Merge branch 'next' of git://git.kernel.org/pub/scm/linux/kernel/git/djbw/async_tx...
[linux-2.6.git] / drivers / md / raid5.c
1 /*
2  * raid5.c : Multiple Devices driver for Linux
3  *         Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4  *         Copyright (C) 1999, 2000 Ingo Molnar
5  *         Copyright (C) 2002, 2003 H. Peter Anvin
6  *
7  * RAID-4/5/6 management functions.
8  * Thanks to Penguin Computing for making the RAID-6 development possible
9  * by donating a test server!
10  *
11  * This program is free software; you can redistribute it and/or modify
12  * it under the terms of the GNU General Public License as published by
13  * the Free Software Foundation; either version 2, or (at your option)
14  * any later version.
15  *
16  * You should have received a copy of the GNU General Public License
17  * (for example /usr/src/linux/COPYING); if not, write to the Free
18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19  */
20
21 /*
22  * BITMAP UNPLUGGING:
23  *
24  * The sequencing for updating the bitmap reliably is a little
25  * subtle (and I got it wrong the first time) so it deserves some
26  * explanation.
27  *
28  * We group bitmap updates into batches.  Each batch has a number.
29  * We may write out several batches at once, but that isn't very important.
30  * conf->bm_write is the number of the last batch successfully written.
31  * conf->bm_flush is the number of the last batch that was closed to
32  *    new additions.
33  * When we discover that we will need to write to any block in a stripe
34  * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35  * the number of the batch it will be in. This is bm_flush+1.
36  * When we are ready to do a write, if that batch hasn't been written yet,
37  *   we plug the array and queue the stripe for later.
38  * When an unplug happens, we increment bm_flush, thus closing the current
39  *   batch.
40  * When we notice that bm_flush > bm_write, we write out all pending updates
41  * to the bitmap, and advance bm_write to where bm_flush was.
42  * This may occasionally write a bit out twice, but is sure never to
43  * miss any bits.
44  */
45
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/async.h>
51 #include <linux/seq_file.h>
52 #include <linux/cpu.h>
53 #include "md.h"
54 #include "raid5.h"
55 #include "bitmap.h"
56
57 /*
58  * Stripe cache
59  */
60
61 #define NR_STRIPES              256
62 #define STRIPE_SIZE             PAGE_SIZE
63 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
64 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
65 #define IO_THRESHOLD            1
66 #define BYPASS_THRESHOLD        1
67 #define NR_HASH                 (PAGE_SIZE / sizeof(struct hlist_head))
68 #define HASH_MASK               (NR_HASH - 1)
69
70 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
71
72 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
73  * order without overlap.  There may be several bio's per stripe+device, and
74  * a bio could span several devices.
75  * When walking this list for a particular stripe+device, we must never proceed
76  * beyond a bio that extends past this device, as the next bio might no longer
77  * be valid.
78  * This macro is used to determine the 'next' bio in the list, given the sector
79  * of the current stripe+device
80  */
81 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
82 /*
83  * The following can be used to debug the driver
84  */
85 #define RAID5_PARANOIA  1
86 #if RAID5_PARANOIA && defined(CONFIG_SMP)
87 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
88 #else
89 # define CHECK_DEVLOCK()
90 #endif
91
92 #ifdef DEBUG
93 #define inline
94 #define __inline__
95 #endif
96
97 #define printk_rl(args...) ((void) (printk_ratelimit() && printk(args)))
98
99 /*
100  * We maintain a biased count of active stripes in the bottom 16 bits of
101  * bi_phys_segments, and a count of processed stripes in the upper 16 bits
102  */
103 static inline int raid5_bi_phys_segments(struct bio *bio)
104 {
105         return bio->bi_phys_segments & 0xffff;
106 }
107
108 static inline int raid5_bi_hw_segments(struct bio *bio)
109 {
110         return (bio->bi_phys_segments >> 16) & 0xffff;
111 }
112
113 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
114 {
115         --bio->bi_phys_segments;
116         return raid5_bi_phys_segments(bio);
117 }
118
119 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
120 {
121         unsigned short val = raid5_bi_hw_segments(bio);
122
123         --val;
124         bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
125         return val;
126 }
127
128 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
129 {
130         bio->bi_phys_segments = raid5_bi_phys_segments(bio) || (cnt << 16);
131 }
132
133 /* Find first data disk in a raid6 stripe */
134 static inline int raid6_d0(struct stripe_head *sh)
135 {
136         if (sh->ddf_layout)
137                 /* ddf always start from first device */
138                 return 0;
139         /* md starts just after Q block */
140         if (sh->qd_idx == sh->disks - 1)
141                 return 0;
142         else
143                 return sh->qd_idx + 1;
144 }
145 static inline int raid6_next_disk(int disk, int raid_disks)
146 {
147         disk++;
148         return (disk < raid_disks) ? disk : 0;
149 }
150
151 /* When walking through the disks in a raid5, starting at raid6_d0,
152  * We need to map each disk to a 'slot', where the data disks are slot
153  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
154  * is raid_disks-1.  This help does that mapping.
155  */
156 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
157                              int *count, int syndrome_disks)
158 {
159         int slot;
160
161         if (idx == sh->pd_idx)
162                 return syndrome_disks;
163         if (idx == sh->qd_idx)
164                 return syndrome_disks + 1;
165         slot = (*count)++;
166         return slot;
167 }
168
169 static void return_io(struct bio *return_bi)
170 {
171         struct bio *bi = return_bi;
172         while (bi) {
173
174                 return_bi = bi->bi_next;
175                 bi->bi_next = NULL;
176                 bi->bi_size = 0;
177                 bio_endio(bi, 0);
178                 bi = return_bi;
179         }
180 }
181
182 static void print_raid5_conf (raid5_conf_t *conf);
183
184 static int stripe_operations_active(struct stripe_head *sh)
185 {
186         return sh->check_state || sh->reconstruct_state ||
187                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
188                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
189 }
190
191 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
192 {
193         if (atomic_dec_and_test(&sh->count)) {
194                 BUG_ON(!list_empty(&sh->lru));
195                 BUG_ON(atomic_read(&conf->active_stripes)==0);
196                 if (test_bit(STRIPE_HANDLE, &sh->state)) {
197                         if (test_bit(STRIPE_DELAYED, &sh->state)) {
198                                 list_add_tail(&sh->lru, &conf->delayed_list);
199                                 blk_plug_device(conf->mddev->queue);
200                         } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
201                                    sh->bm_seq - conf->seq_write > 0) {
202                                 list_add_tail(&sh->lru, &conf->bitmap_list);
203                                 blk_plug_device(conf->mddev->queue);
204                         } else {
205                                 clear_bit(STRIPE_BIT_DELAY, &sh->state);
206                                 list_add_tail(&sh->lru, &conf->handle_list);
207                         }
208                         md_wakeup_thread(conf->mddev->thread);
209                 } else {
210                         BUG_ON(stripe_operations_active(sh));
211                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
212                                 atomic_dec(&conf->preread_active_stripes);
213                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
214                                         md_wakeup_thread(conf->mddev->thread);
215                         }
216                         atomic_dec(&conf->active_stripes);
217                         if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
218                                 list_add_tail(&sh->lru, &conf->inactive_list);
219                                 wake_up(&conf->wait_for_stripe);
220                                 if (conf->retry_read_aligned)
221                                         md_wakeup_thread(conf->mddev->thread);
222                         }
223                 }
224         }
225 }
226
227 static void release_stripe(struct stripe_head *sh)
228 {
229         raid5_conf_t *conf = sh->raid_conf;
230         unsigned long flags;
231
232         spin_lock_irqsave(&conf->device_lock, flags);
233         __release_stripe(conf, sh);
234         spin_unlock_irqrestore(&conf->device_lock, flags);
235 }
236
237 static inline void remove_hash(struct stripe_head *sh)
238 {
239         pr_debug("remove_hash(), stripe %llu\n",
240                 (unsigned long long)sh->sector);
241
242         hlist_del_init(&sh->hash);
243 }
244
245 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
246 {
247         struct hlist_head *hp = stripe_hash(conf, sh->sector);
248
249         pr_debug("insert_hash(), stripe %llu\n",
250                 (unsigned long long)sh->sector);
251
252         CHECK_DEVLOCK();
253         hlist_add_head(&sh->hash, hp);
254 }
255
256
257 /* find an idle stripe, make sure it is unhashed, and return it. */
258 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
259 {
260         struct stripe_head *sh = NULL;
261         struct list_head *first;
262
263         CHECK_DEVLOCK();
264         if (list_empty(&conf->inactive_list))
265                 goto out;
266         first = conf->inactive_list.next;
267         sh = list_entry(first, struct stripe_head, lru);
268         list_del_init(first);
269         remove_hash(sh);
270         atomic_inc(&conf->active_stripes);
271 out:
272         return sh;
273 }
274
275 static void shrink_buffers(struct stripe_head *sh, int num)
276 {
277         struct page *p;
278         int i;
279
280         for (i=0; i<num ; i++) {
281                 p = sh->dev[i].page;
282                 if (!p)
283                         continue;
284                 sh->dev[i].page = NULL;
285                 put_page(p);
286         }
287 }
288
289 static int grow_buffers(struct stripe_head *sh, int num)
290 {
291         int i;
292
293         for (i=0; i<num; i++) {
294                 struct page *page;
295
296                 if (!(page = alloc_page(GFP_KERNEL))) {
297                         return 1;
298                 }
299                 sh->dev[i].page = page;
300         }
301         return 0;
302 }
303
304 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
305 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
306                             struct stripe_head *sh);
307
308 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
309 {
310         raid5_conf_t *conf = sh->raid_conf;
311         int i;
312
313         BUG_ON(atomic_read(&sh->count) != 0);
314         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
315         BUG_ON(stripe_operations_active(sh));
316
317         CHECK_DEVLOCK();
318         pr_debug("init_stripe called, stripe %llu\n",
319                 (unsigned long long)sh->sector);
320
321         remove_hash(sh);
322
323         sh->generation = conf->generation - previous;
324         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
325         sh->sector = sector;
326         stripe_set_idx(sector, conf, previous, sh);
327         sh->state = 0;
328
329
330         for (i = sh->disks; i--; ) {
331                 struct r5dev *dev = &sh->dev[i];
332
333                 if (dev->toread || dev->read || dev->towrite || dev->written ||
334                     test_bit(R5_LOCKED, &dev->flags)) {
335                         printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
336                                (unsigned long long)sh->sector, i, dev->toread,
337                                dev->read, dev->towrite, dev->written,
338                                test_bit(R5_LOCKED, &dev->flags));
339                         BUG();
340                 }
341                 dev->flags = 0;
342                 raid5_build_block(sh, i, previous);
343         }
344         insert_hash(conf, sh);
345 }
346
347 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector,
348                                          short generation)
349 {
350         struct stripe_head *sh;
351         struct hlist_node *hn;
352
353         CHECK_DEVLOCK();
354         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
355         hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
356                 if (sh->sector == sector && sh->generation == generation)
357                         return sh;
358         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
359         return NULL;
360 }
361
362 static void unplug_slaves(mddev_t *mddev);
363 static void raid5_unplug_device(struct request_queue *q);
364
365 static struct stripe_head *
366 get_active_stripe(raid5_conf_t *conf, sector_t sector,
367                   int previous, int noblock, int noquiesce)
368 {
369         struct stripe_head *sh;
370
371         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
372
373         spin_lock_irq(&conf->device_lock);
374
375         do {
376                 wait_event_lock_irq(conf->wait_for_stripe,
377                                     conf->quiesce == 0 || noquiesce,
378                                     conf->device_lock, /* nothing */);
379                 sh = __find_stripe(conf, sector, conf->generation - previous);
380                 if (!sh) {
381                         if (!conf->inactive_blocked)
382                                 sh = get_free_stripe(conf);
383                         if (noblock && sh == NULL)
384                                 break;
385                         if (!sh) {
386                                 conf->inactive_blocked = 1;
387                                 wait_event_lock_irq(conf->wait_for_stripe,
388                                                     !list_empty(&conf->inactive_list) &&
389                                                     (atomic_read(&conf->active_stripes)
390                                                      < (conf->max_nr_stripes *3/4)
391                                                      || !conf->inactive_blocked),
392                                                     conf->device_lock,
393                                                     raid5_unplug_device(conf->mddev->queue)
394                                         );
395                                 conf->inactive_blocked = 0;
396                         } else
397                                 init_stripe(sh, sector, previous);
398                 } else {
399                         if (atomic_read(&sh->count)) {
400                                 BUG_ON(!list_empty(&sh->lru)
401                                     && !test_bit(STRIPE_EXPANDING, &sh->state));
402                         } else {
403                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
404                                         atomic_inc(&conf->active_stripes);
405                                 if (list_empty(&sh->lru) &&
406                                     !test_bit(STRIPE_EXPANDING, &sh->state))
407                                         BUG();
408                                 list_del_init(&sh->lru);
409                         }
410                 }
411         } while (sh == NULL);
412
413         if (sh)
414                 atomic_inc(&sh->count);
415
416         spin_unlock_irq(&conf->device_lock);
417         return sh;
418 }
419
420 static void
421 raid5_end_read_request(struct bio *bi, int error);
422 static void
423 raid5_end_write_request(struct bio *bi, int error);
424
425 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
426 {
427         raid5_conf_t *conf = sh->raid_conf;
428         int i, disks = sh->disks;
429
430         might_sleep();
431
432         for (i = disks; i--; ) {
433                 int rw;
434                 struct bio *bi;
435                 mdk_rdev_t *rdev;
436                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
437                         rw = WRITE;
438                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
439                         rw = READ;
440                 else
441                         continue;
442
443                 bi = &sh->dev[i].req;
444
445                 bi->bi_rw = rw;
446                 if (rw == WRITE)
447                         bi->bi_end_io = raid5_end_write_request;
448                 else
449                         bi->bi_end_io = raid5_end_read_request;
450
451                 rcu_read_lock();
452                 rdev = rcu_dereference(conf->disks[i].rdev);
453                 if (rdev && test_bit(Faulty, &rdev->flags))
454                         rdev = NULL;
455                 if (rdev)
456                         atomic_inc(&rdev->nr_pending);
457                 rcu_read_unlock();
458
459                 if (rdev) {
460                         if (s->syncing || s->expanding || s->expanded)
461                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
462
463                         set_bit(STRIPE_IO_STARTED, &sh->state);
464
465                         bi->bi_bdev = rdev->bdev;
466                         pr_debug("%s: for %llu schedule op %ld on disc %d\n",
467                                 __func__, (unsigned long long)sh->sector,
468                                 bi->bi_rw, i);
469                         atomic_inc(&sh->count);
470                         bi->bi_sector = sh->sector + rdev->data_offset;
471                         bi->bi_flags = 1 << BIO_UPTODATE;
472                         bi->bi_vcnt = 1;
473                         bi->bi_max_vecs = 1;
474                         bi->bi_idx = 0;
475                         bi->bi_io_vec = &sh->dev[i].vec;
476                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
477                         bi->bi_io_vec[0].bv_offset = 0;
478                         bi->bi_size = STRIPE_SIZE;
479                         bi->bi_next = NULL;
480                         if (rw == WRITE &&
481                             test_bit(R5_ReWrite, &sh->dev[i].flags))
482                                 atomic_add(STRIPE_SECTORS,
483                                         &rdev->corrected_errors);
484                         generic_make_request(bi);
485                 } else {
486                         if (rw == WRITE)
487                                 set_bit(STRIPE_DEGRADED, &sh->state);
488                         pr_debug("skip op %ld on disc %d for sector %llu\n",
489                                 bi->bi_rw, i, (unsigned long long)sh->sector);
490                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
491                         set_bit(STRIPE_HANDLE, &sh->state);
492                 }
493         }
494 }
495
496 static struct dma_async_tx_descriptor *
497 async_copy_data(int frombio, struct bio *bio, struct page *page,
498         sector_t sector, struct dma_async_tx_descriptor *tx)
499 {
500         struct bio_vec *bvl;
501         struct page *bio_page;
502         int i;
503         int page_offset;
504         struct async_submit_ctl submit;
505         enum async_tx_flags flags = 0;
506
507         if (bio->bi_sector >= sector)
508                 page_offset = (signed)(bio->bi_sector - sector) * 512;
509         else
510                 page_offset = (signed)(sector - bio->bi_sector) * -512;
511
512         if (frombio)
513                 flags |= ASYNC_TX_FENCE;
514         init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
515
516         bio_for_each_segment(bvl, bio, i) {
517                 int len = bio_iovec_idx(bio, i)->bv_len;
518                 int clen;
519                 int b_offset = 0;
520
521                 if (page_offset < 0) {
522                         b_offset = -page_offset;
523                         page_offset += b_offset;
524                         len -= b_offset;
525                 }
526
527                 if (len > 0 && page_offset + len > STRIPE_SIZE)
528                         clen = STRIPE_SIZE - page_offset;
529                 else
530                         clen = len;
531
532                 if (clen > 0) {
533                         b_offset += bio_iovec_idx(bio, i)->bv_offset;
534                         bio_page = bio_iovec_idx(bio, i)->bv_page;
535                         if (frombio)
536                                 tx = async_memcpy(page, bio_page, page_offset,
537                                                   b_offset, clen, &submit);
538                         else
539                                 tx = async_memcpy(bio_page, page, b_offset,
540                                                   page_offset, clen, &submit);
541                 }
542                 /* chain the operations */
543                 submit.depend_tx = tx;
544
545                 if (clen < len) /* hit end of page */
546                         break;
547                 page_offset +=  len;
548         }
549
550         return tx;
551 }
552
553 static void ops_complete_biofill(void *stripe_head_ref)
554 {
555         struct stripe_head *sh = stripe_head_ref;
556         struct bio *return_bi = NULL;
557         raid5_conf_t *conf = sh->raid_conf;
558         int i;
559
560         pr_debug("%s: stripe %llu\n", __func__,
561                 (unsigned long long)sh->sector);
562
563         /* clear completed biofills */
564         spin_lock_irq(&conf->device_lock);
565         for (i = sh->disks; i--; ) {
566                 struct r5dev *dev = &sh->dev[i];
567
568                 /* acknowledge completion of a biofill operation */
569                 /* and check if we need to reply to a read request,
570                  * new R5_Wantfill requests are held off until
571                  * !STRIPE_BIOFILL_RUN
572                  */
573                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
574                         struct bio *rbi, *rbi2;
575
576                         BUG_ON(!dev->read);
577                         rbi = dev->read;
578                         dev->read = NULL;
579                         while (rbi && rbi->bi_sector <
580                                 dev->sector + STRIPE_SECTORS) {
581                                 rbi2 = r5_next_bio(rbi, dev->sector);
582                                 if (!raid5_dec_bi_phys_segments(rbi)) {
583                                         rbi->bi_next = return_bi;
584                                         return_bi = rbi;
585                                 }
586                                 rbi = rbi2;
587                         }
588                 }
589         }
590         spin_unlock_irq(&conf->device_lock);
591         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
592
593         return_io(return_bi);
594
595         set_bit(STRIPE_HANDLE, &sh->state);
596         release_stripe(sh);
597 }
598
599 static void ops_run_biofill(struct stripe_head *sh)
600 {
601         struct dma_async_tx_descriptor *tx = NULL;
602         raid5_conf_t *conf = sh->raid_conf;
603         struct async_submit_ctl submit;
604         int i;
605
606         pr_debug("%s: stripe %llu\n", __func__,
607                 (unsigned long long)sh->sector);
608
609         for (i = sh->disks; i--; ) {
610                 struct r5dev *dev = &sh->dev[i];
611                 if (test_bit(R5_Wantfill, &dev->flags)) {
612                         struct bio *rbi;
613                         spin_lock_irq(&conf->device_lock);
614                         dev->read = rbi = dev->toread;
615                         dev->toread = NULL;
616                         spin_unlock_irq(&conf->device_lock);
617                         while (rbi && rbi->bi_sector <
618                                 dev->sector + STRIPE_SECTORS) {
619                                 tx = async_copy_data(0, rbi, dev->page,
620                                         dev->sector, tx);
621                                 rbi = r5_next_bio(rbi, dev->sector);
622                         }
623                 }
624         }
625
626         atomic_inc(&sh->count);
627         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
628         async_trigger_callback(&submit);
629 }
630
631 static void mark_target_uptodate(struct stripe_head *sh, int target)
632 {
633         struct r5dev *tgt;
634
635         if (target < 0)
636                 return;
637
638         tgt = &sh->dev[target];
639         set_bit(R5_UPTODATE, &tgt->flags);
640         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
641         clear_bit(R5_Wantcompute, &tgt->flags);
642 }
643
644 static void ops_complete_compute(void *stripe_head_ref)
645 {
646         struct stripe_head *sh = stripe_head_ref;
647
648         pr_debug("%s: stripe %llu\n", __func__,
649                 (unsigned long long)sh->sector);
650
651         /* mark the computed target(s) as uptodate */
652         mark_target_uptodate(sh, sh->ops.target);
653         mark_target_uptodate(sh, sh->ops.target2);
654
655         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
656         if (sh->check_state == check_state_compute_run)
657                 sh->check_state = check_state_compute_result;
658         set_bit(STRIPE_HANDLE, &sh->state);
659         release_stripe(sh);
660 }
661
662 /* return a pointer to the address conversion region of the scribble buffer */
663 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
664                                  struct raid5_percpu *percpu)
665 {
666         return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
667 }
668
669 static struct dma_async_tx_descriptor *
670 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
671 {
672         int disks = sh->disks;
673         struct page **xor_srcs = percpu->scribble;
674         int target = sh->ops.target;
675         struct r5dev *tgt = &sh->dev[target];
676         struct page *xor_dest = tgt->page;
677         int count = 0;
678         struct dma_async_tx_descriptor *tx;
679         struct async_submit_ctl submit;
680         int i;
681
682         pr_debug("%s: stripe %llu block: %d\n",
683                 __func__, (unsigned long long)sh->sector, target);
684         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
685
686         for (i = disks; i--; )
687                 if (i != target)
688                         xor_srcs[count++] = sh->dev[i].page;
689
690         atomic_inc(&sh->count);
691
692         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
693                           ops_complete_compute, sh, to_addr_conv(sh, percpu));
694         if (unlikely(count == 1))
695                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
696         else
697                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
698
699         return tx;
700 }
701
702 /* set_syndrome_sources - populate source buffers for gen_syndrome
703  * @srcs - (struct page *) array of size sh->disks
704  * @sh - stripe_head to parse
705  *
706  * Populates srcs in proper layout order for the stripe and returns the
707  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
708  * destination buffer is recorded in srcs[count] and the Q destination
709  * is recorded in srcs[count+1]].
