, disks = sh->disks;
1746         void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1747         int qd_idx = sh->qd_idx;
1748
1749         pr_debug("compute_block_1, stripe %llu, idx %d\n",
1750                 (unsigned long long)sh->sector, dd_idx);
1751
1752         if ( dd_idx == qd_idx ) {
1753                 /* We're actually computing the Q drive */
1754                 compute_parity6(sh, UPDATE_PARITY);
1755         } else {
1756                 dest = page_address(sh->dev[dd_idx].page);
1757                 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1758                 count = 0;
1759                 for (i = disks ; i--; ) {
1760                         if (i == dd_idx || i == qd_idx)
1761                                 continue;
1762                         p = page_address(sh->dev[i].page);
1763                         if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1764                                 ptr[count++] = p;
1765                         else
1766                                 printk("compute_block() %d, stripe %llu, %d"
1767                                        " not present\n", dd_idx,
1768                                        (unsigned long long)sh->sector, i);
1769
1770                         check_xor();
1771                 }
1772                 if (count)
1773                         xor_blocks(count, STRIPE_SIZE, dest, ptr);
1774                 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1775                 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1776         }
1777 }
1778
1779 /* Compute two missing blocks */
1780 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1781 {
1782         int i, count, disks = sh->disks;
1783         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1784         int d0_idx = raid6_d0(sh);
1785         int faila = -1, failb = -1;
1786         /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1787         void *ptrs[syndrome_disks+2];
1788
1789         for (i = 0; i < disks ; i++)
1790                 ptrs[i] = (void *)raid6_empty_zero_page;
1791         count = 0;
1792         i = d0_idx;
1793         do {
1794                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1795
1796                 ptrs[slot] = page_address(sh->dev[i].page);
1797
1798                 if (i == dd_idx1)
1799                         faila = slot;
1800                 if (i == dd_idx2)
1801                         failb = slot;
1802                 i = raid6_next_disk(i, disks);
1803         } while (i != d0_idx);
1804         BUG_ON(count != syndrome_disks);
1805
1806         BUG_ON(faila == failb);
1807         if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1808
1809         pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1810                  (unsigned long long)sh->sector, dd_idx1, dd_idx2,
1811                  faila, failb);
1812
1813         if (failb == syndrome_disks+1) {
1814                 /* Q disk is one of the missing disks */
1815                 if (faila == syndrome_disks) {
1816                         /* Missing P+Q, just recompute */
1817                         compute_parity6(sh, UPDATE_PARITY);
1818                         return;
1819                 } else {
1820                         /* We're missing D+Q; recompute D from P */
1821                         compute_block_1(sh, ((dd_idx1 == sh->qd_idx) ?
1822                                              dd_idx2 : dd_idx1),
1823                                         0);
1824                         compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1825                         return;
1826                 }
1827         }
1828
1829         /* We're missing D+P or D+D; */
1830         if (failb == syndrome_disks) {
1831                 /* We're missing D+P. */
1832                 raid6_datap_recov(syndrome_disks+2, STRIPE_SIZE, faila, ptrs);
1833         } else {
1834                 /* We're missing D+D. */
1835                 raid6_2data_recov(syndrome_disks+2, STRIPE_SIZE, faila, failb,
1836                                   ptrs);
1837         }
1838
1839         /* Both the above update both missing blocks */
1840         set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1841         set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1842 }
1843
1844 static void
1845 schedule_reconstruction5(struct stripe_head *sh, struct stripe_head_state *s,
1846                          int rcw, int expand)
1847 {
1848         int i, pd_idx = sh->pd_idx, disks = sh->disks;
1849
1850         if (rcw) {
1851                 /* if we are not expanding this is a proper write request, and
1852                  * there will be bios with new data to be drained into the
1853                  * stripe cache
1854                  */
1855                 if (!expand) {
1856                         sh->reconstruct_state = reconstruct_state_drain_run;
1857                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1858                 } else
1859                         sh->reconstruct_state = reconstruct_state_run;
1860
1861                 set_bit(STRIPE_OP_POSTXOR, &s->ops_request);
1862
1863                 for (i = disks; i--; ) {
1864                         struct r5dev *dev = &sh->dev[i];
1865
1866                         if (dev->towrite) {
1867                                 set_bit(R5_LOCKED, &dev->flags);
