Merge commit '3ff195b011d7decf501a4d55aeed312731094796' into for-linus
[linux-2.6.git] / drivers / md / raid5.c
1 /*
2  * raid5.c : Multiple Devices driver for Linux
3  *         Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4  *         Copyright (C) 1999, 2000 Ingo Molnar
5  *         Copyright (C) 2002, 2003 H. Peter Anvin
6  *
7  * RAID-4/5/6 management functions.
8  * Thanks to Penguin Computing for making the RAID-6 development possible
9  * by donating a test server!
10  *
11  * This program is free software; you can redistribute it and/or modify
12  * it under the terms of the GNU General Public License as published by
13  * the Free Software Foundation; either version 2, or (at your option)
14  * any later version.
15  *
16  * You should have received a copy of the GNU General Public License
17  * (for example /usr/src/linux/COPYING); if not, write to the Free
18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19  */
20
21 /*
22  * BITMAP UNPLUGGING:
23  *
24  * The sequencing for updating the bitmap reliably is a little
25  * subtle (and I got it wrong the first time) so it deserves some
26  * explanation.
27  *
28  * We group bitmap updates into batches.  Each batch has a number.
29  * We may write out several batches at once, but that isn't very important.
30  * conf->bm_write is the number of the last batch successfully written.
31  * conf->bm_flush is the number of the last batch that was closed to
32  *    new additions.
33  * When we discover that we will need to write to any block in a stripe
34  * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35  * the number of the batch it will be in. This is bm_flush+1.
36  * When we are ready to do a write, if that batch hasn't been written yet,
37  *   we plug the array and queue the stripe for later.
38  * When an unplug happens, we increment bm_flush, thus closing the current
39  *   batch.
40  * When we notice that bm_flush > bm_write, we write out all pending updates
41  * to the bitmap, and advance bm_write to where bm_flush was.
42  * This may occasionally write a bit out twice, but is sure never to
43  * miss any bits.
44  */
45
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/async.h>
51 #include <linux/seq_file.h>
52 #include <linux/cpu.h>
53 #include <linux/slab.h>
54 #include "md.h"
55 #include "raid5.h"
56 #include "raid0.h"
57 #include "bitmap.h"
58
59 /*
60  * Stripe cache
61  */
62
63 #define NR_STRIPES              256
64 #define STRIPE_SIZE             PAGE_SIZE
65 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
66 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
67 #define IO_THRESHOLD            1
68 #define BYPASS_THRESHOLD        1
69 #define NR_HASH                 (PAGE_SIZE / sizeof(struct hlist_head))
70 #define HASH_MASK               (NR_HASH - 1)
71
72 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
73
74 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
75  * order without overlap.  There may be several bio's per stripe+device, and
76  * a bio could span several devices.
77  * When walking this list for a particular stripe+device, we must never proceed
78  * beyond a bio that extends past this device, as the next bio might no longer
79  * be valid.
80  * This macro is used to determine the 'next' bio in the list, given the sector
81  * of the current stripe+device
82  */
83 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
84 /*
85  * The following can be used to debug the driver
86  */
87 #define RAID5_PARANOIA  1
88 #if RAID5_PARANOIA && defined(CONFIG_SMP)
89 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
90 #else
91 # define CHECK_DEVLOCK()
92 #endif
93
94 #ifdef DEBUG
95 #define inline
96 #define __inline__
97 #endif
98
99 #define printk_rl(args...) ((void) (printk_ratelimit() && printk(args)))
100
101 /*
102  * We maintain a biased count of active stripes in the bottom 16 bits of
103  * bi_phys_segments, and a count of processed stripes in the upper 16 bits
104  */
105 static inline int raid5_bi_phys_segments(struct bio *bio)
106 {
107         return bio->bi_phys_segments & 0xffff;
108 }
109
110 static inline int raid5_bi_hw_segments(struct bio *bio)
111 {
112         return (bio->bi_phys_segments >> 16) & 0xffff;
113 }
114
115 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
116 {
117         --bio->bi_phys_segments;
118         return raid5_bi_phys_segments(bio);
119 }
120
121 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
122 {
123         unsigned short val = raid5_bi_hw_segments(bio);
124
125         --val;
126         bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
127         return val;
128 }
129
130 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
131 {
132         bio->bi_phys_segments = raid5_bi_phys_segments(bio) || (cnt << 16);
133 }
134
135 /* Find first data disk in a raid6 stripe */
136 static inline int raid6_d0(struct stripe_head *sh)
137 {
138         if (sh->ddf_layout)
139                 /* ddf always start from first device */
140                 return 0;
141         /* md starts just after Q block */
142         if (sh->qd_idx == sh->disks - 1)
143                 return 0;
144         else
145                 return sh->qd_idx + 1;
146 }
147 static inline int raid6_next_disk(int disk, int raid_disks)
148 {
149         disk++;
150         return (disk < raid_disks) ? disk : 0;
151 }
152
153 /* When walking through the disks in a raid5, starting at raid6_d0,
154  * We need to map each disk to a 'slot', where the data disks are slot
155  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
156  * is raid_disks-1.  This help does that mapping.
157  */
158 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
159                              int *count, int syndrome_disks)
160 {
161         int slot = *count;
162
163         if (sh->ddf_layout)
164                 (*count)++;
165         if (idx == sh->pd_idx)
166                 return syndrome_disks;
167         if (idx == sh->qd_idx)
168                 return syndrome_disks + 1;
169         if (!sh->ddf_layout)
170                 (*count)++;
171         return slot;
172 }
173
174 static void return_io(struct bio *return_bi)
175 {
176         struct bio *bi = return_bi;
177         while (bi) {
178
179                 return_bi = bi->bi_next;
180                 bi->bi_next = NULL;
181                 bi->bi_size = 0;
182                 bio_endio(bi, 0);
183                 bi = return_bi;
184         }
185 }
186
187 static void print_raid5_conf (raid5_conf_t *conf);
188
189 static int stripe_operations_active(struct stripe_head *sh)
190 {
191         return sh->check_state || sh->reconstruct_state ||
192                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
193                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
194 }
195
196 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
197 {
198         if (atomic_dec_and_test(&sh->count)) {
199                 BUG_ON(!list_empty(&sh->lru));
200                 BUG_ON(atomic_read(&conf->active_stripes)==0);
201                 if (test_bit(STRIPE_HANDLE, &sh->state)) {
202                         if (test_bit(STRIPE_DELAYED, &sh->state)) {
203                                 list_add_tail(&sh->lru, &conf->delayed_list);
204                                 blk_plug_device(conf->mddev->queue);
205                         } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
206                                    sh->bm_seq - conf->seq_write > 0) {
207                                 list_add_tail(&sh->lru, &conf->bitmap_list);
208                                 blk_plug_device(conf->mddev->queue);
209                         } else {
210                                 clear_bit(STRIPE_BIT_DELAY, &sh->state);
211                                 list_add_tail(&sh->lru, &conf->handle_list);
212                         }
213                         md_wakeup_thread(conf->mddev->thread);
214                 } else {
215                         BUG_ON(stripe_operations_active(sh));
216                         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
217                                 atomic_dec(&conf->preread_active_stripes);
218                                 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
219                                         md_wakeup_thread(conf->mddev->thread);
220                         }
221                         atomic_dec(&conf->active_stripes);
222                         if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
223                                 list_add_tail(&sh->lru, &conf->inactive_list);
224                                 wake_up(&conf->wait_for_stripe);
225                                 if (conf->retry_read_aligned)
226                                         md_wakeup_thread(conf->mddev->thread);
227                         }
228                 }
229         }
230 }
231
232 static void release_stripe(struct stripe_head *sh)
233 {
234         raid5_conf_t *conf = sh->raid_conf;
235         unsigned long flags;
236
237         spin_lock_irqsave(&conf->device_lock, flags);
238         __release_stripe(conf, sh);
239         spin_unlock_irqrestore(&conf->device_lock, flags);
240 }
241
242 static inline void remove_hash(struct stripe_head *sh)
243 {
244         pr_debug("remove_hash(), stripe %llu\n",
245                 (unsigned long long)sh->sector);
246
247         hlist_del_init(&sh->hash);
248 }
249
250 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
251 {
252         struct hlist_head *hp = stripe_hash(conf, sh->sector);
253
254         pr_debug("insert_hash(), stripe %llu\n",
255                 (unsigned long long)sh->sector);
256
257         CHECK_DEVLOCK();
258         hlist_add_head(&sh->hash, hp);
259 }
260
261
262 /* find an idle stripe, make sure it is unhashed, and return it. */
263 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
264 {
265         struct stripe_head *sh = NULL;
266         struct list_head *first;
267
268         CHECK_DEVLOCK();
269         if (list_empty(&conf->inactive_list))
270                 goto out;
271         first = conf->inactive_list.next;
272         sh = list_entry(first, struct stripe_head, lru);
273         list_del_init(first);
274         remove_hash(sh);
275         atomic_inc(&conf->active_stripes);
276 out:
277         return sh;
278 }
279
280 static void shrink_buffers(struct stripe_head *sh, int num)
281 {
282         struct page *p;
283         int i;
284
285         for (i=0; i<num ; i++) {
286                 p = sh->dev[i].page;
287                 if (!p)
288                         continue;
289                 sh->dev[i].page = NULL;
290                 put_page(p);
291         }
292 }
293
294 static int grow_buffers(struct stripe_head *sh, int num)
295 {
296         int i;
297
298         for (i=0; i<num; i++) {
299                 struct page *page;
300
301                 if (!(page = alloc_page(GFP_KERNEL))) {
302                         return 1;
303                 }
304                 sh->dev[i].page = page;
305         }
306         return 0;
307 }
308
309 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
310 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
311                             struct stripe_head *sh);
312
313 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
314 {
315         raid5_conf_t *conf = sh->raid_conf;
316         int i;
317
318         BUG_ON(atomic_read(&sh->count) != 0);
319         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
320         BUG_ON(stripe_operations_active(sh));
321
322         CHECK_DEVLOCK();
323         pr_debug("init_stripe called, stripe %llu\n",
324                 (unsigned long long)sh->sector);
325
326         remove_hash(sh);
327
328         sh->generation = conf->generation - previous;
329         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
330         sh->sector = sector;
331         stripe_set_idx(sector, conf, previous, sh);
332         sh->state = 0;
333
334
335         for (i = sh->disks; i--; ) {
336                 struct r5dev *dev = &sh->dev[i];
337
338                 if (dev->toread || dev->read || dev->towrite || dev->written ||
339                     test_bit(R5_LOCKED, &dev->flags)) {
340                         printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
341                                (unsigned long long)sh->sector, i, dev->toread,
342                                dev->read, dev->towrite, dev->written,
343                                test_bit(R5_LOCKED, &dev->flags));
344                         BUG();
345                 }
346                 dev->flags = 0;
347                 raid5_build_block(sh, i, previous);
348         }
349         insert_hash(conf, sh);
350 }
351
352 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector,
353                                          short generation)
354 {
355         struct stripe_head *sh;
356         struct hlist_node *hn;
357
358         CHECK_DEVLOCK();
359         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
360         hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
361                 if (sh->sector == sector && sh->generation == generation)
362                         return sh;
363         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
364         return NULL;
365 }
366
367 static void unplug_slaves(mddev_t *mddev);
368 static void raid5_unplug_device(struct request_queue *q);
369
370 static struct stripe_head *
371 get_active_stripe(raid5_conf_t *conf, sector_t sector,
372                   int previous, int noblock, int noquiesce)
373 {
374         struct stripe_head *sh;
375
376         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
377
378         spin_lock_irq(&conf->device_lock);
379
380         do {
381                 wait_event_lock_irq(conf->wait_for_stripe,
382                                     conf->quiesce == 0 || noquiesce,
383                                     conf->device_lock, /* nothing */);
384                 sh = __find_stripe(conf, sector, conf->generation - previous);
385                 if (!sh) {
386                         if (!conf->inactive_blocked)
387                                 sh = get_free_stripe(conf);
388                         if (noblock && sh == NULL)
389                                 break;
390                         if (!sh) {
391                                 conf->inactive_blocked = 1;
392                                 wait_event_lock_irq(conf->wait_for_stripe,
393                                                     !list_empty(&conf->inactive_list) &&
394                                                     (atomic_read(&conf->active_stripes)
395                                                      < (conf->max_nr_stripes *3/4)
396                                                      || !conf->inactive_blocked),
397                                                     conf->device_lock,
398                                                     raid5_unplug_device(conf->mddev->queue)
399                                         );
400                                 conf->inactive_blocked = 0;
401                         } else
402                                 init_stripe(sh, sector, previous);
403                 } else {
404                         if (atomic_read(&sh->count)) {
405                                 BUG_ON(!list_empty(&sh->lru)
406                                     && !test_bit(STRIPE_EXPANDING, &sh->state));
407                         } else {
408                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
409                                         atomic_inc(&conf->active_stripes);
410                                 if (list_empty(&sh->lru) &&
411                                     !test_bit(STRIPE_EXPANDING, &sh->state))
412                                         BUG();
413                                 list_del_init(&sh->lru);
414                         }
415                 }
416         } while (sh == NULL);
417
418         if (sh)
419                 atomic_inc(&sh->count);
420
421         spin_unlock_irq(&conf->device_lock);
422         return sh;
423 }
424
425 static void
426 raid5_end_read_request(struct bio *bi, int error);
427 static void
428 raid5_end_write_request(struct bio *bi, int error);
429
430 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
431 {
432         raid5_conf_t *conf = sh->raid_conf;
433         int i, disks = sh->disks;
434
435         might_sleep();
436
437         for (i = disks; i--; ) {
438                 int rw;
439                 struct bio *bi;
440                 mdk_rdev_t *rdev;
441                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
442                         rw = WRITE;
443                 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
444                         rw = READ;
445                 else
446                         continue;
447
448                 bi = &sh->dev[i].req;
449
450                 bi->bi_rw = rw;
451                 if (rw == WRITE)
452                         bi->bi_end_io = raid5_end_write_request;
453                 else
454                         bi->bi_end_io = raid5_end_read_request;
455
456                 rcu_read_lock();
457                 rdev = rcu_dereference(conf->disks[i].rdev);
458                 if (rdev && test_bit(Faulty, &rdev->flags))
459                         rdev = NULL;
460                 if (rdev)
461                         atomic_inc(&rdev->nr_pending);
462                 rcu_read_unlock();
463
464                 if (rdev) {
465                         if (s->syncing || s->expanding || s->expanded)
466                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
467
468                         set_bit(STRIPE_IO_STARTED, &sh->state);
469
470                         bi->bi_bdev = rdev->bdev;
471                         pr_debug("%s: for %llu schedule op %ld on disc %d\n",
472                                 __func__, (unsigned long long)sh->sector,
473                                 bi->bi_rw, i);
474                         atomic_inc(&sh->count);
475                         bi->bi_sector = sh->sector + rdev->data_offset;
476                         bi->bi_flags = 1 << BIO_UPTODATE;
477                         bi->bi_vcnt = 1;
478                         bi->bi_max_vecs = 1;
479                         bi->bi_idx = 0;
480                         bi->bi_io_vec = &sh->dev[i].vec;
481                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
482                         bi->bi_io_vec[0].bv_offset = 0;
483                         bi->bi_size = STRIPE_SIZE;
484                         bi->bi_next = NULL;
485                         if (rw == WRITE &&
486                             test_bit(R5_ReWrite, &sh->dev[i].flags))
487                                 atomic_add(STRIPE_SECTORS,
488                                         &rdev->corrected_errors);
489                         generic_make_request(bi);
490                 } else {
491                         if (rw == WRITE)
492                                 set_bit(STRIPE_DEGRADED, &sh->state);
493                         pr_debug("skip op %ld on disc %d for sector %llu\n",
494                                 bi->bi_rw, i, (unsigned long long)sh->sector);
495                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
496                         set_bit(STRIPE_HANDLE, &sh->state);
497                 }
498         }
499 }
500
501 static struct dma_async_tx_descriptor *
502 async_copy_data(int frombio, struct bio *bio, struct page *page,
503         sector_t sector, struct dma_async_tx_descriptor *tx)
504 {
505         struct bio_vec *bvl;
506         struct page *bio_page;
507         int i;
508         int page_offset;
509         struct async_submit_ctl submit;
510         enum async_tx_flags flags = 0;
511
512         if (bio->bi_sector >= sector)
513                 page_offset = (signed)(bio->bi_sector - sector) * 512;
514         else
515                 page_offset = (signed)(sector - bio->bi_sector) * -512;
516
517         if (frombio)
518                 flags |= ASYNC_TX_FENCE;
519         init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
520
521         bio_for_each_segment(bvl, bio, i) {
522                 int len = bio_iovec_idx(bio, i)->bv_len;
523                 int clen;
524                 int b_offset = 0;
525
526                 if (page_offset < 0) {
527                         b_offset = -page_offset;
528                         page_offset += b_offset;
529                         len -= b_offset;
530                 }
531
532                 if (len > 0 && page_offset + len > STRIPE_SIZE)
533                         clen = STRIPE_SIZE - page_offset;
534                 else
535                         clen = len;
536
537                 if (clen > 0) {
538                         b_offset += bio_iovec_idx(bio, i)->bv_offset;
539                         bio_page = bio_iovec_idx(bio, i)->bv_page;
540                         if (frombio)
541                                 tx = async_memcpy(page, bio_page, page_offset,
542                                                   b_offset, clen, &submit);
543                         else
544                                 tx = async_memcpy(bio_page, page, b_offset,
545                                                   page_offset, clen, &submit);
546                 }
547                 /* chain the operations */
548                 submit.depend_tx = tx;
549
550                 if (clen < len) /* hit end of page */
551                         break;
552                 page_offset +=  len;
553         }
554
555         return tx;
556 }
557
558 static void ops_complete_biofill(void *stripe_head_ref)
559 {
560         struct stripe_head *sh = stripe_head_ref;
561         struct bio *return_bi = NULL;
562         raid5_conf_t *conf = sh->raid_conf;
563         int i;
564
565         pr_debug("%s: stripe %llu\n", __func__,
566                 (unsigned long long)sh->sector);
567
568         /* clear completed biofills */
569         spin_lock_irq(&conf->device_lock);
570         for (i = sh->disks; i--; ) {
571                 struct r5dev *dev = &sh->dev[i];
572
573                 /* acknowledge completion of a biofill operation */
574                 /* and check if we need to reply to a read request,
575                  * new R5_Wantfill requests are held off until
576                  * !STRIPE_BIOFILL_RUN
577                  */
578                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
579                         struct bio *rbi, *rbi2;
580
581                         BUG_ON(!dev->read);
582                         rbi = dev->read;
583                         dev->read = NULL;
584                         while (rbi && rbi->bi_sector <
585                                 dev->sector + STRIPE_SECTORS) {
586                                 rbi2 = r5_next_bio(rbi, dev->sector);
587                                 if (!raid5_dec_bi_phys_segments(rbi)) {
588                                         rbi->bi_next = return_bi;
589                                         return_bi = rbi;
590                                 }
591                                 rbi = rbi2;
592                         }
593                 }
594         }
595         spin_unlock_irq(&conf->device_lock);
596         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
597
598         return_io(return_bi);
599
600         set_bit(STRIPE_HANDLE, &sh->state);
601         release_stripe(sh);
602 }
603
604 static void ops_run_biofill(struct stripe_head *sh)
605 {
606         struct dma_async_tx_descriptor *tx = NULL;
607         raid5_conf_t *conf = sh->raid_conf;
608         struct async_submit_ctl submit;
609         int i;
610
611         pr_debug("%s: stripe %llu\n", __func__,
612                 (unsigned long long)sh->sector);
613
614         for (i = sh->disks; i--; ) {
615                 struct r5dev *dev = &sh->dev[i];
616                 if (test_bit(R5_Wantfill, &dev->flags)) {
617                         struct bio *rbi;
618                         spin_lock_irq(&conf->device_lock);
619                         dev->read = rbi = dev->toread;
620                         dev->toread = NULL;
621                         spin_unlock_irq(&conf->device_lock);
622                         while (rbi && rbi->bi_sector <
623                                 dev->sector + STRIPE_SECTORS) {
624                                 tx = async_copy_data(0, rbi, dev->page,
625                                         dev->sector, tx);
626                                 rbi = r5_next_bio(rbi, dev->sector);
627                         }
628                 }
629         }
630
631         atomic_inc(&sh->count);
632         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
633         async_trigger_callback(&submit);
634 }
635
636 static void mark_target_uptodate(struct stripe_head *sh, int target)
637 {
638         struct r5dev *tgt;
639
640         if (target < 0)
641                 return;
642
643         tgt = &sh->dev[target];
644         set_bit(R5_UPTODATE, &tgt->flags);
645         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
646         clear_bit(R5_Wantcompute, &tgt->flags);
647 }
648
649 static void ops_complete_compute(void *stripe_head_ref)
650 {
651         struct stripe_head *sh = stripe_head_ref;
652
653         pr_debug("%s: stripe %llu\n", __func__,
654                 (unsigned long long)sh->sector);
655
656         /* mark the computed target(s) as uptodate */
657         mark_target_uptodate(sh, sh->ops.target);
658         mark_target_uptodate(sh, sh->ops.target2);
659
660         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
661         if (sh->check_state == check_state_compute_run)
662                 sh->check_state = check_state_compute_result;
663         set_bit(STRIPE_HANDLE, &sh->state);
664         release_stripe(sh);
665 }
666
667 /* return a pointer to the address conversion region of the scribble buffer */
668 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
669                                  struct raid5_percpu *percpu)
670 {
671         return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
672 }
673
674 static struct dma_async_tx_descriptor *
675 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
676 {
677         int disks = sh->disks;
678         struct page **xor_srcs = percpu->scribble;
679         int target = sh->ops.target;
680         struct r5dev *tgt = &sh->dev[target];
681         struct page *xor_dest = tgt->page;
682         int count = 0;
683         struct dma_async_tx_descriptor *tx;
684         struct async_submit_ctl submit;
685         int i;
686
687         pr_debug("%s: stripe %llu block: %d\n",
688                 __func__, (unsigned long long)sh->sector, target);
689         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
690
691         for (i = disks; i--; )
692                 if (i != target)
693                         xor_srcs[count++] = sh->dev[i].page;
694
695         atomic_inc(&sh->count);
696
697         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
698                           ops_complete_compute, sh, to_addr_conv(sh, percpu));
699         if (unlikely(count == 1))
700                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
701         else
702                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
703
704         return tx;
705 }
706
707 /* set_syndrome_sources - populate source buffers for gen_syndrome
708  * @srcs - (struct page *) array of size sh->disks
709  * @sh - stripe_head to parse
710  *
711  * Populates srcs in proper layout order for the stripe and returns the
712  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
713  * destination buffer is recorded in srcs[count] and the Q destination
714  * is recorded in srcs[count+1]].