710  */
711 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
712 {
713         int disks = sh->disks;
714         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
715         int d0_idx = raid6_d0(sh);
716         int count;
717         int i;
718
719         for (i = 0; i < disks; i++)
720                 srcs[i] = (void *)raid6_empty_zero_page;
721
722         count = 0;
723         i = d0_idx;
724         do {
725                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
726
727                 srcs[slot] = sh->dev[i].page;
728                 i = raid6_next_disk(i, disks);
729         } while (i != d0_idx);
730         BUG_ON(count != syndrome_disks);
731
732         return count;
733 }
734
735 static struct dma_async_tx_descriptor *
736 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
737 {
738         int disks = sh->disks;
739         struct page **blocks = percpu->scribble;
740         int target;
741         int qd_idx = sh->qd_idx;
742         struct dma_async_tx_descriptor *tx;
743         struct async_submit_ctl submit;
744         struct r5dev *tgt;
745         struct page *dest;
746         int i;
747         int count;
748
749         if (sh->ops.target < 0)
750                 target = sh->ops.target2;
751         else if (sh->ops.target2 < 0)
752                 target = sh->ops.target;
753         else
754                 /* we should only have one valid target */
755                 BUG();
756         BUG_ON(target < 0);
757         pr_debug("%s: stripe %llu block: %d\n",
758                 __func__, (unsigned long long)sh->sector, target);
759
760         tgt = &sh->dev[target];
761         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
762         dest = tgt->page;
763
764         atomic_inc(&sh->count);
765
766         if (target == qd_idx) {
767                 count = set_syndrome_sources(blocks, sh);
768                 blocks[count] = NULL; /* regenerating p is not necessary */
769                 BUG_ON(blocks[count+1] != dest); /* q should already be set */
770                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
771                                   ops_complete_compute, sh,
772                                   to_addr_conv(sh, percpu));
773                 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
774         } else {
775                 /* Compute any data- or p-drive using XOR */
776                 count = 0;
777                 for (i = disks; i-- ; ) {
778                         if (i == target || i == qd_idx)
779                                 continue;
780                         blocks[count++] = sh->dev[i].page;
781                 }
782
783                 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
784                                   NULL, ops_complete_compute, sh,
785                                   to_addr_conv(sh, percpu));
786                 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
787         }
788
789         return tx;
790 }
791
792 static struct dma_async_tx_descriptor *
793 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
794 {
795         int i, count, disks = sh->disks;
796         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
797         int d0_idx = raid6_d0(sh);
798         int faila = -1, failb = -1;
799         int target = sh->ops.target;
800         int target2 = sh->ops.target2;
801         struct r5dev *tgt = &sh->dev[target];
802         struct r5dev *tgt2 = &sh->dev[target2];
803         struct dma_async_tx_descriptor *tx;
804         struct page **blocks = percpu->scribble;
805         struct async_submit_ctl submit;
806
807         pr_debug("%s: stripe %llu block1: %d block2: %d\n",
808                  __func__, (unsigned long long)sh->sector, target, target2);
809         BUG_ON(target < 0 || target2 < 0);
810         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
811         BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
812
813         /* we need to open-code set_syndrome_sources to handle the
814          * slot number conversion for 'faila' and 'failb'
815          */
816         for (i = 0; i < disks ; i++)
817                 blocks[i] = (void *)raid6_empty_zero_page;
818         count = 0;
819         i = d0_idx;
820         do {
821                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
822
823                 blocks[slot] = sh->dev[i].page;
824
825                 if (i == target)
826                         faila = slot;
827                 if (i == target2)
828                         failb = slot;
829                 i = raid6_next_disk(i, disks);
830         } while (i != d0_idx);
831         BUG_ON(count != syndrome_disks);
832
833         BUG_ON(faila == failb);
834         if (failb < faila)
835                 swap(faila, failb);
836         pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
837                  __func__, (unsigned long long)sh->sector, faila, failb);
838
839         atomic_inc(&sh->count);
840
841         if (failb == syndrome_disks+1) {
842                 /* Q disk is one of the missing disks */
843                 if (faila == syndrome_disks) {
844                         /* Missing P+Q, just recompute */
845                         init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
846                                           ops_complete_compute, sh,
847                                           to_addr_conv(sh, percpu));
848                         return async_gen_syndrome(blocks, 0, count+2,
849                                                   STRIPE_SIZE, &submit);
850                 } else {
851                         struct page *dest;
852                         int data_target;
853                         int qd_idx = sh->qd_idx;
854
855                         /* Missing D+Q: recompute D from P, then recompute Q */
856                         if (target == qd_idx)
857                                 data_target = target2;
858                         else
859                                 data_target = target;
860
861                         count = 0;
862                         for (i = disks; i-- ; ) {
863                                 if (i == data_target || i == qd_idx)
864                                         continue;
865                                 blocks[count++] = sh->dev[i].page;
866                         }
867                         dest = sh->dev[data_target].page;
868                         init_async_submit(&submit,
869                                           ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
870                                           NULL, NULL, NULL,
871                                           to_addr_conv(sh, percpu));
872                         tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
873                                        &submit);
874
875                         count = set_syndrome_sources(blocks, sh);
876                         init_async_submit(&submit, ASYNC_TX_FENCE, tx,
877                                           ops_complete_compute, sh,
878                                           to_addr_conv(sh, percpu));
879                         return async_gen_syndrome(blocks, 0, count+2,
880                                                   STRIPE_SIZE, &submit);
881                 }
882         } else {
883                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
884                                   ops_complete_compute, sh,
885                                   to_addr_conv(sh, percpu));
886                 if (failb == syndrome_disks) {
887                         /* We're missing D+P. */
888                         return async_raid6_datap_recov(syndrome_disks+2,
889                                                        STRIPE_SIZE, faila,
890                                                        blocks, &submit);
891                 } else {
892                         /* We're missing D+D. */
893                         return async_raid6_2data_recov(syndrome_disks+2,
894                                                        STRIPE_SIZE, faila, failb,
895                                                        blocks, &submit);
896                 }
897         }
898 }
899
900
901 static void ops_complete_prexor(void *stripe_head_ref)
902 {
903         struct stripe_head *sh = stripe_head_ref;
904
905         pr_debug("%s: stripe %llu\n", __func__,
906                 (unsigned long long)sh->sector);
907 }
908
909 static struct dma_async_tx_descriptor *
910 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
911                struct dma_async_tx_descriptor *tx)
912 {
913         int disks = sh->disks;
914         struct page **xor_srcs = percpu->scribble;
915         int count = 0, pd_idx = sh->pd_idx, i;
916         struct async_submit_ctl submit;
917
918         /* existing parity data subtracted */
919         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
920
921         pr_debug("%s: stripe %llu\n", __func__,
922                 (unsigned long long)sh->sector);
923
924         for (i = disks; i--; ) {
925                 struct r5dev *dev = &sh->dev[i];
926                 /* Only process blocks that are known to be uptodate */
927                 if (test_bit(R5_Wantdrain, &dev->flags))
928                         xor_srcs[count++] = dev->page;
929         }
930
931         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
932                           ops_complete_prexor, sh, to_addr_conv(sh, percpu));
933         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
934
935         return tx;
936 }
937
938 static struct dma_async_tx_descriptor *
939 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
940 {
941         int disks = sh->disks;
942         int i;
943
944         pr_debug("%s: stripe %llu\n", __func__,
945                 (unsigned long long)sh->sector);
946
947         for (i = disks; i--; ) {
948                 struct r5dev *dev = &sh->dev[i];
949                 struct bio *chosen;
950
951                 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
952                         struct bio *wbi;
953
954                         spin_lock(&sh->lock);
955                         chosen = dev->towrite;
956                         dev->towrite = NULL;
957                         BUG_ON(dev->written);
958                         wbi = dev->written = chosen;
959                         spin_unlock(&sh->lock);
960
961                         while (wbi && wbi->bi_sector <
962                                 dev->sector + STRIPE_SECTORS) {
963                                 tx = async_copy_data(1, wbi, dev->page,
964                                         dev->sector, tx);
965                                 wbi = r5_next_bio(wbi, dev->sector);
966                         }
967                 }
968         }
969
970         return tx;
971 }
972
973 static void ops_complete_reconstruct(void *stripe_head_ref)
974 {
975         struct stripe_head *sh = stripe_head_ref;
976         int disks = sh->disks;
977         int pd_idx = sh->pd_idx;
978         int qd_idx = sh->qd_idx;
979         int i;
980
981         pr_debug("%s: stripe %llu\n", __func__,
982                 (unsigned long long)sh->sector);
983
984         for (i = disks; i--; ) {
985                 struct r5dev *dev = &sh->dev[i];
986
987                 if (dev->written || i == pd_idx || i == qd_idx)
988                         set_bit(R5_UPTODATE, &dev->flags);
989         }
990
991         if (sh->reconstruct_state == reconstruct_state_drain_run)
992                 sh->reconstruct_state = reconstruct_state_drain_result;
993         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
994                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
995         else {
996                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
997                 sh->reconstruct_state = reconstruct_state_result;
998         }
999
1000         set_bit(STRIPE_HANDLE, &sh->state);
1001         release_stripe(sh);
1002 }
1003
1004 static void
1005 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1006                      struct dma_async_tx_descriptor *tx)
1007 {
1008         int disks = sh->disks;
1009         struct page **xor_srcs = percpu->scribble;
1010         struct async_submit_ctl submit;
1011         int count = 0, pd_idx = sh->pd_idx, i;
1012         struct page *xor_dest;
1013         int prexor = 0;
1014         unsigned long flags;
1015
1016         pr_debug("%s: stripe %llu\n", __func__,
1017                 (unsigned long long)sh->sector);
1018
1019         /* check if prexor is active which means only process blocks
1020          * that are part of a read-modify-write (written)
1021          */
1022         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1023                 prexor = 1;
1024                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1025                 for (i = disks; i--; ) {
1026                         struct r5dev *dev = &sh->dev[i];
1027                         if (dev->written)
1028                                 xor_srcs[count++] = dev->page;
1029                 }
1030         } else {
1031                 xor_dest = sh->dev[pd_idx].page;
1032                 for (i = disks; i--; ) {
1033                         struct r5dev *dev = &sh->dev[i];
1034                         if (i != pd_idx)
1035                                 xor_srcs[count++] = dev->page;
1036                 }
1037         }
1038
1039         /* 1/ if we prexor'd then the dest is reused as a source
1040          * 2/ if we did not prexor then we are redoing the parity
1041          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1042          * for the synchronous xor case
1043          */
1044         flags = ASYNC_TX_ACK |
1045                 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1046
1047         atomic_inc(&sh->count);
1048
1049         init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1050                           to_addr_conv(sh, percpu));
1051         if (unlikely(count == 1))
1052                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1053         else
1054                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1055 }
1056
1057 static void
1058 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1059                      struct dma_async_tx_descriptor *tx)
1060 {
1061         struct async_submit_ctl submit;
1062         struct page **blocks = percpu->scribble;
1063         int count;
1064
1065         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1066
1067         count = set_syndrome_sources(blocks, sh);
1068
1069         atomic_inc(&sh->count);
1070
1071         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1072                           sh, to_addr_conv(sh, percpu));
1073         async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1074 }
1075
1076 static void ops_complete_check(void *stripe_head_ref)
1077 {
1078         struct stripe_head *sh = stripe_head_ref;
1079
1080         pr_debug("%s: stripe %llu\n", __func__,
1081                 (unsigned long long)sh->sector);
1082
1083         sh->check_state = check_state_check_result;
1084         set_bit(STRIPE_HANDLE, &sh->state);
1085         release_stripe(sh);
1086 }
1087
1088 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1089 {
1090         int disks = sh->disks;
1091         int pd_idx = sh->pd_idx;
1092         int qd_idx = sh->qd_idx;
1093         struct page *xor_dest;
1094         struct page **xor_srcs = percpu->scribble;
1095         struct dma_async_tx_descriptor *tx;
1096         struct async_submit_ctl submit;
1097         int count;
1098         int i;
1099
1100         pr_debug("%s: stripe %llu\n", __func__,
1101                 (unsigned long long)sh->sector);
1102
1103         count = 0;
1104         xor_dest = sh->dev[pd_idx].page;
1105         xor_srcs[count++] = xor_dest;
1106         for (i = disks; i--; ) {
1107                 if (i == pd_idx || i == qd_idx)
1108                         continue;
1109                 xor_srcs[count++] = sh->dev[i].page;
1110         }
1111
1112         init_async_submit(&submit, 0, NULL, NULL, NULL,
1113                           to_addr_conv(sh, percpu));
1114         tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1115                            &sh->ops.zero_sum_result, &submit);
1116
1117         atomic_inc(&sh->count);
1118         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1119         tx = async_trigger_callback(&submit);
1120 }
1121
1122 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1123 {
1124         struct page **srcs = percpu->scribble;
1125         struct async_submit_ctl submit;
1126         int count;
1127
1128         pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1129                 (unsigned long long)sh->sector, checkp);
1130
1131         count = set_syndrome_sources(srcs, sh);
1132         if (!checkp)
1133                 srcs[count] = NULL;
1134
1135         atomic_inc(&sh->count);
1136         init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1137                           sh, to_addr_conv(sh, percpu));
1138         async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1139                            &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1140 }
1141
1142 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1143 {
1144         int overlap_clear = 0, i, disks = sh->disks;
1145         struct dma_async_tx_descriptor *tx = NULL;
1146         raid5_conf_t *conf = sh->raid_conf;
1147         int level = conf->level;
1148         struct raid5_percpu *percpu;
1149         unsigned long cpu;
1150
1151         cpu = get_cpu();
1152         percpu = per_cpu_ptr(conf->percpu, cpu);
1153         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1154                 ops_run_biofill(sh);
1155                 overlap_clear++;
1156         }
1157
1158         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1159                 if (level < 6)
1160                         tx = ops_run_compute5(sh, percpu);
1161                 else {
1162                         if (sh->ops.target2 < 0 || sh->ops.target < 0)
1163                                 tx = ops_run_compute6_1(sh, percpu);
1164                         else
1165                                 tx = ops_run_compute6_2(sh, percpu);
1166                 }
1167                 /* terminate the chain if reconstruct is not set to be run */
1168                 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1169                         async_tx_ack(tx);
1170         }
1171
1172         if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1173                 tx = ops_run_prexor(sh, percpu, tx);
1174
1175         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1176                 tx = ops_run_biodrain(sh, tx);
1177                 overlap_clear++;
1178         }
1179
1180         if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1181                 if (level < 6)
1182                         ops_run_reconstruct5(sh, percpu, tx);
1183                 else
1184                         ops_run_reconstruct6(sh, percpu, tx);
1185         }
1186
1187         if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1188                 if (sh->check_state == check_state_run)
1189                         ops_run_check_p(sh, percpu);
1190                 else if (sh->check_state == check_state_run_q)
1191                         ops_run_check_pq(sh, percpu, 0);
1192                 else if (sh->check_state == check_state_run_pq)
1193                         ops_run_check_pq(sh, percpu, 1);
1194                 else
1195                         BUG();
1196         }
1197
1198         if (overlap_clear)
1199                 for (i = disks; i--; ) {
1200                         struct r5dev *dev = &sh->dev[i];
1201                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
1202                                 wake_up(&sh->raid_conf->wait_for_overlap);
1203                 }
1204         put_cpu();
1205 }
1206
1207 static int grow_one_stripe(raid5_conf_t *conf)
1208 {
1209         struct stripe_head *sh;
1210         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
1211         if (!sh)
1212                 return 0;
1213         memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
1214         sh->raid_conf = conf;
1215         spin_lock_init(&sh->lock);
1216
1217         if (grow_buffers(sh, conf->raid_disks)) {
1218                 shrink_buffers(sh, conf->raid_disks);
1219                 kmem_cache_free(conf->slab_cache, sh);
1220                 return 0;
1221         }
1222         sh->disks = conf->raid_disks;
1223         /* we just created an active stripe so... */
1224         atomic_set(&sh->count, 1);
1225         atomic_inc(&conf->active_stripes);
1226         INIT_LIST_HEAD(&sh->lru);
1227         release_stripe(sh);
1228         return 1;
1229 }
1230
1231 static int grow_stripes(raid5_conf_t *conf, int num)
1232 {
1233         struct kmem_cache *sc;
1234         int devs = conf->raid_disks;
1235
1236         sprintf(conf->cache_name[0],
1237                 "raid%d-%s", conf->level, mdname(conf->mddev));
1238         sprintf(conf->cache_name[1],
1239                 "raid%d-%s-alt", conf->level, mdname(conf->mddev));
1240         conf->active_name = 0;
1241         sc = kmem_cache_create(conf->cache_name[conf->active_name],
1242                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1243                                0, 0, NULL);
1244         if (!sc)
1245                 return 1;
1246         conf->slab_cache = sc;
1247         conf->pool_size = devs;
1248         while (num--)
1249                 if (!grow_one_stripe(conf))
1250                         return 1;
1251         return 0;
1252 }
1253
1254 /**
1255  * scribble_len - return the required size of the scribble region
1256  * @num - total number of disks in the array
1257  *
1258  * The size must be enough to contain:
1259  * 1/ a struct page pointer for each device in the array +2
1260  * 2/ room to convert each entry in (1) to its corresponding dma
1261  *    (dma_map_page()) or page (page_address()) address.
1262  *
1263  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1264  * calculate over all devices (not just the data blocks), using zeros in place
1265  * of the P and Q blocks.
1266  */
1267 static size_t scribble_len(int num)
1268 {
1269         size_t len;
1270
1271         len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1272
1273         return len;
1274 }
1275
1276 static int resize_stripes(raid5_conf_t *conf, int newsize)
1277 {
1278         /* Make all the stripes able to hold 'newsize' devices.
1279          * New slots in each stripe get 'page' set to a new page.
1280          *
1281          * This happens in stages:
1282          * 1/ create a new kmem_cache and allocate the required number of
1283          *    stripe_heads.
1284          * 2/ gather all the old stripe_heads and tranfer the pages across
1285          *    to the new stripe_heads.  This will have the side effect of
1286          *    freezing the array as once all stripe_heads have been collected,
1287          *    no IO will be possible.  Old stripe heads are freed once their
1288          *    pages have been transferred over, and the old kmem_cache is
1289          *    freed when all stripes are done.
1290          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
1291          *    we simple return a failre status - no need to clean anything up.
1292          * 4/ allocate new pages for the new slots in the new stripe_heads.
1293          *    If this fails, we don't bother trying the shrink the
1294          *    stripe_heads down again, we just leave them as they are.
1295          *    As each stripe_head is processed the new one is released into
1296          *    active service.
1297          *
1298          * Once step2 is started, we cannot afford to wait for a write,
1299          * so we use GFP_NOIO allocations.
1300          */
1301         struct stripe_head *osh, *nsh;
1302         LIST_HEAD(newstripes);
1303         struct disk_info *ndisks;
1304         unsigned long cpu;
1305         int err;
1306         struct kmem_cache *sc;
1307         int i;
1308
1309         if (newsize <= conf->pool_size)
1310                 return 0; /* never bother to shrink */
1311
1312         err = md_allow_write(conf->mddev);
1313         if (err)
1314                 return err;
1315
1316         /* Step 1 */
1317         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1318                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1319                                0, 0, NULL);
1320         if (!sc)
1321                 return -ENOMEM;
1322
1323         for (i = conf->max_nr_stripes; i; i--) {
1324                 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1325                 if (!nsh)
1326                         break;
1327
1328                 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1329
1330                 nsh->raid_conf = conf;
1331                 spin_lock_init(&nsh->lock);
1332
1333                 list_add(&nsh->lru, &newstripes);
1334         }
1335         if (i) {
1336                 /* didn't get enough, give up */
1337                 while (!list_empty(&newstripes)) {
1338                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
1339                         list_del(&nsh->lru);
1340                         kmem_cache_free(sc, nsh);
1341                 }
1342                 kmem_cache_destroy(sc);
1343                 return -ENOMEM;
1344         }
1345         /* Step 2 - Must use GFP_NOIO now.
1346          * OK, we have enough stripes, start collecting inactive
1347          * stripes and copying them over
1348          */
1349         list_for_each_entry(nsh, &newstripes, lru) {
1350                 spin_lock_irq(&conf->device_lock);
1351                 wait_event_lock_irq(conf->wait_for_stripe,
1352                                     !list_empty(&conf->inactive_list),
1353                                     conf->device_lock,
1354                                     unplug_slaves(conf->mddev)
1355                         );
1356                 osh = get_free_stripe(conf);
1357                 spin_unlock_irq(&conf->device_lock);
1358                 atomic_set(&nsh->count, 1);
1359                 for(i=0; i<conf->pool_size; i++)
1360                         nsh->dev[i].page = osh->dev[i].page;
1361                 for( ; i<newsize; i++)
1362                         nsh->dev[i].page = NULL;
1363                 kmem_cache_free(conf->slab_cache, osh);
1364         }
1365         kmem_cache_destroy(conf->slab_cache);
1366
1367         /* Step 3.
