md: write bitmap information to devices that are undergoing recovery.
[linux-2.6] / 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/kthread.h>
47 #include "raid6.h"
48
49 #include <linux/raid/bitmap.h>
50 #include <linux/async_tx.h>
51
52 /*
53  * Stripe cache
54  */
55
56 #define NR_STRIPES              256
57 #define STRIPE_SIZE             PAGE_SIZE
58 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
59 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
60 #define IO_THRESHOLD            1
61 #define BYPASS_THRESHOLD        1
62 #define NR_HASH                 (PAGE_SIZE / sizeof(struct hlist_head))
63 #define HASH_MASK               (NR_HASH - 1)
64
65 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
66
67 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
68  * order without overlap.  There may be several bio's per stripe+device, and
69  * a bio could span several devices.
70  * When walking this list for a particular stripe+device, we must never proceed
71  * beyond a bio that extends past this device, as the next bio might no longer
72  * be valid.
73  * This macro is used to determine the 'next' bio in the list, given the sector
74  * of the current stripe+device
75  */
76 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
77 /*
78  * The following can be used to debug the driver
79  */
80 #define RAID5_PARANOIA  1
81 #if RAID5_PARANOIA && defined(CONFIG_SMP)
82 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
83 #else
84 # define CHECK_DEVLOCK()
85 #endif
86
87 #ifdef DEBUG
88 #define inline
89 #define __inline__
90 #endif
91
92 #define printk_rl(args...) ((void) (printk_ratelimit() && printk(args)))
93
94 #if !RAID6_USE_EMPTY_ZERO_PAGE
95 /* In .bss so it's zeroed */
96 const char raid6_empty_zero_page[PAGE_SIZE] __attribute__((aligned(256)));
97 #endif
98
99 /*
100  * We maintain a biased count of active stripes in the bottom 16 bits of
101  * bi_phys_segments, and a count of processed stripes in the upper 16 bits
102  */
103 static inline int raid5_bi_phys_segments(struct bio *bio)
104 {
105         return bio->bi_phys_segments & 0xffff;
106 }
107
108 static inline int raid5_bi_hw_segments(struct bio *bio)
109 {
110         return (bio->bi_phys_segments >> 16) & 0xffff;
111 }
112
113 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
114 {
115         --bio->bi_phys_segments;
116         return raid5_bi_phys_segments(bio);
117 }
118
119 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
120 {
121         unsigned short val = raid5_bi_hw_segments(bio);
122
123         --val;
124         bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
125         return val;
126 }
127
128 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
129 {
130         bio->bi_phys_segments = raid5_bi_phys_segments(bio) || (cnt << 16);
131 }
132
133 static inline int raid6_next_disk(int disk, int raid_disks)
134 {
135         disk++;
136         return (disk < raid_disks) ? disk : 0;
137 }
138
139 static void return_io(struct bio *return_bi)
140 {
141         struct bio *bi = return_bi;
142         while (bi) {
143
144                 return_bi = bi->bi_next;
145                 bi->bi_next = NULL;
146                 bi->bi_size = 0;
147                 bio_endio(bi, 0);
148                 bi = return_bi;
149         }
150 }
151
152 static void print_raid5_conf (raid5_conf_t *conf);
153
154 static int stripe_operations_active(struct stripe_head *sh)
155 {
156         return sh->check_state || sh->reconstruct_state ||
157                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
158                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
159 }
160
161 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
162 {
163         if (atomic_dec_and_test(&sh->count)) {
164                 BUG_ON(!list_empty(&sh->lru));
165                 BUG_ON(atomic_read(&conf->active_stripes)==0);
166                 if (test_bit(STRIPE_HANDLE, &sh->state)) {
167                         if (test_bit(STRIPE_DELAYED, &sh->state)) {
168                                 list_add_tail(&sh->lru, &conf->delayed_list);
169                                 blk_plug_device(conf->mddev->queue);
170                         } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
171                                    sh->bm_seq - conf->seq_write > 0) {
172                                 list_add_tail(&sh->lru, &conf->bitmap_list);
173                                 blk_plug_device(conf->mddev->queue);
174                         } else {
175                                 clear_bit(STRIPE_BIT_DELAY, &sh->state);
176                                 list_add_tail(&sh->lru, &conf->handle_list);
177                         }
178                         md_wakeup_thread(conf->mddev->thread);
179                 } else {
180                         BUG_ON(stripe_operations_active(sh));
181                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
182                                 atomic_dec(&conf->preread_active_stripes);
183                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
184                                         md_wakeup_thread(conf->mddev->thread);
185                         }
186                         atomic_dec(&conf->active_stripes);
187                         if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
188                                 list_add_tail(&sh->lru, &conf->inactive_list);
189                                 wake_up(&conf->wait_for_stripe);
190                                 if (conf->retry_read_aligned)
191                                         md_wakeup_thread(conf->mddev->thread);
192                         }
193                 }
194         }
195 }
196 static void release_stripe(struct stripe_head *sh)
197 {
198         raid5_conf_t *conf = sh->raid_conf;
199         unsigned long flags;
200
201         spin_lock_irqsave(&conf->device_lock, flags);
202         __release_stripe(conf, sh);
203         spin_unlock_irqrestore(&conf->device_lock, flags);
204 }
205
206 static inline void remove_hash(struct stripe_head *sh)
207 {
208         pr_debug("remove_hash(), stripe %llu\n",
209                 (unsigned long long)sh->sector);
210
211         hlist_del_init(&sh->hash);
212 }
213
214 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
215 {
216         struct hlist_head *hp = stripe_hash(conf, sh->sector);
217
218         pr_debug("insert_hash(), stripe %llu\n",
219                 (unsigned long long)sh->sector);
220
221         CHECK_DEVLOCK();
222         hlist_add_head(&sh->hash, hp);
223 }
224
225
226 /* find an idle stripe, make sure it is unhashed, and return it. */
227 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
228 {
229         struct stripe_head *sh = NULL;
230         struct list_head *first;
231
232         CHECK_DEVLOCK();
233         if (list_empty(&conf->inactive_list))
234                 goto out;
235         first = conf->inactive_list.next;
236         sh = list_entry(first, struct stripe_head, lru);
237         list_del_init(first);
238         remove_hash(sh);
239         atomic_inc(&conf->active_stripes);
240 out:
241         return sh;
242 }
243
244 static void shrink_buffers(struct stripe_head *sh, int num)
245 {
246         struct page *p;
247         int i;
248
249         for (i=0; i<num ; i++) {
250                 p = sh->dev[i].page;
251                 if (!p)
252                         continue;
253                 sh->dev[i].page = NULL;
254                 put_page(p);
255         }
256 }
257
258 static int grow_buffers(struct stripe_head *sh, int num)
259 {
260         int i;
261
262         for (i=0; i<num; i++) {
263                 struct page *page;
264
265                 if (!(page = alloc_page(GFP_KERNEL))) {
266                         return 1;
267                 }
268                 sh->dev[i].page = page;
269         }
270         return 0;
271 }
272
273 static void raid5_build_block(struct stripe_head *sh, int i);
274
275 static void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx, int disks)
276 {
277         raid5_conf_t *conf = sh->raid_conf;
278         int i;
279
280         BUG_ON(atomic_read(&sh->count) != 0);
281         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
282         BUG_ON(stripe_operations_active(sh));
283
284         CHECK_DEVLOCK();
285         pr_debug("init_stripe called, stripe %llu\n",
286                 (unsigned long long)sh->sector);
287
288         remove_hash(sh);
289
290         sh->sector = sector;
291         sh->pd_idx = pd_idx;
292         sh->state = 0;
293
294         sh->disks = disks;
295
296         for (i = sh->disks; i--; ) {
297                 struct r5dev *dev = &sh->dev[i];
298
299                 if (dev->toread || dev->read || dev->towrite || dev->written ||
300                     test_bit(R5_LOCKED, &dev->flags)) {
301                         printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
302                                (unsigned long long)sh->sector, i, dev->toread,
303                                dev->read, dev->towrite, dev->written,
304                                test_bit(R5_LOCKED, &dev->flags));
305                         BUG();
306                 }
307                 dev->flags = 0;
308                 raid5_build_block(sh, i);
309         }
310         insert_hash(conf, sh);
311 }
312
313 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector, int disks)
314 {
315         struct stripe_head *sh;
316         struct hlist_node *hn;
317
318         CHECK_DEVLOCK();
319         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
320         hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
321                 if (sh->sector == sector && sh->disks == disks)
322                         return sh;
323         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
324         return NULL;
325 }
326
327 static void unplug_slaves(mddev_t *mddev);
328 static void raid5_unplug_device(struct request_queue *q);
329
330 static struct stripe_head *get_active_stripe(raid5_conf_t *conf, sector_t sector, int disks,
331                                              int pd_idx, int noblock)
332 {
333         struct stripe_head *sh;
334
335         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
336
337         spin_lock_irq(&conf->device_lock);
338
339         do {
340                 wait_event_lock_irq(conf->wait_for_stripe,
341                                     conf->quiesce == 0,
342                                     conf->device_lock, /* nothing */);
343                 sh = __find_stripe(conf, sector, disks);
344                 if (!sh) {
345                         if (!conf->inactive_blocked)
346                                 sh = get_free_stripe(conf);
347                         if (noblock && sh == NULL)
348                                 break;
349                         if (!sh) {
350                                 conf->inactive_blocked = 1;
351                                 wait_event_lock_irq(conf->wait_for_stripe,
352                                                     !list_empty(&conf->inactive_list) &&
353                                                     (atomic_read(&conf->active_stripes)
354                                                      < (conf->max_nr_stripes *3/4)
355                                                      || !conf->inactive_blocked),
356                                                     conf->device_lock,
357                                                     raid5_unplug_device(conf->mddev->queue)
358                                         );
359                                 conf->inactive_blocked = 0;
360                         } else
361                                 init_stripe(sh, sector, pd_idx, disks);
362                 } else {
363                         if (atomic_read(&sh->count)) {
364                           BUG_ON(!list_empty(&sh->lru));
365                         } else {
366                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
367                                         atomic_inc(&conf->active_stripes);
368                                 if (list_empty(&sh->lru) &&
369                                     !test_bit(STRIPE_EXPANDING, &sh->state))
370                                         BUG();
371                                 list_del_init(&sh->lru);
372                         }
373                 }
374         } while (sh == NULL);
375
376         if (sh)
377                 atomic_inc(&sh->count);
378
379         spin_unlock_irq(&conf->device_lock);
380         return sh;
381 }
382
383 static void
384 raid5_end_read_request(struct bio *bi, int error);
385 static void
386 raid5_end_write_request(struct bio *bi, int error);
387
388 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
389 {
390         raid5_conf_t *conf = sh->raid_conf;
391         int i, disks = sh->disks;
392
393         might_sleep();
394
395         for (i = disks; i--; ) {
396                 int rw;
397                 struct bio *bi;
398                 mdk_rdev_t *rdev;
399                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
400                         rw = WRITE;
401                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
402                         rw = READ;
403                 else
404                         continue;
405
406                 bi = &sh->dev[i].req;
407
408                 bi->bi_rw = rw;
409                 if (rw == WRITE)
410                         bi->bi_end_io = raid5_end_write_request;
411                 else
412                         bi->bi_end_io = raid5_end_read_request;
413
414                 rcu_read_lock();
415                 rdev = rcu_dereference(conf->disks[i].rdev);
416                 if (rdev && test_bit(Faulty, &rdev->flags))
417                         rdev = NULL;
418                 if (rdev)
419                         atomic_inc(&rdev->nr_pending);
420                 rcu_read_unlock();
421
422                 if (rdev) {
423                         if (s->syncing || s->expanding || s->expanded)
424                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
425
426                         set_bit(STRIPE_IO_STARTED, &sh->state);
427
428                         bi->bi_bdev = rdev->bdev;
429                         pr_debug("%s: for %llu schedule op %ld on disc %d\n",
430                                 __func__, (unsigned long long)sh->sector,
431                                 bi->bi_rw, i);
432                         atomic_inc(&sh->count);
433                         bi->bi_sector = sh->sector + rdev->data_offset;
434                         bi->bi_flags = 1 << BIO_UPTODATE;
435                         bi->bi_vcnt = 1;
436                         bi->bi_max_vecs = 1;
437                         bi->bi_idx = 0;
438                         bi->bi_io_vec = &sh->dev[i].vec;
439                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
440                         bi->bi_io_vec[0].bv_offset = 0;
441                         bi->bi_size = STRIPE_SIZE;
442                         bi->bi_next = NULL;
443                         if (rw == WRITE &&
444                             test_bit(R5_ReWrite, &sh->dev[i].flags))
445                                 atomic_add(STRIPE_SECTORS,
446                                         &rdev->corrected_errors);
447                         generic_make_request(bi);
448                 } else {
449                         if (rw == WRITE)
450                                 set_bit(STRIPE_DEGRADED, &sh->state);
451                         pr_debug("skip op %ld on disc %d for sector %llu\n",
452                                 bi->bi_rw, i, (unsigned long long)sh->sector);
453                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
454                         set_bit(STRIPE_HANDLE, &sh->state);
455                 }
456         }
457 }
458
459 static struct dma_async_tx_descriptor *
460 async_copy_data(int frombio, struct bio *bio, struct page *page,
461         sector_t sector, struct dma_async_tx_descriptor *tx)
462 {
463         struct bio_vec *bvl;
464         struct page *bio_page;
465         int i;
466         int page_offset;
467
468         if (bio->bi_sector >= sector)
469                 page_offset = (signed)(bio->bi_sector - sector) * 512;
470         else
471                 page_offset = (signed)(sector - bio->bi_sector) * -512;
472         bio_for_each_segment(bvl, bio, i) {
473                 int len = bio_iovec_idx(bio, i)->bv_len;
474                 int clen;
475                 int b_offset = 0;
476
477                 if (page_offset < 0) {
478                         b_offset = -page_offset;
479                         page_offset += b_offset;
480                         len -= b_offset;
481                 }
482
483                 if (len > 0 && page_offset + len > STRIPE_SIZE)
484                         clen = STRIPE_SIZE - page_offset;
485                 else
486                         clen = len;
487
488                 if (clen > 0) {
489                         b_offset += bio_iovec_idx(bio, i)->bv_offset;
490                         bio_page = bio_iovec_idx(bio, i)->bv_page;
491                         if (frombio)
492                                 tx = async_memcpy(page, bio_page, page_offset,
493                                         b_offset, clen,
494                                         ASYNC_TX_DEP_ACK,
495                                         tx, NULL, NULL);
496                         else
497                                 tx = async_memcpy(bio_page, page, b_offset,
498                                         page_offset, clen,
499                                         ASYNC_TX_DEP_ACK,
500                                         tx, NULL, NULL);
501                 }
502                 if (clen < len) /* hit end of page */
503                         break;
504                 page_offset +=  len;
505         }
506
507         return tx;
508 }
509
510 static void ops_complete_biofill(void *stripe_head_ref)
511 {
512         struct stripe_head *sh = stripe_head_ref;
513         struct bio *return_bi = NULL;
514         raid5_conf_t *conf = sh->raid_conf;
515         int i;
516
517         pr_debug("%s: stripe %llu\n", __func__,
518                 (unsigned long long)sh->sector);
519
520         /* clear completed biofills */
521         spin_lock_irq(&conf->device_lock);
522         for (i = sh->disks; i--; ) {
523                 struct r5dev *dev = &sh->dev[i];
524
525                 /* acknowledge completion of a biofill operation */
526                 /* and check if we need to reply to a read request,
527                  * new R5_Wantfill requests are held off until
528                  * !STRIPE_BIOFILL_RUN
529                  */
530                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
531                         struct bio *rbi, *rbi2;
532
533                         BUG_ON(!dev->read);
534                         rbi = dev->read;
535                         dev->read = NULL;
536                         while (rbi && rbi->bi_sector <
537                                 dev->sector + STRIPE_SECTORS) {
538                                 rbi2 = r5_next_bio(rbi, dev->sector);
539                                 if (!raid5_dec_bi_phys_segments(rbi)) {
540                                         rbi->bi_next = return_bi;
541                                         return_bi = rbi;
542                                 }
543                                 rbi = rbi2;
544                         }
545                 }
546         }
547         spin_unlock_irq(&conf->device_lock);
548         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
549
550         return_io(return_bi);
551
552         set_bit(STRIPE_HANDLE, &sh->state);
553         release_stripe(sh);
554 }
555
556 static void ops_run_biofill(struct stripe_head *sh)
557 {
558         struct dma_async_tx_descriptor *tx = NULL;
559         raid5_conf_t *conf = sh->raid_conf;
560         int i;
561
562         pr_debug("%s: stripe %llu\n", __func__,
563                 (unsigned long long)sh->sector);
564
565         for (i = sh->disks; i--; ) {
566                 struct r5dev *dev = &sh->dev[i];
567                 if (test_bit(R5_Wantfill, &dev->flags)) {
568                         struct bio *rbi;
569                         spin_lock_irq(&conf->device_lock);
570                         dev->read = rbi = dev->toread;
571                         dev->toread = NULL;
572                         spin_unlock_irq(&conf->device_lock);
573                         while (rbi && rbi->bi_sector <
574                                 dev->sector + STRIPE_SECTORS) {
575                                 tx = async_copy_data(0, rbi, dev->page,
576                                         dev->sector, tx);
577                                 rbi = r5_next_bio(rbi, dev->sector);
578                         }
579                 }
580         }
581
582         atomic_inc(&sh->count);
583         async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
584                 ops_complete_biofill, sh);
585 }
586
587 static void ops_complete_compute5(void *stripe_head_ref)
588 {
589         struct stripe_head *sh = stripe_head_ref;
590         int target = sh->ops.target;
591         struct r5dev *tgt = &sh->dev[target];
592
593         pr_debug("%s: stripe %llu\n", __func__,
594                 (unsigned long long)sh->sector);
595
596         set_bit(R5_UPTODATE, &tgt->flags);
597         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
598         clear_bit(R5_Wantcompute, &tgt->flags);
599         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
600         if (sh->check_state == check_state_compute_run)
601                 sh->check_state = check_state_compute_result;
602         set_bit(STRIPE_HANDLE, &sh->state);
603         release_stripe(sh);
604 }
605
606 static struct dma_async_tx_descriptor *ops_run_compute5(struct stripe_head *sh)
607 {
608         /* kernel stack size limits the total number of disks */
609         int disks = sh->disks;
610         struct page *xor_srcs[disks];
611         int target = sh->ops.target;
612         struct r5dev *tgt = &sh->dev[target];
613         struct page *xor_dest = tgt->page;
614         int count = 0;
615         struct dma_async_tx_descriptor *tx;
616         int i;
617
618         pr_debug("%s: stripe %llu block: %d\n",
619                 __func__, (unsigned long long)sh->sector, target);
620         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
621
622         for (i = disks; i--; )
623                 if (i != target)
624                         xor_srcs[count++] = sh->dev[i].page;
625
626         atomic_inc(&sh->count);
627
628         if (unlikely(count == 1))
629                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
630                         0, NULL, ops_complete_compute5, sh);
631         else
632                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
633                         ASYNC_TX_XOR_ZERO_DST, NULL,
634                         ops_complete_compute5, sh);
635
636         return tx;
637 }
638
639 static void ops_complete_prexor(void *stripe_head_ref)
640 {
641         struct stripe_head *sh = stripe_head_ref;
642
643         pr_debug("%s: stripe %llu\n", __func__,
644                 (unsigned long long)sh->sector);
645 }
646
647 static struct dma_async_tx_descriptor *
648 ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
649 {
650         /* kernel stack size limits the total number of disks */
651         int disks = sh->disks;
652         struct page *xor_srcs[disks];
653         int count = 0, pd_idx = sh->pd_idx, i;
654
655         /* existing parity data subtracted */
656         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
657
658         pr_debug("%s: stripe %llu\n", __func__,
659                 (unsigned long long)sh->sector);
660
661         for (i = disks; i--; ) {
662                 struct r5dev *dev = &sh->dev[i];
663                 /* Only process blocks that are known to be uptodate */
