declance: Restore tx descriptor ring locking
[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/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/seq_file.h>
51 #include "md.h"
52 #include "raid5.h"
53 #include "bitmap.h"
54
55 /*
56  * Stripe cache
57  */
58
59 #define NR_STRIPES              256
60 #define STRIPE_SIZE             PAGE_SIZE
61 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
62 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
63 #define IO_THRESHOLD            1
64 #define BYPASS_THRESHOLD        1
65 #define NR_HASH                 (PAGE_SIZE / sizeof(struct hlist_head))
66 #define HASH_MASK               (NR_HASH - 1)
67
68 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
69
70 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
71  * order without overlap.  There may be several bio's per stripe+device, and
72  * a bio could span several devices.
73  * When walking this list for a particular stripe+device, we must never proceed
74  * beyond a bio that extends past this device, as the next bio might no longer
75  * be valid.
76  * This macro is used to determine the 'next' bio in the list, given the sector
77  * of the current stripe+device
78  */
79 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
80 /*
81  * The following can be used to debug the driver
82  */
83 #define RAID5_PARANOIA  1
84 #if RAID5_PARANOIA && defined(CONFIG_SMP)
85 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
86 #else
87 # define CHECK_DEVLOCK()
88 #endif
89
90 #ifdef DEBUG
91 #define inline
92 #define __inline__
93 #endif
94
95 #define printk_rl(args...) ((void) (printk_ratelimit() && printk(args)))
96
97 /*
98  * We maintain a biased count of active stripes in the bottom 16 bits of
99  * bi_phys_segments, and a count of processed stripes in the upper 16 bits
100  */
101 static inline int raid5_bi_phys_segments(struct bio *bio)
102 {
103         return bio->bi_phys_segments & 0xffff;
104 }
105
106 static inline int raid5_bi_hw_segments(struct bio *bio)
107 {
108         return (bio->bi_phys_segments >> 16) & 0xffff;
109 }
110
111 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
112 {
113         --bio->bi_phys_segments;
114         return raid5_bi_phys_segments(bio);
115 }
116
117 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
118 {
119         unsigned short val = raid5_bi_hw_segments(bio);
120
121         --val;
122         bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
123         return val;
124 }
125
126 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
127 {
128         bio->bi_phys_segments = raid5_bi_phys_segments(bio) || (cnt << 16);
129 }
130
131 /* Find first data disk in a raid6 stripe */
132 static inline int raid6_d0(struct stripe_head *sh)
133 {
134         if (sh->ddf_layout)
135                 /* ddf always start from first device */
136                 return 0;
137         /* md starts just after Q block */
138         if (sh->qd_idx == sh->disks - 1)
139                 return 0;
140         else
141                 return sh->qd_idx + 1;
142 }
143 static inline int raid6_next_disk(int disk, int raid_disks)
144 {
145         disk++;
146         return (disk < raid_disks) ? disk : 0;
147 }
148
149 /* When walking through the disks in a raid5, starting at raid6_d0,
150  * We need to map each disk to a 'slot', where the data disks are slot
151  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
152  * is raid_disks-1.  This help does that mapping.
153  */
154 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
155                              int *count, int syndrome_disks)
156 {
157         int slot;
158
159         if (idx == sh->pd_idx)
160                 return syndrome_disks;
161         if (idx == sh->qd_idx)
162                 return syndrome_disks + 1;
163         slot = (*count)++;
164         return slot;
165 }
166
167 static void return_io(struct bio *return_bi)
168 {
169         struct bio *bi = return_bi;
170         while (bi) {
171
172                 return_bi = bi->bi_next;
173                 bi->bi_next = NULL;
174                 bi->bi_size = 0;
175                 bio_endio(bi, 0);
176                 bi = return_bi;
177         }
178 }
179
180 static void print_raid5_conf (raid5_conf_t *conf);
181
182 static int stripe_operations_active(struct stripe_head *sh)
183 {
184         return sh->check_state || sh->reconstruct_state ||
185                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
186                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
187 }
188
189 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
190 {
191         if (atomic_dec_and_test(&sh->count)) {
192                 BUG_ON(!list_empty(&sh->lru));
193                 BUG_ON(atomic_read(&conf->active_stripes)==0);
194                 if (test_bit(STRIPE_HANDLE, &sh->state)) {
195                         if (test_bit(STRIPE_DELAYED, &sh->state)) {
196                                 list_add_tail(&sh->lru, &conf->delayed_list);
197                                 blk_plug_device(conf->mddev->queue);
198                         } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
199                                    sh->bm_seq - conf->seq_write > 0) {
200                                 list_add_tail(&sh->lru, &conf->bitmap_list);
201                                 blk_plug_device(conf->mddev->queue);
202                         } else {
203                                 clear_bit(STRIPE_BIT_DELAY, &sh->state);
204                                 list_add_tail(&sh->lru, &conf->handle_list);
205                         }
206                         md_wakeup_thread(conf->mddev->thread);
207                 } else {
208                         BUG_ON(stripe_operations_active(sh));
209                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
210                                 atomic_dec(&conf->preread_active_stripes);
211                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
212                                         md_wakeup_thread(conf->mddev->thread);
213                         }
214                         atomic_dec(&conf->active_stripes);
215                         if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
216                                 list_add_tail(&sh->lru, &conf->inactive_list);
217                                 wake_up(&conf->wait_for_stripe);
218                                 if (conf->retry_read_aligned)
219                                         md_wakeup_thread(conf->mddev->thread);
220                         }
221                 }
222         }
223 }
224
225 static void release_stripe(struct stripe_head *sh)
226 {
227         raid5_conf_t *conf = sh->raid_conf;
228         unsigned long flags;
229
230         spin_lock_irqsave(&conf->device_lock, flags);
231         __release_stripe(conf, sh);
232         spin_unlock_irqrestore(&conf->device_lock, flags);
233 }
234
235 static inline void remove_hash(struct stripe_head *sh)
236 {
237         pr_debug("remove_hash(), stripe %llu\n",
238                 (unsigned long long)sh->sector);
239
240         hlist_del_init(&sh->hash);
241 }
242
243 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
244 {
245         struct hlist_head *hp = stripe_hash(conf, sh->sector);
246
247         pr_debug("insert_hash(), stripe %llu\n",
248                 (unsigned long long)sh->sector);
249
250         CHECK_DEVLOCK();
251         hlist_add_head(&sh->hash, hp);
252 }
253
254
255 /* find an idle stripe, make sure it is unhashed, and return it. */
256 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
257 {
258         struct stripe_head *sh = NULL;
259         struct list_head *first;
260
261         CHECK_DEVLOCK();
262         if (list_empty(&conf->inactive_list))
263                 goto out;
264         first = conf->inactive_list.next;
265         sh = list_entry(first, struct stripe_head, lru);
266         list_del_init(first);
267         remove_hash(sh);
268         atomic_inc(&conf->active_stripes);
269 out:
270         return sh;
271 }
272
273 static void shrink_buffers(struct stripe_head *sh, int num)
274 {
275         struct page *p;
276         int i;
277
278         for (i=0; i<num ; i++) {
279                 p = sh->dev[i].page;
280                 if (!p)
281                         continue;
282                 sh->dev[i].page = NULL;
283                 put_page(p);
284         }
285 }
286
287 static int grow_buffers(struct stripe_head *sh, int num)
288 {
289         int i;
290
291         for (i=0; i<num; i++) {
292                 struct page *page;
293
294                 if (!(page = alloc_page(GFP_KERNEL))) {
295                         return 1;
296                 }
297                 sh->dev[i].page = page;
298         }
299         return 0;
300 }
301
302 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
303 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
304                             struct stripe_head *sh);
305
306 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
307 {
308         raid5_conf_t *conf = sh->raid_conf;
309         int i;
310
311         BUG_ON(atomic_read(&sh->count) != 0);
312         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
313         BUG_ON(stripe_operations_active(sh));
314
315         CHECK_DEVLOCK();
316         pr_debug("init_stripe called, stripe %llu\n",
317                 (unsigned long long)sh->sector);
318
319         remove_hash(sh);
320
321         sh->generation = conf->generation - previous;
322         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
323         sh->sector = sector;
324         stripe_set_idx(sector, conf, previous, sh);
325         sh->state = 0;
326
327
328         for (i = sh->disks; i--; ) {
329                 struct r5dev *dev = &sh->dev[i];
330
331                 if (dev->toread || dev->read || dev->towrite || dev->written ||
332                     test_bit(R5_LOCKED, &dev->flags)) {
333                         printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
334                                (unsigned long long)sh->sector, i, dev->toread,
335                                dev->read, dev->towrite, dev->written,
336                                test_bit(R5_LOCKED, &dev->flags));
337                         BUG();
338                 }
339                 dev->flags = 0;
340                 raid5_build_block(sh, i, previous);
341         }
342         insert_hash(conf, sh);
343 }
344
345 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector,
346                                          short generation)
347 {
348         struct stripe_head *sh;
349         struct hlist_node *hn;
350
351         CHECK_DEVLOCK();
352         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
353         hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
354                 if (sh->sector == sector && sh->generation == generation)
355                         return sh;
356         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
357         return NULL;
358 }
359
360 static void unplug_slaves(mddev_t *mddev);
361 static void raid5_unplug_device(struct request_queue *q);
362
363 static struct stripe_head *
364 get_active_stripe(raid5_conf_t *conf, sector_t sector,
365                   int previous, int noblock)
366 {
367         struct stripe_head *sh;
368
369         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
370
371         spin_lock_irq(&conf->device_lock);
372
373         do {
374                 wait_event_lock_irq(conf->wait_for_stripe,
375                                     conf->quiesce == 0,
376                                     conf->device_lock, /* nothing */);
377                 sh = __find_stripe(conf, sector, conf->generation - previous);
378                 if (!sh) {
379                         if (!conf->inactive_blocked)
380                                 sh = get_free_stripe(conf);
381                         if (noblock && sh == NULL)
382                                 break;
383                         if (!sh) {
384                                 conf->inactive_blocked = 1;
385                                 wait_event_lock_irq(conf->wait_for_stripe,
386                                                     !list_empty(&conf->inactive_list) &&
387                                                     (atomic_read(&conf->active_stripes)
388                                                      < (conf->max_nr_stripes *3/4)
389                                                      || !conf->inactive_blocked),
390                                                     conf->device_lock,
391                                                     raid5_unplug_device(conf->mddev->queue)
392                                         );
393                                 conf->inactive_blocked = 0;
394                         } else
395                                 init_stripe(sh, sector, previous);
396                 } else {
397                         if (atomic_read(&sh->count)) {
398                                 BUG_ON(!list_empty(&sh->lru)
399                                     && !test_bit(STRIPE_EXPANDING, &sh->state));
400                         } else {
401                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
402                                         atomic_inc(&conf->active_stripes);
403                                 if (list_empty(&sh->lru) &&
404                                     !test_bit(STRIPE_EXPANDING, &sh->state))
405                                         BUG();
406                                 list_del_init(&sh->lru);
407                         }
408                 }
409         } while (sh == NULL);
410
411         if (sh)
412                 atomic_inc(&sh->count);
413
414         spin_unlock_irq(&conf->device_lock);
415         return sh;
416 }
417
418 static void
419 raid5_end_read_request(struct bio *bi, int error);
420 static void
421 raid5_end_write_request(struct bio *bi, int error);
422
423 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
424 {
425         raid5_conf_t *conf = sh->raid_conf;
426         int i, disks = sh->disks;
427
428         might_sleep();
429
430         for (i = disks; i--; ) {
431                 int rw;
432                 struct bio *bi;
433                 mdk_rdev_t *rdev;
434                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
435                         rw = WRITE;
436                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
437                         rw = READ;
438                 else
439                         continue;
440
441                 bi = &sh->dev[i].req;
442
443                 bi->bi_rw = rw;
444                 if (rw == WRITE)
445                         bi->bi_end_io = raid5_end_write_request;
446                 else
447                         bi->bi_end_io = raid5_end_read_request;
448
449                 rcu_read_lock();
450                 rdev = rcu_dereference(conf->disks[i].rdev);
451                 if (rdev && test_bit(Faulty, &rdev->flags))
452                         rdev = NULL;
453                 if (rdev)
454                         atomic_inc(&rdev->nr_pending);
455                 rcu_read_unlock();
456
457                 if (rdev) {
458                         if (s->syncing || s->expanding || s->expanded)
459                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
460
461                         set_bit(STRIPE_IO_STARTED, &sh->state);
462
463                         bi->bi_bdev = rdev->bdev;
464                         pr_debug("%s: for %llu schedule op %ld on disc %d\n",
465                                 __func__, (unsigned long long)sh->sector,
466                                 bi->bi_rw, i);
467                         atomic_inc(&sh->count);
468                         bi->bi_sector = sh->sector + rdev->data_offset;
469                         bi->bi_flags = 1 << BIO_UPTODATE;
470                         bi->bi_vcnt = 1;
471                         bi->bi_max_vecs = 1;
472                         bi->bi_idx = 0;
473                         bi->bi_io_vec = &sh->dev[i].vec;
474                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
475                         bi->bi_io_vec[0].bv_offset = 0;
476                         bi->bi_size = STRIPE_SIZE;
477                         bi->bi_next = NULL;
478                         if (rw == WRITE &&
479                             test_bit(R5_ReWrite, &sh->dev[i].flags))
480                                 atomic_add(STRIPE_SECTORS,
481                                         &rdev->corrected_errors);
482                         generic_make_request(bi);
483                 } else {
484                         if (rw == WRITE)
485                                 set_bit(STRIPE_DEGRADED, &sh->state);
486                         pr_debug("skip op %ld on disc %d for sector %llu\n",
487                                 bi->bi_rw, i, (unsigned long long)sh->sector);
488                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
489                         set_bit(STRIPE_HANDLE, &sh->state);
490                 }
491         }
492 }
493
494 static struct dma_async_tx_descriptor *
495 async_copy_data(int frombio, struct bio *bio, struct page *page,
496         sector_t sector, struct dma_async_tx_descriptor *tx)
497 {
498         struct bio_vec *bvl;
499         struct page *bio_page;
500         int i;
501         int page_offset;
502
503         if (bio->bi_sector >= sector)
504                 page_offset = (signed)(bio->bi_sector - sector) * 512;
505         else
506                 page_offset = (signed)(sector - bio->bi_sector) * -512;
507         bio_for_each_segment(bvl, bio, i) {
508                 int len = bio_iovec_idx(bio, i)->bv_len;
509                 int clen;
510                 int b_offset = 0;
511
512                 if (page_offset < 0) {
513                         b_offset = -page_offset;
514                         page_offset += b_offset;
515                         len -= b_offset;
516                 }
517
518                 if (len > 0 && page_offset + len > STRIPE_SIZE)
519                         clen = STRIPE_SIZE - page_offset;
520                 else
521                         clen = len;
522
523                 if (clen > 0) {
524                         b_offset += bio_iovec_idx(bio, i)->bv_offset;
525                         bio_page = bio_iovec_idx(bio, i)->bv_page;
526                         if (frombio)
527                                 tx = async_memcpy(page, bio_page, page_offset,
528                                         b_offset, clen,
529                                         ASYNC_TX_DEP_ACK,
530                                         tx, NULL, NULL);
531                         else
532                                 tx = async_memcpy(bio_page, page, b_offset,
533                                         page_offset, clen,
534                                         ASYNC_TX_DEP_ACK,
535                                         tx, NULL, NULL);
536                 }
537                 if (clen < len) /* hit end of page */
538                         break;
539                 page_offset +=  len;
540         }
541
542         return tx;
543 }
544
545 static void ops_complete_biofill(void *stripe_head_ref)
546 {
547         struct stripe_head *sh = stripe_head_ref;
548         struct bio *return_bi = NULL;
549         raid5_conf_t *conf = sh->raid_conf;
550         int i;
551
552         pr_debug("%s: stripe %llu\n", __func__,
553                 (unsigned long long)sh->sector);
554
555         /* clear completed biofills */
556         spin_lock_irq(&conf->device_lock);
557         for (i = sh->disks; i--; ) {
558                 struct r5dev *dev = &sh->dev[i];
559
560                 /* acknowledge completion of a biofill operation */
561                 /* and check if we need to reply to a read request,
562                  * new R5_Wantfill requests are held off until
563                  * !STRIPE_BIOFILL_RUN
564                  */
565                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
566                         struct bio *rbi, *rbi2;
567
568                         BUG_ON(!dev->read);
569                         rbi = dev->read;
570                         dev->read = NULL;
571                         while (rbi && rbi->bi_sector <
572                                 dev->sector + STRIPE_SECTORS) {
573                                 rbi2 = r5_next_bio(rbi, dev->sector);
574                                 if (!raid5_dec_bi_phys_segments(rbi)) {
575                                         rbi->bi_next = return_bi;
576                                         return_bi = rbi;
577                                 }
578                                 rbi = rbi2;
579                         }
580                 }
581         }
582         spin_unlock_irq(&conf->device_lock);
583         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
584
585         return_io(return_bi);
586
587         set_bit(STRIPE_HANDLE, &sh->state);
588         release_stripe(sh);
589 }
590
591 static void ops_run_biofill(struct stripe_head *sh)
592 {
593         struct dma_async_tx_descriptor *tx = NULL;
594         raid5_conf_t *conf = sh->raid_conf;
595         int i;
596
597         pr_debug("%s: stripe %llu\n", __func__,
598                 (unsigned long long)sh->sector);
599
600         for (i = sh->disks; i--; ) {
601                 struct r5dev *dev = &sh->dev[i];
602                 if (test_bit(R5_Wantfill, &dev->flags)) {
603                         struct bio *rbi;
604                         spin_lock_irq(&conf->device_lock);
605                         dev->read = rbi = dev->toread;
606                         dev->toread = NULL;
607                         spin_unlock_irq(&conf->device_lock);
608                         while (rbi && rbi->bi_sector <
609                                 dev->sector + STRIPE_SECTORS) {
610                                 tx = async_copy_data(0, rbi, dev->page,
611                                         dev->sector, tx);
612                                 rbi = r5_next_bio(rbi, dev->sector);
613                         }
614                 }
615         }
616
617         atomic_inc(&sh->count);
618         async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
619                 ops_complete_biofill, sh);
620 }
621
622 static void ops_complete_compute5(void *stripe_head_ref)
623 {
624         struct stripe_head *sh = stripe_head_ref;
625         int target = sh->ops.target;
626         struct r5dev *tgt = &sh->dev[target];
627
628         pr_debug("%s: stripe %llu\n", __func__,
629                 (unsigned long long)sh->sector);
630
631         set_bit(R5_UPTODATE, &tgt->flags);
632         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
633         clear_bit(R5_Wantcompute, &tgt->flags);
634         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
635         if (sh->check_state == check_state_compute_run)
636                 sh->check_state = check_state_compute_result;
637         set_bit(STRIPE_HANDLE, &sh->state);
638         release_stripe(sh);
639 }
640
641 static struct dma_async_tx_descriptor *ops_run_compute5(struct stripe_head *sh)
642 {
643         /* kernel stack size limits the total number of disks */
644         int disks = sh->disks;
645         struct page *xor_srcs[disks];
646         int target = sh->ops.target;
647         struct r5dev *tgt = &sh->dev[target];
648         struct page *xor_dest = tgt->page;
649         int count = 0;
650         struct dma_async_tx_descriptor *tx;
651         int i;
652
653         pr_debug("%s: stripe %llu block: %d\n",
654                 __func__, (unsigned long long)sh->sector, target);
655         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
656
657         for (i = disks; i--; )
658                 if (i != target)
659                         xor_srcs[count++] = sh->dev[i].page;
660
661         atomic_inc(&sh->count);
662
663         if (unlikely(count == 1))
664                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
665                         0, NULL, ops_complete_compute5, sh);
666         else
667                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
668                         ASYNC_TX_XOR_ZERO_DST, NULL,
669                         ops_complete_compute5, sh);
670
671         return tx;
672 }
673
674 static void ops_complete_prexor(void *stripe_head_ref)
675 {
676         struct stripe_head *sh = stripe_head_ref;
677
678         pr_debug("%s: stripe %llu\n", __func__,
679                 (unsigned long long)sh->sector);
680 }
681
682 static struct dma_async_tx_descriptor *
683 ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
684 {
685         /* kernel stack size limits the total number of disks */
686         int disks = sh->disks;
687         struct page *xor_srcs[disks];
688         int count = 0, pd_idx = sh->pd_idx, i;
689
690         /* existing parity data subtracted */
691         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
692
693         pr_debug("%s: stripe %llu\n", __func__,
694                 (unsigned long long)sh->sector);
695
696         for (i = disks; i--; ) {
697                 struct r5dev *dev = &sh->dev[i];
698                 /* Only process blocks that are known to be uptodate */
699                 if (test_bit(R5_Wantdrain, &dev->flags))
700                         xor_srcs[count++] = dev->page;
701         }
702
703         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
704                 ASYNC_TX_DEP_ACK | ASYNC_TX_XOR_DROP_DST, tx,
705                 ops_complete_prexor, sh);
706
707         return tx;
708 }
709
710 static struct dma_async_tx_descriptor *
711 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
712 {
713         int disks = sh->disks;
714         int i;
715
716         pr_debug("%s: stripe %llu\n", __func__,
717                 (unsigned long long)sh->sector);
718
719         for (i = disks; i--; ) {
720                 struct r5dev *dev = &sh->dev[i];
721                 struct bio *chosen;
722
723                 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
724                         struct bio *wbi;
725
726                         spin_lock(&sh->lock);
727                         chosen = dev->towrite;
728                         dev->towrite = NULL;
729                         BUG_ON(dev->written);
730                         wbi = dev->written = chosen;
731                         spin_unlock(&sh->lock);
732
733                         while (wbi && wbi->bi_sector <
734                                 dev->sector + STRIPE_SECTORS) {
735                                 tx = async_copy_data(1, wbi, dev->page,
736                                         dev->sector, tx);
737                                 wbi = r5_next_bio(wbi, dev->sector);
738                         }
739                 }
740         }
741
742         return tx;
743 }
744
745 static void ops_complete_postxor(void *stripe_head_ref)
746 {
747         struct stripe_head *sh = stripe_head_ref;
748         int disks = sh->disks, i, pd_idx = sh->pd_idx;
749
750         pr_debug("%s: stripe %llu\n", __func__,
751                 (unsigned long long)sh->sector);
752
753         for (i = disks; i--; ) {
754                 struct r5dev *dev = &sh->dev[i];
755                 if (dev->written || i == pd_idx)
756                         set_bit(R5_UPTODATE, &dev->flags);
757         }
758
759         if (sh->reconstruct_state == reconstruct_state_drain_run)
760                 sh->reconstruct_state = reconstruct_state_drain_result;
761         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
762                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
763         else {
764                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
765                 sh->reconstruct_state = reconstruct_state_result;
766         }
767
768         set_bit(STRIPE_HANDLE, &sh->state);
769         release_stripe(sh);
770 }
771
772 static void
773 ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
774 {
775         /* kernel stack size limits the total number of disks */
776         int disks = sh->disks;
777         struct page *xor_srcs[disks];
778
779         int count = 0, pd_idx = sh->pd_idx, i;
780         struct page *xor_dest;
781         int prexor = 0;
782         unsigned long flags;
783
784         pr_debug("%s: stripe %llu\n", __func__,
785                 (unsigned long long)sh->sector);
786
787         /* check if prexor is active which means only process blocks
788          * that are part of a read-modify-write (written)
789          */
790         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
791                 prexor = 1;
792                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
793                 for (i = disks; i--; ) {
794                         struct r5dev *dev = &sh->dev[i];
795                         if (dev->written)
796                                 xor_srcs[count++] = dev->page;
797                 }
798         } else {
799                 xor_dest = sh->dev[pd_idx].page;
800                 for (i = disks; i--; ) {
801                         struct r5dev *dev = &sh->dev[i];
802                         if (i != pd_idx)
803                                 xor_srcs[count++] = dev->page;
804                 }
805         }
806
807         /* 1/ if we prexor'd then the dest is reused as a source
808          * 2/ if we did not prexor then we are redoing the parity
809          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
810          * for the synchronous xor case
811          */
812         flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK |
813                 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
814
815         atomic_inc(&sh->count);
816
817         if (unlikely(count == 1)) {
818                 flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST);
819                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
820                         flags, tx, ops_complete_postxor, sh);
821         } else
822                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
823                         flags, tx, ops_complete_postxor, sh);
824 }
825
826 static void ops_complete_check(void *stripe_head_ref)
827 {
828         struct stripe_head *sh = stripe_head_ref;
829
830         pr_debug("%s: stripe %llu\n", __func__,
831                 (unsigned long long)sh->sector);
832
833         sh->check_state = check_state_check_result;
834         set_bit(STRIPE_HANDLE, &sh->state);
835         release_stripe(sh);
836 }
837
838 static void ops_run_check(struct stripe_head *sh)
839 {
840         /* kernel stack size limits the total number of disks */
841         int disks = sh->disks;
842         struct page *xor_srcs[disks];
843         struct dma_async_tx_descriptor *tx;
844
845         int count = 0, pd_idx = sh->pd_idx, i;
846         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
847
848         pr_debug("%s: stripe %llu\n", __func__,
849                 (unsigned long long)sh->sector);
850
851         for (i = disks; i--; ) {
852                 struct r5dev *dev = &sh->dev[i];
853                 if (i != pd_idx)
854                         xor_srcs[count++] = dev->page;
855         }
856
857         tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
858                 &sh->ops.zero_sum_result, 0, NULL, NULL, NULL);
859
860         atomic_inc(&sh->count);
861         tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
862                 ops_complete_check, sh);
863 }
864
865 static void raid5_run_ops(struct stripe_head *sh, unsigned long ops_request)
866 {
867         int overlap_clear = 0, i, disks = sh->disks;
868         struct dma_async_tx_descriptor *tx = NULL;
869
870         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
871                 ops_run_biofill(sh);
872                 overlap_clear++;
873         }
874
875         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
876                 tx = ops_run_compute5(sh);
877                 /* terminate the chain if postxor is not set to be run */
878                 if (tx && !test_bit(STRIPE_OP_POSTXOR, &ops_request))
879                         async_tx_ack(tx);
880         }
881
882         if (test_bit(STRIPE_OP_PREXOR, &ops_request))
883                 tx = ops_run_prexor(sh, tx);
884
885         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
886                 tx = ops_run_biodrain(sh, tx);
887                 overlap_clear++;
888         }
889
890         if (test_bit(STRIPE_OP_POSTXOR, &ops_request))
891                 ops_run_postxor(sh, tx);
892
893         if (test_bit(STRIPE_OP_CHECK, &ops_request))
894                 ops_run_check(sh);
895
896         if (overlap_clear)
897                 for (i = disks; i--; ) {
898                         struct r5dev *dev = &sh->dev[i];
899                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
900                                 wake_up(&sh->raid_conf->wait_for_overlap);
901                 }
902 }
903
904 static int grow_one_stripe(raid5_conf_t *conf)
905 {
906         struct stripe_head *sh;
907         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
908         if (!sh)
909                 return 0;
910         memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
911         sh->raid_conf = conf;
912         spin_lock_init(&sh->lock);
913
914         if (grow_buffers(sh, conf->raid_disks)) {
915                 shrink_buffers(sh, conf->raid_disks);
916                 kmem_cache_free(conf->slab_cache, sh);
917                 return 0;
918         }
919         sh->disks = conf->raid_disks;
920         /* we just created an active stripe so... */
921         atomic_set(&sh->count, 1);
922         atomic_inc(&conf->active_stripes);
923         INIT_LIST_HEAD(&sh->lru);
924         release_stripe(sh);
925         return 1;
926 }
927
928 static int grow_stripes(raid5_conf_t *conf, int num)
929 {
930         struct kmem_cache *sc;
931         int devs = conf->raid_disks;
932
933         sprintf(conf->cache_name[0],
934                 "raid%d-%s", conf->level, mdname(conf->mddev));
935         sprintf(conf->cache_name[1],
936                 "raid%d-%s-alt", conf->level, mdname(conf->mddev));
937         conf->active_name = 0;
938         sc = kmem_cache_create(conf->cache_name[conf->active_name],
939                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
940                                0, 0, NULL);
941         if (!sc)
942                 return 1;
943         conf->slab_cache = sc;
944         conf->pool_size = devs;
945         while (num--)
946                 if (!grow_one_stripe(conf))
947                         return 1;
948         return 0;
949 }
950
951 static int resize_stripes(raid5_conf_t *conf, int newsize)
952 {
953         /* Make all the stripes able to hold 'newsize' devices.