1368          * At this point, we are holding all the stripes so the array
1369          * is completely stalled, so now is a good time to resize
1370          * conf->disks and the scribble region
1371          */
1372         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1373         if (ndisks) {
1374                 for (i=0; i<conf->raid_disks; i++)
1375                         ndisks[i] = conf->disks[i];
1376                 kfree(conf->disks);
1377                 conf->disks = ndisks;
1378         } else
1379                 err = -ENOMEM;
1380
1381         get_online_cpus();
1382         conf->scribble_len = scribble_len(newsize);
1383         for_each_present_cpu(cpu) {
1384                 struct raid5_percpu *percpu;
1385                 void *scribble;
1386
1387                 percpu = per_cpu_ptr(conf->percpu, cpu);
1388                 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1389
1390                 if (scribble) {
1391                         kfree(percpu->scribble);
1392                         percpu->scribble = scribble;
1393                 } else {
1394                         err = -ENOMEM;
1395                         break;
1396                 }
1397         }
1398         put_online_cpus();
1399
1400         /* Step 4, return new stripes to service */
1401         while(!list_empty(&newstripes)) {
1402                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1403                 list_del_init(&nsh->lru);
1404
1405                 for (i=conf->raid_disks; i < newsize; i++)
1406                         if (nsh->dev[i].page == NULL) {
1407                                 struct page *p = alloc_page(GFP_NOIO);
1408                                 nsh->dev[i].page = p;
1409                                 if (!p)
1410                                         err = -ENOMEM;
1411                         }
1412                 release_stripe(nsh);
1413         }
1414         /* critical section pass, GFP_NOIO no longer needed */
1415
1416         conf->slab_cache = sc;
1417         conf->active_name = 1-conf->active_name;
1418         conf->pool_size = newsize;
1419         return err;
1420 }
1421
1422 static int drop_one_stripe(raid5_conf_t *conf)
1423 {
1424         struct stripe_head *sh;
1425
1426         spin_lock_irq(&conf->device_lock);
1427         sh = get_free_stripe(conf);
1428         spin_unlock_irq(&conf->device_lock);
1429         if (!sh)
1430                 return 0;
1431         BUG_ON(atomic_read(&sh->count));
1432         shrink_buffers(sh, conf->pool_size);
1433         kmem_cache_free(conf->slab_cache, sh);
1434         atomic_dec(&conf->active_stripes);
1435         return 1;
1436 }
1437
1438 static void shrink_stripes(raid5_conf_t *conf)
1439 {
1440         while (drop_one_stripe(conf))
1441                 ;
1442
1443         if (conf->slab_cache)
1444                 kmem_cache_destroy(conf->slab_cache);
1445         conf->slab_cache = NULL;
1446 }
1447
1448 static void raid5_end_read_request(struct bio * bi, int error)
1449 {
1450         struct stripe_head *sh = bi->bi_private;
1451         raid5_conf_t *conf = sh->raid_conf;
1452         int disks = sh->disks, i;
1453         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1454         char b[BDEVNAME_SIZE];
1455         mdk_rdev_t *rdev;
1456
1457
1458         for (i=0 ; i<disks; i++)
1459                 if (bi == &sh->dev[i].req)
1460                         break;
1461
1462         pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1463                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1464                 uptodate);
1465         if (i == disks) {
1466                 BUG();
1467                 return;
1468         }
1469
1470         if (uptodate) {
1471                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1472                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1473                         rdev = conf->disks[i].rdev;
1474                         printk_rl(KERN_INFO "raid5:%s: read error corrected"
1475                                   " (%lu sectors at %llu on %s)\n",
1476                                   mdname(conf->mddev), STRIPE_SECTORS,
1477                                   (unsigned long long)(sh->sector
1478                                                        + rdev->data_offset),
1479                                   bdevname(rdev->bdev, b));
1480                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1481                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1482                 }
1483                 if (atomic_read(&conf->disks[i].rdev->read_errors))
1484                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
1485         } else {
1486                 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1487                 int retry = 0;
1488                 rdev = conf->disks[i].rdev;
1489
1490                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1491                 atomic_inc(&rdev->read_errors);
1492                 if (conf->mddev->degraded)
1493                         printk_rl(KERN_WARNING
1494                                   "raid5:%s: read error not correctable "
1495                                   "(sector %llu on %s).\n",
1496                                   mdname(conf->mddev),
1497                                   (unsigned long long)(sh->sector
1498                                                        + rdev->data_offset),
1499                                   bdn);
1500                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1501                         /* Oh, no!!! */
1502                         printk_rl(KERN_WARNING
1503                                   "raid5:%s: read error NOT corrected!! "
1504                                   "(sector %llu on %s).\n",
1505                                   mdname(conf->mddev),
1506                                   (unsigned long long)(sh->sector
1507                                                        + rdev->data_offset),
1508                                   bdn);
1509                 else if (atomic_read(&rdev->read_errors)
1510                          > conf->max_nr_stripes)
1511                         printk(KERN_WARNING
1512                                "raid5:%s: Too many read errors, failing device %s.\n",
1513                                mdname(conf->mddev), bdn);
1514                 else
1515                         retry = 1;
1516                 if (retry)
1517                         set_bit(R5_ReadError, &sh->dev[i].flags);
1518                 else {
1519                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1520                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1521                         md_error(conf->mddev, rdev);
1522                 }
1523         }
1524         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1525         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1526         set_bit(STRIPE_HANDLE, &sh->state);
1527         release_stripe(sh);
1528 }
1529
1530 static void raid5_end_write_request(struct bio *bi, int error)
1531 {
1532         struct stripe_head *sh = bi->bi_private;
1533         raid5_conf_t *conf = sh->raid_conf;
1534         int disks = sh->disks, i;
1535         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1536
1537         for (i=0 ; i<disks; i++)
1538                 if (bi == &sh->dev[i].req)
1539                         break;
1540
1541         pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1542                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1543                 uptodate);
1544         if (i == disks) {
1545                 BUG();
1546                 return;
1547         }
1548
1549         if (!uptodate)
1550                 md_error(conf->mddev, conf->disks[i].rdev);
1551
1552         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1553         
1554         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1555         set_bit(STRIPE_HANDLE, &sh->state);
1556         release_stripe(sh);
1557 }
1558
1559
1560 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1561         
1562 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1563 {
1564         struct r5dev *dev = &sh->dev[i];
1565
1566         bio_init(&dev->req);
1567         dev->req.bi_io_vec = &dev->vec;
1568         dev->req.bi_vcnt++;
1569         dev->req.bi_max_vecs++;
1570         dev->vec.bv_page = dev->page;
1571         dev->vec.bv_len = STRIPE_SIZE;
1572         dev->vec.bv_offset = 0;
1573
1574         dev->req.bi_sector = sh->sector;
1575         dev->req.bi_private = sh;
1576
1577         dev->flags = 0;
1578         dev->sector = compute_blocknr(sh, i, previous);
1579 }
1580
1581 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1582 {
1583         char b[BDEVNAME_SIZE];
1584         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1585         pr_debug("raid5: error called\n");
1586
1587         if (!test_bit(Faulty, &rdev->flags)) {
1588                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1589                 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1590                         unsigned long flags;
1591                         spin_lock_irqsave(&conf->device_lock, flags);
1592                         mddev->degraded++;
1593                         spin_unlock_irqrestore(&conf->device_lock, flags);
1594                         /*
1595                          * if recovery was running, make sure it aborts.
1596                          */
1597                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1598                 }
1599                 set_bit(Faulty, &rdev->flags);
1600                 printk(KERN_ALERT
1601                        "raid5: Disk failure on %s, disabling device.\n"
1602                        "raid5: Operation continuing on %d devices.\n",
1603                        bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1604         }
1605 }
1606
1607 /*
1608  * Input: a 'big' sector number,
1609  * Output: index of the data and parity disk, and the sector # in them.
1610  */
1611 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1612                                      int previous, int *dd_idx,
1613                                      struct stripe_head *sh)
1614 {
1615         long stripe;
1616         unsigned long chunk_number;
1617         unsigned int chunk_offset;
1618         int pd_idx, qd_idx;
1619         int ddf_layout = 0;
1620         sector_t new_sector;
1621         int algorithm = previous ? conf->prev_algo
1622                                  : conf->algorithm;
1623         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1624                                          : conf->chunk_sectors;
1625         int raid_disks = previous ? conf->previous_raid_disks
1626                                   : conf->raid_disks;
1627         int data_disks = raid_disks - conf->max_degraded;
1628
1629         /* First compute the information on this sector */
1630
1631         /*
1632          * Compute the chunk number and the sector offset inside the chunk
1633          */
1634         chunk_offset = sector_div(r_sector, sectors_per_chunk);
1635         chunk_number = r_sector;
1636         BUG_ON(r_sector != chunk_number);
1637
1638         /*
1639          * Compute the stripe number
1640          */
1641         stripe = chunk_number / data_disks;
1642
1643         /*
1644          * Compute the data disk and parity disk indexes inside the stripe
1645          */
1646         *dd_idx = chunk_number % data_disks;
1647
1648         /*
1649          * Select the parity disk based on the user selected algorithm.
1650          */
1651         pd_idx = qd_idx = ~0;
1652         switch(conf->level) {
1653         case 4:
1654                 pd_idx = data_disks;
1655                 break;
1656         case 5:
1657                 switch (algorithm) {
1658                 case ALGORITHM_LEFT_ASYMMETRIC:
1659                         pd_idx = data_disks - stripe % raid_disks;
1660                         if (*dd_idx >= pd_idx)
1661                                 (*dd_idx)++;
1662                         break;
1663                 case ALGORITHM_RIGHT_ASYMMETRIC:
1664                         pd_idx = stripe % raid_disks;
1665                         if (*dd_idx >= pd_idx)
1666                                 (*dd_idx)++;
1667                         break;
1668                 case ALGORITHM_LEFT_SYMMETRIC:
1669                         pd_idx = data_disks - stripe % raid_disks;
1670                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1671                         break;
1672                 case ALGORITHM_RIGHT_SYMMETRIC:
1673                         pd_idx = stripe % raid_disks;
1674                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1675                         break;
1676                 case ALGORITHM_PARITY_0:
1677                         pd_idx = 0;
1678                         (*dd_idx)++;
1679                         break;
1680                 case ALGORITHM_PARITY_N:
1681                         pd_idx = data_disks;
1682                         break;
1683                 default:
1684                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1685                                 algorithm);
1686                         BUG();
1687                 }
1688                 break;
1689         case 6:
1690
1691                 switch (algorithm) {
1692                 case ALGORITHM_LEFT_ASYMMETRIC:
1693                         pd_idx = raid_disks - 1 - (stripe % raid_disks);
1694                         qd_idx = pd_idx + 1;
1695                         if (pd_idx == raid_disks-1) {
1696                                 (*dd_idx)++;    /* Q D D D P */
1697                                 qd_idx = 0;
1698                         } else if (*dd_idx >= pd_idx)
1699                                 (*dd_idx) += 2; /* D D P Q D */
1700                         break;
1701                 case ALGORITHM_RIGHT_ASYMMETRIC:
1702                         pd_idx = stripe % raid_disks;
1703                         qd_idx = pd_idx + 1;
1704                         if (pd_idx == raid_disks-1) {
1705                                 (*dd_idx)++;    /* Q D D D P */
1706                                 qd_idx = 0;
1707                         } else if (*dd_idx >= pd_idx)
1708                                 (*dd_idx) += 2; /* D D P Q D */
1709                         break;
1710                 case ALGORITHM_LEFT_SYMMETRIC:
1711                         pd_idx = raid_disks - 1 - (stripe % raid_disks);
1712                         qd_idx = (pd_idx + 1) % raid_disks;
1713                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1714                         break;
1715                 case ALGORITHM_RIGHT_SYMMETRIC:
1716                         pd_idx = stripe % raid_disks;
1717                         qd_idx = (pd_idx + 1) % raid_disks;
1718                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1719                         break;
1720
1721                 case ALGORITHM_PARITY_0:
1722                         pd_idx = 0;
1723                         qd_idx = 1;
1724                         (*dd_idx) += 2;
1725                         break;
1726                 case ALGORITHM_PARITY_N:
1727                         pd_idx = data_disks;
1728                         qd_idx = data_disks + 1;
1729                         break;
1730
1731                 case ALGORITHM_ROTATING_ZERO_RESTART:
1732                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
1733                          * of blocks for computing Q is different.
1734                          */
1735                         pd_idx = stripe % raid_disks;
1736                         qd_idx = pd_idx + 1;
1737                         if (pd_idx == raid_disks-1) {
1738                                 (*dd_idx)++;    /* Q D D D P */
1739                                 qd_idx = 0;
1740                         } else if (*dd_idx >= pd_idx)
1741                                 (*dd_idx) += 2; /* D D P Q D */
1742                         ddf_layout = 1;
1743                         break;
1744
1745                 case ALGORITHM_ROTATING_N_RESTART:
1746                         /* Same a left_asymmetric, by first stripe is
1747                          * D D D P Q  rather than
1748                          * Q D D D P
1749                          */
1750                         pd_idx = raid_disks - 1 - ((stripe + 1) % raid_disks);
1751                         qd_idx = pd_idx + 1;
1752                         if (pd_idx == raid_disks-1) {
1753                                 (*dd_idx)++;    /* Q D D D P */
1754                                 qd_idx = 0;
1755                         } else if (*dd_idx >= pd_idx)
1756                                 (*dd_idx) += 2; /* D D P Q D */
1757                         ddf_layout = 1;
1758                         break;
1759
1760                 case ALGORITHM_ROTATING_N_CONTINUE:
1761                         /* Same as left_symmetric but Q is before P */
1762                         pd_idx = raid_disks - 1 - (stripe % raid_disks);
1763                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1764                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1765                         ddf_layout = 1;
1766                         break;
1767
1768                 case ALGORITHM_LEFT_ASYMMETRIC_6:
1769                         /* RAID5 left_asymmetric, with Q on last device */
1770                         pd_idx = data_disks - stripe % (raid_disks-1);
1771                         if (*dd_idx >= pd_idx)
1772                                 (*dd_idx)++;
1773                         qd_idx = raid_disks - 1;
1774                         break;
1775
1776                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1777                         pd_idx = stripe % (raid_disks-1);
1778                         if (*dd_idx >= pd_idx)
1779                                 (*dd_idx)++;
1780                         qd_idx = raid_disks - 1;
1781                         break;
1782
1783                 case ALGORITHM_LEFT_SYMMETRIC_6:
1784                         pd_idx = data_disks - stripe % (raid_disks-1);
1785                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1786                         qd_idx = raid_disks - 1;
1787                         break;
1788
1789                 case ALGORITHM_RIGHT_SYMMETRIC_6:
1790                         pd_idx = stripe % (raid_disks-1);
1791                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1792                         qd_idx = raid_disks - 1;
1793                         break;
1794
1795                 case ALGORITHM_PARITY_0_6:
1796                         pd_idx = 0;
1797                         (*dd_idx)++;
1798                         qd_idx = raid_disks - 1;
1799                         break;
1800
1801
1802                 default:
1803                         printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1804                                algorithm);
1805                         BUG();
1806                 }
1807                 break;
1808         }
1809
1810         if (sh) {
1811                 sh->pd_idx = pd_idx;
1812                 sh->qd_idx = qd_idx;
1813                 sh->ddf_layout = ddf_layout;
1814         }
1815         /*
1816          * Finally, compute the new sector number
1817          */
1818         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1819         return new_sector;
1820 }
1821
1822
1823 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1824 {
1825         raid5_conf_t *conf = sh->raid_conf;
1826         int raid_disks = sh->disks;
1827         int data_disks = raid_disks - conf->max_degraded;
1828         sector_t new_sector = sh->sector, check;
1829         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1830                                          : conf->chunk_sectors;
1831         int algorithm = previous ? conf->prev_algo
1832                                  : conf->algorithm;
1833         sector_t stripe;
1834         int chunk_offset;
1835         int chunk_number, dummy1, dd_idx = i;
1836         sector_t r_sector;
1837         struct stripe_head sh2;
1838
1839
1840         chunk_offset = sector_div(new_sector, sectors_per_chunk);
1841         stripe = new_sector;
1842         BUG_ON(new_sector != stripe);
1843
1844         if (i == sh->pd_idx)
1845                 return 0;
1846         switch(conf->level) {
1847         case 4: break;
1848         case 5:
1849                 switch (algorithm) {
1850                 case ALGORITHM_LEFT_ASYMMETRIC:
1851                 case ALGORITHM_RIGHT_ASYMMETRIC:
1852                         if (i > sh->pd_idx)
1853                                 i--;
1854                         break;
1855                 case ALGORITHM_LEFT_SYMMETRIC:
1856                 case ALGORITHM_RIGHT_SYMMETRIC:
1857                         if (i < sh->pd_idx)
1858                                 i += raid_disks;
1859                         i -= (sh->pd_idx + 1);
1860                         break;
1861                 case ALGORITHM_PARITY_0:
1862                         i -= 1;
1863                         break;
1864                 case ALGORITHM_PARITY_N:
1865                         break;
1866                 default:
1867                         printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1868                                algorithm);
1869                         BUG();
1870                 }
1871                 break;
1872         case 6:
1873                 if (i == sh->qd_idx)
1874                         return 0; /* It is the Q disk */
1875                 switch (algorithm) {
1876                 case ALGORITHM_LEFT_ASYMMETRIC:
1877                 case ALGORITHM_RIGHT_ASYMMETRIC:
1878                 case ALGORITHM_ROTATING_ZERO_RESTART:
1879                 case ALGORITHM_ROTATING_N_RESTART:
1880                         if (sh->pd_idx == raid_disks-1)
1881                                 i--;    /* Q D D D P */
1882                         else if (i > sh->pd_idx)
1883                                 i -= 2; /* D D P Q D */
1884                         break;
1885                 case ALGORITHM_LEFT_SYMMETRIC:
1886                 case ALGORITHM_RIGHT_SYMMETRIC:
1887                         if (sh->pd_idx == raid_disks-1)
1888                                 i--; /* Q D D D P */
1889                         else {
1890                                 /* D D P Q D */
1891                                 if (i < sh->pd_idx)
1892                                         i += raid_disks;
1893                                 i -= (sh->pd_idx + 2);
1894                         }
1895                         break;
1896                 case ALGORITHM_PARITY_0:
1897                         i -= 2;
1898                         break;
1899                 case ALGORITHM_PARITY_N:
1900                         break;
1901                 case ALGORITHM_ROTATING_N_CONTINUE:
1902                         if (sh->pd_idx == 0)
1903                                 i--;    /* P D D D Q */
1904                         else if (i > sh->pd_idx)
1905                                 i -= 2; /* D D Q P D */
1906                         break;
1907                 case ALGORITHM_LEFT_ASYMMETRIC_6:
1908                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1909                         if (i > sh->pd_idx)
1910                                 i--;
1911                         break;
1912                 case ALGORITHM_LEFT_SYMMETRIC_6:
1913                 case ALGORITHM_RIGHT_SYMMETRIC_6:
1914                         if (i < sh->pd_idx)
1915                                 i += data_disks + 1;
1916                         i -= (sh->pd_idx + 1);
1917                         break;
1918                 case ALGORITHM_PARITY_0_6:
1919                         i -= 1;
1920                         break;
1921                 default:
1922                         printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1923                                algorithm);
1924                         BUG();
1925                 }
1926                 break;
1927         }
1928
1929         chunk_number = stripe * data_disks + i;
1930         r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1931
1932         check = raid5_compute_sector(conf, r_sector,
1933                                      previous, &dummy1, &sh2);
1934         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
1935                 || sh2.qd_idx != sh->qd_idx) {
1936                 printk(KERN_ERR "compute_blocknr: map not correct\n");
1937                 return 0;
1938         }
1939         return r_sector;
1940 }
1941
1942
1943 static void
1944 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
1945                          int rcw, int expand)
1946 {
1947         int i, pd_idx = sh->pd_idx, disks = sh->disks;
1948         raid5_conf_t *conf = sh->raid_conf;
1949         int level = conf->level;
1950
1951         if (rcw) {
1952                 /* if we are not expanding this is a proper write request, and
1953                  * there will be bios with new data to be drained into the
1954                  * stripe cache
1955                  */
1956                 if (!expand) {
1957                         sh->reconstruct_state = reconstruct_state_drain_run;
1958                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1959                 } else
1960                         sh->reconstruct_state = reconstruct_state_run;
1961
1962                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
1963
1964                 for (i = disks; i--; ) {
1965                         struct r5dev *dev = &sh->dev[i];
1966
1967                         if (dev->towrite) {
1968                                 set_bit(R5_LOCKED, &dev->flags);
1969                                 set_bit(R5_Wantdrain, &dev->flags);
1970                                 if (!expand)
1971                                         clear_bit(R5_UPTODATE, &dev->flags);
1972                                 s->locked++;
1973                         }
1974                 }
1975                 if (s->locked + conf->max_degraded == disks)
1976                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
1977                                 atomic_inc(&conf->pending_full_writes);
1978         } else {
1979                 BUG_ON(level == 6);
1980                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1981                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1982
1983                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
1984                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
1985                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1986                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
1987
1988                 for (i = disks; i--; ) {
1989                         struct r5dev *dev = &sh->dev[i];
1990                         if (i == pd_idx)
1991                                 continue;
1992
1993                         if (dev->towrite &&
1994                             (test_bit(R5_UPTODATE, &dev->flags) ||
1995                              test_bit(R5_Wantcompute, &dev->flags))) {
1996                                 set_bit(R5_Wantdrain, &dev->flags);
1997                                 set_bit(R5_LOCKED, &dev->flags);
1998                                 clear_bit(R5_UPTODATE, &dev->flags);
1999                                 s->locked++;
2000                         }
2001                 }
2002         }
2003
2004         /* keep the parity disk(s) locked while asynchronous operations
2005          * are in flight
2006          */
2007         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2008         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2009         s->locked++;
2010
2011         if (level == 6) {
2012                 int qd_idx = sh->qd_idx;
2013                 struct r5dev *dev = &sh->dev[qd_idx];
2014
2015                 set_bit(R5_LOCKED, &dev->flags);
2016                 clear_bit(R5_UPTODATE, &dev->flags);
2017                 s->locked++;
2018         }
2019
2020         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2021                 __func__, (unsigned long long)sh->sector,
2022                 s->locked, s->ops_request);
2023 }
2024
2025 /*
2026  * Each stripe/dev can have one or more bion attached.
2027  * toread/towrite point to the first in a chain.
2028  * The bi_next chain must be in order.
2029  */
2030 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2031 {
2032         struct bio **bip;
2033         raid5_conf_t *conf = sh->raid_conf;
2034         int firstwrite=0;
2035
2036         pr_debug("adding bh b#%llu to stripe s#%llu\n",
2037                 (unsigned long long)bi->bi_sector,
2038                 (unsigned long long)sh->sector);
2039
2040
2041         spin_lock(&sh->lock);
2042         spin_lock_irq(&conf->device_lock);
2043         if (forwrite) {
2044                 bip = &sh->dev[dd_idx].towrite;
2045                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2046                         firstwrite = 1;
2047         } else
2048                 bip = &sh->dev[dd_idx].toread;
2049         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2050                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2051                         goto overlap;
2052                 bip = & (*bip)->bi_next;
2053         }
2054         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2055                 goto overlap;
2056
2057         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2058         if (*bip)
2059                 bi->bi_next = *bip;
2060         *bip = bi;
2061         bi->bi_phys_segments++;
2062         spin_unlock_irq(&conf->device_lock);
2063         spin_unlock(&sh->lock);
2064
2065         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2066                 (unsigned long long)bi->bi_sector,
2067                 (unsigned long long)sh->sector, dd_idx);
2068
2069         if (conf->mddev->bitmap && firstwrite) {
2070                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2071                                   STRIPE_SECTORS, 0);
2072                 sh->bm_seq = conf->seq_flush+1;
2073                 set_bit(STRIPE_BIT_DELAY, &sh->state);
2074         }
2075
2076         if (forwrite) {
2077                 /* check if page is covered */
2078                 sector_t sector = sh->dev[dd_idx].sector;
2079                 for (bi=sh->dev[dd_idx].towrite;
2080                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2081                              bi && bi->bi_sector <= sector;
2082                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2083                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2084                                 sector = bi->bi_sector + (bi->bi_size>>9);
2085                 }
2086                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2087                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2088         }
2089         return 1;
2090
2091  overlap:
2092         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2093         spin_unlock_irq(&conf->device_lock);
2094         spin_unlock(&sh->lock);
2095         return 0;
2096 }
2097
2098 static void end_reshape(raid5_conf_t *conf);
2099
2100 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
2101                             struct stripe_head *sh)
2102 {
2103         int sectors_per_chunk =
2104                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2105         int dd_idx;
2106         int chunk_offset = sector_div(stripe, sectors_per_chunk);
2107         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2108
2109         raid5_compute_sector(conf,
2110                              stripe * (disks - conf->max_degraded)
2111                              *sectors_per_chunk + chunk_offset,
2112                              previous,
2113                              &dd_idx, sh);
2114 }
2115
2116 static void
2117 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2118                                 struct stripe_head_state *s, int disks,
2119                                 struct bio **return_bi)
2120 {
2121         int i;
2122         for (i = disks; i--; ) {
2123                 struct bio *bi;
2124                 int bitmap_end = 0;
2125
2126                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2127                         mdk_rdev_t *rdev;
2128                         rcu_read_lock();
2129                         rdev = rcu_dereference(conf->disks[i].rdev);
2130                         if (rdev && test_bit(In_sync, &rdev->flags))
2131                                 /* multiple read failures in one stripe */
2132                                 md_error(conf->mddev, rdev);
2133                         rcu_read_unlock();
2134                 }
2135                 spin_lock_irq(&conf->device_lock);
2136                 /* fail all writes first */
2137                 bi = sh->dev[i].towrite;
2138                 sh->dev[i].towrite = NULL;
2139                 if (bi) {
2140                         s->to_write--;
2141                         bitmap_end = 1;
2142                 }
2143
2144                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2145                         wake_up(&conf->wait_for_overlap);
2146
2147                 while (bi && bi->bi_sector <
2148                         sh->dev[i].sector + STRIPE_SECTORS) {
2149                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2150                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2151                         if (!raid5_dec_bi_phys_segments(bi)) {
2152                                 md_write_end(conf->mddev);
2153                                 bi->bi_next = *return_bi;
2154                                 *return_bi = bi;
2155                         }
2156                         bi = nextbi;
2157                 }
2158                 /* and fail all 'written' */
2159                 bi = sh->dev[i].written;
2160                 sh->dev[i].written = NULL;
2161                 if (bi) bitmap_end = 1;
2162                 while (bi && bi->bi_sector <
2163                        sh->dev[i].sector + STRIPE_SECTORS) {
2164                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2165                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2166                         if (!raid5_dec_bi_phys_segments(bi)) {
2167                                 md_write_end(conf->mddev);
2168                                 bi->bi_next = *return_bi;
2169                                 *return_bi = bi;
2170                         }
2171                         bi = bi2;
2172                 }
2173
2174                 /* fail any reads if this device is non-operational and
2175                  * the data has not reached the cache yet.