1868                                 set_bit(R5_Wantdrain, &dev->flags);
1869                                 if (!expand)
1870                                         clear_bit(R5_UPTODATE, &dev->flags);
1871                                 s->locked++;
1872                         }
1873                 }
1874                 if (s->locked + 1 == disks)
1875                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
1876                                 atomic_inc(&sh->raid_conf->pending_full_writes);
1877         } else {
1878                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1879                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1880
1881                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
1882                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
1883                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1884                 set_bit(STRIPE_OP_POSTXOR, &s->ops_request);
1885
1886                 for (i = disks; i--; ) {
1887                         struct r5dev *dev = &sh->dev[i];
1888                         if (i == pd_idx)
1889                                 continue;
1890
1891                         if (dev->towrite &&
1892                             (test_bit(R5_UPTODATE, &dev->flags) ||
1893                              test_bit(R5_Wantcompute, &dev->flags))) {
1894                                 set_bit(R5_Wantdrain, &dev->flags);
1895                                 set_bit(R5_LOCKED, &dev->flags);
1896                                 clear_bit(R5_UPTODATE, &dev->flags);
1897                                 s->locked++;
1898                         }
1899                 }
1900         }
1901
1902         /* keep the parity disk locked while asynchronous operations
1903          * are in flight
1904          */
1905         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1906         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1907         s->locked++;
1908
1909         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
1910                 __func__, (unsigned long long)sh->sector,
1911                 s->locked, s->ops_request);
1912 }
1913
1914 /*
1915  * Each stripe/dev can have one or more bion attached.
1916  * toread/towrite point to the first in a chain.
1917  * The bi_next chain must be in order.
1918  */
1919 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1920 {
1921         struct bio **bip;
1922         raid5_conf_t *conf = sh->raid_conf;
1923         int firstwrite=0;
1924
1925         pr_debug("adding bh b#%llu to stripe s#%llu\n",
1926                 (unsigned long long)bi->bi_sector,
1927                 (unsigned long long)sh->sector);
1928
1929
1930         spin_lock(&sh->lock);
1931         spin_lock_irq(&conf->device_lock);
1932         if (forwrite) {
1933                 bip = &sh->dev[dd_idx].towrite;
1934                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1935                         firstwrite = 1;
1936         } else
1937                 bip = &sh->dev[dd_idx].toread;
1938         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1939                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1940                         goto overlap;
1941                 bip = & (*bip)->bi_next;
1942         }
1943         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1944                 goto overlap;
1945
1946         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1947         if (*bip)
1948                 bi->bi_next = *bip;
1949         *bip = bi;
1950         bi->bi_phys_segments++;
1951         spin_unlock_irq(&conf->device_lock);
1952         spin_unlock(&sh->lock);
1953
1954         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1955                 (unsigned long long)bi->bi_sector,
1956                 (unsigned long long)sh->sector, dd_idx);
1957
1958         if (conf->mddev->bitmap && firstwrite) {
1959                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1960                                   STRIPE_SECTORS, 0);
1961                 sh->bm_seq = conf->seq_flush+1;
1962                 set_bit(STRIPE_BIT_DELAY, &sh->state);
1963         }
1964
1965         if (forwrite) {
1966                 /* check if page is covered */
1967                 sector_t sector = sh->dev[dd_idx].sector;
1968                 for (bi=sh->dev[dd_idx].towrite;
1969                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1970                              bi && bi->bi_sector <= sector;
1971                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1972                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1973                                 sector = bi->bi_sector + (bi->bi_size>>9);
1974                 }
1975                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1976                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1977         }
1978         return 1;
1979
1980  overlap:
1981         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1982         spin_unlock_irq(&conf->device_lock);
1983         spin_unlock(&sh->lock);
1984         return 0;
1985 }
1986
1987 static void end_reshape(raid5_conf_t *conf);
1988
1989 static int page_is_zero(struct page *p)
1990 {
1991         char *a = page_address(p);
1992         return ((*(u32*)a) == 0 &&
1993                 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1994 }
1995