954          * New slots in each stripe get 'page' set to a new page.
955          *
956          * This happens in stages:
957          * 1/ create a new kmem_cache and allocate the required number of
958          *    stripe_heads.
959          * 2/ gather all the old stripe_heads and tranfer the pages across
960          *    to the new stripe_heads.  This will have the side effect of
961          *    freezing the array as once all stripe_heads have been collected,
962          *    no IO will be possible.  Old stripe heads are freed once their
963          *    pages have been transferred over, and the old kmem_cache is
964          *    freed when all stripes are done.
965          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
966          *    we simple return a failre status - no need to clean anything up.
967          * 4/ allocate new pages for the new slots in the new stripe_heads.
968          *    If this fails, we don't bother trying the shrink the
969          *    stripe_heads down again, we just leave them as they are.
970          *    As each stripe_head is processed the new one is released into
971          *    active service.
972          *
973          * Once step2 is started, we cannot afford to wait for a write,
974          * so we use GFP_NOIO allocations.
975          */
976         struct stripe_head *osh, *nsh;
977         LIST_HEAD(newstripes);
978         struct disk_info *ndisks;
979         int err;
980         struct kmem_cache *sc;
981         int i;
982
983         if (newsize <= conf->pool_size)
984                 return 0; /* never bother to shrink */
985
986         err = md_allow_write(conf->mddev);
987         if (err)
988                 return err;
989
990         /* Step 1 */
991         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
992                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
993                                0, 0, NULL);
994         if (!sc)
995                 return -ENOMEM;
996
997         for (i = conf->max_nr_stripes; i; i--) {
998                 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
999                 if (!nsh)
1000                         break;
1001
1002                 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1003
1004                 nsh->raid_conf = conf;
1005                 spin_lock_init(&nsh->lock);
1006
1007                 list_add(&nsh->lru, &newstripes);
1008         }
1009         if (i) {
1010                 /* didn't get enough, give up */
1011                 while (!list_empty(&newstripes)) {
1012                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
1013                         list_del(&nsh->lru);
1014                         kmem_cache_free(sc, nsh);
1015                 }
1016                 kmem_cache_destroy(sc);
1017                 return -ENOMEM;
1018         }
1019         /* Step 2 - Must use GFP_NOIO now.
1020          * OK, we have enough stripes, start collecting inactive
1021          * stripes and copying them over
1022          */
1023         list_for_each_entry(nsh, &newstripes, lru) {
1024                 spin_lock_irq(&conf->device_lock);
1025                 wait_event_lock_irq(conf->wait_for_stripe,
1026                                     !list_empty(&conf->inactive_list),
1027                                     conf->device_lock,
1028                                     unplug_slaves(conf->mddev)
1029                         );
1030                 osh = get_free_stripe(conf);
1031                 spin_unlock_irq(&conf->device_lock);
1032                 atomic_set(&nsh->count, 1);
1033                 for(i=0; i<conf->pool_size; i++)
1034                         nsh->dev[i].page = osh->dev[i].page;
1035                 for( ; i<newsize; i++)
1036                         nsh->dev[i].page = NULL;
1037                 kmem_cache_free(conf->slab_cache, osh);
1038         }
1039         kmem_cache_destroy(conf->slab_cache);
1040
1041         /* Step 3.
1042          * At this point, we are holding all the stripes so the array
1043          * is completely stalled, so now is a good time to resize
1044          * conf->disks.
1045          */
1046         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1047         if (ndisks) {
1048                 for (i=0; i<conf->raid_disks; i++)
1049                         ndisks[i] = conf->disks[i];
1050                 kfree(conf->disks);
1051                 conf->disks = ndisks;
1052         } else
1053                 err = -ENOMEM;
1054
1055         /* Step 4, return new stripes to service */
1056         while(!list_empty(&newstripes)) {
1057                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1058                 list_del_init(&nsh->lru);
1059                 for (i=conf->raid_disks; i < newsize; i++)
1060                         if (nsh->dev[i].page == NULL) {
1061                                 struct page *p = alloc_page(GFP_NOIO);
1062                                 nsh->dev[i].page = p;
1063                                 if (!p)
1064                                         err = -ENOMEM;
1065                         }
1066                 release_stripe(nsh);
1067         }
1068         /* critical section pass, GFP_NOIO no longer needed */
1069
1070         conf->slab_cache = sc;
1071         conf->active_name = 1-conf->active_name;
1072         conf->pool_size = newsize;
1073         return err;
1074 }
1075
1076 static int drop_one_stripe(raid5_conf_t *conf)
1077 {
1078         struct stripe_head *sh;
1079
1080         spin_lock_irq(&conf->device_lock);
1081         sh = get_free_stripe(conf);
1082         spin_unlock_irq(&conf->device_lock);
1083         if (!sh)
1084                 return 0;
1085         BUG_ON(atomic_read(&sh->count));
1086         shrink_buffers(sh, conf->pool_size);
1087         kmem_cache_free(conf->slab_cache, sh);
1088         atomic_dec(&conf->active_stripes);
1089         return 1;
1090 }
1091
1092 static void shrink_stripes(raid5_conf_t *conf)
1093 {
1094         while (drop_one_stripe(conf))
1095                 ;
1096
1097         if (conf->slab_cache)
1098                 kmem_cache_destroy(conf->slab_cache);
1099         conf->slab_cache = NULL;
1100 }
1101
1102 static void raid5_end_read_request(struct bio * bi, int error)
1103 {
1104         struct stripe_head *sh = bi->bi_private;
1105         raid5_conf_t *conf = sh->raid_conf;
1106         int disks = sh->disks, i;
1107         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1108         char b[BDEVNAME_SIZE];
1109         mdk_rdev_t *rdev;
1110
1111
1112         for (i=0 ; i<disks; i++)
1113                 if (bi == &sh->dev[i].req)
1114                         break;
1115
1116         pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1117                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1118                 uptodate);
1119         if (i == disks) {
1120                 BUG();
1121                 return;
1122         }
1123
1124         if (uptodate) {
1125                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1126                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1127                         rdev = conf->disks[i].rdev;
1128                         printk_rl(KERN_INFO "raid5:%s: read error corrected"
1129                                   " (%lu sectors at %llu on %s)\n",
1130                                   mdname(conf->mddev), STRIPE_SECTORS,
1131                                   (unsigned long long)(sh->sector
1132                                                        + rdev->data_offset),
1133                                   bdevname(rdev->bdev, b));
1134                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1135                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1136                 }
1137                 if (atomic_read(&conf->disks[i].rdev->read_errors))
1138                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
1139         } else {
1140                 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1141                 int retry = 0;
1142                 rdev = conf->disks[i].rdev;
1143
1144                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1145                 atomic_inc(&rdev->read_errors);
1146                 if (conf->mddev->degraded)
1147                         printk_rl(KERN_WARNING
1148                                   "raid5:%s: read error not correctable "
1149                                   "(sector %llu on %s).\n",
1150                                   mdname(conf->mddev),
1151                                   (unsigned long long)(sh->sector
1152                                                        + rdev->data_offset),
1153                                   bdn);
1154                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1155                         /* Oh, no!!! */
1156                         printk_rl(KERN_WARNING
1157                                   "raid5:%s: read error NOT corrected!! "
1158                                   "(sector %llu on %s).\n",
1159                                   mdname(conf->mddev),
1160                                   (unsigned long long)(sh->sector
1161                                                        + rdev->data_offset),
1162                                   bdn);
1163                 else if (atomic_read(&rdev->read_errors)
1164                          > conf->max_nr_stripes)
1165                         printk(KERN_WARNING
1166                                "raid5:%s: Too many read errors, failing device %s.\n",
1167                                mdname(conf->mddev), bdn);
1168                 else
1169                         retry = 1;
1170                 if (retry)
1171                         set_bit(R5_ReadError, &sh->dev[i].flags);
1172                 else {
1173                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1174                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1175                         md_error(conf->mddev, rdev);
1176                 }
1177         }
1178         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1179         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1180         set_bit(STRIPE_HANDLE, &sh->state);
1181         release_stripe(sh);
1182 }
1183
1184 static void raid5_end_write_request(struct bio *bi, int error)
1185 {
1186         struct stripe_head *sh = bi->bi_private;
1187         raid5_conf_t *conf = sh->raid_conf;
1188         int disks = sh->disks, i;
1189         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1190
1191         for (i=0 ; i<disks; i++)
1192                 if (bi == &sh->dev[i].req)
1193                         break;
1194
1195         pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1196                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1197                 uptodate);
1198         if (i == disks) {
1199                 BUG();
1200                 return;
1201         }
1202
1203         if (!uptodate)
1204                 md_error(conf->mddev, conf->disks[i].rdev);
1205
1206         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1207         
1208         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1209         set_bit(STRIPE_HANDLE, &sh->state);
1210         release_stripe(sh);
1211 }
1212
1213
1214 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1215         
1216 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1217 {
1218         struct r5dev *dev = &sh->dev[i];
1219
1220         bio_init(&dev->req);
1221         dev->req.bi_io_vec = &dev->vec;
1222         dev->req.bi_vcnt++;
1223         dev->req.bi_max_vecs++;
1224         dev->vec.bv_page = dev->page;
1225         dev->vec.bv_len = STRIPE_SIZE;
1226         dev->vec.bv_offset = 0;
1227
1228         dev->req.bi_sector = sh->sector;
1229         dev->req.bi_private = sh;
1230
1231         dev->flags = 0;
1232         dev->sector = compute_blocknr(sh, i, previous);
1233 }
1234
1235 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1236 {
1237         char b[BDEVNAME_SIZE];
1238         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1239         pr_debug("raid5: error called\n");
1240
1241         if (!test_bit(Faulty, &rdev->flags)) {
1242                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1243                 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1244                         unsigned long flags;
1245                         spin_lock_irqsave(&conf->device_lock, flags);
1246                         mddev->degraded++;
1247                         spin_unlock_irqrestore(&conf->device_lock, flags);
1248                         /*
1249                          * if recovery was running, make sure it aborts.
1250                          */
1251                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1252                 }
1253                 set_bit(Faulty, &rdev->flags);
1254                 printk(KERN_ALERT
1255                        "raid5: Disk failure on %s, disabling device.\n"
1256                        "raid5: Operation continuing on %d devices.\n",
1257                        bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1258         }
1259 }
1260
1261 /*
1262  * Input: a 'big' sector number,
1263  * Output: index of the data and parity disk, and the sector # in them.
1264  */
1265 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1266                                      int previous, int *dd_idx,
1267                                      struct stripe_head *sh)
1268 {
1269         long stripe;
1270         unsigned long chunk_number;
1271         unsigned int chunk_offset;
1272         int pd_idx, qd_idx;
1273         int ddf_layout = 0;
1274         sector_t new_sector;
1275         int algorithm = previous ? conf->prev_algo
1276                                  : conf->algorithm;
1277         int sectors_per_chunk = previous ? (conf->prev_chunk >> 9)
1278                                          : (conf->chunk_size >> 9);
1279         int raid_disks = previous ? conf->previous_raid_disks
1280                                   : conf->raid_disks;
1281         int data_disks = raid_disks - conf->max_degraded;
1282
1283         /* First compute the information on this sector */
1284
1285         /*
1286          * Compute the chunk number and the sector offset inside the chunk
1287          */
1288         chunk_offset = sector_div(r_sector, sectors_per_chunk);
1289         chunk_number = r_sector;
1290         BUG_ON(r_sector != chunk_number);
1291
1292         /*
1293          * Compute the stripe number
1294          */
1295         stripe = chunk_number / data_disks;
1296
1297         /*
1298          * Compute the data disk and parity disk indexes inside the stripe
1299          */
1300         *dd_idx = chunk_number % data_disks;
1301
1302         /*
1303          * Select the parity disk based on the user selected algorithm.
1304          */
1305         pd_idx = qd_idx = ~0;
1306         switch(conf->level) {
1307         case 4:
1308                 pd_idx = data_disks;
1309                 break;
1310         case 5:
1311                 switch (algorithm) {
1312                 case ALGORITHM_LEFT_ASYMMETRIC:
1313                         pd_idx = data_disks - stripe % raid_disks;
1314                         if (*dd_idx >= pd_idx)
1315                                 (*dd_idx)++;
1316                         break;
1317                 case ALGORITHM_RIGHT_ASYMMETRIC:
1318                         pd_idx = stripe % raid_disks;
1319                         if (*dd_idx >= pd_idx)
1320                                 (*dd_idx)++;
1321                         break;
1322                 case ALGORITHM_LEFT_SYMMETRIC:
1323                         pd_idx = data_disks - stripe % raid_disks;
1324                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1325                         break;
1326                 case ALGORITHM_RIGHT_SYMMETRIC:
1327                         pd_idx = stripe % raid_disks;
1328                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1329                         break;
1330                 case ALGORITHM_PARITY_0:
1331                         pd_idx = 0;
1332                         (*dd_idx)++;
1333                         break;
1334                 case ALGORITHM_PARITY_N:
1335                         pd_idx = data_disks;
1336                         break;
1337                 default:
1338                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1339                                 algorithm);
1340                         BUG();
1341                 }
1342                 break;
1343         case 6:
1344
1345                 switch (algorithm) {
1346                 case ALGORITHM_LEFT_ASYMMETRIC:
1347                         pd_idx = raid_disks - 1 - (stripe % raid_disks);
1348                         qd_idx = pd_idx + 1;
1349                         if (pd_idx == raid_disks-1) {
1350                                 (*dd_idx)++;    /* Q D D D P */
1351                                 qd_idx = 0;
1352                         } else if (*dd_idx >= pd_idx)
1353                                 (*dd_idx) += 2; /* D D P Q D */
1354                         break;
1355                 case ALGORITHM_RIGHT_ASYMMETRIC:
1356                         pd_idx = stripe % raid_disks;
1357                         qd_idx = pd_idx + 1;
1358                         if (pd_idx == raid_disks-1) {
1359                                 (*dd_idx)++;    /* Q D D D P */
1360                                 qd_idx = 0;
1361                         } else if (*dd_idx >= pd_idx)
1362                                 (*dd_idx) += 2; /* D D P Q D */
1363                         break;
1364                 case ALGORITHM_LEFT_SYMMETRIC:
1365                         pd_idx = raid_disks - 1 - (stripe % raid_disks);
1366                         qd_idx = (pd_idx + 1) % raid_disks;
1367                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1368                         break;
1369                 case ALGORITHM_RIGHT_SYMMETRIC:
1370                         pd_idx = stripe % raid_disks;
1371                         qd_idx = (pd_idx + 1) % raid_disks;
1372                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1373                         break;
1374
1375                 case ALGORITHM_PARITY_0:
1376                         pd_idx = 0;
1377                         qd_idx = 1;
1378                         (*dd_idx) += 2;
1379                         break;
1380                 case ALGORITHM_PARITY_N:
1381                         pd_idx = data_disks;
1382                         qd_idx = data_disks + 1;
1383                         break;
1384
1385                 case ALGORITHM_ROTATING_ZERO_RESTART:
1386                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
1387                          * of blocks for computing Q is different.