2176                  */
2177                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2178                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2179                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
2180                         bi = sh->dev[i].toread;
2181                         sh->dev[i].toread = NULL;
2182                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2183                                 wake_up(&conf->wait_for_overlap);
2184                         if (bi) s->to_read--;
2185                         while (bi && bi->bi_sector <
2186                                sh->dev[i].sector + STRIPE_SECTORS) {
2187                                 struct bio *nextbi =
2188                                         r5_next_bio(bi, sh->dev[i].sector);
2189                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2190                                 if (!raid5_dec_bi_phys_segments(bi)) {
2191                                         bi->bi_next = *return_bi;
2192                                         *return_bi = bi;
2193                                 }
2194                                 bi = nextbi;
2195                         }
2196                 }
2197                 spin_unlock_irq(&conf->device_lock);
2198                 if (bitmap_end)
2199                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2200                                         STRIPE_SECTORS, 0, 0);
2201         }
2202
2203         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2204                 if (atomic_dec_and_test(&conf->pending_full_writes))
2205                         md_wakeup_thread(conf->mddev->thread);
2206 }
2207
2208 /* fetch_block5 - checks the given member device to see if its data needs
2209  * to be read or computed to satisfy a request.
2210  *
2211  * Returns 1 when no more member devices need to be checked, otherwise returns
2212  * 0 to tell the loop in handle_stripe_fill5 to continue
2213  */
2214 static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s,
2215                         int disk_idx, int disks)
2216 {
2217         struct r5dev *dev = &sh->dev[disk_idx];
2218         struct r5dev *failed_dev = &sh->dev[s->failed_num];
2219
2220         /* is the data in this block needed, and can we get it? */
2221         if (!test_bit(R5_LOCKED, &dev->flags) &&
2222             !test_bit(R5_UPTODATE, &dev->flags) &&
2223             (dev->toread ||
2224              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2225              s->syncing || s->expanding ||
2226              (s->failed &&
2227               (failed_dev->toread ||
2228                (failed_dev->towrite &&
2229                 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) {
2230                 /* We would like to get this block, possibly by computing it,
2231                  * otherwise read it if the backing disk is insync
2232                  */
2233                 if ((s->uptodate == disks - 1) &&
2234                     (s->failed && disk_idx == s->failed_num)) {
2235                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2236                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2237                         set_bit(R5_Wantcompute, &dev->flags);
2238                         sh->ops.target = disk_idx;
2239                         sh->ops.target2 = -1;
2240                         s->req_compute = 1;
2241                         /* Careful: from this point on 'uptodate' is in the eye
2242                          * of raid_run_ops which services 'compute' operations
2243                          * before writes. R5_Wantcompute flags a block that will
2244                          * be R5_UPTODATE by the time it is needed for a
2245                          * subsequent operation.
2246                          */
2247                         s->uptodate++;
2248                         return 1; /* uptodate + compute == disks */
2249                 } else if (test_bit(R5_Insync, &dev->flags)) {
2250                         set_bit(R5_LOCKED, &dev->flags);
2251                         set_bit(R5_Wantread, &dev->flags);
2252                         s->locked++;
2253                         pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2254                                 s->syncing);
2255                 }
2256         }
2257
2258         return 0;
2259 }
2260
2261 /**
2262  * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2263  */
2264 static void handle_stripe_fill5(struct stripe_head *sh,
2265                         struct stripe_head_state *s, int disks)
2266 {
2267         int i;
2268
2269         /* look for blocks to read/compute, skip this if a compute
2270          * is already in flight, or if the stripe contents are in the
2271          * midst of changing due to a write
2272          */
2273         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2274             !sh->reconstruct_state)
2275                 for (i = disks; i--; )
2276                         if (fetch_block5(sh, s, i, disks))
2277                                 break;
2278         set_bit(STRIPE_HANDLE, &sh->state);
2279 }
2280
2281 /* fetch_block6 - checks the given member device to see if its data needs
2282  * to be read or computed to satisfy a request.
2283  *
2284  * Returns 1 when no more member devices need to be checked, otherwise returns
2285  * 0 to tell the loop in handle_stripe_fill6 to continue
2286  */
2287 static int fetch_block6(struct stripe_head *sh, struct stripe_head_state *s,
2288                          struct r6_state *r6s, int disk_idx, int disks)
2289 {
2290         struct r5dev *dev = &sh->dev[disk_idx];
2291         struct r5dev *fdev[2] = { &sh->dev[r6s->failed_num[0]],
2292                                   &sh->dev[r6s->failed_num[1]] };
2293
2294         if (!test_bit(R5_LOCKED, &dev->flags) &&
2295             !test_bit(R5_UPTODATE, &dev->flags) &&
2296             (dev->toread ||
2297              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2298              s->syncing || s->expanding ||
2299              (s->failed >= 1 &&
2300               (fdev[0]->toread || s->to_write)) ||
2301              (s->failed >= 2 &&
2302               (fdev[1]->toread || s->to_write)))) {
2303                 /* we would like to get this block, possibly by computing it,
2304                  * otherwise read it if the backing disk is insync
2305                  */
2306                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2307                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2308                 if ((s->uptodate == disks - 1) &&
2309                     (s->failed && (disk_idx == r6s->failed_num[0] ||
2310                                    disk_idx == r6s->failed_num[1]))) {
2311                         /* have disk failed, and we're requested to fetch it;
2312                          * do compute it
2313                          */
2314                         pr_debug("Computing stripe %llu block %d\n",
2315                                (unsigned long long)sh->sector, disk_idx);
2316                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2317                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2318                         set_bit(R5_Wantcompute, &dev->flags);
2319                         sh->ops.target = disk_idx;
2320                         sh->ops.target2 = -1; /* no 2nd target */
2321                         s->req_compute = 1;
2322                         s->uptodate++;
2323                         return 1;
2324                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2325                         /* Computing 2-failure is *very* expensive; only
2326                          * do it if failed >= 2
2327                          */
2328                         int other;
2329                         for (other = disks; other--; ) {
2330                                 if (other == disk_idx)
2331                                         continue;
2332                                 if (!test_bit(R5_UPTODATE,
2333                                       &sh->dev[other].flags))
2334                                         break;
2335                         }
2336                         BUG_ON(other < 0);
2337                         pr_debug("Computing stripe %llu blocks %d,%d\n",
2338                                (unsigned long long)sh->sector,
2339                                disk_idx, other);
2340                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2341                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2342                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2343                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
2344                         sh->ops.target = disk_idx;
2345                         sh->ops.target2 = other;
2346                         s->uptodate += 2;
2347                         s->req_compute = 1;
2348                         return 1;
2349                 } else if (test_bit(R5_Insync, &dev->flags)) {
2350                         set_bit(R5_LOCKED, &dev->flags);
2351                         set_bit(R5_Wantread, &dev->flags);
2352                         s->locked++;
2353                         pr_debug("Reading block %d (sync=%d)\n",
2354                                 disk_idx, s->syncing);
2355                 }
2356         }
2357
2358         return 0;
2359 }
2360
2361 /**
2362  * handle_stripe_fill6 - read or compute data to satisfy pending requests.
2363  */
2364 static void handle_stripe_fill6(struct stripe_head *sh,
2365                         struct stripe_head_state *s, struct r6_state *r6s,
2366                         int disks)
2367 {
2368         int i;
2369
2370         /* look for blocks to read/compute, skip this if a compute
2371          * is already in flight, or if the stripe contents are in the
2372          * midst of changing due to a write
2373          */
2374         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2375             !sh->reconstruct_state)
2376                 for (i = disks; i--; )
2377                         if (fetch_block6(sh, s, r6s, i, disks))
2378                                 break;
2379         set_bit(STRIPE_HANDLE, &sh->state);
2380 }
2381
2382
2383 /* handle_stripe_clean_event
2384  * any written block on an uptodate or failed drive can be returned.
2385  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2386  * never LOCKED, so we don't need to test 'failed' directly.
2387  */
2388 static void handle_stripe_clean_event(raid5_conf_t *conf,
2389         struct stripe_head *sh, int disks, struct bio **return_bi)
2390 {
2391         int i;
2392         struct r5dev *dev;
2393
2394         for (i = disks; i--; )
2395                 if (sh->dev[i].written) {
2396                         dev = &sh->dev[i];
2397                         if (!test_bit(R5_LOCKED, &dev->flags) &&
2398                                 test_bit(R5_UPTODATE, &dev->flags)) {
2399                                 /* We can return any write requests */
2400                                 struct bio *wbi, *wbi2;
2401                                 int bitmap_end = 0;
2402                                 pr_debug("Return write for disc %d\n", i);
2403                                 spin_lock_irq(&conf->device_lock);
2404                                 wbi = dev->written;
2405                                 dev->written = NULL;
2406                                 while (wbi && wbi->bi_sector <
2407                                         dev->sector + STRIPE_SECTORS) {
2408                                         wbi2 = r5_next_bio(wbi, dev->sector);
2409                                         if (!raid5_dec_bi_phys_segments(wbi)) {
2410                                                 md_write_end(conf->mddev);
2411                                                 wbi->bi_next = *return_bi;
2412                                                 *return_bi = wbi;
2413                                         }
2414                                         wbi = wbi2;
2415                                 }
2416                                 if (dev->towrite == NULL)
2417                                         bitmap_end = 1;
2418                                 spin_unlock_irq(&conf->device_lock);
2419                                 if (bitmap_end)
2420                                         bitmap_endwrite(conf->mddev->bitmap,
2421                                                         sh->sector,
2422                                                         STRIPE_SECTORS,
2423                                          !test_bit(STRIPE_DEGRADED, &sh->state),
2424                                                         0);
2425                         }
2426                 }
2427
2428         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2429                 if (atomic_dec_and_test(&conf->pending_full_writes))
2430                         md_wakeup_thread(conf->mddev->thread);
2431 }
2432
2433 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2434                 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2435 {
2436         int rmw = 0, rcw = 0, i;
2437         for (i = disks; i--; ) {
2438                 /* would I have to read this buffer for read_modify_write */
2439                 struct r5dev *dev = &sh->dev[i];
2440                 if ((dev->towrite || i == sh->pd_idx) &&
2441                     !test_bit(R5_LOCKED, &dev->flags) &&
2442                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2443                       test_bit(R5_Wantcompute, &dev->flags))) {
2444                         if (test_bit(R5_Insync, &dev->flags))
2445                                 rmw++;
2446                         else
2447                                 rmw += 2*disks;  /* cannot read it */
2448                 }
2449                 /* Would I have to read this buffer for reconstruct_write */
2450                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2451                     !test_bit(R5_LOCKED, &dev->flags) &&
2452                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2453                     test_bit(R5_Wantcompute, &dev->flags))) {
2454                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2455                         else
2456                                 rcw += 2*disks;
2457                 }
2458         }
2459         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2460                 (unsigned long long)sh->sector, rmw, rcw);
2461         set_bit(STRIPE_HANDLE, &sh->state);
2462         if (rmw < rcw && rmw > 0)
2463                 /* prefer read-modify-write, but need to get some data */
2464                 for (i = disks; i--; ) {
2465                         struct r5dev *dev = &sh->dev[i];
2466                         if ((dev->towrite || i == sh->pd_idx) &&
2467                             !test_bit(R5_LOCKED, &dev->flags) &&
2468                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2469                             test_bit(R5_Wantcompute, &dev->flags)) &&
2470                             test_bit(R5_Insync, &dev->flags)) {
2471                                 if (
2472                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2473                                         pr_debug("Read_old block "
2474                                                 "%d for r-m-w\n", i);
2475                                         set_bit(R5_LOCKED, &dev->flags);
2476                                         set_bit(R5_Wantread, &dev->flags);
2477                                         s->locked++;
2478                                 } else {
2479                                         set_bit(STRIPE_DELAYED, &sh->state);
2480                                         set_bit(STRIPE_HANDLE, &sh->state);
2481                                 }
2482                         }
2483                 }
2484         if (rcw <= rmw && rcw > 0)
2485                 /* want reconstruct write, but need to get some data */
2486                 for (i = disks; i--; ) {
2487                         struct r5dev *dev = &sh->dev[i];
2488                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2489                             i != sh->pd_idx &&
2490                             !test_bit(R5_LOCKED, &dev->flags) &&
2491                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2492                             test_bit(R5_Wantcompute, &dev->flags)) &&
2493                             test_bit(R5_Insync, &dev->flags)) {
2494                                 if (
2495                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2496                                         pr_debug("Read_old block "
2497                                                 "%d for Reconstruct\n", i);
2498                                         set_bit(R5_LOCKED, &dev->flags);
2499                                         set_bit(R5_Wantread, &dev->flags);
2500                                         s->locked++;
2501                                 } else {
2502                                         set_bit(STRIPE_DELAYED, &sh->state);
2503                                         set_bit(STRIPE_HANDLE, &sh->state);
2504                                 }
2505                         }
2506                 }
2507         /* now if nothing is locked, and if we have enough data,
2508          * we can start a write request
2509          */
2510         /* since handle_stripe can be called at any time we need to handle the
2511          * case where a compute block operation has been submitted and then a
2512          * subsequent call wants to start a write request.  raid_run_ops only
2513          * handles the case where compute block and reconstruct are requested
2514          * simultaneously.  If this is not the case then new writes need to be
2515          * held off until the compute completes.
2516          */
2517         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2518             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2519             !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2520                 schedule_reconstruction(sh, s, rcw == 0, 0);
2521 }
2522
2523 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2524                 struct stripe_head *sh, struct stripe_head_state *s,
2525                 struct r6_state *r6s, int disks)
2526 {
2527         int rcw = 0, pd_idx = sh->pd_idx, i;
2528         int qd_idx = sh->qd_idx;
2529
2530         set_bit(STRIPE_HANDLE, &sh->state);
2531         for (i = disks; i--; ) {
2532                 struct r5dev *dev = &sh->dev[i];
2533                 /* check if we haven't enough data */
2534                 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2535                     i != pd_idx && i != qd_idx &&
2536                     !test_bit(R5_LOCKED, &dev->flags) &&
2537                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2538                       test_bit(R5_Wantcompute, &dev->flags))) {
2539                         rcw++;
2540                         if (!test_bit(R5_Insync, &dev->flags))
2541                                 continue; /* it's a failed drive */
2542
2543                         if (
2544                           test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2545                                 pr_debug("Read_old stripe %llu "
2546                                         "block %d for Reconstruct\n",
2547                                      (unsigned long long)sh->sector, i);
2548                                 set_bit(R5_LOCKED, &dev->flags);
2549                                 set_bit(R5_Wantread, &dev->flags);
2550                                 s->locked++;
2551                         } else {
2552                                 pr_debug("Request delayed stripe %llu "
2553                                         "block %d for Reconstruct\n",
2554                                      (unsigned long long)sh->sector, i);
2555                                 set_bit(STRIPE_DELAYED, &sh->state);
2556                                 set_bit(STRIPE_HANDLE, &sh->state);
2557                         }
2558                 }
2559         }
2560         /* now if nothing is locked, and if we have enough data, we can start a
2561          * write request
2562          */
2563         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2564             s->locked == 0 && rcw == 0 &&
2565             !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2566                 schedule_reconstruction(sh, s, 1, 0);
2567         }
2568 }
2569
2570 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2571                                 struct stripe_head_state *s, int disks)
2572 {
2573         struct r5dev *dev = NULL;
2574
2575         set_bit(STRIPE_HANDLE, &sh->state);
2576
2577         switch (sh->check_state) {
2578         case check_state_idle:
2579                 /* start a new check operation if there are no failures */
2580                 if (s->failed == 0) {
2581                         BUG_ON(s->uptodate != disks);
2582                         sh->check_state = check_state_run;
2583                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
2584                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2585                         s->uptodate--;
2586                         break;
2587                 }
2588                 dev = &sh->dev[s->failed_num];
2589                 /* fall through */
2590         case check_state_compute_result:
2591                 sh->check_state = check_state_idle;
2592                 if (!dev)
2593                         dev = &sh->dev[sh->pd_idx];
2594
2595                 /* check that a write has not made the stripe insync */
2596                 if (test_bit(STRIPE_INSYNC, &sh->state))
2597                         break;
2598
2599                 /* either failed parity check, or recovery is happening */
2600                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2601                 BUG_ON(s->uptodate != disks);
2602
2603                 set_bit(R5_LOCKED, &dev->flags);
2604                 s->locked++;
2605                 set_bit(R5_Wantwrite, &dev->flags);
2606
2607                 clear_bit(STRIPE_DEGRADED, &sh->state);
2608                 set_bit(STRIPE_INSYNC, &sh->state);
2609                 break;
2610         case check_state_run:
2611                 break; /* we will be called again upon completion */
2612         case check_state_check_result:
2613                 sh->check_state = check_state_idle;
2614
2615                 /* if a failure occurred during the check operation, leave
2616                  * STRIPE_INSYNC not set and let the stripe be handled again
2617                  */
2618                 if (s->failed)
2619                         break;
2620
2621                 /* handle a successful check operation, if parity is correct
2622                  * we are done.  Otherwise update the mismatch count and repair
2623                  * parity if !MD_RECOVERY_CHECK
2624                  */
2625                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2626                         /* parity is correct (on disc,
2627                          * not in buffer any more)
2628                          */
2629                         set_bit(STRIPE_INSYNC, &sh->state);
2630                 else {
2631                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2632                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2633                                 /* don't try to repair!! */
2634                                 set_bit(STRIPE_INSYNC, &sh->state);
2635                         else {
2636                                 sh->check_state = check_state_compute_run;
2637                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2638                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2639                                 set_bit(R5_Wantcompute,
2640                                         &sh->dev[sh->pd_idx].flags);
2641                                 sh->ops.target = sh->pd_idx;
2642                                 sh->ops.target2 = -1;
2643                                 s->uptodate++;
2644                         }
2645                 }
2646                 break;
2647         case check_state_compute_run:
2648                 break;
2649         default:
2650                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2651                        __func__, sh->check_state,
2652                        (unsigned long long) sh->sector);
2653                 BUG();
2654         }
2655 }
2656
2657
2658 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2659                                   struct stripe_head_state *s,
2660                                   struct r6_state *r6s, int disks)
2661 {
2662         int pd_idx = sh->pd_idx;
2663         int qd_idx = sh->qd_idx;
2664         struct r5dev *dev;
2665
2666         set_bit(STRIPE_HANDLE, &sh->state);
2667
2668         BUG_ON(s->failed > 2);
2669
2670         /* Want to check and possibly repair P and Q.
2671          * However there could be one 'failed' device, in which
2672          * case we can only check one of them, possibly using the
2673          * other to generate missing data
2674          */
2675
2676         switch (sh->check_state) {
2677         case check_state_idle:
2678                 /* start a new check operation if there are < 2 failures */
2679                 if (s->failed == r6s->q_failed) {
2680                         /* The only possible failed device holds Q, so it
2681                          * makes sense to check P (If anything else were failed,
2682                          * we would have used P to recreate it).