715  */
716 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
717 {
718         int disks = sh->disks;
719         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
720         int d0_idx = raid6_d0(sh);
721         int count;
722         int i;
723
724         for (i = 0; i < disks; i++)
725                 srcs[i] = NULL;
726
727         count = 0;
728         i = d0_idx;
729         do {
730                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
731
732                 srcs[slot] = sh->dev[i].page;
733                 i = raid6_next_disk(i, disks);
734         } while (i != d0_idx);
735
736         return syndrome_disks;
737 }
738
739 static struct dma_async_tx_descriptor *
740 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
741 {
742         int disks = sh->disks;
743         struct page **blocks = percpu->scribble;
744         int target;
745         int qd_idx = sh->qd_idx;
746         struct dma_async_tx_descriptor *tx;
747         struct async_submit_ctl submit;
748         struct r5dev *tgt;
749         struct page *dest;
750         int i;
751         int count;
752
753         if (sh->ops.target < 0)
754                 target = sh->ops.target2;
755         else if (sh->ops.target2 < 0)
756                 target = sh->ops.target;
757         else
758                 /* we should only have one valid target */
759                 BUG();
760         BUG_ON(target < 0);
761         pr_debug("%s: stripe %llu block: %d\n",
762                 __func__, (unsigned long long)sh->sector, target);
763
764         tgt = &sh->dev[target];
765         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
766         dest = tgt->page;
767
768         atomic_inc(&sh->count);
769
770         if (target == qd_idx) {
771                 count = set_syndrome_sources(blocks, sh);
772                 blocks[count] = NULL; /* regenerating p is not necessary */
773                 BUG_ON(blocks[count+1] != dest); /* q should already be set */
774                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
775                                   ops_complete_compute, sh,
776                                   to_addr_conv(sh, percpu));
777                 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
778         } else {
779                 /* Compute any data- or p-drive using XOR */
780                 count = 0;
781                 for (i = disks; i-- ; ) {
782                         if (i == target || i == qd_idx)
783                                 continue;
784                         blocks[count++] = sh->dev[i].page;
785                 }
786
787                 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
788                                   NULL, ops_complete_compute, sh,
789                                   to_addr_conv(sh, percpu));
790                 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
791         }
792
793         return tx;
794 }
795
796 static struct dma_async_tx_descriptor *
797 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
798 {
799         int i, count, disks = sh->disks;
800         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
801         int d0_idx = raid6_d0(sh);
802         int faila = -1, failb = -1;
803         int target = sh->ops.target;
804         int target2 = sh->ops.target2;
805         struct r5dev *tgt = &sh->dev[target];
806         struct r5dev *tgt2 = &sh->dev[target2];
807         struct dma_async_tx_descriptor *tx;
808         struct page **blocks = percpu->scribble;
809         struct async_submit_ctl submit;
810
811         pr_debug("%s: stripe %llu block1: %d block2: %d\n",
812                  __func__, (unsigned long long)sh->sector, target, target2);
813         BUG_ON(target < 0 || target2 < 0);
814         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
815         BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
816
817         /* we need to open-code set_syndrome_sources to handle the
818          * slot number conversion for 'faila' and 'failb'
819          */
820         for (i = 0; i < disks ; i++)
821                 blocks[i] = NULL;
822         count = 0;
823         i = d0_idx;
824         do {
825                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
826
827                 blocks[slot] = sh->dev[i].page;
828
829                 if (i == target)
830                         faila = slot;
831                 if (i == target2)
832                         failb = slot;
833                 i = raid6_next_disk(i, disks);
834         } while (i != d0_idx);
835
836         BUG_ON(faila == failb);
837         if (failb < faila)
838                 swap(faila, failb);
839         pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
840                  __func__, (unsigned long long)sh->sector, faila, failb);
841
842         atomic_inc(&sh->count);
843
844         if (failb == syndrome_disks+1) {
845                 /* Q disk is one of the missing disks */
846                 if (faila == syndrome_disks) {
847                         /* Missing P+Q, just recompute */
848                         init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
849                                           ops_complete_compute, sh,
850                                           to_addr_conv(sh, percpu));
851                         return async_gen_syndrome(blocks, 0, syndrome_disks+2,
852                                                   STRIPE_SIZE, &submit);
853                 } else {
854                         struct page *dest;
855                         int data_target;
856                         int qd_idx = sh->qd_idx;
857
858                         /* Missing D+Q: recompute D from P, then recompute Q */
859                         if (target == qd_idx)
860                                 data_target = target2;
861                         else
862                                 data_target = target;
863
864                         count = 0;
865                         for (i = disks; i-- ; ) {
866                                 if (i == data_target || i == qd_idx)
867                                         continue;
868                                 blocks[count++] = sh->dev[i].page;
869                         }
870                         dest = sh->dev[data_target].page;
871                         init_async_submit(&submit,
872                                           ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
873                                           NULL, NULL, NULL,
874                                           to_addr_conv(sh, percpu));
875                         tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
876                                        &submit);
877
878                         count = set_syndrome_sources(blocks, sh);
879                         init_async_submit(&submit, ASYNC_TX_FENCE, tx,
880                                           ops_complete_compute, sh,
881                                           to_addr_conv(sh, percpu));
882                         return async_gen_syndrome(blocks, 0, count+2,
883                                                   STRIPE_SIZE, &submit);
884                 }
885         } else {
886                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
887                                   ops_complete_compute, sh,
888                                   to_addr_conv(sh, percpu));
889                 if (failb == syndrome_disks) {
890                         /* We're missing D+P. */
891                         return async_raid6_datap_recov(syndrome_disks+2,
892                                                        STRIPE_SIZE, faila,
893                                                        blocks, &submit);
894                 } else {
895                         /* We're missing D+D. */
896                         return async_raid6_2data_recov(syndrome_disks+2,
897                                                        STRIPE_SIZE, faila, failb,
898                                                        blocks, &submit);
899                 }
900         }
901 }
902
903
904 static void ops_complete_prexor(void *stripe_head_ref)
905 {
906         struct stripe_head *sh = stripe_head_ref;
907
908         pr_debug("%s: stripe %llu\n", __func__,
909                 (unsigned long long)sh->sector);
910 }
911
912 static struct dma_async_tx_descriptor *
913 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
914                struct dma_async_tx_descriptor *tx)
915 {
916         int disks = sh->disks;
917         struct page **xor_srcs = percpu->scribble;
918         int count = 0, pd_idx = sh->pd_idx, i;
919         struct async_submit_ctl submit;
920
921         /* existing parity data subtracted */
922         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
923
924         pr_debug("%s: stripe %llu\n", __func__,
925                 (unsigned long long)sh->sector);
926
927         for (i = disks; i--; ) {
928                 struct r5dev *dev = &sh->dev[i];
929                 /* Only process blocks that are known to be uptodate */
930                 if (test_bit(R5_Wantdrain, &dev->flags))
931                         xor_srcs[count++] = dev->page;
932         }
933
934         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
935                           ops_complete_prexor, sh, to_addr_conv(sh, percpu));
936         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
937
938         return tx;
939 }
940
941 static struct dma_async_tx_descriptor *
942 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
943 {
944         int disks = sh->disks;
945         int i;
946
947         pr_debug("%s: stripe %llu\n", __func__,
948                 (unsigned long long)sh->sector);
949
950         for (i = disks; i--; ) {
951                 struct r5dev *dev = &sh->dev[i];
952                 struct bio *chosen;
953
954                 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
955                         struct bio *wbi;
956
957                         spin_lock(&sh->lock);
958                         chosen = dev->towrite;
959                         dev->towrite = NULL;
960                         BUG_ON(dev->written);
961                         wbi = dev->written = chosen;
962                         spin_unlock(&sh->lock);
963
964                         while (wbi && wbi->bi_sector <
965                                 dev->sector + STRIPE_SECTORS) {
966                                 tx = async_copy_data(1, wbi, dev->page,
967                                         dev->sector, tx);
968                                 wbi = r5_next_bio(wbi, dev->sector);
969                         }
970                 }
971         }
972
973         return tx;
974 }
975
976 static void ops_complete_reconstruct(void *stripe_head_ref)
977 {
978         struct stripe_head *sh = stripe_head_ref;
979         int disks = sh->disks;
980         int pd_idx = sh->pd_idx;
981         int qd_idx = sh->qd_idx;
982         int i;
983
984         pr_debug("%s: stripe %llu\n", __func__,
985                 (unsigned long long)sh->sector);
986
987         for (i = disks; i--; ) {
988                 struct r5dev *dev = &sh->dev[i];
989
990                 if (dev->written || i == pd_idx || i == qd_idx)
991                         set_bit(R5_UPTODATE, &dev->flags);
992         }
993
994         if (sh->reconstruct_state == reconstruct_state_drain_run)
995                 sh->reconstruct_state = reconstruct_state_drain_result;
996         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
997                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
998         else {
999                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1000                 sh->reconstruct_state = reconstruct_state_result;
1001         }
1002
1003         set_bit(STRIPE_HANDLE, &sh->state);
1004         release_stripe(sh);
1005 }
1006
1007 static void
1008 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1009                      struct dma_async_tx_descriptor *tx)
1010 {
1011         int disks = sh->disks;
1012         struct page **xor_srcs = percpu->scribble;
1013         struct async_submit_ctl submit;
1014         int count = 0, pd_idx = sh->pd_idx, i;
1015         struct page *xor_dest;
1016         int prexor = 0;
1017         unsigned long flags;
1018
1019         pr_debug("%s: stripe %llu\n", __func__,
1020                 (unsigned long long)sh->sector);
1021
1022         /* check if prexor is active which means only process blocks
1023          * that are part of a read-modify-write (written)
1024          */
1025         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1026                 prexor = 1;
1027                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1028                 for (i = disks; i--; ) {
1029                         struct r5dev *dev = &sh->dev[i];
1030                         if (dev->written)
1031                                 xor_srcs[count++] = dev->page;
1032                 }
1033         } else {
1034                 xor_dest = sh->dev[pd_idx].page;
1035                 for (i = disks; i--; ) {
1036                         struct r5dev *dev = &sh->dev[i];
1037                         if (i != pd_idx)
1038                                 xor_srcs[count++] = dev->page;
1039                 }
1040         }
1041
1042         /* 1/ if we prexor'd then the dest is reused as a source
1043          * 2/ if we did not prexor then we are redoing the parity
1044          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1045          * for the synchronous xor case
1046          */
1047         flags = ASYNC_TX_ACK |
1048                 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1049
1050         atomic_inc(&sh->count);
1051
1052         init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1053                           to_addr_conv(sh, percpu));
1054         if (unlikely(count == 1))
1055                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1056         else
1057                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1058 }
1059
1060 static void
1061 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1062                      struct dma_async_tx_descriptor *tx)
1063 {
1064         struct async_submit_ctl submit;
1065         struct page **blocks = percpu->scribble;
1066         int count;
1067
1068         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1069
1070         count = set_syndrome_sources(blocks, sh);
1071
1072         atomic_inc(&sh->count);
1073
1074         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1075                           sh, to_addr_conv(sh, percpu));
1076         async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1077 }
1078
1079 static void ops_complete_check(void *stripe_head_ref)
1080 {
1081         struct stripe_head *sh = stripe_head_ref;
1082
1083         pr_debug("%s: stripe %llu\n", __func__,
1084                 (unsigned long long)sh->sector);
1085
1086         sh->check_state = check_state_check_result;
1087         set_bit(STRIPE_HANDLE, &sh->state);
1088         release_stripe(sh);
1089 }
1090
1091 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1092 {
1093         int disks = sh->disks;
1094         int pd_idx = sh->pd_idx;
1095         int qd_idx = sh->qd_idx;
1096         struct page *xor_dest;
1097         struct page **xor_srcs = percpu->scribble;
1098         struct dma_async_tx_descriptor *tx;
1099         struct async_submit_ctl submit;
1100         int count;
1101         int i;
1102
1103         pr_debug("%s: stripe %llu\n", __func__,
1104                 (unsigned long long)sh->sector);
1105
1106         count = 0;
1107         xor_dest = sh->dev[pd_idx].page;
1108         xor_srcs[count++] = xor_dest;
1109         for (i = disks; i--; ) {
1110                 if (i == pd_idx || i == qd_idx)
1111                         continue;
1112                 xor_srcs[count++] = sh->dev[i].page;
1113         }
1114
1115         init_async_submit(&submit, 0, NULL, NULL, NULL,
1116                           to_addr_conv(sh, percpu));
1117         tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1118                            &sh->ops.zero_sum_result, &submit);
1119
1120         atomic_inc(&sh->count);
1121         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1122         tx = async_trigger_callback(&submit);
1123 }
1124
1125 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1126 {
1127         struct page **srcs = percpu->scribble;
1128         struct async_submit_ctl submit;
1129         int count;
1130
1131         pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1132                 (unsigned long long)sh->sector, checkp);
1133
1134         count = set_syndrome_sources(srcs, sh);
1135         if (!checkp)
1136                 srcs[count] = NULL;
1137
1138         atomic_inc(&sh->count);
1139         init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1140                           sh, to_addr_conv(sh, percpu));
1141         async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1142                            &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1143 }
1144
1145 static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1146 {
1147         int overlap_clear = 0, i, disks = sh->disks;
1148         struct dma_async_tx_descriptor *tx = NULL;
1149         raid5_conf_t *conf = sh->raid_conf;
1150         int level = conf->level;
1151         struct raid5_percpu *percpu;
1152         unsigned long cpu;
1153
1154         cpu = get_cpu();
1155         percpu = per_cpu_ptr(conf->percpu, cpu);
1156         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1157                 ops_run_biofill(sh);
1158                 overlap_clear++;
1159         }
1160
1161         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1162                 if (level < 6)
1163                         tx = ops_run_compute5(sh, percpu);
1164                 else {
1165                         if (sh->ops.target2 < 0 || sh->ops.target < 0)
1166                                 tx = ops_run_compute6_1(sh, percpu);
1167                         else
1168                                 tx = ops_run_compute6_2(sh, percpu);
1169                 }
1170                 /* terminate the chain if reconstruct is not set to be run */
1171                 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1172                         async_tx_ack(tx);
1173         }
1174
1175         if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1176                 tx = ops_run_prexor(sh, percpu, tx);
1177
1178         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1179                 tx = ops_run_biodrain(sh, tx);
1180                 overlap_clear++;
1181         }
1182
1183         if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1184                 if (level < 6)
1185                         ops_run_reconstruct5(sh, percpu, tx);
1186                 else
1187                         ops_run_reconstruct6(sh, percpu, tx);
1188         }
1189
1190         if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1191                 if (sh->check_state == check_state_run)
1192                         ops_run_check_p(sh, percpu);
1193                 else if (sh->check_state == check_state_run_q)
1194                         ops_run_check_pq(sh, percpu, 0);
1195                 else if (sh->check_state == check_state_run_pq)
1196                         ops_run_check_pq(sh, percpu, 1);
1197                 else
1198                         BUG();
1199         }
1200
1201         if (overlap_clear)
1202                 for (i = disks; i--; ) {
1203                         struct r5dev *dev = &sh->dev[i];
1204                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
1205                                 wake_up(&sh->raid_conf->wait_for_overlap);
1206                 }
1207         put_cpu();
1208 }
1209
1210 #ifdef CONFIG_MULTICORE_RAID456
1211 static void async_run_ops(void *param, async_cookie_t cookie)
1212 {
1213         struct stripe_head *sh = param;
1214         unsigned long ops_request = sh->ops.request;
1215
1216         clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1217         wake_up(&sh->ops.wait_for_ops);
1218
1219         __raid_run_ops(sh, ops_request);
1220         release_stripe(sh);
1221 }
1222
1223 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1224 {
1225         /* since handle_stripe can be called outside of raid5d context
1226          * we need to ensure sh->ops.request is de-staged before another
1227          * request arrives
1228          */
1229         wait_event(sh->ops.wait_for_ops,
1230                    !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1231         sh->ops.request = ops_request;
1232
1233         atomic_inc(&sh->count);
1234         async_schedule(async_run_ops, sh);
1235 }
1236 #else
1237 #define raid_run_ops __raid_run_ops
1238 #endif
1239
1240 static int grow_one_stripe(raid5_conf_t *conf)
1241 {
1242         struct stripe_head *sh;
1243         int disks = max(conf->raid_disks, conf->previous_raid_disks);
1244         sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
1245         if (!sh)
1246                 return 0;
1247         memset(sh, 0, sizeof(*sh) + (disks-1)*sizeof(struct r5dev));
1248         sh->raid_conf = conf;
1249         spin_lock_init(&sh->lock);
1250         #ifdef CONFIG_MULTICORE_RAID456
1251         init_waitqueue_head(&sh->ops.wait_for_ops);
1252         #endif
1253
1254         if (grow_buffers(sh, disks)) {
1255                 shrink_buffers(sh, disks);
1256                 kmem_cache_free(conf->slab_cache, sh);
1257                 return 0;
1258         }
1259         /* we just created an active stripe so... */
1260         atomic_set(&sh->count, 1);
1261         atomic_inc(&conf->active_stripes);
1262         INIT_LIST_HEAD(&sh->lru);
1263         release_stripe(sh);
1264         return 1;
1265 }
1266
1267 static int grow_stripes(raid5_conf_t *conf, int num)
1268 {
1269         struct kmem_cache *sc;
1270         int devs = max(conf->raid_disks, conf->previous_raid_disks);
1271
1272         sprintf(conf->cache_name[0],
1273                 "raid%d-%s", conf->level, mdname(conf->mddev));
1274         sprintf(conf->cache_name[1],
1275                 "raid%d-%s-alt", conf->level, mdname(conf->mddev));
1276         conf->active_name = 0;
1277         sc = kmem_cache_create(conf->cache_name[conf->active_name],
1278                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1279                                0, 0, NULL);
1280         if (!sc)
1281                 return 1;
1282         conf->slab_cache = sc;
1283         conf->pool_size = devs;
1284         while (num--)
1285                 if (!grow_one_stripe(conf))
1286                         return 1;
1287         return 0;
1288 }
1289
1290 /**
1291  * scribble_len - return the required size of the scribble region
1292  * @num - total number of disks in the array
1293  *
1294  * The size must be enough to contain:
1295  * 1/ a struct page pointer for each device in the array +2
1296  * 2/ room to convert each entry in (1) to its corresponding dma
1297  *    (dma_map_page()) or page (page_address()) address.
1298  *
1299  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1300  * calculate over all devices (not just the data blocks), using zeros in place
1301  * of the P and Q blocks.
1302  */
1303 static size_t scribble_len(int num)
1304 {
1305         size_t len;
1306
1307         len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1308
1309         return len;
1310 }
1311
1312 static int resize_stripes(raid5_conf_t *conf, int newsize)
1313 {
1314         /* Make all the stripes able to hold 'newsize' devices.
1315          * New slots in each stripe get 'page' set to a new page.
1316          *
1317          * This happens in stages:
1318          * 1/ create a new kmem_cache and allocate the required number of
1319          *    stripe_heads.
1320          * 2/ gather all the old stripe_heads and tranfer the pages across
1321          *    to the new stripe_heads.  This will have the side effect of
1322          *    freezing the array as once all stripe_heads have been collected,
1323          *    no IO will be possible.  Old stripe heads are freed once their
1324          *    pages have been transferred over, and the old kmem_cache is
1325          *    freed when all stripes are done.
1326          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
1327          *    we simple return a failre status - no need to clean anything up.
1328          * 4/ allocate new pages for the new slots in the new stripe_heads.
1329          *    If this fails, we don't bother trying the shrink the
1330          *    stripe_heads down again, we just leave them as they are.
1331          *    As each stripe_head is processed the new one is released into
1332          *    active service.
1333          *
1334          * Once step2 is started, we cannot afford to wait for a write,
1335          * so we use GFP_NOIO allocations.
1336          */
1337         struct stripe_head *osh, *nsh;
1338         LIST_HEAD(newstripes);
1339         struct disk_info *ndisks;
1340         unsigned long cpu;
1341         int err;
1342         struct kmem_cache *sc;
1343         int i;
1344
1345         if (newsize <= conf->pool_size)
1346                 return 0; /* never bother to shrink */
1347
1348         err = md_allow_write(conf->mddev);
1349         if (err)
1350                 return err;
1351
1352         /* Step 1 */
1353         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1354                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1355                                0, 0, NULL);
1356         if (!sc)
1357                 return -ENOMEM;
1358
1359         for (i = conf->max_nr_stripes; i; i--) {
1360                 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1361                 if (!nsh)
1362                         break;
1363
1364                 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1365
1366                 nsh->raid_conf = conf;
1367                 spin_lock_init(&nsh->lock);
1368                 #ifdef CONFIG_MULTICORE_RAID456
1369                 init_waitqueue_head(&nsh->ops.wait_for_ops);
1370                 #endif
1371
1372                 list_add(&nsh->lru, &newstripes);
1373         }
1374         if (i) {
1375                 /* didn't get enough, give up */
1376                 while (!list_empty(&newstripes)) {
1377                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
1378                         list_del(&nsh->lru);
1379                         kmem_cache_free(sc, nsh);
1380                 }
1381                 kmem_cache_destroy(sc);
1382                 return -ENOMEM;
1383         }
1384         /* Step 2 - Must use GFP_NOIO now.
1385          * OK, we have enough stripes, start collecting inactive
1386          * stripes and copying them over
1387          */
1388         list_for_each_entry(nsh, &newstripes, lru) {
1389                 spin_lock_irq(&conf->device_lock);
1390                 wait_event_lock_irq(conf->wait_for_stripe,
1391                                     !list_empty(&conf->inactive_list),
1392                                     conf->device_lock,
1393                                     unplug_slaves(conf->mddev)
1394                         );
1395                 osh = get_free_stripe(conf);
1396                 spin_unlock_irq(&conf->device_lock);
1397                 atomic_set(&nsh->count, 1);
1398                 for(i=0; i<conf->pool_size; i++)
1399                         nsh->dev[i].page = osh->dev[i].page;
1400                 for( ; i<newsize; i++)
1401                         nsh->dev[i].page = NULL;
1402                 kmem_cache_free(conf->slab_cache, osh);
1403         }
1404         kmem_cache_destroy(conf->slab_cache);
1405
1406         /* Step 3.