1388                          */
1389                         pd_idx = stripe % raid_disks;
1390                         qd_idx = pd_idx + 1;
1391                         if (pd_idx == raid_disks-1) {
1392                                 (*dd_idx)++;    /* Q D D D P */
1393                                 qd_idx = 0;
1394                         } else if (*dd_idx >= pd_idx)
1395                                 (*dd_idx) += 2; /* D D P Q D */
1396                         ddf_layout = 1;
1397                         break;
1398
1399                 case ALGORITHM_ROTATING_N_RESTART:
1400                         /* Same a left_asymmetric, by first stripe is
1401                          * D D D P Q  rather than
1402                          * Q D D D P
1403                          */
1404                         pd_idx = raid_disks - 1 - ((stripe + 1) % raid_disks);
1405                         qd_idx = pd_idx + 1;
1406                         if (pd_idx == raid_disks-1) {
1407                                 (*dd_idx)++;    /* Q D D D P */
1408                                 qd_idx = 0;
1409                         } else if (*dd_idx >= pd_idx)
1410                                 (*dd_idx) += 2; /* D D P Q D */
1411                         ddf_layout = 1;
1412                         break;
1413
1414                 case ALGORITHM_ROTATING_N_CONTINUE:
1415                         /* Same as left_symmetric but Q is before P */
1416                         pd_idx = raid_disks - 1 - (stripe % raid_disks);
1417                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1418                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1419                         ddf_layout = 1;
1420                         break;
1421
1422                 case ALGORITHM_LEFT_ASYMMETRIC_6:
1423                         /* RAID5 left_asymmetric, with Q on last device */
1424                         pd_idx = data_disks - stripe % (raid_disks-1);
1425                         if (*dd_idx >= pd_idx)
1426                                 (*dd_idx)++;
1427                         qd_idx = raid_disks - 1;
1428                         break;
1429
1430                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1431                         pd_idx = stripe % (raid_disks-1);
1432                         if (*dd_idx >= pd_idx)
1433                                 (*dd_idx)++;
1434                         qd_idx = raid_disks - 1;
1435                         break;
1436
1437                 case ALGORITHM_LEFT_SYMMETRIC_6:
1438                         pd_idx = data_disks - stripe % (raid_disks-1);
1439                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1440                         qd_idx = raid_disks - 1;
1441                         break;
1442
1443                 case ALGORITHM_RIGHT_SYMMETRIC_6:
1444                         pd_idx = stripe % (raid_disks-1);
1445                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1446                         qd_idx = raid_disks - 1;
1447                         break;
1448
1449                 case ALGORITHM_PARITY_0_6:
1450                         pd_idx = 0;
1451                         (*dd_idx)++;
1452                         qd_idx = raid_disks - 1;
1453                         break;
1454
1455
1456                 default:
1457                         printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1458                                algorithm);
1459                         BUG();
1460                 }
1461                 break;
1462         }
1463
1464         if (sh) {
1465                 sh->pd_idx = pd_idx;
1466                 sh->qd_idx = qd_idx;
1467                 sh->ddf_layout = ddf_layout;
1468         }
1469         /*
1470          * Finally, compute the new sector number
1471          */
1472         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1473         return new_sector;
1474 }
1475
1476
1477 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1478 {
1479         raid5_conf_t *conf = sh->raid_conf;
1480         int raid_disks = sh->disks;
1481         int data_disks = raid_disks - conf->max_degraded;
1482         sector_t new_sector = sh->sector, check;
1483         int sectors_per_chunk = previous ? (conf->prev_chunk >> 9)
1484                                          : (conf->chunk_size >> 9);
1485         int algorithm = previous ? conf->prev_algo
1486                                  : conf->algorithm;
1487         sector_t stripe;
1488         int chunk_offset;
1489         int chunk_number, dummy1, dd_idx = i;
1490         sector_t r_sector;
1491         struct stripe_head sh2;
1492
1493
1494         chunk_offset = sector_div(new_sector, sectors_per_chunk);
1495         stripe = new_sector;
1496         BUG_ON(new_sector != stripe);
1497
1498         if (i == sh->pd_idx)
1499                 return 0;
1500         switch(conf->level) {
1501         case 4: break;
1502         case 5:
1503                 switch (algorithm) {
1504                 case ALGORITHM_LEFT_ASYMMETRIC:
1505                 case ALGORITHM_RIGHT_ASYMMETRIC:
1506                         if (i > sh->pd_idx)
1507                                 i--;
1508                         break;
1509                 case ALGORITHM_LEFT_SYMMETRIC:
1510                 case ALGORITHM_RIGHT_SYMMETRIC:
1511                         if (i < sh->pd_idx)
1512                                 i += raid_disks;
1513                         i -= (sh->pd_idx + 1);
1514                         break;
1515                 case ALGORITHM_PARITY_0:
1516                         i -= 1;
1517                         break;
1518                 case ALGORITHM_PARITY_N:
1519                         break;
1520                 default:
1521                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1522                                algorithm);
1523                         BUG();
1524                 }
1525                 break;
1526         case 6:
1527                 if (i == sh->qd_idx)
1528                         return 0; /* It is the Q disk */
1529                 switch (algorithm) {
1530                 case ALGORITHM_LEFT_ASYMMETRIC:
1531                 case ALGORITHM_RIGHT_ASYMMETRIC:
1532                 case ALGORITHM_ROTATING_ZERO_RESTART:
1533                 case ALGORITHM_ROTATING_N_RESTART:
1534                         if (sh->pd_idx == raid_disks-1)
1535                                 i--;    /* Q D D D P */
1536                         else if (i > sh->pd_idx)
1537                                 i -= 2; /* D D P Q D */
1538                         break;
1539                 case ALGORITHM_LEFT_SYMMETRIC:
1540                 case ALGORITHM_RIGHT_SYMMETRIC:
1541                         if (sh->pd_idx == raid_disks-1)
1542                                 i--; /* Q D D D P */
1543                         else {
1544                                 /* D D P Q D */
1545                                 if (i < sh->pd_idx)
1546                                         i += raid_disks;
1547                                 i -= (sh->pd_idx + 2);
1548                         }
1549                         break;
1550                 case ALGORITHM_PARITY_0:
1551                         i -= 2;
1552                         break;
1553                 case ALGORITHM_PARITY_N:
1554                         break;
1555                 case ALGORITHM_ROTATING_N_CONTINUE:
1556                         if (sh->pd_idx == 0)
1557                                 i--;    /* P D D D Q */
1558                         else if (i > sh->pd_idx)
1559                                 i -= 2; /* D D Q P D */
1560                         break;
1561                 case ALGORITHM_LEFT_ASYMMETRIC_6:
1562                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1563                         if (i > sh->pd_idx)
1564                                 i--;
1565                         break;
1566                 case ALGORITHM_LEFT_SYMMETRIC_6:
1567                 case ALGORITHM_RIGHT_SYMMETRIC_6:
1568                         if (i < sh->pd_idx)
1569                                 i += data_disks + 1;
1570                         i -= (sh->pd_idx + 1);
1571                         break;
1572                 case ALGORITHM_PARITY_0_6:
1573                         i -= 1;
1574                         break;
1575                 default:
1576                         printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1577                                algorithm);
1578                         BUG();
1579                 }
1580                 break;
1581         }
1582
1583         chunk_number = stripe * data_disks + i;
1584         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1585
1586         check = raid5_compute_sector(conf, r_sector,
1587                                      previous, &dummy1, &sh2);
1588         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
1589                 || sh2.qd_idx != sh->qd_idx) {
1590                 printk(KERN_ERR "compute_blocknr: map not correct\n");
1591                 return 0;
1592         }
1593         return r_sector;
1594 }
1595
1596
1597
1598 /*
1599  * Copy data between a page in the stripe cache, and one or more bion
1600  * The page could align with the middle of the bio, or there could be
1601  * several bion, each with several bio_vecs, which cover part of the page
1602  * Multiple bion are linked together on bi_next.  There may be extras
1603  * at the end of this list.  We ignore them.
1604  */
1605 static void copy_data(int frombio, struct bio *bio,
1606                      struct page *page,
1607                      sector_t sector)
1608 {
1609         char *pa = page_address(page);
1610         struct bio_vec *bvl;
1611         int i;
1612         int page_offset;
1613
1614         if (bio->bi_sector >= sector)
1615                 page_offset = (signed)(bio->bi_sector - sector) * 512;
1616         else
1617                 page_offset = (signed)(sector - bio->bi_sector) * -512;
1618         bio_for_each_segment(bvl, bio, i) {
1619                 int len = bio_iovec_idx(bio,i)->bv_len;
1620                 int clen;
1621                 int b_offset = 0;
1622
1623                 if (page_offset < 0) {
1624                         b_offset = -page_offset;
1625                         page_offset += b_offset;
1626                         len -= b_offset;
1627                 }
1628
1629                 if (len > 0 && page_offset + len > STRIPE_SIZE)
1630                         clen = STRIPE_SIZE - page_offset;
1631                 else clen = len;
1632
1633                 if (clen > 0) {
1634                         char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1635                         if (frombio)
1636                                 memcpy(pa+page_offset, ba+b_offset, clen);
1637                         else
1638                                 memcpy(ba+b_offset, pa+page_offset, clen);
1639                         __bio_kunmap_atomic(ba, KM_USER0);
1640                 }
1641                 if (clen < len) /* hit end of page */
1642                         break;
1643                 page_offset +=  len;
1644         }
1645 }
1646
1647 #define check_xor()     do {                                              \
1648                                 if (count == MAX_XOR_BLOCKS) {            \
1649                                 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1650                                 count = 0;                                \
1651                            }                                              \
1652                         } while(0)
1653
1654 static void compute_parity6(struct stripe_head *sh, int method)
1655 {
1656         raid5_conf_t *conf = sh->raid_conf;
1657         int i, pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1658         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1659         struct bio *chosen;
1660         /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1661         void *ptrs[syndrome_disks+2];
1662
1663         pd_idx = sh->pd_idx;
1664         qd_idx = sh->qd_idx;
1665         d0_idx = raid6_d0(sh);
1666
1667         pr_debug("compute_parity, stripe %llu, method %d\n",
1668                 (unsigned long long)sh->sector, method);
1669
1670         switch(method) {
1671         case READ_MODIFY_WRITE:
1672                 BUG();          /* READ_MODIFY_WRITE N/A for RAID-6 */
1673         case RECONSTRUCT_WRITE:
1674                 for (i= disks; i-- ;)
1675                         if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1676                                 chosen = sh->dev[i].towrite;
1677                                 sh->dev[i].towrite = NULL;
1678
1679                                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1680                                         wake_up(&conf->wait_for_overlap);
1681
1682                                 BUG_ON(sh->dev[i].written);
1683                                 sh->dev[i].written = chosen;
1684                         }
1685                 break;
1686         case CHECK_PARITY:
1687                 BUG();          /* Not implemented yet */
1688         }
1689
1690         for (i = disks; i--;)
1691                 if (sh->dev[i].written) {
1692                         sector_t sector = sh->dev[i].sector;
1693                         struct bio *wbi = sh->dev[i].written;
1694                         while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1695                                 copy_data(1, wbi, sh->dev[i].page, sector);
1696                                 wbi = r5_next_bio(wbi, sector);
1697                         }
1698
1699                         set_bit(R5_LOCKED, &sh->dev[i].flags);
1700                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
1701                 }
1702
1703         /* Note that unlike RAID-5, the ordering of the disks matters greatly.*/
1704
1705         for (i = 0; i < disks; i++)
1706                 ptrs[i] = (void *)raid6_empty_zero_page;
1707
1708         count = 0;
1709         i = d0_idx;
1710         do {
1711                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1712
1713                 ptrs[slot] = page_address(sh->dev[i].page);
1714                 if (slot < syndrome_disks &&
1715                     !test_bit(R5_UPTODATE, &sh->dev[i].flags)) {
1716                         printk(KERN_ERR "block %d/%d not uptodate "
1717                                "on parity calc\n", i, count);
1718                         BUG();
1719                 }
1720
1721                 i = raid6_next_disk(i, disks);
1722         } while (i != d0_idx);
1723         BUG_ON(count != syndrome_disks);
1724
1725         raid6_call.gen_syndrome(syndrome_disks+2, STRIPE_SIZE, ptrs);
1726
1727         switch(method) {
1728         case RECONSTRUCT_WRITE:
1729                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1730                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1731                 set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
1732                 set_bit(R5_LOCKED,   &sh->dev[qd_idx].flags);
1733                 break;
1734         case UPDATE_PARITY:
1735                 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1736                 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1737                 break;
1738         }
1739 }
1740
1741
1742 /* Compute one missing block */
1743 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1744 {
1745         int i, count, disks = sh->disks;
1746         void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1747         int qd_idx = sh->qd_idx;
1748
1749         pr_debug("compute_block_1, stripe %llu, idx %d\n",
1750                 (unsigned long long)sh->sector, dd_idx);
1751
1752         if ( dd_idx == qd_idx ) {
1753                 /* We're actually computing the Q drive */
1754                 compute_parity6(sh, UPDATE_PARITY);
1755         } else {
1756                 dest = page_address(sh->dev[dd_idx].page);
1757                 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1758                 count = 0;
1759                 for (i = disks ; i--; ) {
1760                         if (i == dd_idx || i == qd_idx)
1761                                 continue;
1762                         p = page_address(sh->dev[i].page);
1763                         if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1764                                 ptr[count++] = p;
1765                         else
1766                                 printk("compute_block() %d, stripe %llu, %d"
1767                                        " not present\n", dd_idx,
1768                                        (unsigned long long)sh->sector, i);
1769
1770                         check_xor();
1771                 }
1772                 if (count)
1773                         xor_blocks(count, STRIPE_SIZE, dest, ptr);
1774                 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1775                 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1776         }
1777 }
1778
1779 /* Compute two missing blocks */
1780 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1781 {
1782         int i, count, disks = sh->disks;
1783         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1784         int d0_idx = raid6_d0(sh);
1785         int faila = -1, failb = -1;
1786         /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1787         void *ptrs[syndrome_disks+2];
1788
1789         for (i = 0; i < disks ; i++)
1790                 ptrs[i] = (void *)raid6_empty_zero_page;
1791         count = 0;
1792         i = d0_idx;
1793         do {
1794                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1795
1796                 ptrs[slot] = page_address(sh->dev[i].page);
1797
1798                 if (i == dd_idx1)
1799                         faila = slot;
1800                 if (i == dd_idx2)
1801                         failb = slot;
1802                 i = raid6_next_disk(i, disks);
1803         } while (i != d0_idx);
1804         BUG_ON(count != syndrome_disks);
1805
1806         BUG_ON(faila == failb);
1807         if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1808
1809         pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1810                  (unsigned long long)sh->sector, dd_idx1, dd_idx2,
1811                  faila, failb);
1812
1813         if (failb == syndrome_disks+1) {
1814                 /* Q disk is one of the missing disks */
1815                 if (faila == syndrome_disks) {
1816                         /* Missing P+Q, just recompute */
1817                         compute_parity6(sh, UPDATE_PARITY);
1818                         return;
1819                 } else {
1820                         /* We're missing D+Q; recompute D from P */
1821                         compute_block_1(sh, ((dd_idx1 == sh->qd_idx) ?
1822                                              dd_idx2 : dd_idx1),
1823                                         0);
1824                         compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1825                         return;
1826                 }
1827         }
1828
1829         /* We're missing D+P or D+D; */
1830         if (failb == syndrome_disks) {
1831                 /* We're missing D+P. */
1832                 raid6_datap_recov(syndrome_disks+2, STRIPE_SIZE, faila, ptrs);
1833         } else {
1834                 /* We're missing D+D. */
1835                 raid6_2data_recov(syndrome_disks+2, STRIPE_SIZE, faila, failb,
1836                                   ptrs);
1837         }
1838
1839         /* Both the above update both missing blocks */
1840         set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1841         set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1842 }
1843
1844 static void
1845 schedule_reconstruction5(struct stripe_head *sh, struct stripe_head_state *s,
1846                          int rcw, int expand)
1847 {
1848         int i, pd_idx = sh->pd_idx, disks = sh->disks;
1849
1850         if (rcw) {
1851                 /* if we are not expanding this is a proper write request, and
1852                  * there will be bios with new data to be drained into the
1853                  * stripe cache
1854                  */
1855                 if (!expand) {
1856                         sh->reconstruct_state = reconstruct_state_drain_run;
1857                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1858                 } else
1859                         sh->reconstruct_state = reconstruct_state_run;
1860
1861                 set_bit(STRIPE_OP_POSTXOR, &s->ops_request);
1862
1863                 for (i = disks; i--; ) {
1864                         struct r5dev *dev = &sh->dev[i];
1865
1866                         if (dev->towrite) {
1867                                 set_bit(R5_LOCKED, &dev->flags);
1868                                 set_bit(R5_Wantdrain, &dev->flags);
1869                                 if (!expand)
1870                                         clear_bit(R5_UPTODATE, &dev->flags);
1871                                 s->locked++;
1872                         }
1873                 }
1874                 if (s->locked + 1 == disks)
1875                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
1876                                 atomic_inc(&sh->raid_conf->pending_full_writes);
1877         } else {
1878                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1879                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1880
1881                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
1882                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
1883                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1884                 set_bit(STRIPE_OP_POSTXOR, &s->ops_request);
1885
1886                 for (i = disks; i--; ) {
1887                         struct r5dev *dev = &sh->dev[i];
1888                         if (i == pd_idx)
1889                                 continue;
1890
1891                         if (dev->towrite &&
1892                             (test_bit(R5_UPTODATE, &dev->flags) ||
1893                              test_bit(R5_Wantcompute, &dev->flags))) {
1894                                 set_bit(R5_Wantdrain, &dev->flags);
1895                                 set_bit(R5_LOCKED, &dev->flags);
1896                                 clear_bit(R5_UPTODATE, &dev->flags);
1897                                 s->locked++;
1898                         }
1899                 }
1900         }
1901
1902         /* keep the parity disk locked while asynchronous operations
1903          * are in flight
1904          */
1905         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1906         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1907         s->locked++;
1908
1909         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
1910                 __func__, (unsigned long long)sh->sector,
1911                 s->locked, s->ops_request);
1912 }
1913
1914 /*
1915  * Each stripe/dev can have one or more bion attached.
1916  * toread/towrite point to the first in a chain.
1917  * The bi_next chain must be in order.
1918  */
1919 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1920 {
1921         struct bio **bip;
1922         raid5_conf_t *conf = sh->raid_conf;
1923         int firstwrite=0;
1924
1925         pr_debug("adding bh b#%llu to stripe s#%llu\n",
1926                 (unsigned long long)bi->bi_sector,
1927                 (unsigned long long)sh->sector);
1928
1929
1930         spin_lock(&sh->lock);
1931         spin_lock_irq(&conf->device_lock);
1932         if (forwrite) {
1933                 bip = &sh->dev[dd_idx].towrite;
1934                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1935                         firstwrite = 1;
1936         } else
1937                 bip = &sh->dev[dd_idx].toread;
1938         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1939                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1940                         goto overlap;
1941                 bip = & (*bip)->bi_next;
1942         }
1943         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1944                 goto overlap;
1945
1946         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1947         if (*bip)
1948                 bi->bi_next = *bip;
1949         *bip = bi;
1950         bi->bi_phys_segments++;
1951         spin_unlock_irq(&conf->device_lock);
1952         spin_unlock(&sh->lock);
1953
1954         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1955                 (unsigned long long)bi->bi_sector,
1956                 (unsigned long long)sh->sector, dd_idx);
1957
1958         if (conf->mddev->bitmap && firstwrite) {
1959                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1960                                   STRIPE_SECTORS, 0);
1961                 sh->bm_seq = conf->seq_flush+1;
1962                 set_bit(STRIPE_BIT_DELAY, &sh->state);
1963         }
1964
1965         if (forwrite) {
1966                 /* check if page is covered */
1967                 sector_t sector = sh->dev[dd_idx].sector;
1968                 for (bi=sh->dev[dd_idx].towrite;
1969                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1970                              bi && bi->bi_sector <= sector;
1971                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1972                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1973                                 sector = bi->bi_sector + (bi->bi_size>>9);
1974                 }
1975                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1976                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1977         }
1978         return 1;
1979
1980  overlap:
1981         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1982         spin_unlock_irq(&conf->device_lock);
1983         spin_unlock(&sh->lock);
1984         return 0;
1985 }
1986
1987 static void end_reshape(raid5_conf_t *conf);
1988
1989 static int page_is_zero(struct page *p)
1990 {
1991         char *a = page_address(p);
1992         return ((*(u32*)a) == 0 &&
1993                 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1994 }
1995
1996 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
1997                             struct stripe_head *sh)
1998 {
1999         int sectors_per_chunk =
2000                 previous ? (conf->prev_chunk >> 9)
2001                          : (conf->chunk_size >> 9);
2002         int dd_idx;
2003         int chunk_offset = sector_div(stripe, sectors_per_chunk);
2004         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2005
2006         raid5_compute_sector(conf,
2007                              stripe * (disks - conf->max_degraded)
2008                              *sectors_per_chunk + chunk_offset,
2009                              previous,
2010                              &dd_idx, sh);
2011 }
2012
2013 static void
2014 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2015                                 struct stripe_head_state *s, int disks,
2016                                 struct bio **return_bi)
2017 {
2018         int i;
2019         for (i = disks; i--; ) {
2020                 struct bio *bi;
2021                 int bitmap_end = 0;
2022
2023                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2024                         mdk_rdev_t *rdev;
2025                         rcu_read_lock();
2026                         rdev = rcu_dereference(conf->disks[i].rdev);
2027                         if (rdev && test_bit(In_sync, &rdev->flags))
2028                                 /* multiple read failures in one stripe */
2029                                 md_error(conf->mddev, rdev);
2030                         rcu_read_unlock();
2031                 }
2032                 spin_lock_irq(&conf->device_lock);
2033                 /* fail all writes first */
2034                 bi = sh->dev[i].towrite;
2035                 sh->dev[i].towrite = NULL;
2036                 if (bi) {
2037                         s->to_write--;
2038                         bitmap_end = 1;
2039                 }
2040
2041                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2042                         wake_up(&conf->wait_for_overlap);
2043
2044                 while (bi && bi->bi_sector <
2045                         sh->dev[i].sector + STRIPE_SECTORS) {
2046                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2047                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2048                         if (!raid5_dec_bi_phys_segments(bi)) {
2049                                 md_write_end(conf->mddev);
2050                                 bi->bi_next = *return_bi;
2051                                 *return_bi = bi;
2052                         }
2053                         bi = nextbi;
2054                 }
2055                 /* and fail all 'written' */
2056                 bi = sh->dev[i].written;
2057                 sh->dev[i].written = NULL;
2058                 if (bi) bitmap_end = 1;
2059                 while (bi && bi->bi_sector <
2060                        sh->dev[i].sector + STRIPE_SECTORS) {
2061                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2062                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2063                         if (!raid5_dec_bi_phys_segments(bi)) {
2064                                 md_write_end(conf->mddev);
2065                                 bi->bi_next = *return_bi;
2066                                 *return_bi = bi;
2067                         }
2068                         bi = bi2;
2069                 }
2070
2071                 /* fail any reads if this device is non-operational and
2072                  * the data has not reached the cache yet.
2073                  */
2074                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2075                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2076                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
2077                         bi = sh->dev[i].toread;
2078                         sh->dev[i].toread = NULL;
2079                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2080                                 wake_up(&conf->wait_for_overlap);
2081                         if (bi) s->to_read--;
2082                         while (bi && bi->bi_sector <
2083                                sh->dev[i].sector + STRIPE_SECTORS) {
2084                                 struct bio *nextbi =
2085                                         r5_next_bio(bi, sh->dev[i].sector);
2086                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2087                                 if (!raid5_dec_bi_phys_segments(bi)) {
2088                                         bi->bi_next = *return_bi;
2089                                         *return_bi = bi;
2090                                 }
2091                                 bi = nextbi;
2092                         }
2093                 }
2094                 spin_unlock_irq(&conf->device_lock);
2095                 if (bitmap_end)
2096                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2097                                         STRIPE_SECTORS, 0, 0);
2098         }
2099
2100         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2101                 if (atomic_dec_and_test(&conf->pending_full_writes))
2102                         md_wakeup_thread(conf->mddev->thread);
2103 }
2104
2105 /* fetch_block5 - checks the given member device to see if its data needs
2106  * to be read or computed to satisfy a request.
2107  *
2108  * Returns 1 when no more member devices need to be checked, otherwise returns
2109  * 0 to tell the loop in handle_stripe_fill5 to continue
2110  */
2111 static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s,
2112                         int disk_idx, int disks)
2113 {
2114         struct r5dev *dev = &sh->dev[disk_idx];
2115         struct r5dev *failed_dev = &sh->dev[s->failed_num];
2116
2117         /* is the data in this block needed, and can we get it? */
2118         if (!test_bit(R5_LOCKED, &dev->flags) &&
2119             !test_bit(R5_UPTODATE, &dev->flags) &&
2120             (dev->toread ||
2121              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2122              s->syncing || s->expanding ||
2123              (s->failed &&
2124               (failed_dev->toread ||
2125                (failed_dev->towrite &&
2126                 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) {
2127                 /* We would like to get this block, possibly by computing it,
2128                  * otherwise read it if the backing disk is insync
2129                  */
2130                 if ((s->uptodate == disks - 1) &&
2131                     (s->failed && disk_idx == s->failed_num)) {
2132                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2133                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2134                         set_bit(R5_Wantcompute, &dev->flags);
2135                         sh->ops.target = disk_idx;
2136                         s->req_compute = 1;
2137                         /* Careful: from this point on 'uptodate' is in the eye
2138                          * of raid5_run_ops which services 'compute' operations
2139                          * before writes. R5_Wantcompute flags a block that will
2140                          * be R5_UPTODATE by the time it is needed for a
2141                          * subsequent operation.