2683                          */
2684                         sh->check_state = check_state_run;
2685                 }
2686                 if (!r6s->q_failed && s->failed < 2) {
2687                         /* Q is not failed, and we didn't use it to generate
2688                          * anything, so it makes sense to check it
2689                          */
2690                         if (sh->check_state == check_state_run)
2691                                 sh->check_state = check_state_run_pq;
2692                         else
2693                                 sh->check_state = check_state_run_q;
2694                 }
2695
2696                 /* discard potentially stale zero_sum_result */
2697                 sh->ops.zero_sum_result = 0;
2698
2699                 if (sh->check_state == check_state_run) {
2700                         /* async_xor_zero_sum destroys the contents of P */
2701                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2702                         s->uptodate--;
2703                 }
2704                 if (sh->check_state >= check_state_run &&
2705                     sh->check_state <= check_state_run_pq) {
2706                         /* async_syndrome_zero_sum preserves P and Q, so
2707                          * no need to mark them !uptodate here
2708                          */
2709                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
2710                         break;
2711                 }
2712
2713                 /* we have 2-disk failure */
2714                 BUG_ON(s->failed != 2);
2715                 /* fall through */
2716         case check_state_compute_result:
2717                 sh->check_state = check_state_idle;
2718
2719                 /* check that a write has not made the stripe insync */
2720                 if (test_bit(STRIPE_INSYNC, &sh->state))
2721                         break;
2722
2723                 /* now write out any block on a failed drive,
2724                  * or P or Q if they were recomputed
2725                  */
2726                 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2727                 if (s->failed == 2) {
2728                         dev = &sh->dev[r6s->failed_num[1]];
2729                         s->locked++;
2730                         set_bit(R5_LOCKED, &dev->flags);
2731                         set_bit(R5_Wantwrite, &dev->flags);
2732                 }
2733                 if (s->failed >= 1) {
2734                         dev = &sh->dev[r6s->failed_num[0]];
2735                         s->locked++;
2736                         set_bit(R5_LOCKED, &dev->flags);
2737                         set_bit(R5_Wantwrite, &dev->flags);
2738                 }
2739                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2740                         dev = &sh->dev[pd_idx];
2741                         s->locked++;
2742                         set_bit(R5_LOCKED, &dev->flags);
2743                         set_bit(R5_Wantwrite, &dev->flags);
2744                 }
2745                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2746                         dev = &sh->dev[qd_idx];
2747                         s->locked++;
2748                         set_bit(R5_LOCKED, &dev->flags);
2749                         set_bit(R5_Wantwrite, &dev->flags);
2750                 }
2751                 clear_bit(STRIPE_DEGRADED, &sh->state);
2752
2753                 set_bit(STRIPE_INSYNC, &sh->state);
2754                 break;
2755         case check_state_run:
2756         case check_state_run_q:
2757         case check_state_run_pq:
2758                 break; /* we will be called again upon completion */
2759         case check_state_check_result:
2760                 sh->check_state = check_state_idle;
2761
2762                 /* handle a successful check operation, if parity is correct
2763                  * we are done.  Otherwise update the mismatch count and repair
2764                  * parity if !MD_RECOVERY_CHECK
2765                  */
2766                 if (sh->ops.zero_sum_result == 0) {
2767                         /* both parities are correct */
2768                         if (!s->failed)
2769                                 set_bit(STRIPE_INSYNC, &sh->state);
2770                         else {
2771                                 /* in contrast to the raid5 case we can validate
2772                                  * parity, but still have a failure to write
2773                                  * back
2774                                  */
2775                                 sh->check_state = check_state_compute_result;
2776                                 /* Returning at this point means that we may go
2777                                  * off and bring p and/or q uptodate again so
2778                                  * we make sure to check zero_sum_result again
2779                                  * to verify if p or q need writeback
2780                                  */
2781                         }
2782                 } else {
2783                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2784                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2785                                 /* don't try to repair!! */
2786                                 set_bit(STRIPE_INSYNC, &sh->state);
2787                         else {
2788                                 int *target = &sh->ops.target;
2789
2790                                 sh->ops.target = -1;
2791                                 sh->ops.target2 = -1;
2792                                 sh->check_state = check_state_compute_run;
2793                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2794                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2795                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2796                                         set_bit(R5_Wantcompute,
2797                                                 &sh->dev[pd_idx].flags);
2798                                         *target = pd_idx;
2799                                         target = &sh->ops.target2;
2800                                         s->uptodate++;
2801                                 }
2802                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2803                                         set_bit(R5_Wantcompute,
2804                                                 &sh->dev[qd_idx].flags);
2805                                         *target = qd_idx;
2806                                         s->uptodate++;
2807                                 }
2808                         }
2809                 }
2810                 break;
2811         case check_state_compute_run:
2812                 break;
2813         default:
2814                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2815                        __func__, sh->check_state,
2816                        (unsigned long long) sh->sector);
2817                 BUG();
2818         }
2819 }
2820
2821 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2822                                 struct r6_state *r6s)
2823 {
2824         int i;
2825
2826         /* We have read all the blocks in this stripe and now we need to
2827          * copy some of them into a target stripe for expand.
2828          */
2829         struct dma_async_tx_descriptor *tx = NULL;
2830         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2831         for (i = 0; i < sh->disks; i++)
2832                 if (i != sh->pd_idx && i != sh->qd_idx) {
2833                         int dd_idx, j;
2834                         struct stripe_head *sh2;
2835                         struct async_submit_ctl submit;
2836
2837                         sector_t bn = compute_blocknr(sh, i, 1);
2838                         sector_t s = raid5_compute_sector(conf, bn, 0,
2839                                                           &dd_idx, NULL);
2840                         sh2 = get_active_stripe(conf, s, 0, 1, 1);
2841                         if (sh2 == NULL)
2842                                 /* so far only the early blocks of this stripe
2843                                  * have been requested.  When later blocks
2844                                  * get requested, we will try again
2845                                  */
2846                                 continue;
2847                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2848                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2849                                 /* must have already done this block */
2850                                 release_stripe(sh2);
2851                                 continue;
2852                         }
2853
2854                         /* place all the copies on one channel */
2855                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
2856                         tx = async_memcpy(sh2->dev[dd_idx].page,
2857                                           sh->dev[i].page, 0, 0, STRIPE_SIZE,
2858                                           &submit);
2859
2860                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2861                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2862                         for (j = 0; j < conf->raid_disks; j++)
2863                                 if (j != sh2->pd_idx &&
2864                                     (!r6s || j != sh2->qd_idx) &&
2865                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
2866                                         break;
2867                         if (j == conf->raid_disks) {
2868                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2869                                 set_bit(STRIPE_HANDLE, &sh2->state);
2870                         }
2871                         release_stripe(sh2);
2872
2873                 }
2874         /* done submitting copies, wait for them to complete */
2875         if (tx) {
2876                 async_tx_ack(tx);
2877                 dma_wait_for_async_tx(tx);
2878         }
2879 }
2880
2881
2882 /*
2883  * handle_stripe - do things to a stripe.
2884  *
2885  * We lock the stripe and then examine the state of various bits
2886  * to see what needs to be done.
2887  * Possible results:
2888  *    return some read request which now have data
2889  *    return some write requests which are safely on disc
2890  *    schedule a read on some buffers
2891  *    schedule a write of some buffers
2892  *    return confirmation of parity correctness
2893  *
2894  * buffers are taken off read_list or write_list, and bh_cache buffers
2895  * get BH_Lock set before the stripe lock is released.
2896  *
2897  */
2898
2899 static bool handle_stripe5(struct stripe_head *sh)
2900 {
2901         raid5_conf_t *conf = sh->raid_conf;
2902         int disks = sh->disks, i;
2903         struct bio *return_bi = NULL;
2904         struct stripe_head_state s;
2905         struct r5dev *dev;
2906         mdk_rdev_t *blocked_rdev = NULL;
2907         int prexor;
2908
2909         memset(&s, 0, sizeof(s));
2910         pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
2911                  "reconstruct:%d\n", (unsigned long long)sh->sector, sh->state,
2912                  atomic_read(&sh->count), sh->pd_idx, sh->check_state,
2913                  sh->reconstruct_state);
2914
2915         spin_lock(&sh->lock);
2916         clear_bit(STRIPE_HANDLE, &sh->state);
2917         clear_bit(STRIPE_DELAYED, &sh->state);
2918
2919         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2920         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2921         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2922
2923         /* Now to look around and see what can be done */
2924         rcu_read_lock();
2925         for (i=disks; i--; ) {
2926                 mdk_rdev_t *rdev;
2927
2928                 dev = &sh->dev[i];
2929                 clear_bit(R5_Insync, &dev->flags);
2930
2931                 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2932                         "written %p\n", i, dev->flags, dev->toread, dev->read,
2933                         dev->towrite, dev->written);
2934
2935                 /* maybe we can request a biofill operation
2936                  *
2937                  * new wantfill requests are only permitted while
2938                  * ops_complete_biofill is guaranteed to be inactive
2939                  */
2940                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2941                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
2942                         set_bit(R5_Wantfill, &dev->flags);
2943
2944                 /* now count some things */
2945                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2946                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2947                 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2948
2949                 if (test_bit(R5_Wantfill, &dev->flags))
2950                         s.to_fill++;
2951                 else if (dev->toread)
2952                         s.to_read++;
2953                 if (dev->towrite) {
2954                         s.to_write++;
2955                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2956                                 s.non_overwrite++;
2957                 }
2958                 if (dev->written)
2959                         s.written++;
2960                 rdev = rcu_dereference(conf->disks[i].rdev);
2961                 if (blocked_rdev == NULL &&
2962                     rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2963                         blocked_rdev = rdev;
2964                         atomic_inc(&rdev->nr_pending);
2965                 }
2966                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2967                         /* The ReadError flag will just be confusing now */
2968                         clear_bit(R5_ReadError, &dev->flags);
2969                         clear_bit(R5_ReWrite, &dev->flags);
2970                 }
2971                 if (!rdev || !test_bit(In_sync, &rdev->flags)
2972                     || test_bit(R5_ReadError, &dev->flags)) {
2973                         s.failed++;
2974                         s.failed_num = i;
2975                 } else
2976                         set_bit(R5_Insync, &dev->flags);
2977         }
2978         rcu_read_unlock();
2979
2980         if (unlikely(blocked_rdev)) {
2981                 if (s.syncing || s.expanding || s.expanded ||
2982                     s.to_write || s.written) {
2983                         set_bit(STRIPE_HANDLE, &sh->state);
2984                         goto unlock;
2985                 }
2986                 /* There is nothing for the blocked_rdev to block */
2987                 rdev_dec_pending(blocked_rdev, conf->mddev);
2988                 blocked_rdev = NULL;
2989         }
2990
2991         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
2992                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
2993                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
2994         }
2995
2996         pr_debug("locked=%d uptodate=%d to_read=%d"
2997                 " to_write=%d failed=%d failed_num=%d\n",
2998                 s.locked, s.uptodate, s.to_read, s.to_write,
2999                 s.failed, s.failed_num);
3000         /* check if the array has lost two devices and, if so, some requests might
3001          * need to be failed
3002          */
3003         if (s.failed > 1 && s.to_read+s.to_write+s.written)
3004                 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3005         if (s.failed > 1 && s.syncing) {
3006                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3007                 clear_bit(STRIPE_SYNCING, &sh->state);
3008                 s.syncing = 0;
3009         }
3010
3011         /* might be able to return some write requests if the parity block
3012          * is safe, or on a failed drive
3013          */
3014         dev = &sh->dev[sh->pd_idx];
3015         if ( s.written &&
3016              ((test_bit(R5_Insync, &dev->flags) &&
3017                !test_bit(R5_LOCKED, &dev->flags) &&
3018                test_bit(R5_UPTODATE, &dev->flags)) ||
3019                (s.failed == 1 && s.failed_num == sh->pd_idx)))
3020                 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3021
3022         /* Now we might consider reading some blocks, either to check/generate
3023          * parity, or to satisfy requests
3024          * or to load a block that is being partially written.
3025          */
3026         if (s.to_read || s.non_overwrite ||
3027             (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3028                 handle_stripe_fill5(sh, &s, disks);
3029
3030         /* Now we check to see if any write operations have recently
3031          * completed
3032          */
3033         prexor = 0;
3034         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3035                 prexor = 1;
3036         if (sh->reconstruct_state == reconstruct_state_drain_result ||
3037             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3038                 sh->reconstruct_state = reconstruct_state_idle;
3039
3040                 /* All the 'written' buffers and the parity block are ready to
3041                  * be written back to disk
3042                  */
3043                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3044                 for (i = disks; i--; ) {
3045                         dev = &sh->dev[i];
3046                         if (test_bit(R5_LOCKED, &dev->flags) &&
3047                                 (i == sh->pd_idx || dev->written)) {
3048                                 pr_debug("Writing block %d\n", i);
3049                                 set_bit(R5_Wantwrite, &dev->flags);
3050                                 if (prexor)
3051                                         continue;
3052                                 if (!test_bit(R5_Insync, &dev->flags) ||
3053                                     (i == sh->pd_idx && s.failed == 0))
3054                                         set_bit(STRIPE_INSYNC, &sh->state);
3055                         }
3056                 }
3057                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
3058                         atomic_dec(&conf->preread_active_stripes);
3059                         if (atomic_read(&conf->preread_active_stripes) <
3060                                 IO_THRESHOLD)
3061                                 md_wakeup_thread(conf->mddev->thread);
3062                 }
3063         }
3064
3065         /* Now to consider new write requests and what else, if anything
3066          * should be read.  We do not handle new writes when:
3067          * 1/ A 'write' operation (copy+xor) is already in flight.
3068          * 2/ A 'check' operation is in flight, as it may clobber the parity
3069          *    block.
3070          */
3071         if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3072                 handle_stripe_dirtying5(conf, sh, &s, disks);
3073
3074         /* maybe we need to check and possibly fix the parity for this stripe
3075          * Any reads will already have been scheduled, so we just see if enough
3076          * data is available.  The parity check is held off while parity
3077          * dependent operations are in flight.
3078          */
3079         if (sh->check_state ||
3080             (s.syncing && s.locked == 0 &&
3081              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3082              !test_bit(STRIPE_INSYNC, &sh->state)))
3083                 handle_parity_checks5(conf, sh, &s, disks);
3084
3085         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3086                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3087                 clear_bit(STRIPE_SYNCING, &sh->state);
3088         }
3089
3090         /* If the failed drive is just a ReadError, then we might need to progress
3091          * the repair/check process
3092          */
3093         if (s.failed == 1 && !conf->mddev->ro &&
3094             test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
3095             && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
3096             && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
3097                 ) {
3098                 dev = &sh->dev[s.failed_num];
3099                 if (!test_bit(R5_ReWrite, &dev->flags)) {
3100                         set_bit(R5_Wantwrite, &dev->flags);
3101                         set_bit(R5_ReWrite, &dev->flags);
3102                         set_bit(R5_LOCKED, &dev->flags);
3103                         s.locked++;
3104                 } else {
3105                         /* let's read it back */
3106                         set_bit(R5_Wantread, &dev->flags);
3107                         set_bit(R5_LOCKED, &dev->flags);
3108                         s.locked++;
3109                 }
3110         }
3111
3112         /* Finish reconstruct operations initiated by the expansion process */
3113         if (sh->reconstruct_state == reconstruct_state_result) {
3114                 struct stripe_head *sh2
3115                         = get_active_stripe(conf, sh->sector, 1, 1, 1);
3116                 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3117                         /* sh cannot be written until sh2 has been read.
3118                          * so arrange for sh to be delayed a little
3119                          */
3120                         set_bit(STRIPE_DELAYED, &sh->state);
3121                         set_bit(STRIPE_HANDLE, &sh->state);
3122                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3123                                               &sh2->state))
3124                                 atomic_inc(&conf->preread_active_stripes);
3125                         release_stripe(sh2);
3126                         goto unlock;
3127                 }
3128                 if (sh2)
3129                         release_stripe(sh2);
3130
3131                 sh->reconstruct_state = reconstruct_state_idle;
3132                 clear_bit(STRIPE_EXPANDING, &sh->state);
3133                 for (i = conf->raid_disks; i--; ) {
3134                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
3135                         set_bit(R5_LOCKED, &sh->dev[i].flags);
3136                         s.locked++;
3137                 }
3138         }
3139
3140         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3141             !sh->reconstruct_state) {
3142                 /* Need to write out all blocks after computing parity */
3143                 sh->disks = conf->raid_disks;
3144                 stripe_set_idx(sh->sector, conf, 0, sh);
3145                 schedule_reconstruction(sh, &s, 1, 1);
3146         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3147                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3148                 atomic_dec(&conf->reshape_stripes);
3149                 wake_up(&conf->wait_for_overlap);
3150                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3151         }
3152
3153         if (s.expanding && s.locked == 0 &&
3154             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3155                 handle_stripe_expansion(conf, sh, NULL);
3156
3157  unlock:
3158         spin_unlock(&sh->lock);
3159
3160         /* wait for this device to become unblocked */
3161         if (unlikely(blocked_rdev))
3162                 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3163
3164         if (s.ops_request)
3165                 raid_run_ops(sh, s.ops_request);
3166
3167         ops_run_io(sh, &s);
3168
3169         return_io(return_bi);
3170
3171         return blocked_rdev == NULL;
3172 }
3173
3174 static bool handle_stripe6(struct stripe_head *sh)
3175 {
3176         raid5_conf_t *conf = sh->raid_conf;
3177         int disks = sh->disks;
3178         struct bio *return_bi = NULL;
3179         int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx;
3180         struct stripe_head_state s;
3181         struct r6_state r6s;
3182         struct r5dev *dev, *pdev, *qdev;
3183         mdk_rdev_t *blocked_rdev = NULL;
3184
3185         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3186                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3187                (unsigned long long)sh->sector, sh->state,
3188                atomic_read(&sh->count), pd_idx, qd_idx,
3189                sh->check_state, sh->reconstruct_state);
3190         memset(&s, 0, sizeof(s));
3191
3192         spin_lock(&sh->lock);
3193         clear_bit(STRIPE_HANDLE, &sh->state);
3194         clear_bit(STRIPE_DELAYED, &sh->state);
3195
3196         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3197         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3198         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3199         /* Now to look around and see what can be done */
3200
3201         rcu_read_lock();
3202         for (i=disks; i--; ) {
3203                 mdk_rdev_t *rdev;
3204                 dev = &sh->dev[i];
3205                 clear_bit(R5_Insync, &dev->flags);
3206
3207                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3208                         i, dev->flags, dev->toread, dev->towrite, dev->written);
3209                 /* maybe we can reply to a read
3210                  *
3211                  * new wantfill requests are only permitted while
3212                  * ops_complete_biofill is guaranteed to be inactive
3213                  */
3214                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3215                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3216                         set_bit(R5_Wantfill, &dev->flags);
3217
3218                 /* now count some things */
3219                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3220                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3221                 if (test_bit(R5_Wantcompute, &dev->flags)) {
3222                         s.compute++;
3223                         BUG_ON(s.compute > 2);
3224                 }
3225
3226                 if (test_bit(R5_Wantfill, &dev->flags)) {
3227                         s.to_fill++;
3228                 } else if (dev->toread)
3229                         s.to_read++;
3230                 if (dev->towrite) {
3231                         s.to_write++;
3232                         if (!test_bit(R5_OVERWRITE, &dev->flags))
3233                                 s.non_overwrite++;
3234                 }
3235                 if (dev->written)
3236                         s.written++;
3237                 rdev = rcu_dereference(conf->disks[i].rdev);
3238                 if (blocked_rdev == NULL &&
3239                     rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3240                         blocked_rdev = rdev;
3241                         atomic_inc(&rdev->nr_pending);
3242                 }
3243                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3244                         /* The ReadError flag will just be confusing now */
3245                         clear_bit(R5_ReadError, &dev->flags);
3246                         clear_bit(R5_ReWrite, &dev->flags);
3247                 }
3248                 if (!rdev || !test_bit(In_sync, &rdev->flags)
3249                     || test_bit(R5_ReadError, &dev->flags)) {
3250                         if (s.failed < 2)
3251                                 r6s.failed_num[s.failed] = i;
3252                         s.failed++;
3253                 } else
3254                         set_bit(R5_Insync, &dev->flags);
3255         }
3256         rcu_read_unlock();
3257
3258         if (unlikely(blocked_rdev)) {
3259                 if (s.syncing || s.expanding || s.expanded ||
3260                     s.to_write || s.written) {
3261                         set_bit(STRIPE_HANDLE, &sh->state);
3262                         goto unlock;
3263                 }
3264                 /* There is nothing for the blocked_rdev to block */
3265                 rdev_dec_pending(blocked_rdev, conf->mddev);
3266                 blocked_rdev = NULL;
3267         }
3268
3269         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3270                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3271                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3272         }
3273
3274         pr_debug("locked=%d uptodate=%d to_read=%d"
3275                " to_write=%d failed=%d failed_num=%d,%d\n",
3276                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3277                r6s.failed_num[0], r6s.failed_num[1]);
3278         /* check if the array has lost >2 devices and, if so, some requests
3279          * might need to be failed
3280          */
3281         if (s.failed > 2 && s.to_read+s.to_write+s.written)
3282                 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3283         if (s.failed > 2 && s.syncing) {
3284                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3285                 clear_bit(STRIPE_SYNCING, &sh->state);
3286                 s.syncing = 0;
3287         }
3288
3289         /*
3290          * might be able to return some write requests if the parity blocks
3291          * are safe, or on a failed drive
3292          */
3293         pdev = &sh->dev[pd_idx];
3294         r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3295                 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3296         qdev = &sh->dev[qd_idx];
3297         r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == qd_idx)
3298                 || (s.failed >= 2 && r6s.failed_num[1] == qd_idx);
3299
3300         if ( s.written &&
3301              ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3302                              && !test_bit(R5_LOCKED, &pdev->flags)
3303                              && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3304              ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3305                              && !test_bit(R5_LOCKED, &qdev->flags)
3306                              && test_bit(R5_UPTODATE, &qdev->flags)))))
3307                 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3308
3309         /* Now we might consider reading some blocks, either to check/generate
3310          * parity, or to satisfy requests
3311          * or to load a block that is being partially written.
3312          */
3313         if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3314             (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3315                 handle_stripe_fill6(sh, &s, &r6s, disks);
3316
3317         /* Now we check to see if any write operations have recently
3318          * completed
3319          */
3320         if (sh->reconstruct_state == reconstruct_state_drain_result) {
3321                 int qd_idx = sh->qd_idx;
3322
3323                 sh->reconstruct_state = reconstruct_state_idle;
3324                 /* All the 'written' buffers and the parity blocks are ready to
3325                  * be written back to disk
3326                  */
3327                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3328                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags));
3329                 for (i = disks; i--; ) {
3330                         dev = &sh->dev[i];
3331                         if (test_bit(R5_LOCKED, &dev->flags) &&
3332                             (i == sh->pd_idx || i == qd_idx ||
3333                              dev->written)) {
3334                                 pr_debug("Writing block %d\n", i);
3335                                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3336                                 set_bit(R5_Wantwrite, &dev->flags);
3337                                 if (!test_bit(R5_Insync, &dev->flags) ||
3338                                     ((i == sh->pd_idx || i == qd_idx) &&
3339                                       s.failed == 0))
3340                                         set_bit(STRIPE_INSYNC, &sh->state);
3341                         }
3342                 }
3343                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
3344                         atomic_dec(&conf->preread_active_stripes);
3345                         if (atomic_read(&conf->preread_active_stripes) <
3346                                 IO_THRESHOLD)
3347                                 md_wakeup_thread(conf->mddev->thread);
3348                 }
3349         }
3350
3351         /* Now to consider new write requests and what else, if anything
3352          * should be read.  We do not handle new writes when:
3353          * 1/ A 'write' operation (copy+gen_syndrome) is already in flight.
3354          * 2/ A 'check' operation is in flight, as it may clobber the parity
3355          *    block.
3356          */
3357         if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3358                 handle_stripe_dirtying6(conf, sh, &s, &r6s, disks);
3359
3360         /* maybe we need to check and possibly fix the parity for this stripe
3361          * Any reads will already have been scheduled, so we just see if enough
3362          * data is available.  The parity check is held off while parity
3363          * dependent operations are in flight.