1407          * At this point, we are holding all the stripes so the array
1408          * is completely stalled, so now is a good time to resize
1409          * conf->disks and the scribble region
1410          */
1411         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1412         if (ndisks) {
1413                 for (i=0; i<conf->raid_disks; i++)
1414                         ndisks[i] = conf->disks[i];
1415                 kfree(conf->disks);
1416                 conf->disks = ndisks;
1417         } else
1418                 err = -ENOMEM;
1419
1420         get_online_cpus();
1421         conf->scribble_len = scribble_len(newsize);
1422         for_each_present_cpu(cpu) {
1423                 struct raid5_percpu *percpu;
1424                 void *scribble;
1425
1426                 percpu = per_cpu_ptr(conf->percpu, cpu);
1427                 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1428
1429                 if (scribble) {
1430                         kfree(percpu->scribble);
1431                         percpu->scribble = scribble;
1432                 } else {
1433                         err = -ENOMEM;
1434                         break;
1435                 }
1436         }
1437         put_online_cpus();
1438
1439         /* Step 4, return new stripes to service */
1440         while(!list_empty(&newstripes)) {
1441                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1442                 list_del_init(&nsh->lru);
1443
1444                 for (i=conf->raid_disks; i < newsize; i++)
1445                         if (nsh->dev[i].page == NULL) {
1446                                 struct page *p = alloc_page(GFP_NOIO);
1447                                 nsh->dev[i].page = p;
1448                                 if (!p)
1449                                         err = -ENOMEM;
1450                         }
1451                 release_stripe(nsh);
1452         }
1453         /* critical section pass, GFP_NOIO no longer needed */
1454
1455         conf->slab_cache = sc;
1456         conf->active_name = 1-conf->active_name;
1457         conf->pool_size = newsize;
1458         return err;
1459 }
1460
1461 static int drop_one_stripe(raid5_conf_t *conf)
1462 {
1463         struct stripe_head *sh;
1464
1465         spin_lock_irq(&conf->device_lock);
1466         sh = get_free_stripe(conf);
1467         spin_unlock_irq(&conf->device_lock);
1468         if (!sh)
1469                 return 0;
1470         BUG_ON(atomic_read(&sh->count));
1471         shrink_buffers(sh, conf->pool_size);
1472         kmem_cache_free(conf->slab_cache, sh);
1473         atomic_dec(&conf->active_stripes);
1474         return 1;
1475 }
1476
1477 static void shrink_stripes(raid5_conf_t *conf)
1478 {
1479         while (drop_one_stripe(conf))
1480                 ;
1481
1482         if (conf->slab_cache)
1483                 kmem_cache_destroy(conf->slab_cache);
1484         conf->slab_cache = NULL;
1485 }
1486
1487 static void raid5_end_read_request(struct bio * bi, int error)
1488 {
1489         struct stripe_head *sh = bi->bi_private;
1490         raid5_conf_t *conf = sh->raid_conf;
1491         int disks = sh->disks, i;
1492         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1493         char b[BDEVNAME_SIZE];
1494         mdk_rdev_t *rdev;
1495
1496
1497         for (i=0 ; i<disks; i++)
1498                 if (bi == &sh->dev[i].req)
1499                         break;
1500
1501         pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1502                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1503                 uptodate);
1504         if (i == disks) {
1505                 BUG();
1506                 return;
1507         }
1508
1509         if (uptodate) {
1510                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1511                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1512                         rdev = conf->disks[i].rdev;
1513                         printk_rl(KERN_INFO "md/raid:%s: read error corrected"
1514                                   " (%lu sectors at %llu on %s)\n",
1515                                   mdname(conf->mddev), STRIPE_SECTORS,
1516                                   (unsigned long long)(sh->sector
1517                                                        + rdev->data_offset),
1518                                   bdevname(rdev->bdev, b));
1519                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1520                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1521                 }
1522                 if (atomic_read(&conf->disks[i].rdev->read_errors))
1523                         atomic_set(&conf->disks[i].rdev->read_errors, 0);
1524         } else {
1525                 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1526                 int retry = 0;
1527                 rdev = conf->disks[i].rdev;
1528
1529                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1530                 atomic_inc(&rdev->read_errors);
1531                 if (conf->mddev->degraded >= conf->max_degraded)
1532                         printk_rl(KERN_WARNING
1533                                   "md/raid:%s: read error not correctable "
1534                                   "(sector %llu on %s).\n",
1535                                   mdname(conf->mddev),
1536                                   (unsigned long long)(sh->sector
1537                                                        + rdev->data_offset),
1538                                   bdn);
1539                 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1540                         /* Oh, no!!! */
1541                         printk_rl(KERN_WARNING
1542                                   "md/raid:%s: read error NOT corrected!! "
1543                                   "(sector %llu on %s).\n",
1544                                   mdname(conf->mddev),
1545                                   (unsigned long long)(sh->sector
1546                                                        + rdev->data_offset),
1547                                   bdn);
1548                 else if (atomic_read(&rdev->read_errors)
1549                          > conf->max_nr_stripes)
1550                         printk(KERN_WARNING
1551                                "md/raid:%s: Too many read errors, failing device %s.\n",
1552                                mdname(conf->mddev), bdn);
1553                 else
1554                         retry = 1;
1555                 if (retry)
1556                         set_bit(R5_ReadError, &sh->dev[i].flags);
1557                 else {
1558                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1559                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1560                         md_error(conf->mddev, rdev);
1561                 }
1562         }
1563         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1564         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1565         set_bit(STRIPE_HANDLE, &sh->state);
1566         release_stripe(sh);
1567 }
1568
1569 static void raid5_end_write_request(struct bio *bi, int error)
1570 {
1571         struct stripe_head *sh = bi->bi_private;
1572         raid5_conf_t *conf = sh->raid_conf;
1573         int disks = sh->disks, i;
1574         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1575
1576         for (i=0 ; i<disks; i++)
1577                 if (bi == &sh->dev[i].req)
1578                         break;
1579
1580         pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1581                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1582                 uptodate);
1583         if (i == disks) {
1584                 BUG();
1585                 return;
1586         }
1587
1588         if (!uptodate)
1589                 md_error(conf->mddev, conf->disks[i].rdev);
1590
1591         rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1592         
1593         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1594         set_bit(STRIPE_HANDLE, &sh->state);
1595         release_stripe(sh);
1596 }
1597
1598
1599 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1600         
1601 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1602 {
1603         struct r5dev *dev = &sh->dev[i];
1604
1605         bio_init(&dev->req);
1606         dev->req.bi_io_vec = &dev->vec;
1607         dev->req.bi_vcnt++;
1608         dev->req.bi_max_vecs++;
1609         dev->vec.bv_page = dev->page;
1610         dev->vec.bv_len = STRIPE_SIZE;
1611         dev->vec.bv_offset = 0;
1612
1613         dev->req.bi_sector = sh->sector;
1614         dev->req.bi_private = sh;
1615
1616         dev->flags = 0;
1617         dev->sector = compute_blocknr(sh, i, previous);
1618 }
1619
1620 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1621 {
1622         char b[BDEVNAME_SIZE];
1623         raid5_conf_t *conf = mddev->private;
1624         pr_debug("raid456: error called\n");
1625
1626         if (!test_bit(Faulty, &rdev->flags)) {
1627                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1628                 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1629                         unsigned long flags;
1630                         spin_lock_irqsave(&conf->device_lock, flags);
1631                         mddev->degraded++;
1632                         spin_unlock_irqrestore(&conf->device_lock, flags);
1633                         /*
1634                          * if recovery was running, make sure it aborts.
1635                          */
1636                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1637                 }
1638                 set_bit(Faulty, &rdev->flags);
1639                 printk(KERN_ALERT
1640                        "md/raid:%s: Disk failure on %s, disabling device.\n"
1641                        KERN_ALERT
1642                        "md/raid:%s: Operation continuing on %d devices.\n",
1643                        mdname(mddev),
1644                        bdevname(rdev->bdev, b),
1645                        mdname(mddev),
1646                        conf->raid_disks - mddev->degraded);
1647         }
1648 }
1649
1650 /*
1651  * Input: a 'big' sector number,
1652  * Output: index of the data and parity disk, and the sector # in them.
1653  */
1654 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1655                                      int previous, int *dd_idx,
1656                                      struct stripe_head *sh)
1657 {
1658         sector_t stripe, stripe2;
1659         sector_t chunk_number;
1660         unsigned int chunk_offset;
1661         int pd_idx, qd_idx;
1662         int ddf_layout = 0;
1663         sector_t new_sector;
1664         int algorithm = previous ? conf->prev_algo
1665                                  : conf->algorithm;
1666         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1667                                          : conf->chunk_sectors;
1668         int raid_disks = previous ? conf->previous_raid_disks
1669                                   : conf->raid_disks;
1670         int data_disks = raid_disks - conf->max_degraded;
1671
1672         /* First compute the information on this sector */
1673
1674         /*
1675          * Compute the chunk number and the sector offset inside the chunk
1676          */
1677         chunk_offset = sector_div(r_sector, sectors_per_chunk);
1678         chunk_number = r_sector;
1679
1680         /*
1681          * Compute the stripe number
1682          */
1683         stripe = chunk_number;
1684         *dd_idx = sector_div(stripe, data_disks);
1685         stripe2 = stripe;
1686         /*
1687          * Select the parity disk based on the user selected algorithm.
1688          */
1689         pd_idx = qd_idx = ~0;
1690         switch(conf->level) {
1691         case 4:
1692                 pd_idx = data_disks;
1693                 break;
1694         case 5:
1695                 switch (algorithm) {
1696                 case ALGORITHM_LEFT_ASYMMETRIC:
1697                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
1698                         if (*dd_idx >= pd_idx)
1699                                 (*dd_idx)++;
1700                         break;
1701                 case ALGORITHM_RIGHT_ASYMMETRIC:
1702                         pd_idx = sector_div(stripe2, raid_disks);
1703                         if (*dd_idx >= pd_idx)
1704                                 (*dd_idx)++;
1705                         break;
1706                 case ALGORITHM_LEFT_SYMMETRIC:
1707                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
1708                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1709                         break;
1710                 case ALGORITHM_RIGHT_SYMMETRIC:
1711                         pd_idx = sector_div(stripe2, raid_disks);
1712                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1713                         break;
1714                 case ALGORITHM_PARITY_0:
1715                         pd_idx = 0;
1716                         (*dd_idx)++;
1717                         break;
1718                 case ALGORITHM_PARITY_N:
1719                         pd_idx = data_disks;
1720                         break;
1721                 default:
1722                         BUG();
1723                 }
1724                 break;
1725         case 6:
1726
1727                 switch (algorithm) {
1728                 case ALGORITHM_LEFT_ASYMMETRIC:
1729                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1730                         qd_idx = pd_idx + 1;
1731                         if (pd_idx == raid_disks-1) {
1732                                 (*dd_idx)++;    /* Q D D D P */
1733                                 qd_idx = 0;
1734                         } else if (*dd_idx >= pd_idx)
1735                                 (*dd_idx) += 2; /* D D P Q D */
1736                         break;
1737                 case ALGORITHM_RIGHT_ASYMMETRIC:
1738                         pd_idx = sector_div(stripe2, raid_disks);
1739                         qd_idx = pd_idx + 1;
1740                         if (pd_idx == raid_disks-1) {
1741                                 (*dd_idx)++;    /* Q D D D P */
1742                                 qd_idx = 0;
1743                         } else if (*dd_idx >= pd_idx)
1744                                 (*dd_idx) += 2; /* D D P Q D */
1745                         break;
1746                 case ALGORITHM_LEFT_SYMMETRIC:
1747                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1748                         qd_idx = (pd_idx + 1) % raid_disks;
1749                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1750                         break;
1751                 case ALGORITHM_RIGHT_SYMMETRIC:
1752                         pd_idx = sector_div(stripe2, raid_disks);
1753                         qd_idx = (pd_idx + 1) % raid_disks;
1754                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1755                         break;
1756
1757                 case ALGORITHM_PARITY_0:
1758                         pd_idx = 0;
1759                         qd_idx = 1;
1760                         (*dd_idx) += 2;
1761                         break;
1762                 case ALGORITHM_PARITY_N:
1763                         pd_idx = data_disks;
1764                         qd_idx = data_disks + 1;
1765                         break;
1766
1767                 case ALGORITHM_ROTATING_ZERO_RESTART:
1768                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
1769                          * of blocks for computing Q is different.
1770                          */
1771                         pd_idx = sector_div(stripe2, raid_disks);
1772                         qd_idx = pd_idx + 1;
1773                         if (pd_idx == raid_disks-1) {
1774                                 (*dd_idx)++;    /* Q D D D P */
1775                                 qd_idx = 0;
1776                         } else if (*dd_idx >= pd_idx)
1777                                 (*dd_idx) += 2; /* D D P Q D */
1778                         ddf_layout = 1;
1779                         break;
1780
1781                 case ALGORITHM_ROTATING_N_RESTART:
1782                         /* Same a left_asymmetric, by first stripe is
1783                          * D D D P Q  rather than
1784                          * Q D D D P
1785                          */
1786                         stripe2 += 1;
1787                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1788                         qd_idx = pd_idx + 1;
1789                         if (pd_idx == raid_disks-1) {
1790                                 (*dd_idx)++;    /* Q D D D P */
1791                                 qd_idx = 0;
1792                         } else if (*dd_idx >= pd_idx)
1793                                 (*dd_idx) += 2; /* D D P Q D */
1794                         ddf_layout = 1;
1795                         break;
1796
1797                 case ALGORITHM_ROTATING_N_CONTINUE:
1798                         /* Same as left_symmetric but Q is before P */
1799                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1800                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1801                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1802                         ddf_layout = 1;
1803                         break;
1804
1805                 case ALGORITHM_LEFT_ASYMMETRIC_6:
1806                         /* RAID5 left_asymmetric, with Q on last device */
1807                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1808                         if (*dd_idx >= pd_idx)
1809                                 (*dd_idx)++;
1810                         qd_idx = raid_disks - 1;
1811                         break;
1812
1813                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1814                         pd_idx = sector_div(stripe2, raid_disks-1);
1815                         if (*dd_idx >= pd_idx)
1816                                 (*dd_idx)++;
1817                         qd_idx = raid_disks - 1;
1818                         break;
1819
1820                 case ALGORITHM_LEFT_SYMMETRIC_6:
1821                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1822                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1823                         qd_idx = raid_disks - 1;
1824                         break;
1825
1826                 case ALGORITHM_RIGHT_SYMMETRIC_6:
1827                         pd_idx = sector_div(stripe2, raid_disks-1);
1828                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1829                         qd_idx = raid_disks - 1;
1830                         break;
1831
1832                 case ALGORITHM_PARITY_0_6:
1833                         pd_idx = 0;
1834                         (*dd_idx)++;
1835                         qd_idx = raid_disks - 1;
1836                         break;
1837
1838                 default:
1839                         BUG();
1840                 }
1841                 break;
1842         }
1843
1844         if (sh) {
1845                 sh->pd_idx = pd_idx;
1846                 sh->qd_idx = qd_idx;
1847                 sh->ddf_layout = ddf_layout;
1848         }
1849         /*
1850          * Finally, compute the new sector number
1851          */
1852         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1853         return new_sector;
1854 }
1855
1856
1857 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1858 {
1859         raid5_conf_t *conf = sh->raid_conf;
1860         int raid_disks = sh->disks;
1861         int data_disks = raid_disks - conf->max_degraded;
1862         sector_t new_sector = sh->sector, check;
1863         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1864                                          : conf->chunk_sectors;
1865         int algorithm = previous ? conf->prev_algo
1866                                  : conf->algorithm;
1867         sector_t stripe;
1868         int chunk_offset;
1869         sector_t chunk_number;
1870         int dummy1, dd_idx = i;
1871         sector_t r_sector;
1872         struct stripe_head sh2;
1873
1874
1875         chunk_offset = sector_div(new_sector, sectors_per_chunk);
1876         stripe = new_sector;
1877
1878         if (i == sh->pd_idx)
1879                 return 0;
1880         switch(conf->level) {
1881         case 4: break;
1882         case 5:
1883                 switch (algorithm) {
1884                 case ALGORITHM_LEFT_ASYMMETRIC:
1885                 case ALGORITHM_RIGHT_ASYMMETRIC:
1886                         if (i > sh->pd_idx)
1887                                 i--;
1888                         break;
1889                 case ALGORITHM_LEFT_SYMMETRIC:
1890                 case ALGORITHM_RIGHT_SYMMETRIC:
1891                         if (i < sh->pd_idx)
1892                                 i += raid_disks;
1893                         i -= (sh->pd_idx + 1);
1894                         break;
1895                 case ALGORITHM_PARITY_0:
1896                         i -= 1;
1897                         break;
1898                 case ALGORITHM_PARITY_N:
1899                         break;
1900                 default:
1901                         BUG();
1902                 }
1903                 break;
1904         case 6:
1905                 if (i == sh->qd_idx)
1906                         return 0; /* It is the Q disk */
1907                 switch (algorithm) {
1908                 case ALGORITHM_LEFT_ASYMMETRIC:
1909                 case ALGORITHM_RIGHT_ASYMMETRIC:
1910                 case ALGORITHM_ROTATING_ZERO_RESTART:
1911                 case ALGORITHM_ROTATING_N_RESTART:
1912                         if (sh->pd_idx == raid_disks-1)
1913                                 i--;    /* Q D D D P */
1914                         else if (i > sh->pd_idx)
1915                                 i -= 2; /* D D P Q D */
1916                         break;
1917                 case ALGORITHM_LEFT_SYMMETRIC:
1918                 case ALGORITHM_RIGHT_SYMMETRIC:
1919                         if (sh->pd_idx == raid_disks-1)
1920                                 i--; /* Q D D D P */
1921                         else {
1922                                 /* D D P Q D */
1923                                 if (i < sh->pd_idx)
1924                                         i += raid_disks;
1925                                 i -= (sh->pd_idx + 2);
1926                         }
1927                         break;
1928                 case ALGORITHM_PARITY_0:
1929                         i -= 2;
1930                         break;
1931                 case ALGORITHM_PARITY_N:
1932                         break;
1933                 case ALGORITHM_ROTATING_N_CONTINUE:
1934                         /* Like left_symmetric, but P is before Q */
1935                         if (sh->pd_idx == 0)
1936                                 i--;    /* P D D D Q */
1937                         else {
1938                                 /* D D Q P D */
1939                                 if (i < sh->pd_idx)
1940                                         i += raid_disks;
1941                                 i -= (sh->pd_idx + 1);
1942                         }
1943                         break;
1944                 case ALGORITHM_LEFT_ASYMMETRIC_6:
1945                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1946                         if (i > sh->pd_idx)
1947                                 i--;
1948                         break;
1949                 case ALGORITHM_LEFT_SYMMETRIC_6:
1950                 case ALGORITHM_RIGHT_SYMMETRIC_6:
1951                         if (i < sh->pd_idx)
1952                                 i += data_disks + 1;
1953                         i -= (sh->pd_idx + 1);
1954                         break;
1955                 case ALGORITHM_PARITY_0_6:
1956                         i -= 1;
1957                         break;
1958                 default:
1959                         BUG();
1960                 }
1961                 break;
1962         }
1963
1964         chunk_number = stripe * data_disks + i;
1965         r_sector = chunk_number * sectors_per_chunk + chunk_offset;
1966
1967         check = raid5_compute_sector(conf, r_sector,
1968                                      previous, &dummy1, &sh2);
1969         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
1970                 || sh2.qd_idx != sh->qd_idx) {
1971                 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
1972                        mdname(conf->mddev));
1973                 return 0;
1974         }
1975         return r_sector;
1976 }
1977
1978
1979 static void
1980 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
1981                          int rcw, int expand)
1982 {
1983         int i, pd_idx = sh->pd_idx, disks = sh->disks;
1984         raid5_conf_t *conf = sh->raid_conf;
1985         int level = conf->level;
1986
1987         if (rcw) {
1988                 /* if we are not expanding this is a proper write request, and
1989                  * there will be bios with new data to be drained into the
1990                  * stripe cache
1991                  */
1992                 if (!expand) {
1993                         sh->reconstruct_state = reconstruct_state_drain_run;
1994                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1995                 } else
1996                         sh->reconstruct_state = reconstruct_state_run;
1997
1998                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
1999
2000                 for (i = disks; i--; ) {
2001                         struct r5dev *dev = &sh->dev[i];
2002
2003                         if (dev->towrite) {
2004                                 set_bit(R5_LOCKED, &dev->flags);
2005                                 set_bit(R5_Wantdrain, &dev->flags);
2006                                 if (!expand)
2007                                         clear_bit(R5_UPTODATE, &dev->flags);
2008                                 s->locked++;
2009                         }
2010                 }
2011                 if (s->locked + conf->max_degraded == disks)
2012                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2013                                 atomic_inc(&conf->pending_full_writes);
2014         } else {
2015                 BUG_ON(level == 6);
2016                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2017                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2018
2019                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2020                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2021                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2022                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2023
2024                 for (i = disks; i--; ) {
2025                         struct r5dev *dev = &sh->dev[i];
2026                         if (i == pd_idx)
2027                                 continue;
2028
2029                         if (dev->towrite &&
2030                             (test_bit(R5_UPTODATE, &dev->flags) ||
2031                              test_bit(R5_Wantcompute, &dev->flags))) {
2032                                 set_bit(R5_Wantdrain, &dev->flags);
2033                                 set_bit(R5_LOCKED, &dev->flags);
2034                                 clear_bit(R5_UPTODATE, &dev->flags);
2035                                 s->locked++;
2036                         }
2037                 }
2038         }
2039
2040         /* keep the parity disk(s) locked while asynchronous operations
2041          * are in flight
2042          */
2043         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2044         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2045         s->locked++;
2046
2047         if (level == 6) {
2048                 int qd_idx = sh->qd_idx;
2049                 struct r5dev *dev = &sh->dev[qd_idx];
2050
2051                 set_bit(R5_LOCKED, &dev->flags);
2052                 clear_bit(R5_UPTODATE, &dev->flags);
2053                 s->locked++;
2054         }
2055
2056         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2057                 __func__, (unsigned long long)sh->sector,
2058                 s->locked, s->ops_request);
2059 }
2060
2061 /*
2062  * Each stripe/dev can have one or more bion attached.
2063  * toread/towrite point to the first in a chain.
2064  * The bi_next chain must be in order.