2142                          */
2143                         s->uptodate++;
2144                         return 1; /* uptodate + compute == disks */
2145                 } else if (test_bit(R5_Insync, &dev->flags)) {
2146                         set_bit(R5_LOCKED, &dev->flags);
2147                         set_bit(R5_Wantread, &dev->flags);
2148                         s->locked++;
2149                         pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2150                                 s->syncing);
2151                 }
2152         }
2153
2154         return 0;
2155 }
2156
2157 /**
2158  * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2159  */
2160 static void handle_stripe_fill5(struct stripe_head *sh,
2161                         struct stripe_head_state *s, int disks)
2162 {
2163         int i;
2164
2165         /* look for blocks to read/compute, skip this if a compute
2166          * is already in flight, or if the stripe contents are in the
2167          * midst of changing due to a write
2168          */
2169         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2170             !sh->reconstruct_state)
2171                 for (i = disks; i--; )
2172                         if (fetch_block5(sh, s, i, disks))
2173                                 break;
2174         set_bit(STRIPE_HANDLE, &sh->state);
2175 }
2176
2177 static void handle_stripe_fill6(struct stripe_head *sh,
2178                         struct stripe_head_state *s, struct r6_state *r6s,
2179                         int disks)
2180 {
2181         int i;
2182         for (i = disks; i--; ) {
2183                 struct r5dev *dev = &sh->dev[i];
2184                 if (!test_bit(R5_LOCKED, &dev->flags) &&
2185                     !test_bit(R5_UPTODATE, &dev->flags) &&
2186                     (dev->toread || (dev->towrite &&
2187                      !test_bit(R5_OVERWRITE, &dev->flags)) ||
2188                      s->syncing || s->expanding ||
2189                      (s->failed >= 1 &&
2190                       (sh->dev[r6s->failed_num[0]].toread ||
2191                        s->to_write)) ||
2192                      (s->failed >= 2 &&
2193                       (sh->dev[r6s->failed_num[1]].toread ||
2194                        s->to_write)))) {
2195                         /* we would like to get this block, possibly
2196                          * by computing it, but we might not be able to
2197                          */
2198                         if ((s->uptodate == disks - 1) &&
2199                             (s->failed && (i == r6s->failed_num[0] ||
2200                                            i == r6s->failed_num[1]))) {
2201                                 pr_debug("Computing stripe %llu block %d\n",
2202                                        (unsigned long long)sh->sector, i);
2203                                 compute_block_1(sh, i, 0);
2204                                 s->uptodate++;
2205                         } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2206                                 /* Computing 2-failure is *very* expensive; only
2207                                  * do it if failed >= 2
2208                                  */
2209                                 int other;
2210                                 for (other = disks; other--; ) {
2211                                         if (other == i)
2212                                                 continue;
2213                                         if (!test_bit(R5_UPTODATE,
2214                                               &sh->dev[other].flags))
2215                                                 break;
2216                                 }
2217                                 BUG_ON(other < 0);
2218                                 pr_debug("Computing stripe %llu blocks %d,%d\n",
2219                                        (unsigned long long)sh->sector,
2220                                        i, other);
2221                                 compute_block_2(sh, i, other);
2222                                 s->uptodate += 2;
2223                         } else if (test_bit(R5_Insync, &dev->flags)) {
2224                                 set_bit(R5_LOCKED, &dev->flags);
2225                                 set_bit(R5_Wantread, &dev->flags);
2226                                 s->locked++;
2227                                 pr_debug("Reading block %d (sync=%d)\n",
2228                                         i, s->syncing);
2229                         }
2230                 }
2231         }
2232         set_bit(STRIPE_HANDLE, &sh->state);
2233 }
2234
2235
2236 /* handle_stripe_clean_event
2237  * any written block on an uptodate or failed drive can be returned.
2238  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2239  * never LOCKED, so we don't need to test 'failed' directly.
2240  */
2241 static void handle_stripe_clean_event(raid5_conf_t *conf,
2242         struct stripe_head *sh, int disks, struct bio **return_bi)
2243 {
2244         int i;
2245         struct r5dev *dev;
2246
2247         for (i = disks; i--; )
2248                 if (sh->dev[i].written) {
2249                         dev = &sh->dev[i];
2250                         if (!test_bit(R5_LOCKED, &dev->flags) &&
2251                                 test_bit(R5_UPTODATE, &dev->flags)) {
2252                                 /* We can return any write requests */
2253                                 struct bio *wbi, *wbi2;
2254                                 int bitmap_end = 0;
2255                                 pr_debug("Return write for disc %d\n", i);
2256                                 spin_lock_irq(&conf->device_lock);
2257                                 wbi = dev->written;
2258                                 dev->written = NULL;
2259                                 while (wbi && wbi->bi_sector <
2260                                         dev->sector + STRIPE_SECTORS) {
2261                                         wbi2 = r5_next_bio(wbi, dev->sector);
2262                                         if (!raid5_dec_bi_phys_segments(wbi)) {
2263                                                 md_write_end(conf->mddev);
2264                                                 wbi->bi_next = *return_bi;
2265                                                 *return_bi = wbi;
2266                                         }
2267                                         wbi = wbi2;
2268                                 }
2269                                 if (dev->towrite == NULL)
2270                                         bitmap_end = 1;
2271                                 spin_unlock_irq(&conf->device_lock);
2272                                 if (bitmap_end)
2273                                         bitmap_endwrite(conf->mddev->bitmap,
2274                                                         sh->sector,
2275                                                         STRIPE_SECTORS,
2276                                          !test_bit(STRIPE_DEGRADED, &sh->state),
2277                                                         0);
2278                         }
2279                 }
2280
2281         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2282                 if (atomic_dec_and_test(&conf->pending_full_writes))
2283                         md_wakeup_thread(conf->mddev->thread);
2284 }
2285
2286 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2287                 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2288 {
2289         int rmw = 0, rcw = 0, i;
2290         for (i = disks; i--; ) {
2291                 /* would I have to read this buffer for read_modify_write */
2292                 struct r5dev *dev = &sh->dev[i];
2293                 if ((dev->towrite || i == sh->pd_idx) &&
2294                     !test_bit(R5_LOCKED, &dev->flags) &&
2295                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2296                       test_bit(R5_Wantcompute, &dev->flags))) {
2297                         if (test_bit(R5_Insync, &dev->flags))
2298                                 rmw++;
2299                         else
2300                                 rmw += 2*disks;  /* cannot read it */
2301                 }
2302                 /* Would I have to read this buffer for reconstruct_write */
2303                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2304                     !test_bit(R5_LOCKED, &dev->flags) &&
2305                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2306                     test_bit(R5_Wantcompute, &dev->flags))) {
2307                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2308                         else
2309                                 rcw += 2*disks;
2310                 }
2311         }
2312         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2313                 (unsigned long long)sh->sector, rmw, rcw);
2314         set_bit(STRIPE_HANDLE, &sh->state);
2315         if (rmw < rcw && rmw > 0)
2316                 /* prefer read-modify-write, but need to get some data */
2317                 for (i = disks; i--; ) {
2318                         struct r5dev *dev = &sh->dev[i];
2319                         if ((dev->towrite || i == sh->pd_idx) &&
2320                             !test_bit(R5_LOCKED, &dev->flags) &&
2321                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2322                             test_bit(R5_Wantcompute, &dev->flags)) &&
2323                             test_bit(R5_Insync, &dev->flags)) {
2324                                 if (
2325                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2326                                         pr_debug("Read_old block "
2327                                                 "%d for r-m-w\n", i);
2328                                         set_bit(R5_LOCKED, &dev->flags);
2329                                         set_bit(R5_Wantread, &dev->flags);
2330                                         s->locked++;
2331                                 } else {
2332                                         set_bit(STRIPE_DELAYED, &sh->state);
2333                                         set_bit(STRIPE_HANDLE, &sh->state);
2334                                 }
2335                         }
2336                 }
2337         if (rcw <= rmw && rcw > 0)
2338                 /* want reconstruct write, but need to get some data */
2339                 for (i = disks; i--; ) {
2340                         struct r5dev *dev = &sh->dev[i];
2341                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2342                             i != sh->pd_idx &&
2343                             !test_bit(R5_LOCKED, &dev->flags) &&
2344                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2345                             test_bit(R5_Wantcompute, &dev->flags)) &&
2346                             test_bit(R5_Insync, &dev->flags)) {
2347                                 if (
2348                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2349                                         pr_debug("Read_old block "
2350                                                 "%d for Reconstruct\n", i);
2351                                         set_bit(R5_LOCKED, &dev->flags);
2352                                         set_bit(R5_Wantread, &dev->flags);
2353                                         s->locked++;
2354                                 } else {
2355                                         set_bit(STRIPE_DELAYED, &sh->state);
2356                                         set_bit(STRIPE_HANDLE, &sh->state);
2357                                 }
2358                         }
2359                 }
2360         /* now if nothing is locked, and if we have enough data,
2361          * we can start a write request
2362          */
2363         /* since handle_stripe can be called at any time we need to handle the
2364          * case where a compute block operation has been submitted and then a
2365          * subsequent call wants to start a write request.  raid5_run_ops only
2366          * handles the case where compute block and postxor are requested
2367          * simultaneously.  If this is not the case then new writes need to be
2368          * held off until the compute completes.
2369          */
2370         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2371             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2372             !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2373                 schedule_reconstruction5(sh, s, rcw == 0, 0);
2374 }
2375
2376 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2377                 struct stripe_head *sh, struct stripe_head_state *s,
2378                 struct r6_state *r6s, int disks)
2379 {
2380         int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2381         int qd_idx = sh->qd_idx;
2382         for (i = disks; i--; ) {
2383                 struct r5dev *dev = &sh->dev[i];
2384                 /* Would I have to read this buffer for reconstruct_write */
2385                 if (!test_bit(R5_OVERWRITE, &dev->flags)
2386                     && i != pd_idx && i != qd_idx
2387                     && (!test_bit(R5_LOCKED, &dev->flags)
2388                             ) &&
2389                     !test_bit(R5_UPTODATE, &dev->flags)) {
2390                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2391                         else {
2392                                 pr_debug("raid6: must_compute: "
2393                                         "disk %d flags=%#lx\n", i, dev->flags);
2394                                 must_compute++;
2395                         }
2396                 }
2397         }
2398         pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2399                (unsigned long long)sh->sector, rcw, must_compute);
2400         set_bit(STRIPE_HANDLE, &sh->state);
2401
2402         if (rcw > 0)
2403                 /* want reconstruct write, but need to get some data */
2404                 for (i = disks; i--; ) {
2405                         struct r5dev *dev = &sh->dev[i];
2406                         if (!test_bit(R5_OVERWRITE, &dev->flags)
2407                             && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2408                             && !test_bit(R5_LOCKED, &dev->flags) &&
2409                             !test_bit(R5_UPTODATE, &dev->flags) &&
2410                             test_bit(R5_Insync, &dev->flags)) {
2411                                 if (
2412                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2413                                         pr_debug("Read_old stripe %llu "
2414                                                 "block %d for Reconstruct\n",
2415                                              (unsigned long long)sh->sector, i);
2416                                         set_bit(R5_LOCKED, &dev->flags);
2417                                         set_bit(R5_Wantread, &dev->flags);
2418                                         s->locked++;
2419                                 } else {
2420                                         pr_debug("Request delayed stripe %llu "
2421                                                 "block %d for Reconstruct\n",
2422                                              (unsigned long long)sh->sector, i);
2423                                         set_bit(STRIPE_DELAYED, &sh->state);
2424                                         set_bit(STRIPE_HANDLE, &sh->state);
2425                                 }
2426                         }
2427                 }
2428         /* now if nothing is locked, and if we have enough data, we can start a
2429          * write request
2430          */
2431         if (s->locked == 0 && rcw == 0 &&
2432             !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2433                 if (must_compute > 0) {
2434                         /* We have failed blocks and need to compute them */
2435                         switch (s->failed) {
2436                         case 0:
2437                                 BUG();
2438                         case 1:
2439                                 compute_block_1(sh, r6s->failed_num[0], 0);
2440                                 break;
2441                         case 2:
2442                                 compute_block_2(sh, r6s->failed_num[0],
2443                                                 r6s->failed_num[1]);
2444                                 break;
2445                         default: /* This request should have been failed? */
2446                                 BUG();
2447                         }
2448                 }
2449
2450                 pr_debug("Computing parity for stripe %llu\n",
2451                         (unsigned long long)sh->sector);
2452                 compute_parity6(sh, RECONSTRUCT_WRITE);
2453                 /* now every locked buffer is ready to be written */
2454                 for (i = disks; i--; )
2455                         if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2456                                 pr_debug("Writing stripe %llu block %d\n",
2457                                        (unsigned long long)sh->sector, i);
2458                                 s->locked++;
2459                                 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2460                         }
2461                 if (s->locked == disks)
2462                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2463                                 atomic_inc(&conf->pending_full_writes);
2464                 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2465                 set_bit(STRIPE_INSYNC, &sh->state);
2466
2467                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2468                         atomic_dec(&conf->preread_active_stripes);
2469                         if (atomic_read(&conf->preread_active_stripes) <
2470                             IO_THRESHOLD)
2471                                 md_wakeup_thread(conf->mddev->thread);
2472                 }
2473         }
2474 }
2475
2476 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2477                                 struct stripe_head_state *s, int disks)
2478 {
2479         struct r5dev *dev = NULL;
2480
2481         set_bit(STRIPE_HANDLE, &sh->state);
2482
2483         switch (sh->check_state) {
2484         case check_state_idle:
2485                 /* start a new check operation if there are no failures */
2486                 if (s->failed == 0) {
2487                         BUG_ON(s->uptodate != disks);
2488                         sh->check_state = check_state_run;
2489                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
2490                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2491                         s->uptodate--;
2492                         break;
2493                 }
2494                 dev = &sh->dev[s->failed_num];
2495                 /* fall through */
2496         case check_state_compute_result:
2497                 sh->check_state = check_state_idle;
2498                 if (!dev)
2499                         dev = &sh->dev[sh->pd_idx];
2500
2501                 /* check that a write has not made the stripe insync */
2502                 if (test_bit(STRIPE_INSYNC, &sh->state))
2503                         break;
2504
2505                 /* either failed parity check, or recovery is happening */
2506                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2507                 BUG_ON(s->uptodate != disks);
2508
2509                 set_bit(R5_LOCKED, &dev->flags);
2510                 s->locked++;
2511                 set_bit(R5_Wantwrite, &dev->flags);
2512
2513                 clear_bit(STRIPE_DEGRADED, &sh->state);
2514                 set_bit(STRIPE_INSYNC, &sh->state);
2515                 break;
2516         case check_state_run:
2517                 break; /* we will be called again upon completion */
2518         case check_state_check_result:
2519                 sh->check_state = check_state_idle;
2520
2521                 /* if a failure occurred during the check operation, leave
2522                  * STRIPE_INSYNC not set and let the stripe be handled again
2523                  */
2524                 if (s->failed)
2525                         break;
2526
2527                 /* handle a successful check operation, if parity is correct
2528                  * we are done.  Otherwise update the mismatch count and repair
2529                  * parity if !MD_RECOVERY_CHECK
2530                  */
2531                 if (sh->ops.zero_sum_result == 0)
2532                         /* parity is correct (on disc,
2533                          * not in buffer any more)
2534                          */
2535                         set_bit(STRIPE_INSYNC, &sh->state);
2536                 else {
2537                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2538                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2539                                 /* don't try to repair!! */
2540                                 set_bit(STRIPE_INSYNC, &sh->state);
2541                         else {
2542                                 sh->check_state = check_state_compute_run;
2543                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2544                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2545                                 set_bit(R5_Wantcompute,
2546                                         &sh->dev[sh->pd_idx].flags);
2547                                 sh->ops.target = sh->pd_idx;
2548                                 s->uptodate++;
2549                         }
2550                 }
2551                 break;
2552         case check_state_compute_run:
2553                 break;
2554         default:
2555                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2556                        __func__, sh->check_state,
2557                        (unsigned long long) sh->sector);
2558                 BUG();
2559         }
2560 }
2561
2562
2563 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2564                                 struct stripe_head_state *s,
2565                                 struct r6_state *r6s, struct page *tmp_page,
2566                                 int disks)
2567 {
2568         int update_p = 0, update_q = 0;
2569         struct r5dev *dev;
2570         int pd_idx = sh->pd_idx;
2571         int qd_idx = sh->qd_idx;
2572
2573         set_bit(STRIPE_HANDLE, &sh->state);
2574
2575         BUG_ON(s->failed > 2);
2576         BUG_ON(s->uptodate < disks);
2577         /* Want to check and possibly repair P and Q.
2578          * However there could be one 'failed' device, in which
2579          * case we can only check one of them, possibly using the
2580          * other to generate missing data
2581          */
2582
2583         /* If !tmp_page, we cannot do the calculations,
2584          * but as we have set STRIPE_HANDLE, we will soon be called
2585          * by stripe_handle with a tmp_page - just wait until then.
2586          */
2587         if (tmp_page) {
2588                 if (s->failed == r6s->q_failed) {
2589                         /* The only possible failed device holds 'Q', so it
2590                          * makes sense to check P (If anything else were failed,
2591                          * we would have used P to recreate it).
2592                          */
2593                         compute_block_1(sh, pd_idx, 1);
2594                         if (!page_is_zero(sh->dev[pd_idx].page)) {
2595                                 compute_block_1(sh, pd_idx, 0);
2596                                 update_p = 1;
2597                         }
2598                 }
2599                 if (!r6s->q_failed && s->failed < 2) {
2600                         /* q is not failed, and we didn't use it to generate
2601                          * anything, so it makes sense to check it
2602                          */
2603                         memcpy(page_address(tmp_page),
2604                                page_address(sh->dev[qd_idx].page),
2605                                STRIPE_SIZE);
2606                         compute_parity6(sh, UPDATE_PARITY);
2607                         if (memcmp(page_address(tmp_page),
2608                                    page_address(sh->dev[qd_idx].page),
2609                                    STRIPE_SIZE) != 0) {
2610                                 clear_bit(STRIPE_INSYNC, &sh->state);
2611                                 update_q = 1;
2612                         }
2613                 }
2614                 if (update_p || update_q) {
2615                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2616                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2617                                 /* don't try to repair!! */
2618                                 update_p = update_q = 0;
2619                 }
2620
2621                 /* now write out any block on a failed drive,
2622                  * or P or Q if they need it
2623                  */
2624
2625                 if (s->failed == 2) {
2626                         dev = &sh->dev[r6s->failed_num[1]];
2627                         s->locked++;
2628                         set_bit(R5_LOCKED, &dev->flags);
2629                         set_bit(R5_Wantwrite, &dev->flags);
2630                 }
2631                 if (s->failed >= 1) {
2632                         dev = &sh->dev[r6s->failed_num[0]];
2633                         s->locked++;
2634                         set_bit(R5_LOCKED, &dev->flags);
2635                         set_bit(R5_Wantwrite, &dev->flags);
2636                 }
2637
2638                 if (update_p) {
2639                         dev = &sh->dev[pd_idx];
2640                         s->locked++;
2641                         set_bit(R5_LOCKED, &dev->flags);
2642                         set_bit(R5_Wantwrite, &dev->flags);
2643                 }
2644                 if (update_q) {
2645                         dev = &sh->dev[qd_idx];
2646                         s->locked++;
2647                         set_bit(R5_LOCKED, &dev->flags);
2648                         set_bit(R5_Wantwrite, &dev->flags);
2649                 }
2650                 clear_bit(STRIPE_DEGRADED, &sh->state);
2651
2652                 set_bit(STRIPE_INSYNC, &sh->state);
2653         }
2654 }
2655
2656 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2657                                 struct r6_state *r6s)
2658 {
2659         int i;
2660
2661         /* We have read all the blocks in this stripe and now we need to
2662          * copy some of them into a target stripe for expand.
2663          */
2664         struct dma_async_tx_descriptor *tx = NULL;
2665         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2666         for (i = 0; i < sh->disks; i++)
2667                 if (i != sh->pd_idx && i != sh->qd_idx) {
2668                         int dd_idx, j;
2669                         struct stripe_head *sh2;
2670
2671                         sector_t bn = compute_blocknr(sh, i, 1);
2672                         sector_t s = raid5_compute_sector(conf, bn, 0,
2673                                                           &dd_idx, NULL);
2674                         sh2 = get_active_stripe(conf, s, 0, 1);
2675                         if (sh2 == NULL)
2676                                 /* so far only the early blocks of this stripe
2677                                  * have been requested.  When later blocks
2678                                  * get requested, we will try again
2679                                  */
2680                                 continue;
2681                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2682                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2683                                 /* must have already done this block */
2684                                 release_stripe(sh2);
2685                                 continue;
2686                         }
2687
2688                         /* place all the copies on one channel */
2689                         tx = async_memcpy(sh2->dev[dd_idx].page,
2690                                 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2691                                 ASYNC_TX_DEP_ACK, tx, NULL, NULL);
2692
2693                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2694                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2695                         for (j = 0; j < conf->raid_disks; j++)
2696                                 if (j != sh2->pd_idx &&
2697                                     (!r6s || j != sh2->qd_idx) &&
2698                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
2699                                         break;
2700                         if (j == conf->raid_disks) {
2701                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2702                                 set_bit(STRIPE_HANDLE, &sh2->state);
2703                         }
2704                         release_stripe(sh2);
2705
2706                 }
2707         /* done submitting copies, wait for them to complete */
2708         if (tx) {
2709                 async_tx_ack(tx);
2710                 dma_wait_for_async_tx(tx);
2711         }
2712 }
2713
2714
2715 /*
2716  * handle_stripe - do things to a stripe.
2717  *
2718  * We lock the stripe and then examine the state of various bits
2719  * to see what needs to be done.
2720  * Possible results:
2721  *    return some read request which now have data
2722  *    return some write requests which are safely on disc
2723  *    schedule a read on some buffers
2724  *    schedule a write of some buffers
2725  *    return confirmation of parity correctness
2726  *
2727  * buffers are taken off read_list or write_list, and bh_cache buffers
2728  * get BH_Lock set before the stripe lock is released.