1996 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
1997                             struct stripe_head *sh)
1998 {
1999         int sectors_per_chunk =
2000                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2001         int dd_idx;
2002         int chunk_offset = sector_div(stripe, sectors_per_chunk);
2003         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2004
2005         raid5_compute_sector(conf,
2006                              stripe * (disks - conf->max_degraded)
2007                              *sectors_per_chunk + chunk_offset,
2008                              previous,
2009                              &dd_idx, sh);
2010 }
2011
2012 static void
2013 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2014                                 struct stripe_head_state *s, int disks,
2015                                 struct bio **return_bi)
2016 {
2017         int i;
2018         for (i = disks; i--; ) {
2019                 struct bio *bi;
2020                 int bitmap_end = 0;
2021
2022                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2023                         mdk_rdev_t *rdev;
2024                         rcu_read_lock();
2025                         rdev = rcu_dereference(conf->disks[i].rdev);
2026                         if (rdev && test_bit(In_sync, &rdev->flags))
2027                                 /* multiple read failures in one stripe */
2028                                 md_error(conf->mddev, rdev);
2029                         rcu_read_unlock();
2030                 }
2031                 spin_lock_irq(&conf->device_lock);
2032                 /* fail all writes first */
2033                 bi = sh->dev[i].towrite;
2034                 sh->dev[i].towrite = NULL;
2035                 if (bi) {
2036                         s->to_write--;
2037                         bitmap_end = 1;
2038                 }
2039
2040                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2041                         wake_up(&conf->wait_for_overlap);
2042
2043                 while (bi && bi->bi_sector <
2044                         sh->dev[i].sector + STRIPE_SECTORS) {
2045                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2046                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2047                         if (!raid5_dec_bi_phys_segments(bi)) {
2048                                 md_write_end(conf->mddev);
2049                                 bi->bi_next = *return_bi;
2050                                 *return_bi = bi;
2051                         }
2052                         bi = nextbi;
2053                 }
2054                 /* and fail all 'written' */
2055                 bi = sh->dev[i].written;
2056                 sh->dev[i].written = NULL;
2057                 if (bi) bitmap_end = 1;
2058                 while (bi && bi->bi_sector <
2059                        sh->dev[i].sector + STRIPE_SECTORS) {
2060                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2061                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2062                         if (!raid5_dec_bi_phys_segments(bi)) {
2063                                 md_write_end(conf->mddev);
2064                                 bi->bi_next = *return_bi;
2065                                 *return_bi = bi;
2066                         }
2067                         bi = bi2;
2068                 }
2069
2070                 /* fail any reads if this device is non-operational and
2071                  * the data has not reached the cache yet.
2072                  */
2073                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2074                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2075                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
2076                         bi = sh->dev[i].toread;
2077                         sh->dev[i].toread = NULL;
2078                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2079                                 wake_up(&conf->wait_for_overlap);
2080                         if (bi) s->to_read--;
2081                         while (bi && bi->bi_sector <
2082                                sh->dev[i].sector + STRIPE_SECTORS) {
2083                                 struct bio *nextbi =
2084                                         r5_next_bio(bi, sh->dev[i].sector);
2085                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2086                                 if (!raid5_dec_bi_phys_segments(bi)) {
2087                                         bi->bi_next = *return_bi;
2088                                         *return_bi = bi;
2089                                 }
2090                                 bi = nextbi;
2091                         }
2092                 }
2093                 spin_unlock_irq(&conf->device_lock);
2094                 if (bitmap_end)
2095                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2096                                         STRIPE_SECTORS, 0, 0);
2097         }
2098
2099         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2100                 if (atomic_dec_and_test(&conf->pending_full_writes))
2101                         md_wakeup_thread(conf->mddev->thread);
2102 }
2103
2104 /* fetch_block5 - checks the given member device to see if its data needs
2105  * to be read or computed to satisfy a request.