2065  */
2066 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2067 {
2068         struct bio **bip;
2069         raid5_conf_t *conf = sh->raid_conf;
2070         int firstwrite=0;
2071
2072         pr_debug("adding bh b#%llu to stripe s#%llu\n",
2073                 (unsigned long long)bi->bi_sector,
2074                 (unsigned long long)sh->sector);
2075
2076
2077         spin_lock(&sh->lock);
2078         spin_lock_irq(&conf->device_lock);
2079         if (forwrite) {
2080                 bip = &sh->dev[dd_idx].towrite;
2081                 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2082                         firstwrite = 1;
2083         } else
2084                 bip = &sh->dev[dd_idx].toread;
2085         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2086                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2087                         goto overlap;
2088                 bip = & (*bip)->bi_next;
2089         }
2090         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2091                 goto overlap;
2092
2093         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2094         if (*bip)
2095                 bi->bi_next = *bip;
2096         *bip = bi;
2097         bi->bi_phys_segments++;
2098         spin_unlock_irq(&conf->device_lock);
2099         spin_unlock(&sh->lock);
2100
2101         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2102                 (unsigned long long)bi->bi_sector,
2103                 (unsigned long long)sh->sector, dd_idx);
2104
2105         if (conf->mddev->bitmap && firstwrite) {
2106                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2107                                   STRIPE_SECTORS, 0);
2108                 sh->bm_seq = conf->seq_flush+1;
2109                 set_bit(STRIPE_BIT_DELAY, &sh->state);
2110         }
2111
2112         if (forwrite) {
2113                 /* check if page is covered */
2114                 sector_t sector = sh->dev[dd_idx].sector;
2115                 for (bi=sh->dev[dd_idx].towrite;
2116                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2117                              bi && bi->bi_sector <= sector;
2118                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2119                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2120                                 sector = bi->bi_sector + (bi->bi_size>>9);
2121                 }
2122                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2123                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2124         }
2125         return 1;
2126
2127  overlap:
2128         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2129         spin_unlock_irq(&conf->device_lock);
2130         spin_unlock(&sh->lock);
2131         return 0;
2132 }
2133
2134 static void end_reshape(raid5_conf_t *conf);
2135
2136 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
2137                             struct stripe_head *sh)
2138 {
2139         int sectors_per_chunk =
2140                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2141         int dd_idx;
2142         int chunk_offset = sector_div(stripe, sectors_per_chunk);
2143         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2144
2145         raid5_compute_sector(conf,
2146                              stripe * (disks - conf->max_degraded)
2147                              *sectors_per_chunk + chunk_offset,
2148                              previous,
2149                              &dd_idx, sh);
2150 }
2151
2152 static void
2153 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2154                                 struct stripe_head_state *s, int disks,
2155                                 struct bio **return_bi)
2156 {
2157         int i;
2158         for (i = disks; i--; ) {
2159                 struct bio *bi;
2160                 int bitmap_end = 0;
2161
2162                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2163                         mdk_rdev_t *rdev;
2164                         rcu_read_lock();
2165                         rdev = rcu_dereference(conf->disks[i].rdev);
2166                         if (rdev && test_bit(In_sync, &rdev->flags))
2167                                 /* multiple read failures in one stripe */
2168                                 md_error(conf->mddev, rdev);
2169                         rcu_read_unlock();
2170                 }
2171                 spin_lock_irq(&conf->device_lock);
2172                 /* fail all writes first */
2173                 bi = sh->dev[i].towrite;
2174                 sh->dev[i].towrite = NULL;
2175                 if (bi) {
2176                         s->to_write--;
2177                         bitmap_end = 1;
2178                 }
2179
2180                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2181                         wake_up(&conf->wait_for_overlap);
2182
2183                 while (bi && bi->bi_sector <
2184                         sh->dev[i].sector + STRIPE_SECTORS) {
2185                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2186                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2187                         if (!raid5_dec_bi_phys_segments(bi)) {
2188                                 md_write_end(conf->mddev);
2189                                 bi->bi_next = *return_bi;
2190                                 *return_bi = bi;
2191                         }
2192                         bi = nextbi;
2193                 }
2194                 /* and fail all 'written' */
2195                 bi = sh->dev[i].written;
2196                 sh->dev[i].written = NULL;
2197                 if (bi) bitmap_end = 1;
2198                 while (bi && bi->bi_sector <
2199                        sh->dev[i].sector + STRIPE_SECTORS) {
2200                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2201                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2202                         if (!raid5_dec_bi_phys_segments(bi)) {
2203                                 md_write_end(conf->mddev);
2204                                 bi->bi_next = *return_bi;
2205                                 *return_bi = bi;
2206                         }
2207                         bi = bi2;
2208                 }
2209
2210                 /* fail any reads if this device is non-operational and
2211                  * the data has not reached the cache yet.
2212                  */
2213                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2214                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2215                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
2216                         bi = sh->dev[i].toread;
2217                         sh->dev[i].toread = NULL;
2218                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2219                                 wake_up(&conf->wait_for_overlap);
2220                         if (bi) s->to_read--;
2221                         while (bi && bi->bi_sector <
2222                                sh->dev[i].sector + STRIPE_SECTORS) {
2223                                 struct bio *nextbi =
2224                                         r5_next_bio(bi, sh->dev[i].sector);
2225                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2226                                 if (!raid5_dec_bi_phys_segments(bi)) {
2227                                         bi->bi_next = *return_bi;
2228                                         *return_bi = bi;
2229                                 }
2230                                 bi = nextbi;
2231                         }
2232                 }
2233                 spin_unlock_irq(&conf->device_lock);
2234                 if (bitmap_end)
2235                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2236                                         STRIPE_SECTORS, 0, 0);
2237         }
2238
2239         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2240                 if (atomic_dec_and_test(&conf->pending_full_writes))
2241                         md_wakeup_thread(conf->mddev->thread);
2242 }
2243
2244 /* fetch_block5 - checks the given member device to see if its data needs
2245  * to be read or computed to satisfy a request.
2246  *
2247  * Returns 1 when no more member devices need to be checked, otherwise returns
2248  * 0 to tell the loop in handle_stripe_fill5 to continue
2249  */
2250 static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s,
2251                         int disk_idx, int disks)
2252 {
2253         struct r5dev *dev = &sh->dev[disk_idx];
2254         struct r5dev *failed_dev = &sh->dev[s->failed_num];
2255
2256         /* is the data in this block needed, and can we get it? */
2257         if (!test_bit(R5_LOCKED, &dev->flags) &&
2258             !test_bit(R5_UPTODATE, &dev->flags) &&
2259             (dev->toread ||
2260              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2261              s->syncing || s->expanding ||
2262              (s->failed &&
2263               (failed_dev->toread ||
2264                (failed_dev->towrite &&
2265                 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) {
2266                 /* We would like to get this block, possibly by computing it,
2267                  * otherwise read it if the backing disk is insync
2268                  */
2269                 if ((s->uptodate == disks - 1) &&
2270                     (s->failed && disk_idx == s->failed_num)) {
2271                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2272                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2273                         set_bit(R5_Wantcompute, &dev->flags);
2274                         sh->ops.target = disk_idx;
2275                         sh->ops.target2 = -1;
2276                         s->req_compute = 1;
2277                         /* Careful: from this point on 'uptodate' is in the eye
2278                          * of raid_run_ops which services 'compute' operations
2279                          * before writes. R5_Wantcompute flags a block that will
2280                          * be R5_UPTODATE by the time it is needed for a
2281                          * subsequent operation.
2282                          */
2283                         s->uptodate++;
2284                         return 1; /* uptodate + compute == disks */
2285                 } else if (test_bit(R5_Insync, &dev->flags)) {
2286                         set_bit(R5_LOCKED, &dev->flags);
2287                         set_bit(R5_Wantread, &dev->flags);
2288                         s->locked++;
2289                         pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2290                                 s->syncing);
2291                 }
2292         }
2293
2294         return 0;
2295 }
2296
2297 /**
2298  * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2299  */
2300 static void handle_stripe_fill5(struct stripe_head *sh,
2301                         struct stripe_head_state *s, int disks)
2302 {
2303         int i;
2304
2305         /* look for blocks to read/compute, skip this if a compute
2306          * is already in flight, or if the stripe contents are in the
2307          * midst of changing due to a write
2308          */
2309         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2310             !sh->reconstruct_state)
2311                 for (i = disks; i--; )
2312                         if (fetch_block5(sh, s, i, disks))
2313                                 break;
2314         set_bit(STRIPE_HANDLE, &sh->state);
2315 }
2316
2317 /* fetch_block6 - checks the given member device to see if its data needs
2318  * to be read or computed to satisfy a request.
2319  *
2320  * Returns 1 when no more member devices need to be checked, otherwise returns
2321  * 0 to tell the loop in handle_stripe_fill6 to continue
2322  */
2323 static int fetch_block6(struct stripe_head *sh, struct stripe_head_state *s,
2324                          struct r6_state *r6s, int disk_idx, int disks)
2325 {
2326         struct r5dev *dev = &sh->dev[disk_idx];
2327         struct r5dev *fdev[2] = { &sh->dev[r6s->failed_num[0]],
2328                                   &sh->dev[r6s->failed_num[1]] };
2329
2330         if (!test_bit(R5_LOCKED, &dev->flags) &&
2331             !test_bit(R5_UPTODATE, &dev->flags) &&
2332             (dev->toread ||
2333              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2334              s->syncing || s->expanding ||
2335              (s->failed >= 1 &&
2336               (fdev[0]->toread || s->to_write)) ||
2337              (s->failed >= 2 &&
2338               (fdev[1]->toread || s->to_write)))) {
2339                 /* we would like to get this block, possibly by computing it,
2340                  * otherwise read it if the backing disk is insync
2341                  */
2342                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2343                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2344                 if ((s->uptodate == disks - 1) &&
2345                     (s->failed && (disk_idx == r6s->failed_num[0] ||
2346                                    disk_idx == r6s->failed_num[1]))) {
2347                         /* have disk failed, and we're requested to fetch it;
2348                          * do compute it
2349                          */
2350                         pr_debug("Computing stripe %llu block %d\n",
2351                                (unsigned long long)sh->sector, disk_idx);
2352                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2353                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2354                         set_bit(R5_Wantcompute, &dev->flags);
2355                         sh->ops.target = disk_idx;
2356                         sh->ops.target2 = -1; /* no 2nd target */
2357                         s->req_compute = 1;
2358                         s->uptodate++;
2359                         return 1;
2360                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2361                         /* Computing 2-failure is *very* expensive; only
2362                          * do it if failed >= 2
2363                          */
2364                         int other;
2365                         for (other = disks; other--; ) {
2366                                 if (other == disk_idx)
2367                                         continue;
2368                                 if (!test_bit(R5_UPTODATE,
2369                                       &sh->dev[other].flags))
2370                                         break;
2371                         }
2372                         BUG_ON(other < 0);
2373                         pr_debug("Computing stripe %llu blocks %d,%d\n",
2374                                (unsigned long long)sh->sector,
2375                                disk_idx, other);
2376                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2377                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2378                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2379                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
2380                         sh->ops.target = disk_idx;
2381                         sh->ops.target2 = other;
2382                         s->uptodate += 2;
2383                         s->req_compute = 1;
2384                         return 1;
2385                 } else if (test_bit(R5_Insync, &dev->flags)) {
2386                         set_bit(R5_LOCKED, &dev->flags);
2387                         set_bit(R5_Wantread, &dev->flags);
2388                         s->locked++;
2389                         pr_debug("Reading block %d (sync=%d)\n",
2390                                 disk_idx, s->syncing);
2391                 }
2392         }
2393
2394         return 0;
2395 }
2396
2397 /**
2398  * handle_stripe_fill6 - read or compute data to satisfy pending requests.
2399  */
2400 static void handle_stripe_fill6(struct stripe_head *sh,
2401                         struct stripe_head_state *s, struct r6_state *r6s,
2402                         int disks)
2403 {
2404         int i;
2405
2406         /* look for blocks to read/compute, skip this if a compute
2407          * is already in flight, or if the stripe contents are in the
2408          * midst of changing due to a write
2409          */
2410         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2411             !sh->reconstruct_state)
2412                 for (i = disks; i--; )
2413                         if (fetch_block6(sh, s, r6s, i, disks))
2414                                 break;
2415         set_bit(STRIPE_HANDLE, &sh->state);
2416 }
2417
2418
2419 /* handle_stripe_clean_event
2420  * any written block on an uptodate or failed drive can be returned.
2421  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2422  * never LOCKED, so we don't need to test 'failed' directly.
2423  */
2424 static void handle_stripe_clean_event(raid5_conf_t *conf,
2425         struct stripe_head *sh, int disks, struct bio **return_bi)
2426 {
2427         int i;
2428         struct r5dev *dev;
2429
2430         for (i = disks; i--; )
2431                 if (sh->dev[i].written) {
2432                         dev = &sh->dev[i];
2433                         if (!test_bit(R5_LOCKED, &dev->flags) &&
2434                                 test_bit(R5_UPTODATE, &dev->flags)) {
2435                                 /* We can return any write requests */
2436                                 struct bio *wbi, *wbi2;
2437                                 int bitmap_end = 0;
2438                                 pr_debug("Return write for disc %d\n", i);
2439                                 spin_lock_irq(&conf->device_lock);
2440                                 wbi = dev->written;
2441                                 dev->written = NULL;
2442                                 while (wbi && wbi->bi_sector <
2443                                         dev->sector + STRIPE_SECTORS) {
2444                                         wbi2 = r5_next_bio(wbi, dev->sector);
2445                                         if (!raid5_dec_bi_phys_segments(wbi)) {
2446                                                 md_write_end(conf->mddev);
2447                                                 wbi->bi_next = *return_bi;
2448                                                 *return_bi = wbi;
2449                                         }
2450                                         wbi = wbi2;
2451                                 }
2452                                 if (dev->towrite == NULL)
2453                                         bitmap_end = 1;
2454                                 spin_unlock_irq(&conf->device_lock);
2455                                 if (bitmap_end)
2456                                         bitmap_endwrite(conf->mddev->bitmap,
2457                                                         sh->sector,
2458                                                         STRIPE_SECTORS,
2459                                          !test_bit(STRIPE_DEGRADED, &sh->state),
2460                                                         0);
2461                         }
2462                 }
2463
2464         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2465                 if (atomic_dec_and_test(&conf->pending_full_writes))
2466                         md_wakeup_thread(conf->mddev->thread);
2467 }
2468
2469 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2470                 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2471 {
2472         int rmw = 0, rcw = 0, i;
2473         for (i = disks; i--; ) {
2474                 /* would I have to read this buffer for read_modify_write */
2475                 struct r5dev *dev = &sh->dev[i];
2476                 if ((dev->towrite || i == sh->pd_idx) &&
2477                     !test_bit(R5_LOCKED, &dev->flags) &&
2478                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2479                       test_bit(R5_Wantcompute, &dev->flags))) {
2480                         if (test_bit(R5_Insync, &dev->flags))
2481                                 rmw++;
2482                         else
2483                                 rmw += 2*disks;  /* cannot read it */
2484                 }
2485                 /* Would I have to read this buffer for reconstruct_write */
2486                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2487                     !test_bit(R5_LOCKED, &dev->flags) &&
2488                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2489                     test_bit(R5_Wantcompute, &dev->flags))) {
2490                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2491                         else
2492                                 rcw += 2*disks;
2493                 }
2494         }
2495         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2496                 (unsigned long long)sh->sector, rmw, rcw);
2497         set_bit(STRIPE_HANDLE, &sh->state);
2498         if (rmw < rcw && rmw > 0)
2499                 /* prefer read-modify-write, but need to get some data */
2500                 for (i = disks; i--; ) {
2501                         struct r5dev *dev = &sh->dev[i];
2502                         if ((dev->towrite || i == sh->pd_idx) &&
2503                             !test_bit(R5_LOCKED, &dev->flags) &&
2504                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2505                             test_bit(R5_Wantcompute, &dev->flags)) &&
2506                             test_bit(R5_Insync, &dev->flags)) {
2507                                 if (
2508                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2509                                         pr_debug("Read_old block "
2510                                                 "%d for r-m-w\n", i);
2511                                         set_bit(R5_LOCKED, &dev->flags);
2512                                         set_bit(R5_Wantread, &dev->flags);
2513                                         s->locked++;
2514                                 } else {
2515                                         set_bit(STRIPE_DELAYED, &sh->state);
2516                                         set_bit(STRIPE_HANDLE, &sh->state);
2517                                 }
2518                         }
2519                 }
2520         if (rcw <= rmw && rcw > 0)
2521                 /* want reconstruct write, but need to get some data */
2522                 for (i = disks; i--; ) {
2523                         struct r5dev *dev = &sh->dev[i];
2524                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2525                             i != sh->pd_idx &&
2526                             !test_bit(R5_LOCKED, &dev->flags) &&
2527                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2528                             test_bit(R5_Wantcompute, &dev->flags)) &&
2529                             test_bit(R5_Insync, &dev->flags)) {
2530                                 if (
2531                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2532                                         pr_debug("Read_old block "
2533                                                 "%d for Reconstruct\n", i);
2534                                         set_bit(R5_LOCKED, &dev->flags);
2535                                         set_bit(R5_Wantread, &dev->flags);
2536                                         s->locked++;
2537                                 } else {
2538                                         set_bit(STRIPE_DELAYED, &sh->state);
2539                                         set_bit(STRIPE_HANDLE, &sh->state);
2540                                 }
2541                         }
2542                 }
2543         /* now if nothing is locked, and if we have enough data,
2544          * we can start a write request
2545          */
2546         /* since handle_stripe can be called at any time we need to handle the
2547          * case where a compute block operation has been submitted and then a
2548          * subsequent call wants to start a write request.  raid_run_ops only
2549          * handles the case where compute block and reconstruct are requested
2550          * simultaneously.  If this is not the case then new writes need to be
2551          * held off until the compute completes.
2552          */
2553         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2554             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2555             !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2556                 schedule_reconstruction(sh, s, rcw == 0, 0);
2557 }
2558
2559 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2560                 struct stripe_head *sh, struct stripe_head_state *s,
2561                 struct r6_state *r6s, int disks)
2562 {
2563         int rcw = 0, pd_idx = sh->pd_idx, i;
2564         int qd_idx = sh->qd_idx;
2565
2566         set_bit(STRIPE_HANDLE, &sh->state);
2567         for (i = disks; i--; ) {
2568                 struct r5dev *dev = &sh->dev[i];
2569                 /* check if we haven't enough data */
2570                 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2571                     i != pd_idx && i != qd_idx &&
2572                     !test_bit(R5_LOCKED, &dev->flags) &&
2573                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2574                       test_bit(R5_Wantcompute, &dev->flags))) {
2575                         rcw++;
2576                         if (!test_bit(R5_Insync, &dev->flags))
2577                                 continue; /* it's a failed drive */
2578
2579                         if (
2580                           test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2581                                 pr_debug("Read_old stripe %llu "
2582                                         "block %d for Reconstruct\n",
2583                                      (unsigned long long)sh->sector, i);
2584                                 set_bit(R5_LOCKED, &dev->flags);
2585                                 set_bit(R5_Wantread, &dev->flags);
2586                                 s->locked++;
2587                         } else {
2588                                 pr_debug("Request delayed stripe %llu "
2589                                         "block %d for Reconstruct\n",
2590                                      (unsigned long long)sh->sector, i);
2591                                 set_bit(STRIPE_DELAYED, &sh->state);
2592                                 set_bit(STRIPE_HANDLE, &sh->state);
2593                         }
2594                 }
2595         }
2596         /* now if nothing is locked, and if we have enough data, we can start a
2597          * write request
2598          */
2599         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2600             s->locked == 0 && rcw == 0 &&
2601             !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2602                 schedule_reconstruction(sh, s, 1, 0);
2603         }
2604 }
2605
2606 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2607                                 struct stripe_head_state *s, int disks)
2608 {
2609         struct r5dev *dev = NULL;
2610
2611         set_bit(STRIPE_HANDLE, &sh->state);
2612
2613         switch (sh->check_state) {
2614         case check_state_idle:
2615                 /* start a new check operation if there are no failures */
2616                 if (s->failed == 0) {
2617                         BUG_ON(s->uptodate != disks);
2618                         sh->check_state = check_state_run;
2619                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
2620                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2621                         s->uptodate--;
2622                         break;
2623                 }
2624                 dev = &sh->dev[s->failed_num];
2625                 /* fall through */
2626         case check_state_compute_result:
2627                 sh->check_state = check_state_idle;
2628                 if (!dev)
2629                         dev = &sh->dev[sh->pd_idx];
2630
2631                 /* check that a write has not made the stripe insync */
2632                 if (test_bit(STRIPE_INSYNC, &sh->state))
2633                         break;
2634
2635                 /* either failed parity check, or recovery is happening */
2636                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2637                 BUG_ON(s->uptodate != disks);
2638
2639                 set_bit(R5_LOCKED, &dev->flags);
2640                 s->locked++;
2641                 set_bit(R5_Wantwrite, &dev->flags);
2642
2643                 clear_bit(STRIPE_DEGRADED, &sh->state);
2644                 set_bit(STRIPE_INSYNC, &sh->state);
2645                 break;
2646         case check_state_run:
2647                 break; /* we will be called again upon completion */
2648         case check_state_check_result:
2649                 sh->check_state = check_state_idle;
2650
2651                 /* if a failure occurred during the check operation, leave
2652                  * STRIPE_INSYNC not set and let the stripe be handled again
2653                  */
2654                 if (s->failed)
2655                         break;
2656
2657                 /* handle a successful check operation, if parity is correct
2658                  * we are done.  Otherwise update the mismatch count and repair
2659                  * parity if !MD_RECOVERY_CHECK
2660                  */
2661                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2662                         /* parity is correct (on disc,
2663                          * not in buffer any more)
2664                          */
2665                         set_bit(STRIPE_INSYNC, &sh->state);
2666                 else {
2667                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2668                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2669                                 /* don't try to repair!! */
2670                                 set_bit(STRIPE_INSYNC, &sh->state);
2671                         else {
2672                                 sh->check_state = check_state_compute_run;
2673                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2674                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2675                                 set_bit(R5_Wantcompute,
2676                                         &sh->dev[sh->pd_idx].flags);
2677                                 sh->ops.target = sh->pd_idx;
2678                                 sh->ops.target2 = -1;
2679                                 s->uptodate++;
2680                         }
2681                 }
2682                 break;
2683         case check_state_compute_run:
2684                 break;
2685         default:
2686                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2687                        __func__, sh->check_state,
2688                        (unsigned long long) sh->sector);
2689                 BUG();
2690         }
2691 }
2692
2693
2694 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2695                                   struct stripe_head_state *s,
2696                                   struct r6_state *r6s, int disks)
2697 {
2698         int pd_idx = sh->pd_idx;
2699         int qd_idx = sh->qd_idx;
2700         struct r5dev *dev;
2701
2702         set_bit(STRIPE_HANDLE, &sh->state);
2703
2704         BUG_ON(s->failed > 2);
2705
2706         /* Want to check and possibly repair P and Q.
2707          * However there could be one 'failed' device, in which
2708          * case we can only check one of them, possibly using the
2709          * other to generate missing data
2710          */
2711
2712         switch (sh->check_state) {
2713         case check_state_idle:
2714                 /* start a new check operation if there are < 2 failures */
2715                 if (s->failed == r6s->q_failed) {
2716                         /* The only possible failed device holds Q, so it
2717                          * makes sense to check P (If anything else were failed,
2718                          * we would have used P to recreate it).