2729  *
2730  */
2731
2732 static bool handle_stripe5(struct stripe_head *sh)
2733 {
2734         raid5_conf_t *conf = sh->raid_conf;
2735         int disks = sh->disks, i;
2736         struct bio *return_bi = NULL;
2737         struct stripe_head_state s;
2738         struct r5dev *dev;
2739         mdk_rdev_t *blocked_rdev = NULL;
2740         int prexor;
2741
2742         memset(&s, 0, sizeof(s));
2743         pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
2744                  "reconstruct:%d\n", (unsigned long long)sh->sector, sh->state,
2745                  atomic_read(&sh->count), sh->pd_idx, sh->check_state,
2746                  sh->reconstruct_state);
2747
2748         spin_lock(&sh->lock);
2749         clear_bit(STRIPE_HANDLE, &sh->state);
2750         clear_bit(STRIPE_DELAYED, &sh->state);
2751
2752         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2753         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2754         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2755
2756         /* Now to look around and see what can be done */
2757         rcu_read_lock();
2758         for (i=disks; i--; ) {
2759                 mdk_rdev_t *rdev;
2760                 struct r5dev *dev = &sh->dev[i];
2761                 clear_bit(R5_Insync, &dev->flags);
2762
2763                 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2764                         "written %p\n", i, dev->flags, dev->toread, dev->read,
2765                         dev->towrite, dev->written);
2766
2767                 /* maybe we can request a biofill operation
2768                  *
2769                  * new wantfill requests are only permitted while
2770                  * ops_complete_biofill is guaranteed to be inactive
2771                  */
2772                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2773                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
2774                         set_bit(R5_Wantfill, &dev->flags);
2775
2776                 /* now count some things */
2777                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2778                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2779                 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2780
2781                 if (test_bit(R5_Wantfill, &dev->flags))
2782                         s.to_fill++;
2783                 else if (dev->toread)
2784                         s.to_read++;
2785                 if (dev->towrite) {
2786                         s.to_write++;
2787                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2788                                 s.non_overwrite++;
2789                 }
2790                 if (dev->written)
2791                         s.written++;
2792                 rdev = rcu_dereference(conf->disks[i].rdev);
2793                 if (blocked_rdev == NULL &&
2794                     rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2795                         blocked_rdev = rdev;
2796                         atomic_inc(&rdev->nr_pending);
2797                 }
2798                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2799                         /* The ReadError flag will just be confusing now */
2800                         clear_bit(R5_ReadError, &dev->flags);
2801                         clear_bit(R5_ReWrite, &dev->flags);
2802                 }
2803                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2804                     || test_bit(R5_ReadError, &dev->flags)) {
2805                         s.failed++;
2806                         s.failed_num = i;
2807                 } else
2808                         set_bit(R5_Insync, &dev->flags);
2809         }
2810         rcu_read_unlock();
2811
2812         if (unlikely(blocked_rdev)) {
2813                 if (s.syncing || s.expanding || s.expanded ||
2814                     s.to_write || s.written) {
2815                         set_bit(STRIPE_HANDLE, &sh->state);
2816                         goto unlock;
2817                 }
2818                 /* There is nothing for the blocked_rdev to block */
2819                 rdev_dec_pending(blocked_rdev, conf->mddev);
2820                 blocked_rdev = NULL;
2821         }
2822
2823         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
2824                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
2825                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
2826         }
2827
2828         pr_debug("locked=%d uptodate=%d to_read=%d"
2829                 " to_write=%d failed=%d failed_num=%d\n",
2830                 s.locked, s.uptodate, s.to_read, s.to_write,
2831                 s.failed, s.failed_num);
2832         /* check if the array has lost two devices and, if so, some requests might
2833          * need to be failed
2834          */
2835         if (s.failed > 1 && s.to_read+s.to_write+s.written)
2836                 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
2837         if (s.failed > 1 && s.syncing) {
2838                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2839                 clear_bit(STRIPE_SYNCING, &sh->state);
2840                 s.syncing = 0;
2841         }
2842
2843         /* might be able to return some write requests if the parity block
2844          * is safe, or on a failed drive
2845          */
2846         dev = &sh->dev[sh->pd_idx];
2847         if ( s.written &&
2848              ((test_bit(R5_Insync, &dev->flags) &&
2849                !test_bit(R5_LOCKED, &dev->flags) &&
2850                test_bit(R5_UPTODATE, &dev->flags)) ||
2851                (s.failed == 1 && s.failed_num == sh->pd_idx)))
2852                 handle_stripe_clean_event(conf, sh, disks, &return_bi);
2853
2854         /* Now we might consider reading some blocks, either to check/generate
2855          * parity, or to satisfy requests
2856          * or to load a block that is being partially written.
2857          */
2858         if (s.to_read || s.non_overwrite ||
2859             (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
2860                 handle_stripe_fill5(sh, &s, disks);
2861
2862         /* Now we check to see if any write operations have recently
2863          * completed
2864          */
2865         prexor = 0;
2866         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
2867                 prexor = 1;
2868         if (sh->reconstruct_state == reconstruct_state_drain_result ||
2869             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
2870                 sh->reconstruct_state = reconstruct_state_idle;
2871
2872                 /* All the 'written' buffers and the parity block are ready to
2873                  * be written back to disk
2874                  */
2875                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2876                 for (i = disks; i--; ) {
2877                         dev = &sh->dev[i];
2878                         if (test_bit(R5_LOCKED, &dev->flags) &&
2879                                 (i == sh->pd_idx || dev->written)) {
2880                                 pr_debug("Writing block %d\n", i);
2881                                 set_bit(R5_Wantwrite, &dev->flags);
2882                                 if (prexor)
2883                                         continue;
2884                                 if (!test_bit(R5_Insync, &dev->flags) ||
2885                                     (i == sh->pd_idx && s.failed == 0))
2886                                         set_bit(STRIPE_INSYNC, &sh->state);
2887                         }
2888                 }
2889                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2890                         atomic_dec(&conf->preread_active_stripes);
2891                         if (atomic_read(&conf->preread_active_stripes) <
2892                                 IO_THRESHOLD)
2893                                 md_wakeup_thread(conf->mddev->thread);
2894                 }
2895         }
2896
2897         /* Now to consider new write requests and what else, if anything
2898          * should be read.  We do not handle new writes when:
2899          * 1/ A 'write' operation (copy+xor) is already in flight.
2900          * 2/ A 'check' operation is in flight, as it may clobber the parity
2901          *    block.
2902          */
2903         if (s.to_write && !sh->reconstruct_state && !sh->check_state)
2904                 handle_stripe_dirtying5(conf, sh, &s, disks);
2905
2906         /* maybe we need to check and possibly fix the parity for this stripe
2907          * Any reads will already have been scheduled, so we just see if enough
2908          * data is available.  The parity check is held off while parity
2909          * dependent operations are in flight.
2910          */
2911         if (sh->check_state ||
2912             (s.syncing && s.locked == 0 &&
2913              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
2914              !test_bit(STRIPE_INSYNC, &sh->state)))
2915                 handle_parity_checks5(conf, sh, &s, disks);
2916
2917         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2918                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2919                 clear_bit(STRIPE_SYNCING, &sh->state);
2920         }
2921
2922         /* If the failed drive is just a ReadError, then we might need to progress
2923          * the repair/check process
2924          */
2925         if (s.failed == 1 && !conf->mddev->ro &&
2926             test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2927             && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2928             && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2929                 ) {
2930                 dev = &sh->dev[s.failed_num];
2931                 if (!test_bit(R5_ReWrite, &dev->flags)) {
2932                         set_bit(R5_Wantwrite, &dev->flags);
2933                         set_bit(R5_ReWrite, &dev->flags);
2934                         set_bit(R5_LOCKED, &dev->flags);
2935                         s.locked++;
2936                 } else {
2937                         /* let's read it back */
2938                         set_bit(R5_Wantread, &dev->flags);
2939                         set_bit(R5_LOCKED, &dev->flags);
2940                         s.locked++;
2941                 }
2942         }
2943
2944         /* Finish reconstruct operations initiated by the expansion process */
2945         if (sh->reconstruct_state == reconstruct_state_result) {
2946                 struct stripe_head *sh2
2947                         = get_active_stripe(conf, sh->sector, 1, 1);
2948                 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
2949                         /* sh cannot be written until sh2 has been read.
2950                          * so arrange for sh to be delayed a little
2951                          */
2952                         set_bit(STRIPE_DELAYED, &sh->state);
2953                         set_bit(STRIPE_HANDLE, &sh->state);
2954                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
2955                                               &sh2->state))
2956                                 atomic_inc(&conf->preread_active_stripes);
2957                         release_stripe(sh2);
2958                         goto unlock;
2959                 }
2960                 if (sh2)
2961                         release_stripe(sh2);
2962
2963                 sh->reconstruct_state = reconstruct_state_idle;
2964                 clear_bit(STRIPE_EXPANDING, &sh->state);
2965                 for (i = conf->raid_disks; i--; ) {
2966                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
2967                         set_bit(R5_LOCKED, &sh->dev[i].flags);
2968                         s.locked++;
2969                 }
2970         }
2971
2972         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
2973             !sh->reconstruct_state) {
2974                 /* Need to write out all blocks after computing parity */
2975                 sh->disks = conf->raid_disks;
2976                 stripe_set_idx(sh->sector, conf, 0, sh);
2977                 schedule_reconstruction5(sh, &s, 1, 1);
2978         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
2979                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2980                 atomic_dec(&conf->reshape_stripes);
2981                 wake_up(&conf->wait_for_overlap);
2982                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2983         }
2984
2985         if (s.expanding && s.locked == 0 &&
2986             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
2987                 handle_stripe_expansion(conf, sh, NULL);
2988
2989  unlock:
2990         spin_unlock(&sh->lock);
2991
2992         /* wait for this device to become unblocked */
2993         if (unlikely(blocked_rdev))
2994                 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
2995
2996         if (s.ops_request)
2997                 raid5_run_ops(sh, s.ops_request);
2998
2999         ops_run_io(sh, &s);
3000
3001         return_io(return_bi);
3002
3003         return blocked_rdev == NULL;
3004 }
3005
3006 static bool handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
3007 {
3008         raid5_conf_t *conf = sh->raid_conf;
3009         int disks = sh->disks;
3010         struct bio *return_bi = NULL;
3011         int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx;
3012         struct stripe_head_state s;
3013         struct r6_state r6s;
3014         struct r5dev *dev, *pdev, *qdev;
3015         mdk_rdev_t *blocked_rdev = NULL;
3016
3017         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3018                 "pd_idx=%d, qd_idx=%d\n",
3019                (unsigned long long)sh->sector, sh->state,
3020                atomic_read(&sh->count), pd_idx, qd_idx);
3021         memset(&s, 0, sizeof(s));
3022
3023         spin_lock(&sh->lock);
3024         clear_bit(STRIPE_HANDLE, &sh->state);
3025         clear_bit(STRIPE_DELAYED, &sh->state);
3026
3027         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3028         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3029         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3030         /* Now to look around and see what can be done */
3031
3032         rcu_read_lock();
3033         for (i=disks; i--; ) {
3034                 mdk_rdev_t *rdev;
3035                 dev = &sh->dev[i];
3036                 clear_bit(R5_Insync, &dev->flags);
3037
3038                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3039                         i, dev->flags, dev->toread, dev->towrite, dev->written);
3040                 /* maybe we can reply to a read */
3041                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
3042                         struct bio *rbi, *rbi2;
3043                         pr_debug("Return read for disc %d\n", i);
3044                         spin_lock_irq(&conf->device_lock);
3045                         rbi = dev->toread;
3046                         dev->toread = NULL;
3047                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
3048                                 wake_up(&conf->wait_for_overlap);
3049                         spin_unlock_irq(&conf->device_lock);
3050                         while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
3051                                 copy_data(0, rbi, dev->page, dev->sector);
3052                                 rbi2 = r5_next_bio(rbi, dev->sector);
3053                                 spin_lock_irq(&conf->device_lock);
3054                                 if (!raid5_dec_bi_phys_segments(rbi)) {
3055                                         rbi->bi_next = return_bi;
3056                                         return_bi = rbi;
3057                                 }
3058                                 spin_unlock_irq(&conf->device_lock);
3059                                 rbi = rbi2;
3060                         }
3061                 }
3062
3063                 /* now count some things */
3064                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3065                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3066
3067
3068                 if (dev->toread)
3069                         s.to_read++;
3070                 if (dev->towrite) {
3071                         s.to_write++;
3072                         if (!test_bit(R5_OVERWRITE, &dev->flags))
3073                                 s.non_overwrite++;
3074                 }
3075                 if (dev->written)
3076                         s.written++;
3077                 rdev = rcu_dereference(conf->disks[i].rdev);
3078                 if (blocked_rdev == NULL &&
3079                     rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3080                         blocked_rdev = rdev;
3081                         atomic_inc(&rdev->nr_pending);
3082                 }
3083                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3084                         /* The ReadError flag will just be confusing now */
3085                         clear_bit(R5_ReadError, &dev->flags);
3086                         clear_bit(R5_ReWrite, &dev->flags);
3087                 }
3088                 if (!rdev || !test_bit(In_sync, &rdev->flags)
3089                     || test_bit(R5_ReadError, &dev->flags)) {
3090                         if (s.failed < 2)
3091                                 r6s.failed_num[s.failed] = i;
3092                         s.failed++;
3093                 } else
3094                         set_bit(R5_Insync, &dev->flags);
3095         }
3096         rcu_read_unlock();
3097
3098         if (unlikely(blocked_rdev)) {
3099                 if (s.syncing || s.expanding || s.expanded ||
3100                     s.to_write || s.written) {
3101                         set_bit(STRIPE_HANDLE, &sh->state);
3102                         goto unlock;
3103                 }
3104                 /* There is nothing for the blocked_rdev to block */
3105                 rdev_dec_pending(blocked_rdev, conf->mddev);
3106                 blocked_rdev = NULL;
3107         }
3108
3109         pr_debug("locked=%d uptodate=%d to_read=%d"
3110                " to_write=%d failed=%d failed_num=%d,%d\n",
3111                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3112                r6s.failed_num[0], r6s.failed_num[1]);
3113         /* check if the array has lost >2 devices and, if so, some requests
3114          * might need to be failed
3115          */
3116         if (s.failed > 2 && s.to_read+s.to_write+s.written)
3117                 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3118         if (s.failed > 2 && s.syncing) {
3119                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3120                 clear_bit(STRIPE_SYNCING, &sh->state);
3121                 s.syncing = 0;
3122         }
3123
3124         /*
3125          * might be able to return some write requests if the parity blocks
3126          * are safe, or on a failed drive
3127          */
3128         pdev = &sh->dev[pd_idx];
3129         r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3130                 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3131         qdev = &sh->dev[qd_idx];
3132         r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == qd_idx)
3133                 || (s.failed >= 2 && r6s.failed_num[1] == qd_idx);
3134
3135         if ( s.written &&
3136              ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3137                              && !test_bit(R5_LOCKED, &pdev->flags)
3138                              && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3139              ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3140                              && !test_bit(R5_LOCKED, &qdev->flags)
3141                              && test_bit(R5_UPTODATE, &qdev->flags)))))
3142                 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3143
3144         /* Now we might consider reading some blocks, either to check/generate
3145          * parity, or to satisfy requests
3146          * or to load a block that is being partially written.
3147          */
3148         if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3149             (s.syncing && (s.uptodate < disks)) || s.expanding)
3150                 handle_stripe_fill6(sh, &s, &r6s, disks);
3151
3152         /* now to consider writing and what else, if anything should be read */
3153         if (s.to_write)
3154                 handle_stripe_dirtying6(conf, sh, &s, &r6s, disks);
3155
3156         /* maybe we need to check and possibly fix the parity for this stripe
3157          * Any reads will already have been scheduled, so we just see if enough
3158          * data is available
3159          */
3160         if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
3161                 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
3162
3163         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3164                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3165                 clear_bit(STRIPE_SYNCING, &sh->state);
3166         }
3167
3168         /* If the failed drives are just a ReadError, then we might need
3169          * to progress the repair/check process
3170          */
3171         if (s.failed <= 2 && !conf->mddev->ro)
3172                 for (i = 0; i < s.failed; i++) {
3173                         dev = &sh->dev[r6s.failed_num[i]];
3174                         if (test_bit(R5_ReadError, &dev->flags)
3175                             && !test_bit(R5_LOCKED, &dev->flags)
3176                             && test_bit(R5_UPTODATE, &dev->flags)
3177                                 ) {
3178                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
3179                                         set_bit(R5_Wantwrite, &dev->flags);
3180                                         set_bit(R5_ReWrite, &dev->flags);
3181                                         set_bit(R5_LOCKED, &dev->flags);
3182                                 } else {
3183                                         /* let's read it back */
3184                                         set_bit(R5_Wantread, &dev->flags);
3185                                         set_bit(R5_LOCKED, &dev->flags);
3186                                 }
3187                         }
3188                 }
3189
3190         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
3191                 struct stripe_head *sh2
3192                         = get_active_stripe(conf, sh->sector, 1, 1);
3193                 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3194                         /* sh cannot be written until sh2 has been read.
3195                          * so arrange for sh to be delayed a little
3196                          */
3197                         set_bit(STRIPE_DELAYED, &sh->state);
3198                         set_bit(STRIPE_HANDLE, &sh->state);
3199                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3200                                               &sh2->state))
3201                                 atomic_inc(&conf->preread_active_stripes);
3202                         release_stripe(sh2);
3203                         goto unlock;
3204                 }
3205                 if (sh2)
3206                         release_stripe(sh2);
3207
3208                 /* Need to write out all blocks after computing P&Q */
3209                 sh->disks = conf->raid_disks;
3210                 stripe_set_idx(sh->sector, conf, 0, sh);
3211                 compute_parity6(sh, RECONSTRUCT_WRITE);
3212                 for (i = conf->raid_disks ; i-- ;  ) {
3213                         set_bit(R5_LOCKED, &sh->dev[i].flags);
3214                         s.locked++;
3215                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
3216                 }
3217                 clear_bit(STRIPE_EXPANDING, &sh->state);
3218         } else if (s.expanded) {
3219                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3220                 atomic_dec(&conf->reshape_stripes);
3221                 wake_up(&conf->wait_for_overlap);
3222                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3223         }
3224
3225         if (s.expanding && s.locked == 0 &&
3226             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3227                 handle_stripe_expansion(conf, sh, &r6s);
3228
3229  unlock:
3230         spin_unlock(&sh->lock);
3231
3232         /* wait for this device to become unblocked */
3233         if (unlikely(blocked_rdev))
3234                 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3235
3236         ops_run_io(sh, &s);
3237
3238         return_io(return_bi);
3239
3240         return blocked_rdev == NULL;
3241 }
3242
3243 /* returns true if the stripe was handled */
3244 static bool handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3245 {
3246         if (sh->raid_conf->level == 6)
3247                 return handle_stripe6(sh, tmp_page);
3248         else
3249                 return handle_stripe5(sh);
3250 }
3251
3252
3253
3254 static void raid5_activate_delayed(raid5_conf_t *conf)
3255 {
3256         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3257                 while (!list_empty(&conf->delayed_list)) {
3258                         struct list_head *l = conf->delayed_list.next;
3259                         struct stripe_head *sh;
3260                         sh = list_entry(l, struct stripe_head, lru);
3261                         list_del_init(l);
3262                         clear_bit(STRIPE_DELAYED, &sh->state);
3263                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3264                                 atomic_inc(&conf->preread_active_stripes);
3265                         list_add_tail(&sh->lru, &conf->hold_list);
3266                 }
3267         } else
3268                 blk_plug_device(conf->mddev->queue);
3269 }
3270
3271 static void activate_bit_delay(raid5_conf_t *conf)
3272 {
3273         /* device_lock is held */
3274         struct list_head head;
3275         list_add(&head, &conf->bitmap_list);
3276         list_del_init(&conf->bitmap_list);
3277         while (!list_empty(&head)) {
3278                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3279                 list_del_init(&sh->lru);
3280                 atomic_inc(&sh->count);
3281                 __release_stripe(conf, sh);
3282         }
3283 }
3284
3285 static void unplug_slaves(mddev_t *mddev)
3286 {
3287         raid5_conf_t *conf = mddev_to_conf(mddev);
3288         int i;
3289
3290         rcu_read_lock();
3291         for (i=0; i<mddev->raid_disks; i++) {
3292                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3293                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3294                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3295
3296                         atomic_inc(&rdev->nr_pending);
3297                         rcu_read_unlock();
3298
3299                         blk_unplug(r_queue);
3300
3301                         rdev_dec_pending(rdev, mddev);
3302                         rcu_read_lock();
3303                 }
3304         }
3305         rcu_read_unlock();
3306 }
3307
3308 static void raid5_unplug_device(struct request_queue *q)
3309 {
3310         mddev_t *mddev = q->queuedata;
3311         raid5_conf_t *conf = mddev_to_conf(mddev);
3312         unsigned long flags;
3313
3314         spin_lock_irqsave(&conf->device_lock, flags);
3315
3316         if (blk_remove_plug(q)) {
3317                 conf->seq_flush++;
3318                 raid5_activate_delayed(conf);
3319         }
3320         md_wakeup_thread(mddev->thread);
3321
3322         spin_unlock_irqrestore(&conf->device_lock, flags);
3323
3324         unplug_slaves(mddev);
3325 }
3326
3327 static int raid5_congested(void *data, int bits)
3328 {
3329         mddev_t *mddev = data;
3330         raid5_conf_t *conf = mddev_to_conf(mddev);
3331
3332         /* No difference between reads and writes.  Just check
3333          * how busy the stripe_cache is
3334          */
3335         if (conf->inactive_blocked)
3336                 return 1;
3337         if (conf->quiesce)
3338                 return 1;
3339         if (list_empty_careful(&conf->inactive_list))
3340                 return 1;
3341
3342         return 0;
3343 }
3344
3345 /* We want read requests to align with chunks where possible,
3346  * but write requests don't need to.
3347  */
3348 static int raid5_mergeable_bvec(struct request_queue *q,
3349                                 struct bvec_merge_data *bvm,
3350                                 struct bio_vec *biovec)
3351 {
3352         mddev_t *mddev = q->queuedata;
3353         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3354         int max;
3355         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3356         unsigned int bio_sectors = bvm->bi_size >> 9;
3357
3358         if ((bvm->bi_rw & 1) == WRITE)
3359                 return biovec->bv_len; /* always allow writes to be mergeable */
3360
3361         if (mddev->new_chunk < mddev->chunk_size)
3362                 chunk_sectors = mddev->new_chunk >> 9;
3363         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3364         if (max < 0) max = 0;
3365         if (max <= biovec->bv_len && bio_sectors == 0)
3366                 return biovec->bv_len;
3367         else
3368                 return max;
3369 }
3370
3371
3372 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3373 {
3374         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3375         unsigned int chunk_sectors = mddev->chunk_size >> 9;
3376         unsigned int bio_sectors = bio->bi_size >> 9;
3377
3378         if (mddev->new_chunk < mddev->chunk_size)
3379                 chunk_sectors = mddev->new_chunk >> 9;
3380         return  chunk_sectors >=
3381                 ((sector & (chunk_sectors - 1)) + bio_sectors);
3382 }
3383
3384 /*
3385  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3386  *  later sampled by raid5d.
3387  */
3388 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3389 {
3390         unsigned long flags;
3391
3392         spin_lock_irqsave(&conf->device_lock, flags);
3393
3394         bi->bi_next = conf->retry_read_aligned_list;
3395         conf->retry_read_aligned_list = bi;
3396
3397         spin_unlock_irqrestore(&conf->device_lock, flags);
3398         md_wakeup_thread(conf->mddev->thread);
3399 }
3400
3401
3402 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3403 {
3404         struct bio *bi;
3405
3406         bi = conf->retry_read_aligned;
3407         if (bi) {
3408                 conf->retry_read_aligned = NULL;
3409                 return bi;
3410         }
3411         bi = conf->retry_read_aligned_list;
3412         if(bi) {
3413                 conf->retry_read_aligned_list = bi->bi_next;
3414                 bi->bi_next = NULL;
3415                 /*
3416                  * this sets the active strip count to 1 and the processed
3417                  * strip count to zero (upper 8 bits)
3418                  */
3419                 bi->bi_phys_segments = 1; /* biased count of active stripes */
3420         }
3421
3422         return bi;
3423 }
3424
3425
3426 /*
3427  *  The "raid5_align_endio" should check if the read succeeded and if it
3428  *  did, call bio_endio on the original bio (having bio_put the new bio
3429  *  first).