2106  *
2107  * Returns 1 when no more member devices need to be checked, otherwise returns
2108  * 0 to tell the loop in handle_stripe_fill5 to continue
2109  */
2110 static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s,
2111                         int disk_idx, int disks)
2112 {
2113         struct r5dev *dev = &sh->dev[disk_idx];
2114         struct r5dev *failed_dev = &sh->dev[s->failed_num];
2115
2116         /* is the data in this block needed, and can we get it? */
2117         if (!test_bit(R5_LOCKED, &dev->flags) &&
2118             !test_bit(R5_UPTODATE, &dev->flags) &&
2119             (dev->toread ||
2120              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2121              s->syncing || s->expanding ||
2122              (s->failed &&
2123               (failed_dev->toread ||
2124                (failed_dev->towrite &&
2125                 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) {
2126                 /* We would like to get this block, possibly by computing it,
2127                  * otherwise read it if the backing disk is insync
2128                  */
2129                 if ((s->uptodate == disks - 1) &&
2130                     (s->failed && disk_idx == s->failed_num)) {
2131                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2132                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2133                         set_bit(R5_Wantcompute, &dev->flags);
2134                         sh->ops.target = disk_idx;
2135                         s->req_compute = 1;
2136                         /* Careful: from this point on 'uptodate' is in the eye
2137                          * of raid5_run_ops which services 'compute' operations
2138                          * before writes. R5_Wantcompute flags a block that will
2139                          * be R5_UPTODATE by the time it is needed for a
2140                          * subsequent operation.
2141                          */
2142                         s->uptodate++;
2143                         return 1; /* uptodate + compute == disks */
2144                 } else if (test_bit(R5_Insync, &dev->flags)) {
2145                         set_bit(R5_LOCKED, &dev->flags);
2146                         set_bit(R5_Wantread, &dev->flags);
2147                         s->locked++;
2148                         pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2149                                 s->syncing);
2150                 }
2151         }
2152
2153         return 0;
2154 }
2155
2156 /**
2157  * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2158  */
2159 static void handle_stripe_fill5(struct stripe_head *sh,
2160                         struct stripe_head_state *s, int disks)
2161 {
2162         int i;
2163
2164         /* look for blocks to read/compute, skip this if a compute
2165          * is already in flight, or if the stripe contents are in the
2166          * midst of changing due to a write
2167          */
2168         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2169             !sh->reconstruct_state)
2170                 for (i = disks; i--; )
2171                         if (fetch_block5(sh, s, i, disks))
2172                                 break;
2173         set_bit(STRIPE_HANDLE, &sh->state);
2174 }
2175
2176 static void handle_stripe_fill6(struct stripe_head *sh,
2177                         struct stripe_head_state *s, struct r6_state *r6s,
2178                         int disks)
2179 {
2180         int i;
2181         for (i = disks; i--; ) {
2182                 struct r5dev *dev = &sh->dev[i];
2183                 if (!test_bit(R5_LOCKED, &dev->flags) &&
2184                     !test_bit(R5_UPTODATE, &dev->flags) &&
2185                     (dev->toread || (dev->towrite &&
2186                      !test_bit(R5_OVERWRITE, &dev->flags)) ||
2187                      s->syncing || s->expanding ||
2188                      (s->failed >= 1 &&
2189                       (sh->dev[r6s->failed_num[0]].toread ||
2190                        s->to_write)) ||
2191                      (s->failed >= 2 &&
2192                       (sh->dev[r6s->failed_num[1]].toread ||
2193                        s->to_write)))) {
2194                         /* we would like to get this block, possibly
2195                          * by computing it, but we might not be able to
2196                          */
2197                         if ((s->uptodate == disks - 1) &&
2198                             (s->failed && (i == r6s->failed_num[0] ||
2199                                            i == r6s->failed_num[1]))) {
2200                                 pr_debug("Computing stripe %llu block %d\n",
2201                                        (unsigned long long)sh->sector, i);
2202                                 compute_block_1(sh, i, 0);
2203                                 s->uptodate++;
2204                         } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2205                                 /* Computing 2-failure is *very* expensive; only
2206                                  * do it if failed >= 2
2207                                  */
2208                                 int other;
2209                                 for (other = disks; other--; ) {
2210                                         if (other == i)
2211                                                 continue;
2212                                         if (!test_bit(R5_UPTODATE,
2213                                               &sh->dev[other].flags))
2214                                                 break;
2215                                 }
2216                                 BUG_ON(other < 0);
2217                                 pr_debug("Computing stripe %llu blocks %d,%d\n",
2218                                        (unsigned long long)sh->sector,
2219                                        i, other);
2220                                 compute_block_2(sh, i, other);
2221                                 s->uptodate += 2;
2222                         } else if (test_bit(R5_Insync, &dev->flags)) {
2223                                 set_bit(R5_LOCKED, &dev->flags);
2224                                 set_bit(R5_Wantread, &dev->flags);
2225                                 s->locked++;
2226                                 pr_debug("Reading block %d (sync=%d)\n",
2227                                         i, s->syncing);
2228                         }
2229                 }
2230         }
2231         set_bit(STRIPE_HANDLE, &sh->state);
2232 }
2233
2234
2235 /* handle_stripe_clean_event
2236  * any written block on an uptodate or failed drive can be returned.