664                 if (test_bit(R5_Wantdrain, &dev->flags))
665                         xor_srcs[count++] = dev->page;
666         }
667
668         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
669                 ASYNC_TX_DEP_ACK | ASYNC_TX_XOR_DROP_DST, tx,
670                 ops_complete_prexor, sh);
671
672         return tx;
673 }
674
675 static struct dma_async_tx_descriptor *
676 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
677 {
678         int disks = sh->disks;
679         int i;
680
681         pr_debug("%s: stripe %llu\n", __func__,
682                 (unsigned long long)sh->sector);
683
684         for (i = disks; i--; ) {
685                 struct r5dev *dev = &sh->dev[i];
686                 struct bio *chosen;
687
688                 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
689                         struct bio *wbi;
690
691                         spin_lock(&sh->lock);
692                         chosen = dev->towrite;
693                         dev->towrite = NULL;
694                         BUG_ON(dev->written);
695                         wbi = dev->written = chosen;
696                         spin_unlock(&sh->lock);
697
698                         while (wbi && wbi->bi_sector <
699                                 dev->sector + STRIPE_SECTORS) {
700                                 tx = async_copy_data(1, wbi, dev->page,
701                                         dev->sector, tx);
702                                 wbi = r5_next_bio(wbi, dev->sector);
703                         }
704                 }
705         }
706
707         return tx;
708 }
709
710 static void ops_complete_postxor(void *stripe_head_ref)
711 {
712         struct stripe_head *sh = stripe_head_ref;
713         int disks = sh->disks, i, pd_idx = sh->pd_idx;
714
715         pr_debug("%s: stripe %llu\n", __func__,
716                 (unsigned long long)sh->sector);
717
718         for (i = disks; i--; ) {
719                 struct r5dev *dev = &sh->dev[i];
720                 if (dev->written || i == pd_idx)
721                         set_bit(R5_UPTODATE, &dev->flags);
722         }
723
724         if (sh->reconstruct_state == reconstruct_state_drain_run)
725                 sh->reconstruct_state = reconstruct_state_drain_result;
726         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
727                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
728         else {
729                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
730                 sh->reconstruct_state = reconstruct_state_result;
731         }
732
733         set_bit(STRIPE_HANDLE, &sh->state);
734         release_stripe(sh);
735 }
736
737 static void
738 ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
739 {
740         /* kernel stack size limits the total number of disks */
741         int disks = sh->disks;
742         struct page *xor_srcs[disks];
743
744         int count = 0, pd_idx = sh->pd_idx, i;
745         struct page *xor_dest;
746         int prexor = 0;
747         unsigned long flags;
748
749         pr_debug("%s: stripe %llu\n", __func__,
750                 (unsigned long long)sh->sector);
751
752         /* check if prexor is active which means only process blocks
753          * that are part of a read-modify-write (written)
754          */
755         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
756                 prexor = 1;
757                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
758                 for (i = disks; i--; ) {
759                         struct r5dev *dev = &sh->dev[i];
760                         if (dev->written)
761                                 xor_srcs[count++] = dev->page;
762                 }
763         } else {
764                 xor_dest = sh->dev[pd_idx].page;
765                 for (i = disks; i--; ) {
766                         struct r5dev *dev = &sh->dev[i];
767                         if (i != pd_idx)
768                                 xor_srcs[count++] = dev->page;
769                 }
770         }
771
772         /* 1/ if we prexor'd then the dest is reused as a source
773          * 2/ if we did not prexor then we are redoing the parity
774          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
775          * for the synchronous xor case
776          */
777         flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK |
778                 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
779
780         atomic_inc(&sh->count);
781
782         if (unlikely(count == 1)) {
783                 flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST);
784                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
785                         flags, tx, ops_complete_postxor, sh);
786         } else
787                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
788                         flags, tx, ops_complete_postxor, sh);
789 }
790
791 static void ops_complete_check(void *stripe_head_ref)
792 {
793         struct stripe_head *sh = stripe_head_ref;
794
795         pr_debug("%s: stripe %llu\n", __func__,
796                 (unsigned long long)sh->sector);
797
798         sh->check_state = check_state_check_result;
799         set_bit(STRIPE_HANDLE, &sh->state);
800         release_stripe(sh);
801 }
802
803 static void ops_run_check(struct stripe_head *sh)
804 {
805         /* kernel stack size limits the total number of disks */
806         int disks = sh->disks;
807         struct page *xor_srcs[disks];
808         struct dma_async_tx_descriptor *tx;
809
810         int count = 0, pd_idx = sh->pd_idx, i;
811         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
812
813         pr_debug("%s: stripe %llu\n", __func__,
814                 (unsigned long long)sh->sector);
815
816         for (i = disks; i--; ) {
817                 struct r5dev *dev = &sh->dev[i];
818                 if (i != pd_idx)
819                         xor_srcs[count++] = dev->page;
820         }
821
822         tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
823                 &sh->ops.zero_sum_result, 0, NULL, NULL, NULL);
824
825         atomic_inc(&sh->count);
826         tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
827                 ops_complete_check, sh);
828 }
829
830 static void raid5_run_ops(struct stripe_head *sh, unsigned long ops_request)
831 {
832         int overlap_clear = 0, i, disks = sh->disks;
833         struct dma_async_tx_descriptor *tx = NULL;
834
835         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
836                 ops_run_biofill(sh);
837                 overlap_clear++;
838         }
839
840         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
841                 tx = ops_run_compute5(sh);
842                 /* terminate the chain if postxor is not set to be run */
843                 if (tx && !test_bit(STRIPE_OP_POSTXOR, &ops_request))
844                         async_tx_ack(tx);
845         }
846
847         if (test_bit(STRIPE_OP_PREXOR, &ops_request))
848                 tx = ops_run_prexor(sh, tx);
849
850         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
851                 tx = ops_run_biodrain(sh, tx);
852                 overlap_clear++;
853         }
854
855         if (test_bit(STRIPE_OP_POSTXOR, &ops_request))
856                 ops_run_postxor(sh, tx);
857
858         if (test_bit(STRIPE_OP_CHECK, &ops_request))
859                 ops_run_check(sh);
860
861         if (overlap_clear)
862                 for (i = disks; i--; ) {
863                         struct r5dev *dev = &sh->dev[i];
864                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
865                                 wake_up(&sh->raid_conf->wait_for_overlap);
866                 }
867 }
868
869 static int grow_one_stripe(raid5_conf_t *conf)
870 {
871         struct stripe_head *sh;
872         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
873         if (!sh)
874                 return 0;
875         memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
876         sh->raid_conf = conf;
877         spin_lock_init(&sh->lock);
878
879         if (grow_buffers(sh, conf->raid_disks)) {
880                 shrink_buffers(sh, conf->raid_disks);
881                 kmem_cache_free(conf->slab_cache, sh);
882                 return 0;
883         }
884         sh->disks = conf->raid_disks;
885         /* we just created an active stripe so... */
886         atomic_set(&sh->count, 1);
887         atomic_inc(&conf->active_stripes);
888         INIT_LIST_HEAD(&sh->lru);
889         release_stripe(sh);
890         return 1;
891 }
892
893 static int grow_stripes(raid5_conf_t *conf, int num)
894 {
895         struct kmem_cache *sc;
896         int devs = conf->raid_disks;
897
898         sprintf(conf->cache_name[0], "raid5-%s", mdname(conf->mddev));
899         sprintf(conf->cache_name[1], "raid5-%s-alt", mdname(conf->mddev));
900         conf->active_name = 0;
901         sc = kmem_cache_create(conf->cache_name[conf->active_name],
902                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
903                                0, 0, NULL);
904         if (!sc)
905                 return 1;
906         conf->slab_cache = sc;
907         conf->pool_size = devs;
908         while (num--)
909                 if (!grow_one_stripe(conf))
910                         return 1;
911         return 0;
912 }
913
914 #ifdef CONFIG_MD_RAID5_RESHAPE
915 static int resize_stripes(raid5_conf_t *conf, int newsize)
916 {
917         /* Make all the stripes able to hold 'newsize' devices.
918          * New slots in each stripe get 'page' set to a new page.
919          *
920          * This happens in stages:
921          * 1/ create a new kmem_cache and allocate the required number of
922          *    stripe_heads.
923          * 2/ gather all the old stripe_heads and tranfer the pages across
924          *    to the new stripe_heads.  This will have the side effect of
925          *    freezing the array as once all stripe_heads have been collected,
926          *    no IO will be possible.  Old stripe heads are freed once their
927          *    pages have been transferred over, and the old kmem_cache is
928          *    freed when all stripes are done.
929          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
930          *    we simple return a failre status - no need to clean anything up.
931          * 4/ allocate new pages for the new slots in the new stripe_heads.
932          *    If this fails, we don't bother trying the shrink the
933          *    stripe_heads down again, we just leave them as they are.
934          *    As each stripe_head is processed the new one is released into
935          *    active service.
936          *
937          * Once step2 is started, we cannot afford to wait for a write,
938          * so we use GFP_NOIO allocations.
939          */
940         struct stripe_head *osh, *nsh;
941         LIST_HEAD(newstripes);
942         struct disk_info *ndisks;
943         int err;
944         struct kmem_cache *sc;
945         int i;
946
947         if (newsize <= conf->pool_size)
948                 return 0; /* never bother to shrink */
949
950         err = md_allow_write(conf->mddev);
951         if (err)
952                 return err;
953
954         /* Step 1 */
955         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
956                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
957                                0, 0, NULL);
958         if (!sc)
959                 return -ENOMEM;
960
961         for (i = conf->max_nr_stripes; i; i--) {
962                 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
963                 if (!nsh)
964                         break;
965
966                 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
967
968                 nsh->raid_conf = conf;
969                 spin_lock_init(&nsh->lock);
970
971                 list_add(&nsh->lru, &newstripes);
972         }
973         if (i) {
974                 /* didn't get enough, give up */
975                 while (!list_empty(&newstripes)) {
976                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
977                         list_del(&nsh->lru);
978                         kmem_cache_free(sc, nsh);
979                 }
980                 kmem_cache_destroy(sc);
981                 return -ENOMEM;
982         }
983         /* Step 2 - Must use GFP_NOIO now.
984          * OK, we have enough stripes, start collecting inactive
985          * stripes and copying them over
986          */
987         list_for_each_entry(nsh, &newstripes, lru) {
988                 spin_lock_irq(&conf->device_lock);
989                 wait_event_lock_irq(conf->wait_for_stripe,
990                                     !list_empty(&conf->inactive_list),
991                                     conf->device_lock,
992                                     unplug_slaves(conf->mddev)
993                         );
994                 osh = get_free_stripe(conf);
995                 spin_unlock_irq(&conf->device_lock);
996                 atomic_set(&nsh->count, 1);
997                 for(i=0; i<conf->pool_size; i++)
998                         nsh->dev[i].page = osh->dev[i].page;
999                 for( ; i<newsize; i++)
1000                         nsh->dev[i].page = NULL;
1001                 kmem_cache_free(conf->slab_cache, osh);
1002         }
1003         kmem_cache_destroy(conf->slab_cache);
1004
1005         /* Step 3.
1006          * At this point, we are holding all the stripes so the array
1007          * is completely stalled, so now is a good time to resize
1008          * conf->disks.
1009          */
1010         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1011         if (ndisks) {
1012                 for (i=0; i<conf->raid_disks; i++)
1013                         ndisks[i] = conf->disks[i];
1014                 kfree(conf->disks);
1015                 conf->disks = ndisks;
1016         } else
1017                 err = -ENOMEM;
1018
1019         /* Step 4, return new stripes to service */
1020         while(!list_empty(&newstripes)) {
1021                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1022                 list_del_init(&nsh->lru);
1023                 for (i=conf->raid_disks; i < newsize; i++)
1024                         if (nsh->dev[i].page == NULL) {
1025                                 struct page *p = alloc_page(GFP_NOIO);
1026                                 nsh->dev[i].page = p;
1027                                 if (!p)
1028                                         err = -ENOMEM;
1029                         }
1030                 release_stripe(nsh);
1031         }
1032         /* critical section pass, GFP_NOIO no longer needed */
1033
1034         conf->slab_cache = sc;
1035         conf->active_name = 1-conf->active_name;
1036         conf->pool_size = newsize;
1037         return err;
1038 }
1039 #endif
1040
1041 static int drop_one_stripe(raid5_conf_t *conf)
1042 {
1043         struct stripe_head *sh;
1044
1045         spin_lock_irq(&conf->device_lock);
1046         sh = get_free_stripe(conf);
1047         spin_unlock_irq(&conf->device_lock);
1048         if (!sh)
1049                 return 0;
1050         BUG_ON(atomic_read(&sh->count));
1051         shrink_buffers(sh, conf->pool_size);
1052         kmem_cache_free(conf->slab_cache, sh);
1053         atomic_dec(&conf->active_stripes);
1054         return 1;
1055 }
1056
1057 static void shrink_stripes(raid5_conf_t *conf)
1058 {
1059         while (drop_one_stripe(conf))
1060                 ;
1061
1062         if (conf->slab_cache)
1063                 kmem_cache_destroy(conf->slab_cache);
1064         conf->slab_cache = NULL;
1065 }
1066
1067 static void raid5_end_read_request(struct bio * bi, int error)
1068 {
1069         struct stripe_head *sh = bi->bi_private;
1070         raid5_conf_t *conf = sh->raid_conf;
1071         int disks = sh->disks, i;
1072         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1073         char b[BDEVNAME_SIZE];
1074         mdk_rdev_t *rdev;
1075
1076
1077         for (i=0 ; i<disks; i++)
1078                 if (bi == &sh->dev[i].req)
1079                         break;
1080
1081         pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1082                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1083                 uptodate);
1084         if (i == disks) {
1085                 BUG();
1086                 return;
1087         }
1088
1089         if (uptodate) {
1090                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1091                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1092                         rdev = conf->disks[i].rdev;
1093                         printk_rl(KERN_INFO "raid5:%s: read error corrected"
1094                                   " (%lu sectors at %llu on %s)\n",
1095                                   mdname(conf->mddev), STRIPE_SECTORS,
1096                                   (unsigned long long)(sh->sector
1097                                                        + rdev->data_offset),
1098                                   bdevname(rdev->bdev, b));
1099                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1100                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1101                 }
1102                 if (atomic_read(&conf->disks[i].rdev->read_errors))
1103                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
1104         } else {
1105                 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1106                 int retry = 0;
1107                 rdev = conf->disks[i].rdev;
1108
1109                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1110                 atomic_inc(&rdev->read_errors);
1111                 if (conf->mddev->degraded)
1112                         printk_rl(KERN_WARNING
1113                                   "raid5:%s: read error not correctable "
1114                                   "(sector %llu on %s).\n",
1115                                   mdname(conf->mddev),
1116                                   (unsigned long long)(sh->sector
1117                                                        + rdev->data_offset),
1118                                   bdn);
1119                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1120                         /* Oh, no!!! */
1121                         printk_rl(KERN_WARNING
1122                                   "raid5:%s: read error NOT corrected!! "
1123                                   "(sector %llu on %s).\n",
1124                                   mdname(conf->mddev),
1125                                   (unsigned long long)(sh->sector
1126                                                        + rdev->data_offset),
1127                                   bdn);
1128                 else if (atomic_read(&rdev->read_errors)
1129                          > conf->max_nr_stripes)
1130                         printk(KERN_WARNING
1131                                "raid5:%s: Too many read errors, failing device %s.\n",
1132                                mdname(conf->mddev), bdn);
1133                 else
1134                         retry = 1;
1135                 if (retry)
1136                         set_bit(R5_ReadError, &sh->dev[i].flags);
1137                 else {
1138                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1139                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1140                         md_error(conf->mddev, rdev);
1141                 }
1142         }
1143         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1144         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1145         set_bit(STRIPE_HANDLE, &sh->state);
1146         release_stripe(sh);
1147 }
1148
1149 static void raid5_end_write_request(struct bio *bi, int error)
1150 {
1151         struct stripe_head *sh = bi->bi_private;
1152         raid5_conf_t *conf = sh->raid_conf;
1153         int disks = sh->disks, i;
1154         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1155
1156         for (i=0 ; i<disks; i++)
1157                 if (bi == &sh->dev[i].req)
1158                         break;
1159
1160         pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1161                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1162                 uptodate);
1163         if (i == disks) {
1164                 BUG();
1165                 return;
1166         }
1167
1168         if (!uptodate)
1169                 md_error(conf->mddev, conf->disks[i].rdev);
1170
1171         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1172         
1173         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1174         set_bit(STRIPE_HANDLE, &sh->state);
1175         release_stripe(sh);
1176 }
1177
1178
1179 static sector_t compute_blocknr(struct stripe_head *sh, int i);
1180         
1181 static void raid5_build_block(struct stripe_head *sh, int i)
1182 {
1183         struct r5dev *dev = &sh->dev[i];
1184
1185         bio_init(&dev->req);
1186         dev->req.bi_io_vec = &dev->vec;
1187         dev->req.bi_vcnt++;
1188         dev->req.bi_max_vecs++;
1189         dev->vec.bv_page = dev->page;
1190         dev->vec.bv_len = STRIPE_SIZE;
1191         dev->vec.bv_offset = 0;
1192
1193         dev->req.bi_sector = sh->sector;
1194         dev->req.bi_private = sh;
1195
1196         dev->flags = 0;
1197         dev->sector = compute_blocknr(sh, i);
1198 }
1199
1200 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1201 {
1202         char b[BDEVNAME_SIZE];
1203         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1204         pr_debug("raid5: error called\n");
1205
1206         if (!test_bit(Faulty, &rdev->flags)) {
1207                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1208                 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1209                         unsigned long flags;
1210                         spin_lock_irqsave(&conf->device_lock, flags);
1211                         mddev->degraded++;
1212                         spin_unlock_irqrestore(&conf->device_lock, flags);
1213                         /*
1214                          * if recovery was running, make sure it aborts.
1215                          */
1216                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1217                 }
1218                 set_bit(Faulty, &rdev->flags);
1219                 printk(KERN_ALERT
1220                        "raid5: Disk failure on %s, disabling device.\n"
1221                        "raid5: Operation continuing on %d devices.\n",
1222                        bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1223         }
1224 }
1225
1226 /*
1227  * Input: a 'big' sector number,
1228  * Output: index of the data and parity disk, and the sector # in them.
1229  */
1230 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
1231                         unsigned int data_disks, unsigned int * dd_idx,
1232                         unsigned int * pd_idx, raid5_conf_t *conf)
1233 {
1234         long stripe;
1235         unsigned long chunk_number;
1236         unsigned int chunk_offset;
1237         sector_t new_sector;
1238         int sectors_per_chunk = conf->chunk_size >> 9;
1239
1240         /* First compute the information on this sector */
1241
1242         /*
1243          * Compute the chunk number and the sector offset inside the chunk
1244          */
1245         chunk_offset = sector_div(r_sector, sectors_per_chunk);
1246         chunk_number = r_sector;
1247         BUG_ON(r_sector != chunk_number);
1248
1249         /*
1250          * Compute the stripe number
1251          */
1252         stripe = chunk_number / data_disks;
1253
1254         /*
1255          * Compute the data disk and parity disk indexes inside the stripe
1256          */
1257         *dd_idx = chunk_number % data_disks;
1258
1259         /*
1260          * Select the parity disk based on the user selected algorithm.