3430  *  If the read failed..
3431  */
3432 static void raid5_align_endio(struct bio *bi, int error)
3433 {
3434         struct bio* raid_bi  = bi->bi_private;
3435         mddev_t *mddev;
3436         raid5_conf_t *conf;
3437         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3438         mdk_rdev_t *rdev;
3439
3440         bio_put(bi);
3441
3442         mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3443         conf = mddev_to_conf(mddev);
3444         rdev = (void*)raid_bi->bi_next;
3445         raid_bi->bi_next = NULL;
3446
3447         rdev_dec_pending(rdev, conf->mddev);
3448
3449         if (!error && uptodate) {
3450                 bio_endio(raid_bi, 0);
3451                 if (atomic_dec_and_test(&conf->active_aligned_reads))
3452                         wake_up(&conf->wait_for_stripe);
3453                 return;
3454         }
3455
3456
3457         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3458
3459         add_bio_to_retry(raid_bi, conf);
3460 }
3461
3462 static int bio_fits_rdev(struct bio *bi)
3463 {
3464         struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3465
3466         if ((bi->bi_size>>9) > q->max_sectors)
3467                 return 0;
3468         blk_recount_segments(q, bi);
3469         if (bi->bi_phys_segments > q->max_phys_segments)
3470                 return 0;
3471
3472         if (q->merge_bvec_fn)
3473                 /* it's too hard to apply the merge_bvec_fn at this stage,
3474                  * just just give up
3475                  */
3476                 return 0;
3477
3478         return 1;
3479 }
3480
3481
3482 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3483 {
3484         mddev_t *mddev = q->queuedata;
3485         raid5_conf_t *conf = mddev_to_conf(mddev);
3486         unsigned int dd_idx;
3487         struct bio* align_bi;
3488         mdk_rdev_t *rdev;
3489
3490         if (!in_chunk_boundary(mddev, raid_bio)) {
3491                 pr_debug("chunk_aligned_read : non aligned\n");
3492                 return 0;
3493         }
3494         /*
3495          * use bio_clone to make a copy of the bio
3496          */
3497         align_bi = bio_clone(raid_bio, GFP_NOIO);
3498         if (!align_bi)
3499                 return 0;
3500         /*
3501          *   set bi_end_io to a new function, and set bi_private to the
3502          *     original bio.
3503          */
3504         align_bi->bi_end_io  = raid5_align_endio;
3505         align_bi->bi_private = raid_bio;
3506         /*
3507          *      compute position
3508          */
3509         align_bi->bi_sector =  raid5_compute_sector(conf, raid_bio->bi_sector,
3510                                                     0,
3511                                                     &dd_idx, NULL);
3512
3513         rcu_read_lock();
3514         rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3515         if (rdev && test_bit(In_sync, &rdev->flags)) {
3516                 atomic_inc(&rdev->nr_pending);
3517                 rcu_read_unlock();
3518                 raid_bio->bi_next = (void*)rdev;
3519                 align_bi->bi_bdev =  rdev->bdev;
3520                 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3521                 align_bi->bi_sector += rdev->data_offset;
3522
3523                 if (!bio_fits_rdev(align_bi)) {
3524                         /* too big in some way */
3525                         bio_put(align_bi);
3526                         rdev_dec_pending(rdev, mddev);
3527                         return 0;
3528                 }
3529
3530                 spin_lock_irq(&conf->device_lock);
3531                 wait_event_lock_irq(conf->wait_for_stripe,
3532                                     conf->quiesce == 0,
3533                                     conf->device_lock, /* nothing */);
3534                 atomic_inc(&conf->active_aligned_reads);
3535                 spin_unlock_irq(&conf->device_lock);
3536
3537                 generic_make_request(align_bi);
3538                 return 1;
3539         } else {
3540                 rcu_read_unlock();
3541                 bio_put(align_bi);
3542                 return 0;
3543         }
3544 }
3545
3546 /* __get_priority_stripe - get the next stripe to process
3547  *
3548  * Full stripe writes are allowed to pass preread active stripes up until
3549  * the bypass_threshold is exceeded.  In general the bypass_count
3550  * increments when the handle_list is handled before the hold_list; however, it
3551  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3552  * stripe with in flight i/o.  The bypass_count will be reset when the
3553  * head of the hold_list has changed, i.e. the head was promoted to the
3554  * handle_list.
3555  */
3556 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3557 {
3558         struct stripe_head *sh;
3559
3560         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3561                   __func__,
3562                   list_empty(&conf->handle_list) ? "empty" : "busy",
3563                   list_empty(&conf->hold_list) ? "empty" : "busy",
3564                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
3565
3566         if (!list_empty(&conf->handle_list)) {
3567                 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3568
3569                 if (list_empty(&conf->hold_list))
3570                         conf->bypass_count = 0;
3571                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3572                         if (conf->hold_list.next == conf->last_hold)
3573                                 conf->bypass_count++;
3574                         else {
3575                                 conf->last_hold = conf->hold_list.next;
3576                                 conf->bypass_count -= conf->bypass_threshold;
3577                                 if (conf->bypass_count < 0)
3578                                         conf->bypass_count = 0;
3579                         }
3580                 }
3581         } else if (!list_empty(&conf->hold_list) &&
3582                    ((conf->bypass_threshold &&
3583                      conf->bypass_count > conf->bypass_threshold) ||
3584                     atomic_read(&conf->pending_full_writes) == 0)) {
3585                 sh = list_entry(conf->hold_list.next,
3586                                 typeof(*sh), lru);
3587                 conf->bypass_count -= conf->bypass_threshold;
3588                 if (conf->bypass_count < 0)
3589                         conf->bypass_count = 0;
3590         } else
3591                 return NULL;
3592
3593         list_del_init(&sh->lru);
3594         atomic_inc(&sh->count);
3595         BUG_ON(atomic_read(&sh->count) != 1);
3596         return sh;
3597 }
3598
3599 static int make_request(struct request_queue *q, struct bio * bi)
3600 {
3601         mddev_t *mddev = q->queuedata;
3602         raid5_conf_t *conf = mddev_to_conf(mddev);
3603         int dd_idx;
3604         sector_t new_sector;
3605         sector_t logical_sector, last_sector;
3606         struct stripe_head *sh;
3607         const int rw = bio_data_dir(bi);
3608         int cpu, remaining;
3609
3610         if (unlikely(bio_barrier(bi))) {
3611                 bio_endio(bi, -EOPNOTSUPP);
3612                 return 0;
3613         }
3614
3615         md_write_start(mddev, bi);
3616
3617         cpu = part_stat_lock();
3618         part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
3619         part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
3620                       bio_sectors(bi));
3621         part_stat_unlock();
3622
3623         if (rw == READ &&
3624              mddev->reshape_position == MaxSector &&
3625              chunk_aligned_read(q,bi))
3626                 return 0;
3627
3628         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3629         last_sector = bi->bi_sector + (bi->bi_size>>9);
3630         bi->bi_next = NULL;
3631         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
3632
3633         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3634                 DEFINE_WAIT(w);
3635                 int disks, data_disks;
3636                 int previous;
3637
3638         retry:
3639                 previous = 0;
3640                 disks = conf->raid_disks;
3641                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3642                 if (unlikely(conf->reshape_progress != MaxSector)) {
3643                         /* spinlock is needed as reshape_progress may be
3644                          * 64bit on a 32bit platform, and so it might be
3645                          * possible to see a half-updated value
3646                          * Ofcourse reshape_progress could change after
3647                          * the lock is dropped, so once we get a reference
3648                          * to the stripe that we think it is, we will have
3649                          * to check again.
3650                          */
3651                         spin_lock_irq(&conf->device_lock);
3652                         if (mddev->delta_disks < 0
3653                             ? logical_sector < conf->reshape_progress
3654                             : logical_sector >= conf->reshape_progress) {
3655                                 disks = conf->previous_raid_disks;
3656                                 previous = 1;
3657                         } else {
3658                                 if (mddev->delta_disks < 0
3659                                     ? logical_sector < conf->reshape_safe
3660                                     : logical_sector >= conf->reshape_safe) {
3661                                         spin_unlock_irq(&conf->device_lock);
3662                                         schedule();
3663                                         goto retry;
3664                                 }
3665                         }
3666                         spin_unlock_irq(&conf->device_lock);
3667                 }
3668                 data_disks = disks - conf->max_degraded;
3669
3670                 new_sector = raid5_compute_sector(conf, logical_sector,
3671                                                   previous,
3672                                                   &dd_idx, NULL);
3673                 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3674                         (unsigned long long)new_sector, 
3675                         (unsigned long long)logical_sector);
3676
3677                 sh = get_active_stripe(conf, new_sector, previous,
3678                                        (bi->bi_rw&RWA_MASK));
3679                 if (sh) {
3680                         if (unlikely(previous)) {
3681                                 /* expansion might have moved on while waiting for a
3682                                  * stripe, so we must do the range check again.
3683                                  * Expansion could still move past after this
3684                                  * test, but as we are holding a reference to
3685                                  * 'sh', we know that if that happens,
3686                                  *  STRIPE_EXPANDING will get set and the expansion
3687                                  * won't proceed until we finish with the stripe.
3688                                  */
3689                                 int must_retry = 0;
3690                                 spin_lock_irq(&conf->device_lock);
3691                                 if (mddev->delta_disks < 0
3692                                     ? logical_sector >= conf->reshape_progress
3693                                     : logical_sector < conf->reshape_progress)
3694                                         /* mismatch, need to try again */
3695                                         must_retry = 1;
3696                                 spin_unlock_irq(&conf->device_lock);
3697                                 if (must_retry) {
3698                                         release_stripe(sh);
3699                                         goto retry;
3700                                 }
3701                         }
3702                         /* FIXME what if we get a false positive because these
3703                          * are being updated.
3704                          */
3705                         if (logical_sector >= mddev->suspend_lo &&
3706                             logical_sector < mddev->suspend_hi) {
3707                                 release_stripe(sh);
3708                                 schedule();
3709                                 goto retry;
3710                         }
3711
3712                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3713                             !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3714                                 /* Stripe is busy expanding or
3715                                  * add failed due to overlap.  Flush everything
3716                                  * and wait a while
3717                                  */
3718                                 raid5_unplug_device(mddev->queue);
3719                                 release_stripe(sh);
3720                                 schedule();
3721                                 goto retry;
3722                         }
3723                         finish_wait(&conf->wait_for_overlap, &w);
3724                         set_bit(STRIPE_HANDLE, &sh->state);
3725                         clear_bit(STRIPE_DELAYED, &sh->state);
3726                         release_stripe(sh);
3727                 } else {
3728                         /* cannot get stripe for read-ahead, just give-up */
3729                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
3730                         finish_wait(&conf->wait_for_overlap, &w);
3731                         break;
3732                 }
3733                         
3734         }
3735         spin_lock_irq(&conf->device_lock);
3736         remaining = raid5_dec_bi_phys_segments(bi);
3737         spin_unlock_irq(&conf->device_lock);
3738         if (remaining == 0) {
3739
3740                 if ( rw == WRITE )
3741                         md_write_end(mddev);
3742
3743                 bio_endio(bi, 0);
3744         }
3745         return 0;
3746 }
3747
3748 static sector_t raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks);
3749
3750 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3751 {
3752         /* reshaping is quite different to recovery/resync so it is
3753          * handled quite separately ... here.
3754          *
3755          * On each call to sync_request, we gather one chunk worth of
3756          * destination stripes and flag them as expanding.
3757          * Then we find all the source stripes and request reads.
3758          * As the reads complete, handle_stripe will copy the data
3759          * into the destination stripe and release that stripe.
3760          */
3761         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3762         struct stripe_head *sh;
3763         sector_t first_sector, last_sector;
3764         int raid_disks = conf->previous_raid_disks;
3765         int data_disks = raid_disks - conf->max_degraded;
3766         int new_data_disks = conf->raid_disks - conf->max_degraded;
3767         int i;
3768         int dd_idx;
3769         sector_t writepos, readpos, safepos;
3770         sector_t stripe_addr;
3771         int reshape_sectors;
3772         struct list_head stripes;
3773
3774         if (sector_nr == 0) {
3775                 /* If restarting in the middle, skip the initial sectors */
3776                 if (mddev->delta_disks < 0 &&
3777                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
3778                         sector_nr = raid5_size(mddev, 0, 0)
3779                                 - conf->reshape_progress;
3780                 } else if (mddev->delta_disks > 0 &&
3781                            conf->reshape_progress > 0)
3782                         sector_nr = conf->reshape_progress;
3783                 sector_div(sector_nr, new_data_disks);
3784                 if (sector_nr) {
3785                         *skipped = 1;
3786                         return sector_nr;
3787                 }
3788         }
3789
3790         /* We need to process a full chunk at a time.
3791          * If old and new chunk sizes differ, we need to process the
3792          * largest of these
3793          */
3794         if (mddev->new_chunk > mddev->chunk_size)
3795                 reshape_sectors = mddev->new_chunk / 512;
3796         else
3797                 reshape_sectors = mddev->chunk_size / 512;
3798
3799         /* we update the metadata when there is more than 3Meg
3800          * in the block range (that is rather arbitrary, should
3801          * probably be time based) or when the data about to be
3802          * copied would over-write the source of the data at
3803          * the front of the range.
3804          * i.e. one new_stripe along from reshape_progress new_maps
3805          * to after where reshape_safe old_maps to
3806          */
3807         writepos = conf->reshape_progress;
3808         sector_div(writepos, new_data_disks);
3809         readpos = conf->reshape_progress;
3810         sector_div(readpos, data_disks);
3811         safepos = conf->reshape_safe;
3812         sector_div(safepos, data_disks);
3813         if (mddev->delta_disks < 0) {
3814                 writepos -= reshape_sectors;
3815                 readpos += reshape_sectors;
3816                 safepos += reshape_sectors;
3817         } else {
3818                 writepos += reshape_sectors;
3819                 readpos -= reshape_sectors;
3820                 safepos -= reshape_sectors;
3821         }
3822
3823         /* 'writepos' is the most advanced device address we might write.
3824          * 'readpos' is the least advanced device address we might read.
3825          * 'safepos' is the least address recorded in the metadata as having
3826          *     been reshaped.
3827          * If 'readpos' is behind 'writepos', then there is no way that we can
3828          * ensure safety in the face of a crash - that must be done by userspace
3829          * making a backup of the data.  So in that case there is no particular
3830          * rush to update metadata.
3831          * Otherwise if 'safepos' is behind 'writepos', then we really need to
3832          * update the metadata to advance 'safepos' to match 'readpos' so that
3833          * we can be safe in the event of a crash.
3834          * So we insist on updating metadata if safepos is behind writepos and
3835          * readpos is beyond writepos.
3836          * In any case, update the metadata every 10 seconds.
3837          * Maybe that number should be configurable, but I'm not sure it is
3838          * worth it.... maybe it could be a multiple of safemode_delay???
3839          */
3840         if ((mddev->delta_disks < 0
3841              ? (safepos > writepos && readpos < writepos)
3842              : (safepos < writepos && readpos > writepos)) ||
3843             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
3844                 /* Cannot proceed until we've updated the superblock... */
3845                 wait_event(conf->wait_for_overlap,
3846                            atomic_read(&conf->reshape_stripes)==0);
3847                 mddev->reshape_position = conf->reshape_progress;
3848                 mddev->curr_resync_completed = mddev->curr_resync;
3849                 conf->reshape_checkpoint = jiffies;
3850                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3851                 md_wakeup_thread(mddev->thread);
3852                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3853                            kthread_should_stop());
3854                 spin_lock_irq(&conf->device_lock);
3855                 conf->reshape_safe = mddev->reshape_position;
3856                 spin_unlock_irq(&conf->device_lock);
3857                 wake_up(&conf->wait_for_overlap);
3858                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3859         }
3860
3861         if (mddev->delta_disks < 0) {
3862                 BUG_ON(conf->reshape_progress == 0);
3863                 stripe_addr = writepos;
3864                 BUG_ON((mddev->dev_sectors &
3865                         ~((sector_t)reshape_sectors - 1))
3866                        - reshape_sectors - stripe_addr
3867                        != sector_nr);
3868         } else {
3869                 BUG_ON(writepos != sector_nr + reshape_sectors);
3870                 stripe_addr = sector_nr;
3871         }
3872         INIT_LIST_HEAD(&stripes);
3873         for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
3874                 int j;
3875                 int skipped = 0;
3876                 sh = get_active_stripe(conf, stripe_addr+i, 0, 0);
3877                 set_bit(STRIPE_EXPANDING, &sh->state);
3878                 atomic_inc(&conf->reshape_stripes);
3879                 /* If any of this stripe is beyond the end of the old
3880                  * array, then we need to zero those blocks
3881                  */
3882                 for (j=sh->disks; j--;) {
3883                         sector_t s;
3884                         if (j == sh->pd_idx)
3885                                 continue;
3886                         if (conf->level == 6 &&
3887                             j == sh->qd_idx)
3888                                 continue;
3889                         s = compute_blocknr(sh, j, 0);
3890                         if (s < raid5_size(mddev, 0, 0)) {
3891                                 skipped = 1;
3892                                 continue;
3893                         }
3894                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3895                         set_bit(R5_Expanded, &sh->dev[j].flags);
3896                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
3897                 }
3898                 if (!skipped) {
3899                         set_bit(STRIPE_EXPAND_READY, &sh->state);
3900                         set_bit(STRIPE_HANDLE, &sh->state);
3901                 }
3902                 list_add(&sh->lru, &stripes);
3903         }
3904         spin_lock_irq(&conf->device_lock);
3905         if (mddev->delta_disks < 0)
3906                 conf->reshape_progress -= reshape_sectors * new_data_disks;
3907         else
3908                 conf->reshape_progress += reshape_sectors * new_data_disks;
3909         spin_unlock_irq(&conf->device_lock);
3910         /* Ok, those stripe are ready. We can start scheduling
3911          * reads on the source stripes.
3912          * The source stripes are determined by mapping the first and last
3913          * block on the destination stripes.
3914          */
3915         first_sector =
3916                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
3917                                      1, &dd_idx, NULL);
3918         last_sector =
3919                 raid5_compute_sector(conf, ((stripe_addr+conf->chunk_size/512)
3920                                             *(new_data_disks) - 1),
3921                                      1, &dd_idx, NULL);
3922         if (last_sector >= mddev->dev_sectors)
3923                 last_sector = mddev->dev_sectors - 1;
3924         while (first_sector <= last_sector) {
3925                 sh = get_active_stripe(conf, first_sector, 1, 0);
3926                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3927                 set_bit(STRIPE_HANDLE, &sh->state);
3928                 release_stripe(sh);
3929                 first_sector += STRIPE_SECTORS;
3930         }
3931         /* Now that the sources are clearly marked, we can release
3932          * the destination stripes
3933          */
3934         while (!list_empty(&stripes)) {
3935                 sh = list_entry(stripes.next, struct stripe_head, lru);
3936                 list_del_init(&sh->lru);
3937                 release_stripe(sh);
3938         }
3939         /* If this takes us to the resync_max point where we have to pause,
3940          * then we need to write out the superblock.
3941          */
3942         sector_nr += reshape_sectors;
3943         if ((sector_nr - mddev->curr_resync_completed) * 2
3944             >= mddev->resync_max - mddev->curr_resync_completed) {
3945                 /* Cannot proceed until we've updated the superblock... */
3946                 wait_event(conf->wait_for_overlap,
3947                            atomic_read(&conf->reshape_stripes) == 0);
3948                 mddev->reshape_position = conf->reshape_progress;
3949                 mddev->curr_resync_completed = mddev->curr_resync;
3950                 conf->reshape_checkpoint = jiffies;
3951                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3952                 md_wakeup_thread(mddev->thread);
3953                 wait_event(mddev->sb_wait,
3954                            !test_bit(MD_CHANGE_DEVS, &mddev->flags)
3955                            || kthread_should_stop());
3956                 spin_lock_irq(&conf->device_lock);
3957                 conf->reshape_safe = mddev->reshape_position;
3958                 spin_unlock_irq(&conf->device_lock);
3959                 wake_up(&conf->wait_for_overlap);
3960                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3961         }
3962         return reshape_sectors;
3963 }
3964
3965 /* FIXME go_faster isn't used */
3966 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3967 {
3968         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3969         struct stripe_head *sh;
3970         sector_t max_sector = mddev->dev_sectors;
3971         int sync_blocks;
3972         int still_degraded = 0;
3973         int i;
3974
3975         if (sector_nr >= max_sector) {
3976                 /* just being told to finish up .. nothing much to do */
3977                 unplug_slaves(mddev);
3978
3979                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3980                         end_reshape(conf);
3981                         return 0;
3982                 }
3983
3984                 if (mddev->curr_resync < max_sector) /* aborted */
3985                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3986                                         &sync_blocks, 1);
3987                 else /* completed sync */
3988                         conf->fullsync = 0;
3989                 bitmap_close_sync(mddev->bitmap);
3990
3991                 return 0;
3992         }
3993
3994         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3995                 return reshape_request(mddev, sector_nr, skipped);
3996
3997         /* No need to check resync_max as we never do more than one
3998          * stripe, and as resync_max will always be on a chunk boundary,
3999          * if the check in md_do_sync didn't fire, there is no chance
4000          * of overstepping resync_max here
4001          */
4002
4003         /* if there is too many failed drives and we are trying
4004          * to resync, then assert that we are finished, because there is
4005          * nothing we can do.
4006          */
4007         if (mddev->degraded >= conf->max_degraded &&
4008             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4009                 sector_t rv = mddev->dev_sectors - sector_nr;
4010                 *skipped = 1;
4011                 return rv;
4012         }
4013         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4014             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4015             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4016                 /* we can skip this block, and probably more */
4017                 sync_blocks /= STRIPE_SECTORS;
4018                 *skipped = 1;
4019                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4020         }
4021
4022
4023         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4024
4025         sh = get_active_stripe(conf, sector_nr, 0, 1);
4026         if (sh == NULL) {
4027                 sh = get_active_stripe(conf, sector_nr, 0, 0);
4028                 /* make sure we don't swamp the stripe cache if someone else
4029                  * is trying to get access
4030                  */
4031                 schedule_timeout_uninterruptible(1);
4032         }
4033         /* Need to check if array will still be degraded after recovery/resync
4034          * We don't need to check the 'failed' flag as when that gets set,
4035          * recovery aborts.
4036          */
4037         for (i=0; i<mddev->raid_disks; i++)
4038                 if (conf->disks[i].rdev == NULL)
4039                         still_degraded = 1;
4040
4041         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4042
4043         spin_lock(&sh->lock);
4044         set_bit(STRIPE_SYNCING, &sh->state);
4045         clear_bit(STRIPE_INSYNC, &sh->state);
4046         spin_unlock(&sh->lock);
4047
4048         /* wait for any blocked device to be handled */
4049         while(unlikely(!handle_stripe(sh, NULL)))
4050                 ;
4051         release_stripe(sh);
4052
4053         return STRIPE_SECTORS;
4054 }
4055
4056 static int  retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
4057 {
4058         /* We may not be able to submit a whole bio at once as there
4059          * may not be enough stripe_heads available.