2237  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2238  * never LOCKED, so we don't need to test 'failed' directly.
2239  */
2240 static void handle_stripe_clean_event(raid5_conf_t *conf,
2241         struct stripe_head *sh, int disks, struct bio **return_bi)
2242 {
2243         int i;
2244         struct r5dev *dev;
2245
2246         for (i = disks; i--; )
2247                 if (sh->dev[i].written) {
2248                         dev = &sh->dev[i];
2249                         if (!test_bit(R5_LOCKED, &dev->flags) &&
2250                                 test_bit(R5_UPTODATE, &dev->flags)) {
2251                                 /* We can return any write requests */
2252                                 struct bio *wbi, *wbi2;
2253                                 int bitmap_end = 0;
2254                                 pr_debug("Return write for disc %d\n", i);
2255                                 spin_lock_irq(&conf->device_lock);
2256                                 wbi = dev->written;
2257                                 dev->written = NULL;
2258                                 while (wbi && wbi->bi_sector <
2259                                         dev->sector + STRIPE_SECTORS) {
2260                                         wbi2 = r5_next_bio(wbi, dev->sector);
2261                                         if (!raid5_dec_bi_phys_segments(wbi)) {
2262                                                 md_write_end(conf->mddev);
2263                                                 wbi->bi_next = *return_bi;
2264                                                 *return_bi = wbi;
2265                                         }
2266                                         wbi = wbi2;
2267                                 }
2268                                 if (dev->towrite == NULL)
2269                                         bitmap_end = 1;
2270                                 spin_unlock_irq(&conf->device_lock);
2271                                 if (bitmap_end)
2272                                         bitmap_endwrite(conf->mddev->bitmap,
2273                                                         sh->sector,
2274                                                         STRIPE_SECTORS,
2275                                          !test_bit(STRIPE_DEGRADED, &sh->state),
2276                                                         0);
2277                         }
2278                 }
2279
2280         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2281                 if (atomic_dec_and_test(&conf->pending_full_writes))
2282                         md_wakeup_thread(conf->mddev->thread);
2283 }
2284
2285 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2286                 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2287 {
2288         int rmw = 0, rcw = 0, i;
2289         for (i = disks; i--; ) {
2290                 /* would I have to read this buffer for read_modify_write */
2291                 struct r5dev *dev = &sh->dev[i];
2292                 if ((dev->towrite || i == sh->pd_idx) &&
2293                     !test_bit(R5_LOCKED, &dev->flags) &&
2294                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2295                       test_bit(R5_Wantcompute, &dev->flags))) {
2296                         if (test_bit(R5_Insync, &dev->flags))
2297                                 rmw++;
2298                         else
2299                                 rmw += 2*disks;  /* cannot read it */
2300                 }
2301                 /* Would I have to read this buffer for reconstruct_write */
2302                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2303                     !test_bit(R5_LOCKED, &dev->flags) &&
2304                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2305                     test_bit(R5_Wantcompute, &dev->flags))) {
2306                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2307                         else
2308                                 rcw += 2*disks;
2309                 }
2310         }
2311         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2312                 (unsigned long long)sh->sector, rmw, rcw);
2313         set_bit(STRIPE_HANDLE, &sh->state);
2314         if (rmw < rcw && rmw > 0)
2315                 /* prefer read-modify-write, but need to get some data */
2316                 for (i = disks; i--; ) {
2317                         struct r5dev *dev = &sh->dev[i];
2318                         if ((dev->towrite || i == sh->pd_idx) &&
2319                             !test_bit(R5_LOCKED, &dev->flags) &&
2320                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2321                             test_bit(R5_Wantcompute, &dev->flags)) &&
2322                             test_bit(R5_Insync, &dev->flags)) {
2323                                 if (
2324                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2325                                         pr_debug("Read_old block "
2326                                                 "%d for r-m-w\n", i);
2327                                         set_bit(R5_LOCKED, &dev->flags);
2328                                         set_bit(R5_Wantread, &dev->flags);
2329                                         s->locked++;
2330                                 } else {