1261          */
1262         switch(conf->level) {
1263         case 4:
1264                 *pd_idx = data_disks;
1265                 break;
1266         case 5:
1267                 switch (conf->algorithm) {
1268                 case ALGORITHM_LEFT_ASYMMETRIC:
1269                         *pd_idx = data_disks - stripe % raid_disks;
1270                         if (*dd_idx >= *pd_idx)
1271                                 (*dd_idx)++;
1272                         break;
1273                 case ALGORITHM_RIGHT_ASYMMETRIC:
1274                         *pd_idx = stripe % raid_disks;
1275                         if (*dd_idx >= *pd_idx)
1276                                 (*dd_idx)++;
1277                         break;
1278                 case ALGORITHM_LEFT_SYMMETRIC:
1279                         *pd_idx = data_disks - stripe % raid_disks;
1280                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1281                         break;
1282                 case ALGORITHM_RIGHT_SYMMETRIC:
1283                         *pd_idx = stripe % raid_disks;
1284                         *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1285                         break;
1286                 default:
1287                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1288                                 conf->algorithm);
1289                 }
1290                 break;
1291         case 6:
1292
1293                 /**** FIX THIS ****/
1294                 switch (conf->algorithm) {
1295                 case ALGORITHM_LEFT_ASYMMETRIC:
1296                         *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1297                         if (*pd_idx == raid_disks-1)
1298                                 (*dd_idx)++;    /* Q D D D P */
1299                         else if (*dd_idx >= *pd_idx)
1300                                 (*dd_idx) += 2; /* D D P Q D */
1301                         break;
1302                 case ALGORITHM_RIGHT_ASYMMETRIC:
1303                         *pd_idx = stripe % raid_disks;
1304                         if (*pd_idx == raid_disks-1)
1305                                 (*dd_idx)++;    /* Q D D D P */
1306                         else if (*dd_idx >= *pd_idx)
1307                                 (*dd_idx) += 2; /* D D P Q D */
1308                         break;
1309                 case ALGORITHM_LEFT_SYMMETRIC:
1310                         *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1311                         *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1312                         break;
1313                 case ALGORITHM_RIGHT_SYMMETRIC:
1314                         *pd_idx = stripe % raid_disks;
1315                         *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1316                         break;
1317                 default:
1318                         printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1319                                conf->algorithm);
1320                 }
1321                 break;
1322         }
1323
1324         /*
1325          * Finally, compute the new sector number
1326          */
1327         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1328         return new_sector;
1329 }
1330
1331
1332 static sector_t compute_blocknr(struct stripe_head *sh, int i)
1333 {
1334         raid5_conf_t *conf = sh->raid_conf;
1335         int raid_disks = sh->disks;
1336         int data_disks = raid_disks - conf->max_degraded;
1337         sector_t new_sector = sh->sector, check;
1338         int sectors_per_chunk = conf->chunk_size >> 9;
1339         sector_t stripe;
1340         int chunk_offset;
1341         int chunk_number, dummy1, dummy2, dd_idx = i;
1342         sector_t r_sector;
1343
1344
1345         chunk_offset = sector_div(new_sector, sectors_per_chunk);
1346         stripe = new_sector;
1347         BUG_ON(new_sector != stripe);
1348
1349         if (i == sh->pd_idx)
1350                 return 0;
1351         switch(conf->level) {
1352         case 4: break;
1353         case 5:
1354                 switch (conf->algorithm) {
1355                 case ALGORITHM_LEFT_ASYMMETRIC:
1356                 case ALGORITHM_RIGHT_ASYMMETRIC:
1357                         if (i > sh->pd_idx)
1358                                 i--;
1359                         break;
1360                 case ALGORITHM_LEFT_SYMMETRIC:
1361                 case ALGORITHM_RIGHT_SYMMETRIC:
1362                         if (i < sh->pd_idx)
1363                                 i += raid_disks;
1364                         i -= (sh->pd_idx + 1);
1365                         break;
1366                 default:
1367                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1368                                conf->algorithm);
1369                 }
1370                 break;
1371         case 6:
1372                 if (i == raid6_next_disk(sh->pd_idx, raid_disks))
1373                         return 0; /* It is the Q disk */
1374                 switch (conf->algorithm) {
1375                 case ALGORITHM_LEFT_ASYMMETRIC:
1376                 case ALGORITHM_RIGHT_ASYMMETRIC:
1377                         if (sh->pd_idx == raid_disks-1)
1378                                 i--;    /* Q D D D P */
1379                         else if (i > sh->pd_idx)
1380                                 i -= 2; /* D D P Q D */
1381                         break;
1382                 case ALGORITHM_LEFT_SYMMETRIC:
1383                 case ALGORITHM_RIGHT_SYMMETRIC:
1384                         if (sh->pd_idx == raid_disks-1)
1385                                 i--; /* Q D D D P */
1386                         else {
1387                                 /* D D P Q D */
1388                                 if (i < sh->pd_idx)
1389                                         i += raid_disks;
1390                                 i -= (sh->pd_idx + 2);
1391                         }
1392                         break;
1393                 default:
1394                         printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1395                                conf->algorithm);
1396                 }
1397                 break;
1398         }
1399
1400         chunk_number = stripe * data_disks + i;
1401         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1402
1403         check = raid5_compute_sector(r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
1404         if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
1405                 printk(KERN_ERR "compute_blocknr: map not correct\n");
1406                 return 0;
1407         }
1408         return r_sector;
1409 }
1410
1411
1412
1413 /*
1414  * Copy data between a page in the stripe cache, and one or more bion
1415  * The page could align with the middle of the bio, or there could be
1416  * several bion, each with several bio_vecs, which cover part of the page
1417  * Multiple bion are linked together on bi_next.  There may be extras
1418  * at the end of this list.  We ignore them.
1419  */
1420 static void copy_data(int frombio, struct bio *bio,
1421                      struct page *page,
1422                      sector_t sector)
1423 {
1424         char *pa = page_address(page);
1425         struct bio_vec *bvl;
1426         int i;
1427         int page_offset;
1428
1429         if (bio->bi_sector >= sector)
1430                 page_offset = (signed)(bio->bi_sector - sector) * 512;
1431         else
1432                 page_offset = (signed)(sector - bio->bi_sector) * -512;
1433         bio_for_each_segment(bvl, bio, i) {
1434                 int len = bio_iovec_idx(bio,i)->bv_len;
1435                 int clen;
1436                 int b_offset = 0;
1437
1438                 if (page_offset < 0) {
1439                         b_offset = -page_offset;
1440                         page_offset += b_offset;
1441                         len -= b_offset;
1442                 }
1443
1444                 if (len > 0 && page_offset + len > STRIPE_SIZE)
1445                         clen = STRIPE_SIZE - page_offset;
1446                 else clen = len;
1447
1448                 if (clen > 0) {
1449                         char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1450                         if (frombio)
1451                                 memcpy(pa+page_offset, ba+b_offset, clen);
1452                         else
1453                                 memcpy(ba+b_offset, pa+page_offset, clen);
1454                         __bio_kunmap_atomic(ba, KM_USER0);
1455                 }
1456                 if (clen < len) /* hit end of page */
1457                         break;
1458                 page_offset +=  len;
1459         }
1460 }
1461
1462 #define check_xor()     do {                                              \
1463                                 if (count == MAX_XOR_BLOCKS) {            \
1464                                 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1465                                 count = 0;                                \
1466                            }                                              \
1467                         } while(0)
1468
1469 static void compute_parity6(struct stripe_head *sh, int method)
1470 {
1471         raid6_conf_t *conf = sh->raid_conf;
1472         int i, pd_idx = sh->pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1473         struct bio *chosen;
1474         /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1475         void *ptrs[disks];
1476
1477         qd_idx = raid6_next_disk(pd_idx, disks);
1478         d0_idx = raid6_next_disk(qd_idx, disks);
1479
1480         pr_debug("compute_parity, stripe %llu, method %d\n",
1481                 (unsigned long long)sh->sector, method);
1482
1483         switch(method) {
1484         case READ_MODIFY_WRITE:
1485                 BUG();          /* READ_MODIFY_WRITE N/A for RAID-6 */
1486         case RECONSTRUCT_WRITE:
1487                 for (i= disks; i-- ;)
1488                         if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1489                                 chosen = sh->dev[i].towrite;
1490                                 sh->dev[i].towrite = NULL;
1491
1492                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1493                                         wake_up(&conf->wait_for_overlap);
1494
1495                                 BUG_ON(sh->dev[i].written);
1496                                 sh->dev[i].written = chosen;
1497                         }
1498                 break;
1499         case CHECK_PARITY:
1500                 BUG();          /* Not implemented yet */
1501         }
1502
1503         for (i = disks; i--;)
1504                 if (sh->dev[i].written) {
1505                         sector_t sector = sh->dev[i].sector;
1506                         struct bio *wbi = sh->dev[i].written;
1507                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1508                                 copy_data(1, wbi, sh->dev[i].page, sector);
1509                                 wbi = r5_next_bio(wbi, sector);
1510                         }
1511
1512                         set_bit(R5_LOCKED, &sh->dev[i].flags);
1513                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
1514                 }
1515
1516 //      switch(method) {
1517 //      case RECONSTRUCT_WRITE:
1518 //      case CHECK_PARITY:
1519 //      case UPDATE_PARITY:
1520                 /* Note that unlike RAID-5, the ordering of the disks matters greatly. */
1521                 /* FIX: Is this ordering of drives even remotely optimal? */
1522                 count = 0;
1523                 i = d0_idx;
1524                 do {
1525                         ptrs[count++] = page_address(sh->dev[i].page);
1526                         if (count <= disks-2 && !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1527                                 printk("block %d/%d not uptodate on parity calc\n", i,count);
1528                         i = raid6_next_disk(i, disks);
1529                 } while ( i != d0_idx );
1530 //              break;
1531 //      }
1532
1533         raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs);
1534
1535         switch(method) {
1536         case RECONSTRUCT_WRITE:
1537                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1538                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1539                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
1540                 set_bit(R5_LOCKED,   &sh->dev[qd_idx].flags);
1541                 break;
1542         case UPDATE_PARITY:
1543                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1544                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1545                 break;
1546         }
1547 }
1548
1549
1550 /* Compute one missing block */
1551 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1552 {
1553         int i, count, disks = sh->disks;
1554         void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1555         int pd_idx = sh->pd_idx;
1556         int qd_idx = raid6_next_disk(pd_idx, disks);
1557
1558         pr_debug("compute_block_1, stripe %llu, idx %d\n",
1559                 (unsigned long long)sh->sector, dd_idx);
1560
1561         if ( dd_idx == qd_idx ) {
1562                 /* We're actually computing the Q drive */
1563                 compute_parity6(sh, UPDATE_PARITY);
1564         } else {
1565                 dest = page_address(sh->dev[dd_idx].page);
1566                 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1567                 count = 0;
1568                 for (i = disks ; i--; ) {
1569                         if (i == dd_idx || i == qd_idx)
1570                                 continue;
1571                         p = page_address(sh->dev[i].page);
1572                         if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1573                                 ptr[count++] = p;
1574                         else
1575                                 printk("compute_block() %d, stripe %llu, %d"
1576                                        " not present\n", dd_idx,
1577                                        (unsigned long long)sh->sector, i);
1578
1579                         check_xor();
1580                 }
1581                 if (count)
1582                         xor_blocks(count, STRIPE_SIZE, dest, ptr);
1583                 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1584                 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1585         }
1586 }
1587
1588 /* Compute two missing blocks */
1589 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1590 {
1591         int i, count, disks = sh->disks;
1592         int pd_idx = sh->pd_idx;
1593         int qd_idx = raid6_next_disk(pd_idx, disks);
1594         int d0_idx = raid6_next_disk(qd_idx, disks);
1595         int faila, failb;
1596
1597         /* faila and failb are disk numbers relative to d0_idx */
1598         /* pd_idx become disks-2 and qd_idx become disks-1 */
1599         faila = (dd_idx1 < d0_idx) ? dd_idx1+(disks-d0_idx) : dd_idx1-d0_idx;
1600         failb = (dd_idx2 < d0_idx) ? dd_idx2+(disks-d0_idx) : dd_idx2-d0_idx;
1601
1602         BUG_ON(faila == failb);
1603         if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1604
1605         pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1606                (unsigned long long)sh->sector, dd_idx1, dd_idx2, faila, failb);
1607
1608         if ( failb == disks-1 ) {
1609                 /* Q disk is one of the missing disks */
1610                 if ( faila == disks-2 ) {
1611                         /* Missing P+Q, just recompute */
1612                         compute_parity6(sh, UPDATE_PARITY);
1613                         return;
1614                 } else {
1615                         /* We're missing D+Q; recompute D from P */
1616                         compute_block_1(sh, (dd_idx1 == qd_idx) ? dd_idx2 : dd_idx1, 0);
1617                         compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1618                         return;
1619                 }
1620         }
1621
1622         /* We're missing D+P or D+D; build pointer table */
1623         {
1624                 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1625                 void *ptrs[disks];
1626
1627                 count = 0;
1628                 i = d0_idx;
1629                 do {
1630                         ptrs[count++] = page_address(sh->dev[i].page);
1631                         i = raid6_next_disk(i, disks);
1632                         if (i != dd_idx1 && i != dd_idx2 &&
1633                             !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1634                                 printk("compute_2 with missing block %d/%d\n", count, i);
1635                 } while ( i != d0_idx );
1636
1637                 if ( failb == disks-2 ) {
1638                         /* We're missing D+P. */
1639                         raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs);
1640                 } else {
1641                         /* We're missing D+D. */
1642                         raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs);
1643                 }
1644
1645                 /* Both the above update both missing blocks */
1646                 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1647                 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1648         }
1649 }
1650
1651 static void
1652 schedule_reconstruction5(struct stripe_head *sh, struct stripe_head_state *s,
1653                          int rcw, int expand)
1654 {
1655         int i, pd_idx = sh->pd_idx, disks = sh->disks;
1656
1657         if (rcw) {
1658                 /* if we are not expanding this is a proper write request, and
1659                  * there will be bios with new data to be drained into the
1660                  * stripe cache
1661                  */
1662                 if (!expand) {
1663                         sh->reconstruct_state = reconstruct_state_drain_run;
1664                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1665                 } else
1666                         sh->reconstruct_state = reconstruct_state_run;
1667
1668                 set_bit(STRIPE_OP_POSTXOR, &s->ops_request);
1669
1670                 for (i = disks; i--; ) {
1671                         struct r5dev *dev = &sh->dev[i];
1672
1673                         if (dev->towrite) {
1674                                 set_bit(R5_LOCKED, &dev->flags);
1675                                 set_bit(R5_Wantdrain, &dev->flags);
1676                                 if (!expand)
1677                                         clear_bit(R5_UPTODATE, &dev->flags);
1678                                 s->locked++;
1679                         }
1680                 }
1681                 if (s->locked + 1 == disks)
1682                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
1683                                 atomic_inc(&sh->raid_conf->pending_full_writes);
1684         } else {
1685                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1686                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1687
1688                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
1689                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
1690                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1691                 set_bit(STRIPE_OP_POSTXOR, &s->ops_request);
1692
1693                 for (i = disks; i--; ) {
1694                         struct r5dev *dev = &sh->dev[i];
1695                         if (i == pd_idx)
1696                                 continue;
1697
1698                         if (dev->towrite &&
1699                             (test_bit(R5_UPTODATE, &dev->flags) ||
1700                              test_bit(R5_Wantcompute, &dev->flags))) {
1701                                 set_bit(R5_Wantdrain, &dev->flags);
1702                                 set_bit(R5_LOCKED, &dev->flags);
1703                                 clear_bit(R5_UPTODATE, &dev->flags);
1704                                 s->locked++;
1705                         }
1706                 }
1707         }
1708
1709         /* keep the parity disk locked while asynchronous operations
1710          * are in flight
1711          */
1712         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1713         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1714         s->locked++;
1715
1716         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
1717                 __func__, (unsigned long long)sh->sector,
1718                 s->locked, s->ops_request);
1719 }
1720
1721 /*
1722  * Each stripe/dev can have one or more bion attached.
1723  * toread/towrite point to the first in a chain.
1724  * The bi_next chain must be in order.
1725  */
1726 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1727 {
1728         struct bio **bip;
1729         raid5_conf_t *conf = sh->raid_conf;
1730         int firstwrite=0;
1731
1732         pr_debug("adding bh b#%llu to stripe s#%llu\n",
1733                 (unsigned long long)bi->bi_sector,
1734                 (unsigned long long)sh->sector);
1735
1736
1737         spin_lock(&sh->lock);
1738         spin_lock_irq(&conf->device_lock);
1739         if (forwrite) {
1740                 bip = &sh->dev[dd_idx].towrite;
1741                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1742                         firstwrite = 1;
1743         } else
1744                 bip = &sh->dev[dd_idx].toread;
1745         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1746                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1747                         goto overlap;
1748                 bip = & (*bip)->bi_next;
1749         }
1750         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1751                 goto overlap;
1752
1753         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1754         if (*bip)
1755                 bi->bi_next = *bip;
1756         *bip = bi;
1757         bi->bi_phys_segments++;
1758         spin_unlock_irq(&conf->device_lock);
1759         spin_unlock(&sh->lock);
1760
1761         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1762                 (unsigned long long)bi->bi_sector,
1763                 (unsigned long long)sh->sector, dd_idx);
1764
1765         if (conf->mddev->bitmap && firstwrite) {
1766                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1767                                   STRIPE_SECTORS, 0);
1768                 sh->bm_seq = conf->seq_flush+1;
1769                 set_bit(STRIPE_BIT_DELAY, &sh->state);
1770         }
1771
1772         if (forwrite) {
1773                 /* check if page is covered */
1774                 sector_t sector = sh->dev[dd_idx].sector;
1775                 for (bi=sh->dev[dd_idx].towrite;
1776                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1777                              bi && bi->bi_sector <= sector;
1778                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1779                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1780                                 sector = bi->bi_sector + (bi->bi_size>>9);
1781                 }
1782                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1783                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1784         }
1785         return 1;
1786
1787  overlap:
1788         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1789         spin_unlock_irq(&conf->device_lock);
1790         spin_unlock(&sh->lock);
1791         return 0;
1792 }
1793
1794 static void end_reshape(raid5_conf_t *conf);
1795
1796 static int page_is_zero(struct page *p)
1797 {
1798         char *a = page_address(p);
1799         return ((*(u32*)a) == 0 &&
1800                 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1801 }
1802
1803 static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks)
1804 {
1805         int sectors_per_chunk = conf->chunk_size >> 9;
1806         int pd_idx, dd_idx;
1807         int chunk_offset = sector_div(stripe, sectors_per_chunk);
1808
1809         raid5_compute_sector(stripe * (disks - conf->max_degraded)
1810                              *sectors_per_chunk + chunk_offset,
1811                              disks, disks - conf->max_degraded,
1812                              &dd_idx, &pd_idx, conf);
1813         return pd_idx;
1814 }
1815
1816 static void
1817 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
1818                                 struct stripe_head_state *s, int disks,
1819                                 struct bio **return_bi)
1820 {
1821         int i;
1822         for (i = disks; i--; ) {
1823                 struct bio *bi;
1824                 int bitmap_end = 0;
1825
1826                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1827                         mdk_rdev_t *rdev;
1828                         rcu_read_lock();
1829                         rdev = rcu_dereference(conf->disks[i].rdev);
1830                         if (rdev && test_bit(In_sync, &rdev->flags))
1831                                 /* multiple read failures in one stripe */
1832                                 md_error(conf->mddev, rdev);
1833                         rcu_read_unlock();
1834                 }
1835                 spin_lock_irq(&conf->device_lock);
1836                 /* fail all writes first */
1837                 bi = sh->dev[i].towrite;
1838                 sh->dev[i].towrite = NULL;
1839                 if (bi) {
1840                         s->to_write--;
1841                         bitmap_end = 1;
1842                 }
1843
1844                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1845                         wake_up(&conf->wait_for_overlap);
1846
1847                 while (bi && bi->bi_sector <
1848                         sh->dev[i].sector + STRIPE_SECTORS) {
1849                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1850                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1851                         if (!raid5_dec_bi_phys_segments(bi)) {
1852                                 md_write_end(conf->mddev);
1853                                 bi->bi_next = *return_bi;
1854                                 *return_bi = bi;
1855                         }
1856                         bi = nextbi;
1857                 }
1858                 /* and fail all 'written' */
1859                 bi = sh->dev[i].written;
1860                 sh->dev[i].written = NULL;
1861                 if (bi) bitmap_end = 1;
1862                 while (bi && bi->bi_sector <
1863                        sh->dev[i].sector + STRIPE_SECTORS) {
1864                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1865                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
1866                         if (!raid5_dec_bi_phys_segments(bi)) {
1867                                 md_write_end(conf->mddev);
1868                                 bi->bi_next = *return_bi;
1869                                 *return_bi = bi;
1870                         }
1871                         bi = bi2;
1872                 }
1873
1874                 /* fail any reads if this device is non-operational and
1875                  * the data has not reached the cache yet.
1876                  */
1877                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
1878                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1879                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
1880                         bi = sh->dev[i].toread;
1881                         sh->dev[i].toread = NULL;
1882                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1883                                 wake_up(&conf->wait_for_overlap);
1884                         if (bi) s->to_read--;
1885                         while (bi && bi->bi_sector <
1886                                sh->dev[i].sector + STRIPE_SECTORS) {
1887                                 struct bio *nextbi =
1888                                         r5_next_bio(bi, sh->dev[i].sector);
1889                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1890                                 if (!raid5_dec_bi_phys_segments(bi)) {
1891                                         bi->bi_next = *return_bi;
1892                                         *return_bi = bi;
1893                                 }
1894                                 bi = nextbi;
1895                         }
1896                 }
1897                 spin_unlock_irq(&conf->device_lock);
1898                 if (bitmap_end)
1899                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1900                                         STRIPE_SECTORS, 0, 0);
1901         }
1902
1903         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
1904                 if (atomic_dec_and_test(&conf->pending_full_writes))
1905                         md_wakeup_thread(conf->mddev->thread);
1906 }
1907
1908 /* fetch_block5 - checks the given member device to see if its data needs
1909  * to be read or computed to satisfy a request.