3364          */
3365         if (sh->check_state ||
3366             (s.syncing && s.locked == 0 &&
3367              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3368              !test_bit(STRIPE_INSYNC, &sh->state)))
3369                 handle_parity_checks6(conf, sh, &s, &r6s, disks);
3370
3371         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3372                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3373                 clear_bit(STRIPE_SYNCING, &sh->state);
3374         }
3375
3376         /* If the failed drives are just a ReadError, then we might need
3377          * to progress the repair/check process
3378          */
3379         if (s.failed <= 2 && !conf->mddev->ro)
3380                 for (i = 0; i < s.failed; i++) {
3381                         dev = &sh->dev[r6s.failed_num[i]];
3382                         if (test_bit(R5_ReadError, &dev->flags)
3383                             && !test_bit(R5_LOCKED, &dev->flags)
3384                             && test_bit(R5_UPTODATE, &dev->flags)
3385                                 ) {
3386                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
3387                                         set_bit(R5_Wantwrite, &dev->flags);
3388                                         set_bit(R5_ReWrite, &dev->flags);
3389                                         set_bit(R5_LOCKED, &dev->flags);
3390                                         s.locked++;
3391                                 } else {
3392                                         /* let's read it back */
3393                                         set_bit(R5_Wantread, &dev->flags);
3394                                         set_bit(R5_LOCKED, &dev->flags);
3395                                         s.locked++;
3396                                 }
3397                         }
3398                 }
3399
3400         /* Finish reconstruct operations initiated by the expansion process */
3401         if (sh->reconstruct_state == reconstruct_state_result) {
3402                 sh->reconstruct_state = reconstruct_state_idle;
3403                 clear_bit(STRIPE_EXPANDING, &sh->state);
3404                 for (i = conf->raid_disks; i--; ) {
3405                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
3406                         set_bit(R5_LOCKED, &sh->dev[i].flags);
3407                         s.locked++;
3408                 }
3409         }
3410
3411         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3412             !sh->reconstruct_state) {
3413                 struct stripe_head *sh2
3414                         = get_active_stripe(conf, sh->sector, 1, 1, 1);
3415                 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3416                         /* sh cannot be written until sh2 has been read.
3417                          * so arrange for sh to be delayed a little
3418                          */
3419                         set_bit(STRIPE_DELAYED, &sh->state);
3420                         set_bit(STRIPE_HANDLE, &sh->state);
3421                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3422                                               &sh2->state))
3423                                 atomic_inc(&conf->preread_active_stripes);
3424                         release_stripe(sh2);
3425                         goto unlock;
3426                 }
3427                 if (sh2)
3428                         release_stripe(sh2);
3429
3430                 /* Need to write out all blocks after computing P&Q */
3431                 sh->disks = conf->raid_disks;
3432                 stripe_set_idx(sh->sector, conf, 0, sh);
3433                 schedule_reconstruction(sh, &s, 1, 1);
3434         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3435                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3436                 atomic_dec(&conf->reshape_stripes);
3437                 wake_up(&conf->wait_for_overlap);
3438                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3439         }
3440
3441         if (s.expanding && s.locked == 0 &&
3442             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3443                 handle_stripe_expansion(conf, sh, &r6s);
3444
3445  unlock:
3446         spin_unlock(&sh->lock);
3447
3448         /* wait for this device to become unblocked */
3449         if (unlikely(blocked_rdev))
3450                 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3451
3452         if (s.ops_request)
3453                 raid_run_ops(sh, s.ops_request);
3454
3455         ops_run_io(sh, &s);
3456
3457         return_io(return_bi);
3458
3459         return blocked_rdev == NULL;
3460 }
3461
3462 /* returns true if the stripe was handled */
3463 static bool handle_stripe(struct stripe_head *sh)
3464 {
3465         if (sh->raid_conf->level == 6)
3466                 return handle_stripe6(sh);
3467         else
3468                 return handle_stripe5(sh);
3469 }
3470
3471 static void raid5_activate_delayed(raid5_conf_t *conf)
3472 {
3473         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3474                 while (!list_empty(&conf->delayed_list)) {
3475                         struct list_head *l = conf->delayed_list.next;
3476                         struct stripe_head *sh;
3477                         sh = list_entry(l, struct stripe_head, lru);
3478                         list_del_init(l);
3479                         clear_bit(STRIPE_DELAYED, &sh->state);
3480                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3481                                 atomic_inc(&conf->preread_active_stripes);
3482                         list_add_tail(&sh->lru, &conf->hold_list);
3483                 }
3484         } else
3485                 blk_plug_device(conf->mddev->queue);
3486 }
3487
3488 static void activate_bit_delay(raid5_conf_t *conf)
3489 {
3490         /* device_lock is held */
3491         struct list_head head;
3492         list_add(&head, &conf->bitmap_list);
3493         list_del_init(&conf->bitmap_list);
3494         while (!list_empty(&head)) {
3495                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3496                 list_del_init(&sh->lru);
3497                 atomic_inc(&sh->count);
3498                 __release_stripe(conf, sh);
3499         }
3500 }
3501
3502 static void unplug_slaves(mddev_t *mddev)
3503 {
3504         raid5_conf_t *conf = mddev->private;
3505         int i;
3506
3507         rcu_read_lock();
3508         for (i = 0; i < conf->raid_disks; i++) {
3509                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3510                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3511                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3512
3513                         atomic_inc(&rdev->nr_pending);
3514                         rcu_read_unlock();
3515
3516                         blk_unplug(r_queue);
3517
3518                         rdev_dec_pending(rdev, mddev);
3519                         rcu_read_lock();
3520                 }
3521         }
3522         rcu_read_unlock();
3523 }
3524
3525 static void raid5_unplug_device(struct request_queue *q)
3526 {
3527         mddev_t *mddev = q->queuedata;
3528         raid5_conf_t *conf = mddev->private;
3529         unsigned long flags;
3530
3531         spin_lock_irqsave(&conf->device_lock, flags);
3532
3533         if (blk_remove_plug(q)) {
3534                 conf->seq_flush++;
3535                 raid5_activate_delayed(conf);
3536         }
3537         md_wakeup_thread(mddev->thread);
3538
3539         spin_unlock_irqrestore(&conf->device_lock, flags);
3540
3541         unplug_slaves(mddev);
3542 }
3543
3544 static int raid5_congested(void *data, int bits)
3545 {
3546         mddev_t *mddev = data;
3547         raid5_conf_t *conf = mddev->private;
3548
3549         /* No difference between reads and writes.  Just check
3550          * how busy the stripe_cache is
3551          */
3552
3553         if (mddev_congested(mddev, bits))
3554                 return 1;
3555         if (conf->inactive_blocked)
3556                 return 1;
3557         if (conf->quiesce)
3558                 return 1;
3559         if (list_empty_careful(&conf->inactive_list))
3560                 return 1;
3561
3562         return 0;
3563 }
3564
3565 /* We want read requests to align with chunks where possible,
3566  * but write requests don't need to.
3567  */
3568 static int raid5_mergeable_bvec(struct request_queue *q,
3569                                 struct bvec_merge_data *bvm,
3570                                 struct bio_vec *biovec)
3571 {
3572         mddev_t *mddev = q->queuedata;
3573         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3574         int max;
3575         unsigned int chunk_sectors = mddev->chunk_sectors;
3576         unsigned int bio_sectors = bvm->bi_size >> 9;
3577
3578         if ((bvm->bi_rw & 1) == WRITE)
3579                 return biovec->bv_len; /* always allow writes to be mergeable */
3580
3581         if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3582                 chunk_sectors = mddev->new_chunk_sectors;
3583         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3584         if (max < 0) max = 0;
3585         if (max <= biovec->bv_len && bio_sectors == 0)
3586                 return biovec->bv_len;
3587         else
3588                 return max;
3589 }
3590
3591
3592 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3593 {
3594         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3595         unsigned int chunk_sectors = mddev->chunk_sectors;
3596         unsigned int bio_sectors = bio->bi_size >> 9;
3597
3598         if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3599                 chunk_sectors = mddev->new_chunk_sectors;
3600         return  chunk_sectors >=
3601                 ((sector & (chunk_sectors - 1)) + bio_sectors);
3602 }
3603
3604 /*
3605  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3606  *  later sampled by raid5d.
3607  */
3608 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3609 {
3610         unsigned long flags;
3611
3612         spin_lock_irqsave(&conf->device_lock, flags);
3613
3614         bi->bi_next = conf->retry_read_aligned_list;
3615         conf->retry_read_aligned_list = bi;
3616
3617         spin_unlock_irqrestore(&conf->device_lock, flags);
3618         md_wakeup_thread(conf->mddev->thread);
3619 }
3620
3621
3622 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3623 {
3624         struct bio *bi;
3625
3626         bi = conf->retry_read_aligned;
3627         if (bi) {
3628                 conf->retry_read_aligned = NULL;
3629                 return bi;
3630         }
3631         bi = conf->retry_read_aligned_list;
3632         if(bi) {
3633                 conf->retry_read_aligned_list = bi->bi_next;
3634                 bi->bi_next = NULL;
3635                 /*
3636                  * this sets the active strip count to 1 and the processed
3637                  * strip count to zero (upper 8 bits)
3638                  */
3639                 bi->bi_phys_segments = 1; /* biased count of active stripes */
3640         }
3641
3642         return bi;
3643 }
3644
3645
3646 /*
3647  *  The "raid5_align_endio" should check if the read succeeded and if it
3648  *  did, call bio_endio on the original bio (having bio_put the new bio
3649  *  first).
3650  *  If the read failed..
3651  */
3652 static void raid5_align_endio(struct bio *bi, int error)
3653 {
3654         struct bio* raid_bi  = bi->bi_private;
3655         mddev_t *mddev;
3656         raid5_conf_t *conf;
3657         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3658         mdk_rdev_t *rdev;
3659
3660         bio_put(bi);
3661
3662         mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3663         conf = mddev->private;
3664         rdev = (void*)raid_bi->bi_next;
3665         raid_bi->bi_next = NULL;
3666
3667         rdev_dec_pending(rdev, conf->mddev);
3668
3669         if (!error && uptodate) {
3670                 bio_endio(raid_bi, 0);
3671                 if (atomic_dec_and_test(&conf->active_aligned_reads))
3672                         wake_up(&conf->wait_for_stripe);
3673                 return;
3674         }
3675
3676
3677         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3678
3679         add_bio_to_retry(raid_bi, conf);
3680 }
3681
3682 static int bio_fits_rdev(struct bio *bi)
3683 {
3684         struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3685
3686         if ((bi->bi_size>>9) > queue_max_sectors(q))
3687                 return 0;
3688         blk_recount_segments(q, bi);
3689         if (bi->bi_phys_segments > queue_max_phys_segments(q))
3690                 return 0;
3691
3692         if (q->merge_bvec_fn)
3693                 /* it's too hard to apply the merge_bvec_fn at this stage,
3694                  * just just give up
3695                  */
3696                 return 0;
3697
3698         return 1;
3699 }
3700
3701
3702 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3703 {
3704         mddev_t *mddev = q->queuedata;
3705         raid5_conf_t *conf = mddev->private;
3706         unsigned int dd_idx;
3707         struct bio* align_bi;
3708         mdk_rdev_t *rdev;
3709
3710         if (!in_chunk_boundary(mddev, raid_bio)) {
3711                 pr_debug("chunk_aligned_read : non aligned\n");
3712                 return 0;
3713         }
3714         /*
3715          * use bio_clone to make a copy of the bio
3716          */
3717         align_bi = bio_clone(raid_bio, GFP_NOIO);
3718         if (!align_bi)
3719                 return 0;
3720         /*
3721          *   set bi_end_io to a new function, and set bi_private to the
3722          *     original bio.
3723          */
3724         align_bi->bi_end_io  = raid5_align_endio;
3725         align_bi->bi_private = raid_bio;
3726         /*
3727          *      compute position
3728          */
3729         align_bi->bi_sector =  raid5_compute_sector(conf, raid_bio->bi_sector,
3730                                                     0,
3731                                                     &dd_idx, NULL);
3732
3733         rcu_read_lock();
3734         rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3735         if (rdev && test_bit(In_sync, &rdev->flags)) {
3736                 atomic_inc(&rdev->nr_pending);
3737                 rcu_read_unlock();
3738                 raid_bio->bi_next = (void*)rdev;
3739                 align_bi->bi_bdev =  rdev->bdev;
3740                 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3741                 align_bi->bi_sector += rdev->data_offset;
3742
3743                 if (!bio_fits_rdev(align_bi)) {
3744                         /* too big in some way */
3745                         bio_put(align_bi);
3746                         rdev_dec_pending(rdev, mddev);
3747                         return 0;
3748                 }
3749
3750                 spin_lock_irq(&conf->device_lock);
3751                 wait_event_lock_irq(conf->wait_for_stripe,
3752                                     conf->quiesce == 0,
3753                                     conf->device_lock, /* nothing */);
3754                 atomic_inc(&conf->active_aligned_reads);
3755                 spin_unlock_irq(&conf->device_lock);
3756
3757                 generic_make_request(align_bi);
3758                 return 1;
3759         } else {
3760                 rcu_read_unlock();
3761                 bio_put(align_bi);
3762                 return 0;
3763         }
3764 }
3765
3766 /* __get_priority_stripe - get the next stripe to process
3767  *
3768  * Full stripe writes are allowed to pass preread active stripes up until
3769  * the bypass_threshold is exceeded.  In general the bypass_count
3770  * increments when the handle_list is handled before the hold_list; however, it
3771  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3772  * stripe with in flight i/o.  The bypass_count will be reset when the
3773  * head of the hold_list has changed, i.e. the head was promoted to the
3774  * handle_list.
3775  */
3776 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3777 {
3778         struct stripe_head *sh;
3779
3780         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3781                   __func__,
3782                   list_empty(&conf->handle_list) ? "empty" : "busy",
3783                   list_empty(&conf->hold_list) ? "empty" : "busy",
3784                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
3785
3786         if (!list_empty(&conf->handle_list)) {
3787                 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3788
3789                 if (list_empty(&conf->hold_list))
3790                         conf->bypass_count = 0;
3791                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3792                         if (conf->hold_list.next == conf->last_hold)
3793                                 conf->bypass_count++;
3794                         else {
3795                                 conf->last_hold = conf->hold_list.next;
3796                                 conf->bypass_count -= conf->bypass_threshold;
3797                                 if (conf->bypass_count < 0)
3798                                         conf->bypass_count = 0;
3799                         }
3800                 }
3801         } else if (!list_empty(&conf->hold_list) &&
3802                    ((conf->bypass_threshold &&
3803                      conf->bypass_count > conf->bypass_threshold) ||
3804                     atomic_read(&conf->pending_full_writes) == 0)) {
3805                 sh = list_entry(conf->hold_list.next,
3806                                 typeof(*sh), lru);
3807                 conf->bypass_count -= conf->bypass_threshold;
3808                 if (conf->bypass_count < 0)
3809                         conf->bypass_count = 0;
3810         } else
3811                 return NULL;
3812
3813         list_del_init(&sh->lru);
3814         atomic_inc(&sh->count);
3815         BUG_ON(atomic_read(&sh->count) != 1);
3816         return sh;
3817 }
3818
3819 static int make_request(struct request_queue *q, struct bio * bi)
3820 {
3821         mddev_t *mddev = q->queuedata;
3822         raid5_conf_t *conf = mddev->private;
3823         int dd_idx;
3824         sector_t new_sector;
3825         sector_t logical_sector, last_sector;
3826         struct stripe_head *sh;
3827         const int rw = bio_data_dir(bi);
3828         int cpu, remaining;
3829
3830         if (unlikely(bio_rw_flagged(bi, BIO_RW_BARRIER))) {
3831                 bio_endio(bi, -EOPNOTSUPP);
3832                 return 0;
3833         }
3834
3835         md_write_start(mddev, bi);
3836
3837         cpu = part_stat_lock();
3838         part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
3839         part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
3840                       bio_sectors(bi));
3841         part_stat_unlock();
3842
3843         if (rw == READ &&
3844              mddev->reshape_position == MaxSector &&
3845              chunk_aligned_read(q,bi))
3846                 return 0;
3847
3848         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3849         last_sector = bi->bi_sector + (bi->bi_size>>9);
3850         bi->bi_next = NULL;
3851         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
3852
3853         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3854                 DEFINE_WAIT(w);
3855                 int disks, data_disks;
3856                 int previous;
3857
3858         retry:
3859                 previous = 0;
3860                 disks = conf->raid_disks;
3861                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3862                 if (unlikely(conf->reshape_progress != MaxSector)) {
3863                         /* spinlock is needed as reshape_progress may be
3864                          * 64bit on a 32bit platform, and so it might be
3865                          * possible to see a half-updated value
3866                          * Ofcourse reshape_progress could change after
3867                          * the lock is dropped, so once we get a reference
3868                          * to the stripe that we think it is, we will have
3869                          * to check again.
3870                          */
3871                         spin_lock_irq(&conf->device_lock);
3872                         if (mddev->delta_disks < 0
3873                             ? logical_sector < conf->reshape_progress
3874                             : logical_sector >= conf->reshape_progress) {
3875                                 disks = conf->previous_raid_disks;
3876                                 previous = 1;
3877                         } else {
3878                                 if (mddev->delta_disks < 0
3879                                     ? logical_sector < conf->reshape_safe
3880                                     : logical_sector >= conf->reshape_safe) {
3881                                         spin_unlock_irq(&conf->device_lock);
3882                                         schedule();
3883                                         goto retry;
3884                                 }
3885                         }
3886                         spin_unlock_irq(&conf->device_lock);
3887                 }
3888                 data_disks = disks - conf->max_degraded;
3889
3890                 new_sector = raid5_compute_sector(conf, logical_sector,
3891                                                   previous,
3892                                                   &dd_idx, NULL);
3893                 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3894                         (unsigned long long)new_sector, 
3895                         (unsigned long long)logical_sector);
3896
3897                 sh = get_active_stripe(conf, new_sector, previous,
3898                                        (bi->bi_rw&RWA_MASK), 0);
3899                 if (sh) {
3900                         if (unlikely(previous)) {
3901                                 /* expansion might have moved on while waiting for a
3902                                  * stripe, so we must do the range check again.
3903                                  * Expansion could still move past after this
3904                                  * test, but as we are holding a reference to
3905                                  * 'sh', we know that if that happens,
3906                                  *  STRIPE_EXPANDING will get set and the expansion
3907                                  * won't proceed until we finish with the stripe.
3908                                  */
3909                                 int must_retry = 0;
3910                                 spin_lock_irq(&conf->device_lock);
3911                                 if (mddev->delta_disks < 0
3912                                     ? logical_sector >= conf->reshape_progress
3913                                     : logical_sector < conf->reshape_progress)
3914                                         /* mismatch, need to try again */
3915                                         must_retry = 1;
3916                                 spin_unlock_irq(&conf->device_lock);
3917                                 if (must_retry) {
3918                                         release_stripe(sh);
3919                                         schedule();
3920                                         goto retry;
3921                                 }
3922                         }
3923
3924                         if (bio_data_dir(bi) == WRITE &&
3925                             logical_sector >= mddev->suspend_lo &&
3926                             logical_sector < mddev->suspend_hi) {
3927                                 release_stripe(sh);
3928                                 /* As the suspend_* range is controlled by
3929                                  * userspace, we want an interruptible
3930                                  * wait.
3931                                  */
3932                                 flush_signals(current);
3933                                 prepare_to_wait(&conf->wait_for_overlap,
3934                                                 &w, TASK_INTERRUPTIBLE);
3935                                 if (logical_sector >= mddev->suspend_lo &&
3936                                     logical_sector < mddev->suspend_hi)
3937                                         schedule();
3938                                 goto retry;
3939                         }
3940
3941                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3942                             !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3943                                 /* Stripe is busy expanding or
3944                                  * add failed due to overlap.  Flush everything
3945                                  * and wait a while
3946                                  */
3947                                 raid5_unplug_device(mddev->queue);
3948                                 release_stripe(sh);
3949                                 schedule();
3950                                 goto retry;
3951                         }
3952                         finish_wait(&conf->wait_for_overlap, &w);
3953                         set_bit(STRIPE_HANDLE, &sh->state);
3954                         clear_bit(STRIPE_DELAYED, &sh->state);
3955                         release_stripe(sh);
3956                 } else {
3957                         /* cannot get stripe for read-ahead, just give-up */
3958                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
3959                         finish_wait(&conf->wait_for_overlap, &w);
3960                         break;
3961                 }
3962                         
3963         }
3964         spin_lock_irq(&conf->device_lock);
3965         remaining = raid5_dec_bi_phys_segments(bi);
3966         spin_unlock_irq(&conf->device_lock);
3967         if (remaining == 0) {
3968
3969                 if ( rw == WRITE )
3970                         md_write_end(mddev);
3971
3972                 bio_endio(bi, 0);
3973         }
3974         return 0;
3975 }
3976
3977 static sector_t raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks);
3978
3979 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3980 {
3981         /* reshaping is quite different to recovery/resync so it is
3982          * handled quite separately ... here.
3983          *
3984          * On each call to sync_request, we gather one chunk worth of
3985          * destination stripes and flag them as expanding.
3986          * Then we find all the source stripes and request reads.
3987          * As the reads complete, handle_stripe will copy the data
3988          * into the destination stripe and release that stripe.
3989          */
3990         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3991         struct stripe_head *sh;
3992         sector_t first_sector, last_sector;
3993         int raid_disks = conf->previous_raid_disks;
3994         int data_disks = raid_disks - conf->max_degraded;
3995         int new_data_disks = conf->raid_disks - conf->max_degraded;
3996         int i;
3997         int dd_idx;
3998         sector_t writepos, readpos, safepos;
3999         sector_t stripe_addr;
4000         int reshape_sectors;
4001         struct list_head stripes;
4002
4003         if (sector_nr == 0) {
4004                 /* If restarting in the middle, skip the initial sectors */
4005                 if (mddev->delta_disks < 0 &&
4006                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4007                         sector_nr = raid5_size(mddev, 0, 0)
4008                                 - conf->reshape_progress;
4009                 } else if (mddev->delta_disks >= 0 &&
4010                            conf->reshape_progress > 0)
4011                         sector_nr = conf->reshape_progress;
4012                 sector_div(sector_nr, new_data_disks);
4013                 if (sector_nr) {
4014                         *skipped = 1;
4015                         return sector_nr;
4016                 }
4017         }
4018
4019         /* We need to process a full chunk at a time.
4020          * If old and new chunk sizes differ, we need to process the
4021          * largest of these
4022          */
4023         if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4024                 reshape_sectors = mddev->new_chunk_sectors;
4025         else
4026                 reshape_sectors = mddev->chunk_sectors;
4027
4028         /* we update the metadata when there is more than 3Meg
4029          * in the block range (that is rather arbitrary, should
4030          * probably be time based) or when the data about to be
4031          * copied would over-write the source of the data at
4032          * the front of the range.
4033          * i.e. one new_stripe along from reshape_progress new_maps
4034          * to after where reshape_safe old_maps to
4035          */
4036         writepos = conf->reshape_progress;
4037         sector_div(writepos, new_data_disks);
4038         readpos = conf->reshape_progress;
4039         sector_div(readpos, data_disks);
4040         safepos = conf->reshape_safe;
4041         sector_div(safepos, data_disks);
4042         if (mddev->delta_disks < 0) {
4043                 writepos -= min_t(sector_t, reshape_sectors, writepos);
4044                 readpos += reshape_sectors;
4045                 safepos += reshape_sectors;
4046         } else {
4047                 writepos += reshape_sectors;
4048                 readpos -= min_t(sector_t, reshape_sectors, readpos);
4049                 safepos -= min_t(sector_t, reshape_sectors, safepos);
4050         }
4051
4052         /* 'writepos' is the most advanced device address we might write.