2331                                         set_bit(STRIPE_DELAYED, &sh->state);
2332                                         set_bit(STRIPE_HANDLE, &sh->state);
2333                                 }
2334                         }
2335                 }
2336         if (rcw <= rmw && rcw > 0)
2337                 /* want reconstruct write, but need to get some data */
2338                 for (i = disks; i--; ) {
2339                         struct r5dev *dev = &sh->dev[i];
2340                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2341                             i != sh->pd_idx &&
2342                             !test_bit(R5_LOCKED, &dev->flags) &&
2343                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2344                             test_bit(R5_Wantcompute, &dev->flags)) &&
2345                             test_bit(R5_Insync, &dev->flags)) {
2346                                 if (
2347                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2348                                         pr_debug("Read_old block "
2349                                                 "%d for Reconstruct\n", i);
2350                                         set_bit(R5_LOCKED, &dev->flags);
2351                                         set_bit(R5_Wantread, &dev->flags);
2352                                         s->locked++;
2353                                 } else {
2354                                         set_bit(STRIPE_DELAYED, &sh->state);
2355                                         set_bit(STRIPE_HANDLE, &sh->state);
2356                                 }
2357                         }
2358                 }
2359         /* now if nothing is locked, and if we have enough data,
2360          * we can start a write request
2361          */
2362         /* since handle_stripe can be called at any time we need to handle the
2363          * case where a compute block operation has been submitted and then a
2364          * subsequent call wants to start a write request.  raid5_run_ops only
2365          * handles the case where compute block and postxor are requested
2366          * simultaneously.  If this is not the case then new writes need to be
2367          * held off until the compute completes.
2368          */
2369         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2370             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2371             !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2372                 schedule_reconstruction5(sh, s, rcw == 0, 0);
2373 }
2374
2375 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2376                 struct stripe_head *sh, struct stripe_head_state *s,
2377                 struct r6_state *r6s, int disks)
2378 {
2379         int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2380         int qd_idx = sh->qd_idx;
2381         for (i = disks; i--; ) {
2382                 struct r5dev *dev = &sh->dev[i];
2383                 /* Would I have to read this buffer for reconstruct_write */
2384                 if (!test_bit(R5_OVERWRITE, &dev->flags)
2385                     && i != pd_idx && i != qd_idx
2386                     && (!test_bit(R5_LOCKED, &dev->flags)
2387                             ) &&
2388                     !test_bit(R5_UPTODATE, &dev->flags)) {
2389                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2390                         else {
2391                                 pr_debug("raid6: must_compute: "
2392                                         "disk %d flags=%#lx\n", i, dev->flags);
2393                                 must_compute++;
2394                         }
2395                 }
2396         }
2397         pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2398                (unsigned long long)sh->sector, rcw, must_compute);
2399         set_bit(STRIPE_HANDLE, &sh->state);
2400
2401         if (rcw > 0)
2402                 /* want reconstruct write, but need to get some data */
2403                 for (i = disks; i--; ) {
2404                         struct r5dev *dev = &sh->dev[i];
2405                         if (!test_bit(R5_OVERWRITE, &dev->flags)
2406                             && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2407                             && !test_bit(R5_LOCKED, &dev->flags) &&
2408                             !test_bit(R5_UPTODATE, &dev->flags) &&
2409                             test_bit(R5_Insync, &dev->flags)) {
2410                                 if (
2411                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2412                                         pr_debug("Read_old stripe %llu "
2413                                                 "block %d for Reconstruct\n",
2414                                              (unsigned long long)sh->sector, i);
2415                                         set_bit(R5_LOCKED, &dev->flags);
2416                                         set_bit(R5_Wantread, &dev->flags);
2417                                         s->locked++;
2418                                 } else {
2419                                         pr_debug("Request delayed stripe %llu "
2420                                                 "block %d for Reconstruct\n",