1910  *
1911  * Returns 1 when no more member devices need to be checked, otherwise returns
1912  * 0 to tell the loop in handle_stripe_fill5 to continue
1913  */
1914 static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s,
1915                         int disk_idx, int disks)
1916 {
1917         struct r5dev *dev = &sh->dev[disk_idx];
1918         struct r5dev *failed_dev = &sh->dev[s->failed_num];
1919
1920         /* is the data in this block needed, and can we get it? */
1921         if (!test_bit(R5_LOCKED, &dev->flags) &&
1922             !test_bit(R5_UPTODATE, &dev->flags) &&
1923             (dev->toread ||
1924              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1925              s->syncing || s->expanding ||
1926              (s->failed &&
1927               (failed_dev->toread ||
1928                (failed_dev->towrite &&
1929                 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) {
1930                 /* We would like to get this block, possibly by computing it,
1931                  * otherwise read it if the backing disk is insync
1932                  */
1933                 if ((s->uptodate == disks - 1) &&
1934                     (s->failed && disk_idx == s->failed_num)) {
1935                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
1936                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
1937                         set_bit(R5_Wantcompute, &dev->flags);
1938                         sh->ops.target = disk_idx;
1939                         s->req_compute = 1;
1940                         /* Careful: from this point on 'uptodate' is in the eye
1941                          * of raid5_run_ops which services 'compute' operations
1942                          * before writes. R5_Wantcompute flags a block that will
1943                          * be R5_UPTODATE by the time it is needed for a
1944                          * subsequent operation.
1945                          */
1946                         s->uptodate++;
1947                         return 1; /* uptodate + compute == disks */
1948                 } else if (test_bit(R5_Insync, &dev->flags)) {
1949                         set_bit(R5_LOCKED, &dev->flags);
1950                         set_bit(R5_Wantread, &dev->flags);
1951                         s->locked++;
1952                         pr_debug("Reading block %d (sync=%d)\n", disk_idx,
1953                                 s->syncing);
1954                 }
1955         }
1956
1957         return 0;
1958 }
1959
1960 /**
1961  * handle_stripe_fill5 - read or compute data to satisfy pending requests.
1962  */
1963 static void handle_stripe_fill5(struct stripe_head *sh,
1964                         struct stripe_head_state *s, int disks)
1965 {
1966         int i;
1967
1968         /* look for blocks to read/compute, skip this if a compute
1969          * is already in flight, or if the stripe contents are in the
1970          * midst of changing due to a write
1971          */
1972         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
1973             !sh->reconstruct_state)
1974                 for (i = disks; i--; )
1975                         if (fetch_block5(sh, s, i, disks))
1976                                 break;
1977         set_bit(STRIPE_HANDLE, &sh->state);
1978 }
1979
1980 static void handle_stripe_fill6(struct stripe_head *sh,
1981                         struct stripe_head_state *s, struct r6_state *r6s,
1982                         int disks)
1983 {
1984         int i;
1985         for (i = disks; i--; ) {
1986                 struct r5dev *dev = &sh->dev[i];
1987                 if (!test_bit(R5_LOCKED, &dev->flags) &&
1988                     !test_bit(R5_UPTODATE, &dev->flags) &&
1989                     (dev->toread || (dev->towrite &&
1990                      !test_bit(R5_OVERWRITE, &dev->flags)) ||
1991                      s->syncing || s->expanding ||
1992                      (s->failed >= 1 &&
1993                       (sh->dev[r6s->failed_num[0]].toread ||
1994                        s->to_write)) ||
1995                      (s->failed >= 2 &&
1996                       (sh->dev[r6s->failed_num[1]].toread ||
1997                        s->to_write)))) {
1998                         /* we would like to get this block, possibly
1999                          * by computing it, but we might not be able to
2000                          */
2001                         if ((s->uptodate == disks - 1) &&
2002                             (s->failed && (i == r6s->failed_num[0] ||
2003                                            i == r6s->failed_num[1]))) {
2004                                 pr_debug("Computing stripe %llu block %d\n",
2005                                        (unsigned long long)sh->sector, i);
2006                                 compute_block_1(sh, i, 0);
2007                                 s->uptodate++;
2008                         } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2009                                 /* Computing 2-failure is *very* expensive; only
2010                                  * do it if failed >= 2
2011                                  */
2012                                 int other;
2013                                 for (other = disks; other--; ) {
2014                                         if (other == i)
2015                                                 continue;
2016                                         if (!test_bit(R5_UPTODATE,
2017                                               &sh->dev[other].flags))
2018                                                 break;
2019                                 }
2020                                 BUG_ON(other < 0);
2021                                 pr_debug("Computing stripe %llu blocks %d,%d\n",
2022                                        (unsigned long long)sh->sector,
2023                                        i, other);
2024                                 compute_block_2(sh, i, other);
2025                                 s->uptodate += 2;
2026                         } else if (test_bit(R5_Insync, &dev->flags)) {
2027                                 set_bit(R5_LOCKED, &dev->flags);
2028                                 set_bit(R5_Wantread, &dev->flags);
2029                                 s->locked++;
2030                                 pr_debug("Reading block %d (sync=%d)\n",
2031                                         i, s->syncing);
2032                         }
2033                 }
2034         }
2035         set_bit(STRIPE_HANDLE, &sh->state);
2036 }
2037
2038
2039 /* handle_stripe_clean_event
2040  * any written block on an uptodate or failed drive can be returned.
2041  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2042  * never LOCKED, so we don't need to test 'failed' directly.
2043  */
2044 static void handle_stripe_clean_event(raid5_conf_t *conf,
2045         struct stripe_head *sh, int disks, struct bio **return_bi)
2046 {
2047         int i;
2048         struct r5dev *dev;
2049
2050         for (i = disks; i--; )
2051                 if (sh->dev[i].written) {
2052                         dev = &sh->dev[i];
2053                         if (!test_bit(R5_LOCKED, &dev->flags) &&
2054                                 test_bit(R5_UPTODATE, &dev->flags)) {
2055                                 /* We can return any write requests */
2056                                 struct bio *wbi, *wbi2;
2057                                 int bitmap_end = 0;
2058                                 pr_debug("Return write for disc %d\n", i);
2059                                 spin_lock_irq(&conf->device_lock);
2060                                 wbi = dev->written;
2061                                 dev->written = NULL;
2062                                 while (wbi && wbi->bi_sector <
2063                                         dev->sector + STRIPE_SECTORS) {
2064                                         wbi2 = r5_next_bio(wbi, dev->sector);
2065                                         if (!raid5_dec_bi_phys_segments(wbi)) {
2066                                                 md_write_end(conf->mddev);
2067                                                 wbi->bi_next = *return_bi;
2068                                                 *return_bi = wbi;
2069                                         }
2070                                         wbi = wbi2;
2071                                 }
2072                                 if (dev->towrite == NULL)
2073                                         bitmap_end = 1;
2074                                 spin_unlock_irq(&conf->device_lock);
2075                                 if (bitmap_end)
2076                                         bitmap_endwrite(conf->mddev->bitmap,
2077                                                         sh->sector,
2078                                                         STRIPE_SECTORS,
2079                                          !test_bit(STRIPE_DEGRADED, &sh->state),
2080                                                         0);
2081                         }
2082                 }
2083
2084         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2085                 if (atomic_dec_and_test(&conf->pending_full_writes))
2086                         md_wakeup_thread(conf->mddev->thread);
2087 }
2088
2089 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2090                 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2091 {
2092         int rmw = 0, rcw = 0, i;
2093         for (i = disks; i--; ) {
2094                 /* would I have to read this buffer for read_modify_write */
2095                 struct r5dev *dev = &sh->dev[i];
2096                 if ((dev->towrite || i == sh->pd_idx) &&
2097                     !test_bit(R5_LOCKED, &dev->flags) &&
2098                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2099                       test_bit(R5_Wantcompute, &dev->flags))) {
2100                         if (test_bit(R5_Insync, &dev->flags))
2101                                 rmw++;
2102                         else
2103                                 rmw += 2*disks;  /* cannot read it */
2104                 }
2105                 /* Would I have to read this buffer for reconstruct_write */
2106                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2107                     !test_bit(R5_LOCKED, &dev->flags) &&
2108                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2109                     test_bit(R5_Wantcompute, &dev->flags))) {
2110                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2111                         else
2112                                 rcw += 2*disks;
2113                 }
2114         }
2115         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2116                 (unsigned long long)sh->sector, rmw, rcw);
2117         set_bit(STRIPE_HANDLE, &sh->state);
2118         if (rmw < rcw && rmw > 0)
2119                 /* prefer read-modify-write, but need to get some data */
2120                 for (i = disks; i--; ) {
2121                         struct r5dev *dev = &sh->dev[i];
2122                         if ((dev->towrite || i == sh->pd_idx) &&
2123                             !test_bit(R5_LOCKED, &dev->flags) &&
2124                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2125                             test_bit(R5_Wantcompute, &dev->flags)) &&
2126                             test_bit(R5_Insync, &dev->flags)) {
2127                                 if (
2128                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2129                                         pr_debug("Read_old block "
2130                                                 "%d for r-m-w\n", i);
2131                                         set_bit(R5_LOCKED, &dev->flags);
2132                                         set_bit(R5_Wantread, &dev->flags);
2133                                         s->locked++;
2134                                 } else {
2135                                         set_bit(STRIPE_DELAYED, &sh->state);
2136                                         set_bit(STRIPE_HANDLE, &sh->state);
2137                                 }
2138                         }
2139                 }
2140         if (rcw <= rmw && rcw > 0)
2141                 /* want reconstruct write, but need to get some data */
2142                 for (i = disks; i--; ) {
2143                         struct r5dev *dev = &sh->dev[i];
2144                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2145                             i != sh->pd_idx &&
2146                             !test_bit(R5_LOCKED, &dev->flags) &&
2147                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2148                             test_bit(R5_Wantcompute, &dev->flags)) &&
2149                             test_bit(R5_Insync, &dev->flags)) {
2150                                 if (
2151                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2152                                         pr_debug("Read_old block "
2153                                                 "%d for Reconstruct\n", i);
2154                                         set_bit(R5_LOCKED, &dev->flags);
2155                                         set_bit(R5_Wantread, &dev->flags);
2156                                         s->locked++;
2157                                 } else {
2158                                         set_bit(STRIPE_DELAYED, &sh->state);
2159                                         set_bit(STRIPE_HANDLE, &sh->state);
2160                                 }
2161                         }
2162                 }
2163         /* now if nothing is locked, and if we have enough data,
2164          * we can start a write request
2165          */
2166         /* since handle_stripe can be called at any time we need to handle the
2167          * case where a compute block operation has been submitted and then a
2168          * subsequent call wants to start a write request.  raid5_run_ops only
2169          * handles the case where compute block and postxor are requested
2170          * simultaneously.  If this is not the case then new writes need to be
2171          * held off until the compute completes.
2172          */
2173         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2174             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2175             !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2176                 schedule_reconstruction5(sh, s, rcw == 0, 0);
2177 }
2178
2179 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2180                 struct stripe_head *sh, struct stripe_head_state *s,
2181                 struct r6_state *r6s, int disks)
2182 {
2183         int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2184         int qd_idx = r6s->qd_idx;
2185         for (i = disks; i--; ) {
2186                 struct r5dev *dev = &sh->dev[i];
2187                 /* Would I have to read this buffer for reconstruct_write */
2188                 if (!test_bit(R5_OVERWRITE, &dev->flags)
2189                     && i != pd_idx && i != qd_idx
2190                     && (!test_bit(R5_LOCKED, &dev->flags)
2191                             ) &&
2192                     !test_bit(R5_UPTODATE, &dev->flags)) {
2193                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2194                         else {
2195                                 pr_debug("raid6: must_compute: "
2196                                         "disk %d flags=%#lx\n", i, dev->flags);
2197                                 must_compute++;
2198                         }
2199                 }
2200         }
2201         pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2202                (unsigned long long)sh->sector, rcw, must_compute);
2203         set_bit(STRIPE_HANDLE, &sh->state);
2204
2205         if (rcw > 0)
2206                 /* want reconstruct write, but need to get some data */
2207                 for (i = disks; i--; ) {
2208                         struct r5dev *dev = &sh->dev[i];
2209                         if (!test_bit(R5_OVERWRITE, &dev->flags)
2210                             && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2211                             && !test_bit(R5_LOCKED, &dev->flags) &&
2212                             !test_bit(R5_UPTODATE, &dev->flags) &&
2213                             test_bit(R5_Insync, &dev->flags)) {
2214                                 if (
2215                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2216                                         pr_debug("Read_old stripe %llu "
2217                                                 "block %d for Reconstruct\n",
2218                                              (unsigned long long)sh->sector, i);
2219                                         set_bit(R5_LOCKED, &dev->flags);
2220                                         set_bit(R5_Wantread, &dev->flags);
2221                                         s->locked++;
2222                                 } else {
2223                                         pr_debug("Request delayed stripe %llu "
2224                                                 "block %d for Reconstruct\n",
2225                                              (unsigned long long)sh->sector, i);
2226                                         set_bit(STRIPE_DELAYED, &sh->state);
2227                                         set_bit(STRIPE_HANDLE, &sh->state);
2228                                 }
2229                         }
2230                 }
2231         /* now if nothing is locked, and if we have enough data, we can start a
2232          * write request
2233          */
2234         if (s->locked == 0 && rcw == 0 &&
2235             !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2236                 if (must_compute > 0) {
2237                         /* We have failed blocks and need to compute them */
2238                         switch (s->failed) {
2239                         case 0:
2240                                 BUG();
2241                         case 1:
2242                                 compute_block_1(sh, r6s->failed_num[0], 0);
2243                                 break;
2244                         case 2:
2245                                 compute_block_2(sh, r6s->failed_num[0],
2246                                                 r6s->failed_num[1]);
2247                                 break;
2248                         default: /* This request should have been failed? */
2249                                 BUG();
2250                         }
2251                 }
2252
2253                 pr_debug("Computing parity for stripe %llu\n",
2254                         (unsigned long long)sh->sector);
2255                 compute_parity6(sh, RECONSTRUCT_WRITE);
2256                 /* now every locked buffer is ready to be written */
2257                 for (i = disks; i--; )
2258                         if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2259                                 pr_debug("Writing stripe %llu block %d\n",
2260                                        (unsigned long long)sh->sector, i);
2261                                 s->locked++;
2262                                 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2263                         }
2264                 if (s->locked == disks)
2265                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2266                                 atomic_inc(&conf->pending_full_writes);
2267                 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2268                 set_bit(STRIPE_INSYNC, &sh->state);
2269
2270                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2271                         atomic_dec(&conf->preread_active_stripes);
2272                         if (atomic_read(&conf->preread_active_stripes) <
2273                             IO_THRESHOLD)
2274                                 md_wakeup_thread(conf->mddev->thread);
2275                 }
2276         }
2277 }
2278
2279 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2280                                 struct stripe_head_state *s, int disks)
2281 {
2282         struct r5dev *dev = NULL;
2283
2284         set_bit(STRIPE_HANDLE, &sh->state);
2285
2286         switch (sh->check_state) {
2287         case check_state_idle:
2288                 /* start a new check operation if there are no failures */
2289                 if (s->failed == 0) {
2290                         BUG_ON(s->uptodate != disks);
2291                         sh->check_state = check_state_run;
2292                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
2293                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2294                         s->uptodate--;
2295                         break;
2296                 }
2297                 dev = &sh->dev[s->failed_num];
2298                 /* fall through */
2299         case check_state_compute_result:
2300                 sh->check_state = check_state_idle;
2301                 if (!dev)
2302                         dev = &sh->dev[sh->pd_idx];
2303
2304                 /* check that a write has not made the stripe insync */
2305                 if (test_bit(STRIPE_INSYNC, &sh->state))
2306                         break;
2307
2308                 /* either failed parity check, or recovery is happening */
2309                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2310                 BUG_ON(s->uptodate != disks);
2311
2312                 set_bit(R5_LOCKED, &dev->flags);
2313                 s->locked++;
2314                 set_bit(R5_Wantwrite, &dev->flags);
2315
2316                 clear_bit(STRIPE_DEGRADED, &sh->state);
2317                 set_bit(STRIPE_INSYNC, &sh->state);
2318                 break;
2319         case check_state_run:
2320                 break; /* we will be called again upon completion */
2321         case check_state_check_result:
2322                 sh->check_state = check_state_idle;
2323
2324                 /* if a failure occurred during the check operation, leave
2325                  * STRIPE_INSYNC not set and let the stripe be handled again
2326                  */
2327                 if (s->failed)
2328                         break;
2329
2330                 /* handle a successful check operation, if parity is correct
2331                  * we are done.  Otherwise update the mismatch count and repair
2332                  * parity if !MD_RECOVERY_CHECK
2333                  */
2334                 if (sh->ops.zero_sum_result == 0)
2335                         /* parity is correct (on disc,
2336                          * not in buffer any more)
2337                          */
2338                         set_bit(STRIPE_INSYNC, &sh->state);
2339                 else {
2340                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2341                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2342                                 /* don't try to repair!! */
2343                                 set_bit(STRIPE_INSYNC, &sh->state);
2344                         else {
2345                                 sh->check_state = check_state_compute_run;
2346                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2347                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2348                                 set_bit(R5_Wantcompute,
2349                                         &sh->dev[sh->pd_idx].flags);
2350                                 sh->ops.target = sh->pd_idx;
2351                                 s->uptodate++;
2352                         }
2353                 }
2354                 break;
2355         case check_state_compute_run:
2356                 break;
2357         default:
2358                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2359                        __func__, sh->check_state,
2360                        (unsigned long long) sh->sector);
2361                 BUG();
2362         }
2363 }
2364
2365
2366 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2367                                 struct stripe_head_state *s,
2368                                 struct r6_state *r6s, struct page *tmp_page,
2369                                 int disks)
2370 {
2371         int update_p = 0, update_q = 0;
2372         struct r5dev *dev;
2373         int pd_idx = sh->pd_idx;
2374         int qd_idx = r6s->qd_idx;
2375
2376         set_bit(STRIPE_HANDLE, &sh->state);
2377
2378         BUG_ON(s->failed > 2);
2379         BUG_ON(s->uptodate < disks);
2380         /* Want to check and possibly repair P and Q.
2381          * However there could be one 'failed' device, in which
2382          * case we can only check one of them, possibly using the
2383          * other to generate missing data
2384          */
2385
2386         /* If !tmp_page, we cannot do the calculations,
2387          * but as we have set STRIPE_HANDLE, we will soon be called
2388          * by stripe_handle with a tmp_page - just wait until then.
2389          */
2390         if (tmp_page) {
2391                 if (s->failed == r6s->q_failed) {
2392                         /* The only possible failed device holds 'Q', so it
2393                          * makes sense to check P (If anything else were failed,
2394                          * we would have used P to recreate it).
2395                          */
2396                         compute_block_1(sh, pd_idx, 1);
2397                         if (!page_is_zero(sh->dev[pd_idx].page)) {
2398                                 compute_block_1(sh, pd_idx, 0);
2399                                 update_p = 1;
2400                         }
2401                 }
2402                 if (!r6s->q_failed && s->failed < 2) {
2403                         /* q is not failed, and we didn't use it to generate
2404                          * anything, so it makes sense to check it
2405                          */
2406                         memcpy(page_address(tmp_page),
2407                                page_address(sh->dev[qd_idx].page),
2408                                STRIPE_SIZE);
2409                         compute_parity6(sh, UPDATE_PARITY);
2410                         if (memcmp(page_address(tmp_page),
2411                                    page_address(sh->dev[qd_idx].page),
2412                                    STRIPE_SIZE) != 0) {
2413                                 clear_bit(STRIPE_INSYNC, &sh->state);
2414                                 update_q = 1;
2415                         }
2416                 }
2417                 if (update_p || update_q) {
2418                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2419                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2420                                 /* don't try to repair!! */
2421                                 update_p = update_q = 0;
2422                 }
2423
2424                 /* now write out any block on a failed drive,
2425                  * or P or Q if they need it
2426                  */
2427
2428                 if (s->failed == 2) {
2429                         dev = &sh->dev[r6s->failed_num[1]];
2430                         s->locked++;
2431                         set_bit(R5_LOCKED, &dev->flags);
2432                         set_bit(R5_Wantwrite, &dev->flags);
2433                 }
2434                 if (s->failed >= 1) {
2435                         dev = &sh->dev[r6s->failed_num[0]];
2436                         s->locked++;
2437                         set_bit(R5_LOCKED, &dev->flags);
2438                         set_bit(R5_Wantwrite, &dev->flags);
2439                 }
2440
2441                 if (update_p) {
2442                         dev = &sh->dev[pd_idx];
2443                         s->locked++;
2444                         set_bit(R5_LOCKED, &dev->flags);
2445                         set_bit(R5_Wantwrite, &dev->flags);
2446                 }
2447                 if (update_q) {
2448                         dev = &sh->dev[qd_idx];
2449                         s->locked++;
2450                         set_bit(R5_LOCKED, &dev->flags);
2451                         set_bit(R5_Wantwrite, &dev->flags);
2452                 }
2453                 clear_bit(STRIPE_DEGRADED, &sh->state);
2454
2455                 set_bit(STRIPE_INSYNC, &sh->state);
2456         }
2457 }
2458
2459 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2460                                 struct r6_state *r6s)
2461 {
2462         int i;
2463
2464         /* We have read all the blocks in this stripe and now we need to
2465          * copy some of them into a target stripe for expand.