4053          * 'readpos' is the least advanced device address we might read.
4054          * 'safepos' is the least address recorded in the metadata as having
4055          *     been reshaped.
4056          * If 'readpos' is behind 'writepos', then there is no way that we can
4057          * ensure safety in the face of a crash - that must be done by userspace
4058          * making a backup of the data.  So in that case there is no particular
4059          * rush to update metadata.
4060          * Otherwise if 'safepos' is behind 'writepos', then we really need to
4061          * update the metadata to advance 'safepos' to match 'readpos' so that
4062          * we can be safe in the event of a crash.
4063          * So we insist on updating metadata if safepos is behind writepos and
4064          * readpos is beyond writepos.
4065          * In any case, update the metadata every 10 seconds.
4066          * Maybe that number should be configurable, but I'm not sure it is
4067          * worth it.... maybe it could be a multiple of safemode_delay???
4068          */
4069         if ((mddev->delta_disks < 0
4070              ? (safepos > writepos && readpos < writepos)
4071              : (safepos < writepos && readpos > writepos)) ||
4072             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4073                 /* Cannot proceed until we've updated the superblock... */
4074                 wait_event(conf->wait_for_overlap,
4075                            atomic_read(&conf->reshape_stripes)==0);
4076                 mddev->reshape_position = conf->reshape_progress;
4077                 mddev->curr_resync_completed = mddev->curr_resync;
4078                 conf->reshape_checkpoint = jiffies;
4079                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4080                 md_wakeup_thread(mddev->thread);
4081                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4082                            kthread_should_stop());
4083                 spin_lock_irq(&conf->device_lock);
4084                 conf->reshape_safe = mddev->reshape_position;
4085                 spin_unlock_irq(&conf->device_lock);
4086                 wake_up(&conf->wait_for_overlap);
4087                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4088         }
4089
4090         if (mddev->delta_disks < 0) {
4091                 BUG_ON(conf->reshape_progress == 0);
4092                 stripe_addr = writepos;
4093                 BUG_ON((mddev->dev_sectors &
4094                         ~((sector_t)reshape_sectors - 1))
4095                        - reshape_sectors - stripe_addr
4096                        != sector_nr);
4097         } else {
4098                 BUG_ON(writepos != sector_nr + reshape_sectors);
4099                 stripe_addr = sector_nr;
4100         }
4101         INIT_LIST_HEAD(&stripes);
4102         for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4103                 int j;
4104                 int skipped_disk = 0;
4105                 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4106                 set_bit(STRIPE_EXPANDING, &sh->state);
4107                 atomic_inc(&conf->reshape_stripes);
4108                 /* If any of this stripe is beyond the end of the old
4109                  * array, then we need to zero those blocks
4110                  */
4111                 for (j=sh->disks; j--;) {
4112                         sector_t s;
4113                         if (j == sh->pd_idx)
4114                                 continue;
4115                         if (conf->level == 6 &&
4116                             j == sh->qd_idx)
4117                                 continue;
4118                         s = compute_blocknr(sh, j, 0);
4119                         if (s < raid5_size(mddev, 0, 0)) {
4120                                 skipped_disk = 1;
4121                                 continue;
4122                         }
4123                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4124                         set_bit(R5_Expanded, &sh->dev[j].flags);
4125                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
4126                 }
4127                 if (!skipped_disk) {
4128                         set_bit(STRIPE_EXPAND_READY, &sh->state);
4129                         set_bit(STRIPE_HANDLE, &sh->state);
4130                 }
4131                 list_add(&sh->lru, &stripes);
4132         }
4133         spin_lock_irq(&conf->device_lock);
4134         if (mddev->delta_disks < 0)
4135                 conf->reshape_progress -= reshape_sectors * new_data_disks;
4136         else
4137                 conf->reshape_progress += reshape_sectors * new_data_disks;
4138         spin_unlock_irq(&conf->device_lock);
4139         /* Ok, those stripe are ready. We can start scheduling
4140          * reads on the source stripes.
4141          * The source stripes are determined by mapping the first and last
4142          * block on the destination stripes.
4143          */
4144         first_sector =
4145                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4146                                      1, &dd_idx, NULL);
4147         last_sector =
4148                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4149                                             * new_data_disks - 1),
4150                                      1, &dd_idx, NULL);
4151         if (last_sector >= mddev->dev_sectors)
4152                 last_sector = mddev->dev_sectors - 1;
4153         while (first_sector <= last_sector) {
4154                 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4155                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4156                 set_bit(STRIPE_HANDLE, &sh->state);
4157                 release_stripe(sh);
4158                 first_sector += STRIPE_SECTORS;
4159         }
4160         /* Now that the sources are clearly marked, we can release
4161          * the destination stripes
4162          */
4163         while (!list_empty(&stripes)) {
4164                 sh = list_entry(stripes.next, struct stripe_head, lru);
4165                 list_del_init(&sh->lru);
4166                 release_stripe(sh);
4167         }
4168         /* If this takes us to the resync_max point where we have to pause,
4169          * then we need to write out the superblock.
4170          */
4171         sector_nr += reshape_sectors;
4172         if ((sector_nr - mddev->curr_resync_completed) * 2
4173             >= mddev->resync_max - mddev->curr_resync_completed) {
4174                 /* Cannot proceed until we've updated the superblock... */
4175                 wait_event(conf->wait_for_overlap,
4176                            atomic_read(&conf->reshape_stripes) == 0);
4177                 mddev->reshape_position = conf->reshape_progress;
4178                 mddev->curr_resync_completed = mddev->curr_resync + reshape_sectors;
4179                 conf->reshape_checkpoint = jiffies;
4180                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4181                 md_wakeup_thread(mddev->thread);
4182                 wait_event(mddev->sb_wait,
4183                            !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4184                            || kthread_should_stop());
4185                 spin_lock_irq(&conf->device_lock);
4186                 conf->reshape_safe = mddev->reshape_position;
4187                 spin_unlock_irq(&conf->device_lock);
4188                 wake_up(&conf->wait_for_overlap);
4189                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4190         }
4191         return reshape_sectors;
4192 }
4193
4194 /* FIXME go_faster isn't used */
4195 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
4196 {
4197         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4198         struct stripe_head *sh;
4199         sector_t max_sector = mddev->dev_sectors;
4200         int sync_blocks;
4201         int still_degraded = 0;
4202         int i;
4203
4204         if (sector_nr >= max_sector) {
4205                 /* just being told to finish up .. nothing much to do */
4206                 unplug_slaves(mddev);
4207
4208                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4209                         end_reshape(conf);
4210                         return 0;
4211                 }
4212
4213                 if (mddev->curr_resync < max_sector) /* aborted */
4214                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4215                                         &sync_blocks, 1);
4216                 else /* completed sync */
4217                         conf->fullsync = 0;
4218                 bitmap_close_sync(mddev->bitmap);
4219
4220                 return 0;
4221         }
4222
4223         /* Allow raid5_quiesce to complete */
4224         wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4225
4226         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4227                 return reshape_request(mddev, sector_nr, skipped);
4228
4229         /* No need to check resync_max as we never do more than one
4230          * stripe, and as resync_max will always be on a chunk boundary,
4231          * if the check in md_do_sync didn't fire, there is no chance
4232          * of overstepping resync_max here
4233          */
4234
4235         /* if there is too many failed drives and we are trying
4236          * to resync, then assert that we are finished, because there is
4237          * nothing we can do.
4238          */
4239         if (mddev->degraded >= conf->max_degraded &&
4240             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4241                 sector_t rv = mddev->dev_sectors - sector_nr;
4242                 *skipped = 1;
4243                 return rv;
4244         }
4245         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4246             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4247             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4248                 /* we can skip this block, and probably more */
4249                 sync_blocks /= STRIPE_SECTORS;
4250                 *skipped = 1;
4251                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4252         }
4253
4254
4255         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4256
4257         sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4258         if (sh == NULL) {
4259                 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4260                 /* make sure we don't swamp the stripe cache if someone else
4261                  * is trying to get access
4262                  */
4263                 schedule_timeout_uninterruptible(1);
4264         }
4265         /* Need to check if array will still be degraded after recovery/resync
4266          * We don't need to check the 'failed' flag as when that gets set,
4267          * recovery aborts.
4268          */
4269         for (i = 0; i < conf->raid_disks; i++)
4270                 if (conf->disks[i].rdev == NULL)
4271                         still_degraded = 1;
4272
4273         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4274
4275         spin_lock(&sh->lock);
4276         set_bit(STRIPE_SYNCING, &sh->state);
4277         clear_bit(STRIPE_INSYNC, &sh->state);
4278         spin_unlock(&sh->lock);
4279
4280         /* wait for any blocked device to be handled */
4281         while (unlikely(!handle_stripe(sh)))
4282                 ;
4283         release_stripe(sh);
4284
4285         return STRIPE_SECTORS;
4286 }
4287
4288 static int  retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
4289 {
4290         /* We may not be able to submit a whole bio at once as there
4291          * may not be enough stripe_heads available.
4292          * We cannot pre-allocate enough stripe_heads as we may need
4293          * more than exist in the cache (if we allow ever large chunks).
4294          * So we do one stripe head at a time and record in
4295          * ->bi_hw_segments how many have been done.
4296          *
4297          * We *know* that this entire raid_bio is in one chunk, so
4298          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4299          */
4300         struct stripe_head *sh;
4301         int dd_idx;
4302         sector_t sector, logical_sector, last_sector;
4303         int scnt = 0;
4304         int remaining;
4305         int handled = 0;
4306
4307         logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4308         sector = raid5_compute_sector(conf, logical_sector,
4309                                       0, &dd_idx, NULL);
4310         last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4311
4312         for (; logical_sector < last_sector;
4313              logical_sector += STRIPE_SECTORS,
4314                      sector += STRIPE_SECTORS,
4315                      scnt++) {
4316
4317                 if (scnt < raid5_bi_hw_segments(raid_bio))
4318                         /* already done this stripe */
4319                         continue;
4320
4321                 sh = get_active_stripe(conf, sector, 0, 1, 0);
4322
4323                 if (!sh) {
4324                         /* failed to get a stripe - must wait */
4325                         raid5_set_bi_hw_segments(raid_bio, scnt);
4326                         conf->retry_read_aligned = raid_bio;
4327                         return handled;
4328                 }
4329
4330                 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4331                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4332                         release_stripe(sh);
4333                         raid5_set_bi_hw_segments(raid_bio, scnt);
4334                         conf->retry_read_aligned = raid_bio;
4335                         return handled;
4336                 }
4337
4338                 handle_stripe(sh);
4339                 release_stripe(sh);
4340                 handled++;
4341         }
4342         spin_lock_irq(&conf->device_lock);
4343         remaining = raid5_dec_bi_phys_segments(raid_bio);
4344         spin_unlock_irq(&conf->device_lock);
4345         if (remaining == 0)
4346                 bio_endio(raid_bio, 0);
4347         if (atomic_dec_and_test(&conf->active_aligned_reads))
4348                 wake_up(&conf->wait_for_stripe);
4349         return handled;
4350 }
4351
4352 #ifdef CONFIG_MULTICORE_RAID456
4353 static void __process_stripe(void *param, async_cookie_t cookie)
4354 {
4355         struct stripe_head *sh = param;
4356
4357         handle_stripe(sh);
4358         release_stripe(sh);
4359 }
4360
4361 static void process_stripe(struct stripe_head *sh, struct list_head *domain)
4362 {
4363         async_schedule_domain(__process_stripe, sh, domain);
4364 }
4365
4366 static void synchronize_stripe_processing(struct list_head *domain)
4367 {
4368         async_synchronize_full_domain(domain);
4369 }
4370 #else
4371 static void process_stripe(struct stripe_head *sh, struct list_head *domain)
4372 {
4373         handle_stripe(sh);
4374         release_stripe(sh);
4375         cond_resched();
4376 }
4377
4378 static void synchronize_stripe_processing(struct list_head *domain)
4379 {
4380 }
4381 #endif
4382
4383
4384 /*
4385  * This is our raid5 kernel thread.
4386  *
4387  * We scan the hash table for stripes which can be handled now.
4388  * During the scan, completed stripes are saved for us by the interrupt
4389  * handler, so that they will not have to wait for our next wakeup.
4390  */
4391 static void raid5d(mddev_t *mddev)
4392 {
4393         struct stripe_head *sh;
4394         raid5_conf_t *conf = mddev->private;
4395         int handled;
4396         LIST_HEAD(raid_domain);
4397
4398         pr_debug("+++ raid5d active\n");
4399
4400         md_check_recovery(mddev);
4401
4402         handled = 0;
4403         spin_lock_irq(&conf->device_lock);
4404         while (1) {
4405                 struct bio *bio;
4406
4407                 if (conf->seq_flush != conf->seq_write) {
4408                         int seq = conf->seq_flush;
4409                         spin_unlock_irq(&conf->device_lock);
4410                         bitmap_unplug(mddev->bitmap);
4411                         spin_lock_irq(&conf->device_lock);
4412                         conf->seq_write = seq;
4413                         activate_bit_delay(conf);
4414                 }
4415
4416                 while ((bio = remove_bio_from_retry(conf))) {
4417                         int ok;
4418                         spin_unlock_irq(&conf->device_lock);
4419                         ok = retry_aligned_read(conf, bio);
4420                         spin_lock_irq(&conf->device_lock);
4421                         if (!ok)
4422                                 break;
4423                         handled++;
4424                 }
4425
4426                 sh = __get_priority_stripe(conf);
4427
4428                 if (!sh)
4429                         break;
4430                 spin_unlock_irq(&conf->device_lock);
4431                 
4432                 handled++;
4433                 process_stripe(sh, &raid_domain);
4434
4435                 spin_lock_irq(&conf->device_lock);
4436         }
4437         pr_debug("%d stripes handled\n", handled);
4438
4439         spin_unlock_irq(&conf->device_lock);
4440
4441         synchronize_stripe_processing(&raid_domain);
4442         async_tx_issue_pending_all();
4443         unplug_slaves(mddev);
4444
4445         pr_debug("--- raid5d inactive\n");
4446 }
4447
4448 static ssize_t
4449 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4450 {
4451         raid5_conf_t *conf = mddev->private;
4452         if (conf)
4453                 return sprintf(page, "%d\n", conf->max_nr_stripes);
4454         else
4455                 return 0;
4456 }
4457
4458 static ssize_t
4459 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4460 {
4461         raid5_conf_t *conf = mddev->private;
4462         unsigned long new;
4463         int err;
4464
4465         if (len >= PAGE_SIZE)
4466                 return -EINVAL;
4467         if (!conf)
4468                 return -ENODEV;
4469
4470         if (strict_strtoul(page, 10, &new))
4471                 return -EINVAL;
4472         if (new <= 16 || new > 32768)
4473                 return -EINVAL;
4474         while (new < conf->max_nr_stripes) {
4475                 if (drop_one_stripe(conf))
4476                         conf->max_nr_stripes--;
4477                 else
4478                         break;
4479         }
4480         err = md_allow_write(mddev);
4481         if (err)
4482                 return err;
4483         while (new > conf->max_nr_stripes) {
4484                 if (grow_one_stripe(conf))
4485                         conf->max_nr_stripes++;
4486                 else break;
4487         }
4488         return len;
4489 }
4490
4491 static struct md_sysfs_entry
4492 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4493                                 raid5_show_stripe_cache_size,
4494                                 raid5_store_stripe_cache_size);
4495
4496 static ssize_t
4497 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4498 {
4499         raid5_conf_t *conf = mddev->private;
4500         if (conf)
4501                 return sprintf(page, "%d\n", conf->bypass_threshold);
4502         else
4503                 return 0;
4504 }
4505
4506 static ssize_t
4507 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4508 {
4509         raid5_conf_t *conf = mddev->private;
4510         unsigned long new;
4511         if (len >= PAGE_SIZE)
4512                 return -EINVAL;
4513         if (!conf)
4514                 return -ENODEV;
4515
4516         if (strict_strtoul(page, 10, &new))
4517                 return -EINVAL;
4518         if (new > conf->max_nr_stripes)
4519                 return -EINVAL;
4520         conf->bypass_threshold = new;
4521         return len;
4522 }
4523
4524 static struct md_sysfs_entry
4525 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4526                                         S_IRUGO | S_IWUSR,
4527                                         raid5_show_preread_threshold,
4528                                         raid5_store_preread_threshold);
4529
4530 static ssize_t
4531 stripe_cache_active_show(mddev_t *mddev, char *page)
4532 {
4533         raid5_conf_t *conf = mddev->private;
4534         if (conf)
4535                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4536         else
4537                 return 0;
4538 }
4539
4540 static struct md_sysfs_entry
4541 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4542
4543 static struct attribute *raid5_attrs[] =  {
4544         &raid5_stripecache_size.attr,
4545         &raid5_stripecache_active.attr,
4546         &raid5_preread_bypass_threshold.attr,
4547         NULL,
4548 };
4549 static struct attribute_group raid5_attrs_group = {
4550         .name = NULL,
4551         .attrs = raid5_attrs,
4552 };
4553
4554 static sector_t
4555 raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks)
4556 {
4557         raid5_conf_t *conf = mddev->private;
4558
4559         if (!sectors)
4560                 sectors = mddev->dev_sectors;
4561         if (!raid_disks) {
4562                 /* size is defined by the smallest of previous and new size */
4563                 if (conf->raid_disks < conf->previous_raid_disks)
4564                         raid_disks = conf->raid_disks;
4565                 else
4566                         raid_disks = conf->previous_raid_disks;
4567         }
4568
4569         sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4570         sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4571         return sectors * (raid_disks - conf->max_degraded);
4572 }
4573
4574 static void raid5_free_percpu(raid5_conf_t *conf)
4575 {
4576         struct raid5_percpu *percpu;
4577         unsigned long cpu;
4578
4579         if (!conf->percpu)
4580                 return;
4581
4582         get_online_cpus();
4583         for_each_possible_cpu(cpu) {
4584                 percpu = per_cpu_ptr(conf->percpu, cpu);
4585                 safe_put_page(percpu->spare_page);
4586                 kfree(percpu->scribble);
4587         }
4588 #ifdef CONFIG_HOTPLUG_CPU
4589         unregister_cpu_notifier(&conf->cpu_notify);
4590 #endif
4591         put_online_cpus();
4592
4593         free_percpu(conf->percpu);
4594 }
4595
4596 static void free_conf(raid5_conf_t *conf)
4597 {
4598         shrink_stripes(conf);
4599         raid5_free_percpu(conf);
4600         kfree(conf->disks);
4601         kfree(conf->stripe_hashtbl);
4602         kfree(conf);
4603 }
4604
4605 #ifdef CONFIG_HOTPLUG_CPU
4606 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4607                               void *hcpu)
4608 {
4609         raid5_conf_t *conf = container_of(nfb, raid5_conf_t, cpu_notify);
4610         long cpu = (long)hcpu;
4611         struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4612
4613         switch (action) {
4614         case CPU_UP_PREPARE:
4615         case CPU_UP_PREPARE_FROZEN:
4616                 if (conf->level == 6 && !percpu->spare_page)
4617                         percpu->spare_page = alloc_page(GFP_KERNEL);
4618                 if (!percpu->scribble)
4619                         percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4620
4621                 if (!percpu->scribble ||
4622                     (conf->level == 6 && !percpu->spare_page)) {
4623                         safe_put_page(percpu->spare_page);
4624                         kfree(percpu->scribble);
4625                         pr_err("%s: failed memory allocation for cpu%ld\n",
4626                                __func__, cpu);
4627                         return NOTIFY_BAD;
4628                 }
4629                 break;
4630         case CPU_DEAD:
4631         case CPU_DEAD_FROZEN:
4632                 safe_put_page(percpu->spare_page);
4633                 kfree(percpu->scribble);
4634                 percpu->spare_page = NULL;
4635                 percpu->scribble = NULL;
4636                 break;
4637         default:
4638                 break;
4639         }
4640         return NOTIFY_OK;
4641 }
4642 #endif
4643
4644 static int raid5_alloc_percpu(raid5_conf_t *conf)
4645 {
4646         unsigned long cpu;
4647         struct page *spare_page;
4648         struct raid5_percpu *allcpus;
4649         void *scribble;
4650         int err;
4651
4652         allcpus = alloc_percpu(struct raid5_percpu);
4653         if (!allcpus)
4654                 return -ENOMEM;
4655         conf->percpu = allcpus;
4656
4657         get_online_cpus();
4658         err = 0;
4659         for_each_present_cpu(cpu) {
4660                 if (conf->level == 6) {
4661                         spare_page = alloc_page(GFP_KERNEL);
4662                         if (!spare_page) {
4663                                 err = -ENOMEM;
4664                                 break;
4665                         }
4666                         per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4667                 }
4668                 scribble = kmalloc(scribble_len(conf->raid_disks), GFP_KERNEL);
4669                 if (!scribble) {
4670                         err = -ENOMEM;
4671                         break;
4672                 }
4673                 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4674         }
4675 #ifdef CONFIG_HOTPLUG_CPU
4676         conf->cpu_notify.notifier_call = raid456_cpu_notify;
4677         conf->cpu_notify.priority = 0;
4678         if (err == 0)
4679                 err = register_cpu_notifier(&conf->cpu_notify);
4680 #endif
4681         put_online_cpus();
4682
4683         return err;
4684 }
4685
4686 static raid5_conf_t *setup_conf(mddev_t *mddev)
4687 {
4688         raid5_conf_t *conf;
4689         int raid_disk, memory;
4690         mdk_rdev_t *rdev;
4691         struct disk_info *disk;
4692
4693         if (mddev->new_level != 5
4694             && mddev->new_level != 4
4695             && mddev->new_level != 6) {
4696                 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4697                        mdname(mddev), mddev->new_level);
4698                 return ERR_PTR(-EIO);
4699         }
4700         if ((mddev->new_level == 5
4701              && !algorithm_valid_raid5(mddev->new_layout)) ||
4702             (mddev->new_level == 6
4703              && !algorithm_valid_raid6(mddev->new_layout))) {
4704                 printk(KERN_ERR "raid5: %s: layout %d not supported\n",
4705                        mdname(mddev), mddev->new_layout);
4706                 return ERR_PTR(-EIO);
4707         }
4708         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4709                 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4710                        mdname(mddev), mddev->raid_disks);
4711                 return ERR_PTR(-EINVAL);
4712         }
4713
4714         if (!mddev->new_chunk_sectors ||
4715             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4716             !is_power_of_2(mddev->new_chunk_sectors)) {
4717                 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4718                        mddev->new_chunk_sectors << 9, mdname(mddev));
4719                 return ERR_PTR(-EINVAL);
4720         }
4721
4722         conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL);
4723         if (conf == NULL)
4724                 goto abort;
4725
4726         conf->raid_disks = mddev->raid_disks;
4727         conf->scribble_len = scribble_len(conf->raid_disks);
4728         if (mddev->reshape_position == MaxSector)
4729                 conf->previous_raid_disks = mddev->raid_disks;
4730         else
4731                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4732
4733         conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4734                               GFP_KERNEL);
4735         if (!conf->disks)
4736                 goto abort;
4737
4738         conf->mddev = mddev;
4739
4740         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4741                 goto abort;
4742
4743         conf->level = mddev->new_level;
4744         if (raid5_alloc_percpu(conf) != 0)
4745                 goto abort;
4746
4747         spin_lock_init(&conf->device_lock);
4748         init_waitqueue_head(&conf->wait_for_stripe);
4749         init_waitqueue_head(&conf->wait_for_overlap);
4750         INIT_LIST_HEAD(&conf->handle_list);
4751         INIT_LIST_HEAD(&conf->hold_list);
4752         INIT_LIST_HEAD(&conf->delayed_list);
4753         INIT_LIST_HEAD(&conf->bitmap_list);
4754         INIT_LIST_HEAD(&conf->inactive_list);
4755         atomic_set(&conf->active_stripes, 0);
4756         atomic_set(&conf->preread_active_stripes, 0);
4757         atomic_set(&conf->active_aligned_reads, 0);
4758         conf->bypass_threshold = BYPASS_THRESHOLD;
4759
4760         pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4761
4762         list_for_each_entry(rdev, &mddev->disks, same_set) {
4763                 raid_disk = rdev->raid_disk;
4764                 if (raid_disk >= conf->raid_disks
4765                     || raid_disk < 0)
4766                         continue;
4767                 disk = conf->disks + raid_disk;
4768
4769                 disk->rdev = rdev;
4770
4771                 if (test_bit(In_sync, &rdev->flags)) {
4772                         char b[BDEVNAME_SIZE];
4773                         printk(KERN_INFO "raid5: device %s operational as raid"
4774                                 " disk %d\n", bdevname(rdev->bdev,b),
4775                                 raid_disk);
4776                 } else
4777                         /* Cannot rely on bitmap to complete recovery */
4778                         conf->fullsync = 1;
4779         }
4780
4781         conf->chunk_sectors = mddev->new_chunk_sectors;
4782         conf->level = mddev->new_level;
4783         if (conf->level == 6)
4784                 conf->max_degraded = 2;
4785         else
4786                 conf->max_degraded = 1;
4787         conf->algorithm = mddev->new_layout;
4788         conf->max_nr_stripes = NR_STRIPES;
4789         conf->reshape_progress = mddev->reshape_position;
4790         if (conf->reshape_progress != MaxSector) {
4791                 conf->prev_chunk_sectors = mddev->chunk_sectors;
4792                 conf->prev_algo = mddev->layout;
4793         }
4794
4795         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4796                  conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4797         if (grow_stripes(conf, conf->max_nr_stripes)) {
4798                 printk(KERN_ERR
4799                         "raid5: couldn't allocate %dkB for buffers\n", memory);
4800                 goto abort;
4801         } else
4802                 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4803                         memory, mdname(mddev));
4804
4805         conf->thread = md_register_thread(raid5d, mddev, NULL);
4806         if (!conf->thread) {
4807                 printk(KERN_ERR
4808                        "raid5: couldn't allocate thread for %s\n",
4809                        mdname(mddev));
4810                 goto abort;
4811         }
4812
4813         return conf;
4814
4815  abort:
4816         if (conf) {
4817                 free_conf(conf);
4818                 return ERR_PTR(-EIO);
4819         } else
4820                 return ERR_PTR(-ENOMEM);
4821 }
4822
4823 static int run(mddev_t *mddev)
4824 {
4825         raid5_conf_t *conf;
4826         int working_disks = 0, chunk_size;
4827         mdk_rdev_t *rdev;
4828
4829         if (mddev->recovery_cp != MaxSector)
4830                 printk(KERN_NOTICE "raid5: %s is not clean"
4831                        " -- starting background reconstruction\n",
4832                        mdname(mddev));
4833         if (mddev->reshape_position != MaxSector) {
4834                 /* Check that we can continue the reshape.