2719                          */
2720                         sh->check_state = check_state_run;
2721                 }
2722                 if (!r6s->q_failed && s->failed < 2) {
2723                         /* Q is not failed, and we didn't use it to generate
2724                          * anything, so it makes sense to check it
2725                          */
2726                         if (sh->check_state == check_state_run)
2727                                 sh->check_state = check_state_run_pq;
2728                         else
2729                                 sh->check_state = check_state_run_q;
2730                 }
2731
2732                 /* discard potentially stale zero_sum_result */
2733                 sh->ops.zero_sum_result = 0;
2734
2735                 if (sh->check_state == check_state_run) {
2736                         /* async_xor_zero_sum destroys the contents of P */
2737                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2738                         s->uptodate--;
2739                 }
2740                 if (sh->check_state >= check_state_run &&
2741                     sh->check_state <= check_state_run_pq) {
2742                         /* async_syndrome_zero_sum preserves P and Q, so
2743                          * no need to mark them !uptodate here
2744                          */
2745                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
2746                         break;
2747                 }
2748
2749                 /* we have 2-disk failure */
2750                 BUG_ON(s->failed != 2);
2751                 /* fall through */
2752         case check_state_compute_result:
2753                 sh->check_state = check_state_idle;
2754
2755                 /* check that a write has not made the stripe insync */
2756                 if (test_bit(STRIPE_INSYNC, &sh->state))
2757                         break;
2758
2759                 /* now write out any block on a failed drive,
2760                  * or P or Q if they were recomputed
2761                  */
2762                 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2763                 if (s->failed == 2) {
2764                         dev = &sh->dev[r6s->failed_num[1]];
2765                         s->locked++;
2766                         set_bit(R5_LOCKED, &dev->flags);
2767                         set_bit(R5_Wantwrite, &dev->flags);
2768                 }
2769                 if (s->failed >= 1) {
2770                         dev = &sh->dev[r6s->failed_num[0]];
2771                         s->locked++;
2772                         set_bit(R5_LOCKED, &dev->flags);
2773                         set_bit(R5_Wantwrite, &dev->flags);
2774                 }
2775                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2776                         dev = &sh->dev[pd_idx];
2777                         s->locked++;
2778                         set_bit(R5_LOCKED, &dev->flags);
2779                         set_bit(R5_Wantwrite, &dev->flags);
2780                 }
2781                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2782                         dev = &sh->dev[qd_idx];
2783                         s->locked++;
2784                         set_bit(R5_LOCKED, &dev->flags);
2785                         set_bit(R5_Wantwrite, &dev->flags);
2786                 }
2787                 clear_bit(STRIPE_DEGRADED, &sh->state);
2788
2789                 set_bit(STRIPE_INSYNC, &sh->state);
2790                 break;
2791         case check_state_run:
2792         case check_state_run_q:
2793         case check_state_run_pq:
2794                 break; /* we will be called again upon completion */
2795         case check_state_check_result:
2796                 sh->check_state = check_state_idle;
2797
2798                 /* handle a successful check operation, if parity is correct
2799                  * we are done.  Otherwise update the mismatch count and repair
2800                  * parity if !MD_RECOVERY_CHECK
2801                  */
2802                 if (sh->ops.zero_sum_result == 0) {
2803                         /* both parities are correct */
2804                         if (!s->failed)
2805                                 set_bit(STRIPE_INSYNC, &sh->state);
2806                         else {
2807                                 /* in contrast to the raid5 case we can validate
2808                                  * parity, but still have a failure to write
2809                                  * back
2810                                  */
2811                                 sh->check_state = check_state_compute_result;
2812                                 /* Returning at this point means that we may go
2813                                  * off and bring p and/or q uptodate again so
2814                                  * we make sure to check zero_sum_result again
2815                                  * to verify if p or q need writeback
2816                                  */
2817                         }
2818                 } else {
2819                         conf->mddev->resync_mismatches += STRIPE_SECTORS;
2820                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2821                                 /* don't try to repair!! */
2822                                 set_bit(STRIPE_INSYNC, &sh->state);
2823                         else {
2824                                 int *target = &sh->ops.target;
2825
2826                                 sh->ops.target = -1;
2827                                 sh->ops.target2 = -1;
2828                                 sh->check_state = check_state_compute_run;
2829                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2830                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2831                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2832                                         set_bit(R5_Wantcompute,
2833                                                 &sh->dev[pd_idx].flags);
2834                                         *target = pd_idx;
2835                                         target = &sh->ops.target2;
2836                                         s->uptodate++;
2837                                 }
2838                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2839                                         set_bit(R5_Wantcompute,
2840                                                 &sh->dev[qd_idx].flags);
2841                                         *target = qd_idx;
2842                                         s->uptodate++;
2843                                 }
2844                         }
2845                 }
2846                 break;
2847         case check_state_compute_run:
2848                 break;
2849         default:
2850                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2851                        __func__, sh->check_state,
2852                        (unsigned long long) sh->sector);
2853                 BUG();
2854         }
2855 }
2856
2857 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2858                                 struct r6_state *r6s)
2859 {
2860         int i;
2861
2862         /* We have read all the blocks in this stripe and now we need to
2863          * copy some of them into a target stripe for expand.
2864          */
2865         struct dma_async_tx_descriptor *tx = NULL;
2866         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2867         for (i = 0; i < sh->disks; i++)
2868                 if (i != sh->pd_idx && i != sh->qd_idx) {
2869                         int dd_idx, j;
2870                         struct stripe_head *sh2;
2871                         struct async_submit_ctl submit;
2872
2873                         sector_t bn = compute_blocknr(sh, i, 1);
2874                         sector_t s = raid5_compute_sector(conf, bn, 0,
2875                                                           &dd_idx, NULL);
2876                         sh2 = get_active_stripe(conf, s, 0, 1, 1);
2877                         if (sh2 == NULL)
2878                                 /* so far only the early blocks of this stripe
2879                                  * have been requested.  When later blocks
2880                                  * get requested, we will try again
2881                                  */
2882                                 continue;
2883                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2884                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2885                                 /* must have already done this block */
2886                                 release_stripe(sh2);
2887                                 continue;
2888                         }
2889
2890                         /* place all the copies on one channel */
2891                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
2892                         tx = async_memcpy(sh2->dev[dd_idx].page,
2893                                           sh->dev[i].page, 0, 0, STRIPE_SIZE,
2894                                           &submit);
2895
2896                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2897                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2898                         for (j = 0; j < conf->raid_disks; j++)
2899                                 if (j != sh2->pd_idx &&
2900                                     (!r6s || j != sh2->qd_idx) &&
2901                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
2902                                         break;
2903                         if (j == conf->raid_disks) {
2904                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2905                                 set_bit(STRIPE_HANDLE, &sh2->state);
2906                         }
2907                         release_stripe(sh2);
2908
2909                 }
2910         /* done submitting copies, wait for them to complete */
2911         if (tx) {
2912                 async_tx_ack(tx);
2913                 dma_wait_for_async_tx(tx);
2914         }
2915 }
2916
2917
2918 /*
2919  * handle_stripe - do things to a stripe.
2920  *
2921  * We lock the stripe and then examine the state of various bits
2922  * to see what needs to be done.
2923  * Possible results:
2924  *    return some read request which now have data
2925  *    return some write requests which are safely on disc
2926  *    schedule a read on some buffers
2927  *    schedule a write of some buffers
2928  *    return confirmation of parity correctness
2929  *
2930  * buffers are taken off read_list or write_list, and bh_cache buffers
2931  * get BH_Lock set before the stripe lock is released.
2932  *
2933  */
2934
2935 static void handle_stripe5(struct stripe_head *sh)
2936 {
2937         raid5_conf_t *conf = sh->raid_conf;
2938         int disks = sh->disks, i;
2939         struct bio *return_bi = NULL;
2940         struct stripe_head_state s;
2941         struct r5dev *dev;
2942         mdk_rdev_t *blocked_rdev = NULL;
2943         int prexor;
2944         int dec_preread_active = 0;
2945
2946         memset(&s, 0, sizeof(s));
2947         pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
2948                  "reconstruct:%d\n", (unsigned long long)sh->sector, sh->state,
2949                  atomic_read(&sh->count), sh->pd_idx, sh->check_state,
2950                  sh->reconstruct_state);
2951
2952         spin_lock(&sh->lock);
2953         clear_bit(STRIPE_HANDLE, &sh->state);
2954         clear_bit(STRIPE_DELAYED, &sh->state);
2955
2956         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2957         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2958         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2959
2960         /* Now to look around and see what can be done */
2961         rcu_read_lock();
2962         for (i=disks; i--; ) {
2963                 mdk_rdev_t *rdev;
2964
2965                 dev = &sh->dev[i];
2966                 clear_bit(R5_Insync, &dev->flags);
2967
2968                 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2969                         "written %p\n", i, dev->flags, dev->toread, dev->read,
2970                         dev->towrite, dev->written);
2971
2972                 /* maybe we can request a biofill operation
2973                  *
2974                  * new wantfill requests are only permitted while
2975                  * ops_complete_biofill is guaranteed to be inactive
2976                  */
2977                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2978                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
2979                         set_bit(R5_Wantfill, &dev->flags);
2980
2981                 /* now count some things */
2982                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2983                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2984                 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2985
2986                 if (test_bit(R5_Wantfill, &dev->flags))
2987                         s.to_fill++;
2988                 else if (dev->toread)
2989                         s.to_read++;
2990                 if (dev->towrite) {
2991                         s.to_write++;
2992                         if (!test_bit(R5_OVERWRITE, &dev->flags))
2993                                 s.non_overwrite++;
2994                 }
2995                 if (dev->written)
2996                         s.written++;
2997                 rdev = rcu_dereference(conf->disks[i].rdev);
2998                 if (blocked_rdev == NULL &&
2999                     rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3000                         blocked_rdev = rdev;
3001                         atomic_inc(&rdev->nr_pending);
3002                 }
3003                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3004                         /* The ReadError flag will just be confusing now */
3005                         clear_bit(R5_ReadError, &dev->flags);
3006                         clear_bit(R5_ReWrite, &dev->flags);
3007                 }
3008                 if (!rdev || !test_bit(In_sync, &rdev->flags)
3009                     || test_bit(R5_ReadError, &dev->flags)) {
3010                         s.failed++;
3011                         s.failed_num = i;
3012                 } else
3013                         set_bit(R5_Insync, &dev->flags);
3014         }
3015         rcu_read_unlock();
3016
3017         if (unlikely(blocked_rdev)) {
3018                 if (s.syncing || s.expanding || s.expanded ||
3019                     s.to_write || s.written) {
3020                         set_bit(STRIPE_HANDLE, &sh->state);
3021                         goto unlock;
3022                 }
3023                 /* There is nothing for the blocked_rdev to block */
3024                 rdev_dec_pending(blocked_rdev, conf->mddev);
3025                 blocked_rdev = NULL;
3026         }
3027
3028         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3029                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3030                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3031         }
3032
3033         pr_debug("locked=%d uptodate=%d to_read=%d"
3034                 " to_write=%d failed=%d failed_num=%d\n",
3035                 s.locked, s.uptodate, s.to_read, s.to_write,
3036                 s.failed, s.failed_num);
3037         /* check if the array has lost two devices and, if so, some requests might
3038          * need to be failed
3039          */
3040         if (s.failed > 1 && s.to_read+s.to_write+s.written)
3041                 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3042         if (s.failed > 1 && s.syncing) {
3043                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3044                 clear_bit(STRIPE_SYNCING, &sh->state);
3045                 s.syncing = 0;
3046         }
3047
3048         /* might be able to return some write requests if the parity block
3049          * is safe, or on a failed drive
3050          */
3051         dev = &sh->dev[sh->pd_idx];
3052         if ( s.written &&
3053              ((test_bit(R5_Insync, &dev->flags) &&
3054                !test_bit(R5_LOCKED, &dev->flags) &&
3055                test_bit(R5_UPTODATE, &dev->flags)) ||
3056                (s.failed == 1 && s.failed_num == sh->pd_idx)))
3057                 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3058
3059         /* Now we might consider reading some blocks, either to check/generate
3060          * parity, or to satisfy requests
3061          * or to load a block that is being partially written.
3062          */
3063         if (s.to_read || s.non_overwrite ||
3064             (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3065                 handle_stripe_fill5(sh, &s, disks);
3066
3067         /* Now we check to see if any write operations have recently
3068          * completed
3069          */
3070         prexor = 0;
3071         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3072                 prexor = 1;
3073         if (sh->reconstruct_state == reconstruct_state_drain_result ||
3074             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3075                 sh->reconstruct_state = reconstruct_state_idle;
3076
3077                 /* All the 'written' buffers and the parity block are ready to
3078                  * be written back to disk
3079                  */
3080                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3081                 for (i = disks; i--; ) {
3082                         dev = &sh->dev[i];
3083                         if (test_bit(R5_LOCKED, &dev->flags) &&
3084                                 (i == sh->pd_idx || dev->written)) {
3085                                 pr_debug("Writing block %d\n", i);
3086                                 set_bit(R5_Wantwrite, &dev->flags);
3087                                 if (prexor)
3088                                         continue;
3089                                 if (!test_bit(R5_Insync, &dev->flags) ||
3090                                     (i == sh->pd_idx && s.failed == 0))
3091                                         set_bit(STRIPE_INSYNC, &sh->state);
3092                         }
3093                 }
3094                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3095                         dec_preread_active = 1;
3096         }
3097
3098         /* Now to consider new write requests and what else, if anything
3099          * should be read.  We do not handle new writes when:
3100          * 1/ A 'write' operation (copy+xor) is already in flight.
3101          * 2/ A 'check' operation is in flight, as it may clobber the parity
3102          *    block.
3103          */
3104         if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3105                 handle_stripe_dirtying5(conf, sh, &s, disks);
3106
3107         /* maybe we need to check and possibly fix the parity for this stripe
3108          * Any reads will already have been scheduled, so we just see if enough
3109          * data is available.  The parity check is held off while parity
3110          * dependent operations are in flight.
3111          */
3112         if (sh->check_state ||
3113             (s.syncing && s.locked == 0 &&
3114              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3115              !test_bit(STRIPE_INSYNC, &sh->state)))
3116                 handle_parity_checks5(conf, sh, &s, disks);
3117
3118         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3119                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3120                 clear_bit(STRIPE_SYNCING, &sh->state);
3121         }
3122
3123         /* If the failed drive is just a ReadError, then we might need to progress
3124          * the repair/check process
3125          */
3126         if (s.failed == 1 && !conf->mddev->ro &&
3127             test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
3128             && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
3129             && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
3130                 ) {
3131                 dev = &sh->dev[s.failed_num];
3132                 if (!test_bit(R5_ReWrite, &dev->flags)) {
3133                         set_bit(R5_Wantwrite, &dev->flags);
3134                         set_bit(R5_ReWrite, &dev->flags);
3135                         set_bit(R5_LOCKED, &dev->flags);
3136                         s.locked++;
3137                 } else {
3138                         /* let's read it back */
3139                         set_bit(R5_Wantread, &dev->flags);
3140                         set_bit(R5_LOCKED, &dev->flags);
3141                         s.locked++;
3142                 }
3143         }
3144
3145         /* Finish reconstruct operations initiated by the expansion process */
3146         if (sh->reconstruct_state == reconstruct_state_result) {
3147                 struct stripe_head *sh2
3148                         = get_active_stripe(conf, sh->sector, 1, 1, 1);
3149                 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3150                         /* sh cannot be written until sh2 has been read.
3151                          * so arrange for sh to be delayed a little
3152                          */
3153                         set_bit(STRIPE_DELAYED, &sh->state);
3154                         set_bit(STRIPE_HANDLE, &sh->state);
3155                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3156                                               &sh2->state))
3157                                 atomic_inc(&conf->preread_active_stripes);
3158                         release_stripe(sh2);
3159                         goto unlock;
3160                 }
3161                 if (sh2)
3162                         release_stripe(sh2);
3163
3164                 sh->reconstruct_state = reconstruct_state_idle;
3165                 clear_bit(STRIPE_EXPANDING, &sh->state);
3166                 for (i = conf->raid_disks; i--; ) {
3167                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
3168                         set_bit(R5_LOCKED, &sh->dev[i].flags);
3169                         s.locked++;
3170                 }
3171         }
3172
3173         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3174             !sh->reconstruct_state) {
3175                 /* Need to write out all blocks after computing parity */
3176                 sh->disks = conf->raid_disks;
3177                 stripe_set_idx(sh->sector, conf, 0, sh);
3178                 schedule_reconstruction(sh, &s, 1, 1);
3179         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3180                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3181                 atomic_dec(&conf->reshape_stripes);
3182                 wake_up(&conf->wait_for_overlap);
3183                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3184         }
3185
3186         if (s.expanding && s.locked == 0 &&
3187             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3188                 handle_stripe_expansion(conf, sh, NULL);
3189
3190  unlock:
3191         spin_unlock(&sh->lock);
3192
3193         /* wait for this device to become unblocked */
3194         if (unlikely(blocked_rdev))
3195                 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3196
3197         if (s.ops_request)
3198                 raid_run_ops(sh, s.ops_request);
3199
3200         ops_run_io(sh, &s);
3201
3202         if (dec_preread_active) {
3203                 /* We delay this until after ops_run_io so that if make_request
3204                  * is waiting on a barrier, it won't continue until the writes
3205                  * have actually been submitted.
3206                  */
3207                 atomic_dec(&conf->preread_active_stripes);
3208                 if (atomic_read(&conf->preread_active_stripes) <
3209                     IO_THRESHOLD)
3210                         md_wakeup_thread(conf->mddev->thread);
3211         }
3212         return_io(return_bi);
3213 }
3214
3215 static void handle_stripe6(struct stripe_head *sh)
3216 {
3217         raid5_conf_t *conf = sh->raid_conf;
3218         int disks = sh->disks;
3219         struct bio *return_bi = NULL;
3220         int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx;
3221         struct stripe_head_state s;
3222         struct r6_state r6s;
3223         struct r5dev *dev, *pdev, *qdev;
3224         mdk_rdev_t *blocked_rdev = NULL;
3225         int dec_preread_active = 0;
3226
3227         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3228                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3229                (unsigned long long)sh->sector, sh->state,
3230                atomic_read(&sh->count), pd_idx, qd_idx,
3231                sh->check_state, sh->reconstruct_state);
3232         memset(&s, 0, sizeof(s));
3233
3234         spin_lock(&sh->lock);
3235         clear_bit(STRIPE_HANDLE, &sh->state);
3236         clear_bit(STRIPE_DELAYED, &sh->state);
3237
3238         s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3239         s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3240         s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3241         /* Now to look around and see what can be done */
3242
3243         rcu_read_lock();
3244         for (i=disks; i--; ) {
3245                 mdk_rdev_t *rdev;
3246                 dev = &sh->dev[i];
3247                 clear_bit(R5_Insync, &dev->flags);
3248
3249                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3250                         i, dev->flags, dev->toread, dev->towrite, dev->written);
3251                 /* maybe we can reply to a read
3252                  *
3253                  * new wantfill requests are only permitted while
3254                  * ops_complete_biofill is guaranteed to be inactive
3255                  */
3256                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3257                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3258                         set_bit(R5_Wantfill, &dev->flags);
3259
3260                 /* now count some things */
3261                 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3262                 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3263                 if (test_bit(R5_Wantcompute, &dev->flags)) {
3264                         s.compute++;
3265                         BUG_ON(s.compute > 2);
3266                 }
3267
3268                 if (test_bit(R5_Wantfill, &dev->flags)) {
3269                         s.to_fill++;
3270                 } else if (dev->toread)
3271                         s.to_read++;
3272                 if (dev->towrite) {
3273                         s.to_write++;
3274                         if (!test_bit(R5_OVERWRITE, &dev->flags))
3275                                 s.non_overwrite++;
3276                 }
3277                 if (dev->written)
3278                         s.written++;
3279                 rdev = rcu_dereference(conf->disks[i].rdev);
3280                 if (blocked_rdev == NULL &&
3281                     rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3282                         blocked_rdev = rdev;
3283                         atomic_inc(&rdev->nr_pending);
3284                 }
3285                 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3286                         /* The ReadError flag will just be confusing now */
3287                         clear_bit(R5_ReadError, &dev->flags);
3288                         clear_bit(R5_ReWrite, &dev->flags);
3289                 }
3290                 if (!rdev || !test_bit(In_sync, &rdev->flags)
3291                     || test_bit(R5_ReadError, &dev->flags)) {
3292                         if (s.failed < 2)
3293                                 r6s.failed_num[s.failed] = i;
3294                         s.failed++;
3295                 } else
3296                         set_bit(R5_Insync, &dev->flags);
3297         }
3298         rcu_read_unlock();
3299
3300         if (unlikely(blocked_rdev)) {
3301                 if (s.syncing || s.expanding || s.expanded ||
3302                     s.to_write || s.written) {
3303                         set_bit(STRIPE_HANDLE, &sh->state);
3304                         goto unlock;
3305                 }
3306                 /* There is nothing for the blocked_rdev to block */
3307                 rdev_dec_pending(blocked_rdev, conf->mddev);
3308                 blocked_rdev = NULL;
3309         }
3310
3311         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3312                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3313                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3314         }
3315
3316         pr_debug("locked=%d uptodate=%d to_read=%d"
3317                " to_write=%d failed=%d failed_num=%d,%d\n",
3318                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3319                r6s.failed_num[0], r6s.failed_num[1]);
3320         /* check if the array has lost >2 devices and, if so, some requests
3321          * might need to be failed
3322          */
3323         if (s.failed > 2 && s.to_read+s.to_write+s.written)
3324                 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3325         if (s.failed > 2 && s.syncing) {
3326                 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3327                 clear_bit(STRIPE_SYNCING, &sh->state);
3328                 s.syncing = 0;
3329         }
3330
3331         /*
3332          * might be able to return some write requests if the parity blocks
3333          * are safe, or on a failed drive
3334          */
3335         pdev = &sh->dev[pd_idx];
3336         r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3337                 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3338         qdev = &sh->dev[qd_idx];
3339         r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == qd_idx)
3340                 || (s.failed >= 2 && r6s.failed_num[1] == qd_idx);
3341
3342         if ( s.written &&
3343              ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3344                              && !test_bit(R5_LOCKED, &pdev->flags)
3345                              && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3346              ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3347                              && !test_bit(R5_LOCKED, &qdev->flags)
3348                              && test_bit(R5_UPTODATE, &qdev->flags)))))
3349                 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3350
3351         /* Now we might consider reading some blocks, either to check/generate
3352          * parity, or to satisfy requests
3353          * or to load a block that is being partially written.
3354          */
3355         if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3356             (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3357                 handle_stripe_fill6(sh, &s, &r6s, disks);
3358
3359         /* Now we check to see if any write operations have recently
3360          * completed
3361          */
3362         if (sh->reconstruct_state == reconstruct_state_drain_result) {
3363
3364                 sh->reconstruct_state = reconstruct_state_idle;
3365                 /* All the 'written' buffers and the parity blocks are ready to
3366                  * be written back to disk
3367                  */
3368                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3369                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags));
3370                 for (i = disks; i--; ) {
3371                         dev = &sh->dev[i];
3372                         if (test_bit(R5_LOCKED, &dev->flags) &&
3373                             (i == sh->pd_idx || i == qd_idx ||
3374                              dev->written)) {
3375                                 pr_debug("Writing block %d\n", i);
3376                                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3377                                 set_bit(R5_Wantwrite, &dev->flags);
3378                                 if (!test_bit(R5_Insync, &dev->flags) ||
3379                                     ((i == sh->pd_idx || i == qd_idx) &&
3380                                       s.failed == 0))
3381                                         set_bit(STRIPE_INSYNC, &sh->state);
3382                         }
3383                 }
3384                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3385                         dec_preread_active = 1;
3386         }
3387
3388         /* Now to consider new write requests and what else, if anything
3389          * should be read.  We do not handle new writes when:
3390          * 1/ A 'write' operation (copy+gen_syndrome) is already in flight.