2466          */
2467         struct dma_async_tx_descriptor *tx = NULL;
2468         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2469         for (i = 0; i < sh->disks; i++)
2470                 if (i != sh->pd_idx && (!r6s || i != r6s->qd_idx)) {
2471                         int dd_idx, pd_idx, j;
2472                         struct stripe_head *sh2;
2473
2474                         sector_t bn = compute_blocknr(sh, i);
2475                         sector_t s = raid5_compute_sector(bn, conf->raid_disks,
2476                                                 conf->raid_disks -
2477                                                 conf->max_degraded, &dd_idx,
2478                                                 &pd_idx, conf);
2479                         sh2 = get_active_stripe(conf, s, conf->raid_disks,
2480                                                 pd_idx, 1);
2481                         if (sh2 == NULL)
2482                                 /* so far only the early blocks of this stripe
2483                                  * have been requested.  When later blocks
2484                                  * get requested, we will try again
2485                                  */
2486                                 continue;
2487                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2488                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2489                                 /* must have already done this block */
2490                                 release_stripe(sh2);
2491                                 continue;
2492                         }
2493
2494                         /* place all the copies on one channel */
2495                         tx = async_memcpy(sh2->dev[dd_idx].page,
2496                                 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2497                                 ASYNC_TX_DEP_ACK, tx, NULL, NULL);
2498
2499                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2500                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2501                         for (j = 0; j < conf->raid_disks; j++)
2502                                 if (j != sh2->pd_idx &&
2503                                     (!r6s || j != raid6_next_disk(sh2->pd_idx,
2504                                                                  sh2->disks)) &&
2505                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
2506                                         break;
2507                         if (j == conf->raid_disks) {
2508                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2509                                 set_bit(STRIPE_HANDLE, &sh2->state);
2510                         }
2511                         release_stripe(sh2);
2512
2513                 }
2514         /* done submitting copies, wait for them to complete */
2515         if (tx) {
2516                 async_tx_ack(tx);
2517                 dma_wait_for_async_tx(tx);
2518         }
2519 }
2520
2521
2522 /*
2523  * handle_stripe - do things to a stripe.
2524  *
2525  * We lock the stripe and then examine the state of various bits
2526  * to see what needs to be done.
2527  * Possible results:
2528  *    return some read request which now have data
2529  *    return some write requests which are safely on disc
2530  *    schedule a read on some buffers
2531  *    schedule a write of some buffers
2532  *    return confirmation of parity correctness
2533  *
2534  * buffers are taken off read_list or write_list, and bh_cache buffers
2535  * get BH_Lock set before the stripe lock is released.
2536  *
2537  */
2538
2539 static bool handle_stripe5(struct stripe_head *sh)
2540 {
2541         raid5_conf_t *conf = sh->raid_conf;
2542         int disks = sh->disks, i;
2543         struct bio *return_bi = NULL;
2544         struct stripe_head_state s;
2545         struct r5dev *dev;
2546         mdk_rdev_t *blocked_rdev = NULL;
2547         int prexor;
2548
2549         memset(&s, 0, sizeof(s));
2550         pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
2551                  "reconstruct:%d\n", (unsigned long long)sh->sector, sh->state,
2552                  atomic_read(&sh->count), sh->pd_idx, sh->check_state,
2553                  sh->reconstruct_state);
2554
2555         spin_lock(&sh->lock);
2556         clear_bit(STRIPE_HANDLE, &sh->state);
2557         clear_bit(STRIPE_DELAYED, &sh->state);
2558
2559         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2560         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2561         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2562
2563         /* Now to look around and see what can be done */
2564         rcu_read_lock();
2565         for (i=disks; i--; ) {
2566                 mdk_rdev_t *rdev;
2567                 struct r5dev *dev = &sh->dev[i];
2568                 clear_bit(R5_Insync, &dev->flags);
2569
2570                 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2571                         "written %p\n", i, dev->flags, dev->toread, dev->read,
2572                         dev->towrite, dev->written);
2573
2574                 /* maybe we can request a biofill operation
2575                  *
2576                  * new wantfill requests are only permitted while
2577                  * ops_complete_biofill is guaranteed to be inactive
2578                  */
2579                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2580                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
2581                         set_bit(R5_Wantfill, &dev->flags);
2582
2583                 /* now count some things */
2584                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2585                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2586                 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2587
2588                 if (test_bit(R5_Wantfill, &dev->flags))
2589                         s.to_fill++;
2590                 else if (dev->toread)
2591                         s.to_read++;
2592                 if (dev->towrite) {
2593                         s.to_write++;
2594                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2595                                 s.non_overwrite++;
2596                 }
2597                 if (dev->written)
2598                         s.written++;
2599                 rdev = rcu_dereference(conf->disks[i].rdev);
2600                 if (blocked_rdev == NULL &&
2601                     rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2602                         blocked_rdev = rdev;
2603                         atomic_inc(&rdev->nr_pending);
2604                 }
2605                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2606                         /* The ReadError flag will just be confusing now */
2607                         clear_bit(R5_ReadError, &dev->flags);
2608                         clear_bit(R5_ReWrite, &dev->flags);
2609                 }
2610                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2611                     || test_bit(R5_ReadError, &dev->flags)) {
2612                         s.failed++;
2613                         s.failed_num = i;
2614                 } else
2615                         set_bit(R5_Insync, &dev->flags);
2616         }
2617         rcu_read_unlock();
2618
2619         if (unlikely(blocked_rdev)) {
2620                 if (s.syncing || s.expanding || s.expanded ||
2621                     s.to_write || s.written) {
2622                         set_bit(STRIPE_HANDLE, &sh->state);
2623                         goto unlock;
2624                 }
2625                 /* There is nothing for the blocked_rdev to block */
2626                 rdev_dec_pending(blocked_rdev, conf->mddev);
2627                 blocked_rdev = NULL;
2628         }
2629
2630         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
2631                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
2632                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
2633         }
2634
2635         pr_debug("locked=%d uptodate=%d to_read=%d"
2636                 " to_write=%d failed=%d failed_num=%d\n",
2637                 s.locked, s.uptodate, s.to_read, s.to_write,
2638                 s.failed, s.failed_num);
2639         /* check if the array has lost two devices and, if so, some requests might
2640          * need to be failed
2641          */
2642         if (s.failed > 1 && s.to_read+s.to_write+s.written)
2643                 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
2644         if (s.failed > 1 && s.syncing) {
2645                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2646                 clear_bit(STRIPE_SYNCING, &sh->state);
2647                 s.syncing = 0;
2648         }
2649
2650         /* might be able to return some write requests if the parity block
2651          * is safe, or on a failed drive
2652          */
2653         dev = &sh->dev[sh->pd_idx];
2654         if ( s.written &&
2655              ((test_bit(R5_Insync, &dev->flags) &&
2656                !test_bit(R5_LOCKED, &dev->flags) &&
2657                test_bit(R5_UPTODATE, &dev->flags)) ||
2658                (s.failed == 1 && s.failed_num == sh->pd_idx)))
2659                 handle_stripe_clean_event(conf, sh, disks, &return_bi);
2660
2661         /* Now we might consider reading some blocks, either to check/generate
2662          * parity, or to satisfy requests
2663          * or to load a block that is being partially written.
2664          */
2665         if (s.to_read || s.non_overwrite ||
2666             (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
2667                 handle_stripe_fill5(sh, &s, disks);
2668
2669         /* Now we check to see if any write operations have recently
2670          * completed
2671          */
2672         prexor = 0;
2673         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
2674                 prexor = 1;
2675         if (sh->reconstruct_state == reconstruct_state_drain_result ||
2676             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
2677                 sh->reconstruct_state = reconstruct_state_idle;
2678
2679                 /* All the 'written' buffers and the parity block are ready to
2680                  * be written back to disk
2681                  */
2682                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2683                 for (i = disks; i--; ) {
2684                         dev = &sh->dev[i];
2685                         if (test_bit(R5_LOCKED, &dev->flags) &&
2686                                 (i == sh->pd_idx || dev->written)) {
2687                                 pr_debug("Writing block %d\n", i);
2688                                 set_bit(R5_Wantwrite, &dev->flags);
2689                                 if (prexor)
2690                                         continue;
2691                                 if (!test_bit(R5_Insync, &dev->flags) ||
2692                                     (i == sh->pd_idx && s.failed == 0))
2693                                         set_bit(STRIPE_INSYNC, &sh->state);
2694                         }
2695                 }
2696                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2697                         atomic_dec(&conf->preread_active_stripes);
2698                         if (atomic_read(&conf->preread_active_stripes) <
2699                                 IO_THRESHOLD)
2700                                 md_wakeup_thread(conf->mddev->thread);
2701                 }
2702         }
2703
2704         /* Now to consider new write requests and what else, if anything
2705          * should be read.  We do not handle new writes when:
2706          * 1/ A 'write' operation (copy+xor) is already in flight.
2707          * 2/ A 'check' operation is in flight, as it may clobber the parity
2708          *    block.
2709          */
2710         if (s.to_write && !sh->reconstruct_state && !sh->check_state)
2711                 handle_stripe_dirtying5(conf, sh, &s, disks);
2712
2713         /* maybe we need to check and possibly fix the parity for this stripe
2714          * Any reads will already have been scheduled, so we just see if enough
2715          * data is available.  The parity check is held off while parity
2716          * dependent operations are in flight.
2717          */
2718         if (sh->check_state ||
2719             (s.syncing && s.locked == 0 &&
2720              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
2721              !test_bit(STRIPE_INSYNC, &sh->state)))
2722                 handle_parity_checks5(conf, sh, &s, disks);
2723
2724         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2725                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2726                 clear_bit(STRIPE_SYNCING, &sh->state);
2727         }
2728
2729         /* If the failed drive is just a ReadError, then we might need to progress
2730          * the repair/check process
2731          */
2732         if (s.failed == 1 && !conf->mddev->ro &&
2733             test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2734             && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2735             && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2736                 ) {
2737                 dev = &sh->dev[s.failed_num];
2738                 if (!test_bit(R5_ReWrite, &dev->flags)) {
2739                         set_bit(R5_Wantwrite, &dev->flags);
2740                         set_bit(R5_ReWrite, &dev->flags);
2741                         set_bit(R5_LOCKED, &dev->flags);
2742                         s.locked++;
2743                 } else {
2744                         /* let's read it back */
2745                         set_bit(R5_Wantread, &dev->flags);
2746                         set_bit(R5_LOCKED, &dev->flags);
2747                         s.locked++;
2748                 }
2749         }
2750
2751         /* Finish reconstruct operations initiated by the expansion process */
2752         if (sh->reconstruct_state == reconstruct_state_result) {
2753                 sh->reconstruct_state = reconstruct_state_idle;
2754                 clear_bit(STRIPE_EXPANDING, &sh->state);
2755                 for (i = conf->raid_disks; i--; ) {
2756                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
2757                         set_bit(R5_LOCKED, &sh->dev[i].flags);
2758                         s.locked++;
2759                 }
2760         }
2761
2762         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
2763             !sh->reconstruct_state) {
2764                 /* Need to write out all blocks after computing parity */
2765                 sh->disks = conf->raid_disks;
2766                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
2767                         conf->raid_disks);
2768                 schedule_reconstruction5(sh, &s, 1, 1);
2769         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
2770                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2771                 atomic_dec(&conf->reshape_stripes);
2772                 wake_up(&conf->wait_for_overlap);
2773                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2774         }
2775
2776         if (s.expanding && s.locked == 0 &&
2777             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
2778                 handle_stripe_expansion(conf, sh, NULL);
2779
2780  unlock:
2781         spin_unlock(&sh->lock);
2782
2783         /* wait for this device to become unblocked */
2784         if (unlikely(blocked_rdev))
2785                 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
2786
2787         if (s.ops_request)
2788                 raid5_run_ops(sh, s.ops_request);
2789
2790         ops_run_io(sh, &s);
2791
2792         return_io(return_bi);
2793
2794         return blocked_rdev == NULL;
2795 }
2796
2797 static bool handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
2798 {
2799         raid6_conf_t *conf = sh->raid_conf;
2800         int disks = sh->disks;
2801         struct bio *return_bi = NULL;
2802         int i, pd_idx = sh->pd_idx;
2803         struct stripe_head_state s;
2804         struct r6_state r6s;
2805         struct r5dev *dev, *pdev, *qdev;
2806         mdk_rdev_t *blocked_rdev = NULL;
2807
2808         r6s.qd_idx = raid6_next_disk(pd_idx, disks);
2809         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
2810                 "pd_idx=%d, qd_idx=%d\n",
2811                (unsigned long long)sh->sector, sh->state,
2812                atomic_read(&sh->count), pd_idx, r6s.qd_idx);
2813         memset(&s, 0, sizeof(s));
2814
2815         spin_lock(&sh->lock);
2816         clear_bit(STRIPE_HANDLE, &sh->state);
2817         clear_bit(STRIPE_DELAYED, &sh->state);
2818
2819         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2820         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2821         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2822         /* Now to look around and see what can be done */
2823
2824         rcu_read_lock();
2825         for (i=disks; i--; ) {
2826                 mdk_rdev_t *rdev;
2827                 dev = &sh->dev[i];
2828                 clear_bit(R5_Insync, &dev->flags);
2829
2830                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
2831                         i, dev->flags, dev->toread, dev->towrite, dev->written);
2832                 /* maybe we can reply to a read */
2833                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
2834                         struct bio *rbi, *rbi2;
2835                         pr_debug("Return read for disc %d\n", i);
2836                         spin_lock_irq(&conf->device_lock);
2837                         rbi = dev->toread;
2838                         dev->toread = NULL;
2839                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
2840                                 wake_up(&conf->wait_for_overlap);
2841                         spin_unlock_irq(&conf->device_lock);
2842                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
2843                                 copy_data(0, rbi, dev->page, dev->sector);
2844                                 rbi2 = r5_next_bio(rbi, dev->sector);
2845                                 spin_lock_irq(&conf->device_lock);
2846                                 if (!raid5_dec_bi_phys_segments(rbi)) {
2847                                         rbi->bi_next = return_bi;
2848                                         return_bi = rbi;
2849                                 }
2850                                 spin_unlock_irq(&conf->device_lock);
2851                                 rbi = rbi2;
2852                         }
2853                 }
2854
2855                 /* now count some things */
2856                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2857                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2858
2859
2860                 if (dev->toread)
2861                         s.to_read++;
2862                 if (dev->towrite) {
2863                         s.to_write++;
2864                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2865                                 s.non_overwrite++;
2866                 }
2867                 if (dev->written)
2868                         s.written++;
2869                 rdev = rcu_dereference(conf->disks[i].rdev);
2870                 if (blocked_rdev == NULL &&
2871                     rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2872                         blocked_rdev = rdev;
2873                         atomic_inc(&rdev->nr_pending);
2874                 }
2875                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2876                         /* The ReadError flag will just be confusing now */
2877                         clear_bit(R5_ReadError, &dev->flags);
2878                         clear_bit(R5_ReWrite, &dev->flags);
2879                 }
2880                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2881                     || test_bit(R5_ReadError, &dev->flags)) {
2882                         if (s.failed < 2)
2883                                 r6s.failed_num[s.failed] = i;
2884                         s.failed++;
2885                 } else
2886                         set_bit(R5_Insync, &dev->flags);
2887         }
2888         rcu_read_unlock();
2889
2890         if (unlikely(blocked_rdev)) {
2891                 if (s.syncing || s.expanding || s.expanded ||
2892                     s.to_write || s.written) {
2893                         set_bit(STRIPE_HANDLE, &sh->state);
2894                         goto unlock;
2895                 }
2896                 /* There is nothing for the blocked_rdev to block */
2897                 rdev_dec_pending(blocked_rdev, conf->mddev);
2898                 blocked_rdev = NULL;
2899         }
2900
2901         pr_debug("locked=%d uptodate=%d to_read=%d"
2902                " to_write=%d failed=%d failed_num=%d,%d\n",
2903                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
2904                r6s.failed_num[0], r6s.failed_num[1]);
2905         /* check if the array has lost >2 devices and, if so, some requests
2906          * might need to be failed
2907          */
2908         if (s.failed > 2 && s.to_read+s.to_write+s.written)
2909                 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
2910         if (s.failed > 2 && s.syncing) {
2911                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2912                 clear_bit(STRIPE_SYNCING, &sh->state);
2913                 s.syncing = 0;
2914         }
2915
2916         /*
2917          * might be able to return some write requests if the parity blocks
2918          * are safe, or on a failed drive
2919          */
2920         pdev = &sh->dev[pd_idx];
2921         r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
2922                 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
2923         qdev = &sh->dev[r6s.qd_idx];
2924         r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == r6s.qd_idx)
2925                 || (s.failed >= 2 && r6s.failed_num[1] == r6s.qd_idx);
2926
2927         if ( s.written &&
2928              ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
2929                              && !test_bit(R5_LOCKED, &pdev->flags)
2930                              && test_bit(R5_UPTODATE, &pdev->flags)))) &&
2931              ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
2932                              && !test_bit(R5_LOCKED, &qdev->flags)
2933                              && test_bit(R5_UPTODATE, &qdev->flags)))))
2934                 handle_stripe_clean_event(conf, sh, disks, &return_bi);
2935
2936         /* Now we might consider reading some blocks, either to check/generate
2937          * parity, or to satisfy requests
2938          * or to load a block that is being partially written.