4060          * We cannot pre-allocate enough stripe_heads as we may need
4061          * more than exist in the cache (if we allow ever large chunks).
4062          * So we do one stripe head at a time and record in
4063          * ->bi_hw_segments how many have been done.
4064          *
4065          * We *know* that this entire raid_bio is in one chunk, so
4066          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4067          */
4068         struct stripe_head *sh;
4069         int dd_idx;
4070         sector_t sector, logical_sector, last_sector;
4071         int scnt = 0;
4072         int remaining;
4073         int handled = 0;
4074
4075         logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4076         sector = raid5_compute_sector(conf, logical_sector,
4077                                       0, &dd_idx, NULL);
4078         last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4079
4080         for (; logical_sector < last_sector;
4081              logical_sector += STRIPE_SECTORS,
4082                      sector += STRIPE_SECTORS,
4083                      scnt++) {
4084
4085                 if (scnt < raid5_bi_hw_segments(raid_bio))
4086                         /* already done this stripe */
4087                         continue;
4088
4089                 sh = get_active_stripe(conf, sector, 0, 1);
4090
4091                 if (!sh) {
4092                         /* failed to get a stripe - must wait */
4093                         raid5_set_bi_hw_segments(raid_bio, scnt);
4094                         conf->retry_read_aligned = raid_bio;
4095                         return handled;
4096                 }
4097
4098                 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4099                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4100                         release_stripe(sh);
4101                         raid5_set_bi_hw_segments(raid_bio, scnt);
4102                         conf->retry_read_aligned = raid_bio;
4103                         return handled;
4104                 }
4105
4106                 handle_stripe(sh, NULL);
4107                 release_stripe(sh);
4108                 handled++;
4109         }
4110         spin_lock_irq(&conf->device_lock);
4111         remaining = raid5_dec_bi_phys_segments(raid_bio);
4112         spin_unlock_irq(&conf->device_lock);
4113         if (remaining == 0)
4114                 bio_endio(raid_bio, 0);
4115         if (atomic_dec_and_test(&conf->active_aligned_reads))
4116                 wake_up(&conf->wait_for_stripe);
4117         return handled;
4118 }
4119
4120
4121
4122 /*
4123  * This is our raid5 kernel thread.
4124  *
4125  * We scan the hash table for stripes which can be handled now.
4126  * During the scan, completed stripes are saved for us by the interrupt
4127  * handler, so that they will not have to wait for our next wakeup.
4128  */
4129 static void raid5d(mddev_t *mddev)
4130 {
4131         struct stripe_head *sh;
4132         raid5_conf_t *conf = mddev_to_conf(mddev);
4133         int handled;
4134
4135         pr_debug("+++ raid5d active\n");
4136
4137         md_check_recovery(mddev);
4138
4139         handled = 0;
4140         spin_lock_irq(&conf->device_lock);
4141         while (1) {
4142                 struct bio *bio;
4143
4144                 if (conf->seq_flush != conf->seq_write) {
4145                         int seq = conf->seq_flush;
4146                         spin_unlock_irq(&conf->device_lock);
4147                         bitmap_unplug(mddev->bitmap);
4148                         spin_lock_irq(&conf->device_lock);
4149                         conf->seq_write = seq;
4150                         activate_bit_delay(conf);
4151                 }
4152
4153                 while ((bio = remove_bio_from_retry(conf))) {
4154                         int ok;
4155                         spin_unlock_irq(&conf->device_lock);
4156                         ok = retry_aligned_read(conf, bio);
4157                         spin_lock_irq(&conf->device_lock);
4158                         if (!ok)
4159                                 break;
4160                         handled++;
4161                 }
4162
4163                 sh = __get_priority_stripe(conf);
4164
4165                 if (!sh)
4166                         break;
4167                 spin_unlock_irq(&conf->device_lock);
4168                 
4169                 handled++;
4170                 handle_stripe(sh, conf->spare_page);
4171                 release_stripe(sh);
4172
4173                 spin_lock_irq(&conf->device_lock);
4174         }
4175         pr_debug("%d stripes handled\n", handled);
4176
4177         spin_unlock_irq(&conf->device_lock);
4178
4179         async_tx_issue_pending_all();
4180         unplug_slaves(mddev);
4181
4182         pr_debug("--- raid5d inactive\n");
4183 }
4184
4185 static ssize_t
4186 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4187 {
4188         raid5_conf_t *conf = mddev_to_conf(mddev);
4189         if (conf)
4190                 return sprintf(page, "%d\n", conf->max_nr_stripes);
4191         else
4192                 return 0;
4193 }
4194
4195 static ssize_t
4196 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4197 {
4198         raid5_conf_t *conf = mddev_to_conf(mddev);
4199         unsigned long new;
4200         int err;
4201
4202         if (len >= PAGE_SIZE)
4203                 return -EINVAL;
4204         if (!conf)
4205                 return -ENODEV;
4206
4207         if (strict_strtoul(page, 10, &new))
4208                 return -EINVAL;
4209         if (new <= 16 || new > 32768)
4210                 return -EINVAL;
4211         while (new < conf->max_nr_stripes) {
4212                 if (drop_one_stripe(conf))
4213                         conf->max_nr_stripes--;
4214                 else
4215                         break;
4216         }
4217         err = md_allow_write(mddev);
4218         if (err)
4219                 return err;
4220         while (new > conf->max_nr_stripes) {
4221                 if (grow_one_stripe(conf))
4222                         conf->max_nr_stripes++;
4223                 else break;
4224         }
4225         return len;
4226 }
4227
4228 static struct md_sysfs_entry
4229 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4230                                 raid5_show_stripe_cache_size,
4231                                 raid5_store_stripe_cache_size);
4232
4233 static ssize_t
4234 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4235 {
4236         raid5_conf_t *conf = mddev_to_conf(mddev);
4237         if (conf)
4238                 return sprintf(page, "%d\n", conf->bypass_threshold);
4239         else
4240                 return 0;
4241 }
4242
4243 static ssize_t
4244 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4245 {
4246         raid5_conf_t *conf = mddev_to_conf(mddev);
4247         unsigned long new;
4248         if (len >= PAGE_SIZE)
4249                 return -EINVAL;
4250         if (!conf)
4251                 return -ENODEV;
4252
4253         if (strict_strtoul(page, 10, &new))
4254                 return -EINVAL;
4255         if (new > conf->max_nr_stripes)
4256                 return -EINVAL;
4257         conf->bypass_threshold = new;
4258         return len;
4259 }
4260
4261 static struct md_sysfs_entry
4262 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4263                                         S_IRUGO | S_IWUSR,
4264                                         raid5_show_preread_threshold,
4265                                         raid5_store_preread_threshold);
4266
4267 static ssize_t
4268 stripe_cache_active_show(mddev_t *mddev, char *page)
4269 {
4270         raid5_conf_t *conf = mddev_to_conf(mddev);
4271         if (conf)
4272                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4273         else
4274                 return 0;
4275 }
4276
4277 static struct md_sysfs_entry
4278 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4279
4280 static struct attribute *raid5_attrs[] =  {
4281         &raid5_stripecache_size.attr,
4282         &raid5_stripecache_active.attr,
4283         &raid5_preread_bypass_threshold.attr,
4284         NULL,
4285 };
4286 static struct attribute_group raid5_attrs_group = {
4287         .name = NULL,
4288         .attrs = raid5_attrs,
4289 };
4290
4291 static sector_t
4292 raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks)
4293 {
4294         raid5_conf_t *conf = mddev_to_conf(mddev);
4295
4296         if (!sectors)
4297                 sectors = mddev->dev_sectors;
4298         if (!raid_disks) {
4299                 /* size is defined by the smallest of previous and new size */
4300                 if (conf->raid_disks < conf->previous_raid_disks)
4301                         raid_disks = conf->raid_disks;
4302                 else
4303                         raid_disks = conf->previous_raid_disks;
4304         }
4305
4306         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4307         sectors &= ~((sector_t)mddev->new_chunk/512 - 1);
4308         return sectors * (raid_disks - conf->max_degraded);
4309 }
4310
4311 static raid5_conf_t *setup_conf(mddev_t *mddev)
4312 {
4313         raid5_conf_t *conf;
4314         int raid_disk, memory;
4315         mdk_rdev_t *rdev;
4316         struct disk_info *disk;
4317
4318         if (mddev->new_level != 5
4319             && mddev->new_level != 4
4320             && mddev->new_level != 6) {
4321                 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4322                        mdname(mddev), mddev->new_level);
4323                 return ERR_PTR(-EIO);
4324         }
4325         if ((mddev->new_level == 5
4326              && !algorithm_valid_raid5(mddev->new_layout)) ||
4327             (mddev->new_level == 6
4328              && !algorithm_valid_raid6(mddev->new_layout))) {
4329                 printk(KERN_ERR "raid5: %s: layout %d not supported\n",
4330                        mdname(mddev), mddev->new_layout);
4331                 return ERR_PTR(-EIO);
4332         }
4333         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4334                 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4335                        mdname(mddev), mddev->raid_disks);
4336                 return ERR_PTR(-EINVAL);
4337         }
4338
4339         if (!mddev->new_chunk || mddev->new_chunk % PAGE_SIZE) {
4340                 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4341                         mddev->new_chunk, mdname(mddev));
4342                 return ERR_PTR(-EINVAL);
4343         }
4344
4345         conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL);
4346         if (conf == NULL)
4347                 goto abort;
4348
4349         conf->raid_disks = mddev->raid_disks;
4350         if (mddev->reshape_position == MaxSector)
4351                 conf->previous_raid_disks = mddev->raid_disks;
4352         else
4353                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4354
4355         conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4356                               GFP_KERNEL);
4357         if (!conf->disks)
4358                 goto abort;
4359
4360         conf->mddev = mddev;
4361
4362         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4363                 goto abort;
4364
4365         if (mddev->new_level == 6) {
4366                 conf->spare_page = alloc_page(GFP_KERNEL);
4367                 if (!conf->spare_page)
4368                         goto abort;
4369         }
4370         spin_lock_init(&conf->device_lock);
4371         init_waitqueue_head(&conf->wait_for_stripe);
4372         init_waitqueue_head(&conf->wait_for_overlap);
4373         INIT_LIST_HEAD(&conf->handle_list);
4374         INIT_LIST_HEAD(&conf->hold_list);
4375         INIT_LIST_HEAD(&conf->delayed_list);
4376         INIT_LIST_HEAD(&conf->bitmap_list);
4377         INIT_LIST_HEAD(&conf->inactive_list);
4378         atomic_set(&conf->active_stripes, 0);
4379         atomic_set(&conf->preread_active_stripes, 0);
4380         atomic_set(&conf->active_aligned_reads, 0);
4381         conf->bypass_threshold = BYPASS_THRESHOLD;
4382
4383         pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4384
4385         list_for_each_entry(rdev, &mddev->disks, same_set) {
4386                 raid_disk = rdev->raid_disk;
4387                 if (raid_disk >= conf->raid_disks
4388                     || raid_disk < 0)
4389                         continue;
4390                 disk = conf->disks + raid_disk;
4391
4392                 disk->rdev = rdev;
4393
4394                 if (test_bit(In_sync, &rdev->flags)) {
4395                         char b[BDEVNAME_SIZE];
4396                         printk(KERN_INFO "raid5: device %s operational as raid"
4397                                 " disk %d\n", bdevname(rdev->bdev,b),
4398                                 raid_disk);
4399                 } else
4400                         /* Cannot rely on bitmap to complete recovery */
4401                         conf->fullsync = 1;
4402         }
4403
4404         conf->chunk_size = mddev->new_chunk;
4405         conf->level = mddev->new_level;
4406         if (conf->level == 6)
4407                 conf->max_degraded = 2;
4408         else
4409                 conf->max_degraded = 1;
4410         conf->algorithm = mddev->new_layout;
4411         conf->max_nr_stripes = NR_STRIPES;
4412         conf->reshape_progress = mddev->reshape_position;
4413         if (conf->reshape_progress != MaxSector) {
4414                 conf->prev_chunk = mddev->chunk_size;
4415                 conf->prev_algo = mddev->layout;
4416         }
4417
4418         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4419                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4420         if (grow_stripes(conf, conf->max_nr_stripes)) {
4421                 printk(KERN_ERR
4422                         "raid5: couldn't allocate %dkB for buffers\n", memory);
4423                 goto abort;
4424         } else
4425                 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4426                         memory, mdname(mddev));
4427
4428         conf->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4429         if (!conf->thread) {
4430                 printk(KERN_ERR
4431                        "raid5: couldn't allocate thread for %s\n",
4432                        mdname(mddev));
4433                 goto abort;
4434         }
4435
4436         return conf;
4437
4438  abort:
4439         if (conf) {
4440                 shrink_stripes(conf);
4441                 safe_put_page(conf->spare_page);
4442                 kfree(conf->disks);
4443                 kfree(conf->stripe_hashtbl);
4444                 kfree(conf);
4445                 return ERR_PTR(-EIO);
4446         } else
4447                 return ERR_PTR(-ENOMEM);
4448 }
4449
4450 static int run(mddev_t *mddev)
4451 {
4452         raid5_conf_t *conf;
4453         int working_disks = 0;
4454         mdk_rdev_t *rdev;
4455
4456         if (mddev->reshape_position != MaxSector) {
4457                 /* Check that we can continue the reshape.
4458                  * Currently only disks can change, it must
4459                  * increase, and we must be past the point where
4460                  * a stripe over-writes itself
4461                  */
4462                 sector_t here_new, here_old;
4463                 int old_disks;
4464                 int max_degraded = (mddev->level == 6 ? 2 : 1);
4465
4466                 if (mddev->new_level != mddev->level) {
4467                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4468                                "required - aborting.\n",
4469                                mdname(mddev));
4470                         return -EINVAL;
4471                 }
4472                 old_disks = mddev->raid_disks - mddev->delta_disks;
4473                 /* reshape_position must be on a new-stripe boundary, and one
4474                  * further up in new geometry must map after here in old
4475                  * geometry.
4476                  */
4477                 here_new = mddev->reshape_position;
4478                 if (sector_div(here_new, (mddev->new_chunk>>9)*
4479                                (mddev->raid_disks - max_degraded))) {
4480                         printk(KERN_ERR "raid5: reshape_position not "
4481                                "on a stripe boundary\n");
4482                         return -EINVAL;
4483                 }
4484                 /* here_new is the stripe we will write to */
4485                 here_old = mddev->reshape_position;
4486                 sector_div(here_old, (mddev->chunk_size>>9)*
4487                            (old_disks-max_degraded));
4488                 /* here_old is the first stripe that we might need to read
4489                  * from */
4490                 if (here_new >= here_old) {
4491                         /* Reading from the same stripe as writing to - bad */
4492                         printk(KERN_ERR "raid5: reshape_position too early for "
4493                                "auto-recovery - aborting.\n");
4494                         return -EINVAL;
4495                 }
4496                 printk(KERN_INFO "raid5: reshape will continue\n");
4497                 /* OK, we should be able to continue; */
4498         } else {
4499                 BUG_ON(mddev->level != mddev->new_level);
4500                 BUG_ON(mddev->layout != mddev->new_layout);
4501                 BUG_ON(mddev->chunk_size != mddev->new_chunk);
4502                 BUG_ON(mddev->delta_disks != 0);
4503         }
4504
4505         if (mddev->private == NULL)
4506                 conf = setup_conf(mddev);
4507         else
4508                 conf = mddev->private;
4509
4510         if (IS_ERR(conf))
4511                 return PTR_ERR(conf);
4512
4513         mddev->thread = conf->thread;
4514         conf->thread = NULL;
4515         mddev->private = conf;
4516
4517         /*
4518          * 0 for a fully functional array, 1 or 2 for a degraded array.
4519          */
4520         list_for_each_entry(rdev, &mddev->disks, same_set)
4521                 if (rdev->raid_disk >= 0 &&
4522                     test_bit(In_sync, &rdev->flags))
4523                         working_disks++;
4524
4525         mddev->degraded = conf->raid_disks - working_disks;
4526
4527         if (mddev->degraded > conf->max_degraded) {
4528                 printk(KERN_ERR "raid5: not enough operational devices for %s"
4529                         " (%d/%d failed)\n",
4530                         mdname(mddev), mddev->degraded, conf->raid_disks);
4531                 goto abort;
4532         }
4533
4534         /* device size must be a multiple of chunk size */
4535         mddev->dev_sectors &= ~(mddev->chunk_size / 512 - 1);
4536         mddev->resync_max_sectors = mddev->dev_sectors;
4537
4538         if (mddev->degraded > 0 &&
4539             mddev->recovery_cp != MaxSector) {
4540                 if (mddev->ok_start_degraded)
4541                         printk(KERN_WARNING
4542                                "raid5: starting dirty degraded array: %s"
4543                                "- data corruption possible.\n",
4544                                mdname(mddev));
4545                 else {
4546                         printk(KERN_ERR
4547                                "raid5: cannot start dirty degraded array for %s\n",
4548                                mdname(mddev));
4549                         goto abort;
4550                 }
4551         }
4552
4553         if (mddev->degraded == 0)
4554                 printk("raid5: raid level %d set %s active with %d out of %d"
4555                        " devices, algorithm %d\n", conf->level, mdname(mddev),
4556                        mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4557                        mddev->new_layout);
4558         else
4559                 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4560                         " out of %d devices, algorithm %d\n", conf->level,
4561                         mdname(mddev), mddev->raid_disks - mddev->degraded,
4562                         mddev->raid_disks, mddev->new_layout);
4563
4564         print_raid5_conf(conf);
4565
4566         if (conf->reshape_progress != MaxSector) {
4567                 printk("...ok start reshape thread\n");
4568                 conf->reshape_safe = conf->reshape_progress;
4569                 atomic_set(&conf->reshape_stripes, 0);
4570                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4571                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4572                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4573                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4574                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4575                                                         "%s_reshape");
4576         }
4577
4578         /* read-ahead size must cover two whole stripes, which is
4579          * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4580          */
4581         {
4582                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4583                 int stripe = data_disks *
4584                         (mddev->chunk_size / PAGE_SIZE);
4585                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4586                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4587         }
4588
4589         /* Ok, everything is just fine now */
4590         if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4591                 printk(KERN_WARNING
4592                        "raid5: failed to create sysfs attributes for %s\n",
4593                        mdname(mddev));
4594
4595         mddev->queue->queue_lock = &conf->device_lock;
4596
4597         mddev->queue->unplug_fn = raid5_unplug_device;
4598         mddev->queue->backing_dev_info.congested_data = mddev;
4599         mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4600
4601         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
4602
4603         blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4604
4605         return 0;
4606 abort:
4607         md_unregister_thread(mddev->thread);
4608         mddev->thread = NULL;
4609         if (conf) {
4610                 shrink_stripes(conf);
4611                 print_raid5_conf(conf);
4612                 safe_put_page(conf->spare_page);
4613                 kfree(conf->disks);
4614                 kfree(conf->stripe_hashtbl);
4615                 kfree(conf);
4616         }
4617         mddev->private = NULL;
4618         printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4619         return -EIO;
4620 }
4621
4622
4623
4624 static int stop(mddev_t *mddev)
4625 {
4626         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4627
4628         md_unregister_thread(mddev->thread);
4629         mddev->thread = NULL;
4630         shrink_stripes(conf);
4631         kfree(conf->stripe_hashtbl);
4632         mddev->queue->backing_dev_info.congested_fn = NULL;
4633         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4634         sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4635         kfree(conf->disks);
4636         kfree(conf);
4637         mddev->private = NULL;
4638         return 0;
4639 }
4640
4641 #ifdef DEBUG
4642 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
4643 {
4644         int i;
4645
4646         seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4647                    (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4648         seq_printf(seq, "sh %llu,  count %d.\n",
4649                    (unsigned long long)sh->sector, atomic_read(&sh->count));
4650         seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4651         for (i = 0; i < sh->disks; i++) {
4652                 seq_printf(seq, "(cache%d: %p %ld) ",
4653                            i, sh->dev[i].page, sh->dev[i].flags);
4654         }
4655         seq_printf(seq, "\n");
4656 }
4657
4658 static void printall(struct seq_file *seq, raid5_conf_t *conf)
4659 {
4660         struct stripe_head *sh;
4661         struct hlist_node *hn;
4662         int i;
4663
4664         spin_lock_irq(&conf->device_lock);
4665         for (i = 0; i < NR_HASH; i++) {
4666                 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4667                         if (sh->raid_conf != conf)
4668                                 continue;
4669                         print_sh(seq, sh);
4670                 }
4671         }
4672         spin_unlock_irq(&conf->device_lock);
4673 }
4674 #endif
4675
4676 static void status(struct seq_file *seq, mddev_t *mddev)
4677 {
4678         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4679         int i;
4680
4681         seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4682         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4683         for (i = 0; i < conf->raid_disks; i++)
4684                 seq_printf (seq, "%s",
4685                                conf->disks[i].rdev &&
4686                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4687         seq_printf (seq, "]");
4688 #ifdef DEBUG
4689         seq_printf (seq, "\n");
4690         printall(seq, conf);
4691 #endif
4692 }
4693
4694 static void print_raid5_conf (raid5_conf_t *conf)
4695 {
4696         int i;
4697         struct disk_info *tmp;
4698
4699         printk("RAID5 conf printout:\n");
4700         if (!conf) {
4701                 printk("(conf==NULL)\n");
4702                 return;
4703         }
4704         printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4705                  conf->raid_disks - conf->mddev->degraded);
4706
4707         for (i = 0; i < conf->raid_disks; i++) {
4708                 char b[BDEVNAME_SIZE];
4709                 tmp = conf->disks + i;
4710                 if (tmp->rdev)
4711                 printk(" disk %d, o:%d, dev:%s\n",
4712                         i, !test_bit(Faulty, &tmp->rdev->flags),
4713                         bdevname(tmp->rdev->bdev,b));
4714         }
4715 }
4716
4717 static int raid5_spare_active(mddev_t *mddev)
4718 {
4719         int i;
4720         raid5_conf_t *conf = mddev->private;
4721         struct disk_info *tmp;
4722
4723         for (i = 0; i < conf->raid_disks; i++) {
4724                 tmp = conf->disks + i;
4725                 if (tmp->rdev
4726                     && !test_bit(Faulty, &tmp->rdev->flags)
4727                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4728                         unsigned long flags;
4729                         spin_lock_irqsave(&conf->device_lock, flags);
4730                         mddev->degraded--;
4731                         spin_unlock_irqrestore(&conf->device_lock, flags);
4732                 }
4733         }
4734         print_raid5_conf(conf);
4735         return 0;
4736 }
4737
4738 static int raid5_remove_disk(mddev_t *mddev, int number)
4739 {
4740         raid5_conf_t *conf = mddev->private;
4741         int err = 0;
4742         mdk_rdev_t *rdev;
4743         struct disk_info *p = conf->disks + number;
4744
4745         print_raid5_conf(conf);
4746         rdev = p->rdev;
4747         if (rdev) {
4748                 if (number >= conf->raid_disks &&
4749                     conf->reshape_progress == MaxSector)
4750                         clear_bit(In_sync, &rdev->flags);
4751
4752                 if (test_bit(In_sync, &rdev->flags) ||
4753                     atomic_read(&rdev->nr_pending)) {
4754                         err = -EBUSY;
4755                         goto abort;
4756                 }
4757                 /* Only remove non-faulty devices if recovery
4758                  * isn't possible.