3391          * 2/ A 'check' operation is in flight, as it may clobber the parity
3392          *    block.
3393          */
3394         if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3395                 handle_stripe_dirtying6(conf, sh, &s, &r6s, disks);
3396
3397         /* maybe we need to check and possibly fix the parity for this stripe
3398          * Any reads will already have been scheduled, so we just see if enough
3399          * data is available.  The parity check is held off while parity
3400          * dependent operations are in flight.
3401          */
3402         if (sh->check_state ||
3403             (s.syncing && s.locked == 0 &&
3404              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3405              !test_bit(STRIPE_INSYNC, &sh->state)))
3406                 handle_parity_checks6(conf, sh, &s, &r6s, disks);
3407
3408         if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3409                 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3410                 clear_bit(STRIPE_SYNCING, &sh->state);
3411         }
3412
3413         /* If the failed drives are just a ReadError, then we might need
3414          * to progress the repair/check process
3415          */
3416         if (s.failed <= 2 && !conf->mddev->ro)
3417                 for (i = 0; i < s.failed; i++) {
3418                         dev = &sh->dev[r6s.failed_num[i]];
3419                         if (test_bit(R5_ReadError, &dev->flags)
3420                             && !test_bit(R5_LOCKED, &dev->flags)
3421                             && test_bit(R5_UPTODATE, &dev->flags)
3422                                 ) {
3423                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
3424                                         set_bit(R5_Wantwrite, &dev->flags);
3425                                         set_bit(R5_ReWrite, &dev->flags);
3426                                         set_bit(R5_LOCKED, &dev->flags);
3427                                         s.locked++;
3428                                 } else {
3429                                         /* let's read it back */
3430                                         set_bit(R5_Wantread, &dev->flags);
3431                                         set_bit(R5_LOCKED, &dev->flags);
3432                                         s.locked++;
3433                                 }
3434                         }
3435                 }
3436
3437         /* Finish reconstruct operations initiated by the expansion process */
3438         if (sh->reconstruct_state == reconstruct_state_result) {
3439                 sh->reconstruct_state = reconstruct_state_idle;
3440                 clear_bit(STRIPE_EXPANDING, &sh->state);
3441                 for (i = conf->raid_disks; i--; ) {
3442                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
3443                         set_bit(R5_LOCKED, &sh->dev[i].flags);
3444                         s.locked++;
3445                 }
3446         }
3447
3448         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3449             !sh->reconstruct_state) {
3450                 struct stripe_head *sh2
3451                         = get_active_stripe(conf, sh->sector, 1, 1, 1);
3452                 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3453                         /* sh cannot be written until sh2 has been read.
3454                          * so arrange for sh to be delayed a little
3455                          */
3456                         set_bit(STRIPE_DELAYED, &sh->state);
3457                         set_bit(STRIPE_HANDLE, &sh->state);
3458                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3459                                               &sh2->state))
3460                                 atomic_inc(&conf->preread_active_stripes);
3461                         release_stripe(sh2);
3462                         goto unlock;
3463                 }
3464                 if (sh2)
3465                         release_stripe(sh2);
3466
3467                 /* Need to write out all blocks after computing P&Q */
3468                 sh->disks = conf->raid_disks;
3469                 stripe_set_idx(sh->sector, conf, 0, sh);
3470                 schedule_reconstruction(sh, &s, 1, 1);
3471         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3472                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3473                 atomic_dec(&conf->reshape_stripes);
3474                 wake_up(&conf->wait_for_overlap);
3475                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3476         }
3477
3478         if (s.expanding && s.locked == 0 &&
3479             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3480                 handle_stripe_expansion(conf, sh, &r6s);
3481
3482  unlock:
3483         spin_unlock(&sh->lock);
3484
3485         /* wait for this device to become unblocked */
3486         if (unlikely(blocked_rdev))
3487                 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3488
3489         if (s.ops_request)
3490                 raid_run_ops(sh, s.ops_request);
3491
3492         ops_run_io(sh, &s);
3493
3494
3495         if (dec_preread_active) {
3496                 /* We delay this until after ops_run_io so that if make_request
3497                  * is waiting on a barrier, it won't continue until the writes
3498                  * have actually been submitted.
3499                  */
3500                 atomic_dec(&conf->preread_active_stripes);
3501                 if (atomic_read(&conf->preread_active_stripes) <
3502                     IO_THRESHOLD)
3503                         md_wakeup_thread(conf->mddev->thread);
3504         }
3505
3506         return_io(return_bi);
3507 }
3508
3509 static void handle_stripe(struct stripe_head *sh)
3510 {
3511         if (sh->raid_conf->level == 6)
3512                 handle_stripe6(sh);
3513         else
3514                 handle_stripe5(sh);
3515 }
3516
3517 static void raid5_activate_delayed(raid5_conf_t *conf)
3518 {
3519         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3520                 while (!list_empty(&conf->delayed_list)) {
3521                         struct list_head *l = conf->delayed_list.next;
3522                         struct stripe_head *sh;
3523                         sh = list_entry(l, struct stripe_head, lru);
3524                         list_del_init(l);
3525                         clear_bit(STRIPE_DELAYED, &sh->state);
3526                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3527                                 atomic_inc(&conf->preread_active_stripes);
3528                         list_add_tail(&sh->lru, &conf->hold_list);
3529                 }
3530         } else
3531                 blk_plug_device(conf->mddev->queue);
3532 }
3533
3534 static void activate_bit_delay(raid5_conf_t *conf)
3535 {
3536         /* device_lock is held */
3537         struct list_head head;
3538         list_add(&head, &conf->bitmap_list);
3539         list_del_init(&conf->bitmap_list);
3540         while (!list_empty(&head)) {
3541                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3542                 list_del_init(&sh->lru);
3543                 atomic_inc(&sh->count);
3544                 __release_stripe(conf, sh);
3545         }
3546 }
3547
3548 static void unplug_slaves(mddev_t *mddev)
3549 {
3550         raid5_conf_t *conf = mddev->private;
3551         int i;
3552         int devs = max(conf->raid_disks, conf->previous_raid_disks);
3553
3554         rcu_read_lock();
3555         for (i = 0; i < devs; i++) {
3556                 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3557                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3558                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3559
3560                         atomic_inc(&rdev->nr_pending);
3561                         rcu_read_unlock();
3562
3563                         blk_unplug(r_queue);
3564
3565                         rdev_dec_pending(rdev, mddev);
3566                         rcu_read_lock();
3567                 }
3568         }
3569         rcu_read_unlock();
3570 }
3571
3572 static void raid5_unplug_device(struct request_queue *q)
3573 {
3574         mddev_t *mddev = q->queuedata;
3575         raid5_conf_t *conf = mddev->private;
3576         unsigned long flags;
3577
3578         spin_lock_irqsave(&conf->device_lock, flags);
3579
3580         if (blk_remove_plug(q)) {
3581                 conf->seq_flush++;
3582                 raid5_activate_delayed(conf);
3583         }
3584         md_wakeup_thread(mddev->thread);
3585
3586         spin_unlock_irqrestore(&conf->device_lock, flags);
3587
3588         unplug_slaves(mddev);
3589 }
3590
3591 static int raid5_congested(void *data, int bits)
3592 {
3593         mddev_t *mddev = data;
3594         raid5_conf_t *conf = mddev->private;
3595
3596         /* No difference between reads and writes.  Just check
3597          * how busy the stripe_cache is
3598          */
3599
3600         if (mddev_congested(mddev, bits))
3601                 return 1;
3602         if (conf->inactive_blocked)
3603                 return 1;
3604         if (conf->quiesce)
3605                 return 1;
3606         if (list_empty_careful(&conf->inactive_list))
3607                 return 1;
3608
3609         return 0;
3610 }
3611
3612 /* We want read requests to align with chunks where possible,
3613  * but write requests don't need to.
3614  */
3615 static int raid5_mergeable_bvec(struct request_queue *q,
3616                                 struct bvec_merge_data *bvm,
3617                                 struct bio_vec *biovec)
3618 {
3619         mddev_t *mddev = q->queuedata;
3620         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3621         int max;
3622         unsigned int chunk_sectors = mddev->chunk_sectors;
3623         unsigned int bio_sectors = bvm->bi_size >> 9;
3624
3625         if ((bvm->bi_rw & 1) == WRITE)
3626                 return biovec->bv_len; /* always allow writes to be mergeable */
3627
3628         if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3629                 chunk_sectors = mddev->new_chunk_sectors;
3630         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3631         if (max < 0) max = 0;
3632         if (max <= biovec->bv_len && bio_sectors == 0)
3633                 return biovec->bv_len;
3634         else
3635                 return max;
3636 }
3637
3638
3639 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3640 {
3641         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3642         unsigned int chunk_sectors = mddev->chunk_sectors;
3643         unsigned int bio_sectors = bio->bi_size >> 9;
3644
3645         if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3646                 chunk_sectors = mddev->new_chunk_sectors;
3647         return  chunk_sectors >=
3648                 ((sector & (chunk_sectors - 1)) + bio_sectors);
3649 }
3650
3651 /*
3652  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3653  *  later sampled by raid5d.
3654  */
3655 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3656 {
3657         unsigned long flags;
3658
3659         spin_lock_irqsave(&conf->device_lock, flags);
3660
3661         bi->bi_next = conf->retry_read_aligned_list;
3662         conf->retry_read_aligned_list = bi;
3663
3664         spin_unlock_irqrestore(&conf->device_lock, flags);
3665         md_wakeup_thread(conf->mddev->thread);
3666 }
3667
3668
3669 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3670 {
3671         struct bio *bi;
3672
3673         bi = conf->retry_read_aligned;
3674         if (bi) {
3675                 conf->retry_read_aligned = NULL;
3676                 return bi;
3677         }
3678         bi = conf->retry_read_aligned_list;
3679         if(bi) {
3680                 conf->retry_read_aligned_list = bi->bi_next;
3681                 bi->bi_next = NULL;
3682                 /*
3683                  * this sets the active strip count to 1 and the processed
3684                  * strip count to zero (upper 8 bits)
3685                  */
3686                 bi->bi_phys_segments = 1; /* biased count of active stripes */
3687         }
3688
3689         return bi;
3690 }
3691
3692
3693 /*
3694  *  The "raid5_align_endio" should check if the read succeeded and if it
3695  *  did, call bio_endio on the original bio (having bio_put the new bio
3696  *  first).
3697  *  If the read failed..
3698  */
3699 static void raid5_align_endio(struct bio *bi, int error)
3700 {
3701         struct bio* raid_bi  = bi->bi_private;
3702         mddev_t *mddev;
3703         raid5_conf_t *conf;
3704         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3705         mdk_rdev_t *rdev;
3706
3707         bio_put(bi);
3708
3709         rdev = (void*)raid_bi->bi_next;
3710         raid_bi->bi_next = NULL;
3711         mddev = rdev->mddev;
3712         conf = mddev->private;
3713
3714         rdev_dec_pending(rdev, conf->mddev);
3715
3716         if (!error && uptodate) {
3717                 bio_endio(raid_bi, 0);
3718                 if (atomic_dec_and_test(&conf->active_aligned_reads))
3719                         wake_up(&conf->wait_for_stripe);
3720                 return;
3721         }
3722
3723
3724         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3725
3726         add_bio_to_retry(raid_bi, conf);
3727 }
3728
3729 static int bio_fits_rdev(struct bio *bi)
3730 {
3731         struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3732
3733         if ((bi->bi_size>>9) > queue_max_sectors(q))
3734                 return 0;
3735         blk_recount_segments(q, bi);
3736         if (bi->bi_phys_segments > queue_max_segments(q))
3737                 return 0;
3738
3739         if (q->merge_bvec_fn)
3740                 /* it's too hard to apply the merge_bvec_fn at this stage,
3741                  * just just give up
3742                  */
3743                 return 0;
3744
3745         return 1;
3746 }
3747
3748
3749 static int chunk_aligned_read(mddev_t *mddev, struct bio * raid_bio)
3750 {
3751         raid5_conf_t *conf = mddev->private;
3752         int dd_idx;
3753         struct bio* align_bi;
3754         mdk_rdev_t *rdev;
3755
3756         if (!in_chunk_boundary(mddev, raid_bio)) {
3757                 pr_debug("chunk_aligned_read : non aligned\n");
3758                 return 0;
3759         }
3760         /*
3761          * use bio_clone to make a copy of the bio
3762          */
3763         align_bi = bio_clone(raid_bio, GFP_NOIO);
3764         if (!align_bi)
3765                 return 0;
3766         /*
3767          *   set bi_end_io to a new function, and set bi_private to the
3768          *     original bio.
3769          */
3770         align_bi->bi_end_io  = raid5_align_endio;
3771         align_bi->bi_private = raid_bio;
3772         /*
3773          *      compute position
3774          */
3775         align_bi->bi_sector =  raid5_compute_sector(conf, raid_bio->bi_sector,
3776                                                     0,
3777                                                     &dd_idx, NULL);
3778
3779         rcu_read_lock();
3780         rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3781         if (rdev && test_bit(In_sync, &rdev->flags)) {
3782                 atomic_inc(&rdev->nr_pending);
3783                 rcu_read_unlock();
3784                 raid_bio->bi_next = (void*)rdev;
3785                 align_bi->bi_bdev =  rdev->bdev;
3786                 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3787                 align_bi->bi_sector += rdev->data_offset;
3788
3789                 if (!bio_fits_rdev(align_bi)) {
3790                         /* too big in some way */
3791                         bio_put(align_bi);
3792                         rdev_dec_pending(rdev, mddev);
3793                         return 0;
3794                 }
3795
3796                 spin_lock_irq(&conf->device_lock);
3797                 wait_event_lock_irq(conf->wait_for_stripe,
3798                                     conf->quiesce == 0,
3799                                     conf->device_lock, /* nothing */);
3800                 atomic_inc(&conf->active_aligned_reads);
3801                 spin_unlock_irq(&conf->device_lock);
3802
3803                 generic_make_request(align_bi);
3804                 return 1;
3805         } else {
3806                 rcu_read_unlock();
3807                 bio_put(align_bi);
3808                 return 0;
3809         }
3810 }
3811
3812 /* __get_priority_stripe - get the next stripe to process
3813  *
3814  * Full stripe writes are allowed to pass preread active stripes up until
3815  * the bypass_threshold is exceeded.  In general the bypass_count
3816  * increments when the handle_list is handled before the hold_list; however, it
3817  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3818  * stripe with in flight i/o.  The bypass_count will be reset when the
3819  * head of the hold_list has changed, i.e. the head was promoted to the
3820  * handle_list.
3821  */
3822 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3823 {
3824         struct stripe_head *sh;
3825
3826         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3827                   __func__,
3828                   list_empty(&conf->handle_list) ? "empty" : "busy",
3829                   list_empty(&conf->hold_list) ? "empty" : "busy",
3830                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
3831
3832         if (!list_empty(&conf->handle_list)) {
3833                 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3834
3835                 if (list_empty(&conf->hold_list))
3836                         conf->bypass_count = 0;
3837                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3838                         if (conf->hold_list.next == conf->last_hold)
3839                                 conf->bypass_count++;
3840                         else {
3841                                 conf->last_hold = conf->hold_list.next;
3842                                 conf->bypass_count -= conf->bypass_threshold;
3843                                 if (conf->bypass_count < 0)
3844                                         conf->bypass_count = 0;
3845                         }
3846                 }
3847         } else if (!list_empty(&conf->hold_list) &&
3848                    ((conf->bypass_threshold &&
3849                      conf->bypass_count > conf->bypass_threshold) ||
3850                     atomic_read(&conf->pending_full_writes) == 0)) {
3851                 sh = list_entry(conf->hold_list.next,
3852                                 typeof(*sh), lru);
3853                 conf->bypass_count -= conf->bypass_threshold;
3854                 if (conf->bypass_count < 0)
3855                         conf->bypass_count = 0;
3856         } else
3857                 return NULL;
3858
3859         list_del_init(&sh->lru);
3860         atomic_inc(&sh->count);
3861         BUG_ON(atomic_read(&sh->count) != 1);
3862         return sh;
3863 }
3864
3865 static int make_request(mddev_t *mddev, struct bio * bi)
3866 {
3867         raid5_conf_t *conf = mddev->private;
3868         int dd_idx;
3869         sector_t new_sector;
3870         sector_t logical_sector, last_sector;
3871         struct stripe_head *sh;
3872         const int rw = bio_data_dir(bi);
3873         int remaining;
3874
3875         if (unlikely(bio_rw_flagged(bi, BIO_RW_BARRIER))) {
3876                 /* Drain all pending writes.  We only really need
3877                  * to ensure they have been submitted, but this is
3878                  * easier.
3879                  */
3880                 mddev->pers->quiesce(mddev, 1);
3881                 mddev->pers->quiesce(mddev, 0);
3882                 md_barrier_request(mddev, bi);
3883                 return 0;
3884         }
3885
3886         md_write_start(mddev, bi);
3887
3888         if (rw == READ &&
3889              mddev->reshape_position == MaxSector &&
3890              chunk_aligned_read(mddev,bi))
3891                 return 0;
3892
3893         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3894         last_sector = bi->bi_sector + (bi->bi_size>>9);
3895         bi->bi_next = NULL;
3896         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
3897
3898         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3899                 DEFINE_WAIT(w);
3900                 int disks, data_disks;
3901                 int previous;
3902
3903         retry:
3904                 previous = 0;
3905                 disks = conf->raid_disks;
3906                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3907                 if (unlikely(conf->reshape_progress != MaxSector)) {
3908                         /* spinlock is needed as reshape_progress may be
3909                          * 64bit on a 32bit platform, and so it might be
3910                          * possible to see a half-updated value
3911                          * Ofcourse reshape_progress could change after
3912                          * the lock is dropped, so once we get a reference
3913                          * to the stripe that we think it is, we will have
3914                          * to check again.
3915                          */
3916                         spin_lock_irq(&conf->device_lock);
3917                         if (mddev->delta_disks < 0
3918                             ? logical_sector < conf->reshape_progress
3919                             : logical_sector >= conf->reshape_progress) {
3920                                 disks = conf->previous_raid_disks;
3921                                 previous = 1;
3922                         } else {
3923                                 if (mddev->delta_disks < 0
3924                                     ? logical_sector < conf->reshape_safe
3925                                     : logical_sector >= conf->reshape_safe) {
3926                                         spin_unlock_irq(&conf->device_lock);
3927                                         schedule();
3928                                         goto retry;
3929                                 }
3930                         }
3931                         spin_unlock_irq(&conf->device_lock);
3932                 }
3933                 data_disks = disks - conf->max_degraded;
3934
3935                 new_sector = raid5_compute_sector(conf, logical_sector,
3936                                                   previous,
3937                                                   &dd_idx, NULL);
3938                 pr_debug("raid456: make_request, sector %llu logical %llu\n",
3939                         (unsigned long long)new_sector, 
3940                         (unsigned long long)logical_sector);
3941
3942                 sh = get_active_stripe(conf, new_sector, previous,
3943                                        (bi->bi_rw&RWA_MASK), 0);
3944                 if (sh) {
3945                         if (unlikely(previous)) {
3946                                 /* expansion might have moved on while waiting for a
3947                                  * stripe, so we must do the range check again.
3948                                  * Expansion could still move past after this
3949                                  * test, but as we are holding a reference to
3950                                  * 'sh', we know that if that happens,
3951                                  *  STRIPE_EXPANDING will get set and the expansion
3952                                  * won't proceed until we finish with the stripe.
3953                                  */
3954                                 int must_retry = 0;
3955                                 spin_lock_irq(&conf->device_lock);
3956                                 if (mddev->delta_disks < 0
3957                                     ? logical_sector >= conf->reshape_progress
3958                                     : logical_sector < conf->reshape_progress)
3959                                         /* mismatch, need to try again */
3960                                         must_retry = 1;
3961                                 spin_unlock_irq(&conf->device_lock);
3962                                 if (must_retry) {
3963                                         release_stripe(sh);
3964                                         schedule();
3965                                         goto retry;
3966                                 }
3967                         }
3968
3969                         if (bio_data_dir(bi) == WRITE &&
3970                             logical_sector >= mddev->suspend_lo &&
3971                             logical_sector < mddev->suspend_hi) {
3972                                 release_stripe(sh);
3973                                 /* As the suspend_* range is controlled by
3974                                  * userspace, we want an interruptible
3975                                  * wait.
3976                                  */
3977                                 flush_signals(current);
3978                                 prepare_to_wait(&conf->wait_for_overlap,
3979                                                 &w, TASK_INTERRUPTIBLE);
3980                                 if (logical_sector >= mddev->suspend_lo &&
3981                                     logical_sector < mddev->suspend_hi)
3982                                         schedule();
3983                                 goto retry;
3984                         }
3985
3986                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3987                             !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3988                                 /* Stripe is busy expanding or
3989                                  * add failed due to overlap.  Flush everything
3990                                  * and wait a while
3991                                  */
3992                                 raid5_unplug_device(mddev->queue);
3993                                 release_stripe(sh);
3994                                 schedule();
3995                                 goto retry;
3996                         }
3997                         finish_wait(&conf->wait_for_overlap, &w);
3998                         set_bit(STRIPE_HANDLE, &sh->state);
3999                         clear_bit(STRIPE_DELAYED, &sh->state);
4000                         if (mddev->barrier && 
4001                             !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4002                                 atomic_inc(&conf->preread_active_stripes);
4003                         release_stripe(sh);
4004                 } else {
4005                         /* cannot get stripe for read-ahead, just give-up */
4006                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
4007                         finish_wait(&conf->wait_for_overlap, &w);
4008                         break;
4009                 }
4010                         
4011         }
4012         spin_lock_irq(&conf->device_lock);
4013         remaining = raid5_dec_bi_phys_segments(bi);
4014         spin_unlock_irq(&conf->device_lock);
4015         if (remaining == 0) {
4016
4017                 if ( rw == WRITE )
4018                         md_write_end(mddev);
4019
4020                 bio_endio(bi, 0);
4021         }
4022
4023         if (mddev->barrier) {
4024                 /* We need to wait for the stripes to all be handled.
4025                  * So: wait for preread_active_stripes to drop to 0.
4026                  */
4027                 wait_event(mddev->thread->wqueue,
4028                            atomic_read(&conf->preread_active_stripes) == 0);
4029         }
4030         return 0;
4031 }
4032
4033 static sector_t raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks);
4034
4035 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
4036 {
4037         /* reshaping is quite different to recovery/resync so it is
4038          * handled quite separately ... here.
4039          *
4040          * On each call to sync_request, we gather one chunk worth of
4041          * destination stripes and flag them as expanding.
4042          * Then we find all the source stripes and request reads.