2939          */
2940         if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
2941             (s.syncing && (s.uptodate < disks)) || s.expanding)
2942                 handle_stripe_fill6(sh, &s, &r6s, disks);
2943
2944         /* now to consider writing and what else, if anything should be read */
2945         if (s.to_write)
2946                 handle_stripe_dirtying6(conf, sh, &s, &r6s, disks);
2947
2948         /* maybe we need to check and possibly fix the parity for this stripe
2949          * Any reads will already have been scheduled, so we just see if enough
2950          * data is available
2951          */
2952         if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
2953                 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
2954
2955         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2956                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2957                 clear_bit(STRIPE_SYNCING, &sh->state);
2958         }
2959
2960         /* If the failed drives are just a ReadError, then we might need
2961          * to progress the repair/check process
2962          */
2963         if (s.failed <= 2 && !conf->mddev->ro)
2964                 for (i = 0; i < s.failed; i++) {
2965                         dev = &sh->dev[r6s.failed_num[i]];
2966                         if (test_bit(R5_ReadError, &dev->flags)
2967                             && !test_bit(R5_LOCKED, &dev->flags)
2968                             && test_bit(R5_UPTODATE, &dev->flags)
2969                                 ) {
2970                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
2971                                         set_bit(R5_Wantwrite, &dev->flags);
2972                                         set_bit(R5_ReWrite, &dev->flags);
2973                                         set_bit(R5_LOCKED, &dev->flags);
2974                                 } else {
2975                                         /* let's read it back */
2976                                         set_bit(R5_Wantread, &dev->flags);
2977                                         set_bit(R5_LOCKED, &dev->flags);
2978                                 }
2979                         }
2980                 }
2981
2982         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
2983                 /* Need to write out all blocks after computing P&Q */
2984                 sh->disks = conf->raid_disks;
2985                 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
2986                                              conf->raid_disks);
2987                 compute_parity6(sh, RECONSTRUCT_WRITE);
2988                 for (i = conf->raid_disks ; i-- ;  ) {
2989                         set_bit(R5_LOCKED, &sh->dev[i].flags);
2990                         s.locked++;
2991                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
2992                 }
2993                 clear_bit(STRIPE_EXPANDING, &sh->state);
2994         } else if (s.expanded) {
2995                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2996                 atomic_dec(&conf->reshape_stripes);
2997                 wake_up(&conf->wait_for_overlap);
2998                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2999         }
3000
3001         if (s.expanding && s.locked == 0 &&
3002             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3003                 handle_stripe_expansion(conf, sh, &r6s);
3004
3005  unlock:
3006         spin_unlock(&sh->lock);
3007
3008         /* wait for this device to become unblocked */
3009         if (unlikely(blocked_rdev))
3010                 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3011
3012         ops_run_io(sh, &s);
3013
3014         return_io(return_bi);
3015
3016         return blocked_rdev == NULL;
3017 }
3018
3019 /* returns true if the stripe was handled */
3020 static bool handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3021 {
3022         if (sh->raid_conf->level == 6)
3023                 return handle_stripe6(sh, tmp_page);
3024         else
3025                 return handle_stripe5(sh);
3026 }
3027
3028
3029
3030 static void raid5_activate_delayed(raid5_conf_t *conf)
3031 {
3032         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3033                 while (!list_empty(&conf->delayed_list)) {
3034                         struct list_head *l = conf->delayed_list.next;
3035                         struct stripe_head *sh;
3036                         sh = list_entry(l, struct stripe_head, lru);
3037                         list_del_init(l);
3038                         clear_bit(STRIPE_DELAYED, &sh->state);
3039                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3040                                 atomic_inc(&conf->preread_active_stripes);
3041                         list_add_tail(&sh->lru, &conf->hold_list);
3042                 }
3043         } else
3044                 blk_plug_device(conf->mddev->queue);
3045 }
3046
3047 static void activate_bit_delay(raid5_conf_t *conf)
3048 {
3049         /* device_lock is held */
3050         struct list_head head;
3051         list_add(&head, &conf->bitmap_list);
3052         list_del_init(&conf->bitmap_list);
3053         while (!list_empty(&head)) {
3054                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3055                 list_del_init(&sh->lru);
3056                 atomic_inc(&sh->count);
3057                 __release_stripe(conf, sh);
3058         }
3059 }
3060
3061 static void unplug_slaves(mddev_t *mddev)
3062 {
3063         raid5_conf_t *conf = mddev_to_conf(mddev);
3064         int i;
3065
3066         rcu_read_lock();
3067         for (i=0; i<mddev->raid_disks; i++) {
3068                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3069                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3070                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3071
3072                         atomic_inc(&rdev->nr_pending);
3073                         rcu_read_unlock();
3074
3075                         blk_unplug(r_queue);
3076
3077                         rdev_dec_pending(rdev, mddev);
3078                         rcu_read_lock();
3079                 }
3080         }
3081         rcu_read_unlock();
3082 }
3083
3084 static void raid5_unplug_device(struct request_queue *q)
3085 {
3086         mddev_t *mddev = q->queuedata;
3087         raid5_conf_t *conf = mddev_to_conf(mddev);
3088         unsigned long flags;
3089
3090         spin_lock_irqsave(&conf->device_lock, flags);
3091
3092         if (blk_remove_plug(q)) {
3093                 conf->seq_flush++;
3094                 raid5_activate_delayed(conf);
3095         }
3096         md_wakeup_thread(mddev->thread);
3097
3098         spin_unlock_irqrestore(&conf->device_lock, flags);
3099
3100         unplug_slaves(mddev);
3101 }
3102
3103 static int raid5_congested(void *data, int bits)
3104 {
3105         mddev_t *mddev = data;
3106         raid5_conf_t *conf = mddev_to_conf(mddev);
3107
3108         /* No difference between reads and writes.  Just check
3109          * how busy the stripe_cache is
3110          */
3111         if (conf->inactive_blocked)
3112                 return 1;
3113         if (conf->quiesce)
3114                 return 1;
3115         if (list_empty_careful(&conf->inactive_list))
3116                 return 1;
3117
3118         return 0;
3119 }
3120
3121 /* We want read requests to align with chunks where possible,
3122  * but write requests don't need to.
3123  */
3124 static int raid5_mergeable_bvec(struct request_queue *q,
3125                                 struct bvec_merge_data *bvm,
3126                                 struct bio_vec *biovec)
3127 {
3128         mddev_t *mddev = q->queuedata;
3129         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3130         int max;
3131         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3132         unsigned int bio_sectors = bvm->bi_size >> 9;
3133
3134         if ((bvm->bi_rw & 1) == WRITE)
3135                 return biovec->bv_len; /* always allow writes to be mergeable */
3136
3137         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3138         if (max < 0) max = 0;
3139         if (max <= biovec->bv_len && bio_sectors == 0)
3140                 return biovec->bv_len;
3141         else
3142                 return max;
3143 }
3144
3145
3146 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3147 {
3148         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3149         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3150         unsigned int bio_sectors = bio->bi_size >> 9;
3151
3152         return  chunk_sectors >=
3153                 ((sector & (chunk_sectors - 1)) + bio_sectors);
3154 }
3155
3156 /*
3157  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3158  *  later sampled by raid5d.
3159  */
3160 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3161 {
3162         unsigned long flags;
3163
3164         spin_lock_irqsave(&conf->device_lock, flags);
3165
3166         bi->bi_next = conf->retry_read_aligned_list;
3167         conf->retry_read_aligned_list = bi;
3168
3169         spin_unlock_irqrestore(&conf->device_lock, flags);
3170         md_wakeup_thread(conf->mddev->thread);
3171 }
3172
3173
3174 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3175 {
3176         struct bio *bi;
3177
3178         bi = conf->retry_read_aligned;
3179         if (bi) {
3180                 conf->retry_read_aligned = NULL;
3181                 return bi;
3182         }
3183         bi = conf->retry_read_aligned_list;
3184         if(bi) {
3185                 conf->retry_read_aligned_list = bi->bi_next;
3186                 bi->bi_next = NULL;
3187                 /*
3188                  * this sets the active strip count to 1 and the processed
3189                  * strip count to zero (upper 8 bits)
3190                  */
3191                 bi->bi_phys_segments = 1; /* biased count of active stripes */
3192         }
3193
3194         return bi;
3195 }
3196
3197
3198 /*
3199  *  The "raid5_align_endio" should check if the read succeeded and if it
3200  *  did, call bio_endio on the original bio (having bio_put the new bio
3201  *  first).
3202  *  If the read failed..
3203  */
3204 static void raid5_align_endio(struct bio *bi, int error)
3205 {
3206         struct bio* raid_bi  = bi->bi_private;
3207         mddev_t *mddev;
3208         raid5_conf_t *conf;
3209         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3210         mdk_rdev_t *rdev;
3211
3212         bio_put(bi);
3213
3214         mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3215         conf = mddev_to_conf(mddev);
3216         rdev = (void*)raid_bi->bi_next;
3217         raid_bi->bi_next = NULL;
3218
3219         rdev_dec_pending(rdev, conf->mddev);
3220
3221         if (!error && uptodate) {
3222                 bio_endio(raid_bi, 0);
3223                 if (atomic_dec_and_test(&conf->active_aligned_reads))
3224                         wake_up(&conf->wait_for_stripe);
3225                 return;
3226         }
3227
3228
3229         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3230
3231         add_bio_to_retry(raid_bi, conf);
3232 }
3233
3234 static int bio_fits_rdev(struct bio *bi)
3235 {
3236         struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3237
3238         if ((bi->bi_size>>9) > q->max_sectors)
3239                 return 0;
3240         blk_recount_segments(q, bi);
3241         if (bi->bi_phys_segments > q->max_phys_segments)
3242                 return 0;
3243
3244         if (q->merge_bvec_fn)
3245                 /* it's too hard to apply the merge_bvec_fn at this stage,
3246                  * just just give up
3247                  */
3248                 return 0;
3249
3250         return 1;
3251 }
3252
3253
3254 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3255 {
3256         mddev_t *mddev = q->queuedata;
3257         raid5_conf_t *conf = mddev_to_conf(mddev);
3258         const unsigned int raid_disks = conf->raid_disks;
3259         const unsigned int data_disks = raid_disks - conf->max_degraded;
3260         unsigned int dd_idx, pd_idx;
3261         struct bio* align_bi;
3262         mdk_rdev_t *rdev;
3263
3264         if (!in_chunk_boundary(mddev, raid_bio)) {
3265                 pr_debug("chunk_aligned_read : non aligned\n");
3266                 return 0;
3267         }
3268         /*
3269          * use bio_clone to make a copy of the bio
3270          */
3271         align_bi = bio_clone(raid_bio, GFP_NOIO);
3272         if (!align_bi)
3273                 return 0;
3274         /*
3275          *   set bi_end_io to a new function, and set bi_private to the
3276          *     original bio.
3277          */
3278         align_bi->bi_end_io  = raid5_align_endio;
3279         align_bi->bi_private = raid_bio;
3280         /*
3281          *      compute position
3282          */
3283         align_bi->bi_sector =  raid5_compute_sector(raid_bio->bi_sector,
3284                                         raid_disks,
3285                                         data_disks,
3286                                         &dd_idx,
3287                                         &pd_idx,
3288                                         conf);
3289
3290         rcu_read_lock();
3291         rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3292         if (rdev && test_bit(In_sync, &rdev->flags)) {
3293                 atomic_inc(&rdev->nr_pending);
3294                 rcu_read_unlock();
3295                 raid_bio->bi_next = (void*)rdev;
3296                 align_bi->bi_bdev =  rdev->bdev;
3297                 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3298                 align_bi->bi_sector += rdev->data_offset;
3299
3300                 if (!bio_fits_rdev(align_bi)) {
3301                         /* too big in some way */
3302                         bio_put(align_bi);
3303                         rdev_dec_pending(rdev, mddev);
3304                         return 0;
3305                 }
3306
3307                 spin_lock_irq(&conf->device_lock);
3308                 wait_event_lock_irq(conf->wait_for_stripe,
3309                                     conf->quiesce == 0,
3310                                     conf->device_lock, /* nothing */);
3311                 atomic_inc(&conf->active_aligned_reads);
3312                 spin_unlock_irq(&conf->device_lock);
3313
3314                 generic_make_request(align_bi);
3315                 return 1;
3316         } else {
3317                 rcu_read_unlock();
3318                 bio_put(align_bi);
3319                 return 0;
3320         }
3321 }
3322
3323 /* __get_priority_stripe - get the next stripe to process
3324  *
3325  * Full stripe writes are allowed to pass preread active stripes up until
3326  * the bypass_threshold is exceeded.  In general the bypass_count
3327  * increments when the handle_list is handled before the hold_list; however, it
3328  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3329  * stripe with in flight i/o.  The bypass_count will be reset when the
3330  * head of the hold_list has changed, i.e. the head was promoted to the
3331  * handle_list.
3332  */
3333 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3334 {
3335         struct stripe_head *sh;
3336
3337         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3338                   __func__,
3339                   list_empty(&conf->handle_list) ? "empty" : "busy",
3340                   list_empty(&conf->hold_list) ? "empty" : "busy",
3341                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
3342
3343         if (!list_empty(&conf->handle_list)) {
3344                 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3345
3346                 if (list_empty(&conf->hold_list))
3347                         conf->bypass_count = 0;
3348                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3349                         if (conf->hold_list.next == conf->last_hold)
3350                                 conf->bypass_count++;
3351                         else {
3352                                 conf->last_hold = conf->hold_list.next;
3353                                 conf->bypass_count -= conf->bypass_threshold;
3354                                 if (conf->bypass_count < 0)
3355                                         conf->bypass_count = 0;
3356                         }
3357                 }
3358         } else if (!list_empty(&conf->hold_list) &&
3359                    ((conf->bypass_threshold &&
3360                      conf->bypass_count > conf->bypass_threshold) ||
3361                     atomic_read(&conf->pending_full_writes) == 0)) {
3362                 sh = list_entry(conf->hold_list.next,
3363                                 typeof(*sh), lru);
3364                 conf->bypass_count -= conf->bypass_threshold;
3365                 if (conf->bypass_count < 0)
3366                         conf->bypass_count = 0;
3367         } else
3368                 return NULL;
3369
3370         list_del_init(&sh->lru);
3371         atomic_inc(&sh->count);
3372         BUG_ON(atomic_read(&sh->count) != 1);
3373         return sh;
3374 }
3375
3376 static int make_request(struct request_queue *q, struct bio * bi)
3377 {
3378         mddev_t *mddev = q->queuedata;
3379         raid5_conf_t *conf = mddev_to_conf(mddev);
3380         unsigned int dd_idx, pd_idx;
3381         sector_t new_sector;
3382         sector_t logical_sector, last_sector;
3383         struct stripe_head *sh;
3384         const int rw = bio_data_dir(bi);
3385         int cpu, remaining;
3386
3387         if (unlikely(bio_barrier(bi))) {
3388                 bio_endio(bi, -EOPNOTSUPP);
3389                 return 0;
3390         }
3391
3392         md_write_start(mddev, bi);
3393
3394         cpu = part_stat_lock();
3395         part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
3396         part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
3397                       bio_sectors(bi));
3398         part_stat_unlock();
3399
3400         if (rw == READ &&
3401              mddev->reshape_position == MaxSector &&
3402              chunk_aligned_read(q,bi))
3403                 return 0;
3404
3405         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3406         last_sector = bi->bi_sector + (bi->bi_size>>9);
3407         bi->bi_next = NULL;
3408         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
3409
3410         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3411                 DEFINE_WAIT(w);
3412                 int disks, data_disks;
3413
3414         retry:
3415                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3416                 if (likely(conf->expand_progress == MaxSector))
3417                         disks = conf->raid_disks;
3418                 else {
3419                         /* spinlock is needed as expand_progress may be
3420                          * 64bit on a 32bit platform, and so it might be
3421                          * possible to see a half-updated value
3422                          * Ofcourse expand_progress could change after
3423                          * the lock is dropped, so once we get a reference
3424                          * to the stripe that we think it is, we will have
3425                          * to check again.
3426                          */
3427                         spin_lock_irq(&conf->device_lock);
3428                         disks = conf->raid_disks;
3429                         if (logical_sector >= conf->expand_progress)
3430                                 disks = conf->previous_raid_disks;
3431                         else {
3432                                 if (logical_sector >= conf->expand_lo) {
3433                                         spin_unlock_irq(&conf->device_lock);
3434                                         schedule();
3435                                         goto retry;
3436                                 }
3437                         }
3438                         spin_unlock_irq(&conf->device_lock);
3439                 }
3440                 data_disks = disks - conf->max_degraded;
3441
3442                 new_sector = raid5_compute_sector(logical_sector, disks, data_disks,
3443                                                   &dd_idx, &pd_idx, conf);
3444                 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3445                         (unsigned long long)new_sector, 
3446                         (unsigned long long)logical_sector);
3447
3448                 sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
3449                 if (sh) {
3450                         if (unlikely(conf->expand_progress != MaxSector)) {
3451                                 /* expansion might have moved on while waiting for a
3452                                  * stripe, so we must do the range check again.
3453                                  * Expansion could still move past after this
3454                                  * test, but as we are holding a reference to
3455                                  * 'sh', we know that if that happens,
3456                                  *  STRIPE_EXPANDING will get set and the expansion
3457                                  * won't proceed until we finish with the stripe.
3458                                  */
3459                                 int must_retry = 0;
3460                                 spin_lock_irq(&conf->device_lock);
3461                                 if (logical_sector <  conf->expand_progress &&
3462                                     disks == conf->previous_raid_disks)
3463                                         /* mismatch, need to try again */
3464                                         must_retry = 1;
3465                                 spin_unlock_irq(&conf->device_lock);
3466                                 if (must_retry) {
3467                                         release_stripe(sh);
3468                                         goto retry;
3469                                 }
3470                         }
3471                         /* FIXME what if we get a false positive because these
3472                          * are being updated.
3473                          */
3474                         if (logical_sector >= mddev->suspend_lo &&
3475                             logical_sector < mddev->suspend_hi) {
3476                                 release_stripe(sh);
3477                                 schedule();
3478                                 goto retry;
3479                         }
3480
3481                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3482                             !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3483                                 /* Stripe is busy expanding or
3484                                  * add failed due to overlap.  Flush everything
3485                                  * and wait a while
3486                                  */
3487                                 raid5_unplug_device(mddev->queue);
3488                                 release_stripe(sh);
3489                                 schedule();
3490                                 goto retry;
3491                         }
3492                         finish_wait(&conf->wait_for_overlap, &w);
3493                         set_bit(STRIPE_HANDLE, &sh->state);
3494                         clear_bit(STRIPE_DELAYED, &sh->state);
3495                         release_stripe(sh);
3496                 } else {
3497                         /* cannot get stripe for read-ahead, just give-up */
3498                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
3499                         finish_wait(&conf->wait_for_overlap, &w);
3500                         break;
3501                 }
3502                         
3503         }
3504         spin_lock_irq(&conf->device_lock);
3505         remaining = raid5_dec_bi_phys_segments(bi);
3506         spin_unlock_irq(&conf->device_lock);
3507         if (remaining == 0) {
3508
3509                 if ( rw == WRITE )
3510                         md_write_end(mddev);
3511
3512                 bio_endio(bi, 0);
3513         }
3514         return 0;
3515 }
3516
3517 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3518 {
3519         /* reshaping is quite different to recovery/resync so it is
3520          * handled quite separately ... here.
3521          *
3522          * On each call to sync_request, we gather one chunk worth of
3523          * destination stripes and flag them as expanding.
3524          * Then we find all the source stripes and request reads.
3525          * As the reads complete, handle_stripe will copy the data
3526          * into the destination stripe and release that stripe.
3527          */
3528         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3529         struct stripe_head *sh;
3530         int pd_idx;
3531         sector_t first_sector, last_sector;
3532         int raid_disks = conf->previous_raid_disks;
3533         int data_disks = raid_disks - conf->max_degraded;
3534         int new_data_disks = conf->raid_disks - conf->max_degraded;
3535         int i;
3536         int dd_idx;
3537         sector_t writepos, safepos, gap;
3538
3539         if (sector_nr == 0 &&
3540             conf->expand_progress != 0) {
3541                 /* restarting in the middle, skip the initial sectors */
3542                 sector_nr = conf->expand_progress;
3543                 sector_div(sector_nr, new_data_disks);
3544                 *skipped = 1;
3545                 return sector_nr;
3546         }
3547
3548         /* we update the metadata when there is more than 3Meg
3549          * in the block range (that is rather arbitrary, should
3550          * probably be time based) or when the data about to be
3551          * copied would over-write the source of the data at
3552          * the front of the range.
3553          * i.e. one new_stripe forward from expand_progress new_maps
3554          * to after where expand_lo old_maps to
3555          */
3556         writepos = conf->expand_progress +
3557                 conf->chunk_size/512*(new_data_disks);
3558         sector_div(writepos, new_data_disks);
3559         safepos = conf->expand_lo;
3560         sector_div(safepos, data_disks);
3561         gap = conf->expand_progress - conf->expand_lo;
3562
3563         if (writepos >= safepos ||
3564             gap > (new_data_disks)*3000*2 /*3Meg*/) {
3565                 /* Cannot proceed until we've updated the superblock... */
3566                 wait_event(conf->wait_for_overlap,
3567                            atomic_read(&conf->reshape_stripes)==0);
3568                 mddev->reshape_position = conf->expand_progress;
3569                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3570                 md_wakeup_thread(mddev->thread);
3571                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3572                            kthread_should_stop());
3573                 spin_lock_irq(&conf->device_lock);
3574                 conf->expand_lo = mddev->reshape_position;
3575                 spin_unlock_irq(&conf->device_lock);
3576                 wake_up(&conf->wait_for_overlap);
3577         }
3578
3579         for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
3580                 int j;
3581                 int skipped = 0;
3582                 pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks);
3583                 sh = get_active_stripe(conf, sector_nr+i,
3584                                        conf->raid_disks, pd_idx, 0);
3585                 set_bit(STRIPE_EXPANDING, &sh->state);
3586                 atomic_inc(&conf->reshape_stripes);
3587                 /* If any of this stripe is beyond the end of the old
3588                  * array, then we need to zero those blocks
3589                  */
3590                 for (j=sh->disks; j--;) {
3591                         sector_t s;
3592                         if (j == sh->pd_idx)
3593                                 continue;
3594                         if (conf->level == 6 &&
3595                             j == raid6_next_disk(sh->pd_idx, sh->disks))
3596                                 continue;
3597                         s = compute_blocknr(sh, j);
3598                         if (s < mddev->array_sectors) {
3599                                 skipped = 1;
3600                                 continue;
3601                         }
3602                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3603                         set_bit(R5_Expanded, &sh->dev[j].flags);
3604                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
3605                 }
3606                 if (!skipped) {
3607                         set_bit(STRIPE_EXPAND_READY, &sh->state);
3608                         set_bit(STRIPE_HANDLE, &sh->state);
3609                 }
3610                 release_stripe(sh);
3611         }
3612         spin_lock_irq(&conf->device_lock);
3613         conf->expand_progress = (sector_nr + i) * new_data_disks;
3614         spin_unlock_irq(&conf->device_lock);
3615         /* Ok, those stripe are ready. We can start scheduling
3616          * reads on the source stripes.