4835                  * Currently only disks can change, it must
4836                  * increase, and we must be past the point where
4837                  * a stripe over-writes itself
4838                  */
4839                 sector_t here_new, here_old;
4840                 int old_disks;
4841                 int max_degraded = (mddev->level == 6 ? 2 : 1);
4842
4843                 if (mddev->new_level != mddev->level) {
4844                         printk(KERN_ERR "raid5: %s: unsupported reshape "
4845                                "required - aborting.\n",
4846                                mdname(mddev));
4847                         return -EINVAL;
4848                 }
4849                 old_disks = mddev->raid_disks - mddev->delta_disks;
4850                 /* reshape_position must be on a new-stripe boundary, and one
4851                  * further up in new geometry must map after here in old
4852                  * geometry.
4853                  */
4854                 here_new = mddev->reshape_position;
4855                 if (sector_div(here_new, mddev->new_chunk_sectors *
4856                                (mddev->raid_disks - max_degraded))) {
4857                         printk(KERN_ERR "raid5: reshape_position not "
4858                                "on a stripe boundary\n");
4859                         return -EINVAL;
4860                 }
4861                 /* here_new is the stripe we will write to */
4862                 here_old = mddev->reshape_position;
4863                 sector_div(here_old, mddev->chunk_sectors *
4864                            (old_disks-max_degraded));
4865                 /* here_old is the first stripe that we might need to read
4866                  * from */
4867                 if (mddev->delta_disks == 0) {
4868                         /* We cannot be sure it is safe to start an in-place
4869                          * reshape.  It is only safe if user-space if monitoring
4870                          * and taking constant backups.
4871                          * mdadm always starts a situation like this in
4872                          * readonly mode so it can take control before
4873                          * allowing any writes.  So just check for that.
4874                          */
4875                         if ((here_new * mddev->new_chunk_sectors != 
4876                              here_old * mddev->chunk_sectors) ||
4877                             mddev->ro == 0) {
4878                                 printk(KERN_ERR "raid5: in-place reshape must be started"
4879                                        " in read-only mode - aborting\n");
4880                                 return -EINVAL;
4881                         }
4882                 } else if (mddev->delta_disks < 0
4883                     ? (here_new * mddev->new_chunk_sectors <=
4884                        here_old * mddev->chunk_sectors)
4885                     : (here_new * mddev->new_chunk_sectors >=
4886                        here_old * mddev->chunk_sectors)) {
4887                         /* Reading from the same stripe as writing to - bad */
4888                         printk(KERN_ERR "raid5: reshape_position too early for "
4889                                "auto-recovery - aborting.\n");
4890                         return -EINVAL;
4891                 }
4892                 printk(KERN_INFO "raid5: reshape will continue\n");
4893                 /* OK, we should be able to continue; */
4894         } else {
4895                 BUG_ON(mddev->level != mddev->new_level);
4896                 BUG_ON(mddev->layout != mddev->new_layout);
4897                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
4898                 BUG_ON(mddev->delta_disks != 0);
4899         }
4900
4901         if (mddev->private == NULL)
4902                 conf = setup_conf(mddev);
4903         else
4904                 conf = mddev->private;
4905
4906         if (IS_ERR(conf))
4907                 return PTR_ERR(conf);
4908
4909         mddev->thread = conf->thread;
4910         conf->thread = NULL;
4911         mddev->private = conf;
4912
4913         /*
4914          * 0 for a fully functional array, 1 or 2 for a degraded array.
4915          */
4916         list_for_each_entry(rdev, &mddev->disks, same_set)
4917                 if (rdev->raid_disk >= 0 &&
4918                     test_bit(In_sync, &rdev->flags))
4919                         working_disks++;
4920
4921         mddev->degraded = conf->raid_disks - working_disks;
4922
4923         if (mddev->degraded > conf->max_degraded) {
4924                 printk(KERN_ERR "raid5: not enough operational devices for %s"
4925                         " (%d/%d failed)\n",
4926                         mdname(mddev), mddev->degraded, conf->raid_disks);
4927                 goto abort;
4928         }
4929
4930         /* device size must be a multiple of chunk size */
4931         mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
4932         mddev->resync_max_sectors = mddev->dev_sectors;
4933
4934         if (mddev->degraded > 0 &&
4935             mddev->recovery_cp != MaxSector) {
4936                 if (mddev->ok_start_degraded)
4937                         printk(KERN_WARNING
4938                                "raid5: starting dirty degraded array: %s"
4939                                "- data corruption possible.\n",
4940                                mdname(mddev));
4941                 else {
4942                         printk(KERN_ERR
4943                                "raid5: cannot start dirty degraded array for %s\n",
4944                                mdname(mddev));
4945                         goto abort;
4946                 }
4947         }
4948
4949         if (mddev->degraded == 0)
4950                 printk("raid5: raid level %d set %s active with %d out of %d"
4951                        " devices, algorithm %d\n", conf->level, mdname(mddev),
4952                        mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4953                        mddev->new_layout);
4954         else
4955                 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4956                         " out of %d devices, algorithm %d\n", conf->level,
4957                         mdname(mddev), mddev->raid_disks - mddev->degraded,
4958                         mddev->raid_disks, mddev->new_layout);
4959
4960         print_raid5_conf(conf);
4961
4962         if (conf->reshape_progress != MaxSector) {
4963                 printk("...ok start reshape thread\n");
4964                 conf->reshape_safe = conf->reshape_progress;
4965                 atomic_set(&conf->reshape_stripes, 0);
4966                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4967                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4968                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4969                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4970                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4971                                                         "reshape");
4972         }
4973
4974         /* read-ahead size must cover two whole stripes, which is
4975          * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4976          */
4977         {
4978                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4979                 int stripe = data_disks *
4980                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
4981                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4982                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4983         }
4984
4985         /* Ok, everything is just fine now */
4986         if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4987                 printk(KERN_WARNING
4988                        "raid5: failed to create sysfs attributes for %s\n",
4989                        mdname(mddev));
4990
4991         mddev->queue->queue_lock = &conf->device_lock;
4992
4993         mddev->queue->unplug_fn = raid5_unplug_device;
4994         mddev->queue->backing_dev_info.congested_data = mddev;
4995         mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4996
4997         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
4998
4999         blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
5000         chunk_size = mddev->chunk_sectors << 9;
5001         blk_queue_io_min(mddev->queue, chunk_size);
5002         blk_queue_io_opt(mddev->queue, chunk_size *
5003                          (conf->raid_disks - conf->max_degraded));
5004
5005         list_for_each_entry(rdev, &mddev->disks, same_set)
5006                 disk_stack_limits(mddev->gendisk, rdev->bdev,
5007                                   rdev->data_offset << 9);
5008
5009         return 0;
5010 abort:
5011         md_unregister_thread(mddev->thread);
5012         mddev->thread = NULL;
5013         if (conf) {
5014                 print_raid5_conf(conf);
5015                 free_conf(conf);
5016         }
5017         mddev->private = NULL;
5018         printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
5019         return -EIO;
5020 }
5021
5022
5023
5024 static int stop(mddev_t *mddev)
5025 {
5026         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
5027
5028         md_unregister_thread(mddev->thread);
5029         mddev->thread = NULL;
5030         mddev->queue->backing_dev_info.congested_fn = NULL;
5031         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
5032         sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
5033         free_conf(conf);
5034         mddev->private = NULL;
5035         return 0;
5036 }
5037
5038 #ifdef DEBUG
5039 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
5040 {
5041         int i;
5042
5043         seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
5044                    (unsigned long long)sh->sector, sh->pd_idx, sh->state);
5045         seq_printf(seq, "sh %llu,  count %d.\n",
5046                    (unsigned long long)sh->sector, atomic_read(&sh->count));
5047         seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
5048         for (i = 0; i < sh->disks; i++) {
5049                 seq_printf(seq, "(cache%d: %p %ld) ",
5050                            i, sh->dev[i].page, sh->dev[i].flags);
5051         }
5052         seq_printf(seq, "\n");
5053 }
5054
5055 static void printall(struct seq_file *seq, raid5_conf_t *conf)
5056 {
5057         struct stripe_head *sh;
5058         struct hlist_node *hn;
5059         int i;
5060
5061         spin_lock_irq(&conf->device_lock);
5062         for (i = 0; i < NR_HASH; i++) {
5063                 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
5064                         if (sh->raid_conf != conf)
5065                                 continue;
5066                         print_sh(seq, sh);
5067                 }
5068         }
5069         spin_unlock_irq(&conf->device_lock);
5070 }
5071 #endif
5072
5073 static void status(struct seq_file *seq, mddev_t *mddev)
5074 {
5075         raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
5076         int i;
5077
5078         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5079                 mddev->chunk_sectors / 2, mddev->layout);
5080         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5081         for (i = 0; i < conf->raid_disks; i++)
5082                 seq_printf (seq, "%s",
5083                                conf->disks[i].rdev &&
5084                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5085         seq_printf (seq, "]");
5086 #ifdef DEBUG
5087         seq_printf (seq, "\n");
5088         printall(seq, conf);
5089 #endif
5090 }
5091
5092 static void print_raid5_conf (raid5_conf_t *conf)
5093 {
5094         int i;
5095         struct disk_info *tmp;
5096
5097         printk("RAID5 conf printout:\n");
5098         if (!conf) {
5099                 printk("(conf==NULL)\n");
5100                 return;
5101         }
5102         printk(" --- rd:%d wd:%d\n", conf->raid_disks,
5103                  conf->raid_disks - conf->mddev->degraded);
5104
5105         for (i = 0; i < conf->raid_disks; i++) {
5106                 char b[BDEVNAME_SIZE];
5107                 tmp = conf->disks + i;
5108                 if (tmp->rdev)
5109                 printk(" disk %d, o:%d, dev:%s\n",
5110                         i, !test_bit(Faulty, &tmp->rdev->flags),
5111                         bdevname(tmp->rdev->bdev,b));
5112         }
5113 }
5114
5115 static int raid5_spare_active(mddev_t *mddev)
5116 {
5117         int i;
5118         raid5_conf_t *conf = mddev->private;
5119         struct disk_info *tmp;
5120
5121         for (i = 0; i < conf->raid_disks; i++) {
5122                 tmp = conf->disks + i;
5123                 if (tmp->rdev
5124                     && !test_bit(Faulty, &tmp->rdev->flags)
5125                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5126                         unsigned long flags;
5127                         spin_lock_irqsave(&conf->device_lock, flags);
5128                         mddev->degraded--;
5129                         spin_unlock_irqrestore(&conf->device_lock, flags);
5130                 }
5131         }
5132         print_raid5_conf(conf);
5133         return 0;
5134 }
5135
5136 static int raid5_remove_disk(mddev_t *mddev, int number)
5137 {
5138         raid5_conf_t *conf = mddev->private;
5139         int err = 0;
5140         mdk_rdev_t *rdev;
5141         struct disk_info *p = conf->disks + number;
5142
5143         print_raid5_conf(conf);
5144         rdev = p->rdev;
5145         if (rdev) {
5146                 if (number >= conf->raid_disks &&
5147                     conf->reshape_progress == MaxSector)
5148                         clear_bit(In_sync, &rdev->flags);
5149
5150                 if (test_bit(In_sync, &rdev->flags) ||
5151                     atomic_read(&rdev->nr_pending)) {
5152                         err = -EBUSY;
5153                         goto abort;
5154                 }
5155                 /* Only remove non-faulty devices if recovery
5156                  * isn't possible.
5157                  */
5158                 if (!test_bit(Faulty, &rdev->flags) &&
5159                     mddev->degraded <= conf->max_degraded &&
5160                     number < conf->raid_disks) {
5161                         err = -EBUSY;
5162                         goto abort;
5163                 }
5164                 p->rdev = NULL;
5165                 synchronize_rcu();
5166                 if (atomic_read(&rdev->nr_pending)) {
5167                         /* lost the race, try later */
5168                         err = -EBUSY;
5169                         p->rdev = rdev;
5170                 }
5171         }
5172 abort:
5173
5174         print_raid5_conf(conf);
5175         return err;
5176 }
5177
5178 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
5179 {
5180         raid5_conf_t *conf = mddev->private;
5181         int err = -EEXIST;
5182         int disk;
5183         struct disk_info *p;
5184         int first = 0;
5185         int last = conf->raid_disks - 1;
5186
5187         if (mddev->degraded > conf->max_degraded)
5188                 /* no point adding a device */
5189                 return -EINVAL;
5190
5191         if (rdev->raid_disk >= 0)
5192                 first = last = rdev->raid_disk;
5193
5194         /*
5195          * find the disk ... but prefer rdev->saved_raid_disk
5196          * if possible.
5197          */
5198         if (rdev->saved_raid_disk >= 0 &&
5199             rdev->saved_raid_disk >= first &&
5200             conf->disks[rdev->saved_raid_disk].rdev == NULL)
5201                 disk = rdev->saved_raid_disk;
5202         else
5203                 disk = first;
5204         for ( ; disk <= last ; disk++)
5205                 if ((p=conf->disks + disk)->rdev == NULL) {
5206                         clear_bit(In_sync, &rdev->flags);
5207                         rdev->raid_disk = disk;
5208                         err = 0;
5209                         if (rdev->saved_raid_disk != disk)
5210                                 conf->fullsync = 1;
5211                         rcu_assign_pointer(p->rdev, rdev);
5212                         break;
5213                 }
5214         print_raid5_conf(conf);
5215         return err;
5216 }
5217
5218 static int raid5_resize(mddev_t *mddev, sector_t sectors)
5219 {
5220         /* no resync is happening, and there is enough space
5221          * on all devices, so we can resize.
5222          * We need to make sure resync covers any new space.
5223          * If the array is shrinking we should possibly wait until
5224          * any io in the removed space completes, but it hardly seems
5225          * worth it.
5226          */
5227         sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5228         md_set_array_sectors(mddev, raid5_size(mddev, sectors,
5229                                                mddev->raid_disks));
5230         if (mddev->array_sectors >
5231             raid5_size(mddev, sectors, mddev->raid_disks))
5232                 return -EINVAL;
5233         set_capacity(mddev->gendisk, mddev->array_sectors);
5234         mddev->changed = 1;
5235         revalidate_disk(mddev->gendisk);
5236         if (sectors > mddev->dev_sectors && mddev->recovery_cp == MaxSector) {
5237                 mddev->recovery_cp = mddev->dev_sectors;
5238                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5239         }
5240         mddev->dev_sectors = sectors;
5241         mddev->resync_max_sectors = sectors;
5242         return 0;
5243 }
5244
5245 static int check_stripe_cache(mddev_t *mddev)
5246 {
5247         /* Can only proceed if there are plenty of stripe_heads.
5248          * We need a minimum of one full stripe,, and for sensible progress
5249          * it is best to have about 4 times that.
5250          * If we require 4 times, then the default 256 4K stripe_heads will
5251          * allow for chunk sizes up to 256K, which is probably OK.
5252          * If the chunk size is greater, user-space should request more
5253          * stripe_heads first.
5254          */
5255         raid5_conf_t *conf = mddev->private;
5256         if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5257             > conf->max_nr_stripes ||
5258             ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5259             > conf->max_nr_stripes) {
5260                 printk(KERN_WARNING "raid5: reshape: not enough stripes.  Needed %lu\n",
5261                        ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5262                         / STRIPE_SIZE)*4);
5263                 return 0;
5264         }
5265         return 1;
5266 }
5267
5268 static int check_reshape(mddev_t *mddev)
5269 {
5270         raid5_conf_t *conf = mddev->private;
5271
5272         if (mddev->delta_disks == 0 &&
5273             mddev->new_layout == mddev->layout &&
5274             mddev->new_chunk_sectors == mddev->chunk_sectors)
5275                 return 0; /* nothing to do */
5276         if (mddev->bitmap)
5277                 /* Cannot grow a bitmap yet */
5278                 return -EBUSY;
5279         if (mddev->degraded > conf->max_degraded)
5280                 return -EINVAL;
5281         if (mddev->delta_disks < 0) {
5282                 /* We might be able to shrink, but the devices must
5283                  * be made bigger first.
5284                  * For raid6, 4 is the minimum size.
5285                  * Otherwise 2 is the minimum
5286                  */
5287                 int min = 2;
5288                 if (mddev->level == 6)
5289                         min = 4;
5290                 if (mddev->raid_disks + mddev->delta_disks < min)
5291                         return -EINVAL;
5292         }
5293
5294         if (!check_stripe_cache(mddev))
5295                 return -ENOSPC;
5296
5297         return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5298 }
5299
5300 static int raid5_start_reshape(mddev_t *mddev)
5301 {
5302         raid5_conf_t *conf = mddev->private;
5303         mdk_rdev_t *rdev;
5304         int spares = 0;
5305         int added_devices = 0;
5306         unsigned long flags;
5307
5308         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5309                 return -EBUSY;
5310
5311         if (!check_stripe_cache(mddev))
5312                 return -ENOSPC;
5313
5314         list_for_each_entry(rdev, &mddev->disks, same_set)
5315                 if (rdev->raid_disk < 0 &&
5316                     !test_bit(Faulty, &rdev->flags))
5317                         spares++;
5318
5319         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5320                 /* Not enough devices even to make a degraded array
5321                  * of that size
5322                  */
5323                 return -EINVAL;
5324
5325         /* Refuse to reduce size of the array.  Any reductions in
5326          * array size must be through explicit setting of array_size
5327          * attribute.
5328          */
5329         if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5330             < mddev->array_sectors) {
5331                 printk(KERN_ERR "md: %s: array size must be reduced "
5332                        "before number of disks\n", mdname(mddev));
5333                 return -EINVAL;
5334         }
5335
5336         atomic_set(&conf->reshape_stripes, 0);
5337         spin_lock_irq(&conf->device_lock);
5338         conf->previous_raid_disks = conf->raid_disks;
5339         conf->raid_disks += mddev->delta_disks;
5340         conf->prev_chunk_sectors = conf->chunk_sectors;
5341         conf->chunk_sectors = mddev->new_chunk_sectors;
5342         conf->prev_algo = conf->algorithm;
5343         conf->algorithm = mddev->new_layout;
5344         if (mddev->delta_disks < 0)
5345                 conf->reshape_progress = raid5_size(mddev, 0, 0);
5346         else
5347                 conf->reshape_progress = 0;
5348         conf->reshape_safe = conf->reshape_progress;
5349         conf->generation++;
5350         spin_unlock_irq(&conf->device_lock);
5351
5352         /* Add some new drives, as many as will fit.
5353          * We know there are enough to make the newly sized array work.
5354          */
5355         list_for_each_entry(rdev, &mddev->disks, same_set)
5356                 if (rdev->raid_disk < 0 &&
5357                     !test_bit(Faulty, &rdev->flags)) {
5358                         if (raid5_add_disk(mddev, rdev) == 0) {
5359                                 char nm[20];
5360                                 set_bit(In_sync, &rdev->flags);
5361                                 added_devices++;
5362                                 rdev->recovery_offset = 0;
5363                                 sprintf(nm, "rd%d", rdev->raid_disk);
5364                                 if (sysfs_create_link(&mddev->kobj,
5365