4043          * As the reads complete, handle_stripe will copy the data
4044          * into the destination stripe and release that stripe.
4045          */
4046         raid5_conf_t *conf = mddev->private;
4047         struct stripe_head *sh;
4048         sector_t first_sector, last_sector;
4049         int raid_disks = conf->previous_raid_disks;
4050         int data_disks = raid_disks - conf->max_degraded;
4051         int new_data_disks = conf->raid_disks - conf->max_degraded;
4052         int i;
4053         int dd_idx;
4054         sector_t writepos, readpos, safepos;
4055         sector_t stripe_addr;
4056         int reshape_sectors;
4057         struct list_head stripes;
4058
4059         if (sector_nr == 0) {
4060                 /* If restarting in the middle, skip the initial sectors */
4061                 if (mddev->delta_disks < 0 &&
4062                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4063                         sector_nr = raid5_size(mddev, 0, 0)
4064                                 - conf->reshape_progress;
4065                 } else if (mddev->delta_disks >= 0 &&
4066                            conf->reshape_progress > 0)
4067                         sector_nr = conf->reshape_progress;
4068                 sector_div(sector_nr, new_data_disks);
4069                 if (sector_nr) {
4070                         mddev->curr_resync_completed = sector_nr;
4071                         sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4072                         *skipped = 1;
4073                         return sector_nr;
4074                 }
4075         }
4076
4077         /* We need to process a full chunk at a time.
4078          * If old and new chunk sizes differ, we need to process the
4079          * largest of these
4080          */
4081         if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4082                 reshape_sectors = mddev->new_chunk_sectors;
4083         else
4084                 reshape_sectors = mddev->chunk_sectors;
4085
4086         /* we update the metadata when there is more than 3Meg
4087          * in the block range (that is rather arbitrary, should
4088          * probably be time based) or when the data about to be
4089          * copied would over-write the source of the data at
4090          * the front of the range.
4091          * i.e. one new_stripe along from reshape_progress new_maps
4092          * to after where reshape_safe old_maps to
4093          */
4094         writepos = conf->reshape_progress;
4095         sector_div(writepos, new_data_disks);
4096         readpos = conf->reshape_progress;
4097         sector_div(readpos, data_disks);
4098         safepos = conf->reshape_safe;
4099         sector_div(safepos, data_disks);
4100         if (mddev->delta_disks < 0) {
4101                 writepos -= min_t(sector_t, reshape_sectors, writepos);
4102                 readpos += reshape_sectors;
4103                 safepos += reshape_sectors;
4104         } else {
4105                 writepos += reshape_sectors;
4106                 readpos -= min_t(sector_t, reshape_sectors, readpos);
4107                 safepos -= min_t(sector_t, reshape_sectors, safepos);
4108         }
4109
4110         /* 'writepos' is the most advanced device address we might write.
4111          * 'readpos' is the least advanced device address we might read.
4112          * 'safepos' is the least address recorded in the metadata as having
4113          *     been reshaped.
4114          * If 'readpos' is behind 'writepos', then there is no way that we can
4115          * ensure safety in the face of a crash - that must be done by userspace
4116          * making a backup of the data.  So in that case there is no particular
4117          * rush to update metadata.
4118          * Otherwise if 'safepos' is behind 'writepos', then we really need to
4119          * update the metadata to advance 'safepos' to match 'readpos' so that
4120          * we can be safe in the event of a crash.
4121          * So we insist on updating metadata if safepos is behind writepos and
4122          * readpos is beyond writepos.
4123          * In any case, update the metadata every 10 seconds.
4124          * Maybe that number should be configurable, but I'm not sure it is
4125          * worth it.... maybe it could be a multiple of safemode_delay???
4126          */
4127         if ((mddev->delta_disks < 0
4128              ? (safepos > writepos && readpos < writepos)
4129              : (safepos < writepos && readpos > writepos)) ||
4130             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4131                 /* Cannot proceed until we've updated the superblock... */
4132                 wait_event(conf->wait_for_overlap,
4133                            atomic_read(&conf->reshape_stripes)==0);
4134                 mddev->reshape_position = conf->reshape_progress;
4135                 mddev->curr_resync_completed = mddev->curr_resync;
4136                 conf->reshape_checkpoint = jiffies;
4137                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4138                 md_wakeup_thread(mddev->thread);
4139                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4140                            kthread_should_stop());
4141                 spin_lock_irq(&conf->device_lock);
4142                 conf->reshape_safe = mddev->reshape_position;
4143                 spin_unlock_irq(&conf->device_lock);
4144                 wake_up(&conf->wait_for_overlap);
4145                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4146         }
4147
4148         if (mddev->delta_disks < 0) {
4149                 BUG_ON(conf->reshape_progress == 0);
4150                 stripe_addr = writepos;
4151                 BUG_ON((mddev->dev_sectors &
4152                         ~((sector_t)reshape_sectors - 1))
4153                        - reshape_sectors - stripe_addr
4154                        != sector_nr);
4155         } else {
4156                 BUG_ON(writepos != sector_nr + reshape_sectors);
4157                 stripe_addr = sector_nr;
4158         }
4159         INIT_LIST_HEAD(&stripes);
4160         for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4161                 int j;
4162                 int skipped_disk = 0;
4163                 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4164                 set_bit(STRIPE_EXPANDING, &sh->state);
4165                 atomic_inc(&conf->reshape_stripes);
4166                 /* If any of this stripe is beyond the end of the old
4167                  * array, then we need to zero those blocks
4168                  */
4169                 for (j=sh->disks; j--;) {
4170                         sector_t s;
4171                         if (j == sh->pd_idx)
4172                                 continue;
4173                         if (conf->level == 6 &&
4174                             j == sh->qd_idx)
4175                                 continue;
4176                         s = compute_blocknr(sh, j, 0);
4177                         if (s < raid5_size(mddev, 0, 0)) {
4178                                 skipped_disk = 1;
4179                                 continue;
4180                         }
4181                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4182                         set_bit(R5_Expanded, &sh->dev[j].flags);
4183                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
4184                 }
4185                 if (!skipped_disk) {
4186                         set_bit(STRIPE_EXPAND_READY, &sh->state);
4187                         set_bit(STRIPE_HANDLE, &sh->state);
4188                 }
4189                 list_add(&sh->lru, &stripes);
4190         }
4191         spin_lock_irq(&conf->device_lock);
4192         if (mddev->delta_disks < 0)
4193                 conf->reshape_progress -= reshape_sectors * new_data_disks;
4194         else
4195                 conf->reshape_progress += reshape_sectors * new_data_disks;
4196         spin_unlock_irq(&conf->device_lock);
4197         /* Ok, those stripe are ready. We can start scheduling
4198          * reads on the source stripes.
4199          * The source stripes are determined by mapping the first and last
4200          * block on the destination stripes.
4201          */
4202         first_sector =
4203                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4204                                      1, &dd_idx, NULL);
4205         last_sector =
4206                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4207                                             * new_data_disks - 1),
4208                                      1, &dd_idx, NULL);
4209         if (last_sector >= mddev->dev_sectors)
4210                 last_sector = mddev->dev_sectors - 1;
4211         while (first_sector <= last_sector) {
4212                 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4213                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4214                 set_bit(STRIPE_HANDLE, &sh->state);
4215                 release_stripe(sh);
4216                 first_sector += STRIPE_SECTORS;
4217         }
4218         /* Now that the sources are clearly marked, we can release
4219          * the destination stripes
4220          */
4221         while (!list_empty(&stripes)) {
4222                 sh = list_entry(stripes.next, struct stripe_head, lru);
4223                 list_del_init(&sh->lru);
4224                 release_stripe(sh);
4225         }
4226         /* If this takes us to the resync_max point where we have to pause,
4227          * then we need to write out the superblock.
4228          */
4229         sector_nr += reshape_sectors;
4230         if ((sector_nr - mddev->curr_resync_completed) * 2
4231             >= mddev->resync_max - mddev->curr_resync_completed) {
4232                 /* Cannot proceed until we've updated the superblock... */
4233                 wait_event(conf->wait_for_overlap,
4234                            atomic_read(&conf->reshape_stripes) == 0);
4235                 mddev->reshape_position = conf->reshape_progress;
4236                 mddev->curr_resync_completed = mddev->curr_resync + reshape_sectors;
4237                 conf->reshape_checkpoint = jiffies;
4238                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4239                 md_wakeup_thread(mddev->thread);
4240                 wait_event(mddev->sb_wait,
4241                            !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4242                            || kthread_should_stop());
4243                 spin_lock_irq(&conf->device_lock);
4244                 conf->reshape_safe = mddev->reshape_position;
4245                 spin_unlock_irq(&conf->device_lock);
4246                 wake_up(&conf->wait_for_overlap);
4247                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4248         }
4249         return reshape_sectors;
4250 }
4251
4252 /* FIXME go_faster isn't used */
4253 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
4254 {
4255         raid5_conf_t *conf = mddev->private;
4256         struct stripe_head *sh;
4257         sector_t max_sector = mddev->dev_sectors;
4258         int sync_blocks;
4259         int still_degraded = 0;
4260         int i;
4261
4262         if (sector_nr >= max_sector) {
4263                 /* just being told to finish up .. nothing much to do */
4264                 unplug_slaves(mddev);
4265
4266                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4267                         end_reshape(conf);
4268                         return 0;
4269                 }
4270
4271                 if (mddev->curr_resync < max_sector) /* aborted */
4272                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4273                                         &sync_blocks, 1);
4274                 else /* completed sync */
4275                         conf->fullsync = 0;
4276                 bitmap_close_sync(mddev->bitmap);
4277
4278                 return 0;
4279         }
4280
4281         /* Allow raid5_quiesce to complete */
4282         wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4283
4284         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4285                 return reshape_request(mddev, sector_nr, skipped);
4286
4287         /* No need to check resync_max as we never do more than one
4288          * stripe, and as resync_max will always be on a chunk boundary,
4289          * if the check in md_do_sync didn't fire, there is no chance
4290          * of overstepping resync_max here
4291          */
4292
4293         /* if there is too many failed drives and we are trying
4294          * to resync, then assert that we are finished, because there is
4295          * nothing we can do.
4296          */
4297         if (mddev->degraded >= conf->max_degraded &&
4298             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4299                 sector_t rv = mddev->dev_sectors - sector_nr;
4300                 *skipped = 1;
4301                 return rv;
4302         }
4303         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4304             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4305             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4306                 /* we can skip this block, and probably more */
4307                 sync_blocks /= STRIPE_SECTORS;
4308                 *skipped = 1;
4309                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4310         }
4311
4312
4313         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4314
4315         sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4316         if (sh == NULL) {
4317                 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4318                 /* make sure we don't swamp the stripe cache if someone else
4319                  * is trying to get access
4320                  */
4321                 schedule_timeout_uninterruptible(1);
4322         }
4323         /* Need to check if array will still be degraded after recovery/resync
4324          * We don't need to check the 'failed' flag as when that gets set,
4325          * recovery aborts.
4326          */
4327         for (i = 0; i < conf->raid_disks; i++)
4328                 if (conf->disks[i].rdev == NULL)
4329                         still_degraded = 1;
4330
4331         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4332
4333         spin_lock(&sh->lock);
4334         set_bit(STRIPE_SYNCING, &sh->state);
4335         clear_bit(STRIPE_INSYNC, &sh->state);
4336         spin_unlock(&sh->lock);
4337
4338         handle_stripe(sh);
4339         release_stripe(sh);
4340
4341         return STRIPE_SECTORS;
4342 }
4343
4344 static int  retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
4345 {
4346         /* We may not be able to submit a whole bio at once as there
4347          * may not be enough stripe_heads available.
4348          * We cannot pre-allocate enough stripe_heads as we may need
4349          * more than exist in the cache (if we allow ever large chunks).
4350          * So we do one stripe head at a time and record in
4351          * ->bi_hw_segments how many have been done.
4352          *
4353          * We *know* that this entire raid_bio is in one chunk, so
4354          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4355          */
4356         struct stripe_head *sh;
4357         int dd_idx;
4358         sector_t sector, logical_sector, last_sector;
4359         int scnt = 0;
4360         int remaining;
4361         int handled = 0;
4362
4363         logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4364         sector = raid5_compute_sector(conf, logical_sector,
4365                                       0, &dd_idx, NULL);
4366         last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4367
4368         for (; logical_sector < last_sector;
4369              logical_sector += STRIPE_SECTORS,
4370                      sector += STRIPE_SECTORS,
4371                      scnt++) {
4372
4373                 if (scnt < raid5_bi_hw_segments(raid_bio))
4374                         /* already done this stripe */
4375                         continue;
4376
4377                 sh = get_active_stripe(conf, sector, 0, 1, 0);
4378
4379                 if (!sh) {
4380                         /* failed to get a stripe - must wait */
4381                         raid5_set_bi_hw_segments(raid_bio, scnt);
4382                         conf->retry_read_aligned = raid_bio;
4383                         return handled;
4384                 }
4385
4386                 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4387                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4388                         release_stripe(sh);
4389                         raid5_set_bi_hw_segments(raid_bio, scnt);
4390                         conf->retry_read_aligned = raid_bio;
4391                         return handled;
4392                 }
4393
4394                 handle_stripe(sh);
4395                 release_stripe(sh);
4396                 handled++;
4397         }
4398         spin_lock_irq(&conf->device_lock);
4399         remaining = raid5_dec_bi_phys_segments(raid_bio);
4400         spin_unlock_irq(&conf->device_lock);
4401         if (remaining == 0)
4402                 bio_endio(raid_bio, 0);
4403         if (atomic_dec_and_test(&conf->active_aligned_reads))
4404                 wake_up(&conf->wait_for_stripe);
4405         return handled;
4406 }
4407
4408
4409 /*
4410  * This is our raid5 kernel thread.
4411  *
4412  * We scan the hash table for stripes which can be handled now.
4413  * During the scan, completed stripes are saved for us by the interrupt
4414  * handler, so that they will not have to wait for our next wakeup.
4415  */
4416 static void raid5d(mddev_t *mddev)
4417 {
4418         struct stripe_head *sh;
4419         raid5_conf_t *conf = mddev->private;
4420         int handled;
4421
4422         pr_debug("+++ raid5d active\n");
4423
4424         md_check_recovery(mddev);
4425
4426         handled = 0;
4427         spin_lock_irq(&conf->device_lock);
4428         while (1) {
4429                 struct bio *bio;
4430
4431                 if (conf->seq_flush != conf->seq_write) {
4432                         int seq = conf->seq_flush;
4433                         spin_unlock_irq(&conf->device_lock);
4434                         bitmap_unplug(mddev->bitmap);
4435                         spin_lock_irq(&conf->device_lock);
4436                         conf->seq_write = seq;
4437                         activate_bit_delay(conf);
4438                 }
4439
4440                 while ((bio = remove_bio_from_retry(conf))) {
4441                         int ok;
4442                         spin_unlock_irq(&conf->device_lock);
4443                         ok = retry_aligned_read(conf, bio);
4444                         spin_lock_irq(&conf->device_lock);
4445                         if (!ok)
4446                                 break;
4447                         handled++;
4448                 }
4449
4450                 sh = __get_priority_stripe(conf);
4451
4452                 if (!sh)
4453                         break;
4454                 spin_unlock_irq(&conf->device_lock);
4455                 
4456                 handled++;
4457                 handle_stripe(sh);
4458                 release_stripe(sh);
4459                 cond_resched();
4460
4461                 spin_lock_irq(&conf->device_lock);
4462         }
4463         pr_debug("%d stripes handled\n", handled);
4464
4465         spin_unlock_irq(&conf->device_lock);
4466
4467         async_tx_issue_pending_all();
4468         unplug_slaves(mddev);
4469
4470         pr_debug("--- raid5d inactive\n");
4471 }
4472
4473 static ssize_t
4474 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4475 {
4476         raid5_conf_t *conf = mddev->private;
4477         if (conf)
4478                 return sprintf(page, "%d\n", conf->max_nr_stripes);
4479         else
4480                 return 0;
4481 }
4482
4483 static ssize_t
4484 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4485 {
4486         raid5_conf_t *conf = mddev->private;
4487         unsigned long new;
4488         int err;
4489
4490         if (len >= PAGE_SIZE)
4491                 return -EINVAL;
4492         if (!conf)
4493                 return -ENODEV;
4494
4495         if (strict_strtoul(page, 10, &new))
4496                 return -EINVAL;
4497         if (new <= 16 || new > 32768)
4498                 return -EINVAL;
4499         while (new < conf->max_nr_stripes) {
4500                 if (drop_one_stripe(conf))
4501                         conf->max_nr_stripes--;
4502                 else
4503                         break;
4504         }
4505         err = md_allow_write(mddev);
4506         if (err)
4507                 return err;
4508         while (new > conf->max_nr_stripes) {
4509                 if (grow_one_stripe(conf))
4510                         conf->max_nr_stripes++;
4511                 else break;
4512         }
4513         return len;
4514 }
4515
4516 static struct md_sysfs_entry
4517 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4518                                 raid5_show_stripe_cache_size,
4519                                 raid5_store_stripe_cache_size);
4520
4521 static ssize_t
4522 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4523 {
4524         raid5_conf_t *conf = mddev->private;
4525         if (conf)
4526                 return sprintf(page, "%d\n", conf->bypass_threshold);
4527         else
4528                 return 0;
4529 }
4530
4531 static ssize_t
4532 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4533 {
4534         raid5_conf_t *conf = mddev->private;
4535         unsigned long new;
4536         if (len >= PAGE_SIZE)
4537                 return -EINVAL;
4538         if (!conf)
4539                 return -ENODEV;
4540
4541         if (strict_strtoul(page, 10, &new))
4542                 return -EINVAL;
4543         if (new > conf->max_nr_stripes)
4544                 return -EINVAL;
4545         conf->bypass_threshold = new;
4546         return len;
4547 }
4548
4549 static struct md_sysfs_entry
4550 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4551                                         S_IRUGO | S_IWUSR,
4552                                         raid5_show_preread_threshold,
4553                                         raid5_store_preread_threshold);
4554
4555 static ssize_t
4556 stripe_cache_active_show(mddev_t *mddev, char *page)
4557 {
4558         raid5_conf_t *conf = mddev->private;
4559         if (conf)
4560                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4561         else
4562                 return 0;
4563 }
4564
4565 static struct md_sysfs_entry
4566 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4567
4568 static struct attribute *raid5_attrs[] =  {
4569         &raid5_stripecache_size.attr,
4570         &raid5_stripecache_active.attr,
4571         &raid5_preread_bypass_threshold.attr,
4572         NULL,
4573 };
4574 static struct attribute_group raid5_attrs_group = {
4575         .name = NULL,
4576         .attrs = raid5_attrs,
4577 };
4578
4579 static sector_t
4580 raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks)
4581 {
4582         raid5_conf_t *conf = mddev->private;
4583
4584         if (!sectors)
4585                 sectors = mddev->dev_sectors;
4586         if (!raid_disks)
4587                 /* size is defined by the smallest of previous and new size */
4588                 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4589
4590         sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4591         sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4592         return sectors * (raid_disks - conf->max_degraded);
4593 }
4594
4595 static void raid5_free_percpu(raid5_conf_t *conf)
4596 {
4597         struct raid5_percpu *percpu;
4598         unsigned long cpu;
4599
4600         if (!conf->percpu)
4601                 return;
4602
4603         get_online_cpus();
4604         for_each_possible_cpu(cpu) {
4605                 percpu = per_cpu_ptr(conf->percpu, cpu);
4606                 safe_put_page(percpu->spare_page);
4607                 kfree(percpu->scribble);
4608         }
4609 #ifdef CONFIG_HOTPLUG_CPU
4610         unregister_cpu_notifier(&conf->cpu_notify);
4611 #endif
4612         put_online_cpus();
4613
4614         free_percpu(conf->percpu);
4615 }
4616
4617 static void free_conf(raid5_conf_t *conf)
4618 {
4619         shrink_stripes(conf);
4620         raid5_free_percpu(conf);
4621         kfree(conf->disks);
4622         kfree(conf->stripe_hashtbl);
4623         kfree(conf);
4624 }
4625
4626 #ifdef CONFIG_HOTPLUG_CPU
4627 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4628                               void *hcpu)
4629 {
4630         raid5_conf_t *conf = container_of(nfb, raid5_conf_t, cpu_notify);
4631         long cpu = (long)hcpu;
4632         struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4633
4634         switch (action) {
4635         case CPU_UP_PREPARE:
4636         case CPU_UP_PREPARE_FROZEN:
4637                 if (conf->level == 6 && !percpu->spare_page)
4638                         percpu->spare_page = alloc_page(GFP_KERNEL);
4639                 if (!percpu->scribble)
4640                         percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4641
4642                 if (!percpu->scribble ||
4643                     (conf->level == 6 && !percpu->spare_page)) {
4644                         safe_put_page(percpu->spare_page);
4645                         kfree(percpu->scribble);
4646                         pr_err("%s: failed memory allocation for cpu%ld\n",
4647                                __func__, cpu);
4648                         return NOTIFY_BAD;
4649                 }
4650                 break;
4651         case CPU_DEAD:
4652         case CPU_DEAD_FROZEN:
4653                 safe_put_page(percpu->spare_page);
4654                 kfree(percpu->scribble);
4655                 percpu->spare_page = NULL;
4656                 percpu->scribble = NULL;
4657                 break;
4658         default:
4659                 break;
4660         }
4661         return NOTIFY_OK;
4662 }
4663 #endif
4664
4665 static int raid5_alloc_percpu(raid5_conf_t *conf)
4666 {
4667         unsigned long cpu;
4668         struct page *spare_page;
4669         struct raid5_percpu __percpu *allcpus;
4670         void *scribble;
4671         int err;
4672
4673         allcpus = alloc_percpu(struct raid5_percpu);
4674         if (!allcpus)
4675                 return -ENOMEM;
4676         conf->percpu = allcpus;
4677
4678         get_online_cpus();
4679         err = 0;
4680         for_each_present_cpu(cpu) {
4681                 if (conf->level == 6) {
4682                         spare_page = alloc_page(GFP_KERNEL);
4683                         if (!spare_page) {
4684                                 err = -ENOMEM;
4685                                 break;
4686                         }
4687                         per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4688                 }
4689                 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4690                 if (!scribble) {
4691                         err = -ENOMEM;
4692                         break;
4693                 }
4694                 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4695         }
4696 #ifdef CONFIG_HOTPLUG_CPU
4697         conf->cpu_notify.notifier_call = raid456_cpu_notify;
4698         conf->cpu_notify.priority = 0;
4699         if (err == 0)
4700                 err = register_cpu_notifier(&conf->cpu_notify);
4701 #endif
4702         put_online_cpus();
4703
4704         return err;
4705 }
4706
4707 static raid5_conf_t *setup_conf(mddev_t *mddev)
4708 {
4709         raid5_conf_t *conf;
4710         int raid_disk, memory, max_disks;
4711         mdk_rdev_t *rdev;
4712         struct disk_info *disk;
4713
4714         if (mddev->new_level != 5
4715             && mddev->new_level != 4
4716             && mddev->new_level != 6) {
4717                 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4718                        mdname(mddev), mddev->new_level);
4719                 return ERR_PTR(-EIO);
4720         }
4721         if ((mddev->new_level == 5
4722              && !algorithm_valid_raid5(mddev->new_layout)) ||
4723             (mddev->new_level == 6
4724              && !algorithm_valid_raid6(mddev->new_layout))) {
4725                 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
4726                        mdname(mddev), mddev->new_layout);
4727                 return ERR_PTR(-EIO);
4728         }
4729         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4730                 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4731                        mdname(mddev), mddev->raid_disks);
4732                 return ERR_PTR(-EINVAL);
4733         }
4734
4735         if (!mddev->new_chunk_sectors ||
4736             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4737             !is_power_of_2(mddev->new_chunk_sectors)) {
4738                 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
4739                        mdname(mddev), mddev->new_chunk_sectors << 9);
4740                 return ERR_PTR(-EINVAL);
4741         }
4742
4743         conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL);
4744         if (conf == NULL)
4745                 goto abort;
4746         spin_lock_init(&conf->device_lock);
4747         init_waitqueue_head(&conf->wait_for_stripe);
4748         init_waitqueue_head(&conf->wait_for_overlap);
4749         INIT_LIST_HEAD(&conf->handle_list);
4750         INIT_LIST_HEAD(&conf->hold_list);
4751         INIT_LIST_HEAD(&conf->delayed_list);
4752         INIT_LIST_HEAD(&conf->bitmap_list);
4753         INIT_LIST_HEAD(&conf->inactive_list);
4754         atomic_set(&conf->active_stripes, 0);
4755         atomic_set(&conf->preread_active_stripes, 0);
4756         atomic_set(&conf->active_aligned_reads, 0);
4757         conf->bypass_threshold = BYPASS_THRESHOLD;
4758
4759         conf->raid_disks = mddev->raid_disks;
4760         if (mddev->reshape_position == MaxSector)
4761                 conf->previous_raid_disks = mddev->raid_disks;
4762         else
4763                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4764         max_disks = max(conf->raid_disks, conf->previous_raid_disks);
4765         conf->scribble_len = scribble_len(max_disks);
4766
4767         conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
4768                               GFP_KERNEL);
4769         if (!conf->disks)
4770                 goto abort;
4771
4772         conf->mddev = mddev;
4773
4774         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4775                 goto abort;
4776
4777         conf->level = mddev->new_level;
4778         if (raid5_alloc_percpu(conf) != 0)
4779                 goto abort;
4780
4781         pr_debug("raid456: run(%s) called.\n", mdname(mddev));
4782
4783         list_for_each_entry(rdev, &mddev->disks, same_set) {
4784                 raid_disk = rdev->raid_disk;
4785                 if (raid_disk >= max_disks
4786                     || raid_disk < 0)
4787                         continue;
4788                 disk = conf->disks + raid_disk;
4789
4790                 disk->rdev = rdev;
4791
4792                 if (test_bit(In_sync, &rdev->flags)) {
4793                         char b[BDEVNAME_SIZE];
4794                         printk(KERN_INFO "md/raid:%s: device %s operational as raid"
4795                                " disk %d\n",
4796                                mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
4797                 } else
4798                         /* Cannot rely on bitmap to complete recovery */
4799                         conf->fullsync = 1;
4800         }
4801
4802         conf->chunk_sectors = mddev->new_chunk_sectors;
4803         conf->level = mddev->new_level;
4804         if (conf->level == 6)
4805                 conf->max_degraded = 2;
4806         else
4807                 conf->max_degraded = 1;
4808         conf->algorithm = mddev->new_layout;
4809         conf->max_nr_stripes = NR_STRIPES;
4810         conf->reshape_progress = mddev->reshape_position;
4811         if (conf->reshape_progress != MaxSector) {
4812                 conf->prev_chunk_sectors = mddev->chunk_sectors;
4813                 conf->prev_algo = mddev->layout;
4814         }
4815
4816         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4817                  max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4818         if (grow_stripes(conf, conf->max_nr_stripes)) {
4819                 printk(KERN_ERR
4820                        "md/raid:%s: couldn't allocate %dkB for buffers\n",
4821                        mdname(mddev), memory);
4822                 goto abort;
4823         } else
4824                 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
4825                        mdname(mddev), memory);
4826
4827         conf->thread = md_register_thread(raid5d, mddev, NULL);
4828         if (!conf->thread) {
4829                 printk(KERN_ERR
4830                        "md/raid:%s: couldn't allocate thread.\n",
4831                        mdname(mddev));
4832                 goto abort;
4833         }
4834
4835         return conf;
4836
4837  abort:
4838         if (conf) {
4839                 free_conf(conf);
4840                 return ERR_PTR(-EIO);
4841         } else
4842                 return ERR_PTR(-ENOMEM);
4843 }
4844
4845
4846 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
4847 {
4848         switch (algo) {
4849         case ALGORITHM_PARITY_0:
4850                 if (raid_disk < max_degraded)
4851                         return 1;
4852                 break;
4853         case ALGORITHM_PARITY_N:
4854                 if (raid_disk >= raid_disks - max_degraded)
4855                         return 1;
4856                 break;
4857         case ALGORITHM_PARITY_0_6:
4858                 if (raid_disk == 0 || 
4859                     raid_disk == raid_disks - 1)
4860                         return 1;
4861                 break;
4862         case ALGORITHM_LEFT_ASYMMETRIC_6:
4863         case ALGORITHM_RIGHT_ASYMMETRIC_6:
4864         case ALGORITHM_LEFT_SYMMETRIC_6:
4865         case ALGORITHM_RIGHT_SYMMETRIC_6:
4866                 if (raid_disk == raid_disks - 1)
4867                         return 1;
4868         }
4869         return 0;
4870 }
4871
4872 static int run(mddev_t *mddev)
4873 {
4874         raid5_conf_t *conf;
4875         int working_disks = 0, chunk_size;
4876         int dirty_parity_disks = 0;
4877         mdk_rdev_t *rdev;
4878         sector_t reshape_offset = 0;
4879
4880         if (mddev->recovery_cp != MaxSector)
4881                 printk(KERN_NOTICE "md/raid:%s: not clean"
4882                        " -- starting background reconstruction\n",
4883                        mdname(mddev));
4884         if (mddev->reshape_position != MaxSector) {
4885                 /* Check that we can continue the reshape.