3617          * The source stripes are determined by mapping the first and last
3618          * block on the destination stripes.
3619          */
3620         first_sector =
3621                 raid5_compute_sector(sector_nr*(new_data_disks),
3622                                      raid_disks, data_disks,
3623                                      &dd_idx, &pd_idx, conf);
3624         last_sector =
3625                 raid5_compute_sector((sector_nr+conf->chunk_size/512)
3626                                      *(new_data_disks) -1,
3627                                      raid_disks, data_disks,
3628                                      &dd_idx, &pd_idx, conf);
3629         if (last_sector >= (mddev->size<<1))
3630                 last_sector = (mddev->size<<1)-1;
3631         while (first_sector <= last_sector) {
3632                 pd_idx = stripe_to_pdidx(first_sector, conf,
3633                                          conf->previous_raid_disks);
3634                 sh = get_active_stripe(conf, first_sector,
3635                                        conf->previous_raid_disks, pd_idx, 0);
3636                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3637                 set_bit(STRIPE_HANDLE, &sh->state);
3638                 release_stripe(sh);
3639                 first_sector += STRIPE_SECTORS;
3640         }
3641         /* If this takes us to the resync_max point where we have to pause,
3642          * then we need to write out the superblock.
3643          */
3644         sector_nr += conf->chunk_size>>9;
3645         if (sector_nr >= mddev->resync_max) {
3646                 /* Cannot proceed until we've updated the superblock... */
3647                 wait_event(conf->wait_for_overlap,
3648                            atomic_read(&conf->reshape_stripes) == 0);
3649                 mddev->reshape_position = conf->expand_progress;
3650                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3651                 md_wakeup_thread(mddev->thread);
3652                 wait_event(mddev->sb_wait,
3653                            !test_bit(MD_CHANGE_DEVS, &mddev->flags)
3654                            || kthread_should_stop());
3655                 spin_lock_irq(&conf->device_lock);
3656                 conf->expand_lo = mddev->reshape_position;
3657                 spin_unlock_irq(&conf->device_lock);
3658                 wake_up(&conf->wait_for_overlap);
3659         }
3660         return conf->chunk_size>>9;
3661 }
3662
3663 /* FIXME go_faster isn't used */
3664 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3665 {
3666         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3667         struct stripe_head *sh;
3668         int pd_idx;
3669         int raid_disks = conf->raid_disks;
3670         sector_t max_sector = mddev->size << 1;
3671         int sync_blocks;
3672         int still_degraded = 0;
3673         int i;
3674
3675         if (sector_nr >= max_sector) {
3676                 /* just being told to finish up .. nothing much to do */
3677                 unplug_slaves(mddev);
3678                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3679                         end_reshape(conf);
3680                         return 0;
3681                 }
3682
3683                 if (mddev->curr_resync < max_sector) /* aborted */
3684                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3685                                         &sync_blocks, 1);
3686                 else /* completed sync */
3687                         conf->fullsync = 0;
3688                 bitmap_close_sync(mddev->bitmap);
3689
3690                 return 0;
3691         }
3692
3693         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3694                 return reshape_request(mddev, sector_nr, skipped);
3695
3696         /* No need to check resync_max as we never do more than one
3697          * stripe, and as resync_max will always be on a chunk boundary,
3698          * if the check in md_do_sync didn't fire, there is no chance
3699          * of overstepping resync_max here
3700          */
3701
3702         /* if there is too many failed drives and we are trying
3703          * to resync, then assert that we are finished, because there is
3704          * nothing we can do.
3705          */
3706         if (mddev->degraded >= conf->max_degraded &&
3707             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3708                 sector_t rv = (mddev->size << 1) - sector_nr;
3709                 *skipped = 1;
3710                 return rv;
3711         }
3712         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3713             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3714             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3715                 /* we can skip this block, and probably more */
3716                 sync_blocks /= STRIPE_SECTORS;
3717                 *skipped = 1;
3718                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3719         }
3720
3721
3722         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3723
3724         pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks);
3725         sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1);
3726         if (sh == NULL) {
3727                 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0);
3728                 /* make sure we don't swamp the stripe cache if someone else
3729                  * is trying to get access
3730                  */
3731                 schedule_timeout_uninterruptible(1);
3732         }
3733         /* Need to check if array will still be degraded after recovery/resync
3734          * We don't need to check the 'failed' flag as when that gets set,
3735          * recovery aborts.
3736          */
3737         for (i=0; i<mddev->raid_disks; i++)
3738                 if (conf->disks[i].rdev == NULL)
3739                         still_degraded = 1;
3740
3741         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
3742
3743         spin_lock(&sh->lock);
3744         set_bit(STRIPE_SYNCING, &sh->state);
3745         clear_bit(STRIPE_INSYNC, &sh->state);
3746         spin_unlock(&sh->lock);
3747
3748         /* wait for any blocked device to be handled */
3749         while(unlikely(!handle_stripe(sh, NULL)))
3750                 ;
3751         release_stripe(sh);
3752
3753         return STRIPE_SECTORS;
3754 }
3755
3756 static int  retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
3757 {
3758         /* We may not be able to submit a whole bio at once as there
3759          * may not be enough stripe_heads available.
3760          * We cannot pre-allocate enough stripe_heads as we may need
3761          * more than exist in the cache (if we allow ever large chunks).
3762          * So we do one stripe head at a time and record in
3763          * ->bi_hw_segments how many have been done.
3764          *
3765          * We *know* that this entire raid_bio is in one chunk, so
3766          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3767          */
3768         struct stripe_head *sh;
3769         int dd_idx, pd_idx;
3770         sector_t sector, logical_sector, last_sector;
3771         int scnt = 0;
3772         int remaining;
3773         int handled = 0;
3774
3775         logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3776         sector = raid5_compute_sector(  logical_sector,
3777                                         conf->raid_disks,
3778                                         conf->raid_disks - conf->max_degraded,
3779                                         &dd_idx,
3780                                         &pd_idx,
3781                                         conf);
3782         last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
3783
3784         for (; logical_sector < last_sector;
3785              logical_sector += STRIPE_SECTORS,
3786                      sector += STRIPE_SECTORS,
3787                      scnt++) {
3788
3789                 if (scnt < raid5_bi_hw_segments(raid_bio))
3790                         /* already done this stripe */
3791                         continue;
3792
3793                 sh = get_active_stripe(conf, sector, conf->raid_disks, pd_idx, 1);
3794
3795                 if (!sh) {
3796                         /* failed to get a stripe - must wait */
3797                         raid5_set_bi_hw_segments(raid_bio, scnt);
3798                         conf->retry_read_aligned = raid_bio;
3799                         return handled;
3800                 }
3801
3802                 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
3803                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
3804                         release_stripe(sh);
3805                         raid5_set_bi_hw_segments(raid_bio, scnt);
3806                         conf->retry_read_aligned = raid_bio;
3807                         return handled;
3808                 }
3809
3810                 handle_stripe(sh, NULL);
3811                 release_stripe(sh);
3812                 handled++;
3813         }
3814         spin_lock_irq(&conf->device_lock);
3815         remaining = raid5_dec_bi_phys_segments(raid_bio);
3816         spin_unlock_irq(&conf->device_lock);
3817         if (remaining == 0)
3818                 bio_endio(raid_bio, 0);
3819         if (atomic_dec_and_test(&conf->active_aligned_reads))
3820                 wake_up(&conf->wait_for_stripe);
3821         return handled;
3822 }
3823
3824
3825
3826 /*
3827  * This is our raid5 kernel thread.
3828  *
3829  * We scan the hash table for stripes which can be handled now.
3830  * During the scan, completed stripes are saved for us by the interrupt
3831  * handler, so that they will not have to wait for our next wakeup.
3832  */
3833 static void raid5d(mddev_t *mddev)
3834 {
3835         struct stripe_head *sh;
3836         raid5_conf_t *conf = mddev_to_conf(mddev);
3837         int handled;
3838
3839         pr_debug("+++ raid5d active\n");
3840
3841         md_check_recovery(mddev);
3842
3843         handled = 0;
3844         spin_lock_irq(&conf->device_lock);
3845         while (1) {
3846                 struct bio *bio;
3847
3848                 if (conf->seq_flush != conf->seq_write) {
3849                         int seq = conf->seq_flush;
3850                         spin_unlock_irq(&conf->device_lock);
3851                         bitmap_unplug(mddev->bitmap);
3852                         spin_lock_irq(&conf->device_lock);
3853                         conf->seq_write = seq;
3854                         activate_bit_delay(conf);
3855                 }
3856
3857                 while ((bio = remove_bio_from_retry(conf))) {
3858                         int ok;
3859                         spin_unlock_irq(&conf->device_lock);
3860                         ok = retry_aligned_read(conf, bio);
3861                         spin_lock_irq(&conf->device_lock);
3862                         if (!ok)
3863                                 break;
3864                         handled++;
3865                 }
3866
3867                 sh = __get_priority_stripe(conf);
3868
3869                 if (!sh)
3870                         break;
3871                 spin_unlock_irq(&conf->device_lock);
3872                 
3873                 handled++;
3874                 handle_stripe(sh, conf->spare_page);
3875                 release_stripe(sh);
3876
3877                 spin_lock_irq(&conf->device_lock);
3878         }
3879         pr_debug("%d stripes handled\n", handled);
3880
3881         spin_unlock_irq(&conf->device_lock);
3882
3883         async_tx_issue_pending_all();
3884         unplug_slaves(mddev);
3885
3886         pr_debug("--- raid5d inactive\n");
3887 }
3888
3889 static ssize_t
3890 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
3891 {
3892         raid5_conf_t *conf = mddev_to_conf(mddev);
3893         if (conf)
3894                 return sprintf(page, "%d\n", conf->max_nr_stripes);
3895         else
3896                 return 0;
3897 }
3898
3899 static ssize_t
3900 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
3901 {
3902         raid5_conf_t *conf = mddev_to_conf(mddev);
3903         unsigned long new;
3904         int err;
3905
3906         if (len >= PAGE_SIZE)
3907                 return -EINVAL;
3908         if (!conf)
3909                 return -ENODEV;
3910
3911         if (strict_strtoul(page, 10, &new))
3912                 return -EINVAL;
3913         if (new <= 16 || new > 32768)
3914                 return -EINVAL;
3915         while (new < conf->max_nr_stripes) {
3916                 if (drop_one_stripe(conf))
3917                         conf->max_nr_stripes--;
3918                 else
3919                         break;
3920         }
3921         err = md_allow_write(mddev);
3922         if (err)
3923                 return err;
3924         while (new > conf->max_nr_stripes) {
3925                 if (grow_one_stripe(conf))
3926                         conf->max_nr_stripes++;
3927                 else break;
3928         }
3929         return len;
3930 }
3931
3932 static struct md_sysfs_entry
3933 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
3934                                 raid5_show_stripe_cache_size,
3935                                 raid5_store_stripe_cache_size);
3936
3937 static ssize_t
3938 raid5_show_preread_threshold(mddev_t *mddev, char *page)
3939 {
3940         raid5_conf_t *conf = mddev_to_conf(mddev);
3941         if (conf)
3942                 return sprintf(page, "%d\n", conf->bypass_threshold);
3943         else
3944                 return 0;
3945 }
3946
3947 static ssize_t
3948 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
3949 {
3950         raid5_conf_t *conf = mddev_to_conf(mddev);
3951         unsigned long new;
3952         if (len >= PAGE_SIZE)
3953                 return -EINVAL;
3954         if (!conf)
3955                 return -ENODEV;
3956
3957         if (strict_strtoul(page, 10, &new))
3958                 return -EINVAL;
3959         if (new > conf->max_nr_stripes)
3960                 return -EINVAL;
3961         conf->bypass_threshold = new;
3962         return len;
3963 }
3964
3965 static struct md_sysfs_entry
3966 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
3967                                         S_IRUGO | S_IWUSR,
3968                                         raid5_show_preread_threshold,
3969                                         raid5_store_preread_threshold);
3970
3971 static ssize_t
3972 stripe_cache_active_show(mddev_t *mddev, char *page)
3973 {
3974         raid5_conf_t *conf = mddev_to_conf(mddev);
3975         if (conf)
3976                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
3977         else
3978                 return 0;
3979 }
3980
3981 static struct md_sysfs_entry
3982 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
3983
3984 static struct attribute *raid5_attrs[] =  {
3985         &raid5_stripecache_size.attr,
3986         &raid5_stripecache_active.attr,
3987         &raid5_preread_bypass_threshold.attr,
3988         NULL,
3989 };
3990 static struct attribute_group raid5_attrs_group = {
3991         .name = NULL,
3992         .attrs = raid5_attrs,
3993 };
3994
3995 static int run(mddev_t *mddev)
3996 {
3997         raid5_conf_t *conf;
3998         int raid_disk, memory;
3999         mdk_rdev_t *rdev;
4000         struct disk_info *disk;
4001         int working_disks = 0;
4002
4003         if (mddev->level != 5 && mddev->level != 4 && mddev->level != 6) {
4004                 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4005                        mdname(mddev), mddev->level);
4006                 return -EIO;
4007         }
4008
4009         if (mddev->chunk_size < PAGE_SIZE) {
4010                 printk(KERN_ERR "md/raid5: chunk_size must be at least "
4011                        "PAGE_SIZE but %d < %ld\n",
4012                        mddev->chunk_size, PAGE_SIZE);
4013                 return -EINVAL;
4014         }
4015
4016         if (mddev->reshape_position != MaxSector) {
4017                 /* Check that we can continue the reshape.
4018                  * Currently only disks can change, it must
4019                  * increase, and we must be past the point where
4020                  * a stripe over-writes itself
4021                  */
4022                 sector_t here_new, here_old;
4023                 int old_disks;
4024                 int max_degraded = (mddev->level == 5 ? 1 : 2);
4025
4026                 if (mddev->new_level != mddev->level ||
4027                     mddev->new_layout != mddev->layout ||
4028                     mddev->new_chunk != mddev->chunk_size) {
4029                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4030                                "required - aborting.\n",
4031                                mdname(mddev));
4032                         return -EINVAL;
4033                 }
4034                 if (mddev->delta_disks <= 0) {
4035                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4036                                "(reduce disks) required - aborting.\n",
4037                                mdname(mddev));
4038                         return -EINVAL;
4039                 }
4040                 old_disks = mddev->raid_disks - mddev->delta_disks;
4041                 /* reshape_position must be on a new-stripe boundary, and one
4042                  * further up in new geometry must map after here in old
4043                  * geometry.
4044                  */
4045                 here_new = mddev->reshape_position;
4046                 if (sector_div(here_new, (mddev->chunk_size>>9)*
4047                                (mddev->raid_disks - max_degraded))) {
4048                         printk(KERN_ERR "raid5: reshape_position not "
4049                                "on a stripe boundary\n");
4050                         return -EINVAL;
4051                 }
4052                 /* here_new is the stripe we will write to */
4053                 here_old = mddev->reshape_position;
4054                 sector_div(here_old, (mddev->chunk_size>>9)*
4055                            (old_disks-max_degraded));
4056                 /* here_old is the first stripe that we might need to read
4057                  * from */
4058                 if (here_new >= here_old) {
4059                         /* Reading from the same stripe as writing to - bad */
4060                         printk(KERN_ERR "raid5: reshape_position too early for "
4061                                "auto-recovery - aborting.\n");
4062                         return -EINVAL;
4063                 }
4064                 printk(KERN_INFO "raid5: reshape will continue\n");
4065                 /* OK, we should be able to continue; */
4066         }
4067
4068
4069         mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
4070         if ((conf = mddev->private) == NULL)
4071                 goto abort;
4072         if (mddev->reshape_position == MaxSector) {
4073                 conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks;
4074         } else {
4075                 conf->raid_disks = mddev->raid_disks;
4076                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4077         }
4078
4079         conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4080                               GFP_KERNEL);
4081         if (!conf->disks)
4082                 goto abort;
4083
4084         conf->mddev = mddev;
4085
4086         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4087                 goto abort;
4088
4089         if (mddev->level == 6) {
4090                 conf->spare_page = alloc_page(GFP_KERNEL);
4091                 if (!conf->spare_page)
4092                         goto abort;
4093         }
4094         spin_lock_init(&conf->device_lock);
4095         mddev->queue->queue_lock = &conf->device_lock;
4096         init_waitqueue_head(&conf->wait_for_stripe);
4097         init_waitqueue_head(&conf->wait_for_overlap);
4098         INIT_LIST_HEAD(&conf->handle_list);
4099         INIT_LIST_HEAD(&conf->hold_list);
4100         INIT_LIST_HEAD(&conf->delayed_list);
4101         INIT_LIST_HEAD(&conf->bitmap_list);
4102         INIT_LIST_HEAD(&conf->inactive_list);
4103         atomic_set(&conf->active_stripes, 0);
4104         atomic_set(&conf->preread_active_stripes, 0);
4105         atomic_set(&conf->active_aligned_reads, 0);
4106         conf->bypass_threshold = BYPASS_THRESHOLD;
4107
4108         pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4109
4110         list_for_each_entry(rdev, &mddev->disks, same_set) {
4111                 raid_disk = rdev->raid_disk;
4112                 if (raid_disk >= conf->raid_disks
4113                     || raid_disk < 0)
4114                         continue;
4115                 disk = conf->disks + raid_disk;
4116
4117                 disk->rdev = rdev;
4118
4119                 if (test_bit(In_sync, &rdev->flags)) {
4120                         char b[BDEVNAME_SIZE];
4121                         printk(KERN_INFO "raid5: device %s operational as raid"
4122                                 " disk %d\n", bdevname(rdev->bdev,b),
4123                                 raid_disk);
4124                         working_disks++;
4125                 } else
4126                         /* Cannot rely on bitmap to complete recovery */
4127                         conf->fullsync = 1;
4128         }
4129
4130         /*
4131          * 0 for a fully functional array, 1 or 2 for a degraded array.