2421                                              (unsigned long long)sh->sector, i);
2422                                         set_bit(STRIPE_DELAYED, &sh->state);
2423                                         set_bit(STRIPE_HANDLE, &sh->state);
2424                                 }
2425                         }
2426                 }
2427         /* now if nothing is locked, and if we have enough data, we can start a
2428          * write request
2429          */
2430         if (s->locked == 0 && rcw == 0 &&
2431             !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2432                 if (must_compute > 0) {
2433                         /* We have failed blocks and need to compute them */
2434                         switch (s->failed) {
2435                         case 0:
2436                                 BUG();
2437                         case 1:
2438                                 compute_block_1(sh, r6s->failed_num[0], 0);
2439                                 break;
2440                         case 2:
2441                                 compute_block_2(sh, r6s->failed_num[0],
2442                                                 r6s->failed_num[1]);
2443                                 break;
2444                         default: /* This request should have been failed? */
2445                                 BUG();
2446                         }
2447                 }
2448
2449                 pr_debug("Computing parity for stripe %llu\n",
2450                         (unsigned long long)sh->sector);
2451                 compute_parity6(sh, RECONSTRUCT_WRITE);
2452                 /* now every locked buffer is ready to be written */
2453                 for (i = disks; i--; )
2454                         if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2455                                 pr_debug("Writing stripe %llu block %d\n",
2456                                        (unsigned long long)sh->sector, i);
2457                                 s->locked++;
2458                                 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2459                         }
2460                 if (s->locked == disks)
2461                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2462                                 atomic_inc(&conf->pending_full_writes);
2463                 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2464                 set_bit(STRIPE_INSYNC, &sh->state);
2465
2466                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2467                         atomic_dec(&conf->preread_active_stripes);
2468                         if (atomic_read(&conf->preread_active_stripes) <
2469                             IO_THRESHOLD)
2470                                 md_wakeup_thread(conf->mddev->thread);
2471                 }
2472         }
2473 }
2474
2475 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2476                                 struct stripe_head_state *s, int disks)
2477 {
2478         struct r5dev *dev = NULL;
2479
2480         set_bit(STRIPE_HANDLE, &sh->state);
2481
2482         switch (sh->check_state) {
2483         case check_state_idle:
2484                 /* start a new check operation if there are no failures */
2485                 if (s->failed == 0) {
2486                         BUG_ON(s->uptodate != disks);
2487                         sh->check_state = check_state_run;
2488                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
2489                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2490                         s->uptodate--;
2491                         break;
2492                 }
2493                 dev = &sh->dev[s->failed_num];
2494                 /* fall through */
2495         case check_state_compute_result:
2496                 sh->check_state = check_state_idle;
2497                 if (!dev)
2498                         dev = &sh->dev[sh->pd_idx];
2499
2500                 /* check that a write has not made the stripe insync */
2501                 if (test_bit(STRIPE_INSYNC, &sh->state))
2502                         break;
2503
2504                 /* either failed parity check, or recovery is happening */
2505                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2506                 BUG_ON(s->uptodate != disks);
2507
2508                 set_bit(R5_LOCKED, &dev->flags);
2509                 s->locked++;
2510                 set_bit(R5_Wantwrite, &dev->flags);
2511
2512                 clear_bit(STRIPE_DEGRADED, &sh->state);
2513                 set_bit(STRIPE_INSYNC, &sh->state);
2514                 break;
2515         case check_state_run:
2516                 break; /* we will be called again upon completion */
2517         case check_state_check_result:
2518                 sh->check_state = check_state_idle;
2519
2520                 /* if a failure occurred during the check operation, leave
2521                  * STRIPE_INSYNC not set and let the stripe be handled again
2522                  */
2523                 if (s->failed)
2524                         break;
2525
2526                 /* handle a successful check operation, if parity is correct
2527                  * we are done.  Otherwise update the mismatch count and repair
2528                  * parity if !MD_RECOVERY_CHECK
2529                  */<