4886                  * Currently only disks can change, it must
4887                  * increase, and we must be past the point where
4888                  * a stripe over-writes itself
4889                  */
4890                 sector_t here_new, here_old;
4891                 int old_disks;
4892                 int max_degraded = (mddev->level == 6 ? 2 : 1);
4893
4894                 if (mddev->new_level != mddev->level) {
4895                         printk(KERN_ERR "md/raid:%s: unsupported reshape "
4896                                "required - aborting.\n",
4897                                mdname(mddev));
4898                         return -EINVAL;
4899                 }
4900                 old_disks = mddev->raid_disks - mddev->delta_disks;
4901                 /* reshape_position must be on a new-stripe boundary, and one
4902                  * further up in new geometry must map after here in old
4903                  * geometry.
4904                  */
4905                 here_new = mddev->reshape_position;
4906                 if (sector_div(here_new, mddev->new_chunk_sectors *
4907                                (mddev->raid_disks - max_degraded))) {
4908                         printk(KERN_ERR "md/raid:%s: reshape_position not "
4909                                "on a stripe boundary\n", mdname(mddev));
4910                         return -EINVAL;
4911                 }
4912                 reshape_offset = here_new * mddev->new_chunk_sectors;
4913                 /* here_new is the stripe we will write to */
4914                 here_old = mddev->reshape_position;
4915                 sector_div(here_old, mddev->chunk_sectors *
4916                            (old_disks-max_degraded));
4917                 /* here_old is the first stripe that we might need to read
4918                  * from */
4919                 if (mddev->delta_disks == 0) {
4920                         /* We cannot be sure it is safe to start an in-place
4921                          * reshape.  It is only safe if user-space if monitoring
4922                          * and taking constant backups.
4923                          * mdadm always starts a situation like this in
4924                          * readonly mode so it can take control before
4925                          * allowing any writes.  So just check for that.
4926                          */
4927                         if ((here_new * mddev->new_chunk_sectors != 
4928                              here_old * mddev->chunk_sectors) ||
4929                             mddev->ro == 0) {
4930                                 printk(KERN_ERR "md/raid:%s: in-place reshape must be started"
4931                                        " in read-only mode - aborting\n",
4932                                        mdname(mddev));
4933                                 return -EINVAL;
4934                         }
4935                 } else if (mddev->delta_disks < 0
4936                     ? (here_new * mddev->new_chunk_sectors <=
4937                        here_old * mddev->chunk_sectors)
4938                     : (here_new * mddev->new_chunk_sectors >=
4939                        here_old * mddev->chunk_sectors)) {
4940                         /* Reading from the same stripe as writing to - bad */
4941                         printk(KERN_ERR "md/raid:%s: reshape_position too early for "
4942                                "auto-recovery - aborting.\n",
4943                                mdname(mddev));
4944                         return -EINVAL;
4945                 }
4946                 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
4947                        mdname(mddev));
4948                 /* OK, we should be able to continue; */
4949         } else {
4950                 BUG_ON(mddev->level != mddev->new_level);
4951                 BUG_ON(mddev->layout != mddev->new_layout);
4952                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
4953                 BUG_ON(mddev->delta_disks != 0);
4954         }
4955
4956         if (mddev->private == NULL)
4957                 conf = setup_conf(mddev);
4958         else
4959                 conf = mddev->private;
4960
4961         if (IS_ERR(conf))
4962                 return PTR_ERR(conf);
4963
4964         mddev->thread = conf->thread;
4965         conf->thread = NULL;
4966         mddev->private = conf;
4967
4968         /*
4969          * 0 for a fully functional array, 1 or 2 for a degraded array.
4970          */
4971         list_for_each_entry(rdev, &mddev->disks, same_set) {
4972                 if (rdev->raid_disk < 0)
4973                         continue;
4974                 if (test_bit(In_sync, &rdev->flags))
4975                         working_disks++;
4976                 /* This disc is not fully in-sync.  However if it
4977                  * just stored parity (beyond the recovery_offset),
4978                  * when we don't need to be concerned about the
4979                  * array being dirty.
4980                  * When reshape goes 'backwards', we never have
4981                  * partially completed devices, so we only need
4982                  * to worry about reshape going forwards.
4983                  */
4984                 /* Hack because v0.91 doesn't store recovery_offset properly. */
4985                 if (mddev->major_version == 0 &&
4986                     mddev->minor_version > 90)
4987                         rdev->recovery_offset = reshape_offset;
4988                         
4989                 if (rdev->recovery_offset < reshape_offset) {
4990                         /* We need to check old and new layout */
4991                         if (!only_parity(rdev->raid_disk,
4992                                          conf->algorithm,
4993                                          conf->raid_disks,
4994                                          conf->max_degraded))
4995                                 continue;
4996                 }
4997                 if (!only_parity(rdev->raid_disk,
4998                                  conf->prev_algo,
4999                                  conf->previous_raid_disks,
5000                                  conf->max_degraded))
5001                         continue;
5002                 dirty_parity_disks++;
5003         }
5004
5005         mddev->degraded = (max(conf->raid_disks, conf->previous_raid_disks)
5006                            - working_disks);
5007
5008         if (mddev->degraded > conf->max_degraded) {
5009                 printk(KERN_ERR "md/raid:%s: not enough operational devices"
5010                         " (%d/%d failed)\n",
5011                         mdname(mddev), mddev->degraded, conf->raid_disks);
5012                 goto abort;
5013         }
5014
5015         /* device size must be a multiple of chunk size */
5016         mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
5017         mddev->resync_max_sectors = mddev->dev_sectors;
5018
5019         if (mddev->degraded > dirty_parity_disks &&
5020             mddev->recovery_cp != MaxSector) {
5021                 if (mddev->ok_start_degraded)
5022                         printk(KERN_WARNING
5023                                "md/raid:%s: starting dirty degraded array"
5024                                " - data corruption possible.\n",
5025                                mdname(mddev));
5026                 else {
5027                         printk(KERN_ERR
5028                                "md/raid:%s: cannot start dirty degraded array.\n",
5029                                mdname(mddev));
5030                         goto abort;
5031                 }
5032         }
5033
5034         if (mddev->degraded == 0)
5035                 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
5036                        " devices, algorithm %d\n", mdname(mddev), conf->level,
5037                        mddev->raid_disks-mddev->degraded, mddev->raid_disks,
5038                        mddev->new_layout);
5039         else
5040                 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
5041                        " out of %d devices, algorithm %d\n",
5042                        mdname(mddev), conf->level,
5043                        mddev->raid_disks - mddev->degraded,
5044                        mddev->raid_disks, mddev->new_layout);
5045
5046         print_raid5_conf(conf);
5047
5048         if (conf->reshape_progress != MaxSector) {
5049                 conf->reshape_safe = conf->reshape_progress;
5050                 atomic_set(&conf->reshape_stripes, 0);
5051                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5052                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5053                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5054                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5055                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5056                                                         "reshape");
5057         }
5058
5059         /* read-ahead size must cover two whole stripes, which is
5060          * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5061          */
5062         {
5063                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
5064                 int stripe = data_disks *
5065                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
5066                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5067                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5068         }
5069
5070         /* Ok, everything is just fine now */
5071         if (mddev->to_remove == &raid5_attrs_group)
5072                 mddev->to_remove = NULL;
5073         else if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5074                 printk(KERN_WARNING
5075                        "md/raid:%s: failed to create sysfs attributes.\n",
5076                        mdname(mddev));
5077
5078         mddev->queue->queue_lock = &conf->device_lock;
5079
5080         mddev->queue->unplug_fn = raid5_unplug_device;
5081         mddev->queue->backing_dev_info.congested_data = mddev;
5082         mddev->queue->backing_dev_info.congested_fn = raid5_congested;
5083
5084         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5085
5086         blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
5087         chunk_size = mddev->chunk_sectors << 9;
5088         blk_queue_io_min(mddev->queue, chunk_size);
5089         blk_queue_io_opt(mddev->queue, chunk_size *
5090                          (conf->raid_disks - conf->max_degraded));
5091
5092         list_for_each_entry(rdev, &mddev->disks, same_set)
5093                 disk_stack_limits(mddev->gendisk, rdev->bdev,
5094                                   rdev->data_offset << 9);
5095
5096         return 0;
5097 abort:
5098         md_unregister_thread(mddev->thread);
5099         mddev->thread = NULL;
5100         if (conf) {
5101                 print_raid5_conf(conf);
5102                 free_conf(conf);
5103         }
5104         mddev->private = NULL;
5105         printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
5106         return -EIO;
5107 }
5108
5109 static int stop(mddev_t *mddev)
5110 {
5111         raid5_conf_t *conf = mddev->private;
5112
5113         md_unregister_thread(mddev->thread);
5114         mddev->thread = NULL;
5115         mddev->queue->backing_dev_info.congested_fn = NULL;
5116         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
5117         free_conf(conf);
5118         mddev->private = NULL;
5119         mddev->to_remove = &raid5_attrs_group;
5120         return 0;
5121 }
5122
5123 #ifdef DEBUG
5124 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
5125 {
5126         int i;
5127
5128         seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
5129                    (unsigned long long)sh->sector, sh->pd_idx, sh->state);
5130         seq_printf(seq, "sh %llu,  count %d.\n",
5131                    (unsigned long long)sh->sector, atomic_read(&sh->count));
5132         seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
5133         for (i = 0; i < sh->disks; i++) {
5134                 seq_printf(seq, "(cache%d: %p %ld) ",
5135                            i, sh->dev[i].page, sh->dev[i].flags);
5136         }
5137         seq_printf(seq, "\n");
5138 }
5139
5140 static void printall(struct seq_file *seq, raid5_conf_t *conf)
5141 {
5142         struct stripe_head *sh;
5143         struct hlist_node *hn;
5144         int i;
5145
5146         spin_lock_irq(&conf->device_lock);
5147         for (i = 0; i < NR_HASH; i++) {
5148                 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
5149                         if (sh->raid_conf != conf)
5150                                 continue;
5151                         print_sh(seq, sh);
5152                 }
5153         }
5154         spin_unlock_irq(&conf->device_lock);
5155 }
5156 #endif
5157
5158 static void status(struct seq_file *seq, mddev_t *mddev)
5159 {
5160         raid5_conf_t *conf = mddev->private;
5161         int i;
5162
5163         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5164                 mddev->chunk_sectors / 2, mddev->layout);
5165         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5166         for (i = 0; i < conf->raid_disks; i++)
5167                 seq_printf (seq, "%s",
5168                                conf->disks[i].rdev &&
5169                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5170         seq_printf (seq, "]");
5171 #ifdef DEBUG
5172         seq_printf (seq, "\n");
5173         printall(seq, conf);
5174 #endif
5175 }
5176
5177 static void print_raid5_conf (raid5_conf_t *conf)
5178 {
5179         int i;
5180         struct disk_info *tmp;
5181
5182         printk(KERN_DEBUG "RAID conf printout:\n");
5183         if (!conf) {
5184                 printk("(conf==NULL)\n");
5185                 return;
5186         }
5187         printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5188                conf->raid_disks,
5189                conf->raid_disks - conf->mddev->degraded);
5190
5191         for (i = 0; i < conf->raid_disks; i++) {
5192                 char b[BDEVNAME_SIZE];
5193                 tmp = conf->disks + i;
5194                 if (tmp->rdev)
5195                         printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5196                                i, !test_bit(Faulty, &tmp->rdev->flags),
5197                                bdevname(tmp->rdev->bdev, b));
5198         }
5199 }
5200
5201 static int raid5_spare_active(mddev_t *mddev)
5202 {
5203         int i;
5204         raid5_conf_t *conf = mddev->private;
5205         struct disk_info *tmp;
5206
5207         for (i = 0; i < conf->raid_disks; i++) {
5208                 tmp = conf->disks + i;
5209                 if (tmp->rdev
5210                     && !test_bit(Faulty, &tmp->rdev->flags)
5211                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5212                         unsigned long flags;
5213                         spin_lock_irqsave(&conf->device_lock, flags);
5214                         mddev->degraded--;
5215                         spin_unlock_irqrestore(&conf->device_lock, flags);
5216                 }
5217         }
5218         print_raid5_conf(conf);
5219         return 0;
5220 }
5221
5222 static int raid5_remove_disk(mddev_t *mddev, int number)
5223 {
5224         raid5_conf_t *conf = mddev->private;
5225         int err = 0;
5226         mdk_rdev_t *rdev;
5227         struct disk_info *p = conf->disks + number;
5228
5229         print_raid5_conf(conf);
5230         rdev = p->rdev;
5231         if (rdev) {
5232                 if (number >= conf->raid_disks &&
5233                     conf->reshape_progress == MaxSector)
5234                         clear_bit(In_sync, &rdev->flags);
5235
5236                 if (test_bit(In_sync, &rdev->flags) ||
5237                     atomic_read(&rdev->nr_pending)) {
5238                         err = -EBUSY;
5239                         goto abort;
5240                 }
5241                 /* Only remove non-faulty devices if recovery
5242                  * isn't possible.
5243                  */
5244                 if (!test_bit(Faulty, &rdev->flags) &&
5245                     mddev->degraded <= conf->max_degraded &&
5246                     number < conf->raid_disks) {
5247                         err = -EBUSY;
5248                         goto abort;
5249                 }
5250                 p->rdev = NULL;
5251                 synchronize_rcu();
5252                 if (atomic_read(&rdev->nr_pending)) {
5253                         /* lost the race, try later */
5254                         err = -EBUSY;
5255                         p->rdev = rdev;
5256                 }
5257         }
5258 abort:
5259
5260         print_raid5_conf(conf);
5261         return err;
5262 }
5263
5264 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
5265 {
5266         raid5_conf_t *conf = mddev->private;
5267         int err = -EEXIST;
5268         int disk;
5269         struct disk_info *p;
5270         int first = 0;
5271         int last = conf->raid_disks - 1;
5272
5273         if (mddev->degraded > conf->max_degraded)
5274                 /* no point adding a device */
5275                 return -EINVAL;
5276
5277         if (rdev->raid_disk >= 0)
5278                 first = last = rdev->raid_disk;
5279
5280         /*
5281          * find the disk ... but prefer rdev->saved_raid_disk
5282          * if possible.
5283          */
5284         if (rdev->saved_raid_disk >= 0 &&
5285             rdev->saved_raid_disk >= first &&
5286             conf->disks[rdev->saved_raid_disk].rdev == NULL)
5287                 disk = rdev->saved_raid_disk;
5288         else
5289                 disk = first;
5290         for ( ; disk <= last ; disk++)
5291                 if ((p=conf->disks + disk)->rdev == NULL) {
5292                         clear_bit(In_sync, &rdev->flags);
5293                         rdev->raid_disk = disk;
5294                         err = 0;
5295                         if (rdev->saved_raid_disk != disk)
5296                                 conf->fullsync = 1;
5297                         rcu_assign_pointer(p->rdev, rdev);
5298                         break;
5299                 }
5300         print_raid5_conf(conf);
5301         return err;
5302 }
5303
5304 static int raid5_resize(mddev_t *mddev, sector_t sectors)
5305 {
5306         /* no resync is happening, and there is enough space
5307          * on all devices, so we can resize.
5308          * We need to make sure resync covers any new space.
5309          * If the array is shrinking we should possibly wait until
5310          * any io in the removed space completes, but it hardly seems
5311          * worth it.
5312          */
5313         sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5314         md_set_array_sectors(mddev, raid5_size(mddev, sectors,
5315                                                mddev->raid_disks));
5316         if (mddev->array_sectors >
5317             raid5_size(mddev, sectors, mddev->raid_disks))
5318                 return -EINVAL;
5319         set_capacity(mddev->gendisk, mddev->array_sectors);
5320         revalidate_disk(mddev->gendisk);
5321         if (sectors > mddev->dev_sectors && mddev->recovery_cp == MaxSector) {
5322                 mddev->recovery_cp = mddev->dev_sectors;
5323                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5324         }
5325         mddev->dev_sectors = sectors;
5326         mddev->resync_max_sectors = sectors;
5327         return 0;
5328 }
5329
5330 static int check_stripe_cache(mddev_t *mddev)
5331 {
5332         /* Can only proceed if there are plenty of stripe_heads.
5333          * We need a minimum of one full stripe,, and for sensible progress
5334          * it is best to have about 4 times that.
5335          * If we require 4 times, then the default 256 4K stripe_heads will
5336          * allow for chunk sizes up to 256K, which is probably OK.
5337          * If the chunk size is greater, user-space should request more
5338          * stripe_heads first.
5339          */
5340         raid5_conf_t *conf = mddev->private;
5341         if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5342             > conf->max_nr_stripes ||
5343             ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5344