4132          */
4133         mddev->degraded = conf->raid_disks - working_disks;
4134         conf->mddev = mddev;
4135         conf->chunk_size = mddev->chunk_size;
4136         conf->level = mddev->level;
4137         if (conf->level == 6)
4138                 conf->max_degraded = 2;
4139         else
4140                 conf->max_degraded = 1;
4141         conf->algorithm = mddev->layout;
4142         conf->max_nr_stripes = NR_STRIPES;
4143         conf->expand_progress = mddev->reshape_position;
4144
4145         /* device size must be a multiple of chunk size */
4146         mddev->size &= ~(mddev->chunk_size/1024 -1);
4147         mddev->resync_max_sectors = mddev->size << 1;
4148
4149         if (conf->level == 6 && conf->raid_disks < 4) {
4150                 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4151                        mdname(mddev), conf->raid_disks);
4152                 goto abort;
4153         }
4154         if (!conf->chunk_size || conf->chunk_size % 4) {
4155                 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4156                         conf->chunk_size, mdname(mddev));
4157                 goto abort;
4158         }
4159         if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
4160                 printk(KERN_ERR 
4161                         "raid5: unsupported parity algorithm %d for %s\n",
4162                         conf->algorithm, mdname(mddev));
4163                 goto abort;
4164         }
4165         if (mddev->degraded > conf->max_degraded) {
4166                 printk(KERN_ERR "raid5: not enough operational devices for %s"
4167                         " (%d/%d failed)\n",
4168                         mdname(mddev), mddev->degraded, conf->raid_disks);
4169                 goto abort;
4170         }
4171
4172         if (mddev->degraded > 0 &&
4173             mddev->recovery_cp != MaxSector) {
4174                 if (mddev->ok_start_degraded)
4175                         printk(KERN_WARNING
4176                                "raid5: starting dirty degraded array: %s"
4177                                "- data corruption possible.\n",
4178                                mdname(mddev));
4179                 else {
4180                         printk(KERN_ERR
4181                                "raid5: cannot start dirty degraded array for %s\n",
4182                                mdname(mddev));
4183                         goto abort;
4184                 }
4185         }
4186
4187         {
4188                 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4189                 if (!mddev->thread) {
4190                         printk(KERN_ERR 
4191                                 "raid5: couldn't allocate thread for %s\n",
4192                                 mdname(mddev));
4193                         goto abort;
4194                 }
4195         }
4196         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4197                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4198         if (grow_stripes(conf, conf->max_nr_stripes)) {
4199                 printk(KERN_ERR 
4200                         "raid5: couldn't allocate %dkB for buffers\n", memory);
4201                 shrink_stripes(conf);
4202                 md_unregister_thread(mddev->thread);
4203                 goto abort;
4204         } else
4205                 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4206                         memory, mdname(mddev));
4207
4208         if (mddev->degraded == 0)
4209                 printk("raid5: raid level %d set %s active with %d out of %d"
4210                         " devices, algorithm %d\n", conf->level, mdname(mddev), 
4211                         mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4212                         conf->algorithm);
4213         else
4214                 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4215                         " out of %d devices, algorithm %d\n", conf->level,
4216                         mdname(mddev), mddev->raid_disks - mddev->degraded,
4217                         mddev->raid_disks, conf->algorithm);
4218
4219         print_raid5_conf(conf);
4220
4221         if (conf->expand_progress != MaxSector) {
4222                 printk("...ok start reshape thread\n");
4223                 conf->expand_lo = conf->expand_progress;
4224                 atomic_set(&conf->reshape_stripes, 0);
4225                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4226                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4227                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4228                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4229                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4230                                                         "%s_reshape");
4231         }
4232
4233         /* read-ahead size must cover two whole stripes, which is
4234          * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4235          */
4236         {
4237                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4238                 int stripe = data_disks *
4239                         (mddev->chunk_size / PAGE_SIZE);
4240                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4241                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4242         }
4243
4244         /* Ok, everything is just fine now */
4245         if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4246                 printk(KERN_WARNING
4247                        "raid5: failed to create sysfs attributes for %s\n",
4248                        mdname(mddev));
4249
4250         mddev->queue->unplug_fn = raid5_unplug_device;
4251         mddev->queue->backing_dev_info.congested_data = mddev;
4252         mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4253
4254         mddev->array_sectors = 2 * mddev->size * (conf->previous_raid_disks -
4255                                             conf->max_degraded);
4256
4257         blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4258
4259         return 0;
4260 abort:
4261         if (conf) {
4262                 print_raid5_conf(conf);
4263                 safe_put_page(conf->spare_page);
4264                 kfree(conf->disks);
4265                 kfree(conf->stripe_hashtbl);
4266                 kfree(conf);
4267         }
4268         mddev->private = NULL;
4269         printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4270         return -EIO;
4271 }
4272
4273
4274
4275 static int stop(mddev_t *mddev)
4276 {
4277         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4278
4279         md_unregister_thread(mddev->thread);
4280         mddev->thread = NULL;
4281         shrink_stripes(conf);
4282         kfree(conf->stripe_hashtbl);
4283         mddev->queue->backing_dev_info.congested_fn = NULL;
4284         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4285         sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4286         kfree(conf->disks);
4287         kfree(conf);
4288         mddev->private = NULL;
4289         return 0;
4290 }
4291
4292 #ifdef DEBUG
4293 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
4294 {
4295         int i;
4296
4297         seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4298                    (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4299         seq_printf(seq, "sh %llu,  count %d.\n",
4300                    (unsigned long long)sh->sector, atomic_read(&sh->count));
4301         seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4302         for (i = 0; i < sh->disks; i++) {
4303                 seq_printf(seq, "(cache%d: %p %ld) ",
4304                            i, sh->dev[i].page, sh->dev[i].flags);
4305         }
4306         seq_printf(seq, "\n");
4307 }
4308
4309 static void printall(struct seq_file *seq, raid5_conf_t *conf)
4310 {
4311         struct stripe_head *sh;
4312         struct hlist_node *hn;
4313         int i;
4314
4315         spin_lock_irq(&conf->device_lock);
4316         for (i = 0; i < NR_HASH; i++) {
4317                 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4318                         if (sh->raid_conf != conf)
4319                                 continue;
4320                         print_sh(seq, sh);
4321                 }
4322         }
4323         spin_unlock_irq(&conf->device_lock);
4324 }
4325 #endif
4326
4327 static void status(struct seq_file *seq, mddev_t *mddev)
4328 {
4329         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4330         int i;
4331
4332         seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4333         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4334         for (i = 0; i < conf->raid_disks; i++)
4335                 seq_printf (seq, "%s",
4336                                conf->disks[i].rdev &&
4337                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4338         seq_printf (seq, "]");
4339 #ifdef DEBUG
4340         seq_printf (seq, "\n");
4341         printall(seq, conf);
4342 #endif
4343 }
4344
4345 static void print_raid5_conf (raid5_conf_t *conf)
4346 {
4347         int i;
4348         struct disk_info *tmp;
4349
4350         printk("RAID5 conf printout:\n");
4351         if (!conf) {
4352                 printk("(conf==NULL)\n");
4353                 return;
4354         }
4355         printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4356                  conf->raid_disks - conf->mddev->degraded);
4357
4358         for (i = 0; i < conf->raid_disks; i++) {
4359                 char b[BDEVNAME_SIZE];
4360                 tmp = conf->disks + i;
4361                 if (tmp->rdev)
4362                 printk(" disk %d, o:%d, dev:%s\n",
4363                         i, !test_bit(Faulty, &tmp->rdev->flags),
4364                         bdevname(tmp->rdev->bdev,b));
4365         }
4366 }
4367
4368 static int raid5_spare_active(mddev_t *mddev)
4369 {
4370         int i;
4371         raid5_conf_t *conf = mddev->private;
4372         struct disk_info *tmp;
4373
4374         for (i = 0; i < conf->raid_disks; i++) {
4375                 tmp = conf->disks + i;
4376                 if (tmp->rdev
4377                     && !test_bit(Faulty, &tmp->rdev->flags)
4378                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4379                         unsigned long flags;
4380                         spin_lock_irqsave(&conf->device_lock, flags);
4381                         mddev->degraded--;
4382                         spin_unlock_irqrestore(&conf->device_lock, flags);
4383                 }
4384         }
4385         print_raid5_conf(conf);
4386         return 0;
4387 }
4388
4389 static int raid5_remove_disk(mddev_t *mddev, int number)
4390 {
4391         raid5_conf_t *conf = mddev->private;
4392         int err = 0;
4393         mdk_rdev_t *rdev;
4394         struct disk_info *p = conf->disks + number;
4395
4396         print_raid5_conf(conf);
4397         rdev = p->rdev;
4398         if (rdev) {
4399                 if (test_bit(In_sync, &rdev->flags) ||
4400                     atomic_read(&rdev->nr_pending)) {
4401                         err = -EBUSY;
4402                         goto abort;
4403                 }
4404                 /* Only remove non-faulty devices if recovery
4405                  * isn't possible.
4406                  */
4407                 if (!test_bit(Faulty, &rdev->flags) &&
4408                     mddev->degraded <= conf->max_degraded) {
4409                         err = -EBUSY;
4410                         goto abort;
4411                 }
4412                 p->rdev = NULL;
4413                 synchronize_rcu();
4414                 if (atomic_read(&rdev->nr_pending)) {
4415                         /* lost the race, try later */
4416                         err = -EBUSY;
4417                         p->rdev = rdev;
4418                 }
4419         }
4420 abort:
4421
4422         print_raid5_conf(conf);
4423         return err;
4424 }
4425
4426 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4427 {
4428         raid5_conf_t *conf = mddev->private;
4429         int err = -EEXIST;
4430         int disk;
4431         struct disk_info *p;
4432         int first = 0;
4433         int last = conf->raid_disks - 1;
4434
4435         if (mddev->degraded > conf->max_degraded)
4436                 /* no point adding a device */
4437                 return -EINVAL;
4438
4439         if (rdev->raid_disk >= 0)
4440                 first = last = rdev->raid_disk;
4441
4442         /*
4443          * find the disk ... but prefer rdev->saved_raid_disk
4444          * if possible.
4445          */
4446         if (rdev->saved_raid_disk >= 0 &&
4447             rdev->saved_raid_disk >= first &&
4448             conf->disks[rdev->saved_raid_disk].rdev == NULL)
4449                 disk = rdev->saved_raid_disk;
4450         else
4451                 disk = first;
4452         for ( ; disk <= last ; disk++)
4453                 if ((p=conf->disks + disk)->rdev == NULL) {
4454                         clear_bit(In_sync, &rdev->flags);
4455                         rdev->raid_disk = disk;
4456                         err = 0;
4457                         if (rdev->saved_raid_disk != disk)
4458                                 conf->fullsync = 1;
4459                         rcu_assign_pointer(p->rdev, rdev);
4460                         break;
4461                 }
4462         print_raid5_conf(conf);
4463         return err;
4464 }
4465
4466 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4467 {
4468         /* no resync is happening, and there is enough space
4469          * on all devices, so we can resize.
4470          * We need to make sure resync covers any new space.
4471          * If the array is shrinking we should possibly wait until
4472          * any io in the removed space completes, but it hardly seems
4473          * worth it.
4474          */
4475         raid5_conf_t *conf = mddev_to_conf(mddev);
4476
4477         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4478         mddev->array_sectors = sectors * (mddev->raid_disks
4479                                           - conf->max_degraded);
4480         set_capacity(mddev->gendisk, mddev->array_sectors);
4481         mddev->changed = 1;
4482         if (sectors/2  > mddev->size && mddev->recovery_cp == MaxSector) {
4483                 mddev->recovery_cp = mddev->size << 1;
4484                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4485         }
4486         mddev->size = sectors /2;
4487         mddev->resync_max_sectors = sectors;
4488         return 0;
4489 }
4490
4491 #ifdef CONFIG_MD_RAID5_RESHAPE
4492 static int raid5_check_reshape(mddev_t *mddev)
4493 {
4494         raid5_conf_t *conf = mddev_to_conf(mddev);
4495         int err;
4496
4497         if (mddev->delta_disks < 0 ||
4498             mddev->new_level != mddev->level)
4499                 return -EINVAL; /* Cannot shrink array or change level yet */
4500         if (mddev->delta_disks == 0)
4501                 return 0; /* nothing to do */
4502         if (mddev->bitmap)
4503                 /* Cannot grow a bitmap yet */
4504                 return -EBUSY;
4505
4506         /* Can only proceed if there are plenty of stripe_heads.
4507          * We need a minimum of one full stripe,, and for sensible progress
4508          * it is best to have about 4 times that.
4509          * If we require 4 times, then the default 256 4K stripe_heads will
4510          * allow for chunk sizes up to 256K, which is probably OK.
4511          * If the chunk size is greater, user-space should request more
4512          * stripe_heads first.
4513          */
4514         if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
4515             (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
4516                 printk(KERN_WARNING "raid5: reshape: not enough stripes.  Needed %lu\n",
4517                        (mddev->chunk_size / STRIPE_SIZE)*4);
4518                 return -ENOSPC;
4519         }
4520
4521         err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4522         if (err)
4523                 return err;
4524
4525         if (mddev->degraded > conf->max_degraded)
4526                 return -EINVAL;
4527         /* looks like we might be able to manage this */
4528         return 0;
4529 }
4530
4531 static int raid5_start_reshape(mddev_t *mddev)
4532 {
4533         raid5_conf_t *conf = mddev_to_conf(mddev);
4534         mdk_rdev_t *rdev;
4535         int spares = 0;
4536         int added_devices = 0;
4537         unsigned long flags;
4538
4539         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4540                 return -EBUSY;
4541
4542         list_for_each_entry(rdev, &mddev->disks, same_set)
4543                 if (rdev->raid_disk < 0 &&
4544                     !test_bit(Faulty, &rdev->flags))
4545                         spares++;
4546
4547         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4548                 /* Not enough devices even to make a degraded array
4549                  * of that size
4550                  */
4551                 return -EINVAL;
4552
4553         atomic_set(&conf->reshape_stripes, 0);
4554         spin_lock_irq(&conf->device_lock);
4555         conf->previous_raid_disks = conf->raid_disks;
4556         conf->raid_disks += mddev->delta_disks;
4557         conf->expand_progress = 0;
4558         conf->expand_lo = 0;
4559         spin_unlock_irq(&conf->device_lock);
4560
4561         /* Add some new drives, as many as will fit.
4562          * We know there are enough to make the newly sized array work.
4563          */
4564         list_for_each_entry(rdev, &mddev->disks, same_set)
4565                 if (rdev->raid_disk < 0 &&
4566                     !test_bit(Faulty, &rdev->flags)) {
4567                         if (raid5_add_disk(mddev, rdev) == 0) {
4568                                 char nm[20];
4569                                 set_bit(In_sync, &rdev->flags);
4570                                 added_devices++;
4571                                 rdev->recovery_offset = 0;
4572                                 sprintf(nm, "rd%d", rdev->raid_disk);
4573                                 if (sysfs_create_link(&mddev->kobj,
4574                                                       &rdev->kobj, nm))
4575                                         printk(KERN_WARNING
4576                                                "raid5: failed to create "
4577                                                " link %s for %s\n",
4578                                                nm, mdname(mddev));
4579                         } else
4580                                 break;
4581                 }
4582
4583         spin_lock_irqsave(&conf->device_lock, flags);
4584         mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
4585         spin_unlock_irqrestore(&conf->device_lock, flags);
4586         mddev->raid_disks = conf->raid_disks;
4587         mddev->reshape_position = 0;
4588         set_bit(MD_CHANGE_DEVS, &mddev->flags);
4589
4590         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4591         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4592         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4593         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4594         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4595                                                 "%s_reshape");
4596         if (!mddev->sync_thread) {
4597                 mddev->recovery = 0;
4598                 spin_lock_irq(&conf->device_lock);
4599                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
4600                 conf->expand_progress = MaxSector;
4601                 spin_unlock_irq(&conf->device_lock);
4602                 return -EAGAIN;
4603         }
4604         md_wakeup_thread(mddev->sync_thread);
4605         md_new_event(mddev);
4606         return 0;
4607 }
4608 #endif
4609
4610 static void end_reshape(raid5_conf_t *conf)
4611 {
4612         struct block_device *bdev;
4613
4614         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
4615                 conf->mddev->array_sectors = 2 * conf->mddev->size *
4616                         (conf->raid_disks - conf->max_degraded);
4617                 set_capacity(conf->mddev->gendisk, conf->mddev->array_sectors);
4618                 conf->mddev->changed = 1;
4619
4620                 bdev = bdget_disk(conf->mddev->gendisk, 0);
4621                 if (bdev) {
4622                         mutex_lock(&bdev->bd_inode->i_mutex);
4623                         i_size_write(bdev->bd_inode,
4624                                      (loff_t)conf->mddev->array_sectors << 9);
4625                         mutex_unlock(&bdev->bd_inode->i_mutex);
4626                         bdput(bdev);
4627                 }
4628                 spin_lock_irq(&conf->device_lock);
4629                 conf->expand_progress = MaxSector;
4630                 spin_unlock_irq(&conf->device_lock);
4631                 conf->mddev->reshape_position = MaxSector;
4632
4633                 /* read-ahead size must cover two whole stripes, which is
4634                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4635                  */
4636                 {
4637                         int data_disks = conf->previous_raid_disks - conf->max_degraded;
4638                         int stripe = data_disks *
4639                                 (conf->mddev->chunk_size / PAGE_SIZE);
4640                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4641                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4642                 }
4643         }
4644 }
4645
4646 static void raid5_quiesce(mddev_t *mddev, int state)
4647 {
4648         raid5_conf_t *conf = mddev_to_conf(mddev);
4649
4650         switch(state) {
4651         case 2: /* resume for a suspend */
4652                 wake_up(&conf->wait_for_overlap);
4653                 break;
4654
4655         case 1: /* stop all writes */
4656                 spin_lock_irq(&conf->device_lock);
4657                 conf->quiesce = 1;
4658                 wait_event_lock_irq(conf->wait_for_stripe,
4659                                     atomic_read(&conf->active_stripes) == 0 &&
4660                                     atomic_read(&conf->active_aligned_reads) == 0,
4661                                     conf->device_lock, /* nothing */);
4662                 spin_unlock_irq(&conf->device_lock);
4663                 break;
4664
4665         case 0: /* re-enable writes */
4666                 spin_lock_irq(&conf->device_lock);
4667                 conf->quiesce = 0;
4668                 wake_up(&conf->wait_for_stripe);
4669                 wake_up(&conf->wait_for_overlap);
4670                 spin_unlock_irq(&conf->device_lock);
4671                 break;
4672         }
4673 }
4674
4675 static struct mdk_personality raid6_personality =
4676 {
4677         .name           = "raid6",
4678         .level          = 6,
4679         .owner          = THIS_MODULE,
4680         .make_request   = make_request,
4681         .run            = run,
4682         .stop           = stop,
4683         .status         = status,
4684         .error_handler  = error,
4685         .hot_add_disk   = raid5_add_disk,
4686         .hot_remove_disk= raid5_remove_disk,
4687         .spare_active   = raid5_spare_active,
4688         .sync_request   = sync_request,
4689         .resize         = raid5_resize,
4690 #ifdef CONFIG_MD_RAID5_RESHAPE
4691         .check_reshape  = raid5_check_reshape,
4692         .start_reshape  = raid5_start_reshape,
4693 #endif
4694         .quiesce        = raid5_quiesce,
4695 };
4696 static struct mdk_personality raid5_personality =
4697 {
4698         .name           = "raid5",
4699         .level          = 5,
4700         .owner          = THIS_MODULE,
4701         .make_request   = make_request,
4702         .run            = run,
4703         .stop           = stop,
4704         .status         = status,
4705         .error_handler  = error,
4706         .hot_add_disk   = raid5_add_disk,
4707         .hot_remove_disk= raid5_remove_disk,
4708         .spare_active   = raid5_spare_active,
4709         .sync_request   = sync_request,
4710         .resize         = raid5_resize,
4711 #ifdef CONFIG_MD_RAID5_RESHAPE
4712         .check_reshape  = raid5_check_reshape,
4713         .start_reshape  = raid5_start_reshape,
4714 #endif
4715         .quiesce        = raid5_quiesce,
4716 };
4717
4718 static struct mdk_personality raid4_personality =
4719 {
4720         .name           = "raid4",
4721         .level          = 4,
4722         .owner          = THIS_MODULE,
4723         .make_request   = make_request,
4724         .run            = run,
4725         .stop           = stop,
4726         .status         = status,
4727         .error_handler  = error,
4728         .hot_add_disk   = raid5_add_disk,
4729         .hot_remove_disk= raid5_remove_disk,
4730         .spare_active   = raid5_spare_active,
4731         .sync_request   = sync_request,
4732         .resize         = raid5_resize,
4733 #ifdef CONFIG_MD_RAID5_RESHAPE
4734         .check_reshape  = raid5_check_reshape,
4735         .start_reshape  = raid5_start_reshape,
4736 #endif
4737         .quiesce        = raid5_quiesce,
4738 };
4739
4740 static int __init raid5_init(void)
4741 {
4742         int e;
4743
4744         e = raid6_select_algo();
4745         if ( e )
4746                 return e;
4747         register_md_personality(&raid6_personality);
4748         register_md_personality(&raid5_personality);
4749         register_md_personality(&raid4_personality);
4750         return 0;
4751 }
4752
4753 static void raid5_exit(void)
4754 {
4755         unregister_md_personality(&raid6_personality);
4756         unregister_md_personality(&raid5_personality);
4757         unregister_md_personality(&raid4_personality);
4758 }
4759
4760 module_init(raid5_init);
4761 module_exit(raid5_exit);
4762 MODULE_LICENSE("GPL");
4763 MODULE_ALIAS("md-personality-4"); /* RAID5 */
4764 MODULE_ALIAS("md-raid5");
4765 MODULE_ALIAS("md-raid4");
4766 MODULE_ALIAS("md-level-5");
4767 MODULE_ALIAS("md-level-4");
4768 MODULE_ALIAS("md-personality-8"); /* RAID6 */
4769 MODULE_ALIAS("md-raid6");
4770 MODULE_ALIAS("md-level-6");
4771
4772 /* This used to be two separate modules, they were: */
4773 MODULE_ALIAS("raid5");
4774 MODULE_ALIAS("raid6");