4759                  */
4760                 if (!test_bit(Faulty, &rdev->flags) &&
4761                     mddev->degraded <= conf->max_degraded &&
4762                     number < conf->raid_disks) {
4763                         err = -EBUSY;
4764                         goto abort;
4765                 }
4766                 p->rdev = NULL;
4767                 synchronize_rcu();
4768                 if (atomic_read(&rdev->nr_pending)) {
4769                         /* lost the race, try later */
4770                         err = -EBUSY;
4771                         p->rdev = rdev;
4772                 }
4773         }
4774 abort:
4775
4776         print_raid5_conf(conf);
4777         return err;
4778 }
4779
4780 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4781 {
4782         raid5_conf_t *conf = mddev->private;
4783         int err = -EEXIST;
4784         int disk;
4785         struct disk_info *p;
4786         int first = 0;
4787         int last = conf->raid_disks - 1;
4788
4789         if (mddev->degraded > conf->max_degraded)
4790                 /* no point adding a device */
4791                 return -EINVAL;
4792
4793         if (rdev->raid_disk >= 0)
4794                 first = last = rdev->raid_disk;
4795
4796         /*
4797          * find the disk ... but prefer rdev->saved_raid_disk
4798          * if possible.
4799          */
4800         if (rdev->saved_raid_disk >= 0 &&
4801             rdev->saved_raid_disk >= first &&
4802             conf->disks[rdev->saved_raid_disk].rdev == NULL)
4803                 disk = rdev->saved_raid_disk;
4804         else
4805                 disk = first;
4806         for ( ; disk <= last ; disk++)
4807                 if ((p=conf->disks + disk)->rdev == NULL) {
4808                         clear_bit(In_sync, &rdev->flags);
4809                         rdev->raid_disk = disk;
4810                         err = 0;
4811                         if (rdev->saved_raid_disk != disk)
4812                                 conf->fullsync = 1;
4813                         rcu_assign_pointer(p->rdev, rdev);
4814                         break;
4815                 }
4816         print_raid5_conf(conf);
4817         return err;
4818 }
4819
4820 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4821 {
4822         /* no resync is happening, and there is enough space
4823          * on all devices, so we can resize.
4824          * We need to make sure resync covers any new space.
4825          * If the array is shrinking we should possibly wait until
4826          * any io in the removed space completes, but it hardly seems
4827          * worth it.
4828          */
4829         sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4830         md_set_array_sectors(mddev, raid5_size(mddev, sectors,
4831                                                mddev->raid_disks));
4832         if (mddev->array_sectors >
4833             raid5_size(mddev, sectors, mddev->raid_disks))
4834                 return -EINVAL;
4835         set_capacity(mddev->gendisk, mddev->array_sectors);
4836         mddev->changed = 1;
4837         if (sectors > mddev->dev_sectors && mddev->recovery_cp == MaxSector) {
4838                 mddev->recovery_cp = mddev->dev_sectors;
4839                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4840         }
4841         mddev->dev_sectors = sectors;
4842         mddev->resync_max_sectors = sectors;
4843         return 0;
4844 }
4845
4846 static int raid5_check_reshape(mddev_t *mddev)
4847 {
4848         raid5_conf_t *conf = mddev_to_conf(mddev);
4849
4850         if (mddev->delta_disks == 0 &&
4851             mddev->new_layout == mddev->layout &&
4852             mddev->new_chunk == mddev->chunk_size)
4853                 return -EINVAL; /* nothing to do */
4854         if (mddev->bitmap)
4855                 /* Cannot grow a bitmap yet */
4856                 return -EBUSY;
4857         if (mddev->degraded > conf->max_degraded)
4858                 return -EINVAL;
4859         if (mddev->delta_disks < 0) {
4860                 /* We might be able to shrink, but the devices must
4861                  * be made bigger first.
4862                  * For raid6, 4 is the minimum size.
4863                  * Otherwise 2 is the minimum
4864                  */
4865                 int min = 2;
4866                 if (mddev->level == 6)
4867                         min = 4;
4868                 if (mddev->raid_disks + mddev->delta_disks < min)
4869                         return -EINVAL;
4870         }
4871
4872         /* Can only proceed if there are plenty of stripe_heads.
4873          * We need a minimum of one full stripe,, and for sensible progress
4874          * it is best to have about 4 times that.
4875          * If we require 4 times, then the default 256 4K stripe_heads will
4876          * allow for chunk sizes up to 256K, which is probably OK.
4877          * If the chunk size is greater, user-space should request more
4878          * stripe_heads first.
4879          */
4880         if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
4881             (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
4882                 printk(KERN_WARNING "raid5: reshape: not enough stripes.  Needed %lu\n",
4883                        (max(mddev->chunk_size, mddev->new_chunk)
4884                         / STRIPE_SIZE)*4);
4885                 return -ENOSPC;
4886         }
4887
4888         return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4889 }
4890
4891 static int raid5_start_reshape(mddev_t *mddev)
4892 {
4893         raid5_conf_t *conf = mddev_to_conf(mddev);
4894         mdk_rdev_t *rdev;
4895         int spares = 0;
4896         int added_devices = 0;
4897         unsigned long flags;
4898
4899         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4900                 return -EBUSY;
4901
4902         list_for_each_entry(rdev, &mddev->disks, same_set)
4903                 if (rdev->raid_disk < 0 &&
4904                     !test_bit(Faulty, &rdev->flags))
4905                         spares++;
4906
4907         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4908                 /* Not enough devices even to make a degraded array
4909                  * of that size
4910                  */
4911                 return -EINVAL;
4912
4913         /* Refuse to reduce size of the array.  Any reductions in
4914          * array size must be through explicit setting of array_size
4915          * attribute.
4916          */
4917         if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
4918             < mddev->array_sectors) {
4919                 printk(KERN_ERR "md: %s: array size must be reduced "
4920                        "before number of disks\n", mdname(mddev));
4921                 return -EINVAL;
4922         }
4923
4924         atomic_set(&conf->reshape_stripes, 0);
4925         spin_lock_irq(&conf->device_lock);
4926         conf->previous_raid_disks = conf->raid_disks;
4927         conf->raid_disks += mddev->delta_disks;
4928         conf->prev_chunk = conf->chunk_size;
4929         conf->chunk_size = mddev->new_chunk;
4930         conf->prev_algo = conf->algorithm;
4931         conf->algorithm = mddev->new_layout;
4932         if (mddev->delta_disks < 0)
4933                 conf->reshape_progress = raid5_size(mddev, 0, 0);
4934         else
4935                 conf->reshape_progress = 0;
4936         conf->reshape_safe = conf->reshape_progress;
4937         conf->generation++;
4938         spin_unlock_irq(&conf->device_lock);
4939
4940         /* Add some new drives, as many as will fit.
4941          * We know there are enough to make the newly sized array work.
4942          */
4943         list_for_each_entry(rdev, &mddev->disks, same_set)
4944                 if (rdev->raid_disk < 0 &&
4945                     !test_bit(Faulty, &rdev->flags)) {
4946                         if (raid5_add_disk(mddev, rdev) == 0) {
4947                                 char nm[20];
4948                                 set_bit(In_sync, &rdev->flags);
4949                                 added_devices++;
4950                                 rdev->recovery_offset = 0;
4951                                 sprintf(nm, "rd%d", rdev->raid_disk);
4952                                 if (sysfs_create_link(&mddev->kobj,
4953                                                       &rdev->kobj, nm))
4954                                         printk(KERN_WARNING
4955                                                "raid5: failed to create "
4956                                                " link %s for %s\n",
4957                                                nm, mdname(mddev));
4958                         } else
4959                                 break;
4960                 }
4961
4962         if (mddev->delta_disks > 0) {
4963                 spin_lock_irqsave(&conf->device_lock, flags);
4964                 mddev->degraded = (conf->raid_disks - conf->previous_raid_disks)
4965                         - added_devices;
4966                 spin_unlock_irqrestore(&conf->device_lock, flags);
4967         }
4968         mddev->raid_disks = conf->raid_disks;
4969         mddev->reshape_position = 0;
4970         set_bit(MD_CHANGE_DEVS, &mddev->flags);
4971
4972         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4973         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4974         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4975         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4976         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4977                                                 "%s_reshape");
4978         if (!mddev->sync_thread) {
4979                 mddev->recovery = 0;
4980                 spin_lock_irq(&conf->device_lock);
4981                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
4982                 conf->reshape_progress = MaxSector;
4983                 spin_unlock_irq(&conf->device_lock);
4984                 return -EAGAIN;
4985         }
4986         conf->reshape_checkpoint = jiffies;
4987         md_wakeup_thread(mddev->sync_thread);
4988         md_new_event(mddev);
4989         return 0;
4990 }
4991
4992 /* This is called from the reshape thread and should make any
4993  * changes needed in 'conf'
4994  */
4995 static void end_reshape(raid5_conf_t *conf)
4996 {
4997
4998         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
4999
5000                 spin_lock_irq(&conf->device_lock);
5001                 conf->previous_raid_disks = conf->raid_disks;
5002                 conf->reshape_progress = MaxSector;
5003                 spin_unlock_irq(&conf->device_lock);
5004                 wake_up(&conf->wait_for_overlap);
5005
5006                 /* read-ahead size must cover two whole stripes, which is
5007                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5008                  */
5009                 {
5010                         int data_disks = conf->raid_disks - conf->max_degraded;
5011                         int stripe = data_disks * (conf->chunk_size
5012                                                    / PAGE_SIZE);
5013                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5014                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5015                 }
5016         }
5017 }
5018
5019 /* This is called from the raid5d thread with mddev_lock held.
5020  * It makes config changes to the device.
5021  */
5022 static void raid5_finish_reshape(mddev_t *mddev)
5023 {
5024         struct block_device *bdev;
5025         raid5_conf_t *conf = mddev_to_conf(mddev);
5026
5027         if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5028
5029                 if (mddev->delta_disks > 0) {
5030                         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5031                         set_capacity(mddev->gendisk, mddev->array_sectors);
5032                         mddev->changed = 1;
5033
5034                         bdev = bdget_disk(mddev->gendisk, 0);
5035                         if (bdev) {
5036                                 mutex_lock(&bdev->bd_inode->i_mutex);
5037                                 i_size_write(bdev->bd_inode,
5038                                              (loff_t)mddev->array_sectors << 9);
5039                                 mutex_unlock(&bdev->bd_inode->i_mutex);
5040                                 bdput(bdev);
5041                         }
5042                 } else {
5043                         int d;
5044                         mddev->degraded = conf->raid_disks;
5045                         for (d = 0; d < conf->raid_disks ; d++)
5046                                 if (conf->disks[d].rdev &&
5047                                     test_bit(In_sync,
5048                                              &conf->disks[d].rdev->flags))
5049                                         mddev->degraded--;
5050                         for (d = conf->raid_disks ;
5051                              d < conf->raid_disks - mddev->delta_disks;
5052                              d++)
5053                                 raid5_remove_disk(mddev, d);
5054                 }
5055                 mddev->layout = conf->algorithm;
5056                 mddev->chunk_size = conf->chunk_size;
5057                 mddev->reshape_position = MaxSector;
5058                 mddev->delta_disks = 0;
5059         }
5060 }
5061
5062 static void raid5_quiesce(mddev_t *mddev, int state)
5063 {
5064         raid5_conf_t *conf = mddev_to_conf(mddev);
5065
5066         switch(state) {
5067         case 2: /* resume for a suspend */
5068                 wake_up(&conf->wait_for_overlap);
5069                 break;
5070
5071         case 1: /* stop all writes */
5072                 spin_lock_irq(&conf->device_lock);
5073                 conf->quiesce = 1;
5074                 wait_event_lock_irq(conf->wait_for_stripe,
5075                                     atomic_read(&conf->active_stripes) == 0 &&
5076                                     atomic_read(&conf->active_aligned_reads) == 0,
5077                                     conf->device_lock, /* nothing */);
5078                 spin_unlock_irq(&conf->device_lock);
5079                 break;
5080
5081         case 0: /* re-enable writes */
5082                 spin_lock_irq(&conf->device_lock);
5083                 conf->quiesce = 0;
5084                 wake_up(&conf->wait_for_stripe);
5085                 wake_up(&conf->wait_for_overlap);
5086                 spin_unlock_irq(&conf->device_lock);
5087                 break;
5088         }
5089 }
5090
5091
5092 static void *raid5_takeover_raid1(mddev_t *mddev)
5093 {
5094         int chunksect;
5095
5096         if (mddev->raid_disks != 2 ||
5097             mddev->degraded > 1)
5098                 return ERR_PTR(-EINVAL);
5099
5100         /* Should check if there are write-behind devices? */
5101
5102         chunksect = 64*2; /* 64K by default */
5103
5104         /* The array must be an exact multiple of chunksize */
5105         while (chunksect && (mddev->array_sectors & (chunksect-1)))
5106                 chunksect >>= 1;
5107
5108         if ((chunksect<<9) < STRIPE_SIZE)
5109                 /* array size does not allow a suitable chunk size */
5110                 return ERR_PTR(-EINVAL);
5111
5112         mddev->new_level = 5;
5113         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5114         mddev->new_chunk = chunksect << 9;
5115
5116         return setup_conf(mddev);
5117 }
5118
5119 static void *raid5_takeover_raid6(mddev_t *mddev)
5120 {
5121         int new_layout;
5122
5123         switch (mddev->layout) {
5124         case ALGORITHM_LEFT_ASYMMETRIC_6:
5125                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5126                 break;
5127         case ALGORITHM_RIGHT_ASYMMETRIC_6:
5128                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5129                 break;
5130         case ALGORITHM_LEFT_SYMMETRIC_6:
5131                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5132                 break;
5133         case ALGORITHM_RIGHT_SYMMETRIC_6:
5134                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5135                 break;
5136         case ALGORITHM_PARITY_0_6:
5137                 new_layout = ALGORITHM_PARITY_0;
5138                 break;
5139         case ALGORITHM_PARITY_N:
5140                 new_layout = ALGORITHM_PARITY_N;
5141                 break;
5142         default:
5143                 return ERR_PTR(-EINVAL);
5144         }
5145         mddev->new_level = 5;
5146         mddev->new_layout = new_layout;
5147         mddev->delta_disks = -1;
5148         mddev->raid_disks -= 1;
5149         return setup_conf(mddev);
5150 }
5151
5152
5153 static int raid5_reconfig(mddev_t *mddev, int new_layout, int new_chunk)
5154 {
5155         /* For a 2-drive array, the layout and chunk size can be changed
5156          * immediately as not restriping is needed.
5157          * For larger arrays we record the new value - after validation
5158          * to be used by a reshape pass.
5159          */
5160         raid5_conf_t *conf = mddev_to_conf(mddev);
5161
5162         if (new_layout >= 0 && !algorithm_valid_raid5(new_layout))
5163                 return -EINVAL;
5164         if (new_chunk > 0) {
5165                 if (new_chunk & (new_chunk-1))
5166                         /* not a power of 2 */
5167                         return -EINVAL;
5168                 if (new_chunk < PAGE_SIZE)
5169                         return -EINVAL;
5170                 if (mddev->array_sectors & ((new_chunk>>9)-1))
5171                         /* not factor of array size */
5172                         return -EINVAL;
5173         }
5174
5175         /* They look valid */
5176
5177         if (mddev->raid_disks == 2) {
5178
5179                 if (new_layout >= 0) {
5180                         conf->algorithm = new_layout;
5181                         mddev->layout = mddev->new_layout = new_layout;
5182                 }
5183                 if (new_chunk > 0) {
5184                         conf->chunk_size = new_chunk;
5185                         mddev->chunk_size = mddev->new_chunk = new_chunk;
5186                 }
5187                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5188                 md_wakeup_thread(mddev->thread);
5189         } else {
5190                 if (new_layout >= 0)
5191                         mddev->new_layout = new_layout;
5192                 if (new_chunk > 0)
5193                         mddev->new_chunk = new_chunk;
5194         }
5195         return 0;
5196 }
5197
5198 static int raid6_reconfig(mddev_t *mddev, int new_layout, int new_chunk)
5199 {
5200         if (new_layout >= 0 && !algorithm_valid_raid6(new_layout))
5201                 return -EINVAL;
5202         if (new_chunk > 0) {
5203                 if (new_chunk & (new_chunk-1))
5204                         /* not a power of 2 */
5205                         return -EINVAL;
5206                 if (new_chunk < PAGE_SIZE)
5207                         return -EINVAL;
5208                 if (mddev->array_sectors & ((new_chunk>>9)-1))
5209                         /* not factor of array size */
5210                         return -EINVAL;
5211         }
5212
5213         /* They look valid */
5214
5215         if (new_layout >= 0)
5216                 mddev->new_layout = new_layout;
5217         if (new_chunk > 0)
5218                 mddev->new_chunk = new_chunk;
5219
5220         return 0;
5221 }
5222
5223 static void *raid5_takeover(mddev_t *mddev)
5224 {
5225         /* raid5 can take over:
5226          *  raid0 - if all devices are the same - make it a raid4 layout
5227          *  raid1 - if there are two drives.  We need to know the chunk size
5228          *  raid4 - trivial - just use a raid4 layout.
5229          *  raid6 - Providing it is a *_6 layout
5230          *
5231          * For now, just do raid1
5232          */
5233
5234         if (mddev->level == 1)
5235                 return raid5_takeover_raid1(mddev);
5236         if (mddev->level == 4) {
5237                 mddev->new_layout = ALGORITHM_PARITY_N;
5238                 mddev->new_level = 5;
5239                 return setup_conf(mddev);
5240         }
5241         if (mddev->level == 6)
5242                 return raid5_takeover_raid6(mddev);
5243
5244         return ERR_PTR(-EINVAL);
5245 }
5246
5247
5248 static struct mdk_personality raid5_personality;
5249
5250 static void *raid6_takeover(mddev_t *mddev)
5251 {
5252         /* Currently can only take over a raid5.  We map the
5253          * personality to an equivalent raid6 personality
5254          * with the Q block at the end.
5255          */
5256         int new_layout;
5257
5258         if (mddev->pers != &raid5_personality)
5259                 return ERR_PTR(-EINVAL);
5260         if (mddev->degraded > 1)
5261                 return ERR_PTR(-EINVAL);
5262         if (mddev->raid_disks > 253)
5263                 return ERR_PTR(-EINVAL);
5264         if (mddev->raid_disks < 3)
5265                 return ERR_PTR(-EINVAL);
5266
5267         switch (mddev->layout) {
5268         case ALGORITHM_LEFT_ASYMMETRIC:
5269                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5270                 break;
5271         case ALGORITHM_RIGHT_ASYMMETRIC:
5272                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5273                 break;
5274         case ALGORITHM_LEFT_SYMMETRIC:
5275                 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5276                 break;
5277         case ALGORITHM_RIGHT_SYMMETRIC:
5278                 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5279                 break;
5280         case ALGORITHM_PARITY_0:
5281                 new_layout = ALGORITHM_PARITY_0_6;
5282                 break;
5283         case ALGORITHM_PARITY_N:
5284                 new_layout = ALGORITHM_PARITY_N;
5285                 break;
5286         default:
5287                 return ERR_PTR(-EINVAL);
5288         }
5289         mddev->new_level = 6;
5290         mddev->new_layout = new_layout;
5291         mddev->delta_disks = 1;
5292         mddev->raid_disks += 1;
5293         return setup_conf(mddev);
5294 }
5295
5296
5297 static struct mdk_personality raid6_personality =
5298 {
5299         .name           = "raid6",
5300         .level          = 6,
5301         .owner          = THIS_MODULE,
5302         .make_request   = make_request,
5303         .run            = run,
5304         .stop           = stop,
5305         .status         = status,
5306         .error_handler  = error,
5307         .hot_add_disk   = raid5_add_disk,
5308         .hot_remove_disk= raid5_remove_disk,
5309         .spare_active   = raid5_spare_active,
5310         .sync_request   = sync_request,
5311         .resize         = raid5_resize,
5312         .size           = raid5_size,
5313         .check_reshape  = raid5_check_reshape,
5314         .start_reshape  = raid5_start_reshape,
5315         .finish_reshape = raid5_finish_reshape,
5316         .quiesce        = raid5_quiesce,
5317         .takeover       = raid6_takeover,
5318         .reconfig       = raid6_reconfig,
5319 };
5320 static struct mdk_personality raid5_personality =
5321 {
5322         .name           = "raid5",
5323         .level          = 5,
5324         .owner          = THIS_MODULE,
5325         .make_request   = make_request,
5326         .run            = run,
5327         .stop           = stop,
5328         .status         = status,
5329         .error_handler  = error,
5330         .hot_add_disk   = raid5_add_disk,
5331         .hot_remove_disk= raid5_remove_disk,
5332         .spare_active   = raid5_spare_active,
5333         .sync_request   = sync_request,
5334         .resize         = raid5_resize,
5335         .size           = raid5_size,
5336         .check_reshape  = raid5_check_reshape,
5337         .start_reshape  = raid5_start_reshape,
5338         .finish_reshape = raid5_finish_reshape,
5339         .quiesce        = raid5_quiesce,
5340         .takeover       = raid5_takeover,
5341         .reconfig       = raid5_reconfig,
5342 };
5343
5344 static struct mdk_personality raid4_personality =
5345 {
5346         .name           = "raid4",
5347         .level          = 4,
5348         .owner          = THIS_MODULE,
5349         .make_request   = make_request,
5350         .run            = run,
5351         .stop           = stop,
5352         .status         = status,
5353         .error_handler  = error,
5354         .hot_add_disk   = raid5_add_disk,
5355         .hot_remove_disk= raid5_remove_disk,
5356         .spare_active   = raid5_spare_active,
5357         .sync_request   = sync_request,
5358         .resize         = raid5_resize,
5359         .size           = raid5_size,
5360         .check_reshape  = raid5_check_reshape,
5361         .start_reshape  = raid5_start_reshape,
5362         .finish_reshape = raid5_finish_reshape,
5363         .quiesce        = raid5_quiesce,
5364 };
5365
5366 static int __init raid5_init(void)
5367 {
5368         register_md_personality(&raid6_personality);
5369         register_md_personality(&raid5_personality);
5370         register_md_personality(&raid4_personality);
5371         return 0;
5372 }
5373
5374 static void raid5_exit(void)
5375 {
5376         unregister_md_personality(&raid6_personality);
5377         unregister_md_personality(&raid5_personality);
5378         unregister_md_personality(&raid4_personality);
5379 }
5380
5381 module_init(raid5_init);
5382 module_exit(raid5_exit);
5383 MODULE_LICENSE("GPL");
5384 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5385 MODULE_ALIAS("md-raid5");
5386 MODULE_ALIAS("md-raid4");
5387 MODULE_ALIAS("md-level-5");
5388 MODULE_ALIAS("md-level-4");
5389 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5390 MODULE_ALIAS("md-raid6");
5391 MODULE_ALIAS("md-level-6");
5392
5393 /* This used to be two separate modules, they were: */
5394 MODULE_ALIAS("raid5");
5395 MODULE_ALIAS("raid6");